diff --git a/.flocks/flockshub/index.json b/.flocks/flockshub/index.json
index 7ac3a1117..3c07841ed 100644
--- a/.flocks/flockshub/index.json
+++ b/.flocks/flockshub/index.json
@@ -14640,7 +14640,7 @@
       "name": "SOC Workspace WebUI",
       "description": "SOC workspace pages for posture, overview, and alert investigation.",
       "descriptionCn": "SOC 工作区页面，包含态势、SOC 总览和告警调查。",
-      "version": "1.0.0",
+      "version": "1.1.4",
       "category": "workflow-automation",
       "tags": [
         "siem",
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/access/soc_alerts_operations.py b/.flocks/flockshub/plugins/webuis/soc_ui/access/soc_alerts_operations.py
index 463212ca4..d8fd8b430 100644
--- a/.flocks/flockshub/plugins/webuis/soc_ui/access/soc_alerts_operations.py
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/access/soc_alerts_operations.py
@@ -84,6 +84,8 @@
         "_source_type",
         "net_type",
         "direction",
+        "threat_severity",
+        "threat_level",
         "threat_name",
         "threat_type",
         "threat_phase",
@@ -228,7 +230,7 @@ def _incident_from_row(row: InternalDataRow) -> dict[str, Any]:
     record = row.raw
     record_id = _record_id(record)
     observed_at = _observed_at(record)
-    threat_name = _first_text(record, "threat_name", "_threat_type", "report_title") or "SOC alert"
+    threat_name = _first_text(record, "threat_name", "report_title") or "SOC alert"
     threat_msg = _first_text(record, "threat_msg", "report_title", "threat_type")
     verdict = _verdict_bucket(record)
     table_cells = _table_cells(record)
@@ -425,7 +427,7 @@ def _verdict_bucket(record: dict[str, Any]) -> str:
         return "success"
     raw = " ".join(
         _text(record.get(key)).lower()
-        for key in ("attack_verdict", "threat_result", "risk_level", "threat_level")
+        for key in ("attack_verdict", "threat_result")
     )
     if any(marker in raw for marker in ("success", "attack_success", "succeeded")):
         return "success"
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/manifest.json b/.flocks/flockshub/plugins/webuis/soc_ui/manifest.json
index c0f50cefc..21a8d74fa 100644
--- a/.flocks/flockshub/plugins/webuis/soc_ui/manifest.json
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/manifest.json
@@ -5,7 +5,7 @@
   "name": "SOC Workspace WebUI",
   "description": "SOC workspace pages for posture, overview, and alert investigation.",
   "descriptionCn": "SOC 工作区页面，包含态势、SOC 总览和告警调查。",
-  "version": "1.0.0",
+  "version": "1.1.4",
   "author": "Flocks Team",
   "license": "MIT",
   "homepage": "",
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/soc_alerts/src/filterValues.ts b/.flocks/flockshub/plugins/webuis/soc_ui/soc_alerts/src/filterValues.ts
new file mode 100644
index 000000000..9ae0a4f8c
--- /dev/null
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/soc_alerts/src/filterValues.ts
@@ -0,0 +1,94 @@
+type Translate = (text: string) => string;
+
+const identityTr: Translate = (text) => text;
+
+const FILTER_VALUE_TEXT: Record<string, Record<string, string>> = {
+  direction: {
+    in: '入站',
+    inbound: '入站',
+    ingress: '入站',
+    out: '出站',
+    outbound: '出站',
+    egress: '出站',
+    lateral: '横向',
+  },
+  threat_severity: {
+    critical: '严重',
+    severe: '严重',
+    high: '高危',
+    medium: '中危',
+    low: '低危',
+    info: '信息',
+    informational: '信息',
+  },
+  threat_level: {
+    critical: '严重',
+    severe: '严重',
+    high: '高危',
+    medium: '中危',
+    low: '低危',
+    info: '信息',
+    informational: '信息',
+  },
+  threat_phase: {
+    recon: '侦察',
+    reconnaissance: '侦察',
+    access: '访问',
+    initial_access: '初始访问',
+    execution: '执行',
+    persistence: '持久化',
+    privilege_escalation: '权限提升',
+    defense_evasion: '防御规避',
+    credential_access: '凭据访问',
+    discovery: '发现',
+    lateral_movement: '横向移动',
+    collection: '收集',
+    command_and_control: '命令与控制',
+    exfiltration: '数据渗出',
+    impact: '影响',
+    exploit: '利用',
+    exploitation: '利用',
+  },
+  threat_result: {
+    attack_success: '攻击成功',
+    success: '成功',
+    succeeded: '成功',
+    attack_failed: '攻击失败',
+    failed: '失败',
+    blocked: '已阻断',
+    detected: '已检测',
+    attack: '攻击行为',
+    benign: '安全',
+    safe: '安全',
+    normal: '正常',
+    unknown: '未知',
+  },
+};
+
+function normalized(value: string) {
+  return value.trim().toLowerCase();
+}
+
+function localizedValueText(key: string, value: string) {
+  return FILTER_VALUE_TEXT[key]?.[normalized(value)] || '';
+}
+
+export function filterOptionText(key: string, value: string, tr: Translate = identityTr) {
+  if (key === 'rsp_status_code') return value || tr('未知响应');
+  if (key === '_source_type' || key === 'net_type') return value || 'unknown';
+  const localizedText = localizedValueText(key, value);
+  return localizedText ? tr(localizedText) : value || tr('空值');
+}
+
+export function matchesFilterOptionSearch(
+  key: string,
+  value: string,
+  search: string,
+  tr: Translate = identityTr,
+) {
+  const query = normalized(search);
+  if (!query) return true;
+  const localizedText = localizedValueText(key, value);
+  return [value, localizedText, localizedText ? tr(localizedText) : '']
+    .some((candidate) => normalized(candidate).includes(query));
+}
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/soc_alerts/src/index.tsx b/.flocks/flockshub/plugins/webuis/soc_ui/soc_alerts/src/index.tsx
index ab12870ab..c681d4aa9 100644
--- a/.flocks/flockshub/plugins/webuis/soc_ui/soc_alerts/src/index.tsx
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/soc_alerts/src/index.tsx
@@ -1,8 +1,9 @@
 import { Fragment, useCallback, useEffect, useMemo, useRef, useState } from 'react';
 import { api } from '@flocks/webui-contract-sdk';
+import { filterOptionText, matchesFilterOptionSearch } from './filterValues';
 
 type Tone = 'red' | 'orange' | 'blue' | 'green' | 'purple' | 'slate';
-type FilterKey = '_source_type' | 'net_type' | 'direction' | 'threat_name' | 'threat_type' | 'threat_phase' | 'threat_result' | 'rsp_status_code' | 'sip' | 'dport' | 'dip' | 'req_host' | 'threat_rule_id';
+type FilterKey = '_source_type' | 'net_type' | 'direction' | 'threat_severity' | 'threat_level' | 'threat_name' | 'threat_type' | 'threat_phase' | 'threat_result' | 'rsp_status_code' | 'sip' | 'dport' | 'dip' | 'req_host' | 'threat_rule_id';
 type TimeRangeKey = '15m' | '1h' | '2h' | '24h' | 'today' | '7d' | '30d';
 type TimeFilterMode = 'relative' | 'custom';
 type TimePanelTab = 'auto' | 'custom';
@@ -201,6 +202,8 @@ const EN_TEXT: Record<string, string> = {
   '数据源': 'Data Source',
   '协议类型': 'Protocol',
   '流量方向': 'Traffic Direction',
+  '严重等级': 'Severity',
+  '威胁级别': 'Threat Level',
   '威胁名称': 'Threat Name',
   '威胁类型': 'Threat Type',
   '攻击阶段': 'Attack Stage',
@@ -261,6 +264,34 @@ const EN_TEXT: Record<string, string> = {
   '展开趋势图': 'Expand timeline',
   '攻击成功': 'Attack Success',
   '攻击失败': 'Attack Failed',
+  '入站': 'Inbound',
+  '出站': 'Outbound',
+  '横向': 'Lateral',
+  '严重': 'Critical',
+  '高危': 'High',
+  '中危': 'Medium',
+  '低危': 'Low',
+  '信息': 'Informational',
+  '侦察': 'Reconnaissance',
+  '初始访问': 'Initial Access',
+  '执行': 'Execution',
+  '持久化': 'Persistence',
+  '权限提升': 'Privilege Escalation',
+  '防御规避': 'Defense Evasion',
+  '凭据访问': 'Credential Access',
+  '发现': 'Discovery',
+  '横向移动': 'Lateral Movement',
+  '收集': 'Collection',
+  '命令与控制': 'Command and Control',
+  '数据渗出': 'Exfiltration',
+  '影响': 'Impact',
+  '利用': 'Exploitation',
+  '成功': 'Success',
+  '失败': 'Failed',
+  '已阻断': 'Blocked',
+  '已检测': 'Detected',
+  '安全': 'Benign',
+  '正常': 'Normal',
   '未知': 'Unknown',
   '暂无可展示的告警数据。': 'No alerts to display.',
   '显示 {start}-{end} / {total} 条，每页 {pageSize} 条': 'Showing {start}-{end} / {total}, {pageSize} per page',
@@ -356,6 +387,8 @@ const BASE_FILTER_CONFIGS: FilterConfig[] = [
   { key: '_source_type', label: '数据源' },
   { key: 'net_type', label: '协议类型' },
   { key: 'direction', label: '流量方向' },
+  { key: 'threat_severity', label: '严重等级' },
+  { key: 'threat_level', label: '威胁级别' },
   { key: 'threat_name', label: '威胁名称' },
 ];
 
@@ -377,6 +410,8 @@ const DEFAULT_FILTER_VALUES: Record<FilterKey, string[]> = {
   _source_type: ['tdp'],
   net_type: ['http'],
   direction: [],
+  threat_severity: [],
+  threat_level: [],
   threat_name: [],
   threat_type: [],
   threat_phase: [],
@@ -990,9 +1025,7 @@ function readFilterValue(incident: IncidentCluster, key: FilterKey) {
 }
 
 function optionText(key: FilterKey, value: string, tr: Translate = identityTr) {
-  if (key === 'rsp_status_code') return value || tr('未知响应');
-  if (key === '_source_type' || key === 'net_type') return value || 'unknown';
-  return value || tr('空值');
+  return filterOptionText(key, value, tr);
 }
 
 function optionLabel(key: FilterKey, value: string | string[], tr: Translate = identityTr) {
@@ -1118,7 +1151,7 @@ function FilterDropdown({
   }, [open, value]);
 
   const choices = options.filter((option) => option && option !== ALL_FILTER_VALUE);
-  const visibleChoices = choices.filter((choice) => optionText(config.key, choice, tr).toLowerCase().includes(search.trim().toLowerCase()));
+  const visibleChoices = choices.filter((choice) => matchesFilterOptionSearch(config.key, choice, search, tr));
   const selected = new Set(draft.map(normalized));
 
   const toggleChoice = (choice: string) => {
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/api/handlers.py b/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/api/handlers.py
index 575d978a0..fb0fe6bc7 100644
--- a/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/api/handlers.py
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/api/handlers.py
@@ -21,7 +21,7 @@
 FACTS_TABLE = "soc_dashboard_alert_facts"
 ACTIVITY_TABLE = "soc_dashboard_activity"
 META_TABLE = "soc_dashboard_meta"
-SCHEMA_VERSION = "2"
+SCHEMA_VERSION = "3"
 ACTIVITY_DEFAULT_LIMIT = 20
 ACTIVITY_MAX_LIMIT = 50
 ACTIVITY_WINDOW_MS = 3000
@@ -66,6 +66,7 @@ class _RecordSource:
     "event_time",
     "source_type",
     "threat_name",
+    "threat_type",
     "is_duplicate",
     "phase",
     "direction",
@@ -97,6 +98,7 @@ def _fact_expressions(prefix):
     source_type_fallback = _json_value(prefix, "source_type")
     threat_name = _json_value(prefix, "threat_name")
     threat_type = _json_value(prefix, "_threat_type")
+    threat_type_fallback = _json_value(prefix, "threat_type")
     phase = _json_value(prefix, "threat_phase")
     attack_phase = _json_value(prefix, "attack_phase")
     kill_chain = _json_value(prefix, "kill_chain_phase")
@@ -108,7 +110,6 @@ def _fact_expressions(prefix):
     app_protocol = _json_value(prefix, "net_app_proto")
     protocol_fallback = _json_value(prefix, "protocol")
     severity = _json_value(prefix, "threat_severity")
-    threat_level = _json_value(prefix, "threat_level")
     risk_level = _json_value(prefix, "risk_level")
     response = _json_value(prefix, "rsp_status_code")
     status_code = _json_value(prefix, "status_code")
@@ -134,15 +135,15 @@ def _fact_expressions(prefix):
         f"{prefix}.event_time",
         f"COALESCE(NULLIF({prefix}.source_type, ''), NULLIF({source_type}, ''), "
         f"NULLIF({source_type_fallback}, ''), 'unknown')",
-        f"COALESCE(NULLIF({prefix}.threat_name, ''), NULLIF({threat_name}, ''), "
-        f"NULLIF({threat_type}, ''), 'unknown')",
+        f"COALESCE(NULLIF({prefix}.threat_name, ''), NULLIF({threat_name}, ''), 'unknown')",
+        f"COALESCE(NULLIF({threat_type}, ''), NULLIF({threat_type_fallback}, ''), 'unknown')",
         f"COALESCE({prefix}.is_duplicate, 0)",
         f"COALESCE(NULLIF({phase}, ''), NULLIF({attack_phase}, ''), NULLIF({kill_chain}, ''), 'unknown')",
         f"COALESCE(NULLIF({direction}, ''), NULLIF({traffic_direction}, ''), 'unknown')",
         f"COALESCE(NULLIF({result}, ''), NULLIF({verdict}, ''), 'unknown')",
         f"COALESCE(NULLIF({protocol}, ''), NULLIF({app_protocol}, ''), "
         f"NULLIF({protocol_fallback}, ''), 'unknown')",
-        f"COALESCE(NULLIF({severity}, ''), NULLIF({threat_level}, ''), NULLIF({risk_level}, ''), 'unknown')",
+        f"COALESCE(NULLIF({severity}, ''), 'unknown')",
         f"COALESCE(NULLIF({response}, ''), NULLIF({status_code}, ''), 'unknown')",
         f"COALESCE(NULLIF({destination_port}, ''), NULLIF({destination_port_fallback}, ''), "
         f"NULLIF({destination_port_legacy}, ''), 'unknown')",
@@ -151,7 +152,7 @@ def _fact_expressions(prefix):
         f"COALESCE({triage_status}, '')",
         f"COALESCE({triage_source}, '')",
         f"COALESCE({verdict}, 'unknown')",
-        f"COALESCE({risk_level}, {threat_level}, {severity}, 'unknown')",
+        f"COALESCE(NULLIF({risk_level}, ''), 'unknown')",
         f"COALESCE(CAST({triage_ms} AS INTEGER), 0)",
         f"CASE WHEN {attack_success} IN (1, '1', 'true') THEN 1 ELSE 0 END",
     )
@@ -360,6 +361,7 @@ def _ensure_sqlite_schema():
                     event_time INTEGER,
                     source_type TEXT,
                     threat_name TEXT,
+                    threat_type TEXT,
                     is_duplicate INTEGER NOT NULL DEFAULT 0,
                     phase TEXT,
                     direction TEXT,
@@ -911,10 +913,7 @@ def _get_workflow_recent_events(
         raw_count = metrics["rawCount"]
         unique_count = metrics["uniqueCount"]
         threat_name = str(
-            preview.get("threat_name")
-            or preview.get("_threat_type")
-            or preview.get("threat_type")
-            or f"降噪批次 · 原始 {raw_count} 条"
+            preview.get("threat_name") or f"降噪批次 · 原始 {raw_count} 条"
         )
         events.append(
             {
@@ -1374,12 +1373,12 @@ def _activity_event(row):
         "alert": {
             "id": _first_activity_text(record, "id", "record_id", "uuid", "event_id", "dedup_key"),
             "sourceType": _first_activity_text(record, "_source_type", "source_type", "device_type"),
-            "threatName": _first_activity_text(record, "threat_name", "_threat_type", "threat_type") or "未知告警",
+            "threatName": _first_activity_text(record, "threat_name") or "未知告警",
             "srcIp": _first_activity_text(record, "sip", "src_ip", "source_ip"),
             "dstIp": _first_activity_text(record, "dip", "dst_ip", "destination_ip"),
             "requestUri": _first_activity_text(record, "req_http_url", "uri", "url"),
             "threatPhase": _first_activity_text(record, "threat_phase"),
-            "threatType": _first_activity_text(record, "threat_type", "_threat_type"),
+            "threatType": _first_activity_text(record, "_threat_type", "threat_type"),
         },
     }
     if stage == "denoise":
@@ -1396,7 +1395,8 @@ def _activity_event(row):
             "durationMs": _safe_int(record.get("triage_ms")),
             "verdict": verdict,
             "verdictLabel": RESULT_LABELS.get(verdict, "待确认"),
-            "riskLevel": _first_activity_text(record, "risk_level", "threat_level"),
+            "threatSeverity": _first_activity_text(record, "threat_severity"),
+            "riskLevel": _first_activity_text(record, "risk_level"),
             "reportTitle": _first_activity_text(record, "report_title"),
             "hasReport": bool(record.get("triage_report")),
         }
@@ -1623,7 +1623,14 @@ def _get_stats(params):
             {"key": "benign", "label": "良性", "value": triage["benign"], "color": "#58a6ff"},
             {"key": "unknown", "label": "未知", "value": triage["unknown"], "color": "#9b8cff"},
         ],
-        "topThreats": _counter_items(triage["threatCounter"] or denoise["threatCounter"], 14),
+        "topThreatTypes": _counter_items(
+            triage["threatTypeCounter"] or denoise["threatTypeCounter"],
+            14,
+        ),
+        "severityLevels": _counter_items(
+            _profile_counter(denoise, triage, "severityCounter"),
+            8,
+        ),
         "riskLevels": _counter_items(triage["riskCounter"], 5),
         "timeline": {
             "labels": denoise.get("_timelineLabels")
@@ -1844,7 +1851,7 @@ def _read_denoise(paths, workflow_call_count: int = 0):
     parse_errors = 0
     headers = []
     source_counter = Counter()
-    threat_counter = Counter()
+    threat_type_counter = Counter()
     profile_counters = _new_profile_counters()
     event_start = None
     event_end = None
@@ -1865,7 +1872,7 @@ def _read_denoise(paths, workflow_call_count: int = 0):
             if obj.get("is_duplicate") is True:
                 file_duplicates += 1
             source_counter[_norm(obj.get("_source_type") or obj.get("source_type") or obj.get("device_type"))] += 1
-            threat_counter[_norm(obj.get("_threat_type") or obj.get("threat_name") or obj.get("threat_type"))] += 1
+            threat_type_counter[_norm(obj.get("_threat_type") or obj.get("threat_type"))] += 1
             _update_profile_counters(obj, profile_counters)
             event_start, event_end = _merge_record_time(event_start, event_end, obj)
         total_raw += file_raw
@@ -1889,7 +1896,7 @@ def _read_denoise(paths, workflow_call_count: int = 0):
         "eventStart": _format_event_time(event_start),
         "eventEnd": _format_event_time(event_end),
         "sourceCounter": source_counter,
-        "threatCounter": threat_counter,
+        "threatTypeCounter": threat_type_counter,
         **profile_counters,
         "seriesRaw": series_raw,
         "seriesUnique": series_unique,
@@ -1928,7 +1935,7 @@ def _read_sqlite_denoise(paths, workflow_call_count):
                 f"WHERE {where_clause} GROUP BY \"source_type\"",
                 query_params,
             ).fetchall()
-            profile_counters, threat_counter = _sqlite_detail_counters(
+            profile_counters, threat_type_counter = _sqlite_detail_counters(
                 conn,
                 settings,
                 where_clause,
@@ -1969,7 +1976,7 @@ def _read_sqlite_denoise(paths, workflow_call_count):
         "eventStart": _format_event_time(min(event_values) if event_values else None),
         "eventEnd": _format_event_time(max(event_values) if event_values else None),
         "sourceCounter": Counter({_norm(key): _safe_int(value) for key, value in source_rows}),
-        "threatCounter": threat_counter,
+        "threatTypeCounter": threat_type_counter,
         **profile_counters,
         "seriesRaw": timeline["raw"],
         "seriesUnique": timeline["unique"],
@@ -2025,8 +2032,12 @@ def _sqlite_timeline(conn, settings, where_clause, query_params, dates, start_ti
         f"COUNT(*) AS raw_count, "
         f"COALESCE(SUM(CASE WHEN is_duplicate = 0 THEN 1 ELSE 0 END), 0) AS unique_count, "
         f"COALESCE(SUM(has_triage), 0) AS triage_count, "
-        f"COALESCE(SUM(CASE WHEN has_triage = 1 AND LOWER(verdict) IN "
-        f"('attack_success', 'attack', 'attack_failed') THEN 1 ELSE 0 END), 0) AS attack_count "
+        f"COALESCE(SUM(CASE WHEN has_triage = 1 "
+        f"AND LOWER(triage_status) NOT IN ('failed', 'error') "
+        f"AND (LOWER(verdict) IN ('attack_success', 'attack', 'attack_failed') "
+        f"OR (attack_success = 1 AND LOWER(verdict) NOT IN "
+        f"('attack_success', 'attack', 'attack_failed', 'benign'))) "
+        f"THEN 1 ELSE 0 END), 0) AS attack_count "
         f"FROM {settings['facts_table']} WHERE {where_clause} AND event_time IS NOT NULL "
         f"GROUP BY bucket_index ORDER BY bucket_index",
         (bucket_start, bucket_seconds, *query_params),
@@ -2079,12 +2090,12 @@ def _sqlite_detail_counters(conn, settings, where_clause, query_params):
     ):
         return (
             {key: Counter(value) for key, value in cached["profileCounters"].items()},
-            Counter(cached["threatCounter"]),
+            Counter(cached["threatTypeCounter"]),
         )
 
     rows = conn.execute(
         f"SELECT phase, direction, result, protocol, severity, response_code, "
-        f"port, threat_name, COUNT(*) AS profile_count FROM {table} "
+        f"port, threat_type, COUNT(*) AS profile_count FROM {table} "
         f"WHERE {where_clause} "
         f"GROUP BY 1, 2, 3, 4, 5, 6, 7, 8",
         query_params,
@@ -2098,7 +2109,7 @@ def _sqlite_detail_counters(conn, settings, where_clause, query_params):
         "severityCounter",
         "responseCounter",
     )
-    threat_counter = Counter()
+    threat_type_counter = Counter()
     for row in rows:
         count = _safe_int(row[8])
         for index, key in enumerate(profile_keys):
@@ -2106,19 +2117,19 @@ def _sqlite_detail_counters(conn, settings, where_clause, query_params):
         port_value = row[6]
         port = str(_safe_int(port_value)) if _safe_int(port_value) > 0 else _norm(port_value)
         profile_counters["portCounter"][port] += count
-        threat_counter[_norm(row[7])] += count
+        threat_type_counter[_norm(row[7])] += count
     with _cache_lock:
         _denoise_detail_cache[cache_key] = {
             "lastRowId": latest_row_id,
             "lastTriagePersistedAt": latest_triage_at,
             "updatedAt": time.monotonic(),
             "profileCounters": {key: Counter(value) for key, value in profile_counters.items()},
-            "threatCounter": Counter(threat_counter),
+            "threatTypeCounter": Counter(threat_type_counter),
         }
         _denoise_detail_cache.move_to_end(cache_key)
         while len(_denoise_detail_cache) > _DENOISE_DETAIL_CACHE_MAX:
             _denoise_detail_cache.popitem(last=False)
-    return profile_counters, threat_counter
+    return profile_counters, threat_type_counter
 
 
 def _read_sqlite_triage(paths):
@@ -2135,24 +2146,37 @@ def _read_sqlite_triage(paths):
         where_clause += " AND event_time BETWEEN ? AND ?"
         query_params.extend((start_time, end_time))
     triage_where = f"{where_clause} AND has_triage = 1"
+    cache_condition = "(LOWER(triage_source) = 'cache' OR LOWER(triage_status) = 'cached')"
+    follower_condition = (
+        "(LOWER(triage_source) IN ('follower', 'followers', 'follower_reused') "
+        "OR LOWER(triage_status) = 'follower_reused')"
+    )
+    failed_condition = "LOWER(triage_status) IN ('failed', 'error')"
+    resolved_condition = f"NOT ({failed_condition})"
+    new_triage_condition = (
+        f"NOT {cache_condition} AND NOT {follower_condition} AND NOT ({failed_condition})"
+    )
     try:
         with sqlite3.connect(settings["db_path"]) as conn:
             row = conn.execute(
                 f"SELECT COUNT(*), "
-                f"COALESCE(SUM(CASE WHEN LOWER(triage_source) = 'cache' "
-                f"OR LOWER(triage_status) = 'cached' THEN 1 ELSE 0 END), 0), "
-                f"COALESCE(SUM(CASE WHEN LOWER(triage_source) IN "
-                f"('follower', 'followers', 'follower_reused') "
-                f"OR LOWER(triage_status) = 'follower_reused' THEN 1 ELSE 0 END), 0), "
-                f"COALESCE(SUM(CASE WHEN LOWER(triage_status) IN ('failed', 'error') "
+                f"COALESCE(SUM(CASE WHEN {cache_condition} THEN 1 ELSE 0 END), 0), "
+                f"COALESCE(SUM(CASE WHEN {follower_condition} THEN 1 ELSE 0 END), 0), "
+                f"COALESCE(SUM(CASE WHEN {failed_condition} "
+                f"THEN 1 ELSE 0 END), 0), "
+                f"COALESCE(SUM(CASE WHEN {new_triage_condition} "
+                f"THEN 1 ELSE 0 END), 0), "
+                f"COALESCE(SUM(CASE WHEN {resolved_condition} AND "
+                f"(LOWER(verdict) = 'attack_success' OR (attack_success = 1 AND LOWER(verdict) NOT IN "
+                f"('attack_success', 'attack', 'attack_failed', 'benign'))) THEN 1 ELSE 0 END), 0), "
+                f"COALESCE(SUM(CASE WHEN {resolved_condition} AND LOWER(verdict) = 'attack' "
+                f"THEN 1 ELSE 0 END), 0), "
+                f"COALESCE(SUM(CASE WHEN {resolved_condition} AND LOWER(verdict) = 'attack_failed' "
+                f"THEN 1 ELSE 0 END), 0), "
+                f"COALESCE(SUM(CASE WHEN {resolved_condition} AND LOWER(verdict) = 'benign' "
                 f"THEN 1 ELSE 0 END), 0), "
-                f"COALESCE(SUM(CASE WHEN LOWER(verdict) = 'attack_success' THEN 1 ELSE 0 END), 0), "
-                f"COALESCE(SUM(CASE WHEN LOWER(verdict) = 'attack' THEN 1 ELSE 0 END), 0), "
-                f"COALESCE(SUM(CASE WHEN LOWER(verdict) = 'attack_failed' THEN 1 ELSE 0 END), 0), "
-                f"COALESCE(SUM(CASE WHEN LOWER(verdict) = 'benign' THEN 1 ELSE 0 END), 0), "
-                f"COALESCE(SUM(CASE WHEN LOWER(verdict) NOT IN "
-                f"('attack_success', 'attack', 'attack_failed', 'benign') THEN 1 ELSE 0 END), 0), "
-                f"COALESCE(SUM(CASE WHEN attack_success = 1 AND LOWER(verdict) <> 'attack_success' "
+                f"COALESCE(SUM(CASE WHEN {resolved_condition} AND LOWER(verdict) NOT IN "
+                f"('attack_success', 'attack', 'attack_failed', 'benign') AND attack_success <> 1 "
                 f"THEN 1 ELSE 0 END), 0), MIN(event_time), MAX(event_time), "
                 f"COALESCE(ROUND(AVG(CASE WHEN triage_ms > 0 THEN triage_ms END)), 0) "
                 f"FROM {settings['facts_table']} WHERE {triage_where}",
@@ -2163,9 +2187,9 @@ def _read_sqlite_triage(paths):
                 f"WHERE {triage_where} GROUP BY source_type",
                 query_params,
             ).fetchall()
-            threat_rows = conn.execute(
-                f"SELECT threat_name, COUNT(*) FROM {settings['facts_table']} "
-                f"WHERE {triage_where} GROUP BY threat_name",
+            threat_type_rows = conn.execute(
+                f"SELECT threat_type, COUNT(*) FROM {settings['facts_table']} "
+                f"WHERE {triage_where} GROUP BY threat_type",
                 query_params,
             ).fetchall()
             risk_rows = conn.execute(
@@ -2191,11 +2215,12 @@ def _read_sqlite_triage(paths):
     cache_hit = _safe_int(row[1])
     followers_reused = _safe_int(row[2])
     triage_failed = _safe_int(row[3])
-    attack_success = _safe_int(row[4]) + _safe_int(row[9])
-    attack = _safe_int(row[5])
-    attack_failed = _safe_int(row[6])
-    benign = _safe_int(row[7])
-    unknown = _safe_int(row[8])
+    new_triaged = _safe_int(row[4])
+    attack_success = _safe_int(row[5])
+    attack = _safe_int(row[6])
+    attack_failed = _safe_int(row[7])
+    benign = _safe_int(row[8])
+    unknown = _safe_int(row[9])
     attack_total = attack_success + attack + attack_failed
     avg_triage_ms = _safe_int(row[12])
     profile_counters = _new_profile_counters()
@@ -2217,7 +2242,7 @@ def _read_sqlite_triage(paths):
     return {
         "totalRecords": total_records,
         "batchTotal": 0,
-        "newTriaged": max(total_records - cache_hit - followers_reused - triage_failed, 0),
+        "newTriaged": new_triaged,
         "cacheHit": cache_hit,
         "triageFailed": triage_failed,
         "followersReused": followers_reused,
@@ -2238,7 +2263,9 @@ def _read_sqlite_triage(paths):
         "eventStart": _format_event_time(_parse_event_time(row[10])),
         "eventEnd": _format_event_time(_parse_event_time(row[11])),
         "sourceCounter": Counter({_norm(key): _safe_int(value) for key, value in source_rows}),
-        "threatCounter": Counter({_norm(key): _safe_int(value) for key, value in threat_rows}),
+        "threatTypeCounter": Counter(
+            {_norm(key): _safe_int(value) for key, value in threat_type_rows}
+        ),
         "riskCounter": Counter({_norm(key): _safe_int(value) for key, value in risk_rows}),
         "statusCounter": Counter({_norm(key): _safe_int(value) for key, value in status_rows}),
         **profile_counters,
@@ -2257,7 +2284,7 @@ def _read_triage(paths):
     headers = []
     verdict_counter = Counter()
     source_counter = Counter()
-    threat_counter = Counter()
+    threat_type_counter = Counter()
     risk_counter = Counter()
     status_counter = Counter()
     profile_counters = _new_profile_counters()
@@ -2268,7 +2295,6 @@ def _read_triage(paths):
     fallback_cache = 0
     fallback_failed = 0
     fallback_followers = 0
-    extra_success = 0
     series_total = []
     series_attack = []
     triage_ms_total = 0
@@ -2298,16 +2324,10 @@ def _read_triage(paths):
             verdict = _norm(obj.get("attack_verdict") or "unknown")
             if verdict not in {"attack_success", "attack", "attack_failed", "benign", "unknown"}:
                 verdict = "unknown"
-            verdict_counter[verdict] += 1
-            if obj.get("attack_success") is True and verdict != "attack_success":
-                extra_success += 1
-            if verdict in {"attack_success", "attack", "attack_failed"}:
-                file_attack += 1
-
             source = _norm(obj.get("_source_type") or obj.get("source_type") or obj.get("device_type"))
             source_counter[source] += 1
-            threat_counter[_norm(obj.get("_threat_type") or obj.get("threat_name") or obj.get("threat_type"))] += 1
-            risk_counter[_norm(obj.get("risk_level") or obj.get("threat_level") or obj.get("threat_severity"))] += 1
+            threat_type_counter[_norm(obj.get("_threat_type") or obj.get("threat_type"))] += 1
+            risk_counter[_norm(obj.get("risk_level"))] += 1
             triage_ms = _safe_int(obj.get("triage_ms"))
             if triage_ms > 0:
                 triage_ms_total += triage_ms
@@ -2317,14 +2337,22 @@ def _read_triage(paths):
             triage_source = _norm(obj.get("triage_source"))
             triage_status = _norm(obj.get("triage_status"))
             status_counter[triage_status or triage_source] += 1
+            triage_failed = triage_status in {"failed", "error"}
+
+            if not triage_failed:
+                if verdict == "unknown" and obj.get("attack_success") is True:
+                    verdict = "attack_success"
+                verdict_counter[verdict] += 1
+                if verdict in {"attack_success", "attack", "attack_failed"}:
+                    file_attack += 1
+            else:
+                fallback_failed += 1
 
             if triage_source == "cache" or triage_status == "cached":
                 fallback_cache += 1
             elif triage_source in {"follower", "followers", "follower_reused"} or triage_status == "follower_reused":
                 fallback_followers += 1
-            elif triage_status in {"failed", "error"}:
-                fallback_failed += 1
-            else:
+            elif not triage_failed:
                 fallback_new += 1
 
         series_total.append(file_total)
@@ -2339,7 +2367,7 @@ def _read_triage(paths):
     triage_failed = fallback_failed
     followers_reused = fallback_followers
 
-    attack_success = verdict_counter["attack_success"] + extra_success
+    attack_success = verdict_counter["attack_success"]
     attack = verdict_counter["attack"]
     attack_failed = verdict_counter["attack_failed"]
     attack_total = attack_success + attack + attack_failed
@@ -2373,7 +2401,7 @@ def _read_triage(paths):
         "eventStart": _format_event_time(event_start),
         "eventEnd": _format_event_time(event_end),
         "sourceCounter": source_counter,
-        "threatCounter": threat_counter,
+        "threatTypeCounter": threat_type_counter,
         "riskCounter": risk_counter,
         "statusCounter": status_counter,
         **profile_counters,
@@ -2399,7 +2427,7 @@ def _update_profile_counters(obj, counters):
     counters["directionCounter"][_norm(obj.get("direction") or obj.get("traffic_direction"))] += 1
     counters["resultCounter"][_norm(obj.get("threat_result") or obj.get("attack_verdict"))] += 1
     counters["protocolCounter"][_norm(obj.get("net_type") or obj.get("net_app_proto") or obj.get("protocol"))] += 1
-    counters["severityCounter"][_norm(obj.get("threat_severity") or obj.get("threat_level") or obj.get("risk_level"))] += 1
+    counters["severityCounter"][_norm(obj.get("threat_severity"))] += 1
     counters["responseCounter"][_norm(obj.get("rsp_status_code") or obj.get("status_code"))] += 1
 
     port_value = obj.get("dport") or obj.get("dst_port") or obj.get("destination_port")
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/src/Page.tsx b/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/src/Page.tsx
index 08092a6c4..9ebc10b18 100644
--- a/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/src/Page.tsx
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/src/Page.tsx
@@ -1,3 +1,5 @@
+import { severityKey, severityRows } from './severityValues';
+
 function getSdk() {
   const sdk = globalThis.__FLOCKS_WEBUI_CONTRACT_SDK__;
   if (!sdk || !sdk.React || !sdk.api) {
@@ -74,7 +76,8 @@ const EMPTY_STATS = {
   closedLoop: { autoClosed: 0, resolved: 0, manualDecision: 0, pending: 0, resolutionRate: 0 },
   verdicts: [],
   attackProfile: [],
-  topThreats: [],
+  topThreatTypes: [],
+  severityLevels: [],
   riskLevels: [],
   timeline: { denoiseRaw: [], denoiseUnique: [], triageTotal: [], triageAttack: [] },
 };
@@ -620,7 +623,7 @@ function CenterColumn({ stats, activity }) {
 }
 
 function RightColumn({ stats }) {
-  const threatRows = stats.topThreats || [];
+  const threatRows = stats.topThreatTypes || [];
   const threatRankLabel = threatRows.length ? `Top ${threatRows.length}` : '排行';
   const threatTotal = threatRows.reduce((sum, item) => sum + (item.value || 0), 0);
   const threatBase = stats.triage.totalRecords || stats.triage.attackTotal || threatTotal;
@@ -631,7 +634,7 @@ function RightColumn({ stats }) {
       h('div', { className: 'loop-diagram', key: 'diagram' }, [
         h('div', { className: 'loop-node primary', key: 'valid' }, [h('b', { key: 'v' }, compactNumber(stats.triage.attackTotal)), h('span', { key: 'l' }, '有效事件')]),
         h('div', { className: 'loop-node', key: 'auto' }, [h('b', { key: 'v' }, compactNumber(stats.closedLoop.autoClosed)), h('span', { key: 'l' }, '自动闭环')]),
-        h('div', { className: 'loop-node warn', key: 'manual' }, [h('b', { key: 'v' }, compactNumber(stats.closedLoop.manualDecision)), h('span', { key: 'l' }, '人工决策')]),
+        h('div', { className: 'loop-node warn', key: 'manual' }, [h('b', { key: 'v' }, compactNumber(stats.closedLoop.manualDecision)), h('span', { key: 'l' }, 'AI转人工')]),
         h('div', { className: 'loop-node hot', key: 'pending' }, [h('b', { key: 'v' }, compactNumber(stats.closedLoop.pending)), h('span', { key: 'l' }, '待处理')]),
       ]),
       h(Gauge, { label: '闭环率', value: stats.closedLoop.resolutionRate, color: '#2ee6a6', key: 'gauge' }),
@@ -864,7 +867,11 @@ function AiCore({ stats, activity }) {
       h('small', { key: 'label' }, coreLabel),
       activeEvent ? h('div', { className: 'ai-operation-window', key: `operation-${activeEvent.eventId}` }, [
         h('div', { className: 'ai-operation-track', key: 'track' }, [...operations, operations[0]].map((operation, index) => h('span', { key: `${operation}-${index}` }, operation))),
-      ]) : h('div', { className: 'ai-operation-idle', key: 'operation-idle' }, '等待新的处理任务'),
+      ]) : h('div', {
+        className: 'ai-operation-idle',
+        title: 'AI 新完成且未使用缓存、复用或失败的研判数量',
+        key: 'operation-idle',
+      }, `AI新增研判 ${compactNumber(stats.triage.newTriaged)}`),
     ]),
     activeEvent ? h('div', { className: 'ai-evidence-field', key: `evidence-${activeEvent.eventId}` }, evidenceItems.map((item, index) => h('div', {
       className: `ai-evidence-card evidence-${index + 1}`,
@@ -1375,26 +1382,12 @@ function CommandActivityLane({ kind, lane }) {
   ]);
 }
 
-function severityRows(stats) {
-  return [
-    { key: 'critical', label: '严重', value: stats.triage.attackSuccess || 0, tone: 'critical' },
-    { key: 'high', label: '高危', value: stats.triage.attack || 0, tone: 'high' },
-    { key: 'medium', label: '中危', value: stats.triage.attackFailed || 0, tone: 'medium' },
-    { key: 'low', label: '低危', value: stats.triage.benign || 0, tone: 'low' },
-  ];
-}
-
 function CommandGraph({ stats, activity }) {
   const denoiseActive = Boolean(activity.denoise.current);
   const triageActive = Boolean(activity.triage.current);
   const severityToneFor = (event) => {
     if (!event) return '';
-    if (event.result?.verdict === 'attack_success') return 'critical';
-    const risk = String(event.result?.riskLevel || '').toLowerCase();
-    if (risk === 'high') return 'high';
-    if (risk === 'medium') return 'medium';
-    if (risk === 'low' || event.result?.verdict === 'benign') return 'low';
-    return '';
+    return severityKey(event.result?.threatSeverity);
   };
   const activeSeverityTone = severityToneFor(activity.triage.current);
   const recentSeverityTone = severityToneFor(activity.triage.last);
@@ -1417,9 +1410,27 @@ function CommandGraph({ stats, activity }) {
       h(AiCore, { stats, activity, key: 'sphere' }),
     ]),
     h('div', { className: 'outcome-stack', key: 'outcomes' }, [
-      h('div', { title: 'AI 新完成且未使用缓存、复用或失败的研判数量', key: 'auto' }, [h(AnimatedNumber, { tag: 'b', value: stats.triage.newTriaged, key: 'value' }), h('span', { key: 'label' }, 'AI自主研判')]),
-      h('div', { className: cx('primary', triageActive && 'processing'), title: '研判结论为攻击成功、攻击行为或攻击失败的事件数量', key: 'events' }, [h(AnimatedNumber, { tag: 'b', value: stats.triage.attackTotal, key: 'value' }), h('span', { key: 'label' }, triageActive ? '结果生成中' : '安全事件')]),
-      h('div', { key: 'manual' }, [h(AnimatedNumber, { tag: 'b', value: stats.closedLoop.manualDecision, key: 'value' }), h('span', { key: 'label' }, '人工研判')]),
+      h('div', { className: cx('primary', triageActive && 'processing'), title: 'AI 研判结论为攻击成功、攻击行为或攻击失败的事件数量', key: 'events' }, [
+        h(AnimatedNumber, { tag: 'b', value: stats.triage.attackTotal, key: 'value' }),
+        h('span', { className: cx('outcome-label', triageActive && 'is-processing'), key: 'label' }, [
+          h('i', { key: 'ai' }, triageActive ? 'AI研判' : 'AI判定'),
+          h('em', { key: 'text' }, triageActive ? '处理中' : '安全事件'),
+        ]),
+      ]),
+      h('div', { title: 'AI 研判结论为良性的事件数量', key: 'benign' }, [
+        h(AnimatedNumber, { tag: 'b', value: stats.triage.benign, key: 'value' }),
+        h('span', { className: 'outcome-label', key: 'label' }, [
+          h('i', { key: 'ai' }, 'AI判定'),
+          h('em', { key: 'text' }, '非安全事件'),
+        ]),
+      ]),
+      h('div', { title: 'AI 判定需要进入人工复核的事件数量', key: 'manual' }, [
+        h(AnimatedNumber, { tag: 'b', value: stats.closedLoop.pending, key: 'value' }),
+        h('span', { className: 'outcome-label', key: 'label' }, [
+          h('i', { key: 'ai' }, 'AI判定'),
+          h('em', { key: 'text' }, '待人工复核'),
+        ]),
+      ]),
     ]),
     h('div', { className: 'severity-stack', key: 'severity' }, severities.map((item) => h('div', {
       className: cx(`severity-node severity-${item.tone}`, item.tone === activeSeverityTone && 'active-target', !activeSeverityTone && item.tone === recentSeverityTone && 'recent-target'),
@@ -3746,6 +3757,28 @@ const CSS = `
 .outcome-stack > div { display: flex; flex-direction: column; min-width: 90px; }
 .outcome-stack b { color: #eef1f0; font-size: 23px; line-height: 1; }
 .outcome-stack span { margin-top: 6px; color: #78837f; font-size: 12px; }
+.outcome-stack span.outcome-label {
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  white-space: nowrap;
+}
+.outcome-label i {
+  padding: 1px 4px;
+  border: 1px solid rgba(43,231,255,.3);
+  border-radius: 3px;
+  color: #62cbe8;
+  background: rgba(43,231,255,.07);
+  font-size: 9px;
+  font-style: normal;
+  line-height: 14px;
+}
+.outcome-label em { color: #86a7b8; font-size: 12px; font-style: normal; }
+.outcome-label.is-processing i {
+  border-color: rgba(155,140,255,.38);
+  color: #aa9cff;
+  background: rgba(155,140,255,.1);
+}
 .outcome-stack .primary b { color: #4bdbae; }
 .severity-stack {
   position: absolute;
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/src/severityValues.ts b/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/src/severityValues.ts
new file mode 100644
index 000000000..d2fcd6f2a
--- /dev/null
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/src/severityValues.ts
@@ -0,0 +1,27 @@
+type SeverityItem = {
+  label?: string;
+  value?: number;
+};
+
+export function severityKey(value: unknown) {
+  const key = String(value || '').trim().toLowerCase();
+  if (key === 'critical' || key === 'severe' || key === '4') return 'critical';
+  if (key === 'high' || key === '3') return 'high';
+  if (key === 'medium' || key === '2') return 'medium';
+  if (key === 'low' || key === '1') return 'low';
+  return '';
+}
+
+export function severityRows(stats: { severityLevels?: SeverityItem[] }) {
+  const counts = { critical: 0, high: 0, medium: 0, low: 0 };
+  for (const item of stats.severityLevels || []) {
+    const key = severityKey(item.label);
+    if (key) counts[key] += Number(item.value || 0);
+  }
+  return [
+    { key: 'critical', label: '严重', value: counts.critical, tone: 'critical' },
+    { key: 'high', label: '高危', value: counts.high, tone: 'high' },
+    { key: 'medium', label: '中危', value: counts.medium, tone: 'medium' },
+    { key: 'low', label: '低危', value: counts.low, tone: 'low' },
+  ];
+}
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/soc_overview/api/handlers.py b/.flocks/flockshub/plugins/webuis/soc_ui/soc_overview/api/handlers.py
index f9c70edcc..6cc44b672 100644
--- a/.flocks/flockshub/plugins/webuis/soc_ui/soc_overview/api/handlers.py
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/soc_overview/api/handlers.py
@@ -271,8 +271,8 @@ def _simulate_triage_from_denoise(paths):
             if verdict in {"attack_success", "attack", "attack_failed"}:
                 file_attack += 1
             source_counter[_norm(obj.get("_source_type") or obj.get("source_type") or obj.get("device_type"))] += 1
-            threat_counter[_norm(obj.get("_threat_type") or obj.get("threat_name") or obj.get("threat_type"))] += 1
-            risk_counter[_norm(obj.get("threat_level") or obj.get("threat_severity") or obj.get("risk_level"))] += 1
+            threat_counter[_norm(obj.get("threat_name"))] += 1
+            risk_counter[_norm(obj.get("risk_level"))] += 1
             _update_profile_counters(obj, profile_counters)
             event_start, event_end = _merge_record_time(event_start, event_end, obj)
         series_total.append(file_total)
@@ -627,7 +627,7 @@ def _read_denoise(paths, workflow_call_count: int = 0):
             if obj.get("is_duplicate") is True:
                 file_duplicates += 1
             source_counter[_norm(obj.get("_source_type") or obj.get("source_type") or obj.get("device_type"))] += 1
-            threat_counter[_norm(obj.get("_threat_type") or obj.get("threat_name") or obj.get("threat_type"))] += 1
+            threat_counter[_norm(obj.get("threat_name"))] += 1
             _update_profile_counters(obj, profile_counters)
             event_start, event_end = _merge_record_time(event_start, event_end, obj)
         total_raw += file_raw
@@ -712,8 +712,8 @@ def _read_triage(paths):
 
             source = _norm(obj.get("_source_type") or obj.get("source_type") or obj.get("device_type"))
             source_counter[source] += 1
-            threat_counter[_norm(obj.get("_threat_type") or obj.get("threat_name") or obj.get("threat_type"))] += 1
-            risk_counter[_norm(obj.get("risk_level") or obj.get("threat_level") or obj.get("threat_severity"))] += 1
+            threat_counter[_norm(obj.get("threat_name"))] += 1
+            risk_counter[_norm(obj.get("risk_level"))] += 1
             _update_profile_counters(obj, profile_counters)
             event_start, event_end = _merge_record_time(event_start, event_end, obj)
             triage_source = _norm(obj.get("triage_source"))
@@ -800,7 +800,7 @@ def _update_profile_counters(obj, counters):
     counters["directionCounter"][_norm(obj.get("direction") or obj.get("traffic_direction"))] += 1
     counters["resultCounter"][_norm(obj.get("threat_result") or obj.get("attack_verdict"))] += 1
     counters["protocolCounter"][_norm(obj.get("net_type") or obj.get("net_app_proto") or obj.get("protocol"))] += 1
-    counters["severityCounter"][_norm(obj.get("threat_severity") or obj.get("threat_level") or obj.get("risk_level"))] += 1
+    counters["severityCounter"][_norm(obj.get("threat_severity"))] += 1
     counters["responseCounter"][_norm(obj.get("rsp_status_code") or obj.get("status_code"))] += 1
 
     port_value = obj.get("dport") or obj.get("dst_port") or obj.get("destination_port")
@@ -892,7 +892,7 @@ def _build_field_stats(paths):
             status_code = _field_text(obj.get("rsp_status_code"))
             direction = _norm(obj.get("direction") or "unknown")
             protocol = _norm(obj.get("net_type") or obj.get("net_app_proto") or "unknown")
-            threat_type = _norm(obj.get("threat_type") or obj.get("_threat_type") or obj.get("threat_name"))
+            threat_type = _norm(obj.get("_threat_type") or obj.get("threat_type"))
             threat_result = _norm(obj.get("threat_result") or "unknown")
             threat_phase = _norm(obj.get("threat_phase") or "unknown")
 
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/soc_overview/src/index.tsx b/.flocks/flockshub/plugins/webuis/soc_ui/soc_overview/src/index.tsx
index dafc71b62..cbf4d3600 100644
--- a/.flocks/flockshub/plugins/webuis/soc_ui/soc_overview/src/index.tsx
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/soc_overview/src/index.tsx
@@ -106,24 +106,19 @@ const EMPTY_STATS: Required<Stats> = {
   fieldStats: EMPTY_FIELD_STATS,
 };
 
-const LABELS: Record<string, string> = {
+const PHASE_LABELS: Record<string, string> = {
   exploit: '漏洞利用',
   recon: '侦察探测',
   post_exploit: '后渗透',
   control: '控制通信',
-  tunneling: '隧道通信',
-  file: '文件风险',
-  c2: '控制通信',
-  trojan: '木马',
-  ransom: '勒索',
-  shell: '命令执行',
-  botnet: '僵尸网络',
-  success: '攻击成功',
-  failed: '攻击失败',
-  unknown: '待确认',
+  unknown: '未知阶段',
+};
+
+const DIRECTION_LABELS: Record<string, string> = {
   in: '入站',
   out: '出站',
   lateral: '横向',
+  unknown: '未知方向',
 };
 
 const TIME_RANGE_OPTIONS: ChoiceOption<TimeRangeKey>[] = [
@@ -271,8 +266,8 @@ function refreshLabel(value: RefreshKey) {
   return REFRESH_OPTIONS.find((option) => option.value === value)?.label || '关闭';
 }
 
-function labelOf(item: CounterItem) {
-  return LABELS[item.key || item.label] || LABELS[item.label] || item.label;
+function labelOf(item: CounterItem, labels?: Record<string, string>) {
+  return labels?.[item.key || item.label] || labels?.[item.label] || item.label;
 }
 
 function truncate(value: string, max = 34) {
@@ -415,8 +410,8 @@ export default function SocOverviewPage() {
       <main className="panel-grid">
         <Panel title="TOP 威胁名称"><RankList rows={stats.topThreats.slice(0, 8)} /></Panel>
         <Panel title="威胁类型分布"><TileList rows={stats.fieldStats.threatTypes} /></Panel>
-        <Panel title="攻击阶段"><ProgressList rows={stats.fieldStats.threatPhases} /></Panel>
-        <Panel title="流量方向与响应"><SplitRanks leftTitle="流量方向" leftRows={stats.fieldStats.directions} rightTitle="响应码" rightRows={stats.fieldStats.statusCodes} /></Panel>
+        <Panel title="攻击阶段"><ProgressList rows={stats.fieldStats.threatPhases} labels={PHASE_LABELS} /></Panel>
+        <Panel title="流量方向与响应"><SplitRanks leftTitle="流量方向" leftRows={stats.fieldStats.directions} leftLabels={DIRECTION_LABELS} rightTitle="响应码" rightRows={stats.fieldStats.statusCodes} /></Panel>
         <Panel title="高频 HTTP 主机"><RankList rows={stats.fieldStats.topHosts} mono /></Panel>
         <Panel title="高频地址与规则"><SplitRanks leftTitle="源地址" leftRows={stats.fieldStats.topSourceIps} rightTitle="规则 ID" rightRows={stats.fieldStats.topRules} mono /></Panel>
       </main>
@@ -563,13 +558,13 @@ function Panel({ title, children }: { title: string; children: React.ReactNode }
   );
 }
 
-function RankList({ rows, mono = false }: { rows: CounterItem[]; mono?: boolean }) {
+function RankList({ rows, mono = false, labels }: { rows: CounterItem[]; mono?: boolean; labels?: Record<string, string> }) {
   return (
     <div className="rank-list">
       {rows.length ? rows.map((item, index) => (
         <div key={`${item.label}-${index}`}>
           <span>{String(index + 1).padStart(2, '0')}</span>
-          <b className={mono ? 'mono' : ''} title={item.label}>{truncate(labelOf(item), mono ? 44 : 30)}</b>
+          <b className={mono ? 'mono' : ''} title={item.label}>{truncate(labelOf(item, labels), mono ? 44 : 30)}</b>
           <em>{formatNumber(item.value)}</em>
         </div>
       )) : <div className="empty">暂无数据</div>}
@@ -582,7 +577,7 @@ function TileList({ rows }: { rows: CounterItem[] }) {
     <div className="tile-list">
       {rows.slice(0, 9).map((item) => (
         <div key={item.label}>
-          <b>{labelOf(item)}</b>
+          <b>{item.label}</b>
           <span>{formatNumber(item.value)}</span>
         </div>
       ))}
@@ -591,13 +586,13 @@ function TileList({ rows }: { rows: CounterItem[] }) {
   );
 }
 
-function ProgressList({ rows }: { rows: CounterItem[] }) {
+function ProgressList({ rows, labels }: { rows: CounterItem[]; labels?: Record<string, string> }) {
   const total = Math.max(1, rows.reduce((value, item) => value + item.value, 0));
   return (
     <div className="progress-list">
       {rows.map((item) => (
         <div key={item.label}>
-          <span>{labelOf(item)}</span>
+          <span>{labelOf(item, labels)}</span>
           <i><b style={{ width: `${Math.max(4, (item.value / total) * 100)}%` }} /></i>
           <em>{formatNumber(item.value)}</em>
         </div>
@@ -607,10 +602,10 @@ function ProgressList({ rows }: { rows: CounterItem[] }) {
   );
 }
 
-function SplitRanks({ leftTitle, leftRows, rightTitle, rightRows, mono = false }: { leftTitle: string; leftRows: CounterItem[]; rightTitle: string; rightRows: CounterItem[]; mono?: boolean }) {
+function SplitRanks({ leftTitle, leftRows, leftLabels, rightTitle, rightRows, mono = false }: { leftTitle: string; leftRows: CounterItem[]; leftLabels?: Record<string, string>; rightTitle: string; rightRows: CounterItem[]; mono?: boolean }) {
   return (
     <div className="split-ranks">
-      <div><h3>{leftTitle}</h3><RankList rows={leftRows.slice(0, 4)} mono={mono} /></div>
+      <div><h3>{leftTitle}</h3><RankList rows={leftRows.slice(0, 4)} mono={mono} labels={leftLabels} /></div>
       <div><h3>{rightTitle}</h3><RankList rows={rightRows.slice(0, 4)} mono={mono} /></div>
     </div>
   );
diff --git a/.flocks/flockshub/plugins/webuis/soc_ui/workspace.json b/.flocks/flockshub/plugins/webuis/soc_ui/workspace.json
index 4988a545e..b17ffc05a 100644
--- a/.flocks/flockshub/plugins/webuis/soc_ui/workspace.json
+++ b/.flocks/flockshub/plugins/webuis/soc_ui/workspace.json
@@ -1,5 +1,6 @@
 {
   "id": "soc_ui",
+  "version": "1.1.4",
   "title": "SOC 工作区",
   "titleEn": "SOC Workspace",
   "icon": "ShieldCheck",
diff --git a/.flocks/plugins/skills/sangfor-edr-use/SKILL.md b/.flocks/plugins/skills/sangfor-edr-use/SKILL.md
index 479c8a18a..616a5b9ec 100644
--- a/.flocks/plugins/skills/sangfor-edr-use/SKILL.md
+++ b/.flocks/plugins/skills/sangfor-edr-use/SKILL.md
@@ -1,37 +1,81 @@
 ---
 name: sangfor-edr-use
-description: 用于处理深信服 EDR（终端检测与响应）相关任务，通过 Flocks browser/CDP 完成登录态复用、终端状态查询、概况统计、失陷设备排查和设备运行状态查看。只要用户提到深信服 EDR、EDR 或 sangfor EDR，必须先加载本 skill；本 skill 是 EDR 任务的唯一入口，未阅读前不要直接使用 browser-use。
+description: 深信服 EDR 登录态管理与首页仪表盘 API 采集。用户提到深信服 EDR、EDR 或 sangfor EDR 时必须先加载本 skill。
 ---
 
 # 深信服 EDR Use
 
-本 skill 只负责任务入口和登录分流；CDP 操作、浏览器启动、验证码、selector、tab/iframe 处理和页面数据提取见 [references/cdp-workflow.md](references/cdp-workflow.md)。
+## 核心功能
 
-## 登录分流
+实现边界：`sangfor_edr_http_login.py` 负责 HTTP 登录、认证探测、
+`auth-state.json` Cookie 及 Secret Manager token bundle；
+`sangfor_edr_dashboard_api.py` 负责仪表盘 API 请求，只读取 HTTP 登录模块
+验证过的同一套 Cookie/token，不从其他状态源拼接凭据。
 
-需要打开 EDR 页面、抓取页面请求或调用 EDR API 采集工具时，按以下顺序执行：
+- 管理同一次登录产生的 Cookie 与 `login_token`。
+- 默认使用 HTTP 登录，开始前必须向用户索取并保存 EDR 地址、用户名和密码。
+- 仅当用户明确选择“打开页面后手动登录”时，才可不索取账密并直接使用 browser/CDP。
+- HTTP 登录连续 3 次失败后，按“browser/CDP 自动化登录（仍需账密）→保留页面供用户手动登录（不需账密）”顺序降级。
+- 每次登录或数据采集前探测现有认证，认证有效则跳过登录，失效则重新登录并更新存储。
+- 通过 API 采集首页终端概况、受影响终端、漏洞、勒索防护、实时病毒、Top 5 终端和设备资源使用率。
 
-1. 调用 `sangfor_edr_auth(action=status_auth_state)`，确认 `validation.valid`、`can_auto_refresh` 和 `has_saved_token`。不要先向用户索要账密。
-2. `validation.valid=true`：
-   - 仅浏览器/CDP任务：直接继续；
-   - 需要 API token：`has_saved_token=true` 才继续，否则调用 `refresh_auth_state` 补齐 token。
-3. `validation.valid=false` 且 `can_auto_refresh=true`：调用 `ensure_auth_state` 自动登录。
-4. 没有可用 state 或账密时：调用 `ensure_auth_state` 打开登录页；返回 `manual_login_required` 后，让用户在该工具打开的同一个 EDR tab 中完成登录、MFA 或 UKey。
-5. 用户完成手动登录后，调用 `complete_manual_login`；只有返回 `manual_login_captured_auth_state` 且 `token_saved=true`，才认为浏览器登录和 API 登录准备均完成。
-6. 返回 `browser_daemon_not_ready` 或 `auth_state_load_failed_browser_daemon_not_ready`：依次执行 `flocks browser --setup`、`flocks browser --doctor`，再重试原 action。
+## 输入与输出
 
-认证工具会在登录提交前监听 `launch_login.php` 的 fetch/XHR 响应，将 `data.token` 保存到 Secret Manager；token 不得回显、写入日志或写入 `auth-state.json`。后续 API 工具应从 Secret Manager 读取对应 token，并同时使用同一 EDR 的 cookies。
+### 认证工具
 
-`bu.port` 是 Flocks browser daemon 的 IPC 端口文件，不是 Chrome 的 remote-debugging 端口；禁止手工创建或修改它。
+调用 `sangfor_edr_auth`：
 
-## 任务边界
+- `status_auth_state`：返回认证文件、凭据和认证探测状态，不返回敏感值。
+- `ensure_auth_state`、`refresh_auth_state`、`http_login`：执行“探测后按需 HTTP 登录”；HTTP 连续 3 次失败时按降级策略进入 browser/CDP，再失败则转为手动登录。
+- `browser_login`：用户明确选择自动化登录时，执行“探测后按需 browser/CDP 登录”；该路径需要账密，除非用户明确要求打开页面后自行手动登录。
+- `validate_auth_state`：仅验证浏览器 state。
+- `complete_manual_login`：保存用户在已打开浏览器中完成的登录。
 
-- 当前 `sangfor_edr_v1_0_0` 的页面业务仍通过浏览器/CDP完成；需要 API 采集时，必须先完成 token readiness 检查。
-- 首页仪表盘可用于设备 CPU/内存/硬盘、终端概况和失陷统计。
-- 查询失陷终端清单时，必须进入“威胁资产分析”并选择“已失陷终端”，不能使用默认“全部”列表。
-- 设备状态抓取脚本位于 `references/fetch_edr_system_state.py`；执行前必须完成上述登录分流。
+成功输出包含 `status`、`valid`、`login_skipped` 和非敏感探测结果；失败输出包含稳定的 `reason` 与恢复建议。
+
+### 仪表盘工具
+
+调用 `sangfor_edr_dashboard`，可输入：
+
+- `sections`：可选采集项；省略时采集全部仪表盘数据。
+- `days`：统计时间范围，允许 1–90 天。
+- `base_url`、`auth_state_path`：可选运行时覆盖。
+
+输出包含 `data`、`errors`、`sections`、`days` 和本次认证是复用还是重登；不得输出 Cookie、密码或 `login_token`。
+
+## 关键配置
+
+- `base_url`：从用户提供的 EDR 地址提取 scheme、host 和 port；不得使用固定示例地址。
+- `username`、`password`：从设备配置或 Secret Manager 读取。
+- `auth_state_path`：Cookie state 文件位置。
+- `auto_ocr_code`、`max_captcha_retry`：HTTP/browser 登录验证码配置。
+- selector 和页面路径配置仅用于 browser/CDP 登录。
+
+## EDR 交互协议
+
+1. 从 Secret Manager 与 `auth-state.json` 加载成套认证，并校验 base URL 和 Cookie 指纹。
+2. 使用 Cookie 与 `login_token` 调用威胁终端概览接口 `get_agent_overview`：
+   - HTTP 200、无登录页重定向、响应成功且包含终端概览数据：认证有效，跳过登录。
+   - Cookie 缺失或不匹配、401/403、重定向、非 200、响应无终端概览数据：认证失效。
+3. 默认重新登录流程：访问登录页，获取 RSA 公钥和验证码，提交 `dlogin`，调用 `launch_login.php`，再 GET `/ui`；HTTP 登录连续 3 次失败后进入 browser/CDP 自动化登录，自动化登录仍需账密，自动化登录失败后保留页面供用户手动登录。
+4. 登录成功后将 Cookie 写入 `auth-state.json`，将 `login_token` 写入 Secret Manager，并更新配对指纹。
+5. 仪表盘 API 只能使用通过上述探测的同一套 Cookie/token；禁止从不同 state 或 Secret 拼接。
+
+## 错误处理
+
+- 缺少地址或账密：返回缺失字段，向用户索取后保存到配置或 Secret Manager。
+- 验证码或 HTTP 登录失败：最多进行 3 次独立 HTTP 登录尝试；仍失败则切换 browser/CDP 自动化登录，切换时仍需账密。
+- browser/CDP 自动化登录失败或用户明确选择打开页面后手动登录：保留浏览器供用户完成登录，不再索取账密，再调用 `complete_manual_login`。
+- 认证探测失败：禁止继续业务 API；先执行 HTTP 重登并再次探测，连续 3 次 HTTP 仍失败则按 browser/CDP 自动化登录→手动登录降级。
+- 仪表盘部分接口失败：保留成功数据，在 `errors` 中按采集项返回失败原因。
+- Cookie、密码和 `login_token` 不得回显、记录日志或混入业务输出。
 
 ## 执行约束
 
-- 运行 skill 内 Python 脚本时必须使用 Flocks 虚拟环境；不要使用系统 Python。
-- 需要具体 CDP 命令、平台启动方式、验证码/selector 配置、tab/iframe 处理或页面关键词时，读取 `references/cdp-workflow.md`，不要在本文件重复展开。
+- 运行本 Skill 提供的 Python 脚本时，必须使用 Flocks 虚拟环境；禁止使用系统 Python。
+- 不得假设 Flocks 项目、插件或虚拟环境的绝对路径；代码必须通过当前运行时加载的模块、`Path.home()`、`~/.flocks` 或显式配置/环境变量解析路径。
+- 需要具体 CDP 命令、浏览器启动方式、验证码识别、selector、tab/iframe 处理或页面关键词时，必须先阅读 [references/cdp-workflow.md](references/cdp-workflow.md)，不要在本文件重复展开。
+- `bu.port` 是 Flocks browser daemon 的 IPC 端口文件，不是 Chrome remote-debugging 端口；禁止手工创建或修改。
+- 默认认证和仪表盘采集开始时不得启动 browser daemon；只有用户明确选择 `browser_login`、执行 `validate_auth_state`/`complete_manual_login`，或 HTTP 登录连续 3 次失败进入降级流程时，才允许使用 browser/CDP。
+- HTTP 登录连续 3 次失败后必须按 browser/CDP 自动化登录→手动登录顺序降级；自动化登录阶段仍需账密，手动登录阶段不得要求账密。
+- 任何 API 采集前必须完成认证探测；认证探测失败时不得继续调用业务接口。
diff --git a/.flocks/plugins/tools/device/sangfor_edr_webcli/_provider.yaml b/.flocks/plugins/tools/device/sangfor_edr_webcli/_provider.yaml
index 5a065f634..f750abcfc 100644
--- a/.flocks/plugins/tools/device/sangfor_edr_webcli/_provider.yaml
+++ b/.flocks/plugins/tools/device/sangfor_edr_webcli/_provider.yaml
@@ -4,13 +4,11 @@ service_id: sangfor_edr
 version: "1.0.0"
 integration_type: device
 description: >
-  Sangfor EDR WebCLI-backed integration. It can reuse an existing browser
-  storage state, or refresh auth-state by driving the real EDR login page
-  through browser daemon / CDP when credentials are configured.
+  Sangfor EDR integration with default HTTP login, explicitly selected
+  browser/CDP login, cookie/token validation, and dashboard API collection.
 description_cn: >
-  Sangfor EDR WebCLI device integration. It reuses full browser auth-state
-  from manual login, or refreshes full browser auth-state through CDP-assisted
-  browser login when credentials are configured.
+  深信服 EDR 集成。默认通过 HTTP 登录，仅用户明确选择时使用 browser/CDP；
+  每次操作验证成套 Cookie/token，并通过 API 采集首页仪表盘。
 credential_fields:
   - key: base_url
     label: Base URL
@@ -33,7 +31,7 @@ credential_fields:
     secret_id: sangfor_edr_username
     input_type: text
     required: false
-    description: Optional. Stored for CDP-assisted browser login when auth-state is missing or expired.
+    description: Optional. Stored for default HTTP login and explicitly selected browser/CDP login.
   - key: password
     label: Password
     storage: secret
@@ -41,7 +39,7 @@ credential_fields:
     secret_id: sangfor_edr_password
     input_type: password
     required: false
-    description: Optional. Stored as a secret and used only to refresh browser auth-state.
+    description: Optional. Stored as a secret and used only to refresh EDR authentication.
   - key: auto_ocr_code
     label: Auto OCR Captcha
     storage: config
@@ -114,10 +112,10 @@ defaults:
   category: custom
   verify_ssl: false
 notes: |
-  The existing manual-login flow saves a full browser storageState containing
-  cookies and localStorage. When credentials are available, the automatic flow
-  also drives the real EDR login page through browser daemon / CDP and then
-  saves the same full browser storageState format.
+  All login methods save cookies to auth-state.json and login_token to Secret
+  Manager. Pairing metadata prevents dashboard APIs from mixing credentials
+  produced by different logins. Device URLs are normalized to scheme, host,
+  and port before use.
 
   If captcha OCR, MFA, page selector matching, or login success detection fails,
   keep the browser/manual login recovery path and save the resulting full
diff --git a/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr.handler.py b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr.handler.py
index 2680c140e..7abff1e6e 100644
--- a/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr.handler.py
+++ b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr.handler.py
@@ -1,38 +1,39 @@
-"""Sangfor EDR browser-state authentication helper.
-
-EDR has no stable Open API in this integration.  This handler follows the same
-browser workflow used by TDP / OneSEC / SkyEye / Qingteng skills:
-
-1. try to load the saved browser auth-state;
-2. if it is still valid, reuse it;
-3. if it is missing or expired, open the real EDR login page through CDP;
-4. fill username, password and captcha in the browser page;
-5. after login succeeds, save the full browser auth-state again.
-
-The handler only replaces the "wait for the user to log in manually" step with
-CDP-assisted form login.  It does not implement EDR business APIs.
-"""
+"""Sangfor EDR authentication and dashboard API helpers."""
 
 from __future__ import annotations
 
 import base64
+import hashlib
 import json
 import os
+import secrets
+import sys
 import time
+from datetime import datetime, timedelta
 from pathlib import Path
 from typing import Any, Optional
 from urllib.parse import urljoin, urlparse
 
+import requests
+import urllib3
+
 from flocks.browser import helpers
 from flocks.browser.admin import ensure_daemon
 from flocks.config.config_writer import ConfigWriter
 from flocks.tool.registry import ToolContext, ToolResult
 
+_PLUGIN_DIR = str(Path(__file__).resolve().parent)
+if _PLUGIN_DIR not in sys.path:
+    sys.path.insert(0, _PLUGIN_DIR)
+import sangfor_edr_dashboard_api as _dashboard_api_module  # noqa: E402
+import sangfor_edr_http_login as _http_login_module  # noqa: E402
+
 SERVICE_ID = "sangfor_edr_v1_0_0"
 LEGACY_SERVICE_ID = "sangfor_edr"
 USERNAME_SECRET_ID = "sangfor_edr_username"
 PASSWORD_SECRET_ID = "sangfor_edr_password"
 TOKEN_SECRET_ID = "sangfor_edr_token"
+TOKEN_BUNDLE_SECRET_ID = "sangfor_edr_token_bundle"
 DEFAULT_AUTH_STATE_PATH = "~/.flocks/browser/sangfor-edr/auth-state.json"
 DEFAULT_LOGIN_PATH = "/ui/login.php"
 DEFAULT_INDEX_PATH = "/ui/#/index"
@@ -241,13 +242,26 @@ def _saved_auto_login_status(params: dict[str, Any]) -> dict[str, Any]:
     has_base_url = bool(str(base_url or "").strip())
     has_username = bool(str(username or "").strip())
     has_password = bool(str(password or "").strip())
+    has_saved_token = bool(secrets.get(TOKEN_SECRET_ID))
+    pair_verified = False
+    auth_probe: dict[str, Any] | None = None
+    if has_base_url and auth_state_path.exists():
+        try:
+            cfg = _resolve_runtime_config({**params, "persist_credentials": False})
+            auth_probe = _probe_auth_pair(cfg)
+            pair_verified = bool(auth_probe.get("valid"))
+        except Exception as exc:
+            auth_probe = {"valid": False, "reason": "auth_probe_failed", "error": str(exc)}
+            pair_verified = False
     return {
         "auth_state_path": str(auth_state_path),
         "auth_state_exists": auth_state_path.exists(),
         "has_base_url": has_base_url,
         "has_saved_username": has_username,
         "has_saved_password": has_password,
-        "has_saved_token": bool(secrets.get(TOKEN_SECRET_ID)),
+        "has_saved_token": has_saved_token,
+        "has_verified_auth_pair": pair_verified,
+        "auth_probe": auth_probe,
         "can_auto_refresh": has_base_url and has_username and has_password,
     }
 
@@ -313,6 +327,196 @@ def _now_ms() -> str:
     return str(int(time.time() * 1000))
 
 
+def _query_id() -> str:
+    return f"Query_{_now_ms()}"
+
+
+def _safe_error(exc: Exception, *sensitive_values: str) -> str:
+    message = str(exc)
+    for value in sensitive_values:
+        if value:
+            message = message.replace(value, "<redacted>")
+    return message
+
+
+def _http_headers(cfg: RuntimeConfig, *, image: bool = False) -> dict[str, str]:
+    headers = {
+        "Pragma": "no-cache",
+        "Cache-Control": "no-cache",
+        "User-Agent": (
+            "Mozilla/5.0 (Windows NT 10.0; Win64; x64) "
+            "AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36"
+        ),
+        "Accept-Language": "zh-CN,zh;q=0.9",
+        "Referer": _login_url(cfg) if image else _url(cfg, "/ui/"),
+    }
+    if image:
+        headers["Accept"] = "image/avif,image/webp,image/apng,image/svg+xml,image/*,*/*;q=0.8"
+    else:
+        headers.update(
+            {
+                "Accept": "application/json, text/plain, */*",
+                "Content-Type": "application/json",
+                "Origin": cfg.base_url,
+            }
+        )
+    return headers
+
+
+def _read_auth_state(path: Path) -> dict[str, Any]:
+    state = json.loads(path.read_text(encoding="utf-8"))
+    if not isinstance(state, dict) or not isinstance(state.get("cookies"), list):
+        raise ValueError("EDR auth-state does not contain a cookie list.")
+    return state
+
+
+def _cookie_fingerprint(cookies: list[dict[str, Any]], base_url: str) -> str:
+    normalized = sorted(
+        (
+            str(cookie.get("name") or ""),
+            str(cookie.get("value") or ""),
+            str(cookie.get("domain") or ""),
+            str(cookie.get("path") or "/"),
+        )
+        for cookie in cookies
+        if isinstance(cookie, dict) and cookie.get("name")
+    )
+    payload = json.dumps(
+        {"base_url": _normalise_base_url(base_url), "cookies": normalized},
+        ensure_ascii=False,
+        separators=(",", ":"),
+    )
+    return hashlib.sha256(payload.encode("utf-8")).hexdigest()
+
+
+def _save_auth_pair(cfg: RuntimeConfig, state: dict[str, Any], token: str) -> dict[str, Any]:
+    cookies = state.get("cookies")
+    if not isinstance(cookies, list) or not cookies or not token:
+        raise ValueError("EDR login did not return a complete cookie/token pair.")
+    cfg.auth_state_path.parent.mkdir(parents=True, exist_ok=True)
+    temp_path = cfg.auth_state_path.with_name(f"{cfg.auth_state_path.name}.tmp")
+    temp_path.write_text(json.dumps(state, ensure_ascii=False, indent=2), encoding="utf-8")
+    temp_path.replace(cfg.auth_state_path)
+    fingerprint = _cookie_fingerprint(cookies, cfg.base_url)
+    secret_manager = _get_secret_manager()
+    secret_manager.set(TOKEN_SECRET_ID, token)
+    secret_manager.set(
+        TOKEN_BUNDLE_SECRET_ID,
+        json.dumps(
+            {
+                "token": token,
+                "base_url": cfg.base_url,
+                "cookie_fingerprint": fingerprint,
+            },
+            ensure_ascii=False,
+            separators=(",", ":"),
+        ),
+    )
+    return {
+        "auth_state_path": str(cfg.auth_state_path),
+        "cookie_count": len(cookies),
+        "pair_verified": True,
+    }
+
+
+def _load_verified_auth_pair(cfg: RuntimeConfig) -> tuple[dict[str, Any], str]:
+    state = _read_auth_state(cfg.auth_state_path)
+    raw_bundle = _get_secret_manager().get(TOKEN_BUNDLE_SECRET_ID)
+    if not raw_bundle:
+        raise RuntimeError("EDR cookie/token pair metadata is missing; refresh authentication.")
+    try:
+        bundle = json.loads(raw_bundle)
+    except (TypeError, json.JSONDecodeError) as exc:
+        raise RuntimeError("EDR cookie/token pair metadata is invalid; refresh authentication.") from exc
+    token = str(bundle.get("token") or "")
+    base_url = str(bundle.get("base_url") or "")
+    fingerprint = str(bundle.get("cookie_fingerprint") or "")
+    actual = _cookie_fingerprint(state["cookies"], cfg.base_url)
+    if (
+        not token
+        or _normalise_base_url(base_url) != cfg.base_url
+        or not secrets.compare_digest(fingerprint, actual)
+    ):
+        raise RuntimeError("EDR cookies and login_token are not from the same login; refresh authentication.")
+    return state, token
+
+
+def _response_contains_agent_overview(payload: Any) -> bool:
+    """Return whether a successful threat-terminal response has data content."""
+    if not isinstance(payload, dict) or not payload.get("success"):
+        return False
+    return any(
+        key in payload and payload.get(key) is not None
+        for key in ("data", "result", "agent_overview", "agent_total", "total", "count")
+    )
+
+
+def _unix_date_range(days: int = 7) -> dict[str, int]:
+    end = datetime.now().replace(hour=23, minute=59, second=59, microsecond=0)
+    start = (end - timedelta(days=max(1, days) - 1)).replace(hour=0, minute=0, second=0)
+    return {"start": int(start.timestamp()), "end": int(end.timestamp())}
+
+
+def _probe_auth_pair(cfg: RuntimeConfig) -> dict[str, Any]:
+    try:
+        state, token = _load_verified_auth_pair(cfg)
+    except Exception as exc:
+        return {"valid": False, "reason": "auth_pair_missing_or_mismatched", "error": str(exc)}
+
+    session = _dashboard_session(cfg, state)
+    try:
+        response = session.post(
+            _url(cfg, f"/launch.php?s={token}&opr=get_agent_overview"),
+            headers=_http_headers(cfg),
+            json={
+                "app_args": {"name": "app.web.event_center.head", "options": {}},
+                "auto": 1,
+                "opr": "get_agent_overview",
+                "date_range": _unix_date_range(),
+                "query_id": _query_id(),
+            },
+            timeout=cfg.timeout,
+            allow_redirects=False,
+        )
+    except Exception as exc:
+        return {"valid": False, "reason": "auth_probe_request_failed", "error": _safe_error(exc, token)}
+
+    location = str(response.headers.get("Location") or "")
+    if response.status_code in {301, 302, 303, 307, 308}:
+        return {
+            "valid": False,
+            "reason": "auth_probe_redirected",
+            "http_status": response.status_code,
+            "location": location,
+        }
+    if response.status_code in {401, 403}:
+        return {
+            "valid": False,
+            "reason": "auth_probe_unauthorized",
+            "http_status": response.status_code,
+        }
+    if response.status_code != 200:
+        return {
+            "valid": False,
+            "reason": "auth_probe_http_error",
+            "http_status": response.status_code,
+        }
+    try:
+        payload = response.json()
+    except Exception as exc:
+        return {"valid": False, "reason": "auth_probe_invalid_json", "error": str(exc)}
+    if not isinstance(payload, dict) or not payload.get("success"):
+        return {"valid": False, "reason": "auth_probe_rejected"}
+    if not _response_contains_agent_overview(payload):
+        return {"valid": False, "reason": "auth_probe_expected_agent_data_missing"}
+    return {
+        "valid": True,
+        "reason": "auth_probe_succeeded",
+        "http_status": 200,
+        "agent_overview_verified": True,
+    }
+
+
 def _login_url(cfg: RuntimeConfig) -> str:
     return _url(cfg, cfg.login_path)
 
@@ -415,18 +619,34 @@ def _install_login_token_capture() -> bool:
         return False
 
 
-def _save_captured_login_token(timeout: float = 2.0) -> bool:
+def _captured_login_token(timeout: float = 2.0) -> str:
     deadline = time.time() + timeout
     while time.time() < deadline:
         try:
             token = str(helpers.js("sessionStorage.getItem('__flocks_sangfor_edr_token') || ''") or "").strip()
             if token:
-                _get_secret_manager().set(TOKEN_SECRET_ID, token)
-                return True
+                return token
         except Exception:
-            return False
+            return ""
         time.sleep(0.1)
-    return False
+    return ""
+
+
+def _save_captured_login_token(timeout: float = 2.0) -> bool:
+    token = _captured_login_token(timeout)
+    if not token:
+        return False
+    _get_secret_manager().set(TOKEN_SECRET_ID, token)
+    return True
+
+
+def _save_browser_auth_pair(cfg: RuntimeConfig) -> dict[str, Any]:
+    token = _captured_login_token()
+    if not token:
+        raise RuntimeError("EDR login succeeded but login_token was not captured.")
+    saved = _save_auth_state(cfg)
+    pair = _save_auth_pair(cfg, _read_auth_state(cfg.auth_state_path), token)
+    return {"state": saved, "pair": pair}
 
 
 def _page_text() -> str:
@@ -929,8 +1149,7 @@ def _refresh_auth_state_with_cdp_login(cfg: RuntimeConfig, captcha_code: str = "
 
             filled = _set_login_form_values(cfg, code)
             if _wait_for_login_success(cfg):
-                token_saved = _save_captured_login_token()
-                saved = _save_auth_state(cfg)
+                saved = _save_browser_auth_pair(cfg)
                 return {
                     "success": True,
                     "status": "browser_cdp_login_refreshed_auth_state",
@@ -939,7 +1158,7 @@ def _refresh_auth_state_with_cdp_login(cfg: RuntimeConfig, captcha_code: str = "
                     "form": form_state,
                     "filled": {key: bool(value) for key, value in filled.items()},
                     "saved": saved,
-                    "token_saved": token_saved,
+                    "token_saved": True,
                 }
             last_error = "login_success_check_timeout"
         except Exception as exc:
@@ -967,14 +1186,27 @@ def _complete_manual_login(cfg: RuntimeConfig) -> dict[str, Any]:
             _open_page(_index_url(cfg))
         if not _is_logged_in(cfg):
             return _manual_login_result(cfg, "manual_login_not_completed")
-        token_saved = _save_captured_login_token()
+        saved = _save_browser_auth_pair(cfg)
+        probe = _probe_auth_pair(cfg)
+        if not probe.get("valid"):
+            return {
+                "success": False,
+                "valid": False,
+                "status": "manual_login_capture_probe_failed",
+                "reason": str(probe.get("reason") or "auth_probe_failed"),
+                "error": "browser login state was saved but HTTP authentication probe failed",
+                "saved": saved,
+                "token_saved": True,
+                "probe": probe,
+            }
         return {
             "success": True,
             "valid": True,
             "status": "manual_login_captured_auth_state",
             "auth_state_path": str(cfg.auth_state_path),
-            "saved": _save_auth_state(cfg),
-            "token_saved": token_saved,
+            "saved": saved,
+            "token_saved": True,
+            "probe": probe,
         }
     except Exception as exc:
         return {
@@ -988,6 +1220,99 @@ def _complete_manual_login(cfg: RuntimeConfig) -> dict[str, Any]:
         }
 
 
+def _dashboard_session(cfg: RuntimeConfig, state: dict[str, Any]) -> requests.Session:
+    session = requests.Session()
+    session.verify = False
+    host = urlparse(cfg.base_url).hostname or ""
+    for cookie in state["cookies"]:
+        if not isinstance(cookie, dict) or not cookie.get("name"):
+            continue
+        domain = str(cookie.get("domain") or host)
+        session.cookies.set(
+            str(cookie["name"]),
+            str(cookie.get("value") or ""),
+            domain=domain,
+            path=str(cookie.get("path") or "/"),
+        )
+    urllib3.disable_warnings(urllib3.exceptions.InsecureRequestWarning)
+    return session
+
+
+def _dashboard_requests(
+    cfg: RuntimeConfig,
+    token: str,
+    *,
+    days: int,
+) -> dict[str, tuple[str, dict[str, Any]]]:
+    end = datetime.now().replace(hour=23, minute=59, second=59, microsecond=0)
+    start = (end - timedelta(days=days - 1)).replace(hour=0, minute=0, second=0)
+    unix_range = _unix_date_range(days)
+    text_range = {
+        "start": start.strftime("%Y-%m-%d %H:%M:%S"),
+        "end": end.strftime("%Y-%m-%d %H:%M:%S"),
+    }
+
+    def launch(opr: str, app_name: str, extra: Optional[dict[str, Any]] = None):
+        payload = {
+            "app_args": {"name": app_name, "options": {}},
+            "auto": 1,
+            "opr": opr,
+            "query_id": _query_id(),
+        }
+        payload.update(extra or {})
+        return f"/launch.php?s={token}&opr={opr}", payload
+
+    return {
+        "agent_overview": launch(
+            "get_agent_overview",
+            "app.web.event_center.head",
+            {"date_range": unix_range},
+        ),
+        "influenced_agent_overview": launch(
+            "get_influenced_agent_overview",
+            "app.web.event_center.head",
+        ),
+        "vulnerability_overview": (
+            f"/api/edrgoweb/v1/vulner/list/homepageVulner?_method=post&s={token}&req_type=polling",
+            {
+                "recentCheckTime": unix_range,
+                "uuid": "sf-id-6",
+                "tid": "0",
+                "uid": cfg.username,
+                "token": token,
+            },
+        ),
+        "ransomware_defense": launch(
+            "ransom_virus_defense_interface",
+            "app.web.event_center.ransom_virus_protection",
+        ),
+        "realtime_virus": launch(
+            "real_time_virus",
+            "app.web.event_center.real_time_events",
+            {"date": text_range, "day_sum": days},
+        ),
+        "top_agents": launch(
+            "get_top5_agents",
+            "app.web.event_center.head",
+            {"day_sum": days, "date_range": unix_range},
+        ),
+        "resource_usage": launch(
+            "get_realtime_resource_usage",
+            "app.web.event_center.head",
+        ),
+    }
+
+
+def _collect_dashboard(
+    cfg: RuntimeConfig,
+    *,
+    sections: list[str],
+    days: int,
+) -> dict[str, Any]:
+    """Compatibility entry point; dashboard API implementation is standalone."""
+    return _dashboard_api_module.collect_dashboard(cfg, sections=sections, days=days)
+
+
 # ── Tool actions ─────────────────────────────────────────────────────────────
 
 def _auth_state_loaded_output(validation: dict[str, Any]) -> dict[str, Any]:
@@ -1003,27 +1328,12 @@ async def handle(ctx: ToolContext) -> ToolResult:
     action = str(params.get("action") or "ensure_auth_state").strip()
 
     try:
-        if action == "status_auth_state":
-            status = _saved_auto_login_status(params)
-            validation: dict[str, Any] | None = None
-            if status.get("has_base_url"):
-                try:
-                    validation = _validate_auth_state(_resolve_runtime_config({**params, "persist_credentials": False}))
-                except Exception as exc:
-                    validation = {
-                        "valid": False,
-                        "reason": "auth_state_validate_failed",
-                        "error": str(exc),
-                        "auth_state_path": status["auth_state_path"],
-                    }
+        if action in {"status_auth_state", "ensure_auth_state", "refresh_auth_state", "http_login"}:
+            result = _http_login_module.run_auth_action(params)
             return ToolResult(
-                success=True,
-                output={
-                    "success": True,
-                    "status": "saved_auto_login_status",
-                    **status,
-                    "validation": validation,
-                },
+                success=bool(result.get("success")),
+                output=result,
+                error=None if result.get("success") else str(result.get("error") or result.get("reason")),
             )
 
         cfg = _resolve_runtime_config(params)
@@ -1044,26 +1354,68 @@ async def handle(ctx: ToolContext) -> ToolResult:
                 error=None if result.get("success") else result.get("reason"),
             )
 
+        if action == "browser_login":
+            probe = _probe_auth_pair(cfg)
+            if probe.get("valid"):
+                result = {
+                    "success": True,
+                    "valid": True,
+                    "status": "browser_login_skipped_valid_auth_pair",
+                    "login_skipped": True,
+                    "probe": probe,
+                }
+            else:
+                result = _refresh_auth_state_with_cdp_login(
+                    cfg,
+                    captcha_code=str(params.get("captcha_code") or ""),
+                )
+                result["previous_probe"] = probe
+                if result.get("success"):
+                    confirmation = _probe_auth_pair(cfg)
+                    result["probe"] = confirmation
+                    if not confirmation.get("valid"):
+                        result.update(
+                            {
+                                "success": False,
+                                "valid": False,
+                                "status": "browser_login_probe_failed",
+                                "reason": str(confirmation.get("reason") or "auth_probe_failed"),
+                            }
+                        )
+            return ToolResult(
+                success=bool(result.get("success")),
+                output=result,
+                error=None if result.get("success") else str(result.get("error") or result.get("reason")),
+            )
+
         if action not in {"ensure_auth_state", "refresh_auth_state"}:
             return ToolResult(
                 success=False,
-                error="Unsupported Sangfor EDR auth action. Use status_auth_state, ensure_auth_state, validate_auth_state, refresh_auth_state, or complete_manual_login.",
+                error=(
+                    "Unsupported Sangfor EDR auth action. Use status_auth_state, "
+                    "ensure_auth_state, validate_auth_state, refresh_auth_state, "
+                    "http_login, browser_login, or complete_manual_login."
+                ),
             )
 
-        force_refresh = action == "refresh_auth_state" or _coerce_bool(params.get("force_refresh"), default=False)
-        if not force_refresh:
-            validation = _validate_auth_state(cfg)
-            if validation.get("valid"):
-                return ToolResult(success=True, output=_auth_state_loaded_output(validation))
-
-        result = _refresh_auth_state_with_cdp_login(
-            cfg,
-            captcha_code=str(params.get("captcha_code") or ""),
+        result = _http_login_module.run_auth_action(params)
+        return ToolResult(
+            success=bool(result.get("success")),
+            output=result,
+            error=None if result.get("success") else str(result.get("error") or result.get("reason")),
         )
+    except Exception as exc:
+        return ToolResult(success=False, error=str(exc))
+
+
+async def handle_dashboard(ctx: ToolContext) -> ToolResult:
+    params = dict(ctx.params)
+    try:
+        result = _dashboard_api_module.run_dashboard(params)
         return ToolResult(
             success=bool(result.get("success")),
             output=result,
-            error=None if result.get("success") else result.get("reason"),
+            error=None if result.get("success") else "dashboard_api_partial_failure",
         )
     except Exception as exc:
         return ToolResult(success=False, error=str(exc))
diff --git a/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_auth.yaml b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_auth.yaml
index 1b3d060ba..3aed97339 100644
--- a/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_auth.yaml
+++ b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_auth.yaml
@@ -1,15 +1,14 @@
 name: sangfor_edr_auth
 description: >
-  Manage Sangfor EDR browser authentication state. Use status_auth_state or
-  validate_auth_state before opening pages; use ensure_auth_state to reuse a
-  valid saved auth-state or automatically refresh full browser state through
-  browser daemon / CDP when credentials are configured. Browser actions start
-  or recover the daemon automatically and return structured recovery guidance.
+  Manage a matched Sangfor EDR cookie/login_token pair. ensure_auth_state and
+  refresh_auth_state use HTTP login by default. browser_login is used only when
+  the user explicitly selects browser/CDP automation. The implementation is in
+  sangfor_edr_http_login.py; every action probes the saved authentication first
+  and skips login while it remains valid.
 description_cn: >
-  管理深信服 EDR 浏览器登录态。先用 status_auth_state 或 validate_auth_state
-  检查已保存 state；当 state 缺失或失效且已配置账密时，ensure_auth_state
-  通过 browser daemon / CDP 驱动真实登录页自动登录并重新保存完整浏览器 state。
-  需要浏览器的 action 会自动启动或恢复 daemon，失败时返回结构化恢复指引。
+  管理深信服 EDR 成套的 Cookie 与 login_token。ensure_auth_state 和
+  refresh_auth_state 默认使用 HTTP 登录；仅用户明确选择 browser_login 时
+  使用 browser/CDP 自动化。每次操作先探测现有认证，有效时跳过登录。
 category: custom
 enabled: true
 requires_confirmation: false
@@ -19,13 +18,14 @@ inputSchema:
   properties:
     action:
       type: string
-      enum: [status_auth_state, ensure_auth_state, validate_auth_state, refresh_auth_state, complete_manual_login]
+      enum: [status_auth_state, ensure_auth_state, validate_auth_state, refresh_auth_state, http_login, browser_login, complete_manual_login]
       default: ensure_auth_state
       description: >
         Authentication action. status_auth_state returns non-sensitive saved
         state/credential availability; ensure_auth_state reuses existing state
         when it appears valid and refreshes when needed; validate_auth_state
-        checks only; refresh_auth_state always performs CDP-assisted browser login;
+        checks only; refresh_auth_state uses HTTP login without browser fallback;
+        http_login and browser_login explicitly select one login method;
         complete_manual_login validates a user-completed login and saves its browser state.
     captcha_code:
       type: string
@@ -39,6 +39,16 @@ inputSchema:
     password:
       type: string
       description: Optional password to save as a secret before running auth.
+    auto_ocr_code:
+      type: boolean
+      default: true
+      description: Whether to recognize the HTTP/browser captcha with OCR when captcha_code is omitted.
+    max_captcha_retry:
+      type: integer
+      minimum: 1
+      maximum: 10
+      default: 5
+      description: Maximum number of HTTP/browser login attempts when captcha validation fails.
     auth_state_path:
       type: string
       description: Optional auth-state path. Defaults to ~/.flocks/browser/sangfor-edr/auth-state.json.
@@ -61,10 +71,6 @@ inputSchema:
       type: string
       default: "#button,input[name='button'],.login-opr-btn,#login,#submit,.login-btn,.btn-login,button[type='submit'],input[type='submit']"
       description: Optional comma-separated selectors for the EDR login submit control.
-    force_refresh:
-      type: boolean
-      default: false
-      description: Force CDP-assisted browser login even if auth-state exists.
   required: [action]
 handler:
   type: script
diff --git a/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_dashboard.yaml b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_dashboard.yaml
new file mode 100644
index 000000000..4a8c27ce3
--- /dev/null
+++ b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_dashboard.yaml
@@ -0,0 +1,46 @@
+name: sangfor_edr_dashboard
+description: >
+  Collect Sangfor EDR home dashboard data through authenticated HTTP APIs.
+  Probes the saved cookie/login_token pair before every collection, skips login
+  when valid, and performs HTTP re-login when missing or expired. API request
+  definitions are implemented in sangfor_edr_dashboard_api.py, which reads the
+  verified pair owned by sangfor_edr_http_login.py.
+description_cn: >
+  通过深信服 EDR HTTP API 采集首页仪表盘。每次采集前探测 Cookie/token，
+  有效时复用，缺失或失效时通过 HTTP 重新登录并更新认证。
+category: custom
+enabled: true
+requires_confirmation: false
+provider: sangfor_edr
+inputSchema:
+  type: object
+  properties:
+    sections:
+      type: array
+      items:
+        type: string
+        enum:
+          - agent_overview
+          - influenced_agent_overview
+          - vulnerability_overview
+          - ransomware_defense
+          - realtime_virus
+          - top_agents
+          - resource_usage
+      description: Optional dashboard sections. Omit to collect every section.
+    days:
+      type: integer
+      minimum: 1
+      maximum: 90
+      default: 7
+      description: Time range in days for dashboard APIs that accept a range.
+    base_url:
+      type: string
+      description: Optional EDR device URL; its scheme, host, and port become the base URL.
+    auth_state_path:
+      type: string
+      description: Optional auth-state path. Defaults to ~/.flocks/browser/sangfor-edr/auth-state.json.
+handler:
+  type: script
+  script_file: sangfor_edr.handler.py
+  function: handle_dashboard
diff --git a/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_dashboard_api.py b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_dashboard_api.py
new file mode 100644
index 000000000..ea6a49a45
--- /dev/null
+++ b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_dashboard_api.py
@@ -0,0 +1,138 @@
+"""Sangfor EDR dashboard API collection using the HTTP auth pair."""
+
+from __future__ import annotations
+
+from datetime import datetime, timedelta
+from typing import Any, Optional
+
+import sangfor_edr_http_login as auth
+
+
+def _dashboard_requests(
+    cfg: auth.RuntimeConfig,
+    token: str,
+    *,
+    days: int,
+) -> dict[str, tuple[str, dict[str, Any]]]:
+    end = datetime.now().replace(hour=23, minute=59, second=59, microsecond=0)
+    start = (end - timedelta(days=max(1, days) - 1)).replace(hour=0, minute=0, second=0)
+    unix_range = auth._unix_date_range(days)
+    text_range = {
+        "start": start.strftime("%Y-%m-%d %H:%M:%S"),
+        "end": end.strftime("%Y-%m-%d %H:%M:%S"),
+    }
+
+    def launch(opr: str, app_name: str, extra: Optional[dict[str, Any]] = None):
+        payload = {
+            "app_args": {"name": app_name, "options": {}},
+            "auto": 1,
+            "opr": opr,
+            "query_id": auth._query_id(),
+        }
+        payload.update(extra or {})
+        return f"/launch.php?s={token}&opr={opr}", payload
+
+    return {
+        "agent_overview": launch(
+            "get_agent_overview",
+            "app.web.event_center.head",
+            {"date_range": unix_range},
+        ),
+        "influenced_agent_overview": launch(
+            "get_influenced_agent_overview",
+            "app.web.event_center.head",
+        ),
+        "vulnerability_overview": (
+            f"/api/edrgoweb/v1/vulner/list/homepageVulner?_method=post&s={token}&req_type=polling",
+            {
+                "recentCheckTime": unix_range,
+                "uuid": "sf-id-6",
+                "tid": "0",
+                "uid": cfg.username,
+                "token": token,
+            },
+        ),
+        "ransomware_defense": launch(
+            "ransom_virus_defense_interface",
+            "app.web.event_center.ransom_virus_protection",
+        ),
+        "realtime_virus": launch(
+            "real_time_virus",
+            "app.web.event_center.real_time_events",
+            {"date": text_range, "day_sum": days},
+        ),
+        "top_agents": launch(
+            "get_top5_agents",
+            "app.web.event_center.head",
+            {"day_sum": days, "date_range": unix_range},
+        ),
+        "resource_usage": launch(
+            "get_realtime_resource_usage",
+            "app.web.event_center.head",
+        ),
+    }
+
+
+def collect_dashboard(
+    cfg: auth.RuntimeConfig,
+    *,
+    sections: list[str],
+    days: int,
+) -> dict[str, Any]:
+    auth_result = auth.ensure_http_auth_pair(cfg)
+    if not auth_result.get("success"):
+        raise RuntimeError(
+            "EDR authentication refresh failed: "
+            f"{auth_result.get('error') or auth_result.get('reason') or auth_result.get('status')}"
+        )
+    state, token = auth.load_verified_auth_pair(cfg)
+    session = auth.dashboard_session(cfg, state)
+    definitions = _dashboard_requests(cfg, token, days=days)
+    selected = sections or list(definitions)
+    unknown = sorted(set(selected) - set(definitions))
+    if unknown:
+        raise ValueError(f"Unsupported EDR dashboard sections: {', '.join(unknown)}")
+
+    data: dict[str, Any] = {}
+    errors: dict[str, str] = {}
+    for section in selected:
+        path, payload = definitions[section]
+        try:
+            response = session.post(
+                auth._url(cfg, path),
+                headers=auth._http_headers(cfg),
+                json=payload,
+                timeout=cfg.timeout,
+            )
+            response.raise_for_status()
+            data[section] = response.json()
+        except Exception as exc:
+            errors[section] = auth._safe_error(exc, token)
+
+    return {
+        "success": not errors,
+        "status": "dashboard_api_collected" if not errors else "dashboard_api_partially_collected",
+        "base_url": cfg.base_url,
+        "days": days,
+        "sections": selected,
+        "data": data,
+        "errors": errors,
+        "auth_pair_verified": True,
+        "authentication": {
+            "status": auth_result.get("status"),
+            "login_skipped": bool(auth_result.get("login_skipped")),
+        },
+    }
+
+
+def run_dashboard(params: dict[str, Any]) -> dict[str, Any]:
+    cfg = auth.resolve_runtime_config({**params, "persist_credentials": False})
+    raw_sections = params.get("sections")
+    if isinstance(raw_sections, str):
+        sections = [item.strip() for item in raw_sections.split(",") if item.strip()]
+    elif isinstance(raw_sections, list):
+        sections = [str(item).strip() for item in raw_sections if str(item).strip()]
+    else:
+        sections = []
+    days = max(1, min(90, auth._coerce_int(params.get("days"), 7)))
+    return collect_dashboard(cfg, sections=sections, days=days)
diff --git a/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_http_login.py b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_http_login.py
new file mode 100644
index 000000000..149dcac01
--- /dev/null
+++ b/.flocks/plugins/tools/device/sangfor_edr_webcli/sangfor_edr_http_login.py
@@ -0,0 +1,523 @@
+"""Standalone Sangfor EDR HTTP authentication and auth-pair storage."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+import secrets
+import time
+from datetime import datetime, timedelta
+from http.cookiejar import CookieJar
+from pathlib import Path
+from typing import Any, Optional
+from urllib.parse import urljoin, urlparse
+
+import requests
+import urllib3
+
+from flocks.config.config_writer import ConfigWriter
+
+SERVICE_ID = "sangfor_edr_v1_0_0"
+LEGACY_SERVICE_ID = "sangfor_edr"
+USERNAME_SECRET_ID = "sangfor_edr_username"
+PASSWORD_SECRET_ID = "sangfor_edr_password"
+TOKEN_SECRET_ID = "sangfor_edr_token"
+TOKEN_BUNDLE_SECRET_ID = "sangfor_edr_token_bundle"
+DEFAULT_AUTH_STATE_PATH = "~/.flocks/browser/sangfor-edr/auth-state.json"
+DEFAULT_LOGIN_PATH = "/ui/login.php"
+DEFAULT_TIMEOUT = 25
+PUBLIC_EXPONENT = 0x10001
+CONFIG_KEYS = (
+    "base_url",
+    "auth_state_path",
+    "auto_ocr_code",
+    "max_captcha_retry",
+    "login_path",
+)
+
+
+class RuntimeConfig:
+    def __init__(
+        self,
+        *,
+        base_url: str,
+        auth_state_path: Path,
+        username: str,
+        password: str,
+        login_path: str,
+        timeout: int,
+        auto_ocr_code: bool,
+        max_captcha_retry: int,
+    ) -> None:
+        self.base_url = base_url
+        self.auth_state_path = auth_state_path
+        self.username = username
+        self.password = password
+        self.login_path = login_path
+        self.timeout = timeout
+        self.auto_ocr_code = auto_ocr_code
+        self.max_captcha_retry = max_captcha_retry
+
+
+def _get_secret_manager():
+    from flocks.security import get_secret_manager
+
+    return get_secret_manager()
+
+
+def _resolve_ref(value: Any) -> Optional[str]:
+    if value is None:
+        return None
+    if not isinstance(value, str):
+        return str(value)
+    if value.startswith("{secret:") and value.endswith("}"):
+        return _get_secret_manager().get(value[8:-1])
+    if value.startswith("{env:") and value.endswith("}"):
+        return os.getenv(value[5:-1])
+    return value
+
+
+def _coerce_bool(value: Any, default: bool = False) -> bool:
+    if value is None:
+        return default
+    if isinstance(value, bool):
+        return value
+    return str(value).strip().lower() in {"1", "true", "yes", "y", "on"}
+
+
+def _coerce_int(value: Any, default: int) -> int:
+    try:
+        return int(str(value).strip())
+    except (TypeError, ValueError):
+        return default
+
+
+def _normalise_base_url(value: str) -> str:
+    candidate = value.strip()
+    if not candidate:
+        raise ValueError("Sangfor EDR base_url is required.")
+    if "://" not in candidate:
+        candidate = f"https://{candidate}"
+    parsed = urlparse(candidate)
+    if parsed.scheme not in {"http", "https"} or not parsed.hostname:
+        raise ValueError(f"Invalid Sangfor EDR base_url: {value!r}")
+    host = parsed.hostname
+    if ":" in host and not host.startswith("["):
+        host = f"[{host}]"
+    port = f":{parsed.port}" if parsed.port else ""
+    return f"{parsed.scheme}://{host}{port}".rstrip("/")
+
+
+def _has_device_context() -> bool:
+    try:
+        from flocks.tool.credential_context import get_active_device_id
+
+        return bool(get_active_device_id())
+    except Exception:
+        return False
+
+
+def _load_service_config() -> dict[str, Any]:
+    primary = ConfigWriter.get_api_service_raw(SERVICE_ID)
+    primary = dict(primary) if isinstance(primary, dict) else {}
+    if _has_device_context():
+        return primary
+    legacy = ConfigWriter.list_api_services_raw()
+    fallback = legacy.get(LEGACY_SERVICE_ID) if isinstance(legacy, dict) else {}
+    for key, value in (fallback or {}).items():
+        if primary.get(key) in (None, "") and value not in (None, ""):
+            primary[key] = value
+    return primary
+
+
+def _save_params_to_service(params: dict[str, Any]) -> dict[str, Any]:
+    service = _load_service_config()
+    persist = _coerce_bool(params.get("persist_credentials"), default=True)
+    for key in CONFIG_KEYS:
+        if params.get(key) not in (None, ""):
+            service[key] = params[key]
+    for key, secret_id in (("username", USERNAME_SECRET_ID), ("password", PASSWORD_SECRET_ID)):
+        value = params.get(key)
+        if isinstance(value, str) and value:
+            if persist:
+                _get_secret_manager().set(secret_id, value)
+                service[key] = f"{{secret:{secret_id}}}"
+            else:
+                service[key] = value
+    if persist and any(params.get(key) not in (None, "") for key in ("base_url", "auth_state_path", "username", "password")):
+        ConfigWriter.set_api_service(SERVICE_ID, service)
+    return service
+
+
+def resolve_runtime_config(params: dict[str, Any]) -> RuntimeConfig:
+    raw = _save_params_to_service(params)
+    secrets_store = _get_secret_manager()
+    base_url = _normalise_base_url(
+        _resolve_ref(raw.get("base_url"))
+        or _resolve_ref(raw.get("host"))
+        or os.getenv("SANGFOR_EDR_BASE_URL")
+        or ""
+    )
+    state_path = Path(
+        _resolve_ref(raw.get("auth_state_path"))
+        or os.getenv("SANGFOR_EDR_AUTH_STATE")
+        or DEFAULT_AUTH_STATE_PATH
+    ).expanduser()
+    username = (
+        _resolve_ref(raw.get("username"))
+        or secrets_store.get(USERNAME_SECRET_ID)
+        or secrets_store.get(f"{SERVICE_ID}_username")
+        or secrets_store.get(f"{LEGACY_SERVICE_ID}_username")
+        or os.getenv("SANGFOR_EDR_USERNAME")
+        or ""
+    ).strip()
+    password = (
+        _resolve_ref(raw.get("password"))
+        or secrets_store.get(PASSWORD_SECRET_ID)
+        or secrets_store.get(f"{SERVICE_ID}_password")
+        or secrets_store.get(f"{LEGACY_SERVICE_ID}_password")
+        or os.getenv("SANGFOR_EDR_PASSWORD")
+        or ""
+    ).strip()
+    return RuntimeConfig(
+        base_url=base_url,
+        auth_state_path=state_path,
+        username=username,
+        password=password,
+        login_path=str(raw.get("login_path") or DEFAULT_LOGIN_PATH),
+        timeout=max(5, _coerce_int(raw.get("timeout"), DEFAULT_TIMEOUT)),
+        auto_ocr_code=_coerce_bool(raw.get("auto_ocr_code"), default=True),
+        max_captcha_retry=max(1, min(10, _coerce_int(raw.get("max_captcha_retry"), 5))),
+    )
+
+
+def _url(cfg: RuntimeConfig, path: str) -> str:
+    return urljoin(cfg.base_url + "/", path.lstrip("/"))
+
+
+def _now_ms() -> str:
+    return str(int(time.time() * 1000))
+
+
+def _query_id() -> str:
+    return f"Query_{_now_ms()}"
+
+
+def _safe_error(exc: Exception, *sensitive_values: str) -> str:
+    message = str(exc)
+    for value in sensitive_values:
+        if value:
+            message = message.replace(value, "<redacted>")
+    return message
+
+
+def _http_headers(cfg: RuntimeConfig, *, image: bool = False) -> dict[str, str]:
+    headers = {
+        "Pragma": "no-cache",
+        "Cache-Control": "no-cache",
+        "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 Chrome/131.0.0.0 Safari/537.36",
+        "Accept-Language": "zh-CN,zh;q=0.9",
+        "Referer": _url(cfg, cfg.login_path) if image else _url(cfg, "/ui/"),
+    }
+    if image:
+        headers["Accept"] = "image/avif,image/webp,image/apng,image/svg+xml,image/*,*/*;q=0.8"
+    else:
+        headers.update({"Accept": "application/json, text/plain, */*", "Content-Type": "application/json", "Origin": cfg.base_url})
+    return headers
+
+
+def http_session(cfg: RuntimeConfig) -> requests.Session:
+    session = requests.Session()
+    session.verify = False
+    session.headers.update({"Accept-Language": "zh-CN,zh;q=0.9"})
+    session.cookies.set("hadSetUkey", "0", domain=urlparse(cfg.base_url).hostname, path="/")
+    urllib3.disable_warnings(urllib3.exceptions.InsecureRequestWarning)
+    return session
+
+
+def _cookie_state(cookies: CookieJar, cfg: RuntimeConfig) -> list[dict[str, Any]]:
+    domain = urlparse(cfg.base_url).hostname or ""
+    result = []
+    for cookie in cookies:
+        item = {
+            "name": cookie.name,
+            "value": cookie.value,
+            "domain": cookie.domain or domain,
+            "path": cookie.path or "/",
+            "secure": bool(cookie.secure),
+            "httpOnly": bool(cookie.has_nonstandard_attr("HttpOnly")),
+            "sameSite": "Lax",
+        }
+        if cookie.expires and cookie.expires > 0:
+            item["expires"] = float(cookie.expires)
+        result.append(item)
+    return result
+
+
+def _cookie_fingerprint(cookies: list[dict[str, Any]], base_url: str) -> str:
+    normalized = sorted(
+        (
+            str(cookie.get("name") or ""),
+            str(cookie.get("value") or ""),
+            str(cookie.get("domain") or ""),
+            str(cookie.get("path") or "/"),
+        )
+        for cookie in cookies
+        if isinstance(cookie, dict) and cookie.get("name")
+    )
+    raw = json.dumps({"base_url": _normalise_base_url(base_url), "cookies": normalized}, separators=(",", ":"))
+    return hashlib.sha256(raw.encode()).hexdigest()
+
+
+def _read_auth_state(path: Path) -> dict[str, Any]:
+    state = json.loads(path.read_text(encoding="utf-8"))
+    if not isinstance(state, dict) or not isinstance(state.get("cookies"), list):
+        raise ValueError("EDR auth-state does not contain a cookie list.")
+    return state
+
+
+def _save_auth_pair(cfg: RuntimeConfig, state: dict[str, Any], token: str) -> dict[str, Any]:
+    cookies = state.get("cookies")
+    if not isinstance(cookies, list) or not cookies or not token:
+        raise ValueError("EDR login did not return a complete cookie/token pair.")
+    cfg.auth_state_path.parent.mkdir(parents=True, exist_ok=True)
+    temp_path = cfg.auth_state_path.with_name(f"{cfg.auth_state_path.name}.tmp")
+    temp_path.write_text(json.dumps(state, ensure_ascii=False, indent=2), encoding="utf-8")
+    temp_path.replace(cfg.auth_state_path)
+    fingerprint = _cookie_fingerprint(cookies, cfg.base_url)
+    manager = _get_secret_manager()
+    manager.set(TOKEN_SECRET_ID, token)
+    manager.set(TOKEN_BUNDLE_SECRET_ID, json.dumps({"token": token, "base_url": cfg.base_url, "cookie_fingerprint": fingerprint}, separators=(",", ":")))
+    return {"auth_state_path": str(cfg.auth_state_path), "cookie_count": len(cookies), "pair_verified": True}
+
+
+def load_verified_auth_pair(cfg: RuntimeConfig) -> tuple[dict[str, Any], str]:
+    state = _read_auth_state(cfg.auth_state_path)
+    raw_bundle = _get_secret_manager().get(TOKEN_BUNDLE_SECRET_ID)
+    if not raw_bundle:
+        raise RuntimeError("EDR cookie/token pair metadata is missing; refresh authentication.")
+    try:
+        bundle = json.loads(raw_bundle)
+    except (TypeError, json.JSONDecodeError) as exc:
+        raise RuntimeError("EDR cookie/token pair metadata is invalid; refresh authentication.") from exc
+    token = str(bundle.get("token") or "")
+    if (
+        not token
+        or _normalise_base_url(str(bundle.get("base_url") or "")) != cfg.base_url
+        or not secrets.compare_digest(str(bundle.get("cookie_fingerprint") or ""), _cookie_fingerprint(state["cookies"], cfg.base_url))
+    ):
+        raise RuntimeError("EDR cookies and login_token are not from the same login; refresh authentication.")
+    return state, token
+
+
+def dashboard_session(cfg: RuntimeConfig, state: dict[str, Any]) -> requests.Session:
+    session = requests.Session()
+    session.verify = False
+    host = urlparse(cfg.base_url).hostname or ""
+    for cookie in state.get("cookies", []):
+        if isinstance(cookie, dict) and cookie.get("name"):
+            session.cookies.set(str(cookie["name"]), str(cookie.get("value") or ""), domain=str(cookie.get("domain") or host), path=str(cookie.get("path") or "/"))
+    urllib3.disable_warnings(urllib3.exceptions.InsecureRequestWarning)
+    return session
+
+
+def _unix_date_range(days: int = 7) -> dict[str, int]:
+    end = datetime.now().replace(hour=23, minute=59, second=59, microsecond=0)
+    start = (end - timedelta(days=max(1, days) - 1)).replace(hour=0, minute=0, second=0)
+    return {"start": int(start.timestamp()), "end": int(end.timestamp())}
+
+
+def _response_contains_agent_overview(payload: Any) -> bool:
+    return isinstance(payload, dict) and bool(payload.get("success")) and any(
+        key in payload and payload.get(key) is not None
+        for key in ("data", "result", "agent_overview", "agent_total", "total", "count")
+    )
+
+
+def probe_auth_pair(cfg: RuntimeConfig) -> dict[str, Any]:
+    try:
+        state, token = load_verified_auth_pair(cfg)
+    except Exception as exc:
+        return {"valid": False, "reason": "auth_pair_missing_or_mismatched", "error": str(exc)}
+    session = dashboard_session(cfg, state)
+    try:
+        response = session.post(
+            _url(cfg, f"/launch.php?s={token}&opr=get_agent_overview"),
+            headers=_http_headers(cfg),
+            json={"app_args": {"name": "app.web.event_center.head", "options": {}}, "auto": 1, "opr": "get_agent_overview", "date_range": _unix_date_range(), "query_id": _query_id()},
+            timeout=cfg.timeout,
+            allow_redirects=False,
+        )
+    except Exception as exc:
+        return {"valid": False, "reason": "auth_probe_request_failed", "error": _safe_error(exc, token)}
+    if response.status_code in {301, 302, 303, 307, 308}:
+        return {"valid": False, "reason": "auth_probe_redirected", "http_status": response.status_code, "location": str(response.headers.get("Location") or "")}
+    if response.status_code in {401, 403}:
+        return {"valid": False, "reason": "auth_probe_unauthorized", "http_status": response.status_code}
+    if response.status_code != 200:
+        return {"valid": False, "reason": "auth_probe_http_error", "http_status": response.status_code}
+    try:
+        payload = response.json()
+    except Exception as exc:
+        return {"valid": False, "reason": "auth_probe_invalid_json", "error": str(exc)}
+    if not isinstance(payload, dict) or not payload.get("success"):
+        return {"valid": False, "reason": "auth_probe_rejected"}
+    if not _response_contains_agent_overview(payload):
+        return {"valid": False, "reason": "auth_probe_expected_agent_data_missing"}
+    return {"valid": True, "reason": "auth_probe_succeeded", "http_status": 200, "agent_overview_verified": True}
+
+
+def _rsa_encrypt_password(rsa_key_hex: str, password: str) -> str:
+    modulus = int(rsa_key_hex.strip().lower().removeprefix("0x"), 16)
+    key_size = (modulus.bit_length() + 7) // 8
+    message = password.encode("utf-8")
+    if key_size < len(message) + 11:
+        raise ValueError("EDR password is too long for the login RSA key.")
+    padding = bytearray()
+    while len(padding) < key_size - len(message) - 3:
+        padding.extend(byte for byte in secrets.token_bytes(key_size) if byte)
+    encoded = b"\x00\x02" + bytes(padding[: key_size - len(message) - 3]) + b"\x00" + message
+    return f"{pow(int.from_bytes(encoded, 'big'), PUBLIC_EXPONENT, modulus):0{key_size * 2}x}"
+
+
+def _ocr_verify_code(image: bytes) -> str:
+    try:
+        import ddddocr
+    except Exception as exc:
+        raise RuntimeError("ddddocr is required for automatic Sangfor EDR captcha recognition.") from exc
+    return ddddocr.DdddOcr(show_ad=False).classification(image).strip()[:4]
+
+
+def _is_captcha_failure(result: dict[str, Any]) -> bool:
+    message = str(result.get("msg") or "")
+    data = result.get("data") if isinstance(result.get("data"), dict) else {}
+    return "验证码" in message or data.get("code") is True
+
+
+def _http_login(cfg: RuntimeConfig, captcha_code: str = "") -> dict[str, Any]:
+    missing = [key for key, value in (("username", cfg.username), ("password", cfg.password)) if not value]
+    if missing:
+        return {"success": False, "status": "http_login_credentials_required", "reason": "missing_http_login_credentials", "missing": missing}
+    session = http_session(cfg)
+    phase = "session_init"
+    try:
+        phase = "login_page"
+        session.get(_url(cfg, cfg.login_path), headers=_http_headers(cfg), timeout=cfg.timeout).raise_for_status()
+        phase = "rsakey"
+        rsa_response = session.post(_url(cfg, "/login"), headers=_http_headers(cfg), json={"opr": "rsakey"}, timeout=cfg.timeout)
+        rsa_response.raise_for_status()
+        rsa_result = rsa_response.json()
+        rsa_key = str(rsa_result.get("key") or "")
+        if not rsa_result.get("success") or not rsa_key:
+            raise RuntimeError("EDR RSA key request failed.")
+        last_error = ""
+        for attempt in range(1, cfg.max_captcha_retry + 1):
+            code = captcha_code.strip()
+            if not code:
+                if not cfg.auto_ocr_code:
+                    return {"success": False, "status": "http_login_captcha_required", "reason": "captcha_code_required"}
+                phase = "captcha"
+                captcha = session.get(_url(cfg, f"/ui/randcode.php?{_now_ms()}"), headers=_http_headers(cfg, image=True), timeout=cfg.timeout)
+                captcha.raise_for_status()
+                code = _ocr_verify_code(captcha.content)
+            phase = "dlogin"
+            login_response = session.post(
+                _url(cfg, "/login"),
+                headers=_http_headers(cfg),
+                json={"opr": "dlogin", "data": {"auth_type": "pwd", "user_name": cfg.username, "code": code, "pwd": _rsa_encrypt_password(rsa_key, cfg.password)}},
+                timeout=cfg.timeout,
+            )
+            login_response.raise_for_status()
+            login_result = login_response.json()
+            if not login_result.get("success") or login_result.get("key") in (None, ""):
+                last_error = str(login_result.get("msg") or "EDR dlogin failed.")
+                if captcha_code or not _is_captcha_failure(login_result):
+                    raise RuntimeError(last_error)
+                continue
+            phase = "launch_login"
+            launch_response = session.post(
+                _url(cfg, "/launch_login.php"),
+                headers=_http_headers(cfg),
+                json={"opr": "dlogin", "app_args": {"name": "app.web.auth.login", "options": {}}, "data": {"key": login_result["key"], "user_aggreement_status": "true"}},
+                timeout=cfg.timeout,
+            )
+            launch_response.raise_for_status()
+            launch_result = launch_response.json()
+            token = str((launch_result.get("data") or {}).get("token") or "")
+            if launch_result.get("success") and token:
+                phase = "ui"
+                session.get(
+                    _url(cfg, "/ui"),
+                    headers=_http_headers(cfg),
+                    timeout=cfg.timeout,
+                ).raise_for_status()
+            cookies = _cookie_state(session.cookies, cfg)
+            if launch_result.get("success") and token and any(cookie["name"].lower() == "sessionid" for cookie in cookies):
+                saved = _save_auth_pair(cfg, {"cookies": cookies, "origins": []}, token)
+                return {"success": True, "valid": True, "status": "http_login_refreshed_auth_state", "attempt": attempt, "token_saved": True, "saved": saved}
+            last_error = str(launch_result.get("msg") or "incomplete EDR cookie/token response")
+            if captcha_code:
+                break
+        raise RuntimeError(last_error or "EDR HTTP login retry limit exceeded.")
+    except Exception as exc:
+        response = getattr(exc, "response", None)
+        status = getattr(response, "status_code", None)
+        suffix = f"; http_status={status}" if status is not None else ""
+        return {"success": False, "valid": False, "status": "http_login_failed", "reason": "http_login_failed", "error": f"phase={phase}{suffix}; detail={_safe_error(exc, cfg.username, cfg.password)}", "phase": phase}
+
+
+def ensure_http_auth_pair(cfg: RuntimeConfig, captcha_code: str = "") -> dict[str, Any]:
+    probe = probe_auth_pair(cfg)
+    if probe.get("valid"):
+        return {"success": True, "valid": True, "status": "http_auth_pair_reused", "login_skipped": True, "probe": probe}
+    result = _http_login(cfg, captcha_code=captcha_code)
+    result["previous_probe"] = probe
+    if result.get("success"):
+        confirmation = probe_auth_pair(cfg)
+        result["probe"] = confirmation
+        if not confirmation.get("valid"):
+            result.update({"success": False, "valid": False, "status": "http_login_probe_failed", "reason": str(confirmation.get("reason") or "http_login_probe_failed")})
+    return result
+
+
+def status_auth_state(params: dict[str, Any]) -> dict[str, Any]:
+    service = _save_params_to_service({**params, "persist_credentials": False})
+    manager = _get_secret_manager()
+    base_url = _resolve_ref(service.get("base_url")) or os.getenv("SANGFOR_EDR_BASE_URL") or ""
+    path = Path(_resolve_ref(service.get("auth_state_path")) or os.getenv("SANGFOR_EDR_AUTH_STATE") or DEFAULT_AUTH_STATE_PATH).expanduser()
+    username = _resolve_ref(service.get("username")) or manager.get(USERNAME_SECRET_ID) or ""
+    password = _resolve_ref(service.get("password")) or manager.get(PASSWORD_SECRET_ID) or ""
+    probe = None
+    if base_url and path.exists():
+        try:
+            probe = probe_auth_pair(resolve_runtime_config({**params, "persist_credentials": False}))
+        except Exception as exc:
+            probe = {"valid": False, "reason": "auth_probe_failed", "error": str(exc)}
+    return {
+        "auth_state_path": str(path),
+        "auth_state_exists": path.exists(),
+        "has_base_url": bool(str(base_url).strip()),
+        "has_saved_username": bool(str(username).strip()),
+        "has_saved_password": bool(str(password).strip()),
+        "has_saved_token": bool(manager.get(TOKEN_SECRET_ID)),
+        "has_verified_auth_pair": bool(probe and probe.get("valid")),
+        "auth_probe": probe,
+        "can_auto_refresh": bool(str(base_url).strip() and str(username).strip() and str(password).strip()),
+    }
+
+
+def run_auth_action(params: dict[str, Any]) -> dict[str, Any]:
+    action = str(params.get("action") or "ensure_auth_state")
+    if action == "status_auth_state":
+        status = status_auth_state(params)
+        return {
+            "success": True,
+            "status": "saved_auto_login_status",
+            **status,
+            "validation": status.get("auth_probe"),
+        }
+    if action not in {"ensure_auth_state", "refresh_auth_state", "http_login"}:
+        raise ValueError(f"Unsupported HTTP EDR auth action: {action}")
+    cfg = resolve_runtime_config(params)
+    return ensure_http_auth_pair(cfg, captcha_code=str(params.get("captcha_code") or ""))
diff --git a/.github/workflows/dispatch-autotest-windows-upgrade.yml b/.github/workflows/dispatch-autotest-windows-upgrade.yml
index 59029c4bf..f300a31b1 100644
--- a/.github/workflows/dispatch-autotest-windows-upgrade.yml
+++ b/.github/workflows/dispatch-autotest-windows-upgrade.yml
@@ -1,17 +1,16 @@
 name: Dispatch autotest Windows upgrade
 
 on:
-  push:
-    branches:
-      - main
+  release:
+    types: [published]
   workflow_dispatch:
     inputs:
       release_tag:
-        description: "Optional Flocks version tag. Defaults to pyproject.toml version."
+        description: "Optional expected version. It must match project.version in pyproject.toml."
         required: false
         type: string
       release_url:
-        description: "Optional source URL. Defaults to the release or commit URL."
+        description: "Optional source URL. Defaults to the published release URL or commit URL."
         required: false
         type: string
       rollback_version:
@@ -48,6 +47,9 @@ jobs:
           INPUT_RELEASE_URL: ${{ inputs.release_url }}
           INPUT_ROLLBACK_VERSION: ${{ inputs.rollback_version }}
           INPUT_RUN_FORCE_FALLBACK: ${{ inputs.run_force_fallback }}
+          RELEASE_EVENT_TAG: ${{ github.event.release.tag_name }}
+          RELEASE_EVENT_URL: ${{ github.event.release.html_url }}
+          RELEASE_EVENT_ASSETS_JSON: ${{ toJson(github.event.release.assets) }}
           GITHUB_SERVER_URL_VALUE: ${{ github.server_url }}
           GITHUB_REPOSITORY_VALUE: ${{ github.repository }}
           GITHUB_SHA_VALUE: ${{ github.sha }}
@@ -64,21 +66,37 @@ jobs:
           PY
           )"
 
-          release_tag="${INPUT_RELEASE_TAG:-}"
-          release_url="${INPUT_RELEASE_URL:-}"
+          requested_release_tag="${INPUT_RELEASE_TAG:-${RELEASE_EVENT_TAG:-}}"
+          release_url="${INPUT_RELEASE_URL:-${RELEASE_EVENT_URL:-}}"
+          installer_asset_name="$(python3 - <<'PY'
+          import json
+          import os
+
+          assets_json = os.environ.get("RELEASE_EVENT_ASSETS_JSON") or "[]"
+          try:
+              assets = json.loads(assets_json)
+          except json.JSONDecodeError:
+              assets = []
+          if not isinstance(assets, list):
+              assets = []
+
+          for asset in assets:
+              name = str(asset.get("name") or "")
+              if name.lower().endswith(".exe"):
+                  print(name)
+                  break
+          PY
+          )"
 
-          if [ -z "$release_tag" ]; then
-            release_tag="$pyproject_version"
+          if [ -n "$requested_release_tag" ] && [ "$requested_release_tag" != "$pyproject_version" ]; then
+            echo "::error::Release version mismatch: requested=$requested_release_tag pyproject=$pyproject_version"
+            exit 1
           fi
+          release_tag="$pyproject_version"
           if [ -z "$release_url" ]; then
             release_url="${GITHUB_SERVER_URL_VALUE}/${GITHUB_REPOSITORY_VALUE}/commit/${GITHUB_SHA_VALUE}"
           fi
 
-          if [ -z "$release_tag" ]; then
-            echo "release_tag is empty" >&2
-            exit 1
-          fi
-
           artifact_time="$(printf "%02d" "${GITHUB_RUN_ATTEMPT_VALUE:-1}")"
           artifact_name="flocks-windows-upgrade-${release_tag}-time${artifact_time}"
 
@@ -88,16 +106,19 @@ jobs:
             echo "rollback_version=${INPUT_ROLLBACK_VERSION:-}"
             echo "run_force_fallback=${INPUT_RUN_FORCE_FALLBACK:-false}"
             echo "artifact_name=$artifact_name"
+            echo "installer_asset_name=$installer_asset_name"
           } >> "$GITHUB_OUTPUT"
 
           {
             echo "### flocks_autotest dispatch payload"
             echo ""
             echo "- Event: ${EVENT_NAME}"
-            echo "- Release tag: ${release_tag}"
+            echo "- Release tag: ${requested_release_tag:-not provided}"
+            echo "- pyproject.toml version: ${release_tag}"
             echo "- Release URL: ${release_url}"
             echo "- Source ref: ${GITHUB_REF_NAME_VALUE}"
             echo "- Source sha: ${GITHUB_SHA_VALUE}"
+            echo "- Installer asset: ${installer_asset_name}"
             echo "- Rollback version: ${INPUT_ROLLBACK_VERSION:-}"
             echo "- Run force fallback: ${INPUT_RUN_FORCE_FALLBACK:-false}"
             echo "- Expected autotest artifact: ${artifact_name}"
@@ -111,6 +132,7 @@ jobs:
           ROLLBACK_VERSION: ${{ steps.payload.outputs.rollback_version }}
           RUN_FORCE_FALLBACK: ${{ steps.payload.outputs.run_force_fallback }}
           ARTIFACT_NAME: ${{ steps.payload.outputs.artifact_name }}
+          INSTALLER_ASSET_NAME: ${{ steps.payload.outputs.installer_asset_name }}
           SOURCE_REPOSITORY: ${{ github.repository }}
           SOURCE_RUN_ID: ${{ github.run_id }}
           SOURCE_SHA: ${{ github.sha }}
@@ -123,11 +145,11 @@ jobs:
           import os
 
           payload = {
-              "event_type": "flocks_main_updated",
+              "event_type": "flocks_release_published",
               "client_payload": {
                   "release_tag": os.environ["RELEASE_TAG"],
                   "release_url": os.environ["RELEASE_URL"],
-                  "installer_asset_name": "",
+                  "installer_asset_name": os.environ.get("INSTALLER_ASSET_NAME", ""),
                   "source_repository": os.environ["SOURCE_REPOSITORY"],
                   "source_run_id": os.environ["SOURCE_RUN_ID"],
                   "source_sha": os.environ["SOURCE_SHA"],
@@ -183,6 +205,7 @@ jobs:
           python3 - <<'PY'
           import json
           import os
+          import re
           import sys
           import time
           import urllib.error
@@ -263,8 +286,15 @@ jobs:
               f"{api_root}/repos/{autotest_repo}/actions/workflows/"
               f"{workflow_ref}/runs?per_page=50"
           )
-          version_title_prefix = f"Windows upgrade {release_tag} "
-          version_artifact_prefix = f"flocks-windows-upgrade-{release_tag}-"
+          version_title_pattern = re.compile(
+              rf"(?<![0-9A-Za-z._+-]){re.escape(release_tag)}"
+              rf"(?![0-9A-Za-z._+-])"
+          )
+          version_artifact_pattern = re.compile(
+              rf"(?<![0-9A-Za-z.]){re.escape(release_tag)}"
+              rf"(?![0-9.]|\+|[-_.](?:rc|alpha|beta|pre|preview|dev|post)(?:[0-9._-]|$))",
+              re.IGNORECASE,
+          )
           dispatch_started = parse_github_time(dispatch_started_at) - timedelta(seconds=30)
           deadline = time.time() + wait_timeout
           source_discovery_deadline = time.time() + 120
@@ -279,7 +309,7 @@ jobs:
               candidates = []
               for run in runs:
                   title = run.get("display_title") or run.get("name") or ""
-                  if not title.startswith(version_title_prefix):
+                  if not version_title_pattern.search(title):
                       continue
                   if run.get("status") != "completed" or run.get("conclusion") != "success":
                       continue
@@ -287,7 +317,13 @@ jobs:
               candidates.sort(key=lambda run: run.get("created_at") or "", reverse=True)
               return candidates[0] if candidates else None
 
-          def find_artifact(run_id, *, exact_name="", name_prefix="", retry_seconds=120):
+          def find_artifact(
+              run_id,
+              *,
+              exact_name="",
+              match_release_version=False,
+              retry_seconds=120,
+          ):
               artifacts_url = (
                   f"{api_root}/repos/{autotest_repo}/actions/runs/{run_id}/artifacts?per_page=100"
               )
@@ -307,13 +343,16 @@ jobs:
                       if not artifact.get("expired")
                       and (
                           (exact_name and (artifact.get("name") or "") == exact_name)
-                          or (name_prefix and (artifact.get("name") or "").startswith(name_prefix))
+                          or (
+                              match_release_version
+                              and version_artifact_pattern.search(artifact.get("name") or "")
+                          )
                       )
                   ]
                   if matching_artifacts or time.time() >= lookup_deadline:
                       break
                   available_text = ", ".join(available_artifact_names) if available_artifact_names else "(none)"
-                  expected_text = exact_name or f"{name_prefix}*"
+                  expected_text = exact_name or f"artifact containing exact version {release_tag}"
                   print(
                       f"Waiting for artifact {expected_text}. Available artifacts: {available_text}",
                       flush=True,
@@ -406,7 +445,7 @@ jobs:
                   fallback_run_id = fallback_run["id"]
                   artifact, available_names = find_artifact(
                       fallback_run_id,
-                      name_prefix=version_artifact_prefix,
+                      match_release_version=True,
                       retry_seconds=120,
                   )
                   if artifact:
@@ -467,7 +506,7 @@ jobs:
               f"- Conclusion: {conclusion}",
               f"- Source run: {source_run_url or 'not found'}",
               f"- Source status: {source_failure or 'ok'}",
-              f"- Version fallback prefix: {version_artifact_prefix}",
+              f"- Version fallback match: exact version {release_tag} anywhere in artifact name",
               f"- Expected artifact: {expected_artifact_name}",
           ]
           if artifact_name:
diff --git a/flocks/acp/agent.py b/flocks/acp/agent.py
index ae2ed6f6f..3025fa1e9 100644
--- a/flocks/acp/agent.py
+++ b/flocks/acp/agent.py
@@ -1334,9 +1334,13 @@ async def _restart_session_state(self, session_id: str) -> None:
         from flocks.session.message import Message
         from flocks.session.features.todo import Todo
         from flocks.session.core.status import SessionStatus
+        from flocks.session.session import Session
 
-        await Message.clear(session_id)
-        await Todo.clear(session_id)
+        async def clear_persisted_state() -> None:
+            await Message.clear(session_id)
+            await Todo.clear(session_id)
+
+        await Session.run_active_write(session_id, clear_persisted_state)
         SessionStatus.clear(session_id)
 
         session = self._session_manager.get(session_id)
diff --git a/flocks/agent/agents/rex/prompt_builder.py b/flocks/agent/agents/rex/prompt_builder.py
index 281234bbc..a4bc0a01c 100644
--- a/flocks/agent/agents/rex/prompt_builder.py
+++ b/flocks/agent/agents/rex/prompt_builder.py
@@ -256,6 +256,11 @@ def _build_command_guidance_section() -> str:
 
 Use `flocks --help` to inspect Flocks CLI commands and usage.
 run_slash_command tool with help command to get the latest slash command guidance.
+
+### Safe Flocks Restart
+
+When you need to restart the running Flocks service yourself, use `flocks restart --server-only`.
+永远不要直接执行 `flocks restart`，这将会导致你杀死自己并且无法自启动。
 </Command_Guidance>"""
 
 
diff --git a/flocks/auth/context.py b/flocks/auth/context.py
index c95d7b6af..7ce35f985 100644
--- a/flocks/auth/context.py
+++ b/flocks/auth/context.py
@@ -10,6 +10,9 @@
 from pydantic import BaseModel, Field
 
 
+API_TOKEN_SERVICE_USER_ID = "api-token-service"
+
+
 class AuthUser(BaseModel):
     """Current authenticated local user."""
 
diff --git a/flocks/auth/service.py b/flocks/auth/service.py
index 3c5d761c0..3db79323d 100644
--- a/flocks/auth/service.py
+++ b/flocks/auth/service.py
@@ -15,7 +15,7 @@
 import aiosqlite
 from pydantic import BaseModel, Field
 
-from flocks.auth.context import AuthUser
+from flocks.auth.context import API_TOKEN_SERVICE_USER_ID, AuthUser
 from flocks.extensions import ensure_callable_methods
 from flocks.storage.storage import Storage
 from flocks.utils.id import Identifier
@@ -256,6 +256,8 @@ async def _create_user_internal(
         normalized_username = username.strip()
         if not normalized_username:
             raise ValueError("用户名不能为空")
+        if normalized_username == API_TOKEN_SERVICE_USER_ID:
+            raise ValueError("该用户名为系统保留用户名")
         if len(password) < 8:
             raise ValueError("密码长度至少 8 位")
 
@@ -635,6 +637,7 @@ async def reassign_orphan_sessions(
                     await Session.update(
                         project_id=session.project_id,
                         session_id=session.id,
+                        allow_inactive=True,
                         owner_user_id=admin_user_id,
                         owner_username=admin_user.username,
                     )
@@ -672,6 +675,7 @@ async def migrate_legacy_sessions_to_admin(cls, admin_user_id: str) -> None:
                 await Session.update(
                     project_id=session.project_id,
                     session_id=session.id,
+                    allow_inactive=True,
                     owner_user_id=admin_user_id,
                     owner_username=admin_username,
                 )
diff --git a/flocks/channel/inbound/dispatcher.py b/flocks/channel/inbound/dispatcher.py
index 44a88f566..1e516c91b 100644
--- a/flocks/channel/inbound/dispatcher.py
+++ b/flocks/channel/inbound/dispatcher.py
@@ -1053,16 +1053,38 @@ async def _get_channel_config(channel_id: str) -> ChannelConfig:
         except Exception:
             return ChannelConfig()
 
+    @staticmethod
+    async def _run_session_loop(binding: Any, loop_callbacks: Any) -> Any:
+        """Run an IM turn with the persisted model preference for its session."""
+        from flocks.session.session import (
+            Session,
+            is_model_auto_session_category,
+        )
+        from flocks.session.session_loop import SessionLoop
+
+        session = await Session.get_by_id(binding.session_id)
+        auto_failover = bool(
+            session
+            and is_model_auto_session_category(
+                getattr(session, "category", "user")
+            )
+            and getattr(session, "model_auto", False)
+        )
+        return await SessionLoop.run(
+            session_id=binding.session_id,
+            agent_name=binding.agent_id,
+            callbacks=loop_callbacks,
+            auto_failover=auto_failover,
+        )
+
     @staticmethod
     async def _run_agent(binding, callbacks: ChannelDeliveryCallbacks) -> None:
         """Run Agent and deliver the final assistant reply."""
         try:
-            from flocks.session.session_loop import SessionLoop
             loop_callbacks = callbacks.to_loop_callbacks()
-            result = await SessionLoop.run(
-                session_id=binding.session_id,
-                agent_name=binding.agent_id,
-                callbacks=loop_callbacks,
+            result = await InboundDispatcher._run_session_loop(
+                binding,
+                loop_callbacks,
             )
 
             if result.last_message:
@@ -1147,11 +1169,9 @@ async def _on_text_delta(delta: str) -> None:
             runner_cbs = RunnerCallbacks(on_text_delta=_on_text_delta)
             loop_callbacks = callbacks.to_loop_callbacks(runner_callbacks=runner_cbs)
 
-            from flocks.session.session_loop import SessionLoop
-            result = await SessionLoop.run(
-                session_id=binding.session_id,
-                agent_name=binding.agent_id,
-                callbacks=loop_callbacks,
+            result = await InboundDispatcher._run_session_loop(
+                binding,
+                loop_callbacks,
             )
 
             final_text = None
@@ -1181,6 +1201,31 @@ async def _append_user_message(
         channel_config: Optional[ChannelConfig] = None,
         model: Optional[dict] = None,
         agent: Optional[str] = None,
+    ) -> None:
+        """Append an inbound message only while its bound session is active."""
+
+        from flocks.session.session import Session
+
+        await Session.run_active_write(
+            session_id,
+            lambda: InboundDispatcher._append_user_message_unchecked(
+                session_id,
+                text,
+                msg,
+                channel_config,
+                model,
+                agent,
+            ),
+        )
+
+    @staticmethod
+    async def _append_user_message_unchecked(
+        session_id: str,
+        text: str,
+        msg: InboundMessage,
+        channel_config: Optional[ChannelConfig] = None,
+        model: Optional[dict] = None,
+        agent: Optional[str] = None,
     ) -> None:
         import mimetypes
         import os
diff --git a/flocks/channel/inbound/session_binding.py b/flocks/channel/inbound/session_binding.py
index c9e49cd2b..3f023a358 100644
--- a/flocks/channel/inbound/session_binding.py
+++ b/flocks/channel/inbound/session_binding.py
@@ -86,8 +86,7 @@ async def resolve_channel_session_owner_kwargs(source_session=None) -> dict[str,
     ``Session.create`` cannot infer the owner from ``current_auth_user``.
     When an existing channel session is being replaced, preserve its owner.
     Otherwise, attach new channel sessions to the local admin if one exists.
-    Installs without local accounts remain ownerless for backward-compatible
-    no-login operation.
+    Installs without local accounts use the explicit system identity.
     """
     owner_user_id = getattr(source_session, "owner_user_id", None) if source_session else None
     owner_username = getattr(source_session, "owner_username", None) if source_session else None
@@ -103,7 +102,12 @@ async def resolve_channel_session_owner_kwargs(source_session=None) -> dict[str,
         from flocks.auth.service import AuthService
 
         if not await AuthService.has_users():
-            return {}
+            from flocks.auth.context import API_TOKEN_SERVICE_USER_ID
+
+            return {
+                "owner_user_id": API_TOKEN_SERVICE_USER_ID,
+                "owner_username": API_TOKEN_SERVICE_USER_ID,
+            }
         users = await AuthService.list_users()
     except Exception as exc:
         log.warn("channel.owner.resolve_failed", {"error": str(exc)})
@@ -288,23 +292,30 @@ async def resolve_or_create(
         existing = await self._find_binding(
             msg.channel_id, msg.account_id, chat_id, thread_id,
         )
+        replaced_session = None
         if existing:
-            # Verify the bound session still exists (user may have deleted it via WebUI)
+            # Archived sessions are immutable history, not live conversation
+            # targets. Replace stale or inactive bindings before the dispatcher
+            # persists the inbound message.
             from flocks.session.session import Session as _Session
-            still_alive = await _Session.get_by_id(existing.session_id)
-            if still_alive:
+            bound_session = await _Session.get_by_id_unfiltered(existing.session_id)
+            if bound_session and bound_session.status == "active":
                 await self._touch(existing.session_id)
                 return existing
-            # Session was deleted — remove stale binding and fall through to create a new one
+            replaced_session = bound_session
             log.info("channel.binding.stale", {
                 "channel": msg.channel_id,
                 "chat_id": chat_id,
                 "old_session_id": existing.session_id,
+                "status": getattr(bound_session, "status", "missing"),
             })
             await self.unbind(existing.session_id)
 
         session_id = await self._create_session(
-            msg, default_agent=default_agent, directory=directory,
+            msg,
+            default_agent=default_agent,
+            directory=directory,
+            source_session=replaced_session,
         )
         now = time.time()
         binding = SessionBinding(
@@ -364,8 +375,11 @@ async def bind_session(
             ValueError: if *session_id* does not exist.
         """
         from flocks.session.session import Session as _Session
-        if not await _Session.get_by_id(session_id):
+        session = await _Session.get_by_id_unfiltered(session_id)
+        if not session:
             raise ValueError(f"Session '{session_id}' not found")
+        if session.status != "active":
+            raise ValueError(f"Session '{session_id}' is not active")
 
         now = time.time()
         binding = SessionBinding(
@@ -556,6 +570,7 @@ async def _create_session(
         msg: InboundMessage,
         default_agent: Optional[str] = None,
         directory: Optional[str] = None,
+        source_session=None,
     ) -> str:
         """Create a new Flocks Session and return its ID.
 
@@ -568,7 +583,7 @@ async def _create_session(
         from flocks.session.session import Session
 
         title = _build_title(msg)
-        owner_kwargs = await resolve_channel_session_owner_kwargs()
+        owner_kwargs = await resolve_channel_session_owner_kwargs(source_session)
         session = await Session.create(
             project_id="channel",
             directory=_resolve_session_directory(directory),
diff --git a/flocks/cli/commands/session.py b/flocks/cli/commands/session.py
index fb13111f3..c23f1384f 100644
--- a/flocks/cli/commands/session.py
+++ b/flocks/cli/commands/session.py
@@ -208,21 +208,20 @@ async def _show_session(session_id: str, project_id: Optional[str]):
 
 @session_app.command("delete")
 def session_delete(
-    session_id: str = typer.Argument(..., help="Session ID to delete"),
+    session_id: str = typer.Argument(..., help="Session ID to permanently delete"),
     project: Optional[str] = typer.Option(
         None, "-p", "--project",
         help="Project ID (uses current project if not specified)"
     ),
     force: bool = typer.Option(
         False, "-f", "--force",
-        help="Skip confirmation prompt"
+        help="Skip permanent deletion confirmation"
     ),
 ):
     """
-    Delete a session
-    
-    This performs a soft delete. The session data is marked as deleted
-    but not permanently removed from storage.
+    Permanently delete a session and its persisted history.
+
+    This operation cannot be restored.
     """
     asyncio.run(_delete_session(session_id, project, force))
 
@@ -241,7 +240,7 @@ async def _delete_session(session_id: str, project_id: Optional[str], force: boo
     # Confirm deletion
     if not force:
         confirm = typer.confirm(
-            f"Delete session '{session.title}'?",
+            f"Permanently delete session '{session.title}' and all of its messages and history?",
             default=False
         )
         if not confirm:
@@ -252,7 +251,7 @@ async def _delete_session(session_id: str, project_id: Optional[str], force: boo
     success = await Session.delete(project_id, session_id)
     
     if success:
-        console.print(f"[green]Deleted session: {session_id}[/green]")
+        console.print(f"[green]Permanently deleted session: {session_id}[/green]")
     else:
         console.print(f"[red]Failed to delete session: {session_id}[/red]")
         raise typer.Exit(1)
@@ -317,9 +316,16 @@ async def _restore_session(session_id: str, project_id: Optional[str]):
         console.print(f"[red]Failed to restore session (not found or not archived): {session_id}[/red]")
         raise typer.Exit(1)
 
-    project_id, _ = resolved
+    project_id, session = resolved
+    if session.parent_id is not None:
+        console.print(f"[red]Restore the root session for this session tree: {session_id}[/red]")
+        raise typer.Exit(1)
 
-    success = await Session.unarchive(project_id, session_id)
+    success = await Session.restore(
+        project_id,
+        session_id,
+        project_owner_id=session.owner_user_id,
+    )
     
     if success:
         console.print(f"[green]Restored session: {session_id}[/green]")
diff --git a/flocks/cli/main.py b/flocks/cli/main.py
index 8fce87905..626b9f5a7 100644
--- a/flocks/cli/main.py
+++ b/flocks/cli/main.py
@@ -41,6 +41,7 @@
     ServiceError,
     resolve_flocks_cli_command,
     restart_all,
+    restart_server,
     runtime_paths,
     show_logs,
     show_status,
@@ -264,6 +265,11 @@ def stop():
 
 @app.command()
 def restart(
+    server_only: bool = typer.Option(
+        False,
+        "--server-only",
+        help="Restart only the backend server without stopping the supervisor daemon",
+    ),
     no_browser: bool = typer.Option(False, "--no-browser", help="Do not open WebUI in a browser"),
     skip_webui_build: bool = typer.Option(
         False,
@@ -281,6 +287,9 @@ def restart(
     Restart Flocks service.
     """
     try:
+        if server_only:
+            restart_server(console)
+            return
         restart_all(
             _restart_service_config(
                 no_browser=no_browser,
diff --git a/flocks/cli/service_manager.py b/flocks/cli/service_manager.py
index cd52842ef..7e90c41c4 100644
--- a/flocks/cli/service_manager.py
+++ b/flocks/cli/service_manager.py
@@ -31,6 +31,7 @@
     read_logs,
     read_supervisor_status,
     request_restart,
+    request_restart_backend,
     request_stop,
     stream_logs,
     supervisor_is_running,
@@ -60,6 +61,7 @@
     "src\\win\\async.c",
     "src/win/async.c",
 )
+WEBUI_BUILD_IGNORED_DIRS = frozenset({"dist", "node_modules", ".vite"})
 WATCHDOG_PID_FILENAME = "watchdog.pid"
 SUPERVISOR_START_TIMEOUT_SECONDS = 180.0
 
@@ -313,6 +315,8 @@ def get_node_major_version() -> int | None:
             check=False,
             capture_output=True,
             text=True,
+            encoding="utf-8",
+            errors="replace",
         )
     except OSError:
         return None
@@ -513,6 +517,8 @@ def _windows_tasklist_process_name(pid: int) -> str | None:
         check=False,
         capture_output=True,
         text=True,
+        encoding="utf-8",
+        errors="replace",
     )
     if completed.returncode != 0:
         return None
@@ -886,6 +892,8 @@ def _process_list_pids() -> list[int]:
             check=False,
             capture_output=True,
             text=True,
+            encoding="utf-8",
+            errors="replace",
         )
     else:
         completed = subprocess.run(
@@ -1105,16 +1113,47 @@ def _build_webui_dist(root: Path, config: ServiceConfig, console) -> None:
             raise ServiceError("WebUI 构建失败。")
 
 
+def _webui_needs_build(webui_dir: Path) -> bool:
+    """Return whether WebUI sources are newer than the production bundle."""
+    index_path = webui_dir / "dist" / "index.html"
+    if not index_path.is_file():
+        return True
+
+    built_at = index_path.stat().st_mtime_ns
+    if webui_dir.stat().st_mtime_ns > built_at:
+        return True
+
+    for current, directories, files in os.walk(webui_dir):
+        directories[:] = [name for name in directories if name not in WEBUI_BUILD_IGNORED_DIRS]
+        current_dir = Path(current)
+        try:
+            if any((current_dir / name).stat().st_mtime_ns > built_at for name in directories + files):
+                return True
+        except FileNotFoundError:
+            # A concurrent source edit is itself sufficient reason to rebuild.
+            return True
+    return False
+
+
 def _ensure_webui_dist(root: Path, config: ServiceConfig, console) -> None:
     """Ensure the FastAPI process can serve the production WebUI bundle."""
     from flocks.server.static_webui import WebUIDistMissingError, ensure_webui_dist_dir
 
+    webui_dir = root / "webui"
     try:
-        ensure_webui_dist_dir()
-        return
+        dist_dir = ensure_webui_dist_dir()
     except WebUIDistMissingError:
         if config.skip_frontend_build:
             raise
+    else:
+        source_dist_dir = webui_dir / "dist"
+        if (
+            config.skip_frontend_build
+            or not (webui_dir / "package.json").is_file()
+            or dist_dir != source_dist_dir.resolve()
+            or not _webui_needs_build(webui_dir)
+        ):
+            return
 
     _build_webui_dist(root, config, console)
     ensure_webui_dist_dir()
@@ -1561,6 +1600,21 @@ def restart_all(config: ServiceConfig, console) -> None:
         _start_all_unlocked(config, console, paths=paths)
 
 
+def restart_server(console) -> None:
+    """Restart only the backend through the running supervisor daemon."""
+    paths = ensure_runtime_dirs()
+    with service_lock(paths):
+        if not supervisor_is_running(paths):
+            raise ServiceError("Flocks daemon 未运行；请执行 `flocks restart` 进行全量重启。")
+        try:
+            status = request_restart_backend(paths=paths)
+        except Exception as error:
+            raise ServiceError(f"Flocks server 重启请求失败：{error}") from error
+        _print_status_payload(status.raw, console, include_daemon_step=False)
+        if not _startup_payload_is_ready(status.raw):
+            raise ServiceError(_startup_failure_message(status.raw))
+
+
 def _print_static_port_migration_hint(config: ServiceConfig, console) -> None:
     """Explain legacy server-port behavior when it differs from public WebUI port."""
     if (
@@ -2105,6 +2159,8 @@ def _run_windows_netstat(port: int) -> str:
         check=False,
         capture_output=True,
         text=True,
+        encoding="utf-8",
+        errors="replace",
     )
     if completed.returncode != 0:
         return ""
diff --git a/flocks/cli/service_process.py b/flocks/cli/service_process.py
index a64d82810..aff0627ba 100644
--- a/flocks/cli/service_process.py
+++ b/flocks/cli/service_process.py
@@ -54,7 +54,7 @@ def probe(self, process: subprocess.Popen | None, host: str, port: int) -> Servi
                 restart=True,
             )
         if not tcp_port_accepts_connections(host, port):
-            return ServiceProbeResult(healthy=False, reason=f"port {port} is not listening", restart=True)
+            return ServiceProbeResult(healthy=False, reason=f"port {port} is not listening")
         return ServiceProbeResult(healthy=True, reason="liveness check passed")
 
 
diff --git a/flocks/cli/service_supervisor.py b/flocks/cli/service_supervisor.py
index 4a615b59a..937558891 100644
--- a/flocks/cli/service_supervisor.py
+++ b/flocks/cli/service_supervisor.py
@@ -27,8 +27,8 @@
 )
 from flocks.cli.service_process import BackendProcessAdapter, ProcessAdapter
 
-SUPERVISOR_CHECK_INTERVAL_SECONDS = 5.0
-SUPERVISOR_HEALTH_FAILURE_THRESHOLD = 2
+SUPERVISOR_CHECK_INTERVAL_SECONDS = 30.0
+SUPERVISOR_HEALTH_FAILURE_THRESHOLD = 10
 SUPERVISOR_BACKOFF_SECONDS = (1.0, 2.0, 5.0, 10.0, 30.0)
 _CLIENT_DISCONNECT_ERRORS = (BrokenPipeError, ConnectionResetError, ConnectionAbortedError)
 
diff --git a/flocks/cli/session_runner.py b/flocks/cli/session_runner.py
index ff57c4b1d..994042259 100644
--- a/flocks/cli/session_runner.py
+++ b/flocks/cli/session_runner.py
@@ -361,7 +361,7 @@ async def _process_message(
             )
 
             async def _clear_history() -> None:
-                await Message.clear(self._session.id)
+                await Message.clear_active(self._session.id)
                 await self._clear_screen()
                 self.console.print("[dim]Conversation history cleared.[/dim]")
 
diff --git a/flocks/config/config.py b/flocks/config/config.py
index a5f234f9c..99fa1f3af 100644
--- a/flocks/config/config.py
+++ b/flocks/config/config.py
@@ -1205,7 +1205,7 @@ async def load_text(cls, text: str, filepath: Path) -> ConfigInfo:
             # Remove comments properly
             # 1. Remove /* */ block comments first
             text_no_comments = re.sub(r'/\*.*?\*/', '', text, flags=re.DOTALL)
-            
+
             # 2. Remove // line comments, but NOT in strings!
             # We need to be careful not to remove // inside quoted strings (like URLs)
             # This regex matches // that are NOT inside quotes
@@ -1218,30 +1218,30 @@ async def load_text(cls, text: str, filepath: Path) -> ConfigInfo:
                 in_string = False
                 escape_next = False
                 comment_start = -1
-                
+
                 for i, char in enumerate(line):
                     if escape_next:
                         escape_next = False
                         continue
-                    
+
                     if char == '\\':
                         escape_next = True
                         continue
-                    
+
                     if char == '"' and not escape_next:
                         in_string = not in_string
-                    
+
                     if not in_string and i < len(line) - 1 and line[i:i+2] == '//':
                         comment_start = i
                         break
-                
+
                 if comment_start >= 0:
                     line = line[:comment_start]
-                
+
                 cleaned_lines.append(line)
-            
+
             text_no_comments = '\n'.join(cleaned_lines)
-            
+
             # Parse JSON
             data = json.loads(text_no_comments)
         except json.JSONDecodeError as e:
diff --git a/flocks/contracts/webui/models.py b/flocks/contracts/webui/models.py
index 912bf6a24..a39205321 100644
--- a/flocks/contracts/webui/models.py
+++ b/flocks/contracts/webui/models.py
@@ -29,6 +29,7 @@ class WebUIWorkspaceManifest(BaseModel):
     model_config = ConfigDict(populate_by_name=True)
 
     id: str = Field(..., description="Stable workspace identifier")
+    version: str = Field("0.0.0", description="Workspace package version")
     title: str = Field(..., description="Navigation label")
     titleEn: Optional[str] = Field(None, description="English navigation label", alias="titleEn")
     icon: str = Field("LayoutDashboard", description="Lucide icon name")
@@ -117,6 +118,7 @@ class WebUIWorkspaceListItem(BaseModel):
     model_config = ConfigDict(populate_by_name=True, by_alias=True)
 
     id: str
+    version: str = "0.0.0"
     title: str
     titleEn: Optional[str] = Field(None, alias="titleEn")
     route: str
diff --git a/flocks/contracts/webui/store.py b/flocks/contracts/webui/store.py
index f47a2f8d9..7222ba9a9 100644
--- a/flocks/contracts/webui/store.py
+++ b/flocks/contracts/webui/store.py
@@ -288,6 +288,7 @@ def list_workspaces(self, *, enabled_only: bool = False) -> list[WebUIWorkspaceL
                 workspaces.append(
                     WebUIWorkspaceListItem(
                         id=manifest.id,
+                        version=manifest.version,
                         title=manifest.title,
                         titleEn=manifest.titleEn,
                         route=webui_contract_workspace_route(manifest.id),
@@ -677,6 +678,8 @@ def _iter_page_dirs(self, root: Path) -> list[tuple[Path, str]]:
             page_dir = manifest_path.parent
             if page_dir == root:
                 continue
+            if self._is_hidden_relative(page_dir, root):
+                continue
             page_id = self._manifest_page_id_at(manifest_path)
             if page_id is None:
                 continue
@@ -701,6 +704,8 @@ def _iter_workspace_dirs(self, root: Path) -> list[tuple[Path, WebUIWorkspaceMan
             workspace_dir = manifest_path.parent
             if workspace_dir == root:
                 continue
+            if self._is_hidden_relative(workspace_dir, root):
+                continue
             manifest = self._read_workspace_manifest_at(workspace_dir)
             if manifest is None:
                 continue
@@ -764,6 +769,23 @@ def _page_dir_in_root(self, root: Path, page_id: str) -> Path:
         self._assert_inside_root(page_path, root)
         return page_path
 
+    @staticmethod
+    def _is_hidden_relative(path: Path, root: Path) -> bool:
+        """True when *path* lives under a dot-prefixed dir relative to *root*.
+
+        The Hub installer stages WebUI packages into sibling scratch dirs
+        named ``.<plugin>.<rand>`` / ``.<plugin>.bak`` inside this very root
+        before the atomic swap. On Windows that swap can fail (WinError 5)
+        while the page watcher holds the tree open, leaving those dot-dirs
+        behind. Scans must skip them so half-written or stale copies never
+        surface as real pages (and duplicate the live ones).
+        """
+        try:
+            rel = path.relative_to(root)
+        except ValueError:
+            return False
+        return any(part.startswith(".") for part in rel.parts)
+
     @staticmethod
     def _assert_inside_root(path: Path, root: Path) -> None:
         try:
diff --git a/flocks/contracts/webui/watcher.py b/flocks/contracts/webui/watcher.py
index 9cbb413f0..14eea47fd 100644
--- a/flocks/contracts/webui/watcher.py
+++ b/flocks/contracts/webui/watcher.py
@@ -143,6 +143,14 @@ def _classify_event(
             return None
         if not rel.parts:
             return None
+        # Ignore events under dot-prefixed dirs. The Hub installer stages
+        # WebUI packages into sibling scratch dirs (``.<plugin>.<rand>`` /
+        # ``.<plugin>.bak``) inside this watched root before its atomic
+        # swap. Reacting to those writes would race the installer's own
+        # build — on Windows the watcher's file handles block the swap
+        # with a WinError 5 access-denied — and surface half-built pages.
+        if any(part.startswith(".") for part in rel.parts):
+            return None
 
         if rel.name == WORKSPACE_MANIFEST_FILE:
             workspace_id = self._store.workspace_id_for_path(src)
diff --git a/flocks/hooks/pipeline.py b/flocks/hooks/pipeline.py
index 817406c57..47ecd0d7c 100644
--- a/flocks/hooks/pipeline.py
+++ b/flocks/hooks/pipeline.py
@@ -2,12 +2,16 @@
 Hook Pipeline
 
 Provides a lightweight hook registry and execution pipeline that mirrors
-oh-my-opencode's lifecycle stages:
-- chat.message
+agent lifecycle stages:
+- user.prompt.submit
+- session.start
 - llm.call.before
 - llm.call.after
 - tool.execute.before
 - tool.execute.after
+- turn.finish
+- subagent.start
+- subagent.stop
 - event
 """
 
@@ -26,11 +30,15 @@
 
 
 class HookStage:
-    CHAT_MESSAGE = "chat.message"
+    USER_PROMPT_SUBMIT = "user.prompt.submit"
+    SESSION_START = "session.start"
     LLM_BEFORE = "llm.call.before"
     LLM_AFTER = "llm.call.after"
     TOOL_BEFORE = "tool.execute.before"
     TOOL_AFTER = "tool.execute.after"
+    TURN_FINISH = "turn.finish"
+    SUBAGENT_START = "subagent.start"
+    SUBAGENT_STOP = "subagent.stop"
     EVENT = "event"
     CHANNEL_INBOUND = "channel.inbound"
     CHANNEL_OUTBOUND_BEFORE = "channel.outbound.before"
@@ -38,11 +46,15 @@ class HookStage:
 
 
 _DEFAULT_STAGE_TIMEOUTS: Dict[str, float] = {
-    HookStage.CHAT_MESSAGE: 5.0,
+    HookStage.USER_PROMPT_SUBMIT: 5.0,
+    HookStage.SESSION_START: 5.0,
     HookStage.LLM_BEFORE: 5.0,
     HookStage.LLM_AFTER: 5.0,
     HookStage.TOOL_BEFORE: 5.0,
     HookStage.TOOL_AFTER: 5.0,
+    HookStage.TURN_FINISH: 5.0,
+    HookStage.SUBAGENT_START: 5.0,
+    HookStage.SUBAGENT_STOP: 5.0,
     HookStage.CHANNEL_INBOUND: 5.0,
     HookStage.CHANNEL_OUTBOUND_BEFORE: 5.0,
     HookStage.CHANNEL_OUTBOUND_AFTER: 5.0,
@@ -58,7 +70,10 @@ class HookContext:
 
 
 class HookBase:
-    async def chat_message(self, ctx: HookContext) -> None:  # pragma: no cover - default no-op
+    async def user_prompt_submit(self, ctx: HookContext) -> None:  # pragma: no cover - default no-op
+        return None
+
+    async def session_start(self, ctx: HookContext) -> None:  # pragma: no cover - default no-op
         return None
 
     async def llm_before(self, ctx: HookContext) -> None:  # pragma: no cover - default no-op
@@ -73,6 +88,15 @@ async def tool_before(self, ctx: HookContext) -> None:  # pragma: no cover - def
     async def tool_after(self, ctx: HookContext) -> None:  # pragma: no cover - default no-op
         return None
 
+    async def turn_finish(self, ctx: HookContext) -> None:  # pragma: no cover - default no-op
+        return None
+
+    async def subagent_start(self, ctx: HookContext) -> None:  # pragma: no cover - default no-op
+        return None
+
+    async def subagent_stop(self, ctx: HookContext) -> None:  # pragma: no cover - default no-op
+        return None
+
     async def event(self, ctx: HookContext) -> None:  # pragma: no cover - default no-op
         return None
 
@@ -218,12 +242,20 @@ async def ensure_initialized(cls, project_dir: Optional[Path] = None) -> None:
             cls._loaded_project_dir = str(load_project_dir.resolve(strict=False))
 
     @classmethod
-    async def run_chat_message(
+    async def run_user_prompt_submit(
         cls,
         input_data: Dict[str, Any],
         output_data: Optional[Dict[str, Any]] = None,
     ) -> HookContext:
-        return await cls._run_stage(HookStage.CHAT_MESSAGE, input_data, output_data)
+        return await cls._run_stage(HookStage.USER_PROMPT_SUBMIT, input_data, output_data)
+
+    @classmethod
+    async def run_session_start(
+        cls,
+        input_data: Dict[str, Any],
+        output_data: Optional[Dict[str, Any]] = None,
+    ) -> HookContext:
+        return await cls._run_stage(HookStage.SESSION_START, input_data, output_data)
 
     @classmethod
     async def run_llm_before(
@@ -257,6 +289,30 @@ async def run_tool_after(
     ) -> HookContext:
         return await cls._run_stage(HookStage.TOOL_AFTER, input_data, output_data)
 
+    @classmethod
+    async def run_turn_finish(
+        cls,
+        input_data: Dict[str, Any],
+        output_data: Optional[Dict[str, Any]] = None,
+    ) -> HookContext:
+        return await cls._run_stage(HookStage.TURN_FINISH, input_data, output_data)
+
+    @classmethod
+    async def run_subagent_start(
+        cls,
+        input_data: Dict[str, Any],
+        output_data: Optional[Dict[str, Any]] = None,
+    ) -> HookContext:
+        return await cls._run_stage(HookStage.SUBAGENT_START, input_data, output_data)
+
+    @classmethod
+    async def run_subagent_stop(
+        cls,
+        input_data: Dict[str, Any],
+        output_data: Optional[Dict[str, Any]] = None,
+    ) -> HookContext:
+        return await cls._run_stage(HookStage.SUBAGENT_STOP, input_data, output_data)
+
     @classmethod
     async def run_event(
         cls,
@@ -411,22 +467,29 @@ async def _invoke_handler(handler: Callable[[HookContext], Awaitable[None]], ctx
 
     @staticmethod
     def _resolve_handler(hook: HookBase, stage: str) -> Optional[Callable[[HookContext], Awaitable[None]]]:
-        if stage == HookStage.CHAT_MESSAGE:
-            return getattr(hook, "chat_message", None)
-        if stage == HookStage.LLM_BEFORE:
-            return getattr(hook, "llm_before", None)
-        if stage == HookStage.LLM_AFTER:
-            return getattr(hook, "llm_after", None)
-        if stage == HookStage.TOOL_BEFORE:
-            return getattr(hook, "tool_before", None)
-        if stage == HookStage.TOOL_AFTER:
-            return getattr(hook, "tool_after", None)
-        if stage == HookStage.EVENT:
-            return getattr(hook, "event", None)
-        if stage == HookStage.CHANNEL_INBOUND:
-            return getattr(hook, "channel_inbound", None)
-        if stage == HookStage.CHANNEL_OUTBOUND_BEFORE:
-            return getattr(hook, "channel_outbound_before", None)
-        if stage == HookStage.CHANNEL_OUTBOUND_AFTER:
-            return getattr(hook, "channel_outbound_after", None)
-        return None
+        method_name = {
+            HookStage.USER_PROMPT_SUBMIT: "user_prompt_submit",
+            HookStage.SESSION_START: "session_start",
+            HookStage.LLM_BEFORE: "llm_before",
+            HookStage.LLM_AFTER: "llm_after",
+            HookStage.TOOL_BEFORE: "tool_before",
+            HookStage.TOOL_AFTER: "tool_after",
+            HookStage.TURN_FINISH: "turn_finish",
+            HookStage.SUBAGENT_START: "subagent_start",
+            HookStage.SUBAGENT_STOP: "subagent_stop",
+            HookStage.EVENT: "event",
+            HookStage.CHANNEL_INBOUND: "channel_inbound",
+            HookStage.CHANNEL_OUTBOUND_BEFORE: "channel_outbound_before",
+            HookStage.CHANNEL_OUTBOUND_AFTER: "channel_outbound_after",
+        }.get(stage)
+        if method_name is None:
+            return None
+
+        handler = getattr(hook, method_name, None)
+        if not callable(handler):
+            return None
+        base_handler = getattr(HookBase, method_name, None)
+        concrete_handler = getattr(type(hook), method_name, None)
+        if base_handler is not None and concrete_handler is base_handler:
+            return None
+        return handler
diff --git a/flocks/hub/catalog.py b/flocks/hub/catalog.py
index 87a45ce51..a999a12b3 100644
--- a/flocks/hub/catalog.py
+++ b/flocks/hub/catalog.py
@@ -86,6 +86,8 @@ def _plugin_manifest_signature(plugin_type: PluginType, root: Path) -> tuple[tup
         candidates = [root / "agent.yaml"]
     elif plugin_type == "workflow":
         candidates = [root / "workflow.json", root / "workflow.md"]
+    elif plugin_type == "webui":
+        candidates = [root / "workspace.json", root / "manifest.json"]
     else:
         try:
             candidates = [
@@ -741,6 +743,7 @@ def _version_tuple(value: str) -> tuple[int, ...]:
 
 
 def _catalog_install_state(
+    plugin_type: PluginType,
     install_path: Optional[Path],
     record: Optional[local.InstalledPluginRecord],
     available_version: str,
@@ -748,10 +751,15 @@ def _catalog_install_state(
     if install_path is None:
         return "available", None
     if record is None:
-        # An inferred install has no trustworthy package version because Hub
-        # manifests are not copied into the install directory. Reconcile it
-        # through the update path instead of assuming it is already current.
-        return "updateAvailable", None
+        installed_version = local.installed_payload_version(plugin_type, install_path)
+        if installed_version is None:
+            # Legacy inferred installs have no trustworthy package version.
+            # Reconcile them through the update path instead of assuming they
+            # are already current.
+            return "updateAvailable", None
+        if _version_tuple(installed_version) < _version_tuple(available_version):
+            return "updateAvailable", installed_version
+        return "installed", installed_version
     if _version_tuple(record.version) < _version_tuple(available_version):
         return "updateAvailable", record.version
     return "installed", record.version
@@ -776,6 +784,7 @@ def _entry_from_manifest(manifest: HubPluginManifest) -> HubCatalogEntry:
     )
 
     state, installed_version = _catalog_install_state(
+        manifest.type,
         install_path,
         record,
         manifest.version,
@@ -838,6 +847,7 @@ def _entry_from_index(
     )
 
     state, installed_version = _catalog_install_state(
+        item.type,
         install_path,
         record,
         item.version,
@@ -920,6 +930,7 @@ def _entry_from_bundled_tool(
     )
 
     state, installed_version = _catalog_install_state(
+        manifest.type,
         install_path,
         record,
         manifest.version,
diff --git a/flocks/hub/installer.py b/flocks/hub/installer.py
index 07b9a30b0..eb894126c 100644
--- a/flocks/hub/installer.py
+++ b/flocks/hub/installer.py
@@ -4,7 +4,9 @@
 
 import asyncio
 import shutil
+import sys
 import tempfile
+import time
 from pathlib import Path
 from typing import Awaitable, Callable
 
@@ -97,17 +99,64 @@ def _remove_path(path: Path) -> None:
         path.unlink()
 
 
+def _purge_stale_scratch(parent: Path, name: str) -> None:
+    """Remove leftover ``.<name>.<rand>`` / ``.<name>.bak`` staging dirs.
+
+    A failed atomic swap (see :func:`_replace_prepared_path`) can leave
+    scratch and backup dirs behind next to *parent*/*name*. They are never
+    valid installs, but on Windows a lingering ``.<name>.bak`` blocks the
+    next swap, so we clear both before staging a fresh copy.
+    """
+    if not parent.is_dir():
+        return
+    for entry in parent.iterdir():
+        stale = entry.name.startswith(f".{name}.") or entry.name == f".{name}.bak"
+        if not stale:
+            continue
+        try:
+            if entry.is_dir() and not entry.is_symlink():
+                shutil.rmtree(entry, ignore_errors=True)
+            else:
+                entry.unlink()
+        except OSError:
+            pass
+
+
+def _replace_with_retry(src: Path, dst: Path) -> None:
+    """``src.replace(dst)`` with a Windows access-denied backoff.
+
+    On Windows an antivirus scan or a directory watcher (e.g. the WebUI
+    page watcher over ``~/.flocks/plugins/contracts/webui``) can hold a
+    transient handle on the freshly written tree, making the atomic swap
+    fail with ``PermissionError`` (WinError 5 / 32). Elsewhere the rename
+    is atomic and never needs retrying.
+    """
+    if sys.platform != "win32":
+        src.replace(dst)
+        return
+    delay = 0.1
+    for attempt in range(6):
+        try:
+            src.replace(dst)
+            return
+        except PermissionError:
+            if attempt == 5:
+                raise
+            time.sleep(delay)
+            delay = min(delay * 2, 1.0)
+
+
 def _replace_prepared_path(prepared: Path, dst: Path) -> Path | None:
     backup: Path | None = None
     if dst.exists() or dst.is_symlink():
         backup = dst.parent / f".{dst.name}.bak"
         _remove_path(backup)
-        dst.replace(backup)
+        _replace_with_retry(dst, backup)
     try:
-        prepared.replace(dst)
+        _replace_with_retry(prepared, dst)
     except Exception:
         if backup is not None and (backup.exists() or backup.is_symlink()):
-            backup.replace(dst)
+            _replace_with_retry(backup, dst)
         raise
     return backup
 
@@ -124,12 +173,13 @@ def _commit_replacement(backup: Path | None) -> None:
 def _rollback_replacement(dst: Path, backup: Path | None) -> None:
     _remove_path(dst)
     if backup is not None and (backup.exists() or backup.is_symlink()):
-        backup.replace(dst)
+        _replace_with_retry(backup, dst)
 
 
 def _copy_package(src: Path, dst: Path, *, retain_backup: bool = False) -> Path | None:
     parent = dst.parent
     parent.mkdir(parents=True, exist_ok=True)
+    _purge_stale_scratch(parent, dst.name)
     tmp = Path(tempfile.mkdtemp(prefix=f".{dst.name}.", dir=str(parent)))
     try:
         _copy_package_contents(src, tmp)
@@ -218,6 +268,7 @@ def _copy_webui_package_with_build(
 ) -> Path | None:
     parent = dst.parent
     parent.mkdir(parents=True, exist_ok=True)
+    _purge_stale_scratch(parent, dst.name)
     tmp = Path(tempfile.mkdtemp(prefix=f".{dst.name}.", dir=str(parent)))
     try:
         _copy_package_contents(src, tmp)
diff --git a/flocks/hub/local.py b/flocks/hub/local.py
index e7f5d2580..963d87e9a 100644
--- a/flocks/hub/local.py
+++ b/flocks/hub/local.py
@@ -122,6 +122,26 @@ def has_install_payload(plugin_type: PluginType, path: Path) -> bool:
     return path.exists()
 
 
+def installed_payload_version(plugin_type: PluginType, path: Path) -> Optional[str]:
+    """Read a version embedded in an installed payload when one is available."""
+    if plugin_type != "webui" or not path.is_dir():
+        return None
+
+    import json
+
+    workspace_path = path / "workspace.json"
+    if not workspace_path.is_file():
+        return None
+    try:
+        raw = json.loads(workspace_path.read_text(encoding="utf-8"))
+    except Exception:
+        return None
+    version = raw.get("version") if isinstance(raw, dict) else None
+    if not isinstance(version, str) or not version.strip():
+        return None
+    return version.strip()
+
+
 def get_record(plugin_type: PluginType, plugin_id: str) -> Optional[InstalledPluginRecord]:
     return load_installed_records().get(f"{plugin_type}:{plugin_id}")
 
diff --git a/flocks/input/events.py b/flocks/input/events.py
index d0fd4e54c..4714c1097 100644
--- a/flocks/input/events.py
+++ b/flocks/input/events.py
@@ -8,6 +8,7 @@
 
 from flocks.command.command import CommandInfo
 from flocks.input.types import InputSourceType, surface_for_source
+from flocks.session.execution_mode import SessionExecutionMode
 
 
 class UserInputEvent(BaseModel):
@@ -31,6 +32,10 @@ class UserInputEvent(BaseModel):
     mock_reply: Optional[str] = Field(None, alias="mockReply")
     system: Optional[str] = None
     tools: Optional[Dict[str, bool]] = None
+    execution_mode: SessionExecutionMode = Field(
+        SessionExecutionMode.BUILD,
+        alias="executionMode",
+    )
 
     @property
     def surface(self) -> str:
diff --git a/flocks/permission/next.py b/flocks/permission/next.py
index 18bfc51af..1bd23a826 100644
--- a/flocks/permission/next.py
+++ b/flocks/permission/next.py
@@ -7,7 +7,7 @@
 
 import asyncio
 from datetime import datetime
-from typing import Optional, Dict, Any, List, Callable, Awaitable
+from typing import Optional, Dict, Any, List, Callable, Awaitable, Iterable
 
 from pydantic import BaseModel, Field
 
@@ -59,6 +59,7 @@ class PermissionNext:
     _session_permissions: Dict[str, Dict[str, str]] = {}
     _permanent_rules: Dict[str, str] = {}
     _state_loaded: bool = False
+    _persistence_tasks: set[asyncio.Task[Any]] = set()
 
     _PENDING_PREFIX = "permission_pending:"
     _REPLY_PREFIX = "permission_reply:"
@@ -114,10 +115,19 @@ async def _ensure_persisted_state_loaded(cls) -> None:
     def _schedule_persist(cls, coro: Awaitable[Any]) -> None:
         try:
             loop = asyncio.get_running_loop()
-            loop.create_task(coro)
+            task = loop.create_task(coro)
+            cls._persistence_tasks.add(task)
+            task.add_done_callback(cls._persistence_tasks.discard)
         except RuntimeError:
             pass
 
+    @classmethod
+    async def _flush_persistence(cls) -> None:
+        """Wait for permission writes already scheduled in this process."""
+
+        while cls._persistence_tasks:
+            await asyncio.gather(*tuple(cls._persistence_tasks))
+
     @classmethod
     async def _persist_pending_request(cls, request_info: PermissionRequestInfo) -> None:
         await Storage.set(
@@ -184,6 +194,53 @@ async def _persist_session_rules(cls, session_id: str) -> None:
             "permission_session",
         )
 
+    @classmethod
+    async def deletion_storage_keys(cls, session_ids: Iterable[str]) -> List[str]:
+        """Return every persisted permission key owned by the given sessions."""
+
+        ids = {session_id for session_id in session_ids if session_id}
+        if not ids:
+            return []
+
+        await cls._flush_persistence()
+        keys = {f"{cls._SESSION_PREFIX}{session_id}" for session_id in ids}
+        request_ids = {
+            request_id
+            for request_id, pending in cls._pending.items()
+            if getattr(pending.get("info"), "session_id", None) in ids
+        }
+
+        for key, value in await Storage.list_entries(prefix=cls._PENDING_PREFIX):
+            if isinstance(value, dict) and value.get("sessionID") in ids:
+                keys.add(key)
+                request_ids.add(key.removeprefix(cls._PENDING_PREFIX))
+
+        for key, value in await Storage.list_entries(prefix=cls._REPLY_PREFIX):
+            if isinstance(value, dict) and value.get("sessionID") in ids:
+                keys.add(key)
+                request_ids.add(key.removeprefix(cls._REPLY_PREFIX))
+
+        for request_id in request_ids:
+            keys.add(f"{cls._PENDING_PREFIX}{request_id}")
+            keys.add(f"{cls._REPLY_PREFIX}{request_id}")
+        return sorted(keys)
+
+    @classmethod
+    def clear_session_runtime(cls, session_ids: Iterable[str]) -> None:
+        """Drop session permission caches and cancel their pending requests."""
+
+        ids = {session_id for session_id in session_ids if session_id}
+        for session_id in ids:
+            cls._session_permissions.pop(session_id, None)
+
+        for request_id, pending in list(cls._pending.items()):
+            if getattr(pending.get("info"), "session_id", None) not in ids:
+                continue
+            future = pending.get("future")
+            if isinstance(future, asyncio.Future) and not future.done():
+                future.cancel()
+            cls._pending.pop(request_id, None)
+
     @classmethod
     async def list_pending_infos(cls) -> List[PermissionRequestInfo]:
         pending_infos = [
diff --git a/flocks/project/project.py b/flocks/project/project.py
index edfa1aa19..8654a8576 100644
--- a/flocks/project/project.py
+++ b/flocks/project/project.py
@@ -8,6 +8,7 @@
 from __future__ import annotations
 
 import asyncio
+import contextvars
 import hashlib
 import json
 import os
@@ -16,13 +17,15 @@
 import tempfile
 import time
 import uuid
-from contextlib import contextmanager
+import weakref
+from contextlib import asynccontextmanager, contextmanager
 from datetime import datetime
 from pathlib import Path
-from typing import Any, ClassVar, Dict, Iterator, List, Literal, Optional, Tuple
+from typing import AsyncIterator, Any, ClassVar, Dict, Iterator, List, Literal, Optional, Tuple
 
 from pydantic import BaseModel, ConfigDict, Field
 
+from flocks.config.config import Config
 from flocks.utils.log import Log
 
 log = Log.create(service="project")
@@ -58,6 +61,24 @@ def _platform_file_unlock(fd: int) -> None:
         fcntl.flock(fd, fcntl.LOCK_UN)
 
 
+async def _acquire_file_lock(fd: int) -> None:
+    """Acquire a blocking OS lock without leaking it when the waiter is cancelled."""
+
+    worker = asyncio.create_task(asyncio.to_thread(_platform_file_lock, fd))
+    try:
+        await asyncio.shield(worker)
+    except asyncio.CancelledError:
+        while not worker.done():
+            try:
+                await asyncio.shield(worker)
+            except asyncio.CancelledError:
+                continue
+        if not worker.cancelled():
+            worker.result()
+            _platform_file_unlock(fd)
+        raise
+
+
 @contextmanager
 def _registry_cross_process_lock(registry_path: Path) -> Iterator[None]:
     """Serialize registry read-modify-write operations across processes."""
@@ -148,6 +169,7 @@ class ProjectRegistryEntry(BaseModel):
     updated_at: int = Field(alias="updatedAt")
     owner_user_id: Optional[str] = Field(None, alias="ownerUserID")
     shared_local: bool = Field(False, alias="sharedLocal")
+    removed_at: Optional[int] = Field(None, alias="removedAt")
 
 
 class ProjectRegistry(BaseModel):
@@ -163,6 +185,12 @@ class Project:
     """User-scoped registry of explicitly created projects."""
 
     _lock = asyncio.Lock()
+    _lifecycle_locks: ClassVar[weakref.WeakValueDictionary[str, asyncio.Lock]] = (
+        weakref.WeakValueDictionary()
+    )
+    _held_lifecycle_guards: ClassVar[
+        contextvars.ContextVar[Dict[str, asyncio.Task[Any]]]
+    ] = contextvars.ContextVar("project_lifecycle_guards", default={})
     _session_stats_cache: ClassVar[
         Dict[Tuple[str, str], Tuple[float, Dict[str, Tuple[int, int, Optional[int]]]]]
     ] = {}
@@ -172,6 +200,89 @@ class Project:
     def _now_ms() -> int:
         return int(datetime.now().timestamp() * 1000)
 
+    @classmethod
+    def lifecycle_lock(cls, project_id: str) -> asyncio.Lock:
+        """Serialize session creation, project removal, and restoration per project."""
+
+        lock = cls._lifecycle_locks.get(project_id)
+        if lock is None:
+            lock = asyncio.Lock()
+            cls._lifecycle_locks[project_id] = lock
+        return lock
+
+    @classmethod
+    @asynccontextmanager
+    async def lifecycle_guard(cls, project_id: str) -> AsyncIterator[None]:
+        """Serialize a project lifecycle transition across tasks and processes.
+
+        The context is re-entrant for the current asyncio task so project deletion
+        can archive several roots through the same public session APIs. ContextVar
+        state records the owning task explicitly because child tasks inherit
+        context and must acquire their own lock.
+        """
+
+        task = asyncio.current_task()
+        if task is None:
+            raise RuntimeError("Project lifecycle guard requires an asyncio task")
+
+        held = cls._held_lifecycle_guards.get()
+        if held.get(project_id) is task:
+            yield
+            return
+
+        async with cls.lifecycle_lock(project_id):
+            lock_dir = Config.get_data_path() / "locks" / "project-lifecycle"
+            lock_dir.mkdir(parents=True, exist_ok=True)
+            lock_name = hashlib.sha256(project_id.encode("utf-8")).hexdigest()
+            fd = os.open(str(lock_dir / f"{lock_name}.lock"), os.O_RDWR | os.O_CREAT, 0o600)
+            acquired = False
+            try:
+                await _acquire_file_lock(fd)
+                acquired = True
+                token = cls._held_lifecycle_guards.set({**held, project_id: task})
+                try:
+                    yield
+                finally:
+                    cls._held_lifecycle_guards.reset(token)
+            finally:
+                try:
+                    if acquired:
+                        _platform_file_unlock(fd)
+                finally:
+                    os.close(fd)
+
+    @classmethod
+    def registry_state(
+        cls,
+        project_id: str,
+        *,
+        owner_id: str,
+    ) -> Literal["virtual", "missing", "active", "removed"]:
+        """Return the registry state used by session create/restore guards."""
+
+        if project_id in {DEFAULT_PROJECT_ID, TASK_SESSION_GROUP_ID}:
+            return "virtual"
+        entry = next(
+            (
+                item
+                for item in cls._read_registry(owner_id).projects
+                if item.id == project_id
+            ),
+            None,
+        )
+        if entry is None:
+            return "missing"
+        return "removed" if entry.removed_at is not None else "active"
+
+    @classmethod
+    def is_removed(cls, project_id: str) -> bool:
+        """Return whether any local registry marks this project as removed."""
+
+        return any(
+            entry.id == project_id and entry.removed_at is not None
+            for entry in cls._all_registry_entries(include_removed=True)
+        )
+
     @staticmethod
     def _flocks_root() -> Path:
         return Path(os.getenv("FLOCKS_ROOT", str(Path.home() / ".flocks"))).expanduser()
@@ -416,6 +527,8 @@ async def create(
                 )
 
                 for entry in registry.projects:
+                    if entry.removed_at is not None:
+                        continue
                     if cls._normalized_worktree(entry.worktree) == normalized_worktree:
                         raise ProjectPathConflictError(cls._entry_to_info(entry))
 
@@ -424,10 +537,35 @@ async def create(
                     raise ValueError("Project name cannot be empty")
                 if normalized_name.casefold() in _RESERVED_PROJECT_NAMES:
                     raise ProjectNameConflictError("Project name is reserved")
-                if any(entry.name.strip().casefold() == normalized_name.casefold() for entry in registry.projects):
+                if any(
+                    entry.removed_at is None
+                    and entry.name.strip().casefold() == normalized_name.casefold()
+                    for entry in registry.projects
+                ):
                     raise ProjectNameConflictError(f"Project name '{normalized_name}' already exists")
 
                 now = cls._now_ms()
+                removed_entry = next(
+                    (
+                        entry
+                        for entry in registry.projects
+                        if entry.removed_at is not None
+                        and cls._normalized_worktree(entry.worktree) == normalized_worktree
+                    ),
+                    None,
+                )
+                if removed_entry is not None:
+                    removed_entry.name = normalized_name
+                    removed_entry.worktree = normalized_worktree
+                    removed_entry.owner_user_id = owner_id
+                    removed_entry.shared_local = False
+                    removed_entry.removed_at = None
+                    removed_entry.updated_at = now
+                    cls._write_registry(owner_id, registry)
+                    cls.invalidate_session_stats(owner_id)
+                    log.info("project.restored", {"id": removed_entry.id})
+                    return cls._entry_to_info(removed_entry)
+
                 project_id = f"prj_{uuid.uuid4()}"
                 entry = ProjectRegistryEntry(
                     id=project_id,
@@ -450,12 +588,20 @@ async def list(
         owner_id: str,
     ) -> List[ProjectInfo]:
         registry = await cls.ensure_registry(owner_id)
-        projects = [cls._entry_to_info(entry) for entry in registry.projects]
+        projects = [
+            cls._entry_to_info(entry)
+            for entry in registry.projects
+            if entry.removed_at is None
+        ]
         projects.sort(key=lambda item: item.time.updated, reverse=True)
         return projects
 
     @classmethod
-    def _all_registry_entries(cls) -> List[ProjectRegistryEntry]:
+    def _all_registry_entries(
+        cls,
+        *,
+        include_removed: bool = False,
+    ) -> List[ProjectRegistryEntry]:
         """Return valid entries across local user registries."""
 
         registry_dir = cls._flocks_root() / "projects"
@@ -465,7 +611,11 @@ def _all_registry_entries(cls) -> List[ProjectRegistryEntry]:
         for path in registry_dir.glob("*.json"):
             registry = cls._read_registry_file(path)
             if registry is not None:
-                entries.extend(registry.projects)
+                entries.extend(
+                    entry
+                    for entry in registry.projects
+                    if include_removed or entry.removed_at is None
+                )
         return entries
 
     @classmethod
@@ -505,7 +655,12 @@ def is_local_shared(cls, owner_id: Optional[str], project_id: Optional[str]) ->
         if not owner_id or not project_id or project_id == DEFAULT_PROJECT_ID:
             return False
         registry = cls._read_registry(owner_id)
-        return any(entry.id == project_id and entry.shared_local for entry in registry.projects)
+        return any(
+            entry.id == project_id
+            and entry.removed_at is None
+            and entry.shared_local
+            for entry in registry.projects
+        )
 
     @classmethod
     async def set_local_shared(
@@ -522,7 +677,14 @@ async def set_local_shared(
         async with cls._lock:
             with _registry_cross_process_lock(cls.registry_path(owner_id)):
                 registry = cls._read_registry(owner_id)
-                entry = next((item for item in registry.projects if item.id == project_id), None)
+                entry = next(
+                    (
+                        item
+                        for item in registry.projects
+                        if item.id == project_id and item.removed_at is None
+                    ),
+                    None,
+                )
                 if entry is None:
                     raise ValueError(f"Project {project_id} not found")
                 entry.owner_user_id = owner_id
@@ -544,7 +706,14 @@ async def get(
         owner_id: str,
     ) -> Optional[ProjectInfo]:
         registry = await cls.ensure_registry(owner_id)
-        entry = next((item for item in registry.projects if item.id == project_id), None)
+        entry = next(
+            (
+                item
+                for item in registry.projects
+                if item.id == project_id and item.removed_at is None
+            ),
+            None,
+        )
         return cls._entry_to_info(entry) if entry else None
 
     @classmethod
@@ -562,11 +731,20 @@ async def update(cls, project_id: str, *, owner_id: str, name: str) -> ProjectIn
         async with cls._lock:
             with _registry_cross_process_lock(cls.registry_path(owner_id)):
                 registry = cls._read_registry(owner_id)
-                entry = next((item for item in registry.projects if item.id == project_id), None)
+                entry = next(
+                    (
+                        item
+                        for item in registry.projects
+                        if item.id == project_id and item.removed_at is None
+                    ),
+                    None,
+                )
                 if entry is None:
                     raise ValueError(f"Project {project_id} not found")
                 if any(
-                    item.id != project_id and item.name.strip().casefold() == normalized_name.casefold()
+                    item.id != project_id
+                    and item.removed_at is None
+                    and item.name.strip().casefold() == normalized_name.casefold()
                     for item in registry.projects
                 ):
                     raise ProjectNameConflictError(f"Project name '{normalized_name}' already exists")
@@ -580,26 +758,114 @@ async def update(cls, project_id: str, *, owner_id: str, name: str) -> ProjectIn
 
     @classmethod
     async def delete(cls, project_id: str, *, owner_id: str) -> bool:
-        """Remove a project registration without changing sessions or files."""
+        """Hide a project registration while preserving metadata for restoration."""
+
+        async with cls.lifecycle_guard(project_id):
+            return await cls._delete_locked(project_id, owner_id=owner_id)
+
+    @classmethod
+    async def _delete_locked(cls, project_id: str, *, owner_id: str) -> bool:
+        """Update the project registry while its lifecycle guard is held."""
 
         if project_id == DEFAULT_PROJECT_ID:
             raise ProjectDeletionError("The default project cannot be deleted")
         async with cls._lock:
             with _registry_cross_process_lock(cls.registry_path(owner_id)):
                 registry = cls._read_registry(owner_id)
-                remaining = [entry for entry in registry.projects if entry.id != project_id]
-                if len(remaining) == len(registry.projects):
+                entry = next(
+                    (
+                        item
+                        for item in registry.projects
+                        if item.id == project_id and item.removed_at is None
+                    ),
+                    None,
+                )
+                if entry is None:
                     raise ValueError(f"Project {project_id} not found")
-                registry.projects = remaining
+                now = cls._now_ms()
+                entry.removed_at = now
+                entry.updated_at = now
+                entry.shared_local = False
                 cls._write_registry(owner_id, registry)
             cls.invalidate_session_stats(owner_id)
-            log.info("project.deleted", {"id": project_id})
+            log.info("project.removed", {"id": project_id})
             return True
 
+    @classmethod
+    async def restore(
+        cls,
+        project_id: str,
+        *,
+        owner_id: str,
+    ) -> Optional[ProjectInfo]:
+        """Restore a soft-removed project when its original directory is usable."""
+
+        async with cls.lifecycle_guard(project_id):
+            return await cls._restore_locked(project_id, owner_id=owner_id)
+
+    @classmethod
+    async def _restore_locked(
+        cls,
+        project_id: str,
+        *,
+        owner_id: str,
+    ) -> Optional[ProjectInfo]:
+        """Restore registry metadata while its lifecycle guard is held."""
+
+        if project_id == DEFAULT_PROJECT_ID:
+            return None
+        async with cls._lock:
+            with _registry_cross_process_lock(cls.registry_path(owner_id)):
+                registry = cls._read_registry(owner_id)
+                entry = next((item for item in registry.projects if item.id == project_id), None)
+                if entry is None:
+                    raise ProjectDeletionError(
+                        f"Project {project_id} restoration metadata is unavailable"
+                    )
+                if entry.removed_at is None:
+                    return cls._entry_to_info(entry)
+                if cls._path_status(entry.worktree) != "available":
+                    raise ProjectDeletionError(
+                        f"Project '{entry.name}' directory is unavailable; the archived task was not restored"
+                    )
+
+                normalized_worktree = cls.validate_worktree(
+                    entry.worktree,
+                    create_if_missing=False,
+                )
+                for other in registry.projects:
+                    if other.id == entry.id or other.removed_at is not None:
+                        continue
+                    if cls._normalized_worktree(other.worktree) == normalized_worktree:
+                        raise ProjectPathConflictError(cls._entry_to_info(other))
+                    if other.name.strip().casefold() == entry.name.strip().casefold():
+                        raise ProjectNameConflictError(
+                            f"Project name '{entry.name}' already exists"
+                        )
+
+                entry.worktree = normalized_worktree
+                entry.owner_user_id = owner_id
+                entry.shared_local = False
+                entry.removed_at = None
+                entry.updated_at = cls._now_ms()
+                cls._write_registry(owner_id, registry)
+            cls.invalidate_session_stats(owner_id)
+            log.info("project.restored", {"id": project_id})
+            return cls._entry_to_info(entry)
+
     @classmethod
     def registered_project_ids(cls, owner_id: str) -> set[str]:
         registry = cls._read_registry(owner_id)
-        return {entry.id for entry in registry.projects}
+        return {entry.id for entry in registry.projects if entry.removed_at is None}
+
+    @classmethod
+    def registered_project_names(cls) -> Dict[str, str]:
+        """Return project display names across local user registries."""
+
+        return {
+            entry.id: entry.name
+            for entry in cls._all_registry_entries(include_removed=True)
+        }
 
     @classmethod
     def effective_project_id(cls, owner_id: str, stored_project_id: Optional[str]) -> str:
diff --git a/flocks/provider/catalog.json b/flocks/provider/catalog.json
index d6af63c94..69f35518f 100644
--- a/flocks/provider/catalog.json
+++ b/flocks/provider/catalog.json
@@ -1229,11 +1229,12 @@
       "MOONSHOT_API_KEY"
     ],
     "models": {
-      "kimi-k2.5": {
-        "name": "Kimi K2.5",
-        "family": "kimi-k2.5",
+      "kimi-k3": {
+        "name": "Kimi K3",
+        "family": "kimi-k3",
         "capabilities": {
           "supports_tools": true,
+          "supports_vision": true,
           "supports_reasoning": true,
           "interleaved": {
             "field": "reasoning_content",
@@ -1244,18 +1245,19 @@
           "supports_streaming": true
         },
         "limits": {
-          "context_window": 262144,
-          "max_output_tokens": 262144
+          "context_window": 1048576,
+          "max_output_tokens": 131072
         },
         "pricing": {
-          "input": 4.0,
-          "output": 21.0,
+          "input": 20.0,
+          "output": 100.0,
+          "cache_read": 2.0,
           "currency": "CNY"
         }
       },
-      "kimi-k2.6": {
-        "name": "Kimi K2.6",
-        "family": "kimi-k2.6",
+      "kimi-k2.7-code": {
+        "name": "Kimi K2.7 Code",
+        "family": "kimi-k2.7-code",
         "capabilities": {
           "supports_tools": true,
           "supports_vision": true,
@@ -1269,9 +1271,8 @@
           "supports_streaming": true
         },
         "limits": {
-          "context_window": 256000,
-          "max_input_tokens": 224000,
-          "max_output_tokens": 16000
+          "context_window": 262144,
+          "max_output_tokens": 32768
         },
         "pricing": {
           "input": 6.5,
@@ -1280,9 +1281,35 @@
           "currency": "CNY"
         }
       },
-      "kimi-k2-thinking": {
-        "name": "Kimi K2 Thinking",
-        "family": "kimi-k2",
+      "kimi-k2.7-code-highspeed": {
+        "name": "Kimi K2.7 Code HighSpeed",
+        "family": "kimi-k2.7-code",
+        "capabilities": {
+          "supports_tools": true,
+          "supports_vision": true,
+          "supports_reasoning": true,
+          "interleaved": {
+            "field": "reasoning_content",
+            "echo": "tool_calls",
+            "placeholder": " ",
+            "cross_provider_policy": "placeholder"
+          },
+          "supports_streaming": true
+        },
+        "limits": {
+          "context_window": 262144,
+          "max_output_tokens": 32768
+        },
+        "pricing": {
+          "input": 13.0,
+          "output": 54.0,
+          "cache_read": 2.6,
+          "currency": "CNY"
+        }
+      },
+      "kimi-k2.5": {
+        "name": "Kimi K2.5",
+        "family": "kimi-k2.5",
         "capabilities": {
           "supports_tools": true,
           "supports_reasoning": true,
@@ -1300,24 +1327,34 @@
         },
         "pricing": {
           "input": 4.0,
-          "output": 16.0,
+          "output": 21.0,
           "currency": "CNY"
         }
       },
-      "kimi-k2": {
-        "name": "Kimi K2 0711",
-        "family": "kimi-k2",
+      "kimi-k2.6": {
+        "name": "Kimi K2.6",
+        "family": "kimi-k2.6",
         "capabilities": {
           "supports_tools": true,
+          "supports_vision": true,
+          "supports_reasoning": true,
+          "interleaved": {
+            "field": "reasoning_content",
+            "echo": "tool_calls",
+            "placeholder": " ",
+            "cross_provider_policy": "placeholder"
+          },
           "supports_streaming": true
         },
         "limits": {
-          "context_window": 131072,
-          "max_output_tokens": 131072
+          "context_window": 256000,
+          "max_input_tokens": 224000,
+          "max_output_tokens": 16000
         },
         "pricing": {
-          "input": 4.0,
-          "output": 16.0,
+          "input": 6.5,
+          "output": 27.0,
+          "cache_read": 1.3,
           "currency": "CNY"
         }
       }
diff --git a/flocks/provider/interleaved.py b/flocks/provider/interleaved.py
index 727628a4e..897833b62 100644
--- a/flocks/provider/interleaved.py
+++ b/flocks/provider/interleaved.py
@@ -6,6 +6,11 @@
 
 REASONING_TRANSPORT_GENERIC_CHAT = "generic_chat"
 REASONING_TRANSPORT_ANTHROPIC_MESSAGES = "anthropic_messages"
+KIMI_K27_CODE_MODELS = frozenset({
+    "kimi-k2.7-code",
+    "kimi-k2.7-code-highspeed",
+})
+KIMI_K3_MODELS = frozenset({"kimi-k3"})
 
 
 _STRICT_REASONING_CONTENT = {
@@ -87,6 +92,18 @@ def _matches_any(text: str, *tokens: str) -> bool:
     return any(token in text for token in tokens)
 
 
+def normalize_model_id(model_id: str) -> str:
+    return model_id.strip().lower().replace("_", "-")
+
+
+def is_kimi_k27_code_model(model_id: str) -> bool:
+    return normalize_model_id(model_id) in KIMI_K27_CODE_MODELS
+
+
+def is_kimi_k3_model(model_id: str) -> bool:
+    return normalize_model_id(model_id) in KIMI_K3_MODELS
+
+
 def infer_interleaved_capability(
     *,
     provider_id: str,
@@ -100,7 +117,7 @@ def infer_interleaved_capability(
     works without user-visible toggles.
     """
     pid = _lower(provider_id)
-    mid = _lower(model_id)
+    mid = normalize_model_id(model_id)
     burl = _lower(base_url)
 
     if "minimax" in mid or pid == "minimax":
@@ -122,6 +139,8 @@ def infer_interleaved_capability(
     ):
         return dict(_PROMOTE_REASONING_CONTENT)
 
+    if is_kimi_k27_code_model(mid) or is_kimi_k3_model(mid):
+        return dict(_STRICT_REASONING_CONTENT)
     if _matches_any(mid, *_STRICT_REASONING_CONTENT_TOKENS):
         return dict(_STRICT_REASONING_CONTENT)
     if (
diff --git a/flocks/provider/options.py b/flocks/provider/options.py
index 07cb66def..733093df5 100644
--- a/flocks/provider/options.py
+++ b/flocks/provider/options.py
@@ -14,6 +14,8 @@
 from flocks.provider.interleaved import (
     REASONING_TRANSPORT_ANTHROPIC_MESSAGES,
     REASONING_TRANSPORT_GENERIC_CHAT,
+    is_kimi_k27_code_model,
+    is_kimi_k3_model,
     resolve_interleaved_capability,
     resolve_reasoning_transport,
 )
@@ -27,6 +29,8 @@
 # ---------------------------------------------------------------------------
 DEFAULT_THINKING_BUDGET = 16000
 DEFAULT_OUTPUT_BUFFER = 8192
+DEFAULT_KIMI_K3_REASONING_EFFORT = "max"
+KIMI_K3_REASONING_EFFORTS = frozenset({"low", "high", "max"})
 
 _GENERIC_CHAT_REASONING_EXTRA_BODY_KEYS = {
     "reasoning_content": "enable_thinking",
@@ -71,6 +75,26 @@ def _resolve_reasoning_enabled(provider_id: str, model_id: str) -> Optional[bool
         return None
 
 
+def _resolve_reasoning_effort(provider_id: str, model_id: str) -> Optional[str]:
+    """Read a model-level reasoning effort from flocks.json."""
+    try:
+        from flocks.provider.model_manager import get_model_manager
+
+        setting = get_model_manager().get_setting(provider_id, model_id)
+        if not setting:
+            return None
+
+        value = (setting.default_parameters or {}).get("reasoning_effort")
+        return value.strip().lower() if isinstance(value, str) else None
+    except Exception as exc:
+        log.debug("options.reasoning_effort_setting_lookup_failed", {
+            "provider_id": provider_id,
+            "model_id": model_id,
+            "error": str(exc),
+        })
+        return None
+
+
 def _resolve_default_extra_body(provider_id: str, model_id: str) -> Optional[Dict[str, Any]]:
     """Read model-level OpenAI-compatible extra_body from flocks.json."""
     try:
@@ -201,12 +225,27 @@ def _build_generic_chat_extra_body(
     model_id: str,
     interleaved_capability: Optional[Dict[str, Any]],
     reasoning_enabled: Optional[bool],
+    reasoning_effort: Optional[str],
 ) -> Optional[Dict[str, Any]]:
     """Build OpenAI-compatible reasoning params for the active replay field."""
     provider_lower = provider_id.lower()
     model_lower = model_id.lower()
     enabled = reasoning_enabled is not False
 
+    if is_kimi_k3_model(model_id):
+        effort = reasoning_effort or DEFAULT_KIMI_K3_REASONING_EFFORT
+        if effort not in KIMI_K3_REASONING_EFFORTS:
+            log.warning(
+                "options.kimi_k3.invalid_reasoning_effort",
+                {
+                    "model_id": model_id,
+                    "reasoning_effort": effort,
+                    "fallback": DEFAULT_KIMI_K3_REASONING_EFFORT,
+                },
+            )
+            effort = DEFAULT_KIMI_K3_REASONING_EFFORT
+        return {"reasoning_effort": effort}
+
     if "deepseek" in model_lower or provider_lower == "deepseek":
         return {
             "thinking": (
@@ -234,6 +273,11 @@ def _build_generic_chat_extra_body(
             )
         }
 
+    if is_kimi_k27_code_model(model_id):
+        # K2.7 cannot disable thinking. Send the enabled form explicitly
+        # because OpenAI-compatible gateways may not apply Moonshot's default.
+        return {"thinking": {"type": "enabled"}}
+
     if "kimi" in model_lower:
         return {
             "thinking": (
@@ -260,6 +304,7 @@ def build_provider_options(
     model_id: str,
     *,
     reasoning_enabled: Optional[bool] = None,
+    reasoning_effort: Optional[str] = None,
     thinking_budget: int = DEFAULT_THINKING_BUDGET,
     resolve_max_tokens: bool = True,
 ) -> Dict[str, Any]:
@@ -273,6 +318,8 @@ def build_provider_options(
         The model being called (e.g. ``"claude-sonnet-4"``).
     thinking_budget:
         Token budget for extended-thinking / reasoning where applicable.
+    reasoning_effort:
+        Kimi K3 reasoning effort (``low``, ``high``, or ``max``).
     resolve_max_tokens:
         If *True*, fall back to the model's configured ``max_tokens`` when
         no provider-specific logic has already set it.
@@ -285,11 +332,17 @@ def build_provider_options(
     model_lower = model_id.lower()
     interleaved_capability = _resolve_interleaved_capability(provider_id, model_id)
     reasoning_transport = _resolve_reasoning_transport(provider_id, model_id)
+    reasoning_effort_explicit = isinstance(reasoning_effort, str)
     reasoning_enabled = (
         _coerce_optional_bool(reasoning_enabled)
         if reasoning_enabled is not None
         else _resolve_reasoning_enabled(provider_id, model_id)
     )
+    reasoning_effort = (
+        reasoning_effort.strip().lower()
+        if isinstance(reasoning_effort, str)
+        else _resolve_reasoning_effort(provider_id, model_id)
+    )
     configured_extra_body = _resolve_default_extra_body(provider_id, model_id)
     interleaved_enabled = interleaved_capability is not None
     if interleaved_enabled and reasoning_enabled is None:
@@ -348,13 +401,37 @@ def build_provider_options(
         configured_extra_body
         or interleaved_enabled
         or reasoning_enabled is True
+        or is_kimi_k27_code_model(model_id)
+        or is_kimi_k3_model(model_id)
     ):
-        extra_body = configured_extra_body or _build_generic_chat_extra_body(
+        automatic_extra_body = _build_generic_chat_extra_body(
             provider_id,
             model_id,
             interleaved_capability,
             reasoning_enabled,
+            reasoning_effort,
         )
+        extra_body = dict(configured_extra_body or automatic_extra_body or {})
+        if is_kimi_k3_model(model_id):
+            # K3 is always a reasoning model and uses the top-level
+            # reasoning_effort field instead of the K2.x thinking object.
+            extra_body.pop("thinking", None)
+            if reasoning_effort_explicit:
+                extra_body["reasoning_effort"] = (automatic_extra_body or {}).get(
+                    "reasoning_effort",
+                    DEFAULT_KIMI_K3_REASONING_EFFORT,
+                )
+            else:
+                configured_effort = extra_body.get("reasoning_effort")
+                if configured_effort not in KIMI_K3_REASONING_EFFORTS:
+                    extra_body["reasoning_effort"] = (automatic_extra_body or {}).get(
+                        "reasoning_effort",
+                        DEFAULT_KIMI_K3_REASONING_EFFORT,
+                    )
+        elif is_kimi_k27_code_model(model_id):
+            # K2.7 rejects disabled thinking. Normalize configured values so
+            # direct and gateway-backed endpoints always request reasoning.
+            extra_body["thinking"] = {"type": "enabled"}
         if extra_body:
             options["extra_body"] = extra_body
             log.debug("options.thinking_params.resolved", {
diff --git a/flocks/provider/provider.py b/flocks/provider/provider.py
index 89c815c71..65465ee5c 100644
--- a/flocks/provider/provider.py
+++ b/flocks/provider/provider.py
@@ -4,11 +4,13 @@
 Manages different AI model providers (Anthropic, OpenAI, Google, etc.)
 """
 
-from typing import Dict, List, Optional, Any, AsyncIterator, Union
-from pydantic import BaseModel, Field, PrivateAttr
 from enum import Enum
+import json
 import os
 import threading
+from typing import Any, AsyncIterator, Dict, List, Optional, Union
+
+from pydantic import BaseModel, Field, PrivateAttr
 
 from flocks.utils.log import Log
 from flocks.config.config import Config
@@ -52,6 +54,11 @@ def _model_info_signature(model: "ModelInfo") -> tuple:
         if pricing is not None
         else None
     )
+    custom_settings = json.dumps(
+        getattr(model, "custom_settings", None) or {},
+        default=str,
+        sort_keys=True,
+    )
     explicit = tuple(sorted(getattr(model, "_explicit_keys", set()) or set()))
     return (
         getattr(model, "id", None),
@@ -59,6 +66,7 @@ def _model_info_signature(model: "ModelInfo") -> tuple:
         getattr(model, "provider_id", None),
         cap_sig,
         pricing_sig,
+        custom_settings,
         explicit,
     )
 
@@ -124,6 +132,7 @@ class ModelInfo(BaseModel):
     provider_id: str = Field(..., description="Provider ID")
     capabilities: ModelCapabilities = Field(default_factory=ModelCapabilities)
     pricing: Optional[Dict[str, Any]] = Field(None, description="Pricing info")
+    custom_settings: Dict[str, Any] = Field(default_factory=dict)
     _explicit_keys: set = PrivateAttr(default_factory=set)
     """Field names explicitly present in flocks.json (not defaults)."""
 
@@ -737,66 +746,72 @@ async def apply_config(cls, config: Optional[Any] = None, provider_id: Optional[
             if not provider:
                 continue
 
+            # Provider credentials/options are optional. Model definitions and
+            # display names below still need to load when credentials are
+            # unresolved or supplied outside the main config.
             options = getattr(pconfig, "options", None)
-            if not options:
-                continue
-
+            options_data: Optional[Dict[str, Any]] = None
             if hasattr(options, "model_dump"):
-                options_data = options.model_dump(exclude_none=True, by_alias=False)
+                options_data = options.model_dump(
+                    exclude_none=True,
+                    by_alias=False,
+                )
             elif isinstance(options, dict):
-                options_data = {k: v for k, v in options.items() if v is not None}
-            else:
-                continue
-
-            # Handle both Python-style (api_key, base_url) and JS-style (apiKey, baseURL)
-            api_key = (
-                options_data.pop("api_key", None)
-                or options_data.pop("apiKey", None)
-            )
-            base_url = (
-                options_data.pop("base_url", None)
-                or options_data.pop("baseURL", None)
-            )
-
-            # Treat empty strings as None (e.g. unresolved {secret:xxx})
-            if isinstance(api_key, str) and not api_key.strip():
-                api_key = None
-            if isinstance(base_url, str) and not base_url.strip():
-                base_url = None
-
-            # Also filter out remaining options that resolved to empty strings
-            options_data = {
-                k: v for k, v in options_data.items()
-                if not (isinstance(v, str) and not v.strip())
-            }
-
-            if api_key is None and base_url is None and not options_data:
-                continue
+                options_data = {
+                    key: value
+                    for key, value in options.items()
+                    if value is not None
+                }
+
+            if options_data is not None:
+                # Handle both Python-style (api_key, base_url) and JS-style
+                # (apiKey, baseURL).
+                api_key = (
+                    options_data.pop("api_key", None)
+                    or options_data.pop("apiKey", None)
+                )
+                base_url = (
+                    options_data.pop("base_url", None)
+                    or options_data.pop("baseURL", None)
+                )
 
-            # ----- Idempotent ProviderConfig update -------------------------------
-            # ``apply_config`` is called from many hot paths: every session
-            # step (``session.runner._step``), every workflow ``llm.ask``,
-            # the ``/session/*`` HTTP routes, plus startup. When session and
-            # workflow run concurrently on different event loops they would
-            # otherwise rewrite the same ``provider._config`` repeatedly and
-            # race on the ``_config_models`` rebuild. Skip mutation whenever
-            # the desired config already matches.
-            desired_cfg = ProviderConfig(
-                provider_id=pid,
-                api_key=api_key,
-                base_url=base_url,
-                custom_settings=options_data,
-            )
-            current_cfg = provider._config
-            current_unchanged = (
-                current_cfg is not None
-                and getattr(current_cfg, "api_key", None) == desired_cfg.api_key
-                and getattr(current_cfg, "base_url", None) == desired_cfg.base_url
-                and (getattr(current_cfg, "custom_settings", None) or {})
-                == (desired_cfg.custom_settings or {})
-            )
-            if not current_unchanged:
-                provider.configure(desired_cfg)
+                # Treat empty strings as None (e.g. unresolved {secret:xxx}).
+                if isinstance(api_key, str) and not api_key.strip():
+                    api_key = None
+                if isinstance(base_url, str) and not base_url.strip():
+                    base_url = None
+
+                # Also filter out remaining options that resolved to empty strings.
+                options_data = {
+                    key: value
+                    for key, value in options_data.items()
+                    if not (isinstance(value, str) and not value.strip())
+                }
+
+                if api_key is not None or base_url is not None or options_data:
+                    # ----- Idempotent ProviderConfig update -------------------
+                    # ``apply_config`` is called from many hot paths: every
+                    # session step, every workflow ``llm.ask``, HTTP routes,
+                    # and startup. Skip mutation whenever the desired config
+                    # already matches.
+                    desired_cfg = ProviderConfig(
+                        provider_id=pid,
+                        api_key=api_key,
+                        base_url=base_url,
+                        custom_settings=options_data,
+                    )
+                    current_cfg = provider._config
+                    current_unchanged = (
+                        current_cfg is not None
+                        and getattr(current_cfg, "api_key", None)
+                        == desired_cfg.api_key
+                        and getattr(current_cfg, "base_url", None)
+                        == desired_cfg.base_url
+                        and (getattr(current_cfg, "custom_settings", None) or {})
+                        == (desired_cfg.custom_settings or {})
+                    )
+                    if not current_unchanged:
+                        provider.configure(desired_cfg)
 
             # Update provider display name from flocks.json only for providers
             # that support custom naming (openai-compatible instances and custom-* providers).
@@ -855,6 +870,16 @@ async def apply_config(cls, config: Optional[Any] = None, provider_id: Optional[
                                     context_window=model_dict.get("context_window"),
                                 ),
                                 pricing=_pricing,
+                                custom_settings={
+                                    key: model_dict[key]
+                                    for key in (
+                                        "stream_first_chunk_timeout_s",
+                                        "streamFirstChunkTimeoutSeconds",
+                                        "stream_ongoing_chunk_timeout_s",
+                                        "streamOngoingChunkTimeoutSeconds",
+                                    )
+                                    if key in model_dict
+                                },
                             )
                             model_info._explicit_keys = _explicit_keys
                             desired_models.append(model_info)
@@ -1036,7 +1061,10 @@ def __init__(self, provider_id: str, name: str):
     
     def configure(self, config: ProviderConfig) -> None:
         """Configure the provider"""
+        config_changed = self._config != config
         self._config = config
+        if config_changed and hasattr(self, "_client"):
+            self._client = None
         self.log.info("provider.configured", {
             "provider_id": self.id,
             "has_api_key": config.api_key is not None,
diff --git a/flocks/provider/sdk/openai.py b/flocks/provider/sdk/openai.py
index ad33a18f3..a360b6e71 100644
--- a/flocks/provider/sdk/openai.py
+++ b/flocks/provider/sdk/openai.py
@@ -18,6 +18,7 @@
 )
 from flocks.provider.sdk.openai_base import (
     DEFAULT_HTTP_TIMEOUT,
+    _normalize_stream_usage,
     build_reasoning_metadata,
     _coerce_bool,
     extract_reasoning_content_with_source,
@@ -219,15 +220,12 @@ async def chat_stream(
         
         # Track usage from final chunk (when stream_options.include_usage is set)
         stream_usage: Optional[Dict[str, int]] = None
+        usage_emitted = False
         
         async for chunk in stream:
             # Capture usage from the final chunk (OpenAI returns it in a chunk with no choices)
             if hasattr(chunk, 'usage') and chunk.usage:
-                stream_usage = {
-                    "prompt_tokens": getattr(chunk.usage, 'prompt_tokens', 0) or 0,
-                    "completion_tokens": getattr(chunk.usage, 'completion_tokens', 0) or 0,
-                    "total_tokens": getattr(chunk.usage, 'total_tokens', 0) or 0,
-                }
+                stream_usage = _normalize_stream_usage(chunk.usage)
             
             if not chunk.choices:
                 continue
@@ -251,6 +249,7 @@ async def chat_stream(
                             finish_reason=choice.finish_reason,
                             usage=stream_usage,
                         )
+                    usage_emitted = stream_usage is not None
                 continue
 
             # Handle reasoning/thinking content (for o1/o3/gpt-5 models)
@@ -315,6 +314,13 @@ async def chat_stream(
                         finish_reason=choice.finish_reason,
                         usage=stream_usage,
                     )
+                usage_emitted = stream_usage is not None
+
+        # OpenAI sends the usage-only chunk after the terminal finish chunk.
+        # Surface it separately when it was not available on the terminal chunk
+        # so the runner can persist usage and nested reasoning tokens.
+        if stream_usage and not usage_emitted:
+            yield StreamChunk(delta="", finish_reason=None, usage=stream_usage)
     
     # Embeddings support (added for memory system)
     async def embed(
diff --git a/flocks/provider/sdk/openai_base.py b/flocks/provider/sdk/openai_base.py
index e99d36536..b56cb05b2 100644
--- a/flocks/provider/sdk/openai_base.py
+++ b/flocks/provider/sdk/openai_base.py
@@ -20,6 +20,7 @@
     ModelInfo,
     StreamChunk,
 )
+from flocks.provider.interleaved import is_kimi_k27_code_model, is_kimi_k3_model
 from flocks.utils.log import Log
 
 log = Log.create(service="provider.openai_base")
@@ -301,7 +302,26 @@ def _normalize_stream_usage(raw_usage: Any) -> Optional[Dict[str, int]]:
     prompt_tokens = (_pt if _pt is not None else getattr(raw_usage, "input_tokens", 0)) or 0
     _ct = getattr(raw_usage, "completion_tokens", None)
     completion_tokens = (_ct if _ct is not None else getattr(raw_usage, "output_tokens", 0)) or 0
-    reasoning_tokens = getattr(raw_usage, "reasoning_tokens", 0) or 0
+    completion_details = getattr(raw_usage, "completion_tokens_details", None)
+    output_details = getattr(raw_usage, "output_tokens_details", None)
+    nested_reasoning_tokens = (
+        getattr(completion_details, "reasoning_tokens", 0)
+        or getattr(output_details, "reasoning_tokens", 0)
+        or 0
+    )
+    reasoning_tokens = (
+        getattr(raw_usage, "reasoning_tokens", 0)
+        or nested_reasoning_tokens
+        or 0
+    )
+    if nested_reasoning_tokens:
+        # OpenAI reports reasoning_tokens as a breakdown already included in
+        # completion_tokens / output_tokens. Flocks stores visible output and
+        # reasoning separately, so remove that nested subset before persisting.
+        completion_tokens = max(
+            0,
+            completion_tokens - nested_reasoning_tokens,
+        )
     total_tokens = getattr(raw_usage, "total_tokens", 0) or (
         prompt_tokens + completion_tokens + reasoning_tokens
     )
@@ -390,6 +410,33 @@ def apply_openai_token_limit(
         params["max_tokens"] = max_tokens
 
 
+def _is_effective_thinking_enabled(
+    model_id: str,
+    thinking: Any,
+    extra_body: Dict[str, Any],
+) -> bool:
+    """Return the effective reasoning state represented by a request."""
+    if is_kimi_k27_code_model(model_id) or is_kimi_k3_model(model_id):
+        return True
+
+    if isinstance(thinking, dict):
+        return thinking.get("type") != "disabled"
+    if thinking:
+        return True
+
+    extra_thinking = extra_body.get("thinking")
+    if isinstance(extra_thinking, dict):
+        return extra_thinking.get("type") != "disabled"
+    if extra_thinking:
+        return True
+
+    if extra_body.get("reasoning_effort") is not None:
+        return True
+    if extra_body.get("reasoning_split") is True:
+        return True
+    return extra_body.get("enable_thinking") is True
+
+
 async def create_chat_completion_with_fallbacks(
     create_call,
     params: Dict[str, Any],
@@ -934,7 +981,9 @@ async def chat(
         apply_openai_token_limit(
             params,
             max_tokens,
-            prefer_completion_tokens=self.PREFER_MAX_COMPLETION_TOKENS,
+            prefer_completion_tokens=(
+                self.PREFER_MAX_COMPLETION_TOKENS or is_kimi_k3_model(model_id)
+            ),
             completion_tokens_explicit=max_completion_tokens_explicit,
         )
         if kwargs.get("tools"):
@@ -945,7 +994,11 @@ async def chat(
         # these request logs are emitted for every model call in long sessions.
         log.info("openai_base.chat.request", {
             "model": model_id,
-            "thinking_enabled": bool(thinking),
+            "thinking_enabled": _is_effective_thinking_enabled(
+                model_id,
+                thinking,
+                extra_body,
+            ),
             "has_extra_body": "extra_body" in params,
             "has_tools": bool(kwargs.get("tools")),
             "max_tokens": max_tokens,
@@ -1021,7 +1074,9 @@ async def chat_stream(
         apply_openai_token_limit(
             params,
             max_tokens,
-            prefer_completion_tokens=self.PREFER_MAX_COMPLETION_TOKENS,
+            prefer_completion_tokens=(
+                self.PREFER_MAX_COMPLETION_TOKENS or is_kimi_k3_model(model_id)
+            ),
             completion_tokens_explicit=max_completion_tokens_explicit,
         )
         if kwargs.get("tools"):
@@ -1031,7 +1086,11 @@ async def chat_stream(
         # without repeating the full message history on every turn.
         log.info("openai_base.stream.request", {
             "model": model_id,
-            "thinking_enabled": bool(thinking),
+            "thinking_enabled": _is_effective_thinking_enabled(
+                model_id,
+                thinking,
+                extra_body,
+            ),
             "has_extra_body": "extra_body" in params,
             "has_tools": bool(kwargs.get("tools")),
             "max_tokens": max_tokens,
@@ -1048,6 +1107,7 @@ async def chat_stream(
             log_prefix="openai_base",
         )
         tool_calls: Dict[int, Dict[str, Any]] = {}
+        started_tool_inputs: set[int] = set()
         emitted_substantive_chunk = False
         stream_usage: Optional[Dict[str, int]] = None
         usage_emitted = False
@@ -1131,6 +1191,7 @@ async def chat_stream(
                 delta_tcs = getattr(delta, "tool_calls", None)
                 if delta_tcs:
                     emitted_substantive_chunk = True
+                    tool_input_markers: List[Dict[str, Any]] = []
                     for tc in delta_tcs:
                         idx = tc.index
                         if idx not in tool_calls:
@@ -1148,6 +1209,28 @@ async def chat_stream(
                                 tool_calls[idx]["function"]["arguments"] += (
                                     tc.function.arguments
                                 )
+                        accumulated_name = tool_calls[idx]["function"]["name"]
+                        if accumulated_name and idx not in started_tool_inputs:
+                            tool_input_markers.append({
+                                "index": idx,
+                                "id": tool_calls[idx]["id"],
+                                "type": "function",
+                                "function": {
+                                    "name": accumulated_name,
+                                    "arguments": "",
+                                },
+                            })
+                            started_tool_inputs.add(idx)
+
+                    # Surface the tool as soon as its name is known, but keep
+                    # partial JSON private. The terminal chunk below publishes
+                    # the complete input once the model finishes generating it.
+                    if tool_input_markers:
+                        yield StreamChunk(
+                            delta="",
+                            finish_reason=None,
+                            tool_calls=tool_input_markers,
+                        )
 
             if choice.finish_reason:
                 # Flush any remaining buffered content from the think-tag extractor
@@ -1173,7 +1256,10 @@ async def chat_stream(
                         yield StreamChunk(delta=seg_text, finish_reason=None)
 
                 if tool_calls:
-                    sorted_calls = [tool_calls[i] for i in sorted(tool_calls.keys())]
+                    sorted_calls = [
+                        {"index": i, **tool_calls[i]}
+                        for i in sorted(tool_calls.keys())
+                    ]
                     tool_calls.clear()
                     # Preserve real finish_reason (e.g. "length" when max_tokens
                     # hit) so the runner can detect truncated tool arguments.
diff --git a/flocks/provider/sdk/openai_compatible.py b/flocks/provider/sdk/openai_compatible.py
index 5f21424f1..812131da3 100644
--- a/flocks/provider/sdk/openai_compatible.py
+++ b/flocks/provider/sdk/openai_compatible.py
@@ -28,6 +28,7 @@
     build_reasoning_metadata,
     create_chat_completion_with_fallbacks,
     _coerce_bool,
+    _is_effective_thinking_enabled,
     _normalize_stream_usage,
     extract_reasoning_content_with_source,
     extract_reasoning_details,
@@ -303,7 +304,11 @@ async def chat_stream(
         
         self.log.info("openai_compatible.stream.request", {
             "model": model_id,
-            "thinking_enabled": bool(thinking),
+            "thinking_enabled": _is_effective_thinking_enabled(
+                model_id,
+                thinking,
+                extra_body,
+            ),
             "has_tools": bool(tools),
             "max_tokens": max_tokens,
             "include_usage": True,
diff --git a/flocks/security/__init__.py b/flocks/security/__init__.py
index 434072168..41f97db7f 100644
--- a/flocks/security/__init__.py
+++ b/flocks/security/__init__.py
@@ -52,7 +52,7 @@ def _resolve_fofa_derived_secret(secret_id: str, secrets: SecretManager) -> Opti
 
 
 def resolve_secret_value(secret_id: str, secrets: Optional[SecretManager] = None) -> Optional[str]:
-    """Resolve a secret by id, including provider-specific derived secrets."""
+    """Resolve a secret by id, including read-only legacy fallbacks."""
     if secrets is None:
         secrets = get_secret_manager()
 
@@ -60,6 +60,12 @@ def resolve_secret_value(secret_id: str, secrets: Optional[SecretManager] = None
     if value is not None:
         return value
 
+    if secret_id.endswith("_llm_key"):
+        provider_id = secret_id.removesuffix("_llm_key")
+        legacy_value = secrets.get(f"{provider_id}_api_key")
+        if legacy_value is not None:
+            return legacy_value
+
     return _resolve_fofa_derived_secret(secret_id, secrets)
 
 
diff --git a/flocks/server/auth.py b/flocks/server/auth.py
index 046fd32f2..47b08e4fc 100644
--- a/flocks/server/auth.py
+++ b/flocks/server/auth.py
@@ -12,13 +12,17 @@
 from fastapi import HTTPException, Request, Response, status
 from starlette.requests import HTTPConnection
 
-from flocks.auth.context import AuthUser, reset_current_auth_user, set_current_auth_user
+from flocks.auth.context import (
+    API_TOKEN_SERVICE_USER_ID,
+    AuthUser,
+    reset_current_auth_user,
+    set_current_auth_user,
+)
 from flocks.auth.service import AuthService
 from flocks.security import get_secret_manager
 
 SESSION_COOKIE_NAME = "flocks_session"
 API_TOKEN_SECRET_ID = "server_api_token"
-API_TOKEN_SERVICE_USER_ID = "api-token-service"
 
 # Paths that never require auth. Everything else is protected by default.
 PUBLIC_PATHS = frozenset({
diff --git a/flocks/server/routes/default_model.py b/flocks/server/routes/default_model.py
index cda00f077..5aa58b862 100644
--- a/flocks/server/routes/default_model.py
+++ b/flocks/server/routes/default_model.py
@@ -4,7 +4,7 @@
 Provides endpoints to get/set default models per model type.
 """
 
-from typing import List, Optional
+from typing import List
 
 from fastapi import APIRouter, HTTPException, status
 from pydantic import BaseModel, Field
@@ -58,6 +58,7 @@ async def get_all_defaults() -> DefaultModelListResponse:
 async def get_resolved_default_model():
     """Return the resolved default LLM model (provider_id + model_id)."""
     from flocks.config.config import Config
+
     result = await Config.resolve_default_llm()
     if not result:
         raise HTTPException(
diff --git a/flocks/server/routes/project.py b/flocks/server/routes/project.py
index 9143f1f0d..ab6ec565e 100644
--- a/flocks/server/routes/project.py
+++ b/flocks/server/routes/project.py
@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import asyncio
 import os
 from pathlib import Path
 from typing import List, Optional
@@ -19,7 +20,7 @@
 )
 from flocks.server.auth import require_user
 from flocks.session.policy import SessionPolicy
-from flocks.session.session import Session
+from flocks.session.session import Session, SessionInfo
 from flocks.utils.log import Log
 
 router = APIRouter()
@@ -76,6 +77,8 @@ async def _list_project_summaries(user: AuthUser, search: Optional[str]) -> List
                 shared_project_ids=shared_project_ids,
             ):
                 continue
+            if session.status != "active":
+                continue
             metadata = session.metadata if isinstance(session.metadata, dict) else {}
             if metadata.get("hideFromSessionManager"):
                 continue
@@ -285,31 +288,56 @@ async def unshare_project_local(project_id: str, request: Request):
 
 @router.delete("/{project_id}", response_model=bool, summary="Delete project")
 async def delete_project(project_id: str, request: Request):
-    """Remove a project and its sessions while preserving project files."""
+    """Archive project tasks and hide the project while preserving project files."""
 
     user = require_user(request)
     try:
-        if await Project.get(project_id, owner_id=user.id) is None:
-            raise ValueError(f"Project {project_id} not found")
-
-        project_sessions = [
-            session
-            for session in await Session.list_all_unfiltered()
-            if session.project_id == project_id
-        ]
-        if any(not SessionPolicy.can_delete(session, user) for session in project_sessions):
-            raise HTTPException(
-                status_code=403,
-                detail="Only session owners can delete project sessions",
-            )
-
-        from flocks.server.routes.session import delete_session_for_user
-
-        for session in project_sessions:
-            if await Session.get(project_id, session.id) is not None:
-                await delete_session_for_user(session.id, user)
-
-        return await Project.delete(project_id, owner_id=user.id)
+        async with Project.lifecycle_guard(project_id):
+            if await Project.get(project_id, owner_id=user.id) is None:
+                raise ValueError(f"Project {project_id} not found")
+
+            project_sessions = [
+                session
+                for session in await Session.list_all_unfiltered()
+                if session.project_id == project_id
+            ]
+            if any(not SessionPolicy.can_delete(session, user) for session in project_sessions):
+                raise HTTPException(
+                    status_code=403,
+                    detail="Only session owners can delete project sessions",
+                )
+
+            root_sessions = [
+                session
+                for session in project_sessions
+                if session.parent_id is None
+            ]
+            if project_sessions and not root_sessions:
+                raise ProjectDeletionError("Project sessions do not contain a restorable root task")
+
+            from flocks.server.routes.session import archive_session_for_user
+
+            newly_archived: list[SessionInfo] = []
+            try:
+                for session in root_sessions:
+                    await archive_session_for_user(session.id, user)
+                    if session.status != "archived":
+                        newly_archived.append(session)
+
+                return await Project.delete(project_id, owner_id=user.id)
+            except BaseException:
+                for session in reversed(newly_archived):
+                    try:
+                        await asyncio.shield(
+                            Session._unarchive_locked(project_id, session.id)
+                        )
+                    except Exception as rollback_exc:
+                        log.warn("project.delete.rollback_failed", {
+                            "project_id": project_id,
+                            "session_id": session.id,
+                            "error": str(rollback_exc),
+                        })
+                raise
     except ProjectDeletionError as exc:
         raise HTTPException(status_code=409, detail=str(exc)) from exc
     except ValueError as exc:
diff --git a/flocks/server/routes/provider.py b/flocks/server/routes/provider.py
index 9ec50d8fe..c35f13b49 100644
--- a/flocks/server/routes/provider.py
+++ b/flocks/server/routes/provider.py
@@ -500,6 +500,7 @@ class ProviderInfo(BaseModel):
     source: str = Field(default="config", description="Provider source: env, config, custom, api")
     env: List[str] = Field(default_factory=list, description="Environment variable names")
     key: Optional[str] = Field(None, description="API key (if configured)")
+    configured: bool = Field(False, description="Whether runtime credentials are configured")
     options: Dict[str, Any] = Field(default_factory=dict, description="Provider options")
     # Flocks expects models as Dict[modelID, Model], not List[Model]
     models: Dict[str, Dict[str, Any]] = Field(default_factory=dict, description="Available models")
@@ -596,6 +597,7 @@ async def list_providers() -> ProviderListResponse:
                 source="config",  # Provider source: env, config, custom, api
                 env=[],  # Environment variable names (can be enhanced later)
                 key=None,  # API key not exposed in list
+                configured=provider.is_configured(),
                 options={},
                 models=models_dict,
             )
@@ -1936,6 +1938,10 @@ async def set_provider_credentials(
                 config_dict["name"] = request.provider_name
             ConfigWriter.add_provider(provider_id, config_dict)
 
+        # Config.get() caches resolved secret values. Invalidate it after
+        # persisting credentials so a later session cannot reapply the old key.
+        Config.clear_cache()
+
         # 3. Configure the provider runtime so is_configured() reflects the change
         await _ensure_provider_initialized()
         provider = Provider.get(provider_id)
diff --git a/flocks/server/routes/session.py b/flocks/server/routes/session.py
index 2f9f2acad..cbc6c3bdd 100644
--- a/flocks/server/routes/session.py
+++ b/flocks/server/routes/session.py
@@ -25,8 +25,20 @@
     build_context_usage_snapshot,
     token_usage_to_dict,
 )
-from flocks.session.session import Session, SessionInfo as SessionModel
+from flocks.session.background_tasks import (
+    has_pending_session_tasks,
+    pending_background_tasks,
+    track_background_task,
+)
+from flocks.session.session import (
+    Session,
+    SessionInactiveError,
+    SessionInfo as SessionModel,
+    SessionNotFoundError,
+    is_model_auto_session_category,
+)
 from flocks.session.policy import SessionPolicy
+from flocks.session.execution_mode import SessionExecutionMode
 from flocks.utils.log import Log
 from flocks.utils.json_repair import parse_json_robust, repair_truncated_json
 from flocks.utils.monitor import get_monitor
@@ -38,7 +50,6 @@
 # Default agent name constant
 DEFAULT_AGENT = "rex"
 DEFAULT_MESSAGE_PAGE_LIMIT = 50
-_DESCENDANT_ABORT_SCAN_LIMIT = 3
 _CONTEXT_USAGE_CACHE_TTL_SECONDS = 5.0
 _context_usage_cache: Dict[Tuple[str, int], Tuple[float, ContextUsageSnapshot]] = {}
 _context_usage_inflight: Dict[Tuple[str, int], asyncio.Task[ContextUsageSnapshot]] = {}
@@ -63,13 +74,9 @@
 
 def _session_uploads_dir(session_id: str) -> Path:
     """Return the application-owned upload directory for one session."""
-    from flocks.config.config import Config
+    from flocks.session.files import session_uploads_dir
 
-    uploads_root = (Config.get_data_path() / "uploads").resolve()
-    target = (uploads_root / session_id).resolve()
-    if target == uploads_root or not target.is_relative_to(uploads_root):
-        raise ValueError(f"Invalid session ID for upload path: {session_id}")
-    return target
+    return session_uploads_dir(session_id)
 
 
 def _materialize_data_url_part(
@@ -191,6 +198,10 @@ class SessionCreateRequest(BaseModel):
     title: Optional[str] = Field(None, description="Session title")
     permission: Optional[List[PermissionRule]] = Field(None, description="Permission rules")
     category: Optional[str] = Field(None, description="Session category (e.g. 'user', 'workflow')")
+    model_auto: bool = Field(
+        False,
+        description="Enable WebUI runtime model failover for this session",
+    )
 
 
 class FileDiff(BaseModel):
@@ -244,9 +255,11 @@ class SessionResponse(BaseModel):
     permission: Optional[List[Dict[str, Any]]] = Field(None, description="Permission rules")
     revert: Optional[Dict[str, Any]] = Field(None, description="Revert state")
     category: str = Field("user", description="Session category: user or task")
+    status: str = Field("active", description="Session status: active or archived")
     provider: Optional[str] = Field(None, description="Pinned provider ID")
     model: Optional[str] = Field(None, description="Pinned model ID")
     model_pinned: bool = Field(False, description="Whether provider/model are pinned for this session")
+    model_auto: bool = Field(False, description="Whether WebUI Auto mode is selected")
     ownerUserID: Optional[str] = Field(None, description="Session owner user id")
     ownerUsername: Optional[str] = Field(None, description="Session owner username")
     canWrite: bool = Field(False, description="Whether current user can continue this session")
@@ -261,15 +274,20 @@ class SessionListItem(BaseModel):
 
     id: str
     projectID: str
+    projectName: Optional[str] = None
     effectiveProjectID: str
     directory: str
     title: str
     time: SessionTime
     category: str = "user"
+    status: str = "active"
     parentID: Optional[str] = None
     provider: Optional[str] = None
     model: Optional[str] = None
     model_pinned: bool = False
+    model_auto: bool = False
+    ownerUserID: Optional[str] = None
+    ownerUsername: Optional[str] = None
     canWrite: bool = False
     canDelete: bool = False
     isShared: bool = False
@@ -284,7 +302,7 @@ def _session_to_response(
     Convert SessionModel to SessionResponse
     """
     current_user = get_current_auth_user()
-    can_write = SessionPolicy.can_write(session, current_user)
+    can_write = session.status == "active" and SessionPolicy.can_write(session, current_user)
     can_delete = SessionPolicy.can_delete(session, current_user)
     is_shared = SessionPolicy.is_shared(session, shared_project_ids)
     from flocks.project.project import Project
@@ -314,9 +332,11 @@ def _session_to_response(
         revert=session.revert.model_dump(by_alias=True) if session.revert else None,
         permission=[p.model_dump() for p in session.permission] if session.permission else None,
         category=session.category,
+        status=session.status,
         provider=session.provider,
         model=session.model,
         model_pinned=session.model_pinned,
+        model_auto=session.model_auto,
         ownerUserID=session.owner_user_id,
         ownerUsername=session.owner_username,
         canWrite=can_write,
@@ -329,6 +349,7 @@ def _session_to_list_item(
     session: SessionModel,
     effective_project_id: Optional[str] = None,
     shared_project_ids: Optional[set[str]] = None,
+    project_name: Optional[str] = None,
 ) -> SessionListItem:
     """Convert a session to the lightweight manager-list response shape."""
     current_user = get_current_auth_user()
@@ -342,6 +363,7 @@ def _session_to_list_item(
     return SessionListItem(
         id=session.id,
         projectID=session.project_id,
+        projectName=project_name,
         effectiveProjectID=effective_project_id,
         directory=session.directory,
         title=session.title,
@@ -352,11 +374,15 @@ def _session_to_list_item(
             archived=session.time.archived,
         ),
         category=session.category,
+        status=session.status,
         parentID=session.parent_id,
         provider=session.provider,
         model=session.model,
         model_pinned=session.model_pinned,
-        canWrite=SessionPolicy.can_write(session, current_user),
+        model_auto=session.model_auto,
+        ownerUserID=session.owner_user_id,
+        ownerUsername=session.owner_username,
+        canWrite=session.status == "active" and SessionPolicy.can_write(session, current_user),
         canDelete=SessionPolicy.can_delete(session, current_user),
         isShared=SessionPolicy.is_shared(session, shared_project_ids),
     )
@@ -394,6 +420,33 @@ def _require_session_read_access(session: SessionModel, user) -> None:
 def _require_session_write_access(session: SessionModel, user) -> None:
     if not SessionPolicy.can_write(session, user):
         raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail="仅会话所有者可写，受邀用户为只读")
+    if session.status != "active":
+        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail="已归档会话不可修改，请先恢复")
+
+
+async def _persist_active_session_write(
+    session_id: str,
+    operation,
+    *,
+    expected_generation: Optional[int] = None,
+):
+    """Linearize a durable route write with archive/delete transitions."""
+    try:
+        return await Session.run_active_write(
+            session_id,
+            operation,
+            expected_generation=expected_generation,
+        )
+    except SessionNotFoundError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail=f"Session {session_id} not found",
+        ) from exc
+    except SessionInactiveError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="已归档会话不可修改，请先恢复",
+        ) from exc
 
 
 async def _require_agent_usable_for_chat(agent_name: Optional[str]) -> None:
@@ -449,11 +502,7 @@ def _share_metadata(session: SessionModel, *, shared: bool, actor_user_id: str)
 
 async def _get_session_by_id_unfiltered(session_id: str) -> Optional[SessionModel]:
     """Fetch session by id while bypassing policy filtering."""
-    token = set_current_auth_user(None)
-    try:
-        return await Session.get_by_id(session_id)
-    finally:
-        reset_current_auth_user(token)
+    return await Session.get_by_id_unfiltered(session_id)
 
 
 async def _resolve_session_working_directory(session: SessionModel) -> str:
@@ -474,6 +523,7 @@ async def _resolve_session_working_directory(session: SessionModel) -> str:
             detail="The session directory and current request directory are unavailable",
         )
     fallback_directory = str(fallback_path.resolve())
+
     log.warn(
         "session.directory.fallback",
         {
@@ -586,6 +636,11 @@ async def list_sessions(
     limit: Optional[int] = Query(None, ge=1, description="Maximum sessions to return"),
     offset: Optional[int] = Query(None, ge=0, description="Number of filtered sessions to skip"),
     category: Optional[str] = Query(None, description="Filter by category: user or task"),
+    session_status: Literal["active", "archived", "all"] = Query(
+        "active",
+        alias="status",
+        description="Filter by session status",
+    ),
 ) -> List[Union[SessionResponse, SessionListItem]]:
     """List all sessions with optional filters"""
     started_at = time.perf_counter()
@@ -597,6 +652,11 @@ async def list_sessions(
     )
     list_started_at = time.perf_counter()
     all_sessions = await Session.list_all_unfiltered()
+    if session_status == "archived":
+        all_sessions.sort(
+            key=lambda item: item.time.archived or item.time.updated,
+            reverse=True,
+        )
     list_elapsed_ms = (time.perf_counter() - list_started_at) * 1000
     visible_project_ids = Project.visible_project_ids(current_user.id)
     shared_project_ids = Project.shared_project_ids()
@@ -605,10 +665,14 @@ async def list_sessions(
     effective_project_ids: Dict[str, str] = {}
     term = search.lower() if search else None
     manager_categories = {"user", "workflow", "entity-config"}
+    project_names = Project.registered_project_names() if view == "list" else {}
     skip_remaining = offset or 0
     
     for session in all_sessions:
-        if not SessionPolicy.can_read(
+        if session.status == "archived":
+            if current_user.role != "admin" and not SessionPolicy.is_owner(session, current_user):
+                continue
+        elif not SessionPolicy.can_read(
             session,
             current_user,
             shared_project_ids=shared_project_ids,
@@ -616,6 +680,8 @@ async def list_sessions(
             continue
         if _is_hidden_from_session_manager(session):
             continue
+        if session_status != "all" and session.status != session_status:
+            continue
         if directory is not None and session.directory != directory:
             continue
         effective_project_id = (
@@ -658,7 +724,12 @@ async def list_sessions(
 
     if view == "list":
         response = [
-            _session_to_list_item(s, effective_project_ids[s.id], shared_project_ids)
+            _session_to_list_item(
+                s,
+                effective_project_ids[s.id],
+                shared_project_ids,
+                project_names.get(s.project_id),
+            )
             for s in filtered
         ]
         log_route_timing(log, "session.list.light.complete", started_at=started_at, extra={
@@ -670,6 +741,7 @@ async def list_sessions(
             "offset": offset,
             "search": bool(search),
             "category": category,
+            "status": session_status,
             "projectID": projectID,
             "list_ms": round(list_elapsed_ms, 2),
         })
@@ -691,6 +763,7 @@ async def list_sessions(
         "offset": offset,
         "search": bool(search),
         "category": category,
+        "status": session_status,
         "projectID": projectID,
         "list_ms": round(list_elapsed_ms, 2),
     })
@@ -710,17 +783,42 @@ async def create_session(http_request: Request, request: Optional[SessionCreateR
     await assert_license_active(feature="session_create")
     if request is None:
         request = SessionCreateRequest()
+    if request.model_auto:
+        category = request.category if request.category is not None else "user"
+        if not is_model_auto_session_category(category):
+            raise HTTPException(
+                status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+                detail=(
+                    "Auto mode is only available for user, entity "
+                    "configuration, and workflow sessions"
+                ),
+            )
+        if current_user.id == API_TOKEN_SERVICE_USER_ID:
+            raise HTTPException(
+                status_code=status.HTTP_403_FORBIDDEN,
+                detail="Auto mode can only be enabled from the WebUI",
+            )
+        from flocks.session.session_loop import SessionLoop
+
+        auto_available, auto_reason = await SessionLoop.validate_auto_configuration()
+        if not auto_available:
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail=f"Auto mode is unavailable: {auto_reason}",
+            )
 
     from flocks.project.instance import Instance
     from flocks.project.project import (
         DEFAULT_PROJECT_ID,
         Project,
+        ProjectDeletionError,
         TASK_SESSION_GROUP_ID,
     )
 
     directory = Instance.get_directory() or os.getcwd()
     instance_project = Instance.get_project()
     project_id = instance_project.id if instance_project else DEFAULT_PROJECT_ID
+    parent_session: Optional[SessionModel] = None
 
     # "default" is accepted for compatibility with older WebUI clients, but
     # ordinary sessions use the current request context rather than a virtual
@@ -742,6 +840,26 @@ async def create_session(http_request: Request, request: Optional[SessionCreateR
             )
         project_id = project.id
         directory = project.worktree
+
+    if request.parentID:
+        parent_session = await _get_session_by_id_unfiltered(request.parentID)
+        if parent_session is None:
+            raise HTTPException(
+                status_code=status.HTTP_404_NOT_FOUND,
+                detail=f"Parent session {request.parentID} not found",
+            )
+        _require_session_write_access(parent_session, current_user)
+        if request.projectID and request.projectID not in {
+            DEFAULT_PROJECT_ID,
+            TASK_SESSION_GROUP_ID,
+            parent_session.project_id,
+        }:
+            raise HTTPException(
+                status_code=status.HTTP_409_CONFLICT,
+                detail="Parent and child sessions must belong to the same project",
+            )
+        project_id = parent_session.project_id
+        directory = parent_session.directory
     
     # Trigger command:new hook if creating from parent (like /new command)
     if request.parentID:
@@ -785,18 +903,34 @@ async def create_session(http_request: Request, request: Optional[SessionCreateR
             for p in request.permission
         ]
 
-    is_api_token_client = current_user.id == API_TOKEN_SERVICE_USER_ID
-    
-    session = await Session.create(
-        project_id=project_id,
-        directory=directory,
-        title=request.title,
-        parent_id=request.parentID,
-        permission=permission,
-        owner_user_id=None if is_api_token_client else current_user.id,
-        owner_username=None if is_api_token_client else current_user.username,
-        **({"category": request.category} if request.category else {}),
-    )
+    if request.projectID and request.projectID not in {
+        DEFAULT_PROJECT_ID,
+        TASK_SESSION_GROUP_ID,
+    }:
+        project = await Project.get(request.projectID, owner_id=current_user.id)
+        if project is None:
+            raise HTTPException(
+                status_code=status.HTTP_409_CONFLICT,
+                detail=f"Project {request.projectID} is no longer available",
+            )
+    try:
+        session = await Session.create(
+            project_id=project_id,
+            directory=directory,
+            title=request.title,
+            parent_id=request.parentID,
+            permission=permission,
+            owner_user_id=(parent_session.owner_user_id if parent_session else current_user.id),
+            owner_username=(parent_session.owner_username if parent_session else current_user.username),
+            model_auto=request.model_auto,
+            model_pinned=False,
+            **({"category": request.category} if request.category else {}),
+        )
+    except ProjectDeletionError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail=str(exc),
+        ) from exc
     Project.invalidate_session_stats()
 
     log.info("session.created", {"session_id": session.id})
@@ -838,6 +972,11 @@ async def get_session(sessionID: str, request: Request) -> SessionResponse:
             detail=f"Session {sessionID} not found"
         )
     _require_session_read_access(session, _current_user)
+    if session.status != "active":
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail=f"Session {sessionID} not found",
+        )
     return await _session_to_response_with_goal(session)
 
 
@@ -941,31 +1080,169 @@ async def update_session_todos(sessionID: str, todos: List[TodoInfo], request: R
             detail=f"Session {sessionID} not found"
         )
     _require_session_write_access(session, _current_user)
+    lifecycle_generation = Session.lifecycle_generation(sessionID)
     try:
-        await Todo.update(
+        normalized_todos = [
+            SessionTodoInfo(**todo.model_dump(exclude_none=True))
+            for todo in todos
+        ]
+        await _persist_active_session_write(
             sessionID,
-            [SessionTodoInfo(**t.model_dump(exclude_none=True)) for t in todos],
+            lambda: Todo.update(sessionID, normalized_todos),
+            expected_generation=lifecycle_generation,
         )
         return todos
+    except HTTPException:
+        raise
     except Exception as e:
         log.error("session.todo.write_error", {"sessionID": sessionID, "error": str(e)})
         raise HTTPException(status_code=500, detail=str(e))
 
 
+@router.post(
+    "/{sessionID}/archive",
+    response_model=SessionResponse,
+    summary="Archive session",
+    description="Archive a session and its descendants without deleting persisted data",
+)
+async def archive_session(sessionID: str, request: Request) -> SessionResponse:
+    current_user = require_user(request)
+    return await archive_session_for_user(sessionID, current_user)
+
+
+@router.post(
+    "/{sessionID}/restore",
+    response_model=SessionResponse,
+    summary="Restore session",
+    description="Restore an archived session and its descendants",
+)
+async def restore_session(sessionID: str, request: Request) -> SessionResponse:
+    current_user = require_user(request)
+    return await restore_session_for_user(sessionID, current_user)
+
+
+async def _manageable_session_tree(session: SessionModel, current_user: AuthUser) -> List[SessionModel]:
+    tree = await Session.collect_tree(session.project_id, session.id)
+    if any(not SessionPolicy.can_delete(item, current_user) for item in tree):
+        raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail="仅会话所有者可管理归档会话")
+    return tree
+
+
+async def archive_session_for_user(session_id: str, current_user: AuthUser) -> SessionResponse:
+    """Stop and archive a complete session tree while retaining persisted data."""
+    session = await _get_session_by_id_unfiltered(session_id)
+    if not session:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail=f"Session {session_id} not found")
+
+    tree = await _manageable_session_tree(session, current_user)
+
+    auth_token = set_current_auth_user(current_user)
+    try:
+        archived_ok = await Session.archive(session.project_id, session.id)
+    finally:
+        reset_current_auth_user(auth_token)
+    if not archived_ok:
+        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail="会话归档失败")
+
+    from flocks.server.routes.event import publish_event
+
+    try:
+        await publish_event("session.updated", {"id": session_id, "status": "archived"})
+    except Exception as exc:
+        log.warn("session.archive.event_error", {"session_id": session_id, "error": str(exc)})
+    try:
+        await emit_audit_event("session_action", {
+            "action": "archive",
+            "actor_id": current_user.username,
+            "actor_name": current_user.username,
+            "user_name": current_user.username,
+            "username": current_user.username,
+            "session_id": session_id,
+            "owner_user_id": current_user.id,
+            "project_id": session.project_id,
+            "affected_sessions": len(tree),
+        })
+    except Exception:
+        pass
+
+    archived = await _get_session_by_id_unfiltered(session_id)
+    if archived is None:
+        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail="会话归档状态不可用")
+    return await _session_to_response_with_goal(archived)
+
+
+async def restore_session_for_user(session_id: str, current_user: AuthUser) -> SessionResponse:
+    """Restore a complete archived session tree."""
+    session = await _get_session_by_id_unfiltered(session_id)
+    if not session:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail=f"Session {session_id} not found")
+
+    tree = await _manageable_session_tree(session, current_user)
+    if session.parent_id is not None:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="只能从根任务恢复完整任务树",
+        )
+    project_owner_id = session.owner_user_id or current_user.id
+    auth_token = set_current_auth_user(current_user)
+    try:
+        try:
+            restored_ok = await Session.restore(
+                session.project_id,
+                session.id,
+                project_owner_id=project_owner_id,
+            )
+        except ValueError as exc:
+            raise HTTPException(
+                status_code=status.HTTP_409_CONFLICT,
+                detail=str(exc),
+            ) from exc
+    finally:
+        reset_current_auth_user(auth_token)
+    if not restored_ok:
+        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail="会话恢复失败")
+
+    from flocks.server.routes.event import publish_event
+
+    try:
+        await publish_event("session.updated", {"id": session_id, "status": "active"})
+    except Exception as exc:
+        log.warn("session.restore.event_error", {"session_id": session_id, "error": str(exc)})
+    try:
+        await emit_audit_event("session_action", {
+            "action": "restore",
+            "actor_id": current_user.username,
+            "actor_name": current_user.username,
+            "user_name": current_user.username,
+            "username": current_user.username,
+            "session_id": session_id,
+            "owner_user_id": current_user.id,
+            "project_id": session.project_id,
+            "affected_sessions": len(tree),
+        })
+    except Exception:
+        pass
+
+    restored = await _get_session_by_id_unfiltered(session_id)
+    if restored is None:
+        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail="会话恢复状态不可用")
+    return await _session_to_response_with_goal(restored)
+
+
 @router.delete(
     "/{sessionID}",
     status_code=status.HTTP_200_OK,
-    summary="Delete session",
-    description="Delete session by ID",
+    summary="Permanently delete session",
+    description="Permanently delete a session tree and its persisted history",
 )
 async def delete_session(sessionID: str, request: Request) -> bool:
-    """Delete session by ID (returns true)"""
+    """Permanently delete a session by ID (returns true)."""
     current_user = require_user(request)
     return await delete_session_for_user(sessionID, current_user)
 
 
 async def delete_session_for_user(session_id: str, current_user: AuthUser) -> bool:
-    """Delete one session using the normal lifecycle cleanup."""
+    """Permanently delete a session tree using the normal lifecycle cleanup."""
     session = await _get_session_by_id_unfiltered(session_id)
     
     if not session:
@@ -974,49 +1251,22 @@ async def delete_session_for_user(session_id: str, current_user: AuthUser) -> bo
             detail=f"Session {session_id} not found"
         )
     
-    if not SessionPolicy.can_delete(session, current_user):
-        raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail="仅会话所有者可删除会话")
-
-    from flocks.session.goal import GoalManager
-    from flocks.session.interaction_queue import InteractionQueue
-
-    await _abort_session_processing(session_id)
-    await InteractionQueue.clear(session_id)
-    await GoalManager.clear(session_id)
-    await _wait_for_sessions_idle([session_id])
-    await _abort_and_wait_descendant_sessions(session.project_id, session_id)
-    await Session.delete(session.project_id, session_id)
-    from flocks.project.project import Project
+    tree = await _manageable_session_tree(session, current_user)
+    tree_ids = [item.id for item in tree]
 
-    Project.invalidate_session_stats()
-
-    # Best-effort cleanup of any image/file uploads materialised for this
-    # session via ``_materialize_data_url_part`` (see prompt_async).
-    # The session DB row is gone, so the on-disk bytes are now orphaned —
-    # remove them to keep application data tidy. We deliberately swallow any
-    # filesystem errors: deletion of the session record is the contract,
-    # the upload cleanup is incidental.
+    auth_token = set_current_auth_user(current_user)
     try:
-        import shutil
-        uploads_root = _session_uploads_dir(session_id)
-        if uploads_root.exists() and uploads_root.is_dir():
-            shutil.rmtree(uploads_root, ignore_errors=True)
-            log.info("session.uploads.cleaned", {
-                "session_id": session_id,
-                "path": str(uploads_root),
-            })
-    except Exception as exc:
-        log.warn("session.uploads.cleanup_failed", {
-            "session_id": session_id,
-            "error": str(exc),
-        })
-
+        deleted_ok = await Session.delete(session.project_id, session_id)
+    finally:
+        reset_current_auth_user(auth_token)
+    if not deleted_ok:
+        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail="会话永久删除失败")
     log.info("session.deleted", {"session_id": session_id})
     try:
         await emit_audit_event(
             "session_action",
             {
-                "action": "delete",
+                "action": "permanent_delete",
                 "actor_id": current_user.username,
                 "actor_name": current_user.username,
                 "user_name": current_user.username,
@@ -1024,6 +1274,7 @@ async def delete_session_for_user(session_id: str, current_user: AuthUser) -> bo
                 "session_id": session_id,
                 "owner_user_id": current_user.id,
                 "project_id": session.project_id,
+                "affected_sessions": len(tree_ids),
             },
         )
     except Exception:
@@ -1031,56 +1282,6 @@ async def delete_session_for_user(session_id: str, current_user: AuthUser) -> bo
     return True
 
 
-async def _collect_descendant_session_ids(project_id: str, session_id: str) -> List[str]:
-    """Return child session IDs in the same order Session.delete will recurse."""
-    sessions = await Session.list(project_id)
-    children_by_parent: Dict[str, List[str]] = {}
-    for child in sessions:
-        if child.parent_id is None:
-            continue
-        children_by_parent.setdefault(child.parent_id, []).append(child.id)
-
-    descendants: List[str] = []
-    seen: set[str] = set()
-
-    def visit(parent_id: str) -> None:
-        for child_id in children_by_parent.get(parent_id, []):
-            if child_id in seen:
-                continue
-            seen.add(child_id)
-            descendants.append(child_id)
-            visit(child_id)
-
-    visit(session_id)
-    return descendants
-
-
-async def _abort_and_wait_descendant_sessions(project_id: str, session_id: str) -> None:
-    """Abort descendants that exist after the parent stops and wait as a batch."""
-    known: set[str] = set()
-    latest: List[str] = []
-
-    for _ in range(_DESCENDANT_ABORT_SCAN_LIMIT):
-        latest = await _collect_descendant_session_ids(project_id, session_id)
-        new_ids = [sid for sid in latest if sid not in known]
-        for descendant_id in new_ids:
-            await _abort_session_processing(descendant_id)
-        known.update(new_ids)
-
-        if latest:
-            await _wait_for_sessions_idle(latest)
-
-        refreshed = await _collect_descendant_session_ids(project_id, session_id)
-        if set(refreshed).issubset(known):
-            return
-        latest = refreshed
-
-    log.warn("session.delete.descendants_unstable", {
-        "session_id": session_id,
-        "descendants": latest,
-    })
-
-
 class SessionUpdateRequest(BaseModel):
     """Request to update session"""
     model_config = ConfigDict(populate_by_name=True)
@@ -1090,6 +1291,15 @@ class SessionUpdateRequest(BaseModel):
     provider: Optional[str] = Field(None, description="Pinned provider ID")
     model: Optional[str] = Field(None, description="Pinned model ID")
     model_pinned: Optional[bool] = Field(None, description="Whether provider/model are pinned for this session")
+    model_auto: Optional[bool] = Field(None, description="Whether WebUI Auto mode is selected")
+
+
+class SessionMoveProjectRequest(BaseModel):
+    """Request to move a complete task tree to a project."""
+
+    model_config = ConfigDict(populate_by_name=True)
+
+    project_id: str = Field(..., alias="projectID", description="Target project ID")
 
 
 @router.patch(
@@ -1114,6 +1324,43 @@ async def update_session(
     current_user = require_user(http_request)
     _require_session_write_access(existing, current_user)
 
+    if request.model_auto is True and (
+        request.provider is not None
+        or request.model is not None
+        or request.model_pinned is True
+    ):
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+            detail="Auto mode cannot be combined with a pinned provider/model",
+        )
+    if request.model_auto is True:
+        if not is_model_auto_session_category(existing.category):
+            raise HTTPException(
+                status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+                detail=(
+                    "Auto mode is only available for user, entity "
+                    "configuration, and workflow sessions"
+                ),
+            )
+        if current_user.id == API_TOKEN_SERVICE_USER_ID:
+            raise HTTPException(
+                status_code=status.HTTP_403_FORBIDDEN,
+                detail="Auto mode can only be enabled from the WebUI",
+            )
+        from flocks.session.session_loop import SessionLoop
+
+        auto_available, auto_reason = await SessionLoop.validate_auto_configuration()
+        if not auto_available:
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail=f"Auto mode is unavailable: {auto_reason}",
+            )
+    if (request.provider is None) != (request.model is None):
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+            detail="provider and model must be updated together",
+        )
+
     updates = {}
     if request.title is not None:
         updates["title"] = request.title
@@ -1125,6 +1372,15 @@ async def update_session(
         updates["model"] = request.model
     if request.model_pinned is not None:
         updates["model_pinned"] = request.model_pinned
+        if request.model_pinned:
+            updates["model_auto"] = False
+    if request.model_auto is not None:
+        updates["model_auto"] = request.model_auto
+        if request.model_auto:
+            updates["model_pinned"] = False
+    if request.provider is not None and request.model is not None:
+        updates["model_auto"] = False
+        updates["model_pinned"] = True
     
     session = await Session.update(
         project_id=existing.project_id,
@@ -1137,11 +1393,119 @@ async def update_session(
             status_code=status.HTTP_404_NOT_FOUND,
             detail=f"Session {sessionID} not found"
         )
+
+    if (
+        request.model_auto is False
+        or request.model_pinned is True
+        or (request.provider is not None and request.model is not None)
+    ):
+        from flocks.session.session_loop import SessionLoop
+
+        SessionLoop.clear_auto_failover_state(sessionID)
     
     log.info("session.updated", {"session_id": sessionID})
     return await _session_to_response_with_goal(session)
 
 
+@router.patch(
+    "/{sessionID}/project",
+    response_model=SessionResponse,
+    summary="Move session to project",
+    description="Move a root session and all descendants to another project",
+)
+async def move_session_to_project(
+    sessionID: str,
+    request: SessionMoveProjectRequest,
+    http_request: Request,
+) -> SessionResponse:
+    """Move an idle root task tree into a writable project."""
+
+    existing = await _get_session_by_id_unfiltered(sessionID)
+    if existing is None:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail=f"Session {sessionID} not found",
+        )
+
+    current_user = require_user(http_request)
+    _require_session_write_access(existing, current_user)
+    if existing.parent_id is not None:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="只能移动根任务及其完整任务树",
+        )
+
+    from flocks.project.project import Project
+
+    target_project = await Project.get(request.project_id, owner_id=current_user.id)
+    if target_project is None:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail=f"Project {request.project_id} not found",
+        )
+    if target_project.path_status != "available":
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="目标项目目录不可用",
+        )
+    if existing.project_id == target_project.id:
+        return await _session_to_response_with_goal(existing)
+
+    tree = await Session.collect_tree(existing.project_id, existing.id)
+    if any(not SessionPolicy.can_write(item, current_user) for item in tree):
+        raise HTTPException(
+            status_code=status.HTTP_403_FORBIDDEN,
+            detail="仅任务所有者可移动完整任务树",
+        )
+
+    if any(_is_session_busy_for_move(item.id) for item in tree):
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="任务正在运行，请等待任务完成或停止后再移动",
+        )
+    if any(item.revert is not None for item in tree):
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="任务正在历史恢复状态，请先完成或取消恢复后再移动",
+        )
+
+    auth_token = set_current_auth_user(current_user)
+    try:
+        moved = await Session.move_to_project(
+            existing.project_id,
+            existing.id,
+            target_project_id=target_project.id,
+            target_directory=target_project.worktree,
+            target_owner_id=current_user.id,
+            additional_busy_check=_is_session_busy_for_move,
+        )
+    finally:
+        reset_current_auth_user(auth_token)
+    if moved is None:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="任务移动失败，请刷新后重试",
+        )
+
+    from flocks.server.routes.event import publish_event
+
+    try:
+        await publish_event("session.updated", {
+            "id": sessionID,
+            "projectID": target_project.id,
+        })
+    except Exception as exc:
+        log.warn("session.move_project.event_error", {
+            "session_id": sessionID,
+            "error": str(exc),
+        })
+
+    return await _session_to_response_with_goal(
+        moved,
+        effective_project_id=target_project.id,
+    )
+
+
 @router.post(
     "/{sessionID}/share-local",
     response_model=SessionResponse,
@@ -1457,8 +1821,11 @@ async def _run_in_background():
                 },
             })
 
-    import asyncio
-    asyncio.create_task(_run_in_background())
+    _schedule_background_coro(
+        _run_in_background(),
+        session_id=sessionID,
+        action="session.summarize",
+    )
     
     log.info("session.summarized", {"session_id": sessionID})
     return True
@@ -1489,11 +1856,17 @@ async def revert_session(sessionID: str, request: RevertRequest, http_request: R
         )
     _require_session_write_access(session, current_user)
 
-    updated = await SessionRevert.revert(
-        session_id=sessionID,
-        message_id=request.messageID,
-        part_id=request.partID,
-    )
+    try:
+        updated = await SessionRevert.revert(
+            session_id=sessionID,
+            message_id=request.messageID,
+            part_id=request.partID,
+        )
+    except (ValueError, SessionInactiveError) as exc:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail=str(exc),
+        ) from exc
     
     log.info("session.reverted", {"session_id": sessionID, "message_id": request.messageID})
     return await _session_to_response_with_goal(updated)
@@ -1584,6 +1957,11 @@ class PromptRequest(BaseModel):
     tools: Optional[Dict[str, bool]] = Field(None, description="Tool settings (deprecated)")
     system: Optional[str] = Field(None, description="System prompt override")
     variant: Optional[str] = Field(None, description="Model variant")
+    execution_mode: SessionExecutionMode = Field(
+        SessionExecutionMode.BUILD,
+        alias="executionMode",
+        description="Execution mode for this user turn",
+    )
 
 
 class UserMessageInfo(BaseModel):
@@ -1612,6 +1990,7 @@ class UserMessageInfo(BaseModel):
     tools: Optional[Dict[str, bool]] = None
     variant: Optional[str] = None
     compacted: Optional[bool] = None
+    executionMode: SessionExecutionMode = SessionExecutionMode.BUILD
 
 
 class AssistantMessageInfo(BaseModel):
@@ -1768,6 +2147,11 @@ async def _message_to_response_info(msg: Any, *, cwd: str) -> MessageInfo:
             agent=getattr(msg, "agent", None) or DEFAULT_AGENT,
             model=model_info,
             compacted=getattr(msg, "compacted", None),
+            executionMode=getattr(
+                msg,
+                "executionMode",
+                SessionExecutionMode.BUILD,
+            ),
         )
 
     tokens_raw = getattr(msg, "tokens", None)
@@ -1914,9 +2298,10 @@ async def get_session_messages(
         return result
     except Exception as e:
         log.error("session.messages.error", {"error": str(e), "sessionID": sessionID})
-        if page or before is not None:
-            return MessagePage(sessionID=sessionID, items=[], hasMore=False, nextBefore=None)
-        return []
+        raise HTTPException(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            detail="Failed to load session messages",
+        ) from e
 
 
 @router.get(
@@ -2087,12 +2472,14 @@ async def _publish_text_part_update(
     })
 
 
-def _track_background_task(task: "asyncio.Task[Any]") -> None:
+def _track_background_task(
+    task: "asyncio.Task[Any]",
+    *,
+    session_id: Optional[str] = None,
+) -> None:
     """Keep background tasks alive until completion."""
-    if not hasattr(router, "_pending_tasks"):
-        router._pending_tasks = set()
-    router._pending_tasks.add(task)
-    task.add_done_callback(lambda t: router._pending_tasks.discard(t))
+    track_background_task(task, session_id=session_id)
+    router._pending_tasks = pending_background_tasks()
 
 
 def _schedule_background_coro(
@@ -2135,13 +2522,13 @@ async def _guarded_coro() -> None:
                     })
 
     task = asyncio.get_running_loop().create_task(_guarded_coro())
-    _track_background_task(task)
+    _track_background_task(task, session_id=session_id)
 
 
 async def _prepare_replay_runtime(
     session_id: str,
     user_message,
-) -> Dict[str, str]:
+) -> Dict[str, Any]:
     """Resolve replay runtime state before mutating session history."""
     from flocks.agent.registry import Agent
     from flocks.config.config import Config
@@ -2149,21 +2536,40 @@ async def _prepare_replay_runtime(
 
     agent_name = getattr(user_message, "agent", None) or await Agent.default_agent()
     agent = await Agent.get(agent_name) or await Agent.get(DEFAULT_AGENT)
-    # Replay should follow the model that is active *now* for this session
-    # (current session pin / current default / current agent override), not the
-    # historical model stored on the original user message being replayed.
-    dummy_request = type(
-        "_MessageReplayRequest",
-        (),
-        {"model": None, "agent": agent_name},
-    )()
-    provider_id, model_id, _ = await _resolve_model(dummy_request, agent, session_id)
+    session = await Session.get_by_id(session_id)
+    auto_failover = bool(
+        session
+        and is_model_auto_session_category(getattr(session, "category", "user"))
+        and getattr(session, "model_auto", False)
+    )
+    if auto_failover:
+        default_llm = await Config.resolve_default_llm()
+        if not default_llm:
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail="Auto mode requires a configured default LLM",
+            )
+        provider_id = default_llm["provider_id"]
+        model_id = default_llm["model_id"]
+    else:
+        # Replay follows the model active *now*, not the model stored on the
+        # historical user message.
+        dummy_request = type(
+            "_MessageReplayRequest",
+            (),
+            {"model": None, "agent": agent_name},
+        )()
+        provider_id, model_id, _ = await _resolve_model(
+            dummy_request,
+            agent,
+            session_id,
+        )
 
     Provider._ensure_initialized()
     config = await Config.get()
     await Provider.apply_config(config, provider_id=provider_id)
     provider = Provider.get(provider_id)
-    if not provider:
+    if not provider and not auto_failover:
         raise HTTPException(
             status_code=status.HTTP_400_BAD_REQUEST,
             detail=f"Provider {provider_id} not found",
@@ -2173,6 +2579,7 @@ async def _prepare_replay_runtime(
         "agent_name": agent_name,
         "provider_id": provider_id,
         "model_id": model_id,
+        "auto_failover": auto_failover,
     }
 
 
@@ -2181,7 +2588,7 @@ async def _run_existing_user_message(
     session,
     user_message,
     working_directory: str,
-    runtime: Optional[Dict[str, str]] = None,
+    runtime: Optional[Dict[str, Any]] = None,
 ):
     """Run SessionLoop using an already-persisted user message."""
     from flocks.server.routes.event import publish_event
@@ -2214,6 +2621,7 @@ async def _on_error(error: str):
         agent_name=agent_name,
         callbacks=loop_callbacks,
         working_directory=working_directory,
+        auto_failover=bool(runtime.get("auto_failover", False)),
     )
 
     if result.action == "queued":
@@ -2233,6 +2641,8 @@ async def _on_error(error: str):
     assistant_message_id = None
     created_ms = end_ms
     final_tokens = {"input": 0, "output": 0, "reasoning": 0, "cache": {"read": 0, "write": 0}}
+    actual_provider_id = result.provider_id or provider_id
+    actual_model_id = result.model_id or model_id
 
     if result.last_message:
         assistant_message_id = result.last_message.id
@@ -2241,6 +2651,14 @@ async def _on_error(error: str):
         finish = getattr(result.last_message, "finish", None)
         if finish:
             finish_reason = finish
+        actual_provider_id = (
+            getattr(result.last_message, "providerID", None)
+            or actual_provider_id
+        )
+        actual_model_id = (
+            getattr(result.last_message, "modelID", None)
+            or actual_model_id
+        )
         result_time = getattr(result.last_message, "time", None)
         if isinstance(result_time, dict):
             created_ms = result_time.get("created", created_ms)
@@ -2260,8 +2678,8 @@ async def _on_error(error: str):
             "role": "assistant",
             "time": {"created": created_ms, "completed": end_ms},
             "parentID": user_message.id,
-            "modelID": model_id,
-            "providerID": provider_id,
+            "modelID": actual_model_id,
+            "providerID": actual_provider_id,
             "mode": agent_name,
             "agent": agent_name,
             "path": {"cwd": working_directory, "root": working_directory},
@@ -2274,8 +2692,8 @@ async def _on_error(error: str):
         publish_event,
         session_id,
         session=session,
-        provider_id=provider_id,
-        model_id=model_id,
+        provider_id=actual_provider_id,
+        model_id=actual_model_id,
     )
 
     log.info("session.message.replay.completed", {
@@ -2343,6 +2761,14 @@ async def resend_session_message(
             detail="Session is currently generating a response",
         )
 
+    try:
+        await SessionRevert.ensure_replayable(session, messageID)
+    except ValueError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail=str(exc),
+        ) from exc
+
     working_directory = await _resolve_session_working_directory(session)
 
     async def _handle_resend() -> None:
@@ -2438,6 +2864,14 @@ async def regenerate_session_message(
             detail="Session is currently generating a response",
         )
 
+    try:
+        await SessionRevert.ensure_replayable(session, parent_message_id)
+    except ValueError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail=str(exc),
+        ) from exc
+
     working_directory = await _resolve_session_working_directory(session)
 
     async def _handle_regenerate() -> None:
@@ -2502,8 +2936,28 @@ async def send_session_message(sessionID: str, request: PromptRequest, http_requ
         )
     current_user = require_user(http_request)
     _require_session_write_access(session, current_user)
+    lifecycle_generation = Session.lifecycle_generation(sessionID)
     
     working_directory = await _resolve_session_working_directory(session)
+    _validate_execution_mode_request(request)
+    if request.execution_mode == SessionExecutionMode.GOAL:
+        from flocks.session.goal import GoalManager
+
+        objective = _extract_text_from_parts(request.parts).strip()
+        state = await GoalManager.set_goal(sessionID, objective)
+        await publish_event("session.goal.updated", {
+            "sessionID": sessionID,
+            "status": state.status,
+            "objective": state.objective,
+            "reason": state.last_reason,
+        })
+        request = request.model_copy(update={
+            "parts": _replace_text_parts(
+                request.parts,
+                GoalManager.goal_prompt(state.objective),
+            ),
+            "display_text": request.display_text or objective,
+        })
     
     log.info("session.message.send.processing", {
         "sessionID": sessionID,
@@ -2517,7 +2971,13 @@ async def send_session_message(sessionID: str, request: PromptRequest, http_requ
         result = await Instance.provide(
             directory=working_directory,
             init=instance_bootstrap,
-            fn=lambda: _process_session_message(sessionID, session, request, working_directory)
+            fn=lambda: _process_session_message(
+                sessionID,
+                session,
+                request,
+                working_directory,
+                lifecycle_generation=lifecycle_generation,
+            )
         )
         log.info("session.message.send.complete", {"sessionID": sessionID})
         return result
@@ -2818,6 +3278,8 @@ async def _process_session_message(
     session,
     request: PromptRequest,
     working_directory: str,
+    *,
+    lifecycle_generation: Optional[int] = None,
 ):
     """
     Process session message within Instance context.
@@ -2837,6 +3299,9 @@ async def _process_session_message(
     from flocks.session.runner import RunnerCallbacks
     import time
     import os
+
+    if lifecycle_generation is None:
+        lifecycle_generation = Session.lifecycle_generation(sessionID)
     
     # Clean up revert state before processing (Flocks compatibility)
     await SessionRevert.cleanup(session)
@@ -2873,9 +3338,27 @@ async def _process_session_message(
     agent_name = request.agent or await Agent.default_agent()
     agent = await Agent.get(agent_name) or await Agent.get(DEFAULT_AGENT)
     
-    provider_id, model_id, model_source = await _resolve_model(
-        request, agent, sessionID
+    auto_failover = bool(
+        is_model_auto_session_category(getattr(session, "category", "user"))
+        and getattr(session, "model_auto", False)
+        and not request.model
     )
+    if auto_failover:
+        from flocks.config.config import Config
+
+        default_llm = await Config.resolve_default_llm()
+        if not default_llm:
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail="Auto mode requires a configured default LLM",
+            )
+        provider_id = default_llm["provider_id"]
+        model_id = default_llm["model_id"]
+        model_source = "auto_primary"
+    else:
+        provider_id, model_id, model_source = await _resolve_model(
+            request, agent, sessionID
+        )
     
     log.info("session.message.model", {
         "provider_id": provider_id,
@@ -2895,15 +3378,17 @@ async def _process_session_message(
             session.provider = provider_id
             session.model = model_id
             session.model_pinned = True
+            session.model_auto = False
+        SessionLoop.clear_auto_failover_state(sessionID)
     
     # Ensure providers are initialized and configured
     Provider._ensure_initialized()
     from flocks.config.config import Config
     config = await Config.get()
     await Provider.apply_config(config, provider_id=provider_id)
-    
+
     provider = Provider.get(provider_id)
-    if not provider:
+    if not provider and not auto_failover:
         raise HTTPException(
             status_code=status.HTTP_400_BAD_REQUEST,
             detail=f"Provider {provider_id} not found"
@@ -2925,17 +3410,22 @@ async def _process_session_message(
     display_metadata = {"displayText": display_text} if display_text else None
 
     _is_no_reply = bool(request.noReply)
-    user_message = await Message.create(
-        session_id=sessionID,
-        role=MessageRole.USER,
-        content=text_content,
-        id=user_message_id,
-        time={"created": now_ms},
-        agent=agent_name,
-        model={"providerID": provider_id, "modelID": model_id},
-        part_id=user_part_id,
-        part_metadata=display_metadata,
-        synthetic=True if _is_no_reply else None,
+    user_message = await _persist_active_session_write(
+        sessionID,
+        lambda: Message.create(
+            session_id=sessionID,
+            role=MessageRole.USER,
+            content=text_content,
+            id=user_message_id,
+            time={"created": now_ms},
+            agent=agent_name,
+            model={"providerID": provider_id, "modelID": model_id},
+            executionMode=request.execution_mode,
+            part_id=user_part_id,
+            part_metadata=display_metadata,
+            synthetic=True if _is_no_reply else None,
+        ),
+        expected_generation=lifecycle_generation,
     )
     user_message_id = user_message.id
     
@@ -2947,6 +3437,7 @@ async def _process_session_message(
             "time": {"created": now_ms},
             "agent": agent_name,
             "model": {"providerID": provider_id, "modelID": model_id},
+            "executionMode": request.execution_mode.value,
         }
     })
     _part_event: dict = {
@@ -3008,7 +3499,11 @@ async def _process_session_message(
                 filename=raw_part.get("filename"),
                 url=url,
             )
-            await Message.add_part(sessionID, user_message_id, file_part)
+            await _persist_active_session_write(
+                sessionID,
+                lambda: Message.add_part(sessionID, user_message_id, file_part),
+                expected_generation=lifecycle_generation,
+            )
             await publish_event("message.part.updated", {
                 "part": {
                     "id": file_part_id,
@@ -3033,15 +3528,19 @@ async def _process_session_message(
             mock_msg_id = Identifier.ascending("message")
             mock_part_id = Identifier.ascending("part")
             mock_now = int(time.time() * 1000)
-            await Message.create(
-                session_id=sessionID,
-                role=MessageRole.ASSISTANT,
-                content=request.mockReply,
-                id=mock_msg_id,
-                time={"created": mock_now, "completed": mock_now},
-                parentID=user_message_id,
-                modelID="mock",
-                part_id=mock_part_id,
+            await _persist_active_session_write(
+                sessionID,
+                lambda: Message.create(
+                    session_id=sessionID,
+                    role=MessageRole.ASSISTANT,
+                    content=request.mockReply,
+                    id=mock_msg_id,
+                    time={"created": mock_now, "completed": mock_now},
+                    parentID=user_message_id,
+                    modelID="mock",
+                    part_id=mock_part_id,
+                ),
+                expected_generation=lifecycle_generation,
             )
             await publish_event("message.updated", {
                 "info": {
@@ -3102,6 +3601,7 @@ async def _on_error(error: str):
         agent_name=agent_name,
         callbacks=loop_callbacks,
         working_directory=working_directory,
+        auto_failover=auto_failover,
     )
 
     # ------------------------------------------------------------------
@@ -3131,6 +3631,8 @@ async def _on_error(error: str):
     final_content = ""
     assistant_message_id = None
     final_tokens = {"input": 0, "output": 0, "reasoning": 0, "cache": {"read": 0, "write": 0}}
+    actual_provider_id = result.provider_id or provider_id
+    actual_model_id = result.model_id or model_id
     
     if result.last_message:
         assistant_message_id = result.last_message.id
@@ -3139,6 +3641,14 @@ async def _on_error(error: str):
         finish = getattr(result.last_message, 'finish', None)
         if finish:
             finish_reason = finish
+        actual_provider_id = (
+            getattr(result.last_message, "providerID", None)
+            or actual_provider_id
+        )
+        actual_model_id = (
+            getattr(result.last_message, "modelID", None)
+            or actual_model_id
+        )
     
     if result.action == "error":
         finish_reason = "error"
@@ -3156,8 +3666,8 @@ async def _on_error(error: str):
             "role": "assistant",
             "time": {"created": now_ms, "completed": end_ms},
             "parentID": user_message_id,
-            "modelID": model_id,
-            "providerID": provider_id,
+            "modelID": actual_model_id,
+            "providerID": actual_provider_id,
             "mode": agent_name,
             "agent": agent_name,
             "path": {"cwd": working_directory, "root": working_directory},
@@ -3170,8 +3680,8 @@ async def _on_error(error: str):
         publish_event,
         sessionID,
         session=session,
-        provider_id=provider_id,
-        model_id=model_id,
+        provider_id=actual_provider_id,
+        model_id=actual_model_id,
     )
     
     # Collect parts for the response
@@ -3199,8 +3709,8 @@ async def _on_error(error: str):
         title_task = loop.create_task(
             SessionTitle.generate_title_after_first_message(
                 session_id=sessionID,
-                model_id=model_id,
-                provider_id=provider_id,
+                model_id=actual_model_id,
+                provider_id=actual_provider_id,
                 event_publish_callback=publish_event,
             )
         )
@@ -3218,8 +3728,8 @@ async def _on_error(error: str):
             "role": "assistant",
             "time": {"created": now_ms, "completed": end_ms},
             "parentID": user_message_id,
-            "modelID": model_id,
-            "providerID": provider_id,
+            "modelID": actual_model_id,
+            "providerID": actual_provider_id,
             "mode": agent_name,
             "agent": agent_name,
             "path": {"cwd": working_directory, "root": working_directory},
@@ -3353,6 +3863,32 @@ def _extract_text_from_parts(parts: List[Dict[str, Any]]) -> str:
     return "".join(part.get("text", "") for part in parts if part.get("type") == "text")
 
 
+def _validate_execution_mode_request(request: PromptRequest) -> None:
+    if request.execution_mode != SessionExecutionMode.GOAL:
+        return
+    objective = _extract_text_from_parts(request.parts).strip()
+    if not objective:
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail="Goal mode requires a non-empty text objective",
+        )
+    if any(part.get("type") != "text" for part in request.parts):
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail="Goal mode does not support attachments",
+        )
+
+
+def _event_text_for_execution_mode(
+    parts: List[Dict[str, Any]],
+    execution_mode: SessionExecutionMode,
+) -> str:
+    text = _extract_text_from_parts(parts)
+    if execution_mode == SessionExecutionMode.GOAL:
+        return f"/goal {text.strip()}"
+    return text
+
+
 def _replace_text_parts(
     parts: Optional[List[Dict[str, Any]]],
     text: str,
@@ -3417,6 +3953,19 @@ def _is_prompt_chain_active(session_id: str) -> bool:
     return session_id in getattr(router, "_prompt_queue_active_sessions", set())
 
 
+def _is_session_busy_for_move(session_id: str) -> bool:
+    """Cover loop execution and route-owned background work during a move."""
+
+    from flocks.session.session_loop import SessionLoop
+
+    return (
+        Session.has_active_operations(session_id)
+        or SessionLoop.is_running(session_id)
+        or _is_prompt_chain_active(session_id)
+        or has_pending_session_tasks(session_id)
+    )
+
+
 def _set_prompt_chain_active(session_id: str, active: bool) -> None:
     if not hasattr(router, "_prompt_queue_active_sessions"):
         router._prompt_queue_active_sessions = set()
@@ -3458,17 +4007,21 @@ def _event_from_queued_prompt(item, working_directory: str):
     return UserInputEvent(
         source_type="webui",
         sessionID=item.sessionID,
-        text=_extract_text_from_parts(item.parts),
+        text=_event_text_for_execution_mode(item.parts, item.executionMode),
         parts=[dict(part) for part in item.parts],
         agent=item.agent,
         model=item.model,
         variant=item.variant,
+        metadata={
+            "_sessionLifecycleGeneration": Session.lifecycle_generation(item.sessionID),
+        },
         display_text=item.display_text,
         messageID=item.messageID,
         noReply=item.noReply,
         mockReply=item.mockReply,
         tools=item.tools,
         system=item.system,
+        executionMode=item.executionMode,
         working_directory=working_directory,
     )
 
@@ -3605,19 +4158,23 @@ def _build_prompt_request_from_event(event, prompt_text: str, display_text: Opti
         noReply=event.no_reply,
         tools=event.tools,
         system=event.system,
+        execution_mode=event.execution_mode,
     )
 
 
 async def _dispatch_sse_input(sessionID: str, session, event, working_directory: str) -> None:
     import time as _time
 
-    from flocks.input.dispatcher import dispatch_user_input
+    from flocks.input.dispatcher import dispatch_user_input, parse_slash_command
     from flocks.input.output import SSEOutputSink
     from flocks.server.routes.event import publish_event
     from flocks.session.message import Message, MessageRole
     from flocks.utils.id import Identifier
 
     agent_name = event.agent or "rex"
+    lifecycle_generation = event.metadata.get("_sessionLifecycleGeneration")
+    if not isinstance(lifecycle_generation, int):
+        lifecycle_generation = Session.lifecycle_generation(sessionID)
 
     async def _create_user_message(
         user_text: str,
@@ -3629,15 +4186,20 @@ async def _create_user_message(
         user_msg_id = event.message_id or Identifier.create("message")
         user_part_id = Identifier.create("part")
         message_agent = agent_override or agent_name
-        await Message.create(
-            session_id=sessionID,
-            role=MessageRole.USER,
-            content=user_text,
-            id=user_msg_id,
-            time={"created": now_ms},
-            agent=message_agent,
-            **({"model": model_info} if model_info else {}),
-            part_id=user_part_id,
+        await _persist_active_session_write(
+            sessionID,
+            lambda: Message.create(
+                session_id=sessionID,
+                role=MessageRole.USER,
+                content=user_text,
+                id=user_msg_id,
+                time={"created": now_ms},
+                agent=message_agent,
+                executionMode=event.execution_mode,
+                **({"model": model_info} if model_info else {}),
+                part_id=user_part_id,
+            ),
+            expected_generation=lifecycle_generation,
         )
         await publish_event("message.updated", {
             "info": {
@@ -3646,6 +4208,7 @@ async def _create_user_message(
                 "role": "user",
                 "time": {"created": now_ms},
                 "agent": message_agent,
+                "executionMode": event.execution_mode.value,
                 **({"model": model_info} if model_info else {}),
             }
         })
@@ -3667,18 +4230,22 @@ async def _publish_direct_response(output_event, text: str) -> None:
         asst_now = int(_time.time() * 1000)
         asst_msg_id = Identifier.ascending("message")
         asst_part_id = Identifier.ascending("part")
-        await Message.create(
-            session_id=sessionID,
-            role=MessageRole.ASSISTANT,
-            content=text,
-            id=asst_msg_id,
-            time={"created": asst_now, "completed": asst_now},
-            parentID=parent_msg_id,
-            modelID="command",
-            providerID="builtin",
-            agent=agent_name,
-            finish="stop",
-            part_id=asst_part_id,
+        await _persist_active_session_write(
+            sessionID,
+            lambda: Message.create(
+                session_id=sessionID,
+                role=MessageRole.ASSISTANT,
+                content=text,
+                id=asst_msg_id,
+                time={"created": asst_now, "completed": asst_now},
+                parentID=parent_msg_id,
+                modelID="command",
+                providerID="builtin",
+                agent=agent_name,
+                finish="stop",
+                part_id=asst_part_id,
+            ),
+            expected_generation=lifecycle_generation,
         )
         await publish_event("message.updated", {
             "info": {
@@ -3719,8 +4286,26 @@ async def _publish_direct_response(output_event, text: str) -> None:
         )
 
     async def _run_llm(output_event, prompt_text: str, display_text: Optional[str] = None) -> None:
+        parsed = parse_slash_command(output_event.text, output_event.metadata)
+        if parsed is not None and parsed.canonical_name == "goal":
+            from flocks.session.goal import GoalManager
+
+            state = await GoalManager.get(sessionID)
+            if state is not None:
+                await publish_event("session.goal.updated", {
+                    "sessionID": sessionID,
+                    "status": state.status,
+                    "objective": state.objective,
+                    "reason": state.last_reason,
+                })
         request = _build_prompt_request_from_event(output_event, prompt_text, display_text)
-        await _process_session_message(sessionID, session, request, working_directory)
+        await _process_session_message(
+            sessionID,
+            session,
+            request,
+            working_directory,
+            lifecycle_generation=lifecycle_generation,
+        )
 
     async def _clear_history() -> None:
         await _clear_session_history(sessionID)
@@ -3763,18 +4348,7 @@ async def _run_session_control(output_event, parsed) -> bool:
         session_control=_run_session_control,
         clear_history=_clear_history,
     )
-    result = await dispatch_user_input(event, sink)
-    if result.command_name == "goal" and result.action == "llm":
-        from flocks.session.goal import GoalManager
-
-        state = await GoalManager.get(sessionID)
-        if state is not None:
-            await publish_event("session.goal.updated", {
-                "sessionID": sessionID,
-                "status": state.status,
-                "objective": state.objective,
-                "reason": state.last_reason,
-            })
+    await dispatch_user_input(event, sink)
 
 
 class PromptQueueUpdateRequest(BaseModel):
@@ -3784,24 +4358,34 @@ class PromptQueueUpdateRequest(BaseModel):
 async def _enqueue_prompt_request(
     session_id: str,
     request: PromptRequest,
+    *,
+    expected_generation: Optional[int] = None,
 ):
     from flocks.session.interaction_queue import InteractionQueue
 
+    _validate_execution_mode_request(request)
+    if expected_generation is None:
+        expected_generation = Session.lifecycle_generation(session_id)
     await _require_agent_usable_for_chat(request.agent)
     model = request.model.model_dump(by_alias=True) if request.model else None
     parts = _materialize_queued_parts(session_id, [dict(part) for part in request.parts])
-    return await InteractionQueue.enqueue(
+    return await _persist_active_session_write(
         session_id,
-        parts=parts,
-        agent=request.agent,
-        model=model,
-        variant=request.variant,
-        display_text=request.display_text,
-        message_id=request.messageID,
-        no_reply=request.noReply,
-        mock_reply=request.mockReply,
-        tools=request.tools,
-        system=request.system,
+        lambda: InteractionQueue.enqueue(
+            session_id,
+            parts=parts,
+            agent=request.agent,
+            model=model,
+            variant=request.variant,
+            display_text=request.display_text,
+            message_id=request.messageID,
+            no_reply=request.noReply,
+            mock_reply=request.mockReply,
+            tools=request.tools,
+            system=request.system,
+            execution_mode=request.execution_mode,
+        ),
+        expected_generation=expected_generation,
     )
 
 
@@ -3948,8 +4532,10 @@ async def send_session_message_async(
     if http_request is not None:
         current_user = require_user(http_request)
         _require_session_write_access(session, current_user)
+    lifecycle_generation = Session.lifecycle_generation(sessionID)
     
     working_directory = await _resolve_session_working_directory(session)
+    _validate_execution_mode_request(request)
     await _require_agent_usable_for_chat(request.agent)
     
     log.info("session.prompt_async.accepted", {
@@ -3957,27 +4543,44 @@ async def send_session_message_async(
         "directory": working_directory,
     })
 
+    event_text = _event_text_for_execution_mode(
+        request.parts,
+        request.execution_mode,
+    )
+    event_display_text = request.display_text
+    if request.execution_mode == SessionExecutionMode.GOAL:
+        event_display_text = (
+            event_display_text
+            or _extract_text_from_parts(request.parts).strip()
+        )
+
     event = UserInputEvent(
         source_type="webui",
         sessionID=sessionID,
-        text=_extract_text_from_parts(request.parts),
+        text=event_text,
         parts=[dict(part) for part in request.parts],
         agent=request.agent,
         model=request.model.model_dump(by_alias=True) if request.model else None,
         variant=request.variant,
-        display_text=request.display_text,
+        metadata={"_sessionLifecycleGeneration": lifecycle_generation},
+        display_text=event_display_text,
         messageID=request.messageID,
         noReply=request.noReply,
         mockReply=request.mockReply,
         tools=request.tools,
         system=request.system,
+        executionMode=request.execution_mode,
         working_directory=working_directory,
     )
 
     existing_queue = await InteractionQueue.list(sessionID)
     if SessionLoop.is_running(sessionID) or existing_queue or _is_prompt_chain_active(sessionID):
         try:
-            item = await _enqueue_prompt_request(sessionID, request)
+            item = await _enqueue_prompt_request(
+                sessionID,
+                request,
+                expected_generation=lifecycle_generation,
+            )
         except QueueFullError as exc:
             raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail=str(exc)) from exc
         await _publish_prompt_queue(sessionID)
@@ -4045,6 +4648,7 @@ async def send_session_command(sessionID: str, request: CommandRequest, http_req
     if http_request is not None:
         current_user = require_user(http_request)
         _require_session_write_access(session, current_user)
+    lifecycle_generation = Session.lifecycle_generation(sessionID)
 
     working_directory = await _resolve_session_working_directory(session)
     await _require_agent_usable_for_chat(request.agent)
@@ -4052,6 +4656,7 @@ async def send_session_command(sessionID: str, request: CommandRequest, http_req
     if not raw_arguments and request.arguments_json is not None:
         raw_arguments = json.dumps(request.arguments_json, ensure_ascii=False)
     command_metadata: Dict[str, Any] = {}
+    command_metadata["_sessionLifecycleGeneration"] = lifecycle_generation
     if request.arguments_json is not None:
         command_metadata["commandArgumentsJson"] = request.arguments_json
 
@@ -4107,10 +4712,7 @@ async def _handle_command() -> None:
 
     loop = asyncio.get_running_loop()
     task = loop.create_task(_handle_command())
-    if not hasattr(router, "_pending_tasks"):
-        router._pending_tasks = set()
-    router._pending_tasks.add(task)
-    task.add_done_callback(lambda t: router._pending_tasks.discard(t))
+    _track_background_task(task, session_id=sessionID)
 
     log.info("session.command.accepted", {
         "sessionID": sessionID,
@@ -4149,12 +4751,24 @@ async def run_shell_command(sessionID: str, request: ShellRequest, http_request:
     if request.model:
         model = {"providerID": request.model.providerID, "modelID": request.model.modelID}
     
-    result = await SessionRunner.shell(
-        session_id=sessionID,
-        agent=request.agent,
-        command=request.command,
-        model=model,
-    )
+    try:
+        async with Session.active_operation(sessionID):
+            result = await SessionRunner.shell(
+                session_id=sessionID,
+                agent=request.agent,
+                command=request.command,
+                model=model,
+            )
+    except SessionNotFoundError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail=str(exc),
+        ) from exc
+    except SessionInactiveError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail=str(exc),
+        ) from exc
     
     log.info("session.shell.executed", {
         "sessionID": sessionID,
@@ -4369,7 +4983,11 @@ async def get_session_statistics(sessionID: str):
         raise HTTPException(status_code=500, detail=f"Failed to get session statistics: {str(e)}")
 
 
-async def _clear_session_history(sessionID: str) -> int:
+async def _clear_session_history(
+    sessionID: str,
+    *,
+    expected_generation: Optional[int] = None,
+) -> int:
     """Clear stored messages for a session and notify subscribed UIs."""
     session_info = await _get_session_by_id_unfiltered(sessionID)
     if not session_info:
@@ -4385,14 +5003,23 @@ async def _clear_session_history(sessionID: str) -> int:
 
     await abort_session(sessionID)
     await InteractionQueue.clear(sessionID)
-    await GoalManager.clear(sessionID)
     try:
         await _publish_prompt_queue(sessionID)
     except Exception as exc:
         log.warn("session.clear.prompt_queue_event_error", {"sessionID": sessionID, "error": str(exc)})
     await _wait_for_session_idle(sessionID)
 
-    deleted_count = await Message.clear(sessionID)
+    async def clear_persisted_history() -> int:
+        await GoalManager.clear(sessionID)
+        deleted_count = await Message.clear(sessionID)
+        await Session._clear_project_move_metadata_locked(sessionID)
+        return deleted_count
+
+    deleted_count = await _persist_active_session_write(
+        sessionID,
+        clear_persisted_history,
+        expected_generation=expected_generation,
+    )
     log.info("session.cleared", {"sessionID": sessionID, "deleted": deleted_count})
 
     try:
@@ -4423,7 +5050,10 @@ async def clear_session(sessionID: str, http_request: Request):
         current_user = require_user(http_request)
         _require_session_write_access(session_info, current_user)
 
-        deleted_count = await _clear_session_history(sessionID)
+        deleted_count = await _clear_session_history(
+            sessionID,
+            expected_generation=Session.lifecycle_generation(sessionID),
+        )
         return {
             "status": "success",
             "sessionID": sessionID,
diff --git a/flocks/session/background_tasks.py b/flocks/session/background_tasks.py
new file mode 100644
index 000000000..3608580a2
--- /dev/null
+++ b/flocks/session/background_tasks.py
@@ -0,0 +1,60 @@
+"""Track cancellable asynchronous work associated with sessions."""
+
+import asyncio
+from typing import Any, Optional
+
+
+_pending_tasks: set[asyncio.Task[Any]] = set()
+_tasks_by_session: dict[str, set[asyncio.Task[Any]]] = {}
+
+
+def track_background_task(
+    task: asyncio.Task[Any],
+    *,
+    session_id: Optional[str] = None,
+) -> None:
+    """Keep a task alive and optionally associate it with a session."""
+
+    _pending_tasks.add(task)
+    if session_id:
+        _tasks_by_session.setdefault(session_id, set()).add(task)
+
+    def discard(completed: asyncio.Task[Any]) -> None:
+        _pending_tasks.discard(completed)
+        if not session_id:
+            return
+        session_tasks = _tasks_by_session.get(session_id)
+        if session_tasks is None:
+            return
+        session_tasks.discard(completed)
+        if not session_tasks:
+            _tasks_by_session.pop(session_id, None)
+
+    task.add_done_callback(discard)
+
+
+async def cancel_session_background_tasks(session_id: str) -> None:
+    """Cancel and join tracked work before a lifecycle transition commits."""
+
+    current_task = asyncio.current_task()
+    tasks = [
+        task
+        for task in _tasks_by_session.get(session_id, ())
+        if task is not current_task and not task.done()
+    ]
+    for task in tasks:
+        task.cancel()
+    if tasks:
+        await asyncio.gather(*tasks, return_exceptions=True)
+
+
+def pending_background_tasks() -> set[asyncio.Task[Any]]:
+    """Return the live task set for compatibility with route-level drains."""
+
+    return _pending_tasks
+
+
+def has_pending_session_tasks(session_id: str) -> bool:
+    """Return whether tracked asynchronous work is still active for a session."""
+
+    return any(not task.done() for task in _tasks_by_session.get(session_id, ()))
diff --git a/flocks/session/callable_state.py b/flocks/session/callable_state.py
index d4e0097e9..e3dbd2a0f 100644
--- a/flocks/session/callable_state.py
+++ b/flocks/session/callable_state.py
@@ -73,5 +73,10 @@ async def clear_session_callable_tools(session_id: str) -> None:
     await Storage.delete(f"{_CALLABLE_PREFIX}{session_id}")
 
 
+def invalidate_session_callable_tools_cache(session_id: str) -> None:
+    """Drop process-local callable-tool state after an external transaction."""
+    _cache.pop(session_id, None)
+
+
 async def session_can_call_tool(session_id: str, tool_name: str) -> bool:
     return tool_name in await get_session_callable_tools(session_id)
diff --git a/flocks/session/execution_mode.py b/flocks/session/execution_mode.py
new file mode 100644
index 000000000..fbad85d6f
--- /dev/null
+++ b/flocks/session/execution_mode.py
@@ -0,0 +1,196 @@
+"""Session execution-mode policy derived from OpenCode and Codex."""
+
+from __future__ import annotations
+
+from enum import Enum
+from typing import Any, Iterable, Optional
+
+from flocks.session.plan_file import (
+    SessionPlanFile,
+    is_current_plan_path,
+    plan_edit_patterns_allowed,
+    plan_file_prompt,
+)
+
+
+class SessionExecutionMode(str, Enum):
+    """Execution mode selected for a user turn."""
+
+    BUILD = "build"
+    PLAN = "plan"
+    GOAL = "goal"
+
+
+PLAN_ONLY_TOOL_NAMES = frozenset({"plan_exit"})
+PLAN_DENIED_TOOL_NAMES = frozenset(
+    {
+        # Explicit slash commands keep their existing direct user-only path.
+        "run_slash_command",
+    }
+)
+PLAN_DELEGATION_TOOL_NAMES = frozenset({"delegate_task", "task"})
+PLAN_DELEGATABLE_AGENT_NAMES = frozenset({"explore", "librarian"})
+PLAN_PATH_SCOPED_TOOL_NAMES = frozenset({"apply_patch", "edit", "write"})
+
+PLAN_MODE_PROMPT = """# Plan Mode
+
+You are in a planning turn. You may inspect files, configuration, types,
+tests, and documentation. Bash is available only for read-only exploration
+and validation. Do not use shell commands to modify files, configuration,
+services, dependencies, version control state, or any other system state.
+
+The only file you may modify is the session plan file named below. The runtime
+enforces this boundary for file-editing tools.
+
+Follow this workflow:
+
+1. Explore first. Ground the plan in the existing environment and resolve
+   discoverable facts through inspection before asking the user.
+   Delegation is limited to the `explore` and `librarian` subagents.
+2. Use the question tool only for material ambiguities, preferences, or
+   trade-offs that cannot be resolved from the environment. After the user
+   answers, continue exploring and planning as needed.
+3. Review the proposed approach for remaining gaps. Ask another focused
+   question if a decision is still required.
+4. Write the decision-complete implementation plan to the session plan file.
+   The plan must be detailed enough for another engineer to execute without
+   making additional design decisions.
+5. Present the final plan to the user, then immediately call plan_exit. That
+   tool asks the user whether to start implementation. If approved, it switches
+   the next turn to Build and starts implementing the approved plan. If
+   declined, remain in Plan and use the feedback to refine it.
+
+Do not ask for implementation approval with ordinary prose or the question
+tool; plan_exit owns that transition. A Plan turn may end only by asking a
+material clarification question or by calling plan_exit after the final plan.
+"""
+
+
+def coerce_execution_mode(value: object) -> SessionExecutionMode:
+    """Return a valid execution mode, defaulting legacy values to Build."""
+
+    if isinstance(value, SessionExecutionMode):
+        return value
+    try:
+        return SessionExecutionMode(str(value or SessionExecutionMode.BUILD.value))
+    except ValueError:
+        return SessionExecutionMode.BUILD
+
+
+def runtime_execution_mode(value: object) -> SessionExecutionMode:
+    """Resolve the permission mode used while executing a turn."""
+
+    mode = coerce_execution_mode(value)
+    if mode == SessionExecutionMode.GOAL:
+        return SessionExecutionMode.BUILD
+    return mode
+
+
+def is_tool_allowed(value: object, tool_name: str) -> bool:
+    """Evaluate tool visibility against OpenCode-style Plan permissions."""
+
+    mode = runtime_execution_mode(value)
+    if tool_name in PLAN_ONLY_TOOL_NAMES:
+        return mode == SessionExecutionMode.PLAN
+    if mode == SessionExecutionMode.BUILD:
+        return True
+    return tool_name not in PLAN_DENIED_TOOL_NAMES
+
+
+def tool_call_denial_reason(
+    value: object,
+    tool_name: str,
+    arguments: dict[str, Any],
+    ctx: Any,
+) -> Optional[str]:
+    """Return a hard Plan-mode denial reason for a concrete tool call."""
+
+    if runtime_execution_mode(value) != SessionExecutionMode.PLAN:
+        return None
+    if tool_name in PLAN_DELEGATION_TOOL_NAMES:
+        subagent_type = str(arguments.get("subagent_type") or "").strip().lower()
+        if (
+            subagent_type in PLAN_DELEGATABLE_AGENT_NAMES
+            and not arguments.get("category")
+            and not arguments.get("session_id")
+        ):
+            return None
+        allowed = ", ".join(sorted(PLAN_DELEGATABLE_AGENT_NAMES))
+        return (
+            f"Tool {tool_name!r} may only delegate to {allowed} via "
+            "subagent_type while Plan mode is active."
+        )
+    if tool_name not in PLAN_PATH_SCOPED_TOOL_NAMES:
+        return None
+
+    if tool_name in {"edit", "write"}:
+        paths = [arguments.get("filePath")]
+    else:
+        try:
+            from flocks.tool.file.apply_patch import parse_patch
+
+            hunks = parse_patch(str(arguments.get("patchText") or ""))
+        except Exception:
+            hunks = []
+        paths = [
+            path
+            for hunk in hunks
+            for path in (getattr(hunk, "path", None), getattr(hunk, "move_path", None))
+            if path
+        ]
+
+    if paths and all(is_current_plan_path(ctx, path) for path in paths):
+        return None
+    return (
+        f"Tool {tool_name!r} may only edit the current session plan file "
+        "while Plan mode is active."
+    )
+
+
+def is_permission_allowed(
+    value: object,
+    permission: str,
+    patterns: Iterable[object],
+    ctx: Any,
+) -> bool:
+    """Apply the same Plan file boundary at the permission entry point."""
+
+    if runtime_execution_mode(value) != SessionExecutionMode.PLAN:
+        return True
+    if permission != "edit":
+        return True
+    return plan_edit_patterns_allowed(ctx, patterns)
+
+
+def is_plan_file_edit(value: object, ctx: Any, path: object) -> bool:
+    """Return whether a read-only sandbox may allow this Plan artifact edit."""
+
+    return (
+        runtime_execution_mode(value) == SessionExecutionMode.PLAN
+        and is_current_plan_path(ctx, path)
+    )
+
+
+def filter_tool_names(value: object, tool_names: Iterable[str]) -> list[str]:
+    """Return only tool names allowed by the selected execution mode."""
+
+    return [name for name in tool_names if is_tool_allowed(value, name)]
+
+
+def execution_mode_prompt(
+    value: object,
+    *,
+    session: Any = None,
+    plan_file: Optional[SessionPlanFile] = None,
+) -> str:
+    """Return the per-turn developer guidance for a mode."""
+
+    mode = runtime_execution_mode(value)
+    if mode == SessionExecutionMode.PLAN:
+        file_prompt = (
+            plan_file_prompt(session, plan=plan_file)
+            if session is not None
+            else ""
+        )
+        return f"{PLAN_MODE_PROMPT.rstrip()}\n\n{file_prompt}".strip()
+    return ""
diff --git a/flocks/session/features/todo.py b/flocks/session/features/todo.py
index f941ac331..a34812ec6 100644
--- a/flocks/session/features/todo.py
+++ b/flocks/session/features/todo.py
@@ -91,23 +91,37 @@ async def update(cls, session_id: str, todos: List[TodoInfo]) -> None:
             for todo in todos
         ]
         
-        # Store in storage
         await Storage.set(
             f"todo:{session_id}",
             [todo.model_dump(exclude_none=True) for todo in validated_todos],
-            "todo"
+            "todo",
         )
-        
-        # Publish event
         await Bus.publish(cls.Updated, {
             "sessionID": session_id,
             "todos": validated_todos,
         })
-        
         log.info("todo.updated", {
             "session_id": session_id,
             "count": len(validated_todos),
         })
+
+    @classmethod
+    async def update_active(
+        cls,
+        session_id: str,
+        todos: List[TodoInfo],
+        *,
+        expected_generation: Optional[int] = None,
+    ) -> None:
+        """Update todos only while the owning session remains active."""
+
+        from flocks.session.session import Session
+
+        await Session.run_active_write(
+            session_id,
+            lambda: cls.update(session_id, todos),
+            expected_generation=expected_generation,
+        )
     
     @classmethod
     async def get(cls, session_id: str) -> List[TodoInfo]:
diff --git a/flocks/session/files.py b/flocks/session/files.py
new file mode 100644
index 000000000..3abf90642
--- /dev/null
+++ b/flocks/session/files.py
@@ -0,0 +1,26 @@
+"""Application-owned files associated with persisted sessions."""
+
+from pathlib import Path
+import shutil
+
+from flocks.config.config import Config
+
+
+def session_uploads_dir(session_id: str) -> Path:
+    """Return the upload directory for one session, constrained to app data."""
+
+    uploads_root = (Config.get_data_path() / "uploads").resolve()
+    target = (uploads_root / session_id).resolve()
+    if target == uploads_root or not target.is_relative_to(uploads_root):
+        raise ValueError(f"Invalid session ID for upload path: {session_id}")
+    return target
+
+
+def remove_session_uploads(session_id: str) -> bool:
+    """Remove one session's upload directory when it exists."""
+
+    target = session_uploads_dir(session_id)
+    if not target.is_dir():
+        return False
+    shutil.rmtree(target)
+    return True
diff --git a/flocks/session/goal.py b/flocks/session/goal.py
index 29c1b5b80..a43572e70 100644
--- a/flocks/session/goal.py
+++ b/flocks/session/goal.py
@@ -199,7 +199,6 @@ async def judge_goal_with_model(
         ],
         **provider_options,
         max_tokens=JUDGE_MAX_TOKENS,
-        temperature=0,
     )
 
     payload = _extract_json_object(response.content)
diff --git a/flocks/session/interaction_queue.py b/flocks/session/interaction_queue.py
index b92464f9e..ba240d4f7 100644
--- a/flocks/session/interaction_queue.py
+++ b/flocks/session/interaction_queue.py
@@ -8,6 +8,7 @@
 
 from pydantic import BaseModel, Field
 
+from flocks.session.execution_mode import SessionExecutionMode
 from flocks.utils.id import Identifier
 
 
@@ -35,6 +36,7 @@ class QueuedPrompt(BaseModel):
     mockReply: Optional[str] = None
     tools: Optional[Dict[str, bool]] = None
     system: Optional[str] = None
+    executionMode: SessionExecutionMode = SessionExecutionMode.BUILD
     status: str = "pending"
     createdAt: int = Field(default_factory=lambda: int(time.time() * 1000))
     updatedAt: int = Field(default_factory=lambda: int(time.time() * 1000))
@@ -45,6 +47,7 @@ class InteractionQueue:
 
     _queues: Dict[str, List[QueuedPrompt]] = {}
     _locks: Dict[str, asyncio.Lock] = {}
+    _paused: set[str] = set()
 
     @classmethod
     def _lock_for(cls, session_id: str) -> asyncio.Lock:
@@ -69,6 +72,7 @@ async def enqueue(
         mock_reply: Optional[str] = None,
         tools: Optional[Dict[str, bool]] = None,
         system: Optional[str] = None,
+        execution_mode: SessionExecutionMode = SessionExecutionMode.BUILD,
     ) -> QueuedPrompt:
         async with cls._lock_for(session_id):
             queue = cls._queues.setdefault(session_id, [])
@@ -87,6 +91,7 @@ async def enqueue(
                 mockReply=mock_reply,
                 tools=dict(tools) if tools else None,
                 system=system,
+                executionMode=execution_mode,
             )
             queue.append(item)
             return item
@@ -133,6 +138,8 @@ async def remove(cls, session_id: str, item_id: str) -> QueuedPrompt:
     @classmethod
     async def pop_next(cls, session_id: str) -> Optional[QueuedPrompt]:
         async with cls._lock_for(session_id):
+            if session_id in cls._paused:
+                return None
             queue = cls._queues.get(session_id, [])
             if not queue:
                 return None
@@ -159,6 +166,18 @@ async def promote(cls, session_id: str, item_id: str) -> QueuedPrompt:
     async def clear(cls, session_id: str) -> None:
         async with cls._lock_for(session_id):
             cls._queues.pop(session_id, None)
+            cls._paused.discard(session_id)
+
+    @classmethod
+    async def pause(cls, session_id: str) -> None:
+        async with cls._lock_for(session_id):
+            if cls._queues.get(session_id):
+                cls._paused.add(session_id)
+
+    @classmethod
+    async def resume(cls, session_id: str) -> None:
+        async with cls._lock_for(session_id):
+            cls._paused.discard(session_id)
 
     @classmethod
     def _find_locked(cls, session_id: str, item_id: str) -> QueuedPrompt:
diff --git a/flocks/session/lifecycle/revert.py b/flocks/session/lifecycle/revert.py
index 3fc7fc84f..ebfaec0cb 100644
--- a/flocks/session/lifecycle/revert.py
+++ b/flocks/session/lifecycle/revert.py
@@ -32,6 +32,29 @@ class SessionRevert:
     
     Provides simplified API for route handlers.
     """
+
+    @classmethod
+    async def ensure_replayable(cls, session: SessionInfo, message_id: str) -> None:
+        """Reject replay across the latest project-move boundary."""
+
+        metadata = session.metadata if isinstance(session.metadata, dict) else {}
+        move_metadata = metadata.get("projectMove")
+        if not isinstance(move_metadata, dict):
+            return
+        boundary_message_id = move_metadata.get("boundaryMessageID")
+        if not boundary_message_id:
+            return
+
+        messages = await Message.list(session.id, include_archived=True)
+        message_ids = [message.id for message in messages]
+        if boundary_message_id not in message_ids or message_id not in message_ids:
+            raise ValueError(
+                "移动项目前的历史消息不能重发或重新生成"
+            )
+        if message_ids.index(message_id) <= message_ids.index(boundary_message_id):
+            raise ValueError(
+                "移动项目前的历史消息不能重发或重新生成"
+            )
     
     @classmethod
     async def revert(
@@ -54,18 +77,36 @@ async def revert(
         session = await Session.get_by_id(session_id)
         if not session:
             raise ValueError(f"Session {session_id} not found")
-        
-        input_obj = RevertInput(
-            session_id=session_id,
-            message_id=message_id,
-            part_id=part_id,
-        )
-        
-        return await SessionRevertManager.revert(
-            project_id=session.project_id,
-            input=input_obj,
-            worktree=session.directory,
-        )
+
+        async with Session.active_operation(session_id) as active_session:
+            if part_id is not None:
+                parts = await Message.parts(message_id, session_id)
+                part_belongs_to_message = any(
+                    (
+                        part.get("id")
+                        if isinstance(part, dict)
+                        else getattr(part, "id", None)
+                    ) == part_id
+                    for part in parts
+                )
+                if not part_belongs_to_message:
+                    raise ValueError(
+                        f"Part {part_id} does not belong to message {message_id}"
+                    )
+
+            await cls.ensure_replayable(active_session, message_id)
+
+            input_obj = RevertInput(
+                session_id=session_id,
+                message_id=message_id,
+                part_id=part_id,
+            )
+
+            return await SessionRevertManager.revert(
+                project_id=active_session.project_id,
+                input=input_obj,
+                worktree=active_session.directory,
+            )
     
     @classmethod
     async def unrevert(cls, session_id: str) -> SessionInfo:
@@ -164,9 +205,15 @@ async def revert(
                     continue
                 
                 # Check if this is the revert point
-                if (msg.id == input.message_id and not input.part_id) or \
-                   (hasattr(part, "id") and part.id == input.part_id) or \
-                   (isinstance(part, dict) and part.get("id") == input.part_id):
+                is_requested_part = (
+                    input.part_id is not None
+                    and msg.id == input.message_id
+                    and (
+                        (hasattr(part, "id") and part.id == input.part_id)
+                        or (isinstance(part, dict) and part.get("id") == input.part_id)
+                    )
+                )
+                if (msg.id == input.message_id and not input.part_id) or is_requested_part:
                     # Check if remaining parts have useful content
                     has_useful = any(
                         (isinstance(p, dict) and p.get("type") in ["text", "tool"]) or
diff --git a/flocks/session/message.py b/flocks/session/message.py
index 7fd6dcbe2..0ca768a1a 100644
--- a/flocks/session/message.py
+++ b/flocks/session/message.py
@@ -21,6 +21,7 @@
 from flocks.utils.id import Identifier
 from flocks.storage.storage import Storage
 from flocks.session.recorder import Recorder
+from flocks.session.execution_mode import SessionExecutionMode
 
 log = Log.create(service="message")
 
@@ -376,6 +377,10 @@ class UserMessageInfo(BaseModel):
     tools: Optional[Dict[str, bool]] = Field(None, description="Tool availability")
     variant: Optional[str] = Field(None, description="Prompt variant")
     compacted: Optional[bool] = Field(None, description="Archived by compaction (soft-deleted)")
+    executionMode: SessionExecutionMode = Field(
+        SessionExecutionMode.BUILD,
+        description="Execution mode used for this user turn",
+    )
 
 
 class AssistantMessageInfo(BaseModel):
@@ -466,6 +471,25 @@ def _cancel_parts_flush_task(cls, session_id: str) -> None:
         if task and not task.done():
             task.cancel()
 
+    @classmethod
+    async def quiesce_parts(cls, session_id: str, *, persist: bool) -> None:
+        """Stop a delayed parts flush, optionally persisting its latest cache."""
+        task = cls._parts_flush_tasks.pop(session_id, None)
+        if task and not task.done():
+            task.cancel()
+            try:
+                await task
+            except asyncio.CancelledError:
+                pass
+
+        async with _session_locks.get(session_id):
+            if persist and session_id in cls._parts_cache:
+                if cls._parts_storage_format.get(session_id) == "legacy":
+                    await cls._persist_parts(session_id)
+                else:
+                    for message_id in list(cls._parts_cache[session_id]):
+                        await cls._persist_parts(session_id, message_id=message_id)
+
     @classmethod
     def _cache_token(cls, session_id: str) -> tuple[int, int]:
         return cls._cache_epoch, cls._session_cache_generations.get(session_id, 0)
@@ -942,6 +966,11 @@ def _normalize_stored_message(
             if not isinstance(model_raw, dict):
                 model_raw = {}
             normalized["agent"] = normalized.get("agent") or "rex"
+            normalized["executionMode"] = (
+                normalized.get("executionMode")
+                or normalized.get("execution_mode")
+                or SessionExecutionMode.BUILD.value
+            )
             normalized["model"] = {
                 "providerID": model_raw.get("providerID")
                 or normalized.get("providerID")
@@ -1863,19 +1892,73 @@ async def delete(cls, session_id: str, message_id: str) -> bool:
                 return False
             messages = cls._messages_cache.get(session_id, [])
             if idx < len(messages) and messages[idx].id == message_id:
-                messages.pop(idx)
+                missing = object()
+                removed_message = messages.pop(idx)
                 cls._rebuild_id_index(session_id)
-                if session_id in cls._parts_cache:
-                    cls._parts_cache[session_id].pop(message_id, None)
-                cls._parts_revision_cache.get(session_id, {}).pop(message_id, None)
-                cls._parts_serialized_cache.get(session_id, {}).pop(message_id, None)
+                parts_cache = cls._parts_cache.get(session_id)
+                removed_parts = (
+                    parts_cache.pop(message_id, missing)
+                    if parts_cache is not None
+                    else missing
+                )
+                parts_revision_cache = cls._parts_revision_cache.get(session_id)
+                removed_revision = (
+                    parts_revision_cache.pop(message_id, missing)
+                    if parts_revision_cache is not None
+                    else missing
+                )
+                parts_serialized_cache = cls._parts_serialized_cache.get(session_id)
+                removed_serialized = (
+                    parts_serialized_cache.pop(message_id, missing)
+                    if parts_serialized_cache is not None
+                    else missing
+                )
+                had_pending_parts_flush = session_id in cls._parts_flush_tasks
                 cls._cancel_parts_flush_task(session_id)
-                await cls._persist_messages(session_id)
-                if cls._parts_storage_format.get(session_id) == "legacy":
-                    await cls._persist_parts(session_id)
-                else:
-                    await Storage.delete(cls._parts_item_key(session_id, message_id))
-                    cls._parts_persisted_mids.setdefault(session_id, set()).discard(message_id)
+                try:
+                    await cls._persist_messages(session_id)
+                except BaseException:
+                    # Message metadata is the deletion commit point. Restore
+                    # every in-memory index/cache if it was not persisted so a
+                    # later retry or process restart sees the same message.
+                    messages.insert(idx, removed_message)
+                    cls._rebuild_id_index(session_id)
+                    if parts_cache is not None and removed_parts is not missing:
+                        parts_cache[message_id] = removed_parts
+                    if (
+                        parts_revision_cache is not None
+                        and removed_revision is not missing
+                    ):
+                        parts_revision_cache[message_id] = removed_revision
+                    if (
+                        parts_serialized_cache is not None
+                        and removed_serialized is not missing
+                    ):
+                        parts_serialized_cache[message_id] = removed_serialized
+                    if had_pending_parts_flush:
+                        cls._schedule_parts_flush(
+                            session_id,
+                            message_id=message_id,
+                        )
+                    raise
+
+                try:
+                    if cls._parts_storage_format.get(session_id) == "legacy":
+                        await cls._persist_parts(session_id)
+                    else:
+                        await Storage.delete(cls._parts_item_key(session_id, message_id))
+                        cls._parts_persisted_mids.setdefault(session_id, set()).discard(
+                            message_id
+                        )
+                except Exception as exc:
+                    # Metadata deletion has committed. Orphaned parts are not
+                    # user-visible and can be cleaned later; restoring the
+                    # message here would make cache and durable metadata diverge.
+                    log.warn("message.delete.parts_cleanup_failed", {
+                        "session_id": session_id,
+                        "message_id": message_id,
+                        "error": str(exc),
+                    })
                 log.info("message.deleted", {"id": message_id, "session_id": session_id})
                 return True
             return False
@@ -1924,7 +2007,7 @@ async def clear(cls, session_id: str) -> int:
             Number of messages cleared
         """
         await cls._ensure_message_cache(session_id)
-        
+
         async with _session_locks.get(session_id):
             count = len(cls._messages_cache.get(session_id, []))
             cls._messages_cache[session_id] = []
@@ -1935,18 +2018,33 @@ async def clear(cls, session_id: str) -> int:
             cls._parts_storage_format[session_id] = "per_message"
             cls._parts_fully_loaded.add(session_id)
             cls._cancel_parts_flush_task(session_id)
-            
-            # Persist changes
+
             await cls._persist_messages(session_id)
             await Storage.clear(prefix=cls._parts_item_prefix(session_id))
             await Storage.delete(cls._parts_blob_key(session_id))
-            
+
             log.info("messages.cleared", {
                 "session_id": session_id,
                 "count": count,
             })
-            
             return count
+
+    @classmethod
+    async def clear_active(
+        cls,
+        session_id: str,
+        *,
+        expected_generation: Optional[int] = None,
+    ) -> int:
+        """Clear history only while the owning session remains active."""
+
+        from flocks.session.session import Session
+
+        return await Session.run_active_write(
+            session_id,
+            lambda: cls.clear(session_id),
+            expected_generation=expected_generation,
+        )
     
     @classmethod
     def invalidate_cache(cls, session_id: Optional[str] = None) -> None:
@@ -2479,7 +2577,7 @@ def delete(cls, session_id: str, message_id: str) -> bool:
     @classmethod
     def clear(cls, session_id: str) -> int:
         """Sync version"""
-        return cls._run_async(Message.clear(session_id))
+        return cls._run_async(Message.clear_active(session_id))
     
     @classmethod
     def get_text_content(cls, message: MessageInfo) -> str:
diff --git a/flocks/session/plan_file.py b/flocks/session/plan_file.py
new file mode 100644
index 000000000..ee650ac6d
--- /dev/null
+++ b/flocks/session/plan_file.py
@@ -0,0 +1,180 @@
+"""Session-scoped plan files modeled after OpenCode's plan artifacts."""
+
+from __future__ import annotations
+
+import os
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Iterable, Optional
+
+
+PLAN_DIRECTORY = Path(".flocks") / "plans"
+_SAFE_COMPONENT_RE = re.compile(r"[^A-Za-z0-9._-]+")
+
+
+@dataclass(frozen=True)
+class SessionPlanFile:
+    """Resolved plan artifact for one session."""
+
+    path: Path
+    relative_path: str
+    permission_path: str
+
+
+def _safe_component(value: object, fallback: str) -> str:
+    normalized = _SAFE_COMPONENT_RE.sub("-", str(value or "")).strip(".-")
+    return normalized or fallback
+
+
+def session_plan_file(
+    session: Any,
+    *,
+    worktree: Optional[str] = None,
+) -> SessionPlanFile:
+    """Return the stable, project-local plan file for a session."""
+
+    session_root = Path(str(session.directory)).expanduser().resolve(strict=False)
+    root = (
+        Path(worktree).expanduser().resolve(strict=False)
+        if worktree and Path(worktree) != Path("/")
+        else session_root
+    )
+    created = int(getattr(getattr(session, "time", None), "created", 0) or 0)
+    slug = _safe_component(getattr(session, "slug", None), "session")
+    filename = f"{created}-{slug}.md"
+    path = root / PLAN_DIRECTORY / filename
+    relative_path = Path(os.path.relpath(path, session_root)).as_posix()
+    return SessionPlanFile(
+        path=path,
+        relative_path=relative_path,
+        permission_path=(PLAN_DIRECTORY / filename).as_posix(),
+    )
+
+
+def context_plan_file(ctx: Any) -> Optional[SessionPlanFile]:
+    """Resolve the plan artifact in the tool's host or sandbox workspace."""
+
+    extra = getattr(ctx, "extra", {}) or {}
+    relative_path = str(extra.get("plan_relative_path") or "").strip()
+    permission_path = str(extra.get("plan_permission_path") or "").strip()
+    sandbox = extra.get("sandbox")
+    if isinstance(sandbox, dict) and sandbox.get("workspace_dir") and relative_path:
+        path = (
+            Path(str(sandbox["workspace_dir"])).expanduser()
+            / Path(relative_path)
+        ).resolve(strict=False)
+    else:
+        raw_path = str(extra.get("plan_file_path") or "").strip()
+        if not raw_path:
+            return None
+        path = Path(raw_path).expanduser().absolute()
+    if not relative_path:
+        return None
+    return SessionPlanFile(
+        path=path,
+        relative_path=relative_path,
+        permission_path=permission_path,
+    )
+
+
+def _validation_root(ctx: Any) -> Path:
+    extra = getattr(ctx, "extra", {}) or {}
+    sandbox = extra.get("sandbox")
+    if isinstance(sandbox, dict) and sandbox.get("workspace_dir"):
+        return Path(str(sandbox["workspace_dir"])).expanduser().resolve(strict=False)
+    workspace_dir = extra.get("workspace_dir")
+    if workspace_dir:
+        return Path(str(workspace_dir)).expanduser().resolve(strict=False)
+    return Path.cwd().resolve(strict=False)
+
+
+def _expected_plan_path(ctx: Any) -> Optional[Path]:
+    extra = getattr(ctx, "extra", {}) or {}
+    sandbox = extra.get("sandbox")
+    relative_path = str(extra.get("plan_relative_path") or "").strip()
+    if isinstance(sandbox, dict) and sandbox.get("workspace_dir") and relative_path:
+        return _validation_root(ctx) / Path(relative_path)
+    absolute_path = str(extra.get("plan_file_path") or "").strip()
+    if absolute_path:
+        return Path(absolute_path).expanduser().absolute()
+    return None
+
+
+def _normalize_tool_path(ctx: Any, raw_path: object) -> Optional[Path]:
+    value = str(raw_path or "").strip()
+    if not value:
+        return None
+    path = Path(value).expanduser()
+    if not path.is_absolute():
+        path = _validation_root(ctx) / path
+    return path.resolve(strict=False)
+
+
+def is_current_plan_path(ctx: Any, raw_path: object) -> bool:
+    """Return whether a tool path is exactly the current session plan file."""
+
+    expected = _expected_plan_path(ctx)
+    candidate = _normalize_tool_path(ctx, raw_path)
+    if expected is None or candidate is None:
+        return False
+    # Keep the expected path lexical. If .flocks/plans or the file itself is a
+    # symlink, resolving the candidate moves it elsewhere and the comparison
+    # fails instead of granting an external write.
+    return os.path.normcase(str(candidate)) == os.path.normcase(str(expected.absolute()))
+
+
+def plan_edit_patterns_allowed(ctx: Any, patterns: Iterable[object]) -> bool:
+    """Validate resolved edit permission patterns for Plan mode."""
+
+    values = list(patterns)
+    expected = _expected_plan_path(ctx)
+    expected_is_safe = bool(
+        expected
+        and os.path.normcase(str(expected.resolve(strict=False)))
+        == os.path.normcase(str(expected.absolute()))
+    )
+    extra = getattr(ctx, "extra", {}) or {}
+    accepted_relative_paths = {
+        Path(str(value)).as_posix()
+        for value in (
+            extra.get("plan_relative_path"),
+            extra.get("plan_permission_path"),
+        )
+        if value
+    }
+    return bool(values) and all(
+        (
+            expected_is_safe
+            and Path(str(value)).as_posix() in accepted_relative_paths
+        )
+        or is_current_plan_path(ctx, value)
+        for value in values
+    )
+
+
+def plan_file_prompt(
+    session: Any,
+    *,
+    plan: Optional[SessionPlanFile] = None,
+) -> str:
+    """Build the OpenCode-style per-turn plan file reminder."""
+
+    plan = plan or session_plan_file(session)
+    if plan.path.is_file():
+        file_guidance = (
+            f"A plan file already exists at `{plan.relative_path}`. Read it and "
+            "update it incrementally with the edit or write tool."
+        )
+    else:
+        file_guidance = (
+            f"No plan file exists yet. Create it at `{plan.relative_path}` with "
+            "the write tool."
+        )
+    return f"""## Plan File
+
+{file_guidance}
+
+This is the only file you may edit in Plan mode. Keep the decision-complete
+implementation plan in this file, then call plan_exit.
+"""
diff --git a/flocks/session/policy.py b/flocks/session/policy.py
index 998779b70..89c07556f 100644
--- a/flocks/session/policy.py
+++ b/flocks/session/policy.py
@@ -11,6 +11,8 @@
 
 from typing import Optional, TYPE_CHECKING
 
+from flocks.auth.context import API_TOKEN_SERVICE_USER_ID
+
 if TYPE_CHECKING:
     from flocks.auth.context import AuthUser
     from flocks.session.session import SessionInfo
@@ -34,6 +36,11 @@ def _resolve_user(user: Optional["AuthUser"]) -> Optional["AuthUser"]:
     def is_owner(session: "SessionInfo", user: Optional["AuthUser"]) -> bool:
         if user is None:
             return False
+        if session.owner_user_id == API_TOKEN_SERVICE_USER_ID or (
+            not session.owner_user_id
+            and session.owner_username == API_TOKEN_SERVICE_USER_ID
+        ):
+            return user.id == API_TOKEN_SERVICE_USER_ID
         if session.owner_user_id and session.owner_user_id == user.id:
             return True
         if session.owner_username and session.owner_username == user.username:
@@ -55,6 +62,17 @@ def is_local_shared(session: "SessionInfo") -> bool:
     def _has_no_owner(session: "SessionInfo") -> bool:
         return not session.owner_user_id and not session.owner_username
 
+    @staticmethod
+    def _is_system_owned(session: "SessionInfo") -> bool:
+        return session.owner_user_id == API_TOKEN_SERVICE_USER_ID or (
+            not session.owner_user_id
+            and session.owner_username == API_TOKEN_SERVICE_USER_ID
+        )
+
+    @classmethod
+    def _is_admin_managed(cls, session: "SessionInfo") -> bool:
+        return cls._has_no_owner(session) or cls._is_system_owned(session)
+
     @classmethod
     def is_shared(
         cls,
@@ -128,14 +146,14 @@ def can_read(
 
         - No auth context (CLI/internal runtime): keep legacy permissive behaviour.
         - Logged-in users: owner, local-shared readers, shared readers, or admins
-          managing ownerless legacy/channel sessions.
+          managing system-owned and ownerless legacy/channel sessions.
         """
         resolved = cls._resolve_user(user)
         if resolved is None:
             return True
         if cls.is_owner(session, resolved):
             return True
-        if cls._has_no_owner(session) and cls.is_admin(resolved):
+        if cls._is_admin_managed(session) and cls.is_admin(resolved):
             return True
         return cls.is_shared_read_only(session, resolved, shared_project_ids)
 
@@ -144,15 +162,15 @@ def can_write(cls, session: "SessionInfo", user: Optional["AuthUser"] = None) ->
         """
         Session write permission.
 
-        Owner can always write. Admins may repair/manage ownerless sessions
-        accumulated before local ownership was available.
+        Owner can always write. Admins may manage system-owned sessions and
+        repair ownerless sessions accumulated before ownership was available.
         """
         resolved = cls._resolve_user(user)
         if resolved is None:
             return False
         if cls.is_owner(session, resolved):
             return True
-        if cls._has_no_owner(session) and cls.is_admin(resolved):
+        if cls._is_admin_managed(session) and cls.is_admin(resolved):
             return True
         return False
 
@@ -163,6 +181,6 @@ def can_delete(cls, session: "SessionInfo", user: Optional["AuthUser"]) -> bool:
             return False
         if cls.is_owner(session, resolved):
             return True
-        if cls._has_no_owner(session) and cls.is_admin(resolved):
+        if cls._is_admin_managed(session) and cls.is_admin(resolved):
             return True
         return False
diff --git a/flocks/session/prompt.py b/flocks/session/prompt.py
index 56466f8cf..21fb0ebb3 100644
--- a/flocks/session/prompt.py
+++ b/flocks/session/prompt.py
@@ -1077,6 +1077,7 @@ async def build_system_prompts(
         agent_prompt: Optional[str],
         provider_id: str,
         model_id: str,
+        execution_mode_prompt: Optional[str] = None,
         prompt_tool_names: Iterable[str] = (),
         tool_revision: Optional[int] = None,
         memory_bootstrap_data: Optional[Dict[str, Any]] = None,
@@ -1165,6 +1166,13 @@ async def build_custom_context() -> Optional[str]:
                 digest_inputs={"agent_name": agent_name, "agent_prompt": agent_prompt or ""},
                 builder=lambda: cls._normalize_prompt_text(agent_prompt),
             ),
+            cls._build_cached_prompt_block(
+                static_cache=static_cache,
+                name="execution_mode",
+                cache_scope="runtime",
+                digest_inputs={"prompt": execution_mode_prompt or ""},
+                builder=lambda: cls._normalize_prompt_text(execution_mode_prompt),
+            ),
             cls._build_cached_prompt_block(
                 static_cache=static_cache,
                 name="memory_snapshot",
diff --git a/flocks/session/runner.py b/flocks/session/runner.py
index 69aec5144..ab388b073 100644
--- a/flocks/session/runner.py
+++ b/flocks/session/runner.py
@@ -47,6 +47,11 @@
     ReasoningDeltaEvent,
     ReasoningEndEvent,
 )
+from flocks.session.streaming.timeouts import (
+    DEFAULT_FIRST_CHUNK_TIMEOUT_S,
+    DEFAULT_ONGOING_CHUNK_TIMEOUT_S,
+    resolve_llm_stream_timeouts,
+)
 from flocks.session.callable_schema import list_session_callable_tool_infos
 from flocks.agent.registry import Agent
 from flocks.agent.agent import AgentInfo
@@ -66,6 +71,13 @@
     is_text_extractable_mime,
     extract_file_text,
 )
+from flocks.session.execution_mode import (
+    SessionExecutionMode,
+    execution_mode_prompt,
+    is_tool_allowed,
+    runtime_execution_mode,
+)
+from flocks.session.plan_file import session_plan_file
 
 
 log = Log.create(service="session.runner")
@@ -102,13 +114,13 @@ def _annotate_with_provider_version(tool_info: Any, description: Optional[str])
 # Maximum seconds to wait for the *first* chunk from the LLM stream.
 # If the model never starts responding, the stream times out and the session
 # surfaces a clear error rather than hanging forever.
-LLM_STREAM_FIRST_CHUNK_TIMEOUT_S = 60
+LLM_STREAM_FIRST_CHUNK_TIMEOUT_S = DEFAULT_FIRST_CHUNK_TIMEOUT_S
 
 # Once the stream has started (at least one chunk received), allow a much
 # longer gap between chunks.  Some models pause for extended periods between
 # reasoning and content generation phases; a tight inter-chunk timeout causes
 # spurious failures in those cases.
-LLM_STREAM_ONGOING_CHUNK_TIMEOUT_S = 300
+LLM_STREAM_ONGOING_CHUNK_TIMEOUT_S = DEFAULT_ONGOING_CHUNK_TIMEOUT_S
 
 _RETRYABLE_TRANSPORT_EXCEPTIONS = (
     httpx.TransportError,
@@ -204,6 +216,43 @@ class ToolCall:
     name: str
     arguments: Dict[str, Any]
 
+
+@dataclass
+class LlmAttemptState:
+    """Observable side effects accumulated across retries for one model."""
+
+    received_chunk: bool = False
+    observable_output_started: bool = False
+    tool_execution_started: bool = False
+
+    @property
+    def replay_safe(self) -> bool:
+        """Whether the same logical LLM call can safely run on another model."""
+        return not self.observable_output_started and not self.tool_execution_started
+
+
+@dataclass(frozen=True)
+class FailoverDecision:
+    """Hermes-aligned retry/failover classification for a provider error."""
+
+    eligible: bool
+    reason: str
+    same_model_retries: int = 3
+
+
+@dataclass
+class StepFailure:
+    """Failure details returned to SessionLoop when finalization is deferred."""
+
+    message: str
+    error_data: Dict[str, Any]
+    assistant_message_id: Optional[str]
+    reason: str
+    allow_fallback: bool
+    attempt_state: LlmAttemptState
+    attempts: int = 0
+
+
 @dataclass
 class StepResult:
     """Result of a single processing step."""
@@ -212,6 +261,7 @@ class StepResult:
     tool_calls: List[ToolCall] = field(default_factory=list)
     error: Optional[str] = None
     usage: Optional[Dict[str, int]] = None
+    failure: Optional[StepFailure] = None
 
 
 @dataclass
@@ -257,6 +307,10 @@ def __init__(
         session_ctx: Optional[Any] = None,  # SessionContext interface
         memory_bootstrap_data: Optional[Dict[str, Any]] = None,
         static_cache: Optional[Dict[str, Any]] = None,
+        defer_step_errors: bool = False,
+        failover_available: bool = False,
+        turn_additional_context: Optional[str] = None,
+        session_start_pending: bool = False,
     ):
         self.session = session
         from flocks.session.core.defaults import fallback_provider_id, fallback_model_id
@@ -271,6 +325,12 @@ def __init__(
         self.session_ctx = session_ctx  # SessionContext interface for decoupled access
         self._memory_bootstrap_data: Optional[Dict[str, Any]] = memory_bootstrap_data
         self._static_cache = static_cache if static_cache is not None else {}
+        self._defer_step_errors = defer_step_errors
+        self._failover_available = failover_available
+        self._turn_additional_context = turn_additional_context
+        self._session_start_pending = session_start_pending
+        self._session_start_fired = False
+        self._attempt_state = LlmAttemptState()
 
     @staticmethod
     def _canonical_tool_signature(tool_name: str, arguments: Dict[str, Any]) -> str:
@@ -308,6 +368,26 @@ def _should_warn_about_tool_loop(self, *, last_user_id: str) -> bool:
         state = self._get_tool_loop_guard_state(last_user_id=last_user_id)
         return int(state.get("exact_count", 0)) >= max(2, REPEATED_EXACT_TOOL_CALL_HALT_THRESHOLD - 1)
 
+    async def _run_session_start_hook(self, agent: Any) -> None:
+        """Run SessionStart after the first system prompt has been built."""
+        if not self._session_start_pending or self._session_start_fired:
+            return
+        self._session_start_fired = True
+        try:
+            from flocks.hooks.pipeline import HookPipeline
+
+            await HookPipeline.run_session_start({
+                "sessionID": self.session.id,
+                "workspace": self.session.directory,
+                "agent": agent.name,
+                "model": {
+                    "providerID": self.provider_id,
+                    "modelID": self.model_id,
+                },
+            })
+        except Exception as exc:
+            log.debug("runner.hook.session_start.error", {"error": str(exc)})
+
     def _build_tool_loop_halt_message(
         self,
         *,
@@ -498,13 +578,43 @@ async def _list_callable_tool_infos_for_turn(
         agent: AgentInfo,
         messages: List[MessageInfo],
     ) -> Tuple[List[Any], Dict[str, Any]]:
+        execution_mode = self._execution_mode_from_messages(messages)
         result = await list_session_callable_tool_infos(
             session_id=self.session.id,
             declared_tool_names=getattr(agent, "tools", None),
             step=self._step,
             event_publish_callback=self.callbacks.event_publish_callback,
         )
-        return result.tool_infos, dict(result.metadata)
+        tool_infos = [
+            tool_info
+            for tool_info in result.tool_infos
+            if is_tool_allowed(execution_mode, tool_info.name)
+        ]
+        if (
+            execution_mode == SessionExecutionMode.PLAN
+            and all(tool_info.name != "plan_exit" for tool_info in tool_infos)
+        ):
+            plan_exit = ToolRegistry.get("plan_exit")
+            if plan_exit is not None and getattr(plan_exit.info, "enabled", True):
+                tool_infos.append(plan_exit.info)
+        metadata = dict(result.metadata)
+        metadata["executionMode"] = execution_mode.value
+        metadata["modeAllowedToolNames"] = sorted(
+            tool_info.name for tool_info in tool_infos
+        )
+        return tool_infos, metadata
+
+    @staticmethod
+    def _execution_mode_from_messages(
+        messages: Optional[List[MessageInfo]],
+    ) -> SessionExecutionMode:
+        for message in reversed(messages or []):
+            if getattr(message, "role", None) != MessageRole.USER:
+                continue
+            return runtime_execution_mode(
+                getattr(message, "executionMode", None)
+            )
+        return runtime_execution_mode(None)
 
     @staticmethod
     def _get_prompt_tool_names_from_schema(tools: List[Dict[str, Any]]) -> Tuple[str, ...]:
@@ -1093,6 +1203,122 @@ def is_aborted(self) -> bool:
         if self._external_abort is not None and self._external_abort.is_set():
             return True
         return False
+
+    @staticmethod
+    def classify_failover_error(error: Dict[str, Any]) -> FailoverDecision:
+        """Classify a provider failure using Hermes-compatible switch timing.
+
+        The classifier deliberately requires an API-shaped error (status code,
+        APIError marker, or a known provider response pattern). Local Python,
+        storage, hook, and tool failures must never move a user turn to another
+        model.
+        """
+        data = error.get("data") or {}
+        status_code = data.get("statusCode")
+        try:
+            status_code = int(status_code) if status_code is not None else None
+        except (TypeError, ValueError):
+            status_code = None
+
+        message = str(data.get("message") or error.get("message") or "")
+        lowered = message.lower()
+        error_name = str(error.get("name") or "")
+
+        if status_code == 413 or any(pattern in lowered for pattern in (
+            "context length", "context_length", "context window", "prompt is too long",
+            "request entity too large", "payload too large",
+        )):
+            return FailoverDecision(False, "context_overflow")
+        if error_name in {"CancelledError", "MessageAbortedError", "AbortedError"}:
+            return FailoverDecision(False, "cancelled")
+
+        quota_or_rate_limited = any(pattern in lowered for pattern in (
+            "rate limit", "too many requests", "quota exceeded", "resource exhausted",
+            "insufficient quota", "billing limit",
+        ))
+        # Some providers report exhausted quota as HTTP 401/403 rather than
+        # 429. Classify the semantic error before the generic auth branch so
+        # the primary receives the same cooldown as other quota failures.
+        if status_code == 429 or quota_or_rate_limited:
+            reason = "billing" if any(
+                pattern in lowered for pattern in ("billing", "insufficient quota")
+            ) else "rate_limit"
+            return FailoverDecision(True, reason, 0)
+        if status_code in {401, 403}:
+            return FailoverDecision(True, "auth", 0)
+        if status_code == 402:
+            return FailoverDecision(True, "billing", 0)
+
+        if "model" in lowered and any(pattern in lowered for pattern in (
+            "not found", "model_not_found", "unknown model", "no such model",
+        )):
+            return FailoverDecision(True, "model_not_found", 0)
+
+        if status_code == 404:
+            if any(pattern in lowered for pattern in (
+                "model not found", "model_not_found", "unknown model", "no such model",
+            )):
+                return FailoverDecision(True, "model_not_found", 0)
+            return FailoverDecision(True, "unknown_api", 3)
+
+        if status_code in {408, 504} or data.get("isConnectionError") is True or any(
+            pattern in lowered
+            for pattern in ("timeout", "timed out", "connection error", "connection reset")
+        ):
+            return FailoverDecision(True, "timeout", 1)
+        if status_code in {503, 529} or any(
+            pattern in lowered for pattern in ("overloaded", "temporarily unavailable")
+        ):
+            return FailoverDecision(True, "overloaded", 1)
+        if status_code in {500, 502}:
+            return FailoverDecision(True, "server_error", 3)
+
+        if any(pattern in lowered for pattern in (
+            "content policy", "content filter", "content_filter", "safety policy",
+            "policy violation",
+        )):
+            return FailoverDecision(True, "content_policy", 0)
+        if error_name == "JSONDecodeError" or any(
+            pattern in lowered for pattern in (
+                "malformed response", "invalid response", "empty choices",
+                "returned choice with null", "null message",
+            )
+        ):
+            return FailoverDecision(True, "invalid_response", 0)
+
+        if status_code is not None and 400 <= status_code < 500:
+            return FailoverDecision(True, "provider_request", 0)
+        if error_name == "APIError" or data.get("isRetryable") is True:
+            return FailoverDecision(True, "unknown_api", 3)
+        return FailoverDecision(False, "local_error")
+
+    def _deferred_failure_result(
+        self,
+        *,
+        message: str,
+        error_data: Dict[str, Any],
+        assistant_message_id: Optional[str],
+        decision: FailoverDecision,
+        attempts: int,
+    ) -> StepResult:
+        state = LlmAttemptState(
+            received_chunk=self._attempt_state.received_chunk,
+            observable_output_started=self._attempt_state.observable_output_started,
+            tool_execution_started=self._attempt_state.tool_execution_started,
+        )
+        return StepResult(
+            action="stop",
+            error=message,
+            failure=StepFailure(
+                message=message,
+                error_data=error_data,
+                assistant_message_id=assistant_message_id,
+                reason=decision.reason,
+                allow_fallback=decision.eligible and state.replay_safe,
+                attempt_state=state,
+                attempts=attempts,
+            ),
+        )
     
     async def _process_step(
         self,
@@ -1100,6 +1326,17 @@ async def _process_step(
         last_user: MessageInfo,
     ) -> StepResult:
         """Process a single step in the loop with retry logic."""
+        self._attempt_state = LlmAttemptState()
+        turn_execution_mode = runtime_execution_mode(
+            getattr(last_user, "executionMode", None)
+        )
+        self._turn_execution_mode = turn_execution_mode
+        from flocks.project.instance import Instance
+
+        self._turn_plan_file = session_plan_file(
+            self.session,
+            worktree=Instance.get_worktree(),
+        )
         # Check for CLI callbacks (if running in CLI mode)
         # Only use CLI fallback if no callbacks were explicitly provided via constructor
         has_explicit_callbacks = any([
@@ -1137,6 +1374,18 @@ async def _process_step(
         provider = Provider.get(self.provider_id)
         if not provider:
             error = f"Provider {self.provider_id} not found"
+            if self._defer_step_errors:
+                return self._deferred_failure_result(
+                    message=error,
+                    error_data={
+                        "name": "ProviderUnavailableError",
+                        "message": error,
+                        "data": {"message": error},
+                    },
+                    assistant_message_id=None,
+                    decision=FailoverDecision(True, "provider_unavailable", 0),
+                    attempts=0,
+                )
             error_dict = self._build_session_error_dict(
                 error,
                 name="ProviderUnavailableError",
@@ -1165,6 +1414,18 @@ async def _process_step(
         
         if not provider.is_configured():
             error = f"Provider {self.provider_id} not configured"
+            if self._defer_step_errors:
+                return self._deferred_failure_result(
+                    message=error,
+                    error_data={
+                        "name": "ProviderUnavailableError",
+                        "message": error,
+                        "data": {"message": error},
+                    },
+                    assistant_message_id=None,
+                    decision=FailoverDecision(True, "provider_unavailable", 0),
+                    attempts=0,
+                )
             error_dict = self._build_session_error_dict(
                 error,
                 name="ProviderConfigurationError",
@@ -1212,6 +1473,11 @@ async def device_asset_prompt_factory() -> Optional[str]:
             agent_prompt=getattr(agent, "prompt", None),
             provider_id=self.provider_id,
             model_id=self.model_id,
+            execution_mode_prompt=execution_mode_prompt(
+                turn_execution_mode,
+                session=self.session,
+                plan_file=self._turn_plan_file,
+            ),
             prompt_tool_names=prompt_tool_names,
             tool_revision=ToolRegistry.revision(),
             memory_bootstrap_data=self._memory_bootstrap_data,
@@ -1225,6 +1491,11 @@ async def device_asset_prompt_factory() -> Optional[str]:
         )
         self._log_perf("runner.process_step.system_prompts_ready", prompts_started_at, prompt_count=len(system_prompts))
 
+        await self._run_session_start_hook(agent)
+
+        if self._turn_additional_context:
+            system_prompts.append(self._turn_additional_context)
+
         if self._should_use_text_tool_call_mode() and tools:
             text_tool_catalog = self._build_text_tool_call_catalog_prompt(tools)
             if text_tool_catalog:
@@ -1259,29 +1530,6 @@ async def device_asset_prompt_factory() -> Optional[str]:
                 from flocks.session.prompt_strings import PROMPT_REPEATED_TOOL_CALLS
                 system_prompts.append(PROMPT_REPEATED_TOOL_CALLS)
 
-        # Hook pipeline: chat.message stage
-        try:
-            from flocks.hooks.pipeline import HookPipeline
-            user_text = await Message.get_text_content(last_user)
-            hook_input = {
-                "sessionID": self.session.id,
-                "workspace": self.session.directory,
-                "agent": agent.name,
-                "model": {"providerID": self.provider_id, "modelID": self.model_id},
-                "message": {
-                    "id": last_user.id,
-                    "role": "user",
-                    "content": user_text,
-                },
-            }
-            hook_output = {"message": {"variant": getattr(last_user, "variant", None)}}
-            ctx = await HookPipeline.run_chat_message(hook_input, hook_output)
-            variant = ctx.output.get("message", {}).get("variant") if ctx else None
-            if variant:
-                await Message.update(self.session.id, last_user.id, variant=variant)
-        except Exception as e:
-            log.debug("runner.hook.chat_message.error", {"error": str(e)})
-        
         # Convert messages to chat format with error handling
         try:
             queued_user_message_ids = self._get_queued_user_message_ids(messages)
@@ -1397,7 +1645,7 @@ async def device_asset_prompt_factory() -> Optional[str]:
             try:
                 # Set status to busy
                 SessionStatus.set(self.session.id, SessionStatusBusy())
-                
+
                 # Call LLM with tools
                 result = await self._call_llm(
                     provider=provider,
@@ -1429,7 +1677,11 @@ async def device_asset_prompt_factory() -> Optional[str]:
                 if (result.action == "stop" and not result.error
                         and not result.content and not result.tool_calls):
                     empty_attempt += 1
-                    if empty_attempt <= MAX_EMPTY_RETRIES:
+                    unsafe_auto_replay = (
+                        self._defer_step_errors
+                        and not self._attempt_state.replay_safe
+                    )
+                    if empty_attempt <= MAX_EMPTY_RETRIES and not unsafe_auto_replay:
                         # Record usage for this empty attempt even though we are
                         # about to retry – the provider may have already charged
                         # for the tokens returned in this response.
@@ -1456,10 +1708,16 @@ async def device_asset_prompt_factory() -> Optional[str]:
                         # All retries exhausted — surface a clear error so the
                         # user knows the model is incompatible, rather than
                         # silently hanging or showing a blank response.
-                        empty_error_msg = (
-                            f"Model '{self.model_id}' returned an empty response "
-                            f"after {MAX_EMPTY_RETRIES} retries."
-                        )
+                        if unsafe_auto_replay:
+                            empty_error_msg = (
+                                f"Model '{self.model_id}' returned no final content after "
+                                "starting observable output; the call was not replayed."
+                            )
+                        else:
+                            empty_error_msg = (
+                                f"Model '{self.model_id}' returned an empty response "
+                                f"after {MAX_EMPTY_RETRIES} retries."
+                            )
                         log.error("runner.step.empty_response_exhausted", {
                             "session_id": self.session.id,
                             "model": self.model_id,
@@ -1474,6 +1732,14 @@ async def device_asset_prompt_factory() -> Optional[str]:
                                 "attempts": empty_attempt,
                             },
                         }
+                        if self._defer_step_errors:
+                            return self._deferred_failure_result(
+                                message=empty_error_msg,
+                                error_data=empty_error_dict,
+                                assistant_message_id=assistant_msg.id,
+                                decision=FailoverDecision(True, "empty_response", 3),
+                                attempts=empty_attempt,
+                            )
                         if self.callbacks.on_error:
                             await self.callbacks.on_error(empty_error_msg)
                         await Message.update(
@@ -1522,12 +1788,12 @@ async def device_asset_prompt_factory() -> Optional[str]:
                 finish = "tool-calls" if result.tool_calls else "stop"
                 await Message.update(self.session.id, assistant_msg.id, finish=finish)
                 await self._record_usage_if_available(result.usage, message_id=assistant_msg.id)
-                
+
                 # Note: Compaction check is now done in the main loop (run()) before processing step
                 # This matches Flocks's logic: check lastFinished.tokens at loop start
 
                 return result
-                
+
             except Exception as e:
                 error_attempt += 1
                 error_log_context = {
@@ -1541,7 +1807,24 @@ async def device_asset_prompt_factory() -> Optional[str]:
 
                 # Check if retryable
                 retry_message = SessionRetry.retryable(error_dict)
-                will_retry = retry_message is not None and error_attempt <= MAX_ERROR_RETRIES
+                failover_decision = self.classify_failover_error(error_dict)
+                retry_limit = MAX_ERROR_RETRIES
+                if (
+                    self._defer_step_errors
+                    and failover_decision.eligible
+                ):
+                    retry_limit = failover_decision.same_model_retries
+                    will_retry = error_attempt <= retry_limit
+                    if will_retry and retry_message is None:
+                        retry_message = (
+                            f"Provider error ({failover_decision.reason}), retrying..."
+                        )
+                else:
+                    will_retry = retry_message is not None and error_attempt <= retry_limit
+                if self._defer_step_errors and not self._attempt_state.replay_safe:
+                    # Retrying after text/reasoning/tool activity can duplicate
+                    # visible output or execute a tool twice.
+                    will_retry = False
 
                 if will_retry:
                     # Error is retryable and we have budget left
@@ -1558,7 +1841,7 @@ async def device_asset_prompt_factory() -> Optional[str]:
                         "attempt": error_attempt,
                         "delay_ms": delay_ms,
                         "reason": retry_message,
-                        "max_retries": MAX_ERROR_RETRIES,
+                        "max_retries": retry_limit,
                     })
                     
                     # Set retry status
@@ -1582,7 +1865,7 @@ async def device_asset_prompt_factory() -> Optional[str]:
                         log.error("runner.step.max_retries_exceeded", {
                             **error_log_context,
                             "attempt": error_attempt,
-                            "max_retries": MAX_ERROR_RETRIES,
+                            "max_retries": retry_limit,
                         })
                     else:
                         log.error("runner.step.not_retryable", {
@@ -1595,6 +1878,15 @@ async def device_asset_prompt_factory() -> Optional[str]:
                         final_error_message = CONNECTION_ERROR_DISPLAY_MESSAGE
                         error_dict["data"]["displayMessage"] = CONNECTION_ERROR_DISPLAY_MESSAGE
 
+                    if self._defer_step_errors:
+                        return self._deferred_failure_result(
+                            message=final_error_message,
+                            error_data=error_dict,
+                            assistant_message_id=assistant_msg.id,
+                            decision=failover_decision,
+                            attempts=error_attempt,
+                        )
+
                     if self.callbacks.on_error:
                         await self.callbacks.on_error(final_error_message)
                     
@@ -1634,6 +1926,66 @@ def _build_tokens_update(stream_usage: Optional[Dict[str, int]]) -> Optional[Dic
             },
         }
 
+    @staticmethod
+    def _serialize_chat_message_for_langfuse(message: ChatMessage) -> Dict[str, Any]:
+        """Serialize the exact provider-bound message payload for Langfuse."""
+        if hasattr(message, "model_dump"):
+            return message.model_dump(exclude_none=True)
+        return {
+            "role": message.role,
+            "content": message.content,
+            "reasoning": getattr(message, "reasoning", None),
+            "tool_calls": getattr(message, "tool_calls", None),
+            "tool_call_id": getattr(message, "tool_call_id", None),
+            "name": getattr(message, "name", None),
+            "custom_settings": getattr(message, "custom_settings", {}),
+        }
+
+    @classmethod
+    def _build_langfuse_request_payload(
+        cls,
+        *,
+        step: int,
+        messages: List[ChatMessage],
+        request_tools: Optional[List[Dict[str, Any]]],
+        available_tools: List[Dict[str, Any]],
+        provider_options: Dict[str, Any],
+    ) -> Dict[str, Any]:
+        payload: Dict[str, Any] = {
+            "step": step,
+            "messages": [cls._serialize_chat_message_for_langfuse(message) for message in messages],
+            "provider_options": provider_options,
+        }
+        if request_tools is not None:
+            payload["request_tools"] = request_tools
+        if available_tools:
+            payload["available_tools"] = available_tools
+        return payload
+
+    @staticmethod
+    def _build_langfuse_response_payload(
+        *,
+        action: str,
+        content: str,
+        reasoning: str,
+        finish_reason: Optional[str],
+        tool_calls: List[ToolCall],
+    ) -> Dict[str, Any]:
+        return {
+            "action": action,
+            "content": content,
+            "reasoning": reasoning,
+            "finish_reason": finish_reason,
+            "tool_calls": [
+                {
+                    "id": tool_call.id,
+                    "name": tool_call.name,
+                    "arguments": tool_call.arguments,
+                }
+                for tool_call in tool_calls
+            ],
+        }
+
     def _resolve_usage_pricing(self) -> Optional[Any]:
         """Resolve pricing config for the current provider/model pair."""
         from flocks.provider.usage_service import resolve_usage_pricing
@@ -2017,9 +2369,42 @@ def _exception_to_error_dict(self, exception: Exception) -> Dict[str, Any]:
                 "transportExceptionModule": type(transport_exception).__module__,
             })
         
-        # Check if it's an API error with specific attributes
-        if hasattr(exception, 'status_code'):
-            status_code = getattr(exception, 'status_code')
+        # Provider SDKs expose HTTP status through several shapes. Walk the
+        # normal exception chain so lightweight wrapper errors do not hide it.
+        status_code = None
+        status_exception = None
+        seen: set[int] = set()
+        current: Optional[BaseException] = exception
+        while current is not None and id(current) not in seen:
+            seen.add(id(current))
+            response = getattr(current, "response", None)
+            status_values = (
+                getattr(current, "status_code", None),
+                getattr(response, "status_code", None),
+                getattr(current, "code", None),
+            )
+            for value in status_values:
+                if callable(value):
+                    try:
+                        value = value()
+                    except TypeError:
+                        continue
+                value = getattr(value, "value", value)
+                if isinstance(value, tuple) and value:
+                    value = value[0]
+                try:
+                    normalized = int(value)
+                except (TypeError, ValueError):
+                    continue
+                if 100 <= normalized <= 599:
+                    status_code = normalized
+                    status_exception = current
+                    break
+            if status_code is not None:
+                break
+            current = current.__cause__ or current.__context__
+
+        if status_code is not None:
             error_dict["name"] = "APIError"
             error_dict["data"]["statusCode"] = status_code
             
@@ -2028,9 +2413,13 @@ def _exception_to_error_dict(self, exception: Exception) -> Dict[str, Any]:
             error_dict["data"]["isRetryable"] = is_retryable
             
             # Extract response headers if available
-            if hasattr(exception, 'response') and hasattr(exception.response, 'headers'):
-                headers = dict(exception.response.headers)
-                error_dict["data"]["responseHeaders"] = headers
+            response = getattr(status_exception, "response", None)
+            headers = getattr(response, "headers", None)
+            if headers is not None:
+                try:
+                    error_dict["data"]["responseHeaders"] = dict(headers)
+                except (TypeError, ValueError):
+                    pass
         
         # Check for common retryable error patterns
         error_msg = str(exception).lower()
@@ -2402,7 +2791,7 @@ async def _to_chat_messages(
                                 "type": "text",
                                 "text": "The following tool was executed by the user",
                             })
-                
+
                 if user_content_blocks and any(
                     block.get("type") == "image"
                     for block in user_content_blocks
@@ -2724,6 +3113,19 @@ def _build_llm_response_payload(
         except Exception as e:
             log.debug("runner.sandbox_context_init_failed", {"error": str(e)})
 
+        async def _on_tool_execution_start(
+            tool_name: str,
+            tool_input: Dict[str, Any],
+        ) -> None:
+            # Mark before hooks/callbacks/execution (including parallel tools)
+            # so a concurrent provider error can never replay side effects.
+            self._attempt_state.tool_execution_started = True
+            if self.callbacks.on_tool_start:
+                await self.callbacks.on_tool_start(tool_name, tool_input)
+
+        turn_plan_file = getattr(self, "_turn_plan_file", None)
+        if turn_plan_file is None:
+            turn_plan_file = session_plan_file(self.session)
         processor = StreamProcessor(
             session_id=self.session.id,
             assistant_message=assistant_msg,
@@ -2732,7 +3134,7 @@ def _build_llm_response_payload(
             permission_callback=self._handle_permission,
             text_delta_callback=self.callbacks.on_text_delta,
             reasoning_delta_callback=self.callbacks.on_reasoning_delta,
-            tool_start_callback=self.callbacks.on_tool_start,
+            tool_start_callback=_on_tool_execution_start,
             tool_end_callback=self.callbacks.on_tool_end,
             event_publish_callback=self.callbacks.event_publish_callback,
             session_key=self.session.id,
@@ -2740,11 +3142,22 @@ def _build_llm_response_payload(
             workspace_dir=self.session.directory,
             langfuse_generation=None,
             step_index=self._step,
+            execution_mode=runtime_execution_mode(
+                getattr(
+                    self,
+                    "_turn_execution_mode",
+                    SessionExecutionMode.BUILD,
+                )
+            ).value,
+            plan_file_path=str(turn_plan_file.path),
+            plan_relative_path=turn_plan_file.relative_path,
+            plan_permission_path=turn_plan_file.permission_path,
         )
         
         # Build provider options (thinking / reasoning / max_tokens)
         from flocks.provider.options import build_provider_options
         provider_options = build_provider_options(self.provider_id, self.model_id)
+        provider_tools = None if self._should_use_text_tool_call_mode() else (tools if tools else None)
 
         # Clean up any leftover reasoning state from a previous (failed) call
         if hasattr(self, '_current_reasoning_id'):
@@ -2766,13 +3179,6 @@ def _build_llm_response_payload(
         generation_ctx = None
         if langfuse_is_active():
             try:
-                input_preview = []
-                for _msg in messages[-12:]:
-                    _mc = _msg.content or ""
-                    input_preview.append(
-                        {"role": _msg.role, "chars": len(_mc), "preview": _mc[:240]}
-                    )
-
                 trace_tags = [
                     f"session:{self.session.id}",
                     f"step:{self._step}",
@@ -2780,36 +3186,47 @@ def _build_llm_response_payload(
                     f"agent:{agent.name}",
                     f"provider:{self.provider_id}",
                 ]
+                request_payload = self._build_langfuse_request_payload(
+                    step=self._step,
+                    messages=messages,
+                    request_tools=provider_tools,
+                    available_tools=tools,
+                    provider_options=provider_options,
+                )
                 trace_ctx = trace_scope(
                     name="SessionRunner.step",
                     session_id=self.session.id,
                     tags=trace_tags,
-                    input={
-                        "step": self._step,
-                        "message_count": len(messages),
-                        "tool_count": len(tools),
-                        "last_user_preview": next(
-                            ((m.content or "")[:280] for m in reversed(messages) if m.role == "user"),
-                            "",
-                        ),
-                    },
+                    input=request_payload,
                     metadata={
                         "provider_id": self.provider_id,
                         "model_id": self.model_id,
                         "agent": agent.name,
                         "workspace": self.session.directory,
+                        "message_count": len(messages),
+                        "available_tool_count": len(tools),
+                        "request_tool_count": len(provider_tools or []),
+                        "tool_transport": (
+                            "provider_param" if provider_tools is not None else "text_prompt"
+                        ),
                     },
                 )
                 generation_ctx = generation_scope(
                     parent=trace_ctx.observation,
                     name="LLM.generate",
                     model=self.model_id,
-                    input=input_preview,
+                    input=request_payload,
                     metadata={
                         "provider_id": self.provider_id,
                         "session_id": self.session.id,
                         "step": self._step,
-                        "tool_names": [t.get("function", {}).get("name", "") for t in tools][:50],
+                        "agent": agent.name,
+                        "workspace": self.session.directory,
+                        "available_tool_count": len(tools),
+                        "request_tool_count": len(provider_tools or []),
+                        "tool_transport": (
+                            "provider_param" if provider_tools is not None else "text_prompt"
+                        ),
                     },
                 )
                 processor._langfuse_generation = generation_ctx.observation
@@ -2842,7 +3259,6 @@ def _build_llm_response_payload(
         stream_usage: Optional[Dict[str, int]] = None
         
         # Stream response and convert chunks to events
-        provider_tools = None if self._should_use_text_tool_call_mode() else (tools if tools else None)
         if provider_tools is None and tools:
             log.info("runner.text_tool_call_mode.enabled", {
                 "session_id": self.session.id,
@@ -2904,6 +3320,14 @@ def _build_llm_response_payload(
         llm_call_started_at = time.perf_counter()
         first_chunk_logged = False
         aborted_during_stream = False
+        stream_timeouts = resolve_llm_stream_timeouts(provider, self.model_id)
+        log.debug("runner.llm.stream_timeouts", {
+            "provider_id": self.provider_id,
+            "model_id": self.model_id,
+            "first_chunk_timeout_s": stream_timeouts.first_chunk_s,
+            "ongoing_chunk_timeout_s": stream_timeouts.ongoing_chunk_s,
+            "local_endpoint": stream_timeouts.is_local,
+        })
         try:
             async for chunk in _iter_with_chunk_timeout(
                 provider.chat_stream(
@@ -2917,10 +3341,11 @@ def _build_llm_response_payload(
                     session_id=self.session.id,
                     **provider_options,
                 ),
-                first_chunk_timeout_s=LLM_STREAM_FIRST_CHUNK_TIMEOUT_S,
-                ongoing_chunk_timeout_s=LLM_STREAM_ONGOING_CHUNK_TIMEOUT_S,
+                first_chunk_timeout_s=stream_timeouts.first_chunk_s,
+                ongoing_chunk_timeout_s=stream_timeouts.ongoing_chunk_s,
             ):
                 chunk_counts["total"] += 1
+                self._attempt_state.received_chunk = True
                 if not first_chunk_logged:
                     first_chunk_logged = True
                     self._log_perf(
@@ -2929,20 +3354,20 @@ def _build_llm_response_payload(
                         provider_id=self.provider_id,
                         model_id=self.model_id,
                     )
-                
+
                 chunk_finish = getattr(chunk, 'finish_reason', None)
                 if chunk_finish:
                     stream_finish_reason = chunk_finish
-                
+
                 # Capture usage from chunk (providers may include it in the final chunk)
                 if hasattr(chunk, 'usage') and chunk.usage:
                     stream_usage = chunk.usage
-                
+
                 # Check for abort
                 if self.is_aborted:
                     aborted_during_stream = True
                     break
-                
+
                 # Determine event type from chunk.  A single chunk may carry any
                 # combination of reasoning / text / tool_calls (e.g. Gemini bundles
                 # them).  We must not drop non-reasoning content when reasoning is
@@ -2959,6 +3384,7 @@ def _build_llm_response_payload(
                     self._current_reasoning_metadata = current_metadata
 
                 if event_type == "reasoning-start" and not hasattr(self, '_current_reasoning_id'):
+                    self._attempt_state.observable_output_started = True
                     reasoning_id_counter += 1
                     self._current_reasoning_id = f"reasoning-{reasoning_id_counter}"
                     self._current_reasoning_metadata = dict(chunk_metadata)
@@ -2999,6 +3425,7 @@ def _build_llm_response_payload(
 
                 # 1) Process reasoning delta (start reasoning block on first sight).
                 if chunk_reasoning or (event_type == 'reasoning' and has_reasoning_metadata):
+                    self._attempt_state.observable_output_started = True
                     reasoning_text = chunk_reasoning or ""
                     chunk_counts["reasoning"] += 1
                     log.debug("runner.reasoning.received", {
@@ -3035,6 +3462,7 @@ def _build_llm_response_payload(
 
                 # 3) Process text delta.
                 if chunk_text:
+                    self._attempt_state.observable_output_started = True
                     chunk_counts["text"] += 1
                     if not text_started:
                         await processor.process_event(TextStartEvent())
@@ -3046,6 +3474,9 @@ def _build_llm_response_payload(
 
                 # 4) Process tool calls.
                 if chunk_tool_calls:
+                    # Tool fragments are persisted/accumulated and may become an
+                    # executable call in this same await, so any replay is unsafe.
+                    self._attempt_state.observable_output_started = True
                     chunk_counts["tool"] += 1
                     for tc in chunk_tool_calls:
                         await tool_accumulator.feed_chunk(tc)
@@ -3159,6 +3590,7 @@ def _build_llm_response_payload(
             )
             for tc_state in list(processor.tool_calls.values())
         ]
+        finish_reason = processor.get_finish_reason()
         result_action = "continue" if tool_calls_for_result else "stop"
         response_payload = _build_llm_response_payload(
             content=content,
@@ -3196,26 +3628,23 @@ def _build_llm_response_payload(
                 log.debug("runner.hook.llm_after.error", {"error": str(exc)})
         
         if tool_calls_for_result:
+            response_payload = self._build_langfuse_response_payload(
+                action="continue",
+                content=content,
+                reasoning=reasoning,
+                finish_reason=finish_reason,
+                tool_calls=tool_calls_for_result,
+            )
             self._end_observability(
                 generation_ctx, trace_ctx,
-                output={
-                    "content_preview": content[:600],
-                    "content_chars": len(content),
-                    "reasoning_chars": len(reasoning),
-                    "tool_calls": [{"id": tc.id, "name": tc.name} for tc in tool_calls_for_result[:30]],
-                },
+                output=response_payload,
                 usage=stream_usage,
                 metadata={
-                    "finish_reason": processor.get_finish_reason(),
+                    "finish_reason": finish_reason,
                     "status": "continue_with_tools",
                     "tool_call_count": len(tool_calls_for_result),
                 },
-                trace_output={
-                    "status": "ok",
-                    "next_action": "continue",
-                    "finish_reason": processor.get_finish_reason(),
-                    "tool_call_count": len(tool_calls_for_result),
-                },
+                trace_output=response_payload,
             )
             return StepResult(
                 action=result_action,
@@ -3224,24 +3653,23 @@ def _build_llm_response_payload(
                 usage=stream_usage,
             )
         
+        response_payload = self._build_langfuse_response_payload(
+            action="stop",
+            content=content,
+            reasoning=reasoning,
+            finish_reason=finish_reason,
+            tool_calls=tool_calls_for_result,
+        )
         self._end_observability(
             generation_ctx, trace_ctx,
-            output={
-                "content_preview": content[:600],
-                "content_chars": len(content),
-                "reasoning_chars": len(reasoning),
-            },
+            output=response_payload,
             usage=stream_usage,
             metadata={
-                "finish_reason": processor.get_finish_reason(),
+                "finish_reason": finish_reason,
                 "status": "stop",
                 "tool_call_count": 0,
             },
-            trace_output={
-                "status": "ok",
-                "next_action": "stop",
-                "finish_reason": processor.get_finish_reason(),
-            },
+            trace_output=response_payload,
         )
         return StepResult(action=result_action, content=content, usage=stream_usage)
     
diff --git a/flocks/session/session.py b/flocks/session/session.py
index b83d898c2..18ab7c2df 100644
--- a/flocks/session/session.py
+++ b/flocks/session/session.py
@@ -5,13 +5,17 @@
 Based on Flocks' ported src/session/index.ts
 """
 
+import asyncio
 import contextvars
 import re
-from typing import List, Dict, Any, Optional
+import weakref
+from contextlib import AsyncExitStack, asynccontextmanager
+from typing import AsyncIterator, Awaitable, Callable, List, Dict, Any, Optional, TypeVar
 from datetime import datetime
 from pydantic import BaseModel, Field, ConfigDict
 
 from flocks.auth.context import (
+    API_TOKEN_SERVICE_USER_ID,
     get_current_auth_user,
     reset_current_auth_user,
     set_current_auth_user,
@@ -19,7 +23,7 @@
 from flocks.storage.storage import Storage
 from flocks.utils.log import Log
 from flocks.utils.id import Identifier
-from flocks.session.message import Message, MessageInfo, AssistantMessageInfo
+from flocks.session.message import Message, MessageInfo, MessageRole, AssistantMessageInfo
 
 # Sentinel for explicitly setting a field to None via Session.update()
 _UNSET = object()
@@ -30,6 +34,21 @@
 # Title prefix patterns for default title detection
 PARENT_TITLE_PREFIX = "New session - "
 CHILD_TITLE_PREFIX = "Child session - "
+MODEL_AUTO_SESSION_CATEGORIES = frozenset({"user", "entity-config", "workflow"})
+_WriteResult = TypeVar("_WriteResult")
+
+
+class SessionNotFoundError(ValueError):
+    """Raised when a lifecycle-aware write cannot find its session."""
+
+
+class SessionInactiveError(RuntimeError):
+    """Raised when a lifecycle-aware write targets a non-active session."""
+
+
+def is_model_auto_session_category(category: Optional[str]) -> bool:
+    """Return whether a session category may use WebUI Auto mode."""
+    return category in MODEL_AUTO_SESSION_CATEGORIES
 
 
 class SessionChangeStats(BaseModel):
@@ -94,6 +113,13 @@ class SessionInfo(BaseModel):
             "Unpinned sessions follow the normal default-model resolution chain."
         ),
     )
+    model_auto: bool = Field(
+        False,
+        description=(
+            "Whether WebUI Auto runtime failover was explicitly selected for "
+            "this session. Other session entry points ignore this flag."
+        ),
+    )
     
     # Session hierarchy
     parent_id: Optional[str] = Field(None, alias="parentID", description="Parent session for branching")
@@ -141,6 +167,15 @@ class Session:
     _id_index: Dict[str, str] = {}
     # Hot-path cache for repeatedly listing sessions in the UI.
     _all_sessions_cache: Optional[List[SessionInfo]] = None
+    # Serializes tree topology changes only. Ordinary writes use a keyed lock,
+    # so unrelated sessions no longer block each other.
+    _tree_lock = asyncio.Lock()
+    _lifecycle_locks: weakref.WeakValueDictionary[str, asyncio.Lock] = (
+        weakref.WeakValueDictionary()
+    )
+    _lifecycle_transition_ids: set[str] = set()
+    _lifecycle_generations: Dict[str, int] = {}
+    _active_operation_counts: Dict[str, int] = {}
 
     @staticmethod
     def _sort_sessions(sessions: List[SessionInfo]) -> List[SessionInfo]:
@@ -175,12 +210,127 @@ def _sync_list_cache(cls, session: SessionInfo) -> None:
             remaining.append(session)
         cls._all_sessions_cache = cls._sort_sessions(remaining)
 
+    @classmethod
+    def _remove_from_list_cache(cls, session_id: str) -> None:
+        """Remove a permanently deleted session from derived caches."""
+        if cls._all_sessions_cache is not None:
+            cls._all_sessions_cache = [
+                session
+                for session in cls._all_sessions_cache
+                if session.id != session_id
+            ]
+
     @classmethod
     def invalidate_cache(cls) -> None:
         """Clear in-memory indexes when the underlying storage changes."""
         cls._id_index.clear()
         cls._all_sessions_cache = None
 
+    @classmethod
+    def is_lifecycle_transitioning(cls, session_id: str) -> bool:
+        """Return whether archive/delete currently owns this session lifecycle."""
+        return session_id in cls._lifecycle_transition_ids
+
+    @classmethod
+    def lifecycle_generation(cls, session_id: str) -> int:
+        """Return a token that changes whenever a lifecycle transition begins."""
+        return cls._lifecycle_generations.get(session_id, 0)
+
+    @classmethod
+    def lifecycle_lock(cls, session_id: str) -> asyncio.Lock:
+        """Return the lock that linearizes durable writes for one session."""
+
+        lock = cls._lifecycle_locks.get(session_id)
+        if lock is None:
+            lock = asyncio.Lock()
+            cls._lifecycle_locks[session_id] = lock
+        return lock
+
+    @classmethod
+    def has_active_operations(cls, session_id: str) -> bool:
+        """Return whether a synchronous operation currently owns this session."""
+
+        return cls._active_operation_counts.get(session_id, 0) > 0
+
+    @classmethod
+    @asynccontextmanager
+    async def active_operation(cls, session_id: str) -> AsyncIterator[SessionInfo]:
+        """Prevent lifecycle transitions while a synchronous operation is running."""
+
+        async with cls.lifecycle_lock(session_id):
+            session = await cls.get_by_id_unfiltered(session_id)
+            if session is None:
+                raise SessionNotFoundError(f"Session {session_id} not found")
+            if session.status != "active" or cls.is_lifecycle_transitioning(session_id):
+                raise SessionInactiveError(f"Session {session_id} is not active")
+            cls._active_operation_counts[session_id] = (
+                cls._active_operation_counts.get(session_id, 0) + 1
+            )
+
+        try:
+            yield session
+        finally:
+            remaining = cls._active_operation_counts.get(session_id, 0) - 1
+            if remaining > 0:
+                cls._active_operation_counts[session_id] = remaining
+            else:
+                cls._active_operation_counts.pop(session_id, None)
+
+    @classmethod
+    async def get_by_id_unfiltered(cls, session_id: str) -> Optional[SessionInfo]:
+        """Get a session by ID without applying the ambient auth policy."""
+
+        token = set_current_auth_user(None)
+        try:
+            return await cls.get_by_id(session_id)
+        finally:
+            reset_current_auth_user(token)
+
+    @classmethod
+    async def run_active_write(
+        cls,
+        session_id: str,
+        operation: Callable[[], Awaitable[_WriteResult]],
+        *,
+        expected_generation: Optional[int] = None,
+    ) -> _WriteResult:
+        """Run one durable write atomically with lifecycle transitions."""
+
+        async with cls.lifecycle_lock(session_id):
+            session = await cls.get_by_id_unfiltered(session_id)
+            if session is None:
+                raise SessionNotFoundError(f"Session {session_id} not found")
+            generation_changed = (
+                expected_generation is not None
+                and cls.lifecycle_generation(session_id) != expected_generation
+            )
+            if (
+                session.status != "active"
+                or cls.is_lifecycle_transitioning(session_id)
+                or generation_changed
+            ):
+                raise SessionInactiveError(f"Session {session_id} is not active")
+            return await operation()
+
+    @classmethod
+    async def _clear_project_move_metadata_locked(cls, session_id: str) -> bool:
+        """Remove a stale replay boundary while the session write lock is held."""
+
+        session = await cls.get_by_id_unfiltered(session_id)
+        if session is None:
+            raise SessionNotFoundError(f"Session {session_id} not found")
+        metadata = dict(getattr(session, "metadata", {}) or {})
+        if "projectMove" not in metadata:
+            return False
+
+        metadata.pop("projectMove")
+        updated_session = session.model_copy(update={"metadata": metadata})
+        storage_key = f"session:{session.project_id}:{session.id}"
+        await Storage.set(storage_key, updated_session, "session")
+        cls._id_index[session.id] = storage_key
+        cls._sync_list_cache(updated_session)
+        return True
+
     @staticmethod
     def has_pinned_model(session: Optional[SessionInfo]) -> bool:
         """Return whether a session has an explicit model lock."""
@@ -198,11 +348,21 @@ def explicit_model_updates(provider_id: str, model_id: str) -> Dict[str, Any]:
             "provider": provider_id,
             "model": model_id,
             "model_pinned": True,
+            "model_auto": False,
         }
 
     @classmethod
     def inherited_model_kwargs(cls, session: Optional[SessionInfo]) -> Dict[str, Any]:
-        """Return pinned model kwargs that should propagate to a child session."""
+        """Return model preference kwargs that should propagate to a new session."""
+        if (
+            session
+            and is_model_auto_session_category(getattr(session, "category", "user"))
+            and getattr(session, "model_auto", False)
+        ):
+            return {
+                "model_auto": True,
+                "model_pinned": False,
+            }
         if not cls.has_pinned_model(session):
             return {}
         return {
@@ -278,27 +438,77 @@ async def create(
             except Exception as e:
                 log.warn("session.memory.default.error", {"error": str(e)})
 
-        # Bind ownership from current auth context unless explicitly provided.
-        if "owner_user_id" not in kwargs or "owner_username" not in kwargs:
+        # Bind root ownership here; children inherit ownership from their parent below.
+        if parent_id is None and (
+            "owner_user_id" not in kwargs or "owner_username" not in kwargs
+        ):
             current_user = get_current_auth_user()
             if current_user:
                 kwargs.setdefault("owner_user_id", current_user.id)
                 kwargs.setdefault("owner_username", current_user.username)
+            else:
+                kwargs.setdefault("owner_user_id", API_TOKEN_SERVICE_USER_ID)
+                kwargs.setdefault("owner_username", API_TOKEN_SERVICE_USER_ID)
         
-        session = SessionInfo(
-            project_id=project_id,
-            directory=directory,
-            title=title or cls._create_default_title(is_child),
-            parent_id=parent_id,
-            permission=permission,
-            **kwargs
-        )
-        
-        # Save to storage
-        storage_key = f"session:{project_id}:{session.id}"
-        await Storage.set(storage_key, session, "session")
-        cls._id_index[session.id] = storage_key
-        cls._sync_list_cache(session)
+        async def persist(parent: Optional[SessionInfo] = None) -> SessionInfo:
+            if parent_id is not None:
+                if parent is None:
+                    raise ValueError(f"Parent session {parent_id} not found")
+                child_owner_id = kwargs.get("owner_user_id")
+                child_owner_username = kwargs.get("owner_username")
+                if parent.owner_user_id and child_owner_id and parent.owner_user_id != child_owner_id:
+                    raise ValueError("Child session owner must match its parent")
+                if parent.owner_username and child_owner_username and parent.owner_username != child_owner_username:
+                    raise ValueError("Child session owner must match its parent")
+                if child_owner_id is None:
+                    kwargs["owner_user_id"] = parent.owner_user_id
+                if child_owner_username is None:
+                    kwargs["owner_username"] = parent.owner_username
+
+            created = SessionInfo(
+                project_id=project_id,
+                directory=directory,
+                title=title or cls._create_default_title(is_child),
+                parent_id=parent_id,
+                permission=permission,
+                **kwargs
+            )
+
+            # The parent status check and child row creation share the archive
+            # lock, so an archive cannot commit between them in this process.
+            storage_key = f"session:{project_id}:{created.id}"
+            await Storage.set(storage_key, created, "session")
+            cls._id_index[created.id] = storage_key
+            cls._sync_list_cache(created)
+            return created
+
+        from flocks.project.project import Project, ProjectDeletionError
+
+        # Every creation path participates in the project lifecycle claim. If
+        # deletion wins the race, a waiting creator observes the removed marker
+        # and cannot recreate an orphan task under the hidden project.
+        async with Project.lifecycle_guard(project_id):
+            if project_id.startswith("prj_") and Project.is_removed(project_id):
+                raise ProjectDeletionError(f"Project {project_id} is no longer available")
+
+            if parent_id is None:
+                session = await persist()
+            else:
+                async with cls._tree_lock:
+                    async with cls.lifecycle_lock(parent_id):
+                        parent = await Storage.get(
+                            f"session:{project_id}:{parent_id}",
+                            SessionInfo,
+                        )
+                        if parent is None or parent.status == "deleted":
+                            raise ValueError(f"Parent session {parent_id} not found")
+                        if cls.is_lifecycle_transitioning(parent_id):
+                            raise ValueError(
+                                f"Parent session {parent_id} is changing lifecycle state"
+                            )
+                        if parent.status != "active":
+                            raise ValueError(f"Parent session {parent_id} is not active")
+                        session = await persist(parent)
 
         try:
             from flocks.agent.registry import Agent
@@ -506,6 +716,8 @@ async def update(
         cls,
         project_id: str,
         session_id: str,
+        *,
+        allow_inactive: bool = False,
         **updates
     ) -> Optional[SessionInfo]:
         """
@@ -519,58 +731,55 @@ async def update(
         Returns:
             Updated session info or None
         """
-        session = await cls.get(project_id, session_id)
-        if not session:
-            return None
-        
-        # Field alias mapping (snake_case -> camelCase)
-        alias_map = {
-            "project_id": "projectID",
-            "parent_id": "parentID",
-            "owner_user_id": "ownerUserID",
-            "owner_username": "ownerUsername",
-        }
-        
-        # Update fields.
-        # Use ``_UNSET`` sentinel to explicitly set a field to None
-        # (plain ``None`` is skipped to preserve existing values).
-        update_data = session.model_dump(by_alias=True)
-        for key, value in updates.items():
-            # Sentinel means "explicitly clear this field"
-            if value is _UNSET:
-                alias_key = alias_map.get(key, key)
-                if alias_key in update_data:
-                    update_data[alias_key] = None
-                elif key in update_data:
-                    update_data[key] = None
-                continue
+        async with cls.lifecycle_lock(session_id):
+            if cls.is_lifecycle_transitioning(session_id):
+                return None
+            session = await cls.get(project_id, session_id)
+            if not session or (session.status != "active" and not allow_inactive):
+                return None
 
-            if value is not None:
-                # Handle nested updates for summary, revert, time
-                if key == "summary" and isinstance(value, dict):
-                    if update_data.get("summary"):
-                        update_data["summary"].update(value)
-                    else:
-                        update_data["summary"] = value
-                elif key == "revert" and isinstance(value, dict):
-                    update_data["revert"] = value
-                else:
-                    # Check both original key and aliased key
+            alias_map = {
+                "project_id": "projectID",
+                "parent_id": "parentID",
+                "owner_user_id": "ownerUserID",
+                "owner_username": "ownerUsername",
+            }
+
+            # Use ``_UNSET`` to explicitly clear a field; ordinary ``None``
+            # remains a no-op for backward compatibility.
+            update_data = session.model_dump(by_alias=True)
+            for key, value in updates.items():
+                if value is _UNSET:
                     alias_key = alias_map.get(key, key)
                     if alias_key in update_data:
-                        update_data[alias_key] = value
+                        update_data[alias_key] = None
                     elif key in update_data:
-                        update_data[key] = value
-        
-        # Update timestamp
-        if "time" not in update_data:
-            update_data["time"] = {}
-        update_data["time"]["updated"] = int(datetime.now().timestamp() * 1000)
-        
-        updated_session = SessionInfo(**update_data)
-        await Storage.set(f"session:{project_id}:{session_id}", updated_session, "session")
-        cls._id_index[session_id] = f"session:{project_id}:{session_id}"
-        cls._sync_list_cache(updated_session)
+                        update_data[key] = None
+                    continue
+
+                if value is not None:
+                    if key == "summary" and isinstance(value, dict):
+                        if update_data.get("summary"):
+                            update_data["summary"].update(value)
+                        else:
+                            update_data["summary"] = value
+                    elif key == "revert" and isinstance(value, dict):
+                        update_data["revert"] = value
+                    else:
+                        alias_key = alias_map.get(key, key)
+                        if alias_key in update_data:
+                            update_data[alias_key] = value
+                        elif key in update_data:
+                            update_data[key] = value
+
+            if "time" not in update_data:
+                update_data["time"] = {}
+            update_data["time"]["updated"] = int(datetime.now().timestamp() * 1000)
+
+            updated_session = SessionInfo(**update_data)
+            await Storage.set(f"session:{project_id}:{session_id}", updated_session, "session")
+            cls._id_index[session_id] = f"session:{project_id}:{session_id}"
+            cls._sync_list_cache(updated_session)
         
         log.info("session.updated", {
             "id": session_id,
@@ -578,11 +787,125 @@ async def update(
         })
         
         return updated_session
+
+    @classmethod
+    async def move_to_project(
+        cls,
+        source_project_id: str,
+        session_id: str,
+        *,
+        target_project_id: str,
+        target_directory: str,
+        target_owner_id: Optional[str] = None,
+        additional_busy_check: Optional[Callable[[str], bool]] = None,
+    ) -> Optional[SessionInfo]:
+        """Move a complete active root session tree to another project."""
+
+        if source_project_id == target_project_id:
+            return await cls.get(source_project_id, session_id)
+
+        from flocks.project.project import Project
+
+        async with AsyncExitStack() as project_guards:
+            for project_id in sorted({source_project_id, target_project_id}):
+                await project_guards.enter_async_context(Project.lifecycle_guard(project_id))
+
+            if (
+                target_owner_id is not None
+                and Project.registry_state(target_project_id, owner_id=target_owner_id) != "active"
+            ):
+                return None
+
+            sessions = await cls._begin_lifecycle_transition(
+                source_project_id,
+                session_id,
+                require_root=True,
+            )
+            if not sessions:
+                return None
+
+            try:
+                if any(session.status != "active" for session in sessions):
+                    return None
+                if any(session.revert is not None for session in sessions):
+                    return None
+
+                from flocks.session.session_loop import SessionLoop
+
+                if any(
+                    SessionLoop.is_running(session.id)
+                    or (
+                        additional_busy_check is not None
+                        and additional_busy_check(session.id)
+                    )
+                    for session in sessions
+                ):
+                    return None
+
+                updated_at = int(datetime.now().timestamp() * 1000)
+                move_boundaries: Dict[str, Optional[str]] = {}
+                for session in sessions:
+                    messages = await Message.list(session.id, include_archived=True)
+                    move_boundaries[session.id] = messages[-1].id if messages else None
+
+                moved_sessions = [
+                    session.model_copy(update={
+                        "project_id": target_project_id,
+                        "directory": target_directory,
+                        "metadata": {
+                            **(
+                                session.metadata
+                                if isinstance(session.metadata, dict)
+                                else {}
+                            ),
+                            "projectMove": {
+                                "sourceProjectID": source_project_id,
+                                "targetProjectID": target_project_id,
+                                "boundaryMessageID": move_boundaries[session.id],
+                                "movedAt": updated_at,
+                            },
+                        },
+                        "time": session.time.model_copy(update={"updated": updated_at}),
+                    })
+                    for session in sessions
+                ]
+                await Storage.mutate_many(
+                    set_entries=[
+                        (f"session:{target_project_id}:{session.id}", session, "session")
+                        for session in moved_sessions
+                    ],
+                    delete_keys=[
+                        f"session:{source_project_id}:{session.id}"
+                        for session in sessions
+                    ],
+                )
+                for session in moved_sessions:
+                    cls._id_index[session.id] = f"session:{target_project_id}:{session.id}"
+                    cls._sync_list_cache(session)
+
+                log.info("session.project_moved", {
+                    "id": session_id,
+                    "source_project_id": source_project_id,
+                    "target_project_id": target_project_id,
+                    "affected_sessions": len(moved_sessions),
+                })
+                return moved_sessions[0]
+            finally:
+                await cls._end_lifecycle_transition(sessions)
     
     @classmethod
     async def delete(cls, project_id: str, session_id: str) -> bool:
+        """Permanently delete a session tree and all application-owned data."""
+
+        from flocks.project.project import Project
+
+        async with Project.lifecycle_guard(project_id):
+            return await cls._delete_locked(project_id, session_id)
+
+    @classmethod
+    async def _delete_locked(cls, project_id: str, session_id: str) -> bool:
         """
-        Delete a session (soft delete)
+        Permanently delete a session tree and all application-owned data.
         
         Also deletes child sessions.
         
@@ -593,52 +916,109 @@ async def delete(cls, project_id: str, session_id: str) -> bool:
         Returns:
             True if deleted
         """
-        session = await cls.get(project_id, session_id)
-        if not session:
+        sessions = await cls._begin_lifecycle_transition(project_id, session_id)
+        if not sessions:
             return False
-        
-        # Delete child sessions first
-        children = await cls.children(project_id, session_id)
-        for child in children:
-            await cls.delete(project_id, child.id)
-
-        # Soft delete
-        await cls.update(project_id, session_id, status="deleted")
-        cls._id_index.pop(session_id, None)
-        
-        # Clear messages
-        await Message.clear(session_id)
+        deleted = False
         try:
-            from flocks.session.callable_state import clear_session_callable_tools
+            if not await cls._stop_session_tree_for_archive(sessions):
+                return False
+            for session in sessions:
+                await Message.quiesce_parts(session.id, persist=False)
 
-            await clear_session_callable_tools(session_id)
-        except Exception as e:
-            log.warn("session.callable_tools.clear_error", {"id": session_id, "error": str(e)})
-        
-        log.info("session.deleted", {
-            "id": session_id,
-            "project_id": project_id,
-        })
+            refreshed = await cls.collect_tree(project_id, session_id)
+            if not refreshed:
+                return False
+            sessions = refreshed
+            session_ids = [session.id for session in sessions]
+
+            from flocks.permission.next import PermissionNext
+
+            permission_keys = await PermissionNext.deletion_storage_keys(session_ids)
+            await Storage.mutate_many(
+                delete_keys=[
+                    key
+                    for session in sessions
+                    for key in (
+                        f"session:{session.project_id}:{session.id}",
+                        f"message:{session.id}",
+                        f"message_parts:{session.id}",
+                        f"todo:{session.id}",
+                        f"goal:{session.id}",
+                        f"session_diff:{session.id}",
+                        f"session_callable_tools:{session.id}",
+                    )
+                ] + permission_keys,
+                delete_prefixes=[
+                    prefix
+                    for session in sessions
+                    for prefix in (
+                        f"message_parts:{session.id}:",
+                        f"message_diff:{session.id}:",
+                        f"system_prompts:{session.id}:",
+                    )
+                ],
+            )
+            PermissionNext.clear_session_runtime(session_ids)
 
-        # Flocks compatibility: clear state and publish event
-        try:
-            from flocks.session.core.session_state import get_main_session_id, set_main_session, remove_subagent_session
-            if get_main_session_id() == session_id:
-                set_main_session(None)
-            remove_subagent_session(session_id)
-        except Exception as e:
-            log.warn("session.state.error", {"error": str(e)})
+            from flocks.session.session_loop import SessionLoop
+            from flocks.session.callable_state import invalidate_session_callable_tools_cache
+            from flocks.project.project import Project
 
-        try:
-            from flocks.bus.bus import Bus
-            from flocks.bus.events import SessionDeleted
-            await Bus.publish(SessionDeleted, {
-                "sessionID": session_id,
+            Project.invalidate_session_stats()
+
+            for session in sessions:
+                cls._id_index.pop(session.id, None)
+                cls._remove_from_list_cache(session.id)
+                SessionLoop.clear_auto_failover_state(session.id)
+                Message.invalidate_cache(session.id)
+                invalidate_session_callable_tools_cache(session.id)
+
+                try:
+                    from flocks.session.files import remove_session_uploads
+
+                    if await asyncio.to_thread(remove_session_uploads, session.id):
+                        log.info("session.uploads.cleaned", {"session_id": session.id})
+                except Exception as exc:
+                    log.warn("session.uploads.cleanup_failed", {
+                        "session_id": session.id,
+                        "error": str(exc),
+                    })
+
+            for session in reversed(sessions):
+                try:
+                    from flocks.session.core.session_state import (
+                        get_main_session_id,
+                        remove_subagent_session,
+                        set_main_session,
+                    )
+
+                    if get_main_session_id() == session.id:
+                        set_main_session(None)
+                    remove_subagent_session(session.id)
+                except Exception as e:
+                    log.warn("session.state.error", {"id": session.id, "error": str(e)})
+
+                try:
+                    from flocks.bus.bus import Bus
+                    from flocks.bus.events import SessionDeleted
+
+                    await Bus.publish(SessionDeleted, {"sessionID": session.id})
+                except Exception as e:
+                    log.warn("session.deleted.event_error", {"id": session.id, "error": str(e)})
+
+            log.info("session.deleted", {
+                "id": session_id,
+                "project_id": project_id,
+                "count": len(sessions),
             })
-        except Exception as e:
-            log.warn("session.deleted.event_error", {"error": str(e)})
-        
-        return True
+            deleted = True
+            return True
+        finally:
+            await cls._end_lifecycle_transition(sessions)
+            if deleted:
+                for session in sessions:
+                    cls._lifecycle_generations.pop(session.id, None)
 
     @classmethod
     async def retain_deleted_user_sessions(cls, user_id: str, username: str) -> int:
@@ -659,15 +1039,89 @@ async def retain_deleted_user_sessions(cls, user_id: str, username: str) -> int:
             await cls.update(
                 project_id=session.project_id,
                 session_id=session.id,
+                allow_inactive=True,
                 owner_user_id=_UNSET,
                 owner_username=username,
             )
             migrated += 1
 
         return migrated
+
+    @classmethod
+    async def _stop_session_tree_for_archive(
+        cls,
+        sessions: List[SessionInfo],
+        *,
+        timeout_s: float = 5.0,
+        clear_prompt_queue: bool = True,
+    ) -> bool:
+        """Stop persisted and in-memory work before committing archive state."""
+        from flocks.session.interaction_queue import InteractionQueue
+        from flocks.session.runner import SessionRunner
+        from flocks.session.session_loop import SessionLoop
+
+        session_ids = [session.id for session in sessions]
+        for session_id in session_ids:
+            SessionLoop.abort(session_id)
+            SessionRunner.cancel(session_id)
+            if clear_prompt_queue:
+                await InteractionQueue.clear(session_id)
+            try:
+                from flocks.server.routes.question import reject_session_questions
+
+                await reject_session_questions(session_id)
+            except Exception as exc:
+                log.warn("session.archive.question_cleanup_error", {
+                    "id": session_id,
+                    "error": str(exc),
+                })
+            try:
+                from flocks.session.background_tasks import cancel_session_background_tasks
+
+                await cancel_session_background_tasks(session_id)
+            except Exception as exc:
+                log.warn("session.archive.background_task_cleanup_error", {
+                    "id": session_id,
+                    "error": str(exc),
+                })
+            try:
+                from flocks.task.background import get_background_manager
+
+                get_background_manager().cancel_by_parent_session_id(session_id)
+            except Exception as exc:
+                log.warn("session.archive.background_cleanup_error", {
+                    "id": session_id,
+                    "error": str(exc),
+                })
+
+        deadline = asyncio.get_running_loop().time() + timeout_s
+        pending = set(session_ids)
+        while pending:
+            running = {session_id for session_id in pending if SessionLoop.is_running(session_id)}
+            if not running:
+                return True
+            now = asyncio.get_running_loop().time()
+            if now >= deadline:
+                log.warn("session.archive.wait_idle_timeout", {
+                    "ids": sorted(running),
+                    "timeout_s": timeout_s,
+                })
+                return False
+            pending = running
+            await asyncio.sleep(min(0.05, deadline - now))
+        return True
     
     @classmethod
     async def archive(cls, project_id: str, session_id: str) -> bool:
+        """Archive a root session and its descendants."""
+
+        from flocks.project.project import Project
+
+        async with Project.lifecycle_guard(project_id):
+            return await cls._archive_locked(project_id, session_id)
+
+    @classmethod
+    async def _archive_locked(cls, project_id: str, session_id: str) -> bool:
         """
         Archive a session
         
@@ -678,23 +1132,111 @@ async def archive(cls, project_id: str, session_id: str) -> bool:
         Returns:
             True if archived
         """
-        session = await cls.get(project_id, session_id)
-        if not session:
+        sessions = await cls._begin_lifecycle_transition(
+            project_id,
+            session_id,
+            require_root=True,
+        )
+        if not sessions:
             return False
+        prompt_queue_paused = False
+        archive_committed = False
+        try:
+            if any(session.status != "archived" for session in sessions):
+                from flocks.session.interaction_queue import InteractionQueue
+
+                prompt_queue_paused = True
+                for session in sessions:
+                    await InteractionQueue.pause(session.id)
+                if not await cls._stop_session_tree_for_archive(
+                    sessions,
+                    clear_prompt_queue=False,
+                ):
+                    return False
+            for session in sessions:
+                await Message.quiesce_parts(session.id, persist=True)
+
+            # Lifecycle claims prevent new in-process children and writes. Read
+            # once more so every child committed before the claim is included.
+            refreshed = await cls.collect_tree(project_id, session_id)
+            if not refreshed:
+                return False
+            sessions = refreshed
+
+            archived_ts = int(datetime.now().timestamp() * 1000)
+            updated_sessions: List[SessionInfo] = []
+            for session in sessions:
+                if session.status == "archived":
+                    updated_sessions.append(session)
+                    continue
+                updated_sessions.append(session.model_copy(update={
+                    "status": "archived",
+                    "time": session.time.model_copy(update={
+                        "archived": session.time.archived or archived_ts,
+                    }),
+                }))
+
+            original_status = {session.id: session.status for session in sessions}
+            changed = [
+                session
+                for session in updated_sessions
+                if original_status[session.id] != session.status
+            ]
+            if changed:
+                await Storage.set_many([
+                    (f"session:{session.project_id}:{session.id}", session, "session")
+                    for session in changed
+                ])
+                for session in changed:
+                    cls._id_index[session.id] = f"session:{session.project_id}:{session.id}"
+                    cls._sync_list_cache(session)
+            archive_committed = True
 
-        archived_ts = int(datetime.now().timestamp() * 1000)
-        time_data = session.time.model_dump()
-        time_data["archived"] = archived_ts
-        await cls.update(project_id, session_id, status="archived", time=time_data)
+            try:
+                from flocks.session.session_loop import SessionLoop
+                from flocks.session.core.session_state import (
+                    get_main_session_id,
+                    remove_subagent_session,
+                    set_main_session,
+                )
+
+                for session in sessions:
+                    SessionLoop.clear_auto_failover_state(session.id)
+                    Message.invalidate_cache(session.id)
+                    if get_main_session_id() == session.id:
+                        set_main_session(None)
+                    remove_subagent_session(session.id)
+            except Exception as exc:
+                log.warn("session.archive.runtime_cleanup_error", {
+                    "id": session_id,
+                    "error": str(exc),
+                })
+
+            log.info("session.archived", {
+                "id": session_id,
+                "project_id": project_id,
+                "count": len(sessions),
+            })
+            return True
+        finally:
+            if prompt_queue_paused and not archive_committed:
+                from flocks.session.interaction_queue import InteractionQueue
 
-        log.info("session.archived", {
-            "id": session_id,
-            "project_id": project_id,
-        })
-        return True
+                for session in sessions:
+                    await InteractionQueue.resume(session.id)
+            await cls._end_lifecycle_transition(sessions)
     
     @classmethod
     async def unarchive(cls, project_id: str, session_id: str) -> bool:
+        """Restore an archived root session and its descendants."""
+
+        from flocks.project.project import Project
+
+        async with Project.lifecycle_guard(project_id):
+            return await cls._unarchive_locked(project_id, session_id)
+
+    @classmethod
+    async def _unarchive_locked(cls, project_id: str, session_id: str) -> bool:
         """
         Restore an archived session
         
@@ -705,27 +1247,174 @@ async def unarchive(cls, project_id: str, session_id: str) -> bool:
         Returns:
             True if restored
         """
-        session = await cls.get(project_id, session_id)
-        if not session:
-            # get() filters deleted; manually check archived
-            raw = await Storage.get(f"session:{project_id}:{session_id}", SessionInfo)
-            if not raw or raw.status != "archived":
-                return False
-            session = raw
-
-        if session.status != "archived":
+        sessions = await cls._begin_lifecycle_transition(
+            project_id,
+            session_id,
+            require_root=True,
+        )
+        if not sessions:
             return False
+        try:
+            changed = [
+                session.model_copy(update={
+                    "status": "active",
+                    "time": session.time.model_copy(update={
+                        "archived": None,
+                    }),
+                })
+                for session in sessions
+                if session.status == "archived"
+            ]
+            if changed:
+                await Storage.set_many([
+                    (f"session:{session.project_id}:{session.id}", session, "session")
+                    for session in changed
+                ])
+                for session in changed:
+                    cls._id_index[session.id] = f"session:{session.project_id}:{session.id}"
+                    cls._sync_list_cache(session)
+
+            from flocks.session.interaction_queue import InteractionQueue
+
+            for session in sessions:
+                await InteractionQueue.resume(session.id)
+
+            log.info("session.unarchived", {
+                "id": session_id,
+                "project_id": project_id,
+                "count": len(sessions),
+            })
+            return True
+        finally:
+            await cls._end_lifecycle_transition(sessions)
 
-        time_data = session.time.model_dump()
-        time_data["archived"] = None
-        await cls.update(project_id, session_id, status="active", time=time_data)
-        
-        log.info("session.unarchived", {
-            "id": session_id,
-            "project_id": project_id,
-        })
-        
-        return True
+    @classmethod
+    async def restore(
+        cls,
+        project_id: str,
+        session_id: str,
+        *,
+        project_owner_id: Optional[str],
+    ) -> bool:
+        """Restore a session tree and its removed project as one operation."""
+
+        from flocks.project.project import (
+            DEFAULT_PROJECT_ID,
+            Project,
+            ProjectDeletionError,
+            TASK_SESSION_GROUP_ID,
+        )
+
+        async with Project.lifecycle_guard(project_id):
+            project_state = (
+                Project.registry_state(project_id, owner_id=project_owner_id)
+                if project_owner_id
+                else (
+                    "virtual"
+                    if project_id in {DEFAULT_PROJECT_ID, TASK_SESSION_GROUP_ID}
+                    else "missing"
+                )
+            )
+            if project_state == "missing" and project_id.startswith("prj_"):
+                raise ProjectDeletionError(
+                    f"Project {project_id} restoration metadata is unavailable"
+                )
+
+            project_was_removed = project_state == "removed"
+            if project_state in {"active", "removed"} and project_owner_id is not None:
+                await Project.restore(project_id, owner_id=project_owner_id)
+
+            try:
+                restored = await cls.unarchive(project_id, session_id)
+            except BaseException:
+                if project_was_removed and project_owner_id is not None:
+                    try:
+                        await asyncio.shield(
+                            Project._delete_locked(project_id, owner_id=project_owner_id)
+                        )
+                    except Exception as rollback_exc:
+                        log.error("session.restore.project_rollback_failed", {
+                            "project_id": project_id,
+                            "error": str(rollback_exc),
+                        })
+                raise
+
+            if not restored and project_was_removed and project_owner_id is not None:
+                await Project.delete(project_id, owner_id=project_owner_id)
+            return restored
+
+    @classmethod
+    async def collect_tree(cls, project_id: str, session_id: str) -> List[SessionInfo]:
+        """Return a root session and all descendants in parent-first order."""
+        token = set_current_auth_user(None)
+        try:
+            root = await cls.get(project_id, session_id)
+        finally:
+            reset_current_auth_user(token)
+        if root is None:
+            return []
+
+        children_by_parent: Dict[str, List[SessionInfo]] = {}
+        for session in await cls.list_all_unfiltered():
+            if session.project_id != project_id:
+                continue
+            if session.parent_id:
+                children_by_parent.setdefault(session.parent_id, []).append(session)
+
+        tree: List[SessionInfo] = []
+        seen: set[str] = set()
+
+        def visit(session: SessionInfo) -> None:
+            if session.id in seen:
+                return
+            seen.add(session.id)
+            tree.append(session)
+            for child in children_by_parent.get(session.id, []):
+                visit(child)
+
+        visit(root)
+        return tree
+
+    @classmethod
+    async def _begin_lifecycle_transition(
+        cls,
+        project_id: str,
+        session_id: str,
+        *,
+        require_root: bool = False,
+    ) -> List[SessionInfo]:
+        """Claim a complete tree so ordinary writes cannot overtake its transition."""
+        async with cls._tree_lock:
+            sessions = await cls.collect_tree(project_id, session_id)
+            if not sessions:
+                return []
+            if require_root and sessions[0].parent_id is not None:
+                log.warn("session.lifecycle.non_root_rejected", {"id": session_id})
+                return []
+
+            async with AsyncExitStack() as locks:
+                for item_id in sorted(session.id for session in sessions):
+                    await locks.enter_async_context(cls.lifecycle_lock(item_id))
+
+                # A child may have committed before the topology lock was
+                # acquired. Re-read under all known keyed locks before claiming.
+                sessions = await cls.collect_tree(project_id, session_id)
+                ids = {session.id for session in sessions}
+                if (
+                    not sessions
+                    or ids & cls._lifecycle_transition_ids
+                    or any(cls.has_active_operations(item_id) for item_id in ids)
+                ):
+                    return []
+                cls._lifecycle_transition_ids.update(ids)
+                for item_id in ids:
+                    cls._lifecycle_generations[item_id] = cls.lifecycle_generation(item_id) + 1
+                return sessions
+
+    @classmethod
+    async def _end_lifecycle_transition(cls, sessions: List[SessionInfo]) -> None:
+        async with cls._tree_lock:
+            cls._lifecycle_transition_ids.difference_update(session.id for session in sessions)
 
     @classmethod
     async def children(cls, project_id: str, parent_id: str) -> List[SessionInfo]:
@@ -760,70 +1449,92 @@ async def fork(
         Returns:
             New forked session
         """
-        # Get original session
-        original = await cls.get(project_id, session_id)
-        if not original:
-            raise ValueError(f"Session {session_id} not found")
-        
-        # Create new session with parent_id set
-        new_session = await cls.create(
-            project_id=project_id,
-            directory=original.directory,
-        )
-        
-        # Update parent_id after creation
-        new_session = await cls.update(
-            project_id=project_id,
-            session_id=new_session.id,
-            parent_id=session_id,
-        )
-        
-        # Copy messages with all parts (include archived so fork preserves full history)
-        messages = await Message.list(session_id, include_archived=True)
-        id_map: Dict[str, str] = {}
-        
-        for msg in messages:
-            if message_id and msg.id >= message_id:
-                break
-            
-            new_id = Identifier.ascending("message")
-            id_map[msg.id] = new_id
-            
-            # Get text content for the initial message creation
-            content = await Message.get_text_content(msg)
-            parent_ref = None
-            if isinstance(msg, AssistantMessageInfo):
-                parent_ref = id_map.get(msg.parentID)
-            
-            # Create the message (this also creates an initial TextPart)
-            await Message.create(
-                session_id=new_session.id,
-                role=msg.role,
-                content=content,
-                id=new_id,
-                parentID=parent_ref or "",
+        from flocks.project.project import Project
+
+        # Keep the parent, new child, and copied history in one project-lifecycle
+        # interval. A concurrent move can only happen before or after the fork.
+        async with Project.lifecycle_guard(project_id):
+            original = await cls.get(project_id, session_id)
+            if not original:
+                raise ValueError(f"Session {session_id} not found")
+
+            messages = await Message.list(session_id, include_archived=True)
+            messages_to_copy: List[MessageInfo] = []
+            for msg in messages:
+                if message_id and msg.id >= message_id:
+                    break
+                messages_to_copy.append(msg)
+            id_map = {
+                msg.id: Identifier.ascending("message")
+                for msg in messages_to_copy
+            }
+
+            fork_move_metadata: Optional[Dict[str, Any]] = None
+            move_metadata = original.metadata.get("projectMove")
+            if isinstance(move_metadata, dict) and move_metadata.get("boundaryMessageID"):
+                original_boundary = move_metadata["boundaryMessageID"]
+                mapped_boundary = id_map.get(original_boundary)
+                if mapped_boundary is None and messages_to_copy:
+                    # A fork ending before the move boundary contains only unsafe
+                    # imported history, so protect the complete copied prefix.
+                    mapped_boundary = id_map[messages_to_copy[-1].id]
+                if mapped_boundary is not None:
+                    fork_move_metadata = {
+                        **move_metadata,
+                        "boundaryMessageID": mapped_boundary,
+                    }
+
+            metadata = (
+                {"projectMove": fork_move_metadata}
+                if fork_move_metadata is not None
+                else None
             )
-            
-            # Copy non-text parts (tool calls, files, patches, etc.)
-            original_parts = await Message.parts(msg.id, session_id)
-            for part in original_parts:
-                if part.type == "text":
-                    continue  # Already created by Message.create above
-                # Clone part with updated session/message IDs
-                part_data = part.model_dump()
-                part_data["id"] = Identifier.ascending("part")
-                part_data["sessionID"] = new_session.id
-                part_data["messageID"] = new_id
-                cloned_part = Message.deserialize_part(part_data)
-                await Message.store_part(new_session.id, new_id, cloned_part)
-        
-        log.info("session.forked", {
-            "from": session_id,
-            "to": new_session.id,
-            "messages": len(id_map),
-        })
-        
-        return new_session
+            new_session = await cls.create(
+                project_id=project_id,
+                directory=original.directory,
+                parent_id=session_id,
+                **({"metadata": metadata} if metadata is not None else {}),
+            )
+
+            # Copy messages with all parts (include archived so fork preserves full history)
+            for msg in messages_to_copy:
+                new_id = id_map[msg.id]
+
+                # Get text content for the initial message creation
+                content = await Message.get_text_content(msg)
+                parent_ref = None
+                if isinstance(msg, AssistantMessageInfo):
+                    parent_ref = id_map.get(msg.parentID)
+
+                # Create the message (this also creates an initial TextPart)
+                await Message.create(
+                    session_id=new_session.id,
+                    role=MessageRole(msg.role),
+                    content=content,
+                    id=new_id,
+                    parentID=parent_ref or "",
+                )
+
+                # Copy non-text parts (tool calls, files, patches, etc.)
+                original_parts = await Message.parts(msg.id, session_id)
+                for part in original_parts:
+                    if part.type == "text":
+                        continue  # Already created by Message.create above
+                    # Clone part with updated session/message IDs
+                    part_data = part.model_dump()
+                    part_data["id"] = Identifier.ascending("part")
+                    part_data["sessionID"] = new_session.id
+                    part_data["messageID"] = new_id
+                    cloned_part = Message.deserialize_part(part_data)
+                    await Message.store_part(new_session.id, new_id, cloned_part)
+
+            log.info("session.forked", {
+                "from": session_id,
+                "to": new_session.id,
+                "messages": len(id_map),
+            })
+
+            return new_session
     
     @classmethod
     async def set_revert(
diff --git a/flocks/session/session_loop.py b/flocks/session/session_loop.py
index adb1e112e..d163d663b 100644
--- a/flocks/session/session_loop.py
+++ b/flocks/session/session_loop.py
@@ -13,6 +13,7 @@
 """
 
 import asyncio
+import hashlib
 import inspect
 import time
 from typing import Optional, List, Dict, Any, Callable, Awaitable
@@ -21,7 +22,11 @@
 
 from flocks.utils.log import Log
 from flocks.utils.id import Identifier
-from flocks.session.session import Session, SessionInfo
+from flocks.session.session import (
+    Session,
+    SessionInfo,
+    is_model_auto_session_category,
+)
 from flocks.session.message import Message, MessageInfo, MessageRole
 from flocks.session.core.status import SessionStatus, SessionStatusBusy, SessionStatusIdle
 from flocks.session.core.task_utils import fire_and_forget
@@ -47,6 +52,26 @@
 
 MAX_OVERFLOW_COMPACTION_ATTEMPTS = 3
 POST_COMPACTION_COOLDOWN_STEPS = 2
+RATE_LIMIT_COOLDOWN_SECONDS = 60.0
+CHAIN_EXHAUSTION_COOLDOWN_SECONDS = 5.0
+
+
+@dataclass(frozen=True)
+class RuntimeModel:
+    """Concrete provider/model candidate used by Auto failover."""
+
+    provider_id: str
+    model_id: str
+
+
+@dataclass
+class AutoFailoverCooldown:
+    """Process-local Hermes-style starting candidate cooldown."""
+
+    model: RuntimeModel
+    primary: RuntimeModel
+    expires_at: float
+    reason: str
 
 
 @dataclass
@@ -81,6 +106,16 @@ class LoopContext:
     # is what the upstream will actually bill us for on the next turn
     # (matches the "observed value wins" rule from docs/design/context-compaction-v2.md §B3).
     last_observed_prompt_tokens: int = 0
+    auto_failover: bool = False
+    # Entrypoint authorization is separate from persisted model_auto. Only a
+    # WebUI message route may set this bit; non-WebUI entrypoints use the default.
+    auto_failover_allowed: bool = False
+    model_candidates: List[RuntimeModel] = field(default_factory=list)
+    candidate_index: int = 0
+    turn_user_id: Optional[str] = None
+    turn_additional_context: Optional[str] = None
+    stop_hook_active: bool = False
+    session_start_pending: bool = False
 
     @property
     def trace_step(self) -> int:
@@ -120,6 +155,8 @@ class LoopResult:
     action: str  # "stop", "continue", "compact", "error", "queued"
     last_message: Optional[MessageInfo] = None
     error: Optional[str] = None
+    provider_id: Optional[str] = None
+    model_id: Optional[str] = None
     metadata: Dict[str, Any] = field(default_factory=dict)
 
 
@@ -137,6 +174,209 @@ class SessionLoop:
     
     # Active loop contexts by session ID
     _active_loops: Dict[str, LoopContext] = {}
+    _auto_failover_cooldowns: Dict[str, AutoFailoverCooldown] = {}
+
+    @classmethod
+    def clear_auto_failover_state(cls, session_id: str) -> None:
+        """Clear process-local routing state when WebUI Auto is disabled."""
+        cls._auto_failover_cooldowns.pop(session_id, None)
+
+    @classmethod
+    async def validate_runtime_model(
+        cls,
+        provider_id: str,
+        model_id: str,
+        *,
+        config: Optional[Any] = None,
+    ) -> tuple[bool, str]:
+        """Validate a configured LLM candidate without a network health probe."""
+        from flocks.config.config import Config
+        from flocks.provider.model_manager import get_model_manager
+        from flocks.provider.types import ModelType
+
+        Provider._ensure_initialized()
+        config = config or await Config.get()
+        if provider_id in (getattr(config, "disabled_providers", None) or []):
+            return False, "provider_disabled"
+        enabled_providers = getattr(config, "enabled_providers", None) or []
+        if enabled_providers and provider_id not in enabled_providers:
+            return False, "provider_disabled"
+        try:
+            await Provider.apply_config(config, provider_id=provider_id)
+        except Exception as exc:
+            log.warn("session.model.candidate_config_failed", {
+                "provider_id": provider_id,
+                "model_id": model_id,
+                "error": str(exc),
+            })
+            return False, "provider_config_error"
+
+        provider = Provider.get(provider_id)
+        if provider is None:
+            return False, "provider_not_found"
+
+        definition = get_model_manager().get_model(provider_id, model_id)
+        if definition is None:
+            return False, "model_not_found"
+        if getattr(definition, "model_type", None) != ModelType.LLM:
+            return False, "not_llm"
+
+        setting = get_model_manager().get_setting(provider_id, model_id)
+        if setting is not None and not setting.enabled:
+            return False, "model_disabled"
+        if not provider.is_configured():
+            return False, "provider_not_configured"
+        return True, "available"
+
+    @classmethod
+    async def _build_model_candidates(
+        cls,
+        primary: RuntimeModel,
+        *,
+        route_seed: str,
+        preferred: Optional[RuntimeModel] = None,
+    ) -> List[RuntimeModel]:
+        """Build a stable per-turn primary, same-provider, cross-provider chain."""
+        from flocks.config.config import Config
+        from flocks.provider.model_manager import get_model_manager
+        from flocks.provider.types import ModelType
+
+        config = await Config.get()
+        await Provider.apply_config(config)
+        definitions = get_model_manager().list_models(
+            model_type=ModelType.LLM,
+            enabled_only=True,
+        )
+        discovered = {
+            RuntimeModel(definition.provider_id, definition.id)
+            for definition in definitions
+        }
+        discovered.discard(primary)
+
+        same_provider: List[RuntimeModel] = []
+        other_providers: List[RuntimeModel] = []
+        for candidate in sorted(
+            discovered,
+            key=lambda item: (item.provider_id, item.model_id),
+        ):
+            available, reason = await cls.validate_runtime_model(
+                candidate.provider_id,
+                candidate.model_id,
+                config=config,
+            )
+            if not available:
+                log.debug("session.model.fallback_skipped", {
+                    "provider_id": candidate.provider_id,
+                    "model_id": candidate.model_id,
+                    "reason": reason,
+                })
+                continue
+
+            if candidate.provider_id == primary.provider_id:
+                same_provider.append(candidate)
+            else:
+                other_providers.append(candidate)
+
+        candidates = [primary]
+        for tier, pool in (
+            ("same_provider", same_provider),
+            ("other_provider", other_providers),
+        ):
+            if not pool:
+                continue
+            selected = (
+                preferred
+                if preferred is not None and preferred in pool
+                else cls._stable_candidate_choice(pool, route_seed, tier)
+            )
+            candidates.append(selected)
+        return candidates
+
+    @staticmethod
+    def _stable_candidate_choice(
+        candidates: List[RuntimeModel],
+        route_seed: str,
+        tier: str,
+    ) -> RuntimeModel:
+        """Choose pseudo-randomly without Python's process-randomized hash()."""
+        ordered = sorted(
+            candidates,
+            key=lambda item: (item.provider_id, item.model_id),
+        )
+        digest = hashlib.sha256(
+            f"{route_seed}\0{tier}".encode("utf-8")
+        ).digest()
+        index = int.from_bytes(digest[:8], "big") % len(ordered)
+        return ordered[index]
+
+    @classmethod
+    async def validate_auto_configuration(cls) -> tuple[bool, str]:
+        """Validate that a newly selected Auto mode has a usable chain."""
+        from flocks.config.config import Config
+
+        default_llm = await Config.resolve_default_llm()
+        if not default_llm:
+            return False, "default_model_missing"
+        primary = RuntimeModel(
+            default_llm["provider_id"],
+            default_llm["model_id"],
+        )
+        available, reason = await cls.validate_runtime_model(
+            primary.provider_id,
+            primary.model_id,
+        )
+        if not available:
+            return False, f"primary_{reason}"
+        return True, "available"
+
+    @classmethod
+    def _active_cooldown_model(
+        cls,
+        session_id: str,
+        primary: RuntimeModel,
+    ) -> Optional[RuntimeModel]:
+        """Return a still-valid cooldown target for the current primary."""
+        cooldown = cls._auto_failover_cooldowns.get(session_id)
+        if cooldown is None:
+            return None
+        if cooldown.expires_at <= time.monotonic() or cooldown.primary != primary:
+            cls._auto_failover_cooldowns.pop(session_id, None)
+            return None
+        return cooldown.model
+
+    @classmethod
+    def _cooldown_candidate_index(
+        cls,
+        session_id: str,
+        candidates: List[RuntimeModel],
+    ) -> int:
+        if not candidates:
+            return 0
+        cooldown_model = cls._active_cooldown_model(session_id, candidates[0])
+        if cooldown_model is None:
+            return 0
+        try:
+            return candidates.index(cooldown_model)
+        except ValueError:
+            cls._auto_failover_cooldowns.pop(session_id, None)
+            return 0
+
+    @classmethod
+    def _select_candidate(cls, ctx: LoopContext, index: int) -> None:
+        candidate = ctx.model_candidates[index]
+        ctx.candidate_index = index
+        ctx.provider_id = candidate.provider_id
+        ctx.model_id = candidate.model_id
+        ctx.session.provider = candidate.provider_id
+        ctx.session.model = candidate.model_id
+        # Prompt and model-capability caches are keyed in most places, but a
+        # fresh dict makes the runtime rebuild guarantee explicit. The tool
+        # loop guard is turn state rather than model state, so it must survive
+        # a provider switch to keep repeated-tool protection effective.
+        tool_loop_guard = ctx.runner_static_cache.get("tool_loop_guard")
+        ctx.runner_static_cache.clear()
+        if tool_loop_guard is not None:
+            ctx.runner_static_cache["tool_loop_guard"] = tool_loop_guard
     
     @classmethod
     def is_running(cls, session_id: str) -> bool:
@@ -268,7 +508,7 @@ async def _detect_queued_user_message(
         _session_id: str,
         post_messages: List[MessageInfo],
         current_user_id: str,
-        last_message: Optional[MessageInfo],
+        _last_message: Optional[MessageInfo],
     ) -> Optional[MessageInfo]:
         if not post_messages:
             return None
@@ -283,11 +523,11 @@ async def _detect_queued_user_message(
             return None
         if newest_user.id <= current_user_id:
             return None
-        if last_message is None:
-            return newest_user
-        if newest_user.id > last_message.id:
-            return newest_user
-        return None
+        # A fallback assistant is created after a user message that arrived
+        # while the primary model was running. Its newer ID must not make that
+        # user message look handled; the current turn's user ID is the stable
+        # boundary for queued work.
+        return newest_user
     
     @classmethod
     async def run(
@@ -298,6 +538,7 @@ async def run(
         agent_name: Optional[str] = None,
         callbacks: Optional[LoopCallbacks] = None,
         working_directory: Optional[str] = None,
+        auto_failover: bool = False,
     ) -> LoopResult:
         """
         Run session loop
@@ -327,6 +568,15 @@ async def run(
         # next iteration once it finishes the current step.
         if cls.is_running(session_id):
             log.info("loop.already_running", {"session_id": session_id})
+            if auto_failover:
+                active_ctx = cls._active_loops.get(session_id)
+                if (
+                    active_ctx is not None
+                    and is_model_auto_session_category(
+                        getattr(active_ctx.session, "category", "user")
+                    )
+                ):
+                    active_ctx.auto_failover_allowed = True
             return LoopResult(
                 action="queued",
                 error="Loop already running",
@@ -340,6 +590,15 @@ async def run(
                 action="error",
                 error=f"Session {session_id} not found",
             )
+        if session.status != "active":
+            log.warning("loop.session_not_active", {
+                "session_id": session_id,
+                "status": session.status,
+            })
+            return LoopResult(
+                action="error",
+                error=f"Session {session_id} is {session.status}",
+            )
         if working_directory:
             session = session.model_copy(update={"directory": working_directory})
         
@@ -351,6 +610,19 @@ async def run(
             provider_id = provider_id or resolved_provider
             model_id = model_id or resolved_model
         
+        primary_model = RuntimeModel(
+            provider_id=provider_id,
+            model_id=model_id,
+        )
+        model_candidates = [primary_model]
+        candidate_index = 0
+        auto_failover = bool(
+            auto_failover
+            and is_model_auto_session_category(
+                getattr(session, "category", "user")
+            )
+        )
+
         # Keep the in-memory session aligned with the runtime model so
         # downstream helpers (title generation, compaction checks, etc.) see
         # the model actually selected for this loop iteration. Unpinned
@@ -382,10 +654,46 @@ async def run(
             agent_name=agent_name or session.agent or "rex",
             session_ctx=session_ctx,
             trace_step_offset=trace_offset,
+            auto_failover=auto_failover,
+            auto_failover_allowed=auto_failover,
+            model_candidates=model_candidates,
+            candidate_index=candidate_index,
+            session_start_pending=trace_offset == 0,
         )
         
-        # Register context
-        cls._active_loops[session_id] = ctx
+        # Register under the same lock used by archive/delete. This closes the
+        # gap where archival could commit after the status check above but
+        # before the loop became visible to the lifecycle stop logic.
+        async with Session.lifecycle_lock(session_id):
+            latest_session = await Session.get_by_id(session_id)
+            if latest_session is None:
+                log.warning("loop.session_not_found_before_register", {
+                    "session_id": session_id,
+                })
+                return LoopResult(
+                    action="error",
+                    error=f"Session {session_id} not found",
+                )
+            if latest_session.status != "active":
+                log.warning("loop.session_not_active_before_register", {
+                    "session_id": session_id,
+                    "status": latest_session.status,
+                })
+                return LoopResult(
+                    action="error",
+                    error=f"Session {session_id} is {latest_session.status}",
+                )
+            if Session.is_lifecycle_transitioning(session_id):
+                return LoopResult(
+                    action="error",
+                    error=f"Session {session_id} is changing lifecycle state",
+                )
+            if cls.is_running(session_id):
+                return LoopResult(
+                    action="queued",
+                    error="Loop already running",
+                )
+            cls._active_loops[session_id] = ctx
         
         # Set status to busy
         SessionStatus.set(session_id, SessionStatusBusy())
@@ -424,7 +732,12 @@ async def run(
                 })
             except Exception as exc:
                 log.warn("loop.error.event_error", {"error": str(exc)})
-            return LoopResult(action="error", error=str(e))
+            return LoopResult(
+                action="error",
+                error=str(e),
+                provider_id=ctx.provider_id,
+                model_id=ctx.model_id,
+            )
         finally:
             # Clean up
             if session_id in cls._active_loops:
@@ -562,6 +875,484 @@ async def _resolve_model(
             return resolved_provider, resolved_model, source
         return resolved_provider, resolved_model
 
+    @classmethod
+    async def _prepare_auto_turn(
+        cls,
+        ctx: LoopContext,
+        last_user: MessageInfo,
+    ) -> bool:
+        """Synchronize routing when the loop advances to a real WebUI turn.
+
+        Returns:
+            True when ``last_user`` starts a new non-synthetic user turn.
+        """
+        if last_user.id == ctx.turn_user_id:
+            return False
+
+        parts = await Message.parts(last_user.id, ctx.session.id)
+        if any(bool(getattr(part, "synthetic", False)) for part in parts):
+            return False
+
+        if ctx.turn_user_id is None:
+            ctx.turn_user_id = last_user.id
+            if ctx.auto_failover and ctx.auto_failover_allowed:
+                primary = ctx.model_candidates[0]
+                preferred = cls._active_cooldown_model(
+                    ctx.session.id,
+                    primary,
+                )
+                ctx.model_candidates = await cls._build_model_candidates(
+                    primary,
+                    route_seed=f"{ctx.session.id}:{last_user.id}",
+                    preferred=preferred,
+                )
+                next_index = cls._cooldown_candidate_index(
+                    ctx.session.id,
+                    ctx.model_candidates,
+                )
+                cls._select_candidate(ctx, next_index)
+            return True
+
+        ctx.turn_user_id = last_user.id
+        persisted_session = await Session.get_by_id(ctx.session.id)
+        persisted_model_auto = bool(
+            persisted_session
+            and is_model_auto_session_category(
+                getattr(persisted_session, "category", "user")
+            )
+            and getattr(persisted_session, "model_auto", False)
+        )
+        persisted_auto = persisted_model_auto and ctx.auto_failover_allowed
+
+        user_model = getattr(last_user, "model", None)
+        user_provider_id = None
+        user_model_id = None
+        if isinstance(user_model, dict):
+            user_provider_id = user_model.get("providerID") or user_model.get("provider_id")
+            user_model_id = user_model.get("modelID") or user_model.get("model_id")
+
+        if not persisted_auto:
+            ctx.auto_failover = False
+            if not persisted_model_auto:
+                cls.clear_auto_failover_state(ctx.session.id)
+                ctx.auto_failover_allowed = False
+            provider_id = (
+                getattr(persisted_session, "provider", None)
+                if Session.has_pinned_model(persisted_session)
+                else user_provider_id
+            ) or ctx.provider_id
+            model_id = (
+                getattr(persisted_session, "model", None)
+                if Session.has_pinned_model(persisted_session)
+                else user_model_id
+            ) or ctx.model_id
+            ctx.model_candidates = [RuntimeModel(provider_id, model_id)]
+            cls._select_candidate(ctx, 0)
+            log.info("session.model.auto_disabled_for_turn", {
+                "session_id": ctx.session.id,
+                "provider_id": provider_id,
+                "model_id": model_id,
+            })
+            return True
+
+        from flocks.config.config import Config
+
+        previous = RuntimeModel(ctx.provider_id, ctx.model_id)
+        default_llm = await Config.resolve_default_llm()
+        primary = RuntimeModel(
+            provider_id=(default_llm or {}).get("provider_id") or user_provider_id or ctx.provider_id,
+            model_id=(default_llm or {}).get("model_id") or user_model_id or ctx.model_id,
+        )
+        # Rebuild once for every real turn. The user message ID makes the
+        # pseudo-random choices stable throughout that turn, while an active
+        # cooldown keeps its valid target in the newly sampled tier.
+        preferred = cls._active_cooldown_model(ctx.session.id, primary)
+        ctx.model_candidates = await cls._build_model_candidates(
+            primary,
+            route_seed=f"{ctx.session.id}:{last_user.id}",
+            preferred=preferred,
+        )
+        ctx.auto_failover = True
+        next_index = cls._cooldown_candidate_index(
+            ctx.session.id,
+            ctx.model_candidates,
+        )
+        cls._select_candidate(ctx, next_index)
+        active = ctx.model_candidates[next_index]
+        log.info("session.model.auto_turn_reset", {
+            "session_id": ctx.session.id,
+            "from_provider_id": previous.provider_id,
+            "from_model_id": previous.model_id,
+            "to_provider_id": active.provider_id,
+            "to_model_id": active.model_id,
+            "cooldown_active": next_index > 0,
+        })
+        return True
+
+    @classmethod
+    async def _run_user_prompt_submit_hook(
+        cls,
+        ctx: LoopContext,
+        last_user: MessageInfo,
+    ) -> None:
+        """Run UserPromptSubmit once for a newly observed real user turn."""
+        try:
+            from flocks.hooks.pipeline import HookPipeline
+
+            prompt = await Message.get_text_content(last_user)
+            hook_ctx = await HookPipeline.run_user_prompt_submit({
+                "sessionID": ctx.session.id,
+                "workspace": ctx.session.directory,
+                "agent": getattr(last_user, "agent", None) or ctx.agent_name,
+                "model": {
+                    "providerID": ctx.provider_id,
+                    "modelID": ctx.model_id,
+                },
+                "messageID": last_user.id,
+                "prompt": prompt,
+            })
+            additional_context = hook_ctx.output.get("additionalContext")
+            if isinstance(additional_context, str) and additional_context.strip():
+                ctx.turn_additional_context = additional_context.strip()
+        except Exception as exc:
+            log.debug("loop.hook.user_prompt_submit.error", {
+                "session_id": ctx.session.id,
+                "message_id": last_user.id,
+                "error": str(exc),
+            })
+
+    @classmethod
+    async def _run_turn_finish_hook(
+        cls,
+        ctx: LoopContext,
+        callbacks: LoopCallbacks,
+        last_user: MessageInfo,
+        last_message: MessageInfo,
+    ) -> bool:
+        """Run TurnFinish and continue the loop when the hook blocks stopping."""
+        try:
+            from flocks.hooks.pipeline import HookPipeline
+
+            hook_user = last_user
+            if ctx.turn_user_id:
+                hook_user = (
+                    await Message.get(ctx.session.id, ctx.turn_user_id)
+                    or last_user
+                )
+            user_text = await Message.get_text_content(hook_user)
+            assistant_text = await Message.get_text_content(last_message)
+            hook_ctx = await HookPipeline.run_turn_finish({
+                "sessionID": ctx.session.id,
+                "workspace": ctx.session.directory,
+                "agent": getattr(last_message, "agent", None) or ctx.agent_name,
+                "model": {
+                    "providerID": ctx.provider_id,
+                    "modelID": ctx.model_id,
+                },
+                "step": ctx.trace_step,
+                "userMessage": {
+                    "id": hook_user.id,
+                    "content": user_text,
+                },
+                "assistantMessage": {
+                    "id": last_message.id,
+                    "content": assistant_text,
+                },
+                "finishReason": "stop",
+                "stopHookActive": ctx.stop_hook_active,
+            })
+        except Exception as exc:
+            log.debug("loop.hook.turn_finish.error", {
+                "session_id": ctx.session.id,
+                "message_id": getattr(last_message, "id", None),
+                "error": str(exc),
+            })
+            return False
+
+        decision = str(hook_ctx.output.get("decision") or "").strip().lower()
+        reason = str(hook_ctx.output.get("reason") or "").strip()
+        if decision != "block":
+            return False
+        if not reason:
+            log.warn("loop.hook.turn_finish.missing_reason", {
+                "session_id": ctx.session.id,
+                "message_id": last_message.id,
+            })
+            return False
+        if ctx.should_abort():
+            log.info("loop.hook.turn_finish.ignored_after_abort", {
+                "session_id": ctx.session.id,
+                "message_id": last_message.id,
+            })
+            return False
+
+        try:
+            if ctx.session_ctx:
+                post_hook_messages = await ctx.session_ctx.get_messages()
+            else:
+                post_hook_messages = await Message.list(ctx.session.id)
+            queued_user = await cls._detect_queued_user_message(
+                ctx.session.id,
+                post_hook_messages,
+                last_user.id,
+                last_message,
+            )
+        except Exception as exc:
+            queued_user = None
+            log.debug("loop.hook.turn_finish.queued_recheck_error", {
+                "session_id": ctx.session.id,
+                "error": str(exc),
+            })
+        if queued_user is not None:
+            turn_state = set_turn_state(
+                ctx.session.id,
+                step=ctx.step,
+                status="continued",
+                continue_reason="queued_message",
+                queued_message_detected=True,
+            )
+            await cls._publish_runtime_event(callbacks, "turn.continued", {
+                **turn_state.model_dump(by_alias=True),
+                "queuedUserMessageID": queued_user.id,
+            })
+            log.info("loop.hook.turn_finish.queued_message_won", {
+                "session_id": ctx.session.id,
+                "queued_user_id": queued_user.id,
+                "source_assistant_message_id": last_message.id,
+            })
+            return True
+
+        from flocks.agent.registry import Agent
+        from flocks.session.core.defaults import DEFAULT_MAX_TOOL_STEPS
+
+        try:
+            agent = await Agent.get(
+                getattr(last_message, "agent", None) or ctx.agent_name
+            )
+        except Exception as exc:
+            log.debug("loop.hook.turn_finish.agent_load_error", {
+                "session_id": ctx.session.id,
+                "error": str(exc),
+            })
+            agent = None
+        max_steps = (
+            agent.steps
+            if agent is not None and getattr(agent, "steps", None) is not None
+            else DEFAULT_MAX_TOOL_STEPS
+        )
+        if ctx.trace_step >= max_steps:
+            log.warn("loop.hook.turn_finish.step_limit", {
+                "session_id": ctx.session.id,
+                "step": ctx.trace_step,
+                "max_steps": max_steps,
+            })
+            return False
+
+        try:
+            continuation = await Message.create(
+                session_id=ctx.session.id,
+                role=MessageRole.USER,
+                content=reason,
+                agent=getattr(hook_user, "agent", None) or ctx.agent_name,
+                model={
+                    "providerID": ctx.provider_id,
+                    "modelID": ctx.model_id,
+                },
+                synthetic=True,
+                part_metadata={
+                    "turnFinishContinuation": True,
+                    "stopHookActive": True,
+                    "sourceAssistantMessageID": last_message.id,
+                },
+            )
+        except Exception as exc:
+            log.error("loop.hook.turn_finish.continuation_error", {
+                "session_id": ctx.session.id,
+                "error": str(exc),
+            })
+            return False
+        ctx.stop_hook_active = True
+        turn_state = set_turn_state(
+            ctx.session.id,
+            step=ctx.step,
+            status="continued",
+            continue_reason="turn_finish_hook",
+            queued_message_detected=False,
+        )
+        await cls._publish_runtime_event(callbacks, "turn.continued", {
+            **turn_state.model_dump(by_alias=True),
+            "turnFinishMessageID": continuation.id,
+        })
+        log.info("loop.continuing_for_turn_finish_hook", {
+            "session_id": ctx.session.id,
+            "continuation_message_id": continuation.id,
+            "source_assistant_message_id": last_message.id,
+        })
+        return True
+
+    @classmethod
+    async def _finalize_deferred_failure(
+        cls,
+        ctx: LoopContext,
+        failure: Any,
+        last_user: MessageInfo,
+    ) -> None:
+        """Persist only the final Auto candidate failure."""
+        if not failure.assistant_message_id:
+            assistant = await Message.create(
+                session_id=ctx.session.id,
+                role=MessageRole.ASSISTANT,
+                content="",
+                agent=getattr(last_user, "agent", None) or ctx.agent_name or "rex",
+                model_id=ctx.model_id,
+                provider_id=ctx.provider_id,
+                parent_id=last_user.id,
+                error=failure.error_data,
+                finish="error",
+            )
+            failure.assistant_message_id = assistant.id
+            return
+        await Message.update(
+            ctx.session.id,
+            failure.assistant_message_id,
+            error=failure.error_data,
+            finish="error",
+        )
+
+    @classmethod
+    async def _process_step_with_failover(
+        cls,
+        ctx: LoopContext,
+        callbacks: LoopCallbacks,
+        messages: List[MessageInfo],
+        last_user: MessageInfo,
+    ) -> Any:
+        """Run one logical step, moving across candidates without replaying output."""
+        from flocks.session.runner import RunnerCallbacks, SessionRunner
+
+        while True:
+            runner_cbs = callbacks.runner_callbacks
+            if runner_cbs is None:
+                runner_cbs = RunnerCallbacks()
+            if callbacks.event_publish_callback and not runner_cbs.event_publish_callback:
+                runner_cbs.event_publish_callback = callbacks.event_publish_callback
+
+            runner = SessionRunner(
+                session=ctx.session,
+                provider_id=ctx.provider_id,
+                model_id=ctx.model_id,
+                agent_name=ctx.agent_name,
+                abort_event=ctx.abort_event,
+                callbacks=runner_cbs,
+                session_ctx=ctx.session_ctx,
+                memory_bootstrap_data=ctx.memory_bootstrap_data,
+                static_cache=ctx.runner_static_cache,
+                defer_step_errors=ctx.auto_failover,
+                failover_available=(
+                    ctx.auto_failover
+                    and ctx.candidate_index + 1 < len(ctx.model_candidates)
+                ),
+                turn_additional_context=ctx.turn_additional_context,
+                session_start_pending=ctx.session_start_pending,
+            )
+            runner._step = ctx.trace_step
+
+            step_result = await runner._process_step(messages, last_user)
+            if runner._session_start_fired:
+                ctx.session_start_pending = False
+            failure = step_result.failure
+            if not ctx.auto_failover or failure is None:
+                return step_result
+
+            next_index = ctx.candidate_index + 1
+            has_next = next_index < len(ctx.model_candidates)
+            if not failure.allow_fallback or not has_next:
+                if (
+                    failure.allow_fallback
+                    and not has_next
+                    and ctx.candidate_index > 0
+                    and failure.reason not in {"rate_limit", "billing"}
+                ):
+                    expires_at = time.monotonic() + CHAIN_EXHAUSTION_COOLDOWN_SECONDS
+                    existing_cooldown = cls._auto_failover_cooldowns.get(ctx.session.id)
+                    if not (
+                        existing_cooldown
+                        and existing_cooldown.expires_at > expires_at
+                    ):
+                        cls._auto_failover_cooldowns[ctx.session.id] = AutoFailoverCooldown(
+                            model=ctx.model_candidates[ctx.candidate_index],
+                            primary=ctx.model_candidates[0],
+                            expires_at=expires_at,
+                            reason="chain_exhausted",
+                        )
+                await cls._finalize_deferred_failure(ctx, failure, last_user)
+                return step_result
+
+            # A candidate may be removed only while its attempt is completely
+            # replay-safe. Failure to delete stops the switch to avoid leaving
+            # two assistant cards for one logical response.
+            if failure.assistant_message_id:
+                try:
+                    deleted = await Message.delete(
+                        ctx.session.id,
+                        failure.assistant_message_id,
+                    )
+                except Exception as exc:
+                    deleted = False
+                    log.error("session.model.fallback_cleanup_failed", {
+                        "session_id": ctx.session.id,
+                        "message_id": failure.assistant_message_id,
+                        "error": str(exc),
+                    })
+                if not deleted:
+                    await cls._finalize_deferred_failure(ctx, failure, last_user)
+                    return step_result
+                await cls._publish_runtime_event(callbacks, "message.removed", {
+                    "sessionID": ctx.session.id,
+                    "messageID": failure.assistant_message_id,
+                })
+
+            previous = ctx.model_candidates[ctx.candidate_index]
+            next_candidate = ctx.model_candidates[next_index]
+
+            if ctx.candidate_index == 0 and failure.reason in {"rate_limit", "billing"}:
+                cls._auto_failover_cooldowns[ctx.session.id] = AutoFailoverCooldown(
+                    model=next_candidate,
+                    primary=ctx.model_candidates[0],
+                    expires_at=time.monotonic() + RATE_LIMIT_COOLDOWN_SECONDS,
+                    reason=failure.reason,
+                )
+            else:
+                cooldown = cls._auto_failover_cooldowns.get(ctx.session.id)
+                if cooldown and cooldown.expires_at > time.monotonic():
+                    cooldown.model = next_candidate
+
+            cls._select_candidate(ctx, next_index)
+            event_payload = {
+                "sessionID": ctx.session.id,
+                "from": {
+                    "providerID": previous.provider_id,
+                    "modelID": previous.model_id,
+                },
+                "to": {
+                    "providerID": next_candidate.provider_id,
+                    "modelID": next_candidate.model_id,
+                },
+                "reason": failure.reason,
+                "candidateIndex": next_index,
+            }
+            log.warn("session.model.fallback", {
+                "from": event_payload["from"],
+                "to": event_payload["to"],
+                "reason": event_payload["reason"],
+                "candidateIndex": event_payload["candidateIndex"],
+            })
+            await cls._publish_runtime_event(
+                callbacks,
+                "session.model.fallback",
+                event_payload,
+            )
+
     @classmethod
     async def _run_loop(
         cls,
@@ -581,6 +1372,7 @@ async def _run_loop(
         7. Loop until complete
         """
         last_message: Optional[MessageInfo] = None
+        loop_error: Optional[str] = None
         
         while not ctx.should_abort():
             # Set status to busy
@@ -678,7 +1470,7 @@ async def _run_loop(
                     stop_reason="no_user_message",
                 )
                 break
-            
+
             last_assistant_parts = (
                 await Message.parts(last_assistant.id, ctx.session.id)
                 if last_assistant
@@ -699,6 +1491,11 @@ async def _run_loop(
                 })
                 last_message = last_assistant
                 break
+
+            if await cls._prepare_auto_turn(ctx, last_user):
+                ctx.turn_additional_context = None
+                ctx.stop_hook_active = False
+                await cls._run_user_prompt_submit_hook(ctx, last_user)
             
             # Bootstrap memory on first step (once per loop, stored in ctx)
             if ctx.step == 1 and ctx.session.memory_enabled and ctx.memory_bootstrap_data is None:
@@ -720,7 +1517,7 @@ async def _run_loop(
             # may cancel this task before it finishes).
             # generate_title_after_first_message is idempotent: if this task saves
             # the title first, the safety-net call returns immediately.
-            if ctx.step == 1:
+            if ctx.step == 1 and not ctx.auto_failover:
                 try:
                     from flocks.session.lifecycle.title import SessionTitle
                     # UserMessageInfo.model is Dict[str, str] {"providerID": ..., "modelID": ...}
@@ -1207,34 +2004,16 @@ async def progress_callback_overflow(stage: str, data: dict) -> None:
                     except Exception as e:
                         log.error("loop.compaction_overflow_check_error", {"error": str(e)})
             
-            # Process step - delegate to runner (matching TUI SessionProcessor.process)
-            from flocks.session.runner import SessionRunner, RunnerCallbacks
-            
-            # Build runner callbacks from loop callbacks
-            runner_cbs = callbacks.runner_callbacks
-            if runner_cbs is None:
-                runner_cbs = RunnerCallbacks()
-            # Ensure event_publish_callback is propagated
-            if callbacks.event_publish_callback and not runner_cbs.event_publish_callback:
-                runner_cbs.event_publish_callback = callbacks.event_publish_callback
-            
-            runner = SessionRunner(
-                session=ctx.session,
-                provider_id=ctx.provider_id,
-                model_id=ctx.model_id,
-                agent_name=ctx.agent_name,
-                abort_event=ctx.abort_event,
-                callbacks=runner_cbs,
-                session_ctx=ctx.session_ctx,
-                memory_bootstrap_data=ctx.memory_bootstrap_data,
-                static_cache=ctx.runner_static_cache,
-            )
-            # Use session-cumulative step number for observability.
-            runner._step = ctx.trace_step
-            
             # Process single step — wrap in a Task so abort() can cancel it immediately
             # rather than waiting for the current tool call to finish.
-            step_task = asyncio.create_task(runner._process_step(messages, last_user))
+            step_task = asyncio.create_task(
+                cls._process_step_with_failover(
+                    ctx,
+                    callbacks,
+                    messages,
+                    last_user,
+                )
+            )
             ctx._current_step_task = step_task
             step_started_at = asyncio.get_event_loop().time()
             try:
@@ -1256,6 +2035,7 @@ async def progress_callback_overflow(stage: str, data: dict) -> None:
             
             # Handle result
             if step_result.action == "stop":
+                loop_error = step_result.error
                 # Report error if step failed
                 if step_result.error and callbacks.on_error:
                     await callbacks.on_error(step_result.error)
@@ -1266,7 +2046,13 @@ async def progress_callback_overflow(stage: str, data: dict) -> None:
                 else:
                     post_messages = await Message.list(ctx.session.id)
                 for msg in reversed(post_messages):
-                    if msg.role == MessageRole.ASSISTANT:
+                    if (
+                        msg.role == MessageRole.ASSISTANT
+                        and (
+                            not ctx.auto_failover
+                            or getattr(msg, "parentID", None) == last_user.id
+                        )
+                    ):
                         last_message = msg
                         break
 
@@ -1375,6 +2161,20 @@ async def progress_callback_overflow(stage: str, data: dict) -> None:
                         })
                         continue
 
+                if (
+                    not step_result.error
+                    and not ctx.should_abort()
+                    and last_message is not None
+                    and getattr(last_message, "finish", None) == "stop"
+                    and await cls._run_turn_finish_hook(
+                        ctx,
+                        callbacks,
+                        last_user,
+                        last_message,
+                    )
+                ):
+                    continue
+
                 stop_reason = step_result.error or (getattr(last_message, "finish", None) if last_message else None) or "stop"
                 turn_state = set_turn_state(
                     ctx.session.id,
@@ -1429,12 +2229,16 @@ async def progress_callback_overflow(stage: str, data: dict) -> None:
         
         # Return result
         return LoopResult(
-            action="stop",
+            action="error" if ctx.auto_failover and loop_error else "stop",
             last_message=last_message,
+            error=loop_error if ctx.auto_failover else None,
+            provider_id=ctx.provider_id,
+            model_id=ctx.model_id,
             metadata={
                 "steps": ctx.step,
                 "session_id": ctx.session.id,
                 "last_compaction_step": ctx.last_compaction_step,
+                **({"aborted": True} if ctx.should_abort() else {}),
             },
         )
     
@@ -1709,6 +2513,7 @@ async def _execute_subtask(
             agent=last_user.agent if hasattr(last_user, 'agent') else agent_name,
             model=last_user.model if hasattr(last_user, 'model') else model_id,
             provider=last_user.provider if hasattr(last_user, 'provider') else provider_id,
+            synthetic=True,
         )
         
         log.info("loop.subtask.completed", {
diff --git a/flocks/session/streaming/stream_processor.py b/flocks/session/streaming/stream_processor.py
index 1835e0ff8..ca1666474 100644
--- a/flocks/session/streaming/stream_processor.py
+++ b/flocks/session/streaming/stream_processor.py
@@ -119,6 +119,10 @@ def __init__(
         workspace_dir: Optional[str] = None,
         langfuse_generation: Optional[Any] = None,
         step_index: Optional[int] = None,
+        execution_mode: str = "build",
+        plan_file_path: Optional[str] = None,
+        plan_relative_path: Optional[str] = None,
+        plan_permission_path: Optional[str] = None,
     ):
         self.session_id = session_id
         self.assistant_message = assistant_message
@@ -136,6 +140,10 @@ def __init__(
         self._workspace_dir = workspace_dir
         self._langfuse_generation = langfuse_generation
         self._step_index = step_index
+        self._execution_mode = execution_mode
+        self._plan_file_path = plan_file_path
+        self._plan_relative_path = plan_relative_path
+        self._plan_permission_path = plan_permission_path
         self._sandbox_runtime_cache = None
         self._sandbox_config_cache = None
         self._sandbox_context_cache = None
@@ -657,9 +665,11 @@ async def _handle_tool_call(self, event: ToolCallEvent) -> None:
             except Exception as e:
                 log.error("stream.tool_start_callback.error", {"error": str(e)})
         
-        # Hook pipeline: tool.execute.before
+        # Hook pipeline: tool_before
+        post_hook_fired = False
         try:
             from flocks.hooks.pipeline import HookPipeline
+
             hook_ctx = await HookPipeline.run_tool_before({
                 "sessionID": self.session_id,
                 "workspace": self._workspace_dir,
@@ -674,10 +684,28 @@ async def _handle_tool_call(self, event: ToolCallEvent) -> None:
                 updated = hook_ctx.input.get("tool", {}).get("input")
                 if isinstance(updated, dict):
                     tool_input = updated
-            hook_skip = hook_ctx.output.get("skip") if hook_ctx else False
+            decision = str(
+                (hook_ctx.output.get("decision") if hook_ctx else "") or ""
+            ).strip().lower()
+            hook_blocked = decision == "block"
+            hook_block_reason = str(
+                (hook_ctx.output.get("reason") if hook_ctx else "") or ""
+            ).strip()
+        except asyncio.CancelledError:
+            await self._finalize_interrupted_tool_call(
+                tool_state=tool_state,
+                tool_name=tool_name,
+                tool_input=tool_input,
+                tool_call_id=tool_call_id,
+                tool_start_time=tool_start_time,
+                post_hook_fired=False,
+            )
+            raise
         except Exception as e:
             log.error("stream.tool_before_hook.error", {"error": str(e)})
-            hook_skip = False
+            hook_blocked = False
+            hook_block_reason = ""
+        tool_state.input = tool_input
 
         # Execute tool synchronously
         tool_span_ctx = None
@@ -699,10 +727,11 @@ async def _handle_tool_call(self, event: ToolCallEvent) -> None:
             except Exception as exc:
                 log.debug("stream.tool_span.init_failed", {"error": str(exc)})
         try:
-            if hook_skip:
+            if hook_blocked:
                 result = ToolResult(
                     success=False,
-                    error="Tool execution blocked by hook",
+                    error=hook_block_reason or "Tool execution blocked by hook",
+                    metadata={"blocked_by_hook": True},
                 )
             else:
                 sandbox_meta = await self._resolve_sandbox_meta(tool_name)
@@ -818,6 +847,26 @@ def _mark_finished() -> None:
                         _cb.mark_finished = _mark_finished
                         return _cb
 
+                    tool_extra = {
+                        **sandbox_meta["extra"],
+                        "execution_mode": self._execution_mode,
+                        "workspace_dir": self._workspace_dir,
+                        "model": {
+                            "providerID": getattr(
+                                self.assistant_message,
+                                "providerID",
+                                None,
+                            ),
+                            "modelID": getattr(
+                                self.assistant_message,
+                                "modelID",
+                                None,
+                            ),
+                        },
+                        "plan_file_path": self._plan_file_path,
+                        "plan_relative_path": self._plan_relative_path,
+                        "plan_permission_path": self._plan_permission_path,
+                    }
                     ctx = ToolContext(
                         session_id=self.session_id,
                         message_id=self.assistant_message.id,
@@ -825,7 +874,7 @@ def _mark_finished() -> None:
                         call_id=tool_call_id,
                         abort_event=self.abort_event,
                         permission_callback=self.permission_callback,
-                        extra=sandbox_meta["extra"],
+                        extra=tool_extra,
                         metadata_callback=_make_metadata_cb(),
                         event_publish_callback=self.event_publish_callback,
                     )
@@ -843,26 +892,23 @@ def _mark_finished() -> None:
                         if cb and hasattr(cb, 'mark_finished'):
                             cb.mark_finished()
 
-            # Hook pipeline: tool.execute.after
-            try:
-                from flocks.hooks.pipeline import HookPipeline
-                hook_ctx = await HookPipeline.run_tool_after({
-                    "sessionID": self.session_id,
-                    "workspace": self._workspace_dir,
-                    "agent": self.agent.name,
-                    "tool": {
-                        "name": tool_name,
-                        "input": tool_input,
-                        "callID": tool_call_id,
-                    },
-                    "result": result.model_dump(),
-                })
-                if hook_ctx and isinstance(hook_ctx.output, dict):
-                    override = hook_ctx.output.get("result")
-                    if isinstance(override, dict):
-                        result = ToolResult(**override)
-            except Exception as e:
-                log.error("stream.tool_after_hook.error", {"error": str(e)})
+            if result.success:
+                hook_status = "completed"
+            elif (result.metadata or {}).get("blocked_by_hook"):
+                hook_status = "blocked"
+            elif (result.metadata or {}).get("blocked_by_policy"):
+                hook_status = "blocked"
+            else:
+                hook_status = "error"
+            post_hook_fired = True
+            result = await self._run_tool_after_hook(
+                tool_name=tool_name,
+                tool_input=tool_input,
+                tool_call_id=tool_call_id,
+                result=result,
+                status=hook_status,
+                tool_start_time=tool_start_time,
+            )
             
             # Update tool state
             tool_state.status = "completed" if result.success else "error"
@@ -874,14 +920,10 @@ def _mark_finished() -> None:
                 "tool_name": tool_name,
                 "success": result.success,
             })
-            output_preview = result.output if result.success else result.error
-            if isinstance(output_preview, str):
-                output_preview = output_preview[:600]
-
             try:
                 if tool_span_ctx is not None:
                     end_kwargs: Dict[str, Any] = {
-                        "output": output_preview,
+                        "output": result.output if result.success else result.error,
                         "metadata": {
                             "success": result.success,
                             "title": result.title,
@@ -968,65 +1010,15 @@ def _mark_finished() -> None:
                     log.error("stream.tool_end_callback.error", {"error": str(e)})
             
         except asyncio.CancelledError:
-            interrupt_msg = "Tool execution was interrupted"
-            log.info("stream.tool_call.cancelled", {
-                "tool_call_id": tool_call_id,
-                "tool_name": tool_name,
-            })
-            try:
-                if tool_span_ctx is not None:
-                    tool_span_ctx.end(
-                        output=interrupt_msg,
-                        metadata={"success": False},
-                        level="ERROR",
-                        status_message="tool_cancelled",
-                    )
-            except Exception as _span_err:
-                log.debug("stream.tool_span.cancel_end_failed", {"error": str(_span_err)})
-
-            tool_state.status = "error"
-            tool_state.error = interrupt_msg
-
-            try:
-                tool_end_time = int(datetime.now().timestamp() * 1000)
-                error_state = ToolStateError(
-                    status="error",
-                    input=tool_input,
-                    error=interrupt_msg,
-                    time={"start": tool_start_time if 'tool_start_time' in locals() else tool_end_time, "end": tool_end_time},
-                )
-
-                error_part = ToolPart(
-                    id=tool_state.part_id,
-                    sessionID=self.session_id,
-                    messageID=self.assistant_message.id,
-                    type="tool",
-                    callID=tool_call_id,
-                    tool=tool_name,
-                    state=error_state,
-                )
-                await Message.store_part(self.session_id, self.assistant_message.id, error_part)
-
-                if self.event_publish_callback:
-                    await self.event_publish_callback("message.part.updated", {
-                        "part": {
-                            "id": tool_state.part_id,
-                            "messageID": self.assistant_message.id,
-                            "sessionID": self.session_id,
-                            "type": "tool",
-                            "callID": tool_call_id,
-                            "tool": tool_name,
-                            "state": {
-                                "status": "error",
-                                "input": tool_input,
-                                "error": interrupt_msg,
-                                "time": {"start": tool_start_time if 'tool_start_time' in locals() else tool_end_time, "end": tool_end_time},
-                            }
-                        }
-                    })
-            except Exception as store_e:
-                log.error("stream.tool_call.cancelled_update_failed", {"error": str(store_e)})
-
+            await self._finalize_interrupted_tool_call(
+                tool_state=tool_state,
+                tool_name=tool_name,
+                tool_input=tool_input,
+                tool_call_id=tool_call_id,
+                tool_start_time=tool_start_time,
+                post_hook_fired=post_hook_fired,
+                tool_span_ctx=tool_span_ctx,
+            )
             raise
         except Exception as e:
             log.error("stream.tool_call.error", {
@@ -1034,6 +1026,21 @@ def _mark_finished() -> None:
                 "tool_name": tool_name,
                 "error": str(e),
             })
+            if not post_hook_fired:
+                post_hook_fired = True
+                exception_result = ToolResult(
+                    success=False,
+                    error=str(e),
+                )
+                await self._run_tool_after_hook(
+                    tool_name=tool_name,
+                    tool_input=tool_input,
+                    tool_call_id=tool_call_id,
+                    result=exception_result,
+                    status="error",
+                    tool_start_time=tool_start_time,
+                    allow_result_override=False,
+                )
             try:
                 if tool_span_ctx is not None:
                     tool_span_ctx.end(
@@ -1107,6 +1114,148 @@ def _mark_finished() -> None:
                 except Exception as e2:
                     log.error("stream.tool_end_callback.error", {"error": str(e2)})
 
+    async def _finalize_interrupted_tool_call(
+        self,
+        *,
+        tool_state: ToolCallState,
+        tool_name: str,
+        tool_input: Dict[str, Any],
+        tool_call_id: str,
+        tool_start_time: int,
+        post_hook_fired: bool,
+        tool_span_ctx: Optional[Any] = None,
+    ) -> None:
+        """Emit and persist the terminal state for an interrupted tool call."""
+        interrupt_msg = "Tool execution was interrupted"
+        log.info("stream.tool_call.cancelled", {
+            "tool_call_id": tool_call_id,
+            "tool_name": tool_name,
+        })
+        try:
+            if not post_hook_fired:
+                interrupted_result = ToolResult(
+                    success=False,
+                    error=interrupt_msg,
+                    metadata={"interrupted": True},
+                )
+                await self._run_tool_after_hook(
+                    tool_name=tool_name,
+                    tool_input=tool_input,
+                    tool_call_id=tool_call_id,
+                    result=interrupted_result,
+                    status="interrupted",
+                    tool_start_time=tool_start_time,
+                    allow_result_override=False,
+                )
+        finally:
+            try:
+                if tool_span_ctx is not None:
+                    tool_span_ctx.end(
+                        output=interrupt_msg,
+                        metadata={"success": False},
+                        level="ERROR",
+                        status_message="tool_cancelled",
+                    )
+            except Exception as span_error:
+                log.debug(
+                    "stream.tool_span.cancel_end_failed",
+                    {"error": str(span_error)},
+                )
+
+            tool_state.status = "error"
+            tool_state.error = interrupt_msg
+            tool_end_time = int(datetime.now().timestamp() * 1000)
+            error_state = ToolStateError(
+                status="error",
+                input=tool_input,
+                error=interrupt_msg,
+                time={"start": tool_start_time, "end": tool_end_time},
+            )
+            error_part = ToolPart(
+                id=tool_state.part_id,
+                sessionID=self.session_id,
+                messageID=self.assistant_message.id,
+                type="tool",
+                callID=tool_call_id,
+                tool=tool_name,
+                state=error_state,
+            )
+            try:
+                await Message.store_part(
+                    self.session_id,
+                    self.assistant_message.id,
+                    error_part,
+                )
+                if self.event_publish_callback:
+                    await self.event_publish_callback(
+                        "message.part.updated",
+                        {
+                            "part": {
+                                "id": tool_state.part_id,
+                                "messageID": self.assistant_message.id,
+                                "sessionID": self.session_id,
+                                "type": "tool",
+                                "callID": tool_call_id,
+                                "tool": tool_name,
+                                "state": {
+                                    "status": "error",
+                                    "input": tool_input,
+                                    "error": interrupt_msg,
+                                    "time": {
+                                        "start": tool_start_time,
+                                        "end": tool_end_time,
+                                    },
+                                },
+                            }
+                        },
+                    )
+            except Exception as store_error:
+                log.error(
+                    "stream.tool_call.cancelled_update_failed",
+                    {"error": str(store_error)},
+                )
+
+    async def _run_tool_after_hook(
+        self,
+        *,
+        tool_name: str,
+        tool_input: Dict[str, Any],
+        tool_call_id: str,
+        result: ToolResult,
+        status: str,
+        tool_start_time: int,
+        allow_result_override: bool = True,
+    ) -> ToolResult:
+        """Run tool_after with a normalized result for every tool outcome."""
+        try:
+            from flocks.hooks.pipeline import HookPipeline
+
+            duration_ms = max(
+                0,
+                int(datetime.now().timestamp() * 1000) - tool_start_time,
+            )
+            hook_ctx = await HookPipeline.run_tool_after({
+                "sessionID": self.session_id,
+                "workspace": self._workspace_dir,
+                "agent": self.agent.name,
+                "tool": {
+                    "name": tool_name,
+                    "input": tool_input,
+                    "callID": tool_call_id,
+                },
+                "status": status,
+                "durationMs": duration_ms,
+                "result": result.model_dump(),
+                "error": result.error if not result.success else None,
+            })
+            if allow_result_override and isinstance(hook_ctx.output, dict):
+                override = hook_ctx.output.get("result")
+                if isinstance(override, dict):
+                    return ToolResult(**override)
+        except Exception as exc:
+            log.error("stream.tool_after_hook.error", {"error": str(exc)})
+        return result
+
     async def _load_config_data(self) -> Dict[str, Any]:
         """Load and cache config as plain dict."""
         if isinstance(self._config_data, dict):
diff --git a/flocks/session/streaming/timeouts.py b/flocks/session/streaming/timeouts.py
new file mode 100644
index 000000000..4cc4babc7
--- /dev/null
+++ b/flocks/session/streaming/timeouts.py
@@ -0,0 +1,156 @@
+"""Resolve adaptive timeout budgets for LLM streaming responses."""
+
+from __future__ import annotations
+
+import ipaddress
+import os
+from dataclasses import dataclass
+from typing import Any, Mapping, Optional
+from urllib.parse import urlparse
+
+
+DEFAULT_FIRST_CHUNK_TIMEOUT_S = 120.0
+DEFAULT_LOCAL_FIRST_CHUNK_TIMEOUT_S = 1800.0
+DEFAULT_ONGOING_CHUNK_TIMEOUT_S = 300.0
+
+FIRST_CHUNK_TIMEOUT_ENV = "FLOCKS_LLM_STREAM_FIRST_CHUNK_TIMEOUT_S"
+ONGOING_CHUNK_TIMEOUT_ENV = "FLOCKS_LLM_STREAM_ONGOING_CHUNK_TIMEOUT_S"
+
+_FIRST_CHUNK_SETTING_KEYS = (
+    "stream_first_chunk_timeout_s",
+    "streamFirstChunkTimeoutSeconds",
+)
+_ONGOING_CHUNK_SETTING_KEYS = (
+    "stream_ongoing_chunk_timeout_s",
+    "streamOngoingChunkTimeoutSeconds",
+)
+_LOCAL_PROVIDER_IDS = frozenset({"local", "ollama"})
+_LOCAL_HOSTNAMES = frozenset({"localhost", "host.docker.internal"})
+
+
+@dataclass(frozen=True)
+class LlmStreamTimeouts:
+    """Effective timeout budgets for one LLM stream."""
+
+    first_chunk_s: float
+    ongoing_chunk_s: float
+    is_local: bool
+
+
+def _positive_float(value: Any) -> Optional[float]:
+    """Return a positive finite float, or ``None`` for an invalid value."""
+    try:
+        parsed = float(value)
+    except (TypeError, ValueError):
+        return None
+    if parsed <= 0 or parsed == float("inf") or parsed != parsed:
+        return None
+    return parsed
+
+
+def _setting(
+    settings: Mapping[str, Any],
+    keys: tuple[str, ...],
+) -> Optional[float]:
+    for key in keys:
+        if key in settings:
+            value = _positive_float(settings[key])
+            if value is not None:
+                return value
+    return None
+
+
+def _provider_base_url(provider: Any) -> str:
+    config = getattr(provider, "_config", None)
+    candidates = (
+        getattr(config, "base_url", None),
+        getattr(provider, "_base_url", None),
+        getattr(provider, "DEFAULT_BASE_URL", None),
+    )
+    return next(
+        (value.strip() for value in candidates if isinstance(value, str) and value.strip()),
+        "",
+    )
+
+
+def _is_local_endpoint(provider: Any) -> bool:
+    provider_id = str(getattr(provider, "id", "") or "").strip().lower()
+    if provider_id in _LOCAL_PROVIDER_IDS:
+        return True
+
+    hostname = (urlparse(_provider_base_url(provider)).hostname or "").lower()
+    if not hostname:
+        return False
+    if hostname in _LOCAL_HOSTNAMES or hostname.endswith(".local"):
+        return True
+
+    try:
+        address = ipaddress.ip_address(hostname)
+    except ValueError:
+        return False
+    return address.is_loopback or address.is_private or address.is_link_local
+
+
+def _provider_settings(provider: Any) -> Mapping[str, Any]:
+    config = getattr(provider, "_config", None)
+    settings = getattr(config, "custom_settings", None)
+    return settings if isinstance(settings, Mapping) else {}
+
+
+def _model_settings(provider: Any, model_id: str) -> Mapping[str, Any]:
+    try:
+        models = provider.get_models()
+    except Exception:
+        return {}
+    for model in models or []:
+        if getattr(model, "id", None) != model_id:
+            continue
+        settings = getattr(model, "custom_settings", None)
+        return settings if isinstance(settings, Mapping) else {}
+    return {}
+
+
+def _resolve_timeout(
+    *,
+    model_settings: Mapping[str, Any],
+    provider_settings: Mapping[str, Any],
+    setting_keys: tuple[str, ...],
+    environment_name: str,
+    default: float,
+) -> float:
+    candidates = (
+        _setting(model_settings, setting_keys),
+        _setting(provider_settings, setting_keys),
+        _positive_float(os.getenv(environment_name)),
+    )
+    return next((value for value in candidates if value is not None), default)
+
+
+def resolve_llm_stream_timeouts(provider: Any, model_id: str) -> LlmStreamTimeouts:
+    """Resolve model, provider, environment, and endpoint-aware timeouts.
+
+    Precedence is model configuration, provider configuration, environment,
+    then the built-in endpoint-aware default.
+    """
+    is_local = _is_local_endpoint(provider)
+    provider_settings = _provider_settings(provider)
+    model_settings = _model_settings(provider, model_id)
+    first_chunk_default = DEFAULT_LOCAL_FIRST_CHUNK_TIMEOUT_S if is_local else DEFAULT_FIRST_CHUNK_TIMEOUT_S
+
+    return LlmStreamTimeouts(
+        first_chunk_s=_resolve_timeout(
+            model_settings=model_settings,
+            provider_settings=provider_settings,
+            setting_keys=_FIRST_CHUNK_SETTING_KEYS,
+            environment_name=FIRST_CHUNK_TIMEOUT_ENV,
+            default=first_chunk_default,
+        ),
+        ongoing_chunk_s=_resolve_timeout(
+            model_settings=model_settings,
+            provider_settings=provider_settings,
+            setting_keys=_ONGOING_CHUNK_SETTING_KEYS,
+            environment_name=ONGOING_CHUNK_TIMEOUT_ENV,
+            default=DEFAULT_ONGOING_CHUNK_TIMEOUT_S,
+        ),
+        is_local=is_local,
+    )
diff --git a/flocks/session/streaming/tool_accumulator.py b/flocks/session/streaming/tool_accumulator.py
index 1957ed259..6e266e95d 100644
--- a/flocks/session/streaming/tool_accumulator.py
+++ b/flocks/session/streaming/tool_accumulator.py
@@ -43,12 +43,17 @@ def __init__(self, processor: Any) -> None:
     async def feed_chunk(self, tc: dict[str, Any]) -> None:
         """Process a single tool-call chunk from the provider stream."""
         tc_index = tc.get("index", 0)
-        tc_id = tc.get("id")
+        provider_tc_id = tc.get("id")
 
-        if tc_id:
-            self._index_to_id[tc_index] = tc_id
-        elif tc_index in self._index_to_id:
+        # Keep one stable internal id for the lifetime of a streamed tool call.
+        # Some providers reveal the tool name before their call id; replacing a
+        # generated id later would disconnect the already-published pending UI
+        # part from the eventual ToolCallEvent.
+        if tc_index in self._index_to_id:
             tc_id = self._index_to_id[tc_index]
+        elif provider_tc_id:
+            tc_id = provider_tc_id
+            self._index_to_id[tc_index] = tc_id
         else:
             tc_id = Identifier.create("call")
             self._index_to_id[tc_index] = tc_id
@@ -79,6 +84,17 @@ async def feed_chunk(self, tc: dict[str, Any]) -> None:
 
         accumulated = self._accumulator[tc_id]["arguments_str"]
         final_name = self._accumulator[tc_id]["name"]
+
+        # Publish the tool step as soon as its name is known. Parameters can be
+        # large (especially write/edit content), so waiting for valid JSON here
+        # would hide the tool until input generation — and often execution —
+        # was nearly complete.
+        if final_name and not self._accumulator[tc_id].get("input_started"):
+            await self._processor.process_event(
+                ToolInputStartEvent(id=tc_id, tool_name=final_name)
+            )
+            self._accumulator[tc_id]["input_started"] = True
+
         if accumulated and final_name:
             arguments, ok = _parse_json_robust(accumulated)
             if ok:
@@ -89,12 +105,6 @@ async def feed_chunk(self, tc: dict[str, Any]) -> None:
                         self._accumulator[tc_id]["awaiting_required"] = True
                         return
 
-                if not self._accumulator[tc_id].get("input_started") and final_name:
-                    await self._processor.process_event(
-                        ToolInputStartEvent(id=tc_id, tool_name=final_name)
-                    )
-                    self._accumulator[tc_id]["input_started"] = True
-
                 if final_name:
                     await self._processor.process_event(
                         ToolCallEvent(
diff --git a/flocks/storage/storage.py b/flocks/storage/storage.py
index debf1bf16..43a8724cc 100644
--- a/flocks/storage/storage.py
+++ b/flocks/storage/storage.py
@@ -13,7 +13,7 @@
 from contextlib import asynccontextmanager
 from pathlib import Path
 import sqlite3
-from typing import Any, AsyncIterator, Awaitable, Callable, Dict, List, Optional, Tuple, Type, TypeVar
+from typing import Any, AsyncIterator, Awaitable, Callable, Dict, List, Optional, Sequence, Tuple, Type, TypeVar
 import json
 import aiosqlite
 from datetime import datetime
@@ -1543,6 +1543,105 @@ async def _write() -> None:
 
         cls._log.debug("storage.set", {"key": key, "type": value_type})
 
+    @classmethod
+    async def set_many(cls, entries: Sequence[Tuple[str, Any, str]]) -> None:
+        """Store multiple values atomically in one SQLite transaction."""
+        await cls.mutate_many(set_entries=entries)
+
+    @classmethod
+    async def mutate_many(
+        cls,
+        *,
+        set_entries: Sequence[Tuple[str, Any, str]] = (),
+        delete_keys: Sequence[str] = (),
+        delete_prefixes: Sequence[str] = (),
+    ) -> int:
+        """Apply related set/delete operations in one SQLite transaction."""
+        entries = list(set_entries)
+        keys_to_delete = list(delete_keys)
+        prefixes_to_delete = list(delete_prefixes)
+        if not entries and not keys_to_delete and not prefixes_to_delete:
+            return 0
+
+        routing_paths = {
+            *(cls.route_db_path_for_key(key) for key, _value, _value_type in entries),
+            *(cls.route_db_path_for_key(key) for key in keys_to_delete),
+            *(cls.route_db_path_for_prefix(prefix) for prefix in prefixes_to_delete),
+        }
+        if len(routing_paths) != 1:
+            raise ValueError("Storage.mutate_many operations must target the same database")
+
+        if not entries:
+            serialized_entries = []
+        else:
+            serialized_entries = [
+                (
+                    key,
+                    value.model_dump_json() if isinstance(value, BaseModel) else json.dumps(value),
+                    value_type,
+                )
+                for key, value, value_type in entries
+            ]
+
+        await cls._ensure_init()
+        db_path = next(iter(routing_paths))
+
+        from datetime import UTC
+
+        now = datetime.now(UTC).isoformat()
+
+        async def _write() -> int:
+            async with cls.connect(db_path) as db:
+                try:
+                    await db.execute("BEGIN IMMEDIATE")
+                    for key, serialized, value_type in serialized_entries:
+                        await db.execute(
+                            """
+                            INSERT OR REPLACE INTO storage (key, value, type, created_at, updated_at)
+                            VALUES (?, ?, ?,
+                                COALESCE((SELECT created_at FROM storage WHERE key = ?), ?),
+                                ?)
+                            """,
+                            (key, serialized, value_type, key, now, now),
+                        )
+                    deleted = 0
+                    for key in keys_to_delete:
+                        cursor = await db.execute("DELETE FROM storage WHERE key = ?", (key,))
+                        deleted += max(cursor.rowcount, 0)
+                    for prefix in prefixes_to_delete:
+                        cursor = await db.execute(
+                            f"DELETE FROM storage WHERE {cls._like_prefix_clause()}",
+                            (cls._like_prefix_pattern(prefix),),
+                        )
+                        deleted += max(cursor.rowcount, 0)
+                    await db.commit()
+                    return deleted
+                except BaseException:
+                    await db.rollback()
+                    raise
+
+        deleted = await cls._run_with_corruption_recovery(
+            lambda: cls._run_write_with_retry(
+                _write,
+                action="mutate_many",
+                target=f"{len(entries)} sets/{len(keys_to_delete)} keys/{len(prefixes_to_delete)} prefixes",
+            ),
+            db_path=db_path,
+            action="mutate_many",
+        )
+
+        for key in keys_to_delete:
+            cls._invalidate_runtime_caches(key)
+        for prefix in prefixes_to_delete:
+            cls._invalidate_runtime_caches(prefix)
+        cls._log.debug("storage.mutate_many", {
+            "set_count": len(entries),
+            "delete_key_count": len(keys_to_delete),
+            "delete_prefix_count": len(prefixes_to_delete),
+            "deleted": deleted,
+        })
+        return deleted
+
     @classmethod
     async def get(cls, key: str, model: Optional[Type[T]] = None) -> Optional[T | Any]:
         """
diff --git a/flocks/tool/agent/delegate_task.py b/flocks/tool/agent/delegate_task.py
index 88c42d52b..cf00616be 100644
--- a/flocks/tool/agent/delegate_task.py
+++ b/flocks/tool/agent/delegate_task.py
@@ -4,6 +4,8 @@
 
 from __future__ import annotations
 
+import asyncio
+import time
 from typing import Optional, List, Dict, Any
 
 from flocks.tool.registry import (
@@ -26,12 +28,152 @@
 from flocks.agent.registry import is_delegatable
 from flocks.skill.skill import Skill
 from flocks.config.config import Config
-from flocks.tool.subagent_result import format_sync_subagent_result
+from flocks.tool.subagent_result import (
+    _extract_message_error,
+    format_sync_subagent_result,
+)
 from flocks.utils.log import Log
 
 log = Log.create(service="tool.delegate_task")
 
 
+async def _run_subagent_with_hooks(
+    *,
+    ctx: ToolContext,
+    child_session_id: str,
+    child_agent: str,
+    workspace: str,
+    prompt: str,
+    description: str,
+    resumed: bool,
+    provider_id: Optional[str] = None,
+    model_id: Optional[str] = None,
+    callbacks: Optional[Any] = None,
+) -> Any:
+    """Run one child session with paired SubagentStart/SubagentStop hooks."""
+    from flocks.hooks.pipeline import HookPipeline
+
+    common_payload = {
+        "sessionID": ctx.session_id,
+        "workspace": workspace,
+        "parentSessionID": ctx.session_id,
+        "parentMessageID": ctx.message_id,
+        "childSessionID": child_session_id,
+        "agentType": child_agent,
+        "prompt": prompt,
+        "description": description,
+        "resumed": resumed,
+    }
+    try:
+        await HookPipeline.run_subagent_start(common_payload)
+    except Exception as exc:
+        log.debug("delegate_task.hook.subagent_start.error", {
+            "child_session_id": child_session_id,
+            "error": str(exc),
+        })
+
+    started_at = time.perf_counter()
+    try:
+        result = await SessionLoop.run(
+            child_session_id,
+            provider_id=provider_id,
+            model_id=model_id,
+            callbacks=callbacks,
+        )
+    except asyncio.CancelledError:
+        duration_ms = int((time.perf_counter() - started_at) * 1000)
+        try:
+            await HookPipeline.run_subagent_stop({
+                **common_payload,
+                "status": "interrupted",
+                "durationMs": duration_ms,
+                "summary": None,
+                "error": "Sub-agent execution was interrupted",
+            })
+        except Exception as exc:
+            log.debug("delegate_task.hook.subagent_stop.error", {
+                "child_session_id": child_session_id,
+                "error": str(exc),
+            })
+        raise
+    except Exception as exc:
+        duration_ms = int((time.perf_counter() - started_at) * 1000)
+        try:
+            await HookPipeline.run_subagent_stop({
+                **common_payload,
+                "status": "error",
+                "durationMs": duration_ms,
+                "summary": None,
+                "error": str(exc),
+            })
+        except Exception as hook_exc:
+            log.debug("delegate_task.hook.subagent_stop.error", {
+                "child_session_id": child_session_id,
+                "error": str(hook_exc),
+            })
+        raise
+
+    summary = None
+    last_message = getattr(result, "last_message", None)
+    if last_message is not None:
+        try:
+            summary = await Message.get_text_content(last_message)
+        except Exception as exc:
+            log.debug("delegate_task.hook.subagent_summary.error", {
+                "child_session_id": child_session_id,
+                "error": str(exc),
+            })
+    result_error = getattr(result, "error", None)
+    message_error = (
+        _extract_message_error(last_message)
+        if last_message is not None
+        else None
+    )
+    message_finish = (
+        getattr(last_message, "finish", None)
+        if last_message is not None
+        else None
+    )
+    result_metadata = getattr(result, "metadata", None)
+    interrupted = (
+        isinstance(result_metadata, dict)
+        and bool(result_metadata.get("aborted"))
+    )
+    if interrupted:
+        status = "interrupted"
+        stop_error = result_error or "Sub-agent execution was interrupted"
+    elif (
+        getattr(result, "action", None) == "error"
+        or result_error
+        or message_error
+        or message_finish == "error"
+    ):
+        status = "error"
+        stop_error = (
+            result_error
+            or message_error
+            or "Sub-agent execution failed"
+        )
+    else:
+        status = "completed"
+        stop_error = None
+    duration_ms = int((time.perf_counter() - started_at) * 1000)
+    try:
+        await HookPipeline.run_subagent_stop({
+            **common_payload,
+            "status": status,
+            "durationMs": duration_ms,
+            "summary": summary,
+            "error": stop_error,
+        })
+    except Exception as exc:
+        log.debug("delegate_task.hook.subagent_stop.error", {
+            "child_session_id": child_session_id,
+            "error": str(exc),
+        })
+    return result
+
+
 async def _subagent_session_permissions(agent_name: str) -> list:
     """Build session permission rules for a delegated subagent."""
     from flocks.agent.registry import Agent
@@ -395,8 +537,15 @@ async def delegate_task_tool(
             agent=session.agent or ctx.agent,
         )
         from flocks.session.session_loop import LoopCallbacks
-        result = await SessionLoop.run(
-            session.id,
+
+        result = await _run_subagent_with_hooks(
+            ctx=ctx,
+            child_session_id=session.id,
+            child_agent=session.agent or ctx.agent,
+            workspace=getattr(session, "directory", None) or "",
+            prompt=prompt,
+            description=description,
+            resumed=True,
             callbacks=LoopCallbacks(
                 event_publish_callback=ctx.event_publish_callback,
             ),
@@ -494,8 +643,14 @@ async def delegate_task_tool(
         description=description,
     )
     ctx.metadata({"title": description, "metadata": {"sessionId": created.id, "status": "running"}})
-    result = await SessionLoop.run(
-        created.id,
+    result = await _run_subagent_with_hooks(
+        ctx=ctx,
+        child_session_id=created.id,
+        child_agent=agent_to_use,
+        workspace=runtime_directory,
+        prompt=full_prompt,
+        description=description,
+        resumed=False,
         provider_id=(category_model or {}).get("providerID"),
         model_id=(category_model or {}).get("modelID"),
         callbacks=forwarder.build_callbacks(
diff --git a/flocks/tool/file/apply_patch.py b/flocks/tool/file/apply_patch.py
index af46ef698..e7ef476e1 100644
--- a/flocks/tool/file/apply_patch.py
+++ b/flocks/tool/file/apply_patch.py
@@ -293,7 +293,12 @@ async def apply_patch_tool(
         )
     
     sandbox = ctx.extra.get("sandbox") if ctx.extra else None
-    if isinstance(sandbox, dict) and sandbox.get("workspace_access") == "ro":
+    sandbox_read_only = (
+        isinstance(sandbox, dict)
+        and sandbox.get("workspace_access") == "ro"
+    )
+    execution_mode = ctx.extra.get("execution_mode") if ctx.extra else None
+    if sandbox_read_only and execution_mode != "plan":
         return ToolResult(
             success=False,
             error=(
@@ -397,6 +402,25 @@ async def apply_patch_tool(
                 success=False,
                 error=f"Failed to process hunk for {hunk.path}: {str(e)}"
             )
+
+    if sandbox_read_only:
+        from flocks.session.execution_mode import is_plan_file_edit
+
+        if not all(
+            is_plan_file_edit(execution_mode, ctx, change["filePath"])
+            and (
+                not change.get("movePath")
+                or is_plan_file_edit(execution_mode, ctx, change["movePath"])
+            )
+            for change in file_changes
+        ):
+            return ToolResult(
+                success=False,
+                error=(
+                    "Patch is blocked in sandbox read-only workspace mode. "
+                    "Only the current session plan file may be changed in Plan mode."
+                ),
+            )
     
     # Request permission
     await ctx.ask(
diff --git a/flocks/tool/file/edit.py b/flocks/tool/file/edit.py
index abc0566cc..377ab212d 100644
--- a/flocks/tool/file/edit.py
+++ b/flocks/tool/file/edit.py
@@ -526,6 +526,20 @@ async def edit_tool(
 
     sandbox = ctx.extra.get("sandbox") if ctx.extra else None
     if isinstance(sandbox, dict) and sandbox.get("workspace_access") == "ro":
+        from flocks.session.execution_mode import is_plan_file_edit
+
+        plan_file_edit = is_plan_file_edit(
+            ctx.extra.get("execution_mode"),
+            ctx,
+            filepath,
+        )
+    else:
+        plan_file_edit = False
+    if (
+        isinstance(sandbox, dict)
+        and sandbox.get("workspace_access") == "ro"
+        and not plan_file_edit
+    ):
         return ToolResult(
             success=False,
             error=(
diff --git a/flocks/tool/file/write.py b/flocks/tool/file/write.py
index eee0379cf..50aee858e 100644
--- a/flocks/tool/file/write.py
+++ b/flocks/tool/file/write.py
@@ -273,6 +273,20 @@ async def write_tool(
 
     sandbox = ctx.extra.get("sandbox") if ctx.extra else None
     if isinstance(sandbox, dict) and sandbox.get("workspace_access") == "ro":
+        from flocks.session.execution_mode import is_plan_file_edit
+
+        plan_file_edit = is_plan_file_edit(
+            ctx.extra.get("execution_mode"),
+            ctx,
+            filepath,
+        )
+    else:
+        plan_file_edit = False
+    if (
+        isinstance(sandbox, dict)
+        and sandbox.get("workspace_access") == "ro"
+        and not plan_file_edit
+    ):
         return ToolResult(
             success=False,
             error=(
diff --git a/flocks/tool/registry.py b/flocks/tool/registry.py
index b172a847d..35d1c3cd7 100644
--- a/flocks/tool/registry.py
+++ b/flocks/tool/registry.py
@@ -267,6 +267,20 @@ async def ask(
             always: Always-allow patterns
             metadata: Additional metadata
         """
+        execution_mode = self.extra.get("execution_mode")
+        if execution_mode:
+            from flocks.session.execution_mode import is_permission_allowed
+
+            if not is_permission_allowed(
+                execution_mode,
+                permission,
+                patterns,
+                self,
+            ):
+                raise PermissionError(
+                    "Plan mode may only edit the current session plan file."
+                )
+
         request = PermissionRequest(
             permission=permission,
             patterns=patterns,
@@ -870,6 +884,40 @@ async def execute(
                 message_id="default"
             )
 
+        execution_mode = ctx.extra.get("execution_mode")
+        if execution_mode:
+            from flocks.session.execution_mode import (
+                is_tool_allowed,
+                tool_call_denial_reason,
+            )
+
+            if not is_tool_allowed(execution_mode, tool_name):
+                log.warn("tool.execute.execution_mode_denied", {
+                    "name": tool_name,
+                    "execution_mode": str(execution_mode),
+                    "session_id": ctx.session_id,
+                })
+                return ToolResult(
+                    success=False,
+                    error=(
+                        f"Tool {tool_name!r} is not available in "
+                        f"{str(execution_mode)!r} execution mode."
+                    ),
+                )
+            denial_reason = tool_call_denial_reason(
+                execution_mode,
+                tool_name,
+                kwargs,
+                ctx,
+            )
+            if denial_reason:
+                log.warn("tool.execute.execution_mode_call_denied", {
+                    "name": tool_name,
+                    "execution_mode": str(execution_mode),
+                    "session_id": ctx.session_id,
+                })
+                return ToolResult(success=False, error=denial_reason)
+
         log.info("tool.execute", {
             "name": tool_name,
             "params": list(kwargs.keys()),
@@ -1616,7 +1664,7 @@ def _register_builtin_tools(cls) -> None:
             # security/ — SSH forensics + threat intelligence (optional: asyncssh)
             ("flocks.tool.security", ["ssh_host_cmd", "ssh_run_script"]),
             # system/ — questions, model config, memory, MCP management, session management, slash commands
-            ("flocks.tool.system", ["question", "model_config", "memory", "flocks_mcp", "session_manage", "slash_command", "tool_search"]),
+            ("flocks.tool.system", ["question", "plan_exit", "model_config", "memory", "flocks_mcp", "session_manage", "slash_command", "tool_search"]),
             # skill/ — skill management (search, install, status, deps, remove, load)
             ("flocks.tool.skill", ["flocks_skills", "skill_load"]),
             # device/ — security device asset context and status probes
@@ -1650,16 +1698,43 @@ def _register_builtin_tools(cls) -> None:
         if "get_time" not in cls._tools:
             @cls.register_function(
                 name="get_time",
-                description="Get current date and time",
+                description="Get current date and time in ISO 8601 or Unix timestamp format",
                 category=ToolCategory.SYSTEM,
                 native=True,
-                parameters=[]
+                parameters=[
+                    ToolParameter(
+                        name="format",
+                        type=ParameterType.STRING,
+                        description=(
+                            "Output format: 'iso' for ISO 8601, 'unix' for Unix seconds, "
+                            "or 'unix_ms' for Unix milliseconds. Defaults to 'iso'."
+                        ),
+                        required=False,
+                        default="iso",
+                        enum=["iso", "unix", "unix_ms"],
+                    )
+                ]
             )
-            async def get_time(ctx: ToolContext) -> ToolResult:
+            async def get_time(ctx: ToolContext, format: str = "iso") -> ToolResult:
                 from datetime import datetime
+
+                if format not in {"iso", "unix", "unix_ms"}:
+                    return ToolResult(
+                        success=False,
+                        error="format must be one of: iso, unix, unix_ms",
+                    )
+
+                now = datetime.now()
+                if format == "unix":
+                    output = str(int(now.timestamp()))
+                elif format == "unix_ms":
+                    output = str(int(now.timestamp() * 1000))
+                else:
+                    output = now.isoformat()
+
                 return ToolResult(
                     success=True,
-                    output=datetime.now().isoformat()
+                    output=output,
                 )
 
     @classmethod
diff --git a/flocks/tool/system/plan_exit.py b/flocks/tool/system/plan_exit.py
new file mode 100644
index 000000000..2dacda008
--- /dev/null
+++ b/flocks/tool/system/plan_exit.py
@@ -0,0 +1,186 @@
+"""Plan completion and Build handoff modeled after OpenCode's plan_exit tool."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from flocks.session.execution_mode import SessionExecutionMode
+from flocks.session.message import Message, MessageRole
+from flocks.session.plan_file import context_plan_file
+from flocks.tool.registry import (
+    ToolCategory,
+    ToolContext,
+    ToolRegistry,
+    ToolResult,
+)
+from flocks.tool.system.question import question_tool
+
+
+START_IMPLEMENTING = "开始实施"
+CONTINUE_PLANNING = "调整计划"
+
+DESCRIPTION = """Finish a completed plan and ask the user whether to implement it.
+
+Call this only after presenting a decision-complete plan. Approval starts a new
+Build turn in the current session loop to implement the approved plan.
+Declining keeps the session in Plan so the plan can be refined.
+"""
+
+
+async def _publish(ctx: ToolContext, event_type: str, properties: dict[str, Any]) -> None:
+    if ctx.event_publish_callback:
+        await ctx.event_publish_callback(event_type, properties)
+
+
+async def _turn_model_and_variant(
+    ctx: ToolContext,
+) -> tuple[dict[str, str] | None, str | None]:
+    """Copy the active Plan turn model like OpenCode's synthetic Build turn."""
+
+    try:
+        messages = await Message.list(ctx.session_id)
+        last_user = next(
+            (
+                message
+                for message in reversed(messages)
+                if getattr(message, "role", None) == MessageRole.USER
+            ),
+            None,
+        )
+    except Exception:
+        last_user = None
+
+    model = getattr(last_user, "model", None)
+    if not isinstance(model, dict) or not all(
+        model.get(key) for key in ("providerID", "modelID")
+    ):
+        model = ctx.extra.get("model")
+    if not isinstance(model, dict) or not all(
+        model.get(key) for key in ("providerID", "modelID")
+    ):
+        model = None
+    return model, getattr(last_user, "variant", None)
+
+
+@ToolRegistry.register_function(
+    name="plan_exit",
+    description=DESCRIPTION,
+    category=ToolCategory.SYSTEM,
+    parameters=[],
+)
+async def plan_exit_tool(ctx: ToolContext) -> ToolResult:
+    """Ask for plan approval and continue immediately in Build mode."""
+
+    plan = context_plan_file(ctx)
+    if plan is None or not plan.path.is_file():
+        return ToolResult(
+            success=False,
+            error="Write the session plan file before calling plan_exit.",
+        )
+    try:
+        if not plan.path.read_text(encoding="utf-8").strip():
+            return ToolResult(
+                success=False,
+                error="The session plan file is empty. Complete it before calling plan_exit.",
+            )
+    except OSError as exc:
+        return ToolResult(success=False, error=f"Could not read the session plan file: {exc}")
+
+    confirmation = await question_tool(
+        ctx,
+        questions=[
+            {
+                "header": "Plan complete",
+                "question": (
+                    f"The plan at {plan.relative_path} is complete. Would you like "
+                    "to switch to Build and start implementing?"
+                ),
+                "type": "choice",
+                "options": [
+                    {
+                        "label": START_IMPLEMENTING,
+                        "description": "Switch to Build and implement the approved plan now.",
+                    },
+                    {
+                        "label": CONTINUE_PLANNING,
+                        "description": "Stay in Plan and describe what should be changed.",
+                        "allowText": True,
+                    },
+                ],
+                "multiple": False,
+                "custom": False,
+            }
+        ],
+    )
+    if not confirmation.success:
+        return confirmation
+    if confirmation.metadata.get("deferred"):
+        return confirmation
+
+    answers = confirmation.metadata.get("answers") or []
+    selected = answers[0] if answers else []
+    if START_IMPLEMENTING not in selected:
+        feedback = "\n".join(
+            str(value).strip()
+            for value in selected
+            if str(value).strip() and value != CONTINUE_PLANNING
+        )
+        output = (
+            "The user chose to remain in Plan. Continue refining the plan "
+            "using their feedback."
+        )
+        metadata = {
+            "approved": False,
+            "executionMode": SessionExecutionMode.PLAN.value,
+        }
+        if feedback:
+            output = f"{output}\n\nUser feedback:\n{feedback}"
+            metadata["feedback"] = feedback
+        return ToolResult(
+            success=True,
+            output=output,
+            title="Continue planning",
+            metadata=metadata,
+        )
+
+    build_model, build_variant = await _turn_model_and_variant(ctx)
+    build_message = await Message.create(
+        session_id=ctx.session_id,
+        role=MessageRole.USER,
+        content=(
+            f"The plan at {plan.relative_path} has been approved. "
+            "Switch to Build mode, read that file, and implement it now."
+        ),
+        agent=ctx.agent,
+        model=build_model,
+        variant=build_variant,
+        executionMode=SessionExecutionMode.BUILD,
+        synthetic=True,
+        part_metadata={
+            "planImplementation": True,
+            "planPath": plan.relative_path,
+        },
+    )
+    await _publish(
+        ctx,
+        "session.execution_mode.changed",
+        {
+            "sessionID": ctx.session_id,
+            "executionMode": SessionExecutionMode.BUILD.value,
+            "reason": "plan-approved",
+        },
+    )
+    return ToolResult(
+        success=True,
+        output=(
+            "The plan was approved. Continue immediately in Build mode and "
+            "implement the approved plan."
+        ),
+        title="Plan approved",
+        metadata={
+            "approved": True,
+            "executionMode": SessionExecutionMode.BUILD.value,
+            "buildMessageID": build_message.id,
+            "planPath": plan.relative_path,
+        },
+    )
diff --git a/flocks/tool/system/question.py b/flocks/tool/system/question.py
index 4fb80273e..e83af3201 100644
--- a/flocks/tool/system/question.py
+++ b/flocks/tool/system/question.py
@@ -107,7 +107,7 @@ def _first_non_empty_string(data: Dict[str, Any], keys: tuple[str, ...]) -> str:
     return ""
 
 
-def normalize_question_option(opt: Any) -> Optional[Dict[str, str]]:
+def normalize_question_option(opt: Any) -> Optional[Dict[str, Any]]:
     """Normalize LLM-produced choice options into the UI's label/description shape."""
     if isinstance(opt, str):
         label = opt.strip()
@@ -123,7 +123,13 @@ def normalize_question_option(opt: Any) -> Optional[Dict[str, str]]:
         label, description = description, ""
     if not label:
         return None
-    return {"label": label, "description": description}
+    normalized: Dict[str, Any] = {
+        "label": label,
+        "description": description,
+    }
+    if opt.get("allowText") is True:
+        normalized["allowText"] = True
+    return normalized
 
 
 def _format_channel_question_text(questions: List[Dict[str, Any]]) -> str:
@@ -296,6 +302,13 @@ async def default_question_handler(
                                         "properties": {
                                             "label": {"type": "string"},
                                             "description": {"type": "string"},
+                                            "allowText": {
+                                                "type": "boolean",
+                                                "description": (
+                                                    "Show a text input for this option and "
+                                                    "return both its label and entered text."
+                                                ),
+                                            },
                                         },
                                         "required": ["label"],
                                         "additionalProperties": False,
diff --git a/flocks/tool/system/session_manage.py b/flocks/tool/system/session_manage.py
index 7bf2968d2..9eb070f67 100644
--- a/flocks/tool/system/session_manage.py
+++ b/flocks/tool/system/session_manage.py
@@ -40,8 +40,8 @@
   model, provider, and memory_enabled. For requests like "change this session to
   gpt-5" or "update the session model", use action=update with model, and set
   provider too when the user specifies one.
-- delete: soft-delete a session and its child sessions; requires session_id and
-  confirmation.
+- delete: permanently delete a session, its child sessions, messages, and
+  history; requires session_id and confirmation. This cannot be restored.
 - archive: archive or restore a session; requires session_id. Set archive=false
   to restore an archived session to active.
 """
@@ -221,7 +221,11 @@ async def session_manage(
             permission="session_manage",
             patterns=[f"delete:{session_id}"],
             always=[],
-            metadata={"action": "delete", "session_id": session_id},
+            metadata={
+                "action": "permanent_delete",
+                "session_id": session_id,
+                "destructive": True,
+            },
         )
         return await _session_delete_impl(ctx, session_id=session_id)
     if action == "archive":
@@ -482,7 +486,7 @@ async def _session_delete_impl(ctx: ToolContext, session_id: str) -> ToolResult:
 
     return ToolResult(
         success=True,
-        output=f"Session '{session_id}'（{session.title}）已删除",
+        output=f"Session '{session_id}'（{session.title}）已永久删除",
     )
 
 
@@ -495,27 +499,32 @@ async def _session_archive_impl(
     session_id: str,
     archive: Optional[bool] = True,
 ) -> ToolResult:
-    from flocks.session.session import Session, SessionInfo
-    from flocks.storage.storage import Storage
+    from flocks.session.session import Session
 
-    # get_by_id 会跳过 archived session，需直接扫 Storage
-    session = None
-    keys = await Storage.list_keys(prefix="session:")
-    for key in keys:
-        try:
-            s = await Storage.get(key, SessionInfo)
-            if s and s.id == session_id and s.status != "deleted":
-                session = s
-                break
-        except Exception:
-            continue
+    session = await Session.get_by_id_unfiltered(session_id)
 
     if not session:
         return ToolResult(success=False, error=f"未找到 session '{session_id}'")
 
+    if archive is not False and session_id == ctx.session_id:
+        return ToolResult(
+            success=False,
+            error="不能在当前会话的工具调用尚未结束时归档自身，请从工作台归档该会话",
+        )
+    if archive is False and getattr(session, "parent_id", None) is not None:
+        return ToolResult(
+            success=False,
+            error="只能从根任务恢复完整任务树",
+        )
+
     try:
         if archive is False:
-            ok = await Session.unarchive(session.project_id, session_id)
+            owner_user_id = getattr(session, "owner_user_id", None)
+            ok = await Session.restore(
+                session.project_id,
+                session_id,
+                project_owner_id=owner_user_id,
+            )
             action = "取消归档"
         else:
             ok = await Session.archive(session.project_id, session_id)
diff --git a/flocks/tool/task/todo.py b/flocks/tool/task/todo.py
index a9e5f04b0..dd275858d 100644
--- a/flocks/tool/task/todo.py
+++ b/flocks/tool/task/todo.py
@@ -237,9 +237,12 @@ async def todo_tool(
     old_todos = await Todo.get(ctx.session_id)
     normalized_todos = _normalize_todos(todos)
     if _all_terminal(normalized_todos):
-        await Todo.update(ctx.session_id, [])
+        await Todo.update_active(ctx.session_id, [])
     else:
-        await Todo.update(ctx.session_id, normalized_todos)
+        await Todo.update_active(
+            ctx.session_id,
+            normalized_todos,
+        )
 
     old_serialized = _serialize_todos(old_todos)
     new_serialized = _serialize_todos(normalized_todos)
diff --git a/flocks/utils/langfuse.py b/flocks/utils/langfuse.py
index 53564a43b..52d7939f9 100644
--- a/flocks/utils/langfuse.py
+++ b/flocks/utils/langfuse.py
@@ -42,17 +42,47 @@ def end(self, **_: Any) -> None:
     "langfuse_current_observation",
     default=None,
 )
+_CAPTURE_MODE_ENV = "FLOCKS_LANGFUSE_CAPTURE_MODE"
+_MAX_CHARS_ENV = "FLOCKS_LANGFUSE_MAX_CHARS"
+_DEFAULT_CAPTURE_MODE = "full"
+_DEFAULT_MAX_CHARS = 8000
+_VALID_CAPTURE_MODES = {"full", "truncated"}
+_TRACE_NAME_ATTR = "langfuse.trace.name"
+_TRACE_USER_ID_ATTR = "user.id"
+_TRACE_SESSION_ID_ATTR = "session.id"
+_TRACE_TAGS_ATTR = "langfuse.trace.tags"
 
 
 def _filter_none(values: Dict[str, Any]) -> Dict[str, Any]:
     return {k: v for k, v in values.items() if v is not None}
 
 
-def _truncate_value(value: Any, max_chars: int = 8000) -> Any:
+def _get_capture_mode() -> str:
+    mode = os.getenv(_CAPTURE_MODE_ENV, _DEFAULT_CAPTURE_MODE).strip().lower()
+    if mode in _VALID_CAPTURE_MODES:
+        return mode
+    return _DEFAULT_CAPTURE_MODE
+
+
+def _get_capture_max_chars() -> int:
+    raw_value = os.getenv(_MAX_CHARS_ENV, str(_DEFAULT_MAX_CHARS)).strip()
+    try:
+        max_chars = int(raw_value)
+    except (TypeError, ValueError):
+        return _DEFAULT_MAX_CHARS
+    if max_chars <= 0:
+        return _DEFAULT_MAX_CHARS
+    return max_chars
+
+
+def _truncate_value(value: Any, max_chars: Optional[int] = None) -> Any:
     if isinstance(value, str):
-        if len(value) <= max_chars:
+        if _get_capture_mode() == "full":
             return value
-        return value[:max_chars] + f"...[truncated:{len(value) - max_chars}]"
+        limit = max_chars if max_chars is not None else _get_capture_max_chars()
+        if len(value) <= limit:
+            return value
+        return value[:limit] + f"...[truncated:{len(value) - limit}]"
     return value
 
 
@@ -67,6 +97,58 @@ def _sanitize_payload(value: Any) -> Any:
     return _truncate_value(value)
 
 
+def _set_trace_attribute(observation: Any, key: str, value: Any) -> None:
+    otel_span = getattr(observation, "_otel_span", None)
+    if otel_span is None or not hasattr(otel_span, "is_recording"):
+        return
+    try:
+        if otel_span.is_recording():
+            otel_span.set_attribute(key, value)
+    except Exception:
+        return
+
+
+def _propagate_trace_dimensions(
+    observation: Any,
+    *,
+    trace_name: Optional[str] = None,
+    session_id: Optional[str] = None,
+    user_id: Optional[str] = None,
+    tags: Optional[list[str]] = None,
+    parent: Any = None,
+) -> Any:
+    inherited_trace_name = trace_name
+    inherited_session_id = session_id
+    inherited_user_id = user_id
+    inherited_tags = tags
+
+    parent_span = getattr(parent, "_otel_span", None)
+    parent_attrs = getattr(parent_span, "attributes", None) if parent_span is not None else None
+    if parent_attrs:
+        if inherited_trace_name is None:
+            inherited_trace_name = parent_attrs.get(_TRACE_NAME_ATTR)
+        if inherited_session_id is None:
+            inherited_session_id = parent_attrs.get(_TRACE_SESSION_ID_ATTR)
+        if inherited_user_id is None:
+            inherited_user_id = parent_attrs.get(_TRACE_USER_ID_ATTR)
+        if inherited_tags is None:
+            parent_tags = parent_attrs.get(_TRACE_TAGS_ATTR)
+            if isinstance(parent_tags, list):
+                inherited_tags = parent_tags
+            elif isinstance(parent_tags, tuple):
+                inherited_tags = list(parent_tags)
+
+    if inherited_trace_name is not None:
+        _set_trace_attribute(observation, _TRACE_NAME_ATTR, inherited_trace_name)
+    if inherited_session_id is not None:
+        _set_trace_attribute(observation, _TRACE_SESSION_ID_ATTR, inherited_session_id)
+    if inherited_user_id is not None:
+        _set_trace_attribute(observation, _TRACE_USER_ID_ATTR, inherited_user_id)
+    if inherited_tags is not None:
+        _set_trace_attribute(observation, _TRACE_TAGS_ATTR, inherited_tags)
+    return observation
+
+
 def initialize() -> None:
     """Initialize Langfuse client once (no-op when unavailable)."""
     global _client, _initialized
@@ -140,26 +222,48 @@ def create_trace(
     )
     try:
         # Old SDKs may expose .trace(), newer SDKs are OTEL-native and expose
-        # .start_span() / .start_observation().
+        # .start_observation().
         if hasattr(client, "trace"):
             return client.trace(**payload)
-        trace_obs = client.start_span(
-            name=name,
-            input=payload.get("input"),
-            metadata=payload.get("metadata"),
-        )
-        # Best-effort enrich current trace with session/user dimensions.
-        try:
-            client.update_current_trace(
+        if hasattr(client, "start_observation"):
+            trace_metadata = dict(payload.get("metadata") or {})
+            if session_id is not None:
+                trace_metadata.setdefault("session_id", session_id)
+            if user_id is not None:
+                trace_metadata.setdefault("user_id", user_id)
+            if tags is not None:
+                trace_metadata.setdefault("tags", tags)
+            trace_obs = client.start_observation(
+                name=name,
+                as_type="span",
+                input=payload.get("input"),
+                metadata=trace_metadata or None,
+            )
+            return _propagate_trace_dimensions(
+                trace_obs,
+                trace_name=name,
                 session_id=session_id,
                 user_id=user_id,
                 tags=tags,
+            )
+        if hasattr(client, "start_span"):
+            trace_obs = client.start_span(
+                name=name,
                 input=payload.get("input"),
                 metadata=payload.get("metadata"),
             )
-        except Exception:
-            pass
-        return trace_obs
+            # Best-effort enrich current trace with session/user dimensions.
+            try:
+                client.update_current_trace(
+                    session_id=session_id,
+                    user_id=user_id,
+                    tags=tags,
+                    input=payload.get("input"),
+                    metadata=payload.get("metadata"),
+                )
+            except Exception:
+                pass
+            return trace_obs
     except Exception as exc:
         log.warn("langfuse.trace_failed", {"error": str(exc), "name": name})
         return _NoopObservation("trace")
@@ -189,15 +293,21 @@ def create_generation(
     )
     try:
         if parent and hasattr(parent, "generation"):
-            return parent.generation(**payload)
+            return _propagate_trace_dimensions(parent.generation(**payload), parent=parent)
         if parent and hasattr(parent, "start_generation"):
-            return parent.start_generation(**payload)
+            return _propagate_trace_dimensions(parent.start_generation(**payload), parent=parent)
         if parent and hasattr(parent, "start_observation"):
-            return parent.start_observation(as_type="generation", **payload)
+            return _propagate_trace_dimensions(
+                parent.start_observation(as_type="generation", **payload),
+                parent=parent,
+            )
         if hasattr(client, "start_generation"):
-            return client.start_generation(**payload)
+            return _propagate_trace_dimensions(client.start_generation(**payload), parent=parent)
         if hasattr(client, "start_observation"):
-            return client.start_observation(as_type="generation", **payload)
+            return _propagate_trace_dimensions(
+                client.start_observation(as_type="generation", **payload),
+                parent=parent,
+            )
     except Exception as exc:
         log.warn("langfuse.generation_failed", {"error": str(exc), "name": name})
     return _NoopObservation("generation")
@@ -225,15 +335,21 @@ def create_span(
     )
     try:
         if parent and hasattr(parent, "span"):
-            return parent.span(**payload)
+            return _propagate_trace_dimensions(parent.span(**payload), parent=parent)
         if parent and hasattr(parent, "start_span"):
-            return parent.start_span(**payload)
+            return _propagate_trace_dimensions(parent.start_span(**payload), parent=parent)
         if parent and hasattr(parent, "start_observation"):
-            return parent.start_observation(as_type="span", **payload)
+            return _propagate_trace_dimensions(
+                parent.start_observation(as_type="span", **payload),
+                parent=parent,
+            )
         if hasattr(client, "start_span"):
-            return client.start_span(**payload)
+            return _propagate_trace_dimensions(client.start_span(**payload), parent=parent)
         if hasattr(client, "start_observation"):
-            return client.start_observation(as_type="span", **payload)
+            return _propagate_trace_dimensions(
+                client.start_observation(as_type="span", **payload),
+                parent=parent,
+            )
     except Exception as exc:
         log.warn("langfuse.span_failed", {"error": str(exc), "name": name})
     return _NoopObservation("span")
@@ -267,20 +383,32 @@ def end_observation(
             observation.end(**payload)
             return
     except TypeError:
-        try:
-            observation.end()
-            return
-        except Exception:
-            pass
+        pass
     except Exception:
         pass
 
     try:
         if hasattr(observation, "update"):
-            observation.update(**payload)
+            update_payload = dict(payload)
+            if usage:
+                update_payload["usage_details"] = usage
+            try:
+                observation.update(**update_payload)
+            except TypeError:
+                fallback_payload = dict(payload)
+                if usage:
+                    fallback_payload["usage"] = usage
+                    fallback_payload.pop("usage_details", None)
+                observation.update(**fallback_payload)
     except Exception as exc:
         log.debug("langfuse.end_fallback_update_failed", {"error": str(exc)})
 
+    try:
+        if hasattr(observation, "end"):
+            observation.end()
+    except Exception as exc:
+        log.debug("langfuse.end_fallback_end_failed", {"error": str(exc)})
+
 
 def is_active() -> bool:
     """Return True when Langfuse is initialized and has a live client."""
diff --git a/flocks/workflow/tool_context.py b/flocks/workflow/tool_context.py
index e06ad3b42..b803f9810 100644
--- a/flocks/workflow/tool_context.py
+++ b/flocks/workflow/tool_context.py
@@ -127,23 +127,7 @@ async def cleanup_workflow_tool_context(tool_context: Optional[ToolContext]) ->
         if extra.get("workflow_child_session_created") is True:
             return False
 
-        await Session.delete(session.project_id, session.id)
-
-        # Session.delete is intentionally a soft delete. Trigger parents without
-        # child tasks are implementation details, so remove their residual rows
-        # to keep high-frequency trigger traffic storage-bounded.
-        from flocks.storage.storage import Storage
-
-        await Storage.delete(f"session:{session.project_id}:{session.id}")
-        await Storage.delete(f"message:{session.id}")
-        await Storage.delete(f"todo:{session.id}")
-        await Storage.delete(f"goal:{session.id}")
-        await Storage.delete(f"session_diff:{session.id}")
-        await Storage.clear(prefix=f"message_diff:{session.id}:")
-        await Storage.clear(prefix=f"system_prompts:{session.id}:")
-
-        Message.invalidate_cache(session.id)
-        return True
+        return await Session.delete(session.project_id, session.id)
     except Exception as exc:
         log.warning(
             "workflow.tool_context.cleanup_failed",
diff --git a/pyproject.toml b/pyproject.toml
index 939c3b233..384249fa7 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "flocks"
-version = "v2026.7.23"
+version = "v2026.7.29"
 description = "AI-Native SecOps platform with multi-agent collaboration"
 authors = [
     {name = "Flocks Team", email = "team@example.com"}
diff --git a/scripts/dev.sh b/scripts/dev.sh
index 80474efb6..03a4fab5b 100644
--- a/scripts/dev.sh
+++ b/scripts/dev.sh
@@ -25,7 +25,6 @@ if [ "${BACKEND_ACCESS_HOST}" = "0.0.0.0" ] || [ "${BACKEND_ACCESS_HOST}" = "::"
 fi
 
 BACKEND_BASE_URL="http://${BACKEND_ACCESS_HOST}:${BACKEND_PORT}"
-BACKEND_WS_URL="ws://${BACKEND_ACCESS_HOST}:${BACKEND_PORT}"
 BACKEND_PID=""
 BACKEND_PGID=""
 CLEANUP_DONE=0
@@ -223,8 +222,7 @@ start_frontend() {
 
     echo -e "${GREEN}🎨 启动前端服务: http://${FRONTEND_HOST}:${FRONTEND_PORT}${NC}"
     cd "${PROJECT_ROOT}/webui"
-    VITE_API_BASE_URL="${BACKEND_BASE_URL}" \
-    VITE_WS_BASE_URL="${BACKEND_WS_URL}" \
+    FLOCKS_API_PROXY_TARGET="${BACKEND_BASE_URL}" \
     npm run dev -- --host "${FRONTEND_HOST}" --port "${FRONTEND_PORT}"
 }
 
diff --git a/scripts/validate_flockshub.py b/scripts/validate_flockshub.py
index 7068f7cdc..ffdd3208e 100644
--- a/scripts/validate_flockshub.py
+++ b/scripts/validate_flockshub.py
@@ -62,6 +62,8 @@ def main() -> int:
         manifest = load_json(manifest_path)
         if manifest.get("id") != plugin_id or manifest.get("type") != plugin_type:
             fail(f"Manifest id/type mismatch: {manifest_rel}")
+        if manifest.get("version") != entry.get("version"):
+            fail(f"Index/manifest version mismatch: {manifest_rel}")
         if manifest.get("category") not in categories:
             fail(f"Unknown category in {manifest_rel}: {manifest.get('category')}")
         unknown_tags = set(manifest.get("tags", [])) - tags
@@ -75,6 +77,11 @@ def main() -> int:
             fail(f"Unknown risk level in {manifest_rel}: {risk_level}")
 
         package_dir = manifest_path.parent
+        workspace_path = package_dir / "workspace.json"
+        if plugin_type == "webui" and workspace_path.is_file():
+            workspace = load_json(workspace_path)
+            if workspace.get("version") != manifest.get("version"):
+                fail(f"WebUI workspace/manifest version mismatch: {manifest_rel}")
         for entrypoint in manifest.get("entrypoints", []):
             ensure_relative(entrypoint)
             if not (package_dir / entrypoint).exists():
diff --git a/tests/agent/test_unified_session_loop.py b/tests/agent/test_unified_session_loop.py
index 255882ccd..44aaa8fa0 100644
--- a/tests/agent/test_unified_session_loop.py
+++ b/tests/agent/test_unified_session_loop.py
@@ -8,6 +8,8 @@
 4. _resolve_model implements 5-level priority correctly
 """
 
+import asyncio
+
 import pytest
 from unittest.mock import AsyncMock, MagicMock, patch
 from dataclasses import dataclass
@@ -279,7 +281,226 @@ async def test_process_session_message_pins_explicit_request_model(self, monkeyp
             provider="anthropic",
             model="claude-sonnet-4-5",
             model_pinned=True,
+            model_auto=False,
+        )
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("category", ["user", "entity-config", "workflow"])
+    async def test_webui_auto_uses_default_primary_and_returns_actual_model(
+        self,
+        monkeypatch,
+        category,
+    ):
+        """Supported WebUI chats use Auto and report the recovered model."""
+        from types import SimpleNamespace
+
+        from flocks.server.routes import session as session_routes
+        from flocks.session.session_loop import LoopResult, SessionLoop
+
+        request = session_routes.PromptRequest(
+            parts=[{"type": "text", "text": "hello"}],
         )
+        session = SimpleNamespace(
+            id="ses_auto",
+            project_id="proj",
+            directory="/tmp/project",
+            agent="rex",
+            provider="stale",
+            model="stale-model",
+            model_pinned=False,
+            model_auto=True,
+            category=category,
+        )
+        agent = SimpleNamespace(
+            name="rex",
+            model={"providerID": "agent-provider", "modelID": "agent-model"},
+        )
+        fallback_message = SimpleNamespace(
+            id="msg_assistant",
+            providerID="fallback",
+            modelID="fallback-model",
+            finish="stop",
+            tokens=None,
+        )
+        loop_run = AsyncMock(return_value=LoopResult(
+            action="stop",
+            last_message=fallback_message,
+            provider_id="fallback",
+            model_id="fallback-model",
+        ))
+        message_create = AsyncMock(return_value=SimpleNamespace(id="msg_user"))
+        context_usage_update = AsyncMock()
+        title_generation = AsyncMock()
+
+        monkeypatch.setattr(session_routes, "_require_agent_usable_for_chat", AsyncMock())
+        monkeypatch.setattr(
+            "flocks.agent.registry.Agent.default_agent",
+            AsyncMock(return_value="rex"),
+        )
+        monkeypatch.setattr("flocks.agent.registry.Agent.get", AsyncMock(return_value=agent))
+        monkeypatch.setattr(
+            "flocks.config.config.Config.resolve_default_llm",
+            AsyncMock(return_value={
+                "provider_id": "primary",
+                "model_id": "primary-model",
+            }),
+        )
+        monkeypatch.setattr(
+            "flocks.config.config.Config.get",
+            AsyncMock(return_value=SimpleNamespace()),
+        )
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_runtime_model",
+            AsyncMock(return_value=(True, "available")),
+        )
+        monkeypatch.setattr(SessionLoop, "run", loop_run)
+        monkeypatch.setattr("flocks.provider.provider.Provider._ensure_initialized", lambda: None)
+        monkeypatch.setattr("flocks.provider.provider.Provider.apply_config", AsyncMock())
+        monkeypatch.setattr("flocks.provider.provider.Provider.get", lambda _provider_id: object())
+        monkeypatch.setattr("flocks.tool.registry.ToolRegistry.init", lambda: None)
+        monkeypatch.setattr(
+            "flocks.session.lifecycle.revert.SessionRevert.cleanup",
+            AsyncMock(),
+        )
+        monkeypatch.setattr("flocks.session.message.Message.create", message_create)
+        monkeypatch.setattr(
+            "flocks.session.message.Message.get_text_content",
+            AsyncMock(return_value="recovered"),
+        )
+        monkeypatch.setattr("flocks.session.message.Message.parts", AsyncMock(return_value=[]))
+        monkeypatch.setattr("flocks.server.routes.event.publish_event", AsyncMock())
+        monkeypatch.setattr(
+            session_routes,
+            "_publish_context_usage_update",
+            context_usage_update,
+        )
+        monkeypatch.setattr(
+            "flocks.session.lifecycle.title.SessionTitle.generate_title_after_first_message",
+            title_generation,
+        )
+
+        response = await session_routes._process_session_message(
+            session.id,
+            session,
+            request,
+            session.directory,
+        )
+
+        assert message_create.await_args.kwargs["model"] == {
+            "providerID": "primary",
+            "modelID": "primary-model",
+        }
+        assert loop_run.await_args.kwargs["auto_failover"] is True
+        assert loop_run.await_args.kwargs["provider_id"] == "primary"
+        assert response["info"]["providerID"] == "fallback"
+        assert response["info"]["modelID"] == "fallback-model"
+        assert context_usage_update.await_args.kwargs["provider_id"] == "fallback"
+        assert context_usage_update.await_args.kwargs["model_id"] == "fallback-model"
+        await asyncio.sleep(0)
+        assert title_generation.await_args.kwargs["provider_id"] == "fallback"
+        assert title_generation.await_args.kwargs["model_id"] == "fallback-model"
+
+    @pytest.mark.asyncio
+    async def test_unsupported_session_ignores_auto_flag(self, monkeypatch):
+        """A corrupt task Auto flag cannot activate WebUI Auto."""
+        from types import SimpleNamespace
+
+        from flocks.server.routes import session as session_routes
+        from flocks.session.session_loop import SessionLoop
+
+        request = session_routes.PromptRequest(
+            parts=[{"type": "text", "text": "task input"}],
+            noReply=True,
+        )
+        session = SimpleNamespace(
+            id="ses_task",
+            project_id="proj",
+            directory="/tmp/project",
+            agent="rex",
+            provider="direct",
+            model="direct-model",
+            model_pinned=True,
+            model_auto=True,
+            category="task",
+        )
+        agent = SimpleNamespace(name="rex", model=None)
+        resolve = AsyncMock(return_value=("direct", "direct-model", "session"))
+        default_llm = AsyncMock()
+        validate = AsyncMock()
+        message_create = AsyncMock(return_value=SimpleNamespace(id="msg_user"))
+
+        monkeypatch.setattr(
+            session_routes,
+            "_require_agent_usable_for_chat",
+            AsyncMock(),
+        )
+        monkeypatch.setattr(
+            "flocks.agent.registry.Agent.default_agent",
+            AsyncMock(return_value="rex"),
+        )
+        monkeypatch.setattr(
+            "flocks.agent.registry.Agent.get",
+            AsyncMock(return_value=agent),
+        )
+        monkeypatch.setattr(session_routes, "_resolve_model", resolve)
+        monkeypatch.setattr(
+            "flocks.config.config.Config.resolve_default_llm",
+            default_llm,
+        )
+        monkeypatch.setattr(
+            "flocks.config.config.Config.get",
+            AsyncMock(return_value=SimpleNamespace()),
+        )
+        monkeypatch.setattr(SessionLoop, "validate_runtime_model", validate)
+        monkeypatch.setattr(
+            "flocks.provider.provider.Provider._ensure_initialized",
+            lambda: None,
+        )
+        monkeypatch.setattr(
+            "flocks.provider.provider.Provider.apply_config",
+            AsyncMock(),
+        )
+        monkeypatch.setattr(
+            "flocks.provider.provider.Provider.get",
+            lambda _provider_id: object(),
+        )
+        monkeypatch.setattr("flocks.tool.registry.ToolRegistry.init", lambda: None)
+        monkeypatch.setattr(
+            "flocks.session.lifecycle.revert.SessionRevert.cleanup",
+            AsyncMock(),
+        )
+        monkeypatch.setattr(
+            "flocks.session.message.Message.create",
+            message_create,
+        )
+        monkeypatch.setattr(
+            "flocks.server.routes.event.publish_event",
+            AsyncMock(),
+        )
+        context_usage = AsyncMock()
+        monkeypatch.setattr(
+            session_routes,
+            "_publish_context_usage_update",
+            context_usage,
+        )
+
+        await session_routes._process_session_message(
+            session.id,
+            session,
+            request,
+            session.directory,
+        )
+
+        resolve.assert_awaited_once()
+        default_llm.assert_not_awaited()
+        validate.assert_not_awaited()
+        assert message_create.await_args.kwargs["model"] == {
+            "providerID": "direct",
+            "modelID": "direct-model",
+        }
+        assert context_usage.await_args.kwargs["provider_id"] == "direct"
+        assert context_usage.await_args.kwargs["model_id"] == "direct-model"
 
     @pytest.mark.asyncio
     async def test_display_text_does_not_replace_model_prompt(self, monkeypatch):
diff --git a/tests/channel/test_channel.py b/tests/channel/test_channel.py
index 34a7830cf..5b6c57430 100644
--- a/tests/channel/test_channel.py
+++ b/tests/channel/test_channel.py
@@ -342,6 +342,62 @@ def test_empty_input(self):
 
 
 class TestFeishuNativeCommands:
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("category", "model_auto", "expected"),
+        [
+            ("user", True, True),
+            ("user", False, False),
+            ("task", True, False),
+        ],
+    )
+    async def test_im_loop_uses_supported_persisted_auto_mode(
+        self,
+        monkeypatch,
+        category,
+        model_auto,
+        expected,
+    ):
+        from flocks.channel.inbound.dispatcher import InboundDispatcher
+        from flocks.channel.inbound.session_binding import SessionBinding
+
+        binding = SessionBinding(
+            channel_id="slack",
+            account_id="T1",
+            chat_id="C123",
+            chat_type=ChatType.CHANNEL,
+            thread_id=None,
+            session_id="session_1",
+            agent_id="rex",
+            created_at=0,
+            last_message_at=0,
+        )
+        monkeypatch.setattr(
+            "flocks.session.session.Session.get_by_id",
+            AsyncMock(
+                return_value=SimpleNamespace(
+                    id="session_1",
+                    category=category,
+                    model_auto=model_auto,
+                )
+            ),
+        )
+        run_mock = AsyncMock(return_value=SimpleNamespace(last_message=None))
+        monkeypatch.setattr(
+            "flocks.session.session_loop.SessionLoop.run",
+            run_mock,
+        )
+
+        loop_callbacks = SimpleNamespace()
+        await InboundDispatcher._run_session_loop(binding, loop_callbacks)
+
+        run_mock.assert_awaited_once_with(
+            session_id="session_1",
+            agent_name="rex",
+            callbacks=loop_callbacks,
+            auto_failover=expected,
+        )
+
     @pytest.mark.asyncio
     async def test_status_command_reports_session_state(self, monkeypatch):
         from flocks.channel.inbound.dispatcher import InboundDispatcher
@@ -496,6 +552,7 @@ async def fake_update(project_id, session_id, **updates):
                     "provider": "anthropic",
                     "model": "claude-sonnet-4-20250514",
                     "model_pinned": True,
+                    "model_auto": False,
                 },
             )
         ]
@@ -599,6 +656,90 @@ async def fake_deliver(ctx, session_id=None):
         assert update_mock.await_args.args == ("channel", "session_old")
         assert update_mock.await_args.kwargs["status"] == "archived"
 
+    @pytest.mark.asyncio
+    async def test_new_command_inherits_auto_model_mode(self, monkeypatch):
+        from flocks.channel.inbound.dispatcher import InboundDispatcher
+        from flocks.channel.inbound.session_binding import SessionBinding
+
+        dispatcher = InboundDispatcher()
+        dispatcher._trigger_command_hook = AsyncMock()
+        dispatcher.binding_service.rebind = AsyncMock(
+            return_value=SessionBinding(
+                channel_id="wecom",
+                account_id="default",
+                chat_id="room_1",
+                chat_type=ChatType.DIRECT,
+                thread_id=None,
+                session_id="session_new",
+                agent_id="rex",
+                created_at=0,
+                last_message_at=0,
+            )
+        )
+        binding = SessionBinding(
+            channel_id="wecom",
+            account_id="default",
+            chat_id="room_1",
+            chat_type=ChatType.DIRECT,
+            thread_id=None,
+            session_id="session_old",
+            agent_id="rex",
+            created_at=0,
+            last_message_at=0,
+        )
+        msg = InboundMessage(
+            channel_id="wecom",
+            account_id="default",
+            message_id="msg_1",
+            sender_id="user_1",
+            chat_id="room_1",
+            chat_type=ChatType.DIRECT,
+            text="/new",
+            mention_text="/new",
+        )
+        monkeypatch.setattr(
+            "flocks.channel.outbound.deliver.OutboundDelivery.deliver",
+            AsyncMock(),
+        )
+        monkeypatch.setattr(
+            "flocks.session.session.Session.get_by_id",
+            AsyncMock(
+                return_value=SimpleNamespace(
+                    id="session_old",
+                    project_id="channel",
+                    directory="/tmp/project",
+                    agent="rex",
+                    category="user",
+                    model_auto=True,
+                    model_pinned=False,
+                    provider=None,
+                    model=None,
+                )
+            ),
+        )
+        create_mock = AsyncMock(
+            return_value=SimpleNamespace(id="session_new", agent="rex")
+        )
+        monkeypatch.setattr("flocks.session.session.Session.create", create_mock)
+        monkeypatch.setattr(
+            "flocks.session.session.Session.update",
+            AsyncMock(return_value=None),
+        )
+
+        await dispatcher._handle_session_command(
+            binding=binding,
+            msg=msg,
+            callbacks=dispatcher._build_callbacks(binding, msg),
+            scope_override=None,
+            channel_config=ChannelConfig(enabled=True),
+        )
+
+        create_kwargs = create_mock.await_args.kwargs
+        assert create_kwargs["model_auto"] is True
+        assert create_kwargs["model_pinned"] is False
+        assert "provider" not in create_kwargs
+        assert "model" not in create_kwargs
+
     @pytest.mark.asyncio
     async def test_new_command_with_args_sends_args_to_new_session(self, monkeypatch):
         from flocks.channel.inbound.dispatcher import InboundDispatcher
diff --git a/tests/channel/test_session_binding.py b/tests/channel/test_session_binding.py
index 45e7ad366..0a09d8e43 100644
--- a/tests/channel/test_session_binding.py
+++ b/tests/channel/test_session_binding.py
@@ -3,7 +3,8 @@
 
 import pytest
 
-from flocks.channel.inbound.session_binding import SessionBindingService
+from flocks.channel.base import ChatType, InboundMessage
+from flocks.channel.inbound.session_binding import SessionBinding, SessionBindingService
 
 
 @pytest.mark.asyncio
@@ -25,3 +26,52 @@ async def test_latest_active_user_binding_returns_none_when_channel_is_ambiguous
         result = await service.latest_active_user_binding(channel_id="wecom")
 
     assert result is None
+
+
+@pytest.mark.asyncio
+async def test_resolve_or_create_replaces_archived_binding() -> None:
+    service = SessionBindingService()
+    existing = SessionBinding(
+        channel_id="slack",
+        account_id="default",
+        chat_id="chat-1",
+        chat_type=ChatType.DIRECT,
+        thread_id=None,
+        session_id="ses_archived",
+        agent_id="rex",
+        created_at=1,
+        last_message_at=2,
+    )
+    archived = SimpleNamespace(
+        status="archived",
+        owner_user_id="usr_1",
+        owner_username="alice",
+    )
+    service._find_binding = AsyncMock(return_value=existing)
+    service.unbind = AsyncMock()
+    service._create_session = AsyncMock(return_value="ses_replacement")
+    service._insert = AsyncMock()
+    msg = InboundMessage(
+        channel_id="slack",
+        account_id="default",
+        message_id="msg-1",
+        sender_id="alice",
+        chat_id="chat-1",
+        chat_type=ChatType.DIRECT,
+        text="hello again",
+    )
+
+    with patch(
+        "flocks.session.session.Session.get_by_id_unfiltered",
+        AsyncMock(return_value=archived),
+    ):
+        binding = await service.resolve_or_create(msg, default_agent="rex")
+
+    assert binding.session_id == "ses_replacement"
+    service.unbind.assert_awaited_once_with("ses_archived")
+    service._create_session.assert_awaited_once_with(
+        msg,
+        default_agent="rex",
+        directory=None,
+        source_session=archived,
+    )
diff --git a/tests/channel/test_unified_prompt_context.py b/tests/channel/test_unified_prompt_context.py
index ac09461b3..b1c105b41 100644
--- a/tests/channel/test_unified_prompt_context.py
+++ b/tests/channel/test_unified_prompt_context.py
@@ -257,7 +257,7 @@ async def _fake_create(**kwargs):
         assert captured["owner_username"] == "admin"
 
     @pytest.mark.asyncio
-    async def test_create_session_stays_ownerless_without_local_accounts(self):
+    async def test_create_session_uses_system_owner_without_local_accounts(self):
         captured = {}
 
         class _StubSession:
@@ -277,8 +277,8 @@ async def _fake_create(**kwargs):
             )
 
         assert sid == "ses_ownerless"
-        assert "owner_user_id" not in captured
-        assert "owner_username" not in captured
+        assert captured["owner_user_id"] == "api-token-service"
+        assert captured["owner_username"] == "api-token-service"
         list_users.assert_not_awaited()
 
     @pytest.mark.asyncio
@@ -407,6 +407,15 @@ async def _boom(_session, _p, _m):
 
 
 class TestAppendUserMessagePersistsModel:
+    @pytest.fixture(autouse=True)
+    def _active_session_write(self, monkeypatch):
+        from flocks.session.session import Session
+
+        async def run_active_write(_cls, _session_id, operation, **_kwargs):
+            return await operation()
+
+        monkeypatch.setattr(Session, "run_active_write", classmethod(run_active_write))
+
     @pytest.mark.asyncio
     async def test_model_is_passed_to_message_create(self):
         captured = {}
@@ -498,6 +507,29 @@ async def _fake_create(**kwargs):
         assert "agent" not in captured
 
 
+@pytest.mark.asyncio
+async def test_append_user_message_rejects_inactive_session_before_persisting(monkeypatch):
+    from flocks.session.message import Message
+    from flocks.session.session import Session, SessionInactiveError
+
+    create = AsyncMock()
+    monkeypatch.setattr(Message, "create", create)
+    monkeypatch.setattr(
+        Session,
+        "run_active_write",
+        AsyncMock(side_effect=SessionInactiveError("archived")),
+    )
+
+    with pytest.raises(SessionInactiveError, match="archived"):
+        await InboundDispatcher._append_user_message(
+            "ses_archived",
+            "must not be stored",
+            _msg(),
+        )
+
+    create.assert_not_awaited()
+
+
 # ---------------------------------------------------------------------------
 # 4. Legacy channel_bindings.agent_id one-shot normalisation
 # ---------------------------------------------------------------------------
diff --git a/tests/cli/test_service_manager.py b/tests/cli/test_service_manager.py
index 42f3736ad..3ab93e061 100644
--- a/tests/cli/test_service_manager.py
+++ b/tests/cli/test_service_manager.py
@@ -1,4 +1,5 @@
 import json
+import os
 import shutil
 import subprocess
 import sys
@@ -17,6 +18,8 @@
     make_short_runtime_root,
 )
 
+_REAL_ENSURE_WEBUI_DIST = service_manager._ensure_webui_dist
+
 
 class DummyConsole:
     def __init__(self) -> None:
@@ -520,6 +523,87 @@ def test_run_windows_netstat_handles_missing_stdout(monkeypatch) -> None:
     assert service_manager._run_windows_netstat(5173) == ""
 
 
+def test_run_windows_netstat_decodes_with_replacement(monkeypatch) -> None:
+    """netstat output is decoded as utf-8/replace so GBK console output on
+    Chinese Windows never raises UnicodeDecodeError in the reader thread."""
+    seen: dict[str, object] = {}
+
+    def fake_run(*_args, **kwargs):
+        seen.update(kwargs)
+        return SimpleNamespace(returncode=0, stdout="  TCP 127.0.0.1:5173 0.0.0.0:0 LISTENING 42\n")
+
+    monkeypatch.setattr(service_manager.subprocess, "run", fake_run)
+
+    assert "42" in service_manager._run_windows_netstat(5173)
+    assert seen.get("encoding") == "utf-8"
+    assert seen.get("errors") == "replace"
+
+
+def test_windows_tasklist_process_name_decodes_with_replacement(monkeypatch) -> None:
+    monkeypatch.setattr(service_manager.sys, "platform", "win32")
+    seen: dict[str, object] = {}
+
+    def fake_run(*_args, **kwargs):
+        seen.update(kwargs)
+        return SimpleNamespace(returncode=0, stdout='"python.exe","9436"\n')
+
+    monkeypatch.setattr(service_manager.subprocess, "run", fake_run)
+
+    assert service_manager._windows_tasklist_process_name(9436) == "python.exe"
+    assert seen.get("encoding") == "utf-8"
+    assert seen.get("errors") == "replace"
+
+
+def test_process_list_pids_windows_decodes_with_replacement(monkeypatch) -> None:
+    monkeypatch.setattr(service_manager.sys, "platform", "win32")
+    seen: dict[str, object] = {}
+
+    def fake_run(*_args, **kwargs):
+        seen.update(kwargs)
+        return SimpleNamespace(returncode=0, stdout="9436\n36056\n")
+
+    monkeypatch.setattr(service_manager.subprocess, "run", fake_run)
+
+    assert service_manager._process_list_pids() == [9436, 36056]
+    assert seen.get("encoding") == "utf-8"
+    assert seen.get("errors") == "replace"
+
+
+def test_get_node_major_version_decodes_with_replacement(monkeypatch) -> None:
+    monkeypatch.setattr(service_manager, "resolve_node_executable", lambda: "node")
+    seen: dict[str, object] = {}
+
+    def fake_run(*_args, **kwargs):
+        seen.update(kwargs)
+        return SimpleNamespace(returncode=0, stdout="v20.11.1\n")
+
+    monkeypatch.setattr(service_manager.subprocess, "run", fake_run)
+
+    assert service_manager.get_node_major_version() == 20
+    assert seen.get("encoding") == "utf-8"
+    assert seen.get("errors") == "replace"
+
+
+def test_windows_process_probes_survive_gbk_bytes(monkeypatch) -> None:
+    """End-to-end: real GBK-encoded bytes flowing through the decode path
+    must not raise UnicodeDecodeError (the restart-time reader-thread crash)."""
+    monkeypatch.setattr(service_manager.sys, "platform", "win32")
+    # 0xbb is the byte that crashed utf-8 decode in the original bug report.
+    gbk_stdout = "映像名称: python.exe 拒绝访问".encode("gbk").decode("utf-8", errors="replace")
+
+    def fake_run(*_args, **kwargs):
+        assert kwargs.get("encoding") == "utf-8"
+        assert kwargs.get("errors") == "replace"
+        return SimpleNamespace(returncode=0, stdout=gbk_stdout)
+
+    monkeypatch.setattr(service_manager.subprocess, "run", fake_run)
+
+    # None of these should raise, even with mojibake stdout.
+    service_manager._windows_tasklist_process_name(123)
+    service_manager._run_windows_netstat(5173)
+    service_manager._process_list_pids()
+
+
 def test_port_owner_pids_warns_when_no_tool_found(monkeypatch) -> None:
     monkeypatch.setattr(service_manager.sys, "platform", "linux")
     monkeypatch.setattr(service_manager, "which", lambda _name: None)
@@ -1112,6 +1196,40 @@ def test_restart_all_stops_then_starts_daemon(monkeypatch) -> None:
     assert call_order == ["stop", "start"]
 
 
+def test_restart_server_requests_backend_restart(monkeypatch) -> None:
+    calls: list[str] = []
+    console = DummyConsole()
+    paths = _make_runtime_paths(Path("/tmp/flocks-test"))
+    status = _supervisor_status(_supervisor_status_payload())
+
+    monkeypatch.setattr(service_manager, "ensure_runtime_dirs", lambda: paths)
+    monkeypatch.setattr(service_manager, "supervisor_is_running", lambda _paths: True)
+    monkeypatch.setattr(
+        service_manager,
+        "request_restart_backend",
+        lambda **_kwargs: calls.append("backend") or status,
+    )
+    monkeypatch.setattr(
+        service_manager,
+        "_print_status_payload",
+        lambda *_args, **_kwargs: calls.append("status"),
+    )
+
+    service_manager.restart_server(console)
+
+    assert calls == ["backend", "status"]
+
+
+def test_restart_server_requires_running_supervisor(monkeypatch) -> None:
+    paths = _make_runtime_paths(Path("/tmp/flocks-test"))
+
+    monkeypatch.setattr(service_manager, "ensure_runtime_dirs", lambda: paths)
+    monkeypatch.setattr(service_manager, "supervisor_is_running", lambda _paths: False)
+
+    with pytest.raises(service_manager.ServiceError, match="请执行 `flocks restart` 进行全量重启"):
+        service_manager.restart_server(DummyConsole())
+
+
 def test_start_all_without_stop_starts_supervisor_daemon(monkeypatch, tmp_path: Path) -> None:
     paths = _make_runtime_paths(tmp_path)
     calls: list[str] = []
@@ -1518,7 +1636,10 @@ def _fake_process(pid: int, args: list[str] | None = None, returncode: int | Non
     return SimpleNamespace(pid=pid, args=args or [str(pid)], returncode=returncode, poll=lambda: returncode)
 
 
-def test_supervisor_recovers_backend_when_port_disappears(monkeypatch, tmp_path: Path) -> None:
+def test_supervisor_restarts_backend_after_tenth_consecutive_port_failure(
+    monkeypatch,
+    tmp_path: Path,
+) -> None:
     paths = _make_runtime_paths(tmp_path)
     calls: list[str] = []
     monkeypatch.setattr(service_manager, "ensure_runtime_dirs", lambda: paths)
@@ -1535,12 +1656,90 @@ def test_supervisor_recovers_backend_when_port_disappears(monkeypatch, tmp_path:
         lambda *_args, **_kwargs: calls.append("start:backend") or _fake_process(333, ["backend-new"]),
     )
 
-    daemon.tick()
+    assert daemon.interval == 30.0
+    assert daemon.failure_threshold == 10
+
+    for expected_failure_count in range(1, 10):
+        daemon.tick()
+        assert calls == []
+        assert daemon.backend.state == "degraded"
+        assert daemon.backend.health_failure_count == expected_failure_count
 
+    daemon.tick()
     assert calls == ["stop:后端", "start:backend"]
     assert daemon.backend.pid == 333
 
 
+def test_supervisor_resets_backend_port_failures_after_recovery(
+    monkeypatch,
+    tmp_path: Path,
+) -> None:
+    paths = _make_runtime_paths(tmp_path)
+    calls: list[str] = []
+    port_states = iter([False] * 9 + [True] + [False] * 9)
+    monkeypatch.setattr(service_manager, "ensure_runtime_dirs", lambda: paths)
+    daemon = service_supervisor.SupervisorDaemon(service_manager.ServiceConfig(backend_port=9995))
+    daemon.paths = paths
+    daemon.backend.process = _fake_process(111, ["backend"])
+
+    monkeypatch.setattr(
+        service_process,
+        "tcp_port_accepts_connections",
+        lambda *_args: next(port_states),
+    )
+    monkeypatch.setattr(
+        service_manager,
+        "_terminate_process",
+        lambda *_args, **_kwargs: calls.append("stop"),
+    )
+
+    for _ in range(9):
+        daemon.tick()
+
+    assert calls == []
+    assert daemon.backend.health_failure_count == 9
+
+    daemon.tick()
+
+    assert daemon.backend.state == "healthy"
+    assert daemon.backend.health_failure_count == 0
+
+    for _ in range(9):
+        daemon.tick()
+
+    assert calls == []
+    assert daemon.backend.state == "degraded"
+    assert daemon.backend.health_failure_count == 9
+
+
+def test_supervisor_restarts_backend_on_first_probe_after_process_exits(
+    monkeypatch,
+    tmp_path: Path,
+) -> None:
+    paths = _make_runtime_paths(tmp_path)
+    calls: list[str] = []
+    monkeypatch.setattr(service_manager, "ensure_runtime_dirs", lambda: paths)
+    daemon = service_supervisor.SupervisorDaemon(service_manager.ServiceConfig(backend_port=9995))
+    daemon.paths = paths
+    daemon.backend.process = _fake_process(111, ["backend"], returncode=7)
+
+    monkeypatch.setattr(
+        service_manager,
+        "_terminate_process",
+        lambda *_args, **_kwargs: calls.append("stop"),
+    )
+    monkeypatch.setattr(
+        service_manager,
+        "_start_backend_process",
+        lambda *_args, **_kwargs: calls.append("start") or _fake_process(333, ["backend-new"]),
+    )
+
+    daemon.tick()
+
+    assert calls == ["stop", "start"]
+    assert daemon.backend.pid == 333
+
+
 def test_supervisor_liveness_probe_keeps_backend_healthy(monkeypatch, tmp_path: Path) -> None:
     """Liveness-only probe: process alive + TCP port open = healthy, no HTTP check needed."""
     paths = _make_runtime_paths(tmp_path)
@@ -1655,6 +1854,108 @@ def fake_run(command, **_kwargs):
     assert build_calls[0][0] == r"C:\Users\flocks\AppData\Local\Programs\Flocks\tools\node\npm.cmd"
 
 
+def test_webui_needs_build_when_source_is_newer_than_dist(tmp_path: Path) -> None:
+    webui_dir = tmp_path / "webui"
+    source_path = webui_dir / "src" / "main.tsx"
+    index_path = webui_dir / "dist" / "index.html"
+    source_path.parent.mkdir(parents=True)
+    index_path.parent.mkdir()
+    source_path.write_text("before", encoding="utf-8")
+    index_path.write_text("<html></html>", encoding="utf-8")
+    built_at = index_path.stat().st_mtime_ns
+    os.utime(source_path, ns=(built_at + 1_000_000_000, built_at + 1_000_000_000))
+
+    assert service_manager._webui_needs_build(webui_dir) is True
+
+
+def test_webui_needs_build_ignores_generated_directories(tmp_path: Path) -> None:
+    webui_dir = tmp_path / "webui"
+    source_path = webui_dir / "src" / "main.tsx"
+    index_path = webui_dir / "dist" / "index.html"
+    generated_path = webui_dir / "node_modules" / "package" / "index.js"
+    source_path.parent.mkdir(parents=True)
+    index_path.parent.mkdir()
+    generated_path.parent.mkdir(parents=True)
+    source_path.write_text("source", encoding="utf-8")
+    index_path.write_text("<html></html>", encoding="utf-8")
+    generated_path.write_text("generated", encoding="utf-8")
+    built_at = max(path.stat().st_mtime_ns for path in (webui_dir, source_path.parent, source_path))
+    os.utime(index_path, ns=(built_at + 1_000_000_000, built_at + 1_000_000_000))
+    os.utime(generated_path, ns=(built_at + 2_000_000_000, built_at + 2_000_000_000))
+
+    assert service_manager._webui_needs_build(webui_dir) is False
+
+
+def test_webui_needs_build_when_source_is_deleted(tmp_path: Path) -> None:
+    webui_dir = tmp_path / "webui"
+    source_dir = webui_dir / "src"
+    source_path = source_dir / "main.tsx"
+    index_path = webui_dir / "dist" / "index.html"
+    source_dir.mkdir(parents=True)
+    index_path.parent.mkdir()
+    source_path.write_text("source", encoding="utf-8")
+    index_path.write_text("<html></html>", encoding="utf-8")
+    built_at = max(path.stat().st_mtime_ns for path in (webui_dir, source_dir, source_path))
+    os.utime(index_path, ns=(built_at + 1_000_000_000, built_at + 1_000_000_000))
+    source_path.unlink()
+    os.utime(source_dir, ns=(built_at + 2_000_000_000, built_at + 2_000_000_000))
+
+    assert service_manager._webui_needs_build(webui_dir) is True
+
+
+def test_ensure_webui_dist_rebuilds_when_source_is_newer(monkeypatch, tmp_path: Path) -> None:
+    from flocks.server import static_webui
+
+    webui_dir = tmp_path / "webui"
+    source_path = webui_dir / "src" / "main.tsx"
+    index_path = webui_dir / "dist" / "index.html"
+    source_path.parent.mkdir(parents=True)
+    index_path.parent.mkdir()
+    (webui_dir / "package.json").write_text("{}", encoding="utf-8")
+    source_path.write_text("before", encoding="utf-8")
+    index_path.write_text("<html></html>", encoding="utf-8")
+    built_at = index_path.stat().st_mtime_ns
+    os.utime(source_path, ns=(built_at + 1_000_000_000, built_at + 1_000_000_000))
+    builds: list[Path] = []
+    monkeypatch.setattr(static_webui, "ensure_webui_dist_dir", lambda: index_path.parent.resolve())
+    monkeypatch.setattr(
+        service_manager,
+        "_build_webui_dist",
+        lambda root, _config, _console: builds.append(root),
+    )
+
+    _REAL_ENSURE_WEBUI_DIST(tmp_path, service_manager.ServiceConfig(), DummyConsole())
+
+    assert builds == [tmp_path]
+
+
+def test_ensure_webui_dist_skips_stale_source_when_requested(monkeypatch, tmp_path: Path) -> None:
+    from flocks.server import static_webui
+
+    webui_dir = tmp_path / "webui"
+    source_path = webui_dir / "src" / "main.tsx"
+    index_path = webui_dir / "dist" / "index.html"
+    source_path.parent.mkdir(parents=True)
+    index_path.parent.mkdir()
+    (webui_dir / "package.json").write_text("{}", encoding="utf-8")
+    source_path.write_text("before", encoding="utf-8")
+    index_path.write_text("<html></html>", encoding="utf-8")
+    built_at = index_path.stat().st_mtime_ns
+    os.utime(source_path, ns=(built_at + 1_000_000_000, built_at + 1_000_000_000))
+    builds: list[Path] = []
+    monkeypatch.setattr(static_webui, "ensure_webui_dist_dir", lambda: index_path.parent.resolve())
+    monkeypatch.setattr(
+        service_manager,
+        "_build_webui_dist",
+        lambda root, _config, _console: builds.append(root),
+    )
+
+    config = service_manager.ServiceConfig(skip_frontend_build=True)
+    _REAL_ENSURE_WEBUI_DIST(tmp_path, config, DummyConsole())
+
+    assert builds == []
+
+
 def test_start_backend_raises_when_port_has_listener(monkeypatch, tmp_path: Path) -> None:
     paths = service_manager.RuntimePaths(
         root=tmp_path,
diff --git a/tests/cli/test_session_commands.py b/tests/cli/test_session_commands.py
index a664095c7..4bd75558a 100644
--- a/tests/cli/test_session_commands.py
+++ b/tests/cli/test_session_commands.py
@@ -1,3 +1,5 @@
+from unittest.mock import AsyncMock
+
 import pytest
 from typer.testing import CliRunner
 
@@ -83,6 +85,30 @@ async def fake_delete(project_id: str, session_id: str):
     await session_cmd._delete_session(session.id, None, force=True)
 
 
+@pytest.mark.asyncio
+async def test_delete_session_confirmation_calls_out_permanent_deletion(monkeypatch) -> None:
+    session = _build_session(project_id="proj_delete_prompt", title="Delete Forever")
+    prompts: list[str] = []
+
+    async def fake_get_by_id(_session_id: str):
+        return session
+
+    def fake_confirm(prompt: str, *, default: bool):
+        prompts.append(prompt)
+        assert default is False
+        return False
+
+    monkeypatch.setattr(session_cmd.Storage, "init", _noop_storage_init)
+    monkeypatch.setattr(session_cmd.Session, "get_by_id", fake_get_by_id)
+    monkeypatch.setattr(session_cmd.typer, "confirm", fake_confirm)
+
+    await session_cmd._delete_session(session.id, None, force=False)
+
+    assert prompts == [
+        "Permanently delete session 'Delete Forever' and all of its messages and history?"
+    ]
+
+
 @pytest.mark.asyncio
 async def test_archive_session_resolves_project_from_session(monkeypatch) -> None:
     session = _build_session(project_id="proj_archive")
@@ -105,19 +131,34 @@ async def fake_archive(project_id: str, session_id: str):
 
 @pytest.mark.asyncio
 async def test_restore_session_resolves_project_from_session(monkeypatch) -> None:
-    session = _build_session(project_id="proj_restore")
+    session = _build_session(project_id="proj_restore").model_copy(
+        update={"owner_user_id": "usr_cli"}
+    )
 
     async def fake_get_by_id(session_id: str):
         assert session_id == session.id
         return session
 
-    async def fake_unarchive(project_id: str, session_id: str):
+    async def fake_restore(
+        project_id: str,
+        session_id: str,
+        *,
+        project_owner_id: str,
+    ):
         assert project_id == session.project_id
         assert session_id == session.id
+        assert project_owner_id == "usr_cli"
         return True
 
     monkeypatch.setattr(session_cmd.Storage, "init", _noop_storage_init)
     monkeypatch.setattr(session_cmd.Session, "get_by_id", fake_get_by_id)
-    monkeypatch.setattr(session_cmd.Session, "unarchive", fake_unarchive)
+    restore_session = AsyncMock(side_effect=fake_restore)
+    monkeypatch.setattr(session_cmd.Session, "restore", restore_session)
 
     await session_cmd._restore_session(session.id, None)
+
+    restore_session.assert_awaited_once_with(
+        session.project_id,
+        session.id,
+        project_owner_id="usr_cli",
+    )
diff --git a/tests/conftest.py b/tests/conftest.py
index ecb8e12c9..651e714fc 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -3,6 +3,7 @@
 """
 
 import os
+from pathlib import Path
 
 import pytest
 
@@ -37,3 +38,51 @@ def pytest_runtest_setup(item):
     for marker_name, (env_var, reason) in _API_KEY_MARKERS.items():
         if item.get_closest_marker(marker_name) and not os.getenv(env_var):
             pytest.skip(reason)
+
+
+@pytest.fixture(autouse=True)
+def _home_honors_env(monkeypatch):
+    """Make ``Path.home()`` follow the ``HOME`` env var on every platform.
+
+    Many suites isolate their filesystem by ``monkeypatch.setenv("HOME", ...)``
+    and then rely on production code that resolves ``Path.home() / ".flocks"``
+    (hub installs, plugin roots, the WebUI contract store, ...). That works on
+    Linux CI, where ``Path.home()`` honors ``$HOME`` — but on Windows
+    ``Path.home()`` reads ``%USERPROFILE%`` and ignores ``HOME`` entirely, so
+    those tests silently leak into (and assert against) the *real* user home.
+
+    This autouse fixture patches ``pathlib.Path.home`` to honor ``HOME`` (then
+    ``USERPROFILE``) at call time, aligning Windows with Linux CI. It is
+    behavior-preserving when ``HOME`` already points at the real home (the
+    default), so tests that never touch ``HOME`` see no change.
+    """
+    def _home() -> Path:
+        candidate = os.environ.get("HOME") or os.environ.get("USERPROFILE")
+        if candidate:
+            return Path(candidate)
+        return Path(os.path.expanduser("~"))
+
+    monkeypatch.setattr(Path, "home", staticmethod(_home))
+
+    # ``Path.home()`` covers explicit home lookups, but ``~``/``~user`` tilde
+    # expansion goes through ``os.path.expanduser`` (and ``Path.expanduser``),
+    # which on Windows keys off ``%USERPROFILE%``/``%HOMEDRIVE%%HOMEPATH%`` and
+    # likewise ignores ``HOME``. Route a bare ``~`` / ``~/...`` through the same
+    # ``HOME``-aware root so tilde-based tests are hermetic too. Anything else
+    # (``~otheruser``) falls back to the real implementation untouched.
+    _real_expanduser = os.path.expanduser
+
+    def _expanduser(path):
+        text = os.fspath(path)
+        # Only rewrite plain ``str`` tilde paths; leave bytes and ``~user``
+        # forms to the real implementation so we never change its type
+        # contract or raise on a bytes path.
+        if isinstance(text, str):
+            home = os.environ.get("HOME") or os.environ.get("USERPROFILE")
+            if home and (text == "~" or text.startswith("~/") or text.startswith("~\\")):
+                return home + text[1:]
+        return _real_expanduser(path)
+
+    monkeypatch.setattr(os.path, "expanduser", _expanduser)
+
+
diff --git a/tests/contracts/webui/test_store.py b/tests/contracts/webui/test_store.py
index d88f35175..78d4c9ede 100644
--- a/tests/contracts/webui/test_store.py
+++ b/tests/contracts/webui/test_store.py
@@ -50,6 +50,7 @@ def _write_workspace(
     workspace_id: str,
     title: str,
     order: int = 100,
+    version: str | None = None,
     default_page_id: str | None = None,
     sections: list[dict] | None = None,
 ) -> None:
@@ -63,6 +64,8 @@ def _write_workspace(
         "enabled": True,
         "placement": "sceneWorkspace",
     }
+    if version is not None:
+        payload["version"] = version
     if default_page_id is not None:
         payload["defaultPageId"] = default_page_id
     if sections is not None:
@@ -109,6 +112,40 @@ def test_list_pages_scans_user_and_project_roots_with_user_priority(tmp_path):
     assert store.page_dir("project-page").is_relative_to(project_root)
 
 
+def test_list_pages_skips_installer_scratch_dirs(tmp_path):
+    """Pages that exist only under the Hub installer's ``.<name>.<rand>`` /
+    ``.<name>.bak`` scratch dirs must never surface as real pages.
+
+    Reproduces the Windows WinError 5 aftermath where a failed atomic swap
+    left the SOC pages' manifests stuck inside scratch dirs while the real
+    install was incomplete.
+    """
+    user_root = tmp_path / "user" / "contracts" / "webui"
+    _write_page(user_root, "real-page", "Real Page")
+    # Manifests stranded in leftover scratch/backup dirs (no real install).
+    _write_page_at(user_root, ".soc_ui.abc123/soc_overview", "soc-overview", "Stranded Overview")
+    _write_page_at(user_root, ".soc_ui.bak/soc_dashboard", "soc-dashboard", "Stranded Dashboard")
+
+    store = WebUIPagesStore(root=user_root, project_root=None, legacy_root=None)
+
+    pages = store.list_pages()
+    assert [page.id for page in pages] == ["real-page"]
+    assert pages[0].title == "Real Page"
+
+
+def test_list_workspaces_skips_installer_scratch_dirs(tmp_path):
+    user_root = tmp_path / "user" / "contracts" / "webui"
+    _write_workspace(user_root, "real_ws", "Real Workspace")
+    # A workspace manifest stranded under a scratch dir must be ignored.
+    _write_workspace(user_root / ".soc_ui.abc123", "soc_ui", "Stranded Workspace")
+
+    store = WebUIPagesStore(root=user_root, project_root=None, legacy_root=None)
+
+    workspaces = store.list_workspaces()
+    assert [workspace.id for workspace in workspaces] == ["real_ws"]
+    assert workspaces[0].title == "Real Workspace"
+
+
 def test_grouped_page_directory_uses_manifest_id_for_lookup(tmp_path):
     user_root = tmp_path / "user" / "contracts" / "webui"
     _write_workspace(user_root, "scene_workspace", "场景工作区")
@@ -140,6 +177,7 @@ def test_list_workspaces_returns_grouped_pages(tmp_path):
         "scene_workspace",
         "场景工作区",
         order=5,
+        version="2.3.4",
         default_page_id="ops-overview",
         sections=[
             {
@@ -159,6 +197,7 @@ def test_list_workspaces_returns_grouped_pages(tmp_path):
     workspaces = store.list_workspaces()
 
     assert [workspace.id for workspace in workspaces] == ["scene_workspace"]
+    assert workspaces[0].version == "2.3.4"
     assert workspaces[0].title == "场景工作区"
     assert workspaces[0].route == "/contracts/webui/workspaces/scene_workspace"
     assert workspaces[0].placement == "sceneWorkspace"
diff --git a/tests/contracts/webui/test_watcher.py b/tests/contracts/webui/test_watcher.py
index a5043c6a1..271bbfc48 100644
--- a/tests/contracts/webui/test_watcher.py
+++ b/tests/contracts/webui/test_watcher.py
@@ -104,3 +104,39 @@ def test_watcher_classifies_workspace_manifest_change(tmp_path):
 
     assert page_id == "scene_workspace"
     assert pending.manifest_changed
+
+
+def test_watcher_ignores_installer_scratch_dir_events(tmp_path):
+    """Events under the Hub installer's ``.<name>.<rand>`` / ``.<name>.bak``
+    scratch dirs must be dropped *before* any store lookup — reacting there
+    races the installer's atomic swap and on Windows blocks it with a
+    WinError 5 access-denied.
+
+    The store is stubbed to raise if consulted, proving the dot-dir guard
+    short-circuits ahead of path resolution rather than relying on the
+    store's own scratch-dir filtering.
+    """
+    root = tmp_path / "webui_pages"
+    root.mkdir(parents=True)
+
+    class _RaisingStore:
+        def workspace_id_for_path(self, _path):
+            raise AssertionError("store should not be consulted for scratch-dir events")
+
+        def page_id_for_path(self, _path):
+            raise AssertionError("store should not be consulted for scratch-dir events")
+
+    watcher = WebUIPagesWatcher(
+        store=_RaisingStore(), builder=_BuilderStub(), api_runtime=_RuntimeStub()
+    )
+
+    for candidate in (
+        root / ".soc_ui.abc123" / "soc_overview" / "src" / "index.tsx",
+        root / ".soc_ui.abc123" / "soc_overview" / "manifest.json",
+        root / ".soc_ui.abc123" / "soc_overview" / "dist" / "page.js",
+        root / ".soc_ui.bak" / "soc_overview" / "workspace.json",
+    ):
+        assert (
+            watcher._classify_event(candidate, root, event_type="modified", is_directory=False)
+            is None
+        )
diff --git a/tests/hooks/test_pipeline.py b/tests/hooks/test_pipeline.py
index 807a43853..5e285b2bb 100644
--- a/tests/hooks/test_pipeline.py
+++ b/tests/hooks/test_pipeline.py
@@ -10,7 +10,7 @@
 
 import pytest
 
-from flocks.hooks.pipeline import HookBase, HookPipeline
+from flocks.hooks.pipeline import HookBase, HookPipeline, HookStage
 from flocks.plugin.loader import PluginLoader
 
 
@@ -111,3 +111,60 @@ async def test_pipeline_resolves_project_dir_from_session_when_workspace_missing
     )
 
     assert "hook-session" in set(HookPipeline.list_hooks())
+
+
+def test_hook_base_exposes_only_canonical_lifecycle_names() -> None:
+    assert not hasattr(HookBase, "chat_message")
+    assert hasattr(HookBase, "tool_before")
+    assert hasattr(HookBase, "tool_after")
+    assert not hasattr(HookBase, "pre_tool_use")
+    assert not hasattr(HookBase, "post_tool_use")
+    assert not hasattr(HookPipeline, "run_chat_message")
+    assert hasattr(HookPipeline, "run_tool_before")
+    assert hasattr(HookPipeline, "run_tool_after")
+    assert not hasattr(HookPipeline, "run_pre_tool_use")
+    assert not hasattr(HookPipeline, "run_post_tool_use")
+
+
+def test_default_hook_methods_are_not_stage_handlers() -> None:
+    hook = HookBase()
+
+    assert HookPipeline._resolve_handler(
+        hook,
+        HookStage.USER_PROMPT_SUBMIT,
+    ) is None
+    assert HookPipeline._resolve_handler(hook, HookStage.LLM_BEFORE) is None
+    assert HookPipeline._resolve_handler(hook, HookStage.TOOL_BEFORE) is None
+
+
+@pytest.mark.asyncio
+async def test_pipeline_runs_canonical_lifecycle_handlers() -> None:
+    seen: list[str] = []
+
+    class _CanonicalHook(HookBase):
+        async def user_prompt_submit(self, ctx) -> None:
+            seen.append(ctx.stage)
+
+        async def tool_before(self, ctx) -> None:
+            seen.append(ctx.stage)
+
+        async def tool_after(self, ctx) -> None:
+            seen.append(ctx.stage)
+
+        async def turn_finish(self, ctx) -> None:
+            seen.append(ctx.stage)
+
+    HookPipeline.register("canonical-hook", _CanonicalHook())
+    HookPipeline._initialized = True
+
+    await HookPipeline.run_user_prompt_submit({"sessionID": "ses_test"})
+    await HookPipeline.run_tool_before({"sessionID": "ses_test"})
+    await HookPipeline.run_tool_after({"sessionID": "ses_test"})
+    await HookPipeline.run_turn_finish({"sessionID": "ses_test"})
+
+    assert seen == [
+        HookStage.USER_PROMPT_SUBMIT,
+        HookStage.TOOL_BEFORE,
+        HookStage.TOOL_AFTER,
+        HookStage.TURN_FINISH,
+    ]
diff --git a/tests/hub/test_hub_catalog.py b/tests/hub/test_hub_catalog.py
index bdb479b2e..0eaab0640 100644
--- a/tests/hub/test_hub_catalog.py
+++ b/tests/hub/test_hub_catalog.py
@@ -270,6 +270,45 @@ def test_catalog_ignores_stale_record_when_legacy_install_is_inferred(
     assert local.get_record("device", plugin_id) is None
 
 
+def test_catalog_uses_webui_workspace_version_for_inferred_installs(
+    isolated_hub_env,
+):
+    from flocks.hub.catalog import clear_catalog_caches
+
+    webui_dir = (
+        isolated_hub_env["home"]
+        / ".flocks"
+        / "plugins"
+        / "contracts"
+        / "webui"
+        / "soc_ui"
+    )
+    webui_dir.mkdir(parents=True)
+    workspace_path = webui_dir / "workspace.json"
+    workspace = {
+        "id": "soc_ui",
+        "version": "1.0.0",
+        "title": "SOC 工作区",
+        "placement": "sceneWorkspace",
+    }
+    workspace_path.write_text(json.dumps(workspace), encoding="utf-8")
+    clear_catalog_caches()
+
+    entry = {item.id: item for item in list_catalog(plugin_type="webui")}["soc_ui"]
+
+    assert entry.version == "1.1.4"
+    assert entry.state == "updateAvailable"
+    assert entry.installedVersion == "1.0.0"
+
+    workspace["version"] = "1.1.4"
+    workspace_path.write_text(json.dumps(workspace), encoding="utf-8")
+
+    refreshed = {item.id: item for item in list_catalog(plugin_type="webui")}["soc_ui"]
+
+    assert refreshed.state == "installed"
+    assert refreshed.installedVersion == "1.1.4"
+
+
 def test_pentest_agents_are_listed_in_agent_catalog():
     entries = list_catalog(plugin_type="agent")
     ids = {entry.id for entry in entries}
@@ -381,6 +420,7 @@ async def noop_refresh(_plugin_type, _changed_path=None):
     webui_dir = home_plugins / "contracts" / "webui" / "soc_ui"
     access_dir = home_plugins / "contracts" / "access" / "soc_ui"
     dashboard_manifest = json.loads((webui_dir / "soc_dashboard" / "manifest.json").read_text(encoding="utf-8"))
+    workspace_manifest = json.loads((webui_dir / "workspace.json").read_text(encoding="utf-8"))
 
     assert (webui_dir / "workspace.json").is_file()
     assert (webui_dir / "soc_alerts" / "dist" / "page.js").is_file()
@@ -390,6 +430,7 @@ async def noop_refresh(_plugin_type, _changed_path=None):
     assert (access_dir / "soc_alerts_operations.py").is_file()
     assert set(built_pages) == {"soc-alerts", "soc-dashboard", "soc-overview"}
     assert dashboard_manifest["id"] == "soc-dashboard"
+    assert workspace_manifest["version"] == record.version == load_manifest("webui", "soc_ui").version
     assert record.installPath == str(webui_dir)
 
     removed = await uninstall_plugin("webui", "soc_ui")
@@ -475,6 +516,8 @@ async def noop_refresh(_plugin_type, _changed_path=None):
         "threat_rule_id": "D1181087257",
         "threat_name": "SQL injection",
         "threat_msg": "Detected SQL injection attempt.",
+        "threat_severity": "critical",
+        "threat_level": "high",
         "threat_phase": "exploit",
         "threat_type": "exploit",
         "threat_result": "failed",
@@ -545,6 +588,44 @@ async def noop_refresh(_plugin_type, _changed_path=None):
     assert response.body["summary"]["attackFailed"] == 1
     assert response.body["incidents"][0]["id"] == "alert-1"
     assert response.body["incidents"][0]["tableCells"]["_source_type"]["value"] == "tdp"
+    assert response.body["incidents"][0]["tableCells"]["threat_severity"]["value"] == "critical"
+    assert response.body["incidents"][0]["tableCells"]["threat_level"]["value"] == "high"
+
+    filtered = runtime.execute(
+        page_id="soc-alerts",
+        contract_id="soc.alerts.operations",
+        operation_name="list",
+        payload={
+            "params": {
+                "filters": {
+                    "threat_severity": ["critical"],
+                    "threat_level": ["high"],
+                },
+                "limit": 10,
+            }
+        },
+        principal=AuthUser(id="u1", username="admin", role="admin"),
+    )
+
+    assert filtered.status_code == 200
+    assert filtered.body["summary"]["representativeCount"] == 1
+    assert [incident["id"] for incident in filtered.body["incidents"]] == ["alert-1"]
+
+    for filters in (
+        {"threat_severity": ["low"]},
+        {"threat_level": ["low"]},
+    ):
+        excluded = runtime.execute(
+            page_id="soc-alerts",
+            contract_id="soc.alerts.operations",
+            operation_name="list",
+            payload={"params": {"filters": filters, "limit": 10}},
+            principal=AuthUser(id="u1", username="admin", role="admin"),
+        )
+
+        assert excluded.status_code == 200
+        assert excluded.body["summary"]["representativeCount"] == 0
+        assert excluded.body["incidents"] == []
 
 
 async def test_hub_installs_soc_workspace_component_children(isolated_hub_env, monkeypatch: pytest.MonkeyPatch):
diff --git a/tests/hub/test_installer_swap.py b/tests/hub/test_installer_swap.py
new file mode 100644
index 000000000..2ce6bbbbd
--- /dev/null
+++ b/tests/hub/test_installer_swap.py
@@ -0,0 +1,156 @@
+"""Unit tests for the Hub installer's Windows-safe directory swap helpers.
+
+These exercise the atomic-swap path in isolation (explicit ``tmp_path``
+dirs, no ``Path.home()`` dependency), covering the WinError 5 aftermath
+where a directory watcher or AV scan holds a transient handle on the
+freshly staged tree and blocks ``src.replace(dst)``.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from flocks.hub import installer
+
+
+def _make_tree(path: Path, marker: str) -> None:
+    path.mkdir(parents=True, exist_ok=True)
+    (path / "manifest.json").write_text(marker, encoding="utf-8")
+
+
+def test_replace_prepared_path_swaps_into_place(tmp_path):
+    src = tmp_path / ".soc_ui.scratch"
+    dst = tmp_path / "soc_ui"
+    _make_tree(src, "new")
+
+    backup = installer._replace_prepared_path(src, dst)
+
+    assert backup is None
+    assert (dst / "manifest.json").read_text(encoding="utf-8") == "new"
+    assert not src.exists()
+    assert not (tmp_path / ".soc_ui.bak").exists()
+
+
+def test_replace_prepared_path_overwrites_existing_and_returns_backup(tmp_path):
+    src = tmp_path / ".soc_ui.scratch"
+    dst = tmp_path / "soc_ui"
+    _make_tree(src, "new")
+    _make_tree(dst, "old")
+
+    backup = installer._replace_prepared_path(src, dst)
+
+    assert (dst / "manifest.json").read_text(encoding="utf-8") == "new"
+    assert backup == tmp_path / ".soc_ui.bak"
+    assert (backup / "manifest.json").read_text(encoding="utf-8") == "old"
+
+    installer._commit_replacement(backup)
+
+    assert not (tmp_path / ".soc_ui.bak").exists()
+
+
+def test_replace_prepared_path_restores_existing_after_new_swap_fails(tmp_path, monkeypatch):
+    monkeypatch.setattr(installer.sys, "platform", "win32")
+    monkeypatch.setattr(installer.time, "sleep", lambda _s: None)
+
+    src = tmp_path / ".soc_ui.scratch"
+    dst = tmp_path / "soc_ui"
+    _make_tree(src, "new")
+    _make_tree(dst, "old")
+
+    real_replace = Path.replace
+    failed_attempts = 0
+
+    def deny_new_swap(self, target):
+        nonlocal failed_attempts
+        if self == src and Path(target) == dst:
+            failed_attempts += 1
+            raise PermissionError("[WinError 5] Access is denied")
+        return real_replace(self, target)
+
+    monkeypatch.setattr(Path, "replace", deny_new_swap)
+    with pytest.raises(PermissionError):
+        installer._replace_prepared_path(src, dst)
+
+    assert failed_attempts == 6
+    assert (dst / "manifest.json").read_text(encoding="utf-8") == "old"
+    assert (src / "manifest.json").read_text(encoding="utf-8") == "new"
+    assert not (tmp_path / ".soc_ui.bak").exists()
+
+
+def test_replace_with_retry_recovers_from_transient_permission_error(tmp_path, monkeypatch):
+    """A first ``PermissionError`` (WinError 5) is retried, not surfaced."""
+    monkeypatch.setattr(installer.sys, "platform", "win32")
+    monkeypatch.setattr(installer.time, "sleep", lambda _s: None)
+
+    src = tmp_path / ".soc_ui.scratch"
+    dst = tmp_path / "soc_ui"
+    _make_tree(src, "new")
+
+    real_replace = Path.replace
+    calls = {"n": 0}
+
+    def flaky_replace(self, target):
+        calls["n"] += 1
+        if calls["n"] == 1:
+            raise PermissionError("[WinError 5] Access is denied")
+        return real_replace(self, target)
+
+    monkeypatch.setattr(Path, "replace", flaky_replace)
+    installer._replace_with_retry(src, dst)
+
+    assert calls["n"] == 2
+    assert (dst / "manifest.json").read_text(encoding="utf-8") == "new"
+
+
+def test_replace_with_retry_reraises_after_exhausting_attempts(tmp_path, monkeypatch):
+    monkeypatch.setattr(installer.sys, "platform", "win32")
+    monkeypatch.setattr(installer.time, "sleep", lambda _s: None)
+
+    src = tmp_path / ".soc_ui.scratch"
+    dst = tmp_path / "soc_ui"
+    _make_tree(src, "new")
+
+    def always_denied(self, target):
+        raise PermissionError("[WinError 5] Access is denied")
+
+    monkeypatch.setattr(Path, "replace", always_denied)
+    with pytest.raises(PermissionError):
+        installer._replace_with_retry(src, dst)
+
+
+def test_purge_stale_scratch_removes_leftovers(tmp_path):
+    parent = tmp_path
+    _make_tree(parent / ".soc_ui.55ram7wo" / "soc_overview", "stranded")
+    _make_tree(parent / ".soc_ui.bak" / "soc_dashboard", "stranded")
+    _make_tree(parent / "soc_ui", "live")
+    # Unrelated dot-dir for a different plugin must be left untouched.
+    _make_tree(parent / ".other.bak", "keep")
+
+    installer._purge_stale_scratch(parent, "soc_ui")
+
+    assert not (parent / ".soc_ui.55ram7wo").exists()
+    assert not (parent / ".soc_ui.bak").exists()
+    assert (parent / "soc_ui" / "manifest.json").read_text(encoding="utf-8") == "live"
+    assert (parent / ".other.bak").exists()
+
+
+def test_copy_package_purges_stale_scratch_before_staging(tmp_path):
+    """A prior failed install's leftovers self-heal on the next install."""
+    src = tmp_path / "bundled" / "soc_ui"
+    _make_tree(src, "payload")
+    (src / "manifest.json").write_text('{"id": "soc_ui"}', encoding="utf-8")
+    (src / "src").mkdir()
+    (src / "src" / "index.tsx").write_text("export default 1;\n", encoding="utf-8")
+
+    dst = tmp_path / "install" / "soc_ui"
+    dst.parent.mkdir(parents=True)
+    _make_tree(dst.parent / ".soc_ui.leftover" / "soc_overview", "stranded")
+
+    installer._copy_package(src, dst)
+
+    assert (dst / "src" / "index.tsx").is_file()
+    assert not (dst.parent / ".soc_ui.leftover").exists()
+    # ``manifest.json`` is intentionally not copied by the installer.
+    assert not (dst / "manifest.json").exists()
diff --git a/tests/hub/test_soc_dashboard_schema.py b/tests/hub/test_soc_dashboard_schema.py
index fc4257162..30e87ce49 100644
--- a/tests/hub/test_soc_dashboard_schema.py
+++ b/tests/hub/test_soc_dashboard_schema.py
@@ -30,11 +30,59 @@ def _load_dashboard_handlers():
     return module
 
 
+def _load_overview_handlers():
+    handler_path = (
+        Path(__file__).resolve().parents[2]
+        / ".flocks"
+        / "flockshub"
+        / "plugins"
+        / "webuis"
+        / "soc_ui"
+        / "soc_overview"
+        / "api"
+        / "handlers.py"
+    )
+    spec = importlib.util.spec_from_file_location("soc_overview_fields_test", handler_path)
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    previous = sys.dont_write_bytecode
+    try:
+        sys.dont_write_bytecode = True
+        spec.loader.exec_module(module)
+    finally:
+        sys.dont_write_bytecode = previous
+    return module
+
+
+def _load_alert_operations():
+    operations_path = (
+        Path(__file__).resolve().parents[2]
+        / ".flocks"
+        / "flockshub"
+        / "plugins"
+        / "webuis"
+        / "soc_ui"
+        / "access"
+        / "soc_alerts_operations.py"
+    )
+    spec = importlib.util.spec_from_file_location("soc_alert_operations_test", operations_path)
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    previous = sys.dont_write_bytecode
+    try:
+        sys.dont_write_bytecode = True
+        spec.loader.exec_module(module)
+    finally:
+        sys.dont_write_bytecode = previous
+    return module
+
+
 def test_soc_dashboard_migrates_legacy_alert_records_schema(tmp_path: Path):
     db_path = tmp_path / "soc.db"
     first_record = {
         "_source_type": "tdp",
         "threat_name": "SQL injection",
+        "_threat_type": "web_attack",
         "triage_status": "ok",
         "_triage_persisted_at": "2026-07-14T13:00:00",
         "attack_verdict": "attack",
@@ -74,6 +122,7 @@ def test_soc_dashboard_migrates_legacy_alert_records_schema(tmp_path: Path):
     second_record = {
         "_source_type": "hids",
         "threat_name": "Malware download",
+        "threat_type": "malware",
         "triage_status": "ok",
         "_triage_persisted_at": "2026-07-14T13:01:00",
     }
@@ -127,7 +176,7 @@ def test_soc_dashboard_migrates_legacy_alert_records_schema(tmp_path: Path):
         columns = {row[1] for row in conn.execute("PRAGMA table_info(alert_records)")}
         indexes = {row[1] for row in conn.execute("PRAGMA index_list(alert_records)")}
         facts = conn.execute(
-            "SELECT alert_row_id, row_key, source_type, threat_name, has_triage "
+            "SELECT alert_row_id, row_key, source_type, threat_name, threat_type, has_triage "
             "FROM soc_dashboard_alert_facts ORDER BY alert_row_id"
         ).fetchall()
         source_rows = conn.execute(
@@ -163,12 +212,189 @@ def test_soc_dashboard_migrates_legacy_alert_records_schema(tmp_path: Path):
     } <= indexes
     assert source_rows == [(1, "1", 0), (2, "2", 0)]
     assert facts == [
-        (1, "1", "tdp", "SQL injection", 1),
-        (2, "2", "hids", "Malware download", 1),
+        (1, "1", "tdp", "SQL injection", "web_attack", 1),
+        (2, "2", "hids", "Malware download", "malware", 1),
     ]
     assert "COALESCE(NULLIF(NEW.row_id, ''), CAST(NEW.rowid AS TEXT))" in trigger_sql
     assert updated_fact == ("2026-07-14T13:02:00", "attack")
-    assert schema_version == "2"
+    assert schema_version == "3"
+
+
+def test_soc_dashboard_triage_outcomes_partition_records(tmp_path: Path):
+    db_path = tmp_path / "soc.db"
+    asset_date = "2026-07-14"
+    records = [
+        {"triage_status": "ok", "attack_verdict": "attack_success", "attack_success": True},
+        {"triage_status": "ok", "attack_verdict": "attack"},
+        {"triage_status": "ok", "attack_verdict": "attack_failed"},
+        {"triage_status": "ok", "attack_verdict": "benign"},
+        {"triage_status": "ok", "attack_verdict": "unknown"},
+        {"triage_status": "failed", "attack_verdict": "unknown"},
+        {"triage_status": "failed", "attack_verdict": "benign"},
+        {"triage_status": "ok", "attack_verdict": "legacy", "attack_success": True},
+    ]
+    severity_values = ["low", "critical", "high", "medium", "low", "critical", "high", "medium"]
+    for index, record in enumerate(records):
+        record.update(
+            {
+                "threat_name": f"threat-name-{index}",
+                "_threat_type": f"threat-type-{index}",
+                "threat_type": f"fallback-type-{index}",
+                "threat_severity": severity_values[index],
+                "threat_level": "ignored-threat-level",
+                "risk_level": "High",
+            }
+        )
+    with sqlite3.connect(db_path) as conn:
+        conn.execute(
+            """
+            CREATE TABLE alert_records (
+                id INTEGER PRIMARY KEY AUTOINCREMENT,
+                record_json TEXT NOT NULL,
+                asset_date TEXT NOT NULL,
+                event_time INTEGER NOT NULL
+            )
+            """
+        )
+        conn.executemany(
+            "INSERT INTO alert_records(record_json, asset_date, event_time) VALUES (?, ?, ?)",
+            [
+                (json.dumps(record), asset_date, 1784014800 + index)
+                for index, record in enumerate(records)
+            ],
+        )
+        conn.commit()
+
+    handlers = _load_dashboard_handlers()
+    handlers.DEFAULT_SQLITE_DB = db_path
+    handlers._schema_ready.clear()
+
+    assert handlers._ensure_sqlite_schema() is True
+    sources = handlers._find_sqlite_sources(
+        asset_date,
+        asset_date,
+        1784014800,
+        1784014800 + len(records),
+    )
+    triage = handlers._read_triage(sources)
+    closed_loop = handlers._build_closed_loop(triage)
+    with sqlite3.connect(db_path) as conn:
+        timeline = handlers._sqlite_timeline(
+            conn,
+            handlers._sqlite_settings(),
+            "asset_date = ? AND event_time BETWEEN ? AND ?",
+            [asset_date, 1784014800, 1784014800 + len(records)],
+            [asset_date],
+            1784014800,
+            1784014800 + len(records),
+        )
+        first_fact = conn.execute(
+            "SELECT threat_name, threat_type, severity, risk_level "
+            "FROM soc_dashboard_alert_facts ORDER BY alert_row_id LIMIT 1"
+        ).fetchone()
+
+    assert triage["totalRecords"] == 8
+    assert triage["newTriaged"] == 6
+    assert triage["attackSuccess"] == 2
+    assert triage["attack"] == 1
+    assert triage["attackFailed"] == 1
+    assert triage["attackTotal"] == 4
+    assert triage["benign"] == 1
+    assert triage["unknown"] == 1
+    assert triage["triageFailed"] == 2
+    assert first_fact == ("threat-name-0", "threat-type-0", "low", "High")
+    assert dict(triage["threatTypeCounter"]) == {
+        f"threat-type-{index}": 1 for index in range(len(records))
+    }
+    assert dict(triage["severityCounter"]) == {
+        "critical": 2,
+        "high": 2,
+        "low": 2,
+        "medium": 2,
+    }
+    assert dict(triage["riskCounter"]) == {"high": 8}
+    assert closed_loop["pending"] == 3
+    assert triage["attackTotal"] + triage["benign"] + closed_loop["pending"] == 8
+    assert sum(timeline["attack"]) == triage["attackTotal"]
+
+
+def test_soc_overview_keeps_threat_names_and_types_separate(tmp_path: Path):
+    db_path = tmp_path / "soc.db"
+    asset_date = "2026-07-14"
+    records = [
+        {
+            "threat_name": "SQL injection attempt",
+            "_threat_type": "web_attack",
+            "threat_type": "ignored_fallback",
+        },
+        {
+            "threat_name": "Suspicious crawler",
+        },
+    ]
+    with sqlite3.connect(db_path) as conn:
+        conn.execute(
+            "CREATE TABLE alert_records "
+            "(record_json TEXT NOT NULL, asset_date TEXT NOT NULL, event_time INTEGER NOT NULL)"
+        )
+        conn.executemany(
+            "INSERT INTO alert_records VALUES (?, ?, ?)",
+            [
+                (json.dumps(record), asset_date, 1784014800 + index)
+                for index, record in enumerate(records)
+            ],
+        )
+        conn.commit()
+
+    handlers = _load_overview_handlers()
+    handlers.DEFAULT_SQLITE_DB = db_path
+    source = handlers._RecordSource(
+        path=db_path,
+        role="denoise",
+        date=asset_date,
+        data_source="sqlite",
+    )
+    denoise = handlers._read_denoise([source])
+    field_stats = handlers._build_field_stats([source])
+
+    assert dict(denoise["threatCounter"]) == {
+        "sql injection attempt": 1,
+        "suspicious crawler": 1,
+    }
+    assert {
+        item["label"]: item["value"] for item in field_stats["threatTypes"]
+    } == {"web_attack": 1, "unknown": 1}
+
+
+def test_soc_dashboard_activity_does_not_mix_name_type_or_risk_fields():
+    handlers = _load_dashboard_handlers()
+    row = {
+        "activity_row_id": 1,
+        "activity_event_time": 1784014800,
+        "sample_count": 1,
+        "record_json": json.dumps(
+            {
+                "_threat_type": "web_attack",
+                "triage_status": "ok",
+                "threat_level": "critical",
+            }
+        ),
+    }
+
+    event = handlers._activity_event(row)
+
+    assert event["alert"]["threatName"] == "未知告警"
+    assert event["alert"]["threatType"] == "web_attack"
+    assert event["result"]["threatSeverity"] == ""
+    assert event["result"]["riskLevel"] == ""
+
+
+def test_soc_alert_verdict_does_not_fall_back_to_risk_or_threat_level():
+    operations = _load_alert_operations()
+
+    assert operations._verdict_bucket(
+        {"risk_level": "attack_success", "threat_level": "attack_failed"}
+    ) == "unknown"
+    assert operations._verdict_bucket({"attack_verdict": "attack_success"}) == "success"
 
 
 def test_soc_dashboard_activity_exposes_live_denoise_workflow_progress(tmp_path: Path):
diff --git a/tests/observability/test_langfuse_observability.py b/tests/observability/test_langfuse_observability.py
index 8473ac802..8dcc02ef8 100644
--- a/tests/observability/test_langfuse_observability.py
+++ b/tests/observability/test_langfuse_observability.py
@@ -27,6 +27,15 @@ def trace(self, **kwargs):
         return _FakeObservation("trace", kwargs)
 
 
+class _NewSdkTraceClient:
+    def __init__(self):
+        self.start_observation_payload = None
+
+    def start_observation(self, **kwargs):
+        self.start_observation_payload = kwargs
+        return _TrackingObservation("trace", kwargs)
+
+
 class _FakeObservation:
     """Fake Langfuse observation with end/generation/span support."""
 
@@ -45,6 +54,41 @@ def end(self, **kwargs):
         self.end_payload = kwargs
 
 
+class _TrackingSpan:
+    def __init__(self) -> None:
+        self.attributes = {}
+
+    def is_recording(self):
+        return True
+
+    def set_attribute(self, key, value):
+        self.attributes[key] = value
+
+
+class _TrackingObservation(_FakeObservation):
+    def __init__(self, kind: str, payload: dict):
+        super().__init__(kind, payload)
+        self._otel_span = _TrackingSpan()
+
+    def generation(self, **kwargs):
+        return _TrackingObservation("generation", kwargs)
+
+    def span(self, **kwargs):
+        return _TrackingObservation("span", kwargs)
+
+
+class _NewSdkLikeObservation:
+    def __init__(self) -> None:
+        self.update_payload = None
+        self.end_calls = 0
+
+    def update(self, **kwargs):
+        self.update_payload = kwargs
+
+    def end(self):
+        self.end_calls += 1
+
+
 def test_create_span_uses_current_observation(monkeypatch):
     monkeypatch.setattr(lf, "_get_client", lambda: object())
     parent = _FakeParent()
@@ -92,6 +136,55 @@ def test_create_trace_forwards_tags(monkeypatch):
     assert client.trace_payload["tags"] == ["session:s1", "step:2", "session_step:s1:2"]
 
 
+def test_create_trace_uses_start_observation_for_new_sdk(monkeypatch):
+    client = _NewSdkTraceClient()
+    monkeypatch.setattr(lf, "_get_client", lambda: client)
+
+    obs = lf.create_trace(
+        name="SessionRunner.step",
+        session_id="s1",
+        user_id="u1",
+        tags=["session:s1", "step:2"],
+        input={"step": 2},
+        metadata={"provider_id": "openai"},
+    )
+
+    assert obs.kind == "trace"
+    assert client.start_observation_payload is not None
+    assert client.start_observation_payload["as_type"] == "span"
+    assert client.start_observation_payload["input"] == {"step": 2}
+    assert client.start_observation_payload["metadata"]["provider_id"] == "openai"
+    assert client.start_observation_payload["metadata"]["session_id"] == "s1"
+    assert client.start_observation_payload["metadata"]["user_id"] == "u1"
+    assert client.start_observation_payload["metadata"]["tags"] == ["session:s1", "step:2"]
+    assert obs._otel_span.attributes["langfuse.trace.name"] == "SessionRunner.step"
+    assert obs._otel_span.attributes["session.id"] == "s1"
+    assert obs._otel_span.attributes["user.id"] == "u1"
+    assert obs._otel_span.attributes["langfuse.trace.tags"] == ["session:s1", "step:2"]
+
+
+def test_generation_and_span_inherit_trace_dimensions_from_parent(monkeypatch):
+    monkeypatch.setattr(lf, "_get_client", lambda: object())
+
+    parent = _TrackingObservation("trace", {"name": "trace"})
+    parent._otel_span.attributes["langfuse.trace.name"] = "SessionRunner.step"
+    parent._otel_span.attributes["session.id"] = "s1"
+    parent._otel_span.attributes["user.id"] = "u1"
+    parent._otel_span.attributes["langfuse.trace.tags"] = ["session:s1", "step:2"]
+
+    gen = lf.create_generation(parent=parent, name="LLM.generate", model="gpt-5", input={"x": 1})
+    span = lf.create_span(parent=parent, name="Tool.execute.read", input={"path": "/tmp/a"})
+
+    assert gen._otel_span.attributes["langfuse.trace.name"] == "SessionRunner.step"
+    assert gen._otel_span.attributes["session.id"] == "s1"
+    assert gen._otel_span.attributes["user.id"] == "u1"
+    assert gen._otel_span.attributes["langfuse.trace.tags"] == ["session:s1", "step:2"]
+    assert span._otel_span.attributes["langfuse.trace.name"] == "SessionRunner.step"
+    assert span._otel_span.attributes["session.id"] == "s1"
+    assert span._otel_span.attributes["user.id"] == "u1"
+    assert span._otel_span.attributes["langfuse.trace.tags"] == ["session:s1", "step:2"]
+
+
 def test_initialize_supports_langfuse_base_url(monkeypatch):
     class _FakeLangfuseClient:
         def __init__(self, **kwargs):
@@ -156,6 +249,29 @@ def test_end_observation_passes_usage(monkeypatch):
     assert gen_obs.end_payload.get("output") == "result"
 
 
+def test_end_observation_updates_before_end_for_new_sdk():
+    """Langfuse v4-style observations require update(...), then end()."""
+    obs = _NewSdkLikeObservation()
+    usage = {"prompt_tokens": 100, "completion_tokens": 50}
+
+    lf.end_observation(
+        obs,
+        output={"content": "result"},
+        metadata={"status": "ok"},
+        usage=usage,
+        level="ERROR",
+        status_message="done",
+    )
+
+    assert obs.update_payload is not None
+    assert obs.update_payload["output"] == {"content": "result"}
+    assert obs.update_payload["metadata"] == {"status": "ok"}
+    assert obs.update_payload["usage_details"] == usage
+    assert obs.update_payload["level"] == "ERROR"
+    assert obs.update_payload["status_message"] == "done"
+    assert obs.end_calls == 1
+
+
 def test_scope_end_is_idempotent():
     """Calling end() twice on a scope should not raise."""
     noop = lf._NoopObservation("test")
@@ -164,11 +280,21 @@ def test_scope_end_is_idempotent():
     scope.end(output="second")
 
 
-def test_sanitize_truncates_long_strings():
+def test_sanitize_keeps_full_strings_in_full_mode(monkeypatch):
+    long_str = "a" * 10000
+    monkeypatch.setenv("FLOCKS_LANGFUSE_CAPTURE_MODE", "full")
+    result = lf._sanitize_payload(long_str)
+    assert result == long_str
+
+
+def test_sanitize_truncates_long_strings_in_truncated_mode(monkeypatch):
     long_str = "a" * 10000
+    monkeypatch.setenv("FLOCKS_LANGFUSE_CAPTURE_MODE", "truncated")
+    monkeypatch.setenv("FLOCKS_LANGFUSE_MAX_CHARS", "128")
     result = lf._sanitize_payload(long_str)
     assert len(result) < 10000
     assert "truncated" in result
+    assert result.startswith("a" * 128)
 
 
 def test_observation_scope_exception_handling():
diff --git a/tests/project/test_project.py b/tests/project/test_project.py
index bd9e5e7a2..364e97c1e 100644
--- a/tests/project/test_project.py
+++ b/tests/project/test_project.py
@@ -1,5 +1,7 @@
 import asyncio
 import json
+import os
+import sys
 import uuid
 from unittest.mock import AsyncMock, patch
 
@@ -13,6 +15,8 @@
     ProjectPathConflictError,
     TASK_SESSION_GROUP_ID,
 )
+from flocks.config.config import Config
+from flocks.session.session import Session
 from flocks.storage.storage import Storage
 
 
@@ -24,6 +28,49 @@ def project_root(tmp_path, monkeypatch):
     return tmp_path
 
 
+@pytest.mark.asyncio
+async def test_lifecycle_guard_is_reentrant_and_cross_process(tmp_path, monkeypatch):
+    monkeypatch.setattr(
+        Config,
+        "get_data_path",
+        classmethod(lambda _cls: tmp_path),
+    )
+    marker = tmp_path / "child-acquired"
+    project_id = "prj_cross_process_guard"
+    script = """
+import os
+import sys
+from pathlib import Path
+from flocks.project.project import _platform_file_lock, _platform_file_unlock
+
+fd = os.open(sys.argv[1], os.O_RDWR)
+try:
+    _platform_file_lock(fd)
+    Path(sys.argv[2]).touch()
+    _platform_file_unlock(fd)
+finally:
+    os.close(fd)
+"""
+
+    async with Project.lifecycle_guard(project_id):
+        async with Project.lifecycle_guard(project_id):
+            pass
+        lock_path = next((tmp_path / "locks" / "project-lifecycle").glob("*.lock"))
+        process = await asyncio.create_subprocess_exec(
+            sys.executable,
+            "-c",
+            script,
+            str(lock_path),
+            str(marker),
+            cwd=os.getcwd(),
+        )
+        await asyncio.sleep(0.1)
+        assert not marker.exists()
+
+    assert await asyncio.wait_for(process.wait(), timeout=5) == 0
+    assert marker.exists()
+
+
 @pytest.mark.asyncio
 async def test_list_projects_uses_json_registry_without_virtual_default(project_root):
     labs = project_root / "labs"
@@ -115,7 +162,7 @@ async def test_update_renames_registered_project(project_root):
 
 
 @pytest.mark.asyncio
-async def test_delete_only_removes_registry_entry(project_root):
+async def test_delete_soft_removes_registry_entry_and_restore_preserves_identity(project_root):
     labs = project_root / "labs"
     labs.mkdir()
     created = await Project.create(owner_id="user-1", name="Labs", worktree=str(labs))
@@ -126,8 +173,51 @@ async def test_delete_only_removes_registry_entry(project_root):
     assert result is True
     assert labs.exists()
     assert await Project.get(created.id, owner_id="user-1") is None
+    assert await Project.list(owner_id="user-1") == []
+    assert Project.registered_project_names()[created.id] == "Labs"
+    registry = json.loads(Project.registry_path("user-1").read_text(encoding="utf-8"))
+    assert registry["projects"][0]["id"] == created.id
+    assert registry["projects"][0]["removedAt"] > 0
     storage_delete.assert_not_awaited()
 
+    restored = await Project.restore(created.id, owner_id="user-1")
+
+    assert restored is not None
+    assert restored.id == created.id
+    assert restored.name == "Labs"
+    assert [project.id for project in await Project.list(owner_id="user-1")] == [created.id]
+
+
+@pytest.mark.asyncio
+async def test_restore_removed_project_requires_original_directory(project_root):
+    labs = project_root / "labs"
+    labs.mkdir()
+    created = await Project.create(owner_id="user-1", name="Labs", worktree=str(labs))
+    await Project.delete(created.id, owner_id="user-1")
+    labs.rmdir()
+
+    with pytest.raises(ProjectDeletionError, match="directory is unavailable"):
+        await Project.restore(created.id, owner_id="user-1")
+
+    assert await Project.get(created.id, owner_id="user-1") is None
+
+
+@pytest.mark.asyncio
+async def test_removed_project_rejects_direct_session_creation(project_root):
+    labs = project_root / "removed"
+    labs.mkdir()
+    created = await Project.create(owner_id="user-1", name="Removed", worktree=str(labs))
+    await Project.delete(created.id, owner_id="user-1")
+
+    with pytest.raises(ProjectDeletionError, match="no longer available"):
+        await Session.create(project_id=created.id, directory=str(labs))
+
+
+@pytest.mark.asyncio
+async def test_restore_missing_project_metadata_is_rejected(project_root):
+    with pytest.raises(ProjectDeletionError, match="metadata is unavailable"):
+        await Project.restore("prj_missing", owner_id="user-1")
+
 
 @pytest.mark.asyncio
 async def test_delete_rejects_default_project(project_root):
diff --git a/tests/provider/test_chinese_providers.py b/tests/provider/test_chinese_providers.py
index af892d611..1245b52fe 100644
--- a/tests/provider/test_chinese_providers.py
+++ b/tests/provider/test_chinese_providers.py
@@ -169,12 +169,43 @@ def test_alibaba_catalog(self):
     def test_moonshot_catalog(self):
         models = get_provider_model_definitions("moonshot")
         assert {m.id for m in models} == {
+            "kimi-k3",
+            "kimi-k2.7-code",
+            "kimi-k2.7-code-highspeed",
             "kimi-k2.5",
             "kimi-k2.6",
-            "kimi-k2-thinking",
-            "kimi-k2",
         }
 
+        k3 = next(m for m in models if m.id == "kimi-k3")
+        assert k3.capabilities.supports_vision is True
+        assert k3.capabilities.supports_reasoning is True
+        assert k3.capabilities.interleaved["field"] == "reasoning_content"
+        assert k3.pricing.input == 20.0
+        assert k3.pricing.output == 100.0
+        assert k3.pricing.cache_read == 2.0
+        assert k3.limits.context_window == 1048576
+        assert k3.limits.max_output_tokens == 131072
+
+        k27 = next(m for m in models if m.id == "kimi-k2.7-code")
+        assert k27.capabilities.supports_vision is True
+        assert k27.capabilities.supports_reasoning is True
+        assert k27.capabilities.interleaved["field"] == "reasoning_content"
+        assert k27.pricing.input == 6.5
+        assert k27.pricing.output == 27.0
+        assert k27.pricing.cache_read == 1.3
+        assert k27.limits.context_window == 262144
+        assert k27.limits.max_output_tokens == 32768
+
+        k27_highspeed = next(m for m in models if m.id == "kimi-k2.7-code-highspeed")
+        assert k27_highspeed.capabilities.supports_vision is True
+        assert k27_highspeed.capabilities.supports_reasoning is True
+        assert k27_highspeed.capabilities.interleaved["field"] == "reasoning_content"
+        assert k27_highspeed.pricing.input == 13.0
+        assert k27_highspeed.pricing.output == 54.0
+        assert k27_highspeed.pricing.cache_read == 2.6
+        assert k27_highspeed.limits.context_window == 262144
+        assert k27_highspeed.limits.max_output_tokens == 32768
+
         k26 = next(m for m in models if m.id == "kimi-k2.6")
         assert k26.capabilities.supports_vision is True
         assert k26.capabilities.supports_reasoning is True
@@ -184,8 +215,6 @@ def test_moonshot_catalog(self):
         assert k26.pricing.cache_read == 1.3
         assert k26.limits.context_window == 256000
 
-        thinking = next(m for m in models if m.id == "kimi-k2-thinking")
-        assert thinking.capabilities.supports_reasoning is True
         k25 = next(m for m in models if m.id == "kimi-k2.5")
         assert k25.capabilities.supports_reasoning is True
         assert k25.capabilities.interleaved["field"] == "reasoning_content"
diff --git a/tests/provider/test_openai_base_provider.py b/tests/provider/test_openai_base_provider.py
index 5c2202411..6f7fbd72f 100644
--- a/tests/provider/test_openai_base_provider.py
+++ b/tests/provider/test_openai_base_provider.py
@@ -624,6 +624,54 @@ async def test_chat_logs_compact_message_summary(self):
             "assistant": 1,
         }
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("model_id", "extra_body"),
+        [
+            ("kimi-k2.6", {"thinking": {"type": "enabled"}}),
+            ("kimi-k2.7-code", {"thinking": {"type": "enabled"}}),
+            ("kimi-k2.7-code-highspeed", {"thinking": {"type": "enabled"}}),
+            ("kimi-k3", {"reasoning_effort": "max"}),
+        ],
+    )
+    async def test_chat_logs_effective_extra_body_thinking_state(
+        self,
+        model_id,
+        extra_body,
+    ):
+        provider, create = self._build_provider_with_client()
+        create.return_value = self._mock_chat_response()
+        info = Mock()
+
+        from flocks.provider.provider import ChatMessage
+
+        with patch.object(openai_base_module, "log", Mock(info=info)):
+            await provider.chat(
+                model_id,
+                [ChatMessage(role="user", content="hello")],
+                extra_body=extra_body,
+            )
+
+        _event, payload = info.call_args.args
+        assert payload["thinking_enabled"] is True
+
+    @pytest.mark.asyncio
+    async def test_kimi_k3_uses_max_completion_tokens(self):
+        provider, create = self._build_provider_with_client()
+        create.return_value = self._mock_chat_response()
+
+        from flocks.provider.provider import ChatMessage
+
+        await provider.chat(
+            "kimi-k3",
+            [ChatMessage(role="user", content="hello")],
+            max_tokens=131072,
+        )
+
+        kwargs = create.await_args.kwargs
+        assert kwargs["max_completion_tokens"] == 131072
+        assert "max_tokens" not in kwargs
+
 
 class TestExtractReasoningContent:
     """Regression: some proxies send stream chunks with ``delta is None``."""
@@ -657,6 +705,108 @@ async def _stream_from_chunks(*chunks):
         yield chunk
 
 
+class TestOpenAIBaseProviderStreamingToolCalls:
+    @pytest.mark.asyncio
+    async def test_chat_stream_emits_name_only_marker_before_complete_tool_input(self):
+        provider = MockProviderWithoutCatalog()
+        create = AsyncMock()
+        provider._client = MagicMock()
+        provider._client.chat.completions.create = create
+
+        first_arguments = '{"filePath":"/tmp/report.md","content":"'
+        remaining_arguments = f'{"long content " * 500}"}}'
+        tool_start_chunk = SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=SimpleNamespace(
+                        content=None,
+                        tool_calls=[
+                            SimpleNamespace(
+                                index=0,
+                                id="call_write",
+                                function=SimpleNamespace(
+                                    name="write",
+                                    arguments=first_arguments,
+                                ),
+                            )
+                        ],
+                    ),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        tool_arguments_chunk = SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=SimpleNamespace(
+                        content=None,
+                        tool_calls=[
+                            SimpleNamespace(
+                                index=0,
+                                id=None,
+                                function=SimpleNamespace(
+                                    name=None,
+                                    arguments=remaining_arguments,
+                                ),
+                            )
+                        ],
+                    ),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        finish_chunk = SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=SimpleNamespace(content=None, tool_calls=None),
+                    finish_reason="tool_calls",
+                )
+            ],
+            usage=None,
+        )
+        create.return_value = _stream_from_chunks(
+            tool_start_chunk,
+            tool_arguments_chunk,
+            finish_chunk,
+        )
+
+        from flocks.provider.provider import ChatMessage
+
+        chunks = [
+            chunk
+            async for chunk in provider.chat_stream(
+                "kimi-k2.7-code",
+                [ChatMessage(role="user", content="write a long file")],
+                tools=[{"type": "function", "function": {"name": "write"}}],
+            )
+        ]
+
+        assert len(chunks) == 2
+        assert chunks[0].finish_reason is None
+        assert chunks[0].tool_calls == [
+            {
+                "index": 0,
+                "id": "call_write",
+                "type": "function",
+                "function": {"name": "write", "arguments": ""},
+            }
+        ]
+        assert chunks[1].finish_reason == "tool_calls"
+        assert chunks[1].tool_calls == [
+            {
+                "index": 0,
+                "id": "call_write",
+                "type": "function",
+                "function": {
+                    "name": "write",
+                    "arguments": first_arguments + remaining_arguments,
+                },
+            }
+        ]
+
+
 class TestOpenAIBaseProviderStreamingUsage:
     @staticmethod
     def _build_provider_with_stream():
@@ -716,6 +866,53 @@ async def test_chat_stream_includes_usage_and_attaches_to_terminal_chunk(self):
             "total_tokens": 18,
         }
 
+    @pytest.mark.asyncio
+    async def test_chat_stream_preserves_nested_reasoning_tokens(self):
+        provider, create = self._build_provider_with_stream()
+
+        content_chunk = SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=SimpleNamespace(content="hello", tool_calls=None),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        usage_chunk = SimpleNamespace(
+            choices=[],
+            usage=SimpleNamespace(
+                prompt_tokens=11,
+                completion_tokens=7,
+                total_tokens=18,
+                completion_tokens_details=SimpleNamespace(reasoning_tokens=5),
+            ),
+        )
+        finish_chunk = SimpleNamespace(
+            choices=[SimpleNamespace(delta=None, finish_reason="stop")],
+            usage=None,
+        )
+        create.return_value = _stream_from_chunks(
+            content_chunk, usage_chunk, finish_chunk
+        )
+
+        from flocks.provider.provider import ChatMessage
+
+        chunks = [
+            chunk
+            async for chunk in provider.chat_stream(
+                "gpt-5.6-luna",
+                [ChatMessage(role="user", content="hello")],
+            )
+        ]
+
+        assert chunks[-1].usage == {
+            "prompt_tokens": 11,
+            "completion_tokens": 2,
+            "total_tokens": 18,
+            "reasoning_tokens": 5,
+        }
+
     @pytest.mark.asyncio
     async def test_chat_stream_emits_trailing_usage_when_usage_only_chunk_arrives_after_finish(self):
         provider, create = self._build_provider_with_stream()
diff --git a/tests/provider/test_openai_compatible_provider.py b/tests/provider/test_openai_compatible_provider.py
index ee8389fe6..92efbe651 100644
--- a/tests/provider/test_openai_compatible_provider.py
+++ b/tests/provider/test_openai_compatible_provider.py
@@ -183,6 +183,49 @@ async def test_chat_does_not_fallback_for_completion_token_value_errors(self):
 
         assert create.await_count == 1
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("model_id", "extra_body"),
+        [
+            ("kimi-k2.7-code-highspeed", {"thinking": {"type": "enabled"}}),
+            ("kimi-k3", {"reasoning_effort": "max"}),
+        ],
+    )
+    async def test_stream_logs_effective_kimi_thinking_state(
+        self,
+        model_id,
+        extra_body,
+    ):
+        provider, create = _build_provider_with_client()
+        provider.log = MagicMock()
+        create.return_value = _stream_from_chunks(
+            SimpleNamespace(
+                choices=[
+                    SimpleNamespace(
+                        delta=SimpleNamespace(content="ok", tool_calls=None),
+                        finish_reason="stop",
+                    )
+                ],
+                usage=None,
+            )
+        )
+
+        _ = [
+            chunk
+            async for chunk in provider.chat_stream(
+                model_id,
+                [ChatMessage(role="user", content="hello")],
+                extra_body=extra_body,
+            )
+        ]
+
+        request_logs = [
+            call.args
+            for call in provider.log.info.call_args_list
+            if call.args[0] == "openai_compatible.stream.request"
+        ]
+        assert request_logs[0][1]["thinking_enabled"] is True
+
 
 class TestOpenAICompatibleProviderMiniMaxFallback:
     def test_is_minimax_empty_response_target_matches_all_minimax_aliases(self):
diff --git a/tests/provider/test_openai_provider.py b/tests/provider/test_openai_provider.py
index cc78a6972..10f6211d4 100644
--- a/tests/provider/test_openai_provider.py
+++ b/tests/provider/test_openai_provider.py
@@ -1,9 +1,17 @@
-from unittest.mock import MagicMock, patch
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock, patch
 
-from flocks.provider.provider import ProviderConfig
+import pytest
+
+from flocks.provider.provider import ChatMessage, ProviderConfig
 from flocks.provider.sdk.openai import OpenAIProvider
 
 
+async def _stream_from_chunks(*chunks):
+    for chunk in chunks:
+        yield chunk
+
+
 class TestOpenAIProviderConfiguration:
     @patch("httpx.AsyncClient")
     @patch("openai.AsyncOpenAI")
@@ -40,3 +48,57 @@ def test_get_client_respects_verify_ssl_false(self, mock_async_openai, mock_http
             base_url="https://gateway.internal/v1",
             http_client=http_client,
         )
+
+    def test_configure_invalidates_existing_client_when_credentials_change(self):
+        provider = OpenAIProvider()
+        stale_client = MagicMock()
+        provider._client = stale_client
+
+        provider.configure(
+            ProviderConfig(
+                provider_id=provider.id,
+                api_key="new-api-key",
+                base_url="https://new-gateway.internal/v1",
+            )
+        )
+
+        assert provider._client is None
+
+
+class TestOpenAIProviderStreamingUsage:
+    @pytest.mark.asyncio
+    async def test_chat_stream_preserves_nested_reasoning_tokens(self):
+        provider = OpenAIProvider()
+        create = AsyncMock()
+        provider._client = MagicMock()
+        provider._client.chat.completions.create = create
+
+        usage_chunk = SimpleNamespace(
+            choices=[],
+            usage=SimpleNamespace(
+                prompt_tokens=11,
+                completion_tokens=7,
+                total_tokens=18,
+                completion_tokens_details=SimpleNamespace(reasoning_tokens=5),
+            ),
+        )
+        finish_chunk = SimpleNamespace(
+            choices=[SimpleNamespace(delta=None, finish_reason="stop")],
+            usage=None,
+        )
+        create.return_value = _stream_from_chunks(finish_chunk, usage_chunk)
+
+        chunks = [
+            chunk
+            async for chunk in provider.chat_stream(
+                "gpt-5.6-luna",
+                [ChatMessage(role="user", content="hello")],
+            )
+        ]
+
+        assert chunks[-1].usage == {
+            "prompt_tokens": 11,
+            "completion_tokens": 2,
+            "total_tokens": 18,
+            "reasoning_tokens": 5,
+        }
diff --git a/tests/provider/test_provider.py b/tests/provider/test_provider.py
index 6309bb141..c4d9fa843 100644
--- a/tests/provider/test_provider.py
+++ b/tests/provider/test_provider.py
@@ -221,6 +221,7 @@ def test_resolve_model_does_not_infer_interleaved_for_non_reasoning_model(monkey
     [
         ("openai-compatible", "qwen3-235b-a22b-thinking", "https://api.example.com/v1", "reasoning_content"),
         ("openai-compatible", "kimi-k2-thinking-turbo", "https://api.example.com/v1", "reasoning_content"),
+        ("openai-compatible", "kimi-k3", "https://api.example.com/v1", "reasoning_content"),
         ("openai-compatible", "deepseek-v4-pro", "https://api.deepseek.com/v1", "reasoning_content"),
         ("openai-compatible", "glm-4.7", "https://api.example.com/v1", "reasoning_content"),
         ("openai-compatible", "minimax-m3", "https://api.example.com/v1", "reasoning_details"),
diff --git a/tests/provider/test_provider_apply_config_idempotent.py b/tests/provider/test_provider_apply_config_idempotent.py
index b00af163d..3f8aa54c8 100644
--- a/tests/provider/test_provider_apply_config_idempotent.py
+++ b/tests/provider/test_provider_apply_config_idempotent.py
@@ -67,7 +67,11 @@ def __exit__(self, *_args, **_kwargs):
         self.real_provider.configure = self._original_configure
 
 
-def _build_fake_config(provider_id: str) -> Any:
+def _build_fake_config(
+    provider_id: str,
+    *,
+    model_first_chunk_timeout_s: int = 480,
+) -> Any:
     """Return a SimpleNamespace mimicking ``ConfigInfo`` shape needed by
     ``apply_config``: ``.provider`` dict -> ``.options`` (with model_dump)
     and ``.models``.
@@ -88,6 +92,7 @@ def _build_fake_config(provider_id: str) -> Any:
                 "supports_vision": False,
                 "supports_reasoning": False,
                 "max_tokens": 4096,
+                "stream_first_chunk_timeout_s": model_first_chunk_timeout_s,
             }
         )
     }
@@ -128,6 +133,9 @@ async def test_apply_config_is_idempotent_for_unchanged_input() -> None:
             "apply_config must not rebuild _config_models when the desired "
             "model list matches the existing one"
         )
+        assert provider.get_models()[0].custom_settings == {
+            "stream_first_chunk_timeout_s": 480,
+        }
 
 
 @pytest.mark.asyncio
@@ -171,3 +179,87 @@ async def test_apply_config_still_mutates_when_input_changes() -> None:
     assert rec.configure_calls == 1, (
         f"expected exactly 1 configure call on api_key change, got {rec.configure_calls}"
     )
+
+
+@pytest.mark.asyncio
+async def test_apply_config_rebuilds_models_when_stream_timeout_changes() -> None:
+    """A model timeout change must invalidate the model-list signature."""
+@pytest.mark.parametrize(
+    "options",
+    [
+        None,
+        SimpleNamespace(
+            model_dump=lambda exclude_none, by_alias: {
+                "api_key": " ",
+                "base_url": "",
+            }
+        ),
+    ],
+    ids=["missing-options", "empty-credentials"],
+)
+async def test_apply_config_loads_name_and_models_without_credentials(
+    options: Any,
+) -> None:
+    """Provider metadata must not depend on resolved credentials."""
+    Provider._ensure_initialized()
+    provider = Provider.get("openai-compatible")
+    assert provider is not None
+
+    first_cfg = _build_fake_config(
+        "openai-compatible",
+        model_first_chunk_timeout_s=480,
+    )
+    with patch("flocks.provider.provider.Config.get", return_value=first_cfg):
+        await Provider.apply_config(provider_id="openai-compatible")
+
+    second_cfg = _build_fake_config(
+        "openai-compatible",
+        model_first_chunk_timeout_s=600,
+    )
+    with patch(
+        "flocks.provider.provider.Config.get",
+        return_value=second_cfg,
+    ), _Recorder(provider) as rec:
+        await Provider.apply_config(provider_id="openai-compatible")
+
+    assert rec.models_assignments == 1
+    assert provider.get_models()[0].custom_settings == {
+        "stream_first_chunk_timeout_s": 600,
+    }
+    original_config = provider._config
+    original_models = list(getattr(provider, "_config_models", []) or [])
+    original_name = provider.name
+    model_id = f"credentialless-model-{id(options)}"
+    fake_cfg = SimpleNamespace(
+        provider={
+            "openai-compatible": SimpleNamespace(
+                name="Credentialless Provider",
+                options=options,
+                models={
+                    model_id: {
+                        "name": "Credentialless Model",
+                        "supports_tools": True,
+                    }
+                },
+            )
+        }
+    )
+
+    try:
+        with _Recorder(provider) as first:
+            await Provider.apply_config(fake_cfg, provider_id="openai-compatible")
+
+        assert first.configure_calls == 0
+        assert first.models_assignments == 1
+        assert provider.name == "Credentialless Provider"
+        assert [model.id for model in provider._config_models] == [model_id]
+
+        with _Recorder(provider) as second:
+            await Provider.apply_config(fake_cfg, provider_id="openai-compatible")
+
+        assert second.configure_calls == 0
+        assert second.models_assignments == 0
+    finally:
+        provider._config = original_config
+        provider._config_models = original_models
+        provider.name = original_name
diff --git a/tests/provider/test_provider_options.py b/tests/provider/test_provider_options.py
index 4e8f11be0..45f2af82b 100644
--- a/tests/provider/test_provider_options.py
+++ b/tests/provider/test_provider_options.py
@@ -58,6 +58,125 @@ def test_moonshot_kimi_hybrid_models_use_official_thinking_payload(self):
 
         assert options["extra_body"] == KIMI_THINKING_EXTRA_BODY
 
+    def test_moonshot_kimi_k27_forces_official_thinking_payload(self):
+        options = provider_options.build_provider_options(
+            "moonshot",
+            "kimi-k2.7-code",
+            reasoning_enabled=False,
+            resolve_max_tokens=False,
+        )
+
+        assert options["extra_body"] == KIMI_THINKING_EXTRA_BODY
+
+    def test_moonshot_kimi_k27_highspeed_forces_official_thinking_payload(self):
+        options = provider_options.build_provider_options(
+            "moonshot",
+            "kimi-k2.7-code-highspeed",
+            reasoning_enabled=False,
+            resolve_max_tokens=False,
+        )
+
+        assert options["extra_body"] == KIMI_THINKING_EXTRA_BODY
+
+    def test_moonshot_kimi_k3_uses_default_reasoning_effort(self):
+        options = provider_options.build_provider_options(
+            "moonshot",
+            "kimi-k3",
+            reasoning_enabled=False,
+            resolve_max_tokens=False,
+        )
+
+        assert options["extra_body"] == {"reasoning_effort": "max"}
+
+    def test_kimi_k27_forces_thinking_even_when_toggle_is_disabled(self):
+        options = provider_options.build_provider_options(
+            "threatbook-cn-llm",
+            "kimi-k2.7-code-highspeed",
+            reasoning_enabled=False,
+            resolve_max_tokens=False,
+        )
+
+        assert options["extra_body"] == KIMI_THINKING_EXTRA_BODY
+
+    def test_kimi_k27_overrides_configured_disabled_thinking(self, monkeypatch):
+        monkeypatch.setattr(
+            provider_options,
+            "_resolve_default_extra_body",
+            lambda *_args: {
+                "thinking": {"type": "disabled"},
+                "custom_option": True,
+            },
+        )
+
+        options = provider_options.build_provider_options(
+            "openai-compatible",
+            "kimi-k2.7-code",
+            resolve_max_tokens=False,
+        )
+
+        assert options["extra_body"] == {
+            "thinking": {"type": "enabled"},
+            "custom_option": True,
+        }
+
+    def test_kimi_k3_uses_reasoning_effort_instead_of_thinking(self):
+        options = provider_options.build_provider_options(
+            "openai-compatible",
+            "kimi-k3",
+            reasoning_enabled=False,
+            resolve_max_tokens=False,
+        )
+
+        assert options["extra_body"] == {"reasoning_effort": "max"}
+        assert "thinking" not in options["extra_body"]
+
+    def test_kimi_k3_respects_supported_reasoning_effort(self):
+        options = provider_options.build_provider_options(
+            "openai-compatible",
+            "kimi-k3",
+            reasoning_effort="low",
+            resolve_max_tokens=False,
+        )
+
+        assert options["extra_body"] == {"reasoning_effort": "low"}
+
+    def test_kimi_k3_preserves_configured_extra_body_reasoning_effort(
+        self,
+        monkeypatch,
+    ):
+        monkeypatch.setattr(
+            provider_options,
+            "_resolve_default_extra_body",
+            lambda *_args: {"reasoning_effort": "high"},
+        )
+
+        options = provider_options.build_provider_options(
+            "openai-compatible",
+            "kimi-k3",
+            resolve_max_tokens=False,
+        )
+
+        assert options["extra_body"] == {"reasoning_effort": "high"}
+
+    def test_kimi_k3_explicit_reasoning_effort_overrides_configured_default(
+        self,
+        monkeypatch,
+    ):
+        monkeypatch.setattr(
+            provider_options,
+            "_resolve_default_extra_body",
+            lambda *_args: {"reasoning_effort": "high"},
+        )
+
+        options = provider_options.build_provider_options(
+            "openai-compatible",
+            "kimi-k3",
+            reasoning_effort="low",
+            resolve_max_tokens=False,
+        )
+
+        assert options["extra_body"] == {"reasoning_effort": "low"}
+
     def test_openai_compatible_qwen_models_enable_thinking_by_default(self, monkeypatch):
         monkeypatch.setattr(
             provider_options,
diff --git a/tests/provider/test_thinking_params.py b/tests/provider/test_thinking_params.py
index db82214f1..d9c638214 100644
--- a/tests/provider/test_thinking_params.py
+++ b/tests/provider/test_thinking_params.py
@@ -47,6 +47,7 @@
 
 from flocks.provider import model_catalog
 from flocks.provider import options as provider_options
+from flocks.provider.interleaved import is_kimi_k27_code_model, is_kimi_k3_model
 
 DEEPSEEK_THINKING_EXTRA_BODY = {"thinking": {"type": "enabled"}}
 GLM_THINKING_EXTRA_BODY = {"thinking": {"type": "enabled", "clear_thinking": False}}
@@ -98,6 +99,10 @@ def _expected_generic_chat_extra_body(
         return GLM_THINKING_EXTRA_BODY
     if "mimo" in model_lower:
         return MIMO_THINKING_EXTRA_BODY
+    if is_kimi_k3_model(model_id):
+        return {"reasoning_effort": "max"}
+    if is_kimi_k27_code_model(model_id):
+        return KIMI_THINKING_EXTRA_BODY
     if "kimi" in model_lower:
         return KIMI_THINKING_EXTRA_BODY
     if "minimax" in model_lower or provider_lower == "minimax":
@@ -437,6 +442,9 @@ def test_anthropic_transport_still_uses_thinking_field(
             ("qwen3-7b-uncatalogued", {"enable_thinking": True}),
             ("glm-5-uncatalogued", GLM_THINKING_EXTRA_BODY),
             ("kimi-k2.6-uncatalogued", KIMI_THINKING_EXTRA_BODY),
+            ("kimi-k2.7-code", KIMI_THINKING_EXTRA_BODY),
+            ("kimi-k2.7-code-highspeed", KIMI_THINKING_EXTRA_BODY),
+            ("kimi-k3", {"reasoning_effort": "max"}),
             ("mimo-v2.5-pro-uncatalogued", MIMO_THINKING_EXTRA_BODY),
             ("minimax-m4-uncatalogued", {"reasoning_split": True}),
             ("step-3.5-flash-uncatalogued", {"enable_thinking": True}),
@@ -465,6 +473,16 @@ def test_series_token_fallback_emits_expected_extra_body(
             f"emitted {expected_extra_body}. options={options!r}"
         )
 
+    def test_kimi_k27_dispatch_does_not_match_unknown_suffix(self) -> None:
+        options = provider_options.build_provider_options(
+            "openai-compatible",
+            "kimi-k2.7-code-future",
+            reasoning_enabled=False,
+            resolve_max_tokens=False,
+        )
+
+        assert "extra_body" not in options
+
     @pytest.mark.parametrize(
         "provider_id,model_id,expected_extra_body",
         [
diff --git a/tests/sandbox/test_sandbox_runtime_integration.py b/tests/sandbox/test_sandbox_runtime_integration.py
index 734e0959c..7223845af 100644
--- a/tests/sandbox/test_sandbox_runtime_integration.py
+++ b/tests/sandbox/test_sandbox_runtime_integration.py
@@ -297,6 +297,13 @@ async def fake_resolve_sandbox_context(**_kwargs):
 
     assert len(captured_extras) == 2
     assert captured_extras[0][0] == "bash"
-    assert captured_extras[0][1] == {}
+    assert captured_extras[0][1] == {
+        "execution_mode": "build",
+        "workspace_dir": "/tmp",
+        "model": {"providerID": "test-provider", "modelID": "test-model"},
+        "plan_file_path": None,
+        "plan_relative_path": None,
+        "plan_permission_path": None,
+    }
     assert "sandbox" in captured_extras[1][1]
     assert captured_extras[1][1]["sandbox"]["container_name"] == "flocks-sbx-test"
diff --git a/tests/security/test_secret_resolution.py b/tests/security/test_secret_resolution.py
index c29d261e1..dd8d54f7a 100644
--- a/tests/security/test_secret_resolution.py
+++ b/tests/security/test_secret_resolution.py
@@ -25,6 +25,20 @@ def test_resolve_secret_value_prefers_real_fofa_split_secrets():
     assert resolve_secret_value("fofa_api_key", secrets) == "override-api-key"
 
 
+def test_resolve_secret_value_reads_legacy_llm_key_without_mutating_secrets():
+    secrets = MagicMock()
+    secrets.get.side_effect = lambda key: {
+        "custom-codex_api_key": "legacy-provider-key",
+    }.get(key)
+
+    assert (
+        resolve_secret_value("custom-codex_llm_key", secrets)
+        == "legacy-provider-key"
+    )
+    secrets.set.assert_not_called()
+    secrets.delete.assert_not_called()
+
+
 def test_resolve_secret_refs_returns_empty_for_invalid_fofa_canonical_secret():
     secrets = MagicMock()
     secrets.get.side_effect = lambda key: {
diff --git a/tests/server/routes/test_admin_users_routes.py b/tests/server/routes/test_admin_users_routes.py
index 1ed56409e..b813d7f12 100644
--- a/tests/server/routes/test_admin_users_routes.py
+++ b/tests/server/routes/test_admin_users_routes.py
@@ -96,3 +96,16 @@ async def test_admin_routes_delete_user_not_allowed(client: AsyncClient):
 
     response = await client.delete(f"/api/admin/users/{user_id}")
     assert response.status_code == 404, response.text
+
+
+@pytest.mark.asyncio
+async def test_service_owner_username_is_reserved():
+    from flocks.auth.context import API_TOKEN_SERVICE_USER_ID
+    from flocks.auth.service import AuthService
+
+    with pytest.raises(ValueError, match="系统保留"):
+        await AuthService._create_user_internal(
+            username=API_TOKEN_SERVICE_USER_ID,
+            password="Password123!",
+            role="member",
+        )
diff --git a/tests/server/routes/test_project_routes.py b/tests/server/routes/test_project_routes.py
index f147ebf13..f269b879b 100644
--- a/tests/server/routes/test_project_routes.py
+++ b/tests/server/routes/test_project_routes.py
@@ -1,7 +1,8 @@
+import asyncio
 from pathlib import Path
 
 import pytest
-from fastapi import status
+from fastapi import HTTPException, status
 from httpx import AsyncClient
 
 from flocks.auth.context import AuthUser
@@ -48,6 +49,12 @@ async def test_project_crud_uses_registry_without_project_database_rows(
     )
     assert session_response.status_code == status.HTTP_200_OK
     session_id = session_response.json()["id"]
+    other_session_response = await client.post(
+        "/api/session",
+        json={"title": "Keep archived", "projectID": project["id"]},
+    )
+    assert other_session_response.status_code == status.HTTP_200_OK
+    other_session_id = other_session_response.json()["id"]
 
     delete_response = await client.delete(f"/api/project/{project['id']}")
     assert delete_response.status_code == status.HTTP_200_OK
@@ -55,6 +62,25 @@ async def test_project_crud_uses_registry_without_project_database_rows(
     assert retained_file.read_text(encoding="utf-8") == "project files stay"
     assert (await client.get("/api/project")).json() == []
     assert (await client.get(f"/api/session/{session_id}")).status_code == status.HTTP_404_NOT_FOUND
+    archived = (
+        await client.get(
+            "/api/session",
+            params={"status": "archived", "view": "list", "manager": "true", "roots": "true"},
+        )
+    ).json()
+    archived_by_id = {item["id"]: item for item in archived}
+    assert archived_by_id[session_id]["projectName"] == "Security Labs"
+    assert other_session_id in archived_by_id
+    assert (await Session.get(project["id"], session_id)).status == "archived"
+
+    restore_response = await client.post(f"/api/session/{session_id}/restore")
+    assert restore_response.status_code == status.HTTP_200_OK
+    assert restore_response.json()["status"] == "active"
+    restored_projects = (await client.get("/api/project")).json()
+    assert [item["id"] for item in restored_projects] == [project["id"]]
+    assert restored_projects[0]["name"] == "Security Labs"
+    assert (await Session.get(project["id"], session_id)).status == "active"
+    assert (await Session.get(project["id"], other_session_id)).status == "archived"
 
 
 @pytest.mark.asyncio
@@ -74,6 +100,309 @@ async def test_create_project_creates_missing_directory(
     assert worktree.is_dir()
 
 
+@pytest.mark.asyncio
+async def test_move_regular_task_into_project(
+    client: AsyncClient,
+    tmp_path: Path,
+):
+    worktree = tmp_path / "move-target"
+    worktree.mkdir()
+    project = (
+        await client.post(
+            "/api/project",
+            json={"name": "Move Target", "worktree": str(worktree)},
+        )
+    ).json()
+    session = (
+        await client.post(
+            "/api/session",
+            json={"title": "Move me"},
+        )
+    ).json()
+
+    response = await client.patch(
+        f"/api/session/{session['id']}/project",
+        json={"projectID": project["id"]},
+    )
+
+    assert response.status_code == status.HTTP_200_OK
+    assert response.json()["projectID"] == project["id"]
+    assert response.json()["effectiveProjectID"] == project["id"]
+    assert response.json()["directory"] == str(worktree.resolve())
+    assert await Session.get(session["projectID"], session["id"]) is None
+    moved = await Session.get(project["id"], session["id"])
+    assert moved is not None and moved.directory == str(worktree.resolve())
+
+
+@pytest.mark.asyncio
+async def test_move_rejects_active_prompt_chain(
+    client: AsyncClient,
+    tmp_path: Path,
+):
+    from flocks.server.routes import session as session_routes
+
+    worktree = tmp_path / "busy-move-target"
+    worktree.mkdir()
+    project = (
+        await client.post(
+            "/api/project",
+            json={"name": "Busy Move Target", "worktree": str(worktree)},
+        )
+    ).json()
+    session = (await client.post("/api/session", json={"title": "Busy"})).json()
+
+    session_routes._set_prompt_chain_active(session["id"], True)
+    try:
+        response = await client.patch(
+            f"/api/session/{session['id']}/project",
+            json={"projectID": project["id"]},
+        )
+    finally:
+        session_routes._set_prompt_chain_active(session["id"], False)
+
+    assert response.status_code == status.HTTP_409_CONFLICT
+    assert await Session.get(session["projectID"], session["id"]) is not None
+
+
+@pytest.mark.asyncio
+async def test_move_blocks_replay_of_messages_from_previous_project(
+    client: AsyncClient,
+    tmp_path: Path,
+):
+    worktree = tmp_path / "history-move-target"
+    worktree.mkdir()
+    project = (
+        await client.post(
+            "/api/project",
+            json={"name": "History Move Target", "worktree": str(worktree)},
+        )
+    ).json()
+    session = (await client.post("/api/session", json={"title": "History"})).json()
+    assert (
+        await client.post(
+            f"/api/session/{session['id']}/message",
+            json={
+                "parts": [{"type": "text", "text": "Before moving"}],
+                "noReply": True,
+                "mockReply": "Old project reply",
+            },
+        )
+    ).status_code == status.HTTP_200_OK
+    messages = (await client.get(f"/api/session/{session['id']}/message")).json()
+    user_message = next(item for item in messages if item["info"]["role"] == "user")
+    text_part = next(part for part in user_message["parts"] if part["type"] == "text")
+    assert (
+        await client.patch(
+            f"/api/session/{session['id']}/project",
+            json={"projectID": project["id"]},
+        )
+    ).status_code == status.HTTP_200_OK
+
+    response = await client.post(
+        f"/api/session/{session['id']}/message/{user_message['info']['id']}/resend",
+        json={"text": "Do not replay", "partID": text_part["id"]},
+    )
+
+    assert response.status_code == status.HTTP_409_CONFLICT
+    assert "移动项目前" in response.text
+    revert_response = await client.post(
+        f"/api/session/{session['id']}/revert",
+        json={"messageID": user_message["info"]["id"]},
+    )
+    assert revert_response.status_code == status.HTTP_409_CONFLICT
+
+
+@pytest.mark.asyncio
+async def test_clear_resets_move_boundary_for_new_message_replay(
+    client: AsyncClient,
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+):
+    from flocks.server.routes import session as session_routes
+
+    worktree = tmp_path / "clear-move-target"
+    worktree.mkdir()
+    project = (
+        await client.post(
+            "/api/project",
+            json={"name": "Clear Move Target", "worktree": str(worktree)},
+        )
+    ).json()
+    session = (await client.post("/api/session", json={"title": "Clear History"})).json()
+    assert (
+        await client.post(
+            f"/api/session/{session['id']}/message",
+            json={"parts": [{"type": "text", "text": "Before moving"}], "noReply": True},
+        )
+    ).status_code == status.HTTP_200_OK
+    assert (
+        await client.patch(
+            f"/api/session/{session['id']}/project",
+            json={"projectID": project["id"]},
+        )
+    ).status_code == status.HTTP_200_OK
+    assert (
+        await client.post(f"/api/session/{session['id']}/clear")
+    ).status_code == status.HTTP_200_OK
+
+    moved = await Session.get(project["id"], session["id"])
+    assert moved is not None
+    assert "projectMove" not in moved.metadata
+
+    assert (
+        await client.post(
+            f"/api/session/{session['id']}/message",
+            json={"parts": [{"type": "text", "text": "After clearing"}], "noReply": True},
+        )
+    ).status_code == status.HTTP_200_OK
+    messages = (await client.get(f"/api/session/{session['id']}/message")).json()
+    user_message = next(item for item in messages if item["info"]["role"] == "user")
+    text_part = next(part for part in user_message["parts"] if part["type"] == "text")
+    scheduled_coroutines = []
+
+    def close_scheduled_coro(coro, **_kwargs):
+        scheduled_coroutines.append(coro)
+        coro.close()
+
+    monkeypatch.setattr(session_routes, "_schedule_background_coro", close_scheduled_coro)
+
+    replay_response = await client.post(
+        f"/api/session/{session['id']}/message/{user_message['info']['id']}/resend",
+        json={"text": "Replay after clearing", "partID": text_part["id"]},
+    )
+
+    assert replay_response.status_code == status.HTTP_202_ACCEPTED
+    assert len(scheduled_coroutines) == 1
+
+
+@pytest.mark.asyncio
+async def test_restore_archived_task_keeps_project_removed_when_directory_is_missing(
+    client: AsyncClient,
+    tmp_path: Path,
+):
+    worktree = tmp_path / "missing-after-remove"
+    worktree.mkdir()
+    project = (
+        await client.post(
+            "/api/project",
+            json={"name": "Missing Project", "worktree": str(worktree)},
+        )
+    ).json()
+    session = (
+        await client.post(
+            "/api/session",
+            json={"title": "Still Archived", "projectID": project["id"]},
+        )
+    ).json()
+    assert (await client.delete(f"/api/project/{project['id']}")).status_code == status.HTTP_200_OK
+    worktree.rmdir()
+
+    restore_response = await client.post(f"/api/session/{session['id']}/restore")
+
+    assert restore_response.status_code == status.HTTP_409_CONFLICT
+    assert "directory is unavailable" in restore_response.text
+    assert (await client.get("/api/project")).json() == []
+    stored = await Session.get(project["id"], session["id"])
+    assert stored is not None and stored.status == "archived"
+
+
+@pytest.mark.asyncio
+async def test_project_removal_waits_for_inflight_session_creation(
+    client: AsyncClient,
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+):
+    worktree = tmp_path / "concurrent-remove"
+    worktree.mkdir()
+    project = (
+        await client.post(
+            "/api/project",
+            json={"name": "Concurrent", "worktree": str(worktree)},
+        )
+    ).json()
+    create_started = asyncio.Event()
+    allow_create = asyncio.Event()
+    original_set = Storage.set
+
+    async def blocked_set(key, value, *args, **kwargs):
+        if key.startswith(f"session:{project['id']}:"):
+            create_started.set()
+            await allow_create.wait()
+        return await original_set(key, value, *args, **kwargs)
+
+    monkeypatch.setattr(Storage, "set", blocked_set)
+    create_task = asyncio.create_task(
+        client.post(
+            "/api/session",
+            json={"title": "Concurrent task", "projectID": project["id"]},
+        )
+    )
+    await create_started.wait()
+    delete_task = asyncio.create_task(client.delete(f"/api/project/{project['id']}"))
+
+    with pytest.raises(asyncio.TimeoutError):
+        await asyncio.wait_for(asyncio.shield(delete_task), timeout=0.05)
+
+    allow_create.set()
+    create_response, delete_response = await asyncio.gather(create_task, delete_task)
+
+    assert create_response.status_code == status.HTTP_200_OK
+    assert delete_response.status_code == status.HTTP_200_OK
+    created = await Session.get(project["id"], create_response.json()["id"])
+    assert created is not None and created.status == "archived"
+
+
+@pytest.mark.asyncio
+async def test_project_delete_restores_earlier_tasks_when_later_archive_fails(
+    client: AsyncClient,
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+):
+    from flocks.server.routes import session as session_routes
+
+    worktree = tmp_path / "archive-rollback"
+    worktree.mkdir()
+    project = (
+        await client.post(
+            "/api/project",
+            json={"name": "Archive rollback", "worktree": str(worktree)},
+        )
+    ).json()
+    session_ids = [
+        (
+            await client.post(
+                "/api/session",
+                json={"title": title, "projectID": project["id"]},
+            )
+        ).json()["id"]
+        for title in ("First", "Second")
+    ]
+    original_archive = session_routes.archive_session_for_user
+    calls = 0
+
+    async def fail_second_archive(session_id, user):
+        nonlocal calls
+        calls += 1
+        if calls == 2:
+            raise HTTPException(status_code=409, detail="archive failed")
+        return await original_archive(session_id, user)
+
+    monkeypatch.setattr(
+        session_routes,
+        "archive_session_for_user",
+        fail_second_archive,
+    )
+
+    response = await client.delete(f"/api/project/{project['id']}")
+
+    assert response.status_code == status.HTTP_409_CONFLICT
+    assert [item["id"] for item in (await client.get("/api/project")).json()] == [
+        project["id"]
+    ]
+    stored = [await Session.get(project["id"], session_id) for session_id in session_ids]
+    assert all(session is not None and session.status == "active" for session in stored)
+
+
 @pytest.mark.asyncio
 async def test_duplicate_project_directory_returns_existing_project(
     client: AsyncClient,
@@ -273,6 +602,13 @@ async def test_project_counts_and_session_lists_are_user_scoped(
         "Second Alice session",
     }
 
+    archive_response = await client.post(
+        f"/api/session/{alice_session.json()['id']}/archive",
+    )
+    assert archive_response.status_code == status.HTTP_200_OK
+    after_archive = (await client.get("/api/project")).json()
+    assert after_archive[0]["sessionCount"] == 1
+
 
 @pytest.mark.asyncio
 async def test_session_title_update_invalidates_project_search_stats(
diff --git a/tests/server/routes/test_provider_optional_api_key.py b/tests/server/routes/test_provider_optional_api_key.py
index 5aa342202..dd244acef 100644
--- a/tests/server/routes/test_provider_optional_api_key.py
+++ b/tests/server/routes/test_provider_optional_api_key.py
@@ -32,6 +32,8 @@ def patched_runtime(monkeypatch: pytest.MonkeyPatch):
     monkeypatch.setattr("flocks.security.get_secret_manager", lambda: fake_secrets)
     monkeypatch.setattr(provider_routes.Provider, "_ensure_initialized", MagicMock())
     monkeypatch.setattr(provider_routes.Provider, "get", lambda _pid: runtime_provider)
+    clear_config_cache = MagicMock()
+    monkeypatch.setattr(provider_routes.Config, "clear_cache", clear_config_cache)
 
     # Pretend the provider already exists in flocks.json so set_provider_credentials
     # follows the "update existing" path and never calls add_provider() with real
@@ -58,10 +60,30 @@ def patched_runtime(monkeypatch: pytest.MonkeyPatch):
         lambda _provider: {},
     )
 
-    return {"secrets": fake_secrets, "provider": runtime_provider}
+    return {
+        "secrets": fake_secrets,
+        "provider": runtime_provider,
+        "clear_config_cache": clear_config_cache,
+    }
 
 
 class TestOptionalApiKey:
+    @pytest.mark.asyncio
+    async def test_saving_credentials_invalidates_resolved_config_cache(
+        self, patched_runtime
+    ):
+        """The next request must not reapply a secret cached before this update."""
+        result = await provider_routes.set_provider_credentials(
+            "custom-local-gateway",
+            provider_routes.ProviderCredentialRequest(
+                api_key="new-secret",
+                base_url="http://127.0.0.1:8317/v1",
+            ),
+        )
+
+        assert result["success"] is True
+        patched_runtime["clear_config_cache"].assert_called_once_with()
+
     @pytest.mark.asyncio
     async def test_openai_compatible_accepts_empty_api_key(self, patched_runtime):
         """openai-compatible: empty api_key -> success, placeholder persisted."""
diff --git a/tests/server/routes/test_provider_route_responsiveness.py b/tests/server/routes/test_provider_route_responsiveness.py
index ba6967e37..ba1f4a09f 100644
--- a/tests/server/routes/test_provider_route_responsiveness.py
+++ b/tests/server/routes/test_provider_route_responsiveness.py
@@ -211,3 +211,37 @@ async def _config():
 
     assert heartbeat_ticks >= 3
     assert response.all == []
+
+
+async def test_list_providers_exposes_runtime_credential_state(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    from types import SimpleNamespace
+
+    from flocks.server.routes import provider as provider_routes
+
+    async def _initialized() -> None:
+        return None
+
+    async def _config():
+        raise RuntimeError("no merged config needed for focused route test")
+
+    runtime_provider = SimpleNamespace(
+        name="Missing Credentials",
+        is_configured=lambda: False,
+    )
+    monkeypatch.setattr(provider_routes, "_ensure_provider_initialized", _initialized)
+    monkeypatch.setattr(provider_routes.Config, "get", _config)
+    monkeypatch.setattr(
+        provider_routes.ConfigWriter,
+        "list_provider_ids",
+        lambda: ["missing-credentials"],
+    )
+    monkeypatch.setattr(provider_routes.Provider, "get", lambda _provider_id: runtime_provider)
+    monkeypatch.setattr(provider_routes.Provider, "list_models", lambda _provider_id: [])
+
+    response = await provider_routes.list_providers()
+
+    assert len(response.all) == 1
+    assert response.all[0].configured is False
+    assert response.connected == ["missing-credentials"]
diff --git a/tests/server/routes/test_session_routes.py b/tests/server/routes/test_session_routes.py
index 05d0470f1..e20a70d07 100644
--- a/tests/server/routes/test_session_routes.py
+++ b/tests/server/routes/test_session_routes.py
@@ -14,12 +14,12 @@
 
 import asyncio
 from types import SimpleNamespace
-from unittest.mock import AsyncMock
+from unittest.mock import AsyncMock, MagicMock
 
 import pytest
 from fastapi import HTTPException, status
 from httpx import AsyncClient
-from flocks.auth.context import AuthUser
+from flocks.auth.context import API_TOKEN_SERVICE_USER_ID, AuthUser
 from flocks.session.core.status import SessionStatus, SessionStatusBusy
 from flocks.session.message import (
     Message,
@@ -31,6 +31,16 @@
 from flocks.session.orphan_tools import INTERRUPTED_TOOL_ERROR
 from flocks.session.session import Session
 
+
+def _use_webui_admin(monkeypatch: pytest.MonkeyPatch) -> AuthUser:
+    """Authenticate a route test as a browser user instead of the API token."""
+    from flocks.server.routes import session as session_routes
+
+    user = AuthUser(id="usr_admin", username="admin", role="admin", status="active")
+    monkeypatch.setattr(session_routes, "require_user", lambda _request: user)
+    return user
+
+
 # ===========================================================================
 # CRUD
 # ===========================================================================
@@ -114,15 +124,110 @@ async def test_create_session_with_category(self, client: AsyncClient):
         assert resp.json()["category"] == "workflow"
 
     @pytest.mark.asyncio
-    async def test_create_session_with_api_token_is_ownerless(self, client: AsyncClient):
-        """Sessions created by API-token clients remain manageable by WebUI admins."""
+    @pytest.mark.parametrize("category", [None, "entity-config", "workflow"])
+    async def test_create_session_with_manual_auto_mode(
+        self,
+        client: AsyncClient,
+        monkeypatch: pytest.MonkeyPatch,
+        category: str | None,
+    ):
+        """Auto is persisted for supported WebUI conversation sessions."""
+        from flocks.session.session_loop import SessionLoop
+
+        _use_webui_admin(monkeypatch)
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_auto_configuration",
+            AsyncMock(return_value=(True, "available")),
+        )
+        payload = {"title": "Auto Session", "model_auto": True}
+        if category is not None:
+            payload["category"] = category
+        resp = await client.post("/api/session", json=payload)
+
+        assert resp.status_code == status.HTTP_200_OK
+        assert resp.json()["model_auto"] is True
+        assert resp.json()["model_pinned"] is False
+        assert resp.json()["category"] == (category or "user")
+
+    @pytest.mark.asyncio
+    async def test_create_unsupported_session_rejects_auto(
+        self,
+        client: AsyncClient,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        from flocks.session.session_loop import SessionLoop
+
+        validate_auto = AsyncMock(return_value=(True, "available"))
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_auto_configuration",
+            validate_auto,
+        )
+
+        resp = await client.post(
+            "/api/session",
+            json={"category": "task", "model_auto": True},
+        )
+
+        assert resp.status_code == status.HTTP_422_UNPROCESSABLE_ENTITY
+        assert (
+            "only available for user, entity configuration, and workflow sessions"
+            in str(resp.json())
+        )
+        validate_auto.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_create_session_rejects_unavailable_auto(
+        self,
+        client: AsyncClient,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        from flocks.session.session_loop import SessionLoop
+
+        _use_webui_admin(monkeypatch)
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_auto_configuration",
+            AsyncMock(return_value=(False, "fallback_unavailable")),
+        )
+
+        resp = await client.post("/api/session", json={"model_auto": True})
+
+        assert resp.status_code == status.HTTP_400_BAD_REQUEST
+        assert "fallback_unavailable" in str(resp.json())
+
+    @pytest.mark.asyncio
+    async def test_api_token_cannot_create_auto_session(
+        self,
+        client: AsyncClient,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        from flocks.session.session_loop import SessionLoop
+
+        validate_auto = AsyncMock(return_value=(True, "available"))
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_auto_configuration",
+            validate_auto,
+        )
+
+        resp = await client.post("/api/session", json={"model_auto": True})
+
+        assert resp.status_code == status.HTTP_403_FORBIDDEN
+        assert "only be enabled from the WebUI" in str(resp.json())
+        validate_auto.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_create_session_with_api_token_is_system_owned(self, client: AsyncClient):
+        """API-token sessions have an explicit owner manageable by WebUI admins."""
         resp = await client.post("/api/session", json={"title": "TUI Session"})
 
         assert resp.status_code == status.HTTP_200_OK
         session = await Session.get_by_id(resp.json()["id"])
         assert session is not None
-        assert session.owner_user_id is None
-        assert session.owner_username is None
+        assert session.owner_user_id == API_TOKEN_SERVICE_USER_ID
+        assert session.owner_username == API_TOKEN_SERVICE_USER_ID
 
     @pytest.mark.asyncio
     async def test_create_session_with_local_user_keeps_owner(
@@ -176,6 +281,16 @@ async def test_webui_admin_can_manage_api_token_session(
         )
         assert message_resp.status_code == status.HTTP_200_OK
 
+        child_resp = await client.post(
+            "/api/session",
+            json={"title": "WebUI child", "parentID": session_id},
+        )
+        assert child_resp.status_code == status.HTTP_200_OK
+        child = await Session.get_by_id(child_resp.json()["id"])
+        assert child is not None
+        assert child.owner_user_id == API_TOKEN_SERVICE_USER_ID
+        assert child.owner_username == API_TOKEN_SERVICE_USER_ID
+
         rename_resp = await client.patch(
             f"/api/session/{session_id}",
             json={"title": "Renamed in WebUI"},
@@ -284,24 +399,255 @@ async def test_list_sessions_light_manager_filters_and_omits_heavy_fields(self,
         assert set(row) == {
             "id",
             "projectID",
+            "projectName",
             "effectiveProjectID",
             "directory",
             "title",
             "time",
             "category",
+            "status",
             "parentID",
             "provider",
             "model",
             "model_pinned",
+            "model_auto",
+            "ownerUserID",
+            "ownerUsername",
             "canWrite",
             "canDelete",
             "isShared",
         }
         assert row["projectID"]
         assert row["directory"]
+        assert "ownerUsername" in row
         assert "goal" not in row
         assert "summary" not in row
 
+    @pytest.mark.asyncio
+    async def test_archive_hides_session_preserves_history_and_restores_tree(self, client: AsyncClient):
+        parent_resp = await client.post("/api/session", json={"title": "Archive Parent"})
+        parent_id = parent_resp.json()["id"]
+        project_id = parent_resp.json()["projectID"]
+        child_resp = await client.post(
+            "/api/session",
+            json={"title": "Archive Child", "parentID": parent_id},
+        )
+        child_id = child_resp.json()["id"]
+        await client.post(
+            f"/api/session/{parent_id}/message",
+            json={"parts": [{"type": "text", "text": "keep this history"}], "noReply": True},
+        )
+
+        archive_resp = await client.post(f"/api/session/{parent_id}/archive")
+
+        assert archive_resp.status_code == status.HTTP_200_OK
+        assert archive_resp.json()["status"] == "archived"
+        archived_parent = await Session.get(project_id, parent_id)
+        archived_child = await Session.get(project_id, child_id)
+        assert archived_parent is not None and archived_parent.status == "archived"
+        assert archived_child is not None and archived_child.status == "archived"
+        assert len(await Message.list(parent_id)) == 1
+
+        active_list = await client.get("/api/session", params={"status": "active"})
+        archived_list = await client.get(
+            "/api/session",
+            params={"status": "archived", "view": "list", "manager": "true", "roots": "true"},
+        )
+        assert parent_id not in {item["id"] for item in active_list.json()}
+        assert parent_id in {item["id"] for item in archived_list.json()}
+        assert child_id not in {item["id"] for item in archived_list.json()}
+        archived_row = next(item for item in archived_list.json() if item["id"] == parent_id)
+        assert archived_row["canWrite"] is False
+
+        get_resp = await client.get(f"/api/session/{parent_id}")
+        update_resp = await client.patch(f"/api/session/{parent_id}", json={"title": "Blocked"})
+        message_resp = await client.post(
+            f"/api/session/{parent_id}/message",
+            json={"parts": [{"type": "text", "text": "blocked"}], "noReply": True},
+        )
+        assert get_resp.status_code == status.HTTP_404_NOT_FOUND
+        assert update_resp.status_code == status.HTTP_409_CONFLICT
+        assert message_resp.status_code == status.HTTP_409_CONFLICT
+
+        child_restore_resp = await client.post(f"/api/session/{child_id}/restore")
+        late_child_resp = await client.post(
+            "/api/session",
+            json={"title": "Late Child", "parentID": parent_id},
+        )
+        assert child_restore_resp.status_code == status.HTTP_409_CONFLICT
+        assert late_child_resp.status_code == status.HTTP_409_CONFLICT
+
+        restore_resp = await client.post(f"/api/session/{parent_id}/restore")
+        assert restore_resp.status_code == status.HTTP_200_OK
+        assert restore_resp.json()["status"] == "active"
+        restored_child = await Session.get(project_id, child_id)
+        assert restored_child is not None and restored_child.status == "active"
+        assert len(await Message.list(parent_id)) == 1
+
+    @pytest.mark.asyncio
+    async def test_archive_cancels_route_background_work(
+        self,
+        client: AsyncClient,
+        session_id: str,
+    ):
+        from flocks.server.routes import session as session_routes
+
+        started = asyncio.Event()
+        stopped = asyncio.Event()
+
+        async def background_work() -> None:
+            started.set()
+            try:
+                await asyncio.Event().wait()
+            finally:
+                stopped.set()
+
+        session_routes._schedule_background_coro(
+            background_work(),
+            session_id=session_id,
+            action="test.background",
+        )
+        await started.wait()
+
+        response = await client.post(f"/api/session/{session_id}/archive")
+
+        assert response.status_code == status.HTTP_200_OK
+        assert stopped.is_set()
+
+    @pytest.mark.asyncio
+    async def test_list_status_filter_applies_before_pagination(self, client: AsyncClient):
+        archived_resp = await client.post("/api/session", json={"title": "Archived First"})
+        active_resp = await client.post("/api/session", json={"title": "Active Second"})
+        assert (await client.post(f"/api/session/{archived_resp.json()['id']}/archive")).status_code == 200
+
+        response = await client.get(
+            "/api/session",
+            params={"status": "active", "limit": "1", "offset": "0"},
+        )
+
+        assert response.status_code == status.HTTP_200_OK
+        assert [item["id"] for item in response.json()] == [active_resp.json()["id"]]
+
+    @pytest.mark.asyncio
+    async def test_archived_manager_admin_sees_all_and_member_only_owns(
+        self,
+        client: AsyncClient,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        from flocks.server.routes import session as session_routes
+
+        alice = AuthUser(id="usr_alice", username="alice", role="member", status="active")
+        bob = AuthUser(id="usr_bob", username="bob", role="member", status="active")
+        admin = AuthUser(id="usr_admin", username="admin", role="admin", status="active")
+        alice_session = await Session.create(
+            project_id="archive_permissions",
+            directory="/tmp",
+            title="Alice archive",
+            owner_user_id=alice.id,
+            owner_username=alice.username,
+        )
+        bob_session = await Session.create(
+            project_id="archive_permissions",
+            directory="/tmp",
+            title="Bob shared archive",
+            owner_user_id=bob.id,
+            owner_username=bob.username,
+            metadata={"shared_local": True},
+        )
+        assert await Session.archive(alice_session.project_id, alice_session.id) is True
+        assert await Session.archive(bob_session.project_id, bob_session.id) is True
+
+        params = {
+            "status": "archived",
+            "view": "list",
+            "manager": "true",
+            "roots": "true",
+        }
+        monkeypatch.setattr(session_routes, "require_user", lambda _request: admin)
+        admin_response = await client.get("/api/session", params=params)
+
+        assert admin_response.status_code == status.HTTP_200_OK
+        admin_rows = {row["id"]: row for row in admin_response.json()}
+        assert {alice_session.id, bob_session.id}.issubset(admin_rows)
+        assert admin_rows[alice_session.id]["ownerUsername"] == "alice"
+        assert admin_rows[bob_session.id]["ownerUsername"] == "bob"
+
+        monkeypatch.setattr(session_routes, "require_user", lambda _request: alice)
+        member_response = await client.get("/api/session", params=params)
+
+        assert member_response.status_code == status.HTTP_200_OK
+        assert [row["id"] for row in member_response.json()] == [alice_session.id]
+
+        ordinary_archived_response = await client.get(
+            "/api/session",
+            params={"status": "archived", "view": "list", "roots": "true"},
+        )
+        assert ordinary_archived_response.status_code == status.HTTP_200_OK
+        assert [row["id"] for row in ordinary_archived_response.json()] == [
+            alice_session.id
+        ]
+
+        all_status_response = await client.get(
+            "/api/session",
+            params={"status": "all", "view": "list", "roots": "true"},
+        )
+        assert bob_session.id not in {
+            row["id"] for row in all_status_response.json()
+        }
+
+    @pytest.mark.asyncio
+    async def test_archive_rejects_tree_with_descendant_owned_by_another_user(
+        self,
+        client: AsyncClient,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        from flocks.auth.context import reset_current_auth_user, set_current_auth_user
+        from flocks.server.routes import session as session_routes
+        from flocks.session.session import SessionInfo
+        from flocks.storage.storage import Storage
+
+        alice = AuthUser(id="usr_alice", username="alice", role="member", status="active")
+        parent = await Session.create(
+            project_id="archive_mixed_owner_tree",
+            directory="/tmp",
+            title="Alice parent",
+            owner_user_id=alice.id,
+            owner_username=alice.username,
+        )
+        child = SessionInfo(
+            projectID=parent.project_id,
+            directory=parent.directory,
+            title="Bob child",
+            parentID=parent.id,
+            ownerUserID="usr_bob",
+            ownerUsername="bob",
+        )
+        await Storage.set(
+            f"session:{child.project_id}:{child.id}",
+            child,
+            "session",
+        )
+        Session.invalidate_cache()
+        monkeypatch.setattr(session_routes, "require_user", lambda _request: alice)
+
+        token = set_current_auth_user(alice)
+        try:
+            response = await client.post(f"/api/session/{parent.id}/archive")
+        finally:
+            reset_current_auth_user(token)
+
+        assert response.status_code == status.HTTP_403_FORBIDDEN
+        stored_parent = await Storage.get(
+            f"session:{parent.project_id}:{parent.id}",
+            SessionInfo,
+        )
+        stored_child = await Storage.get(
+            f"session:{child.project_id}:{child.id}",
+            SessionInfo,
+        )
+        assert stored_parent is not None and stored_parent.status == "active"
+        assert stored_child is not None and stored_child.status == "active"
+
     @pytest.mark.asyncio
     async def test_create_session_in_user_managed_project(
         self,
@@ -333,6 +679,7 @@ async def test_create_session_in_user_managed_project(
         )
         row = next(item for item in list_resp.json() if item["id"] == session_resp.json()["id"])
         assert row["projectID"] == project["id"]
+        assert row["projectName"] == "Labs"
         assert row["effectiveProjectID"] == project["id"]
         assert row["directory"] == project["worktree"]
 
@@ -368,12 +715,12 @@ async def test_legacy_session_is_grouped_under_tasks_without_rewrite(
         assert stored.project_id == "legacy-git-project"
 
     @pytest.mark.asyncio
-    async def test_deleted_project_sessions_are_removed(
+    async def test_deleted_project_sessions_are_archived_and_restorable(
         self,
         client: AsyncClient,
         tmp_path,
     ):
-        """Deleting a project removes its sessions but preserves its directory."""
+        """Deleting a project archives its sessions and restoring one revives the project."""
         worktree = tmp_path / "removable-project"
         worktree.mkdir()
         project_response = await client.post(
@@ -399,7 +746,65 @@ async def test_deleted_project_sessions_are_removed(
             params={"view": "list", "manager": "true", "projectID": "tasks"},
         )
         assert all(item["id"] != session_id for item in tasks_response.json())
-        assert await Session.get(project["id"], session_id) is None
+        archived = await Session.get(project["id"], session_id)
+        assert archived is not None and archived.status == "archived"
+
+        restore_response = await client.post(f"/api/session/{session_id}/restore")
+        assert restore_response.status_code == status.HTTP_200_OK
+        assert restore_response.json()["status"] == "active"
+        assert [item["id"] for item in (await client.get("/api/project")).json()] == [project["id"]]
+
+    @pytest.mark.asyncio
+    async def test_restore_rolls_project_back_when_session_restore_fails(
+        self,
+        client: AsyncClient,
+        tmp_path,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        worktree = tmp_path / "restore-rollback"
+        worktree.mkdir()
+        project = (
+            await client.post(
+                "/api/project",
+                json={"name": "Rollback", "worktree": str(worktree)},
+            )
+        ).json()
+        session_id = (
+            await client.post(
+                "/api/session",
+                json={"title": "Rollback Me", "projectID": project["id"]},
+            )
+        ).json()["id"]
+        assert (await client.delete(f"/api/project/{project['id']}")).status_code == 200
+
+        async def fail_unarchive(_project_id: str, _session_id: str) -> bool:
+            return False
+
+        monkeypatch.setattr(Session, "unarchive", fail_unarchive)
+        response = await client.post(f"/api/session/{session_id}/restore")
+
+        assert response.status_code == status.HTTP_409_CONFLICT
+        assert (await client.get("/api/project")).json() == []
+        archived = await Session.get(project["id"], session_id)
+        assert archived is not None and archived.status == "archived"
+
+    @pytest.mark.asyncio
+    async def test_restore_rejects_missing_managed_project_metadata(
+        self,
+        client: AsyncClient,
+    ):
+        session = await Session.create(
+            project_id="prj_missing_restore_metadata",
+            directory="/tmp",
+            title="Missing project metadata",
+        )
+        assert await Session.archive(session.project_id, session.id) is True
+
+        response = await client.post(f"/api/session/{session.id}/restore")
+
+        assert response.status_code == status.HTTP_409_CONFLICT
+        archived = await Session.get(session.project_id, session.id)
+        assert archived is not None and archived.status == "archived"
 
     @pytest.mark.asyncio
     async def test_get_session(self, client: AsyncClient, session_id: str):
@@ -554,193 +959,213 @@ async def test_update_session_title(self, client: AsyncClient, session_id: str):
         assert resp.json()["title"] == "Updated Title"
 
     @pytest.mark.asyncio
-    async def test_update_session_not_found(self, client: AsyncClient):
-        """PATCH for unknown session returns 404."""
-        resp = await client.patch(
-            "/api/session/ses_nonexistent00000000000000",
-            json={"title": "X"},
-        )
-        assert resp.status_code == status.HTTP_404_NOT_FOUND
-
-    @pytest.mark.asyncio
-    async def test_delete_session(self, client: AsyncClient, session_id: str):
-        """DELETE /api/session/{id} removes the session."""
-        resp = await client.delete(f"/api/session/{session_id}")
-        assert resp.status_code == status.HTTP_200_OK
-        assert resp.json() is True
-
-        # Confirm it is gone
-        get_resp = await client.get(f"/api/session/{session_id}")
-        assert get_resp.status_code == status.HTTP_404_NOT_FOUND
-
-    @staticmethod
-    def _patch_delete_session_dependencies(
-        monkeypatch,
-        session_routes,
-        *,
+    async def test_update_session_auto_and_concrete_model_are_exclusive(
+        self,
+        client: AsyncClient,
         session_id: str,
-        order: list[str],
-        session_list,
-    ) -> None:
-        async def fake_abort_session_processing(abort_session_id: str) -> bool:
-            order.append(f"abort:{abort_session_id}")
-            return True
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        from flocks.session.session_loop import SessionLoop
 
-        async def fake_wait_for_sessions_idle(session_ids: list[str], timeout_s: float = 5.0) -> None:
-            order.append(f"wait:{','.join(session_ids)}")
+        _use_webui_admin(monkeypatch)
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_auto_configuration",
+            AsyncMock(return_value=(True, "available")),
+        )
+        auto_resp = await client.patch(
+            f"/api/session/{session_id}",
+            json={"model_auto": True},
+        )
+        assert auto_resp.status_code == status.HTTP_200_OK
+        assert auto_resp.json()["model_auto"] is True
+        assert auto_resp.json()["model_pinned"] is False
 
-        async def fake_interaction_queue_clear(_session_id: str) -> None:
-            order.append("queue_clear")
+        invalid_resp = await client.patch(
+            f"/api/session/{session_id}",
+            json={
+                "model_auto": True,
+                "provider": "openai",
+                "model": "gpt-4o",
+                "model_pinned": True,
+            },
+        )
+        assert invalid_resp.status_code == status.HTTP_422_UNPROCESSABLE_ENTITY
 
-        async def fake_goal_clear(_session_id: str) -> None:
-            order.append("goal_clear")
+        concrete_resp = await client.patch(
+            f"/api/session/{session_id}",
+            json={"provider": "openai", "model": "gpt-4o"},
+        )
+        assert concrete_resp.status_code == status.HTTP_200_OK
+        assert concrete_resp.json()["model_auto"] is False
+        assert concrete_resp.json()["model_pinned"] is True
 
-        async def fake_session_delete(_project_id: str, delete_session_id: str) -> bool:
-            assert delete_session_id == session_id
-            order.append("delete")
-            return True
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("category", ["entity-config", "workflow"])
+    async def test_update_supported_sidebar_session_allows_auto(
+        self,
+        client: AsyncClient,
+        monkeypatch: pytest.MonkeyPatch,
+        category: str,
+    ):
+        from flocks.session.session_loop import SessionLoop
 
-        monkeypatch.setattr(session_routes.Session, "list", session_list)
-        monkeypatch.setattr(
-            session_routes,
-            "_abort_session_processing",
-            fake_abort_session_processing,
-        )
-        monkeypatch.setattr(
-            session_routes,
-            "_wait_for_sessions_idle",
-            fake_wait_for_sessions_idle,
+        create_resp = await client.post(
+            "/api/session",
+            json={"title": "Sidebar Chat", "category": category},
         )
+        assert create_resp.status_code == status.HTTP_200_OK
+
+        _use_webui_admin(monkeypatch)
         monkeypatch.setattr(
-            "flocks.session.interaction_queue.InteractionQueue.clear",
-            fake_interaction_queue_clear,
+            SessionLoop,
+            "validate_auto_configuration",
+            AsyncMock(return_value=(True, "available")),
         )
-        monkeypatch.setattr(
-            "flocks.session.goal.GoalManager.clear",
-            fake_goal_clear,
+        resp = await client.patch(
+            f"/api/session/{create_resp.json()['id']}",
+            json={"model_auto": True},
         )
-        monkeypatch.setattr(session_routes.Session, "delete", fake_session_delete)
+
+        assert resp.status_code == status.HTTP_200_OK
+        assert resp.json()["category"] == category
+        assert resp.json()["model_auto"] is True
+        assert resp.json()["model_pinned"] is False
 
     @pytest.mark.asyncio
-    async def test_delete_session_aborts_and_waits_before_delete(
+    async def test_update_session_rejects_unavailable_auto(
         self,
         client: AsyncClient,
         session_id: str,
-        monkeypatch,
+        monkeypatch: pytest.MonkeyPatch,
     ):
-        """DELETE waits for active processing to stop before clearing messages."""
-        from flocks.server.routes import session as session_routes
+        from flocks.session.session_loop import SessionLoop
 
-        order: list[str] = []
-
-        async def fake_session_list(_project_id: str):
-            return []
-
-        self._patch_delete_session_dependencies(
-            monkeypatch,
-            session_routes,
-            session_id=session_id,
-            order=order,
-            session_list=fake_session_list,
+        _use_webui_admin(monkeypatch)
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_auto_configuration",
+            AsyncMock(return_value=(False, "primary_provider_not_configured")),
         )
 
-        resp = await client.delete(f"/api/session/{session_id}")
+        resp = await client.patch(
+            f"/api/session/{session_id}",
+            json={"model_auto": True},
+        )
 
-        assert resp.status_code == status.HTTP_200_OK
-        assert resp.json() is True
-        assert order == [
-            f"abort:{session_id}",
-            "queue_clear",
-            "goal_clear",
-            f"wait:{session_id}",
-            "delete",
-        ]
+        assert resp.status_code == status.HTTP_400_BAD_REQUEST
+        assert "primary_provider_not_configured" in str(resp.json())
 
     @pytest.mark.asyncio
-    async def test_delete_session_waits_for_descendant_loops_before_delete(
+    async def test_pinning_existing_auto_session_disables_auto(
         self,
         client: AsyncClient,
         session_id: str,
-        monkeypatch,
+        monkeypatch: pytest.MonkeyPatch,
     ):
-        """DELETE waits for child and grandchild loops before recursive delete."""
-        from flocks.server.routes import session as session_routes
+        from flocks.session.session_loop import SessionLoop
 
-        child_id = "ses_delete_child_wait"
-        grandchild_id = "ses_delete_grandchild_wait"
-        order: list[str] = []
+        _use_webui_admin(monkeypatch)
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_auto_configuration",
+            AsyncMock(return_value=(True, "available")),
+        )
+        clear_state = MagicMock()
+        monkeypatch.setattr(SessionLoop, "clear_auto_failover_state", clear_state)
+        auto_resp = await client.patch(
+            f"/api/session/{session_id}",
+            json={"model_auto": True},
+        )
+        assert auto_resp.status_code == status.HTTP_200_OK
 
-        async def fake_session_list(_project_id: str):
-            return [
-                SimpleNamespace(id=child_id, parent_id=session_id),
-                SimpleNamespace(id=grandchild_id, parent_id=child_id),
-            ]
+        pin_resp = await client.patch(
+            f"/api/session/{session_id}",
+            json={"model_pinned": True},
+        )
 
-        self._patch_delete_session_dependencies(
-            monkeypatch,
-            session_routes,
-            session_id=session_id,
-            order=order,
-            session_list=fake_session_list,
+        assert pin_resp.status_code == status.HTTP_200_OK
+        assert pin_resp.json()["model_pinned"] is True
+        assert pin_resp.json()["model_auto"] is False
+        clear_state.assert_called_once_with(session_id)
+
+    @pytest.mark.asyncio
+    async def test_api_token_cannot_enable_auto_on_existing_session(
+        self,
+        client: AsyncClient,
+        session_id: str,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        from flocks.session.session_loop import SessionLoop
+
+        validate_auto = AsyncMock(return_value=(True, "available"))
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_auto_configuration",
+            validate_auto,
         )
 
-        resp = await client.delete(f"/api/session/{session_id}")
+        resp = await client.patch(
+            f"/api/session/{session_id}",
+            json={"model_auto": True},
+        )
 
-        assert resp.status_code == status.HTTP_200_OK
-        assert resp.json() is True
-        assert order == [
-            f"abort:{session_id}",
-            "queue_clear",
-            "goal_clear",
-            f"wait:{session_id}",
-            f"abort:{child_id}",
-            f"abort:{grandchild_id}",
-            f"wait:{child_id},{grandchild_id}",
-            "delete",
-        ]
+        assert resp.status_code == status.HTTP_403_FORBIDDEN
+        assert "only be enabled from the WebUI" in str(resp.json())
+        validate_auto.assert_not_awaited()
 
     @pytest.mark.asyncio
-    async def test_delete_session_aborts_descendant_that_appears_after_parent_wait(
+    async def test_update_unsupported_session_rejects_auto(
         self,
         client: AsyncClient,
-        session_id: str,
-        monkeypatch,
+        monkeypatch: pytest.MonkeyPatch,
     ):
-        """DELETE re-collects descendants after parent abort to catch late children."""
-        from flocks.server.routes import session as session_routes
+        from flocks.session.session_loop import SessionLoop
 
-        child_id = "ses_delete_late_child_wait"
-        list_calls = 0
-        order: list[str] = []
+        create_resp = await client.post(
+            "/api/session",
+            json={"title": "Non-user Session", "category": "task"},
+        )
+        assert create_resp.status_code == status.HTTP_200_OK
+        session_id = create_resp.json()["id"]
+        validate_auto = AsyncMock(return_value=(True, "available"))
+        monkeypatch.setattr(
+            SessionLoop,
+            "validate_auto_configuration",
+            validate_auto,
+        )
 
-        async def fake_session_list(_project_id: str):
-            nonlocal list_calls
-            list_calls += 1
-            if list_calls == 1:
-                return []
-            return [SimpleNamespace(id=child_id, parent_id=session_id)]
+        resp = await client.patch(
+            f"/api/session/{session_id}",
+            json={"model_auto": True},
+        )
 
-        self._patch_delete_session_dependencies(
-            monkeypatch,
-            session_routes,
-            session_id=session_id,
-            order=order,
-            session_list=fake_session_list,
+        assert resp.status_code == status.HTTP_422_UNPROCESSABLE_ENTITY
+        assert (
+            "only available for user, entity configuration, and workflow sessions"
+            in str(resp.json())
         )
+        validate_auto.assert_not_awaited()
 
-        resp = await client.delete(f"/api/session/{session_id}")
+    @pytest.mark.asyncio
+    async def test_update_session_not_found(self, client: AsyncClient):
+        """PATCH for unknown session returns 404."""
+        resp = await client.patch(
+            "/api/session/ses_nonexistent00000000000000",
+            json={"title": "X"},
+        )
+        assert resp.status_code == status.HTTP_404_NOT_FOUND
 
+    @pytest.mark.asyncio
+    async def test_delete_session(self, client: AsyncClient, session_id: str):
+        """DELETE /api/session/{id} removes the session."""
+        resp = await client.delete(f"/api/session/{session_id}")
         assert resp.status_code == status.HTTP_200_OK
         assert resp.json() is True
-        assert order == [
-            f"abort:{session_id}",
-            "queue_clear",
-            "goal_clear",
-            f"wait:{session_id}",
-            f"abort:{child_id}",
-            f"wait:{child_id}",
-            "delete",
-        ]
+
+        # Confirm it is gone
+        get_resp = await client.get(f"/api/session/{session_id}")
+        assert get_resp.status_code == status.HTTP_404_NOT_FOUND
 
     @pytest.mark.asyncio
     async def test_delete_session_not_found(self, client: AsyncClient):
@@ -765,6 +1190,24 @@ async def test_list_messages_empty(self, client: AsyncClient, session_id: str):
         assert isinstance(data, list)
         assert len(data) == 0
 
+    @pytest.mark.asyncio
+    async def test_list_messages_reports_storage_failure(
+        self,
+        client: AsyncClient,
+        session_id: str,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        """A read failure must not look like a successfully empty history."""
+        monkeypatch.setattr(
+            Message,
+            "list_with_parts",
+            AsyncMock(side_effect=RuntimeError("storage unavailable")),
+        )
+
+        resp = await client.get(f"/api/session/{session_id}/message")
+
+        assert resp.status_code == status.HTTP_500_INTERNAL_SERVER_ERROR
+
     @pytest.mark.asyncio
     async def test_send_message_noReply(self, client: AsyncClient, session_id: str):
         """POST /api/session/{id}/message with noReply=True stores without triggering LLM."""
@@ -1057,6 +1500,47 @@ async def test_message_to_unknown_session_returns_404(self, client: AsyncClient)
         )
         assert resp.status_code == status.HTTP_404_NOT_FOUND
 
+    @pytest.mark.asyncio
+    async def test_message_is_not_persisted_if_session_archives_after_preflight(
+        self,
+        client: AsyncClient,
+        session_id: str,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        from flocks.provider.provider import Provider
+        from flocks.server.routes import session as session_routes
+
+        apply_config_entered = asyncio.Event()
+        release_apply_config = asyncio.Event()
+        original_apply_config = Provider.apply_config
+
+        async def blocked_apply_config(*args, **kwargs):
+            await original_apply_config(*args, **kwargs)
+            apply_config_entered.set()
+            await release_apply_config.wait()
+
+        monkeypatch.setattr(Provider, "apply_config", blocked_apply_config)
+
+        message_task = asyncio.create_task(
+            client.post(
+                f"/api/session/{session_id}/message",
+                json={
+                    "parts": [{"type": "text", "text": "must not persist"}],
+                    "noReply": True,
+                },
+            )
+        )
+        await apply_config_entered.wait()
+        archive_response = await client.post(f"/api/session/{session_id}/archive")
+        restore_response = await client.post(f"/api/session/{session_id}/restore")
+        release_apply_config.set()
+        message_response = await message_task
+
+        assert archive_response.status_code == status.HTTP_200_OK
+        assert restore_response.status_code == status.HTTP_200_OK
+        assert message_response.status_code == status.HTTP_409_CONFLICT
+        assert await Message.list(session_id) == []
+
     @pytest.mark.asyncio
     async def test_resend_user_message_updates_text_and_truncates_followups(
         self,
@@ -1444,8 +1928,191 @@ async def test_prepare_replay_runtime_uses_current_model_resolution(
             "agent_name": "rex",
             "provider_id": "openai",
             "model_id": "gpt-4.1",
+            "auto_failover": False,
         }
 
+    @pytest.mark.asyncio
+    async def test_auto_replay_reports_actual_fallback_model(
+        self,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        """Replay passes Auto authorization and publishes the recovered model."""
+        from flocks.server.routes import session as session_routes
+        from flocks.session.session_loop import LoopResult, SessionLoop
+
+        user_message = SimpleNamespace(id="msg_user", agent="rex")
+        session = SimpleNamespace(id="ses_auto", directory="/tmp/project")
+        assistant = SimpleNamespace(
+            id="msg_assistant",
+            providerID="fallback",
+            modelID="fallback-model",
+            finish="stop",
+            tokens=None,
+            time={"created": 123},
+        )
+        run = AsyncMock(return_value=LoopResult(
+            action="stop",
+            last_message=assistant,
+            provider_id="fallback",
+            model_id="fallback-model",
+        ))
+        publish = AsyncMock()
+        context_usage = AsyncMock()
+
+        monkeypatch.setattr(SessionLoop, "run", run)
+        monkeypatch.setattr(
+            "flocks.session.lifecycle.revert.SessionRevert.cleanup",
+            AsyncMock(),
+        )
+        monkeypatch.setattr(
+            "flocks.session.message.Message.get_text_content",
+            AsyncMock(return_value="recovered"),
+        )
+        monkeypatch.setattr("flocks.server.routes.event.publish_event", publish)
+        monkeypatch.setattr(
+            session_routes,
+            "_publish_context_usage_update",
+            context_usage,
+        )
+
+        result = await session_routes._run_existing_user_message(
+            session.id,
+            session,
+            user_message,
+            session.directory,
+            runtime={
+                "agent_name": "rex",
+                "provider_id": "primary",
+                "model_id": "primary-model",
+                "auto_failover": True,
+            },
+        )
+
+        assert result["status"] == "completed"
+        assert run.await_args.kwargs["auto_failover"] is True
+        completion = next(
+            call.args[1]
+            for call in publish.await_args_list
+            if call.args[0] == "message.updated"
+        )
+        assert completion["info"]["providerID"] == "fallback"
+        assert completion["info"]["modelID"] == "fallback-model"
+        assert context_usage.await_args.kwargs["provider_id"] == "fallback"
+        assert context_usage.await_args.kwargs["model_id"] == "fallback-model"
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("category", ["user", "entity-config", "workflow"])
+    async def test_prepare_auto_replay_defers_unavailable_primary_to_failover(
+        self,
+        monkeypatch: pytest.MonkeyPatch,
+        category: str,
+    ):
+        from flocks.server.routes import session as session_routes
+        from flocks.session.session_loop import SessionLoop
+
+        user_message = SimpleNamespace(agent="rex")
+        monkeypatch.setattr(
+            "flocks.agent.registry.Agent.get",
+            AsyncMock(return_value=SimpleNamespace(name="rex", model=None)),
+        )
+        monkeypatch.setattr(
+            "flocks.session.session.Session.get_by_id",
+            AsyncMock(return_value=SimpleNamespace(
+                model_auto=True,
+                category=category,
+            )),
+        )
+        monkeypatch.setattr(
+            "flocks.config.config.Config.resolve_default_llm",
+            AsyncMock(return_value={
+                "provider_id": "primary",
+                "model_id": "primary-model",
+            }),
+        )
+        monkeypatch.setattr(
+            "flocks.config.config.Config.get",
+            AsyncMock(return_value=SimpleNamespace()),
+        )
+        validate = AsyncMock(return_value=(False, "provider_not_configured"))
+        monkeypatch.setattr(SessionLoop, "validate_runtime_model", validate)
+        monkeypatch.setattr("flocks.provider.provider.Provider._ensure_initialized", lambda: None)
+        monkeypatch.setattr("flocks.provider.provider.Provider.apply_config", AsyncMock())
+        monkeypatch.setattr("flocks.provider.provider.Provider.get", lambda _provider_id: None)
+        resolve = AsyncMock()
+        monkeypatch.setattr(session_routes, "_resolve_model", resolve)
+
+        runtime = await session_routes._prepare_replay_runtime("ses_auto", user_message)
+
+        assert runtime == {
+            "agent_name": "rex",
+            "provider_id": "primary",
+            "model_id": "primary-model",
+            "auto_failover": True,
+        }
+        validate.assert_not_awaited()
+        resolve.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_prepare_replay_ignores_auto_on_unsupported_session(
+        self,
+        monkeypatch: pytest.MonkeyPatch,
+    ):
+        from flocks.server.routes import session as session_routes
+        from flocks.session.session_loop import SessionLoop
+
+        user_message = SimpleNamespace(agent="rex")
+        monkeypatch.setattr(
+            "flocks.agent.registry.Agent.get",
+            AsyncMock(return_value=SimpleNamespace(name="rex", model=None)),
+        )
+        monkeypatch.setattr(
+            "flocks.session.session.Session.get_by_id",
+            AsyncMock(return_value=SimpleNamespace(
+                model_auto=True,
+                category="task",
+            )),
+        )
+        resolve = AsyncMock(return_value=("direct", "direct-model", "session"))
+        monkeypatch.setattr(session_routes, "_resolve_model", resolve)
+        default_llm = AsyncMock()
+        monkeypatch.setattr(
+            "flocks.config.config.Config.resolve_default_llm",
+            default_llm,
+        )
+        monkeypatch.setattr(
+            "flocks.config.config.Config.get",
+            AsyncMock(return_value=SimpleNamespace()),
+        )
+        validate = AsyncMock()
+        monkeypatch.setattr(SessionLoop, "validate_runtime_model", validate)
+        monkeypatch.setattr(
+            "flocks.provider.provider.Provider._ensure_initialized",
+            lambda: None,
+        )
+        monkeypatch.setattr(
+            "flocks.provider.provider.Provider.apply_config",
+            AsyncMock(),
+        )
+        monkeypatch.setattr(
+            "flocks.provider.provider.Provider.get",
+            lambda _provider_id: object(),
+        )
+
+        runtime = await session_routes._prepare_replay_runtime(
+            "ses_workflow",
+            user_message,
+        )
+
+        assert runtime == {
+            "agent_name": "rex",
+            "provider_id": "direct",
+            "model_id": "direct-model",
+            "auto_failover": False,
+        }
+        resolve.assert_awaited_once()
+        default_llm.assert_not_awaited()
+        validate.assert_not_awaited()
+
     @pytest.mark.asyncio
     async def test_resend_uses_current_model_for_replay(
         self,
@@ -1679,7 +2346,11 @@ async def fake_publish_event(_event: str, _payload: dict) -> None:
         monkeypatch.setattr(
             session_routes.Session,
             "get_by_id",
-            AsyncMock(return_value=SimpleNamespace(id=session_id, directory="/tmp/project")),
+            AsyncMock(return_value=SimpleNamespace(
+                id=session_id,
+                directory="/tmp/project",
+                status="active",
+            )),
         )
         monkeypatch.setattr(session_routes, "abort_session", fake_abort_session)
         monkeypatch.setattr(session_routes, "_wait_for_session_idle", fake_wait_for_session_idle)
diff --git a/tests/server/routes/test_workflow_poller_routes.py b/tests/server/routes/test_workflow_poller_routes.py
index f683b4057..66bf57eab 100644
--- a/tests/server/routes/test_workflow_poller_routes.py
+++ b/tests/server/routes/test_workflow_poller_routes.py
@@ -23,6 +23,13 @@ async def _fake_restart(workflow_id: str) -> dict[str, Any]:
         assert workflow_id == "wf-1"
         return {"workflowId": workflow_id, "state": "running", "lastStatus": None}
 
+    async def _fake_persist(
+        _workflow_id: str,
+        workflow_data: dict[str, Any],
+        _triggers: list[Any],
+    ) -> dict[str, Any]:
+        return workflow_data
+
     monkeypatch.setattr(
         workflow_routes,
         "_read_workflow_from_fs",
@@ -31,6 +38,7 @@ async def _fake_restart(workflow_id: str) -> dict[str, Any]:
         ),
     )
     monkeypatch.setattr(workflow_routes.WorkflowStore, "put_config", _fake_put_config)
+    monkeypatch.setattr(workflow_routes, "_persist_workflow_triggers", _fake_persist)
     monkeypatch.setattr(
         "flocks.workflow.poller_manager.default_manager",
         SimpleNamespace(restart_workflow=_fake_restart),
diff --git a/tests/server/test_auth_service_reassign.py b/tests/server/test_auth_service_reassign.py
index 1bcb37307..159d94e9a 100644
--- a/tests/server/test_auth_service_reassign.py
+++ b/tests/server/test_auth_service_reassign.py
@@ -60,7 +60,8 @@ async def test_reassign_orphan_sessions_skips_owned_and_rewrites_orphans(monkeyp
     async def _list_all():
         return listed
 
-    async def _update(*, project_id, session_id, owner_user_id, owner_username):
+    async def _update(*, project_id, session_id, owner_user_id, owner_username, allow_inactive=False):
+        assert allow_inactive is True
         update_calls.append({
             "project_id": project_id,
             "session_id": session_id,
@@ -119,7 +120,8 @@ async def test_reassign_orphan_sessions_continues_on_partial_failure(monkeypatch
     async def _list_all():
         return listed
 
-    async def _update(*, project_id, session_id, owner_user_id, owner_username):
+    async def _update(*, project_id, session_id, owner_user_id, owner_username, allow_inactive=False):
+        assert allow_inactive is True
         update_calls.append(session_id)
         if session_id == "ses_b":
             raise RuntimeError("storage write failed")
diff --git a/tests/server/test_input_dispatcher.py b/tests/server/test_input_dispatcher.py
index 17a11d8f0..3a2c767cc 100644
--- a/tests/server/test_input_dispatcher.py
+++ b/tests/server/test_input_dispatcher.py
@@ -288,6 +288,63 @@ async def test_channel_unsafe_command_is_rejected(self):
 
 
 class TestSessionRoutesUseDispatcher:
+    @pytest.mark.asyncio
+    async def test_goal_mode_publishes_active_goal_before_llm(self, monkeypatch):
+        from flocks.input.events import UserInputEvent
+        from flocks.server.routes import session as session_routes
+
+        order = []
+        goal_state = SimpleNamespace(
+            status="active",
+            objective="fix tests",
+            last_reason=None,
+        )
+
+        async def publish(event_type, _properties):
+            if event_type == "session.goal.updated":
+                order.append("goal")
+
+        async def process(*_args, **_kwargs):
+            order.append("llm")
+
+        monkeypatch.setattr(
+            "flocks.command.direct.GoalManager.set_goal",
+            AsyncMock(return_value=goal_state),
+        )
+        monkeypatch.setattr(
+            "flocks.command.direct.GoalManager.goal_prompt",
+            MagicMock(return_value="goal prompt"),
+        )
+        monkeypatch.setattr(
+            "flocks.session.goal.GoalManager.get",
+            AsyncMock(return_value=goal_state),
+        )
+        monkeypatch.setattr(
+            "flocks.server.routes.event.publish_event",
+            publish,
+        )
+        monkeypatch.setattr(
+            session_routes,
+            "_process_session_message",
+            process,
+        )
+
+        await session_routes._dispatch_sse_input(
+            "ses_goal_mode",
+            SimpleNamespace(id="ses_goal_mode"),
+            UserInputEvent(
+                source_type="webui",
+                sessionID="ses_goal_mode",
+                text="/goal fix tests",
+                parts=[{"type": "text", "text": "fix tests"}],
+                display_text="fix tests",
+                executionMode="goal",
+            ),
+            "/tmp/project",
+        )
+
+        assert order == ["goal", "llm"]
+
     @pytest.mark.asyncio
     async def test_prompt_async_routes_through_dispatcher(self, monkeypatch):
         from flocks.server.routes import session as session_routes
@@ -389,7 +446,9 @@ def test_session_upload_path_rejects_traversal(self, monkeypatch, tmp_path):
             session_routes._session_uploads_dir("../outside")
 
     @pytest.mark.asyncio
-    async def test_prompt_async_queues_when_session_running_without_creating_message(self, monkeypatch):
+    async def test_prompt_async_queues_when_session_running_without_creating_message(
+        self, monkeypatch, tmp_path
+    ):
         from flocks.server.routes import session as session_routes
         from flocks.session.interaction_queue import InteractionQueue
 
@@ -399,7 +458,13 @@ async def test_prompt_async_queues_when_session_running_without_creating_message
         message_create = AsyncMock()
         monkeypatch.setattr(
             "flocks.session.session.Session.get_by_id",
-            AsyncMock(return_value=SimpleNamespace(id=session_id, directory="/tmp/project")),
+            AsyncMock(
+                return_value=SimpleNamespace(
+                    id=session_id,
+                    directory=str(tmp_path),
+                    status="active",
+                )
+            ),
         )
         monkeypatch.setattr("flocks.session.session_loop.SessionLoop.is_running", lambda _sid: True)
         monkeypatch.setattr("flocks.session.message.Message.create", message_create)
@@ -417,7 +482,7 @@ async def test_prompt_async_queues_when_session_running_without_creating_message
         message_create.assert_not_called()
 
     @pytest.mark.asyncio
-    async def test_prompt_queue_rejects_when_full(self, monkeypatch):
+    async def test_prompt_queue_rejects_when_full(self, monkeypatch, tmp_path):
         from fastapi import HTTPException
 
         from flocks.server.routes import session as session_routes
@@ -427,7 +492,13 @@ async def test_prompt_queue_rejects_when_full(self, monkeypatch):
         await InteractionQueue.clear(session_id)
         monkeypatch.setattr(
             "flocks.session.session.Session.get_by_id",
-            AsyncMock(return_value=SimpleNamespace(id=session_id, directory="/tmp/project")),
+            AsyncMock(
+                return_value=SimpleNamespace(
+                    id=session_id,
+                    directory=str(tmp_path),
+                    status="active",
+                )
+            ),
         )
         monkeypatch.setattr("flocks.session.session_loop.SessionLoop.is_running", lambda _sid: True)
         monkeypatch.setattr(session_routes, "_publish_prompt_queue", AsyncMock())
@@ -445,7 +516,7 @@ async def test_prompt_queue_rejects_when_full(self, monkeypatch):
         assert exc_info.value.status_code == 409
 
     @pytest.mark.asyncio
-    async def test_run_now_aborts_and_schedules_drain(self, monkeypatch):
+    async def test_run_now_aborts_and_schedules_drain(self, monkeypatch, tmp_path):
         from flocks.server.routes import session as session_routes
         from flocks.session.interaction_queue import InteractionQueue
 
@@ -461,7 +532,13 @@ async def test_run_now_aborts_and_schedules_drain(self, monkeypatch):
         drain_mock = AsyncMock()
         monkeypatch.setattr(
             "flocks.session.session.Session.get_by_id",
-            AsyncMock(return_value=SimpleNamespace(id=session_id, directory="/tmp/project")),
+            AsyncMock(
+                return_value=SimpleNamespace(
+                    id=session_id,
+                    directory=str(tmp_path),
+                    status="active",
+                )
+            ),
         )
         monkeypatch.setattr("flocks.session.session_loop.SessionLoop.is_running", lambda _sid: True)
         monkeypatch.setattr(session_routes, "abort_session", abort_mock)
@@ -474,7 +551,7 @@ async def test_run_now_aborts_and_schedules_drain(self, monkeypatch):
         assert resp["status"] == "accepted"
         abort_mock.assert_awaited_once_with(session_id)
         wait_mock.assert_awaited_once_with(session_id)
-        drain_mock.assert_awaited_once_with(session_id, "/tmp/project")
+        drain_mock.assert_awaited_once_with(session_id, str(tmp_path.resolve()))
 
     @pytest.mark.asyncio
     async def test_scheduled_drain_retries_until_session_idle(self, monkeypatch):
diff --git a/tests/server/test_server_port_config.py b/tests/server/test_server_port_config.py
index 3964ff3d2..fb6b1f867 100644
--- a/tests/server/test_server_port_config.py
+++ b/tests/server/test_server_port_config.py
@@ -307,6 +307,22 @@ def fake_restart_all(config, _console):
         assert captured["config"].frontend_port == 5273
         assert captured["config"].legacy_backend_port == 9100
 
+    def test_restart_server_only_does_not_restart_daemon(self, monkeypatch):
+        """Test server-only restart leaves the supervisor daemon running."""
+        calls = []
+
+        monkeypatch.setattr(cli_main, "restart_server", lambda _console: calls.append("server"))
+        monkeypatch.setattr(
+            cli_main,
+            "restart_all",
+            lambda _config, _console: calls.append("all"),
+        )
+
+        result = CliRunner().invoke(cli_main.app, ["restart", "--server-only"])
+
+        assert result.exit_code == 0
+        assert calls == ["server"]
+
     def test_restart_accepts_public_host_and_port(self, monkeypatch):
         """Test restart command accepts the unified public host/port options."""
         captured = {}
diff --git a/tests/session/test_auto_model_failover.py b/tests/session/test_auto_model_failover.py
new file mode 100644
index 000000000..98541ec3c
--- /dev/null
+++ b/tests/session/test_auto_model_failover.py
@@ -0,0 +1,1380 @@
+"""Focused tests for WebUI Auto runtime model failover."""
+
+import time
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from flocks.session.message import Message, MessageRole
+from flocks.session.runner import (
+    LlmAttemptState,
+    SessionRunner,
+    StepFailure,
+    StepResult,
+)
+from flocks.session.session import Session, SessionInfo
+from flocks.session.session_loop import (
+    AutoFailoverCooldown,
+    LoopCallbacks,
+    LoopContext,
+    LoopResult,
+    RuntimeModel,
+    SessionLoop,
+)
+
+
+def _session(**updates) -> SessionInfo:
+    values = {
+        "id": "ses_auto",
+        "projectID": "project",
+        "directory": "/tmp/project",
+        "agent": "rex",
+        "provider": "primary",
+        "model": "primary-model",
+        "model_pinned": False,
+        "model_auto": True,
+    }
+    values.update(updates)
+    return SessionInfo.model_construct(**values)
+
+
+def _ctx(
+    *,
+    auto: bool = True,
+    index: int = 0,
+    category: str = "user",
+) -> LoopContext:
+    candidates = [
+        RuntimeModel("primary", "primary-model"),
+        RuntimeModel("fallback", "fallback-model"),
+    ]
+    active = candidates[index]
+    return LoopContext(
+        session=_session(
+            provider=active.provider_id,
+            model=active.model_id,
+            category=category,
+        ),
+        provider_id=active.provider_id,
+        model_id=active.model_id,
+        agent_name="rex",
+        auto_failover=auto,
+        auto_failover_allowed=auto,
+        model_candidates=candidates if auto else [active],
+        candidate_index=index if auto else 0,
+    )
+
+
+def _failure(
+    *,
+    assistant_id: str,
+    reason: str = "server_error",
+    safe: bool = True,
+) -> StepResult:
+    state = LlmAttemptState(observable_output_started=not safe)
+    message = "provider failed"
+    return StepResult(
+        action="stop",
+        error=message,
+        failure=StepFailure(
+            message=message,
+            error_data={"name": "APIError", "data": {"message": message}},
+            assistant_message_id=assistant_id,
+            reason=reason,
+            allow_fallback=safe,
+            attempt_state=state,
+            attempts=1,
+        ),
+    )
+
+
+@pytest.fixture(autouse=True)
+def _clear_cooldowns():
+    SessionLoop._auto_failover_cooldowns.clear()
+    yield
+    SessionLoop._auto_failover_cooldowns.clear()
+
+
+@pytest.mark.parametrize(
+    ("status_code", "message", "reason", "same_model_retries"),
+    [
+        (401, "Unauthorized", "auth", 0),
+        (402, "Payment required", "billing", 0),
+        (429, "Too many requests", "rate_limit", 0),
+        (403, "Quota exceeded", "rate_limit", 0),
+        (403, "Insufficient quota", "billing", 0),
+        (408, "Request timeout", "timeout", 1),
+        (404, "Route not found", "unknown_api", 3),
+        (500, "Internal server error", "server_error", 3),
+        (502, "Bad gateway", "server_error", 3),
+        (529, "Provider overloaded", "overloaded", 1),
+    ],
+)
+def test_failover_classifier_retry_thresholds(
+    status_code: int,
+    message: str,
+    reason: str,
+    same_model_retries: int,
+):
+    decision = SessionRunner.classify_failover_error({
+        "name": "APIError",
+        "data": {"message": message, "statusCode": status_code},
+    })
+
+    assert decision.eligible is True
+    assert decision.reason == reason
+    assert decision.same_model_retries == same_model_retries
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("status_code", "expected_calls"),
+    [
+        (401, 1),
+        (402, 1),
+        (429, 1),
+        (408, 2),
+        (404, 4),
+        (500, 4),
+        (502, 4),
+        (529, 2),
+    ],
+)
+async def test_runner_applies_failover_retry_thresholds(
+    monkeypatch,
+    status_code: int,
+    expected_calls: int,
+):
+    runner = SessionRunner(
+        session=_session(),
+        provider_id="primary",
+        model_id="primary-model",
+        defer_step_errors=True,
+        failover_available=True,
+    )
+    last_user = SimpleNamespace(id="msg_user", agent="rex", role="user")
+    provider = MagicMock()
+    provider.is_configured.return_value = True
+    assistant = SimpleNamespace(id="msg_assistant")
+    failure = RuntimeError(f"Provider HTTP {status_code}")
+    failure.status_code = status_code
+    call_llm = AsyncMock(side_effect=failure)
+
+    monkeypatch.setattr(
+        "flocks.session.runner.Agent.get",
+        AsyncMock(return_value=SimpleNamespace(
+            name="rex",
+            steps=None,
+            mode="primary",
+            prompt="",
+            tools=[],
+        )),
+    )
+    monkeypatch.setattr("flocks.session.runner.Provider.get", lambda _provider_id: provider)
+    monkeypatch.setattr("flocks.session.runner.Provider.apply_config", AsyncMock())
+    monkeypatch.setattr(
+        "flocks.session.runner.SessionPrompt.build_system_prompts",
+        AsyncMock(return_value=[]),
+    )
+    monkeypatch.setattr(runner, "_build_callable_tool_schema", AsyncMock(return_value=[]))
+    monkeypatch.setattr(
+        runner,
+        "_to_chat_messages",
+        AsyncMock(return_value=[SimpleNamespace(role="user", content="hello")]),
+    )
+    monkeypatch.setattr(Message, "get_text_content", AsyncMock(return_value="hello"))
+    monkeypatch.setattr(Message, "parts", AsyncMock(return_value=[]))
+    monkeypatch.setattr(Message, "create", AsyncMock(return_value=assistant))
+    monkeypatch.setattr(Message, "update", AsyncMock())
+    monkeypatch.setattr(runner, "_call_llm", call_llm)
+    monkeypatch.setattr("flocks.session.runner.SessionRetry.sleep", AsyncMock())
+
+    result = await runner._process_step([last_user], last_user)
+
+    assert result.failure is not None
+    assert call_llm.await_count == expected_calls
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("status_code", "expected_calls"),
+    [
+        (429, 1),
+        (503, 2),
+    ],
+)
+async def test_last_auto_candidate_keeps_hermes_retry_thresholds(
+    monkeypatch,
+    status_code: int,
+    expected_calls: int,
+):
+    """The last candidate still has Auto's per-model retry budget."""
+    runner = SessionRunner(
+        session=_session(),
+        provider_id="fallback",
+        model_id="fallback-model",
+        defer_step_errors=True,
+        failover_available=False,
+    )
+    last_user = SimpleNamespace(id="msg_user", agent="rex", role="user")
+    provider = MagicMock()
+    provider.is_configured.return_value = True
+    assistant = SimpleNamespace(id="msg_assistant")
+    failure = RuntimeError(f"Provider HTTP {status_code}")
+    failure.status_code = status_code
+    call_llm = AsyncMock(side_effect=failure)
+
+    monkeypatch.setattr(
+        "flocks.session.runner.Agent.get",
+        AsyncMock(return_value=SimpleNamespace(
+            name="rex",
+            steps=None,
+            mode="primary",
+            prompt="",
+            tools=[],
+        )),
+    )
+    monkeypatch.setattr("flocks.session.runner.Provider.get", lambda _provider_id: provider)
+    monkeypatch.setattr("flocks.session.runner.Provider.apply_config", AsyncMock())
+    monkeypatch.setattr(
+        "flocks.session.runner.SessionPrompt.build_system_prompts",
+        AsyncMock(return_value=[]),
+    )
+    monkeypatch.setattr(runner, "_build_callable_tool_schema", AsyncMock(return_value=[]))
+    monkeypatch.setattr(
+        runner,
+        "_to_chat_messages",
+        AsyncMock(return_value=[SimpleNamespace(role="user", content="hello")]),
+    )
+    monkeypatch.setattr(Message, "get_text_content", AsyncMock(return_value="hello"))
+    monkeypatch.setattr(Message, "parts", AsyncMock(return_value=[]))
+    monkeypatch.setattr(Message, "create", AsyncMock(return_value=assistant))
+    monkeypatch.setattr(Message, "update", AsyncMock())
+    monkeypatch.setattr(runner, "_call_llm", call_llm)
+    monkeypatch.setattr("flocks.session.runner.SessionRetry.sleep", AsyncMock())
+
+    result = await runner._process_step([last_user], last_user)
+
+    assert result.failure is not None
+    assert call_llm.await_count == expected_calls
+
+
+@pytest.mark.parametrize(
+    ("exception", "status_code", "reason"),
+    [
+        (
+            type("GoogleSdkError", (RuntimeError,), {"code": 429})(
+                "Resource exhausted"
+            ),
+            429,
+            "rate_limit",
+        ),
+        (
+            type(
+                "ResponseSdkError",
+                (RuntimeError,),
+                {
+                    "response": SimpleNamespace(
+                        status_code=503,
+                        headers={"retry-after": "1"},
+                    )
+                },
+            )("Service unavailable"),
+            503,
+            "overloaded",
+        ),
+    ],
+)
+def test_exception_status_is_normalized_from_sdk_shapes(
+    exception: Exception,
+    status_code: int,
+    reason: str,
+):
+    runner = SessionRunner(
+        session=_session(),
+        provider_id="primary",
+        model_id="primary-model",
+    )
+
+    error = runner._exception_to_error_dict(exception)
+
+    assert error["data"]["statusCode"] == status_code
+    assert SessionRunner.classify_failover_error(error).reason == reason
+
+
+def test_exception_status_is_normalized_from_cause_chain():
+    inner = type("GoogleSdkError", (RuntimeError,), {"code": 401})(
+        "Unauthenticated"
+    )
+    outer = RuntimeError("Provider wrapper failed")
+    outer.__cause__ = inner
+    runner = SessionRunner(
+        session=_session(),
+        provider_id="primary",
+        model_id="primary-model",
+    )
+
+    error = runner._exception_to_error_dict(outer)
+
+    assert error["data"]["statusCode"] == 401
+    assert SessionRunner.classify_failover_error(error).reason == "auth"
+
+
+def test_local_validation_error_never_fails_over():
+    decision = SessionRunner.classify_failover_error({
+        "name": "ValidationError",
+        "data": {"message": "Local prompt schema validation failed"},
+    })
+
+    assert decision.eligible is False
+    assert decision.reason == "local_error"
+
+
+def test_model_not_found_without_status_fails_over():
+    decision = SessionRunner.classify_failover_error({
+        "name": "ValueError",
+        "data": {"message": "Model acme-v2 not found for provider custom"},
+    })
+
+    assert decision.eligible is True
+    assert decision.reason == "model_not_found"
+    assert decision.same_model_retries == 0
+
+
+def test_content_filter_error_fails_over_immediately():
+    decision = SessionRunner.classify_failover_error({
+        "name": "BadRequestError",
+        "data": {"message": "Response blocked by content_filter"},
+    })
+
+    assert decision.eligible is True
+    assert decision.reason == "content_policy"
+    assert decision.same_model_retries == 0
+
+
+def test_candidate_switch_keeps_tool_loop_guard_only():
+    ctx = _ctx()
+    tool_loop_guard = {
+        "last_user_id": "msg_user",
+        "signature": "same-tool-call",
+        "count": 2,
+    }
+    ctx.runner_static_cache.update({
+        "tool_loop_guard": tool_loop_guard,
+        "tool_schema_cache": {"primary": "schema"},
+        "chat_context_cache": {"primary": "context"},
+        "system_prompt": "primary prompt",
+    })
+
+    SessionLoop._select_candidate(ctx, 1)
+
+    assert ctx.runner_static_cache == {"tool_loop_guard": tool_loop_guard}
+    assert ctx.runner_static_cache["tool_loop_guard"] is tool_loop_guard
+
+
+@pytest.mark.asyncio
+async def test_reasoning_only_empty_response_is_not_replayed(monkeypatch):
+    runner = SessionRunner(
+        session=_session(),
+        provider_id="primary",
+        model_id="primary-model",
+        defer_step_errors=True,
+        failover_available=True,
+    )
+    last_user = SimpleNamespace(id="msg_user", agent="rex", role="user")
+    provider = MagicMock()
+    provider.is_configured.return_value = True
+    assistant = SimpleNamespace(id="msg_assistant")
+    call_count = 0
+
+    async def call_llm(*_args, **_kwargs):
+        nonlocal call_count
+        call_count += 1
+        runner._attempt_state.observable_output_started = True
+        return StepResult(action="stop", content="")
+
+    monkeypatch.setattr(
+        "flocks.session.runner.Agent.get",
+        AsyncMock(return_value=SimpleNamespace(
+            name="rex",
+            steps=None,
+            mode="primary",
+            prompt="",
+            tools=[],
+        )),
+    )
+    monkeypatch.setattr("flocks.session.runner.Provider.get", lambda _provider_id: provider)
+    monkeypatch.setattr("flocks.session.runner.Provider.apply_config", AsyncMock())
+    monkeypatch.setattr(
+        "flocks.session.runner.SessionPrompt.build_system_prompts",
+        AsyncMock(return_value=[]),
+    )
+    monkeypatch.setattr(runner, "_build_callable_tool_schema", AsyncMock(return_value=[]))
+    monkeypatch.setattr(
+        runner,
+        "_to_chat_messages",
+        AsyncMock(return_value=[SimpleNamespace(role="user", content="hello")]),
+    )
+    monkeypatch.setattr(Message, "get_text_content", AsyncMock(return_value="hello"))
+    monkeypatch.setattr(Message, "parts", AsyncMock(return_value=[]))
+    monkeypatch.setattr(Message, "create", AsyncMock(return_value=assistant))
+    monkeypatch.setattr(Message, "update", AsyncMock())
+    monkeypatch.setattr(runner, "_call_llm", call_llm)
+    sleep = AsyncMock()
+    monkeypatch.setattr("flocks.session.runner.SessionRetry.sleep", sleep)
+
+    result = await runner._process_step([last_user], last_user)
+
+    assert call_count == 1
+    assert result.failure is not None
+    assert result.failure.allow_fallback is False
+    sleep.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("chunk_kind", ["text", "reasoning", "tool"])
+async def test_real_stream_activity_prevents_retry_and_fallback(
+    monkeypatch,
+    chunk_kind: str,
+):
+    """Exercise the real stream loop, then fail after an observable fragment."""
+
+    class FakeStreamProcessor:
+        def __init__(self, *args, tool_start_callback=None, **kwargs):
+            self.tool_start_callback = tool_start_callback
+            self.tool_calls = {}
+            self._text = []
+            self._reasoning = []
+
+        async def process_event(self, event):
+            event_name = type(event).__name__
+            if event_name == "TextDeltaEvent":
+                self._text.append(event.text)
+            elif event_name == "ReasoningDeltaEvent":
+                self._reasoning.append(event.text)
+            elif event_name == "ToolCallEvent":
+                if self.tool_start_callback:
+                    await self.tool_start_callback(event.tool_name, event.input)
+                self.tool_calls[event.tool_call_id] = SimpleNamespace(
+                    id=event.tool_call_id,
+                    name=event.tool_name,
+                    input=event.input,
+                )
+
+        def get_text_content(self):
+            return "".join(self._text)
+
+        def get_reasoning_content(self):
+            return "".join(self._reasoning)
+
+    if chunk_kind == "text":
+        chunk = SimpleNamespace(
+            delta="visible text",
+            reasoning=None,
+            event_type=None,
+            metadata={},
+            tool_calls=None,
+            finish_reason=None,
+            usage=None,
+        )
+    elif chunk_kind == "reasoning":
+        chunk = SimpleNamespace(
+            delta="visible reasoning",
+            reasoning=None,
+            event_type="reasoning",
+            metadata={},
+            tool_calls=None,
+            finish_reason=None,
+            usage=None,
+        )
+    else:
+        chunk = SimpleNamespace(
+            delta="",
+            reasoning=None,
+            event_type=None,
+            metadata={},
+            tool_calls=[{
+                "index": 0,
+                "id": "call_1",
+                "function": {"name": "example_tool", "arguments": "{}"},
+            }],
+            finish_reason=None,
+            usage=None,
+        )
+
+    class FailingStreamProvider:
+        def __init__(self):
+            self.calls = 0
+
+        def is_configured(self):
+            return True
+
+        def chat_stream(self, **_kwargs):
+            self.calls += 1
+
+            async def stream():
+                yield chunk
+                failure = RuntimeError("Provider HTTP 500")
+                failure.status_code = 500
+                raise failure
+
+            return stream()
+
+    provider = FailingStreamProvider()
+    runner = SessionRunner(
+        session=_session(),
+        provider_id="primary",
+        model_id="primary-model",
+        defer_step_errors=True,
+        failover_available=True,
+    )
+    last_user = SimpleNamespace(id="msg_user", agent="rex", role="user")
+    assistant = SimpleNamespace(id="msg_assistant")
+
+    monkeypatch.setattr(
+        "flocks.session.runner.Agent.get",
+        AsyncMock(return_value=SimpleNamespace(
+            name="rex",
+            steps=None,
+            mode="primary",
+            prompt="",
+            tools=[],
+        )),
+    )
+    monkeypatch.setattr("flocks.session.runner.Provider.get", lambda _provider_id: provider)
+    monkeypatch.setattr("flocks.session.runner.Provider.apply_config", AsyncMock())
+    monkeypatch.setattr(
+        "flocks.session.runner.SessionPrompt.build_system_prompts",
+        AsyncMock(return_value=[]),
+    )
+    monkeypatch.setattr(runner, "_build_callable_tool_schema", AsyncMock(return_value=[]))
+    monkeypatch.setattr(
+        runner,
+        "_to_chat_messages",
+        AsyncMock(return_value=[SimpleNamespace(role="user", content="hello")]),
+    )
+    monkeypatch.setattr(runner, "_should_use_text_tool_call_mode", lambda: False)
+    monkeypatch.setattr(Message, "get_text_content", AsyncMock(return_value="hello"))
+    monkeypatch.setattr(Message, "parts", AsyncMock(return_value=[]))
+    monkeypatch.setattr(Message, "create", AsyncMock(return_value=assistant))
+    monkeypatch.setattr(Message, "update", AsyncMock())
+    monkeypatch.setattr("flocks.session.runner.StreamProcessor", FakeStreamProcessor)
+    monkeypatch.setattr(
+        "flocks.session.runner.HookPipeline.has_stage_handlers",
+        AsyncMock(return_value=False),
+    )
+    monkeypatch.setattr("flocks.session.runner.langfuse_is_active", lambda: False)
+    monkeypatch.setattr(
+        "flocks.provider.options.build_provider_options",
+        lambda _provider_id, _model_id: {},
+    )
+    sleep = AsyncMock()
+    monkeypatch.setattr("flocks.session.runner.SessionRetry.sleep", sleep)
+
+    result = await runner._process_step([last_user], last_user)
+
+    assert provider.calls == 1
+    assert result.failure is not None
+    assert result.failure.allow_fallback is False
+    assert result.failure.attempt_state.received_chunk is True
+    assert result.failure.attempt_state.observable_output_started is True
+    assert result.failure.attempt_state.tool_execution_started is (
+        chunk_kind == "tool"
+    )
+    sleep.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_safe_failure_switches_candidate_and_removes_blank_message(monkeypatch):
+    ctx = _ctx()
+    last_user = SimpleNamespace(id="msg_user", agent="rex")
+    events = []
+
+    async def process_step(runner, _messages, _last_user):
+        if runner.provider_id == "primary":
+            return _failure(assistant_id="msg_failed")
+        return StepResult(action="stop", content="recovered")
+
+    monkeypatch.setattr(SessionRunner, "_process_step", process_step)
+    delete = AsyncMock(return_value=True)
+    monkeypatch.setattr(Message, "delete", delete)
+
+    async def publish(event, payload):
+        events.append((event, payload))
+
+    result = await SessionLoop._process_step_with_failover(
+        ctx,
+        LoopCallbacks(event_publish_callback=publish),
+        [last_user],
+        last_user,
+    )
+
+    assert result.content == "recovered"
+    assert (ctx.provider_id, ctx.model_id) == ("fallback", "fallback-model")
+    delete.assert_awaited_once_with("ses_auto", "msg_failed")
+    assert any(event == "message.removed" for event, _ in events)
+    assert any(event == "session.model.fallback" for event, _ in events)
+
+
+@pytest.mark.asyncio
+async def test_queued_user_is_detected_before_replacement_assistant():
+    current_user = SimpleNamespace(id="msg_001", role=MessageRole.USER)
+    queued_user = SimpleNamespace(id="msg_002", role=MessageRole.USER)
+    replacement_assistant = SimpleNamespace(
+        id="msg_003",
+        role=MessageRole.ASSISTANT,
+    )
+
+    detected = await SessionLoop._detect_queued_user_message(
+        "ses_auto",
+        [current_user, queued_user, replacement_assistant],
+        current_user.id,
+        replacement_assistant,
+    )
+
+    assert detected is queued_user
+
+
+@pytest.mark.asyncio
+async def test_preflight_chain_exhaustion_persists_final_error(monkeypatch):
+    ctx = _ctx()
+    last_user = SimpleNamespace(
+        id="msg_user",
+        agent="rex",
+        role=MessageRole.USER,
+    )
+
+    async def preflight_failure(_runner, _messages, _last_user):
+        message = "Provider not configured"
+        return StepResult(
+            action="stop",
+            error=message,
+            failure=StepFailure(
+                message=message,
+                error_data={
+                    "name": "ProviderUnavailableError",
+                    "data": {"message": message},
+                },
+                assistant_message_id=None,
+                reason="provider_unavailable",
+                allow_fallback=True,
+                attempt_state=LlmAttemptState(),
+                attempts=0,
+            ),
+        )
+
+    final_assistant = SimpleNamespace(id="msg_final_error")
+    create = AsyncMock(return_value=final_assistant)
+    monkeypatch.setattr(SessionRunner, "_process_step", preflight_failure)
+    monkeypatch.setattr(Message, "create", create)
+
+    result = await SessionLoop._process_step_with_failover(
+        ctx,
+        LoopCallbacks(),
+        [last_user],
+        last_user,
+    )
+
+    assert result.error == "Provider not configured"
+    assert result.failure is not None
+    assert result.failure.assistant_message_id == final_assistant.id
+    create.assert_awaited_once_with(
+        session_id="ses_auto",
+        role=MessageRole.ASSISTANT,
+        content="",
+        agent="rex",
+        model_id="fallback-model",
+        provider_id="fallback",
+        parent_id=last_user.id,
+        error={
+            "name": "ProviderUnavailableError",
+            "data": {"message": "Provider not configured"},
+        },
+        finish="error",
+    )
+
+
+@pytest.mark.asyncio
+async def test_failed_blank_message_deletion_stops_switch(monkeypatch):
+    ctx = _ctx()
+    last_user = SimpleNamespace(id="msg_user", agent="rex")
+    monkeypatch.setattr(
+        SessionRunner,
+        "_process_step",
+        AsyncMock(return_value=_failure(assistant_id="msg_failed")),
+    )
+    monkeypatch.setattr(Message, "delete", AsyncMock(return_value=False))
+    update = AsyncMock()
+    monkeypatch.setattr(Message, "update", update)
+
+    result = await SessionLoop._process_step_with_failover(
+        ctx,
+        LoopCallbacks(),
+        [last_user],
+        last_user,
+    )
+
+    assert result.error == "provider failed"
+    assert ctx.provider_id == "primary"
+    update.assert_awaited_once_with(
+        "ses_auto",
+        "msg_failed",
+        error={"name": "APIError", "data": {"message": "provider failed"}},
+        finish="error",
+    )
+
+
+@pytest.mark.asyncio
+async def test_fallbacks_are_attempted_in_candidate_order(monkeypatch):
+    ctx = _ctx()
+    ctx.model_candidates = [
+        RuntimeModel("primary", "primary-model"),
+        RuntimeModel("fallback-1", "model-1"),
+        RuntimeModel("fallback-2", "model-2"),
+    ]
+    last_user = SimpleNamespace(id="msg_user", agent="rex")
+    attempts = []
+
+    async def process_step(runner, _messages, _last_user):
+        attempts.append((runner.provider_id, runner.model_id))
+        if runner.provider_id != "fallback-2":
+            return _failure(assistant_id=f"msg_{runner.provider_id}")
+        return StepResult(action="stop", content="recovered")
+
+    monkeypatch.setattr(SessionRunner, "_process_step", process_step)
+    delete = AsyncMock(return_value=True)
+    monkeypatch.setattr(Message, "delete", delete)
+
+    result = await SessionLoop._process_step_with_failover(
+        ctx,
+        LoopCallbacks(),
+        [last_user],
+        last_user,
+    )
+
+    assert result.content == "recovered"
+    assert attempts == [
+        ("primary", "primary-model"),
+        ("fallback-1", "model-1"),
+        ("fallback-2", "model-2"),
+    ]
+    assert delete.await_count == 2
+
+
+@pytest.mark.asyncio
+async def test_chain_exhaustion_finalizes_only_last_candidate(monkeypatch):
+    ctx = _ctx()
+    ctx.model_candidates = [
+        RuntimeModel("primary", "primary-model"),
+        RuntimeModel("fallback-1", "model-1"),
+        RuntimeModel("fallback-2", "model-2"),
+    ]
+    last_user = SimpleNamespace(id="msg_user", agent="rex")
+
+    async def process_step(runner, _messages, _last_user):
+        return _failure(assistant_id=f"msg_{runner.provider_id}")
+
+    monkeypatch.setattr(SessionRunner, "_process_step", process_step)
+    delete = AsyncMock(return_value=True)
+    update = AsyncMock()
+    monkeypatch.setattr(Message, "delete", delete)
+    monkeypatch.setattr(Message, "update", update)
+
+    result = await SessionLoop._process_step_with_failover(
+        ctx,
+        LoopCallbacks(),
+        [last_user],
+        last_user,
+    )
+
+    assert result.error == "provider failed"
+    assert delete.await_count == 2
+    update.assert_awaited_once()
+    assert update.await_args.args[1] == "msg_fallback-2"
+    cooldown = SessionLoop._auto_failover_cooldowns[ctx.session.id]
+    assert cooldown.model == RuntimeModel("fallback-2", "model-2")
+    assert cooldown.reason == "chain_exhausted"
+
+
+@pytest.mark.asyncio
+async def test_full_loop_reports_chain_exhaustion_once(monkeypatch):
+    ctx = _ctx()
+    ctx.session.memory_enabled = False
+    user = SimpleNamespace(
+        id="msg_user",
+        role=MessageRole.USER,
+        agent="rex",
+        model={"providerID": "primary", "modelID": "primary-model"},
+    )
+    # This test exercises exhaustion of an already fixed per-turn chain.
+    ctx.turn_user_id = user.id
+    final_assistant = SimpleNamespace(
+        id="msg_fallback",
+        role=MessageRole.ASSISTANT,
+        parentID=user.id,
+        finish="error",
+    )
+    ctx.session_ctx = SimpleNamespace(
+        get_messages=AsyncMock(side_effect=[
+            [user],
+            [user, final_assistant],
+        ])
+    )
+    attempts = []
+
+    async def process_step(runner, _messages, _last_user):
+        attempts.append((runner.provider_id, runner.model_id))
+        return _failure(assistant_id=f"msg_{runner.provider_id}")
+
+    monkeypatch.setattr(SessionRunner, "_process_step", process_step)
+    monkeypatch.setattr(Message, "parts", AsyncMock(return_value=[]))
+    monkeypatch.setattr(Message, "delete", AsyncMock(return_value=True))
+    update = AsyncMock()
+    monkeypatch.setattr(Message, "update", update)
+    on_error = AsyncMock()
+
+    result = await SessionLoop._run_loop(
+        ctx,
+        LoopCallbacks(
+            on_error=on_error,
+            event_publish_callback=AsyncMock(),
+        ),
+    )
+
+    assert result.action == "error"
+    assert result.error == "provider failed"
+    assert result.last_message is final_assistant
+    assert (result.provider_id, result.model_id) == (
+        "fallback",
+        "fallback-model",
+    )
+    assert attempts == [
+        ("primary", "primary-model"),
+        ("fallback", "fallback-model"),
+    ]
+    on_error.assert_awaited_once_with("provider failed")
+    update.assert_awaited_once()
+    assert update.await_args.args[1] == "msg_fallback"
+
+
+@pytest.mark.asyncio
+async def test_observable_failure_is_finalized_without_replay(monkeypatch):
+    ctx = _ctx()
+    last_user = SimpleNamespace(id="msg_user", agent="rex")
+    monkeypatch.setattr(
+        SessionRunner,
+        "_process_step",
+        AsyncMock(return_value=_failure(assistant_id="msg_partial", safe=False)),
+    )
+    delete = AsyncMock(return_value=True)
+    update = AsyncMock()
+    monkeypatch.setattr(Message, "delete", delete)
+    monkeypatch.setattr(Message, "update", update)
+
+    result = await SessionLoop._process_step_with_failover(
+        ctx,
+        LoopCallbacks(),
+        [last_user],
+        last_user,
+    )
+
+    assert result.error == "provider failed"
+    assert ctx.provider_id == "primary"
+    delete.assert_not_awaited()
+    update.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_rate_limit_switch_sets_primary_cooldown(monkeypatch):
+    ctx = _ctx()
+    last_user = SimpleNamespace(id="msg_user", agent="rex")
+
+    async def process_step(runner, _messages, _last_user):
+        if runner.provider_id == "primary":
+            return _failure(assistant_id="msg_rate", reason="rate_limit")
+        return StepResult(action="stop", content="recovered")
+
+    monkeypatch.setattr(SessionRunner, "_process_step", process_step)
+    monkeypatch.setattr(Message, "delete", AsyncMock(return_value=True))
+
+    await SessionLoop._process_step_with_failover(
+        ctx,
+        LoopCallbacks(),
+        [last_user],
+        last_user,
+    )
+
+    cooldown = SessionLoop._auto_failover_cooldowns[ctx.session.id]
+    assert cooldown.model == RuntimeModel("fallback", "fallback-model")
+    assert cooldown.reason == "rate_limit"
+    assert SessionLoop._cooldown_candidate_index(
+        ctx.session.id,
+        ctx.model_candidates,
+    ) == 1
+
+
+@pytest.mark.asyncio
+async def test_403_quota_failure_sets_primary_cooldown(monkeypatch):
+    decision = SessionRunner.classify_failover_error({
+        "name": "APIError",
+        "data": {"message": "Quota exceeded", "statusCode": 403},
+    })
+    ctx = _ctx()
+    last_user = SimpleNamespace(id="msg_user", agent="rex")
+
+    async def process_step(runner, _messages, _last_user):
+        if runner.provider_id == "primary":
+            return _failure(
+                assistant_id="msg_quota",
+                reason=decision.reason,
+            )
+        return StepResult(action="stop", content="recovered")
+
+    monkeypatch.setattr(SessionRunner, "_process_step", process_step)
+    monkeypatch.setattr(Message, "delete", AsyncMock(return_value=True))
+
+    await SessionLoop._process_step_with_failover(
+        ctx,
+        LoopCallbacks(),
+        [last_user],
+        last_user,
+    )
+
+    cooldown = SessionLoop._auto_failover_cooldowns[ctx.session.id]
+    assert cooldown.reason == "rate_limit"
+    assert cooldown.model == RuntimeModel("fallback", "fallback-model")
+    assert cooldown.expires_at > time.monotonic() + 50
+
+
+@pytest.mark.asyncio
+async def test_chain_exhaustion_does_not_shorten_rate_limit_cooldown(monkeypatch):
+    ctx = _ctx()
+    last_user = SimpleNamespace(id="msg_user", agent="rex")
+
+    async def process_step(runner, _messages, _last_user):
+        reason = "rate_limit" if runner.provider_id == "primary" else "server_error"
+        return _failure(
+            assistant_id=f"msg_{runner.provider_id}",
+            reason=reason,
+        )
+
+    monkeypatch.setattr(SessionRunner, "_process_step", process_step)
+    monkeypatch.setattr(Message, "delete", AsyncMock(return_value=True))
+    monkeypatch.setattr(Message, "update", AsyncMock())
+
+    await SessionLoop._process_step_with_failover(
+        ctx,
+        LoopCallbacks(),
+        [last_user],
+        last_user,
+    )
+
+    cooldown = SessionLoop._auto_failover_cooldowns[ctx.session.id]
+    assert cooldown.reason == "rate_limit"
+    assert cooldown.model == RuntimeModel("fallback", "fallback-model")
+    # A 5s anti-replay window must not replace the primary's 60s cooldown.
+    assert cooldown.expires_at > time.monotonic() + 50
+
+
+@pytest.mark.asyncio
+async def test_candidate_builder_discovers_one_model_per_tier_with_stable_seed(
+    monkeypatch,
+):
+    config = SimpleNamespace()
+    model_manager = MagicMock()
+    model_manager.list_models.return_value = [
+        SimpleNamespace(provider_id="primary", id="primary-model"),
+        SimpleNamespace(provider_id="primary", id="same-a"),
+        SimpleNamespace(provider_id="primary", id="same-b"),
+        SimpleNamespace(provider_id="other-a", id="other-a-model"),
+        SimpleNamespace(provider_id="other-b", id="other-b-model"),
+        SimpleNamespace(provider_id="missing", id="missing-model"),
+    ]
+    monkeypatch.setattr(
+        "flocks.config.config.Config.get",
+        AsyncMock(return_value=config),
+    )
+    monkeypatch.setattr(
+        "flocks.provider.provider.Provider.apply_config",
+        AsyncMock(),
+    )
+    monkeypatch.setattr(
+        "flocks.provider.model_manager.get_model_manager",
+        lambda: model_manager,
+    )
+
+    async def validate(provider_id, _model_id, **_kwargs):
+        available = provider_id != "missing"
+        return available, "available" if available else "provider_not_configured"
+
+    monkeypatch.setattr(SessionLoop, "validate_runtime_model", validate)
+    primary = RuntimeModel("primary", "primary-model")
+
+    first = await SessionLoop._build_model_candidates(
+        primary,
+        route_seed="ses_auto:msg_1",
+    )
+    repeated = await SessionLoop._build_model_candidates(
+        primary,
+        route_seed="ses_auto:msg_1",
+    )
+
+    assert first == repeated
+    assert first[0] == primary
+    assert len(first) == 3
+    assert first[1].provider_id == "primary"
+    assert first[1].model_id in {"same-a", "same-b"}
+    assert first[2].provider_id in {"other-a", "other-b"}
+    assert all(candidate.provider_id != "missing" for candidate in first)
+
+    selections = {
+        tuple(await SessionLoop._build_model_candidates(
+            primary,
+            route_seed=f"ses_auto:msg_{index}",
+        ))
+        for index in range(12)
+    }
+    assert len(selections) > 1
+
+
+@pytest.mark.asyncio
+async def test_candidate_builder_keeps_active_cooldown_model_in_its_tier(
+    monkeypatch,
+):
+    config = SimpleNamespace()
+    model_manager = MagicMock()
+    model_manager.list_models.return_value = [
+        SimpleNamespace(provider_id="primary", id="primary-model"),
+        SimpleNamespace(provider_id="primary", id="same-a"),
+        SimpleNamespace(provider_id="primary", id="same-b"),
+        SimpleNamespace(provider_id="other", id="other-a"),
+        SimpleNamespace(provider_id="other", id="other-b"),
+    ]
+    monkeypatch.setattr(
+        "flocks.config.config.Config.get",
+        AsyncMock(return_value=config),
+    )
+    monkeypatch.setattr(
+        "flocks.provider.provider.Provider.apply_config",
+        AsyncMock(),
+    )
+    monkeypatch.setattr(
+        "flocks.provider.model_manager.get_model_manager",
+        lambda: model_manager,
+    )
+    monkeypatch.setattr(
+        SessionLoop,
+        "validate_runtime_model",
+        AsyncMock(return_value=(True, "available")),
+    )
+    primary = RuntimeModel("primary", "primary-model")
+    cooldown_model = RuntimeModel("other", "other-b")
+
+    candidates = await SessionLoop._build_model_candidates(
+        primary,
+        route_seed="ses_auto:new-turn",
+        preferred=cooldown_model,
+    )
+
+    assert candidates[0] == primary
+    assert candidates[1].provider_id == "primary"
+    assert candidates[2] == cooldown_model
+
+
+@pytest.mark.asyncio
+async def test_auto_configuration_only_requires_available_primary(monkeypatch):
+    monkeypatch.setattr(
+        "flocks.config.config.Config.resolve_default_llm",
+        AsyncMock(return_value={
+            "provider_id": "primary",
+            "model_id": "primary-model",
+        }),
+    )
+    monkeypatch.setattr(
+        SessionLoop,
+        "validate_runtime_model",
+        AsyncMock(return_value=(True, "available")),
+    )
+    build_candidates = AsyncMock()
+    monkeypatch.setattr(SessionLoop, "_build_model_candidates", build_candidates)
+
+    assert await SessionLoop.validate_auto_configuration() == (True, "available")
+    build_candidates.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_candidate_builder_allows_primary_only_chain(monkeypatch):
+    model_manager = MagicMock()
+    model_manager.list_models.return_value = [
+        SimpleNamespace(provider_id="primary", id="primary-model"),
+    ]
+    monkeypatch.setattr(
+        "flocks.config.config.Config.get",
+        AsyncMock(return_value=SimpleNamespace()),
+    )
+    monkeypatch.setattr(
+        "flocks.provider.provider.Provider.apply_config",
+        AsyncMock(),
+    )
+    monkeypatch.setattr(
+        "flocks.provider.model_manager.get_model_manager",
+        lambda: model_manager,
+    )
+
+    primary = RuntimeModel("primary", "primary-model")
+
+    assert await SessionLoop._build_model_candidates(
+        primary,
+        route_seed="ses_auto:msg_primary_only",
+    ) == [primary]
+
+
+def test_cooldown_is_cleared_when_primary_changes():
+    candidates = [
+        RuntimeModel("new-primary", "new-model"),
+        RuntimeModel("fallback", "fallback-model"),
+    ]
+    SessionLoop._auto_failover_cooldowns["ses_auto"] = AutoFailoverCooldown(
+        model=RuntimeModel("fallback", "fallback-model"),
+        primary=RuntimeModel("old-primary", "old-model"),
+        expires_at=float("inf"),
+        reason="rate_limit",
+    )
+
+    assert SessionLoop._cooldown_candidate_index("ses_auto", candidates) == 0
+    assert "ses_auto" not in SessionLoop._auto_failover_cooldowns
+
+
+@pytest.mark.asyncio
+async def test_synthetic_subtask_continuation_keeps_fallback(monkeypatch):
+    ctx = _ctx(index=1)
+    ctx.turn_user_id = "msg_real"
+    synthetic_user = SimpleNamespace(
+        id="msg_subtask_continue",
+        model={"providerID": "primary", "modelID": "primary-model"},
+    )
+    monkeypatch.setattr(
+        Message,
+        "parts",
+        AsyncMock(return_value=[SimpleNamespace(synthetic=True)]),
+    )
+
+    await SessionLoop._prepare_auto_turn(ctx, synthetic_user)
+
+    assert ctx.auto_failover is True
+    assert ctx.turn_user_id == "msg_real"
+    assert (ctx.provider_id, ctx.model_id) == ("fallback", "fallback-model")
+
+
+@pytest.mark.asyncio
+async def test_first_real_turn_builds_stable_chain_from_user_id(monkeypatch):
+    ctx = _ctx()
+    ctx.turn_user_id = None
+    ctx.model_candidates = [RuntimeModel("primary", "primary-model")]
+    first_user = SimpleNamespace(
+        id="msg_first",
+        model={"providerID": "primary", "modelID": "primary-model"},
+    )
+    rebuilt = [
+        RuntimeModel("primary", "primary-model"),
+        RuntimeModel("primary", "same-provider-model"),
+        RuntimeModel("other", "other-provider-model"),
+    ]
+    build = AsyncMock(return_value=rebuilt)
+    monkeypatch.setattr(SessionLoop, "_build_model_candidates", build)
+
+    await SessionLoop._prepare_auto_turn(ctx, first_user)
+
+    assert ctx.turn_user_id == "msg_first"
+    assert ctx.model_candidates == rebuilt
+    build.assert_awaited_once_with(
+        RuntimeModel("primary", "primary-model"),
+        route_seed="ses_auto:msg_first",
+        preferred=None,
+    )
+
+
+@pytest.mark.asyncio
+async def test_queued_explicit_model_disables_auto(monkeypatch):
+    ctx = _ctx(index=1)
+    ctx.turn_user_id = "msg_real"
+    queued_user = SimpleNamespace(
+        id="msg_explicit",
+        model={"providerID": "explicit", "modelID": "explicit-model"},
+    )
+    persisted = _session(
+        provider="explicit",
+        model="explicit-model",
+        model_pinned=True,
+        model_auto=False,
+    )
+    monkeypatch.setattr(Message, "parts", AsyncMock(return_value=[]))
+    monkeypatch.setattr(
+        "flocks.session.session.Session.get_by_id",
+        AsyncMock(return_value=persisted),
+    )
+
+    await SessionLoop._prepare_auto_turn(ctx, queued_user)
+
+    assert ctx.auto_failover is False
+    assert ctx.model_candidates == [RuntimeModel("explicit", "explicit-model")]
+    assert (ctx.provider_id, ctx.model_id) == ("explicit", "explicit-model")
+
+
+@pytest.mark.asyncio
+async def test_non_webui_loop_cannot_activate_persisted_auto(monkeypatch):
+    ctx = _ctx(auto=False)
+    ctx.turn_user_id = "msg_real"
+    ctx.auto_failover_allowed = False
+    queued_user = SimpleNamespace(
+        id="msg_non_webui",
+        model={"providerID": "direct", "modelID": "direct-model"},
+    )
+    monkeypatch.setattr(Message, "parts", AsyncMock(return_value=[]))
+    monkeypatch.setattr(
+        "flocks.session.session.Session.get_by_id",
+        AsyncMock(return_value=_session(model_auto=True)),
+    )
+
+    await SessionLoop._prepare_auto_turn(ctx, queued_user)
+
+    assert ctx.auto_failover is False
+    assert ctx.auto_failover_allowed is False
+    assert ctx.model_candidates == [RuntimeModel("direct", "direct-model")]
+
+
+@pytest.mark.asyncio
+async def test_queued_webui_turn_rebuilds_auto_chain(monkeypatch):
+    ctx = _ctx(auto=False)
+    ctx.turn_user_id = "msg_real"
+    ctx.auto_failover_allowed = True
+    queued_user = SimpleNamespace(
+        id="msg_auto",
+        model={"providerID": "primary", "modelID": "primary-model"},
+    )
+    rebuilt = [
+        RuntimeModel("primary", "primary-model"),
+        RuntimeModel("new-fallback", "new-fallback-model"),
+    ]
+    monkeypatch.setattr(Message, "parts", AsyncMock(return_value=[]))
+    monkeypatch.setattr(
+        "flocks.session.session.Session.get_by_id",
+        AsyncMock(return_value=_session(model_auto=True)),
+    )
+    monkeypatch.setattr(
+        "flocks.config.config.Config.resolve_default_llm",
+        AsyncMock(return_value={
+            "provider_id": "primary",
+            "model_id": "primary-model",
+        }),
+    )
+    build = AsyncMock(return_value=rebuilt)
+    monkeypatch.setattr(SessionLoop, "_build_model_candidates", build)
+
+    await SessionLoop._prepare_auto_turn(ctx, queued_user)
+
+    assert ctx.auto_failover is True
+    assert ctx.model_candidates == rebuilt
+    build.assert_awaited_once_with(
+        RuntimeModel("primary", "primary-model"),
+        route_seed="ses_auto:msg_auto",
+        preferred=None,
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("category", ["user", "entity-config", "workflow"])
+async def test_queued_webui_auto_authorizes_active_loop(category):
+    ctx = _ctx(auto=False, category=category)
+    SessionLoop._active_loops[ctx.session.id] = ctx
+    try:
+        result = await SessionLoop.run(ctx.session.id, auto_failover=True)
+    finally:
+        SessionLoop._active_loops.pop(ctx.session.id, None)
+
+    assert result.action == "queued"
+    assert ctx.auto_failover_allowed is True
+
+
+@pytest.mark.asyncio
+async def test_unsupported_session_loop_ignores_auto_authorization(
+    monkeypatch,
+):
+    task_session = _session(category="task")
+    captured_ctx = None
+
+    async def run_loop(ctx, _callbacks):
+        nonlocal captured_ctx
+        captured_ctx = ctx
+        return LoopResult(action="stop")
+
+    build_candidates = AsyncMock()
+    monkeypatch.setattr(
+        "flocks.session.session.Session.get_by_id",
+        AsyncMock(return_value=task_session),
+    )
+    monkeypatch.setattr(SessionLoop, "_build_model_candidates", build_candidates)
+    monkeypatch.setattr(SessionLoop, "_run_loop", run_loop)
+    monkeypatch.setattr(SessionLoop, "_publish_session_status", AsyncMock())
+    monkeypatch.setattr(Message, "list", AsyncMock(return_value=[]))
+    monkeypatch.setattr(
+        "flocks.session.orphan_tools.abort_orphan_running_parts",
+        AsyncMock(),
+    )
+    monkeypatch.setattr(
+        "flocks.session.session.Session.touch",
+        AsyncMock(),
+    )
+    monkeypatch.setattr("flocks.bus.bus.Bus.publish", AsyncMock())
+
+    await SessionLoop.run(
+        task_session.id,
+        provider_id="primary",
+        model_id="primary-model",
+        auto_failover=True,
+    )
+
+    assert captured_ctx is not None
+    assert captured_ctx.auto_failover is False
+    assert captured_ctx.auto_failover_allowed is False
+    assert captured_ctx.model_candidates == [
+        RuntimeModel("primary", "primary-model")
+    ]
+    build_candidates.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_active_unsupported_loop_rejects_auto_authorization():
+    ctx = _ctx(auto=False, category="task")
+    SessionLoop._active_loops[ctx.session.id] = ctx
+    try:
+        result = await SessionLoop.run(ctx.session.id, auto_failover=True)
+    finally:
+        SessionLoop._active_loops.pop(ctx.session.id, None)
+
+    assert result.action == "queued"
+    assert ctx.auto_failover_allowed is False
+
+
+@pytest.mark.asyncio
+async def test_session_delete_clears_auto_failover_cooldown(monkeypatch):
+    session = _session()
+    SessionLoop._auto_failover_cooldowns[session.id] = AutoFailoverCooldown(
+        model=RuntimeModel("fallback", "fallback-model"),
+        primary=RuntimeModel("primary", "primary-model"),
+        expires_at=float("inf"),
+        reason="rate_limit",
+    )
+    monkeypatch.setattr(Session, "get", AsyncMock(return_value=session))
+    monkeypatch.setattr(Session, "children", AsyncMock(return_value=[]))
+    monkeypatch.setattr(Session, "update", AsyncMock(return_value=session))
+    monkeypatch.setattr(Message, "clear", AsyncMock(return_value=0))
+    monkeypatch.setattr(
+        "flocks.session.callable_state.clear_session_callable_tools",
+        AsyncMock(),
+    )
+    monkeypatch.setattr("flocks.bus.bus.Bus.publish", AsyncMock())
+
+    assert await Session.delete("project", session.id) is True
+    assert session.id not in SessionLoop._auto_failover_cooldowns
diff --git a/tests/session/test_execution_mode.py b/tests/session/test_execution_mode.py
new file mode 100644
index 000000000..005042d3c
--- /dev/null
+++ b/tests/session/test_execution_mode.py
@@ -0,0 +1,482 @@
+from types import SimpleNamespace
+
+import pytest
+from fastapi import HTTPException
+from pydantic import ValidationError
+
+from flocks.server.routes.session import (
+    PromptRequest,
+    _event_text_for_execution_mode,
+    _validate_execution_mode_request,
+)
+from flocks.session.execution_mode import (
+    SessionExecutionMode,
+    execution_mode_prompt,
+    is_tool_allowed,
+    runtime_execution_mode,
+    tool_call_denial_reason,
+)
+from flocks.session.interaction_queue import InteractionQueue
+from flocks.session.plan_file import is_current_plan_path, session_plan_file
+from flocks.session.session import SessionInfo, SessionTime
+from flocks.tool.registry import (
+    Tool,
+    ToolCategory,
+    ToolContext,
+    ToolInfo,
+    ToolParameter,
+    ToolRegistry,
+    ToolResult,
+    ParameterType,
+)
+from flocks.tool.file.write import write_tool
+
+
+def test_prompt_request_defaults_to_build_and_accepts_plan() -> None:
+    default_request = PromptRequest(parts=[{"type": "text", "text": "hello"}])
+    plan_request = PromptRequest.model_validate({
+        "parts": [{"type": "text", "text": "hello"}],
+        "executionMode": "plan",
+    })
+
+    assert default_request.execution_mode == SessionExecutionMode.BUILD
+    assert plan_request.execution_mode == SessionExecutionMode.PLAN
+
+
+def test_prompt_request_rejects_removed_ask_mode() -> None:
+    with pytest.raises(ValidationError):
+        PromptRequest.model_validate({
+            "parts": [{"type": "text", "text": "hello"}],
+            "executionMode": "ask",
+        })
+
+
+def test_goal_transport_uses_build_permissions_and_slash_dispatch() -> None:
+    parts = [{"type": "text", "text": "  finish the feature  "}]
+
+    assert runtime_execution_mode("goal") == SessionExecutionMode.BUILD
+    assert _event_text_for_execution_mode(
+        parts,
+        SessionExecutionMode.GOAL,
+    ) == "/goal finish the feature"
+
+
+def test_goal_requires_text_only_objective() -> None:
+    empty = PromptRequest.model_validate({
+        "parts": [],
+        "executionMode": "goal",
+    })
+    attachment = PromptRequest.model_validate({
+        "parts": [
+            {"type": "text", "text": "inspect this"},
+            {"type": "file", "url": "file:///tmp/report.txt"},
+        ],
+        "executionMode": "goal",
+    })
+
+    with pytest.raises(HTTPException, match="non-empty text objective"):
+        _validate_execution_mode_request(empty)
+    with pytest.raises(HTTPException, match="does not support attachments"):
+        _validate_execution_mode_request(attachment)
+
+
+def test_plan_uses_read_only_permission_rules() -> None:
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "read")
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "grep")
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "question")
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "plan_exit")
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "bash")
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "edit")
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "write")
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "unknown_plugin_tool")
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "task")
+    assert is_tool_allowed(SessionExecutionMode.PLAN, "delegate_task")
+    assert not is_tool_allowed(SessionExecutionMode.PLAN, "run_slash_command")
+
+    assert is_tool_allowed(SessionExecutionMode.BUILD, "bash")
+    assert not is_tool_allowed(SessionExecutionMode.BUILD, "plan_exit")
+    assert "decision-complete implementation plan" in execution_mode_prompt("plan")
+    assert "material clarification question" in execution_mode_prompt("plan")
+    assert "call plan_exit" in execution_mode_prompt("plan")
+    assert "`explore` and `librarian`" in execution_mode_prompt("plan")
+    assert execution_mode_prompt("build") == ""
+
+
+@pytest.mark.parametrize("tool_name", ["task", "delegate_task"])
+def test_plan_delegation_only_allows_explore_and_librarian(tool_name) -> None:
+    ctx = ToolContext(session_id="session-1", message_id="message-1")
+
+    for subagent_type in ("explore", "librarian"):
+        assert tool_call_denial_reason(
+            SessionExecutionMode.PLAN,
+            tool_name,
+            {"subagent_type": subagent_type},
+            ctx,
+        ) is None
+
+    for arguments in (
+        {"subagent_type": "general"},
+        {"category": "quick"},
+        {"session_id": "child-session"},
+        {},
+    ):
+        reason = tool_call_denial_reason(
+            SessionExecutionMode.PLAN,
+            tool_name,
+            arguments,
+            ctx,
+        )
+        assert reason is not None
+        assert "explore, librarian" in reason
+
+
+def test_plan_file_is_stable_and_session_scoped(tmp_path) -> None:
+    first = SessionInfo(
+        id="session-1",
+        slug="first-plan",
+        projectID="project-1",
+        directory=str(tmp_path),
+        time=SessionTime(created=1234, updated=1234),
+    )
+    second = first.model_copy(update={
+        "slug": "second-plan",
+        "time": SessionTime(created=5678, updated=5678),
+    })
+
+    first_plan = session_plan_file(first)
+    second_plan = session_plan_file(second)
+
+    assert first_plan.path == tmp_path / ".flocks" / "plans" / "1234-first-plan.md"
+    assert first_plan.relative_path == ".flocks/plans/1234-first-plan.md"
+    assert first_plan.permission_path == ".flocks/plans/1234-first-plan.md"
+    assert session_plan_file(first) == first_plan
+    assert second_plan.path != first_plan.path
+
+
+def test_plan_file_uses_worktree_root_and_directory_relative_tool_path(
+    tmp_path,
+) -> None:
+    worktree = tmp_path / "repo"
+    directory = worktree / "packages" / "app"
+    directory.mkdir(parents=True)
+    session = SessionInfo(
+        slug="nested-plan",
+        projectID="project-1",
+        directory=str(directory),
+        time=SessionTime(created=1234, updated=1234),
+    )
+
+    plan = session_plan_file(session, worktree=str(worktree))
+
+    assert plan.path == worktree / ".flocks" / "plans" / "1234-nested-plan.md"
+    assert plan.relative_path == "../../.flocks/plans/1234-nested-plan.md"
+    assert plan.permission_path == ".flocks/plans/1234-nested-plan.md"
+
+
+def test_plan_prompt_describes_create_then_incremental_edit(tmp_path) -> None:
+    session = SessionInfo(
+        slug="prompt-plan",
+        projectID="project-1",
+        directory=str(tmp_path),
+        time=SessionTime(created=1234, updated=1234),
+    )
+    plan = session_plan_file(session)
+
+    create_prompt = execution_mode_prompt("plan", session=session)
+    assert plan.relative_path in create_prompt
+    assert "No plan file exists yet" in create_prompt
+    assert "Bash is available only for read-only exploration" in create_prompt
+
+    plan.path.parent.mkdir(parents=True)
+    plan.path.write_text("# Plan\n", encoding="utf-8")
+
+    edit_prompt = execution_mode_prompt("plan", session=session)
+    assert plan.relative_path in edit_prompt
+    assert "already exists" in edit_prompt
+    assert "update it incrementally" in edit_prompt
+
+
+def test_plan_path_guard_rejects_symlink_escape(tmp_path) -> None:
+    plan_relative = ".flocks/plans/1234-plan.md"
+    plan_path = tmp_path / plan_relative
+    external = tmp_path / "external.md"
+    external.write_text("outside\n", encoding="utf-8")
+    plan_path.parent.mkdir(parents=True)
+    plan_path.symlink_to(external)
+    ctx = ToolContext(
+        session_id="session-1",
+        message_id="message-1",
+        extra={
+            "execution_mode": "plan",
+            "workspace_dir": str(tmp_path),
+            "plan_file_path": str(plan_path),
+            "plan_relative_path": plan_relative,
+            "plan_permission_path": plan_relative,
+        },
+    )
+
+    assert not is_current_plan_path(ctx, plan_relative)
+
+
+@pytest.mark.asyncio
+async def test_prompt_queue_preserves_execution_mode() -> None:
+    session_id = "execution-mode-queue"
+    await InteractionQueue.clear(session_id)
+
+    item = await InteractionQueue.enqueue(
+        session_id,
+        parts=[{"type": "text", "text": "plan this"}],
+        execution_mode=SessionExecutionMode.PLAN,
+    )
+
+    queued = await InteractionQueue.list(session_id)
+    assert item.executionMode == SessionExecutionMode.PLAN
+    assert queued[0].executionMode == SessionExecutionMode.PLAN
+
+    await InteractionQueue.clear(session_id)
+
+
+@pytest.mark.asyncio
+async def test_registry_scopes_plan_delegation_before_handler(monkeypatch) -> None:
+    calls: list[dict] = []
+
+    async def handler(_ctx, **_kwargs):
+        calls.append(_kwargs)
+        return ToolResult(success=True, output="ok")
+
+    tool = Tool(
+        info=ToolInfo(
+            name="task",
+            description="Delegation test tool",
+            category=ToolCategory.FILE,
+        ),
+        handler=handler,
+    )
+    monkeypatch.setattr(
+        ToolRegistry,
+        "get",
+        classmethod(lambda _cls, _name: tool),
+    )
+
+    explore = await ToolRegistry.execute(
+        "task",
+        ctx=ToolContext(
+            session_id="session-1",
+            message_id="message-1",
+            extra={"execution_mode": "plan"},
+        ),
+        subagent_type="explore",
+    )
+    denied = await ToolRegistry.execute(
+        "delegate_task",
+        ctx=ToolContext(
+            session_id="session-1",
+            message_id="message-1",
+            extra={"execution_mode": "plan"},
+        ),
+        subagent_type="general",
+    )
+
+    assert explore.success
+    assert not denied.success
+    assert "explore, librarian" in (denied.error or "")
+    assert calls == [{"subagent_type": "explore"}]
+
+
+@pytest.mark.asyncio
+async def test_registry_scopes_plan_edits_to_current_plan_file(
+    monkeypatch,
+    tmp_path,
+) -> None:
+    calls: list[str] = []
+
+    async def handler(_ctx, **kwargs):
+        calls.append(kwargs["filePath"])
+        return ToolResult(success=True, output="ok")
+
+    tool = Tool(
+        info=ToolInfo(
+            name="write",
+            description="Write test",
+            category=ToolCategory.FILE,
+            parameters=[
+                ToolParameter(
+                    name="filePath",
+                    type=ParameterType.STRING,
+                    required=True,
+                )
+            ],
+        ),
+        handler=handler,
+    )
+    monkeypatch.setattr(
+        ToolRegistry,
+        "get",
+        classmethod(lambda _cls, _name: tool),
+    )
+    plan_relative = ".flocks/plans/1234-plan.md"
+    ctx = ToolContext(
+        session_id="session-1",
+        message_id="message-1",
+        extra={
+            "execution_mode": "plan",
+            "workspace_dir": str(tmp_path),
+            "plan_file_path": str(tmp_path / plan_relative),
+            "plan_relative_path": plan_relative,
+            "plan_permission_path": plan_relative,
+        },
+    )
+
+    allowed = await ToolRegistry.execute(
+        "write",
+        ctx=ctx,
+        filePath=plan_relative,
+    )
+    denied = await ToolRegistry.execute(
+        "write",
+        ctx=ctx,
+        filePath=".flocks/plans/other.md",
+    )
+    traversal = await ToolRegistry.execute(
+        "write",
+        ctx=ctx,
+        filePath=".flocks/plans/../outside.md",
+    )
+
+    assert allowed.success
+    assert not denied.success
+    assert not traversal.success
+    assert calls == [plan_relative]
+
+
+@pytest.mark.asyncio
+async def test_tool_context_rechecks_plan_edit_permission(tmp_path) -> None:
+    plan_relative = ".flocks/plans/1234-plan.md"
+    ctx = ToolContext(
+        session_id="session-1",
+        message_id="message-1",
+        extra={
+            "execution_mode": "plan",
+            "workspace_dir": str(tmp_path),
+            "plan_file_path": str(tmp_path / plan_relative),
+            "plan_relative_path": plan_relative,
+            "plan_permission_path": plan_relative,
+        },
+    )
+
+    await ctx.ask(permission="edit", patterns=[plan_relative])
+    with pytest.raises(PermissionError, match="current session plan file"):
+        await ctx.ask(permission="edit", patterns=["src/main.py"])
+
+
+@pytest.mark.asyncio
+async def test_read_only_sandbox_allows_only_plan_artifact_write(tmp_path) -> None:
+    plan_relative = ".flocks/plans/1234-plan.md"
+    plan_path = tmp_path / plan_relative
+    ctx = ToolContext(
+        session_id="session-1",
+        message_id="message-1",
+        extra={
+            "execution_mode": "plan",
+            "workspace_dir": str(tmp_path),
+            "plan_file_path": str(plan_path),
+            "plan_relative_path": plan_relative,
+            "plan_permission_path": plan_relative,
+            "sandbox": {
+                "workspace_dir": str(tmp_path),
+                "workspace_access": "ro",
+            },
+        },
+    )
+
+    allowed = await write_tool(ctx, "# Plan\n", plan_relative)
+    denied = await write_tool(ctx, "bad\n", "src/main.py")
+
+    assert allowed.success
+    assert plan_path.read_text(encoding="utf-8") == "# Plan\n"
+    assert not denied.success
+    assert not (tmp_path / "src" / "main.py").exists()
+
+
+@pytest.mark.asyncio
+async def test_runner_filters_tools_with_message_mode(monkeypatch) -> None:
+    from flocks.session.runner import SessionRunner
+
+    runner = object.__new__(SessionRunner)
+    runner.session = SimpleNamespace(id="session-1")
+    runner._step = 1
+    runner.callbacks = SimpleNamespace(event_publish_callback=None)
+    agent = SimpleNamespace(
+        tools=[
+            "read",
+            "bash",
+            "write",
+            "edit",
+            "task",
+            "delegate_task",
+            "run_slash_command",
+        ]
+    )
+
+    result = SimpleNamespace(
+        tool_infos=[
+            SimpleNamespace(name="read"),
+            SimpleNamespace(name="bash"),
+            SimpleNamespace(name="write"),
+            SimpleNamespace(name="edit"),
+            SimpleNamespace(name="task"),
+            SimpleNamespace(name="delegate_task"),
+            SimpleNamespace(name="run_slash_command"),
+        ],
+        metadata={},
+    )
+
+    async def list_tools(**_kwargs):
+        return result
+
+    monkeypatch.setattr(
+        "flocks.session.runner.list_session_callable_tool_infos",
+        list_tools,
+    )
+    monkeypatch.setattr(
+        ToolRegistry,
+        "get",
+        classmethod(
+            lambda _cls, name: (
+                SimpleNamespace(info=SimpleNamespace(name="plan_exit", enabled=True))
+                if name == "plan_exit"
+                else None
+            )
+        ),
+    )
+    messages = [
+        SimpleNamespace(
+            role="user",
+            executionMode=SessionExecutionMode.PLAN,
+        )
+    ]
+
+    tools, metadata = await runner._list_callable_tool_infos_for_turn(
+        agent,
+        messages,
+    )
+
+    assert [tool.name for tool in tools] == [
+        "read",
+        "bash",
+        "write",
+        "edit",
+        "task",
+        "delegate_task",
+        "plan_exit",
+    ]
+    assert metadata["executionMode"] == "plan"
+    assert metadata["modeAllowedToolNames"] == [
+        "bash",
+        "delegate_task",
+        "edit",
+        "plan_exit",
+        "read",
+        "task",
+        "write",
+    ]
diff --git a/tests/session/test_goal.py b/tests/session/test_goal.py
index 7e845d7f3..9a4c189ad 100644
--- a/tests/session/test_goal.py
+++ b/tests/session/test_goal.py
@@ -291,7 +291,7 @@ async def test_goal_model_judge_uses_provider_options_without_main_token_budget(
     kwargs = provider.chat.await_args.kwargs
     assert kwargs["extra_body"] == {"reasoning_split": True}
     assert kwargs["max_tokens"] == JUDGE_MAX_TOKENS
-    assert kwargs["temperature"] == 0
+    assert "temperature" not in kwargs
     assert decision.verdict == "complete"
 
 
diff --git a/tests/session/test_lifecycle_hooks.py b/tests/session/test_lifecycle_hooks.py
new file mode 100644
index 000000000..b08c3581e
--- /dev/null
+++ b/tests/session/test_lifecycle_hooks.py
@@ -0,0 +1,699 @@
+"""Focused tests for the Python lifecycle-hook seams."""
+
+from __future__ import annotations
+
+import asyncio
+from types import SimpleNamespace
+from unittest.mock import ANY, AsyncMock, MagicMock, patch
+
+import pytest
+
+from flocks.hooks.pipeline import HookContext, HookStage
+from flocks.session.goal import GoalDecision
+from flocks.session.runner import SessionRunner, StepResult
+from flocks.session.session import SessionInfo
+from flocks.session.session_loop import LoopCallbacks, LoopContext, SessionLoop
+
+
+def _session(session_id: str = "ses_lifecycle_hooks") -> SessionInfo:
+    return SessionInfo.model_construct(
+        id=session_id,
+        slug="hooks",
+        project_id="project",
+        directory="/tmp/project",
+        title="Lifecycle Hooks",
+        agent="rex",
+    )
+
+
+def _loop_context(session_id: str = "ses_lifecycle_hooks") -> LoopContext:
+    return LoopContext(
+        session=_session(session_id),
+        provider_id="test-provider",
+        model_id="test-model",
+        agent_name="rex",
+    )
+
+
+@pytest.mark.asyncio
+async def test_real_user_turn_is_detected_once_and_synthetic_is_ignored() -> None:
+    ctx = _loop_context()
+    first_user = SimpleNamespace(id="msg_user", model={})
+    synthetic_user = SimpleNamespace(id="msg_synthetic", model={})
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.parts",
+            AsyncMock(
+                side_effect=[
+                    [],
+                    [SimpleNamespace(synthetic=True)],
+                ]
+            ),
+        ),
+        patch.object(
+            SessionLoop,
+            "_run_user_prompt_submit_hook",
+            AsyncMock(),
+        ) as submit_hook,
+    ):
+        for user in (first_user, first_user, synthetic_user):
+            if await SessionLoop._prepare_auto_turn(ctx, user):
+                await SessionLoop._run_user_prompt_submit_hook(ctx, user)
+
+    assert ctx.turn_user_id == first_user.id
+    submit_hook.assert_awaited_once_with(ctx, first_user)
+
+
+@pytest.mark.asyncio
+async def test_user_prompt_submit_adds_ephemeral_turn_context() -> None:
+    ctx = _loop_context()
+    user = SimpleNamespace(id="msg_user", agent="rex")
+    run_hook = AsyncMock(
+        return_value=HookContext(
+            stage=HookStage.USER_PROMPT_SUBMIT,
+            input={},
+            output={"additionalContext": "  current sprint context  "},
+        )
+    )
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.get_text_content",
+            AsyncMock(return_value="implement hooks"),
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_user_prompt_submit",
+            run_hook,
+        ),
+    ):
+        await SessionLoop._run_user_prompt_submit_hook(ctx, user)
+
+    assert ctx.turn_additional_context == "current sprint context"
+    payload = run_hook.await_args.args[0]
+    assert payload["messageID"] == user.id
+    assert payload["prompt"] == "implement hooks"
+    assert payload["model"] == {
+        "providerID": "test-provider",
+        "modelID": "test-model",
+    }
+
+
+@pytest.mark.asyncio
+async def test_session_start_runs_only_when_pending() -> None:
+    runner = SessionRunner(
+        session=_session("ses_session_start"),
+        provider_id="test-provider",
+        model_id="test-model",
+        session_start_pending=True,
+    )
+    run_hook = AsyncMock()
+
+    with patch(
+        "flocks.session.runner.HookPipeline.run_session_start",
+        run_hook,
+    ):
+        await runner._run_session_start_hook(SimpleNamespace(name="rex"))
+        await runner._run_session_start_hook(SimpleNamespace(name="rex"))
+
+    run_hook.assert_awaited_once()
+    assert runner._session_start_fired is True
+    assert run_hook.await_args.args[0]["sessionID"] == "ses_session_start"
+
+
+@pytest.mark.asyncio
+async def test_turn_finish_block_creates_synthetic_continuation() -> None:
+    ctx = _loop_context("ses_turn_finish")
+    ctx.turn_user_id = "msg_user"
+    user = SimpleNamespace(
+        id="msg_user",
+        agent="rex",
+        model={"providerID": "test-provider", "modelID": "test-model"},
+    )
+    assistant = SimpleNamespace(
+        id="msg_assistant",
+        agent="rex",
+        finish="stop",
+    )
+    continuation = SimpleNamespace(id="msg_continuation")
+    callbacks = LoopCallbacks(event_publish_callback=AsyncMock())
+    create_message = AsyncMock(return_value=continuation)
+    run_hook = AsyncMock(
+        return_value=HookContext(
+            stage=HookStage.TURN_FINISH,
+            input={},
+            output={
+                "decision": "block",
+                "reason": "Run the test suite before finishing.",
+            },
+        )
+    )
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.get",
+            AsyncMock(return_value=user),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.get_text_content",
+            AsyncMock(side_effect=["implement hooks", "implementation complete"]),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.create",
+            create_message,
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_turn_finish",
+            run_hook,
+        ),
+        patch(
+            "flocks.agent.registry.Agent.get",
+            AsyncMock(return_value=SimpleNamespace(steps=10)),
+        ),
+    ):
+        continued = await SessionLoop._run_turn_finish_hook(
+            ctx,
+            callbacks,
+            user,
+            assistant,
+        )
+
+    assert continued is True
+    assert ctx.stop_hook_active is True
+    assert create_message.await_args.kwargs["content"] == ("Run the test suite before finishing.")
+    assert create_message.await_args.kwargs["synthetic"] is True
+    assert create_message.await_args.kwargs["part_metadata"] == {
+        "turnFinishContinuation": True,
+        "stopHookActive": True,
+        "sourceAssistantMessageID": assistant.id,
+    }
+    hook_payload = run_hook.await_args.args[0]
+    assert hook_payload["finishReason"] == "stop"
+    assert hook_payload["stopHookActive"] is False
+    callbacks.event_publish_callback.assert_awaited_once()
+    assert callbacks.event_publish_callback.await_args.args[0] == "turn.continued"
+
+
+@pytest.mark.asyncio
+async def test_queued_prompt_arriving_during_turn_finish_wins() -> None:
+    ctx = _loop_context("ses_turn_finish_queue_race")
+    ctx.turn_user_id = "msg_001"
+    user = SimpleNamespace(id="msg_001", agent="rex", role="user")
+    assistant = SimpleNamespace(
+        id="msg_002",
+        agent="rex",
+        role="assistant",
+        finish="stop",
+    )
+    queued_user = SimpleNamespace(id="msg_003", agent="rex", role="user")
+    ctx.session_ctx = SimpleNamespace(
+        get_messages=AsyncMock(
+            return_value=[user, assistant, queued_user],
+        )
+    )
+    callbacks = LoopCallbacks(event_publish_callback=AsyncMock())
+    create_message = AsyncMock()
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.get",
+            AsyncMock(return_value=user),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.get_text_content",
+            AsyncMock(side_effect=["prompt", "response"]),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.create",
+            create_message,
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_turn_finish",
+            AsyncMock(
+                return_value=HookContext(
+                    stage=HookStage.TURN_FINISH,
+                    input={},
+                    output={"decision": "block", "reason": "continue"},
+                )
+            ),
+        ),
+    ):
+        continued = await SessionLoop._run_turn_finish_hook(
+            ctx,
+            callbacks,
+            user,
+            assistant,
+        )
+
+    assert continued is True
+    create_message.assert_not_awaited()
+    callbacks.event_publish_callback.assert_awaited_once()
+    event_name, payload = callbacks.event_publish_callback.await_args.args
+    assert event_name == "turn.continued"
+    assert payload["queuedUserMessageID"] == queued_user.id
+
+
+@pytest.mark.asyncio
+async def test_turn_finish_block_is_ignored_at_agent_step_limit() -> None:
+    ctx = _loop_context("ses_turn_finish_limit")
+    ctx.turn_user_id = "msg_user"
+    ctx.trace_step_offset = 2
+    ctx.step = 1
+    user = SimpleNamespace(id="msg_user", agent="rex")
+    assistant = SimpleNamespace(
+        id="msg_assistant",
+        agent="rex",
+        finish="stop",
+    )
+    create_message = AsyncMock()
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.get",
+            AsyncMock(return_value=user),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.get_text_content",
+            AsyncMock(side_effect=["prompt", "response"]),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.create",
+            create_message,
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_turn_finish",
+            AsyncMock(
+                return_value=HookContext(
+                    stage=HookStage.TURN_FINISH,
+                    input={},
+                    output={"decision": "block", "reason": "continue"},
+                )
+            ),
+        ),
+        patch(
+            "flocks.agent.registry.Agent.get",
+            AsyncMock(return_value=SimpleNamespace(steps=3)),
+        ),
+    ):
+        continued = await SessionLoop._run_turn_finish_hook(
+            ctx,
+            LoopCallbacks(),
+            user,
+            assistant,
+        )
+
+    assert continued is False
+    create_message.assert_not_awaited()
+
+
+def _message(
+    message_id: str,
+    role: str,
+    *,
+    finish: str | None = None,
+) -> SimpleNamespace:
+    return SimpleNamespace(
+        id=message_id,
+        role=role,
+        finish=finish,
+        tokens=None,
+        summary=False,
+        agent="rex",
+        model={"providerID": "test-provider", "modelID": "test-model"},
+    )
+
+
+@pytest.mark.asyncio
+async def test_turn_finish_runs_only_after_persisted_stop() -> None:
+    ctx = _loop_context("ses_turn_finish_integration")
+    user = _message("msg_001", "user")
+    assistant = _message("msg_002", "assistant", finish="stop")
+    ctx.session_ctx = SimpleNamespace(
+        get_messages=AsyncMock(
+            side_effect=[
+                [user],
+                [user, assistant],
+            ]
+        )
+    )
+    run_turn_finish = AsyncMock(return_value=False)
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.parts",
+            AsyncMock(return_value=[]),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.get_text_content",
+            AsyncMock(return_value="final response"),
+        ),
+        patch(
+            "flocks.session.session_loop.Provider.resolve_model_info",
+            return_value=(0, 0, None),
+        ),
+        patch(
+            "flocks.session.session_loop.GoalManager.evaluate_after_turn",
+            AsyncMock(
+                return_value=GoalDecision(
+                    status="inactive",
+                    verdict="inactive",
+                )
+            ),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_user_prompt_submit_hook",
+            AsyncMock(),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_turn_finish_hook",
+            run_turn_finish,
+        ),
+        patch(
+            "flocks.session.runner.SessionRunner._process_step",
+            AsyncMock(return_value=StepResult(action="stop")),
+        ),
+        patch(
+            "flocks.session.lifecycle.title.SessionTitle.ensure_title",
+            MagicMock(return_value=None),
+        ),
+        patch(
+            "flocks.session.session_loop.fire_and_forget",
+            MagicMock(),
+        ),
+    ):
+        result = await SessionLoop._run_loop(ctx, LoopCallbacks())
+
+    assert result.action == "stop"
+    run_turn_finish.assert_awaited_once_with(
+        ctx,
+        ANY,
+        user,
+        assistant,
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("step_result", "assistant_finish"),
+    [
+        (StepResult(action="stop", error="provider failed"), "error"),
+        (StepResult(action="continue"), "tool-calls"),
+    ],
+)
+async def test_turn_finish_skips_errors_and_tool_calls(
+    step_result: StepResult,
+    assistant_finish: str,
+) -> None:
+    ctx = _loop_context(f"ses_turn_finish_{assistant_finish}")
+    user = _message("msg_001", "user")
+    assistant = _message("msg_002", "assistant", finish=assistant_finish)
+    ctx.session_ctx = SimpleNamespace(
+        get_messages=AsyncMock(
+            side_effect=[
+                [user],
+                [user, assistant],
+            ]
+        )
+    )
+    run_turn_finish = AsyncMock(return_value=False)
+
+    async def process_step(*_args, **_kwargs):
+        if step_result.action == "continue":
+            ctx.signal_abort()
+        return step_result
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.parts",
+            AsyncMock(return_value=[]),
+        ),
+        patch(
+            "flocks.session.session_loop.Provider.resolve_model_info",
+            return_value=(0, 0, None),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_user_prompt_submit_hook",
+            AsyncMock(),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_turn_finish_hook",
+            run_turn_finish,
+        ),
+        patch(
+            "flocks.session.runner.SessionRunner._process_step",
+            AsyncMock(side_effect=process_step),
+        ),
+        patch(
+            "flocks.session.lifecycle.title.SessionTitle.ensure_title",
+            MagicMock(return_value=None),
+        ),
+        patch(
+            "flocks.session.session_loop.fire_and_forget",
+            MagicMock(),
+        ),
+    ):
+        result = await SessionLoop._run_loop(ctx, LoopCallbacks())
+
+    assert result.action == "stop"
+    run_turn_finish.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_queued_user_message_takes_priority_over_turn_finish() -> None:
+    ctx = _loop_context("ses_turn_finish_queue")
+    user = _message("msg_001", "user")
+    assistant = _message("msg_002", "assistant", finish="stop")
+    queued_user = _message("msg_003", "user")
+    ctx.session_ctx = SimpleNamespace(
+        get_messages=AsyncMock(
+            side_effect=[
+                [user],
+                [user, assistant, queued_user],
+            ]
+        )
+    )
+    run_turn_finish = AsyncMock(return_value=False)
+
+    async def process_step(*_args, **_kwargs):
+        ctx.signal_abort()
+        return StepResult(action="stop")
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.parts",
+            AsyncMock(return_value=[]),
+        ),
+        patch(
+            "flocks.session.session_loop.Provider.resolve_model_info",
+            return_value=(0, 0, None),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_user_prompt_submit_hook",
+            AsyncMock(),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_turn_finish_hook",
+            run_turn_finish,
+        ),
+        patch(
+            "flocks.session.runner.SessionRunner._process_step",
+            AsyncMock(side_effect=process_step),
+        ),
+        patch(
+            "flocks.session.lifecycle.title.SessionTitle.ensure_title",
+            MagicMock(return_value=None),
+        ),
+        patch(
+            "flocks.session.session_loop.fire_and_forget",
+            MagicMock(),
+        ),
+    ):
+        await SessionLoop._run_loop(ctx, LoopCallbacks())
+
+    run_turn_finish.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_goal_continuation_takes_priority_over_turn_finish() -> None:
+    ctx = _loop_context("ses_turn_finish_goal")
+    user = _message("msg_001", "user")
+    assistant = _message("msg_002", "assistant", finish="stop")
+    goal_user = _message("msg_003", "user")
+    ctx.session_ctx = SimpleNamespace(
+        get_messages=AsyncMock(
+            side_effect=[
+                [user],
+                [user, assistant],
+            ]
+        )
+    )
+    run_turn_finish = AsyncMock(return_value=False)
+
+    async def process_step(*_args, **_kwargs):
+        ctx.signal_abort()
+        return StepResult(action="stop")
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.parts",
+            AsyncMock(return_value=[]),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.get_text_content",
+            AsyncMock(return_value="not done"),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.create",
+            AsyncMock(return_value=goal_user),
+        ),
+        patch(
+            "flocks.session.session_loop.Provider.resolve_model_info",
+            return_value=(0, 0, None),
+        ),
+        patch(
+            "flocks.session.session_loop.GoalManager.evaluate_after_turn",
+            AsyncMock(
+                return_value=GoalDecision(
+                    status="active",
+                    verdict="continue",
+                    should_continue=True,
+                    continuation_prompt="continue the goal",
+                )
+            ),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_user_prompt_submit_hook",
+            AsyncMock(),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_turn_finish_hook",
+            run_turn_finish,
+        ),
+        patch(
+            "flocks.session.runner.SessionRunner._process_step",
+            AsyncMock(side_effect=process_step),
+        ),
+        patch(
+            "flocks.session.lifecycle.title.SessionTitle.ensure_title",
+            MagicMock(return_value=None),
+        ),
+        patch(
+            "flocks.session.session_loop.fire_and_forget",
+            MagicMock(),
+        ),
+    ):
+        await SessionLoop._run_loop(ctx, LoopCallbacks())
+
+    run_turn_finish.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_abort_does_not_trigger_turn_finish() -> None:
+    ctx = _loop_context("ses_turn_finish_abort")
+    user = _message("msg_001", "user")
+    ctx.session_ctx = SimpleNamespace(get_messages=AsyncMock(return_value=[user]))
+    run_turn_finish = AsyncMock(return_value=False)
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.parts",
+            AsyncMock(return_value=[]),
+        ),
+        patch(
+            "flocks.session.session_loop.Provider.resolve_model_info",
+            return_value=(0, 0, None),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_user_prompt_submit_hook",
+            AsyncMock(),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_turn_finish_hook",
+            run_turn_finish,
+        ),
+        patch(
+            "flocks.session.runner.SessionRunner._process_step",
+            AsyncMock(side_effect=asyncio.CancelledError()),
+        ),
+        patch(
+            "flocks.session.lifecycle.title.SessionTitle.ensure_title",
+            MagicMock(return_value=None),
+        ),
+        patch(
+            "flocks.session.session_loop.fire_and_forget",
+            MagicMock(),
+        ),
+    ):
+        await SessionLoop._run_loop(ctx, LoopCallbacks())
+
+    run_turn_finish.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_late_abort_after_step_completion_skips_turn_finish() -> None:
+    ctx = _loop_context("ses_turn_finish_late_abort")
+    user = _message("msg_001", "user")
+    assistant = _message("msg_002", "assistant", finish="stop")
+    ctx.session_ctx = SimpleNamespace(
+        get_messages=AsyncMock(
+            side_effect=[
+                [user],
+                [user, assistant],
+            ]
+        )
+    )
+    run_turn_finish = AsyncMock(return_value=False)
+
+    async def abort_after_step(_step: int) -> None:
+        ctx.abort_event.set()
+
+    with (
+        patch(
+            "flocks.session.session_loop.Message.parts",
+            AsyncMock(return_value=[]),
+        ),
+        patch(
+            "flocks.session.session_loop.Message.get_text_content",
+            AsyncMock(return_value="final response"),
+        ),
+        patch(
+            "flocks.session.session_loop.Provider.resolve_model_info",
+            return_value=(0, 0, None),
+        ),
+        patch(
+            "flocks.session.session_loop.GoalManager.evaluate_after_turn",
+            AsyncMock(
+                return_value=GoalDecision(
+                    status="inactive",
+                    verdict="inactive",
+                )
+            ),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_user_prompt_submit_hook",
+            AsyncMock(),
+        ),
+        patch(
+            "flocks.session.session_loop.SessionLoop._run_turn_finish_hook",
+            run_turn_finish,
+        ),
+        patch(
+            "flocks.session.runner.SessionRunner._process_step",
+            AsyncMock(return_value=StepResult(action="stop")),
+        ),
+        patch(
+            "flocks.session.lifecycle.title.SessionTitle.ensure_title",
+            MagicMock(return_value=None),
+        ),
+        patch(
+            "flocks.session.session_loop.fire_and_forget",
+            MagicMock(),
+        ),
+    ):
+        result = await SessionLoop._run_loop(
+            ctx,
+            LoopCallbacks(on_step_end=abort_after_step),
+        )
+
+    run_turn_finish.assert_not_awaited()
+    assert result.metadata["aborted"] is True
diff --git a/tests/session/test_message_parts_persistence.py b/tests/session/test_message_parts_persistence.py
index 4cb058253..61dafc1ff 100644
--- a/tests/session/test_message_parts_persistence.py
+++ b/tests/session/test_message_parts_persistence.py
@@ -1,6 +1,7 @@
 """Persistence tests for message parts storage formats."""
 
 import asyncio
+from unittest.mock import AsyncMock
 
 import pytest
 
@@ -180,6 +181,61 @@ async def test_delete_removes_parts_using_session_storage_format() -> None:
     assert await Storage.get(f"message_parts:{per_message_session_id}") is None
 
 
+@pytest.mark.asyncio
+async def test_delete_restores_caches_when_message_persistence_fails(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    session_id = "ses_parts_delete_message_failure"
+    await Message.create(
+        session_id,
+        MessageRole.USER,
+        "keep me",
+        id="msg_a",
+        part_id="part_a",
+    )
+    persist_messages = AsyncMock(
+        side_effect=RuntimeError("message storage unavailable")
+    )
+    monkeypatch.setattr(Message, "_persist_messages", persist_messages)
+
+    with pytest.raises(RuntimeError, match="message storage unavailable"):
+        await Message.delete(session_id, "msg_a")
+
+    restored = await Message.get_with_parts_lazy(session_id, "msg_a")
+    assert restored is not None
+    assert restored.info.id == "msg_a"
+    assert [part.text for part in restored.parts] == ["keep me"]
+    stored_messages = await Storage.get(f"message:{session_id}")
+    assert [message["id"] for message in stored_messages] == ["msg_a"]
+
+
+@pytest.mark.asyncio
+async def test_delete_commits_when_parts_cleanup_fails(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    session_id = "ses_parts_delete_parts_failure"
+    await Message.create(
+        session_id,
+        MessageRole.USER,
+        "delete me",
+        id="msg_a",
+        part_id="part_a",
+    )
+    original_delete = Storage.delete
+
+    async def fail_parts_delete(key: str) -> None:
+        if key == f"message_parts:{session_id}:msg_a":
+            raise RuntimeError("parts storage unavailable")
+        await original_delete(key)
+
+    monkeypatch.setattr(Storage, "delete", fail_parts_delete)
+
+    assert await Message.delete(session_id, "msg_a") is True
+    assert await Message.get(session_id, "msg_a") is None
+    assert await Storage.get(f"message:{session_id}") == []
+    assert await Storage.get(f"message_parts:{session_id}:msg_a") is not None
+
+
 @pytest.mark.asyncio
 async def test_clear_removes_legacy_blob_and_per_message_keys() -> None:
     legacy_session_id = "ses_parts_clear_legacy"
diff --git a/tests/session/test_runner_langfuse_payloads.py b/tests/session/test_runner_langfuse_payloads.py
new file mode 100644
index 000000000..129ee8b3a
--- /dev/null
+++ b/tests/session/test_runner_langfuse_payloads.py
@@ -0,0 +1,81 @@
+from flocks.provider.provider import ChatMessage
+from flocks.session.runner import SessionRunner, ToolCall
+
+
+def test_build_langfuse_request_payload_keeps_full_messages_and_system_prompt() -> None:
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "read_file",
+                "parameters": {"type": "object"},
+            },
+        }
+    ]
+    messages = [
+        ChatMessage(role="system", content="system prompt"),
+        ChatMessage(role="user", content="user question"),
+        ChatMessage(
+            role="assistant",
+            content="calling tool",
+            tool_calls=[
+                {
+                    "id": "call_1",
+                    "type": "function",
+                    "function": {
+                        "name": "read_file",
+                        "arguments": "{\"path\":\"/tmp/demo.txt\"}",
+                    },
+                }
+            ],
+        ),
+    ]
+
+    payload = SessionRunner._build_langfuse_request_payload(
+        step=3,
+        messages=messages,
+        request_tools=tools,
+        available_tools=tools,
+        provider_options={"temperature": 0.1, "max_tokens": 1024},
+    )
+
+    assert payload["step"] == 3
+    assert payload["messages"][0]["role"] == "system"
+    assert payload["messages"][0]["content"] == "system prompt"
+    assert payload["messages"][1]["content"] == "user question"
+    assert payload["messages"][2]["tool_calls"][0]["function"]["arguments"] == "{\"path\":\"/tmp/demo.txt\"}"
+    assert payload["request_tools"] == tools
+    assert payload["available_tools"] == tools
+    assert payload["provider_options"] == {"temperature": 0.1, "max_tokens": 1024}
+
+
+def test_build_langfuse_response_payload_keeps_full_content_reasoning_and_tool_arguments() -> None:
+    full_content = "assistant output " * 80
+    full_reasoning = "reasoning output " * 60
+    tool_calls = [
+        ToolCall(
+            id="call_1",
+            name="read_file",
+            arguments={"path": "/tmp/demo.txt", "offset": 1, "limit": 1000},
+        )
+    ]
+
+    payload = SessionRunner._build_langfuse_response_payload(
+        action="continue",
+        content=full_content,
+        reasoning=full_reasoning,
+        finish_reason="tool_calls",
+        tool_calls=tool_calls,
+    )
+
+    assert payload["action"] == "continue"
+    assert payload["content"] == full_content
+    assert payload["reasoning"] == full_reasoning
+    assert payload["finish_reason"] == "tool_calls"
+    assert payload["tool_calls"] == [
+        {
+            "id": "call_1",
+            "name": "read_file",
+            "arguments": {"path": "/tmp/demo.txt", "offset": 1, "limit": 1000},
+        }
+    ]
diff --git a/tests/session/test_runner_llm_hooks.py b/tests/session/test_runner_llm_hooks.py
index de691d02f..7a3233e36 100644
--- a/tests/session/test_runner_llm_hooks.py
+++ b/tests/session/test_runner_llm_hooks.py
@@ -59,6 +59,9 @@ def get_reasoning_content(self) -> str:
     def get_finish_reason(self):
         return self.finish_reason
 
+    async def drain_parallel_tool_calls(self) -> None:
+        return None
+
 
 class _FakeToolAccumulator:
     def __init__(self, processor):
diff --git a/tests/session/test_session.py b/tests/session/test_session.py
index 84892fecb..b470ec3b5 100644
--- a/tests/session/test_session.py
+++ b/tests/session/test_session.py
@@ -7,7 +7,12 @@
 from unittest.mock import AsyncMock
 
 import pytest
-from flocks.auth.context import AuthUser, reset_current_auth_user, set_current_auth_user
+from flocks.auth.context import (
+    API_TOKEN_SERVICE_USER_ID,
+    AuthUser,
+    reset_current_auth_user,
+    set_current_auth_user,
+)
 from flocks.session.session import Session
 from flocks.session.message import Message, MessageInfo, MessageRole, TokenUsage
 from flocks.session.callable_state import add_session_callable_tools, get_session_callable_tools
@@ -29,6 +34,8 @@ async def test_session_create():
     assert session.directory == "/test/dir"
     assert session.title == "Test Session"
     assert session.status == "active"
+    assert session.owner_user_id == API_TOKEN_SERVICE_USER_ID
+    assert session.owner_username == API_TOKEN_SERVICE_USER_ID
 
 
 @pytest.mark.asyncio
diff --git a/tests/session/test_session_advanced.py b/tests/session/test_session_advanced.py
index d6dcf0bcd..f2004e487 100644
--- a/tests/session/test_session_advanced.py
+++ b/tests/session/test_session_advanced.py
@@ -11,15 +11,29 @@
 - SessionInfo model fields
 """
 
+import asyncio
+from types import SimpleNamespace
+from unittest.mock import AsyncMock
+
 import pytest
 
+from flocks.auth.context import (
+    API_TOKEN_SERVICE_USER_ID,
+    AuthUser,
+    reset_current_auth_user,
+    set_current_auth_user,
+)
+from flocks.session.message import Message, MessageRole, ToolPart, ToolStatePending
+from flocks.session.features.todo import Todo, TodoInfo
 from flocks.session.session import (
     PermissionRule,
     Session,
+    SessionInactiveError,
     SessionInfo,
     SessionRevert,
     SessionTime,
 )
+from flocks.storage.storage import Storage
 
 
 # ---------------------------------------------------------------------------
@@ -41,7 +55,9 @@ async def test_archive_sets_status(self):
         result = await Session.archive("proj_arch_1", session.id)
         assert result is True
         raw = await Session.get("proj_arch_1", session.id)
-        assert raw is None or (raw and raw.status == "archived")
+        assert raw is not None
+        assert raw.status == "archived"
+        assert raw.time.archived is not None
 
     @pytest.mark.asyncio
     async def test_archive_nonexistent_returns_false(self):
@@ -51,22 +67,540 @@ async def test_archive_nonexistent_returns_false(self):
     @pytest.mark.asyncio
     async def test_unarchive_restores_session(self):
         session = await _create(project_id="proj_arch_2")
+        await asyncio.sleep(0.01)
         await Session.archive("proj_arch_2", session.id)
+        archived = await Session.get("proj_arch_2", session.id)
+        assert archived is not None
+        assert archived.time.updated == session.time.updated
         result = await Session.unarchive("proj_arch_2", session.id)
         assert result is True
+        restored = await Session.get("proj_arch_2", session.id)
+        assert restored is not None
+        assert restored.status == "active"
+        assert restored.time.archived is None
+        assert restored.time.updated == session.time.updated
 
     @pytest.mark.asyncio
-    async def test_unarchive_nonarchived_returns_false(self):
+    async def test_unarchive_nonarchived_is_idempotent(self):
         session = await _create(project_id="proj_arch_3")
-        # Active session should not be unarchiveable
         result = await Session.unarchive("proj_arch_3", session.id)
-        assert result is False
+        assert result is True
 
     @pytest.mark.asyncio
     async def test_unarchive_nonexistent_returns_false(self):
         result = await Session.unarchive("proj_x", "ses_does_not_exist")
         assert result is False
 
+    @pytest.mark.asyncio
+    async def test_archive_and_unarchive_apply_to_descendants(self):
+        parent = await _create(project_id="proj_arch_tree", title="Parent")
+        child = await Session.fork("proj_arch_tree", parent.id)
+
+        assert await Session.archive("proj_arch_tree", parent.id) is True
+        archived_parent = await Session.get("proj_arch_tree", parent.id)
+        archived_child = await Session.get("proj_arch_tree", child.id)
+        assert archived_parent is not None and archived_parent.status == "archived"
+        assert archived_child is not None and archived_child.status == "archived"
+
+        assert await Session.unarchive("proj_arch_tree", parent.id) is True
+        restored_parent = await Session.get("proj_arch_tree", parent.id)
+        restored_child = await Session.get("proj_arch_tree", child.id)
+        assert restored_parent is not None and restored_parent.status == "active"
+        assert restored_child is not None and restored_child.status == "active"
+
+    @pytest.mark.asyncio
+    async def test_archive_and_unarchive_preserve_prompt_queues(self):
+        from flocks.session.interaction_queue import InteractionQueue
+
+        parent = await _create(project_id="proj_arch_queue", title="Parent")
+        child = await Session.create(
+            project_id=parent.project_id,
+            directory=parent.directory,
+            title="Child",
+            parent_id=parent.id,
+        )
+        queued_text = {
+            parent.id: ["1", "11", "2", "22", "3"],
+            child.id: ["child-1", "child-2"],
+        }
+
+        try:
+            before = {}
+            for queued_session_id, prompts in queued_text.items():
+                for prompt in prompts:
+                    await InteractionQueue.enqueue(
+                        queued_session_id,
+                        parts=[{"type": "text", "text": prompt}],
+                    )
+                before[queued_session_id] = [
+                    item.model_dump()
+                    for item in await InteractionQueue.list(queued_session_id)
+                ]
+
+            assert await Session.archive(parent.project_id, parent.id) is True
+            for queued_session_id in queued_text:
+                assert await InteractionQueue.pop_next(queued_session_id) is None
+                assert [
+                    item.model_dump()
+                    for item in await InteractionQueue.list(queued_session_id)
+                ] == before[queued_session_id]
+
+            assert await Session.unarchive(parent.project_id, parent.id) is True
+            for queued_session_id, prompts in queued_text.items():
+                restored = await InteractionQueue.list(queued_session_id)
+                assert [item.model_dump() for item in restored] == before[queued_session_id]
+                assert [
+                    item.parts[0]["text"]
+                    for item in restored
+                ] == prompts
+                first = await InteractionQueue.pop_next(queued_session_id)
+                assert first is not None
+                assert first.id == before[queued_session_id][0]["id"]
+                assert [
+                    item.id
+                    for item in await InteractionQueue.list(queued_session_id)
+                ] == [
+                    item["id"]
+                    for item in before[queued_session_id][1:]
+                ]
+        finally:
+            await InteractionQueue.clear(parent.id)
+            await InteractionQueue.clear(child.id)
+
+    @pytest.mark.asyncio
+    async def test_archive_is_idempotent_and_preserves_first_timestamp(self):
+        session = await _create(project_id="proj_arch_idempotent")
+
+        assert await Session.archive(session.project_id, session.id) is True
+        first = await Session.get(session.project_id, session.id)
+        assert first is not None and first.time.archived is not None
+
+        assert await Session.archive(session.project_id, session.id) is True
+        second = await Session.get(session.project_id, session.id)
+        assert second is not None
+        assert second.time.archived == first.time.archived
+
+    @pytest.mark.asyncio
+    async def test_unarchive_child_is_rejected_to_preserve_tree_state(self):
+        parent = await _create(project_id="proj_arch_child_restore", title="Parent")
+        child = await Session.create(
+            project_id=parent.project_id,
+            directory=parent.directory,
+            title="Child",
+            parent_id=parent.id,
+        )
+        assert await Session.archive(parent.project_id, parent.id) is True
+
+        assert await Session.unarchive(parent.project_id, child.id) is False
+        archived_parent = await Session.get(parent.project_id, parent.id)
+        archived_child = await Session.get(parent.project_id, child.id)
+        assert archived_parent is not None and archived_parent.status == "archived"
+        assert archived_child is not None and archived_child.status == "archived"
+
+    @pytest.mark.asyncio
+    async def test_create_child_under_archived_parent_is_rejected(self):
+        parent = await _create(project_id="proj_arch_child_create", title="Parent")
+        assert await Session.archive(parent.project_id, parent.id) is True
+
+        with pytest.raises(ValueError, match="is not active"):
+            await Session.create(
+                project_id=parent.project_id,
+                directory=parent.directory,
+                title="Late child",
+                parent_id=parent.id,
+            )
+
+    @pytest.mark.asyncio
+    async def test_child_inherits_parent_owner_when_explicit_values_are_none(self):
+        parent = await Session.create(
+            project_id="proj_arch_child_owner",
+            directory="/tmp",
+            title="Parent",
+            owner_user_id="owner-1",
+            owner_username="alice",
+        )
+
+        child = await Session.create(
+            project_id=parent.project_id,
+            directory=parent.directory,
+            title="Child",
+            parent_id=parent.id,
+            owner_user_id=None,
+            owner_username=None,
+        )
+
+        assert child.owner_user_id == parent.owner_user_id
+        assert child.owner_username == parent.owner_username
+
+    @pytest.mark.asyncio
+    async def test_child_inherits_system_parent_owner_under_admin_context(self):
+        parent = await Session.create(
+            project_id="proj_system_parent_owner",
+            directory="/tmp",
+            title="System parent",
+        )
+        admin = AuthUser(id="usr_admin", username="admin", role="admin")
+        token = set_current_auth_user(admin)
+        try:
+            child = await Session.create(
+                project_id=parent.project_id,
+                directory=parent.directory,
+                title="Child",
+                parent_id=parent.id,
+            )
+        finally:
+            reset_current_auth_user(token)
+
+        assert child.owner_user_id == API_TOKEN_SERVICE_USER_ID
+        assert child.owner_username == API_TOKEN_SERVICE_USER_ID
+
+    @pytest.mark.asyncio
+    async def test_archived_session_loop_cannot_restart(self):
+        from flocks.session.session_loop import SessionLoop
+
+        session = await _create(project_id="proj_arch_loop_guard")
+        assert await Session.archive(session.project_id, session.id) is True
+
+        result = await SessionLoop.run(session.id)
+
+        assert result.action == "error"
+        assert "archived" in (result.error or "")
+
+    @pytest.mark.asyncio
+    async def test_lifecycle_guard_rejects_todo_write_after_archive(self):
+        session = await _create(project_id="proj_arch_todo_guard")
+        assert await Session.archive(session.project_id, session.id) is True
+
+        with pytest.raises(SessionInactiveError):
+            await Todo.update_active(
+                session.id,
+                [TodoInfo(id="late", content="must not persist")],
+            )
+
+        assert await Todo.get(session.id) == []
+
+    @pytest.mark.asyncio
+    async def test_archive_wins_before_session_loop_registration(self, monkeypatch):
+        from flocks.session.session_loop import SessionLoop
+
+        session = await _create(project_id="proj_arch_loop_race")
+
+        async def archive_during_loop_start(_session_id: str):
+            assert await Session.archive(session.project_id, session.id) is True
+            return []
+
+        monkeypatch.setattr(Message, "list", archive_during_loop_start)
+
+        result = await SessionLoop.run(
+            session.id,
+            provider_id="test-provider",
+            model_id="test-model",
+        )
+
+        assert result.action == "error"
+        assert "archived" in (result.error or "")
+        assert SessionLoop.is_running(session.id) is False
+
+    @pytest.mark.asyncio
+    async def test_archive_cannot_be_overwritten_by_an_inflight_update(self, monkeypatch):
+        session = await _create(project_id="proj_arch_update_race")
+        storage_key = f"session:{session.project_id}:{session.id}"
+        update_reached_write = asyncio.Event()
+        release_update = asyncio.Event()
+        original_set = Storage.set
+
+        async def delayed_set(key, value, *args, **kwargs):
+            if key == storage_key and getattr(value, "title", None) == "Racing update":
+                update_reached_write.set()
+                await release_update.wait()
+            return await original_set(key, value, *args, **kwargs)
+
+        monkeypatch.setattr(Storage, "set", delayed_set)
+
+        update_task = asyncio.create_task(
+            Session.update(session.project_id, session.id, title="Racing update")
+        )
+        await update_reached_write.wait()
+        archive_task = asyncio.create_task(Session.archive(session.project_id, session.id))
+        await asyncio.sleep(0)
+        assert archive_task.done() is False
+        release_update.set()
+
+        assert await update_task is not None
+        assert await archive_task is True
+        stored = await Storage.get(storage_key, SessionInfo)
+        assert stored is not None
+        assert stored.status == "archived"
+        assert stored.title == "Racing update"
+
+    @pytest.mark.asyncio
+    async def test_unrelated_session_updates_do_not_share_a_global_lock(self, monkeypatch):
+        first = await _create(project_id="proj_keyed_locks", title="First")
+        second = await _create(project_id="proj_keyed_locks", title="Second")
+        first_write_started = asyncio.Event()
+        release_first_write = asyncio.Event()
+        original_set = Storage.set
+
+        async def delayed_set(key, value, *args, **kwargs):
+            if key.endswith(f":{first.id}") and getattr(value, "title", None) == "Slow":
+                first_write_started.set()
+                await release_first_write.wait()
+            return await original_set(key, value, *args, **kwargs)
+
+        monkeypatch.setattr(Storage, "set", delayed_set)
+        first_update = asyncio.create_task(
+            Session.update(first.project_id, first.id, title="Slow")
+        )
+        await first_write_started.wait()
+
+        try:
+            second_update = await asyncio.wait_for(
+                Session.update(second.project_id, second.id, title="Fast"),
+                timeout=0.5,
+            )
+            assert second_update is not None and second_update.title == "Fast"
+        finally:
+            release_first_write.set()
+            await first_update
+
+    @pytest.mark.asyncio
+    async def test_archive_flushes_debounced_tool_parts_before_cache_invalidation(self):
+        session = await _create(project_id="proj_arch_parts_flush")
+        message = await Message.create(
+            session_id=session.id,
+            role=MessageRole.ASSISTANT,
+            content="working",
+        )
+        tool_part = ToolPart(
+            sessionID=session.id,
+            messageID=message.id,
+            callID="call-pending",
+            tool="read",
+            state=ToolStatePending(input={"path": "/tmp/a"}, raw='{"path":"/tmp/a"}'),
+        )
+        await Message.store_part(session.id, message.id, tool_part)
+
+        assert session.id in Message._parts_flush_tasks
+        assert await Session.archive(session.project_id, session.id) is True
+
+        reloaded_parts = await Message.parts(message.id, session.id)
+        reloaded_tool = next(part for part in reloaded_parts if part.id == tool_part.id)
+        assert isinstance(reloaded_tool, ToolPart)
+        assert reloaded_tool.state.status == "pending"
+
+    @pytest.mark.asyncio
+    async def test_permanent_delete_clears_prompt_queue(self):
+        from flocks.session.interaction_queue import InteractionQueue
+
+        session = await _create(project_id="proj_delete_queue")
+        await InteractionQueue.enqueue(
+            session.id,
+            parts=[{"type": "text", "text": "discard on delete"}],
+        )
+
+        assert await Session.delete(session.project_id, session.id) is True
+        assert await InteractionQueue.list(session.id) == []
+
+    @pytest.mark.asyncio
+    async def test_permanent_delete_removes_tree_data_in_one_mutation(
+        self,
+        tmp_path,
+        monkeypatch,
+    ):
+        from flocks.config.config import Config
+        from flocks.permission.next import PermissionNext, PermissionRequestInfo
+        from flocks.session.files import session_uploads_dir
+
+        monkeypatch.setattr(
+            Config,
+            "get_data_path",
+            classmethod(lambda _cls: tmp_path),
+        )
+        parent = await _create(project_id="proj_delete_atomic", title="Parent")
+        child = await Session.create(
+            project_id=parent.project_id,
+            directory=parent.directory,
+            title="Child",
+            parent_id=parent.id,
+        )
+        for session in (parent, child):
+            upload_dir = session_uploads_dir(session.id)
+            upload_dir.mkdir(parents=True)
+            (upload_dir / "attachment.txt").write_text("remove", encoding="utf-8")
+            await Storage.set(f"message:{session.id}", [{"id": "message"}], "message")
+            await Storage.set(f"message_parts:{session.id}", {"legacy": []}, "message_parts")
+            await Storage.set(f"message_parts:{session.id}:message", [], "message_parts")
+            await Storage.set(f"todo:{session.id}", [{"content": "remove"}], "todo")
+            await Storage.set(f"goal:{session.id}", {"objective": "retain until delete"}, "goal")
+            await Storage.set(f"session_diff:{session.id}", {"files": []}, "session_diff")
+            await Storage.set(f"message_diff:{session.id}:message", {"diff": "remove"}, "message_diff")
+            await Storage.set(f"system_prompts:{session.id}:default", {"text": "remove"}, "system_prompt")
+            await Storage.set(f"session_callable_tools:{session.id}", {"tools": ["read"]}, "session_callable_tools")
+            await Storage.set(
+                f"{PermissionNext._SESSION_PREFIX}{session.id}",
+                {"bash": "allow"},
+                "permission_session",
+            )
+            PermissionNext._session_permissions[session.id] = {"bash": "allow"}
+
+        pending = PermissionRequestInfo(
+            id="per_delete_tree",
+            sessionID=child.id,
+            permission="write",
+            patterns=["*"],
+        )
+        pending_future = asyncio.get_running_loop().create_future()
+        PermissionNext._pending[pending.id] = {
+            "info": pending,
+            "future": pending_future,
+        }
+        await Storage.set(
+            f"{PermissionNext._PENDING_PREFIX}{pending.id}",
+            pending.model_dump(by_alias=True),
+            "permission_pending",
+        )
+        await Storage.set(
+            f"{PermissionNext._REPLY_PREFIX}{pending.id}",
+            {"reply": "allow", "sessionID": child.id},
+            "permission_reply",
+        )
+
+        assert await Session.delete(parent.project_id, parent.id) is True
+
+        for session in (parent, child):
+            deleted_session = await Storage.get(
+                f"session:{session.project_id}:{session.id}",
+                SessionInfo,
+            )
+            assert deleted_session is None
+            assert await Storage.get(f"message:{session.id}") is None
+            assert await Storage.get(f"message_parts:{session.id}") is None
+            assert await Storage.list_keys(prefix=f"message_parts:{session.id}:") == []
+            assert await Storage.get(f"todo:{session.id}") is None
+            assert await Storage.get(f"goal:{session.id}") is None
+            assert await Storage.get(f"session_diff:{session.id}") is None
+            assert await Storage.list_keys(prefix=f"message_diff:{session.id}:") == []
+            assert await Storage.list_keys(prefix=f"system_prompts:{session.id}:") == []
+            assert await Storage.get(f"session_callable_tools:{session.id}") is None
+            assert await Storage.get(f"{PermissionNext._SESSION_PREFIX}{session.id}") is None
+            assert session.id not in PermissionNext._session_permissions
+            assert not session_uploads_dir(session.id).exists()
+        assert await Storage.get(f"{PermissionNext._PENDING_PREFIX}{pending.id}") is None
+        assert await Storage.get(f"{PermissionNext._REPLY_PREFIX}{pending.id}") is None
+        assert pending.id not in PermissionNext._pending
+        assert pending_future.cancelled()
+
+
+class TestMoveToProject:
+    @pytest.mark.asyncio
+    async def test_move_updates_complete_tree_and_removes_source_keys(self):
+        parent = await _create(project_id="proj_move_source", title="Parent", directory="/tmp/source")
+        child = await Session.create(
+            project_id=parent.project_id,
+            directory=parent.directory,
+            title="Child",
+            parent_id=parent.id,
+        )
+
+        moved = await Session.move_to_project(
+            parent.project_id,
+            parent.id,
+            target_project_id="proj_move_target",
+            target_directory="/tmp/target",
+        )
+
+        assert moved is not None
+        assert moved.project_id == "proj_move_target"
+        assert moved.directory == "/tmp/target"
+        assert await Session.get("proj_move_source", parent.id) is None
+        assert await Session.get("proj_move_source", child.id) is None
+        moved_parent = await Session.get("proj_move_target", parent.id)
+        moved_child = await Session.get("proj_move_target", child.id)
+        assert moved_parent is not None and moved_parent.directory == "/tmp/target"
+        assert moved_child is not None and moved_child.directory == "/tmp/target"
+        assert moved_child.parent_id == moved_parent.id
+
+    @pytest.mark.asyncio
+    async def test_move_rejects_a_child_as_the_tree_root(self):
+        parent = await _create(project_id="proj_move_child_source", title="Parent")
+        child = await Session.create(
+            project_id=parent.project_id,
+            directory=parent.directory,
+            title="Child",
+            parent_id=parent.id,
+        )
+
+        moved = await Session.move_to_project(
+            parent.project_id,
+            child.id,
+            target_project_id="proj_move_child_target",
+            target_directory="/tmp/target",
+        )
+
+        assert moved is None
+        assert await Session.get(parent.project_id, parent.id) is not None
+        assert await Session.get(parent.project_id, child.id) is not None
+
+    @pytest.mark.asyncio
+    async def test_move_records_replay_boundary(self, monkeypatch):
+        parent = await _create(project_id="proj_move_history_source")
+        monkeypatch.setattr(
+            Message,
+            "list",
+            AsyncMock(return_value=[SimpleNamespace(id="msg_old")]),
+        )
+
+        moved = await Session.move_to_project(
+            parent.project_id,
+            parent.id,
+            target_project_id="proj_move_history_target",
+            target_directory="/tmp/target",
+        )
+
+        assert moved is not None
+        assert moved.revert is None
+        assert moved.metadata["projectMove"] == {
+            "sourceProjectID": "proj_move_history_source",
+            "targetProjectID": "proj_move_history_target",
+            "boundaryMessageID": "msg_old",
+            "movedAt": moved.time.updated,
+        }
+
+    @pytest.mark.asyncio
+    async def test_move_rejects_active_revert(self):
+        parent = await _create(project_id="proj_move_revert_source")
+        await Session.update(
+            parent.project_id,
+            parent.id,
+            revert={"messageID": "msg_old", "snapshot": "tree_old"},
+        )
+
+        moved = await Session.move_to_project(
+            parent.project_id,
+            parent.id,
+            target_project_id="proj_move_revert_target",
+            target_directory="/tmp/target",
+        )
+
+        assert moved is None
+        original = await Session.get(parent.project_id, parent.id)
+        assert original is not None and original.revert is not None
+
+    @pytest.mark.asyncio
+    async def test_move_rejects_active_synchronous_operation(self):
+        parent = await _create(project_id="proj_move_active_source")
+
+        async with Session.active_operation(parent.id):
+            moved = await Session.move_to_project(
+                parent.project_id,
+                parent.id,
+                target_project_id="proj_move_active_target",
+                target_directory="/tmp/target",
+            )
+
+        assert moved is None
+        assert await Session.get(parent.project_id, parent.id) is not None
+
 
 # ---------------------------------------------------------------------------
 # Fork / Children
@@ -81,6 +615,94 @@ async def test_fork_creates_child_session(self):
         assert child.parent_id == parent.id
         assert child.project_id == parent.project_id
 
+    @pytest.mark.asyncio
+    async def test_fork_preserves_project_move_replay_boundary(self):
+        from flocks.session.lifecycle.revert import SessionRevert as LifecycleSessionRevert
+
+        parent = await _create(project_id="proj_fork_move_source")
+        await Message.create(
+            parent.id,
+            MessageRole.USER,
+            "First message before moving",
+        )
+        boundary_message = await Message.create(
+            parent.id,
+            MessageRole.USER,
+            "Last message before moving",
+        )
+        moved = await Session.move_to_project(
+            parent.project_id,
+            parent.id,
+            target_project_id="proj_fork_move_target",
+            target_directory="/tmp/fork-target",
+        )
+        assert moved is not None
+        await Message.create(moved.id, MessageRole.USER, "After moving")
+
+        child = await Session.fork(moved.project_id, moved.id)
+        child_messages = await Message.list(child.id, include_archived=True)
+
+        assert child.metadata["projectMove"]["boundaryMessageID"] == child_messages[1].id
+        with pytest.raises(ValueError, match="移动项目前"):
+            await LifecycleSessionRevert.ensure_replayable(child, child_messages[0].id)
+        with pytest.raises(ValueError, match="移动项目前"):
+            await LifecycleSessionRevert.ensure_replayable(child, child_messages[1].id)
+        await LifecycleSessionRevert.ensure_replayable(child, child_messages[2].id)
+
+        prefix_child = await Session.fork(
+            moved.project_id,
+            moved.id,
+            message_id=boundary_message.id,
+        )
+        prefix_messages = await Message.list(prefix_child.id, include_archived=True)
+        assert len(prefix_messages) == 1
+        assert (
+            prefix_child.metadata["projectMove"]["boundaryMessageID"]
+            == prefix_messages[0].id
+        )
+        with pytest.raises(ValueError, match="移动项目前"):
+            await LifecycleSessionRevert.ensure_replayable(
+                prefix_child,
+                prefix_messages[0].id,
+            )
+
+    @pytest.mark.asyncio
+    async def test_concurrent_move_waits_for_complete_fork(self, monkeypatch):
+        parent = await _create(project_id="proj_fork_race_source")
+        await Message.create(parent.id, MessageRole.USER, "Copy me")
+        copy_started = asyncio.Event()
+        allow_copy = asyncio.Event()
+        original_get_text_content = Message.get_text_content
+
+        async def blocking_get_text_content(message):
+            copy_started.set()
+            await allow_copy.wait()
+            return await original_get_text_content(message)
+
+        monkeypatch.setattr(Message, "get_text_content", blocking_get_text_content)
+
+        fork_task = asyncio.create_task(Session.fork(parent.project_id, parent.id))
+        await asyncio.wait_for(copy_started.wait(), timeout=1)
+        move_task = asyncio.create_task(Session.move_to_project(
+            parent.project_id,
+            parent.id,
+            target_project_id="proj_fork_race_target",
+            target_directory="/tmp/fork-race-target",
+        ))
+        await asyncio.sleep(0)
+        assert not move_task.done()
+
+        allow_copy.set()
+        child = await asyncio.wait_for(fork_task, timeout=1)
+        moved = await asyncio.wait_for(move_task, timeout=1)
+
+        assert moved is not None
+        assert await Session.get(parent.project_id, parent.id) is None
+        assert await Session.get(parent.project_id, child.id) is None
+        moved_child = await Session.get("proj_fork_race_target", child.id)
+        assert moved_child is not None
+        assert moved_child.parent_id == parent.id
+
     @pytest.mark.asyncio
     async def test_fork_nonexistent_raises_or_returns_none(self):
         # fork() may raise ValueError or return None for nonexistent sessions
@@ -134,6 +756,24 @@ async def test_set_revert_persisted(self):
         assert updated.revert is not None
         assert updated.revert.message_id == "msg_003"
 
+    @pytest.mark.asyncio
+    async def test_revert_rejects_part_from_a_different_message(self):
+        from flocks.session.lifecycle.revert import SessionRevert as LifecycleSessionRevert
+
+        session = await _create(project_id="proj_revert_part_binding")
+        first_message = await Message.create(session.id, MessageRole.USER, "First")
+        second_message = await Message.create(session.id, MessageRole.USER, "Second")
+        first_parts = await Message.parts(first_message.id, session.id)
+
+        with pytest.raises(ValueError, match="does not belong"):
+            await LifecycleSessionRevert.revert(
+                session.id,
+                second_message.id,
+                first_parts[0].id,
+            )
+
+        assert not Session.has_active_operations(session.id)
+
 
 # ---------------------------------------------------------------------------
 # set_current / get_current
diff --git a/tests/session/test_session_policy.py b/tests/session/test_session_policy.py
index eaa698c5d..3281eace9 100644
--- a/tests/session/test_session_policy.py
+++ b/tests/session/test_session_policy.py
@@ -2,7 +2,7 @@
 
 from __future__ import annotations
 
-from flocks.auth.context import AuthUser
+from flocks.auth.context import API_TOKEN_SERVICE_USER_ID, AuthUser
 from flocks.session.session import SessionInfo
 from flocks.session.policy import SessionPolicy
 
@@ -57,6 +57,22 @@ def test_can_delete_requires_owner_only():
     assert SessionPolicy.can_delete(session, stranger) is False
 
 
+def test_service_owned_session_cannot_be_claimed_by_matching_username():
+    session = _make_session(
+        owner_user_id=API_TOKEN_SERVICE_USER_ID,
+        owner_username=API_TOKEN_SERVICE_USER_ID,
+    )
+    colliding_member = AuthUser(
+        id="usr_real_member",
+        username=API_TOKEN_SERVICE_USER_ID,
+        role="member",
+    )
+
+    assert SessionPolicy.is_owner(session, colliding_member) is False
+    assert SessionPolicy.can_read(session, colliding_member) is False
+    assert SessionPolicy.can_write(session, colliding_member) is False
+
+
 def test_can_read_requires_owner_for_private_session():
     owner = _make_user()
     admin = _make_user(user_id="usr_admin", username="root", role="admin")
@@ -94,3 +110,20 @@ def test_ownerless_session_admin_can_manage_but_member_cannot():
     assert SessionPolicy.can_read(session, member) is False
     assert SessionPolicy.can_write(session, member) is False
     assert SessionPolicy.can_delete(session, member) is False
+
+
+def test_system_owned_session_admin_can_manage_but_member_cannot():
+    admin = _make_user(user_id="usr_admin", username="admin", role="admin")
+    member = _make_user(user_id="usr_member", username="member", role="member")
+    session = _make_session(
+        owner_user_id=API_TOKEN_SERVICE_USER_ID,
+        owner_username=API_TOKEN_SERVICE_USER_ID,
+    )
+
+    assert SessionPolicy.can_read(session, admin) is True
+    assert SessionPolicy.can_write(session, admin) is True
+    assert SessionPolicy.can_delete(session, admin) is True
+
+    assert SessionPolicy.can_read(session, member) is False
+    assert SessionPolicy.can_write(session, member) is False
+    assert SessionPolicy.can_delete(session, member) is False
diff --git a/tests/session/test_stream_processor.py b/tests/session/test_stream_processor.py
index 186de7c3f..914f16d30 100644
--- a/tests/session/test_stream_processor.py
+++ b/tests/session/test_stream_processor.py
@@ -16,6 +16,7 @@
 import pytest
 from unittest.mock import AsyncMock, MagicMock, patch
 
+from flocks.hooks.pipeline import HookContext, HookStage
 from flocks.tool.registry import ToolResult
 from flocks.session.streaming.stream_processor import StreamProcessor, ToolCallState
 from flocks.session.streaming.stream_events import (
@@ -399,6 +400,333 @@ async def test_tool_call_state_created(self):
 # ---------------------------------------------------------------------------
 
 class TestToolCallExecution:
+    @pytest.mark.asyncio
+    async def test_tool_before_can_update_input_and_tool_after_can_replace_result(self):
+        proc = _make_processor()
+        execute = AsyncMock(return_value=ToolResult(
+            success=True,
+            output="original result",
+        ))
+
+        async def update_input(payload):
+            payload["tool"]["input"] = {"command": "echo changed"}
+            return HookContext(
+                stage=HookStage.TOOL_BEFORE,
+                input=payload,
+            )
+
+        post_hook = AsyncMock(return_value=HookContext(
+            stage=HookStage.TOOL_AFTER,
+            input={},
+            output={
+                "result": {
+                    "success": True,
+                    "output": "hook result",
+                },
+            },
+        ))
+
+        with (
+            patch(
+                "flocks.session.streaming.stream_processor.Message.store_part",
+                new=AsyncMock(),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_before",
+                new=AsyncMock(side_effect=update_input),
+            ) as pre_hook,
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_after",
+                new=post_hook,
+            ),
+            patch(
+                "flocks.session.streaming.stream_processor.ToolRegistry.execute",
+                new=execute,
+            ),
+        ):
+            await proc.process_event(
+                ToolInputStartEvent(id="tc_changed", tool_name="bash")
+            )
+            await proc.process_event(ToolCallEvent(
+                tool_call_id="tc_changed",
+                tool_name="bash",
+                input={"command": "echo original"},
+            ))
+
+        pre_hook.assert_awaited_once()
+        assert execute.await_args.kwargs["command"] == "echo changed"
+        post_hook.assert_awaited_once()
+        post_payload = post_hook.await_args.args[0]
+        assert post_payload["status"] == "completed"
+        assert post_payload["tool"]["input"] == {"command": "echo changed"}
+        assert post_payload["result"]["output"] == "original result"
+        assert post_payload["durationMs"] >= 0
+        assert proc.tool_calls["tc_changed"].input == {
+            "command": "echo changed"
+        }
+        assert proc.tool_calls["tc_changed"].output == "hook result"
+
+    @pytest.mark.asyncio
+    async def test_tool_error_still_emits_tool_after(self):
+        proc = _make_processor()
+        post_hook = AsyncMock(return_value=HookContext(
+            stage=HookStage.TOOL_AFTER,
+            input={},
+        ))
+
+        with (
+            patch(
+                "flocks.session.streaming.stream_processor.Message.store_part",
+                new=AsyncMock(),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_before",
+                new=AsyncMock(return_value=HookContext(
+                    stage=HookStage.TOOL_BEFORE,
+                    input={
+                        "tool": {
+                            "name": "bash",
+                            "input": {"command": "false"},
+                            "callID": "tc_error",
+                        }
+                    },
+                )),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_after",
+                new=post_hook,
+            ),
+            patch(
+                "flocks.session.streaming.stream_processor.ToolRegistry.execute",
+                new=AsyncMock(return_value=ToolResult(
+                    success=False,
+                    error="command failed",
+                )),
+            ),
+        ):
+            await proc.process_event(
+                ToolInputStartEvent(id="tc_error", tool_name="bash")
+            )
+            await proc.process_event(ToolCallEvent(
+                tool_call_id="tc_error",
+                tool_name="bash",
+                input={"command": "false"},
+            ))
+
+        post_hook.assert_awaited_once()
+        post_payload = post_hook.await_args.args[0]
+        assert post_payload["status"] == "error"
+        assert post_payload["error"] == "command failed"
+
+    @pytest.mark.asyncio
+    async def test_tool_before_can_block_and_tool_after_reports_blocked(self):
+        proc = _make_processor()
+        execute = AsyncMock(
+            return_value=ToolResult(success=True, output="should not run")
+        )
+        post_hook = AsyncMock(return_value=HookContext(
+            stage=HookStage.TOOL_AFTER,
+            input={},
+        ))
+
+        with (
+            patch(
+                "flocks.session.streaming.stream_processor.Message.store_part",
+                new=AsyncMock(),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_before",
+                new=AsyncMock(return_value=HookContext(
+                    stage=HookStage.TOOL_BEFORE,
+                    input={
+                        "tool": {
+                            "name": "bash",
+                            "input": {"command": "rm -rf target"},
+                            "callID": "tc_blocked",
+                        }
+                    },
+                    output={
+                        "decision": "block",
+                        "reason": "Destructive command is not allowed",
+                    },
+                )),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_after",
+                new=post_hook,
+            ),
+            patch(
+                "flocks.session.streaming.stream_processor.ToolRegistry.execute",
+                new=execute,
+            ),
+        ):
+            await proc.process_event(
+                ToolInputStartEvent(id="tc_blocked", tool_name="bash")
+            )
+            await proc.process_event(ToolCallEvent(
+                tool_call_id="tc_blocked",
+                tool_name="bash",
+                input={"command": "rm -rf target"},
+            ))
+
+        execute.assert_not_awaited()
+        assert proc.tool_calls["tc_blocked"].status == "error"
+        assert proc.tool_calls["tc_blocked"].error == (
+            "Destructive command is not allowed"
+        )
+        post_payload = post_hook.await_args.args[0]
+        assert post_payload["status"] == "blocked"
+        assert post_payload["error"] == "Destructive command is not allowed"
+
+    @pytest.mark.asyncio
+    async def test_sandbox_block_still_emits_tool_after(self):
+        proc = _make_processor()
+        post_hook = AsyncMock(return_value=HookContext(
+            stage=HookStage.TOOL_AFTER,
+            input={},
+        ))
+
+        with (
+            patch(
+                "flocks.session.streaming.stream_processor.Message.store_part",
+                new=AsyncMock(),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_before",
+                new=AsyncMock(return_value=HookContext(
+                    stage=HookStage.TOOL_BEFORE,
+                    input={
+                        "tool": {
+                            "name": "bash",
+                            "input": {"command": "pwd"},
+                            "callID": "tc_sandbox",
+                        }
+                    },
+                )),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_after",
+                new=post_hook,
+            ),
+            patch.object(
+                proc,
+                "_resolve_sandbox_meta",
+                new=AsyncMock(return_value={
+                    "blocked": True,
+                    "error": "Sandbox denied tool",
+                    "extra": {},
+                }),
+            ),
+            patch(
+                "flocks.session.streaming.stream_processor.ToolRegistry.execute",
+                new=AsyncMock(),
+            ) as execute,
+        ):
+            await proc.process_event(
+                ToolInputStartEvent(id="tc_sandbox", tool_name="bash")
+            )
+            await proc.process_event(ToolCallEvent(
+                tool_call_id="tc_sandbox",
+                tool_name="bash",
+                input={"command": "pwd"},
+            ))
+
+        execute.assert_not_awaited()
+        assert post_hook.await_args.args[0]["status"] == "blocked"
+
+    @pytest.mark.asyncio
+    async def test_cancelled_tool_emits_interrupted_tool_after(self):
+        proc = _make_processor()
+        post_hook = AsyncMock(return_value=HookContext(
+            stage=HookStage.TOOL_AFTER,
+            input={},
+        ))
+
+        with (
+            patch(
+                "flocks.session.streaming.stream_processor.Message.store_part",
+                new=AsyncMock(),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_before",
+                new=AsyncMock(return_value=HookContext(
+                    stage=HookStage.TOOL_BEFORE,
+                    input={
+                        "tool": {
+                            "name": "bash",
+                            "input": {"command": "sleep 10"},
+                            "callID": "tc_interrupted",
+                        }
+                    },
+                )),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_after",
+                new=post_hook,
+            ),
+            patch(
+                "flocks.session.streaming.stream_processor.ToolRegistry.execute",
+                new=AsyncMock(side_effect=asyncio.CancelledError()),
+            ),
+        ):
+            await proc.process_event(
+                ToolInputStartEvent(id="tc_interrupted", tool_name="bash")
+            )
+            with pytest.raises(asyncio.CancelledError):
+                await proc.process_event(ToolCallEvent(
+                    tool_call_id="tc_interrupted",
+                    tool_name="bash",
+                    input={"command": "sleep 10"},
+                ))
+
+        post_hook.assert_awaited_once()
+        assert post_hook.await_args.args[0]["status"] == "interrupted"
+
+    @pytest.mark.asyncio
+    async def test_cancelled_tool_before_persists_interrupted_state(self):
+        proc = _make_processor()
+        store_part = AsyncMock()
+        post_hook = AsyncMock(return_value=HookContext(
+            stage=HookStage.TOOL_AFTER,
+            input={},
+        ))
+
+        with (
+            patch(
+                "flocks.session.streaming.stream_processor.Message.store_part",
+                new=store_part,
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_before",
+                new=AsyncMock(side_effect=asyncio.CancelledError()),
+            ),
+            patch(
+                "flocks.hooks.pipeline.HookPipeline.run_tool_after",
+                new=post_hook,
+            ),
+            patch(
+                "flocks.session.streaming.stream_processor.ToolRegistry.execute",
+                new=AsyncMock(),
+            ) as execute,
+        ):
+            await proc.process_event(
+                ToolInputStartEvent(id="tc_before_cancel", tool_name="bash")
+            )
+            with pytest.raises(asyncio.CancelledError):
+                await proc.process_event(ToolCallEvent(
+                    tool_call_id="tc_before_cancel",
+                    tool_name="bash",
+                    input={"command": "sleep 10"},
+                ))
+
+        execute.assert_not_awaited()
+        post_hook.assert_awaited_once()
+        assert post_hook.await_args.args[0]["status"] == "interrupted"
+        assert proc.tool_calls["tc_before_cancel"].status == "error"
+        final_part = store_part.await_args.args[2]
+        assert final_part.state.status == "error"
+        assert final_part.state.error == "Tool execution was interrupted"
+
     @pytest.mark.asyncio
     async def test_tool_call_executes_tool(self):
         proc = _make_processor()
@@ -659,6 +987,68 @@ async def test_tool_start_callback_called(self):
 
         callback.assert_called_once()
 
+    @pytest.mark.asyncio
+    async def test_tool_span_records_full_output(self):
+        proc = _make_processor()
+        proc._langfuse_generation = object()
+        long_output = "tool output " * 120
+        span_ctx = MagicMock()
+
+        successful_result = ToolResult(
+            success=True,
+            output=long_output,
+            title="bash",
+            metadata={},
+        )
+
+        with (
+            patch("flocks.session.streaming.stream_processor.Message.store_part", new=AsyncMock()),
+            patch("flocks.session.streaming.stream_processor.Message.update_part", new=AsyncMock()),
+            patch(
+                "flocks.session.streaming.stream_processor.ToolRegistry.execute",
+                new=AsyncMock(return_value=successful_result),
+            ),
+            patch("flocks.session.streaming.stream_processor.span_scope", return_value=span_ctx),
+        ):
+            await proc.process_event(ToolInputStartEvent(id="tc_span_output", tool_name="bash"))
+            await proc.process_event(
+                ToolCallEvent(tool_call_id="tc_span_output", tool_name="bash", input={"command": "ls -la"})
+            )
+
+        assert span_ctx.end.call_count == 1
+        assert span_ctx.end.call_args.kwargs["output"] == long_output
+
+    @pytest.mark.asyncio
+    async def test_tool_span_records_full_error(self):
+        proc = _make_processor()
+        proc._langfuse_generation = object()
+        long_error = "tool error " * 120
+        span_ctx = MagicMock()
+
+        failed_result = ToolResult(
+            success=False,
+            error=long_error,
+            metadata={},
+        )
+
+        with (
+            patch("flocks.session.streaming.stream_processor.Message.store_part", new=AsyncMock()),
+            patch("flocks.session.streaming.stream_processor.Message.update_part", new=AsyncMock()),
+            patch(
+                "flocks.session.streaming.stream_processor.ToolRegistry.execute",
+                new=AsyncMock(return_value=failed_result),
+            ),
+            patch("flocks.session.streaming.stream_processor.span_scope", return_value=span_ctx),
+        ):
+            await proc.process_event(ToolInputStartEvent(id="tc_span_error", tool_name="bash"))
+            await proc.process_event(
+                ToolCallEvent(tool_call_id="tc_span_error", tool_name="bash", input={"command": "false"})
+            )
+
+        assert span_ctx.end.call_count == 1
+        assert span_ctx.end.call_args.kwargs["output"] == long_error
+        assert span_ctx.end.call_args.kwargs["level"] == "ERROR"
+
     @pytest.mark.asyncio
     async def test_tool_error_falls_back_to_metadata_output(self):
         event_callback = AsyncMock()
diff --git a/tests/session/test_stream_timeouts.py b/tests/session/test_stream_timeouts.py
new file mode 100644
index 000000000..9aedb7b09
--- /dev/null
+++ b/tests/session/test_stream_timeouts.py
@@ -0,0 +1,111 @@
+"""Tests for adaptive LLM stream timeout resolution."""
+
+from types import SimpleNamespace
+
+from flocks.session.streaming.timeouts import resolve_llm_stream_timeouts
+
+
+def _provider(
+    *,
+    provider_id: str = "openai",
+    base_url: str = "https://api.example.com/v1",
+    provider_settings: dict | None = None,
+    model_settings: dict | None = None,
+):
+    model = SimpleNamespace(
+        id="test-model",
+        custom_settings=model_settings or {},
+    )
+    return SimpleNamespace(
+        id=provider_id,
+        _base_url=base_url,
+        _config=SimpleNamespace(
+            base_url=base_url,
+            custom_settings=provider_settings or {},
+        ),
+        get_models=lambda: [model],
+    )
+
+
+def test_cloud_provider_uses_safe_defaults(monkeypatch):
+    monkeypatch.delenv("FLOCKS_LLM_STREAM_FIRST_CHUNK_TIMEOUT_S", raising=False)
+    monkeypatch.delenv("FLOCKS_LLM_STREAM_ONGOING_CHUNK_TIMEOUT_S", raising=False)
+
+    timeouts = resolve_llm_stream_timeouts(_provider(), "test-model")
+
+    assert timeouts.first_chunk_s == 120.0
+    assert timeouts.ongoing_chunk_s == 300.0
+    assert timeouts.is_local is False
+
+
+def test_local_provider_gets_long_prefill_budget(monkeypatch):
+    monkeypatch.delenv("FLOCKS_LLM_STREAM_FIRST_CHUNK_TIMEOUT_S", raising=False)
+
+    timeouts = resolve_llm_stream_timeouts(
+        _provider(base_url="http://127.0.0.1:11434/v1"),
+        "test-model",
+    )
+
+    assert timeouts.first_chunk_s == 1800.0
+    assert timeouts.is_local is True
+
+
+def test_private_network_endpoint_is_treated_as_local(monkeypatch):
+    monkeypatch.delenv("FLOCKS_LLM_STREAM_FIRST_CHUNK_TIMEOUT_S", raising=False)
+
+    timeouts = resolve_llm_stream_timeouts(
+        _provider(base_url="http://192.168.1.20:8000/v1"),
+        "test-model",
+    )
+
+    assert timeouts.first_chunk_s == 1800.0
+    assert timeouts.is_local is True
+
+
+def test_environment_overrides_default(monkeypatch):
+    monkeypatch.setenv("FLOCKS_LLM_STREAM_FIRST_CHUNK_TIMEOUT_S", "240")
+    monkeypatch.setenv("FLOCKS_LLM_STREAM_ONGOING_CHUNK_TIMEOUT_S", "420")
+
+    timeouts = resolve_llm_stream_timeouts(_provider(), "test-model")
+
+    assert timeouts.first_chunk_s == 240.0
+    assert timeouts.ongoing_chunk_s == 420.0
+
+
+def test_provider_config_overrides_environment(monkeypatch):
+    monkeypatch.setenv("FLOCKS_LLM_STREAM_FIRST_CHUNK_TIMEOUT_S", "240")
+    provider = _provider(
+        provider_settings={"stream_first_chunk_timeout_s": 360},
+    )
+
+    timeouts = resolve_llm_stream_timeouts(provider, "test-model")
+
+    assert timeouts.first_chunk_s == 360.0
+
+
+def test_model_config_overrides_provider_config(monkeypatch):
+    monkeypatch.setenv("FLOCKS_LLM_STREAM_FIRST_CHUNK_TIMEOUT_S", "240")
+    provider = _provider(
+        provider_settings={"stream_first_chunk_timeout_s": 360},
+        model_settings={
+            "stream_first_chunk_timeout_s": 480,
+            "stream_ongoing_chunk_timeout_s": 600,
+        },
+    )
+
+    timeouts = resolve_llm_stream_timeouts(provider, "test-model")
+
+    assert timeouts.first_chunk_s == 480.0
+    assert timeouts.ongoing_chunk_s == 600.0
+
+
+def test_invalid_overrides_fall_back_to_defaults(monkeypatch):
+    monkeypatch.setenv("FLOCKS_LLM_STREAM_FIRST_CHUNK_TIMEOUT_S", "invalid")
+    provider = _provider(
+        provider_settings={"stream_first_chunk_timeout_s": -1},
+        model_settings={"stream_first_chunk_timeout_s": 0},
+    )
+
+    timeouts = resolve_llm_stream_timeouts(provider, "test-model")
+
+    assert timeouts.first_chunk_s == 120.0
diff --git a/tests/session/test_tool_accumulator.py b/tests/session/test_tool_accumulator.py
index e470253bc..0ef761f6f 100644
--- a/tests/session/test_tool_accumulator.py
+++ b/tests/session/test_tool_accumulator.py
@@ -108,18 +108,20 @@ async def test_index_to_id_mapping_without_explicit_id(self):
 
 class TestFeedChunkIncremental:
     @pytest.mark.asyncio
-    async def test_partial_json_does_not_fire(self):
+    async def test_partial_json_emits_start_without_executing_tool(self):
         acc, proc = _make_accumulator()
         with patch("flocks.session.streaming.tool_accumulator.ToolRegistry") as mock_reg:
             mock_reg.get_schema.return_value = None
 
-            # Incomplete JSON - no closing brace
             await acc.feed_chunk(_make_chunk(
-                tc_id="call_inc", name="my_tool", arguments='{"command": "ls'
+                tc_id="call_inc", name="write", arguments='{"filePath": "/tmp/f", "content": "long'
             ))
 
-        # Incomplete JSON should not fire
-        assert proc.process_event.call_count == 0
+        events = [call.args[0] for call in proc.process_event.call_args_list]
+        assert len(events) == 1
+        assert isinstance(events[0], ToolInputStartEvent)
+        assert events[0].id == "call_inc"
+        assert events[0].tool_name == "write"
 
     @pytest.mark.asyncio
     async def test_incremental_chunks_complete_and_fire(self):
@@ -128,12 +130,46 @@ async def test_incremental_chunks_complete_and_fire(self):
             mock_reg.get_schema.return_value = None
 
             await acc.feed_chunk(_make_chunk(tc_id="call_inc2", name="tool_x", arguments='{"key":'))
-            assert proc.process_event.call_count == 0
+            assert proc.process_event.call_count == 1
+            assert isinstance(proc.process_event.call_args_list[0].args[0], ToolInputStartEvent)
 
             await acc.feed_chunk(_make_chunk(tc_id="call_inc2", arguments='"value"}'))
 
-        event_types = [c.args[0].type for c in proc.process_event.call_args_list]
-        assert "tool-call" in event_types
+        events = [call.args[0] for call in proc.process_event.call_args_list]
+        assert [event.type for event in events] == ["tool-input-start", "tool-call"]
+        assert events[1].input == {"key": "value"}
+
+    @pytest.mark.asyncio
+    async def test_name_only_chunk_emits_start_before_arguments_arrive(self):
+        acc, proc = _make_accumulator()
+        with patch("flocks.session.streaming.tool_accumulator.ToolRegistry") as mock_reg:
+            mock_reg.get_schema.return_value = None
+
+            await acc.feed_chunk(_make_chunk(tc_id="call_name_first", name="edit"))
+
+        events = [call.args[0] for call in proc.process_event.call_args_list]
+        assert len(events) == 1
+        assert isinstance(events[0], ToolInputStartEvent)
+        assert events[0].tool_name == "edit"
+
+    @pytest.mark.asyncio
+    async def test_generated_id_stays_stable_when_provider_id_arrives_late(self):
+        acc, proc = _make_accumulator()
+        with patch("flocks.session.streaming.tool_accumulator.ToolRegistry") as mock_reg:
+            mock_reg.get_schema.return_value = None
+
+            await acc.feed_chunk(_make_chunk(index=3, name="write"))
+            start_event = proc.process_event.call_args_list[0].args[0]
+
+            await acc.feed_chunk(_make_chunk(
+                index=3,
+                tc_id="provider_call_id",
+                arguments='{"filePath": "/tmp/f", "content": "done"}',
+            ))
+
+        events = [call.args[0] for call in proc.process_event.call_args_list]
+        assert [event.type for event in events] == ["tool-input-start", "tool-call"]
+        assert events[1].tool_call_id == start_event.id
 
     @pytest.mark.asyncio
     async def test_completed_call_ignored_on_re_feed(self):
diff --git a/tests/storage/test_storage.py b/tests/storage/test_storage.py
index 063731c2b..36f644d2d 100644
--- a/tests/storage/test_storage.py
+++ b/tests/storage/test_storage.py
@@ -70,6 +70,85 @@ async def test_storage_with_model(storage):
     assert retrieved.value == 42
 
 
+@pytest.mark.asyncio
+async def test_storage_set_many_commits_all_entries(storage):
+    """Multiple values should be persisted by one batch operation."""
+    await storage.set_many([
+        ("batch-set:key1", {"value": 1}, "test"),
+        ("batch-set:key2", StorageTestModel(id="m2", name="Beta", value=2), "test_model"),
+    ])
+
+    assert await storage.get("batch-set:key1") == {"value": 1}
+    model = await storage.get("batch-set:key2", StorageTestModel)
+    assert model == StorageTestModel(id="m2", name="Beta", value=2)
+
+
+@pytest.mark.asyncio
+async def test_storage_set_many_rolls_back_all_entries(storage, monkeypatch):
+    """A failed entry must not leave an earlier entry committed."""
+    original_connect = Storage.connect
+
+    @asynccontextmanager
+    async def failing_connect(db_path=None):
+        async with original_connect(db_path) as connection:
+            original_execute = connection.execute
+            calls = 0
+
+            async def fail_second_insert(sql, parameters=None):
+                nonlocal calls
+                if "INSERT OR REPLACE INTO storage" in sql:
+                    calls += 1
+                    if calls == 2:
+                        raise sqlite3.OperationalError("injected batch failure")
+                return await original_execute(sql, parameters)
+
+            monkeypatch.setattr(connection, "execute", fail_second_insert)
+            yield connection
+
+    with monkeypatch.context() as patcher:
+        patcher.setattr(Storage, "connect", failing_connect)
+        with pytest.raises(sqlite3.OperationalError, match="injected batch failure"):
+            await storage.set_many([
+                ("rollback:key1", {"value": 1}, "test"),
+                ("rollback:key2", {"value": 2}, "test"),
+            ])
+
+    assert await storage.get("rollback:key1") is None
+    assert await storage.get("rollback:key2") is None
+
+
+@pytest.mark.asyncio
+async def test_storage_mutate_many_rolls_back_sets_and_deletes(storage, monkeypatch):
+    """A failed mixed mutation must preserve every pre-transaction value."""
+    await storage.set("mixed:replace", {"value": "before"}, "test")
+    await storage.set("mixed:delete", {"value": "keep"}, "test")
+    original_connect = Storage.connect
+
+    @asynccontextmanager
+    async def failing_connect(db_path=None):
+        async with original_connect(db_path) as connection:
+            original_execute = connection.execute
+
+            async def fail_delete(sql, parameters=None):
+                if "DELETE FROM storage" in sql:
+                    raise sqlite3.OperationalError("injected mixed mutation failure")
+                return await original_execute(sql, parameters)
+
+            monkeypatch.setattr(connection, "execute", fail_delete)
+            yield connection
+
+    with monkeypatch.context() as patcher:
+        patcher.setattr(Storage, "connect", failing_connect)
+        with pytest.raises(sqlite3.OperationalError, match="injected mixed mutation failure"):
+            await storage.mutate_many(
+                set_entries=[("mixed:replace", {"value": "after"}, "test")],
+                delete_keys=["mixed:delete"],
+            )
+
+    assert await storage.get("mixed:replace") == {"value": "before"}
+    assert await storage.get("mixed:delete") == {"value": "keep"}
+
+
 @pytest.mark.asyncio
 async def test_storage_delete(storage):
     """Test delete operation"""
diff --git a/tests/tool/test_plan_exit.py b/tests/tool/test_plan_exit.py
new file mode 100644
index 000000000..47e257252
--- /dev/null
+++ b/tests/tool/test_plan_exit.py
@@ -0,0 +1,173 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from flocks.session.execution_mode import SessionExecutionMode
+from flocks.session.interaction_queue import InteractionQueue
+from flocks.session.message import Message, MessageRole
+from flocks.tool.registry import ToolContext, ToolResult
+from flocks.tool.system import plan_exit
+
+
+@pytest.fixture(autouse=True)
+async def clear_queue():
+    session_id = "plan-exit-session"
+    await InteractionQueue.clear(session_id)
+    await Message.clear(session_id)
+    yield
+    await InteractionQueue.clear(session_id)
+    await Message.clear(session_id)
+
+
+def _context(
+    events: list[tuple[str, dict]],
+    tmp_path,
+    *,
+    plan_content: str = "# Approved plan\n",
+) -> ToolContext:
+    async def publish(event_type: str, properties: dict) -> None:
+        events.append((event_type, properties))
+
+    plan_path = tmp_path / ".flocks" / "plans" / "1234-plan.md"
+    plan_path.parent.mkdir(parents=True, exist_ok=True)
+    plan_path.write_text(plan_content, encoding="utf-8")
+    return ToolContext(
+        session_id="plan-exit-session",
+        message_id="message-1",
+        agent="rex",
+        extra={
+            "execution_mode": "plan",
+            "workspace_dir": str(tmp_path),
+            "model": {"providerID": "openai", "modelID": "gpt-test"},
+            "plan_file_path": str(plan_path),
+            "plan_relative_path": ".flocks/plans/1234-plan.md",
+            "plan_permission_path": ".flocks/plans/1234-plan.md",
+        },
+        event_publish_callback=publish,
+    )
+
+
+@pytest.mark.asyncio
+async def test_plan_exit_approval_continues_immediately_in_build(
+    monkeypatch,
+    tmp_path,
+) -> None:
+    events: list[tuple[str, dict]] = []
+
+    async def approve(*_args, **_kwargs):
+        return ToolResult(
+            success=True,
+            output="approved",
+            metadata={"answers": [[plan_exit.START_IMPLEMENTING]]},
+        )
+
+    monkeypatch.setattr(plan_exit, "question_tool", approve)
+
+    result = await plan_exit.plan_exit_tool(_context(events, tmp_path))
+    messages = await Message.list("plan-exit-session")
+    build_message = messages[-1]
+    parts = await Message.parts(build_message.id, "plan-exit-session")
+
+    assert result.success
+    assert result.metadata["approved"] is True
+    assert await InteractionQueue.list("plan-exit-session") == []
+    assert build_message.role == MessageRole.USER
+    assert build_message.executionMode == SessionExecutionMode.BUILD
+    assert build_message.agent == "rex"
+    assert build_message.model == {"providerID": "openai", "modelID": "gpt-test"}
+    assert parts[0].synthetic is True
+    assert ".flocks/plans/1234-plan.md" in parts[0].text
+    assert result.metadata["planPath"] == ".flocks/plans/1234-plan.md"
+    assert [event_type for event_type, _ in events] == [
+        "session.execution_mode.changed",
+    ]
+
+
+@pytest.mark.asyncio
+async def test_plan_exit_decline_stays_in_plan(monkeypatch, tmp_path) -> None:
+    async def decline(*_args, **_kwargs):
+        return ToolResult(
+            success=True,
+            output="declined",
+            metadata={"answers": [[plan_exit.CONTINUE_PLANNING]]},
+        )
+
+    monkeypatch.setattr(plan_exit, "question_tool", decline)
+
+    result = await plan_exit.plan_exit_tool(_context([], tmp_path))
+
+    assert result.success
+    assert result.metadata == {"approved": False, "executionMode": "plan"}
+    assert await InteractionQueue.list("plan-exit-session") == []
+
+
+@pytest.mark.asyncio
+async def test_plan_exit_returns_continue_planning_feedback(
+    monkeypatch,
+    tmp_path,
+) -> None:
+    async def provide_feedback(*_args, **_kwargs):
+        return ToolResult(
+            success=True,
+            output="feedback",
+            metadata={
+                "answers": [
+                    [
+                        plan_exit.CONTINUE_PLANNING,
+                        "Keep the public API unchanged.",
+                    ]
+                ]
+            },
+        )
+
+    monkeypatch.setattr(plan_exit, "question_tool", provide_feedback)
+
+    result = await plan_exit.plan_exit_tool(_context([], tmp_path))
+
+    assert result.success
+    assert result.metadata == {
+        "approved": False,
+        "executionMode": "plan",
+        "feedback": "Keep the public API unchanged.",
+    }
+    assert "Keep the public API unchanged." in result.output
+    assert await InteractionQueue.list("plan-exit-session") == []
+
+
+@pytest.mark.asyncio
+async def test_plan_exit_does_not_approve_deferred_channel_question(
+    monkeypatch,
+    tmp_path,
+) -> None:
+    async def deferred(*_args, **_kwargs):
+        return ToolResult(
+            success=True,
+            output="sent",
+            metadata={"deferred": True},
+        )
+
+    monkeypatch.setattr(plan_exit, "question_tool", deferred)
+
+    result = await plan_exit.plan_exit_tool(_context([], tmp_path))
+
+    assert result.metadata["deferred"] is True
+    assert await InteractionQueue.list("plan-exit-session") == []
+
+
+@pytest.mark.asyncio
+async def test_plan_exit_requires_non_empty_plan_file(tmp_path) -> None:
+    missing_ctx = _context([], tmp_path)
+    missing_path = missing_ctx.extra["plan_file_path"]
+    Path(missing_path).unlink()
+
+    missing = await plan_exit.plan_exit_tool(missing_ctx)
+    empty = await plan_exit.plan_exit_tool(
+        _context([], tmp_path, plan_content=" \n")
+    )
+
+    assert not missing.success
+    assert "Write the session plan file" in (missing.error or "")
+    assert not empty.success
+    assert "empty" in (empty.error or "")
diff --git a/tests/tool/test_question_channel.py b/tests/tool/test_question_channel.py
index 50d5b6fee..963335f24 100644
--- a/tests/tool/test_question_channel.py
+++ b/tests/tool/test_question_channel.py
@@ -21,6 +21,14 @@ def test_normalize_question_option_accepts_common_llm_shapes() -> None:
         "label": "Only descriptive text",
         "description": "",
     }
+    assert normalize_question_option({
+        "label": "调整计划",
+        "allowText": True,
+    }) == {
+        "label": "调整计划",
+        "description": "",
+        "allowText": True,
+    }
     assert normalize_question_option({"label": ""}) is None
 
 
diff --git a/tests/tool/test_sangfor_edr_handler.py b/tests/tool/test_sangfor_edr_handler.py
index 3eb286367..dd18988ab 100644
--- a/tests/tool/test_sangfor_edr_handler.py
+++ b/tests/tool/test_sangfor_edr_handler.py
@@ -120,8 +120,8 @@ def test_complete_manual_login_saves_state(tmp_path, monkeypatch):
     monkeypatch.setattr(handler, "_ensure_browser_daemon", lambda: None)
     monkeypatch.setattr(handler.helpers, "page_info", lambda: {"url": "https://edr.example.com/ui/#/index"})
     monkeypatch.setattr(handler, "_is_logged_in", lambda cfg: True)
-    monkeypatch.setattr(handler, "_save_auth_state", lambda cfg: {"cookies": 1})
-    monkeypatch.setattr(handler, "_save_captured_login_token", lambda: True)
+    monkeypatch.setattr(handler, "_save_browser_auth_pair", lambda cfg: {"cookies": 1})
+    monkeypatch.setattr(handler, "_probe_auth_pair", lambda cfg: {"valid": True})
 
     result = handler._complete_manual_login(cfg)
 
@@ -181,3 +181,310 @@ def test_login_token_capture_hooks_fetch_and_xhr(monkeypatch):
     assert "/launch_login.php" in scripts[0]
     assert "window.fetch" in scripts[0]
     assert "XMLHttpRequest.prototype" in scripts[0]
+
+
+def test_normalise_base_url_extracts_origin_from_page_url():
+    handler = _load_handler()
+
+    assert (
+        handler._normalise_base_url("https://edr.example.com:8443/ui/#/index")
+        == "https://edr.example.com:8443"
+    )
+    assert handler._normalise_base_url("edr.example.com/ui/login.php") == "https://edr.example.com"
+
+
+def test_auth_pair_rejects_cookie_token_mismatch(tmp_path, monkeypatch):
+    handler = _load_handler()
+    state_path = tmp_path / "auth-state.json"
+    state_path.write_text(
+        '{"cookies":[{"name":"sessionid","value":"new","domain":"edr.example.com","path":"/"}],"origins":[]}',
+        encoding="utf-8",
+    )
+    old_cookies = [{"name": "sessionid", "value": "old", "domain": "edr.example.com", "path": "/"}]
+    bundle = {
+        "token": "old-token",
+        "base_url": "https://edr.example.com",
+        "cookie_fingerprint": handler._cookie_fingerprint(old_cookies, "https://edr.example.com"),
+    }
+    monkeypatch.setattr(
+        handler,
+        "_get_secret_manager",
+        lambda: type("Secrets", (), {"get": lambda self, key: __import__("json").dumps(bundle)})(),
+    )
+
+    try:
+        handler._load_verified_auth_pair(_cfg(handler, state_path))
+    except RuntimeError as exc:
+        assert "not from the same login" in str(exc)
+    else:
+        raise AssertionError("mismatched EDR authentication must be rejected")
+
+
+def test_http_login_saves_matched_cookie_and_token(tmp_path, monkeypatch):
+    handler = _load_handler()
+    http = handler._http_login_module
+    cfg = _cfg(handler, tmp_path / "auth-state.json")
+
+    class Response:
+        def __init__(self, data=None, content=b"captcha"):
+            self._data = data or {}
+            self.content = content
+
+        def raise_for_status(self):
+            return None
+
+        def json(self):
+            return self._data
+
+    class Session:
+        def __init__(self):
+            self.cookies = http.requests.cookies.RequestsCookieJar()
+            self.cookies.set("sessionid", "cookie-value", domain="edr.example.com", path="/")
+            self.posts = []
+
+        def get(self, url, **kwargs):
+            return Response()
+
+        def post(self, url, **kwargs):
+            self.posts.append((url, kwargs.get("json")))
+            payload = kwargs.get("json") or {}
+            if payload.get("opr") == "rsakey":
+                return Response({"success": True, "key": "c7" * 64})
+            if url.endswith("/login"):
+                return Response({"success": True, "key": 7})
+            return Response({"success": True, "data": {"token": "token-value"}})
+
+    saved = {}
+    monkeypatch.setattr(http, "http_session", lambda cfg: Session())
+    monkeypatch.setattr(http, "_ocr_verify_code", lambda content: "1234")
+    monkeypatch.setattr(
+        http,
+        "_save_auth_pair",
+        lambda cfg, state, token: saved.update({"state": state, "token": token}) or {"pair_verified": True},
+    )
+
+    result = http._http_login(cfg)
+
+    assert result["status"] == "http_login_refreshed_auth_state"
+    assert saved["token"] == "token-value"
+    assert saved["state"]["cookies"][0]["value"] == "cookie-value"
+
+
+def test_dashboard_request_definitions_use_dynamic_base_inputs(tmp_path):
+    handler = _load_handler()
+    cfg = _cfg(handler, tmp_path / "auth-state.json")
+
+    definitions = handler._dashboard_requests(cfg, "token-value", days=7)
+
+    agent_path, agent_payload = definitions["agent_overview"]
+    assert agent_path.endswith("opr=get_agent_overview")
+    assert agent_payload["app_args"]["name"] == "app.web.event_center.head"
+    vulner_path, vulner_payload = definitions["vulnerability_overview"]
+    assert "s=token-value" in vulner_path
+    assert vulner_payload["uid"] == "admin"
+    assert vulner_payload["token"] == "token-value"
+
+
+def test_auth_probe_requires_http_200_and_agent_overview_data(tmp_path, monkeypatch):
+    handler = _load_handler()
+    cfg = _cfg(handler, tmp_path / "auth-state.json")
+    monkeypatch.setattr(handler, "_load_verified_auth_pair", lambda cfg: ({"cookies": []}, "token"))
+
+    class Response:
+        status_code = 200
+        headers = {}
+
+        def json(self):
+            return {"success": True, "data": {"total": 12, "online": 10}}
+
+    class Session:
+        def post(self, *args, **kwargs):
+            assert kwargs["allow_redirects"] is False
+            assert args[0].endswith("opr=get_agent_overview")
+            assert kwargs["json"]["app_args"]["options"] == {}
+            assert kwargs["json"]["opr"] == "get_agent_overview"
+            assert "date_range" in kwargs["json"]
+            return Response()
+
+    monkeypatch.setattr(handler, "_dashboard_session", lambda cfg, state: Session())
+
+    result = handler._probe_auth_pair(cfg)
+
+    assert result == {
+        "valid": True,
+        "reason": "auth_probe_succeeded",
+        "http_status": 200,
+        "agent_overview_verified": True,
+    }
+
+
+def test_auth_probe_rejects_success_without_agent_overview_data(tmp_path, monkeypatch):
+    handler = _load_handler()
+    cfg = _cfg(handler, tmp_path / "auth-state.json")
+    monkeypatch.setattr(handler, "_load_verified_auth_pair", lambda cfg: ({"cookies": []}, "token"))
+
+    class Response:
+        status_code = 200
+        headers = {}
+
+        def json(self):
+            return {"success": True, "message": "ok"}
+
+    class Session:
+        def post(self, *args, **kwargs):
+            return Response()
+
+    monkeypatch.setattr(handler, "_dashboard_session", lambda cfg, state: Session())
+
+    result = handler._probe_auth_pair(cfg)
+
+    assert result == {
+        "valid": False,
+        "reason": "auth_probe_expected_agent_data_missing",
+    }
+
+
+def test_auth_probe_rejects_login_redirect(tmp_path, monkeypatch):
+    handler = _load_handler()
+    cfg = _cfg(handler, tmp_path / "auth-state.json")
+    monkeypatch.setattr(handler, "_load_verified_auth_pair", lambda cfg: ({"cookies": []}, "token"))
+
+    class Response:
+        status_code = 302
+        headers = {"Location": "/ui/login.php"}
+
+    class Session:
+        def post(self, *args, **kwargs):
+            return Response()
+
+    monkeypatch.setattr(handler, "_dashboard_session", lambda cfg, state: Session())
+
+    result = handler._probe_auth_pair(cfg)
+
+    assert result["valid"] is False
+    assert result["reason"] == "auth_probe_redirected"
+
+
+def test_http_auth_reuses_valid_pair_without_login(tmp_path, monkeypatch):
+    handler = _load_handler()
+    http = handler._http_login_module
+    cfg = _cfg(handler, tmp_path / "auth-state.json")
+    monkeypatch.setattr(
+        http,
+        "probe_auth_pair",
+        lambda cfg: {"valid": True, "reason": "auth_probe_succeeded"},
+    )
+    monkeypatch.setattr(
+        http,
+        "_http_login",
+        lambda *args, **kwargs: (_ for _ in ()).throw(AssertionError("login must be skipped")),
+    )
+
+    result = http.ensure_http_auth_pair(cfg)
+
+    assert result["status"] == "http_auth_pair_reused"
+    assert result["login_skipped"] is True
+
+
+def test_http_auth_relogs_and_confirms_when_probe_fails(tmp_path, monkeypatch):
+    handler = _load_handler()
+    http = handler._http_login_module
+    cfg = _cfg(handler, tmp_path / "auth-state.json")
+    probes = iter(
+        [
+            {"valid": False, "reason": "auth_probe_unauthorized"},
+            {"valid": True, "reason": "auth_probe_succeeded"},
+        ]
+    )
+    monkeypatch.setattr(http, "probe_auth_pair", lambda cfg: next(probes))
+    monkeypatch.setattr(
+        http,
+        "_http_login",
+        lambda cfg, captcha_code="": {"success": True, "status": "http_login_refreshed_auth_state"},
+    )
+
+    result = http.ensure_http_auth_pair(cfg)
+
+    assert result["success"] is True
+    assert result["previous_probe"]["reason"] == "auth_probe_unauthorized"
+    assert result["probe"]["valid"] is True
+
+
+def test_dashboard_error_redacts_login_token():
+    handler = _load_handler()
+
+    error = handler._safe_error(
+        RuntimeError("401 Client Error for url: https://edr.test/launch.php?s=secret-token&opr=check"),
+        "secret-token",
+    )
+
+    assert "secret-token" not in error
+    assert "<redacted>" in error
+
+
+def test_http_login_reports_failure_phase(tmp_path, monkeypatch):
+    handler = _load_handler()
+    http = handler._http_login_module
+    cfg = _cfg(handler, tmp_path / "auth-state.json")
+
+    class Session:
+        def get(self, *args, **kwargs):
+            raise http.requests.ConnectionError("connection refused")
+
+    monkeypatch.setattr(http, "http_session", lambda cfg: Session())
+
+    result = http._http_login(cfg)
+
+    assert result["status"] == "http_login_failed"
+    assert result["phase"] == "login_page"
+    assert "phase=login_page" in result["error"]
+    assert "connection refused" in result["error"]
+
+
+def test_http_login_retries_captcha_dlogin_failure(tmp_path, monkeypatch):
+    handler = _load_handler()
+    http = handler._http_login_module
+    cfg = _cfg(handler, tmp_path / "auth-state.json")
+    cfg.max_captcha_retry = 2
+
+    class Response:
+        def __init__(self, payload):
+            self.payload = payload
+            self.content = b"captcha"
+
+        def raise_for_status(self):
+            return None
+
+        def json(self):
+            return self.payload
+
+    class Session:
+        def __init__(self):
+            self.cookies = http.requests.cookies.RequestsCookieJar()
+            self.cookies.set("sessionid", "cookie-value", domain="edr.example.com", path="/")
+            self.dlogin_count = 0
+
+        def get(self, url, **kwargs):
+            return Response({})
+
+        def post(self, url, **kwargs):
+            payload = kwargs.get("json") or {}
+            if payload.get("opr") == "rsakey":
+                return Response({"success": True, "key": "c7" * 64})
+            if url.endswith("/login"):
+                self.dlogin_count += 1
+                if self.dlogin_count == 1:
+                    return Response({"success": False, "msg": "验证码错误"})
+                return Response({"success": True, "key": 7})
+            return Response({"success": True, "data": {"token": "token-value"}})
+
+    session = Session()
+    monkeypatch.setattr(http, "http_session", lambda cfg: session)
+    monkeypatch.setattr(http, "_ocr_verify_code", lambda content: "1234")
+    monkeypatch.setattr(http, "_save_auth_pair", lambda cfg, state, token: {"pair_verified": True})
+
+    result = http._http_login(cfg)
+
+    assert result["success"] is True
+    assert result["attempt"] == 2
+    assert session.dlogin_count == 2
diff --git a/tests/tool/test_session_manage_tool.py b/tests/tool/test_session_manage_tool.py
index cae00eb23..dff530441 100644
--- a/tests/tool/test_session_manage_tool.py
+++ b/tests/tool/test_session_manage_tool.py
@@ -1,10 +1,11 @@
+from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 
 from flocks.tool.registry import ToolContext, ToolRegistry, ToolResult
 import flocks.tool.system.session_manage  # noqa: F401 - ensure tool registration
-from flocks.tool.system.session_manage import session_manage
+from flocks.tool.system.session_manage import _session_archive_impl, session_manage
 
 
 def make_ctx() -> ToolContext:
@@ -83,5 +84,62 @@ async def test_session_manage_delete_requests_confirmation():
     ctx.ask.assert_awaited_once()
     ask_kwargs = ctx.ask.await_args.kwargs
     assert ask_kwargs["permission"] == "session_manage"
-    assert ask_kwargs["metadata"] == {"action": "delete", "session_id": "ses_123"}
+    assert ask_kwargs["metadata"] == {
+        "action": "permanent_delete",
+        "session_id": "ses_123",
+        "destructive": True,
+    }
     delete_impl.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_session_manage_rejects_archiving_the_current_running_session():
+    ctx = make_ctx()
+    ctx.session_id = "ses_current"
+    session = SimpleNamespace(
+        id="ses_current",
+        project_id="project",
+        title="Current",
+        status="active",
+    )
+
+    with (
+        patch("flocks.storage.storage.Storage.list_keys", AsyncMock(return_value=["session:project:ses_current"])),
+        patch("flocks.storage.storage.Storage.get", AsyncMock(return_value=session)),
+        patch("flocks.session.session.Session.archive", AsyncMock()) as archive,
+    ):
+        result = await _session_archive_impl(ctx, "ses_current", True)
+
+    assert result.success is False
+    assert "不能在当前会话" in (result.error or "")
+    archive.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_session_manage_restore_revives_removed_project_first():
+    ctx = make_ctx()
+    ctx.session_id = "ses_operator"
+    session = SimpleNamespace(
+        id="ses_archived",
+        project_id="project_removed",
+        title="Archived",
+        status="archived",
+        owner_user_id="usr_owner",
+    )
+
+    with (
+        patch(
+            "flocks.storage.storage.Storage.list_keys",
+            AsyncMock(return_value=["session:project_removed:ses_archived"]),
+        ),
+        patch("flocks.storage.storage.Storage.get", AsyncMock(return_value=session)),
+        patch("flocks.session.session.Session.restore", AsyncMock(return_value=True)) as restore_session,
+    ):
+        result = await _session_archive_impl(ctx, session.id, False)
+
+    assert result.success is True
+    restore_session.assert_awaited_once_with(
+        session.project_id,
+        session.id,
+        project_owner_id="usr_owner",
+    )
diff --git a/tests/tool/test_subagent_hooks.py b/tests/tool/test_subagent_hooks.py
new file mode 100644
index 000000000..11b83aad0
--- /dev/null
+++ b/tests/tool/test_subagent_hooks.py
@@ -0,0 +1,261 @@
+"""Tests for paired delegate_task subagent lifecycle hooks."""
+
+from __future__ import annotations
+
+import asyncio
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, patch
+
+import pytest
+
+from flocks.tool.agent.delegate_task import _run_subagent_with_hooks
+from flocks.tool.registry import ToolContext
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("resumed", [False, True])
+async def test_subagent_hooks_wrap_child_run(resumed: bool) -> None:
+    ctx = ToolContext(
+        session_id="ses_parent",
+        message_id="msg_parent",
+        agent="rex",
+    )
+    last_message = SimpleNamespace(id="msg_child_final")
+    loop_result = SimpleNamespace(
+        action="stop",
+        error=None,
+        last_message=last_message,
+    )
+    start_hook = AsyncMock()
+    stop_hook = AsyncMock()
+
+    with (
+        patch(
+            "flocks.tool.agent.delegate_task.SessionLoop.run",
+            AsyncMock(return_value=loop_result),
+        ) as run_loop,
+        patch(
+            "flocks.tool.agent.delegate_task.Message.get_text_content",
+            AsyncMock(return_value="child summary"),
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_start",
+            start_hook,
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_stop",
+            stop_hook,
+        ),
+    ):
+        result = await _run_subagent_with_hooks(
+            ctx=ctx,
+            child_session_id="ses_child",
+            child_agent="explore",
+            workspace="/tmp/project",
+            prompt="inspect hooks",
+            description="Inspect hooks",
+            resumed=resumed,
+        )
+
+    assert result is loop_result
+    run_loop.assert_awaited_once_with(
+        "ses_child",
+        provider_id=None,
+        model_id=None,
+        callbacks=None,
+    )
+    start_payload = start_hook.await_args.args[0]
+    assert start_payload["parentSessionID"] == "ses_parent"
+    assert start_payload["childSessionID"] == "ses_child"
+    assert start_payload["agentType"] == "explore"
+    assert start_payload["resumed"] is resumed
+    stop_payload = stop_hook.await_args.args[0]
+    assert stop_payload["status"] == "completed"
+    assert stop_payload["summary"] == "child summary"
+    assert stop_payload["durationMs"] >= 0
+
+
+@pytest.mark.asyncio
+async def test_subagent_stop_reports_child_error() -> None:
+    ctx = ToolContext(
+        session_id="ses_parent",
+        message_id="msg_parent",
+        agent="rex",
+    )
+    start_hook = AsyncMock()
+    stop_hook = AsyncMock()
+
+    with (
+        patch(
+            "flocks.tool.agent.delegate_task.SessionLoop.run",
+            AsyncMock(side_effect=RuntimeError("child failed")),
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_start",
+            start_hook,
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_stop",
+            stop_hook,
+        ),
+    ):
+        with pytest.raises(RuntimeError, match="child failed"):
+            await _run_subagent_with_hooks(
+                ctx=ctx,
+                child_session_id="ses_child",
+                child_agent="explore",
+                workspace="/tmp/project",
+                prompt="inspect hooks",
+                description="Inspect hooks",
+                resumed=False,
+            )
+
+    start_hook.assert_awaited_once()
+    stop_hook.assert_awaited_once()
+    stop_payload = stop_hook.await_args.args[0]
+    assert stop_payload["status"] == "error"
+    assert stop_payload["summary"] is None
+    assert stop_payload["error"] == "child failed"
+
+
+@pytest.mark.asyncio
+async def test_subagent_stop_reports_persisted_message_error() -> None:
+    ctx = ToolContext(
+        session_id="ses_parent",
+        message_id="msg_parent",
+        agent="rex",
+    )
+    last_message = SimpleNamespace(
+        id="msg_child_error",
+        finish="error",
+        error={"message": "provider authentication failed"},
+    )
+    loop_result = SimpleNamespace(
+        action="stop",
+        error=None,
+        last_message=last_message,
+        metadata={},
+    )
+    stop_hook = AsyncMock()
+
+    with (
+        patch(
+            "flocks.tool.agent.delegate_task.SessionLoop.run",
+            AsyncMock(return_value=loop_result),
+        ),
+        patch(
+            "flocks.tool.agent.delegate_task.Message.get_text_content",
+            AsyncMock(return_value=""),
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_start",
+            AsyncMock(),
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_stop",
+            stop_hook,
+        ),
+    ):
+        await _run_subagent_with_hooks(
+            ctx=ctx,
+            child_session_id="ses_child",
+            child_agent="explore",
+            workspace="/tmp/project",
+            prompt="inspect hooks",
+            description="Inspect hooks",
+            resumed=False,
+        )
+
+    stop_hook.assert_awaited_once()
+    stop_payload = stop_hook.await_args.args[0]
+    assert stop_payload["status"] == "error"
+    assert stop_payload["summary"] == ""
+    assert stop_payload["error"] == "provider authentication failed"
+
+
+@pytest.mark.asyncio
+async def test_subagent_stop_reports_interruption() -> None:
+    ctx = ToolContext(
+        session_id="ses_parent",
+        message_id="msg_parent",
+        agent="rex",
+    )
+    stop_hook = AsyncMock()
+
+    with (
+        patch(
+            "flocks.tool.agent.delegate_task.SessionLoop.run",
+            AsyncMock(side_effect=asyncio.CancelledError()),
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_start",
+            AsyncMock(),
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_stop",
+            stop_hook,
+        ),
+    ):
+        with pytest.raises(asyncio.CancelledError):
+            await _run_subagent_with_hooks(
+                ctx=ctx,
+                child_session_id="ses_child",
+                child_agent="explore",
+                workspace="/tmp/project",
+                prompt="inspect hooks",
+                description="Inspect hooks",
+                resumed=True,
+            )
+
+    stop_hook.assert_awaited_once()
+    stop_payload = stop_hook.await_args.args[0]
+    assert stop_payload["status"] == "interrupted"
+    assert stop_payload["summary"] is None
+    assert stop_payload["error"] == "Sub-agent execution was interrupted"
+
+
+@pytest.mark.asyncio
+async def test_subagent_stop_reports_normalized_loop_abort() -> None:
+    ctx = ToolContext(
+        session_id="ses_parent",
+        message_id="msg_parent",
+        agent="rex",
+    )
+    stop_hook = AsyncMock()
+    loop_result = SimpleNamespace(
+        action="stop",
+        error=None,
+        last_message=None,
+        metadata={"aborted": True},
+    )
+
+    with (
+        patch(
+            "flocks.tool.agent.delegate_task.SessionLoop.run",
+            AsyncMock(return_value=loop_result),
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_start",
+            AsyncMock(),
+        ),
+        patch(
+            "flocks.hooks.pipeline.HookPipeline.run_subagent_stop",
+            stop_hook,
+        ),
+    ):
+        result = await _run_subagent_with_hooks(
+            ctx=ctx,
+            child_session_id="ses_child",
+            child_agent="explore",
+            workspace="/tmp/project",
+            prompt="inspect hooks",
+            description="Inspect hooks",
+            resumed=False,
+        )
+
+    assert result is loop_result
+    stop_hook.assert_awaited_once()
+    stop_payload = stop_hook.await_args.args[0]
+    assert stop_payload["status"] == "interrupted"
+    assert stop_payload["summary"] is None
+    assert stop_payload["error"] == "Sub-agent execution was interrupted"
diff --git a/tests/tool/test_tools.py b/tests/tool/test_tools.py
index 895cf0f80..dfba07484 100644
--- a/tests/tool/test_tools.py
+++ b/tests/tool/test_tools.py
@@ -17,6 +17,7 @@
 import os
 import tempfile
 import shutil
+import time
 import uuid
 from pathlib import Path
 from typing import Dict, Any, List
@@ -920,6 +921,46 @@ async def test_get_time_tool_after_init(self, tool_context):
         # Should return ISO format datetime
         assert "T" in result.output  # ISO format contains 'T'
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("output_format", "scale"),
+        [("unix", 1), ("unix_ms", 1000)],
+    )
+    async def test_get_time_tool_timestamp_formats(
+        self,
+        tool_context,
+        output_format,
+        scale,
+    ):
+        """Test Unix timestamp output in seconds and milliseconds."""
+        ToolRegistry.init()
+        before = int(time.time() * scale)
+
+        result = await ToolRegistry.execute(
+            "get_time",
+            ctx=tool_context,
+            format=output_format,
+        )
+
+        after = int(time.time() * scale)
+        assert result.success
+        assert result.output.isdigit()
+        assert before <= int(result.output) <= after
+
+    @pytest.mark.asyncio
+    async def test_get_time_tool_rejects_invalid_format(self, tool_context):
+        """Test invalid output formats return a clear error."""
+        ToolRegistry.init()
+
+        result = await ToolRegistry.execute(
+            "get_time",
+            ctx=tool_context,
+            format="invalid",
+        )
+
+        assert not result.success
+        assert result.error == "format must be one of: iso, unix, unix_ms"
+
 
 # =============================================================================
 # ToolContext Tests
diff --git a/uv.lock b/uv.lock
index 12a019c9d..5eb156e92 100644
--- a/uv.lock
+++ b/uv.lock
@@ -553,7 +553,7 @@ wheels = [
 
 [[package]]
 name = "flocks"
-version = "2026.7.23"
+version = "2026.7.29"
 source = { editable = "." }
 dependencies = [
     { name = "aiofiles" },
diff --git a/webui/src/api/provider.ts b/webui/src/api/provider.ts
index e0f3062ef..de9467005 100644
--- a/webui/src/api/provider.ts
+++ b/webui/src/api/provider.ts
@@ -249,6 +249,7 @@ export const defaultModelAPI = {
   /** Delete default model for a type */
   delete: (modelType: string) =>
     client.delete(`/api/default-model/${modelType}`),
+
 };
 
 // ==================== Usage API ====================
diff --git a/webui/src/api/session.test.ts b/webui/src/api/session.test.ts
index 3303e361e..8f10c48f2 100644
--- a/webui/src/api/session.test.ts
+++ b/webui/src/api/session.test.ts
@@ -94,4 +94,14 @@ describe('sessionApi message actions', () => {
     expect(mockDelete).toHaveBeenCalledWith('/api/session/session-1/prompt_queue/queue-1');
     expect(mockPost).toHaveBeenCalledWith('/api/session/session-1/prompt_queue/queue-2/run_now');
   });
+
+  it('calls archive and restore endpoints', async () => {
+    const { sessionApi } = await import('./session');
+
+    await sessionApi.archive('session-1');
+    await sessionApi.restore('session-1');
+
+    expect(mockPost).toHaveBeenCalledWith('/api/session/session-1/archive');
+    expect(mockPost).toHaveBeenCalledWith('/api/session/session-1/restore');
+  });
 });
diff --git a/webui/src/api/session.ts b/webui/src/api/session.ts
index a421cb219..fd7fd63cb 100644
--- a/webui/src/api/session.ts
+++ b/webui/src/api/session.ts
@@ -1,4 +1,6 @@
 import client from './client';
+import type { Session } from '@/types';
+import type { SessionExecutionMode } from '@/utils/sessionExecutionMode';
 
 export interface SessionMessagePartPayload {
   id: string;
@@ -29,6 +31,7 @@ export interface QueuedPrompt {
   status: 'pending' | 'executing' | string;
   createdAt: number;
   updatedAt: number;
+  executionMode?: SessionExecutionMode;
 }
 
 export interface PromptQueueResponse {
@@ -82,6 +85,7 @@ export interface SessionListParams {
   start?: number;
   search?: string;
   category?: string;
+  status?: 'active' | 'archived' | 'all';
 }
 
 export interface SessionMessagePage {
@@ -102,8 +106,8 @@ export const sessionApi = {
   /**
    * 获取会话列表
    */
-  list: async (params?: SessionListParams) => {
-    const response = await client.get('/api/session', { params });
+  list: async (params?: SessionListParams): Promise<Session[]> => {
+    const response = await client.get<Session[]>('/api/session', { params });
     return response.data;
   },
 
@@ -126,27 +130,57 @@ export const sessionApi = {
   /**
    * 创建会话
    */
-  create: async (data?: { title?: string; parentID?: string; projectID?: string }) => {
+  create: async (data?: { title?: string; parentID?: string; projectID?: string; model_auto?: boolean }) => {
     const response = await client.post('/api/session', data || {});
     return response.data;
   },
 
   /**
-   * 删除会话
+   * 永久删除会话（普通工作台应使用 archive）
    */
-  delete: async (sessionId: string) => {
-    const response = await client.delete(`/api/session/${sessionId}`);
+  delete: async (sessionId: string): Promise<boolean> => {
+    const response = await client.delete<boolean>(`/api/session/${sessionId}`);
+    return response.data;
+  },
+
+  /**
+   * 归档会话并保留全部持久化数据
+   */
+  archive: async (sessionId: string): Promise<SessionResponse> => {
+    const response = await client.post(`/api/session/${sessionId}/archive`);
+    return response.data;
+  },
+
+  /**
+   * 恢复已归档会话
+   */
+  restore: async (sessionId: string): Promise<SessionResponse> => {
+    const response = await client.post(`/api/session/${sessionId}/restore`);
     return response.data;
   },
 
   /**
    * 更新会话
    */
-  update: async (sessionId: string, data: { title?: string; provider?: string; model?: string; model_pinned?: boolean }) => {
+  update: async (sessionId: string, data: {
+    title?: string;
+    provider?: string;
+    model?: string;
+    model_pinned?: boolean;
+    model_auto?: boolean;
+  }) => {
     const response = await client.patch(`/api/session/${sessionId}`, data);
     return response.data;
   },
 
+  /**
+   * 将任务及其子任务移动到指定项目
+   */
+  moveToProject: async (sessionId: string, projectID: string): Promise<SessionResponse> => {
+    const response = await client.patch(`/api/session/${sessionId}/project`, { projectID });
+    return response.data;
+  },
+
   /**
    * 本地共享会话（所有本地账号可见，只读）
    */
@@ -215,6 +249,7 @@ export const sessionApi = {
     model?: Record<string, unknown>;
     variant?: string;
     displayText?: string;
+    executionMode?: SessionExecutionMode;
   }) => {
     const response = await client.post(`/api/session/${sessionId}/prompt_queue`, data);
     return response.data;
diff --git a/webui/src/api/webuiContractPages.ts b/webui/src/api/webuiContractPages.ts
index 5cb3f3846..abdee30af 100644
--- a/webui/src/api/webuiContractPages.ts
+++ b/webui/src/api/webuiContractPages.ts
@@ -29,6 +29,7 @@ export interface WebUIContractWorkspaceSection {
 
 export interface WebUIContractWorkspaceListItem {
   id: string;
+  version?: string;
   title: string;
   titleEn?: string | null;
   route: string;
diff --git a/webui/src/components/common/ChatPromptSelectors.test.tsx b/webui/src/components/common/ChatPromptSelectors.test.tsx
index 60ea93cbe..8bbe976fb 100644
--- a/webui/src/components/common/ChatPromptSelectors.test.tsx
+++ b/webui/src/components/common/ChatPromptSelectors.test.tsx
@@ -37,6 +37,7 @@ vi.mock('react-i18next', () => ({
         'modelPicker.empty': '暂无模型',
         'modelPicker.count': `${params?.count ?? 0}`,
         'modelPicker.vision': '视觉',
+        'modelPicker.auto': 'Auto',
         loading: '加载中',
       };
       return translations[key] ?? key;
@@ -104,7 +105,7 @@ beforeEach(() => {
 });
 
 describe('ChatModelPicker', () => {
-  it('opens the model menu toward the left edge of the trigger', async () => {
+  it('opens the model menu from the left edge of the trigger', async () => {
     const user = userEvent.setup();
 
     render(
@@ -120,13 +121,157 @@ describe('ChatModelPicker', () => {
 
     const menu = screen.getByText('选择模型').closest('.absolute');
     expect(menu).not.toBeNull();
-    expect(menu).toHaveClass('right-0');
+    expect(menu).toHaveClass('left-0');
     expect(menu).toHaveClass('bottom-full');
-    expect(menu).not.toHaveClass('left-0');
+    expect(menu).not.toHaveClass('right-0');
+    expect(menu).toHaveStyle({ transform: 'translateX(0px)' });
+  });
+
+  it('shifts the left-anchored menu only when the right edge would overflow', async () => {
+    const user = userEvent.setup();
+
+    render(
+      <ChatModelPicker
+        groupedOptions={groupedOptions}
+        loading={false}
+        selectedModelOption={groupedOptions[0].models[0]}
+        onSelectModel={vi.fn()}
+      />,
+    );
+
+    const trigger = screen.getByRole('button', { name: /minimax-m3/i });
+    const selector = trigger.closest('[data-model-selector]');
+    vi.spyOn(selector!, 'getBoundingClientRect').mockReturnValue({
+      bottom: 0,
+      height: 0,
+      left: 700,
+      right: 700,
+      top: 0,
+      width: 0,
+      x: 700,
+      y: 0,
+      toJSON: () => ({}),
+    });
+    const originalViewportWidth = window.innerWidth;
+    Object.defineProperty(window, 'innerWidth', { configurable: true, value: 1000 });
+
+    await user.click(trigger);
+
+    const menu = screen.getByText('选择模型').closest('.absolute');
+    expect(menu).toHaveStyle({ transform: 'translateX(-36px)' });
+    Object.defineProperty(window, 'innerWidth', { configurable: true, value: originalViewportWidth });
+  });
+
+  it('renders Auto as an opt-in single-line item with its hint in the info tooltip', async () => {
+    const user = userEvent.setup();
+    const onSelectAuto = vi.fn();
+
+    render(
+      <ChatModelPicker
+        groupedOptions={groupedOptions}
+        loading={false}
+        selectedModelOption={groupedOptions[0].models[0]}
+        onSelectModel={vi.fn()}
+        autoOption={{
+          selected: false,
+          disabled: false,
+          statusLabel: 'Primary then fallback',
+          onSelect: onSelectAuto,
+        }}
+      />,
+    );
+
+    await user.click(screen.getByRole('button', { name: /minimax-m3/i }));
+    const autoButton = screen.getByRole('button', { name: 'Auto' });
+    expect(screen.queryByText('Primary then fallback')).not.toBeInTheDocument();
+
+    const info = autoButton.querySelector('.lucide-info')?.parentElement;
+    expect(info).toBeInTheDocument();
+    await user.hover(info!);
+    expect(await screen.findByText('Primary then fallback')).toBeInTheDocument();
+    await user.click(autoButton);
+    expect(onSelectAuto).toHaveBeenCalledOnce();
+  });
+
+  it('shows only Auto as selected when Auto mode is active', async () => {
+    const user = userEvent.setup();
+
+    render(
+      <ChatModelPicker
+        groupedOptions={groupedOptions}
+        loading={false}
+        selectedModelOption={groupedOptions[0].models[0]}
+        onSelectModel={vi.fn()}
+        autoOption={{
+          selected: true,
+          disabled: false,
+          statusLabel: 'Primary then fallback',
+          onSelect: vi.fn(),
+        }}
+      />,
+    );
+
+    const trigger = screen.getByRole('button', { name: /^Auto/i });
+    expect(trigger).toHaveAttribute('title', 'Auto: Primary then fallback');
+    await user.click(trigger);
+
+    expect(screen.getAllByRole('button', { name: 'Auto' })[1]).toHaveClass('shadow-[inset_2px_0_0_#a1a1aa]');
+    expect(screen.getByRole('button', { name: /minimax-m3/i })).not.toHaveClass('shadow-[inset_2px_0_0_#a1a1aa]');
+  });
+
+  it('does not show Auto unless the caller opts in', async () => {
+    const user = userEvent.setup();
+
+    render(
+      <ChatModelPicker
+        groupedOptions={groupedOptions}
+        loading={false}
+        selectedModelOption={groupedOptions[0].models[0]}
+        onSelectModel={vi.fn()}
+      />,
+    );
+
+    await user.click(screen.getByRole('button', { name: /minimax-m3/i }));
+    expect(screen.queryByRole('button', { name: 'Auto' })).not.toBeInTheDocument();
   });
 });
 
 describe('useChatModelOptions', () => {
+  it('keeps Auto opt-in and clears it when a concrete model is selected', async () => {
+    listDefinitionsMock.mockResolvedValue({
+      data: { models: [makeModelDefinition()] },
+    });
+    getResolvedMock.mockResolvedValue({
+      data: { provider_id: 'provider-1', model_id: 'model-1' },
+    });
+
+    const { result } = renderHook(() => useChatModelOptions({ enableAuto: true }));
+
+    await waitFor(() => {
+      expect(result.current.canSelectAuto).toBe(true);
+      expect(result.current.selectedModelKey).toBe('provider-1::model-1');
+    });
+    expect(result.current.selectedModelAuto).toBe(false);
+    expect(result.current.selectedPromptModel).toEqual({
+      providerID: 'provider-1',
+      modelID: 'model-1',
+    });
+
+    act(() => result.current.selectAuto());
+
+    expect(result.current.selectedModelAuto).toBe(true);
+    expect(result.current.selectedPromptModel).toBeNull();
+    expect(result.current.effectiveModelOption).toEqual(result.current.primaryModelOption);
+
+    act(() => result.current.selectModelKey('provider-1::model-1'));
+
+    expect(result.current.selectedModelAuto).toBe(false);
+    expect(result.current.selectedPromptModel).toEqual({
+      providerID: 'provider-1',
+      modelID: 'model-1',
+    });
+  });
+
   it('shares enabled model and default model requests across concurrent hook instances', async () => {
     let resolveDefinitions: (value: { data: { models: any[] } }) => void = () => {};
     listDefinitionsMock.mockReturnValue(new Promise((resolve) => {
diff --git a/webui/src/components/common/ChatPromptSelectors.tsx b/webui/src/components/common/ChatPromptSelectors.tsx
index 0df85d411..c39b6bec2 100644
--- a/webui/src/components/common/ChatPromptSelectors.tsx
+++ b/webui/src/components/common/ChatPromptSelectors.tsx
@@ -1,5 +1,5 @@
-import { useCallback, useEffect, useMemo, useState } from 'react';
-import { Bot, ChevronDown, Cpu, Info } from 'lucide-react';
+import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+import { Bot, ChevronDown, Cpu, Info, Sparkles } from 'lucide-react';
 import { useTranslation } from 'react-i18next';
 
 import type { Agent } from '@/api/agent';
@@ -33,6 +33,13 @@ export type ChatModelProviderGroup = {
   models: ChatModelOption[];
 };
 
+export type ChatModelPickerAutoOption = {
+  selected: boolean;
+  disabled: boolean;
+  statusLabel: string;
+  onSelect: () => void;
+};
+
 type SelectorTooltip = {
   title: string;
   lines: string[];
@@ -40,6 +47,23 @@ type SelectorTooltip = {
   y: number;
 };
 
+const MODEL_MENU_WIDTH_PX = 320;
+const MODEL_MENU_VIEWPORT_PADDING_PX = 16;
+
+export function getAnchoredMenuLeftOffset(
+  anchorLeft: number,
+  viewportWidth: number,
+  menuWidth = MODEL_MENU_WIDTH_PX,
+  viewportPadding = MODEL_MENU_VIEWPORT_PADDING_PX,
+): number {
+  const constrainedWidth = Math.min(
+    menuWidth,
+    Math.max(0, viewportWidth - viewportPadding * 2),
+  );
+  const rightOverflow = anchorLeft + constrainedWidth + viewportPadding - viewportWidth;
+  return rightOverflow > 0 ? -rightOverflow : 0;
+}
+
 export function __resetChatModelOptionsResourcesForTesting(): void {
   __resetChatModelResourcesForTesting();
 }
@@ -73,7 +97,7 @@ export function useChatAgentOptions(options: { allowedAgentNames?: string[] } =
   };
 }
 
-export function useChatModelOptions() {
+export function useChatModelOptions({ enableAuto = false }: { enableAuto?: boolean } = {}) {
   const { t } = useTranslation('session');
   const { providers, loading: loadingProviders } = useProviders();
   const {
@@ -81,6 +105,7 @@ export function useChatModelOptions() {
     loading: loadingEnabledModels,
   } = useEnabledChatModelDefinitions();
   const [selectedModelKey, setSelectedModelKey] = useState<string | null>(null);
+  const [selectedModelAuto, setSelectedModelAuto] = useState(false);
 
   const options = useMemo<ChatModelOption[]>(() => {
     const providerById = new Map(
@@ -158,6 +183,34 @@ export function useChatModelOptions() {
     data: resolvedDefaultModel,
     initialized: resolvedDefaultModelInitialized,
   } = useResolvedDefaultModel(options.length > 0);
+  const primaryModelOption = useMemo(() => {
+    if (!resolvedDefaultModel) return null;
+    return options.find((option) => (
+      option.providerID === resolvedDefaultModel.providerID
+      && option.modelID === resolvedDefaultModel.modelID
+    )) ?? null;
+  }, [options, resolvedDefaultModel]);
+  const canSelectAuto = Boolean(enableAuto && primaryModelOption);
+  const effectiveModelOption = selectedModelAuto ? primaryModelOption : selectedModelOption;
+  const selectModelKey = useCallback((key: string) => {
+    setSelectedModelAuto(false);
+    setSelectedModelKey(key);
+  }, []);
+  const selectAuto = useCallback(() => {
+    if (canSelectAuto) setSelectedModelAuto(true);
+  }, [canSelectAuto]);
+  const modelPickerAutoOption = useMemo<ChatModelPickerAutoOption | undefined>(() => (
+    enableAuto
+      ? {
+          selected: selectedModelAuto,
+          disabled: !canSelectAuto,
+          statusLabel: canSelectAuto
+            ? t('modelPicker.autoHint')
+            : t('modelPicker.autoUnavailable'),
+          onSelect: selectAuto,
+        }
+      : undefined
+  ), [canSelectAuto, enableAuto, selectAuto, selectedModelAuto, t]);
 
   useEffect(() => {
     if (selectedModelKey || options.length === 0 || !resolvedDefaultModelInitialized) return;
@@ -178,11 +231,19 @@ export function useChatModelOptions() {
     groupedOptions,
     loading: loadingProviders || loadingEnabledModels,
     options,
+    canSelectAuto,
+    effectiveModelOption,
+    modelPickerAutoOption,
+    primaryModelOption,
+    selectAuto,
+    selectModelKey,
+    selectedModelAuto,
     selectedModelKey,
     selectedModelOption,
-    selectedPromptModel: selectedModelOption
+    selectedPromptModel: !selectedModelAuto && selectedModelOption
       ? { providerID: selectedModelOption.providerID, modelID: selectedModelOption.modelID }
       : null,
+    setSelectedModelAuto,
     setSelectedModelKey,
   };
 }
@@ -417,16 +478,28 @@ export function ChatModelPicker({
   loading,
   selectedModelOption,
   onSelectModel,
+  autoOption,
 }: {
   groupedOptions: ChatModelProviderGroup[];
   loading: boolean;
   selectedModelOption: ChatModelOption | null;
   onSelectModel: (option: ChatModelOption) => void;
+  autoOption?: ChatModelPickerAutoOption;
 }) {
   const { t } = useTranslation('session');
   const [open, setOpen] = useState(false);
+  const selectorRef = useRef<HTMLDivElement>(null);
+  const [menuLeftOffset, setMenuLeftOffset] = useState(0);
   const { tooltip, showTooltip, hideTooltip } = useSelectorTooltip();
   const hasOptions = groupedOptions.some((group) => group.models.length > 0);
+  const updateMenuLeftOffset = useCallback(() => {
+    const selector = selectorRef.current;
+    if (!selector) return;
+    setMenuLeftOffset(getAnchoredMenuLeftOffset(
+      selector.getBoundingClientRect().left,
+      window.innerWidth,
+    ));
+  }, []);
 
   useEffect(() => {
     if (!open) return;
@@ -442,23 +515,44 @@ export function ChatModelPicker({
     if (!open) hideTooltip();
   }, [hideTooltip, open]);
 
+  useEffect(() => {
+    if (!open) return;
+    updateMenuLeftOffset();
+    window.addEventListener('resize', updateMenuLeftOffset);
+    return () => window.removeEventListener('resize', updateMenuLeftOffset);
+  }, [open, updateMenuLeftOffset]);
+
   return (
-    <div className="relative" data-model-selector>
+    <div ref={selectorRef} className="relative" data-model-selector>
       <button
         type="button"
-        onClick={() => setOpen((value) => !value)}
+        onClick={() => {
+          if (!open) updateMenuLeftOffset();
+          setOpen((value) => !value);
+        }}
         disabled={loading || !hasOptions}
         className="flex h-7 w-[132px] min-w-0 items-center gap-1.5 rounded-lg px-2 text-xs text-zinc-600 transition-colors hover:bg-zinc-200/60 hover:text-zinc-900 disabled:cursor-not-allowed disabled:opacity-50 dark:text-zinc-300 dark:hover:bg-zinc-800 dark:hover:text-zinc-50"
-        title={selectedModelOption ? `${selectedModelOption.providerName} / ${selectedModelOption.modelID}` : t('modelPicker.empty')}
+        title={autoOption?.selected
+          ? `${t('modelPicker.auto')}: ${autoOption.statusLabel}`
+          : selectedModelOption
+            ? `${selectedModelOption.providerName} / ${selectedModelOption.modelID}`
+            : t('modelPicker.empty')}
       >
-        <Cpu className="h-3 w-3 shrink-0" />
+        {autoOption?.selected
+          ? <Sparkles className="h-3 w-3 shrink-0" />
+          : <Cpu className="h-3 w-3 shrink-0" />}
         <span className="truncate font-medium">
-          {selectedModelOption?.label ?? (loading ? t('loading') : t('modelPicker.empty'))}
+          {autoOption?.selected
+            ? t('modelPicker.auto')
+            : selectedModelOption?.label ?? (loading ? t('loading') : t('modelPicker.empty'))}
         </span>
         <ChevronDown className={`h-3 w-3 shrink-0 transition-transform ${open ? 'rotate-180' : ''}`} />
       </button>
       {open && (
-        <div className="absolute right-0 bottom-full z-50 mb-2 w-80 max-w-[calc(100vw-2rem)] rounded-lg border border-zinc-200 bg-white shadow-sm dark:border-zinc-800 dark:bg-zinc-900 dark:shadow-xl dark:shadow-black/30">
+        <div
+          className="absolute left-0 bottom-full z-50 mb-2 w-80 max-w-[calc(100vw-2rem)] rounded-lg border border-zinc-200 bg-white shadow-sm dark:border-zinc-800 dark:bg-zinc-900 dark:shadow-xl dark:shadow-black/30"
+          style={{ transform: `translateX(${menuLeftOffset}px)` }}
+        >
           <div className="border-b border-zinc-100 px-2.5 py-1.5 dark:border-zinc-800">
             <div className="text-xs font-semibold text-zinc-700 dark:text-zinc-100">{t('modelPicker.title')}</div>
             <div className="truncate text-[10px] text-zinc-400 dark:text-zinc-500">{t('modelPicker.hint')}</div>
@@ -466,60 +560,98 @@ export function ChatModelPicker({
           <div className="h-[13.5rem] overflow-y-auto p-1.5">
             {loading ? (
               <div className="p-3 text-center text-xs text-zinc-500">{t('loading')}</div>
-            ) : groupedOptions.length > 0 ? (
-              groupedOptions.map((group) => (
-                <div key={group.providerID} className="py-1 first:pt-0 last:pb-0">
-                  <div className="sticky top-0 z-10 flex items-center justify-between gap-2 bg-white/95 px-1.5 py-1 text-[10px] font-semibold text-zinc-500 backdrop-blur dark:bg-zinc-900/95 dark:text-zinc-400">
-                    <span className="truncate">{group.providerName}</span>
-                    <span className="shrink-0 rounded bg-zinc-50 px-1.5 py-0.5 text-[9px] text-zinc-500 dark:bg-zinc-800 dark:text-zinc-300">
-                      {t('modelPicker.count', { count: group.models.length })}
-                    </span>
+            ) : (
+              <>
+                {autoOption && (
+                  <div className="mb-1 border-b border-zinc-100 pb-1.5 dark:border-zinc-800">
+                    <button
+                      type="button"
+                      onClick={() => {
+                        autoOption.onSelect();
+                        setOpen(false);
+                      }}
+                      disabled={autoOption.disabled}
+                      className={`w-full rounded-md px-2 py-1.5 text-left transition-colors disabled:cursor-not-allowed disabled:opacity-45 ${
+                        autoOption.selected
+                          ? 'bg-zinc-50 text-zinc-900 shadow-[inset_2px_0_0_#a1a1aa] dark:bg-zinc-800 dark:text-zinc-50 dark:shadow-[inset_2px_0_0_#539bf5]'
+                          : 'text-zinc-700 hover:bg-zinc-50 dark:text-zinc-300 dark:hover:bg-zinc-800 dark:hover:text-zinc-50'
+                      }`}
+                    >
+                      <div className="flex min-w-0 items-center gap-2">
+                        <Sparkles className={`h-3 w-3 shrink-0 ${autoOption.selected ? 'text-zinc-600 dark:text-zinc-200' : 'text-zinc-400 dark:text-zinc-500'}`} />
+                        <span className="min-w-0 flex-1 truncate text-xs font-medium text-zinc-900 dark:text-zinc-100">
+                          {t('modelPicker.auto')}
+                        </span>
+                        <span
+                          className="group relative rounded p-0.5 transition-colors hover:bg-zinc-200 dark:hover:bg-zinc-700"
+                          onMouseDown={(event) => { event.preventDefault(); event.stopPropagation(); }}
+                          onClick={(event) => { event.preventDefault(); event.stopPropagation(); }}
+                          onPointerEnter={(event) => showTooltip(event.currentTarget, t('modelPicker.auto'), [autoOption.statusLabel])}
+                          onMouseEnter={(event) => showTooltip(event.currentTarget, t('modelPicker.auto'), [autoOption.statusLabel])}
+                          onMouseOver={(event) => showTooltip(event.currentTarget, t('modelPicker.auto'), [autoOption.statusLabel])}
+                          onMouseLeave={hideTooltip}
+                          onPointerLeave={hideTooltip}
+                        >
+                          <Info className="h-3 w-3 text-zinc-300 transition-colors group-hover:text-zinc-500 dark:text-zinc-600 dark:group-hover:text-zinc-300" />
+                        </span>
+                      </div>
+                    </button>
                   </div>
-                  <div className="space-y-0.5">
-                    {group.models.map((option) => (
-                      <button
-                        key={option.key}
-                        type="button"
-                        onClick={() => {
-                          onSelectModel(option);
-                          setOpen(false);
-                        }}
-                        className={`w-full rounded-md px-2 py-1.5 text-left transition-colors ${
-                          selectedModelOption?.key === option.key
-                            ? 'bg-zinc-50 text-zinc-900 shadow-[inset_2px_0_0_#a1a1aa] dark:bg-zinc-800 dark:text-zinc-50 dark:shadow-[inset_2px_0_0_#539bf5]'
-                            : 'text-zinc-700 hover:bg-zinc-50 dark:text-zinc-300 dark:hover:bg-zinc-800 dark:hover:text-zinc-50'
-                        }`}
-                      >
-                        <div className="flex min-w-0 items-center gap-2">
-                          <Cpu className={`h-3 w-3 shrink-0 ${selectedModelOption?.key === option.key ? 'text-zinc-600 dark:text-zinc-200' : 'text-zinc-400 dark:text-zinc-500'}`} />
-                          <span className="min-w-0 flex-1 truncate text-xs font-medium text-zinc-900 dark:text-zinc-100">{option.label}</span>
-                          {option.supportsVision === true && (
-                            <span className="shrink-0 rounded bg-zinc-100 px-1.5 py-0.5 text-[9px] font-medium text-zinc-600 dark:bg-zinc-800 dark:text-zinc-300">
-                              {t('modelPicker.vision')}
-                            </span>
-                          )}
-                          <div className="ml-auto flex shrink-0 items-center gap-1">
-                            <span
-                              className="group relative rounded p-0.5 transition-colors hover:bg-zinc-200 dark:hover:bg-zinc-700"
-                              onMouseDown={(event) => { event.preventDefault(); event.stopPropagation(); }}
-                              onClick={(event) => { event.preventDefault(); event.stopPropagation(); }}
-                              onPointerEnter={(event) => showTooltip(event.currentTarget, option.label, [option.pricingLabel, option.contextLabel])}
-                              onMouseEnter={(event) => showTooltip(event.currentTarget, option.label, [option.pricingLabel, option.contextLabel])}
-                              onMouseOver={(event) => showTooltip(event.currentTarget, option.label, [option.pricingLabel, option.contextLabel])}
-                              onMouseLeave={hideTooltip}
-                              onPointerLeave={hideTooltip}
-                            >
-                              <Info className="h-3 w-3 text-zinc-300 transition-colors group-hover:text-zinc-500 dark:text-zinc-600 dark:group-hover:text-zinc-300" />
-                            </span>
+                )}
+                {groupedOptions.length > 0 ? groupedOptions.map((group) => (
+                  <div key={group.providerID} className="py-1 first:pt-0 last:pb-0">
+                    <div className="sticky top-0 z-10 flex items-center justify-between gap-2 bg-white/95 px-1.5 py-1 text-[10px] font-semibold text-zinc-500 backdrop-blur dark:bg-zinc-900/95 dark:text-zinc-400">
+                      <span className="truncate">{group.providerName}</span>
+                      <span className="shrink-0 rounded bg-zinc-50 px-1.5 py-0.5 text-[9px] text-zinc-500 dark:bg-zinc-800 dark:text-zinc-300">
+                        {t('modelPicker.count', { count: group.models.length })}
+                      </span>
+                    </div>
+                    <div className="space-y-0.5">
+                      {group.models.map((option) => (
+                        <button
+                          key={option.key}
+                          type="button"
+                          onClick={() => {
+                            onSelectModel(option);
+                            setOpen(false);
+                          }}
+                          className={`w-full rounded-md px-2 py-1.5 text-left transition-colors ${
+                            !autoOption?.selected && selectedModelOption?.key === option.key
+                              ? 'bg-zinc-50 text-zinc-900 shadow-[inset_2px_0_0_#a1a1aa] dark:bg-zinc-800 dark:text-zinc-50 dark:shadow-[inset_2px_0_0_#539bf5]'
+                              : 'text-zinc-700 hover:bg-zinc-50 dark:text-zinc-300 dark:hover:bg-zinc-800 dark:hover:text-zinc-50'
+                          }`}
+                        >
+                          <div className="flex min-w-0 items-center gap-2">
+                            <Cpu className={`h-3 w-3 shrink-0 ${!autoOption?.selected && selectedModelOption?.key === option.key ? 'text-zinc-600 dark:text-zinc-200' : 'text-zinc-400 dark:text-zinc-500'}`} />
+                            <span className="min-w-0 flex-1 truncate text-xs font-medium text-zinc-900 dark:text-zinc-100">{option.label}</span>
+                            {option.supportsVision === true && (
+                              <span className="shrink-0 rounded bg-zinc-100 px-1.5 py-0.5 text-[9px] font-medium text-zinc-600 dark:bg-zinc-800 dark:text-zinc-300">
+                                {t('modelPicker.vision')}
+                              </span>
+                            )}
+                            <div className="ml-auto flex shrink-0 items-center gap-1">
+                              <span
+                                className="group relative rounded p-0.5 transition-colors hover:bg-zinc-200 dark:hover:bg-zinc-700"
+                                onMouseDown={(event) => { event.preventDefault(); event.stopPropagation(); }}
+                                onClick={(event) => { event.preventDefault(); event.stopPropagation(); }}
+                                onPointerEnter={(event) => showTooltip(event.currentTarget, option.label, [option.pricingLabel, option.contextLabel])}
+                                onMouseEnter={(event) => showTooltip(event.currentTarget, option.label, [option.pricingLabel, option.contextLabel])}
+                                onMouseOver={(event) => showTooltip(event.currentTarget, option.label, [option.pricingLabel, option.contextLabel])}
+                                onMouseLeave={hideTooltip}
+                                onPointerLeave={hideTooltip}
+                              >
+                                <Info className="h-3 w-3 text-zinc-300 transition-colors group-hover:text-zinc-500 dark:text-zinc-600 dark:group-hover:text-zinc-300" />
+                              </span>
+                            </div>
                           </div>
-                        </div>
-                      </button>
-                    ))}
+                        </button>
+                      ))}
+                    </div>
                   </div>
-                </div>
-              ))
-            ) : (
-              <div className="p-3 text-center text-xs text-zinc-500">{t('modelPicker.empty')}</div>
+                )) : (
+                  <div className="p-3 text-center text-xs text-zinc-500">{t('modelPicker.empty')}</div>
+                )}
+              </>
             )}
           </div>
         </div>
diff --git a/webui/src/components/common/DelegateDetailSheet.test.tsx b/webui/src/components/common/DelegateDetailSheet.test.tsx
new file mode 100644
index 000000000..4074768b5
--- /dev/null
+++ b/webui/src/components/common/DelegateDetailSheet.test.tsx
@@ -0,0 +1,51 @@
+import { render } from '@testing-library/react';
+import { describe, expect, it, vi } from 'vitest';
+
+import DelegateDetailSheet from './DelegateDetailSheet';
+
+const { sessionChatPropsMock } = vi.hoisted(() => ({
+  sessionChatPropsMock: vi.fn(),
+}));
+
+vi.mock('react-i18next', () => ({
+  useTranslation: () => ({
+    t: (key: string) => key,
+  }),
+}));
+
+vi.mock('./SessionChat', () => ({
+  default: (props: Record<string, unknown>) => {
+    sessionChatPropsMock(props);
+    return null;
+  },
+}));
+
+describe('DelegateDetailSheet', () => {
+  it('uses the session-management process timeline inside the child conversation', () => {
+    render(
+      <DelegateDetailSheet
+        open
+        onClose={vi.fn()}
+        sessionId="ses-child"
+        agentName="Librarian"
+        description="调研 OpenClaw 最新版本"
+        status="running"
+      />,
+    );
+
+    expect(sessionChatPropsMock).toHaveBeenCalledWith(expect.objectContaining({
+      sessionId: 'ses-child',
+      live: true,
+      hideInput: true,
+      agentName: 'Librarian',
+      display: {
+        compact: false,
+        pageCanvas: true,
+        showTimestamp: true,
+        collapseIntermediateSteps: true,
+        processGroupsDefaultOpen: false,
+        processGroupsOpenWhileActive: true,
+      },
+    }));
+  });
+});
diff --git a/webui/src/components/common/DelegateDetailSheet.tsx b/webui/src/components/common/DelegateDetailSheet.tsx
index 420c981b3..274eb6922 100644
--- a/webui/src/components/common/DelegateDetailSheet.tsx
+++ b/webui/src/components/common/DelegateDetailSheet.tsx
@@ -178,7 +178,15 @@ export default function DelegateDetailSheet({
             sessionId={sessionId}
             live
             hideInput
-            display={{ compact: false, showTimestamp: true }}
+            display={{
+              compact: false,
+              pageCanvas: true,
+              showTimestamp: true,
+              collapseIntermediateSteps: true,
+              processGroupsDefaultOpen: false,
+              processGroupsOpenWhileActive: true,
+            }}
+            agentName={agentName}
             className="h-full"
             emptyText={t('delegate.emptyChat')}
           />
diff --git a/webui/src/components/common/DelegateTaskCard.test.tsx b/webui/src/components/common/DelegateTaskCard.test.tsx
index a50dbbefb..3db628e49 100644
--- a/webui/src/components/common/DelegateTaskCard.test.tsx
+++ b/webui/src/components/common/DelegateTaskCard.test.tsx
@@ -1,8 +1,48 @@
-import { describe, expect, it } from 'vitest';
+import { fireEvent, render, screen } from '@testing-library/react';
+import { describe, expect, it, vi } from 'vitest';
 
-import { shouldRenderDelegateTaskCard } from './DelegateTaskCard';
+import DelegateTaskCard, { shouldRenderDelegateTaskCard } from './DelegateTaskCard';
 import type { MessagePart } from '../../types';
 
+vi.mock('react-i18next', () => ({
+  useTranslation: () => ({
+    t: (key: string, options?: Record<string, unknown>) => {
+      const translations: Record<string, string> = {
+        'delegate.delegatedTo': '委派给 {{agent}}',
+        'delegate.pending': '等待中',
+        'delegate.running': '执行中',
+        'delegate.completed': '已完成',
+        'delegate.background': '后台',
+        'delegate.elapsedRunning': '已运行',
+        'delegate.elapsedDone': '耗时',
+        'delegate.steps': '步',
+        'delegate.resultSummary': '结果摘要',
+        'delegate.viewExecution': '查看执行过程',
+        'delegate.subTask': '子任务',
+      };
+      return (translations[key] ?? key).replace(
+        /\{\{(\w+)\}\}/g,
+        (_, name: string) => String(options?.[name] ?? ''),
+      );
+    },
+  }),
+}));
+
+vi.mock('./DelegateDetailSheet', async () => {
+  const ReactModule = await import('react');
+  return {
+    default: ({
+      open,
+      sessionId,
+    }: {
+      open: boolean;
+      sessionId: string;
+    }) => open
+      ? ReactModule.createElement('div', { 'data-testid': 'delegate-detail-sheet' }, sessionId)
+      : null,
+  };
+});
+
 describe('shouldRenderDelegateTaskCard', () => {
   it('does not treat generic tool category fields as delegate tasks', () => {
     const part = {
@@ -84,3 +124,48 @@ describe('shouldRenderDelegateTaskCard', () => {
     expect(shouldRenderDelegateTaskCard(part)).toBe(false);
   });
 });
+
+describe('DelegateTaskCard process step', () => {
+  it('shows delegated agent context and opens the child execution sheet', () => {
+    const part = {
+      id: 'part-delegate',
+      type: 'tool',
+      tool: 'delegate_task',
+      state: {
+        status: 'running',
+        input: {
+          description: '调研 OpenClaw 最新版本',
+          prompt: '调研 GitHub 上的最新发布',
+          subagent_type: 'librarian',
+          run_in_background: true,
+        },
+        metadata: {
+          sessionId: 'ses-child',
+          status: 'running',
+          background: true,
+          steps: [
+            {
+              tool: 'websearch',
+              title: '搜索最新版本',
+              status: 'running',
+            },
+          ],
+          stepCount: 1,
+        },
+      },
+    } as MessagePart;
+
+    render(<DelegateTaskCard part={part} processStep />);
+
+    const processStep = screen.getByTestId('chat-process-delegate-step');
+    expect(processStep).toHaveTextContent('委派给 Librarian');
+    expect(processStep).toHaveTextContent('调研 OpenClaw 最新版本');
+    expect(processStep).toHaveTextContent('执行中');
+    expect(processStep).toHaveTextContent('后台');
+    expect(processStep.querySelector('summary')).toHaveClass('text-sm');
+
+    fireEvent.click(screen.getByRole('button', { name: /查看执行过程/ }));
+
+    expect(screen.getByTestId('delegate-detail-sheet')).toHaveTextContent('ses-child');
+  });
+});
diff --git a/webui/src/components/common/DelegateTaskCard.tsx b/webui/src/components/common/DelegateTaskCard.tsx
index 9110fb61f..1b0a28278 100644
--- a/webui/src/components/common/DelegateTaskCard.tsx
+++ b/webui/src/components/common/DelegateTaskCard.tsx
@@ -6,7 +6,7 @@
  */
 
 import { useState, useEffect, useRef } from 'react';
-import { ChevronRight, ExternalLink, XCircle } from 'lucide-react';
+import { Bot, ChevronDown, ChevronRight, ExternalLink, XCircle } from 'lucide-react';
 import { useTranslation } from 'react-i18next';
 import type { MessagePart, ToolState } from '@/types';
 import DelegateDetailSheet from './DelegateDetailSheet';
@@ -211,9 +211,10 @@ const STATUS_STYLE: Record<string, {
 
 export interface DelegateTaskCardProps {
   part: MessagePart;
+  processStep?: boolean;
 }
 
-export default function DelegateTaskCard({ part }: DelegateTaskCardProps) {
+export default function DelegateTaskCard({ part, processStep = false }: DelegateTaskCardProps) {
   const { t } = useTranslation('common');
   let state: Partial<ToolState> = {};
   let info: DelegateInfo;
@@ -253,6 +254,123 @@ export default function DelegateTaskCard({ part }: DelegateTaskCardProps) {
     setElapsed(info.durationMs);
   }, [info.status, state.time?.start, state.time?.end, info.durationMs]);
 
+  if (processStep) {
+    return (
+      <>
+        <details data-testid="chat-process-delegate-step" className="group/delegate min-w-0">
+          <summary className="flex min-h-7 cursor-pointer list-none items-center gap-2 text-sm font-medium text-[#747a78] transition-colors hover:text-zinc-900 dark:text-zinc-400 dark:hover:text-zinc-200 [&::-webkit-details-marker]:hidden">
+            <span className="inline-grid h-[18px] w-[18px] flex-[0_0_18px] place-items-center text-violet-500 dark:text-violet-400">
+              <Bot className="h-4 w-4" />
+            </span>
+            <span className="min-w-0 flex-shrink-0 text-violet-700 dark:text-violet-300">
+              {t('delegate.delegatedTo', { agent: info.agentName })}
+            </span>
+            <span className="min-w-0 truncate font-normal text-[#9a9f9c] dark:text-zinc-500">
+              {info.description}
+            </span>
+            {info.isBackground && (
+              <span className="flex-shrink-0 rounded bg-violet-100 px-1 py-0.5 text-[9px] font-medium leading-none text-violet-600 dark:bg-violet-950/70 dark:text-violet-300">
+                {t('delegate.background')}
+              </span>
+            )}
+            <span className={`ml-auto flex flex-shrink-0 items-center gap-1 text-[11px] ${cfg.textColor}`}>
+              {cfg.pulse && (
+                <span className="h-1.5 w-1.5 rounded-full bg-sky-500 animate-pulse" />
+              )}
+              {t(`delegate.${info.status}`, { defaultValue: info.status })}
+            </span>
+            <ChevronDown className="ml-0.5 h-3 w-3 flex-shrink-0 text-[#9da29f] transition-transform group-open/delegate:rotate-180 dark:text-zinc-500" />
+          </summary>
+
+          <div className="mb-[9px] ml-2 mt-[3px] space-y-2 border-l border-[#e3e6e3] py-1.5 pl-[26px] pr-0 text-xs text-[#686e6c] dark:border-zinc-700 dark:text-zinc-400">
+            {elapsed !== null && elapsed > 0 && (
+              <div className="text-[11px] text-zinc-400 dark:text-zinc-500">
+                {info.status === 'running' ? t('delegate.elapsedRunning') : t('delegate.elapsedDone')}
+                {' '}
+                {formatDuration(elapsed)}
+                {info.stepCount > 0 && ` · ${info.stepCount} ${t('delegate.steps')}`}
+              </div>
+            )}
+
+            {info.status === 'running' && info.steps.length > 0 && (
+              <div className="space-y-1">
+                {info.steps.map((step, index) => (
+                  <div key={`${step.tool}-${index}`} className="flex min-w-0 items-center gap-1.5 text-[11px] leading-snug">
+                    <span className={
+                      step.status === 'completed'
+                        ? 'text-emerald-500'
+                        : step.status === 'error'
+                          ? 'text-red-500'
+                          : 'text-sky-400 animate-pulse'
+                    }>
+                      {step.status === 'completed' ? '✓' : step.status === 'error' ? '✗' : '◌'}
+                    </span>
+                    <span className="w-16 flex-shrink-0 truncate font-mono text-[10px] text-zinc-500">
+                      {step.tool}
+                    </span>
+                    <span className="truncate text-zinc-400">{step.title}</span>
+                  </div>
+                ))}
+                {info.currentText && (
+                  <div className="truncate pl-4 text-[10px] italic text-zinc-400">
+                    ⋯ {info.currentText.slice(-80)}
+                  </div>
+                )}
+              </div>
+            )}
+
+            {info.status === 'error' && info.error && (
+              <div className="flex items-start gap-1.5 rounded-md bg-red-50 p-2 text-[11px] text-red-700 dark:bg-red-950/40 dark:text-red-300">
+                <XCircle className="mt-0.5 h-3.5 w-3.5 flex-shrink-0" />
+                <span className="break-words">{info.error}</span>
+              </div>
+            )}
+
+            {info.status === 'completed' && info.output && (
+              <details>
+                <summary className="cursor-pointer select-none text-[11px] font-medium text-zinc-500 hover:text-zinc-700 dark:text-zinc-400 dark:hover:text-zinc-200">
+                  {t('delegate.resultSummary')}
+                </summary>
+                <div className="mt-1 max-h-40 overflow-y-auto whitespace-pre-wrap break-words rounded-md border border-zinc-200 bg-white/70 p-2 text-[11px] leading-relaxed text-zinc-700 dark:border-zinc-700 dark:bg-zinc-900/70 dark:text-zinc-300">
+                  {truncateOutput(info.output)}
+                </div>
+              </details>
+            )}
+
+            <button
+              type="button"
+              onClick={() => {
+                if (!info.childSessionId) return;
+                setSheetOpen(true);
+              }}
+              disabled={!info.childSessionId}
+              className={`flex items-center gap-1 text-[11px] font-medium transition-colors group ${
+                info.childSessionId
+                  ? 'cursor-pointer text-violet-600 hover:text-violet-800 dark:text-violet-400 dark:hover:text-violet-300'
+                  : 'cursor-not-allowed text-zinc-400 dark:text-zinc-600'
+              }`}
+            >
+              <ExternalLink className="h-3 w-3" />
+              {t('delegate.viewExecution')}
+              <ChevronRight className="h-3 w-3 transition-transform group-hover:translate-x-0.5" />
+            </button>
+          </div>
+        </details>
+
+        {info.childSessionId && (
+          <DelegateDetailSheet
+            open={sheetOpen}
+            onClose={() => setSheetOpen(false)}
+            sessionId={info.childSessionId}
+            agentName={info.agentName}
+            description={info.description}
+            status={info.status}
+          />
+        )}
+      </>
+    );
+  }
+
   return (
     <>
       <div className={`mt-2 rounded-lg border ${cfg.borderColor} ${cfg.bgColor} overflow-hidden`}>
diff --git a/webui/src/components/common/EntitySheet.test.tsx b/webui/src/components/common/EntitySheet.test.tsx
index ffde9c541..a8a85b394 100644
--- a/webui/src/components/common/EntitySheet.test.tsx
+++ b/webui/src/components/common/EntitySheet.test.tsx
@@ -200,6 +200,48 @@ describe('EntitySheet', () => {
       );
     });
 
+    it('propagates manually selected Auto mode to session creation and sending', async () => {
+      const createAndSend = vi.fn().mockResolvedValue('rex-session-auto');
+      mockUseSessionChat.mockReturnValue({
+        sessionId: null,
+        loading: false,
+        error: null,
+        create: vi.fn().mockResolvedValue(undefined),
+        createAndSend,
+        retry: vi.fn().mockResolvedValue(undefined),
+        reset: vi.fn(),
+      });
+
+      render(
+        <EntitySheet
+          {...defaultProps}
+          initialTab="rex"
+          rexAgentName="rex"
+          rexModel={null}
+          rexModelAuto
+        >
+          <div>Form content</div>
+        </EntitySheet>,
+      );
+
+      expect(mockUseSessionChat).toHaveBeenCalledWith(expect.objectContaining({
+        category: 'entity-config',
+        modelAuto: true,
+      }));
+      const sessionChatProps = vi.mocked(SessionChat).mock.calls.at(-1)?.[0] as any;
+      expect(sessionChatProps).toEqual(expect.objectContaining({
+        model: null,
+        modelAuto: true,
+      }));
+
+      await sessionChatProps.onCreateAndSend('hello', [], undefined, undefined);
+      expect(createAndSend).toHaveBeenCalledWith(expect.objectContaining({
+        text: 'hello',
+        model: null,
+        modelAuto: true,
+      }));
+    });
+
     it('renders extract from Rex as a guide action instead of a standalone footer action', async () => {
       const user = userEvent.setup();
       const onExtractFromRex = vi.fn().mockResolvedValue(undefined);
@@ -327,6 +369,30 @@ describe('EntitySheet', () => {
       });
     });
 
+    it('restores Auto selection from a persisted Rex session', async () => {
+      const onRexModelAutoChange = vi.fn();
+      window.localStorage.setItem(
+        'flocks:entity-sheet:rex-session:v1:agent-edit:auto-agent',
+        'persisted-auto-session',
+      );
+      mockClientGet.mockResolvedValueOnce({ data: { model_auto: true } });
+
+      render(
+        <EntitySheet
+          {...defaultProps}
+          initialTab="rex"
+          rexSessionStorageKey="agent-edit:auto-agent"
+          onRexModelAutoChange={onRexModelAutoChange}
+        >
+          <div>Form content</div>
+        </EntitySheet>,
+      );
+
+      await waitFor(() => {
+        expect(onRexModelAutoChange).toHaveBeenCalledWith(true);
+      });
+    });
+
     it('clears a stored Rex session when validation reports it missing', async () => {
       window.localStorage.setItem(
         'flocks:entity-sheet:rex-session:v1:agent-edit:audit-agent',
diff --git a/webui/src/components/common/EntitySheet.tsx b/webui/src/components/common/EntitySheet.tsx
index 37e39b752..355f73ff0 100644
--- a/webui/src/components/common/EntitySheet.tsx
+++ b/webui/src/components/common/EntitySheet.tsx
@@ -164,6 +164,8 @@ export interface EntitySheetProps {
   rexAgentName?: string;
   rexMentionAgents?: Agent[];
   rexModel?: { providerID: string; modelID: string } | null;
+  rexModelAuto?: boolean;
+  onRexModelAutoChange?: (enabled: boolean) => void;
   rexSupportsVision?: boolean | null;
   rexContextWindowTokens?: number | null;
   /** Persist and resume the Rex conversation across refreshes when provided. */
@@ -212,6 +214,8 @@ export default function EntitySheet({
   rexAgentName,
   rexMentionAgents,
   rexModel,
+  rexModelAuto = false,
+  onRexModelAutoChange,
   rexSupportsVision,
   rexContextWindowTokens,
   rexSessionStorageKey,
@@ -267,6 +271,7 @@ export default function EntitySheet({
   } = useSessionChat({
     title: `${title} — ${t('entity.rexAssist')}`,
     category: 'entity-config',
+    modelAuto: rexModelAuto,
     contextMessage: rexSystemContext,
     welcomeMessage: rexWelcomeMessage,
     initialSessionId: storedRexSessionId,
@@ -288,8 +293,9 @@ export default function EntitySheet({
 
     (async () => {
       try {
-        await client.get(`/api/session/${stored}`);
+        const response = await client.get(`/api/session/${stored}`);
         if (cancelled) return;
+        onRexModelAutoChange?.(Boolean(response.data?.model_auto));
         setStoredRexSessionId(stored);
       } catch {
         if (cancelled) return;
@@ -305,7 +311,7 @@ export default function EntitySheet({
     return () => {
       cancelled = true;
     };
-  }, [rexSessionStorageKey]);
+  }, [onRexModelAutoChange, rexSessionStorageKey]);
 
   useEffect(() => {
     if (!rexSessionHydrated || !sessionId) return;
@@ -374,6 +380,7 @@ export default function EntitySheet({
 
   useEffect(() => {
     if (!open) {
+      onRexModelAutoChange?.(false);
       setActiveTab(getDefaultTab());
       if (!rexSessionStorageKey) {
         resetRexSession();
@@ -387,7 +394,7 @@ export default function EntitySheet({
       setTestPrompt(effectiveDefaultTestPrompt);
       setDrawerWidth(resolvedInitialWidth());
     }
-  }, [open, mode, defaultTestPrompt, resetRexSession, initialWidth, showTabs, hideRex, hideForm, initialTab, rexSessionStorageKey]);
+  }, [open, mode, defaultTestPrompt, resetRexSession, initialWidth, showTabs, hideRex, hideForm, initialTab, onRexModelAutoChange, rexSessionStorageKey]);
 
   // ── Tab handling ──────────────────────────────────────────────────────────
 
@@ -432,18 +439,15 @@ export default function EntitySheet({
   const openRex = useCallback(
     (msg?: string) => {
       setActiveTab('rex');
-      if (activeRexSessionId && msg) {
-        const payload: Record<string, unknown> = {
-          parts: [{ type: 'text', text: msg }],
-        };
-        if (rexAgentName) payload.agent = rexAgentName;
-        if (rexModel) payload.model = rexModel;
-        client.post(`/api/session/${activeRexSessionId}/prompt_async`, payload);
-      } else if (msg) {
-        createAndSendRex({ text: msg, agent: rexAgentName, model: rexModel }).catch(() => {});
-      }
+      if (!msg) return;
+      createAndSendRex({
+        text: msg,
+        agent: rexAgentName,
+        model: rexModel,
+        modelAuto: rexModelAuto,
+      }).catch(() => {});
     },
-    [activeRexSessionId, createAndSendRex, rexAgentName, rexModel],
+    [createAndSendRex, rexAgentName, rexModel, rexModelAuto],
   );
 
   // ── openTest (exposed via context) ────────────────────────────────────────
@@ -499,9 +503,10 @@ export default function EntitySheet({
       text: prompt,
       agent: rexAgentName,
       model: rexModel,
+      modelAuto: rexModelAuto,
       displayText: buildInstructionDisplayText(label),
     }).catch(() => {});
-  }, [createAndSendRex, handleExtract, rexAgentName, rexModel]);
+  }, [createAndSendRex, handleExtract, rexAgentName, rexModel, rexModelAuto]);
 
   if (!open) return null;
 
@@ -698,6 +703,7 @@ export default function EntitySheet({
                     agentName={rexAgentName}
                     mentionAgents={rexMentionAgents}
                     model={rexModel}
+                    modelAuto={rexModelAuto}
                     supportsVision={rexSupportsVision ?? supportsVision}
                     contextWindowTokens={rexContextWindowTokens}
                     toolbarSlot={rexToolbarSlot}
@@ -709,6 +715,7 @@ export default function EntitySheet({
                       imageParts,
                       agent: agentOverride || rexAgentName,
                       model: modelOverride === undefined ? rexModel : modelOverride,
+                      modelAuto: rexModelAuto,
                       displayText: options?.displayText,
                     }) : undefined}
                     welcomeContent={(
diff --git a/webui/src/components/common/QuestionTool.test.tsx b/webui/src/components/common/QuestionTool.test.tsx
index 5399e521b..4be6c0cc6 100644
--- a/webui/src/components/common/QuestionTool.test.tsx
+++ b/webui/src/components/common/QuestionTool.test.tsx
@@ -195,6 +195,37 @@ describe('QuestionTool', () => {
     expect(screen.queryByRole('button', { name: /自定义 \/ 补充说明/ })).not.toBeInTheDocument();
   });
 
+  it('accepts feedback through the continue-planning option', async () => {
+    const user = userEvent.setup();
+    const onAnswer = vi.fn().mockResolvedValue(undefined);
+
+    render(
+      <QuestionTool
+        questions={[{
+          question: 'Would you like to start implementing?',
+          type: 'choice',
+          custom: false,
+          options: [
+            '开始实施',
+            {
+              label: '调整计划',
+              allowText: true,
+            },
+          ],
+        }]}
+        onAnswer={onAnswer}
+      />,
+    );
+
+    await user.click(screen.getByRole('button', { name: /调整计划/ }));
+    await user.type(screen.getByRole('textbox'), 'Keep the public API unchanged.');
+    await user.click(screen.getByRole('button', { name: /确认/ }));
+
+    expect(onAnswer).toHaveBeenCalledWith([
+      ['调整计划', 'Keep the public API unchanged.'],
+    ]);
+  });
+
   it('falls back to text input when a choice question has no visible options', async () => {
     const user = userEvent.setup();
     const onAnswer = vi.fn().mockResolvedValue(undefined);
diff --git a/webui/src/components/common/QuestionTool.tsx b/webui/src/components/common/QuestionTool.tsx
index 99ac8f631..79117e314 100644
--- a/webui/src/components/common/QuestionTool.tsx
+++ b/webui/src/components/common/QuestionTool.tsx
@@ -23,6 +23,8 @@ export type QuestionType = 'choice' | 'text' | 'number' | 'file' | 'confirm' | '
 export interface QuestionOption {
   label?: string;
   description?: string;
+  /** Show a text input and submit both the option label and entered text. */
+  allowText?: boolean;
   [key: string]: unknown;
 }
 
@@ -95,10 +97,14 @@ function optionDescription(opt: QuestionOption | string): string {
   return '';
 }
 
+function optionAllowsText(opt: QuestionOption | string): boolean {
+  return typeof opt !== 'string' && opt.allowText === true;
+}
+
 const CUSTOM_CHOICE_PREFIX = '__flocks_custom_choice__:';
 
 function isCustomChoiceLabel(label: string): boolean {
-  return /^(其他|其它|自定义|补充)|\b(other|custom)\b|请补充|补充说明|type your answer/i.test(label.trim());
+  return /^(其他|其它|自定义|补充)|\b(other|custom|feedback)\b|请补充|补充说明|type your answer/i.test(label.trim());
 }
 
 function customChoiceValue(text: string): string {
@@ -187,22 +193,26 @@ function ChoiceInput({
     .map(opt => ({
       label: optionLabel(opt),
       description: optionDescription(opt),
+      allowText: optionAllowsText(opt),
       custom: false,
     }))
     .filter(opt => opt.label);
-  const hasProvidedCustomOption = visibleOptions.some(opt => isCustomChoiceLabel(opt.label));
+  const hasProvidedCustomOption = visibleOptions.some(
+    opt => opt.allowText || isCustomChoiceLabel(opt.label),
+  );
   const options = shouldOfferCustomChoice(q) && !hasProvidedCustomOption
     ? [
         ...visibleOptions,
         {
           label: t('question.customAnswer'),
           description: t('question.textPlaceholder'),
+          allowText: false,
           custom: true,
         },
       ]
     : visibleOptions.map(opt => ({
         ...opt,
-        custom: isCustomChoiceLabel(opt.label),
+        custom: opt.allowText || isCustomChoiceLabel(opt.label),
       }));
   const customSelected = hasCustomChoice(answer);
   const customText = customChoiceText(answer);
@@ -213,7 +223,15 @@ function ChoiceInput({
       onChange([label]);
     }
   };
-  const toggleCustom = () => {
+  const toggleCustom = (label: string, preserveLabel: boolean) => {
+    if (preserveLabel) {
+      onChange(
+        customSelected && answer.includes(label)
+          ? []
+          : [label, customChoiceValue(customText)],
+      );
+      return;
+    }
     if (multiple) {
       if (customSelected) {
         onChange(answer.filter(value => !isCustomChoiceValue(value)));
@@ -224,8 +242,12 @@ function ChoiceInput({
     }
     onChange(customSelected ? [] : [customChoiceValue(customText)]);
   };
-  const setCustomText = (text: string) => {
+  const setCustomText = (label: string, preserveLabel: boolean, text: string) => {
     const nextCustom = customChoiceValue(text);
+    if (preserveLabel) {
+      onChange([label, nextCustom]);
+      return;
+    }
     if (multiple) {
       const withoutCustom = answer.filter(value => !isCustomChoiceValue(value));
       onChange([...withoutCustom, nextCustom]);
@@ -244,11 +266,17 @@ function ChoiceInput({
         {options.map(opt => {
           const label = opt.label;
           const desc = opt.description;
-          const selected = opt.custom ? customSelected : answer.includes(label);
+          const selected = opt.custom
+            ? customSelected && (!opt.allowText || answer.includes(label))
+            : answer.includes(label);
           return (
             <div key={label}>
               <button
-                onClick={() => (opt.custom ? toggleCustom() : toggle(label))}
+                onClick={() => (
+                  opt.custom
+                    ? toggleCustom(label, opt.allowText)
+                    : toggle(label)
+                )}
                 disabled={disabled}
                 className={`w-full text-left rounded-lg border px-3 py-2 text-sm transition-all flex items-start gap-2.5 ${
                   selected
@@ -287,7 +315,7 @@ function ChoiceInput({
                   <TextInput
                     q={{ ...q, type: 'text', placeholder: q.placeholder || t('question.textPlaceholder'), multiline: q.multiline ?? true }}
                     answer={[customText]}
-                    onChange={(value) => setCustomText(value[0] ?? '')}
+                    onChange={(value) => setCustomText(label, opt.allowText, value[0] ?? '')}
                     disabled={disabled}
                     compact={compact}
                   />
diff --git a/webui/src/components/common/SessionChat.test.ts b/webui/src/components/common/SessionChat.test.ts
index 5dcd06532..20788ca0f 100644
--- a/webui/src/components/common/SessionChat.test.ts
+++ b/webui/src/components/common/SessionChat.test.ts
@@ -1,5 +1,5 @@
 import React from 'react';
-import { act, fireEvent, render, screen, waitFor } from '@testing-library/react';
+import { act, fireEvent, render, screen, waitFor, within } from '@testing-library/react';
 import userEvent from '@testing-library/user-event';
 import { beforeEach, describe, expect, it, vi } from 'vitest';
 
@@ -48,16 +48,20 @@ const sessionApiResendMessageMock = vi.fn();
 const sessionApiRegenerateMessageMock = vi.fn();
 const sessionApiGetContextUsageMock = vi.fn();
 const sessionApiGetMock = vi.fn();
+const sessionApiUpdateMock = vi.fn();
 const useSessionMessagesMock = vi.fn();
 const useSSEOptionsRef = vi.hoisted(() => ({ current: null as any }));
 const tMock = (key: string, options?: Record<string, unknown>) => {
   const value = ({
   'chat.placeholder': '请输入消息',
   'chat.emptyText': '暂无消息',
+  'chat.loadFailed': '消息加载失败',
+  'chat.retry': '重试',
   'chat.sending': '发送中...',
   'chat.thinking': '思考中...',
   'chat.streaming': '继续输出中...',
-  'chat.process.title': '过程（{{count}} 项）',
+  'chat.process.title': '查看 {{count}} 个步骤',
+  'chat.process.deepThinking': '深度思考',
   'chat.process.reasoningCount': '{{count}} 段思考',
   'chat.process.toolCount': '{{count}} 次工具调用',
   'chat.process.textCount': '{{count}} 段中间回复',
@@ -83,6 +87,13 @@ const tMock = (key: string, options?: Record<string, unknown>) => {
   'chat.goal.status.completed': 'Completed',
   'chat.goal.status.blocked': 'Blocked',
   'chat.goal.status.paused': 'Paused',
+  'chat.addMenu.title': '添加',
+  'chat.addMenu.files': '文件和图片',
+  'chat.references.selected': '已选择的资源',
+  'chat.references.subagent': '子智能体',
+  'chat.references.skill': '技能',
+  'chat.references.workflow': '工作流',
+  'chat.references.remove': '移除 {{type}} {{name}}',
   'chat.mention.title': '选择 Agent',
   'chat.mention.navigate': '导航',
   'chat.mention.select': '选择',
@@ -90,7 +101,19 @@ const tMock = (key: string, options?: Record<string, unknown>) => {
   'chat.tool.running': '执行中',
   'chat.tool.completed': '已完成',
   'chat.tool.error': '失败',
+  'chat.tool.loadSkill': '加载技能',
+  'chat.tool.actions.readFile': '读取文件',
+  'chat.tool.actions.writeFile': '写入文件',
+  'chat.tool.actions.editFile': '编辑文件',
+  'chat.tool.actions.executeCommand': '执行命令',
+  'chat.tool.actions.askQuestion': '向用户提问',
+  'chat.tool.actions.installSkill': '安装技能',
+  'chat.tool.actions.addProvider': '添加模型服务',
+  'chat.tool.todoUpdated': '已更新待办',
   'chat.tool.inputParams': '输入参数',
+  'chat.tool.progress.writingFile': '正在写入文件…',
+  'chat.tool.progress.editingFile': '正在编辑文件…',
+  'chat.tool.progress.working': '正在执行此操作…',
   'chat.tool.outputResult': '输出结果',
   'chat.tool.todoStages': 'Todo 阶段',
   'chat.tool.todoStatus.pending': '待办',
@@ -190,6 +213,7 @@ vi.mock('@/api/client', () => ({
 vi.mock('@/api/session', () => ({
   sessionApi: {
     get: (...args: unknown[]) => sessionApiGetMock(...args),
+    update: (...args: unknown[]) => sessionApiUpdateMock(...args),
     listPromptQueue: (...args: unknown[]) => sessionApiListPromptQueueMock(...args),
     enqueuePrompt: (...args: unknown[]) => sessionApiEnqueuePromptMock(...args),
     updateQueuedPrompt: (...args: unknown[]) => sessionApiUpdateQueuedPromptMock(...args),
@@ -234,6 +258,7 @@ beforeEach(() => {
   sessionApiResendMessageMock.mockResolvedValue({});
   sessionApiRegenerateMessageMock.mockResolvedValue({});
   sessionApiGetMock.mockResolvedValue({});
+  sessionApiUpdateMock.mockResolvedValue({});
   sessionApiGetContextUsageMock.mockResolvedValue({
     sessionID: 'sess-1',
     usedTokens: 0,
@@ -251,10 +276,13 @@ beforeEach(() => {
   useSessionMessagesMock.mockReturnValue({
     messages: [],
     loading: false,
+    error: null,
     refetch: vi.fn(),
     addMessage: vi.fn(),
     updateMessage: vi.fn(),
     updateMessagePart: vi.fn(),
+    removeMessage: vi.fn(),
+    clearMessages: vi.fn(),
     replaceMessageText: vi.fn(),
     truncateAfterMessage: vi.fn(),
   });
@@ -271,6 +299,83 @@ function makeMessage(overrides: Partial<Message> & { id: string }): Message {
   } as Message;
 }
 
+type FetchedMessageFixture = {
+  info: {
+    id: string;
+    sessionID: string;
+    role: 'user' | 'assistant';
+    parentID?: string;
+    finish?: string | null;
+  };
+  parts: Message['parts'];
+};
+
+function makeFetchedMessage(
+  info: Omit<FetchedMessageFixture['info'], 'sessionID'>,
+  parts: Message['parts'] = [],
+): FetchedMessageFixture {
+  return {
+    info: { sessionID: 'sess-1', ...info },
+    parts,
+  };
+}
+
+function mockFallbackPolling({
+  localMessages,
+  fetchedMessages,
+  refetch,
+  status = { 'sess-1': { type: 'busy' } },
+}: {
+  localMessages: Message[];
+  fetchedMessages: FetchedMessageFixture[];
+  refetch: () => unknown;
+  status?: Record<string, { type: string }>;
+}) {
+  useSessionMessagesMock.mockReturnValue({
+    messages: localMessages,
+    loading: false,
+    refetch,
+    addMessage: vi.fn(),
+    updateMessage: vi.fn(),
+    updateMessagePart: vi.fn(),
+    replaceMessageText: vi.fn(),
+    truncateAfterMessage: vi.fn(),
+  });
+  clientGetMock.mockImplementation((url: string) => {
+    if (url === '/api/session/sess-1/message') {
+      return Promise.resolve({
+        data: {
+          items: fetchedMessages,
+          hasMore: false,
+          nextBefore: null,
+        },
+      });
+    }
+    if (url === '/api/session/status') {
+      return Promise.resolve({ data: status });
+    }
+    return Promise.resolve({ data: {} });
+  });
+}
+
+async function startFallbackPolling(onStreamingDone: () => void) {
+  render(React.createElement(SessionChat, {
+    sessionId: 'sess-1',
+    live: true,
+    onStreamingDone,
+  }));
+  await act(async () => {
+    await Promise.resolve();
+  });
+  clientGetMock.mockClear();
+  act(() => {
+    useSSEOptionsRef.current.onEvent({
+      type: 'session.status',
+      properties: { sessionID: 'sess-1', status: { type: 'busy' } },
+    });
+  });
+}
+
 function mockStatefulSessionMessages() {
   useSessionMessagesMock.mockImplementation(() => {
     const [messages, setMessages] = React.useState<Message[]>([]);
@@ -302,10 +407,15 @@ function mockStatefulSessionMessages() {
     return {
       messages,
       loading: false,
+      error: null,
       refetch: vi.fn(),
       addMessage: (message: Message) => setMessages((prev) => [...prev, message]),
       updateMessage: upsertMessage,
       updateMessagePart: vi.fn(),
+      removeMessage: (messageId: string) => setMessages((prev) => prev.filter(
+        (message) => message.id !== messageId,
+      )),
+      clearMessages: () => setMessages([]),
       replaceMessageText: vi.fn(),
       markMessageStopped: (messageId: string) => setMessages((prev) => prev.map(
         (message) => (message.id === messageId ? { ...message, finish: 'stop' } : message),
@@ -315,6 +425,120 @@ function mockStatefulSessionMessages() {
   });
 }
 
+describe('SessionChat message loading state', () => {
+  it('renders a retry action instead of an empty conversation after loading fails', async () => {
+    const user = userEvent.setup();
+    const refetch = vi.fn();
+    useSessionMessagesMock.mockReturnValue({
+      messages: [],
+      loading: false,
+      error: 'server unavailable',
+      refetch,
+      addMessage: vi.fn(),
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      removeMessage: vi.fn(),
+      clearMessages: vi.fn(),
+      replaceMessageText: vi.fn(),
+      markMessageStopped: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
+
+    expect(screen.getByRole('alert')).toHaveTextContent('消息加载失败');
+    expect(screen.queryByText('暂无消息')).not.toBeInTheDocument();
+    await user.click(screen.getByRole('button', { name: '重试' }));
+    expect(refetch).toHaveBeenCalledTimes(1);
+  });
+});
+
+describe('ChatToolPart file operation titles', () => {
+  it('shows only the filename in the header while preserving the full path in details', () => {
+    const fullPath = '/Users/example/.flocks/workspace/outputs/2026-07-24/gold_price_retrieval_plan.md';
+    const { container } = render(React.createElement(ChatToolPart, {
+      part: {
+        id: 'tool-write-file',
+        type: 'tool',
+        tool: 'write',
+        state: {
+          status: 'completed',
+          input: { filePath: fullPath, content: '# Plan' },
+          output: 'Wrote file successfully.',
+          title: fullPath,
+        },
+      } as Message['parts'][number],
+    }));
+
+    const header = container.querySelector('summary');
+    expect(header).toHaveTextContent('gold_price_retrieval_plan.md');
+    expect(header).not.toHaveTextContent(fullPath);
+    expect(container).toHaveTextContent(fullPath);
+  });
+
+  it('handles Windows-style file paths in file tool headers', () => {
+    const { container } = render(React.createElement(ChatToolPart, {
+      part: {
+        id: 'tool-read-file',
+        type: 'tool',
+        tool: 'read_file',
+        state: {
+          status: 'completed',
+          input: { path: 'C:\\workspace\\reports\\summary.txt' },
+          output: 'summary',
+        },
+      } as Message['parts'][number],
+    }));
+
+    expect(container.querySelector('summary')).toHaveTextContent('summary.txt');
+  });
+
+  it('shows only the first filename and file count for multi-file patches', () => {
+    const firstPath = '/repo/src/a.ts';
+    const { container } = render(React.createElement(ChatToolPart, {
+      part: {
+        id: 'tool-apply-patch',
+        type: 'tool',
+        tool: 'apply_patch',
+        state: {
+          status: 'completed',
+          input: {
+            patchText: [
+              '*** Begin Patch',
+              `*** Update File: ${firstPath}`,
+              '*** Add File: /repo/src/b.ts',
+              '*** End Patch',
+            ].join('\n'),
+          },
+          output: 'Done!',
+        },
+      } as Message['parts'][number],
+    }));
+
+    const header = container.querySelector('summary');
+    expect(header).toHaveTextContent('a.ts +1');
+    expect(header).not.toHaveTextContent(firstPath);
+    expect(container).toHaveTextContent(firstPath);
+  });
+
+  it('does not change summaries for non-file tools', () => {
+    const { container } = render(React.createElement(ChatToolPart, {
+      part: {
+        id: 'tool-custom',
+        type: 'tool',
+        tool: 'custom_tool',
+        state: {
+          status: 'completed',
+          input: { path: '/api/v1/incidents' },
+          output: 'ok',
+        },
+      } as Message['parts'][number],
+    }));
+
+    expect(container.querySelector('summary')).toHaveTextContent('path=/api/v1/incidents');
+  });
+});
+
 describe('dedupeUploadedDocumentAttachments', () => {
   it('keeps the latest successful document for a workspace path', () => {
     const items = dedupeUploadedDocumentAttachments([
@@ -454,6 +678,10 @@ describe('getMessageBubbleClassName', () => {
 
     expect(className).toContain('w-auto');
     expect(className.split(' ')).not.toContain('w-full');
+    expect(className).toContain('rounded-[18px]');
+    expect(className).toContain('py-3');
+    expect(className).not.toContain('rounded-[24px]');
+    expect(className).not.toContain('py-4');
   });
 
   it('expands editing user bubbles to full width in full layout', () => {
@@ -475,6 +703,7 @@ describe('getMessageBubbleClassName', () => {
     });
 
     expect(className).toContain('w-full');
+    expect(className).toContain('text-[15px]');
   });
 
   it('fills the fixed compact assistant message column', () => {
@@ -488,6 +717,21 @@ describe('getMessageBubbleClassName', () => {
     expect(className).toContain('max-w-full');
   });
 
+  it.each([false, true])('keeps assistant replies in the transparent content flow when compact=%s', (compact) => {
+    const className = getMessageBubbleClassName({
+      compact,
+      isUser: false,
+      isEditing: false,
+    });
+
+    expect(className).toContain('bg-transparent');
+    expect(className).not.toContain('bg-white');
+    expect(className).not.toContain('border-zinc-200/90');
+    expect(className).not.toContain('shadow-sm');
+    expect(className).not.toContain('rounded-[20px]');
+    expect(className).not.toContain('rounded-[24px]');
+  });
+
   it('keeps compact user bubbles content-sized when not editing', () => {
     const className = getMessageBubbleClassName({
       compact: true,
@@ -498,6 +742,19 @@ describe('getMessageBubbleClassName', () => {
     expect(className).toContain('max-w-full');
     expect(className.split(/\s+/)).not.toContain('w-full');
   });
+
+  it.each([false, true])('uses a neutral user bubble background when compact=%s', (compact) => {
+    const className = getMessageBubbleClassName({
+      compact,
+      isUser: true,
+      isEditing: false,
+    });
+
+    expect(className).toContain('bg-zinc-50');
+    expect(className).toContain(compact ? 'border-black/[0.07]' : 'border-black/[0.09]');
+    expect(className).not.toContain('bg-sky-50');
+    expect(className).not.toContain('border-sky-100');
+  });
 });
 
 describe('getMessageGroupClassName', () => {
@@ -544,7 +801,139 @@ describe('getMessageGroupClassName', () => {
   });
 });
 
+describe('SessionChat embedded full layout', () => {
+  it('reserves horizontal space for avatars in full-width panels', () => {
+    useSessionMessagesMock.mockReturnValue({
+      messages: [
+        makeMessage({
+          id: 'assistant-contained-avatar',
+          role: 'assistant',
+          parts: [{ id: 'text-1', type: 'text', text: 'Task complete.' } as any],
+        }),
+      ],
+      loading: false,
+      refetch: vi.fn(),
+      addMessage: vi.fn(),
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      removeMessage: vi.fn(),
+      replaceMessageText: vi.fn(),
+      markMessageStopped: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    const { container } = render(React.createElement(SessionChat, {
+      sessionId: 'sess-1',
+      hideInput: true,
+      display: {
+        compact: false,
+        fullWidth: true,
+        pageCanvas: true,
+      },
+    }));
+
+    expect(container.querySelector('.space-y-5.px-12')).toBeInTheDocument();
+  });
+});
+
 describe('SessionChat copy action', () => {
+  it('excludes hidden task metadata from copied text', async () => {
+    const user = userEvent.setup();
+    const writeText = vi.fn().mockResolvedValue(undefined);
+
+    Object.defineProperty(window, 'isSecureContext', {
+      configurable: true,
+      value: true,
+    });
+    Object.defineProperty(navigator, 'clipboard', {
+      configurable: true,
+      value: { writeText },
+    });
+
+    useSessionMessagesMock.mockReturnValue({
+      messages: [
+        makeMessage({
+          id: 'assistant-copy-filtered',
+          role: 'assistant',
+          parts: [{
+            id: 'text-1',
+            type: 'text',
+            text: 'copy this result\n\n<task_metadata>\nsession_id: ses-child\n</task_metadata>',
+          }],
+        }),
+      ],
+      loading: false,
+      refetch: vi.fn(),
+      addMessage: vi.fn(),
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      replaceMessageText: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    render(React.createElement(SessionChat, {
+      sessionId: 'sess-1',
+      display: { compact: false, showActions: true },
+    }));
+
+    await user.click(screen.getByRole('button', { name: 'chat.copy' }));
+    await waitFor(() => expect(writeText).toHaveBeenCalledOnce());
+
+    const copiedText = writeText.mock.calls[0][0];
+    expect(copiedText).toContain('copy this result');
+    expect(copiedText).not.toContain('task_metadata');
+    expect(copiedText).not.toContain('ses-child');
+  });
+
+  it('preserves user-authored task metadata in the bubble and copied text', async () => {
+    const user = userEvent.setup();
+    const writeText = vi.fn().mockResolvedValue(undefined);
+    const text = [
+      'Please explain this payload:',
+      '',
+      '```xml',
+      '<task_metadata>',
+      'session_id: ses-user-example',
+      '</task_metadata>',
+      '```',
+    ].join('\n');
+
+    Object.defineProperty(window, 'isSecureContext', {
+      configurable: true,
+      value: true,
+    });
+    Object.defineProperty(navigator, 'clipboard', {
+      configurable: true,
+      value: { writeText },
+    });
+
+    useSessionMessagesMock.mockReturnValue({
+      messages: [
+        makeMessage({
+          id: 'user-task-metadata-example',
+          role: 'user',
+          parts: [{ id: 'text-1', type: 'text', text }],
+        }),
+      ],
+      loading: false,
+      refetch: vi.fn(),
+      addMessage: vi.fn(),
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      replaceMessageText: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    render(React.createElement(SessionChat, {
+      sessionId: 'sess-1',
+      display: { compact: false, showActions: true },
+    }));
+
+    expect(screen.getByText(/session_id: ses-user-example/)).toBeInTheDocument();
+    await user.click(screen.getByRole('button', { name: 'chat.copy' }));
+    await waitFor(() => expect(writeText).toHaveBeenCalledWith(text));
+  });
+
   it('falls back when async clipboard is unavailable', async () => {
     const user = userEvent.setup();
     const execCommand = vi.fn().mockReturnValue(true);
@@ -882,9 +1271,104 @@ describe('SessionChat standalone thinking indicator', () => {
       consoleError.mockRestore();
     }
   });
+
+  it('clears local messages without refetching after session.cleared', () => {
+    const clearMessages = vi.fn();
+    const refetch = vi.fn();
+    useSessionMessagesMock.mockReturnValue({
+      messages: [makeMessage({ id: 'message-before-clear', role: 'assistant', parts: [] })],
+      loading: false,
+      refetch,
+      addMessage: vi.fn(),
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      removeMessage: vi.fn(),
+      clearMessages,
+      replaceMessageText: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
+
+    act(() => {
+      useSSEOptionsRef.current.onEvent({
+        type: 'session.cleared',
+        properties: { sessionID: 'sess-1' },
+      });
+    });
+
+    expect(clearMessages).toHaveBeenCalledOnce();
+    expect(refetch).not.toHaveBeenCalled();
+  });
+
+  it('removes an intermediate assistant when message.removed arrives', () => {
+    const removeMessage = vi.fn();
+    useSessionMessagesMock.mockReturnValue({
+      messages: [makeMessage({ id: 'assistant-failed', role: 'assistant', parts: [] })],
+      loading: false,
+      refetch: vi.fn(),
+      addMessage: vi.fn(),
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      removeMessage,
+      replaceMessageText: vi.fn(),
+      markMessageStopped: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
+
+    act(() => {
+      useSSEOptionsRef.current.onEvent({
+        type: 'message.removed',
+        properties: {
+          sessionID: 'sess-1',
+          messageID: 'assistant-failed',
+        },
+      });
+    });
+
+    expect(removeMessage).toHaveBeenCalledWith('assistant-failed');
+  });
 });
 
 describe('SessionChat instruction display text', () => {
+  it('filters internal task metadata from visible message text', () => {
+    useSessionMessagesMock.mockReturnValue({
+      messages: [
+        makeMessage({
+          id: 'assistant-with-task-metadata',
+          role: 'assistant',
+          parts: [{
+            id: 'assistant-with-task-metadata-part',
+            type: 'text',
+            text: [
+              '查询完成。',
+              '<task_metadata>',
+              'session_id: ses-child',
+              '</task_metadata>',
+              '消息发送失败。',
+            ].join('\n'),
+          }] as Message['parts'],
+        }),
+      ],
+      loading: false,
+      refetch: vi.fn(),
+      addMessage: vi.fn(),
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      replaceMessageText: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
+
+    expect(screen.getByText('查询完成。')).toBeInTheDocument();
+    expect(screen.getByText('消息发送失败。')).toBeInTheDocument();
+    expect(screen.queryByText(/task_metadata/)).not.toBeInTheDocument();
+    expect(screen.queryByText(/ses-child/)).not.toBeInTheDocument();
+  });
+
   it('renders metadata displayText while keeping the raw prompt out of the bubble', () => {
     useSessionMessagesMock.mockReturnValue({
       messages: [
@@ -931,18 +1415,136 @@ describe('SessionChat instruction display text', () => {
       );
     });
   });
-});
+});
+
+describe('SessionChat composer controls', () => {
+  it('enables Auto on an existing session before sending without a model override', async () => {
+    const user = userEvent.setup();
+    render(React.createElement(SessionChat, {
+      sessionId: 'sess-1',
+      modelAuto: true,
+      model: null,
+    }));
+
+    await user.type(screen.getByPlaceholderText('请输入消息'), 'continue{enter}');
+
+    await waitFor(() => {
+      expect(sessionApiUpdateMock).toHaveBeenCalledWith('sess-1', {
+        model_auto: true,
+        model_pinned: false,
+      });
+      expect(clientPostMock).toHaveBeenCalledWith(
+        '/api/session/sess-1/prompt_async',
+        expect.objectContaining({
+          executionMode: 'build',
+          messageID: expect.any(String),
+          parts: [{ type: 'text', text: 'continue' }],
+        }),
+      );
+    });
+    expect(sessionApiUpdateMock.mock.invocationCallOrder[0]).toBeLessThan(
+      clientPostMock.mock.invocationCallOrder[0],
+    );
+  });
+
+  it('keeps the disabled send button visible in dark mode', () => {
+    const { container } = render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
+
+    const disabledButtons = Array.from(container.querySelectorAll('button:disabled'));
+    const sendButton = disabledButtons.find((button) => button.querySelector('svg'));
+
+    expect(sendButton?.className).toContain('dark:bg-[#46515e]');
+    expect(sendButton?.className).toContain('dark:text-[#b8c2cc]');
+    expect(sendButton?.className).toContain('dark:border-[#5a6573]');
+  });
+
+  it('groups attachment and injected actions in the Add menu', async () => {
+    const user = userEvent.setup();
+    const onOpenChange = vi.fn();
+    render(React.createElement(SessionChat, {
+      sessionId: 'sess-1',
+      onComposerAddMenuOpenChange: onOpenChange,
+      composerAddMenuSlot: ({ insertMention, insertReference }) => React.createElement(
+        React.Fragment,
+        null,
+        React.createElement(
+          'button',
+          { type: 'button', onClick: () => insertMention('explore') },
+          '智能体',
+        ),
+        React.createElement(
+          'button',
+          { type: 'button', onClick: () => insertReference('diagnose', 'skill') },
+          '技能',
+        ),
+      ),
+    }));
+
+    await user.click(screen.getByRole('button', { name: '添加' }));
+
+    expect(screen.getByRole('menu', { name: '添加' })).toBeInTheDocument();
+    expect(screen.getByText('文件和图片')).toBeInTheDocument();
+    await user.click(screen.getByRole('button', { name: '智能体' }));
+    expect(within(screen.getByLabelText('已选择的资源')).getByText('explore')).toBeInTheDocument();
+    await user.click(screen.getByRole('button', { name: '技能' }));
+    expect(within(screen.getByLabelText('已选择的资源')).getByText('diagnose')).toBeInTheDocument();
+    expect(screen.getByPlaceholderText('请输入消息')).toHaveValue('');
+    expect(onOpenChange).toHaveBeenLastCalledWith(true);
+
+    await user.keyboard('{Escape}');
+
+    expect(screen.queryByRole('menu', { name: '添加' })).not.toBeInTheDocument();
+    expect(onOpenChange).toHaveBeenLastCalledWith(false);
+  });
 
-describe('SessionChat composer controls', () => {
-  it('keeps the disabled send button visible in dark mode', () => {
-    const { container } = render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
+  it('keeps a selected subagent reference when the default agent changes', async () => {
+    const user = userEvent.setup();
+    const mentionAgents = [
+      {
+        name: 'rex',
+        description: 'Main orchestrator',
+        mode: 'primary',
+        permission: [],
+        options: {},
+        skills: [],
+        tools: [],
+      },
+      {
+        name: 'explore',
+        description: 'Explore the codebase',
+        mode: 'subagent',
+        native: true,
+        permission: [],
+        options: {},
+        skills: [],
+        tools: [],
+      },
+    ];
+    function AgentMenuHarness() {
+      const [agentName, setAgentName] = React.useState('rex');
+      return React.createElement(SessionChat, {
+        sessionId: 'sess-1',
+        agentName,
+        mentionAgents,
+        composerAddMenuSlot: ({ insertMention }) => React.createElement(
+          'button',
+          {
+            type: 'button',
+            onClick: () => {
+              setAgentName('explore');
+              insertMention('explore');
+            },
+          },
+          '选择 Explore',
+        ),
+      });
+    }
+    render(React.createElement(AgentMenuHarness));
 
-    const disabledButtons = Array.from(container.querySelectorAll('button:disabled'));
-    const sendButton = disabledButtons.find((button) => button.querySelector('svg'));
+    await user.click(screen.getByRole('button', { name: '添加' }));
+    await user.click(screen.getByRole('button', { name: '选择 Explore' }));
 
-    expect(sendButton?.className).toContain('dark:bg-[#46515e]');
-    expect(sendButton?.className).toContain('dark:text-[#b8c2cc]');
-    expect(sendButton?.className).toContain('dark:border-[#5a6573]');
+    expect(within(screen.getByLabelText('已选择的资源')).getByText('explore')).toBeInTheDocument();
   });
 });
 
@@ -1450,14 +2052,20 @@ describe('SessionChat intermediate process collapse', () => {
 
     const processGroup = screen.getByTestId('chat-process-group') as HTMLDetailsElement;
     expect(processGroup.open).toBe(false);
-    expect(screen.getByText('过程（2 项）')).toBeInTheDocument();
-    expect(screen.getByText('1 段思考 · 1 次工具调用')).toBeInTheDocument();
+    expect(screen.getByText('查看 2 个步骤')).toBeInTheDocument();
+    expect(processGroup.querySelector('summary')).toHaveClass('text-sm', 'font-medium');
+    expect(processGroup.querySelector('summary')).not.toHaveClass('font-semibold');
+    expect(processGroup.className).not.toContain('rounded-lg');
+    expect(processGroup.closest('[data-process-output="true"]')?.className).not.toContain('bg-white');
     expect(screen.getByText('已读取当前 workflow.md。')).toBeInTheDocument();
 
-    await user.click(screen.getByText('过程（2 项）'));
+    await user.click(screen.getByText('查看 2 个步骤'));
 
     expect(processGroup.open).toBe(true);
-    expect(screen.getByText('read')).toBeInTheDocument();
+    expect(screen.getByTestId('chat-process-timeline')).toBeInTheDocument();
+    expect(screen.getByTestId('chat-process-reasoning-step')).toHaveTextContent('深度思考');
+    expect(screen.getByTestId('chat-process-reasoning-step').querySelector('button')).toHaveClass('text-sm');
+    expect(screen.getByTestId('chat-process-tool-step')).toHaveTextContent('读取文件');
   });
 
   it('opens process groups while an assistant message is active and collapses after completion', () => {
@@ -1507,7 +2115,8 @@ describe('SessionChat intermediate process collapse', () => {
     expect(processGroup.open).toBe(false);
   });
 
-  it('uses one global process group and leaves only the final reply visible', () => {
+  it('uses one global process group and renders intermediate summaries as peer steps', async () => {
+    const user = userEvent.setup();
     useSessionMessagesMock.mockReturnValue({
       messages: [
         makeMessage({
@@ -1589,9 +2198,18 @@ describe('SessionChat intermediate process collapse', () => {
     const processGroup = screen.getByTestId('chat-process-group') as HTMLDetailsElement;
     expect(screen.getAllByTestId('chat-process-group')).toHaveLength(1);
     expect(processGroup.open).toBe(false);
-    expect(screen.getByText('过程（5 项）')).toBeInTheDocument();
-    expect(screen.getByText('2 段思考 · 2 次工具调用 · 1 段中间回复')).toBeInTheDocument();
+    expect(screen.getByText('查看 5 个步骤')).toBeInTheDocument();
     expect(screen.getByText('最终结果已生成。')).toBeVisible();
+
+    await user.click(screen.getByText('查看 5 个步骤'));
+
+    const summaryStep = screen.getByTestId('chat-process-text-step');
+    const reasoningStep = screen.getAllByTestId('chat-process-reasoning-step')[0];
+    expect(summaryStep).toHaveTextContent('已经读取文件，继续检查相关配置。');
+    expect(summaryStep.querySelector('svg')).not.toBeInTheDocument();
+    expect(summaryStep.className).not.toContain('pl-');
+    expect(summaryStep.querySelector('.border-l')).not.toBeInTheDocument();
+    expect(summaryStep.parentElement?.parentElement).toBe(reasoningStep.parentElement?.parentElement);
   });
 
   it('keeps a user-opened process group open when new tool parts arrive', async () => {
@@ -1656,7 +2274,7 @@ describe('SessionChat intermediate process collapse', () => {
     }));
 
     const processGroup = screen.getByTestId('chat-process-group') as HTMLDetailsElement;
-    await user.click(screen.getByText('过程（2 项）'));
+    await user.click(screen.getByText('查看 2 个步骤'));
     expect(processGroup.open).toBe(true);
 
     useSessionMessagesMock.mockReturnValue({
@@ -1675,10 +2293,10 @@ describe('SessionChat intermediate process collapse', () => {
       display: { collapseIntermediateSteps: true },
     }));
 
-    expect(screen.getByText('过程（3 项）')).toBeInTheDocument();
+    expect(screen.getByText('查看 3 个步骤')).toBeInTheDocument();
     expect(screen.getByTestId('chat-process-group')).toBe(processGroup);
     expect(processGroup.open).toBe(true);
-    expect(screen.getByText('write')).toBeVisible();
+    expect(screen.getAllByTestId('chat-process-tool-step')).toHaveLength(2);
   });
 
   it('restores a user-opened process group after switching away from a keyed session chat', async () => {
@@ -1743,7 +2361,7 @@ describe('SessionChat intermediate process collapse', () => {
     });
     const { rerender } = render(renderKeyedChat('sess-1'));
 
-    await user.click(screen.getByText('过程（2 项）'));
+    await user.click(screen.getByText('查看 2 个步骤'));
     expect(screen.getByTestId('chat-process-group')).toHaveProperty('open', true);
 
     rerender(renderKeyedChat('sess-2'));
@@ -1753,7 +2371,7 @@ describe('SessionChat intermediate process collapse', () => {
 
     const restoredProcessGroup = screen.getByTestId('chat-process-group') as HTMLDetailsElement;
     expect(restoredProcessGroup.open).toBe(true);
-    expect(screen.getByText('read')).toBeVisible();
+    expect(screen.getByTestId('chat-process-tool-step').querySelector('summary')).toBeVisible();
   });
 
   it('keeps pending questions visible and folds answered questions into the process group', async () => {
@@ -1841,7 +2459,7 @@ describe('SessionChat intermediate process collapse', () => {
 
     expect(screen.getByText('需要你的回答')).toBeVisible();
     expect(screen.getByText('选择范围')).toBeVisible();
-    expect(screen.getByText('过程（1 项）')).toBeInTheDocument();
+    expect(screen.getByText('查看 1 个步骤')).toBeInTheDocument();
 
     pendingQuestionsHookMock.pendingQuestions = {};
     useSessionMessagesMock.mockReturnValue({
@@ -1861,26 +2479,24 @@ describe('SessionChat intermediate process collapse', () => {
     }));
 
     expect(screen.queryByText('需要你的回答')).not.toBeInTheDocument();
-    expect(screen.getByText('过程（2 项）')).toBeInTheDocument();
-    expect(screen.getByText('1 段思考 · 1 次工具调用')).toBeInTheDocument();
-    expect(screen.queryByText('question')).not.toBeVisible();
+    expect(screen.getByText('查看 2 个步骤')).toBeInTheDocument();
+    expect(screen.getByText('向用户提问')).not.toBeVisible();
     expect(screen.queryByText('输入参数')).not.toBeInTheDocument();
     expect(screen.queryByText('输出结果')).not.toBeInTheDocument();
 
-    await user.click(screen.getByText('过程（2 项）'));
+    await user.click(screen.getByText('查看 2 个步骤'));
 
-    expect(screen.getByText('question')).toBeVisible();
-    expect(screen.getByText('已完成')).toBeVisible();
+    expect(screen.getByText('向用户提问')).toBeVisible();
     expect(screen.getByText('问题')).not.toBeVisible();
 
-    await user.click(screen.getByText('question'));
+    await user.click(screen.getByText('向用户提问'));
 
     expect(screen.getByText('问题')).toBeVisible();
     expect(screen.getByText('回答')).toBeVisible();
     expect(screen.getAllByText('测试范围').length).toBeGreaterThan(0);
     expect(screen.getByText('是')).toBeVisible();
 
-    await user.click(screen.getByText('过程（2 项）'));
+    await user.click(screen.getByText('查看 2 个步骤'));
 
     useSessionMessagesMock.mockReturnValue({
       messages: [makeQuestionMessage('completed', true)],
@@ -1899,8 +2515,8 @@ describe('SessionChat intermediate process collapse', () => {
     }));
 
     expect(screen.getByText('已按你的选择继续处理。')).toBeVisible();
-    expect(screen.getByText('过程（2 项）')).toBeInTheDocument();
-    expect(screen.getByText('question')).not.toBeVisible();
+    expect(screen.getByText('查看 2 个步骤')).toBeInTheDocument();
+    expect(screen.getByText('向用户提问')).not.toBeVisible();
   });
 
   it('renders collapsed process groups inside the full compact assistant column', () => {
@@ -2000,7 +2616,7 @@ describe('SessionChat intermediate process collapse', () => {
     const processGroup = screen.getByTestId('chat-process-group') as HTMLDetailsElement;
     expect(processGroup.open).toBe(true);
 
-    await user.click(screen.getByText('过程（2 项）'));
+    await user.click(screen.getByText('查看 2 个步骤'));
 
     expect(processGroup.open).toBe(false);
   });
@@ -2097,8 +2713,7 @@ describe('SessionChat intermediate process collapse', () => {
     }));
 
     expect(screen.getAllByTestId('chat-process-group')).toHaveLength(1);
-    expect(screen.getByText('过程（4 项）')).toBeInTheDocument();
-    expect(screen.getByText('2 段思考 · 2 次工具调用')).toBeInTheDocument();
+    expect(screen.getByText('查看 4 个步骤')).toBeInTheDocument();
   });
 
   it('keeps the compact compaction bubble at the full assistant column width', async () => {
@@ -2149,7 +2764,122 @@ describe('SessionChat intermediate process collapse', () => {
   });
 });
 
-describe('SessionChat agent mentions', () => {
+describe('SessionChat optimistic message identity', () => {
+  it.each([
+    ['prompt', 'message that fails', '/api/session/sess-1/prompt_async'],
+    ['slash command', '/tools', '/api/session/sess-1/command'],
+  ])('removes the optimistic message when a %s request fails', async (_kind, input, endpoint) => {
+    const user = userEvent.setup();
+    const addMessage = vi.fn();
+    const removeMessage = vi.fn();
+    const alertSpy = vi.spyOn(window, 'alert').mockImplementation(() => {});
+    clientPostMock.mockRejectedValueOnce(new Error('network down'));
+    useSessionMessagesMock.mockReturnValue({
+      messages: [],
+      loading: false,
+      refetch: vi.fn(),
+      addMessage,
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      removeMessage,
+      replaceMessageText: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    try {
+      render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
+
+      await user.type(screen.getByPlaceholderText('请输入消息'), `${input}{enter}`);
+
+      await waitFor(() => expect(clientPostMock).toHaveBeenCalledWith(
+        endpoint,
+        expect.any(Object),
+      ));
+      await waitFor(() => expect(removeMessage).toHaveBeenCalledOnce());
+
+      const optimisticMessage = addMessage.mock.calls[0][0] as Message;
+      expect(removeMessage).toHaveBeenCalledWith(optimisticMessage.id);
+    } finally {
+      alertSpy.mockRestore();
+    }
+  });
+
+  it('uses the optimistic user message ID for the persisted prompt', async () => {
+    const user = userEvent.setup();
+    const addMessage = vi.fn();
+    useSessionMessagesMock.mockReturnValue({
+      messages: [],
+      loading: false,
+      refetch: vi.fn(),
+      addMessage,
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      replaceMessageText: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
+
+    await user.type(screen.getByPlaceholderText('请输入消息'), '测试 question 工具{enter}');
+
+    await waitFor(() => {
+      expect(clientPostMock).toHaveBeenCalledWith(
+        '/api/session/sess-1/prompt_async',
+        expect.objectContaining({ messageID: expect.stringMatching(/^msg_/) }),
+      );
+    });
+
+    expect(addMessage).toHaveBeenCalledTimes(1);
+    expect(clientPostMock.mock.calls.filter(
+      ([url]) => url === '/api/session/sess-1/prompt_async',
+    )).toHaveLength(1);
+    const optimisticMessage = addMessage.mock.calls[0][0] as Message;
+    const promptCall = clientPostMock.mock.calls.find(
+      ([url]) => url === '/api/session/sess-1/prompt_async',
+    );
+    const payload = promptCall?.[1] as { messageID?: string } | undefined;
+    expect(payload?.messageID).toBe(optimisticMessage.id);
+  });
+
+  it('uses the optimistic user message ID for slash commands', async () => {
+    const user = userEvent.setup();
+    const addMessage = vi.fn();
+    useSessionMessagesMock.mockReturnValue({
+      messages: [],
+      loading: false,
+      refetch: vi.fn(),
+      addMessage,
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      replaceMessageText: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
+
+    await user.type(screen.getByPlaceholderText('请输入消息'), '/tools{enter}');
+
+    await waitFor(() => {
+      expect(clientPostMock).toHaveBeenCalledWith(
+        '/api/session/sess-1/command',
+        expect.objectContaining({ messageID: expect.stringMatching(/^msg_/) }),
+      );
+    });
+
+    expect(addMessage).toHaveBeenCalledTimes(1);
+    expect(clientPostMock.mock.calls.filter(
+      ([url]) => url === '/api/session/sess-1/command',
+    )).toHaveLength(1);
+    const optimisticMessage = addMessage.mock.calls[0][0] as Message;
+    const commandCall = clientPostMock.mock.calls.find(
+      ([url]) => url === '/api/session/sess-1/command',
+    );
+    const payload = commandCall?.[1] as { messageID?: string } | undefined;
+    expect(payload?.messageID).toBe(optimisticMessage.id);
+  });
+});
+
+describe('SessionChat composer references', () => {
   const mentionAgents = [
     {
       name: 'rex',
@@ -2174,20 +2904,20 @@ describe('SessionChat agent mentions', () => {
     },
   ];
 
-  it('shows matching agents when typing @', async () => {
+  it('opens the same Add menu when typing @', async () => {
     const user = userEvent.setup();
     render(React.createElement(SessionChat, {
       sessionId: 'sess-1',
       mentionAgents,
     }));
 
-    await user.type(screen.getByPlaceholderText('请输入消息'), '@ex');
+    await user.type(screen.getByPlaceholderText('请输入消息'), '@');
 
-    expect(screen.getByText('@explore')).toBeInTheDocument();
-    expect(screen.getByText('探索代码库')).toBeInTheDocument();
+    expect(screen.getByRole('menu', { name: '添加' })).toBeInTheDocument();
+    expect(screen.getByPlaceholderText('请输入消息')).toHaveValue('');
   });
 
-  it('routes one message to the mentioned agent without changing the default agent', async () => {
+  it('routes one message to a structured subagent reference', async () => {
     const user = userEvent.setup();
     render(React.createElement(SessionChat, {
       sessionId: 'sess-1',
@@ -2195,7 +2925,10 @@ describe('SessionChat agent mentions', () => {
       mentionAgents,
     }));
 
-    await user.type(screen.getByPlaceholderText('请输入消息'), '@explore summarize this file{enter}');
+    await user.type(
+      screen.getByPlaceholderText('请输入消息'),
+      'subagent:explore summarize this file{enter}',
+    );
 
     await waitFor(() => {
       expect(clientPostMock).toHaveBeenCalledWith(
@@ -2208,6 +2941,61 @@ describe('SessionChat agent mentions', () => {
     });
   });
 
+  it('sends selected resources using the unified structured syntax', async () => {
+    const user = userEvent.setup();
+    const onCreateAndSend = vi.fn().mockResolvedValue('sess-created');
+    render(React.createElement(SessionChat, {
+      sessionId: null,
+      agentName: 'rex',
+      mentionAgents,
+      onCreateAndSend,
+      composerAddMenuSlot: ({ insertMention, insertReference }) => React.createElement(
+        React.Fragment,
+        null,
+        React.createElement(
+          'button',
+          { type: 'button', onClick: () => insertMention('explore') },
+          '选择子智能体',
+        ),
+        React.createElement(
+          'button',
+          { type: 'button', onClick: () => insertReference('diagnose', 'skill') },
+          '选择技能',
+        ),
+        React.createElement(
+          'button',
+          { type: 'button', onClick: () => insertReference('code-review', 'skill') },
+          '选择第二个技能',
+        ),
+        React.createElement(
+          'button',
+          { type: 'button', onClick: () => insertReference('triage', 'workflow') },
+          '选择工作流',
+        ),
+      ),
+    }));
+
+    await user.click(screen.getByRole('button', { name: '添加' }));
+    await user.click(screen.getByRole('button', { name: '选择子智能体' }));
+    await user.click(screen.getByRole('button', { name: '选择技能' }));
+    await user.click(screen.getByRole('button', { name: '选择第二个技能' }));
+    await user.click(screen.getByRole('button', { name: '选择工作流' }));
+    expect(within(screen.getByLabelText('已选择的资源')).getByText('diagnose')).toBeInTheDocument();
+    expect(within(screen.getByLabelText('已选择的资源')).getByText('code-review')).toBeInTheDocument();
+    await user.type(screen.getByPlaceholderText('请输入消息'), 'check this{enter}');
+
+    await waitFor(() => {
+      expect(onCreateAndSend).toHaveBeenCalledWith(
+        'workflow:triage skill:code-review skill:diagnose subagent:explore check this',
+        [],
+        'explore',
+        undefined,
+        undefined,
+        'build',
+      );
+    });
+  });
+
   it('uses the selected agent when creating a session from the first message', async () => {
     const user = userEvent.setup();
     const onCreateAndSend = vi.fn().mockResolvedValue('sess-created');
@@ -2226,6 +3014,8 @@ describe('SessionChat agent mentions', () => {
         [],
         'explore',
         undefined,
+        undefined,
+        'build',
       );
     });
   });
@@ -2247,7 +3037,7 @@ describe('SessionChat agent mentions', () => {
     });
 
     sessionApiEnqueuePromptMock.mockClear();
-    await user.type(screen.getByRole('textbox'), '@explore queued message{enter}');
+    await user.type(screen.getByRole('textbox'), 'subagent:explore queued message{enter}');
 
     await waitFor(() => {
       expect(sessionApiEnqueuePromptMock).toHaveBeenCalledWith(
@@ -2345,28 +3135,324 @@ describe('buildTodoSummary', () => {
     })).toBe('Progress 1/3 · In progress 1');
   });
 
-  it('prefers current metadata todos when available', () => {
-    expect(buildTodoSummary({
-      metadata: {
-        oldTodos: [
-          { id: '1', content: '定位 todo 摘要问题', status: 'pending' },
-          { id: '2', content: '补充回归测试', status: 'pending' },
-        ],
-        newTodos: [
-          { id: '1', content: '定位 todo 摘要问题', status: 'completed' },
-          { id: '3', content: '验证 Web UI 展示', status: 'completed' },
-        ],
-      },
-    })).toBe('Completed 2/2');
+  it('prefers current metadata todos when available', () => {
+    expect(buildTodoSummary({
+      metadata: {
+        oldTodos: [
+          { id: '1', content: '定位 todo 摘要问题', status: 'pending' },
+          { id: '2', content: '补充回归测试', status: 'pending' },
+        ],
+        newTodos: [
+          { id: '1', content: '定位 todo 摘要问题', status: 'completed' },
+          { id: '3', content: '验证 Web UI 展示', status: 'completed' },
+        ],
+      },
+    })).toBe('Completed 2/2');
+  });
+
+  it('renders a readable fallback for todo actions without structured entries', () => {
+    expect(buildTodoSummary({
+      input: {
+        action: 'write',
+        todos: [],
+      },
+    })).toBe('Update todos');
+  });
+});
+
+describe('ChatToolPart delegate rendering', () => {
+  it('keeps the specialized delegate view inside a process timeline', () => {
+    render(
+      React.createElement(ChatToolPart, {
+        processStep: true,
+        part: {
+          id: 'delegate-process-step',
+          type: 'tool',
+          tool: 'delegate_task',
+          state: {
+            status: 'running',
+            input: {
+              subagent_type: 'explore',
+              description: '排查会话页面',
+            },
+          },
+        } as any,
+      }),
+    );
+
+    expect(screen.getByTestId('chat-process-delegate-step')).toBeInTheDocument();
+    expect(screen.queryByTestId('chat-process-tool-step')).not.toBeInTheDocument();
+  });
+});
+
+describe('ChatToolPart load skill rendering', () => {
+  const part = {
+    id: 'load-skill-part',
+    type: 'tool',
+    tool: 'skill_load',
+    callID: 'call-load-skill',
+    state: {
+      status: 'completed',
+      input: {
+        name: 'agent-builder',
+      },
+      title: 'Loaded skill: agent-builder',
+      output: 'Skill loaded',
+    },
+  } as any;
+
+  it('uses the localized action name in the process timeline', () => {
+    render(React.createElement(ChatToolPart, { part, processStep: true }));
+
+    const processStep = screen.getByTestId('chat-process-tool-step');
+    expect(processStep.querySelector('summary')).toHaveTextContent('加载技能');
+    expect(processStep.querySelector('summary')).toHaveTextContent('agent-builder');
+    expect(processStep.querySelector('summary')).not.toHaveTextContent('Loaded skill');
+    expect(processStep.querySelector('summary')).not.toHaveTextContent('已完成');
+  });
+
+  it('uses the localized action name in the expanded tool card', () => {
+    const { container } = render(React.createElement(ChatToolPart, { part }));
+
+    expect(container.querySelector('summary')).toHaveTextContent('加载技能');
+    expect(container.querySelector('summary')).not.toHaveTextContent('load skill');
+  });
+});
+
+describe('ChatToolPart semantic tool presentation', () => {
+  it('shows the tool immediately and fills in parameters when they arrive', () => {
+    const initialPart = {
+      id: 'streaming-write-part',
+      type: 'tool',
+      tool: 'write',
+      state: {
+        status: 'pending',
+        input: {},
+      },
+    } as any;
+    const { rerender } = render(
+      React.createElement(ChatToolPart, {
+        processStep: true,
+        part: initialPart,
+      }),
+    );
+
+    let summary = screen.getByTestId('chat-process-tool-step').querySelector('summary');
+    expect(summary).toHaveTextContent('写入文件');
+    expect(summary).not.toHaveTextContent('filePath');
+    expect(screen.queryByText('输入参数')).not.toBeInTheDocument();
+    expect(screen.getByTestId('chat-tool-action-progress')).toHaveTextContent(
+      '正在写入文件…',
+    );
+
+    rerender(
+      React.createElement(ChatToolPart, {
+        processStep: true,
+        part: {
+          ...initialPart,
+          state: {
+            status: 'running',
+            input: {
+              filePath: '/repo/report.md',
+              content: '# Report',
+            },
+          },
+        },
+      }),
+    );
+
+    summary = screen.getByTestId('chat-process-tool-step').querySelector('summary');
+    expect(summary).toHaveTextContent('写入文件');
+    expect(summary).toHaveTextContent('report.md');
+    expect(summary).not.toHaveTextContent('/repo/report.md');
+    expect(screen.getByText('输入参数')).toBeInTheDocument();
+    expect(screen.queryByTestId('chat-tool-action-progress')).not.toBeInTheDocument();
+  });
+
+  it.each(['edit', 'apply_patch'])('describes a pending %s tool as editing the file', (tool) => {
+    render(
+      React.createElement(ChatToolPart, {
+        processStep: true,
+        part: {
+          id: `streaming-${tool}-part`,
+          type: 'tool',
+          tool,
+          state: {
+            status: 'pending',
+            input: {},
+          },
+        } as any,
+      }),
+    );
+
+    expect(screen.getByTestId('chat-tool-action-progress')).toHaveTextContent(
+      '正在编辑文件…',
+    );
+  });
+
+  it('shows a localized action and concise target in the process timeline', () => {
+    render(
+      React.createElement(ChatToolPart, {
+        processStep: true,
+        part: {
+          id: 'read-file-part',
+          type: 'tool',
+          tool: 'read',
+          state: {
+            status: 'completed',
+            input: {
+              filePath: '/repo/webui/src/components/common/SessionChat.tsx',
+              offset: 5200,
+              limit: 200,
+            },
+          },
+        } as any,
+      }),
+    );
+
+    const summary = screen.getByTestId('chat-process-tool-step').querySelector('summary');
+    expect(summary).toHaveTextContent('读取文件');
+    expect(summary).toHaveTextContent('SessionChat.tsx');
+    expect(summary).not.toHaveTextContent('/repo/webui/src/components/common/SessionChat.tsx');
+    expect(summary).not.toHaveTextContent('已完成');
+    expect(summary).not.toHaveTextContent('filePath=');
+  });
+
+  it('uses a subcommand-specific action for skill management', () => {
+    render(
+      React.createElement(ChatToolPart, {
+        processStep: true,
+        part: {
+          id: 'install-skill-part',
+          type: 'tool',
+          tool: 'flocks_skills',
+          state: {
+            status: 'running',
+            input: {
+              subcommand: 'install',
+              args: 'agent-builder',
+            },
+          },
+        } as any,
+      }),
+    );
+
+    const summary = screen.getByTestId('chat-process-tool-step').querySelector('summary');
+    expect(summary).toHaveTextContent('安装技能');
+    expect(summary).toHaveTextContent('agent-builder');
+    expect(summary).not.toHaveTextContent('执行中');
+  });
+
+  it('redacts sensitive fields in expanded input parameters', () => {
+    const { container } = render(
+      React.createElement(ChatToolPart, {
+        part: {
+          id: 'add-provider-part',
+          type: 'tool',
+          tool: 'add_provider',
+          state: {
+            status: 'completed',
+            input: {
+              name: 'Internal LLM',
+              api_key: 'super-secret-key',
+              config: {
+                password: 'super-secret-password',
+                base_url: 'https://models.example.com/v1',
+              },
+            },
+          },
+        } as any,
+      }),
+    );
+
+    expect(container.textContent).toContain('添加模型服务');
+    const payload = container.querySelector('pre')?.textContent || '';
+    expect(payload).toContain('••••••');
+    expect(payload).toContain('https://models.example.com/v1');
+    expect(payload).not.toContain('super-secret-key');
+    expect(payload).not.toContain('super-secret-password');
+  });
+});
+
+describe('ChatMessageBubble session typography', () => {
+  it('gives the full-layout agent label a clear heading level', () => {
+    render(React.createElement(ChatMessageBubble, {
+      message: makeMessage({
+        id: 'assistant-agent-label',
+        role: 'assistant',
+        parts: [{
+          id: 'assistant-text',
+          type: 'text',
+          text: '已加载技能。',
+        } as any],
+      }),
+      compact: false,
+    }));
+
+    expect(screen.getByText('Rex')).toHaveClass('text-base', 'font-semibold', 'text-[#3f444a]');
+  });
+});
+
+describe('ChatMessageBubble footer layout', () => {
+  it('places assistant actions immediately before the timestamp', () => {
+    render(React.createElement(ChatMessageBubble, {
+      message: makeMessage({
+        id: 'assistant-footer',
+        role: 'assistant',
+        timestamp: Date.now(),
+        parts: [{
+          id: 'assistant-text',
+          type: 'text',
+          text: '任务已完成。',
+        } as any],
+      }),
+      compact: false,
+      showActions: true,
+      showTimestamp: true,
+    }));
+
+    const regenerateButton = screen.getByRole('button', { name: 'chat.regenerate' });
+    const actionGroup = regenerateButton.parentElement;
+    const footer = actionGroup?.parentElement;
+
+    expect(footer).toHaveClass('justify-start', 'gap-1.5');
+    expect(footer?.children[0]).toBe(actionGroup);
+    expect(footer?.children[1]).toHaveClass('text-[11px]', 'text-[#8b929d]');
+    expect(regenerateButton).toHaveClass('text-[#8b929d]');
+    expect(regenerateButton).toHaveClass(
+      'border-transparent',
+      'bg-transparent',
+      'hover:bg-white',
+      'active:bg-white',
+      'focus-visible:bg-white',
+    );
+    expect(regenerateButton).not.toHaveClass('border-gray-200/80', 'bg-white/80');
   });
 
-  it('renders a readable fallback for todo actions without structured entries', () => {
-    expect(buildTodoSummary({
-      input: {
-        action: 'write',
-        todos: [],
-      },
-    })).toBe('Update todos');
+  it('keeps the user timestamp before the action group', () => {
+    render(React.createElement(ChatMessageBubble, {
+      message: makeMessage({
+        id: 'user-footer',
+        role: 'user',
+        timestamp: Date.now(),
+        parts: [{
+          id: 'user-text',
+          type: 'text',
+          text: '继续处理。',
+        } as any],
+      }),
+      compact: false,
+      showActions: true,
+      showTimestamp: true,
+    }));
+
+    const copyButton = screen.getByRole('button', { name: 'chat.copy' });
+    const actionGroup = copyButton.parentElement;
+    const footer = actionGroup?.parentElement;
+
+    expect(footer).toHaveClass('justify-between');
+    expect(footer?.children[0]).toHaveClass('text-[11px]');
+    expect(footer?.children[1]).toBe(actionGroup);
   });
 });
 
@@ -2444,7 +3530,7 @@ describe('ChatToolPart bash rendering', () => {
     );
 
     const text = container.textContent || '';
-    expect(text).toContain('bash');
+    expect(text).toContain('执行命令');
     expect(text).toContain('运行会话组件测试');
     expect(text).toContain('npm run test:run -- SessionChat.test.ts');
     expect(text).toContain('命令');
@@ -2526,7 +3612,9 @@ describe('SessionChat context usage popover', () => {
     render(React.createElement(SessionChat, { sessionId: 'sess-1' }));
 
     const contextButton = await screen.findByRole('button', { name: 'chat.contextUsageTitle' });
-    expect(contextButton).toHaveClass('h-6', 'w-6');
+    expect(contextButton).toHaveClass('h-6');
+    expect(contextButton).not.toHaveClass('w-6');
+    expect(contextButton).toHaveTextContent('12%');
     await user.click(contextButton);
 
     expect(screen.getByText('System prompt')).toBeInTheDocument();
@@ -2727,6 +3815,28 @@ describe('SessionChat context usage popover', () => {
 });
 
 describe('SessionChat goal banner', () => {
+  it('shows a Goal-mode objective as soon as the request is accepted', async () => {
+    const user = userEvent.setup();
+    render(React.createElement(SessionChat, {
+      sessionId: 'sess-1',
+      executionMode: 'goal',
+    }));
+
+    await user.type(screen.getByPlaceholderText('请输入消息'), '实现解析项目下.flocks{enter}');
+
+    await waitFor(() => {
+      expect(clientPostMock).toHaveBeenCalledWith(
+        '/api/session/sess-1/prompt_async',
+        expect.objectContaining({
+          executionMode: 'goal',
+          parts: [{ type: 'text', text: '实现解析项目下.flocks' }],
+        }),
+      );
+    });
+    expect(await screen.findByText('Goal')).toBeInTheDocument();
+    expect(screen.getByText('实现解析项目下.flocks')).toBeInTheDocument();
+  });
+
   it('hydrates a persisted goal banner when the session loads', async () => {
     sessionApiGetMock.mockResolvedValue({
       id: 'sess-1',
@@ -3001,6 +4111,126 @@ describe('streaming activity helpers', () => {
   });
 });
 
+describe('SessionChat SSE reconnect recovery', () => {
+  async function renderStreamingSession(onStreamingDone = vi.fn()) {
+    const refetch = vi.fn();
+    useSessionMessagesMock.mockReturnValue({
+      messages: [
+        makeMessage({
+          id: 'assistant-1',
+          finish: null,
+          parts: [{
+            id: 'tool-1',
+            type: 'tool',
+            tool: 'bash',
+            state: {
+              status: 'error',
+              error: 'Interrupted by server restart',
+            },
+          }] as Message['parts'],
+        }),
+      ],
+      loading: false,
+      refetch,
+      addMessage: vi.fn(),
+      updateMessage: vi.fn(),
+      updateMessagePart: vi.fn(),
+      replaceMessageText: vi.fn(),
+      truncateAfterMessage: vi.fn(),
+    });
+
+    render(React.createElement(SessionChat, {
+      sessionId: 'sess-1',
+      live: true,
+      onStreamingDone,
+    }));
+
+    await waitFor(() => {
+      expect(clientGetMock).toHaveBeenCalledWith('/api/session/status');
+    });
+    clientGetMock.mockClear();
+
+    act(() => {
+      useSSEOptionsRef.current.onEvent({
+        type: 'session.status',
+        properties: { sessionID: 'sess-1', status: { type: 'busy' } },
+      });
+    });
+    expect(screen.getByTitle('chat.stopTitle')).toBeInTheDocument();
+
+    return { onStreamingDone, refetch };
+  }
+
+  it('clears streaming after reconnect when the restarted server reports the session idle', async () => {
+    const { onStreamingDone, refetch } = await renderStreamingSession();
+
+    act(() => {
+      useSSEOptionsRef.current.onReconnect();
+    });
+
+    await waitFor(() => {
+      expect(clientGetMock).toHaveBeenCalledWith('/api/session/status');
+      expect(screen.queryByTitle('chat.stopTitle')).not.toBeInTheDocument();
+    });
+    expect(refetch).toHaveBeenCalled();
+    expect(onStreamingDone).toHaveBeenCalledTimes(1);
+  });
+
+  it('keeps streaming after reconnect while a long-running session is still busy', async () => {
+    const { onStreamingDone } = await renderStreamingSession();
+    clientGetMock.mockImplementation((url: string) => {
+      if (url === '/api/session/status') {
+        return Promise.resolve({ data: { 'sess-1': { type: 'busy' } } });
+      }
+      return Promise.resolve({ data: {} });
+    });
+
+    act(() => {
+      useSSEOptionsRef.current.onReconnect();
+    });
+
+    await waitFor(() => {
+      expect(clientGetMock).toHaveBeenCalledWith('/api/session/status');
+    });
+    expect(screen.getByTitle('chat.stopTitle')).toBeInTheDocument();
+    expect(onStreamingDone).not.toHaveBeenCalled();
+  });
+
+  it('ignores a stale idle response after a newer busy event arrives', async () => {
+    let resolveStatus: ((value: { data: Record<string, unknown> }) => void) | undefined;
+    const { onStreamingDone } = await renderStreamingSession();
+    clientGetMock.mockImplementation((url: string) => {
+      if (url === '/api/session/status') {
+        return new Promise((resolve) => {
+          resolveStatus = resolve;
+        });
+      }
+      return Promise.resolve({ data: {} });
+    });
+
+    act(() => {
+      useSSEOptionsRef.current.onReconnect();
+    });
+    await waitFor(() => {
+      expect(resolveStatus).toBeDefined();
+    });
+
+    act(() => {
+      useSSEOptionsRef.current.onEvent({
+        type: 'session.status',
+        properties: { sessionID: 'sess-1', status: { type: 'busy' } },
+      });
+      resolveStatus?.({ data: {} });
+    });
+
+    await act(async () => {
+      await Promise.resolve();
+    });
+    expect(screen.getByTitle('chat.stopTitle')).toBeInTheDocument();
+    expect(onStreamingDone).not.toHaveBeenCalled();
+  });
+});
+
 describe('SessionChat fallback polling', () => {
   it('reconciles pending questions while the session is busy', async () => {
     vi.useFakeTimers();
@@ -3033,65 +4263,228 @@ describe('SessionChat fallback polling', () => {
     }
   });
 
-  it('does not finish streaming while fetched messages still contain a running tool', async () => {
+  it('refetches when polling finds a pending question missing from local messages', async () => {
+    vi.useFakeTimers();
+    let resolveRefetch: (() => void) | undefined;
+    const refetch = vi.fn(() => new Promise<void>((resolve) => {
+      resolveRefetch = resolve;
+    }));
+    const onStreamingDone = vi.fn();
+    try {
+      pendingQuestionsHookMock.pendingQuestions = {
+        'call-question-1': {
+          requestId: 'request-question-1',
+          questions: [{ question: 'Continue?' }],
+        },
+      };
+      mockFallbackPolling({
+        localMessages: [
+          makeMessage({ id: 'user-1', role: 'user' }),
+          makeMessage({ id: 'assistant-1', parentID: 'user-1' }),
+        ],
+        fetchedMessages: [
+          makeFetchedMessage({ id: 'user-1', role: 'user' }),
+          makeFetchedMessage(
+            { id: 'assistant-1', role: 'assistant', parentID: 'user-1', finish: 'tool-calls' },
+            [{
+              id: 'question-tool-1',
+              messageID: 'assistant-1',
+              sessionID: 'sess-1',
+              type: 'tool',
+              tool: 'question',
+              callID: 'call-question-1',
+              state: {
+                status: 'completed',
+                input: { questions: [{ question: 'Continue?' }] },
+              },
+            } as Message['parts'][number]],
+          ),
+        ],
+        refetch,
+      });
+      await startFallbackPolling(onStreamingDone);
+
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(5_000);
+      });
+
+      expect(refetch).toHaveBeenCalledTimes(1);
+      expect(onStreamingDone).not.toHaveBeenCalled();
+      expect(clientGetMock).not.toHaveBeenCalledWith('/api/session/status');
+
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(5_000);
+      });
+      expect(refetch).toHaveBeenCalledTimes(1);
+      expect(clientGetMock.mock.calls.filter(
+        ([url]) => url === '/api/session/sess-1/message',
+      )).toHaveLength(2);
+
+      await act(async () => {
+        resolveRefetch?.();
+        await Promise.resolve();
+      });
+    } finally {
+      vi.useRealTimers();
+    }
+  });
+
+  it('ignores a missing running question from a historical turn while the session is busy', async () => {
     vi.useFakeTimers();
     const refetch = vi.fn();
     const onStreamingDone = vi.fn();
     try {
-      useSessionMessagesMock.mockReturnValue({
-        messages: [
-          makeMessage({
-            id: 'assistant-1',
-            finish: 'tool-calls',
-            parts: [
-              { id: 'tool-1', type: 'tool', state: { status: 'running' } } as Message['parts'][number],
-            ],
+      const localMessages = [
+        makeMessage({ id: 'old-user', role: 'user' }),
+        makeMessage({ id: 'old-assistant', parentID: 'old-user', finish: 'tool-calls' }),
+        makeMessage({ id: 'current-user', role: 'user' }),
+        makeMessage({ id: 'current-assistant', parentID: 'current-user' }),
+      ];
+      mockFallbackPolling({
+        localMessages,
+        fetchedMessages: [
+          makeFetchedMessage({ id: 'old-user', role: 'user' }),
+          makeFetchedMessage(
+            {
+              id: 'old-assistant',
+              role: 'assistant',
+              parentID: 'old-user',
+              finish: 'tool-calls',
+            },
+            [{
+              id: 'old-tool',
+              type: 'tool',
+              tool: 'question',
+              state: { status: 'running' },
+            } as Message['parts'][number]],
+          ),
+          makeFetchedMessage({ id: 'current-user', role: 'user' }),
+          makeFetchedMessage({
+            id: 'current-assistant',
+            role: 'assistant',
+            parentID: 'current-user',
+            finish: null,
           }),
         ],
-        loading: false,
         refetch,
-        addMessage: vi.fn(),
-        updateMessage: vi.fn(),
-        updateMessagePart: vi.fn(),
-        replaceMessageText: vi.fn(),
-        truncateAfterMessage: vi.fn(),
       });
-      clientGetMock.mockResolvedValueOnce({
-        data: {
-          items: [
+      await startFallbackPolling(onStreamingDone);
+
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(5_000);
+      });
+
+      expect(clientGetMock).toHaveBeenCalledWith('/api/session/sess-1/message', {
+        params: { page: true, limit: 50, include_archived: true },
+      });
+      expect(clientGetMock).not.toHaveBeenCalledWith('/api/session/status');
+      expect(refetch).not.toHaveBeenCalled();
+      expect(onStreamingDone).not.toHaveBeenCalled();
+    } finally {
+      vi.useRealTimers();
+    }
+  });
+
+  it('finishes streaming when only a historical turn has a running question', async () => {
+    vi.useFakeTimers();
+    const refetch = vi.fn();
+    const onStreamingDone = vi.fn();
+    try {
+      const localMessages = [
+        makeMessage({ id: 'old-user', role: 'user' }),
+        makeMessage({ id: 'old-assistant', parentID: 'old-user', finish: 'tool-calls' }),
+        makeMessage({ id: 'current-user', role: 'user' }),
+        makeMessage({ id: 'current-assistant', parentID: 'current-user', finish: 'stop' }),
+      ];
+      mockFallbackPolling({
+        localMessages,
+        fetchedMessages: [
+          makeFetchedMessage({ id: 'old-user', role: 'user' }),
+          makeFetchedMessage(
             {
-              info: {
-                id: 'assistant-1',
-                sessionID: 'sess-1',
-                role: 'assistant',
-                finish: 'tool-calls',
-              },
-              parts: [
-                { id: 'tool-1', type: 'tool', state: { status: 'running' } },
-              ],
+              id: 'old-assistant',
+              role: 'assistant',
+              parentID: 'old-user',
+              finish: 'tool-calls',
             },
-          ],
-          hasMore: false,
-          nextBefore: null,
-        },
+            [{
+              id: 'old-tool',
+              type: 'tool',
+              tool: 'question',
+              state: { status: 'running' },
+            } as Message['parts'][number]],
+          ),
+          makeFetchedMessage({ id: 'current-user', role: 'user' }),
+          makeFetchedMessage(
+            {
+              id: 'current-assistant',
+              role: 'assistant',
+              parentID: 'current-user',
+              finish: 'stop',
+            },
+            [{ id: 'current-text', type: 'text', text: 'done' } as Message['parts'][number]],
+          ),
+        ],
+        refetch,
+        status: {},
       });
+      await startFallbackPolling(onStreamingDone);
 
-      render(React.createElement(SessionChat, {
-        sessionId: 'sess-1',
-        live: true,
-        onStreamingDone,
-      }));
-      act(() => {
-        useSSEOptionsRef.current.onEvent({
-          type: 'session.status',
-          properties: { sessionID: 'sess-1', status: { type: 'busy' } },
-        });
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(5_000);
+      });
+
+      expect(clientGetMock).toHaveBeenCalledWith('/api/session/status');
+      expect(refetch).toHaveBeenCalledTimes(1);
+      expect(onStreamingDone).toHaveBeenCalledTimes(1);
+
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(5_000);
+      });
+      expect(refetch).toHaveBeenCalledTimes(1);
+    } finally {
+      vi.useRealTimers();
+    }
+  });
+
+  it('does not refetch a missing non-question tool while it is still running', async () => {
+    vi.useFakeTimers();
+    const refetch = vi.fn();
+    const onStreamingDone = vi.fn();
+    try {
+      mockFallbackPolling({
+        localMessages: [
+          makeMessage({ id: 'user-1', role: 'user' }),
+          makeMessage({ id: 'assistant-1', parentID: 'user-1', finish: 'tool-calls' }),
+        ],
+        fetchedMessages: [
+          makeFetchedMessage({ id: 'user-1', role: 'user' }),
+          makeFetchedMessage(
+            {
+              id: 'assistant-1',
+              role: 'assistant',
+              parentID: 'user-1',
+              finish: 'tool-calls',
+            },
+            [{
+              id: 'tool-1',
+              type: 'tool',
+              tool: 'bash',
+              state: { status: 'running' },
+            } as Message['parts'][number]],
+          ),
+        ],
+        refetch,
       });
+      await startFallbackPolling(onStreamingDone);
 
-      await vi.advanceTimersByTimeAsync(5_000);
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(5_000);
+      });
 
       expect(refetch).not.toHaveBeenCalled();
       expect(onStreamingDone).not.toHaveBeenCalled();
+      expect(clientGetMock).not.toHaveBeenCalledWith('/api/session/status');
       expect(clientGetMock).toHaveBeenCalledWith('/api/session/sess-1/message', {
         params: { page: true, limit: 50, include_archived: true },
       });
diff --git a/webui/src/components/common/SessionChat.tsx b/webui/src/components/common/SessionChat.tsx
index 90a94f391..693b04cca 100644
--- a/webui/src/components/common/SessionChat.tsx
+++ b/webui/src/components/common/SessionChat.tsx
@@ -17,7 +17,7 @@
  */
 
 import { useState, useCallback, useRef, useEffect, useMemo, memo } from 'react';
-import { Send, Loader2, ChevronDown, Square, Copy, User, FileText, AlertCircle, X, RefreshCw, Pencil, Save, ImageIcon, Paperclip, ArrowUp, Clock, CheckCircle2, XCircle, Brain, Trash2, Bot, Check, ListTree } from 'lucide-react';
+import { Send, Loader2, ChevronDown, Square, Copy, User, FileText, AlertCircle, X, RefreshCw, Pencil, Save, ImageIcon, Paperclip, Plus, ArrowUp, Clock, CheckCircle2, XCircle, Brain, Trash2, Bot, Check, Eye, ListTree, BookOpen, Workflow as WorkflowIcon } from 'lucide-react';
 import { StreamingMarkdown, useStreamingContent } from './StreamingMarkdown';
 import { useTranslation } from 'react-i18next';
 import LoadingSpinner from './LoadingSpinner';
@@ -34,11 +34,13 @@ import type { Command } from '@/api/skill';
 import type { Agent } from '@/api/agent';
 import { useToast } from './Toast';
 import { buildRunWorkflowHeaderSummary } from './toolStageSummary';
+import { getFileOperationDisplayName, redactToolInput, resolveToolPresentation } from './toolPresentation';
 import { areChatMessagePartsRenderEqual } from './sessionChatRenderEquality';
 import { workspaceAPI } from '@/api/workspace';
 import { formatSmartTime } from '@/utils/time';
 import { getAgentDisplayDescription } from '@/utils/agentDisplay';
 import { copyText } from '@/utils/clipboard';
+import { createMessageId } from '@/utils/messageId';
 import {
   FILE_INPUT_ACCEPT_IMAGES,
   batchCompressOptions,
@@ -50,6 +52,7 @@ import {
   type ImagePartData,
 } from '@/utils/imageUpload';
 import type { Message, MessagePart, SessionGoalState, ToolState } from '@/types';
+import type { SessionExecutionMode } from '@/utils/sessionExecutionMode';
 import {
   buildInstructionDisplayText,
   fetchSessionChatCommands,
@@ -57,6 +60,7 @@ import {
   parseInstructionDisplayText,
   resolveSessionChatSSEAction,
   shouldForwardSSEEventToParent,
+  stripTaskMetadata,
   type CompactionStage,
   usePendingQuestions,
   useSessionContextUsage,
@@ -109,11 +113,25 @@ export interface ConversationBottomSlotActions {
   hasMessages: boolean;
 }
 
-function getMessagePartDisplayText(part: MessagePart): string {
+export interface ComposerAddMenuActions {
+  closeMenu: () => void;
+  insertMention: (agentName: string) => void;
+  insertReference: (value: string, kind: 'workflow' | 'skill') => void;
+}
+
+type ComposerReferenceKind = 'subagent' | 'skill' | 'workflow';
+
+interface ComposerReference {
+  kind: ComposerReferenceKind;
+  value: string;
+}
+
+function getMessagePartDisplayText(part: MessagePart, hideTaskMetadata = false): string {
   const metadataDisplayText = part.metadata?.displayText ?? part.metadata?.display_text;
-  return typeof metadataDisplayText === 'string' && metadataDisplayText
+  const displayText = typeof metadataDisplayText === 'string' && metadataDisplayText
     ? metadataDisplayText
     : part.text || '';
+  return hideTaskMetadata ? stripTaskMetadata(displayText) : displayText;
 }
 
 export interface SessionChatProps {
@@ -147,6 +165,12 @@ export interface SessionChatProps {
   agentName?: string;
   /** Model override to include in prompt_async requests */
   model?: { providerID: string; modelID: string } | null;
+  /** Execution mode to include in prompt and queue requests. */
+  executionMode?: SessionExecutionMode;
+  /** Called after a prompt using the current execution mode is accepted. */
+  onExecutionModeAccepted?: (mode: SessionExecutionMode) => void;
+  /** Persist Auto failover before sending through an existing session. */
+  modelAuto?: boolean;
   /** Agents available for one-turn @mention routing. */
   mentionAgents?: Agent[];
   /** Display configuration (compact, showActions, showTimestamp) */
@@ -163,6 +187,10 @@ export interface SessionChatProps {
   onError?: (message: string) => void;
   /** Extra content injected into the left side of the composer toolbar */
   toolbarSlot?: React.ReactNode;
+  /** Extra actions shown below the file picker in the composer's Add menu. */
+  composerAddMenuSlot?: React.ReactNode | ((actions: ComposerAddMenuActions) => React.ReactNode);
+  /** Called when the composer's Add menu opens or closes. */
+  onComposerAddMenuOpenChange?: (open: boolean) => void;
   /** Minimum textarea height in px. Defaults to the compact single-line composer height. */
   composerTextareaMinHeight?: number;
   /** Maximum textarea height in px. Defaults to the existing compact/full-page values. */
@@ -190,6 +218,7 @@ export interface SessionChatProps {
     agentOverride?: string,
     modelOverride?: { providerID: string; modelID: string } | null,
     options?: PromptDisplayOptions,
+    executionModeOverride?: SessionExecutionMode,
   ) => Promise<unknown> | unknown;
   /** Called when the user sends "/new" to create a new session */
   onCreateNewSession?: () => Promise<void> | void;
@@ -558,18 +587,18 @@ function ContextUsageRing({
   return (
     <div
       ref={wrapperRef}
-      className="relative inline-flex h-6 w-6 shrink-0 items-center justify-center"
+      className="relative inline-flex h-6 shrink-0 items-center justify-center"
     >
       <button
         type="button"
-        className="relative inline-flex h-6 w-6 items-center justify-center rounded-full transition-colors hover:bg-zinc-200/60 focus-visible:outline focus-visible:outline-2 focus-visible:outline-offset-2 focus-visible:outline-sky-500"
+        className="relative inline-flex h-6 items-center gap-1 rounded-md px-1 text-[10px] font-medium tabular-nums text-zinc-500 transition-colors hover:bg-zinc-200/60 hover:text-zinc-700 focus-visible:outline focus-visible:outline-2 focus-visible:outline-offset-2 focus-visible:outline-sky-500 dark:text-zinc-400 dark:hover:bg-zinc-800 dark:hover:text-zinc-200"
         title={title}
         aria-label={title}
         aria-haspopup="menu"
         aria-expanded={open}
         onClick={() => setOpen((value) => !value)}
       >
-        <svg className="absolute inset-0 h-6 w-6 -rotate-90" viewBox="0 0 24 24" aria-hidden="true">
+        <svg className="h-5 w-5 shrink-0 -rotate-90" viewBox="0 0 24 24" aria-hidden="true">
           <circle cx="12" cy="12" r={radius} fill="none" strokeWidth="2" className="stroke-zinc-200 dark:stroke-zinc-800" />
           <circle
             cx="12"
@@ -583,6 +612,7 @@ function ContextUsageRing({
             strokeDashoffset={strokeDashoffset}
           />
         </svg>
+        <span aria-hidden="true">{clamped}%</span>
       </button>
 
       {open && (
@@ -817,27 +847,21 @@ export function getMessageBubbleClassName({
   isUser: boolean;
   isEditing: boolean;
 }): string {
+  if (!isUser) {
+    const typographyClass = compact ? 'text-sm' : 'text-[15px]';
+
+    return `w-full max-w-full min-w-0 bg-transparent py-1 ${typographyClass} text-[#34393e] break-words dark:text-zinc-100`;
+  }
+
   if (compact) {
-    const widthClass = isUser
-      ? (isEditing ? 'w-full max-w-full' : 'max-w-full')
-      : 'w-full max-w-full';
-
-    return `${widthClass} min-w-0 px-4 py-3 rounded-[20px] text-sm break-words shadow-sm ${
-      isUser
-        ? 'bg-sky-50 border border-sky-100 text-zinc-900 dark:border-zinc-700 dark:bg-zinc-800 dark:text-zinc-50 dark:shadow-none'
-        : 'bg-white border border-zinc-200/90 dark:border-zinc-800 dark:bg-zinc-900 dark:text-zinc-100 dark:shadow-none'
-    }`;
+    const widthClass = isEditing ? 'w-full max-w-full' : 'max-w-full';
+
+    return `${widthClass} min-w-0 px-4 py-3 rounded-[20px] text-sm break-words shadow-sm border border-black/[0.07] bg-zinc-50 text-[#30343a] dark:border-white/[0.08] dark:bg-[#303842] dark:text-zinc-50 dark:shadow-none`;
   }
 
-  const widthClass = isUser
-    ? (isEditing ? 'w-full' : 'w-auto')
-    : 'w-full';
+  const widthClass = isEditing ? 'w-full' : 'w-auto';
 
-  return `${widthClass} min-w-0 max-w-full px-5 py-4 rounded-[24px] text-sm break-words shadow-sm ${
-    isUser
-      ? 'bg-sky-50 border border-sky-100 text-zinc-900 dark:border-zinc-700 dark:bg-zinc-800 dark:text-zinc-50 dark:shadow-none'
-      : 'bg-white border border-zinc-200/90 dark:border-zinc-800 dark:bg-zinc-900 dark:text-zinc-100 dark:shadow-none'
-  }`;
+  return `${widthClass} min-w-0 max-w-full px-5 py-3 rounded-[18px] text-sm break-words shadow-sm border border-black/[0.09] bg-zinc-50 text-[#30343a] dark:border-white/[0.10] dark:bg-[#303842] dark:text-zinc-50 dark:shadow-none`;
 }
 
 export function getInstructionDisplayBubbleClassName(compact: boolean): string {
@@ -901,6 +925,54 @@ export function hasActiveToolPart(parts?: Array<Pick<MessagePart, 'type' | 'stat
   return parts?.some(isActiveToolPart) ?? false;
 }
 
+type FetchedMessageWithParts = {
+  info?: {
+    id?: string;
+    role?: string;
+    parentID?: string | null;
+    finish?: string | null;
+    time?: { completed?: number | null };
+  };
+  parts?: MessagePart[];
+};
+
+function getCurrentTurnAssistantMessages(
+  messages: FetchedMessageWithParts[],
+): FetchedMessageWithParts[] | null {
+  let latestUserIndex = -1;
+  for (let index = messages.length - 1; index >= 0; index -= 1) {
+    if (messages[index]?.info?.role === 'user') {
+      latestUserIndex = index;
+      break;
+    }
+  }
+
+  let turnParentID = latestUserIndex >= 0
+    ? messages[latestUserIndex]?.info?.id
+    : undefined;
+  let turnStartIndex = latestUserIndex + 1;
+
+  // A single tool-heavy turn can exceed the latest-message page. If its user
+  // message is outside the page, recover the turn from the newest assistant's
+  // parent instead of falling back to every historical tool in the page.
+  if (!turnParentID) {
+    for (let index = messages.length - 1; index >= 0; index -= 1) {
+      const info = messages[index]?.info;
+      if (info?.role === 'assistant' && info.parentID) {
+        turnParentID = info.parentID;
+        turnStartIndex = 0;
+        break;
+      }
+    }
+  }
+
+  if (!turnParentID) return null;
+  return messages.slice(turnStartIndex).filter((message) => (
+    message.info?.role === 'assistant'
+    && message.info.parentID === turnParentID
+  ));
+}
+
 export function isActiveSessionStatus(status?: { type?: string } | null): boolean {
   return status?.type === 'busy' || status?.type === 'compacting' || status?.type === 'retry';
 }
@@ -1417,15 +1489,19 @@ function findMentionTrigger(text: string, cursor: number): { start: number; end:
   };
 }
 
-function resolveMentionAgentName(text: string, agents: Agent[]): string | null {
+function resolveReferencedAgentName(text: string, agents: Agent[]): string | null {
   const sorted = [...agents].sort((a, b) => b.name.length - a.name.length);
   for (const agent of sorted) {
-    const pattern = new RegExp(`(^|\\s)@${escapeRegExp(agent.name)}(?=$|\\s|[,.!?;:，。！？；：])`, 'i');
+    const pattern = new RegExp(`(^|\\s)subagent:${escapeRegExp(agent.name)}(?=$|\\s|[,.!?;，。！？；])`, 'i');
     if (pattern.test(text)) return agent.name;
   }
   return null;
 }
 
+function formatComposerReference(reference: ComposerReference): string {
+  return `${reference.kind}:${reference.value}`;
+}
+
 export default function SessionChat({
   sessionId,
   live = false,
@@ -1441,6 +1517,9 @@ export default function SessionChat({
   initialDisplayText,
   agentName,
   model,
+  executionMode = 'build',
+  onExecutionModeAccepted,
+  modelAuto = false,
   display,
   welcomeContent,
   conversationBottomSlot,
@@ -1452,6 +1531,8 @@ export default function SessionChat({
   onInitialMessageConsumed,
   supportsVision,
   toolbarSlot,
+  composerAddMenuSlot,
+  onComposerAddMenuOpenChange,
   composerTextareaMinHeight,
   composerTextareaMaxHeight,
   centerToolbarSlot,
@@ -1462,6 +1543,7 @@ export default function SessionChat({
   const toast = useToast();
   const compact = display?.compact ?? true;
   const fullWidth = display?.fullWidth ?? false;
+  const pageCanvas = display?.pageCanvas ?? false;
   const showActions = display?.showActions ?? false;
   const showTimestamp = display?.showTimestamp ?? false;
   const collapseIntermediateSteps = display?.collapseIntermediateSteps ?? false;
@@ -1471,15 +1553,29 @@ export default function SessionChat({
   const effectiveComposerTextareaMaxHeight = composerTextareaMaxHeight ?? (compact ? 96 : 200);
   const effectivePlaceholder = placeholder ?? t('chat.placeholder');
   const effectiveEmptyText = emptyText ?? t('chat.emptyText');
+  const autoModelSessionRef = useRef<string | null>(null);
+  useEffect(() => {
+    if (!modelAuto) autoModelSessionRef.current = null;
+  }, [modelAuto]);
+  const ensureAutoModelSession = useCallback(async () => {
+    if (!sessionId || !modelAuto || autoModelSessionRef.current === sessionId) return;
+    await sessionApi.update(sessionId, {
+      model_auto: true,
+      model_pinned: false,
+    });
+    autoModelSessionRef.current = sessionId;
+  }, [modelAuto, sessionId]);
   // Restore any persisted draft on first mount so navigating away (e.g.
   // sidebar → Agents → back to Sessions) doesn't wipe the user's half-typed
   // message. Subsequent session changes are re-hydrated by the effect below.
   const [input, setInput] = useState<string>(() => readChatDraft(sessionId));
+  const [composerReferences, setComposerReferences] = useState<ComposerReference[]>([]);
   const [sending, setSending] = useState(false);
   const [isStreaming, setIsStreaming] = useState(false);
   const activeToolPartIdsRef = useRef<Set<string>>(new Set());
   const [attachments, setAttachments] = useState<ComposerAttachment[]>([]);
   const [isDragOver, setIsDragOver] = useState(false);
+  const [showComposerAddMenu, setShowComposerAddMenu] = useState(false);
   // Lightbox preview for composer thumbnails. Shares the same overlay
   // component used by message bubbles so the click-to-enlarge gesture is
   // consistent across the upload tray and the rendered chat history.
@@ -1581,6 +1677,7 @@ export default function SessionChat({
   const initialMessageSentRef = useRef('');
   const abortingRef = useRef(false);
   const sessionBusyRef = useRef(false);
+  const sessionStatusRevisionRef = useRef(0);
   const goalHydrationVersionRef = useRef(0);
   // ID of the assistant message that was aborted; used to ignore its finish event
   const abortedMessageIdRef = useRef<string | null>(null);
@@ -1604,6 +1701,45 @@ export default function SessionChat({
   const fileInputRef = useRef<HTMLInputElement>(null);
   const isComposingRef = useRef(false);
 
+  const closeComposerAddMenu = useCallback(() => {
+    setShowComposerAddMenu(false);
+    onComposerAddMenuOpenChange?.(false);
+  }, [onComposerAddMenuOpenChange]);
+
+  const openComposerAddMenu = useCallback(() => {
+    setShowComposerAddMenu(true);
+    onComposerAddMenuOpenChange?.(true);
+  }, [onComposerAddMenuOpenChange]);
+
+  const toggleComposerAddMenu = useCallback(() => {
+    setShowComposerAddMenu((open) => {
+      const nextOpen = !open;
+      onComposerAddMenuOpenChange?.(nextOpen);
+      return nextOpen;
+    });
+  }, [onComposerAddMenuOpenChange]);
+
+  useEffect(() => {
+    if (!showComposerAddMenu) return;
+    const handlePointerDown = (event: MouseEvent) => {
+      const target = event.target as HTMLElement;
+      if (!target.closest('[data-composer-add-menu]')) closeComposerAddMenu();
+    };
+    const handleKeyDown = (event: KeyboardEvent) => {
+      if (event.key === 'Escape') closeComposerAddMenu();
+    };
+    document.addEventListener('mousedown', handlePointerDown);
+    document.addEventListener('keydown', handleKeyDown);
+    return () => {
+      document.removeEventListener('mousedown', handlePointerDown);
+      document.removeEventListener('keydown', handleKeyDown);
+    };
+  }, [closeComposerAddMenu, showComposerAddMenu]);
+
+  useEffect(() => {
+    if (sending && showComposerAddMenu) closeComposerAddMenu();
+  }, [closeComposerAddMenu, sending, showComposerAddMenu]);
+
   // Slash command autocomplete state
   const [commands, setCommands] = useState<Command[]>([]);
   const [showCommandDropdown, setShowCommandDropdown] = useState(false);
@@ -1630,7 +1766,12 @@ export default function SessionChat({
   );
   const hasUploadingFiles = attachments.some((attachment) => attachment.status === 'uploading');
   const canSend = !sending && !hasUploadingFiles &&
-    (!!input.trim() || successfulDocAttachments.length > 0 || successfulImageAttachments.length > 0);
+    (
+      !!input.trim()
+      || composerReferences.length > 0
+      || successfulDocAttachments.length > 0
+      || successfulImageAttachments.length > 0
+    );
   const filteredMentionAgents = useMemo(() => {
     const q = mentionQuery.trim().toLowerCase();
     return mentionAgents
@@ -1654,9 +1795,6 @@ export default function SessionChat({
     }
   }, []);
 
-  const loadOlderMessagesRef = useRef<(() => Promise<void>) | null>(null);
-  const hasMoreMessagesRef = useRef(false);
-  const loadingOlderMessagesRef = useRef(false);
   const rafScheduledRef = useRef(false);
   const handleScroll = useCallback(() => {
     if (rafScheduledRef.current) return;
@@ -1665,18 +1803,6 @@ export default function SessionChat({
       const el = scrollContainerRef.current;
       if (el) {
         isAtBottomRef.current = el.scrollHeight - el.scrollTop - el.clientHeight <= SCROLL_BOTTOM_THRESHOLD_PX;
-        if (el.scrollTop <= 80 && hasMoreMessagesRef.current && !loadingOlderMessagesRef.current) {
-          const previousHeight = el.scrollHeight;
-          const previousTop = el.scrollTop;
-          const loadPromise = loadOlderMessagesRef.current?.();
-          if (loadPromise) void loadPromise.finally(() => {
-            requestAnimationFrame(() => {
-              const current = scrollContainerRef.current;
-              if (!current) return;
-              current.scrollTop = current.scrollHeight - previousHeight + previousTop;
-            });
-          });
-        }
       }
       rafScheduledRef.current = false;
     });
@@ -1685,21 +1811,18 @@ export default function SessionChat({
   const {
     messages,
     loading,
-    loadingOlder,
-    hasMore: hasMoreMessages,
+    error: messagesError,
     refetch,
-    loadOlder,
     addMessage,
     updateMessage,
     updateMessagePart,
+    removeMessage,
+    clearMessages,
     replaceMessageText,
     markMessageStopped,
     truncateAfterMessage,
   } =
     useSessionMessages(sessionId || undefined);
-  useEffect(() => { loadOlderMessagesRef.current = loadOlder; }, [loadOlder]);
-  useEffect(() => { hasMoreMessagesRef.current = hasMoreMessages; }, [hasMoreMessages]);
-  useEffect(() => { loadingOlderMessagesRef.current = loadingOlder; }, [loadingOlder]);
   const contextUsageMessages = contextUsageRefreshing && !contextUsageSnapshot ? [] : messages;
   const contextUsageBreakdown = useMemo(
     () => buildContextUsageBreakdown(contextUsageMessages, input, contextUsageSnapshot),
@@ -1725,6 +1848,8 @@ export default function SessionChat({
   // Keep a ref to latest messages so handleAbort can read it without stale closure
   const messagesRef = useRef(messages);
   useEffect(() => { messagesRef.current = messages; }, [messages]);
+  const pendingQuestionsRef = useRef(pendingQuestions);
+  useEffect(() => { pendingQuestionsRef.current = pendingQuestions; }, [pendingQuestions]);
 
   const hasUserMessage = useMemo(() => messages.some((m) => m.role === 'user'), [messages]);
 
@@ -1767,6 +1892,7 @@ export default function SessionChat({
         case 'ignore':
           return;
         case 'session-cleared':
+          sessionStatusRevisionRef.current += 1;
           abortingRef.current = false;
           sessionBusyRef.current = false;
           activeToolPartIdsRef.current.clear();
@@ -1775,10 +1901,11 @@ export default function SessionChat({
           setIsStreaming(false);
           setGoalBanner(null);
           setDismissedGoalKey('');
-          refetch();
+          clearMessages();
           void refreshContextUsage({ clear: true });
           return;
         case 'session-status':
+          sessionStatusRevisionRef.current += 1;
           if (action.statusType === 'busy') {
             sessionBusyRef.current = true;
             if (
@@ -1854,6 +1981,17 @@ export default function SessionChat({
           }
           return;
         }
+        case 'message-removed': {
+          const removedMessage = messagesRef.current.find((message) => message.id === action.messageID);
+          removedMessage?.parts.forEach((part) => {
+            if (part.id) activeToolPartIdsRef.current.delete(part.id);
+          });
+          if (abortedMessageIdRef.current === action.messageID) {
+            abortedMessageIdRef.current = null;
+          }
+          removeMessage(action.messageID);
+          return;
+        }
         case 'message-part-updated': {
           const part = action.part as Pick<MessagePart, 'id' | 'type' | 'state'>;
           if (part.id) {
@@ -1931,6 +2069,8 @@ export default function SessionChat({
       sessionId,
       updateMessage,
       updateMessagePart,
+      removeMessage,
+      clearMessages,
       refetch,
       refreshContextUsage,
       applyContextUsagePushSnapshot,
@@ -1977,11 +2117,52 @@ export default function SessionChat({
     [onError, submitReject, t, toast],
   );
 
+  const reconcileSessionStatusAfterReconnect = useCallback(async () => {
+    if (!sessionId) return;
+    const statusRevision = sessionStatusRevisionRef.current;
+
+    try {
+      const response = await client.get('/api/session/status');
+      if (statusRevision !== sessionStatusRevisionRef.current) return;
+
+      const status = response.data?.[sessionId];
+      if (isActiveSessionStatus(status)) {
+        sessionBusyRef.current = true;
+        if (!abortingRef.current && !suppressStreamingUntilIdleRef.current) {
+          setIsStreaming(true);
+        }
+        if (status?.type === 'compacting') {
+          setIsCompacting(true);
+          isCompactingRef.current = true;
+          setCompactingMessage(status.message || t('chat.compacting'));
+        } else {
+          setIsCompacting(false);
+          isCompactingRef.current = false;
+          setCompactingMessage('');
+          setCompactionStages([]);
+        }
+        return;
+      }
+
+      sessionBusyRef.current = false;
+      activeToolPartIdsRef.current.clear();
+      setSending(false);
+      setIsStreaming(false);
+      setIsCompacting(false);
+      isCompactingRef.current = false;
+      setCompactingMessage('');
+      setCompactionStages([]);
+    } catch {
+      // Keep the current activity state when reconnect status cannot be verified.
+    }
+  }, [sessionId, t]);
+
   const { status: sseStatus } = useSSE({
     url: `${getApiBase()}/api/event`,
     onEvent: handleSSEEvent,
     onReconnect: () => {
       if (!sessionId) return;
+      void reconcileSessionStatusAfterReconnect();
       refetch();
       refreshContextUsage();
       fetchPromptQueue();
@@ -2047,11 +2228,11 @@ export default function SessionChat({
     setMentionRange(null);
     setMentionQuery('');
     setSelectedMentionIndex(0);
-    setPendingAgentName(agentName || 'rex');
     abortingRef.current = false;
     abortedMessageIdRef.current = null;
     suppressStreamingUntilIdleRef.current = false;
     sessionBusyRef.current = false;
+    sessionStatusRevisionRef.current += 1;
     statusCheckedRef.current = null;
     isAtBottomRef.current = true;
     clearPendingQuestions();
@@ -2059,8 +2240,9 @@ export default function SessionChat({
     // don't force a remount (Session/index.tsx does, but other consumers
     // such as WorkflowDetail/ChatTab may swap sessionId without a remount).
     setInput(readChatDraft(sessionId));
+    setComposerReferences([]);
     setProcessGroupOpenState(readProcessGroupOpenState(sessionId));
-  }, [sessionId, agentName, clearPendingQuestions]);
+  }, [sessionId, clearPendingQuestions]);
 
   const handleProcessGroupOpenChange = useCallback((key: string, open: boolean) => {
     setProcessGroupOpenState(prev => {
@@ -2422,20 +2604,22 @@ export default function SessionChat({
     setIsStreaming(true);
 
     const displayText = args ? `/${command} ${args}` : `/${command}`;
-    const tempId = `temp-${Date.now()}`;
+    const messageId = createMessageId();
     addMessage({
-      id: tempId,
+      id: messageId,
       sessionID: sessionId,
       role: 'user',
-      parts: [{ id: `${tempId}-part`, type: 'text', text: displayText }],
+      parts: [{ id: `temp-${messageId}-part`, type: 'text', text: displayText }],
       timestamp: Date.now(),
     } as Message);
 
     try {
+      await ensureAutoModelSession();
       await client.post(`/api/session/${sessionId}/command`, {
         command,
         arguments: args,
         agent: agentName,
+        messageID: messageId,
       });
       if (command === 'goal' && args.trim()) {
         goalHydrationVersionRef.current += 1;
@@ -2445,6 +2629,7 @@ export default function SessionChat({
       }
     } catch (err: unknown) {
       setIsStreaming(false);
+      removeMessage(messageId);
       const axiosErr = err as any;
       if (axiosErr?.response?.status === 404) {
         onError?.('Session not found. Please start a new session.');
@@ -2465,6 +2650,7 @@ export default function SessionChat({
     options?: PromptDisplayOptions,
   ) => {
     if (!sessionId) return;
+    await ensureAutoModelSession();
     const effectiveAgent = agentOverride || agentName;
     const visibleText = options?.displayText || text;
     // Clear abort state immediately so SSE events for the new stream are not suppressed
@@ -2477,18 +2663,18 @@ export default function SessionChat({
     setIsStreaming(true);
     setPendingAgentName(effectiveAgent || 'rex');
 
-    const tempId = `temp-${Date.now()}`;
+    const messageId = createMessageId();
     const tempParts: MessagePart[] = [];
-    if (visibleText) tempParts.push({ id: `${tempId}-text`, type: 'text', text: visibleText });
+    if (visibleText) tempParts.push({ id: `temp-${messageId}-text`, type: 'text', text: visibleText });
     imageParts.forEach((img, i) => {
-      tempParts.push({ id: `${tempId}-img-${i}`, type: 'file', url: img.url, mime: img.mime, filename: img.filename });
+      tempParts.push({ id: `temp-${messageId}-img-${i}`, type: 'file', url: img.url, mime: img.mime, filename: img.filename });
     });
 
     addMessage({
-      id: tempId,
+      id: messageId,
       sessionID: sessionId,
       role: 'user',
-      parts: tempParts.length > 0 ? tempParts : [{ id: `${tempId}-part`, type: 'text', text: visibleText }],
+      parts: tempParts.length > 0 ? tempParts : [{ id: `temp-${messageId}-part`, type: 'text', text: visibleText }],
       timestamp: Date.now(),
       agent: effectiveAgent,
     } as Message);
@@ -2496,14 +2682,24 @@ export default function SessionChat({
     try {
       const payload: Record<string, unknown> = {
         parts: buildPromptParts(text, imageParts),
+        messageID: messageId,
       };
       if (effectiveAgent) payload.agent = effectiveAgent;
       if (model) payload.model = model;
       if (options?.displayText) payload.displayText = options.displayText;
+      payload.executionMode = executionMode;
 
       await client.post(`/api/session/${sessionId}/prompt_async`, payload);
+      if (executionMode === 'goal' && text.trim()) {
+        goalHydrationVersionRef.current += 1;
+        writeDismissedGoalKey(sessionId, '');
+        setGoalBanner({ objective: text.trim(), status: 'active' });
+        setDismissedGoalKey('');
+      }
+      onExecutionModeAccepted?.(executionMode);
     } catch (err: unknown) {
       setIsStreaming(false);
+      removeMessage(messageId);
       const axiosErr = err as any;
       if (axiosErr?.response?.status === 404) {
         onError?.(`Session not found. Please start a new session.`);
@@ -2525,12 +2721,15 @@ export default function SessionChat({
     if (!sessionId) return;
     const effectiveAgent = agentOverride || agentName;
     try {
+      await ensureAutoModelSession();
       await enqueuePrompt({
         parts: buildPromptParts(text, imageParts),
         ...(effectiveAgent ? { agent: effectiveAgent } : {}),
         ...(model ? { model } : {}),
         ...(options?.displayText ? { displayText: options.displayText } : {}),
+        executionMode,
       });
+      onExecutionModeAccepted?.(executionMode);
     } catch (err: any) {
       const statusCode = err?.response?.status;
       const detail = err?.response?.data?.detail;
@@ -2569,7 +2768,15 @@ export default function SessionChat({
       setSending(true);
       try {
         setPendingAgentName(agentName || 'rex');
-        await onCreateAndSend(trimmed, [], agentName, model, options);
+        await onCreateAndSend(
+          trimmed,
+          [],
+          agentName,
+          model,
+          options,
+          executionMode,
+        );
+        onExecutionModeAccepted?.(executionMode);
       } catch {
         setInput(trimmed);
       } finally {
@@ -2587,16 +2794,24 @@ export default function SessionChat({
 
   const handleSend = async () => {
     if (!canSend) return;
-    const rawText = input.trim();
+    const draftText = input.trim();
+    const referencesToSend = [...composerReferences];
+    const referenceText = referencesToSend.map(formatComposerReference).join(' ');
+    const rawText = [referenceText, draftText].filter(Boolean).join(' ');
     const docAttachmentsToSend = [...successfulDocAttachments];
     const imageAttachmentsToSend = [...successfulImageAttachments];
     const text = buildMessageText(rawText, docAttachmentsToSend);
-    const mentionedAgent = resolveMentionAgentName(rawText, mentionAgents);
+    const mentionedAgent = resolveReferencedAgentName(rawText, mentionAgents);
+    const restoreDraft = () => {
+      setInput(draftText);
+      setComposerReferences(referencesToSend);
+    };
 
     // Need either text content or image attachments
     if (!text && imageAttachmentsToSend.length === 0) return;
 
     setInput('');
+    setComposerReferences([]);
     setShowCommandDropdown(false);
     setMentionRange(null);
 
@@ -2621,7 +2836,7 @@ export default function SessionChat({
         await enqueueText(text, imageParts, mentionedAgent || undefined);
         setAttachments([]);
       } catch {
-        setInput(rawText);
+        restoreDraft();
         setAttachments([...docAttachmentsToSend, ...imageAttachmentsToSend]);
       }
       return;
@@ -2631,13 +2846,13 @@ export default function SessionChat({
     if (parsed) {
       if (!sessionId) {
         // Slash commands need an existing session; restore input and do nothing
-        setInput(rawText);
+        restoreDraft();
         return;
       }
       try {
         await sendCommand(parsed.command, parsed.args);
       } catch {
-        setInput(rawText);
+        restoreDraft();
       }
       return;
     }
@@ -2648,13 +2863,21 @@ export default function SessionChat({
         try {
           const effectiveAgent = mentionedAgent || agentName;
           setPendingAgentName(effectiveAgent || 'rex');
-          await onCreateAndSend(text, imageParts, effectiveAgent || undefined, model);
+          await onCreateAndSend(
+            text,
+            imageParts,
+            effectiveAgent || undefined,
+            model,
+            undefined,
+            executionMode,
+          );
+          onExecutionModeAccepted?.(executionMode);
           setAttachments([]);
         } catch {
           // Restore both the text and the attachment list so the user can
           // retry without re-uploading images. Image data URLs are already
           // in memory, so restoring the array is safe and cheap.
-          setInput(rawText);
+          restoreDraft();
           setAttachments(imageAttachmentsToSend);
         } finally {
           setSending(false);
@@ -2667,7 +2890,7 @@ export default function SessionChat({
       await sendText(text, imageParts, mentionedAgent || undefined);
       setAttachments([]);
     } catch {
-      setInput(rawText);
+      restoreDraft();
       setAttachments(imageAttachmentsToSend);
     }
   };
@@ -2706,6 +2929,43 @@ export default function SessionChat({
     });
   }, [input, mentionRange]);
 
+  const insertAgentMention = useCallback((name: string) => {
+    setComposerReferences((current) => [
+      { kind: 'subagent', value: name },
+      ...current.filter((reference) => reference.kind !== 'subagent'),
+    ]);
+    setMentionRange(null);
+    setMentionQuery('');
+    setSelectedMentionIndex(0);
+    requestAnimationFrame(() => {
+      textareaRef.current?.focus();
+    });
+  }, []);
+
+  const insertComposerReference = useCallback((
+    value: string,
+    kind: 'workflow' | 'skill',
+  ) => {
+    setComposerReferences((current) => {
+      if (kind === 'skill') {
+        if (current.some((reference) => reference.kind === 'skill' && reference.value === value)) {
+          return current;
+        }
+        return [{ kind, value }, ...current];
+      }
+      return [
+        { kind, value },
+        ...current.filter((reference) => reference.kind !== kind),
+      ];
+    });
+    setMentionRange(null);
+    setMentionQuery('');
+    setSelectedMentionIndex(0);
+    requestAnimationFrame(() => {
+      textareaRef.current?.focus();
+    });
+  }, []);
+
   const handleKeyDown = (e: React.KeyboardEvent) => {
     const currentValue = e.currentTarget instanceof HTMLTextAreaElement ? e.currentTarget.value : input;
     const activeMention = mentionRange
@@ -2864,24 +3124,81 @@ export default function SessionChat({
   // Fallback polling to detect completion when SSE events are missed
   useEffect(() => {
     if (!isStreaming || !sessionId) return;
+    let questionRecoveryInFlight = false;
     const timer = setInterval(async () => {
       try {
         const res = await client.get(`/api/session/${sessionId}/message`, {
           params: { page: true, limit: 50, include_archived: true },
         });
-        const msgs: any[] = Array.isArray(res.data) ? res.data : (res.data?.items || []);
+        const msgs: FetchedMessageWithParts[] = Array.isArray(res.data)
+          ? res.data
+          : (res.data?.items || []);
+        const currentTurnAssistantMessages = getCurrentTurnAssistantMessages(msgs);
+        const shouldRecoverQuestionPart = (part: MessagePart) => (
+          isQuestionToolName(part.tool || '')
+          && (
+            isActiveToolPart(part)
+            || !!(part.callID && pendingQuestionsRef.current[part.callID])
+          )
+        );
+        const hasFetchedPendingQuestion = currentTurnAssistantMessages?.some((message) => (
+          (message.parts || []).some((part) => (
+            shouldRecoverQuestionPart(part)
+          ))
+        )) ?? false;
+        if (currentTurnAssistantMessages && hasFetchedPendingQuestion) {
+          const localQuestionPartIds = new Set<string>();
+          const localQuestionCallIds = new Set<string>();
+          for (const message of messagesRef.current) {
+            for (const part of message.parts || []) {
+              if (
+                (part.type !== 'tool' && part.type !== 'toolCall')
+                || !isQuestionToolName(part.tool || '')
+              ) continue;
+              if (part.id) localQuestionPartIds.add(part.id);
+              if (part.callID) localQuestionCallIds.add(part.callID);
+            }
+          }
+          const hasMissingFetchedQuestion = currentTurnAssistantMessages.some((msg) => (
+            (msg.parts || []).some((part: MessagePart) => (
+              shouldRecoverQuestionPart(part)
+              && !(
+                (part.id && localQuestionPartIds.has(part.id))
+                || (part.callID && localQuestionCallIds.has(part.callID))
+              )
+            ))
+          ));
+          if (hasMissingFetchedQuestion && !questionRecoveryInFlight) {
+            questionRecoveryInFlight = true;
+            try {
+              await refetch();
+            } finally {
+              questionRecoveryInFlight = false;
+            }
+          }
+        }
+
         const lastMsg = msgs[msgs.length - 1];
         if (lastMsg?.info?.role === 'assistant' && (lastMsg.info.finish || lastMsg.info.time?.completed)) {
-          const hasFetchedActiveTool = msgs.some((msg) => hasActiveToolPart(msg.parts));
-          if (hasFetchedActiveTool) {
-            return;
-          }
+          const currentTurnMessages = currentTurnAssistantMessages
+            ? new Set(currentTurnAssistantMessages)
+            : null;
+          const hasFetchedActiveTool = msgs.some((msg) => (
+            (msg.parts || []).some((part: MessagePart) => {
+              if (!isQuestionToolName(part.tool || '')) return isActiveToolPart(part);
+              const isPendingQuestion = isActiveToolPart(part)
+                || !!(part.callID && pendingQuestionsRef.current[part.callID]);
+              if (!isPendingQuestion) return false;
+              return currentTurnMessages === null || currentTurnMessages.has(msg);
+            })
+          ));
+          if (hasFetchedActiveTool) return;
+
           activeToolPartIdsRef.current.clear();
           const statusRes = await client.get('/api/session/status');
           const status = statusRes.data?.[sessionId];
-          if (isActiveSessionStatus(status)) {
-            return;
-          }
+          if (isActiveSessionStatus(status)) return;
+
           refetch();
           setIsStreaming(false);
         }
@@ -3050,11 +3367,17 @@ export default function SessionChat({
   // ── Styling based on compact mode ──
   const msgAreaClass = compact
     ? 'relative flex flex-col flex-1 min-h-0 overflow-y-auto bg-gray-50 px-4 py-4 dark:bg-zinc-950'
-    : 'relative flex flex-col flex-1 min-h-0 overflow-y-auto bg-gray-50 py-6 dark:bg-zinc-950';
+    : pageCanvas
+      ? 'relative flex flex-col flex-1 min-h-0 overflow-y-auto bg-transparent py-5'
+      : 'relative flex flex-col flex-1 min-h-0 overflow-y-auto bg-gray-50 py-6 dark:bg-zinc-950';
 
   const msgListClass = compact
     ? fullWidth ? 'space-y-3 w-full px-4' : 'space-y-3'
-    : fullWidth ? 'space-y-5 w-full px-5' : 'space-y-5 w-[min(76%,64rem)] mx-auto px-6';
+    : fullWidth
+      ? 'space-y-5 w-full px-12'
+      : pageCanvas
+        ? 'space-y-[18px] w-full max-w-[760px] mx-auto px-7'
+        : 'space-y-5 w-[min(76%,64rem)] mx-auto px-6';
   const visibleGoalBanner = goalBanner && getGoalBannerKey(goalBanner) !== dismissedGoalKey
     ? goalBanner
     : null;
@@ -3075,7 +3398,26 @@ export default function SessionChat({
       >
         {loading && messages.length === 0 ? (
           <div className="flex justify-center py-8">
-            <LoadingSpinner />
+            <LoadingSpinner
+              size="sm"
+              delayMs={180}
+              className="opacity-60 [&_svg]:text-zinc-400 dark:[&_svg]:text-zinc-500"
+            />
+          </div>
+        ) : messagesError && messages.length === 0 ? (
+          <div className="flex min-h-40 items-center justify-center px-6" role="alert">
+            <div className="flex flex-col items-center gap-3 text-center">
+              <span className="text-sm text-zinc-500 dark:text-zinc-400">
+                {t('chat.loadFailed')}
+              </span>
+              <button
+                type="button"
+                onClick={() => void refetch()}
+                className="rounded-lg border border-zinc-200 bg-white px-3 py-1.5 text-xs font-medium text-zinc-700 transition-colors hover:bg-zinc-50 dark:border-zinc-700 dark:bg-zinc-900 dark:text-zinc-200 dark:hover:bg-zinc-800"
+              >
+                {t('chat.retry')}
+              </button>
+            </div>
           </div>
         ) : messages.length === 0 ? (
           welcomeContent ? (
@@ -3093,19 +3435,6 @@ export default function SessionChat({
           )
         ) : (
           <div ref={messagesContentRef} className={msgListClass}>
-            {hasMoreMessages && (
-              <div className="flex justify-center pb-2">
-                <button
-                  type="button"
-                  onClick={() => void loadOlder()}
-                  disabled={loadingOlder}
-                  className="inline-flex items-center gap-1.5 rounded-full border border-zinc-200 bg-white px-3 py-1.5 text-xs font-medium text-zinc-500 transition-colors hover:bg-zinc-50 disabled:cursor-not-allowed disabled:opacity-60"
-                >
-                  {loadingOlder ? <Loader2 className="h-3.5 w-3.5 animate-spin" /> : <ChevronDown className="h-3.5 w-3.5 rotate-180" />}
-                  <span>{loadingOlder ? t('chat.loadingOlder', 'Loading...') : t('chat.loadOlder', 'Load earlier messages')}</span>
-                </button>
-              </div>
-            )}
             <ChatMessageTimeline
               items={historyItems}
               pendingQuestions={pendingQuestions}
@@ -3170,7 +3499,7 @@ export default function SessionChat({
                   </span>
                   <div className="flex flex-col items-start flex-1 min-w-0">
                     <div className={`flex items-center gap-2 ${compact ? 'h-7' : 'h-8'}`}>
-                      <span className="text-xs font-semibold text-zinc-700 dark:text-zinc-300">{formatAgentName(pendingAgentName)}</span>
+                      <span className="text-sm font-semibold text-zinc-700 dark:text-zinc-300">{formatAgentName(pendingAgentName)}</span>
                     </div>
                     <div className="flex flex-col min-w-0 w-full">
                       <div className={`${compact ? 'w-full max-w-full px-4 py-3 rounded-[20px]' : 'w-full px-5 py-4 rounded-[24px]'} text-sm break-words shadow-sm bg-amber-50 border border-amber-200 dark:border-amber-500/35 dark:bg-amber-950/30 dark:shadow-none`}>
@@ -3228,7 +3557,7 @@ export default function SessionChat({
                   </span>
                   <div className="flex flex-col items-start flex-1 min-w-0">
                     <div className={`flex items-center gap-2 ${compact ? 'h-7' : 'h-8'}`}>
-                      <span className="text-xs font-semibold text-zinc-700 dark:text-zinc-300">{formatAgentName(pendingAgentName)}</span>
+                      <span className="text-sm font-semibold text-zinc-700 dark:text-zinc-300">{formatAgentName(pendingAgentName)}</span>
                     </div>
                     <div className="flex flex-col min-w-0 w-full">
                       <div className={getStandaloneThinkingBubbleClassName(compact)}>
@@ -3273,8 +3602,26 @@ export default function SessionChat({
 
       {/* Follow-up input */}
       {!hideInput && (
-        <div className={`flex-shrink-0 bg-white ${compact ? 'px-4 py-3' : 'py-4'} dark:bg-zinc-950`}>
-          <div className={`relative min-w-0 ${!compact ? (fullWidth ? 'w-full px-5' : 'w-[min(76%,64rem)] mx-auto px-6') : ''}`}>
+        <div
+          className={`flex-shrink-0 ${
+            compact
+              ? 'bg-white px-4 py-3 dark:bg-zinc-950'
+              : pageCanvas
+                ? 'bg-transparent px-7 pb-6 pt-2'
+                : 'bg-white py-4 dark:bg-zinc-950'
+          }`}
+        >
+          <div
+            className={`relative min-w-0 ${
+              !compact
+                ? fullWidth
+                  ? 'w-full px-5'
+                  : pageCanvas
+                    ? 'mx-auto w-full max-w-[760px]'
+                    : 'w-[min(76%,64rem)] mx-auto px-6'
+                : ''
+            }`}
+          >
             {conversationBottomSlot && (
               <div className="mb-2 min-w-0">
                 {typeof conversationBottomSlot === 'function'
@@ -3348,14 +3695,16 @@ export default function SessionChat({
               onDragOver={handleComposerDragOver}
               onDragLeave={handleComposerDragLeave}
               onDrop={handleComposerDrop}
-              className={`rounded-2xl border transition-all ${
+              className={`${pageCanvas && !compact ? 'min-h-[124px] rounded-[20px] shadow-[0_3px_12px_rgba(22,27,34,0.045)]' : 'rounded-2xl'} border transition-all ${
                 isCompacting
                   ? 'border-amber-200 bg-amber-50/30 dark:border-amber-500/35 dark:bg-amber-950/25'
                   : isDragOver
                     ? 'border-sky-300 bg-sky-50/60 ring-4 ring-sky-100 dark:border-sky-500/50 dark:bg-sky-950/35 dark:ring-sky-500/10'
                     : isStreaming
                       ? 'border-zinc-200 bg-zinc-50 dark:border-zinc-800 dark:bg-zinc-900/70'
-                      : 'border-zinc-200 bg-zinc-50 hover:border-zinc-300 focus-within:border-zinc-300 focus-within:bg-white focus-within:ring-4 focus-within:ring-zinc-100 dark:border-zinc-800 dark:bg-zinc-900/70 dark:hover:border-zinc-700 dark:focus-within:border-zinc-700 dark:focus-within:bg-zinc-900 dark:focus-within:ring-zinc-800/60'
+                      : pageCanvas && !compact
+                        ? 'border-black/[0.09] bg-zinc-50 hover:border-black/[0.14] focus-within:border-black/[0.14] focus-within:bg-white focus-within:ring-4 focus-within:ring-black/[0.025] dark:border-white/[0.10] dark:bg-[#303842] dark:hover:border-white/[0.16] dark:focus-within:border-white/[0.16] dark:focus-within:bg-[#343d48] dark:focus-within:ring-white/[0.03]'
+                        : 'border-zinc-200 bg-zinc-50 hover:border-zinc-300 focus-within:border-zinc-300 focus-within:bg-white focus-within:ring-4 focus-within:ring-zinc-100 dark:border-zinc-800 dark:bg-zinc-900/70 dark:hover:border-zinc-700 dark:focus-within:border-zinc-700 dark:focus-within:bg-zinc-900 dark:focus-within:ring-zinc-800/60'
               }`}
             >
                 {/* Node reference chip */}
@@ -3477,14 +3826,77 @@ export default function SessionChat({
                     {t('chat.upload.dropHint')}
                   </div>
                 )}
-                <div className="px-4 pt-3 pb-1">
+                {composerReferences.length > 0 && (
+                  <div
+                    className="flex flex-wrap gap-1.5 px-4 pt-3"
+                    aria-label={t('chat.references.selected')}
+                  >
+                    {composerReferences.map((reference) => {
+                      const referenceStyle = reference.kind === 'subagent'
+                        ? 'border-sky-200/80 bg-sky-50 text-sky-700 dark:border-sky-400/20 dark:bg-sky-400/[0.10] dark:text-sky-200'
+                        : reference.kind === 'skill'
+                          ? 'border-emerald-200/80 bg-emerald-50 text-emerald-700 dark:border-emerald-400/20 dark:bg-emerald-400/[0.10] dark:text-emerald-200'
+                          : 'border-amber-200/80 bg-amber-50 text-amber-700 dark:border-amber-400/20 dark:bg-amber-400/[0.10] dark:text-amber-200';
+                      const ReferenceIcon = reference.kind === 'subagent'
+                        ? Bot
+                        : reference.kind === 'skill'
+                          ? BookOpen
+                          : WorkflowIcon;
+                      return (
+                        <span
+                          key={`${reference.kind}:${reference.value}`}
+                          className={`inline-flex h-7 max-w-full items-center gap-1.5 rounded-lg border px-2 text-[11px] font-medium shadow-[0_1px_2px_rgba(22,27,34,0.03)] ${referenceStyle}`}
+                        >
+                          <ReferenceIcon className="h-3 w-3 shrink-0" />
+                          <span className="shrink-0 opacity-65">
+                            {t(`chat.references.${reference.kind}`)}
+                          </span>
+                          <span className="truncate font-semibold">{reference.value}</span>
+                          <button
+                            type="button"
+                            onClick={() => {
+                              setComposerReferences((current) => current.filter(
+                                (item) => (
+                                  item.kind !== reference.kind
+                                  || item.value !== reference.value
+                                ),
+                              ));
+                              textareaRef.current?.focus();
+                            }}
+                            className="-mr-1 grid h-5 w-5 shrink-0 place-items-center rounded-md opacity-55 transition hover:bg-black/[0.06] hover:opacity-100 dark:hover:bg-white/[0.08]"
+                            aria-label={t('chat.references.remove', {
+                              type: t(`chat.references.${reference.kind}`),
+                              name: reference.value,
+                            })}
+                          >
+                            <X className="h-3 w-3" />
+                          </button>
+                        </span>
+                      );
+                    })}
+                  </div>
+                )}
+                <div className={`px-4 pb-1 ${composerReferences.length > 0 ? 'pt-2' : 'pt-3'}`}>
                   <textarea
                     ref={textareaRef}
                     value={input}
                     onChange={(e) => {
                       const val = e.target.value;
-                      setInput(val);
                       const cursor = e.target.selectionStart ?? val.length;
+                      const beforeCursor = val.slice(0, cursor);
+                      if (/(^|\s)@$/.test(beforeCursor)) {
+                        const next = `${val.slice(0, cursor - 1)}${val.slice(cursor)}`;
+                        setInput(next);
+                        setShowCommandDropdown(false);
+                        setMentionRange(null);
+                        openComposerAddMenu();
+                        requestAnimationFrame(() => {
+                          textareaRef.current?.focus();
+                          textareaRef.current?.setSelectionRange(cursor - 1, cursor - 1);
+                        });
+                        return;
+                      }
+                      setInput(val);
                       const mention = mentionAgents.length > 0 ? findMentionTrigger(val, cursor) : null;
                       const trimmed = val.trimStart();
                       const slashQuery = trimmed.startsWith('/') ? trimmed.slice(1) : '';
@@ -3523,7 +3935,7 @@ export default function SessionChat({
                             ? t('chat.placeholderNodeRef', { nodeId: nodeRef.id })
                             : effectivePlaceholder
                     }
-                    className={`w-full resize-none outline-none bg-transparent text-sm placeholder-zinc-400 dark:placeholder-zinc-600 ${
+                    className={`w-full resize-none bg-transparent text-sm outline-none placeholder:text-[#7b838e] dark:placeholder:text-[#9aa7b4] ${
                       sending ? 'text-zinc-400 cursor-not-allowed dark:text-zinc-500' : 'text-zinc-900 dark:text-zinc-100'
                     }`}
                     style={{
@@ -3537,15 +3949,61 @@ export default function SessionChat({
 
                 {/* Bottom toolbar inside the composer card */}
                 <div className="flex items-center gap-1 px-2 pb-2">
-                  <button
-                    type="button"
-                    onClick={() => fileInputRef.current?.click()}
-                    disabled={sending}
-                    title={t('chat.upload.selectWithImage')}
-                    className="inline-flex h-8 w-8 items-center justify-center rounded-lg text-zinc-500 transition-colors hover:bg-zinc-200/60 hover:text-zinc-800 disabled:cursor-not-allowed disabled:opacity-40"
-                  >
-                    <Paperclip className="h-4 w-4" />
-                  </button>
+                  <div className="relative" data-composer-add-menu>
+                    <button
+                      type="button"
+                      onClick={toggleComposerAddMenu}
+                      disabled={sending}
+                      title={t('chat.addMenu.title')}
+                      aria-label={t('chat.addMenu.title')}
+                      aria-haspopup="menu"
+                      aria-expanded={showComposerAddMenu}
+                      className={`inline-flex h-8 w-8 items-center justify-center rounded-full border transition-all duration-150 disabled:cursor-not-allowed disabled:opacity-40 ${
+                        showComposerAddMenu
+                          ? 'border-zinc-300 bg-white text-zinc-900 shadow-[0_2px_8px_rgba(22,27,34,0.08)] dark:border-white/[0.14] dark:bg-white/[0.09] dark:text-white'
+                          : 'border-transparent text-zinc-500 hover:border-zinc-200 hover:bg-white hover:text-zinc-900 dark:text-zinc-400 dark:hover:border-white/[0.10] dark:hover:bg-white/[0.07] dark:hover:text-white'
+                      }`}
+                    >
+                      <Plus className="h-[17px] w-[17px]" strokeWidth={2} />
+                    </button>
+
+                    {showComposerAddMenu && (
+                      <div
+                        role="menu"
+                        aria-label={t('chat.addMenu.title')}
+                        className="absolute bottom-full left-0 z-50 mb-2 w-[228px] max-w-[calc(100vw-2rem)] overflow-visible rounded-[14px] border border-black/[0.09] bg-white/95 p-1.5 shadow-[0_18px_46px_rgba(22,27,34,0.14),0_2px_8px_rgba(22,27,34,0.06)] backdrop-blur-xl dark:border-white/[0.11] dark:bg-[#252c35]/95 dark:shadow-[0_20px_48px_rgba(8,10,13,0.44)]"
+                      >
+                        <div className="px-2.5 pb-1.5 pt-1 text-[11px] font-medium text-zinc-400 dark:text-zinc-500">
+                          {t('chat.addMenu.title')}
+                        </div>
+                        <button
+                          type="button"
+                          role="menuitem"
+                          onClick={() => {
+                            closeComposerAddMenu();
+                            fileInputRef.current?.click();
+                          }}
+                          className="group flex h-10 w-full items-center gap-2.5 rounded-[9px] px-2 text-left text-[13px] font-medium text-zinc-700 transition-colors hover:bg-zinc-100/90 hover:text-zinc-950 dark:text-zinc-200 dark:hover:bg-white/[0.07] dark:hover:text-white"
+                        >
+                          <span className="grid h-7 w-7 shrink-0 place-items-center rounded-lg border border-zinc-200/80 bg-white text-zinc-500 shadow-[0_1px_2px_rgba(22,27,34,0.04)] transition-colors group-hover:text-zinc-800 dark:border-white/[0.10] dark:bg-white/[0.05] dark:text-zinc-400 dark:group-hover:text-white">
+                            <Paperclip className="h-3.5 w-3.5" />
+                          </span>
+                          <span className="min-w-0 flex-1 truncate">{t('chat.addMenu.files')}</span>
+                        </button>
+                        {composerAddMenuSlot && (
+                          <div className="mt-0.5 border-t border-zinc-100 pt-0.5 dark:border-white/[0.07]">
+                            {typeof composerAddMenuSlot === 'function'
+                              ? composerAddMenuSlot({
+                                closeMenu: closeComposerAddMenu,
+                                insertMention: insertAgentMention,
+                                insertReference: insertComposerReference,
+                              })
+                              : composerAddMenuSlot}
+                          </div>
+                        )}
+                      </div>
+                    )}
+                  </div>
 
                   <div className="mx-1 h-4 w-px shrink-0 bg-zinc-200 dark:bg-zinc-800" />
 
@@ -3810,14 +4268,14 @@ function ProcessGroupDetails({
     <details
       open={effectiveOpen}
       data-testid="chat-process-group"
-      className="group/process mt-2 first:mt-0 overflow-hidden rounded-lg border border-zinc-200/90 bg-white/80 shadow-none"
+      className="group/process mt-2 w-full first:mt-0 text-[#626874] dark:text-[#aab4bf]"
     >
       <summary
         onClick={handleSummaryClick}
-        className="flex cursor-pointer list-none items-center gap-2 px-2.5 py-2 text-xs text-zinc-600 transition-colors hover:bg-zinc-50"
+        className="flex min-h-7 cursor-pointer list-none items-center gap-2 text-sm font-medium transition-colors hover:text-zinc-900 dark:hover:text-zinc-100 [&::-webkit-details-marker]:hidden"
       >
         {summary}
-        <ChevronDown className="ml-auto h-3 w-3 flex-shrink-0 text-zinc-400 transition-transform group-open/process:rotate-180" />
+        <ChevronDown className="ml-0.5 h-3 w-3 flex-shrink-0 text-[#9da29f] transition-transform group-open/process:rotate-180 dark:text-zinc-500" />
       </summary>
       {children}
     </details>
@@ -3850,7 +4308,7 @@ function ChatMessageBubbleInner({
   onEditSend,
   onRegenerate,
 }: ChatMessageBubbleProps) {
-  const { t } = useTranslation('session');
+  const { t, i18n } = useTranslation('session');
   const isUser = message.role === 'user';
   const parts: MessagePart[] = Array.isArray(message.parts) ? message.parts : [];
   const { getPartExpanded, togglePart } = useReasoningToggle(parts, message.finish);
@@ -3873,11 +4331,13 @@ function ChatMessageBubbleInner({
   const rawAgentName = message.agent || 'rex';
   const agentName = rawAgentName.charAt(0).toUpperCase() + rawAgentName.slice(1);
 
-  const getTextContent = () =>
-    parts
+  const getTextContent = () => {
+    const text = parts
       .filter((p) => p.type === 'text' && p.text)
       .map((p) => p.text)
       .join('\n\n');
+    return isUser ? text : stripTaskMetadata(text);
+  };
 
   const editableTextParts = parts.filter((part): part is MessagePart & { text: string } =>
     part.type === 'text' && typeof part.text === 'string',
@@ -3888,17 +4348,26 @@ function ChatMessageBubbleInner({
   const editableRawText = latestEditablePart?.text || '';
   const isEditing = !!targetPartId && editingMessageId === targetMessageId;
   const isActionPending = actionMessageId === targetMessageId;
+  const hasProcessOutput = collapseIntermediateSteps && !isUser && parts.some((part) => {
+    if (part.type === 'reasoning' || part.type === 'thinking') {
+      return !!getRenderableThinkingText(part);
+    }
+    if (part.type !== 'tool') return false;
+    return !(part.callID && pendingQuestions?.[part.callID]);
+  });
   const instructionDisplayLabel = isUser && !isEditing && editableTextParts.length === 1
-    ? parseInstructionDisplayText(getMessagePartDisplayText(editableTextParts[0]))
+    ? parseInstructionDisplayText(getMessagePartDisplayText(editableTextParts[0], false))
     : null;
 
   const bubbleClass = instructionDisplayLabel
     ? getInstructionDisplayBubbleClassName(compact)
-    : getMessageBubbleClassName({ compact, isUser, isEditing });
+    : hasProcessOutput
+      ? `w-full max-w-full min-w-0 break-words ${compact ? 'text-sm' : 'text-[15px]'} text-[#34393e] dark:text-zinc-100`
+      : getMessageBubbleClassName({ compact, isUser, isEditing });
   const messageGroupClass = getMessageGroupClassName({ compact, isUser, isEditing });
   const actionBarClass = `flex items-center gap-1.5`;
   const editingActionBarClass = getEditingActionBarClassName();
-  const iconButtonClass = 'group/action relative inline-flex h-6 w-6 items-center justify-center rounded-full border border-gray-200/80 bg-white/80 text-gray-400 transition-colors duration-150 hover:border-gray-300 hover:text-gray-700 disabled:opacity-40 disabled:cursor-not-allowed dark:border-zinc-800 dark:bg-zinc-900/80 dark:text-zinc-500 dark:hover:border-zinc-700 dark:hover:text-zinc-200';
+  const iconButtonClass = 'group/action relative inline-flex h-6 w-6 items-center justify-center rounded-full border border-transparent bg-transparent text-[#8b929d] transition-colors duration-150 hover:bg-white hover:text-[#4f5660] active:bg-white focus-visible:bg-white focus-visible:outline-none disabled:cursor-not-allowed disabled:opacity-40 dark:text-[#9aa7b4] dark:hover:bg-white/[0.08] dark:hover:text-zinc-100 dark:active:bg-white/[0.08] dark:focus-visible:bg-white/[0.08]';
   const tooltipClass = 'pointer-events-none absolute bottom-full left-1/2 z-10 mb-1.5 -translate-x-1/2 whitespace-nowrap rounded-md bg-gray-900 px-2 py-1 text-[11px] font-medium text-white opacity-0 shadow-sm transition-opacity duration-150 group-hover/action:opacity-100';
   const messageErrorText = isUser ? '' : getMessageErrorText(message);
   const hasOnlyBlankTextParts = parts.length > 0 && parts.every((part) =>
@@ -3921,7 +4390,11 @@ function ChatMessageBubbleInner({
 
   const headerHeight = compact ? 'h-7' : 'h-8';
   const bubble = (
-    <div className={`${bubbleClass} relative`} style={{ overflowWrap: 'anywhere' }}>
+    <div
+      className={`${bubbleClass} relative`}
+      data-process-output={hasProcessOutput ? 'true' : undefined}
+      style={{ overflowWrap: 'anywhere' }}
+    >
 
       {/* Empty / loading state */}
       {(parts.length === 0 || shouldRenderAssistantErrorState) && (
@@ -3980,7 +4453,7 @@ function ChatMessageBubbleInner({
             return true;
           };
           const isRenderableTextPart = (part: MessagePart): boolean => (
-            part.type === 'text' && !!getMessagePartDisplayText(part).trim()
+            part.type === 'text' && !!getMessagePartDisplayText(part, !isUser).trim()
           );
           const isRenderableDisplayPart = (part: MessagePart): boolean => {
             if (isIntermediateProcessPart(part)) return true;
@@ -3992,21 +4465,21 @@ function ChatMessageBubbleInner({
           const activeTailPart = isActive
             ? [...displayParts].reverse().find(isRenderableDisplayPart)
             : undefined;
-          const renderPart = (part: MessagePart, i: number, isVisible = true) => (
+          const renderPart = (part: MessagePart, i: number, isVisible = true, processStep = false) => (
             // Spacing between consecutive parts is owned by this wrapper,
             // not by individual part components. Each part used to set its
             // own `mt-2 first:mt-0`, but since every part lives in its own
             // wrapper div, `first:` always matched and the gap collapsed
             // to zero between, e.g., a tool card and the next thinking
             // block, making them look glued together.
-            <div key={part.id || i} className="mt-2 first:mt-0">
+            <div key={part.id || i} className={processStep ? 'min-w-0' : 'mt-2 first:mt-0'}>
               {/* Text */}
               {part.type === 'text' && (() => {
                 const rawText = part.text || '';
                 const nodeRefMatch = isUser
                   ? rawText.match(/^@@node:([^|\n]+)\|([^\n]+)\n([\s\S]*)$/)
                   : null;
-                const partDisplayText = getMessagePartDisplayText(part);
+                const partDisplayText = getMessagePartDisplayText(part, !isUser);
                 if (!partDisplayText.trim()) return null;
                 const displayText = nodeRefMatch && partDisplayText === rawText ? nodeRefMatch[3] : partDisplayText;
                 const instructionLabel = isUser ? parseInstructionDisplayText(displayText) : null;
@@ -4026,10 +4499,22 @@ function ChatMessageBubbleInner({
                         <span className="text-[9px] text-gray-500 flex-shrink-0">{nodeRefMatch[2]}</span>
                       </div>
                     )}
-                    <StreamingMarkdown
-                      content={displayText}
-                      isStreaming={isActive && !isUser}
-                    />
+                    {processStep ? (
+                      <div
+                        data-testid="chat-process-text-step"
+                        className="min-w-0 py-1 text-sm leading-7 text-[#686e6c] dark:text-zinc-400"
+                      >
+                        <StreamingMarkdown
+                          content={displayText}
+                          isStreaming={isActive && !isUser}
+                        />
+                      </div>
+                    ) : (
+                      <StreamingMarkdown
+                        content={displayText}
+                        isStreaming={isActive && !isUser}
+                      />
+                    )}
                   </>
                 );
               })()}
@@ -4045,6 +4530,7 @@ function ChatMessageBubbleInner({
                   onReject={onQuestionReject && part.callID
                     ? () => onQuestionReject(part.callID!, pendingQuestions![part.callID!].requestId)
                     : undefined}
+                  processStep={processStep}
                 />
               )}
 
@@ -4055,6 +4541,37 @@ function ChatMessageBubbleInner({
                 const partKey = part.id || `reasoning-${i}`;
                 const isThinking = part === activeTailPart;
                 const isExpanded = isThinking || getPartExpanded(partKey);
+                if (processStep) {
+                  return (
+                    <div data-testid="chat-process-reasoning-step">
+                      <button
+                        type="button"
+                        aria-expanded={isExpanded}
+                        onClick={() => togglePart(partKey)}
+                        disabled={isThinking}
+                        className="flex min-h-7 w-full cursor-pointer items-center gap-2 text-left text-sm font-medium text-[#747a78] transition-colors hover:text-zinc-900 disabled:cursor-default dark:text-zinc-400 dark:hover:text-zinc-200"
+                      >
+                        <span className="inline-grid h-[18px] w-[18px] flex-[0_0_18px] place-items-center text-violet-500">
+                          {isThinking ? (
+                            <Loader2 className="h-[15px] w-[15px] animate-spin" />
+                          ) : (
+                            <Brain className="h-[15px] w-[15px]" />
+                          )}
+                        </span>
+                        <span className="min-w-0">{t('chat.process.deepThinking')}</span>
+                        <ChevronDown className={`ml-0.5 h-3 w-3 flex-shrink-0 text-[#9da29f] transition-transform dark:text-zinc-500 ${isExpanded ? 'rotate-180' : ''}`} />
+                      </button>
+                      {isExpanded && isVisible && (
+                        <div className="mb-[9px] ml-2 mt-[3px] max-h-52 overflow-y-auto whitespace-pre-wrap border-l border-[#e3e6e3] py-1.5 pl-[26px] pr-0 text-sm leading-7 text-[#686e6c] dark:border-zinc-700 dark:text-zinc-400">
+                          <StreamingReasoningText
+                            content={thinkingText}
+                            isStreaming={isThinking}
+                          />
+                        </div>
+                      )}
+                    </div>
+                  );
+                }
                 return (
                   // Vertical spacing is provided by the parent part wrapper
                   // (see `otherParts.map` above); keep this container neutral
@@ -4101,14 +4618,6 @@ function ChatMessageBubbleInner({
             </div>
           );
           const renderProcessGroup = (group: Array<{ part: MessagePart; index: number }>, groupIndex: number) => {
-            const reasoningCount = group.filter(({ part }) => part.type === 'reasoning' || part.type === 'thinking').length;
-            const toolCount = group.filter(({ part }) => part.type === 'tool').length;
-            const textCount = group.filter(({ part }) => part.type === 'text').length;
-            const summary = [
-              reasoningCount > 0 ? t('chat.process.reasoningCount', { count: reasoningCount }) : '',
-              toolCount > 0 ? t('chat.process.toolCount', { count: toolCount }) : '',
-              textCount > 0 ? t('chat.process.textCount', { count: textCount }) : '',
-            ].filter(Boolean).join(' · ');
             const processGroupOpen = processGroupsDefaultOpen || (processGroupsOpenWhileActive && isActive);
             const processGroupKey = `${message.id}:process:${groupIndex}`;
             const hasStoredOpenState = !!processGroupOpenState
@@ -4124,20 +4633,17 @@ function ChatMessageBubbleInner({
                 onOpenChange={(open) => onProcessGroupOpenChange?.(processGroupKey, open)}
                 summary={(
                   <>
-                    <ListTree className="h-3.5 w-3.5 flex-shrink-0 text-zinc-400" />
-                    <span className="flex-shrink-0 font-semibold text-zinc-700">
+                    <span className="inline-grid h-[18px] w-[18px] flex-[0_0_18px] place-items-center text-[#8d9a95] dark:text-zinc-500">
+                      <Eye className="h-4 w-4" />
+                    </span>
+                    <span className="min-w-0">
                       {t('chat.process.title', { count: group.length })}
                     </span>
-                    {summary && (
-                      <span className="min-w-0 truncate text-zinc-500">
-                        {summary}
-                      </span>
-                    )}
                   </>
                 )}
               >
-                <div className="border-t border-zinc-200/70 px-2.5 py-2">
-                  {group.map(({ part, index }) => renderPart(part, index, effectiveProcessGroupOpen))}
+                <div data-testid="chat-process-timeline" className="mt-3 grid gap-1 pl-1">
+                  {group.map(({ part, index }) => renderPart(part, index, effectiveProcessGroupOpen, true))}
                 </div>
               </ProcessGroupDetails>
             );
@@ -4266,11 +4772,14 @@ function ChatMessageBubbleInner({
       )}
     </div>
   );
+  const footerTimestamp = showTimestamp && message.timestamp
+    ? <span className="select-none text-[11px] text-[#8b929d] dark:text-[#9aa7b4]">{formatSmartTime(message.timestamp, i18n.language)}</span>
+    : null;
   const footer = !compact && showActions && parts.length > 0 && !isEditing ? (
-    <div className="flex items-center justify-between mt-1.5">
-      {showTimestamp && message.timestamp
-        ? <span className="text-[11px] text-zinc-400 select-none">{formatSmartTime(message.timestamp)}</span>
-        : <span />}
+    <div className={`mt-1.5 flex items-center ${
+      isUser ? 'justify-between' : `justify-start${footerTimestamp ? ' gap-1.5' : ''}`
+    }`}>
+      {isUser && (footerTimestamp || <span />)}
       <div className={actionBarClass}>
         {isUser ? (
           <>
@@ -4318,6 +4827,7 @@ function ChatMessageBubbleInner({
           </>
         )}
       </div>
+      {!isUser && footerTimestamp}
     </div>
   ) : null;
 
@@ -4373,7 +4883,7 @@ function ChatMessageBubbleInner({
         </div>
         <div className="flex w-full min-w-0 flex-col items-start">
           <div className={`flex items-center gap-2 ${headerHeight}`}>
-            <span className="text-xs font-semibold text-zinc-700 dark:text-zinc-300">
+            <span className="text-base font-semibold text-[#3f444a] dark:text-[#d7dee8]">
               {agentName}
             </span>
           </div>
@@ -4399,7 +4909,7 @@ function ChatMessageBubbleInner({
         {avatar}
         <div className="flex flex-col items-start flex-1 min-w-0">
           <div className={`flex items-center gap-2 ${headerHeight}`}>
-            <span className="text-xs font-semibold text-zinc-700 dark:text-zinc-300">
+            <span className="text-sm font-semibold text-zinc-700 dark:text-zinc-300">
               {agentName}
             </span>
           </div>
@@ -4626,15 +5136,19 @@ function formatQuestionAnswerValue(question: QuestionItem, value: string, t: Tod
 
 function ChatQuestionResult({
   state,
+  actionLabel,
   statusLabel,
   statusIcon,
   statusIconColor,
+  processStep = false,
   t,
 }: {
   state: Partial<ToolState>;
+  actionLabel: string;
   statusLabel: string;
   statusIcon: React.ReactNode;
   statusIconColor: string;
+  processStep?: boolean;
   t: TodoTranslator;
 }) {
   const questions = readQuestionItems(state.input?.questions);
@@ -4660,6 +5174,67 @@ function ChatQuestionResult({
     ? 'bg-red-50 text-red-500'
     : 'bg-zinc-100 text-zinc-500';
   const firstQuestion = questions[0]?.header || questions[0]?.question || '';
+  const questionDetails = questions.map((question, index) => {
+    const answers = normalizeQuestionAnswer(rawAnswers[index]);
+    const displayAnswers = answers.length > 0
+      ? answers.map((answer) => formatQuestionAnswerValue(question, answer, t))
+      : [t('chat.questionResult.unanswered')];
+    return (
+      <div
+        key={`${question.question}-${index}`}
+        className={`space-y-2 py-2 last:pb-0 ${processStep ? 'first:pt-0' : 'first:pt-2'}`}
+      >
+        <div className="grid grid-cols-[32px_minmax(0,1fr)] gap-2">
+          <span className="pt-0.5 text-[11px] font-medium text-zinc-400">
+            {t('chat.questionResult.questionLabel')}
+          </span>
+          <div className="min-w-0">
+            {question.header && (
+              <div className="mb-0.5 text-[11px] font-medium text-zinc-500">{question.header}</div>
+            )}
+            <div className="text-xs leading-5 text-zinc-700 dark:text-zinc-300">{question.question}</div>
+          </div>
+        </div>
+        <div className="grid grid-cols-[32px_minmax(0,1fr)] gap-2">
+          <span className={`pt-0.5 text-[11px] font-medium ${answerLabelClass}`}>
+            {t('chat.questionResult.answerLabel')}
+          </span>
+          <div className="flex min-w-0 flex-wrap gap-1.5">
+            {displayAnswers.map((answer, answerIndex) => (
+              <span
+                key={`${answer}-${answerIndex}`}
+                className={`rounded-md border px-1.5 py-0.5 text-[11px] font-medium ${answerChipClass}`}
+              >
+                {answer}
+              </span>
+            ))}
+          </div>
+        </div>
+      </div>
+    );
+  });
+
+  if (processStep) {
+    return (
+      <details data-testid="chat-process-tool-step" className="group/tool min-w-0">
+        <summary className="flex min-h-7 cursor-pointer list-none items-center gap-2 text-sm font-medium text-[#747a78] transition-colors hover:text-zinc-900 dark:text-zinc-400 dark:hover:text-zinc-200 [&::-webkit-details-marker]:hidden">
+          <span className={`inline-grid h-[18px] w-[18px] flex-[0_0_18px] place-items-center ${statusIconColor}`}>
+            {statusIcon}
+          </span>
+          <span className="min-w-0 flex-shrink-0">{actionLabel}</span>
+          {firstQuestion && (
+            <span className="ml-0.5 min-w-0 truncate font-normal text-[#9a9f9c] dark:text-zinc-500">
+              {firstQuestion}
+            </span>
+          )}
+          <ChevronDown className="ml-0.5 h-3 w-3 flex-shrink-0 text-[#9da29f] transition-transform group-open/tool:rotate-180 dark:text-zinc-500" />
+        </summary>
+        <div className="mb-[9px] ml-2 mt-[3px] space-y-1.5 border-l border-[#e3e6e3] py-1.5 pl-[26px] pr-0 text-xs text-[#686e6c] dark:border-zinc-700 dark:text-zinc-400">
+          {questionDetails}
+        </div>
+      </details>
+    );
+  }
 
   return (
     <details className="group/tool rounded-lg bg-zinc-50 overflow-hidden">
@@ -4667,7 +5242,7 @@ function ChatQuestionResult({
         <span className={`${statusIconColor} flex-shrink-0 mt-0.5`}>{statusIcon}</span>
         <div className="min-w-0 flex-1">
           <div className="flex items-center gap-2 min-w-0">
-            <span className="font-medium text-zinc-700 text-xs whitespace-nowrap flex-shrink-0">question</span>
+            <span className="font-medium text-zinc-700 text-xs whitespace-nowrap flex-shrink-0">{actionLabel}</span>
             {firstQuestion && (
               <span className="text-[11px] text-zinc-400 truncate min-w-0">
                 {firstQuestion}
@@ -4683,45 +5258,7 @@ function ChatQuestionResult({
         </div>
       </summary>
       <div className="border-t border-zinc-200/60 px-2.5 py-2 space-y-1.5 text-xs">
-        {questions.map((question, index) => {
-          const answers = normalizeQuestionAnswer(rawAnswers[index]);
-          const displayAnswers = answers.length > 0
-            ? answers.map((answer) => formatQuestionAnswerValue(question, answer, t))
-            : [t('chat.questionResult.unanswered')];
-          return (
-            <div
-              key={`${question.question}-${index}`}
-              className="space-y-2 py-2 first:pt-2 last:pb-0"
-            >
-              <div className="grid grid-cols-[32px_minmax(0,1fr)] gap-2">
-                <span className="pt-0.5 text-[11px] font-medium text-zinc-400">
-                  {t('chat.questionResult.questionLabel')}
-                </span>
-                <div className="min-w-0">
-                  {question.header && (
-                    <div className="mb-0.5 text-[11px] font-medium text-zinc-500">{question.header}</div>
-                  )}
-                  <div className="text-xs leading-5 text-zinc-700">{question.question}</div>
-                </div>
-              </div>
-              <div className="grid grid-cols-[32px_minmax(0,1fr)] gap-2">
-                <span className={`pt-0.5 text-[11px] font-medium ${answerLabelClass}`}>
-                  {t('chat.questionResult.answerLabel')}
-                </span>
-                <div className="flex min-w-0 flex-wrap gap-1.5">
-                  {displayAnswers.map((answer, answerIndex) => (
-                    <span
-                      key={`${answer}-${answerIndex}`}
-                      className={`rounded-md border px-1.5 py-0.5 text-[11px] font-medium ${answerChipClass}`}
-                    >
-                      {answer}
-                    </span>
-                  ))}
-                </div>
-              </div>
-            </div>
-          );
-        })}
+        {questionDetails}
       </div>
     </details>
   );
@@ -4954,16 +5491,17 @@ export interface ChatToolPartProps {
   pendingQuestion?: PendingQuestion;
   onAnswer?: (answers: string[][]) => Promise<void>;
   onReject?: () => Promise<void>;
+  processStep?: boolean;
 }
 
-export function ChatToolPart({ part, pendingQuestion, onAnswer, onReject }: ChatToolPartProps) {
+export function ChatToolPart({ part, pendingQuestion, onAnswer, onReject, processStep = false }: ChatToolPartProps) {
   const { t } = useTranslation('session');
   const toolName = part.tool || 'unknown';
 
   // Keep the delegate fallback narrow: many MCP tools also carry a generic
   // `category` field (for example wecom_mcp category="doc").
   if (shouldRenderDelegateTaskCard(part)) {
-    return <DelegateTaskCard part={part} />;
+    return <DelegateTaskCard part={part} processStep={processStep} />;
   }
 
   const state: Partial<ToolState> = part.state || {};
@@ -5013,20 +5551,56 @@ export function ChatToolPart({ part, pendingQuestion, onAnswer, onReject }: Chat
     : [];
   const showGenericToolPayload = toolName !== 'todo' && !isBashTool;
   const isTodoTool = toolName === 'todo';
+  const toolPresentation = resolveToolPresentation(toolName, state, t);
+  const hasToolInput = Boolean(state.input && Object.keys(state.input).length > 0);
+  const safeToolInput = hasToolInput
+    ? redactToolInput(state.input) as Record<string, unknown>
+    : undefined;
+  const showActionProgress = showGenericToolPayload
+    && !safeToolInput
+    && (status === 'pending' || status === 'running');
+  const actionProgressLabel = toolName === 'write'
+    ? t('chat.tool.progress.writingFile')
+    : toolName === 'edit' || toolName === 'apply_patch'
+      ? t('chat.tool.progress.editingFile')
+      : t('chat.tool.progress.working');
 
   // Reuse the shared helpers so the truncation rules stay in sync with the
   // delegate-task card and any other places that render tool input previews.
-  const inputSummary = state.input
+  const inputSummary = hasToolInput
     ? truncateToolDisplayText(
         toolName === 'todo'
           ? buildTodoSummary(state, t)
           : isBashTool
           ? buildBashHeaderSummary(state)
-          : buildToolInputSummary(state.input),
+          : buildToolInputSummary(safeToolInput || {}),
       )
     : '';
   const displayTitle = state.title ? truncateToolDisplayText(state.title) : '';
   const workflowHeaderSummary = truncateToolDisplayText(buildRunWorkflowHeaderSummary(toolName, state, t));
+  const semanticDetail = truncateToolDisplayText(toolPresentation.detail);
+  const fileOperationDisplayName = truncateToolDisplayText(getFileOperationDisplayName(toolName, state));
+  const toolDisplayName = toolPresentation.label;
+  const processStepLabel = isTodoTool
+    ? t('chat.tool.todoUpdated')
+    : toolPresentation.known
+      ? toolDisplayName
+      : config.label;
+  const processStepDetail = fileOperationDisplayName
+    || workflowHeaderSummary
+    || semanticDetail
+    || displayTitle
+    || inputSummary
+    || (toolPresentation.known ? '' : toolDisplayName);
+  const cardHeaderDetail = fileOperationDisplayName
+    || workflowHeaderSummary
+    || semanticDetail
+    || inputSummary
+    || displayTitle;
+  const processStepIcon = isTodoTool
+    ? <ListTree className="h-[15px] w-[15px]" />
+    : config.icon;
+  const processStepIconColor = isTodoTool ? 'text-slate-500' : config.iconColor;
   const statusBadgeClass = isTodoTool
     ? 'text-[11px] font-medium text-zinc-500'
     : `text-[11px] font-medium px-1.5 py-0.5 rounded-md ${config.pill}`;
@@ -5049,14 +5623,123 @@ export function ChatToolPart({ part, pendingQuestion, onAnswer, onReject }: Chat
     return (
       <ChatQuestionResult
         state={state}
+        actionLabel={toolDisplayName}
         statusLabel={config.label}
         statusIcon={config.icon}
         statusIconColor={config.iconColor}
+        processStep={processStep}
         t={t}
       />
     );
   }
 
+  const toolDetails = (
+    <>
+      {isTodoTool && todoEntries.length > 0 && (
+        <div className="space-y-1.5">
+          <div className="flex items-center justify-between gap-3 text-[11px] font-medium text-zinc-500">
+            <span>{t('chat.tool.todoStages')}</span>
+            <span className="font-normal text-zinc-400">{todoEntries.length}</span>
+          </div>
+          <div className="divide-y divide-zinc-100 dark:divide-zinc-800">
+            {todoEntries.map((todo, index) => (
+              <div
+                key={todo.id || index}
+                className="grid grid-cols-[16px_minmax(0,1fr)_auto] items-start gap-2 py-1.5 text-[11px] first:pt-0 last:pb-0"
+              >
+                <span className="mt-0.5 flex h-4 w-4 flex-shrink-0 items-center justify-center">
+                  {todoStatusIcon(todo.status)}
+                </span>
+                <span className={`min-w-0 leading-5 ${todoTextClass(todo.status)}`}>
+                  {todo.activeForm && todo.status === 'in_progress' ? todo.activeForm : todo.content}
+                </span>
+                <span
+                  className={`flex-shrink-0 whitespace-nowrap leading-5 ${todoStatusLabelClass(todo.status)}`}
+                >
+                  {todoStatusLabel(todo.status, t)}
+                </span>
+              </div>
+            ))}
+          </div>
+        </div>
+      )}
+
+      {isBashTool && (
+        <ChatBashPayload state={state} t={t} />
+      )}
+
+      {showActionProgress && (
+        <div
+          data-testid="chat-tool-action-progress"
+          className="flex min-h-6 items-center gap-2 py-0.5 text-xs text-[#8d9390] dark:text-zinc-500"
+        >
+          <span className="inline-flex items-center gap-0.5" aria-hidden="true">
+            <span className="h-1 w-1 animate-pulse rounded-full bg-current" />
+            <span className="h-1 w-1 animate-pulse rounded-full bg-current [animation-delay:150ms]" />
+            <span className="h-1 w-1 animate-pulse rounded-full bg-current [animation-delay:300ms]" />
+          </span>
+          <span>{actionProgressLabel}</span>
+        </div>
+      )}
+
+      {showGenericToolPayload && safeToolInput && (
+        <details>
+          <summary className="mb-1 cursor-pointer text-[11px] font-medium text-zinc-500 transition-colors hover:text-zinc-700 dark:hover:text-zinc-300">
+            {t('chat.tool.inputParams')}
+          </summary>
+          <pre className="overflow-x-auto rounded-md bg-zinc-950 p-2 font-mono text-[11px] leading-relaxed text-zinc-300">
+            {JSON.stringify(safeToolInput, null, 2)}
+          </pre>
+        </details>
+      )}
+
+      {showGenericToolPayload && status === 'completed' && state.output !== undefined && (
+        <details open>
+          <summary className="mb-1 cursor-pointer text-[11px] font-medium text-zinc-500 transition-colors hover:text-zinc-700 dark:hover:text-zinc-300">
+            {t('chat.tool.outputResult')}
+          </summary>
+          <pre className="max-h-48 overflow-x-auto overflow-y-auto rounded-md bg-zinc-950 p-2 font-mono text-[11px] leading-relaxed text-green-400">
+            {formatToolPayload(state.output)}
+          </pre>
+        </details>
+      )}
+
+      {status === 'error' && state.error && (
+        <div className="rounded-md border border-red-100 bg-red-50 px-2.5 py-1.5 text-[11px] text-red-600 dark:border-red-900/60 dark:bg-red-950/40 dark:text-red-300">
+          {state.error}
+        </div>
+      )}
+
+      {state.time?.start && state.time?.end && (
+        <div className="text-right text-[10px] text-zinc-400">
+          {((state.time.end - state.time.start) / 1000).toFixed(2)}s
+        </div>
+      )}
+    </>
+  );
+
+  if (processStep) {
+    return (
+      <details data-testid="chat-process-tool-step" className="group/tool min-w-0">
+        <summary className="flex min-h-7 cursor-pointer list-none items-center gap-2 text-sm font-medium text-[#747a78] transition-colors hover:text-zinc-900 dark:text-zinc-400 dark:hover:text-zinc-200 [&::-webkit-details-marker]:hidden">
+          <span className={`inline-grid h-[18px] w-[18px] flex-[0_0_18px] place-items-center ${processStepIconColor}`}>
+            {processStepIcon}
+          </span>
+          <span className="min-w-0 flex-shrink-0">{processStepLabel}</span>
+          {processStepDetail && (
+            <span className="ml-0.5 min-w-0 truncate font-normal text-[#9a9f9c] dark:text-zinc-500">
+              {processStepDetail}
+            </span>
+          )}
+          <ChevronDown className="ml-0.5 h-3 w-3 flex-shrink-0 text-[#9da29f] transition-transform group-open/tool:rotate-180 dark:text-zinc-500" />
+        </summary>
+        <div className="mb-[9px] ml-2 mt-[3px] space-y-1.5 border-l border-[#e3e6e3] py-1.5 pl-[26px] pr-0 text-xs text-[#686e6c] dark:border-zinc-700 dark:text-zinc-400">
+          {toolDetails}
+        </div>
+      </details>
+    );
+  }
+
   return (
     // No top margin here — the part wrapper in SessionChat owns vertical
     // spacing so every adjacent tool / thinking / text part is separated by a
@@ -5066,28 +5749,15 @@ export function ChatToolPart({ part, pendingQuestion, onAnswer, onReject }: Chat
         <span className={`${config.iconColor} flex-shrink-0 mt-0.5`}>{config.icon}</span>
         <div className="min-w-0 flex-1">
           <div className="flex items-center gap-2 min-w-0">
-            <span className="font-medium text-zinc-700 text-xs whitespace-nowrap flex-shrink-0">{toolName.replace(/_/g, ' ')}</span>
-            {workflowHeaderSummary ? (
-              <span className="text-[11px] text-emerald-700 truncate min-w-0">
-                {workflowHeaderSummary}
+            <span className="font-medium text-zinc-700 text-xs whitespace-nowrap flex-shrink-0">{toolDisplayName}</span>
+            {cardHeaderDetail && (
+              <span
+                className={`text-[11px] truncate min-w-0 ${
+                  workflowHeaderSummary ? 'text-emerald-700' : 'text-zinc-400'
+                }`}
+              >
+                {cardHeaderDetail}
               </span>
-            ) : (
-              <>
-                {inputSummary && (
-                  <span
-                    className="text-[11px] text-zinc-400 font-mono truncate min-w-0"
-                  >
-                    {inputSummary}
-                  </span>
-                )}
-                {displayTitle && !inputSummary && (
-                  <span
-                    className="text-[11px] text-zinc-400 truncate min-w-0"
-                  >
-                    {displayTitle}
-                  </span>
-                )}
-              </>
             )}
           </div>
         </div>
@@ -5099,73 +5769,8 @@ export function ChatToolPart({ part, pendingQuestion, onAnswer, onReject }: Chat
         </div>
       </summary>
 
-      <div className="border-t border-zinc-200/60 px-2.5 py-2 space-y-1.5 text-xs">
-        {isTodoTool && todoEntries.length > 0 && (
-          <div className="space-y-1.5">
-            <div className="flex items-center justify-between gap-3 text-[11px] font-medium text-zinc-500">
-              <span>{t('chat.tool.todoStages')}</span>
-              <span className="font-normal text-zinc-400">{todoEntries.length}</span>
-            </div>
-            <div className="divide-y divide-zinc-100">
-              {todoEntries.map((todo, index) => (
-                <div
-                  key={todo.id || index}
-                  className="grid grid-cols-[16px_minmax(0,1fr)_auto] items-start gap-2 py-1.5 text-[11px] first:pt-0 last:pb-0"
-                >
-                  <span className="mt-0.5 flex h-4 w-4 flex-shrink-0 items-center justify-center">
-                    {todoStatusIcon(todo.status)}
-                  </span>
-                  <span className={`min-w-0 leading-5 ${todoTextClass(todo.status)}`}>
-                    {todo.activeForm && todo.status === 'in_progress' ? todo.activeForm : todo.content}
-                  </span>
-                  <span
-                    className={`flex-shrink-0 whitespace-nowrap leading-5 ${todoStatusLabelClass(todo.status)}`}
-                  >
-                    {todoStatusLabel(todo.status, t)}
-                  </span>
-                </div>
-              ))}
-            </div>
-          </div>
-        )}
-
-        {isBashTool && (
-          <ChatBashPayload state={state} t={t} />
-        )}
-
-        {showGenericToolPayload && state.input && (
-          <details>
-            <summary className="cursor-pointer text-[11px] text-zinc-500 font-medium hover:text-zinc-700 transition-colors mb-1">
-              {t('chat.tool.inputParams')}
-            </summary>
-            <pre className="p-2 bg-zinc-950 text-zinc-300 rounded-md text-[11px] overflow-x-auto font-mono leading-relaxed">
-              {JSON.stringify(state.input, null, 2)}
-            </pre>
-          </details>
-        )}
-
-        {showGenericToolPayload && status === 'completed' && state.output !== undefined && (
-          <details open>
-            <summary className="cursor-pointer text-[11px] text-zinc-500 font-medium hover:text-zinc-700 transition-colors mb-1">
-              {t('chat.tool.outputResult')}
-            </summary>
-            <pre className="p-2 bg-zinc-950 text-green-400 rounded-md text-[11px] overflow-x-auto max-h-48 overflow-y-auto font-mono leading-relaxed">
-              {formatToolPayload(state.output)}
-            </pre>
-          </details>
-        )}
-
-        {status === 'error' && state.error && (
-          <div className="px-2.5 py-1.5 bg-red-50 border border-red-100 rounded-md text-[11px] text-red-600">
-            {state.error}
-          </div>
-        )}
-
-        {state.time?.start && state.time?.end && (
-          <div className="text-zinc-400 text-right text-[10px]">
-            {((state.time.end - state.time.start) / 1000).toFixed(2)}s
-          </div>
-        )}
+      <div className="space-y-1.5 border-t border-zinc-200/60 px-2.5 py-2 text-xs">
+        {toolDetails}
       </div>
     </details>
   );
diff --git a/webui/src/components/common/StreamingMarkdown.test.tsx b/webui/src/components/common/StreamingMarkdown.test.tsx
index 08c2437ed..d4bd29f2e 100644
--- a/webui/src/components/common/StreamingMarkdown.test.tsx
+++ b/webui/src/components/common/StreamingMarkdown.test.tsx
@@ -330,4 +330,18 @@ describe('StreamingMarkdown', () => {
     expect(paragraph?.querySelectorAll('br')).toHaveLength(2);
     expect(paragraph?.textContent).toBe('first line\nsecond line\nthird line');
   });
+
+  it('renders inline code as a lightweight chip', () => {
+    const { container } = render(
+      <StreamingMarkdown content={'Saved to `.flocks/plans/plan.md`.'} isStreaming={false} />,
+    );
+
+    expect(container.querySelector('code')).toHaveClass(
+      'rounded-md',
+      'bg-zinc-100/80',
+      'font-normal',
+      'text-zinc-700',
+    );
+    expect(container.querySelector('code')).not.toHaveClass('font-semibold');
+  });
 });
diff --git a/webui/src/components/common/StreamingMarkdown.tsx b/webui/src/components/common/StreamingMarkdown.tsx
index 3691267eb..bb4e81cd0 100644
--- a/webui/src/components/common/StreamingMarkdown.tsx
+++ b/webui/src/components/common/StreamingMarkdown.tsx
@@ -265,7 +265,7 @@ const MarkdownContent = memo(function MarkdownContent({ content }: { content: st
             if (!isBlock) {
               return (
                 <code
-                  className="bg-gray-100 text-gray-800 px-1 py-0.5 rounded text-[0.85em] font-mono"
+                  className="rounded-md bg-zinc-100/80 px-1.5 py-0.5 font-mono text-[0.85em] font-normal text-zinc-700 dark:bg-white/[0.07] dark:text-zinc-200"
                   {...props}
                 >
                   {children}
diff --git a/webui/src/components/common/toolPresentation.test.ts b/webui/src/components/common/toolPresentation.test.ts
new file mode 100644
index 000000000..b87b64ab1
--- /dev/null
+++ b/webui/src/components/common/toolPresentation.test.ts
@@ -0,0 +1,197 @@
+import { describe, expect, it } from 'vitest';
+
+import enSession from '@/locales/en-US/session.json';
+import zhSession from '@/locales/zh-CN/session.json';
+
+import {
+  getFileOperationDisplayName,
+  redactToolInput,
+  resolveToolPresentation,
+  type ToolPresentationTranslator,
+} from './toolPresentation';
+
+function translator(locale: Record<string, unknown>): ToolPresentationTranslator {
+  return (key) => {
+    const value = key.split('.').reduce<unknown>((current, segment) => {
+      if (!current || typeof current !== 'object') return undefined;
+      return (current as Record<string, unknown>)[segment];
+    }, locale);
+    return typeof value === 'string' ? value : key;
+  };
+}
+
+const zh = translator(zhSession);
+const en = translator(enSession);
+
+const BUILT_IN_TOOLS = [
+  'read',
+  'write',
+  'edit',
+  'apply_patch',
+  'glob',
+  'doc_parser',
+  'bash',
+  'grep',
+  'lsp',
+  'websearch',
+  'webfetch',
+  'delegate_task',
+  'task',
+  'schedule_task_create',
+  'schedule_task_list',
+  'schedule_task_status',
+  'schedule_task_update',
+  'schedule_task_delete',
+  'schedule_task_rerun',
+  'todo',
+  'run_workflow',
+  'run_workflow_node',
+  'workflow_config_manage',
+  'ssh_host_cmd',
+  'ssh_run_script',
+  'question',
+  'list_providers',
+  'add_provider',
+  'add_model',
+  'memory_search',
+  'memory_get',
+  'memory_write',
+  'flocks_mcp',
+  'session_manage',
+  'run_slash_command',
+  'tool_search',
+  'flocks_skills',
+  'skill_load',
+  'device_manage',
+  'channel_message',
+  'im_send_message',
+  'wecom_mcp',
+  'get_time',
+];
+
+describe('resolveToolPresentation', () => {
+  it('provides localized action names for every user-facing built-in tool', () => {
+    for (const toolName of BUILT_IN_TOOLS) {
+      const zhPresentation = resolveToolPresentation(toolName, {}, zh);
+      const enPresentation = resolveToolPresentation(toolName, {}, en);
+
+      expect(zhPresentation.known, toolName).toBe(true);
+      expect(enPresentation.known, toolName).toBe(true);
+      expect(zhPresentation.label, toolName).not.toContain('chat.tool.');
+      expect(enPresentation.label, toolName).not.toContain('chat.tool.');
+    }
+  });
+
+  it('uses dynamic skill, MCP, session, and device action names', () => {
+    expect(resolveToolPresentation(
+      'flocks_skills',
+      { input: { subcommand: 'install', args: 'agent-builder' } },
+      zh,
+    )).toMatchObject({ label: '安装技能', detail: 'agent-builder' });
+
+    expect(resolveToolPresentation(
+      'flocks_mcp',
+      { input: { subcommand: 'connect', name: 'brave-search' } },
+      zh,
+    )).toMatchObject({ label: '连接 MCP 服务', detail: 'brave-search' });
+
+    expect(resolveToolPresentation(
+      'session_manage',
+      { input: { action: 'archive', archive: false, session_id: 'ses-1' } },
+      zh,
+    )).toMatchObject({ label: '恢复任务', detail: 'ses-1' });
+
+    expect(resolveToolPresentation(
+      'device_manage',
+      { input: { action: 'connectivity_test', device_name: 'SOC 主设备' } },
+      zh,
+    )).toMatchObject({ label: '检测设备连接', detail: 'SOC 主设备' });
+  });
+
+  it('extracts concise targets instead of exposing raw input summaries', () => {
+    expect(resolveToolPresentation(
+      'read',
+      { input: { filePath: '/repo/SessionChat.tsx', offset: 5200 } },
+      zh,
+    )).toMatchObject({
+      label: '读取文件',
+      detail: '/repo/SessionChat.tsx · 5200',
+    });
+
+    expect(resolveToolPresentation(
+      'webfetch',
+      { input: { url: 'https://docs.example.com/releases/latest' } },
+      zh,
+    )).toMatchObject({
+      label: '读取网页',
+      detail: 'docs.example.com/releases/latest',
+    });
+
+    expect(resolveToolPresentation(
+      'apply_patch',
+      {
+        input: {
+          patchText: [
+            '*** Begin Patch',
+            '*** Update File: src/a.ts',
+            '*** Add File: src/b.ts',
+            '*** End Patch',
+          ].join('\n'),
+        },
+      },
+      zh,
+    )).toMatchObject({
+      label: '修改文件',
+      detail: 'src/a.ts +1',
+    });
+  });
+
+  it('keeps unknown plugin tools on the generic fallback', () => {
+    expect(resolveToolPresentation('custom_plugin_action', {}, zh)).toEqual({
+      known: false,
+      label: 'custom plugin action',
+      detail: '',
+    });
+  });
+});
+
+describe('getFileOperationDisplayName', () => {
+  it('uses the first filename and changed-file count for apply_patch', () => {
+    expect(getFileOperationDisplayName('apply_patch', {
+      input: {
+        patchText: [
+          '*** Begin Patch',
+          '*** Update File: /repo/src/a.ts',
+          '*** Add File: C:\\repo\\src\\b.ts',
+          '*** End Patch',
+        ].join('\n'),
+      },
+    })).toBe('a.ts +1');
+  });
+});
+
+describe('redactToolInput', () => {
+  it('redacts sensitive fields recursively without hiding ordinary parameters', () => {
+    expect(redactToolInput({
+      api_key: 'secret-api-key',
+      config: {
+        password: 'secret-password',
+        base_url: 'https://example.com',
+      },
+      headers: [
+        { Authorization: 'Bearer secret-token' },
+      ],
+      key_path: '/Users/rex/.ssh/id_ed25519',
+    })).toEqual({
+      api_key: '••••••',
+      config: {
+        password: '••••••',
+        base_url: 'https://example.com',
+      },
+      headers: [
+        { Authorization: '••••••' },
+      ],
+      key_path: '/Users/rex/.ssh/id_ed25519',
+    });
+  });
+});
diff --git a/webui/src/components/common/toolPresentation.ts b/webui/src/components/common/toolPresentation.ts
new file mode 100644
index 000000000..4c0d06945
--- /dev/null
+++ b/webui/src/components/common/toolPresentation.ts
@@ -0,0 +1,347 @@
+import type { ToolState } from '@/types';
+
+export type ToolPresentationTranslator = (
+  key: string,
+  options?: Record<string, unknown>,
+) => string;
+
+export interface ToolPresentation {
+  known: boolean;
+  label: string;
+  detail: string;
+}
+
+const STATIC_LABEL_KEYS: Record<string, string> = {
+  read: 'chat.tool.actions.readFile',
+  write: 'chat.tool.actions.writeFile',
+  edit: 'chat.tool.actions.editFile',
+  apply_patch: 'chat.tool.actions.editFile',
+  glob: 'chat.tool.actions.findFiles',
+  doc_parser: 'chat.tool.actions.parseDocument',
+  bash: 'chat.tool.actions.executeCommand',
+  grep: 'chat.tool.actions.searchContent',
+  lsp: 'chat.tool.actions.analyzeCode',
+  websearch: 'chat.tool.actions.searchWeb',
+  webfetch: 'chat.tool.actions.fetchWeb',
+  delegate_task: 'chat.tool.actions.delegateTask',
+  task: 'chat.tool.actions.delegateTask',
+  schedule_task_create: 'chat.tool.actions.createScheduledTask',
+  schedule_task_list: 'chat.tool.actions.listScheduledTasks',
+  schedule_task_status: 'chat.tool.actions.viewScheduledTask',
+  schedule_task_update: 'chat.tool.actions.updateScheduledTask',
+  schedule_task_delete: 'chat.tool.actions.deleteScheduledTask',
+  schedule_task_rerun: 'chat.tool.actions.rerunScheduledTask',
+  todo: 'chat.tool.todoUpdated',
+  run_workflow: 'chat.tool.actions.runWorkflow',
+  run_workflow_node: 'chat.tool.actions.runWorkflowNode',
+  question: 'chat.tool.actions.askQuestion',
+  memory_search: 'chat.tool.actions.searchMemory',
+  memory_get: 'chat.tool.actions.readMemory',
+  memory_write: 'chat.tool.actions.writeMemory',
+  list_providers: 'chat.tool.actions.listProviders',
+  add_provider: 'chat.tool.actions.addProvider',
+  add_model: 'chat.tool.actions.addModel',
+  run_slash_command: 'chat.tool.actions.runSlashCommand',
+  tool_search: 'chat.tool.actions.searchTools',
+  skill_load: 'chat.tool.loadSkill',
+  flocks_skills: 'chat.tool.actions.manageSkills',
+  flocks_mcp: 'chat.tool.actions.manageMcp',
+  session_manage: 'chat.tool.actions.manageTasks',
+  workflow_config_manage: 'chat.tool.actions.manageWorkflowConfig',
+  device_manage: 'chat.tool.actions.manageDevices',
+  ssh_host_cmd: 'chat.tool.actions.runRemoteCommand',
+  ssh_run_script: 'chat.tool.actions.runRemoteScript',
+  channel_message: 'chat.tool.actions.sendMessage',
+  im_send_message: 'chat.tool.actions.sendMessage',
+  wecom_mcp: 'chat.tool.actions.useWeCom',
+  get_time: 'chat.tool.actions.getTime',
+};
+
+const ACTION_LABEL_KEYS: Record<string, Record<string, string>> = {
+  lsp: {
+    goToDefinition: 'chat.tool.actions.findDefinition',
+    findReferences: 'chat.tool.actions.findReferences',
+    hover: 'chat.tool.actions.inspectSymbol',
+    documentSymbol: 'chat.tool.actions.listSymbols',
+    workspaceSymbol: 'chat.tool.actions.searchSymbols',
+    goToImplementation: 'chat.tool.actions.findImplementations',
+    prepareCallHierarchy: 'chat.tool.actions.analyzeCallHierarchy',
+    incomingCalls: 'chat.tool.actions.findCallers',
+    outgoingCalls: 'chat.tool.actions.findCallees',
+  },
+  flocks_skills: {
+    find: 'chat.tool.actions.searchSkills',
+    install: 'chat.tool.actions.installSkill',
+    status: 'chat.tool.actions.checkSkills',
+    'install-deps': 'chat.tool.actions.installSkillDependencies',
+    remove: 'chat.tool.actions.removeSkill',
+  },
+  flocks_mcp: {
+    list: 'chat.tool.actions.listMcpServers',
+    add: 'chat.tool.actions.addMcpServer',
+    remove: 'chat.tool.actions.removeMcpServer',
+    connect: 'chat.tool.actions.connectMcpServer',
+    disconnect: 'chat.tool.actions.disconnectMcpServer',
+  },
+  session_manage: {
+    list: 'chat.tool.actions.listTasks',
+    get: 'chat.tool.actions.viewTask',
+    create: 'chat.tool.actions.createTask',
+    update: 'chat.tool.actions.updateTask',
+    delete: 'chat.tool.actions.deleteTask',
+    archive: 'chat.tool.actions.archiveTask',
+  },
+  workflow_config_manage: {
+    get: 'chat.tool.actions.readWorkflowConfig',
+    status: 'chat.tool.actions.viewWorkflowConfigStatus',
+    sync: 'chat.tool.actions.syncWorkflowConfig',
+    diff: 'chat.tool.actions.compareWorkflowConfig',
+    put: 'chat.tool.actions.updateWorkflowConfig',
+  },
+  device_manage: {
+    list: 'chat.tool.actions.listDevices',
+    list_templates: 'chat.tool.actions.listDeviceTemplates',
+    create: 'chat.tool.actions.createDevice',
+    update: 'chat.tool.actions.updateDevice',
+    connectivity_test: 'chat.tool.actions.testDeviceConnection',
+  },
+  wecom_mcp: {
+    list: 'chat.tool.actions.listWeComTools',
+    call: 'chat.tool.actions.useWeCom',
+  },
+};
+
+const SENSITIVE_KEY_PATTERN =
+  /(?:api[_-]?key|password|passwd|token|secret|authorization|credential|private[_-]?key|access[_-]?key|cookie)/i;
+const FILE_OPERATION_TOOL_NAMES = new Set([
+  'read',
+  'write',
+  'edit',
+  'apply_patch',
+  'read_file',
+  'write_file',
+  'edit_file',
+  'create_file',
+  'delete_file',
+]);
+
+function stringValue(
+  input: Record<string, unknown>,
+  ...keys: string[]
+): string {
+  for (const key of keys) {
+    const value = input[key];
+    if (typeof value === 'string' && value.trim()) return value.trim();
+    if (typeof value === 'number' && Number.isFinite(value)) return String(value);
+  }
+  return '';
+}
+
+function joinDetail(...parts: Array<string | undefined>): string {
+  return parts.filter((part): part is string => Boolean(part?.trim())).join(' · ');
+}
+
+function urlDetail(value: string): string {
+  if (!value) return '';
+  try {
+    const url = new URL(value);
+    return `${url.hostname}${url.pathname === '/' ? '' : url.pathname}`;
+  } catch {
+    return value;
+  }
+}
+
+function workflowName(value: unknown): string {
+  if (typeof value !== 'string' || !value.trim()) return '';
+  const normalized = value.trim().replace(/\\/g, '/');
+  const lastSegment = normalized.split('/').filter(Boolean).pop() || normalized;
+  return lastSegment.replace(/\.json$/i, '');
+}
+
+function patchPaths(value: unknown): string[] {
+  if (typeof value !== 'string') return [];
+  return Array.from(
+    value.matchAll(/^\*\*\* (?:Add|Update|Delete) File:\s*(.+)$/gm),
+    (match) => match[1]?.trim(),
+  ).filter((path): path is string => Boolean(path));
+}
+
+function patchDetail(value: unknown): string {
+  const paths = patchPaths(value);
+  if (paths.length === 0) return '';
+  if (paths.length === 1) return paths[0];
+  return `${paths[0]} +${paths.length - 1}`;
+}
+
+function fileName(value: string): string {
+  const normalized = value.trim().replace(/[\\/]+$/g, '');
+  if (!normalized) return '';
+  return normalized.split(/[\\/]/).filter(Boolean).pop() || normalized;
+}
+
+export function getFileOperationDisplayName(
+  toolName: string,
+  state: Partial<ToolState>,
+): string {
+  if (!FILE_OPERATION_TOOL_NAMES.has(toolName)) return '';
+
+  if (toolName === 'apply_patch') {
+    const paths = patchPaths(state.input?.patchText);
+    if (paths.length > 0) {
+      const firstFileName = fileName(paths[0]);
+      return paths.length === 1 ? firstFileName : `${firstFileName} +${paths.length - 1}`;
+    }
+  }
+
+  const path = stringValue(
+    state.input || {},
+    'filePath',
+    'filepath',
+    'file_path',
+    'path',
+  ) || (typeof state.title === 'string' ? state.title : '');
+  return fileName(path);
+}
+
+function resolveLabelKey(toolName: string, input: Record<string, unknown>): string | undefined {
+  if (toolName === 'session_manage' && input.action === 'archive' && input.archive === false) {
+    return 'chat.tool.actions.restoreTask';
+  }
+  const action = stringValue(input, 'action', 'subcommand', 'operation');
+  return ACTION_LABEL_KEYS[toolName]?.[action] || STATIC_LABEL_KEYS[toolName];
+}
+
+function buildDetail(
+  toolName: string,
+  state: Partial<ToolState>,
+): string {
+  const input = state.input || {};
+  const metadata = state.metadata || {};
+
+  switch (toolName) {
+    case 'read':
+      return joinDetail(
+        stringValue(input, 'filePath'),
+        stringValue(input, 'offset', 'from_line'),
+      );
+    case 'write':
+    case 'edit':
+      return stringValue(input, 'filePath');
+    case 'apply_patch':
+      return patchDetail(input.patchText);
+    case 'glob':
+      return joinDetail(stringValue(input, 'pattern'), stringValue(input, 'path'));
+    case 'grep':
+      return joinDetail(stringValue(input, 'pattern'), stringValue(input, 'path'));
+    case 'doc_parser':
+      return stringValue(input, 'input_path');
+    case 'bash':
+      return stringValue(input, 'description', 'command');
+    case 'lsp':
+      return joinDetail(
+        stringValue(input, 'filePath'),
+        input.line === undefined ? '' : `L${String(input.line)}`,
+      );
+    case 'websearch':
+    case 'tool_search':
+    case 'memory_search':
+      return stringValue(input, 'query');
+    case 'webfetch':
+      return urlDetail(stringValue(input, 'url'));
+    case 'skill_load':
+      return stringValue(input, 'name', 'skill_name');
+    case 'flocks_skills':
+      return stringValue(input, 'args');
+    case 'memory_get':
+    case 'memory_write':
+      return stringValue(input, 'path');
+    case 'schedule_task_create':
+      return joinDetail(
+        stringValue(input, 'title'),
+        stringValue(input, 'cron_description', 'run_at', 'cron', 'schedule'),
+      );
+    case 'schedule_task_list':
+      return stringValue(input, 'status', 'type');
+    case 'schedule_task_status':
+    case 'schedule_task_delete':
+    case 'schedule_task_rerun':
+      return stringValue(input, 'task_id');
+    case 'schedule_task_update':
+      return stringValue(input, 'title', 'task_id');
+    case 'run_workflow':
+    case 'run_workflow_node':
+      return joinDetail(
+        stringValue(metadata, 'workflow_name') || workflowName(input.workflow),
+        stringValue(input, 'node_id'),
+      );
+    case 'workflow_config_manage':
+      return joinDetail(
+        stringValue(input, 'workflow_id'),
+        stringValue(input, 'config_type'),
+      );
+    case 'flocks_mcp':
+      return stringValue(input, 'name');
+    case 'session_manage':
+      return stringValue(input, 'title', 'session_id', 'project_id');
+    case 'device_manage':
+      return stringValue(input, 'device_name', 'device_id', 'storage_key');
+    case 'ssh_host_cmd':
+      return joinDetail(stringValue(input, 'host'), stringValue(input, 'command'));
+    case 'ssh_run_script':
+      return joinDetail(stringValue(input, 'host'), stringValue(input, 'script_path'));
+    case 'channel_message':
+      return joinDetail(
+        stringValue(input, 'channel_type'),
+        stringValue(input, 'chat_id', 'session_id'),
+      );
+    case 'im_send_message':
+      return joinDetail(
+        stringValue(input, 'channel_type'),
+        stringValue(input, 'target', 'session_id'),
+      );
+    case 'wecom_mcp':
+      return joinDetail(stringValue(input, 'category'), stringValue(input, 'method'));
+    case 'list_providers':
+      return stringValue(input, 'provider_id');
+    case 'add_provider':
+      return stringValue(input, 'name');
+    case 'add_model':
+      return joinDetail(stringValue(input, 'name', 'model_id'), stringValue(input, 'provider_id'));
+    case 'run_slash_command':
+      return joinDetail(stringValue(input, 'command'), stringValue(input, 'arguments'));
+    default:
+      return '';
+  }
+}
+
+export function resolveToolPresentation(
+  rawToolName: string,
+  state: Partial<ToolState>,
+  t: ToolPresentationTranslator,
+): ToolPresentation {
+  const toolName = rawToolName === 'load_skill' ? 'skill_load' : rawToolName;
+  const input = state.input || {};
+  const labelKey = resolveLabelKey(toolName, input);
+  return {
+    known: Boolean(labelKey),
+    label: labelKey ? t(labelKey) : rawToolName.replace(/_/g, ' '),
+    detail: buildDetail(toolName, state),
+  };
+}
+
+export function redactToolInput(value: unknown, key = ''): unknown {
+  if (key && SENSITIVE_KEY_PATTERN.test(key)) return '••••••';
+  if (Array.isArray(value)) {
+    return value.map((item) => redactToolInput(item));
+  }
+  if (value && typeof value === 'object') {
+    return Object.fromEntries(
+      Object.entries(value as Record<string, unknown>)
+        .map(([childKey, childValue]) => [
+          childKey,
+          redactToolInput(childValue, childKey),
+        ]),
+    );
+  }
+  return value;
+}
diff --git a/webui/src/components/common/useRexComposerControls.tsx b/webui/src/components/common/useRexComposerControls.tsx
index 2a8c602d0..f2bfbf010 100644
--- a/webui/src/components/common/useRexComposerControls.tsx
+++ b/webui/src/components/common/useRexComposerControls.tsx
@@ -18,17 +18,23 @@ export function useRexComposerControls() {
   const {
     groupedOptions,
     loading,
+    effectiveModelOption,
+    modelPickerAutoOption,
+    selectModelKey,
+    selectedModelAuto,
     selectedModelOption,
     selectedPromptModel,
-    setSelectedModelKey,
-  } = useChatModelOptions();
+    setSelectedModelAuto,
+  } = useChatModelOptions({ enableAuto: true });
 
   return useMemo(() => ({
     rexAgentName: REX_AGENT_NAME,
     rexMentionAgents: agents,
     rexModel: selectedPromptModel,
-    rexSupportsVision: selectedModelOption?.supportsVision ?? defaultSupportsVision,
-    rexContextWindowTokens: selectedModelOption?.contextWindowTokens ?? null,
+    rexModelAuto: selectedModelAuto,
+    onRexModelAutoChange: setSelectedModelAuto,
+    rexSupportsVision: effectiveModelOption?.supportsVision ?? defaultSupportsVision,
+    rexContextWindowTokens: effectiveModelOption?.contextWindowTokens ?? null,
     rexComposerTextareaMinHeight: 48,
     rexComposerTextareaMaxHeight: 120,
     rexToolbarSlot: (
@@ -42,16 +48,20 @@ export function useRexComposerControls() {
         groupedOptions={groupedOptions}
         loading={loading}
         selectedModelOption={selectedModelOption}
-        onSelectModel={(option) => setSelectedModelKey(option.key)}
+        onSelectModel={(option) => selectModelKey(option.key)}
+        autoOption={modelPickerAutoOption}
       />
     ),
   }), [
     agents,
     defaultSupportsVision,
+    effectiveModelOption,
     groupedOptions,
     loading,
+    modelPickerAutoOption,
+    selectModelKey,
+    selectedModelAuto,
     selectedModelOption,
     selectedPromptModel,
-    setSelectedModelKey,
   ]);
 }
diff --git a/webui/src/features/session-chat/display.test.ts b/webui/src/features/session-chat/display.test.ts
index 2b97867eb..acdf87457 100644
--- a/webui/src/features/session-chat/display.test.ts
+++ b/webui/src/features/session-chat/display.test.ts
@@ -1,6 +1,10 @@
 import { describe, expect, it } from 'vitest';
 
-import { buildInstructionDisplayText, parseInstructionDisplayText } from './display';
+import {
+  buildInstructionDisplayText,
+  parseInstructionDisplayText,
+  stripTaskMetadata,
+} from './display';
 
 describe('session chat display metadata', () => {
   it('round-trips instruction display labels', () => {
@@ -13,4 +17,28 @@ describe('session chat display metadata', () => {
   it('ignores regular display text', () => {
     expect(parseInstructionDisplayText('普通消息')).toBeNull();
   });
+
+  it('removes task metadata while preserving the visible result', () => {
+    const text = [
+      '查询完成。',
+      '',
+      '<task_metadata>',
+      'session_id: ses-child',
+      '</task_metadata>',
+      '',
+      '消息发送失败。',
+    ].join('\n');
+
+    const visibleText = stripTaskMetadata(text);
+
+    expect(visibleText).toContain('查询完成。');
+    expect(visibleText).toContain('消息发送失败。');
+    expect(visibleText).not.toContain('task_metadata');
+    expect(visibleText).not.toContain('ses-child');
+  });
+
+  it('hides a trailing task metadata block while it is still streaming', () => {
+    expect(stripTaskMetadata('查询完成。\n\n<task_metadata>\nsession_id: ses-child'))
+      .toBe('查询完成。\n\n');
+  });
 });
diff --git a/webui/src/features/session-chat/display.ts b/webui/src/features/session-chat/display.ts
index 3e1961110..fa294127a 100644
--- a/webui/src/features/session-chat/display.ts
+++ b/webui/src/features/session-chat/display.ts
@@ -3,6 +3,8 @@ export interface PromptDisplayOptions {
 }
 
 const INSTRUCTION_DISPLAY_PREFIX = '@@flocks-instruction:';
+const TASK_METADATA_BLOCK_PATTERN = /<task_metadata\b[^>]*>[\s\S]*?<\/task_metadata\s*>/gi;
+const TRAILING_TASK_METADATA_PATTERN = /<task_metadata\b[^>]*>[\s\S]*$/i;
 
 export function buildInstructionDisplayText(label: string): string {
   return `${INSTRUCTION_DISPLAY_PREFIX}${label}`;
@@ -14,12 +16,20 @@ export function parseInstructionDisplayText(text: string): string | null {
     : null;
 }
 
+export function stripTaskMetadata(text: string): string {
+  return text
+    .replace(TASK_METADATA_BLOCK_PATTERN, '')
+    .replace(TRAILING_TASK_METADATA_PATTERN, '');
+}
+
 /** Display-related options grouped to reduce prop surface. */
 export interface SessionChatDisplay {
   /** Compact mode for panels/dialogs (default: true). Set false for full-page. */
   compact?: boolean;
   /** Let embedded chats use the full available message width. */
   fullWidth?: boolean;
+  /** Use the quieter, centered canvas treatment for the session-management page. */
+  pageCanvas?: boolean;
   /** Show copy action on assistant messages */
   showActions?: boolean;
   /** Show timestamp below each message */
diff --git a/webui/src/features/session-chat/index.ts b/webui/src/features/session-chat/index.ts
index c64a7155d..3965b836b 100644
--- a/webui/src/features/session-chat/index.ts
+++ b/webui/src/features/session-chat/index.ts
@@ -1,6 +1,7 @@
 export {
   buildInstructionDisplayText,
   parseInstructionDisplayText,
+  stripTaskMetadata,
   type PromptDisplayOptions,
   type SessionChatDisplay,
 } from './display';
diff --git a/webui/src/features/session-chat/sseActions.test.ts b/webui/src/features/session-chat/sseActions.test.ts
index ca2874ac8..2213a5c45 100644
--- a/webui/src/features/session-chat/sseActions.test.ts
+++ b/webui/src/features/session-chat/sseActions.test.ts
@@ -58,6 +58,17 @@ describe('resolveSessionChatSSEAction', () => {
       info: { id: 'msg-1' },
     });
 
+    expect(resolveSessionChatSSEAction({
+      type: 'message.removed',
+      properties: {
+        sessionID: 'sess-1',
+        messageID: 'msg-1',
+      },
+    }, 'sess-1')).toEqual({
+      kind: 'message-removed',
+      messageID: 'msg-1',
+    });
+
     expect(resolveSessionChatSSEAction({
       type: 'message.part.updated',
       properties: {
@@ -75,6 +86,14 @@ describe('resolveSessionChatSSEAction', () => {
       part: { id: 'part-1' },
       delta: 'hello',
     });
+
+    expect(resolveSessionChatSSEAction({
+      type: 'message.removed',
+      properties: {
+        sessionID: 'other-session',
+        messageID: 'msg-1',
+      },
+    }, 'sess-1')).toEqual({ kind: 'ignore' });
   });
 
   it('resolves questions only when request and call ids are present', () => {
diff --git a/webui/src/features/session-chat/sseActions.ts b/webui/src/features/session-chat/sseActions.ts
index 58d5f3172..70fc3338d 100644
--- a/webui/src/features/session-chat/sseActions.ts
+++ b/webui/src/features/session-chat/sseActions.ts
@@ -24,6 +24,7 @@ export type SessionChatSSEAction =
   | { kind: 'session-cleared' }
   | { kind: 'session-status'; statusType?: string; message?: string }
   | { kind: 'message-updated'; info: SessionMessageInfo }
+  | { kind: 'message-removed'; messageID: string }
   | { kind: 'message-part-updated'; part: MessagePart; delta?: string }
   | { kind: 'question-asked'; callID: string; requestId: string; questions: unknown[] }
   | { kind: 'question-resolved'; requestId: string }
@@ -82,6 +83,13 @@ export function resolveSessionChatSSEAction(
     return { kind: 'message-updated', info: properties.info as SessionMessageInfo };
   }
 
+  if (type === 'message.removed') {
+    if (!isCurrentSession(properties, sessionId) || typeof properties.messageID !== 'string') {
+      return { kind: 'ignore' };
+    }
+    return { kind: 'message-removed', messageID: properties.messageID };
+  }
+
   if (type === 'message.part.updated') {
     if (!isRecord(properties.part) || properties.part.sessionID !== sessionId) return { kind: 'ignore' };
     return {
diff --git a/webui/src/features/session-chat/sseRouting.test.ts b/webui/src/features/session-chat/sseRouting.test.ts
index a95681412..4c899396e 100644
--- a/webui/src/features/session-chat/sseRouting.test.ts
+++ b/webui/src/features/session-chat/sseRouting.test.ts
@@ -27,6 +27,10 @@ describe('shouldForwardSSEEventToParent', () => {
       type: 'message.part.updated',
       properties: { part: { sessionID: 'session-1' } },
     }, 'session-1')).toBe(true);
+    expect(shouldForwardSSEEventToParent({
+      type: 'message.removed',
+      properties: { sessionID: 'session-1', messageID: 'message-1' },
+    }, 'session-1')).toBe(true);
     expect(shouldForwardSSEEventToParent({
       type: 'session.status',
       properties: { sessionID: 'session-1' },
diff --git a/webui/src/features/session-chat/useSessionPromptQueue.ts b/webui/src/features/session-chat/useSessionPromptQueue.ts
index 4ecf0af6f..2101ceb63 100644
--- a/webui/src/features/session-chat/useSessionPromptQueue.ts
+++ b/webui/src/features/session-chat/useSessionPromptQueue.ts
@@ -1,6 +1,7 @@
 import { useCallback, useLayoutEffect, useRef, useState } from 'react';
 
 import { sessionApi, type QueuedPrompt } from '@/api/session';
+import type { SessionExecutionMode } from '@/utils/sessionExecutionMode';
 
 export interface EnqueuePromptPayload {
   parts: Array<Record<string, unknown>>;
@@ -8,6 +9,7 @@ export interface EnqueuePromptPayload {
   model?: Record<string, unknown>;
   variant?: string;
   displayText?: string;
+  executionMode?: SessionExecutionMode;
 }
 
 export function useSessionPromptQueue(sessionId?: string | null) {
diff --git a/webui/src/hooks/useProviders.test.tsx b/webui/src/hooks/useProviders.test.tsx
index 9a493efb0..cba9d1e27 100644
--- a/webui/src/hooks/useProviders.test.tsx
+++ b/webui/src/hooks/useProviders.test.tsx
@@ -60,6 +60,29 @@ describe('useProviders', () => {
     ]);
   });
 
+  it('uses the backend credential state instead of treating every connected provider as configured', async () => {
+    listMock.mockResolvedValue({
+      data: {
+        all: [
+          { ...makeProvider('openai', { 'gpt-4o': {} }), configured: false },
+          { ...makeProvider('deepseek'), configured: true },
+        ],
+        connected: ['openai', 'deepseek'],
+      },
+    });
+
+    const { result } = renderHook(() => useProviders());
+
+    await waitFor(() => {
+      expect(result.current.loading).toBe(false);
+    });
+
+    expect(result.current.providers).toMatchObject([
+      { id: 'openai', configured: false, category: 'international' },
+      { id: 'deepseek', configured: true, category: 'connected' },
+    ]);
+  });
+
   it('shares provider list requests across concurrent hook instances', async () => {
     let resolveList: (value: { data: any }) => void = () => {};
     listMock.mockReturnValue(new Promise((resolve) => {
diff --git a/webui/src/hooks/useProviders.ts b/webui/src/hooks/useProviders.ts
index d671e5b4b..e702e0831 100644
--- a/webui/src/hooks/useProviders.ts
+++ b/webui/src/hooks/useProviders.ts
@@ -31,7 +31,9 @@ function normalizeProvidersPayload(data: unknown): ProvidersResourceData {
   const connectedSet = new Set(connectedIds);
   const providers: EnrichedProvider[] = rawProviders.map((provider) => {
     const modelCount = provider.models ? Object.keys(provider.models).length : 0;
-    const configured = connectedSet.has(provider.id);
+    const configured = typeof provider.configured === 'boolean'
+      ? provider.configured
+      : connectedSet.has(provider.id);
     const category: ProviderCategory = configured ? 'connected' : getProviderCategory(provider.id);
     return {
       ...provider,
diff --git a/webui/src/hooks/useSessionChat.test.ts b/webui/src/hooks/useSessionChat.test.ts
index 74a4698e9..5b961b8f3 100644
--- a/webui/src/hooks/useSessionChat.test.ts
+++ b/webui/src/hooks/useSessionChat.test.ts
@@ -13,8 +13,12 @@
 import { describe, expect, it, vi, beforeEach } from 'vitest';
 
 const mockPost = vi.fn();
+const mockPatch = vi.fn();
 vi.mock('@/api/client', () => ({
-  default: { post: (...args: unknown[]) => mockPost(...args) },
+  default: {
+    post: (...args: unknown[]) => mockPost(...args),
+    patch: (...args: unknown[]) => mockPatch(...args),
+  },
 }));
 
 import { renderHook, act } from '@testing-library/react';
@@ -25,6 +29,7 @@ const SESSION_ID = 'sess-abc';
 
 beforeEach(() => {
   vi.clearAllMocks();
+  mockPatch.mockResolvedValue({ data: {} });
   // /api/session creates a new session
   mockPost.mockImplementation((url: string) => {
     if (url === '/api/session') return Promise.resolve({ data: { id: SESSION_ID } });
@@ -135,7 +140,119 @@ describe('useSessionChat.createAndSend — image forwarding', () => {
     expect(mockPost.mock.calls.some(([url]) => url === '/api/session')).toBe(false);
     expect(mockPost).toHaveBeenCalledWith(
       '/api/session/existing-session/prompt_async',
-      { parts: [{ type: 'text', text: 'continue' }] },
+      {
+        executionMode: 'build',
+        parts: [{ type: 'text', text: 'continue' }],
+      },
+    );
+  });
+});
+
+describe('useSessionChat — Auto session creation', () => {
+  it('forwards the hook modelAuto option when creating a session', async () => {
+    const { result } = renderHook(() =>
+      useSessionChat({ title: 'Auto chat', category: 'entity-config', modelAuto: true }),
+    );
+
+    await act(async () => {
+      await result.current.create();
+    });
+
+    expect(mockPost).toHaveBeenCalledWith('/api/session', {
+      title: 'Auto chat',
+      category: 'entity-config',
+      model_auto: true,
+    });
+  });
+
+  it('lets create explicitly override the hook modelAuto option', async () => {
+    const { result } = renderHook(() =>
+      useSessionChat({ title: 'Auto chat', modelAuto: true }),
+    );
+
+    await act(async () => {
+      await result.current.create({ modelAuto: false });
+    });
+
+    expect(mockPost).toHaveBeenCalledWith('/api/session', {
+      title: 'Auto chat',
+      model_auto: false,
+    });
+  });
+
+  it('lets createAndSend enable Auto without adding a synthetic prompt model', async () => {
+    const { result } = renderHook(() =>
+      useSessionChat({ title: 'Auto chat' }),
+    );
+
+    await act(async () => {
+      await result.current.createAndSend({
+        text: 'hello',
+        model: null,
+        modelAuto: true,
+      });
+    });
+
+    expect(mockPost).toHaveBeenCalledWith('/api/session', {
+      title: 'Auto chat',
+      model_auto: true,
+    });
+    expect(mockPost).toHaveBeenCalledWith(
+      `/api/session/${SESSION_ID}/prompt_async`,
+      {
+        executionMode: 'build',
+        parts: [{ type: 'text', text: 'hello' }],
+      },
+    );
+  });
+
+  it('enables Auto before sending through a resumed session', async () => {
+    const { result } = renderHook(() =>
+      useSessionChat({
+        title: 'Auto chat',
+        initialSessionId: 'existing-session',
+        modelAuto: true,
+      }),
+    );
+
+    await act(async () => {
+      await result.current.createAndSend({ text: 'continue' });
+    });
+
+    expect(mockPatch).toHaveBeenCalledWith('/api/session/existing-session', {
+      model_auto: true,
+      model_pinned: false,
+    });
+    expect(mockPatch.mock.invocationCallOrder[0]).toBeLessThan(
+      mockPost.mock.invocationCallOrder[0],
+    );
+    expect(mockPost).toHaveBeenCalledWith(
+      '/api/session/existing-session/prompt_async',
+      {
+        executionMode: 'build',
+        parts: [{ type: 'text', text: 'continue' }],
+      },
+    );
+  });
+
+  it('sends the selected execution mode with the turn', async () => {
+    const { result } = renderHook(() =>
+      useSessionChat({ title: 'Plan chat', initialSessionId: 'existing-session' }),
+    );
+
+    await act(async () => {
+      await result.current.createAndSend({
+        text: 'inspect the implementation',
+        executionMode: 'plan',
+      });
+    });
+
+    expect(mockPost).toHaveBeenCalledWith(
+      '/api/session/existing-session/prompt_async',
+      {
+        executionMode: 'plan',
+        parts: [{ type: 'text', text: 'inspect the implementation' }],
+      },
     );
   });
 });
diff --git a/webui/src/hooks/useSessionChat.ts b/webui/src/hooks/useSessionChat.ts
index f936d6cc2..295dbcf88 100644
--- a/webui/src/hooks/useSessionChat.ts
+++ b/webui/src/hooks/useSessionChat.ts
@@ -1,10 +1,13 @@
 import { useState, useCallback, useRef, useEffect } from 'react';
 import client from '@/api/client';
 import { buildPromptParts, type ImagePartData } from '@/utils/imageUpload';
+import type { SessionExecutionMode } from '@/utils/sessionExecutionMode';
 
 export interface UseSessionChatOptions {
   title: string;
   category?: string;
+  /** Enable runtime model failover when creating a new session. */
+  modelAuto?: boolean;
   /** Context injected via noReply (not visible as user message) */
   contextMessage?: string;
   /** Mock welcome message from assistant */
@@ -21,12 +24,16 @@ export interface CreateAndSendOptions {
   imageParts?: ImagePartData[];
   agent?: string;
   model?: { providerID: string; modelID: string } | null;
+  /** Override Auto mode for the new session created by this send. */
+  modelAuto?: boolean;
   displayText?: string;
+  executionMode?: SessionExecutionMode;
 }
 
 export function useSessionChat({
   title,
   category,
+  modelAuto,
   contextMessage,
   welcomeMessage,
   initialSessionId = null,
@@ -38,8 +45,8 @@ export function useSessionChat({
 
   const sessionIdRef = useRef<string | null>(initialSessionId);
   const createPromiseRef = useRef<Promise<string> | null>(null);
-  const optionsRef = useRef({ title, category, contextMessage, welcomeMessage });
-  optionsRef.current = { title, category, contextMessage, welcomeMessage };
+  const optionsRef = useRef({ title, category, modelAuto, contextMessage, welcomeMessage });
+  optionsRef.current = { title, category, modelAuto, contextMessage, welcomeMessage };
 
   const create = useCallback(
     async (overrides?: Partial<UseSessionChatOptions>): Promise<string> => {
@@ -53,8 +60,9 @@ export function useSessionChat({
       const opts = { ...optionsRef.current, ...overrides };
 
       const doCreate = async (): Promise<string> => {
-        const payload: Record<string, string> = { title: opts.title };
+        const payload: Record<string, string | boolean> = { title: opts.title };
         if (opts.category) payload.category = opts.category;
+        if (typeof opts.modelAuto === 'boolean') payload.model_auto = opts.modelAuto;
 
         const res = await client.post('/api/session', payload);
         const sid: string = res.data.id;
@@ -120,15 +128,30 @@ export function useSessionChat({
       imageParts,
       agent,
       model,
+      modelAuto: createModelAuto,
       displayText,
+      executionMode = 'build',
     }: CreateAndSendOptions): Promise<string> => {
-      const sid = await create();
+      const resumedExistingSession = Boolean(sessionIdRef.current);
+      const effectiveModelAuto = typeof createModelAuto === 'boolean'
+        ? createModelAuto
+        : optionsRef.current.modelAuto;
+      const sid = await create(
+        typeof createModelAuto === 'boolean' ? { modelAuto: createModelAuto } : undefined,
+      );
+      if (resumedExistingSession && effectiveModelAuto) {
+        await client.patch(`/api/session/${sid}`, {
+          model_auto: true,
+          model_pinned: false,
+        });
+      }
       const payload: Record<string, unknown> = {
         parts: buildPromptParts(text, imageParts),
       };
       if (agent) payload.agent = agent;
       if (model) payload.model = model;
       if (displayText) payload.displayText = displayText;
+      payload.executionMode = executionMode;
       client.post(`/api/session/${sid}/prompt_async`, payload).catch(() => {});
       return sid;
     },
diff --git a/webui/src/hooks/useSessions.test.ts b/webui/src/hooks/useSessions.test.ts
index 499489c6a..7ca37967d 100644
--- a/webui/src/hooks/useSessions.test.ts
+++ b/webui/src/hooks/useSessions.test.ts
@@ -260,6 +260,22 @@ describe('updateMessagePart scheduling', () => {
     expect((msg!.parts as any[])[0].text).toBe('hello world');
   });
 
+  it('removes a failed assistant placeholder by message id', async () => {
+    const { result } = renderHook(() => useSessionMessages('sess-1'));
+    await act(async () => {});
+
+    await act(async () => {
+      result.current.addMessage(makeMsg({ id: 'keep' }));
+      result.current.addMessage(makeMsg({ id: 'remove-me' }));
+    });
+
+    await act(async () => {
+      result.current.removeMessage('remove-me');
+    });
+
+    expect(result.current.messages.map(message => message.id)).toEqual(['keep']);
+  });
+
   it('applies every known part update without waiting for an animation frame', async () => {
     const { result } = renderHook(() => useSessionMessages('sess-1'));
     await act(async () => {});
@@ -389,225 +405,6 @@ describe('updateMessagePart scheduling', () => {
     expect(result.current.loading).toBe(false);
   });
 
-  it('ignores an older-page response after switching sessions', async () => {
-    let resolveOlderSessionA: (value: unknown) => void = () => {};
-    vi.mocked(client.get).mockImplementation((url: string, config?: any) => {
-      if (url.includes('sess-a') && config?.params?.before) {
-        return new Promise((resolve) => {
-          resolveOlderSessionA = resolve;
-        }) as any;
-      }
-      if (url.includes('sess-a')) {
-        return Promise.resolve({
-          data: {
-            items: [{
-              info: {
-                id: 'msg-a-new',
-                sessionID: 'sess-a',
-                role: 'assistant',
-                time: { created: 200 },
-              },
-              parts: [],
-            }],
-            hasMore: true,
-            nextBefore: 'msg-a-new',
-          },
-        }) as any;
-      }
-      return Promise.resolve({
-        data: [{
-          info: {
-            id: 'msg-b',
-            sessionID: 'sess-b',
-            role: 'assistant',
-            time: { created: 300 },
-          },
-          parts: [],
-        }],
-      }) as any;
-    });
-
-    const { result, rerender } = renderHook(
-      ({ id }: { id: string }) => useSessionMessages(id),
-      { initialProps: { id: 'sess-a' } },
-    );
-    await act(async () => {});
-
-    let olderRequest: Promise<void> = Promise.resolve();
-    act(() => {
-      olderRequest = result.current.loadOlder();
-    });
-
-    rerender({ id: 'sess-b' });
-    await act(async () => {});
-    expect(result.current.messages.map((message) => message.id)).toEqual(['msg-b']);
-
-    await act(async () => {
-      resolveOlderSessionA({
-        data: {
-          items: [{
-            info: {
-              id: 'msg-a-old',
-              sessionID: 'sess-a',
-              role: 'user',
-              time: { created: 100 },
-            },
-            parts: [],
-          }],
-          hasMore: false,
-          nextBefore: null,
-        },
-      });
-      await olderRequest;
-    });
-
-    expect(result.current.messages.map((message) => message.id)).toEqual(['msg-b']);
-    expect(result.current.loadingOlder).toBe(false);
-  });
-
-  it('clears older-page loading when a first-page refetch invalidates it', async () => {
-    let resolveOlderPage: (value: unknown) => void = () => {};
-    let firstPageRequestCount = 0;
-    vi.mocked(client.get).mockImplementation((_url: string, config?: any) => {
-      if (config?.params?.before) {
-        return new Promise((resolve) => {
-          resolveOlderPage = resolve;
-        }) as any;
-      }
-
-      firstPageRequestCount += 1;
-      const messageId = firstPageRequestCount === 1 ? 'msg-current' : 'msg-refreshed';
-      return Promise.resolve({
-        data: {
-          items: [{
-            info: {
-              id: messageId,
-              sessionID: 'sess-1',
-              role: 'assistant',
-              time: { created: 200 + firstPageRequestCount },
-            },
-            parts: [],
-          }],
-          hasMore: true,
-          nextBefore: messageId,
-        },
-      }) as any;
-    });
-
-    const { result } = renderHook(() => useSessionMessages('sess-1'));
-    await act(async () => {});
-
-    let olderRequest: Promise<void> = Promise.resolve();
-    act(() => {
-      olderRequest = result.current.loadOlder();
-    });
-    expect(result.current.loadingOlder).toBe(true);
-
-    await act(async () => {
-      await result.current.refetch();
-    });
-    expect(result.current.loadingOlder).toBe(false);
-    expect(result.current.messages.map((message) => message.id)).toEqual([
-      'msg-current',
-      'msg-refreshed',
-    ]);
-
-    await act(async () => {
-      resolveOlderPage({
-        data: {
-          items: [{
-            info: {
-              id: 'msg-stale-older',
-              sessionID: 'sess-1',
-              role: 'user',
-              time: { created: 100 },
-            },
-            parts: [],
-          }],
-          hasMore: false,
-          nextBefore: null,
-        },
-      });
-      await olderRequest;
-    });
-
-    expect(result.current.loadingOlder).toBe(false);
-    expect(result.current.messages.map((message) => message.id)).toEqual([
-      'msg-current',
-      'msg-refreshed',
-    ]);
-  });
-
-  it('does not start an older-page request while a first-page refetch is pending', async () => {
-    let resolveRefetch: (value: unknown) => void = () => {};
-    let firstPageRequests = 0;
-    let olderPageRequests = 0;
-    vi.mocked(client.get).mockImplementation((_url: string, config?: any) => {
-      if (config?.params?.before) {
-        olderPageRequests += 1;
-        return Promise.resolve({ data: { items: [], hasMore: false, nextBefore: null } }) as any;
-      }
-
-      firstPageRequests += 1;
-      if (firstPageRequests === 1) {
-        return Promise.resolve({
-          data: {
-            items: [{
-              info: {
-                id: 'msg-current',
-                sessionID: 'sess-1',
-                role: 'assistant',
-                time: { created: 200 },
-              },
-              parts: [],
-            }],
-            hasMore: true,
-            nextBefore: 'old-cursor',
-          },
-        }) as any;
-      }
-
-      return new Promise((resolve) => {
-        resolveRefetch = resolve;
-      }) as any;
-    });
-
-    const { result } = renderHook(() => useSessionMessages('sess-1'));
-    await act(async () => {});
-
-    let refetchRequest: Promise<void> = Promise.resolve();
-    act(() => {
-      refetchRequest = result.current.refetch();
-    });
-    await act(async () => {
-      await result.current.loadOlder();
-    });
-
-    expect(olderPageRequests).toBe(0);
-
-    await act(async () => {
-      resolveRefetch({
-        data: {
-          items: [{
-            info: {
-              id: 'msg-refreshed',
-              sessionID: 'sess-1',
-              role: 'assistant',
-              time: { created: 300 },
-            },
-            parts: [],
-          }],
-          hasMore: true,
-          nextBefore: 'new-cursor',
-        },
-      });
-      await refetchRequest;
-    });
-
-    expect(result.current.loading).toBe(false);
-    expect(olderPageRequests).toBe(0);
-  });
-
   it('replaceMessageText updates the targeted text part by partId', async () => {
     const { result } = renderHook(() => useSessionMessages('sess-1'));
     await act(async () => {});
@@ -714,6 +511,88 @@ describe('updateMessagePart scheduling', () => {
     expect((result.current.messages[1].parts as any[])[0].text).toBe('new reply');
   });
 
+  it('keeps one user message when refetch and delayed SSE reconcile its optimistic ID', async () => {
+    const optimisticId = 'msg_000000000001abcdefghijklmn';
+    vi.mocked(client.get).mockResolvedValueOnce({ data: [] });
+    const { result } = renderHook(() => useSessionMessages('sess-1'));
+    await act(async () => {});
+
+    await act(async () => {
+      result.current.addMessage(makeMsg({
+        id: optimisticId,
+        role: 'user',
+        timestamp: 100,
+        parts: [{
+          id: `temp-${optimisticId}-text`,
+          type: 'text',
+          text: '测试 question 工具',
+        }] as Message['parts'],
+      }));
+    });
+
+    vi.mocked(client.get).mockResolvedValueOnce({
+      data: {
+        items: [
+          {
+            info: {
+              id: optimisticId,
+              sessionID: 'sess-1',
+              role: 'user',
+              time: { created: 200 },
+            },
+            parts: [{
+              id: 'user-real-text',
+              messageID: optimisticId,
+              sessionID: 'sess-1',
+              type: 'text',
+              text: '测试 question 工具',
+            }],
+          },
+          {
+            info: {
+              id: 'assistant-question',
+              sessionID: 'sess-1',
+              role: 'assistant',
+              parentID: optimisticId,
+              time: { created: 201 },
+            },
+            parts: [{
+              id: 'question-part',
+              messageID: 'assistant-question',
+              sessionID: 'sess-1',
+              type: 'tool',
+              tool: 'question',
+              callID: 'question-call',
+              state: { status: 'running' },
+            }],
+          },
+        ],
+        hasMore: false,
+        nextBefore: null,
+      },
+    });
+
+    await act(async () => {
+      await result.current.refetch();
+      result.current.updateMessage({
+        id: optimisticId,
+        sessionID: 'sess-1',
+        role: 'user',
+        time: { created: 200 },
+      });
+    });
+
+    expect(result.current.messages.filter((message) => message.role === 'user'))
+      .toHaveLength(1);
+    expect(result.current.messages.map((message) => message.id)).toEqual([
+      optimisticId,
+      'assistant-question',
+    ]);
+    expect(result.current.messages[0].parts.map((part) => part.id)).toEqual([
+      'user-real-text',
+    ]);
+  });
+
   it('truncateAfterMessage keeps the target by default', async () => {
     const { result } = renderHook(() => useSessionMessages('sess-1'));
     await act(async () => {});
@@ -823,103 +702,280 @@ describe('updateMessagePart scheduling', () => {
     expect((msg?.parts as any[])[0].state.status).toBe('completed');
   });
 
-  it('fetches the first message page and prepends older messages', async () => {
+  it('fetches the complete message history in one request', async () => {
+    vi.mocked(client.get).mockResolvedValueOnce({
+      data: [
+        {
+          info: {
+            id: 'msg-old',
+            sessionID: 'sess-1',
+            role: 'user',
+            time: { created: 100 },
+          },
+          parts: [{ id: 'part-old', type: 'text', text: 'old' }],
+        },
+        {
+          info: {
+            id: 'msg-new',
+            sessionID: 'sess-1',
+            role: 'assistant',
+            time: { created: 200 },
+          },
+          parts: [],
+        },
+      ],
+    } as any);
+
+    const { result } = renderHook(() => useSessionMessages('sess-1'));
+    await act(async () => {});
+
+    expect(result.current.messages.map((msg) => msg.id)).toEqual(['msg-old', 'msg-new']);
+    expect(client.get).toHaveBeenCalledOnce();
+    expect(client.get).toHaveBeenCalledWith('/api/session/sess-1/message', {
+      params: { include_archived: true },
+    });
+  });
+
+  it('removes messages that are absent from a complete refetch', async () => {
     vi.mocked(client.get)
       .mockResolvedValueOnce({
-        data: {
-          items: [{
-            info: {
-              id: 'msg-new',
-              sessionID: 'sess-1',
-              role: 'assistant',
-              time: { created: 200 },
-            },
-            parts: [],
-          }],
-          hasMore: true,
-          nextBefore: 'msg-new',
-        },
+        data: [{
+          info: {
+            id: 'msg-before-clear',
+            sessionID: 'sess-1',
+            role: 'assistant',
+            time: { created: 100 },
+          },
+          parts: [{ id: 'part-before-clear', type: 'text', text: 'old result' }],
+        }],
       } as any)
+      .mockResolvedValueOnce({ data: [] });
+
+    const { result } = renderHook(() => useSessionMessages('sess-1'));
+    await act(async () => {});
+    expect(result.current.messages.map((message) => message.id)).toEqual(['msg-before-clear']);
+
+    await act(async () => {
+      await result.current.refetch();
+    });
+
+    expect(result.current.messages).toEqual([]);
+  });
+
+  it('keeps the current history when a complete refetch fails', async () => {
+    vi.mocked(client.get)
       .mockResolvedValueOnce({
-        data: {
-          items: [{
-            info: {
-              id: 'msg-old',
-              sessionID: 'sess-1',
-              role: 'user',
-              time: { created: 100 },
-              model: { providerID: 'openai', modelID: 'gpt-4o' },
-            },
-            parts: [{ id: 'part-old', type: 'text', text: 'old' }],
-          }],
-          hasMore: false,
-          nextBefore: null,
-        },
-      } as any);
+        data: [{
+          info: {
+            id: 'msg-before-error',
+            sessionID: 'sess-1',
+            role: 'assistant',
+            time: { created: 100 },
+          },
+          parts: [{ id: 'part-before-error', type: 'text', text: 'keep me' }],
+        }],
+      } as any)
+      .mockRejectedValueOnce(new Error('storage unavailable'));
 
     const { result } = renderHook(() => useSessionMessages('sess-1'));
     await act(async () => {});
 
-    expect(result.current.messages.map((msg) => msg.id)).toEqual(['msg-new']);
-    expect(result.current.hasMore).toBe(true);
-    expect(client.get).toHaveBeenCalledWith('/api/session/sess-1/message', {
-      params: { page: true, limit: 50, include_archived: true },
+    await act(async () => {
+      await result.current.refetch();
     });
 
+    expect(result.current.messages.map((message) => message.id)).toEqual(['msg-before-error']);
+    expect(result.current.error).toBe('storage unavailable');
+  });
+
+  it('keeps an optimistic local message until the complete history confirms it', async () => {
+    vi.mocked(client.get)
+      .mockResolvedValueOnce({ data: [] })
+      .mockResolvedValueOnce({ data: [] });
+
+    const { result } = renderHook(() => useSessionMessages('sess-1'));
+    await act(async () => {});
+
     await act(async () => {
-      await result.current.loadOlder();
+      result.current.addMessage(makeMsg({
+        id: 'msg-optimistic',
+        role: 'user',
+        parts: [{ id: 'temp-msg-optimistic-text', type: 'text', text: 'hello' } as any],
+      }));
+      await result.current.refetch();
     });
 
-    expect(result.current.messages.map((msg) => msg.id)).toEqual(['msg-old', 'msg-new']);
-    expect(result.current.hasMore).toBe(false);
-    expect(client.get).toHaveBeenLastCalledWith('/api/session/sess-1/message', {
-      params: { page: true, limit: 50, before: 'msg-new', include_archived: true },
+    expect(result.current.messages.map((message) => message.id)).toEqual(['msg-optimistic']);
+  });
+
+  it('keeps an optimistic message when an older refetch resolves after SSE confirmation', async () => {
+    const olderRefetch = deferred<any>();
+    vi.mocked(client.get)
+      .mockResolvedValueOnce({ data: [] })
+      .mockReturnValueOnce(olderRefetch.promise);
+
+    const { result } = renderHook(() => useSessionMessages('sess-1'));
+    await act(async () => {});
+
+    await act(async () => {
+      result.current.addMessage(makeMsg({
+        id: 'msg-racing',
+        role: 'user',
+        parts: [{ id: 'temp-msg-racing-text', type: 'text', text: 'hello' } as any],
+      }));
+    });
+
+    let refetchPromise: Promise<void> = Promise.resolve();
+    act(() => {
+      refetchPromise = result.current.refetch();
+    });
+    expect(client.get).toHaveBeenCalledTimes(2);
+
+    act(() => {
+      result.current.updateMessage({
+        id: 'msg-racing',
+        sessionID: 'sess-1',
+        role: 'user',
+        time: { created: 200 },
+      });
     });
+
+    await act(async () => {
+      olderRefetch.resolve({ data: [] });
+      await refetchPromise;
+    });
+
+    expect(result.current.messages.map((message) => message.id)).toEqual(['msg-racing']);
   });
 
-  it('starts only one older-page request when loadOlder is called twice before rerender', async () => {
-    let resolveOlderPage: (value: unknown) => void = () => {};
-    let olderPageRequests = 0;
-    vi.mocked(client.get).mockImplementation((_url: string, config?: any) => {
-      if (config?.params?.before) {
-        olderPageRequests += 1;
-        return new Promise((resolve) => {
-          resolveOlderPage = resolve;
-        }) as any;
-      }
+  it('keeps newer SSE message state when an older complete refetch resolves', async () => {
+    const olderRefetch = deferred<any>();
+    const staleMessage = {
+      info: {
+        id: 'msg-existing',
+        sessionID: 'sess-1',
+        role: 'assistant',
+        time: { created: 100 },
+      },
+      parts: [{
+        id: 'part-existing',
+        messageID: 'msg-existing',
+        sessionID: 'sess-1',
+        type: 'text',
+        text: 'stale text',
+      }],
+    };
+    vi.mocked(client.get)
+      .mockResolvedValueOnce({ data: [staleMessage] })
+      .mockReturnValueOnce(olderRefetch.promise);
 
-      return Promise.resolve({
-        data: {
-          items: [],
-          hasMore: true,
-          nextBefore: 'older-cursor',
-        },
-      }) as any;
+    const { result } = renderHook(() => useSessionMessages('sess-1'));
+    await act(async () => {});
+
+    let refetchPromise: Promise<void> = Promise.resolve();
+    act(() => {
+      refetchPromise = result.current.refetch();
+    });
+
+    act(() => {
+      result.current.updateMessagePart({
+        id: 'part-existing',
+        messageID: 'msg-existing',
+        sessionID: 'sess-1',
+        type: 'text',
+        text: 'new streamed text',
+      }, 'new streamed text');
+      result.current.updateMessage({
+        id: 'msg-new-assistant',
+        sessionID: 'sess-1',
+        role: 'assistant',
+        time: { created: 200 },
+      });
+    });
+
+    await act(async () => {
+      olderRefetch.resolve({ data: [staleMessage] });
+      await refetchPromise;
     });
 
+    expect(result.current.messages.map((message) => message.id)).toEqual([
+      'msg-existing',
+      'msg-new-assistant',
+    ]);
+    expect((result.current.messages[0].parts as any[])[0].text).toBe('new streamed text');
+  });
+
+  it('clears optimistic messages before applying an empty session history', async () => {
+    vi.mocked(client.get)
+      .mockResolvedValueOnce({ data: [] })
+      .mockResolvedValueOnce({ data: [] });
+
     const { result } = renderHook(() => useSessionMessages('sess-1'));
     await act(async () => {});
 
-    let firstRequest: Promise<void> = Promise.resolve();
-    let secondRequest: Promise<void> = Promise.resolve();
+    await act(async () => {
+      result.current.addMessage(makeMsg({
+        id: 'msg-pending-clear',
+        role: 'user',
+        parts: [{ id: 'temp-msg-pending-clear-text', type: 'text', text: 'hello' } as any],
+      }));
+      result.current.clearMessages();
+      await result.current.refetch();
+    });
+
+    expect(result.current.messages).toEqual([]);
+  });
+
+  it('keeps messages empty when a request started before clear resolves late', async () => {
+    const pendingRequest = deferred<any>();
+    vi.mocked(client.get).mockReturnValueOnce(pendingRequest.promise);
+
+    const { result } = renderHook(() => useSessionMessages('sess-1'));
+    expect(result.current.loading).toBe(true);
+
     act(() => {
-      firstRequest = result.current.loadOlder();
-      secondRequest = result.current.loadOlder();
+      result.current.clearMessages();
     });
 
-    expect(olderPageRequests).toBe(1);
+    expect(result.current.messages).toEqual([]);
+    expect(result.current.error).toBeNull();
+    expect(result.current.loading).toBe(false);
 
     await act(async () => {
-      resolveOlderPage({
-        data: {
-          items: [],
-          hasMore: false,
-          nextBefore: null,
-        },
+      pendingRequest.resolve({
+        data: [{
+          info: {
+            id: 'msg-deleted-before-response',
+            sessionID: 'sess-1',
+            role: 'assistant',
+            time: { created: 100 },
+          },
+          parts: [{ id: 'part-deleted-before-response', type: 'text', text: 'stale' }],
+        }],
       });
-      await Promise.all([firstRequest, secondRequest]);
+      await pendingRequest.promise;
+    });
+
+    expect(result.current.messages).toEqual([]);
+    expect(result.current.error).toBeNull();
+    expect(result.current.loading).toBe(false);
+  });
+
+  it('resets a previous loading error when messages are cleared', async () => {
+    vi.mocked(client.get).mockRejectedValueOnce(new Error('storage unavailable'));
+
+    const { result } = renderHook(() => useSessionMessages('sess-1'));
+    await act(async () => {});
+    expect(result.current.error).toBe('storage unavailable');
+
+    act(() => {
+      result.current.clearMessages();
     });
-    expect(result.current.loadingOlder).toBe(false);
+
+    expect(result.current.messages).toEqual([]);
+    expect(result.current.error).toBeNull();
+    expect(result.current.loading).toBe(false);
+    expect(client.get).toHaveBeenCalledOnce();
   });
 
 });
@@ -1025,16 +1081,19 @@ describe('useSessions list loading', () => {
     await act(async () => {});
 
     expect(result.current.loading).toBe(false);
+    expect(result.current.refreshing).toBe(false);
     expect(result.current.sessions.map((session) => session.id)).toEqual(['session-1']);
 
     rerender({ search: 'triage' });
 
     expect(result.current.loading).toBe(false);
+    expect(result.current.refreshing).toBe(true);
     expect(result.current.sessions.map((session) => session.id)).toEqual(['session-1']);
 
     await act(async () => {
       resolveSearch([]);
     });
+    expect(result.current.refreshing).toBe(false);
   });
 
   it('loads and tracks pages independently for each project', async () => {
diff --git a/webui/src/hooks/useSessions.ts b/webui/src/hooks/useSessions.ts
index 519aa382a..7f958485f 100644
--- a/webui/src/hooks/useSessions.ts
+++ b/webui/src/hooks/useSessions.ts
@@ -6,7 +6,6 @@ import type { Session, Message } from '@/types';
 const VISIBLE_CATEGORIES = new Set(['user', 'workflow', 'entity-config']);
 const ABORTED_TOOL_ERROR = 'Tool execution was interrupted';
 const SESSION_LIST_PAGE_SIZE = 100;
-const MESSAGE_PAGE_SIZE = 50;
 
 function finalizeStoppedMessageParts(parts: Message['parts'], stoppedAt = Date.now()): Message['parts'] {
   return parts.map((part) => {
@@ -92,52 +91,51 @@ function mergeFetchedMessages(prev: Message[], fetched: Message[]): Message[] {
   }));
 }
 
-function mergeLatestFetchedMessages(prev: Message[], fetched: Message[]): Message[] {
-  if (prev.length === 0) return normalizeMessageOrder(fetched);
+function mergeCompleteFetchedMessages(
+  prev: Message[],
+  fetched: Message[],
+  provisionalMessageIds: Set<string>,
+  protectedMessageIds: Set<string>,
+): Message[] {
   const fetchedIds = new Set(fetched.map((message) => message.id));
+  const previousById = new Map(prev.map((message) => [message.id, message]));
   const mergedFetched = mergeFetchedMessages(prev, fetched);
-  const firstFetchedTimestamp = mergedFetched[0]?.timestamp ?? Number.POSITIVE_INFINITY;
-  const retainedOlder = prev.filter(
-    (message) => !fetchedIds.has(message.id) && message.timestamp <= firstFetchedTimestamp,
-  );
-  const retainedNewer = prev.filter(
-    (message) => !fetchedIds.has(message.id) && message.timestamp > firstFetchedTimestamp,
+  const protectedFetched = mergedFetched.map((message) => (
+    protectedMessageIds.has(message.id)
+      ? previousById.get(message.id) ?? message
+      : message
+  ));
+  const retainedLocal = prev.filter(
+    (message) => (
+      !fetchedIds.has(message.id)
+      && (
+        provisionalMessageIds.has(message.id)
+        || protectedMessageIds.has(message.id)
+      )
+    ),
   );
-  return normalizeMessageOrder([...retainedOlder, ...mergedFetched, ...retainedNewer]);
-}
-
-function prependOlderMessages(prev: Message[], older: Message[]): Message[] {
-  const existingIds = new Set(prev.map((message) => message.id));
-  return normalizeMessageOrder([...older.filter((message) => !existingIds.has(message.id)), ...prev]);
+  return normalizeMessageOrder([...protectedFetched, ...retainedLocal]);
 }
 
-function transformMessageResponse(data: any): {
-  messages: Message[];
-  hasMore: boolean;
-  nextBefore: string | null;
-} {
+function transformMessageResponse(data: any): Message[] {
   const items = Array.isArray(data) ? data : (data?.items ?? []);
-  return {
-    messages: items.map((msg: any) => ({
-      id: msg.info.id,
-      sessionID: msg.info.sessionID,
-      role: msg.info.role,
-      parts: msg.parts || [],
-      parentID: msg.info.parentID,
-      agent: msg.info.agent,
-      model: msg.info.model,
-      modelID: msg.info.modelID,
-      providerID: msg.info.providerID,
-      cost: msg.info.cost,
-      tokens: msg.info.tokens,
-      timestamp: msg.info.time?.created || Date.now(),
-      finish: msg.info.finish || null,
-      error: msg.info.error || null,
-      compacted: msg.info.compacted || null,
-    })),
-    hasMore: Array.isArray(data) ? false : Boolean(data?.hasMore),
-    nextBefore: Array.isArray(data) ? null : (data?.nextBefore ?? null),
-  };
+  return items.map((msg: any) => ({
+    id: msg.info.id,
+    sessionID: msg.info.sessionID,
+    role: msg.info.role,
+    parts: msg.parts || [],
+    parentID: msg.info.parentID,
+    agent: msg.info.agent,
+    model: msg.info.model,
+    modelID: msg.info.modelID,
+    providerID: msg.info.providerID,
+    cost: msg.info.cost,
+    tokens: msg.info.tokens,
+    timestamp: msg.info.time?.created || Date.now(),
+    finish: msg.info.finish || null,
+    error: msg.info.error || null,
+    compacted: msg.info.compacted || null,
+  }));
 }
 
 function markMeasure(name: string, startMark: string) {
@@ -229,6 +227,7 @@ function sessionEffectiveProjectId(session: Session): string {
 export function useSessions(search = '', options?: UseSessionsOptions) {
   const [sessions, setSessions] = useState<Session[]>([]);
   const [loading, setLoading] = useState(true);
+  const [refreshing, setRefreshing] = useState(true);
   const [loadingMore, setLoadingMore] = useState(false);
   const [hasMore, setHasMore] = useState(false);
   const [hasMoreByProject, setHasMoreByProject] = useState<Record<string, boolean>>({});
@@ -292,6 +291,7 @@ export function useSessions(search = '', options?: UseSessionsOptions) {
         setLoading(true);
       }
       if (!append) {
+        setRefreshing(true);
         activeAppendProjectsRef.current.clear();
         setLoadingMore(false);
         setLoadingMoreProjectIds(new Set());
@@ -390,6 +390,7 @@ export function useSessions(search = '', options?: UseSessionsOptions) {
         }
       } else if (isCurrentRequest()) {
         setLoading(false);
+        setRefreshing(false);
         initializedRef.current = true;
       }
     }
@@ -437,6 +438,7 @@ export function useSessions(search = '', options?: UseSessionsOptions) {
   return {
     sessions,
     loading,
+    refreshing,
     error,
     refetch: fetchSessions,
     updateSessionTitle,
@@ -454,122 +456,81 @@ export function useSessions(search = '', options?: UseSessionsOptions) {
 export function useSessionMessages(sessionId?: string) {
   const [messages, setMessages] = useState<Message[]>([]);
   const [loading, setLoading] = useState(false);
-  const [loadingOlder, setLoadingOlder] = useState(false);
-  const [hasMore, setHasMore] = useState(false);
-  const [nextBefore, setNextBefore] = useState<string | null>(null);
   const [error, setError] = useState<string | null>(null);
   const activeSessionIdRef = useRef(sessionId);
-  const firstPageRequestIdRef = useRef(0);
-  const firstPageInFlightRequestIdRef = useRef<number | null>(null);
-  const olderPageRequestIdRef = useRef(0);
-  const olderPageInFlightRequestIdRef = useRef<number | null>(null);
+  const requestIdRef = useRef(0);
+  const provisionalMessageIdsRef = useRef<Set<string>>(new Set());
+  const messageMutationVersionsRef = useRef<Map<string, number>>(new Map());
+  const mutationVersionRef = useRef(0);
+
+  const markMessageMutation = (messageId: string) => {
+    mutationVersionRef.current += 1;
+    messageMutationVersionsRef.current.set(messageId, mutationVersionRef.current);
+  };
 
   const fetchMessages = useCallback(async () => {
     if (!sessionId || activeSessionIdRef.current !== sessionId) return;
 
     const requestSessionId = sessionId;
-    const requestId = ++firstPageRequestIdRef.current;
-    firstPageInFlightRequestIdRef.current = requestId;
-    // A fresh first page invalidates pagination based on an older snapshot.
-    olderPageRequestIdRef.current += 1;
-    olderPageInFlightRequestIdRef.current = null;
-    setLoadingOlder(false);
+    const requestId = ++requestIdRef.current;
+    const requestMutationVersion = mutationVersionRef.current;
     const isCurrentRequest = () => (
       activeSessionIdRef.current === requestSessionId
-      && firstPageRequestIdRef.current === requestId
+      && requestIdRef.current === requestId
     );
     
     try {
       setLoading(true);
       setError(null);
-      const startMark = 'session:messages:first-page-start';
+      const startMark = 'session:messages:start';
       if (typeof performance !== 'undefined') performance.mark(startMark);
       const response = await client.get(`/api/session/${sessionId}/message`, {
-        params: { page: true, limit: MESSAGE_PAGE_SIZE, include_archived: true },
+        params: { include_archived: true },
       });
-      markMeasure('session:messages:first-page', startMark);
+      markMeasure('session:messages', startMark);
       if (!isCurrentRequest()) return;
-      const { messages: messagesData, hasMore, nextBefore } = transformMessageResponse(response.data);
-      setMessages(prev => mergeLatestFetchedMessages(prev, messagesData));
-      setHasMore(hasMore);
-      setNextBefore(nextBefore);
+      const messagesData = transformMessageResponse(response.data);
+      const protectedMessageIds = new Set(
+        Array.from(messageMutationVersionsRef.current.entries())
+          .filter(([, version]) => version > requestMutationVersion)
+          .map(([messageId]) => messageId),
+      );
+      messagesData.forEach((message) => {
+        if (!protectedMessageIds.has(message.id)) {
+          provisionalMessageIdsRef.current.delete(message.id);
+        }
+      });
+      messageMutationVersionsRef.current.forEach((version, messageId) => {
+        if (version <= requestMutationVersion) {
+          messageMutationVersionsRef.current.delete(messageId);
+        }
+      });
+      setMessages(prev => mergeCompleteFetchedMessages(
+        prev,
+        messagesData,
+        provisionalMessageIdsRef.current,
+        protectedMessageIds,
+      ));
     } catch (err: any) {
       if (isCurrentRequest()) {
         setError(err.message || 'Failed to fetch messages');
       }
     } finally {
-      if (firstPageInFlightRequestIdRef.current === requestId) {
-        firstPageInFlightRequestIdRef.current = null;
-      }
       if (isCurrentRequest()) {
         setLoading(false);
       }
     }
   }, [sessionId]);
 
-  const loadOlder = useCallback(async () => {
-    if (
-      !sessionId
-      || activeSessionIdRef.current !== sessionId
-      || firstPageInFlightRequestIdRef.current !== null
-      || olderPageInFlightRequestIdRef.current !== null
-      || !hasMore
-      || !nextBefore
-    ) return;
-
-    const requestSessionId = sessionId;
-    const requestId = ++olderPageRequestIdRef.current;
-    olderPageInFlightRequestIdRef.current = requestId;
-    const firstPageRequestId = firstPageRequestIdRef.current;
-    const isCurrentRequest = () => (
-      activeSessionIdRef.current === requestSessionId
-      && olderPageRequestIdRef.current === requestId
-      && firstPageRequestIdRef.current === firstPageRequestId
-    );
-
-    try {
-      setLoadingOlder(true);
-      setError(null);
-      const startMark = 'session:messages:older-page-start';
-      if (typeof performance !== 'undefined') performance.mark(startMark);
-      const response = await client.get(`/api/session/${sessionId}/message`, {
-        params: {
-          page: true,
-          limit: MESSAGE_PAGE_SIZE,
-          before: nextBefore,
-          include_archived: true,
-        },
-      });
-      markMeasure('session:messages:older-page', startMark);
-      if (!isCurrentRequest()) return;
-      const page = transformMessageResponse(response.data);
-      setMessages(prev => prependOlderMessages(prev, page.messages));
-      setHasMore(page.hasMore);
-      setNextBefore(page.nextBefore);
-    } catch (err: any) {
-      if (isCurrentRequest()) {
-        setError(err.message || 'Failed to fetch older messages');
-      }
-    } finally {
-      if (olderPageInFlightRequestIdRef.current === requestId) {
-        olderPageInFlightRequestIdRef.current = null;
-      }
-      if (isCurrentRequest()) {
-        setLoadingOlder(false);
-      }
-    }
-  }, [hasMore, nextBefore, sessionId]);
-
   // Reset state synchronously before paint when session changes
   // to prevent flash of welcome screen (useEffect runs AFTER paint)
   useLayoutEffect(() => {
     activeSessionIdRef.current = sessionId;
+    provisionalMessageIdsRef.current.clear();
+    messageMutationVersionsRef.current.clear();
+    mutationVersionRef.current = 0;
     setMessages([]);
     setError(null);
-    setHasMore(false);
-    setNextBefore(null);
-    olderPageInFlightRequestIdRef.current = null;
-    setLoadingOlder(false);
     if (sessionId) {
       setLoading(true);
     } else {
@@ -584,15 +545,15 @@ export function useSessionMessages(sessionId?: string) {
   return {
     messages,
     loading,
-    loadingOlder,
-    hasMore,
     error,
     refetch: fetchMessages,
-    loadOlder,
     addMessage: (message: Message) => {
+      provisionalMessageIdsRef.current.add(message.id);
+      markMessageMutation(message.id);
       setMessages(prev => [...prev, message]);
     },
     updateMessage: (messageInfo: any) => {
+      markMessageMutation(messageInfo.id);
       setMessages(prev => {
         const existingIndex = prev.findIndex(m => m.id === messageInfo.id);
         if (existingIndex >= 0) {
@@ -626,6 +587,9 @@ export function useSessionMessages(sessionId?: string) {
           );
           if (tempIndex >= 0) {
             const updated = [...prev];
+            provisionalMessageIdsRef.current.delete(updated[tempIndex].id);
+            messageMutationVersionsRef.current.delete(updated[tempIndex].id);
+            provisionalMessageIdsRef.current.add(messageInfo.id);
             const nextUser = {
               id: messageInfo.id,
               sessionID: messageInfo.sessionID,
@@ -645,6 +609,7 @@ export function useSessionMessages(sessionId?: string) {
         }
 
         // Add new message
+        provisionalMessageIdsRef.current.add(messageInfo.id);
         const nextMessage = {
           id: messageInfo.id,
           sessionID: messageInfo.sessionID,
@@ -675,9 +640,30 @@ export function useSessionMessages(sessionId?: string) {
      * frame-level typing cadence and Markdown parse budget.
      */
     updateMessagePart: (partInfo: any, delta?: string) => {
-      setMessages(prev => applyMessagePartUpdate(prev, partInfo, delta));
+      markMessageMutation(partInfo.messageID);
+      setMessages(prev => {
+        if (!prev.some((message) => message.id === partInfo.messageID)) {
+          provisionalMessageIdsRef.current.add(partInfo.messageID);
+        }
+        return applyMessagePartUpdate(prev, partInfo, delta);
+      });
+    },
+    removeMessage: (messageId: string) => {
+      provisionalMessageIdsRef.current.delete(messageId);
+      messageMutationVersionsRef.current.delete(messageId);
+      setMessages(prev => prev.filter((message) => message.id !== messageId));
+    },
+    clearMessages: () => {
+      requestIdRef.current += 1;
+      provisionalMessageIdsRef.current.clear();
+      messageMutationVersionsRef.current.clear();
+      mutationVersionRef.current = 0;
+      setMessages([]);
+      setError(null);
+      setLoading(false);
     },
     replaceMessageText: (messageId: string, partId: string, text: string) => {
+      markMessageMutation(messageId);
       setMessages(prev => prev.map((message) => {
         if (message.id !== messageId) return message;
 
@@ -698,6 +684,7 @@ export function useSessionMessages(sessionId?: string) {
       }));
     },
     markMessageStopped: (messageId: string) => {
+      markMessageMutation(messageId);
       setMessages(prev => prev.map((message) => {
         if (message.id !== messageId) return message;
         if (message.finish === 'stop') return message;
@@ -713,7 +700,12 @@ export function useSessionMessages(sessionId?: string) {
       setMessages(prev => {
         const targetIndex = prev.findIndex((message) => message.id === messageId);
         if (targetIndex < 0) return prev;
-        return prev.slice(0, options?.includeTarget ? targetIndex : targetIndex + 1);
+        const keepCount = options?.includeTarget ? targetIndex : targetIndex + 1;
+        prev.slice(keepCount).forEach((message) => {
+          provisionalMessageIdsRef.current.delete(message.id);
+          messageMutationVersionsRef.current.delete(message.id);
+        });
+        return prev.slice(0, keepCount);
       });
     },
   };
diff --git a/webui/src/i18n.test.ts b/webui/src/i18n.test.ts
index ed475f047..c7e9a9dd0 100644
--- a/webui/src/i18n.test.ts
+++ b/webui/src/i18n.test.ts
@@ -1,5 +1,11 @@
 import { describe, expect, it } from 'vitest';
 import i18n, { preloadI18nNamespaces } from './i18n';
+import enHome from './locales/en-US/home.json';
+import enNav from './locales/en-US/nav.json';
+import enSession from './locales/en-US/session.json';
+import zhHome from './locales/zh-CN/home.json';
+import zhNav from './locales/zh-CN/nav.json';
+import zhSession from './locales/zh-CN/session.json';
 
 describe('i18n lazy namespaces', () => {
   it('loads route namespaces on demand instead of bundling them up front', async () => {
@@ -14,4 +20,32 @@ describe('i18n lazy namespaces', () => {
     expect(i18n.hasResourceBundle('en-US', 'workflow')).toBe(true);
     expect(Object.keys(i18n.getResourceBundle('en-US', 'workflow'))).not.toHaveLength(0);
   });
+
+  it('uses workbench and task terminology for session management', () => {
+    expect(zhNav.sessions).toBe('工作台');
+    expect(zhNav.workspace).toBe('文件目录');
+    expect(zhHome.quickActions.sessions.title).toBe('工作台');
+    expect(zhSession.managementTitle).toBe('工作台');
+    expect(zhSession.sessionCount).toBe('{{count}} 个任务');
+    expect(zhSession.collapseLoaded).toBe('收起');
+    expect(zhSession.newSession).toBe('新建任务');
+    expect(zhSession.createSession).toBe('新建任务');
+    expect(zhSession.createSessionInProject).toBe('在项目 {{project}} 中新建任务');
+    expect(zhSession.projectDialog.newSessionAction).toBe('新建任务');
+    expect(zhSession.filterConversations).toBe('搜索任务');
+    expect(zhSession.noResults).toBe('没有匹配的任务');
+
+    expect(enNav.sessions).toBe('Workbench');
+    expect(enNav.workspace).toBe('File Directory');
+    expect(enHome.quickActions.sessions.title).toBe('Workbench');
+    expect(enSession.managementTitle).toBe('Workbench');
+    expect(enSession.sessionCount).toBe('{{count}} tasks');
+    expect(enSession.collapseLoaded).toBe('Show less');
+    expect(enSession.newSession).toBe('New Task');
+    expect(enSession.createSession).toBe('New Task');
+    expect(enSession.createSessionInProject).toBe('New task in project {{project}}');
+    expect(enSession.projectDialog.newSessionAction).toBe('New task');
+    expect(enSession.filterConversations).toBe('Search tasks');
+    expect(enSession.noResults).toBe('No matching tasks');
+  });
 });
diff --git a/webui/src/locales/en-US/common.json b/webui/src/locales/en-US/common.json
index 6fe7a67a5..94d7ab996 100644
--- a/webui/src/locales/en-US/common.json
+++ b/webui/src/locales/en-US/common.json
@@ -77,6 +77,7 @@
     "stopped": "Stopped"
   },
   "delegate": {
+    "delegatedTo": "Delegated to {{agent}}",
     "pending": "Waiting",
     "running": "Executing",
     "completed": "Completed",
@@ -88,6 +89,7 @@
     "resultSummary": "Result Summary",
     "viewFullChat": "View Full Conversation",
     "viewDialog": "View Dialog",
+    "viewExecution": "View Execution",
     "emptyChat": "Sub-agent conversation is empty",
     "subTask": "Sub-task"
   },
diff --git a/webui/src/locales/en-US/home.json b/webui/src/locales/en-US/home.json
index 8c9c9702f..43bd86af3 100644
--- a/webui/src/locales/en-US/home.json
+++ b/webui/src/locales/en-US/home.json
@@ -43,7 +43,7 @@
   },
   "quickActions": {
     "sessions": {
-      "title": "Sessions",
+      "title": "Workbench",
       "description": "Chat with AI to complete complex security tasks"
     },
     "workflows": {
diff --git a/webui/src/locales/en-US/model.json b/webui/src/locales/en-US/model.json
index fbc54f05e..367976e2a 100644
--- a/webui/src/locales/en-US/model.json
+++ b/webui/src/locales/en-US/model.json
@@ -18,6 +18,12 @@
     "noCost": "No cost",
     "toggleCurrency": "Click to switch between USD and CNY"
   },
+  "modelSelection": {
+    "info": "Model information",
+    "free": "Free",
+    "unavailable": "Not provided",
+    "closeDefault": "Close default model selector"
+  },
   "providerList": {
     "empty": "No providers configured",
     "emptyHint": "Click \"Add Provider\" to get started",
diff --git a/webui/src/locales/en-US/nav.json b/webui/src/locales/en-US/nav.json
index 72e713ef5..8c42318b3 100644
--- a/webui/src/locales/en-US/nav.json
+++ b/webui/src/locales/en-US/nav.json
@@ -3,9 +3,9 @@
   "flocksHome": "Home",
   "aiWorkbench": "AI Workbench",
   "sceneWorkspaces": "Scene Workspaces",
-  "sessions": "Sessions",
+  "sessions": "Workbench",
   "tasks": "Task Center",
-  "workspace": "Workspace",
+  "workspace": "File Directory",
   "agentHub": "Agent Studio",
   "agents": "Agents",
   "workflows": "Workflows",
@@ -29,6 +29,8 @@
   "settingsPreferences": "Preferences",
   "settingsPreferencesDescription": "Adjust workspace display, language, theme, and tool failure handling.",
   "settingsGroupPreferences": "Preferences",
+  "settingsGroupData": "Data Management",
+  "archivedData": "Archived Tasks",
   "settingsGroupSystem": "Account & System",
   "settingsGroupIntegrations": "Models & Channels",
   "displayName": "Workspace Name",
diff --git a/webui/src/locales/en-US/session.json b/webui/src/locales/en-US/session.json
index 2263f1f67..4b39163cf 100644
--- a/webui/src/locales/en-US/session.json
+++ b/webui/src/locales/en-US/session.json
@@ -1,13 +1,22 @@
 {
-  "newSession": "New Session",
+  "managementTitle": "Workbench",
+  "sessionCount": "{{count}} tasks",
+  "clearSearch": "Clear search",
+  "newSession": "New Task",
+  "loadMore": "Load more",
+  "collapseLoaded": "Show less",
   "noSessions": "No sessions yet",
   "selectSession": "Select a session",
   "clickToStart": "Click the list on the left to start a conversation",
-  "createSession": "New Session",
+  "createSession": "New Task",
   "createFailed": "Failed to create session",
   "deleteFailed": "Delete failed",
   "deleteAction": "Delete",
   "confirmDelete": "Are you sure you want to delete this session?",
+  "archiveAction": "Archive",
+  "confirmArchive": "Archive this session? You can restore it later from Archived Tasks in Settings.",
+  "archiveSuccess": "Archived. You can restore it from Archived Tasks in Settings.",
+  "archiveFailed": "Archive failed",
   "rename": "Rename",
   "renamePlaceholder": "Enter a new session name",
   "renameFailed": "Rename failed",
@@ -20,6 +29,13 @@
   "shareEnabled": "Session is shared to all local accounts as read-only",
   "shareDisabled": "Session sharing canceled",
   "shareUpdateFailed": "Failed to update sharing status",
+  "moveToProjectAction": "Move to Project",
+  "moveToProjectTitle": "Choose a Project",
+  "moveToProjectBack": "Back to task actions",
+  "moveToProjectEmpty": "No available projects",
+  "moveToProjectRunning": "A running task cannot be moved",
+  "moveToProjectSuccess": "Moved to project “{{project}}”",
+  "moveToProjectFailed": "Failed to move task",
   "moreActions": "More actions",
   "selectMode": "Batch select",
   "cancelSelect": "Cancel",
@@ -30,6 +46,41 @@
   "deleteSelected": "Delete ({{count}})",
   "confirmBatchDelete": "Are you sure you want to delete {{count}} selected session(s)?",
   "batchDeleteFailed": "Failed to delete {{count}} session(s)",
+  "archiveSelected": "Archive ({{count}})",
+  "confirmBatchArchive": "Archive {{count}} selected session(s)?",
+  "batchArchiveFailed": "Failed to archive {{count}} session(s)",
+  "archivedData": {
+    "title": "Archived Tasks",
+    "description": "Manage archived tasks. Restored tasks return to the workbench; permanently deleted data cannot be recovered.",
+    "searchPlaceholder": "Search archived tasks",
+    "refresh": "Refresh",
+    "selected": "{{count}} selected",
+    "restoreSelected": "Restore selected",
+    "deleteSelected": "Permanently delete selected",
+    "loading": "Loading archived tasks…",
+    "loadFailed": "Failed to load archived tasks",
+    "retry": "Retry",
+    "empty": "No archived tasks",
+    "emptyDescription": "Tasks archived from the workbench will appear here.",
+    "selectAll": "Select all manageable tasks",
+    "selectSession": "Select task {{title}}",
+    "session": "Task",
+    "owner": "Owner",
+    "systemOwner": "System task",
+    "project": "Original project",
+    "archivedAt": "Archived at",
+    "actions": "Actions",
+    "restore": "Restore",
+    "delete": "Delete",
+    "restoreSuccess": "Restored {{count}} session(s)",
+    "restoreFailed": "Failed to restore {{count}} session(s)",
+    "deleteSuccess": "Permanently deleted {{count}} session(s)",
+    "deleteFailed": "Failed to permanently delete {{count}} session(s)",
+    "loadMore": "Load more",
+    "deleteTitle": "Permanently delete archived tasks",
+    "deleteDescription": "This will permanently delete messages, attachments, and related state for {{count}} task(s). This action cannot be undone.",
+    "confirmDelete": "Permanently delete"
+  },
   "workflowSession": "Workflow Session",
   "configSession": "Config Session",
   "showHistory": "Show conversation history",
@@ -39,11 +90,31 @@
   "connectionFailed": "Connection failed",
   "notConnected": "Not connected",
   "loading": "Loading...",
+  "refreshingWorkbench": "Refreshing workbench",
+  "restoringTask": "Restoring task",
   "noAgents": "No agents available",
   "smartAssistant": "AI Smart Assistant",
+  "executionMode": {
+    "title": "Execution mode",
+    "hint": "Controls how Rex handles the next turn",
+    "options": {
+      "build": {
+        "label": "Build",
+        "description": "Use Rex's full tools to complete the task"
+      },
+      "plan": {
+        "label": "Plan",
+        "description": "Research and produce an implementation plan"
+      },
+      "goal": {
+        "label": "Goal",
+        "description": "Start a persistent goal, then return to Build"
+      }
+    }
+  },
   "agentPicker": {
     "title": "Choose Chat Agent",
-    "hint": "New chats start with Rex; type @Agent to route one message",
+    "hint": "Rex is the default; selections route the current message as a subagent reference",
     "filter": {
       "all": "All",
       "builtin": "Built-in",
@@ -59,6 +130,10 @@
     "title": "Choose Model",
     "hint": "Overrides the model used when sending this chat message",
     "empty": "No available models",
+    "auto": "Auto",
+    "autoHint": "If the primary model fails, randomly try up to one available model from the same provider, then one from another provider",
+    "autoUnavailable": "Configure an available primary model",
+    "autoUserSessionsOnly": "Auto is available only for WebUI chat sessions",
     "count": "{{count}}",
     "vision": "Vision",
     "free": "Free",
@@ -95,6 +170,8 @@
     "placeholderStreaming": "AI is processing, please wait...",
     "placeholderNodeRef": "Describe the change intent for node {{nodeId}}...",
     "emptyText": "No execution records",
+    "loadFailed": "Failed to load messages",
+    "retry": "Retry",
     "suggestions": "Example questions",
     "stopTitle": "Stop execution",
     "contextUsageTitle": "Context used: {{used}} / {{total}} ({{percent}}%)",
@@ -139,7 +216,8 @@
     "thinking": "Thinking...",
     "streaming": "Streaming...",
     "process": {
-      "title": "Process ({{count}})",
+      "title": "View {{count}} steps",
+      "deepThinking": "Deep thinking",
       "reasoningCount": "{{count}} reasoning",
       "toolCount": "{{count}} tool calls",
       "textCount": "{{count}} intermediate replies"
@@ -220,12 +298,114 @@
       "imageNotSupported": "The current model does not support image analysis. Please select a vision-capable model in the model settings.",
       "imageNotSupportedBanner": "Image analysis is not supported by the current model"
     },
+    "addMenu": {
+      "title": "Add",
+      "files": "Files and images",
+      "agent": "Agent",
+      "workflows": "Workflows",
+      "skills": "Skills",
+      "selectWorkflow": "Select workflow",
+      "selectSkill": "Select skill",
+      "noWorkflows": "No workflows available",
+      "noSkills": "No skills available"
+    },
+    "references": {
+      "selected": "Selected resources",
+      "subagent": "Subagent",
+      "skill": "Skill",
+      "workflow": "Workflow",
+      "remove": "Remove {{type}} {{name}}"
+    },
     "tool": {
       "pending": "Pending",
       "running": "Running",
       "completed": "Completed",
       "error": "Failed",
+      "loadSkill": "Load skill",
+      "actions": {
+        "readFile": "Read file",
+        "writeFile": "Write file",
+        "editFile": "Edit file",
+        "findFiles": "Find files",
+        "parseDocument": "Parse document",
+        "executeCommand": "Run command",
+        "searchContent": "Search content",
+        "analyzeCode": "Analyze code",
+        "findDefinition": "Find definition",
+        "findReferences": "Find references",
+        "inspectSymbol": "Inspect symbol",
+        "listSymbols": "List code symbols",
+        "searchSymbols": "Search code symbols",
+        "findImplementations": "Find implementations",
+        "analyzeCallHierarchy": "Analyze call hierarchy",
+        "findCallers": "Find callers",
+        "findCallees": "Find callees",
+        "searchWeb": "Search the web",
+        "fetchWeb": "Read webpage",
+        "delegateTask": "Delegate to subagent",
+        "createScheduledTask": "Create scheduled task",
+        "listScheduledTasks": "List scheduled tasks",
+        "viewScheduledTask": "View task status",
+        "updateScheduledTask": "Update scheduled task",
+        "deleteScheduledTask": "Delete scheduled task",
+        "rerunScheduledTask": "Rerun task",
+        "runWorkflow": "Run workflow",
+        "runWorkflowNode": "Run workflow node",
+        "askQuestion": "Ask user",
+        "searchMemory": "Search memory",
+        "readMemory": "Read memory",
+        "writeMemory": "Save memory",
+        "listProviders": "List model providers",
+        "addProvider": "Add model provider",
+        "addModel": "Add model",
+        "runSlashCommand": "Run system command",
+        "searchTools": "Find tools",
+        "manageSkills": "Manage skills",
+        "searchSkills": "Search skills",
+        "installSkill": "Install skill",
+        "checkSkills": "Check skill status",
+        "installSkillDependencies": "Install skill dependencies",
+        "removeSkill": "Remove skill",
+        "manageMcp": "Manage MCP servers",
+        "listMcpServers": "List MCP servers",
+        "addMcpServer": "Add MCP server",
+        "removeMcpServer": "Remove MCP server",
+        "connectMcpServer": "Connect MCP server",
+        "disconnectMcpServer": "Disconnect MCP server",
+        "manageTasks": "Manage tasks",
+        "listTasks": "List tasks",
+        "viewTask": "View task",
+        "createTask": "Create task",
+        "updateTask": "Update task",
+        "deleteTask": "Delete task",
+        "archiveTask": "Archive task",
+        "restoreTask": "Restore task",
+        "manageWorkflowConfig": "Manage workflow config",
+        "readWorkflowConfig": "Read workflow config",
+        "viewWorkflowConfigStatus": "View workflow config status",
+        "syncWorkflowConfig": "Sync workflow config",
+        "compareWorkflowConfig": "Compare workflow config",
+        "updateWorkflowConfig": "Update workflow config",
+        "manageDevices": "Manage devices",
+        "listDevices": "List devices",
+        "listDeviceTemplates": "List device templates",
+        "createDevice": "Add device",
+        "updateDevice": "Update device",
+        "testDeviceConnection": "Test device connection",
+        "runRemoteCommand": "Run remote command",
+        "runRemoteScript": "Run remote script",
+        "sendMessage": "Send message",
+        "useWeCom": "Use WeCom",
+        "listWeComTools": "List WeCom tools",
+        "getTime": "Get current time"
+      },
+      "todoUpdated": "Todos updated",
       "inputParams": "Input Parameters",
+      "progress": {
+        "writingFile": "Writing file…",
+        "editingFile": "Editing file…",
+        "working": "Working…"
+      },
       "outputResult": "Output Result",
       "todoStages": "Todo stages",
       "todoStatus": {
@@ -265,11 +445,16 @@
   "projectsSection": "Projects",
   "toggleProjects": "Expand or collapse projects",
   "tasksSection": "Tasks",
+  "projectPicker": {
+    "title": "Project context",
+    "none": "Work without a project",
+    "create": "New project"
+  },
   "toggleTasks": "Expand or collapse tasks",
   "selectTasks": "Select tasks",
   "createTaskSession": "New task",
   "selectProject": "Select project {{project}}",
-  "createSessionInProject": "New session in project {{project}}",
+  "createSessionInProject": "New task in project {{project}}",
   "projectActions": "Project actions",
   "defaultProject": "Default project",
   "projectPathUnavailable": "Project directory unavailable",
@@ -281,14 +466,14 @@
     "createTitle": "New Project",
     "renameTitle": "Rename Project",
     "renameAction": "Rename",
-    "newSessionAction": "New session",
+    "newSessionAction": "New task",
     "copyPathAction": "Copy path",
     "deleteAction": "Delete",
     "deleteTitle": "Delete Project",
-    "deleteDescription": "Delete project “{{project}}”? Its conversations will be deleted, but the project folder and files will be kept.",
+    "deleteDescription": "Remove project “{{project}}”? Its tasks will be archived and its folder and files kept. Restoring any archived task will also restore the project.",
     "confirmDelete": "Delete",
     "defaultDeleteDisabled": "The default project cannot be deleted",
-    "deleteSuccess": "Project deleted",
+    "deleteSuccess": "Project removed and its tasks archived",
     "deleteFailed": "Failed to delete project",
     "nameLabel": "Project name",
     "namePlaceholder": "Enter project name",
@@ -305,8 +490,8 @@
     "copyPathFailed": "Failed to copy project path",
     "saveFailed": "Failed to save project"
   },
-  "filterConversations": "Filter conversations...",
-  "noResults": "No conversations found",
+  "filterConversations": "Search tasks",
+  "noResults": "No matching tasks",
   "showMore": "Show {{count}} more",
   "showLess": "Show less"
 }
diff --git a/webui/src/locales/zh-CN/common.json b/webui/src/locales/zh-CN/common.json
index a4b7c5c44..421f74016 100644
--- a/webui/src/locales/zh-CN/common.json
+++ b/webui/src/locales/zh-CN/common.json
@@ -77,6 +77,7 @@
     "stopped": "已停止"
   },
   "delegate": {
+    "delegatedTo": "委派给 {{agent}}",
     "pending": "等待中",
     "running": "执行中",
     "completed": "已完成",
@@ -88,6 +89,7 @@
     "resultSummary": "结果摘要",
     "viewFullChat": "查看完整对话",
     "viewDialog": "查看对话框",
+    "viewExecution": "查看执行过程",
     "emptyChat": "子 Agent 对话为空",
     "subTask": "子任务"
   },
diff --git a/webui/src/locales/zh-CN/home.json b/webui/src/locales/zh-CN/home.json
index 23ff02a2c..317b45d42 100644
--- a/webui/src/locales/zh-CN/home.json
+++ b/webui/src/locales/zh-CN/home.json
@@ -43,7 +43,7 @@
   },
   "quickActions": {
     "sessions": {
-      "title": "会话管理",
+      "title": "工作台",
       "description": "与 AI 对话，完成复杂的安全任务"
     },
     "workflows": {
diff --git a/webui/src/locales/zh-CN/model.json b/webui/src/locales/zh-CN/model.json
index 4761c8ff1..aef45a275 100644
--- a/webui/src/locales/zh-CN/model.json
+++ b/webui/src/locales/zh-CN/model.json
@@ -18,6 +18,12 @@
     "noCost": "暂无费用",
     "toggleCurrency": "点击切换人民币 / 美元"
   },
+  "modelSelection": {
+    "info": "模型信息",
+    "free": "免费",
+    "unavailable": "暂无",
+    "closeDefault": "关闭默认模型选择"
+  },
   "providerList": {
     "empty": "尚未添加模型供应商",
     "emptyHint": "点击右上角「添加模型供应商」开始配置",
diff --git a/webui/src/locales/zh-CN/nav.json b/webui/src/locales/zh-CN/nav.json
index 752a117b6..59590281e 100644
--- a/webui/src/locales/zh-CN/nav.json
+++ b/webui/src/locales/zh-CN/nav.json
@@ -3,9 +3,9 @@
   "flocksHome": "首页",
   "aiWorkbench": "AI 工作台",
   "sceneWorkspaces": "场景工作区",
-  "sessions": "会话管理",
+  "sessions": "工作台",
   "tasks": "任务中心",
-  "workspace": "工作空间",
+  "workspace": "文件目录",
   "agentHub": "智能体工作室",
   "agents": "智能体",
   "workflows": "工作流",
@@ -29,6 +29,8 @@
   "settingsPreferences": "偏好设置",
   "settingsPreferencesDescription": "调整当前工作台的显示、语言、主题和工具失败处理。",
   "settingsGroupPreferences": "偏好",
+  "settingsGroupData": "数据管理",
+  "archivedData": "已归档任务",
   "settingsGroupSystem": "账号与系统",
   "settingsGroupIntegrations": "模型与通道",
   "displayName": "工作台名称",
diff --git a/webui/src/locales/zh-CN/session.json b/webui/src/locales/zh-CN/session.json
index 6e2ff9a60..329fd3640 100644
--- a/webui/src/locales/zh-CN/session.json
+++ b/webui/src/locales/zh-CN/session.json
@@ -1,13 +1,22 @@
 {
-  "newSession": "新建会话",
+  "managementTitle": "工作台",
+  "sessionCount": "{{count}} 个任务",
+  "clearSearch": "清空搜索",
+  "newSession": "新建任务",
+  "loadMore": "加载更多",
+  "collapseLoaded": "收起",
   "noSessions": "暂无会话",
   "selectSession": "选择一个会话",
   "clickToStart": "点击左侧列表开始对话",
-  "createSession": "新建会话",
+  "createSession": "新建任务",
   "createFailed": "创建会话失败",
   "deleteFailed": "删除失败",
   "deleteAction": "删除",
   "confirmDelete": "确定要删除这个会话吗？",
+  "archiveAction": "归档",
+  "confirmArchive": "确定归档这个会话吗？归档后可在设置的“已归档任务”中恢复。",
+  "archiveSuccess": "已归档，可在设置的“已归档任务”中恢复",
+  "archiveFailed": "归档失败",
   "rename": "重命名",
   "renamePlaceholder": "输入新的会话名称",
   "renameFailed": "重命名失败",
@@ -20,6 +29,13 @@
   "shareEnabled": "会话已共享给本地账号（只读）",
   "shareDisabled": "已取消会话共享",
   "shareUpdateFailed": "更新共享状态失败",
+  "moveToProjectAction": "移动到项目",
+  "moveToProjectTitle": "选择目标项目",
+  "moveToProjectBack": "返回任务操作",
+  "moveToProjectEmpty": "暂无可用项目",
+  "moveToProjectRunning": "任务运行时无法移动",
+  "moveToProjectSuccess": "已移动到项目“{{project}}”",
+  "moveToProjectFailed": "移动任务失败",
   "moreActions": "更多操作",
   "selectMode": "批量选择",
   "cancelSelect": "取消",
@@ -30,6 +46,41 @@
   "deleteSelected": "删除 ({{count}})",
   "confirmBatchDelete": "确定要删除选中的 {{count}} 个会话吗？",
   "batchDeleteFailed": "{{count}} 个会话删除失败",
+  "archiveSelected": "归档 ({{count}})",
+  "confirmBatchArchive": "确定归档选中的 {{count}} 个会话吗？",
+  "batchArchiveFailed": "{{count}} 个会话归档失败",
+  "archivedData": {
+    "title": "已归档任务",
+    "description": "管理已归档任务。恢复后任务将重新出现在工作台；永久删除后无法恢复。",
+    "searchPlaceholder": "搜索已归档任务",
+    "refresh": "刷新",
+    "selected": "已选择 {{count}} 项",
+    "restoreSelected": "恢复所选",
+    "deleteSelected": "永久删除所选",
+    "loading": "正在加载已归档任务…",
+    "loadFailed": "加载已归档任务失败",
+    "retry": "重试",
+    "empty": "暂无已归档任务",
+    "emptyDescription": "从工作台归档的任务会显示在这里。",
+    "selectAll": "选择全部可管理任务",
+    "selectSession": "选择任务 {{title}}",
+    "session": "任务",
+    "owner": "所属用户",
+    "systemOwner": "系统任务",
+    "project": "原项目",
+    "archivedAt": "归档时间",
+    "actions": "操作",
+    "restore": "恢复",
+    "delete": "删除",
+    "restoreSuccess": "已恢复 {{count}} 个会话",
+    "restoreFailed": "{{count}} 个会话恢复失败",
+    "deleteSuccess": "已永久删除 {{count}} 个会话",
+    "deleteFailed": "{{count}} 个会话永久删除失败",
+    "loadMore": "加载更多",
+    "deleteTitle": "永久删除已归档任务",
+    "deleteDescription": "将永久删除 {{count}} 个任务的消息、附件和相关状态。此操作无法恢复。",
+    "confirmDelete": "永久删除"
+  },
   "workflowSession": "工作流会话",
   "configSession": "配置会话",
   "showHistory": "显示对话历史",
@@ -39,11 +90,31 @@
   "connectionFailed": "连接失败",
   "notConnected": "未连接",
   "loading": "加载中...",
+  "refreshingWorkbench": "正在刷新工作台",
+  "restoringTask": "正在恢复任务",
   "noAgents": "暂无可用的Agent",
   "smartAssistant": "智能AI助手",
+  "executionMode": {
+    "title": "执行模式",
+    "hint": "控制 Rex 如何处理下一轮消息",
+    "options": {
+      "build": {
+        "label": "Build",
+        "description": "使用 Rex 的完整工具完成任务"
+      },
+      "plan": {
+        "label": "Plan",
+        "description": "调研需求并输出可实施计划"
+      },
+      "goal": {
+        "label": "Goal",
+        "description": "启动持久目标，然后返回 Build"
+      }
+    }
+  },
   "agentPicker": {
     "title": "选择对话 Agent",
-    "hint": "新会话首条默认由 Rex 处理；输入 @Agent 可路由单条消息",
+    "hint": "Rex 为默认智能体；选择后以子智能体标签路由当前消息",
     "filter": {
       "all": "全部",
       "builtin": "内置",
@@ -59,6 +130,10 @@
     "title": "选择模型",
     "hint": "作为本次对话发送时的模型覆盖",
     "empty": "暂无可用模型",
+    "auto": "Auto",
+    "autoHint": "主模型失败后，依次随机尝试同 Provider 和其他 Provider 的可用模型（各最多一个）",
+    "autoUnavailable": "请先配置可用的主模型",
+    "autoUserSessionsOnly": "Auto 仅适用于 WebUI 对话会话",
     "count": "{{count}} 个",
     "vision": "视觉",
     "free": "免费",
@@ -95,6 +170,8 @@
     "placeholderStreaming": "AI 正在处理中，请稍候...",
     "placeholderNodeRef": "描述对节点 {{nodeId}} 的修改意图...",
     "emptyText": "暂无执行记录",
+    "loadFailed": "消息加载失败",
+    "retry": "重试",
     "suggestions": "示例问题",
     "stopTitle": "停止执行",
     "contextUsageTitle": "上下文已用：{{used}} / {{total}}（{{percent}}%）",
@@ -139,7 +216,8 @@
     "thinking": "思考中...",
     "streaming": "继续输出中...",
     "process": {
-      "title": "过程（{{count}} 项）",
+      "title": "查看 {{count}} 个步骤",
+      "deepThinking": "深度思考",
       "reasoningCount": "{{count}} 段思考",
       "toolCount": "{{count}} 次工具调用",
       "textCount": "{{count}} 段中间回复"
@@ -220,12 +298,114 @@
       "imageNotSupported": "当前模型不支持图片分析，请在模型配置中选择支持视觉的模型",
       "imageNotSupportedBanner": "当前模型不支持图片分析"
     },
+    "addMenu": {
+      "title": "添加",
+      "files": "文件和图片",
+      "agent": "智能体",
+      "workflows": "工作流",
+      "skills": "技能",
+      "selectWorkflow": "选择工作流",
+      "selectSkill": "选择技能",
+      "noWorkflows": "暂无可用工作流",
+      "noSkills": "暂无可用技能"
+    },
+    "references": {
+      "selected": "已选择的资源",
+      "subagent": "子智能体",
+      "skill": "技能",
+      "workflow": "工作流",
+      "remove": "移除 {{type}} {{name}}"
+    },
     "tool": {
       "pending": "等待中",
       "running": "执行中",
       "completed": "已完成",
       "error": "失败",
+      "loadSkill": "加载技能",
+      "actions": {
+        "readFile": "读取文件",
+        "writeFile": "写入文件",
+        "editFile": "修改文件",
+        "findFiles": "查找文件",
+        "parseDocument": "解析文档",
+        "executeCommand": "执行命令",
+        "searchContent": "搜索内容",
+        "analyzeCode": "分析代码",
+        "findDefinition": "查找定义",
+        "findReferences": "查找引用",
+        "inspectSymbol": "查看符号信息",
+        "listSymbols": "列出代码符号",
+        "searchSymbols": "搜索代码符号",
+        "findImplementations": "查找实现",
+        "analyzeCallHierarchy": "分析调用关系",
+        "findCallers": "查找调用方",
+        "findCallees": "查找被调用项",
+        "searchWeb": "搜索网页",
+        "fetchWeb": "读取网页",
+        "delegateTask": "委派子智能体",
+        "createScheduledTask": "创建定时任务",
+        "listScheduledTasks": "查看定时任务",
+        "viewScheduledTask": "查看任务状态",
+        "updateScheduledTask": "更新定时任务",
+        "deleteScheduledTask": "删除定时任务",
+        "rerunScheduledTask": "重新运行任务",
+        "runWorkflow": "运行工作流",
+        "runWorkflowNode": "运行工作流节点",
+        "askQuestion": "向用户提问",
+        "searchMemory": "检索记忆",
+        "readMemory": "读取记忆",
+        "writeMemory": "保存记忆",
+        "listProviders": "查看模型服务",
+        "addProvider": "添加模型服务",
+        "addModel": "添加模型",
+        "runSlashCommand": "运行系统命令",
+        "searchTools": "查找工具",
+        "manageSkills": "管理技能",
+        "searchSkills": "搜索技能",
+        "installSkill": "安装技能",
+        "checkSkills": "检查技能状态",
+        "installSkillDependencies": "安装技能依赖",
+        "removeSkill": "移除技能",
+        "manageMcp": "管理 MCP 服务",
+        "listMcpServers": "查看 MCP 服务",
+        "addMcpServer": "添加 MCP 服务",
+        "removeMcpServer": "移除 MCP 服务",
+        "connectMcpServer": "连接 MCP 服务",
+        "disconnectMcpServer": "断开 MCP 服务",
+        "manageTasks": "管理任务",
+        "listTasks": "查看任务",
+        "viewTask": "查看任务详情",
+        "createTask": "创建任务",
+        "updateTask": "更新任务",
+        "deleteTask": "删除任务",
+        "archiveTask": "归档任务",
+        "restoreTask": "恢复任务",
+        "manageWorkflowConfig": "管理工作流配置",
+        "readWorkflowConfig": "读取工作流配置",
+        "viewWorkflowConfigStatus": "查看工作流配置状态",
+        "syncWorkflowConfig": "同步工作流配置",
+        "compareWorkflowConfig": "对比工作流配置",
+        "updateWorkflowConfig": "更新工作流配置",
+        "manageDevices": "管理设备",
+        "listDevices": "查看设备",
+        "listDeviceTemplates": "查看设备模板",
+        "createDevice": "添加设备",
+        "updateDevice": "更新设备",
+        "testDeviceConnection": "检测设备连接",
+        "runRemoteCommand": "执行远程命令",
+        "runRemoteScript": "运行远程脚本",
+        "sendMessage": "发送消息",
+        "useWeCom": "操作企业微信",
+        "listWeComTools": "查看企业微信工具",
+        "getTime": "获取当前时间"
+      },
+      "todoUpdated": "已更新待办",
       "inputParams": "输入参数",
+      "progress": {
+        "writingFile": "正在写入文件…",
+        "editingFile": "正在编辑文件…",
+        "working": "正在执行此操作…"
+      },
       "outputResult": "输出结果",
       "todoStages": "Todo 阶段",
       "todoStatus": {
@@ -265,11 +445,16 @@
   "projectsSection": "项目",
   "toggleProjects": "展开或收起项目",
   "tasksSection": "任务",
+  "projectPicker": {
+    "title": "项目上下文",
+    "none": "不在项目中工作",
+    "create": "新建项目"
+  },
   "toggleTasks": "展开或收起任务",
   "selectTasks": "选择任务",
   "createTaskSession": "新建任务",
   "selectProject": "选择项目 {{project}}",
-  "createSessionInProject": "在项目 {{project}} 中新建会话",
+  "createSessionInProject": "在项目 {{project}} 中新建任务",
   "projectActions": "项目操作",
   "defaultProject": "默认项目",
   "projectPathUnavailable": "项目目录不可用",
@@ -281,14 +466,14 @@
     "createTitle": "新建项目",
     "renameTitle": "重命名项目",
     "renameAction": "重命名",
-    "newSessionAction": "新建会话",
+    "newSessionAction": "新建任务",
     "copyPathAction": "复制路径",
     "deleteAction": "删除",
     "deleteTitle": "删除项目",
-    "deleteDescription": "确定删除项目“{{project}}”吗？项目中的会话将被删除，但项目文件夹和文件会保留。",
+    "deleteDescription": "确定移除项目“{{project}}”吗？项目中的任务将进入归档，项目文件夹和文件会保留；恢复任一归档任务时会同时恢复该项目。",
     "confirmDelete": "删除",
     "defaultDeleteDisabled": "默认项目不能删除",
-    "deleteSuccess": "项目已删除",
+    "deleteSuccess": "项目已移除，相关任务已归档",
     "deleteFailed": "删除项目失败",
     "nameLabel": "项目名称",
     "namePlaceholder": "输入项目名称",
@@ -305,8 +490,8 @@
     "copyPathFailed": "复制项目路径失败",
     "saveFailed": "保存项目失败"
   },
-  "filterConversations": "搜索会话...",
-  "noResults": "没有匹配的会话",
+  "filterConversations": "搜索任务",
+  "noResults": "没有匹配的任务",
   "showMore": "显示更多 {{count}} 条",
   "showLess": "收起"
 }
diff --git a/webui/src/pages/DeviceIntegration/index.tsx b/webui/src/pages/DeviceIntegration/index.tsx
index 121c388ec..061ac9646 100644
--- a/webui/src/pages/DeviceIntegration/index.tsx
+++ b/webui/src/pages/DeviceIntegration/index.tsx
@@ -578,8 +578,9 @@ function DeviceAddRexPanel({
       text: prompt,
       agent: rexComposerControls.rexAgentName,
       model: rexComposerControls.rexModel,
+      modelAuto: rexComposerControls.rexModelAuto,
     }).catch(() => {});
-  }, [createAndSend, rexComposerControls.rexAgentName, rexComposerControls.rexModel]);
+  }, [createAndSend, rexComposerControls.rexAgentName, rexComposerControls.rexModel, rexComposerControls.rexModelAuto]);
 
   const vendorGroups = useMemo(() => {
     const groups = new Map<string, { vendor: DeviceVendor; templates: DeviceTemplate[] }>();
@@ -719,6 +720,7 @@ function DeviceAddRexPanel({
           agentName={rexComposerControls.rexAgentName}
           mentionAgents={rexComposerControls.rexMentionAgents}
           model={rexComposerControls.rexModel}
+          modelAuto={rexComposerControls.rexModelAuto}
           supportsVision={rexComposerControls.rexSupportsVision}
           contextWindowTokens={rexComposerControls.rexContextWindowTokens}
           composerTextareaMinHeight={rexComposerControls.rexComposerTextareaMinHeight}
@@ -879,6 +881,7 @@ function DeviceAddRexPanel({
           imageParts,
           agent: agentOverride || rexComposerControls.rexAgentName,
           model: modelOverride === undefined ? rexComposerControls.rexModel : modelOverride,
+          modelAuto: rexComposerControls.rexModelAuto,
         }) : undefined}
       />
       </Suspense>
@@ -1985,6 +1988,7 @@ export default function DeviceIntegrationPage() {
   } = useSessionChat({
     title: t('wizard.rex.title'),
     category: 'entity-config',
+    modelAuto: rexComposerControls.rexModelAuto,
     contextMessage: rexContextMessage,
     welcomeMessage: t('wizard.rex.welcome'),
   });
@@ -2221,6 +2225,7 @@ export default function DeviceIntegrationPage() {
         ...buildDeviceTestGuidePrompt(createdDevice, panel.template),
         agent: rexComposerControls.rexAgentName,
         model: rexComposerControls.rexModel,
+        modelAuto: rexComposerControls.rexModelAuto,
       }).catch(() => {});
       pollRexTestStatus(createdDevice);
       await fetchData(true);
@@ -2258,12 +2263,13 @@ export default function DeviceIntegrationPage() {
       ...prompt,
       agent: rexComposerControls.rexAgentName,
       model: rexComposerControls.rexModel,
+      modelAuto: rexComposerControls.rexModelAuto,
     });
     if (input.action === 'test' && input.device) {
       pollRexTestStatus(input.device);
     }
     setPanel({ kind: 'wizard' });
-  }, [createAndSendRex, pollRexTestStatus, rexComposerControls.rexAgentName, rexComposerControls.rexModel]);
+  }, [createAndSendRex, pollRexTestStatus, rexComposerControls.rexAgentName, rexComposerControls.rexModel, rexComposerControls.rexModelAuto]);
 
   // ──────────────────────────────────────────────────────────────────────────
   // Group to use when adding a new device (follows sidebar selection).
diff --git a/webui/src/pages/Model/index.test.tsx b/webui/src/pages/Model/index.test.tsx
index cbd42e1d7..dbcac372b 100644
--- a/webui/src/pages/Model/index.test.tsx
+++ b/webui/src/pages/Model/index.test.tsx
@@ -1,5 +1,5 @@
 import React from 'react';
-import { screen, waitFor } from '@testing-library/react';
+import { screen, waitFor, within } from '@testing-library/react';
 import userEvent from '@testing-library/user-event';
 import { vi } from 'vitest';
 
@@ -30,6 +30,9 @@ vi.mock('react-i18next', () => ({
   useTranslation: () => ({
     t: (key: string, params?: Record<string, unknown>) => {
       if (key === 'status.models') return `${params?.count ?? 0} models`;
+      if (key === 'modelSelection.info') {
+        return 'modelSelection.info';
+      }
       const translations: Record<string, string> = {
         pageTitle: 'Models',
         pageDescription: 'Manage providers',
@@ -308,6 +311,15 @@ describe('ModelPage configure provider dialog', () => {
     expect(mocks.revealCredentials).toHaveBeenCalledWith('openai');
   }
 
+  it('does not test provider credentials when the page loads', async () => {
+    renderWithRouter(<ModelPage />);
+
+    await waitFor(() => {
+      expect(mocks.getCredentials).toHaveBeenCalledWith('openai');
+    });
+    expect(mocks.testCredentials).not.toHaveBeenCalled();
+  });
+
   it('displays and preserves the existing API key when saving a new base URL', async () => {
     const user = userEvent.setup();
     await openConfigureDialog(user);
@@ -324,6 +336,7 @@ describe('ModelPage configure provider dialog', () => {
     await user.click(screen.getByRole('button', { name: 'Save' }));
 
     await waitFor(() => expect(mocks.setCredentials).toHaveBeenCalled());
+    expect(mocks.testCredentials).not.toHaveBeenCalled();
     const payload = mocks.setCredentials.mock.calls.at(-1)?.[1];
     expect(payload).toEqual(expect.objectContaining({
       base_url: 'https://new.example.com/v1',
@@ -368,3 +381,107 @@ describe('ModelPage configure provider dialog', () => {
     }));
   });
 });
+
+describe('ModelPage default model selector', () => {
+  const providers = [
+    {
+      id: 'openai',
+      name: 'OpenAI Gateway',
+      source: 'config',
+      env: [],
+      key: null,
+      options: {},
+      models: {},
+      configured: true,
+      modelCount: 1,
+      category: 'connected',
+    },
+    {
+      id: 'minimax',
+      name: 'MiniMax Cloud',
+      source: 'config',
+      env: [],
+      key: null,
+      options: {},
+      models: {},
+      configured: true,
+      modelCount: 1,
+      category: 'connected',
+    },
+  ];
+  const models = [
+    {
+      id: 'gpt-4o',
+      name: 'GPT-4o',
+      provider_id: 'openai',
+      model_type: 'llm',
+      status: 'active',
+      capabilities: { features: [], supports_streaming: true, supports_tools: true },
+    },
+    {
+      id: 'minimax-m3',
+      name: 'MiniMax Vision M3',
+      provider_id: 'minimax',
+      model_type: 'llm',
+      status: 'active',
+      capabilities: {
+        features: [],
+        supports_streaming: true,
+        supports_tools: true,
+        supports_vision: false,
+        modalities: { input: ['text', 'image'], output: ['text'] },
+      },
+      limits: { context_window: 200000, max_output_tokens: 8192 },
+      pricing: { input: 1.25, output: 5, unit: 1000000, currency: 'USD' },
+    },
+  ];
+
+  beforeEach(() => {
+    vi.clearAllMocks();
+    localStorage.clear();
+    sessionStorage.clear();
+    mocks.useProviders.mockReturnValue({
+      providers,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: mocks.refetch,
+    });
+    mocks.getSummary.mockResolvedValue({ data: null });
+    mocks.getResolved.mockResolvedValue({ data: { provider_id: 'openai', model_id: 'gpt-4o' } });
+    mocks.listDefinitions.mockResolvedValue({ data: { models, total: models.length } });
+    mocks.getCredentials.mockResolvedValue({ data: null });
+    mocks.testCredentials.mockResolvedValue({ data: { success: true, latency_ms: 10 } });
+  });
+
+  it('shows provider groups, model identity, vision, and details', async () => {
+    const user = userEvent.setup();
+    renderWithRouter(<ModelPage />);
+
+    await user.click(await screen.findByTitle('dashboard.setDefaultModel'));
+    const heading = await screen.findByRole('heading', { name: 'dashboard.setDefaultModel' });
+    const selector = within(heading.parentElement?.parentElement as HTMLElement);
+
+    expect(selector.getByText('MiniMax Cloud')).toBeInTheDocument();
+    expect(selector.getByText('MiniMax Vision M3')).toBeInTheDocument();
+    expect(selector.getByText('minimax-m3')).toBeInTheDocument();
+    expect(selector.getByText('form.vision')).toBeInTheDocument();
+    expect(selector.getByRole('button', {
+      name: 'modelSelection.info GPT-4o',
+    })).toBeInTheDocument();
+
+    await user.click(selector.getByRole('button', {
+      name: 'modelSelection.info MiniMax Vision M3',
+    }));
+
+    const tooltip = await screen.findByRole('tooltip');
+    expect(within(tooltip).getByText('form.modelId')).toBeInTheDocument();
+    expect(within(tooltip).getByText('form.contextWindow')).toBeInTheDocument();
+    expect(within(tooltip).getByText('form.pricing')).toBeInTheDocument();
+    expect(tooltip).toHaveTextContent('minimax-m3');
+    expect(tooltip).toHaveTextContent(/200(?:K|,?000)/);
+    expect(tooltip).toHaveTextContent(/1\.25/);
+    expect(tooltip).toHaveTextContent(/\b5(?:\.0+)?\b/);
+    expect(tooltip).toHaveTextContent(/USD|\$/);
+  });
+});
diff --git a/webui/src/pages/Model/index.tsx b/webui/src/pages/Model/index.tsx
index b34dc23f1..5d506f3d3 100644
--- a/webui/src/pages/Model/index.tsx
+++ b/webui/src/pages/Model/index.tsx
@@ -1,4 +1,5 @@
-import { useState, useEffect, useMemo, useCallback, useRef } from 'react';
+import { useState, useEffect, useMemo, useCallback, useRef, useId } from 'react';
+import { createPortal } from 'react-dom';
 import { useTranslation } from 'react-i18next';
 import {
   Brain, Cog, TestTube, Trash2, Search,
@@ -8,6 +9,7 @@ import {
   ChevronDown, Check, AlertCircle, Loader2,
   X, Shield, Pencil, Star, AlertTriangle,
   CheckCircle2,
+  Info,
 } from 'lucide-react';
 import PageHeader from '@/components/common/PageHeader';
 import LoadingSpinner from '@/components/common/LoadingSpinner';
@@ -159,12 +161,13 @@ export default function ModelPage() {
 
     Promise.all([
       defaultModelAPI.getResolved().catch(() => ({ data: null })),
-      modelV2API.listDefinitions({ enabled_only: true }).catch(() => ({ data: { models: [] } })),
+      modelV2API.listDefinitions({ enabled_only: true }),
     ]).then(([defaultRes, modelsRes]) => {
+      const availableModels = modelsRes.data.models || [];
+
       const dm = defaultRes.data;
       if (!dm) return;
 
-      const availableModels = modelsRes.data.models || [];
       const isValid = availableModels.some(
         m => m.provider_id === dm.provider_id && m.id === dm.model_id
       );
@@ -177,14 +180,11 @@ export default function ModelPage() {
         defaultModelAPI.delete('llm').catch(() => {});
         setDefaultModel(null);
       }
+    }).catch(() => {
+      // Keep the current default model when model definitions cannot be loaded.
     });
   }, []);
 
-  // Auto-test all configured providers on initial load, with cache (connected: 1h, failed: 5min)
-  const [autoTested, setAutoTested] = useState(false);
-  const handleTestConnectionRef = useRef<(id: string, silent?: boolean) => Promise<void>>(null!);
-
-
   // Only show configured (connected) providers
   const configuredProviders = useMemo(() => {
     return providers.filter(p => p.configured).sort((a, b) => a.name.localeCompare(b.name));
@@ -265,42 +265,28 @@ export default function ModelPage() {
   }, []);
   handleSelectProviderRef.current = handleSelectProvider;
 
-  const handleTestConnection = async (providerId: string, silent = false) => {
-    try {
-      const response = await providerAPI.testCredentials(providerId);
-      if (response.data.success) {
-        setConnectionStatus(prev => ({ ...prev, [providerId]: 'connected' }));
-        saveConnectionCache(providerId, 'connected');
-        if (!silent) toast.success(t('testSuccess'), `${t('latency')}: ${response.data.latency_ms}ms`);
-      } else {
-        setConnectionStatus(prev => ({ ...prev, [providerId]: 'failed' }));
-        saveConnectionCache(providerId, 'failed');
-        if (!silent) toast.error(t('testFailed'), response.data.message);
-      }
-    } catch (err: any) {
-      setConnectionStatus(prev => ({ ...prev, [providerId]: 'failed' }));
-      saveConnectionCache(providerId, 'failed');
-      if (!silent) toast.error(t('testFailed'), err.message);
-    }
-  };
-  handleTestConnectionRef.current = handleTestConnection;
-
-  // Auto-test effect (runs once after handlers are defined)
+  // Restore statuses from explicit connection tests without issuing API calls.
   useEffect(() => {
-    if (!loading && !autoTested && providers.length > 0) {
-      setAutoTested(true);
-      const cache = loadConnectionCache();
-      const configuredList = providers.filter(p => p.configured);
-      configuredList.forEach(p => {
-        const cached = getCachedStatus(cache, p.id);
-        if (cached !== null) {
-          setConnectionStatus(prev => ({ ...prev, [p.id]: cached }));
+    if (!loading && providers.length > 0) {
+      const cachedStatuses = loadConnectionCache();
+      setConnectionStatus(prev => {
+        let changed = false;
+        const next = { ...prev };
+        providers.filter(p => p.configured).forEach(p => {
+          const cached = getCachedStatus(cachedStatuses, p.id);
+          if (cached !== null && next[p.id] !== cached) {
+            next[p.id] = cached;
+            changed = true;
+          }
+        });
+        if (changed) {
+          return next;
         } else {
-          handleTestConnectionRef.current(p.id, true);
+          return prev;
         }
       });
     }
-  }, [loading, providers, autoTested]);
+  }, [loading, providers]);
 
   const handleDeleteProvider = async (providerId: string) => {
     const isDefaultAffected = defaultModel && defaultModel.provider_id === providerId;
@@ -358,10 +344,7 @@ export default function ModelPage() {
     if (addedProviderId) {
       setConnectionStatus(prev => ({ ...prev, [addedProviderId]: 'unknown' }));
       setPendingSelectId(addedProviderId);
-      setTimeout(() => {
-        refetch();
-        handleTestConnection(addedProviderId, true);
-      }, 500);
+      setTimeout(() => refetch(), 500);
     } else {
       setTimeout(() => refetch(), 300);
     }
@@ -572,8 +555,6 @@ export default function ModelPage() {
               setCredentials(res.data);
               // Refresh model list in right panel
               handleSelectProvider(selectedProvider);
-              // Auto-test after credential update
-              handleTestConnection(selectedProvider.id, true);
             }
             refetch();
             setShowConfigDialog(false);
@@ -623,6 +604,7 @@ export default function ModelPage() {
       {showDefaultModelDialog && (
         <SetDefaultModelDialog
           current={defaultModel}
+          providers={providers}
           onClose={() => setShowDefaultModelDialog(false)}
           onSaved={(m) => {
             setDefaultModel(m);
@@ -633,6 +615,7 @@ export default function ModelPage() {
           }}
         />
       )}
+
     </div>
   );
 }
@@ -666,7 +649,7 @@ function DashboardStrip({
   }, [i18n.language]);
 
   return (
-    <div className="grid grid-cols-5 gap-3 mb-4">
+    <div className="grid grid-cols-2 gap-3 mb-4 lg:grid-cols-5">
       {/* Default Model Card */}
       <div className="rounded-lg border p-3 bg-purple-50 text-purple-700 border-purple-200">
         <div className="flex items-center justify-between mb-1 opacity-70">
@@ -2889,15 +2872,216 @@ function ModelDetailSheet({
   );
 }
 
+// ==================== Shared Model Selection ====================
+
+type ModelSelectionGroup = {
+  providerId: string;
+  providerName: string;
+  models: ModelDefinitionV2[];
+};
+
+function modelSupportsVision(model: ModelDefinitionV2): boolean {
+  const capabilities = model.capabilities;
+  return capabilities?.supports_vision === true
+    || capabilities?.features?.includes('vision') === true
+    || capabilities?.modalities?.input?.includes('image') === true;
+}
+
+function groupModelsForSelection(
+  models: ModelDefinitionV2[],
+  providerNames: Map<string, string>,
+): ModelSelectionGroup[] {
+  const grouped = new Map<string, ModelDefinitionV2[]>();
+  models.forEach(model => {
+    const entries = grouped.get(model.provider_id) ?? [];
+    entries.push(model);
+    grouped.set(model.provider_id, entries);
+  });
+  return Array.from(grouped.entries())
+    .map(([providerId, providerModels]) => ({
+      providerId,
+      providerName: providerNames.get(providerId) || providerId,
+      models: providerModels.sort((a, b) => (a.name || a.id).localeCompare(b.name || b.id)),
+    }))
+    .sort((a, b) => a.providerName.localeCompare(b.providerName));
+}
+
+function formatModelContextWindow(model: ModelDefinitionV2): string {
+  const contextWindow = model.limits?.context_window;
+  if (!contextWindow) return '—';
+  if (contextWindow >= 1_000_000) {
+    return `${Number((contextWindow / 1_000_000).toFixed(1))}M`;
+  }
+  if (contextWindow >= 1_000) {
+    return `${Number((contextWindow / 1_000).toFixed(1))}K`;
+  }
+  return String(contextWindow);
+}
+
+function formatModelPricing(
+  model: ModelDefinitionV2,
+  freeLabel: string,
+  unavailableLabel: string,
+): string {
+  const pricing = model.pricing;
+  if (!pricing) return unavailableLabel;
+  if (pricing.input === 0 && pricing.output === 0) return freeLabel;
+  const symbol = pricing.currency === 'CNY' ? '¥' : pricing.currency === 'USD' ? '$' : `${pricing.currency} `;
+  return `${symbol}${pricing.input} / ${symbol}${pricing.output} / 1M`;
+}
+
+function ModelSelectionInfo({ model }: { model: ModelDefinitionV2 }) {
+  const { t } = useTranslation('model');
+  const tooltipId = useId();
+  const [position, setPosition] = useState<{ x: number; y: number } | null>(null);
+
+  const show = useCallback((target: HTMLElement) => {
+    const rect = target.getBoundingClientRect();
+    const tooltipHalfWidth = 128;
+    const viewportPadding = 8;
+    setPosition({
+      x: Math.min(
+        window.innerWidth - tooltipHalfWidth - viewportPadding,
+        Math.max(tooltipHalfWidth + viewportPadding, rect.left + rect.width / 2),
+      ),
+      y: rect.top - 8,
+    });
+  }, []);
+  const hide = useCallback(() => setPosition(null), []);
+  const label = `${t('modelSelection.info')} ${model.name || model.id}`;
+
+  return (
+    <>
+      <button
+        type="button"
+        aria-label={label}
+        aria-describedby={position ? tooltipId : undefined}
+        onClick={event => {
+          event.stopPropagation();
+          show(event.currentTarget);
+        }}
+        onFocus={event => show(event.currentTarget)}
+        onBlur={hide}
+        onPointerEnter={event => show(event.currentTarget)}
+        onPointerLeave={hide}
+        onMouseEnter={event => show(event.currentTarget)}
+        onMouseLeave={hide}
+        className="mr-3 inline-flex h-7 w-7 shrink-0 items-center justify-center rounded-md text-gray-300 transition-colors hover:bg-gray-100 hover:text-gray-600 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-slate-400 focus-visible:ring-offset-1"
+      >
+        <Info className="h-4 w-4" aria-hidden="true" />
+      </button>
+      {position && typeof document !== 'undefined' && createPortal(
+        <div
+          id={tooltipId}
+          role="tooltip"
+          className="pointer-events-none fixed z-[1000] w-64 -translate-x-1/2 -translate-y-full rounded-lg border border-gray-200 bg-white px-3 py-2.5 text-xs leading-relaxed text-gray-600 shadow-lg"
+          style={{ left: position.x, top: position.y }}
+        >
+          <div className="mb-1.5 truncate font-semibold text-gray-900">{model.name || model.id}</div>
+          <dl className="space-y-1">
+            <div className="flex gap-2">
+              <dt className="shrink-0 text-gray-400">{t('form.modelId')}</dt>
+              <dd className="min-w-0 flex-1 break-all text-right font-mono text-[11px] text-gray-700">{model.id}</dd>
+            </div>
+            <div className="flex gap-2">
+              <dt className="shrink-0 text-gray-400">{t('form.contextWindow')}</dt>
+              <dd className="min-w-0 flex-1 text-right text-gray-700">{formatModelContextWindow(model)}</dd>
+            </div>
+            <div className="flex gap-2">
+              <dt className="shrink-0 text-gray-400">{t('form.pricing')}</dt>
+              <dd className="min-w-0 flex-1 text-right text-gray-700">
+                {formatModelPricing(model, t('modelSelection.free'), t('modelSelection.unavailable'))}
+              </dd>
+            </div>
+          </dl>
+          <div className="absolute left-1/2 top-full -translate-x-1/2 border-4 border-transparent border-t-gray-200" />
+        </div>,
+        document.body,
+      )}
+    </>
+  );
+}
+
+function ModelSelectionList({
+  groups,
+  selectedKey,
+  pendingKey,
+  disabled = false,
+  onSelect,
+}: {
+  groups: ModelSelectionGroup[];
+  selectedKey?: string | null;
+  pendingKey?: string | null;
+  disabled?: boolean;
+  onSelect: (model: ModelDefinitionV2) => void;
+}) {
+  const { t } = useTranslation('model');
+
+  return (
+    <div className="overflow-hidden rounded-xl border border-gray-200 bg-white">
+      {groups.map(group => (
+        <div key={group.providerId}>
+          <div className="border-b border-gray-100 bg-gray-50 px-4 py-2 text-xs font-semibold uppercase tracking-wide text-gray-500">
+            {group.providerName}
+          </div>
+          {group.models.map(model => {
+            const key = `${model.provider_id}/${model.id}`;
+            const selected = selectedKey === key;
+            return (
+              <div
+                key={key}
+                className={`flex min-w-0 items-center border-b border-gray-100 transition-colors last:border-0 ${
+                  selected ? 'bg-slate-50' : 'bg-white hover:bg-gray-50'
+                }`}
+              >
+                <button
+                  type="button"
+                  onClick={() => onSelect(model)}
+                  disabled={disabled}
+                  aria-pressed={selected}
+                  className="flex min-w-0 flex-1 items-center gap-3 px-4 py-3 text-left text-sm text-gray-700 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-slate-400 disabled:cursor-not-allowed disabled:opacity-50"
+                >
+                  <span className={`min-w-0 flex-1 truncate ${selected ? 'font-semibold text-slate-900' : 'font-medium text-gray-800'}`}>
+                    {model.name || model.id}
+                  </span>
+                  {modelSupportsVision(model) && (
+                    <span
+                      title={t('form.vision')}
+                      className="shrink-0 rounded-md bg-gray-100 px-2 py-0.5 text-[10px] font-semibold text-gray-600"
+                    >
+                      {t('form.vision')}
+                    </span>
+                  )}
+                  <span className="max-w-[42%] shrink-0 truncate text-xs text-gray-400" title={model.id}>
+                    {model.id}
+                  </span>
+                  {pendingKey === key ? (
+                    <Loader2 className="h-4 w-4 shrink-0 animate-spin text-gray-400" />
+                  ) : selected ? (
+                    <Check className="h-4 w-4 shrink-0 text-slate-700" />
+                  ) : null}
+                </button>
+                <ModelSelectionInfo model={model} />
+              </div>
+            );
+          })}
+        </div>
+      ))}
+    </div>
+  );
+}
+
 // ==================== Set Default Model Dialog ====================
 
 function SetDefaultModelDialog({
   current,
+  providers,
   onClose,
   onSaved,
   onCleared,
 }: {
   current: { provider_id: string; model_id: string } | null;
+  providers: EnrichedProvider[];
   onClose: () => void;
   onSaved: (m: { provider_id: string; model_id: string }) => void;
   onCleared?: () => void;
@@ -2908,12 +3092,33 @@ function SetDefaultModelDialog({
   const [loading, setLoading] = useState(true);
   const [saving, setSaving] = useState<string | null>(null);
   const [invalidWarning, setInvalidWarning] = useState<string | null>(null);
+  const closeButtonRef = useRef<HTMLButtonElement>(null);
+
+  useEffect(() => {
+    const previouslyFocused = document.activeElement instanceof HTMLElement
+      ? document.activeElement
+      : null;
+    closeButtonRef.current?.focus();
+    return () => previouslyFocused?.focus();
+  }, []);
+
+  useEffect(() => {
+    const handleKeyDown = (event: KeyboardEvent) => {
+      if (event.key !== 'Escape' || saving) return;
+      event.preventDefault();
+      onClose();
+    };
+    window.addEventListener('keydown', handleKeyDown);
+    return () => window.removeEventListener('keydown', handleKeyDown);
+  }, [onClose, saving]);
 
   useEffect(() => {
     setLoading(true);
     setInvalidWarning(null);
     modelV2API.listDefinitions({ enabled_only: true }).then(r => {
-      const loadedModels = r.data.models || [];
+      const loadedModels = (r.data.models || []).filter(
+        model => model.model_type === 'llm',
+      );
       setModels(loadedModels);
 
       // 校验当前默认模型是否仍在可用列表中
@@ -2931,15 +3136,14 @@ function SetDefaultModelDialog({
     }).catch(() => {}).finally(() => setLoading(false));
   }, []);
 
-  // Group by provider_id
-  const grouped = useMemo(() => {
-    const map: Record<string, ModelDefinitionV2[]> = {};
-    for (const m of models) {
-      if (!map[m.provider_id]) map[m.provider_id] = [];
-      map[m.provider_id].push(m);
-    }
-    return map;
-  }, [models]);
+  const providerNames = useMemo(
+    () => new Map(providers.map(provider => [provider.id, provider.name || provider.id])),
+    [providers],
+  );
+  const grouped = useMemo(
+    () => groupModelsForSelection(models, providerNames),
+    [models, providerNames],
+  );
 
   const handleSelect = async (providerId: string, modelId: string) => {
     setSaving(`${providerId}/${modelId}`);
@@ -2962,21 +3166,36 @@ function SetDefaultModelDialog({
   };
 
   return (
-    <div className="fixed inset-0 bg-black/50 flex items-center justify-center z-50" onClick={onClose}>
+    <div
+      className="fixed inset-0 bg-black/50 flex items-center justify-center z-50"
+      onClick={() => {
+        if (!saving) onClose();
+      }}
+    >
       <div
-        className="bg-white rounded-xl shadow-xl w-full max-w-md max-h-[70vh] flex flex-col"
+        role="dialog"
+        aria-modal="true"
+        aria-labelledby="default-model-dialog-title"
+        className="bg-white rounded-xl shadow-xl w-full max-w-xl max-h-[78vh] flex flex-col"
         onClick={e => e.stopPropagation()}
       >
         <div className="flex items-center justify-between px-5 py-4 border-b border-gray-200">
-          <h2 className="text-lg font-semibold text-gray-900">{t('dashboard.setDefaultModel')}</h2>
-          <button onClick={onClose} className="text-gray-400 hover:text-gray-600">
+          <h2 id="default-model-dialog-title" className="text-lg font-semibold text-gray-900">{t('dashboard.setDefaultModel')}</h2>
+          <button
+            ref={closeButtonRef}
+            type="button"
+            onClick={onClose}
+            disabled={saving !== null}
+            aria-label={t('modelSelection.closeDefault')}
+            className="text-gray-400 hover:text-gray-600 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-slate-400 disabled:opacity-50"
+          >
             <X className="w-5 h-5" />
           </button>
         </div>
 
-        <div className="flex-1 overflow-y-auto">
+        <div className="min-h-0 flex-1 overflow-y-auto px-5 py-4">
           {invalidWarning && (
-            <div className="mx-4 mt-3 flex items-start gap-2 rounded-lg border border-amber-200 bg-amber-50 px-3 py-2 text-xs text-amber-700">
+            <div className="mb-3 flex items-start gap-2 rounded-lg border border-amber-200 bg-amber-50 px-3 py-2 text-xs text-amber-700">
               <AlertTriangle className="mt-0.5 w-3.5 h-3.5 flex-shrink-0" />
               <span>{invalidWarning}</span>
             </div>
@@ -2985,35 +3204,16 @@ function SetDefaultModelDialog({
             <div className="flex justify-center py-10">
               <Loader2 className="w-6 h-6 animate-spin text-gray-400" />
             </div>
-          ) : Object.keys(grouped).length === 0 ? (
+          ) : grouped.length === 0 ? (
             <div className="py-10 text-center text-sm text-gray-500">{t('detail.noModels')}</div>
           ) : (
-            Object.entries(grouped).map(([providerId, provModels]) => (
-              <div key={providerId}>
-                <div className="px-5 py-2 text-xs font-semibold text-gray-500 bg-gray-50 border-b border-gray-100 uppercase tracking-wide">
-                  {providerId}
-                </div>
-                {provModels.map(m => {
-                  const isActive = current?.provider_id === providerId && current?.model_id === m.id;
-                  const key = `${providerId}/${m.id}`;
-                  return (
-                    <button
-                      key={m.id}
-                      onClick={() => handleSelect(providerId, m.id)}
-                      disabled={saving !== null}
-                      className={`w-full flex items-center justify-between px-5 py-3 text-sm text-left hover:bg-gray-50 transition-colors border-b border-gray-100 last:border-0 ${isActive ? 'text-purple-700 font-medium' : 'text-gray-700'}`}
-                    >
-                      <span className="truncate">{m.name || m.id}</span>
-                      {saving === key ? (
-                        <Loader2 className="w-4 h-4 animate-spin text-gray-400 flex-shrink-0" />
-                      ) : isActive ? (
-                        <Check className="w-4 h-4 text-purple-600 flex-shrink-0" />
-                      ) : null}
-                    </button>
-                  );
-                })}
-              </div>
-            ))
+            <ModelSelectionList
+              groups={grouped}
+              selectedKey={current ? `${current.provider_id}/${current.model_id}` : null}
+              pendingKey={saving}
+              disabled={saving !== null}
+              onSelect={model => void handleSelect(model.provider_id, model.id)}
+            />
           )}
         </div>
       </div>
diff --git a/webui/src/pages/Session/index.test.tsx b/webui/src/pages/Session/index.test.tsx
index 0b2b857ed..5db9684c1 100644
--- a/webui/src/pages/Session/index.test.tsx
+++ b/webui/src/pages/Session/index.test.tsx
@@ -4,8 +4,16 @@ import { act, fireEvent, render, screen, waitFor, within } from '@testing-librar
 import userEvent from '@testing-library/user-event';
 import { MemoryRouter, useNavigate } from 'react-router-dom';
 import { __resetChatModelResourcesForTesting } from '@/hooks/useChatModelResources';
+import { formatRelativeTime } from '@/utils/time';
 import SessionPage from './index';
 
+const sessionStatusSSEOptionsRef = vi.hoisted(() => ({
+  current: null as null | {
+    onEvent: (event: { type: string; properties?: Record<string, unknown> }) => void;
+    onReconnect?: () => void;
+  },
+}));
+
 const {
   client,
   sessionApi,
@@ -20,6 +28,8 @@ const {
   defaultModelAPI,
   modelV2API,
   hubAPI,
+  workflowAPI,
+  skillAPI,
   toast,
 } = vi.hoisted(() => ({
   client: {
@@ -29,9 +39,11 @@ const {
     post: vi.fn(),
   },
   sessionApi: {
+    archive: vi.fn(),
     delete: vi.fn(),
     get: vi.fn(),
     getMessages: vi.fn(),
+    moveToProject: vi.fn(),
     update: vi.fn(),
   },
   updateSessionTitle: vi.fn(),
@@ -53,6 +65,12 @@ const {
     install: vi.fn(),
     installStream: vi.fn(),
   },
+  workflowAPI: {
+    listSummaries: vi.fn(),
+  },
+  skillAPI: {
+    status: vi.fn(),
+  },
   toast: {
     error: vi.fn(),
     info: vi.fn(),
@@ -64,6 +82,7 @@ const {
 vi.mock('@/api/client', () => ({
   __esModule: true,
   default: client,
+  getApiBase: () => '',
 }));
 
 vi.mock('@/api/session', () => ({
@@ -74,6 +93,14 @@ vi.mock('@/api/hub', () => ({
   hubAPI,
 }));
 
+vi.mock('@/api/workflow', () => ({
+  workflowAPI,
+}));
+
+vi.mock('@/api/skill', () => ({
+  skillAPI,
+}));
+
 vi.mock('@/hooks/useSessions', () => ({
   useSessions,
 }));
@@ -86,6 +113,18 @@ vi.mock('@/hooks/useProviders', () => ({
   useProviders,
 }));
 
+vi.mock('@/hooks/useSSE', () => ({
+  useSSE: (options: typeof sessionStatusSSEOptionsRef.current) => {
+    sessionStatusSSEOptionsRef.current = options;
+    return {
+      status: 'connected',
+      retryCount: 0,
+      reconnect: vi.fn(),
+      disconnect: vi.fn(),
+    };
+  },
+}));
+
 vi.mock('@/api/provider', () => ({
   defaultModelAPI,
   modelV2API,
@@ -110,6 +149,8 @@ vi.mock('@/components/common/SessionChat', () => ({
     sessionId,
     mentionAgents,
     toolbarSlot,
+    composerAddMenuSlot,
+    onComposerAddMenuOpenChange,
     centerToolbarSlot,
     welcomeContent,
     initialMessage,
@@ -118,6 +159,10 @@ vi.mock('@/components/common/SessionChat', () => ({
     onSSEEvent,
     agentName,
     model,
+    executionMode,
+    onExecutionModeAccepted,
+    supportsVision,
+    contextWindowTokens,
     display,
     hideInput,
   }: {
@@ -125,11 +170,21 @@ vi.mock('@/components/common/SessionChat', () => ({
     agentName?: string;
     mentionAgents?: Array<{ name: string }>;
     toolbarSlot?: React.ReactNode;
+    composerAddMenuSlot?: React.ReactNode | ((actions: {
+      closeMenu: () => void;
+      insertMention: (agentName: string) => void;
+      insertReference: (value: string, kind: 'workflow' | 'skill') => void;
+    }) => React.ReactNode);
+    onComposerAddMenuOpenChange?: (open: boolean) => void;
     centerToolbarSlot?: React.ReactNode;
     welcomeContent?: React.ReactNode | ((setInput: (text: string) => void) => React.ReactNode);
     initialMessage?: string | null;
     initialDisplayText?: string | null;
     model?: { providerID: string; modelID: string } | null;
+    executionMode?: 'build' | 'plan' | 'goal';
+    onExecutionModeAccepted?: (mode: 'build' | 'plan' | 'goal') => void;
+    supportsVision?: boolean;
+    contextWindowTokens?: number | null;
     hideInput?: boolean;
     display?: {
       compact?: boolean;
@@ -145,6 +200,7 @@ vi.mock('@/components/common/SessionChat', () => ({
       agentOverride?: string,
       modelOverride?: unknown,
       options?: { displayText?: string },
+      executionModeOverride?: 'build' | 'plan' | 'goal',
     ) => Promise<unknown> | unknown;
     onSSEEvent?: (event: { type: string; properties?: Record<string, unknown> }) => void;
   }) {
@@ -155,6 +211,9 @@ vi.mock('@/components/common/SessionChat', () => ({
         data-agent-name={agentName ?? ''}
         data-mention-agents={(mentionAgents ?? []).map((a) => a.name).join(',')}
         data-model={model ? `${model.providerID}/${model.modelID}` : ''}
+        data-execution-mode={executionMode ?? ''}
+        data-supports-vision={String(Boolean(supportsVision))}
+        data-context-window={contextWindowTokens ?? ''}
         data-collapse-intermediate={String(Boolean(display?.collapseIntermediateSteps))}
         data-process-groups-default-open={String(Boolean(display?.processGroupsDefaultOpen))}
         data-process-groups-open-while-active={String(Boolean(display?.processGroupsOpenWhileActive))}
@@ -163,15 +222,43 @@ vi.mock('@/components/common/SessionChat', () => ({
         data-initial-display={initialDisplayText ?? ''}
       >
         {sessionId ?? 'no-session'}
+        <button type="button" onClick={() => onComposerAddMenuOpenChange?.(true)}>
+          mock-open-add-menu
+        </button>
+        {typeof composerAddMenuSlot === 'function'
+          ? composerAddMenuSlot({
+            closeMenu: vi.fn(),
+            insertMention: (agentName) => setInput(`subagent:${agentName} `),
+            insertReference: (value, kind) => {
+              setInput(`${kind}:${value} `);
+            },
+          })
+          : composerAddMenuSlot}
         {toolbarSlot}
         {centerToolbarSlot}
         {!sessionId && welcomeContent ? (
           typeof welcomeContent === 'function' ? welcomeContent(setInput) : welcomeContent
         ) : null}
         <div data-testid="mock-chat-input">{input}</div>
-        <button type="button" onClick={() => void onCreateAndSend?.('hello from empty session', [], agentName)}>
+        <button
+          type="button"
+          onClick={() => void onCreateAndSend?.(
+            'hello from empty session',
+            [],
+            agentName,
+            undefined,
+            undefined,
+            executionMode,
+          )}
+        >
           mock-create-and-send
         </button>
+        <button
+          type="button"
+          onClick={() => executionMode && onExecutionModeAccepted?.(executionMode)}
+        >
+          mock-accept-mode
+        </button>
         <button
           type="button"
           onClick={() => onSSEEvent?.({
@@ -181,6 +268,19 @@ vi.mock('@/components/common/SessionChat', () => ({
         >
           mock-session-updated
         </button>
+        <button
+          type="button"
+          onClick={() => onSSEEvent?.({
+            type: 'session.execution_mode.changed',
+            properties: {
+              sessionID: sessionId,
+              executionMode: 'build',
+              reason: 'plan-approved',
+            },
+          })}
+        >
+          mock-plan-approved
+        </button>
       </div>
     );
   },
@@ -199,6 +299,7 @@ vi.mock('@/utils/agentDisplay', () => ({
 
 vi.mock('@/utils/time', () => ({
   formatSessionDate: () => 'formatted-date',
+  formatRelativeTime: vi.fn(() => '17小时前'),
 }));
 
 vi.mock('react-i18next', () => ({
@@ -282,6 +383,7 @@ describe('SessionPage session actions menu', () => {
   beforeEach(() => {
     vi.clearAllMocks();
     __resetChatModelResourcesForTesting();
+    sessionStatusSSEOptionsRef.current = null;
     localStorage.clear();
     sessionStorage.clear();
 
@@ -311,6 +413,8 @@ describe('SessionPage session actions menu', () => {
     });
     defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: '', model_id: '' } });
     modelV2API.listDefinitions.mockResolvedValue({ data: { models: [] } });
+    workflowAPI.listSummaries.mockResolvedValue({ data: [] });
+    skillAPI.status.mockResolvedValue({ data: [] });
     client.get.mockResolvedValue({
       data: [{
         id: 'default',
@@ -328,6 +432,12 @@ describe('SessionPage session actions menu', () => {
     hubAPI.installStream.mockResolvedValue(undefined);
 
     sessionApi.update.mockResolvedValue({ ...session, title: 'Renamed Session' });
+    sessionApi.moveToProject.mockResolvedValue({
+      ...session,
+      projectID: 'prj_labs',
+      effectiveProjectID: 'prj_labs',
+      directory: '/tmp/labs',
+    });
     client.patch.mockResolvedValue({ data: { id: 'prj_project2', worktree: '/tmp/labs', name: 'Renamed Project' } });
     client.post.mockResolvedValue({ data: secondSession });
     sessionApi.get.mockResolvedValue(session);
@@ -342,11 +452,123 @@ describe('SessionPage session actions menu', () => {
         parts: [{ id: 'part-1', type: 'text', text: 'hello export' }],
       },
     ]);
+    sessionApi.archive.mockResolvedValue({ id: 'session-1', status: 'archived' });
     sessionApi.delete.mockResolvedValue(true);
 
     vi.stubGlobal('confirm', vi.fn(() => true));
   });
 
+  it('renders Build by default and keeps Agent in the Add menu', async () => {
+    renderSessionPage();
+
+    await screen.findByRole('button', { name: 'executionMode.title' });
+    const agentButton = screen.getByRole('button', { name: 'chat.addMenu.agent' });
+
+    expect(screen.getByTestId('session-chat')).toHaveAttribute('data-execution-mode', 'build');
+    expect(agentButton).toHaveAttribute('aria-haspopup', 'menu');
+  });
+
+  it('persists Plan per session', async () => {
+    const user = userEvent.setup();
+    renderSessionPage('/sessions?session=session-1');
+
+    const modeButton = await screen.findByRole('button', { name: 'executionMode.title' });
+    await user.click(modeButton);
+    await user.click(screen.getByRole('menuitemradio', {
+      name: /executionMode.options.plan.label/,
+    }));
+
+    expect(screen.getByTestId('session-chat')).toHaveAttribute('data-execution-mode', 'plan');
+    expect(localStorage.getItem('flocks:session-execution-mode:session-1')).toBe('plan');
+  });
+
+  it('restores a persisted session execution mode', async () => {
+    localStorage.setItem('flocks:session-execution-mode:session-1', 'plan');
+    renderSessionPage('/sessions?session=session-1');
+
+    await waitFor(() => {
+      expect(screen.getByTestId('session-chat')).toHaveAttribute('data-execution-mode', 'plan');
+    });
+  });
+
+  it('switches an approved Plan to Build from the session event', async () => {
+    const user = userEvent.setup();
+    renderSessionPage('/sessions?session=session-1');
+
+    await user.click(await screen.findByRole('button', { name: 'executionMode.title' }));
+    await user.click(screen.getByRole('menuitemradio', {
+      name: /executionMode.options.plan.label/,
+    }));
+    expect(screen.getByTestId('session-chat')).toHaveAttribute('data-execution-mode', 'plan');
+
+    await user.click(screen.getByRole('button', { name: 'mock-plan-approved' }));
+
+    expect(screen.getByTestId('session-chat')).toHaveAttribute('data-execution-mode', 'build');
+    expect(localStorage.getItem('flocks:session-execution-mode:session-1')).toBeNull();
+  });
+
+  it('resets Goal to Build after the prompt is accepted', async () => {
+    const user = userEvent.setup();
+    renderSessionPage('/sessions?session=session-1');
+
+    await user.click(await screen.findByRole('button', { name: 'executionMode.title' }));
+    await user.click(screen.getByRole('menuitemradio', {
+      name: /executionMode.options.goal.label/,
+    }));
+    expect(screen.getByTestId('session-chat')).toHaveAttribute('data-execution-mode', 'goal');
+
+    await user.click(screen.getByRole('button', { name: 'mock-accept-mode' }));
+
+    expect(screen.getByTestId('session-chat')).toHaveAttribute('data-execution-mode', 'build');
+    expect(localStorage.getItem('flocks:session-execution-mode:session-1')).toBeNull();
+  });
+
+  it('promotes a draft Plan mode when the first message creates a session', async () => {
+    const user = userEvent.setup();
+    renderSessionPage();
+
+    await user.click(await screen.findByRole('button', { name: 'executionMode.title' }));
+    await user.click(screen.getByRole('menuitemradio', {
+      name: /executionMode.options.plan.label/,
+    }));
+    await user.click(screen.getByRole('button', { name: 'mock-create-and-send' }));
+
+    await waitFor(() => {
+      expect(client.post).toHaveBeenCalledWith(
+        '/api/session/session-2/prompt_async',
+        expect.objectContaining({ executionMode: 'plan' }),
+      );
+    });
+    expect(localStorage.getItem('flocks:session-execution-mode:session-2')).toBe('plan');
+    expect(localStorage.getItem('flocks:session-execution-mode:draft')).toBeNull();
+  });
+
+  it('keeps the workbench visible and shows a page refresh state while sessions load', () => {
+    useSessions.mockReturnValue({
+      sessions: [],
+      loading: true,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+
+    renderSessionPage();
+
+    expect(screen.getByLabelText('managementTitle')).toBeInTheDocument();
+    expect(screen.getByTestId('workbench-refresh-status')).toHaveTextContent('refreshingWorkbench');
+    expect(screen.getByTestId('session-list-skeleton')).toBeInTheDocument();
+    expect(screen.getByTestId('session-chat')).toHaveTextContent('no-session');
+    expect(screen.queryByText('loading-spinner')).not.toBeInTheDocument();
+    expect(screen.getByTestId('session-list-scroll')).toHaveClass(
+      'session-sidebar-scrollbar',
+      'overflow-y-auto',
+    );
+    expect(screen.getByTestId('session-list-scroll')).not.toHaveClass('scrollbar-hide');
+  });
+
   it('shows default sessions under tasks without a default project row', async () => {
     const user = userEvent.setup();
     renderSessionPage();
@@ -355,6 +577,15 @@ describe('SessionPage session actions menu', () => {
     const projectsHeading = screen.getByText('projectsSection');
     const tasksSection = tasksHeading.closest('section');
     const projectsSection = projectsHeading.closest('section');
+    const newSessionButton = screen.getByRole('button', { name: 'newSession' });
+    const searchInput = screen.getByPlaceholderText('filterConversations');
+    expect(newSessionButton.previousElementSibling).toHaveClass('left-2', 'h-3.5', 'w-3.5');
+    expect(searchInput.previousElementSibling).toHaveClass('left-2', 'h-3.5', 'w-3.5');
+    expect(searchInput).toHaveClass('text-sm', 'font-medium');
+    expect(tasksHeading.closest('div')).toHaveClass('px-2', 'text-xs', 'text-zinc-500');
+    expect(projectsHeading.closest('div')).toHaveClass('px-2', 'text-xs', 'text-zinc-500');
+    expect(tasksHeading.nextElementSibling).toHaveTextContent('(1)');
+    expect(projectsHeading.nextElementSibling).toHaveTextContent('(0)');
     expect(tasksSection).not.toBeNull();
     expect(projectsSection).not.toBeNull();
     expect(tasksSection?.parentElement).toBe(projectsSection?.parentElement);
@@ -364,27 +595,182 @@ describe('SessionPage session actions menu', () => {
       pageSize: 20,
     });
     expect(screen.queryByText('defaultProjectName')).not.toBeInTheDocument();
-    expect(screen.getByText('Original Session')).toBeInTheDocument();
+    const taskTitle = screen.getByText('Original Session').closest('h3');
+    expect(taskTitle).toHaveClass('text-sm', 'font-medium');
+    expect(taskTitle?.parentElement?.parentElement).toHaveClass('px-3');
 
-    await user.click(screen.getByRole('button', { name: 'toggleTasks' }));
+    const tasksToggle = screen.getByRole('button', { name: 'toggleTasks' });
+    expect(tasksToggle).toContainElement(tasksHeading);
+    expect(tasksToggle.querySelector('svg')).toHaveClass('h-3.5', 'w-3.5');
+    expect(screen.queryByRole('button', { name: 'selectTasks' })).not.toBeInTheDocument();
+    expect(screen.queryByRole('button', { name: 'createTaskSession' })).not.toBeInTheDocument();
+
+    await user.click(tasksToggle);
     expect(screen.queryByText('Original Session')).not.toBeInTheDocument();
 
-    await user.click(screen.getByRole('button', { name: 'selectTasks' }));
+    await user.click(tasksToggle);
     expect(screen.getByText('Original Session')).toBeInTheDocument();
   });
 
-  it('creates a new session from the tasks row', async () => {
-    const user = userEvent.setup();
+  it('keeps the workbench canvas, sidebar, selected row, and dark palette classes stable', async () => {
+    renderSessionPage('/sessions?session=session-1');
+
+    const workbenchSidebar = screen.getByLabelText('managementTitle');
+    const workbenchCanvas = workbenchSidebar.parentElement;
+    const mainCanvas = workbenchSidebar.nextElementSibling;
+    const sessionTitle = await within(workbenchSidebar).findByText('Original Session');
+    const selectedRow = sessionTitle.closest('div.group');
+
+    expect(workbenchCanvas).toHaveClass('bg-[#fcfcfd]', 'dark:bg-[#303842]');
+    expect(workbenchSidebar).toHaveClass('bg-gray-50', 'dark:bg-[#252c35]');
+    expect(workbenchSidebar).toHaveClass('h-full', 'border-r');
+    expect(workbenchSidebar).toHaveClass('border-black/[0.10]');
+    expect(workbenchSidebar).not.toHaveClass('rounded-2xl');
+    expect(workbenchSidebar.className).not.toContain('shadow-');
+    expect(mainCanvas).toHaveClass('bg-[#fcfcfd]', 'dark:bg-[#303842]');
+    expect(within(mainCanvas as HTMLElement).getByRole('heading', { level: 2 })).toHaveClass('text-[#555a61]');
+    await waitFor(() => {
+      expect(selectedRow).toHaveClass('bg-zinc-200/70', 'dark:bg-[#3a434e]');
+    });
+  });
+
+  it('shows and clears the sidebar running state from recovered and live session status', async () => {
+    useSessions.mockReturnValue({
+      sessions: [session, secondSession],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    client.get.mockImplementation((url: string) => Promise.resolve({
+      data: url === '/api/session/status'
+        ? {
+            [session.id]: { type: 'busy' },
+            [secondSession.id]: { type: 'busy' },
+          }
+        : [{
+            id: 'default',
+            worktree: '/tmp/project',
+            name: '默认',
+            isDefault: true,
+            pathStatus: 'available',
+            sessionCount: 2,
+          }],
+    }));
+
     renderSessionPage();
 
-    await screen.findByText('tasksSection');
-    await user.click(screen.getByRole('button', { name: 'createTaskSession' }));
+    const runningStatuses = await screen.findAllByRole('status', { name: 'chat.tool.running' });
+    expect(runningStatuses.map((status) => status.getAttribute('data-session-running')))
+      .toEqual(expect.arrayContaining([session.id, secondSession.id]));
 
-    await waitFor(() => {
-      expect(client.post).toHaveBeenCalledWith('/api/session', {
-        title: 'New Session',
+    act(() => {
+      sessionStatusSSEOptionsRef.current?.onEvent({
+        type: 'session.status',
+        properties: {
+          sessionID: session.id,
+          status: { type: 'idle' },
+        },
+      });
+    });
+
+    expect(screen.getByRole('status', { name: 'chat.tool.running' }))
+      .toHaveAttribute('data-session-running', secondSession.id);
+
+    act(() => {
+      sessionStatusSSEOptionsRef.current?.onEvent({
+        type: 'session.status',
+        properties: {
+          sessionID: session.id,
+          status: { type: 'retry' },
+        },
       });
     });
+
+    expect(screen.getAllByRole('status', { name: 'chat.tool.running' }))
+      .toHaveLength(2);
+  });
+
+  it('shows load more as text without an idle arrow', async () => {
+    useSessions.mockReturnValue({
+      sessions: [session],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+      hasMoreByProject: { tasks: true },
+      loadingMoreProjectIds: new Set(),
+      loadMore: vi.fn(),
+    });
+
+    renderSessionPage();
+
+    const loadMoreButton = await screen.findByRole('button', { name: 'loadMore' });
+    expect(loadMoreButton.querySelector('svg')).toBeNull();
+  });
+
+  it('collapses loaded task pages and reopens cached tasks without another request', async () => {
+    const user = userEvent.setup();
+    const loadMore = vi.fn();
+    const loadedTasks = Array.from({ length: 8 }, (_, index) => ({
+      ...session,
+      id: `task-${index + 1}`,
+      slug: `task-${index + 1}`,
+      title: `Task ${index + 1}`,
+    }));
+    client.get.mockResolvedValue({
+      data: [
+        {
+          id: 'default',
+          worktree: '/tmp/project',
+          name: '默认',
+          isDefault: true,
+          pathStatus: 'available',
+          sessionCount: 8,
+        },
+        {
+          id: 'prj_labs',
+          worktree: '/tmp/labs',
+          name: 'Labs',
+          isDefault: false,
+          pathStatus: 'available',
+          sessionCount: 0,
+        },
+      ],
+    });
+    useSessions.mockReturnValue({
+      sessions: loadedTasks,
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+      hasMoreByProject: { tasks: false },
+      loadingMoreProjectIds: new Set(),
+      loadMore,
+    });
+
+    renderSessionPage();
+
+    expect(await screen.findByText('Task 8')).toBeInTheDocument();
+    await user.click(screen.getByRole('button', { name: 'collapseLoaded' }));
+
+    expect(screen.queryByText('Task 7')).not.toBeInTheDocument();
+    expect(screen.queryByText('Task 8')).not.toBeInTheDocument();
+
+    await user.click(screen.getByRole('button', { name: 'loadMore' }));
+
+    expect(screen.getByText('Task 7')).toBeInTheDocument();
+    expect(screen.getByText('Task 8')).toBeInTheDocument();
+    expect(loadMore).not.toHaveBeenCalled();
   });
 
   it('collapses the projects section and restores it after remounting', async () => {
@@ -398,6 +784,9 @@ describe('SessionPage session actions menu', () => {
     const firstRender = renderSessionPage();
 
     await screen.findByText('Labs');
+    const projectButton = screen.getByRole('button', { name: 'selectProject' });
+    expect(projectButton.querySelector('svg')).toHaveClass('h-3.5', 'w-3.5');
+    expect(projectButton.parentElement).toHaveClass('px-3');
     await user.click(screen.getByRole('button', { name: 'toggleProjects' }));
     expect(screen.queryByText('Labs')).not.toBeInTheDocument();
 
@@ -518,7 +907,7 @@ describe('SessionPage session actions menu', () => {
     expect(screen.queryByText('Original Session')).not.toBeInTheDocument();
   });
 
-  it('renders project sessions with the same card outline as task sessions', async () => {
+  it('renders project sessions with the compact conversation row treatment', async () => {
     client.get.mockResolvedValue({
       data: [
         { id: 'default', worktree: '/tmp/project', name: '默认', isDefault: true },
@@ -546,7 +935,7 @@ describe('SessionPage session actions menu', () => {
     const sessionTitle = await screen.findByText('Original Session');
     const sessionCard = sessionTitle.closest('[class*="cursor-pointer"]');
     expect(sessionCard).not.toBeNull();
-    expect(sessionCard).toHaveClass('border-gray-100');
+    expect(sessionCard).toHaveClass('min-h-[34px]', 'rounded-lg', 'border-transparent');
   });
 
   it('groups legacy sessions by the effective project returned by the backend', async () => {
@@ -616,7 +1005,7 @@ describe('SessionPage session actions menu', () => {
 
     await waitFor(() => {
       expect(client.post).toHaveBeenCalledWith('/api/project', { name: 'Labs', worktree: '/tmp/labs' });
-      expect(screen.getByText('Labs')).toBeInTheDocument();
+      expect(within(screen.getByLabelText('managementTitle')).getByText('Labs')).toBeInTheDocument();
     });
   });
 
@@ -823,7 +1212,7 @@ describe('SessionPage session actions menu', () => {
     await user.type(folderInput, '/tmp/labs');
     await user.click(screen.getByRole('button', { name: 'save' }));
 
-    expect(await screen.findByText('Labs')).toBeInTheDocument();
+    expect(await within(screen.getByLabelText('managementTitle')).findByText('Labs')).toBeInTheDocument();
     expect(screen.getByText('noProjectSessions')).toBeInTheDocument();
   });
 
@@ -853,6 +1242,7 @@ describe('SessionPage session actions menu', () => {
     const projectLabel = await screen.findByText('Labs');
     const projectRow = projectLabel.closest('[class*="group/project"]');
     expect(projectRow).not.toBeNull();
+    expect(projectRow?.firstElementChild).toHaveClass('text-sm');
     await user.click(within(projectRow as HTMLElement).getByRole('button', { name: 'projectActions' }));
     await user.click(within(projectRow as HTMLElement).getByRole('menuitem', { name: 'projectDialog.renameAction' }));
     const input = screen.getByLabelText('projectDialog.nameLabel');
@@ -864,7 +1254,8 @@ describe('SessionPage session actions menu', () => {
       expect(client.patch).toHaveBeenCalledWith('/api/project/prj_project2', { name: 'Renamed Project' });
     });
     const renamedProject = await screen.findByText('Renamed Project');
-    expect(renamedProject.closest('[class*="group/project"]')).toHaveTextContent('8');
+    expect(renamedProject.closest('[class*="group/project"]')).not.toHaveTextContent('8');
+    expect(screen.getByText('projectsSection').nextElementSibling).toHaveTextContent('(1)');
   });
 
   it('shares and unshares a project from the sidebar', async () => {
@@ -928,7 +1319,7 @@ describe('SessionPage session actions menu', () => {
 
     const projectRow = (await screen.findByText('Shared Labs')).closest('[class*="group/project"]');
     expect(projectRow).not.toBeNull();
-    expect(within(projectRow as HTMLElement).getByRole('button', { name: 'createSessionInProject' })).toBeDisabled();
+    expect(within(projectRow as HTMLElement).queryByRole('button', { name: 'createSessionInProject' })).not.toBeInTheDocument();
     await user.click(within(projectRow as HTMLElement).getByRole('button', { name: 'projectActions' }));
     expect(within(projectRow as HTMLElement).getByRole('menuitem', { name: 'projectDialog.copyPathAction' })).toBeInTheDocument();
     expect(within(projectRow as HTMLElement).queryByRole('menuitem', { name: 'shareAction' })).not.toBeInTheDocument();
@@ -1021,38 +1412,18 @@ describe('SessionPage session actions menu', () => {
     });
   });
 
-  it('creates a session from a specific project row', async () => {
-    const user = userEvent.setup();
+  it('does not show a create-session button on a project row', async () => {
     const currentProject = { id: 'default', worktree: '/tmp/project', name: '默认', isDefault: true };
     client.get.mockResolvedValue({
       data: [currentProject, { id: 'prj_project2', worktree: '/tmp/labs', name: 'Labs' }],
     });
-    client.post.mockResolvedValue({
-      data: {
-        ...secondSession,
-        id: 'session-labs',
-        projectID: 'prj_project2',
-        title: 'New Session',
-      },
-    });
 
     renderSessionPage();
 
     const projectLabel = await screen.findByText('Labs');
     const projectRow = projectLabel.closest('[class*="group/project"]');
     expect(projectRow).not.toBeNull();
-    await user.click(within(projectRow as HTMLElement).getByRole('button', { name: 'createSessionInProject' }));
-
-    await waitFor(() => {
-      expect(client.post).toHaveBeenCalledWith('/api/session', {
-        title: 'New Session',
-        projectID: 'prj_project2',
-      });
-    });
-    expect(addSession).toHaveBeenCalledWith(expect.objectContaining({
-      id: 'session-labs',
-      projectID: 'prj_project2',
-    }));
+    expect(within(projectRow as HTMLElement).queryByRole('button', { name: 'createSessionInProject' })).not.toBeInTheDocument();
   });
 
   it('opens the actions menu for a session item', async () => {
@@ -1064,68 +1435,249 @@ describe('SessionPage session actions menu', () => {
     await user.click(screen.getByRole('button', { name: 'moreActions' }));
 
     const menu = document.querySelector('[data-session-menu-portal]');
-    expect(menu).toHaveClass('min-w-28', 'rounded-md', 'p-1');
+    expect(menu).toHaveClass('w-[132px]', 'rounded-[10px]', 'p-1');
     expect(menu).not.toHaveClass('w-36', 'rounded-lg');
     expect(screen.getByRole('button', { name: 'rename' })).toBeInTheDocument();
     expect(screen.getByRole('button', { name: 'downloadJson' })).toBeInTheDocument();
-    expect(screen.getByRole('button', { name: 'deleteAction' })).toBeInTheDocument();
+    expect(screen.getByRole('button', { name: 'moveToProjectAction' })).toBeInTheDocument();
+    expect(screen.getByRole('button', { name: 'archiveAction' })).toBeInTheDocument();
   });
 
-  it('renames a session inline from the actions menu', async () => {
+  it('moves a regular task into a selected project', async () => {
     const user = userEvent.setup();
+    client.get.mockResolvedValue({
+      data: [
+        { id: 'prj_labs', worktree: '/tmp/labs', name: 'Labs', canWrite: true, pathStatus: 'available' },
+      ],
+    });
 
     renderSessionPage();
 
     await screen.findByText('Original Session');
     await user.click(screen.getByRole('button', { name: 'moreActions' }));
-    await user.click(screen.getByRole('button', { name: 'rename' }));
-
-    const input = screen.getByRole('textbox', { name: 'rename' });
-    await user.clear(input);
-    await user.type(input, 'Renamed Session{enter}');
+    await user.click(screen.getByRole('button', { name: 'moveToProjectAction' }));
+    await user.click(screen.getByRole('menuitem', { name: 'Labs' }));
 
     await waitFor(() => {
-      expect(sessionApi.update).toHaveBeenCalledWith('session-1', { title: 'Renamed Session' });
+      expect(sessionApi.moveToProject).toHaveBeenCalledWith('session-1', 'prj_labs');
     });
-    expect(updateSessionTitle).toHaveBeenCalledWith('session-1', 'Renamed Session');
-    expect(sessionApi.update).toHaveBeenCalledTimes(1);
+    expect(refetchSessions).toHaveBeenCalled();
+    expect(toast.success).toHaveBeenCalledWith('moveToProjectSuccess');
   });
 
-  it('coalesces bursty session.updated events into one sidebar refetch', () => {
-    vi.useFakeTimers();
-    try {
-      renderSessionPage();
-
-      const emitSessionUpdated = screen.getByRole('button', { name: 'mock-session-updated' });
-      act(() => {
-        emitSessionUpdated.click();
-        emitSessionUpdated.click();
-      });
-
-      expect(updateSessionTitle).toHaveBeenCalledTimes(2);
-      expect(refetchSessions).not.toHaveBeenCalled();
+  it('keeps the project picker inside the viewport near the bottom edge', async () => {
+    const user = userEvent.setup();
+    client.get.mockResolvedValue({
+      data: [
+        { id: 'prj_labs', worktree: '/tmp/labs', name: 'Labs', canWrite: true, pathStatus: 'available' },
+      ],
+    });
+    const originalInnerHeight = window.innerHeight;
+    Object.defineProperty(window, 'innerHeight', { configurable: true, value: 400 });
 
-      act(() => {
-        vi.advanceTimersByTime(499);
-      });
-      expect(refetchSessions).not.toHaveBeenCalled();
+    renderSessionPage();
 
-      act(() => {
-        vi.advanceTimersByTime(1);
-      });
-      expect(refetchSessions).toHaveBeenCalledTimes(1);
-    } finally {
-      vi.useRealTimers();
-    }
+    await screen.findByText('Original Session');
+    const actionsTrigger = screen.getByRole('button', { name: 'moreActions' });
+    vi.spyOn(actionsTrigger, 'getBoundingClientRect').mockReturnValue({
+      bottom: 380,
+      right: 300,
+    } as DOMRect);
+    await user.click(actionsTrigger);
+    await user.click(screen.getByRole('button', { name: 'moveToProjectAction' }));
+
+    const menu = document.querySelector('[data-session-menu-portal]') as HTMLElement;
+    expect(menu.style.top).toBe('300px');
+    expect(menu.style.maxHeight).toBe('calc(100vh - 16px)');
+
+    Object.defineProperty(window, 'innerHeight', {
+      configurable: true,
+      value: originalInnerHeight,
+    });
   });
 
-  it('downloads session data as CLI-compatible JSON', async () => {
+  it('shows the backend detail when moving a task fails', async () => {
     const user = userEvent.setup();
-    const OriginalBlob = Blob;
-    const originalCreateElement = document.createElement.bind(document);
-    let createdAnchor: HTMLAnchorElement | null = null;
-    let blobArg: Blob | null = null;
-    let blobParts: BlobPart[] = [];
+    client.get.mockResolvedValue({
+      data: [
+        { id: 'prj_labs', worktree: '/tmp/labs', name: 'Labs', canWrite: true, pathStatus: 'available' },
+      ],
+    });
+    sessionApi.moveToProject.mockRejectedValue({
+      message: 'Request failed with status code 409',
+      response: { data: { detail: '任务正在运行，请稍后再移动' } },
+    });
+
+    renderSessionPage();
+
+    await screen.findByText('Original Session');
+    await user.click(screen.getByRole('button', { name: 'moreActions' }));
+    await user.click(screen.getByRole('button', { name: 'moveToProjectAction' }));
+    await user.click(screen.getByRole('menuitem', { name: 'Labs' }));
+
+    await waitFor(() => {
+      expect(toast.error).toHaveBeenCalledWith(
+        'moveToProjectFailed',
+        '任务正在运行，请稍后再移动',
+      );
+    });
+  });
+
+  it('shows the move spinner only on the selected target project', async () => {
+    const user = userEvent.setup();
+    client.get.mockResolvedValue({
+      data: [
+        { id: 'prj_labs', worktree: '/tmp/labs', name: 'Labs', canWrite: true, pathStatus: 'available' },
+        { id: 'prj_docs', worktree: '/tmp/docs', name: 'Docs', canWrite: true, pathStatus: 'available' },
+      ],
+    });
+    sessionApi.moveToProject.mockImplementation(() => new Promise(() => {}));
+
+    renderSessionPage();
+
+    await screen.findByText('Original Session');
+    await user.click(screen.getByRole('button', { name: 'moreActions' }));
+    await user.click(screen.getByRole('button', { name: 'moveToProjectAction' }));
+    const labsTarget = screen.getByRole('menuitem', { name: 'Labs' });
+    const docsTarget = screen.getByRole('menuitem', { name: 'Docs' });
+    await user.click(labsTarget);
+
+    await waitFor(() => {
+      expect(labsTarget.querySelector('.animate-spin')).not.toBeNull();
+    });
+    expect(docsTarget.querySelector('.animate-spin')).toBeNull();
+  });
+
+  it('moves a project task to a different project and marks its current project', async () => {
+    const user = userEvent.setup();
+    client.get.mockResolvedValue({
+      data: [
+        { id: 'prj_source', worktree: '/tmp/source', name: 'Source', canWrite: true, pathStatus: 'available' },
+        { id: 'prj_target', worktree: '/tmp/target', name: 'Target', canWrite: true, pathStatus: 'available' },
+      ],
+    });
+    useSessions.mockReturnValue({
+      sessions: [{
+        ...session,
+        projectID: 'prj_source',
+        effectiveProjectID: 'prj_source',
+        directory: '/tmp/source',
+      }],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+
+    renderSessionPage();
+
+    await screen.findByText('Original Session');
+    await user.click(screen.getByRole('button', { name: 'moreActions' }));
+    await user.click(screen.getByRole('button', { name: 'moveToProjectAction' }));
+    expect(screen.getByRole('menuitem', { name: 'Source' })).toBeDisabled();
+    await user.click(screen.getByRole('menuitem', { name: 'Target' }));
+
+    await waitFor(() => {
+      expect(sessionApi.moveToProject).toHaveBeenCalledWith('session-1', 'prj_target');
+    });
+  });
+
+  it('shows a compact relative session timestamp and keeps the actions trigger background-free', async () => {
+    const user = userEvent.setup();
+
+    renderSessionPage();
+
+    const timestamp = await screen.findByText('17小时前');
+    const actionsTrigger = screen.getByRole('button', { name: 'moreActions' });
+
+    expect(timestamp).not.toHaveClass('group-hover:opacity-0');
+    expect(timestamp).toHaveClass('text-zinc-500');
+    expect(timestamp).toHaveAttribute('title', 'formatted-date');
+    expect(actionsTrigger).not.toHaveClass('hover:bg-white/80');
+
+    await user.click(actionsTrigger);
+
+    expect(actionsTrigger).not.toHaveClass('bg-white/80');
+  });
+
+  it('refreshes relative session timestamps every minute', () => {
+    vi.useFakeTimers();
+    let relativeTimeLabel = '17小时前';
+    vi.mocked(formatRelativeTime).mockImplementation(() => relativeTimeLabel);
+    try {
+      renderSessionPage();
+      expect(screen.getByText('17小时前')).toBeInTheDocument();
+
+      relativeTimeLabel = '18小时前';
+      act(() => {
+        vi.advanceTimersByTime(60_000);
+      });
+
+      expect(screen.getByText('18小时前')).toBeInTheDocument();
+    } finally {
+      vi.mocked(formatRelativeTime).mockImplementation(() => '17小时前');
+      vi.useRealTimers();
+    }
+  });
+
+  it('renames a session inline from the actions menu', async () => {
+    const user = userEvent.setup();
+
+    renderSessionPage();
+
+    await screen.findByText('Original Session');
+    await user.click(screen.getByRole('button', { name: 'moreActions' }));
+    await user.click(screen.getByRole('button', { name: 'rename' }));
+
+    const input = screen.getByRole('textbox', { name: 'rename' });
+    await user.clear(input);
+    await user.type(input, 'Renamed Session{enter}');
+
+    await waitFor(() => {
+      expect(sessionApi.update).toHaveBeenCalledWith('session-1', { title: 'Renamed Session' });
+    });
+    expect(updateSessionTitle).toHaveBeenCalledWith('session-1', 'Renamed Session');
+    expect(sessionApi.update).toHaveBeenCalledTimes(1);
+  });
+
+  it('coalesces bursty session.updated events into one sidebar refetch', () => {
+    vi.useFakeTimers();
+    try {
+      renderSessionPage();
+
+      const emitSessionUpdated = screen.getByRole('button', { name: 'mock-session-updated' });
+      act(() => {
+        emitSessionUpdated.click();
+        emitSessionUpdated.click();
+      });
+
+      expect(updateSessionTitle).toHaveBeenCalledTimes(2);
+      expect(refetchSessions).not.toHaveBeenCalled();
+
+      act(() => {
+        vi.advanceTimersByTime(499);
+      });
+      expect(refetchSessions).not.toHaveBeenCalled();
+
+      act(() => {
+        vi.advanceTimersByTime(1);
+      });
+      expect(refetchSessions).toHaveBeenCalledTimes(1);
+    } finally {
+      vi.useRealTimers();
+    }
+  });
+
+  it('downloads session data as CLI-compatible JSON', async () => {
+    const user = userEvent.setup();
+    const OriginalBlob = Blob;
+    const originalCreateElement = document.createElement.bind(document);
+    let createdAnchor: HTMLAnchorElement | null = null;
+    let blobArg: Blob | null = null;
+    let blobParts: BlobPart[] = [];
 
     class BlobMock extends OriginalBlob {
       constructor(parts: BlobPart[], options?: BlobPropertyBag) {
@@ -1190,20 +1742,35 @@ describe('SessionPage session actions menu', () => {
     vi.stubGlobal('Blob', OriginalBlob);
   });
 
-  it('deletes a session from the actions menu', async () => {
+  it('archives a session from the actions menu', async () => {
     const user = userEvent.setup();
 
     renderSessionPage();
 
     await screen.findByText('Original Session');
     await user.click(screen.getByRole('button', { name: 'moreActions' }));
-    await user.click(screen.getByRole('button', { name: 'deleteAction' }));
+    await user.click(screen.getByRole('button', { name: 'archiveAction' }));
 
     await waitFor(() => {
-      expect(sessionApi.delete).toHaveBeenCalledWith('session-1');
+      expect(sessionApi.archive).toHaveBeenCalledWith('session-1');
     });
     expect(removeSession).toHaveBeenCalledWith('session-1');
-    expect(global.confirm).toHaveBeenCalledWith('confirmDelete');
+    expect(global.confirm).toHaveBeenCalledWith('confirmArchive');
+  });
+
+  it('keeps a successful archive successful when project-count refresh fails', async () => {
+    const user = userEvent.setup();
+    renderSessionPage();
+
+    await screen.findByText('Original Session');
+    await waitFor(() => expect(client.get).toHaveBeenCalled());
+    client.get.mockRejectedValueOnce(new Error('project refresh failed'));
+    await user.click(screen.getByRole('button', { name: 'moreActions' }));
+    await user.click(screen.getByRole('button', { name: 'archiveAction' }));
+
+    await waitFor(() => expect(toast.success).toHaveBeenCalledWith('archiveSuccess'));
+    expect(removeSession).toHaveBeenCalledWith('session-1');
+    expect(toast.error).not.toHaveBeenCalledWith('archiveFailed', expect.anything());
   });
 
   it('does not auto-attach any session on first load without history', () => {
@@ -1500,7 +2067,8 @@ describe('SessionPage session actions menu', () => {
       expect(sessionApi.get).toHaveBeenCalledWith('session-missing-from-list');
     });
     expect(screen.queryByTestId('session-chat')).not.toBeInTheDocument();
-    expect(screen.getByText('loading-spinner')).toBeInTheDocument();
+    expect(screen.getByTestId('session-chat-skeleton')).not.toHaveClass('animate-pulse');
+    expect(screen.getByTestId('workbench-refresh-status')).toHaveTextContent('restoringTask');
 
     await act(async () => {
       request.resolve(fetchedSession);
@@ -1609,14 +2177,14 @@ describe('SessionPage session actions menu', () => {
 
     expect(screen.getByTestId('session-chat')).toHaveAttribute('data-mention-agents', 'rex,explore');
 
-    await user.click(screen.getByRole('button', { name: /Rex/i }));
+    await user.click(screen.getByRole('button', { name: 'chat.addMenu.agent' }));
 
     expect(screen.getByRole('button', { name: /Explore/i })).toBeInTheDocument();
     expect(screen.queryByRole('button', { name: /hidden-system/i })).not.toBeInTheDocument();
     expect(screen.queryByRole('button', { name: /Oracle/i })).not.toBeInTheDocument();
   });
 
-  it('resets the chat agent to Rex when creating a new session', async () => {
+  it('opens a blank new-session draft without creating and resets the agent to Rex', async () => {
     const user = userEvent.setup();
     useAgents.mockReturnValue({
       agents: [
@@ -1645,18 +2213,23 @@ describe('SessionPage session actions menu', () => {
       refetch: vi.fn(),
     });
 
-    renderSessionPage();
+    renderSessionPage('/sessions?session=session-1');
 
-    await user.click(screen.getByRole('button', { name: /Rex/i }));
+    await waitFor(() => {
+      expect(screen.getByTestId('session-chat')).toHaveTextContent('session-1');
+    });
+    await user.click(screen.getByRole('button', { name: 'chat.addMenu.agent' }));
     await user.click(screen.getByRole('button', { name: /Explore/i }));
-    expect(screen.getByRole('button', { name: /Explore/i })).toBeInTheDocument();
+    expect(screen.getByTestId('mock-chat-input')).toHaveTextContent('subagent:explore');
+    expect(screen.getByTestId('session-chat')).toHaveAttribute('data-agent-name', 'rex');
 
     await user.click(screen.getByRole('button', { name: 'newSession' }));
 
-    await waitFor(() => {
-      expect(addSession).toHaveBeenCalledWith(secondSession);
-    });
-    expect(screen.getByRole('button', { name: /Rex/i })).toBeInTheDocument();
+    expect(screen.getByTestId('session-chat')).toHaveTextContent('no-session');
+    expect(client.post).not.toHaveBeenCalledWith('/api/session', expect.anything());
+    expect(screen.getByRole('button', { name: 'projectPicker.title' })).toHaveAttribute('aria-expanded', 'false');
+    expect(screen.queryByRole('menu', { name: 'projectPicker.title' })).not.toBeInTheDocument();
+    expect(screen.getByRole('button', { name: 'chat.addMenu.agent' })).toBeInTheDocument();
   });
 
   it('shows the pinned model for the selected session on load', async () => {
@@ -1690,7 +2263,7 @@ describe('SessionPage session actions menu', () => {
     await waitFor(() => {
       expect(screen.getByTestId('session-chat')).toHaveAttribute('data-model', 'minimax/minimax-m3');
     });
-    expect(defaultModelAPI.getResolved).not.toHaveBeenCalled();
+    expect(defaultModelAPI.getResolved).toHaveBeenCalledTimes(1);
   });
 
   it('persists model changes to the selected session', async () => {
@@ -1735,12 +2308,330 @@ describe('SessionPage session actions menu', () => {
         provider: 'minimax',
         model: 'minimax-m3',
         model_pinned: true,
+        model_auto: false,
       });
     });
     expect(refetchSessions).toHaveBeenCalled();
   });
 
-  it('resets the selected model to the default when creating a new session', async () => {
+  it('switches a pinned session to Auto without sending a synthetic model', async () => {
+    const user = userEvent.setup();
+    useSessions.mockReturnValue({
+      sessions: [{
+        ...session,
+        provider: 'minimax',
+        model: 'minimax-m3',
+        model_pinned: true,
+      }],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    useProviders.mockReturnValue({
+      providers: modelProviders,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
+    });
+    defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: 'openai', model_id: 'gpt-4o' } });
+    modelV2API.listDefinitions.mockResolvedValue({ data: { models: modelDefinitions } });
+
+    renderSessionPage('/sessions?session=session-1');
+
+    await waitFor(() => {
+      expect(screen.getByTestId('session-chat')).toHaveAttribute('data-model', 'minimax/minimax-m3');
+    });
+    await user.click(screen.getByRole('button', { name: /MiniMax M3/i }));
+    const autoButton = await screen.findByRole('button', { name: 'modelPicker.auto' });
+    await user.click(autoButton);
+
+    await waitFor(() => {
+      expect(sessionApi.update).toHaveBeenCalledWith('session-1', {
+        model_auto: true,
+        model_pinned: false,
+      });
+    });
+    expect(screen.getByTestId('session-chat')).toHaveAttribute('data-model', '');
+  });
+
+  it('does not offer Auto for a non-WebUI task session', async () => {
+    const user = userEvent.setup();
+    useSessions.mockReturnValue({
+      sessions: [{ ...session, category: 'task' }],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    useProviders.mockReturnValue({
+      providers: modelProviders,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
+    });
+    defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: 'openai', model_id: 'gpt-4o' } });
+    modelV2API.listDefinitions.mockResolvedValue({ data: { models: modelDefinitions } });
+
+    renderSessionPage('/sessions?session=session-1');
+
+    await user.click(await screen.findByRole('button', { name: /GPT-4o/i }));
+    const autoButton = await screen.findByRole('button', { name: 'modelPicker.auto' });
+    expect(autoButton).toBeDisabled();
+    expect(sessionApi.update).not.toHaveBeenCalled();
+  });
+
+  it('offers Auto for an entity configuration WebUI session', async () => {
+    const user = userEvent.setup();
+    useSessions.mockReturnValue({
+      sessions: [{ ...session, category: 'entity-config' }],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    useProviders.mockReturnValue({
+      providers: modelProviders,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
+    });
+    defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: 'openai', model_id: 'gpt-4o' } });
+    modelV2API.listDefinitions.mockResolvedValue({ data: { models: modelDefinitions } });
+
+    renderSessionPage('/sessions?session=session-1');
+
+    await user.click(await screen.findByRole('button', { name: /GPT-4o/i }));
+    const autoButton = await screen.findByRole('button', { name: 'modelPicker.auto' });
+    expect(autoButton).toBeEnabled();
+    await user.click(autoButton);
+    expect(sessionApi.update).toHaveBeenCalledWith('session-1', {
+      model_auto: true,
+      model_pinned: false,
+    });
+  });
+
+  it('does not offer Auto without a valid primary model', async () => {
+    const user = userEvent.setup();
+    useSessions.mockReturnValue({
+      sessions: [session],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    useProviders.mockReturnValue({
+      providers: modelProviders,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
+    });
+    defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: '', model_id: '' } });
+    modelV2API.listDefinitions.mockResolvedValue({ data: { models: modelDefinitions } });
+
+    renderSessionPage('/sessions?session=session-1');
+
+    await user.click(await screen.findByRole('button', { name: /GPT-4o/i }));
+    const autoButton = await screen.findByRole('button', { name: 'modelPicker.auto' });
+    expect(autoButton).toBeDisabled();
+    expect(sessionApi.update).not.toHaveBeenCalled();
+  });
+
+  it('keeps Auto available for a historical user session without a category field', async () => {
+    const user = userEvent.setup();
+    useSessions.mockReturnValue({
+      sessions: [{ ...session, category: undefined }],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    useProviders.mockReturnValue({
+      providers: modelProviders,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
+    });
+    defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: 'openai', model_id: 'gpt-4o' } });
+    modelV2API.listDefinitions.mockResolvedValue({ data: { models: modelDefinitions } });
+
+    renderSessionPage('/sessions?session=session-1');
+
+    await user.click(await screen.findByRole('button', { name: /GPT-4o/i }));
+    expect(await screen.findByRole('button', { name: 'modelPicker.auto' })).toBeEnabled();
+  });
+
+  it('uses primary model capabilities while Auto is selected', async () => {
+    useSessions.mockReturnValue({
+      sessions: [{ ...session, model_auto: true, model_pinned: false }],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    useProviders.mockReturnValue({
+      providers: modelProviders,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
+    });
+    defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: 'openai', model_id: 'gpt-4o' } });
+    modelV2API.listDefinitions.mockResolvedValue({
+      data: {
+        models: modelDefinitions.map((definition) => definition.id === 'gpt-4o'
+          ? {
+              ...definition,
+              capabilities: { supports_vision: true },
+              limits: { context_window: 200000 },
+            }
+          : {
+              ...definition,
+              capabilities: { supports_vision: false },
+              limits: { context_window: 32000 },
+            }),
+      },
+    });
+
+    renderSessionPage('/sessions?session=session-1');
+
+    await waitFor(() => {
+      expect(screen.getByTestId('session-chat')).toHaveAttribute('data-supports-vision', 'true');
+      expect(screen.getByTestId('session-chat')).toHaveAttribute('data-context-window', '200000');
+      expect(screen.getByTestId('session-chat')).toHaveAttribute('data-model', '');
+    });
+  });
+
+  it('turns Auto off when a concrete model is selected', async () => {
+    const user = userEvent.setup();
+    useSessions.mockReturnValue({
+      sessions: [{ ...session, model_auto: true, model_pinned: false }],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    useProviders.mockReturnValue({
+      providers: modelProviders,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
+    });
+    defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: 'openai', model_id: 'gpt-4o' } });
+    modelV2API.listDefinitions.mockResolvedValue({ data: { models: modelDefinitions } });
+
+    renderSessionPage('/sessions?session=session-1');
+
+    await user.click(await screen.findByRole('button', { name: /^modelPicker\.auto/i }));
+    await user.click(screen.getByRole('button', { name: /MiniMax M3/i }));
+
+    await waitFor(() => {
+      expect(sessionApi.update).toHaveBeenCalledWith('session-1', {
+        provider: 'minimax',
+        model: 'minimax-m3',
+        model_pinned: true,
+        model_auto: false,
+      });
+    });
+  });
+
+  it('keeps an existing Auto session selected with a valid primary model', async () => {
+    useSessions.mockReturnValue({
+      sessions: [{ ...session, model_auto: true, model_pinned: false }],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    useProviders.mockReturnValue({
+      providers: modelProviders,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
+    });
+    defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: 'openai', model_id: 'gpt-4o' } });
+    modelV2API.listDefinitions.mockResolvedValue({ data: { models: modelDefinitions } });
+
+    renderSessionPage('/sessions?session=session-1');
+
+    await waitFor(() => {
+      expect(screen.getByRole('button', { name: /^modelPicker\.auto/i })).toBeInTheDocument();
+      expect(screen.getByTestId('session-chat')).toHaveAttribute('data-model', '');
+    });
+  });
+
+  it('creates a blank-session Auto chat without a model override', async () => {
+    const user = userEvent.setup();
+    useSessions.mockReturnValue({
+      sessions: [],
+      loading: false,
+      error: null,
+      refetch: refetchSessions,
+      updateSessionTitle,
+      removeSession,
+      removeSessions,
+      addSession,
+    });
+    useProviders.mockReturnValue({
+      providers: modelProviders,
+      connectedIds: ['openai', 'minimax'],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
+    });
+    defaultModelAPI.getResolved.mockResolvedValue({ data: { provider_id: 'openai', model_id: 'gpt-4o' } });
+    modelV2API.listDefinitions.mockResolvedValue({ data: { models: modelDefinitions } });
+
+    renderSessionPage();
+
+    await user.click(await screen.findByRole('button', { name: /GPT-4o/i }));
+    await user.click(await screen.findByRole('button', { name: 'modelPicker.auto' }));
+    await user.click(screen.getByRole('button', { name: 'mock-create-and-send' }));
+
+    await waitFor(() => {
+      expect(client.post).toHaveBeenCalledWith('/api/session', {
+        title: 'New Session',
+        model_auto: true,
+      });
+      expect(client.post).toHaveBeenCalledWith(
+        '/api/session/session-2/prompt_async',
+        expect.not.objectContaining({ model: expect.anything() }),
+      );
+    });
+  });
+
+  it('resets the selected model to the default when starting a new session', async () => {
     const user = userEvent.setup();
     useSessions.mockReturnValue({
       sessions: [{
@@ -1776,12 +2667,42 @@ describe('SessionPage session actions menu', () => {
     await user.click(screen.getByRole('button', { name: 'newSession' }));
 
     await waitFor(() => {
-      expect(addSession).toHaveBeenCalledWith(secondSession);
+      expect(screen.getByTestId('session-chat')).toHaveTextContent('no-session');
       expect(screen.getByTestId('session-chat')).toHaveAttribute('data-model', 'openai/gpt-4o');
     });
+    expect(client.post).not.toHaveBeenCalledWith('/api/session', expect.anything());
+  });
+
+  it('lets the user choose a project before the first message creates the session', async () => {
+    const user = userEvent.setup();
+    client.get.mockResolvedValue({
+      data: [
+        { id: 'default', worktree: '/tmp/project', name: '默认', isDefault: true },
+        { id: 'prj_labs', worktree: '/tmp/labs', name: 'Labs', canWrite: true },
+      ],
+    });
+
+    renderSessionPage('/sessions?session=session-1');
+
+    await waitFor(() => {
+      expect(screen.getByTestId('session-chat')).toHaveTextContent('session-1');
+    });
+    await user.click(screen.getByRole('button', { name: 'newSession' }));
+
+    expect(client.post).not.toHaveBeenCalledWith('/api/session', expect.anything());
+    await user.click(screen.getByRole('button', { name: 'projectPicker.title' }));
+    await user.click(screen.getByRole('menuitemradio', { name: 'Labs' }));
+    await user.click(screen.getByRole('button', { name: 'mock-create-and-send' }));
+
+    await waitFor(() => {
+      expect(client.post).toHaveBeenCalledWith('/api/session', {
+        title: 'New Session',
+        projectID: 'prj_labs',
+      });
+    });
   });
 
-  it('uses the selected agent for the first message when an empty session is created by sending', async () => {
+  it('keeps Rex as the default while selecting a one-turn subagent reference', async () => {
     const user = userEvent.setup();
     useAgents.mockReturnValue({
       agents: [
@@ -1812,17 +2733,101 @@ describe('SessionPage session actions menu', () => {
 
     renderSessionPage();
 
-    await user.click(screen.getByRole('button', { name: /Rex/i }));
+    await user.click(screen.getByRole('button', { name: 'chat.addMenu.agent' }));
+    expect(screen.queryByRole('button', { name: /Rex/i })).not.toBeInTheDocument();
     await user.click(screen.getByRole('button', { name: /Explore/i }));
-    expect(screen.getByRole('button', { name: /Explore/i })).toBeInTheDocument();
-    await user.click(screen.getByRole('button', { name: 'mock-create-and-send' }));
+    expect(screen.getByTestId('mock-chat-input')).toHaveTextContent('subagent:explore');
+    expect(screen.getByTestId('session-chat')).toHaveAttribute('data-agent-name', 'rex');
+  });
 
-    await waitFor(() => {
-      expect(client.post).toHaveBeenCalledWith(
-        '/api/session/session-2/prompt_async',
-        expect.objectContaining({ agent: 'explore' }),
-      );
+  it('inserts the selected agent as a structured subagent reference', async () => {
+    const user = userEvent.setup();
+    useAgents.mockReturnValue({
+      agents: [
+        {
+          name: 'rex',
+          description: 'Rex',
+          mode: 'primary',
+          permission: [],
+          options: {},
+          skills: [],
+          tools: [],
+        },
+        {
+          name: 'explore',
+          description: 'Explore',
+          mode: 'subagent',
+          native: true,
+          permission: [],
+          options: {},
+          skills: [],
+          tools: [],
+        },
+      ],
+      loading: false,
+      error: null,
+      refetch: vi.fn(),
     });
-    expect(screen.getByRole('button', { name: /Explore/i })).toBeInTheDocument();
+
+    renderSessionPage();
+
+    await user.click(screen.getByRole('button', { name: 'chat.addMenu.agent' }));
+    await user.click(screen.getByRole('button', { name: /Explore/i }));
+
+    expect(screen.getByTestId('mock-chat-input')).toHaveTextContent('subagent:explore');
+  });
+
+  it('inserts selected workflows and skills as composer references', async () => {
+    const user = userEvent.setup();
+    workflowAPI.listSummaries.mockResolvedValue({
+      data: [{
+        id: 'alert-triage',
+        name: 'Alert triage',
+        description: 'Investigate security alerts',
+        category: 'security',
+        status: 'active',
+        source: 'project',
+        createdAt: 1,
+        updatedAt: 1,
+        nodeCount: 2,
+        stats: {
+          callCount: 0,
+          successCount: 0,
+          errorCount: 0,
+          totalRuntime: 0,
+          avgRuntime: 0,
+          thumbsUp: 0,
+          thumbsDown: 0,
+        },
+      }],
+    });
+    skillAPI.status.mockResolvedValue({
+      data: [{
+        name: 'diagnose',
+        description: 'Debug difficult failures',
+        location: '/skills/diagnose',
+        source: 'project',
+        eligible: true,
+      }],
+    });
+
+    renderSessionPage();
+
+    await user.click(screen.getByRole('button', { name: 'mock-open-add-menu' }));
+    const skillButton = screen.getByRole('button', { name: 'chat.addMenu.skills' });
+    const workflowButton = screen.getByRole('button', { name: 'chat.addMenu.workflows' });
+    const menuButtons = screen.getAllByRole('button');
+    expect(menuButtons.indexOf(skillButton)).toBeLessThan(menuButtons.indexOf(workflowButton));
+
+    await user.click(workflowButton);
+    expect(screen.queryByText('security')).not.toBeInTheDocument();
+    await user.click(await screen.findByRole('button', { name: /Alert triage/i }));
+    expect(screen.getByTestId('mock-chat-input')).toHaveTextContent('workflow:alert-triage');
+
+    await user.click(screen.getByRole('button', { name: 'mock-open-add-menu' }));
+    await user.click(screen.getByRole('button', { name: 'chat.addMenu.skills' }));
+    await user.click(await screen.findByRole('button', { name: /diagnose/i }));
+    expect(screen.getByTestId('mock-chat-input')).toHaveTextContent('skill:diagnose');
+    expect(screen.getByText('chat.addMenu.selectSkill')).toBeInTheDocument();
   });
 });
diff --git a/webui/src/pages/Session/index.tsx b/webui/src/pages/Session/index.tsx
index f9b27aba7..7999b42ce 100644
--- a/webui/src/pages/Session/index.tsx
+++ b/webui/src/pages/Session/index.tsx
@@ -1,18 +1,20 @@
 import { memo, useState, useEffect, useMemo, useCallback, useRef, type RefObject } from 'react';
 import {
-  Plus, Trash2,
-  ChevronDown, ChevronRight, Sparkles, Shield, Search, AlertTriangle,
+  Plus, Trash2, Archive,
+  ChevronDown, ChevronLeft, ChevronRight, Sparkles, Shield, Search, AlertTriangle,
   PanelLeftClose, PanelLeft, Bot, Loader2,
   Workflow as WorkflowIcon, Settings2, CheckSquare,
-  MoreHorizontal, PencilLine, Download, Share2, Cpu, Info,
-  FolderGit2, FolderPlus, FolderOpen, Copy, ArrowUp, HardDrive,
+  MoreHorizontal, PencilLine, Download, Share2, Cpu, Info, X, Check,
+  FolderGit2, FolderPlus, FolderOpen, Copy, ArrowUp, HardDrive, BookOpen,
+  Hammer, ClipboardList, Target,
 } from 'lucide-react';
 import { useTranslation } from 'react-i18next';
-import i18n from '@/i18n';
 import { useLocation, useNavigate, useSearchParams } from 'react-router-dom';
+import { getAnchoredMenuLeftOffset } from '@/components/common/ChatPromptSelectors';
 import LoadingSpinner from '@/components/common/LoadingSpinner';
 import { useToast } from '@/components/common/Toast';
 import SessionChat, { buildInstructionDisplayText, type PromptDisplayOptions, type SSEChatEvent, type SSEConnectionStatus } from '@/components/common/SessionChat';
+import { useSSE } from '@/hooks/useSSE';
 import SuiteInstallProgressPanel, {
   applySuiteInstallProgressEvent,
   createSuiteInstallProgressState,
@@ -21,18 +23,33 @@ import SuiteInstallProgressPanel, {
 } from '@/components/hub/SuiteInstallProgressPanel';
 import { sessionApi } from '@/api/session';
 import { hubAPI, type HubInstallProgressEvent } from '@/api/hub';
+import { skillAPI, type Skill } from '@/api/skill';
+import { workflowAPI, type WorkflowSummary } from '@/api/workflow';
 import type { Agent } from '@/api/agent';
 import { useSessions } from '@/hooks/useSessions';
 import { useAgents } from '@/hooks/useAgents';
 import { useProviders } from '@/hooks/useProviders';
-import { useEnabledChatModelDefinitions, useResolvedDefaultModel } from '@/hooks/useChatModelResources';
-import client from '@/api/client';
+import {
+  useEnabledChatModelDefinitions,
+  useResolvedDefaultModel,
+} from '@/hooks/useChatModelResources';
+import client, { getApiBase } from '@/api/client';
 import { useDefaultModelVision } from '@/hooks/useDefaultModelVision';
 import { buildPromptParts, type ImagePartData } from '@/utils/imageUpload';
 import { getAgentDisplayDescription, getAgentDisplayName, isAgentUsableInChat } from '@/utils/agentDisplay';
-import { formatSessionDate } from '@/utils/time';
+import { formatRelativeTime, formatSessionDate } from '@/utils/time';
+import { getWorkflowDisplayName } from '@/utils/workflowDisplay';
 import type { ModelDefinitionV2, Session } from '@/types';
 import { useAuth } from '@/contexts/AuthContext';
+import {
+  DEFAULT_SESSION_EXECUTION_MODE,
+  promoteDraftExecutionMode,
+  readSessionExecutionMode,
+  resetDraftExecutionMode,
+  writeSessionExecutionMode,
+  type PersistentSessionExecutionMode,
+  type SessionExecutionMode,
+} from '@/utils/sessionExecutionMode';
 
 function sanitizeSessionExportName(value: string) {
   const trimmed = value.trim();
@@ -48,8 +65,22 @@ const SESSION_PAGE_VISITED_STORAGE_KEY = 'flocks:sessions:visited';
 const SOC_WORKSPACE_COMPONENT_ID = 'soc-workspace';
 const INSTALLED_HUB_STATES = new Set(['installed', 'localOnly', 'updateAvailable']);
 const SESSION_UPDATE_REFETCH_DEBOUNCE_MS = 500;
+const AUTO_MODEL_KEY = '__flocks_auto__';
 const TASK_SESSION_GROUP_ID = 'tasks';
+const SESSION_EXECUTION_MODES: SessionExecutionMode[] = ['build', 'plan', 'goal'];
 type AgentSourceFilter = 'all' | 'builtin' | 'custom';
+
+function ExecutionModeIcon({
+  mode,
+  className = 'h-3 w-3',
+}: {
+  mode: SessionExecutionMode;
+  className?: string;
+}) {
+  if (mode === 'plan') return <ClipboardList className={className} />;
+  if (mode === 'goal') return <Target className={className} />;
+  return <Hammer className={className} />;
+}
 type ProjectSummary = {
   id: string;
   worktree: string;
@@ -84,8 +115,146 @@ type FolderBrowserResponse = {
   roots: FolderEntry[];
   entries: FolderEntry[];
 };
+
+type ComposerResourceItem = {
+  id: string;
+  name: string;
+  description?: string;
+  badge?: string;
+  source: 'builtin' | 'custom';
+};
+
+function ComposerResourcePicker({
+  label,
+  title,
+  emptyText,
+  icon: Icon,
+  items,
+  loading,
+  open,
+  onToggle,
+  onSelect,
+}: {
+  label: string;
+  title: string;
+  emptyText: string;
+  icon: React.ComponentType<{ className?: string }>;
+  items: ComposerResourceItem[];
+  loading: boolean;
+  open: boolean;
+  onToggle: () => void;
+  onSelect: (item: ComposerResourceItem) => void;
+}) {
+  const { t } = useTranslation('session');
+  const [sourceFilter, setSourceFilter] = useState<AgentSourceFilter>('all');
+  const filteredItems = items.filter((item) => (
+    sourceFilter === 'all' || item.source === sourceFilter
+  ));
+
+  return (
+    <div className="relative">
+      <button
+        type="button"
+        onClick={onToggle}
+        className="group flex h-10 w-full items-center gap-2.5 rounded-[9px] px-2 text-left text-[13px] font-medium text-zinc-700 transition-colors hover:bg-zinc-100/90 hover:text-zinc-950 dark:text-zinc-200 dark:hover:bg-white/[0.07] dark:hover:text-white"
+        aria-label={label}
+        aria-haspopup="menu"
+        aria-expanded={open}
+      >
+        <span className="grid h-7 w-7 shrink-0 place-items-center rounded-lg border border-zinc-200/80 bg-white text-zinc-500 shadow-[0_1px_2px_rgba(22,27,34,0.04)] transition-colors group-hover:text-zinc-800 dark:border-white/[0.10] dark:bg-white/[0.05] dark:text-zinc-400 dark:group-hover:text-white">
+          <Icon className="h-3.5 w-3.5" />
+        </span>
+        <span className="min-w-0 flex-1 truncate">{label}</span>
+        <ChevronRight className={`h-3.5 w-3.5 shrink-0 text-zinc-400 transition-transform ${open ? '-rotate-90' : ''}`} />
+      </button>
+      {open && (
+        <div className="absolute bottom-0 left-[calc(100%+0.5rem)] z-[60] w-[300px] max-w-[calc(100vw-2rem)] overflow-hidden rounded-[14px] border border-black/[0.09] bg-white/95 shadow-[0_18px_46px_rgba(22,27,34,0.14),0_2px_8px_rgba(22,27,34,0.06)] backdrop-blur-xl dark:border-white/[0.11] dark:bg-[#252c35]/95 dark:shadow-[0_20px_48px_rgba(8,10,13,0.44)]">
+          <div className="flex min-h-12 items-center justify-between gap-3 border-b border-zinc-100 px-3 py-2 dark:border-white/[0.08]">
+            <span className="min-w-0 truncate text-[12px] font-semibold text-zinc-800 dark:text-zinc-100">
+              {title}
+            </span>
+            <div className="inline-flex shrink-0 items-center rounded-lg bg-zinc-100 p-0.5 text-[10px] dark:bg-black/15">
+              {(['all', 'builtin', 'custom'] as AgentSourceFilter[]).map((filter) => (
+                <button
+                  key={filter}
+                  type="button"
+                  onClick={() => setSourceFilter(filter)}
+                  className={`rounded-md px-2 py-1 transition-colors ${
+                    sourceFilter === filter
+                      ? 'bg-white text-zinc-900 shadow-[0_1px_3px_rgba(22,27,34,0.08)] dark:bg-white/[0.12] dark:text-zinc-50'
+                      : 'text-zinc-500 hover:text-zinc-800 dark:text-zinc-400 dark:hover:text-zinc-100'
+                  }`}
+                >
+                  {t(`agentPicker.filter.${filter}`)}
+                </button>
+              ))}
+            </div>
+          </div>
+          <div className="h-64 space-y-0.5 overflow-y-auto p-1.5">
+            {loading ? (
+              <div className="flex h-full items-center justify-center text-zinc-400">
+                <Loader2 className="h-4 w-4 animate-spin" />
+              </div>
+            ) : filteredItems.length > 0 ? (
+              filteredItems.map((item) => (
+                <button
+                  key={item.id}
+                  type="button"
+                  onClick={() => onSelect(item)}
+                  title={item.description}
+                  className="flex w-full min-w-0 items-center gap-2 rounded-lg px-2.5 py-2 text-left text-zinc-700 transition-colors hover:bg-zinc-50 hover:text-zinc-950 dark:text-zinc-300 dark:hover:bg-white/[0.06] dark:hover:text-white"
+                >
+                  <span className="grid h-6 w-6 shrink-0 place-items-center rounded-md bg-zinc-100 text-zinc-400 dark:bg-white/[0.05] dark:text-zinc-500">
+                    <Icon className="h-3 w-3" />
+                  </span>
+                  <span className="min-w-0 flex-1 truncate text-xs font-medium">{item.name}</span>
+                  <span className={`shrink-0 rounded px-1.5 py-0.5 text-[9px] font-medium ${
+                    item.source === 'builtin'
+                      ? 'bg-zinc-100 text-zinc-600 dark:bg-white/[0.08] dark:text-zinc-300'
+                      : 'bg-teal-50 text-teal-600 dark:bg-teal-950/40 dark:text-teal-300'
+                  }`}>
+                    {t(`agentPicker.badge.${item.source}`)}
+                  </span>
+                  {item.badge && (
+                    <span className="max-w-20 shrink-0 truncate rounded bg-zinc-100 px-1.5 py-0.5 text-[9px] font-medium text-zinc-500 dark:bg-white/[0.07] dark:text-zinc-400">
+                      {item.badge}
+                    </span>
+                  )}
+                </button>
+              ))
+            ) : (
+              <div className="flex h-full items-center justify-center px-4 text-center text-xs text-zinc-400">
+                {emptyText}
+              </div>
+            )}
+          </div>
+        </div>
+      )}
+    </div>
+  );
+}
 const MULTI_PROJECT_SESSION_PAGE_SIZE = 6;
 const SINGLE_PROJECT_SESSION_PAGE_SIZE = 20;
+
+function readSessionStatusType(status: unknown): string | undefined {
+  if (typeof status === 'string') return status;
+  if (!status || typeof status !== 'object' || !('type' in status)) return undefined;
+  return typeof status.type === 'string' ? status.type : undefined;
+}
+
+function isRunningSessionStatus(status: unknown): boolean {
+  const statusType = readSessionStatusType(status);
+  return statusType === 'busy' || statusType === 'compacting' || statusType === 'retry';
+}
+
+function readRunningSessionIds(statuses: unknown): Set<string> {
+  if (!statuses || typeof statuses !== 'object' || Array.isArray(statuses)) return new Set();
+  return new Set(
+    Object.entries(statuses)
+      .filter(([, status]) => isRunningSessionStatus(status))
+      .map(([sessionId]) => sessionId),
+  );
+}
 type ChatModelOption = {
   key: string;
   providerID: string;
@@ -109,10 +278,6 @@ type SelectorTooltip = {
   y: number;
 };
 
-function formatAgentName(name: string): string {
-  return name ? name.charAt(0).toUpperCase() + name.slice(1) : name;
-}
-
 function readLastSelectedSessionId(): string | null {
   try {
     return window.localStorage.getItem(LAST_SELECTED_SESSION_STORAGE_KEY);
@@ -182,9 +347,12 @@ interface SessionSidebarItemProps {
   selectMode: boolean;
   checked: boolean;
   menuOpen: boolean;
+  running: boolean;
   renaming: boolean;
   renameValue: string;
   renameSubmitting: boolean;
+  language: string;
+  relativeTimeClock: number;
   t: (key: string, options?: Record<string, unknown>) => string;
   renameInputRef: RefObject<HTMLInputElement | null>;
   onSelect: (sessionId: string) => void;
@@ -202,9 +370,11 @@ function SessionSidebarItemInner({
   selectMode,
   checked,
   menuOpen,
+  running,
   renaming,
   renameValue,
   renameSubmitting,
+  language,
   t,
   renameInputRef,
   onSelect,
@@ -217,19 +387,17 @@ function SessionSidebarItemInner({
   return (
     <div
       onClick={() => onSelect(session.id)}
-      className={`group relative mb-1 cursor-pointer border px-3 transition-all duration-150 ${
-        nested ? 'ml-7 mr-2 rounded-lg py-2' : 'mx-2 rounded-xl py-2.5'
+      className={`group relative mb-0.5 min-h-[34px] cursor-pointer rounded-lg border px-3 py-1 transition-colors duration-100 ${
+        nested ? 'ml-5' : ''
       } ${
         !selectMode && selected
-          ? 'border-gray-300 bg-gray-100 shadow-sm dark:border-zinc-700 dark:bg-zinc-900 dark:shadow-none'
+          ? 'border-transparent bg-zinc-200/70 text-[#202328] dark:bg-[#3a434e] dark:text-white'
           : selectMode && checked
-          ? 'bg-blue-50 border-blue-200 dark:border-blue-500/40 dark:bg-blue-950/30'
-          : nested
-            ? 'border-gray-100 hover:border-gray-200 hover:bg-gray-50 dark:border-zinc-800 dark:hover:bg-zinc-900'
-            : 'border-gray-100 hover:border-gray-200 hover:bg-gray-50 hover:shadow-sm dark:border-transparent dark:hover:border-zinc-800 dark:hover:bg-zinc-900 dark:hover:shadow-none'
+          ? 'border-blue-200 bg-blue-50 text-[#202328] dark:border-blue-400/40 dark:bg-blue-500/10 dark:text-white'
+          : 'border-transparent text-[#5b6067] hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#b8c2cc] dark:hover:bg-white/[0.06] dark:hover:text-white'
       }`}
     >
-      <div className="flex items-center gap-1.5 min-w-0 pr-7">
+      <div className="flex min-w-0 items-center gap-1.5 pr-6">
         {selectMode && (
           <input
             type="checkbox"
@@ -241,12 +409,12 @@ function SessionSidebarItemInner({
         )}
         {session.category === 'workflow' && (
           <span title={t('workflowSession')} className="flex-shrink-0">
-            <WorkflowIcon className="w-3 h-3 text-orange-400" />
+            <WorkflowIcon className="h-3.5 w-3.5 text-orange-400" />
           </span>
         )}
         {session.category === 'entity-config' && (
           <span title={t('configSession')} className="flex-shrink-0">
-            <Settings2 className="w-3 h-3 text-purple-400" />
+            <Settings2 className="h-3.5 w-3.5 text-purple-400" />
           </span>
         )}
         {renaming ? (
@@ -262,28 +430,47 @@ function SessionSidebarItemInner({
             }}
             placeholder={t('renamePlaceholder')}
             disabled={renameSubmitting}
-            className="w-full min-w-0 rounded border border-blue-300 bg-white px-1.5 py-0.5 text-sm text-gray-900 outline-none focus:border-blue-400 dark:border-blue-500/50 dark:bg-zinc-950 dark:text-zinc-100"
+            className="h-6 w-full min-w-0 rounded-md border border-blue-300 bg-white px-1.5 text-sm text-gray-900 outline-none focus:border-blue-400 dark:border-blue-500/50 dark:bg-[#252c35] dark:text-white"
             aria-label={t('rename')}
             data-session-rename-input
           />
         ) : (
-          <h3 className="font-semibold text-gray-900 truncate text-sm flex items-center gap-1.5 dark:text-zinc-100">
+          <h3 className="flex min-w-0 flex-1 items-center gap-1.5 truncate text-sm font-medium">
             <span className="truncate">{session.title}</span>
             {session.isShared && (
-              <span className="inline-flex items-center rounded-full border border-blue-200 bg-blue-50 px-1.5 py-0.5 text-[10px] font-medium text-blue-700">
+              <span className="inline-flex shrink-0 items-center rounded-full border border-blue-200 bg-blue-50 px-1.5 py-0.5 text-[10px] font-medium text-blue-700 dark:border-blue-400/35 dark:bg-blue-500/10 dark:text-blue-200">
                 {t('sharedTag')}
               </span>
             )}
           </h3>
         )}
+        {session.time?.updated && !renaming && (
+          <time
+            dateTime={new Date(session.time.updated).toISOString()}
+            title={formatSessionDate(session.time.updated)}
+            className="ml-1 shrink-0 whitespace-nowrap text-[11px] font-normal tabular-nums text-zinc-500 dark:text-[#8f9ba8]"
+          >
+            {formatRelativeTime(session.time.updated, language)}
+          </time>
+        )}
       </div>
-      {session.time?.updated && !renaming && (
-        <p className="mt-1 text-xs text-gray-400 truncate pl-0.5 dark:text-zinc-500">
-          {formatSessionDate(session.time.updated)}
-        </p>
-      )}
       {!selectMode && (
-        <div className="absolute right-1.5 top-2" data-session-actions>
+        <div className="absolute right-1 top-1/2 h-6 w-6 -translate-y-1/2" data-session-actions>
+          {running && (
+            <span
+              role="status"
+              aria-label={t('chat.tool.running')}
+              title={t('chat.tool.running')}
+              data-session-running={session.id}
+              className={`pointer-events-none absolute inset-0 grid place-items-center text-[#5f8fcb] transition-opacity dark:text-[#8ab4e8] ${
+                menuOpen
+                  ? 'opacity-0'
+                  : 'opacity-100 group-hover:opacity-0 group-focus-within:opacity-0'
+              }`}
+            >
+              <Loader2 className="h-[14px] w-[14px] animate-spin" />
+            </span>
+          )}
           <button
             onClick={(e) => {
               e.stopPropagation();
@@ -292,11 +479,13 @@ function SessionSidebarItemInner({
             title={t('moreActions')}
             aria-label={t('moreActions')}
             aria-expanded={menuOpen}
-            className={`p-1 rounded text-gray-400 hover:text-gray-600 hover:bg-gray-200 transition-all dark:text-zinc-500 dark:hover:bg-zinc-800 dark:hover:text-zinc-200 ${
-              menuOpen ? 'opacity-100 text-gray-600 bg-gray-200 dark:bg-zinc-800 dark:text-zinc-200' : 'opacity-0 group-hover:opacity-100'
+            className={`absolute inset-0 grid h-6 w-6 place-items-center rounded-lg text-[#989ca1] transition-[color,opacity,background-color] hover:bg-white hover:text-[#202328] dark:text-[#8f9ba8] dark:hover:bg-white/[0.08] dark:hover:text-white ${
+              menuOpen || (selected && !running)
+                ? 'text-[#202328] opacity-100 dark:text-white'
+                : 'opacity-0 group-hover:opacity-100'
             }`}
           >
-            <MoreHorizontal className="w-3.5 h-3.5" />
+            <MoreHorizontal className="h-[15px] w-[15px]" />
           </button>
         </div>
       )}
@@ -315,12 +504,89 @@ const SessionSidebarItem = memo(SessionSidebarItemInner, (prev, next) => (
   prev.selectMode === next.selectMode &&
   prev.checked === next.checked &&
   prev.menuOpen === next.menuOpen &&
+  prev.running === next.running &&
   prev.renaming === next.renaming &&
   prev.renameValue === next.renameValue &&
   prev.renameSubmitting === next.renameSubmitting &&
+  prev.language === next.language &&
+  prev.relativeTimeClock === next.relativeTimeClock &&
   prev.t === next.t
 ));
 
+function WorkbenchRefreshStatus({ label }: { label: string }) {
+  return (
+    <div
+      role="status"
+      aria-live="polite"
+      data-testid="workbench-refresh-status"
+      className="ml-auto flex min-w-0 items-center gap-2 rounded-full border border-black/[0.06] bg-white/55 px-2.5 py-1 text-[11px] font-medium text-[#737980] shadow-[0_1px_2px_rgba(22,27,34,0.03)] dark:border-white/[0.08] dark:bg-white/[0.04] dark:text-[#aeb8c3]"
+    >
+      <span className="h-1.5 w-1.5 shrink-0 animate-pulse rounded-full bg-[#4f92e8]" />
+      <span className="truncate">{label}</span>
+      <span className="h-px w-8 shrink-0 overflow-hidden rounded-full bg-[#d9e5f4] dark:bg-[#46515e]">
+        <span className="block h-full w-1/2 animate-pulse rounded-full bg-[#4f92e8]" />
+      </span>
+    </div>
+  );
+}
+
+function SessionListSkeleton() {
+  return (
+    <div
+      data-testid="session-list-skeleton"
+      aria-hidden="true"
+      className="animate-pulse px-2 pt-1"
+    >
+      <div className="mb-2 h-3 w-16 rounded-full bg-black/[0.07] dark:bg-white/[0.08]" />
+      <div className="space-y-1.5">
+        {[72, 88, 64].map((width, index) => (
+          <div
+            key={`${width}-${index}`}
+            className="flex h-[34px] items-center gap-2 rounded-lg px-2"
+          >
+            <div className="h-3.5 w-3.5 shrink-0 rounded bg-black/[0.06] dark:bg-white/[0.07]" />
+            <div
+              className="h-3 rounded-full bg-black/[0.06] dark:bg-white/[0.07]"
+              style={{ width: `${width}%` }}
+            />
+          </div>
+        ))}
+      </div>
+      <div className="mb-2 mt-5 h-3 w-12 rounded-full bg-black/[0.07] dark:bg-white/[0.08]" />
+      <div className="space-y-1.5">
+        {[82, 68, 76, 58].map((width, index) => (
+          <div
+            key={`${width}-${index}`}
+            className="flex h-[34px] items-center gap-2 rounded-lg px-2"
+          >
+            <div className="h-3.5 w-3.5 shrink-0 rounded-full bg-black/[0.06] dark:bg-white/[0.07]" />
+            <div
+              className="h-3 rounded-full bg-black/[0.06] dark:bg-white/[0.07]"
+              style={{ width: `${width}%` }}
+            />
+          </div>
+        ))}
+      </div>
+    </div>
+  );
+}
+
+function SessionChatSkeleton() {
+  return (
+    <div
+      data-testid="session-chat-skeleton"
+      aria-hidden="true"
+      className="flex min-h-0 flex-1 justify-center overflow-hidden px-7 pt-16"
+    >
+      <LoadingSpinner
+        size="sm"
+        delayMs={180}
+        className="opacity-60 [&_svg]:text-zinc-400 dark:[&_svg]:text-zinc-500"
+      />
+    </div>
+  );
+}
+
 export default function SessionPage() {
   const { t, i18n } = useTranslation('session');
   const { user } = useAuth();
@@ -334,9 +600,27 @@ export default function SessionPage() {
   const [sidebarCollapsed, setSidebarCollapsed] = useState(false);
   const [selectedAgent, setSelectedAgent] = useState('rex');
   const [showAgentOptions, setShowAgentOptions] = useState(false);
+  const [showWorkflowOptions, setShowWorkflowOptions] = useState(false);
+  const [showSkillOptions, setShowSkillOptions] = useState(false);
+  const [composerWorkflows, setComposerWorkflows] = useState<WorkflowSummary[]>([]);
+  const [composerSkills, setComposerSkills] = useState<Skill[]>([]);
+  const [loadingComposerResources, setLoadingComposerResources] = useState(false);
+  const [selectedExecutionMode, setSelectedExecutionMode] = useState<SessionExecutionMode>(
+    () => readSessionExecutionMode(null),
+  );
+  const [showExecutionModeOptions, setShowExecutionModeOptions] = useState(false);
+  const executionModeHandoffRef = useRef<{
+    sessionId: string;
+    mode: SessionExecutionMode;
+  } | null>(null);
+  const [showProjectOptions, setShowProjectOptions] = useState(false);
   const [selectedModelKey, setSelectedModelKey] = useState<string | null>(null);
   const [showModelOptions, setShowModelOptions] = useState(false);
+  const modelSelectorRef = useRef<HTMLDivElement>(null);
+  const [modelMenuLeftOffset, setModelMenuLeftOffset] = useState(0);
   const [sseStatus, setSseStatus] = useState<SSEConnectionStatus>('disconnected');
+  const [runningSessionIds, setRunningSessionIds] = useState<Set<string>>(new Set());
+  const [relativeTimeClock, setRelativeTimeClock] = useState(0);
   const [creating, setCreating] = useState(false);
   const [installingSocWorkspace, setInstallingSocWorkspace] = useState(false);
   const [suiteInstallProgress, setSuiteInstallProgress] = useState<SuiteInstallProgressState | null>(null);
@@ -345,6 +629,7 @@ export default function SessionPage() {
   const [selectMode, setSelectMode] = useState(false);
   const [checkedIds, setCheckedIds] = useState<Set<string>>(new Set());
   const [projects, setProjects] = useState<ProjectSummary[]>([]);
+  const [refreshingProjects, setRefreshingProjects] = useState(true);
   const [selectedProjectId, setSelectedProjectId] = useState<string | null>(null);
   const [collapsedProjectIds, setCollapsedProjectIds] = useState<Set<string>>(() => {
     try {
@@ -354,6 +639,7 @@ export default function SessionPage() {
       return new Set();
     }
   });
+  const [collapsedLoadedSessionGroupIds, setCollapsedLoadedSessionGroupIds] = useState<Set<string>>(new Set());
   const [projectsSectionCollapsed, setProjectsSectionCollapsed] = useState(() => {
     try {
       return window.localStorage.getItem(projectsSectionCollapsedStorageKey) === 'true';
@@ -372,9 +658,12 @@ export default function SessionPage() {
   const [folderBrowser, setFolderBrowser] = useState<FolderBrowserResponse | null>(null);
   const [folderBrowserLoading, setFolderBrowserLoading] = useState(false);
   const [projectSubmitting, setProjectSubmitting] = useState(false);
-  const [batchDeleting, setBatchDeleting] = useState(false);
+  const [batchArchiving, setBatchArchiving] = useState(false);
   const [openMenuSessionId, setOpenMenuSessionId] = useState<string | null>(null);
   const [menuAnchor, setMenuAnchor] = useState<{ top: number; right: number } | null>(null);
+  const [movePickerSessionId, setMovePickerSessionId] = useState<string | null>(null);
+  const [movingSessionId, setMovingSessionId] = useState<string | null>(null);
+  const [movingTargetProjectId, setMovingTargetProjectId] = useState<string | null>(null);
   const [renamingSessionId, setRenamingSessionId] = useState<string | null>(null);
   const [renameValue, setRenameValue] = useState('');
   const [renameSubmitting, setRenameSubmitting] = useState(false);
@@ -384,22 +673,36 @@ export default function SessionPage() {
   const [agentSourceFilter, setAgentSourceFilter] = useState<AgentSourceFilter>('all');
   const [selectedSessionFallback, setSelectedSessionFallback] = useState<Session | null>(null);
   const [selectorTooltip, setSelectorTooltip] = useState<SelectorTooltip | null>(null);
+  const updateModelMenuLeftOffset = useCallback(() => {
+    const selector = modelSelectorRef.current;
+    if (!selector) return;
+    setModelMenuLeftOffset(getAnchoredMenuLeftOffset(
+      selector.getBoundingClientRect().left,
+      window.innerWidth,
+    ));
+  }, []);
   const renameInputRef = useRef<HTMLInputElement | null>(null);
   const renameSubmitInFlightRef = useRef(false);
   const projectSubmitInFlightRef = useRef(false);
   const folderBrowserRequestIdRef = useRef(0);
   const folderBrowserInputPathRef = useRef<string | null>(null);
   const sessionUpdateRefetchTimerRef = useRef<number | null>(null);
+  const sessionStatusEventVersionRef = useRef(0);
   const projectListRequestSeqRef = useRef(0);
+  const composerResourcesLoadedRef = useRef(false);
   const toast = useToast();
 
   const sessionProjectIds = useMemo(
     () => [TASK_SESSION_GROUP_ID, ...projects.map((project) => project.id)],
     [projects],
   );
+  const sessionListPageSize = projects.length >= 1
+    ? MULTI_PROJECT_SESSION_PAGE_SIZE
+    : SINGLE_PROJECT_SESSION_PAGE_SIZE;
   const {
     sessions,
     loading: loadingSessions,
+    refreshing: refreshingSessions = loadingSessions,
     refetch: refetchSessions,
     updateSessionTitle,
     removeSession,
@@ -410,9 +713,7 @@ export default function SessionPage() {
     loadMore: loadMoreSessions,
   } = useSessions(searchQuery, {
     projectIds: sessionProjectIds,
-    pageSize: projects.length >= 1
-      ? MULTI_PROJECT_SESSION_PAGE_SIZE
-      : SINGLE_PROJECT_SESSION_PAGE_SIZE,
+    pageSize: sessionListPageSize,
   });
   const { agents, loading: loadingAgents } = useAgents();
   const { providers, loading: loadingProviders } = useProviders();
@@ -428,16 +729,46 @@ export default function SessionPage() {
   const chatAgents = useMemo(() => [...primaryAgents, ...subAgents], [primaryAgents, subAgents]);
   const filteredChatAgents = useMemo(
     () => chatAgents.filter((agent) => {
+      if (agent.name.toLowerCase() === 'rex') return false;
       if (agentSourceFilter === 'builtin') return agent.native;
       if (agentSourceFilter === 'custom') return !agent.native;
       return true;
     }),
     [chatAgents, agentSourceFilter],
   );
-  const selectedAgentInfo = useMemo(
-    () => chatAgents.find((agent) => agent.name === selectedAgent),
-    [chatAgents, selectedAgent],
-  );
+  const loadComposerResources = useCallback(async () => {
+    if (composerResourcesLoadedRef.current || loadingComposerResources) return;
+    setLoadingComposerResources(true);
+    try {
+      const [workflowResult, skillResult] = await Promise.allSettled([
+        workflowAPI.listSummaries({ status: 'active' }),
+        skillAPI.status(),
+      ]);
+      if (workflowResult.status === 'fulfilled') {
+        setComposerWorkflows(
+          workflowResult.value.data
+            .filter((workflow) => workflow.status === 'active')
+            .sort((a, b) => getWorkflowDisplayName(a, i18n.language)
+              .localeCompare(getWorkflowDisplayName(b, i18n.language))),
+        );
+      }
+      if (skillResult.status === 'fulfilled') {
+        setComposerSkills(
+          skillResult.value.data
+            .filter((skill) => (
+              !skill.ui_hidden
+              && !skill.disabled
+              && skill.eligible !== false
+              && skill.category !== 'system'
+            ))
+            .sort((a, b) => a.name.localeCompare(b.name)),
+        );
+      }
+      composerResourcesLoadedRef.current = true;
+    } finally {
+      setLoadingComposerResources(false);
+    }
+  }, [i18n.language, loadingComposerResources]);
   const chatModelOptions = useMemo<ChatModelOption[]>(() => {
     const providerById = new Map(
       providers
@@ -461,6 +792,7 @@ export default function SessionPage() {
     };
 
     return enabledModelDefinitions.flatMap((model) => {
+      if (model.model_type !== 'llm') return [];
       const provider = providerById.get(model.provider_id);
       if (!provider) return [];
       return [{
@@ -515,18 +847,46 @@ export default function SessionPage() {
     ? makeModelKey(selectedSession.provider, selectedSession.model)
     : null;
   const hasPinnedModelOption = !!pinnedModelKey && chatModelOptions.some((option) => option.key === pinnedModelKey);
+  const selectedModelAuto = selectedModelKey === AUTO_MODEL_KEY;
   const selectedModelOption = useMemo(
     () => chatModelOptions.find((option) => option.key === selectedModelKey) ?? (selectedModelKey ? null : chatModelOptions[0] ?? null),
     [chatModelOptions, selectedModelKey],
   );
-  const selectedPromptModel = selectedModelOption
-    ? { providerID: selectedModelOption.providerID, modelID: selectedModelOption.modelID }
-    : null;
   const {
     data: resolvedDefaultModel,
     initialized: resolvedDefaultModelInitialized,
-  } = useResolvedDefaultModel(chatModelOptions.length > 0 && !hasPinnedModelOption);
-  const effectiveSupportsVision = selectedModelOption?.supportsVision ?? supportsVision;
+  } = useResolvedDefaultModel(chatModelOptions.length > 0);
+  const primaryModelOption = useMemo(() => {
+    if (!resolvedDefaultModel) return null;
+    const key = makeModelKey(resolvedDefaultModel.providerID, resolvedDefaultModel.modelID);
+    return chatModelOptions.find((option) => option.key === key) ?? null;
+  }, [chatModelOptions, resolvedDefaultModel]);
+  const autoSelectionAllowed = !selectedSessionId || Boolean(
+    selectedSession && ['user', 'entity-config', 'workflow'].includes(selectedSession.category ?? 'user'),
+  );
+  const canSelectAuto = Boolean(
+    autoSelectionAllowed && primaryModelOption,
+  );
+  const effectiveModelOption = selectedModelAuto ? primaryModelOption : selectedModelOption;
+  const selectedPromptModel = selectedModelAuto
+    ? null
+    : selectedModelOption
+      ? { providerID: selectedModelOption.providerID, modelID: selectedModelOption.modelID }
+      : null;
+  const effectiveSupportsVision = effectiveModelOption?.supportsVision ?? supportsVision;
+  const autoStatusLabel = !autoSelectionAllowed
+    ? t('modelPicker.autoUserSessionsOnly')
+    : primaryModelOption
+      ? t('modelPicker.autoHint')
+      : t('modelPicker.autoUnavailable');
+  const firstChatModelKey = chatModelOptions[0]?.key ?? null;
+  const resolvedDefaultKey = resolvedDefaultModel
+    ? makeModelKey(resolvedDefaultModel.providerID, resolvedDefaultModel.modelID)
+    : null;
+  const defaultSelectionKey = resolvedDefaultKey
+    && chatModelOptions.some(option => option.key === resolvedDefaultKey)
+    ? resolvedDefaultKey
+    : firstChatModelKey;
 
   const toggleProjectCollapsed = useCallback((projectId: string) => {
     setCollapsedProjectIds(prev => {
@@ -536,6 +896,15 @@ export default function SessionPage() {
     });
   }, []);
 
+  const setLoadedSessionGroupCollapsed = useCallback((groupId: string, collapsed: boolean) => {
+    setCollapsedLoadedSessionGroupIds((current) => {
+      const next = new Set(current);
+      if (collapsed) next.add(groupId);
+      else next.delete(groupId);
+      return next;
+    });
+  }, []);
+
   useEffect(() => {
     try {
       window.localStorage.setItem(
@@ -611,18 +980,27 @@ export default function SessionPage() {
         worktree: '',
         sessions: taskSessions,
         sessionCount: taskSessions.length,
-        pathStatus: 'available' as const,
       };
     },
     [projects, sessions, t],
   );
   const taskGroupCollapsed = collapsedProjectIds.has(TASK_SESSION_GROUP_ID);
-  const taskGroupSelected = selectedProjectId === TASK_SESSION_GROUP_ID;
-  const hasMoreTaskSessions = hasMoreByProject[TASK_SESSION_GROUP_ID] ?? false;
+  const taskSessionsCollapsedToFirstPage = collapsedLoadedSessionGroupIds.has(TASK_SESSION_GROUP_ID);
+  const visibleTaskSessions = taskSessionsCollapsedToFirstPage
+    ? taskSessionGroup.sessions.slice(0, sessionListPageSize)
+    : taskSessionGroup.sessions;
+  const hasMoreRemoteTaskSessions = hasMoreByProject[TASK_SESSION_GROUP_ID] ?? false;
+  const canShowMoreTaskSessions = taskSessionsCollapsedToFirstPage || hasMoreRemoteTaskSessions;
+  const canCollapseTaskSessions = !taskSessionsCollapsedToFirstPage
+    && taskSessionGroup.sessions.length > sessionListPageSize;
 
   const selectedProjectIDForCreate = selectedProjectId && selectedProjectId !== TASK_SESSION_GROUP_ID
     ? selectedProjectId
     : null;
+  const selectedProjectContextLabel = selectedProjectId === TASK_SESSION_GROUP_ID
+    ? t('projectPicker.none')
+    : projectSessionGroups.find((group) => group.id === selectedProjectId)?.label
+      ?? t('projectPicker.none');
 
   useEffect(() => {
     const selectableProjectIds = new Set([
@@ -665,22 +1043,29 @@ export default function SessionPage() {
 
   const fetchProjects = useCallback(async (ensureProject?: ProjectSummary, query = '') => {
     const requestSeq = ++projectListRequestSeqRef.current;
-    const listResult = await client.get('/api/project', {
-      params: { search: query.trim() || undefined },
-    });
-    if (requestSeq !== projectListRequestSeqRef.current) return;
-    const nextProjects = Array.isArray(listResult.data)
-      ? listResult.data.filter((project: ProjectSummary) => (
-        project.id !== TASK_SESSION_GROUP_ID && !project.isDefault
-      ))
-      : [];
-    setProjects((currentProjects) => {
-      if (!ensureProject?.id || nextProjects.some((project) => project.id === ensureProject.id)) {
-        return nextProjects;
+    setRefreshingProjects(true);
+    try {
+      const listResult = await client.get('/api/project', {
+        params: { search: query.trim() || undefined },
+      });
+      if (requestSeq !== projectListRequestSeqRef.current) return;
+      const nextProjects = Array.isArray(listResult.data)
+        ? listResult.data.filter((project: ProjectSummary) => (
+          project.id !== TASK_SESSION_GROUP_ID && !project.isDefault
+        ))
+        : [];
+      setProjects((currentProjects) => {
+        if (!ensureProject?.id || nextProjects.some((project) => project.id === ensureProject.id)) {
+          return nextProjects;
+        }
+        const currentProject = currentProjects.find((project) => project.id === ensureProject.id);
+        return [{ ...currentProject, ...ensureProject }, ...nextProjects];
+      });
+    } finally {
+      if (requestSeq === projectListRequestSeqRef.current) {
+        setRefreshingProjects(false);
       }
-      const currentProject = currentProjects.find((project) => project.id === ensureProject.id);
-      return [{ ...currentProject, ...ensureProject }, ...nextProjects];
-    });
+    }
   }, []);
 
   const scheduleSessionListRefetch = useCallback(() => {
@@ -702,6 +1087,19 @@ export default function SessionPage() {
   }, []);
 
   const handleSSEEvent = useCallback((event: SSEChatEvent) => {
+    if (
+      event.type === 'session.execution_mode.changed'
+      && event.properties?.sessionID === selectedSessionId
+      && event.properties?.executionMode === 'build'
+    ) {
+      setSelectedExecutionMode(DEFAULT_SESSION_EXECUTION_MODE);
+      writeSessionExecutionMode(
+        selectedSessionId,
+        DEFAULT_SESSION_EXECUTION_MODE,
+      );
+      setShowExecutionModeOptions(false);
+      return;
+    }
     if (event.type === 'session.notice' && event.properties?.kind === 'directory-fallback') {
       toast.warning(
         t('projectDirectoryFallbackTitle'),
@@ -721,12 +1119,67 @@ export default function SessionPage() {
       // those bursts don't turn into a request/re-render storm.
       scheduleSessionListRefetch();
     }
-  }, [scheduleSessionListRefetch, t, toast, updateSessionTitle]);
+  }, [
+    scheduleSessionListRefetch,
+    selectedSessionId,
+    t,
+    toast,
+    updateSessionTitle,
+  ]);
 
   useEffect(() => {
     void fetchProjects(undefined, searchQuery);
   }, [fetchProjects, searchQuery]);
 
+  const setSessionRunning = useCallback((sessionId: string, running: boolean) => {
+    setRunningSessionIds((current) => {
+      if (current.has(sessionId) === running) return current;
+      const next = new Set(current);
+      if (running) next.add(sessionId);
+      else next.delete(sessionId);
+      return next;
+    });
+  }, []);
+
+  const handleSessionStatusEvent = useCallback((event: SSEChatEvent) => {
+    if (event.type !== 'session.status') return;
+    const sessionId = event.properties?.sessionID;
+    if (typeof sessionId !== 'string') return;
+    sessionStatusEventVersionRef.current += 1;
+    setSessionRunning(sessionId, isRunningSessionStatus(event.properties?.status));
+  }, [setSessionRunning]);
+
+  const refreshRunningSessionIds = useCallback(async () => {
+    const eventVersion = sessionStatusEventVersionRef.current;
+    try {
+      const response = await client.get('/api/session/status');
+      if (sessionStatusEventVersionRef.current !== eventVersion) return;
+      setRunningSessionIds(readRunningSessionIds(response.data));
+    } catch {
+      // The sidebar status is progressive enhancement; chat remains usable
+      // when status recovery is temporarily unavailable.
+    }
+  }, []);
+
+  useSSE({
+    url: `${getApiBase()}/api/event`,
+    onEvent: handleSessionStatusEvent,
+    onReconnect: refreshRunningSessionIds,
+    enabled: true,
+    reconnect: { enabled: true, maxRetries: 5, initialDelay: 1000, maxDelay: 10000 },
+  });
+
+  useEffect(() => {
+    void refreshRunningSessionIds();
+  }, [refreshRunningSessionIds]);
+
+  useEffect(() => {
+    const intervalId = window.setInterval(() => {
+      setRelativeTimeClock((tick) => tick + 1);
+    }, 60_000);
+    return () => window.clearInterval(intervalId);
+  }, []);
+
   useEffect(() => {
     if (!openProjectMenuId) return;
     const closeMenu = () => setOpenProjectMenuId(null);
@@ -780,6 +1233,16 @@ export default function SessionPage() {
     writeLastSelectedSessionId(selectedSessionId);
   }, [selectedSession?.id, selectedSessionId]);
 
+  useEffect(() => {
+    const handoff = executionModeHandoffRef.current;
+    if (selectedSessionId && handoff?.sessionId === selectedSessionId) {
+      executionModeHandoffRef.current = null;
+      setSelectedExecutionMode(handoff.mode);
+      return;
+    }
+    setSelectedExecutionMode(readSessionExecutionMode(selectedSessionId));
+  }, [selectedSessionId]);
+
   useEffect(() => {
     if (!selectedSessionId) {
       setSelectedSessionFallback(null);
@@ -825,6 +1288,42 @@ export default function SessionPage() {
     return () => document.removeEventListener('mousedown', handle);
   }, [showAgentOptions]);
 
+  useEffect(() => {
+    if (!showExecutionModeOptions) return;
+    const handlePointerDown = (event: MouseEvent) => {
+      const target = event.target as HTMLElement;
+      if (!target.closest('[data-execution-mode-selector]')) {
+        setShowExecutionModeOptions(false);
+      }
+    };
+    const handleKeyDown = (event: KeyboardEvent) => {
+      if (event.key === 'Escape') setShowExecutionModeOptions(false);
+    };
+    document.addEventListener('mousedown', handlePointerDown);
+    document.addEventListener('keydown', handleKeyDown);
+    return () => {
+      document.removeEventListener('mousedown', handlePointerDown);
+      document.removeEventListener('keydown', handleKeyDown);
+    };
+  }, [showExecutionModeOptions]);
+
+  useEffect(() => {
+    if (!showProjectOptions) return;
+    const handle = (e: MouseEvent) => {
+      const target = e.target as HTMLElement;
+      if (!target.closest('[data-project-selector]')) setShowProjectOptions(false);
+    };
+    document.addEventListener('mousedown', handle);
+    return () => document.removeEventListener('mousedown', handle);
+  }, [showProjectOptions]);
+
+  useEffect(() => {
+    if (!showModelOptions) return;
+    updateModelMenuLeftOffset();
+    window.addEventListener('resize', updateModelMenuLeftOffset);
+    return () => window.removeEventListener('resize', updateModelMenuLeftOffset);
+  }, [showModelOptions, updateModelMenuLeftOffset]);
+
   useEffect(() => {
     if (!showModelOptions) return;
     const handle = (e: MouseEvent) => {
@@ -836,7 +1335,12 @@ export default function SessionPage() {
   }, [showModelOptions]);
 
   useEffect(() => {
-    if (chatModelOptions.length === 0) {
+    if (selectedSession?.model_auto) {
+      setSelectedModelKey(AUTO_MODEL_KEY);
+      return;
+    }
+
+    if (!firstChatModelKey) {
       setSelectedModelKey(null);
       return;
     }
@@ -848,30 +1352,37 @@ export default function SessionPage() {
 
     setSelectedModelKey(null);
     if (!resolvedDefaultModelInitialized) return;
-    const defaultKey = resolvedDefaultModel
-      ? makeModelKey(resolvedDefaultModel.providerID, resolvedDefaultModel.modelID)
-      : null;
-    const fallbackKey = chatModelOptions[0]?.key ?? null;
-    setSelectedModelKey(defaultKey && chatModelOptions.some((option) => option.key === defaultKey) ? defaultKey : fallbackKey);
+    setSelectedModelKey(defaultSelectionKey);
   }, [
-    chatModelOptions,
+    defaultSelectionKey,
+    firstChatModelKey,
     hasPinnedModelOption,
     pinnedModelKey,
-    resolvedDefaultModel,
     resolvedDefaultModelInitialized,
+    selectedSession?.model_auto,
     selectedSessionId,
   ]);
 
   useEffect(() => {
-    if (loadingEnabledModels || chatModelOptions.length === 0 || !selectedModelKey) return;
+    if (loadingEnabledModels || chatModelOptions.length === 0 || !selectedModelKey || selectedModelAuto) return;
     if (chatModelOptions.some((option) => option.key === selectedModelKey)) return;
     setSelectedModelKey(chatModelOptions[0].key);
-  }, [chatModelOptions, loadingEnabledModels, selectedModelKey]);
+  }, [chatModelOptions, loadingEnabledModels, selectedModelAuto, selectedModelKey]);
 
   useEffect(() => {
-    if (showAgentOptions || showModelOptions) return;
+    if (
+      showAgentOptions
+      || showExecutionModeOptions
+      || showModelOptions
+      || showProjectOptions
+    ) return;
     setSelectorTooltip(null);
-  }, [showAgentOptions, showModelOptions]);
+  }, [
+    showAgentOptions,
+    showExecutionModeOptions,
+    showModelOptions,
+    showProjectOptions,
+  ]);
 
   useEffect(() => {
     if (!openMenuSessionId) return;
@@ -880,6 +1391,7 @@ export default function SessionPage() {
       if (!target.closest('[data-session-actions]') && !target.closest('[data-session-menu-portal]')) {
         setOpenMenuSessionId(null);
         setMenuAnchor(null);
+        setMovePickerSessionId(null);
       }
     };
     document.addEventListener('mousedown', handle);
@@ -895,19 +1407,56 @@ export default function SessionPage() {
   useEffect(() => {
     if (!selectMode) return;
     setOpenMenuSessionId(null);
+    setMovePickerSessionId(null);
     setRenamingSessionId(null);
     setRenameValue('');
   }, [selectMode]);
 
+  const handleSelectExecutionMode = useCallback((mode: SessionExecutionMode) => {
+    setSelectedExecutionMode(mode);
+    setShowExecutionModeOptions(false);
+    if (mode !== 'goal') {
+      writeSessionExecutionMode(
+        selectedSessionId,
+        mode as PersistentSessionExecutionMode,
+      );
+    }
+  }, [selectedSessionId]);
+
+  const handleExecutionModeAccepted = useCallback((mode: SessionExecutionMode) => {
+    if (mode !== 'goal') return;
+    setSelectedExecutionMode(DEFAULT_SESSION_EXECUTION_MODE);
+    writeSessionExecutionMode(
+      selectedSessionId,
+      DEFAULT_SESSION_EXECUTION_MODE,
+    );
+  }, [selectedSessionId]);
+
+  const handleStartNewSession = useCallback(() => {
+    writeLastSelectedSessionId(null);
+    setSelectedSessionId(null);
+    setSelectedSessionFallback(null);
+    setPendingInitialMessage(null);
+    setPendingInitialDisplayText(null);
+    setSelectedAgent('rex');
+    setSelectedModelKey(null);
+    setSseStatus('disconnected');
+    setShowAgentOptions(false);
+    setShowModelOptions(false);
+    setShowProjectOptions(false);
+  }, []);
+
   const handleCreateSession = useCallback(async (projectIdOverride?: string) => {
     if (creating) return;
     const targetGroupId = projectIdOverride ?? selectedProjectId ?? TASK_SESSION_GROUP_ID;
     const projectID = targetGroupId === TASK_SESSION_GROUP_ID ? null : targetGroupId;
+    const carryAutoSelection = !selectedSessionId && selectedModelAuto;
     setCreating(true);
     try {
       const response = await client.post('/api/session', {
         title: 'New Session',
         ...(projectID ? { projectID } : {}),
+        ...(carryAutoSelection ? { model_auto: true } : {}),
       });
       addSession(response.data);
       await fetchProjects(undefined, searchQuery);
@@ -919,20 +1468,28 @@ export default function SessionPage() {
         return next;
       });
       setSelectedAgent('rex');
-      setSelectedModelKey(null);
+      executionModeHandoffRef.current = {
+        sessionId: response.data.id,
+        mode: DEFAULT_SESSION_EXECUTION_MODE,
+      };
+      resetDraftExecutionMode();
+      writeSessionExecutionMode(
+        response.data.id,
+        DEFAULT_SESSION_EXECUTION_MODE,
+      );
+      setSelectedExecutionMode(DEFAULT_SESSION_EXECUTION_MODE);
+      setShowExecutionModeOptions(false);
+      setSelectedModelKey(carryAutoSelection ? AUTO_MODEL_KEY : null);
       setSelectedSessionId(response.data.id);
     } catch (err: any) {
       toast.error(t('createFailed'), err.message);
     } finally {
       setCreating(false);
     }
-  }, [creating, selectedProjectId, addSession, fetchProjects, searchQuery, toast, t]);
-
-  const handleCreateSessionInProject = useCallback((projectId: string) => {
-    void handleCreateSession(projectId);
-  }, [handleCreateSession]);
+  }, [creating, selectedProjectId, selectedSessionId, selectedModelAuto, addSession, fetchProjects, searchQuery, toast, t]);
 
   const handleSelectModel = useCallback(async (option: ChatModelOption) => {
+    const previousModelKey = selectedModelKey;
     setSelectedModelKey(option.key);
     setShowModelOptions(false);
     if (!selectedSessionId) return;
@@ -942,12 +1499,33 @@ export default function SessionPage() {
         provider: option.providerID,
         model: option.modelID,
         model_pinned: true,
+        model_auto: false,
+      });
+      refetchSessions();
+    } catch (err: any) {
+      setSelectedModelKey(previousModelKey);
+      toast.error(t('chat.error', 'Error'), err.message);
+    }
+  }, [refetchSessions, selectedModelKey, selectedSessionId, toast, t]);
+
+  const handleSelectAutoModel = useCallback(async () => {
+    if (!canSelectAuto) return;
+    const previousModelKey = selectedModelKey;
+    setSelectedModelKey(AUTO_MODEL_KEY);
+    setShowModelOptions(false);
+    if (!selectedSessionId) return;
+
+    try {
+      await sessionApi.update(selectedSessionId, {
+        model_auto: true,
+        model_pinned: false,
       });
       refetchSessions();
     } catch (err: any) {
+      setSelectedModelKey(previousModelKey);
       toast.error(t('chat.error', 'Error'), err.message);
     }
-  }, [refetchSessions, selectedSessionId, toast, t]);
+  }, [canSelectAuto, refetchSessions, selectedModelKey, selectedSessionId, toast, t]);
 
   const handleCreateAndSend = useCallback(async (
     text: string,
@@ -955,18 +1533,37 @@ export default function SessionPage() {
     agentOverride?: string,
     modelOverride?: { providerID: string; modelID: string } | null,
     options?: PromptDisplayOptions,
+    executionModeOverride?: SessionExecutionMode,
   ) => {
     try {
+      const effectiveExecutionMode = executionModeOverride || selectedExecutionMode;
       const response = await client.post('/api/session', {
         title: 'New Session',
         ...(selectedProjectIDForCreate ? { projectID: selectedProjectIDForCreate } : {}),
+        ...(selectedModelAuto ? { model_auto: true } : {}),
       });
       const newSessionId = response.data.id;
 
       addSession(response.data);
       await fetchProjects(undefined, searchQuery);
       setSelectedSessionFallback(response.data);
-      setSelectedModelKey(null);
+      executionModeHandoffRef.current = {
+        sessionId: newSessionId,
+        mode: effectiveExecutionMode,
+      };
+      if (effectiveExecutionMode === 'goal') {
+        writeSessionExecutionMode(
+          newSessionId,
+          DEFAULT_SESSION_EXECUTION_MODE,
+        );
+        resetDraftExecutionMode();
+      } else {
+        promoteDraftExecutionMode(
+          newSessionId,
+          effectiveExecutionMode as PersistentSessionExecutionMode,
+        );
+      }
+      setSelectedModelKey(selectedModelAuto ? AUTO_MODEL_KEY : null);
       setSelectedSessionId(newSessionId);
 
       const payload: Record<string, unknown> = {
@@ -974,15 +1571,32 @@ export default function SessionPage() {
       };
       const effectiveAgent = agentOverride || selectedAgent || 'rex';
       if (effectiveAgent) payload.agent = effectiveAgent;
-      if (modelOverride) payload.model = modelOverride;
+      if (!selectedModelAuto && modelOverride) payload.model = modelOverride;
       if (options?.displayText) payload.displayText = options.displayText;
-      client.post(`/api/session/${newSessionId}/prompt_async`, payload).catch((err: any) => {
-        toast.error(t('chat.sendFailed', 'Send failed'), err.message);
-      });
+      payload.executionMode = effectiveExecutionMode;
+      await client.post(`/api/session/${newSessionId}/prompt_async`, payload);
+      if (effectiveExecutionMode === 'goal') {
+        setSelectedExecutionMode(DEFAULT_SESSION_EXECUTION_MODE);
+        writeSessionExecutionMode(
+          newSessionId,
+          DEFAULT_SESSION_EXECUTION_MODE,
+        );
+      }
     } catch (err: any) {
-      toast.error(t('createFailed'), err.message);
+      toast.error(t('chat.sendFailed', 'Send failed'), err.message);
+      throw err;
     }
-  }, [addSession, fetchProjects, searchQuery, selectedAgent, selectedProjectIDForCreate, toast, t]);
+  }, [
+    addSession,
+    fetchProjects,
+    searchQuery,
+    selectedAgent,
+    selectedExecutionMode,
+    selectedModelAuto,
+    selectedProjectIDForCreate,
+    toast,
+    t,
+  ]);
 
   const handleSuiteInstallProgress = useCallback((progress: HubInstallProgressEvent) => {
     setSuiteInstallProgress(current => applySuiteInstallProgressEvent(current, progress));
@@ -1045,22 +1659,28 @@ export default function SessionPage() {
     });
   }, []);
 
-  const handleDeleteSession = useCallback(async (sessionId: string) => {
+  const handleArchiveSession = useCallback(async (sessionId: string) => {
     const target = sessions.find((s) => s.id === sessionId);
     if (target?.canDelete === false) {
-      toast.error(t('deleteFailed'), i18n.t('auth:error.noPermissionToDeleteSession') as string);
+      toast.error(t('archiveFailed'), i18n.t('auth:error.noPermissionToDeleteSession') as string);
       return;
     }
-    if (!confirm(t('confirmDelete'))) return;
+    if (!confirm(t('confirmArchive'))) return;
     try {
-      await sessionApi.delete(sessionId);
-      // Remove from local state first so auto-select won't pick the deleted session.
-      // No need to refetchSessions — removeSession already keeps the list accurate.
-      if (selectedSessionId === sessionId) setSelectedSessionId(null);
-      removeSession(sessionId);
-      await fetchProjects(undefined, searchQuery);
+      await sessionApi.archive(sessionId);
     } catch (err: any) {
-      toast.error(t('deleteFailed'), err.message);
+      toast.error(t('archiveFailed'), err.message);
+      return;
+    }
+    // The mutation is complete. A secondary project-count refresh must not
+    // turn a successful archive into an error or re-enable the removed row.
+    if (selectedSessionId === sessionId) setSelectedSessionId(null);
+    removeSession(sessionId);
+    toast.success(t('archiveSuccess'));
+    try {
+      await fetchProjects(undefined, searchQuery);
+    } catch {
+      // SSE/reconnect refreshes will reconcile project counts later.
     }
   }, [fetchProjects, removeSession, searchQuery, selectedSessionId, toast, t]);
 
@@ -1353,6 +1973,41 @@ export default function SessionPage() {
     }
   }, [refetchSessions, t, toast]);
 
+  const handleMoveSessionToProject = useCallback(async (
+    sessionId: string,
+    project: ProjectSummary,
+  ) => {
+    if (movingSessionId) return;
+    setMovingSessionId(sessionId);
+    setMovingTargetProjectId(project.id);
+    try {
+      await sessionApi.moveToProject(sessionId, project.id);
+      setOpenMenuSessionId(null);
+      setMenuAnchor(null);
+      setMovePickerSessionId(null);
+      setProjectsSectionCollapsed(false);
+      setCollapsedProjectIds((current) => {
+        const next = new Set(current);
+        next.delete(project.id);
+        return next;
+      });
+      toast.success(t('moveToProjectSuccess', { project: getProjectLabel(project) }));
+      await Promise.allSettled([
+        refetchSessions(),
+        fetchProjects(undefined, searchQuery),
+      ]);
+    } catch (err: any) {
+      const detail = err?.response?.data?.detail;
+      toast.error(
+        t('moveToProjectFailed'),
+        typeof detail === 'string' ? detail : err?.message,
+      );
+    } finally {
+      setMovingSessionId(null);
+      setMovingTargetProjectId(null);
+    }
+  }, [fetchProjects, movingSessionId, refetchSessions, searchQuery, t, toast]);
+
   const handleEnterSelectMode = useCallback(() => {
     setSelectMode(true);
     setCheckedIds(new Set());
@@ -1379,6 +2034,7 @@ export default function SessionPage() {
     }
   }, [handleToggleCheck, selectMode]);
   const handleToggleSessionMenu = useCallback((sessionId: string, trigger: HTMLElement) => {
+    setMovePickerSessionId(null);
     setOpenMenuSessionId((current) => {
       if (current === sessionId) {
         setMenuAnchor(null);
@@ -1398,37 +2054,44 @@ export default function SessionPage() {
     }
   }, [checkedIds.size, sessions]);
 
-  const handleBatchDelete = useCallback(async () => {
-    if (checkedIds.size === 0 || batchDeleting) return;
-    if (!confirm(t('confirmBatchDelete', { count: checkedIds.size }))) return;
-    setBatchDeleting(true);
-    const ids = Array.from(checkedIds);
-    const succeeded: string[] = [];
-    const failed: string[] = [];
-    await Promise.all(ids.map(async (id) => {
-      try {
-        await client.delete(`/api/session/${id}`);
-        succeeded.push(id);
-      } catch {
-        failed.push(id);
+  const handleBatchArchive = useCallback(async () => {
+    if (checkedIds.size === 0 || batchArchiving) return;
+    if (!confirm(t('confirmBatchArchive', { count: checkedIds.size }))) return;
+    setBatchArchiving(true);
+    try {
+      const ids = Array.from(checkedIds);
+      const succeeded: string[] = [];
+      const failed: string[] = [];
+      await Promise.all(ids.map(async (id) => {
+        try {
+          await sessionApi.archive(id);
+          succeeded.push(id);
+        } catch {
+          failed.push(id);
+        }
+      }));
+      if (succeeded.length > 0) {
+        removeSessions(succeeded);
+        if (selectedSessionId && succeeded.includes(selectedSessionId)) {
+          setSelectedSessionId(null);
+        }
+        try {
+          await fetchProjects(undefined, searchQuery);
+        } catch {
+          // The archive results remain authoritative; counts reconcile later.
+        }
       }
-    }));
-    if (succeeded.length > 0) {
-      removeSessions(succeeded);
-      await fetchProjects(undefined, searchQuery);
-      if (selectedSessionId && succeeded.includes(selectedSessionId)) {
-        setSelectedSessionId(null);
+      if (failed.length > 0) {
+        setCheckedIds(new Set(failed));
+        toast.error(t('batchArchiveFailed', { count: failed.length }));
+      } else {
+        setCheckedIds(new Set());
+        setSelectMode(false);
       }
+    } finally {
+      setBatchArchiving(false);
     }
-    if (failed.length > 0) {
-      setCheckedIds(new Set(failed));
-      toast.error(t('batchDeleteFailed', { count: failed.length }));
-    } else {
-      setCheckedIds(new Set());
-      setSelectMode(false);
-    }
-    setBatchDeleting(false);
-  }, [batchDeleting, checkedIds, fetchProjects, removeSessions, searchQuery, selectedSessionId, toast, t]);
+  }, [batchArchiving, checkedIds, fetchProjects, removeSessions, searchQuery, selectedSessionId, toast, t]);
 
   const renderSessionListItem = (session: Session) => (
     <div
@@ -1524,59 +2187,94 @@ export default function SessionPage() {
     </div>
   );
 
-  if (loadingSessions) {
-    return (
-      <div className="flex items-center justify-center h-full">
-        <LoadingSpinner delayMs={180} />
-      </div>
-    );
-  }
+  const workbenchRefreshing = refreshingSessions || refreshingProjects || resolvingSelectedSession;
+  const workbenchRefreshLabel = resolvingSelectedSession
+    ? t('restoringTask')
+    : t('refreshingWorkbench');
+  const showSessionListSkeleton = loadingSessions && sessions.length === 0;
 
   return (
-    <div className="h-full w-full flex overflow-hidden bg-gray-50 dark:bg-zinc-950">
+    <div className="flex h-full w-full overflow-hidden bg-[#fcfcfd] text-[#202328] dark:bg-[#303842] dark:text-[#d7dee8]">
       {/* ── Sidebar ── */}
       <div
-        className={`bg-white border-r border-gray-100 flex flex-col transition-all duration-300 flex-shrink-0 h-full overflow-hidden dark:border-zinc-800 dark:bg-zinc-950 ${
-          sidebarCollapsed ? 'w-0' : 'w-64'
+        className={`flex h-full flex-shrink-0 flex-col overflow-hidden border-r bg-gray-50 transition-[width,opacity] duration-200 dark:bg-[#252c35] ${
+          sidebarCollapsed
+            ? 'w-0 border-transparent opacity-0'
+            : 'w-[282px] border-black/[0.10] opacity-100 dark:border-white/[0.10]'
         }`}
+        aria-label={t('managementTitle')}
       >
-        {/* Header：始终显示新建 + 搜索 */}
-        <div className="px-3 pt-3 pb-2 flex-shrink-0 space-y-2">
-          <div className="relative">
-            {creating
-              ? <Loader2 className="absolute left-2.5 top-1/2 -translate-y-1/2 w-3.5 h-3.5 animate-spin text-gray-400 pointer-events-none" />
-              : <Plus className="absolute left-2.5 top-1/2 -translate-y-1/2 w-3.5 h-3.5 text-gray-500 pointer-events-none" />}
-            <button
-              onClick={() => void handleCreateSession()}
-              disabled={creating}
-              className="w-full pl-8 pr-3 py-2 text-left bg-white border border-gray-200 text-gray-700 rounded-lg hover:bg-gray-50 hover:border-gray-300 shadow-sm hover:shadow transition-all disabled:opacity-60 disabled:cursor-not-allowed text-sm font-medium dark:border-zinc-800 dark:bg-zinc-900 dark:text-zinc-200 dark:hover:border-zinc-700 dark:hover:bg-zinc-800"
-            >
-              {t('newSession')}
-            </button>
+        {/* Header：始终显示标题、新建与搜索 */}
+        <div className="flex-shrink-0 px-3 pb-2 pt-3.5">
+          <div className="mb-2 flex h-8 items-center justify-between px-1">
+            <div className="min-w-0">
+              <strong className="text-sm font-semibold text-[#202328] dark:text-[#f0f3f7]">
+                {t('managementTitle')}
+              </strong>
+              <span className="ml-2 text-[11px] text-[#7b8087] dark:text-[#9aa7b4]">
+                {t('sessionCount', { count: sessions.length })}
+              </span>
+            </div>
           </div>
-          <div className="relative">
-            <Search className="absolute left-2.5 top-1/2 -translate-y-1/2 w-3.5 h-3.5 text-gray-400 pointer-events-none" />
-            <input
-              value={searchQuery}
-              onChange={(e) => setSearchQuery(e.target.value)}
-              placeholder={t('filterConversations', 'Filter conversations...')}
-              className="w-full pl-8 pr-3 py-1.5 text-sm bg-gray-100 rounded-lg border-0 outline-none focus:bg-gray-200 transition-colors placeholder:text-gray-400 text-gray-700 dark:bg-zinc-900 dark:text-zinc-200 dark:placeholder:text-zinc-600 dark:focus:bg-zinc-800"
-            />
+
+          <div className="space-y-0.5">
+            <div className="relative h-[34px] rounded-lg transition-colors hover:bg-black/[0.04] dark:hover:bg-white/[0.06]">
+              {creating
+                ? <Loader2 className="pointer-events-none absolute left-2 top-1/2 h-3.5 w-3.5 -translate-y-1/2 animate-spin text-[#7b8087]" />
+                : <PencilLine className="pointer-events-none absolute left-2 top-1/2 h-3.5 w-3.5 -translate-y-1/2 text-[#7b8087] dark:text-[#9aa7b4]" />}
+              <button
+                onClick={handleStartNewSession}
+                disabled={creating}
+                className="h-full w-full rounded-lg border-0 bg-transparent pl-9 pr-3 text-left text-sm font-medium text-[#474b51] transition-colors hover:text-[#202328] disabled:cursor-not-allowed disabled:opacity-60 dark:text-[#c3ccd6] dark:hover:text-white"
+              >
+                {t('newSession')}
+              </button>
+            </div>
+
+            <div className="relative h-[34px] rounded-lg transition-colors hover:bg-black/[0.04] focus-within:bg-black/[0.04] dark:hover:bg-white/[0.06] dark:focus-within:bg-white/[0.06]">
+              <Search className="pointer-events-none absolute left-2 top-1/2 h-3.5 w-3.5 -translate-y-1/2 text-[#7b8087] dark:text-[#9aa7b4]" />
+              <input
+                value={searchQuery}
+                onChange={(e) => setSearchQuery(e.target.value)}
+                placeholder={t('filterConversations', 'Search tasks')}
+                className="h-full w-full rounded-lg border-0 bg-transparent pl-9 pr-8 text-sm font-medium text-[#474b51] outline-none placeholder:text-[#474b51] focus:bg-transparent dark:text-[#c3ccd6] dark:placeholder:text-[#c3ccd6]"
+              />
+              {searchQuery && (
+                <button
+                  type="button"
+                  onClick={() => setSearchQuery('')}
+                  className="absolute right-1.5 top-1/2 grid h-5 w-5 -translate-y-1/2 place-items-center rounded-md text-[#7b8087] transition-colors hover:bg-black/[0.065] hover:text-[#202328] dark:text-[#9aa7b4] dark:hover:bg-white/[0.08] dark:hover:text-white"
+                  title={t('clearSearch')}
+                  aria-label={t('clearSearch')}
+                >
+                  <X className="h-3.5 w-3.5" />
+                </button>
+              )}
+            </div>
           </div>
         </div>
 
         {/* Session list */}
-        <div className="flex-1 overflow-y-auto overflow-x-hidden scrollbar-hide pb-2">
+        <div
+          data-testid="session-list-scroll"
+          className="session-sidebar-scrollbar min-h-0 flex-1 overflow-y-auto overflow-x-hidden px-2 pb-[18px] pt-1.5"
+        >
+          {showSessionListSkeleton ? (
+            <SessionListSkeleton />
+          ) : (
           <div>
             <section className="group/projects-section">
-                <div className="flex items-center justify-between px-4 pt-4 pb-1 text-xs font-semibold text-gray-400 uppercase tracking-wide select-none dark:text-zinc-600">
+                <div className="flex h-7 select-none items-center gap-1.5 px-2 text-xs font-semibold uppercase tracking-[0.02em] text-zinc-500 dark:text-[#8f9ba8]">
                   <button
                     type="button"
                     onClick={() => setProjectsSectionCollapsed((collapsed) => !collapsed)}
-                    className="flex items-center gap-1 rounded text-left transition-colors hover:text-gray-700 dark:hover:text-zinc-300"
+                    className="flex h-6 items-center gap-1 rounded-lg px-1 text-left transition-colors hover:bg-black/[0.04] hover:text-[#474b51] dark:hover:bg-white/[0.06] dark:hover:text-white"
                     aria-label={t('toggleProjects')}
                   >
                     <span>{t('projectsSection')}</span>
+                    <span className="font-normal tabular-nums text-[#858a91] dark:text-[#8f9ba8]">
+                      ({managedProjectSessionGroups.length})
+                    </span>
                     {projectsSectionCollapsed
                       ? <ChevronRight className="h-3.5 w-3.5 shrink-0" />
                       : <ChevronDown className="h-3.5 w-3.5 shrink-0" />}
@@ -1584,7 +2282,7 @@ export default function SessionPage() {
                   <button
                     type="button"
                     onClick={handleOpenCreateProject}
-                    className="rounded p-0.5 text-gray-400 transition-colors hover:bg-gray-100 hover:text-gray-700 dark:text-zinc-600 dark:hover:bg-zinc-900 dark:hover:text-zinc-300"
+                    className="ml-auto grid h-6 w-6 place-items-center rounded-lg text-[#8a8e94] transition-colors hover:bg-black/[0.04] hover:text-[#474b51] dark:text-[#8f9ba8] dark:hover:bg-white/[0.06] dark:hover:text-white"
                     title={t('projectDialog.createTitle')}
                     aria-label={t('projectDialog.createTitle')}
                   >
@@ -1592,22 +2290,30 @@ export default function SessionPage() {
                   </button>
                 </div>
                 {!projectsSectionCollapsed && (
-                  <div className="space-y-2">
+                  <div>
                 {managedProjectSessionGroups.map((group) => {
                   const collapsed = collapsedProjectIds.has(group.id);
                   const isSelectedProject = selectedProjectId === group.id;
-                  const hasMoreProjectSessions = (
+                  const sessionsCollapsedToFirstPage = collapsedLoadedSessionGroupIds.has(group.id);
+                  const visibleProjectSessions = sessionsCollapsedToFirstPage
+                    ? group.sessions.slice(0, sessionListPageSize)
+                    : group.sessions;
+                  const hasMoreRemoteProjectSessions = (
                     group.sessions.length < group.sessionCount
                     || hasMoreByProject[group.id]
                   );
+                  const canShowMoreProjectSessions = sessionsCollapsedToFirstPage
+                    || hasMoreRemoteProjectSessions;
+                  const canCollapseProjectSessions = !sessionsCollapsedToFirstPage
+                    && group.sessions.length > sessionListPageSize;
                   const persistedProject = projects.find((project) => project.id === group.id);
                   return (
                     <div key={group.id} className="group/project relative">
                       <div
-                        className={`mx-2 flex items-center gap-1 rounded-lg px-1.5 py-1 text-sm transition-colors ${
+                        className={`flex h-[34px] items-center gap-1 rounded-lg px-3 text-sm transition-colors ${
                           isSelectedProject
-                            ? 'bg-gray-100 text-gray-900 dark:bg-zinc-900 dark:text-zinc-100'
-                            : 'text-gray-600 hover:bg-gray-50 hover:text-gray-900 dark:text-zinc-400 dark:hover:bg-zinc-900 dark:hover:text-zinc-100'
+                            ? 'font-semibold text-[#202328] dark:text-white'
+                            : 'font-medium text-[#474b51] hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white'
                         }`}
                         title={group.worktree}
                       >
@@ -1617,12 +2323,12 @@ export default function SessionPage() {
                             setSelectedProjectId(group.id);
                             toggleProjectCollapsed(group.id);
                           }}
-                          className="flex min-w-0 flex-1 items-center gap-2 rounded px-1 py-0.5 text-left"
+                          className="flex h-full min-w-0 flex-1 items-center gap-2 rounded-lg text-left"
                           aria-label={t('selectProject', { project: group.label })}
                           aria-expanded={!collapsed}
                         >
-                          <FolderGit2 className="h-3.5 w-3.5 shrink-0 text-gray-500 dark:text-zinc-500" />
-                          <span className="min-w-0 flex-1 truncate font-medium">{group.label}</span>
+                          <FolderGit2 className="h-3.5 w-3.5 shrink-0 text-[#6f757c] dark:text-[#9aa7b4]" />
+                          <span className="min-w-0 flex-1 truncate">{group.label}</span>
                           {group.isShared && (
                             <span className="inline-flex shrink-0 items-center rounded-full border border-blue-200 bg-blue-50 px-1.5 py-0.5 text-[10px] font-medium text-blue-700 dark:border-blue-500/30 dark:bg-blue-950/30 dark:text-blue-300">
                               {t('sharedTag')}
@@ -1635,19 +2341,6 @@ export default function SessionPage() {
                             />
                           )}
                         </button>
-                        <button
-                          type="button"
-                          onClick={(event) => {
-                            event.stopPropagation();
-                            handleCreateSessionInProject(group.id);
-                          }}
-                          disabled={creating || !group.canWrite || group.pathStatus !== 'available'}
-                          className="rounded p-1 text-gray-400 transition-colors hover:bg-gray-200 hover:text-gray-700 disabled:cursor-not-allowed disabled:opacity-50 dark:text-zinc-600 dark:hover:bg-zinc-800 dark:hover:text-zinc-200"
-                          title={t('createSessionInProject', { project: group.label })}
-                          aria-label={t('createSessionInProject', { project: group.label })}
-                        >
-                          {creating && isSelectedProject ? <Loader2 className="h-3 w-3 animate-spin" /> : <Plus className="h-3 w-3" />}
-                        </button>
                         {persistedProject && (
                           <button
                             type="button"
@@ -1655,7 +2348,11 @@ export default function SessionPage() {
                               event.stopPropagation();
                               setOpenProjectMenuId((current) => current === group.id ? null : group.id);
                             }}
-                            className="rounded p-1 text-gray-400 transition-colors hover:bg-gray-200 hover:text-gray-700 dark:text-zinc-600 dark:hover:bg-zinc-800 dark:hover:text-zinc-200"
+                            className={`grid h-[26px] w-[26px] place-items-center rounded-lg text-[#7b8087] transition-all hover:bg-black/[0.065] hover:text-[#202328] dark:text-[#9aa7b4] dark:hover:bg-white/[0.08] dark:hover:text-white ${
+                              openProjectMenuId === group.id || isSelectedProject
+                                ? 'opacity-100'
+                                : 'opacity-0 group-hover/project:opacity-100 group-focus-within/project:opacity-100'
+                            }`}
                             title={t('projectActions')}
                             aria-label={t('projectActions')}
                             aria-expanded={openProjectMenuId === group.id}
@@ -1663,13 +2360,10 @@ export default function SessionPage() {
                             <MoreHorizontal className="h-3.5 w-3.5" />
                           </button>
                         )}
-                        <span className="inline-flex min-w-5 shrink-0 items-center justify-center rounded-full bg-white/80 px-1.5 py-0.5 text-[10px] tabular-nums text-gray-400 dark:bg-zinc-800 dark:text-zinc-500">
-                          {group.sessionCount}
-                        </span>
                       </div>
                       {persistedProject && openProjectMenuId === group.id && (
                         <div
-                          className="absolute right-8 top-8 z-30 min-w-28 rounded-md border border-zinc-200 bg-white p-1 shadow-lg dark:border-zinc-700 dark:bg-zinc-900"
+                          className="absolute right-7 top-8 z-30 grid w-[140px] gap-0.5 rounded-[10px] border border-black/[0.11] bg-[#fdfdfc] p-1 shadow-[0_8px_24px_rgba(22,27,34,0.10)] dark:border-white/[0.10] dark:bg-[#303842] dark:shadow-[0_12px_32px_rgba(15,18,22,0.35)]"
                           role="menu"
                           onClick={(event) => event.stopPropagation()}
                         >
@@ -1677,7 +2371,7 @@ export default function SessionPage() {
                             type="button"
                             role="menuitem"
                             onClick={() => void handleCopyProjectPath(group)}
-                            className="flex w-full items-center gap-2 rounded px-2 py-1.5 text-left text-xs text-zinc-700 hover:bg-zinc-100 dark:text-zinc-200 dark:hover:bg-zinc-800"
+                            className="flex h-[30px] w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
                           >
                             <Copy className="h-3.5 w-3.5" />
                             {t('projectDialog.copyPathAction')}
@@ -1687,7 +2381,7 @@ export default function SessionPage() {
                               type="button"
                               role="menuitem"
                               onClick={() => void handleShareProject(group, !group.isShared)}
-                              className="flex w-full items-center gap-2 rounded px-2 py-1.5 text-left text-xs text-zinc-700 hover:bg-zinc-100 dark:text-zinc-200 dark:hover:bg-zinc-800"
+                              className="flex h-[30px] w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
                             >
                               <Share2 className="h-3.5 w-3.5" />
                               {t(group.isShared ? 'unshareAction' : 'shareAction')}
@@ -1698,21 +2392,21 @@ export default function SessionPage() {
                               type="button"
                               role="menuitem"
                               onClick={() => handleOpenRenameProject(group)}
-                              className="flex w-full items-center gap-2 rounded px-2 py-1.5 text-left text-xs text-zinc-700 hover:bg-zinc-100 dark:text-zinc-200 dark:hover:bg-zinc-800"
+                              className="flex h-[30px] w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
                             >
                               <PencilLine className="h-3.5 w-3.5" />
                               {t('projectDialog.renameAction')}
                             </button>
                           )}
                           {group.canDelete && (
-                            <div className="mx-2 my-1 border-t border-zinc-100 dark:border-zinc-800" />
+                            <div className="mx-2 my-0.5 border-t border-black/[0.07] dark:border-white/[0.08]" />
                           )}
                           {group.canDelete && (
                             <button
                               type="button"
                               role="menuitem"
                               onClick={() => handleOpenDeleteProject(group)}
-                              className="flex w-full items-center gap-2 rounded px-2 py-1.5 text-left text-xs text-red-600 hover:bg-red-50 disabled:cursor-not-allowed disabled:opacity-40 dark:text-red-300 dark:hover:bg-red-950/40"
+                              className="flex h-[30px] w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#c33c36] transition-colors hover:bg-[#fff0ef] hover:text-[#a92520] disabled:cursor-not-allowed disabled:opacity-40 dark:text-red-300 dark:hover:bg-red-500/10 dark:hover:text-red-200"
                             >
                               <Trash2 className="h-3.5 w-3.5" />
                               {t('projectDialog.deleteAction')}
@@ -1721,9 +2415,9 @@ export default function SessionPage() {
                         </div>
                       )}
                       {!collapsed && (
-                        <div className="mt-1.5">
+                        <div className="mb-1 mt-0.5">
                           {group.sessions.length > 0 ? (
-                            group.sessions.map((session) => (
+                            visibleProjectSessions.map((session) => (
                               <SessionSidebarItem
                                 key={session.id}
                                 session={session}
@@ -1732,9 +2426,12 @@ export default function SessionPage() {
                                 selectMode={selectMode}
                                 checked={checkedIds.has(session.id)}
                                 menuOpen={openMenuSessionId === session.id}
+                                running={runningSessionIds.has(session.id)}
                                 renaming={renamingSessionId === session.id}
                                 renameValue={renameValue}
                                 renameSubmitting={renameSubmitting}
+                                language={i18n.language}
+                                relativeTimeClock={relativeTimeClock}
                                 t={t}
                                 renameInputRef={renameInputRef}
                                 onSelect={handleSelectSessionRow}
@@ -1746,23 +2443,40 @@ export default function SessionPage() {
                               />
                             ))
                           ) : (
-                            <div className="ml-7 mr-2 rounded-lg px-3 py-2 text-xs text-gray-400 dark:text-zinc-600">
+                            <div className="ml-[19px] px-2.5 py-2 text-[11px] text-[#a0a4aa] dark:text-[#8f9ba8]">
                               {t('noProjectSessions')}
                             </div>
                           )}
-                          {hasMoreProjectSessions && (
-                            <div className="ml-7 mr-2 py-1">
+                          {(canShowMoreProjectSessions || canCollapseProjectSessions) && (
+                            <div className="ml-[19px] flex h-7 items-center justify-center gap-2 py-0.5">
+                              {canShowMoreProjectSessions && (
                               <button
                                 type="button"
-                                onClick={() => void loadMoreSessions(group.id)}
+                                onClick={() => {
+                                  if (sessionsCollapsedToFirstPage) {
+                                    setLoadedSessionGroupCollapsed(group.id, false);
+                                    return;
+                                  }
+                                  void loadMoreSessions(group.id);
+                                }}
                                 disabled={loadingMoreProjectIds.has(group.id)}
-                                className="flex w-full items-center justify-center gap-1 rounded px-2 py-1 text-xs text-gray-400 hover:bg-gray-50 hover:text-gray-700 disabled:cursor-not-allowed disabled:opacity-50 dark:hover:bg-zinc-900 dark:hover:text-zinc-300"
+                                className="flex h-7 items-center justify-center gap-1 rounded-lg px-2 text-[11px] text-[#969aa0] transition-colors hover:bg-black/[0.04] hover:text-[#474b51] disabled:cursor-not-allowed disabled:opacity-50 dark:text-[#8f9ba8] dark:hover:bg-white/[0.06] dark:hover:text-white"
                               >
-                                {loadingMoreProjectIds.has(group.id)
-                                  ? <Loader2 className="h-3 w-3 animate-spin" />
-                                  : <ChevronDown className="h-3 w-3" />}
-                                {t('loadMore', 'Load more')}
+                                {loadingMoreProjectIds.has(group.id) && (
+                                  <Loader2 className="h-3 w-3 animate-spin" />
+                                )}
+                                {t('loadMore')}
                               </button>
+                              )}
+                              {canCollapseProjectSessions && (
+                                <button
+                                  type="button"
+                                  onClick={() => setLoadedSessionGroupCollapsed(group.id, true)}
+                                  className="flex h-7 items-center justify-center rounded-lg px-2 text-[11px] text-[#969aa0] transition-colors hover:bg-black/[0.04] hover:text-[#474b51] dark:text-[#8f9ba8] dark:hover:bg-white/[0.06] dark:hover:text-white"
+                                >
+                                  {t('collapseLoaded')}
+                                </button>
+                              )}
                             </div>
                           )}
                         </div>
@@ -1775,53 +2489,28 @@ export default function SessionPage() {
               </section>
               {taskSessionGroup && (
                 <section className="group/tasks-section mt-3">
-                  <div className="flex items-center gap-1 px-4 pb-1 pt-2 text-xs font-semibold uppercase tracking-wide text-gray-400 select-none dark:text-zinc-600">
+                  <div className="flex h-7 select-none items-center gap-1 px-2 text-xs font-semibold uppercase tracking-[0.02em] text-zinc-500 dark:text-[#8f9ba8]">
                     <button
                       type="button"
                       onClick={() => {
                         setSelectedProjectId(taskSessionGroup.id);
                         toggleProjectCollapsed(taskSessionGroup.id);
                       }}
-                      className="min-w-0 truncate rounded text-left transition-colors hover:text-gray-700 dark:hover:text-zinc-300"
-                      aria-label={t('selectTasks')}
-                    >
-                      {t('tasksSection')}
-                    </button>
-                    <button
-                      type="button"
-                      onClick={(event) => {
-                        event.stopPropagation();
-                        toggleProjectCollapsed(taskSessionGroup.id);
-                      }}
-                      className="mr-auto rounded p-0.5 text-gray-400 transition-colors hover:bg-gray-100 hover:text-gray-700 dark:text-zinc-600 dark:hover:bg-zinc-900 dark:hover:text-zinc-300"
+                      className="flex h-6 items-center gap-1 rounded-lg px-1 text-left transition-colors hover:bg-black/[0.04] hover:text-[#474b51] dark:hover:bg-white/[0.06] dark:hover:text-white"
                       aria-label={t('toggleTasks')}
                     >
+                      <span>{t('tasksSection')}</span>
+                      <span className="font-normal tabular-nums text-[#858a91] dark:text-[#8f9ba8]">
+                        ({taskSessionGroup.sessionCount})
+                      </span>
                       {taskGroupCollapsed
                         ? <ChevronRight className="h-3.5 w-3.5 shrink-0" />
                         : <ChevronDown className="h-3.5 w-3.5 shrink-0" />}
                     </button>
-                    <button
-                      type="button"
-                      onClick={(event) => {
-                        event.stopPropagation();
-                        void handleCreateSession(TASK_SESSION_GROUP_ID);
-                      }}
-                      disabled={creating || taskSessionGroup.pathStatus !== 'available'}
-                      className="rounded p-0.5 text-gray-400 transition-colors hover:bg-gray-100 hover:text-gray-700 disabled:cursor-not-allowed disabled:opacity-50 dark:text-zinc-600 dark:hover:bg-zinc-900 dark:hover:text-zinc-300"
-                      title={t('createTaskSession')}
-                      aria-label={t('createTaskSession')}
-                    >
-                      {creating && taskGroupSelected
-                        ? <Loader2 className="h-3 w-3 animate-spin" />
-                        : <Plus className="h-3 w-3" />}
-                    </button>
-                    <span className="w-5 shrink-0 text-right tabular-nums">
-                      {taskSessionGroup.sessionCount}
-                    </span>
                   </div>
                   {!taskGroupCollapsed && (
-                    <div className="mt-1 space-y-1">
-                      {taskSessionGroup.sessions.map((session) => (
+                    <div className="mt-0.5">
+                      {visibleTaskSessions.map((session) => (
                         <SessionSidebarItem
                           key={session.id}
                           session={session}
@@ -1829,9 +2518,12 @@ export default function SessionPage() {
                           selectMode={selectMode}
                           checked={checkedIds.has(session.id)}
                           menuOpen={openMenuSessionId === session.id}
+                          running={runningSessionIds.has(session.id)}
                           renaming={renamingSessionId === session.id}
                           renameValue={renameValue}
                           renameSubmitting={renameSubmitting}
+                          language={i18n.language}
+                          relativeTimeClock={relativeTimeClock}
                           t={t}
                           renameInputRef={renameInputRef}
                           onSelect={handleSelectSessionRow}
@@ -1842,19 +2534,36 @@ export default function SessionPage() {
                           onToggleMenu={handleToggleSessionMenu}
                         />
                       ))}
-                      {hasMoreTaskSessions && (
-                        <div className="mx-4 py-1">
+                      {(canShowMoreTaskSessions || canCollapseTaskSessions) && (
+                        <div className="flex h-7 items-center justify-center gap-2 py-0.5">
+                          {canShowMoreTaskSessions && (
                           <button
                             type="button"
-                            onClick={() => void loadMoreSessions(taskSessionGroup.id)}
+                            onClick={() => {
+                              if (taskSessionsCollapsedToFirstPage) {
+                                setLoadedSessionGroupCollapsed(taskSessionGroup.id, false);
+                                return;
+                              }
+                              void loadMoreSessions(taskSessionGroup.id);
+                            }}
                             disabled={loadingMoreProjectIds.has(taskSessionGroup.id)}
-                            className="flex w-full items-center justify-center gap-1 rounded px-2 py-1 text-xs text-gray-400 hover:bg-gray-50 hover:text-gray-700 disabled:cursor-not-allowed disabled:opacity-50 dark:hover:bg-zinc-900 dark:hover:text-zinc-300"
+                            className="flex h-7 items-center justify-center gap-1 rounded-lg px-2 text-[11px] text-[#969aa0] transition-colors hover:bg-black/[0.04] hover:text-[#474b51] disabled:cursor-not-allowed disabled:opacity-50 dark:text-[#8f9ba8] dark:hover:bg-white/[0.06] dark:hover:text-white"
                           >
-                            {loadingMoreProjectIds.has(taskSessionGroup.id)
-                              ? <Loader2 className="h-3 w-3 animate-spin" />
-                              : <ChevronDown className="h-3 w-3" />}
-                            {t('loadMore', 'Load more')}
+                            {loadingMoreProjectIds.has(taskSessionGroup.id) && (
+                              <Loader2 className="h-3 w-3 animate-spin" />
+                            )}
+                            {t('loadMore')}
                           </button>
+                          )}
+                          {canCollapseTaskSessions && (
+                            <button
+                              type="button"
+                              onClick={() => setLoadedSessionGroupCollapsed(taskSessionGroup.id, true)}
+                              className="flex h-7 items-center justify-center rounded-lg px-2 text-[11px] text-[#969aa0] transition-colors hover:bg-black/[0.04] hover:text-[#474b51] dark:text-[#8f9ba8] dark:hover:bg-white/[0.06] dark:hover:text-white"
+                            >
+                              {t('collapseLoaded')}
+                            </button>
+                          )}
                         </div>
                       )}
                     </div>
@@ -1862,41 +2571,48 @@ export default function SessionPage() {
                 </section>
               )}
           </div>
+          )}
         </div>
 
         {/* Bottom：批量操作栏 / 批量选择入口 */}
         {sessions.length > 0 && (
-          <div className="border-t border-gray-100 px-3 pt-3 pb-4 flex-shrink-0 dark:border-zinc-800">
+          <div className="flex min-h-11 flex-shrink-0 items-center justify-center border-t border-black/[0.07] px-3 dark:border-white/[0.08]">
             {selectMode ? (
               <div className="grid grid-cols-3 gap-1.5">
                 <button
                   onClick={handleSelectAll}
-                  className="flex items-center justify-center py-2 text-sm text-blue-600 bg-blue-50 hover:bg-blue-100 rounded-lg transition-colors"
+                  className="grid h-[30px] w-[30px] place-items-center rounded-lg text-blue-600 transition-colors hover:bg-blue-50 dark:text-blue-300 dark:hover:bg-blue-500/10"
+                  title={checkedIds.size === sessions.length && sessions.length > 0 ? t('deselectAll') : t('selectAll')}
+                  aria-label={checkedIds.size === sessions.length && sessions.length > 0 ? t('deselectAll') : t('selectAll')}
                 >
-                  {checkedIds.size === sessions.length && sessions.length > 0 ? t('deselectAll') : t('selectAll')}
+                  <CheckSquare className="h-[15px] w-[15px]" />
                 </button>
                 <button
                   onClick={handleExitSelectMode}
-                  className="flex items-center justify-center py-2 text-sm text-gray-600 bg-gray-100 hover:bg-gray-200 rounded-lg transition-colors"
+                  className="grid h-[30px] w-[30px] place-items-center rounded-lg text-[#60656b] transition-colors hover:bg-black/[0.04] dark:text-[#b8c2cc] dark:hover:bg-white/[0.06]"
+                  title={t('cancelSelect')}
+                  aria-label={t('cancelSelect')}
                 >
-                  {t('cancelSelect')}
+                  <X className="h-[15px] w-[15px]" />
                 </button>
                 <button
-                  onClick={handleBatchDelete}
-                  disabled={checkedIds.size === 0 || batchDeleting}
-                  className="flex items-center justify-center py-2 text-sm text-red-600 bg-red-50 hover:bg-red-100 rounded-lg disabled:opacity-40 disabled:cursor-not-allowed transition-colors"
-                  title={t('deleteSelected', { count: checkedIds.size })}
+                  onClick={handleBatchArchive}
+                  disabled={checkedIds.size === 0 || batchArchiving}
+                  className="grid h-[30px] w-[30px] place-items-center rounded-lg text-red-600 transition-colors hover:bg-red-50 disabled:cursor-default disabled:opacity-35 dark:text-red-300 dark:hover:bg-red-500/10"
+                  title={t('archiveSelected', { count: checkedIds.size })}
+                  aria-label={t('archiveSelected', { count: checkedIds.size })}
                 >
-                  {batchDeleting ? <Loader2 className="w-4 h-4 animate-spin" /> : <Trash2 className="w-4 h-4" />}
+                  {batchArchiving ? <Loader2 className="h-[15px] w-[15px] animate-spin" /> : <Archive className="h-[15px] w-[15px]" />}
                 </button>
               </div>
             ) : (
               <button
                 onClick={handleEnterSelectMode}
-                className="w-full flex items-center justify-center gap-1.5 py-1.5 text-xs text-gray-400 hover:text-gray-600 hover:bg-gray-50 rounded-lg transition-colors"
+                className="grid h-[30px] w-[30px] place-items-center rounded-lg text-[#7b8087] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#9aa7b4] dark:hover:bg-white/[0.06] dark:hover:text-white"
+                title={t('selectMode')}
+                aria-label={t('selectMode')}
               >
-                <CheckSquare className="w-3.5 h-3.5" />
-                <span>{t('selectMode')}</span>
+                <CheckSquare className="h-[15px] w-[15px]" />
               </button>
             )}
           </div>
@@ -1904,32 +2620,34 @@ export default function SessionPage() {
       </div>
 
       {/* ── Main area ── */}
-      <div className="flex-1 flex flex-col overflow-hidden h-full min-w-0">
+      <div className="flex h-full min-w-0 flex-1 flex-col overflow-hidden bg-[#fcfcfd] dark:bg-[#303842]">
         {/* Header */}
-        <div className="px-6 h-12 border-b border-gray-200 bg-white flex items-center justify-between flex-shrink-0 relative dark:border-zinc-800 dark:bg-zinc-950/95">
-          <div className="absolute left-4 top-1/2 -translate-y-1/2">
+        <header className="relative flex h-[52px] flex-shrink-0 items-center gap-2 px-4 text-[13px]">
+          <div className="shrink-0">
             <button
               onClick={() => setSidebarCollapsed(!sidebarCollapsed)}
-              className="p-2 bg-white border border-gray-300 rounded-lg hover:bg-gray-50 shadow-sm hover:shadow-md transition-all duration-200 dark:border-zinc-800 dark:bg-zinc-900 dark:text-zinc-300 dark:hover:bg-zinc-800 dark:hover:shadow-none"
+              className="grid h-[30px] w-[30px] place-items-center rounded-lg border-0 bg-transparent text-[#7b8087] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#9aa7b4] dark:hover:bg-white/[0.06] dark:hover:text-white"
               title={sidebarCollapsed ? t('showHistory') : t('hideHistory')}
+              aria-label={sidebarCollapsed ? t('showHistory') : t('hideHistory')}
             >
-              {sidebarCollapsed ? <PanelLeft className="w-5 h-5" /> : <PanelLeftClose className="w-5 h-5" />}
+              {sidebarCollapsed ? <PanelLeft className="h-4 w-4" /> : <PanelLeftClose className="h-4 w-4" />}
             </button>
           </div>
 
-          <div className="flex items-center gap-3 ml-14">
-            <h2 className="text-base font-semibold text-gray-900 dark:text-zinc-100">
+          <div className="flex min-w-0 items-center">
+            <h2 className="truncate text-sm font-semibold text-[#555a61] dark:text-[#c3ccd6]">
               {selectedSession?.title || t('newSession')}
             </h2>
           </div>
 
-        </div>
+          {workbenchRefreshing && (
+            <WorkbenchRefreshStatus label={workbenchRefreshLabel} />
+          )}
+        </header>
 
         {/* Chat — powered by unified SessionChat */}
         {resolvingSelectedSession ? (
-          <div className="flex min-h-0 flex-1 items-center justify-center">
-            <LoadingSpinner delayMs={180} />
-          </div>
+          <SessionChatSkeleton />
         ) : (
           <SessionChat
             key={activeChatSessionId ?? 'empty-session'}
@@ -1938,6 +2656,7 @@ export default function SessionPage() {
           hideInput={selectedSession?.canWrite === false}
           display={{
             compact: false,
+            pageCanvas: true,
             showActions: true,
             showTimestamp: true,
             collapseIntermediateSteps: true,
@@ -1945,8 +2664,11 @@ export default function SessionPage() {
             processGroupsOpenWhileActive: true,
           }}
           agentName={selectedAgent}
+          executionMode={selectedExecutionMode}
+          onExecutionModeAccepted={handleExecutionModeAccepted}
           mentionAgents={chatAgents}
           className="flex-1 min-h-0"
+          composerTextareaMinHeight={56}
           initialMessage={pendingInitialMessage}
           initialDisplayText={pendingInitialDisplayText}
           onInitialMessageConsumed={() => {
@@ -1957,13 +2679,13 @@ export default function SessionPage() {
           onSSEEvent={handleSSEEvent}
           onError={handleChatError}
           onCreateAndSend={handleCreateAndSend}
-          onCreateNewSession={handleCreateSession}
+          onCreateNewSession={handleStartNewSession}
           onStreamingDone={() => {
             setPendingInitialMessage(null);
             setPendingInitialDisplayText(null);
           }}
           supportsVision={effectiveSupportsVision}
-          contextWindowTokens={selectedModelOption?.contextWindowTokens ?? null}
+          contextWindowTokens={effectiveModelOption?.contextWindowTokens ?? null}
           model={selectedPromptModel}
           welcomeContent={(setInput) => (
             <WelcomeScreen
@@ -1972,25 +2694,44 @@ export default function SessionPage() {
               alertOperationsBusy={installingSocWorkspace}
             />
           )}
-          toolbarSlot={
+          onComposerAddMenuOpenChange={(open) => {
+            setShowAgentOptions(false);
+            setShowWorkflowOptions(false);
+            setShowSkillOptions(false);
+            if (!open) return;
+            setShowProjectOptions(false);
+            setShowModelOptions(false);
+            void loadComposerResources();
+          }}
+          composerAddMenuSlot={({ closeMenu, insertMention, insertReference }) => (
+            <>
             <div className="relative" data-agent-selector>
               <button
                 type="button"
-                onClick={() => setShowAgentOptions(!showAgentOptions)}
-                className="flex h-7 w-auto max-w-[150px] min-w-0 items-center gap-1.5 rounded-lg px-2 text-xs text-zinc-600 transition-colors hover:bg-zinc-200/60 hover:text-zinc-900 dark:text-zinc-400 dark:hover:bg-zinc-800 dark:hover:text-zinc-100"
+                onClick={() => {
+                  setShowAgentOptions((open) => !open);
+                  setShowWorkflowOptions(false);
+                  setShowSkillOptions(false);
+                  setShowProjectOptions(false);
+                  setShowModelOptions(false);
+                }}
+                className="group flex h-10 w-full items-center gap-2.5 rounded-[9px] px-2 text-left text-[13px] font-medium text-zinc-700 transition-colors hover:bg-zinc-100/90 hover:text-zinc-950 dark:text-zinc-200 dark:hover:bg-white/[0.07] dark:hover:text-white"
                 title={t('agentPicker.title')}
+                aria-label={t('chat.addMenu.agent')}
+                aria-haspopup="menu"
+                aria-expanded={showAgentOptions}
               >
-                <Bot className="h-3 w-3 shrink-0" />
-                <span className="truncate font-medium">
-                  {selectedAgentInfo ? getAgentDisplayName(selectedAgentInfo, i18n.language) : formatAgentName(selectedAgent)}
+                <span className="grid h-7 w-7 shrink-0 place-items-center rounded-lg border border-zinc-200/80 bg-white text-zinc-500 shadow-[0_1px_2px_rgba(22,27,34,0.04)] transition-colors group-hover:text-zinc-800 dark:border-white/[0.10] dark:bg-white/[0.05] dark:text-zinc-400 dark:group-hover:text-white">
+                  <Bot className="h-3.5 w-3.5" />
                 </span>
-                <ChevronDown className={`h-3 w-3 shrink-0 transition-transform ${showAgentOptions ? 'rotate-180' : ''}`} />
+                <span className="min-w-0 flex-1 truncate">{t('chat.addMenu.agent')}</span>
+                <ChevronRight className={`h-3.5 w-3.5 shrink-0 text-zinc-400 transition-transform ${showAgentOptions ? '-rotate-90' : ''}`} />
               </button>
               {showAgentOptions && (
-                <div className="absolute left-0 bottom-full z-50 mb-2 w-80 max-w-[calc(100vw-2rem)] rounded-lg border border-zinc-200 bg-white shadow-sm dark:border-zinc-800 dark:bg-zinc-900 dark:shadow-xl dark:shadow-black/30">
-                  <div className="flex items-center justify-between gap-2 border-b border-zinc-100 px-2.5 py-1.5 dark:border-zinc-800">
+                <div className="absolute bottom-0 left-[calc(100%+0.5rem)] z-[60] w-[300px] max-w-[calc(100vw-2rem)] overflow-hidden rounded-[14px] border border-black/[0.09] bg-white/95 shadow-[0_18px_46px_rgba(22,27,34,0.14),0_2px_8px_rgba(22,27,34,0.06)] backdrop-blur-xl dark:border-white/[0.11] dark:bg-[#252c35]/95 dark:shadow-[0_20px_48px_rgba(8,10,13,0.44)]">
+                  <div className="flex min-h-12 items-center justify-between gap-3 border-b border-zinc-100 px-3 py-2 dark:border-white/[0.08]">
                     <div className="min-w-0">
-                      <div className="text-xs font-semibold text-zinc-700 dark:text-zinc-100">{t('agentPicker.title')}</div>
+                      <div className="text-[12px] font-semibold text-zinc-800 dark:text-zinc-100">{t('agentPicker.title')}</div>
                       <div
                         className="truncate text-[10px] text-zinc-400 dark:text-zinc-500"
                         onPointerEnter={(event) => showSelectorTooltip(event.currentTarget, t('agentPicker.title'), [t('agentPicker.hint')])}
@@ -2002,16 +2743,16 @@ export default function SessionPage() {
                         {t('agentPicker.hint')}
                       </div>
                     </div>
-                    <div className="inline-flex shrink-0 items-center rounded-md border border-zinc-200 bg-white p-0.5 text-[10px] dark:border-zinc-800 dark:bg-zinc-950">
+                    <div className="inline-flex shrink-0 items-center rounded-lg bg-zinc-100 p-0.5 text-[10px] dark:bg-black/15">
                       {(['all', 'builtin', 'custom'] as AgentSourceFilter[]).map((filter) => (
                         <button
                           key={filter}
                           type="button"
                           onClick={() => setAgentSourceFilter(filter)}
-                          className={`rounded px-1.5 py-0.5 transition-colors ${
+                          className={`rounded-md px-2 py-1 transition-colors ${
                             agentSourceFilter === filter
-                              ? 'bg-zinc-100 text-zinc-900 dark:bg-zinc-800 dark:text-zinc-50'
-                              : 'text-zinc-500 hover:bg-zinc-50 hover:text-zinc-800 dark:text-zinc-400 dark:hover:bg-zinc-800 dark:hover:text-zinc-100'
+                              ? 'bg-white text-zinc-900 shadow-[0_1px_3px_rgba(22,27,34,0.08)] dark:bg-white/[0.12] dark:text-zinc-50'
+                              : 'text-zinc-500 hover:text-zinc-800 dark:text-zinc-400 dark:hover:text-zinc-100'
                           }`}
                         >
                           {t(`agentPicker.filter.${filter}`)}
@@ -2027,37 +2768,47 @@ export default function SessionPage() {
                         const displayName = getAgentDisplayName(agent, i18n.language);
                         const primaryDesc = getAgentDisplayDescription(agent, i18n.language) || t('smartAssistant');
                         return (
-                        <button
-                          key={agent.name}
-                          onClick={() => { setSelectedAgent(agent.name); setShowAgentOptions(false); }}
-                          className={`w-full min-w-0 rounded-md px-2 py-1.5 text-left transition-colors ${
-                            selectedAgent === agent.name
-                              ? 'bg-zinc-50 text-zinc-900 shadow-[inset_2px_0_0_#a1a1aa] dark:bg-zinc-800 dark:text-zinc-50 dark:shadow-[inset_2px_0_0_#539bf5]'
-                              : 'hover:bg-zinc-50 text-zinc-700 dark:text-zinc-300 dark:hover:bg-zinc-800 dark:hover:text-zinc-50'
-                          }`}
-                        >
-                          <div className="flex min-w-0 items-center gap-2">
-                            <Bot className={`h-3 w-3 shrink-0 ${selectedAgent === agent.name ? 'text-zinc-600 dark:text-zinc-200' : 'text-zinc-400 dark:text-zinc-500'}`} />
-                            <span className="min-w-0 flex-1 truncate text-xs font-medium text-zinc-900 dark:text-zinc-100">
-                              {displayName}
-                            </span>
-                            <span className={`shrink-0 rounded px-1.5 py-0.5 text-[9px] font-medium ${
-                              agent.mode === 'primary'
-                                ? 'bg-zinc-100 text-zinc-600 dark:bg-zinc-800 dark:text-zinc-300'
-                                : agent.native
-                                  ? 'bg-zinc-100 text-zinc-600 dark:bg-zinc-800 dark:text-zinc-300'
-                                  : 'bg-teal-50 text-teal-600 dark:bg-teal-950/40 dark:text-teal-300'
-                            }`}>
-                              {agent.mode === 'primary'
-                                ? t('agentPicker.badge.primary')
-                                : agent.native
-                                  ? t('agentPicker.badge.builtin')
-                                  : t('agentPicker.badge.custom')}
-                            </span>
-                            <div className="ml-auto flex shrink-0 items-center gap-1">
+                          <button
+                            key={agent.name}
+                            type="button"
+                            onClick={() => {
+                              setShowAgentOptions(false);
+                              insertMention(agent.name);
+                              closeMenu();
+                            }}
+                            className={`w-full min-w-0 rounded-lg px-2.5 py-2 text-left transition-colors ${
+                              selectedAgent === agent.name
+                                ? 'bg-zinc-100/90 text-zinc-900 dark:bg-white/[0.09] dark:text-zinc-50'
+                                : 'text-zinc-700 hover:bg-zinc-50 dark:text-zinc-300 dark:hover:bg-white/[0.06] dark:hover:text-zinc-50'
+                            }`}
+                          >
+                            <div className="flex min-w-0 items-center gap-2">
+                              <span className={`grid h-6 w-6 shrink-0 place-items-center rounded-md ${
+                                selectedAgent === agent.name
+                                  ? 'bg-white text-zinc-700 shadow-[0_1px_3px_rgba(22,27,34,0.08)] dark:bg-white/[0.10] dark:text-white'
+                                  : 'bg-zinc-100 text-zinc-400 dark:bg-white/[0.05] dark:text-zinc-500'
+                              }`}>
+                                <Bot className="h-3 w-3" />
+                              </span>
+                              <span className="min-w-0 flex-1 truncate text-xs font-medium text-zinc-900 dark:text-zinc-100">
+                                {displayName}
+                              </span>
+                              <span className={`shrink-0 rounded px-1.5 py-0.5 text-[9px] font-medium ${
+                                agent.mode === 'primary'
+                                  ? 'bg-zinc-100 text-zinc-600 dark:bg-white/[0.08] dark:text-zinc-300'
+                                  : agent.native
+                                    ? 'bg-zinc-100 text-zinc-600 dark:bg-white/[0.08] dark:text-zinc-300'
+                                    : 'bg-teal-50 text-teal-600 dark:bg-teal-950/40 dark:text-teal-300'
+                              }`}>
+                                {agent.mode === 'primary'
+                                  ? t('agentPicker.badge.primary')
+                                  : agent.native
+                                    ? t('agentPicker.badge.builtin')
+                                    : t('agentPicker.badge.custom')}
+                              </span>
                               {primaryDesc && (
                                 <span
-                                  className="group relative rounded p-0.5 transition-colors hover:bg-zinc-200 dark:hover:bg-zinc-700"
+                                  className="group relative rounded p-0.5 transition-colors hover:bg-zinc-200 dark:hover:bg-white/[0.08]"
                                   onMouseDown={(event) => { event.preventDefault(); event.stopPropagation(); }}
                                   onClick={(event) => { event.preventDefault(); event.stopPropagation(); }}
                                   onPointerEnter={(event) => showSelectorTooltip(event.currentTarget, displayName, [primaryDesc])}
@@ -2070,8 +2821,7 @@ export default function SessionPage() {
                                 </span>
                               )}
                             </div>
-                          </div>
-                        </button>
+                          </button>
                         );
                       })
                     ) : (
@@ -2081,33 +2831,282 @@ export default function SessionPage() {
                 </div>
               )}
             </div>
+            <ComposerResourcePicker
+              label={t('chat.addMenu.skills')}
+              title={t('chat.addMenu.selectSkill')}
+              emptyText={t('chat.addMenu.noSkills')}
+              icon={BookOpen}
+              items={composerSkills.map((skill) => ({
+                id: skill.name,
+                name: skill.name,
+                description: skill.description,
+                badge: skill.category,
+                source: skill.source === 'project' ? 'builtin' : 'custom',
+              }))}
+              loading={loadingComposerResources}
+              open={showSkillOptions}
+              onToggle={() => {
+                setShowSkillOptions((open) => !open);
+                setShowAgentOptions(false);
+                setShowWorkflowOptions(false);
+              }}
+              onSelect={(skill) => {
+                insertReference(skill.id, 'skill');
+              }}
+            />
+            <ComposerResourcePicker
+              label={t('chat.addMenu.workflows')}
+              title={t('chat.addMenu.selectWorkflow')}
+              emptyText={t('chat.addMenu.noWorkflows')}
+              icon={WorkflowIcon}
+              items={composerWorkflows.map((workflow) => ({
+                id: workflow.id,
+                name: getWorkflowDisplayName(workflow, i18n.language),
+                description: workflow.description,
+                source: workflow.source === 'project' ? 'builtin' : 'custom',
+              }))}
+              loading={loadingComposerResources}
+              open={showWorkflowOptions}
+              onToggle={() => {
+                setShowWorkflowOptions((open) => !open);
+                setShowAgentOptions(false);
+                setShowSkillOptions(false);
+              }}
+              onSelect={(workflow) => {
+                insertReference(workflow.id, 'workflow');
+                setShowWorkflowOptions(false);
+                closeMenu();
+              }}
+            />
+            </>
+          )}
+          toolbarSlot={
+            <div className="flex min-w-0 items-center gap-0.5">
+              <div className="relative" data-execution-mode-selector>
+                <button
+                  type="button"
+                  onClick={() => {
+                    setShowExecutionModeOptions(!showExecutionModeOptions);
+                    setShowAgentOptions(false);
+                    setShowModelOptions(false);
+                    setShowProjectOptions(false);
+                  }}
+                  className="flex h-7 items-center gap-1.5 rounded-lg px-2 text-xs text-zinc-600 transition-colors hover:bg-zinc-200/60 hover:text-zinc-900 dark:text-zinc-400 dark:hover:bg-zinc-800 dark:hover:text-zinc-100"
+                  title={t(`executionMode.options.${selectedExecutionMode}.description`)}
+                  aria-label={t('executionMode.title')}
+                  aria-haspopup="menu"
+                  aria-expanded={showExecutionModeOptions}
+                >
+                  <ExecutionModeIcon mode={selectedExecutionMode} />
+                  <span className="font-medium">
+                    {t(`executionMode.options.${selectedExecutionMode}.label`)}
+                  </span>
+                  <ChevronDown className={`h-3 w-3 shrink-0 transition-transform ${showExecutionModeOptions ? 'rotate-180' : ''}`} />
+                </button>
+                {showExecutionModeOptions && (
+                  <div
+                    role="menu"
+                    aria-label={t('executionMode.title')}
+                    className="absolute bottom-full left-0 z-50 mb-2 w-64 overflow-hidden rounded-lg border border-zinc-200 bg-white shadow-sm dark:border-zinc-800 dark:bg-zinc-900 dark:shadow-xl dark:shadow-black/30"
+                  >
+                    <div className="border-b border-zinc-100 px-3 py-2 dark:border-zinc-800">
+                      <div className="text-xs font-semibold text-zinc-700 dark:text-zinc-100">
+                        {t('executionMode.title')}
+                      </div>
+                      <div className="mt-0.5 text-[10px] text-zinc-400 dark:text-zinc-500">
+                        {t('executionMode.hint')}
+                      </div>
+                    </div>
+                    <div className="space-y-0.5 p-1.5">
+                      {SESSION_EXECUTION_MODES.map((mode) => {
+                        const selected = selectedExecutionMode === mode;
+                        return (
+                          <button
+                            key={mode}
+                            type="button"
+                            role="menuitemradio"
+                            aria-checked={selected}
+                            onClick={() => handleSelectExecutionMode(mode)}
+                            className={`flex w-full items-start gap-2 rounded-md px-2 py-2 text-left transition-colors ${
+                              selected
+                                ? 'bg-zinc-50 text-zinc-900 shadow-[inset_2px_0_0_#a1a1aa] dark:bg-zinc-800 dark:text-zinc-50 dark:shadow-[inset_2px_0_0_#539bf5]'
+                                : 'text-zinc-700 hover:bg-zinc-50 dark:text-zinc-300 dark:hover:bg-zinc-800 dark:hover:text-zinc-50'
+                            }`}
+                          >
+                            <ExecutionModeIcon
+                              mode={mode}
+                              className={`mt-0.5 h-3.5 w-3.5 shrink-0 ${
+                                selected
+                                  ? 'text-zinc-700 dark:text-zinc-100'
+                                  : 'text-zinc-400 dark:text-zinc-500'
+                              }`}
+                            />
+                            <span className="min-w-0 flex-1">
+                              <span className="block text-xs font-medium">
+                                {t(`executionMode.options.${mode}.label`)}
+                              </span>
+                              <span className="mt-0.5 block text-[10px] leading-4 text-zinc-400 dark:text-zinc-500">
+                                {t(`executionMode.options.${mode}.description`)}
+                              </span>
+                            </span>
+                            {selected && <Check className="mt-0.5 h-3.5 w-3.5 shrink-0" />}
+                          </button>
+                        );
+                      })}
+                    </div>
+                  </div>
+                )}
+              </div>
+              {!activeChatSessionId && (
+              <div className="relative" data-project-selector>
+                <button
+                  type="button"
+                  onClick={() => {
+                    setShowProjectOptions((open) => !open);
+                    setShowAgentOptions(false);
+                    setShowExecutionModeOptions(false);
+                    setShowModelOptions(false);
+                  }}
+                  className="flex h-7 max-w-[160px] min-w-0 items-center gap-1.5 rounded-lg px-2 text-xs text-[#5e646b] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#b8c2cc] dark:hover:bg-white/[0.06] dark:hover:text-white"
+                  title={t('projectPicker.title')}
+                  aria-label={t('projectPicker.title')}
+                  aria-expanded={showProjectOptions}
+                >
+                  <FolderGit2 className="h-3 w-3 shrink-0" />
+                  <span className="truncate font-medium">{selectedProjectContextLabel}</span>
+                  <ChevronDown className={`h-3 w-3 shrink-0 transition-transform ${showProjectOptions ? 'rotate-180' : ''}`} />
+                </button>
+                {showProjectOptions && (
+                  <div
+                    className="absolute bottom-full left-0 z-50 mb-2 grid w-64 gap-0.5 rounded-[10px] border border-black/[0.11] bg-[#fdfdfc] p-1 shadow-[0_8px_24px_rgba(22,27,34,0.10)] dark:border-white/[0.10] dark:bg-[#303842] dark:shadow-[0_12px_32px_rgba(15,18,22,0.35)]"
+                    role="menu"
+                    aria-label={t('projectPicker.title')}
+                  >
+                    {projectSessionGroups.map((group) => (
+                      <button
+                        key={group.id}
+                        type="button"
+                        role="menuitemradio"
+                        aria-checked={selectedProjectId === group.id}
+                        onClick={() => {
+                          setSelectedProjectId(group.id);
+                          setShowProjectOptions(false);
+                        }}
+                        className="flex h-8 w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
+                      >
+                        <FolderGit2 className="h-3.5 w-3.5 shrink-0 text-[#7b8087] dark:text-[#9aa7b4]" />
+                        <span className="min-w-0 flex-1 truncate">{group.label}</span>
+                        {selectedProjectId === group.id && <Check className="h-3.5 w-3.5 shrink-0 text-[#4f92e8]" />}
+                      </button>
+                    ))}
+                    <div className="mx-2 my-0.5 border-t border-black/[0.07] dark:border-white/[0.08]" />
+                    <button
+                      type="button"
+                      role="menuitemradio"
+                      aria-checked={selectedProjectId === TASK_SESSION_GROUP_ID}
+                      onClick={() => {
+                        setSelectedProjectId(TASK_SESSION_GROUP_ID);
+                        setShowProjectOptions(false);
+                      }}
+                      className="flex h-8 w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
+                    >
+                      <X className="h-3.5 w-3.5 shrink-0 text-[#7b8087] dark:text-[#9aa7b4]" />
+                      <span className="min-w-0 flex-1 truncate">{t('projectPicker.none')}</span>
+                      {selectedProjectId === TASK_SESSION_GROUP_ID && <Check className="h-3.5 w-3.5 shrink-0 text-[#4f92e8]" />}
+                    </button>
+                    <button
+                      type="button"
+                      role="menuitem"
+                      onClick={() => {
+                        setShowProjectOptions(false);
+                        handleOpenCreateProject();
+                      }}
+                      className="flex h-8 w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
+                    >
+                      <FolderPlus className="h-3.5 w-3.5 shrink-0 text-[#7b8087] dark:text-[#9aa7b4]" />
+                      <span>{t('projectPicker.create')}</span>
+                    </button>
+                  </div>
+                )}
+              </div>
+            )}
+            </div>
           }
           centerToolbarSlot={
-            <div className="relative" data-model-selector>
+            <div ref={modelSelectorRef} className="relative" data-model-selector>
               <button
                 type="button"
-                onClick={() => setShowModelOptions(!showModelOptions)}
+                onClick={() => {
+                  if (!showModelOptions) updateModelMenuLeftOffset();
+                  setShowModelOptions(!showModelOptions);
+                  setShowExecutionModeOptions(false);
+                  setShowProjectOptions(false);
+                  setShowAgentOptions(false);
+                }}
                 disabled={loadingProviders || loadingEnabledModels || chatModelOptions.length === 0}
                 className="flex h-7 w-[132px] min-w-0 items-center gap-1.5 rounded-lg px-2 text-xs text-zinc-600 transition-colors hover:bg-zinc-200/60 hover:text-zinc-900 disabled:cursor-not-allowed disabled:opacity-50 dark:text-zinc-400 dark:hover:bg-zinc-800 dark:hover:text-zinc-100"
-                title={selectedModelOption ? `${selectedModelOption.providerName} / ${selectedModelOption.modelID}` : t('modelPicker.empty')}
+                title={selectedModelAuto
+                  ? `${t('modelPicker.auto')}: ${autoStatusLabel}`
+                  : selectedModelOption
+                    ? `${selectedModelOption.providerName} / ${selectedModelOption.modelID}`
+                    : t('modelPicker.empty')}
               >
-                <Cpu className="h-3 w-3 shrink-0" />
+                {selectedModelAuto
+                  ? <Sparkles className="h-3 w-3 shrink-0" />
+                  : <Cpu className="h-3 w-3 shrink-0" />}
                 <span className="truncate font-medium">
-                  {selectedModelOption?.label ?? (loadingProviders || loadingEnabledModels ? t('loading') : t('modelPicker.empty'))}
+                  {selectedModelAuto
+                    ? t('modelPicker.auto')
+                    : selectedModelOption?.label ?? (loadingProviders || loadingEnabledModels ? t('loading') : t('modelPicker.empty'))}
                 </span>
                 <ChevronDown className={`h-3 w-3 shrink-0 transition-transform ${showModelOptions ? 'rotate-180' : ''}`} />
               </button>
               {showModelOptions && (
-                <div className="absolute right-0 bottom-full z-50 mb-2 w-80 max-w-[calc(100vw-2rem)] rounded-lg border border-zinc-200 bg-white shadow-sm dark:border-zinc-800 dark:bg-zinc-900 dark:shadow-xl dark:shadow-black/30">
+                <div
+                  className="absolute left-0 bottom-full z-50 mb-2 w-80 max-w-[calc(100vw-2rem)] rounded-lg border border-zinc-200 bg-white shadow-sm dark:border-zinc-800 dark:bg-zinc-900 dark:shadow-xl dark:shadow-black/30"
+                  style={{ transform: `translateX(${modelMenuLeftOffset}px)` }}
+                >
                   <div className="border-b border-zinc-100 px-2.5 py-1.5 dark:border-zinc-800">
                     <div className="text-xs font-semibold text-zinc-700 dark:text-zinc-100">{t('modelPicker.title')}</div>
                     <div className="truncate text-[10px] text-zinc-400 dark:text-zinc-500">{t('modelPicker.hint')}</div>
                   </div>
-                  <div className="h-[13.5rem] overflow-y-auto p-1.5">
+                  <div className="h-[15.5rem] overflow-y-auto p-1.5">
                     {loadingProviders || loadingEnabledModels ? (
                       <div className="p-3 text-center text-xs text-zinc-500">{t('loading')}</div>
-                    ) : groupedChatModelOptions.length > 0 ? (
-                      groupedChatModelOptions.map((group) => (
+                    ) : (
+                      <>
+                        <div className="mb-1 border-b border-zinc-100 pb-1.5 dark:border-zinc-800">
+                          <button
+                            type="button"
+                            onClick={() => void handleSelectAutoModel()}
+                            disabled={!canSelectAuto}
+                            className={`w-full rounded-md px-2 py-1.5 text-left transition-colors disabled:cursor-not-allowed disabled:opacity-45 ${
+                              selectedModelAuto
+                                ? 'bg-zinc-50 text-zinc-900 shadow-[inset_2px_0_0_#a1a1aa] dark:bg-zinc-800 dark:text-zinc-50 dark:shadow-[inset_2px_0_0_#539bf5]'
+                                : 'text-zinc-700 hover:bg-zinc-50 dark:text-zinc-300 dark:hover:bg-zinc-800 dark:hover:text-zinc-50'
+                            }`}
+                          >
+                            <div className="flex min-w-0 items-center gap-2">
+                              <Sparkles className={`h-3 w-3 shrink-0 ${selectedModelAuto ? 'text-zinc-600 dark:text-zinc-200' : 'text-zinc-400 dark:text-zinc-500'}`} />
+                              <span className="min-w-0 flex-1 truncate text-xs font-medium text-zinc-900 dark:text-zinc-100">
+                                {t('modelPicker.auto')}
+                              </span>
+                              <span
+                                className="group relative rounded p-0.5 transition-colors hover:bg-zinc-200 dark:hover:bg-zinc-700"
+                                onMouseDown={(event) => { event.preventDefault(); event.stopPropagation(); }}
+                                onClick={(event) => { event.preventDefault(); event.stopPropagation(); }}
+                                onPointerEnter={(event) => showSelectorTooltip(event.currentTarget, t('modelPicker.auto'), [autoStatusLabel])}
+                                onMouseEnter={(event) => showSelectorTooltip(event.currentTarget, t('modelPicker.auto'), [autoStatusLabel])}
+                                onMouseOver={(event) => showSelectorTooltip(event.currentTarget, t('modelPicker.auto'), [autoStatusLabel])}
+                                onMouseLeave={() => setSelectorTooltip(null)}
+                                onPointerLeave={() => setSelectorTooltip(null)}
+                              >
+                                <Info className="h-3 w-3 text-zinc-300 transition-colors group-hover:text-zinc-500 dark:text-zinc-600 dark:group-hover:text-zinc-300" />
+                              </span>
+                            </div>
+                          </button>
+                        </div>
+                        {groupedChatModelOptions.length > 0 ? groupedChatModelOptions.map((group) => (
                         <div key={group.providerID} className="py-1 first:pt-0 last:pb-0">
                           <div className="sticky top-0 z-10 flex items-center justify-between gap-2 bg-white/95 px-1.5 py-1 text-[10px] font-semibold text-zinc-500 backdrop-blur dark:bg-zinc-900/95 dark:text-zinc-400">
                             <span className="truncate">{group.providerName}</span>
@@ -2154,9 +3153,10 @@ export default function SessionPage() {
                             ))}
                           </div>
                         </div>
-                      ))
-                    ) : (
-                      <div className="p-3 text-center text-xs text-zinc-500">{t('modelPicker.empty')}</div>
+                        )) : (
+                          <div className="p-3 text-center text-xs text-zinc-500">{t('modelPicker.empty')}</div>
+                        )}
+                      </>
                     )}
                   </div>
                   <div className="border-t border-zinc-100 p-1.5 dark:border-zinc-800">
@@ -2182,15 +3182,15 @@ export default function SessionPage() {
       </div>
 
       {projectDialogMode && (
-        <div className="fixed inset-0 z-[70] flex items-center justify-center bg-black/30 px-4 backdrop-blur-sm">
-          <div className="w-full max-w-lg rounded-lg border border-zinc-200 bg-white shadow-xl dark:border-zinc-800 dark:bg-zinc-950">
-            <div className="border-b border-zinc-100 px-4 py-3 dark:border-zinc-800">
-              <h3 className="text-sm font-semibold text-zinc-900 dark:text-zinc-100">
+        <div className="fixed inset-0 z-[70] flex items-center justify-center bg-[#1e23283d] px-6 backdrop-blur-[2px]">
+          <div className="w-full max-w-[420px] overflow-hidden rounded-2xl border border-black/[0.11] bg-[#fdfdfc] shadow-[0_8px_24px_rgba(22,27,34,0.10)] dark:border-white/[0.10] dark:bg-[#303842] dark:shadow-[0_12px_32px_rgba(15,18,22,0.35)]">
+            <div className="border-b border-black/[0.07] px-4 py-[13px] dark:border-white/[0.08]">
+              <h3 className="text-sm font-semibold text-[#202328] dark:text-[#f0f3f7]">
                 {projectDialogMode === 'create' ? t('projectDialog.createTitle') : t('projectDialog.renameTitle')}
               </h3>
             </div>
-            <div className="space-y-3 px-4 py-4">
-              <label className="block text-xs font-medium text-zinc-500 dark:text-zinc-400" htmlFor="session-project-name">
+            <div className="space-y-3 px-4 py-4 text-[13px] text-[#676c73] dark:text-[#b8c2cc]">
+              <label className="block text-[11px] font-semibold text-[#777c82] dark:text-[#9aa7b4]" htmlFor="session-project-name">
                 {t('projectDialog.nameLabel')}
               </label>
               <input
@@ -2220,12 +3220,12 @@ export default function SessionPage() {
                 }}
                 disabled={projectSubmitting}
                 placeholder={t('projectDialog.namePlaceholder')}
-                className="w-full rounded-lg border border-zinc-200 bg-white px-3 py-2 text-sm text-zinc-900 outline-none transition-colors placeholder:text-zinc-400 focus:border-blue-400 dark:border-zinc-800 dark:bg-zinc-900 dark:text-zinc-100 dark:placeholder:text-zinc-600 dark:focus:border-blue-500"
+                className="h-9 w-full rounded-[10px] border border-black/[0.11] bg-white px-2.5 text-[13px] text-[#202328] outline-none transition-colors placeholder:text-[#969aa0] focus:border-[#4f92e8] dark:border-white/[0.10] dark:bg-[#252c35] dark:text-white dark:placeholder:text-[#8f9ba8]"
                 autoFocus
               />
               {projectDialogMode === 'create' && (
                 <>
-                  <label className="block pt-1 text-xs font-medium text-zinc-500 dark:text-zinc-400" htmlFor="session-project-worktree">
+                  <label className="block pt-1 text-[11px] font-semibold text-[#777c82] dark:text-[#9aa7b4]" htmlFor="session-project-worktree">
                     {t('projectDialog.folderLabel')}
                   </label>
                   <div className="flex gap-2">
@@ -2243,13 +3243,13 @@ export default function SessionPage() {
                       }}
                       disabled={projectSubmitting}
                       placeholder={t('projectDialog.folderPlaceholder')}
-                      className="min-w-0 flex-1 rounded-lg border border-zinc-200 bg-white px-3 py-2 font-mono text-xs text-zinc-900 outline-none transition-colors placeholder:text-zinc-400 focus:border-blue-400 dark:border-zinc-800 dark:bg-zinc-900 dark:text-zinc-100 dark:placeholder:text-zinc-600 dark:focus:border-blue-500"
+                      className="h-9 min-w-0 flex-1 rounded-[10px] border border-black/[0.11] bg-white px-2.5 font-mono text-xs text-[#202328] outline-none transition-colors placeholder:text-[#969aa0] focus:border-[#4f92e8] dark:border-white/[0.10] dark:bg-[#252c35] dark:text-white dark:placeholder:text-[#8f9ba8]"
                     />
                     <button
                       type="button"
                       onClick={() => void loadFolderBrowser(projectWorktreeValue)}
                       disabled={projectSubmitting || folderBrowserLoading}
-                      className="inline-flex shrink-0 items-center gap-1.5 rounded-lg border border-zinc-200 px-3 text-xs font-medium text-zinc-600 hover:bg-zinc-50 disabled:cursor-not-allowed disabled:opacity-50 dark:border-zinc-800 dark:text-zinc-300 dark:hover:bg-zinc-900"
+                      className="inline-flex h-9 shrink-0 items-center gap-1.5 rounded-[10px] border border-black/[0.11] px-3 text-xs font-medium text-[#60656b] transition-colors hover:bg-black/[0.04] disabled:cursor-not-allowed disabled:opacity-50 dark:border-white/[0.10] dark:text-[#c3ccd6] dark:hover:bg-white/[0.06]"
                     >
                       {folderBrowserLoading ? <Loader2 className="h-3.5 w-3.5 animate-spin" /> : <FolderOpen className="h-3.5 w-3.5" />}
                       {t('projectDialog.chooseFolder')}
@@ -2315,12 +3315,12 @@ export default function SessionPage() {
                 </>
               )}
             </div>
-            <div className="flex items-center justify-end gap-2 border-t border-zinc-100 px-4 py-3 dark:border-zinc-800">
+            <div className="flex items-center justify-end gap-2 border-t border-black/[0.07] px-4 py-[13px] dark:border-white/[0.08]">
               <button
                 type="button"
                 onClick={handleCloseProjectDialog}
                 disabled={projectSubmitting}
-                className="rounded-lg px-3 py-1.5 text-sm text-zinc-500 transition-colors hover:bg-zinc-100 hover:text-zinc-800 disabled:cursor-not-allowed disabled:opacity-50 dark:text-zinc-400 dark:hover:bg-zinc-900 dark:hover:text-zinc-100"
+                className="h-8 min-w-[66px] rounded-[10px] bg-[#f1f1ef] px-3 text-xs text-[#5d6269] transition-colors hover:bg-[#e8e8e5] disabled:cursor-not-allowed disabled:opacity-50 dark:bg-white/[0.07] dark:text-[#c3ccd6] dark:hover:bg-white/[0.10]"
               >
                 {t('cancel')}
               </button>
@@ -2328,7 +3328,7 @@ export default function SessionPage() {
                 type="button"
                 onClick={() => void handleSubmitProject()}
                 disabled={projectSubmitting}
-                className="inline-flex min-w-16 items-center justify-center rounded-lg bg-zinc-900 px-3 py-1.5 text-sm font-medium text-white transition-colors hover:bg-zinc-700 disabled:cursor-not-allowed disabled:opacity-60 dark:bg-zinc-100 dark:text-zinc-950 dark:hover:bg-zinc-300"
+                className="inline-flex h-8 min-w-[66px] items-center justify-center rounded-[10px] bg-[#24272b] px-3 text-xs font-medium text-white transition-colors hover:bg-[#35393e] disabled:cursor-not-allowed disabled:opacity-60 dark:bg-[#eef2f6] dark:text-[#252c35] dark:hover:bg-white"
               >
                 {projectSubmitting ? <Loader2 className="h-4 w-4 animate-spin" /> : t('save')}
               </button>
@@ -2338,22 +3338,22 @@ export default function SessionPage() {
       )}
 
       {projectPendingDelete && (
-        <div className="fixed inset-0 z-[70] flex items-center justify-center bg-black/30 px-4 backdrop-blur-sm">
-          <div className="w-full max-w-sm rounded-lg border border-zinc-200 bg-white shadow-xl dark:border-zinc-800 dark:bg-zinc-950">
-            <div className="border-b border-zinc-100 px-4 py-3 dark:border-zinc-800">
-              <h3 className="text-sm font-semibold text-zinc-900 dark:text-zinc-100">
+        <div className="fixed inset-0 z-[70] flex items-center justify-center bg-[#1e23283d] px-6 backdrop-blur-[2px]">
+          <div className="w-full max-w-[420px] overflow-hidden rounded-2xl border border-black/[0.11] bg-[#fdfdfc] shadow-[0_8px_24px_rgba(22,27,34,0.10)] dark:border-white/[0.10] dark:bg-[#303842] dark:shadow-[0_12px_32px_rgba(15,18,22,0.35)]">
+            <div className="border-b border-black/[0.07] px-4 py-[13px] dark:border-white/[0.08]">
+              <h3 className="text-sm font-semibold text-[#202328] dark:text-[#f0f3f7]">
                 {t('projectDialog.deleteTitle')}
               </h3>
             </div>
-            <div className="px-4 py-4 text-sm leading-6 text-zinc-600 dark:text-zinc-300">
+            <div className="px-4 py-4 text-[13px] leading-6 text-[#676c73] dark:text-[#b8c2cc]">
               {t('projectDialog.deleteDescription', { project: projectPendingDelete.label })}
             </div>
-            <div className="flex items-center justify-end gap-2 border-t border-zinc-100 px-4 py-3 dark:border-zinc-800">
+            <div className="flex items-center justify-end gap-2 border-t border-black/[0.07] px-4 py-[13px] dark:border-white/[0.08]">
               <button
                 type="button"
                 onClick={() => setProjectPendingDelete(null)}
                 disabled={projectDeleting}
-                className="rounded-lg px-3 py-1.5 text-sm text-zinc-500 transition-colors hover:bg-zinc-100 hover:text-zinc-800 disabled:cursor-not-allowed disabled:opacity-50 dark:text-zinc-400 dark:hover:bg-zinc-900 dark:hover:text-zinc-100"
+                className="h-8 min-w-[66px] rounded-[10px] bg-[#f1f1ef] px-3 text-xs text-[#5d6269] transition-colors hover:bg-[#e8e8e5] disabled:cursor-not-allowed disabled:opacity-50 dark:bg-white/[0.07] dark:text-[#c3ccd6] dark:hover:bg-white/[0.10]"
               >
                 {t('cancel')}
               </button>
@@ -2361,7 +3361,7 @@ export default function SessionPage() {
                 type="button"
                 onClick={() => void handleDeleteProject()}
                 disabled={projectDeleting}
-                className="inline-flex min-w-16 items-center justify-center gap-1.5 rounded-lg bg-red-600 px-3 py-1.5 text-sm font-medium text-white transition-colors hover:bg-red-500 disabled:cursor-not-allowed disabled:opacity-60"
+                className="inline-flex h-8 min-w-[66px] items-center justify-center gap-1.5 rounded-[10px] bg-[#dc3f39] px-3 text-xs font-medium text-white transition-colors hover:bg-[#c9342f] disabled:cursor-not-allowed disabled:opacity-60"
               >
                 {projectDeleting ? <Loader2 className="h-4 w-4 animate-spin" /> : <Trash2 className="h-4 w-4" />}
                 {t('projectDialog.confirmDelete')}
@@ -2399,16 +3399,84 @@ export default function SessionPage() {
         const sid = openMenuSessionId;
         const session = sessions.find(s => s.id === sid);
         if (!session) return null;
+        const currentProjectId = session.effectiveProjectID || session.projectID;
+        const showMovePicker = movePickerSessionId === session.id;
+        const moveTargetProjects = projects.filter((project) => (
+          project.id.startsWith('prj_')
+          && project.canWrite !== false
+          && project.pathStatus !== 'missing'
+          && project.pathStatus !== 'unreadable'
+        ));
+        const movePickerHeight = Math.min(
+          272,
+          48 + Math.max(44, moveTargetProjects.length * 34),
+        );
+        const menuTop = showMovePicker
+          ? Math.max(8, Math.min(menuAnchor.top, window.innerHeight - movePickerHeight - 8))
+          : menuAnchor.top;
         return (
           <div
-            className="fixed z-50 min-w-28 rounded-md border border-zinc-200 bg-white p-1 shadow-lg dark:border-zinc-700 dark:bg-zinc-900"
-            style={{ top: menuAnchor.top, right: menuAnchor.right }}
+            className={`fixed z-50 grid gap-0.5 overflow-hidden rounded-[10px] border border-black/[0.11] bg-[#fdfdfc] p-1 shadow-[0_8px_24px_rgba(22,27,34,0.10)] dark:border-white/[0.10] dark:bg-[#303842] dark:shadow-[0_12px_32px_rgba(15,18,22,0.35)] ${
+              showMovePicker ? 'w-[220px]' : 'w-[132px]'
+            }`}
+            style={{
+              top: menuTop,
+              right: menuAnchor.right,
+              maxHeight: showMovePicker ? 'calc(100vh - 16px)' : undefined,
+            }}
             data-session-menu-portal
             onClick={(e) => e.stopPropagation()}
           >
+            {showMovePicker ? (
+              <>
+                <div className="flex h-8 items-center gap-1 border-b border-black/[0.07] px-1 pb-1 dark:border-white/[0.08]">
+                  <button
+                    type="button"
+                    onClick={() => setMovePickerSessionId(null)}
+                    className="grid h-6 w-6 shrink-0 place-items-center rounded-md text-[#6f757c] transition-colors hover:bg-black/[0.05] hover:text-[#202328] dark:text-[#aeb8c3] dark:hover:bg-white/[0.07] dark:hover:text-white"
+                    aria-label={t('moveToProjectBack')}
+                  >
+                    <ChevronLeft className="h-3.5 w-3.5" />
+                  </button>
+                  <span className="min-w-0 truncate text-xs font-semibold text-[#34383d] dark:text-[#e3e8ee]">
+                    {t('moveToProjectTitle')}
+                  </span>
+                </div>
+                <div className="session-sidebar-scrollbar min-h-0 max-h-56 overflow-y-auto py-0.5" role="menu" aria-label={t('moveToProjectTitle')}>
+                  {moveTargetProjects.length === 0 ? (
+                    <div className="px-2 py-3 text-center text-xs leading-5 text-[#858a91] dark:text-[#8f9ba8]">
+                      {t('moveToProjectEmpty')}
+                    </div>
+                  ) : moveTargetProjects.map((project) => {
+                    const isCurrent = project.id === currentProjectId;
+                    const isMovingHere = (
+                      movingSessionId === session.id
+                      && movingTargetProjectId === project.id
+                    );
+                    return (
+                      <button
+                        key={project.id}
+                        type="button"
+                        role="menuitem"
+                        onClick={() => void handleMoveSessionToProject(session.id, project)}
+                        disabled={isCurrent || Boolean(movingSessionId)}
+                        title={project.worktree}
+                        className="flex min-h-[34px] w-full items-center gap-2 rounded-lg px-2 py-1.5 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] disabled:cursor-default disabled:opacity-60 dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
+                      >
+                        <FolderGit2 className="h-3.5 w-3.5 shrink-0 text-[#7b8087] dark:text-[#9aa7b4]" />
+                        <span className="min-w-0 flex-1 truncate">{getProjectLabel(project)}</span>
+                        {isCurrent && <Check className="h-3.5 w-3.5 shrink-0 text-blue-500" />}
+                        {!isCurrent && isMovingHere && <Loader2 className="h-3.5 w-3.5 shrink-0 animate-spin text-blue-500" />}
+                      </button>
+                    );
+                  })}
+                </div>
+              </>
+            ) : (
+              <>
             <button
               onClick={(e) => { e.stopPropagation(); handleStartRename(session.id, session.title); setOpenMenuSessionId(null); setMenuAnchor(null); }}
-              className="flex w-full items-center gap-2 rounded px-2 py-1.5 text-left text-xs text-zinc-700 transition-colors hover:bg-zinc-100 dark:text-zinc-200 dark:hover:bg-zinc-800"
+              className="flex h-[30px] w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
             >
               <PencilLine className="w-3.5 h-3.5" />
               <span>{t('rename')}</span>
@@ -2416,7 +3484,7 @@ export default function SessionPage() {
             <button
               onClick={(e) => { e.stopPropagation(); void handleDownloadSession(session.id, session.title); setOpenMenuSessionId(null); setMenuAnchor(null); }}
               disabled={downloadingSessionId === session.id}
-              className="flex w-full items-center gap-2 rounded px-2 py-1.5 text-left text-xs text-zinc-700 transition-colors hover:bg-zinc-100 disabled:cursor-not-allowed disabled:opacity-50 dark:text-zinc-200 dark:hover:bg-zinc-800"
+              className="flex h-[30px] w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] disabled:cursor-not-allowed disabled:opacity-50 dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
             >
               <Download className="w-3.5 h-3.5" />
               <span>{t('downloadJson')}</span>
@@ -2424,20 +3492,31 @@ export default function SessionPage() {
             <button
               onClick={(e) => { e.stopPropagation(); setOpenMenuSessionId(null); setMenuAnchor(null); void handleShareSession(session.id, !session.isShared); }}
               disabled={session.canWrite === false}
-              className="flex w-full items-center gap-2 rounded px-2 py-1.5 text-left text-xs text-zinc-700 transition-colors hover:bg-zinc-100 disabled:cursor-not-allowed disabled:opacity-50 dark:text-zinc-200 dark:hover:bg-zinc-800"
+              className="flex h-[30px] w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] disabled:cursor-not-allowed disabled:opacity-50 dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
             >
               <Share2 className="w-3.5 h-3.5" />
               <span>{session.isShared ? t('unshareAction') : t('shareAction')}</span>
             </button>
-            <div className="mx-2 my-1 border-t border-zinc-100 dark:border-zinc-800" />
             <button
-              onClick={(e) => { e.stopPropagation(); setOpenMenuSessionId(null); setMenuAnchor(null); void handleDeleteSession(session.id); }}
+              onClick={(e) => { e.stopPropagation(); setMovePickerSessionId(session.id); }}
+              disabled={session.canWrite === false || runningSessionIds.has(session.id)}
+              title={runningSessionIds.has(session.id) ? t('moveToProjectRunning') : undefined}
+              className="flex h-[30px] w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#4e5359] transition-colors hover:bg-black/[0.04] hover:text-[#202328] disabled:cursor-not-allowed disabled:opacity-50 dark:text-[#c3ccd6] dark:hover:bg-white/[0.06] dark:hover:text-white"
+            >
+              <FolderOpen className="h-3.5 w-3.5" />
+              <span className="truncate">{t('moveToProjectAction')}</span>
+            </button>
+            <div className="mx-2 my-0.5 border-t border-black/[0.07] dark:border-white/[0.08]" />
+            <button
+              onClick={(e) => { e.stopPropagation(); setOpenMenuSessionId(null); setMenuAnchor(null); void handleArchiveSession(session.id); }}
               disabled={session.canDelete === false}
-              className="flex w-full items-center gap-2 rounded px-2 py-1.5 text-left text-xs text-red-600 transition-colors hover:bg-red-50 disabled:cursor-not-allowed disabled:opacity-50 dark:text-red-300 dark:hover:bg-red-950/40"
+              className="flex h-[30px] w-full items-center gap-2 rounded-lg px-2 text-left text-xs text-[#c33c36] transition-colors hover:bg-[#fff0ef] hover:text-[#a92520] disabled:cursor-not-allowed disabled:opacity-50 dark:text-red-300 dark:hover:bg-red-500/10 dark:hover:text-red-200"
             >
-              <Trash2 className="w-3.5 h-3.5" />
-              <span>{t('deleteAction')}</span>
+              <Archive className="w-3.5 h-3.5" />
+              <span>{t('archiveAction')}</span>
             </button>
+              </>
+            )}
           </div>
         );
       })()}
@@ -2458,39 +3537,39 @@ function WelcomeScreen({
 }) {
   const { t } = useTranslation('session');
   return (
-    <div className="text-center max-w-2xl px-8">
-      <div className="w-20 h-20 mx-auto mb-6 rounded-full bg-gradient-to-br from-slate-700 to-slate-900 flex items-center justify-center shadow-lg">
-        <Sparkles className="w-10 h-10 text-white" />
+    <div className="w-full max-w-[660px] px-9 text-center">
+      <div className="mx-auto mb-5 flex h-16 w-16 items-center justify-center rounded-full bg-slate-800 text-white shadow-[0_3px_10px_rgba(30,41,59,0.10)] dark:bg-[#46515e]">
+        <Sparkles className="h-[30px] w-[30px]" strokeWidth={1.7} />
       </div>
-      <h3 className="text-xl font-bold text-gray-900 mb-3 dark:text-zinc-50">{t('welcome.title')}</h3>
-      <p className="text-sm text-gray-600 mb-8 dark:text-zinc-400">{t('welcome.description')}</p>
+      <h3 className="text-[22px] font-semibold tracking-[-0.01em] text-[#1f2937] dark:text-[#f0f3f7]">{t('welcome.title')}</h3>
+      <p className="mt-3 text-sm leading-6 text-[#647084] dark:text-[#aeb9c5]">{t('welcome.description')}</p>
 
-      <div className="flex flex-wrap gap-3 justify-center">
+      <div className="mt-6 flex flex-wrap justify-center gap-2.5">
         <button
           onClick={onAlertOperationsSetup}
           disabled={alertOperationsBusy}
-          className="flex items-center gap-2 px-5 py-3 bg-white border-2 border-gray-200 rounded-xl hover:border-slate-400 hover:bg-slate-50 transition-all duration-200 shadow-sm hover:shadow-md disabled:cursor-not-allowed disabled:opacity-70 dark:border-zinc-800 dark:bg-zinc-900 dark:hover:border-slate-500/70 dark:hover:bg-zinc-800 dark:hover:shadow-none"
+          className="inline-flex h-11 items-center gap-2.5 rounded-[10px] border border-black/[0.11] bg-zinc-50 px-4 text-sm font-semibold text-[#374151] transition-colors hover:border-[#c2c9d2] hover:bg-white hover:text-[#111827] disabled:cursor-not-allowed disabled:opacity-70 dark:border-white/[0.10] dark:bg-[#303842] dark:text-[#d7dee8] dark:hover:border-white/[0.18] dark:hover:bg-[#3a434e] dark:hover:text-white"
         >
           {alertOperationsBusy ? (
-            <Loader2 className="w-5 h-5 animate-spin text-slate-600" />
+            <Loader2 className="h-[18px] w-[18px] animate-spin text-[#657489] dark:text-[#aeb9c5]" />
           ) : (
-            <Shield className="w-5 h-5 text-slate-600" />
+            <Shield className="h-[18px] w-[18px] text-[#657489] dark:text-[#aeb9c5]" />
           )}
-          <span className="font-medium text-gray-700 dark:text-zinc-200">{t('welcome.alertOperations')}</span>
+          <span>{t('welcome.alertOperations')}</span>
         </button>
         <button
           onClick={() => onSuggestion(t('welcome.threatHuntingSuggestion'))}
-          className="flex items-center gap-2 px-5 py-3 bg-white border-2 border-gray-200 rounded-xl hover:border-orange-400 hover:bg-orange-50 transition-all duration-200 shadow-sm hover:shadow-md dark:border-zinc-800 dark:bg-zinc-900 dark:hover:border-orange-500/70 dark:hover:bg-orange-950/30 dark:hover:shadow-none"
+          className="inline-flex h-11 items-center gap-2.5 rounded-[10px] border border-black/[0.11] bg-zinc-50 px-4 text-sm font-semibold text-[#374151] transition-colors hover:border-[#c2c9d2] hover:bg-white hover:text-[#111827] dark:border-white/[0.10] dark:bg-[#303842] dark:text-[#d7dee8] dark:hover:border-white/[0.18] dark:hover:bg-[#3a434e] dark:hover:text-white"
         >
-          <Search className="w-5 h-5 text-orange-600" />
-          <span className="font-medium text-gray-700 dark:text-zinc-200">{t('welcome.threatHunting')}</span>
+          <Search className="h-[18px] w-[18px] text-[#f4511e]" />
+          <span>{t('welcome.threatHunting')}</span>
         </button>
         <button
           onClick={() => onSuggestion(t('welcome.incidentResponseSuggestion'))}
-          className="flex items-center gap-2 px-5 py-3 bg-white border-2 border-gray-200 rounded-xl hover:border-amber-400 hover:bg-amber-50 transition-all duration-200 shadow-sm hover:shadow-md dark:border-zinc-800 dark:bg-zinc-900 dark:hover:border-amber-500/70 dark:hover:bg-amber-950/30 dark:hover:shadow-none"
+          className="inline-flex h-11 items-center gap-2.5 rounded-[10px] border border-black/[0.11] bg-zinc-50 px-4 text-sm font-semibold text-[#374151] transition-colors hover:border-[#c2c9d2] hover:bg-white hover:text-[#111827] dark:border-white/[0.10] dark:bg-[#303842] dark:text-[#d7dee8] dark:hover:border-white/[0.18] dark:hover:bg-[#3a434e] dark:hover:text-white"
         >
-          <AlertTriangle className="w-5 h-5 text-amber-600" />
-          <span className="font-medium text-gray-700 dark:text-zinc-200">{t('welcome.incidentResponse')}</span>
+          <AlertTriangle className="h-[18px] w-[18px] text-amber-500" />
+          <span>{t('welcome.incidentResponse')}</span>
         </button>
       </div>
     </div>
diff --git a/webui/src/pages/Settings/ArchivedDataPanel.test.tsx b/webui/src/pages/Settings/ArchivedDataPanel.test.tsx
new file mode 100644
index 000000000..cd5756afe
--- /dev/null
+++ b/webui/src/pages/Settings/ArchivedDataPanel.test.tsx
@@ -0,0 +1,181 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+import { act, render, screen, waitFor } from '@testing-library/react';
+import userEvent from '@testing-library/user-event';
+import ArchivedDataPanel from './ArchivedDataPanel';
+
+const { sessionApi, toast } = vi.hoisted(() => ({
+  sessionApi: {
+    list: vi.fn(),
+    restore: vi.fn(),
+    delete: vi.fn(),
+  },
+  toast: {
+    success: vi.fn(),
+    error: vi.fn(),
+  },
+}));
+
+vi.mock('react-i18next', () => ({
+  useTranslation: () => ({
+    t: (key: string) => key,
+  }),
+}));
+
+vi.mock('@/api/session', () => ({ sessionApi }));
+vi.mock('@/components/common/Toast', () => ({ useToast: () => toast }));
+
+const archivedSession = {
+  id: 'session-archived',
+  slug: 'archived',
+  projectID: 'project-1',
+  projectName: 'Security Project',
+  effectiveProjectID: 'project-1',
+  directory: '/tmp/project',
+  title: 'Archived Session',
+  version: '1.0.0',
+  category: 'user',
+  status: 'archived',
+  canDelete: true,
+  ownerUserID: 'user-alice',
+  ownerUsername: 'alice',
+  time: {
+    created: 1710000000000,
+    updated: 1710000100000,
+    archived: 1710000200000,
+  },
+};
+
+describe('ArchivedDataPanel', () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    sessionApi.list.mockResolvedValue([archivedSession]);
+    sessionApi.restore.mockResolvedValue({ ...archivedSession, status: 'active' });
+    sessionApi.delete.mockResolvedValue(true);
+  });
+
+  it('loads archived sessions and restores one', async () => {
+    const user = userEvent.setup();
+    render(<ArchivedDataPanel />);
+
+    expect(await screen.findByText('Archived Session')).toBeInTheDocument();
+    expect(sessionApi.list).toHaveBeenCalledWith(expect.objectContaining({
+      status: 'archived',
+      manager: true,
+      roots: true,
+    }));
+    expect(screen.getByText('alice')).toBeInTheDocument();
+    expect(screen.getByText('Security Project')).toBeInTheDocument();
+    expect(screen.queryByText('project-1')).not.toBeInTheDocument();
+    expect(screen.queryByText('user')).not.toBeInTheDocument();
+
+    const tableHeader = screen.getByText('archivedData.session').parentElement;
+    expect(tableHeader).toHaveClass('text-sm', 'font-normal');
+    expect(tableHeader).not.toHaveClass('text-xs', 'font-semibold');
+    expect(screen.getByText('Archived Session')).toHaveClass('text-sm', 'font-normal');
+    expect(screen.getByText('alice')).toHaveClass('text-sm', 'font-normal');
+    expect(screen.getByText('Security Project')).toHaveClass('text-sm', 'font-normal');
+
+    const refreshButton = screen.getByRole('button', { name: 'archivedData.refresh' });
+    expect(refreshButton.textContent).toBe('');
+
+    const restoreButton = screen.getByRole('button', { name: 'archivedData.restore' });
+    expect(restoreButton).toHaveTextContent('archivedData.restore');
+    expect(restoreButton).toHaveClass('text-sm', 'font-normal', 'text-blue-600', 'hover:text-blue-700');
+    expect(screen.getByRole('button', { name: 'archivedData.delete' })).toHaveClass('text-sm', 'font-normal');
+    await user.click(restoreButton);
+
+    await waitFor(() => expect(sessionApi.restore).toHaveBeenCalledWith('session-archived'));
+    expect(screen.queryByText('Archived Session')).not.toBeInTheDocument();
+  });
+
+  it('requires a dialog confirmation before permanent deletion', async () => {
+    const user = userEvent.setup();
+    render(<ArchivedDataPanel />);
+
+    await screen.findByText('Archived Session');
+    const deleteButton = screen.getByRole('button', { name: 'archivedData.delete' });
+    expect(deleteButton).toHaveTextContent('archivedData.delete');
+    await user.click(deleteButton);
+
+    expect(screen.getByRole('dialog')).toBeInTheDocument();
+    expect(sessionApi.delete).not.toHaveBeenCalled();
+    await user.click(screen.getByRole('button', { name: 'archivedData.confirmDelete' }));
+
+    await waitFor(() => expect(sessionApi.delete).toHaveBeenCalledWith('session-archived'));
+    expect(screen.queryByText('Archived Session')).not.toBeInTheDocument();
+  });
+
+  it('keeps failed rows after a partial bulk restore', async () => {
+    const user = userEvent.setup();
+    const failedSession = {
+      ...archivedSession,
+      id: 'session-failed',
+      title: 'Failed Session',
+    };
+    sessionApi.list.mockResolvedValue([archivedSession, failedSession]);
+    sessionApi.restore.mockImplementation((id: string) => (
+      id === failedSession.id
+        ? Promise.reject(new Error('restore failed'))
+        : Promise.resolve({ ...archivedSession, status: 'active' })
+    ));
+    render(<ArchivedDataPanel />);
+
+    await screen.findByText('Failed Session');
+    await user.click(screen.getByRole('checkbox', { name: 'archivedData.selectAll' }));
+    await user.click(screen.getByRole('button', { name: 'archivedData.restoreSelected' }));
+
+    await waitFor(() => expect(screen.queryByText('Archived Session')).not.toBeInTheDocument());
+    expect(screen.getByText('Failed Session')).toBeInTheDocument();
+    expect(toast.success).toHaveBeenCalled();
+    expect(toast.error).toHaveBeenCalled();
+  });
+
+  it('ignores an older response after a newer search completes', async () => {
+    const user = userEvent.setup();
+    let resolveInitial!: (value: unknown[]) => void;
+    let resolveSearch!: (value: unknown[]) => void;
+    sessionApi.list
+      .mockImplementationOnce(() => new Promise((resolve) => { resolveInitial = resolve; }))
+      .mockImplementationOnce(() => new Promise((resolve) => { resolveSearch = resolve; }));
+    render(<ArchivedDataPanel />);
+
+    await user.type(screen.getByRole('textbox', { name: 'archivedData.searchPlaceholder' }), 'new');
+    await waitFor(() => expect(sessionApi.list).toHaveBeenCalledTimes(2));
+    await act(async () => {
+      resolveSearch([{ ...archivedSession, id: 'session-new', title: 'New Result' }]);
+    });
+    expect(await screen.findByText('New Result')).toBeInTheDocument();
+
+    await act(async () => {
+      resolveInitial([{ ...archivedSession, id: 'session-old', title: 'Old Result' }]);
+    });
+    expect(screen.queryByText('Old Result')).not.toBeInTheDocument();
+    expect(screen.getByText('New Result')).toBeInTheDocument();
+  });
+
+  it('labels legacy ownerless sessions as system tasks', async () => {
+    sessionApi.list.mockResolvedValue([{
+      ...archivedSession,
+      ownerUserID: undefined,
+      ownerUsername: undefined,
+    }]);
+
+    render(<ArchivedDataPanel />);
+
+    expect(await screen.findByText('archivedData.systemOwner')).toBeInTheDocument();
+    expect(screen.queryByText('user')).not.toBeInTheDocument();
+  });
+
+  it('localizes the API service owner as a system task', async () => {
+    sessionApi.list.mockResolvedValue([{
+      ...archivedSession,
+      ownerUserID: 'api-token-service',
+      ownerUsername: 'api-token-service',
+    }]);
+
+    render(<ArchivedDataPanel />);
+
+    expect(await screen.findByText('archivedData.systemOwner')).toBeInTheDocument();
+    expect(screen.queryByText('api-token-service')).not.toBeInTheDocument();
+  });
+});
diff --git a/webui/src/pages/Settings/ArchivedDataPanel.tsx b/webui/src/pages/Settings/ArchivedDataPanel.tsx
new file mode 100644
index 000000000..cf7a5e97f
--- /dev/null
+++ b/webui/src/pages/Settings/ArchivedDataPanel.tsx
@@ -0,0 +1,351 @@
+import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+import { ArchiveRestore, Loader2, RefreshCw, Search, Trash2 } from 'lucide-react';
+import { useTranslation } from 'react-i18next';
+import { sessionApi } from '@/api/session';
+import { useToast } from '@/components/common/Toast';
+import type { Session } from '@/types';
+
+const PAGE_SIZE = 50;
+const SYSTEM_SESSION_OWNER_USER_ID = 'api-token-service';
+
+function getProjectLabel(session: Session): string {
+  const projectName = session.projectName?.trim();
+  if (projectName) return projectName;
+  const normalizedDirectory = session.directory.replace(/[\\/]+$/g, '');
+  const directoryName = normalizedDirectory.split(/[\\/]/).filter(Boolean).pop();
+  return directoryName || session.projectID;
+}
+
+async function runBatchAction(
+  ids: string[],
+  action: (id: string) => Promise<unknown>,
+): Promise<{ succeeded: string[]; failed: string[] }> {
+  const results = await Promise.all(ids.map(async (id) => {
+    try {
+      await action(id);
+      return { id, succeeded: true };
+    } catch {
+      return { id, succeeded: false };
+    }
+  }));
+  return {
+    succeeded: results.filter((result) => result.succeeded).map((result) => result.id),
+    failed: results.filter((result) => !result.succeeded).map((result) => result.id),
+  };
+}
+
+export default function ArchivedDataPanel() {
+  const { t } = useTranslation('session');
+  const toast = useToast();
+  const [sessions, setSessions] = useState<Session[]>([]);
+  const [searchValue, setSearchValue] = useState('');
+  const [search, setSearch] = useState('');
+  const [loading, setLoading] = useState(true);
+  const [loadingMore, setLoadingMore] = useState(false);
+  const [hasMore, setHasMore] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+  const [selectedIds, setSelectedIds] = useState<Set<string>>(new Set());
+  const [restoringIds, setRestoringIds] = useState<Set<string>>(new Set());
+  const [pendingDeleteIds, setPendingDeleteIds] = useState<string[]>([]);
+  const [deleting, setDeleting] = useState(false);
+  const requestSeqRef = useRef(0);
+
+  useEffect(() => {
+    const timer = window.setTimeout(() => setSearch(searchValue.trim()), 250);
+    return () => window.clearTimeout(timer);
+  }, [searchValue]);
+
+  const fetchSessions = useCallback(async (append = false) => {
+    const requestSeq = ++requestSeqRef.current;
+    append ? setLoadingMore(true) : setLoading(true);
+    setError(null);
+    try {
+      const offset = append ? sessions.length : 0;
+      const response = await sessionApi.list({
+        status: 'archived',
+        view: 'list',
+        manager: true,
+        roots: true,
+        search: search || undefined,
+        limit: PAGE_SIZE,
+        offset,
+      });
+      if (requestSeq !== requestSeqRef.current) return;
+      const rows = Array.isArray(response) ? response : [];
+      setSessions((current) => {
+        if (!append) return rows;
+        const knownIds = new Set(current.map((session) => session.id));
+        return [...current, ...rows.filter((session) => !knownIds.has(session.id))];
+      });
+      setHasMore(rows.length >= PAGE_SIZE);
+      if (!append) setSelectedIds(new Set());
+    } catch (err: any) {
+      if (requestSeq !== requestSeqRef.current) return;
+      setError(err?.response?.data?.detail || err?.message || t('archivedData.loadFailed'));
+    } finally {
+      if (requestSeq === requestSeqRef.current) {
+        setLoading(false);
+        setLoadingMore(false);
+      }
+    }
+  }, [search, sessions.length, t]);
+
+  useEffect(() => {
+    void fetchSessions(false);
+  }, [search]);
+
+  const manageableSessions = useMemo(
+    () => sessions.filter((session) => session.canDelete !== false),
+    [sessions],
+  );
+  const allSelected = manageableSessions.length > 0
+    && manageableSessions.every((session) => selectedIds.has(session.id));
+
+  const removeRows = useCallback((ids: string[]) => {
+    const removed = new Set(ids);
+    setSessions((current) => current.filter((session) => !removed.has(session.id)));
+    setSelectedIds((current) => {
+      const next = new Set(current);
+      ids.forEach((id) => next.delete(id));
+      return next;
+    });
+  }, []);
+
+  const restoreSessions = useCallback(async (ids: string[]) => {
+    if (ids.length === 0) return;
+    setRestoringIds((current) => new Set([...current, ...ids]));
+    const { succeeded, failed } = await runBatchAction(ids, sessionApi.restore);
+    removeRows(succeeded);
+    setRestoringIds((current) => {
+      const next = new Set(current);
+      ids.forEach((id) => next.delete(id));
+      return next;
+    });
+    if (succeeded.length > 0) toast.success(t('archivedData.restoreSuccess', { count: succeeded.length }));
+    if (failed.length > 0) toast.error(t('archivedData.restoreFailed', { count: failed.length }));
+  }, [removeRows, t, toast]);
+
+  const permanentlyDelete = useCallback(async () => {
+    if (pendingDeleteIds.length === 0 || deleting) return;
+    setDeleting(true);
+    const { succeeded, failed } = await runBatchAction(
+      pendingDeleteIds,
+      sessionApi.delete,
+    );
+    removeRows(succeeded);
+    setDeleting(false);
+    setPendingDeleteIds([]);
+    if (succeeded.length > 0) toast.success(t('archivedData.deleteSuccess', { count: succeeded.length }));
+    if (failed.length > 0) toast.error(t('archivedData.deleteFailed', { count: failed.length }));
+  }, [deleting, pendingDeleteIds, removeRows, t, toast]);
+
+  const toggleSelection = (id: string) => {
+    setSelectedIds((current) => {
+      const next = new Set(current);
+      next.has(id) ? next.delete(id) : next.add(id);
+      return next;
+    });
+  };
+
+  return (
+    <div className="mx-auto w-full max-w-7xl">
+      <header className="border-b border-zinc-200 pb-6 dark:border-zinc-800">
+        <h1 className="text-2xl font-bold tracking-normal text-zinc-950 dark:text-zinc-50">
+          {t('archivedData.title')}
+        </h1>
+        <p className="mt-2 text-sm text-zinc-500 dark:text-zinc-400">
+          {t('archivedData.description')}
+        </p>
+      </header>
+
+      <div className="mt-6 flex items-center gap-3">
+        <label className="relative block w-full sm:max-w-sm">
+          <Search className="pointer-events-none absolute left-3 top-1/2 h-4 w-4 -translate-y-1/2 text-zinc-400" />
+          <input
+            value={searchValue}
+            onChange={(event) => setSearchValue(event.target.value)}
+            placeholder={t('archivedData.searchPlaceholder')}
+            aria-label={t('archivedData.searchPlaceholder')}
+            className="h-10 w-full rounded-lg border border-zinc-200 bg-white pl-9 pr-3 text-sm outline-none focus:border-zinc-400 dark:border-zinc-800 dark:bg-zinc-900 dark:text-zinc-100"
+          />
+        </label>
+        <button
+          type="button"
+          onClick={() => void fetchSessions(false)}
+          disabled={loading}
+          aria-label={t('archivedData.refresh')}
+          title={t('archivedData.refresh')}
+          className="ml-auto inline-flex h-10 w-10 shrink-0 items-center justify-center rounded-lg text-zinc-500 transition-colors hover:bg-zinc-100 hover:text-zinc-900 disabled:opacity-50 dark:text-zinc-400 dark:hover:bg-zinc-800 dark:hover:text-zinc-100"
+        >
+          <RefreshCw className={`h-4 w-4 ${loading ? 'animate-spin' : ''}`} />
+        </button>
+      </div>
+
+      {selectedIds.size > 0 && (
+        <div className="mt-4 flex flex-wrap items-center gap-2 rounded-lg border border-zinc-200 bg-zinc-50 p-3 dark:border-zinc-800 dark:bg-zinc-900/60">
+          <span className="mr-auto text-sm text-zinc-600 dark:text-zinc-300">
+            {t('archivedData.selected', { count: selectedIds.size })}
+          </span>
+          <button
+            type="button"
+            onClick={() => void restoreSessions(Array.from(selectedIds))}
+            aria-label={t('archivedData.restoreSelected')}
+            title={t('archivedData.restoreSelected')}
+            className="inline-flex h-9 w-9 items-center justify-center rounded-md text-zinc-600 transition-colors hover:bg-white hover:text-zinc-950 dark:text-zinc-300 dark:hover:bg-zinc-800 dark:hover:text-zinc-50"
+          >
+            <ArchiveRestore className="h-4 w-4" />
+          </button>
+          <button
+            type="button"
+            onClick={() => setPendingDeleteIds(Array.from(selectedIds))}
+            className="inline-flex h-9 items-center gap-2 rounded-md bg-red-600 px-3 text-sm font-semibold text-white hover:bg-red-700"
+          >
+            <Trash2 className="h-4 w-4" />
+            {t('archivedData.deleteSelected')}
+          </button>
+        </div>
+      )}
+
+      <div className="mt-4 overflow-hidden rounded-xl border border-zinc-200 dark:border-zinc-800">
+        {loading ? (
+          <div className="flex min-h-48 items-center justify-center gap-2 text-sm text-zinc-500">
+            <Loader2 className="h-5 w-5 animate-spin" />
+            {t('archivedData.loading')}
+          </div>
+        ) : error ? (
+          <div className="flex min-h-48 flex-col items-center justify-center gap-3 p-6 text-center">
+            <p className="text-sm text-red-600 dark:text-red-300">{error}</p>
+            <button type="button" onClick={() => void fetchSessions(false)} className="text-sm font-semibold underline">
+              {t('archivedData.retry')}
+            </button>
+          </div>
+        ) : sessions.length === 0 ? (
+          <div className="flex min-h-48 flex-col items-center justify-center p-6 text-center text-zinc-500">
+            <ArchiveRestore className="mb-3 h-8 w-8 text-zinc-300 dark:text-zinc-700" />
+            <p className="text-sm font-semibold text-zinc-700 dark:text-zinc-300">{t('archivedData.empty')}</p>
+            <p className="mt-1 text-xs">{t('archivedData.emptyDescription')}</p>
+          </div>
+        ) : (
+          <>
+            <div className="hidden grid-cols-[2.25rem_minmax(10rem,2fr)_minmax(6.5rem,0.8fr)_minmax(7rem,1fr)_9.5rem_7rem] items-center gap-4 border-b border-zinc-200 bg-zinc-50 px-4 py-2.5 text-sm font-normal text-zinc-500 dark:border-zinc-800 dark:bg-zinc-900 dark:text-zinc-400 md:grid">
+              <input
+                type="checkbox"
+                checked={allSelected}
+                onChange={() => setSelectedIds(allSelected ? new Set() : new Set(manageableSessions.map((session) => session.id)))}
+                aria-label={t('archivedData.selectAll')}
+                className="h-4 w-4 justify-self-center"
+              />
+              <span>{t('archivedData.session')}</span>
+              <span>{t('archivedData.owner')}</span>
+              <span>{t('archivedData.project')}</span>
+              <span>{t('archivedData.archivedAt')}</span>
+              <span className="text-center">{t('archivedData.actions')}</span>
+            </div>
+            {sessions.map((session) => {
+              const canManage = session.canDelete !== false;
+              const restoring = restoringIds.has(session.id);
+              const isSystemOwned = session.ownerUserID === SYSTEM_SESSION_OWNER_USER_ID
+                || (!session.ownerUserID && session.ownerUsername === SYSTEM_SESSION_OWNER_USER_ID);
+              const ownerLabel = isSystemOwned
+                ? t('archivedData.systemOwner')
+                : session.ownerUsername || t('archivedData.systemOwner');
+              const projectLabel = getProjectLabel(session);
+              return (
+                <div key={session.id} className="grid gap-3 border-b border-zinc-100 px-4 py-3.5 text-sm font-normal last:border-b-0 dark:border-zinc-900 md:grid-cols-[2.25rem_minmax(10rem,2fr)_minmax(6.5rem,0.8fr)_minmax(7rem,1fr)_9.5rem_7rem] md:items-center md:gap-4">
+                  <input
+                    type="checkbox"
+                    checked={selectedIds.has(session.id)}
+                    onChange={() => toggleSelection(session.id)}
+                    disabled={!canManage}
+                    aria-label={t('archivedData.selectSession', { title: session.title })}
+                    className="h-4 w-4 justify-self-center"
+                  />
+                  <div className="min-w-0">
+                    <p className="truncate text-sm font-normal text-zinc-900 dark:text-zinc-100">{session.title}</p>
+                  </div>
+                  <p className="truncate text-sm font-normal text-zinc-600 dark:text-zinc-400" title={ownerLabel}>
+                    {ownerLabel}
+                  </p>
+                  <p className="truncate text-sm font-normal text-zinc-600 dark:text-zinc-400" title={projectLabel}>
+                    {projectLabel}
+                  </p>
+                  <p className="text-sm font-normal text-zinc-500">
+                    {session.time.archived ? new Date(session.time.archived).toLocaleString() : '—'}
+                  </p>
+                  <div className="flex items-center justify-center gap-3">
+                    <button
+                      type="button"
+                      onClick={() => void restoreSessions([session.id])}
+                      disabled={!canManage || restoring}
+                      aria-label={t('archivedData.restore')}
+                      title={t('archivedData.restore')}
+                      className="text-sm font-normal text-blue-600 transition-colors hover:text-blue-700 disabled:opacity-40 dark:text-blue-300 dark:hover:text-blue-200"
+                    >
+                      {t('archivedData.restore')}
+                    </button>
+                    <button
+                      type="button"
+                      onClick={() => setPendingDeleteIds([session.id])}
+                      disabled={!canManage}
+                      aria-label={t('archivedData.delete')}
+                      title={t('archivedData.delete')}
+                      className="text-sm font-normal text-red-600 transition-colors hover:text-red-700 disabled:opacity-40 dark:text-red-300 dark:hover:text-red-200"
+                    >
+                      {t('archivedData.delete')}
+                    </button>
+                  </div>
+                </div>
+              );
+            })}
+          </>
+        )}
+      </div>
+
+      {hasMore && !loading && (
+        <div className="mt-4 flex justify-center">
+          <button
+            type="button"
+            onClick={() => void fetchSessions(true)}
+            disabled={loadingMore}
+            className="inline-flex h-9 items-center gap-2 rounded-md border border-zinc-200 px-4 text-sm font-semibold hover:bg-zinc-50 disabled:opacity-50 dark:border-zinc-800 dark:hover:bg-zinc-900"
+          >
+            {loadingMore && <Loader2 className="h-4 w-4 animate-spin" />}
+            {t('archivedData.loadMore')}
+          </button>
+        </div>
+      )}
+
+      {pendingDeleteIds.length > 0 && (
+        <div className="fixed inset-0 z-[100] flex items-center justify-center bg-black/45 p-4" role="presentation">
+          <div role="dialog" aria-modal="true" aria-labelledby="permanent-delete-title" className="w-full max-w-md rounded-xl bg-white p-6 shadow-2xl dark:bg-zinc-900">
+            <h2 id="permanent-delete-title" className="text-lg font-bold text-zinc-950 dark:text-zinc-50">
+              {t('archivedData.deleteTitle')}
+            </h2>
+            <p className="mt-2 text-sm leading-6 text-zinc-600 dark:text-zinc-300">
+              {t('archivedData.deleteDescription', { count: pendingDeleteIds.length })}
+            </p>
+            <div className="mt-6 flex justify-end gap-2">
+              <button
+                type="button"
+                onClick={() => setPendingDeleteIds([])}
+                disabled={deleting}
+                className="h-9 rounded-md border border-zinc-200 px-4 text-sm font-semibold dark:border-zinc-700"
+              >
+                {t('cancel')}
+              </button>
+              <button
+                type="button"
+                onClick={() => void permanentlyDelete()}
+                disabled={deleting}
+                className="inline-flex h-9 items-center gap-2 rounded-md bg-red-600 px-4 text-sm font-semibold text-white hover:bg-red-700 disabled:opacity-50"
+              >
+                {deleting && <Loader2 className="h-4 w-4 animate-spin" />}
+                {t('archivedData.confirmDelete')}
+              </button>
+            </div>
+          </div>
+        </div>
+      )}
+    </div>
+  );
+}
diff --git a/webui/src/pages/Settings/index.test.tsx b/webui/src/pages/Settings/index.test.tsx
index 5f88be28d..a9ded5cb3 100644
--- a/webui/src/pages/Settings/index.test.tsx
+++ b/webui/src/pages/Settings/index.test.tsx
@@ -53,6 +53,10 @@ vi.mock('@/pages/AuditLogs', () => ({
   default: () => <div>audit logs page</div>,
 }));
 
+vi.mock('./ArchivedDataPanel', () => ({
+  default: () => <div>archived data page</div>,
+}));
+
 vi.mock('@/pages/FlocksproUpgrade', () => ({
   default: () => <div>flocks pro page</div>,
 }));
@@ -206,6 +210,16 @@ describe('SettingsPage', () => {
     expect(flocksproUsersApi.hasCapability).toHaveBeenCalled();
   });
 
+  it('renders archived data management for local users', async () => {
+    renderSettings('/settings/archived-data');
+
+    expect(await screen.findByText('archived data page')).toBeInTheDocument();
+    expect(screen.getAllByRole('link', { name: 'archivedData' })[0]).toHaveAttribute(
+      'href',
+      '/settings/archived-data',
+    );
+  });
+
   it('hides audit logs when Flocks Pro capability is unavailable', async () => {
     flocksproUsersApi.hasCapability.mockResolvedValue(false);
 
diff --git a/webui/src/pages/Settings/index.tsx b/webui/src/pages/Settings/index.tsx
index 21d184c2a..7ab6bbd67 100644
--- a/webui/src/pages/Settings/index.tsx
+++ b/webui/src/pages/Settings/index.tsx
@@ -5,6 +5,7 @@ import { useTranslation } from 'react-i18next';
 import {
   ArrowLeft,
   ArrowUpCircle,
+  Archive,
   Check,
   ImageIcon,
   Languages,
@@ -45,8 +46,9 @@ const ConfigPage = lazySettingsPage(() => import('@/pages/Config'));
 const SystemLogPage = lazySettingsPage(() => import('@/pages/SystemLog'));
 const FlocksproUpgradePage = lazySettingsPage(() => import('@/pages/FlocksproUpgrade'), ['flockspro']);
 const AuditLogsPage = lazySettingsPage(() => import('@/pages/AuditLogs'), ['flockspro']);
+const ArchivedDataPage = lazySettingsPage(() => import('./ArchivedDataPanel'), ['session']);
 
-type SettingsSectionId = 'preferences' | 'account' | 'system-logs' | 'audit-logs' | 'flockspro';
+type SettingsSectionId = 'preferences' | 'archived-data' | 'account' | 'system-logs' | 'audit-logs' | 'flockspro';
 
 interface ReturnLocation {
   pathname: string;
@@ -74,6 +76,7 @@ interface SettingsGroup {
 function isSettingsSectionId(value: string | undefined): value is SettingsSectionId {
   return (
     value === 'preferences' ||
+    value === 'archived-data' ||
     value === 'account' ||
     value === 'system-logs' ||
     value === 'audit-logs' ||
@@ -477,6 +480,7 @@ function SettingsContent({ sectionId }: { sectionId: SettingsSectionId }) {
   return (
     <Suspense fallback={<RoutePageSkeleton delayMs={180} />}>
       {sectionId === 'account' && <ConfigPage />}
+      {sectionId === 'archived-data' && <ArchivedDataPage />}
       {sectionId === 'system-logs' && <SystemLogPage />}
       {sectionId === 'audit-logs' && <AuditLogsPage />}
       {sectionId === 'flockspro' && <FlocksproUpgradePage />}
@@ -544,6 +548,12 @@ export default function SettingsPage() {
           { id: 'preferences', name: t('settingsPreferences'), icon: SettingsIcon },
         ],
       },
+      {
+        name: t('settingsGroupData'),
+        items: [
+          { id: 'archived-data', name: t('archivedData'), icon: Archive },
+        ],
+      },
       {
         name: t('settingsGroupSystem'),
         items: [
diff --git a/webui/src/pages/Task/QueuedSection.test.tsx b/webui/src/pages/Task/QueuedSection.test.tsx
index bdbc6ebf4..932ba89ea 100644
--- a/webui/src/pages/Task/QueuedSection.test.tsx
+++ b/webui/src/pages/Task/QueuedSection.test.tsx
@@ -344,6 +344,15 @@ describe('QueuedSection', () => {
         sessionId: 'ses-task-1',
         live: true,
         hideInput: true,
+        display: {
+          compact: false,
+          fullWidth: true,
+          pageCanvas: true,
+          showTimestamp: true,
+          collapseIntermediateSteps: true,
+          processGroupsDefaultOpen: false,
+          processGroupsOpenWhileActive: true,
+        },
       }),
     );
   });
diff --git a/webui/src/pages/Task/QueuedSection.tsx b/webui/src/pages/Task/QueuedSection.tsx
index 2ebc15f5c..754a813db 100644
--- a/webui/src/pages/Task/QueuedSection.tsx
+++ b/webui/src/pages/Task/QueuedSection.tsx
@@ -467,6 +467,15 @@ function QueuedDetailPanel({ task, onClose, onAction, onRefresh }: {
               sessionId={sessionId}
               live={shouldStreamSession}
               hideInput
+              display={{
+                compact: false,
+                fullWidth: true,
+                pageCanvas: true,
+                showTimestamp: true,
+                collapseIntermediateSteps: true,
+                processGroupsDefaultOpen: false,
+                processGroupsOpenWhileActive: true,
+              }}
               emptyText={emptyText}
               className="flex-1 min-h-0"
               onSSEEvent={(event) => {
diff --git a/webui/src/pages/WorkflowCreate/CreateChatTab.test.tsx b/webui/src/pages/WorkflowCreate/CreateChatTab.test.tsx
index 0e6b1d791..a80820db1 100644
--- a/webui/src/pages/WorkflowCreate/CreateChatTab.test.tsx
+++ b/webui/src/pages/WorkflowCreate/CreateChatTab.test.tsx
@@ -9,14 +9,26 @@ import { workflowAPI } from '@/api/workflow';
 
 const {
   capturedSessionChatProps,
+  capturedSessionOptions,
+  capturedModelPickerProps,
   mockCreateAndSend,
+  mockClientGet,
+  mockClientPatch,
   mockSendPrompt,
+  mockSetSelectedModelAuto,
   mockSetSelectedModelKey,
+  workflowModelState,
 } = vi.hoisted(() => ({
   capturedSessionChatProps: [] as any[],
+  capturedSessionOptions: [] as any[],
+  capturedModelPickerProps: [] as any[],
   mockCreateAndSend: vi.fn(),
+  mockClientGet: vi.fn(),
+  mockClientPatch: vi.fn(),
   mockSendPrompt: vi.fn(),
+  mockSetSelectedModelAuto: vi.fn(),
   mockSetSelectedModelKey: vi.fn(),
+  workflowModelState: { auto: false },
 }));
 
 const selectedPromptModel = { providerID: 'provider-1', modelID: 'model-1' };
@@ -37,12 +49,19 @@ vi.mock('@/hooks/useDefaultModelVision', () => ({
 }));
 
 vi.mock('@/hooks/useSessionChat', () => ({
-  useSessionChat: (options: any) => ({
-    sessionId: options.initialSessionId ?? null,
-    error: null,
-    createAndSend: mockCreateAndSend,
-    retry: vi.fn(),
-  }),
+  useSessionChat: (options: any) => {
+    capturedSessionOptions.push(options);
+    return {
+      sessionId: options.initialSessionId ?? null,
+      error: null,
+      createAndSend: mockCreateAndSend,
+      retry: vi.fn(),
+    };
+  },
+}));
+
+vi.mock('@/api/client', () => ({
+  default: { get: mockClientGet, patch: mockClientPatch },
 }));
 
 vi.mock('@/api/workflow', () => ({
@@ -60,14 +79,30 @@ vi.mock('@/components/common/ChatPromptSelectors', () => ({
   useChatModelOptions: () => ({
     groupedOptions: [],
     loading: false,
+    effectiveModelOption: selectedModelOption,
+    modelPickerAutoOption: {
+      selected: workflowModelState.auto,
+      disabled: false,
+      statusLabel: 'Auto',
+      onSelect: () => { workflowModelState.auto = true; },
+    },
+    selectModelKey: (key: string) => {
+      workflowModelState.auto = false;
+      mockSetSelectedModelKey(key);
+    },
+    selectedModelAuto: workflowModelState.auto,
     selectedModelOption,
-    selectedPromptModel,
+    selectedPromptModel: workflowModelState.auto ? null : selectedPromptModel,
+    setSelectedModelAuto: mockSetSelectedModelAuto,
     setSelectedModelKey: mockSetSelectedModelKey,
   }),
   ChatAgentDisplay: ({ selectedAgent }: { selectedAgent: string }) => (
     <div>Agent:{selectedAgent}</div>
   ),
-  ChatModelPicker: () => <div>ModelPicker</div>,
+  ChatModelPicker: (props: any) => {
+    capturedModelPickerProps.push(props);
+    return <div>ModelPicker</div>;
+  },
 }));
 
 vi.mock('@/components/common/SessionChat', () => ({
@@ -141,7 +176,15 @@ describe('WorkflowCreate CreateChatTab', () => {
   beforeEach(() => {
     vi.clearAllMocks();
     capturedSessionChatProps.length = 0;
+    capturedSessionOptions.length = 0;
+    capturedModelPickerProps.length = 0;
+    workflowModelState.auto = false;
+    mockSetSelectedModelAuto.mockImplementation((enabled: boolean) => {
+      workflowModelState.auto = enabled;
+    });
     mockCreateAndSend.mockResolvedValue('session-1');
+    mockClientGet.mockResolvedValue({ data: {} });
+    mockClientPatch.mockResolvedValue({ data: {} });
     vi.mocked(workflowAPI.listSummaries).mockResolvedValue({ data: [] });
     vi.mocked(workflowAPI.get).mockResolvedValue({ data: null as any });
   });
@@ -170,6 +213,7 @@ describe('WorkflowCreate CreateChatTab', () => {
         imageParts: [],
         agent: 'rex',
         model: selectedPromptModel,
+        modelAuto: false,
         displayText: '@@flocks-instruction:如何创建工作流',
       });
     });
@@ -201,6 +245,7 @@ describe('WorkflowCreate CreateChatTab', () => {
     expect(capturedSessionChatProps[0].agentName).toBe('rex');
     expect(capturedSessionChatProps[0].mentionAgents.map((agent: any) => agent.name)).toEqual(['rex']);
     expect(capturedSessionChatProps[0].model).toEqual(selectedPromptModel);
+    expect(capturedSessionChatProps[0].modelAuto).toBe(false);
     expect(capturedSessionChatProps[0].supportsVision).toBe(true);
     expect(capturedSessionChatProps[0].contextWindowTokens).toBe(128000);
     expect(capturedSessionChatProps[0].display).toEqual({
@@ -224,9 +269,45 @@ describe('WorkflowCreate CreateChatTab', () => {
         imageParts: [],
         agent: 'rex',
         model: selectedPromptModel,
+        modelAuto: false,
         displayText: '@@flocks-instruction:IP 情报',
       });
     });
+    expect(capturedSessionOptions[0]).toEqual(expect.objectContaining({
+      category: 'workflow',
+      modelAuto: false,
+    }));
+    expect(capturedModelPickerProps[0].autoOption).toBeDefined();
+  });
+
+  it('uses Auto only after manual selection and sends no concrete model', async () => {
+    const view = renderCreateChatTab();
+
+    act(() => {
+      capturedModelPickerProps[capturedModelPickerProps.length - 1].autoOption.onSelect();
+    });
+    view.rerender(
+      <MemoryRouter>
+        <CreateChatTab onWorkflowCreated={vi.fn()} />
+      </MemoryRouter>,
+    );
+
+    const latestProps = capturedSessionChatProps[capturedSessionChatProps.length - 1];
+    expect(latestProps.model).toBeNull();
+    expect(latestProps.modelAuto).toBe(true);
+    expect(latestProps.supportsVision).toBe(true);
+    expect(capturedSessionOptions[capturedSessionOptions.length - 1].modelAuto).toBe(true);
+
+    await latestProps.onCreateAndSend('创建一个工作流', []);
+
+    expect(mockCreateAndSend).toHaveBeenLastCalledWith({
+      text: '创建一个工作流',
+      imageParts: [],
+      agent: 'rex',
+      model: null,
+      modelAuto: true,
+      displayText: undefined,
+    });
   });
 
   it('resumes a persisted create session', () => {
@@ -241,6 +322,45 @@ describe('WorkflowCreate CreateChatTab', () => {
     expect(onSessionChange).toHaveBeenCalledWith('session-restored');
   });
 
+  it('restores Auto from a persisted create session', async () => {
+    mockClientGet.mockResolvedValueOnce({ data: { model_auto: true } });
+
+    renderCreateChatTab({ initialSessionId: 'session-auto' });
+
+    await waitFor(() => {
+      expect(mockSetSelectedModelAuto).toHaveBeenCalledWith(true);
+      const latestProps = capturedSessionChatProps[capturedSessionChatProps.length - 1];
+      expect(latestProps.model).toBeNull();
+      expect(latestProps.modelAuto).toBe(true);
+    });
+  });
+
+  it('persists picker changes for a resumed create session', async () => {
+    renderCreateChatTab({ initialSessionId: 'session-picker' });
+
+    await waitFor(() => {
+      expect(mockClientGet).toHaveBeenCalledWith('/api/session/session-picker');
+    });
+    act(() => {
+      capturedModelPickerProps[capturedModelPickerProps.length - 1].autoOption.onSelect();
+    });
+    expect(mockClientPatch).toHaveBeenCalledWith('/api/session/session-picker', {
+      model_auto: true,
+      model_pinned: false,
+    });
+
+    act(() => {
+      capturedModelPickerProps[capturedModelPickerProps.length - 1]
+        .onSelectModel(selectedModelOption);
+    });
+    expect(mockClientPatch).toHaveBeenCalledWith('/api/session/session-picker', {
+      provider: 'provider-1',
+      model: 'model-1',
+      model_pinned: true,
+      model_auto: false,
+    });
+  });
+
   it('does not attach an already-known workflow just because it was created recently', async () => {
     const creationStartedAt = Date.now();
     const recentKnownWorkflow = {
diff --git a/webui/src/pages/WorkflowCreate/CreateChatTab.tsx b/webui/src/pages/WorkflowCreate/CreateChatTab.tsx
index 324cec695..41b69ee08 100644
--- a/webui/src/pages/WorkflowCreate/CreateChatTab.tsx
+++ b/webui/src/pages/WorkflowCreate/CreateChatTab.tsx
@@ -11,12 +11,14 @@ import {
   ChatModelPicker,
   useChatAgentOptions,
   useChatModelOptions,
+  type ChatModelOption,
 } from '@/components/common/ChatPromptSelectors';
 import ChatGuideDock, { type ChatGuideAction } from '@/components/common/ChatGuideDock';
 import GuidedCreatePanel from '@/components/common/GuidedCreatePanel';
 import { useSessionChat } from '@/hooks/useSessionChat';
 import { useDefaultModelVision } from '@/hooks/useDefaultModelVision';
 import { workflowAPI, Workflow } from '@/api/workflow';
+import client from '@/api/client';
 import type { ImagePartData } from '@/utils/imageUpload';
 
 const FALLBACK_POLL_MS = 10_000;
@@ -84,11 +86,15 @@ export default function CreateChatTab({
   const {
     groupedOptions: groupedChatModelOptions,
     loading: loadingChatModels,
+    effectiveModelOption,
+    modelPickerAutoOption,
+    selectModelKey,
+    selectedModelAuto,
     selectedModelOption,
     selectedPromptModel,
-    setSelectedModelKey,
-  } = useChatModelOptions();
-  const supportsVision = selectedModelOption?.supportsVision ?? defaultSupportsVision;
+    setSelectedModelAuto,
+  } = useChatModelOptions({ enableAuto: true });
+  const supportsVision = effectiveModelOption?.supportsVision ?? defaultSupportsVision;
   const exampleActions = useMemo(() => (
     (Array.isArray(exampleQuestions) ? exampleQuestions : []).map((question, index) => ({
       label: Array.isArray(exampleQuestionLabels) && exampleQuestionLabels[index]
@@ -106,11 +112,55 @@ export default function CreateChatTab({
   const { sessionId, error, createAndSend, retry } = useSessionChat({
     title: t('create.chat.sessionTitle'),
     category: 'workflow',
+    modelAuto: selectedModelAuto,
     contextMessage: t('create.chat.contextMessage'),
     welcomeMessage: t('create.chat.welcomeMessage'),
     initialSessionId,
   });
 
+  useEffect(() => {
+    setSelectedModelAuto(false);
+    if (!initialSessionId) return;
+    let cancelled = false;
+    client.get(`/api/session/${initialSessionId}`)
+      .then((response) => {
+        if (cancelled) return;
+        setSelectedModelAuto(Boolean(response.data?.model_auto));
+      })
+      .catch(() => {
+        if (cancelled) return;
+        setSelectedModelAuto(false);
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [initialSessionId, setSelectedModelAuto]);
+
+  const handleSelectModel = useCallback((option: ChatModelOption) => {
+    selectModelKey(option.key);
+    if (!sessionId) return;
+    client.patch(`/api/session/${sessionId}`, {
+      provider: option.providerID,
+      model: option.modelID,
+      model_pinned: true,
+      model_auto: false,
+    }).catch((err) => {
+      console.warn('[WorkflowCreate] failed to pin workflow chat model', err);
+    });
+  }, [selectModelKey, sessionId]);
+
+  const handleSelectAuto = useCallback(() => {
+    if (!modelPickerAutoOption || modelPickerAutoOption.disabled) return;
+    modelPickerAutoOption.onSelect();
+    if (!sessionId) return;
+    client.patch(`/api/session/${sessionId}`, {
+      model_auto: true,
+      model_pinned: false,
+    }).catch((err) => {
+      console.warn('[WorkflowCreate] failed to enable Auto for workflow chat', err);
+    });
+  }, [modelPickerAutoOption, sessionId]);
+
   const knownIdsRef = useRef<Set<string>>(new Set());
   const snapshotStartedAtRef = useRef<number | null>(null);
   const createdWorkflowRef = useRef<string | null>(null);
@@ -250,10 +300,11 @@ export default function CreateChatTab({
         imageParts,
         agent: agentOverride || WORKFLOW_CHAT_AGENT_NAME,
         model: modelOverride === undefined ? selectedPromptModel : modelOverride,
+        modelAuto: selectedModelAuto,
         displayText: options?.displayText,
       });
     },
-    [createAndSend, selectedPromptModel],
+    [createAndSend, selectedModelAuto, selectedPromptModel],
   );
 
   const handleWelcomeGuidePrompt = useCallback(
@@ -292,8 +343,9 @@ export default function CreateChatTab({
       agentName={WORKFLOW_CHAT_AGENT_NAME}
       mentionAgents={workflowChatAgents}
       supportsVision={supportsVision}
-      contextWindowTokens={selectedModelOption?.contextWindowTokens ?? null}
+      contextWindowTokens={effectiveModelOption?.contextWindowTokens ?? null}
       model={selectedPromptModel}
+      modelAuto={selectedModelAuto}
       onStreamingDone={handleStreamingDone}
       onSSEEvent={handleSSEEvent}
       onCreateAndSend={!sessionId ? handleCreateAndSend : undefined}
@@ -324,7 +376,10 @@ export default function CreateChatTab({
           groupedOptions={groupedChatModelOptions}
           loading={loadingChatModels}
           selectedModelOption={selectedModelOption}
-          onSelectModel={(option) => setSelectedModelKey(option.key)}
+          onSelectModel={handleSelectModel}
+          autoOption={modelPickerAutoOption
+            ? { ...modelPickerAutoOption, onSelect: handleSelectAuto }
+            : undefined}
         />
       }
       conversationBottomSlot={({ sendPrompt, sending, streaming }) => (
diff --git a/webui/src/pages/WorkflowDetail/tabs/ChatTab.test.tsx b/webui/src/pages/WorkflowDetail/tabs/ChatTab.test.tsx
index 72cad565f..3f9184fd8 100644
--- a/webui/src/pages/WorkflowDetail/tabs/ChatTab.test.tsx
+++ b/webui/src/pages/WorkflowDetail/tabs/ChatTab.test.tsx
@@ -209,6 +209,10 @@ vi.mock('react-i18next', () => ({
         'detail.chat.welcome.guideSampleShort': '验证样例数据',
         'detail.chat.welcome.guideSampleDesc': '验证输入输出',
         'detail.chat.welcome.samplePrompt': '请先读取 {{guidePath}} 后验证样例。工作流 ID 是 {{id}}，工作流目录是 {{dir}}。',
+        'modelPicker.auto': 'Auto',
+        'modelPicker.title': '选择模型',
+        'modelPicker.hint': '作为本次对话发送时的模型覆盖',
+        'modelPicker.empty': '暂无模型',
         'detail.run.guideApiShort': '发布为 API',
         'detail.run.guideApiDesc': '配置 API 发布',
         'detail.run.guideApiInstruction': '围绕 API 发布读取 guide.md，先 GET {{configEndpoint}}，确认后 PUT {{configEndpoint}}，config.json 和 workflow.json 不是直接写入目标；后端接口不可用时必须停止配置流程',
@@ -266,6 +270,31 @@ function renderChatTab(props: Partial<ComponentProps<typeof ChatTab>> = {}) {
   );
 }
 
+function configureMinimaxChatModel(supportsVision = false) {
+  mockUseProviders.mockReturnValue({
+    providers: [{ id: 'minimax', name: 'MiniMax', configured: true }],
+    connectedIds: ['minimax'],
+    loading: false,
+    error: null,
+    refetch: vi.fn(),
+  });
+  mockModelListDefinitions.mockResolvedValue({
+    data: {
+      models: [{
+        provider_id: 'minimax',
+        id: 'minimax-m3',
+        name: 'minimax-m3',
+        model_type: 'chat',
+        capabilities: { supports_vision: supportsVision },
+        limits: { context_window: 128000 },
+      }],
+    },
+  });
+  mockDefaultModelGetResolved.mockResolvedValue({
+    data: { provider_id: 'minimax', model_id: 'minimax-m3' },
+  });
+}
+
 describe('WorkflowDetail ChatTab', () => {
   beforeEach(() => {
     vi.clearAllMocks();
@@ -709,6 +738,7 @@ describe('WorkflowDetail ChatTab', () => {
       provider: 'minimax',
       model: 'minimax-m3',
       model_pinned: true,
+      model_auto: false,
     });
     await waitFor(() => {
       const latestProps = capturedSessionChatProps[capturedSessionChatProps.length - 1];
@@ -716,6 +746,63 @@ describe('WorkflowDetail ChatTab', () => {
     });
   });
 
+  it('restores Auto from the active workflow chat session', async () => {
+    setStoredSessions(workflow.id, [
+      { id: 'workflow-session-auto', title: 'Existing', createdAt: Date.now() },
+    ]);
+    configureMinimaxChatModel(true);
+    mockClientGet.mockResolvedValue({
+      data: {
+        id: 'workflow-session-auto',
+        model_auto: true,
+        model_pinned: false,
+      },
+    });
+
+    renderChatTab();
+
+    await waitFor(() => {
+      const latestProps = capturedSessionChatProps[capturedSessionChatProps.length - 1];
+      expect(latestProps.sessionId).toBe('workflow-session-auto');
+      expect(latestProps.modelAuto).toBe(true);
+      expect(latestProps.model).toBeNull();
+      expect(latestProps.supportsVision).toBe(true);
+    });
+    expect(capturedSessionOptions[capturedSessionOptions.length - 1].modelAuto).toBe(true);
+  });
+
+  it('manually enables Auto for the active workflow chat session', async () => {
+    const user = userEvent.setup();
+    setStoredSessions(workflow.id, [
+      { id: 'workflow-session-enable-auto', title: 'Existing', createdAt: Date.now() },
+    ]);
+    configureMinimaxChatModel();
+    mockClientGet.mockResolvedValue({
+      data: {
+        id: 'workflow-session-enable-auto',
+        provider: 'minimax',
+        model: 'minimax-m3',
+        model_pinned: true,
+        model_auto: false,
+      },
+    });
+
+    renderChatTab();
+
+    await user.click(await screen.findByRole('button', { name: /minimax-m3/i }));
+    await user.click(await screen.findByRole('button', { name: 'Auto' }));
+
+    expect(mockClientPatch).toHaveBeenCalledWith('/api/session/workflow-session-enable-auto', {
+      model_auto: true,
+      model_pinned: false,
+    });
+    await waitFor(() => {
+      const latestProps = capturedSessionChatProps[capturedSessionChatProps.length - 1];
+      expect(latestProps.modelAuto).toBe(true);
+      expect(latestProps.model).toBeNull();
+    });
+  });
+
   it('keeps the workflow composer compact enough for guide shortcuts above it', () => {
     renderChatTab();
 
diff --git a/webui/src/pages/WorkflowDetail/tabs/ChatTab.tsx b/webui/src/pages/WorkflowDetail/tabs/ChatTab.tsx
index 46bdb7a86..192c23c61 100644
--- a/webui/src/pages/WorkflowDetail/tabs/ChatTab.tsx
+++ b/webui/src/pages/WorkflowDetail/tabs/ChatTab.tsx
@@ -120,10 +120,17 @@ export default function ChatTab({
     groupedOptions: groupedChatModelOptions,
     loading: loadingChatModels,
     options: chatModelOptions,
+    canSelectAuto,
+    effectiveModelOption,
+    modelPickerAutoOption,
+    selectAuto,
+    selectModelKey,
+    selectedModelAuto,
     selectedModelOption,
     selectedPromptModel,
+    setSelectedModelAuto,
     setSelectedModelKey,
-  } = useChatModelOptions();
+  } = useChatModelOptions({ enableAuto: true });
   const sessionPinnedModelKey = sessionPinnedModel
     ? `${sessionPinnedModel.providerID}::${sessionPinnedModel.modelID}`
     : null;
@@ -131,17 +138,26 @@ export default function ChatTab({
     sessionPinnedModelKey && chatModelOptions.some((option) => option.key === sessionPinnedModelKey),
   );
   const effectiveSessionPinnedModel = hasSessionPinnedModelOption ? sessionPinnedModel : null;
-  const effectiveSupportsVision = selectedModelOption?.supportsVision ?? defaultSupportsVision;
+  const effectiveSupportsVision = effectiveModelOption?.supportsVision ?? defaultSupportsVision;
 
   const applySessionModel = useCallback((session: unknown) => {
     const data = session as {
       provider?: unknown;
       model?: unknown;
+      model_auto?: unknown;
       model_pinned?: unknown;
     } | null;
     const providerID = typeof data?.provider === 'string' ? data.provider : '';
     const modelID = typeof data?.model === 'string' ? data.model : '';
 
+    if (data?.model_auto) {
+      setSessionPinnedModel(null);
+      setSelectedModelAuto(true);
+      return;
+    }
+
+    setSelectedModelAuto(false);
+
     if (data?.model_pinned && providerID && modelID) {
       setSessionPinnedModel({ providerID, modelID });
       setSelectedModelKey(`${providerID}::${modelID}`);
@@ -150,7 +166,7 @@ export default function ChatTab({
 
     setSessionPinnedModel(null);
     setSelectedModelKey(null);
-  }, [setSelectedModelKey]);
+  }, [setSelectedModelAuto, setSelectedModelKey]);
 
   useEffect(() => {
     workflowRevisionRef.current = workflowRevisionKey(workflow);
@@ -190,6 +206,7 @@ export default function ChatTab({
   } = useSessionChat({
     title: t('detail.chat.sessionTitle', { name: workflowDisplayName }),
     category: 'workflow',
+    modelAuto: selectedModelAuto,
     contextMessage: [
       t('detail.chat.contextMessage', workflowChatPromptParams),
       workflowChatPromptParams.backendConfigAccessGuide,
@@ -214,6 +231,7 @@ export default function ChatTab({
 
     if (!sessionId) {
       setSessionPinnedModel(null);
+      setSelectedModelAuto(false);
       setSelectedModelKey(null);
       return;
     }
@@ -227,13 +245,14 @@ export default function ChatTab({
       .catch(() => {
         if (cancelled || hydrationId !== sessionModelHydrationRef.current) return;
         setSessionPinnedModel(null);
+        setSelectedModelAuto(false);
         setSelectedModelKey(null);
       });
 
     return () => {
       cancelled = true;
     };
-  }, [applySessionModel, sessionId]);
+  }, [applySessionModel, sessionId, setSelectedModelAuto, setSelectedModelKey]);
 
   // Load stored sessions and validate only the active one (lightweight check)
   useEffect(() => {
@@ -339,6 +358,7 @@ export default function ChatTab({
             imageParts,
             agent: effectiveAgent,
             model: effectiveModel,
+            modelAuto: selectedModelAuto,
             displayText: effectiveDisplayText,
           });
         } else {
@@ -350,7 +370,7 @@ export default function ChatTab({
         setInitialMessage(null);
       }
     },
-    [onFirstMessageSent, selectedPromptModel, createSession, createAndSendSession],
+    [onFirstMessageSent, selectedModelAuto, selectedPromptModel, createSession, createAndSendSession],
   );
 
   const handleWelcomeGuidePrompt = useCallback(
@@ -385,7 +405,7 @@ export default function ChatTab({
     sessionModelHydrationRef.current += 1;
     const nextModel = { providerID: option.providerID, modelID: option.modelID };
     setSessionPinnedModel(nextModel);
-    setSelectedModelKey(option.key);
+    selectModelKey(option.key);
 
     if (!sessionId) return;
 
@@ -393,10 +413,26 @@ export default function ChatTab({
       provider: option.providerID,
       model: option.modelID,
       model_pinned: true,
+      model_auto: false,
     }).catch((err) => {
       console.warn('[WorkflowDetail] failed to pin workflow chat model', err);
     });
-  }, [sessionId, setSelectedModelKey]);
+  }, [selectModelKey, sessionId]);
+
+  const handleSelectAuto = useCallback(() => {
+    if (!canSelectAuto) return;
+    sessionModelHydrationRef.current += 1;
+    setSessionPinnedModel(null);
+    selectAuto();
+
+    if (!sessionId) return;
+    client.patch(`/api/session/${sessionId}`, {
+      model_auto: true,
+      model_pinned: false,
+    }).catch((err) => {
+      console.warn('[WorkflowDetail] failed to enable Auto for workflow chat', err);
+    });
+  }, [canSelectAuto, selectAuto, sessionId]);
 
   // Helper: fetch fresh workflow and notify parent if updated
   const checkWorkflowUpdate = useCallback(async () => {
@@ -561,8 +597,9 @@ export default function ChatTab({
           initialMessage={initialMessage}
           onSSEEvent={handleSSEEvent}
           supportsVision={effectiveSupportsVision}
-          contextWindowTokens={selectedModelOption?.contextWindowTokens ?? null}
+          contextWindowTokens={effectiveModelOption?.contextWindowTokens ?? null}
           model={effectiveSessionPinnedModel ?? selectedPromptModel}
+          modelAuto={selectedModelAuto}
           onCreateAndSend={!sessionId ? handleCreateAndSend : undefined}
           composerTextareaMinHeight={48}
           composerTextareaMaxHeight={120}
@@ -578,6 +615,9 @@ export default function ChatTab({
               loading={loadingChatModels}
               selectedModelOption={selectedModelOption}
               onSelectModel={handleSelectModel}
+              autoOption={modelPickerAutoOption
+                ? { ...modelPickerAutoOption, onSelect: handleSelectAuto }
+                : undefined}
             />
           }
           conversationBottomSlot={({ sendPrompt, sending, streaming }) => (
diff --git a/webui/src/styles/index.css b/webui/src/styles/index.css
index 00366a553..2d8ab4abd 100644
--- a/webui/src/styles/index.css
+++ b/webui/src/styles/index.css
@@ -1388,6 +1388,54 @@
   background: #687584;
 }
 
+.session-sidebar-scrollbar {
+  scrollbar-color: #c4c8cc transparent;
+  scrollbar-width: thin;
+}
+
+.session-sidebar-scrollbar::-webkit-scrollbar {
+  width: 7px;
+}
+
+.session-sidebar-scrollbar::-webkit-scrollbar-track {
+  background: transparent;
+}
+
+.session-sidebar-scrollbar::-webkit-scrollbar-thumb {
+  background: #c4c8cc;
+  border: 2px solid transparent;
+  border-radius: 999px;
+  background-clip: padding-box;
+}
+
+.session-sidebar-scrollbar::-webkit-scrollbar-thumb:hover {
+  background: #a9aeb4;
+  background-clip: padding-box;
+}
+
+.dark .session-sidebar-scrollbar {
+  scrollbar-color: #626e7c transparent !important;
+}
+
+.dark .session-sidebar-scrollbar::-webkit-scrollbar {
+  width: 7px;
+}
+
+.dark .session-sidebar-scrollbar::-webkit-scrollbar-track {
+  background: transparent;
+}
+
+.dark .session-sidebar-scrollbar::-webkit-scrollbar-thumb {
+  background: #626e7c;
+  border-color: transparent;
+  background-clip: padding-box;
+}
+
+.dark .session-sidebar-scrollbar::-webkit-scrollbar-thumb:hover {
+  background: #758292;
+  background-clip: padding-box;
+}
+
 .dark table {
   color: var(--flocks-dark-text-muted);
 }
diff --git a/webui/src/types/index.ts b/webui/src/types/index.ts
index 7314c7725..62a72882e 100644
--- a/webui/src/types/index.ts
+++ b/webui/src/types/index.ts
@@ -5,22 +5,26 @@
  */
 export interface Session {
   id: string;
-  slug: string;
+  slug?: string;
   projectID: string;
+  projectName?: string;
   effectiveProjectID?: string;
   directory: string;
   parentID?: string;
   summary?: SessionSummary;
   title: string;
-  version: string;
+  version?: string;
   time: SessionTime;
   permission?: PermissionRule[];
   revert?: SessionRevert;
   /** Session category: 'user' | 'workflow' | 'task' | 'entity-config' | ... */
   category?: string;
+  status?: 'active' | 'archived';
   provider?: string;
   model?: string;
   model_pinned?: boolean;
+  /** Runtime provider failover is enabled only after an explicit WebUI selection. */
+  model_auto?: boolean;
   ownerUserID?: string;
   ownerUsername?: string;
   canWrite?: boolean;
@@ -432,10 +436,11 @@ export interface ProviderInfoV2 {
   source: string;
   env: string[];
   key: string | null;
+  /** True only when the runtime has the credentials required by this provider. */
+  configured?: boolean;
   options: Record<string, any>;
   models: Record<string, ProviderModelInfo>;
   // Derived on frontend
-  configured?: boolean;
   modelCount?: number;
   category?: ProviderCategory;
 }
diff --git a/webui/src/utils/messageId.test.ts b/webui/src/utils/messageId.test.ts
new file mode 100644
index 000000000..b068b795b
--- /dev/null
+++ b/webui/src/utils/messageId.test.ts
@@ -0,0 +1,23 @@
+import { describe, expect, it } from 'vitest';
+
+import { createMessageId } from './messageId';
+
+describe('createMessageId', () => {
+  it('creates a canonical message ID with its timestamp bits encoded', () => {
+    const timestamp = 1_700_000_000_000;
+    const id = createMessageId(timestamp);
+
+    expect(id).toMatch(/^msg_[0-9a-f]{12}[0-9A-Za-z]{14}$/);
+    const encodedTime = BigInt(`0x${id.slice(4, 16)}`);
+    expect(encodedTime >> 12n).toBe(BigInt(timestamp) & ((1n << 36n) - 1n));
+  });
+
+  it('keeps IDs unique and ordered within the same millisecond', () => {
+    const timestamp = 1_700_000_000_001;
+    const first = createMessageId(timestamp);
+    const second = createMessageId(timestamp);
+
+    expect(second).not.toBe(first);
+    expect(second > first).toBe(true);
+  });
+});
diff --git a/webui/src/utils/messageId.ts b/webui/src/utils/messageId.ts
new file mode 100644
index 000000000..0ab7bd224
--- /dev/null
+++ b/webui/src/utils/messageId.ts
@@ -0,0 +1,24 @@
+const BASE62 = '0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz';
+const RANDOM_SUFFIX_LENGTH = 14;
+
+let lastTimestamp = 0;
+let counter = 0;
+
+export function createMessageId(timestamp = Date.now()): string {
+  if (timestamp !== lastTimestamp) {
+    lastTimestamp = timestamp;
+    counter = 0;
+  }
+  counter += 1;
+
+  const encodedTime = BigInt(timestamp) * 0x1000n + BigInt(counter);
+  const timeHex = encodedTime.toString(16).slice(-12).padStart(12, '0');
+  const randomBytes = new Uint8Array(RANDOM_SUFFIX_LENGTH);
+  globalThis.crypto.getRandomValues(randomBytes);
+  const randomSuffix = Array.from(
+    randomBytes,
+    (byte) => BASE62[byte % BASE62.length],
+  ).join('');
+
+  return `msg_${timeHex}${randomSuffix}`;
+}
diff --git a/webui/src/utils/restartPolling.test.ts b/webui/src/utils/restartPolling.test.ts
index 5e456d8d2..b2aae340f 100644
--- a/webui/src/utils/restartPolling.test.ts
+++ b/webui/src/utils/restartPolling.test.ts
@@ -17,6 +17,7 @@ describe('checkRestartReadiness', () => {
       configurable: true,
       value: originalLocation,
     });
+    vi.useRealTimers();
     vi.restoreAllMocks();
   });
 
@@ -34,8 +35,55 @@ describe('checkRestartReadiness', () => {
       ready: false,
       reason: 'health check returned HTTP 404',
     });
-    expect(fetchMock).toHaveBeenCalledWith('/api/health', { cache: 'no-store' });
-    expect(fetchMock).toHaveBeenCalledWith('/', { cache: 'no-store' });
-    expect(fetchMock).not.toHaveBeenCalledWith('http://127.0.0.1:8000/api/health', { cache: 'no-store' });
+    expect(fetchMock.mock.calls.map(([input]) => String(input))).toEqual([
+      '/api/health',
+      '/',
+    ]);
+    expect(fetchMock).toHaveBeenNthCalledWith(
+      1,
+      '/api/health',
+      expect.objectContaining({
+        cache: 'no-store',
+        signal: expect.anything(),
+      }),
+    );
+    expect(fetchMock).toHaveBeenNthCalledWith(
+      2,
+      '/',
+      expect.objectContaining({
+        cache: 'no-store',
+        signal: expect.anything(),
+      }),
+    );
+  });
+
+  it('aborts stalled readiness requests so polling can continue', async () => {
+    vi.useFakeTimers();
+    const requestSignals: AbortSignal[] = [];
+    const fetchMock = vi.fn((_input: RequestInfo | URL, init?: RequestInit) => {
+      const signal = init?.signal;
+      if (!signal) {
+        throw new Error('expected an abort signal');
+      }
+      requestSignals.push(signal);
+      return new Promise<Response>((_resolve, reject) => {
+        signal.addEventListener('abort', () => {
+          reject(new DOMException('The operation was aborted.', 'AbortError'));
+        }, { once: true });
+      });
+    });
+    vi.stubGlobal('fetch', fetchMock);
+
+    const readinessPromise = checkRestartReadiness();
+    await vi.advanceTimersByTimeAsync(3_000);
+    await vi.advanceTimersByTimeAsync(3_000);
+
+    const readiness = await readinessPromise;
+    expect(readiness.ready).toBe(false);
+    expect(readiness.reason).toContain('health check failed');
+    expect(readiness.reason).toContain('root page check failed');
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+    expect(requestSignals).toHaveLength(2);
+    expect(requestSignals.every((signal) => signal.aborted)).toBe(true);
   });
 });
diff --git a/webui/src/utils/restartPolling.ts b/webui/src/utils/restartPolling.ts
index dc142d5c1..19ccdad9b 100644
--- a/webui/src/utils/restartPolling.ts
+++ b/webui/src/utils/restartPolling.ts
@@ -1,4 +1,5 @@
 const UPGRADE_PAGE_MARKER = 'flocks-upgrade-in-progress';
+const RESTART_READINESS_REQUEST_TIMEOUT_MS = 3_000;
 
 export interface RestartReadiness {
   ready: boolean;
@@ -11,9 +12,25 @@ function errorMessage(error: unknown): string {
   return 'request failed';
 }
 
+async function fetchWithTimeout(url: string): Promise<Response> {
+  const controller = new AbortController();
+  const timeoutId = window.setTimeout(() => {
+    controller.abort();
+  }, RESTART_READINESS_REQUEST_TIMEOUT_MS);
+
+  try {
+    return await fetch(url, {
+      cache: 'no-store',
+      signal: controller.signal,
+    });
+  } finally {
+    window.clearTimeout(timeoutId);
+  }
+}
+
 async function readUpgradePageState(): Promise<string | null> {
   try {
-    const rootResponse = await fetch('/', { cache: 'no-store' });
+    const rootResponse = await fetchWithTimeout('/');
     if (!rootResponse.ok) {
       return `root page returned HTTP ${rootResponse.status}`;
     }
@@ -31,7 +48,7 @@ async function readUpgradePageState(): Promise<string | null> {
 
 async function checkHealth(url: string): Promise<Response | null> {
   try {
-    return await fetch(url, { cache: 'no-store' });
+    return await fetchWithTimeout(url);
   } catch {
     return null;
   }
diff --git a/webui/src/utils/sessionExecutionMode.test.ts b/webui/src/utils/sessionExecutionMode.test.ts
new file mode 100644
index 000000000..98b3ca40d
--- /dev/null
+++ b/webui/src/utils/sessionExecutionMode.test.ts
@@ -0,0 +1,54 @@
+import { beforeEach, describe, expect, it } from 'vitest';
+
+import {
+  DEFAULT_SESSION_EXECUTION_MODE,
+  EXECUTION_MODE_DRAFT_STORAGE_KEY,
+  EXECUTION_MODE_STORAGE_PREFIX,
+  promoteDraftExecutionMode,
+  readSessionExecutionMode,
+  resetDraftExecutionMode,
+  writeSessionExecutionMode,
+} from './sessionExecutionMode';
+
+describe('sessionExecutionMode storage', () => {
+  beforeEach(() => {
+    localStorage.clear();
+  });
+
+  it('defaults new and existing composers to Build', () => {
+    expect(readSessionExecutionMode()).toBe(DEFAULT_SESSION_EXECUTION_MODE);
+    expect(readSessionExecutionMode('session-1')).toBe(DEFAULT_SESSION_EXECUTION_MODE);
+  });
+
+  it('persists Plan independently by session', () => {
+    writeSessionExecutionMode('session-1', 'plan');
+
+    expect(readSessionExecutionMode('session-1')).toBe('plan');
+  });
+
+  it('promotes the draft mode to a newly created session', () => {
+    writeSessionExecutionMode(null, 'plan');
+    promoteDraftExecutionMode('session-new', 'plan');
+
+    expect(readSessionExecutionMode('session-new')).toBe('plan');
+    expect(localStorage.getItem(EXECUTION_MODE_DRAFT_STORAGE_KEY)).toBeNull();
+  });
+
+  it('stores Build by removing the override', () => {
+    localStorage.setItem(`${EXECUTION_MODE_STORAGE_PREFIX}session-1`, 'plan');
+    writeSessionExecutionMode('session-1', 'build');
+    writeSessionExecutionMode(null, 'plan');
+    resetDraftExecutionMode();
+
+    expect(localStorage.getItem(`${EXECUTION_MODE_STORAGE_PREFIX}session-1`)).toBeNull();
+    expect(localStorage.getItem(EXECUTION_MODE_DRAFT_STORAGE_KEY)).toBeNull();
+  });
+
+  it('ignores removed and one-shot persisted modes', () => {
+    localStorage.setItem(`${EXECUTION_MODE_STORAGE_PREFIX}session-1`, 'ask');
+    expect(readSessionExecutionMode('session-1')).toBe('build');
+
+    localStorage.setItem(`${EXECUTION_MODE_STORAGE_PREFIX}session-1`, 'goal');
+    expect(readSessionExecutionMode('session-1')).toBe('build');
+  });
+});
diff --git a/webui/src/utils/sessionExecutionMode.ts b/webui/src/utils/sessionExecutionMode.ts
new file mode 100644
index 000000000..386eb56b5
--- /dev/null
+++ b/webui/src/utils/sessionExecutionMode.ts
@@ -0,0 +1,62 @@
+export type SessionExecutionMode = 'build' | 'plan' | 'goal';
+export type PersistentSessionExecutionMode = Exclude<SessionExecutionMode, 'goal'>;
+
+export const DEFAULT_SESSION_EXECUTION_MODE: PersistentSessionExecutionMode = 'build';
+export const EXECUTION_MODE_STORAGE_PREFIX = 'flocks:session-execution-mode:';
+export const EXECUTION_MODE_DRAFT_STORAGE_KEY = `${EXECUTION_MODE_STORAGE_PREFIX}draft`;
+
+function isPersistentMode(value: unknown): value is PersistentSessionExecutionMode {
+  return value === 'build' || value === 'plan';
+}
+
+function storageKey(sessionId?: string | null): string {
+  return sessionId
+    ? `${EXECUTION_MODE_STORAGE_PREFIX}${sessionId}`
+    : EXECUTION_MODE_DRAFT_STORAGE_KEY;
+}
+
+export function readSessionExecutionMode(
+  sessionId?: string | null,
+): PersistentSessionExecutionMode {
+  if (typeof window === 'undefined') return DEFAULT_SESSION_EXECUTION_MODE;
+  try {
+    const stored = window.localStorage.getItem(storageKey(sessionId));
+    return isPersistentMode(stored) ? stored : DEFAULT_SESSION_EXECUTION_MODE;
+  } catch {
+    return DEFAULT_SESSION_EXECUTION_MODE;
+  }
+}
+
+export function writeSessionExecutionMode(
+  sessionId: string | null | undefined,
+  mode: PersistentSessionExecutionMode,
+): void {
+  if (typeof window === 'undefined') return;
+  try {
+    const key = storageKey(sessionId);
+    if (mode === DEFAULT_SESSION_EXECUTION_MODE) {
+      window.localStorage.removeItem(key);
+    } else {
+      window.localStorage.setItem(key, mode);
+    }
+  } catch {
+    // Composer preferences must never block the chat flow.
+  }
+}
+
+export function promoteDraftExecutionMode(
+  sessionId: string,
+  mode: PersistentSessionExecutionMode,
+): void {
+  writeSessionExecutionMode(sessionId, mode);
+  if (typeof window === 'undefined') return;
+  try {
+    window.localStorage.removeItem(EXECUTION_MODE_DRAFT_STORAGE_KEY);
+  } catch {
+    // Composer preferences must never block the chat flow.
+  }
+}
+
+export function resetDraftExecutionMode(): void {
+  writeSessionExecutionMode(null, DEFAULT_SESSION_EXECUTION_MODE);
+}
diff --git a/webui/src/utils/socAlertFilterValues.test.ts b/webui/src/utils/socAlertFilterValues.test.ts
new file mode 100644
index 000000000..4e9a037e7
--- /dev/null
+++ b/webui/src/utils/socAlertFilterValues.test.ts
@@ -0,0 +1,30 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  filterOptionText,
+  matchesFilterOptionSearch,
+} from '../../../.flocks/flockshub/plugins/webuis/soc_ui/soc_alerts/src/filterValues';
+
+const EN_TEXT: Record<string, string> = {
+  '访问': 'Access',
+  '高危': 'High',
+};
+const english = (text: string) => EN_TEXT[text] || text;
+
+describe('SOC alert filter value localization', () => {
+  it('localizes fixed values and preserves dynamic threat types', () => {
+    expect(filterOptionText('threat_phase', 'access')).toBe('访问');
+    expect(filterOptionText('threat_phase', 'access', english)).toBe('Access');
+    expect(filterOptionText('threat_type', 'SQL injection')).toBe('SQL injection');
+    expect(filterOptionText('threat_type', 'SQL injection', english)).toBe('SQL injection');
+  });
+
+  it('matches raw, Chinese, and English option text regardless of locale', () => {
+    expect(matchesFilterOptionSearch('threat_level', 'high', 'high')).toBe(true);
+    expect(matchesFilterOptionSearch('threat_level', 'high', '高危', english)).toBe(true);
+    expect(matchesFilterOptionSearch('threat_level', 'high', 'High', english)).toBe(true);
+    expect(matchesFilterOptionSearch('threat_type', 'SQL injection', 'SQL injection', english)).toBe(true);
+    expect(matchesFilterOptionSearch('threat_type', 'SQL injection', 'SQL 注入', english)).toBe(false);
+    expect(matchesFilterOptionSearch('threat_type', 'SQL injection', 'crawler', english)).toBe(false);
+  });
+});
diff --git a/webui/src/utils/socDashboardSeverity.test.ts b/webui/src/utils/socDashboardSeverity.test.ts
new file mode 100644
index 000000000..65966a26d
--- /dev/null
+++ b/webui/src/utils/socDashboardSeverity.test.ts
@@ -0,0 +1,31 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  severityKey,
+  severityRows,
+} from '../../../.flocks/flockshub/plugins/webuis/soc_ui/soc_dashboard/src/severityValues';
+
+describe('SOC dashboard severity values', () => {
+  it('builds severity rows only from threat severity values', () => {
+    expect(severityRows({
+      severityLevels: [
+        { label: 'critical', value: 2 },
+        { label: '3', value: 3 },
+        { label: 'medium', value: 5 },
+        { label: 'low', value: 7 },
+        { label: 'unknown', value: 11 },
+      ],
+    }).map((item) => [item.key, item.value])).toEqual([
+      ['critical', 2],
+      ['high', 3],
+      ['medium', 5],
+      ['low', 7],
+    ]);
+  });
+
+  it('does not derive severity from attack verdicts', () => {
+    expect(severityRows({ severityLevels: [] }).every((item) => item.value === 0)).toBe(true);
+    expect(severityKey('attack_success')).toBe('');
+    expect(severityKey('benign')).toBe('');
+  });
+});
diff --git a/webui/src/utils/time.test.ts b/webui/src/utils/time.test.ts
new file mode 100644
index 000000000..afd3f22cd
--- /dev/null
+++ b/webui/src/utils/time.test.ts
@@ -0,0 +1,44 @@
+import { describe, expect, it } from 'vitest';
+
+import { formatRelativeTime, formatSmartTime } from './time';
+
+describe('formatSmartTime', () => {
+  const at = (year: number, month: number, day: number, hour = 15, minute = 34) =>
+    new Date(year, month - 1, day, hour, minute).getTime();
+
+  it('uses a compact 24-hour clock for messages from today', () => {
+    expect(formatSmartTime(at(2026, 7, 24), 'zh-CN', at(2026, 7, 24, 18, 0))).toBe('15:34');
+  });
+
+  it('labels messages from yesterday', () => {
+    expect(formatSmartTime(at(2026, 7, 23), 'zh-CN', at(2026, 7, 24, 18, 0))).toBe('昨天 15:34');
+  });
+
+  it('uses the weekday for messages within the previous week', () => {
+    expect(formatSmartTime(at(2026, 7, 23), 'zh-CN', at(2026, 7, 25, 18, 0))).toBe('星期四 15:34');
+  });
+
+  it('uses a Chinese month and day for older messages from the same year', () => {
+    expect(formatSmartTime(at(2026, 7, 10), 'zh-CN', at(2026, 7, 24, 18, 0))).toBe('7月10日 15:34');
+    expect(formatSmartTime(at(2026, 6, 20), 'zh-CN', at(2026, 7, 24, 18, 0))).toBe('6月20日 15:34');
+  });
+});
+
+describe('formatRelativeTime', () => {
+  const now = new Date('2026-07-24T12:00:00+08:00').getTime();
+
+  it('formats recent session activity as compact Chinese relative time', () => {
+    expect(formatRelativeTime(now - 17 * 60 * 60 * 1000, 'zh-CN', now)).toBe('17小时前');
+    expect(formatRelativeTime(now - 12 * 60 * 1000, 'zh-CN', now)).toBe('12分钟前');
+    expect(formatRelativeTime(now - 3 * 24 * 60 * 60 * 1000, 'zh-CN', now)).toBe('3天前');
+  });
+
+  it('uses the requested locale', () => {
+    expect(formatRelativeTime(now - 17 * 60 * 60 * 1000, 'en-US', now)).toBe('17 hours ago');
+  });
+
+  it('uses a compact just-now label for activity within the last minute', () => {
+    expect(formatRelativeTime(now - 20 * 1000, 'zh-CN', now)).toBe('刚刚');
+    expect(formatRelativeTime(now - 20 * 1000, 'en-US', now)).toBe('just now');
+  });
+});
diff --git a/webui/src/utils/time.ts b/webui/src/utils/time.ts
index c0ee42085..fc3eaed40 100644
--- a/webui/src/utils/time.ts
+++ b/webui/src/utils/time.ts
@@ -5,25 +5,44 @@
  * duplicate implementations across components.
  */
 
-export function formatSmartTime(timestamp: number): string {
+export function formatSmartTime(
+  timestamp: number,
+  locale = 'zh-CN',
+  nowTimestamp = Date.now(),
+): string {
   const d = new Date(timestamp);
-  const now = new Date();
-  const diffMs = now.getTime() - d.getTime();
-  const diffDays = Math.floor(diffMs / (1000 * 60 * 60 * 24));
+  const now = new Date(nowTimestamp);
+  const diffDays = localCalendarDayNumber(now) - localCalendarDayNumber(d);
+  const time = formatClockTime(d, locale);
 
   if (diffDays === 0) {
-    // 今天：只显示时间
-    return formatTime12h(d);
+    return time;
   }
   if (diffDays === 1) {
-    // 昨天
-    return `昨天 ${formatTime12h(d)}`;
+    return locale.toLowerCase().startsWith('zh')
+      ? `昨天 ${time}`
+      : `Yesterday ${time}`;
+  }
+  if (diffDays > 1 && diffDays < 7) {
+    const weekday = new Intl.DateTimeFormat(locale, { weekday: 'long' }).format(d);
+    return `${weekday} ${time}`;
   }
   if (now.getFullYear() === d.getFullYear()) {
-    // 今年：不显示年
-    return `${d.getMonth() + 1}/${d.getDate()} ${formatTime12h(d)}`;
+    if (locale.toLowerCase().startsWith('zh')) {
+      return `${d.getMonth() + 1}月${d.getDate()}日 ${time}`;
+    }
+    const date = new Intl.DateTimeFormat(locale, { month: 'short', day: 'numeric' }).format(d);
+    return `${date} ${time}`;
   }
-  return `${d.getFullYear()}/${d.getMonth() + 1}/${d.getDate()} ${formatTime12h(d)}`;
+  if (locale.toLowerCase().startsWith('zh')) {
+    return `${d.getFullYear()}年${d.getMonth() + 1}月${d.getDate()}日 ${time}`;
+  }
+  const date = new Intl.DateTimeFormat(locale, {
+    year: 'numeric',
+    month: 'short',
+    day: 'numeric',
+  }).format(d);
+  return `${date} ${time}`;
 }
 
 export function formatSessionDate(ts: number): string {
@@ -32,6 +51,38 @@ export function formatSessionDate(ts: number): string {
   return `${d.getFullYear()}/${d.getMonth() + 1}/${d.getDate()} ${formatTime12h(d)}`;
 }
 
+export function formatRelativeTime(
+  timestamp: number,
+  locale = 'zh-CN',
+  now = Date.now(),
+): string {
+  if (!timestamp || !Number.isFinite(timestamp)) return '';
+
+  const diffMs = timestamp - now;
+  const absDiffMs = Math.abs(diffMs);
+  const isChinese = locale.toLowerCase().startsWith('zh');
+
+  if (absDiffMs < 60_000) {
+    return isChinese ? '刚刚' : 'just now';
+  }
+
+  const units: Array<{
+    unit: Intl.RelativeTimeFormatUnit;
+    milliseconds: number;
+    upperBound: number;
+  }> = [
+    { unit: 'minute', milliseconds: 60_000, upperBound: 60 * 60_000 },
+    { unit: 'hour', milliseconds: 60 * 60_000, upperBound: 24 * 60 * 60_000 },
+    { unit: 'day', milliseconds: 24 * 60 * 60_000, upperBound: 30 * 24 * 60 * 60_000 },
+    { unit: 'month', milliseconds: 30 * 24 * 60 * 60_000, upperBound: 365 * 24 * 60 * 60_000 },
+    { unit: 'year', milliseconds: 365 * 24 * 60 * 60_000, upperBound: Number.POSITIVE_INFINITY },
+  ];
+  const selected = units.find(({ upperBound }) => absDiffMs < upperBound) || units[units.length - 1];
+  const value = Math.trunc(diffMs / selected.milliseconds) || (diffMs < 0 ? -1 : 1);
+
+  return new Intl.RelativeTimeFormat(locale, { numeric: 'always' }).format(value, selected.unit);
+}
+
 function formatTime12h(d: Date): string {
   let hours = d.getHours();
   const minutes = d.getMinutes().toString().padStart(2, '0');
@@ -40,3 +91,17 @@ function formatTime12h(d: Date): string {
   if (hours === 0) hours = 12;
   return `${hours}:${minutes} ${ampm}`;
 }
+
+function formatClockTime(d: Date, locale: string): string {
+  if (locale.toLowerCase().startsWith('zh')) {
+    return `${d.getHours().toString().padStart(2, '0')}:${d.getMinutes().toString().padStart(2, '0')}`;
+  }
+  return new Intl.DateTimeFormat(locale, {
+    hour: 'numeric',
+    minute: '2-digit',
+  }).format(d);
+}
+
+function localCalendarDayNumber(d: Date): number {
+  return Math.floor(Date.UTC(d.getFullYear(), d.getMonth(), d.getDate()) / 86_400_000);
+}