diff --git a/README.en.md b/README.en.md
index 8540752..9799dca 100644
--- a/README.en.md
+++ b/README.en.md
@@ -24,6 +24,8 @@ QuietHarness does not include a task database, background automation, or a team
 
 If you are exploring how multiple long-lived owners/workers can share canonical state, use semantic handoffs, reject false completion, and continue the same task after worker failover, see the experimental [OPC Company Layer](labs/opc-company/README.en.md). It is a governance lab in this same repository, not a QuietHarness installation requirement.
 
+If your agents can already keep working but keep interrupting you for the wrong reasons, see the experimental [Attention Inbox Lab](labs/attention-inbox/README.md). It keeps routine completions and internal waiting states quiet, and surfaces only concrete human decisions or meaningful risks. It is a derived attention layer over existing task/receipt state, not another queue or scheduler.
+
 ## Try it inside one isolated project
 
 ```bash
diff --git a/README.md b/README.md
index 1268d0c..28678c1 100644
--- a/README.md
+++ b/README.md
@@ -24,6 +24,8 @@ QuietHarness 不会给每个请求套一层仪式，也不是要求你复制我
 
 如果你关心多个长期 Owner / Worker 如何共享 canonical state、做 semantic handoff、拒绝假完成并在换 worker 后继续同一任务，可以看实验性的 [OPC Company Layer](labs/opc-company/README.md)。它是同一仓库里的治理实验层，不是 QuietHarness 的安装前提。
 
+如果你的 Agent 已经能持续工作，但经常因为错误的 `waiting_for_user`、常规完成或重复状态打断你，可以看实验性的 [Attention Inbox Lab](labs/attention-inbox/README.md)。它只把真正的人工决策和有意义的风险送到注意力入口；任务事实、调度和执行仍留在现有系统里，不另造一套队列或控制面。
+
 ## 先在一个隔离项目试用
 
 ```bash
@@ -273,6 +275,7 @@ templates/                      # 当前三端小 Core
 docs/                           # 我的系统架构、任务、收尾与日报边界
 examples/first-success/         # 单客户端隔离试用与行为验收
 examples/leo-system/            # 脱敏的真实结构示例
+labs/attention-inbox/           # 实验性的 Agent 注意力协议与策略
 scripts/                        # 盘点、安装、验证
 tests/                          # 隔离 HOME 冒烟测试
 media/launch-v3/                # 系统地图与发布素材
diff --git a/labs/attention-inbox/PRODUCT.md b/labs/attention-inbox/PRODUCT.md
new file mode 100644
index 0000000..4851279
--- /dev/null
+++ b/labs/attention-inbox/PRODUCT.md
@@ -0,0 +1,59 @@
+# Product wedge: One Inbox for Agent Attention
+
+## Positioning
+
+**Your agents work everywhere. You only need one inbox.**
+
+The wedge is not another multi-agent framework. It is the boundary between autonomous work and human attention.
+
+## User promise
+
+Agents continue routine work silently. The user is interrupted only when:
+
+1. a concrete human decision is required; or
+2. a meaningful risk needs awareness.
+
+Everything else is suppressed or left in history.
+
+## V0 scope
+
+- a portable event contract
+- deterministic escalation
+- evidence-aware rendering
+- duplicate suppression
+- adapters for existing agent/task systems
+
+Not V0:
+
+- another canonical task database
+- another scheduler
+- another orchestration engine
+
+## Public-core boundary
+
+The public layer can include:
+
+- protocol/schema
+- local reference logic
+- adapters
+- basic inbox policy
+- synthetic fixtures
+- conformance tests
+
+Future product work can focus on reliable delivery, cross-device history, team routing, and observability without changing the event contract.
+
+## Activation metric
+
+Do not optimize for GitHub stars first.
+
+A useful pilot user should satisfy all three in one week:
+
+- connects at least 2 agents/workers;
+- receives at least 5 meaningful inbox events;
+- completes at least 3 decisions from phone without opening the laptop.
+
+## Differentiation
+
+Coding-agent workspaces optimize how agents work together.
+
+Attention Inbox optimizes when a human must re-enter the loop.
diff --git a/labs/attention-inbox/README.md b/labs/attention-inbox/README.md
new file mode 100644
index 0000000..2be9a93
--- /dev/null
+++ b/labs/attention-inbox/README.md
@@ -0,0 +1,122 @@
+# Attention Inbox Lab
+
+A small, deterministic human-attention layer for AI agents.
+
+The problem is not that agents cannot keep working. The problem is that humans get interrupted by the wrong things: routine completions, internal `waiting_for_user` flags, stale status changes, or duplicate alerts. Meanwhile real decisions and real risks can be buried.
+
+Attention Inbox turns agent/task events into one of three outcomes:
+
+```text
+SUPPRESS
+RISK_ALERT
+ACTION_REQUIRED
+```
+
+It is deliberately **not** a task database, scheduler, queue, router, or new control plane. Canonical task state stays where it already lives. This lab is only a derived attention surface over existing task/receipt/event outputs.
+
+## 10-minute first success
+
+Requirements: Python 3 only.
+
+From `labs/attention-inbox`:
+
+```bash
+python3 scripts/inbox.py check fixtures/decision-required.json
+# exit 0
+# ATTENTION_EVENT_VALID
+
+python3 scripts/inbox.py render fixtures/internal-waiting.json
+# exit 0
+# SUPPRESS reason=no_concrete_user_decision
+
+python3 scripts/inbox.py render fixtures/decision-required.json
+# exit 0
+# ACTION_REQUIRED task=synthetic-release gate=publish_release
+
+python3 scripts/inbox.py render fixtures/risk-alert.json
+# exit 0
+# RISK_ALERT task=synthetic-worker-health severity=high
+
+python3 scripts/inbox.py render fixtures/completion.json
+# exit 0
+# SUPPRESS reason=routine_completion
+
+python3 scripts/inbox.py batch fixtures/events.jsonl
+# exit 0
+# output contains one duplicate suppression token
+
+python3 tests/run.py
+# exit 0
+# ATTENTION_INBOX_TEST_OK
+```
+
+## Core rule
+
+`waiting_for_user`, `blocked`, or similar internal statuses are not proof that a user decision exists.
+
+An event becomes `ACTION_REQUIRED` only when `decision.required=true` **and** the event contains a concrete gate and a concrete question/action. Otherwise it stays suppressed unless it is a meaningful risk alert.
+
+## Event shape
+
+See [`schema/inbox-event.schema.json`](schema/inbox-event.schema.json).
+
+The reference event has:
+
+- stable `event_id`
+- canonical `task_ref`
+- source/owner
+- event `kind`
+- severity
+- concise summary
+- evidence references
+- optional concrete `decision`
+- deterministic `dedupe_key`
+- source revision/freshness
+
+The schema intentionally does not store credentials, private payloads, or execution tokens.
+
+## Adapter contract
+
+Any agent, CLI, daemon, MCP adapter, or workflow may emit this event shape. The adapter owns translation; Attention Inbox owns only validation, escalation, rendering, and dedupe.
+
+Recommended flow:
+
+```text
+Agent / existing worker
+        ↓
+existing receipt / task writeback
+        ↓
+adapter → AttentionEvent
+        ↓
+Attention Inbox policy
+   ├─ SUPPRESS
+   ├─ RISK_ALERT
+   └─ ACTION_REQUIRED
+        ↓
+existing notification / phone surface
+        ↓
+human decision
+        ↓
+existing owner/writeback/resume path
+```
+
+The final resume/writeback remains part of the existing system. This lab must not invent another execution path.
+
+## Public boundary
+
+Safe to open-source:
+
+- schema
+- policy
+- synthetic fixtures
+- reference CLI
+- tests
+
+Keep private:
+
+- production routing
+- push/relay credentials
+- machine identities
+- real task data
+- account/wallet/trading execution
+- private memory
diff --git a/labs/attention-inbox/examples/adapter-output.jsonl b/labs/attention-inbox/examples/adapter-output.jsonl
new file mode 100644
index 0000000..026338b
--- /dev/null
+++ b/labs/attention-inbox/examples/adapter-output.jsonl
@@ -0,0 +1,2 @@
+{"schema_version":1,"event_id":"evt-example-1","task_ref":"task-123","source":"codex-adapter","owner":"products","kind":"blocked","severity":"low","summary":"Worker is paused internally; no user decision exists.","occurred_at":"2026-10-06T15:00:00Z","dedupe_key":"task-123:pause","source_revision":"rev-8","decision":{"required":false}}
+{"schema_version":1,"event_id":"evt-example-2","task_ref":"task-456","source":"claude-adapter","owner":"products","kind":"decision_required","severity":"medium","summary":"Validated artifact reached an external publish gate.","occurred_at":"2026-10-06T15:02:00Z","dedupe_key":"task-456:publish","source_revision":"rev-3","evidence":[{"type":"artifact","ref":"artifact://release-review"}],"decision":{"required":true,"gate_id":"public_publish","question":"Publish the validated artifact?","allowed_actions":["approve_publish","hold"]}}
diff --git a/labs/attention-inbox/fixtures/completion.json b/labs/attention-inbox/fixtures/completion.json
new file mode 100644
index 0000000..eb392f6
--- /dev/null
+++ b/labs/attention-inbox/fixtures/completion.json
@@ -0,0 +1,15 @@
+{
+  "schema_version": 1,
+  "event_id": "evt-done-001",
+  "task_ref": "synthetic-background-sync",
+  "source": "synthetic-background-worker",
+  "owner": "daily_rhythm",
+  "kind": "completion",
+  "severity": "low",
+  "summary": "Routine background sync completed successfully.",
+  "occurred_at": "2026-10-06T14:03:00Z",
+  "dedupe_key": "synthetic-background-sync:done",
+  "source_revision": 11,
+  "evidence": [],
+  "decision": {"required": false}
+}
diff --git a/labs/attention-inbox/fixtures/decision-required.json b/labs/attention-inbox/fixtures/decision-required.json
new file mode 100644
index 0000000..d065902
--- /dev/null
+++ b/labs/attention-inbox/fixtures/decision-required.json
@@ -0,0 +1,22 @@
+{
+  "schema_version": 1,
+  "event_id": "evt-decision-001",
+  "task_ref": "synthetic-release",
+  "source": "synthetic-codex-worker",
+  "owner": "product_distribution",
+  "kind": "decision_required",
+  "severity": "medium",
+  "summary": "Release artifact is validated and ready for a public publish gate.",
+  "occurred_at": "2026-10-06T14:00:00Z",
+  "dedupe_key": "synthetic-release:publish_release",
+  "source_revision": 4,
+  "evidence": [
+    {"type": "artifact", "ref": "./artifacts/release-review.md", "label": "validated review"}
+  ],
+  "decision": {
+    "required": true,
+    "gate_id": "publish_release",
+    "question": "Publish the validated release publicly?",
+    "allowed_actions": ["approve_publish", "hold"]
+  }
+}
diff --git a/labs/attention-inbox/fixtures/events.jsonl b/labs/attention-inbox/fixtures/events.jsonl
new file mode 100644
index 0000000..aed4fa2
--- /dev/null
+++ b/labs/attention-inbox/fixtures/events.jsonl
@@ -0,0 +1,3 @@
+{"schema_version":1,"event_id":"evt-batch-1","task_ref":"synthetic-release","source":"synthetic-worker","owner":"product_distribution","kind":"decision_required","severity":"medium","summary":"Validated release reached public gate.","occurred_at":"2026-10-06T14:10:00Z","dedupe_key":"synthetic-release:publish","source_revision":1,"decision":{"required":true,"gate_id":"publish_release","question":"Publish now?","allowed_actions":["approve_publish","hold"]}}
+{"schema_version":1,"event_id":"evt-batch-2","task_ref":"synthetic-release","source":"synthetic-worker","owner":"product_distribution","kind":"decision_required","severity":"medium","summary":"Duplicate transport delivery.","occurred_at":"2026-10-06T14:10:01Z","dedupe_key":"synthetic-release:publish","source_revision":1,"decision":{"required":true,"gate_id":"publish_release","question":"Publish now?","allowed_actions":["approve_publish","hold"]}}
+{"schema_version":1,"event_id":"evt-batch-3","task_ref":"synthetic-routine","source":"synthetic-worker","owner":"daily_rhythm","kind":"completion","severity":"low","summary":"Routine work completed.","occurred_at":"2026-10-06T14:11:00Z","dedupe_key":"synthetic-routine:done","source_revision":3,"decision":{"required":false}}
diff --git a/labs/attention-inbox/fixtures/internal-waiting.json b/labs/attention-inbox/fixtures/internal-waiting.json
new file mode 100644
index 0000000..eaf025e
--- /dev/null
+++ b/labs/attention-inbox/fixtures/internal-waiting.json
@@ -0,0 +1,15 @@
+{
+  "schema_version": 1,
+  "event_id": "evt-wait-001",
+  "task_ref": "synthetic-research",
+  "source": "synthetic-worker",
+  "owner": "research_dd",
+  "kind": "blocked",
+  "severity": "low",
+  "summary": "Worker emitted an internal waiting_for_user-like state but no concrete decision exists.",
+  "occurred_at": "2026-10-06T14:01:00Z",
+  "dedupe_key": "synthetic-research:internal-wait",
+  "source_revision": 2,
+  "evidence": [],
+  "decision": {"required": false}
+}
diff --git a/labs/attention-inbox/fixtures/risk-alert.json b/labs/attention-inbox/fixtures/risk-alert.json
new file mode 100644
index 0000000..47fafea
--- /dev/null
+++ b/labs/attention-inbox/fixtures/risk-alert.json
@@ -0,0 +1,17 @@
+{
+  "schema_version": 1,
+  "event_id": "evt-risk-001",
+  "task_ref": "synthetic-worker-health",
+  "source": "synthetic-host-sentinel",
+  "owner": "personal_ops",
+  "kind": "risk_alert",
+  "severity": "high",
+  "summary": "Execution host disk usage crossed the configured risk threshold.",
+  "occurred_at": "2026-10-06T14:02:00Z",
+  "dedupe_key": "synthetic-worker-health:disk-high",
+  "source_revision": 9,
+  "evidence": [
+    {"type": "receipt", "ref": "./artifacts/host-health.json", "label": "health receipt"}
+  ],
+  "decision": {"required": false}
+}
diff --git a/labs/attention-inbox/schema/inbox-event.schema.json b/labs/attention-inbox/schema/inbox-event.schema.json
new file mode 100644
index 0000000..8ae1d97
--- /dev/null
+++ b/labs/attention-inbox/schema/inbox-event.schema.json
@@ -0,0 +1,63 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://example.invalid/attention-inbox/inbox-event.schema.json",
+  "title": "AttentionEvent",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "schema_version",
+    "event_id",
+    "task_ref",
+    "source",
+    "owner",
+    "kind",
+    "severity",
+    "summary",
+    "occurred_at",
+    "dedupe_key",
+    "source_revision"
+  ],
+  "properties": {
+    "schema_version": {"const": 1},
+    "event_id": {"type": "string", "minLength": 1},
+    "task_ref": {"type": "string", "minLength": 1},
+    "source": {"type": "string", "minLength": 1},
+    "owner": {"type": "string", "minLength": 1},
+    "kind": {
+      "enum": ["decision_required", "blocked", "failure", "risk_alert", "completion", "info"]
+    },
+    "severity": {"enum": ["low", "medium", "high", "critical"]},
+    "summary": {"type": "string", "minLength": 1, "maxLength": 500},
+    "occurred_at": {"type": "string", "minLength": 1},
+    "dedupe_key": {"type": "string", "minLength": 1},
+    "source_revision": {"type": ["string", "integer"]},
+    "evidence": {
+      "type": "array",
+      "items": {
+        "type": "object",
+        "additionalProperties": false,
+        "required": ["type", "ref"],
+        "properties": {
+          "type": {"enum": ["artifact", "receipt", "log", "diff", "url", "other"]},
+          "ref": {"type": "string", "minLength": 1},
+          "label": {"type": "string"}
+        }
+      }
+    },
+    "decision": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["required"],
+      "properties": {
+        "required": {"type": "boolean"},
+        "gate_id": {"type": "string"},
+        "question": {"type": "string"},
+        "allowed_actions": {
+          "type": "array",
+          "items": {"type": "string", "minLength": 1},
+          "uniqueItems": true
+        }
+      }
+    }
+  }
+}
diff --git a/labs/attention-inbox/scripts/inbox.py b/labs/attention-inbox/scripts/inbox.py
new file mode 100644
index 0000000..de34394
--- /dev/null
+++ b/labs/attention-inbox/scripts/inbox.py
@@ -0,0 +1,146 @@
+#!/usr/bin/env python3
+import argparse
+import json
+import sys
+
+KINDS = {"decision_required", "blocked", "failure", "risk_alert", "completion", "info"}
+SEVERITIES = {"low", "medium", "high", "critical"}
+REQUIRED = {
+    "schema_version", "event_id", "task_ref", "source", "owner", "kind",
+    "severity", "summary", "occurred_at", "dedupe_key", "source_revision"
+}
+
+
+def load_json(path: str):
+    with open(path, "r", encoding="utf-8") as f:
+        return json.load(f)
+
+
+def validate_event(event):
+    errors = []
+    missing = sorted(REQUIRED - set(event))
+    if missing:
+        errors.append("missing=" + ",".join(missing))
+    if event.get("schema_version") != 1:
+        errors.append("schema_version")
+    if event.get("kind") not in KINDS:
+        errors.append("kind")
+    if event.get("severity") not in SEVERITIES:
+        errors.append("severity")
+    for key in ("event_id", "task_ref", "source", "owner", "summary", "occurred_at", "dedupe_key"):
+        if key in event and (not isinstance(event[key], str) or not event[key].strip()):
+            errors.append(key)
+    decision = event.get("decision")
+    if decision is not None:
+        if not isinstance(decision, dict) or "required" not in decision or not isinstance(decision.get("required"), bool):
+            errors.append("decision")
+    return errors
+
+
+def concrete_decision(event):
+    decision = event.get("decision") or {}
+    if decision.get("required") is not True:
+        return False
+    if not str(decision.get("gate_id", "")).strip():
+        return False
+    if not str(decision.get("question", "")).strip():
+        return False
+    actions = decision.get("allowed_actions")
+    return isinstance(actions, list) and len(actions) > 0 and all(isinstance(a, str) and a.strip() for a in actions)
+
+
+def classify(event):
+    if concrete_decision(event):
+        return "ACTION_REQUIRED", "concrete_decision"
+
+    if event.get("kind") == "risk_alert" and event.get("severity") in {"high", "critical"}:
+        return "RISK_ALERT", "meaningful_risk"
+
+    if event.get("kind") == "failure" and event.get("severity") in {"high", "critical"}:
+        return "RISK_ALERT", "meaningful_failure"
+
+    if event.get("kind") == "completion":
+        return "SUPPRESS", "routine_completion"
+
+    if event.get("kind") in {"blocked", "decision_required"}:
+        return "SUPPRESS", "no_concrete_user_decision"
+
+    return "SUPPRESS", "no_attention_needed"
+
+
+def render(event):
+    outcome, reason = classify(event)
+    task = event["task_ref"]
+    if outcome == "ACTION_REQUIRED":
+        gate = event["decision"]["gate_id"]
+        return f"ACTION_REQUIRED task={task} gate={gate}"
+    if outcome == "RISK_ALERT":
+        return f"RISK_ALERT task={task} severity={event['severity']}"
+    return f"SUPPRESS reason={reason}"
+
+
+def cmd_check(path):
+    event = load_json(path)
+    errors = validate_event(event)
+    if errors:
+        for err in errors:
+            print(f"ATTENTION_EVENT_INVALID {err}")
+        return 1
+    print("ATTENTION_EVENT_VALID")
+    return 0
+
+
+def cmd_render(path):
+    event = load_json(path)
+    errors = validate_event(event)
+    if errors:
+        for err in errors:
+            print(f"ATTENTION_EVENT_INVALID {err}")
+        return 1
+    print(render(event))
+    return 0
+
+
+def cmd_batch(path):
+    seen = set()
+    invalid = 0
+    with open(path, "r", encoding="utf-8") as f:
+        for lineno, line in enumerate(f, 1):
+            if not line.strip():
+                continue
+            try:
+                event = json.loads(line)
+            except json.JSONDecodeError:
+                print(f"ATTENTION_EVENT_INVALID line={lineno} json")
+                invalid += 1
+                continue
+            errors = validate_event(event)
+            if errors:
+                print(f"ATTENTION_EVENT_INVALID line={lineno} {';'.join(errors)}")
+                invalid += 1
+                continue
+            dedupe_identity = (event["dedupe_key"], str(event["source_revision"]))
+            if dedupe_identity in seen:
+                print(f"SUPPRESS reason=duplicate event={event['event_id']}")
+                continue
+            seen.add(dedupe_identity)
+            print(render(event))
+    return 1 if invalid else 0
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Reference Attention Inbox policy CLI")
+    sub = parser.add_subparsers(dest="command", required=True)
+    for name in ("check", "render", "batch"):
+        p = sub.add_parser(name)
+        p.add_argument("path")
+    args = parser.parse_args()
+    if args.command == "check":
+        return cmd_check(args.path)
+    if args.command == "render":
+        return cmd_render(args.path)
+    return cmd_batch(args.path)
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/labs/attention-inbox/tests/run.py b/labs/attention-inbox/tests/run.py
new file mode 100644
index 0000000..5ce0871
--- /dev/null
+++ b/labs/attention-inbox/tests/run.py
@@ -0,0 +1,43 @@
+#!/usr/bin/env python3
+import importlib.util
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+SCRIPT = ROOT / "scripts" / "inbox.py"
+spec = importlib.util.spec_from_file_location("attention_inbox", SCRIPT)
+mod = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(mod)
+
+
+def load(name):
+    return mod.load_json(str(ROOT / "fixtures" / name))
+
+
+def expect(condition, message):
+    if not condition:
+        raise AssertionError(message)
+
+
+def main():
+    decision = load("decision-required.json")
+    waiting = load("internal-waiting.json")
+    risk = load("risk-alert.json")
+    completion = load("completion.json")
+
+    for event in (decision, waiting, risk, completion):
+        expect(mod.validate_event(event) == [], f"invalid fixture: {event['event_id']}")
+
+    expect(mod.render(decision) == "ACTION_REQUIRED task=synthetic-release gate=publish_release", "decision should escalate")
+    expect(mod.render(waiting) == "SUPPRESS reason=no_concrete_user_decision", "internal waiting must not masquerade as approval")
+    expect(mod.render(risk) == "RISK_ALERT task=synthetic-worker-health severity=high", "high risk should surface")
+    expect(mod.render(completion) == "SUPPRESS reason=routine_completion", "routine completion should stay quiet")
+
+    fake = dict(decision)
+    fake["decision"] = {"required": True, "gate_id": "publish_release", "question": "Publish?", "allowed_actions": []}
+    expect(mod.render(fake) == "SUPPRESS reason=no_concrete_user_decision", "decision without actions is not concrete")
+
+    print("ATTENTION_INBOX_TEST_OK")
+
+
+if __name__ == "__main__":
+    main()