Skip to content

[PR] fix: separate parallel CC subagent runs into their own lanes via msgid store matching (#6011) - #6067

Open
ranxianglei wants to merge 4 commits into
masterfrom
2026-11-10_cc-subagent-sidechain-msgid-lanes
Open

ranxianglei wants to merge 4 commits into
masterfrom
2026-11-10_cc-subagent-sidechain-msgid-lanes

Conversation

@ranxianglei

@ranxianglei ranxianglei commented Nov 9, 2026 •

Copy link
Copy Markdown
Owner

Closes #6011

Problem

Claude Code standalone (proxy mode) runs many parallel Task subagents under one conversation id. All of them share byte-identical first user cells (Task boilerplate), so the D1-B discriminator (first user message fragment) is identical across runs — every subagent request lands on ONE persona lane <rawId>|sub:<fp16(tag)>:<fp16(disc)>. Concurrent runs merge into a single session: cache-prefix thrashing (in the reporter's log the sub lane is ~31% of input volume and ~74% of non-cached cost), birth-time preflight stalls, cross-run contamination.

How (three parts)

1. Acquisition — the Claude adapter projects sidechain threads out of the transcript JSONL: isSidechain: true rows are grouped into chains by parentUuid back-walk from tail rows, and a sidechain request binds to exactly one chain only when its full-tail content fingerprints match that chain's tail (and the main-thread tail does not). Ambiguity or mismatch ⇒ whole turn degrades (no ids), never guesses. Gates unchanged: rollout identityChannelHarvest (currently 0%) or env BILI_IDENTITY_HARVEST, loopback only.

2. Pin-gate extension — verified-HARVEST ids now pin onto canonical rows regardless of pluginMode / pluginLaneServerHistory cohort. The harvest registry already enforced rollout-or-env × loopback × full content verification upstream, so harvested ids are as trustworthy as stamped ones; without this, CC-standalone (proxy mode) lanes never accumulated pinned rows and store-overlap rungs could never fire for them. Stamped (body/header) traffic keeps the old double gate unchanged.

3. Consumption = standard sessionId+msgid store matching — a new sub-lane rung (trySiblingLane) fires at all three fork exits AFTER the raw-key rescue and BEFORE the legacy mint. Candidates = skeleton-table <rawId>|sub:* keys ∪ unified-store session list prefix-filtered the same way; per candidate it reads the PINNED rows (live AND retired — retire-don't-delete, so client compaction cannot erase matchability) with the existing indexed session_id + msg_id IN (...) query. Exactly one candidate overlapping ≥1 incoming id ⇒ continue on that lane; several overlap ⇒ warn + degrade to fresh mint (never guess); none ⇒ mint <rawId>|sub:m:<first inbound id> — the FULL untruncated id (host-minted uuids are unique by construction; no fp16 truncation). Requests without ids take the legacy key shapes byte-identically. No parallel ledger: the skeleton table supplies candidate KEYS only; the id truth stays in the unified store (the documented "no parallel JSON ledger is kept" principle in handle.ts).

Behavior changes (old → new)

  1. CC subagent lanes separate when harvest is active (the fix). Old: all parallel Task runs shared one <rawId>|sub:<fp16(tag)>:<fp16(disc)> lane → merged histories, cache thrash. New: each run gets its own <rawId>|sub:m:<first host id> lane from round 1 and continues on it via store-overlap. Gated behind the harvest opt-in (off everywhere today).
  2. Fork key shape for stamped-id traffic changes immediately (ungated). Old: persona/task forks minted by <rawId>|sub:<fp16(tag)>:<fp16(disc)> or <rawId>|sub:task:<skel8> even when msgids were present (ids fed only the raw rescue). New: <rawId>|sub:m:<first inbound id>. Previously-colliding strangers now separate; previously-stable lanes get new keys on their next fork. Restart-stable (same first id ⇒ same key). D1-C token lanes unaffected (the token rung precedes).
  3. Claude adapter is stricter whenever the session file contains sidechain rows. Old: role-only tail alignment could bind a sidechain request onto MAIN-thread row ids (the misbinding path). New: content fingerprints must verify the exact thread; a transient wire/transcript text drift costs one turn's pinning instead of risking a misbind.
  4. Verified-harvest requests now pin msg_ids onto canonical rows (additive column; see the §5 note in docs/msgid-infrastructure.md). Plugin lanes: pin identity to client msgids, render outbound from the server-owned history | 插件车道服务端正史:消息身份钉在客户端 id 上,出站前缀由正史渲染 #4491 observation accounting becomes non-empty best-effort logs for these lanes; the raw-key rescue becomes effective for CC main sessions too.

Config surface

None. Rides the existing rollout features.identityChannelHarvest (currently 0%) + existing env BILI_IDENTITY_HARVEST + loopback-only rule. Activation is an owner action; until then nothing changes except behavior change 2 above (stamped plugin lanes).

Tests

  • NEW tests/issue6011-msgid-fork-lanes.test.ts (T1–T10): three parallel runs → three distinct full-id mints; no-ids control pins the legacy collision; round-2 continuation via store-overlap; compaction erasing the first id still continues (retired rows stay matchable); total amnesia → fresh coherent mint; multi-lane ambiguity → warn + degrade-mint; raw-rescue precedence over sibling scan; restart survives (continuation is purely store-backed); malformed ids → legacy shapes; aux blobs unrecorded; store-list-only discovery; unavailable store degrades without crashing.
  • EXTENDED tests/identity-harvest.test.ts: interleaved sidechain threads bind per fingerprint set; ambiguous identical-text threads → null; role-match/fingerprint-mismatch → null (misbinding path closed); main-request regression intact; unstructured rows (no parentUuid) → main verifies, subagent-shaped admits nothing; two-round continuity; rewrite/orphan detection via the overlap window.
  • Adjacent suites green locally on this head: issue5719, issue4352, issue5095, issue3397, issue5686, issue2944, msgid-rule-parity, msgid-pin (93/93 targeted). Full suite = remote CI on this head.

kernel-change: N/A (host-side only; kernel/ untouched)

Identity-logic change — human review required (§7.4), no auto-merge.

@github-actions

github-actions Bot commented Nov 9, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-11-10_cc-subagent-sidechain-msgid-lanes (2acea50)

Latest build — install this exact version

npm install -g billion-context@0.0.8898-pr.6067.8f88acc

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Recent builds (commit → version)

commit version build
2acea50 billion-context@0.0.8898-pr.6067.8f88acc build
cb2c9ce billion-context@0.0.8875-pr.6067.bb13c66 build
9943e44 billion-context@0.0.8867-pr.6067.5fe3772 build
11c4c18 billion-context@0.0.8865-pr.6067.5dff430 build
ff1de9d billion-context@0.0.8858-pr.6067.36445f5 build
e607000 billion-context@0.0.8856-pr.6067.770497b build

Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr6067.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei ranxianglei changed the title fix: separate parallel CC subagent runs into their own lanes via msgid anchors (#6011) [PR] fix: separate parallel CC subagent runs into their own lanes via msgid anchors (#6011) Nov 9, 2026
ranxianglei pushed a commit that referenced this pull request Nov 9, 2026
…s gate (#6067)

ClaudeSidechainRow (identity-claude.ts) and fingerprintMessage (identity-harvest.ts) are used only inside their own files; the CI gate requires dropping the leading export for INTERNAL_ONLY symbols.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

Duplicate screening: not a duplicate — a complementary follow-up. GitHub #6067 (this PR, closes #6011) and GitHub #5726 (external contributor feriaref, closes #5724) attack the same symptom family (parallel CC Task-subagent runs colliding onto one compression lane → sub-lane cache-hit collapse), but through different mechanisms: #5726 widens the content discriminator (approach "A", works today, no client dependency, separates from the first divergent cell); #6067 keys forked lanes off host-minted msgids + adds a Claude sidechain projection (approach "C-wire + C-jsonl", separates from round 1 even on byte-identical content, but needs msgid carriers ⇒ harvest on). They are complementary and stack per the KDD #11 ladder (msgid > order > hash): the id rung fires first when ids are present, the content widening stays as the fallback when ids are absent. #5726's own review thread concluded "merge A first, C-wire as a follow-up" — #6067 is exactly that follow-up. Both should land; neither supersedes the other.

Labels: bug · claude-code · scope/wiring · p1 (parallel CC subagents collide onto one lane → sub-lane hit-rate collapse ~68.7% vs main ~95.0%; cost explosion plus concurrent-run state trampling ⇒ top of the p1/p2 band, consistent with the #5726 family rating).

Independent re-review (read all 848+/67− diffs; did not copy existing thread conclusions)

Direction — correct, root-cause (疏 not 堵). The lane-key derivation never consumed msgids: fork-exit keys were content-derived (<rawId>|sub:<fp16(tag)>[:<fp16(disc)>]), and tryMsgIdRescue only had a merge direction (overlap ⇒ re-anchor; zero overlap ⇒ no evidence). #6067 adds the missing rung: id-based fork keys + a sibling-continuation rung + a strict sidechain attribution that verifies by content fingerprint or returns null (never falls through to a role-only misbind). Fix sits at the decision point, not a payload filter.

Regression sweep — clean (verified against live master 444630170):

  • No-ids path is byte-identical: trySiblingLane returns null immediately when ids are absent; the mint falls back to the legacy key; noteLaneAnchors is not invoked. Malformed ids ([""], [id,42], []) pin the legacy key (test T6).
  • All downstream |sub: consumers are suffix-agnostic — they use .includes("|sub:") or .split("|sub:")[0] only, never the suffix format. New |sub:m:<fp> shape is fully compatible: src/session.ts:1596,1605; src/plugin.ts:1410,1476,1495-1499. splitSessionWarnings likewise unaffected.
  • Role-check extraction is behavior-preserving: rolesAlign is the exact same loop pulled out of alignTailIds (guard messageRoles.length === tail.length ⟺ === expectedLen). The old header's "4-position spot-check" was a stale comment; the code always did a full-sequence check — the PR corrects the comment, changes no behavior (pinned by the pre-existing "role flip outside the old window degrades" test).
  • Raw-rescue precedence preserved at all three fork exits (tag-changed / diverged-order / shorter-divergent): tryMsgIdRescue runs before the sibling rung (test T4). D1-C token lanes unaffected (token rung precedes the fork exits).
  • Persistence: anchors is an additive optional field; malformed values degrade to undefined on hydrate (no version bump/migration); snapshot round-trips anchors (test T5).
  • claudeVisibleRows return-type change is safe: no production consumer outside its own file (only the test imports it).

Test teeth — both halves covered, positive + negative: resolver suite T1 (fails pre-fix: 3 runs collide) / T1-control (pins the exact bug without ids) / T2 (no cross-bind) / T3 (bounded split on compaction) / T4 (raw-rescue precedence) / T5 (restart) / T6 (malformed degrade) / T7 (aux never anchors); sidechain suite covers bind-by-fingerprint, wrong-content→null, identical-text-ambiguous→whole-turn-null, unstructured-degrade-while-main-intact, and per-thread continuity + store-rewrite detection. Meets gate discipline #3617 (both directions) and §7.4 criterion 2.

CI (live, head ff1de9dda9): 0 failures so far — 11 test shards + 2 static-gate legs still in progress; remaining legs green or correctly skipped for this diff. The initial unused-exports red (on e6070005) is fixed in ff1de9dd (un-exported the two internal symbols; verified locally: gate exit 0).

Findings (non-blocking)

  1. Eager hashing on the hot path (p3, disclose-or-tighten). src/server/handle.ts:1204 computes fingerprintMessageTexts(openaiLike) (sha256 per message) before calling harvestIdentity, while the actual gate (featureEnabled("identityChannelHarvest") + loopback) lives inside harvestIdentity (src/identity-harvest.ts:245,249). So with harvesting off (everywhere today), the hashing is wasted on every qualifying request (anthropic/openai + client-provided parent identity + no stamped ids). The PR description's "gated / zero cost until opted in" is inaccurate at this point. Recommend computing it lazily (inside the gate) or noting the hot-path cost in the description. Not a blocker: dormant feature, modest magnitude, no availability impact.
  2. Merge-coordination heads-up (not a defect): [PR] feat(session): split Claude Code SDK subagents into isolated session lanes #5726 and [PR] fix: separate parallel CC subagent runs into their own lanes via msgid store matching (#6011) #6067 both modify src/session-resolve.ts / resolveSessionOrder (hot file). Whichever merges second needs a rebase (§7.5 rework driver fix: keep system prompt byte-stable for prefix cache #1). Semantically they layer cleanly (id rung before content rung), so it's ordering, not a conflict.
  3. Disclosure precision (nit): item fix: keep system prompt byte-stable for prefix cache #1's blanket "zero behavior change until opted in" is slightly too strong — the stamped-id mint-key-shape change activates for native-stamped (non-token) lanes independent of the harvest gate. It is disclosed in item fix: address all code review findings (condense, routing, session-id, dead code) #2, and it's safe (internal key, deterministic+unique-per-run, fresh-fork-only, token lanes untouched) — just worth the distinction.

结构化八问

  1. 是什么问题:同一会话下并行 ≥2 个 CC Task subagent 共享字节相同的样板前导,所有字节档与 D1-B 内容判别撞键 → 全部并进一条 persona 车道,每次新运行从零重新 preflight 合并历史(Claude Code sub-agents are getting a lot of cache misses, but the main session is fine. | Claudecode 子代理大量 miss 缓存,主会话没事 #6011 报告的缓存风暴;子车道命中 ~68.7% vs 主 ~95.0%)。
  2. 什么原因导致的:车道键派生从不消费 msgid——fork 出口键为内容派生,并行同类运行的判别片段字节相同 → fp16 相同 → 同键;tryMsgIdRescue 只有合并方向(重叠⇒回锚、零重叠⇒不作证据),缺「id 不相交⇒拆车道」的档。根因 = KDD feat: self-update (auto + manual) #11 身份阶梯(msgid>顺序>hash)中 msgid 从未接入车道创建。
  3. 影响范围与严重程度:未装 hook、同会话并发 ≥2 个 CC Task subagent 的用户;子车道缓存命中崩塌 → token 成本爆炸 + 并发运行间压缩状态互相踩踏(正确性相邻)。严重度 p1(成本崩塌+并发互扰,非数据丢失/不可用;处 p1/p2 边界取上沿)。修复默认灰度关,当前线上零实际波及。
  4. 怎么修复的:fork 车道键锚到宿主铸造 msgid(|sub:m:<fp16(首个入站 id)>,每运行唯一、跨轮稳定)+ sibling 续接档(入站 id 与该车道锚定见证集重叠⇒继续该车道)+ claude 适配器从本机 JSONL sidechain 行做内容指纹对齐归属(恰一候选且主尾不匹配⇒绑该线程;歧义/无证据⇒整轮 null,绝不落 role-only 误绑)。打在判定点,非症状层过滤。
  5. 是否有回归问题:逐项过——无 id 路径逐字节保持;下游 |sub: 消费者全以后缀无关方式解析(includes/split)→ 新 m: 键兼容;role-check 抽取行为保持;三 fork 出口 raw-rescue 优先序保留;D1-C token 车道不受(token 档在前);splitSessionWarnings 后缀无关;order-persist anchors 可选字段畸形即降级、无版本迁移。无 fail-fast 变吞错、无阈值/时序/输出格式漂移。
  6. 是否破坏缓存:模型侧前缀缓存不受(出站字节不变,仅改 bili 内部车道簿记);目标即恢复命中。唯一一次性代价=受影响存量 stamped/headerless 车道升级后重键一轮冷启动(披露fix: address all code review findings (condense, routing, session-id, dead code) #2),无数据丢失。
  7. 是否改了设计方向:无转向——沿 KDD feat: self-update (auto + manual) #11/fix: suppress duplicate auto-update notices within one process #17/release v0.1.6 #18 推进「msgid 参与车道创建」属方向强化;sidechain 投影沿用 Investigate and integrate a stable "session + message ID" identity channel on a per-client basis. Current status: only f | 逐客户端调查并接入稳定"会话+消息ID"身份信道:现状仅 dsh/omp/pi/opencode 四车道在跑,其余(codebuddy/hermes/claude/codex/zcode…)按清单逐个研究 #5825 采集缝纪律(自包含适配器+整轮退化)。
  8. 是否引入新配置:无新 env/config;复用既有 harvest 门(BILI_IDENTITY_HARVEST / rollout identityChannelHarvest,默认关 × 仅回环)。

收工判定:方向正确、diff 干净(全为本 PR 目的)、回归面受控且已披露、测试两半均有牙、CI 进行中 0 失败。触碰身份/会话绑定逻辑 → §7.4 人工审批范畴,不适用自动合并。CI 转绿后可合,由你执行;Finding 1 建议顺手惰性化或补披露,不强求本轮。

@ranxianglei
ranxianglei force-pushed the 2026-11-10_cc-subagent-sidechain-msgid-lanes branch from ff1de9d to 11c4c18 Compare November 10, 2026 01:46
ranxianglei pushed a commit that referenced this pull request Nov 10, 2026
…s gate (#6067)

ClaudeSidechainRow (identity-claude.ts) and fingerprintMessage (identity-harvest.ts) are used only inside their own files; the CI gate requires dropping the leading export for INTERNAL_ONLY symbols.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

Conflict resolved and rebased onto current master — merge state went dirty → blocked (conflict gone; "blocked" is just the branch-protection gate waiting on required checks). No design decision was needed; this was a purely mechanical rebase + conflict resolution.

七问 summary

  1. 是什么问题:PR [PR] fix: separate parallel CC subagent runs into their own lanes via msgid store matching (#6011) #6067 的分支落后 master 67 个提交,与 base 产生合并冲突(merge state = dirty),无法合并。工作期间 master 推进很快(前进过三次)。

  2. 什么原因导致的:master 侧合入了 Existing JSON State Files: Inventory and First-Batch Unified Storage Migration | 存量 JSON 状态文件:盘点与第一批统一存储迁移 #5753(session-order skeleton 迁到统一 kv_store),把 readDiskEntries() 的内联校验重构为共享 helper orderEntryOf(e);本 PR 在同一位置为新 anchors 字段追加了 .map()。双方改了同一段代码 → 内容冲突。

  3. 影响范围与严重程度:仅影响该 PR 自身的可合并性,不影响任何已发布行为。修复前 PR 不可合并(dirty);修复后可合并(blocked,等 CI)。

  4. 怎么修复的:rebase 到最新 master(base 541cf260c4)。唯一真实冲突在 src/order-persist.ts 的 readDiskEntries(),按「意图并集」解决:保留 master 的 DRY 重构(用 orderEntryOf(e) helper 做 filter 校验——其校验逻辑与 PR 原内联谓词逐条等价),同时保留 PR 新增的 anchors .map()。这是由冲突驱动的一处机械简化(helper 替代等价内联谓词),非行为变更。src/server/handle.ts 两侧改动重叠但落在不同区域,自动合并干净。

  5. 是否有回归问题:无。本地验证:typecheck(tsc --noEmit)clean、build 成功、PR 定向测试 29/29 全绿(tests/issue6011-msgid-fork-lanes.test.ts + tests/identity-harvest.test.ts)、全仓无残留冲突标记。rebase 不改变 PR 自身贡献的内容,diff 与 rebase 前一致(仅 base 前移)。

  6. 是否破坏缓存:不涉及——未触碰任何缓存 / 重试 / 增量游标 / 去重键逻辑。

  7. 当前状态与判定:已 force-push(force-with-lease)到 origin/2026-11-10_cc-subagent-sidechain-msgid-lanes,新 head 11c4c18602(push 前已 fetch 确认无并发推送者)。CI 在新 SHA 上重跑:29/30 已结算(16 success + 13 skipped)、0 failed;剩余 1 个 Windows 测试分片(test windows-latest 22 #3)in_progress 约 40 分钟(慢 runner,非红),会自行结算。结论:冲突解决完成、无红 CI 需修;若最后那个分片稍后转红,由后续触发单独处理。

@ranxianglei
ranxianglei force-pushed the 2026-11-10_cc-subagent-sidechain-msgid-lanes branch from 11c4c18 to 9943e44 Compare November 10, 2026 02:37
@ranxianglei ranxianglei changed the title [PR] fix: separate parallel CC subagent runs into their own lanes via msgid anchors (#6011) [PR] fix: separate parallel CC subagent runs into their own lanes via msgid store matching (#6011) Nov 10, 2026
@ranxianglei

Copy link
Copy Markdown
Owner Author

(ai: vllm-qwen/qwen3.8-27b)

Verdict: REQUEST_CHANGES — the head implements the design the owner rejected in-thread (2026-11-10T01:17Z); independent verification confirms all three rejection points against the code. The agreed redo (same branch, in place) is the right path; this review pins down what the new head must satisfy.

Independent verification of the owner's three points

  1. Parallel ledger instead of the standard store-matching chain — confirmed. Recorded.anchors (src/session-resolve.ts) plus the persisted anchors field (src/order-persist.ts) form a second id ledger; trySiblingLane reads only that table, and the diff/docs state it deliberately ("SKELETON anchor table ONLY — never reads the unified store"). This contradicts the [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolve… #5719 principle documented at src/server/handle.ts:1229-1236 ("the RECORDED side … lives in the UNIFIED STORE … no parallel JSON ledger is kept"), the T5 ruling in docs/msgid-infrastructure.md §5 ("the ONLY id truth … no parallel ledger") — which this PR's own consumer-⑥ row overrides in the same file — and KDD fix: suppress duplicate auto-update notices within one process #17 (unified storage as the single id fact source).
  2. fp16 truncation in the mint key — confirmed. The mint sites use <rawId>|sub:m:${fp16(inboundIds[0])}: a 64-bit truncation of host UUIDs that are globally unique by contract. An unnecessary collision class; full-id keys cost nothing.
  3. Compaction fragility — confirmed. Anchors keep only the 16 most-recent ids and harvested inbound ids are the sliding tail window. Compaction erasing every anchored id ⇒ zero overlap ⇒ fresh mint ⇒ the run splits mid-execution and re-collides/re-cold-starts — intermittently reproducing the exact storm Claude Code sub-agents are getting a lot of cache misses, but the main session is fine. | Claudecode 子代理大量 miss 缓存,主会话没事 #6011 fixes. Store-based matching keeps retired ids matchable (tombstones, Four storage-layer defects: identity hash contaminated with positional metadata (truncation triggers identity swap) / sw | 存储层四缺陷:身份哈希掺位置元数据(截断即换血)/清扫删除回退历史/渲染接缝图片只比长度/fork 谱系死列 #4390) until physical archive, so the split disappears structurally.

Additional findings (required in the redo)

  1. Eager fingerprint hashing on the hot path. src/server/handle.ts computes fingerprintMessageTexts(openaiLike) (sha256 per message over full text) before harvestIdentity, whose gates live inside it (featureEnabled("identityChannelHarvest") src/identity-harvest.ts:189, loopback :193). Every openai/anthropic request with client-provided parent identity and non-empty messages therefore pays sha256 over its whole history TODAY — i.e. every Claude Code request in production, with the feature off everywhere. The PR body's "zero behavior change until opted in" is inaccurate about CPU cost. Move the computation behind the gate (cheap pre-check or lazy provider).
  2. Write-only anchors on the raw key. The public wrapper records anchors for every resolution including raw-key main-session ones; trySiblingLane scans only |sub: keys, so raw-lane anchors are never read, yet a persist is scheduled every round (even when the set is unchanged). Vanishes with the anchor revert.
  3. Docs nit. docs/session-identity-resolution.md: list item "3. Availability" is re-indented under item 2 (markdown nesting artifact). Superseded by the docs rewrite.

No fix commits were pushed: the branch is mid-redo by the author (in-place rewrite announced in-thread ~1h ago) and pushing now would fork the active working tree. Findings 1–3, 5, 6 are resolved by the planned redo itself; finding 4 belongs in the seam files the redo rewrites. All are carried here as conditions and will be verified line-by-line when the new head lands.

What is sound (keep in the redo)

  • Data layer: JSONL sidechain projection, parentUuid chain grouping, strict full-tail content verification, never-guess ambiguity/degrade matrix — a correct implementation of the owner's C-jsonl ruling. The adapter tests have teeth (content drift→null, identical-text ambiguity→null, unstructured rows degrade subagent-shaped requests while keeping the main pass, per-thread continuity catches store rewrites).
  • Regression hygiene of the current diff: no-ids paths byte-identical (T1c/T6), raw-lane rescue precedence preserved (T4), downstream |sub: consumers suffix-agnostic, aux lanes excluded from anchoring (T7), malformed ids degrade.
  • CI: green on this head (all completed shards; one windows test shard still running).

Conditions for the new head

  1. Consumption rebuilt as standard sessionId+msgid unified-store matching per the in-thread plan (per-candidate-key storedMsgIdsOf overlap: unique→continue / multi-hit→warn+demote-to-mint, never guess / zero→mint); ALL skeleton-anchor machinery reverted (Recorded.anchors, the order-persist field, the wrapper chokepoint).
  2. Mint key uses the FULL first inbound id (no fp16).
  3. Fingerprint computation moved behind the harvest gate (finding 4).
  4. Pin-gate extension decided — the open 🔶 in-thread (A: pin only |sub: derived lanes vs B: all validated harvest requests incl. the raw main key; author recommends B). The store-based sibling rung depends on rows actually being pinned for harvested lanes.
  5. Tests/docs rewritten to pin the new mechanism — note current T3 pins the rejected fragility as intended behavior ("compaction erasing every anchored id → fresh coherent lane"); it must flip to post-compaction continuation via retired ids.

Hard rules: no kernel/ changes (verified — the body's kernel-change: N/A holds; zero files under kernel/). Merge stays human-gated per §7.4 (identity/session-binding logic) regardless of this review.


中文八问

  1. 是什么问题:PR 声称修复 CC 并行 Task 子代理挤在同一条 persona 车道导致的缓存命中风暴(Claude Code sub-agents are getting a lot of cache misses, but the main session is fine. | Claudecode 子代理大量 miss 缓存,主会话没事 #6011),但实现方案被 owner 判定需要重做。
  2. 什么原因导致的:修复为子车道续接建了「骨架锚点平行账本」(Recorded.anchors + order-persist 持久化),违反 [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolve… #5719/T5「统一存储是唯一 id 事实源、不保留平行账本」的原则(src/server/handle.ts:1229-1236、docs/msgid-infrastructure.md §5 —— PR 自己在同一文档新增的消费者⑥行里直接推翻了该裁决);mint 键用 fp16 截断引入不必要的冲突面;16 锚点滑动窗口在压缩后丢失重叠、回退到分裂。三点均已对代码独立核实成立。
  3. 影响范围与严重程度:尚未合并(blocked awaiting review)。若按现状合入:双 id 账本破坏身份架构的单一事实源不变量;fp16 冲突可致两个 run 误绑同一车道(corruption 级);压缩后 run 再次分裂→原风暴间歇复发;eager sha256 给全部生产环境 CC 请求(功能关闭状态下)增加常态 CPU 成本。架构级回归 + 性能漂移,不可合。
  4. 怎么修复的(已同意的 redo 计划):保留数据层(JSONL sidechain 投影 + 内容指纹验证,测试保留);消费侧重建为标准 sessionId+msgid 统一存储匹配(逐候选键 storedMsgIdsOf 重叠:唯一命中→continue / 多命中→告警并降级 mint、绝不猜 / 零命中→mint);mint 用完整 UUID;删除全部锚点机制;pin 门禁扩展到已验证 harvest 来源(A/B 范围待 owner 拍板,作者推荐 B);fingerprint 计算移到门禁之后惰性执行。
  5. 是否有回归问题:对当前 diff 触碰路径的既有行为清点——无 id 路径字节级不变(T1c/T6 钉住)、raw-lane rescue 优先级保持(T4)、下游 |sub: 消费者后缀无关、aux 车道不参与锚点(T7)、D1-C token 车道不受影响;因消费层将整体重写,以上结论在新 head 上需重新核验。
  6. 是否破坏缓存:改动的目的即修复缓存命中;当前设计的压缩分裂路径恰恰会再次打断命中(退回冷 preflight),store 方案(retired id 在物理归档前持续可匹配)从结构上消除该分裂;既有缓存失效策略本身不受影响。
  7. 是否改了设计方向:原 PR 引入了平行账本方向(与既定方向相悖);redo 回到统一存储标准匹配链,与 [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolve… #5719/[PR] release review #5826 方向一致。
  8. 是否引入了新的配置:无(沿用 BILI_IDENTITY_HARVEST env / rollout features.identityChannelHarvest,默认关 × 仅回环)。

判定:REQUEST_CHANGES —— 当前 head 不可合;等原地重做的新 head 落地后按上述条件逐项复审。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete → REQUEST_CHANGES (head 11c4c18): the shipped design is the owner-rejected one — parallel skeleton-anchor ledger instead of the standard sessionId+msgid unified-store matching chain, fp16-truncated mint key, compaction-fragile 16-anchor window; all three rejection points independently verified in code, plus one extra finding (eager per-message sha256 ahead of the harvest gate costs every production CC request CPU while the feature is off). No fix commits pushed — branch is mid-redo by the author; full review with five redo conditions posted on the PR: #6067 (comment) . Awaiting the in-place redo head for re-review. No kernel/ changes; merge stays human-gated per §7.4.

ci-check added 2 commits November 10, 2026 11:25
… review)

The handle.ts caller computed fingerprintMessageTexts(openaiLike) eagerly on every client-provided-id request before calling harvestIdentity, so all CC traffic paid sha256 over full message texts even with harvesting off everywhere (rollout 0%, no env). IdentityHarvestCtx.messageFingerprints becomes computeMessageFingerprints, a lazy provider invoked only after every registry gate passed AND the claude adapter matched a file containing sidechain rows; files without sidechain rows never hash at all. Pure lazification, zero semantic change; adds the negative-direction gate test (provider must not run while gates are off, #3617 both-directions discipline).
…review)

The redo indented item 3 'Availability.' by three spaces, so GFM parses it as a nested list under item 2 'Data locality' — a rendering regression of the very list this section documents. Restore column-0 sibling indent as on master.
@ranxianglei

Copy link
Copy Markdown
Owner Author

(ai: vllm-qwen/qwen3.8-27b)

Re-review at head cb2c9ce (after the in-place redo + two review-side fixes) — verdict: APPROVE. This supersedes my earlier review comment on the pre-redo head.

Merge stays human-gated per AGENTS.md §7.4 — identity/session-binding logic is never auto-merge territory. Note: this account IS the PR author, so GitHub rejects a formal review verdict here (self-review forbidden); this comment carries the full review in its place.

All five v1 findings verified against the new head (9943e44)

  1. Store-matching consumption; parallel ledger fully reverted ✓ — trySiblingLane (src/session-resolve.ts) fires at all three fork exits, after raw-key rescue and before legacy mint. Candidates = skeleton <rawId>|sub:* keys ∪ same-prefix ids from the unified store; per-candidate overlap via indexed session_id+msg_id query over live AND retired rows (retire-don't-delete, Four storage-layer defects: identity hash contaminated with positional metadata (truncation triggers identity swap) / sw | 存储层四缺陷:身份哈希掺位置元数据(截断即换血)/清扫删除回退历史/渲染接缝图片只比长度/fork 谱系死列 #4390). Exactly-one overlap → continue; several → warn + degrade to fresh mint (never guess); none → mint. No Recorded.anchors anywhere, src/order-persist.ts untouched — the [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolve… #5719/T5 "unified store is the sole id fact source" violation is gone.
  2. Full-id mint key ✓ — <rawId>|sub:m:<FULL first inbound id>, no fp16 truncation of globally-unique host UUIDs; id-less traffic keeps legacy shapes byte-identical (pinned by T1c/T7).
  3. Eager hot-path hashing ✗ → fixed below (my commit 3495d00).
  4. Compaction fragility ✓ resolved by design: matching reads retired rows too, so T3 pins continuation after compaction erases the FIRST id; total amnesia mints a fresh coherent lane (bounded split, honest degrade) instead of re-splitting forever.
  5. Write-only anchor table / docs ✓ — anchor table gone; docs now consistent (consumer-⑥ row corrected, boundary B rewritten, log strings match code exactly; one indent regression the redo itself introduced fixed below).

Pin-gate extension (v1 condition C3, option B)

Both prep fns pin harvest-sourced ids regardless of pluginMode; stamped traffic keeps the old double gate; watermark hit still skips pinning. Truth table checked against the old code: the sole behavioral delta is proxy-mode + harvest-sourced ids now pinning (additive column, best-effort accounting) — plugin-mode behavior identical in every case.

Test quality (④)

Strong teeth: T2 asserts the reason matches /#6011/, which FORCES the sibling rung (every subagent round re-hits the tag-changed exit because the raw skeleton belongs to the main conversation; raw rescue misses by construction); T4 pins the never-guess branch; T6 proves restart continuity comes purely from store rows (no parallel ledger to lose); T10 pins no-crash degrade when the store is unavailable. The harvest suite pins the misbind path closed: a drifted-content subagent-shaped request returned the MAIN conversation's tail ids pre-fix — that exact fixture now returns null — and identical-text parallel threads → null (ambiguous, never guess).

Behavior changes (re-confirmed against code, as disclosed in the PR body)

  • (1) CC lanes separate when harvest active — gated, rollout 0%.
  • (2) UNGATED: fork key shape for STAMPED-id traffic becomes <rawId>|sub:m:<first inbound id>, replacing the legacy <rawId>|sub:<fp16(tag)>:<fp16(disc)> / |sub:task: shapes even for stamped plugin lanes; previously-stable lanes take NEW keys on their next fork; restart-stable; D1-C token lanes unaffected (token rung precedes). The one always-on delta — disclosed, intentional, restart-stable.
  • (3) claude adapter stricter whenever the session file contains sidechain rows (files without them stay on the exact legacy main pass).
  • (4) Verified-harvest requests pin msg_ids onto canonical rows (additive column).
  • Config surface: none — reuses the existing rollout key + env, both default-off.

Non-blocking observations (accepted, no action)

  • Per fork-exit cost when harvest is active = 1 sessions-table scan + N point queries (N = sub-lanes under the raw id); CC subagent rounds hit the tag-changed exit every request. Fork-exits-only, gated, SQLite point lookups — acceptable at current scale.
  • Deterministic mint keys make the pin/ingest race self-healing: a fast round-2 arriving before round-1's rows persist re-mints the SAME key (record() idempotent) rather than splitting.

Verification

CI fully green at pre-fix head 9943e44 (16 success / 0 failed, mirror cache 2026-11-10T02:52Z). On the two fix commits below: targeted suites 34/34 pass + typecheck clean locally; full-suite verdict rides CI on new head cb2c9ce. kernel/: zero files touched — kernel-change: N/A holds.

已顺手修:harvest 指纹哈希从「每个 client-provided-id 请求在注册表门控之前急切 sha256」改为门控全部通过且 JSONL 含 sidechain 行之后才惰性计算——特性全关时零哈希开销,PR 描述「opt-in 前除行为变更 2 外无变化」由此成立;附负向门控测试(门全关时 provider 绝不执行,#3617 双向纪律)+ commit 3495d00
已顺手修:docs/session-identity-resolution.md 第 3 项「Availability」列表缩进恢复为顶层项(redo 误加 3 空格缩进,GFM 将其解析为第 2 项的嵌套子项——渲染回归)+ commit cb2c9ce


八问(中文)

  1. 是什么问题:CC 并行子代理运行共用同一个宿主会话 id,fork 后全部塌进同一条 persona 车道互相污染;本 PR 让它们各进各的车道。
  2. 什么原因导致的:身份采集此前只有 stamped-id 通道(插件车道),普通 CC 子代理流量不带任何客户端自报 id,同一 raw id 下所有 fork 都落到同一个旧式 persona 键上。
  3. 影响范围与严重程度:触发条件 = 同一宿主会话内并行跑多个 CC 子代理(问题今天就存在:车道互相污染属 p1 级数据混线);修复生效需开启 harvest 通道(rollout 当前 0%,env 默认关)。另有一处未门控变更:stamped-id 流量的 fork 键形状立即切换(见第 6 问)。
  4. 怎么修复的:三段式——JSONL sidechain 投影 + 内容指纹绑定(歧义整轮降级、绝不猜)、pin 门控扩展到 harvest 来源 id、消费走统一存储标准 sessionId+msgid 匹配(唯一重叠→续用;多重叠→告警降级重新铸键;无重叠→铸全量 id 键)。
  5. 是否有回归问题:逐一审过既有行为——无 id 流量旧键形状字节不变(T1c/T7 钉住)、raw-key 救援优先级不变(T5)、D1-C token 档先于 sibling 档、watermark 命中跳过 pin、无 sidechain 行的文件走原样 legacy 主路径、存储不可用时降级铸键不崩溃(T10)。未发现无意回归。
  6. 是否破坏缓存:前缀缓存策略本身不受影响;未门控的键形状切换会让既有 stamped 车道在下一次 fork 时换新键(一次性车道分裂,重启稳定)——已在 PR 正文行为变更 2 披露。
  7. 是否改了设计方向:无架构转向——刻意放弃 v1 的并行锚点账本(违反 [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolveSessionOrder] Full compaction and persona tag drift occurring in the same turn → inevitably forks to |sub: side | [resolve… #5719/T5「统一存储是 id 唯一事实源」),改走既有统一存储 msgid 匹配消费。
  8. 是否引入了新的配置:无——复用既有 rollout 键 identityChannelHarvest(0%)与 env BILI_IDENTITY_HARVEST,均默认关。

判定:可合并(APPROVE);合并按 §7.4 留给人工(身份/会话绑定逻辑不自动合并)。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

Review complete at new head cb2c9ce: APPROVE — all five v1 findings verified resolved against the redo (unified-store msgid matching with full anchor-ledger revert, full-id mint keys, pin-gate extension, compaction-proof retired-row matching, consistent docs/tests); two small findings fixed on spot on the PR branch (lazy fingerprint hashing behind registry gates + docs list-indent restore, commits 3495d00/cb2c9ce; targeted 34/34 + typecheck clean, CI green at pre-fix head 9943e44); full review posted as PR comment (formal review API blocked: bot account is the PR author) and merge stays human-gated per §7.4 — #6067 (comment)

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

Status closeout — nothing new to act on. This wake was the mirror-sync echo of the APPROVE review posted above (full eight-question review lives in that comment / on PR #6067); no new owner input arrived. Re-verified against live state just now:

Review position stands: APPROVE — all five v1 findings verified resolved against the redo (unified-store msgid matching with the parallel anchor ledger fully reverted, full-id mint keys, pin-gate option B, compaction-proof retired-row matching, docs/tests consistent). Merge stays human-gated per §7.4 (identity/session-binding logic never auto-merges) — yours to execute once CI is green.

….md (#6067 review)

The shared-discipline bullet described the role check as a bounded
"spot-check up to 4 positions". The 4-position window existed only at
file creation (34dbcc2) and was removed by d29c73c (#5865); current
behavior is a full role-sequence check over every position where both
sides expose a role (rolesAlign, src/identity-harvest.ts). Doc-only.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

Verdict: APPROVE (head 2acea50 — doc-only commit on top of reviewed cb2c9ce)

Process note: the bot account is this PR's author, so the formal GitHub review API rejects self-review; as in the previous round, this comment carries the full review.

Independent end-to-end re-review (merge-base 541cf26..head; 11 files, +1078/−132; no kernel/, no pi-subagents/, no src/order-persist.ts — the v1 parallel anchor ledger is fully reverted per the owner ruling):

  • src/session-resolve.ts — trySiblingLane(exit) fires at ALL THREE fork exits after tryMsgIdRescue (raw-rescue precedence preserved) and before the legacy mint. Candidates = skeleton <rawId>|sub:* keys ∪ store-listed sessions, prefix-filtered; per-candidate stored-msgid overlap vs inbound ids: exactly one → continue (record(hitKey, …)), multiple → warn + demote to mint (never guess), zero → mint. Mint uses the FULL untruncated first inbound id (<rawId>|sub:m:<id>). No-id paths stay byte-identical to legacy shapes. New listStoredSessions params are lazy and fork-exit-scoped; store unavailable → skeleton candidates only → mint, no crash (T10).
  • src/identity-claude.ts — sidechain projection: parentUuid chains walked from tail rows + FULL-tail sha256 content-fingerprint match; exactly-one candidate && main tail unverified → bind composite key <sessionId>|sc:<chain-head>; zero candidates && main tail verified → legacy main pass; ambiguous/drifted/conflicting → null (the role-only fallthrough misbind path is closed). Files without sidechain rows take the exact legacy pass.
  • src/identity-harvest.ts — rolesAlign extracted verbatim from alignTailIds (behavior-preserving); computeMessageFingerprints is a lazy provider invoked only after every registry gate passed AND the claude adapter matched a file containing sidechain rows (negative gate test pins invoked === 0).
  • src/server/handle.ts + prepare-anthropic/openai — provenance tracking (body|header|harvest); pin gate extended to (pluginMode || msgIdSource === "harvest"). Truth table: sole delta = proxy-mode harvest-sourced ids now pin (additive column); plugin-mode identical in every case; openai watermark-hit skip unchanged.
  • Tests — both directions asserted per Test Blind Spot Retrospective: Reinforcing Negative Assertion Discipline, Tightening Fake Upstream Strictness, Thin Inst | 测试盲区复盘:补反向断言纪律、假上游严格化、薄安装冒烟、形态矩阵与生命周期覆盖 #3617: T1 three parallel runs → three distinct |sub:m: lanes vs T1-control pinning the pre-fix collision; T2 continues own lane (reason string forces the sibling rung); T3 retired-row match survives compaction; T4 never-guess; T5 raw-rescue outranks sibling; T6 restart from store rows alone; T7 malformed ids → legacy key; T9/T10 store edge cases; misbind fixture closed (the exact pre-fix shape now returns null); per-thread continuity catches a store rewrite.
  • Docs — ladder rung 6, consumer map (5→6, sibling scan fires before the raw-rescue call sites), boundary-B rewrite, pin-gate note, trace strings — all match the code and the actual logger calls.

Behavior changes disclosed (old → new):

  1. CC parallel subagent runs get separate lanes once harvest is active — gated by rollout.features.identityChannelHarvest (default off, rollout 0% → zero users affected today).
  2. UNGATED: fork key shape for stamped-id traffic switches from <rawId>|sub:<fp16(tag)>[:<fp16(disc)>] / |sub:task:<skel8> to <rawId>|sub:m:<first inbound id> when body-stamped ids exist — one-time lane split at the next fork exit, restart-stable (store rows carry it), D1-C token lanes unaffected.
  3. Claude adapter gets stricter whenever sidechain rows exist: unverified/ambiguous requests degrade instead of binding.
  4. Proxy-mode harvest-sourced requests now pin msg_ids (additive accounting column).

已顺手修:docs/identity-channels.md:85 把共享纪律写成 “boundary role spot-check up to 4 positions” —— 4 位窗口只存在于文件创建时(34dbcc2d04),d29c73c1cf(#5865)已移除,现行行为是全序列角色检查;已改为与 rolesAlign 文档注释一致的表述 + commit 2acea50。(历史准确的两处保留不动:src/identity-harvest.ts:188 “was a 4-position spot check”、tests/identity-harvest.test.ts:574 测试名。)

CI (mirror cache, new head 2acea50): 14 success + 3 in progress + 13 skipped, 0 failures; state “behind” = master moved on, informational. The new commit is doc-only — no code paths changed.

Merge: human-gated per §7.4 (identity/session-binding logic — auto-merge forbidden). No config surface added; no kernel changes.


八问:

  1. 是什么问题:Claude Code 并行的 Task 子代理共用同一宿主会话 id、请求头字节级相同,fork 车道键(<rawId>|sub:<fp16(tag)>:<fp16(disc)>)无法区分它们 → 全部塌缩到同一条 persona 车道,互相污染前缀缓存(v0.1.213 实测:碰撞车道加权缓存命中 68.7% vs 主会话 95.0%,约 31% 输入量背了约 74% 的非缓存成本)。
  2. 什么原因导致的:fork 出口的键只由内容指纹派生,并行运行在该维度上结构性不可区分;宿主 msgid(全局唯一 UUID)一直存在但没在 fork 决策点被消费。根因修复 = 在 fork 出口消费统一存储的 msgid 行做重叠匹配,mint 用完整 id。
  3. 影响范围与严重程度:启用面 = claude-code 插件宿主跑并行 Task(harvest 路径有灰度门,rollout 0%,当前零用户受影响);未加门的键形变更影响所有带 body 戳 id 的 fork 流量(一次性车道分裂,无数据丢失,重启稳定)。严重度:功能启用后属 p1 级缓存污染;当前被门控关闭。
  4. 怎么修复的:三件套——①三个 fork 出口加 sibling rung:候选 = skeleton 子车道 ∪ 存储列表前缀过滤,逐候选比对存储 msgid 重叠(唯一→继续 / 多个→告警+降级 mint,绝不猜 / 零→mint);②mint 键用完整首个 inbound id(去掉 fp16 截断);③claude 适配器 sidechain 投影:parentUuid 链 + 全尾内容指纹绑定复合键 <sessionId>|sc:<链头 uuid>,严格降级矩阵(歧义/未验证→null,永不落回 role-only)。指纹哈希全懒加载,置于所有注册表门之后。
  5. 是否有回归问题:逐项清点既有行为——无 id 路径与旧键形字节一致;raw-rescue 优先级保留(sibling 在其后触发);alignTailIds 重构行为保持(rolesAlign 原样抽出);openai 水位跳过保留;plugin-mode pin 各情形不变;存储不可用→mint 不崩(T10)。未发现无意回归;有意变更即上面披露的 (1)-(4)。
  6. 是否破坏缓存:否——修复本身恢复缓存效率(车道分离消除跨运行污染);lastAligned 连续性状态按会话 id 隔离(复合 sc 键保持每线程独立);pin 的 msg_ids 列纯增量。
  7. 是否改了设计方向:无新架构——扩展现有身份信道缝(Investigate and integrate a stable "session + message ID" identity channel on a per-client basis. Current status: only f | 逐客户端调查并接入稳定"会话+消息ID"身份信道:现状仅 dsh/omp/pi/opencode 四车道在跑,其余(codebuddy/hermes/claude/codex/zcode…)按清单逐个研究 #5825/[PR] release review #5826),新增一个消费者(trySiblingLane);统一存储仍是唯一 id 事实源(v1 的并行锚点账本按 owner 裁决已整体移除)。方向是强化而非转向。
  8. 是否引入了新的配置:无——不加 env/config/CLI;灰度走既有 rollout.features.identityChannelHarvest(默认关,0%)。

判定:可合并(APPROVE),合并按 §7.4 由人工执行(身份/会话绑定逻辑,禁自动合并)。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b) Status after process restart: nothing lost or blocked — the full APPROVE review (with 八问 and the 已顺手修 receipt, commit 2acea50) is posted above and the coordination line is filed on #4599. CI on head 2acea50: 0 failures. Only remaining step is the human merge (§7.4, identity/session-binding logic). No further action needed from anyone else.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Claude Code sub-agents are getting a lot of cache misses, but the main session is fine. | Claudecode 子代理大量 miss 缓存,主会话没事

1 participant