Skip to content

[PR] fix: judge recognizes plugin-lane compress pairs as summary carriers (#4401) - #4410

Open
ranxianglei wants to merge 8 commits into
masterfrom
2026-10-23_issue4401-judge-plugin-carrier
Open

ranxianglei wants to merge 8 commits into
masterfrom
2026-10-23_issue4401-judge-plugin-carrier

Conversation

@ranxianglei

@ranxianglei ranxianglei commented Oct 23, 2026 •

Copy link
Copy Markdown
Owner

Closes #4401

What this changes

The invariant judge (#3842, armed default-on since #3938) did not know the plugin-lane carrier regime (KDD#7/#1567): in plugin lanes the client's own re-sent compress pair IS the wire summary carrier — while the pair rides, stripKernelSummaries removes the kernel anchor, so a steady-state folded plugin session carries zero acp_summary markers on the wire. The judge then declared folded=false and flagged every covered inbound message as mid-history-drop — one warn per request. That is exactly the reporter's 1104/day I1 mid-history-drop storm in long-lived IM-bridge sessions. No bytes were ever lost: the full summary text rides in the re-sent call arguments; the warnings were false positives.

Fix (advisory layer only, zero traffic change):

  • A successful compress receipt (tool-result, toolName=compress, text starting [Compressed ) is now a summary carrier of the same rank as the anchor. Refused/failed receipts fold nothing and deliberately do NOT mark the regime.
  • Entry cells covered by ACTIVE blocks (host state via new activeBlockCoveredIds, wired at the relay seam) are expected absences: counted as covered K in the verdict/log instead of violations. Honored only when the exit actually carries fold evidence — the zero-carrier anomaly (no anchor AND no pair) stays red.

Behavior change disclosure (old → new)

old new
Plugin-folding session, every post-fold request I1 violation: mid-history-drop … (N finding(s), matched≈0, exempt carrier=0) WARN clean debug-level verdict … matched X/Y, covered K
Genuine defect: covered cell missing AND no carrier at all I1 warn unchanged (I1 warn)
Genuine defect: UNCOVERED cell missing from outbound I1/I2 warn unchanged (still fires)
Refused compress call + drop behind it I1 warn unchanged (refused receipts don't mark the regime)
Proxy-mode sessions (anchor kept) green unchanged (no pairs ride there)

Tests (local targeted; full suite = CI on pushed head)

  • tests/invariant-judge.test.ts 34/34 — 8 new: steady-state green w/ covered count; negative control without coveredIds; uncovered-drop teeth; zero-carrier stays red; refused receipts non-regime; multi-fold behind first fold point free; activeBlockCoveredIds unions ACTIVE blocks only.
  • tests/invariant-judge-e2e.test.ts 6/6 — new live path: real server + real /__bili/plugin/tool fold, full-history resend with riding pair judges clean (covered=3: turns 1–3 excused, pinned first user message rides as matched twin per kernel/src/prune.ts §8.1), then an uncovered cell vanishing from the outbound still fires.
  • Companions green: unified-acp-invariants, side-lane-invariants, cache-friendly-plugin (30/30). typecheck + build clean.

Open items handed back to the issue thread

  • The "answers previous question" symptom at ~350K-token folded contexts also occurs in sessions with ZERO violations — attention-level territory, not this wire defect (Q2/Q3/Q4 discussion in the issue).
  • Requesting a log excerpt from the reporter around 2026-10-22T16:10:44Z for sess=session-268a4f93 (grep [plugin] tool compress executed via plugin) to confirm lane attribution in their deployment.

Review updates (rebased + follow-up fixes)

Rebased onto master 44067a5 — the original base was ~1151 commits stale; conflicts with the I3 egress-pairing pass (#4862) and the render-lane reconcile (#4241) resolved as union of intent. Test counts above reflect the pre-rebase head.

Review found two sub-defects of the same carrier-recognition mechanism, fixed on this branch:

  1. Exemption cap vs riding pairs. The overflow ledger counted every riding compress pair as 2 carrier messages; with N active blocks (2N messages) the default cap of 128 tripped at N=65 — the reporter's 86-block session shape would have converted the mid-history-drop storm into an exemption-overflow storm. Pair volume is now bounded by 2 x |coveredIds| (one pair per active block, each covering >=1 message); excess over that ceiling still counts toward the cap, and without host state the full pair volume counts (over-warn direction only). The overflow detail string gains an additive pairExcess= field.
    • Behavior change (old -> new): plugin-lane folded session, post-fold request — old: exemption-overflow warn once riding pairs exceed 64; new: clean unless carrier volume beyond the bound trips the cap. Non-plugin sessions and state-less judge seams: unchanged.
  2. MCP-dialect compress pairs (mcp__bili__compress, zcode lane, [Bug] MCP compress 历史调用未被清理,长会话累积大量旧摘要与回执 #2970/[Bug] MCP compress 未接入 reasoning 清理及未完成调用保护 #3222) now mark the regime like bare compress (previously unrecognized by the new logic); refused/failed receipts still do not mark it.

Post-rebase verification: invariant-judge unit 52/52 (+4 new), e2e 6/6, companions (unified-acp-invariants / side-lane-invariants / cache-friendly-plugin) 33/33, typecheck + build clean. Full suite = CI on pushed head.

@ranxianglei ranxianglei changed the title fix: judge recognizes plugin-lane compress pairs as summary carriers (#4401) [PR] fix: judge recognizes plugin-lane compress pairs as summary carriers (#4401) Oct 23, 2026
@github-actions

github-actions Bot commented Oct 23, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-10-23_issue4401-judge-plugin-carrier (35f9448)

Latest build — install this exact version

npm install -g billion-context@0.0.9712-pr.4410.665f12f

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
35f9448 billion-context@0.0.9712-pr.4410.665f12f build
a4f9573 billion-context@0.0.8546-pr.4410.81adfc5 build
ffcb8a1 billion-context@0.0.8520-pr.4410.55c812d build
9d442eb billion-context@0.0.7266-pr.4410.64641fd build

Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr4410.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

查重: 不重复。 本 PR 是 #4401「invariant judge 不认识 plugin-lane carrier 制度」机制的唯一在途修复;检索 open+closed issue/PR 无同机制并行修复(相关的 #3991/#3992 为不同机制——stale cache 渗漏 vs judge 误报,不并入)。

标签: bug, dsh, scope/kernel, p2(advisory 层误报风暴,无真实字节丢失;与原 issue #4401 定级一致)

审核进行中,先报初步发现:

  1. 分支基线落后当前 master 约 9 天(base fe5ba5912 @ 2026-10-21 → master 44067a5bc @ 2026-10-30,中间 1151 个提交),期间 src/invariant-judge.ts 被多个 PR 改过——尤其 [Bug][claude] 0.1.207: subagent requests fail with 400 "tool_use ids were found without tool_result blocks" right after a fold in the parent session #4862(I3 egress pairing 不变量)恰好改动了本 PR 要动的同一行 warn 日志格式,直接合并必冲突。
  2. 按审核纪律先 rebase 到当前 master,核对与 I3 的交互(I3 对 compress pair 的豁免路径与本 PR 的 carrier 认定是否一致),再跑定向测试 + CI 后给完整结论。

@ranxianglei ranxianglei added bug Something isn't working dsh p2 scope/kernel labels Oct 30, 2026
@ranxianglei
ranxianglei force-pushed the 2026-10-23_issue4401-judge-plugin-carrier branch from 1cd1134 to 9d442eb Compare October 30, 2026 23:58
ranxianglei pushed a commit that referenced this pull request Oct 30, 2026
…dialect compress pairs (#4401)

Review follow-up on #4410 (two sub-defects of the same carrier-recognition
mechanism):

1. The exemption-overflow cap (maxExempt, default 128) counted every riding
   compress pair as 2 carrier messages. A steady-state plugin session with N
   active blocks carries 2N such messages, so the cap tripped at N=65 even
   though every pair is legitimate host state — the very sessions this PR
   exists to silence (86 active blocks in the reporter's logs) would have
   converted the mid-history-drop storm into an exemption-overflow storm.
   Pair volume is now bounded by 2 x |coveredIds| (one pair per active block,
   each covering >=1 message); excess over that ceiling still counts toward
   the cap, and without host state the full pair volume counts (safe
   direction: over-warn, never under-warn). The overflow detail string gains
   an additive pairExcess= field.

2. isCompressToolMsg matched only the bare "compress" tool name; the MCP
   dialect form "mcp__bili__compress" (#2970/#3222, zcode lane) was not
   recognized by the regime logic. It now uses the same dual-name predicate
   as the I3 pairing pass and hideConsumedCompressCalls.

Tests: +4 unit tests (86-pair steady state green at carrier=172; fake-carrier
excess still trips overflow with host state present; dialect pair marks the
regime; refused dialect receipt does not).
@ranxianglei

Copy link
Copy Markdown
Owner Author

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

Premise flag from #5866 triage (plugin-lane array-form folds end with zero full summaries on the wire):

The normalization premise "No bytes were ever lost: the full summary text rides in the re-sent call arguments" does NOT hold for native-array-form plugin folds. hideConsumedCompressCalls stubs kept compress-call summaries >200 chars (SUMMARY_STUB_CHARS, kernel/src/hide-consumed.ts), and stripKernelSummaries strips the rendered acp_summary anchor whenever a compress call quoting the same range pair EXISTS (existence+range match only, no completeness check — src/server.ts buildCarrierIndex/stripKernelSummaries). Net effect on the outbound wire: zero full copies (call stubbed + render stripped). Verified empirically on master df0a402.

Two consequences for this PR:

  1. The test fixture uses <200-char summaries, so both sides of the resent pair are byte-identical and the stub asymmetry is never exercised. Consider adding a >200-char summary case — on current master that shape judges red (ok=false, violations=[mid-history-rewrite×2, mid-history-drop×3]), i.e. After repeated folding of long conversations, I1 mid-history-drop occurs (1104 times/day), manifesting as the model an | 长会话反复折叠后出现 I1 mid-history-drop(1104 次/天),表现为模型答上一个问题 #4401's storm persists for exactly this population.
  2. Plugin-lane array-form folds end with zero full summaries on the wire after the first turn (call stubbed + render stripped) #5866 proposes option A (completeness-aware anchor strip: keep the anchor unless the outbound call summary is byte-identical to the embedded block summary). If A lands, the anchor stays on the wire ⇒ folded=true ⇒ these pairs fall into the free zone without coveredIds. The mechanisms are complementary (coveredIds still covers shapes where the anchor is legitimately absent while the call changes), but A shrinks the population this PR must defend. Owner decision pending on Plugin-lane array-form folds end with zero full summaries on the wire after the first turn (call stubbed + render stripped) #5866 — flagging so the merge order isn't locked in on the current premise.

Related: #5866

awork and others added 5 commits November 7, 2026 20:15
…invariant judge

summary carrier (KDD#7/#1567): while the pair rides, stripKernelSummaries
removes the kernel anchor, so a steady-state folded plugin session carries
zero acp_summary markers on the wire. The armed-default judge (#3938) then
declares folded=false and flags every covered inbound message as
mid-history-drop — one warn per request (1104/day in the reporter's
long-lived IM-bridge sessions). No bytes were ever lost: the full summary
text rides in the re-sent call arguments.

Teach the judge the regime: a successful compress receipt (tool-result,
toolName=compress, text starting with "[Compressed ") is a summary carrier
of the same rank as the anchor; refused/failed receipts fold nothing and do
not mark the regime. Entry cells covered by ACTIVE blocks (host state via
new activeBlockCoveredIds, wired at the relay seam) are expected absences,
counted as "covered K" in the verdict — honored only when the exit actually
carries fold evidence, so the zero-carrier anomaly (no anchor AND no pair)
stays red.
Unit: steady-state shape green with covered count; negative control without
coveredIds red; uncovered-drop teeth fire; zero-carrier anomaly stays red;
refused receipts do not mark the regime; multi-fold behind the first fold
point is free I2 territory; activeBlockCoveredIds unions ACTIVE blocks only.

Live e2e: real server + real /__bili/plugin/tool fold, full-history resend
with the riding pair judges clean (covered=3 — turns 1-3 excused, pinned
first user message rides as matched twin per kernel/src/prune.ts), then an
uncovered cell vanishing from the outbound still fires.
…dialect compress pairs (#4401)

Review follow-up on #4410 (two sub-defects of the same carrier-recognition
mechanism):

1. The exemption-overflow cap (maxExempt, default 128) counted every riding
   compress pair as 2 carrier messages. A steady-state plugin session with N
   active blocks carries 2N such messages, so the cap tripped at N=65 even
   though every pair is legitimate host state — the very sessions this PR
   exists to silence (86 active blocks in the reporter's logs) would have
   converted the mid-history-drop storm into an exemption-overflow storm.
   Pair volume is now bounded by 2 x |coveredIds| (one pair per active block,
   each covering >=1 message); excess over that ceiling still counts toward
   the cap, and without host state the full pair volume counts (safe
   direction: over-warn, never under-warn). The overflow detail string gains
   an additive pairExcess= field.

2. isCompressToolMsg matched only the bare "compress" tool name; the MCP
   dialect form "mcp__bili__compress" (#2970/#3222, zcode lane) was not
   recognized by the regime logic. It now uses the same dual-name predicate
   as the I3 pairing pass and hideConsumedCompressCalls.

Tests: +4 unit tests (86-pair steady state green at carrier=172; fake-carrier
excess still trips overflow with host state present; dialect pair marks the
regime; refused dialect receipt does not).
…dge (#4401)

The PR's normalization premise — full summary text rides in the re-sent call arguments — does not hold on the outbound wire for native-array-form plugin folds: hideConsumedCompressCalls stubs kept-call summaries over 200 chars and stripKernelSummaries strips the acp_summary anchor on existence+range match only, so a fold whose summary exceeds the stub threshold left both halves of the re-sent pair differing byte-for-byte (call half stubbed) while the anchor was gone. The judge read that as mid-history-rewrite/mid-history-drop — #4401's storm persisted for exactly the long-summary population, which the short-summary fixtures never exercised.

stableEqual now accepts a directional sanctioned-compress-twin proof: same raw prefix, same non-content fields, and the exit content list an order-preserving subsequence of the entry list where each surviving entry matches by exact range key (startId/endId, messageId fallback) and verbatim-or-exact-stub summary (199-char prefix + U+2026, surrogate-safe). Mirrors kernel/src/hide-consumed.ts lockstep (not imported — judge stays a leaf); unparseable args fail safe to plain byte compare. The range-filter facet (dead-range entries dropped outbound) is covered by the subsequence match.

Tests: unit POS (>200-char stub, multi-range drop) + NEG teeth (non-exact clip, rewritten summary, planted range all stay red) + e2e live path asserting the forwarded call half really is the stubbed form upstream while the armed judge reads the steady-state request as a clean fold. Related: #5866 (option A would keep the anchor for this shape; complementary, merge-order independent).
@ranxianglei
ranxianglei force-pushed the 2026-10-23_issue4401-judge-plugin-carrier branch from 9d442eb to ffcb8a1 Compare November 7, 2026 13:27
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

查重: 不重复——你的 flag 与 #4401 同属一个机制家族(judge 不认 plugin-lane compress pair),按「一个机制一个 PR」纪律直接在本 PR 内修掉,不另立 issue。
标签: bug + p2 + scope/kernel + dsh + pr(维持现状)——确定性日志噪音风暴、有绕行 = p2;修复跨车道但症状人群是 dsh IM-bridge 会话。

前提确认:你说得对,已实证

代码级 + 线上双确认:稳态下 >200 字符摘要的 plugin fold 出站线上零份全文——kernel hideConsumedCompressCalls.compactEntry(SUMMARY_STUB_CHARS=200,clampPrefix(199)+"…")把 call 半截打桩,host stripKernelSummaries 按存在性+range 匹配(无完整性检查)剥掉渲染锚点。新增 e2e 直接抓取转发 body 证明:重发的 call arguments ≠ 客户端原文且含 \u2026 桩。PR 原前提「全文随 call arguments 重发」对该人群不成立。

修法:judge 侧识别 kernel 授权的 call-半截投影(commit 1f0f0c2)

不动 host 侧 strip(那是 #5866 option A 的领地,owner 未拍板,本 PR 零触碰)。在 judge 的 stableEqual 里加一条方向性结构证明(与既有 isTruncatedTwin/isCcrStoredResult 同一模式):exit call 半截当且仅当等于 entry 半截在 kernel 授权变换下的投影时视为等价——① 长摘要被替换为精确的 clampPrefix(199)+"…" 桩;② 非存活 range 条目被保序移除(rewriteCompressText facet,tier-merge 会话里多 range call 引到死 range 时可达)。解析失败回退纯字节比较(fail-safe,永不掩盖)。负控钉住:假桩(非精确前缀)、改写摘要、植入 range 三者仍红。测试:4 个单测 POS/NEG + e2e 活路径(桩在线上可证 + 判决 ok/covered=5)。

行为变更披露(旧→新):旧:>200 字符摘要的 plugin fold 稳态判红(mid-history-rewrite on stubbed half;裸 master 上还有 mid-history-drop×N)→ 新:判绿,pair 计入 covered carrier。仅 advisory 层(warn 日志/面板),零流量变更;#4401 风暴对长摘要人群止住。≤200 字符路径逐字节不变。

与 #5866 option A 的关系:互补,合入顺序互不锁死

A 落地后锚点留在线上 ⇒ sumIdx=0 ⇒ pair 落自由区,根本走不到 twin 检查(照样绿);A 不落地则由 twin 检查兜住。本修复两条路都成立,不预设 A 的取舍。注意 A 本身改的是 host 侧 wire 形状(completeness-aware strip),属 #5866 的 owner 决策项,这里不越界。

八问

  1. 是什么问题:长摘要 plugin fold 出站线零份全文(kernel 打桩 + host 剥锚各毁一份),judge 把桩误读成历史改写/丢消息 → After repeated folding of long conversations, I1 mid-history-drop occurs (1104 times/day), manifesting as the model an | 长会话反复折叠后出现 I1 mid-history-drop(1104 次/天),表现为模型答上一个问题 #4401 误报风暴覆盖该人群。
  2. 什么原因:两个独立的授权变换各自销毁了归一化前提假设会在线上的那份副本;judge 对 call 通道只有字节相等,没有「授权投影」概念。
  3. 影响范围与严重程度:长寿命 plugin 会话中摘要超 200 字符的 fold;仅 advisory 噪音,无功能影响;p2(确定性噪音类)。
  4. 怎么修的:judge 学会结构识别 kernel 授权的 call-半截投影(方向性 twin 证明),而非放宽字节相等。
  5. 是否有回归:审过既有行为——≤200 字节相同路径不变、refused receipt 仍非 regime、MCP-dialect 不变、pair-volume cap 不变;负控证明不可掩盖(假桩/改写/植入均仍红);无阈值/时序/输出格式漂移。
  6. 是否破坏缓存:不涉及(advisory only)。
  7. 是否改了设计方向:没有——沿用 judge 既有 sanctioned-transform 模式;option A 仍是 Plugin-lane array-form folds end with zero full summaries on the wire after the first turn (call stubbed + render stripped) #5866 的独立决策项。
  8. 是否引入新配置:没有。

CI(head ffcb8a1,已 rebase 到 master 5af5f9f + merge ci-fixes)

11 条 e2e/gate 车道全绿;核心矩阵 6/12 shard 红 = 全部为 #5860 F3 继承签名(silent-backend-nudge ×3 + fork-nudge-trigger ×2 + issue4241 render-baseline,逐条比对日志与 #4599 floor 1103 清单吻合,#728B 通过),等 #5867 owner A/B/C 拍板,全舰队分支同红,已在 #4599 留一行。本 PR 自身测试在对应 shard 内全绿,零 PR 专属失败。本地定向验证:invariant-judge 族 64/64 + companion 24/24,typecheck/build 干净。

判定:ready for review/merge;残余红为继承性、已登记,不阻塞本 PR 的审查结论。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

本轮结论:6 个红分片全部是 master 继承红(#5860 波次的 nudge 行为测试钉),与本 PR 代码无关;master 侧的裁定修复(#5882)已并入本分支并推送,等 CI 重跑。

本轮做了什么(判断,非流水账)

  1. 先复核活状态:run 42377083141 实际已 completed/failure @ ffcb8a18dc(13:44Z 结束);唤醒时镜像缓存显示 queued 只是运行中快照,不是已转绿。
  2. 日志逐条对账(拉了 6 个失败分片中 4 个的完整日志):失败项与 CI Control Center: Centralized CI Failure Handling & Status Sync (Comment all CI issues here; do not open new issues) | CI 总控: CI 失败集中处理与状态同步(所有 CI 问题在此评论, 勿开新 issue) #4599 floor 1103 的归因清单逐条吻合——silent-backend-nudge ×3(#728A/#728C/effectiveTokenCount's text-only raw bound caps away the image term from the nudge numerator (silent-backend sessions) #1119)、fork-nudge-trigger ×2、issue4241-judge-render-baseline ×1;断言指纹(expected trailing user nudge, got last message: [tool schema])逐字一致,#728B 按文档通过。各分片失败数与文档分布精确匹配,零 PR 专属失败;全部静态门禁在本 head 上绿。
  3. 关键变化:owner 已在 The storage-render golden snapshot on master went stale after kernel 0.0.184 (#5860): the next scheduled CI run will def | master 的 storage-render 黄金快照在 kernel 0.0.184(#5860)之后过期:定时 CI 下次必红 #5867 拍板 B(有界延迟)——修复以 [PR] Infinite Wait Issue in Stress Injection: N-Round Fallback (Side Effect Introduced by Fix #5860) | 压力注入无限等待问题:N轮兜底(修复#5860引入的副作用) #5882 落地 master(kernel 0.0.185 maxEstimateHoldRounds estimate-hold backstop + 行为测试重钉 + golden 对 0.0.185 再生),ci-fixes 同步至 d7769ca5b0(tree 与 master 逐字节相同,哨兵实测 run 42384566808 30/30 全绿)。
  4. 在 PR 分支 merge origin/ci-fixes:零冲突,新 head a4f9573662。合并后 tree = ci-fixes tip + 仅本 PR 自己的 4 个文件(src/invariant-judge.ts、src/server/relay.ts、tests/invariant-judge.test.ts、tests/invariant-judge-e2e.test.ts);golden 文件与 master 的 0.0.185 版逐字节一致。
  5. 本地定向验证(全量以 CI 为准):typecheck clean;本 PR 自有 invariant-judge 单元 57/57 + e2e 7/7;原红三族 silent-backend-nudge 4/4、fork-nudge-trigger 3/3、issue4241-judge-render-baseline 1/1 全部转绿。
  6. 推送:ffcb8a18dc → a4f9573662(fast-forward;推送前 fetch 核对无并发推送)。已在 CI Control Center: Centralized CI Failure Handling & Status Sync (Comment all CI issues here; do not open new issues) | CI 总控: CI 失败集中处理与状态同步(所有 CI 问题在此评论, 勿开新 issue) #4599 留一行登记。

结构化八问

  1. 是什么问题:PR 检查矩阵里 6 个 test 分片红,全是 nudge 行为测试的旧断言挂掉——master 带病(born-red),每个 PR head 都继承。
  2. 什么原因导致的:今日 [PR] fix(nudge): no billing sample this round → hold pressure-band injection one round (phantom-pressure guard) #5860(kernel 0.0.184)把「本地估算读数时压力带 nudge」延后一轮等新语义,但没重钉既有行为测试,master push 不触发 CI 所以一直隐形,定时 cron 暴露;本分支基线早于修复。属产品契约项,已由 owner 拍板 B 并经 [PR] Infinite Wait Issue in Stress Injection: N-Round Fallback (Side Effect Introduced by Fix #5860) | 压力注入无限等待问题:N轮兜底(修复#5860引入的副作用) #5882 修复。
  3. 影响范围与严重程度:仅本 PR 的检查矩阵(合并阻塞信号级问题);用户侧零影响、无数据丢失;本 PR 自身零失败。
  4. 怎么修复的:源头收敛——把含 [PR] Infinite Wait Issue in Stress Injection: N-Round Fallback (Side Effect Introduced by Fix #5860) | 压力注入无限等待问题:N轮兜底(修复#5860引入的副作用) #5882 裁定修复的 ci-fixes 合进 PR 分支,分支侧不重复修 master 级失败(政策要求修复只收敛到 ci-fixes)。
  5. 是否有回归问题:无。合并是纯 master 内容集成、零冲突、本 PR 4 个文件零改动;本地 typecheck + 定向套件全绿;本轮未引入任何语义/阈值/输出格式漂移。
  6. 是否破坏缓存:不涉及。
  7. 是否改了设计方向:没有——本 PR 维持 advisory-only 审计层修复;judge 载体识别机制不变。
  8. 是否引入了新的配置:本 PR 无。(kernel 0.0.185 的 maxEstimateHoldRounds 等是 master 侧已发布内容,非本轮引入。)

备注:floor 11 的 >200 字符 stub 不对称前提标记仍在 #5866 跟踪(待 owner 拍板),不阻塞本 PR 的 CI。

收工判定:CI 重跑转绿后 ready-to-merge(合并由人执行)。

The suite baseline preload (preload-storage-baseline.mjs) pins
BILI_STORAGE_UNIFIED="legacy" for the broad corpus. After #6480 flipped the
storage default to pure-sqlite, the notice-registry suite's visibleNotices
assertions expect the pure-sqlite default notice, which is gated on
unifiedSelectionIsSqlite (unified "sqlite" or absent) — the preload's
"legacy" made the notice silent, so the suite was born-red on master. Pin
the selection explicitly (#4011 lesson: a test that depends on the production
default passes locally but fails in CI).
# Conflicts:
#	src/invariant-judge.ts
#	src/server/relay.ts
#	tests/invariant-judge.test.ts
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

Merge conflict resolved — branch is back at current master, pushed as 35f9448eb6.

What happened: master advanced ~40 commits past this branch's base (kernel 0.0.195, #6003 egress-chokepoint relay refactor, #5478 invariant-provenance / #564/#2627 separator provenance, #4862 compress-pairing, pure-SQLite default, announcement policy, …) and the PR was flagged dirty.

What I did — merged origin/ci-fixes (c6369fb6a1 = master e1f31ad9d5 + sentinel's tests/notice-registry.test.ts fix) into the branch. 3 files conflicted, all three touched by both sides; every hunk was additive on both sides, so all resolved as union of intent:

No semantic change to either side: the merged tree is exactly ci-fixes tip + this PR's 4 files (225/5/144/305 lines — same shape as before the merge).

Verification on the merged head: npm run typecheck clean; targeted suites green 71/71 (tests/invariant-judge.test.ts + tests/invariant-judge-e2e.test.ts, covering both this PR's tests and the newly merged #2627/#4862 tests). Fetched before push — no concurrent pushes, fast-forward push, no force-push. CI re-runs on the new head; the full-suite verdict is the remote CI lanes.

中文八问

  1. 是什么问题:PR 分支与 master 冲突(dirty)——master 前进了约 40 个提交,与 PR 在同一批文件上各有新增。
  2. 什么原因导致的:PR 分支基于旧 master;期间 master 合入了 # [Feature] Support user-registerable egress request body rewriting plugin pipeline | # [Feature] 支持用户可注册的出站(egress)请求体改写插件管线 #6003(relay egress 重构)、[PR] fix(invariants): bind separator exemptions to producer-owned wire items #5478/Compression breaks Responses assistant history ordering, causing repeated invalid_assistant_history 400s #564/[Bug][Responses] acp_turn_sep 合成分隔符进入持久压缩覆盖集合,导致重复历史漂移误报 #2627(invariant-judge 分隔符溯源)、[Bug][claude] 0.1.207: subagent requests fail with 400 "tool_use ids were found without tool_result blocks" right after a fold in the parent session #4862(compress 配对)等,与本 PR 在 src/invariant-judge.ts、src/server/relay.ts、tests/invariant-judge.test.ts 三处同位置追加代码,产生文本冲突。
  3. 影响范围与严重程度:仅阻塞本 PR 合并;不影响任何已发布行为(冲突全部是两侧各自的新增可选字段/新测试段)。
  4. 怎么修复的:把 origin/ci-fixes(= 最新 master + 哨兵测试修复)merge 进 PR 分支,6 个冲突 hunk 全部按「意图并集」解决——两侧新增内容都保留,调用点合并为单次调用同时传 coveredIds 与分隔符溯源字段;未改动任何一侧的语义。
  5. 是否有回归问题:无。合并树 == ci-fixes tip + 本 PR 的 4 个文件(行数形状与合并前一致);typecheck 干净;定向测试 71/71 全绿(含本 PR 全部测试与 master 新并入的 [Bug][Responses] acp_turn_sep 合成分隔符进入持久压缩覆盖集合,导致重复历史漂移误报 #2627/[Bug][claude] 0.1.207: subagent requests fail with 400 "tool_use ids were found without tool_result blocks" right after a fold in the parent session #4862 测试)。
  6. 是否破坏缓存:不涉及(未触碰任何缓存/重试/游标逻辑)。
  7. 是否改了设计方向:不涉及(纯冲突消解,无方向变更)。
  8. 是否引入了新的配置:不涉及。

收工判定:冲突已解、分支已对齐当前 master 并推送(35f9448eb6),等 CI 在新 head 上转绿即可合并。

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

Labels

bug Something isn't working dsh p2 scope/kernel

Projects

None yet

1 participant