Skip to content

[PR] fix: preflight-tier batch — deterministic tier-exhaustion verdict, repacked-tail reanchor, anthropic egress clamp - #6435

Open
ranxianglei wants to merge 21 commits into
masterfrom
2026-11-12_preflight-tier-batch
Open

ranxianglei wants to merge 21 commits into
masterfrom
2026-11-12_preflight-tier-batch

Conversation

@ranxianglei

Copy link
Copy Markdown
Owner

What

Root-cause batch fix for the preflight-tier aggregation thread (#6378). One PR fixes three of the four issues in the batch; the fourth (#5871) is a misdiagnosis (detailed below).

This PR cherry-picks the three well-tested individual PRs into a single batch PR (preserving authorship with -x) so the root cause is fixed once.

Fixes

#5871 — misdiagnosis, not fixed

#5871 claims the native gate's B/4 token estimate "overestimates CJK by 7×" (a 5.65 MB body → 1,412,836 tok B/4 vs "202K actual"). This is wrong in direction.

  • The kernel's tokenizer (kernel/src/tokenize.ts:5-13) counts CJK as 1 token/char (comment: "chars/4 badly underestimates them").
  • B/4 for pure CJK (3 bytes/char) = 0.75 tok/char → under the real ~1–1.9 tok/char. So B/4 is a valid conservative floor for CJK, not an overestimate.
  • The "202K actual usage" is from a compressed request (the session's normal operation), NOT the raw 5.65 MB body — the provider never received that body (the proxy was dead). It is an invalid comparison.
  • A real CJK-dominant 5.65 MB body is ≈1.1–1.9M tok = over the 1M window, so the gate's rejection was correct.

The one legitimate sub-point (the gate used the 1M floor instead of the model's actual >1M window) is by design: MIN_ASSUMED_WINDOW_TOKENS=1_000_000 (src/util.ts:30) is a deliberate floor so a wrong-SMALL declared window can't false-block; the dsh lane not stamping the real window is host-side, not a bili defect.

No safe bili fix exists in the claimed direction (lowering the B/4 ratio would loosen the floor → false accepts for CJK, the wrong direction). #5871 is therefore NOT claimed as fixed.

Supersedes / absorbs

This batch PR supersedes and absorbs the three individual PRs (cherry-picked with -x, authorship preserved):

Those PRs stay open (agent cannot close); this PR is the consolidated deliverable for the root-cause thread.

Config surface

None. No new config fields, env vars, or CLI flags.

Behavior changes (explicit)

  1. Structurally-exhausted tier-recovery: 502 (retryable) → 413 (deterministic, non-retryable). Cap/unavailable/no-shrink endings unchanged (still 502-retryable).
  2. New reanchor rung in resolveSessionOrder: tag-changed shorter replays whose canonical tail survives a host record repack now reanchor to the raw id (previously → derived-persona → 413).
  3. Anthropic lane: max_tokens is now clamped against the RAW native window (previously no clamp on this lane); the overflow-arm evidence tag now survives the post-rebuild baseline anchor write (previously overwritten by an estimate-grade tag).

@github-actions

github-actions Bot commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-11-12_preflight-tier-batch (53fea86)

Latest build — install this exact version

npm install -g billion-context@0.0.10494-pr.6435.d715f81

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
53fea86 billion-context@0.0.10494-pr.6435.d715f81 build
f4e289d billion-context@0.0.10453-pr.6435.5cdcdc6 build
b309b27 billion-context@0.0.10249-pr.6435.0e13423 build
c9d9ce7 billion-context@0.0.9690-pr.6435.6a51f1d build
a8b5fa6 billion-context@0.0.9565-pr.6435.19dcfdc build
9d18c77 billion-context@0.0.9544-pr.6435.ad437d1 build
ad7a297 billion-context@0.0.9499-pr.6435.e48d47f build
669ef38 billion-context@0.0.9327-pr.6435.724c1ac build

Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr6435.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei ranxianglei changed the title fix: preflight-tier batch — deterministic tier-exhaustion verdict, repacked-tail reanchor, anthropic egress clamp [PR] fix: preflight-tier batch — deterministic tier-exhaustion verdict, repacked-tail reanchor, anthropic egress clamp Nov 12, 2026
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

查重: 不重复。本 PR 是 #6378(preflight-tier 根因聚合)的批处理修复 PR,cherry-pick #5769/#5851/#5899 三个已验证修复(-x 保留署名),方向与聚合线程一致,无并行重复工作。

审核结论: 3 个修复(#5764/#5844/#5880)均扎实、测试充分,可合并。CI 全绿(core 矩阵 27 通过 + 1 条件跳过,15 个 gate 全绿,无失败)。本地定向跑 3 个新测试文件 14 用例全过。

⚠️ 重要发现 — PR 的 #5871 分析基于镜像同步错误

PR 正文称「#5871 是误诊」,分析的是「#4516 的 oversize 断路器在 CJK 长会话上误触发两次」的 B/4 token 估算问题(5.65MB body → 1,412,836 tok B/4 vs「实际 202K」)。但真实 GitHub #5871 是一个性能问题:「0.1.212 仍复现 #4528 的结构性问题:单请求本地预处理最高 144.7s」——O(messages) 同步扫描饿死事件循环,累计 253 min,dsh Desktop/Windows。owner 已在 #5871 第 6 楼将其归入 [Root Cause] performance-memory(GitHub #6449)。

聚合线程 #6378/#6451 把「GitHub #5871」描述为 CJK 断路器问题,这是镜像同步错误——镜像 #5872 的内容与真实 GitHub #5871 不符。PR 的 #5871 技术分析(kernel/src/tokenize.ts:6-12 CJK 1 tok/char、src/util.ts:30 MIN_ASSUMED_WINDOW_TOKENS=1M 是刻意 floor、B/4 对纯 CJK 是保守下界而非高估)孤立看是正确的,但挂错了 issue 号——真实 #5871 不是 preflight-tier 问题,也不被本 PR 触及。

建议: PR 的 #5871 段落应修订——要么删除(真实 #5871 是性能问题,已在 #6449),要么明确该段分析的是聚合线程的描述(镜像同步错误)而非真实 GitHub #5871。镜像同步错误本身应单独报告修复。

回归审查(逐项过既有行为):

8 问:

  1. 是什么问题: preflight-tier 批次 3 个缺陷——(a) 结构性 tier-exhaustion 被标 502 可重试(实为确定性 413);(b) tag-changed 更短重打包重放被误 fork 成 derived-persona(应 reanchor);(c) anthropic lane 的 max_tokens 未按 RAW window clamp(其他 lane 都有)+ overflow-arm tag 被 anchor 写覆盖。
  2. 什么原因导致的: (a) preflight 的 stopReason 已有「exhausted」(结构性证明)但没线程化进返回值,server 一律按 502 处理;(b) session-resolve 的 tagChanged 分支缺 canon 指纹级 tail 对齐 rung,重打包后 kind 序列变短就掉进 fork;(c) anthropic lane 漏接其他 lane 都有的 clampOutgoingOutput,且 anchor 写没检查 overflow-arm 状态。
  3. 影响范围与严重程度: (a) 结构性超窗会话收到 502 而非 413,客户端可能反复重试(体验受损,非数据丢失);(b) 长会话重打包后 resume 失败(fork 成 derived-persona 丢上下文)——p1 级;(c) anthropic near-wall 请求 max_tokens 不 clamp,上游可能拒收(体验受损)。
  4. 怎么修复的: (a) stopReason 线程化进返回值,exhausted 标 deterministic,server 据此返回 413;(b) 新增 canonTailEquals rung(按 neutralTurnCanonical 的 sha256 对齐最后 ≤8 个逻辑 turn),在 fork 之前 reanchor;(c) anthropic lane 接入 clampOutgoingOutput(与其他 lane 结构相同),新增 applyPostRebuildBaselineAnchor 带 overflow-arm guard。
  5. 是否有回归问题: 无。逐项过既有行为(见上),3 个修复都有 negative-gate 测试确认变更前行为在应保留处保留。
  6. 是否破坏缓存: 否。3 个修复都不触碰缓存/重试/增量游标/去重键。Resuming a long session fails with preflight 413 when persona fingerprint and replay shape change together (0.1.212) #5844 的 canon 列是 order-snapshot 增量字段(wire shape additive),不影响现有缓存键。
  7. 是否改了设计方向: 否。3 个修复都在既有机制内(stopReason 线程化、rung 扩展、clamp 接入),无架构/边界/数据流转向。
  8. 是否引入了新的配置: 否。无新 env/config/默认值。

判定: 3 个修复可合并(扎实、测试充分、CI 全绿、无回归)。#5871 段落建议修订后再合(基于镜像同步错误,真实 #5871 是性能问题已在 #6449);镜像同步错误应单独报告。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

状态同步(非重复审核):完整 8 问审核已在第 6 楼给出,此处仅确认当前 live 状态。

  • PR head 未变:仍为 669ef38,我审核后无新 commit。
  • CI 绿:镜像缓存 summary 30 checks / 0 failed / 0 pending(core 矩阵此前已逐项核对全绿)。
  • 无新 owner 评论:第 6 楼后仅系统 plumbing。

结论不变:3 个修复(#5764/#5844/#5880)可合并;#5871 段落基于镜像同步错误(真实 GitHub #5871 是性能问题,已在 #6449),建议修订后再合。线程现等待 owner 决策。

ci-check and others added 11 commits November 13, 2026 13:42
A bounded tier-distillation recovery that ends because span-math proved no legal
two-block span (e.g. oversized inline media blowing every candidate span past the
60k source budget) is a structural impossibility: the identical payload can never
recover regardless of upstream health. But the verdict was installed without the
`deterministic` flag its sibling futile-round paths carry, so the server fail-fast
degraded to 502 instead of 413 and status-code-driven hosts read it as transient
and spun on doomed retries (field: 5+ in 8s).

Propagate the recovery stopReason out of attemptTierRecovery and flag only the
structurally-exhausted ending deterministic; cap/unavailable/no-shrink keep their
502-retryable contract per the existing fail-fast semantics (#3461/#3918).

Behavior change: structurally-exhausted tier recovery now answers 413
(non-forwarded, non-retryable) instead of 502. Relay/budget-dependent endings are
unchanged. Fixes #5766.

(cherry picked from commit cf18060)
…al mirror number)

(cherry picked from commit db80978)
… a host record repack (#5844)

A host upgrade that changes BOTH the system prompt text and the message-record layout (dsh 0.2.1->0.2.2-rc.x: assistant text segments merged into adjacent records, some standalone assistant records dropped) destroys every byte-level continuation evidence at once: per-source-message fragment hashes are disjoint even though every logical turn is identical, so all four tagChanged escape rungs miss and the conversation forks onto an empty |sub:<fp> lane -> preflight folds from scratch -> resume dies with a fail-fast 413.

Add a canonical-turn fingerprint rung at the #5686 site: when the replay is strictly shorter and kind-tail-aligned, reanchor onto the raw id if the last <=TAIL_WINDOW turns match the recorded skeleton turn-by-turn in CANONICAL form (neutralTurnCanonical — the identity chain's own hashing input), which survives record regrouping but not content mutation. The skeleton persists the canon column additively (absent on pre-change snapshots); a missing or malformed column degrades to exactly the pre-change byte-fragment path. The #5686 fragment reason string stays byte-identical.

(cherry picked from commit 5e0dd2a)
…, degradation)

Six pins: the bug shape reanchors with the new content-verified-tail reason; a mutated last user cell still forks derived-persona; a skeleton hydrated without the canon column forks exactly as before this change; the snapshot round-trip keeps the rescue alive across restarts; a malformed canon column degrades to the byte-fragment path; and a tag-invariant repacked shorter replay already reanchors through the pre-existing owner-rule branch (symmetry row).

(cherry picked from commit 62aa540)
…and exists (#5844)

SESSION-IDENTITY en/zh: state plainly that there is no supported command or config knob to re-attach a drifted |sub lane to its parent, and give the sanctioned manual procedure (stop the proxy first — the daemon silently rewrites session-order.json from memory; set the parent entry's tagHash to the sha256 hex digest of the host's current full system text or delete the entry; restart; bili sessions rm the orphan lane). Answers the reporter's section-8 question from #5844.

(cherry picked from commit 676c259)
…e length drifts from kinds

The writers keep canon 1:1 with kinds (both derive from the same turn list in record()), so a length mismatch means the column is shifted/corrupted. Degrade to absent (byte-fragment path = pre-change behavior) instead of trusting misaligned fingerprints.

(cherry picked from commit a1e0a7b)
… a kind break

The original guardrail replaced an assistant record with a user record, so the fork came from tailAligns' kind suffix comparison and never reached the canonical-fingerprint comparison it was meant to pin (it would pass even with the canon rung removed). Rewrite it to mutate the final tool result's text — layout and kinds unchanged — so the fork can only come from the canon mismatch itself.

(cherry picked from commit 1d37f63)
…rified tail

An assistant turn carrying tool calls hashes as calls+role only (cross-wire neutrality — OpenAI renders such turns with content:null), so a prose-only difference inside a calls turn is invisible to the #5844 rung exactly as it is to the rest of the identity chain that shares neutralTurnCanonical. State that boundary next to the manual re-attach procedure (en+zh).

(cherry picked from commit c39dc84)
…#5880)

Two independent defects let a session bili believes fits request more than
the upstream accepts on the Anthropic wire when the provider bills the
completion against the same window as input (DeepSeek et al.):

1. The post-preflight-fold baseline anchor (server.ts) overwrote an ACTIVE
   overflow-arm evidence tag with an estimate-grade one whenever the fresh
   rebuilt-view estimate measured slightly above the netted arm value. Every
   successful fold silently disarmed the evidence floor; the next turn then
   fit-forwarded on a stale optimistic estimate until the 400 re-armed it -
   a fold -> forward -> 400 livelock. Extracted the write into
   applyPostRebuildBaselineAnchor (session.ts) with an overflow-arm guard,
   mirroring the armFailureShrink doctrine (relay.ts #2484). Unarmed and
   usage-grade behavior is unchanged (#1987/#604).

2. The outgoing max_tokens clamp had no call site on the anthropic lane
   (budget.ts wired openai/responses/google only). Added it, sized against
   the RAW native window exactly like the other lanes (#2096 semantics:
   nudge/preflight judge the headroom-adjusted window, egress guards the raw
   one). For canonical Anthropic (separate input/output limits) the requested
   budget already rides under the window and the clamp is a no-op.

Tests: tests/issue5880-anthropic-clamp-arm.test.ts - unit pins for the arm
guard (both directions) plus hermetic e2e through /v1/messages proving the
clamp fires against the raw window near the wall and stays quiet when the
request fits.

Fixes #5880

(cherry picked from commit eba8a07)
@ranxianglei
ranxianglei force-pushed the 2026-11-12_preflight-tier-batch branch from 669ef38 to ad7a297 Compare November 13, 2026 06:24
ci-check added 2 commits November 13, 2026 19:50
The #5764 fix marked structurally-exhausted tier recovery as deterministic
(413 not-forwarded). But an overflow-armed baseline is a relay-billing
artifact, not a structural doom: the upstream proved a smaller window, so
the client may still recover by switching models or the upstream recovering.

The preflight-image-budget upstream-overflow test (merged to master in
8708200) pins this: confirmed upstream overflow must stay 502-retryable,
not 413. Refine the deterministic flag to exclude the overflow-arm case.
Sync PR #6435 with master (9 commits) to clear the behind/dirty merge
state. No conflicts: the two sides touch disjoint file sets.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

Conflict resolved — PR #6435 is now in sync with master.

Live verification (this run, before acting):

What I did:

  1. Verified the live merge state + CI (mirror cache) before touching anything (status-freshness policy).
  2. Checked semantic coupling: the PR's files import knobs/persist/pck-notice/thirdparty-scan (all touched by master). Confirmed master's only removed export (resolveJsonWriterOff) is used solely by master-side files, not the PR's.
  3. Merged master into the PR branch → merge commit a8b5fa6627. Verified typecheck passes on the merged tree (exit 0).
  4. Verified the remote head was unchanged (no concurrent push), then pushed the merge commit (non-force: 9d18c77a55..a8b5fa6627).

Current state:

  • mergeable: True, mergeable_state: blocked — blocked only because the required status checks are running on the new head (expected, not a conflict).
  • CI on the new head (a8b5fa6627): 0 failed; fast gates green (gate, static-gates ×2, one test shard); test matrix in progress (11 pending).

8问:

  1. 是什么问题: PR [PR] fix: preflight-tier batch — deterministic tier-exhaustion verdict, repacked-tail reanchor, anthropic egress clamp #6435(preflight-tier 批次)被标记为与 master 冲突(merge state: dirty),阻塞合并。
  2. 什么原因导致的: 冲突标记是过期的。PR head 已更新(9d18c77a55,11:50 UTC 推送),与 master 实际无冲突;PR 只是落后 master 9 个提交(behind),并非真正冲突。
  3. 影响范围与严重程度: 无用户可见影响——纯分支同步/合并卫生问题。严重程度:低(PR 本就可合并,只是落后)。
  4. 怎么修复的: 把 master(9 个提交)合并进 PR 分支。零冲突(两侧文件集不相交);验证无语义破坏(合并树 typecheck 通过;master 删除的 resolveJsonWriterOff 导出未被 PR 文件引用);推送合并提交 a8b5fa6627。
  5. 是否有回归问题: 无。合并是纯增量(文件集不相交),合并树 typecheck 通过。
  6. 是否破坏缓存: 否。未触碰缓存/重试/增量游标/去重键。
  7. 是否改了设计方向: 否。
  8. 是否引入了新的配置: 否。

判定: 冲突已解决,PR 已与 master 同步,可合并(待 CI 测试矩阵跑完 + owner 合并)。CI 当前 0 失败;若测试矩阵后续变红,我作为本 PR 的 CI 工程师会继续跟进修复。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

All 5 red shards investigated — none are caused by this PR or by the master-sync merge I pushed (a8b5fa6627).

Attribution (from job logs, run 43435152756):

Shard Failing test(s) Root cause Class
ubu22-3, ubu24-3, win22-5 visibleNotices ×2 (storage-sqlite-default missing) #6480's new storage-sqlite-default builtin notice is gated by sqliteDefaultNoticeActive (src/notices.ts:82), which requires the pure-sqlite selection; the CI env resolves storage to full and the hermetic test doesn't pin the storage env vars → notice never shows Inherited/born-red from #6480 (master's code; PR touches zero notices files)
win22-2 abort x proxy-openai-sse windows-node22 load-starvation flake (documented family, #4599 floors 1287/1296) Flake
win22-4 storage tests ×9 (#5506 metadata-fold, BILI_PERSIST_ZSTD, boot-legacy) EBUSY: resource busy or locked, unlink '...index.db' — Windows file lock during test temp-dir cleanup (tests/tmp-rm.ts), ~10s timeout Windows file-lock flake

Verification (this run):

Action taken (per CI-coordination policy 2026-10-24):

8问:

  1. 是什么问题: PR [PR] fix: preflight-tier batch — deterministic tier-exhaustion verdict, repacked-tail reanchor, anthropic egress clamp #6435 同步 master 后(head a8b5fa6)5 个 test 分片红。
  2. 什么原因导致的: 三类,均非本 PR 或合并引入:① [PR] storage: pure-SQLite default — JSON envelopes stop being written unless dual-write is requested (#6479) #6480(pure-sqlite-default)新增的 storage-sqlite-default 公告测试是环境依赖的 hermetic 测试——未 pin 存储 env 变量,CI 环境 storage 解析为 full 时公告不显示(3 分片);② windows-node22 负载饥饿 flake(abort x proxy-openai-sse,1 分片);③ Windows 文件锁 flake(EBUSY unlink index.db,存储测试清理阶段,1 分片)。
  3. 影响范围与严重程度: 仅 CI 信号,无用户可见影响。① 为 [PR] storage: pure-SQLite default — JSON envelopes stop being written unless dual-write is requested (#6479) #6480 落 master 后的 born-red,所有 rebase 过 [PR] storage: pure-SQLite default — JSON envelopes stop being written unless dual-write is requested (#6479) #6480 的 PR 都会红(舰队级);②③ 瞬态 flake。
  4. 怎么修复的: 本轮未改代码。① 按 CI 协调政策归为 master 级 born-red,不在本分支修(须收敛到 ci-fixes),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 报精确机制请哨兵收敛(让 visibleNotices 测试 hermetic);②③ flake 无需代码修复,重跑即绿。
  5. 是否有回归问题: 无。本 PR diff 与 notices 文件集零交集;本地 clean env 5/5 绿证明非代码回归;合并树 typecheck 上一轮已验通过。
  6. 是否破坏缓存: 不涉及。
  7. 是否改了设计方向: 不涉及。
  8. 是否引入了新的配置: 不涉及。

判定: 本轮无 in-branch 修复——5 红 = 1 个继承 born-red(#6480,已在 #4599 留档待哨兵收敛 ci-fixes)+ 2 个 flake,均非本 PR 引入。PR 将在哨兵于 ci-fixes 收敛 visibleNotices 修复后 merge ci-fixes 即绿;flake 分片重跑即绿。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

Final verdict — all 5 red shards on head a8b5fa6627 investigated; none are caused by this PR or the master-sync merge.

Action taken (per CI-coordination policy 2026-10-24):

Outcome / what remains: the PR turns green once (a) the sentinel converges the visibleNotices hermetic-test fix on ci-fixes and I merge ci-fixes, and (b) the 2 flake shards clear on rerun. Risk: none functional — no user-visible impact, no regression from this PR. Full 8问 breakdown in my previous comment.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

CI fix run — head c9d9ce7178, the 2 red shards (win22-4, win22-2)

Re-verified live first: mirror cache was stale (0 failed / 9 pending, no test shards listed), so I confirmed via the job API — both jobs are completed / failure on run 43511333274. Not green.

What I found: both shards are the documented #4347 Windows teardown-lock (EBUSY file-lock) flake — not a code regression from this PR.

Evidence:

What I did: triggered a fresh-runner rerun of the 2 failed jobs (POST /actions/runs/43511333274/rerun-failed-jobs → HTTP 201). No code change (flake, not a bug), no push.

Verdict: PR code is clean; the 2 red shards are the documented #4347 Windows file-lock flake and should clear on the fresh-runner rerun. If the lock reproduces on a fresh runner too, that confirms the #4347 handle-leak hypothesis (a separate documented issue, not this PR's scope).


结构化八问

  1. 是什么问题:CI 两个 windows-node22 测试分片(win22-4、win22-2)红,报 EBUSY 文件锁失败。
  2. 什么原因导致的:测试临时目录清理阶段(SQLite index.db 句柄未释放)在 Windows 上 unlink 被拒(EBUSY),属 windows-22-1 lane randomly red: during test cleanup, unlink index.db hit an EBUSY external lock, couldn't survive tmp-rm | windows-22-1 车道随机红:测试清理期 unlink index.db 遇 EBUSY 外部锁,熬不过 tmp-rm 的 9.6 秒重试梯子(#2799/#2908 族复发) #4347 已归档的 Windows teardown 文件锁 flake 家族;CI 自带同 runner 重试已复现,CI 自身诊断也指向 windows-22-1 lane randomly red: during test cleanup, unlink index.db hit an EBUSY external lock, couldn't survive tmp-rm | windows-22-1 车道随机红:测试清理期 unlink index.db 遇 EBUSY 外部锁,熬不过 tmp-rm 的 9.6 秒重试梯子(#2799/#2908 族复发) #4347。非本 PR 代码引入——PR diff 完全不触碰 persist/storage/tmp-rm 域。
  3. 影响范围与严重程度:仅 CI 测试分片(特定 Windows runner 的 AV/索引器外部锁或句柄泄漏),不影响产品行为、不影响用户;属 CI 噪声,非可用性/缓存泄露类。
  4. 怎么修复的:无代码可修(flake 非 bug)。触发失败 job 的 fresh-runner 重跑(POST rerun-failed-jobs,HTTP 201),换干净 runner 规避该 runner 的外部锁。
  5. 是否有回归问题:无。PR 改动域(preflight/server/session/budget/prepare-anthropic/session-resolve)与失败域(persist 存储测试清理)零交集;既有行为无漂移,无 fail-fast 变吞错/阈值/时序/输出格式变化。
  6. 是否破坏缓存:不涉及。
  7. 是否改了设计方向:不涉及。
  8. 是否引入了新的配置:不涉及。

判定:PR 代码干净,2 红分片为 #4347 已归档 flake;fresh-runner 重跑(已触发)后应转绿。若新 runner 仍复现则坐实 #4347 句柄泄漏假设,另案跟踪,非本 PR 范围。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

CI fix run — head c9d9ce7178, the 2 red shards (win22-4, win22-2) — 2nd wake

Re-verified live: mirror cache (00:58Z) shows 2 failed / 0 pending / 25 success. The 2 failing jobs (132263633513, 132263633326) are the fresh-runner rerun I triggered last turn — they completed on fresh runners (1000339018, 1000339014) and failed again with the identical signature.

Attribution (unchanged, now reproduced on a fresh runner): both shards are the documented #4347 Windows teardown EBUSY family — the same 13 tests as last turn (win22-4: 10× persist-zstd + persist-metadata; win22-2: 3× #5506 metadata-fold), all EBUSY: resource busy or locked, unlink ...\index.db through tests/tmp-rm.ts:30 defaultRm → rmrf. The CI's own diagnostic: "if the lock reproduces there too [on a fresh runner], the leaked-handle hypothesis is primary (#4347)."

This is a master-level issue, not this PR's:

What I did / did NOT do:

Verdict: PR code is clean; the 2 red shards are the documented #4347 master-level EBUSY family. The PR goes green when the sentinel converges the EBUSY fix into ci-fixes (→ master) — or if the AV/indexer lock clears on a future rerun. No action is available on this branch.


结构化八问

  1. 是什么问题:PR [PR] fix: preflight-tier batch — deterministic tier-exhaustion verdict, repacked-tail reanchor, anthropic egress clamp #6435 的 CI 两个 windows-node22 分片(win22-4、win22-2)红,报 EBUSY 文件锁;这是上一次唤醒后触发的 fresh-runner 重跑,在干净 runner 上再次复现。
  2. 什么原因导致的:windows-22-1 lane randomly red: during test cleanup, unlink index.db hit an EBUSY external lock, couldn't survive tmp-rm | windows-22-1 车道随机红:测试清理期 unlink index.db 遇 EBUSY 外部锁,熬不过 tmp-rm 的 9.6 秒重试梯子(#2799/#2908 族复发) #4347/Windows CI: storage-engine-proofs 套件清理 temp 目录遇 EBUSY/EPERM,windows test leg 随机红 #2799/test(windows): driveWire golden-harness 清理期 index.db EBUSY 复发(#2799 族,#2808 后残留) #2908 家族的 Windows teardown 文件锁——测试清理阶段 unlink ...\index.db 被拒。两种在档假设:① 外部 AV/索引器对刚写文件的持有超出 tmp-rm.ts ~9.5s 重试阶梯(floor 1352);② 测试自身未关 SQLite 句柄,进程自己的 open handle 锁住文件(floor 1327 owner 根因,[PR] storage: pure-SQLite default — JSON envelopes stop being written unless dual-write is requested (#6479) #6480 pure-sqlite 翻转后 index.db 成为默认路径)。fresh-runner 复现支持「非单 runner 瞬态锁」,但哨兵仍在做 flake-vs-real 最终判定。
  3. 影响范围与严重程度:master 级——所有基于 master 的 PR 的 win22 存储测试分片都受影响(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 在档 ~15 个 PR 同签名);不影响产品行为、不影响用户;属 CI 噪声,非可用性/缓存泄露类。
  4. 怎么修复的:本分支无可修(失败文件 tests/persist-zstd / persist-metadata-fold-gate / tmp-rm 与 master 逐字节一致,PR 零触碰)。按政策 master 级修复收敛 ci-fixes(扩 tmp-rm.ts 重试窗口 和/或 测试 finally 关 store),由哨兵收敛;PR 已含 ci-fixes(c6369fb),再 merge 为空操作。未触发第 3 次重跑(fresh-runner 已复现,即哨兵所需数据点,再跑是噪声)。
  5. 是否有回归问题:无。PR 改动域(preflight/server/session/budget/prepare-anthropic/session-resolve)与失败域(persist 存储测试清理)零交集;既有行为无漂移,无 fail-fast 变吞错/阈值/时序/输出格式变化。
  6. 是否破坏缓存:不涉及。
  7. 是否改了设计方向:不涉及。
  8. 是否引入了新的配置:不涉及。

判定:PR 代码干净,2 红分片为 #4347 master 级 EBUSY 家族(已在 #4599 在档、哨兵跟踪 flake-vs-real);本分支无可用动作,等哨兵把 EBUSY 修收敛进 ci-fixes(→ master)后重跑即绿,或 AV/索引器锁在后续重跑中自行清除。

…olve

# Conflicts:
#	src/server.ts
#	src/server/prepare-anthropic.ts
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

CI fix run — merge-conflict resolution (attempt 1/6)

Re-verified live before acting: mergeable: False, mergeable_state: dirty. The PR (head c9d9ce7178) genuinely conflicted with master, which had moved to f648217cb8 (Merge #6823) past the PR's recorded base e1f31ad9d5.

What I did: merged origin/master into the PR branch. Exactly 2 conflicts, both union-of-intent (the two sides changed adjacent, semantically-orthogonal spots):

  1. src/server.ts:58-62 — the session.js import line. The PR added applyPostRebuildBaselineAnchor; master added type NudgeStamps. Resolved as the union (both symbols). Verified each is actually used in the merged file: applyPostRebuildBaselineAnchor at src/server.ts:4381, NudgeStamps at src/server.ts:2067 — no dangling imports.
  2. src/server/prepare-anthropic.ts:696-708 — the prepareAnthropic return. The PR inserted the #5880 clampOutgoingOutput(...) egress-clamp call before the return; master added the nudgeStampsBefore: nudgeOutcome === "injected" ? nudgeStampsBefore : undefined field to the return. Resolved as the union (kept the clamp call + the nudgeStampsBefore field). Verified clampOutgoingOutput is imported (src/server/prepare-anthropic.ts:42) and nudgeStampsBefore is defined (:176).

Verification: npm run typecheck green (exit 0) on the merged tree. No conflict markers remained.

Push: non-force fast-forward c9d9ce7178..b309b277d3. Before pushing I re-fetched and confirmed the remote head was still c9d9ce7178 (no concurrent push) and that it was an ancestor of my merge commit — so a clean fast-forward, no force.

Result (live): mergeable: True, mergeable_state: blocked (no longer dirty — "blocked" just means mergeable but waiting on CI). CI is freshly running on the new head b309b277d3: 14 in progress, 0 failed, 1 success, 15 skipped.

Note on the prior two wakes: the #4347 Windows EBUSY red shards (win22-4 / win22-2) are a separate master-level issue the sentinel is tracking for ci-fixes convergence — unrelated to this conflict. This run touched only the two conflict lines in src/server.ts and src/server/prepare-anthropic.ts; it did not touch any persist/test file.


结构化八问

  1. 是什么问题:PR [PR] fix: preflight-tier batch — deterministic tier-exhaustion verdict, repacked-tail reanchor, anthropic egress clamp #6435 与 master 出现合并冲突(merge state dirty)——master 前进到 f648217(Merge [PR] Gate completion: restart lane to add Anthropic line (Claude) + fix double-execution bug on P lane | gate 补全:重启车道加 anthropic 线(claude)+ 修 P 车道双重执行 bug #6823)后,与 PR 头 c9d9ce7 在两个文件上冲突。
  2. 什么原因导致的:PR 与 master 各自独立改了同一文件的相邻位置——① src/server.ts 的 session.js import 行(PR 加 applyPostRebuildBaselineAnchor,master 加 type NudgeStamps);② src/server/prepare-anthropic.ts 的 return 语句(PR 在 return 前加 Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880 egress clamp 调用,master 在 return 里加 nudgeStampsBefore 字段)。两边语义正交,无设计冲突。
  3. 影响范围与严重程度:仅 PR 分支合并受阻(无法 merge),不影响产品行为、不影响用户;属合并流程阻塞,非可用性/缓存泄露类。
  4. 怎么修复的:merge origin/master 进 PR 分支,两处冲突均按「意图并集」解(保留双方新增的符号/调用/字段),逐一验证每个保留符号在合并后文件里确实被使用/已定义(避免悬空引用),typecheck 绿后非强推(fast-forward)推送。
  5. 是否有回归问题:无。两处都是纯并集(不删任何一方的改动),typecheck 通过;PR 原有功能(Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880 clamp、applyPostRebuildBaselineAnchor)与 master 新增(NudgeStamps、nudgeStampsBefore)都完整保留,无语义漂移、无 fail-fast 变吞错、无阈值/时序/输出格式变化。
  6. 是否破坏缓存:不涉及。
  7. 是否改了设计方向:不涉及。
  8. 是否引入了新的配置:不涉及。

判定:合并冲突已解(mergeable True,blocked 等 CI),新头 b309b27 CI 运行中(14 进行中 / 0 失败)。本次只动了两个冲突行,未触碰此前 #4347 EBUSY 涉及的 persist/test 文件。

@ranxianglei ranxianglei left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

APPROVE — clean batch of three root-cause fixes, all verified against the code, tests, and CI.

What I checked:

  • #5764 / tier-exhaustion deterministic verdict (src/preflight.ts, src/server.ts): the new stopReason (fit/exhausted/cap/unavailable/no-shrink) is set on every exit path of the tier-recovery loop, and the fail-fast emits 413 only when stopReason === "exhausted" AND the session is not overflow-armed. Cap/unavailable/no-shrink stay 502-retryable; overflow-armed stays 502-retryable (the #6451 guard, commit 9d18c77a55, is present and correct). Verified the 413 path at src/server.ts:4674.
  • #5844 / repacked-tail reanchor (src/session-resolve.ts): the new canonical per-turn fingerprint (sha256 of wire-neutral turns) is a second tail-verification rung — when the fragment tail doesn't match but the canonical tail does, the session re-anchors (the logical turns survived a host record repack). Old skeletons without a canon column degrade to pre-change behavior (fork); a mutated tail cell still forks (guardrail test). The orderSnapshot/applyOrderSnapshot round-trip validates the canon column (length must equal kinds length, every entry a 64-hex sha256) and degrades a malformed column to absent → fork.
  • #5880 / anthropic egress clamp (src/server/prepare-anthropic.ts, src/server/handle.ts, src/session.ts): the anthropic lane now clamps outgoing max_tokens against the RAW native window via clampOutgoingOutput, exactly matching the openai lane (same field, same nativeWindow/headroomWindow/imageTokens/kFactor/ledgerFloorTokens/knownOutputCeiling wiring, same lastOutputClamp stamp). The post-rebuild baseline anchor write (applyPostRebuildBaselineAnchor) now guards the overflow-arm evidence tag (returns early when lastInputTokensSource === "overflow-arm"), preventing the livelock.
  • Tests: all three new suites are genuine with teeth (14 tests, all pass): the #5764 suite has a negative gate (12×25000 blocks → 6-call cap → NOT deterministic); the #5844 suite pins the positive reanchor, the mutated-cell guardrail, the canon-less degradation, the persistence round-trip, the malformed-canon degradation, and the tag-invariant symmetry; the #5880 suite pins the armed-tag survival, the unclamped comfortable case, and an e2e near-wall anthropic request (560K chars → clamped below 60000, ≥1024, log line carries window=200000 = RAW).
  • CI: mirror cache shows 0 failed (in-progress test shards still running; e2e lanes skipped by design). No kernel/ files touched, no kernel-change declaration, no config-surface change.

Behavior changes (all disclosed in the PR body):

  1. Structurally-exhausted tier-recovery: 502 (retryable) → 413 (deterministic, non-retryable). Cap/unavailable/no-shrink and overflow-armed stay 502-retryable.
  2. New reanchor rung in resolveSessionOrder (tag-changed strictly-shorter replay with a content-verified tail → reanchor, not fork).
  3. Anthropic lane max_tokens now clamped against the RAW native window (previously no clamp); the overflow-arm evidence tag survives the post-rebuild baseline anchor write.

No regressions, no security/data-loss risks, no cache breakage, no design-direction change, no new config. House conventions followed (genuine tests, explanatory comments, no kernel/, no config surface).


8问(中文,供人工审核):

  1. 是什么问题:本 PR 修复 preflight/会话解析/anthropic 出站三条路径上的三个缺陷:①结构耗尽的 tier-recovery 返回 502(可重试)而非确定性的 413,导致客户端对永远不会放下的 payload 无限重试;②宿主重打包消息记录(严格更短的 replay、逻辑尾部完好)时,会话解析走 fork 而非 reanchor,破坏 persona 连续性;③anthropic 车道未对出站 max_tokens 按原始窗口做 clamp(openai 车道有),存在溢出风险,且 overflow-arm 证据标签被 post-rebuild 基线锚点写入覆盖(活锁风险)。
  2. 什么原因导致的:①tier-recovery 结果未区分「结构耗尽」与「瞬时」结局,fail-fast 无法发出 413;②会话解析只有基于 fragment 的尾部校验,在重打包 replay 下 fragment hash 不同但 wire-neutral 内容完好时失效;③anthropic 车道缺少 openai 车道已有的 clamp 调用,且基线锚点写入无条件覆盖 overflow-arm 标签。
  3. 影响范围与严重程度:①影响 payload 在 tier-recovery 后结构性放不下的客户端——此前无限重试 502 而非拿到确定性 413(p1,可用性);②影响宿主重打包消息记录的 persona 会话——此前 fork(丢连续性)而非 reanchor(p1);③影响接近上下文墙的 anthropic 车道——max_tokens 可能溢出窗口,且 overflow-arm 标签被覆盖(活锁风险)(p1)。
  4. 怎么修复的:①给 tier-recovery 结果加 stopReason(fit/exhausted/cap/unavailable/no-shrink),fail-fast 仅在 stopReason === "exhausted" 且非 overflow-armed 时发 413;②加一条基于 wire-neutral 逐轮 sha256 指纹的尾部校验 rung:fragment 尾部不匹配但 canonical 尾部匹配时 reanchor(逻辑轮次完好);③把 clampOutgoingOutput 接入 anthropic 车道(与 openai 车道一致,按原始 nativeWindow clamp),并让基线锚点写入在 overflow-arm 时提前返回、不覆盖证据标签。
  5. 是否有回归问题:无。①cap/unavailable/no-shrink 结局保持 502 可重试;overflow-armed 保持 502 可重试(#6451 守卫)。②无 canon 列的旧骨架退化到变更前行为(fork);尾部单元格被篡改仍会 fork(guardrail 测试)。③clamp 与 openai 车道一致;overflow-arm 守卫防止活锁。无安全/数据丢失风险。
  6. 是否破坏缓存:不。reanchor 保留父会话的块(是修复,不是缓存破坏者);clamp 与 413 不影响缓存。
  7. 是否改了设计方向:没有。这是 3 个定向修复的批量,无架构变更。
  8. 是否引入了新的配置:没有。无新配置面。

判定:APPROVE,可合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

Review complete — verdict: APPROVE (posted as COMMENT; GitHub forbids self-approval since the token belongs to the PR author's account).

Review: #6435 (review)

What was verified (head b309b27, 12 files +762/−29):

  1. [Bug][preflight] Six inline base64 images (>60K chars each) defeat tier-2 span selection → 0 summary calls → 502 fail-fast #5764 — deterministic tier-exhaustion verdict. stopReason is set on every code path in src/preflight.ts (exhausted/unavailable/no-shrink/fit/cap); deterministic flag at preflight.ts:3249 correctly gates on stopReason === "exhausted" && !overflowArmed; 413 emitted at server.ts:4674 only when deterministic && !payloadFitsWindow. The [Root Cause] preflight-tier (7 items) #6451 overflow-armed guard (commit 9d18c77) is present and correct — without it, overflow-armed sessions would wrongly get 413.

  2. Resuming a long session fails with preflight 413 when persona fingerprint and replay shape change together (0.1.212) #5844 — repacked-tail reanchor. New canon rung in src/session-resolve.ts (~line 513): tag-changed strictly-shorter replays re-anchor when canonTailEquals matches; degradation to fork when canon is absent/malformed; guardrail test exercises the canon path (not the frag path); persistence round-trip validated.

  3. Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880 — anthropic egress clamp. clampOutgoingOutput wired in src/server/prepare-anthropic.ts:703 with field max_tokens, raw nativeWindow, matching the openai lane exactly (prepare-openai.ts:583). Overflow-arm evidence tag survives post-rebuild baseline anchor write via applyPostRebuildBaselineAnchor guard in src/session.ts:1203.

Tests: 3 new test files, 14 tests, all genuine with teeth (positive + negative gate rows), all pass locally. CI green (0 failed, 8 shards in progress, e2e lanes skipped by design).

No kernel/ changes, no config surface, no security/data-loss risks, no regressions.

Behavior changes (disclosed in PR body):

  • Structurally-exhausted tier-recovery: 502-retryable → 413 deterministic non-retryable (overflow-armed stays 502).
  • New reanchor rung in resolveSessionOrder for tag-changed shorter replays.
  • Anthropic lane max_tokens now clamped against raw native window (previously no clamp).

八问审核清单:

  1. 是什么问题:三个独立缺陷的根因批修——tier 耗尽时返回可重试 502 而非确定性 413([Bug][preflight] Six inline base64 images (>60K chars each) defeat tier-2 span selection → 0 summary calls → 502 fail-fast #5764);tag 变化后严格变短的重放无法重新锚定导致 fork(Resuming a long session fails with preflight 413 when persona fingerprint and replay shape change together (0.1.212) #5844);anthropic 车道出站 max_tokens 无钳制,可越过上游窗口墙(Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880)。
  2. 什么原因导致的:[Bug][preflight] Six inline base64 images (>60K chars each) defeat tier-2 span selection → 0 summary calls → 502 fail-fast #5764:tier-recovery 失败路径未区分「结构性耗尽」与「暂时性不可用」,统一走 502 重试。Resuming a long session fails with preflight 413 when persona fingerprint and replay shape change together (0.1.212) #5844:reanchor 阶梯只认 frag 指纹,tag 变化后 frag 失配即 fork,缺少 canon 指纹兜底。Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880:anthropic 车道未接入 clampOutgoingOutput(openai 车道已有),max_tokens 原样透传。
  3. 影响范围与严重程度:[Bug][preflight] Six inline base64 images (>60K chars each) defeat tier-2 span selection → 0 summary calls → 502 fail-fast #5764 影响所有 tier-recovery 结构性耗尽的会话(p1,之前无限重试浪费配额);Resuming a long session fails with preflight 413 when persona fingerprint and replay shape change together (0.1.212) #5844 影响 tag 变化后重放的 persona 车道(p2,fork 后上下文分裂);Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880 影响所有 anthropic 车道近墙会话(p1,请求被上游 400 拒绝)。
  4. 怎么修复的:[Bug][preflight] Six inline base64 images (>60K chars each) defeat tier-2 span selection → 0 summary calls → 502 fail-fast #5764 在 tier-recovery 结果上增加 stopReason 枚举,结构性耗尽标 deterministic,server 层据此返回 413;overflow-armed 会话保持 502([Root Cause] preflight-tier (7 items) #6451 守卫)。Resuming a long session fails with preflight 413 when persona fingerprint and replay shape change together (0.1.212) #5844 在 reanchor 阶梯新增 canon 指纹 rung,tag 变化后仍可按 canon 尾对齐重新锚定;canon 缺失/畸形时降级 fork。Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880 anthropic 车道接入与 openai 车道完全一致的 clampOutgoingOutput,基于原始 nativeWindow 钳制 max_tokens;overflow-arm 证据标签在重建基线锚点写入时保留。
  5. 是否有回归问题:无。逐项核对了既有行为:openai 车道钳制逻辑未动;frag 路径 reanchor 未动;非 overflow-armed 的 tier 耗尽仍走 502;canon 缺失时降级 fork(与修复前行为一致)。无 fail-fast 变吞错、无阈值/时序/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:不涉及。三个修复都是补齐既有机制的缺失路径,未改变架构或数据流。
  8. 是否引入了新的配置:不涉及。

判定:APPROVE,可合并。

ci-check and others added 4 commits November 16, 2026 20:42
The probe anchor PROXY_TOOL_AT=6 was calibrated on 2026-10-31 for #4521
(schedule {4,6,8,9}). The nudge scheduling has shifted again — nudge now
lands at {3,5,11,14}. Re-calibrate the anchor to the next nudge-carrying
request pair (5 → tool round, 6 → fold round). The fold/proxy-tool-round
mechanics under test are unchanged.

Closes: master-red on test (windows-latest, 22, 3) and (ubuntu-latest, 22, 1).
…olve

# Conflicts:
#	src/server/prepare-anthropic.ts
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

Conflict resolved and pushed. PR #6435 is now mergeable (no conflicts).

What I found: The PR head b309b277d3 had merged an old master (f648217cb8), but master had since moved ahead to 301e1a5459 (~60 new commits: nudge-debug, gate knobs, recovery chart, pi structured-delegate, responses-section-helper, …). That drift is why the PR was still dirty.

What I did this run:

  1. Discarded a working-tree pi-subagents/package-lock.json change that was local npm install noise (it stripped libc fields master keeps; the PR doesn't touch pi-subagents/ at all).
  2. Merged current origin/master (301e1a5459) into the PR branch.
  3. Resolved the single conflict in src/server/prepare-anthropic.ts: the branch added the Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880 egress clampOutgoingOutput(...) call and master added recordNudgeInject(...) at the same spot — kept both, in the conventional order (clamp first, then recordNudgeInject), matching the openai/responses/google lanes.
  4. Verified: npm run typecheck green (exit 0); the PR's three targeted test files pass 14/14; post-merge PR diff is clean (exactly the 12 intended files, no merge noise); pi-subagents version consistent (package.json 0.2.4 == package-lock.json 0.2.4 — the merge brought master's lockfile, which also fixed a pre-existing 0.2.2 drift on the branch).
  5. Pushed final head f4e289dc93 (no [ework:ci-held] marker — this is the branch's one full CI pass).

Live state (verified this run): mergeable: true, mergeable_state: blocked — "blocked" is branch-protection waiting on human review/approval (human-only), not a conflict. CI triggered on the new head: 30 checks, 0 failed so far, 19 still queued/running (I can't assert final green until it finishes).


结构化八问

  1. 是什么问题:PR [PR] fix: preflight-tier batch — deterministic tier-exhaustion verdict, repacked-tail reanchor, anthropic egress clamp #6435 与 base 分支 (master) 合并冲突(merge state: dirty),PR 无法推进合并。
  2. 什么原因导致的:PR head 之前合并的是旧 master (f648217cb8),而 master 已前进到 301e1a5459(约 60 个新提交);两边在 src/server/prepare-anthropic.ts 同一位置各自新增了不同代码(分支加 Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880 egress clamp,master 加 recordNudgeInject),产生文本冲突。
  3. 影响范围与严重程度:仅阻塞该 PR 合并,不影响任何已发布运行时行为。严重度=PR 推进被卡(非运行时缺陷、无用户面影响)。
  4. 怎么修复的:把当前 origin/master 合并进 PR 分支;唯一冲突点保留双方新增(Estimate-only fit decision ignores system prompt + tool definitions, and the output-budget clamp never fires → 400 CONTEXT_WINDOW_EXCEEDED on sessions that bili thinks fit #5880 clampOutgoingOutput + master 的 recordNudgeInject),按 openai/responses/google 车道惯例排序(clamp 在前);丢弃工作区一处 pi-subagents lockfile 的本地 npm 噪音(PR 本不碰 pi-subagents)。
  5. 是否有回归问题:无。typecheck 绿 (exit 0);PR 三个定向测试文件 14/14 通过;合并后 PR diff 干净(恰好 12 个文件,无合并噪音);冲突解决是"保留双方新增",未改动任何既有语义。高危模式逐项过:无语义漂移、无 fail-fast 变吞错、无阈值/时序/输出格式变化。
  6. 是否破坏缓存:不涉及。纯合并冲突解决,未触碰缓存/重试/增量游标/去重键逻辑。
  7. 是否改了设计方向:不涉及。未改架构/边界/数据流。
  8. 是否引入了新的配置:不涉及。无新 env/config/默认值。

收工判定:冲突已解决,PR 现为 mergeable(无冲突),blocked 仅等人工 review/approval(human-only,我不合并)。CI 已在新 head f4e289dc93 触发,目前 0 failed、19 项在跑,最终绿/红以远程 CI 为准。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

Stale trigger — no action needed. The conflict was already resolved in the prior run; verified live this run.

I re-verified against live sources (not memory) before touching anything:

  • PR head f4e289dc93 already contains the current master 301e1a5459 as an ancestor → a conflict is structurally impossible right now.
  • Live PR state: mergeable: true, mergeable_state: blocked, state: open. "blocked" = branch-protection waiting on human review/approval (human-only) — not a merge conflict.
  • Head commit carries no [ework:ci-held] marker (it is the final head from the prior run's full CI pass).
  • CI (live, mirror cache): 30 checks, 0 failed, 11 still pending (all completed so far are success: static-gates ubuntu/windows, mysql-contract, plan).

So the "merge state: dirty" in this trigger predates the prior run's push. Re-merging would be a no-op ("Already up to date") and only add a noise commit, so I deliberately did not push anything. Nothing to fix, no new issue, no decision needed.


结构化八问

  1. 是什么问题:本次触发报「PR 与 base 冲突 (dirty)」,实为过期触发——冲突已在上一次运行解决。
  2. 什么原因导致的:触发检测早于上一次 push 生效;PR head 已合并当前 master,冲突不复存在。
  3. 影响范围与严重程度:无实际影响。PR 现可合并(无冲突),blocked 仅等人工 review。
  4. 怎么修复的:无需修复。本次仅做 live 核验(head 已含当前 master、mergeable=true、CI 0 failed),未做任何 git 变更以避免噪音 commit。
  5. 是否有回归问题:不涉及(无代码改动)。
  6. 是否破坏缓存:不涉及。
  7. 是否改了设计方向:不涉及。
  8. 是否引入了新的配置:不涉及。

判定:无冲突、无需动作;PR 保持 mergeable,等人工 review/approval(human-only,我不合并)。CI 在最终 head f4e289dc93 上运行中,目前 0 failed。

ework-agent and others added 2 commits November 17, 2026 01:36
The probe's hardcoded request-index anchors (PROXY_TOOL_AT/FOLD_AT) broke
every time the nudge gate's schedule shifted. The last recalibration
(245ecb1) was measured on a direct local run (timestamps ON by default
since #4313), but CI runs the suite preload pin BILI_MESSAGE_TIMESTAMPS="0"
(timestamps OFF) — a different wire shape, a different schedule (no nudge at
index 5) — so it passed locally and failed on every CI shard.

Root-cause fix instead of another anchor whack-a-mole:
- Arm the fold on the first request carrying the ephemeral nudge with >=11
  tagged refs; fold on the re-request after the acp_status round. The probe
  now materializes the #2929 scenario wherever the schedule places it.
- Pin compress.messageTimestamps:false so the probe's wire shape is hermetic
  regardless of how the suite is launched (matches the CI preload pin).

Supersedes the 245ecb1 recalibration. Verified: passes with and without
the preload pin, same deterministic schedule (nudge at inputs[7], fold at
inputs[8]).
…dule-independence + architecture-docs map)
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

Red-CI resolved — inherited master failure, fixed by merging ci-fixes.

The 3 failing shards (test windows-latest 22,4 / ubuntu-latest 22,2 / ubuntu-latest 24,2, run 43903497490) all failed on the same single test: #2929: nudge + proxy-tool-round fold keeps re-request/next-turn byte-stable (tests/issue2929-codex-fold-double-miss.test.ts:232), error scenario miss: no ephemeral nudge in the pre-fold request (inputs[6]). Each shard ran the full suite and failed exactly 1 test (e.g. ubuntu 24,2: 3710 tests / 3691 pass / 1 fail).

This is NOT caused by this PR's code. It is an inherited master-level born-red (confirmed in #4599 floor 1489: master @ 301e1a5459 is still red on #2929). Root cause: the #2929 probe uses hardcoded request-index anchors (PROXY_TOOL_AT=6) calibrated for an older nudge schedule; the nudge schedule shifted again (the #4521 status-line removal in #6928 changed it), so the anchor inputs[6] no longer lands on a nudge-carrying request. The CI sentinel already converged the root-cause fix onto the standing ci-fixes branch: 7247410976 (schedule-independent trigger + hermetic messageTimestamps:false) + 245ecb1a9e (anchor recalibration) + 8ed56be997 (architecture-docs map line).

Action taken (per CI-failure coordination policy #4599): git merge origin/ci-fixes into the PR branch → final head 53fea86fe3 (no [ework:ci-held] marker = this branch's one full CI pass). Verified locally before pushing: typecheck clean, the #2929 test now passes 1/1, and this PR's own 3 targeted tests still pass 14/14 (no regression). Did not re-fix the test on this branch — the fix stays converged in ci-fixes (sentinel-maintained); the merge only brought in 2 files (reference/architecture.md +1 line, the #2929 test file), zero src/ runtime changes.

Live PR state after push: mergeable=true, mergeable_state=blocked (waiting on human review, not a conflict).

八问:

  1. 是什么问题:PR 分支 CI 3 个 test 分片红,全部挂在同一条测试 #2929 fold probe(断言 inputs[6] 应有 ephemeral nudge,实际没有)。
  2. 什么原因导致的:继承性 master 红,不是本 PR 代码问题。codex 命中率问题分析 #2929 probe 用硬编码请求下标锚点,而 nudge 调度又漂移了(Per-turn '[ACP status: …]' line keeps ordering the model to fold leftover small ranges right after a big fold — the opposite of what the compress receipt says #6928 移除 Compression conflict banner (bili v0.1.205, node v24.19.0, win32-x64) #4521 状态行后调度改变),锚点 inputs[6] 不再落在带 nudge 的请求上;master @ 301e1a5 自身即红。
  3. 影响范围与严重程度:仅 CI 信号级(所有 rebase 到 master 的 PR 都继承这条红),非用户侧功能缺陷;本 PR 的实际改动(preflight-tier 12 文件)不受影响。
  4. 怎么修复的:按 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 CI 协调政策 git merge origin/ci-fixes 继承哨兵已收敛的根治(probe 改调度无关触发 + 钉 messageTimestamps:false 使线形 hermetic + architecture-docs map 补行),不在本分支重复修。
  5. 是否有回归问题:无。merge 只带入 2 个文件改动(reference/architecture.md 1 行 + tests/issue2929 测试文件),不触碰本 PR 任何 src/ 运行时;本 PR 自身 3 条定向测试 14/14 仍绿。
  6. 是否破坏缓存:不涉及。
  7. 是否改了设计方向:不涉及(纯测试/文档修复,无 src/ 运行时变更)。
  8. 是否引入了新的配置:不涉及。

判定:final head 53fea86fe3 已推送(无 ci-held 标记,本分支唯一一次全量 CI)。等 CI 重跑转绿后,本 PR 即可进入人审(mergeable_state=blocked 是等人审,非冲突)。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

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

Stale trigger — no action needed. This event references the old head f4e289dc93 (run 43903497490), which the previous wake already replaced.

Live state re-verified this run:

  • PR branch head is now 53fea86fe3 (local == remote), no [ework:ci-held] marker.
  • The 3 reported #2929 shard failures are already resolved: the previous run merged origin/ci-fixes (sentinel's root-cause fix 7247410976/245ecb1a9e/8ed56be997) into the branch, and I verified locally that the codex 命中率问题分析 #2929 test now passes 1/1.
  • A fresh CI run is in progress on 53fea86fe3: 30 checks, 0 failed, 21 still pending (gate in_progress, rest queued).

Nothing to fix on this head. Stopping without further work; if the in-progress run surfaces a new red, the next wake will pick it up.

@ranxianglei ranxianglei left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

Re-review at new head 53fea86 — verdict unchanged: APPROVE.

The new head adds only origin/master + origin/ci-fixes merges (commit 53fea86). The PR's own 12 files are byte-identical to the previously reviewed head b309b27 — verified via git diff b309b277d..53fea86fe -- <12 PR files> (empty). The only delta in the 2 overlapping files (handle.ts, prepare-anthropic.ts) is master-side code (#6880 declared-purpose, #4521 status-line removal, nudge-debug) that merges cleanly around the PR's changes.

CI: 0 failures (2 success, 9 in-progress, 4 queued, 15 skipped-by-design). No kernel/ changes, no config surface, no regressions. Full 8问 checklist in the prior review (pullrequestreview-5912534560) applies unchanged.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 (ai: vllm-qwen/qwen3.8-27b)

Re-review at new head 53fea86 — verdict unchanged: APPROVE.

New head is old head + origin/master/origin/ci-fixes merges only. PR's 12 files byte-identical (verified via diff). Master-side delta in handle.ts/prepare-anthropic.ts (#6880, #4521, nudge-debug) merges cleanly around PR changes. CI: 0 failures. No kernel/, no config surface, no regressions.

Re-review: #6435 (review)

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