Repository navigation
[PR] EN | acp_status: stop recommending BLOCKED sentinel range endpoints (#6168) - #6305
ranxianglei wants to merge 5 commits into
Conversation
…nt (#6168) kernel-change: acp_status' uncompressed range view used each segment group's first/last entry verbatim as candidate endpoints; policy-protected messages carry the BLOCKED sentinel ref, so groups containing them rendered rows like 'BLOCKED-m02518' that could never be submitted to compress and were uniformly flagged 'refs unresolvable'. The recommend side already screens BLOCKED out at base-set level (buildCompressibleRanges); the report view violated the documented identical-segmentation contract. Fix is display-layer only: every BLOCKED entry in a group moves to an aggregate protected stats line under the header (totals unchanged), candidates span the first..last normal members of the group positionally (no ref arithmetic, non-monotonic refs kept), all-BLOCKED groups emit no candidate row, and flagFor now shows the real projection error (cleaned, capped) instead of the flat misdiagnosis. No change to apply/recommend/segmentation semantics. Behavior changes (disclosed): (1) acp_status uncompressed range rows can no longer start/end at BLOCKED; protected content now appears once in a 'protected: N msgs, T (BLOCKED refs - not submittable to compress)' line plus the per-message drilldown. (2) A group whose members are all protected renders '(no candidate ranges - every visible message is protected)' instead of an unresolvable row. (3) The foldability warning for genuinely failing projections prints the underlying boundary/projection error truncated to 100 chars instead of 'refs unresolvable'. Header totals and per-message counts are byte-identical to before for the same session state. Fixes #6168
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.11011-pr.6305.656721dThis exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr6305.tgz
npm install -g packageInstall instructions are refreshed on each push. |
[bot] (ai: vllm-qwen/qwen3.8-27b) 进度汇报(进程曾被强制重启,恢复后已完成冲突处理): PR #6305 冲突已解决并推送,等待 CI。
|
|
该PR已按根因归类纳入 [dsh-lane] 根因汇总 #6294,由汇总单统一跟踪处理;本单不再单独跟进,故关闭。如认为归类有误或问题仍未解决,请直接回复,会重新打开。 |
|
状态更正(自动化脚本误操作): 本单此前被「纳入根因聚合」流程误关闭, 现重新打开。机器人工作已停, 本单由根因聚合线程统一跟踪, 根因修复落地后请手动关闭或在本单回复「close」。 |
|
重新打开:此前被归类聚合脚本误关。按 owner 口径,纳入根因聚合的 issue 只停机器人、不关闭,由根因聚合线程统一跟踪,根因修复落地后统一关闭。 |
|
(更正)此前按"清理"口径误关闭;按 owner 口径纠正:被根因聚合纳入的 issue 只停机器人、不关闭。本 issue 重新打开,机器人工作已停。 |
[bot] (ai: vllm-qwen/qwen3.8-27b) 收工汇报(本会话曾被强制重启两次——原因是在会话内长时间等待 CI;已改为「推送后立即汇报、不再会话内守候」,CI 若红会由系统再次唤醒): PR #6305 冲突已清除,当前 head 八问:
透明披露:第二次推送( 判定:CI 工程师部分收工——冲突已清、无红项遗留,等新 head 的 CI 全绿 + owner 审批即可合并。 |
[bot] 🏷 查重: 不重复——镜像 [PR] 行已查,无针对本线程的并行修复 PR。分支上已含另一位 worker 推上去的 master 同步(4e39b71),我独立复核而非另开 PR;本地独立重做的合并结果与远端 head 逐字节相同( 代码审核结论:修法正确,纯展示层。 根因在源码层面证实:base 的 验证证据(均在远端 head 4e39b71 的同一棵树上):
CI: 当前 head 4e39b71 上 30 项 check 运行中,0 失败、14 pending(镜像 09:02Z 抓取)。原 head 9c73677 上零 check 的异常已随新推送自行消解,无需在 #4599 记录。 线程状态: PR OPEN,mergeable_state=blocked(checks 进行中 + kernel 变更属 human-only 审核档,等人工 review/merge)。本单已纳入根因聚合 #6294,按 owner 口径(2026-11-12)聚合单只停机器人、不关闭——我不关闭,也不再推送,等人工合并。 八问
收工判定:代码正确、分支状态正确(已验证)、CI 运行中无失败;待人工 review/merge(kernel 变更 human-only 档),机器人侧无待办。 |
[bot] 🏷 两条均为已确认批次的重复投递(06:26Z/06:29Z),无新信息、无状态变化,不重复核验;结论以本线程最近两条 bot 确认为准,继续等人工 review/merge。 |
[bot] 🏷 06:32Z 一条,为该重复系列(06:15Z–06:32Z)最后一项,无新信息;结论不变,等人工 review/merge。 |
[bot] 🏷 重复投递仍在继续(新时间戳 06:38Z/06:42Z,内容同前),无新信息、无状态变化。若持续出现,建议查一下上游同步侧的转发动作;本线程结论不变,等人工 review/merge。 |
[bot] 🏷 同文重复投递(06:42Z),无新信息;结论以本线程 bot 确认为准,等人工 review/merge。 |
[bot] 🏷 这条与前面数条是同一条自动化更正通知的重复投递(带 真正需要处理的是同步侧重复转发这一条——它每转一次就唤醒我一次,我只能逐条回确认,等于在本线程持续刷噪音。我无法修平台同步逻辑,请你在同步侧停掉这条的循环转发(或确认要我继续逐条确认)。PR 本身结论不变,等人工 review/merge。 |
[bot] 🏷 同文重复投递仍在继续,无新信息;已请你在同步侧停掉循环转发,结论不变,等人工 review/merge。 |
[bot] (ai: vllm-qwen/qwen3.8-27b) 冲突已再次清除并推送。本轮做的事:master 在 11-12 之后又前进了 300 个提交(含 batch4 kernel 修复与 #6709),PR 分支与基线再次失步(系统报 dirty;本次 live 核验时镜像态为 behind,两种状态同一处置——合入基线)。 八问:
判定:新 head |
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Findings-only review — no approval (kernel gate). No blocking findings; the fix is verified correct.
This PR touches kernel/ and carries a kernel-change: declaration → human-only review tier; I post findings only, and the merge decision waits on the owner (see the 🔶 需要决策 thread comment).
Direction (issue #6168): confirmed. The repro (BLOCKED–m02518 ⚠ not foldable: refs unresolvable) is real and the root cause is verified at source: renderUncompressedRanges (kernel/src/report.ts) took each segmentGroups group's first/last entry verbatim as candidate endpoints, while collectVisible correctly keeps BLOCKED entries in visible for the totals; parseBoundary (kernel/src/boundaries.ts) returns null for "BLOCKED", so resolveBoundaries threw and the old catch flattened it to "refs unresolvable". The recommend side already screens BLOCKED out at base-set level (kernel/src/recommend.ts:333) — the report view violated the documented identical-segmentation contract. Fixing the display layer (not the apply path) is the right layer: apply was never affected, matching the reporter's observation that valid ranges compress fine.
Correctness (verified against the head tree, not just the diff):
- Candidates now span the first..last NORMAL member positionally (no ref arithmetic; non-monotonic refs preserved via
orderedRefPair); all-BLOCKED groups emit no candidate row; every BLOCKED entry lands exactly once in theprotected:aggregate (invariant: candidate counts + protected = visible total). - Header totals and the per-message drilldown are byte-identical for the same state; sort/limit semantics untouched (the protected line is header-adjacent and limit-immune).
flagForendpoints are now always resolvable numeric refs, so a throw is a genuine boundary/projection failure;cleanProjectionError(whitespace-collapsed, 100-char cap, empty → "projection failed") surfaces the real reason — strictly more informative, no new failure mode.- Host-side report parsers are safe: the two new lines are 2-space-indented and match the existing continuation grammar in
src/marker/status-body.ts(/^ {2,}\S/) andsrc/acp-panel.ts(/^\s/); no new column-0 section start, so no parser list needs updating. - Regression sweep of existing behaviors on the touched path: header totals unchanged, compressed scope unchanged, recommend/apply/segmentation unchanged,
startIndex(size-sort tie-break only) now references the first normal member — no ordering change for distinct token counts. One pre-existing nuance, unchanged by this PR and out of scope: a BLOCKED user message still splits a group in the report view (it stays a segment item) while recommend's base set excludes it (coarser grouping) — both resulting candidates are submittable and the PR strictly improves the previously broken row. - Kernel boundary:
kernel-change:declaration present; version discipline satisfied — kernel 0.0.198 → 0.0.199 consistent acrosskernel/package.json,kernel/package-lock.json, and rootpackage-lock.jsonin the final tree (the PR body's "0.0.193 → 0.0.194" wording predates the final rebase); mechanical-bug exemption properly invoked (reproducing regression test + declaration).
Tests: 9 new tests in kernel/tests/blocked-status-ranges.test.ts map 1:1 to the issue's acceptance matrix (leading/trailing/mid-group/all-BLOCKED, non-monotonic refs, genuine resolution failure, sort=time+limit, drilldown, normal-only regression guard); they assert on real rendered output with teeth (line formats, the count invariant, the "refs unresolvable" negation). Local targeted runs on this head: new suite 9/9, related kernel suites (ranges-consistency, recommend-range-integrity, reversed-ranges, status-surface-meta, range-coverage) 34/34, host tests/acp-status.test.ts 9/9, kernel typecheck + build clean. Full-suite verdict = CI (mirror: 0 failed at last fetch; test shards in flight).
Behavior changes — all three disclosed in the PR body, all verified in code: (1) no BLOCKED endpoints — protected content moves to the aggregate line; (2) an all-protected group renders the explicit "no candidate ranges" note; (3) genuinely failing projections print the real error (100-char cap) instead of "refs unresolvable".
Verdict: no blocking findings. Merge is gated on the owner's human-tier approval for the kernel change (kernel boundary rule 3).
八问
- 是什么问题:
acp_status未压缩区间视图把策略保护消息的BLOCKED哨兵渲染成候选行端点(BLOCKED–m02518),并配一刀切的refs unresolvable误诊——模型拿到的是永远提交不了的假候选。 - 什么原因导致的:report 层
renderUncompressedRanges直接取 segment group 首/末条目当候选端点,缺 BLOCKED 守卫;recommend 层早在 base-set 级就筛掉 BLOCKED(recommend.ts:333),两视图违反 identical-segmentation 契约。纯 report 层守卫缺失,apply 路径不受影响。 - 影响范围与严重程度:仅
acp_status展示层——用户/模型看到假候选行与错误诊断;实际压缩行为不受影响。体验/诊断瑕疵级(p2 档)。 - 怎么修复的:展示层守卫——BLOCKED 条目收进 header 下聚合统计行(总量口径不变);候选行按位置跨 group 内首..末正常成员(无 ref 算术);全保护 group 渲染显式说明;
flagFor打印真实投影错误(空白折叠、100 字符截断)。不动 apply/recommend/segmentation。 - 是否有回归问题:无。逐项过:header 总量与逐消息计数不变;drilldown 不变;sort/limit 语义不变;宿主两个报告解析器(status-body.ts / acp-panel.ts)的既有续行文法覆盖新行;recommend/apply 未触碰;既有 report/ranges 套件全绿。
- 是否破坏缓存:不涉及——无缓存/重试/游标/去重逻辑变更。
- 是否改了设计方向:不涉及——仅展示层守卫,无架构变更。
- 是否引入了新的配置:不涉及。
判定:代码正确、测试有牙、披露完整;因 kernel 变更属 human-only 审核档,本 review 仅作 findings 记录,不 approve——等 owner 人工批准。
🔶 需要决策 PR #6305(kernel 变更,human-only 审核档)——审核完成,等 owner 授权合并。
备注:needs-decision 标签未能打上——本工作区 tools/ 下没有 add-label.sh,镜像 API shim 也未实现 labels 端点(按 Problem Reporting 如实上报)。 review lane 一行汇报:PR #6305(head 03a2c5b)审核完成——findings-only、无阻塞发现、无顺手修、无新 issue;等 owner kernel 档批准。 |
[bot] (ai: vllm-qwen/qwen3.8-27b) 冲突已第三次清除并推送。本轮:master 自 11-20 后又前进 286 个提交(kernel 版本 0.0.199→0.0.200,其中 #7252 的版本 bump 与本 PR 的 0.0.199 撞车、master 侧已自行 bump 到 0.0.200 解掉那次碰撞),PR 分支与基线再次失步。 八问:
流程说明:唤醒时已检查旧 head 提交信息,无 判定:新 head |
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Re-review of new head ba00f79 — master sync only, substantive diff unchanged. Findings-only, no approval (kernel gate). No blocking findings.
The new head is a pure master sync (merge commit ba00f79; the PR's own commit remains 9c73677, unchanged). Verified on this head:
kernel/src/report.tsandkernel/tests/blocked-status-ranges.test.tsare byte-identical to the previously reviewed head (03a2c5b); the 3-dot diff vs master is the same 5 files with the same stats.- Master has not touched any affected path since the previous sync (kernel/src/report.ts, recommend.ts, segment.ts, boundaries.ts, src/marker/status-body.ts, src/acp-panel.ts, src/acp-status.ts).
- Kernel version: 0.0.199 → 0.0.201 (collision with master's 0.0.200 resolved in the merge commit), consistent across
kernel/package.json,kernel/package-lock.json, and rootpackage-lock.json— version discipline satisfied. - Targeted suites re-run on this head: new suite + related kernel suites 34/34, host
tests/acp-status.test.ts9/9, kernel build clean. CI mirror: 0 failed (test shards in flight).
All conclusions of the previous review (id 5909661389) carry over: the fix is a display-layer guard, apply/recommend/segmentation untouched, all three behavior changes disclosed in the PR body, tests assert on real rendered output with teeth.
八问
- 是什么问题:同前一轮——
acp_status未压缩区间视图把BLOCKED哨兵渲染成候选端点并配 "refs unresolvable" 误诊。 - 什么原因导致的:同前一轮——report 层直接取 group 首/末条目当候选端点、缺 BLOCKED 守卫;recommend 层早在 base-set 级就筛掉 BLOCKED。
- 影响范围与严重程度:同前一轮——仅展示层,压缩行为不受影响,体验/诊断瑕疵级。
- 怎么修复的:同前一轮——展示层守卫(BLOCKED 收进聚合统计行 + 候选行按位置跨正常成员 + 全保护 group 显式说明 +
flagFor打印真实错误)。 - 是否有回归问题:无。本轮额外核验:新 head 为纯 master 同步、PR 文件逐字节不变、master 未触碰受影响路径、版本冲突已解为 0.0.201(三处一致)、受影响套件全绿。
- 是否破坏缓存:不涉及。
- 是否改了设计方向:不涉及。
- 是否引入了新的配置:不涉及。
判定:新 head 仅 master 同步,实质内容不变,重验全绿;结论与前一轮相同——findings-only、不 approve,等 owner kernel 档人工批准。
复审核完成(新 head ba00f79):新 head 为纯 master 同步(PR 自身 commit 仍只有 9c73677),实质 diff 与上一 head 逐字节一致,master 未触碰任何受影响路径;kernel 版本冲突已解为 0.0.201(三处版本文件一致)。新 head 上重跑定向套件:kernel 34/34 + 宿主 acp-status 9/9 + build 绿;CI 镜像 0 失败(shards 运行中)。 结论不变:findings-only、无阻塞发现、无顺手修、无新 issue——等 owner kernel 档人工批准(见前一条 🔶 需要决策 评论 + GitHub review id 5979784640)。 |
What & why
Fixes #6168 (follow-up empirical report to #6167 — Windows / Codex CLI evidence).
acp_status's uncompressed range view rendered candidate rows whose endpoints were theBLOCKEDsentinel (e.g.BLOCKED–m02518 ⚠ not foldable: refs unresolvable). Root cause:collectVisible()correctly includes policy-protected messages in the totals, butrenderUncompressedRanges()used each segment group's first/last entry verbatim as candidate endpoints — while the recommend side already screens BLOCKED out at base-set level (buildCompressibleRanges). The report view violated the documented identical-segmentation contract. The apply path is unaffected, matching the reporter's observation that valid ranges compress fine.Changes (display layer only)
BLOCKEDentry in a group moves to one aggregate stats line under the header:protected: N msgs, T (BLOCKED refs — not submittable to compress); header totals and per-message counts unchanged.flagForprints the real projection/boundary error (whitespace-collapsed, capped at 100 chars) instead of the flat "refs unresolvable" misdiagnosis (coordinate point with [kernel] Three state-ledger distortions in nested folding: supersedes_block_id has no writer / bare tags spanning alread | [kernel] 嵌套折叠三处状态账目失真:supersedes_block_id 无 writer / 跨已覆盖区间裸标签诱发冗余重折 / 投影异常吞成 flat verdict(v0.1.212 实证) #5667).projectUnit()on the exact apply-view class.Behavior changes (old → new, disclosed)
BLOCKED— protected content now appears once in the stats line plus the per-message drilldown (was: bogus row + flat warning).(no candidate ranges — every visible message is protected)instead of an unresolvable row.Tests
New
kernel/tests/blocked-status-ranges.test.ts— 9 tests over the issue's acceptance matrix: leading/trailing/mid-group/all-BLOCKED, non-monotonic refs, genuine resolution failure, sort=time+limit, per-message drilldown, normal-only regression guard. Targeted suites green locally (69/69 incl. the pre-existing report/ranges/reversed-ranges/status-surface suites); full-suite verdict = CI on this head.Kernel boundary
kernel-change: missing guard against the BLOCKED sentinel in report-layer candidate construction — violates the documented identical-segmentation contract between recommend and report views (recommend already carries the contrasting implementation). Mechanical-bug exemption applied (owner policy 2026-11-06): reproducing regression test added, declaration present; the human-only review tier still gates merge. Kernel version 0.0.193 → 0.0.194 (CI-enforced version discipline).
Closes #6168