Repository navigation
[PR] fix: honest compose surfaces — [not foldable] tags on UNCOVERED HISTORY runs + historical-refs note on compress success receipts - #5311
Conversation
(ai: vllm-qwen/qwen3.8-27b) 查重: 不重复。 这是 #5186 compose-surface honesty 的 host 侧子件(代理车道),对应 GitHub PR #5311(分支 标签: bug + scope/wiring + p2 —— 缺陷类(模型把不可折叠行抄进 compress 批次导致整批被拒);改动面在请求管道/回执与 acp_status 工具表面,属协议接线;严重度 p2(有绕行:模型仍可只从 acp_status 组批,属浪费回合而非数据丢失/缓存泄露)。 正在对 PR #5311 做独立审核(方向、回归、代码、测试有效性),结论按结构化八问跟进。 |
📦 Built Package ArtifactBranch: Latest build — install this exact versionnpm install -g billion-context@0.0.10984-pr.5311.8774576This exact version is immutable and reproducible. The rolling Recent builds (commit → version)
Download artifact
tar xzf billion-context-pr5311.tgz
npm install -g packageInstall instructions are refreshed on each push. |
140c4c5 to
205213e
Compare
|
(ai: vllm-qwen/qwen3.8-27b) The verdict-backed tag follow-up for the UNCOVERED HISTORY runs (the blanket Filed as #5784 (kernel authorization request: pure re-export, zero behavior change; options A approve / B keep the blanket tag). Verified against master @ be611a8: all four functions + the three types exist and are module-level exported, the package boundary genuinely lacks them, and Waiting on the owner's A/B ruling; the verdict-tag follow-up lands on this branch after approval. |
…surfaces The compress success receipt gained the historical-refs note (#5311), which enters the wire once per fold as an ephemeral tool payload and shifts the session's input-token accounting by its length. The pinned outbound-body digest of the fold turn, the persisted-state digests, and the stats/metadata (+ nudge watermark) section digests move accordingly. Regenerated via scripts/update-storage-render-goldens.ts under the documented parity pins; A/B drive against the pre-change code shows exactly one differing wire body (the fold turn carrying the note) and token-count-only state deltas.
|
(ai: vllm-qwen/qwen3.8-27b) Addendum to my earlier note here: the verdict-backed screen no longer needs the #5784 kernel export — PR #5785 implements it host-side ( The owner delegated the #5784 ruling and it is shelved (option B) as an optional optimization for any future surface that wants the direct-projection path. Nothing on this branch waits on #5784. |
(ai: vllm-qwen/qwen3.8-27b) 红 CI 已修复并全绿:新增 commit 结构化八问
备注:owner 评论里提到的 verdict-backed 标签后续项(需 kernel 边界导出,已立 #5784 待 A/B 拍板)不在本次红 CI 范围内,本次未触碰;golden 重生成不依赖该决策,blanket 标签版本即可先行合入。 判定:可合并。 |
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Review verdict: APPROVE ✅ — clean host-side fix, correctly scoped, well-tested, green CI (30 checks, 0 failed/pending), zero kernel/ changes.
(Posted as a comment-review: GitHub forbids the author-identity from setting a formal APPROVE state on its own PR; the recommendation below stands and merge stays human.)
Reviewed the full diff (9 files) in source context at head 62e2595b1, both compression modes, and the golden-render regression surface.
Correctness (verified in source, not just the diff):
[not foldable]rides the shareddescribeUncoveredRuns; its only two callers aresrc/acp-status.ts:329(UNCOVERED HISTORY section) andsrc/stream.ts:694(success-receipt "Uncovered history remains" variant) → both surfaces tagged, exactly as claimed. Runs come fromuncoveredHistoryRuns, which skips anything inside the actionable window (uncovered.ts:166), so the tag is true for every listed row.HISTORICAL_REFS_NOTEis gated onchangedBlocks.length > 0(success-only) and rides bothmsg+logMsg(stream.ts:1595-1598) — consistent with #1718 (static text, no excerpts → log-safe; same pattern as the subagent note directly above it). Failure/no-op receipts untouched.
Regression audit:
- Both modes: proxy-lane only. The proxy executes compress server-side (
applyRanges); plugin-mode hosts execute locally and render their own receipts (KDD #7), so they get neither change. The tag is additive cosmetic text regardless. - Behavior drift old→new disclosed for both changes ✓.
- Storage-render goldens: only the fold-round body (
i:9) grew +292B (= the note);i:10/11/12are byte-identical → the note appears once at the fold and does NOT propagate into the re-sent prefix → no recurring prefix-cache break.blockContentsdigest unchanged (stored summaries intact); state/stats/metadata/nudge digests moved as the expected consequence of the larger receipt. No wire-shape or persistence-format change.
Tests: real teeth — P1/P2/N1/P3/N2 positive+negative pairs (N1 pins the tag does NOT leak into the offered/actionable list; N2 pins failure receipts stay clean); compress-fingerprint-refold updated to pin one-note-per-line and that no-op carries neither note. These catch the regressions (tag removed / leaked into offers / note on failure).
Non-blocking note (transparency only): the PR body describes one new test file, but the diff also adds tests/issue5186-batch-coverage-staleness.test.ts (+171) pinning kernel applyCompression batch-vs-sequential coverage (tied to "#5186 checklist item ④ / #5780 body ④"; documents a probed non-reproduction + a nested-double-block wart). It is test-only, zero production-code change, within the #5186 family, and harmless — but it isn't mentioned in the body. Worth a one-line mention; not blocking.
Coordination (per owner addendum): noted — #5784 kernel export is moot (host-side via #5785, shelved option B); nothing on this branch waits on it. The shared render hunk with #5785 is a merge-ordering matter (second-merger rebases; marker composes after any prefix) and does not affect this PR's own diff.
八问(结构化结论)
- 是什么问题:模型会把「UNCOVERED HISTORY」里活像可折叠的 ref 段、以及旧摘要里引用的过期 ref,抄进 compress batch → 整批被拒、0 折叠。
- 什么原因导致的:两个诊断面(acp_status 未覆盖段 + 成功回执)把不可折叠的 span 用活 ref 呈现,且没告诉模型「摘要里引用的 ref 是随代际重置的历史快照」。
- 影响范围与严重程度:仅 proxy 车道;正确性/体验瑕疵(模型偶尔组错 batch),非崩溃、非数据丢失。
- 怎么修复的:在两个面上加诚实标注——每个未覆盖 run 前缀
[not foldable](共用 formatter,两处都生效);成功回执追加一行「引用是历史快照,range 只从 acp_status/本回执取」。改的是表面措辞,不是过滤模型输出。 - 是否有回归问题:无。仅折叠轮出站体 +292B(=该注记),后续轮字节不变 → 无重复前缀缓存破坏;blockContents 摘要字节不变;无 wire 形状/格式变更。
- 是否破坏缓存:不破坏前缀缓存(重发前缀不变);持久化字节随回执略增(已披露、有界、非格式变更)。
- 是否改了设计方向:否。仍是同源诊断面诚实化,无架构/边界转向。
- 是否引入了新的配置:否。无新 env/config/默认值。
合并判定:可合(APPROVE)。 唯一建议:PR 描述补一句第二个测试文件(非阻塞)。
(ai: vllm-qwen/qwen3.8-27b) Reviewed end-to-end; verdict: APPROVE. Posted the full review on the PR as a comment-review — GitHub won't let the author identity set a formal APPROVE state on its own PR, so the recommendation lives in the review body and merge stays yours. Your addendum is incorporated and changes nothing blocking:
Regression checks I ran beyond CI-green (30 checks, 0 failed/pending):
One non-blocking transparency nit: the PR body documents one new test file, but the diff also adds Full 八问 structured conclusion is in the review body. 合并判定:可合(APPROVE)。 |
62e2595 to
21a6bec
Compare
…surfaces The compress success receipt gained the historical-refs note (#5311), which enters the wire once per fold as an ephemeral tool payload and shifts the session's input-token accounting by its length. The pinned outbound-body digest of the fold turn, the persisted-state digests, and the stats/metadata (+ nudge watermark) section digests move accordingly. Regenerated via scripts/update-storage-render-goldens.ts under the documented parity pins; A/B drive against the pre-change code shows exactly one differing wire body (the fold turn carrying the note) and token-count-only state deltas. Rebased onto master 33104b1 (conflict in these same four files: master had regenerated them for its own changes). Post-rebase delta vs the pre-rebase goldens is the metadata section digest only: master bumped the package version 0.1.212 -> 0.1.213, which the persisted session metadata embeds. Every per-turn wire body sha/shaCiteNorm is unchanged.
21a6bec to
12b4912
Compare
…surfaces The compress success receipt gained the historical-refs note (#5311), which enters the wire once per fold as an ephemeral tool payload and shifts the session's input-token accounting by its length. The pinned outbound-body digest of the fold turn, the persisted-state digests, and the stats/metadata (+ nudge watermark) section digests move accordingly. Regenerated via scripts/update-storage-render-goldens.ts under the documented parity pins; A/B drive against the pre-change code shows exactly one differing wire body (the fold turn carrying the note) and token-count-only state deltas. Rebased onto master 33104b1 (conflict in these same four files: master had regenerated them for its own changes). Post-rebase delta vs the pre-rebase goldens is the metadata section digest only: master bumped the package version 0.1.212 -> 0.1.213, which the persisted session metadata embeds. Every per-turn wire body sha/shaCiteNorm is unchanged.
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Re-review at new head 12b49120f — Verdict: APPROVE.
(Posted as a comment-review: GitHub blocks the author identity from setting a formal APPROVE on its own PR; the recommendation stands and merge stays human.)
What changed since my last review (62e2595b1): only a master merge into the branch (kernel → 0.0.195, #5765 blockOps gate, #6132 compressAtFoldPoint removal, …). The PR's own contribution vs master is byte-identical — same 9 files, same line deltas, same body, no kernel/ files. This is exactly the "second-merger rebases" coordination noted earlier.
Verified at the new head (re-checked, not assumed):
describeUncoveredRunshunk byte-identical; still exactly two callers (src/acp-status.ts:339,src/stream.ts:711) — master added no third caller, so[not foldable]reaches precisely the two intended surfaces.HISTORICAL_REFS_NOTEstill gated onchangedBlocks.length > 0(success-only), ridingmsg+logMsg— consistent with #1718 (static text, no excerpts → log-safe). Failure/no-op receipts untouched.- The merged master introduced no semantic interaction with either surface (no new formatter caller;
applyRangesgate intact).
Regression posture (unchanged from last round): proxy-lane only; plugin-mode hosts render their own receipts; the note appears once at the fold round (golden delta is the note-sized change at the single fold index; later bodies untouched) → no recurring prefix-cache break; stored-summary bytes intact; no wire-shape or persistence-format change. Behavior drift old→new disclosed in the body.
Tests: P1/P2/N1/P3/N2 positive+negative pairs pin the tag-on-every-run / tag-not-in-offer / note-on-success / note-not-on-failure directions; fingerprint-refold pins one-note-per-line and that no-op carries neither. These catch the regressions.
CI: at review time the new head shows 0 failures — 8 passed / 12 skipped / 10 guard-lanes still queued (suite in progress). The PR's own diff was fully green at the prior head and this merges onto green master.
No open findings. 合并判定:可合(APPROVE)。
八问(结构化结论)
- 是什么问题:模型会把「UNCOVERED HISTORY」里活像可折叠的 ref 段、以及旧摘要里引用的过期 ref,抄进 compress batch → 整批被拒、0 折叠。
- 什么原因导致的:两个诊断面(acp_status 未覆盖段 + 成功回执)把不可折叠 span 用活 ref 呈现,且未告知模型「摘要里引用的 ref 是随代际重置的历史快照」。
- 影响范围与严重程度:仅 proxy 车道;正确性/体验瑕疵(模型偶尔组错 batch),非崩溃、非数据丢失。
- 怎么修复的:在两个面上加诚实标注——每个未覆盖 run 前缀
[not foldable](共用 formatter,两处生效);成功回执追加一行「引用是历史快照,range 只从 acp_status/本回执取」。改的是表面措辞,不是过滤模型输出。 - 是否有回归问题:无。master 合并后逐点复核:formatter 仍仅两处调用、note 仍受成功门控、无新调用方/语义交互;仅折叠轮出站体增该注记、后续轮字节不变 → 无重复前缀缓存破坏;摘要字节不变;无 wire/格式变更。
- 是否破坏缓存:不破坏前缀缓存(重发前缀不变);持久化字节随回执略增(已披露、有界、非格式变更)。
- 是否改了设计方向:否。仍是同源诊断面诚实化,无架构/边界转向。
- 是否引入了新的配置:否。无新 env/config/默认值。
(ai: vllm-qwen/qwen3.8-27b) Re-reviewed at the new head The only change since my last pass is the master merge you coordinated (kernel → 0.0.195, #5765 blockOps gate, #6132 compressAtFoldPoint removal). The PR's own 9-file diff vs master is byte-identical to what I reviewed — same hunks, same body, no
Regression posture unchanged: proxy-lane only; the note lands once at the fold round (later bodies untouched) → no recurring prefix-cache break; stored-summary / wire / persistence-format all intact. No open findings. CI on this head: 0 failures at review time (8 passed / 12 skipped / 10 guard-lanes still queued — suite in progress); the PR's own diff was fully green at the prior head and merges onto green master. Full 八问 structured conclusion is in the review. 合并判定:可合(APPROVE)。 |
12b4912 to
82439e1
Compare
…surfaces The compress success receipt gained the historical-refs note (#5311), which enters the wire once per fold as an ephemeral tool payload and shifts the session's input-token accounting by its length. The pinned outbound-body digest of the fold turn, the persisted-state digests, and the stats/metadata (+ nudge watermark) section digests move accordingly. Regenerated via scripts/update-storage-render-goldens.ts under the documented parity pins; A/B drive against the pre-change code shows exactly one differing wire body (the fold turn carrying the note) and token-count-only state deltas. Rebased onto master 33104b1 (conflict in these same four files: master had regenerated them for its own changes). Post-rebase delta vs the pre-rebase goldens is the metadata section digest only: master bumped the package version 0.1.212 -> 0.1.213, which the persisted session metadata embeds. Every per-turn wire body sha/shaCiteNorm is unchanged.
82439e1 to
14acdda
Compare
…surfaces The compress success receipt gained the historical-refs note (#5311), which enters the wire once per fold as an ephemeral tool payload and shifts the session's input-token accounting by its length. The pinned outbound-body digest of the fold turn, the persisted-state digests, and the stats/metadata (+ nudge watermark) section digests move accordingly. Regenerated via scripts/update-storage-render-goldens.ts under the documented parity pins; A/B drive against the pre-change code shows exactly one differing wire body (the fold turn carrying the note) and token-count-only state deltas. Rebased onto master 33104b1 (conflict in these same four files: master had regenerated them for its own changes). Post-rebase delta vs the pre-rebase goldens is the metadata section digest only: master bumped the package version 0.1.212 -> 0.1.213, which the persisted session metadata embeds. Every per-turn wire body sha/shaCiteNorm is unchanged.
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Re-review at head 14acddaa4 — Verdict: APPROVE (posted as a comment-review; a formal self-APPROVE is blocked because the posting PAT identity is the PR author ranxianglei).
What changed since my last review (head 12b49120f): another master re-merge into the branch — kernel 0.0.197, #6480 pure-SQLite default, #6640 watermark fastpath (#5403/#5418), #6508 announcement policy, #6568 release-gate, and more. The PR's own 3 commits (fix + 2 tests) were replayed on the new base.
The PR's contribution is unchanged (third confirmation). PR-vs-master (GitHub merge-base, base e5e1647ba0) is the same 9 files with the same line deltas, no kernel/:
src/stream.ts+7,src/uncovered.ts+6/-1,tests/compress-fingerprint-refold.test.ts+8/-3, 4tests/golden/storage-render/*.json, 2 new test files (issue5186-compose-surface-honesty+157,issue5186-batch-coverage-staleness+171).
Re-verified at the new head (byte-identical hunks):
describeUncoveredRunsstill has exactly two callers (src/acp-status.ts:339,src/stream.ts:711); the[not foldable]tag stays true for every listed run (they are outside the actionable offer by construction).HISTORICAL_REFS_NOTEstill success-gated —if (changedBlocks.length > 0)atsrc/stream.ts:1670, appended tomsg+logMsg(1672/1673). Failure/no-op receipts untouched.- The 4 goldens still differ from the new master base by only the note-sized delta at the single fold index — master's storage/kernel changes (pure-SQLite default, kernel 0.0.197) are in the base, not the PR, so the note is the only PR-side persisted-state change. No recurring prefix-cache break (later round bodies byte-identical).
Regression posture unchanged: no wire/persistence-format change; proxy-lane only (plugin hosts render their own receipts); the note is static, excerpt-free, log-safe (#1718-consistent). The re-merge introduced no new regression.
⚠ CI state (honest note): at review time, head 14acddaa4 shows 25 cancelled / 3 success / 2 skipped / 0 failed / 0 pending in the check cache — the run looks cancelled/interrupted, not a clean green pass (the trigger header says "green", but the authoritative check source does not show one). The code verdict is on merit (byte-identical to a diff that was fully green at 62e2595b1); please confirm a green CI run on this head before merging.
No open code findings (NO OPEN ENDS): nothing to fix on the spot, nothing to convert to a decision, nothing parked.
8问 (Chinese, for human reviewers):
- 是什么问题: 让压缩回执对「不可折叠」与「历史引用」说真话——给 UNCOVERED HISTORY 每段加
[not foldable]前缀,并在压缩成功回执里加一段「摘要里引用的 ref 是历史快照」的说明。 - 什么原因导致的: 模型把 UNCOVERED HISTORY 的 ref 区间误当折叠建议复制进 compress 范围;且会话重置(ref 从 m00001 重启)后,旧摘要里的 ref 引用可能失效,回执未提示。
- 影响范围与严重程度: 仅代理车道(proxy-lane)的压缩成功回执与 acp_status 的 UNCOVERED HISTORY 展示;属体验/诚实性瑕疵,无数据丢失、无 wire 变更。
- 怎么修复的: 在
describeUncoveredRuns每行加[not foldable]前缀;新增HISTORICAL_REFS_NOTE常量,仅在压缩成功(changedBlocks.length > 0)时拼进 msg+logMsg。 - 是否有回归问题: 无。
describeUncoveredRuns恰好两个调用方,[not foldable]对每行恒真;note 仅成功路径、静态无摘录、日志安全(#1718 一致);goldens 仅折叠轮 +292B,后续轮字节一致 → 无周期性前缀缓存破坏;rebase 后 master 侧存储/kernel 变更都在 base,PR 自身 diff 三次确认字节一致。 - 是否破坏缓存: 无。note 只落在单个折叠轮,后续轮字节不变。
- 是否改了设计方向: 无。不改架构/边界/数据流。
- 是否引入了新的配置: 无。
判定:APPROVE(代码层面);合并前请确认该 head 的 CI 为绿色(当前缓存显示 25 cancelled / 0 failed / 0 pending)。
[bot] 🏷 PR #5311 re-review (round 3) at head This is the third review of this PR. Since my last review ( The contribution is unchanged — third byte-identical confirmation. PR-vs-master (merge-base ⚠ CI is NOT green at this head. The check cache shows 25 cancelled / 3 success / 2 skipped / 0 failed / 0 pending (confirmed stable on a second read) — a cancelled/interrupted run, not the "green" the trigger header claims. The code verdict stands on merit (byte-identical to a diff that was fully green at Full review posted on the PR (review id 5901736025). No 8问 (Chinese):
判定:APPROVE(代码层面);合并前请确认该 head 的 CI 为绿色(当前缓存 25 cancelled / 0 failed / 0 pending)。 |
14acdda to
ea67419
Compare
…surfaces The compress success receipt gained the historical-refs note (#5311), which enters the wire once per fold as an ephemeral tool payload and shifts the session's input-token accounting by its length. The pinned outbound-body digest of the fold turn, the persisted-state digests, and the stats/metadata (+ nudge watermark) section digests move accordingly. Regenerated via scripts/update-storage-render-goldens.ts under the documented parity pins; A/B drive against the pre-change code shows exactly one differing wire body (the fold turn carrying the note) and token-count-only state deltas. Rebased onto master 33104b1 (conflict in these same four files: master had regenerated them for its own changes). Post-rebase delta vs the pre-rebase goldens is the metadata section digest only: master bumped the package version 0.1.212 -> 0.1.213, which the persisted session metadata embeds. Every per-turn wire body sha/shaCiteNorm is unchanged.
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Re-review at head ea674196e — Verdict: APPROVE (posted as a comment-review; a formal self-APPROVE is blocked because the posting PAT identity is the PR author ranxianglei).
What changed since my last review (head 14acddaa4): another master re-merge into the branch — #6996 compress-nudge-example, #7054 attribution-diagnostics topic, #6938 release-v0.1.216, #6958 Claude Code deferred-tools, #6922 architecture-docs map fix, #6953 reply-stash, wire-contract, and more (304 files). The PR's own 3 commits (fix + 2 tests) were replayed on the new base.
The PR's contribution is unchanged (fourth confirmation). PR-vs-master (GitHub merge-base, base a259e0b8fa) is the same 9 files with the same line deltas, no kernel/:
src/stream.ts+7,src/uncovered.ts+6/-1,tests/compress-fingerprint-refold.test.ts+8/-3, 4tests/golden/storage-render/*.json, 2 new test files (issue5186-compose-surface-honesty+157,issue5186-batch-coverage-staleness+171).
Re-verified at the new head (byte-identical hunks; only line numbers shifted by master inserts):
describeUncoveredRunsstill has exactly two callers (src/acp-status.ts:339,src/stream.ts:760); the[not foldable]tag stays true for every listed run (they are outside the actionable offer by construction).HISTORICAL_REFS_NOTEstill success-gated —if (changedBlocks.length > 0)atsrc/stream.ts:1791, appended tomsg+logMsg(1793/1794). Failure/no-op receipts untouched.- The 4 goldens still differ from the new master base by only the note-sized delta at the single fold index (master's storage/kernel/wire changes are in the base, not the PR). No recurring prefix-cache break (later round bodies byte-identical).
Regression posture unchanged: no wire/persistence-format change; proxy-lane only (plugin hosts render their own receipts); the note is static, excerpt-free, log-safe (#1718-consistent). The re-merge introduced no new regression.
CI state (improved vs last round): head ea674196e is in progress — 14 pending (5 queued + 9 in_progress), 4 success, 12 skipped, 0 failed. The round-3 cancelled state is resolved; this is a live run with no failures so far. It should land green on completion; please confirm the final green before merging.
No open code findings (NO OPEN ENDS): nothing to fix on the spot, nothing to convert to a decision, nothing parked.
8问 (Chinese, for human reviewers):
- 是什么问题: 让压缩回执对「不可折叠」与「历史引用」说真话——给 UNCOVERED HISTORY 每段加
[not foldable]前缀,并在压缩成功回执里加「摘要里引用的 ref 是历史快照」说明。 - 什么原因导致的: 模型把 UNCOVERED HISTORY 的 ref 区间误当折叠建议复制进 compress 范围;会话重置(ref 从 m00001 重启)后旧摘要 ref 可能失效,回执未提示。
- 影响范围与严重程度: 仅代理车道(proxy-lane)压缩成功回执与 acp_status 的 UNCOVERED HISTORY 展示;体验/诚实性瑕疵,无数据丢失、无 wire 变更。
- 怎么修复的:
describeUncoveredRuns每行加[not foldable]前缀;新增HISTORICAL_REFS_NOTE,仅压缩成功(changedBlocks.length > 0)时拼进 msg+logMsg。 - 是否有回归问题: 无。
describeUncoveredRuns恰好两个调用方,[not foldable]每行恒真;note 仅成功路径、静态无摘录、日志安全(#1718 一致);goldens 仅折叠轮 +292B,后续轮字节一致 → 无周期性前缀缓存破坏;PR 自身 diff 四次确认字节一致。 - 是否破坏缓存: 无。note 只落在单个折叠轮,后续轮字节不变。
- 是否改了设计方向: 无。
- 是否引入了新的配置: 无。
判定:APPROVE(代码层面);CI 进行中(0 failed),完成后确认绿色即可合并。
[bot] 🏷 PR #5311 re-review (round 4) at head Fourth review of this PR. Since my last review ( The contribution is unchanged — fourth byte-identical confirmation. PR-vs-master (merge-base CI is healthy this round (improved vs last round's cancelled state): head Full review posted on the PR (review id 5946257807). No 8问 (Chinese):
判定:APPROVE(代码层面);CI 进行中(0 failed),完成后确认绿色即可合并。 |
ea67419 to
77ff8de
Compare
…surfaces The compress success receipt gained the historical-refs note (#5311), which enters the wire once per fold as an ephemeral tool payload and shifts the session's input-token accounting by its length. The pinned outbound-body digest of the fold turn, the persisted-state digests, and the stats/metadata (+ nudge watermark) section digests move accordingly. Regenerated via scripts/update-storage-render-goldens.ts under the documented parity pins; A/B drive against the pre-change code shows exactly one differing wire body (the fold turn carrying the note) and token-count-only state deltas. Rebased onto master 33104b1 (conflict in these same four files: master had regenerated them for its own changes). Post-rebase delta vs the pre-rebase goldens is the metadata section digest only: master bumped the package version 0.1.212 -> 0.1.213, which the persisted session metadata embeds. Every per-turn wire body sha/shaCiteNorm is unchanged.
(ai: vllm-qwen/qwen3.8-27b) CI 工程师本轮结论:上报的冲突已解决,本轮无需再 rebase / push。 触发器报「PR 与 base 冲突 (merge state: dirty)」,但该状态来自 11-17 的过期镜像缓存(缓存 sha 仍是旧 head 本轮核实的证据(全部 live,非记忆):
CI 状态说明(诚实边界): 镜像缓存仍停在旧 head 结构化八问
判定:冲突已解决(由前序/并行 run 完成),分支干净且已 rebase 到当前 master,本轮无遗留 rebase/push 动作;新 head CI 绿/红待镜像缓存刷新后确认。 |
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Re-review at head 77ff8deaf — Verdict: APPROVE (posted as a comment-review; a formal self-APPROVE is blocked because the posting PAT identity is the PR author ranxianglei).
What changed since my last review (head ea674196e): another master re-merge into the branch. The PR's own 3 commits (fix + 2 tests) were replayed on the new base.
The PR's contribution is unchanged (fifth confirmation). PR-vs-master (GitHub merge-base, base a446a3c1c8) is the same 9 files with the same line deltas, no kernel/:
src/stream.ts+7,src/uncovered.ts+6/-1,tests/compress-fingerprint-refold.test.ts+8/-3, 4tests/golden/storage-render/*.json, 2 new test files (issue5186-compose-surface-honesty+157,issue5186-batch-coverage-staleness+171).
Re-verified at the new head (byte-identical hunks; only line numbers shifted by master inserts):
describeUncoveredRunsstill has exactly two callers (src/acp-status.ts:339,src/stream.ts:774); the[not foldable]tag stays true for every listed run (they are outside the actionable offer by construction).HISTORICAL_REFS_NOTEstill success-gated —if (changedBlocks.length > 0)atsrc/stream.ts:1815, appended tomsg+logMsg(1817/1818). Failure/no-op receipts untouched.- The 4 goldens still differ from the new master base by only the note-sized delta at the single fold index (master's changes are in the base, not the PR). No recurring prefix-cache break (later round bodies byte-identical).
Regression posture unchanged: no wire/persistence-format change; proxy-lane only (plugin hosts render their own receipts); the note is static, excerpt-free, log-safe (#1718-consistent). The re-merge introduced no new regression.
CI is GREEN this round: head 77ff8deaf shows 17 success / 13 skipped / 0 failed / 0 pending — a clean green pass. This is the first fully-green CI state across my five reviews (round 3 was cancelled, round 4 in-progress). Ready to merge from a CI standpoint.
No open code findings (NO OPEN ENDS): nothing to fix on the spot, nothing to convert to a decision, nothing parked.
8问 (Chinese, for human reviewers):
- 是什么问题: 让压缩回执对「不可折叠」与「历史引用」说真话——给 UNCOVERED HISTORY 每段加
[not foldable]前缀,并在压缩成功回执里加「摘要里引用的 ref 是历史快照」说明。 - 什么原因导致的: 模型把 UNCOVERED HISTORY 的 ref 区间误当折叠建议复制进 compress 范围;会话重置(ref 从 m00001 重启)后旧摘要 ref 可能失效,回执未提示。
- 影响范围与严重程度: 仅代理车道(proxy-lane)压缩成功回执与 acp_status 的 UNCOVERED HISTORY 展示;体验/诚实性瑕疵,无数据丢失、无 wire 变更。
- 怎么修复的:
describeUncoveredRuns每行加[not foldable]前缀;新增HISTORICAL_REFS_NOTE,仅压缩成功(changedBlocks.length > 0)时拼进 msg+logMsg。 - 是否有回归问题: 无。
describeUncoveredRuns恰好两个调用方,[not foldable]每行恒真;note 仅成功路径、静态无摘录、日志安全(#1718 一致);goldens 仅折叠轮 +292B,后续轮字节一致 → 无周期性前缀缓存破坏;PR 自身 diff 五次确认字节一致。 - 是否破坏缓存: 无。note 只落在单个折叠轮,后续轮字节不变。
- 是否改了设计方向: 无。
- 是否引入了新的配置: 无。
判定:APPROVE(代码层面);CI 已全绿(17 success / 13 skipped / 0 failed),可合并。
[bot] 🏷 PR #5311 re-review (round 5) at head Fifth review of this PR. Since my last review ( The contribution is unchanged — fifth byte-identical confirmation. PR-vs-master (merge-base CI is GREEN this round: head Full review posted on the PR (review id 5965795449). No 8问 (Chinese):
判定:APPROVE(代码层面);CI 已全绿(17 success / 13 skipped / 0 failed),可合并。 |
…d refs as historical on compress success receipts (#5186)
…: no observable staleness divergence; one nested-double-block wart pinned) (#5186)
…surfaces The compress success receipt gained the historical-refs note (#5311), which enters the wire once per fold as an ephemeral tool payload and shifts the session's input-token accounting by its length. The pinned outbound-body digest of the fold turn, the persisted-state digests, and the stats/metadata (+ nudge watermark) section digests move accordingly. Regenerated via scripts/update-storage-render-goldens.ts under the documented parity pins; A/B drive against the pre-change code shows exactly one differing wire body (the fold turn carrying the note) and token-count-only state deltas. Rebased onto master 33104b1 (conflict in these same four files: master had regenerated them for its own changes). Post-rebase delta vs the pre-rebase goldens is the metadata section digest only: master bumped the package version 0.1.212 -> 0.1.213, which the persisted session metadata embeds. Every per-turn wire body sha/shaCiteNorm is unchanged.
77ff8de to
138fa54
Compare
(ai: vllm-qwen/qwen3.8-27b) CI 工程师本轮:冲突已解决,rebase + golden 重生后 force-push 新 head 本轮的冲突是真实的(与上轮不同):master 前进了 21 个 commit,其中 本轮做的(全部 live 核实):
CI 状态(诚实边界): 镜像缓存仍停在 rebase 前的旧 head 结构化八问
判定:冲突已解决(rebase + 官方脚本 golden 重生),分支干净 fast-forward,已 force-push 新 head |
ranxianglei
left a comment
There was a problem hiding this comment.
(ai: vllm-qwen/qwen3.8-27b)
Re-review at head 138fa54d4 — Verdict: APPROVE (posted as a comment-review; a formal self-APPROVE is blocked because the posting PAT identity is the PR author ranxianglei).
What changed since my last review (head 77ff8deaf): another master re-merge into the branch. The PR's own 3 commits (fix + 2 tests) were replayed on the new base.
The PR's contribution is unchanged (sixth confirmation). PR-vs-master (GitHub merge-base, base 6453feeeb8) is the same 9 files with the same line deltas, no kernel/:
src/stream.ts+7,src/uncovered.ts+6/-1,tests/compress-fingerprint-refold.test.ts+8/-3, 4tests/golden/storage-render/*.json, 2 new test files (issue5186-compose-surface-honesty+157,issue5186-batch-coverage-staleness+171).
Re-verified at the new head (byte-identical hunks; line numbers stable vs last round):
describeUncoveredRunsstill has exactly two callers (src/acp-status.ts:339,src/stream.ts:774); the[not foldable]tag stays true for every listed run (they are outside the actionable offer by construction).HISTORICAL_REFS_NOTEstill success-gated —if (changedBlocks.length > 0)atsrc/stream.ts:1815, appended tomsg+logMsg(1817/1818). Failure/no-op receipts untouched.- The 4 goldens still differ from the new master base by only the note-sized delta at the single fold index (master's changes are in the base, not the PR). No recurring prefix-cache break (later round bodies byte-identical).
Regression posture unchanged: no wire/persistence-format change; proxy-lane only (plugin hosts render their own receipts); the note is static, excerpt-free, log-safe (#1718-consistent). The re-merge introduced no new regression.
CI is GREEN: head 138fa54d4 shows 17 success / 13 skipped / 0 failed / 0 pending — a clean green pass. Ready to merge.
No open code findings (NO OPEN ENDS): nothing to fix on the spot, nothing to convert to a decision, nothing parked.
8问 (Chinese, for human reviewers):
- 是什么问题: 让压缩回执对「不可折叠」与「历史引用」说真话——给 UNCOVERED HISTORY 每段加
[not foldable]前缀,并在压缩成功回执里加「摘要里引用的 ref 是历史快照」说明。 - 什么原因导致的: 模型把 UNCOVERED HISTORY 的 ref 区间误当折叠建议复制进 compress 范围;会话重置(ref 从 m00001 重启)后旧摘要 ref 可能失效,回执未提示。
- 影响范围与严重程度: 仅代理车道(proxy-lane)压缩成功回执与 acp_status 的 UNCOVERED HISTORY 展示;体验/诚实性瑕疵,无数据丢失、无 wire 变更。
- 怎么修复的:
describeUncoveredRuns每行加[not foldable]前缀;新增HISTORICAL_REFS_NOTE,仅压缩成功(changedBlocks.length > 0)时拼进 msg+logMsg。 - 是否有回归问题: 无。
describeUncoveredRuns恰好两个调用方,[not foldable]每行恒真;note 仅成功路径、静态无摘录、日志安全(#1718 一致);goldens 仅折叠轮 +292B,后续轮字节一致 → 无周期性前缀缓存破坏;PR 自身 diff 六次确认字节一致。 - 是否破坏缓存: 无。note 只落在单个折叠轮,后续轮字节不变。
- 是否改了设计方向: 无。
- 是否引入了新的配置: 无。
判定:APPROVE(代码层面);CI 已全绿(17 success / 13 skipped / 0 failed),可合并。
[bot] 🏷 PR #5311 re-review (round 6) at head Sixth review of this PR. Since my last review ( The contribution is unchanged — sixth byte-identical confirmation. PR-vs-master (merge-base CI is GREEN: head Full review posted on the PR (review id 5975321993). No 8问 (Chinese):
判定:APPROVE(代码层面);CI 已全绿(17 success / 13 skipped / 0 failed),可合并。 |
Part of #5186 — compose-surface honesty (host side, proxy lane).
What
Two small host-side fixes for the "model composes a bad batch" family, both at the surfaces where ranges get composed:
UNCOVERED HISTORY rows are tagged
[not foldable](src/uncovered.ts). The section prints live-looking ref spans (m02042–m02068 (27 msgs, ~7.1K tok)) with the "not as fold suggestions" caveat only in a trailing parenthetical — models copied the spans into compress batches anyway (independent codex repro in the CI queue starved for 8 hours by zombie runs: queued/running timeouts never cancelled, runs from deleted branches hold sl | CI 队列被僵尸 run 饿死 8 小时:排队/运行超时不取消、删分支的 run 永远占位 #5186 thread: a batch composed out of this exact line, whole batch rejected, 0 blocks folded). Every listed run now leads with[not foldable], matching the diagnostic-view marker family (⚠ not foldable: …). Both surfaces inherit it through the shared formatter: the acp_status section and the success-receipt "Uncovered history remains" verdict. Runs are by construction outside the actionable offer, so the tag is true for every row.Success receipts teach the source-of-truth split (
src/stream.ts). Stored summaries quote refs verbatim and those quotes stay in context forever while the ledger can reset underneath them (session switch / native compaction restarts refs at m00001). Every fingerprinted compress success receipt now carries one line after the [Bug][pi] compress 回执指纹行被模型误读为摘要截断,触发防御性 decompress 浪费 token #2544 steering: "Refs quoted in stored summaries are historical snapshots: … Compose ranges only from acp_status output or the live ranges listed in this receipt." Rides the model-facing receipt AND the log copy (static text, no excerpts — 日志清理隐私信息 #1718 unaffected). Failure/no-op receipts are untouched.Behavior changes (disclosed)
UNCOVERED HISTORYsection + success-receipt uncovered variant: each run line gains a[not foldable]prefix. Old:m02042–m02068 (27 msgs, ~7.1K tok)→ New:[not foldable] m02042–m02068 (27 msgs, ~7.1K tok).Tests
tests/issue5186-compose-surface-honesty.test.ts: P1 formatter unit (tag on every listed run, cap intact), P2 acp_status surface renders the tag next to the hole span, N1 offered runs stay untagged (tag never leaks into the actionable list), P3 success receipt carries the note, N2 total-failure receipts don't.tests/compress-fingerprint-refold.test.ts: receipt layout pins updated for the new line (one-note-per-line contract preserved); no-op receipt pinned to carry neither note.