fix(cli): prevent /rewind from dropping conversation history after /compress-fast - #9331
Merged
Conversation
computeApiTruncationIndex treated every successful compression marker as a summarizing boundary. /compress-fast emits the same marker but removes no user prompts from the API history and inserts no summary prefix, so rewinding to a post-marker turn collapsed the truncation anchor to the startup context and silently dropped the whole pre-marker conversation. Tag compression UI items with the path that produced them (CompressionProps.compressionKind) and make the rewind mapping ignore 'fast' markers as boundaries. Legacy markers without a kind keep prior behavior. Fixes #9320
Collaborator
Collaborator
|
Thanks for the continued work — this re-run covers the push through Template still good ✓ Problem: unchanged and solid — #9320 is an observed bug with measured evidence (llama.cpp prefill logs, ~87k tokens resumed as ~27k). The scope growth since the first pass is review-driven: each increment answers a named finding from the /review rounds (placeholder count drift, marker-less auto-compaction), not scope creep. The structural residue of those findings is tracked in #9437 rather than folded in here — the right split for a bugfix PR. Direction: aligned, as before. Silent context loss in Size: core paths are touched now, so correcting this section from the first pass. Breakdown: ~109 production-logic lines — core 30 (placeholder predicate + export + a resume-hazard note) and cli 79 (mapping guard, kind tags, type field; the ACP Session.ts delta is comment-only) — vs ~444 test lines. Well under the 500-line maintainer-awareness bar; the Tier-2 100%-confidence bar for core changes applies and is addressed in the code review below. Approach: every edit earns its place. The placeholder predicate is derived from the producer constant so the two cannot drift, the entrance-3 guard converts a silent-truncation corner into a loud abort, and the exact-match collision corner is documented and pinned as fail-safe rather than hidden. Legacy kind-less markers keep today's behavior — disclosed in the PR body. The increments land in the one function where the loss originates, and the diff carries nothing else. Risk: Stage 1e matches Moving on to code review. 🔍 中文说明感谢持续投入——本次 re-run 覆盖到 模板完整 ✓ 问题：不变且扎实——#9320 是已观测、有量化证据的 bug（llama.cpp prefill 日志：约 87k token 恢复成约 27k）。首轮之后的范围增长由评审驱动：每个增量都对应 /review 轮次中的具名发现（占位符计数漂移、无标记自动压缩），不是范围蔓延；结构性残余记录在 #9437 而非并入本 PR——对 bugfix PR 而言拆分合理。 方向：同前，对齐。/rewind 的静默上下文丢失是核心功能的可靠性漏洞。 规模：现已触及核心路径，此处更正首轮的这一节。拆分：约 109 行生产逻辑——core 30（占位符谓词 + 导出 + resume 风险注释）、cli 79（映射守卫、kind 标记、类型字段；ACP Session.ts 增量仅为注释）——对约 444 行测试。远低于 500 行维护者关注线；核心改动的 Tier-2 百分百信心标准适用，并在下方代码审查中落实。 方案：每处改动都有必要。占位符谓词从产生方常量推导，两者不会漂移；entrance-3 守卫把静默截断角落变成响亮中止；精确匹配碰撞角落被记录并钉为失败安全形态而非掩盖。无 kind 的旧标记保持现状行为——PR 正文已披露。增量都落在丢失发生的那一个函数里，diff 没有夹带其他内容。 风险：Stage 1e 命中 进入代码审查 🔍 — Qwen Code · qwen3.8-max Reviewed at |
Collaborator
Code reviewMy independent proposal for the increments pushed since the first pass: for the placeholder hazard, match the FULL generated shape with a pattern derived from the producer's own constant (a prefix match would miscount pasted placeholders), and for the marker-less auto-compaction corner, fail loud in the Verified along the way (on
No correctness, security, or convention blockers. One non-blocking observation, which round 8 already recorded as a deferred probe: the placeholder regex's interior class admits Files changed (13 of 13 shown)
TestingUnattended run — no PR code executed locally. Evidence is the PR's own CI on the reviewed commit, fetched via the API; all checks have completed. Final CI results for
Beyond CI, two evidence sources with honest attribution:
Not verified by this run: wire-level behavior of the two increments added after the verified head — that is the gap the lane below names. Sandboxed verification would settle this: 中文说明代码审查首轮之后新增增量的独立思路：占位符隐患应匹配完整生成形态、且模式从产生方常量推导（前缀匹配会误数粘贴进来的占位符）；无标记自动压缩角落应在 核实过程（基于 未发现正确性、安全或规范阻断项。一条非阻断观察（round 8 已记录为延后探针）：占位符正则内部字符类接受产生方会剥离的 测试无人值守运行——本地未执行任何 PR 代码。证据为被审 commit 上 PR 自身的 CI（经 API 获取）：全部检查已完成且无失败；运行新增 historyMapping 与 microcompact 测试的是 CI 之外的两个证据来源（如实标注）：8 月 17 日的沙箱 沙箱验证可以一锤定音： — Qwen Code · qwen3.8-max Reviewed at |
Collaborator
|
Confidence: 4/5 — the observed bug is fixed and A/B-proven, every review-round increment earned its place and closes a named hazard; the two reservations below are non-blocking. Stepping back: this is the fix the gate should wave through, and the eight review rounds did their job without burying it. The problem was real and measured (#9320's ~87k-becomes-~27k prefill logs), the core fix is the explicit one and carries a genuine A/B load-bearing proof from the sandboxed run, and each increment pushed afterwards answers a named finding — the placeholder exclusion keeps the UI-turn and API-prompt counts aligned, and the entrance-3 guard turns round 8's silent-loss scenario into a loud abort. I re-derived the guard's trigger condition against Two reservations, both named, neither blocking:
CI is fully green on the reviewed head (no pending 中文说明置信度：4/5 —— 已观测的 bug 已修复且有 A/B 承重证明，评审轮次的每个增量都名副其实、关闭了具名隐患；以下两点保留意见均为非阻断。 整体来看：这是门禁应当放行的修复，八轮评审也尽到了职责而没有把它拖死。问题真实且可量化（#9320 约 87k 变约 27k 的 prefill 日志），核心修法是显式的并带沙箱运行给出的真正 A/B 承重证明，其后推送的每个增量都对应一条具名发现——占位符排除保持 UI 轮次与 API prompt 计数对齐，entrance-3 守卫把 round 8 的静默丢失场景变成响亮中止。我对照 两点保留意见，均已点名、均非阻断：
被审 head 的 CI 已全绿（无未完成的 pull_request 运行），批准护栏不适用（内部分支、fix 类型），作者是维护者——现在批准，并钉在被审 commit 上。 — Qwen Code · qwen3.8-max Reviewed at |
Contributor
Code Coverage Summary
CLI Package - Full Text ReportCore Package - Full Text ReportFor detailed HTML reports, please see the 'coverage-reports-22.x-ubuntu-latest' artifact from the main CI run. |
qwen-code-ci-bot
approved these changes
Aug 17, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
LGTM, looks ready to ship — CI landed green after the review. ✅
Collaborator
Author
|
@qwen-code /verify |
Collaborator
|
Sandboxed verification: ✅ passed — merge-ready (agent verdict) - workflow run Ran the PR in an isolated, token-free container: A/B against the base build, mock-free harness assertions, targeted gates. Advisory evidence for human reviewers — not a review, an approval, or a CI check. Scripted assertions: 37 passed · 0 failed · 37 total 中文 — 判定：✅ 通过 · 可合入（agent 判定）沙箱验证在隔离、无凭证的容器中执行了该 PR 的代码（与 base 构建 A/B 对照、无 mock harness 断言、定向门禁）。仅作为评审证据，不构成评审、批准或 CI 检查。 脚本断言：37 通过 · 0 失败 · 37 总计 Verification reportVerification report — PR #9331Verdict: 中文摘要
Central claim + A/BCentral claim: after Harness:
The base-side reds are encoded as passing assertions (expected control failures), per the counting rule. Mutation/vacuity check (PR's own test files run against base source in the worktree): Secondary claims verified:
CorrectionsNone — the PR description's mechanism account (fast compression removes no user prompts, inserts no summary prefix) was checked against Findings (non-blocking)
Not covered
MethodologyEnvironment: CI verify container ( Evidence images
Harness scripts and raw logs are in the workflow run artifacts (7-day retention). — Qwen Code · sandboxed verification |
Collaborator
Author
|
@qwen-code /review --timeout=360 |
Contributor
|
Qwen Code review request accepted. Review is queued in workflow run. |
qwen-code-ci-bot
requested changes
Aug 17, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Not explored to full depth (tool budget reached): "agent 1c": could not verify whether the TUI submit path permits an image-only prompt with empty text (reachability of the media-clearing count desync above).; "agent 6a": verify whether any submission path (TUI image paste with empty text, ACP content blocks) can produce a user-role Content with top-level media and no text part —…; "agent 6a": verify whether UI compression marker items (and compressionKind ) are restored on session resume — if markers are reconstructed from the chat_compression sys….
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
未探索到全部深度（达到工具调用预算）："agent 1c"：could not verify whether the TUI submit path permits an image-only prompt with empty text (reachability of the media-clearing count desync above).；"agent 6a"：verify whether any submission path (TUI image paste with empty text, ACP content blocks) can produce a user-role Content with top-level media and no text part —…；"agent 6a"：verify whether UI compression marker items (and compressionKind ) are restored on session resume — if markers are reconstructed from the chat_compression sys…。
— qwen3.8-max via Qwen Code /review (v0.21.13)
qwen-code-ci-bot
requested changes
Aug 17, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Not explored to full depth (tool budget reached): "agent 1c": confirm whether the TUI can submit a user turn whose API Content has no text part (image-only paste / bare @image.png mention), which decides reachability o….
Deferred under the convergence posture (round 2, not a blocker) — recorded, not requested in this round:
packages/cli/src/ui/types.ts:112 — [review] R1-3 compressionKind stamped only on ephemeral UI items; persistence discards it — still stands, deferred (convergence posture)packages/cli/src/ui/commands/compressFastCommand.ts:141 — [review] D2-1 CompressionMessage renders fast and summarize markers identically; nothing user-visible distinguishes the two pathspackages/cli/src/ui/utils/historyMapping.test.ts:355 — [probe] D2-2 no test pins the summarize->fast marker ordering; a plausible variant passes all 31 existing tests yet breaks it
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
未探索到全部深度（达到工具调用预算）："agent 1c"：confirm whether the TUI can submit a user turn whose API Content has no text part (image-only paste / bare @image.png mention), which decides reachability o…。
收敛姿态下延后（第 2 轮，非阻断）——已记录，本轮不要求修改：共 3 条（原文未翻译，列表见上方英文部分）。
— qwen3.8-max via Qwen Code /review (v0.21.13)
…count Microcompaction's media clearing replaces the top-level inlineData/fileData parts of user entries with text placeholders. A media-only user entry (e.g. an image-only ACP prompt) never produced a UI user turn, but once cleared it satisfied isUserTextContent's naive text check, desynchronizing the API prompt count from the UI turn count. Rewinding across a fast marker then returned one prompt early and truncateHistory silently dropped a turn the UI still shows — the same hazard this PR fixes, newly reachable through the cross-fast-marker rewinds it enables. Exclude placeholder-prefixed text parts from the prompt count (entries mixing placeholders with real prompt text still count), and mirror the compressionKind persistence caveat onto ChatCompressionRecordPayload. Addresses review R1-1 (Critical) and R1-3 (minimal option).
yiliang114
requested review from
LaZzyMan,
doudouOUC,
tanzhenxin and
wenshao
as code owners
August 17, 2026 09:43
yiliang114
enabled auto-merge
August 17, 2026 09:59
Collaborator
Author
Local E2E verification (macOS) — before / afterRan the manual repro from the Reviewer Test Plan against two bundled builds ( Scenario: 3 short turns →
Both builds render an identical TUI throughout the flow; the loss surfaces only in the request body, which is why the bug stays silent until the next request — consistent with the #9320 report. This also covers the macOS column: ✅ verified end-to-end on macOS. |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Partially reviewed — gaps disclosed. Suggestions are inline.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
中文说明
仅完成部分审查，审查缺口已披露。 建议见行内评论。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
— qwen3.8-max via Qwen Code /review (v0.21.13)
- R3-1: exclude media-clear placeholders by the full generated shape (`[Old inline media cleared: <mime>]`) instead of the bare prefix, so a genuine user prompt that merely begins with the prefix still counts as a prompt. Adds a shared isClearedMediaPlaceholder() helper next to the producer constant and regression tests for both the truncation-late and spurious-compressed-error witness cases. - R3-2: document why the ACP session's private #isUserTextContent twin deliberately keeps the prefix check (ACP rewind maps against per-prompt file-history snapshots, which include media-only prompts). - R3-3: drop the zero-consumer MICROCOMPACT_CLEARED_MESSAGE and MICROCOMPACT_CLEARED_IMAGE_PREFIX barrel exports added by the prior commit; export only the new isClearedMediaPlaceholder() helper.
Collaborator
Author
|
Closeout for the R3 review round (pushed 6dd5b6b):
Verification: core typecheck clean, touched files lint-clean, microcompact (79) and historyMapping (35) suites green; local review converged with no evidence-backed findings. All three threads resolved. |
qwen-code-ci-bot
requested changes
Aug 17, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Not explored to full depth (tool budget reached): "agent 1a": run the PR's new vitest suites (historyMapping/compress/microcompact) — the review worktree has no node_modules installed and no parent checkout node_modules to….
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
未探索到全部深度（达到工具调用预算）："agent 1a"：run the PR's new vitest suites (historyMapping/compress/microcompact) — the review worktree has no node_modules installed and no parent checkout node_modules to…。
— qwen3.8-max via Qwen Code /review (v0.21.13)
sanitizeMimeForPlaceholder returns '' for empty/whitespace-only/ bracket-only mimeTypes and the producer's ?? fallback only covers null/undefined, so microcompaction can emit `[Old inline media cleared: ]`. The consumer regex required one-or-more mime chars and rejected that shape, leaving a cleared media-only entry counted as a genuine prompt and desynchronizing the rewind prompt count. Accept the empty-mime shape and pin the producer/consumer round-trip for degenerate mimeTypes. Also correct the ACP-twin note in historyMapping.ts (the twin keeps the bare text-presence check, not a prefix check) and document the deliberate divergence reciprocally on Session.ts #isUserTextContent, with a Session.test.ts pin that cleared placeholders stay rewindable on the ACP side.
Collaborator
Author
|
Patrol closeout for the new review findings on 6dd5b6b (fixed in 5298003):
Verification: microcompact 81/81, historyMapping 35/35, Session |
…nd mapping Records the known exact-match collision corner next to the media-clear placeholder exclusion and pins its fail-safe behavior with a test: a genuine prompt whose entire text equals a generated placeholder is excluded from the API prompt count, so later rewind targets abort loudly (-1) instead of truncating against a misaligned count. A durable fix needs a structural sentinel on cleared parts, which changes the persisted API history shape and stays out of scope for this PR.
qwen-code-ci-bot
requested changes
Aug 17, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Deferred under the convergence posture (round 5, not a blocker) — recorded, not requested in this round:
packages/cli/src/ui/utils/historyMapping.ts:111 — [probe] /restore checkpoints reinstate kind-less legacy fast markers misclassified as summarizing boundaries — silent-loss walk reproduced for restored sessions (probe: returns 1, expected 3…
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
收敛姿态下延后（第 5 轮，非阻断）——已记录，本轮不要求修改：共 1 条（原文未翻译，列表见上方英文部分）。
— qwen3.8-max via Qwen Code /review (v0.21.13)
Address the round-5 review thread within the scope that is not pending the maintainer structural decision: - The media-clear placeholder regex no longer accepts \r/\n/\t interiors: sanitizeMimeForPlaceholder normalizes them to spaces, so the producer can never emit them, and accepting them let multi-line user text that merely starts with the prefix be misclassified as a cleared-media placeholder. Every producible shape (including the empty-mime one) still matches. - The exact-match collision doc no longer claims "never silent history loss": mid-history, rewinding TO a colliding turn truncates one turn late (its prompt+response stays in model context while the UI removes the turn - under-deletion, not loss of context the UI keeps). The mid-history shape is now pinned next to the first-turn pin so a future structural sentinel updates both expectations.
qwen-code-ci-bot
requested changes
Aug 18, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Deferred under the convergence posture (round 6, not a blocker) — recorded, not requested in this round:
packages/core/src/services/microcompaction/microcompact.ts:29 — [probe] CLEARED_MEDIA_PLACEHOLDER_RE's leading ^ anchor is pinned by no test — the anchor-removal mutant survives all 82+37 predicate/walk testspackages/cli/src/ui/utils/historyMapping.ts:111 — [probe] /restore checkpoints reinstate kind-less legacy fast markers misclassified as summarizing boundaries — silent-loss walk reproduced for restored sessions (round-5 deferral, still stan…
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
收敛姿态下延后（第 6 轮，非阻断）——已记录，本轮不要求修改：共 2 条（原文未翻译，列表见上方英文部分）。
— qwen3.8-max via Qwen Code /review (v0.21.13)
Collaborator
Author
|
@qwen-code /review --timeout=360 |
qwen-code-ci-bot
requested changes
Aug 18, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Deferred under the convergence posture (round 7, not a blocker) — recorded, not requested in this round:
packages/core/src/services/microcompaction/microcompact.ts:29 — [probe] CLEARED_MEDIA_PLACEHOLDER_RE interior accepts '[' / leading-whitespace / >128-char interiors the producer cannot emit — collision corner wider than the documented emitt…
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
收敛姿态下延后（第 7 轮，非阻断）——已记录，本轮不要求修改：共 1 条（原文未翻译，列表见上方英文部分）。
— qwen3.8-max via Qwen Code /review (v0.21.13)
qwen-code-ci-bot
requested changes
Aug 18, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Not explored to full depth (tool budget reached): "agent 6b": verify whether extractTextFromParts (resumeHistoryUtils.ts:50) strips <system-reminder> blocks — decides whether a [reminder, media] record rebuilds with …; "agent 6b": confirm the TUI session picker offers ACP-originated sessions for resume (cross-origin reachability for Finding 2).
Deferred under the convergence posture (round 8, not a blocker) — recorded, not requested in this round:
packages/core/src/services/microcompaction/microcompact.ts:33 — [probe] CLEARED_MEDIA_PLACEHOLDER_RE accepts '[' / >128-char / untrimmed interiors the producer cannot emit — collision surface wider than the documented limitationpackages/core/src/services/chatRecordingService.ts:484 — [review] new NOTE's 'exists only on ephemeral UI items today' claim is wrong — compression markers persist via outputHistoryItems and survive resume with kind intactpackages/cli/src/ui/utils/historyMapping.test.ts:355 — [probe] no test pins the summarize→fast marker ordering; a last-marker-of-any-kind mutant passes all 37 existing testspackages/core/src/services/microcompaction/microcompact.test.ts:171 — [probe] the ^ anchor and \] exclusion of CLEARED_MEDIA_PLACEHOLDER_RE are pinned by no negative test; two one-line mutants survive the suite
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
未探索到全部深度（达到工具调用预算）："agent 6b"：verify whether extractTextFromParts (resumeHistoryUtils.ts:50) strips <system-reminder> blocks — decides whether a [reminder, media] record rebuilds with …；"agent 6b"：confirm the TUI session picker offers ACP-originated sessions for resume (cross-origin reachability for Finding 2)。
收敛姿态下延后（第 8 轮，非阻断）——已记录，本轮不要求修改：共 4 条（原文未翻译，列表见上方英文部分）。
— qwen3.8-max via Qwen Code /review (v0.21.13)
… compressed prefix (#9437 entrance 3)
Collaborator
Author
|
R5-1 entrance (3) is now closed in |
Collaborator
Author
|
@qwen-code /triage |
Collaborator
|
Sandboxed verification: The verification job did not complete (checkout, runner, or setup error) and produced no report. See the workflow run for details. 中文 — 判定：
|
qwen-code-ci-bot
requested changes
Aug 19, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Deferred under the convergence posture (round 9, not a blocker) — recorded, not requested in this round:
packages/core/src/services/microcompaction/microcompact.ts:30 — [probe] CLEARED_MEDIA_PLACEHOLDER_RE accepts '[' and unbounded interiors the producer can never emitpackages/core/src/services/microcompaction/microcompact.test.ts:171 — [probe] the regex's leading ^ anchor is pinned by no negative test; the anchor-removal mutant survives the suitepackages/cli/src/ui/utils/historyMapping.ts:177 — [review] the new guard's -1 surfaces as the generic 'turn that was compressed' message, misleading for marker-less auto-compactionpackages/cli/src/acp-integration/session/Session.ts:3589 — [probe] the ACP twin's turn-0 shortcut has no counterpart to the new TUI fail-loud guard; probe-verified silent truncation
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
收敛姿态下延后（第 9 轮，非阻断）——已记录，本轮不要求修改：共 4 条（原文未翻译，列表见上方英文部分）。
— qwen3.8-max via Qwen Code /review (v0.21.14)
Comment on lines
+113
to
+114
| item.compression.compressionStatus === CompressionStatus.COMPRESSED && | ||
| item.compression.compressionKind !== 'fast', |
Collaborator
There was a problem hiding this comment.
[Critical] R5-1: Round-9 status of the class finding — the rewind mapping still re-derives UI/API alignment from two independent representations (turn count vs prompt count, marker boundaries, text-shape matching), and the surface closes only structurally. The new commit acbfaf2 closed entrance (3)'s first-turn variant — the uiUserTurnCount === 0 shortcut now fails loud when a compressed prefix has no UI boundary, and the added test pins it — but three entrances remain open at this commit, each probe-verified this round:
(1) The same marker-less auto-compaction hazard on the sibling walk branch — the new guard covers only the shortcut. After auto-compaction absorbs M turns without a UI marker, the user sends K ≥ 2 new prompts and rewinds to an absorbed turn j (2 ≤ j ≤ K): the walk returns a silent index ≥ 0, so the UI truncates to the old turns while the model keeps [prelude, summary, ack, new prompts] — context silently diverges from everything displayed.
(2) Resumed cross-surface sessions — a media-only prompt cleared by /compress-fast is rebuilt by TUI resume as a counted [User message with attachments] turn while the rebuilt API entry is the placeholder now excluded from the walk, so the rewind mapping desyncs by one: rewinding to a later turn truncates one turn late (keeping prompt+response the UI removed), and rewinding to the last turn aborts with a spurious "was compressed" error.
(3) Restored kind-less legacy fast markers — resume reinstates recorded slash-command outputHistoryItems verbatim, so pre-PR /compress-fast markers (shipped since v0.18.0) come back without compressionKind and act as summarizing boundaries: rewinding a post-marker turn silently truncates early (observed index 0 where 2 is expected — dropping everything while the UI still shows the turns), and the kind-less marker bypasses the new guard because compressionIndex ≠ -1. The NOTE added on ChatCompressionRecordPayload warns about a reconstruction that already happens through this path.
Witness (probes against the real code at this commit):
(1) rewind to second absorbed turn (walk branch, unguarded):
AssertionError: expected 5 to be -1
hoisting the guard above the shortcut branch: 41/41 tests pass
(2) rewind to first text turn: 4 (pre-PR counting model: 2)
rewind to last turn: -1 (pre-PR: 4)
(3) restored marker: {compressionStatus: 1, no compressionKind}
first post-marker rewind: 0 (expected 2)
second post-marker rewind: 2 (expected 4)
Rounds 1–9 have each surfaced a state of this mechanism; per-entrance patches do not converge it. The structural closure (single authoritative source / stable-identity mapping) is tracked in #9437. If the rework stays out of scope, the probes validate per-entrance stopgaps: hoist the marker-less guard above the uiUserTurnCount === 0 branch so the walk fails loud too; exclude rebuilt attachment-only turns (no promptId) from the rewind count or mirror the placeholder exclusion on the UI side; treat a kind-less marker as a boundary only when the API history actually carries a compressed summary prefix.
中文说明
[严重] R5-1：类发现的第 9 轮状态——回退映射仍然从两个独立表示重新推导 UI/API 对齐（轮次计数对 prompt 计数、标记边界、文本形态匹配），该表面只能从结构上关闭。新提交 acbfaf2 关闭了入口 (3) 的第一轮变体——当存在压缩前缀而 UI 无摘要边界时，uiUserTurnCount === 0 捷径现在响亮失败，新增测试已钉住——但本提交上仍有三个入口敞开，均在本轮经探针验证：
(1) 同一"无标记自动压缩"危害存在于同函数的遍历分支——新守护只覆盖捷径。自动压缩吸收 M 轮（无 UI 标记）后，用户发送 K ≥ 2 个新 prompt 并回退到被吸收的轮次 j（2 ≤ j ≤ K）：遍历返回静默的 ≥ 0 索引，UI 截断到旧轮次，而模型保留 [prelude, summary, ack, 新 prompt]——上下文与界面显示静默背离。
(2) 恢复的跨表面会话——被 /compress-fast 清理的纯媒体 prompt 在 TUI 恢复时被重建为计数的 [User message with attachments] 轮次，而重建后的 API 条目是被本次排除出遍历的占位符，回退映射因此错位一轮：回退到更晚轮次时延迟一轮截断（保留 UI 已移除的 prompt+回复），回退到最后一轮时报出虚假的"已压缩"错误。
(3) 还原的无 kind 旧版 fast 标记——会话恢复会原样还原记录的 slash 命令 outputHistoryItems，因此 PR 之前产生的 /compress-fast 标记（自 v0.18.0 发布）还原后不带 compressionKind，被当作摘要式边界：回退到标记后轮次时静默提前截断（实测索引 0，期望 2——UI 仍显示这些轮次时丢弃全部内容），且无 kind 标记因 compressionIndex ≠ -1 绕过新守护。ChatCompressionRecordPayload 上新增的 NOTE 警告的重建其实已经通过该路径发生。
证据（在本提交上对真实代码的探针）：(1) 回退到第二个被吸收轮次（遍历分支，无守护）：AssertionError: expected 5 to be -1；把守护提升到捷径分支之上后 41/41 测试通过。(2) 回退到第一个文本轮次：4（改动前计数模型为 2）；回退到最后一轮：-1（改动前为 4）。(3) 还原的标记：{compressionStatus: 1, 无 compressionKind}；标记后第一轮回退：0（期望 2）；标记后第二轮回退：2（期望 4）。
第 1–9 轮每轮都发现该机制的一种状态；逐入口修补并不收敛。结构性关闭（单一权威来源/稳定标识映射）已在 #9437 跟踪。若重构暂不在范围内，探针已验证各入口的缓解措施：把无标记守护提升到 uiUserTurnCount === 0 分支之上使遍历也响亮失败；把重建出的仅附件轮次（无 promptId）排除出回退计数，或在 UI 侧镜像占位符排除；仅当 API 历史确实携带压缩摘要前缀时才把无 kind 标记当作边界。
— qwen3.8-max via Qwen Code /review (v0.21.14)
qwen-code-ci-bot
approved these changes
Aug 19, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
LGTM, looks ready to ship. ✅
doudouOUC
approved these changes
Aug 19, 2026
doudouOUC
left a comment
doudouOUC
left a comment
Collaborator
There was a problem hiding this comment.
Review summary
14 review agents examined the full diff (13 files, +551/-2, 217 source lines). Coverage: 3/3 chunks reviewed, all 14 agents completed work, no budget gaps.
Findings
No Critical findings. All code paths are correctly implemented, well-tested, and follow existing conventions. The two Critical findings from Agent 6a (attacker mindset) were verified against the producer code and determined to be false positives: microcompaction keeps the functionResponse part structure rather than replacing it with a text part, so the hasFunctionResponse guard correctly catches cleared tool result entries.
Verification
- Build (packages/core): Completed successfully
- Build (packages/cli): Blocked by pre-existing audio-capture node-gyp failure (missing Python) — not a PR defect
- Tests: Not run due to the pre-existing build chain blockage
Verdict
This PR correctly fixes the root cause of #9320 (silent context loss after compress-fast + rewind). The remaining structural class finding about the dual-representation rewind mapping is a known larger rework deferred to #9437.
Collaborator
|
Released in v0.21.15. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this PR does
/rewindmaps the selected UI turn to an API-history truncation point viacomputeApiTruncationIndex, and it treats every successful compression marker as a summarizing boundary./compress-fastemits the sameCompressionStatus.COMPRESSEDmarker, but it is purely rule-based: it removes no user prompts from the API history and inserts no summary prefix. So rewinding to any turn after a/compress-fastmarker collapsed the truncation anchor to just the startup context entry and silently dropped the entire pre-marker conversation — the UI history still rendered everything, so nothing looked wrong until the next request.This PR tags compression UI items with the path that produced them (
CompressionProps.compressionKind: 'summarize' | 'fast', set by/compressand/compress-fastrespectively) and makes the rewind mapping ignore'fast'markers as boundaries — turns are then mapped against the full history exactly as if no compression happened. Legacy markers without a kind keep the previous behavior.Why it's needed
Silent context loss. After
/compress-fast, rewinding to a post-marker turn truncates the model history to the startup context plus the re-entered prompt; everything before is gone without any warning. Reported in #9320 with llama.cpp prefill logs: resumed context was ~27k tokens instead of the expected ~87k.Reviewer Test Plan
How to verify
cd packages/cli && npx vitest run src/ui/utils/historyMapping.test.ts— the newwith fast (non-summarizing) compression markersblock covers the reported shape (rewind to first/later post-marker turn keeps the full pre-marker history, pre-marker turns stay reachable, a later summarizing compression still absorbs turns, legacy markers behave as before). Mutation check: reverting only thehistoryMapping.tspredicate makes 3 of the new tests fail.compressCommand.test.ts/compressFastCommand.test.tsassert the newcompressionKindon the added compression item./compress-fast→/rewindto the first turn after the compression entry → before this fix the next request carries only startup context + the new prompt; after it, the full pre-marker history is retained.Evidence (Before & After)
N/A — no rendering change; the compression banner and rewind selector UI are untouched. Behavior verified at the mapping function, which is where the loss originates.
Tested on
Environment (optional)
Unit tests only (
vitest runonpackages/cli).Risk & Scope
compressionKindand keep being treated as summarizing boundaries (today's behavior) — only markers created by this build onward get the corrected mapping.Session.#computeApiTruncationIndexForUserTurn) maps purely from API history and is not affected; file-history rewind interplay is unchanged.Linked Issues
Fixes #9320
Related: #9309 (same reporter; diagnosed in-thread as expected estimator-vs-API banner scaling, not context loss — the rewind interaction fixed here was the actual loss path).
中文说明
本 PR 做了什么
/rewind通过computeApiTruncationIndex把选中的 UI 轮次映射为 API 历史的截断点，并把每一个成功的压缩标记都当作"摘要式压缩"的边界。/compress-fast会发出相同的CompressionStatus.COMPRESSED标记，但它是纯规则压缩：既不会从 API 历史中移除任何用户 prompt，也不会插入摘要前缀。因此，回退到/compress-fast标记之后的任意轮次时，截断锚点会坍缩到只剩启动上下文条目，整段标记前的对话被静默丢弃——UI 历史仍然完整渲染，所以在下一次请求之前看不出任何异常。本 PR 为压缩 UI 条目打上产生来源标记（
CompressionProps.compressionKind: 'summarize' | 'fast'，分别由/compress与/compress-fast设置），并让回退映射忽略'fast'标记的边界作用——轮次映射将基于完整历史进行，等同于没有发生过压缩。没有该字段的旧标记保持原有行为。为什么需要
静默上下文丢失。
/compress-fast之后回退到标记后的轮次，会把模型历史截断到只剩启动上下文 + 重新输入的 prompt，之前的全部内容无声丢失。#9320 附带 llama.cpp prefill 日志：恢复后的上下文只有约 27k token，而预期是约 87k。评审验证计划
如何验证
cd packages/cli && npx vitest run src/ui/utils/historyMapping.test.ts——新增的with fast (non-summarizing) compression markers用例覆盖报告中的形态（回退到标记后第一/后续轮次时保留完整标记前历史、标记前轮次仍可回退、其后的摘要式压缩仍会吸收轮次、旧标记行为不变）。变异验证：仅还原historyMapping.ts中的判定条件，新增用例中有 3 个会失败。compressCommand.test.ts/compressFastCommand.test.ts断言了新增的compressionKind字段。/compress-fast→/rewind选择压缩条目后的第一个轮次 → 修复前下一次请求只带启动上下文 + 新 prompt；修复后完整保留标记前历史。前后证据
N/A——无渲染变化；压缩横幅与回退选择器 UI 均未改动。行为在映射函数层验证，丢失正是发生在该处。
测试环境
环境（可选）
仅单测（
packages/cli下vitest run）。风险与范围
compressionKind，仍按摘要式边界处理（即现状行为）——只有本版本之后产生的标记才获得修正后的映射。Session.#computeApiTruncationIndexForUserTurn）纯粹基于 API 历史映射，不受影响；文件历史回退的交互未改动。关联 Issue
Fixes #9320
相关：#9309（同一 reporter；线程中已诊断为估算值与 API 计数的口径差异，并非上下文丢失——本 PR 修复的回退交互才是真正的丢失路径）。