feat(autofix): stop a PR that fails to push for N rounds in a row - #7482
Merged
Conversation
Under takeover the round cap is 100, which is right for a PR that needs many PRODUCTIVE rounds. It is wrong for one that fails every round: #6723 ran 7 consecutive failed rounds (3 agent timeouts at 50 min, 4 gate rejections whose fix broke tests) over 8 hours, heading for round 100, because it is a 5700-line, 47-file, 5-day-old PR racing a fast-moving main — every round re-resolves a conflict it cannot finish or that fails the gate. Retrying at the same per-round budget will not converge; a human has to rebase or split it. Adds CONSECUTIVE_FAILURE_CAP (5), distinct from the total round cap. The handoff step already runs only when a round did NOT push, so it counts the unbroken run of prior failure markers — stopping at the first push ("Addressed the latest review feedback") or legitimate no-op ("no changes needed"), either of which proves progress and resets the streak. At the cap it forces the terminal round even under takeover, with a handoff that names the real fix (rebase/split, then /retry). Cause- agnostic: a timeout and a gate rejection both count.
Collaborator
Author
|
@qwen-code /takeover |
qwen-code-dev-bot
added
the
autofix/takeover
Collaborator
|
🤝 Takeover engaged: the autofix loop now manages this PR — it will address new review feedback and resolve base conflicts until the label is removed or the round cap is reached. Remove the 中文说明🤝 已接管：autofix 循环现在管理此 PR —— 将持续处理新的评审反馈与 base 冲突，直到移除标签或达到轮次上限。移除 |
Collaborator
|
Thanks for the PR! Template looks good ✓ — all sections present, bilingual, with a concrete reviewer test plan. Problem: observed, not theoretical. The description documents a real case (#6723): seven straight failed rounds under takeover — three ~50-min agent timeouts plus four gate rejections — over eight hours with nothing pushed, heading for round 100. The round cap bounds productive iteration; nothing today bounds an unbroken run of failures. That gap is real. Direction: aligned. This is autofix-loop reliability infrastructure, squarely in the same line as #7330 / #7351 / #7368 / #7412 / #7416 / #7438. It touches no auth, sandbox, model-selection, telemetry, or public-contract surface — just the takeover handoff step. Size: not applicable — no core-module paths. Two files (a workflow + its test), 142+/1-. Approach: the scope feels right and the diff is minimal. A consecutive-failure sub-cap that resets on any push or legitimate no-op is about as simple as this gets given that GitHub Actions runs are stateless — reconstructing the streak from the bot's own eval-comment history is the natural move, and forcing the existing terminal-marker path ( Moving on to code review. 🔍 中文说明感谢贡献！ 模板完整 ✓ —— 各节齐全、双语、并附带具体的评审验证方案。 问题： 已观测到，非理论性。PR 描述了真实案例（#6723）：在 takeover 下连续 7 轮失败——3 次约 50 分钟的 agent 超时加 4 次门拒绝——历时 8 小时零推送，正冲着 round 100 去。轮次上限约束的是有产出的迭代；而目前没有任何机制约束连续不断的失败。这个缺口是真实存在的。 方向： 对齐。这是 autofix 循环的可靠性基础设施，与 #7330 / #7351 / #7368 / #7412 / #7416 / #7438 同属一条主线。它不触及 auth、sandbox、模型选择、telemetry 或公共契约——只动 takeover 的交接步骤。 规模： 不适用 —— 未触及核心模块路径。两个文件（workflow + 其测试），142+/1-。 方案： 范围合理，diff 极简。在 GitHub Actions 运行无状态的前提下，一个"任意推送或合法 no-op 即重置"的连续失败子上限,已是最简方案——从 bot 自己的 eval 评论历史重建连续段是自然选择,而强制走既有的终态标记路径（ 进入代码审查 🔍 — Qwen Code · qwen3.8-max-preview 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. |
Collaborator
Code reviewMy independent take before reading the diff: with GitHub Actions runs being stateless, the only durable record of prior rounds is the bot's own eval comments, so I'd reconstruct the failure streak from those, cap it around 5, reset on any sign of progress (a push or a deliberate no-op), and force the existing terminal path so future scans skip. The PR does essentially this, and does it cleanly — it reuses the proven I verified the load-bearing assumptions against the actual workflow, not just the diff:
One substantive observation (non-blocking). The streak is counted across all of the bot's eval markers — the One minor robustness nit. The reset keys off hardcoded headline substrings, and the new test hardcodes those same strings rather than extracting them from the workflow — so a future reword of the push/no-op headline could silently stop resetting the streak without the test catching it. Low priority. Real-scenario testingThis change is entirely inside a GitHub Actions workflow step — there's no CLI/TUI surface, so a All six documented scenarios behave as claimed, plus fail-open on API error. The re-arm case that surfaces the observation above: I also ran the PR's own suite and the static checks the description claims, in a clean checkout at the reviewed commit: 92/92 green (the 中文说明代码审查读 diff 前我的独立想法：由于 GitHub Actions 各轮运行无状态,先前轮次唯一持久记录就是 bot 自己的 eval 评论,因此我会从这些评论重建失败连续段、上限设 5 左右、在任意进展信号（推送或刻意 no-op）处重置、并强制走既有终态路径让后续扫描跳过。PR 基本如此,且做得干净——复用久经验证的 我对照真实 workflow（而非仅凭 diff）核验了关键假设：
一个实质性观察（非阻塞）。 连续段统计遍历 bot 的所有 eval 标记—— 一个小健壮性瑕疵。 重置依赖硬编码的标题子串,而新测试也硬编码了同样的字符串、并非从 workflow 提取——因此未来若改写推送/no-op 标题,可能悄悄不再重置连续段而测试无法捕获。优先级低。 真实场景测试本改动完全位于一个 GitHub Actions workflow 步骤内——没有 CLI/TUI 界面,因此在 tmux 里跑 （回放输出见上方英文代码块：六种已记录场景全部如预期,外加 API 错误 fail-open;re-arm 场景见第二段回放。） 我还在被审 commit 的干净检出中跑了 PR 自带测试套件与描述声称的静态检查：92/92 全绿（描述提醒的 — Qwen Code · qwen3.8-max-preview Reviewed at |
Collaborator
|
Confidence: 4/5 — a focused, well-tested fix for a real, evidenced problem; the only reservation is a design-consistency question (re-arm window scoping) that I can't resolve from the diff alone but that doesn't break the intended recovery path. Stepping back: this earns its place. The motivating case is concrete (#6723 burning seven agent rounds under takeover with nothing pushed), the fix is the minimal thing that addresses it — a consecutive-failure sub-cap that reuses the existing terminal-marker path — and it's careful where it matters: it resets on any real progress (push or deliberate no-op) so a healthy iterating PR never trips it, it's cause-agnostic (a timeout and a gate rejection both count, because both are "failed, nothing pushed"), and it fails open if the comment fetch dies. My independent proposal and the PR converged on the same approach, which is usually a sign the design is the obvious-right one. The test is the strongest part. It replays the actual extracted bash block under a stubbed The one thing I'd want the author or a maintainer to settle, named in the review above: the streak isn't scoped to the re-arm window the way the round counter is, so 中文说明置信度：4/5 —— 针对一个真实且有证据的问题、聚焦且测试充分的修复;唯一的保留是一个我无法仅凭 diff 判定的设计一致性疑问（re-arm 窗口限定),但它不破坏预期的恢复路径。 退一步看：这个 PR 有其价值。动机案例具体（#6723 在 takeover 下烧掉 7 个 agent 轮次却零推送),修复是解决它的最小手段——一个复用既有终态标记路径的连续失败子上限——并且在关键处很谨慎：在任意真实进展（推送或刻意 no-op）处重置,因此健康迭代的 PR 永不触发;不区分成因（超时与门拒都算,因为两者都是"失败且未推送");评论拉取失败时 fail-open。我的独立方案与 PR 收敛于同一做法,这通常意味着设计就是那个"显然正确"的选择。 测试是最强的部分。它在桩 唯一希望作者或 maintainer 拍板的点（已在上方审查中点名）：连续段没有像轮次计数器那样限定到 re-arm 窗口,因此 — Qwen Code · qwen3.8-max-preview Reviewed at |
qwen-code-ci-bot
previously approved these changes
Jul 22, 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. ✅
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Reviewed — no blockers. Suggestions are inline.
— qwen3.7-max via Qwen Code /review
…breaker (#7482) - Fix misleading comment: the walk is oldest-first (API order) with reset-on-success, not newest-first with early stop - Prefer the already-fetched ic.json over a redundant gh api call, falling back to the API only when the file is missing - Filter eval markers by re-arm window (win=) so pre-re-arm failures do not immediately re-terminate a re-armed PR - Add test coverage for the MARK_ROUND == MAX_ROUNDS guard and for window-scoped streak counting
Collaborator
|
🤖 Addressed the latest review feedback (round 1/100). What changed, and what I pushed back on: Review feedback addressedAll four suggestions from the automated reviewer were implemented in a single commit. 1. Misleading comment (rc:3627860245) — FixedThe comment described the walk as "newest-first" with an early stop, but the code iterates oldest-first (GitHub API default order) with reset-on-success and no 2. Redundant
|
Collaborator
|
🤖 Reviewed the latest feedback — no changes needed. Why, point by point: No changes neededThe latest evaluation found no actionable feedback on PR #7482:
The only signal was a Verification: all 92 tests in 中文说明无需修改最新一轮评估未发现 PR #7482 有任何需要处理的反馈：
唯一的信号是 验证： 当前 HEAD（ Base-conflict check: no conflict with main. 🧠 Handled by Qwen Code · model/模型 |
Collaborator
Author
Code review — consecutive-failure sub-capVerified the mechanics by extracting the new block from the PR head and replaying it under 1. It silently overrides the "transient model error keeps the full budget" decision right above itThe breaker is cause-agnostic, and the mixed-cause test confirms That path produces a non-terminal RETRY headline ( Two clean resolutions — either is fine, but it's worth a conscious call:
2. The reset detector ↔ emit-headline coupling is untestedThe reset keys on the literal substrings 3. Minor
Correctness, fail-open behavior, window scoping, and test fidelity (real extracted block replayed under bash) all check out. #1 is the only one I'd want resolved before merge; #2/#3 are hardening. |
Collaborator
|
🤖 Addressed the latest review feedback (round 2/100). What changed, and what I pushed back on: Addressed review feedback on the consecutive-failure circuit breaker1. Transient model errors now exempt from the breaker (Required)Feedback: The breaker is cause-agnostic and counts transient model errors (429/5xx) toward the streak, but the sibling Decision: Implemented the exemption. The breaker's Changes:
2. Reset detector ↔ emit-headline coupling now pinned by test (Suggestion)Feedback: The reset keys on literal substrings Decision: Implemented. Added test assertions that match the actual 3a. Explicit
|
qwen-code-ci-bot
approved these changes
Jul 22, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
No issues found. LGTM! ✅
— qwen3.7-max via Qwen Code /review
yiliang114
approved these changes
Jul 22, 2026
yiliang114
left a comment
yiliang114
left a comment
Collaborator
There was a problem hiding this comment.
LGTM. Clean circuit breaker design — the transient-error exemption is well-reasoned (provider outage must not mass-terminate PRs), and the streak reset on push/no-op is correct. Test coverage is exceptional: 10+ cases including window filtering, auth vs transient distinction, and pinning the reset substrings to actual emit lines.
wenshao
pushed a commit
that referenced
this pull request
Jul 22, 2026
…block CI reddened `retries a skipped-Prepare` after main's consecutive-failure cap (#7482) merged into this branch: that block was inserted between this decision block and the report `{`, and it calls `gh api`. The test's `{`-anchored regex over-captured through it, so the extracted script ran the unstubbed `gh api` and failed. Anchor the end on the same `# Consecutive-failure` comment the sibling gate-crash test already uses, so the extraction stops at this decision block's own closing `fi`.
pull Bot
pushed a commit
to bit-cook/qwen-code
that referenced
this pull request
Jul 23, 2026
…minal (QwenLM#7490) * fix(autofix): retry a skipped-Prepare instead of stranding the PR terminal A base/infra failure BEFORE the agent runs was misread as an agent crash and terminated the PR forever. When an early step fails — installing or building the trusted base, checkout, node setup — the `Prepare branch and feedback` step is skipped, so NEWEST is empty, and the report step's "crashed before reading feedback" branch fired: MARK_ROUND=MAX_ROUNDS, terminal, scan skips it on every future tick. Observed: a web-shell TypeScript break on `main` failed `Install dependencies and build` (which builds the trusted base) across a whole scan batch, and SIX healthy PRs were stranded terminal at round=100 in one run — including ones at round 9 and 11 that had nothing to do with the break. `round=100` there is a terminal sentinel, not 100 attempts. NEWEST-empty now splits on steps.prepare.outcome: - 'skipped' (an earlier step failed, the agent never ran) is infra/base and transient: retry with a sentinel ts so the feedback stays live, incrementing the round so a PERSISTENTLY broken base is still bounded and stops at the cap (recoverable with /retry). - 'success'/'failure' (Prepare ran, no feedback produced) is a genuine pre-read agent crash: unchanged terminal behaviour. This is the reverse of the asymmetry QwenLM#7482 addresses: that bounds a crash AFTER reading that retried forever; this stops a transient failure BEFORE reading from going terminal after one. * docs(autofix): note a pre-Prepare cancel also retries intentionally (QwenLM#7490) * fix(autofix): also retry a cancelled/empty prepare outcome, not just skipped A previous review comment on this PR noted that a job cancelled before Prepare should retry too. It was right about the intent but the code did not do it: `steps.prepare.outcome` is 'cancelled' for a cancel and '' for a job that stopped before Prepare entered the step context — both DISTINCT from 'skipped', so `== 'skipped'` sent them to the terminal branch, the same over-termination this PR exists to fix. Match on "not a real Prepare run" (`!= 'success' && != 'failure'`) instead, so skipped, cancelled, and empty all retry; only a Prepare that actually ran to a verdict (success/failure) with no feedback stays terminal — the genuine pre-read agent crash. Test extended to drive the cancelled and empty cases (retry) and both real-run outcomes (terminal); mutation-verified that reverting to `== 'skipped'` reddens the cancelled case. * test(autofix): update the pre-read-crash case for the broadened retry The prior commit broadened NEWEST-empty retry to skipped/cancelled/empty but left the older 'replays the handoff decision' test asserting the old terminal behaviour for an unset PREPARE_OUTCOME (which now retries). That test's terminal cases now set PREPARE_OUTCOME=success/failure explicitly — the only outcomes that still terminate — so it exercises the genuine pre-read agent crash rather than the infra/cancel path. * test(autofix): anchor the skipped-Prepare extraction past the CONSEC block CI reddened `retries a skipped-Prepare` after main's consecutive-failure cap (QwenLM#7482) merged into this branch: that block was inserted between this decision block and the report `{`, and it calls `gh api`. The test's `{`-anchored regex over-captured through it, so the extracted script ran the unstubbed `gh api` and failed. Anchor the end on the same `# Consecutive-failure` comment the sibling gate-crash test already uses, so the extraction stops at this decision block's own closing `fi`. * fix(autofix): exempt skipped-Prepare from the consecutive-failure breaker A broken base build skips Prepare, producing no API error file — so the consecutive-failure breaker ran on the new retry path and, after 5 scans, re-introduced the exact mass-stranding this PR exists to prevent. Exempt pre-agent infra failures (skipped/cancelled/empty outcome) from the breaker, mirroring the transient 429/5xx exemption: same failure class (not the PR's fault, self-heals, hits the whole batch). The round cap + sentinel-ts /retry recovery already bounds a persistently broken base. Also trim "checkout" from the retry headlines (checkout failures do not land in this branch) and hoist the duplicated MARK_TS assignment. * fix(autofix): reset the consecutive-failure streak on prior infra-failure markers The streak walker counted prior infra-failure headlines ("AutoFix could not start —…") as failures, inflating the consecutive-failure count on subsequent rounds. A PR with 3 real agent failures, then 3 rounds of base-build infra failures, then 1 more real failure would trip the cap-5 breaker even though only 4 rounds were the PR's fault. Add the two infra-failure headline patterns as reset strings in the streak walker, alongside the existing push and no-op resets. The genuine agent-crash headline ("AutoFix could not start evaluation —…") is deliberately excluded — it is a real failure and must still count. * fix(autofix): clarify infra-failure headlines and else-branch comment (QwenLM#7490) Address review nits: the retry headline now mentions cancelled runs, the cap headline says 'reached the round cap' instead of overstating 'could not start for N rounds', the else-branch comment says 'prepare itself crashed' instead of 'agent crash', and the streak-reset pattern is simplified now that both infra headlines share the same prefix. --------- Co-authored-by: wenshao <wenshao@example.com> Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com>
chiga0
pushed a commit
that referenced
this pull request
Jul 23, 2026
) * feat(autofix): stop a PR that fails to push for N rounds in a row Under takeover the round cap is 100, which is right for a PR that needs many PRODUCTIVE rounds. It is wrong for one that fails every round: #6723 ran 7 consecutive failed rounds (3 agent timeouts at 50 min, 4 gate rejections whose fix broke tests) over 8 hours, heading for round 100, because it is a 5700-line, 47-file, 5-day-old PR racing a fast-moving main — every round re-resolves a conflict it cannot finish or that fails the gate. Retrying at the same per-round budget will not converge; a human has to rebase or split it. Adds CONSECUTIVE_FAILURE_CAP (5), distinct from the total round cap. The handoff step already runs only when a round did NOT push, so it counts the unbroken run of prior failure markers — stopping at the first push ("Addressed the latest review feedback") or legitimate no-op ("no changes needed"), either of which proves progress and resets the streak. At the cap it forces the terminal round even under takeover, with a handoff that names the real fix (rebase/split, then /retry). Cause- agnostic: a timeout and a gate rejection both count. * fix(autofix): address review feedback on consecutive-failure circuit breaker (#7482) - Fix misleading comment: the walk is oldest-first (API order) with reset-on-success, not newest-first with early stop - Prefer the already-fetched ic.json over a redundant gh api call, falling back to the API only when the file is missing - Filter eval markers by re-arm window (win=) so pre-re-arm failures do not immediately re-terminate a re-armed PR - Add test coverage for the MARK_ROUND == MAX_ROUNDS guard and for window-scoped streak counting * fix(autofix): exempt transient model errors from consecutive-failure breaker (#7482) --------- Co-authored-by: wenshao <wenshao@example.com> Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com>
chiga0
pushed a commit
that referenced
this pull request
Jul 23, 2026
…minal (#7490) * fix(autofix): retry a skipped-Prepare instead of stranding the PR terminal A base/infra failure BEFORE the agent runs was misread as an agent crash and terminated the PR forever. When an early step fails — installing or building the trusted base, checkout, node setup — the `Prepare branch and feedback` step is skipped, so NEWEST is empty, and the report step's "crashed before reading feedback" branch fired: MARK_ROUND=MAX_ROUNDS, terminal, scan skips it on every future tick. Observed: a web-shell TypeScript break on `main` failed `Install dependencies and build` (which builds the trusted base) across a whole scan batch, and SIX healthy PRs were stranded terminal at round=100 in one run — including ones at round 9 and 11 that had nothing to do with the break. `round=100` there is a terminal sentinel, not 100 attempts. NEWEST-empty now splits on steps.prepare.outcome: - 'skipped' (an earlier step failed, the agent never ran) is infra/base and transient: retry with a sentinel ts so the feedback stays live, incrementing the round so a PERSISTENTLY broken base is still bounded and stops at the cap (recoverable with /retry). - 'success'/'failure' (Prepare ran, no feedback produced) is a genuine pre-read agent crash: unchanged terminal behaviour. This is the reverse of the asymmetry #7482 addresses: that bounds a crash AFTER reading that retried forever; this stops a transient failure BEFORE reading from going terminal after one. * docs(autofix): note a pre-Prepare cancel also retries intentionally (#7490) * fix(autofix): also retry a cancelled/empty prepare outcome, not just skipped A previous review comment on this PR noted that a job cancelled before Prepare should retry too. It was right about the intent but the code did not do it: `steps.prepare.outcome` is 'cancelled' for a cancel and '' for a job that stopped before Prepare entered the step context — both DISTINCT from 'skipped', so `== 'skipped'` sent them to the terminal branch, the same over-termination this PR exists to fix. Match on "not a real Prepare run" (`!= 'success' && != 'failure'`) instead, so skipped, cancelled, and empty all retry; only a Prepare that actually ran to a verdict (success/failure) with no feedback stays terminal — the genuine pre-read agent crash. Test extended to drive the cancelled and empty cases (retry) and both real-run outcomes (terminal); mutation-verified that reverting to `== 'skipped'` reddens the cancelled case. * test(autofix): update the pre-read-crash case for the broadened retry The prior commit broadened NEWEST-empty retry to skipped/cancelled/empty but left the older 'replays the handoff decision' test asserting the old terminal behaviour for an unset PREPARE_OUTCOME (which now retries). That test's terminal cases now set PREPARE_OUTCOME=success/failure explicitly — the only outcomes that still terminate — so it exercises the genuine pre-read agent crash rather than the infra/cancel path. * test(autofix): anchor the skipped-Prepare extraction past the CONSEC block CI reddened `retries a skipped-Prepare` after main's consecutive-failure cap (#7482) merged into this branch: that block was inserted between this decision block and the report `{`, and it calls `gh api`. The test's `{`-anchored regex over-captured through it, so the extracted script ran the unstubbed `gh api` and failed. Anchor the end on the same `# Consecutive-failure` comment the sibling gate-crash test already uses, so the extraction stops at this decision block's own closing `fi`. * fix(autofix): exempt skipped-Prepare from the consecutive-failure breaker A broken base build skips Prepare, producing no API error file — so the consecutive-failure breaker ran on the new retry path and, after 5 scans, re-introduced the exact mass-stranding this PR exists to prevent. Exempt pre-agent infra failures (skipped/cancelled/empty outcome) from the breaker, mirroring the transient 429/5xx exemption: same failure class (not the PR's fault, self-heals, hits the whole batch). The round cap + sentinel-ts /retry recovery already bounds a persistently broken base. Also trim "checkout" from the retry headlines (checkout failures do not land in this branch) and hoist the duplicated MARK_TS assignment. * fix(autofix): reset the consecutive-failure streak on prior infra-failure markers The streak walker counted prior infra-failure headlines ("AutoFix could not start —…") as failures, inflating the consecutive-failure count on subsequent rounds. A PR with 3 real agent failures, then 3 rounds of base-build infra failures, then 1 more real failure would trip the cap-5 breaker even though only 4 rounds were the PR's fault. Add the two infra-failure headline patterns as reset strings in the streak walker, alongside the existing push and no-op resets. The genuine agent-crash headline ("AutoFix could not start evaluation —…") is deliberately excluded — it is a real failure and must still count. * fix(autofix): clarify infra-failure headlines and else-branch comment (#7490) Address review nits: the retry headline now mentions cancelled runs, the cap headline says 'reached the round cap' instead of overstating 'could not start for N rounds', the else-branch comment says 'prepare itself crashed' instead of 'agent crash', and the streak-reset pattern is simplified now that both infra headlines share the same prefix. --------- Co-authored-by: wenshao <wenshao@example.com> Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com>
yiliang114
added a commit
to he-yufeng/qwen-code
that referenced
this pull request
Jul 23, 2026
) * fix(cli): correct queued message display style and ordering Mid-turn steer messages (user input queued while the model is responding) had two display bugs: 1. They rendered with notification styling (● icon) instead of user-input styling (> prefix) because accept() added them to UI history as MessageType.NOTIFICATION. 2. They appeared below the model's reply because accept() was only called in the finally block after the entire response stream completed, appending the user message after all model response items. Fix: use MessageType.USER with sentToModel: true for steer messages, and settle the steer input on the first stream event (after the user-content push lands but before model-response events are committed to UI history). Pass steer inputs through to recursive sendMessageStream calls so all takeSteerInput paths benefit from early settlement. Add a WeakSet guard to settleSteerInput for idempotency across recursive invocations. * test(core): add ordering test for early steer settlement Verify that accept() is called after the first stream event is pulled but before subsequent events reach the consumer, pinning the settle-before-content timing that ensures queued user messages render above the model's reply. * fix(cli): use sentToModel: false for steer messages, address review - Use sentToModel: false instead of true: steer messages are injected into an existing tool-result turn, not standalone user turns. sentToModel: true would make isRealUserTurn() count them as real turns, inflating the rewind turn index. - Remove unnecessary as HistoryItemWithoutId cast. - Add post-cleanup assertion in ordering test to verify the WeakSet guard prevents double-settlement. * fix(cli): align resumed mid-turn steer display with live session (#7381) Resume path now renders mid_turn_user_message as MessageType.USER with sentToModel: false, matching the live-session styling. Add a comment documenting the intentional sentToModel: false choice. * fix(cli): exclude steer messages from user-turn filters (#7381) Steer messages (sentToModel: false) were counted as real user turns by five downstream consumers that filter on type === 'user' without checking sentToModel, breaking cancel auto-restore, telemetry turn count, prompt recall, away-recap thresholds, and resume collapse boundaries. Add sentToModel !== false guards at each site. * test(cli): add coverage for sentToModel !== false guards (#7381) * test(cli): add coverage for sentToModel !== false guard in input-history filter (#7381) * test(cli): add coverage for sentToModel !== false guard in YOLO turn-count telemetry (#7381) * fix(cli): restore corrupted docs and classify steer items as synthetic (#7381) * fix(docs): restore corrupted autogenerated input names in GitHub Action docs (#7381) * fix(cli): deduplicate findLastUserItemIndex and add steerInput forwarding test (#7381) * fix(cli): keep code-block copy numbering continuous across steer items (#7381) * test(core): add Hook continuation steerInput forwarding test Verify that steerInput is forwarded through the Stop-hook continuation path and settled early on the first content event of the continuation turn, matching the existing Steer continuation coverage. * fix(cli): sync selection test fixtures with ink FrameCell/ReadonlyFrame types (#7381) * fix(core): align cron day wildcard semantics (#7464) Co-authored-by: destire-mio <248462155+destire-mio@users.noreply.github.com> * feat(core): keep completed background agents resident (#7426) * feat(core): keep background agents resident * fix(core): harden background continuation boundaries * docs(core): move per-spawn cleanup comment to subagentDispose The comment describing the per-spawn cleanup (which stays undefined on the fork-resume path) had drifted above the launchModel declaration, where it no longer applied and could mislead readers. Relocate it to the subagentDispose assignment in the non-fork branch it actually documents. * fix(core): close finishing window and release resident on error in background GOAL path - Non-worktree GOAL completion drained the message queue but never called registry.beginFinishing(), unlike the worktree path. A send_message racing the terminal transition could be accepted (status still running, finishingAgents empty) and then orphaned by complete(). Call beginFinishing() after the empty drain to reject the racing message instead. - The completion catch block never reset keepResident, so a throw from patchAgentMeta/registry.complete left the runtime resident but finalized as failed — a zombie that cleanupRuntime never reclaimed. Reset keepResident in the catch so the finally block disposes it. --------- Co-authored-by: Claude <noreply@anthropic.com> * ci(autofix): continue environment-specific fixes (#7444) * ci(autofix): continue environment-specific fixes * docs(autofix): align verification wording * docs(autofix): require bundle before integration tests * docs(autofix): scope surrogate verification rules * docs(autofix): require focused tests before integration checks * docs(autofix): clarify review verification guidance * fix(acp-bridge): close prompt-terminal follow-ups from the PR #7400 self-review (#7453) * fix(acp-bridge): close prompt-terminal follow-ups from PR #7400 self-review Keep a removed RUNNING prompt visible to the teardown flush via a removed flag so its terminal still publishes when the session closes before the agent cooperates; gate broadcastTurnError's session turn-state mutation to running prompts; propagate the typed PromptDeadlineExceededError from the pre-dispatch abort check; document the deadline FIFO-release overlap trade-off, the trailing prompt_cancelled after flush, and the result.then/finally ordering invariant; route the dedup log to the debug channel; drop the prompt-deadline re-export that pulled the bridge into a leaf module. Fixes #7451 * test(acp-bridge): cover promote-then-remove-then-settle duplicate completed guard (#7453) --------- Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com> * fix(core): strip Qwen-internal daemon secrets from agent-spawned child env (#7256) * fix(core): strip Qwen-internal daemon secrets from agent-spawned child env Shell subprocesses (and the monitor tool and stdio MCP servers) inherited the full daemon process.env, including QWEN_SERVER_TOKEN (the serve-daemon bearer credential), so an agent-run command like printenv QWEN_SERVER_TOKEN could read an internal secret. Add a shared sanitizeChildEnv() that removes Qwen-internal daemon/server tokens (QWEN_SERVER_TOKEN, QWEN_DAEMON_TOKEN) before spawning, and apply it at the shell child_process + PTY paths, monitor.ts, and the mcp-client stdio transport. The denylist is deliberately narrow: it does NOT strip third-party credentials (GH_TOKEN, AWS_*, NPM_TOKEN, ...) that real shell workflows legitimately inherit -- only Qwen-internal secrets. Exported from the package root so the desktop denylists can consolidate onto it later. Fixes #6601. * test(core): cover daemon-secret stripping on monitor and mcp-client spawn sites * test(core): replace process.env instead of mutating in shell sanitization tests The file restores process.env by reference in afterEach, so in-place key mutations leaked into later tests. Use the replacement pattern already used by setupConflictingPathEnv. * docs(core): align JSDoc @param names with actual function signatures (#7492) Fix 6 instances where JSDoc @param tags had drifted from their corresponding function signatures — parameters were renamed, removed, or undocumented over time but the doc blocks were not updated. Closes #7446 * feat(serve): support forced MCP reconnects (#7488) * feat(serve): support forced MCP reconnects * test(serve): cover forced MCP reconnect options --------- Co-authored-by: 克竟 <dingbingzhi.dbz@alibaba-inc.com> * fix(cli): insert newline on Shift+Enter and stop streaming thinking-block flicker (#7397) * fix(cli): re-push Kitty keyboard flags onto the alternate screen in VP mode In VP mode the app renders on the alternate screen (`alternateScreen: true`), but the Kitty keyboard progressive-enhancement flags were pushed only once at startup on the main screen. The Kitty spec tracks these flags per screen buffer, so the alternate screen's stack stays empty and the terminal never reports modifiers: Shift+Enter arrives as a bare Enter (submit) or, when the terminal emits an ESC-prefixed variant, as an orphaned Escape that trips the empty-buffer double-Esc rewind prompt — so Shift+Enter can never insert a newline in VP mode even on Kitty-capable terminals (e.g. cmux). Re-push the flags onto the alternate screen right after Ink enters it (Ink writes the enter-alt-screen sequence synchronously inside render(), so the push is correctly ordered). Ink discards the alternate screen and its flag stack on unmount, leaving the startup main-screen push balanced by the existing disableKittyProtocol() on cleanup. Generated with AI Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix(cli): stabilize streaming thinking block height to stop flicker The pending "Thinking…" block renders the tail of the reasoning stream in a content-sized box. As the model emits paragraph separators, a blank line enters and leaves the tail window (and `trimEnd` drops trailing blanks), so the visible line count oscillates and the block flickers 2→3→5 rows during streaming. Track the tallest height the block has reached for the current thought and never render fewer rows than that (capped at the streaming window size), padding at the top so the newest line stays pinned to the bottom. The tracker resets when streaming ends or when the buffer shrinks (a new thought replaced it), so height is monotonic within a thought without leaking across thoughts. Generated with AI Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix(cli): decode xterm modifyOtherKeys Shift/Ctrl/Alt+Enter so it inserts a newline Terminals such as Ghostty report Shift+Enter as the xterm modifyOtherKeys sequence `ESC [ 27 ; <mods> ; <key> ~` (e.g. `ESC [ 27 ; 2 ; 13 ~`) when the Kitty keyboard protocol is not negotiated — which is the default, since Kitty detection does not always succeed. Two bugs kept this from inserting a newline: 1. The CSI-u parser read the leading `27` marker as the key code (matching the Escape key code 27) instead of the real key code in the third parameter, so with Kitty enabled Shift+Enter was mistaken for Escape and tripped the double-Esc rewind prompt. 2. The reassembly path that stitches readline's shredded CSI fragments back together was gated behind `kittyProtocolEnabled`, so with Kitty disabled the `ESC [ 27 ; 2 ;` head plus the stray `13~` tail leaked into the composer as literal text and no newline was inserted. Decode the third parameter as the real key code for the `27;…~` form, and route those sequences through the reassembly buffer even when Kitty is off (only the `ESC [ 27` marker opts in, so keys readline already parses cleanly are untouched). Shift/Ctrl/Alt+Enter now insert a newline in both VP and non-VP mode regardless of Kitty negotiation. Generated with AI Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix(cli): anchor VP viewport to the top until a conversation turn exists On a fresh VP-mode session the virtualized list holds the banner plus startup notices (tips / MOTD / info), so it is longer than one item. Keying the initial scroll anchor off list length alone selected scroll-to-end, which pinned the banner to the bottom of the full-height viewport and left the top half of the screen blank. Anchor to the top until there is an actual conversation turn (a user/user_shell history item or a pending response), then resume scroll-to-end so the latest output stays in view. Startup notices no longer count as content that forces bottom alignment. Generated with AI Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix(cli): stabilize streaming thinking window against availableTerminalHeight drift The grow-only streaming thinking window still flickered because its line cap was derived from availableTerminalHeight. While a thought streams the terminal keeps constrainHeight on, so availableTerminalHeight (and the derived maxLines) drifts up and down as sibling pending content grows, and the grow-only clamp `min(maxLines, …)` shrank the block whenever it dipped. Use a constant window height (MAX_STREAMING_THINKING_VISUAL_LINES) for the pending window instead. The window is only a few lines, so a fixed cap cannot meaningfully overflow (VP scrolls anyway), and the height stays stable while still growing monotonically within a thought. Generated with AI Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * Revert "fix(cli): anchor VP viewport to the top until a conversation turn exists" This reverts commit fbe86a9e159b75ea1f5b689cc327599c9dc91090. * fix(cli): guard modifyOtherKeys detection against keypresses without a sequence The modifyOtherKeys prefix check ran on every keypress, but some synthetic keypresses (and the useKeypress test harness) emit a key with no `sequence`, so `key.sequence.startsWith(...)` threw an unhandled rejection. Use optional chaining so a missing sequence is simply not a modifyOtherKeys start. Generated with AI Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * test(cli): mock pushKittyProtocolFlags in gemini.test.tsx kitty mock The kittyProtocolDetector mock omitted the newly added pushKittyProtocolFlags export. Add it so the mock stays in sync with the real module and a VP-mode startup path exercised through this suite cannot hit an undefined call. Generated with AI Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> --------- Co-authored-by: 秦奇 <gary.gq@alibaba-inc.com> Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix(web-shell): open singleton subagent details (#7495) Co-authored-by: ytahdn <ytahdn@gmail.com> * fix(web-shell): avoid redundant git status requests (#7496) Co-authored-by: ytahdn <ytahdn@gmail.com> * fix(agent): ignore empty working_dir placeholders (#7343) * fix(agent): ignore empty working_dir placeholders * test(agent): align empty working_dir expectations * feat(prompts): allow overriding core identity via QWEN_SYSTEM_IDENTITY_MD (#7478) * feat(prompts): update prompts.ts for QWEN_SYSTEM_IDENTITY_MD * feat(prompts): update prompts.test.ts for QWEN_SYSTEM_IDENTITY_MD * fix(prompts): address CR on QWEN_SYSTEM_IDENTITY_MD Keep getDefaultCoreIdentitySentence private, fail loud on path resolution errors, use trimEnd, and resolve identity only on the default-prompt branch. * test(prompts): align identity override tests with CR feedback Sample default identity from live prompt, cover trimEnd trailing whitespace, and assert homedir resolution failures throw. --------- Co-authored-by: 易良 <1204183885@qq.com> * fix(cli): yield to single-slot background agents (#7258) Co-authored-by: hogeheer <267467744+hogeheer499-commits@users.noreply.github.com> * docs(autofix): require evidenced pre-commit verification, not a bare "verified" (#7486) * docs(autofix): require evidenced pre-commit verification, not a bare "verified" The skill already said to run build/typecheck/lint/Vitest before committing, but softly — and #7408 committed a fix with a TS error the gate then rejected while its summary claimed "verified all 3 commits". A self-assessment the gate contradicts wastes a whole round. Strengthens the address-review contract from "run the checks" to: - actually run them, do not assert them from reading the diff; - if typecheck or a touched-package test fails, do NOT commit — treat the feedback as unresolved (failure.md); - end address-summary.md with a `## Verification` section listing each command run and its result; a bare "verified" is not acceptable. The framing is structural, not etiquette: the deterministic gate re-runs the same commands and discards the round on any failure, so skipping them only moves the rejection later. Pinned by a test so it cannot soften back. This is the checkable half of "audit before committing" — the undirected/reverse-audit-until-clean practice does not transfer to an unsupervised agent (no verifiable stopping condition, and it worsens the timeouts seen on large PRs), but "run the gate's own checks first and show the evidence" does. * fix(autofix): clarify Verification section precedes collapsed Chinese translation (#7486) --------- Co-authored-by: wenshao <wenshao@example.com> Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com> Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com> * feat(autofix): stop a PR that fails to push for N rounds in a row (#7482) * feat(autofix): stop a PR that fails to push for N rounds in a row Under takeover the round cap is 100, which is right for a PR that needs many PRODUCTIVE rounds. It is wrong for one that fails every round: #6723 ran 7 consecutive failed rounds (3 agent timeouts at 50 min, 4 gate rejections whose fix broke tests) over 8 hours, heading for round 100, because it is a 5700-line, 47-file, 5-day-old PR racing a fast-moving main — every round re-resolves a conflict it cannot finish or that fails the gate. Retrying at the same per-round budget will not converge; a human has to rebase or split it. Adds CONSECUTIVE_FAILURE_CAP (5), distinct from the total round cap. The handoff step already runs only when a round did NOT push, so it counts the unbroken run of prior failure markers — stopping at the first push ("Addressed the latest review feedback") or legitimate no-op ("no changes needed"), either of which proves progress and resets the streak. At the cap it forces the terminal round even under takeover, with a handoff that names the real fix (rebase/split, then /retry). Cause- agnostic: a timeout and a gate rejection both count. * fix(autofix): address review feedback on consecutive-failure circuit breaker (#7482) - Fix misleading comment: the walk is oldest-first (API order) with reset-on-success, not newest-first with early stop - Prefer the already-fetched ic.json over a redundant gh api call, falling back to the API only when the file is missing - Filter eval markers by re-arm window (win=) so pre-re-arm failures do not immediately re-terminate a re-armed PR - Add test coverage for the MARK_ROUND == MAX_ROUNDS guard and for window-scoped streak counting * fix(autofix): exempt transient model errors from consecutive-failure breaker (#7482) --------- Co-authored-by: wenshao <wenshao@example.com> Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> * feat(core): restore background agent roster (#7459) * feat(core): restore background agent roster * fix(web-shell): add list_agents to TOOL_DISPLAY_NAMES The new list_agents core wire tool was added to core's ToolNames but not to the web-shell TOOL_DISPLAY_NAMES map, causing toolFormatting.drift.test.ts to fail (expected ['list_agents'] to deeply equal []). Add the missing 'ListAgents' display-name entry so the browser panel shows a friendly name instead of the raw wire name and the drift guard passes. * fix(cli): reload old-session background agents on failed resume rollback When /resume fails after core has swapped but before the UI swap, the catch block rolls core back to the old session via startNewSession(oldSessionId). However the forward path already called resetBackgroundStateForSessionSwitch, which cleared the old session's in-memory background agents. The rollback did not reload them, so list_agents returned empty for the old session (whose sidecars are still on disk) until the next process start or successful resume. Reload the old session's paused background agents after rolling core back, so the restored roster matches on-disk state. Placed after startNewSession so the loadPausedBackgroundAgents current-session guard is satisfied; best-effort via .catch so it never blocks the rollback path. * fix(web-shell): add zh translation for list_agents tool name The toolFormatting test 'has a zh translation for every tool in the display-name map' failed with expected ['list_agents'] to deeply equal [] because list_agents was added to TOOL_DISPLAY_NAMES without a matching toolName.list_agents zh-CN entry. Add the translation to restore parity. * fix(cli): resolve CI failures for background-agent roster restore - Add toolDisplayName.ListAgents translations (en, zh, zh-TW, ca) so the new list_agents tool has a zh entry; fixes i18n/index.test.ts. - Add loadPausedBackgroundAgents and consumePendingRecoveredAgentsNotice to the acpAgent worktree test config mock, which loadSession now calls via #restoreBackgroundAgentsOnResume; fixes acpAgent.worktree.test.ts. * refactor(core): extract incompatible-isolation blocked reason to a const Move the incompatible-isolation blocked-reason string out of an inline literal into a module-level INCOMPATIBLE_ISOLATION_BLOCKED_REASON const, matching its four sibling reasons so the text is discoverable by constant-name grep and edited alongside the others. * fix(core): preserve retained activity state on failed agent revive Address review feedback on the background-agent roster restore: - On a failed completed-agent revive, restore UI state with a non-empty guard instead of `??`. Because `restorePausedEntry` resets the paused entry's `recentActivities` to `[]`, the previous `failedEntry?.field ?? completedEntry.field` kept that empty array and dropped the pre-revive snapshot (the UI Progress section rendered empty). Applied consistently to pendingMessages, recentActivities, and pendingApprovals. Add regression coverage for previously untested paths: - failed revive preserves pre-revive recentActivities - terminal-agent cap admits only the newest MAX_RETAINED_TERMINAL_AGENTS completed sidecars on restore - /resume rollback reloads the old session's background agents - headless resume prepends the recovered-agents notice to the prompt * test(cli): cover interrupted-turn continuation not consuming recovered-agents notice Add ACP and headless regression tests asserting an interrupted-turn continuation does not consume the one-shot recovered-agents notice (the !isContinue / !continueInterrupted guards), so it is delivered on the user's next ordinary prompt. Mirrors the existing slash-command coverage. --------- Co-authored-by: Claude <noreply@anthropic.com> * feat(cli): support custom skill directories via settings (#7395) * feat(cli): support custom skill directories via settings (#7394) Add skills.directories setting that accepts an array of additional directory paths to scan for skills (SKILL.md files). Paths support ~ expansion. Directories are scanned recursively at user level, after the default ~/.qwen/skills/ directory. Example settings.json: { "skills": { "directories": ["~/.agent/skills", "~/.claude/skills"] } } Changes: - settingsSchema.ts: add skills.directories array setting - core Config: add customSkillDirs param and getCustomSkillDirs() - SkillManager: append custom dirs to user-level skill base dirs - CLI config: read skills.directories and pass to core Config * fix(cli): regenerate settings schema for skills.directories (#7394) * fix(core): address review feedback for custom skill directories (#7395) - Use optional chaining for getCustomSkillDirs() to prevent TypeError on partial Config mocks (workspace-skill-management, workspace-skills-status) - Reuse expandHomeDir utility instead of inline tilde expansion - Fix inaccurate 'scanned recursively' wording to 'one level deep' - Correct JSDoc: paths are raw, expansion happens in SkillManager - Trim whitespace from custom dir entries in CLI layer - Add tests for custom dir expansion, dedup, and partial config safety * fix(core): address review feedback for custom skill directories (#7395) * fix(core): address review feedback for custom skill directories (#7395) * test(core): add relative path resolution test for custom skill dirs (#7395) * fix(cli): add Array.isArray guard for skills.directories and safe mode test (#7395) * fix(skills): address review feedback on custom skill directories (#7395) - Add bare mode test for skills.directories guard - Include resolved absolute path in relative directory warning - Clarify that dedup applies to default user dirs, not bundled skills - Regenerate settings schema --------- Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com> Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com> Co-authored-by: Qwen Code Autofix <qwen-code-autofix@users.noreply.github.com> * fix(core): add image modality support for qwen3.8-max and kimi-k3 models (#7491) * fix(core): add image modality support for qwen3.8-max models qwen3.8-max-preview supports image input but was falling through to the catch-all text-only rule because no pattern matched it. This caused the vision bridge to unnecessarily transcribe images via a secondary model instead of sending them directly to the primary model. * fix(core): also add image modality for kimi-k3 Kimi K3 officially supports image + video input but was falling through to the catch-all text-only rule, same issue as qwen3.8-max. * fix(dingtalk): preserve non-bot mention context (#7473) * fix(dingtalk): preserve non-bot mention context * test(dingtalk): cover plural mentions, staffId fallback, and edge cases (#7473) --------- Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> * fix(core): harden the usage salvage around session deletion (#7425) Post-merge review follow-ups on #7391 (three findings): - Salvage the archived transcript in the active-branch deletion too: when both copies co-exist (an interrupted archive) and the fresh active transcript carries no telemetry, the archived copy holds the session's usage history and was deleted unsalvaged. The dedup guard makes the extra call a no-op whenever the active copy already wrote. - Enforce the "never blocks deletion" contract at the call site: a salvageUsageBestEffort wrapper catches and warns, so the guarantee is structural rather than an implementation detail of persistUsageBeforeTranscriptDeletion. The new failure-tolerance test (salvage rejects -> deletion still succeeds) fails without the wrapper — the bare await let the rejection escape through removeSessionFiles' rethrowing catch. - Clear the salvage module mock in beforeEach so the wiring test's invocationCallOrder assertions can never read stale calls. Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * fix(core): make fork subagents discoverable (#7460) * test(core): cover Shell truncation without an artifact (#7470) Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix(ci): autofix route checks existing labels on non-trigger label events (#7481) * fix(ci): autofix route checks existing labels on non-trigger label events When triage adds multiple labels in sequence, per-issue concurrency cancels earlier runs. If the last label is not a trigger label (e.g. scope/build-system), the surviving run skips the issue phase even though the issue already has autofix/approved + status/ready-for-agent. Before ignoring a non-trigger label event, check ISSUE_LABELS_JSON for both required labels. If present and the issue is open, proceed with the issue phase. Trust was already established when the trigger labels were applied (both require triage+ permission). * fix(ci): require trusted sender for label fallback * feat(cli): preserve semantic text when copying VP selections (#7286) * docs(cli): define semantic copy fidelity scope * docs(cli): address semantic frame review gaps * docs(cli): preserve soft-wrap source separators * feat(cli): preserve semantic selection copy * fix(cli): address semantic copy review findings * fix(cli): preserve clipped semantic boundaries * fix(cli): limit separator carrier joiner to visible width in wrap metadata The greedy /\s+/ match in wrapTextWithMetadata could capture more source whitespace than the separator carrier row actually consumed (e.g. a tab following a space), causing duplicated whitespace in semantic copy. Limit the match to visibleLine.length characters and add a mixed space/tab regression test. --------- Co-authored-by: 秦奇 <gary.gq@alibaba-inc.com> * test(core): stub the registry methods agent.ts actually calls (#7538) The shared stubRegistry in agent.test.ts was missing six methods that agent.ts reaches: bridgeApprovalEvents, getQueuedCount, registerResidentAgent, restartCompletedAgent, unregisterResidentAgent and waitForMessages. That is not a benign omission. The background body wraps its work in a try/catch that routes any throw into registry.fail(), so a missing method never surfaces as 'not a function' — it silently converts a successful run into a failed one. On the GOAL completion path unregisterResidentAgent is called immediately before complete(), so the TypeError replaced the completion entirely: registry.fail('fork-...', 'registry2.unregisterResidentAgent is not a function', ...) That is what broke 'runs a non-interactive fork through the background registry' on main. #7460 added the registry.complete assertion, which exposed the incomplete stub — before it, nothing checked whether the background body finished successfully and the TypeError was swallowed. Stub all six with their real return shapes (unregisterResidentAgent returns boolean, bridgeApprovalEvents returns the unsubscribe callback agent.ts later invokes, waitForMessages resolves to a list) and assert registry.fail was not called before asserting completion, so a future gap reports the actual error instead of 'complete: 0 calls'. * perf(startup): lazy-load Google GenAI SDK on first use (#7512) * perf(startup): lazy-load Google GenAI SDK on first use Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * codex: address PR review feedback (#7512) Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * codex: address PR review feedback (#7512) Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> --------- Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix(vscode): use file picker image paths for vision input (#7493) * fix(vscode): use image paths from file picker * fix(vscode): keep image picker paths raw * fix(vscode): resolve image picker paths on submit * fix(vscode): send picked images as vision context * fix(vscode): encode prompt image file URIs * fix(vscode): address image path review comments * test(vscode): cover image file reference edge cases * fix(cli): open the actual serve fallback port (#7501) * fix(cli): open actual serve fallback port * test(cli): match serve URL to fallback listener * docs(cli): clarify serve listen error handling --------- Co-authored-by: Shaojin Wen <shaojin.wensj@alibaba-inc.com> * fix(ci): don't let one failing scenario sink the whole visual preview (#7511) The web-shell visuals render runs every screenshot and flow in a single `test:e2e:visuals`, and that step had no `continue-on-error`, while the compose and upload steps had no `if: always()`. So one failing or timing-out scenario failed the job, the artifact was never uploaded, and the publish workflow had nothing to post — the entire preview vanished even when every other scenario passed and its PNG was already on disk. A flow (a long multi-click sequence) is the most fragile scenario kind, so the fragile one silently takes down the deterministic screenshots. PR #7498 hit exactly this: 29 scenarios passed, one new channel-management flow timed out, and the PR got no preview and no comment at all. Make the after-capture step `continue-on-error` so the passing captures survive and the later steps still compose and upload them. The publish job only runs on a `success` conclusion, so the job must stay green — but a masked failure must not read as a clean preview. Ship the step's real `.outcome` (which continue-on-error does NOT mask, unlike `.conclusion`) to the publisher as `render-status.txt`, and have the comment builder use it: an empty preview whose render failed says "one or more scenarios failed to render" and is explicitly NOT the reassuring green check or the coverage-gap prompt (both imply the render ran); a partial preview is labelled partial above the shots that did render. A missing status file (older run) defaults to complete, so this only ever adds a warning, never suppresses a real preview. The failing scenario still needs fixing — it's now surfaced in the comment rather than by silently deleting everyone else's preview. Co-authored-by: wenshao <wenshao@example.com> * feat(web-shell): add selective shadow DOM isolation (#7551) Co-authored-by: 钉萁 <dingqi.jww@alibaba-inc.com> * feat(web-shell): add renderChatHeader slot for custom session header (#7553) * fix(cli): say review coverage gaps in the author's units, not chunk ids (#7550) The posted review body rendered coverage disclosures with the run's own bookkeeping as subjects: bare chunk ids, unsorted, one per subject. On a run that certified nothing (PR #7268) the body enumerated all 49 chunk ids across two sentences while opening with "Reviewed. Suggestions are inline." — the opener certified the exact thing every following sentence took back, and nothing on the PR page maps a chunk id to code. Three changes, all render-time — the structural entries, the caps, the caller-echo dedup and the stderr remediation still key on chunk ids, which is where the id is the selector a reader can act on: - Coverage now returns the plan's chunk→files table (DiffChunk.files was already in the plan JSON; the coverage type slice dropped it). - compose-review renders chunk gaps through describeChunkGap: every planned chunk collapses to "the entire diff", a narrow gap with known files names the files, and anything wider is counted against the plan's total. Applied to the receipt sentence, the uncoverable sentence (bare CLI entries only — caller-authored entries render verbatim) and the grouped per-cause sentences. - The COMMENT opener may no longer say "Reviewed." over a disclosure set that denies it: when no chunk is both covered and undisclosed — or no chunk universe could be read at all — it opens with a zero-certified warning instead. A rewritten launch demonstrably read its chunk, so coverage alone is not the test; certified is covered with no disclosure against it. Co-authored-by: verify <verify@local> * fix(autofix): retry a skipped-Prepare instead of stranding the PR terminal (#7490) * fix(autofix): retry a skipped-Prepare instead of stranding the PR terminal A base/infra failure BEFORE the agent runs was misread as an agent crash and terminated the PR forever. When an early step fails — installing or building the trusted base, checkout, node setup — the `Prepare branch and feedback` step is skipped, so NEWEST is empty, and the report step's "crashed before reading feedback" branch fired: MARK_ROUND=MAX_ROUNDS, terminal, scan skips it on every future tick. Observed: a web-shell TypeScript break on `main` failed `Install dependencies and build` (which builds the trusted base) across a whole scan batch, and SIX healthy PRs were stranded terminal at round=100 in one run — including ones at round 9 and 11 that had nothing to do with the break. `round=100` there is a terminal sentinel, not 100 attempts. NEWEST-empty now splits on steps.prepare.outcome: - 'skipped' (an earlier step failed, the agent never ran) is infra/base and transient: retry with a sentinel ts so the feedback stays live, incrementing the round so a PERSISTENTLY broken base is still bounded and stops at the cap (recoverable with /retry). - 'success'/'failure' (Prepare ran, no feedback produced) is a genuine pre-read agent crash: unchanged terminal behaviour. This is the reverse of the asymmetry #7482 addresses: that bounds a crash AFTER reading that retried forever; this stops a transient failure BEFORE reading from going terminal after one. * docs(autofix): note a pre-Prepare cancel also retries intentionally (#7490) * fix(autofix): also retry a cancelled/empty prepare outcome, not just skipped A previous review comment on this PR noted that a job cancelled before Prepare should retry too. It was right about the intent but the code did not do it: `steps.prepare.outcome` is 'cancelled' for a cancel and '' for a job that stopped before Prepare entered the step context — both DISTINCT from 'skipped', so `== 'skipped'` sent them to the terminal branch, the same over-termination this PR exists to fix. Match on "not a real Prepare run" (`!= 'success' && != 'failure'`) instead, so skipped, cancelled, and empty all retry; only a Prepare that actually ran to a verdict (success/failure) with no feedback stays terminal — the genuine pre-read agent crash. Test extended to drive the cancelled and empty cases (retry) and both real-run outcomes (terminal); mutation-verified that reverting to `== 'skipped'` reddens the cancelled case. * test(autofix): update the pre-read-crash case for the broadened retry The prior commit broadened NEWEST-empty retry to skipped/cancelled/empty but left the older 'replays the handoff decision' test asserting the old terminal behaviour for an unset PREPARE_OUTCOME (which now retries). That test's terminal cases now set PREPARE_OUTCOME=success/failure explicitly — the only outcomes that still terminate — so it exercises the genuine pre-read agent crash rather than the infra/cancel path. * test(autofix): anchor the skipped-Prepare extraction past the CONSEC block CI reddened `retries a skipped-Prepare` after main's consecutive-failure cap (#7482) merged into this branch: that block was inserted between this decision block and the report `{`, and it calls `gh api`. The test's `{`-anchored regex over-captured through it, so the extracted script ran the unstubbed `gh api` and failed. Anchor the end on the same `# Consecutive-failure` comment the sibling gate-crash test already uses, so the extraction stops at this decision block's own closing `fi`. * fix(autofix): exempt skipped-Prepare from the consecutive-failure breaker A broken base build skips Prepare, producing no API error file — so the consecutive-failure breaker ran on the new retry path and, after 5 scans, re-introduced the exact mass-stranding this PR exists to prevent. Exempt pre-agent infra failures (skipped/cancelled/empty outcome) from the breaker, mirroring the transient 429/5xx exemption: same failure class (not the PR's fault, self-heals, hits the whole batch). The round cap + sentinel-ts /retry recovery already bounds a persistently broken base. Also trim "checkout" from the retry headlines (checkout failures do not land in this branch) and hoist the duplicated MARK_TS assignment. * fix(autofix): reset the consecutive-failure streak on prior infra-failure markers The streak walker counted prior infra-failure headlines ("AutoFix could not start —…") as failures, inflating the consecutive-failure count on subsequent rounds. A PR with 3 real agent failures, then 3 rounds of base-build infra failures, then 1 more real failure would trip the cap-5 breaker even though only 4 rounds were the PR's fault. Add the two infra-failure headline patterns as reset strings in the streak walker, alongside the existing push and no-op resets. The genuine agent-crash headline ("AutoFix could not start evaluation —…") is deliberately excluded — it is a real failure and must still count. * fix(autofix): clarify infra-failure headlines and else-branch comment (#7490) Address review nits: the retry headline now mentions cancelled runs, the cap headline says 'reached the round cap' instead of overstating 'could not start for N rounds', the else-branch comment says 'prepare itself crashed' instead of 'agent crash', and the streak-reset pattern is simplified now that both infra headlines share the same prefix. --------- Co-authored-by: wenshao <wenshao@example.com> Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com> * fix(cli): keep role codenames and brief paths out of the posted review body (#7560) The posted body still carried two operator registers #7550 left in place: roster role subjects rendered their internal codenames ("Agent 1c: Cross-file tracer", "Test coverage matrix (whole-diff)"), and an unread brief's disclosure interpolated its filesystem path. And when verify and the reverse audit failed the same way, the body said it twice, in two near-identical sentences. - Every Brief now carries a publicLabel — the dimension said as what it checks ("the cross-file consistency pass") — and coverage's structural disclosures carry it as publicSubject beside the internal subject, plus a path-free publicReason for unread briefs. The internal label and the path stay on stderr, where they are the selector an operator acts on; every dedup and certification check still keys on the internal subject. - compose-review renders the public fields and groups by the reason the body PRINTS, so two unread briefs share one path-free sentence instead of repeating it per role. - verificationGaps merges verify and reverse-audit failures of the same delivery shape into one sentence with both subjects and both consequences; mixed shapes keep their precise per-role texts, and the per-role rebuild commands stay on stderr either way. Co-authored-by: verify <verify@local> * fix(autofix): retry an agent timeout instead of advancing past its feedback (#7563) A timeout evaluated NOTHING — the agent ran out of budget before finishing, so nothing was committed and the feedback is unaddressed. It was treated as an evaluated verdict (real ts, watermark advances), which strands that feedback: the next scan sees "nothing new" and never retries. Observed on #7471 (round 13/100), a heavily-reviewed 1871-line PR: rounds 11 and 13 timed out, but round 12 pushed — so a timeout is transient far more often than not, and advancing past it left the round-13 feedback unhandled. run-agent.mjs now drops an `agent-timeout` signal on result.timedOut, and the handoff routes it like a pre-verdict crash: sentinel ts (feedback stays live) and a retry, with a headline that names the real fix at the cap (split the PR or raise the budget). A PR that PERSISTENTLY times out is bounded by the round cap and the consecutive-failure cap, so this cannot loop forever — it just stops treating a one-off budget blip as a verdict. The loop guard stays terminal (a tool-call loop is a real defect, not a budget blip). An API error still routes to its own model-key handoff; the timeout signal is written only when NOT an API error. Co-authored-by: wenshao <wenshao@example.com> * feat(serve): add workspace-level generation (#7552) * feat(serve): add workspace-level generation * docs(serve): document workspace generation capability * fix(serve): align workspace generation contracts --------- Co-authored-by: ytahdn <ytahdn@gmail.com> * ci: matrix ECS runner update + sudo install + repository_dispatch trigger (#7513) * ci: matrix ECS runner update with sudo install - Use matrix strategy (ecs-update-sg, ecs-update-64c) to update both physical ECS hosts in parallel (fail-fast: false). - Always use sudo npm install -g so the package lands in /usr/local (system-wide PATH) instead of the runner user's home directory. - Move concurrency to job level (matrix context not available at workflow level per actionlint). - Add repository_dispatch trigger for release-driven updates. - Register new runner labels in actionlint.yaml. * fix(ci): use dispatch version for runner update --------- Co-authored-by: Shaojin Wen <shaojin.wensj@alibaba-inc.com> * fix(web-shell): include managed id in artifact open requests (#7570) Co-authored-by: 钉萁 <dingqi.jww@alibaba-inc.com> * feat(serve): persist workspace channel configuration (#7514) * feat(serve): persist workspace channel configuration * fix(serve): harden channel settings snapshots * fix(serve): validate startup channel names * fix(serve): reserve all channel name --------- Co-authored-by: Shaojin Wen <shaojin.wensj@alibaba-inc.com> * fix(sdk-python): require canonical form in validate_session_id (#7532) uuid.UUID() accepts several non-canonical spellings — braced {...}, urn:uuid:..., and dash-less hex — so validate_session_id let them through after the RFC 4122 variant check. The value is then forwarded to the CLI verbatim as --session-id/--resume, producing a malformed session id downstream rather than a clear error at the SDK boundary. Reject anything whose canonical form differs from the input. Case is deliberately not part of the comparison: UUID() lowercases, and an all-uppercase spelling is still valid canonical input. Co-authored-by: Shaojin Wen <shaojin.wensj@alibaba-inc.com> * fix(web-shell): sync background agent status (#7561) * fix(web-shell): sync background agent status * fix(web-shell): harden background agent reconciliation --------- Co-authored-by: ytahdn <ytahdn@gmail.com> * feat(core): propagate trusted daemon invocation context (#7279) * feat(core): propagate trusted daemon invocation context * test(cli): update ACP startup expectation * refactor(core): centralize ACP capability env key * test(cli): update worktree ACP core mock * test(integration): run daemon context smoke on PRs * test(ci): update no-AK smoke expectation * test(core): cover invocation context isolation * fix(cli): compare ACP capability safely * fix(docs): restore GitHub action input names * fix(core): sanitize private ACP capability from child env * fix(core): reuse private ACP capability env constant * test(cli): cover malformed trusted invocation context * test(acp-bridge): assert exact child environment --------- Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> Co-authored-by: 易良 <1204183885@qq.com> * fix(feishu): await stream cancels in media download teardown (#7465) * fix(feishu): await stream cancels in media download teardown downloadMedia left two reject paths' stream teardown unawaited: - the oversize-stream path called reader.cancel() without awaiting, so a cancel error during teardown became an unhandled rejection (fatal under Node's default --unhandled-rejections=throw); - the Content-Length reject path returned without cancelling resp.body, leaving the connection pinned until GC. Both were already fixed for the sibling DingTalk downloader in #7361 (which was itself modelled on this Feishu code), so this brings Feishu to parity. Adds a regression test that pins the reader.cancel() await via a rejecting cancel, plus an assertion that the Content-Length path releases the body. * test(feishu): cover a rejecting body.cancel() on the Content-Length path Mirrors the existing reader.cancel() teardown test for the other reject path, per review feedback. Removing the await on resp.body?.cancel() flips execution onto the 'rejected: size ... exceeds' branch and the test fails. * fix(autofix): make the review-address report wrapper lines bilingual (#7569) The agent's address-summary.md / no-action.md already ends with a collapsed Chinese translation, but the workflow-appended wrapper lines around it — the "Addressed/Reviewed the latest feedback" lead-in, the "Base-conflict check" line, and the "Re-review when you have a moment" footer — were English-only and sat outside that block. So the posted comment was only half translated, unlike the takeover-ack comments (full collapsed Chinese block) and the "model/模型" sign-off in this same report (already inline-bilingual). Give each wrapper line an inline Chinese translation, matching the model/模型 idiom. The English halves are preserved verbatim — the streak-reset detector globs on "Addressed the latest review feedback" and "no changes needed", and a test extracts these lines — so behaviour is unchanged and old English-only comments still match. A new test pins each English-Chinese pair so a future reword that drops the Chinese fails. The terminal handoff/failure comment is left English-only for now (SKILL.md keeps it so by design); that is a separate change. Co-authored-by: wenshao <wenshao@example.com> * feat(cli): post the review body bilingually when the PR description is Chinese (#7564) When the PR author writes Chinese, the posted /review body was English-only. fetch-pr now records whether the PR description contains Han characters (prDescriptionHasHan, detected from the same gh pr view call and stamped into the plan report), and compose-review renders the body bilingually off that flag: the English body leads, the complete Chinese version rides collapsed in a <details><summary>中文说明</summary> block, and the model footer stays outside the fold. The signal is the CLI's own — the caller cannot toggle the register of a certified body — and a local plan has no field, so nothing changes for terminal-only reviews. Every deterministic body fragment carries an en/zh pair end to end: compose-review's clause templates and describeChunkGap phrases, the coverage disclosures (reasons, publicLabel role subjects via a new publicLabelZh, the path-free unread-brief reason) and the Step 4/5 gap texts including the combined same-shape sentence. Fragments with no deterministic translation — model-written findings, caller echoes, interpolated errors — ride verbatim in both halves. verificationGaps now returns structural {subject, reason, subjectZh, reasonZh} entries, which also removes compose-review's last recover-the-boundary-from-prose parse. SKILL.md instructs the same format for the model-authored inline comments: English finding first (marker and suggestion block stay in the English half — tooling filters on them), full Chinese translation collapsed beneath, footer last. Co-authored-by: verify <verify@local> * feat(autofix): auto-rerun a check that died on infrastructure, once (#7562) * feat(autofix): auto-rerun a check that died on infrastructure, once A failed check can be red because the machine died, not the code — a self-hosted runner losing the server, the disk filling. #7490's E2E failed with "runner lost communication with the server" and went green on a rerun. The scan now reruns such a check's failed jobs automatically. Detection is a conservative annotation whitelist (INFRA_FAILURE_SIGNATURES) — only unambiguous machine failures, never a test-level timeout, which could be a real regression. The one-shot guard is run_attempt, not a marker: a run already retried to attempt 2 and still infra-failing is persistent, so it is left for a human; after a rerun the attempt increments, so the next scan will not rerun it. Every step is fail-safe (any API error → no rerun), it runs only when the PR actually has a failed check, and the gate carries the same review-address carve-out as the other check selectors so the loop never reruns its own runs. This is the transient-infra sibling of #7554 (stale-base): that merges current main when a check is base-inherited; this reruns when a check died on the runner. Neither touches a check that is a genuine failure. Note: rerun-failed-jobs needs the PAT to hold `actions: write`. * fix(autofix): use POSIX ERE groups in infra-failure regex, cover all signatures in tests (#7562) * fix(autofix): also treat a git fetch/clone transport death as infra #6506's checkout died mid-transfer — "fetch-pack: invalid index-pack output" and "RPC failed; curl 92 ... CANCEL" — which then hung the job into the 20m limit. That is infra, not the PR (it only touches a doc), and a re-run made it green. But the infra-signature whitelist did not cover it, so the auto-rerun did not fire and it waited on a human. Add `invalid index-pack output` and `RPC failed` — the two canonical git-transport-death phrases — to INFRA_FAILURE_SIGNATURES. A co-present job-timeout line does not block the match (one matching line classifies the run), and a BARE timeout with no transport signature is still left alone, since it can be a real regression. Both new signatures are pinned in the test's per-signature loop, plus a case on #6506's real composite annotation and a bare-timeout-is-not-rerun guard. * fix(autofix): paginate annotations and filter Autofix runs in infra-rerun loop (#7562) --------- Co-authored-by: wenshao <wenshao@example.com> Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com> Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com> * fix(serve): detect stale SSE cursors across daemon restarts via epoch token; preserve turn attribution and surface compaction failures in replay (#7458) * fix(daemon): epoch-token restart detection, compaction attribution, and degraded-snapshot signaling (DAEMON-001/007/008) * fix(acp-bridge): field-level turn attribution merge and replayDegraded bridge test (#7458) * fix(serve): skip bus epoch lookup for virtual subagent SSE streams (#7458) The REST SSE route looked up the bus epoch for every session id, but virtual subagent sessions ride their own bus and their compound ids are not in the bridge's byId map, so the lookup threw and aborted the subscription — breaking subagent event streams. Skip the lookup for the virtual path and degrade a torn-down real session to a headerless stream (mirrors the /acp route). Also bumps the daemon browser SDK bundle budget (167KB -> 168KB) for the epoch fields and declares eventEpoch on DaemonSession so the create/attach path drops its inline type cast. * fix(serve): stamp eventEpoch on accepted continuations and surface replayDegraded in the SDK (#7458) Address three review suggestions: - POST /session/:id/continue now returns eventEpoch alongside lastEventId, mirroring the prompt 202 envelope so continuation-seeded SSE cursors detect daemon restarts (DAEMON-001) - DaemonSessionClient exposes replayDegraded from the load response so SDK consumers can prefer the full transcript over a degraded snapshot - add /acp dispatch-level regression test for the degraded-snapshot stderr breadcrumb (fires only when snapshot.degraded is set) * test(cli): fix load-reply race in the degraded-breadcrumb transport test Await each session/load reply frame before opening the session stream so the GET cannot race conn.ownSession() into a 403; addresses the review Critical on the deg-0 arm. * fix(serve): allow and expose X-Qwen-Event-Epoch in CORS headers Cross-origin SSE clients must send the epoch header through preflight and read it from the response, or stale-cursor detection (DAEMON-001) is silently disabled for every CORS client. --------- Co-authored-by: qwen-code-bot <qwen-code-bot@users.noreply.github.com> Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> Co-authored-by: Qwen Autofix <qwen-autofix[bot]@users.noreply.github.com> * feat(core): Align GenAI telemetry with ARMS (#7536) * feat(core): align GenAI telemetry with ARMS Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix(core): remove estimated token usage splits Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix(core): address GenAI telemetry review feedback Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> --------- Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> Co-authored-by: Shaojin Wen <shaojin.wensj@alibaba-inc.com> * fix(serve): avoid TOCTOU race dropping live sessions from list response (#7556) * Initial plan * fix(serve): avoid TOCTOU race dropping live sessions from list response --------- Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: 易良 <1204183885@qq.com> * fix(cli): prevent monitor turns after task_stop (#7573) --------- Co-authored-by: 秦奇 <gary.gq@alibaba-inc.com> Co-authored-by: Qwen Code Autofix <qwen-code-autofix@users.noreply.github.com> Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com> Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com> Co-authored-by: Qwen Code Autofix <qwen-code-autofix[bot]@users.noreply.github.com> Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> Co-authored-by: destire-mio <qppque@gmail.com> Co-authored-by: destire-mio <248462155+destire-mio@users.noreply.github.com> Co-authored-by: Dragon <52599892+DragonnZhang@users.noreply.github.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: 易良 <1204183885@qq.com> Co-authored-by: jinye <djy1989418@126.com> Co-authored-by: chinesepowered <nlai@rediffmail.com> Co-authored-by: ovochouovo <18212194+ovochouovo@users.noreply.github.com> Co-authored-by: Edenman <67549719+BZ-D@users.noreply.github.com> Co-authored-by: 克竟 <dingbingzhi.dbz@alibaba-inc.com> Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> Co-authored-by: ytahdn <1294726970@qq.com> Co-authored-by: ytahdn <ytahdn@gmail.com> Co-authored-by: Truraly <94105924+Truraly@users.noreply.github.com> Co-authored-by: zjgzx1988 <zjgzx1988@hotmail.com> Co-authored-by: hogeheer499-commits <hogeheer499@gmail.com> Co-authored-by: hogeheer <267467744+hogeheer499-commits@users.noreply.github.com> Co-authored-by: Shaojin Wen <shaojin.wensj@alibaba-inc.com> Co-authored-by: wenshao <wenshao@example.com> Co-authored-by: qwen-code-dev-bot <qwen-code-dev@service.alibaba.com> Co-authored-by: Nothing Chan <chenliu.cl@alibaba-inc.com> Co-authored-by: 钉萁 <dingqi.jww@alibaba-inc.com> Co-authored-by: yuanyuanAli <135116774+yuanyuanAli@users.noreply.github.com> Co-authored-by: verify <verify@local> Co-authored-by: qqqys <qys177@gmail.com> Co-authored-by: callmeYe <512217680@qq.com> Co-authored-by: Qwen Autofix <qwen-autofix[bot]@users.noreply.github.com> Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com>
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
4 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
Adds a consecutive-failure sub-cap, separate from the total round cap, so a PR that fails to push every round stops instead of retrying up to 100 times under takeover.
Why it's needed
The round cap answers "how many productive rounds may a PR take?" — and takeover raises it to 100 deliberately, for PRs that need many rounds of address-feedback-then-push. It does not answer "how many rounds may fail in a row?"
Observed on #6723 (a 5700-line, 47-file, 5-day-old PR that conflicts with a fast-moving
main):tests failed in packages/coretests failed in packages/clitests failedSeven straight failures over eight hours, nothing pushed, heading for round 100 — each round a ~50-minute agent run or a full gate cycle. Every round re-resolves a conflict the agent cannot finish inside the budget, or produces a fix that breaks tests. Retrying at the same per-round budget does not converge; a human has to rebase or split the PR.
The two failure kinds are handled elsewhere individually (a timeout, a gate rejection), but nothing notices that they are happening back to back with no progress.
How
CONSECUTIVE_FAILURE_CAP: '5', distinct fromTAKEOVER_MAX_ROUNDS: '100'.Reaching the handoff step at all means this round did not push — the push and no-op paths report from a different step. So the step counts the unbroken run of prior failure markers by walking the bot's own eval-comment headlines newest-first and stopping at the first that pushed (
Addressed the latest review feedback) or was a deliberate no-op (no changes needed) — either proves the loop was making progress, so the streak resets there. If the run (this round included) reaches the cap, it forces the terminal round even under takeover, with a handoff that names the actual fix: rebase / split / reduce the PR, then@qwen-code /retryto re-arm.It is cause-agnostic: a timeout and a gate rejection both count, because both are "failed, nothing pushed." A single push or no-op anywhere in the run resets it, so this only ever fires on a genuinely stuck PR.
Reviewer Test Plan
How to verify
npx vitest run scripts/tests/qwen-autofix-workflow.test.js— 92/92 (a run in three hits the pre-existingtakeover-command toggleload flake; see Risk). The new test replays the real extracted counter+override block under bash, with a stubbedghserving a synthetic comment history:/retryMutation-verified, four reverts each turning it red: neutering the override; making a no-op not reset; making a push not reset; setting the cap equal to the takeover cap (so it never binds).
One existing test needed its extraction anchor updated:
retries a verification-gate crashgrabbed the decision block by its trailingfi+ report{, which this block now sits between. It now anchors on the new block's comment, so it drives only the crash decision — unchanged behaviour, noghcall.Static, run locally with the exact CI toolchain:
js-yamlparses; all 37run:blocks passbash -n; actionlint 1.7.12 clean; prettier clean.Post-merge smoke: fix(prompt-cache): stabilize deferred tool calls #6723's next scan should post the structural handoff and go terminal instead of a 8th failed round.
Evidence (Before & After)
Tested on
Risk & Scope
/retryre-arms it after the human intervenes, opening a fresh window.gh api …/commentsfetch per failing round. On the fetch failing,PRIOR_HEADSis empty and the streak is 1 — so an API blip cannot force a spurious terminal (fail-open).takeover-command togglefails under full-suite subprocess load — A/B on this machine shows unmodifiedmainfailing the same test at a comparable rate, and it passes in isolation./retryrecovers it today.Linked Issues
Found while analysing why #6723 had burned 7 agent rounds under takeover with nothing pushed. Part of the autofix-reliability line: #7330, #7351, #7368, #7412, #7416, #7438.
中文说明
本 PR 做了什么
新增一个连续失败子上限,独立于总轮次上限,让"每轮都推送失败"的 PR 停下来,而不是在 takeover 下一路重试到 100。
为什么需要
轮次上限回答的是"一个 PR 可以用多少有产出的轮次",takeover 特意把它提到 100,是给需要多轮"处理反馈→推送"的 PR。它没有回答"可以连续失败多少轮"。
在 #6723(5700 行、47 文件、跑了 5 天、与飞速前进的
main冲突)上实测:7 轮全败(3 次 agent 超时 50 分钟、4 次门拒绝——修复过不了测试),历时 8 小时零推送,正冲着 round 100 去,每轮要么 50 分钟 agent 运行、要么一整轮门检查。每轮都在重解一个解不完/过不了测试的冲突。以相同的单轮预算重试不会收敛,只能靠人 rebase 或拆分。两类失败(超时、门拒绝)各自都有处理,但没有任何机制注意到它们正在背靠背、毫无进展地连续发生。
怎么做
CONSECUTIVE_FAILURE_CAP: '5',独立于TAKEOVER_MAX_ROUNDS: '100'。能走到交接步骤,本身就意味着这一轮没有推送(推送与 no-op 路径从另一个步骤汇报)。于是该步骤统计"未被打断的连续失败":倒序遍历 bot 自己的 eval 评论标题,遇到第一个推送(
Addressed the latest review feedback)或合法 no-op(no changes needed)即停——两者都证明循环在推进,连续计数在此重置。若这条连续段(含本轮)达到上限,即使在 takeover 下也强制终态,并给出真正的修复指引:rebase/拆分/精简该 PR,再@qwen-code /retry重新武装。它不区分成因:超时和门拒绝都算"失败且未推送"。连续段中任意一次推送或 no-op 都会重置它,因此只会在真正卡住的 PR 上触发。
评审验证
npx vitest run scripts/tests/qwen-autofix-workflow.test.js—— 92/92。新测试在 bash 下回放真实提取的计数+override 块,用桩gh喂合成评论历史,覆盖 6 种组合(见上方英文表格:无历史=1 不终态;4 败+本轮=5 终态;3 败+本轮=4 不终态;推送重置;no-op 重置;超时与门拒混合仍达 5 终态)。retries a verification-gate crash原来靠"结尾fi+ 报告{"抓取决策块,而本块正插在两者之间。现改为锚定本块的注释,只驱动 crash 决策——行为不变,不调用gh。run:块过bash -n;actionlint 1.7.12 clean;prettier clean。风险与范围
/retry会重新武装、开启新窗口。gh api …/comments。拉取失败时PRIOR_HEADS为空、连续数为 1——因此一次 API 抖动不会误判终态(fail-open)。takeover-command toggle在全量子进程负载下失败——本机 A/B 显示未修改的main以相当频率红同一测试,单独跑通过。/retry恢复。关联 Issue
在分析 #6723 为何在 takeover 下烧掉 7 个 agent 轮次却零推送时发现。属 autofix 可靠性主线:#7330、#7351、#7368、#7412、#7416、#7438。