perf(review): remove serial wall-clock the audit loop and tail never needed - #8642
Merged
Conversation
wenshao
requested review from
LaZzyMan,
doudouOUC,
tanzhenxin and
yiliang114
as code owners
August 6, 2026 12:41
Collaborator
|
🔄 Qwen Triage is running — watch live progress. Stage results will post in this thread as they complete. 🔄 Qwen Triage 正在运行 —— 查看实时进度。各阶段结果完成后会更新在本线程。 |
Collaborator
|
Thanks for the PR! Template looks good ✓ — all sections present, with an unusually thorough evidence base. Problem: observed, not theoretical. Six real CI review runs (with run IDs), stage-level timings extracted from the Direction: aligned. A median-141-minute review run that barely correlates with PR size is a real merge-throughput bottleneck, and this complements rather than overlaps #8602 (hangs) and #8631 (fan-out serialization). CHANGELOG: no direct reference — this is our own internal review pipeline, so that's expected. Size: touches Approach: the scope feels right. Each of the four cuts is anchored to a measured incident, the quality-relevant invariants are explicitly preserved (two independent dry audits before convergence, every finding verified, whiff handling, 3B retirement), and the out-of-scope list is honest (effort-tier policy, per-agent tool budgets, 3B pairing). I looked for a simpler path and didn't find one — the convergence pair is the minimal change that removes the serial second round, and dropping the env-var preference entirely (rather than reordering it) is right given it was wrong 100% of the time it was consulted. No unrelated edits in the diff. Risk: no elevated risk signals — none of the changed files match the high-risk paths from the revert-history analysis. Moving on to code review. 🔍 中文说明感谢贡献！ 模板完整 ✓ —— 各节齐全，证据基础扎实。 问题：已观测而非理论。6 个真实 CI review run（附 run ID）、从 方向：对齐。中位 141 分钟且与 PR 大小几乎无关的 review 耗时是真实的合入吞吐瓶颈，本 PR 与 #8602（挂死）、#8631（扇出串行化）互补而非重叠。CHANGELOG：无直接引用——这是我们自己的内部 review 流水线，符合预期。 规模：触及 方案：范围合理。四处削减各自锚定一个实测事故，质量相关的不变量被显式保留（收敛前两次独立 dry 审计、每个 finding 都过验证、whiff 处理、3B retirement），范围外清单也诚实（effort 档位政策、单 agent 工具预算、3B 配对）。我找过更简路径，没找到——收敛对是移除串行第二轮的最小改动；鉴于该 env 偏好被咨询时 100% 是错的，彻底弃用（而非调整优先级）是对的。diff 无无关改动。 风险：无升级风险信号——改动文件均未命中 revert 历史分析的高风险路径。 进入代码审查 🔍 — Qwen Code · qwen3.8-max Reviewed at |
Collaborator
Code reviewNo blockers found. I verified the load-bearing mechanics claims against the base code rather than taking the description at face value:
Independent proposal matched: my pre-diff plan was the same four cuts, and the diff handles the edge cases I'd have asked about (pair dedup, convergence-count restart after a reporting pair, retroactive-dry timing one round late under pipelining) explicitly in the text. Nothing simpler occurred to me that the PR missed. Test evidence — the PR's own CI, read via APIFetched once at the reviewed commit; no polling. At fetch time: 0 failures, Final CI results for
One row per check name (latest run); skipped checks omitted; failures sort first. / 每个检查名一行（取最新一次运行），省略 skipped，失败项排在最前。 What the suite pins (per the diff, not re-run here): the save-artifact suite pins explicit Sandboxed verification would settle this: 中文说明代码审查：未发现阻塞项。关键机制声明均对照 base 代码核实，而非仅凭描述采信：
独立提案吻合：我看 diff 前的方案就是这四处削减；diff 对我会追问的边界情形（配对去重、报告对后收敛计数重启、流水线下的追认 dry 晚一轮到达）都在文本中显式处理。没有更简路径被遗漏。 测试证据：在被审 commit 上经 API 一次性读取，不轮询。读取时：0 失败， 沙箱验证可落定此事： — Qwen Code · qwen3.8-max Reviewed at |
Collaborator
Author
Code Review —
|
Collaborator
|
Confidence: 4/5 — clean review across every stage; the one open item is inherent to the change: the wall-clock win is counterfactual until live review runs execute the new skill text. Stepping back: this is what a good perf PR looks like. Every cut is anchored to a named, measured incident — not "this looks parallelizable", but "run X spent 22 minutes waiting on verdicts its launch never consumed", "run Y shipped a capped verdict for want of a second dry audit it had time to launch in parallel but not in series". I independently proposed the same four cuts before reading the diff, and the diff goes further than my proposal in exactly the right places: the edge cases (reporting-pair dedup, convergence-count restart, retroactive-dry arriving one round late under pipelining) are handled in the text instead of being left to orchestrator discretion — which is precisely the failure mode fix 1 removes. The quality invariants hold up under scrutiny: two independent substantively-dry audits before convergence, every finding verified, whiffed scopes staying not-audited, 3B untouched. The tag backstop plus Reservations, named so they aren't forgotten: the "after" numbers are counterfactual (the PR says so itself), so the −5% to −30% range is a promise until the next CI reviews land — the Approving — deferred until CI lands green on 中文说明退一步看整体：这是一个 perf PR 该有的样子。每处削减都锚定一个有名有姓的实测事故——不是"这里看起来可以并行"，而是"run X 等了 22 分钟一个它的发射根本不消费的验证结果"、"run Y 为一个来得及并行、来不及串行的第二次 dry 审计背上了 capped verdict"。我在读 diff 前独立提出了同样的四处削减，而 diff 在正确的方向上超出了我的提案：边界情形（报告对去重、收敛计数重启、流水线下追认 dry 晚一轮到达）被写进文本，而不是留给编排器自由裁量——后者正是修复点 1 要消灭的失败模式。 质量不变量经得起推敲：收敛前两次独立、有实质收据的 dry 审计、每个 finding 都过验证、whiff 范围保持未审计、3B 未触碰。标签兜底加 保留意见，点名以免被遗忘："后"的数字是反事实的（PR 自己也这么说），所以 −5% ~ −30% 区间是一个承诺，要等下一批 CI review 落地才兑现——上面点名的 批准——延迟至 CI 在被审 commit 上落绿后执行。 — 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. |
Collaborator
Author
|
All five suggestions addressed in 71a2ba3:
Also in flight: a live qwen3.8-max run of the updated skill against PR 8619 (same CI invocation shape) to convert the counterfactual savings into a measured before/after; will post the stage table when it completes. — Claude Code · Opus 5 |
wenshao
force-pushed
the
perf/review-pipeline-latency
branch
from
August 6, 2026 13:45
71a2ba3 to
e0e4d50
Compare
Contributor
|
Please do not rebase or force-push to an active PR as it invalidates existing review comments. Note for future reference, the bots always squash all changes into a single commit automatically as part of the integration. 中文请勿对活跃的 PR 执行 rebase 或 force-push，因为这会使已有的评审评论失效。另外，供日后参考：作为集成流程的一部分，机器人始终会自动将所有改动压缩（squash）为单个提交。 |
added 4 commits
August 6, 2026 22:49
…needed Four latency cuts, each backed by stage-level timings extracted from six real CI review runs (2026-08-05/06); none changes what a review covers or what evidence a verdict requires. - Step 4's initial verification rides with the first reverse-audit launch (the k=0 case of the existing pipelined-loop rule). One measured run held its round-1 auditor 22 minutes behind verdicts the launch never consumed, while a sibling run the same day launched the two together. - Reverse-audit rounds 1 and 2 become the convergence pair on 3A plans: two independent auditors launched together, both-dry = the two dry audits the criterion already demands. A dry round leaves the findings list unchanged, so serial round 2's input was byte-identical to round 1's; the pair removes the wall clock between them, not either audit. Also spares the budget-stop cap a run took for want of a second dry round it had time to run in parallel but not in series. - A reporting round whose every finding the verifier rejected is retroactively dry: rejections already remove the entries, but not the reset they applied to the dry counter, and one measured run spent two extra rounds re-proving what round 2 plus a rejection had already answered. - save-artifact resolves its workspace root from cwd (or an explicit --workspace-root), not QWEN_CODE_PROJECT_DIR - the harness exports that variable as the session-storage dir under the runtime base, never the main checkout, and all six runs burned minutes rediscovering it before improvising workarounds. Steps 8/9 also batch the tail bookkeeping into four responses instead of one command per model turn.
- Pin the cwd default with the trap armed: the regression test now chdirs into the root, passes no workspace root, and points QWEN_CODE_PROJECT_DIR at a decoy - reintroducing the removed env preference turns it red (verified by negative control). - Have the skill pass --workspace-root explicitly in the Step 8 block, so the containment anchor never depends on where the command happened to run. - Pin the retroactive-dry pairing direction: forward only, consulted on a round's return; the upgrade never ends the loop by itself and never discards an in-flight round. - Make the Step 4 batch self-contained (findings files written before either prompt set is built), and gate tail group 4 on group 3's success so cleanup cannot destroy save-artifact's only inputs after a failure. - Add the defensive rule for a deadline-refused pair member: launch round 1 alone and treat the refusal as the budget stop.
…t review Ten findings from a 9-agent review round plus one reverse-audit round over the PR's own diff; all verified against the cited lines before fixing. - Step 4 k=0 batch: pre-confirmed [build]/[test] entries enter the cumulative list untagged (the every-entry-tagged wording would have sent deterministic findings into the tag backstop and capped the verdict); the no-shard clause covers the all-pre-confirmed case, not just zero findings. - Convergence pair x retroactive-dry: the two rules now compose instead of conflicting - a pair retired in full counts as the dry predecessor at round 3's return; round k-1 counts launches, not labels; verdicts are applied at a round's return before the dry consultation; the pair's verification is one build over the deduped union passed as --round 2. - Tail batching: a group (3) that cannot be resolved no longer gates cleanup forever - disclose, skip record_artifact, preserve the inputs beside the report, and still run cleanup so the bypass audit and the completion line always happen; the old before-cleanup ordering sentence now names the group (3)/(4) split instead of contradicting it. - Accuracy: "byte-identical" weakened to what the pipelined merge rule actually guarantees, and the pair's round-2 suppression window named as the already-accepted trade; the #8353 incident recast to the forward pairing the rule licenses (round 5 was the waste, not rounds 4-5) with the ledger-vs-audit round numbering disambiguated. - save-artifact: comment block and test comments no longer describe the flag as embedder-only - it is the containment anchor the skill passes on every run; new CLI option contract test drives --workspace-root through real yargs into saveReviewArtifact (mirrors the test-plan/test-delta precedent for camel-case flag regressions).
Round-2 reverse audit: the pair bullet's "one verify build over the deduped union" contradicted Step 4's verifyShard sharding (quality collapses past 8 findings per verifier) and its own plural "the pair's verifiers". The rule now shards the deduped union exactly as any reporting round's findings are, every shard passed as --round 2; the per-member ban is unchanged.
wenshao
force-pushed
the
perf/review-pipeline-latency
branch
from
August 6, 2026 14:49
e06ef40 to
2f1b763
Compare
yiliang114
approved these changes
Aug 6, 2026
yiliang114
left a comment
yiliang114
left a comment
Collaborator
There was a problem hiding this comment.
LGTM, no blockers. Code: removing the QWEN_CODE_PROJECT_DIR preference is correct (measured wrong 100% of the time — it names the session-storage dir, not the checkout); the explicit --workspace-root + cwd fallback is a trust improvement, and the decoy test (env never consulted) plus the yargs-contract test (parsed object fed straight in, catching camelCase flag bugs) are exactly the right pins. Orchestration: pipelining the initial verification (k=0), the 3A convergence pair, the retroactively-dry rule, and Step 8/9 batching all remove serial wall-clock while preserving the safety invariants (last-round verification before Step 6, tag backstop, budget gate, unreviewedDimensions cap); each is backed by measured DESIGN.md data and the retroactively-dry rule is carefully constrained (forward-only, never ends the loop alone, a launched auditor's return is always taken). No P0/P1.
Collaborator
Author
|
@qwen-code /triage |
Collaborator
|
🔬 Sandboxed verification is running — watch live progress. The report will be posted here when the run completes. 🔬 沙箱验证正在运行 —— 查看实时进度。运行结束后验证报告会发布在这里。 |
wenshao
enabled auto-merge
August 6, 2026 15:10
Collaborator
Author
|
Closing the loop on the promised validation run (live qwen3.8-max, CI-faithful headless replay against PR 8619, updated skill bundled): Confirmed executing as specified — the launch shapes the counterfactual savings depend on:
Caveat, disclosed: the run's subcommands were contaminated by a local-only version skew (launching via The real before/after lands with the next few CI reviews now that this is on main — happy to compare stage tables against the six-run baseline once a couple complete. — Claude Code · Opus 5 |
Collaborator
|
Released in v0.21.7. |
This was referenced Aug 9, 2026
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
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
Four latency cuts to the
/reviewpipeline, each derived from stage-level timings extracted from six real GitHub-triggered review runs (2026-08-05/06). None of them changes what a review covers, which agents run, or what evidence a verdict requires — they remove wall-clock that was serial by accident, not by design.— [unverified]tags for exactly this state), while a sibling run the same day launched the two together. Now the initial verification is the k=0 case of the same rule, on both topologies.--round 1and--round 2, same--findingsfile; the CLI already keys the two records apart by round, so this needs zero CLI change. Both dry = the two independent, substantively-dry audits the two-consecutive-dry criterion already demands: a dry round leaves the cumulative list unchanged, so serial round 2's launch input was substantively identical to round 1's — the same entries, at most with verification tags cleared in between — an independent rerun bought at the price of a full round of wall clock. If either member reports, the pair is one reporting round and the loop continues serially from round 3 with convergence counting restarted, so the non-dry path loses nothing. 3B is untouched (rounds already fan out per chunk; the retirement schedule is the convergence ledger there).save-artifactresolves its workspace root from cwd (or an explicit--workspace-root), never fromQWEN_CODE_PROJECT_DIR— plus Steps 8/9 batch the tail bookkeeping into four responses instead of one command per model turn. The env preference was wrong 100% of the time it was consulted: the harness exports that variable as the session-storage directory under the runtime base (Storage.getProjectDir(), where the harness's transcripts live), never the main checkout, in every environment. All six measured runs refused their own inputs at Step 8 and burned 1.5–3 minutes improvising workarounds (override the env var; copy inputs into the qwen-home mirror; retry path shapes). The skill runs every subcommand from the main checkout, so cwd is right by construction.Five new measured-incident entries in DESIGN.md anchor the new rules, enforced by the existing SKILL.test.ts pointer↔heading consistency gate.
Why it's needed
/reviewhas become the bottleneck on merge throughput: of the last 81 executed review runs, the 33 that ran a real review (≥30 min, all successful) took median 141 / p75 236 / p90 277 minutes, and duration barely correlates with PR size — a 29-line PR took 65 minutes, a 72-line PR 138 minutes, a 1-line docs PR 57 minutes.Stage decomposition of two small-PR runs (timestamps from the CI job logs):
The audit-loop tail and bookkeeping stages above are serial waste: #8619 and #8607 both converged at the minimum two rounds, run serially with byte-identical inputs; the #8501 round-5 review had its round 2 refused by the deadline gate and shipped a capped verdict for want of a second dry audit it had time to run in parallel but not in series (fix 2 also removes that cap scenario); the round-15 review of #8353 ran a fifth audit round (~15–20 min) after round 4's dry return had already completed the two-dry evidence alongside a verifier rejection (fix 3). Counterfactual totals for the six runs: −5% to −30% wall clock, with the quality-relevant loop structure — two independent dry audits before convergence, every finding verified, whiff handling, 3B retirement — unchanged.
Complements #8602 (stream lifetime cap + slim launch blocks) and #8631 (daemon fan-out concurrency): those fix hangs and serialization; this removes work-shaped waiting that remains after them.
Reviewer Test Plan
How to verify
cd packages/cli && npx vitest run src/commands/review/— 56 files, 1991 passed. The save-artifact suite now pins: explicit--workspace-rootresolution, cwd default, and a regression test thatQWEN_CODE_PROJECT_DIRis ignored entirely (decoy-root test).cd packages/core && npx vitest run src/skills/bundled/review/SKILL.test.ts— the pointer↔heading gate covers the five new DESIGN.md incidents.agent-promptkeys findings-role records asreverse-audit--round-<k>--<digest>(agent-prompt.ts), so--round 1/--round 2with the same findings file are two records, two briefs, two receipts; the deadline gate prices a paired round-2 admission at the 600-second observed floor (deadline.tsexpectedRoundSecondstakes the max span), so a pair admits whenever round 1 does and round 3's admission measures the pair's real cost.Evidence (Before & After)
The six dissected runs above are the before. After is counterfactual until the next CI reviews land: fix 1 saves 8–22 min on the ~half of runs that serialized the initial verify; fix 2 saves 13–25 min on min-convergence runs (2 of 6) and converts #8501-shaped budget caps into clean convergences; fix 3 saves one round (~15–20 min) on rejected-finding shapes; fix 4 saves 5–8 min on every run.
Tested on
Environment (optional)
Vitest suites from
packages/cli/packages/core; stage timings extracted from thereview-prjob logs of runs 31089304225 (#8619), 31080454889 (#8628), 31082388591 (#8607), 31080978821 (#8353), 31089156499 (#8594), 31085374670 (#8501).Risk & Scope
save-artifactbehavior change: an embedder that deliberately exportedQWEN_CODE_PROJECT_DIRas a workaround (what the measured CI runs improvised) now gets cwd resolution instead — correct in every observed case, since those runs ran from the main checkout;--workspace-rootis the explicit escape hatch.--comment, a policy change), per-agent tool budgets (the 11-vs-41-minute fan-out wave variance — named by fix(cli): Run ACP agent fan-outs concurrently and past the tool-call cap #8631 as the next target), and 3B pairing of rounds 1+2 (would interact with retirement/deadline pricing; not measured as a bottleneck).Linked Issues
Related: #8602, #8631 (complementary
/reviewperformance work).中文说明
这个 PR 做了什么
对
/review流水线的四处延迟削减，全部来自对 6 个真实 GitHub 触发 review 运行（2026-08-05/06）的日志逐阶段计时。它们不改变 review 的覆盖范围、agent 名单或裁决所需的证据——移除的是"偶然串行"而非"设计串行"的墙钟。— [unverified]标签支持这个状态），同一天的另一个 run 则把两者一起发。现在首次验证是同一规则的 k=0 情形，两种拓扑都适用。--round 1与--round 2、同一--findings文件；CLI 的记录键本就按轮号区分，零 CLI 改动。双双 dry = 两连 dry 判据要求的"两次独立、有实质收据的 dry 审计"：dry 轮不改变累计清单，串行第 2 轮的输入与第 1 轮实质相同——同一组条目，至多验证标签在间隙被清除——是一次用整轮墙钟买来的独立重跑。若任一成员报告 finding，收敛对算一个报告轮，从第 3 轮起恢复串行、收敛计数重新起算，非 dry 路径无任何损失。3B 不动（轮次本就按 chunk 扇出，retirement 调度就是那里的收敛台账）。save-artifact的 workspace 根从 cwd（或显式--workspace-root）解析，不再读QWEN_CODE_PROJECT_DIR；同时 Step 8/9 把尾部记账合批为 4 个响应（原来一命令一模型轮）。旧的 env 偏好 100% 是错的：harness 导出的该变量是 runtime base 下的会话存储目录（Storage.getProjectDir()，transcripts 所在），任何环境下都不是主 checkout。6 个实测 run 全部在 Step 8 被拒收自己的输入、烧 1.5–3 分钟各自摸索绕法。skill 本就要求所有子命令从主 checkout 运行，cwd 天然正确。DESIGN.md 新增 5 条 measured-incident 条目锚定新规则，由既有 SKILL.test.ts 指针↔标题一致性测试强制。
为什么需要
/review已成为合入吞吐率的瓶颈：最近 81 个实际执行的 review run 中，33 个真正跑完整流程的（≥30 分钟，全部成功）耗时中位 141 / p75 236 / p90 277 分钟，且与 PR 大小几乎无关——29 行的 PR 65 分钟，72 行的 138 分钟，1 行的 docs PR 57 分钟。两个小 PR run 的逐阶段分解（时间戳取自 CI job 日志）：#8619（29 行，59 分钟）setup 3.5 / 扇出 11.7 / verify 与 RA1 并行 15 / RA2 12.8（修复点 2）/ 提交 7.8 / 尾部 5.3（修复点 4）；#8628（72 行，130 分钟）setup 3.2 / 扇出 41 / verify 串行 22.4（修复点 1）/ RA1 22 / RA2 18 / 提交 12 / 尾部 4.6。
其中的串行浪费：#8619 与 #8607 都以最小两轮收敛，却用逐字节相同的输入串行跑了两轮；#8501 的 round-5 review 第 2 轮被 deadline gate 拒绝，为一个"来得及并行、来不及串行"的第二次 dry 审计背上了 capped verdict（修复点 2 同时消掉这个场景）；#8353 的 round-15 review 在第 4 轮 dry 返回加一次驳回已凑齐两连 dry 证据后，仍多跑了第 5 轮（约 15–20 分钟）（修复点 3）。6 个 run 的反事实合计：墙钟 −5% ~ −30%，而质量相关的循环结构——收敛前两次独立 dry 审计、每个 finding 都过验证、whiff 处理、3B retirement——原封不动。
与 #8602（流总时长上限 + 发射块瘦身）、#8631（daemon 扇出并发）互补：那两个修挂死与串行化，本 PR 移除其后仍剩的"长得像工作的等待"。
审查者验证计划
如何验证：
cd packages/cli && npx vitest run src/commands/review/——56 个文件、1991 通过；save-artifact 套件钉住显式--workspace-root解析、cwd 默认值、以及"完全忽略QWEN_CODE_PROJECT_DIR"的诱饵根回归测试。cd packages/core && npx vitest run src/skills/bundled/review/SKILL.test.ts——指针↔标题一致性测试覆盖 5 条新 DESIGN.md 条目。收敛对的机制核查（零 CLI 改动，可自行确认成立）：agent-prompt对 findings 角色的记录键为reverse-audit--round-<k>--<digest>，同一 findings 文件的--round 1/--round 2是两条记录、两份 brief、两张收据；deadline gate 对成对第 2 轮的准入按 600 秒观测下限计价（expectedRoundSeconds取最大跨度），第 1 轮能准入则第 2 轮必然能，第 3 轮准入实测收敛对的真实成本。SKILL 语义：读 Step 4 "the verifiers do not launch alone"、Step 5 "The convergence pair (3A only)" 与两条新终止规则，确认两次独立 dry 审计要求、whiff 规则、3B retirement/CONVERGED 路径均未被削弱。前后证据：前=上述 6 个解剖 run；后=反事实，待后续 CI review 落地实测。修复点 1 在约一半 run 上省 8–22 分钟；修复点 2 在最小收敛 run（6 中 2）省 13–25 分钟、并把 #8501 型 budget cap 变为干净收敛；修复点 3 在驳回型 run 省一轮（约 15–20 分钟）；修复点 4 每个 run 省 5–8 分钟。
测试平台：Linux ✅（macOS/Windows：CI）。测试环境：
packages/cli/packages/core下 vitest；阶段计时提取自 run 31089304225（#8619）、31080454889（#8628）、31082388591（#8607）、31080978821（#8353）、31089156499（#8594）、31085374670（#8501）的review-prjob 日志。风险与范围
save-artifact行为变化：曾把导出QWEN_CODE_PROJECT_DIR当绕法的接入方（即实测 CI run 的即兴做法）现在得到 cwd 解析——在所有观测场景下均正确（它们本就从主 checkout 运行）；--workspace-root是显式逃生口。--comment，属政策变更）、单 agent 工具预算（扇出波 11 vs 41 分钟的方差，fix(cli): Run ACP agent fan-outs concurrently and past the tool-call cap #8631 已点名为下一目标）、3B 的 1+2 轮配对（与 retirement/deadline 计价有交互，且未实测为瓶颈）。关联 Issue
相关：#8602、#8631（互补的
/review性能工作）。