Skip to content

fix(ci): preserve Maven suite results across cache write-back failures - #15274

Merged
yiliang114 merged 43 commits into
mainfrom
autofix/issue-15262
Nov 7, 2026
Merged

yiliang114 merged 43 commits into
mainfrom
autofix/issue-15262

Conversation

@qwen-code-dev-bot

@qwen-code-dev-bot qwen-code-dev-bot commented Oct 22, 2026 •

Copy link
Copy Markdown
Collaborator

What this PR does

This PR hardens the SDK Java Maven cache path after the suite/write-back lock split. Failed MySQL integration suites can still publish downloaded artifacts while preserving the original suite status, cache setup failures preserve the already-known suite result, and the final error annotation identifies the failed Maven stage. Cache seeds report copy failures, tolerate an unavailable cache or failed best-effort prune, and detach mutable resolver metadata before Maven runs — including single-link resolver state inherited from a previous job, since RUNNER_TEMP persists across jobs on the self-hosted pool. Write-back excludes the same metadata and partial-download files.

Why it's needed

The lock relocation associated with #15262 already landed on main through #15128 and #15254. This branch is the remaining hardening delta: a failed suite should not throw away reusable downloads, a best-effort cache operation should not replace the suite result, and mutable resolver state should not be published as an immutable shared artifact. The test repair in 11c0f4b also includes the deferred-status reporter in extracted shell snippets so those tests can exercise their intended paths.

Reviewer Test Plan

How to verify

  • Fail either Maven stage in the self-hosted MySQL integration lane. Reusable downloaded artifacts should reach write-back, later suites should not run, and the step should retain the failed stage's status and annotation.
  • Run a green and a failed suite with an unusable cache or missing job-local repository. The cache fault should produce a warning and preserve the suite result.
  • Seed a warm cache containing resolver markers and version metadata. Mutable hardlinks should be detached before Maven runs; write-back should publish artifacts while excluding resolver markers, metadata and partial-download files.
  • Exercise a cold cache and a seed prune failure. Both should still reach Maven; an actual copy failure should remain visible in the log.
  • Plant a plain single-link resolver-status.properties / x.jar.lastUpdated directly in ${RUNNER_TEMP}/m2-repo before the seed body runs. The seed prune must delete both while the host-cache peers stay intact. Restoring -links +1 to the seed prunes must red the clears inherited single-link resolver state from the job-local repository witness.

Evidence (Before & After)

Head 26960a72d1 merges main a5fec03684 into 89d3898f4d; the PR delta is unchanged (the same seven files, 1116 insertions / 198 deletions), and the merged workflow's .size-baseline entry is re-recorded to 162545. Earlier measurements below are labeled with the commit they were taken at; the four required suites were re-run after this merge. Prior heads: a0d162dbcc (main 90eb13e4c6), d0f6312b82 (main 91a6274075), 7618d49a (main fe30b5daaa).

Check Result Measured at
sdk-java-workflow + sdk-java-lock + workflow-size + main-ci-failure-issue-workflow 505 passed | 47 skipped (552), re-run after the latest merge 26960a7, local macOS (a0d162d: 502/47)
Wider set: six other sdk-java.yml parsers + qwen-triage-workflow 241 passed | 0 skipped; 191 passed | 10 skipped (triage's pre-existing 5 s subprocess timeouts on a tmux-less Mac pass with --testTimeout=30000; no test modified) 26960a7, local macOS (a0d162d ten-suite set: 743/47 + 191/10)
Maintainer Linux verification: 34 replay cells + author guard suite + 18-row mutation matrix 61/61 assertions; guard suite 142/142 on Linux; 18/18 mutations killed, no survivors — verdict merge-ready (report in the PR thread) 7618d49, Linux node:22-bookworm non-root, 2026-10-31
SDK Java (fast) / ubuntu-latest / Java 21 — real self-hosted pool lane pass, 30m25s 7618d49 CI
SDK Java (full) tier on the real pool — Runtime Broker and Managed Agent MySQL (1h13m), Hosted process fault (1h54m), EventTransport broker / multi-node, Java 11/17/21, Real daemon E2E, Flyway all pass 7618d49 CI
Six other suites parsing sdk-java.yml (spotbugs-gate, hosted-process-ci, managed-agent-gate-workflow, managed-agent-server-e2e, event-transport-multi-node-ci, event-transport-broker-ci) 236 passed | 0 skipped c774661, local macOS; PR delta unchanged since
Mutation: restore -links +1 at the four seed prunes Witness red, vitest exit 1 (inherited single-link marker survived the seed prune: expected true to be false); fix restored, green. Re-killed on Linux as row m5b at 7618d49 c774661 local macOS; 7618d49 Linux
.size-baseline sdk-java.yml entry 162545 = exact git show HEAD:.github/workflows/sdk-java.yml | wc -c (159271 at a0d162d) 26960a7
prettier / eslint / git diff --check Clean; pre-commit hooks ran c774661 and d0f6312
Lint & Static / Test (ubuntu) / web-shell E2E Smoke CI pass after re-run (first attempts failed on fleet network: actionlint download curl timeouts; corepack pnpm download ETIMEDOUT/ENETUNREACH at install; smoke's unrelated previews qwen3.8-max reasoning flake). CI re-runs at the merged head 7618d49 CI; d0f6312 pending

The 47 macOS skips are the file's Linux-only behavioral pins; the maintainer's Linux round executes them (142/142), so the macOS counts alone do not certify the changed behavior — the Linux verification does. The new inherited-state witness is cross-platform and ran on both systems. Prior Linux A/B and real-pool reports remain historical evidence at their recorded heads.

Tested on

OS Status
macOS 502 passed | 47 skipped across the four required suites at d0f6312 (re-run after the merge); Linux-only behavioral pins skip here
Linux Maintainer replay + guard suite 142/142 + 18/18 mutations at 7618d49; CI fast Java 21 and the full SDK Java tier pass at 7618d49; merged head CI pending
Windows Not run locally

Current closeout

Round of 2026-10-31 (local takeover after the AutoFix loop reached its round cap): merged main 812a8c7ccb, resolving the write-back conflicts semantically — main's nine-coordinate prune list (from the #15871 Maven-module split) plus this PR's seven-name ! -name exclusions, with the matching test pins moved along — and fixed R13-1 by dropping -links +1 from all four seed prunes so inherited single-link resolver state is cleared again; the write-backs' -links 1 delta filter is unchanged. Both design-doc languages are synchronized and .size-baseline is re-recorded to the exact byte count. Unlike the previous static-only closeout, this round ran the suites and the mutation locally (evidence table above). The merge-order decision R2-6 asked for is recorded: this PR is the chosen landing PR of the family; #15275, #15284 and #15233 are closed, #15254 merged, and #15286 remains the separately owned follow-up.

Later on 2026-10-31: the maintainer merged current main (7618d49a, PR files untouched) and published an independent Linux verification at that head — verdict merge-ready, 61/61 assertions, 18/18 mutations killed. One non-blocking Suggestion from that report is deferred and recorded here per the repository's post-five-round convergence rule: when the managed-agent-server checkout is missing, the deferred-failure annotation names the previous successful Maven stage instead of the checkout (both lanes). The report carries a measured minimal fix (set mvn_stage before the cd; the test pin counting mvn_stage occurrences must move with it). It is follow-up work, not a merge blocker; the deferral is recorded so nothing is silently dropped.

2026-11-01: merged main 91a6274075 (d0f6312b82) to clear the base conflict, keeping main's Hosted-admission change (#16010) alongside this PR's hardening; delta unchanged, baseline re-recorded (155529), required suites re-run green. At 7618d49a CI was fully green (including the full SDK Java tier on the real pool) and the PR held two maintainer approvals (wenshao, yiliang114); CI re-runs at the merged head, and stale-approval dismissal rules may require the two approvals to be refreshed. No review thread is unresolved.

2026-11-02: the maintainer merged main once more (702da7a21a, merging 60f5c1ed95); the PR delta is unchanged apart from main-side comment edits inside sdk-java.yml, which now measures 156105 bytes while .size-baseline still records 155529 — a 576-byte record nit the CI size check tolerates, to be re-recorded with the next content change. All checks at this head are green, including the re-run review-pr whose round-15 review posted zero findings. The two approvals from 7618d49a were dismissed on push by the stale-approval rule; the PR is MERGEABLE and now waits on two fresh maintainer approvals at 702da7a21a. The round-15 deferred items stay recorded follow-ups: the baseline record, the MySQL cd-deferral/stage-label pins, the classifier fixture's lock-wait line, and the MySQL write-back comment wording.

2026-11-03: merged main 90eb13e4c6 (a0d162dbcc) to clear the new base conflict, keeping main's optional-runtime-backend Maven modules (#16141) alongside this PR's hardening; delta unchanged, .size-baseline re-recorded to 159271 (which also clears the 576-byte nit above), and the four required suites re-run green (502 passed | 47 skipped) plus the wider ten-suite set (743 passed | 47 skipped). CI re-runs at the merged head; the two approvals must be refreshed at a0d162dbcc since every push dismisses the previous ones.

2026-11-04: merged main a5fec03684 (26960a72d1) to clear the new base conflict, keeping main's disk-exhaustion guards (#16068) alongside this PR's hardening; delta unchanged, .size-baseline re-recorded to 162545, four required suites re-run green (505 passed | 47 skipped) plus six other parsers (241 passed | 0 skipped) and triage (191 passed | 10 skipped). The two web-shell smoke failures before the merge were the SSE-handshake race on a busy pool host and were re-run; the push briefly hit a GitHub-side Internal Server Error during the day's minor outage and succeeded after a short wait (the watch's worktree now pushes over HTTPS with gh auth git-credential, GitHub SSH having refused connections since this morning). CI re-runs at the merged head; two fresh approvals needed at 26960a72d1.

2026-11-07: merged main cb282ef5cfe (54edbe30b9a) to clear the new .size-baseline conflict, keeping the runner-local Java tool-cache repair from main alongside this PR cache hardening. The effective PR delta is unchanged: the same seven files with identical per-file addition/deletion counts (1116 insertions / 198 deletions). .size-baseline is re-recorded to 165241, exactly the merged workflow byte count. The four required suites re-run green (506 passed | 47 skipped), six other sdk-java.yml parsers pass (242/242), the new Java tool-cache suite passes (8/8), and prettier plus git diff --check are clean. The latest independent Linux verification at f4981199200a covers the identical PR delta and reports merge-ready with 122/122 assertions; PR CI at 54edbe30b9a completed with 48 passing and 35 skipped checks: one substantive red lane remains, full-tier Hosted process fault gates, plus its aggregate. The failures are two test assertions, not a Maven build failure: HostedPublicWorkspaceIT.structuredQuestionSurvivesJavaAndHarnessRestart and HostedWorkspaceToolTurnIT.recordsLatencyWithDelayedRuntimeProvisioning. Both also failed in independent PR #16469 full-tier run 42242651788 on shared ancestor 7cb46d9fed; the shared regression is tracked by #14724. No PR-specific regression was identified, no unrelated Java change was added here, and no speculative 97-minute retry was launched. Fresh approvals are required after this push.

Risk & Scope

The host cache remains best-effort. Failure to seed can increase network resolution. An unexpected bare-flock failure remains fatal. Real Maven resolver behavior and contention on the self-hosted pool require CI evidence beyond shell fixtures. The seed prune now also deletes inherited single-link markers in the job-local repository; deleting a job-local hardlink never removes the host peer.

#15286 contains distinct follow-up work: existing-destination seed nesting, removal of old in-house artifacts from the host cache, self-hosted Checkstyle repository selection, and reference-release cache isolation. Those changes remain separately owned. Shared-helper and test-fixture extraction are deferred rather than widening this repair.

Related to #15262; the original lock relocation is already on main.

Latest repair

c7746613ec clears inherited resolver state in the Maven cache seed prune (R13-1): the four seed prunes drop -links +1, the deAlias pin moves with them, and a cross-platform regression witness plants nlink-1 markers before the seed body — mutation-verified red (exit 1) with -links +1 restored and green with the fix. The previous repair (300-second failed-MySQL wait clamp, metadata checksum sidecar and partial-download exclusions) is unchanged.

The Managed Agent fast tier can fail without reporting a test result after waiting out the per-host SDK Java lock. #15128 scoped the MariaDB lane's lock to its cache write-back, but the test job still held the same lock across the SDK and Managed Agent Maven suites, so a long full-tier suite could starve the fast tier on that host.

Seed the test job's Maven repository from the host cache, run both suites unlocked from that job-local copy, and take the ticketed host lock only for the cache write-back. The write-back children close fd 9 so a cancelled job's survivor cannot retain the lock. Pin the new order and the write-back fd closures in the workflow tests, and update the fair-lock design docs and the workflow size baseline.

Fixes #15262

Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
@qwen-code-dev-bot

Copy link
Copy Markdown
Collaborator Author

E2E test report

What changed

The self-hosted SDK Java test lane now seeds a job-local Maven repository from the host cache, runs its SDK and Managed Agent Maven suites from that copy without the per-host lock, and takes the ticketed host lock only for the bounded cache write-back. The write-back's children close the inherited lock descriptor so a cancelled job's survivor cannot retain the lock.

Why

A fast-tier job can fail before reporting a test result when it waits out its 90-minute host-lock budget behind a long suite hold. The MariaDB lane was already scoped to its write-back; the test lane was the remaining long hold, so a roughly 90-minute full-tier Java 21 leg on the same host could starve the fast tier — the failure shape reported in #15262.

E2E test plan

  1. On a self-hosted pool, run a touched PR's fast tier so the self-hosted Java 21 leg executes.
  2. Read the self-hosted SDK Java test step's log. Expected order: job-local repository seeded, both Maven suites run, then the lock wait line and the cache write-back. There must be no Waiting for the sdk-java host lock line during the suites.
  3. Run a second fast-tier job on the same host while the first is still in Maven. Expected: the second job starts Maven without waiting for the first suite; it may wait only for the short write-back.
  4. Confirm the fast tier reports test results instead of the pre-Maven lock timeout, and that the write-back leaves no held lock after the step.

The real pool run is left to CI; this checkout has no Java/Maven toolchain and no self-hosted pool.

Local evidence

  • Static guards pin that both lock sites seed a job-local repository and run every Maven command before the acquisition, and that the write-back's mkdir and bounded find/cp close fd 9.
  • The dynamic guard runs the test lane's suite slice with a stubbed mvn and asserts the host lock stays free; the surviving-child test now covers both lock sites.
  • Mutation probes, each reverted and re-run green:
Probe Mutation Witness that failed
A Re-acquire the host lock before the seed/Maven suite keeps the host lock out of every locked child
B Drop 9>&- from the write-back mkdir keeps the host lock out of every locked child
C Drop the job-local Maven repository from the test lane keeps setup-java Maven files job-local in the test job
D Drop 9>&- from the write-back find/cp releases the host lock when a write-back child survives the step

Verification

  • npx vitest run scripts/tests/sdk-java-workflow.test.js — 116 passed
  • npx vitest run scripts/tests/sdk-java-workflow.test.js scripts/tests/workflow-size.test.js scripts/tests/sdk-java-lock.test.js — 368 passed, 1 skipped
  • npx vitest run scripts/tests/event-transport-broker-ci.test.js scripts/tests/event-transport-multi-node-ci.test.js scripts/tests/hosted-process-ci.test.js scripts/tests/main-ci-failure-issue-workflow.test.js scripts/tests/managed-agent-gate-workflow.test.js scripts/tests/spotbugs-gate.test.js scripts/tests/workflow-size.test.js — 469 passed, 1 skipped (7 files)
  • npm run build — passed; the first attempt hit EACCES on the read-only corepack cache, and the rerun with a writable COREPACK_HOME passed
  • npm run typecheck — passed
  • npm run lint — passed
  • npm run check-i18n — passed
  • npx prettier --check .github/workflows/sdk-java.yml docs/design/2026-10-20-sdk-java-fair-lock-queue.md docs/design/2026-10-20-sdk-java-fair-lock-queue.zh-CN.md scripts/tests/sdk-java-workflow.test.js — passed
  • Mutation probes A–D — each failed the intended witness, then restored and the suite returned green
  • Real self-hosted pool — not run locally; CI owns the post-merge pool verification
中文说明

E2E 测试报告

改动内容

自托管 SDK Java 测试通道现在先从宿主机缓存播种一个 job 本地 Maven 仓库,在不持每宿主机锁的情况下用该副本运行 SDK 和 Managed Agent 两套 Maven 套件,最后只为有界缓存写回获取排队锁。写回的子进程会关闭继承来的锁描述符,因此被取消 job 的幸存进程无法继续持锁。

原因

当 fast 档 job 排在长时间套件持锁之后、等满 90 分钟宿主机锁预算时,它会在报告任何测试结果之前失败。MariaDB 通道已经把锁收窄到写回;测试通道是剩下的长时间持锁点,因此同一宿主机上一条约 90 分钟的 full 档 Java 21 leg 可能饿死 fast 档——这正是 #15262 报告的失败形态。

E2E 测试计划

  1. 在自托管 pool 上运行一个被改动 PR 的 fast 档,使自托管 Java 21 leg 执行。
  2. 阅读自托管 SDK Java 测试步骤的日志。预期顺序:先播种 job 本地仓库,再运行两套 Maven 套件,然后才出现取锁日志和缓存写回。套件运行期间不得出现 Waiting for the sdk-java host lock 日志行。
  3. 在第一个 job 仍在运行 Maven 时,在同一宿主机上运行第二个 fast 档 job。预期:第二个 job 无需等待第一个套件就开始 Maven;它最多只等待短暂的写回。
  4. 确认 fast 档报告测试结果,而不是在 Maven 之前发生锁超时;并确认写回结束后没有残留的持锁。

真实 pool 运行留给 CI;本 checkout 没有 Java/Maven 工具链,也没有自托管 pool。

本地证据

  • 静态 guard 固定两处取锁点都先播种 job 本地仓库、所有 Maven 命令都在取锁之前,并且写回的 mkdir 和有界 find/cp 都关闭 fd 9。
  • 动态 guard 用 stub 的 mvn 运行测试通道的套件片段,断言宿主机锁保持空闲;幸存子进程测试现在覆盖两处取锁点。
  • 变异探针,每个都已还原并重新跑绿:
探针 变异 失败的 witness
A 在播种/Maven 套件之前重新获取宿主机锁 keeps the host lock out of every locked child
B 去掉写回 mkdir 的 9>&- keeps the host lock out of every locked child
C 去掉测试通道的 job 本地 Maven 仓库 keeps setup-java Maven files job-local in the test job
D 去掉写回 find/cp 的 9>&- releases the host lock when a write-back child survives the step

验证

  • npx vitest run scripts/tests/sdk-java-workflow.test.js — 116 通过
  • npx vitest run scripts/tests/sdk-java-workflow.test.js scripts/tests/workflow-size.test.js scripts/tests/sdk-java-lock.test.js — 368 通过,1 跳过
  • npx vitest run scripts/tests/event-transport-broker-ci.test.js scripts/tests/event-transport-multi-node-ci.test.js scripts/tests/hosted-process-ci.test.js scripts/tests/main-ci-failure-issue-workflow.test.js scripts/tests/managed-agent-gate-workflow.test.js scripts/tests/spotbugs-gate.test.js scripts/tests/workflow-size.test.js — 469 通过,1 跳过(7 个文件)
  • npm run build — 通过;首次因 corepack 缓存只读报 EACCES,改用可写的 COREPACK_HOME 重跑后通过
  • npm run typecheck — 通过
  • npm run lint — 通过
  • npm run check-i18n — 通过
  • npx prettier --check .github/workflows/sdk-java.yml docs/design/2026-10-20-sdk-java-fair-lock-queue.md docs/design/2026-10-20-sdk-java-fair-lock-queue.zh-CN.md scripts/tests/sdk-java-workflow.test.js — 通过
  • 变异探针 A–D — 每个都使目标 witness 失败,还原后套件恢复绿色
  • 真实自托管 pool — 未在本地运行;合并后的 pool 验证由 CI 负责
--- 🧠 Handled by **Qwen Code** · model/模型 `deepseek-v4.1-flash-1017` · CLI `0.26.0`

@qwen-code-review-bot

qwen-code-review-bot commented Oct 22, 2026 •

Copy link
Copy Markdown
Collaborator

✅ Qwen Triage finished — view run. See the stage comments in this thread for the result.

✅ Qwen Triage 已完成 —— 查看运行。结果见本线程中的各阶段评论。

@github-actions github-actions Bot added the review/self-reported The linked issue was opened by the PR author (self-reported) label Oct 22, 2026
@qwen-code-review-bot

Copy link
Copy Markdown
Collaborator

Thanks for the PR!

Template looks good ✓ — every required heading is filled in, both design-doc languages are linked, and the two versions carry the same decisions and acceptance criteria.

Problem: observed, not theoretical. #15262 is an automated main-CI failure report with a run (39780882803), a commit, and named jobs: SDK Java (fast) / ubuntu-latest / Java 21 failed in Run Java SDK tests (self-hosted), and the fast tier failed in Report the tier result — i.e. it died before reporting a single test result. That is the starvation shape the description claims.

Direction: aligned. The design doc already commits to "hold scope: the write-back, not the suite" and landed it on the MariaDB lane with run-id evidence; this extends that same decision to the second lock site instead of inventing a second mechanism. Worth saying plainly, because the sibling PRs make the cheaper option look tempting: raising the wait budget (#15232, #15257) is a smaller diff, but the design doc rejects exactly that — "a wait budget bounds one attempt, not the wait… no constant budget bounds the queue." So the larger diff here is the correct one, not scope creep.

Size: not applicable — no core paths. Production changes are ~83 lines (the workflow step plus one baseline number); the remainder is the guard test and two design docs.

Approach: the scope feels right and it reuses what's already there. Two things I'd raise before this merges.

First, and this is the one I can't settle from the diff: there are seven open autofix PRs from the same bot editing this same region of sdk-java.yml — #15229, #15232, #15233, #15254, #15257, #15275 and this one. Three of them (#15254, #15275, #15274) propose the same test-lane lock scope. They cannot all land, and each records a different .size-baseline byte count for the same file, so whichever merges second trips the size ratchet until it is rebased. The body names only #15254 as overlapping. Picking the survivor is a maintainer call, not something this gate should choose.

Second, a tradeoff the body doesn't name: unlocking the suite means two Maven suites can now run concurrently on one ECS host, where the test job's suite was previously serialized. The cold-host download cost is mentioned; peak resource contention is not. The MariaDB lane already runs its multi-hour verify unlocked and the design doc lists reducing occupancy as an explicit non-goal, so I don't think this blocks anything — it's the thing to watch on the first pool runs.

Risk: no elevated risk signals — none of the revert-correlated paths are touched (only .github/workflows/, docs/design/, scripts/tests/).

The duplicate family doesn't block the code review, so I'm continuing to Stage 2 — but it does need a human decision before merge. 🔍

中文说明

感谢贡献!

模板完整 ✓ —— 所有必需标题都已填写,设计文档中英文版本均已链接,两个版本的决策与验收标准一致。

问题: 已观测到,非理论性。#15262 是自动生成的 main CI 失败报告,带有 run(39780882803)、commit 和具体 job 名:SDK Java (fast) / ubuntu-latest / Java 21 在 Run Java SDK tests (self-hosted) 步骤失败,fast 档在 Report the tier result 失败——即在报告任何测试结果之前就失败了。这正是描述中所说的饿死形态。

方向: 对齐。设计文档已经确立"持锁范围:写回,而非套件",并已在 MariaDB 通道落地且附有 run-id 证据;本 PR 是把同一决策扩展到第二处取锁点,而不是另造一套机制。需要明确说明的是(因为兄弟 PR 让更便宜的方案看起来很诱人):提高等待预算(#15232、#15257)diff 更小,但设计文档恰恰否定了这种做法——"等待预算只约束一次尝试,而不约束整个等待……没有任何常量预算能约束队列"。所以这里较大的 diff 是正确的,不是范围蔓延。

规模: 不适用——未触及核心路径。生产代码改动约 83 行(workflow 步骤加一个 baseline 数字);其余是 guard 测试和两份设计文档。

方案: 范围合理,且复用了已有实现。合并前我想提两点。

第一点,也是我从 diff 无法判断的:同一 bot 有 七 个开放 autofix PR 在改 sdk-java.yml 的同一区域——#15229、#15232、#15233、#15254、#15257、#15275 和本 PR。其中三个(#15254、#15275、#15274)提出相同的测试通道锁范围。它们不可能全部合并,且每个都为同一文件记录了不同的 .size-baseline 字节数,因此第二个合并的 PR 会在 rebase 之前触发尺寸棘轮。PR 正文只提到 #15254 重叠。选出保留哪一个属于维护者决定,不应由本 gate 选择。

第二点,正文未提及的取舍:套件移出锁后,同一台 ECS 宿主机上现在可能并发运行两套 Maven 套件,而此前 test job 的套件是串行的。正文提到了冷宿主机的下载成本,但没有提到峰值资源争用。MariaDB 通道已经在不持锁的情况下运行数小时的 verify,且设计文档明确把降低占用列为非目标,所以我认为这不构成阻塞——但这是合并后首批 pool 运行需要观察的点。

风险: 无升级风险信号——未触及任何与 revert 相关的路径(仅 .github/workflows/、docs/design/、scripts/tests/)。

重复 PR 家族不阻塞代码审查,因此我继续进入 Stage 2——但合并前确实需要人工决定。🔍

— Qwen Code · qwen3.8-max-2026-09-02

Reviewed at 80598d6aa445293bbf624a9cc79f6ec5270f5178 · re-run with @qwen-code /triage

@qwen-code-review-bot

qwen-code-review-bot commented Oct 22, 2026 •

Copy link
Copy Markdown
Collaborator

Code review

No critical blockers. I read the workflow step, the guard test, and the MariaDB lane it mirrors, and verified the parts that can be checked statically.

What I checked and cleared:

  • Every mvn call resolves job-local. This was the main way the change could be wrong — one Maven invocation left on host-shared ~/.m2 would reintroduce the exact race the lock existed for. -Dmaven.repo.local is added to the step-level MAVEN_ARGS, so all four invocations in the step (the SDK suite, the install, and both managed-agent branches) inherit it. None touches the shared cache.
  • It reuses the landed pattern rather than inventing one. Same cp -aln seed, same -links 1 delta filter, same two com/alibaba prune excludes, same timeout 300, same 9>&- on mkdir and on the find/cp, same warning arms as the MariaDB lane. That is the strongest correctness signal here: this shape has already survived real pool runs.
  • Best-effort under bash -e. The step runs with the Actions default bash -e, so an unguarded failure would abort it. The seed carries || true and the write-back carries || write_back=$? with warning arms, so a cache failure cannot redden the gate after mvn has already passed. Conversely a genuine test failure still aborts before the write-back — correct.
  • fd 9 closures on both write-back children, so a cancelled job's survivor cannot retain the host lock. The guard test now loops both lock sites instead of only the MariaDB one.
  • .size-baseline is exactly right. I fetched sdk-java.yml raw at the head commit: 119874 bytes, matching the recorded number precisely. Main is 118095 actual against 114901 recorded, which sits inside the 4096-byte growth allowance — so this is the correct same-PR bump, not a stale one.
  • The timeout math still closes. 5400 s wait + 925 s suite + 300 s write-back + ~11 min setup ≈ 121 min, under the 150 min ceiling. The assertion was widened to seconds + 925 + writeBackBound + 11 * 60 rather than left stale against the old hold definition.
  • The renamed guard is not a dropped invariant. "serializes latency-sensitive tests" sounded like it encoded a real constraint that unlocking the suite would violate, so I went looking. latency-sensitive appears nowhere else in the repo — not in the workflow, not in the lock helper, not in the design doc, which frames the lock purely as host-shared-state protection under a heading "Hold scope: the write-back, not the suite". Stale name, correctly renamed.
  • The other eight tests that parse this workflow. AGENTS.md asks that every scripts/tests/*.test.js parsing a changed workflow be updated. Nine parse sdk-java.yml; this PR updates one. I checked the rest statically: spotbugs-gate.test.js's global invariant only forbids spotbugs.skip in env blobs, so the added -Dmaven.repo.local doesn't trip it; sdk-java-lock.test.js exercises the ticket-queue helper, not the workflow's hold scope; the remainder key off job and step names this PR doesn't touch. Nothing looks broken, and CI will confirm.

Non-blocking observations:

  • The runs the Maven suite before taking the host lock guard slices the run block between two literal strings — the cp -aln seed and the comment # Only the write-back mutates host-shared state. Coupling an assertion to prose means rewording that comment silently changes what the slice covers. The load-bearing part (the lock is still free after the suite runs) does still get asserted.
  • On a cold host nothing is seeded, so everything Maven downloads is -links 1 and the whole tree becomes write-back candidate, bounded by the 300 s timeout. If it times out the cache stays cold and the next job repeats the work. The warning arm makes it visible and the body acknowledges the download cost — worth watching on the first pool runs, not worth blocking.

CI test evidence

The PR's own CI, read through the API at the reviewed commit. No check is red. 55 check-runs exist on this SHA; the substantive ones had not started yet at review time, including SDK Java (fast) / ubuntu-latest / Java 21 — the exact job that failed in #15262. I did not wait on them, so this review carries no test-pass evidence: what landed green is only the cheap pre-flight jobs. Treat the results below as incomplete, not as a pass.

Final CI results for 80598d6 (auto-updated by the triage finalize job after CI completed):

Check Conclusion
Classify PR ✅ success
Integration Tests (no-AK, No Sandbox) ✅ success
Lint & Static (ubuntu-latest, Node 22.x) ✅ success
Managed Agent fast tier ✅ success
SDK Java (fast) / Flyway migration version uniqueness ✅ success
SDK Java (fast) / Real daemon E2E / Java 11 ✅ success
SDK Java (fast) / ubuntu-latest / Java 21 ✅ success
Select the Managed Agent tier ✅ success
Test (ubuntu-latest, Node 22.x) ✅ success
web-shell E2E Smoke (ubuntu-latest, Node 22.x) ✅ success

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。

Not verified: real self-hosted pool behaviour. The guard tests are static shape pins plus a dynamic guard that runs the suite slice against a stubbed mvn — they prove the script text has the right ordering and that the lock stays free under a stub, not that a real host stops starving. No lane available here can drive two concurrent jobs on an actual ECS host, so that part genuinely waits for the post-merge pool runs, as the body says.

Sandboxed verification would settle part of this: @qwen-code /verify — specifically, that the new guards are load-bearing rather than decorative. The author reports mutation probes A–D each reddening an intended witness, but that is the author's claim, not independently re-run; an A/B against the base build would confirm the suite actually fails when the lock is re-acquired before the suite or when a 9>&- closure is dropped. It cannot settle the pool-contention question, which needs real hosts.

中文说明

代码审查

无阻塞性问题。我阅读了 workflow 步骤、guard 测试以及它所镜像的 MariaDB 通道,并对可静态验证的部分做了核实。

已核查并排除的问题:

  • 每个 mvn 调用都使用 job 本地仓库。 这是本改动最可能出错的地方——只要有一个 Maven 调用仍解析宿主机共享的 ~/.m2,就会重新引入锁本要防止的竞态。-Dmaven.repo.local 加在步骤级 MAVEN_ARGS 上,因此该步骤中全部四个调用(SDK 套件、install、以及两个 managed-agent 分支)都会继承它。没有一个触及共享缓存。
  • 复用已落地的模式,而非另造一套。 与 MariaDB 通道相同的 cp -aln 播种、相同的 -links 1 增量过滤、相同的两个 com/alibaba prune 排除、相同的 timeout 300、mkdir 与 find/cp 上相同的 9>&-、相同的 warning 分支。这是此处最强的正确性信号:这一形态已经历过真实 pool 运行。
  • 在 bash -e 下是尽力而为。 该步骤使用 Actions 默认的 bash -e,未加保护的失败会中止步骤。播种带 || true,写回带 || write_back=$? 与 warning 分支,因此在 mvn 已通过之后,缓存失败不会让 gate 变红。反过来,真实的测试失败仍会在写回前中止——这是正确的。
  • 两个写回子进程都关闭了 fd 9,因此被取消 job 的幸存进程无法继续持有宿主机锁。guard 测试现在遍历两处取锁点,而不只是 MariaDB 那一处。
  • .size-baseline 完全正确。 我在 head commit 上原样拉取了 sdk-java.yml:119874 字节,与记录数字精确一致。main 上实际 118095 对记录 114901,落在 4096 字节增长余量内——所以这是正确的同 PR 更新,而非过期值。
  • 超时计算仍然成立。 5400 秒等待 + 925 秒套件 + 300 秒写回 + 约 11 分钟安装 ≈ 121 分钟,低于 150 分钟上限。断言被放宽为 seconds + 925 + writeBackBound + 11 * 60,而不是按旧的持锁定义留下过期值。
  • 重命名的 guard 不是丢失的不变量。 "serializes latency-sensitive tests" 听起来像是编码了某个真实约束,而把套件移出锁会违反它,所以我去查了。latency-sensitive 在仓库中别处均未出现——workflow、锁辅助脚本、设计文档都没有;设计文档把锁纯粹定义为宿主机共享状态保护,标题即"持锁范围:写回,而非套件"。名称过期,重命名正确。
  • 其余八个解析该 workflow 的测试。 AGENTS.md 要求更新每个解析被改 workflow 的 scripts/tests/*.test.js。共有九个解析 sdk-java.yml,本 PR 更新了一个。我静态检查了其余的:spotbugs-gate.test.js 的全局不变量只禁止 env 中出现 spotbugs.skip,新增的 -Dmaven.repo.local 不会触发;sdk-java-lock.test.js 测的是票据队列辅助脚本,而非 workflow 的持锁范围;其余依赖本 PR 未触及的 job 与 step 名。没有发现被破坏的地方,CI 会最终确认。

非阻塞观察:

  • runs the Maven suite before taking the host lock 这个 guard 用两个字面字符串切割 run 块——cp -aln 播种与注释 # Only the write-back mutates host-shared state。把断言耦合到散文上意味着改写该注释会悄悄改变切片覆盖的范围。承重的部分(套件运行后锁仍空闲)确实仍被断言。
  • 冷宿主机上没有播种,因此 Maven 下载的一切都是 -links 1,整棵树都成为写回候选,受 300 秒超时限制。若超时,缓存仍为冷态,下一个 job 重复该工作。warning 分支使其可见,正文也承认了下载成本——首批 pool 运行值得观察,但不构成阻塞。

CI 测试证据

在被审查的 commit 上通过 API 读取本 PR 自身的 CI。没有 check 变红。 该 SHA 上共有 55 个 check-run;审查时实质性的那些尚未开始,包括 SDK Java (fast) / ubuntu-latest / Java 21——正是 #15262 中失败的那个 job。我没有等待它们,因此本次审查不携带测试通过的证据:变绿的只有廉价的前置 job。下表应视为不完整,而不是通过。

未验证:真实自托管 pool 行为。guard 测试是静态形态 pin,加上一个用 stub 的 mvn 运行套件切片的动态 guard——它们证明脚本文本顺序正确、且在 stub 下锁保持空闲,而不证明真实宿主机不再饿死。此处没有任何通道能在真实 ECS 宿主机上驱动两个并发 job,因此这部分确实要等合并后的 pool 运行,正如正文所说。

沙箱验证可以解决其中一部分:@qwen-code /verify——具体说是验证新 guard 是承重的而非装饰性的。作者报告变异探针 A–D 各自让目标 witness 变红,但那是作者的声明,未经独立重跑;与 base 构建做 A/B 可以确认:当锁在套件之前被重新获取、或某个 9>&- 关闭被去掉时,套件确实会失败。它无法解决 pool 争用问题,那需要真实宿主机。

— Qwen Code · qwen3.8-max-2026-09-02

Reviewed at 80598d6aa445293bbf624a9cc79f6ec5270f5178 · re-run with @qwen-code /triage

@qwen-code-review-bot

qwen-code-review-bot commented Oct 22, 2026 •

Copy link
Copy Markdown
Collaborator

Confidence: 3/5 — the diff itself is clean and I would merge it on its own merits; what I can't decide is which of seven overlapping autofix PRs should be the one that lands.

Going into this I expected to talk myself out of it, and the code didn't give me a reason to. My independent proposal before reading the diff was: point Maven at a job-local repository so concurrent jobs share no mutable state, then lock only a bounded write-back. That is precisely what this does, and it does it by copying a pattern already living in the same file rather than designing a new one. The five things most likely to be wrong — an mvn call left resolving from the shared cache, an unguarded failure aborting the step under bash -e, a write-back child inheriting the lock fd, a stale size baseline, a timeout ceiling that no longer adds up — I checked each one and each is handled. I went hunting specifically for a dropped invariant behind the renamed "latency-sensitive" guard, because that name sounded load-bearing, and it isn't: the string exists nowhere else in the repo and the design doc scopes the lock to host-shared state.

So this is not a "the code is bad" defer. Two things stop me approving.

The duplicate family is the real issue. Seven open autofix PRs from the same bot edit this same region of sdk-java.yml: #15229, #15232, #15233, #15254, #15257, #15275 and this one. Three of them propose the same test-lane lock scope. The body names one sibling; there are six. They are mutually exclusive in practice — each records a different .size-baseline byte count for the same file, so the second to merge fails the size ratchet until rebased, and they will textually conflict on the lines they all rewrite. Approving this one is an arbitrary choice among three near-identical candidates, and leaving a deferred-approval instruction here would be worse: the finalize job would approve whichever sibling happens to go green first, with no human having chosen. That is exactly the failure mode such an instruction shouldn't serve, so I'm not leaving one.

CI hadn't started. The substantive checks were all still queued at review time, including SDK Java (fast) / ubuntu-latest / Java 21 — the exact job that failed in #15262. Nothing is red, and nothing is proven either. The unit and workflow suites that would pin this change had not run.

On the reflection questions: yes, the problem is real and observed (#15262 carries a run ID and a named failing step, not a hypothesis). Yes, six months from now I'd thank the author — the write-back pattern is now consistent across both lock sites instead of one lane being special. And the one reservation I hold about the code itself is that unlocking the suite raises peak concurrency on a shared host, a tradeoff the body doesn't name; I don't think it's blocking, since the MariaDB lane already runs unlocked for hours and the design doc calls reduced occupancy a non-goal, but it's the first thing I'd look at if the pool gets noisier after this lands.

⏸️ Deferring rather than approving — this needs a human call. Specifically: which of #15274 / #15275 / #15254 should land for the test lane, and who closes the other six. Nothing in the diff needs changing to answer that.

I could not resolve an accountable owner to @mention, and I'd rather say so than guess a login: no maintainer handle is configured, the PR carries no labels and the owner map is label-driven with every changed path outside packages/core/, and no human has reviewed yet. A maintainer passing by — please take the merge-order decision above, and re-run @qwen-code /triage once the family is pruned if you want a fresh verdict on the survivor.

中文说明

Confidence: 3/5 —— diff 本身是干净的,单看它我会合并;我无法判断的是七个相互重叠的 autofix PR 中应该由哪一个落地。

开始时我预期会说服自己否掉它,但代码没有给我理由。在读 diff 之前,我的独立方案是:让 Maven 指向 job 本地仓库,使并发 job 之间不共享可变状态,然后只为一次有界写回取锁。这正是本 PR 所做的,而且它是通过复制同一文件中已有的模式实现的,而非另做设计。最可能出错的五点——某个 mvn 调用仍从共享缓存解析、某个未加保护的失败在 bash -e 下中止步骤、某个写回子进程继承锁 fd、过期的 size baseline、不再成立的超时上限——我逐一核查,每一点都得到了处理。我特意去查找重命名的 "latency-sensitive" guard 背后是否丢失了某个不变量,因为那个名字听起来是承重的,而它不是:该字符串在仓库别处并不存在,设计文档把锁的范围限定为宿主机共享状态。

所以这不是"代码有问题"式的 defer。有两点让我不予批准。

重复 PR 家族是真正的问题。 同一 bot 有七个开放的 autofix PR 在改 sdk-java.yml 的同一区域:#15229、#15232、#15233、#15254、#15257、#15275 和本 PR。其中三个提出相同的测试通道锁范围。正文只提到一个兄弟 PR;实际有六个。它们在实践上互斥——每个都为同一文件记录了不同的 .size-baseline 字节数,因此第二个合并的会在 rebase 之前触发尺寸棘轮失败,而且它们会在彼此都要重写的行上产生文本冲突。批准本 PR 等于在三个几乎相同的候选中做任意选择,而在此处留下延迟批准指令会更糟:finalize job 会批准恰好先变绿的那个兄弟 PR,而没有人工做过选择。这正是此类指令不应服务的失败模式,所以我没有留下它。

CI 尚未开始。 审查时实质性 check 全部仍在排队,包括 SDK Java (fast) / ubuntu-latest / Java 21——正是 #15262 中失败的那个 job。没有变红的,但也没有任何被证明的。本应 pin 住该改动的单元与 workflow 套件尚未运行。

关于反思问题:是的,问题真实且已观测(#15262 带有 run ID 和明确的失败步骤名,而非假设)。是的,六个月后我会感谢作者——写回模式现在在两处取锁点上保持一致,而不是让某一条通道成为特例。我对代码本身保留的一点意见是:把套件移出锁会提高共享宿主机上的峰值并发,而正文没有提及这一取舍;我不认为它构成阻塞,因为 MariaDB 通道已经不持锁运行数小时,且设计文档把降低占用列为非目标,但如果合并后 pool 变得更嘈杂,这是我会首先查看的地方。

⏸️ 选择 defer 而非批准——这需要人工决定。 具体是:测试通道应由 #15274 / #15275 / #15254 中哪一个落地,以及由谁关闭其余六个。回答这个问题不需要修改 diff 中的任何内容。

我无法解析出应当 @ 的负责人,与其猜测一个 login,不如直接说明:未配置维护者 handle,PR 不带任何 label 而 owner 映射由 label 驱动、且所有改动路径都在 packages/core/ 之外,同时尚无人工 review 记录。路过的维护者——请接管上述合并顺序决定;若在精简该 PR 家族后希望对保留者获得新结论,请重新运行 @qwen-code /triage。

— Qwen Code · qwen3.8-max-2026-09-02

Reviewed at 80598d6aa445293bbf624a9cc79f6ec5270f5178 · re-run with @qwen-code /triage

@qwen-code-review-bot qwen-code-review-bot left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[Critical] Blocking finding(s) follow.

Reviewed. Suggestions are inline.

Unresolved, please confirm:

Not explored to full depth (tool budget reached): "agent 6a": none — node_modules is absent in this worktree, so I could not execute npx vitest run scripts/tests/sdk-java-workflow.test.js scripts/tests/workflow-size.tes…; "agent 4": npx vitest run scripts/tests/sdk-java-workflow.test.js scripts/tests/workflow-size.test.js — the review worktree has no node_modules , so the rewritten pins ….

Not linted (tool limitation, not a blocker): .github/workflows/sdk-java.yml — actionlint embedded-shell source mapping is not yet supported.

中文说明

已审查。 建议见行内评论。

未决,请确认:共 1 条(原文未翻译,列表见上方英文部分)。

未探索到全部深度(达到工具调用预算):"agent 6a":none — node_modules is absent in this worktree, so I could not execute npx vitest run scripts/tests/sdk-java-workflow.test.js scripts/tests/workflow-size.tes…;"agent 4":npx vitest run scripts/tests/sdk-java-workflow.test.js scripts/tests/workflow-size.test.js — the review worktree has no node_modules , so the rewritten pins …。

未检查(工具限制,非阻断):.github/workflows/sdk-java.yml——actionlint 对 workflow 内嵌 shell 的源映射尚未支持。

— qwen3.8-max via Qwen Code /review (v0.26.0)

Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread scripts/tests/sdk-java-workflow.test.js Resolved
Comment thread scripts/tests/sdk-java-workflow.test.js Resolved
@qwen-code-dev-bot

Copy link
Copy Markdown
Collaborator Author

🕐 Review received — an automatic review of the current head is still running, so this round is held until it lands (a push now would cancel it and discard its work, #8888). Your feedback stays queued for the next eligible round.

中文说明

🕐 已收到评审 —— 当前 head 上仍有一轮自动 review 在运行,本轮暂缓(现在推送会取消该 review 并丢弃其工作,#8888)。反馈保持排队,等待下一次可运行的轮次处理。

@qwen-code-dev-bot

qwen-code-dev-bot commented Oct 22, 2026 •

Copy link
Copy Markdown
Collaborator Author

⚠️ AutoFix round 10 ended without publishing a report — view run.

中文说明

⚠️ AutoFix 第 10 轮结束但未发布报告 —— 查看运行。

@qwen-code-dev-bot

Copy link
Copy Markdown
Collaborator Author

🤖 Addressed the latest review feedback (round 1/10). What changed, and what I pushed back on: · 已处理最新评审反馈(第 1/10 轮)。改动内容与我反驳保留之处如下:

Autofix review round — PR #15274 (issue #15262)

All nine inline suggestions from the automated review were verified against the current code, fixed, and are resolved in code. The review-body merge-order note is a maintainer decision and remains open; it has no inline thread, so there is nothing to resolve on it. No conflicts were present: the round ran with --conflict false and performed no merge.

Findings and dispositions

Finding Disposition Change
[rc:4354291537] R1-1 — seed failures are silent Fixed All four cp -aln seed sites now end with `2>/dev/null
[rc:4354291552] R1-2 — retiring #10552's per-host suite serialization is unrecorded Fixed The English and Chinese design docs now record the decision in the hold-scope section, amend the sharding bullet so it no longer rejects concurrent suites as an untested assumption, and add a residual-failure-mode entry naming the MariaDB lane's unlocked suite since #14903 as the evidence.
[rc:4354291563] R1-3 — a failed suite skips the write-back Fixed The test job's suite now runs under set +e with suite_status=$?, the acquisition and write-back still run, and the step ends with exit "${suite_status}". Both write-backs filter *.lastUpdated and resolver-status.properties. A new Linux test runs a failing mvn, plants an nlink == 1 file, and asserts the file is published while the step still exits 1.
[rc:4354291576] R1-4 — a lock wait timeout fails a green suite Fixed Both the helper arm and the bare-flock fallback are now if ! ...; then guards that warn and skip the best-effort write-back. The test job exits ${suite_status}; the MariaDB step exits 0 because a failed suite cannot reach it under bash -e. Tests pin both branches and the warning text.
[rc:4354291584] R1-5 — the two write-back copies are pinned unequally Fixed A shared m2WriteBack literal and warning-arm pins now cover both copies, plus the cd "${RUNNER_TEMP}/m2-repo" position. Deleting the cd, the prune clause, -links 1, or either warning arm turns keeps the host lock out of every locked child red for the affected copy.
[rc:4354291589] R1-6 — -n non-portability is documented as a guarantee Fixed via the comment-softening alternative -n stays because older coreutils versions cannot be ruled out, and the comments in both copies now say it skips rather than replaces an existing host file and is non-portable across coreutils versions.
[rc:4354291597] R1-7 — the "roughly 90 minutes" hold figure is wrong Fixed The workflow comment and both design docs now state the 925 s suite hold, the measured 65m47s/86m50s queue positions, and that shortening every hold to a ≤300 s write-back shortens the queue; the single 90-minute-holder claim is removed.
[rc:4354291608] R1-8 — the ordering loop uses first-occurrence offsets Fixed The loop now checks seed presence, then asserts that the region after the acquisition contains no mvn line at all. Appending a duplicate mvn below the lock turns keeps the host lock out of every locked child red.
[rc:4354291614] R1-9 — the jobLocal guard asserts a count, not a site Fixed The test job branch now resolves Run Java SDK tests (self-hosted) and asserts its MAVEN_ARGS, pins the full seed string, and creates a real ${HOME}/.m2/repository marker file in the behavioral fixture and asserts the seeded copy has the same inode. Moving the suffix to the hosted step and retargeting the seed both turn tests red.

Open question for maintainers

[rv:5621767429] The PR-level triage defer asks for a human merge-order decision across the overlapping autofix family: #15233, #15254, #15257 and #15275 remain open. This is a product/merge-order decision, not a code defect, and it is not ours to settle. It is a review-body item with no inline thread, so it is left exactly as raised and is recorded here for the maintainer.

Changed files

  • .github/workflows/sdk-java.yml
  • .github/workflows/.size-baseline
  • docs/design/2026-10-20-sdk-java-fair-lock-queue.md
  • docs/design/2026-10-20-sdk-java-fair-lock-queue.zh-CN.md
  • scripts/tests/sdk-java-workflow.test.js

Verification

  • npm run build — passed.
  • npm run typecheck — passed.
  • npm run lint — passed.
  • COREPACK_HOME=/tmp/corepack-home npx vitest run --config ./scripts/tests/vitest.config.ts scripts/tests/sdk-java-workflow.test.js scripts/tests/sdk-java-lock.test.js scripts/tests/workflow-size.test.js — 370 passed, 1 skipped (3 files).
  • npx prettier --check .github/workflows/sdk-java.yml docs/design/2026-10-20-sdk-java-fair-lock-queue.md docs/design/2026-10-20-sdk-java-fair-lock-queue.zh-CN.md scripts/tests/sdk-java-workflow.test.js — passed. .size-baseline is extensionless and has no Prettier parser; it was not reformatted.

Mutation probes run against the same workflow content that was committed; each mutated/negated branch turned the named test red:

  • Seed warning arm removed — RED (keeps the host lock out of every locked child, plus the all-seed test).
  • Test-job suite status capture removed (set +e → set -e) — RED (publishes a failed suite's downloads and exits with its status).
  • Final suite status swallowed (exit "${suite_status}" → exit 0) — RED (same test).
  • Test-job acquisition guard removed — RED (serializes the host-shared Maven cache write-back on each physical ECS host).
  • MariaDB acquisition guard removed — RED (holds the per-host sdk-java lock once around the cache write-back).
  • Test-job -links 1 dropped — RED (keeps the host lock out of every locked child).
  • MariaDB -links 1 dropped — RED (installs and verifies against one job-local Maven repository).
  • Duplicate mvn appended below the test-job lock — RED (keeps the host lock out of every locked child).
  • Job-local suffix moved to the hosted step — RED (keeps setup-java Maven files job-local in the test job).
  • Test-job seed destination retargeted to m2-cache — RED (runs the Maven suite from the seeded job-local repository before taking the host lock).
  • Test-job cd "${RUNNER_TEMP}/m2-repo" removed — RED (keeps the host lock out of every locked child).
  • Test-job prune clause dropped — RED (keeps the host lock out of every locked child).
  • Test-job write-back warning arms removed — RED (keeps the host lock out of every locked child).
中文说明

Autofix 审查轮次——PR #15274(issue #15262)

自动审查提出的九条行内建议都已对当前代码核实、修复,并在代码中解决。审查正文中的合并顺序说明属于维护者决定,保持未决;它没有行内线程,因此没有可解决的线程。本轮没有冲突:以 --conflict false 运行,未执行合并。

发现与处理

发现 处理 改动
[rc:4354291537] R1-1——种子失败被静默 已修复 四个 cp -aln 种子位点现在都以 `2>/dev/null
[rc:4354291552] R1-2——撤销 #10552 的每宿主机套件串行化没有记录 已修复 中英文设计文档现在在持锁范围一节记录该决定,修改分片条目使其不再把并发套件当作未经验证的假设否定,并新增一条有意接受的残余失败模式,引用 MariaDB 通道自 #14903 起已不持锁运行作为证据。
[rc:4354291563] R1-3——套件失败会跳过写回 已修复 test job 的套件现在在 set +e 下运行,用 suite_status=$? 记录状态,取锁与写回仍会执行,步骤最后以 exit "${suite_status}" 结束。两份写回都会过滤 *.lastUpdated 与 resolver-status.properties。新增的 Linux 测试让 mvn 失败、预置一个 nlink == 1 文件,断言文件仍被发布且步骤仍以 1 退出。
[rc:4354291576] R1-4——取锁等待超时会让全绿套件失败 已修复 helper 分支与裸 flock fallback 现在都是 if ! ...; then 守卫,超时时告警并跳过尽力而为的写回。test job 以 ${suite_status} 退出;MariaDB 步骤以 0 退出,因为在 bash -e 下失败的套件根本到不了这里。测试钉住两个分支与告警文本。
[rc:4354291584] R1-5——两份写回副本的测试钉住不对等 已修复 共享的 m2WriteBack 精确字面量与告警分支断言现在覆盖两份副本,并钉住 cd "${RUNNER_TEMP}/m2-repo" 的位置。删除 cd、prune 子句、-links 1 或任一告警分支,都会使受影响副本的 keeps the host lock out of every locked child 变红。
[rc:4354291589] R1-6——-n 的不可移植性被写成保证 采纳注释软化方案修复 保留 -n,因为无法排除更旧的 coreutils 版本;两份副本的注释现在说明它是"跳过"而非"替换"宿主机已有文件,并说明该选项跨 coreutils 版本不可移植。
[rc:4354291597] R1-7——"约 90 分钟"持锁数字错误 已修复 工作流注释与两份设计文档现在写明 925 秒的套件持锁、实测的 65分47秒/86分50秒排队位置,以及把每次持锁缩短到 ≤300 秒写回缩短的是队列;删除了单一 90 分钟持锁者的说法。
[rc:4354291608] R1-8——排序循环使用首次出现偏移 已修复 循环现在先检查种子存在,再断言取锁之后的区域完全不出现 mvn 行。在锁下追加重复的 mvn 会使 keeps the host lock out of every locked child 变红。
[rc:4354291614] R1-9——jobLocal 守卫只断言数量而非位置 已修复 test job 分支现在解析 Run Java SDK tests (self-hosted) 并断言其 MAVEN_ARGS,钉住完整种子字符串,并在行为测试的 fixture 中创建真实的 ${HOME}/.m2/repository 标记文件,断言播种副本具有相同 inode。把后缀移到托管步骤、或改掉种子目标,都会使测试变红。

维护者未决问题

[rv:5621767429] PR 级 triage defer 要求维护者在重叠的 autofix 家族之间做合并顺序决定:#15233、#15254、#15257、#15275 仍未合并。这是产品/合并顺序决定,不是代码缺陷,也不由本轮决定。它是审查正文条目,没有行内线程,因此按原样保留,并在此记录供维护者处理。

改动文件

  • .github/workflows/sdk-java.yml
  • .github/workflows/.size-baseline
  • docs/design/2026-10-20-sdk-java-fair-lock-queue.md
  • docs/design/2026-10-20-sdk-java-fair-lock-queue.zh-CN.md
  • scripts/tests/sdk-java-workflow.test.js

验证

  • npm run build——通过。
  • npm run typecheck——通过。
  • npm run lint——通过。
  • COREPACK_HOME=/tmp/corepack-home npx vitest run --config ./scripts/tests/vitest.config.ts scripts/tests/sdk-java-workflow.test.js scripts/tests/sdk-java-lock.test.js scripts/tests/workflow-size.test.js——370 通过,1 跳过(3 个文件)。
  • npx prettier --check .github/workflows/sdk-java.yml docs/design/2026-10-20-sdk-java-fair-lock-queue.md docs/design/2026-10-20-sdk-java-fair-lock-queue.zh-CN.md scripts/tests/sdk-java-workflow.test.js——通过。.size-baseline 没有扩展名,Prettier 没有对应解析器,未被重新格式化。

变异探针针对与提交内容相同的工作流运行;每个被删除或取反的分支都使对应测试变红:

  • 删除种子告警分支——变红(keeps the host lock out of every locked child,以及全种子测试)。
  • 删除 test job 的套件状态捕获(set +e → set -e)——变红(publishes a failed suite's downloads and exits with its status)。
  • 吞掉最终套件状态(exit "${suite_status}" → exit 0)——变红(同一测试)。
  • 删除 test job 的取锁守卫——变红(serializes the host-shared Maven cache write-back on each physical ECS host)。
  • 删除 MariaDB 的取锁守卫——变红(holds the per-host sdk-java lock once around the cache write-back)。
  • 删除 test job 的 -links 1——变红(keeps the host lock out of every locked child)。
  • 删除 MariaDB 的 -links 1——变红(installs and verifies against one job-local Maven repository)。
  • 在 test job 的锁下追加重复 mvn——变红(keeps the host lock out of every locked child)。
  • 把 job 本地后缀移到托管步骤——变红(keeps setup-java Maven files job-local in the test job)。
  • 把 test job 的种子目标改为 m2-cache——变红(runs the Maven suite from the seeded job-local repository before taking the host lock)。
  • 删除 test job 的 cd "${RUNNER_TEMP}/m2-repo"——变红(keeps the host lock out of every locked child)。
  • 删除 test job 的 prune 子句——变红(keeps the host lock out of every locked child)。
  • 删除 test job 的写回告警分支——变红(keeps the host lock out of every locked child)。

Base-conflict check · 基分支冲突检查: no conflict with main. · 与 main 无冲突。

🧵 Resolved all 9 selected review thread(s). · 已关闭全部选中的 9 条评审线程。

Re-review when you have a moment. After round 10 this bot stops and leaves the PR for a human. · 有空请复审;第 10 轮后本 bot 停止并将 PR 交给人工。


🧠 Handled by Qwen Code · model/模型 deepseek-v4.1-flash-1017 · CLI 0.26.0

@qwen-code-review-bot

Copy link
Copy Markdown
Collaborator

Qwen Code review did not complete successfully. The review pipeline failed before a review could be posted. A transient error is retried automatically; if you are seeing this, retry with @qwen-code /review. See workflow logs.

@qwen-code-dev-bot

Copy link
Copy Markdown
Collaborator Author

🤖 Reviewed the latest feedback — no changes needed. Why, point by point: · 已审阅最新反馈——无需改动。逐点说明原因如下:

No code changes were made this round: the new feedback contains no reviewable findings.

What arrived this round

  • Reviews: none.
  • Inline comments: none.
  • Issue-level comments: one, from the automated reviewer (qwen-code-review-bot), stating that its review pipeline did not complete and that a transient error is retried automatically, with a pointer to retry via @qwen-code /review. This is a pipeline-status notice, not a finding about the change, so there is nothing to implement or rebut.
  • Failed checks: none.
  • Still-red checks: none. All reported check runs are SUCCESS or SKIPPED.

Disposition

The feedback cannot be verified as a code defect, and no claim about current behavior was made, so no probe or fix applies. The review that failed will be retried by the review pipeline; once it posts actual findings they will be triaged on their merits like any other feedback. The branch is unchanged.

中文说明

本轮没有做任何代码改动:新的反馈中不包含任何可评审的问题项。

本轮收到的内容

  • Review: 无。
  • 行内评论: 无。
  • Issue 级评论: 一条,来自自动评审机器人(qwen-code-review-bot),说明其评审流水线未能完成,临时性错误会自动重试,并提示可通过 @qwen-code /review 重试。这是流水线状态通知,而不是针对本次改动的评审意见,因此没有需要实现或反驳的内容。
  • 失败的检查: 无。
  • 仍然失败的检查: 无。所有报告的检查运行均为 SUCCESS 或 SKIPPED。

处理结论

该反馈无法被验证为代码缺陷,也没有对当前行为提出任何断言,因此不适用复现探针或修复。失败的评审将由评审流水线自动重试;一旦它给出实际的评审意见,将像其他反馈一样按其实质内容进行分诊。分支保持不变。

Base-conflict check · 基分支冲突检查: no conflict with main. · 与 main 无冲突。

🧠 Handled by Qwen Code · model/模型 deepseek-v4.1-flash-1017 · CLI 0.26.0

@qwen-code-review-bot qwen-code-review-bot left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Partially reviewed — gaps disclosed.

Not explored to full depth (tool budget reached): "agent 6c": executing scripts/tests/sdk-java-workflow.test.js — the review worktree has no node_modules , so every assertion above was traced by hand against the parsed ….

Not reviewed: reverse audit — stopped before round 5 by the review time budget.

Not linted (tool limitation, not a blocker): .github/workflows/sdk-java.yml — actionlint embedded-shell source mapping is not yet supported.

Deferred under the convergence posture (round 2, not a blocker) — recorded, not requested in this round:

  • .github/workflows/sdk-java.yml:303 — [probe] mkdir -p and cd in the write-back are bare under the restored set -e , so a host-state failure reddens a step whose suite already passed
  • scripts/tests/sdk-java-workflow.test.js:242 — [probe] the executed ordering test's slice end-marker indexOf is never asserted > -1, so rewording one workflow comment silently turns its flock witness into dead code

[Critical] PR-level merge-order blocker carried from the previous round — still stands, and no code change in this diff closes it. Approval was deferred pending a human decision on which of the overlapping autofix PRs should land for the sdk-java test lane. Verified at the reviewed commit: #15233, #15254, #15257 and #15275 all remain OPEN (#15229 and #15232 have since closed), and the mechanism the deferral names still holds — #15254, #15257 and #15275 each edit .github/workflows/sdk-java.yml and each records a different .size-baseline byte count for it (119651, 118481 and 123176 against this PR's 121709), so whichever merges second fails the size ratchet until rebased, and #15233 rewrites the same scripts/tests/sdk-java-workflow.test.js this PR rewrites. No human has reviewed: the only review on the PR is qwen-code-review-bot's COMMENTED, and every comment in the thread is from qwen-code-review-bot or qwen-code-dev-bot. This is a decision for a maintainer — which of #15274 / #15275 / #15254 lands, and who closes the others — not a defect in the diff; nothing in the code needs changing to answer it, and this review's own findings are all Suggestion-level.

中文说明

仅完成部分审查,审查缺口已披露。

未探索到全部深度(达到工具调用预算):"agent 6c":executing scripts/tests/sdk-java-workflow.test.js — the review worktree has no node_modules , so every assertion above was traced by hand against the parsed …。

未审查:反向审计——评审时间预算不足,未能开始第 5 轮。

未检查(工具限制,非阻断):.github/workflows/sdk-java.yml——actionlint 对 workflow 内嵌 shell 的源映射尚未支持。

收敛姿态下延后(第 2 轮,非阻断)——已记录,本轮不要求修改:共 2 条(原文未翻译,列表见上方英文部分)。

[Critical] PR-level merge-order blocker carried from the previous round — still stands, and no code change in this diff closes it. Approval was deferred pending a human decision on which of the overlapping autofix PRs should land for the sdk-java test lane. Verified at the reviewed commit: #15233, #15254, #15257 and #15275 all remain OPEN (#15229 and #15232 have since closed), and the mechanism the deferral names still holds — #15254, #15257 and #15275 each edit .github/workflows/sdk-java.yml and each records a different .size-baseline byte count for it (119651, 118481 and 123176 against this PR's 121709), so whichever merges second fails the size ratchet until rebased, and #15233 rewrites the same scripts/tests/sdk-java-workflow.test.js this PR rewrites. No human has reviewed: the only review on the PR is qwen-code-review-bot's COMMENTED, and every comment in the thread is from qwen-code-review-bot or qwen-code-dev-bot. This is a decision for a maintainer — which of #15274 / #15275 / #15254 lands, and who closes the others — not a defect in the diff; nothing in the code needs changing to answer it, and this review's own findings are all Suggestion-level.

— qwen3.8-max via Qwen Code /review (v0.26.0)

Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread .github/workflows/sdk-java.yml Resolved
Comment thread scripts/tests/sdk-java-workflow.test.js Resolved
@qwen-code-dev-bot

Copy link
Copy Markdown
Collaborator Author

🕐 Review received — an automatic review of the current head is still running, so this round is held until it lands (a push now would cancel it and discard its work, #8888). Your feedback stays queued for the next eligible round.

中文说明

🕐 已收到评审 —— 当前 head 上仍有一轮自动 review 在运行,本轮暂缓(现在推送会取消该 review 并丢弃其工作,#8888)。反馈保持排队,等待下一次可运行的轮次处理。

qwen-code-dev-bot and others added 4 commits October 24, 2026 02:34
…15262)

The write-back runs after the suite, so the step's exit code must come from
the suite: both lanes now track the Maven status, demote only a lock-wait
timeout with a step-summary record, and redden when the queue itself is
unusable. The wait drops to 1800 s because it guards only the bounded copy;
the seed stops hiding cp's diagnosis while staying quiet on a fresh host;
the design docs record the head-of-queue stall the shorter hold does not
bound.
The previous round changed .github/scripts/sdk-java-lock.sh, which this PR never touched; the deterministic gate rejects that expansion. Restore the helper and keep both call sites on the pre-round `if !` shape, retaining the 1800 s wait and the step-summary record. Drop the status-2 pins and tests that only existed for the reverted helper, and record the unusable-queue gap as a follow-up in the design docs. The timeout-returns test stays.

Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
@qwen-code-dev-bot

Copy link
Copy Markdown
Collaborator Author

🤖 Addressed the latest review feedback (round 2/10). What changed, and what I pushed back on: · 已处理最新评审反馈(第 2/10 轮)。改动内容与我反驳保留之处如下:

Round summary — PR #15274 (issue #15262)

Round: same-run verification repair, plus re-triage of the nine inline findings already fixed in the rejected round.

Deterministic rejection repaired

The rejected commit added .github/scripts/sdk-java-lock.sh to the diff, an area this PR never touched. Deterministic verification rejects that expansion regardless of which feedback requested it. This follow-up commit restores the helper byte-for-byte to its pre-round version and keeps the fix inside the PR's footprint:

  • Both call sites are back to the pre-round if ! sdk_java_lock_acquire ... shape (still 1800 s, still writing the skip to $GITHUB_STEP_SUMMARY), so no code depends on the reverted status-2 behavior.
  • The status-2 pins and the shortcut test that only existed for the reverted helper are removed; the timeout-returns test is kept, because the demotion still relies on the helper returning rather than exiting.
  • The design docs (English and Chinese) now record that a queue which cannot be created or written returns the same status as a consumed budget, so it also skips the write-back; separating the two needs a follow-up on the helper.
  • .size-baseline is lowered from 124102 to 123812 to match the workflow after the revert.

Inline findings

  • rc:4362748672 (R1-1) — resolved: all four seeds are guarded by [ -d "${HOME}/.m2/repository" ], 2>/dev/null is gone, and the warning carries the cp exit status.
  • rc:4362748681 (R2-1) — resolved: the MariaDB step tracks both Maven runs and exits ${suite_status}; the failed-suite cases assert status 1 and the published marker.
  • rc:4362748696 (R1-3) — resolved: cd ../../managed-agent-server || suite_status=$? gates the managed-agent runs; the missing-checkout case asserts exit 1 and exactly two Maven invocations.
  • rc:4362748712 (R1-4) — escalated, not implemented: the fix requires changing .github/scripts/sdk-java-lock.sh, which is outside this PR's footprint and is the exact change the gate rejected. The residual risk is documented and the open question is on the thread.
  • rc:4362748732 (R2-2) — resolved: both sites wait 1800 s, re-derive the job ceilings, and record every skip in $GITHUB_STEP_SUMMARY.
  • rc:4362748744 (R2-3) — deferred: extracting the shared write-back would create or edit .github/scripts/, also outside this PR's footprint. The thread carries the reason.
  • rc:4362748748 (R1-7) — resolved: the paragraph records the queue-position and head-of-queue figures, the date is 2026-10-22, and the un-ticketed-holder residual risk is stated, in both languages.
  • rc:4362748759 (R2-4) — resolved: the invariant is narrowed to the two suites; the checkstyle/package steps remain ~/.m2 readers, in both languages.
  • rc:4362748771 (R2-5) — resolved: the failed-suite test is parameterised over Java 11/21 and Maven run 1/2/3, with the Java-21 fixture running from the step's working directory.

Maintainer decisions still open

Verification

  • npm run build — passed (COREPACK_HOME=/tmp/corepack-home for the sandbox's read-only cache).
  • npm run typecheck — passed.
  • npm run lint — passed.
  • npx vitest run --config ./scripts/tests/vitest.config.ts scripts/tests/sdk-java-workflow.test.js scripts/tests/sdk-java-lock.test.js scripts/tests/workflow-size.test.js — 377 passed, 1 skipped; re-run clean after the commit hook.
  • bash .github/scripts/check-workflow-size.sh — passed.
  • npx prettier --check on the changed workflow, docs and tests — passed.
中文说明

本轮摘要 — PR #15274(issue #15262)

本轮为同一轮验证修复,并对上一轮(被拒绝的提交)已修复的 9 条行内 finding 重新分诊。

已修复的确定性验证拒绝

被拒绝的提交把 .github/scripts/sdk-java-lock.sh 加入了 diff,而本 PR 从未碰过该区域。无论反馈如何要求,确定性验证都会拒绝这种扩张。本次后续提交把辅助脚本逐字节恢复为轮前版本,并把修复保持在 PR 足迹之内:

  • 两个调用点恢复为轮前的 if ! sdk_java_lock_acquire ... 形态(仍为 1800 秒,仍把跳过写入 $GITHUB_STEP_SUMMARY),因此没有代码依赖被回退的状态 2 行为。
  • 仅存在于被回退辅助脚本的 status-2 断言与快捷测试已移除;超时返回测试保留,因为降级仍依赖辅助脚本“返回”而非“退出”。
  • 设计文档(英文与中文)现在记录:无法创建或写入的队列返回与预算耗尽相同的状态,因此同样跳过写回;区分两者需要在辅助脚本上做后续修改。
  • .size-baseline 从 124102 下调到 123812,与回退后的工作流一致。

行内 finding

  • rc:4362748672(R1-1)— 已解决:四处种子复制都由 [ -d "${HOME}/.m2/repository" ] 守卫,2>/dev/null 已移除,告警携带 cp 的退出状态。
  • rc:4362748681(R2-1)— 已解决:MariaDB 步骤跟踪两次 Maven 运行并以 ${suite_status} 退出;失败套件用例断言状态 1 且 marker 已发布。
  • rc:4362748696(R1-3)— 已解决:cd ../../managed-agent-server || suite_status=$? 为 managed-agent 运行把关;缺少 checkout 的用例断言退出 1 且恰好两次 Maven 调用。
  • rc:4362748712(R1-4)— 升级,未实现:修复需要修改 .github/scripts/sdk-java-lock.sh,该文件在本 PR 足迹之外,而这正是验证门拒绝的改动。残余风险已记录;未决问题已留在讨论串。
  • rc:4362748732(R2-2)— 已解决:两处都等待 1800 秒、重新推导 job 上限,并把每次跳过记录到 $GITHUB_STEP_SUMMARY。
  • rc:4362748744(R2-3)— 延后:抽取共享写回会创建或修改 .github/scripts/,同样在本 PR 足迹之外。讨论串中已记录原因。
  • rc:4362748748(R1-7)— 已解决:段落记录了队列位置与队首停摆的数字,日期为 2026-10-22,并陈述了未取号持有者的残余风险,两种语言同步。
  • rc:4362748759(R2-4)— 已解决:不变量收窄到两套套件;checkstyle/package 步骤仍直接读取 ~/.m2,两种语言同步。
  • rc:4362748771(R2-5)— 已解决:失败套件测试按 Java 11/21 与 Maven 第 1/2/3 次运行参数化,Java-21 fixture 从步骤的 working-directory 运行。

仍需维护者决定

验证

  • npm run build —— 通过(沙箱 corepack 缓存只读,设置 COREPACK_HOME=/tmp/corepack-home)。
  • npm run typecheck —— 通过。
  • npm run lint —— 通过。
  • npx vitest run --config ./scripts/tests/vitest.config.ts scripts/tests/sdk-java-workflow.test.js scripts/tests/sdk-java-lock.test.js scripts/tests/workflow-size.test.js —— 377 通过、1 跳过;提交钩子后再次运行通过。
  • bash .github/scripts/check-workflow-size.sh —— 通过。
  • 对改动的工作流、文档与测试执行 npx prettier --check —— 通过。
Base-conflict check · 基分支冲突检查: no conflict with main. · 与 main 无冲突。

⚠️ The branch received new commits while this round ran; they were merged into this push, but this round's verification predates that merge — re-check anything that landed mid-run. · 本轮运行期间分支收到了新的提交;本次推送已将其合并,但本轮验证在合并之前完成——请复查运行期间落地的改动。

⚠️ Review-thread resolution skipped — guard: salvage merge; resolved 0 of 7 selected thread(s), 7 left for a later round. · 评审线程关闭被跳过——守卫:salvage merge;选中 7 条,本轮关闭 0 条,其余 7 条留待后续轮次。

Re-review when you have a moment. After round 10 this bot stops and leaves the PR for a human. · 有空请复审;第 10 轮后本 bot 停止并将 PR 交给人工。


🧠 Handled by Qwen Code · model/模型 deepseek-v4.1-flash-1017 · CLI 0.26.0

@qwen-code-review-bot

Copy link
Copy Markdown
Collaborator

Qwen Code review timed out. Qwen review timed out after 21600 seconds (of the 360-minute budget). This run already used the maximum 360 minute timeout. See workflow logs.

@github-actions

github-actions Bot commented Nov 4, 2026

Copy link
Copy Markdown
Contributor

Managed Agent gate failed

Run 41826134057 · commit 26960a72d1c6

Failing tests

  • com.alibaba.qwen.code.managedagent.persistence.WorkspaceMigrationMySqlIT.legacyCwdSettlementWaitingForMigrationSeesCommittedFence

@qwen-code-review-bot qwen-code-review-bot left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No blocking issues. LGTM! ✅

Not fully linted (tool limitation, not a blocker): .github/workflows/sdk-java.yml — actionlint ran workflow-level checks; embedded run: shell deferred (source mapping not yet supported).

Deferred under the convergence posture (round 20, not a blocker) — recorded, not requested in this round:

  • .github/workflows/sdk-java.yml:469 — [review] New write-back setup arms bypass lock_setup_fault, and the cd arm drops the reason

— deepseek-v4.1-flash-1017@1f214e4e via Qwen Code /review (v0.26.7)

@github-actions

github-actions Bot commented Nov 5, 2026

Copy link
Copy Markdown
Contributor

Managed Agent gate failed

Run 41955829162 · commit dd55b8f5f89d

Failing tests

  • com.alibaba.qwen.code.managedagent.HostedCleanupRetentionMySqlIT.cleanupAdmissionConflictIsReturnedOverHttpOnMySql

@qwen-code-review-bot qwen-code-review-bot left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No blocking issues. LGTM! ✅

Not fully linted (tool limitation, not a blocker): .github/workflows/sdk-java.yml — actionlint ran workflow-level checks; embedded run: shell deferred (source mapping not yet supported).

Deferred under the convergence posture (round 21, not a blocker) — recorded, not requested in this round:

  • .github/workflows/sdk-java.yml:1018 — [review] MySQL missing-checkout guard is untested: reverting cd packages/managed-agent-server || mvn_status=$? to a bare cd reds nothing while dropping the deferred status, the 300 s clamp, the stag…
  • .github/workflows/sdk-java.yml:1021 — [review] MySQL stage labels are pinned by no test: a stale sdk-java install label or a renamed mysql integration suite reds nothing and mislabels a MySQL failure

— deepseek-v4.1-flash-1017@1f214e4e via Qwen Code /review (v0.26.7)

@github-actions

github-actions Bot commented Nov 5, 2026 •

Copy link
Copy Markdown
Contributor

Managed Agent gate failed

Run 42043047475 · commit 16efc1bf44a0

Failing tests

  • com.alibaba.qwen.code.managedagent.ToolPublicationLifecycleMySqlIT.receiptAndReplayDoNotDeadlockWithSameTenantLifecycleCompletion(boolean)[1]
  • com.alibaba.qwen.code.managedagent.ToolPublicationLifecycleMySqlIT.receiptAndReplayDoNotDeadlockWithSameTenantLifecycleCompletion(boolean)[2]

@wenshao

wenshao commented Nov 5, 2026

Copy link
Copy Markdown
Collaborator

autofix round — 2026-11-05: MySQL lane red again after re-run — needs a human decision

SDK Java (full) / Runtime Broker and Managed Agent MySQL / Java 21 failed twice at this head (initial run + the re-run this watch triggered), so no further automatic re-run will be attempted here. Both failures are the same test: ToolPublicationLifecycleMySqlIT.receiptAndReplayDoNotDeadlockWithSameTenantLifecycleCompletion (both parameterized arms, :161, Expecting value to be true but was false), with the aggregate Managed Agent gate failing alongside.

This is a known main-side test-fixture defect, not something this PR introduced: #16405 ("[Test] Publication lifecycle fixture misses the tenant guard it waits for") is open for exactly this fixture, and this PR changes no Java code. It is the second distinct Java fixture failure seen on this PR's lanes in two days (#16265's WorkspaceMigrationMySqlIT was the other) — the managed-agent suite is currently flaky on main.

Decision for a maintainer: wait for #16405 (and #16265) to be fixed and let the lane go green on a later head, or merge accepting these known main-side fixture flakes (risk call). The watch will not modify the Java fixtures (out of this PR's scope) and will not spend another re-run at this head. Two fresh approvals at 16efc1bf44 remain the other merge gate.

中文说明

autofix 轮次 — 2026-11-05:MySQL 通道重跑后再败——需要人工决定

Runtime Broker and Managed Agent MySQL / Java 21 在本 head 两次失败(首跑 + 重跑),不再自动重试。两次均为同一测试:ToolPublicationLifecycleMySqlIT.receiptAndReplayDoNotDeadlockWithSameTenantLifecycleCompletion(两个参数化臂,:161),聚合 Managed Agent gate 随之失败。

这是已知 main 侧测试夹具缺陷,非本 PR 引入:#16405 正是该夹具的跟踪 issue(fixture 缺 tenant guard),且本 PR 不改 Java 代码。这是两天内该 PR 通道上第二个不同的 Java 夹具失败(另一个为 #16265 的 WorkspaceMigrationMySqlIT)——managed-agent 套件在 main 上目前不稳定。

请维护者决定: 等 #16405(及 #16265)修复后在后续 head 转绿,或接受这些已知 main 侧夹具 flake 直接合并(风险判断)。监视不会修改 Java 夹具(超出本 PR 范围),也不再消耗本 head 的重跑额度。16efc1bf44 上两个新批准仍是另一道合并门槛。

@qwen-code-review-bot qwen-code-review-bot left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reviewed.

1 Suggestion-level finding(s) this review confirmed are already reported on this PR and are not repeated:

  • D22-3 failed-checkout stage label — already reported (round 21 deferral 'MySQL stage labels are pinned by no test …'; round-15 closeout item 'the MySQL cd-deferral/stage-label pins')

Not explored to full depth (tool budget reached): "agent reverse-audit (round 5)": did not execute the vitest suites ( scripts/tests/sdk-java-workflow.test.js , sdk-java-lock.test.js , main-ci-failure-issue-workflow.test.js ) or a real floc….

Not fully linted (tool limitation, not a blocker): .github/workflows/sdk-java.yml — actionlint ran workflow-level checks; embedded run: shell deferred (source mapping not yet supported).

Deferred under the convergence posture (round 22, not a blocker) — recorded, not requested in this round:

  • .github/workflows/sdk-java.yml:302 — [review] Seed nests the host cache at m2-repo/repository when RUNNER_TEMP survives a previous job, so the seed silently stops warming Maven (fix: cp -aln SRC/. DST/)
  • .github/workflows/.size-baseline:63 — [review] Entry records 162545 for a 162758-byte sdk-java.yml (213 over, inside the 4096 allowance), while the design doc claims re-measurement after the last edit

— deepseek-v4.1-flash-1017@1f214e4e via Qwen Code /review (v0.26.8)

@qwen-code-dev-bot

Copy link
Copy Markdown
Collaborator Author

🕐 Review received — an automatic review of the current head is still running, so this round is held until it lands (a push now would cancel it and discard its work, #8888). Your feedback stays queued for the next eligible round.

中文说明

🕐 已收到评审 —— 当前 head 上仍有一轮自动 review 在运行,本轮暂缓(现在推送会取消该 review 并丢弃其工作,#8888)。反馈保持排队,等待下一次可运行的轮次处理。

@qwen-code-review-bot qwen-code-review-bot left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

⚠️ Downgraded from Approve to Comment: CI failing: web-shell E2E Smoke (shard 2/4), web-shell E2E Smoke (shard 4/4). Partially reviewed — gaps disclosed.

Not reviewed: build-and-test — the SDK Java lanes that run this workflow were skipped in CI at this head (tier resolved to none), so the real-pool behaviour of the changed seed, prune and write-back was not exercised; the suites that pin their text did run green locally.

Not explored to full depth (tool budget reached): "agent 6b": did not execute scripts/tests/sdk-java-workflow.test.js , scripts/tests/sdk-java-lock.test.js or scripts/tests/main-ci-failure-issue-workflow.test.js under….

Not fully linted (tool limitation, not a blocker): .github/workflows/sdk-java.yml — actionlint ran workflow-level checks; embedded run: shell deferred (source mapping not yet supported).

Deferred under the convergence posture (round 23, not a blocker) — recorded, not requested in this round:

  • .github/workflows/sdk-java.yml:311 — [review] the seed prune is silent when it cannot delete a hardlinked resolver marker, so the unlocked Maven run can still rewrite the host cache through it
  • .github/workflows/sdk-java.yml:469 — [review] the two new write-back prelude arms bypass lock_setup_fault, and the cd arm's warning names no cause
  • scripts/tests/sdk-java-workflow.test.js:501 — [review] the new marker-set test duplicates the neighbouring pin and adds a third copy of the seven-name literal

— deepseek-v4.1-flash-1017 via Qwen Code /review (v0.26.8)

@wenshao

wenshao commented Nov 6, 2026

Copy link
Copy Markdown
Collaborator

Verdict: merge-ready — 122/122 scripted assertions passed across 46 replay cells (23 scenarios × 2 arms), the Linux guard suites, a 26-test non-vacuity cross-run, and a 24-row mutation matrix (16 suite kills + 4 harness kills + 2 adjudicated survivors + 2 adjudication rows). Verified head f4981199200a3ca4e76512991c1a2d50460ebd91; base 07a467c38a51f1d7d51ea4c12e00e24f825d4f0b (current main, which the head merges and contains). The previous round's two non-blocking Suggestions are re-measured and stand; both are author-recorded deferrals. No new blocking finding.

中文摘要

结论:merge-ready(122/122 断言通过;上一轮两条非阻塞建议复测仍在,均为作者已记录的延期项;无新阻塞)。

增量分类:上轮已验 head 7618d49a → 本轮 f4981199 之间 PR 没有自有生产提交,全部是 main 合并加手写解冲突——写回 prune 列表与 #16141 的 16 坐标取并集、.size-baseline 逐次重记。main 动过闭包(sdk-java.yml +163 行、测试 +100/+33 行),故整套重测而非沿用旧数字。最终合并 f4981199 经 merge-tree 反证为纯机械(树 ed2711fb… 相等)。

A/B 结论(两臂同一份逐字提取的 run: 步骤、真实 bash/flock/find/cp/timeout、仅桩 mvn;46 格):

  • 失败的 MySQL 套件仍写回并保持失败状态与阶段注解:head 2/2 红格写回 1、2 个文件;base 0/2(main 内联失败)。
  • 缓存故障不改写套件结果:不可用宿主缓存 / 缺失 job-local 仓库,head 保持退出 0 并 warning,base 变红(4/4 格翻转)。
  • 写回排除全部 7 类可变名:head 所有格 markers=0;base 每格发布 3–12 个标记。
  • seed 尽力而为(10-27 的 S2 阻塞回归已修):不可删/不可读夹具下 head 两车道都退出 0、Maven 跑满;把 || true 去掉的变异体(h2)复现 exit 1、Maven 0 次。
  • 继承的单链接解析器状态:head 清掉全部 7 种、宿主 peer 完好;base 留下 3 种新名(metadata/part/part.lock)——扩展名集承重。
  • 元数据原地改写:3 次 Maven 后宿主 maven-metadata-local.xml head 仅 ORIGINAL,base 被改写 3 次。
  • 失败腿锁预算:持锁实测 head MySQL 红格 wait=300、绿格 5400;base 红格根本到不了锁。test 车道 300 钳制在 main 已有(两臂相同)。

复测上轮两条建议(仍在,非阻塞):S1 缺失 checkout 时注解仍写成上一成功阶段 'sdk-java install'(两车道复现);S2 seed 剪枝失败仍静默——不可删标记以 nlink=2 硬链接存活且 0 warning。变异 m12 证实 MySQL cd 守卫承重但无测试钉住(h6:去掉后下载丢失、无注解),与作者记录的延期项一致。

CI:SDK Java (fast) / ubuntu-latest / Java 21 在 pool 主机 ecs-qwen-hk3-25 通过,其中被改的 self-hosted 测试步骤真实执行 23 分钟(hosted 变体被跳过);MySQL 集成车道在此 head 被 fast tier 跳过,其 pool 证据仍是上轮 head 的 full-tier 全绿。两条 smoke 红(shard 2/4、4/4)同 run 内 shard 3/4 在另一主机通过——同代码不同主机,失败签名是页面根节点 never rendered + daemon ECONNREFUSED,归因为 hz2 主机环境,diff 不触碰任何 web-shell 文件。轮内到达的第 23 轮 bot 评审因这两条 smoke 红把 Approve 降为 Comment,其「车道被跳过/tier 为 none」的说法与 job 步骤账本不符(已更正)。

未覆盖:真实 Maven 解析语义(桩模拟)、hosted 车道 seed/actions.cache 执行路径(静态四站点计数+套件钉住)、Windows、真实 pool 的受控 A/B。

Previous-round status

The newest substantive earlier round is the 2026-10-31 Linux verification at 7618d49a (comment 6545084664, merge-ready 61/61). Every carried-forward item was re-measured at f4981199, not quoted.

# Finding Severity Status at f498119
1 S1 — missing-checkout annotation names the previous successful stage (both lanes) Suggestion stands (author-recorded deferral; re-measured: test and MySQL cells both annotate 'sdk-java install'; agreed non-blocking)
2 S2 — seed prune failure is silent while the sibling cp warns; an undeletable _remote.repositories survives as an nlink-2 host hardlink Suggestion stands (author-recorded deferral; re-measured: survived=true nlink=2 warnings=0 on both lanes; agreed non-blocking)
3 10-27 S2 blocker — fatal seed prune reddens the leg before Maven blocked (round 1) fixed, re-confirmed: fault cells reach Maven on both lanes (3/3 and 2/2 calls); the fatal shape reproduces only in mutant h2 (exit 1, Maven 0)
4 Round-1 correction — lock relocation already on main description resolved in the body; the scope section still states it correctly
5 Round-15 deferred prose items (baseline record, classifier lock-wait line, write-back comment wording) nit baseline drift re-measured: 162758 actual vs 162545 recorded = +213 bytes, inside the 4096 allowance; the two prose items not re-measured (no behavioral surface)

Delta classification (why a full re-run)

Between 7618d49a and f4981199 the PR authored no production commits — the first-parent path is nine Merge branch 'main' commits. A direct mechanical merge of the two endpoints is not tree-equal (git merge-tree --write-tree --merge-base=fe30b5daaa … yields c5143a6b… with a .size-baseline conflict, vs the head tree ed2711fb…), so the chain carries hand-written resolutions: the write-back coordinate union at a0d162dbcc (main's 16 in-house coordinates from #16141 plus this PR's seven ! -name exclusions) and successive .size-baseline re-records. A per-file added/removed line-multiset comparison of the two effective diffs (base→P′ vs base→H) shows the PR contribution identical in all seven files except those find lines and the baseline numbers. The final merge f4981199 itself is purely mechanical (merge-tree of its two parents equals its tree, ed2711fb…). Because main touched the measurement closure (sdk-java.yml +163 lines, sdk-java-workflow.test.js +100, main-ci-failure-issue-workflow.test.js +33 since fe30b5daaa), the whole matrix was rebuilt and re-run rather than carried forward.

Central claim and A/B

The claim this head is held to: a failed sdk-java Maven suite still reaches the cache write-back and the step keeps Maven's status with a stage annotation; cache setup faults demote to the suite result; the write-back publishes only the job-local delta with all seven mutable name classes excluded and all 16 in-house coordinates pruned; the seed stays best-effort, observable, and clears inherited single-link resolver state; a red MySQL leg waits only the 300 s copy budget.

Both arms ran the verbatim run: blocks of Run Java SDK tests (self-hosted) and the MySQL integration step (repo yaml parser, no retyping) under bash --noprofile --norc -e, with the real sdk-java-lock.sh, real util-linux flock/find/cp/timeout, and a stub mvn that records calls, plants the downloads and all seven marker classes, and simulates the in-place metadata rewrite. Every cell had a fresh HOME/RUNNER_TEMP/workspace; budget cells held the host lock with a real background flock so the timed wait path executed. Witness: 01-ab-cells.png (raw per-cell logs under logs/).

Cell (both lanes unless noted) base head expected
MySQL suite fails at install / integration exit 1, publishes 0 / 0 exit 1, publishes 1 / 2, names the failed stage head keeps status + publishes
Test suite fails at run 1 / run 3 exit 1, publishes 4 / 12 (3 / 9 markers) exit 1, publishes 1 / 3, 0 markers, names stage main already defers here; head adds exclusions
Green write-back 8–12 published incl. 6–9 markers downloads only, 0 markers head excludes all 7 name classes
Green + unusable host cache / missing job-local repo exit 1 (both lanes) exit 0 + warning (both lanes) cache fault must not redden
Seed source undeletable / unreadable exit 0, Maven full, silent cp exit 0, Maven full, cp warns S2 fixed + observable
Missing managed-agent-server checkout exit 1, publishes 0 exit 1, publishes 2 / 1, annotation (wrong stage — S1) downloads survive
Inherited single-link markers (test lane) 3 new-name survivors, host peers intact 0 survivors, host peers intact extended prune set load-bearing
maven-metadata-local.xml through the seed host ORIGINAL+REWRITTEN×3 host ORIGINAL only F1 fixed
Held lock, MySQL red / green red never reaches the lock; green wait=5400 red wait=300; green wait=5400 failed leg waits only the copy budget
Held lock, test red wait=300 wait=300 clamp pre-existing on main
Lock path is a directory + red suite exit 1, no annotation exit 1 + warning + stage annotation suite status survives the lock fault

Lock ordering held in every cell that took a lock on either arm: the first flock call follows the last mvn call (37 lock-taking cells, 0 violations).

Findings

No new blocking finding. The two previous Suggestions stand, re-measured at this head; both are recorded deferrals under the repository's post-five-round convergence rule, and this round agrees with the non-blocking severity.

S1 (Suggestion, stands) — missing-checkout annotation names the previous stage

Re-measured on both lanes: with no managed-agent-server checkout, the step exits 1, keeps the downloads (test 2, MySQL 1), and annotates ::error::Maven suite 'sdk-java install' failed with status 1 although install succeeded. The previous round's measured minimal fix (set mvn_stage before the cd) still applies; the test pin counting mvn_stage occurrences must move with it. The mutation matrix adds the matching coverage fact: m12 (MySQL cd … || mvn_status=$? → bare cd) survived all 143 tests, and harness row h6 drove the mutant build — downloads lost (published 0) and no annotation — so the guard is load-bearing but unpinned, exactly the author-recorded "MySQL cd-deferral/stage-label pins" deferral. The test-lane sibling (m11) is pinned.

S2 (Suggestion, stands) — the seed prune failure is silent

Re-measured: with an undeletable seed entry, both lanes reach Maven (exit 0, full calls) and the marker survives in the job-local copy as an nlink-2 hardlink of the host file with zero warnings, while the sibling cp warns on its failure. The non-fatal shape is correct (mutant h2 proves the fatal version kills the leg before Maven); the missing diagnostic is the gap. A find … -delete || echo "::warning::…" keeps the contract and makes the condition observable.

Nits

  • .size-baseline records 162545 while the workflow measures 162758 (+213 bytes, inside the 4096 growth allowance; the workflow-size suite passes 239 tests). Main-side comment edits since the last re-record; the body's own policy is to re-record with the next content change.
  • m6 (seed prune directory guard → if true) survived the suite but is redundant defence, adjudicated by driving the mutant build (h5): with no job-local repo, find errors and || true swallows them, byte-identical outcome to head. No action needed; listed for completeness.

Mutation matrix (24 rows, all executed)

Baseline control: the unmutated head workflow runs the guard file through the exact mutation runner as 143 passed, and the m0 control (final exit "${mvn_status}" → exit 0) was killed by 10 named assertions — the runner demonstrably kills in the same file it mutates. 18 suite rows: 16 killed, 2 survived (m6, m12, both adjudicated above by replaying the mutant step, not by reading). 6 harness rows: h1 (-links +1 restored → all 7 inherited markers survive), h2 (fatal prune → exit 1, Maven 0, both lanes — the 10-27 regression shape), h3 (MySQL deferral off → red publishes 0 while the test-lane control still publishes 1), h4 (final exit 0 → red MySQL exits 0 while the control exits 1), plus the two adjudication rows h5/h6. Every anchor matched its intended site count exactly; every row restored byte-exact. Witness: 02-mutation-matrix.png; ledgers mutations-suite.json, mutations-harness.json, mutations.json.

Coverage is claimed per-guard over the effective diff, not per-commit: since the previous verified head the PR authored no production commits (merge resolutions only), so a per-commit enumeration would be empty and prove nothing.

Non-vacuity

The head's three changed test files were copied verbatim onto the base production workflow (vac-tree): 26 tests went red (525 passed), all in sdk-java-workflow.test.js, spanning every guard family — failed-suite publish, demotion arms, seed best-effort, inherited-state witness, exclusions, coordinate union, lock scope. The new pins are load-bearing, not decorative.

CI at this head

Job Result Attribution
SDK Java (fast) / ubuntu-latest / Java 21 pass (job 127836870037, pool host ecs-qwen-hk3-25) the changed self-hosted test step executed for 23 minutes and passed (the hosted variant was skipped); the MySQL integration lane was skipped at this head by fast-tier selection
SDK Java (fast) / Real daemon E2E / Java 11, Flyway pass
web-shell E2E Smoke (shard 2/4) fail, 24 failed / 35 passed environment: runner ecs-qwen-hz2-9; every failure is the page-root/composer never rendered within 20 s (all 3 retries), with vite proxy ECONNREFUSED 127.0.0.1:4170 at startup
web-shell E2E Smoke (shard 4/4) fail, 18 failed / 52 passed environment: runner ecs-qwen-hz2-26, same signature across the whole 30-minute run
web-shell E2E Smoke (shard 3/4) pass on ecs-qwen-hz3-6 same-run control: identical merged code passes on a different host
web-shell E2E Smoke (shard 1/4), review-pr, review-scan pending at report time

The same-run shard-3 control is the attribution evidence: one code, three hosts, the two hz2 hosts fail with a runner-level page-load signature while hz3 passes. The diff touches no web-shell file. Main's own ci.yml runs skip the smoke lane, so no main-side comparison was available; the within-run control stands in its place.

Round-23 review arrived during this round (review 5799105222, COMMENTED, downgraded from Approve citing the two smoke reds above). Its "not reviewed" note says the SDK Java lanes were skipped at this head ("tier resolved to none") — the job's step ledger corrects that: the fast tier ran the self-hosted test step on a pool host for 23 minutes; what was skipped is the full-tier MySQL lane. Its three deferred probes map onto this round's measurements: (1) the silent seed prune is S2 above, re-measured; (2) the two write-back prelude arms inline their warn-and-exit instead of calling lock_setup_fault, and the cd arm's warning names no cause — the demotion behavior itself is measured correct (both fault cells keep the suite status), the wording is a diagnostics nit; (3) the marker-set test adds a third copy of the seven-name literal — a test-maintenance nit. None is blocking; all three are recorded, not requested.

Not covered

  • Real Maven resolver semantics. The stub plants the files a real resolution leaves and simulates the in-place metadata rewrite (append through the hardlink); the 10-25 round measured the rewrite with real Maven 3.9.11. This round measures the mechanism the fix relies on, not the resolver.
  • Hosted-lane seeds and actions/cache. The two hosted seed sites are counted and shape-pinned by the suite (four seeds, two write-backs, all asserted at this head) but not executed; no actions/cache restore in this rig.
  • The real self-hosted pool as a controlled A/B. The green SDK Java (fast) run exercised the changed self-hosted test step once on real infrastructure (23 minutes, pool host hk3-25), not a before/after measurement; the changed MySQL step was not pool-executed at this head (fast-tier selection skipped the full lanes) — its pool evidence is the previous head's full-tier green plus this round's replay.
  • Windows and the Java 11/17 toolchain matrix (the replay ran Java 11 and 21 legs as env settings, not real toolchains).
  • The two prose deferrals (classifier fixture lock-wait line, MySQL write-back comment wording) — no behavioral surface to re-measure.
  • Existing-destination seed nesting (cp -aln into an existing m2-repo nests the cache): both arms show it, owned by fix(ci): keep sdk-java resolver bookkeeping and in-house jars out of the host cache (#15281) #15286, unchanged by this diff. The inherited-state cells plant markers around it; the prune still reaches them.

Methodology

Isolation: node:22-bookworm container, --network none, env -i, non-root uid 54754, no credential mounts; verify-pr-prepare.mjs --require-isolation exited 0 inside (/.dockerenv; credential-free per attestation). Arms are git archive trees at the head and at baseRefOid (which is the current main tip and the head's second parent, so no trial merge was needed); node_modules hardlink-copied per tree. The replay extracted both steps verbatim with the repo's yaml parser and ran 23 scenarios per arm with per-cell fresh rigs; budget cells held the lock with a real background flock, captured the first flock --wait N argument off a wrapper ledger, then released. Gates ran through a round-local vitest config (the repo config with globalSetup disabled for the air-gapped container and workers capped at 4); the one skipped test is a base-SHA-dependent workflow-size pin whose git fetch cannot resolve without network. Suite mutations ran in a third tree with byte-level pristine restore (no git inside the container); harness mutations re-extracted pristine steps per row. The assertion ledger (assertions.json, logs/assertions.log) counts 122 scripted checks: 88 replay, 3 gate, 21 mutation, 6 harness-mutation, 2 static, 2 suggestion-reproductions — expected base-arm reds are encoded as assertions that base exhibits the old behavior and PASS when it does. Raw logs: logs/replay-{head,base}.txt, logs/gate-{head,base,vac}.log, logs/mut-*.log, logs/ci-smoke-shard{2,4}.log; result JSONs results-{head,base}.json, results-mut-*.json. Evidence figures are captures of this round's recorded outputs (01-ab-cells.png, 02-mutation-matrix.png, 03-gates-and-ci.png).

Evidence

A/B replay cells

Mutation matrix

Gates and CI attribution

— qwen-latest-series-invite-202610-m6@e6bf8ffe via Qwen Code /verify-pr (v0.26.8; skill c5273cc80df5)

wenshao
wenshao previously approved these changes Nov 6, 2026
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
@wenshao
wenshao dismissed stale reviews from qwen-code-review-bot and themself via 54edbe3 November 6, 2026 16:54
@wenshao

wenshao commented Nov 6, 2026

Copy link
Copy Markdown
Collaborator

Verdict: merge-ready — 123/123 scripted assertions passed at the new head: 46 replay cells re-run, Linux guard suites green, the 26-test non-vacuity cross-run still red on base production, and the 24-row mutation matrix reproduced (16 suite kills + 4 harness kills + 2 adjudicated survivors). Verified head 54edbe30b9aefafcf43bef4dc4672ca866c72204; base cb282ef5cfe0ba77e2bf8dd3a0bcba302415f924 (current main, which the head merges and contains). The delta since the r2-verified head is one main merge with no PR-authored commits; the two deferred Suggestions stand, and the r2 size-baseline nit is resolved.

中文摘要

结论:merge-ready(123/123 断言;自上轮已验 head 以来 PR 零自有提交,两条延期建议仍在,r2 的基线小疵已消除)。

增量分类:f4981199 → 54edbe30 只有一个 merge 提交(父:旧 head + 新 main tip cb282ef5)。merge-tree 反证:直接机械合并会在 .size-baseline 冲突(树 c70741a8… ≠ 实际树 206f52ba…),remerge-diff 显示唯一手写部分就是基线重记为 165241。main 侧新增独立的「Restore runner-local Java tool cache」步骤(5 处)及其测试 pin(+40 行);本 PR 的两个目标步骤与锁脚本与 r2 字节一致(cmp 证据在案)。因生产文件 blob 变了,仍整套重测。

重测结果(与 r2 逐格一致):46 格回放全翻转如预期——失败 MySQL 套件写回 1/2 个文件并保持状态与阶段注解;缓存故障不改写套件结果;7 类可变名全排除;继承单链接状态 head 清 7 种/base 留 3 种;元数据硬链接改写被挡;持锁实测红腿 wait=300、绿腿 5400。门禁 head 552/553、base 529/530(各 +1 是 main 新步骤的 pin);head 测试对 base 生产码仍 26 条红。变异基线 144 绿,16 杀 + m6(冗余防御)/m12(承重无测试,作者延期项)由 harness 仲裁。

两条建议复测仍在(非阻塞):S1 缺失 checkout 注解仍指上一成功阶段;S2 seed 剪枝失败仍静默(nlink=2 存活、0 warning)。r2 基线小疵已修:165241 记录 = 165241 实测,精确相等。

CI:此 head 0 失败(28 过 / 4 pending / 18 跳过)——r2 的两条 smoke 红在此 head 不再失败,佐证 hz2 主机归因。wenshao 在旧 head 的批准被 head 前进按过期 dismiss,需在新 head 重新批准(流程状态,非发现)。

Previous-round status

The previous round is r2 at f4981199 (comment 6683409904, merge-ready 122/122). Every item re-measured at 54edbe30:

# Finding Severity Status at 54edbe3
1 S1 — missing-checkout annotation names the previous successful stage (both lanes) Suggestion stands (re-measured: both lanes annotate 'sdk-java install'; author-recorded deferral, non-blocking)
2 S2 — seed prune failure silent; undeletable _remote.repositories survives as nlink-2 host hardlink, 0 warnings Suggestion stands (re-measured on both lanes; author-recorded deferral, non-blocking)
3 m12 — MySQL cd guard load-bearing but unpinned by the suite Suggestion stands (m12 survived 144 tests again; h6 replay: mutant loses the downloads and the annotation)
4 r2 nit — .size-baseline +213 bytes drift nit resolved: the merge re-recorded 165241, exactly the measured byte count
5 10-27 S2 blocker — fatal seed prune blocked (r1) fixed, re-confirmed: fault cells reach Maven on both lanes; the fatal shape reproduces only in mutant h2

Delta classification

f4981199..54edbe30 contains exactly one commit — Merge branch 'main' into autofix/issue-15262 (parents: the r2 head and cb282ef5cf, the current main tip). A direct mechanical merge of those parents is not tree-equal (git merge-tree --write-tree yields c70741a8… with a .size-baseline conflict; the head tree is 206f52ba…); git show --remerge-diff 54edbe30 shows the only hand-written resolution is the baseline re-record (162545 vs 159831 → 165241). The workflow and test files merged without conflict.

Main-side closure changes since the r2 base: sdk-java.yml +48 lines — a new independent Restore runner-local Java tool cache step at 5 sites calling .github/scripts/restore-java-tool-cache.sh — plus +40 test lines pinning it, and the baseline. The PR's own measurement closure is byte-identical to r2: both target run: blocks (Run Java SDK tests (self-hosted), the MySQL integration step) and sdk-java-lock.sh compare equal by cmp/blob (evidence: logs-step-compare.txt). Because the production file blobs changed, the full matrix was re-run rather than carried forward; every replay cell reproduced its r2 outcome, as the closure identity predicts.

A/B and gates (all re-run at this head)

Same rig as r2: verbatim-extracted run: blocks, real bash/flock/find/cp/timeout and sdk-java-lock.sh, stub mvn, fresh per-cell rigs, held-lock cells for the wait budgets; node:22-bookworm, --network none, non-root, --require-isolation exit 0. 23 scenarios × 2 arms = 46 cells, 37 lock-taking cells with zero ordering violations. Witness: 01-ab-cells.png.

Measurement r3 result
Failed MySQL suite (install / integration) head exit 1, publishes 1 / 2, names the failed stage; base publishes 0
Green write-back head 0 markers (7 name classes excluded, 16 coordinates pruned); base 6–9 markers
Cache faults (unusable host cache / missing job-local repo) head exit 0 + warning; base exit 1 (4/4 flips)
Seed faults (undeletable / unreadable) both arms reach Maven; head warns on cp failure, base silent
Inherited single-link markers head clears all 7, base leaves the 3 new names; host peers intact
Metadata rewrite through the seed head host file ORIGINAL only; base ORIGINAL+REWRITTEN×3
Held-lock budgets MySQL head red 300 / green 5400; base red never reaches the lock; test-lane 300 pre-existing on main
Guard suites (Linux) head 552 passed | 1 skipped (553), base 529 passed | 1 skipped (530) — the +1 each is main's new tool-cache pin
Non-vacuity head's 3 test files on base production: 26 red (526 passed), same 26 pins as r2
Mutation matrix baseline 144 green through the exact runner; 18 suite rows: 16 killed, m6/m12 survived and adjudicated by replay (h5: redundant defence; h6: load-bearing, unpinned); 4 harness kills (h1–h4). Witness: 02-mutation-matrix.png
Static shape 2 write-backs × (16 coords, 7 exclusions); 4 seeds × 7 names, no -links; baseline 165241 = measured 165241

CI at this head

0 failures: 28 pass / 4 pending / 18 skipping at report time. The two r2 smoke reds (shards 2/4 and 4/4, attributed to the hz2 hosts via the same-run shard-3 control) are no longer failing at this head — shards 3/4 and 4/4 passed, 1/4 and 2/4 pending. Pending: the two smoke shards, review-pr, and SDK Java (fast) / ubuntu-latest / Java 21 (the lane that executes the changed self-hosted test step on a pool host; at r2 it ran 23 minutes green). SDK Java (fast) / Real daemon E2E / Java 11 and Flyway already pass. Witness: 03-gates-and-ci.png.

Review state (process, not a finding): the head advance dismissed wenshao's approval on f4981199 (review 5800991434, submitted after the r2 report); reviewDecision is REVIEW_REQUIRED. Fresh maintainer approvals are needed at 54edbe30; this round's delta evidence above is what a re-approval can rely on.

Not covered

Unchanged from r2: real Maven resolver semantics (stub simulates the in-place rewrite; measured with real Maven in the 10-25 round); hosted-lane seed execution and actions/cache (shape-pinned by the suite, not executed); the pool as a controlled A/B (the green ubuntu lane is one real-infrastructure pass); Windows; the two prose deferrals; existing-destination seed nesting (owned by #15286, both arms). Main's new tool-cache step is base behavior on both arms and outside this PR's delta; its pins ran green in the head gate.

Methodology

Round artifacts: tmp/pr15274-verify-20261107-011138/. The r2 harness was reused verbatim (driver, cells, stubs, mutation runners, assertion ledger) with only the round OIDs and the gate-count anchors updated; the assertion ledger counts 123 checks by source (88 replay incl. 5 budget, 3 gate, 21 mutation-suite, 6 harness-mutation, 3 static incl. the closure-identity check, 2 suggestion-reproductions). Arms are git archive trees at the head and base with hardlinked node_modules; mutations ran in a third tree with byte-level pristine restore; the vacuity cross-run used a fourth tree (base production + head tests). Raw logs under logs/, result JSONs results-*.json, ledgers assertions.json / mutations*.json. Evidence figures capture this round's recorded outputs; the three PNGs are hosted on the assets branch and linked below.

Evidence

A/B replay cells

Mutation matrix

Gates and CI

— qwen-latest-series-invite-202610-m6@e6bf8ffe via Qwen Code /verify-pr (v0.26.8; skill c5273cc80df5)

@wenshao

wenshao commented Nov 6, 2026

Copy link
Copy Markdown
Collaborator

CI attribution at 54edbe30 — the three smoke reds are runner-level page-load flakes, not a regression; failed jobs re-run (attempt 2). The r3 verification verdict (merge-ready, comment 6691055600) stands; no code change is warranted.

中文摘要

CI 归因(head 54edbe30,r3 已验 head,无新提交):三条 smoke 红都是 runner 级页面加载 flake,非回归;已重跑失败 job(attempt 2),不改代码。r3 的 merge-ready 结论不变。

  • 「web-shell E2E Smoke (ubuntu-latest, Node 22.x)」不是独立失败——它是单行聚合门禁 test "$SMOKE_RESULT" = success,因 shard 红而红(10 秒 job)。
  • shard 1/4(hz4-2):12 failed / 14 flaky / 95 passed,日志 134 行 ECONNREFUSED 127.0.0.1:4170;shard 2/4(hz4-8):16 failed / 16 flaky / 80 passed,190 行同签名——已知的「mock daemon 未接受连接」变体:页面根/组件整 20s 不渲染,失败横跨 generated-ui、files、excel-preview 等规格。
  • 同 run 对照:shard 3/4(hz5-5)与 4/4(hz2-30)用同一份代码与同一个 unit-test-tree artifact 全绿。
  • 失败规格与本 PR diff(仅 workflow/测试/文档)交集为空。
  • 部分重跑安全:shard 以显式 run-id 下载 unit-test-tree,该 artifact 无 attempt 后缀且未过期。

Evidence

Check Runner Result Signature
Smoke shard 1/4 ecs-qwen-hz4-2 12 failed / 14 flaky / 95 passed 134× ECONNREFUSED 127.0.0.1:4170, page.waitForFunction 20 s timeouts
Smoke shard 2/4 ecs-qwen-hz4-8 16 failed / 16 flaky / 80 passed 190× ECONNREFUSED 127.0.0.1:4170, element(s) not found
Smoke (ubuntu-latest, Node 22.x) ecs-qwen-hz2-25 fail in 10 s aggregation gate only: test "$SMOKE_RESULT" = success with SMOKE_RESULT: failure
Smoke shard 3/4 ecs-qwen-hz5-5 pass same-run control
Smoke shard 4/4 ecs-qwen-hz2-30 pass same-run control
  1. Known flake family. The connection-refused counts (134 / 190 lines of [WebServer] Error: connect ECONNREFUSED 127.0.0.1:4170, starting at dev-server startup and continuing through the run) match the recorded "mock daemon never accepted connections" variant: the workspace never reaches ready, so page roots/composers time out across unrelated specs. The 14 / 16 flaky counts show the same hosts recovering intermittently. This is the same family attributed at the previous head (then on hz2 hosts); the failing hosts rotate, the signature does not.
  2. Same-run control. Shards 3/4 and 4/4 ran the identical merged code and downloaded the same unit-test-tree artifact (run-id pinned) and passed on different hosts. One code, four hosts, two fail with a connection-level signature ⇒ host-side cause.
  3. Zero diff intersection. The failing specs (generated-ui.*, web-shell.files.*, excel-preview, copy-markdown) share no file with this PR's delta (.github/workflows/sdk-java.yml, scripts/tests/*, design docs, .size-baseline).
  4. Partial re-run safety. The shards download unit-test-tree with an explicit run-id; the artifact has no attempt suffix and is not expired, so re-running only the failed jobs keeps the download resolvable.

Action taken: re-ran the failed jobs of run 42256680321 (attempt 2, no code change). The aggregation gate re-evaluates after the shards. Result will be noted here once settled.

— qwen-latest-series-invite-202610-m6@e6bf8ffe via Qwen Code /verify-pr (v0.26.8; skill c5273cc80df5)

@qwen-code-review-bot qwen-code-review-bot left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No blocking issues. LGTM! ✅

Not fully linted (tool limitation, not a blocker): .github/workflows/sdk-java.yml — actionlint ran workflow-level checks; embedded run: shell deferred (source mapping not yet supported).

Deferred under the convergence posture (round 24, not a blocker) — recorded, not requested in this round:

  • .github/workflows/sdk-java.yml:314 — [review] seed nests into an existing m2-repo; warm cache unreachable
  • .github/workflows/sdk-java.yml:320 — [review] seed prune silent; comment contradicts the test comment
  • .github/workflows/sdk-java.yml:366 — [review] checkout failure misattributed to a passed Maven suite
  • .github/workflows/sdk-java.yml:1152 — [review] stale 'Maven already passed' premise in the mysql write-back
  • scripts/tests/sdk-java-workflow.test.js:381 — [review] byte-identity test neutralized by a two-sided no-op
  • scripts/tests/sdk-java-workflow.test.js:1361 — [review] seeded-alias assertion is vacuous without a seed-landed…
  • scripts/tests/sdk-java-workflow.test.js:1529 — [review] checkout guard unpinned; mutation survives
  • scripts/tests/sdk-java-workflow.test.js:3009 — [review] mysql stage names unpinned

— deepseek-v4.1-flash-1017 via Qwen Code /review (v0.26.8)

@yiliang114
yiliang114 merged commit e992868 into main Nov 7, 2026
139 of 145 checks passed
@github-actions

github-actions Bot commented Nov 7, 2026 •

Copy link
Copy Markdown
Contributor

Managed Agent gate failed

Run 42268556452 · commit 54edbe30b9ae

Failing tests

  • com.alibaba.qwen.code.managedagent.HostedWorkspaceToolTurnIT.recordsLatencyWithDelayedRuntimeProvisioning

wenshao added a commit that referenced this pull request Nov 7, 2026
…n-hook

main advanced 11 commits (stacked-skill overflow #16404, workflow agent
approvals #16432, daemon UI tool blocks #16306, sibling model settings
#15829, late snapshot publication claim #16489, inline FQN imports in
runtime modules #16504, cancelled shell as interrupted #15836, CSI statement
helper fold #16503, Maven suite cache write-back #15274, sed char match
#15701, Ink TUI tool-batch round #16428). One conflict, the workflow size
ratchet:

- .size-baseline: sdk-java.yml auto-merged to 167124 bytes (HEAD recorded
  161927 / main 165241, both below the combined actual), so record the actual
  merged size 167124 per the ratchet's "record the real size in the same PR"
  rule. The absolute gate (470000) and the 4096-byte baseline allowance both
  pass.

No source conflict: main's commits did not touch this branch's A3 source
files, so RuntimeBrokerService / RuntimeExecutionRecovery (R18-1/R19-1
observe-only mount-handback guard) and the cli serve fence are unchanged.

Verification: check-workflow-size.sh green (bash5); workflow meta tests
(workflow-size + sdk-java-workflow + hosted-process-ci) 428 green; workspace
build and ClassSizeRatchetTest re-run after the merge.
@qwen-code-review-bot

Copy link
Copy Markdown
Collaborator

Released in v0.27.0.

@github-actions

github-actions Bot commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

Managed Agent gate failed

Run 42314638208 · commit 54edbe30b9ae

Failing tests

  • com.alibaba.qwen.code.managedagent.HostedWorkspaceToolTurnIT.recordsLatencyWithDelayedRuntimeProvisioning

Tracking issue

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

autofix/needs-human The autofix loop stopped on this PR — a human must re-arm, split, merge, or close it

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants