E2E report / 真机验证报告

Real local macOS daemon and real shell processes, captured on 2026-09-20 during implementation; screenshots are linked in the PR body. No mocked daemon was used. These captures precede the final hook compatibility fix and are not claimed as a fresh final-commit E2E run.

Limits: the initial success command was submitted from a separate real client and required browser reload, so it does not prove automatic live completion. Reconnection navigated home and required reopening the session. Intermediate output of the interrupted run was not recorded. Its restored step summary showed 3 seconds rather than the observed approximately 180 seconds, so that summary is not elapsed-time evidence. Existing export-template artifacts were reused for CLI startup because the full export renderer build exceeds its budget. A new screenshot attempt during PR creation was blocked by browser navigation policy.

Final hook fix verification: a 40,022-character successful stdout retains its middle sentinel; PostToolUse receives all 40,103 display characters including the saved-output notice. History/recording still cap each text field at 32,000 characters and ACP rawOutput remains within 65,536 UTF-8 JSON bytes. Four focused checks passed using actual producer/handler code with mocked execution service, filesystem and hook runner; this is not an OS hook E2E test. The 599-test core suite passed. Workspace typechecks passed; full build and root integration typecheck retain the blockers listed in the PR.

中文

2026-09-20 实现过程中使用 macOS 本地真实 daemon 和 Shell 进程验证，截图见 PR 正文，没有模拟 daemon。图片采集早于最终 hook 兼容修复，不代表最终提交的新一轮真机 E2E。

• 运行中：等待门控文件的前台命令显示运行状态、耗时及明确超时。现有 ACP 链路未提供普通中间输出，因此卡片显示“等待输出”。
• 成功：真实健康检查退出 0，展示三行输出，长运行提示单独呈现，检查了复制后的勾号。
• 失败：浏览器提交的命令等待 12 秒、输出 stdout/stderr 后退出 7，页面无需刷新即自动进入失败态并保留输出。
• 完成后重启：成功和失败记录的结构化字段与重启前一致。
• 运行中重启：240 秒命令在 180 秒心跳后被中断。重启后没有活跃请求或残留探针进程，卡片显示缺失结果的失败回退，没有虚构输出或退出码。

限制：最初成功命令由另一个真实客户端提交，浏览器需要刷新，因此不能证明自动实时完成。重连曾回到首页，需要重新打开会话。被中断命令的中间输出没有录制；恢复后的步骤摘要显示 3 秒，与约 180 秒实际观察不符，不能作为耗时证据。因完整导出渲染包超限，启动 CLI 时复用了已有导出模板产物。本次创建 PR 时的新截图尝试被浏览器导航策略阻塞。

最终 hook 修复验证：40,022 字符成功 stdout 的中间标记完整保留，PostToolUse 收到包含文件提示的全部 40,103 字符展示文本。历史/录制文本字段仍限制为 32,000 字符，ACP rawOutput 仍不超过 65,536 UTF-8 JSON 字节。4 项针对性检查使用真实生产端/handler，底层执行服务、文件系统和 hook runner 使用 mock，不是操作系统 hook E2E。599 项 core 测试通过；各 workspace 类型检查通过；完整构建及根集成类型检查仍有 PR 中列明的阻塞。

Update: rebased onto main as f492336; PR now contains only the 55 feature files and is mergeable. Reran 599 core + 132 Web + 38 ACP tests: all 769 passed. The independent 4-check hook verification and full-build/typecheck results above were collected before this rebase; they were not rerun.

更新：已迁移到 main，提交 f492336；PR 仅包含 55 个功能文件，无合并冲突。重新执行 599 项 core、132 项 Web 和 38 项 ACP 测试，合计 769 项通过。上面的 4 项独立 hook 验证及完整构建/类型检查结果采集于此次 rebase 前，未重新执行。

Please do not rebase or force-push to an active PR as it invalidates existing review comments. Note for future reference, the bots always squash all changes into a single commit automatically as part of the integration.

Maintainer verification — real local environment, PR #12311

Verified f492336b26 (base be9ed41afd, 0 commits behind) on macOS 15 / Node 24.18.1, in two isolated worktrees built from scratch (pnpm install --frozen-lockfile + root npm run build).

Rig — everything is real except the model. Two production daemons, each serving its own built packages/web-shell/dist (not vite dev, not a mocked daemon): node packages/cli/dist/index.js serve --port 5311|5312 --workspace <ws> --web, each with an isolated QWEN_HOME. The only stub is a local OpenAI-compatible server on 127.0.0.1:5399 that returns scripted run_shell_command tool calls, so scenarios are deterministic. Every shell process, the ACP child, the SSE transport, chat recording, the export pipeline and the browser (Playwright/Chromium against the daemon origin) are real. Arm A = this PR, arm B = origin/main, same commands, same scripted calls.

Verdict

The feature works as described on a real daemon, and I found no correctness regression in any consumer I could exercise. Two statements in the PR description are stale and should be corrected, one CI job is red for a reason that is not this PR, and I have one Minor finding worth a decision before merge (exported documents), plus four nits.

1. The red Test (ubuntu-latest) job is inherited from main

packages/web-shell → client/live/messages.test.ts > Live Voice messages > are defined and used nowhere outside client/live fails with:

AssertionError: expected [ 'components/composer/AddMenu.tsx' ] to deeply equal []

I reproduced it on origin/main (be9ed41afd) with byte-identical output. AddMenu.tsx:684 calls t('live.open'), added by #12252; the guard test landed one commit later in #12305, so main is currently red. Full packages/web-shell suite, same machine, back to back:

BranchPickerPopover.test.tsx passes in isolation on both arms (141/141) and CI's run of this PR reported only the one failure, so it is full-suite load flake, not PR-induced. This PR adds no new test failures; main needs a separate fix (drop t('live.open') from AddMenu.tsx, or move the key out of the live. namespace).

2. Two stale claims in the PR description

Please update the description — as written it reads as if the PR is build-blocked, which it is not.

3. Reviewer Test Plan — all five steps, real daemon

4. Compatibility A/B against main (the part I would most want proven)

5. Findings

F1 — Minor — exported HTML documents get a half-populated card. sanitizeResultPreview maps shell_result → { kind: 'text', text }, so the exported envelope carries no structured fields and ShellToolOutput runs with result === null. Measured on a real export rendered with this build's renderer:

• Execution details contains only Timeout / Use default — no exit code, directory, signal, notices or output files.
• A failed run puts the whole Command: / Directory: / Output: / Error: / Exit Code: / Signal: / Process Group PGID: envelope inside the Output pane — re-creating exactly the output/metadata mixing the card removes in the live UI.
• The same run is labelled Succeeded live but Completed in the export, because result?.exitCode === 0 cannot be evaluated.

docs/design/structured-shell-results.md does say the v1 document schema "uses its sanitized text fallback", so the mechanism is intentional — but the visible result is a card that looks empty/inconsistent. Either carry the structured result through the document schema, or suppress the Execution details disclosure (and the failure-envelope Output) when result is null. See figure 6.

F2 — Minor — directory changed meaning and the (root) branch is dead. shell.ts sets directory: cwd (resolved absolute path), but ShellToolOutput renders result.directory === '(root)' ? t('shell.result.defaultDirectory') : result.directory. The producer never emits '(root)' — that is the legacy text label, built from this.params.directory || '(root)' at shell.ts:2835 — so shell.result.defaultDirectory is unreachable and the card always shows an absolute host path where main showed Directory: (root) or the requested relative subdirectory. Verified in the transcript: "directory":"/private/tmp/.../rig/ws". Either emit the requested-directory semantics, or drop the dead comparison.

F3 — Minor — Exit code 0 is shown for timed-out and cancelled runs. Measured: timeout → {outcome:'timed_out', exitCode:0, signal:15}; user cancel → {outcome:'cancelled', exitCode:0, signal:15}. The card then prints Exit code 0 in Execution details right next to Timed out / Cancelled. main never surfaced an exit code in either case (its display text was Command timed out after 2000ms… / Command cancelled by user.), so this is newly misleading. Suggest suppressing Exit code when signal is set or outcome is timed_out / cancelled. See figure 7.

F4 — Nit — three test-coverage gaps, found by mutation. 19 mutants over the new logic, 15 killed. Survivors:

Both directory and outputFiles are correct in the live rig, so these are pinning gaps rather than defects — but they are exactly the two fields F2 is about.

F5 — Nit — container-execution-environment.ts:702. The "Container cleanup failed after tool execution" notice is still appended only if (typeof result.returnDisplay === 'string'), so with a structured shell result it is now silently dropped from the display (it survives in llmContent and error.message). Code read only — I did not exercise a container runtime.

F6 — Nit — small cleanups. shell.result.elapsed is added to both locales but referenced nowhere (dead weight in a byte-budgeted renderer); transcriptToMessages.ts re-implements isShellToolName as an inline regex in two places instead of importing the existing helper (same four names today, easy to drift); the error <pre> (styles.shellFailure) has no max-height/scroll unlike Command and Output (bounded in practice by the ACP budget, so cosmetic).

6. Test runs on this head

7. Evidence

All captures are from the real daemons above; the "before" images come from the same rig running origin/main.

Successful command — before / after

Running command — before / after. main renders the raw tool input JSON as the card body; the PR suppresses it and shows Running with live elapsed time.

Long-run notice — before / after. This is the clearest demonstration of the PR's premise: process metadata leaves the stdout blob.

Outcomes: exit 7, grep exit 1, timeout, cancellation

Truncation and restart recovery

F1 — exported HTML transcript, before / after

F3 — Exit code 0 on a timed-out run

Narrow viewport and copy feedback

8. Not verified

Windows and Linux (macOS only). Live PTY ansiOutput streaming into the card — the current ACP path does not deliver ordinary intermediate stdout, so parseShellLiveOutput was never reached in this rig. Container execution environment (F5). The model was a local scripted stub, so nothing here exercises real model behaviour. Real-world rendering of an exported document still pins the renderer to the published 0.24.1 asset on unpkg; I rendered the export against this build's renderer instead.

Recommendation

Merge-ready on correctness, with a decision needed on F1 (exported documents) and, ideally, one-line fixes for F2 and F3. The red Test (ubuntu-latest) job should not block this PR — it needs a separate fix on main. Please also refresh the "build blocked" and "typecheck fails" paragraphs in the description.

AssertionError: expected [ 'components/composer/AddMenu.tsx' ] to deeply equal []

Review follow-up in c34536c2fe (additive commit; no rebase):

• F1 fixed: text-only document exports render the complete fallback without the live categorized card or inferred execution details. Kept the v1 document schema and its sanitization intact.
• F2 fixed: retain resolved absolute execution-directory semantics; remove the unreachable (root) comparison and unused default-directory translations.
• F3 fixed: hide exit code for cancellation, timeout and signal termination; preserve outcome and signal.
• F4 covered: pin the actual directory, nonempty outputFiles and completed outcome when background promotion loses the race. No action for the equivalent mutant.
• F5 fixed: cleanup warnings are appended to compatible text and structured notices, leaving stdout unchanged. Verification uses a mocked worker; no real container-runtime claim.
• F6 deferred: cosmetic cleanup only; no further unrelated refactoring in this review cycle. The dedicated visual-preview workflow coverage gap is also a follow-up, separate from the existing component/browser regression tests.

Validation: the new assertions first failed on the old implementation (4 UI + 1 container case), then passed. The complete affected runs passed 385 core + 136 Web = 521 tests, plus 5 independent focused checks. Formatting, lint and whitespace checks passed. Two post-fix source/test/call-path audit passes found no further issue in this correction.

Root build now exits 0; all workspace typechecks pass. On local Node 22.14.0, root integration typecheck still fails at process-registry.ts:773 with string/NonSharedBuffer incompatibility; the generated-template error is gone. This is reported separately from the maintainer's successful full check on Node 24.18.1. Updated both languages in the PR description accordingly.

The PR body includes a new browser screenshot of fixed component fixtures, explicitly marked as fixed test data, not a daemon or real shell execution. Existing real-daemon screenshots remain labelled separately. Images are in the fork's independent asset branch, not the implementation branch.

Known deferred compatibility difference: structured ACP text and output share 64KB. Daemon CLI long display text may truncate earlier than the former string payload (40,021 characters became 32,667 in the targeted projection check). Normal Ink rendering parity does not cover this boundary. We retain the existing transport limit in this PR and record CLI long-output budget compatibility as a follow-up, rather than claiming exact parity.

中文：追加提交 c34536c2fe 已处理 F1–F5：导出纯文本回退、明确实际目录并删死分支、隐藏异常终止退出码、补强三处断言、恢复容器清理提示。F6 外观清理及专用视觉预览覆盖留待后续，避免继续扩大改动。

新增断言先在旧实现上失败（4 项 UI、1 项容器），修复后通过；完整相关测试 385 core + 136 Web = 521 项，另有 5 项独立验证通过。格式、Lint、空白检查通过；两轮修复后源码/测试/调用链复查未发现本次修正的新问题。容器验证使用 mock worker，没有声称真机容器测试。

完整构建已通过，各 workspace 类型检查通过。本机 Node 22.14.0 的根集成类型检查仍报 process-registry.ts:773 string/NonSharedBuffer 不兼容，缺失生成模板错误已消失；与维护者 Node 24.18.1 全量通过的结果分别说明。PR 中英文描述已更新。新增浏览器截图明确标为固定组件测试数据，不是真实 Shell/daemon；原有真机截图单独保留，图片不进入实现分支。

明确后续项：结构化 ACP 的 text/output 共用 64KB，daemon CLI 长文本可能提前截断（定向检查中 40,021 字符变为 32,667）。普通 Ink 显示一致性不覆盖此边界。本 PR 保留当前传输限额，将 CLI 长输出预算兼容记录为后续工作，不声称完全等价。

Fixed the two failing scripts tests from the Ubuntu Node 22 test job in 6c2ee93. The new shellResult export was missing from the integration-test source aliases and the review capture loader's named source map. Added both mappings (4 lines); no product behavior changes.

Validation: reproduced the two failures before the fix; the same two test files now pass all 4 tests. Prettier, ESLint, diff check, and full build pass. All workspace typechecks pass. Local integration typecheck still reports the previously documented packages/acp-bridge/src/process-registry.ts:773 string/NonSharedBuffer mismatch on Node 22.14.0; full typecheck is therefore not claimed green. Remote CI for the new commit is pending.

已修复本轮 Ubuntu Node 22 Test 中的两项失败：新增 shellResult 导出时漏同步集成测试源码别名和 review capture loader 映射。仅补齐 4 行映射。修复前复现失败，修复后相同测试 4 项全部通过；格式、lint、构建通过。workspace 类型检查全部通过，本地集成类型检查仍存在此前记录的 process-registry.ts:773 类型错误。新提交的远端 CI 待确认。

@qwen-code /takeover

🤝 Takeover engaged: the autofix loop now manages this PR — it will address new review feedback and resolve base conflicts until the label is removed or the round cap is reached. This is a fork PR, so the first round comes from the next scheduled scan (usually within minutes). Remove the autofix/takeover label (or comment @qwen-code /takeover stop) to release.

@qwen-code /resolve

Qwen Code resolved the merge conflicts and pushed the branch update.

Merge resolution — PR #12311 ← main

Merge commit ac01af9d1d (PR head 6c2ee93ae0 × origin/main b85264be79). One conflict: integration-tests/tsconfig.json.

Root cause. Main's b3dee94b64 (PR #12304, transcript-replay timing) added a telemetryConstants entry to that file's paths map; this PR added shellResult at the same anchor, after transcriptRecords.

Textual, not semantic — adjacent insertions, no shared logic. Union (both targets exist; JSONC parses):

"@qwen-code/qwen-code-core/shellResult": ["../packages/core/src/utils/shell-result.ts"],
"@qwen-code/qwen-code-core/telemetryConstants": ["../packages/core/src/telemetry/constants.ts"],

Load-bearing.

• Both explicit entries must stay. The map ends in a @qwen-code/qwen-code-core/* wildcard, so dropping one does not fail the typecheck — it silently resolves against dist instead of source. A future "dedupe" here is a break. Order is irrelevant.
• The same union holds in the sibling registries (core exports, cli/sdk tsconfigs and vitest aliases, text-capture.tsx), all auto-merged with both keys and verified. acp-bridge's carry only telemetryConstants, correctly: the PR's new test there adds no imports. Otherwise the merge is an exact union both ways; only the conflicted file was edited.

Not verified — no build/typecheck/lint/tests run. All three sit in NON-conflicted files, so any break is CI's:

1. The PR widens ToolResultDisplay with ShellResultDisplay. feat(acp-bridge): surface request and tool timing on paged transcript replay #12304's acp-bridge/src/transcript-replay.ts is the only base change touching resultDisplay in production code; it types it unknown, adds 261 pure insertions and never rewrites the pass-through, so no new consumer assumes string.
2. core/src/tools/shell.test.ts: both sides added tests to one suite — main's per-shell description/schema length budgets vs the PR's structured returnDisplay. The PR touches neither description, so budgets should hold; unconfirmed.
3. The PR's new acp-bridge test asserts toHaveLength(1); main's machine now emits timing frames, but only for ui_telemetry records, and that test feeds one tool_result.

Maintainer re-verification (round 2) — PR #12311 @ ac01af9d1d

Re-ran the same real-daemon rig against the updated head (ac01af9d1d, merge-base b85264be79; the earlier round was f492336b26 on be9ed41afd). Both worktrees were reinstalled from the lockfile and rebuilt from scratch; arm A = this head, arm B = b85264be79 (the PR's own merge-base, so the A/B is fair).

All five findings from the previous round are fixed and pinned by tests. I found no fix-induced regression in any consumer. Two small export-only nits remain from the F1 fix, and the deferred-compatibility note in the description overstates the real impact — my measured number is much smaller than the one quoted.

Findings from round 1 — re-verified on a real daemon

I also mutated the three new fixes to confirm the new tests actually pin them, rather than just passing:

Two small nits left over from the F1 fix (export only)

N1 — the command is printed twice in an exported document, three times on a failure. The new documentMode && !result branch renders <pre>{command}</pre> above the text fallback, but the collapsed row header already shows the command, and on a failure the fallback text itself opens with Command: …. docs/design/shell-command-card.md opens by naming command repetition as the problem this PR fixes, so it is worth closing the loop — dropping the extra <pre> (the header already carries it) or restoring getToolDescription(tool) there, as the base build showed, would do it.

N2 — a successful command with no output now exports as a card containing only the command. true → base exported Command: true / Directory: (root) / Output: (empty) / Exit Code: 0 / …; this head exports just true and nothing else. The live card correctly says No output; the export says nothing at all, so a reader cannot tell "ran and produced nothing" from "we lost the result". grep exit 1 is fine (Command exited with code: 1 still carries the status).

Neither blocks merge; both are one-line changes in the same branch you just added.

The deferred ACP-budget note is much smaller than stated

The description says "the structured ACP text/output fields share the 64KB budget, so daemon CLI long display text can be truncated earlier than the former string result (40,021 characters became 32,667 in the targeted projection check)". Measured end-to-end on the real daemon, same 204,000-char stdout, both arms:

So the display-text loss is 100 characters (0.3 %), not ~18 %: an upstream 32,000-char display cap already binds before the ACP budget does, and raising tools.truncateToolOutputThreshold to 200,000 changed neither arm. The real cost is the other half of that table — because text and output are near-duplicates for a successful command (text is the output plus the "saved to" line), the ACP frame for a large result doubles, from 32,613 B to exactly the full 65,536 B budget. That is worth a sentence in the description in place of the current one; if you ever want it back, giving output priority in allocatePayloadBudgets rather than an equal share would be the lever.

The process-registry.ts:773 typecheck failure is a local tree artifact

It does not reproduce here and it does not reproduce in CI:

• npm run typecheck (all workspaces + typecheck:integration) on Node 24.18.1: exit 0, zero error TS.
• tsc -p integration-tests/tsconfig.json on Node 22.23.2: exit 0.
• CI runs npm run typecheck:integration inside Integration Tests (no-AK, No Sandbox) on a Node 22.x runner — pass on this head.

Line 773 is return result.stdout in a string-returning helper. spawnSync(...).stdout only widens to NonSharedBuffer when @types/node ≥ 22 is resolved at that import; this lockfile pins @types/node 20.19.1 at the root (TypeScript 5.8.3), which is what CI and both of my trees resolve. Your tree is almost certainly hoisting a newer @types/node. Nothing to fix in the PR — I'd just drop the caveat from the description.

Regression sweep on this head

The live/messages.test.ts failure I reported last round is gone: #12316 and #12319 fixed it on main, and every CI job on this head now passes except the still-running review-pr.

Verdict

The corrections are real, minimal and well pinned, and nothing else moved. From my side this is ready to merge; N1/N2 are cosmetic and export-only, and the two description sentences above are worth a touch-up. One housekeeping note: the branch is now 5 commits behind main (dfdbbace1f) again — no conflicts expected, just re-run the gate after the next merge.

Round-2 evidence (same rig; "before the fix" images are this PR at f492336b26):

F3 — exit code suppressed for timeout and cancellation

F1 — exported document, before and after the fix (N1 visible: the command appears twice)

N2 — grep exit 1 and a no-output success in the export, base vs this head

⚠️ AutoFix round 5 ended without publishing a report — view run.

🤖 AutoFix ran out of time before finishing (timeout (7200000ms)) (attempt 1/100) — it will retry on the next scan.

What I found before stopping:
Qwen failed during address-review: timeout (7200000ms).

See the Qwen Autofix agent step logs for model/tool output.

Run log: https://github.com/QwenLM/qwen-code/actions/runs/35520279908

🧠 Handled by Qwen Code · model/模型 kimi-k3 · CLI 0.24.1

🤖 Addressed the latest review feedback (round 3/100). What changed, and what I pushed back on: · 已处理最新评审反馈（第 3/100 轮）。改动内容与我反驳保留之处如下：

Address-review round — PR #12311

Feedback triage

• Reviews / inline comments / issue-level comments: none new since the last evaluation.
• Failed checks — "Signal the reviewed fork PR: CANCELLED" (×21): workflow signal jobs cancelled by the harness, not code failures on this branch. Nothing actionable in the PR's code.
• Same-run verification repair — "core rebuild failed on the agent-committed fix": the gate's scoped npm run build --workspace packages/core died in copyBrowserUseAssets because packages/browser-use/dist did not exist in the verify environment.

Root cause of the deterministic rejection

The rejection was environmental, not caused by the rejected commit (a test-only change): the review-address job installs with QWEN_SKIP_PREPARE=1 and restores only the repo-root dist/ and packages/core/dist artifacts, so packages/browser-use/dist is never present there. Core's build (scripts/build_package.js → copy_files.js) unconditionally stages the browser-use runtime when the bundled skill exists, so any scoped core build without that runtime fails. The dependency was expressed only through the root scripts/build.js ordering (browser-use builds first), which a scoped --workspace packages/core invocation never traverses. Reproduced locally before the fix: identical Browser-use runtime not found failure with packages/browser-use/dist absent.

Fix (commit d23096e)

packages/core/package.json gains a one-line prebuild that builds @qwen-code/browser-use only when its dist/index.js sentinel is missing. Scoped core builds are now self-sufficient; the root build still builds browser-use first, so the prebuild is a no-op skip on that path and adds no cost there. One file, one line; the manifest-scripts area class is already within this PR's own footprint (the PR earlier added the ./shellResult export to the same manifest).

Mutation probe (build wiring — no unit-test surface applies)

• Guard absent + runtime absent → scoped core build fails (the gate rejection, reproduced locally pre-fix).
• Guard present + runtime absent → scoped core build passes; the fallback staged packages/browser-use/dist.
• Guard present + runtime present → scoped core build passes on the skip path.
• No test files touched; nothing deleted, skipped, or weakened.

Verification

• npm run build --workspace packages/core with packages/browser-use/dist removed — passed (the gate's exact failing command).
• npm run build --workspace packages/core with packages/browser-use/dist present — passed (skip path).
• npm run build (full) — passed, exit 0; working tree stayed clean apart from the intended manifest change.
• npm run typecheck — passed, exit 0.
• npm run lint — passed, exit 0.
• npx prettier --check packages/core/package.json — passed.
• Core package tests, mirroring the gate (npm run test --workspace packages/core -- --changed origin/main --passWithNoTests with the gate's vitest load clamps) under a gate-equivalent environment (env -i, writable HOME, no SANDBOX/QWEN_* variables): 29,295 passed, 10 skipped, 0 failed across 772 files.
  • Note: the same command fails inside this agent sandbox (155 failures in workflow-snapshot, config, agent-headless, editor, and four other files) because the sandbox provides a read-only $HOME and exports SANDBOX/QWEN_HOME; none of those files' subjects are in the PR diff, all pass with a writable clean environment, and the gate's env -i allowlist strips those variables on the runner.

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

Re-review when you have a moment. After round 100 this bot stops and leaves the PR for a human. · 有空请复审；第 100 轮后本 bot 停止并将 PR 交给人工。

🧠 Handled by Qwen Code · model/模型 kimi-k3 · CLI 0.24.2

@qwen-code /resolve

Qwen Code did not run conflict resolution for this request.

PR #12311 does not currently have merge conflicts with main.

Maintainer verification (round 3) — PR #12311 @ d23096e857

Re-verified the head that the three post-round-2 commits produced (69a6f7cd4a → af4e1217f1 → d23096e857), merge-base b85264be79 (unchanged since round 2). macOS 15 / Node 24.18.1 / pnpm 11.24.0.

Rig. Three git worktrees — main b85264be79, previous head ac01af9d1d, current head d23096e857 — each fully built and each running its own production daemon self-hosting its own packages/web-shell/dist (node packages/cli/dist/index.js serve --web), with an isolated QWEN_HOME and a scripted fake OpenAI endpoint that returns a deterministic run_shell_command per scenario marker. Real shell processes, real ACP/SSE transport, real Playwright captures, real export endpoint rendered offline against each arm's own export-transcript-document.{js,css}. Pre-PR ("legacy") records were produced by the main daemon and then replayed by the PR daemons from a copy of the same QWEN_HOME.

Verdict: mergeable. Everything the three new commits claim reproduces independently. Two new Minor issues, both cosmetic/completeness rather than correctness, one with a verified one-word patch. Nothing blocking.

1. Round-3 fixes — reproduced

Control/bidi escaping (R1-8 / R1-38). A real command whose argument carries U+202E … U+202C plus BEL. On ac01af9d1d the card renders the override live, so gnp.exe is displayed as exe.png — a textbook trojan-source spoof. On d23096e857 it renders as inert \u202e…\u202c.

Clipboard stays byte-exact. Mutating either copy handler to copy the sanitized string is killed by the suite (both mutants), so the "escape for display, copy raw bytes" contract is pinned.

N1 / N2 (my round-2 nits) — both fixed, confirmed in a rendered export. Left: ac01af9d1d repeats the command above the fallback (3 occurrences per turn: header, <pre>, and the envelope's own Command: line) and exports a successful-but-silent run as a bare command line. Right: d23096e857 prints it twice (header + envelope) and exports No output.

String rawOutput display (R1-50). Background-start now shows the human display string instead of the model-facing blob (id / pid / output-file / status-file paths).

Unadvertised win: pre-PR history renders better. Replaying a main-produced session through the PR daemon, the same persisted record stops dumping the whole Command:/Directory:/Output:/Exit Code:/PGID envelope into Output and shows the command once plus clean output. This is the single biggest visible improvement for existing transcripts and it is not called out in the PR description.

hermeticEnv() (af4e1217f1). Reproduced the gate failure rather than taking it on trust: with HOME poisoned by remote.pushDefault=gd, the pre-fix tree fails exactly 3/107 with fatal: 'gd' does not appear to be a git repository; the fixed tree is 107/107 in both the poisoned and the clean environment.

prebuild (d23096e857). Full matrix re-run locally, including the case the commit message does not cover — the repo's declared package manager:

Root npm run build also confirms the skip path: browser-use builds first, then core's prebuild no-ops.

2. New findings

F1 (Minor) — a pre-PR record whose display string is empty falls back to the model-facing envelope.
ShellToolOutput.tsx:43 gates the new string branch on typeof raw === 'string' && raw, so an empty display string ('', which is what a silent command persists) is treated as "no string" and the card falls back to extractText(tool) — the envelope. The new legacyOutputRepeatsCommand guard then also suppresses the Command panel, because the envelope starts with Command: <cmd>. Net effect: in one transcript, sibling cards render clean output while the silent one renders a raw envelope with no Command section and no copy button.

Verified candidate patch — one word, A/B/A on the running daemon:

-    typeof raw === 'string' && raw
+    typeof raw === 'string'

The focused suites (145 tests) stay green both with and without the patch, so neither behaviour is pinned.

F2 (Minor) — the new finished-command elapsed is live-only; it disappears on reload.
Same session, same command: while the page stays open the card reads Succeeded 3s; after F5 it reads Succeeded. Cause is not in the component — the daemon's compacted replay emits one tool_call event per tool, already in its terminal status (a 6-tool session replays as exactly 6 events, zero in_progress), so the client has no start timestamp and formatElapsed returns ''. Not a regression (main never showed elapsed at all), but R1-39 is only half delivered, and the displayed duration is inconsistent between a live and a reloaded view of the same run.

F3 (informational, not introduced here) — the trojan-source hardening covers the expanded card only.
The collapsed tool summary — the line a user sees before expanding anything, and the header line in exported documents — still renders model-supplied text with the raw override. Byte-identical on both arms, so this PR did not introduce it; flagging it because it is the surface the escaping was meant to protect.

F4 (Minor, coverage) — 4 of the 6 new sanitizeControlChars call sites are unpinned.
12 targeted mutants against the round-3 code, 8 killed / 4 survived:

3. Regression & gate status on this head

• Full packages/web-shell suite, same machine, both arms: head 349 files / 9207 tests, 0 failed; main b85264be79 348 / 9171, 0 failed. Round 2's 2 failures and main's own live/messages.test.ts red are both gone.
• npm run build (root) exit 0; npm run typecheck (all workspaces + typecheck:integration) exit 0 on Node 24.18.1 — the description's remaining process-registry.ts:773 note still does not reproduce here.
• eslint + prettier --check clean on all five files changed since ac01af9d1d.
• All PR checks green; mergeStateStatus is BLOCKED only on REVIEW_REQUIRED.
• Cancel/timeout headlines stay correct: a SIGTERM-killed run reads Failed / Signal 15 with no synthetic exit code, where main shows Exit Code: 0 next to Signal: 15.

4. Checked and not reproduced

• R1-30 (an empty shell display no longer collapsing to {type:'none'} in the TUI): drove the real TUI through @lydell/node-pty on both arms with a silent command. Final frames are the same — ✓ Shell true (Produce no output), one line, no empty box. I could not produce the symptom.

5. Still open, for the record

The autofix bot's round-2 comment defers 21 of its own review findings as "real and scoped". Two of them are visible in this run and worth a follow-up rather than a block:

• R1-1 — exported documents of failed runs still render the model-facing envelope (Command:/Directory:/Exit Code:), while the live card for the same run shows clean output. A live/export inconsistency, not a correctness bug.
• R1-31 — measured end to end in round 2: the display-text cost is ~0.3 %, not the 40,021→32,667 the description implies; the real cost is the ACP frame going to the full 64 KiB because text and output are near-duplicates.

Recommendation: merge. F1 is a one-word change worth taking first since it is already verified; F2, F3 and F4 are follow-ups.

-    typeof raw === 'string' && raw
+    typeof raw === 'string'

Evidence: figures at wenshao/qwen-code@236dbbb06c524add571be5a6740a2aceb71277af under pr12311-r3/. Rounds 1–2 reports: #issuecomment-5748505797, #issuecomment-5750340118.

@qwen-code /triage

🤖 Addressed the latest review feedback (round 4/100). What changed, and what I pushed back on: · 已处理最新评审反馈（第 4/100 轮）。改动内容与我反驳保留之处如下：

Autofix round — PR #12311 @ 22f6db30d5

Round-3 maintainer verification (@wenshao, [ic:5754429976]) reported the PR
mergeable with one recommended in-PR fix (F1) and three follow-ups (F2–F4).

F1 (Minor) — empty display string fell back to the model-facing envelope — FIXED

A pre-PR record of a silent command persists an empty display string
(rawOutput: ''). ShellToolOutput.tsx gated the string branch on
typeof raw === 'string' && raw, so the empty string was treated as "no
string" and the card fell back to extractText(tool) — the
Command:/Directory:/Output:/Exit Code: envelope — and the
legacyOutputRepeatsCommand guard then also suppressed the Command panel,
leaving a bare envelope with no copy button.

Reproduced before fixing with a new focused test
(ShellToolOutput.test.tsx › treats an empty string rawOutput as authoritative empty output): on the pre-fix code the card rendered the
envelope (Directory: (root) present, no Copy command button) — the test
failed exactly as the report describes.

Fix (the maintainer's verified one-word patch): gate on the type alone,
typeof raw === 'string'. A persisted empty display string now renders the
Command section plus a No output label, matching its sibling cards. The
new test passes post-fix; since it fails with the pre-fix expression and
passes with the fix, it pins the branch (mutation probe: removing the fix
reverts the test to red).

Changes: ShellToolOutput.tsx (+1/−1), ShellToolOutput.test.tsx (+13).

F2 (Minor) — finished-command elapsed is live-only — deferred to follow-up

Root cause is not in the component: the daemon's compacted replay emits one
terminal tool_call event per tool with no in_progress/start timestamp,
so after reload formatElapsed has no inputs. The fix belongs in the replay
path (packages/acp-bridge/src/transcript-replay.ts), outside this PR's
web-shell footprint, and the maintainer explicitly assigned it as a
follow-up. Recorded in deferred-findings.json.

F3 (informational) — collapsed summary renders model text unsanitized — deferred

Byte-identical on main and this head; this PR did not introduce it. The
collapsed row (ToolGroup.tsx) and export headers are a separate surface
from the expanded card this PR hardens. Recorded in
deferred-findings.json per the maintainer's follow-up assignment.

F4 (Minor, coverage) — 4 of 6 new sanitize call sites unpinned — deferred

The surviving mutants (document fallback, notices, result.directory,
result.outputFiles rendering raw) are a coverage gap on round-3 code, not
a defect; the maintainer explicitly assigned it as a follow-up. Recorded in
deferred-findings.json.

Notes

• R1-1 and R1-31 remain tracked in the deferred-findings issue from round 2;
  the maintainer confirmed both as follow-ups, not blockers.
• The maintainer observed that this PR also improves rendering of pre-PR
  (legacy) transcript records — the biggest visible win for existing
  sessions — and that it is not called out in the PR description. Flagging
  for the author; the PR body is not this round's to edit.
• The second issue-level comment ([ic:5754462890]) is a CI-bot status note
  (sandboxed verification running); no action.
• No review threads are resolved by this round (the feedback arrived as one
  issue-level comment), and no inline replies are owed.

Verification

All commands were actually run locally on this head (22f6db30d5). The
sandbox caps any single command at ~120s and reaps detached processes, so
the monorepo-wide commands were run as equivalent per-package sequences in
dependency order; the deterministic gate re-runs the full commands post-push.

• Reproduction probe (pre-fix): npx vitest run client/components/messages/tools/ShellToolOutput.test.tsx in packages/web-shell — 1 failed / 35 passed, failing exactly on the new F1 test (envelope rendered, no copy button), confirming the defect.
• Post-fix: same command — 36/36 passed.
• Focused related suites: ToolGroup.test.tsx, transcriptToMessages.test.ts, toolFormatting.test.ts — 389 passed.
• Full packages/web-shell suite via 6 shards (npx vitest run --shard=i/6): 8229 passed, 0 failed across 349 files — except App.test.tsx, which alone exceeds the sandbox's 120s command cap and could not be run locally. It never uses an empty-string rawOutput (verified by grep), so the change cannot affect it; the maintainer's run of the same suite on this head was fully green (349 files / 9207 tests).
• npm run build equivalent — every workspace built in the scripts/build.js order (generate, browser-use, core, channels, audio-capture, node-repl, acp-bridge, sdk-typescript, web-shell, web-templates, cli, generate-settings-schema, qwen-live, vscode-ide-companion, chrome-extension, both integrations) — all exit 0, with web-shell/web-templates/cli rebuilt after the fix.
• npm run typecheck equivalent — tsc in every workspace with a typecheck script (acp-bridge, audio-capture, browser-use, chrome-extension, cli, core, node-repl, qwen-live, sdk-typescript, web-shell, both integrations) plus npm run typecheck:integration — all exit 0.
• npm run lint equivalent — eslint --ext .ts,.tsx over packages/*, packages/channels, integrations, scripts, integration-tests, root vitest.config.ts and eslint-rules (patches/ is globally ignored by the eslint config) — all exit 0.
• npx prettier --check on both changed files — clean.
• Settings schema untouched (no settings source changed); generate-settings-schema run as part of the build sequence above — exit 0.

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

Re-review when you have a moment. After round 100 this bot stops and leaves the PR for a human. · 有空请复审；第 100 轮后本 bot 停止并将 PR 交给人工。

🧠 Handled by Qwen Code · model/模型 kimi-k3 · CLI 0.24.2

@qwen-code /triage

🤖 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：

Autofix round: no actionable feedback

No code changes were made in this round.

What was reviewed

• Reviews / inline comments: none newer than the last evaluation (2026-09-21T02:04:15Z).
• Issue-level comments: the only new item is an automated status comment from @qwen-code-ci-bot announcing that sandboxed verification is running. It is a progress notification from the verification workflow, not a request for changes, so there is nothing to address in code.
• Failed checks: none. All 25 checks on the current head are green (22 SUCCESS, 3 SKIPPED).

Outcome

With no review feedback, no inline findings, and no failing checks, there is no in-scope work for this round. The branch is left untouched and nothing was committed.

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 kimi-k3 · CLI 0.24.2

Addressed the latest review's two code findings and verified the failure-path compatibility end to end through the scheduler and hook handler.

• R1-1: The producer now always retains its original human-facing display text. Tests cover exit 3 with/without output and signal termination. The scheduler intentionally retains its final error-text override: the pre-PR createErrorResponse already used resultDisplay ?? error.message, so removing that override would change terminal error details and failure-hook context. Both design languages now distinguish producer text from the scheduler error fallback.
• R1-4: recorder_output now projects structured display text with the same helper as recorder_input. Tests cover empty, small and compacted 40,000-character inputs, comparing the output diagnostic with the actual saved display. This fixes diagnostic visibility; persisted tool results were not lost.
• Correction to the sandbox report's Finding A: The claimed failed-PostToolBatch undefined → string change is not supported by the complete path. Base 1cc63cfe already has resultDisplay: resultDisplay ?? error.message. Independently captured actual scheduler batch payloads for legacy and structured results, then passed those captures through the real hook handler: both produce the same error string. Feeding undefined directly to the handler bypasses the scheduler fallback. Added legacy/structured scheduler and failed-batch normalization coverage rather than changing compatible behavior.
• Documented that background acknowledgments and pre-execution cancellation retain their string fallback. Previously deferred suggestions remain deferred.

Validation: 1,156 tests passed across the four affected core suites; 4 independent boundary checks passed. Temporarily restoring both old production implementations made all 6 targeted regression cases fail; restoring the fixes made all six pass. Full build and bundle passed; the bundled CLI starts and reports 0.24.2. All workspace typechecks passed. Root integration typecheck still reports the previously documented process-registry.ts:773 string/NonSharedBuffer error on local Node 22.14.0. ESLint, Prettier and diff checks passed. Two self-audit passes and an independent incremental review found no further defect.

Browser evidence: 1 Chromium command-card E2E passed, using the existing mock daemon. Success/failure screenshots were visually inspected, uploaded to the separate asset branch, and replace the earlier supplemental fixtures in the PR description. These are UI checks, not real shell/OS hook E2E or proof of the recorder fix. The independent boundary checks use mocked execution/filesystem/hook runner; the scheduler and hook handler under observation are actual implementation code.

中文：已修复生产端失败展示文本混入模型外壳，以及录制输出边界漏掉结构化展示文本的问题。调度器保留旧有错误详情与失败 hook 上下文。沙箱报告关于失败 PostToolBatch 从缺省变为字符串的结论漏看了基线中的 resultDisplay ?? error.message 默认值；完整 scheduler→序列化→hook 链路验证旧、新结果最终字符串一致，因此补兼容测试而不修改正确行为。已同步中英文设计说明和提前返回/后台场景范围。

验证：1,156 项测试、4 项独立边界检查、1 项使用模拟 daemon 的 Chromium E2E 通过；撤回修复时 6 项回归测试全部失败，恢复后全部通过。构建、bundle、workspace 类型检查、lint/格式通过；根集成类型检查仍有已记录的本地 process-registry:773 类型错误。成功/失败截图已上传独立素材分支并更新 PR，明确仅作为模拟 daemon 的 UI 证据。