feat(web-shell): improve session overview navigation and details - #11238
Conversation
E2E verification reportVerified on macOS with Chromium, Node.js 22.22.2 and an isolated daemon serving the actual locally built
Reproduction after building, with a local Web Shell server running: cd packages/web-shell
PLAYWRIGHT_BASE_URL=http://127.0.0.1:<port> npx playwright test --config playwright.config.ts client/e2e/web-shell.session-overview.spec.ts --project=chromiumWindows, Linux, other browsers and real model execution were not validated. 中文测试报告在 macOS、Chromium、Node.js 22.22.2 上验证，隔离 daemon 提供本地实际构建的
构建并启动本地 Web Shell 后，可用上方命令复现测试。未验证 Windows、Linux、其他浏览器和真实模型执行。 |
🖼️ web-shell visual previewRendered against a mock daemon (no real backend): the PR base vs this PR head Screenshots · before / afterFull-resolution recordings (.webm) are attached to the workflow run. — Qwen Code · web-shell visuals |
|
Thanks for the PR! Template looks good ✓ — and the design doc under Problem: real, and I checked it against Direction: aligned. Session-list state visibility and filter menus are somewhere the reference agents keep investing — Claude Code's CHANGELOG carries "Added Open and Closed to the session list's status filter menu", "Added a hollow ring in the session list for sessions open in a terminal … so they no longer look closed", and a fix for "a tab opened from the session list jumping back to that session". Status filters, distinct state indicators, and unambiguous row-click semantics are the same direction. Size: not applicable — no core paths touched ( Approach: close to what I'd have proposed independently — reuse the sidebar's
Risk: no elevated risk signals — none of the revert-correlated paths are touched. One thing worth knowing up front: the new Moving on to code review. 🔍 中文说明感谢贡献！ 模板完整 ✓ —— 问题： 真实存在，我对照 方向： 对齐。会话列表的状态可见性和筛选菜单是参考 agent 持续投入的方向——Claude Code 的 CHANGELOG 里有"Added Open and Closed to the session list's status filter menu"、"Added a hollow ring in the session list for sessions open in a terminal … so they no longer look closed"，以及"a tab opened from the session list jumping back to that session"的修复。状态筛选、可区分的状态标识、明确的行点击语义都是同一方向。 规模： 不适用——未触及核心路径（ 方案： 与我自己独立想到的做法很接近——复用侧栏的
风险： 无升级风险信号——未触及与回滚相关的文件路径。有一点先说明：新增的 进入代码审查 🔍 — Qwen Code · qwen3.8-max-2026-09-02 Reviewed at |
Code reviewI wrote down my own approach before opening the diff — reuse the sidebar popover rather than grow a second one in the overview, two-line session cell, checkbox-only selection, extend the existing search. The PR landed on essentially the same shape, and in two places it is better than what I had in mind: it reuses the 1. Blocker — the shared popover change breaks an existing
|
| File | What changed |
|---|---|
docs/design/web-shell-session-overview.md |
New design doc; states the problem, the reuse-the-sidebar-popover decision, and that no daemon or SDK surface changes. |
packages/web-shell/client/components/SessionOverviewPanel.tsx |
The bulk of it. Deletes the SessionIdCell and the git/sessionId/workspace columns, folds workspace, branch and PR under the title, adds a status column and a four-way status filter, moves the workspace filter into the toolbar, switches row click from select to open, and hides batch actions when nothing is selected. |
packages/web-shell/client/components/SessionOverviewPanel.module.css |
One new .attention rule colouring the two blocked states with the warning token. |
packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx |
Shared component: new openOnClick mode using PopoverTrigger instead of PopoverAnchor, full path instead of basename, wrapping instead of truncation, status derived from the waiting flags, wider collision boundary, and stopPropagation on portaled clicks. |
packages/web-shell/client/components/sidebar/SessionDetailsTooltip.test.tsx |
Adds the ref-forwarding, keyboard-dismissal and host-portal case for the new click trigger. |
packages/web-shell/client/components/SessionOverviewPanel.test.tsx |
Rewrites the tests that asserted the old columns and row-click semantics; adds status priority, filter and search reset, details entry, ID copy without navigation, and batch-action visibility. |
packages/web-shell/client/e2e/web-shell.session-overview.spec.ts |
New four-case Playwright spec over a mocked daemon. Not tagged @smoke, so it does not run on PR CI. |
packages/web-shell/client/i18n.tsx |
Twelve new keys per locale for the status column, filters, details label and session count; widens the search placeholder. Does not remove the three keys finding 3 orphans. |
Testing evidence
This is an unattended CI run, so I did not build or execute anything from this PR — the evidence below is the PR's own CI, read through the API for commit 965c72bd4d71550c71afc8a045c5ddd52f5fba96. Nothing has failed yet, but the two checks that matter most for this diff have not finished: Test (ubuntu-latest, Node 22.x) and Lint & Static (ubuntu-latest, Node 22.x) are both still in progress, and web-shell E2E Smoke has not been created because it is queued behind Test. That smoke job is the one finding 1 predicts will go red, so the CI signal for this PR is genuinely incomplete right now rather than green.
Capture web-shell visuals passed, which is mild reassurance that nothing throws at render, but that workflow captures screenshots and does not assert the truncation behaviour the smoke spec pins. Test (macos-latest) and Test (windows-latest) are skipped by the CI profile, so the unit suite is Linux-only on this PR.
Not verified, and why: the browser-level claims — popover placement inside the Web Shell boundary at constrained viewport sizes, focus returning to the details button on Escape, external PR and issue links opening a popup without closing the overview — cannot be settled from the diff, and the spec that would settle them is not tagged @smoke, so this PR's CI never runs it. The PR description's "four Chromium browser regression cases passed" and "146 focused unit tests passed" are the author's own local results; I have not re-run them and they are not evidence in this section.
Final CI results for 965c72b (auto-updated by the triage finalize job after CI completed):
| Check | Conclusion |
|---|---|
Lint & Static (ubuntu-latest, Node 22.x) |
❌ failure |
Test (ubuntu-latest, Node 22.x) |
🚫 cancelled |
web-shell E2E Smoke (ubuntu-latest, Node 22.x) |
🚫 cancelled |
Capture web-shell visuals (ubuntu-latest, Node 22.x) |
✅ success |
Classify PR |
✅ success |
Desktop Shell (ubuntu-22.04) |
✅ success |
Desktop Shell (windows-2022) |
✅ success |
Integration Tests (no-AK, No Sandbox) |
✅ 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。
Sandboxed verification would settle this: @qwen-code /verify — that row click navigates while checkbox click only selects, that the status priority holds against a real daemon rather than a mocked one, and that the details popover stays inside the Web Shell boundary at constrained sizes, are none of them observable from the diff, and the new Playwright spec that would demonstrate them never runs on this PR's CI because it is untagged.
中文说明
代码审查
我在看 diff 之前先写下了自己的方案——复用侧栏浮窗而不是在总览里再造一个、双行会话单元格、仅由复选框控制选择、扩展已有搜索。这个 PR 基本落在同一形状上，有两处比我想的更好：复用了已有的 SessionCardStatus 排序而不是另造状态机；并通过 DaemonSessionSummary 本来就定义的 isWaitingForPermission / isWaitingForUserQuestion 回传状态，因此不需要改动 SDK 接口。一个阻塞项和两个清理项。
1. 阻塞项——共享浮窗的改动破坏了本 PR 未修改的现有 @smoke 浏览器测试
web-shell.collapsed-groups-persist.spec.ts 里的 keeps long session details inside a constrained WebShell @smoke 会悬浮一个侧栏会话，其名称是 'longunbrokensessiontitle'.repeat(24)，并断言 details.getByTitle(longTitle) 存在且带 title 属性，同时断言标题是单行截断（clientHeight <= lineHeight + 1 且 scrollWidth > clientWidth）。
本 PR 去掉了 .sessionDetailsTitle 上的 title={label}，并用 !whitespace-normal break-words 覆盖了该类的 white-space: nowrap / text-overflow: ellipsis。两条断言都会失败：getByTitle(longTitle) 匹配到 0 个元素（我核对过 fixture，createOrganizedScenario 的这些会话没有 prs、branch、worktree，所以该浮窗内剩下的 title 只有目录路径和会话 ID）；换行后的标题是多行且没有横向溢出。
这不是可以忽略的过期测试——它跑在本 PR 自己的 CI 上。ci.yml 的 web-shell E2E Smoke (ubuntu-latest, Node 22.x) 执行 npm run test:e2e:smoke（--grep @smoke），它 needs: test，而 test 仍在进行中，所以这个 check 还没被创建。预计会红。
解法是一个设计决定，而不是改测试。要么侧栏保留省略号截断、只在总览换行——这正是共享浮窗上表现层 prop 的用途，而 worktreeOnly（见下）已经是这个接缝；要么认定换行在两处都正确，并在本 PR 内更新该 smoke spec，明确锁定新行为。不应该发生的是：合并时让一个被 diff 悄悄失效的 gating check 留在那里。
2. worktreeOnly 变成死开关
它唯一的调用方是本 PR 删掉的总览 Git 列；prop、默认值和三处读取点都还留在 SessionDetailsTooltip.tsx。改动之后 align 永远是 'start'，两个 !worktreeOnly && 分支永远渲染——正是 AGENTS.md 审查规则要求 grep 的"声明并读取但没有任何调用方设置"的可选标志。删掉它，或者把这个接缝用于第 1 点里的侧栏/总览分化。
3. 三个 i18n key 成为孤儿
sessionsOverview.sessionId、sessionsOverview.folder、sessionsOverview.worktree 只有三处调用，全在被删除的列头里。新增的 e2e spec 匹配的是字面量 'Session ID' 而不是 key，所以现在没有读取方——中英文共 6 条。
已核对且没问题的部分：合成的 session 对象能通过类型检查（DaemonSessionSummary 只要求 sessionId 和 workspaceCwd，其余含两个 waiting 标志均为可选）；碰撞边界改动比看起来窄（helper 本来就返回 closest('[data-web-shell-root]') ?? sidebar，侧栏结果不变，只是让不在 aside 内的总览第一次有了边界）；PopoverContent 上的 stopPropagation 是必需的而非装饰（浮窗走 portal，点击会沿 React 树冒泡到行的 onClick）；复选框不会触发导航（select 列本来就有 stopRowClick: true）；筛选与搜索变更会同时重置页码和选择（effect 依赖数组已补上 statusFilter）；单测确实在锁定行为而非复述实现。
非阻塞小点：各筛选按钮的计数在 JSX 里内联调用 searchedCards.filter(...)，每次渲染跑四遍——50 行时无所谓，目录变大后值得 useMemo；浮窗的 idle 状态把两个独立翻译的字符串拼在一起（Idle · 3 clients），无法按语言调整语序。
测试证据
这是无人值守的 CI 运行，因此我没有构建或执行本 PR 的任何代码——以上证据来自 API 读取的本 PR 自身 CI（commit 965c72bd4d71550c71afc8a045c5ddd52f5fba96）。目前没有失败项，但对这份 diff 最关键的 check 还没跑完，因此本 PR 的 CI 信号现在是"不完整"而不是"绿"。
未验证的部分及原因：浏览器层面的断言（受限视口下浮窗是否留在 Web Shell 边界内、Escape 后焦点是否回到详情按钮、外部 PR/Issue 链接是否在打开弹窗的同时保留总览）无法从 diff 判定，而能判定它们的 spec 没有 @smoke 标签，本 PR 的 CI 不会运行它。PR 描述里的"四个 Chromium 浏览器回归用例通过""146 个聚焦单测通过"是作者本地的结果，我没有重跑，也不作为本节的证据。
沙箱验证可以定这件事：@qwen-code /verify——行点击导航而复选框只选择、状态优先级对真实 daemon（而非 mock）成立、详情浮窗在受限尺寸下留在 Web Shell 边界内，这三点都无法从 diff 看出，而本可证明它们的新 Playwright spec 因为没打标签，在本 PR 的 CI 上从不运行。
— Qwen Code · qwen3.8-max-2026-09-02
Reviewed at 965c72bd4d71550c71afc8a045c5ddd52f5fba96 · re-run with @qwen-code /triage
|
Confidence: 2/5 — the design and the tests are good and I'd merge the overview half today, but the diff silently invalidates an existing browser test that runs on this PR's own CI, and that needs resolving first. @wenshao Requesting changes on one item only — the rest of this is genuinely well put together. I went looking for whether the problem is real rather than accepting the framing, and it is: on The approach matched what I'd have proposed independently, and beat it in two places — deriving the column from the The blocker reads as an oversight at a seam, not as carelessness, which is why this is request-changes rather than a rethink. The diff reaches beyond the overview into a popover the sidebar also renders, and the shared-component work is otherwise careful — the collision-boundary edit is narrower than it looks and is correct, One correction to how I described that in the review comment, because it changes what a maintainer should do about it. I called the smoke job a gating check. Not blocking, and not the reason for the verdict: the now-dead I'm not requesting changes because the scope felt ambitious or because the PR is large. It is neither — the scope is right for the stated goal and the size is mostly tests. It is one specific, verified, deterministic breakage with two clean ways out, both named in the review comment. 中文说明Confidence: 2/5 —— 设计和测试都不错，总览这部分我今天就可以合；但这份 diff 悄悄让一个跑在本 PR 自己 CI 上的现有浏览器测试失效，需要先解决。 @wenshao 只为一项请求修改——其余部分确实做得很好。 我去核实了问题是否真实存在，而不是采信 PR 的表述：在 方案与我自己独立想到的做法一致，有两处还更好——基于已有的 这个阻塞项看起来是接缝处的疏漏，而不是草率，所以是"请求修改"而不是"重新设计"。diff 越过总览改到了侧栏同样渲染的浮窗，而共享组件这部分其余工作都很细致——碰撞边界的改动比看起来窄而且是正确的，portal 内容上的 对我在审查评论里的一个说法做更正，因为它会改变维护者该采取的行动。我把那个 smoke job 称作 gating check。实际上 不阻塞、也不是本次结论的原因：已成死开关的 我请求修改，不是因为范围显得过大或 PR 太大。两者都不是——范围与所述目标相称，体量主要在测试。原因是一个具体、已验证、确定性的破坏，且在审查评论里给出了两条干净的解法。 — Qwen Code · qwen3.8-max-2026-09-02 Reviewed at |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
One blocker before this can merge.
The shared session-details popover now wraps long titles and drops its title attribute. That breaks the existing @smoke browser test keeps long session details inside a constrained WebShell @smoke in packages/web-shell/client/e2e/web-shell.collapsed-groups-persist.spec.ts, which pins single-line ellipsized truncation and matches the title via getByTitle. That spec is not in this diff, and it does run in this PR's CI through the web-shell E2E Smoke job (queued behind Test, so it hasn't reported yet).
Two clean ways out, both written up in my review comment: keep truncation in the sidebar and wrap only in the overview — worktreeOnly is already that seam — or decide wrapping is right everywhere and update the smoke spec in this PR. Worth knowing that main's ruleset has no required-status-check rule, so CI won't mechanically stop the merge if this is left as is.
Two cleanups can ride along with the fix: worktreeOnly is now a dead switch (its only caller was the deleted Git column), and sessionsOverview.sessionId / .folder / .worktree are orphaned in both locales.
Everything else looks good — the design doc, the reuse of the existing status ranking and DaemonSessionSummary fields, and unit tests that genuinely pin the row-click and status-priority changes. 🙏
— Qwen Code · qwen3.8-max-2026-09-02
CI fix verificationFixed the constrained session-details regression reported by CI. The smoke test now checks complete wrapped titles instead of the removed native title attribute; the popover stays within the available height and scrolls to keep metadata and the copy action accessible. The four overview regression cases are now included in PR smoke runs.
The bundled-page run used an isolated local daemon with deterministic session responses. Remote CI is rerunning for this fix; these results are local verification. 已修复 CI 报出的会话详情回归：测试改为验证完整标题换行，浮窗限制在可用高度内，超长内容可滚动查看并复制会话 ID。四项总览回归已加入 PR 的 smoke 测试范围。 整库 build、typecheck、bundle、146 项定向单元测试和本地全部 54 项 Chromium smoke 均通过。另通过本地打包 CLI 运行了五项浏览器测试，确认受限视口中的标题完整性、浮窗边界、滚动复制，以及总览筛选、悬浮详情、键盘焦点和行导航行为。ESLint、Prettier、diff 检查也通过。远端 CI 将重新验证该修复。 |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed. Suggestions are inline.
2 Suggestion-level finding(s) this review confirmed are already reported on this PR and are not repeated:
- D1-1 packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx:34 worktreeOnly dead switch — already reported (comment 5564498785; review 5127758460)
- D1-2 packages/web-shell/client/i18n.tsx:3096/3098/3100 and 6477/6479/6481 orphaned sessionsOverview.sessionId / .folder / .worktree keys — already reported (comment 5564498785; review 5127758460)
Not reviewed: reverse audit — stopped before round 3 by the review time budget.
中文说明
仅完成部分审查，审查缺口已披露。 建议见行内评论。
本轮确认的 2 条建议级发现已在 PR 上报告过，不再重复发布（列表见上方英文部分）。
未审查：反向审计——评审时间预算不足，未能开始第 3 轮。
— qwen3.8-max via Qwen Code /review (v0.23.0)
Review follow-up and E2E reportAddressed the 15 inline review threads in Details now replace one another without stealing focus, close when their row leaves the visible filter/page, keep the focused copy action visible, and wrap long branches as well as titles, paths and IDs. Dragging text does not navigate; a plain-cell click during rename preserves the draft. Compact metadata and status have recoverable full text, and the workspace filter exposes its visible state in its accessible name. Before committing, reverse audits found and fixed keyboard-selection, anchor-remount, stale-open-state and focus-restoration issues. The final unchanged code then completed two consecutive clean reverse-audit passes, including an independent review and verification against the actual bundled Web Shell. The commit hook made no changes to that audited tree. Validation on macOS with Chromium and deterministic daemon responses:
GitHub reports the PR as mergeable with respect to conflicts. Fresh CI is running for this commit; these results describe completed local verification. Real model execution and non-Chromium browsers remain outside this UI verification. 中文说明已在 提交前多轮反向审计发现并修复了文字选择影响键盘激活、悬浮锚点重新挂载、详情状态残留和旧浮窗抢回焦点等问题；最终代码连续两轮没有发现新问题，且通过实际打包页面的独立验证后才提交。提交钩子没有改变审计过的代码。 整库 build、typecheck、bundle，208 项定向单元测试，56 项 smoke，打包页面的八项浏览器回归，以及深浅色两项截图测试均通过。PR 原有 wenshao fork 图床配图保留。GitHub 已确认没有合并冲突，新一轮 CI 正在运行。 |
|
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 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. |
|
@qwen-code /takeover |
qwen-code-dev-bot
added
the
autofix/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. Remove the 中文说明🤝 已接管：autofix 循环现在管理此 PR —— 将持续处理新的评审反馈与 base 冲突，直到移除标签或达到轮次上限。移除 |
CI follow-up and reverse-audit reportFixed the remaining session-details replacement smoke failure in The test now verifies the clicked session, waits for the intended hover session and outgoing dialog removal, then confirms the new dialog remains open before checking dismissal. Real animations and existing timeouts are preserved. Only this browser test changed. The remote branch advanced with another main merge during verification; this commit preserves that merge, and the build, typecheck, unit, smoke and bundled-page checks below were rerun on the merged version before committing. Reverse auditing found an additional false-positive window: uniqueness alone could accept a target already closing after focus theft. The final open-state assertion rejects that case. After this correction, the unchanged final diff completed two consecutive clean reverse-audit passes before commit. Commit hooks preserved the exact audited tree. Validation:
All 15 previous inline threads remain resolved, with no new unresolved thread. The latest bot comment reports its review workflow timeout and contains no new code finding. Existing PR screenshots on the wenshao fork asset branch are preserved. These are completed local results; fresh remote CI is verifying this commit. 中文说明已在 反向审计进一步发现并补上了焦点回抢可能导致测试错误通过的窗口，最终代码在提交前连续两轮未发现新问题，提交钩子未改变审计过的代码。整库 build、typecheck、bundle、208 项单元测试、关闭重试的 CI 模式下全部 59 项 smoke，以及实际打包页面的 8 项回归均通过。正常和延长动画通过；新浮窗未打开、旧浮窗未退出、旧浮窗抢回焦点三类故障均被识别。 15 条原有行内评论仍全部解决，没有新增未解决评论；最新机器人评论仅报告其审查流程超时。PR 中 wenshao fork 资产分支上的配图保留。以上是本地验证结果，远端 CI 正在验证新提交。 |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not explored to full depth (tool budget reached): chunk 10: none — no check was cut short. (I did not run the repo-wide npm run typecheck ; packages/web-shell/tsconfig.json excludes client/e2e/** and *.test.* , so …; "agent reverse-audit (round 1)": did not execute the new Playwright spec (it needs a web-shell build plus the dev-server webServer boot); every assertion above was validated statically agai…; "agent reverse-audit (round 1)": the new Playwright visual case in client/e2e/visuals/screenshots.spec.ts was walked statically only — this review worktree has no built bundle or browser, so ….
Not reviewed: reverse audit — stopped before round 5 by the review time budget.
Deferred under the convergence posture (round 2, not a blocker) — recorded, not requested in this round:
packages/web-shell/client/components/SessionOverviewPanel.tsx:74 — [review] Inert TooltipProvider and an unreachable, untested meta.tooltip branch in the shared data-tablepackages/web-shell/client/components/SessionOverviewPanel.test.tsx:582 — [probe] The filter/search reset test survives reducing the reset effect's dependency list to []packages/web-shell/client/components/SessionOverviewPanel.test.tsx:563 — [probe] A live status change silently shrinks a batch selection under any non-all filter, and no test can see the prunepackages/web-shell/client/components/SessionOverviewPanel.test.tsx:1140 — [probe] The rewritten footer test drops the base's data-variant assertion, leaving batch-button variants unpinnedpackages/web-shell/client/components/SessionOverviewPanel.test.tsx:520 — [probe] Substring search is unpinned on the title, session-ID and branch arms (a startsWith refactor ships green)
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）：chunk 10：none — no check was cut short. (I did not run the repo-wide npm run typecheck ; packages/web-shell/tsconfig.json excludes client/e2e/** and *.test.* , so …；"agent reverse-audit (round 1)"：did not execute the new Playwright spec (it needs a web-shell build plus the dev-server webServer boot); every assertion above was validated statically agai…；"agent reverse-audit (round 1)"：the new Playwright visual case in client/e2e/visuals/screenshots.spec.ts was walked statically only — this review worktree has no built bundle or browser, so …。
未审查：反向审计——评审时间预算不足，未能开始第 5 轮。
收敛姿态下延后（第 2 轮，非阻断）——已记录，本轮不要求修改：共 5 条（原文未翻译，列表见上方英文部分）。
— qwen3.8-max via Qwen Code /review (v0.23.0)
|
🤖 AutoFix ran out of time before finishing (timeout (7200000ms)) (attempt 2/100) — it will retry on the next scan.
See the Qwen Autofix agent step logs for model/tool output. 中文说明🤖 AutoFix 在完成前耗尽了时间（timeout (7200000ms)）（第 2/100 次尝试）—— 将在下次扫描时重试。 Run log: https://github.com/QwenLM/qwen-code/actions/runs/34231179206 🧠 Handled by Qwen Code · model/模型 |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
9 Suggestion-level finding(s) this review confirmed are already reported on this PR and are not repeated:
- overview visual fixture row-action coverage (e2e/visuals/screenshots.spec.ts:57) — already reported (round-3 deferral list, review 5137868769)
- batch-button data-variant coverage (SessionOverviewPanel.test.tsx:1397) — already reported (round-2 deferral list, review 5135733031)
- filter/search reset coverage (SessionOverviewPanel.test.tsx:587) — already reported (round-2 deferral list, review 5135733031)
- details-slot click-term witness (SessionOverviewPanel.test.tsx:810) — already reported (round-4 deferral list, review 5141881786)
- overview PR-number search matching (SessionOverviewPanel.tsx:535) — already reported (round-4 deferral list, review 5141881786)
- workspace span max-w cap coverage (SessionOverviewPanel.tsx:1184) — already reported (round-4 deferral list, review 5141881786)
- R3-4 status/cue parity loop coverage (SessionOverviewPanel.test.tsx:663) — still stands, already reported (round-3 inline thread)
- R4-2 rename draft retention on in-table controls (SessionOverviewPanel.tsx:1637) — still stands, already reported (round-4 inline thread; author reproduced and deferred to issue 11390)
- R4-3 restoreFocusRef reset witness (sidebar/SessionDetailsTooltip.tsx:155) — still stands, already reported (round-4 inline thread)
Not reviewed: test-efficacy probe — the harness control produced no verdict (harnessValidated null), so revert/mutant/hunk efficacy for this delta is unmeasured.
Not explored to full depth (tool budget reached): "agent reverse-audit (round 1)": did not read the bodies of canArchiveCard / canRenameCard in SessionOverviewPanel.tsx , so I could not decide whether the absence of any positive (enabled-…; "agent reverse-audit (round 1)": whether the pre-PR search filter also matched card.workspaceCwd — the new placeholder/aria-label enumerates only title, branch, PR and ID, and git show 1a73f…; "agent reverse-audit (round 2)": verifying per-browser document.getSelection() behaviour for a selection inside an open shadow root (needs a Playwright browser run against session-overview-s…; "agent 1c": npm run typecheck / focused vitest runs for SessionOverviewPanel.test.tsx and SessionDetailsTooltip.test.tsx were not executed; type-level and runtime c…; "agent reverse-audit (round 1)": executing the two e2e specs in my chunk ( web-shell.collapsed-groups-persist.spec.ts , visuals/screenshots.spec.ts ) under Playwright — every assertion was ver…, and 3 more.
Not reviewed: reverse audit — stopped before round 4 by the review time budget.
Deferred under the convergence posture (round 5, not a blocker) — recorded, not requested in this round:
packages/web-shell/client/components/SessionOverviewPanel.test.tsx:874 — [probe] nothing pins that blur cancels a rename draft rather than saving itpackages/web-shell/client/components/SessionOverviewPanel.test.tsx:1239 — [probe] rewritten title-cell test dropped the load-bearing min-w-0 assertionpackages/web-shell/client/e2e/web-shell.session-overview.spec.ts:281 — [probe] blank-row-space click addressed by positional cell index plus a hardcoded pixel offsetpackages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx:119 — [probe] new reopen guard is unwitnessed; deleting it leaves 135/135 greenpackages/web-shell/client/components/SessionOverviewPanel.test.tsx:2031 — [probe] deleted contrast loop leaves the Time cell's text-current and the new idle dimming unpinnedpackages/web-shell/client/components/SessionOverviewPanel.tsx:1152 — [probe] title button lost the only focus-openable disclosure of a truncated labelpackages/web-shell/client/e2e/web-shell.collapsed-groups-persist.spec.ts:145 — [probe] overscroll check identifies the scroller by child position inside a shared primitivepackages/web-shell/client/components/SessionOverviewPanel.test.tsx:1270 — [probe] popover align changed from center to start with no pin left anywherepackages/web-shell/client/e2e/web-shell.session-overview.spec.ts:199 — [probe] link-keeps-popover-open promise asserted only after a re-hover that re-opens it
Convergence: round 5 posted 3 inline comment(s), 3 of them reported for the first time; the previous round posted 4 (4 new). Findings keep coming back to the same files: packages/web-shell/client/components/SessionOverviewPanel.tsx (findings in round 4; 2 more now); packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx (findings in round 4; 1 more now). A cluster that keeps producing siblings usually means the fixes are treating instances of a shared root cause — triaging that cause before the next round, or splitting an independent cluster into its own pull request, tends to end the loop faster than fixing them one at a time. (Observation only — nothing was withheld from this review because of this observation.)
Mechanism health: this round did not close cleanly, so it withholds the incremental anchor — and the round it recovered had no anchor this round could use either — none at all, one with no certifier, one certified by an identity other than the one this round runs under, or one this round's fetch refused or resolved to the head — so the next review re-reads the whole diff unless recovery grafts an earlier own anchor that the round running it can use onto the complete work list this round leaves behind, and keeps doing so until a round's marker carries an anchor again or a graft lands that the round running it can use. (Stated, not acted on — this changes nothing about what the round posts.)
中文说明
仅完成部分审查，审查缺口已披露。
本轮确认的 9 条建议级发现已在 PR 上报告过，不再重复发布（列表见上方英文部分）。
未审查（原文为英文）：test-efficacy probe — the harness control produced no verdict (harnessValidated null), so revert/mutant/hunk efficacy for this delta is unmeasured.
未探索到全部深度（达到工具调用预算）："agent reverse-audit (round 1)"：did not read the bodies of canArchiveCard / canRenameCard in SessionOverviewPanel.tsx , so I could not decide whether the absence of any positive (enabled-…；"agent reverse-audit (round 1)"：whether the pre-PR search filter also matched card.workspaceCwd — the new placeholder/aria-label enumerates only title, branch, PR and ID, and git show 1a73f…；"agent reverse-audit (round 2)"：verifying per-browser document.getSelection() behaviour for a selection inside an open shadow root (needs a Playwright browser run against session-overview-s…；"agent 1c"：npm run typecheck / focused vitest runs for SessionOverviewPanel.test.tsx and SessionDetailsTooltip.test.tsx were not executed; type-level and runtime c…；"agent reverse-audit (round 1)"：executing the two e2e specs in my chunk ( web-shell.collapsed-groups-persist.spec.ts , visuals/screenshots.spec.ts ) under Playwright — every assertion was ver…，另有 3 条。
未审查：反向审计——评审时间预算不足，未能开始第 4 轮。
收敛姿态下延后（第 5 轮，非阻断）——已记录，本轮不要求修改：共 9 条（原文未翻译，列表见上方英文部分）。
收敛情况：第 5 轮发布了 3 条行内评论，其中 3 条是首次提出；上一轮发布了 4 条（其中 4 条首次提出）。发现反复回到同一批文件：packages/web-shell/client/components/SessionOverviewPanel.tsx（第 4 轮已出过发现，本轮又有 2 条）；packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx（第 4 轮已出过发现，本轮又有 1 条）。一个不断再生兄弟发现的簇，通常意味着逐条修复只在处理同一根因的实例——先定位并处理该根因，或把独立的簇拆成单独的 PR，通常比逐条修复更快结束循环。（仅为观察——本轮评审未因此扣留任何内容。）
机制健康：本轮未能干净收尾，因而扣留了增量锚点，而它恢复到的那一轮也没有留下本轮可用的锚点——要么完全没有、要么没有认证者、要么由本轮运行身份之外的身份认证、要么被本轮的获取拒绝或解析为头提交——因此下一次评审将重读整个 diff，除非恢复流程把本轮能使用的更早自有锚点嫁接到本轮留下的完整工作清单上；并会一直如此，直到某一轮的标记重新带上锚点，或落地的嫁接能被运行该轮的评审使用。（仅陈述，不据此行动——这不改变本轮发布的任何内容。）
— qwen3.8-max via Qwen Code /review (v0.23.0)
… session overview (#11238) Merge main and resolve the SessionDetailsTooltip conflict by combining the overview's controlled-open/ownerToken props with main's side prop and split-pane collision boundary. - deriveSessionCards now counts activeWorkState 'active' as running, so the status column/filter no longer labels sessions with live background work "Idle" (which also left Archive/Delete enabled for them) while the details popover and sidebar reported active work. - A click-pinned details popover whose row leaves the page mid-poll now hands focus to the panel root from a layout-effect cleanup (before React removes the row's DOM) instead of dropping it to <body>. - The row Details button is disabled while an inline rename draft is open, so its popover can no longer steal focus and blur-cancel the draft. - Repair main's split-view App test: it referenced the undeclared mockUseDaemonActivePromptBridge (a ChatPane hook App never calls); the render witness is mockUseDaemonSessionActivityBridge. Each new guard is pinned by a witness verified red without it: removing the activeWorkState term, the focus handoff, or the disabled gate turns the respective new test red.
|
🤖 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： Review round — PR #11238 (commit 85c5e61)Merged Merge conflict resolution
Findings
Merge fallout repairedMain's #11250 added Mutation probes (each guard proven load-bearing)
Verification
中文说明评审轮次 — PR #11238(提交 85c5e61)合并了 合并冲突解决
发现处理
合并引入问题的修复main 的 #11250 在 变异探针(每个守卫都验证为关键路径)
验证
Base-conflict check · 基分支冲突检查: conflicted with main — resolved in this push. · 与 main 有冲突——已在本次推送中解决。 🧵 Resolved all 4 selected review thread(s). · 已关闭全部选中的 4 条评审线程。 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 |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
1 Suggestion-level finding(s) this review confirmed are already reported on this PR and are not repeated:
- filter/search reset coverage (SessionOverviewPanel.test.tsx:587) — already reported (round-2 deferral list, review 5135733031; re-recorded in the round-4 and round-5 bodies)
Not reviewed: test-efficacy probe — the harness control produced no verdict (harnessValidated null), so revert/mutant/hunk efficacy for this delta is unmeasured.
Not reviewed: issue-fidelity — linked-issue evidence could not be fetched (closing-issue discovery failed: gh >= 2.72.0 required), so the closing-issue set is unknown rather than empty.
Not explored to full depth (tool budget reached): "agent reverse-audit (round 1)": I did not execute packages/web-shell/client/components/SessionOverviewPanel.test.tsx in this worktree, so the finding rests on reading the code paths rather t…; "agent reverse-audit (round 3)": executing the two changed Playwright specs ( web-shell.collapsed-groups-persist.spec.ts , visuals/screenshots.spec.ts ) and the new shadow-DOM harness page in …; chunk 10: the two Playwright specs were reasoned against the component/CSS source but never executed here (no dev server or browser run), so the wrap/no-clip metric asser….
Not reviewed: reverse audit — stopped before round 4 by the review time budget.
Deferred under the convergence posture (round 6, not a blocker) — recorded, not requested in this round; 2 Critical(s) among them are deferred by their axes — fails-closed on new surface, where no wrong result is certified and the merge base had neither the surface nor the defect — and remain follow-up work recorded in the findings artifact:
packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx:143 — [probe] Critical [fails-closed] [new-surface] R5-2: (fix-induced) the new unmount focus handoff excludes…packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx:255 — [probe] Critical [fails-closed] [new-surface] A parent-driven close of a click-pinned popover drops…packages/web-shell/client/components/SessionOverviewPanel.test.tsx:593 — [probe] No test asserts a status-filter count while a workspace is…packages/web-shell/client/components/SessionOverviewPanel.test.tsx:1299 — [probe] The new PR-badge click test pins only the isolation half,…packages/web-shell/client/components/SessionOverviewPanel.test.tsx:1837 — [probe] The only witness for the new N/M workspaces label cannot…packages/web-shell/client/components/SessionOverviewPanel.tsx:225 — [probe] activeWorkState 'unknown' reads Idle in the new Status…packages/web-shell/client/components/SessionOverviewPanel.tsx:782 — [probe] Clicking a row's own Rename button while its draft is open…packages/web-shell/client/components/SessionOverviewPanel.tsx:1313 — [review] The new Details button is the only action button with no…packages/web-shell/client/components/SessionOverviewPanel.tsx:1443 — [probe] The clamp-window early return preserves hover-triggered…packages/web-shell/client/components/SessionOverviewPanel.tsx:1448 — [probe] After the reconcile effect discards a rename draft,…packages/web-shell/client/components/SessionOverviewPanel.tsx:1678 — [probe] The status-filter button renders its label and count with…packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx:145 — [probe] The new focus rescue resolves its target through an…packages/web-shell/client/e2e/web-shell.session-overview.spec.ts:142 — [probe] The e2e step meant to witness 'a filter change clears the…
Convergence: round 6 posted 1 inline comment(s), 1 of them reported for the first time; the previous round posted 3 (3 new). Findings keep coming back to the same files: packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx (findings in round 5; 1 more now). A cluster that keeps producing siblings usually means the fixes are treating instances of a shared root cause — triaging that cause before the next round, or splitting an independent cluster into its own pull request, tends to end the loop faster than fixing them one at a time. (Observation only — nothing was withheld from this review because of this observation.)
Mechanism health: this round did not close cleanly, so it withholds the incremental anchor — and the round it recovered had no anchor this round could use either — none at all, one with no certifier, one certified by an identity other than the one this round runs under, or one this round's fetch refused or resolved to the head — so the next review re-reads the whole diff unless recovery grafts an earlier own anchor that the round running it can use onto the complete work list this round leaves behind, and keeps doing so until a round's marker carries an anchor again or a graft lands that the round running it can use. (Stated, not acted on — this changes nothing about what the round posts.)
中文说明
仅完成部分审查，审查缺口已披露。
本轮确认的 1 条建议级发现已在 PR 上报告过，不再重复发布（列表见上方英文部分）。
未审查（原文为英文）：test-efficacy probe — the harness control produced no verdict (harnessValidated null), so revert/mutant/hunk efficacy for this delta is unmeasured.
未审查（原文为英文）：issue-fidelity — linked-issue evidence could not be fetched (closing-issue discovery failed: gh >= 2.72.0 required), so the closing-issue set is unknown rather than empty.
未探索到全部深度（达到工具调用预算）："agent reverse-audit (round 1)"：I did not execute packages/web-shell/client/components/SessionOverviewPanel.test.tsx in this worktree, so the finding rests on reading the code paths rather t…；"agent reverse-audit (round 3)"：executing the two changed Playwright specs ( web-shell.collapsed-groups-persist.spec.ts , visuals/screenshots.spec.ts ) and the new shadow-DOM harness page in …；chunk 10：the two Playwright specs were reasoned against the component/CSS source but never executed here (no dev server or browser run), so the wrap/no-clip metric asser…。
未审查：反向审计——评审时间预算不足，未能开始第 4 轮。
收敛姿态下延后（第 6 轮，非阻断）——已记录，本轮不要求修改；其中 2 条 Critical 按其失败方向与对照基线延后——fails-closed 且 new-surface：未认证任何错误结果，且 merge base 既无该功能面也无该缺陷——作为后续工作记录在 findings 工件中：共 13 条（原文未翻译，列表见上方英文部分）。
收敛情况：第 6 轮发布了 1 条行内评论，其中 1 条是首次提出；上一轮发布了 3 条（其中 3 条首次提出）。发现反复回到同一批文件：packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx（第 5 轮已出过发现，本轮又有 1 条）。一个不断再生兄弟发现的簇，通常意味着逐条修复只在处理同一根因的实例——先定位并处理该根因，或把独立的簇拆成单独的 PR，通常比逐条修复更快结束循环。（仅为观察——本轮评审未因此扣留任何内容。）
机制健康：本轮未能干净收尾，因而扣留了增量锚点，而它恢复到的那一轮也没有留下本轮可用的锚点——要么完全没有、要么没有认证者、要么由本轮运行身份之外的身份认证、要么被本轮的获取拒绝或解析为头提交——因此下一次评审将重读整个 diff，除非恢复流程把本轮能使用的更早自有锚点嫁接到本轮留下的完整工作清单上；并会一直如此，直到某一轮的标记重新带上锚点，或落地的嫁接能被运行该轮的评审使用。（仅陈述，不据此行动——这不改变本轮发布的任何内容。）
— qwen3.8-max via Qwen Code /review (v0.23.0)
|
🔀 Base updated: red check(s) [web-shell E2E Smoke (ubuntu-latest, Node 22.x)] pass on current main — merged current main via update-branch; CI will re-run. 中文说明🔀 已更新 base：红色检查 [web-shell E2E Smoke (ubuntu-latest, Node 22.x)] 在当前 main 上通过 —— 已通过 update-branch 合入当前 main，CI 将重新运行。 |
|
CI repair for 8bee8cc The failed run at 1044b1c had two independent causes:
Validation on macOS, Node.js 22.22.2, Chromium:
本轮修复了完整路径断言过时，以及测试超时重试时未等待 supervisor 退出便清理目录的问题。55 项相关单测、70 项 smoke 和实际打包 CLI 上的 5 项分屏测试通过；延迟、元数据替换后提前失败、真实写入错误三条故障路径均已验证。审计发现的原 token 清理问题在提交前修正，随后重新完成连续两轮反向审计。 |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
11 Suggestion-level finding(s) this review confirmed are already reported on this PR and are not repeated:
- activeWorkState 'unknown' status-column versus popover disagreement (SessionOverviewPanel.tsx:566) — already reported (round-6 deferral list, review 5148947616)
- workspace funnel open flag outliving its unmounted popover (SessionOverviewPanel.tsx:1541) — already reported (round-4 deferral list, review 5141881786)
- inert TooltipProvider left by the column removals (SessionOverviewPanel.tsx:1724) — already reported (round-2 deferral list, review 5135733031)
- overview visual fixture row-action and PR-badge coverage (e2e/visuals/screenshots.spec.ts:73) — already reported (round-3 deferral list, review 5137868769; re-recorded in the round-5 body)
- details reopen-guard and copy-badge witness (sidebar/SessionDetailsTooltip.tsx:155) — already reported (round-5 deferral list, review 5146099714; round-4 inline thread R4-3 still stands)
- overview git search matching versus the shared exact-match rule (SessionOverviewPanel.test.tsx:526) — already reported (round-4 deferral list, review 5141881786; re-recorded in the round-5 body)
- filter-clears-selection e2e witness (e2e/web-shell.session-overview.spec.ts:142) — already reported (round-6 deferral list, review 5148947616)
- workspace and branch programmatic field labels (i18n.tsx:3133) — already reported (round-3 deferral list, review 5137868769; re-recorded in the round-4 and round-5 bodies)
- footer pagination coverage in the selected state (SessionOverviewPanel.test.tsx:1447) — already reported (round-4 deferral list, review 5141881786)
- batch-button data-variant coverage (SessionOverviewPanel.test.tsx:1449) — already reported (round-2 deferral list, review 5135733031; re-recorded in the round-4 and round-5 bodies)
- rename-draft discard without focus restore (SessionOverviewPanel.tsx:1443) — already reported (round-6 deferral list, review 5148947616)
Not reviewed: test-efficacy probe — the harness control produced no verdict (harnessValidated null), so revert/mutant/hunk efficacy for this delta is unmeasured.
Not explored to full depth (tool budget reached): "agent reverse-audit (round 2)": confirming by reading the session-resolution bootstrap ( WorkspaceSessionProvider / connection setup) that page.goto('/') leaves no session loaded, rather th…; "agent reverse-audit (round 2)": did not execute either Playwright spec — the new layout assertions in web-shell.collapsed-groups-persist.spec.ts ( clientHeight > lineHeight , scrollHeight <…; "agent reverse-audit (round 2)": did not verify that the { name, baseBranch } literals added to createOrganizedScenario and to the new visuals/screenshots.spec.ts fixture satisfy every re…; "agent reverse-audit (round 2)": did not confirm that vite dev actually serves and transforms the new harness's inline module script (bare react / react-dom/client specifiers plus ../ind….
Not reviewed: reverse audit — stopped before round 3 by the review time budget.
2 Suggestion(s) were drafted inline past the resolved critical posting floor; the CLI moved them into the deferral list below (floor enforcement).
Deferred under the convergence posture (round 7, not a blocker) — recorded, not requested in this round; 2 Critical(s) among them are deferred by their axes — fails-closed on new surface, where no wrong result is certified and the merge base had neither the surface nor the defect — and remain follow-up work recorded in the findings artifact:
packages/web-shell/client/components/sidebar/SessionDetailsTooltip.tsx:143 — [probe] Critical [fails-closed] [new-surface] R5-2: (fix-induced) the new unmount focus handoff excludes the anchor-parked focus, so a hover-replaced row that then…packages/web-shell/client/components/SessionOverviewPanel.tsx:1159 — [probe] Critical [fails-closed] [new-surface] the drag-select guard's document-global point-in-time Selection predicate swallows a deliberate title click after any earlier…packages/cli/src/agent-view/supervisor-runner.test.ts:427 — [review] The teardown path this CI-repair commit exists to add — registering a live runner for shutdown *before* readiness is awaited, and shutting it down with the retained token …packages/cli/src/agent-view/supervisor-runner.test.ts:38 — [review] Making the supervisor cleanup the first await in this hook, with Promise.all (which rejects), means a rejecting supervisor short-circuits afterEach before the fake se…packages/web-shell/client/components/SessionOverviewPanel.tsx:1307 — [review] the two new session-details entry points ignore the host's sessionActions.items embed config that every other surface honoursdocs/design/web-shell-session-overview.md:1 — [review] new design doc ships English-only, with no zh-CN sibling and no reciprocal language linksdocs/design/web-shell-session-overview.md:66 — [review] the two Critical focus-loss findings round 6 deferred are in neither issue 11390 nor this doc's out-of-scope list, so no owned record carries thempackages/cli/src/agent-view/supervisor-runner.test.ts:436 — [review] the teardown's vi.waitFor budget (10s) equals the resolved afterEach hookTimeout on non-ECS runners, so vitest kills the hook firstpackages/web-shell/client/e2e/session-overview-shadow-dom.html:16 — [review] the shadow-DOM harness is the only app-level render entry point without React.StrictMode, so its smoke test never sees the effect replay
Mechanism health: this round did not close cleanly, so it withholds the incremental anchor — and the round it recovered had no anchor this round could use either — none at all, one with no certifier, one certified by an identity other than the one this round runs under, or one this round's fetch refused or resolved to the head — so the next review re-reads the whole diff unless recovery grafts an earlier own anchor that the round running it can use onto the complete work list this round leaves behind, and keeps doing so until a round's marker carries an anchor again or a graft lands that the round running it can use. (Stated, not acted on — this changes nothing about what the round posts.)
中文说明
仅完成部分审查，审查缺口已披露。
本轮确认的 11 条建议级发现已在 PR 上报告过，不再重复发布（列表见上方英文部分）。
未审查（原文为英文）：test-efficacy probe — the harness control produced no verdict (harnessValidated null), so revert/mutant/hunk efficacy for this delta is unmeasured.
未探索到全部深度（达到工具调用预算）："agent reverse-audit (round 2)"：confirming by reading the session-resolution bootstrap ( WorkspaceSessionProvider / connection setup) that page.goto('/') leaves no session loaded, rather th…；"agent reverse-audit (round 2)"：did not execute either Playwright spec — the new layout assertions in web-shell.collapsed-groups-persist.spec.ts ( clientHeight > lineHeight , scrollHeight <…；"agent reverse-audit (round 2)"：did not verify that the { name, baseBranch } literals added to createOrganizedScenario and to the new visuals/screenshots.spec.ts fixture satisfy every re…；"agent reverse-audit (round 2)"：did not confirm that vite dev actually serves and transforms the new harness's inline module script (bare react / react-dom/client specifiers plus ../ind…。
未审查：反向审计——评审时间预算不足，未能开始第 3 轮。
2 条 Suggestion 在已解析的 critical 发布下限之外被起草为行内评论；CLI 已将其移入下方延后清单（下限强制执行）。
收敛姿态下延后（第 7 轮，非阻断）——已记录，本轮不要求修改；其中 2 条 Critical 按其失败方向与对照基线延后——fails-closed 且 new-surface：未认证任何错误结果，且 merge base 既无该功能面也无该缺陷——作为后续工作记录在 findings 工件中：共 9 条（原文未翻译，列表见上方英文部分）。
机制健康：本轮未能干净收尾，因而扣留了增量锚点，而它恢复到的那一轮也没有留下本轮可用的锚点——要么完全没有、要么没有认证者、要么由本轮运行身份之外的身份认证、要么被本轮的获取拒绝或解析为头提交——因此下一次评审将重读整个 diff，除非恢复流程把本轮能使用的更早自有锚点嫁接到本轮留下的完整工作清单上；并会一直如此，直到某一轮的标记重新带上锚点，或落地的嫁接能被运行该轮的评审使用。（仅陈述，不据此行动——这不改变本轮发布的任何内容。）
— qwen3.8-max via Qwen Code /review (v0.23.1)
|
🤖 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 change requiredActionable feedback this roundOnly one actionable item was newer than the last evaluation: the inline Everything else in this round's feedback requests no work:
Why no commit was madeThe defect was already repaired at HEAD by
The thread VerificationCommands actually run in this round, with results:
No source, test, or spec file was modified this round; the working tree is 中文说明Autofix 本轮——无需改动本轮可执行的反馈比上次评估更新的反馈中只有一条可执行项：行内 Critical 本轮反馈中的其余内容都不要求改动：
为什么本轮没有提交该缺陷已在 HEAD 由
线程 验证本轮实际运行的命令及结果：
本轮未修改任何源码、测试或 spec 文件；工作树在 Base-conflict check · 基分支冲突检查: no conflict with main. · 与 main 无冲突。 🧵 Resolved all 1 selected review thread(s). · 已关闭全部选中的 1 条评审线程。 🧠 Handled by Qwen Code · model/模型 |
|
@qwen-code /triage |
|
Sandboxed verification: ❌ not passed — findings reported (agent verdict) - workflow run Ran the PR in an isolated, token-free container: A/B against the base build, mock-free harness assertions, targeted gates. Advisory evidence for human reviewers — not a review, an approval, or a CI check. Scripted assertions: 58 passed · 0 failed · 58 total Flakiness gate: ✅ 4 changed test file(s) x 5 identical rounds, no divergence 中文 — 判定：❌ 不通过 · 报告了发现（agent 判定）沙箱验证在隔离、无凭证的容器中执行了该 PR 的代码（与 base 构建 A/B 对照、无 mock harness 断言、定向门禁）。仅作为评审证据，不构成评审、批准或 CI 检查。 脚本断言：58 通过 · 0 失败 · 58 总计 抖动门：✅ 4 changed test file(s) x 5 identical rounds, no divergence Verification reportPR 11238 deep verification —
|
| fixture | base (HEAD^1) |
head | flip |
|---|---|---|---|
A1 activeWorkState:'active', no prompt signal |
idle | running | FLIP |
A2 activeWorkState:'active', hasActivePrompt:false |
idle | running | FLIP |
A3 'active' + status report hasActivePrompt:true |
running | running | – |
A4 sibling activeWorkState:'unknown' |
idle | idle | – |
A5 sibling activeWorkState:'unsupported' |
idle | idle | – |
A6 activeWorkState:'idle' |
idle | idle | – |
A7 activeWorkState absent (older daemon) |
idle | idle | – |
A8 'active' + pending permission via status report |
needsApproval | needsApproval | – |
A9 'active' + isWaitingForUserQuestion |
askUserQuestion | askUserQuestion | – |
A10 hasActivePrompt:true, activeWorkState absent |
running | running | – |
Oracle: the status field of the derived card. 2 of 10 cells flip; the other 8 are
byte-identical, so the change is confined to exactly the value it names and does not disturb
the approval/question precedence it sits below.
Part B — the rendered consequence (jsdom, real component, no stub of the unit under test)
| fixture | base Archive/Delete disabled |
head Archive/Delete disabled |
flip |
|---|---|---|---|
B1 activeWorkState:'active' only |
false / false (both enabled) | true / true (both disabled) | FLIP |
B2 sibling 'unknown' only |
false / false | false / false | – |
B3 sibling 'unsupported' only |
false / false | false / false | – |
B4 truly idle ('idle') |
false / false | false / false | – |
Oracle: the disabled property of the buttons carrying aria-label="Archive" /
"Delete". This is why the status term matters — canArchiveCard and canDeleteCard are
byte-identical on both arms (both require card.status === 'idle'), so the PR's one-term
change to the status derivation is the only thing standing between a session doing background
work and a destructive mutation. B4 proves the fixture isolates that term: a genuinely idle
row keeps both mutations enabled on head.
Part C — the two status surfaces for one row
The PR deletes the popover's worktreeOnly prop, so head is the first version where the
overview renders both a Status column and a popover status line for the same row. In base
both were absent from the overview (measured: no [data-web-shell-session-status] element, no
[data-web-shell-session-details] button), because base's overview popover was the git-only
variant whose status and session-ID rows sat inside {!worktreeOnly && …}.
activeWorkState |
column attr | column text | popover text | agree | Archive/Delete disabled | narrow cue |
|---|---|---|---|---|---|---|
'active' |
running | Running | Running | yes | true / true | running |
'unknown' |
idle | Idle | Background activity unknown | NO | false / false | null |
'unsupported' |
idle | Idle | Idle · 0 client(s) | yes | false / false | null |
'idle' |
idle | Idle | Idle · 0 client(s) | yes | false / false | null |
| absent | idle | Idle | Idle · 0 client(s) | yes | false / false | null |
Real browser renders of the head build are in 03-overview-head-chromium-dark.png (overview,
four distinct states) and 04-details-popover-head-chromium-light.png (the reused details
popover).
Control validation
The monorepo symlink trap was checked rather than assumed. The diff touches no
package.json/lockfile (git diff --name-only HEAD^1..HEAD | grep -E 'package(-lock)?\.json'
→ none), so sharing the root node_modules is dependency-clean. Each arm's harness recorded
its own import.meta.url and the driver asserted it resolved inside the expected tree
(base: component resolved inside the base tree → PASS; same for head). ./SessionOverviewPanel
is a relative specifier, so Vite resolves it against the harness's own path; the one internal
package that does cross into the head tree (@qwen-code/sdk) is reached only through
import type, which esbuild strips, and @qwen-code/web-shell/daemon-react-sdk is both
aliased to each tree's own client/daemon-react-sdk.ts and mocked outright by the harness.
Mutation matrix — the PR's own claimed witnesses, re-measured
The merge commit states: "Each new guard is pinned by a witness verified red without it:
removing the activeWorkState term, the focus handoff, or the disabled gate turns the
respective new test red." That is a falsifiable claim, so it was re-measured rather than
accepted. Cells captured live as 02-mutation-matrix.png. Sources were restored after every
mutant and the restore verified by sha256 (client/components/SessionOverviewPanel.tsx restored byte-for-byte → PASS, likewise the tooltip).
| mutant | guard | suite result | killed | first failing test | assertion quoted |
|---|---|---|---|---|---|
| M0 control | none (unmutated) | 141 passed | green as required | – | – |
| M1 | deriveSessionCards: drop || session.activeWorkState === 'active' |
1 failed | yes | counts background work as running even without an active prompt | expected 'Idle' to be 'Running' |
| M2 | details button: disabled={Boolean(editingIdentity)} → disabled={false} |
1 failed | yes | keeps an inline rename draft when clicking the details button | expected false to be true |
| M3 | tooltip: drop the layout-effect focus handoff to [data-web-shell-session-panel] |
2 failed | yes | closes details when live state removes its anchor through a filter / …through a page | expected <body>…</body> to be <div tabindex="-1" …> |
| M4 | tooltip: focus rescue no longer scoped to the shared ownerToken |
1 failed | yes | leaves another owner's pinned details open and focused on hover | expected 'closed' to be 'open' |
| M5 harness control | STATUS_PRIORITY inverted (idle first) |
8 failed | yes | ranks needs-approval above user-input above running above idle, then by recency +2 | expected ['s-idle','s-run','s-q','s-appr'] to deeply equal ['s-appr','s-q','s-run','s-idle'] |
Every mutant died on a behavioural assertion naming expected-versus-actual values; none
died on a compile or import break (compileError=false on all rows). Attribution is correct:
each witness test is named after the guard it pins, not an unrelated bystander. M5 is the
harness-level positive control landed in the same file as M1/M2, so "the suite does not cover
this" and "my harness never ran your suite" are distinguishable. No survivor — all four
guards the PR claims are genuinely pinned.
Corrections to the PR description
These are corrections to the description, not requests to change code.
"both dark/light visual cases passed" reads as a visual-regression result. It is not. The
visuals suite contains zero toHaveScreenshot assertions anywhere under client/e2e/, and
captureScreenshot (client/e2e/visuals/harness.ts:164) only calls page.screenshot({ path })
after freezing animations. I ran the two new cases and they did pass — 2 passed (9.3s),
producing session-overview-{dark,light}.png and session-overview-details-{dark,light}.png —
but "passed" here means the page rendered, the expected text and dialog were present, and the
PNG was written, not that pixels were compared against a committed baseline. The before/after
threshold the harness comments mention lives in separate tooling, not in these tests. The
substantive assertions in those two cases (the overview contains the fixture title; the details
dialog is visible with the accessible name Review release approval) did execute and did pass.
Everything else in the description's validation block that I could check held up exactly:
"all 70 Chromium smoke cases" → 70/70; "220 focused unit tests" is a superset count I did not
reproduce, but the three changed test files give 141 and the whole package gives 6,769;
"repository build, typecheck and bundle passed" → typecheck re-run here and green.
Findings
1. activeWorkState: 'unknown' is labelled Idle with Archive/Delete enabled — the overview fails open where its own data source documents fail-closed (Suggestion)
Measured, from Part A / Part B / Part C above: a session whose only live-work signal is
activeWorkState: 'unknown' derives status: 'idle', renders "Idle" in the new Status
column, renders "Background activity unknown" in the details popover for the same row,
shows no narrow-layout cue, and offers Archive and Delete enabled.
Why 'unknown' is not a theoretical value — the producer's own documentation
(packages/acp-bridge/src/bridge.ts):
entryActiveWorkStatereturns'unknown'whenchildWorkIsUnknown(entry), i.e. "the
channel negotiated reporting, but no snapshot recent enough to grade has arrived.
Never-reported and gone-quiet are the same state on purpose — a snapshot from ten minutes
ago says nothing about whether a background agent started since" — or when the channel's
capability does not cover all fiveACTIVE_WORK_HOLD_CATEGORIES(agent,notification,
shell,session,workflow).entryHasActiveWork, the bridge's health surface, is documented as "Fails closed on
ignorance: unknown reads the same as busy, because a controller must not be able to
mistake 'nobody told me' for 'nothing is running'."- The bridge's automatic teardown acts on that: it routes through
closeIfChildUnheld→
confirmChildUnheld, asking the child before destroying.
The overview is a controller, and it does the thing the comment says a controller must not: it
reads "nobody told me" as "nothing is running" and enables destructive mutations on that
basis. The user-initiated path has no server-side backstop — I traced it rather than assumed
it: deleteSessions (routes/session.ts) → deleteDaemonSessions (server/session-archive.ts,
which contains no hasActivePrompt / activeWork / busy refusal) → bridge.closeSession →
closeSessionImpl directly; closeSession does not call closeIfChildUnheld (its only
two call sites, bridge.ts:3419 and :4097, are the automatic-teardown paths). So
canDeleteCard's card.status === 'idle' is the only active-work gate on that path.
Pre-existing cause vs this PR's contribution, named separately:
- Pre-existing (not this PR's fault): base derived
'unknown'→idleidentically, and
canArchiveCard/canDeleteCardare byte-identical on both arms. The permissive gate for
'unknown'predates this PR. - This PR's contribution: (a) it took on making the derived status truthful and closed
'active'only, leaving the sibling of the same bug class open; (b) by deleting
worktreeOnlyit made the overview render a Status column and a popover status for the same
row for the first time, so the two now visibly contradict each other ("Idle" vs "Background
activity unknown") in one surface — measured as theagree = NOrow of Part C. That
contradicts the design doc's own stated goal, "Make the popover status agree with the
overview's derived live state."
What this finding is NOT (bounded, per the evidence):
- It is not a regression: base behaved identically for
'unknown', and the PR strictly
improved the'active'case (proven by the B1 flip). 'unsupported'is not part of this finding. The bridge explicitly scopes it out: "A
channel that never negotiated is not 'unknown', it is out of scope: treating it as unknown
would make every legacy Session unreapable." Treating'unsupported'as deletable is a
documented, deliberate choice, and Part C shows its two surfaces agree ("Idle" / "Idle · 0
client(s)"). I am not asking for it to change.- A confirmation dialog (
sessionsOverview.confirmDeleteTitle, "Delete session?") stands
between the button and the call, and Archive is reversible. - I did not demonstrate end-to-end loss of a genuinely held background task: that needs a
live runtime with a stale-holds child, which this container does not have. The chain above is
proven by reading the four call sites plus measuring the UI gate, not by observing a deletion.
No measured patch is offered, deliberately. Either fix direction changes a contract the
author owns: relabelling needs a new SessionCardStatus value (which ripples into
STATUS_PRIORITY, matchesStatus, the filter enum, and four new i18n keys in two locales),
while gating the mutations needs activeWorkState carried onto SessionCard (which today
holds only the derived status). Both are product decisions — blocking Delete on 'unknown'
could over-block whenever a runtime's hold reporting goes stale. What I did measure is the
constraint that applies to whichever is chosen: see Finding 2 — nothing in the repo pins this
axis today, so the fix must ship with its own fixture.
Reproduce:
node tmp/pr11238-verify-20260909-120107/ab-driver.mjs tmp/pr11238-verify-20260909-120107
# Part A row "A4 SIBLING activeWorkState=unknown": base=idle head=idle
# Part C row "unknown": column "Idle" vs popover "Background activity unknown", deleteDisabled=false2. activeWorkState appears in no browser-level fixture, so the guard the merge commit calls its point is unpinned above unit level (Suggestion)
The merge commit frames the activeWorkState term as the fix that stops Archive/Delete being
enabled for sessions with live background work. It is pinned at unit level (M1 above kills
counts background work as running even without an active prompt). It is pinned at no
other level:
- Removing the term leaves the new 12-test browser spec fully green —
12 passed (39.1s),
loggate-e2e-m1-mutant.log. - The reason is a fixture gap, proven by census rather than by reading: 0 of 26 e2e
artifacts (*.spec.tsand*.htmlunderclient/e2e/) contain the string
activeWorkState. The overview spec drives its four states purely through
hasActivePrompt/isWaitingForPermission/isWaitingForUserQuestion; so does the new
visual-capture scenario. The PR's own screenshot evidence therefore never contains a session
in the state the fix exists for.
The fixture that would go red: add a fifth row to the web-shell.session-overview.spec.ts
state matrix — displayName: 'Background work fixture', hasActivePrompt: false,
activeWorkState: 'active' — and assert [data-web-shell-session-status="running"] plus
Archive/Delete toBeDisabled(), mirroring the loop the spec already runs for the other four
states. That single row is what turns M1 from a unit-only witness into a browser-level one.
Reproduce:
node tmp/pr11238-verify-20260909-120107/coverage-crosscheck.mjs tmp/pr11238-verify-20260909-120107
# "census: NO e2e artifact anywhere mentions activeWorkState" -> PASS (26 scanned, 0 matches)3. The sibling axis is pinned in neither direction — a reverse-mutation probe (completeness reporting, not a merge condition)
Per the rule that a suite unable to tell head from head-plus-fix has its coverage gap exactly
where the next regression will land: I applied a candidate change in a scratch copy (treat
'unknown' as 'running' in deriveSessionCards) and re-ran both layers.
| layer | head | head + probe |
|---|---|---|
| unit (3 changed test files) | 141 passed | 141 passed |
e2e (web-shell.session-overview.spec.ts) |
12 passed | 12 passed |
Green on both sides is not reassurance — it is proof that no test in the repo asserts anything
about how the overview treats activeWorkState: 'unknown', in either direction. The tooltip's
label is pinned (SessionDetailsTooltip.test.tsx:76 → 'Background activity unknown'); the
overview's contradicting treatment of the same session is not. Whichever way Finding 1 is
resolved, it needs a new fixture, because today a regression here is invisible. Source restored
byte-for-byte (f30b806804c01355 before and after) and git status --porcelain is empty.
4. Design doc has no .zh-CN.md pair and no language links (Suggestion — repo-wide gap, not this PR's)
docs/design/README.md requires "Maintain two complete versions in the same directory:
<name>.md for English and <name>.zh-CN.md for Simplified Chinese" and "Add relative links
to both versions immediately below each document title." The PR adds
docs/design/web-shell-session-overview.md with neither: no sibling .zh-CN.md, and the title
is followed directly by ## Problem.
Context that should decide how this is handled, measured rather than assumed: docs/design
holds 485 non-.zh-CN Markdown documents (recursively, across 37 subdirectories) and
exactly 2 .zh-CN.md files — README.zh-CN.md, and
customize-banner-area/customize-banner-area.zh-CN.md. So one design document in 485 has
the required Chinese counterpart. And even that single complying pair does not carry the
reciprocal language links below its title: grep -rl 简体中文 docs/design matches only
README.md and README.zh-CN.md. Neither do the six most recently touched design docs. This
PR's doc follows the same shape as essentially every neighbour. Per the project's own review
rule a translation gap is a Suggestion, not a Critical; recording it here is recording the rule,
not singling out this author. The real decision is whether to start enforcing the convention on
new docs or to treat it as a repo-wide hygiene item.
Reviewer Test Plan — walked step by step
I ran the automated specs that encode each step rather than driving a browser by hand; the
mapping below says which evidence covers which step. No step turned out to be structurally
unreachable — every one maps onto a spec that exists and passes.
| step | claim | evidence | result |
|---|---|---|---|
| 1 | four distinct states, spinner only for running; filters combine with title/branch/PR (#123)/ID search; changing a filter clears selection and returns to page 1 |
e2e compact overview distinguishes states… asserts all four data-web-shell-session-status labels + title attrs and Session ID column count 0; the reset effect (SessionOverviewPanel.tsx:1026) zeroes pageIndex, clears rowSelection, closes details, keyed on exactly [excludedWorkspaceCwds, searchQuery, statusFilter]; search covers label/sessionId/gitBranch/#<pr.number> |
covered, passed |
| 2 | hover for full details; Enter on the details button; copy ID; PR/issue links; Escape returns focus; small-viewport wrap and scroll to the copy action | e2e title hover exposes full metadata, permits copy, and links keep overview open, details button supports keyboard and Escape restores focus without navigating, overview details stay inside an embedded shell; my Part C read the popover's [data-web-shell-session-id] and status row directly |
covered, passed |
| 3 | row click opens the session; drag-select keeps the overview open; plain-cell click during rename preserves the draft; checkboxes reveal batch actions; restrictions on active sessions and cross-workspace ops remain | e2e row opens a session without selecting it, keeps text selection and rename drafts in the overview; M2 mutant killed the draft-preservation witness; canUseSessionMutation gates cross-workspace; Part B/Part C measured the active-session restriction and found its 'unknown' residual (Finding 1) |
covered, passed, with Finding 1 |
| 4 | sorting during rename cancels the draft and keeps focus; Shadow DOM hover replacement keeps focus on the incoming title; focus preserved from a search field or a different popover owner | e2e sorting during rename preserves keyboard navigation with N sessions (2 params), shadow portal details keep focus in the overview after hover replacement, hover details preserve a focused search field; M4 mutant killed the owner-token witness | covered, passed |
| 5 | rename-then-cancel leaves details closed; a live update moving the row off-page discards the hidden draft; a catalog shrink preserves visible details/draft; at 499px each active/attention row keeps its title icon | e2e attention cues stay visible in the pinned title at narrow widths; M3 mutant killed closes details when live state removes its anchor through a filter and …through a page | covered, passed |
Gates
| gate | command | result |
|---|---|---|
| web-shell unit (whole package) | npx vitest run in packages/web-shell |
290 files / 6,769 tests passed, 0 failed (89.8s) |
| the three changed test files | npx vitest run client/components/SessionOverviewPanel.test.tsx client/components/sidebar/SessionDetailsTooltip.test.tsx client/components/sidebar/sessionDetailsCollisionBoundary.test.ts |
141 passed (120 + 17 + 4) |
| CLI test touched by the follow-up commit | npx vitest run src/agent-view/supervisor-runner.test.ts in packages/cli |
10 passed (817 ms) |
| Chromium smoke | npx playwright test --grep @smoke --project=chromium |
70 passed (1.7 m) |
| new overview spec | npx playwright test client/e2e/web-shell.session-overview.spec.ts |
12 passed (39.1 s) |
| other changed specs | … web-shell.collapsed-groups-persist.spec.ts web-shell.split-persist.spec.ts |
7 passed (24.8 s) |
| new visual captures | npx playwright test --config playwright.visuals.config.ts --grep "session overview" |
2 passed, 4 PNGs written (see Correction above) |
| typecheck | npm run typecheck (repo-wide) |
pass |
| eslint | npx eslint on all 12 changed .ts/.tsx files |
exit 0, zero output bytes |
| prettier | npx prettier --experimental-cli --check packages/web-shell docs/design/web-shell-session-overview.md packages/cli/src/agent-view/supervisor-runner.test.ts |
All matched files use Prettier code style! |
Both lint gates were proven live before being cited. Appending const plantedUnusedVariable = 42;
to sessionDetailsCollisionBoundary.ts made eslint report
11:7 error 'plantedUnusedVariable' is assigned a value but never used @typescript-eslint/no-unused-vars,
and appending const spaced = 1; made prettier report
[warn] Code style issues found in 1 file. The file was restored to sha 1efc2137830aaf6a,
matching its pre-plant value. An unproven green gate would have been an assumption here, since
eslint exits 0 both when it passes and when it matches nothing.
Not covered
- Per-commit attribution. The checkout is depth 2:
git rev-parse --is-shallow-repository
→true, andgit rev-list HEAD^1..HEAD^2returns 1 commit while
$QWEN_VERIFY_CONTEXTlists 14. Only the aggregateHEAD^1..HEADdiff was verified. The
four bot/authored fix commits described in the snapshot (rename-time control clicks, details
focus ownership, shadow-portal focus, the status/focus/rename merge) were verified only as
they survive in the aggregate head, not individually. - Base OID drift. The snapshot's
baseRefOidiseb383fd3bbd4b55f1c5ba8d7da7048193f018230,
which is not present locally (git cat-file -tfails);mainadvanced between the
snapshot fetch and the merge-ref computation. The A/B usedHEAD^1
(cf987cdc59997b9ea935226f592bc5be1e121a06), the base tip GitHub actually merged into. No
trial merge into a newermainwas performed — with only three commits reachable there is no
newermainlocally to merge into. - End-to-end demonstration of Finding 1. No live runtime, model, or stale-holds child was
available, so no actual deletion of a session holding background work was observed. The chain
is a code-path proof plus a measured UI gate. - Non-Chromium browsers and the
mobile-chromiumproject. Only--project=chromiumwas
run; Firefox/WebKit are not installed. The PR itself scopes these out. - Repo-wide test suite. Only
packages/web-shellin full and one CLI test file.packages/core,
the rest ofpackages/cli, and the integration suites were not run. npm run build/npm run bundlewere not re-run by this round; the workflow built at
HEAD before the clock started, and I rannpm run typecheckmyself instead. The e2e specs ran
against the vite dev server (source transformed live), not the bundleddist/cli.js, so the
PR's claim of "against the actual bundled CLI" was not reproduced at that level.- The PR's separate E2E comment and the six external screenshot URLs in the description were
not fetched — this job has no GitHub token and treats the local tree plus the snapshot as the
whole world. - Real daemon responses. Every browser and jsdom run used the specs' own intercepted fake
daemon. The[WebServer] connect ECONNREFUSED 127.0.0.1:4170lines in the e2e logs are the
vite dev proxy finding no real daemon; they are expected noise, and all specs passed anyway. - Interactive manual walkthrough. The Reviewer Test Plan was verified through the specs that
encode it (table above), not by hand-driving a browser. - The
sessionDetailsCollisionBoundaryfallback change was traced but not A/B'd. Its
behaviour for every caller configuration was reasoned through rather than measured: the
WorkspaceDetailsTooltipcall site is byte-identical between arms, and for aSessionDetailsTooltip
anchor the newclosest('[data-web-shell-root]') ?? closest('aside') ?? nullchain matches
base'sclosest('aside') ?? nullresult in the sidebar cases (with and without a shell root)
and differs only for the overview anchor, where it now constrains to the Web Shell instead of
returningnull— which is the intended change and is pinned by the new test resolves an
overview anchor inside a shell without an aside. Flagging it as reasoned, not measured.
Methodology
Everything ran in the CI verify container (node:22-bookworm, node v22.23.2, npm 10.9.8) on
the merge-ref checkout at 3582b7787a, with npm ci and npm run build already completed by
the workflow. The base arm was a scratch git worktree at tmp/base-tree (HEAD^1), nested
under the repo so it resolved node_modules by walking up to the root — no per-tree install —
and removed with git worktree remove --force once the A/B cells were captured; the tree is
currently clean (git status --porcelain empty, asserted by the coverage cross-check).
Three drivers, all kept in this artifact dir so a maintainer can rerun them verbatim:
ab-driver.mjs copies the same verify-ab.test.tsx into both trees and runs it under each
tree's own packages/web-shell/vitest.config.ts, so the arms differ only by the code under
test; the harness mocks only the daemon-facing seams (daemon-react-sdk, useScopedSessions,
the live-state hook, the catalog controller) and never the unit under test — deriveSessionCards
and SessionOverviewPanel are the real compiled components, rendered through react-dom/client
in jsdom, with the real SessionDetailsTooltip portalling into the document.
mutation-driver.mjs applies each mutant as an exact single-occurrence string replacement
(throwing if the anchor is not unique), runs the relevant suites, classifies the red as
behavioural versus compile breakage, and restores with sha256 verification.
coverage-crosscheck.mjs asserts over the recorded run logs plus a filesystem census.
Assertion bookkeeping follows the skill contract: an A/B control cell asserts that the base
arm shows the pre-fix behaviour, so base going red as predicted counts as a pass; fail
counts only unexpected outcomes. Two early failures were harness defects, not PR defects, and
were fixed and re-run rather than shipped: Part B initially executed at module-collection time
before beforeEach installed the mocks, so the ErrorBoundary swallowed the render and the test
still passed while reporting found: 0 (now guarded by a positive control requiring exactly one
Archive and one Delete button per row), and the control-validation step compared a file:// URL
against a filesystem path. Repo suite counts are reported separately as gates and are not
folded into assertions.json, so every number there is one of my own scripted checks.
Raw logs: ab-base.json, ab-head.json, ab-base.log, ab-head.log, ab-driver-stdout.log,
ab-tally.json, mutation-M*.log, mutation-tally.json, mutation-stdout.log,
probe-run.log, coverage-stdout.log, coverage-tally.json, gate-web-shell-unit.log,
gate-e2e-session-overview.log, gate-e2e-smoke-head.log, gate-e2e-other-specs.log,
gate-e2e-m1-mutant.log, gate-visuals.log, plus the two component diffs.
Images: 01-ab-status-and-mutation-gate.png (the A/B cells as they printed),
02-mutation-matrix.png (the matrix as it printed),
03-overview-head-chromium-dark.png and 04-details-popover-head-chromium-light.png (real
Chromium renders of the head build, produced by the PR's own visual spec),
05-coverage-crosscheck.png (the coverage census and probe results).
Flakiness gate log
e2e suite, out of gate scope: packages/web-shell/client/e2e/visuals/screenshots.spec.ts
e2e suite, out of gate scope: packages/web-shell/client/e2e/web-shell.collapsed-groups-persist.spec.ts
e2e suite, out of gate scope: packages/web-shell/client/e2e/web-shell.session-overview.spec.ts
e2e suite, out of gate scope: packages/web-shell/client/e2e/web-shell.split-persist.spec.ts
rounds=5 files=4 skipped=4
file packages/cli/src/agent-view/supervisor-runner.test.ts: (cd packages/cli) npx --no-install vitest run ./src/agent-view/supervisor-runner.test.ts
file packages/web-shell/client/components/SessionOverviewPanel.test.tsx: (cd packages/web-shell) npx --no-install vitest run ./client/components/SessionOverviewPanel.test.tsx
file packages/web-shell/client/components/sidebar/SessionDetailsTooltip.test.tsx: (cd packages/web-shell) npx --no-install vitest run ./client/components/sidebar/SessionDetailsTooltip.test.tsx
file packages/web-shell/client/components/sidebar/sessionDetailsCollisionBoundary.test.ts: (cd packages/web-shell) npx --no-install vitest run ./client/components/sidebar/sessionDetailsCollisionBoundary.test.ts
per-file results (P=pass F=fail I=infra-exit, one letter per run):
packages/cli/src/agent-view/supervisor-runner.test.ts: PPPPP
packages/web-shell/client/components/SessionOverviewPanel.test.tsx: PPPPP
packages/web-shell/client/components/sidebar/SessionDetailsTooltip.test.tsx: PPPPP
packages/web-shell/client/components/sidebar/sessionDetailsCollisionBoundary.test.ts: PPPPP
verdict: pass
summary: 4 changed test file(s) x 5 identical rounds, no divergence
--- per-invocation detail (full copy in the artifact) ---
round 1 · packages/cli/src/agent-view/supervisor-runner.test.ts: P (exit 0)
round 1 · packages/web-shell/client/components/SessionOverviewPanel.test.tsx: P (exit 0)
round 1 · packages/web-shell/client/components/sidebar/SessionDetailsTooltip.test.tsx: P (exit 0)
round 1 · packages/web-shell/client/components/sidebar/sessionDetailsCollisionBoundary.test.ts: P (exit 0)
round 2 · packages/cli/src/agent-view/supervisor-runner.test.ts: P (exit 0)
round 2 · packages/web-shell/client/components/SessionOverviewPanel.test.tsx: P (exit 0)
round 2 · packages/web-shell/client/components/sidebar/SessionDetailsTooltip.test.tsx: P (exit 0)
round 2 · packages/web-shell/client/components/sidebar/sessionDetailsCollisionBoundary.test.ts: P (exit 0)
round 3 · packages/cli/src/agent-view/supervisor-runner.test.ts: P (exit 0)
round 3 · packages/web-shell/client/components/SessionOverviewPanel.test.tsx: P (exit 0)
round 3 · packages/web-shell/client/components/sidebar/SessionDetailsTooltip.test.tsx: P (exit 0)
round 3 · packages/web-shell/client/components/sidebar/sessionDetailsCollisionBoundary.test.ts: P (exit 0)
round 4 · packages/cli/src/agent-view/supervisor-runner.test.ts: P (exit 0)
round 4 · packages/web-shell/client/components/SessionOverviewPanel.test.tsx: P (exit 0)
round 4 · packages/web-shell/client/components/sidebar/SessionDetailsTooltip.test.tsx: P (exit 0)
round 4 · packages/web-shell/client/components/sidebar/sessionDetailsCollisionBoundary.test.ts: P (exit 0)
round 5 · packages/cli/src/agent-view/supervisor-runner.test.ts: P (exit 0)
round 5 · packages/web-shell/client/components/SessionOverviewPanel.test.tsx: P (exit 0)
round 5 · packages/web-shell/client/components/sidebar/SessionDetailsTooltip.test.tsx: P (exit 0)
round 5 · packages/web-shell/client/components/sidebar/sessionDetailsCollisionBoundary.test.ts: P (exit 0)
Evidence images
Harness scripts and raw logs are in the workflow run artifacts (7-day retention).
— Qwen Code · sandboxed verification
qqqys
left a comment
qqqys
left a comment
There was a problem hiding this comment.
Critical-only review at head 8bee8cca.
Previously blocking findings — all verified fixed in the code at this head
| Finding | Verification at 8bee8cca |
|---|---|
R6-1 (full workspace path falsified the untouched split-persist assertion; smoke red) |
SessionDetailsTooltip.tsx:274-276 renders and titles workspaceLabel ?? folderPath; web-shell.split-persist.spec.ts now asserts details.getByText(WORKSPACE_CWD, { exact: true }), so both specs are satisfied by the same contract. web-shell.collapsed-groups-persist.spec.ts still locates the cell by title, which the popover keeps. Status check rollup at this head: SUCCESS, no failing or pending run. |
| R5-1 (background work classified as Idle, enabling Archive/Delete on a working session) | deriveSessionCards (SessionOverviewPanel.tsx:224-226) now reads (session.hasActivePrompt ?? status?.hasActivePrompt) || session.activeWorkState === 'active' behind needsApproval/askUserQuestion, and does not fold 'unknown' in. sessionDetailsProps:566-568 derives the popover flags from card.status, so the column, the filter counts, the popover ladder and the card.status === 'idle' gates on canArchiveCard/canDeleteCard now consult one source. |
R5-2 (pinned popover unmounted with focus, dropping activeElement to <body>) |
Relocated into the tooltip's own layout cleanup as the finding required: SessionDetailsTooltip.tsx:132-149 runs on unmount, pierces shadow roots, and hands focus to [data-web-shell-session-panel], which exists with tabIndex={-1} at SessionOverviewPanel.tsx:1644-1648. The panel effect no longer has to win a race against the row's removal. |
| R5-3 (Details button destroyed an in-progress rename via Radix autofocus) | The actions-column trigger is disabled={Boolean(editingIdentity)} (SessionOverviewPanel.tsx:1312), so rename and click-pinned details are mutually exclusive in both directions (startRename still clears detailsOpen at :782). The mousedown guard's pinned contract is untouched. |
R4-1 (focus handoff resolved through ownerDocument.activeElement, retargeting to the shadow host) |
Both the handoff (:167-177) and the unmount cleanup (:139-146) now walk active.shadowRoot.activeElement before matching, and the rescue matches only popovers carrying this instance's data-web-shell-session-details-content owner token, which also closes the fix-induced R2-5. |
| R3-1 (first in-row control click swallowed while a draft is open) | The capture guard exempts only input (SessionOverviewPanel.tsx:1650-1659), so a pressed button/link no longer blurs the draft and remounts before mouseup; onRowClick still bails out while editingCard is set or a selection is active (:1717-1721). |
R1-8 / R2-1 (hoisted detailsOpen slot not reconciled) |
startRename clears the slot, and the reconcile effect is gated on clampPending so a shrinking catalog cannot clear state for a row that is still about to be visible (:1441-1455). |
Critical-only scan of the current diff
Read the full diff plus the current contents of SessionOverviewPanel.tsx, SessionDetailsTooltip.tsx, sessionDetailsCollisionBoundary.ts, i18n.tsx and the touched Playwright specs. No blocking correctness, security, data-loss, regression or compatibility defect is provable at this head:
- Row navigation is bounded:
onRowClickignores clicks while renaming or while a text selection is active, checkbox and control clicks stay selection/action-only, and archive/delete/split keep their existing per-card and cross-workspace gates. - Every locale key the new UI reads (
sessionsOverview.filter.*,sessionsOverview.status*,sessionsOverview.details,sidebar.activeWork,sidebar.activityUnknown,sidebar.clients) exists in both the English and Chinese tables. - The follow-up commit is test-only (
supervisor-runner.test.tsteardown, split-pane details assertion); production behavior is unchanged and the check rollup is green.
All 38 review threads on this PR are marked resolved, and each previously blocking one was re-verified against the code above rather than against the flag. The remaining reported items are Suggestion-level (truncation recovery, extra coverage, visual capture of the popover, overscroll containment) and are tracked in #11390; they do not block merge.
|
🤖 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 review round — no action neededNo actionable feedback was found for this round, so no changes were made. What was checked:
The PR remains as it stands, awaiting the in-flight sandboxed verification report and further human review. 中文说明Autofix 审查轮次 —— 无需处理本轮未发现需要处理的反馈，因此未做任何改动。 已检查的内容：
该 PR 保持现状，等待正在进行中的沙箱验证报告以及进一步的人工审查。 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 |
yiliang114
left a comment
yiliang114
left a comment
There was a problem hiding this comment.
Approving as the second vote. qqqys's Critical-only re-verification at this head covers the seven previously-blocking findings (R1-8/R2-1 detailsOpen reconciliation, R3-1 draft-swallow guard, R4-1 shadow-host focus handoff, R5-1 idle classification of active work, R5-2 unmount focus drop, R5-3 rename/Details mutual exclusion, R6-1 workspace-path assertion), each re-checked in the code rather than by thread flag, and I concur.
My independent scan of the current diff found no blocking defect: row navigation is bounded (rename/selection/checkbox all short-circuit onRowClick), archive/delete keep their per-card and cross-workspace gates, and every locale key the new UI reads exists in both en and zh tables. The follow-up commit is test-only. All 38 threads are resolved; the remaining Suggestion-level items are tracked in #11390 and don't block.
CI is fully green (Lint, Test, web-shell E2E Smoke, Capture visuals, no-AK integration, OpenTUI no-flicker, TUI parity, review-pr).
Stale — filed against a prior head. All blocking findings are fixed at 8bee8cc (qqqys Critical-only verification) and two approvals are on the current head.
ytahdn
left a comment
ytahdn
left a comment
There was a problem hiding this comment.
主旨总结 / PR Summary
中文 — 本 PR 重构 Web Shell 的「会话总览」面板，让它更易扫读和导航：
- 每个会话在标题下方直接展示工作区、分支和 PR，并把「待审批 / 需用户输入 / 运行中 / 空闲」拆成互不相同的状态（此前审批、提问、运行共用一个 spinner，无法区分谁被卡住）；新增状态筛选按钮与「标题/分支/PR 号(含
#123)/会话 ID」搜索。 - 标题悬停复用侧栏那套完整详情浮层（完整标题、工作区路径、会话 ID、状态、PR/issue 链接、可复制 ID），并新增一个显式「详情」按钮供键盘/触摸访问；窄屏时在标题旁保留紧凑状态图标。
- 交互契约改为：点击行=打开会话，勾选框=选择，有选择时才出现批量操作（此前点击行是切换选择）。
- 覆盖模块：
SessionOverviewPanel.tsx、sidebar/SessionDetailsTooltip.tsx、sessionDetailsCollisionBoundary.ts、i18n.tsx及一处 CLI 测试基建修复（supervisor-runner.test.ts），并配套大量单测/E2E/视觉用例。
English — This PR reworks the Web Shell Session Overview panel for scannability and navigation:
- Each session now shows workspace, branch and PR under its title, with distinct approval / question / running / idle states (previously approval, question and running shared one spinner, so blocked work was indistinguishable). Adds status filter buttons plus title/branch/PR-number(incl.
#123)/session-ID search. - Title hover reuses the sidebar's full details popover (full title, workspace path, session ID, status, PR/issue links, copyable ID), plus an explicit details button for keyboard/touch; compact status cues stay beside titles on narrow layouts.
- The interaction contract changes to: click a row = open the session, checkboxes = selection, batch actions appear only when a selection exists (previously a row click toggled selection).
- Modules touched:
SessionOverviewPanel.tsx,sidebar/SessionDetailsTooltip.tsx,sessionDetailsCollisionBoundary.ts,i18n.tsx, plus one CLI test-infra fix (supervisor-runner.test.ts) and extensive unit/E2E/visual coverage.
结论 / Verdict: APPROVE
中文 — 我按纯静态方式审查了 head（8bee8ccad）的生产代码，并对每处跨文件影响做了 head tree 复核，未发现 Critical 或 Important 问题。以下是我特意核实并排除的几处疑点，以及两条非阻塞 Nit。
已核实、确认非缺陷：
resolveSessionDetailsCollisionBoundary签名语义变更：参数由sidebar改为anchor，回退逻辑由?? sidebar改为?? closest('[data-web-shell-root]') ?? closest('aside') ?? null。全仓唯一另一个调用方WorkspaceDetailsTooltip.tsx（本 PR 未改）传入的正是anchorRef.current?.closest('aside')；对<aside>元素而言新旧回退结果完全一致（aside.closest('aside')命中自身），故其行为未变。SessionDetailsTooltip.tsx自身调用点也已同步更新。- 删除的 i18n key（
sessionsOverview.folder / worktree / sessionId）：head tree 中已无任何引用。 - 新增 i18n key（
status.idle/running/needsApproval/askUserQuestion、filter.*、details、sessionCount、workspaceAll、workspacesSelected、statusColumn、statusFilter）：EN 与 ZH 均已齐备；模板字面量 keysessionsOverview.status.${status}的四种取值全部存在（这类 key tsc 静态查不到，专门核对过）。 - 删除的
worktreeOnlyprop 与SessionIdCell：head 无残留引用。 - 选择/打开职责分离：
select列与actions列均带stopRowClick: true，勾选框/详情按钮不会误触发「打开会话」；onRowClick在重命名中或存在文本选区时提前 return，避免拖选元数据时误开。 .attention { color: var(--warning-color) }：--warning-color是既有主题变量，与仓库多处 module.css 用法一致。- columns
useMemo依赖数组：与 memo 体内实际引用一致，无遗漏。 - 焦点管理（受控 open + ownerToken 单例替换、
useLayoutEffect卸载兜底把焦点交还[data-web-shell-session-panel]（本 PR 已给该根加tabIndex={-1}）、restoreFocusRef与onCloseAutoFocus的配合、shadowRoot activeElement 递归下钻）逻辑自洽，且有针对性单测/E2E（含 shadow-dom 用例）覆盖。
Nit（非阻塞，可选）：
- PR 号搜索用
`#${pr.number}`.includes(query)，短数字查询会宽泛命中（如1命中#123）。这与既有的标题/ID 子串搜索行为一致，可接受；若希望精确按 PR 号定位，可考虑数值相等或前缀匹配。 detailsOpenRef.current = detailsOpen（渲染期写 ref 镜像 state）是社区常见的 React 反模式；此处因在同一次渲染体内、子单元格读取之前同步赋值而安全，但仍可考虑用 effect 同步或直接读 state 以更显式。
CI：发布时全绿 —— Test (ubuntu, Node 22)、Lint & Static、web-shell E2E Smoke、Integration Tests (no-AK)、Capture web-shell visuals、Desktop Shell 均 pass；其余为 skipping 的自动化 review/authorize 作业。
English — I reviewed the head (8bee8ccad) production code statically and cross-checked every cross-file impact against the head tree. No Critical or Important issues. Below are the specific suspicions I verified and cleared, plus two non-blocking nits.
Verified as non-defects:
resolveSessionDetailsCollisionBoundarysignature/semantics change: the param went fromsidebartoanchor, and the fallback from?? sidebarto?? closest('[data-web-shell-root]') ?? closest('aside') ?? null. The only other caller repo-wide,WorkspaceDetailsTooltip.tsx(untouched by this PR), passesanchorRef.current?.closest('aside'); for an<aside>element the old and new fallbacks resolve identically (aside.closest('aside')matches itself), so its behavior is unchanged.SessionDetailsTooltip.tsx's own call site was updated accordingly.- Removed i18n keys (
sessionsOverview.folder / worktree / sessionId): no remaining references in the head tree. - New i18n keys (
status.idle/running/needsApproval/askUserQuestion,filter.*,details,sessionCount,workspaceAll,workspacesSelected,statusColumn,statusFilter): present in both EN and ZH; all four values of the template-literal keysessionsOverview.status.${status}exist (tsc can't statically check these, so I verified by hand). - Removed
worktreeOnlyprop andSessionIdCell: no dangling references at head. - Selection vs. open separation: both the
selectandactionscolumns carrystopRowClick: true, so checkbox/details clicks don't trigger "open session";onRowClickearly-returns while renaming or when a text selection exists, avoiding accidental opens during metadata drag-select. .attention { color: var(--warning-color) }:--warning-coloris an existing theme variable, consistent with its use across many module.css files.- columns
useMemodependency array: matches the values actually referenced inside; nothing missing. - Focus management (controlled open + single-owner-token replacement, the
useLayoutEffectunmount rescue handing focus back to[data-web-shell-session-panel]— which this PR givestabIndex={-1}— therestoreFocusRef/onCloseAutoFocusinterplay, and recursive shadowRoot activeElement descent) is internally consistent and covered by targeted unit/E2E tests (including a shadow-dom case).
Nits (non-blocking, optional):
- PR-number search uses
`#${pr.number}`.includes(query), so short numeric queries match broadly (e.g.1hits#123). This is consistent with the existing title/ID substring behavior and acceptable; if precise PR lookup is desired, consider numeric-equality or prefix matching. detailsOpenRef.current = detailsOpen(writing a ref to mirror state during render) is a commonly-cited React anti-pattern; it's safe here because it's assigned synchronously in the same render body before child cells read it, but an effect-synced ref or reading state directly would be more explicit.
CI: green at publish time — Test (ubuntu, Node 22), Lint & Static, web-shell E2E Smoke, Integration Tests (no-AK), Capture web-shell visuals, and Desktop Shell all pass; the rest are skipping automation review/authorize jobs.
|
Released in v0.23.3. |
6 participants
What this PR does
Make the session overview easier to scan and navigate. Each session now shows its workspace, branch and PR below the title, with distinct approval, question, running and idle states. Status filters and branch/PR search help locate sessions that need attention. Compact status icons remain beside pinned titles when narrow layouts hide the middle columns.
Reuse the sidebar's full details popover on title hover and provide a details button for keyboard and touch access. The popover shows the complete title, workspace path and session ID, supports copying the ID and opening linked PRs/issues, and keeps these actions separate from row navigation. Long details scroll within the available viewport height so every field and the copy action remain accessible. Opening another details popover replaces the previous one, and keyboard focus stays visible even when the details overflow. If hover replaces focused details, focus moves to the incoming title so Tab continues in the table; ordinary hover preserves a focused search field, including when details portals live in a ShadowRoot. Clicking a row opens its session; checkboxes control selection, and batch actions appear when a selection exists.
Why it's needed
Previously, approval requests, questions and running sessions shared a spinner, making blocked work hard to distinguish. Separate metadata columns competed with session titles, and clicking different parts of a row alternated between opening and selecting. The overview also exposed only the Git portion of the sidebar details, so users could not inspect the same information from both places.
Reviewer Test Plan
How to verify
Open Session Overview with approval, question, running and idle sessions. Confirm distinct labels/icons, with a spinner only for running sessions. Combine status and workspace filters with title, branch, PR number (including
#123) or ID search; changing a filter should clear selection and return to the first page.Hover a title to inspect full details. Open the explicit details button using Enter, copy the session ID, and follow a PR or issue link. These actions should keep the overview open and leave selection unchanged. Escape should close explicit details and return focus to its button. In a small viewport, long titles should wrap completely, the popover should stay within Web Shell, and scrolling should reach the copy action.
Click a row to open its owning session. Dragging to select metadata should keep the overview open; clicking a plain cell while renaming should preserve the draft, with Enter to save and Escape to cancel. Select checkboxes to reveal batch actions and clear selection to hide them. Pagination should remain available, and existing restrictions on active sessions and cross-workspace operations should remain in force.
While renaming, click Time once. Sorting should cancel the unsaved rename, focus Time, and let Tab continue to the first row checkbox, whether the edited session stays visible or moves off the page. In a Shadow DOM host, open Details and hover another title: the incoming title should retain focus after the old popover closes, and Tab should continue within that row. Hovering from a focused search field or from a different popover owner should preserve that focus.
Start renaming immediately after hovering a title, then cancel: details should stay closed. While renaming, a live update that filters or reorders the row off the current page should discard the hidden draft and keep other rows usable. A catalog shrink that moves the current session onto the last page should preserve its visible details or draft. At 499px width, scroll the table to both ends and confirm each active/attention row keeps its title icon.
Evidence (Before & After)
Screenshots use synthetic sessions from the browser regression tests. The before image includes a selected row; the after image shows the overview with selection cleared.
The shared details popover with the complete workspace path, PR/issue links, status and copyable session ID:
Light theme and narrow-screen details
Light theme at 1440px:
At 499px, pinned title cues keep each active/attention row identifiable:
Details at 499px, with the complete branch wrapping and the copy action accessible:
Observed against the globally installed
qwenbaseline and the locally built Web Shell, using deterministic daemon responses for four session states:Validation: repository build, typecheck and bundle passed; 220 focused unit tests, all 67 Chromium smoke cases and both dark/light visual cases passed. Fourteen affected regressions passed against the actual bundled CLI, with four additional independent cases covering light/shadow portal ownership and search focus, and sorting during rename with 4/11 sessions. The shadow host adapter only enables the public Shadow DOM option in the standalone host props; product logic remains identical to the bundle. Two targeted deliberate regressions failed at the intended focus assertions, and sources were restored byte-for-byte. ESLint, Prettier and two consecutive clean reverse-audit passes completed before commit. E2E details are posted in a separate comment.
Tested on
Environment (optional)
Node.js 22.22.2; isolated local daemon launched with
node dist/cli.js serve --web --no-open; Chromium with mocked daemon responses for reproducible UI state. Baseline used globalqwen0.23.0.Risk & Scope
Linked Issues
Follow-up coverage, the existing footer Archive edge-click issue, and the rename draft retention design are tracked in #11390.
中文说明
此 PR 的作用
让会话总览更便于浏览和导航。每个会话的工作区、分支和 PR 显示在标题下方，并明确区分待审批、待回答、运行中和空闲状态。状态筛选以及分支、PR 搜索帮助定位需要处理的会话。窄屏中间列被遮挡时，固定标题旁的紧凑状态图标仍可见。
在标题悬浮时复用侧栏的完整详情浮窗，并提供支持键盘和触屏的详情按钮。浮窗展示完整标题、工作区路径和会话 ID，支持复制 ID 和打开关联 PR、Issue，且这些操作不会触发行导航。详情过长时在视口可用高度内滚动，所有字段与复制操作仍可访问。打开另一个详情时替换已有浮窗，即使内容溢出也保持键盘焦点可见。悬浮替换正在使用的详情时，焦点移到新标题，Tab 从该行继续；普通悬浮保留搜索框的焦点，包括详情门户位于 ShadowRoot 的嵌入场景。点击行打开对应会话，复选框控制选择，有选择时才显示批量操作。
为什么需要
此前，审批请求、提问和运行中的会话使用相同的加载动画，难以分辨等待用户处理的工作。独立元数据列挤占标题空间，点击一行的不同位置分别执行打开或选择操作。总览也仅展示侧栏详情中的 Git 信息，无法在两处查看相同信息。
审阅者测试计划
如何验证
在会话总览中准备待审批、待回答、运行中和空闲会话，确认状态文字和图标可区分，仅运行中显示加载动画。组合状态、工作区筛选与标题、分支、PR 编号（包括
#123）或 ID 搜索；修改筛选时应清空选择并返回第一页。悬浮标题查看完整详情。用 Enter 打开详情按钮，复制会话 ID，并访问 PR 或 Issue 链接。这些操作应保留总览页面且不改变选择。Escape 应关闭显式打开的详情，并将焦点返回详情按钮。在小视口中，长标题应完整换行，浮窗应保持在 Web Shell 内，并可通过滚动到达复制按钮。
点击行打开其所属会话。拖选元信息应保留总览；重命名期间点击普通单元格应保留草稿，Enter 保存、Escape 取消。勾选复选框时显示批量操作，清空选择时隐藏。分页保持可用，运行中会话和跨工作区操作的原有限制继续生效。
重命名期间点击一次“时间”：应执行一次排序、取消未保存草稿并聚焦“时间”，Tab 继续到首行复选框；编辑行仍可见和被排到下一页两种情况都应成立。在 Shadow DOM 宿主中打开详情再悬浮另一标题：旧浮窗关闭后，新标题仍应保持焦点，Tab 从该行继续。搜索框已有焦点或不同浮窗实例间悬浮时，原焦点应保持。
悬浮标题后立即开始重命名再取消，详情应保持关闭。实时更新通过筛选或重排将编辑行移出当前页时，应丢弃隐藏草稿且其他行仍可操作；目录收缩后若该会话仍在最终页，则保留其详情或草稿。在 499 像素宽度下横向滚到两端，运行中和待处理行的标题图标应始终可见。
修改前后证据
截图使用浏览器回归测试中的模拟会话。修改前截图包含一个已选中行；修改后截图展示清空选择后的总览。
共享详情浮窗包含完整工作区路径、PR/Issue 链接、状态和可复制的会话 ID：
浅色主题与窄屏详情
1440 像素宽度下的浅色主题：
499 像素宽度下，固定标题旁的图标仍可辨认每个运行中和待处理会话：
499 像素宽度下的详情，完整分支名换行且复制按钮可访问：
使用全局安装的
qwen作为基线，与本地构建的 Web Shell 对比；通过确定的 daemon 响应构造四种会话状态：验证结果：整库 build、typecheck、bundle 通过；220 项定向单元测试、全部 67 项 Chromium smoke 及两项深浅色视觉用例通过。14 项受影响回归在实际打包 CLI 上通过，另有 4 项独立用例验证普通/Shadow DOM 浮窗的归属和搜索保焦，以及 4/11 个会话下重命名期间的排序。Shadow DOM 宿主适配仅启用公开的宿主选项，产品逻辑与构建产物相同。两项刻意回退均在预期焦点断言失败，源码随后逐字节恢复。ESLint、Prettier 及提交前连续两轮无新问题的反向审计均完成。E2E 详情另发评论。
已测试平台
环境
Node.js 22.22.2；通过
node dist/cli.js serve --web --no-open启动隔离的本地 daemon；Chromium 使用模拟 daemon 响应复现 UI 状态。基线为全局qwen0.23.0。风险与范围
关联 Issue
测试补强、既有的页脚 Archive 底缘漏点击问题及重命名草稿保留规则，记录于 #11390。
CI follow-up
Align the split-pane details regression with the complete workspace path shown by the shared popover. The supervisor tests now wait for startup and shut down their live supervisor before removing its temporary directory, preventing a timed-out attempt from leaving an asynchronous write behind during retries. Production behavior is unchanged by this follow-up.
Validation: build, typecheck and bundle; 55 related CLI unit tests; all 70 Chromium smoke cases; all 5 split-pane cases against the actual bundled CLI. A first-start delay beyond the readiness deadline now retries successfully without an unhandled rejection, while a real filesystem write failure still fails the test. Cleanup also retains the authenticated token of the runner it started, so replacing metadata before an assertion failure cannot leave the old runner alive. Two consecutive reverse-audit passes completed before commit. The detailed CI reproduction and verification report is posted separately.
CI 修复：分屏详情测试改为验证完整工作区路径；supervisor 测试在清理临时目录前关闭并等待实例退出，避免超时重试留下未完成的写入。此次修改仅涉及测试。build、typecheck、bundle、55 项 CLI 单测、70 项 smoke 及实际打包 CLI 上的 5 项分屏用例通过；延迟启动重试不再产生未处理拒绝，真实写入错误仍使测试失败。清理保留已验证的原实例 token，元数据替换后的提前失败也能关闭原实例。提交前完成连续两轮反向审计。