[LIHUA919]

Goal And Delivered Outcome

• Outcome basis / optional anchor: the required dashboard-browser job of Python Tests (feeds dashboard-acceptance → checks → merge-gate) is red on main.
• Goal/source and gap: [Bug]: the required dashboard-browser lane is red on main — a steward view switch drops the live turn's elapsed receipt #7384 reported two shard failures from feat(chat): open the selected original steward from App Conversation #7320. Later main commits already repaired those two (conversation-activity's elapsed receipt and composer-session-admission now pass), and test(workspace): preserve fixture failures and observe paged history #7494 made the managed fixture report its own failures. Running the shards on current main exposes the next failures that keep the lane red:
  1. Shard 1, managed-agent-sessions: managed_agent_fixture.py imports pytest, but the dashboard-browser job installs the checkout without test extras (pip install -e .). The fixture exits before the scenario starts. Before test(workspace): preserve fixture failures and observe paged history #7494 this surfaced as an unhandled EPIPE; now it surfaces as ModuleNotFoundError: No module named 'pytest'. Both PRs that introduced and changed this scenario (feat(chat): open and continue original managed Agent Sessions #7284, fix(chat): retain one observer for active managed turns #7354) had dashboard-browser skipped, so it has not passed in hosted CI before.
  2. Shard 2, goal-workspace-m2: two races. The scenario counts visible work immediately after toggling every owner group (hosted: 126 !== 200, "All work remains reachable in the original owner order"). It also focuses the drawer's help before the drawer's next-frame focus handoff to its title, which blurs and dismisses the tooltip (local: about 1 in 3 runs at line 261).
  3. Shard 2, scenarios hosted CI had not reached (the shard stopped at goal-workspace-m2). Each one asserts UI that a later, deliberate product change moved:
     • research-finding-reader: fix(research): restore historical decision reading context #7501 keeps the decision a reader just closed expanded and focused, so the question no longer collapses to three records after an older finding was read (20 !== 3). fix(research): restore historical decision reading context #7501's own dashboard-browser shards were cancelled. The scenario now asserts that shipped contract: the decision just read stays in view with focus. The default three-record view stays covered by research-history (lines 182-183).
     • goal-suggestions: since 63784cfa9, the work vital shows partial coverage in its contextual help (labelled with the unknown total), not as inline text. The scenario now opens that help and asserts the same "已列出 5 项中的 0 项" counts in the tooltip.
• Observable before → after, with the validation row that proves it: before, both scenarios fail as above; after, both pass repeatedly (see Validation). The decisive check is this PR's own hosted dashboard-browser (1/2/3) run.
• Issue/task and intended base: Related to [Bug]: the required dashboard-browser lane is red on main — a steward view switch drops the live turn's elapsed receipt #7384; base main.

Author Declaration

• Written by: model_agent (Claude Opus, Anthropic), at the account owner's direction.

Implemented against

• Specification and revision: [Bug]: the required dashboard-browser lane is red on main — a steward view switch drops the live turn's elapsed receipt #7384 "Expected behavior" (the required dashboard-browser job is green on main for an unchanged checkout), plus each scenario's own stated assertions, which stay unchanged.
• Criteria:

Criterion (spec clause)	Disposition	Symbol / path	Test or command
Managed-session acceptance runs in the lane's install (no test extras)	implemented	examples/personal-workspace-browser/managed_agent_fixture.py (ProcessPatch)	LOOPX_PERSONAL_WORKSPACE_SCENARIO=managed-agent-sessions node examples/personal-workspace-browser-smoke.mjs with no pytest installed
"All work remains reachable" is checked after the expansion actually renders, still within the 2s budget	implemented	examples/personal-workspace-browser/goal-workspace-m2.mjs	LOOPX_PERSONAL_WORKSPACE_SCENARIO=goal-workspace-m2 node examples/personal-workspace-browser-smoke.mjs
Drawer help tooltip is exercised after the drawer's documented focus handoff	implemented	same file, verifyRemoteSelection	same command
Assertions follow deliberate UI moves without weakening the fact checked	implemented	research-finding-reader.mjs, goal-suggestions.mjs	single-scenario runs, 2/2 each
Lane green on hosted runners	deferred to this PR's CI	—	dashboard-browser (1/2/3)

• Self-check before submission: reproduced each failure on the base before changing it (negative control below), reran the changed scenarios in an environment matching the lane (checkout installed without test extras), ran all three shards locally, and ran the upstream managed-fixture-process.test.mjs. Deliberately left out: product code changes (see Scope).

Scope And Continuation

• Completed scope and remaining work: test- and fixture-only. The scenarios' assertions and budgets are unchanged. Only what they wait for changed: the fixture's import surface, a poll until all 200 items are visible inside the existing 2s budget, and a wait for the drawer title to take focus.
• Not in this PR, found while reproducing:
  • Possible product race (macOS, local): in conversation-activity, right after switching the Goal view Tasks → Chat, the live Turn's pending receipt is briefly absent: the Chat-exit cleanup retires it and recovery re-creates it only after createChatSession. A Command+Enter steer in that gap takes the ordinary-send path and clears the composer draft. It reproduced in about 3 of 4 local macOS runs at conversation-activity.mjs:253, but hosted Linux shard 1 passes this scenario. A real fix changes composer turn-admission or Goal-view observer retention, so it needs its own review rather than a test change here.
  • Local-only flakes that pass on hosted runners: research-worker-readback (ECONNRESET from the disposable backend in my environment) and research-record-search (focus, passes in isolation).
• Slice boundary / successor: if this PR's hosted shards are green, [Bug]: the required dashboard-browser lane is red on main — a steward view switch drops the live turn's elapsed receipt #7384 can close. A follow-up issue for the composer race above is proposed, not yet filed.

Validation

• Tested revision: bae64527c (on main 1b6dd6a42). An earlier hosted run on a73a070c8 passed dashboard-browser (1) and (3) and failed (2) at research-finding-reader, which the later commits fix.
• Run state: finished (local); hosted CI running on this PR.
• Input classes: synthetic

Check kind	Result	Public-safe evidence / limitation
regression_parity	passed	Negative control on the base: managed-agent-sessions fails with ModuleNotFoundError: No module named 'pytest' when the checkout is installed without test extras; with this branch it passes (2/2).
real_entrypoint	passed	goal-workspace-m2: before the fix, fails 1 of 3 local runs (line 261); after, 4/4 and 3/3 on the rebased head. The hosted line-155 failure did not reproduce locally; the fix waits for the same visible count the assertion requires.
real_entrypoint	passed	Full local shards with the fixes: shard 1 passes all 27 scenarios (with a local-only wait for the macOS race above, not committed); shards 2/3 stop only at the local-only flakes listed in Scope, which pass on hosted runners.
real_entrypoint	passed	Shard-2 scenarios after the previous hosted stop point, run individually: research-finding-reader (2/2), goal-suggestions (2/2), research-display-language, replan-cadence, since-you-looked, first-goal-agent-unavailable, attention-notifications, completed-history-isolation.
unit	passed	node --test examples/personal-workspace-browser/managed-fixture-process.test.mjs (6/6).
static	passed	py_compile; node --check; scripts/generate_semantic_inventory.py --changed-from HEAD reports no new vocabulary carriers.
real_entrypoint	running	Hosted dashboard-browser (1/2/3) on this PR is the decisive check.

• Coverage and gaps: the changed paths are the scenarios' waits and assertion targets plus the fixture's process setup, and both are exercised by the scenario runs above. Hosted CI is the only evidence for Linux timing. Shard 2's scenarios after goal-workspace-m2 had not run on hosted main before (the shard stopped there), so new failures there are possible and will be addressed on this PR if they appear.

Frontend / Visual Evidence

• UI impact: none (test harness and fixture only).

Type of Change

• Test update

LoopX Area

• Build, packaging, installer, or CI

Shared-authority RFC fixture impact

N/A

Boundary Checklist

• Neither the diff nor this PR body/comments/attachments disclose private state, credentials, raw traces or verifier output, internal links, or local machine paths.
• I did not duplicate maintainer-owned benchmark work.
• I kept the change scoped to the linked issue/task.
• No paired RFC changed.
• UI impact none.
• Every commit includes a DCO Signed-off-by trailer.

🤖 Generated with Claude Code

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=max

APPROVE，精确 head bae64527cca4e3378a29162b54ef121e163cc753。

动机 / Motivation

维护打包工作区回归的开发者，以及依赖这些检查发现界面、会话与权限退化的用户。无 test extras 的浏览器任务无法启动托管会话 fixture；异步渲染和焦点交接尚未完成时读取 DOM 会产生竞态，研究返回与部分覆盖断言还停留在旧界面。原打包入口可启动真实隔离 SQLite/Chat fixture；测试在原预算内等到实际渲染／焦点状态，再核对准确旧结论与部分工作数量，而不把展开恢复当成初始折叠。没有产品代码、UI、权限、默认值、状态协议、provider 或模型调用变化；不以测试修复宣称新产品能力或最终发布资格。 这是修复发布验证入口的独立批次，不能把恢复测试运行本身当成最终发布或新功能证明。

改动思路 / Approach

复用既有五个场景、真实状态 owner 与本机浏览器入口，只移除 fixture 的测试依赖并修正观察时点和语义位置。 按 1b6dd6a42ab7f9620a4655061952adb58038a65b 的 docs/development/good-smokes.md 核对 NO-TEST-EXTRAS、BOUNDED-RENDER、DRAWER-FOCUS、EXACT-RESEARCH-CONTEXT、PARTIAL-COVERAGE 五项判据。先判断产品实际契约，再修复测试，不让旧快照或测试运行器依赖反过来决定产品语义。真实 Chat/SQLite、原字段、权限负例、完整记录和原两秒预算保留；不需要新增 UI、状态 owner 或通用等待框架。

具体改动 / Changes

四个原文件共 +27/-7。托管会话 fixture 用仅含 setattr/setenv 的进程级 adapter 复用 isolate_sqlite_runtime；临时目录、环境和 tempfile cache 的生命周期仍归该独立 fixture 进程。禁止导入 pytest 的同条件对照中，基线明确 ModuleNotFoundError，head 启动实际服务器并读回三个原受管会话。不是靠给浏览器任务补 test extras 隐藏依赖。

200 项工作展开等待实际 DOM 收敛，再沿原性能时钟检查总耗时严格小于 2 秒；轮询最长 5 秒不会放宽这个通过条件。Drawer 本来在下一帧将焦点交给标题，现在先观察 title 实际获得焦点，再打开并核对原 tooltip，避免被后续焦点交接关闭。部分覆盖仍核对同一五项中的零项，按真实 masthead 的未知总数帮助入口读取，没有删除数量或把未知说成完整。

研究返回沿 TrailQuestion 的 selectedIndex 展开到刚读过的旧结论；此时要求三条会与恢复原记录和焦点的契约矛盾。head 核对具体「比较 2」按钮的焦点，research-history 仍独立断言初始问题数和每题记录数均为三，旧结论选择、刷新、Back、窄屏和全文断言保留。两个研究 UI owner 的 Git blob 与实际触发 20≠3 的 early3 版本完全相同，差异来自测试观察而非产品改写；实际桌面与窄屏截图也已检查。

精确 head 上五个完整 packaged 场景通过，包括四个受影响场景和 research-history 的初始三条覆盖；禁止 pytest 导入仍得到真实三会话 ready，基线在同条件明确拒绝。premerge 5 direct+3 selected、规定 Ruff 范围、17 模块 mypy、三个 Node 语法检查及四路径公开边界通过。 首次辅助观察器误读启动的人类提示为 JSON，修复读取后同 head 通过，原失败保留。额外扩大到整个旧 fixture 的 Ruff 检查得到 10 条存量样式问题；相同命令比较 base/head 的 code、message 和出错原行完全一致，仅行位置改变。规定 Ruff 范围通过，不能将这个辅助失败改称整个旧文件 lint 通过。作者描述的 hosted CI 未作为独立验证，按 Goal 的 wait_for_ci=false 没有查询 CI。

对主干的风险 / Risks

主要风险是等待和定位变化掩盖迟到渲染或旧记录丢失。原 200 项与两秒断言、具体旧结论焦点、独立初始三条、未知覆盖的准确计数都保留，并在完整 packaged 场景上执行。ProcessPatch 不做进程内恢复，因此其正确范围是这个独立、由测试关闭的 fixture 进程，不能直接复制给共享 pytest 进程。源身份在三条资格 lane 前后未变且清理成功；没有真实用户 Goal、私有工作区或外部 provider 操作。

我的整体评价 / Verdict

这是合理的测试修复，复用现有真实 owner 和独立负例，移除了浏览器 fixture 不应依赖的 pytest，并把观察顺序与当前 UI 生命周期对齐。现有场景保留比增建镜像测试或放宽预算更直接的回归价值。同域简化已完成，无需无关格式清理、生产重构或通用重试层。正式发布仍须在包含其他主线修复的统一候选上完成全仓、完整打包目录、真实模型、安装和渠道验证。 当前批准只覆盖此精确测试修复提交，公开前后路径与 DCO 也已核对，正式发布的统一 source 资格仍单独完成。

English verdict: APPROVE bae64527cca4e3378a29162b54ef121e163cc753. Five complete packaged scenarios, a real no-pytest startup counterfactual and required local gates pass. The original two-second budget, exact older-record focus, fresh default-three view, partial counts and authority assertions remain. Ten optional file-wide lint diagnostics are unchanged at base/head; no product or provider behavior changes.