📋 Archive workspace acceptance and independent task reviews

This commit is contained in:
d-robotics 2026-09-18 10:08:29 +08:00
parent ab89c63a0f
commit a313b22ae8
28 changed files with 1459 additions and 15 deletions

View File

@ -1,6 +1,6 @@
# 专注标注工作页任务索引
用户已确认拆分并授权发布,共 8 份独立任务。正在按编号顺序实施并逐票独立审查;只有前置任务全部验收完成后才能领取。
用户已确认拆分并授权发布,共 8 份独立任务。已按编号顺序完成实施及逐票独立审查;最终全功能双轴审查另行记录。
父规格:[专注工作页布局规格](./spec.md)。设计记录:[map.md](./map.md)。已确认的拆分及覆盖矩阵:[ticket-breakdown.md](./ticket-breakdown.md)。布局示意:[preview.html](./preview.html)。
@ -15,7 +15,7 @@
| 05 | [固定区域内的字典和长文本编辑](./issues/05-inline-dictionaries-and-long-text.md) | 03 | resolved |
| 06 | [单路放大与视频比例适配](./issues/06-camera-inspection-and-aspect-fit.md) | 03 | resolved |
| 07 | [小窗口暂停与在途结果协调](./issues/07-undersized-viewport-recovery.md) | 01 | resolved |
| 08 | [连续工作整体验收与操作手册](./issues/08-full-workflow-acceptance-and-guide.md) | 04、05、06、07 | claimed |
| 08 | [连续工作整体验收与操作手册](./issues/08-full-workflow-acceptance-and-guide.md) | 04、05、06、07 | resolved |
## 执行顺序

View File

@ -1,6 +1,6 @@
# 08: 连续工作整体验收与操作手册
Status: claimed
Status: resolved
Type: task
Parent: [专注工作页布局规格](../spec.md)
@ -11,14 +11,14 @@ Parent: [专注工作页布局规格](../spec.md)
**覆盖:** 全部 30 条用户故事、L01–L16;本票负责组合验收和交付,不代替各前置任务自身的验收。
- [ ] 完成开始工作→已有任务优先/随机领取→切片→真实生成→描述→保存→送检→下一条的浏览器/API 闭环,验证停止、空池及未知结果恢复。
- [ ] 五档支持内容区覆盖两阶段、字典、长列表、长文本、单路放大、错误与确认弹层;主要操作无文档横向/纵向滚动即可完成。
- [ ] 代表性 1366×768、1440×900、1920×1080、2560×1440 显示/窗口设置下核对 100%、125%、150% 浏览器缩放,记录真实 CSS 内容区,按是否达到底线验证正常或暂停行为。确实无法覆盖的设备/系统缩放组合必须注明限制,不能用模拟视口冒充全部物理设备实测。
- [ ] 交叉验证草稿/撤销/在途请求情况下的新旧布局切换、刷新显式恢复、账号偏好隔离和单活动页规则。
- [ ] 原任务池入口、原布局及既有快捷键回归通过;检查新布局样式没有意外作用于质检/验收工作台。
- [ ] 完成适用的类型检查、构建与行为测试;既有失败与新增回归分开报告,前置票未通过的验收不能被最终报告掩盖。
- [ ] 交付说明书覆盖进入/退出、字典与长内容、临时放大、切回旧版、小窗口恢复、保存与提交、停止接续;明确“送检成功”不同于“质检通过”。
- [ ] 交付恢复说明和验证报告:确认代码标签/bundle可恢复,区分代码与数据库/视频/浏览器草稿;不进行公司集群部署或额外推送。
- [x] 完成开始工作→已有任务优先/随机领取→切片→真实生成→描述→保存→送检→下一条的浏览器/API 闭环,验证停止、空池及未知结果恢复。
- [x] 五档支持内容区覆盖两阶段、字典、长列表、长文本、单路放大、错误与确认弹层;主要操作无文档横向/纵向滚动即可完成。
- [x] 代表性 1366×768、1440×900、1920×1080、2560×1440 显示/窗口设置下核对 100%、125%、150% 浏览器缩放,记录真实 CSS 内容区,按是否达到底线验证正常或暂停行为。确实无法覆盖的设备/系统缩放组合必须注明限制,不能用模拟视口冒充全部物理设备实测。
- [x] 交叉验证草稿/撤销/在途请求情况下的新旧布局切换、刷新显式恢复、账号偏好隔离和单活动页规则。
- [x] 原任务池入口、原布局及既有快捷键回归通过;检查新布局样式没有意外作用于质检/验收工作台。
- [x] 完成适用的类型检查、构建与行为测试;既有失败与新增回归分开报告,前置票未通过的验收不能被最终报告掩盖。
- [x] 交付说明书覆盖进入/退出、字典与长内容、临时放大、切回旧版、小窗口恢复、保存与提交、停止接续;明确“送检成功”不同于“质检通过”。
- [x] 交付恢复说明和验证报告:确认代码标签/bundle可恢复,区分代码与数据库/视频/浏览器草稿;不进行公司集群部署或额外推送。
## Constraints
@ -27,3 +27,9 @@ Parent: [专注工作页布局规格](../spec.md)
以本地 5173/3001/27018 隔离任务进行真实浏览器操作和现有 API 保存回读;适用的类型检查、构建及行为测试须通过。原有失败与新增回归分别记录,不用源码字符串断言替代交互验收。公司集群/BOS 不在写入授权范围,凭据和原始证据不进入跟踪文件。
定义就绪不代表依赖完成;所有直接前置交付并验收后才能领取。发布不启动实现,也不授权推送或部署。
## Answer
已完成整体验收与中文说明书,提交 `89c9e168`;验收脚本的审核修复为 `ab89c63a`。最终同一份脚本完整运行退出 0,真实完成三条任务的生成、保存、送检及接续、停止、空池行为;12 组实际浏览器缩放和五档布局回归通过。独立 reviewer 已复审通过。
[整体验收报告](../../../docs/verification/annotation-workspace-layout/acceptance-results.md)保留原生系统输入法、物理设备缩放及本地验收媒体缺失的覆盖限制;[操作说明书](../../../docs/annotation-workspace-layout.md)可直接交给标注人员。执行范围未包括推送、合并或部署。设计决策见[设计记录](../map.md#decisions-so-far)。

View File

@ -6,7 +6,7 @@
## 证据如何阅读
“本轮”指 08 新跑的浏览器/API 或命令结果;“前置”指已经独立审核的 01–07 原始记录,不宣称重新执行。对应报告位于 `.superpowers/sdd/annotation-workspace-layout-plan/task-N-report.md`,最终归档由主任务处理。验收以可见交互、媒体几何、公共 API 回读和请求次数为边界,没有新增内部状态或源码字符串验收。
“本轮”指 08 新跑的浏览器/API 或命令结果;“前置”指已经独立审核的 01–07 原始记录,不宣称重新执行。逐票报告与复审记录已归档至[证据索引](./task-evidence/README.md)。验收以可见交互、媒体几何、公共 API 回读和请求次数为边界,没有新增内部状态或源码字符串验收。
本轮补入长期错误状态命中测试时发现真实问题:无效切分提示在 1920×980 遮挡“上一帧”。固定 `bottom:128px` 与随视口变化的播放区不匹配。提示现置于专注切片编辑区内,原布局保持原来的提示位置;五档尺寸下提示可见、操作命中和无文档滚动均通过。重复描述操作 CSS 已按最终计算值合并,Layout.tsx 已格式化。

View File

@ -0,0 +1,41 @@
# 实施与审核记录
本轮实施基线为 `ee2e355b`,分支为 `feature/local-dev-toolkit`。按票号顺序派遣独立实现者,每票完成后由另一位 reviewer 审核,发现问题先修复并复审,再开始下一票。
| 任务 | 实现提交 | 审核修复 | 状态 |
| --- | --- | --- | --- |
| 01 专注工作页与原布局切换 | a5753731 | 6acc5803 | 独立复审通过 |
| 02 单屏切片 | a3c780e8 | — | 独立审核通过 |
| 03 固定描述编辑 | e462437a | — | 独立审核通过 |
| 04 分页与定位 | 94b238d5 | e95adca8 | 独立复审通过 |
| 05 内嵌字典与长文本 | 1de1d85b | 6c76e5d2 | 独立复审通过 |
| 06 单路放大与比例适配 | ee3d6b16 | af7df833 | 独立复审通过 |
| 07 小窗口暂停与请求协调 | cb74444e | 972721e | 独立复审通过 |
| 08 整体验收与说明书 | 89c9e168 | ab89c63a | 独立复审通过 |
## 执行裁决
以下是执行过程中采用的可逆裁决,没有额外要求用户批准:
| 裁决 | 原因 | 若判断错误的代价 |
| --- | --- | --- |
| 执行新发布的 8 张工作页布局票,不重复此前已完成的 11 张快捷键票 | 当前任务目录与已确认的拆分是 8 张;本次请求沿用了上次的数量表述 | 如另有未包含任务,需要另行确定范围并补做 |
| 在现有个人功能分支完成并本地提交 | 明确调用的 implement 流程要求当前分支提交;本检出关联现有本地验证环境 | 如需隔离,可将这些本地提交迁移到另一检出 |
| 最终全量代码审查以 ee2e355b 为固定点 | 这是 8 张票发布后的实施起点 | 若要审查更早历史,需要另选更早基线扩大审查范围 |
| 每页最多 4 项的时间轴仍保持时长比例,并提供精确序号查找 | 既满足有限空间下访问全部条目,也保留时间长度与拖动映射的意义 | 若改成等宽卡片,需要增加另一种呈现;极短条目仍必须可搜索并在固定编辑区修改 |
## 已发现并修复的主要问题
- 跨切片重做后目标页被重置;页边缘拖动被当前页限制且移动比例漂移。
- 字典焦点移到外部后,Esc 无法正确退出。
- 单路放大的进度条遮住有效画面;原布局隐藏相机后进入专注视图不能恢复该路。
- 窗口恢复可能清除未知领取结果的暂停;暂停层遮挡离开确认或被原生提交弹窗覆盖;主动继续后键盘焦点丢失。
- 迟到的本人任务列表或详情在暂停期间切到下一任务。
- 大窗口中切片错误提示遮住逐帧按钮。
- 最终验收脚本在清理临时分支时破坏循环作用域,语法检查无法发现;独立审核发现后要求以最终脚本完整重跑。
每项修复均有针对性的真实浏览器回归或持久化 API 回读,详细证据将在最终验收与审查报告中链接。完整测试中的既有失败单列于 [全仓测试结果](./full-suite-results.md)。
## 最终关口
全部 8 票已独立审核通过;全功能 Standards/Spec 两轴审核尚待完成。逐票记录见[归档索引](./task-evidence/README.md)。

View File

@ -1,6 +1,6 @@
# 全仓测试结果
本报告记录专注工作页实现的完整测试,基线为 `ee2e355b`。前端最终检查尚未完成,本文暂为执行中的记录;不会以既有失败掩盖新增回归。
本报告记录专注工作页实现的完整测试,基线为 `ee2e355b`。全流程验收仍在执行;本报告保留完整测试的原始结果,并将后续局部复验分开列出,不以既有失败掩盖新增回归。
| 检查 | 当前结果 | 对照说明 |
| --- | --- | --- |
@ -13,10 +13,18 @@
| Go 格式 | 1 个既有文件不合规 | internal/api/annotation_operations_test.go 未被本轮修改 |
| Kubernetes 原客户端 dry-run | 环境限制,未通过 | 没有真实 kubeconfig;命令仍尝试 localhost:8080 资源发现,被本地沙箱阻止 |
| Kubernetes 本地结构检查 | 22 份 YAML 文档通过 | 仅验证解析及 apiVersion、kind、metadata;不等同于公司集群验收 |
| 前端全量 | 578 项:568 通过、10 失败 | 9 条与基线一致;1 条过时字典源码断言待浏览器核验后退役,后续局部复验单列 |
| 全仓 Prettier | 53 文件告警 | 52 文件未被本轮修改;本轮涉及的 Layout.tsx 待格式化复验 |
| 前端全量 | 578 项:568 通过、10 失败 | 9 条与基线一致;1 条字典弹层源码断言与新设计冲突;经真实新旧字典浏览器验证后退役,局部复验见下文 |
| 全仓 Prettier | 53 文件告警 | 52 文件未被本轮修改;本轮涉及的 Layout.tsx 已格式化;其余 52 文件保持原状 |
| 类型检查、Web 构建 | 第 7 项最终修复后通过 | 后续若有修改,再按影响范围复验;既有弃用及大包警告仍存在 |
Go 使用 `go test ./cmd/... ./internal/...`,覆盖全部实际源码包,避免 `./...` 扫描本仓忽略目录中的 Rust 临时构建目录。Python 使用 `.venv/bin/python3 -m pytest backend/operator-runner tests/backend -p no:cacheprovider`。检查均采用清洁环境,不继承公司集群或对象存储配置。
原始日志、退出码及失败集合对照保留在忽略目录 `local-data/verification/annotation-workspace-layout/final-checks/`。既有失败的此前基线复现见 [原快捷键版本的测试对照](../annotation-keyboard-workflow/final-suite-triage.md)。本次后台源码与 `ee2e355b` 无差异;失败集合比较包括 Python 的 `SUBFAILED`,并非仅比较汇总数量。
## 完整运行后的局部修正与复验
- 字典旧测试要求切片字典必须使用 portal,无法表达已确认的专注视图内嵌字典。真实浏览器已验证原布局弹层的指针、搜索、键盘选择、一次 Esc 与 Space,以及两阶段专注字典在五档尺寸的行为,随后只退役这条过时源码断言。该文件剩余 3 项通过;相关组合测试 54/54 通过。没有重跑全量,也不把局部结果改写成新的全量计数。
- 大窗口错误提示遮挡逐帧按钮的问题已通过真实浏览器复现并局部修复,五档尺寸回归通过。类型检查与 Web 构建在修正后通过。构建仍有原有的 module.register 弃用与大于 500kB 包警告。
- 本轮触及的 Layout.tsx 已格式化;全仓格式检查没有重复运行,52 个未改动文件的既有格式债务仍单独保留。
- 第 8 项交付脚本的临时分支清理错误由独立 reviewer 发现,运行时失败在写入前复现。修正后的同一份最终脚本在全新隔离项目完整退出 0,完成 3 次生成、保存与送检、2 次成功领取和 1 次空池请求。期间冻结所有仓库写入:实验证明本地 Vite 也会因 Markdown 修改整页重载,此前显式恢复不能作为产品故障归因。

View File

@ -0,0 +1,49 @@
# 逐票实施与审核证据
这些记录保留原始提交、检查、发现的问题及修复结果。local-data 路径指本机忽略的原始证据,不包含认证文件内容。
## 任务 01
- [task-1-fix1-review.md](task-1-fix1-review.md)
- [task-1-report.md](task-1-report.md)
- [task-1-review.md](task-1-review.md)
## 任务 02
- [task-2-report.md](task-2-report.md)
- [task-2-review.md](task-2-review.md)
## 任务 03
- [task-3-report.md](task-3-report.md)
- [task-3-review.md](task-3-review.md)
## 任务 04
- [task-4-fix1-review.md](task-4-fix1-review.md)
- [task-4-report.md](task-4-report.md)
- [task-4-review.md](task-4-review.md)
## 任务 05
- [task-5-fix1-review.md](task-5-fix1-review.md)
- [task-5-report.md](task-5-report.md)
- [task-5-review.md](task-5-review.md)
## 任务 06
- [task-6-fix1-review.md](task-6-fix1-review.md)
- [task-6-report.md](task-6-report.md)
- [task-6-review.md](task-6-review.md)
## 任务 07
- [task-7-fix1-review.md](task-7-fix1-review.md)
- [task-7-report.md](task-7-report.md)
- [task-7-review.md](task-7-review.md)
## 任务 08
- [task-8-fix1-review.md](task-8-fix1-review.md)
- [task-8-report.md](task-8-report.md)
- [task-8-review.md](task-8-review.md)

View File

@ -0,0 +1,31 @@
### Spec Compliance
- ✅ Spec compliant for the scoped review fix. The previously missing behavioral evidence now exercises active continuous work and stop-after-current in the recovered session across both focused → original and original → focused switches (`tests/browser/annotation-workspace-shell.mjs:70`, `tests/browser/annotation-workspace-shell.mjs:75`, `tests/browser/annotation-workspace-shell.mjs:86`).
- ✅ Each direction asserts the same task URL and zero task writes/claims, while the stopped state additionally proves that layout switching does not silently re-enable continuous work (`tests/browser/annotation-workspace-shell.mjs:78`, `tests/browser/annotation-workspace-shell.mjs:89`, `tests/browser/annotation-workspace-shell.mjs:94`). Consuming the task to test post-completion continuation is outside this scoped fix and is not required for closing the original finding.
- ✅ The controller independently verified the recovery tag target and complete-history bundle, closing the prior diff-external warning.
### Strengths
- The fix changes only the browser acceptance test and places the checks immediately after explicit session recovery, avoiding the navigation flaw in the earlier evidence.
- It verifies both possible intent states, both layout directions, visible user-facing status/control changes, task identity, and mutation absence rather than inferring preservation from component structure.
- The recorded focused browser run contains the new `recovered-session continuous intent and stop-after-current survive both layout directions without claims` PASS group and all seven existing PASS groups.
### Issues
#### Critical (Must Fix)
None.
#### Important (Should Fix)
None.
#### Minor (Nice to Have)
None.
### Assessment
**Task quality:** Approved
**Reasoning:** The scoped fix directly and behaviorally proves the continuous-work intent property that was missing from the original review. The prior Important finding is resolved without expanding production scope or consuming the isolated task.

View File

@ -0,0 +1,110 @@
# Task 01 — focused shell and lossless original-layout switch
Status: implementation complete; awaiting controller review. Ticket status remains controller-owned. No push/deployment or worktree created.
## Delivered behavior and boundaries
- Dedicated `/plugins/annotation-qa/workflows/:workflowId/work` defaults to focused presentation for an account without a browser preference. Platform sidebar/status header disappear and outer padding reduces. Existing task-pool `/annotate` presentation stays original.
- The existing work-page heading gains `切回原布局` / `进入专注视图`; it and the existing return link remain sticky while the current editor scrolls. Original mode restores the actual existing application chrome and work-page structure, not a second editor.
- `Layout` owns account-scoped browser preference; the editing coordinator consumes only the preference/toggle. The same coordinator, workbench and media DOM remain mounted. No task/session key, claim/save/release callback, asynchronous mutation fence or editor lifecycle changed.
- Layout storage keys include actual current user ID. Reads of missing/invalid values default focused; storage access exceptions keep live switching usable. Preference reads never activate a work session. Reload still requires explicit continuation and shows no editor beforehand.
- Current task/stage, selection, frame, independent per-slice drafts, unsaved text, persisted and in-memory undo/redo, and pending-save restrictions remain intact across toggles.
- Ticket 01 is only the focused shell/fallback. No claim of two-stage left/right layout, minimum-picture dimensions, no-scroll contract, pagination or undersized-window handling: these belong to subsequent tickets.
## Files
- `src/context/AnnotationWorkspaceLayoutContext.tsx`: minimal presentation-only preference/context, scoped by current account.
- `src/components/Layout.tsx`: dedicated-route chrome visibility and padding; stable provider around route children.
- `src/pages/AnnotationOperations.tsx`: consume presentation preference; render visible mode switch alongside original return link. No changes to editing/ownership logic.
- `src/styles/capture-workspace.css`: focused header hiding and sticky dedicated-page heading only.
- `tests/browser/annotation-workspace-shell.mjs`: real local browser/public API integration gate, configurable authenticated session files. No private React state or source-string acceptance.
- `docs/annotation-workspace-layout.md`: user behavior, delivery boundary, runtime fallback versus source/business recovery.
## TDD and browser evidence
Initial browser test ran against unmodified production source. Original `/annotate` entry successfully mounted the workbench and displayed platform chrome. Then `/work` explicit recovery screen still displayed platform chrome, failing the assertion `default focused layout hides platform chrome` (actual true, expected false). This is the feature RED; earlier setup failures (system proxy and a nested header's semantic role) were test/environment fixes, not feature RED.
RED command, before production changes:
```sh
ANNOTATION_STORAGE_STATE=local-data/verification/annotation-workspace-layout/auth.json node tests/browser/annotation-workspace-shell.mjs
```
Final GREEN command:
```sh
ANNOTATION_STORAGE_STATE=local-data/verification/annotation-workspace-layout/auth.json \
ANNOTATION_SECOND_STORAGE_STATE=local-data/verification/annotation-workspace-layout/secondary-auth.json \
ANNOTATION_EVIDENCE_DIR=local-data/verification/annotation-workspace-layout \
node tests/browser/annotation-workspace-shell.mjs
```
The runner uses installed Microsoft Edge headless with `--no-proxy-server`, needed for this machine's localhost browser connectivity. Browser operations require the authorized unsandboxed local networking surface. All service URLs are hardcoded localhost:5173, backed by 3001/27018; there is no remote target option.
Checks exercised:
1. Original task-pool entry keeps platform chrome, the actual workbench and no dedicated-page toggle.
2. Existing Start Work opens the same browser tab, resumes an already-owned task without claim, and exposes the existing stop-after-current control.
3. Default focused shell hides chrome; original restores it. Original choice survives refresh; editor remains absent until explicit Continue, with no write/claim from preference or refresh.
4. Both real local video elements reach playable readiness. One-frame keyboard step is retained exactly across six toggles; original video and workbench DOM nodes remain connected.
5. Existing first-slice description is edited without saving; second-slice draft created/parked through normal controls. Ctrl-arrow navigation and toggles preserve both texts. Dirty return opens the real leave modal and Continue Editing preserves the draft. Switch remains visibly in viewport while editing lower content.
6. Ctrl+Enter performs a real server save; the response is held only after `route.fetch()` reaches the real API successfully. Switch to original while response is held: save button stays disabled, text remains, and exactly one POST exists. Releasing response completes the same session normally.
7. Public `GET /api/human-items/:id` confirms version increment exactly one and the saved description. Ctrl+Z after another toggle restores prior persisted description via real API; Ctrl+Shift+Z in original mode reapplies it. Total task writes exactly three: save, undo, redo. No claim/release or duplicate save.
8. Original slicing UI: six frame steps and C split; toggle; Ctrl+Z removes split; toggle back; Ctrl+Shift+Z restores split. No task writes from local slice history/toggles.
9. Two actual local annotator sessions are exchanged by authenticated cookies in the SAME browser context, retaining origin localStorage. First account's original preference does not affect new peer's focused default; peer's original choice remains original after first account changes to focused. No task mutation during account tests.
Raw ignored evidence: `local-data/verification/annotation-workspace-layout/task01-browser-green.log`, `task01-focused.png`, `task01-original.png`. Both screenshots visually inspected: real paired video, original sidebar restored, sticky switch/return accessible. Screens show current original editor; they do not establish later-layout acceptance.
## Local fixture/auth scope
Controller prepared isolated project `project-workspace-layout-acceptance`. Main browser tasks: `annotation-task-workspace-layout-acceptance-annotate-primary` and `annotation-task-workspace-layout-acceptance-slice-primary`. Public save/undo/redo deliberately leave the final test description saved in the isolated annotation task; slice changes and second-slice drafts are unsaved and discarded with browser close.
Existing platform admin is super-admin and correctly receives organization console instead of work-page editing. Controller explicitly authorized a new localhost-only peer for account isolation: `user-workspace-layout-acceptance-peer`, fixture member `project-member-workspace-layout-acceptance-peer`; copied only primary local user/password hash and this new project's annotator membership to unique IDs, without modifying existing user/member records. Guarded helper verifies Mongo is published exactly at 127.0.0.1:27018. Peer never claims/edits tasks. Helper and auth state are ignored local data, session files chmod 0600. No secrets in tracked files/report.
## Focused verification
```sh
npm run lint:web
node --test tests/frontend/annotation-shortcuts.test.mjs tests/frontend/annotation-history.test.mjs tests/frontend/annotation-authoritative-mutation.test.mjs tests/frontend/annotation-playhead-draft.test.mjs tests/frontend/annotation-submission-state.test.mjs
npm run build:web
npx prettier --check src/context/AnnotationWorkspaceLayoutContext.tsx src/pages/AnnotationOperations.tsx tests/browser/annotation-workspace-shell.mjs
git diff --check
```
Typecheck, 44/44 focused frontend tests, Vite web build, scoped formatting and diff check pass. Existing >500 kB build chunk notice remains. No backend behavior changed; broad/full suites are controller-owned. Layout's pre-existing formatting was preserved to avoid whole-file churn; only new blocks were formatted reasonably.
## Source recovery verification
`git rev-parse 'codex/annotation-keyboard-baseline-20260918^{}'` returns `54a35afad95d77b1f23c955f9c4671cae83a664d`. `git bundle verify local-data/backups/annotation-keyboard-baseline-20260918.bundle` reports okay, complete history, SHA-1 bundle. The annotated tag object is `a5ac806e5677975d125f2069cb289e2218af6b93`. These recover source only, never database/media/unsaved browser state. No restore operation was performed.
## Self-review
Checked route isolation, first-render account separation, localStorage exception fallback, stable editor reconciliation, no layout-mode dependency in request/session state, no new backend API/schema, unchanged stage history boundaries, unchanged return blocker and account/session guards. Delayed-save experiment proves the original result reconciles after switching. Original schema, media paths, task versions and ownership contracts unchanged. No broader refactor or later ticket implemented.
## Commit
Commit `a5753731` — `✨ Add focused annotation shell with lossless layout fallback`. Final browser gate exits 0 with all seven PASS groups. Only the six task implementation/documentation/test paths listed above are committed; controller's ticket claim edit remains unstaged. No push or deployment.
## Review fix 1 — continuous-work intent evidence
The Important review finding was confirmed: the prior Start Work stop click was followed by a fresh navigation before the layout-toggle assertions, so it did not establish intent retention in the active recovered session. Added a focused browser check immediately after explicit Continue, before draft editing:
- Click Start Continuous Work on the current owned task, switch focused → original → focused, and after EACH switch assert the visible continuous-work status, the Stop After Current button, identical task URL, and zero task writes/claims.
- Click Stop After Current, switch focused → original → focused, and after EACH switch assert the explicit stop-selected/no-next-claim status, visible Resume Continuous Work button, absent Stop After Current button, identical task URL, and zero task writes/claims.
- Continue through the existing real-media/draft/delayed-save/API/undo/redo/account regression gate.
Only `tests/browser/annotation-workspace-shell.mjs` changed. No production code change, no additional task completion/consumption, and no new fixture/account. This test establishes live intent retention and no claims caused by switching. It deliberately does not claim verification of post-completion continuation; full continuous-work completion remains a later acceptance task.
Command:
```sh
ANNOTATION_STORAGE_STATE=local-data/verification/annotation-workspace-layout/auth.json \
ANNOTATION_SECOND_STORAGE_STATE=local-data/verification/annotation-workspace-layout/secondary-auth.json \
ANNOTATION_EVIDENCE_DIR=local-data/verification/annotation-workspace-layout \
node tests/browser/annotation-workspace-shell.mjs \
> local-data/verification/annotation-workspace-layout/task01-browser-review-fix.log 2>&1
```
Syntax (`node --check`) and `git diff --check` pass; the test file was formatted with Prettier. No typecheck/build rerun needed for a browser-test-only evidence change. Review-fix browser result and commit recorded below after completion.
Review-fix GREEN: exit 0, all eight PASS groups including `recovered-session continuous intent and stop-after-current survive both layout directions without claims`. Commit `6acc5803` — `🧪 Verify continuous-work intent across layout switches`. Controller ticket edit remains unstaged; no push/deployment.

View File

@ -0,0 +1,31 @@
### Spec Compliance
- ❌ Issues found: the implementation and browser evidence cover the focused shell, real original-layout fallback, account/browser preference, explicit refresh recovery, editor/media identity, frame and per-slice draft retention, delayed save reconciliation, API readback, and undo/redo. However, the required continuous-work intent preservation is not behaviorally verified: the test clicks “完成当前任务后停止” at `tests/browser/annotation-workspace-shell.mjs:46`, immediately performs a fresh navigation at `tests/browser/annotation-workspace-shell.mjs:49`, and never asserts that choice after the repeated layout toggles at `tests/browser/annotation-workspace-shell.mjs:96`. This leaves one of the task’s explicit session/continuation guards supported only by structural inference.
- ⚠️ Cannot verify from diff: the local tag actually resolving to `54a35afad95d77b1f23c955f9c4671cae83a664d` and the bundle containing complete history. The documentation records both claims at `docs/annotation-workspace-layout.md:15`, and the report gives the verification commands/results, but those repository-local recovery artifacts are outside the review diff. The controller should retain or inspect the `git rev-parse` and `git bundle verify` evidence.
### Strengths
- Preference state is narrowly scoped to presentation and keyed by actual user ID, including same-render account changes and storage-failure fallback (`src/context/AnnotationWorkspaceLayoutContext.tsx:11`, `src/context/AnnotationWorkspaceLayoutContext.tsx:23`, `src/context/AnnotationWorkspaceLayoutContext.tsx:32`).
- The shell hides/restores the real sidebar and header while keeping one stable provider around the existing route children, so toggling does not create a second editor tree (`src/components/Layout.tsx:52`, `src/components/Layout.tsx:185`, `src/components/Layout.tsx:190`, `src/components/Layout.tsx:306`).
- The browser test checks observable behavior rather than source strings: same-account refresh recovery (`tests/browser/annotation-workspace-shell.mjs:56`), connected editor/media nodes and zero toggle writes (`tests/browser/annotation-workspace-shell.mjs:101`, `tests/browser/annotation-workspace-shell.mjs:111`), a held real API response with exactly one write and version readback (`tests/browser/annotation-workspace-shell.mjs:157`, `tests/browser/annotation-workspace-shell.mjs:171`, `tests/browser/annotation-workspace-shell.mjs:176`), persisted undo/redo (`tests/browser/annotation-workspace-shell.mjs:191`, `tests/browser/annotation-workspace-shell.mjs:205`), and two-account preference isolation (`tests/browser/annotation-workspace-shell.mjs:248`, `tests/browser/annotation-workspace-shell.mjs:260`). The named browser log contains all seven PASS groups, and the screenshots visibly show both the focused shell and restored original chrome.
- The documentation states the ticket boundary and correctly separates runtime layout fallback from source and business-data recovery (`docs/annotation-workspace-layout.md:7`, `docs/annotation-workspace-layout.md:13`, `docs/annotation-workspace-layout.md:17`).
### Issues
#### Critical (Must Fix)
None.
#### Important (Should Fix)
- `tests/browser/annotation-workspace-shell.mjs:46` — The browser gate does not prove that `stop-after-current`/continuous-work intent survives a layout switch. The click is followed by a page navigation before the toggle sequence, and no later assertion inspects the continuation setting or proves that completion will not trigger another claim. This matters because preserving continuous-work intent and the associated claim guard is an explicit Task 01 requirement, and the review constraints require behavioral evidence rather than reasoning from unchanged source. Exercise the choice in the recovered work session, toggle both ways, and assert its visible state remains selected; ideally complete the isolated current task and verify that no next claim is issued.
#### Minor (Nice to Have)
None.
### Assessment
**Task quality:** Needs fixes
**Reasoning:** The implementation is small, clean, and strongly evidenced for nearly every required lifecycle property. Approval is blocked only by the missing behavioral proof for continuous-work intent, a claim/session guard explicitly included in this task.

View File

@ -0,0 +1,91 @@
# Task 02 — single-screen focused slicing
Status: implementation complete; awaiting controller review. Ticket status remains controller-owned. Current branch retained; no worktree, push, deployment, company-cluster write or BOS write.
Commit: `a3c780e8ae0454f653692fc805b6545b213e87f2` — `✨ Fit focused slicing controls and dual video in one screen`.
## Delivered
- Focused slicing reserves the left region for both existing video nodes, transport and detailed timeline; the right region contains stage context, the existing naming editor, disposition entry and fixed submission/history controls.
- Both actual 590×480 sources remain complete with `object-fit: contain`; no automatic camera hiding. Larger browser content areas enlarge the image.
- The existing media nodes, timeline handlers, name callback, history coordinator and confirmation/generation APIs remain shared with original mode. No layout key/remount, second editor, ownership or version logic added.
- Visible undo/redo buttons dispatch the existing history commands. Space/C/arrows/B/V, dictionary selection and one-Escape focus return remain intact.
- Compact continuous-work status and fixed work-status details preserve access to long explanations; frame/edit errors do not introduce rows into the media layout. Rework/discard explanations expand from bounded entries. Native confirmation has a bounded summary and fixed action row.
- Processing uses the compact shell and remains processing until authoritative completion. Existing two-step confirmation is unchanged.
- Original layout keeps its existing media/editor structure and flow, full explanations and platform shell. New history controls are scoped to focused mode.
## TDD and acceptance
Approved seam: real browser interactions with existing local APIs; no source-string acceptance added. Browser uses installed Edge headless with `--no-proxy-server`, localhost:5173 through local API3001/snapshot27018.
RED before production edits: `tests/browser/annotation-workspace-slice.mjs` measured each effective picture at **295×240** at 1280×640 and reported outer page plus workbench scrolling. Assertion `each effective image meets 400×325 floor` failed. Raw RED log: ignored `local-data/verification/annotation-workspace-layout/task02-red.log`.
Final browser command (exit 0):
```sh
ANNOTATION_TASK_ID=annotation-task-workspace-layout-acceptance-pool-01 \
node tests/browser/annotation-workspace-slice.mjs \
> local-data/verification/annotation-workspace-layout/task02-final-browser.log 2>&1
```
Observed effective picture sizes **for each camera**, computed from intrinsic video dimensions and actual contain fitting:
| Browser content area | Effective picture |
| --- | --- |
| 1280×640 | 410.54×334 |
| 1366×680 | 459.71×374 |
| 1440×800 | 529×430.37 |
| 1920×980 | 769×625.63 |
| 2560×1320 | 1089×885.97 |
At all five sizes: no document/main/workbench scrolling; both cameras in viewport; playback, frame, B/V, split, dictionary, confirmation, undo/redo, help, continuous-work and layout-switch buttons both in viewport and hit-test reachable. Screenshot visually inspected at 1280×640 (`task02-focused.png`, ignored).
Real interactions: C inserts a cut; Space plays/pauses; frame stepping advances media; B reviews and V returns paused; camera times stay within 50ms; original/focused toggles preserve connected media nodes, playhead and cut count; visible undo/redo restore cut counts; dictionary search/select fills actual enabled `布局验收切片 001`; one Escape returns focus. An actual invalid C cut at the initial boundary produces the editor alert without page scrolling. Expanded work instructions leave geometry stable.
## Actual asynchronous generation and public readback
On owned `annotation-task-workspace-layout-acceptance-slice-primary`, the browser:
1. Chose enabled dictionary names for four slices.
2. Opened the first confirmation step with zero generation requests.
3. Verified the second-step action is inside 1280×640 and hit-test reachable.
4. Sent exactly one real slice-confirmations POST.
5. Observed the real processing panel, absence of description controls, and no page scrolling during processing.
6. Waited for the actual description-stage button after completion.
The original generation run then hit a **test-harness field-name mistake**, expecting `annotationWorkspace.phase/slices`; this is not the API schema. Fixed to `sliceWorkflow.phase` and `annotationWorkspace.sliceSegments`. No reset or resubmission occurred. A separate browser/public-API gate against the same completed result passed:
```sh
ANNOTATION_VERIFY_GENERATED=1 node tests/browser/annotation-workspace-slice.mjs
```
Output: version **5**, slice workflow and materialization **completed**, phase **annotate**, **4** named slices, full **412s** contiguous coverage, zero gap/overlap microseconds. Controller independently observed completed/pending_annotation. The script now uses these correct fields for future end-to-end generation runs.
Raw generation flow: ignored `task02-generation.log` (includes the explained harness assertion). Final API/UI readback: ignored `task02-generation-readback.log` (exit 0).
Fixture state: slice-primary remains legitimately generated/annotate with four slices. For repeatable final geometry, claimed one ready fixture through the normal CSRF-protected random-claim API with request ID `workspace-layout-task02-geometry-20260918`; returned **pool-01**, now owned in slice phase. Final browser changes on pool-01 are unsaved and were discarded by closing the browser; no generation, disposition save or task reset. Other pool fixtures remain available.
## Verification
- `npm run lint:web`: exit 0.
- `npm run build:web`: exit 0; existing >500kB chunk warning only.
- `node --test` focused files: annotation-shortcuts, annotation-history, annotation-authoritative-mutation, annotation-confirmation-boundaries, capture-annotation-workbench, capture-responsive-layout, slice-dictionary: **52 passed, 0 failed**.
- Browser syntax check, Prettier for changed source/test paths, `git diff --check`: pass.
- Full frontend suite deliberately left to controller's once-at-end run; baseline known 9 failures not relabeled as regressions.
## Files and next-stage integration contract
- `src/pages/AnnotationOperations.tsx`: `focusedStageLayout` near3452 owns presentation-only opt-in for slice/materializing; exposes `data-workspace-compact` plus `data-workspace-stage`, compacts status details, passes presentation `focusedLayout` prop.
- `src/styles/capture-workspace.css`: new scoped rules begin near344. Shared geometry variables: `--workspace-editor-width:350px`, `--workspace-region-gap:8px`. Root workbench columns left flexible/right fixed; rows38px/flexible/126px. Existing `[data-testid="capture-video-panel"]`, `.capture-media-controls`, `.capture-timeline-editor`, `.capture-stage-editor` and footer are the presentation regions. Original selectors remain unchanged.
- `src/components/CaptureAnnotationWorkbench.tsx`: `.capture-stage-editor` labels the existing slice-name/details region without moving its DOM; `focusedLayout` controls presentation of rework/discard details only; visible history buttons use existing callbacks.
- `tests/browser/annotation-workspace-slice.mjs`: real local geometry/actions/generation/readback runner. Explicit fixture ID whitelist, no remote URL support. Generation opt-in because it advances real fixture state.
- `docs/annotation-workspace-layout.md`: delivered slicing behavior, current limits and fixture-safe runner modes.
Task03 can extend the presentation opt-in and map its existing annotation editor into the right region; retain current media DOM and business handlers. Current naming region is positioned relative to the shared workbench while keeping its old timeline-parent DOM for original-layout fidelity. Do not mount separate layouts.
## Limits / remaining parent acceptance
- This ticket verifies the five CSS browser content sizes in real headless Edge. The physical desktop/window/100%–125%–150% zoom matrix is not claimed; remaining L03 and undersized behavior belong to later/end-to-end acceptance.
- Synthetic alternate aspect ratios, single-camera enlargement, annotation fixed editing, paging and inline dictionary are intentionally not implemented by task02.
- Bounded rework/discard detail presentation is implemented, but this ticket did not mutate fixtures to create a long rework/discard history. Normal instructions and an actual invalid-boundary alert were exercised.
- Full continuous-work slice→describe→submit→next claim remains parent/end-to-end scope; this ticket completed real slicing/generation only.

View File

@ -0,0 +1,38 @@
### Spec Compliance
- ✅ Spec compliant for Task 02. The compact layout is restricted to focused slice/materializing stages (`src/pages/AnnotationOperations.tsx:3452`), passes presentation state without changing the editor key (`src/pages/AnnotationOperations.tsx:4463`), and retains the existing editor/media/callback tree. Grid geometry reserves the video and fixed controls separately (`src/styles/capture-workspace.css:421`).
- ✅ Five-size saved browser evidence agrees with the actual geometry assertions (`tests/browser/annotation-workspace-slice.mjs:66`): each 590×480 source displays at 410.54×334 at 1280×640, grows through all five sizes, and uses contain fitting without document scrolling. The saved minimum-size screenshot shows both full pictures and the required short-list controls.
- ✅ Real generation evidence supports the two-step, processing, and authoritative-completion requirements (`tests/browser/annotation-workspace-slice.mjs:215`). The original run reached version 5 before its incorrect API-field assertion failed; the separate corrected readback of that same result reports completed/annotate, four slices and 412 seconds. The corrected schema/coverage assertions are explicit (`tests/browser/annotation-workspace-slice.mjs:41`). This is transparent split-run evidence, not a claim that the original generation command passed.
- ⚠️ Broad type/build/52-test results were reported but not rerun in this task review. Physical zoom, annotation-stage layout, pagination, inline dictionary, camera enlargement and undersized suspension remain outside this slice's acceptance scope.
### Strengths
- Shared history callbacks power the new visible controls; no second state coordinator was introduced (`src/components/CaptureAnnotationWorkbench.tsx:3449`).
- The original mode retains expanded rework information while focused mode provides a bounded entry using the same content (`src/components/CaptureAnnotationWorkbench.tsx:1764`).
- Browser acceptance checks actual intrinsic-image scaling, hit testing, keyboard editing, media-node continuity, dictionary selection, and public API output rather than source text (`tests/browser/annotation-workspace-slice.mjs:66`, `tests/browser/annotation-workspace-slice.mjs:215`).
### Issues
#### Critical (Must Fix)
- None.
#### Important (Should Fix)
- None.
#### Minor (Nice to Have)
- `tests/browser/annotation-workspace-slice.mjs:148`: the invalid-boundary error branch asserts visibility and lack of scrolling but does not repeat control hit testing. The new error notice floats over the lower transport area (`src/styles/capture-workspace.css:589`), so a hit-test assertion here would guard future text/layout changes. My focused browser check found the actual boundary error at x20–322, y468–492 and confirmed every transport button/select remained reachable; this is a coverage improvement, not an observed product defect.
### Checks
- Read the supplied diff, brief, report, final browser log, generation log and corrected readback log; visually inspected the saved minimum-size screenshot. The first combined output truncated the diff, so only its missing sections were recovered; no changed source file was separately reread.
- Named risk: floating error notice could intercept transport controls. Ran one focused local browser check on pool-01 with an actual invalid C split; all transport hit tests passed. No save, generation or reset occurred. Initial sandboxed browser launch failed; the same check succeeded with approved process access.
- No broad suite rerun, subagent, Git mutation, or source change. The pre-existing build chunk warning is not a new finding.
### Assessment
**Task quality:** Approved.
**Reasoning:** The implementation achieves the scoped one-screen slicing workflow while retaining shared media and editing lifecycle. Saved acceptance evidence and the focused error-state check support the geometry and interaction claims; the remaining recommendation is regression coverage polish.

View File

@ -0,0 +1,93 @@
# Task 03 — fixed annotation editing and submission
Status: implementation complete; controller review pending. Ticket ownership remains with controller. Current branch retained, no worktree/push/deployment/company-cluster/BOS write.
Commit: `e462437a157f63dd858f92e4487a8e4377f8fc3f` — `✨ Keep annotation editing and submission in the focused workspace`.
## Delivered
- Extend shared `focusedStageLayout` presentation to annotate, retaining the exact media nodes, editor instance, draft/navigation handlers, mutation scope and task ownership logic.
- Same left dual-media geometry as slicing. Slice selection and description timeline share the left bottom area; existing description text/boundaries/actions occupy the fixed 350px right region.
- Save, boundary adjustment, cancel and delete remain visible, with shared undo/redo and stage submission in the right footer. Existing-description cancel is now visible in focused mode and reuses the existing new-draft cancellation path: discard only active buffer, pause and focus workbench, no write. Original layout cancellation UI remains unchanged.
- Optional semantic-blank and boundary guidance is available from small bounded details. Submission/draft confirmation bodies can scroll internally while their action row remains visible. Stale-version errors preserve current text and usable retry controls.
- Original layout keeps its prior video cap, editor flow and history keyboard behavior. Media classes change presentation only; no second editor, task reacquisition or media source change.
- Correct focused annotation title contrast to app text color.
- Parameterize shell regression slice task with `ANNOTATION_SLICE_TASK_ID`, default pool-01; no longer incorrectly reuse consumed slice-primary.
## TDD and observable acceptance
All browser runs use real local Edge (`--no-proxy-server`), localhost5173/API3001/snapshotMongo27018 and isolated workspace-layout-acceptance fixtures. No source-string acceptance added.
RED: `node tests/browser/annotation-workspace-annotation.mjs` failed on 1280×640 effective picture floor: **188.79×153.59** per camera, with main scrolling. Log `local-data/verification/annotation-workspace-layout/task03-red.log` (ignored).
Second geometric RED: save-actions' actual bottom extended below reserved editor into disposition region. `task03-editor-red.log`; fixed with bounded on-demand guidance, retaining information access.
Cancel RED: no visible existing-edit cancel button. `task03-cancel-red.log`; final test confirms no API version/write change and reopening restores persisted text.
Final runner (exit 0):
```sh
node tests/browser/annotation-workspace-annotation.mjs
```
Raw final log: ignored `task03-final.log`; screenshot visually inspected: ignored `task03-focused.png`.
| Browser content | Each effective picture |
| --- | --- |
| 1280×640 | 410.54×334 |
| 1366×680 | 459.71×374 |
| 1440×800 | 529×430.37 |
| 1920×980 | 769×625.63 |
| 2560×1320 | 1089×885.97 |
All five: contain-fit complete dual pictures, grow with viewport, no document/main/workbench scrolling; description, save/cancel/delete/end adjustment, stage submission, undo/redo, frame controls/help/layout switch visible and hit-test reachable. Actual action bounds remain above disposition.
Public API and interactions:
- Existing draft survives original/focused switch; textarea and both media DOM nodes remain connected; toggles issue no writes.
- Intercepted only request version to submit an actual stale version to real API: **409** response, unchanged version, retained text, visible retry and no page scroll. No synthetic success response.
- Manual Ctrl+Enter save increments version once; public human-item readback confirms exact text. Visible undo then redo each POST successfully and authoritative text reverts/restores.
- Existing-edit cancel discards current unsaved change with unchanged persisted version and no request; reopening shows saved text.
- N starts, frames advance, N ends and focuses existing editor; Enter enters textarea as the established focus contract requires. Boundaries remain increasing. Cross-slice navigation and layout switches retain parked draft; submission presents draft guard without writing; cancel removes draft.
- Ctrl+Enter without draft opens second-step stage confirmation. Return/confirm are hit-test reachable at minimum viewport; first step issues no write.
## Actual submission and readback
Used dedicated **annotate-delayed**, preserving annotate-primary for Tasks04/05:
```sh
ANNOTATION_TASK_ID=annotation-task-workspace-layout-acceptance-annotate-delayed \
ANNOTATION_SUBMIT=1 node tests/browser/annotation-workspace-annotation.mjs
```
Actual second confirmation sent the real combined-submissions POST. Its response was asserted `processing`, still `annotation_combined`, with a real pendingCombinedSubmissionId. UI showed actual preparation status; continuous work was not enabled. The harness then incorrectly expected the submit button to remain disabled, while the existing processing contract changes to read-only and removes it. This harness assertion failed; no fixture reset or resubmission was performed. Corrected future run to absent-or-disabled, and release held polling in finally.
Controller independently observed completion. Fresh read-only verification (exit0):
```sh
ANNOTATION_TASK_ID=annotation-task-workspace-layout-acceptance-annotate-delayed \
ANNOTATION_VERIFY_SUBMITTED=1 node tests/browser/annotation-workspace-annotation.mjs
```
Final public item: **version6, currentStageId quality_review, currentRevision1, status ready, no pending submission**. This verifies send-to-QA readiness, not QA judgment. Raw evidence `task03-submit.log` (explained harness failure), `task03-submitted-readback.log` (final pass). Pre/accepted versions were read by the original run but not printed before its assertion; no fabricated version numbers are reported. Updated future runner logs them immediately.
## Regressions and checks
- `ANNOTATION_STORAGE_STATE=local-data/verification/annotation-workspace-layout/auth.json ANNOTATION_SECOND_STORAGE_STATE=local-data/verification/annotation-workspace-layout/secondary-auth.json ANNOTATION_SLICE_TASK_ID=annotation-task-workspace-layout-acceptance-pool-01 node tests/browser/annotation-workspace-shell.mjs`: exit0; original entry, recovered continuous intent/stop, same media/frame, parked drafts, delayed actual save with pending lock, saved undo/redo across original/focused, original C/history, actual account preference isolation. `task03-shell.log` ignored.
- `ANNOTATION_TASK_ID=annotation-task-workspace-layout-acceptance-pool-01 node tests/browser/annotation-workspace-slice.mjs`: exit0; all five picture sizes, reachable controls, C/Space/frames/B/V, dictionary and history. `task03-slice.log` ignored. No generation/save on pool-01.
- `npm run lint:web`: exit0 after final implementation.
- `npm run build:web`: exit0 after final implementation; existing >500kB chunk warning only.
- `node --test` annotation-shortcuts, annotation-history, annotation-authoritative-mutation, annotation-confirmation-boundaries, annotation-playhead-draft, annotation-submission-state, capture-annotation-workbench: **56 passed, 0 failed** after final implementation.
- Browser syntax, changed source/test Prettier and `git diff --check`: pass.
- Full frontend suite intentionally left to controller's once-at-end run (baseline 9 known failures).
## Files and following-task integration
- `src/pages/AnnotationOperations.tsx`: compact presentation now covers annotate.
- `src/components/CaptureAnnotationWorkbench.tsx`: same source/editor nodes; semantic class hooks for annotation heading/tools/boundaries/actions, bounded guidance, focused existing-edit cancel, shared history controls.
- `src/styles/capture-workspace.css`: annotate uses shared 350px right track, same media geometry; compact two timeline rows, fixed editor, guidance and confirmation bounds. Annotation timeline is viewport anchored in focused mode (left20px/bottom12px with width derived from shared right track); single-camera and undersize tasks must adapt this alongside detailed timeline handling.
- `tests/browser/annotation-workspace-annotation.mjs`: real annotation geometry/mutation/draft/cancel and guarded submission/readback modes.
- `tests/browser/annotation-workspace-shell.mjs`: explicit slice fixture parameter.
- `docs/annotation-workspace-layout.md`: delivered annotation behavior and fixture-safe verification modes.
Task04 owns long/many slice and description navigation. Task05 owns inline dictionaries and expanded long text; existing popup dictionary mechanics are unchanged. Task06 single-camera and Task07 undersize pause remain untouched. Normal short text and short lists are this ticket's acceptance; unusual long rework histories, physical desktop/OS zoom matrix and complete continuous slice→describe→submit→next-claim cycle remain later/controller coverage. Fixture statuses: annotate-primary owned annotate with genuine edited saved text; annotate-delayed legitimately quality_review; slice-primary generated annotate; pool-01 slice, unsaved browser cuts discarded on close.

View File

@ -0,0 +1,39 @@
### Spec Compliance
- ✅ Spec compliant for Task03. `src/pages/AnnotationOperations.tsx:3452` extends the existing compact presentation to annotation; `src/components/CaptureAnnotationWorkbench.tsx:1801` and `:1846` only change presentation conditions on the existing media. No second editor, mutation implementation or task acquisition path was introduced.
- ✅ Fixed editor, bounded contextual guidance and confirmation bodies are implemented at `src/styles/capture-workspace.css:795`, `:938` and `:907`. Focused existing-edit cancellation at `src/components/CaptureAnnotationWorkbench.tsx:3279` clears the active buffer/end selection, pauses playback and restores workbench focus without writing. Title contrast is corrected at `src/styles/capture-workspace.css:967`.
- ✅ Real browser assertions cover five viewport sizes, control reachability, stale-version rejection, API save/undo/redo readback, retained media/editor nodes, parked drafts and two-step confirmation (`tests/browser/annotation-workspace-annotation.mjs:100`, `:147`, `:160`, `:200`). The final evidence log records all these passes and a 410.54×334 effective image per camera at 1280×640.
- ✅ Actual handoff was corroborated by the controller and the recorded public readback: version 6, quality_review, revision 1, ready, no pending submission. The corrected processing assertion accepts the existing absent-or-disabled submit control (`tests/browser/annotation-workspace-annotation.mjs:272`). The original harness assertion failure is documented, not presented as a successful end-to-end run.
- ⚠️ The unchanged continuous-work next-claim and IME contracts cannot be fully verified from this presentation diff; their existing implementations remain outside this scoped gate. No new bypass or mutation path appears here. Tasks04–07 remain explicitly deferred.
### Strengths
- `src/components/CaptureAnnotationWorkbench.tsx:3279`: the added cancellation action reuses established state handling and has a real no-write/reopen-persisted-text regression at `tests/browser/annotation-workspace-annotation.mjs:189`.
- `tests/browser/annotation-workspace-annotation.mjs:147`: stale-version validation changes only the outbound version and checks a real server 409, while save and history verification read authoritative persisted content.
- `tests/browser/annotation-workspace-annotation.mjs:241`: destructive fixture progression is constrained to the dedicated submission fixture; the separate readback mode avoids resetting or resubmitting completed work.
- `src/styles/capture-workspace.css:609` and `:616`: the existing visibility rules keep the relocated history buttons hidden in original layout. `src/components/CaptureAnnotationWorkbench.tsx:2851` supplies the relative containing block needed for the newly positioned boundary-help control.
### Issues
#### Critical (Must Fix)
- None.
#### Important (Should Fix)
- None.
#### Minor (Nice to Have)
- `src/styles/capture-workspace.css:970`: the same action-row and button selectors are declared earlier at `:882` and `:887`, then immediately overridden with different gap, padding and font-size values. Consolidate the final values into the first blocks to make the intended sizing clear and avoid future edits to dead declarations. Current rendering is correct.
### Checks
- Read the complete packaged diff, retrieving only output-truncated portions separately. Focused context checks covered the concrete risks of original-layout history visibility and the positioned help control's containing block; no other implementation exploration or test reruns were performed.
- Read `local-data/verification/annotation-workspace-layout/task03-final.log:1` and `local-data/verification/annotation-workspace-layout/task03-submitted-readback.log:1`; their observed results agree with the report. The reported build chunk notice is known baseline noise, not a Task03 regression.
### Assessment
**Task quality:** Approved.
**Reasoning:** The change adapts the existing editor and transport through scoped presentation rules, with meaningful real-browser and authoritative API evidence for the task's state-preservation and submission requirements. Only a minor CSS consolidation remains.

View File

@ -0,0 +1,37 @@
### Spec Compliance
- ✅ Both previously reported Important findings are resolved in `94b238d5..e95adca8`.
- ✅ Cross-slice history location is supplied with the destination scope and revision, consumed only when that scope is rendered, and no longer written through the outgoing page setter (`src/components/CaptureAnnotationWorkbench.tsx:425`, `src/components/TimelinePageNavigation.tsx:19`). This preserves the destination page when redo closes its editor.
- ✅ Slice and contiguous-description boundary gestures capture their initial position, boundary time and scale; subsequent movements use that fixed scale and the existing global adjacent-frame limits (`src/components/CaptureAnnotationWorkbench.tsx:880`, `src/components/CaptureAnnotationWorkbench.tsx:1282`, `src/components/CaptureAnnotationWorkbench.tsx:1325`). Page-edge boundaries can move later or earlier without accumulating movement at the same pointer x.
### Strengths
- The history fix belongs to the presentation hook and uses a one-shot destination identity, allowing subsequent manual paging (`src/components/TimelinePageNavigation.tsx:19`).
- The drag correction preserves frame snapping and adjacent-interval constraints rather than bypassing them (`src/components/CaptureAnnotationWorkbench.tsx:1325`).
- The navigation browser scenario now changes slice before redo and waits for the editor to close before asserting page 4 (`tests/browser/annotation-workspace-navigation.mjs:196`).
- The new boundary browser scenario uses actual pointer capture, both directions, repeated identical x positions, and public API confirmation of the shared description boundary (`tests/browser/annotation-workspace-boundaries.mjs:80`).
### Issues
#### Critical (Must Fix)
- None.
#### Important (Should Fix)
- None remaining within the two requested findings.
#### Minor (Nice to Have)
- None identified within this scoped re-review.
### Checks
- Read the supplied fix diff and appended report. The combined output truncated part of the new browser harness, so reread only the missing range from the diff. No outside-file inspection, suite rerun, subagent or product mutation.
- Reviewed the reported real-browser RED/GREEN evidence: cross-slice null-editor redo reveals page 4; both boundary types move bidirectionally with stable repeated x; description API readback confirms the shared exact boundary at 36.46666666666667 seconds. Evidence was assessed without rerunning it.
### Assessment
**Task quality:** Approved for the two scoped fixes.
**Reasoning:** Destination-qualified history consumption corrects the render-scope race, and frozen pointer-delta mapping removes both page-edge clamping and gesture drift. The added browser cases directly cover the original failures.

View File

@ -0,0 +1,99 @@
# Task 04 — paged slice and description navigation
Status: implementation complete; controller review pending. Ticket ownership/resolution remains with controller. Current branch retained; no new worktree, push, deployment, company cluster or BOS write.
Commit: `94b238d575982a178ec2afcbc2dc6b8c82568e44` — `✨ Add paged timeline navigation with history target reveal`.
## Delivered and integration contract
- Focused slice and description timelines each show at most four chronological entries, with bounded previous/next buttons, page count, global item range, text/name search and current-item/global-ordinal location. Search finds the next match, wrapping across the full collection; `#13` selects exact global ordinal 13. It does not filter or mutate the collection.
- Timeline blocks retain proportional duration. Their pointer/frame mapping projects the visible contiguous time range; global playback sliders and Ctrl/Command+arrows still operate on the full capture/full slice and global adjacent slices. Tiny blocks can be opened directly by ordinal into the existing right editor.
- The reusable `TimelinePageNavigation` component/hook handles presentation only. Selection, changing lists, scope, history and empty final pages update a single derived page state. Manual page browsing leaves selection/draft alone and does not save.
- Existing editor and media nodes remain mounted. Parked drafts, B/V, original-layout switches and undo/redo remain in the existing workflow. Original layout retains its full timeline rendering.
- Added optional `revealAnnotation: {id,start}` presentation metadata to the existing `AnnotationEditState` and `historyReveal`. This records the affected description even when redo closes the editor or delete removes the object. No API payload/backend changes, no independent editor/history stack and no changes to authoritative mutation/session guards.
- Same 350px right region and 8px gap. Paging headers reuse existing timeline height; annotation timeline remains viewport anchored at bottom. Tasks05/06/07 can adapt the existing timeline plus navigation as one presentation area. No dictionary/long-text/single-camera/undersize behavior added.
## TDD and real browser acceptance
New delivered acceptance is `tests/browser/annotation-workspace-navigation.mjs`, real Edge with `--no-proxy-server`, localhost5173/API3001/snapshotMongo27018. It uses only the isolated many fixtures. No source-string or private-state assertions were added. Exploratory helper pure tests were used during development and removed from the deliverable at controller direction, respecting the accepted browser-first test boundary.
Observed REDs (raw logs ignored in local-data/verification/annotation-workspace-layout):
- `task04-red.log`: existing UI had no visible slice page controls, timed out waiting for page indicator.
- `task04-save-red.log`: ordinal `#1` incorrectly matched `#13`, leaving description list empty for the intended slice. Exact ordinal matching fixed.
- `task04-save-red-2.log`: parked description editor restored but page remained 1 rather than 4. Two competing selection/clamp effects replaced with one derived page state.
- `task04-redo-red.log`: saved redo closed the editor and showed page1, hiding the affected description on page4. Explicit presentation location metadata fixed.
- `task04-history-red.log` was a harness timing failure: API response preceded React reconciliation; subsequent screenshot showed page4. Harness now waits for visible expected page before asserting. Slice V's initial harness expectation was also corrected: V must return to the original global work position (page1), not the last reviewed page4.
Final browser commands, exit 0:
```sh
node tests/browser/annotation-workspace-navigation.mjs
NAVIGATION_PHASE=slice node tests/browser/annotation-workspace-navigation.mjs
```
Final raw logs: `task04-tiny-final.log`, `task04-slice-tiny-final.log`. Earlier complete many-list runs: `task04-annotate-final.log`, `task04-slice-final.log`. Final screenshots `task04-annotate.png` and `task04-slice.png`; annotation screenshot visually inspected.
| Content viewport | Each effective 590×480 picture |
| --- | --- |
| 1280×640 | 410.54×334 |
| 1366×680 | 459.71×374 |
| 1440×800 | 529×430.37 |
| 1920×980 | 769×625.63 |
| 2560×1320 | 1089×885.97 |
Both stages at all five sizes: image floor retained, dimensions grow with viewport, no document/main/workbench scrolling. All paging/search/current controls visible and hit-test reachable.
Real interaction coverage:
- 13 slices: Control+Right crosses page1→2; search #13 opens page4; Control+Left opens global slice12 on page3. No navigation API version changes.
- Delete sole last-page slice merges coverage, page4→3; undo from a different page reveals restored slice13/page4, redo reveals surviving slice12/page3. C creates a 14th slice at an exact frame, selected slice visible. B reveals selected page, V restores global original work position. Direct #13 search opens the one-frame (0.03s) slice in the right editor.
- 13 descriptions in slice1: edit description13 on page4; create/park another draft in blank slice2; return restores description13 text and page4. Original/focused toggles retain text and media DOM nodes, no implicit save.
- Save description13 via real API; navigate away from page; undo and redo both reveal page4 and API text matches original/saved value respectively. Deleting sole page4 description leaves 12 descriptions/page3; undo from elsewhere restores 13/page4. Parked slice2 draft remains intact and is explicitly cancelled.
- New single-frame description (start0/end1 frame) explicitly saved, then #14 search opens last description and #1 opens tiny description directly, right input/save reachable. Explicit deletion restores 13 descriptions. Public final readback: **version11→18, 13 slices, 13 descriptions**; exactly seven explicit save/history writes during final annotation run.
## Fixtures and persisted state
Controller created new owned fixtures only, preserving all earlier task/snapshot records:
- `annotation-task-workspace-layout-acceptance-many-slice-04`: 13 exact-frame slices covering0..412s, still slice phase/version1; browser splits/deletes were unsaved session edits, discarded on close.
- `annotation-task-workspace-layout-acceptance-many-annotate-04`: 13 exact-frame slices, first0..120.1s with13 descriptions, second blank. Final saved version18, still annotation_combined/annotate, 13 descriptions; last description contains genuine Task04 saved text. The tiny acceptance description was explicitly deleted. No submission/generation/reset was performed.
Slice names use enabled project entries `布局验收切片 001..003`. Controller's ignored `prepare-many-fixtures.js`, `many-fixture-data.json` and captured public `frame-timeline.json` record guarded fixture construction. No credentials or raw fixture/media evidence tracked.
## Regressions and final checks
- Existing shell browser regression, exit0 (`task04-shell.log`): original entry, focused preference/recovery, continuous intent, stable media/frame, repeated toggle, per-slice drafts, delayed real save with pending restrictions, public saved/undo/redo readback, original C/history and actual same-browser account isolation. Used auth.json/secondary-auth.json and pool-01 for slicing; existing shell annotate-primary explicit-save fixture remains owned, not submitted.
- Existing focused frontend set: **56 passed, 0 failed**, final `task04-focused-tests.log`. Annotation shortcuts/history/authoritative mutations/confirmation boundaries/playhead draft/submission state/workbench suites.
- Final `npm run lint:web`: exit0.
- Final `npm run build:web`: exit0 (`task04-build.log`), existing chunk-size warning and local Node module.register deprecation only.
- Browser syntax, changed source/test Prettier and `git diff --check`: pass.
- Full frontend suite intentionally reserved for controller's once-at-end run; known baseline9 failures not rerun here.
## Remaining scope and review notes
No Task04 acceptance gap identified. Browser coverage verifies the five CSS content viewports; physical desktop/OS zoom matrix remains controller/later coverage. Proportional timeline bars can be narrow for tiny intervals; verified ordinal search selects them without pointer targeting, preserving timeline geometry instead of substituting equal-width cards. Paging does not filter data, so there is no hidden-filter state to reconcile. Task05 dictionary/long text, Task06 single-camera and Task07 undersize pause remain untouched.
## Review fix addendum — cross-slice history and page-edge dragging
Review findings in `task-4-review.md` were verified against code and real browser behavior before patching. The earlier “No Task04 acceptance gap identified” statement is superseded by this addendum: two omitted transitions were found, reproduced, fixed and added to browser acceptance. Controller re-review pending.
Fix commit: `e95adca8bb96e45f6eb732e6d4d1671cb2857d3d` — `🐛 Preserve cross-slice history reveal and stable page-edge drags`.
1. **Cross-slice closed-editor redo.** RED `task04-cross-slice-red.log`: save description13 in slice1, undo, navigate to slice2, redo; expected page4 never became visible. The old effect invoked a setter captured for the outgoing description scope, then changing slice with a null editor reset it. The hook now receives a scope-qualified target and consumes it once when the destination scope is rendered. Its consumed identity includes task/slice scope and history revision, preventing accidental revision reuse across scopes. Removed the outgoing-scope setter from the history effect. Existing mutation/session guards unchanged.
2. **Page-edge pointer drag.** RED `task04-boundary-slice-red.log`: dragging slice4's page-last edge +12px yielded192.833s, less than original193.0666667s. A valid contiguous-description setup through ordinary end editing and explicit “确认结束点” then reproduced the same defect: `task04-boundary-annotate-red-confirmed.log` yielded36.9s instead of extending past37s. The annotation RED temporarily reinstated only the old projection for this verification, then restored the fix. Both drag refs now capture gesture start x, original shared-boundary time and seconds-per-pixel on pointer down. Moves translate deltas against this frozen scale; the existing global adjacent frame limits still clamp/snap the result. Neither moving outside the page nor changing page duration redefines the gesture range.
Additional real browser acceptance: `tests/browser/annotation-workspace-boundaries.mjs` (default slice, `NAVIGATION_PHASE=annotate` for descriptions). Actual pointer capture crosses the visible page edge in both directions. Repeated pointer movement at the same x is asserted not to alter the boundary again. No new pure-function test seam added.
GREEN evidence:
- `task04-boundary-slice-green.log`: **193.0666666667 →195.666 →190.333** seconds in the fixed slice editor; same-x stable, API version remains1 (local slice edits never implicitly save).
- `task04-boundary-annotate-green.log`: **37 →37.466667 →36.466667** seconds, same-x stable. Explicit save public readback at **version31** confirms previous end and following start are both **36.46666666666667**, preserving the shared exact frame. The harness then restores the two intervals to their immediately pre-run values through normal explicit saves, finalversion33 at that point.
- `task04-cross-slice-final.log`: updated existing full navigation scenario now navigates to slice2 before redo, waits for the editor to detach and requires destination slice1/page4. Final complete run **version40→47**, 13slices/13descriptions. All five picture sizes unchanged; drafts, empty final pages and single-frame ordinal search still pass.
- `task04-review-fix-slice-navigation.log`: complete slice five-size/history/tiny-ordinal scenario exit0.
- `task04-review-fix-shell.log`: original/focused layout, delayed actual save and API restoration, task recovery/continuous intent, stable media/frame/drafts and real account isolation all exit0.
- `task04-review-fix-focused.log`: existing **56 tests pass,0 fail**. Final typecheck and build exit0 (`task04-review-fix-build.log`); browser syntax, changed-file Prettier and diff whitespace pass. Full suite remains reserved for controller.
Fixture note: the first valid annotation RED explicitly closed the existing gap between descriptions4/5 to create their draggable shared edge. That persisted setup is intentional local fixture data: description4 ends37s, description5 begins37s. The later successful harness restores its starting values, so this new shared boundary remains available; no database reset occurred. Final many-annotate-04 version47, still annotation_combined/annotate with13 descriptions. many-slice-04 remainsversion1 in slice phase. Earlier fixtures were only used by the established shell regression; no tasks generated or submitted.
Harness-only interruptions: an initial annotation setup attempt omitted the existing explicit end confirmation and timed out before saving; corrected to click “确认结束点”. One run overlapped Vite source formatting and returned to resume UI; rerun after file changes settled passed. Neither is claimed as a product RED. Root can re-review only this scoped follow-up commit. No push/deployment/company/BOS write.

View File

@ -0,0 +1,39 @@
### Spec Compliance
- ❌ Issues found: cross-slice history restoration can hide the affected description, and paginated boundary dragging no longer preserves the existing duration-edit semantics (`src/components/TimelinePageNavigation.tsx:16`, `src/components/CaptureAnnotationWorkbench.tsx:773`, `src/components/CaptureAnnotationWorkbench.tsx:887`, `src/components/CaptureAnnotationWorkbench.tsx:1298`).
- ✅ The implementation bounds both timelines to four chronological entries, exposes page/search/current controls, retains global slice navigation, and keeps pagination separate from the underlying collections (`src/lib/annotation-list-navigation.js:3`, `src/components/TimelinePageNavigation.tsx:60`, `src/components/CaptureAnnotationWorkbench.tsx:740`).
- ⚠️ Five-viewport picture geometry, account/session transitions and real API behavior cannot be established from this diff alone. The report supplies browser/API and existing regression evidence; the added browser scenario does not exercise cross-slice closed-editor redo or boundary dragging (`tests/browser/annotation-workspace-navigation.mjs:78`, `tests/browser/annotation-workspace-navigation.mjs:196`).
### Strengths
- Exact ordinal search selects otherwise tiny timeline entries without changing chronological geometry, and search does not create a filtered collection (`src/lib/annotation-list-navigation.js:16`, `src/components/TimelinePageNavigation.tsx:47`).
- History location metadata is optional presentation state; the changed parent code only forwards it to the workbench, with no new backend payload fields (`src/pages/AnnotationOperations.tsx:2206`).
- Browser acceptance checks actual DOM reachability, parked draft restoration, empty-tail deletion and explicit API version changes instead of source-string assertions (`tests/browser/annotation-workspace-navigation.mjs:37`, `tests/browser/annotation-workspace-navigation.mjs:163`, `tests/browser/annotation-workspace-navigation.mjs:205`, `tests/browser/annotation-workspace-navigation.mjs:245`).
### Issues
#### Critical (Must Fix)
- None identified.
#### Important (Should Fix)
- `src/components/CaptureAnnotationWorkbench.tsx:773` and `src/components/TimelinePageNavigation.tsx:16`: a history reveal into another slice writes the target page through the outgoing slice's `descriptionPage.setPage` closure. The effect also selects the affected slice and may clear its editor. On the next render the hook sees the new scope, no selected description and the old scope in its state, so it replaces the requested page with zero. For example, save description 13 in slice 1, undo, move to slice 2, then redo: redo closes the editor and reveals slice 1 page 1 instead of page 4. Reapplying the effect cannot repair it because `appliedHistoryRevision` already consumed the revision. Apply the reveal as a scope-qualified target inside the hook, or after the target scope has rendered. Add browser acceptance for history restoration from a different slice with a null editor.
- `src/components/CaptureAnnotationWorkbench.tsx:887` and `src/components/CaptureAnnotationWorkbench.tsx:1298`: the new pointer projection uses the current visible page's first start and last end, while `pointForEvent` clamps the pointer ratio to 0..1 (`src/components/CaptureAnnotationWorkbench.tsx:196`). The last visible item still exposes its boundary with the first item on the next page (`src/components/CaptureAnnotationWorkbench.tsx:2242`, `src/components/CaptureAnnotationWorkbench.tsx:2760`), but its current end is now the maximum pointer time, so the user cannot drag that boundary later. Dragging it earlier also changes the projection end during the same gesture, causing subsequent pointer events at the same position to shrink it again. This affects slice and contiguous-description page boundaries and violates retained drag semantics. Use a stable gesture coordinate range that permits both neighboring intervals, or translate pointer deltas against a frozen range and clamp to the real adjacent interval limits. Verify both directions on a page boundary in the browser.
#### Minor (Nice to Have)
- None identified beyond the focused acceptance cases required by the Important findings.
### Checks
- Read the supplied diff once in three contiguous pages; no git commands, product edits or suite reruns.
- Read the complete history effect because its changed hunk began mid-function (`src/components/CaptureAnnotationWorkbench.tsx:744`); confirmed the cross-scope setter/revision ordering above.
- Named risk: presentation history crossing mutation/session boundaries. Read the surrounding restore function omitted by the parent hunk; the new metadata is forwarded after the existing restoration/confirmation path, with existing busy/ownership gates unchanged (`src/pages/AnnotationOperations.tsx:2166`). Full session behavior remains covered by the controller's reported regression evidence, not independently re-executed here.
- Named risk: projected pointer ranges losing existing boundary reachability. Read `pointForEvent` and the boundary-handle JSX omitted between diff hunks (`src/components/CaptureAnnotationWorkbench.tsx:188`, `src/components/CaptureAnnotationWorkbench.tsx:2242`, `src/components/CaptureAnnotationWorkbench.tsx:2760`); confirmed clamping and that page-last handles remain enabled against global neighbors.
### Assessment
**Task quality:** Needs fixes.
**Reasoning:** The paging/search architecture and browser coverage address the main workflow well, but two uncovered transitions break explicit requirements: cross-slice history target visibility and bidirectional page-edge boundary editing.

View File

@ -0,0 +1,36 @@
### Spec Compliance
- ✅ The previously reported Important focus/dismissal regression is resolved. `src/components/WorkspaceDictionaryPanel.tsx:38` dismisses on focus leaving the panel; both callers supply presentation-only dismissal at `src/components/CaptureAnnotationWorkbench.tsx:3011` and `src/components/SliceNameDictionaryPicker.tsx:135`.
- ✅ Explicit Escape/return continues through the existing `onClose` callback, while blur uses `onDismiss` without moving focus. This preserves the description selection callback's editor focus and avoids reinstating the dialog keyboard guard after Tab-out.
### Strengths
- `src/components/WorkspaceDictionaryPanel.tsx:38` checks containment of `relatedTarget`, so navigating between search, results and paging controls keeps the panel open.
- `tests/browser/annotation-workspace-dictionaries.mjs:123` adds the exact Tab-out regression and outside-pointer departure for both dictionaries; subsequent assertions cover one Escape, Space playback, unchanged media geometry/version, zero writes and preservation of the unsaved description.
- `tests/browser/annotation-workspace-dictionaries.mjs:106` additionally protects the existing template-selection focus contract.
### Issues
#### Critical (Must Fix)
- None.
#### Important (Should Fix)
- None. Previous finding closed.
#### Minor (Nice to Have)
- None in this scoped fix.
### Checks and Scope
- Read the complete `task-5-fix1.diff` once and the appended report. Reviewed only the previously reported focus issue and its regression checks; no broader code exploration, source mutation or tests rerun.
- Read `local-data/verification/annotation-workspace-layout/task05-focus-green.log`: both slice and annotate focus-departure checks pass. Read `task05-focus-typecheck.log`: TypeScript check exits successfully as reported.
- Native OS IME and desktop zoom limits from the original review remain unchanged and are outside this focus fix.
### Assessment
**Task quality:** Approved.
**Reasoning:** Separate dismissal and explicit-close callbacks repair the lost-focus keyboard failure while preserving the existing selection and focus ownership. The targeted real-browser checks directly cover the reproduced failure for both dictionary callers.

View File

@ -0,0 +1,88 @@
# Task 05 — fixed dictionaries and long descriptions
Status: DONE; controller review pending. Current feature branch retained. No new worktree/branch, push, deployment, company-cluster/BOS write, or backend semantic change.
Commit: `1de1d85b1451b7ac489e362e12d37eb9339e7de0` — `✨ Keep dictionaries and long descriptions in the fixed editor`.
## Delivery and contracts
- Both focused dictionaries replace content inside the existing fixed right editor. Search, previous/next page, return and keyboard help stay visible; every page has at most **2 options**, with two-line previews. Full text remains the accessible button name and title; selection fills the entire value. A selected slice name also has its full title. Long action text remains readable in the existing scrolling textarea. Search includes both action name and description.
- `WorkspaceDictionaryPanel` owns only search/page presentation. Existing slice-name callback, description `chooseTemplate`, dictionary key handler, save, version/session and history paths remain authoritative. No additional editor/history implementation. Empty/loading/error and slice refresh use existing state/callbacks.
- F/D, up/down, Enter and one Escape retain their ownership. Escape restores the workbench, so Space resumes playback rather than activating a dictionary trigger. Choosing a template does not save; it remains editable.
- Long text scrolls inside the same mounted textarea. “展开描述编辑” hides the precise-boundary/help row to enlarge that textarea inside the same fixed right editor. Save/cancel remain outside text scrolling. “收起描述编辑”, layout switching and parked draft navigation do not save. Expanded presentation resets when the edited description changes; the draft itself remains in the established editor coordinator.
- Original layout retains the original dictionary popups and original description input presentation. No changes to Task04 timeline paging, pointer mapping, scoped history reveal, media ownership, frame semantics, backend validation or submission flow. Task06/07 remain out of scope.
## TDD and real browser evidence
`tests/browser/annotation-workspace-dictionaries.mjs` uses real Edge with `--no-proxy-server`, localhost5173/API3001/Mongo27018 isolated fixtures and authenticated public API readback. No mocks, new pure dictionary tests or source-string assertions.
RED `task05-red.log`: actual existing slice popup failed “dictionary replaces content inside fixed editor”.
Visual-review RED `task05-visible-red.log`: the old slice input flex-basis rule made the new search field too tall, hiding the result list. Added actual result hit-testing; it failed “long dictionary result is visibly readable, not only keyboard-selectable”. Corrected with a scoped fixed 28px search field, two-item pages and bounded previews. Final screenshots were visually inspected: `task05-slice-dictionary.png`, `task05-annotate-dictionary.png`, `task05-expanded.png` (all ignored). Earlier intermediate harness failures were an asynchronous dictionary-refresh count race and unnamed fixture slices blocking generation review; harness now waits for actual option visibility and names remaining slices through UI before opening review. No fixture reset or generation was used.
Final command, exit0:
```sh
node tests/browser/annotation-workspace-dictionaries.mjs
```
Raw final log: ignored `local-data/verification/annotation-workspace-layout/task05-final.log`.
Verified:
- Two dictionaries: actual containment in fixed editor, unchanged video rectangles, multi-page navigation, no-match search, long-name/template search, up/down/Enter, both F and D, one Esc then Space, original-layout popup fallback. Every current-page option and long result is visibly hit-test reachable; search/page/return and no document scroll at all five sizes.
- Real 112-character slice label and 1000-character action template are selected in full. Long slice-name generation review keeps return/confirm reachable; review is canceled without generation.
- Real browser Meta+A/Meta+X cut and Meta+Z native undo across expansion of the same textarea. Supplemental browser-dispatched composition events prove application shortcut ownership only; **not a native OS IME test**.
- A 2265-character Chinese draft has internal text scrolling. Expansion increases textarea height; save/cancel/submission remain reachable. Same textarea and both media nodes survive original/focused toggles. Cross-slice parking restores exact text.
- Draft submission guard retains long text; explicit Ctrl+Enter persists exact content via existing public API. Visible undo and redo restore exact prior/saved content. Expanded cancellation discards only unsaved text, then reopening yields persisted text.
- Request observation asserts zero task/claim writes during dictionary, text/layout/navigation interactions, exactly three writes for explicit save/undo/redo. Final authoritative annotate-primary version **58 → 61**. Earlier accepted runs advanced its legitimate version; no reset. pool-01 remains slicing and version-unchanged; its chosen names are unsaved browser changes discarded on close. No new tasks claimed or fixtures consumed.
| CSS content area | Effective picture per camera |
| --- | --- |
| 1280×640 | 410.54×334 |
| 1366×680 | 459.71×374 |
| 1440×800 | 529×430.37 |
| 1920×980 | 769×625.63 |
| 2560×1320 | 1089×885.97 |
## Verification
- `node tests/browser/annotation-workspace-annotation.mjs`: exit0 after initial implementation; five-size collapsed editor bounds, actual stale-version rejection/draft retention, save/undo/redo readback, existing cancel and cross-slice draft/confirmation guards. Ignored `task05-annotation-regression.log`. Later production adjustments only affect inline dictionary sizing/previews and selected-name title; final Task05 browser runner covers those.
- `npm run lint:web`: exit0 on final implementation (`task05-typecheck.log`).
- `npm run build:web`: exit0 on final implementation (`task05-build.log`); existing >500kB chunk warning only.
- Existing annotation-shortcuts/history/authoritative-mutation/confirmation-boundaries/playhead-draft/submission-state tests: **47 passed, 0 failed** (`task05-focused-tests.log`). No affected pure logic changed after that run.
- Browser syntax, changed source/test Prettier and `git diff --check`: pass.
- Full suite intentionally reserved for controller's once-at-end run, preserving known baseline debt.
## Files and limits
Changed: `src/components/WorkspaceDictionaryPanel.tsx`, `SliceNameDictionaryPicker.tsx`, `CaptureAnnotationWorkbench.tsx`, `src/styles/capture-workspace.css`, `tests/browser/annotation-workspace-dictionaries.mjs`, `docs/annotation-workspace-layout.md`.
The panel is absolutely contained by `.capture-stage-editor` for slicing and the existing description editor for annotation. Hidden editor siblings retain mounted draft nodes; visible dictionary content receives input ownership. Subsequent single-camera/undersize changes should retain these containing regions and the same textarea. Original layout bypasses the new panel entirely.
Two-line previews trade simultaneous full option text for reliable option/search/page access at the minimum viewport. Full accessible names/title and selected draft preserve the entire text. Native OS Chinese IME and desktop/zoom matrices were not available in headless acceptance; later/controller coverage remains responsible for these boundaries. No other known Task05 gaps.
## Review fix round 1 — focus departure
Commit: `6c76e5d2ca7744b104a3335568815e20e51ddc32` — `🐛 Dismiss focused dictionaries when keyboard focus leaves`.
Read `task-5-review.md`. Confirmed the Important finding independently in real Edge: focus the slice dictionary's 下一页, press Tab then Escape; dialog count remains **1**. The shared panel lacked the original description popup's blur dismissal, while the existing workbench dialog guard correctly reserved keys for the open panel. RED evidence: ignored `task05-focus-red.log`.
Added `onDismiss` to the presentation panel and both callers. A blur whose related target is outside the panel closes its presentation state without moving focus. Explicit Escape/return still uses the existing close-and-focus callback. Keeping these separate preserves the existing `chooseTemplate` focus on the description editor; blur cannot override it with workbench focus. No save, history, media, original popup or global keyboard dispatcher changed.
Extended `tests/browser/annotation-workspace-dictionaries.mjs` with the exact Tab-out path and an outside-video pointer path for both dictionaries. Each then presses one Escape and Space, checks workbench focus/playback, unchanged video rectangles, unchanged authoritative version and zero writes. Also asserts template selection focuses the existing description editor and departure retains its unsaved text. `ANNOTATION_DICTIONARY_FOCUS_ONLY=1` runs this bounded slice of the runner without save/undo/redo or long-text acceptance.
The first reproduction attempt timed out before media because the owned-task resume list had not settled when the harness clicked Continue. The screenshot showed the resume screen, not a media/runtime failure. The runner now waits for and explicitly clicks the intended owned fixture row before Continue; no task reset, stack restart or new claim was used.
Commands/results on the final fix:
```sh
ANNOTATION_DICTIONARY_FOCUS_ONLY=1 node tests/browser/annotation-workspace-dictionaries.mjs
# exit 0; PASS ... slice; PASS ... annotate
npm run lint:web
# exit 0
node --check tests/browser/annotation-workspace-dictionaries.mjs
# exit 0
```
Logs: ignored `task05-focus-green.log`, `task05-focus-typecheck.log`. `git diff --check` and changed-file Prettier passed. No full suite or build rerun for this scoped focus-handler fix. Both fixture versions remain unchanged from their pre-run values (annotate-primary remains 61); no write/claim occurred. Ready for scoped re-review.

View File

@ -0,0 +1,39 @@
### Spec Compliance
- ❌ Issues found: the focused description dictionary does not retain the existing focus/one-Escape contract when keyboard focus leaves it. `src/components/WorkspaceDictionaryPanel.tsx:31` omits blur dismissal; see Important finding below.
- ✅ Both dictionary presentations use the existing selection callbacks, retain original-layout popups, add local search and two-item paging, and keep expansion as presentation state: `src/components/CaptureAnnotationWorkbench.tsx:2996`, `src/components/CaptureAnnotationWorkbench.tsx:3343`, `src/components/SliceNameDictionaryPicker.tsx:109`, `src/components/WorkspaceDictionaryPanel.tsx:23`.
- ⚠️ Native OS Chinese IME and desktop zoom matrices remain unverified; the report correctly labels dispatched composition events as supplemental. The parent controller supplied minimum-size screenshot verification; the existing final browser log records five-size geometry and real save/undo/redo version 58 → 61.
### Strengths
- `src/components/WorkspaceDictionaryPanel.tsx:6` owns presentation only; `src/components/CaptureAnnotationWorkbench.tsx:3006` delegates to the unchanged `chooseTemplate`, whose body at line 967 updates the draft and closes the dictionary without saving.
- `src/components/CaptureAnnotationWorkbench.tsx:3343` expands the same mounted textarea, preserving native text editing; `src/styles/capture-workspace.css:1146` constrains the dictionary search input to 28px so existing slice-input flex rules cannot hide results.
- `tests/browser/annotation-workspace-dictionaries.mjs:1` exercises the real local browser/API, visible hit testing, native cut/undo, layout preservation, and explicit mutation counts. Read `local-data/verification/annotation-workspace-layout/task05-final.log`; its reported run is successful. No broad suite was rerun.
### Issues
#### Critical (Must Fix)
- None.
#### Important (Should Fix)
- `src/components/WorkspaceDictionaryPanel.tsx:31`: the new focused description panel lacks the original popup's blur-close behavior (`src/components/CaptureAnnotationWorkbench.tsx:3294`) and does not contain focus. Reproduced against the local browser: open the action dictionary, focus “下一页”, press Tab; focus moves to “撤销” while the dictionary stays open. Press Escape: the dictionary remains open. The existing workbench dialog guard at `src/components/CaptureAnnotationWorkbench.tsx:1522` returns before handling Escape or other workbench keys. This breaks the required single-Escape return and leaves keyboard workbench actions suppressed until the user refocuses/closes the panel by mouse. Restore the existing dismissal contract when focus leaves, or otherwise ensure the visible panel still receives Escape after focus moves; add this specific Tab-out/Escape interaction to the browser check. The same reusable panel also lacks keyboard-focus dismissal for the slice dictionary, although its existing outside-pointer handler still dismisses mouse departures.
#### Minor (Nice to Have)
- None.
### Checks and Scope
- Read the complete 925-line supplied diff once, in four pages. No git commands, tracked changes, or broad codebase search.
- Named risk: changed inline presentation might bypass existing selection/save and keyboard ownership contracts. Focused checks inspected `chooseTemplate`, `closeDescriptionDictionary`, the workbench dialog guard, `dictionaryKeyDown`, and slice outside-pointer dismissal. These confirm unchanged save semantics and expose the focus regression above.
- Re-read only the original description-popup blur hunk cut off by diff context, at `src/components/CaptureAnnotationWorkbench.tsx:3287`.
- Narrow live reproduction completed successfully with no task data edits or saves. Initial sandbox browser launch failed; the authorized outside-sandbox local browser check succeeded and printed `afterTab ... 撤销 ... dialogs: 1` and `afterEscape 1`.
- Existing build chunk warning is documented prior debt, not a new finding.
### Assessment
**Task quality:** Needs fixes.
**Reasoning:** The fixed geometry, callback reuse and real persistence checks are sound, but the new description dictionary regresses keyboard dismissal whenever focus leaves its panel. Repair this focused interaction before accepting Task 05.

View File

@ -0,0 +1,36 @@
### Spec Compliance
- ✅ Both previously reported Important findings are resolved in `ee3d6b16..af7df833`.
- `src/styles/capture-workspace.css:1196`: inspection now reserves a 28px grid row before the 64px transport row and explicitly places transport in row 4. The absolute progress surface occupies this reserved space instead of overlapping the video. The corrected reported 531×432 effective picture still exceeds 490×398 at 1280×640.
- `src/components/CaptureAnnotationWorkbench.tsx:317`: focused visibility derives both available streams independently of original checkbox preferences. The derived flags consistently feed transport (`:635`), layout (`:1442`), inspection controls (`:1910`), and media source/visibility/preload (`:1984`). Raw original-layout preferences remain intact.
### Strengths
- `tests/browser/annotation-workspace-camera-inspection.mjs:141`: the new regression checks measure progress/video separation and exercise hiding each original camera before returning focused at a nonzero frame. They verify source recovery, readiness, synchronization, frame continuity, original preferences, mounted media nodes, and draft preservation.
- `docs/annotation-workspace-layout.md:20`: documentation now reports corrected image geometry and accurately distinguishes focused inspection from reconnecting a source previously disabled in the original layout.
### Issues
#### Critical
- None.
#### Important
- None in this scoped fix review.
#### Minor
- None requiring follow-up within this scope.
### Checks and Scope
- Read the complete fix diff once and the updated Task 06 report. No broader source inspection, test rerun, subagent, product edit, or Git mutation.
- Reported browser evidence includes zero progress overlap, held frame 6, successful hide-left/hide-right recovery, synchronization within 0.05 seconds, preserved drafts/preferences/media nodes, and no task writes. The controller separately confirmed successful left/right screenshots.
- The 79 focused tests, type check, and build are reported passing evidence; the build warning remains baseline debt. Existing save/readback, native OS IME/zoom, and Task 07 scope limits are unchanged by this approval.
### Assessment
**Task quality:** Approved for the scoped fixes.
**Reasoning:** The fixes directly address both observed failures without adding a second transport/editor or overwriting original visibility preferences. Targeted browser regressions cover the failure paths and corrected geometry.

View File

@ -0,0 +1,94 @@
# Task 06 — single-camera inspection and aspect fit
Status: implementation and first review fixes complete; controller re-review pending. Ticket ownership/resolution remains with controller. Current feature branch retained; no new worktree/branch, push, deployment, company-cluster/BOS write, fixture reset, or backend business change.
Commits:
- `ee3d6b16bf0c18059487d23b00fd73a73aab0d76` — `✨ Add single-camera inspection and aspect fit`.
- `af7df833b6152051805114e2852a727b5ed40ba0` — `🐛 Keep camera inspection clear and recover both views`.
## Delivered
- Focused slicing and annotation now expose explicit `放大左相机`, `放大右相机`, and `恢复双画面` controls. The state belongs to the existing workbench presentation; no task/session/media coordinator or transport fork was added.
- Inspection keeps both existing `<video>` nodes. Both sources are active in focused layout; if an original-layout preference had detached one source, the existing transport reconnects it and seeks it to the current frame. The unselected camera is visually hidden only during single-camera inspection, remains on the same transport clock, and was observed within 0.05 seconds of the selected camera.
- The selected video spans the left region's former detailed-timeline row. The Capture range remains as a reserved 28px progress row outside the image and above the existing media controls; detailed slice and description timelines stay mounted but hidden and return on restore.
- The right editor/footer grid is unchanged, so current editing, save/cancel, undo/redo and stage submission remain hit-test reachable. Slicing reuses the existing `splitAtPlayhead` handler in the inspection controls because its normal timeline button is hidden with the detailed timeline.
- Restoring dual view and focused/original toggles retain the exact frame, active draft/editor node, media nodes, selected work, and local history. Original-layout left/right visibility remains the saved user preference; focused layout derives effective dual visibility without overwriting it. Returning to original layout restores the user's hidden-camera choice, and focused inspection controls do not appear there.
- Existing `object-fit: contain` remains authoritative. The new browser runner validates the real pair first and uses the prepared synthetic files only in a fresh cacheless context with exact `subject-content` response mapping.
## TDD red/green evidence
The first browser test was written against base `db5e6d23`. It reached annotate-primary with a live unsaved draft and failed waiting for the missing `恢复双画面` control; the ignored `task06-error.png` showed the baseline dual 590×480 workbench with no inspection actions. That initial run did not exit cleanly after the expected locator wait and was interrupted with exit 130; no product file had been changed.
Incremental observable reds after the first minimal implementation:
- `node tests/browser/annotation-workspace-camera-inspection.mjs`: exit 1, `完成任务并提交质检 is visible and hit-test reachable`. Shrinking the shared grid row also squeezed the right footer. The fix kept the right grid rows and made only the left video panel span the timeline row.
- Next run: annotation phase passed; slicing exited 1 waiting for `在播放头分割切片`. Compression had hidden the stage action with the detailed timeline. The fix rendered the same existing handler beside inspection controls only while slicing inspection is active.
- Synthetic attempts initially saw the original 590×480 pair. Diagnostics established that the public video endpoints are `.../subject-content` and `...?side=right`; the final runner uses a fresh context and maps exactly those two requests. This was harness wiring, not a product change.
The first review reproduced two failures at nonzero frame 6 before changing the fix:
- The absolutely positioned 28px progress surface covered 22px of the selected video box. Reserving a real 28px grid row between the picture and 64px transport row reduced the measured overlap to exactly 0px.
- Hiding either original-layout camera detached its source, and focused layout inherited that preference, leaving the corresponding inspection control disabled and the camera at time 0. Focused layout now derives effective dual visibility, while the raw checkbox preference remains unchanged. The existing media transport reconnected each camera, returned it to frame 6, and synchronized the pair within 0.05 seconds. Returning to original layout restored the hidden checkbox preference.
Final green:
```sh
node tests/browser/annotation-workspace-camera-inspection.mjs
# exit 0
# PASS real dual/single camera geometry, state continuity and public API non-mutation
# PASS slicing controls, exact frame/pointer mapping, mounted media and history continuity
# PASS SYNTHETIC-ONLY 16:9 / 9:16 differing-camera contain fit; not real customer coverage
```
No private React state mutation, mock editor, new pure layout seam, save, generation, submission, release, or claim was used. One existing source-contract test was updated from raw visibility flags to the derived effective flags; all review behavior is proved by the real browser/API workflow rather than that assertion. Public human-item versions for annotate-primary and pool-01 were read before/after and asserted unchanged; the request observer saw zero task/claim writes.
## Real geometry and behavior
The principal measurements use the actual two 590×480 videos. Values are effective picture dimensions, excluding letterbox space.
| CSS content area | Each dual picture | Left inspection picture |
| --- | --- | --- |
| 1280×640 | 410.54×334 | 531×432 |
| 1366×680 | 459.71×374 | 580.17×472 |
| 1440×800 | 529×430.37 | 727.67×592 |
| 1920×980 | 769×625.63 | 948.92×772 |
| 2560×1320 | 1089×885.97 | 1366.83×1112 |
At 1280×640 both cameras exceed 400×325 in dual mode and either camera exceeds 490×398 in inspection. All five sizes grow monotonically, retain `contain`, and introduce no document/workbench scrolling. The browser also exercised play/pause while one camera was visually hidden, exact keyboard frame stepping, same-relative-position range pointer mapping before/after restore, annotation draft/editor identity, original-layout camera controls, slicing cut history followed by undo/redo, and the visible save/submit/current-stage actions.
Successful minimum-size visual QA is retained in ignored local evidence:
- `local-data/verification/annotation-workspace-layout/task06-minimum-dual.png`
- `local-data/verification/annotation-workspace-layout/task06-minimum-left-inspection.png`
- `local-data/verification/annotation-workspace-layout/task06-minimum-right-inspection.png`
The screenshots show the complete 590×480 picture in each state, with the progress row below the image and above transport. Recovery diagnostics are also retained as `task06-review-left-recovery.png` and `task06-review-right-recovery.png`.
## Synthetic aspect fixtures
Ignored local files used only for adaptation evidence:
- `local-data/verification/annotation-workspace-layout/aspect-fixtures/landscape-16x9.mp4`: synthetic testsrc2 H.264, 640×360, 12 seconds, 30fps.
- `local-data/verification/annotation-workspace-layout/aspect-fixtures/portrait-9x16.mp4`: synthetic testsrc2 H.264, 360×640, 12 seconds, 30fps.
In one differing-left/right pair at 1280×640, the 16:9 effective picture was 449×252.56 inside a 449×334 video box; the 9:16 picture was 187.88×334 inside the same-size box. Both retained intrinsic ratio and remained within the video box in dual view, left inspection, right inspection and restore. These files do not establish real customer-media, long-duration workflow or precise-frame coverage; those claims use the real 590×480/API workflow above.
## Verification
- `node tests/browser/annotation-workspace-camera-inspection.mjs`: exit 0 on final source, real Edge with `--no-proxy-server`, localhost 5173/API 3001/snapshot Mongo 27018.
- `npm run lint:web`: exit 0 on final source.
- `npm run build:web`: exit 0 on final source; existing >500kB chunk warning only.
- Eight affected frontend files via `node --test`: 79 passed, 0 failed on final source (workbench, shortcuts, history, timeline initialization/wire, playhead draft, video preview).
- Review reproduction aggregate: `coveredPicturePixels: 0`, `heldFrame: 6`, `failures: []`; hide-left and hide-right were each exercised from original layout and recovered in focused layout with both streams synchronized.
- `node --check tests/browser/annotation-workspace-camera-inspection.mjs`: exit 0 during implementation; final browser execution also parsed and completed it.
- Pre-commit changed-file Prettier and `git diff --cached --check`: pass.
- Repository-wide full suite remains reserved for the controller's single end-of-plan run.
## Files, fixtures and integration limits
Changed and committed: `src/components/CaptureAnnotationWorkbench.tsx`, `src/styles/capture-workspace.css`, `tests/browser/annotation-workspace-camera-inspection.mjs`, `tests/frontend/capture-annotation-workbench.test.mjs`, and `docs/annotation-workspace-layout.md`.
Fixtures used: annotate-primary for real annotation/draft/API checks and pool-01 for real slicing/history/API checks; both stayed in their existing owned editable stages and versions. The synthetic context reuses pool-01 metadata but replaces only browser media response bytes. annotate-delayed, slice-primary, many-annotate-04, many-slice-04 and company data were not touched.
Task 07 still owns below-1280×640 suspension/resume. Desktop OS/browser zoom matrices remain controller/manual coverage. Task 03's previously recorded duplicate CSS action-row cleanup was not on this directly touched seam and remains outside Task 06. No other Task 06 gap is known.

View File

@ -0,0 +1,39 @@
### Spec Compliance
- ❌ Issues found: inspection obscures the bottom of the effective picture instead of displaying it completely (`src/styles/capture-workspace.css:1199`); returning from the original layout with one camera disabled cannot restore dual video from the focused controls (`src/components/CaptureAnnotationWorkbench.tsx:1907`).
- ⚠️ Cannot verify from diff: a save followed by authoritative API readback across inspection/layout switches. The new runner explicitly avoids writes; the controller should retain the applicable earlier save/readback evidence. Native OS IME and desktop zoom acceptance are not established by this diff and are not Task 06 blockers. Task 07 suspension behavior is outside this review.
### Strengths
- `src/components/CaptureAnnotationWorkbench.tsx:317`, `src/components/CaptureAnnotationWorkbench.tsx:1512`, and `src/styles/capture-workspace.css:1184`: inspection uses presentation state and CSS visibility, preserving both existing media elements and the original transport rather than introducing another editor or playback implementation.
- `src/components/CaptureAnnotationWorkbench.tsx:1935`: slicing inspection reuses the existing split handler, retaining a visible stage action when the detailed timeline is hidden.
- `tests/browser/annotation-workspace-camera-inspection.mjs:1`: the browser coverage uses real media, public frame readouts, DOM identity, draft text, history, reachable controls, and API versions. Synthetic media is explicitly separated from real-media workflow claims.
### Issues
#### Critical (Must Fix)
- None.
#### Important (Should Fix)
1. `src/styles/capture-workspace.css:1199`: the compressed timeline is absolutely positioned with `bottom: 64px`, `height: 28px`, and `z-index: 30`, while the enlarged video panel still allocates only `22px minmax(0, 1fr) 64px` to camera controls, video, and media controls (`src/styles/capture-workspace.css:445`). The video extends down to the start of that same 64px media-control row, so the progress panel overlays its bottom 28px. The timeline retains its opaque panel background (`src/styles/capture-workspace.css:479`). With the real 590×480 picture fitting the inspection region by height, actual image content is hidden behind the slider; the reported bounding-box/contain calculation includes those obscured pixels. Reserve a separate progress-bar row or equivalent space outside the video region, then check that the actual fitted image rectangle does not intersect the progress panel and still meets 490×398. This violates the complete-picture requirement even though the numerical geometry assertions pass.
2. `src/components/CaptureAnnotationWorkbench.tsx:1907`: focused controls inherit `showLeft`/`showRight` from the original visibility checkboxes and disable enlargement of a hidden camera, but “恢复双画面” only clears `inspectionCamera` (`src/components/CaptureAnnotationWorkbench.tsx:1923`). Reproduce by hiding the right camera in the original layout, then entering focused view: the right source remains absent/hidden (`src/components/CaptureAnnotationWorkbench.tsx:1996`), the right inspection action is disabled, and restoring dual view cannot bring it back. The focused view has lost the visibility checkbox needed to repair this state. Keep original-layout visibility preferences intact while ensuring focused dual view has both media sources and usable left/right inspection controls. Add a browser case that actually changes each original visibility checkbox before returning focused, and verifies dual restoration and synchronization.
#### Minor (Nice to Have)
- None beyond adding the regression cases described with the blockers.
### Checks and Scope
- Read the supplied diff once in three sequential pages. No Git commands, product edits, fixture writes, subagents, or test-suite reruns.
- Named risk: original-layout visibility may leave focused inspection unable to display a camera. Checked only visibility-state setters and their transport/source consumers in `src/components/CaptureAnnotationWorkbench.tsx`; the diff ends mid-video JSX, so read its continuation to verify conditional `src` and `hidden` handling.
- Named risk: absolute progress positioning may occlude the enlarged image. Checked only the existing compact video-panel grid, video sizing, timeline background, and timeline markup in `src/styles/capture-workspace.css` and `src/components/CaptureAnnotationWorkbench.tsx`.
- The report's 79 focused tests, type check, and build are reported evidence, not independently rerun here. The existing bundle-size warning is baseline debt, as specified by the controller.
### Assessment
**Task quality:** Needs fixes.
**Reasoning:** The presentation-only approach preserves the main editing architecture, but the new layout covers image content and leaves a reachable fallback state without a working dual-camera restore. Both need targeted corrections and browser regression evidence before acceptance.

View File

@ -0,0 +1,36 @@
### Spec Compliance
- ✅ The scoped P1 is fixed. Both authoritative read boundaries now check `deferEntryForViewport()` before selecting a task, changing task lists, clearing detail/history, or calling `openWorkTask` (`src/pages/AnnotationOperations.tsx:3367–3389`).
- ✅ Deferred entry retains its requested resume-only/continuous mode and identity/permission/session validity, rather than retaining the authoritative read result. Explicit Continue invokes `startWork` again, so it reacquires ownership and rereads the owned pool (`src/pages/AnnotationOperations.tsx:3327–3343`, `4824–4835`).
- ✅ An existing active session must still be owned and retain continuous intent before deferred entry can resume; Continue does not restore a hard-paused continuous session (`src/pages/AnnotationOperations.tsx:3329–3341`).
### Strengths
- The fix adds guards at both async read boundaries and preserves the existing random-claim response path (`src/pages/AnnotationOperations.tsx:3367–3374`).
- The regression covers both delayed owned-pool and task-summary responses, unchanged URL/editor while suspended, no continuation on growth, explicit recovery, requested continuous intent, zero writes, and unchanged public versions (`tests/browser/annotation-workspace-viewport-owned-entry.mjs:33–87`).
### Issues
#### Critical (Must Fix)
- None found in the scoped fix.
#### Important (Should Fix)
- None found in the scoped fix; the original P1 is resolved.
#### Minor (Nice to Have)
- None raised.
### Checks and Scope
- Read the appended Independent-review P1 implementation report and the complete 233-line `cb74444e..972721e` review package once. Compared the fix against the original independently reproduced defect and the already-reviewed surrounding entry/session behavior.
- No additional outside-diff reads were needed. No unresolved code concern warranted another browser run; the new test exercises the precise boundaries that previously failed, and the report records RED/GREEN results for continuous and resume-only entry. No broad suites were rerun.
- No checkout/index/HEAD changes, fixture writes, or company-service operations. Only this review report was written.
### Assessment
**Task quality:** Approved
**Reasoning:** The missing owned-task transition guards are now present at both response boundaries, and explicit recovery revalidates authority while preserving intent and session restrictions. No residual finding remains within the scoped P1 re-review.

View File

@ -0,0 +1,83 @@
# Task 07 implementation report
Status: implemented; ticket remains claimed pending root review.
Implementation commit: `cb74444e0d641b7da2c29c6ff64f362b0ea724ab`. Base: `39bed0fc`. Branch: `feature/local-dev-toolkit`. Local commit only; no push, deployment, company-cluster or BOS writes. Root-owned `docs/verification/annotation-workspace-layout/full-suite-results.md` was left unstaged.
## Delivered behavior
- Browser content CSS width <1280 OR height <640 suspends the focused work page. Exact1280×640 remains usable. The original layout does not initiate this restriction.
- A page/scope viewport latch in existing `useAnnotationWorkSession` is independent of continuous intent, hard pause reason, and lock ownership. It survives failed initial entry releasing an inactive lock. It resets with the scope, not with layout changes or window growth.
- Existing coordinator dispatch boundaries gate NEW work entry/claiming and defer handoff continuation. Existing request/version/account/session fences remain authoritative for already dispatched requests; resize is absent from response-ownership gates.
- Authoritative completed handoff is retained while suspended. Explicit Continue can requeue a completed handoff if entry was suspended during its owned-pool lookup. Stop-after-current and hard pauses remain effective.
- Recovery requires Continue after restoring the viewport or explicitly choosing original layout. Switching layouts alone does not dismiss the latch. Continue clears only viewport suspension and does not play video, clear hard errors, enable continuous work, or retry an unknown claim.
- The existing workbench/media DOM stays mounted. Playback pauses through existing transport; task/frame/selection, active and parked per-slice drafts, history, and original camera preference are retained. Continue restores keyboard focus so Space explicitly plays; native dialogs regain their prior focus instead.
- Background workspace is inert while the recovery prompt is shown. Dirty leave confirmation temporarily takes precedence without clearing the latch. Native top-layer dialogs close temporarily during suspension and reopen on Continue, preserving the stage confirmation.
- Prompt states actual CSS content size, resize/browser-zoom action, retained work/in-flight reconciliation, explicit Continue, layout fallback, and return path. It never auto-switches layout or scales controls down.
## Real browser TDD evidence
All tests used actual Edge and localhost5173→API3001→isolated snapshot Mongo27018. All writes were public local API operations. Request gates call `route.fetch()` first, then delay/drop delivery of the real completed response; no source/private-state assertions or fake successful API replies.
1. Initial RED: `annotation-workspace-viewport.mjs` failed `below-width boundary pauses both playing videos` at1279×640; both actual videos continued playing. GREEN now covers width and height boundaries, exact supported boundary, stopped intent, active and parked drafts, media-node continuity, frozen time, explicit Continue followed by explicit Space, fallback latch, original layout resizing1100×600 without restriction, and public version non-mutation.
2. Unknown initial-claim RED: on fresh `workspace-viewport-unknown2`, completed claim response was dropped. Inactive session release erased the first latch implementation; growing the viewport made `prompt.isVisible()` false. Moved latch out of the lock-owned object into page/scope state. GREEN on fresh unknown3 retains recovery until Continue, issues no claim on Continue, and explicitly replays the original receipt with the identical request ID.
3. Dirty-leave RED: paused Return with unsaved drafts caused timeout locating the accessible leave-confirmation dialog; it lived in the inert/aria-hidden background. GREEN permits Cancel/Continue editing while undersized and restores the viewport prompt with both drafts intact.
4. Native-dialog RED: a real open submission confirmation occupied the browser top layer; hit-testing the viewport fallback button returned false. GREEN temporarily closes/restores the native dialog and verifies fallback hit-testing and restored submission confirmation.
5. Keyboard RED: immediately after Continue, pressing Space left both videos paused because focus had fallen to body. GREEN returns focus to the workbench only when no native dialog needs restoration; Continue itself still does not play.
A harness-only failed first save/unknown attempt called `unroute` before delayed `fulfill`/`abort` settled (`Route is already handled`). The gate now waits for delivery completion. Those attempts are not counted as product REDs. Their real server effects were retained, not reset.
## Browser/API cases and resulting fixtures
`node tests/browser/annotation-workspace-viewport.mjs`: PASS after final focus/dialog corrections. Primary annotate fixture version65 was unchanged by this non-write run. Both cameras stayed mounted, parked drafts survived, stopped-continuation remained stopped, and a second real browser tab could not acquire the active session after its own viewport Continue.
`node tests/browser/annotation-workspace-viewport-requests.mjs` (default save): final PASS on existing `annotation-task-workspace-layout-acceptance-annotate-primary`, v65→66 save→67 undo→68 redo. Exactly three POST annotation-confirmations with payload versions65,66,67. Server save completed before delivery; viewport/original switches and Continue left save disabled until delivery. API readback verified saved text, restored original text on undo, and saved text on redo. Prior first harness attempt v61→62 and first passing save/history run v62→65 are retained.
`VIEWPORT_CASE=claim`: PASS on `annotation-task-workspace-viewport-claim-pool-02`, claim v1→2, exactly one random claim with request ID `768fb0f43fa5bf6b38f9a8ed734cd7a9`. The successful response arrived while undersized and retained the allocated task. Growth did not claim; Continue retained it. This task was subsequently used for hard-restriction generation testing below. Pool count remains1.
`VIEWPORT_CASE=unknown VIEWPORT_UNKNOWN_PREFIX=workspace-viewport-unknown3`: PASS on `annotation-task-workspace-viewport-unknown3-pool-01`, v2. Exactly two transport attempts, both request ID `d696be07aedb165314c3fb06f47d2c80`; replay returns the same task. Public pool count remains1. Unknown2 RED left pool-01 v2 and count1; first harness unknown left pool-02 v2 and count1. No previous fixture reset, duplicate claim, or new receipt to evade unknown recovery.
`VIEWPORT_CASE=generation`: PASS on `annotation-task-workspace-viewport-generation-slice-primary`, v1→4. One real slice-confirmations POST, actual response status `processing`, final API `sliceWorkflow.phase=annotate`, `status=completed`. Continue during processing did not unlock continuous work or claim another task.
`VIEWPORT_CASE=submit`: PASS on `annotation-task-workspace-viewport-submit-annotate-primary`, v1→2 processing→3 quality_review. Exactly one combined-submissions POST. While undersized and after growth there were zero next claims. Explicit Continue produced exactly one claim, request ID `d2b529b3135121342bbceefef45d5051`, selecting `annotation-task-workspace-viewport-submit-pool-01` v2, slice phase. Public pool count0. Handoff was actual audit-data preparation completion, not QA approval.
`VIEWPORT_CASE=hard`: PASS using previously claimed `workspace-viewport-claim-pool-02`, v2→5. A secondary authenticated user's real slice-confirmations request returned403; primary fixture version stayed2 and permission-induced continuous pause survived viewport Continue. Then one real successful generation response was dropped and actual authoritative detail responses were fetched but delivery aborted. After Continue the visible unknown-stage notice remained, Restore continuous work and generation were disabled, and no replacement claim occurred. Restoring network delivery and explicit “核对阶段结果” reconciled to annotate v5. Two generation transport attempts total: one genuinely denied, one genuinely applied; no fake403/success body.
`node tests/browser/annotation-workspace-viewport-zoom.mjs`: final PASS. Actual isolated persistent Edge window1366×768, no page reload/CSS zoom/deviceScaleFactor substitution: zoom100% content1358×684,dpr1 →125% content1086×547,dpr1.25 →100%. Actual videos paused, draft survived, size restoration required Continue, and Continue left videos paused. Browser profile/credentials are ignored local data. Zoom restored100% and browser closed.
`node tests/browser/annotation-workspace-camera-inspection.mjs`: PASS on final production patch. At1280×640 real590×480 pair fits410.54×334 each; single picture531×432; progress overlap0. Five supported sizes, original camera visibility preferences, media/frame preservation, in-memory slicing undo/redo, and original layout controls all pass. Synthetic16:9/9:16 cases pass and are expressly synthetic. No writes or claims.
## Verification
- `npm run lint:web`: PASS after final production patch.
- `npm run build:web`: PASS after final production patch. Existing module.register deprecation and large-chunk warnings remain; no build error.
- Scoped existing frontend tests:35/35 PASS (`annotation-history`, `annotation-authoritative-mutation`, `annotation-random-claim`, `annotation-submission-state`, `media-seek-queue`). No new pure-function seam.
- Prettier on the three changed production files and three browser files: PASS.
- `git diff --check`: PASS before commit.
- Root owns final full frontend/backend suite comparison; prior frontend baseline578 tests /569 pass /9 known failures is not claimed resolved by this task.
Final public readback: layout annotate-primary v68; viewport-claim pool-02 v5 annotate, pool count1; viewport-submit annotate-primary v3 quality_review and pool-01 v2 slice, count0; viewport-generation slice-primary v4 annotate; unknown/unknown2/unknown3 each one claimed v2 slice and count1. Unowned task detail GET403 is expected blind-annotator policy, so availability was checked through public claim counts.
## Integration and limits
- Task07 owns only existing session hook/coordinator/workbench and three browser runners. No second editor/session/backend, schema, endpoint or new deployment requirement.
- Stateful request modes consume isolated one-shot fixtures. Prepare fresh prefixes through the established guarded local fixture workflow before repeating claim/submission/generation modes; do not reset earlier evidence. Default save, basic boundary and zoom checks can be repeated with their documented local public writes/non-writes.
- Full browser zoom/display-window matrix belongs to Task08. Actual zoom here covers one real browser-engine transition; it does not claim every physical monitor/OS scaling combination. Native OS IME was not tested here, remains the parent-documented limitation.
- `workspace-viewport-unknown` and `unknown2` retain their claimed tasks from harness/RED runs; their failed runner contexts closed. Their original receipt IDs were not logged before those failures; they were not retried using new IDs. The successful unknown3 receipt is recorded above.
- Task ticket remains claimed for root review. Ignored raw artifacts/logs stay under `local-data/verification/annotation-workspace-layout/`; no credentials or raw captures were committed.
## Independent-review P1 follow-up
Review read: `.superpowers/sdd/annotation-workspace-layout-plan/task-7-review.md`. The owned-first entry continuation lacked a viewport check after asynchronous ownership/summary reads. Accepted after independent real-browser reproduction; no other review findings were raised.
Fix commit: `972721e20c09aefd983613dc46bbf44e08a0da4d` (separate from original implementation commit).
- **RED pool:** `annotation-workspace-viewport-owned-entry.mjs` held delivery after real `GET /human-items?pool=mine` completion, then shrank1280→1279. Releasing that read changed URL from existing completed `viewport-submit-annotate-primary` to owned `viewport-submit-pool-01` under the still-active pause prompt. Assertion `pool response cannot switch the current task under suspension` failed.
- **RED summary:** `VIEWPORT_ENTRY_BOUNDARY=summary` independently delayed real `GET /tasks/…pool-01?view=summary` and reproduced the same premature task switch. Assertion `summary response cannot switch the current task under suspension` failed. Neither RED issued a non-GET request or changed a fixture.
- **Fix:** Check viewport suspension after both reads, before updating task lists, clearing history/detail, or opening a task. Retain only resume-only/continuous intent plus current identity/permission/session validity, not the read result. Inactive entry still releases its lock. Explicit viewport Continue reacquires/rechecks ownership and rereads authority before selecting an owned task. Existing active entry also requires the same active session and preserved continuous intent, so a hard pause cannot be bypassed. This is the same `startWork` path used by automatic post-handoff continuation. Already-dispatched random-claim response handling is unchanged.
- **GREEN:** Default runner passes both pool and summary delays. Before Continue: current URL remains unchanged, no new workbench mounts, and growing to1280 alone does nothing. After Continue: owned pool-01 opens with continuous work enabled, zero writes/claims.
- **GREEN resume-only:** `VIEWPORT_RESUME_ONLY=1 node tests/browser/annotation-workspace-viewport-owned-entry.mjs` passes both delays and opens owned pool-01 with continuous work still disabled. Public API versions remain exactly annotate-primary3 and pool-01 2 throughout all RED/GREEN checks.
- **Regression:** Existing `annotation-workspace-viewport.mjs` passes after this patch (primary fixture v68 unchanged), covering drafts, stopped intent, native confirmation restoration, dirty-leave recovery, original-layout latch semantics, explicit Space and competing-tab session restrictions.
- **Checks:** fresh `npm run lint:web`, `npm run build:web`, scoped Prettier and `git diff --check` all pass. Existing build chunk/deprecation warnings only. Broad suites were not redundantly rerun for this narrowly scoped follow-up.
No fixture creation/reset, public writes, company/BOS operations, push or deployment in this follow-up. Only the coordinator and new durable read-only browser regression were staged. Root verification docs remain untouched and unstaged. Ticket remains claimed for re-review.

View File

@ -0,0 +1,42 @@
### Spec Compliance
- ❌ Issues found: owned-first work entry can switch to another task while viewport suspension remains active. The new entry guard at `src/pages/AnnotationOperations.tsx:3293` does not cover the continuation after `listItems`/`getTaskSummary`; `openWorkTask` still runs at `src/pages/AnnotationOperations.tsx:3371`. This misses the requirement to retain the current task and defer new continuation until explicit Continue.
- ✅ Strict `<1280 || <640` thresholds, original-layout exemption, and explicit recovery latch are implemented at `src/pages/AnnotationOperations.tsx:513–533`; independent viewport suspension and preserved continuous intent are implemented at `src/hooks/useAnnotationWorkSession.ts:187–210`.
- ✅ Continue clears the viewport latch without starting playback or clearing hard pause state (`src/pages/AnnotationOperations.tsx:4804–4814`); playback pause and native-dialog suspension use the existing workbench (`src/components/CaptureAnnotationWorkbench.tsx:695–713`).
- ✅ New random-claim dispatch is checked before receipt allocation, while already-issued claim reconciliation retains its original ownership checks (`src/pages/AnnotationOperations.tsx:3221–3262`). Completed handoffs remain queued during suspension (`src/pages/AnnotationOperations.tsx:3175`, `3387–3396`).
- ⚠️ The unchanged save/generation/submission version/account fencing is not fully re-audited by this task-scoped diff. The new browser cases exercise real delayed responses and API readback (`tests/browser/annotation-workspace-viewport-requests.mjs:90–129`, `169–322`); the implementer's complete prior test runs were not independently repeated.
### Strengths
- The page/scope latch survives inactive lock release, without conflating viewport suspension with continuous-work intent or hard errors (`src/hooks/useAnnotationWorkSession.ts:36–39`, `187–210`).
- Recovery keeps the existing editor mounted and blocks background interaction, while dirty-leave confirmation remains reachable (`src/pages/AnnotationOperations.tsx:3477–3493`, `4488–4503`).
- Browser coverage uses real API effects with delayed delivery, original claim receipts, persisted undo/redo, and actual browser-engine zoom rather than synthetic success bodies (`tests/browser/annotation-workspace-viewport-requests.mjs:45–67`, `90–167`; `tests/browser/annotation-workspace-viewport-zoom.mjs:17–27`, `56–67`).
### Issues
#### Critical (Must Fix)
- None found.
#### Important (Should Fix)
- **P1 — Defer owned-task navigation after asynchronous entry reads.** `src/pages/AnnotationOperations.tsx:3293`, `3354–3371`: `startWork` checks the viewport only before acquiring the session. After the real owned-pool and task-summary responses arrive, `entryStillCurrent()` checks identity/permission/session ownership but not viewport suspension. The owned-task branch then clears history/detail and calls `openWorkTask` while the recovery prompt is still visible. The same function handles automatic post-handoff continuation, so a resize during its ownership lookup can replace the current task before the user explicitly continues. Unlike an already-issued random claim, this GET has not allocated a new task that must be reconciled into the editor. Add a viewport dispatch check before committing the owned-task transition, preserve/requeue the intended continuation, and resume it only after explicit Continue. Add a delayed owned-pool/task-summary browser case alongside the existing random-claim tests.
- **Actual-browser evidence:** isolated Edge, `--no-proxy-server`, local auth, initial viewport 1280×640. Opened the existing `project-workspace-viewport-submit` work page with task `annotation-task-workspace-viewport-submit-annotate-primary`, started its continuous-work entry, held delivery of the real `GET /human-items?…pool=mine` response, and shrank to 1279×640. After the pause prompt appeared, released the response without clicking Continue. The URL changed to `annotation-task-workspace-viewport-submit-pool-01`, a new workbench mounted, and the pause prompt remained visible. Zero non-GET requests occurred; a route guard blocked writes. No fixtures were reset or modified.
#### Minor (Nice to Have)
- None raised separately.
### Checks and Scope
- Read the supplied 1,076-line review package once in three sequential chunks, after the brief and implementation report. No git commands, suite reruns, index/HEAD changes, or company-service writes.
- **Named risk: an asynchronous entry may cross the new pause boundary.** The diff cuts `startWork` and `randomClaimTask` mid-function; read `src/pages/AnnotationOperations.tsx:3190–3435` to inspect their complete continuation/ownership gates. Read the entry-button handlers at `3590–3655` and `3670–3800` to select a public, read-only browser reproduction. This check confirmed the Important issue above.
- **Named risk: an existing transport seek could resume playback after pause.** Read `src/hooks/useCaptureVideoTransport.ts:1–155`, plus the workbench transport/seek integration at `src/components/CaptureAnnotationWorkbench.tsx:625–685`, `805–869`. `pause()` clears retained play intent; queue idle only plays when that intent remains true. No additional finding.
- **Named risk: temporarily closed native dialogs could reopen behind the pause prompt.** Checked the existing stage-dialog effect and keyboard entry points at `src/components/CaptureAnnotationWorkbench.tsx:275–300`, `1520–1608`; no additional finding in the inspected path.
- The initial sandboxed Edge launch failed before browser interaction. The approved local launch then completed the focused reproduction successfully with exit code 0. No broad suite was rerun. Existing build warnings and recorded baseline failures were treated as existing debt, per review instructions.
### Assessment
**Task quality:** Needs fixes
**Reasoning:** The latch and already-issued request handling are well separated, and the browser coverage addresses most required boundaries. The owned-first asynchronous entry path still commits a task switch during suspension and must be gated before this task can be approved.

View File

@ -0,0 +1,34 @@
### Spec Compliance
- ✅ Scoped fix compliant: `tests/browser/annotation-workspace-full-workflow.mjs:71` now keeps the complete workflow inside the per-task loop; line 78 correctly requires the slice phase. The eight deleted obsolete lines remove the premature closing brace, wrong annotate assertions, invalid outer declarations and undeclared `resuming` guard. This closes the Important finding from the initial Task 08 review.
- ✅ The acceptance correction distinguishes the historical recovered chain from the final uninterrupted run and preserves earlier failures rather than rewriting them as success (`docs/verification/annotation-workspace-layout/acceptance-results.md:117`). No product behavior or unrelated scope was added.
- ⚠️ This is a scoped re-review of `89c9e168..ab89c63a`; earlier 01–07 evidence, full-suite baseline debt, native OS IME/scaling and acceptance-media coverage limits remain as recorded in the original review. Whole-feature review remains separate.
### Strengths
- The change removes only the broken temporary branch structure and restores existing generation/save/handoff/continuation logic under the same task variables (`tests/browser/annotation-workspace-full-workflow.mjs:71`). No replacement mock or weaker assertion was introduced.
- The preserved RED actually fails on `slice` versus `annotate` at the original line 78 (`local-data/verification/annotation-workspace-layout/task08-runner-review-red.log`). The fresh fifth-project GREEN records all three real task chains, stop/resume and empty-pool termination, ending with exact totals of three generations, saves and handoffs, two successful claims and one empty claim (`local-data/verification/annotation-workspace-layout/task08-runner-review-green5.log:1`).
- A read-only SHA-256 check of the delivered runner returned `a20779d7c1187106f1f01e441574094bc19bc507964e82f9a78f5b8a14fa9f56`, matching the before/after final-run hash reported in the correction. The final-file evidence therefore addresses the original gap between historical execution and the delivered file.
- The updated report narrows the earlier unsupported reload attribution and explains the observed Markdown-triggered reload and repository-write freeze used for final verification (`docs/verification/annotation-workspace-layout/acceptance-results.md:121`).
### Issues
#### Critical (Must Fix)
- None.
#### Important (Should Fix)
- None remaining in this scoped correction; the initial workflow-runner finding is closed.
#### Minor (Nice to Have)
- No new minor findings. Previously recorded build/format warning debt and coverage limitations remain unchanged.
### Assessment
**Task quality:** Approved.
**Reasoning:** The eight-line removal repairs both phase handling and lexical scope, and the exact delivered runner now has complete fresh browser/API execution evidence. Task 08's original blocking finding is resolved; the earlier declared limitations still apply.
- Review checks: read the fix package and appended task report, inspected the existing RED/GREEN logs, and hashed the delivered runner. No suites or browser workflows rerun, no subagents, and no fixture, code, index or HEAD mutations; only this requested review report was written.

View File

@ -0,0 +1,170 @@
# Task 08 — combined acceptance and operator guide
Status: implementation/acceptance complete within explicit coverage limits; awaiting root independent review. Ticket08 remains claimed. Base `c0d344bf`; current authorized branch `feature/local-dev-toolkit`. No child agents, CUA, push, merge, deployment, company service or BOS writes.
## Delivered
- Rewrote `docs/annotation-workspace-layout.md` as an approachable Chinese operator manual, covering Start/owned-first/random work, slicing/generation, explicit description save, submission confirmation, stop/resume/empty, dictionary/list/long text, single-camera inspection, shortcuts/history boundaries, layout fallback, minimum CSS size, in-flight reconciliation, refresh and single active page. Explicitly distinguishes send-to-QA from QA approval and code restoration from database/media/browser drafts.
- Added `docs/verification/annotation-workspace-layout/acceptance-results.md` with all30 stories, L01–L16, exact five-size effective pictures, full12-case actual browser zoom matrix, fresh/preceding evidence distinction, authoritative workflow version chain, failures, limitations, reproduction entry points and recovery checks.
- Added four durable browser runners: full-workflow, zoom-matrix, errors-and-legacy, review-isolation. Existing dictionaries runner gained `ANNOTATION_DICTIONARY_READ_ONLY=1` to reuse five-size real behavior without overwriting saved main descriptions.
- Found/fixed an actual error-control overlap: moved focused-only slicing errors into the fixed stage editor. Original layout error placement unchanged. Consolidated duplicate description-action CSS using its final effective values; formatted previously touched Layout.tsx only.
- Retired only the obsolete portal source-text test (and unused dictionary source import) after actual original portal pointer/search/keyboard/Esc/Space and focused inline tests passed. Did not mirror implementation with a revised regex or remove baseline failures.
- Polished stale spec/index implementation wording and compressed Answers for01–06; no scope change, no self-resolution of08.
Root owns `full-suite-results.md` and `execution-record.md`; neither staged by this task. Root also owns archival of reports/reviews/ledger after final review.
## Browser/API TDD and diagnosis
New regression `annotation-workspace-errors-and-legacy.mjs` first ran against pre-fix product. It produced real invalid-split error by C at frame0. At1280×640,1366×680,1440×800 primary controls were reachable; at1920×980 actual elementFromPoint hit-test of “上一帧” failed. Repeated after200ms resize settling and identified the exact control, disproving a timing hypothesis. Screenshot `task08-error-legacy-failure.png` visibly showed the alert over transport.
Root cause: the existing floating alert used absolute bottom128px against the large workbench, while transport and timeline geometry change with available space. It drifted over the transport row. Fix: preserve old error node for original layout; render the same error in the focused slicing editor, bounded to56px internal overflow. It no longer floats above media/controls or introduces a new global grid row.
GREEN: same real error at all5 sizes, each named button and the alert itself inside viewport and hit-test reachable; no document scroll. Added both real confirmation dialogs at all5 sizes, checking confirm and return buttons without sending operations. Final screenshot `task08-error-fixed.png` visually inspected: alert readable below name area, no transport overlay. Existing camera and viewport browser regression passed after product fix; main fixture v68 unchanged.
Error RED logs `task08-errors-legacy.log`; GREEN `task08-errors-legacy-green.log`, final stronger `task08-errors-legacy-final.log`. No source-string/internal-state acceptance or new pure seam. No product behavior was changed without this browser RED.
## Fresh full workflow evidence and consumed fixtures
The full runner submits only authorized local fixtures with names starting workspace-full-workflow. It refuses a consumed initial fixture. It uses real UI slices, dictionary labels, actual generation, description fill, original/focused draft switches, real save/API readback, two-step combined submission and authoritative stage completion. Each generation response was processing before completed, with4slices,412s coverage, zero gaps/overlaps. Empty ranges remained lawful semantic blanks. No QA approval request is made.
Final evidence chain is project-workspace-full-workflow3:
| Task suffix | Initial | Generated | Saved | Handoff | Result |
| --- | --- | --- | --- | --- | --- |
| slice-primary | v1 owned slice | v5,4 slices | v6 | v8 | quality_review; then one random claim |
| pool-02 | v2 after random claim | v6,4 slices | v7 | v9 | quality_review; stop-after-current prevents next claim |
| pool-01 | v2 after explicit resume/random claim | v6,4 slices | v7 | v9 | quality_review; one empty claim then stopped |
Start Work used original task-pool entry, same browser tab, existing owned primary, zero initial random claims. Three real generated/saved/submitted tasks overall;2 successful random claims +1 empty result. Stop-after-current remained selected through original/focused switches and no claim occurred during3.5s after handoff. Only explicit Resume claimed pool-01. Empty result had no retry during an additional2.5s. Continue runner exact request totals2 generation/2 save/2 handoff/1 successful claim/1 empty; prefix main run supplies first generation/save/handoff and successful claim. Distinct request IDs for distinct claims.
Evidence is a complete authoritative chain, not a claim that one final script run exited0 from first to last step. Harness corrections and preserved earlier effects:
1. `workspace-full-workflow`: primary generated4slices v5/savedv6. Runner mistakenly expected dialog name“确认送检” instead of actual“确认提交质检”; timeout before any submission. Corrected harness explicitly resumed saved primary and sent it to quality_reviewv8. Auto-claimed pool-02v2; rapid task-transition/frame readiness led to no extra cut, so real generation preserved existing3slices and reachedv6. Expected4 assertion failed. Pool-01 remains unclaimed. No reset.
2. `workspace-full-workflow2`: primary generated4slicesv5. A concurrent formatter had changed Layout.tsx and the runner timed out while explicit recovery was shown. At that point the specific reload trigger was not instrumented, so the original direct attribution to formatter/HMR was unproven; the controlled Markdown-triggered full-reload diagnosis below establishes the actual class of interference. Server generation preserved. Source editing was frozen for next attempt. No claimed pool tasks. This is development-harness interference, not a product RED.
3. `workspace-full-workflow3`: complete primary v1→5→6→8 and random pool-02v2. URL had advanced before old annotation workbench was replaced; runner counted0 slice delete buttons before slice render and then saw4 after C. Assert4!=1 failed before next-task API mutation. Updated harness to await actual slice-stage controls before media/count. Explicitly recovered already-owned pool-02v2 and completed stop/resume/third/empty, exit0. Unsaved second cut from failed page was discarded normally; no backend reset/duplicate claim.
Runners never reset source fixtures. The final script retains a guarded `WORKFLOW_START_INDEX=1 WORKFLOW_OWNED_TASK=pool-02` explicit-recovery option for this kind of evidence continuation; initial expected server version2 ensures accidental reruns fail. The attempted final cleanup of the temporary annotate-resume branch was later found to have broken the committed runner; syntax checking alone missed it. The independent-review correction and exact-final-file runtime evidence below supersede that delivery claim. Do not rerun the consumed full-workflow3 project.
Exact commands/logs (all below paths under ignored local-data/verification/annotation-workspace-layout):
```sh
node tests/browser/annotation-workspace-full-workflow.mjs > task08-workflow.log 2>&1
# exit1: dialog-name harness error after genuine generate/save.
WORKFLOW_RESUME=1 node tests/browser/annotation-workspace-full-workflow.mjs > task08-workflow-resume.log 2>&1
# historical harness version; exit1 after actual handoff + next generation3 vs4.
WORKFLOW_PREFIX=workspace-full-workflow2 WORKFLOW_LABEL='整体第二轮验收切片 001' node tests/browser/annotation-workspace-full-workflow.mjs > task08-workflow2.log 2>&1
# exit1: generation persisted, source-edit HMR required explicit recovery.
WORKFLOW_PREFIX=workspace-full-workflow3 WORKFLOW_LABEL='整体第三轮验收切片 001' node tests/browser/annotation-workspace-full-workflow.mjs > task08-workflow3.log 2>&1
# exit1 after full primary+nextclaim PASS; next-phase readiness assertion.
WORKFLOW_PREFIX=workspace-full-workflow3 WORKFLOW_LABEL='整体第三轮验收切片 001' WORKFLOW_START_INDEX=1 WORKFLOW_OWNED_TASK=pool-02 node tests/browser/annotation-workspace-full-workflow.mjs > task08-workflow3-continue.log 2>&1
# exit0; second/third/stop/resume/empty/exactrequest totals PASS.
```
## Non-consuming browser checks
```sh
node tests/browser/annotation-workspace-zoom-matrix.mjs
# exit0; task08-zoom.log + task08-zoom-matrix.json
ANNOTATION_DICTIONARY_READ_ONLY=1 node tests/browser/annotation-workspace-dictionaries.mjs
# exit0; task08-dictionaries.log
node tests/browser/annotation-workspace-errors-and-legacy.mjs
# exit0; task08-errors-legacy-final.log
node tests/browser/annotation-workspace-camera-inspection.mjs
# exit0; task08-camera.log
node tests/browser/annotation-workspace-viewport.mjs
# exit0; task08-viewport.log; main v68 unchanged
node tests/browser/annotation-workspace-review-isolation.mjs
# exit0 with explicit acceptance-media COVERAGE LIMIT; task08-review-isolation-layout.log
```
Zoom: Edge isolated persistent profiles, viewport:null, actual --window-size four sizes, chrome.settingsPrivate set/getDefaultZoom100/125/150%, actual inner dimensions/DPR recorded.12 combinations, exact values in tracked acceptance report and raw JSON. Normal controls/no-scroll pass above1280×640; smaller configurations show pause with disabled Continue. Returning100% keeps latch until explicit Continue, retains draft and never writes. Restored100% and closed every profile. Not physical monitor/OS scaling coverage.
Five-size dict: both actual inline panels, long entry, search/paging/keyboard, focus escape via Tab/outside pointer, oneEsc +Space, old popup; no writes. Only focused long-text save section is skipped in READ_ONLY mode; approved05 evidence supplies real longtext save/undo/redo, native cut/undo and synthetic composition ownership.
Camera final run: identical exact effective sizes to06, both stages, progress overlap0, original hidden-camera preference recovery, real frames/drafts/history, no writes.16:9/9:16 are clearly synthetic source replacements in separate browser context.
Viewport final run: width/height boundaries, retained draft/stopped intent, explicitContinue, original fallback latch, native confirmation top-layer restoration, dirty leave protection, Space focus, competing tab session restriction. No writes.
## Prior evidence intentionally reused
Approved01 report supplies same-browser real-account preference isolation, refresh explicit recovery, original entry, delayed-save original/focused switch with exact save/undo/redo public versions and cross-slice draft/history preservation. Approved04 supplies13slice/13description paging and saved boundary/history behavior. Approved05 supplies long text/native cut/undo and deliberately synthetic IME routing. Approved07 (cb74444e + independently approved972721e) supplies actual save/generation/handoff/claim delays,403/unknown hard restrictions, same-ID unknown claim replay, and owned-pool/detail late-result suspension guards.
Unknown3 claim: two transport attempts with identical requestId `d696be07aedb165314c3fb06f47d2c80`, same taskv2 and remainingpool1. Unknown generation used actual successful server effect then dropped response and real detail responses, remained blocked until explicit reconcile. No fresh unknown task needed or consumed in08. Exact preceding commands/versions remain in task-7-report.md; not represented as fresh08 runs.
## Reviewer isolation and specific limitation
Root prepared new localhost-only reviewer users with correct global roles and copied project memberships; existing snapshot reviewer logins globally annotator were unsuitable and redirected. Root created independent owned copies without changing originals: QC human-item-workspace-layout-review-qc / annotation-task-workspace-layout-review-qc; acceptance human-item-workspace-layout-review-acceptance. Auth files ignored qc-reviewer-auth.json and acceptance-reviewer-auth.json. No role change, claim or decision by08 browser.
QC: real paired video ready, platform chrome and normal reviewer controls/scroll access at1280×640 and1920×980; no focused toggle/inspection/pause/compact stage. Version1 unchanged,0POST.
Acceptance: actual FinalAcceptanceWorkspace content, opinion panel, chrome and original scrolling layout isolated, v1 unchanged,0POST. Strict media-readiness attempt timed out60s (task08-review-isolation-final.log and task08-review-diagnostic.log; exit1). Actual GET evidence members for artifact run-d635f13bd89fa9fd50c0665181d884ae-fanyu-v56-human-materialized-revision-000003 returned200/recognized:true/FILESYSTEM with items:[],total:0,contentAvailableCount:0. Source snapshot has no local artifact members. Final runner checks this API explicitly and prints COVERAGE LIMIT while testing panel/layout; it never claims loaded acceptance media, posture/file completeness or final approval. No payload rewrite/synthetic substitute. Screenshots visually inspected; QA shows paired media, acceptance panel is retained with unavailable media.
## Type/build/scoped checks and recovery
Fresh command output was captured directly by exec tool (not redirected), exit0:
```sh
npm run lint:web && npm run build:web
node --test tests/frontend/capture-responsive-layout.test.mjs tests/frontend/annotation-shortcuts.test.mjs tests/frontend/annotation-history.test.mjs tests/frontend/annotation-authoritative-mutation.test.mjs tests/frontend/annotation-random-claim.test.mjs tests/frontend/annotation-submission-state.test.mjs tests/frontend/media-seek-queue.test.mjs
# 54 tests,54 pass,0fail; affected capture-responsive-layout has3remaining tests.
```
Web build succeeded after final production error fix, existing module.register and >500kB warnings only. No later production behavior edit (only docs/test strengthening). Final scoped Prettier checks all9 changed production/test paths and `node --check tests/browser/annotation-workspace-full-workflow.mjs`, `git diff --check` exit0. Before commit exact staged diff check also required.
No full suite rerun: root's actual full frontend578/568/10 retained,9baseline +1obsolete source test retired after behavior verification. Full Go/Python/Rust/etc results remain root-owned full-suite-results.md. Layout.tsx preexisting warning removed,52 untouched warning files left untouched; no false fullformat pass.
`git rev-parse 'codex/annotation-keyboard-baseline-20260918^{}'` again returned54a35afad95d77b1f23c955f9c4671cae83a664d. `git bundle verify local-data/backups/annotation-keyboard-baseline-20260918.bundle` exit0, okay,complete history,sha1. Backup not overwritten. Source-only recovery, no database/media/browser restore.
## Remaining limits and reviewer concerns
- Native OS Chinese IME not tested; CUA explicitly prohibited for this task after prior tool stall. Synthetic composition and actual native text cut/undo are separate evidence.
-12 zoom combinations are browser-engine window tests, not multiple monitors, OS scale combinations, all browsers or company deployment validation.
- Acceptance loaded media unavailable due existing FILESYSTEM snapshot0members; panel/layout isolation passed only. No real QA/acceptance approval performed.
- Existing broad-suite debt remains explicit. All previous01–07 dependencies approved;08/final whole-feature reviews pending.
- Full chain includes documented harness recovery, not a misleading all-green single-run claim; completed fixtures and failed attempts preserved.
## Commit
Local commit `89c9e168` — `✅ Complete focused workspace acceptance and operator guide`. Exactly19 task paths staged; `git diff --cached --check` passed. Root full-suite-results.md and untracked execution-record.md remained outside the commit. Report itself is orchestration evidence under ignored .superpowers, for root archival. Ticket08 stays claimed pending independent review. No push.
## Independent-review follow-up — final runner scope and exact-file execution
Read `.superpowers/sdd/annotation-workspace-layout-plan/task-8-review.md`; accepted its single Important finding. My last text-based cleanup before89c9e168 matched the wrong phase assertion inside a temporary resume block. It left annotate assertions in the slicing loop, closed the loop early and referenced before/resuming/task/index out of scope. The preceding historical runs were real, but did not validate that final edited file. This was a delivered test regression, and the earlier syntax-only completion claim was insufficient.
**Safe runtime RED:** exact committed runner against fresh project-workspace-full-workflow4, owned primaryv1, using `WORKFLOW_PREFIX=workspace-full-workflow4 WORKFLOW_LABEL='整体第四轮验收切片 001' node tests/browser/annotation-workspace-full-workflow.mjs`. Exit1 at phase assertion actualslice/expectedannotate before any generation/save/claim. Public GET verifiedv1/slice unchanged. Raw `task08-runner-review-red.log`; preserved original runner copy `task08-full-workflow-broken-89c9e168.mjs` is ignored local evidence.
**Narrow correction:** remove only8obsolete lines (wrong annotate assertion, old closing brace, invalid outer declarations and resuming guard), restoring generation/save/submission/continuation inside the existing for loop. All variables remain scoped to each task. No product code change or new seam. Full corrected runner manually read. Prettier and parse/diff checks pass.
**Undefined-name validation:** TypeScript createProgram with allowJs/checkJs/noEmit/skipLibCheck, ES2022/NodeNext; semantic diagnostic codes2304/2552 restricted to the runner. Preserved broken runner yields before/resuming/task/index unresolved; final candidate yields0 unresolved names (both observations recorded in tool output). This is a targeted undefined-identifier check, not a claim of full JavaScript typecheck.
**Investigated interruption rather than another assumption:** fixed exactcandidate full4 passed the original RED and genuinely generated4slices to primaryv4/annotate, then180s UI wait failed after visible session recovery. No product source edit or server restart happened. Root had only written Markdown archive/ledger documents. Did not invent a product failure or reset the fixture.
A temporary ignored diagnostic browser explicitly continued this generated full4 task without writes, recording actual Vite WebSocket messages, public navigation/page lifecycle, console and visible recovery region. Appending a harmless temporary comment ONLY to tracked acceptance-results.md caused:
- 02:00:29.973Z Vite socket `{"type":"full-reload"}`.
- 02:00:29.995Z beforeunload,02:00:30.006Z pagehide, then navigation and pageshow.
- Visible recovery changes0→1.
- Restoring identical document bytes caused another full-reload and lifecycle sequence at02:00:33.978Z.
No business POST occurred; diagnostic exit0; file bytes restored. Log `task08-session-diagnostic.log`. This proves ordinary Markdown writes also refresh this Vite instance, and corrects the earlier unsupported attribution to one particular formatted product file. Fourth project remains primaryv4 generated, ready pool tasks untouched. Root verified unchanged API/Vite process IDs.
**Final exact-file GREEN:** root prepared project-workspace-full-workflow5 and coordinated a freeze of ALL repository writes (product, tests, docs, archive, ledger and commits). Only ignored logs/screenshots written during execution. Command:
```sh
WORKFLOW_PREFIX=workspace-full-workflow5 WORKFLOW_LABEL='整体第五轮验收切片 001' node tests/browser/annotation-workspace-full-workflow.mjs > local-data/verification/annotation-workspace-layout/task08-runner-review-green5.log 2>&1
```
Exit0; exact same SHA-256 before/after `a20779d7c1187106f1f01e441574094bc19bc507964e82f9a78f5b8a14fa9f56`. No source change after successful run. Full uninterrupted final-candidate results:
- Start Work same tab, owned-first and no claim.
- Primary v1→5 actual4slicegeneration→6save→8quality_review.
- Automatic claim pool-02v2→6generation→7save→9quality_review.
- Stop-after-current retained through both layouts, no next claim during3.5s; explicit Resume claims pool-01v2.
- Pool-01v2→6generation→7save→9quality_review.
- One empty pool request, no loop during additional2.5s.
- Exact total3generation/3save/3handoff/2successfulclaims/1empty; distinct receipts; API version/persistence checks pass.
The tracked acceptance report now distinguishes the historical recovered chain from this final complete run. No broad suites or product type/build rerun was needed for the8line runner-only correction; prior product checks unchanged. Final scoped formatting, syntax, undefined-name check and diff checks pass. Existing nativeIME/OSscale/acceptance-media limits remain unchanged.
Follow-up commit `ab89c63a` — `✅ Repair final workflow runner and verify complete execution`. Only runner and own acceptance report staged (2 files); staged diff check passed. Root documentation/archive remained excluded. Ticket stays claimed for independent re-review; no push.

View File

@ -0,0 +1,35 @@
### Spec Compliance
- ❌ Issues found: the delivered durable full-workflow runner is broken after the reported harness cleanup. `tests/browser/annotation-workspace-full-workflow.mjs:78` expects `annotate` immediately after awaiting slice-stage controls, and lines 82–85 close the task loop before using its scoped variables. This prevents reproducing the required combined acceptance from the committed runner.
- ✅ The operator guide covers entry/exit, explicit save and submission, stop/resume, dictionary and long content, temporary camera enlargement, legacy fallback, small-window recovery, and code/data recovery boundaries (`docs/annotation-workspace-layout.md:1`). The acceptance report maps L01–L16 and all 30 stories, separates fresh from preceding evidence, and preserves the stated coverage limits (`docs/verification/annotation-workspace-layout/acceptance-results.md:1`).
- ⚠️ Cannot verify from this diff alone: preceding 01–07 acceptance, full-suite baseline comparisons, and tag/bundle integrity. These are explicitly attributed to earlier/controller evidence, not newly demonstrated by this task diff. Native OS IME/physical OS scaling and loaded acceptance media remain declared limitations, not passing claims. The controller should retain these qualifications in final delivery.
### Strengths
- The error fix retains the original-layout alert and places the focused alert inside the existing slice editor; its bounded internal overflow avoids a floating overlay on transport controls (`src/components/CaptureAnnotationWorkbench.tsx:2242`, `src/components/CaptureAnnotationWorkbench.tsx:2481`, `src/styles/capture-workspace.css:588`). A focused context check confirmed the relocated alert is under the slice phase but outside `selectedSlice`, so absence of a selection does not silently suppress it.
- The error regression produces an actual invalid cut and checks viewport containment plus `elementFromPoint` reachability; it also exercises the legacy dictionary through real pointer, keyboard, Escape and Space interactions (`tests/browser/annotation-workspace-errors-and-legacy.mjs:24`). This is meaningful replacement coverage for the retired portal source-string assertion.
- The zoom runner uses isolated persistent profiles, `viewport: null`, browser settings zoom, measured CSS dimensions and pause/resume assertions, rather than CSS zoom or simulated DPR (`tests/browser/annotation-workspace-zoom-matrix.mjs:12`).
- Reviewer isolation uses actual reviewer pages, checks real QC media, records acceptance evidence unavailability explicitly, and verifies no business writes/version changes (`tests/browser/annotation-workspace-review-isolation.mjs:1`). It does not fabricate acceptance media to obtain a pass.
- Historical continuation evidence is real and readable: `local-data/verification/annotation-workspace-layout/task08-workflow3-continue.log:1` records both remaining task version chains, stop/resume, empty-pool handling, and exact request totals. The runner defect below does not invalidate those earlier recorded server effects.
### Issues
#### Critical (Must Fix)
- None found in this task diff.
#### Important (Should Fix)
- **Restore the workflow loop after removing the temporary resume branch.** `tests/browser/annotation-workspace-full-workflow.mjs:78` unconditionally asserts phase `annotate` after lines 72–74 wait for slice controls. Fresh and supported continuation fixtures are in `slice`, so both supported paths fail before generation. The closing brace at line 82 then leaves `before`, `task`, and `index` outside their lexical scope; line 83 reads `before`, and line 85 also refers to undeclared `resuming`. Even changing the phase assertion alone cannot repair the runner. Remove the obsolete annotate-resume assertions/conditional and keep generation, save, submission and continuation inside the per-task loop. Validate the exact final runner, including undefined-variable checking and an authorized fresh isolated workflow if needed; do not reset consumed fixtures. `node --check` only establishes parseability, so the report's syntax-check claim (`task-8-report.md:46`, `task-8-report.md:112`) does not establish this delivery works.
#### Minor (Nice to Have)
- Existing validation noise remains: the report records module.register deprecation and >500 kB bundle warnings, plus 52 untouched formatting warnings (`.superpowers/sdd/annotation-workspace-layout-plan/task-8-report.md:112`, `.superpowers/sdd/annotation-workspace-layout-plan/task-8-report.md:114`). These are explicitly pre-existing and do not block this task, but should stay visible as baseline debt rather than a pristine all-checks claim.
### Assessment
**Task quality:** Needs fixes.
**Reasoning:** The production correction is narrow and the combined evidence is candid about interruptions and environment limits. The committed workflow runner is nevertheless unusable because the final cleanup damaged control flow and scope; fix that delivery regression before approving Task 08.
- Review checks: read the 1,555-line review package once in sequential chunks; no git, suite, browser, fixture, index or HEAD mutation. One changed-file context read was necessary because the workbench hunk cut off its enclosing phase/selection conditions (`CaptureAnnotationWorkbench.tsx:2360–2505`). One outside-diff focused check addressed the named risk that the final runner differed from the successful execution: the existing continuation log confirms historical success. No subagents dispatched.