feat(workspace): focus first screens and inspect team work in Goal chat

[huangruiteng]

Goal and result

Related to #4574, roadmap R2/R3 and #4731. Owners can inspect durable member work and actual task/runtime prerequisites from the original Goal conversation. The workspace puts decisions, conversation and results first, with relevant controls one step away.

Changes

• Goal conversation: a compact LoopX bar opens team/settings dialogs. Escape restores focus and the conversation position. Pausing explicitly names the coordinator; dispatched members continue. Rejected, unavailable, uncertain and recovery states stay visible.
• Home: populated active lanes fit the viewport; empty groups and duplicate summaries are omitted. Stopped Goals remain in the sidebar archive and no longer occupy the right-side home. The digest appears for new run results.
• Overview, tasks and outputs: compact usage preserves unknown values; decisions and errors remain direct. Existing Kanban/list behavior stays intact. A rendered reply preview leads directly to the conversation or an editable draft. Output provenance remains in its existing drawer.
• Settings: remove duplicate headings while preserving source, activation consequences and preview/apply. The selected category stays visible on narrow screens.
• Shared inspection: CLI delegation inspect, enabled MCP and new Goal Chat tools reuse the delegation owner. TypeScript classifies actual Turn, runtime and pinned-acceptance facts. Inspection creates no work, calls no model and grants no execution permission. Original native Sessions/tool schemas are preserved.
• Development workflow: design, AGENTS, the PR template and the bilingual presentation RFC require each visible element to provide high-value information OR an essential interaction OR expressive visual presentation, judged across the whole viewport.

The bounded refactor shares execution arguments, verifier pin checks and the execution-summary predicate; moves the existing Turn result projection intact; and reuses Markdown, dialogs, drawers and configuration ownership. No new state store or scheduler.

Before / after

All five comparisons use the same synthetic data and viewport. Left is before the attention redesign; right is the approved final version. No owner-local project data is included.

Home — active work fits; stopped history stays in the sidebar

Goal overview — decisions and results replace empty chrome

Tasks — one readable reply, direct continuation, existing Kanban

Settings — less repetition, same configuration controls

Narrow settings — selected category remains visible

Validation

Reviewed head: 6d3fe3bd1429c3dbfb1c1aeeffccca567270d08a; base: cc8e28d8b58a9b15e928d7e2cee16097172567d1.

• Python Turn/readback/preflight/managed-step suite: 232 passed. The final selector-hardening rerun: 14 passed. TypeScript delegation tests: 5 passed. Ruff and control-plane TypeScript typecheck passed.
• Dashboard production build passed. Eight packaged workspace browser scenarios passed, including direct decisions, unknown usage, delivery errors, provenance, editable replies, mobile categories, dialog focus/scroll return, long Goal-title overflow and existing Kanban behavior.
• Desktop/mobile/Chinese usage browser checks, the fresh Frontstage share-bundle smoke and the public-navigation migration browser smoke passed. The existing large-bundle build warning remains.
• Final local premerge canary passed: 19 selected checks, zero failures, zero warnings and zero manual holds. Public/private boundary scan and git diff --check passed for the complete 64-file diff.
• One initial Frontstage smoke used a stale temporary export and a system Python outside this checkout; rebuilding the export with this worktree's virtual environment passed both the standalone smoke and the final canary. No required local validation was skipped.
• The pre-rebase quality receipt cqr_40149b4f0588220f279f is superseded and is not evidence for this rebased head. No replacement receipt is claimed because this review has no explicit Goal id; the exact-head diff is instead qualified by the validation and canary evidence above.

The future-facing pass was applied at the touched boundary: execution readback now has one focused projection, Goal-owned execution configuration remains authoritative after rebase, and local inspection cannot retarget its project or scan root through host arguments. Larger assignment-receipt integration, provisioning, two-cycle coordinator continuation, remote capacity probes, idle-host wake and Lark parity remain with the existing roadmap owners. This is a complete, independently reviewable local inspection and presentation increment, not completion of the full roadmap. Product/runtime merge remains maintainer-owned.

[huangruiteng]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机

审查 head 62cf816982f988b8883cf760ac19918e99b211aa，基线 74b62ebb55c9f9fefb209f0351593ba8f9c5d32d。本批是 整体路线图 R2/R3 的合理增量：在原 Goal 对话直接检查持久团队工作和成员启动条件。它没有完成整个持续团队目标；计划分配回执整合、经授权的创建、两轮主协调员持续推进及 Lark 等价仍未验收。

改动思路

最强反对理由是“再造一份 readiness 状态，反而把登记、模型可用、已执行和已验收混为一谈”。当前实现通过复用原 owner 避免这个问题：目录继续使用 #4731 的 requester-scoped journal；检查使用实际 Turn dry-run、所选执行配置和 canonical acceptance；TypeScript 只分类这些事实。没有新的调度器、持久状态或执行授权。未知的通用/云端运行时保持 unknown。

比较了只补文档、从登记推断状态、增建 fleet registry 三种替代方案；前者不能解决原对话的直接检查体验，后两者会重复权威。现有 #4733 的宿主容量配置是相邻能力，并不能替代本次任务级的准入、验收检查；#4376 的独立 workspace 未被引入。

具体改动

• Delegations.inspect 和共享 _execution_arguments 保持任务、workspace、host/model/effort、validator 参数一致。先用真实 parser 拒绝执行开关、缩写执行开关及身份重定向，再运行只读预检；检查前后核对绑定/验收来源。已有 Turn 拒绝会显式返回错误。
• goal_task_validation_files_current 复用抽出的原 verifier pin matcher，读取已绑定校验文件的当前哈希，不执行产物验证。真正的执行完成仍走现有独立验收。
• delegationPreflight 在 TS 边界区分任务准入、验收绑定和三态 runtime availability。launchable 只表示本机前置条件，不代表执行许可、容量预留或正在运行。
• CLI、已启用 MCP、新挂载的 Chat tool 共用检查。旧 native conversation 的 schema 与共享 GUIDANCE 保持；owner HTTP 入口沿用保存的 sender/config，暂停时可查，但不提交模型 Turn。
• GoalTeamWork 在原控制栏下面按需展开：分页查看 accepted/unavailable/recovery，展示产物哈希和实际 profile，切换身份/配置清掉旧观察，异步请求带代次隔离。普通 snapshot polling 不触发验收/预检。源码和随包前端同时更新。
• 现有 Turn result builder 整体移动到 execution_readback.py，全部活动调用点继续共用它；managed_step 从 canonical transaction 导入常量。执行/恢复回执保持原形，只有 dry-run 增加三项紧凑 route 事实。文档、两张合成截图和双语 roadmap 明确边界。

对主干的风险

未发现阻塞性问题。风险主要是把时点检查误解为长期承诺，以及反复验收大量结果的成本；检查保持按需、分页、未知状态和原工作身份，文档明确不提供 fleet snapshot 或远端容量保证。

验证：

• 最终 Turn/readback/preflight/managed-step 套件 118 passed；实际 stdio MCP 检查、执行、断线重连 2 passed；持久目录边界 6 passed。真实 CLI/HTTP 路径使用隔离 File/SQLite，拒绝越权/执行标志，检查不启动模型。
• 验收/委派回归 46 passed，Chat/preflight 27 passed；最终套件补充覆盖 host 拒绝。修改 pinned verifier 的真实 CLI 用例先复现 false-ready，再验证修复。
• 不变性：5 个原 CLI 场景完整输出/退出码与基线一致；禁用 MCP 的完整 schema、旧 enabled schema 和共享 Chat GUIDANCE 一致；相同 File/SQLite 输入的 dry-run 仅增加已声明 compact route。无新持久化迁移。
• 当前 CLI 回读此前真实 DSH/Ark 合成团队的 3 个根请求 + 1 个下层请求，4 份当前 accepted 产物与哈希重新核验；零启动、Goal 仍 active。这不是新一轮真实模型协作。
• 源码及打包版浏览器检查通过；最终完整打包版 8 个场景包含桌面/窄屏、未知 runtime、不可读结果、恢复和分页。浏览器使用合成 API fixture 验证交互，真实后端证据由独立 CLI/HTTP/stdio 测试承担。
• TS typecheck 与 5 个 TS 用例、Ruff、配置范围内 mypy、前端构建和公开边界扫描通过。最终完整 canary 5 个直接检查 + 19 个选中检查通过，无失败/跳过；精确范围质量回执有效。

早期 CLI 输出增长和 executor 模块规模检查确实失败过：减少 route 到消费者所需的三个字段，并把既有 result projection 拆到相邻模块后修复，没有提高预算上限。按 review policy revision 7 的 wait_for_ci=false，没有查询、轮询或等待远端 CI。

我的整体评价

APPROVE，作为可独立使用和回滚的本地检查增量。它让 owner 在原对话判断“原工作在哪里、当前能核验什么、哪里还不能确定”，同时保留多层协调可复用的接口。未来适配继续沿原注册、runtime、计划回执和 host continuation owner 推进，不在管家里增加第二套控制面。该 PR 涉及运行时与产品面，留给维护者合并。

English verdict: APPROVE - head 62cf816. Shared on-demand inspection preserves current authority, unknown runtime and existing execution receipts. Real CLI/HTTP/stdio File/SQLite, packaged UI, base/head parity and final local canary passed. This is a bounded local-inspection increment; provisioning, autonomous two-cycle continuation and Lark remain unqualified. Maintainer merge required.

[huangruiteng]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机

审查 head b8c34e41baf7779dc900831dbb66d0e2d8417255，base 31525ca92a16ee0e06b7e88ba129f94771d11179。当前交付同时回应两项已确认需求：从原 Goal 对话检查持久团队工作；让整个工作区首屏集中于高价值信息、必要交互或有表现力的展示。整体路线图 R2/R3 仍是持续团队目标的依据，本 PR 是其中可独立使用的检查与体验增量。

相对上次审阅，本次新增了跨栏目注意力优化、同数据前后对比和流程规范，并按最新反馈将右侧主页的「已停止」移除，保留侧栏归档入口。首屏展示已获得 owner 确认；本结论重新审查当前完整 diff，不沿用旧 head 的批准。

改动思路

最强反对理由是：压缩界面可能隐藏阻塞、未知状态或必要操作，而新 readiness 层可能重复执行权威。实现采用现有 typed state 和组件：只收起明确的常规活动，保留等待、失败、中断、拒绝、未知用量和不可核验信息；团队与设置一键打开，关闭回到原位置。单独整理某个 feature 不能消除首页、概览、任务与设置的重复信息，因此本次按整个工作区调整，而没有增加新的注意力引擎。

后端继续复用 #4731 的 requester-scoped 工作目录、实际 Turn dry-run、现有执行 profile 与 canonical acceptance。TS 分类观察事实，Python 负责现有 IO/适配；没有新调度器、持久状态、Agent 创建或执行授权。空 runtime availability 不被解释为可用。

具体改动

• Delegations.inspect 与 _execution_arguments 共用实际任务、workspace、host/model/effort、validator 参数。真实 parser 先拒绝执行/缩写执行开关和身份重定向，检查前后核对来源；失败明确返回。
• goal_task_validation_files_current 与正常验收复用原 verifier pin matcher。delegationPreflight 区分任务准入、验收就绪和三态 runtime 可用性；检查不是产物验收，也不是容量承诺。
• CLI、启用的 MCP、新挂载的 Chat tool 共享检查。Owner HTTP 查询暂停中的团队，不提交模型 Turn；原 native Session、旧工具 schema 和共享 GUIDANCE 保持。
• GoalLoopXMode 使用紧凑控制栏和原生 dialog；GoalTeamWork 保持按需分页、身份代次隔离、显式未知/恢复/不可读状态。正常 polling 不触发团队检查；「暂停协调员」说明已派发成员继续运行。
• ManagerHomeBoard 只展示有内容的活跃分组；停止归档只在左侧。执行摘要谓词共享，安静 Session 不丢失。ChannelTimeline 按明确状态折叠常规活动，移除重复 runtime/status 文案和误导性 Live/0–1 进度。
• 概览保留直接决策、异常和未知用量；任务复用 Markdown renderer，提供回复预览、直接进入对话和可编辑草稿，原 Kanban/list 不变；成果沿用来源 drawer。设置减少重复标题，保留来源、影响、preview/apply 和主题选择，窄屏选中栏目自动进入视口。
• 既有 Turn result builder 整体迁入 execution_readback.py；执行回执不变，只有 preview 增加三项 compact route。设计规范、AGENTS、PR 模板、双语展示 RFC 和操作文档同步更新；PR 正文包含五组同数据同视口对比图。

对主干的风险

未发现阻塞问题。主要剩余风险是将时点检查误解为长期可用性，或用浏览器 fixture 代替真实模型协作证明；文档和交付范围明确排除了这两种结论。

验证与证据边界：

• 当前 attention 版本的 8 个打包浏览器场景通过。随后仅涉及 Settings 的修复重新通过源码和打包导航；最后移除主页停止归档后，打包导航、构建/typecheck 和 workspace contract 再次通过。覆盖决策、未知用量、读取失败、来源 drawer、可编辑回复、窄屏设置、dialog 焦点/滚动返回，以及停止 Goal 的侧栏访问。
• 桌面/移动端/中文 usage 验证通过；五组比较图使用相同合成资料与视口，无本机私有项目数据。既有 bundle-size 构建提示仍存在。
• 先前 head 62cf816982f988b8883cf760ac19918e99b211aa 的 118 个 Turn/readback/preflight/managed-step、2 个真实 stdio 重连、6 个目录边界、46 个验收/委派、27 个 Chat/preflight 用例保留原来源记录。已实际比较 Git blobs：后端、测试及依赖未变化，base 更新仅涉及文档 ledger/governance，所以这些结果继续适用，并未重标成新执行。
• 同样保留原来源的完整 CLI 输出、禁用 MCP/旧 enabled schema、共享 GUIDANCE 及 File/SQLite preview 对照。此前 pinned verifier 修改导致 false-ready 的真实 CLI 反例已修复。历史真实 DSH/Ark accepted 产物回读仅证明既有结果可重新核验，不宣称本次启动了新一轮自主团队。
• 最终完整本地 canary：5 个直接检查、19 个选中检查通过，零失败、零跳过；公开边界扫描 61 个现存候选路径通过。质量回执 cqr_40149b4f0588220f279f 对当前 64 路径精确范围验证有效。此前输出和模块预算问题通过缩小投影、抽出既有 owner 修复，没有提高阈值。
• 使用 review policy revision 7、wait_for_ci=false；未查询、轮询或等待远端 CI。没有以 CI 状态替代本地证据。

我的整体评价

APPROVE。当前交付让用户留在原对话处理团队工作，并让活跃目标、决策和结果重新成为首屏主体；跨栏目复用现有权威和组件，后续改动更易定位。计划分配回执整合、授权 provisioning、两轮持续协调验证、远端探测/唤醒及 Lark 等价仍归原路线图后续边界。本 PR 涉及产品与运行时，留给维护者合并。

English verdict: APPROVE - head b8c34e4. Shared on-demand inspection and approved workspace first screens preserve authority, persistent Sessions, unknown/error states and direct actions. Eight packaged scenarios, final affected navigation/build checks, exact quality and full local canary passed; unchanged backend evidence was explicitly requalified for source continuity. No fresh autonomous model-team cycle is claimed. Maintainer merge required.

[huangruiteng]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机

结论先行：exact head 6d3fe3bd1429c3dbfb1c1aeeffccca567270d08a 没有阻塞性 finding。这个 PR 没有声称完成 #4574 的整个 R2/R3 路线，而是交付一个完整且可独立回滚的增量：Goal owner 可以在原 Goal 对话里按需查看当前协调身份的持久委派工作，并基于真实 Turn 路由、Goal 验收 pin 和 managed executor 观测检查成员启动前提；与此同时，个人工作区首屏把决策、结果和直接操作放回主要视觉层级。

不合入的最强理由是 64 个文件同时覆盖运行时、CLI/MCP/HTTP、前端、文档、测试和打包资产，表面较宽。但只做 UI badge 会复制执行真相，只做底层协议又会留下无产品入口的半成品。当前边界复用了已有 DelegationService、Turn preview、Goal acceptance 和 packaged workspace，没有新建状态库或 scheduler，因而比平行状态缓存或 Python 重写准入规则更小、更符合现有 ownership。

改动思路

入口链路是 Goal Chat team dialog / delegation inspect / collaboration MCP → ChatLoopXMode.read_team / DelegationService → TypeScript delegationPreflight 与 inventory projection → 既有 journal、Turn admission、Goal acceptance pins 和 managed executor。TypeScript 继续拥有状态分类，Python 只绑定 Goal-owned execution_config、协调身份并适配入口，React 只呈现结果。

inspect 是派生的 point-in-time projection：它强制 dry-run 且 host_invoked/state_written/quota_spent/scheduler_acknowledged 全为 false；launchable 也不是执行许可，真正 start 会重新核验准入。operations 只分页当前 requester 的现有 journal，unavailable、不完整页和 recovery_required 不会被折叠成“团队空闲”。普通未配置会话不会显示 team trigger，也不会发起 team read，因此 default-off 路径与原行为隔离。

具体改动

• 新增只读的 delegation preflight/inventory，并通过现有 CLI、MCP、Goal Chat HTTP 和打包 UI 暴露；原有 list/start/read/wait/resume、原生 Session/tool schema、Turn receipt 均保持兼容。
• 将 Turn public execution readback 从大 executor.py 抽成单一 execution_payload projection，preview、settlement、recovery 共用同一事实来源。
• Goal Chat 用现有 dialog 展示成员检查、分页 operation、恢复/不可核验状态和 artifact hash；异步 generation token 丢弃过时回包，错误与 refresh 归 UI 所有。
• 首屏删除/折叠重复和低价值 chrome，保留直接决策、未知用量、delivery error、provenance、editable reply、Kanban、移动端 category 与 stopped Goal sidebar archive。
• rebase 后修复了 _execution 四元组调用不一致、缺失 coordinator 的 KeyError 风险、fixture 对 Session-owned 配置的过时假设、浏览器 projection 瞬态，以及 host args 可覆盖 --project/--scan-root 的范围问题。

关键代码讲解

1. loopx/chat_loopx_mode.py:250 ChatLoopXMode.read_team：只接受 inspect/operations 的精确 body，使用本 Goal 的协调身份和配置创建服务；暂停协调员不阻断读取，但不会提交模型 Turn。
2. loopx/control_plane/collaboration/delegation.ts:38 delegationPreflight：按 turn_blocked → acceptance_unavailable → runtime_unavailable → runtime_unverified → launchable 的 typed precedence 分类，并验证 preview 无副作用。
3. loopx/control_plane/turn_driver/execution_readback.py:18 execution_payload：集中构造 preview/execute/recovery 的公开回读；route 只在 preview 暴露，quota、settlement、lane fence、host failure 与 recovery 仍来自原 owner。
4. apps/presentation/dashboard/src/features/personal-workspace/goal-team-work.tsx:7 GoalTeamWork：显式触发 scoped readback，保留分页不是 fleet snapshot、无记录不等于 Goal 完成、不可核验不应盲目重派等语义。

对主干的风险

最强回归场景是：读操作因错误配置或 host 参数检查了另一个 workspace，或者把 stale pin / unknown runtime 误报为可执行，最终诱导重复派工。exact head 通过 Goal-owned config、注册 coordinator guard、selector override 拒绝、typed precedence、零 effect preview 和显式 incomplete/unavailable 状态阻止该路径；启动仍由现有 Turn admission 重新核验。影响面限于本地已配置 Goal owner，没有获得 remote/fleet authority。

验证覆盖完整 diff：Python Turn/readback/preflight/managed-step 232 passed；最终 selector/preflight 14 passed；TypeScript delegation 5/5；Ruff、control-plane typecheck、dashboard production build；8 个 packaged workspace browser 场景；desktop/mobile/Chinese usage、fresh Frontstage share bundle 和 public navigation；git diff --check；premerge canary 19/19，0 failure、0 warning、0 manual hold，public/private boundary clean。第一次 Frontstage smoke 误用了 stale /tmp export 与 system Python；用当前 worktree venv 重建后 standalone 和最终 canary 均通过，没有跳过 required local validation。最终 GitHub CI 在发布本 review 前也已等待到全部非 skipped checks 通过。

语义与 CI 对齐

该变更扩展既有 delegation/Turn/Goal Chat 词汇，不创建新的 actor lifecycle 或 authority 模型；“preflight / inspect / local launch prerequisites”与实现范围一致，文档明确排除执行许可、运行证明、fleet snapshot、Goal 完成与 remote capacity guarantee。repository-native CI 的 DCO、Python shards/stage2c、Node compatibility、dashboard acceptance、real PostgreSQL、Ark adapter、desktop 和 build checks 均在 exact head 上通过；release/presentation 类 job 的 GitHub skipped 状态不是本 PR 的失败或人工 hold。

剩余风险是组合 diff 较宽，以及没有新跑 autonomous remote team / Lark journey；但这两者不属于本 PR 声称的本地只读增量，browser fixture 也没有被当成真实模型团队证据。若 scoped/no-effect invariant 回归，最小修复是恢复 Goal-owned selectors 与 zero-effect preview，然后重跑 preflight 和 packaged browser suites。

我的整体评价

这是一个 justified increment：它在不增加第二权威或持久状态的前提下闭合“真实事实 → 公共入口 → packaged UI → 失败反馈”的本地用户路径；关联 refactor 也已把 execution readback 集中到单一 owner，并加固了 rebase 暴露的范围与配置问题。代码体量与跨层目标相称，既有入口和 disabled path 有实证兼容性，未发现 typed-state、domain-neutrality、behavior-disclosure 或 guidance-vs-obligation 问题。

我对 exact head 给出 APPROVE 结论。由于这是 author-owned PR，GitHub 不允许 formal self-approval，所以本记录以 COMMENTED review 发布；并且仓库政策明确规定 loopx/**、apps/**、CLI/API 与控制面行为变更不能由作者自合并，即使本地 canary 和 exact-head readiness 均通过也必须由另一位 maintainer 合并。

English verdict: APPROVE - exact head 6d3fe3bd1429c3dbfb1c1aeeffccca567270d08a has no blocking findings and is ready for a different maintainer to merge after the repository-required exact-head readiness check; repository policy forbids author self-merge for these product and control-plane surfaces.

[huangruiteng]

Post-merge process audit

This PR was squash-merged at 2026-09-19T15:52:48Z, before the exact-head COMMENTED approval review was published at 2026-09-19T16:05:41Z. That ordering is a process gap for a product/control-plane change even though the code verdict is APPROVE.

• Reviewed PR head: 6d3fe3bd1429c3dbfb1c1aeeffccca567270d08a
• Squash merge commit: 39247551d88f473034ea4dac97356c0a1bc9200b
• Exact-head review: feat(workspace): focus first screens and inspect team work in Goal chat #4736 (review)
• Patch readback: the stable patch id of cc8e28d8..6d3fe3bd exactly matches 8fd32d4b..39247551 (0b9891a2a590c3391f422490b5268cd1caea1d8a), so the reviewed PR delta is the delta that landed; the newer squash parent only contains unrelated concurrent main changes.
• Final CI: all 33 readiness checks are successful after rerunning the initially zero-step cancelled pytest job; no pending or failing checks and no unresolved review threads.

The active protect main ruleset now requires code-owner review, last-push approval, one approval, review-thread resolution, Sign-off, and merge-gate. The repository owner remains an always-bypass actor, so policy still depends on the maintainer not using bypass for normal product/control-plane self-merges. Narrowing that bypass or reserving it for separately audited emergencies is an owner-level repository-setting decision, not a code change to retrofit into this merged PR.