feat(app): let the conversation lead the Goal Chat view

[huangruiteng]

Stacked on #5071. The Goal conversation is now the main content of the Goal Chat view. Background machinery no longer stacks up above it.

Delivery brief

• Goal/source: The App conversation RFC (docs/architecture/rfcs/app-conversation-and-async-inbox-v0.md) asks for the conversation to dominate the view, with delegation and returns shown inline.
• Current gap: Opening a busy Goal's Chat showed a column of cards before any message: every scheduled monitor and heartbeat, routine runs, and every draft proposal. Drafts were labelled with raw protocol strings such as monitor.create · ready.
• Observable result:
  • Scheduled checks, heartbeats and routine runs fold into one 后台工作 / Background work line, for example 4 个定时任务已启用 · 1 次执行已结束.
  • Drafts use owner language, for example 新定时检查 · 待你确认 / New scheduled check · Awaiting you. Only the newest draft awaiting the owner stays open; the others fold into 另有 N 个待确认提议.
• Owning boundary: presentation only (channel-timeline.tsx). No change to the action, proposal or schedule contracts.

What stays visible

These stay visible on purpose, and no text-based inference decides that something is safe to hide:

• items that need you;
• waiting, interrupted and failed runs;
• applying, applied, failed and stale proposals;
• the gated (needs-authorization) history, which keeps its existing fold.

Only proposals whose status is ready fold behind the newest one.

Behavior change disclosure

• Schedule rows are no longer in the main timeline by default. They open from the Background work fold, and their drawers and actions are unchanged.
• The proposal row label no longer shows the raw action kind. The kind moves to data-action-kind for tooling and tests. The typed-actions, steward-journey and team-plan scenarios now select by that attribute and assert the owner-facing label.

Entry points

• Frontend: changed.
• CLI and Lark: no change. They do not render the App timeline.

Validation

• tsc --noEmit: passed.
• npm run build, including the packaged chat bundle: passed.
• Personal workspace browser suite, all 17 scenarios: passed in development and in packaged mode.
• typed-actions now asserts that a schedule is hidden until Background work is opened, then drives the schedule drawer actions (run now, pause, edit, stop) from the fold.

Future-facing pass

Applied: kind and status labels come from one i18n map (proposal.kind.*, proposal.status.*) instead of string concatenation. Deferred: delegation and returned results inline in the conversation (async inbox) will be the next slice.

First-screen gate

This changes the Goal Chat first viewport. The owner reviews before/after screenshots before merge; leave the merge to the maintainer.

Maintainer Repair At The Merged Head

Reviewed and merged head 727b9de1470ff192518337748bd2b7441a9b90b6 on base 60e1d030a725e212e49a889324a016bb8c052053, merged into codex/app-honest-work-surface as e11bc8d146035ce9001b907dd2d51112b8190ced with an identical tree.

The review's P2 is fixed the way it asked. Recency now comes from the stored record instead of the array position: the typed adapter carries created_at/updated_at, a focused proposal-recency.ts owns the comparison (mirroring ChatActionStore.list's (updated_at, proposal_id) descending order, ties included), ChannelTimeline selects the newest ready draft with it, and dedupeProposals keeps the newest record for a subject instead of the last one. The fold markup, wording, count, gated/applied visibility and drawer keyboard path are unchanged.

The proof had to be made real first: the browser fixture answered /api/actions in insertion order, which is why seventeen scenarios passed while the first screen offered the oldest draft. It now serves the store's actual order, and a new newest-draft scenario covers a restore, a reload and re-entry, a deliberately reversed list, and a draft created in this session. Reverting the fold change makes that scenario fail with the newest row folded, so the assertion is not vacuous.

First-Screen Presentation

Approved by the owner before this head was pushed, per the repository first-screen gate. Before → after at the same viewport, from the same two stored drafts:

• Before: Older pending draft: 调整发布前的验收清单 led the first screen and the newest draft was folded.
• After: Newest pending draft: 确认候选版本的回滚条件 leads, with 另有 1 个待确认提议 folding the older one.

Desktop (1512) and mobile (390) captures were reviewed in both states.

Validation At The Merged Head

18/18 browser scenarios in development and 18/18 against the packaged chat bundle (build:chat plus chat_bundle.py verify --source); smoke:proposal-recency, smoke:goal-order, smoke:delegation-preflight, theme and contract smokes; tsc --noEmit; loopx canary premerge --from-git-diff with 0 direct and 0 run failures; public boundary scan clean; exact-diff change-quality receipt valid. Remote CI was not consulted (wait_for_ci=false). With a symlinked dependency tree the development conversation-input scenario logs an extra 403 for a font resolved outside the Vite root; it reproduces on the unrepaired head and disappears once the dependency sits inside the worktree, so it is an environment artifact, not a PR regression.

[huangruiteng]

Request changes conclusion (author-owned PR; GitHub blocks formal self-review)

Exact head: 3afdfe4504d468582d0a1e1a3f76e97336629627; base: 60e1d030a725e212e49a889324a016bb8c052053 on codex/app-honest-work-surface.

动机

按 App conversation RFC 与现行 design 标准，Goal Chat 应把用户对话、结果和待处理决定放在主要位置，后台运行及 schedule 可保持安静。此 PR 折叠后台工作并把 action/status 显示为中英文用户用语，方向成立。必须同时保证重载之后主视图仍展示真正的最新草稿。

改动思路

沿用现有 ChannelTimeline、typed action adapter、drawer 和 i18n owner，仅改变展示投影，不添加动作执行或 schedule 状态写入。保留失败、中断、等待及非 ready 提议的既有入口，后台项可展开后继续管理。新增 ready 提议折叠目前依赖数组位置，与实际存储列表顺序冲突。

具体改动

关键代码讲解

已阅读全部 6 个变更文件，以及 workspaceProposal、dedupeProposals、list restore、schedule producer、action store 排序与抽屉调用。

ChannelTimeline 根据既有 typed status 选择后台项和 ready 草稿，再由原 row 进入 drawer；workspaceProposal 将存储记录转成展示模型，但没有保留用于比较新旧的时间；ChatActionStore.list 则按权威 updated_at 新到旧排序。这三个可达符号之间的顺序假设造成下面的回归，i18n/CSS 本身没有改变动作权限。

[P2] 重载后把最新草稿折叠、留下最旧草稿。 channel-timeline.tsx:165–167 的 slice(0, -1) 假设最后一条是最新提议。然而 ChatActionStore.list 按 updated_at 降序返回，restore 保持该顺序，adapter 也没有提供用于排序的时间。两条不同目标的 ready 草稿按真实 /api/actions 返回的 [newest, oldest] 进入 Goal Chat 时，新草稿进入“另有 1 个待确认提议”，旧草稿留在主视图；只倒置展示顺序就会倒置结果。用户重载或重新进入 Goal 后，会优先看到过时的可确认动作。

最小修复：在现有 typed adapter 保留/使用权威 proposal 时间或等价的稳定新旧关系，显式选出最新 ready 提议；不要让恢复顺序或插入顺序决定“最新”。补充真实 store 的 newest-first 列表恢复、刷新/重新进入、后续新建草稿及顺序倒置的浏览器断言，确保最新仍可见、其余仍可通过键盘展开进入 drawer。

本地验证：Dashboard build/typecheck、开发与打包版各 17 个现有浏览器场景通过；Goal order、delegation preflight、workspace theme、workspace contract smokes、diff 检查及 6 文件 public boundary scan 通过。额外用真实本地 action store 与 HTTP 路由取得两条持久化 preview 的列表，再将相同返回值送入仓库原生浏览器 fixture：base 两条均可见；head 在 newest-first 顺序下隐藏最新；倒序后隐藏最旧。开发版与打包版一致。桌面/移动首屏均观察到旧草稿留在主要区域；键盘展开及 drawer 路径仍有效。该分层验证没有执行 monitor 或改变生产 Goal 状态，也不宣称整个后端动作流程已通过 live 验收。

对主干的风险

语义与 CI 对齐

具体回归发生在已有 typed-action 恢复路径，影响后续访问和累积草稿的用户体验，现有 17 场景缺少这一排序反例，因此正常场景全绿不能覆盖它。默认展示变化已在 PR 中说明；没有 default-off 承诺，没有新增权限、公共协议或第二状态 owner。类型状态的显式匹配方向正确，但排序判断必须回到其权威数据。

未来改动准备已应用于统一 i18n action/status 标签；排序应作为同一展示边界内的修复，不需要新框架。完整异步 App 协作/结果返回仍是既有后续工作。首屏 owner presentation approval 未由这次技术评审认证，仍需按项目原有 gate 保留。

我的整体评价

REQUEST_CHANGES：展示压缩是有价值的范围，但最新草稿恢复顺序已出现可复现的回归。先修复上面的 P2 并加入重载排序反例，再对整个不变 head 复评。没有合并或自合并授权。

English verdict: REQUEST_CHANGES

[huangruiteng]

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

详细中文评审

结论：APPROVE。Exact head：727b9de1470ff192518337748bd2b7441a9b90b6；base：60e1d030a725e212e49a889324a016bb8c052053（codex/app-honest-work-surface）。上一轮评审提出的 P2 已修复并补上真实反例，首屏展示也按项目首屏 gate 取得了 owner 确认。

动机

按 App conversation RFC 与 design 标准，Goal Chat 应把用户对话、结果和待处理决定放在主要位置，后台运行与 schedule 收进可展开的安静区域，action/status 用用户语言展示。上一轮指出：这种“一条待确认草稿领衔、其余折叠”的展示必须保证重载后领衔的是真正最新的草稿。原实现用 readyProposalItems.slice(0, -1) 折叠数组末尾之前的所有草稿，而 ChatActionStore.list 是按 updated_at 倒序返回、restore 原样保留该顺序的，于是最新草稿被折叠、最旧草稿留在主视图；会话内新建草稿则被追加到末尾，同一个切片又给出了相反结果——展示结论实际由数组如何拼装决定。

改动思路

把“哪条更新”从位置假设改回记录本身。适配器把存储的 created_at/updated_at 带进行模型，新增聚焦模块 proposal-recency.ts 负责比较（按 store 自身的 (updated_at, proposal_id) 倒序，连并列时的次序都与真实列表一致），折叠逻辑改为按该规则选出最新一条、其余按同一规则倒序折叠；dedupeProposals 也从“后写覆盖”改为“保留最新记录”，避免同一修复出现两处不同强度。浏览器 fixture 原先按插入顺序返回 /api/actions，与真实 store 不符，正是 17 个场景全绿却测不出该回归的原因，因此改为按 store 顺序返回，并新增 newest-draft 场景覆盖恢复、刷新重进、会话内新建草稿以及故意倒序四种到达顺序。

具体改动

关键代码讲解

1. proposal-recency.ts:14 proposalRecencyKey / :23 compareProposalRecency：以 (updatedAt ?? createdAt ?? "", previewId) 为键，按 store 的字符串比较与并列次序倒序排列；缺时间时退化为确定性顺序而非抛错。
2. channel-timeline.tsx:170 readyByRecency：[...readyProposalItems].sort(compareProposalRecency) 后 slice(1) 即为折叠集合，首个即首屏草稿；折叠项仍按同一规则倒序渲染，foldedProposalIds 继续驱动可见性过滤，applying/applied/failed/gated 行完全不受影响。
3. personal-workspace-page.tsx:78 dedupeProposals：同一 subject 下保留较新记录（compareProposalRecency(proposal, current) < 0），键与顺序不变。
4. personal-workspace-page.tsx:694 workspaceProposal：透传 updatedAt/createdAt；personal-workspace-model.ts 增加两个可选字段。
5. 测试侧：proposal-recency.test.mjs 钉住恢复序、倒序、并列、宿主回调无时间与单条草稿；newest-draft.mjs 通过真实入口断言首屏行、折叠计数与键盘展开进入 drawer；fixture.mjs 改为按 store 排序；team-plan.mjs 的两条草稿原本时间戳相同（真实存储不会如此），现明确区分先后。

对主干的风险

范围只在这块展示投影：不写状态、不改变 apply/cancel/regenerate/defer 的权限与路径，未新增持久化字段或公共协议，折叠的标记、文案与计数语义均未改动，因此没有权限或数据兼容风险。首屏可见变化只有“领衔的草稿换成最新那条”，已由 owner 通过 before/after 首屏预览确认。残余风险有两处：宿主直接返回且完全不含时间的 preview 会退化为确定性而非时间顺序（存储路径永远带时间，单测已钉住该回退）；本轮未查询远端 CI（goal 策略 wait_for_ci=false）。验证覆盖：开发版与打包版各 18/18 浏览器场景通过（含 review 要求的 typed-actions/team-plan/steward-journey），smoke:proposal-recency、smoke:goal-order、smoke:delegation-preflight、theme 与 contract smoke、tsc --noEmit、build:chat + chat_bundle.py verify --source 均通过，loopx canary premerge --from-git-diff 直接检查与选中检查 0 失败，public boundary scan 干净，exact-diff quality receipt 有效。另附一个环境说明：本机用软链依赖树跑开发版时 conversation-input 会因 Vite 拒绝服务 Vite 根之外的字体而多出一条 403，把该依赖放进工作树后同场景通过，且在未修复的 head 上同样复现，与本次改动无关。

我的整体评价

APPROVE。problem_context 的 long_horizon 与 user_experience 均为 improved：重复进入 Goal 时不再随机退化到某一条历史草稿，首屏给出的是用户最近动过的那条，折叠计数与键盘可达性保持原样。observable_semantics 为 intentional_change_validated，并单独记录了本轮相对上一 head 的真实差异（领衔草稿改变、折叠形状不变）。repository_reuse 为 reused：复用既有 typed adapter、既有折叠标记与 i18n key，并把顺序规则收口到一处；scope_fit、code_volume、change_proportionality 通过（约 57 行生产与配置、独立 33 行规则模块、48 行单测与约 164 行浏览器 fixture/场景）；authority_semantics 为 not_applicable，default_off_isolation 不适用（未新增可选能力）。合并授权来自维护者，本评审只说明该 exact head 的批准条件已经满足。

English verdict: APPROVE - exact head 727b9de on base 60e1d03. The reviewed P2 is repaired the way the review asked: recency now comes from the stored record instead of the array position, so a restore, a reload, a session-created draft and a deliberately reversed list all lead with the newest unconfirmed draft. The proof is real: the browser fixture now serves /api/actions in the store's actual (updated_at, proposal_id) newest-first order -- which is why 17 scenarios could pass while this was broken -- and a new newest-draft scenario fails on the reviewed head with the newest row folded, then passes here. dedupeProposals was made order-independent by the same rule, one fixture that gave two stored drafts identical timestamps was corrected, and the fold markup, wording, counts, gated/applied visibility and drawer keyboard path are unchanged. Validation: 18/18 scenarios in development and 18/18 packaged against the built bundle, tsc --noEmit, the recency unit smoke, the goal-order/delegation/theme/contract smokes, canary premerge with 0 direct and 0 run failures, a clean public boundary scan and a valid exact-diff quality receipt. The first-viewport presentation carries the owner's approval. Unrelated and non-blocking: with a symlinked dependency tree the development conversation-input scenario logs an extra 403 for a font outside the Vite root, identically on the unrepaired head. Merge authority comes from the maintainer, not from this review.

[huangruiteng]

Merge decision (maintainer)

Reviewed exact head 727b9de1470ff192518337748bd2b7441a9b90b6 on base 60e1d030a725e212e49a889324a016bb8c052053. Published review: #pullrequestreview-5326610755 — five-block Chinese review, English verdict: APPROVE, with the author-owned titled fallback because GitHub blocks formal self-approval on this PR. Merged as e11bc8d146035ce9001b907dd2d51112b8190ced; the merged tree is identical to the reviewed head tree.

Changed surfaces: channel-timeline.tsx (which pending draft leads the Goal Chat first screen), the view model and adapter, a new proposal-recency.ts plus its unit smoke, the browser fixture's list ordering, a new newest-draft scenario, one fixture timestamp, and the dashboard smoke scripts.

Checks named in the review: 18/18 scenarios in development and 18/18 packaged, tsc --noEmit, smoke:proposal-recency, smoke:goal-order, smoke:delegation-preflight, theme and contract smokes, build:chat with chat_bundle.py verify --source, loopx canary premerge --from-git-diff with 0 direct and 0 run failures, public boundary scan clean, and a valid exact-diff change-quality receipt for base origin/codex/app-honest-work-surface.

Failures and skips: remote CI was not consulted (wait_for_ci=false for this goal). The development conversation-input scenario logs an extra 403 for @fontsource-variable/geist when the dependency tree is symlinked outside the Vite root; it fails identically on the unrepaired head and passes once the dependency is inside the worktree, so it is an environment artifact. No manual holds.

Why the coverage is enough: the diff changes which draft is rendered inline and the rule that decides it, with a unit smoke for the rule, a browser scenario that exercises four arrival orders through the real surface, and a mutation check proving the new assertions fail on the reviewed head. The unrelated proposal behaviors (failed recovery rows, gated rows, team-plan cards, monitor controls, drawer keyboard path) are covered by the existing scenarios, which still pass.

First-screen gate: the owner approved the before/after first viewport before this head was pushed. Merge authority came from the maintainer, not from the review.