[wanglongan587]

Summary

在普通聊天和工作流对话中展示只读运行诊断：独立的父/子活动接收时间、按来源分组的当前工具、已知待处理请求和明确上报的 provider 重试/压缩状态。历史重放不会伪造最近活动，静默不被称作卡死。

live ACP → bounded observation → run details
child projection → visible transcript; does not renew parent timeout

Refs ora-space/circus#301

ADR 状态

实现 specs/decisions/node/agent-runtime/20261020-child-projection-history-and-diagnostics.md（ora-space/specs#145，head 23de5ef，merge commit e473b2581ec174953db8573d884d5996d0abad49）；该 ADR 已以 approved 合并到 specs main。

本 PR 实现 D3 的运行观察与 D4 的父子存活隔离：

• D3：只展示客户端实际收到的活动与接收时间；工具按本次尝试观察到的来源分为父、子、来源未确认，省略 metadata 的子工具增量仍归子；重试清除旧尝试的观察；缺报告显示未知，不从历史加载时间、静默或父回合状态推断健康。加载流重放未结束回合时仍标未知，不会被说成「本次尝试没有活动」。
• D4：投影来的子文本、思考、工具、plan 与子权限往返都不重置或暂停父 prompt 的既有无活动窗口；父委托工具自身的 in_progress、Timeout/Wait、取消与错误收尾不变。

不在本 PR 范围：history 来源保真由 #695 承接；只读子代理详情树与「委托已返回结果 / 后续未知（投影已停止）」两种标注由 #697 承接。三份 PR 合并并验收后，由落地运行在 specs 开 PR 把该 ADR 改为 implemented。

Evidence

• Before: OpenCode 投影子工具可以暂停父计时；界面无法区分缺报告、静默与等待请求。
• After: 生产 ACP 路由测试验证子文本、工具及省略 metadata 的增量仍可见，子权限往返也不会独立续期，但父 prompt 在既有 45 秒窗口到期；同一路由下父工具权限往返会重新计时。父委托、Wait、取消的原测试保持通过。
• Runtime library：123 passed（rebase 后复跑，含新增的 2 个权限路由用例）；clippy -D warnings 通过。
• Chat：49 passed；app-shell 全量 2140 tests / 225 files passed（其中诊断面板 6 tests、workflow chat 15 tests）；变更文件的 ESLint、Prettier 通过。
• 观察快照的语义：sendMessage 与重试开始一个空的实时快照，因此「本次尝试尚未观察到父回合活动」只用于确有实时观察的尝试；加载流重放未结束回合时保持「未知；历史重放不代表刚发生的活动」，已结束且未观察到活动时去掉「尚未」。加载流本身不记录实时活动。
• 中英文文档与翻译同步。分支已 rebase 到 desktop main be72c252（含 fix(history): preserve projected child attribution across reload #695），crates/history 与 main 无差异；当前 head e988bee9。

Merge Danger

Door: two-way

Blast Radius: runtime

不更改 Timeout/Wait 配置、窗口、重试次数或 JSONL。无法证明被等待的投影子活动不能独立保活；真正等待的父委托应继续上报 in_progress。诊断只读，不增加子代理控制或健康判定。

[ObsisMc]

Reviewed head f8ab7e06 (base 3fa5fc03), against ora-space/specs#145 head 09d5774a.

The runtime half looks right: projected child text/plan/tool/tool-delta and child permission round trips no longer reset or pause the parent window, a projected tool id stays excluded across metadata-free deltas, and the parent delegate's own in_progress path is untouched. The production-route test in prompt_policy_tests.rs exercises exactly the 45 s scenario. ADR status in the description is accurate (proposed → not mergeable yet).

One issue I'd like fixed before this goes further:

1. "Latest live activity" counts child projections, so the panel contradicts the runtime's own timeout.
observeDiagnostics (packages/chat/src/diagnostics.ts:26-41) updates lastActivity for any message/thought/tool/plan update, including ones carrying _meta["opencode/child-session"] (only session_info_update checks for it). Since this PR makes the runtime ignore exactly those events for liveness, an OpenCode v2 run where only a background child is streaming will show "No new activity for 2 seconds" under "Parent turn: Processing" while the parent prompt is about to be cancelled and re-sent at 45 s. That is the "background output read as parent progress" confusion #283's acceptance criteria and ADR D3/D4 want to avoid. Suggested fix: don't let projected updates advance the parent lastActivity (or record them as a separately labelled child activity), and add a store test feeding a projected chunk. "Reported active tools" (run-diagnostics.tsx:26-34) mixes child and parent tools the same way; at minimum mark the child ones.

Smaller notes (non-blocking):

2. store.ts:435-436: the doc comment "Flushes one buffered text batch…" now attaches to let pendingDiagnostics, not flushPendingTextChunk.
3. Tests cover reload clearing and session isolation; ADR acceptance 5 also lists provider switch, failure, cancel and the workflow chat consumer. The clears are there in code (store.ts agent-switch / error / abort / finally) but untested.
4. "Pending requests" can never show anything today because the actor auto-allows every permission before the UI sees it. You already say so in the run issue; maybe say it in the panel too (or in the doc section) so "No known pending requests" isn't read as "nothing is waiting on you".
5. The child-session handling is validated only with synthetic frames. Worth one real OpenCode v2 capture in the run issue confirming the parent task tool's in-progress updates don't carry opencode/child-session (if they did, the delegate pause would be lost).

[wanglongan587]

Addressed in a4e0c848:

• Parent observations and projected-child observations now have separate clocks and labels. Metadata-free child tool deltas retain their known source; retry clears the attempt-scoped ownership cache.
• Tools are separated into parent, child and unconfirmed source groups. Historical/unobserved tools are not assumed to be parent work.
• Fixed the flush comment placement and documented the host's current automatic permission handling in the panel.
• Added store coverage for projected chunks, metadata-free deltas, retry ID reuse, provider switch, failure, cancellation and new prompts; added the actual workflow chat consumer test.

Chat: 48 passed. Diagnostic/workflow UI: 18 passed. Type/lint checks pass. The linked run also contains SHA-256-backed real v2 captures verifying 7 parent delegates / 14 in-progress updates carry no child projection marker. Runtime code is unchanged in this revision.

[ObsisMc]

Re-reviewed head a4e0c848 (delta from f8ab7e06).

All points from the previous review are addressed:

• Parent lastActivity no longer takes projected child updates; child activity has its own clock and label, and the silence counter only measures the parent stream.
• Tool sources follow the runtime rule: a known child tool stays child across metadata-free deltas, the cache is cleared on retry, and unobserved tools show as "unconfirmed" instead of being counted as parent.
• The flush comment is back in place, and the panel and docs now say that auto-handled permissions and unsupported forms are not listed.
• New store tests cover provider switch, failure, cancel and new-prompt clearing. There is also a workflow consumer test.
• The real OpenCode v2.0.32 captures (7 parent delegates, 14 in_progress updates, none with opencode/child-session) settle my concern about the delegate pause.

Locally on a4e0c848: packages/chat deno tests 48 passed; app-shell run-diagnostics + run-node-session-chat 18 passed. I did not re-run the Rust suite (runtime unchanged in this revision).

Non-blocking nit: after an Ora inactivity retry, the diagnostics are cleared, so the panel shows "Unknown; replay is not new activity" during the new attempt. Nothing is being replayed at that point, so a neutral "no activity observed in this attempt yet" would be more accurate.

Merge is still gated on ora-space/specs#145 reaching approved, as the description says.

[ObsisMc]

审查 head 08295ae。Rebase 后前 6 个提交与此前批准的 a4e0c84 一致（range-diff 无差异），D3/D4 的主体结论不变。新提交 34d0621 有一处需要修改：

阻塞：run-diagnostics.tsx:71-76 只用 isResponding 区分"新尝试"和"回放"。但 loadSession 回放未结束的回合时，preview() 会设置 isResponding: hasOpenTurn（store.ts:1031），且不写入 diagnostics。工作流自动首个 prompt、重开运行中的会话都走这条路径（store.ts:352、run-node-session-chat.tsx:81）。结果是回合内容在逐段出现，面板却显示"本次尝试尚未观察到活动"，把"未知"说成了"没有"，与 ADR 要求的加载流标明未知不符，也与本提交保留历史重放文案的意图相反。
建议：面板同时考虑 isLoading（isResponding && !isLoading）；或由 store 在 sendMessage 和重试时把 diagnostics 设为 {}，面板用 diagnostics === undefined 表示未知/回放。请补一个 load 流中有未结束回合时仍显示回放文案的测试。

非阻塞：

• 只有子代理活动时，"最近父回合活动：本次尝试尚未观察到活动"容易被读成整个尝试都没有活动，建议改为"本次尝试尚未观察到父回合活动"。
• 子权限往返不续期目前只有 PromptLiveness 单元测试，actor 路由测试没有覆盖权限；建议补一个用例，或调整 PR 描述里的说法。

[ObsisMc]

Approve @ e988bee。

增量复审：range-diff 显示原 8 个提交与 08295ae 一致，只新增 9e1d0d4 和 e988bee；已追平 main be72c25，crates/history 与 main 无差异。CI 8 项全绿。

• 阻塞项已修复：sendMessage 与不活动重试把 diagnostics 置为 {}，loadSession 回放保持 undefined。面板先看 diagnostics 是否存在，回放（包括未结束回合、工作流自动首个 prompt）显示“未知；历史重放不代表刚发生的活动”，实时尝试才显示“尚未观察到父回合活动”。RunDiagnostics 字段全可选，{} 合法；observeDiagnostics 与 retry/compaction 分支对 {} 都按已有快照合并，没有行为回退。store 和面板都补了加载流用例。
• 文案：“本次尝试尚未观察到父回合活动”/“本次尝试未观察到父回合活动”区分进行中与已结束，避免与“未在处理”矛盾。
• 权限往返：e988bee9 让 fake provider 走真实 ACP session/request_permission，actor 路由用例证明子工具权限往返不续期、父工具权限往返重新计时，PR 描述的说法现在有直接证据。