[loopx-agent]

A steward request could finish tool calls and then hit model quota without producing a final answer. The old recovery path correctly refused to replay uncertain actions, but also had no way to finish the answer, leaving the original question permanently owed.

Add one source-verified, read-only answer review through the existing original reply worker and the configured Codex account pool. Preserve completed tool outputs, the failed Turn and its notice; never replay the original action. Keep pending questions visible when a review fails or returns only a gate. Newer independent questions cannot invalidate an already-formed answer.

The review only retains reads already granted by the current steward runtime and disables writes, network, apps and action metadata. Restricted runtimes remain workspace-only. Existing native-adapter owners absorb the mechanical profile/start/model helpers without increasing module budgets or changing worker account priority.

Validation: 25 focused original-worker and native-proof cases; 53 TypeScript cases; 621 agent, store and recovery cases after owner extraction; 48 focused alternate-route and original-worker cases after the final route changes. The same original-worker regression fails against the immutable base. Final standard premerge receipts qualify the committed head. Real installation, original-service execution and original-provider reply readback remain separate acceptance.

[loopx-agent]

Reviewer: actor_kind=model_agent; model=GPT-6 (Codex); provider=OpenAI; self_reported (host execution observation unavailable).

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

当前 exact head d89294bbf9a9900ca429a9e180e753a834c057aa 没有阻塞发现。审查覆盖完整 12 文件差异、原生证据读取、权限下限、一次恢复与原消息回传，未把源代码测试当成真实历史消息已答。

动机 / Motivation

等待原私聊管家回答直接问题的用户，以及原回复协调 worker。 先执行三个工具读取再遇额度错误：旧路径只能留下失败通知；新路径在原输入、执行和来源证明匹配后，仅用已有结果和当前获准读取形成一次答案。 同一 Core Session、原受众收到与原问题绑定的补分析；原失败 Turn 和失败通知保持原样。 不替接收 Agent 执行业务，不重新运行原动作，不改 worker 账户优先级，不保证未知或缺失证据的问题已有答案。 源代码和隔离原 worker 协议已验证；安装、真实原问题的实质答案和原受众实际读回仍须独立验收。其他历史问题、委托结论及未知效果不由本 PR 自动核销。

改动思路 / Approach

复用原请求账本、原 TypeScript 准入、原回复 worker、原配置账户池及 native adapter。只新增一次受限补分析，不重放原业务，不增加消费者、候选账户或调度 owner。 原生输入 hash、闭合区间、工具完成状态和实际执行归属必须匹配；只取已完成的可见工具结果，排除私有思考。不存在正常最终答案才进入补分析；原有完整答案仍沿仅补正文路径返回。

具体改动 / Changes

规范依据 docs/architecture/rfcs/capable-manager-semantic-handoff-v0.md，不可变 revision b4216d8c79f7f31785cf3bb6845f6847fcf0a6a6：request-level-coverage、no-action-replay 和 failed-notice-not-completion 分别由原 worker、精确 native proof、原记录不变及 gate 仍 pending 断言覆盖。installed-original-return 的真实安装和原 provider 完整答案保留为合并后独立验收。

关键代码讲解

interrupted_answer_observation（loopx/capabilities/native_chat/antigravity_adapter.py:384）两次核对原闭合片段，只接受已完成的已知工具和终止错误；终止错误缺少 executionId 只在已锚定片段允许，外来 executionId 仍拒绝。planInterruptedStewardAnswer（loopx/control_plane/collaboration/conversation_binding.ts:1103）复用原 CAS 和 fresh route 只保留一次补答，失败和 gate 仍为 pending，禁止落回普通业务重试。_checked_answer_review / execute_turn（loopx/capabilities/native_chat/external_conversations.py:695）派发前再验来源和原结果，独立后续问题或无关 native 尾部增长不会丢掉已形成答案，所有 action metadata 均拒绝。native_conversation_policy（loopx/capabilities/native_chat/codex_context.py:220）只保留当前原 runtime 已有的读取权限：restricted 仍 workspace-only，已有 trusted_owner 才保留宿主读取；统一禁写、网络、apps 和 patch。普通后问清除 review 标记并恢复原适配器。

对主干的风险 / Risks

最强反例是把补答变成重放或给 restricted runtime 新宿主权限。原 worker 集成实际验证 native wire profile、dynamicTools、文件账本和 provider 回复，断言原失败 Turn、原失败通知、原工具产物不变、原 native send 恰好一次；新增错误、async/未知工具、source/stop/queue/容量变化均保守拒绝。丢失 ACK、enqueue 后崩溃、已经切到 Codex、补答中收到新问题和补答返回动作或 gate 也覆盖。

25 focused native-proof/original-worker cases; 53 typed cases; 621 agent/store/recovery cases after owner extraction; 48 alternate-route/original-worker cases after the route change; final committed-head premerge 6 direct and 10 selected checks pass. The identical positive original-worker regression fails on immutable base. 621 项是在 owner extraction 后运行；最终路由变化另外运行了相关 48 项及最终 25 项，未冒称所有 621 项都在最后一次小改后重跑。初次 premerge 发现模块超出已有 ceiling，现将机械配置/启动/模型归一化放回既有 native owners，旧 normalize_agent_model_id 导出保留，预算没有调高，最终 maintainability 和标准 premerge 通过。

语义与 CI 对齐

本地新增 answer_review 与权限 profile 复用现有 recovery/metadata 语义；没有新队列、调度、账户选择器、公开 CLI 或额外权限授予。当前 API-primary alternate 已新鲜可用时，不让无关 Argon quota 读失败再次卡住补答；不把这个路由观察当成账户余额证明。没有默认关闭能力或加载指令，default-off 不适用；普通路径保留，restricted 反例有线级断言。按当前 wait_for_ci=false，没有查询、轮询或等待远端 CI。

我的整体评价 / Verdict

long_horizon 与 user_experience 均改善：已执行动作和缺失答案分开恢复，用户不必重复提问，后续输入不再挤掉旧答案。代码沿原 owner，修改范围与实际故障相称；合成协议已覆盖主要安全和生命周期反例。没有页面结构或布局变化，视觉审查不适用。源代码和隔离原 worker 协议已验证；安装、真实原问题的实质答案和原受众实际读回仍须独立验收。其他历史问题、委托结论及未知效果不由本 PR 自动核销。

English verdict: APPROVE d89294b; the original owner can form one read-only answer from verified interrupted results without replaying actions or expanding current grants. Installed original-provider delivery remains separate acceptance.