fix(pr-review): skip duplicate work covered by independent reviews

[loopx-agent]

A contributor's ordinary GitHub review currently fails the agent's bilingual conclusion format and gets selected for another full review. This change separates submitted independent review activity from earned approval: source-complete activity by a non-author GitHub User associated as OWNER, MEMBER, COLLABORATOR or CONTRIBUTOR can cover duplicate review work on the current head.

• Reuse the TypeScript review-history owner and expose an optional independent_review fact. Formal COMMENTED (including inline-only), APPROVED and CHANGES_REQUESTED submissions count as activity; issue comments, the author, bots, unknown provenance, old heads, pending-only and dismissed opinions do not.
• Fetch the full paginated GitHub history and recheck the head/base after reading provenance. Dismissal retires that reviewer's coverage; missing or malformed metadata keeps the ordinary review runnable. Share the complete history reader with approval closeout.
• An inventory-only covered row has no executable review plan. An explicit fresh exact-head audit can restore one for an open, non-Draft or merged PR. Strict approval body qualification, unresolved dissent, validation, readiness and merge authority retain their existing gates.
• Update the existing CLI help, capability contract and thin host skill. There is no new configuration or UI; existing capability settings and their inheritance are unchanged.

This intentionally changes default queue selection for qualifying independent reviews. It does not authenticate a biological human or certify review quality. GitHub association is the provider-owned eligibility input; no reviewer-name allowlist or prose classification is added.

Validation at final abbc1f4b627310b240af4ace18b9d8d235654285, against target main 82f9c09bd61e26e458cecfd4cffeab2350359381:

• Complete offline review/source modules: 527 passed, 43 skipped at the preceding implementation commit; the final commit changes only the existing smoke expectation, with all seven production file hashes unchanged. The skipped tests explicitly opt into paid live-model qualification and were not executed; no model-quality claim is made.
• Ordinary noneditable base/head wheels through the isolated public CLI: 26 invocations, three newly covered activity states, nine negative parity cases, and the fresh-audit override. All seven changed installed production files match the reviewed source bytes. The final wheel also passed three CLI readbacks; eight additional exact-target cases cover inline reviews, unrelated inventory beyond the display cap, new subjects, prose/author impersonation resistance and coverage loss/recovery.
• Provider-path tests cover pagination, later dismissal, account provenance, source failure/recovery and post-read head drift. Closeout tests preserve approval and dissent holds.
• Configured Ruff and mypy, control-plane TypeScript typecheck, Chat/wheel builds, staged bilingual RFC gate, diff hygiene and public-boundary scan passed. Chat build retains its existing large-chunk warning.
• Exact-scope change-quality receipt passes. Final standard Goal-bound premerge passed 6 direct checks and all 19 selected checks, without manual holds. An earlier premerge run found the old no-action token asserted by the existing command smoke; the final commit corrects that expectation, retaining the assertion and every selected check. CI observation is disabled by the managed review configuration; local validation and exact-head gates remain required.

Bounded refactoring: consolidate review pagination and extend the existing history adapter; defer broader review-policy restructuring. Full provenance adds a paginated history read and version readback for foreign-review PRs, reusing bounded fanout. No queue-latency or throughput improvement is claimed.

中文：按当前提交上、来源完整的非作者贡献者/维护者正式评审跳过重复审查，不要求对方使用本 Agent 的双语模板。作者自评、机器人、旧提交、待提交/已撤销评审及未知来源不触发跳过。该事实只覆盖重复工作，不构成通过结论，也不解除异议、验证或合并权限门槛；显式重新审查仍可执行。

[loopx-agent]

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

Reviewer: model_agent | gpt-6.1-sol | OpenAI | runtime_reported | reasoning_effort=max

动机

处理 PR 队列的维护者和 Agent 会遇到同一提交被重复完整审查的问题。例如贡献者已提交普通行内评审，旧流程仍因正文不符合本 Agent 的双语模板而再次生成全文审查计划；新流程读取合格账号的完整当前提交评审后，仅回读已有记录并继续队列。

普通安装包的 CLI 已验证合格记录不再生成执行计划，而作者自评、机器人、旧提交、未知身份和来源不全仍保留原处理路径。本改动不把第三方评审视为批准，不解除阻塞、验证或合并权限，也不声称已验证真实模型收益。本 PR 已完成代码与普通安装包验证；合并后的本机升级和实际来源回读尚待执行。

改动思路

应保留这套有界设计：GitHub provider 只提供完整事实，TypeScript 历史 owner 决定覆盖，队列消费投影，批准与合并仍由原有门负责。本 PR 边界是打通现有评审入口的重复工作覆盖、失败恢复和文档适配；不新增配置、权限、持久状态或独立页面。已有历史选择和 approval closeout 继续共用类型化 owner；本次只是增加“已有独立评审活动”的派生事实，不把正文格式校验改成宽松批准。该事实每次从 GitHub 计算，不需要人维护另一份状态或设置。普通的 COMMENTED、APPROVED、CHANGES_REQUESTED 都可能覆盖重复工作；批准结论与工作覆盖是不同判断。

下表对应审查提交 abbc1f4b627310b240af4ace18b9d8d235654285，是源码与执行结果的简要对照：

条件	队列结果	批准/合并影响
来源完整、当前提交、非作者 User、合格仓库关联身份	仅保留记录，不生成全文审查计划	不授予批准或合并权限
作者、Bot、未知身份、旧提交、来源不全	沿原流程处理	原有门保持
最近已提交活动被撤销或 head 改变	原覆盖失效	不清除其他评审者异议
明确请求 fresh audit	允许范围内重新生成证据计划	仍须独立验证和授权

不改动会继续因格式重复审查；放宽原批准模板会误改权限语义；用户名白名单无法表达通用来源。因此保留现有 owner 是更小且可回退的修复。额外分页读取有实际成本，已经复用原有有界并发，不宣称测得吞吐或延迟收益。

具体改动

关键代码讲解

1. _attach_review_provenance 发现他人评审后读取完整 REST 历史，补齐账号类型和仓库关联身份，再核对 head/base。普通 gh pr view 不含这些身份事实；测试替换的只是 GitHub IO，真实 scan、归一化与类型化决策仍执行。读取失败或字段损坏不产生覆盖，版本漂移则拒绝混合快照。
2. independentReview 按真实 review id 和提交时间选择各账号最近的已提交事件，仅接受当前 head、非作者、明确非 Bot 且关联身份为 OWNER/MEMBER/COLLABORATOR/CONTRIBUTOR 的记录。PENDING 不抹去已提交活动，DISMISSED 只撤回该账号覆盖；缺失或无法验证的来源不会跳过工作。这是 GitHub 账号来源判断，不是自然人认证。
3. select_review_history 继续调用原有 RPC，返回原 index 和可选事实；_review_conclusion 原严格正文/作者 fallback 判断保留。审批 closeout 改用同一分页 reader，避免复制来源逻辑；活动与正式意见的 COMMENTED 差别有意分开，下一次资格政策只需改同一 TS owner。
4. materialize_review_execution 消费 native action：覆盖行没有 plan/template/证据命令，并回读评审者、状态、原记录链接。已有严格批准仍走 readiness 路径。显式 fresh audit 可以重启覆盖行，新增 peer 路径限制为非 Draft 的 OPEN 或 MERGED，不把 CLOSED/Draft 变成绕过入口。原 CLI 帮助、双语能力文档与薄 skill 同步说明默认变化。

规范依据是改动前的 loopx/capabilities/pr_review_queue/README.md，固定修订 82f9c09bd61e26e458cecfd4cffeab2350359381 的 Source Reads 段，而不是本 PR 改写后的文字。C1：严格批准格式与作者 COMMENTED 标题仍由原正文测试和 closeout 约束。C2：同一类型化历史 owner、各账号意见和撤销语义保持；分页和顺序变换的三种来源形状得到相同完整行。C3：换 head 需新证据，无格式的异议仍可见，来源记录不授予撤回或合并权限。新的重复活动覆盖是当前维护者要求的默认改变，并已用普通安装包作反事实验证。

验证包括完整离线评审/来源模块 527 通过、43 个明确 opt-in 的实时模型资格项跳过；基线与实现普通不可编辑 wheel 的 26 次 CLI 调用、最终 wheel 三次读回、八项公开 exact-target 范围/恢复验证和三种生产来源分页/顺序形状。基线会对三类合格活动再次生成全文计划，改后不再生成；九类不合格/来源不全输入保持原结论诊断和动作。最终 head 与模块测试修订的七个生产文件字节完全一致，仅追加既有 smoke 期望修复。

最终标准 premerge 的 6 项直接检查、19 项选定检查全部通过，无人工 hold；包括双语 RFC、公开边界及现有端到端命令 smoke。当前范围的 change-quality receipt 有效，配置化 Ruff/mypy、TS 类型检查及正常 Chat/wheel 构建通过。更早的 ce99d83 premerge 暴露旧 no-action token 的断言遗漏，最终提交只修正该期望，未删断言、改限额或跳过检查。按解析后的配置未查询、轮询或等待 CI。

对主干的风险

主要风险是把不完整、旧的或伪装的记录误当覆盖。来源层的完整分页、账号类型、关联身份和版本复核，配合类型化精确匹配及否定用例限制了它。标题/正文声称“维护者已通过”、作者即使具有 OWNER 身份，都不能建立覆盖；新增 PR 和超过显示上限的无关项目不会改变明确目标的结果。源丢失时恢复原处理，来源恢复或新提交的合格评审可继续，显式 fresh audit 也保留可控恢复路径。

GitHub User 可能由自动化使用；这次按账号事实去重，不证明评审质量。CHANGES_REQUESTED 可表示“已有独立评审活动”，但仍是阻塞意见，closeout 的三类状态用例都保留 exact approval 缺失/异议/授权门。没有新增配置、数据库、quota、权限、外部写入或 UI 布局；已检查既有机器/Goal 能力编辑器，沿原会话/CLI 入口交付。额外 REST 读取的延迟未作量化，构建原有大 chunk 警告保留；实时模型行为、长期成本和实际本机采用不在这些离线证据中。

我的整体评价

没有发现阻塞性的代码或当前验证问题。当前交付是有用且有界的增量：代码和普通安装包已经实现可用的重复审查覆盖，撤销、换头、未知来源、明确重审及批准边界均有反例支持；本机升级与实际来源读回仍是合并后的同一授权交付步骤。本轮没有把测试数、记录或 context delivery 当成长期收益证明。已有恢复类经验仅采纳“从普通用户任务检查可用结果”的方法，其恢复 RFC 场景对本 PR 不适用，不继承历史结论或声称记忆效用已证实。

设计保持同域复用：共用历史 reader 与 owner，不引入另一套状态或设置；更大规模的评审架构调整没有当前必要性。合并仍需要精确 head 的公开评审、有效本地验证、readiness 回读和维护者授权，覆盖事实不能替代这些步骤。

English verdict: APPROVE — abbc1f4b627310b240af4ace18b9d8d235654285. Independent current-head contributor/maintainer review activity covers duplicate work without earning approval or clearing dissent. Final local premerge passed 6 direct and 19 selected checks; offline review/source tests passed 527 with 43 opt-in live-model skips, and ordinary-wheel counterfactuals passed. Native host upgrade/readback follows merge; CI was not consulted under the configured policy.