[loopx-agent]

Goal And Delivered Outcome

修复大 PR 文件清单与精确 head 混源的问题。GraphQL 详情只返回前 100 个文件时，现有 provider 会通过 REST 分页补全；PR 若在分页期间换 head、更新 base 或改变统计，旧实现仅检查文件数量，仍可能把另一版本的文件作为完整来源。

本改动在现有 github_source 入口复用 snapshot 回读，核对 head、base、文件数和增删行数；重复路径也拒绝。普通队列保留 source-incomplete 结果，精确目标读取沿原错误路径退出，调用方重新读取当前版本即可。没有新增审批、配置或审查协议。

• Outcome basis: reproduced provider-boundary regression; source completeness in the existing pull-request-review capability.
• Intended base: main, rebased onto 71970fb03eb2ab90b353f0895bb1463ca7eaf236 before submission.

Author Declaration

• Written by: model_agent — gpt-6.1-sol / OpenAI, runtime_reported, reasoning_effort=max.
• Specification: existing exact-head and complete-source rules in loopx/capabilities/pr_review_queue/README.md and skills/loopx-pr-review/SKILL.md at e7a6c8d5f09a8c7e5d6021aabce00223c9324fce.

Criterion	Disposition	Implementation / evidence
Keep source details bound to the selected exact head	implemented	Reuse _fetch_complete_pr_files.same_snapshot after ordinary paginated reads; queue and exact-target drift regressions
Incomplete source is not a complete review inventory	implemented	Reject duplicate file paths and unavailable version readback through existing incomplete-read paths
Preserve configured no-CI behavior	implemented	New cases reject any statusCheckRollup query; no CI polling or waiting

Scope And Continuation

Complete within this provider fix. The same existing Python IO owner serves queue discovery and exact targets; TypeScript review decisions and publication/merge authority are unchanged. No new owner, schema, persistence or compatibility branch is introduced. Reusing the existing version check is the bounded simplification; a new snapshot framework is unnecessary.

The version readback adds one small metadata read only when ordinary REST file completion is needed. It detects observed source drift, without claiming server-side snapshot isolation. Final review publication still requires the existing fresh-head check.

Validation

• Tested revision: d69857a
• Run state: finished
• Input classes: synthetic, public_fixture

Check	Result	Evidence / limitation
Regression before / after	passed	Before the fix, 14 drift/unavailable/duplicate cases were wrongly accepted; 3 positive controls passed. The same cases pass after the fix.
Related Python suites	passed	tests/test_pr_review_github_scan.py plus Git inventory recovery, independent-review and approval-closeout suites: 197 passed. Includes existing >3000-file Git recovery and ordinary non-paginated reads.
Static	passed	Changed-file Ruff, provider mypy, diff hygiene and diff-scoped semantic advisory; no supported new production vocabulary carrier.
Real entrypoint / backend	passed	Source-checkout CLI with the real GitHub provider read PR #7430 at d1ed715198bc3d9eafb0fc7ff38afa63f2e79510: complete inventory of 188 files, matching 18,289 additions / 1,979 deletions. This verifies stable read success; concurrent mutation was simulated at the provider boundary.
Standard premerge	passed	Same final head and immutable base 71970fb03eb2ab90b353f0895bb1463ca7eaf236: all direct hygiene/bilingual/compile checks and both selected canaries passed; maintainability 7.33s, semantic 31.28s. The original final-head semantic timeout at 120.553s and earlier pre-rebase pass at 88.53s remain historical results. No source, workload or 120-second limit changed.
Controlled timeout diagnosis	passed_with_limits	Declared base/head/head/base sequence on the same checkout and Python 3.14.8: base 26.96s / 33.07s, head 26.43s / 26.45s; all four full checks passed with identical stdout. No candidate regression was observed in these samples. A separate Pyinstrument 5.1.3 capture completed successfully and points to existing semantic scans; it is diagnostic, not ordinary latency. Host concurrency was uncontrolled and the original failure lacks resource counters, so the historical cause and sustained reliability remain unproven. One full premerge followed this diagnosis; no repeated-until-green loop, timeout relaxation or performance-improvement claim.
Remote CI / installed adoption	not_run	No remote CI queried, triggered or awaited. No installed runtime changed.

Frontend / Visual Evidence

UI impact: none. This fixes an existing read-only provider's source-integrity check; existing capability settings and CLI failure/retry paths remain the journey. No rendered surface changed.

Type of Change

• Bug fix
• Test update

LoopX Area

• Capability or extension (providers, adapters, skills)

Shared-authority RFC fixture impact

N/A: no authority-store, runtime routing or persistence change.

Boundary Checklist

• Public-safe source, synthetic fixtures and aggregate evidence; no private state or raw traces.
• No benchmark work duplicated.
• Scope remains the demonstrated file-inventory defect.
• No RFC pair changed.
• No UI change.
• DCO sign-off on every commit.

[loopx-agent]

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; reasoning_effort=max; declaration_source=runtime_reported

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

动机 / Motivation

通过 CLI 读取大型 PR 的评审者，需要文件清单与所选提交版本一致。旧版本在 GraphQL 只返回前100个文件时转用 REST 分页；即使分页期间提交、基线或增删统计改变，只要文件数相同，仍可能把混合版本当成完整来源，导致审错内容和重复返工。

本次验证的改善是：分页完成后重新核对版本和统计，并拒绝重复路径；稳定版本继续返回完整清单，变化或读取失败明确保留为不完整来源。本次不改变审查权限、队列顺序、合并条件或调度，也不承诺跨请求的 GitHub 原子快照。

改动思路 / Approach

已有 provider 统一负责文件来源完整性；复用同一版本比较，将新增成本限制在必要的只读回查和局部去重，不引入第二套状态或权限判断。本次交付分页来源完整性修复及两种现有读取入口的回归验证；不扩展审批、合并、调度或远端原子性。

现有入口先读取 PR 版本与详情，详情不足时沿同一个 helper 分页。原实现仅在超过 API 上限、需要本地 Git 恢复时核对版本；本次让普通完整 REST 结果也使用已有的五字段比较，并在收集过程中拒绝重复路径。失败仍由原调用方处理：普通队列保留不完整结果，指定提交的读取明确报错，随后重新获取当前来源。稳定输入不增加人工确认、配置或复制步骤。

最强反对理由是额外一次 metadata 请求可能把瞬时网络故障转成 incomplete。但在需要精确提交证据的入口，明确重读比把另一版本标成完整更符合既有契约；无需为此增加缓存、第二个判断器或持久状态。分页后的采样仍不是远端原子快照，不能声称消除了所有并发变化。

具体改动 / Changes

完整 diff 为2文件，生产代码6+/2-，测试71+；没有文档、生成物、UI 或状态迁移。审查版本为 d69857a112be42abbf3d5ea24b5ceb7da2146e27，实际 diff 的不可变 merge base 为 71970fb03eb2ab90b353f0895bb1463ca7eaf236，区别于 GitHub 当前基线指针。

关键代码讲解

• _fetch_complete_pr_files：在 REST 每页收集时用局部路径集合拒绝重复项；达到声明数量后，只要调用者提供了 snapshot，就重新读取 head、base、文件数和增删总量进行完整比较。任一不符或读回异常返回 None，不能靠相同数量通过。
• attach_pr_review_details：现有调用者已经把版本化 row 传给分页 helper；本次保留它的失败、列表顺序和结果装配规则，没有另设调用者分类逻辑。
• scan_github_pull_request_targets：指定准确 head 的公开读取复用同一个详情路径。新拒绝分支沿已有 detail read was incomplete 错误退出，没有把 source completeness 当成 review 或合并许可。

规格依据为 loopx/capabilities/pr_review_queue/README.md，spec_revision=71970fb03eb2ab90b353f0895bb1463ca7eaf236，以改动前的来源完整性与权限条款为准。原文没有编号，以下标签对应既有条款：INV-VERSION（head/base/统计围栏，版本变化为 incomplete）implemented；INV-COMPLETE（失败保留为不完整、稳定完整读取保持形状）implemented；INV-AUTHORITY（完整 API、已有本地 Git 恢复和 closeout 不扩权）implemented。对应证据分别是同一不可变 fixture 的两种入口反例、完整稳定观察逐值对照，以及 inventory/approval-closeout 全模块验证。

已搜索全部 helper 调用及现有 snapshot/Git 恢复实现；当前主干 provider 与这里的 base 仍是同一 Git blob，未把已合入能力重复当作新成果。复用成立，额外抽象没有当前调用者价值。集合属于本次调用内的派生值，不是新的状态词表，也不按字符串片段或文案猜测执行义务。

对主干的风险 / Risks

没有发现本 diff 的阻塞性正确性问题。关键反例是“文件数未变，但分页后版本已变”：同一16例 harness 通过真实生产 scan 入口运行，在 base 为2 PASS／14 FAIL，在 head 为16 PASS。七类负例覆盖 head、base、数量、增删统计、回读不可用和重复路径，两种入口均拒绝；普通队列保持可见的 incomplete，指定提交沿原错误退出。两个稳定正例的全部101行和完整输出逐值相等，没有忽略字段来制造一致。

验证	结果与边界
四个完整 provider、inventory、独立 review 和 closeout 模块	PASS，197项
同一 base/head fixture 的普通／准确提交 scan	PASS；base14负例失败，head16例全部通过；传输被控制，完整性判断未被模拟
实际 source-checkout CLI 与真实 GitHub API	PASS；同一稳定192文件 PR 的来源版本、规模及 complete 一致；packet 只展示10个 key_files，仅这10行做了相等比较，未声称比对完整192行
两端 changed-file Ruff、配置的17文件 mypy 和 provider mypy	PASS
配置的必需 CI Ruff 命令	FAIL；两端同一 E731，位置 tests/test_manager_context_roundtrip.py:289，完整697字节失败输出完全相同，源文件、workflow、配置和依赖锁字节未变
标准 exact-diff premerge	PASS；直接检查和所选 canary 通过，无手动 hold；不授予合并权限

必需 Ruff 失败仍是失败，严格 change-quality 资格保留 FAILED 和合并阻塞；独立证据表明它在未修改的路径上，当前 policy26 允许保留本次正确性 APPROVE，不能把它写成全绿或要求本 PR 顺手修无关代码。更宽的额外 Ruff 诊断也在两端以同一120项失败、34个失败路径逐个 Git blob 未变，未混充配置的 CI 范围。

作者此前记录的语义 smoke 120秒超时保留为历史观察；本次 exact-head 标准检查在原限制下通过，未由一次成功推断历史原因或长期性能。按当前契约没有查询、轮询或等待远端 CI。真实 GitHub 实测是稳定来源集成，未模拟网络中的真实并发变更；101行反例测试承担这部分受控判别。恢复方式是按当前 source 重读并重新建立准确提交证据，没有后台重试或额外扣额。

我的整体评价 / Verdict

goal_achieved 仅指本次分页来源完整性修复的可验收任务。long_horizon=improved，依据是防止混合版本反复进入后续评审、失败后仍有原重读路线；user_experience=improved，依据是同一 CLI 在稳定输入下不增加人工步骤，而错误来源明确可见。这些是受影响路径的行为判断，没有建立模型采用、净 token、持续时延或完整人工成本收益。

架构 retain，范围 proportionate：一个既有 provider、一次必要的版本回查及局部集合，完整稳定结果保持；default_off_isolation=not_applicable，因为这是公开说明的默认正确性修复，没有新增可选能力或静默开启配置。权限 aligned，未改变 review、closeout、合并或调度授权。局部复用已经是可审查的简化，不需要为了减少行数移动别的 owner。

visual_review=not_applicable：检查的是 CLI JSON/provider 和未变化的展示调用者，没有渲染页面修改；渲染范围无待验项。远端原子快照、安装态 Host／模型采用、长期效率和完整192行现场相等均未验证。当前准确 head 的 APPROVE 与必需 lint／严格资格、GitHub readiness 及维护者合并权限分开，后者继续保留。

English verdict: APPROVE — d69857a; verified version-fenced pagination and duplicate rejection, stable-path parity retained; pre-existing required lint and strict qualification remain failed, with merge authority separate.