headless marker-reading 系统性统一:worker marker 落 artifact 而非 log 致 detection/revalidation/readiness/review 多处误判

[loning]

问题

headless 下 worker codex(implement / reviewer)经常把终止 marker(IMPLEMENT_DONE:issue-<N>:<status> / REVIEW_DONE:<pr>:<role>:<verdict>)只写进 run artifact(.refactor-loop/runs/*.md)末尾,不在 log 末尾 echo(codex stdout marker 落点不可靠,同批 worker 有的在 log、有的只在 artifact)。系统多个消费点只扫 log 末尾 marker,导致 markerless worker 被系统性误判:

1. detection(wakeup_plan.completed_marker_actions):markerless → 不生成 publish/review action。已 hotfix(commit 8aeea89,implement artifact fallback)。
2. revalidation(wakeup_runner._source_log_has_clean_marker):markerless → reject publish(clean_exit_marker_missing)。已 hotfix(commit dfacc64)。
3. readiness(consensus-implementation readiness / suppressed_reason):最破坏的一处。系统扫 log 不知道 implement 已完成(marker 在 artifact)→ 判定该派 implement → 重派覆盖已完成的 worktree 工作,原本 IMPLEMENT_DONE:ok 退化成 :partial、工作丢失 + 浪费 codex(dogfood 实测 测试纪律修宪(CLAUDE.md):behavior-first;source-regression 仅锁 doc↔code 授权边界;禁同义反复实现字面断言 #493 :ok→:partial、发行前 clean-room 集成测试制度:独立干净 host 仓库完整 dogfood 验证后才发版 #421/同义反复 source-literal 测试普查与替换为 behavior test #494 被重派)。未修。
4. review verdict(review gate truth table):markerless reviewer verdict 在 artifact、log 无 → 真值表判 missing role → WAIT_OR_REDISPATCH,review→merge 卡(dogfood 实测 让 review comment 与 refactor marker 冲突可收敛 #402 tests/quality REVIEW_DONE:approve 只在 artifact)。未修。

连续单点 hotfix(detection + revalidation)已证明是打地鼠:readiness 重派还会丢失已完成工作,review 还卡。implement→PR、review→merge 两条链都被 markerless 卡死(open PR 数长期只有 2)。

新原则(待共识收窄)

统一 marker-reading robustness:一个共享的 marker reader(契约待定:log-first + artifact-fallback,或与 #205 rule 4 的 review-verdict-artifact-first 对齐),所有终止-marker 消费点(detection / revalidation / readiness / review-verdict)都经它,保证 worker marker 落 log 或 artifact 都被一致识别。可双管齐下同时强化 worker prompt 要求 marker 落 log 末尾,但不能只靠 prompt(worker 行为不可靠是前提)。

不变量 / 边界

• 仍只接受 standalone 终止 marker(沿用既有 marker 真值表语义);artifact 读取限定 runs/<worker>.md 同名伴随面,clean-exit-scoped。
• 不放宽 review 真值表、merge gate、consensus gate 的判定语义;只修「marker 在哪读」,不改「marker 意味着什么」。
• 与 headless daemon fail-closed 点的 codex 兜底裁决/生成机制(narrow allowlist,非通用 escape hatch) #498(daemon fail-closed 派 codex 兜底)是姊妹问题(都应对 worker 行为不可靠),但本 issue 是 detection robustness,不引入 codex 兜底。

decision questions

1. 共享 reader 契约:log-first+artifact-fallback / artifact-first / 两者一致性校验?
2. readiness 端如何识别「已完成 implement,不重派」(避免丢工作)而不与「真需重派的失败 implement」混淆?
3. 与 design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 rule 4(review verdict from artifact frontmatter)如何收口为单一来源?
4. 是否同时加 worker-prompt 强化 + source-regression 锁住所有消费点都用共享 reader?

证据

• commit 8aeea89(detection fix)、dfacc64(revalidation fix)、71edf91(tick 日志详细化,暴露被掩盖的 blocked)
• dogfood 实测:测试纪律修宪(CLAUDE.md):behavior-first;source-regression 仅锁 doc↔code 授权边界;禁同义反复实现字面断言 #493 IMPLEMENT_DONE:ok→重派→:partial、发行前 clean-room 集成测试制度:独立干净 host 仓库完整 dogfood 验证后才发版 #421/同义反复 source-literal 测试普查与替换为 behavior test #494 重派、让 review comment 与 refactor marker 冲突可收敛 #402 reviewer verdict 仅在 artifact、open PR 数长期=2

⟦AI:AUTO-LOOP⟧

[loning]

🤖 minimal solver 建议：用共享 reader 收口 marker 读取

TL;DR

• 这是什么: headless marker-reading 系统性统一:worker marker 落 artifact 而非 log 致 detection/revalidation/readiness/review 多处误判 #499 的 minimal-change solver 方案。
• 现在到哪一步 / 结论是什么: 建议新增一个 clean-exit + 同名 artifact fallback 的共享 worker terminal marker reader，并让 detection / revalidation / readiness / review 统一使用。
• 需要 maintainer 做什么 OR controller 下一步: 不需要 maintainer 决策；controller 可等待本轮其他 solver 与 meta-judge 收敛。

---

详细说明

当前 checkout 里已能看到两个局部热修：wakeup_plan.py 只给 implement detection 做 artifact fallback，wakeup_runner.py 只给 revalidation 做 implement fallback；但 implement_lifecycle.py 的 readiness 仍只读 log marker，review completed-marker detection 也仍会漏掉 artifact-only REVIEW_DONE。最小修复不是改 truth table，而是把终止 marker 的读取位置统一成一个 owner-local reader。

建议的 reader 契约是：clean EXIT=0 的 worker log 优先读 log 中 standalone marker；log 无 marker 时，只读同名 .refactor-loop/runs/<log-stem>.md 尾部 standalone marker。review gate 的 verdict 仍以 artifact frontmatter 为准，REVIEW_DONE marker 只证明 worker 完成并触发路由，不替代 verdict 语义。

---

📎 完整 codex 原始输出(存档备查)

---

solver: minimal
issue: 499
cluster: issue-499
verdict: propose

Recommended framing

最小可行边界是新增一个很小的 worker_terminal_markers.py 共享 reader，把现有 wakeup_plan.py 的 implement artifact fallback 和 wakeup_runner.py 的 revalidation fallback 合并为同一契约：clean EXIT=0 的 worker log 先读 log 中 standalone 终止 marker；若 log 无 marker，则只读同名 companion run artifact .refactor-loop/runs/<log-stem>.md 的尾部 standalone marker。随后只改四个消费点接入该 reader：completed-marker detection、runner clean-exit revalidation、implement readiness 的 classify_implement_attempt、review completed-marker / review evidence head 读取。review gate 的 verdict 语义不放宽：artifact frontmatter 仍是 verdict 权威，marker 只作为完成/路由证据；无 clean EXIT=0 仍一律不接受 artifact marker。

Concrete plan

• Files:
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/worker_terminal_markers.py: 新增共享 reader，提供 terminal_exit_line、is_clean_exit_log、extract_standalone_done_marker、marker_from_clean_worker(log_path, prefixes=None)、clean_worker_has_marker(log_path, marker)；artifact 路径机械映射为 log_path.parent.parent / "runs" / (log_path.stem + ".md")，只读尾部 MARKER_TAIL_LINES，不扫描任意 artifact。
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/wakeup_plan.py: 删除 _implement_artifact_marker_fallback、_reviewer_log_has_valid_marker 等局部 marker 读取；completed_marker_actions() 改用共享 reader，使 REVIEW_DONE 只在 artifact 时也能生成 review action；latest_reviewer_heads() 用 artifact/frontmatter+clean log 读取 head，log-only fallback 也走共享 reader。
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/wakeup_runner.py: _source_log_has_clean_marker() 改用共享 reader，删除 _implement_run_artifact_has_marker()；_review_evidence_from_log() 用共享 reader 取 REVIEW_DONE，_review_evidence_from_artifact() 保持 frontmatter verdict 权威但用共享 reader 校验 marker 计数/clean companion。
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/implement_lifecycle.py: classify_implement_attempt() 的 marker 查找从 clean_implement_done_marker(lines) 改为共享 reader，接受 clean log + 同名 artifact 中的 IMPLEMENT_DONE:*:ok；失败/partial/nonzero 仍 redispatch，publish_ready 仍要求 canonical branch、worktree、base、non-empty diff。
  • skills/codex-refactor-loop/SKILL.md: 在 structured-consumption / phase routing 附近补一条局部合同：终止 marker 消费必须走共享 reader；clean EXIT=0 是 artifact fallback 的前置条件；companion artifact 只能是同名 .refactor-loop/runs/<worker>.md；review verdict artifact frontmatter 不被 marker 替代。
  • Tests: 修改现有反向测试，并补行为/源码回归测试，见下。
• LOC delta: ~+110 / -70，净增约 +40；触及 5 个源码/文档文件 + 3 个测试文件。
• Tests to add/modify:
  • skills/codex-refactor-loop/scripts/test_worker_terminal_markers.py：新增，覆盖 log-first、artifact fallback、no clean EXIT=0 fail closed、wrong companion artifact 不接受、duplicate/malformed marker 不接受。
  • skills/codex-refactor-loop/scripts/test_wakeup_plan.py：把 test_consensus_implementation_readiness_redispatches_markerless_local_attempt 改为 “artifact marker ok suppresses redispatch / implementation_ready_to_publish”；新增 completed_marker_actions 可从 review-pr<N>-<role>-r<R>.md 的 REVIEW_DONE 生成 action。
  • skills/codex-refactor-loop/scripts/test_wakeup_runner.py：把 test_dispatch_consensus_implementation_redispatches_markerless_implement_log 改为 artifact marker 存在时 block dispatch 且理由为 consensus_implementation_not_ready:implementation_ready_to_publish；保留无 artifact marker 时可重派。
  • skills/codex-refactor-loop/scripts/test_wakeup_runner_review_gate.py：新增 reviewer log clean exit、artifact frontmatter verdict + artifact-only REVIEW_DONE 可进入 truth table；无 clean exit 仍 WAIT；frontmatter verdict 与 artifact marker 不一致时 frontmatter verdict 生效或 invalid，按现有 design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 rule 4 断言。
  • source-regression：在 test_wakeup_plan.py 或新增 test_worker_terminal_marker_source.py 断言 wakeup_plan.py、wakeup_runner.py、implement_lifecycle.py 不再定义 implement-only fallback，且 import/use worker_terminal_markers；SKILL.md 含 companion artifact/clean exit/review frontmatter 三个锚点。
• Governance/ruleset change (if any): skills/codex-refactor-loop/SKILL.md 小幅补充局部合同即可；不改 CLAUDE.md/AGENTS.md。当前规则要求 structured consumption 和 phase routing 只消费 clean exit + standalone marker，这次只是把 “marker 在哪读” 写成 owner-local 事实源，值得增加 trusted-base 成本，因为已存在两个热修 fallback 且 readiness/review 仍分裂；深共识可达，因为不新增 lifecycle authority、不放宽 truth table、不改变 marker 含义。
• Migration path: no migration needed。既有 logs/runs 文件名已经同名配对；新 reader 只让过去 artifact-only 的 clean worker marker 被识别，未 clean exit 或无同名 artifact 的旧记录维持 fail closed。

Risks

• companion artifact fallback 若过宽会把草稿 artifact 当完成证据；用 clean EXIT=0 + 同名 artifact + standalone allowlisted marker + tail window 收窄。
• review 端存在两类事实：frontmatter verdict 与 REVIEW_DONE marker；计划保持 frontmatter verdict 为 truth-table verdict，marker 只证明该 worker 完成，避免 design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 rule 4 回退。
• implement readiness 接受 artifact marker 后会更早进入 publish-ready 分支；仍需 canonical branch/worktree/base/diff 校验，避免把空/错 worktree 发布。
• source-regression 若只检查字符串会脆弱；应检查 “消费点导入共享 reader” 和 “旧 fallback 函数名不存在” 两类稳定事实。

Escalation triggers (if any)

• none

Reasoning trace (internal — short, for meta-judge)

• Why this is the minimum: 当前 checkout 已有两处局部 fallback：wakeup_plan.py:684 的 implement detection fallback 和 wakeup_runner.py:1086 的 revalidation fallback；但 implement_lifecycle.py:65 仍只读 log lines，导致 readiness 把 markerless clean implement 判 redispatch；review completed-marker detection 也只从 log 取 marker，虽然 review gate 本身已有 artifact-frontmatter verdict 读取。把这几处合并到一个小 reader 是删除重复而不是引入新主链路。
• What I considered but rejected: artifact-first 全面替换；会让 draft artifact 在无 clean log 时更容易误入 truth table。两者一致性强校验；会继续拒绝 artifact-only marker，无法解决 issue。只改 prompt 要求 echo marker；issue 前提就是 worker stdout marker 落点不可靠。
• What I cannot decide alone: 是否未来把 solver/meta-judge/router marker 也迁到同一 reader；本 issue 最小边界先覆盖 issue 明确列出的 implement/review worker 终止 marker 消费点。

⟦AI:AUTO-LOOP⟧
SOLVER_DONE:minimal:propose:新增共享 clean-exit companion artifact marker reader，并接入 detection/revalidation/readiness/review 四个消费点

⟦AI:AUTO-LOOP⟧

[loning]

🤖 structural solver: 收口为共享 worker marker reader

TL;DR

• 这是什么: structural solver 建议把 clean-exit worker 终止证据抽成脚本内共享读模型。
• 现在到哪一步 / 结论是什么: 建议新增 worker_markers.py,让 detection / revalidation / readiness / review gate 都走同一 reader。
• 需要 maintainer 做什么 OR controller 下一步: maintainer 暂不需要动作,等 3 solver + meta-judge 收敛后再进入 implement。

---

详细说明

我建议新增 WorkerCompletionEvidence(worker 完成证据的结构化对象) 和 worker_markers.py(只读同名 log / runs artifact / review frontmatter 的共享模块)。它不做 git/gh/lifecycle 操作,只回答一个问题:这个 worker 是否在 clean EXIT=0 下产出了 standalone 终止 marker 或 review verdict。

方案保留现有语义:implement 必须 clean-exit 且只接受 standalone IMPLEMENT_DONE:*:ok;review gate 仍按 #205 artifact frontmatter verdict: approve|comment|reject 优先,artifact 缺失才 fallback clean log-tail REVIEW_DONE。变化只在读取位置统一:artifact-only implement ok 不再被 readiness 判成 markerless 后重派,artifact-only reviewer verdict 不再让 truth table 缺 role。

方案	trade-off
共享 worker_markers.py	触碰 3 个 runtime 文件,但一次性消除 4 个 consumer 的 parser 分叉。
继续局部 hotfix	改动少,但 detection/revalidation/readiness/review gate 会继续漂移。
全局 artifact-first	表面统一,但会改变 implement clean-exit scope,风险更高。

---

📎 完整 codex 原始输出(存档备查)

---

solver: structural
issue: 499
cluster: issue-499
verdict: propose

PROJECT_RULES clause violated (quoted verbatim)

• 事实源唯一:同一约束禁止在多处平行声明。版本号 → .version-bump.json;host 运行时事实 → host.env;skill 行为 → 该 skill 的 SKILL.md 与 scripts/test_*.py。

• 变更必须可验证:行为约束必须落到机械验证手段(behavior test / source-regression test / 段落 lint);仅靠"agent 应该记得"承载的约束视为未落地。

| Structured consumption | Steady-state controller/daemon paths consume only clean-exit status plus final allowlisted standalone marker/verdict lines, artifact frontmatter, CLI JSON/action fields, and artifact paths. Raw log prose is diagnostic-only. | Do not read, understand, quote, relay, or transcribe worker log prose, review reject prose, judge reasoning, or progress raw tails during normal routing. Raw logs are allowed only for EXIT!=0, stream disconnect/503, stuck/crash, missing/invalid structured artifact, router fallback, or worker self-post failure diagnostics. | structured-consumption boundary, phase routing details | wakeup_plan.py, phase9/router.py, monitors/progress.py |

4. Review verdict authority for merge-readiness starts from .refactor-loop/runs/review-pr<N>-<role>-r<R>.md frontmatter verdict: approve|comment|reject; only missing or invalid review artifacts fall back to clean log-tail REVIEW_DONE markers.

Recommended framing

把“clean-exit worker 终止证据”提升为 codex_refactor_loop.worker_markers 的脚本内共享读模型:它只读取同名 log、同名 runs artifact、review frontmatter 和 prompt/log 中的 head 信息,返回带 source/path/marker/verdict/head 的结构化 evidence。wakeup_plan.py 的 completed-marker detection、wakeup_runner.py 的 revalidation、implement_lifecycle.py 的 readiness、review gate/latest reviewer head 全部导入同一读模型,消除现在 detection/revalidation/readiness/review 各自解析同一事实导致的分叉;这不是新增 lifecycle authority,只是把已授权 #396 action projection 和 review truth table 的证据读取口收敛为单一事实源。

Concrete plan

• New abstractions (if any): WorkerCompletionEvidence + ReviewEvidenceReader + ImplementCompletionReader, Layer: skills/codex-refactor-loop/scripts/codex_refactor_loop, Project: codex-refactor-loop, File: worker_markers.py。接口:
  • read_implement_completion(repo_root, log_path) -> WorkerCompletionEvidence | None: 要求同名 implement-*.log 有 terminal EXIT=0;优先接受 log 中 standalone ok 终止 marker;否则读 .refactor-loop/runs/<same-basename>.md 尾部 standalone ok 终止 marker;返回 marker、log path、artifact path、source surface。
  • source_has_clean_marker(repo_root, source_artifact, marker) -> bool: 供 wakeup-runner revalidation 使用,同一规则判断 source evidence 是否满足 clean_exit_source_marker。
  • read_review_evidences(repo_root, pr_number) -> list[ReviewEvidence]: artifact-first,frontmatter verdict 为权威;artifact 缺失时 fallback 到 clean log-tail marker;head_sha 从 artifact/prompt/log 结构化字段读;重复/invalid verdict fail closed。
• Files:
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/worker_markers.py: 新增共享读模型,不含 git/gh/subprocess/lifecycle 操作。
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/wakeup_plan.py: 删除本地 implement artifact fallback、review head fallback 的重复 parser,改用 worker_markers 产出 completed-marker action 和 reviewer heads。
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/wakeup_runner.py: 删除 _source_log_has_clean_marker / review evidence 本地重复 parser,改用 worker_markers。
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/implement_lifecycle.py: classify_implement_attempt() 改从 read_implement_completion() 取 ok marker;artifact-only ok 不再返回 markerless。
  • skills/codex-refactor-loop/SKILL.md: 在 design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 dogfood anti-rules 后追加 headless marker-reading 系统性统一:worker marker 落 artifact 而非 log 致 detection/revalidation/readiness/review 多处误判 #499 局部合同:所有 clean-exit worker completion consumers use worker_markers.py;implement readiness 不得因 artifact-only marker 重派;review gate artifact-first remains unchanged。
  • skills/codex-refactor-loop/scripts/test_worker_markers.py: 新增 behavior tests。
  • skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py: source-regression 锁住 headless marker-reading 系统性统一:worker marker 落 artifact 而非 log 致 detection/revalidation/readiness/review 多处误判 #499 段落、四个 consumer 名称、artifact fallback 范围和 no lifecycle authority。
  • 现有测试补强: test_review_gate_e2e.py, test_wakeup_runner_review_gate.py, test_phase9_router_daemon.py 或 test_wakeup_plan.py 中相关 fixture 改/增 artifact-only 场景。
• LOC delta: ~+220 / -120。新增共享模块约 +170 LOC,新测试约 +170 LOC,三处 parser 删除/替换约 -120 LOC,SKILL source-regression 文本约 +35 LOC;净增约 +255 LOC。
• Tests to add:
  • test_worker_markers.py::test_implement_artifact_only_ok_is_publish_ready: log 只有 EXIT=0,同名 artifact 尾部有 standalone ok marker,classify_implement_attempt() 返回 publish_ready 而不是 redispatch:markerless。
  • test_worker_markers.py::test_implement_artifact_marker_requires_clean_exit: artifact 有 marker 但 log 无 terminal EXIT=0 或 EXIT=1,reader 返回 empty,readiness 可重派。
  • test_wakeup_plan.py::test_completed_marker_action_uses_artifact_source_for_markerless_log: completed-marker action 能从 artifact-only implement completion 生成 publish action,source_artifact 仍指向 log,并带 marker source metadata。
  • test_wakeup_runner.py::test_revalidation_accepts_artifact_only_implement_marker: runner 对同一 action 不再报 clean_exit_marker_missing。
  • test_review_gate_e2e.py::test_review_gate_accepts_artifact_frontmatter_when_log_marker_missing: 三 reviewer log 仅 EXIT=0,runs artifact frontmatter verdict/head 完整,truth table 可进入 merge/fix/wait 对应结果。
  • test_review_gate_e2e.py::test_review_gate_falls_back_to_log_only_when_artifact_missing: 保持 design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 compatibility。
  • test_skill_entrypoint_contract.py::test_worker_marker_reader_contract_issue499: 字面断言 SKILL.md 写明 detection/revalidation/readiness/review gate 都经共享 reader,且不改变 review truth table/merge gate。
• Schema/protocol changes (if any): none。marker family、action schema、review truth table、merge gate 均不变;只改变同一终止证据的读取路径。
• Governance/ruleset changes (if any): skills/codex-refactor-loop/SKILL.md 局部合同从“review verdict authority artifact-first”扩展为“worker completion evidence reader 是四个消费者的脚本事实源”。不改 CLAUDE.md,因为现有 事实源唯一、Structured consumption、design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 review artifact-first 已足够授权;这是 skill-local implementation contract 收口。deep consensus 可达,因为它不新增 controller authority、不改 marker semantics、不碰 host.env/production SSOT。
• Runtime cost: 每个 wakeup 对候选 log 增加至多 1 次同名 artifact tail read,每个文件 tail 约 80 行;100 个 log 时额外读取 <1 MB,本地延迟约 5-20 ms。review gate 每 PR 最多读取 3-9 个 artifact/log/prompt 小文件,与现状同量级;无网络、无 git/gh、无额外 actor hop、无 long-running allocation。新增 dataclass 对象约每 evidence 1 个,典型 <30 KB/wakeup。
• If the structural answer is issue decomposition, use the controller-private IssueDecompositionPlan boundary and child body artifacts; forbid solver/judge/worker GitHub lifecycle, public issue factories, and any wakeup_plan.py decompose action/projection.

Risks

• 相比最小补丁,共享 reader 会触碰 3 个 runtime 文件和 review gate tests,短期改动面更大;但问题已经发生在 4 个 consumer,继续局部 fallback 会制造第五套 parser。
• 过度抽象风险在于把所有 marker family 做成通用 registry。边界应收窄为 worker completion evidence reader,只覆盖现有 implement/review 两条链;solver/meta/router marker 先不迁入,除非后续出现同类 ≥2 consumer 分叉。
• artifact/log 冲突时必须 fail closed:同一 role/round 出现 duplicate marker、frontmatter invalid、artifact marker 与 verdict 不一致时不“猜”,返回 invalid evidence 让 truth table WAIT/REDISPATCH。

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why this structure beats a documented exception: issue 里 detection/revalidation 两个 hotfix 已经证明“给当前 caller 加 fallback”会漏 readiness/review gate;共享 reader 才能让新增 consumer 自动继承同一 clean-exit/artifact-fallback 语义。
• Why I picked this abstraction over alternatives: artifact-first 全局化会冲突 implement 的 clean-exit log scope;log-first 全局化会继续背离 design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 review artifact frontmatter authority。分成 implement completion reader 与 review evidence reader,再由 worker_markers.py 统一源文件,能保留两条语义差异但消除 parser 分叉。
• What I cannot decide alone: artifact/log 两边同时存在且 verdict/status 冲突时,我建议 fail closed 并暴露 invalid evidence;若 maintainer 想采用 artifact-overrides-log,需要 meta-judge 明确写入 truth table 兼容规则。

⟦AI:AUTO-LOOP⟧
SOLVER_DONE:structural:propose:worker_markers_shared_clean_exit_evidence_reader_for_detection_revalidation_readiness_review

⟦AI:AUTO-LOOP⟧

[loning]

🤖 delete solver 建议收敛 marker reader，删除重复 fallback

TL;DR

• 这是什么: headless marker-reading 系统性统一:worker marker 落 artifact 而非 log 致 detection/revalidation/readiness/review 多处误判 #499 的 delete-framing 评估，检查 markerless worker 是否该靠删除或合并现有路径解决。
• 结论: 不删除 marker robustness 功能；建议删掉分散在 detection/revalidation/readiness/review 的重复局部 reader，收敛到一个 clean EXIT=0 + 同名 artifact fallback 的共享读取器。
• 下一步: meta-judge 可把该方案与 minimal/structural solver 对齐，implementation 只改读取事实面，不放宽 review truth table 或 merge gate。

---

详细说明

delete 视角下，这不是“功能多余”，而是“同一能力被拆成多套私有判断”。当前 wakeup_plan.py 已有 implement artifact fallback，wakeup_runner.py 又有 revalidation fallback，implement_lifecycle.py 仍用 log-only clean_implement_done_marker 导致 readiness 把 artifact-only IMPLEMENT_DONE:*:ok 当 markerless，review gate 也有独立的 artifact/log verdict 解析路径。

建议的删除点是这些局部 fallback/helper，而不是删除 marker 检测本身：保留 #205 的 review artifact-frontmatter-first，implement/readiness/revalidation/detection 统一用一个 owner-local completion reader。它只接受 clean EXIT=0 绑定的 standalone marker，并且 artifact fallback 只能读 logs/<stem>.log 对应的 runs/<stem>.md；log 与 artifact 冲突时 fail closed。

---

📎 完整 codex 原始输出(存档备查)

---

solver: delete
issue: 499
cluster: issue-499
verdict: propose

Classification

d) Genuinely needed but over-built — marker reading is a real controller capability, but it is currently split across several local readers and hotfix fallbacks.

Recommended action

不要删除 marker robustness 这个能力；删除重复的局部 marker reader/fallback，把 detection、revalidation、consensus-implementation readiness、review gate 统一到一个 owner-local completion reader。该 reader 只接受 clean EXIT=0 绑定的 standalone 终止 marker，并只读取同名伴随 artifact：logs/<stem>.log 对应 runs/<stem>.md。review 继续沿用 #205 的 artifact-frontmatter-first 规则，implement 用同一 reader 的 log-first + artifact-fallback；如果 log 与 artifact 同时给出冲突 marker/verdict，fail closed，不放宽 truth table 或 merge gate。

Concrete plan (if propose)

• Files to delete: none as whole files.
• Functions/local branches to delete or collapse:
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/wakeup_plan.py: delete _implement_artifact_marker_fallback; replace marker_from_completed_log/_extract_completed_marker_line call sites with shared completion-reader calls; delete review-only marker validity helper duplication where it can use the same reader.
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/wakeup_runner.py: delete _source_log_has_clean_marker and _implement_run_artifact_has_marker; replace review evidence log/artifact marker parsing with shared reader primitives while keeping ReviewGateSnapshot truth-table logic unchanged.
  • skills/codex-refactor-loop/scripts/codex_refactor_loop/implement_lifecycle.py: delete log-only clean_implement_done_marker as the readiness source; use shared reader so marker-in-artifact IMPLEMENT_DONE:*:ok is not classified as markerless.
• Caller migrations:
  • completed_marker_actions(...) -> worker_completion.marker_for_completed_worker(repo_root, log_path); action source stays the log path, but marker may be recovered from the same-stem run artifact.
  • WakeupRunner._precondition_error(... clean_exit_source_marker ...) -> shared has_clean_source_marker(repo_root, log_path, marker).
  • classify_implement_attempt(...) -> shared implement_done_ok_marker(repo_root, log_path) before canonical worktree/branch/base/diff checks.
  • WakeupRunner._review_evidences(...) and latest_reviewer_heads(...) -> shared review_evidence_for(pr, role, round) that reads review-pr<N>-<role>-r<R>.md frontmatter first, falls back to clean log-tail REVIEW_DONE only when artifact is absent, and binds head SHA from artifact/prompt/log exactly as today.
• Tests to delete: no whole test files. Delete/replace tests that assert markerless clean implement logs are redispatchable, specifically the current behavior in test_consensus_implementation_readiness_redispatches_markerless_local_attempt and test_dispatch_consensus_implementation_redispatches_markerless_implement_log.
• Tests to add/retarget:
  • implement readiness: clean implement-issue-20.log with EXIT=0 and runs/implement-issue-20.md final IMPLEMENT_DONE:issue-20:ok plus valid worktree/branch/diff suppresses redispatch as implementation_ready_to_publish.
  • detection/revalidation symmetry: the same fixture creates one publish action and passes clean_exit_source_marker.
  • review detection: clean review-pr402-tests-r1.log with no REVIEW_DONE, plus runs/review-pr402-tests-r1.md frontmatter verdict: approve and final REVIEW_DONE:402:tests:approve, is eligible for review-gate projection.
  • review truth table: artifact frontmatter verdict with clean companion log but no log marker counts as present; conflicting log/artifact verdict fails closed as invalid evidence.
  • source-regression: wakeup_plan.py, wakeup_runner.py, and implement_lifecycle.py import the shared completion reader and do not define private duplicate artifact-fallback marker readers.
• LOC delta: -40 to -80 net expected after deleting duplicate helper branches and keeping the shared reader narrow.
• Philosophy/CLAUDE.md/SPEC/Tier changes (if any): none. The current structured-consumption boundary already allows clean EXIT=0, standalone marker/verdict lines, artifact frontmatter, and artifact paths. This plan only makes that boundary mechanically consistent.
• If deletion is unsafe because the issue is merely too large, you may propose a bounded IssueDecompositionPlan artifact path instead of deletion: not needed; the scope is bounded to one reader plus four call-site classes.

Reverse-evidence (why this is safe to delete)

• No public API breaks: verified by rg over skills/codex-refactor-loop/scripts; affected symbols are private helpers or module-local methods, not plugin manifest or public CLI surfaces.
• No external repo dependency on the code: $EXTERNAL_REPOS untouched; this is skill-private .refactor-loop controller logic.
• Tests covering the path are themselves not load-bearing when they assert the old bug: current tests explicitly expect markerless implement logs to redispatch, which contradicts issue headless marker-reading 系统性统一:worker marker 落 artifact 而非 log 致 detection/revalidation/readiness/review 多处误判 #499's dogfood evidence once the marker exists in the companion artifact.
• Existing design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 contract already says review verdict authority is review-pr<N>-<role>-r<R>.md frontmatter first, with clean log-tail REVIEW_DONE only as fallback. The proposed collapse aligns implement marker handling with that precedent instead of inventing a second review source.
• Existing hotfixes 8aeea89 and dfacc64 are evidence that the detection/revalidation fallback is needed but duplicated; deleting only one fallback would regress one side.

Risks

• If a run artifact can be written without the companion log's clean EXIT=0, artifact fallback would be unsafe. The plan avoids this by requiring clean log exit before reading the artifact.
• If log and artifact disagree, picking either side would be unsafe. The plan requires conflict to fail closed.
• If review artifacts sometimes omit frontmatter and only include prose, the shared reader must keep failing closed; it should not parse prose for verdicts.
• If implement partial is meant to preserve worktree state, this plan does not treat it as publish-ready; only IMPLEMENT_DONE:*:ok plus canonical worktree/branch/fresh-base/non-empty-diff suppresses redispatch.

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why I claim this can be deleted: the problem is caused by multiple private readers implementing overlapping "clean exit + marker + artifact" logic. wakeup_plan.py has marker_from_completed_log plus _implement_artifact_marker_fallback; wakeup_runner.py has _source_log_has_clean_marker plus _implement_run_artifact_has_marker; implement_lifecycle.py still uses log-only clean_implement_done_marker; review evidence parsing is another local path. These are replaceable with a single owner-local reader.
• What I checked to verify safety: rg for IMPLEMENT_DONE, REVIEW_DONE, completed_marker_actions, _source_log_has_clean_marker, consensus_implementation_suppressed_reason, review gate, and design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 source-regression text; read the relevant code in wakeup_plan.py, wakeup_runner.py, implement_lifecycle.py, and tests; checked git log showing 8aeea89 detection and dfacc64 revalidation hotfixes plus 57a6adc introducing the current log-only implement lifecycle classifier.
• What I cannot decide alone: exact helper/module name and whether implementation keeps the reader in a new worker_completion.py module or folds it into an existing module. The deletion requirement is the important part: one owner-local reader, no duplicate private fallbacks.

⟦AI:AUTO-LOOP⟧
SOLVER_DONE:delete:propose:collapse duplicate log/artifact marker readers into one clean-exit companion reader; delete local fallback branches that misclassify marker-in-artifact workers

⟦AI:AUTO-LOOP⟧

[loning]

🤖 meta-judge 判定：#499 已达成设计共识

TL;DR

• 这是什么: headless marker-reading 系统性统一:worker marker 落 artifact 而非 log 致 detection/revalidation/readiness/review 多处误判 #499 design-consensus round 1 的 meta-judge 仲裁结果。
• 现在到哪一步 / 结论是什么: 三个 solver 均同意用共享 worker marker reader 收口 detection / revalidation / readiness / review 的 marker 读取。
• 需要 maintainer 做什么 OR controller 下一步: maintainer 不需要决策；controller 可按 judge artifact 派 implement codex，并给 issue 加 crnd:triage:resume-requested。

---

详细说明

三份方案不是同名同量级完全一致，但 implementation-bearing 边界一致：保留 clean EXIT=0 与 standalone marker 语义，只把“marker 从哪里读”统一到一个 owner-local reader。implement 链路采用 log-first + 同名 .refactor-loop/runs/<stem>.md fallback；review gate 仍以 artifact frontmatter 的 verdict 为真值表权威，REVIEW_DONE 只作为 worker completion / routing evidence，不替代 verdict。

方案	仲裁取舍
minimal	给出最小接入面和测试清单，适合作为实现边界。
structural	给出共享模块命名 worker_markers.py 和结构化 evidence 方向，作为实现命名。
delete	明确删除重复局部 fallback，而不是删除 marker robustness 功能。

共识授权的实现范围只包含 worker_markers.py、wakeup_plan.py、wakeup_runner.py、implement_lifecycle.py、SKILL.md 和对应 behavior/source-regression tests；不改 review truth table、merge gate、consensus gate、CLAUDE.md、AGENTS.md 或任何 lifecycle authority。

---

📎 完整 codex 原始输出(存档备查)

---

issue: 499
cluster: issue-499
convergence_round: 1
solver_verdicts:
minimal: propose
structural: propose
delete: propose
decision: consensus

Decision

三份 solver 都给出同一 implementation-bearing framing:不要继续在 detection / revalidation / readiness / review gate 各自热修 marker 读取,而是新增一个 skill-private 共享 worker marker reader,把 clean EXIT=0、standalone terminal marker、同名 .refactor-loop/runs/<stem>.md companion artifact fallback、review frontmatter verdict authority 收口到同一脚本事实源。命名和 LOC 估算有差异,但不改变边界或行为语义;实现采用 structural 的 worker_markers.py 命名,并吸收 delete/minimal 的删除重复 fallback 要求。

If consensus

• Chosen framing: hybrid-minimal-structural-delete
• Implement plan (structured fields read by wakeup-plan from this judge artifact only, not from solver artifacts or prompt-body free text):
  • scope_paths:
    • skills/codex-refactor-loop/scripts/codex_refactor_loop/worker_markers.py
    • skills/codex-refactor-loop/scripts/codex_refactor_loop/wakeup_plan.py
    • skills/codex-refactor-loop/scripts/codex_refactor_loop/wakeup_runner.py
    • skills/codex-refactor-loop/scripts/codex_refactor_loop/implement_lifecycle.py
    • skills/codex-refactor-loop/SKILL.md
    • skills/codex-refactor-loop/scripts/test_worker_markers.py
    • skills/codex-refactor-loop/scripts/test_wakeup_plan.py
    • skills/codex-refactor-loop/scripts/test_wakeup_runner.py
    • skills/codex-refactor-loop/scripts/test_wakeup_runner_review_gate.py
    • skills/codex-refactor-loop/scripts/test_review_gate_e2e.py
    • skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py
  • old_pattern: detection、runner revalidation、implement readiness、review evidence/head 读取各自解析 log tail 或局部 artifact fallback,导致 worker marker 只落同名 run artifact 时被误判为 markerless/missing role。
  • new_principle: 所有 clean-exit worker terminal marker 消费点必须通过 codex_refactor_loop.worker_markers;implement 使用 log-first + same-stem companion artifact fallback 识别 IMPLEMENT_DONE:*:ok;review gate 保持 design: dogfood 运维经验写回 SKILL.md(并行audit覆盖/prompt注册/verdict源/daemon重启) #205 artifact-frontmatter-first verdict authority,REVIEW_DONE 只作为 clean worker completion / routing evidence;artifact fallback 必须绑定同名 log 且 terminal EXIT=0,只接受 standalone allowlisted marker,duplicate/malformed/conflicting evidence fail closed。
  • verification_hints: python3 -m pytest skills/codex-refactor-loop/scripts/test_worker_markers.py skills/codex-refactor-loop/scripts/test_wakeup_plan.py skills/codex-refactor-loop/scripts/test_wakeup_runner.py skills/codex-refactor-loop/scripts/test_wakeup_runner_review_gate.py skills/codex-refactor-loop/scripts/test_review_gate_e2e.py skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py
• Philosophy/CLAUDE.md/SPEC/Tier changes included: none; only skills/codex-refactor-loop/SKILL.md gets a skill-local contract/source-regression anchor for headless marker-reading 系统性统一:worker marker 落 artifact 而非 log 致 detection/revalidation/readiness/review 多处误判 #499 worker marker reading.
• Implementation owner: dispatch implement codex with cluster_id=issue-499, design_decision_path=.refactor-loop/runs/phase9-issue499-r1-judge.md
• Add crnd:triage:resume-requested label to issue 499

Round audit trail (links to local artifacts)

• solver-minimal: .refactor-loop/logs/phase9-issue499-r1-minimal.log
• solver-structural: .refactor-loop/logs/phase9-issue499-r1-structural.log
• solver-delete: .refactor-loop/logs/phase9-issue499-r1-delete.log
• work-unit-producer: manual-issue (prompt-only provenance)
• work-unit-source-ref: gh-issue-499

⟦AI:AUTO-LOOP⟧
META_JUDGE_DONE:consensus:hybrid-minimal-structural-delete:shared worker_markers reader for clean-exit same-stem artifact fallback across detection revalidation readiness and review

⟦AI:AUTO-LOOP⟧