[refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007)

[loning]

audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007)(cluster-008)

iter-2 audit 产出(由 issue #1 reflector re-cluster 拆分而来)。requires_design → Phase 9 共识。
⚠️ 分析/方案完全由 solver codex 产出;controller 只 framing + 监控。

完整规格(solver 必须读取)

本 cluster 的 evidence / fix boundary / human_brief 完整规格见仓库:
.refactor-loop/runs/audit-iter-2.md 的 cluster-008 节。solver 先读该节再提案。
背景见已关闭的 #1(通用化脊柱整体)与其 reflector re-cluster 分析。

⏸️ blocked-on #3

本 cluster 依赖 cluster-007 的 work-unit 契约定型;#3 共识达成后 controller 自动解锁(加 phase9-auto-solve + 🔍,移除 ⏸️)进 Phase 9。

📢 cc @loning

⟦AI:AUTO-LOOP⟧

[loning]

📊 当前状态 — 解锁 + Phase 9 r1(不需要人介入)

#3(007)work-unit 契约已共识 → 本簇(008 audit-producer 适配)解锁,携契约 + maintainer「别过度泛化」指令进 Phase 9 r1 三 solver。

🤖 controller status banner

⟦AI:AUTO-LOOP⟧

[loning]

🤖 minimal solver: 建议用 manual-issue 薄适配让 audit 成为默认 producer

TL;DR

• 这是什么: issue [refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007) #4 / cluster-008 的 minimal solver 独立方案。
• 结论: 违反点真实存在;最小方案是保留 audit 默认入口,新增 manual-issue 作为 WorkUnitV1 producer,不引入新 envelope 或 state 迁移。
• 下一步: meta-judge 对比 3 个 solver 输出;maintainer 暂不需要介入。

📢 cc 原作者:@loning 请 sanity-check 这个边界是否符合“别过度泛化,薄适配优先”。

---

详细说明

我核验了 README.md:72、SKILL.md Phase 0/1/7、prompts/audit.md 和 prompts/triage-external-issue.md: 当前启动路径仍强制 audit,Phase 1 只验收 audit-iter-N.md 与 candidate NDJSON,外部 triage 仍只接受 CLAUDE/AGENTS 生产代码违规并一刀切拒绝 docs/tooling-only。因此 cluster-008 的违反不是 stale evidence。

最小改动是 3 个文件、约 +70/-35 LOC: SKILL.md 把 Phase 1 改成 “work-unit production, audit default”,并说明手动 issue 经 triage 后也可进入;triage-external-issue.md 放宽为“具体仓库 work unit”,但继续拒绝产品需求、无实现提案的 bug、外部依赖和超大范围;REFERENCE.md 在 issue #3 的 WorkUnitV1 契约后补 Producers in v1,只列 audit 与 manual-issue。非 audit entry 写 work_unit_id/kind/producer/source_ref 到现有 clusters_planned,不伪造 cluster_id,也不迁移到 work_units_*。

---

📎 完整 codex 原始输出(存档备查)

---

solver: minimal
issue: 4
cluster: cluster-008-audit-producer-adapter
verdict: propose

Recommended framing

The violation is real, but the minimum viable boundary is producer intake wording plus one thin manual-work-unit path that writes the already-agreed WorkUnitV1 item shape into the existing authoritative clusters_planned queue. Keep audit as the default first producer and keep its current artifact validation. Do not introduce WorkUnitEnvelopeV1, work_units_*, a producer registry, or docs/spec/marketing producer families. The only new behavior is that triage can accept a maintainer-labeled issue that already contains a concrete repository work unit, normalize it as producer: manual-issue, and send it through the same Phase 9 spine without fabricating cluster_id.

Concrete plan (English)

• Files: skills/codex-refactor-loop/SKILL.md — change Phase 0 step 4 from "dispatch audit codex" to "dispatch the configured/default work-unit producer; v1 default is audit". Add one paragraph saying audit is the built-in producer, external/manual issues can also enter after triage, and both feed WorkUnitV1 records into clusters_planned.
• Files: skills/codex-refactor-loop/SKILL.md — rename the Phase 1 heading from Audit to Work-unit production (audit default) and keep the existing audit validation as the producer: audit subsection. After validation, say audit clusters are normalized as kind: audit-cluster, producer: audit, source_ref: .refactor-loop/runs/audit-iter-N.md#<cluster-id>.
• Files: skills/codex-refactor-loop/SKILL.md — update the Phase 7 external issue intake section. Path B should no longer ask only "violates CLAUDE/AGENTS?" It should ask whether the issue describes a concrete repository work unit with files/scope, non-product intent, and an implementation decision suitable for Phase 9. Accepted manual issues become WorkUnitV1 entries with kind: manual-work-unit, producer: manual-issue, source_ref: gh-issue-<N>, work_unit_id: issue-<N> or wu-issue-<N>, and no cluster_id / legacy_cluster_id.
• Files: skills/codex-refactor-loop/prompts/triage-external-issue.md — relax accept criteria from "CLAUDE/AGENTS production-code violation only" to "concrete repo work unit suitable for consensus"; still reject product feature requests, runtime bug reports without an implementation proposal, external dependency requests, unclear issues, and scope >50 files. Remove the docs-only/tooling-only blanket reject; accept docs/spec/tooling only when they cite concrete repo files and require a codebase-owned change. Keep evidence and human_brief requirements.
• Files: skills/codex-refactor-loop/prompts/triage-external-issue.md — change accepted body shaping from "refactor cluster" to "WorkUnitV1-backed design issue"; require work_unit_id, kind, producer, source_ref, scope_paths, old_pattern or problem statement, new_principle or desired invariant, verification_hints, and optional audit compatibility fields only if the source is audit.
• Files: skills/codex-refactor-loop/REFERENCE.md — after issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3's WorkUnitV1 contract lands, add a short Producers in v1 paragraph: audit and manual-issue are the only supported v1 producers; both write WorkUnitV1 items into clusters_planned; non-audit producers must not invent cluster_id or depend on audit candidate NDJSON.
• LOC delta: ~+70 / -35 across 3 files, assuming issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3's WorkUnitV1 text already exists. Without issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 implemented first, add ~+10 LOC to reference the agreed contract by name but do not duplicate it.
• Tests to add/modify: no new executable tests; this repo has no skill test harness and the change is prompt/SKILL contract text. Add grep verification in the implementation summary: rg -n "manual-work-unit|manual-issue|Producers in v1|producer: audit|producer: manual-issue|work_unit_id: issue-" skills/codex-refactor-loop/{SKILL.md,REFERENCE.md,prompts/triage-external-issue.md}; rg -n "不是 docs-only 或 tooling-only|属于本 refactor loop 范畴\\(违反 CLAUDE/AGENTS 条款\\)|Phase 1 — Audit \\(one codex" skills/codex-refactor-loop should no longer find the old exclusive wording.
• Philosophy/CLAUDE.md/SPEC/Tier change: none. This implements the README generalization route inside the skill contract and stays under issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3's v1 decision: clusters_* remains authoritative, WorkUnitV1 is the item contract, and WORK_UNIT_ID is the prompt identity. No top-level CLAUDE/AGENTS rule needs a new exception, no SPEC.md / conformance/ / trusted_base.lock is touched, and no GPG or reinstall step is required. Deep consensus should hold because the plan follows the maintainer instruction: thin adapter first, no over-general producer framework.
• Migration path: no migration needed. Existing audit loops continue to start with audit and validate audit-iter-N.md plus candidate NDJSON. New manual issues only enter through triage and use WorkUnitV1 fields in the existing queue.

Concrete plan (中文)

• 文件: skills/codex-refactor-loop/SKILL.md — 把 Phase 0 第 4 步从“派 audit codex”改为“派 configured/default work-unit producer; v1 默认是 audit”。新增一段说明 audit 是内置 producer,外部/手动 issue 经过 triage 后也可以进入,两者都写 WorkUnitV1 record 到 clusters_planned。
• 文件: skills/codex-refactor-loop/SKILL.md — 把 Phase 1 标题从 Audit 改为 Work-unit production (audit default),保留现有 audit 校验作为 producer: audit 子段。校验通过后说明 audit cluster 会规范化为 kind: audit-cluster, producer: audit, source_ref: .refactor-loop/runs/audit-iter-N.md#<cluster-id>。
• 文件: skills/codex-refactor-loop/SKILL.md — 更新 Phase 7 外部 issue 接入。Path B 不再只问“是否违反 CLAUDE/AGENTS”,而是判断 issue 是否描述了一个具体仓库 work unit:有文件/范围、不是产品需求本身、且需要 Phase 9 做实现决策。接受的手动 issue 生成 WorkUnitV1 entry: kind: manual-work-unit, producer: manual-issue, source_ref: gh-issue-<N>, work_unit_id: issue-<N> 或 wu-issue-<N>,且不写 cluster_id / legacy_cluster_id。
• 文件: skills/codex-refactor-loop/prompts/triage-external-issue.md — 把 accept 标准从“只能是 CLAUDE/AGENTS production-code violation”放宽为“适合共识流程的具体仓库 work unit”;仍拒绝产品 feature request、没有实现提案的 runtime bug report、外部依赖请求、不清晰 issue、超过 50 文件的范围。删除 docs-only/tooling-only 的一刀切拒绝;docs/spec/tooling 只有在引用具体仓库文件且需要仓库内改动时才接受。保留 evidence 和 human_brief 要求。
• 文件: skills/codex-refactor-loop/prompts/triage-external-issue.md — 把接受后的 body reshape 从 “refactor cluster” 改成 “WorkUnitV1-backed design issue”;要求写出 work_unit_id, kind, producer, source_ref, scope_paths, old_pattern 或问题陈述,new_principle 或目标不变量,verification_hints;只有来源是 audit 时才允许 audit 兼容字段。
• 文件: skills/codex-refactor-loop/REFERENCE.md — 等 issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 的 WorkUnitV1 契约落地后,新增简短 Producers in v1 段落: v1 只支持 audit 和 manual-issue 两个 producer;二者都把 WorkUnitV1 item 写入 clusters_planned;非 audit producer 禁止伪造 cluster_id,也不能依赖 audit candidate NDJSON。
• LOC delta: 预计约 +70 / -35,共 3 个文件,前提是 issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 的 WorkUnitV1 文本已先落地。若 issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 尚未实现,只额外加约 +10 LOC 引用已共识的契约名称,不要复制整套契约。
• 需要新增/修改的测试: 不新增可执行测试;本仓库没有 skill test harness,本改动是 prompt/SKILL 契约文本。实现总结里加入 grep 验证: rg -n "manual-work-unit|manual-issue|Producers in v1|producer: audit|producer: manual-issue|work_unit_id: issue-" skills/codex-refactor-loop/{SKILL.md,REFERENCE.md,prompts/triage-external-issue.md}; rg -n "不是 docs-only 或 tooling-only|属于本 refactor loop 范畴\\(违反 CLAUDE/AGENTS 条款\\)|Phase 1 — Audit \\(one codex" skills/codex-refactor-loop 不应再找到旧的唯一入口措辞。
• Philosophy/CLAUDE.md/SPEC/Tier change: 无。该方案在 skill 契约内部实现 README 的泛化路线,并遵守 issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 的 v1 决策:clusters_* 仍是权威队列,WorkUnitV1 是 item 契约,WORK_UNIT_ID 是 prompt 身份。不需要新增顶层 CLAUDE/AGENTS 例外,不触碰 SPEC.md / conformance/ / trusted_base.lock,也不需要 GPG 或 reinstall。深度共识应可达,因为方案符合 maintainer 指令:先做薄适配,不搭过度泛化的 producer 框架。
• Migration path: 不需要迁移。现有 audit loop 继续默认从 audit 开始,继续校验 audit-iter-N.md 和 candidate NDJSON。新的手动 issue 只通过 triage 进入,并在现有队列里使用 WorkUnitV1 字段。

Risks

• The persisted queue is still named clusters_planned, so manual work units will live inside a legacy-named container until a later v2.
• Manual issue triage can admit vague work if the prompt wording is too loose; keep the concrete file/scope and verification hints requirements mandatory.
• Docs/spec/tooling acceptance is deliberately narrow. Pure prose requests with no repository-owned implementation decision still reject.
• Audit remains the default bootstrap producer, so this is adapter support rather than a full producer plugin system.

Escalation triggers (if any)

• none

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

• Why this is the minimum: The cited violation is that audit is the only producer and triage rejects non-refactor repository work. Issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 already supplies the WorkUnitV1 item contract, so the smallest fix is to let one non-audit producer write that item shape and to document audit as the default producer.
• What I considered but rejected: A generic producer registry, per-kind docs/spec/tooling producers, WorkUnitEnvelopeV1, work_units_*, and a normalizer helper. All add abstraction before a second implemented producer family exists and conflict with the maintainer's thin-adapter instruction.
• What I cannot decide alone: Whether future v2 should split manual-issue into separate docs-task, spec-change, and tooling-task producers. That should wait until real accepted issues show distinct validation needs.

SOLVER_DONE:minimal:propose:add manual-issue WorkUnitV1 intake and make audit the default producer, not the only producer
⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 structural solver 提案:用 WorkUnitProducerV1 薄适配 audit/triage

TL;DR

• 这是什么: issue [refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007) #4 的 structural solver 方案, 基于 [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 已定的 clusters_* + WorkUnitV1 v1 契约。
• 结论: 建议新增 WorkUnitProducerV1(producer 输出 WorkUnitV1 的标准合同), audit 和 external-issue triage 两个入口走同一薄 adapter。
• 下一步: meta-judge 对比 3 个 solver 输出; maintainer 暂不需要介入。

📢 cc 原作者:@loning 请 sanity-check 这个边界是否足够薄,且没有越过 #3 的 v1 契约。

---

详细说明

违反点是 README.md:72 要求 audit 可替换为任意 work-unit 来源,但 SKILL.md:573-585 首轮强制派 audit,SKILL.md:645-676 只验证 audit artifact 后写 clusters_planned,triage-external-issue.md:19-24 又把 docs/tooling/spec 类仓库工作拒掉。我的方案不迁移 state、不引入 WorkUnitEnvelopeV1: 只在 producer 边界新增 WorkUnitProducerV1,要求 producer 输出 #3 定下的 WorkUnitV1 item。

落地边界是 7 个文件,估算 +330/-75 LOC: REFERENCE.md 写 producer 合同,SKILL.md 把 Phase 1 表述为 producer intake,audit prompt 补齐 WorkUnitV1 字段,triage prompt 放宽到“有具体路径 + maintainer scope + verification hints”的 docs/spec/tooling 工作,新增 work_unit_producer.py 做 schema validation/projection。运行成本是每个 producer 结果一次本地校验,约 20-50 ms,无额外 GitHub hop、无新 daemon。

---

📎 完整 codex 原始输出(存档备查)

---

solver: structural
issue: 4
cluster: cluster-008-audit-producer-adapter
verdict: propose

CLAUDE clause violated (quoted verbatim)

定位与共识构建引擎的设计哲学见 README.md。

Delegated README invariant:

1. 抽出引擎脊柱(solve → consensus → implement → verify),让 audit 这一步(产生工作单元的种子)可替换 —— 换成任意 work-unit 来源(设计提案 / 文档任务 / 营销产出 / spec 变更),其余整套共识机制原样复用。

Recommended framing (English)

Add a producer-side contract, WorkUnitProducerV1, and make audit one adapter that emits WorkUnitV1 items into the existing v1 authoritative clusters_* queues. The second concrete caller is the existing external-issue triage path: it should accept maintainer-scoped repository work such as docs/spec/tooling changes when it can produce a complete WorkUnitV1, without fabricating cluster_id. This eliminates the violation by construction because the engine spine consumes only WorkUnitV1 records; audit-specific candidate files, AUDIT_DONE, and legacy_cluster_id stay inside the audit producer adapter.

Recommended framing (中文)

新增 producer 侧契约 WorkUnitProducerV1,并把 audit 改成一个向现有 v1 权威 clusters_* 队列写入 WorkUnitV1 item 的 adapter。第二个具体调用方是已经存在的外部 issue triage 路径:当 maintainer 标记的 docs/spec/tooling 等仓库工作能产出完整 WorkUnitV1 时,它可以接入,且不得伪造 cluster_id。这样结构上消除违反点:引擎脊柱只消费 WorkUnitV1 记录;audit 专有的 candidate 文件、AUDIT_DONE、legacy_cluster_id 都被限制在 audit producer adapter 内。

Concrete plan

• New abstractions (if any): WorkUnitProducerV1, a documented producer output contract in the Skill contract layer, project skills/codex-refactor-loop, documented in skills/codex-refactor-loop/REFERENCE.md; and a thin work_unit_producer.py adapter/validator in skills/codex-refactor-loop/scripts/. Interface: producer output contains producer, items[]; each item is the [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 WorkUnitV1 shape with work_unit_schema_version: 1, work_unit_id, kind, producer, source_ref, scope_paths, old_pattern, new_principle, verification_hints, dependencies, risk, leverage, optional legacy_cluster_id. Non-audit producers must omit cluster_id/legacy_cluster_id.
• Existing repo primitives that apply: .refactor-loop/state.json remains the durable queue; clusters_planned, clusters_active, clusters_done, and clusters_failed remain authoritative v1 containers; prompt materialization under .refactor-loop/prompts/ is the actor inbox for audit/triage/implement/verify codexes; GitHub issues are the external event envelope; audit/triage output to clusters_planned is the projection pipeline. No IAsyncEnumerable, Channel, or in-process actor inbox exists in this repo, so adding those would be architecture cosplay rather than leverage.
• Files:
  • skills/codex-refactor-loop/REFERENCE.md: add WorkUnitProducerV1 after the WorkUnitV1 contract; define two v1 producers: audit and external-issue; state that all producer outputs are projected into existing clusters_* queues, not work_units_*.
  • skills/codex-refactor-loop/SKILL.md: rename Phase 1 wording to "Producer intake"; first wakeup still dispatches audit by default; validation calls the producer adapter after existing audit checks; accepted producer items populate clusters_planned as WorkUnitV1; non-audit items enter without cluster_id.
  • skills/codex-refactor-loop/prompts/audit.md: require accepted cluster YAML to include WorkUnitV1-compatible fields (work_unit_id, kind: audit-cluster, producer: audit, source_ref, legacy_cluster_id), while preserving current audit artifacts and AUDIT_DONE for compatibility.
  • skills/codex-refactor-loop/prompts/triage-external-issue.md: change accept criteria from "CLAUDE/AGENTS refactor violation only" to "repo-concrete work unit with files, governing host policy or maintainer scope, and verification hints"; accept docs/spec/tooling work only when it has concrete paths and checks; continue rejecting product feature requests, runtime bug reports, external dependency work, unclear issues, and scope >50 files.
  • skills/codex-refactor-loop/prompts/design-issue-body.md: include a work_unit block as durable routing metadata; keep cluster_id only when the source producer is audit.
  • skills/codex-refactor-loop/scripts/work_unit_producer.py: new narrow adapter/validator; reads audit cluster sections or standardized triage issue bodies and emits canonical WorkUnitV1 JSON records for insertion into clusters_planned.
  • skills/codex-refactor-loop/scripts/tests/test_work_unit_producer.py: new behavior tests for producer normalization and rejection rules.
• LOC delta: ~+330 / -75 across 7 files. Split estimate: REFERENCE.md +55/-5, SKILL.md +65/-25, audit.md +25/-8, triage-external-issue.md +55/-30, design-issue-body.md +15/-7, new adapter +75, new tests +140.
• Tests to add:
  • test_audit_cluster_normalizes_to_work_unit_v1: audit cluster input becomes a WorkUnitV1 record with producer: audit, kind: audit-cluster, source_ref pointing to audit-iter-N.md#cluster-id, and legacy_cluster_id; target queue remains clusters_planned.
  • test_external_issue_can_emit_non_audit_work_unit: a maintainer-scoped docs/spec/tooling issue with concrete paths and verification hints becomes producer: external-issue, kind != audit-cluster, and has no cluster_id or legacy_cluster_id.
  • test_external_issue_rejects_feature_or_bug_without_repo_contract: product feature requests, runtime bug reports, external dependency requests, and pathless docs/tooling reports fail validation with a reject category.
  • test_v1_does_not_introduce_envelope_or_work_units_queues: adapter output and docs use WorkUnitV1 items in clusters_*; no WorkUnitEnvelopeV1 or work_units_* authority appears.
• proto changes (if any): none.
• Philosophy/CLAUDE.md/SPEC/Tier changes (if any): none. Current invariant is already delegated from CLAUDE.md to README.md, and issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 has fixed the v1 contract as clusters_* plus WorkUnitV1. This plan implements that invariant inside the skill contract without changing CLAUDE.md, AGENTS.md, Tier boundaries, SPEC.md, conformance, package identity, public labels, branch names, or markers. Deep consensus should be reachable because it honors the maintainer instruction: one thin adapter layer, no state migration, no over-generalized envelope.
• Runtime cost: audit path adds one local adapter/validator process after existing audit validation, about 20-50 ms Python startup plus O(number of emitted items) JSON parsing; for 30 items, expected memory allocation is <150 KB and wall time <100 ms on a laptop. External issue path reuses the existing triage codex and adds the same local validation cost only on accepted issues. No extra GitHub/network hop, no extra background daemon, and no added latency in implement/verify after items are queued.

Risks

• This keeps clusters_* as the visible state container for non-audit work units, which is semantically awkward but required by [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3's v1 contract.
• Adding work_unit_producer.py could grow into a hidden engine. Keep it bounded to schema validation and projection only: no scheduling, no consensus decisions, no label changes, no marker migration.
• Broadening triage can accidentally admit feature requests. The guardrail is concrete: non-audit acceptance requires maintainer scope, repo paths, expected artifact kind, and verification hints; otherwise reject.
• Audit remains the default first-wakeup producer, so a host-selected producer switch still needs explicit configuration wording or a maintainer-labeled issue path.

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why this structure beats a documented exception: A documented exception would say "audit clusters count as work units" while leaving every intake path audit-shaped. WorkUnitProducerV1 makes the boundary explicit: producers emit the [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 item contract, and the engine consumes that item contract from the legacy v1 queue.
• Why I picked this abstraction over alternatives: It has two concrete callers now, audit and external-issue triage, plus the README extension point. WorkUnitEnvelopeV1, work_units_*, Channel, and actor runtime types are cleaner in a different system but contradict the [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 consensus and add migration cost before evidence demands it.
• What I cannot decide alone: whether future producers get configured by a host env var, a maintainer-labeled issue, or a new prompt template registry. This issue should only define the producer contract and land audit/external-issue adapters.

SOLVER_DONE:structural:propose:define WorkUnitProducerV1 plus thin audit/external-issue adapters into existing WorkUnitV1 clusters_* queues
⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 delete solver — 建议删除 producer 框架范围,只做薄适配

TL;DR

• 这是什么:issue [refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007) #4 的 delete solver 结论,专门质疑是否需要新增抽象。
• 结论:保留 audit producer,但删除 producer registry / 新 envelope / 多类 producer taxonomy 的范围。
• 下一步:meta-judge 可把本方案与 minimal/structural 对齐,优先选择现有 audit + triage/manual issue 的 WorkUnitV1 薄适配。

📢 cc 原作者:@loning 请 sanity-check triage/manual issue 作为第二个 v1 producer 是否符合 maintainer 的“薄适配优先”。

---

详细说明

我不建议删除 audit 本身。SKILL.md:645-676 仍把 audit 作为首次自动发现 work unit 的真实入口,删掉会让 first wakeup 失去自动 cluster 发现能力。README.md:72 要求 audit 可替换,但不要求 v1 先造 producer framework。

建议删除的是过度范围:不新增 producer registry,不引入 WorkUnitEnvelopeV1,不引入 work_units_*,不拆 docs/spec/tooling 专用 producer 类,不改 Phase 9 solver 语义。现有 auto-loop-triage 路径已经能把外部 issue reshape 后送入 Phase 9,所以它可以作为第二个 producer:接受条件改成“能表达为 repo-local、scoped、带 evidence/source_ref/verification hints 的 WorkUnitV1 item”,而不是继续硬拒 docs-only/tooling-only。

这样 #4 只需要把 audit 输出和 triage/manual issue 输出都适配到 #3 已共识的 v1 契约:clusters_* 仍是权威队列,队列元素是 WorkUnitV1,producer 字段区分 audit 与 triage/manual-issue。

---

📎 完整 codex 原始输出(存档备查)

---

solver: delete
issue: 4
cluster: cluster-008-audit-producer-adapter
verdict: propose

Classification

(d) Genuinely needed but over-built. The current audit path is load-bearing, and README requires audit to become a replaceable work-unit source. Deleting producer adaptation entirely would delete the only automatic discovery path and contradict the already accepted issue #3 v1 contract. The over-built part is any new producer registry, new producer framework, new envelope, docs/spec/tooling-specific producer taxonomy, or state migration. Existing audit and external triage already form the two practical intake paths; they only need to emit or describe WorkUnitV1 items in the existing authoritative clusters_* queues.

Recommended action (English)

Keep audit as a first-class producer, but collapse issue #4 to a thin adapter: audit output is normalized into WorkUnitV1 items with producer: "audit" and kind: "audit-cluster" before entering clusters_planned; external triage/manual issue intake becomes the second producer by reshaping accepted issues into the same WorkUnitV1-compatible design issue shape with producer: "triage" or producer: "manual-issue". Delete the broader scope: no new producer registry, no WorkUnitEnvelopeV1, no work_units_*, no separate docs/spec/tooling producer classes, no Phase 9 solver changes, and no new daemon beyond the existing triage monitor.

Recommended action (中文)

保留 audit 作为一等 producer,但把 issue #4 压成薄适配:audit 产物进入 clusters_planned 前归一为 WorkUnitV1 item,字段使用 producer: "audit" 与 kind: "audit-cluster";外部 triage/手动 issue 接入作为第二个 producer,把接受的 issue reshape 成同一套 WorkUnitV1-compatible design issue 形态,字段使用 producer: "triage" 或 producer: "manual-issue"。删除更大的范围:不新增 producer registry,不引入 WorkUnitEnvelopeV1,不引入 work_units_*,不拆 docs/spec/tooling 专用 producer 类,不改 Phase 9 solver 语义,也不新增现有 triage monitor 之外的 daemon。

Concrete plan (if propose)

• Files to delete: none from the current repo; delete the proposed new producer-framework surface from the implementation scope.
• Caller migrations:
  • skills/codex-refactor-loop/SKILL.md:645-676 Phase 1 audit validation -> keep the validation, then explicitly map accepted audit clusters to WorkUnitV1 items stored in clusters_planned.
  • skills/codex-refactor-loop/SKILL.md:1128-1162 external issue intake -> describe Path A/Path B as producer intake into the same WorkUnitV1 item contract, not as a separate refactor-only side path.
  • skills/codex-refactor-loop/prompts/triage-external-issue.md:19-24 accept criteria -> delete the automatic docs-only / tooling-only rejection as a category gate, and replace it with a narrower rule: accept repo-local work that can be expressed as a scoped WorkUnitV1 item with evidence, source_ref, scope_paths, and verification hints; still reject product feature requests, runtime bug reports, external dependencies, unclear tasks, and oversize scope.
  • skills/codex-refactor-loop/prompts/audit.md -> add only the WorkUnitV1 producer metadata expectation; keep AUDIT_DONE and audit artifact validation as compatibility.
  • skills/codex-refactor-loop/REFERENCE.md -> document the two v1 producers: audit and triage/manual-issue; state that new producers must write WorkUnitV1 items into authoritative clusters_* queues and must not create a second queue family.
• Tests to delete: none; do not add producer-registry tests. Verification should be grep-level conformance for absence of WorkUnitEnvelopeV1, work_units_*, and producer registry wording, plus presence of producer: "audit" / producer: "triage" guidance.
• LOC delta: -120 to -220 versus a multi-producer framework plan; likely net implementation is about +35 / -25 across 4 prompt/docs files.
• Philosophy/CLAUDE.md/SPEC/Tier changes (if any): none. README.md:72 already says audit should become replaceable; issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 already fixed the v1 state contract by keeping clusters_* authoritative. This plan follows both by making producer identity an item-level metadata field instead of a new architecture vocabulary or Tier boundary. Consensus can hold because it satisfies the maintainer directive "thin adapter first" and does not require GPG ratification, SPEC changes, conformance trusted-base edits, or package renaming.
• Tracking issue (if defer): none. This is not a deferral plan; it deletes over-built scope now and accepts the thin adapter now.

Reverse-evidence (why this is safe to delete)

• No public API breaks: shipped callers still route through clusters_planned, AUDIT_DONE, audit artifacts, GitHub issue labels, and Phase 9 design issues. The proposal preserves those surfaces.
• No external repo dependency on the code: the deleted scope is unimplemented design surface, not a published producer API. Existing installed skill users still see audit and triage behavior.
• Tests covering the path are themselves not load-bearing: there are no shipped tests for producer registry, docs/spec/tooling producer classes, work_units_*, or WorkUnitEnvelopeV1; those should not be added for issue [refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007) #4.
• README.md:72 requires replaceable work-unit sources, but it does not require a general producer framework in v1.
• SKILL.md:645-676 proves audit is the live automatic producer; deleting audit would remove first-wakeup discovery and AUDIT_DONE routing.
• SKILL.md:1128-1162 and prompts/triage-external-issue.md:36-44 prove an existing manual/external intake path already reshapes issues into Phase 9; this is the simplest second producer.
• Git history shows the current audit/cluster state came from the initial verbatim migration commit 1722f1f, whose commit message already named future generalization as "让 audit work-unit 来源可替换". That supports adapting the existing intake paths rather than replacing them with a new framework.

Risks

• If implementers leave the docs-only / tooling-only hard reject in triage, the second producer remains narrower than README's promised work-unit source model.
• If implementers broaden triage too far, feature requests and runtime bug reports may enter Phase 9. The acceptance rule must require repo-local evidence, scoped files, source_ref, and verification hints.
• The queue names clusters_* remain misleading. Issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 deliberately accepted that compatibility cost, so [refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007) #4 must repeat the invariant: queue names are legacy; WorkUnitV1 item fields are the v1 protocol.
• A future third producer may need richer metadata. That should extend WorkUnitV1 only when there is a concrete producer and caller, not in this cluster.

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why I claim this can be deleted: the extra framework is not needed to satisfy the value chain. Audit already produces work; triage/manual issue intake already exists as an issue-to-Phase-9 path. The missing piece is item metadata and acceptance wording, not a new dispatch spine.
• What I checked to verify safety: issue [refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007) #4 body and comment, .refactor-loop/runs/audit-iter-2.md cluster-008, README generalization route, CLAUDE/AGENTS repo constraints, issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 r3 consensus, SKILL.md Phase 0/1 audit path, external issue intake, prompts/audit.md, prompts/triage-external-issue.md, REFERENCE.md, grep call sites for clusters_planned, AUDIT_DONE, auto-loop-triage, WorkUnitV1, WorkUnitEnvelope, and git history for clusters_planned / AUDIT_DONE.
• What I cannot decide alone: the exact future vocabulary for producer kinds beyond audit and triage/manual-issue. That is not necessary for issue [refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007) #4 because v1 only needs these two existing intake paths.

SOLVER_DONE:delete:propose:collapse audit-producer adaptation to existing audit plus triage/manual WorkUnitV1 producers; delete new producer framework scope
⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

📊 当前状态 — Phase 9 r1 meta-judge 仲裁中(不需要人介入)

🤖 controller status banner

⟦AI:AUTO-LOOP⟧

[loning]

🤖 meta-judge: issue #4 第 1 轮未达成 3/3 framing,进入 round 2 收敛

TL;DR

• 这是什么: issue [refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007) #4 / cluster-008 的 Phase 9 meta-judge 仲裁结果。
• 结论: 3 个 solver 都 propose,方向一致,但 WorkUnitProducerV1 + script/tests 是否入 v1 仍分歧,不能判 consensus。
• 下一步: controller 进入 round 2,只问 exact files、producer 命名、verification scope; maintainer 暂不需要介入。

---

详细说明

3 个方案的共同点已经很清楚:继续用 clusters_* 作为权威队列,让 audit 默认产出 WorkUnitV1,让 external/manual triage 作为第二入口产出非 audit 的 WorkUnitV1,同时明确不做 WorkUnitEnvelopeV1、work_units_*、producer registry、docs/spec/tooling 专用 producer 类、Phase 9 solver 语义变化和 state migration。

不能直接判 consensus 的原因是范围差异已经超过“同 framing”:minimal 是 3 个文件的 prompt/docs 薄适配; delete 是 4 个文件的薄适配并保留 triage/manual-issue 命名空间; structural 是 7 个文件,新增 WorkUnitProducerV1(producer 输出 WorkUnitV1 的命名合同)、work_unit_producer.py 和测试。round 2 只收敛这个差异。

方案	文件/LOC 估算	关键取舍	仲裁结果
minimal	3 文件,约 +70/-35	最少 prompt/docs 改动,用 manual-issue	可行但未获 3/3
structural	7 文件,约 +330/-75	新增 WorkUnitProducerV1、script、tests	抽象/文件范围未获 3/3
delete	4 文件,约 +35/-25	删除框架范围,用 audit + triage/manual 薄适配	可行但命名/文件范围未定

round 2 的具体问题:issue #4 应该落成 prompt/docs-only 薄适配,还是命名的 WorkUnitProducerV1 + work_unit_producer.py + tests 合同;请给出 exact files、唯一 producer vocabulary、以及 verification scope。

---

📎 完整 codex 原始输出(存档备查)

---

issue: 4
cluster: cluster-008-audit-producer-adapter
convergence_round: 1
solver_verdicts:
minimal: propose
structural: propose
delete: propose
decision: converge

Decision (English)

Convergence round needed. All 3 solvers agree on the main invariant: keep clusters_* authoritative, keep audit as the default producer, normalize audit into WorkUnitV1, allow external/manual triage to emit non-audit WorkUnitV1 items, reject WorkUnitEnvelopeV1, work_units_*, producer registries, docs/spec/tooling producer taxonomies, Phase 9 semantic changes, and state migration. But they do not yet meet the 3/3 same-framing gate: minimal proposes a 3-file prompt/docs change with manual-issue; delete proposes a 4-file thin adapter with triage/manual-issue and no new framework; structural proposes 7 files, a named WorkUnitProducerV1 contract, work_unit_producer.py, and tests. That is a real abstraction/file/LOC split, plus unresolved producer vocabulary, so consensus cannot be declared in round 1.

Decision (中文)

需要收敛下一轮。3 个 solver 对核心不变量已经一致: clusters_* 继续作为权威队列,audit 继续是默认 producer,audit 输出归一成 WorkUnitV1,外部/manual triage 可以生成非 audit 的 WorkUnitV1,并且不引入 WorkUnitEnvelopeV1、work_units_*、producer registry、docs/spec/tooling 专用 producer 分类、Phase 9 语义变化或 state migration。但这还没有达到 3/3 同 framing: minimal 是 3 个文件的 prompt/docs 改动并使用 manual-issue; delete 是 4 个文件的薄适配并使用 triage/manual-issue,不新增框架; structural 是 7 个文件,新增命名契约 WorkUnitProducerV1、work_unit_producer.py 和测试。这是实际的抽象层/文件范围/LOC 差异,同时 producer 命名也未定,所以第 1 轮不能判定 consensus。

If consensus

• Chosen framing: n/a
• Implement plan (verbatim copy from the winning solver's "Concrete plan" section): n/a
• Philosophy/CLAUDE.md/SPEC/Tier changes included: n/a
• Implementation owner: n/a
• Add auto-loop-resume label to issue 4: n/a

If converge

• Convergence question (specific): Should issue [refactor-design] cluster-008: audit 作为 work-unit producer 适配(脊柱 2/4,依赖 007) #4 land as a prompt/docs-only thin adapter over existing audit plus triage/manual issue intake, or as a named WorkUnitProducerV1 contract with work_unit_producer.py and tests; answer with exact files, one producer vocabulary, and verification scope.
• What each solver should address explicitly:
  • Whether WorkUnitProducerV1 is needed in v1, or whether WorkUnitV1 plus a Producers in v1 docs section is the entire contract.
  • Whether any new script/test files are required now; if yes, name the exact files and why grep/prompt conformance is insufficient.
  • The exact non-audit producer name: choose one of manual-issue, external-issue, or triage/manual-issue, and state whether aliases are allowed.
  • The exact file set: decide whether prompts/audit.md and prompts/design-issue-body.md are in scope, or whether SKILL.md, REFERENCE.md, and prompts/triage-external-issue.md are sufficient.
  • The agreed no-go list: no WorkUnitEnvelopeV1, no work_units_*, no producer registry, no docs/spec/tooling-specific producer classes, no Phase 9 solver semantics change, no state migration.
• Round number this fires: 2

If escalate

• Trigger category: n/a
• Why consensus failed to progress: n/a
• Suggested next step: n/a

Round audit trail (links to local artifacts)

• solver-minimal: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue4-r1-minimal.md
• solver-structural: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue4-r1-structural.md
• solver-delete: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue4-r1-delete.md

META_JUDGE_DONE:converge:round-2:Should issue #4 use a prompt/docs-only thin adapter or a named WorkUnitProducerV1 script-backed contract; decide exact files, producer vocabulary, and verification scope.
⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧