[refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4)

[loning]

work-unit 契约 schema(通用化脊柱拆分 1/4)(cluster-007)

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-007 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。

📢 cc @loning

⟦AI:AUTO-LOOP⟧

[loning]

📊 当前状态 — Phase 9 round 1(不需要人介入)

维度	值
阶段	Phase 9 r1 三 solver
Cluster	cluster-007-work-unit-contract-schema(脊柱基础,008/009/010 依赖本契约)
codex	minimal/structural/delete 读 audit-iter-2.md cluster-007 节 + grep 代码,独立提案 work-unit 契约 schema
是否需要人介入	❌ 否

🤖 controller status banner

⟦AI:AUTO-LOOP⟧

[loning]

🤖 minimal solver 提案: WorkUnit v1 兼容契约, 不迁移 state

TL;DR

• 这是什么: issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 的 minimal solver 方案, 已验证 evidence 仍真实存在。
• 结论: 建议新增 WorkUnit v1 契约, 但 v1 继续把现有 clusters_* 作为持久队列, cluster_id 作为 legacy alias。
• 下一步: meta-judge 对比 3 个 solver 输出; 本方案不需要 maintainer 介入。

📢 cc 原作者:@loning 请 sanity-check 这个最小边界是否符合 cluster-007 的设计问题。

---

详细说明

最小改动边界是 5 个文件、约 +55 / -18 LOC: REFERENCE.md 增加 WorkUnit v1 最小字段和兼容映射, SKILL.md 初始 state 增加 work_unit_schema_version: 1, Phase 2/3/9 的 prompt 改成以 WORK_UNIT_ID 为主身份。CLUSTER_ID 保留, 但只作为 v1 下 audit cluster 的 legacy source id。

我没有建议现在新增 work_units_planned/active/done/failed, 因为那会引入 state 迁移、恢复兼容和双写问题。对 cluster-007 来说, 先把协议名字和兼容映射钉住即可; 非 audit producer 留给 cluster-008, marker/label 迁移留给 cluster-009/010。

---

📎 完整 codex 原始输出(存档备查)

---

solver: minimal
issue: 3
cluster: cluster-007-work-unit-contract-schema
verdict: propose

Recommended framing

The violation is real, but the minimum viable boundary is a contract clarification plus prompt boundary rename, not a storage migration. Define WorkUnit v1 in REFERENCE.md as the canonical engine contract, with current cluster-* records as the only v1 legacy encoding: work_unit_id == cluster_id, producer == audit, and source_ref == .refactor-loop/runs/audit-iter-N.md#<cluster-id>. Keep clusters_* state fields for v1 recovery compatibility, add work_unit_schema_version: 1, and update Phase 2/3/9 prompts and SKILL dispatch text to pass/use WORK_UNIT_ID while preserving CLUSTER_ID as a legacy alias. This removes the audit's protocol violation without renaming labels, changing producers, or inventing a new runtime scheduler.

Concrete plan (English)

• Files: skills/codex-refactor-loop/REFERENCE.md — add a short WorkUnit v1 contract section before the state schema; document required fields (work_unit_id, kind, producer, source_ref, scope_paths, old_pattern, new_principle, verification_hints, dependencies, risk, leverage) and the v1 compatibility mapping from clusters_* / cluster_id.
• Files: skills/codex-refactor-loop/REFERENCE.md — update the state schema example to include "work_unit_schema_version": 1; keep clusters_planned, clusters_active, clusters_done, and clusters_failed, but explicitly label them as the v1 persisted work-unit queues; update design_pending to show "work_unit_id": "cluster-NNN" plus "cluster_id": "cluster-NNN" as legacy alias.
• Files: skills/codex-refactor-loop/SKILL.md — update initial state.json with "work_unit_schema_version": 1; in Phase 1 validation/transition, say each accepted audit cluster is normalized to a WorkUnit v1 before being stored in clusters_planned; in Phase 2 dispatch, export WORK_UNIT_ID=$CLUSTER_ID; in Phase 9 wording, describe state.design_pending[i] as a work unit and keep cluster as v1 legacy encoding.
• Files: skills/codex-refactor-loop/prompts/implement.md — change the primary framing from "implement ${CLUSTER_ID}" to "implement work unit ${WORK_UNIT_ID}"; keep the audit section lookup under a clearly named v1 compatibility paragraph using ${CLUSTER_ID}.
• Files: skills/codex-refactor-loop/prompts/verify.md — same boundary change as implement: primary identity is ${WORK_UNIT_ID}, legacy audit artifacts still use ${CLUSTER_ID} for v1.
• Files: skills/codex-refactor-loop/prompts/meta-judge.md — change the first line and output guidance from cluster-only to work-unit-first, with cluster as the legacy v1 source id.
• LOC delta: ~+55 / -18 across 5 files.
• Tests to add/modify: no new test files; this repo has no test harness. Verification should be grep-based: rg -n "WorkUnit v1|work_unit_schema_version|WORK_UNIT_ID" skills/codex-refactor-loop/{REFERENCE.md,SKILL.md,prompts/implement.md,prompts/verify.md,prompts/meta-judge.md} and rg -n "任务：实施 \\$\\{CLUSTER_ID\\}|任务：验证 \\$\\{CLUSTER_ID\\}|design-issue .*cluster \\\$\{CLUSTER_ID\}\" skills/codex-refactor-loop/prompts should show the old cluster-only prompt openings are gone.
• Philosophy/CLAUDE.md/SPEC/Tier change: none. This is inside the README generalization route and the Phase 9 design gate; it does not need a CLAUDE.md exception. It is worth no trusted-base cost because it preserves current state recovery and only names the contract that the current fields already implement. Deep consensus should be reachable because it answers the audit's design question directly: v1 keeps clusters_* as persisted fields and documents them as work-unit queues, instead of introducing work_units_* now.
• Migration path: no migration needed. Existing state files without work_unit_schema_version are interpreted as WorkUnit v1 legacy state; future writes include the field.

Concrete plan (中文)

• 文件: skills/codex-refactor-loop/REFERENCE.md — 在 state schema 前新增简短的 WorkUnit v1 contract 小节; 写清最小必填字段(work_unit_id, kind, producer, source_ref, scope_paths, old_pattern, new_principle, verification_hints, dependencies, risk, leverage), 并明确 clusters_* / cluster_id 到 WorkUnit v1 的兼容映射。
• 文件: skills/codex-refactor-loop/REFERENCE.md — 更新 state schema 示例, 增加 "work_unit_schema_version": 1; 保留 clusters_planned, clusters_active, clusters_done, clusters_failed, 但把它们标注为 v1 持久化 work-unit 队列; design_pending 示例同时展示 "work_unit_id": "cluster-NNN" 与 legacy alias "cluster_id": "cluster-NNN"。
• 文件: skills/codex-refactor-loop/SKILL.md — 初始 state.json 增加 "work_unit_schema_version": 1; Phase 1 审计通过后说明每个 audit cluster 先规范化为 WorkUnit v1 再进入 clusters_planned; Phase 2 派发时导出 WORK_UNIT_ID=$CLUSTER_ID; Phase 9 文案改成 state.design_pending[i] 里的 work unit 需要设计决策, cluster 只是 v1 legacy encoding。
• 文件: skills/codex-refactor-loop/prompts/implement.md — 主身份从“实施 ${CLUSTER_ID}”改为“实施 work unit ${WORK_UNIT_ID}”; audit 段读取继续保留, 但放到 v1 compatibility 段落里, 明确 ${CLUSTER_ID} 是 legacy source id。
• 文件: skills/codex-refactor-loop/prompts/verify.md — 与 implement 同步: 主身份是 ${WORK_UNIT_ID}, v1 下的旧审计 artifact 仍用 ${CLUSTER_ID} 定位。
• 文件: skills/codex-refactor-loop/prompts/meta-judge.md — 首行和输出指引改为 work-unit-first, cluster 作为 v1 legacy source id 保留。
• LOC delta: 预计约 +55 / -18, 共 5 个文件。
• 需要新增/修改的测试: 不新增测试文件; 本仓库没有测试 harness。验证用 grep: rg -n "WorkUnit v1|work_unit_schema_version|WORK_UNIT_ID" skills/codex-refactor-loop/{REFERENCE.md,SKILL.md,prompts/implement.md,prompts/verify.md,prompts/meta-judge.md}; 另跑 rg -n "任务：实施 \\$\\{CLUSTER_ID\\}|任务：验证 \\$\\{CLUSTER_ID\\}|design-issue .*cluster \\\$\{CLUSTER_ID\}\" skills/codex-refactor-loop/prompts, 确认旧的 cluster-only prompt 开头已消失。
• Philosophy/CLAUDE.md/SPEC/Tier change: 无。该改动落在 README 的泛化路线和 Phase 9 设计 gate 内, 不需要 CLAUDE.md 例外。它不增加 trusted-base 成本, 因为当前 state 恢复语义不变, 只是把现有字段已经承担的协议命名清楚。深度共识应可达成, 因为它直接回答 audit 的设计问题: v1 保留 clusters_* 作为持久字段, 并文档化为 work-unit 队列, 暂不引入 work_units_*。
• Migration path: 不需要迁移。缺少 work_unit_schema_version 的既有 state 文件按 WorkUnit v1 legacy state 读取; 后续新写入带上该字段。

Risks

• Keeps clusters_* in persisted state for v1, so the internal storage names remain audit/refactor-flavored until a later migration.
• Requires discipline in later clusters: producer adapter and marker/label rename work must not reinterpret this as permission to keep adding cluster-only protocol fields.
• A pure documentation/template contract cannot prove runtime behavior because current controller behavior is mostly written as SKILL instructions, not executable code with tests.
• WORK_UNIT_ID=$CLUSTER_ID is intentionally narrow; non-audit producers still need cluster-008 to define how they populate source-specific fields.

Escalation triggers (if any)

• none

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

• Why this is the minimum: The cited violation is that the protocol has no generic identity. Adding a versioned contract and passing WORK_UNIT_ID at prompt boundaries removes that violation while preserving all existing state queues and recovery behavior.
• What I considered but rejected: Adding work_units_planned/active/done/failed now. That would be cleaner long-term but creates migration, recovery, and dual-write questions across the controller. It belongs after the producer and marker compatibility clusters converge, not in the minimal contract fix.
• What I cannot decide alone: Whether a later v2 should actually rename persisted queues to work_units_*; this proposal deliberately leaves that as a future migration.

SOLVER_DONE:minimal:propose:document WorkUnit v1 over legacy cluster state; no storage migration

⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 structural solver 提案:用 WorkUnitEnvelope 固化工作单元契约

TL;DR

• 这是什么:issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 的 structural solver 独立方案。
• 结论:建议新增 WorkUnitEnvelopeV1(producer 到 solve/implement/verify 的标准工作单元信封),让 cluster 只作为旧审计来源的兼容投影。
• 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 schema 边界是否应只覆盖数据契约,不顺手做 rename/label 迁移。

---

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 REFERENCE.md、SKILL.md、implement/verify/meta-judge prompts 仍把 cluster_id 和 clusters_* 当唯一协议。我的结构性方案是新增 WorkUnitEnvelopeV1:一个带版本号的工作单元信封,记录 id/kind/source/spec/scope/dependencies/risk/leverage,再把旧 cluster-* 映射进 legacy.cluster_id。

计划触达 8 个文件,估算 ~+360/-140 LOC;新增 4 个行为测试。运行成本是每个 work unit 一次本地 normalizer 调用,约 20-40 ms Python 启动和 1-3 KB JSON 对象分配;没有额外 GitHub/network hop。这个方案不改 CLAUDE.md/SPEC/Tier 文件,也不处理 issue #9/#10 的 marker/rename,避免把数据契约和外部命名迁移混在一个 PR。

---

📎 完整 codex 原始输出(存档备查)

---

solver: structural
issue: 3
cluster: cluster-007-work-unit-contract-schema
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)

Introduce a versioned WorkUnitEnvelope as the durable contract between producers and the existing solve -> consensus -> implement -> verify spine. Audit clusters become one producer projection (source.kind = "audit-cluster") instead of the protocol itself. The envelope should live in REFERENCE.md, be materialized into state.json as work_units_*, and be passed to implement/verify/meta-judge prompts through a small normalizer helper so legacy cluster-* fields remain readable aliases. This removes the violation by construction: every phase consumes the same work-unit identity/spec/dependency contract, and audit-only names are confined to the compatibility adapter.

Recommended framing (中文)

引入有版本号的 WorkUnitEnvelope 作为 producer 与现有 solve -> consensus -> implement -> verify 脊柱之间的持久契约。审计产出的 cluster 只是一种 producer 投影(source.kind = "audit-cluster"),不再是协议本身。该 envelope 写入 REFERENCE.md,在 state.json 中落为 work_units_*,并通过一个小的 normalizer helper 注入 implement / verify / meta-judge prompts,旧的 cluster-* 字段只保留为可读兼容 alias。这样违反点按结构消失:所有 phase 都消费同一个 work-unit 身份 / spec / dependency 契约,audit 专用命名被限制在兼容 adapter 内。

Concrete plan

• New abstractions (if any): WorkUnitEnvelopeV1 in Skill runtime contract layer, project skills/codex-refactor-loop, documented in skills/codex-refactor-loop/REFERENCE.md and normalized by skills/codex-refactor-loop/scripts/work_unit_contract.py. Interface fields: schema_version, id, kind, title, source.kind, source.ref, spec.path, spec.section, scope_paths, risk, leverage, dependencies, legacy.cluster_id?.
• Files:
  • skills/codex-refactor-loop/REFERENCE.md: add WorkUnitEnvelopeV1; update state schema from clusters_* primary to work_units_* primary; document clusters_* as derived legacy aliases for state v1 recovery only.
  • skills/codex-refactor-loop/SKILL.md: change bootstrap schema to include schema_version: 2, work_units_planned/active/done/failed; describe Phase 1 as a replaceable producer; update Phase 2/3/9 dispatch wording to use work units and mention cluster_id only under compatibility.
  • skills/codex-refactor-loop/prompts/implement.md: replace required ${CLUSTER_ID} + audit-section contract with ${WORK_UNIT_ID}, ${WORK_UNIT_SPEC_PATH}, ${WORK_UNIT_SPEC_SECTION}, ${WORK_UNIT_KIND}; keep ${CLUSTER_ID} as optional legacy alias when source.kind=audit-cluster.
  • skills/codex-refactor-loop/prompts/verify.md: same work-unit input contract; summary output path becomes verify-${WORK_UNIT_ID}.md.
  • skills/codex-refactor-loop/prompts/meta-judge.md: frame Phase 9 as a design issue for ${WORK_UNIT_ID}; cluster appears only as a legacy source kind.
  • skills/codex-refactor-loop/scripts/controller_lib.sh: extend render_template to render WORK_UNIT_* placeholders and fail if neither WORK_UNIT_ID nor legacy CLUSTER_ID is available.
  • skills/codex-refactor-loop/scripts/work_unit_contract.py: new helper that reads audit cluster YAML or existing state entries, emits canonical envelope JSON, and can project legacy cluster_id for old prompt/log names.
  • skills/codex-refactor-loop/prompts/design-issue-body.md: use work-unit terminology in the durable issue body while including legacy.cluster_id when present.
• LOC delta: ~+360 / -140. Split: REFERENCE.md +120/-45, SKILL.md +110/-60, prompts +80/-35, controller_lib.sh +25/-0, new helper +25/-0 in first pass if kept deliberately narrow.
• Tests to add:
  • skills/codex-refactor-loop/scripts/tests/test_work_unit_contract.py: audit cluster input normalizes to WorkUnitEnvelopeV1 with id, source.kind=audit-cluster, spec.path, dependencies, and legacy.cluster_id.
  • skills/codex-refactor-loop/scripts/tests/test_work_unit_contract.py: non-audit input such as kind=docs-task validates without cluster_id and projects no cluster-only fields.
  • skills/codex-refactor-loop/scripts/tests/test_prompt_contract.py: rendered implement and verify prompts contain no required ${CLUSTER_ID} or mandatory audit section when given a non-audit work unit.
  • skills/codex-refactor-loop/scripts/tests/test_state_projection.py: state v1 with clusters_planned is read as canonical work_units_planned, and writing state v2 keeps legacy aliases derived rather than authoritative.
• proto changes (if any): none.
• Philosophy/CLAUDE.md/SPEC/Tier changes (if any): none. Current CLAUDE invariant delegates engine philosophy to README and already says de-shelling follows README's generalization route. This plan implements that route inside the skill contract; it does not change Tier files, SPEC, conformance, trusted base, or top-level package boundaries.
• Runtime cost: one local normalizer call per dispatched work unit, about 20-40 ms Python startup on macOS plus one JSON object allocation of ~1-3 KB per work unit. No extra GitHub/network hop. If the controller later caches normalized envelopes in state.json, per-wakeup overhead is O(number of open work units) JSON parsing, typically <5 ms for 30 units.

Risks

• The structural framing touches more text than a pure alias rename, so issue cluster-007: v1 work-unit 契约(#3 共识) #8/cluster-008: audit-as-producer 适配(#4 共识) #9/cluster-010: 澄清通用定位(#6 共识 minimal,不重命名) #10 must not simultaneously rename labels or package names in the same PR. Keep this issue to the data contract only.
• WorkUnitEnvelopeV1 can become over-engineering if it starts modeling every future producer. Bound it to fields already consumed by at least two callers: identity, source reference, spec location, scope, risk/leverage, dependencies, and legacy projection.
• State v2 migration can create ambiguity if both clusters_* and work_units_* are hand-edited. Make work_units_* authoritative and document clusters_* as read-only derived aliases for recovery.

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why this structure beats a documented exception: An exception would say "non-audit producers may fake cluster_id"; that preserves the same audit-only spine. A canonical envelope makes audit and non-audit producers enter through the same contract and prevents future prompts from adding new cluster-only assumptions.
• Why I picked this abstraction over alternatives: WorkUnitEnvelopeV1 has explicit justification from the README extension point and at least four concrete callers: implement prompt, verify prompt, meta-judge prompt, and design issue/state projection. A full actor inbox or event-sourcing layer is unnecessary here because this repo's durable coordination primitives are GitHub comments/labels, log markers, prompt files, and state.json debug projection, not an in-process Channel/IAsyncEnumerable.
• What I cannot decide alone: Exact kind vocabulary beyond the initial audit-cluster, docs-task, spec-change, and design-proposal should be finalized by consensus across issue cluster-007: v1 work-unit 契约(#3 共识) #8 producer adaptation, but the envelope field names above are sufficient for issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3.

SOLVER_DONE:structural:propose:WorkUnitEnvelopeV1 as canonical state/prompt contract with legacy cluster projection

⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 Delete solver：建议删掉通用 work-unit schema 要求

TL;DR

• 这是什么：issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 的 delete 视角 solver 输出，结论是 propose。
• 结论：不要新增 work_units_* schema；删除/收窄“任意 work-unit 来源是当前不变量”的 README/插件承诺。
• 下一步：meta-judge 对三 solver 仲裁；若采纳此路，implement 只改文档/manifest，不改运行中 cluster 控制面。
  📢 cc 原作者:@loning 请 sanity-check：是否确实要把 generic R&D 定位当作当前强约束。

---

详细说明

我把这个点归类为 orphan feature：work-unit 是 README 和插件文案里的产品承诺，但仓库里没有实现的 work_unit schema、没有第二个 producer、没有测试、也没有公共调用方。当前真实链路仍是 audit -> cluster -> solve -> consensus -> implement -> verify。

反向证据是：work_units_ / work_unit / WORK_UNIT / WorkUnit 在 shipped Markdown/shell/Python/JSON 里 0 个实现命中；cluster_id / CLUSTER_ID / clusters_* 有 70 个真实控制面命中；git log -p -S 显示 cluster 契约来自初始 verbatim 移植，而 generic 化当时只是“后续泛化”。所以最低代码量方案不是加 schema，而是删除这条 roadmap-as-invariant，让 cluster_id / clusters_* 保持 v1 支持契约。

---

📎 完整 codex 原始输出(存档备查)

---

solver: delete
issue: 3
cluster: cluster-007-work-unit-contract-schema
verdict: propose

Classification

(b) Orphan feature — the generic work-unit contract is a README/plugin promise and an audit-derived design issue, but there is no implemented generic work_unit schema, no generic producer, no public API, no tests, and no caller outside the current audit/cluster path. The real shipped capability is still audit -> cluster -> solve -> consensus -> implement -> verify.

Recommended action (English)

Delete the requirement to introduce a generic work-unit schema now. Keep cluster_id and clusters_* as the supported v1 contract for the shipped codex-refactor-loop, and delete or narrow the repository/product wording that makes arbitrary work-unit sources a current invariant. This avoids adding work_units_*, compatibility shims, schema migration rules, and prompt aliases before a second producer exists. If a future non-audit producer becomes real, it should arrive with its own concrete producer and tests, not as a speculative schema layer.

Recommended action (中文)

现在删除“必须引入通用 work-unit schema”的要求。保留 cluster_id 和 clusters_* 作为已发布 codex-refactor-loop 的 v1 契约，并删除或收窄仓库 / 产品文案中“任意 work-unit 来源已是当前不变量”的表述。这样可以避免在没有第二个 producer、没有测试、没有公共调用方时先加 work_units_*、兼容 shim、schema migration 和 prompt alias。未来如果真的出现非 audit producer，应由那个 producer 连同测试一起提出新契约，而不是现在预铺一层抽象。

Concrete plan (if propose)

• Files to delete: none. Delete the orphan schema requirement and product promise, not the working skill files.
• Caller migrations:
  • README.md:3,7,11-19 -> narrow positioning from "generic consensus R&D for any git-state-producing activity" to "consensus-driven autonomous refactor/rule-compliance loop".
  • README.md:72 -> delete the binding "任意 work-unit 来源" roadmap item, or rewrite it as a non-goal for the shipped v1 product.
  • .claude-plugin/plugin.json, .codex-plugin/plugin.json, .cursor-plugin/plugin.json, gemini-extension.json, package.json -> narrow descriptions that imply arbitrary R&D/work-unit support; keep consensus, multi-agent, codex, orchestration, refactor, workflow.
  • .codex-plugin/plugin.json:27 -> remove "把任意持续向仓库提交状态的活动当作研发"; keep the actual codex-refactor-loop three-stage audit/implement/verify description.
  • CLAUDE.md:39 -> change the current invariant from "泛化路线见 README;脱壳 / 重名前不要改正文逻辑" to "current shipped contract is the migrated refactor loop; generic work-unit sources are not required for repository correctness."
  • IMPROVEMENT-BACKLOG.md:F7 -> mark the work-unit generalization item rejected or non-binding, not pending work.
  • skills/codex-refactor-loop/REFERENCE.md, SKILL.md, prompts/implement.md, prompts/verify.md, prompts/meta-judge.md -> no migration; retain existing cluster_id / clusters_* / audit-section contract.
• Tests to delete: none. There are no tests covering a generic work-unit contract.
• LOC delta: -15 to -45 documentation/manifest lines, depending on whether README "泛化路线" is deleted outright or narrowed to a one-sentence non-goal.
• Philosophy/CLAUDE.md/SPEC/Tier changes (if any): README.md changes from "any committed repository activity is R&D and can become a work-unit source" to "this repository currently ships a consensus refactor/rule-compliance loop; arbitrary work-unit sources are outside v1 scope." CLAUDE.md changes from treating README genericization as the current route to treating the existing refactor loop as the supported contract. This is worth it because it deletes a roadmap-as-invariant that is forcing a schema with 0 current generic producers. Consensus can hold because the existing skill, labels, markers, prompts, and scripts already agree on the narrower cluster contract.
• Tracking issue (if defer): none. This is a delete-now plan, not a deferral.

Reverse-evidence (why this is safe to delete)

• No public API breaks (verified by git grep on public surface): the only positive generic work-unit promise is in README/plugin-facing wording; the shipped skill surface uses cluster_id, CLUSTER_ID, and clusters_*.
• No C# callers exist: rg -l 'cluster_id|CLUSTER_ID|clusters_' --type cs searched no files because this repository has no C# source surface.
• No external repo dependency on the code: external repos were untouched, and no generic work_unit interface exists for them to depend on.
• Tests covering the path are themselves not load-bearing: there are no tests for generic work-unit intake or schema migration.
• rg -n 'work_units_|work_unit|WORK_UNIT|WorkUnit' over Markdown/shell/Python/JSON returned 0 implemented schema hits.
• rg -n 'cluster_id|CLUSTER_ID|clusters_' returned 70 hits in the real control plane, all aligned with the current audit/cluster path.
• Git blame/history shows clusters_planned and CLUSTER_ID were introduced in the initial import commit as part of the verbatim codex-refactor-loop; the same commit called genericization "后续泛化", not a shipped interface.
• The value chain remains intact after deletion: audit still produces clusters; design issues still run Phase 9; implement/verify still consume cluster prompts; PR/reviewer/CI flow remains unchanged.
• Cluster 008/009/010 only require this schema if the generic product promise survives. If issue 3 deletes the promise, those downstream clusters should be re-scoped or dropped rather than implemented on a speculative base.

Risks

• The maintainer may intentionally want the README aspiration to remain binding even before any second producer exists.
• Users may have installed the plugin because the marketing promised arbitrary R&D instead of the concrete refactor loop.
• Narrowing the contract lowers product ambition and may force a future non-audit producer to reopen design from first principles.
• Hidden downstream automation could parse README's "泛化路线" as a task source; I found no parser or caller in this repository.
• The prompt asks for CLAUDE.md "## 架构哲学" / "删除优先", but the current CLAUDE.md has no such heading or clause; I treated the active deletion-first rule as coming from the solver/reviewer prompts.

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why I claim this can be deleted: the requested schema is not implemented behavior; it is a consequence of a broad README/product invariant. The existing engine has exactly one work source shape: audit clusters. Adding a generic schema now creates a compatibility layer for zero current producers.
• What I checked to verify safety: gh issue view 3 --json; .refactor-loop/runs/audit-iter-2.md cluster-007; README.md; CLAUDE.md; plugin manifests; GEMINI.md; IMPROVEMENT-BACKLOG.md; REFERENCE.md; SKILL.md; implement/verify/meta-judge prompt call sites; rg for cluster_id, CLUSTER_ID, clusters_*, and work_unit; git grep over public surfaces; rg -l ... --type cs; git log --oneline; git log -p --follow -S.
• What I cannot decide alone: whether the maintainer wants the broad generic R&D positioning to remain a binding product goal despite the missing producer. If yes, delete solver should abstain and let minimal/structural define the schema.

SOLVER_DONE:delete:propose:delete generic work-unit schema requirement; keep cluster contract as shipped v1
⟦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)

维度	值
阶段	Phase 9 r1 meta-judge
Cluster	cluster-007 work-unit 契约 schema
是否需要人介入	❌ 否

🤖 controller status banner

⟦AI:AUTO-LOOP⟧

[loning]

🤖 meta-judge：issue #3 进入第 2 轮收敛

TL;DR

• 这是什么:issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3 / cluster-007 的 Phase 9 r1 meta-judge 仲裁结果。
• 结论:没有 3/3 共识;minimal、structural、delete 三个方案在产品边界和 state schema 上互相冲突。
• 下一步:controller 应发起 round 2,用下面的窄问题要求三个 solver 二选一/三选一。

---

详细说明

本轮 3 个 solver 都是 propose,但不是同一 framing。minimal 选择在现有 clusters_* 上文档化 WorkUnit v1,只加 WORK_UNIT_ID 边界 alias;structural 选择新增 WorkUnitEnvelopeV1(producer 到 solve/implement/verify 的标准工作单元信封)、work_units_*、normalizer helper 和测试;delete 选择删除/收窄 README、manifest、CLAUDE 中的通用 work-unit 产品承诺,继续把 cluster_id / clusters_* 作为 v1 契约。

方案	state 权威字段	文件/成本	trade-off
minimal	clusters_*	约 +55/-18, 5 文件	最小兼容,但通用化主要停在文档和 prompt 边界
structural	work_units_*	约 +360/-140, 8 文件 + 4 测试	契约更实,但立刻承担迁移和兼容成本
delete	clusters_*	约 -15/-45, 文档/manifest	删除抽象债,但降低 README 的通用 R&D 野心

收敛问题:issue #3 的 binding contract 到底要保留 README 的通用 work-unit 野心,还是删除/收窄这条野心;如果保留,v1 应保持 legacy-backed clusters_* + WORK_UNIT_ID alias,还是现在迁移到权威 WorkUnitEnvelopeV1 / work_units_* + normalizer + tests?

---

📎 完整 codex 原始输出(存档备查)

---

issue: 3
cluster: cluster-007-work-unit-contract-schema
convergence_round: 1
solver_verdicts:
minimal: propose
structural: propose
delete: propose
decision: converge

Decision (English)

Convergence is required. All 3 solvers produced concrete proposals, but they do not share the same framing: minimal keeps clusters_* as the WorkUnit v1 persisted queues and adds WORK_UNIT_ID as a prompt boundary alias; structural introduces an authoritative WorkUnitEnvelopeV1, work_units_* state, a normalizer helper, and tests; delete rejects the generic work-unit requirement and narrows README/manifest/CLAUDE wording to the shipped cluster contract. That is a direct contradiction on product ambition, state authority, migration cost, and file scope, not a 3/3 consensus. Because this is convergence round 1, the stall trigger cannot fire, and there is no maintainer tie-breaker in the issue comments, the next action is a narrowed convergence round.

Decision (中文)

需要进入收敛轮。3 个 solver 都给出了具体方案,但 framing 不一致: minimal 保留 clusters_* 作为 WorkUnit v1 的持久队列,只在 prompt 边界新增 WORK_UNIT_ID alias; structural 引入权威的 WorkUnitEnvelopeV1、work_units_* state、normalizer helper 和测试; delete 则拒绝当前引入通用 work-unit 要求,改为收窄 README / manifest / CLAUDE 中的产品承诺,继续支持现有 cluster 契约。这些分歧直接落在产品野心、state 权威字段、迁移成本和文件范围上,不是 3/3 深度共识。当前只是第 1 轮,stall 条件不成立,issue 评论里也没有 maintainer tie-breaker,所以下一步应发起带窄问题的第 2 轮收敛。

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 3: n/a

If converge

• Convergence question (specific): For issue [refactor-design] cluster-007: work-unit 契约 schema(通用化脊柱拆分 1/4) #3, should the binding contract preserve the README generic work-unit ambition or delete/narrow that ambition; if preserved, should v1 keep legacy clusters_* as authoritative queues with WORK_UNIT_ID aliases, or migrate now to authoritative WorkUnitEnvelopeV1 / work_units_* with a normalizer and tests?
• What each solver should address explicitly:
  • Choose exactly one of the existing framings: minimal legacy-backed WorkUnit v1, structural WorkUnitEnvelope/state migration, or delete/narrow generic ambition; do not introduce a fourth hybrid.
  • State whether README generic work-unit positioning remains a binding product invariant for v1, and cite the exact README/CLAUDE/manifest lines to keep or change.
  • If preserving generic work units, specify which state fields are authoritative in v1: clusters_* or work_units_*, and how old state is read.
  • If deleting/narrowing the ambition, explain why downstream clusters 008/009/010 should be dropped or re-scoped instead of implemented.
  • Give the exact file set, approximate LOC delta, tests or grep verification, and compatibility behavior for existing cluster_id / CLUSTER_ID prompts.
• 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-issue3-r1-minimal.md
• solver-structural: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue3-r1-structural.md
• solver-delete: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue3-r1-delete.md

META_JUDGE_DONE:converge:round-2:Should issue #3 preserve generic work-unit ambition and choose legacy-backed WorkUnit v1 vs WorkUnitEnvelope migration, or delete/narrow the ambition?

⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

📊 当前状态 — Phase 9 round 2(收敛中,不需要人介入)

维度	值
收敛问题	保留通用 work-unit 抱负+选 v1/Envelope 迁移,还是删/收窄
是否需要人介入	❌ 否

🤖 controller status banner

⟦AI:AUTO-LOOP⟧