design: 程序检测不应依赖自然语言字面 — 支持中英切换

[loning]

背景

当前 source-regression tests / banner filters / forbidden-token lists / comment filters 大量直接 grep 中文字面(e.g. `"ensure all 5 daemons"`、`"需-maintainer-决策"`、`"## 🤖"` + 后续中文段标题)。一旦把面向人的文案切换到英文(或反之),所有这些检测会:

• 假绿(检测目标找不到了,跳过)
• 假红(detector 期望的字面不在新语言里)

初步 grep:~52 处 NL-anchored detection site(只统计了几个高频 token)。

目标

把"程序检测"和"人类文案"解耦,使中英文一键切换不影响:

• 共识引擎(marker / sentinel / truth-table)
• daemon/controller 守恒(allowlist / forbidden-token / lifecycle 边界)
• Phase 0/9 fixed-point 检测
• Comment / banner 识别

不可动

• marker 词表(`SOLVER_DONE`/`META_JUDGE_DONE`/...)— 已经 language-agnostic
• sentinel `⟦AI:AUTO-LOOP⟧` — 已经 language-agnostic
• 真值表 / Phase routing / 3/3 unanimous

可以砍 / 替换

• `test_ensure_project_rules_fixed_points.py` 字面断言中文段标题 → 改断言 marker token / heading ID / 结构 anchor
• banner filter `## 🤖` + 中文文字 → 改纯 emoji prefix(已经 emoji-only 可工作)+ 用 sentinel 做权威识别
• forbidden token list `"ensure all 6 daemons"` → 改抽象 invariant ID(e.g. `DAEMON_COUNT_DRIFT`)
• 中文 banner role 推断(reviewer/judge/...)→ 改 marker role suffix 推断

验证标准

• 全部 skill 文档可切换到 English(或反之)而不破 behavior test / source-regression test
• 共识引擎语义零变更
• detection 覆盖率不降
• Token 总量大致不变(目标是 detection-loose-coupling,不是减量)

Phase 9 framing

• minimal: 逐 detection site 把中文字面改成 marker / token / emoji / heading-id anchor,不引入新抽象
• structural: 引入 `DetectionInvariantV1`(invariant_id + canonical anchor list,language-agnostic),所有 detector 通过 invariant_id 查
• delete: 删除不必要 detection,改靠 schema / exit code / marker 推断;只保留 invariant 必须的最小集

⟦AI:AUTO-LOOP⟧

[loning]

🔁 Phase 9 r1 dispatched — 3 framing 独立 reasoning

• minimal: 逐 detection site 把中文字面换 marker/token/emoji/heading-id anchor,无新抽象
• structural: 引入 DetectionInvariantV1(invariant_id → canonical anchor list),detector 通过 ID 查
• delete: 删多余 detection,改靠 schema/exit-code/marker 推断

r1 三独立 → meta-judge consensus / converge:r2 / escalate:stalled。

⟦AI:AUTO-LOOP⟧

[loning]

🤖 delete solver: 删除 NL 字面检测锚点

TL;DR

• 这是什么:delete solver 视角,结论是删“自然语言文案字面检测层”,不是删检测能力本身。
• 当前结论:propose,复用现有 ⟦AI:AUTO-LOOP⟧ / marker / EXIT=0 / label / daemon token,不新增 DetectionInvariantV1。
• 下一步:meta-judge 可把本方案和 minimal/structural 对齐,若达成共识则让 implement 删除/替换相关 source-regression 断言。

---

详细说明

我反向追了 comment-monitor.sh、peek.sh、post_banner.py、test_ensure_project_rules_fixed_points.py 和 test_skill_entrypoint_contract.py。真正 load-bearing 的能力是识别 AI 评论、Phase marker、human label 语义、daemon bootstrap 顺序;这些已有稳定机器面:⟦AI:AUTO-LOOP⟧、SOLVER_DONE / META_JUDGE_DONE、EXIT=0、label 常量、daemon 脚本名和 heartbeat JSON。

因此 delete 方案是 collapse-and-delete:删掉测试里要求中文标题/中文句子/emoji banner 英文短句必须存在的断言,迁移到现有机器 token 和 schema/key/function 检查。保留 legacy emoji-prefix fallback,但不再让 TL;DR、详细说明、完整 codex 原始输出、ensure all 5 restart-helper-managed daemons 这类人类文案成为程序真值源。

---

📎 完整 codex 原始输出(存档备查)

---

solver: delete
issue: 106
cluster: issue-106
verdict: propose

Classification

(d) Genuinely needed but over-built — feature is real, but the current failure is caused by source-regression/tests treating human-readable prose as machine contract. The right delete framing is to delete NL-literal detection sites and collapse them onto existing machine tokens: ⟦AI:AUTO-LOOP⟧, terminal role markers, EXIT=0, label constants, daemon heartbeat JSON, and script/schema fields.

Recommended action

建议删除“自然语言文案字面必须存在/不存在”的检测层,不要引入新的 DetectionInvariantV1 注册表。保留能力本身:AI 评论识别继续靠 ⟦AI:AUTO-LOOP⟧ 末尾 sentinel,Phase 路由继续靠 SOLVER_DONE / META_JUDGE_DONE 等 marker 加 EXIT=0,human label 语义继续靠现有 label 常量和 helper guard,daemon 数量继续靠 restart helper allowlist / heartbeat JSON。对 test_ensure_project_rules_fixed_points.py 和 test_skill_entrypoint_contract.py 中绑定中文标题、中文段落、emoji heading、英文短句的 source-regression 断言做删除或迁移:只断言结构性 token、函数名、schema key、marker allowlist、sentinel 末行,不再断言“当前中文怎么写”。

Concrete plan (if propose)

• Files to delete: none as whole files. Delete selected NL-literal assertions/allowlist branches inside skills/codex-refactor-loop/scripts/test_ensure_project_rules_fixed_points.py and skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py.
• Caller migrations:
  • test_skill_entrypoint_contract.py::test_first_wakeup_bootstrap_obligations_are_ordered_in_skill_alone → stop asserting exact phrase ensure all 5 restart-helper-managed daemons; assert ordered machine anchors instead: restart-daemons.sh, the five daemon executable names, then daemon-event Monitor.
  • test_skill_entrypoint_contract.py::test_heavy_reference_material_is_not_in_entrypoint → replace exact Chinese banner headings (## 📊 当前状态 — <phase>, ## 🆘 状态卡片 — 共识机制无法继续收敛) with anchor/category checks that SKILL links to status-banner/escalation reference while heavy templates remain out of SKILL.
  • test_ensure_project_rules_fixed_points.py human-label tests → keep CANONICAL_HUMAN_LABELS, META_RESOLVED:escalate-human, helper function names, and gh label create checks; delete assertions that require exact Chinese section titles such as ## \👤 human:需-maintainer-决策` 严格语义(强制)` or historical Chinese heading prose.
  • test_ensure_project_rules_fixed_points.py AI-post tests → keep final independent ⟦AI:AUTO-LOOP⟧, --body-file, and terminal marker checks; remove assertions that require 末尾独立一行, 本 skill 的 ..., controller progress reporter, 反模式, and similar human prose.
  • comment-monitor.sh / peek.sh call sites → no capability deletion. Treat emoji-prefix checks (## 🤖, ## 📊, ## ✅, ## 🆘) as legacy fallback only; tests should assert sentinel-first behavior and bounded fallback list, not exact Chinese banner titles.
  • prompts/_github-post-rules.md consumers → keep post format as user-facing convention for now, but tests should not use TL;DR/详细说明/完整 codex 原始输出 as machine truth. The machine truth is first-line AI prefix plus final sentinel.
• Tests to delete: no full test file. Delete/replace NL-literal subtests in test_ensure_project_rules_fixed_points.py and test_skill_entrypoint_contract.py; add focused behavior tests only where a source-regression assertion was pretending to test behavior, e.g. comment monitor skips any body with sentinel regardless of language.
• LOC delta: about -60 to -120 net. Deletion comes from removing per-phrase assertIn / assertNotIn lists and their exception branches; a few replacements may add smaller token/schema assertions.
• Philosophy/CLAUDE.md/SPEC/Tier changes (if any): none. Existing facts already support this: CLAUDE.md says behavior constraints must land in mechanical verification, and the issue narrows what counts as mechanical verification. The proposed invariant is local to this skill: “machine detection must target stable markers/schema/exit status/sentinel, not natural-language prose”. Deep consensus should hold because it reduces false red/false green risk without weakening marker, sentinel, label, or daemon contracts.

Reverse-evidence (why this is safe to delete)

• No public API breaks: find . -name '*.cs' returns no C# surface; this repo is skill/docs/scripts, and the affected paths are tests/prompts/scripts under skills/codex-refactor-loop.
• Existing machine tokens already cover the important behavior: SKILL.md names final independent ⟦AI:AUTO-LOOP⟧; REFERENCE.md requires EXIT=0 before terminal markers; apply_triage_decision.py already validates final sentinel by splitlines()[-1:]; comment-monitor.sh sentinel-checks the whole body before legacy heading prefixes.
• Git blame shows exact phrase anchoring is churn-prone: ensure all 5 restart-helper-managed daemons was introduced by issue design: phase9_router_daemon 漏 detect SOLVER_DONE triplet → 手动派 meta-judge 3 次 #100 only to align helper/SKILL/tests wording after a 5/6-daemon mismatch. The load-bearing fact is the daemon allowlist and Monitor separation, not that English sentence.
• Tests covering the path are not load-bearing when they assert prose. A Chinese heading assertion proves copy text survived; it does not prove route semantics, no-lifecycle boundaries, or marker validity.
• No external repo dependency on the wording was found in tracked source. Host-facing behavior is GitHub comments and skill contracts; those remain governed by sentinel/markers and current post rules.
• The plan avoids a new registry abstraction. DetectionInvariantV1 would add another fact source for names that already exist as constants, function names, markers, schema keys, and sentinel.

Risks

• Some source-regression assertions contain useful smoke coverage accidentally. When deleting them, replacements must preserve the underlying invariant: label count/creation, final sentinel, lifecycle ban, ordered Phase 0 obligations, and daemon roster.
• The legacy emoji-prefix fallback in comment-monitor.sh still depends on ## 🤖 / ## 📊 for old comments without sentinel. Removing that fallback immediately could cause self-react loops on historical AI posts, so this proposal deletes tests' dependence on full prose but does not delete the fallback code yet.
• If GitHub post rules intentionally require specific Chinese section names for reader ergonomics, those can remain as prompt style rules; they just should not be treated as machine detection truth.
• Over-pruning source-regression could make docs drift easier. The guardrail is to replace exact prose assertions with semantic anchors, not to remove all coverage.

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why I claim this can be deleted: the value chain for the violating sites ends at tests and filters that want to distinguish AI/controller posts, Phase routing, label semantics, or daemon bootstrap obligations. Those capabilities do not require Chinese/English prose strings; they already have stable machine surfaces: sentinel, markers, EXIT=0, helper names, labels, and daemon/script names.
• What I checked to verify safety: gh issue view 106 --comments; .refactor-loop/runs/audit-iter-.md; deletion-first clause in CLAUDE.md/AGENTS.md; grep for ## 🤖, 需-maintainer-决策, ensure all 5/6 daemons, TL;DR/details strings, sentinel and marker usage; comment-monitor.sh, post_banner.py, peek.sh, test_ensure_project_rules_fixed_points.py, test_skill_entrypoint_contract.py; git history for solver-delete.md, comment-monitor.sh, and issue design: phase9_router_daemon 漏 detect SOLVER_DONE triplet → 手动派 meta-judge 3 次 #100 daemon wording.
• What I cannot decide alone: exact replacement wording for every assertion if implementers choose a minimal patch, but the deletion boundary is clear: prose is not a detector; stable tokens/schema are.

SOLVER_DONE:delete:propose:delete NL-literal source-regression anchors and collapse detection onto existing sentinel/marker/EXIT/label/daemon-token surfaces, no new DetectionInvariantV1

⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 structural solver — issue 106 DetectionInvariantV1 framing

TL;DR

• 这是什么:structural solver 给 issue 106 的独立设计方案。
• 结论:建议引入 DetectionInvariantV1,让检测读 semantic id / stable anchor,文案只做 display。
• 下一步:meta-judge 汇总 3 个 solver 后决定 consensus / converge / stalled。

---

详细说明

我复核了当前证据: comment-monitor.sh 用 ## 🤖 / ## 📊 等 heading prefix 跳过 AI 评论,peek.sh 用相同前缀和 emoji label prefix 做状态检测,auto_release_gate.py / concurrency_monitor.py 直接依赖 👤 human:需-maintainer-决策,source-regression tests 也有多处中文 heading / 段落字面断言。这些都把“检测语义”和“人类文案”绑在一起。

推荐方案是新增一个 skill-local registry: DetectionInvariantV1。首次出现说明:它是检测不变量注册表,用 invariant_id 表示程序要检测的语义,再映射到稳定 marker/sentinel/anchor/operational token 和 zh/en display text。这样中英切换只改 display mapping,不会改 daemon/filter/source-regression 的判断入口。

方案点	结构取舍
新 helper	detection_invariants.py 提供 lookup/validate/CLI,5+ 个当前 caller 共用
规则变更	CLAUDE.md 增加“机器锚点与人类文案解耦”一条,窄化既有事实源/验证规则
成本	约 +620 / -240 LOC,新增 3 组行为测试,无新 daemon/无 GitHub mutation
风险控制	registry 只存 selector/display,行为事实仍锚到 SKILL/REFERENCE canonical anchors

---

📎 完整 codex 原始输出(存档备查)

---

solver: structural
issue: 106
cluster: issue-106
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。

治理前置:架构性 / 流程性规则与对应机械验证手段同时进仓库,缺一不补口径。

Recommended framing

引入 DetectionInvariantV1:一个 skill-local 的机器检测注册表,把“检测什么语义”固定为 invariant_id,把“当前中文/英文怎么展示”降级为 display text。comment-monitor.sh / peek.sh / auto_release_gate.py / concurrency_monitor.py / check_skill_degradation.py 都从同一注册表取 stable selector、label name、legacy AI prefix、source anchor,不再 grep 中文正文。这样语言切换只改文案渲染或 registry display mapping,不会改 detector 的语义入口;违规由结构消失,不是靠“以后别写中文 assertIn”。

Concrete plan

• New abstractions (if any): DetectionInvariantV1。
  • Interface: schema_version, invariants[].id, kind, stable_tokens, canonical_anchors, display.zh, display.en, operational_token, legacy_matchers, owners.
  • Helper API: load_detection_invariants(repo_root), token(id), anchors(id), legacy_prefixes(group), label_name(id, language), assert_invariant_present(id, text).
  • CLI: detection_invariants.py token <id>, prefixes <group>, labels --language zh|en, check-source <id> <path>.
  • Layer + Project: skill runtime support, skills/codex-refactor-loop.
• Files:
  • CLAUDE.md / AGENTS.md: add repo-level rule “机器锚点与人类文案解耦” under 设计哲学.
  • skills/codex-refactor-loop/SKILL.md: add Controller Contract row for DetectionInvariantV1; add $LOOP_OUTPUT_LANGUAGE=zh|en host field; update language policy wording.
  • skills/codex-refactor-loop/REFERENCE.md: document registry schema, allowed matcher kinds, migration rule, and “display text is not detector authority”.
  • skills/codex-refactor-loop/detection-invariants.json: new registry with ids for ai_generated_comment, controller_status_comment, human.awaiting_maintainer_decision, human.auto_continue, phase labels, skill_degradation_watch, daemon roster, GitHub sentinel, and source-regression anchors.
  • skills/codex-refactor-loop/scripts/detection_invariants.py: new stdlib-only loader/validator/CLI.
  • skills/codex-refactor-loop/scripts/comment-monitor.sh: load AI/sentinel/prefix config once per tick; keep sentinel as primary; use registry for legacy prefixes.
  • skills/codex-refactor-loop/scripts/peek.sh: replace hardcoded AI heading prefixes and phase/human label prefix jq fragments with registry-provided tokens.
  • skills/codex-refactor-loop/scripts/auto_release_gate.py: replace literal 👤 human:需-maintainer-决策 with semantic id lookup.
  • skills/codex-refactor-loop/scripts/concurrency_monitor.py: replace literal human/phase label matching with registry label sets.
  • skills/codex-refactor-loop/scripts/check_skill_degradation.py: replace REQUIRED_*_MARKERS natural-language literals with invariant ids + canonical anchors.
  • skills/codex-refactor-loop/scripts/test_ensure_project_rules_fixed_points.py: shrink NL literal assertions to registry-id assertions; keep tests that verify semantics through helper behavior.
  • skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py, test_statusline_daemon_health_regression.py, test_marker_emission_contract.py: update source-regression expectations to use ids/anchors instead of localized headings.
• LOC delta: ~+620 / -240.
  • New registry: ~140 LOC JSON.
  • New helper: ~180 LOC Python.
  • New behavior tests: ~260 LOC.
  • Script rewiring: ~+120 / -120.
  • Rule/docs update: ~+80 / -40.
• Tests to add:
  • scripts/test_detection_invariants.py: 8 behavior tests: registry schema valid; ids unique; missing id fails closed; label_name() switches zh/en without changing id; legacy prefix group returns only configured prefixes; assert_invariant_present() passes on anchor not prose; removing anchor fails; CLI exits nonzero on malformed registry.
  • scripts/test_comment_origin_detection.py: 5 behavior tests: sentinel-only AI comment skipped; English AI heading skipped via registry; Chinese heading skipped via registry; bot author skipped without localized body; maintainer comment without sentinel/prefix is not skipped.
  • scripts/test_language_switch_detection.py: 6 behavior tests using temp registry language en: concurrency_monitor.compute_expected() still treats maintainer-decision label as waiting; auto_release_gate queries configured label; peek.sh non-AI filter ignores AI comments after English banner; phase label filtering survives English display names; source-regression anchor check survives rewritten English prose; marker/sentinel remain unchanged.
  • Update existing behavior tests that construct literal labels to ask the registry for the operational label.
• proto changes (if any): none.
• Philosophy/PROJECT_RULES/SPEC/Tier changes (if any):
  • CLAUDE.md / AGENTS.md, section 设计哲学(跨 skill 不动点), add:
    - **机器锚点与人类文案解耦**:程序检测、source-regression、daemon/comment filters 只能依赖 marker、sentinel、invariant_id、结构化 schema、heading ID 或 registry 中的 operational token;不得依赖可本地化自然语言正文。语言切换只能改 display text,不得改 detector 语义。
    Current invariant: fact-source and verification are required, but no clause forbids localized prose as detector input.
    Proposed invariant: detector authority is a stable machine anchor, not human copy.
    Trusted-base cost: one repo-level bullet plus tests; no lifecycle authority, no daemon authority, no Tier file, no GPG.
    Why consensus can hold: it directly narrows an abuse-prone surface and strengthens existing “事实源唯一 / 变更必须可验证” clauses.
  • skills/codex-refactor-loop/SKILL.md, language policy from:
    Policy: Source files are English-only; external user-facing artifacts are 中文 by default. No mandatory parallel English section.
    to:
    Policy: Source files are English-only. External user-facing artifacts render in the configured loop output language, default 中文. No mandatory parallel English section. Program detection must use markers, sentinel, DetectionInvariantV1 ids, structured schema, heading ids, or operational tokens; it must not depend on localized display text.
  • skills/codex-refactor-loop/SKILL.md, host config add:
    | $LOOP_OUTPUT_LANGUAGE | External artifact display language for generated GitHub/body/status text | zh(default) or en; machine markers/sentinel/invariant ids do not change |
  • SPEC/conformance/trusted_base/Tier changes: none.
• Runtime cost:
  • comment-monitor.sh: +1 registry load per 30s tick via Python or jq, ~5-20 ms local, 0 GitHub API calls added, no per-comment Python fork.
  • peek.sh: +1 registry load per invocation, ~5-20 ms; still dominated by gh calls.
  • Python daemons/helpers: one JSON parse at startup or lazy singleton, ~20-60 KB heap, <1 ms lookup per label/prefix.
  • No actor inbox hop, no Channel/IAsyncEnumerable/projection pipeline change, no extra daemon, no additional GitHub mutation.

Risks

• 比 minimal 方案多一个 registry/helper,初始实现成本高约 400 LOC。边界控制: registry 只放 detector selectors 和 display mapping,不放工作流语义;SKILL/REFERENCE 仍是人类语义源。
• registry 可能变成第二事实源。控制方式: DetectionInvariantV1 的 canonical_anchors 只指向 SKILL/REFERENCE 中的 stable anchors;测试断言 registry id 与文档 anchor 同时存在,不允许 registry 自己解释行为。
• shell 脚本读取 registry 会增加一点复杂度。控制方式:每个脚本 tick 只读一次,失败 fail closed,并保留 sentinel 作为 AI comment 主判定。

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why this structure beats a documented exception: exception 只能说“这些中文 grep 先保留”,但语言切换仍会继续破坏 detector。Registry + invariant id 把 detector 输入类型改掉,新检测点默认没有机会依赖中文正文。
• Why I picked this abstraction over alternatives: 逐点改 token 会留下 50+ 个分散约定;删除 detection 会降低 coverage。DetectionInvariantV1 有至少 5 个当前 caller: comment-monitor、peek、auto_release_gate、concurrency_monitor、check_skill_degradation/source-regression。
• What I cannot decide alone: registry 初始 id 命名和英文 display 文案可由 implement worker按现有 label/phase taxonomy落地;但结构边界应固定为 semantic id + stable anchor + display mapping,不能退回自然语言 grep。

SOLVER_DONE:structural:propose:use DetectionInvariantV1 registry so detectors consume semantic ids and stable anchors, while zh/en text becomes display-only
⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 minimal solver: issue 106 最小修法

TL;DR

• 这是什么: issue 106 的 minimal 独立 solver 结论。
• 结论:违规真实,但只需把测试里的中文文案锚点换成协议 token / markdown anchor,不引入新 invariant registry。
• 下一步:meta-judge 可比较 minimal/structural/delete,若收敛则派 implement 改 SKILL.md 与现有 source-regression tests。

---

详细说明

我核了当前源码: comment-monitor.sh 已经用 ⟦AI:AUTO-LOOP⟧ sentinel 做主判定,## 🤖 只是 legacy emoji-prefix 兜底,所以这里不需要重做评论过滤。仍然真实的违规在测试/契约层:例如 test_skill_entrypoint_contract.py 断言 ensure all 5 restart-helper-managed daemons、## 📊 当前状态 — <phase>、## 🆘 状态卡片 — 共识机制无法继续收敛;test_ensure_project_rules_fixed_points.py 断言中文 human label 标题。这些都会在文案切换语言时假红或假绿。

最小边界是新增一条 SKILL.md 规则:程序检测只可依赖 public markers、AI sentinel、canonical label names、emoji banner prefixes、显式 markdown/html anchor id;不得依赖人类可读标题/句子。实现时只改现有测试断言和少量 anchor,LOC 约 +40 / -30,无迁移,无 CLAUDE/AGENTS/Tier 改动。

---

📎 完整 codex 原始输出(存档备查)

---

solver: minimal
issue: 106
cluster: issue-106
verdict: propose

Recommended framing

违规仍然真实,但范围比 issue 正文里的粗 grep 更窄: comment-monitor.sh 已经以 sentinel 为主判定,## 🤖 只是 legacy emoji-prefix 兜底,不需要重做;真正需要改的是 source-regression tests 和少量契约文档把中文标题/句子当成检测事实源。最小方案是在 SKILL.md 增加一条 language-neutral detection rule,明确 public markers / sentinel / label names / emoji banner prefixes / markdown anchor ids 是稳定协议 token,可字面断言;人类可读文案如 TL;DR、详细说明、状态卡片、下一步自动会做、ensure all 5 ... 不能作为程序检测锚点。随后只改现有测试断言,不引入 DetectionInvariantV1、新 schema、registry 或第二套检测系统。

Concrete plan

• Files:
  • skills/codex-refactor-loop/SKILL.md: 在 ## 工作语言规则(源码内英文,源码外中文) 下新增 Language-neutral detection anchors 小段;同时给中文语义标题前加稳定 anchor,例如 <a id="human-label-escalation-contract"></a>。
  • skills/codex-refactor-loop/REFERENCE.md: 如测试需要,给 status/escalation 模板中的人类标题加稳定 anchor 或保留现有 status-and-escalation-templates anchor;不改模板正文语义。
  • skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py: 把 ensure all 5 restart-helper-managed daemons、## 📊 当前状态 — <phase>、## 🆘 状态卡片 — 共识机制无法继续收敛 这类人类文案断言改成 script name / section anchor / protocol token 断言。
  • skills/codex-refactor-loop/scripts/test_ensure_project_rules_fixed_points.py: 把 ## \👤 human:需-maintainer-决策` 严格语义(强制)这类中文标题断言改为 anchor + canonical label token + required semantic tokens;新增一个小的 source-regression test,禁止assertIn/assertNotIn/assertEqual` 直接锚定 GitHub body prose headings/phrases,并 allowlist public protocol tokens。
• LOC delta: ~+40 / -30。
• Tests to add/modify:
  • Modify SkillEntrypointContractTests.test_first_wakeup_bootstrap_obligations_are_ordered_in_skill_alone。
  • Modify SkillEntrypointContractTests.test_heavy_reference_material_is_not_in_entrypoint。
  • Modify HumanLabelSemanticsTests.test_skill_md_has_strict_human_label_semantics_section。
  • Add one source-regression test in test_ensure_project_rules_fixed_points.py, roughly test_program_detection_does_not_anchor_on_human_prose。
• Philosophy/PROJECT_RULES/SPEC/Tier change: SKILL.md only, no CLAUDE.md / AGENTS.md / Tier file change. Current rule is Source files are English-only; external user-facing artifacts are 中文 by default. No mandatory parallel English section. Proposed additive text: Program detection must be language-neutral. Allowed literal anchors are public markers, the AI sentinel, canonical label names, stable emoji banner prefixes, and explicit markdown/html anchor ids. Tests and daemon filters must not depend on human-readable prose headings or sentences; when prose needs to move between Chinese and English, add or reuse a stable anchor instead. This is worth the trusted-base cost because it narrows an existing language rule without changing routing, labels, markers, or truth tables; deep consensus should hold because it preserves current Chinese default while making future English switch mechanically safe.
• Migration path: one PR rewrites the cited assertions to anchors/tokens and adds the guard test; no runtime migration needed.

Risks

• Public label names still contain Chinese (👤 human:需-maintainer-决策, 🤖 human:auto-推进). I treat them as protocol tokens under the existing “do not rename labels during this split” rule, not as switchable prose.
• A broad regex ban on Chinese in tests would create false positives because protocol labels and historical notes intentionally contain Chinese; the guard must be scoped to assertion arguments and known GitHub prose phrases.
• This does not implement a full i18n system for prompts; it only ensures tests/filters do not break when human-facing prose is later translated.
• Adding anchors to headings is a small trusted-base change; overusing anchors everywhere would become another registry, so only cited detection points should get them.

Escalation triggers (if any)

• none

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

• Why this is the minimum: the issue examples are current at test_skill_entrypoint_contract.py:87,108,109, test_ensure_project_rules_fixed_points.py:1826, and related source-regression assertions; changing these assertions to existing stable tokens/anchors removes the violation without changing runtime routing.
• What I considered but rejected: DetectionInvariantV1 is unnecessary for ~4-8 active assertion sites; deleting comment/banner filtering is unsafe because sentinel fallback for old comments and emoji status prefixes still protect comment sweeps.
• What I cannot decide alone: whether a future separate issue should rename public labels for full English UX; this issue should not rename labels because SKILL.md explicitly freezes labels during this split.

SOLVER_DONE:minimal:propose:replace human-prose detection anchors with stable protocol tokens and anchors, no new invariant registry

⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 meta-judge: issue 106 r1 收敛到 r2

TL;DR

• 三个 solver 都是 propose,违规真实。
• delete/minimal 同意不加 registry,迁移到 stable tokens / anchors。
• structural 提出 DetectionInvariantV1 registry + 更宽 runtime/doc scope。
• 因抽象、文件范围、LOC 和 runtime scope 不一致,本轮不是 consensus。
• 下一轮只收敛 registry 是否必要及具体边界。

---

详细说明

r1 的共同点是:程序检测不应依赖可本地化自然语言正文。分歧是实现边界。delete/minimal 都选择复用现有 ⟦AI:AUTO-LOOP⟧、role markers、EXIT=0、label token、daemon/script token 和 stable anchors;structural 则要求新增 DetectionInvariantV1 注册表、helper/CLI、$LOOP_OUTPUT_LANGUAGE 和多处 runtime rewiring。

因此按 Phase 9 meta-judge 规则,这是 3/3 propose but framings disagree,round 1 未触发 stall,应派 r2 收敛问题:是否必须新增 registry,还是用无 registry 的 stable token/anchor 边界解决。

---

📎 完整 meta-judge artifact

---

issue: 106
cluster: issue-106
convergence_round: 1
solver_verdicts:
minimal: propose
structural: propose
delete: propose
decision: converge

Decision

三位 solver 都确认 issue #106 真实存在,也都同意程序检测不应依赖 TL;DR、详细说明、中文状态标题、英文说明句等可本地化自然语言正文。尚未达成 3/3 consensus,因为 delete/minimal 的边界是“不新增 registry,把现有断言迁移到 sentinel / marker / EXIT=0 / label token / stable anchor”,而 structural 要新增 DetectionInvariantV1、registry/helper、$LOOP_OUTPUT_LANGUAGE、CLAUDE/AGENTS 规则和多处 runtime rewiring。文件范围、抽象命名、LOC 和 runtime scope 都不一致;当前是 round 1,没有 stall 条件,所以进入 round 2 收敛。

If converge

• Convergence question (specific): issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 是否应以“无 DetectionInvariantV1、只新增/复用 stable anchors + existing protocol tokens 并替换 source-regression prose assertions”的边界解决;若认为必须引入 registry,请列出当前已存在的 runtime caller 证据和为什么现有 sentinel/marker/label/anchor 不能覆盖。
• What each solver should address explicitly:
  • 是否新增 DetectionInvariantV1 / detection-invariants.json / helper CLI;必须回答 yes/no。
  • runtime scope 是否只限 source-regression tests + SKILL/REFERENCE anchor/rule,还是必须改 comment-monitor.sh、peek.sh、auto_release_gate.py、concurrency_monitor.py、check_skill_degradation.py。
  • rule scope 是 none、SKILL.md only,还是 CLAUDE.md / AGENTS.md 仓库级规则;必须给 exact clause/text。
  • legacy ## 🤖 / status emoji prefix 是否保留为 fallback,以及测试应断言 prefix、sentinel、anchor 还是 registry id。
  • LOC/files estimate must be comparable enough for meta-judge to apply the 30% variance rule.
• Round number this fires: 2

Round audit trail (links to local artifacts)

• solver-minimal: .refactor-loop/logs/phase9-issue106-r1-minimal.log
• solver-structural: .refactor-loop/logs/phase9-issue106-r1-structural.log
• solver-delete: .refactor-loop/logs/phase9-issue106-r1-delete.log

META_JUDGE_DONE:converge:round-2:decide-no-registry-stable-token-anchor-boundary-vs-DetectionInvariantV1-runtime-registry-scope

⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 meta-judge: issue 106 进入 r2 收敛

TL;DR

• 这是什么:Phase 9 meta-judge 对 design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 三个 r1 solver 的仲裁。
• 当前结论:不达成 consensus,进入 r2 convergence;分歧是是否需要 DetectionInvariantV1 registry + CLAUDE.md 规则。
• 下一步:controller 重新派 r2 三 solver,只回答 registry vs existing token/anchor boundary。

---

详细说明

3 个 solver 都是 propose,也都同意检测不应依赖中文/英文自然语言正文,应转向 marker、sentinel、stable token、anchor 或 schema。但 framing 没有 3/3 一致:minimal/delete 走现有 token/anchor 删除或替换 NL-literal assertion,structural 要新增 DetectionInvariantV1 registry/helper、迁移 5 个以上 caller,并改 CLAUDE.md / $LOOP_OUTPUT_LANGUAGE。LOC delta 从 +40/-30 或 -60~-120 net 到 +620/-240,超过 consensus 允许范围。

r1 小于 stalled 阈值 3,且分歧是具名技术问题,所以本轮不能 escalate。r2 收敛问题已收窄为:是否必须引入 DetectionInvariantV1,还是现有 marker/sentinel/label/script/schema/heading-anchor surfaces 已足以保持 detection 覆盖率不降。

---

📎 完整 codex 原始输出(存档备查)

---

issue: 106
cluster: issue-106
convergence_round: 1
solver_verdicts:
minimal: propose
structural: propose
delete: propose
decision: converge

Decision

本轮判 converge。3 个 solver 都确认 #106 违规真实,且都同意目标是让程序检测依赖 marker / sentinel / stable token / anchor / schema,而不是依赖中文或英文自然语言正文;但没有达到 3/3 同一 framing。minimal 与 delete 都拒绝新增 DetectionInvariantV1,主张在 SKILL.md/现有测试中把 NL-literal assertion 删除或迁移到现有机器 token/anchor;structural 则要求新增 registry/helper、迁移 5 个以上 caller,并修改 CLAUDE.md/AGENTS.md 与 $LOOP_OUTPUT_LANGUAGE。这是具名技术分歧,且 LOC delta 从约 +40/-30 或 -60~-120 net 到 +620/-240 差异超过 30%;当前 round=1,不满足 stalled 条件,所以不能 consensus,也不能 escalate。

If consensus

• Chosen framing: none
• Implement plan: none
• Philosophy/CLAUDE.md/SPEC/Tier changes included: none
• Implementation owner: not dispatched
• Add auto-loop-resume label to issue 97: no

If converge

• Convergence question (specific): 下一轮请只回答一个边界问题:design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 是否必须引入 DetectionInvariantV1 registry/helper 并修改 CLAUDE.md,还是现有 marker/sentinel/label/script/schema/heading-anchor surfaces 足以保持检测覆盖率不降;请给出同一最终文件清单、caller 迁移范围和测试口径。
• What each solver should address explicitly:
  • 逐项列出现有检测点分类:source-regression prose assertion、comment/banner filter、label/phase inference、daemon roster / forbidden-token drift check;说明哪些必须迁移,哪些只需删除或保留 legacy fallback。
  • 若选择 DetectionInvariantV1,必须说明为什么现有 marker/sentinel/label/script/schema/anchor surfaces 不够,并给出 registry 是否会成为第二事实源的防护测试。
  • 若拒绝 DetectionInvariantV1,必须说明如何防止未来新增 detection site 继续锚定自然语言正文,并给出 source-regression guard 的精确 allowlist / denylist。
  • 明确 comment-monitor.sh、peek.sh、auto_release_gate.py、concurrency_monitor.py、check_skill_degradation.py 是否需要 runtime 迁移;不能只讨论测试断言。
  • 明确规则落点:只改 skills/codex-refactor-loop/SKILL.md / REFERENCE.md,还是也改 CLAUDE.md / AGENTS.md;如改,给出被替换 invariant 与新 invariant。
  • 给出最终 LOC delta、behavior tests、source-regression tests,并确认 marker 词表、AI sentinel、Phase truth table、3/3 unanimous gate、public label 当前名字不变。
• Round number this fires: 2

If escalate

• Trigger category: none
• Why consensus failed to progress: not stalled; this is round 1 and the disagreement is the named implementation boundary between a new DetectionInvariantV1 registry plus repo-level clause, versus deleting/replacing NL-literal assertions on existing stable token/anchor surfaces.
• Suggested next step: dispatch round 2 solvers with the convergence question above.

Round audit trail (links to local artifacts)

• solver-minimal: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue106-r1-minimal.md
• solver-structural: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue106-r1-structural.md
• solver-delete: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue106-r1-delete.md

META_JUDGE_DONE:converge:round-2:Choose DetectionInvariantV1 registry plus CLAUDE rule versus existing token-anchor deletion boundary, with caller migration and coverage tests fixed
⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 minimal solver r2: 选择无 registry 的 token/anchor 边界

TL;DR

• 这是什么:issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 r2 minimal solver 对 registry 分歧的收敛回答。
• 结论:propose;不引入 DetectionInvariantV1,用现有 sentinel/marker/label/script/schema/emoji-prefix/anchor 替换 prose detection。
• 下一步:meta-judge 可与 structural/delete r2 对齐,若 3/3 收敛则派 implement。

---

详细说明

我复核了 issue body、r1 artifacts 和当前源码证据。当前需要迁移的核心是 source-regression tests 把 TL;DR、详细说明、状态卡片、ensure all 5 ... 等人类文案当成检测事实源。runtime 侧并不需要新 registry:comment-monitor.sh 已经 sentinel-first,emoji heading 只是旧评论 fallback;auto_release_gate.py / concurrency_monitor.py 用的是当前 GitHub public label token,不是可本地化正文。

最小落点是 SKILL.md 加一条 language-neutral detection rule,REFERENCE.md/section 加 stable anchor,测试改断言 marker/sentinel/label/script/schema/anchor,再加一个 source-regression guard 防止未来新增 prose assertion。这样不改 marker 词表、AI sentinel、Phase truth table、3/3 gate,也不改 public label 当前名字。

---

📎 完整 codex 原始输出(存档备查)

---

solver: minimal
issue: 106
cluster: issue-106
verdict: propose
round: 2
convergence_marker: META_JUDGE_DONE:converge:round-2:decide-no-registry-stable-token-anchor-boundary-vs-DetectionInvariantV1-runtime-registry-scope

Recommended framing

建议选择“无 registry 的 stable token / anchor 边界”。当前违规真实,但证据集中在 source-regression 把人类可读句子当契约,以及少量 comment/status 检测未把“sentinel 优先、emoji prefix 只是 legacy protocol fallback”写清楚;现有 ⟦AI:AUTO-LOOP⟧、SOLVER_DONE / META_JUDGE_DONE、EXIT=0、public label token、script name、schema key、markdown anchor 已足够覆盖检测语义。DetectionInvariantV1 会新增 JSON/helper/CLI 和第二事实源,但当前没有一个 runtime caller 需要动态语言映射才能保持正确性。

Concrete plan

• Files:
  • skills/codex-refactor-loop/SKILL.md: 在工作语言规则下新增 Language-neutral detection anchors 小段;明确程序检测只能依赖 marker、sentinel、public label token、script/schema token、stable emoji banner prefix、explicit markdown/html anchor id,不得依赖可本地化正文句子。给 strict human-label section 增加稳定 anchor,例如 <a id="human-label-escalation-contract"></a>。
  • skills/codex-refactor-loop/REFERENCE.md: 仅在需要被 source-regression 引用的 status / escalation 模板附近增加或复用 stable anchor;保留 ## 📊 / ## 🤖 / ## ✅ / ## 🆘 作为 legacy banner protocol prefix,不把后续中文标题作为检测事实源。
  • skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py: 把 ensure all 5 restart-helper-managed daemons 改为按顺序断言 restart-daemons.sh、五个 daemon executable names、daemon-event Monitor;把 status/escalation heavy-template 检查改为 anchor/category token,不再断言完整中文标题。
  • skills/codex-refactor-loop/scripts/test_ensure_project_rules_fixed_points.py: 把 strict human-label section 的中文 heading literal 改为 anchor + 👤 human:需-maintainer-决策 public label token + META_RESOLVED:escalate-human / helper guard tokens;把 sentinel tests 保留为 ⟦AI:AUTO-LOOP⟧ final-line behavior,删掉对 末尾独立一行、controller progress reporter、反模式 等人类文案的 detection 依赖。
  • skills/codex-refactor-loop/scripts/comment-monitor.sh: 不需要 registry/runtime migration;保留 sentinel-first 识别和 bounded AI prefix fallback。可把注释改成“legacy protocol prefixes”,并加一个小行为测试覆盖 sentinel-only AI comment 被跳过。
  • skills/codex-refactor-loop/scripts/peek.sh: 不需要 registry;如果改动,只把 duplicated prefix tuple 收敛为本文件内常量或注释,仍以 sentinel + bounded emoji prefixes 判定 AI/controller posts。
  • skills/codex-refactor-loop/scripts/auto_release_gate.py 和 skills/codex-refactor-loop/scripts/concurrency_monitor.py: 不迁移到 registry;👤 human:需-maintainer-决策 是当前 GitHub public label protocol token,不是可本地化正文。允许在局部常量命名为 HUMAN_DECISION_LABEL 以降低误解,但不改 label 名。
  • skills/codex-refactor-loop/scripts/check_skill_degradation.py: 不引入 invariant registry;若有 prose drift checks,只允许查 forbidden architecture tokens / protocol tokens / anchor ids,不查中文正文。
• LOC delta: ~+75 / -55。文档 anchors/rule 约 +20;测试替换和 guard 约 +55/-55;可选局部常量/comment 约 +0-20。
• Tests to add/modify:
  • Modify test_first_wakeup_bootstrap_obligations_are_ordered_in_skill_alone to assert daemon roster/script anchors instead of the English sentence.
  • Modify test_heavy_reference_material_is_not_in_entrypoint to assert reference anchors/categories instead of full Chinese banner titles.
  • Modify test_skill_md_has_strict_human_label_semantics_section to assert human-label-escalation-contract plus label/marker/helper tokens.
  • Modify sentinel/post tests to assert final sentinel token and --body-file, not Chinese explanatory prose.
  • Add one source-regression guard, e.g. test_program_detection_does_not_anchor_on_human_prose, scanning tests/scripts for assertIn/assertNotIn/assertEqual against known GitHub prose headings/sentences (TL;DR, 详细说明, 完整 codex 原始输出, 状态卡片, 当前状态, ensure all 5 restart-helper-managed daemons) while allowlisting protocol tokens (⟦AI:AUTO-LOOP⟧, role markers, public label names, emoji prefixes, script names, anchor ids).
  • Add or extend one behavior test for comment origin detection: body ending with sentinel is skipped even when the first heading is English or arbitrary prose; legacy ## 🤖 prefix remains skipped for old posts without sentinel.
• Philosophy/PROJECT_RULES/SPEC/Tier change (if any): skills/codex-refactor-loop/SKILL.md only; no CLAUDE.md / AGENTS.md, no SPEC/Tier/GPG. Current clause: Source files are English-only; external user-facing artifacts are 中文 by default. No mandatory parallel English section. Proposed additive text: Program detection is language-neutral: detectors, daemons, and source-regression tests may depend on public markers, the AI sentinel, public label tokens, stable emoji banner prefixes, script/schema tokens, and explicit markdown/html anchor ids; they must not depend on localized human-readable headings or prose. When prose needs to change language, add or reuse a stable anchor instead of asserting the prose. This is worth the trusted-base cost because it narrows the existing language rule without changing routing, labels, marker vocabulary, or lifecycle authority; consensus should hold because it preserves detection coverage and avoids a new fact source.
• Migration path: one implementation PR rewrites the cited assertions, adds anchors/rule/guard, and optionally names local constants for label/prefix tokens; no runtime data migration and no GitHub label rename.

Risks

• Public labels still contain Chinese text. Minimal framing treats them as protocol tokens because renaming labels would be a separate lifecycle migration and would violate the “public label 当前名字不变” convergence constraint.
• A broad ban on non-ASCII or Chinese in tests would be wrong; the guard must target detection assertions over human-facing prose, with an explicit allowlist for labels, markers, sentinels, emoji prefixes, scripts, schema keys, and anchors.
• Keeping legacy emoji-prefix fallback means some display-looking strings remain in filters. The bounded first-token prefix is a protocol compatibility surface; the human prose after it is not.
• This does not build full i18n for generated GitHub comments. It only prevents behavior/source-regression from depending on the current Chinese prose.

Escalation triggers (if any)

• none

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

• Why this is the minimum: the current concrete breakage can be removed by replacing prose assertions at the cited test sites and documenting the allowed detection anchors. No caller needs dynamic lookup: comment-monitor.sh already uses sentinel-first plus bounded prefixes; peek.sh can keep the same protocol prefixes; release/concurrency checks use public labels; phase routing already uses markers and EXIT=0.
• What I considered but rejected: DetectionInvariantV1 registry/helper/CLI because it adds another source for tokens already owned by SKILL/REFERENCE/tests/scripts, expands runtime scope to 5+ callers, and introduces registry drift tests just to prevent the abstraction from becoming a second truth source.
• What I cannot decide alone: exact final anchor ids and guard regex names; implementer can choose them as long as they are stable and the public marker/sentinel/phase truth table/3-of-3 consensus/public label names remain unchanged.

SOLVER_DONE:minimal:propose:no DetectionInvariantV1; replace prose detection with existing sentinel/marker/label/script/schema/emoji-prefix/anchor tokens plus a guard against future prose assertions

⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 delete solver r2: 选择无 registry 的删除边界

TL;DR

• 这是什么:issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 r2 delete solver 对 registry vs stable-token/anchor 的收敛回答。
• 结论:propose,删除/替换自然语言文案检测锚点,不新增 DetectionInvariantV1 registry。
• 下一步:meta-judge 可与 r2 minimal/structural 对齐,若 consensus 则派 implement 做 tests/docs 的窄迁移。

---

详细说明

我复核了当前 runtime caller:comment-monitor.sh 已经 sentinel-first,peek.sh 只是保留有界 legacy banner prefix,auto_release_gate.py / concurrency_monitor.py 使用 canonical label token,daemon roster 也有脚本名、DAEMON_NAMES、heartbeat/schema 面。它们不需要新增 registry 才能保持检测能力。

推荐 delete/collapse 边界是:只删除或替换 source-regression tests 中把 TL;DR、详细说明、中文状态标题、ensure all 5... 等人类文案当作机器事实源的断言;改断言现有 sentinel / marker / EXIT=0 / label / script / schema / explicit anchor。保留 legacy emoji-prefix fallback,但不把 prefix 后的本地化文案作为 detector。

---

📎 完整 codex 原始输出(存档备查)

---

solver: delete
issue: 106
cluster: issue-106
verdict: propose
convergence_marker: META_JUDGE_DONE:converge:round-2:decide-no-registry-stable-token-anchor-boundary-vs-DetectionInvariantV1-runtime-registry-scope

Classification

(d) Genuinely needed but over-built. The needed capability is language-neutral detection of controller/daemon/Phase contracts; the over-built part is treating human prose as detector authority, and the r1 structural proposal would over-correct by adding a second registry surface for facts already represented by existing protocol tokens, script names, schema keys, labels, sentinel, and markers.

Recommended action

建议选择无 DetectionInvariantV1 的 delete/collapse 边界:删除或替换 source-regression tests 里的自然语言文案字面锚点,把检测收敛到现有机器面。不要新增 detection-invariants.json、helper CLI、$LOOP_OUTPUT_LANGUAGE 或 runtime registry。保留现有 runtime 能力:AI 评论主判定靠 ⟦AI:AUTO-LOOP⟧,legacy 老评论靠有界 ## <emoji> prefix fallback;Phase 路由靠 EXIT=0 后的 marker;human/phase state 靠当前 canonical label token;daemon roster 靠脚本文件名、DAEMON_NAMES、heartbeat JSON 和 restart helper allowlist。

Concrete plan (if propose)

• Files to delete: none as whole files. Delete selected NL-prose assertion clauses inside skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py and skills/codex-refactor-loop/scripts/test_ensure_project_rules_fixed_points.py.
• New abstractions: none. Explicitly do not add DetectionInvariantV1, detection-invariants.json, detection_invariants.py, a registry CLI, or a new schema.
• Caller migrations:
  • Source-regression prose assertions: migrate. Replace exact prose checks such as ensure all 5 restart-helper-managed daemons, ## 📊 当前状态 — <phase>, ## 🆘 状态卡片 — 共识机制无法继续收敛, and ## \👤 human:需-maintainer-决策` 严格语义(强制)with stable anchors/tokens:restart-daemons.sh, the five daemon script names, daemon-event Monitor, Status Banners, status-and-escalation-templates, CANONICAL_HUMAN_LABELS, META_RESOLVED:escalate-human`, helper names, and marker allowlists.
  • Comment/banner filters: do not migrate to a registry. comment-monitor.sh is already sentinel-first via body contains ⟦AI:AUTO-LOOP⟧; keep its legacy ## 🤖 / ## 📊 / ## ✅ / ## 🆘 prefix fallback for historical posts without sentinel. Tests should assert sentinel-first behavior plus a bounded fallback prefix set, not localized headings after the emoji.
  • Label/phase inference: do not migrate to a registry. Current labels such as 👤 human:需-maintainer-决策 and 🤖 human:auto-推进 are protocol tokens in this repository, not switchable display prose. Keep exact label constants and helper guard tests.
  • Daemon roster / forbidden-token drift checks: collapse from prose phrases to existing machine identifiers. Assert script names, DAEMON_NAMES, heartbeat JSON keys, check_skill_degradation.py --static, ALERT_LOG, PENDING_EVENTS, and forbidden surface constants; remove checks whose only value is that a sentence or Chinese heading survived.
  • auto_release_gate.py, concurrency_monitor.py, peek.sh, check_skill_degradation.py: no runtime rewrite for issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106. Only adjust tests where they mistake prose for behavior. Runtime code already consumes label tokens, daemon names, signal keys, sentinel, or bounded legacy prefixes.
• Tests to delete: no full test file. Delete individual assertIn / assertNotIn / assertEqual checks that anchor on reader prose (TL;DR, 详细说明, 完整 codex 原始输出, 末尾独立一行, controller progress reporter, 反模式, exact Chinese banner titles, and English sentence fragments) unless the checked string is a protocol token.
• Tests to add/modify:
  • Modify SkillEntrypointContractTests.test_first_wakeup_bootstrap_obligations_are_ordered_in_skill_alone to assert the ordered daemon/script/Monitor anchors instead of the full English sentence.
  • Modify SkillEntrypointContractTests.test_heavy_reference_material_is_not_in_entrypoint to assert section/reference anchors rather than exact localized banner template text.
  • Modify HumanLabelSemanticsTests.test_skill_md_has_strict_human_label_semantics_section to assert a stable anchor plus label token and required route tokens, not the whole localized heading.
  • Add one source-regression guard, for example test_program_detection_does_not_anchor_on_human_prose, scoped to test assertion arguments. Allowlist public protocol tokens: SOLVER_DONE, META_JUDGE_DONE, EXIT=0, ⟦AI:AUTO-LOOP⟧, canonical label names, script names, schema keys, explicit markdown/html anchors, and bounded ## <emoji> legacy prefixes. Denylist known reader-prose anchors: TL;DR, 详细说明, 完整 codex 原始输出, 下一步, 状态卡片, and full prose sentences such as ensure all 5 restart-helper-managed daemons.
  • Add or keep a behavior test that a comment containing the sentinel is skipped regardless of language, and that a maintainer comment without sentinel/bot author/legacy AI prefix is not skipped.
• LOC delta: approximately +45 / -85, net -40. The net deletion comes from removing phrase lists and per-phrase assertions; replacements are smaller token/anchor assertions plus one scoped guard test.
• Philosophy/CLAUDE.md/SPEC/Tier changes (if any): none at repository level. Optional skill-local text only under skills/codex-refactor-loop/SKILL.md language policy:
  Program detection must use stable protocol tokens: markers, the AI sentinel, canonical label names, script/schema keys, bounded legacy AI banner prefixes, or explicit anchors. Tests and daemon filters must not depend on localized human-readable prose headings or sentences.
  This is not a Tier/SPEC change and does not alter marker vocabulary, AI sentinel, Phase truth table, 3/3 gate, or public label names.

Reverse-evidence (why this is safe to delete)

• comment-monitor.sh already uses ⟦AI:AUTO-LOOP⟧ as the primary AI-post detector before the legacy heading-prefix case. A registry would merely rename the sentinel and prefix constants without adding behavior.
• peek.sh filters maintainer comments by sentinel, bot author, and legacy banner prefixes; it does not require localized heading text after the emoji to preserve the safety property.
• auto_release_gate.py and concurrency_monitor.py depend on canonical labels and signal names. Those are protocol tokens under the current label taxonomy, not natural-language body copy.
• check_skill_degradation.py has some NL-ish required marker strings, but the load-bearing checks are named files, alert paths, CI job names, forbidden surface constants, and hook names. Issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 can remove reader-prose anchors there without introducing a registry.
• The structural registry would create another place to list facts already present in SKILL/REFERENCE/tests/scripts. That conflicts with the repository preference for a single mainline and narrow extension points, and it increases drift risk unless guarded by more source-regression tests.
• No public API or external host surface is removed. The GitHub comment style can remain Chinese by default; only machine detection stops treating that style as truth.

Risks

• Some prose assertions accidentally cover useful contracts. Replacements must preserve the underlying invariant: final sentinel, marker allowlist, EXIT=0 gate, human-label guard path, daemon roster, and forbidden lifecycle boundaries.
• The legacy emoji-prefix fallback should not be deleted in this issue because older AI comments may lack the sentinel. The deletion boundary is "no localized prose after the prefix as detector", not "remove all prefixes".
• A broad "no Chinese in tests" rule would be wrong because canonical labels and historical refactor comments intentionally contain Chinese. The new guard must be scoped to assertion arguments used as reader-prose anchors and must have a protocol-token allowlist.
• If the project later wants full output-language switching, that can be a separate design issue. This issue should not rename public labels or add display-language configuration.

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why I claim this can be deleted: every r2 detection category has an existing stable surface. Source-regression prose assertions can be deleted/replaced with anchors/tokens; runtime comment detection already has sentinel-first behavior; labels are protocol tokens; daemon count has script names and heartbeat/schema surfaces.
• What I checked to verify safety: issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 body/comments; r1 minimal/structural/delete/judge artifacts; comment-monitor.sh; peek.sh; auto_release_gate.py; concurrency_monitor.py; check_skill_degradation.py; test_skill_entrypoint_contract.py; test_ensure_project_rules_fixed_points.py; SKILL Phase 0, Sentinel, Label, language, and Phase routing clauses; git history showing recent churn around daemon wording.
• What I cannot decide alone: exact anchor names for the implement patch, but the delete boundary is precise: no registry, no runtime rewiring, delete prose-as-detector and replace with current protocol tokens/anchors.

SOLVER_DONE:delete:propose:delete localized prose detection anchors and use existing sentinel/marker/EXIT/label/script/schema/anchor surfaces; no DetectionInvariantV1 registry
⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 delete solver r2: 选择无 registry 的删除边界

TL;DR

• 这是什么:issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 r2 delete solver 对 registry vs stable-token/anchor 的收敛回答。
• 结论:propose,删除/替换自然语言文案检测锚点,不新增 DetectionInvariantV1 registry。
• 下一步:meta-judge 可与 r2 minimal/structural 对齐,若 consensus 则派 implement 做 tests/docs 的窄迁移。

---

详细说明

我复核了当前 runtime caller:comment-monitor.sh 已经 sentinel-first,peek.sh 只是保留有界 legacy banner prefix,auto_release_gate.py / concurrency_monitor.py 使用 canonical label token,daemon roster 也有脚本名、DAEMON_NAMES、heartbeat/schema 面。它们不需要新增 registry 才能保持检测能力。

推荐 delete/collapse 边界是:只删除或替换 source-regression tests 中把 TL;DR、详细说明、中文状态标题、ensure all 5... 等人类文案当作机器事实源的断言;改断言现有 sentinel / marker / EXIT=0 / label / script / schema / explicit anchor。保留 legacy emoji-prefix fallback,但不把 prefix 后的本地化文案作为 detector。

---

📎 完整 codex 原始输出(存档备查)

---

solver: delete
issue: 106
cluster: issue-106
verdict: propose
convergence_marker: META_JUDGE_DONE:converge:round-2:decide-no-registry-stable-token-anchor-boundary-vs-DetectionInvariantV1-runtime-registry-scope

Classification

(d) Genuinely needed but over-built. The needed capability is language-neutral detection of controller/daemon/Phase contracts; the over-built part is treating human prose as detector authority, and the r1 structural proposal would over-correct by adding a second registry surface for facts already represented by existing protocol tokens, script names, schema keys, labels, sentinel, and markers.

Recommended action

建议选择无 DetectionInvariantV1 的 delete/collapse 边界:删除或替换 source-regression tests 里的自然语言文案字面锚点,把检测收敛到现有机器面。不要新增 detection-invariants.json、helper CLI、$LOOP_OUTPUT_LANGUAGE 或 runtime registry。保留现有 runtime 能力:AI 评论主判定靠 ⟦AI:AUTO-LOOP⟧,legacy 老评论靠有界 ## <emoji> prefix fallback;Phase 路由靠 EXIT=0 后的 marker;human/phase state 靠当前 canonical label token;daemon roster 靠脚本文件名、DAEMON_NAMES、heartbeat JSON 和 restart helper allowlist。

Concrete plan (if propose)

• Files to delete: none as whole files. Delete selected NL-prose assertion clauses inside skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py and skills/codex-refactor-loop/scripts/test_ensure_project_rules_fixed_points.py.
• New abstractions: none. Explicitly do not add DetectionInvariantV1, detection-invariants.json, detection_invariants.py, a registry CLI, or a new schema.
• Caller migrations:
  • Source-regression prose assertions: migrate. Replace exact prose checks such as ensure all 5 restart-helper-managed daemons, ## 📊 当前状态 — <phase>, ## 🆘 状态卡片 — 共识机制无法继续收敛, and ## \👤 human:需-maintainer-决策` 严格语义(强制)with stable anchors/tokens:restart-daemons.sh, the five daemon script names, daemon-event Monitor, Status Banners, status-and-escalation-templates, CANONICAL_HUMAN_LABELS, META_RESOLVED:escalate-human`, helper names, and marker allowlists.
  • Comment/banner filters: do not migrate to a registry. comment-monitor.sh is already sentinel-first via body contains ⟦AI:AUTO-LOOP⟧; keep its legacy ## 🤖 / ## 📊 / ## ✅ / ## 🆘 prefix fallback for historical posts without sentinel. Tests should assert sentinel-first behavior plus a bounded fallback prefix set, not localized headings after the emoji.
  • Label/phase inference: do not migrate to a registry. Current labels such as 👤 human:需-maintainer-决策 and 🤖 human:auto-推进 are protocol tokens in this repository, not switchable display prose. Keep exact label constants and helper guard tests.
  • Daemon roster / forbidden-token drift checks: collapse from prose phrases to existing machine identifiers. Assert script names, DAEMON_NAMES, heartbeat JSON keys, check_skill_degradation.py --static, ALERT_LOG, PENDING_EVENTS, and forbidden surface constants; remove checks whose only value is that a sentence or Chinese heading survived.
  • auto_release_gate.py, concurrency_monitor.py, peek.sh, check_skill_degradation.py: no runtime rewrite for issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106. Only adjust tests where they mistake prose for behavior. Runtime code already consumes label tokens, daemon names, signal keys, sentinel, or bounded legacy prefixes.
• Tests to delete: no full test file. Delete individual assertIn / assertNotIn / assertEqual checks that anchor on reader prose (TL;DR, 详细说明, 完整 codex 原始输出, 末尾独立一行, controller progress reporter, 反模式, exact Chinese banner titles, and English sentence fragments) unless the checked string is a protocol token.
• Tests to add/modify:
  • Modify SkillEntrypointContractTests.test_first_wakeup_bootstrap_obligations_are_ordered_in_skill_alone to assert the ordered daemon/script/Monitor anchors instead of the full English sentence.
  • Modify SkillEntrypointContractTests.test_heavy_reference_material_is_not_in_entrypoint to assert section/reference anchors rather than exact localized banner template text.
  • Modify HumanLabelSemanticsTests.test_skill_md_has_strict_human_label_semantics_section to assert a stable anchor plus label token and required route tokens, not the whole localized heading.
  • Add one source-regression guard, for example test_program_detection_does_not_anchor_on_human_prose, scoped to test assertion arguments. Allowlist public protocol tokens: SOLVER_DONE, META_JUDGE_DONE, EXIT=0, ⟦AI:AUTO-LOOP⟧, canonical label names, script names, schema keys, explicit markdown/html anchors, and bounded ## <emoji> legacy prefixes. Denylist known reader-prose anchors: TL;DR, 详细说明, 完整 codex 原始输出, 下一步, 状态卡片, and full prose sentences such as ensure all 5 restart-helper-managed daemons.
  • Add or keep a behavior test that a comment containing the sentinel is skipped regardless of language, and that a maintainer comment without sentinel/bot author/legacy AI prefix is not skipped.
• LOC delta: approximately +45 / -85, net -40. The net deletion comes from removing phrase lists and per-phrase assertions; replacements are smaller token/anchor assertions plus one scoped guard test.
• Philosophy/CLAUDE.md/SPEC/Tier changes (if any): none at repository level. Optional skill-local text only under skills/codex-refactor-loop/SKILL.md language policy:
  Program detection must use stable protocol tokens: markers, the AI sentinel, canonical label names, script/schema keys, bounded legacy AI banner prefixes, or explicit anchors. Tests and daemon filters must not depend on localized human-readable prose headings or sentences.
  This is not a Tier/SPEC change and does not alter marker vocabulary, AI sentinel, Phase truth table, 3/3 gate, or public label names.

Reverse-evidence (why this is safe to delete)

• comment-monitor.sh already uses ⟦AI:AUTO-LOOP⟧ as the primary AI-post detector before the legacy heading-prefix case. A registry would merely rename the sentinel and prefix constants without adding behavior.
• peek.sh filters maintainer comments by sentinel, bot author, and legacy banner prefixes; it does not require localized heading text after the emoji to preserve the safety property.
• auto_release_gate.py and concurrency_monitor.py depend on canonical labels and signal names. Those are protocol tokens under the current label taxonomy, not natural-language body copy.
• check_skill_degradation.py has some NL-ish required marker strings, but the load-bearing checks are named files, alert paths, CI job names, forbidden surface constants, and hook names. Issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 can remove reader-prose anchors there without introducing a registry.
• The structural registry would create another place to list facts already present in SKILL/REFERENCE/tests/scripts. That conflicts with the repository preference for a single mainline and narrow extension points, and it increases drift risk unless guarded by more source-regression tests.
• No public API or external host surface is removed. The GitHub comment style can remain Chinese by default; only machine detection stops treating that style as truth.

Risks

• Some prose assertions accidentally cover useful contracts. Replacements must preserve the underlying invariant: final sentinel, marker allowlist, EXIT=0 gate, human-label guard path, daemon roster, and forbidden lifecycle boundaries.
• The legacy emoji-prefix fallback should not be deleted in this issue because older AI comments may lack the sentinel. The deletion boundary is "no localized prose after the prefix as detector", not "remove all prefixes".
• A broad "no Chinese in tests" rule would be wrong because canonical labels and historical refactor comments intentionally contain Chinese. The new guard must be scoped to assertion arguments used as reader-prose anchors and must have a protocol-token allowlist.
• If the project later wants full output-language switching, that can be a separate design issue. This issue should not rename public labels or add display-language configuration.

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why I claim this can be deleted: every r2 detection category has an existing stable surface. Source-regression prose assertions can be deleted/replaced with anchors/tokens; runtime comment detection already has sentinel-first behavior; labels are protocol tokens; daemon count has script names and heartbeat/schema surfaces.
• What I checked to verify safety: issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 body/comments; r1 minimal/structural/delete/judge artifacts; comment-monitor.sh; peek.sh; auto_release_gate.py; concurrency_monitor.py; check_skill_degradation.py; test_skill_entrypoint_contract.py; test_ensure_project_rules_fixed_points.py; SKILL Phase 0, Sentinel, Label, language, and Phase routing clauses; git history showing recent churn around daemon wording.
• What I cannot decide alone: exact anchor names for the implement patch, but the delete boundary is precise: no registry, no runtime rewiring, delete prose-as-detector and replace with current protocol tokens/anchors.

SOLVER_DONE:delete:propose:delete localized prose detection anchors and use existing sentinel/marker/EXIT/label/script/schema/anchor surfaces; no DetectionInvariantV1 registry
⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 structural solver r2 — stable token/anchor boundary

TL;DR

• 这是什么:issue design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 Phase 9 r2 structural solver 输出。
• 结论:不引入 DetectionInvariantV1 runtime registry;用现有 protocol token + explicit anchor + source-regression guard 解耦检测与文案。
• 下一步:meta-judge 可把 r2 三个 solver 对齐到同一实现边界。

---

详细说明

我复核了当前 detection sites:comment-monitor.sh / peek.sh 已经以 ⟦AI:AUTO-LOOP⟧ 为主判定,emoji heading 只是历史 fallback;auto_release_gate.py / concurrency_monitor.py 使用的是 canonical GitHub label token;check_skill_degradation.py 检查的是 named exception、forbidden protocol token 和 check/script name。把这些再包进 registry 会复制事实源,还会产生新的 drift 面。

推荐结构是 ProgramDetectionAnchorPolicy:SKILL/REFERENCE 写清 allowed detector classes,tests 增加 guard 拦截未来 assertIn / filter 继续锚定 TL;DR、中文标题、英文说明句等 localized prose。runtime 热路径不新增文件读取、不新增 daemon、不改 marker/label/truth table。

---

📎 完整 codex 原始输出(存档备查)

---

solver: structural
issue: 106
cluster: issue-106
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。

治理前置:架构性 / 流程性规则与对应机械验证手段同时进仓库,缺一不补口径。

Recommended framing

不引入 DetectionInvariantV1 runtime registry。结构性解法是把检测边界收敛到已经存在的 protocol surface:AI sentinel、role marker、EXIT=0、canonical GitHub label、daemon/script 名、heartbeat/state 文件、显式 markdown/html anchor;自然语言 heading / TL;DR / 中文说明句只作为 display prose。新增一个 skill-local source-regression guard,禁止未来 test / daemon filter 把人类文案句子作为检测真值;必要时在 SKILL.md / REFERENCE.md 放稳定 anchor,检测只断言 anchor + protocol token。这样避免 registry 成为第二事实源,同时把“不能 grep prose”机械落地。

Concrete plan

• New abstractions (if any): ProgramDetectionAnchorPolicy(文档规则 + source-regression guard,不是 runtime registry)。
  • Interface: allowed anchor classes = marker, sentinel, canonical_label, script_or_check_name, state_schema_key, markdown_anchor_id, legacy_ai_banner_prefix。
  • Layer + Project: skills/codex-refactor-loop skill contract / tests。
• Files:
  • skills/codex-refactor-loop/SKILL.md:在 Sentinel and Comment Sweep 或工作语言规则下增加一条:program detection must not depend on localized prose; allowed anchors are sentinel, markers, canonical labels, script/check names, state schema keys, explicit anchor ids, and bounded legacy AI banner prefixes.
  • skills/codex-refactor-loop/REFERENCE.md:给 status banner / human label semantics / daemon roster / skill degradation watch sections 补或复用显式 anchors;保留中文模板,但说明模板文案不是 detector authority。
  • skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py:把 ensure all 5 restart-helper-managed daemons、## 📊 当前状态 — <phase>、## 🆘 状态卡片 — 共识机制无法继续收敛 这类 prose literal 断言改为 daemon 文件名顺序、reference anchor、sentinel、controller-owned banner contract token。
  • skills/codex-refactor-loop/scripts/test_ensure_project_rules_fixed_points.py:把 ## \👤 human:需-maintainer-决策` 严格语义(强制)断言改为 anchor + canonical label + required semantic tokens;新增 guard 扫描assertIn/assertNotIn/assertEqual/assertRegex的字符串参数,拒绝TL;DR、详细说明、完整 codex 原始输出、当前状态、状态卡片`、长中文/英文 prose heading 等 detector anchors,allowlist marker/sentinel/labels/script/check names/anchor ids。
  • skills/codex-refactor-loop/scripts/comment-monitor.sh:不做 registry 迁移;保留 sentinel 主判定 + bounded legacy emoji-prefix fallback。可只更新注释,说明 prefix fallback 是历史 AI 评论兼容面,不是 prose detector authority。
  • skills/codex-refactor-loop/scripts/peek.sh:不做 registry 迁移;保留 sentinel + bounded prefix skip;将 tests 断言 sentinel-first / bot filter / bounded prefix tuple,不要求中文 banner 句子。
  • skills/codex-refactor-loop/scripts/auto_release_gate.py:不迁移。👤 human:需-maintainer-决策 和 ⏸️ phase:blocked 是 canonical protocol labels,不是可本地化 prose。
  • skills/codex-refactor-loop/scripts/concurrency_monitor.py:不迁移。PHASE_EXPECTED keys 和 human label equality 是 canonical label protocol;测试继续断言 label taxonomy。
  • skills/codex-refactor-loop/scripts/check_skill_degradation.py:不引入 registry;保留 named exception / forbidden-surface checks。若有长 prose marker,替换为 section anchor + forbidden token + script/check name;plugin registry、DegradationCheck、standalone watchdog 是 forbidden protocol token,不是 display copy。
• LOC delta:约 +90 / -55。
  • SKILL/REFERENCE anchor/rule text +25 / -5。
  • test assertion rewrites +35 / -45。
  • new source-regression guard +30 / -5。
  • runtime scripts 0 或注释级 +5 / 0。
• Tests to add:
  • Modify test_skill_entrypoint_contract.py::test_first_wakeup_bootstrap_obligations_are_ordered_in_skill_alone:断言五个 daemon executable names + restart-daemons.sh + daemon-event Monitor 的相对顺序,不再断言英文 prose phrase。
  • Modify test_skill_entrypoint_contract.py::test_heavy_reference_material_is_not_in_entrypoint:断言 REFERENCE.md#status-and-escalation-templates anchor 和 sentinel/banner role tokens,不再断言中文模板标题。
  • Modify HumanLabelSemanticsTests.test_skill_md_has_strict_human_label_semantics_section:使用 explicit anchor id 或 canonical label + Apply only when / DO NOT apply when / META_RESOLVED:escalate-human,不要求中文 heading literal。
  • Add test_program_detection_does_not_anchor_on_human_prose:扫描 source-regression tests 和 shell/python filters,fail on new natural-language prose anchors except allowlisted protocol labels, sentinel, markers, check names, script names, and bounded legacy prefixes.
  • Keep behavior tests for comment-monitor.sh / peek.sh sentinel and bot filtering unchanged or strengthen sentinel-first cases; no new registry behavior tests.
• proto changes (if any): none.
• Philosophy/PROJECT_RULES/SPEC/Tier changes (if any):skills/codex-refactor-loop/SKILL.md only。Current invariant:external artifacts 中文 by default and filters skip sentinel/banner prefixes。Proposed invariant text: Program detection is language-neutral: tests, daemon filters, and source-regression guards must target markers, the AI sentinel, canonical labels, script/check names, state schema keys, explicit anchor ids, or the bounded legacy AI banner-prefix set; they must not depend on localized human prose headings or sentences. Worth it because it narrows a skill-local detection rule without changing CLAUDE.md, marker vocabulary, labels, truth table, or lifecycle authority; consensus should hold because all r1 solvers agreed prose is not machine truth and the remaining disagreement was registry scope.
• Runtime cost:0 new daemon,0 new runtime file reads,0 latency hops,0 allocations on hot paths. Test-time guard cost:single-file AST/text scan,约 <50 ms in local pytest/unittest.

Risks

• 这比 r1 structural registry 少一个 central lookup,所以新增检测点仍可能手写 bad prose literal。用 guard 测试拦 assert* 和 filter prose anchors,并要求新增 anchor id 才能通过,把风险落到 CI。
• Public label names 含中文,但它们是 canonical protocol labels,不能在本 issue 顺手 i18n/rename;把 label token 放入 allowlist 是刻意保留协议稳定性。
• Legacy ## 🤖 / ## 📊 prefix fallback 仍是 display-looking token。保留原因是历史 AI 评论可能没有 sentinel;边界是 bounded prefix tuple,测试只断言 bounded set 和 sentinel-first,不依赖后续中文标题。
• 如果未来真的需要多语言输出选择器,应另开 issue 设计 rendering/i18n;本 issue 只解决 detection 与 prose 解耦。

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why this structure beats a documented exception:只写“不要依赖中文”会退化成 agent 记忆;本方案把 allowed detector classes 写入 SKILL,并用 source-regression guard 拦新增 prose anchors。
• Why I picked this abstraction over alternatives:DetectionInvariantV1 会复制 SKILL/REFERENCE/test 中已存在的事实源,还会碰到 goal: skills 规范化 + 防退化(脚本检测 SKILL.md / scripts / prompts 漂移) #66 明确禁止的 registry/protocol 扩张气味;当前 caller 已能用 sentinel、marker、label、script/check name、anchor id 覆盖。逐点裸改又缺未来防线,所以选“stable token/anchor + guard”的中间结构。
• What I cannot decide alone:是否未来把 public labels 也做语言无关 id + localized display;这会触及 GitHub label migration 和 user-facing UX,不属于 design: 程序检测不应依赖自然语言字面 — 支持中英切换 #106 的检测解耦最小闭环。

SOLVER_DONE:structural:propose:no DetectionInvariantV1 runtime registry; solve with stable protocol tokens, explicit anchors, and a source-regression guard against localized prose detectors

⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧

[loning]

🤖 issue 106 r2 meta-judge: consensus reached

TL;DR

• 这是什么:Phase 9 meta-judge 对 issue 106 的 r2 仲裁结果。
• 当前结论:3/3 solver unanimous,选择无 DetectionInvariantV1 的 stable token / explicit anchor / guard 边界。
• 下一步:controller 可按 artifact dispatch implement,并给 issue 106 加 auto-loop-resume。

---

详细说明

3 个 solver 在 r2 都从上一轮 registry 分歧收敛到同一实现边界:不新增 DetectionInvariantV1、registry/helper CLI、$LOOP_OUTPUT_LANGUAGE,也不迁移 comment-monitor.sh / peek.sh / release gate / concurrency monitor 到新事实源。程序检测只依赖现有机器面:⟦AI:AUTO-LOOP⟧、role marker、EXIT=0、canonical label token、daemon/script/schema token、explicit markdown/html anchor、bounded legacy emoji banner prefix。中文/英文标题、TL;DR、说明句只保留为 reader-facing prose,不能作为 detector authority。

solver	verdict	framing
minimal	propose	无 registry;用 sentinel/marker/label/script/schema/anchor 替换 prose 断言
structural	propose	ProgramDetectionAnchorPolicy:文档规则 + source-regression guard,非 runtime registry
delete	propose	删除/替换 localized prose detector anchors,保留现有检测能力

实施授权范围是 skills/codex-refactor-loop skill-local 变更:更新 SKILL.md 检测语言规则,给 REFERENCE.md 必要 section 加 anchor,重写两个 source-regression test 的自然语言字面断言,新增 guard 防止未来继续 grep prose。无 CLAUDE.md / AGENTS.md,无 SPEC/Tier,无 public label rename,无 marker/truth-table 变化。

---

📎 完整 codex 原始输出(存档备查)

---

issue: 106
cluster: issue-106
convergence_round: 2
solver_verdicts:
minimal: propose
structural: propose
delete: propose
decision: consensus

Decision

本轮判定 consensus reached。3/3 solver 在 r2 都选择同一边界:不引入 DetectionInvariantV1 / registry / helper CLI / $LOOP_OUTPUT_LANGUAGE,把程序检测收敛到现有稳定机器面,包括 marker、AI sentinel、EXIT=0、canonical label token、daemon/script/schema token、explicit markdown/html anchor、bounded legacy emoji banner prefix;同时改写 source-regression 里依赖自然语言正文/标题的断言,并新增 guard 防止未来继续把 TL;DR、中文标题、英文说明句等 localized prose 当检测事实源。minimal、structural、delete 对 runtime 文件迁移、label 不改名、legacy prefix fallback 保留、SKILL.md skill-local policy 和测试覆盖的选择没有实质冲突,LOC 估算差异只来自删除/替换断言的计数口径。

If consensus

• Chosen framing: structural
• Implement plan (verbatim copy from the winning solver's "Concrete plan" section)

Concrete plan

• New abstractions (if any): ProgramDetectionAnchorPolicy(文档规则 + source-regression guard,不是 runtime registry)。

  • Interface: allowed anchor classes = marker, sentinel, canonical_label, script_or_check_name, state_schema_key, markdown_anchor_id, legacy_ai_banner_prefix。
  • Layer + Project: skills/codex-refactor-loop skill contract / tests。

• Files:

  • skills/codex-refactor-loop/SKILL.md:在 Sentinel and Comment Sweep 或工作语言规则下增加一条:program detection must not depend on localized prose; allowed anchors are sentinel, markers, canonical labels, script/check names, state schema keys, explicit anchor ids, and bounded legacy AI banner prefixes.
  • skills/codex-refactor-loop/REFERENCE.md:给 status banner / human label semantics / daemon roster / skill degradation watch sections 补或复用显式 anchors;保留中文模板,但说明模板文案不是 detector authority。
  • skills/codex-refactor-loop/scripts/test_skill_entrypoint_contract.py:把 ensure all 5 restart-helper-managed daemons、## 📊 当前状态 — <phase>、## 🆘 状态卡片 — 共识机制无法继续收敛 这类 prose literal 断言改为 daemon 文件名顺序、reference anchor、sentinel、controller-owned banner contract token。
  • skills/codex-refactor-loop/scripts/test_ensure_project_rules_fixed_points.py:把 ## \👤 human:需-maintainer-决策` 严格语义(强制)断言改为 anchor + canonical label + required semantic tokens;新增 guard 扫描assertIn/assertNotIn/assertEqual/assertRegex的字符串参数,拒绝TL;DR、详细说明、完整 codex 原始输出、当前状态、状态卡片`、长中文/英文 prose heading 等 detector anchors,allowlist marker/sentinel/labels/script/check names/anchor ids。
  • skills/codex-refactor-loop/scripts/comment-monitor.sh:不做 registry 迁移;保留 sentinel 主判定 + bounded legacy emoji-prefix fallback。可只更新注释,说明 prefix fallback 是历史 AI 评论兼容面,不是 prose detector authority。
  • skills/codex-refactor-loop/scripts/peek.sh:不做 registry 迁移;保留 sentinel + bounded prefix skip;将 tests 断言 sentinel-first / bot filter / bounded prefix tuple,不要求中文 banner 句子。
  • skills/codex-refactor-loop/scripts/auto_release_gate.py:不迁移。👤 human:需-maintainer-决策 和 ⏸️ phase:blocked 是 canonical protocol labels,不是可本地化 prose。
  • skills/codex-refactor-loop/scripts/concurrency_monitor.py:不迁移。PHASE_EXPECTED keys 和 human label equality 是 canonical label protocol;测试继续断言 label taxonomy。
  • skills/codex-refactor-loop/scripts/check_skill_degradation.py:不引入 registry;保留 named exception / forbidden-surface checks。若有长 prose marker,替换为 section anchor + forbidden token + script/check name;plugin registry、DegradationCheck、standalone watchdog 是 forbidden protocol token,不是 display copy。

• LOC delta:约 +90 / -55。

  • SKILL/REFERENCE anchor/rule text +25 / -5。
  • test assertion rewrites +35 / -45。
  • new source-regression guard +30 / -5。
  • runtime scripts 0 或注释级 +5 / 0。

• Tests to add:

  • Modify test_skill_entrypoint_contract.py::test_first_wakeup_bootstrap_obligations_are_ordered_in_skill_alone:断言五个 daemon executable names + restart-daemons.sh + daemon-event Monitor 的相对顺序,不再断言英文 prose phrase。
  • Modify test_skill_entrypoint_contract.py::test_heavy_reference_material_is_not_in_entrypoint:断言 REFERENCE.md#status-and-escalation-templates anchor 和 sentinel/banner role tokens,不再断言中文模板标题。
  • Modify HumanLabelSemanticsTests.test_skill_md_has_strict_human_label_semantics_section:使用 explicit anchor id 或 canonical label + Apply only when / DO NOT apply when / META_RESOLVED:escalate-human,不要求中文 heading literal。
  • Add test_program_detection_does_not_anchor_on_human_prose:扫描 source-regression tests 和 shell/python filters,fail on new natural-language prose anchors except allowlisted protocol labels, sentinel, markers, check names, script names, and bounded legacy prefixes。
  • Keep behavior tests for comment-monitor.sh / peek.sh sentinel and bot filtering unchanged or strengthen sentinel-first cases; no new registry behavior tests。

• proto changes (if any): none。

• Philosophy/PROJECT_RULES/SPEC/Tier changes (if any):skills/codex-refactor-loop/SKILL.md only。Current invariant:external artifacts 中文 by default and filters skip sentinel/banner prefixes。Proposed invariant text: Program detection is language-neutral: tests, daemon filters, and source-regression guards must target markers, the AI sentinel, canonical labels, script/check names, state schema keys, explicit anchor ids, or the bounded legacy AI banner-prefix set; they must not depend on localized human prose headings or sentences. Worth it because it narrows a skill-local detection rule without changing CLAUDE.md, marker vocabulary, labels, truth table, or lifecycle authority; consensus should hold because all r1 solvers agreed prose is not machine truth and the remaining disagreement was registry scope。

• Runtime cost:0 new daemon,0 new runtime file reads,0 latency hops,0 allocations on hot paths. Test-time guard cost:single-file AST/text scan,约 <50 ms in local pytest/unittest。

• Philosophy/CLAUDE.md/SPEC/Tier changes included: skills/codex-refactor-loop/SKILL.md only, adding the skill-local invariant quoted above; no CLAUDE.md / AGENTS.md, no SPEC/Tier, no public label rename, no marker/truth-table change.

• Implementation owner: dispatch implement codex with cluster_id=issue-106, design_decision_path=/Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue106-r2-judge.md

• Add auto-loop-resume label to issue 106

Round audit trail (links to local artifacts)

• solver-minimal: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue106-r2-minimal.md
• solver-structural: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue106-r2-structural.md
• solver-delete: /Users/auric/consensus-rnd/.refactor-loop/runs/phase9-issue106-r2-delete.md

META_JUDGE_DONE:consensus:structural:no DetectionInvariantV1; use stable protocol tokens, explicit anchors, and source-regression guard against localized prose detectors

⟦AI:AUTO-LOOP⟧

⟦AI:AUTO-LOOP⟧