aloekun · aloekun · Apr 29, 2026 · Apr 29, 2026 · Apr 29, 2026 · Apr 29, 2026
diff --git a/.takt/facets/instructions/analyze-coderabbit.md b/.takt/facets/instructions/analyze-coderabbit.md
@@ -22,17 +22,21 @@ CodeRabbit sometimes raises findings that are not applicable to this project. Be
 2. For each finding, check:
    - **Platform scope**: This project targets Windows only. Findings about cross-platform compatibility (e.g., `.exe` hardcoding) are NOT applicable -- downgrade to `Info`
    - **Intentional design**: Check if the finding contradicts an ADR decision. If so, mark as `not_applicable`
-   - **Scope mismatch**: If the finding targets a read-only zone (`.takt/`, `docs/adr/`, `templates/`), mark as `not_applicable`
+   - **Sensitive-file protection** (Edit-blocked): If the finding targets `.claude/` (Claude Code sensitive-file protected — Edit/Write tool will refuse), mark as `user_decision_path` (NOT `not_applicable` — the issue may be real, but auto-fix cannot apply it)
+   - **Scope mismatch**: If the finding targets a read-only zone (`.takt/`, `docs/adr/`, `templates/`) or a non-source path (`.git/`, `.jj/`, `node_modules/`, `target/`), mark as `not_applicable`
    - **False positive**: If the finding misunderstands the code logic, mark as `not_applicable`
 
 Mark each finding as:
 - `applicable` -- genuine issue that should be addressed
+- `user_decision_path` -- finding is real but auto-fix is blocked by sensitive-file protection (`.claude/`); user decides
 - `not_applicable` -- does not apply to this project (with reason)
 
 ### Step 3: Severity classification
-For `applicable` findings only, classify by severity:
+For both `applicable` and `user_decision_path` findings, take the severity from CodeRabbit's `severity` field (do not reclassify):
 - Critical > High > Major > Medium > Minor > Low > Info
 
+The severity is preserved on `user_decision_path` findings so the user can prioritize their manual decision (a Critical `.claude/` finding still warrants attention even though auto-fix cannot apply it). For `not_applicable` findings, severity is irrelevant and may be omitted from the report.
+
 ### Step 4: Produce report and verdict
 
 ## Output Format
@@ -50,6 +54,11 @@ For `applicable` findings only, classify by severity:
 |---|-------------|-------|---------------|
 | 1 | path:line   | ...   | Platform scope: Windows only |
 
+### User Decision Path (sensitive-file protected)
+| # | File (Line) | Severity | Issue | Path Reason |
+|---|-------------|----------|-------|-------------|
+| 1 | .claude/... | Major    | ...   | sensitive-file protection — auto-fix blocked |
+
 ### Applicable Findings by Severity
 
 #### Critical / High / Major
@@ -69,12 +78,13 @@ For `applicable` findings only, classify by severity:
 
 - **approved**: No applicable findings, OR all applicable findings are Info/Low severity
   - Output: `approved` condition
-- **needs_fix**: Any applicable Critical, High, or Major finding exists
+- **needs_fix**: Any applicable Critical, High, or Major finding exists (excluding `user_decision_path`)
   - Output: `needs_fix` condition
   - These will be automatically fixed in the next step
-- **user_decision**: Only Medium or lower applicable findings exist (no Critical/High/Major)
+- **user_decision**: Only Medium or lower applicable findings exist, OR all remaining findings are `user_decision_path` (sensitive-file protected) regardless of severity
   - Output: `user_decision` condition
   - These are reported but NOT auto-fixed; the user decides
+  - **Important**: A `.claude/` finding of any severity routes here to prevent fix loop pathology (auto-fix would attempt 4+ Edit calls all blocked by sensitive-file protection, wasting iterations)
 
 ## Important
 

diff --git a/.takt/facets/instructions/review-security.md b/.takt/facets/instructions/review-security.md
@@ -31,3 +31,20 @@ Before evaluating the change, **read the following project documents** using the
    - Whether the change enables new privilege, data access, code execution, or prompt modification
 3. For each detected issue, classify it as blocking or non-blocking
 4. If there is even one blocking issue, judge as REJECT
+
+## Docs-only changes: trust boundary criterion
+
+For changes that touch **only documentation** (`docs/**`, ADRs, README, comments) and no executable code or configuration:
+
+- **Pass criterion**: If the change does NOT alter a trust boundary, judge as APPROVE without further security analysis
+- **Trust boundary unchanged** (APPROVE immediately):
+  - Policy explanations, terminology definitions, design rationale
+  - Workflow descriptions, ADR records of past decisions
+  - Reformatting, hierarchy reorganization, cross-reference updates
+- **Trust boundary changed** (apply full security review):
+  - Documentation of new authentication / authorization policies
+  - Redefinition of permission scopes or privilege boundaries
+  - Changes to documented secret handling, credential storage, or trust assumptions
+  - Specifications that other systems will rely on (API contracts, security guarantees)
+
+Rationale: documentation that does not redefine who-can-do-what cannot introduce security vulnerabilities by itself. Treating descriptive docs as security-relevant produces false-positive iterations and erodes review signal.
diff --git a/.takt/facets/instructions/review-simplicity.md b/.takt/facets/instructions/review-simplicity.md
@@ -20,6 +20,21 @@ Review ONLY the lines changed in the diff. Do NOT explore cross-file dependencie
 - **YAGNI violations**: Flag speculative abstractions, unused parameters, or over-engineered patterns that serve no current need
 - **Naming clarity**: Flag ambiguous variable/function names that obscure intent
 
+## Scope of DRY / YAGNI (do NOT raise findings outside this scope)
+
+The DRY and YAGNI criteria above apply **only to executable code logic**.
+
+- **DRY scope**: Flag duplicated *code logic* (copy-paste functions, repeated control flow, redundant computations). Do NOT flag:
+  - Documentation hierarchies that intentionally restate context (e.g., a summary table followed by detailed bullet points)
+  - Repetition between docs and code (docs explain, code executes — they serve different audiences)
+  - Test code mirroring production code structure (test independence > test DRY)
+- **YAGNI scope**: Flag *speculative code abstractions* (unused parameters, premature interfaces, over-engineered patterns in production code). Do NOT flag:
+  - Planning documents listing "future candidates", "Phase 2 検討", or "out of scope but worth considering" sections — these capture design intent for shared understanding, not speculative implementation
+  - ADR alternatives sections describing rejected options — these document the decision rationale
+  - Comments documenting *known constraints or limitations* of the current implementation (these are not speculation; they are recorded reality)
+
+If a finding cannot be tied to executable code logic, it is out of scope — do not raise it.
+
 ## Judgment procedure
 
 1. Read the diff from `.takt/review-diff.txt`

diff --git a/docs/adr/adr-030-deterministic-post-merge-feedback.md b/docs/adr/adr-030-deterministic-post-merge-feedback.md
@@ -281,6 +281,44 @@ GitHub 上に観測可能な成果物 (PR / tag / commit description) は一切
 
 `pnpm merge-pr` 自体は ADR-028 の対象 (PR マージは外部可視) だが、これは既存ゲートで管理済み。本 ADR で追加するのは merge **後** の post_steps のみ。
 
+### post-pr-review fix loop の対象外パス
+
+post-pr-review workflow の `analyze` step が CodeRabbit findings を分類する際、以下のパスに該当する finding は `user_decision_path` (severity に関わらず `user_decision` verdict 経路) に分類し、自動 fix loop に流さない。
+
+#### 対象外条件
+
+| カテゴリ | パスパターン | 理由 |
+|---------|-------------|------|
+| Claude Code sensitive-file protection | `.claude/**` | Edit/Write tool が refuse する。fix loop が回ると `fix.1` / `fix_supervisor.1-3` の 4 step が空費される pathological loop に陥る |
+| VCS 内部 | `.git/**`, `.jj/**` | バージョン管理系の内部ファイルは手作業対象外 |
+| 依存物 | `node_modules/**`, `target/**` | ビルド成果物 / 外部依存。リポジトリ内のソース変更で対応すべき |
+
+#### 採用根拠
+
+PR #91 で実証された pathological loop:
+
+- CodeRabbit が `.claude/` 配下のファイルへの finding を生成
+- `analyze` step が `needs_fix` 判定 → `fix` step 起動
+- `fix` step の Edit tool 呼び出しが Claude Code の sensitive-file protection でブロック
+- supervisor が「fix 失敗」と判定して再 fix トリガー、最大 4 iteration まで全失敗
+- 結果: 8 step (analyze + fix×4 + supervisor×3) が空費され、rate-limit を浪費し review feedback が遅延
+
+実装は `.takt/facets/instructions/analyze-coderabbit.md` の "Sensitive-file protection" / "Scope mismatch" filter として配置 (本 ADR は仕様のみ規定)。
+
+#### Verdict ルールの整合
+
+| Severity | 通常 path | `.claude/` etc. (対象外 path) |
+|----------|----------|-------------------------------|
+| Critical/High/Major | `needs_fix` (auto-fix) | `user_decision` (報告のみ、user に委ねる) |
+| Medium 以下 | `user_decision` | `user_decision` (同左) |
+
+`user_decision` 経路に流すことで、findings 自体は report に含まれユーザーが判断できる一方、fix loop は走らないため pathological loop が発生しない。findings の握りつぶしではなく **責任の所在を auto-fix から user に移す** 設計である。
+
+#### 関連 ADR
+
+- ADR-022 (責務分離) — Edit-blocked path を auto-fix 対象から除外することで、自動化と手動操作の境界が明確化
+- ADR-018 (post-pr-monitor 移行) — post-pr-review workflow が `.takt/facets/` 配下の facet で挙動制御される設計の前提
+
 ## 実装タスク
 
 詳細な実装手順は [`docs/todo.md`](../todo.md) の「マージ後フィードバック機構の決定論化」セクション Phase B-F を参照。本 ADR は仕様のみを規定する。

diff --git a/docs/todo.md b/docs/todo.md
@@ -25,17 +25,15 @@
 | 6 | 🚀 Tier 1 | ADR-032 PR-pre: GitHub Branch Protection 整備 | todo2.md | 設定のみ | なし (依存タスクは完了済) |
 | 7 | 🚀 Tier 1 | **PowerShell custom-lint-rule の `(?i)` フラグ自動検証 (PR #91 T1-1)** | todo3.md | S | なし (PR #91 直接対策、code-review.md 追記も同 PR で land) |
 | 8 | 🔧 Tier 2 | 週次レビュー (ADR-031) Phase B 実装 | todo.md | 中-高 | なし (順位 20 の compensating check 前提) |
-| 9 | 🔧 Tier 2 | reviewer facet 改善 (review-simplicity / review-security の DRY/YAGNI/security 軸明文化) | todo2.md | S | なし |
 | 10 | 🔧 Tier 2 | ADR-032 PR-broken-link: broken-link-check + 内部アンカー検査 統合 | todo2.md | Small-中 | なし (clean baseline 確立済) |
 | 11 | 🔧 Tier 2 | `cli-pr-monitor` プロセス正常終了の integration test (PR #85 T2-2) | todo2.md | S | なし |
 | 12 | 🔧 Tier 2 | **`cli-pr-monitor` ポーリング延長 + 重複起動ロック (PR #88 T2-4)** ★ rate-limit critical | todo3.md | Medium | なし (Polling anti-pattern 検出 (PR #86 T1-1, 完了済) と補完) |
 | 13 | 🔧 Tier 2 | **post-pr-review に rate-limit 自動検出 + 再トリガー (PR #89 T2-1)** ★ rate-limit critical | todo3.md | Medium | なし (順位 12 と補完) |
-| 14 | 🔧 Tier 2 | **post-pr-review fix loop の `.claude/` filter + ADR-030 制約明記 (PR #91 T2-1 + T3-2 Bundle)** ★ convergence | todo3.md | S + XS | なし (PR #91 直接対策、analyze facet + ADR 追記の同 PR Bundle) |
 | 15 | 🔧 Tier 2 | **cli-pr-monitor 通知 Recovery 経路 (SessionStart hook 拡張)** ★ silent loss prevention | todo3.md | S/M | なし (ADR-030 L2 recovery パターンを cli-pr-monitor に適用) |
 | 16 | 🔧 Tier 2 | **`vitest` を devDependencies に固定 (PR #88 T2-3)** | todo3.md | Small | なし |
 | 17 | 🔧 Tier 2 | **`pnpm create-pr` 必須引数ヘルプ改善 (PR #88 T2-5)** | todo3.md | Small | なし |
 | 18 | 🔧 Tier 2 | **`.failed` marker への recovery 手順自己文書化 (PR #90 T2-2)** | todo3.md | S | なし |
-| 19 | 🔧 Tier 2 | **takt ハーネスの `REJECT-ESCALATE` terminal verdict 実装 (PR #91 T2-2)** | todo3.md | M | 順位 14 (path-based 解決) land 後推奨 |
+| 19 | 🔧 Tier 2 | **takt ハーネスの `REJECT-ESCALATE` terminal verdict 実装 (PR #91 T2-2)** | todo3.md | M | post-pr-review fix loop の `.claude/` filter (Bundle T、完了済) land 後推奨 |
 | 20 | 💎 Tier 3 | ADR-032 PR-β: 実装 (enabled=false default) | todo2.md | 中-高 | 6, 8, 10 |
 | 21 | 💎 Tier 3 | ADR-032 PR-γ: enablement (1 行 flip) | todo2.md | XS | 順位 8 dogfood + 順位 20 |
 | 22 | 💎 Tier 3 | ADR-032 PR-δ: dogfood + メトリクス検証 | todo2.md | (運用) | 順位 21 |
@@ -45,18 +43,21 @@
 | 26 | 💎 Tier 3 | **post-pr-monitor polling 禁止のグローバル明文化 (PR #86 T3-2)** | todo2.md | XS | なし |
 | 27 | 🧹 Tier 4 | ADR-030 Phase E/F: 旧機構廃止 + dogfood | todo.md | 中 | なし (cleanup) |
 | 28 | ⏳ Tier 5 | (追って) ADR-030 の takt-test-vc 反映 | todo.md | 中 | 順位 27 Phase F |
+| 29 | 🚀 Tier 1 | **非 docs ファイル `docs/todo` 参照検出 lint rule (PR #94 T1-1) ★ Bundle U** | todo3.md | S | なし (PR #94 直接対策、Cross-File Reference Lifecycle の決定論的防止層) |
+| 30 | 💎 Tier 3 | **Cross-File Reference Lifecycle ルールに具体例追記 (PR #94 T3-2) ★ Bundle U** | todo3.md | XS | なし (Bundle U で 29 と並行 land 推奨) |
 
 **戦略**: Tier 1 を 2〜3 セッションで片付け → Tier 2 で ADR-032 の前提 + rate-limit + convergence cost 削減を進める → Tier 3 で ADR-032 を land + ドキュメント整備。Tier 4-5 は cleanup / 外部展開で daily efficiency への直接効果は小さい。
 
 **Bundle 1 完了 + post-merge-feedback 反映 (2026-04-29)**: PR #91 (Bundle 1: PowerShell + Markdown anchor lint rules) merge 後の post-merge-feedback で **4 件の新規 task を追加** (PowerShell `(?i)` 自動検証 / `.claude/` filter + ADR-030 制約 / cli-pr-monitor 通知 Recovery 経路 / takt REJECT-ESCALATE)。**前 2 件は本 PR で実証された「fix iteration の根因」に対する決定論的防止策で最優先候補**。**日付ベース見出しアンカーのグローバル明文化 task は決定論的防止 (no-mutable-anchor rule) との二重防衛として継続有効**。
 
-**reviewer facet 改善 task は全 PR の review 精度を即時向上させ、Tier 2 内で 週次レビュー Phase B / ADR-032 PR-broken-link / cli-pr-monitor exit test と並列実施可能**。
+**reviewer facet 改善 (Bundle T で land 済)** + **post-pr-review fix loop の `.claude/` filter (Bundle T で land 済)** が完了し、reviewer 精度向上 + convergence cost 削減の二段構えが成立。残る Tier 2 では rate-limit critical 系 (cli-pr-monitor ポーリング延長 / post-pr-review rate-limit 自動検出) を最優先候補とする。
 **rate-limit 系の 2 タスク (cli-pr-monitor ポーリング延長 + 重複起動ロック / post-pr-review rate-limit 自動検出 + 再トリガー) は rate-limit 直撃のため Tier 2 内で最優先候補**。前者 = ポーリング頻度全体の削減、後者 = review 単位での自動再トリガー、Polling anti-pattern 検出 (PR #86 T1-1、完了済) を含む 3 層で rate-limit を抑制する設計。
-**post-pr-review fix loop の `.claude/` filter + Recovery 経路 (SessionStart hook 拡張) は本 PR #91 の直接観測知見**。前者 = path-based filter で 8 step 空費の pathological loop を防止 / 後者 = SessionStart hook で再起動跨ぎの通知ロスト防止。
+**cli-pr-monitor 通知 Recovery 経路 (SessionStart hook 拡張) は PR #91 の直接観測知見**。SessionStart hook で再起動跨ぎの通知ロスト防止。post-pr-review fix loop の `.claude/` filter (path-based 解決) は Bundle T で land 済。
 **Stop hook の `pnpm lint:md` 統合 task は Markdown linter hook 統合 (PR #88 で merged) の gap closure**。**AI 生成一時スクリプト pattern 検出は push 前 untracked `__*` hook (PR #85 T1-4) と関連** (実装前に擦り合わせ要)。
 **`.failed` marker 自己文書化 task は ADR-030 soft-fail 機構の運用負荷削減** (PR #89 セッションで recovery が機能した実証から派生、Effort S)。
-**takt REJECT-ESCALATE は post-pr-review fix loop の `.claude/` filter task の verdict-based 一般解**。path-based 解決の land 後に着手することで、補完関係になる。
+**takt REJECT-ESCALATE は post-pr-review fix loop の `.claude/` filter (Bundle T で land 済) の verdict-based 一般解**。path-based 解決が完了したので、本 task 着手で補完関係を完成させる。
 **T3 グローバルルール 4 件 (日付ベース見出しアンカー / jj conflict リカバリ / `__` prefix scratch / post-pr-monitor polling 禁止) は `~/.claude/` 配下への XS 追記なので並列実施推奨**。
+**Bundle U (Cross-File Reference Lifecycle 強化) は PR #94 post-merge-feedback 直接対策**。非 docs ファイル `docs/todo` 参照検出 lint rule (Tier 1 = 決定論的防止) と Cross-File Reference Lifecycle ルール具体例追記 (Tier 3 = preventive guidance) を 1 PR で land 推奨。両者は同一テーマ (永続成果物→ephemeral 参照禁止の二層化) で補完関係、effort 合計 S+XS。
 
 ---