feat: rules/ を現行モデル観点でスリム化（#117、v0.67.0）

CHANGELOG.md

@@ -4,6 +4,17 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/), and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [0.67.0] - 2026-06-12
+
+LLM 性能監査（#121）P4 の第 1 弾（#117）。常時注入される rules/ から現行モデルに情報量のない行を削除した（10 → 8 ファイル）。
+
+### Removed
+- **rules/patterns.md / rules/hooks.md を廃止（#117）**: patterns.md は「既存パターン踏襲」（harness 標準と重複）と「詳細例は skills へ」（ルール作者向けメタ指示の誤配置）を削除し、有効な 2 行（抽象化の閾値・構造化パーサ優先）を coding-style.md へ統合。hooks.md は hook 作成時しか意味を持たないリファレンスの常時注入だったため、要点をリポジトリ CLAUDE.md の Hook Fragment Assembly 節へ移設（SessionStart/SessionEnd の用途限定行はキット自身の利用実態と矛盾するため削除）。既存配備分は update の retired 掃除で除去
+- **行単位の削除（#117）**: anti-patterns.md「Load only task-relevant files…」（小コンテキスト時代の遺物）、agents.md の 8 領域列挙と並列実行・スコープ最小化行（agents/*.md の description と harness 標準の再記述）、coding-style.md の immutability 選好（JS テンプレ由来の越境。TS/JS スコープでは skills/coding-standards がカバー）とデフォルト挙動の再掲 3 行、permissions-guide.md の回帰テスト義務（キット開発規約のグローバル漏出 — リポジトリ CLAUDE.md へ移設）
+
+### Changed
+- **rules/agents.md を委譲閾値基準に書換（#117）**: 「Delegate to a subagent only when the task is large enough to amortize the overhead, or benefits from isolation or parallelism」+ 出力照合の 2 行に縮小（反証検証で「残す」と判定された L6 は維持）
+
 ## [0.66.0] - 2026-06-12
 
 LLM 性能監査（#121）P3 の最終弾（#116）。skills の一律強制を縮小し内容の鮮度を更新した。

CLAUDE.md

@@ -244,6 +244,8 @@ Rules to remember: verify fresh install, `setup.sh --update`/`/update-kit`, and
 
 Multiple features can safely use the same hook type (e.g., `PreCompact`, `PostCompact`) — `merge_deep()` concatenates arrays instead of replacing them.
 
+**Hook authoring conventions** (formerly distributed as `rules/hooks.md`): treat hook stdout/stderr as user-visible control surfaces and keep output concise; prefer external hook scripts over inline JSON shell when logic is nontrivial; filter by matcher first, then validate the real tool input schema inside the script. New hooks.json fragments and permissions.json changes need a regression test under `tests/unit/` (follow the `test-hook-fixtures.sh` style).
+
 **Hook ordering matters**: `safety-net` must be the **first** entry in `hook_fragments[]` so its `PreToolUse` entry appears at index 0 (runs before other PreToolUse hooks). When adding new PreToolUse hooks, append after safety-net.
 
 **`build_settings_file()`** in `lib/deploy.sh` is the single unified settings.json builder. It accepts an output path parameter and is called by both fresh install and update paths.

README.en.md

@@ -51,7 +51,7 @@ Claude Code Starter Kit bootstraps a consistent, high-quality Claude Code enviro
 
 - **3 profiles**: Minimal, Standard (recommended), Full
 - **9 agents**: planner, architect, tdd-guide, code-reviewer, security-reviewer, build-error-resolver, e2e-runner, refactor-cleaner, doc-updater
-- **10 rules**: coding-style, git-workflow, hooks, patterns, performance, security, testing, agents, anti-patterns, permissions-guide
+- **8 rules**: coding-style, git-workflow, performance, security, testing, agents, anti-patterns, permissions-guide
 - **20 slash commands**: /plan, /tdd, /build-fix, /e2e, /verify, /research, /web-article, /oss-analyze, /web-source-review, /handover, /update-kit, and more
 - **12 skill modules**: backend-patterns, frontend-patterns, security-review, tdd-workflow, prompt-patterns, and more
 - **11 optional hooks/settings**: safety net (cc-safety-net), auto update, web content update, tmux reminder, doc blocker, Prettier or Biome formatting, PR creation log, pre-compact snapshot (opt-in), statusline, doc size guard, feature recommendation
@@ -392,7 +392,7 @@ claude-code-starter-kit/
 │   ├── en/                 # English templates & strings
 │   └── ja/                 # Japanese templates & strings
 ├── agents/                 # Agent definitions (9 files)
-├── rules/                  # Rule files (10 files)
+├── rules/                  # Rule files (8 files)
 ├── commands/               # Slash commands (20 files)
 └── skills/                 # Skill modules (12 dirs)
 ```

README.md

@@ -1103,7 +1103,7 @@ claude-code-starter-kit/
 │   ├── en/                 # 英語
 │   └── ja/                 # 日本語
 ├── agents/                 # AI エージェント定義（9種）
-├── rules/                  # コーディングルール（10種）
+├── rules/                  # コーディングルール（8種）
 ├── commands/               # スラッシュコマンド（20個）
 └── skills/                 # スキルモジュール（12個）
 ```

rules/agents.md

@@ -1,6 +1,4 @@
 # Agents
 
-- Use specialized agents for clearly complex planning, architecture, review, security, TDD, E2E, cleanup, or docs work.
-- Run independent agent/review work in parallel when it reduces wait time.
-- Keep each agent prompt scoped to one responsibility and the minimum context needed.
+- Delegate to a subagent only when the task is large enough to amortize the overhead, or benefits from isolation or parallelism.
 - After agent output, reconcile findings yourself before changing code or reporting completion.

rules/anti-patterns.md

@@ -4,4 +4,3 @@
 - Implement the requested scope; ask before adding adjacent features.
 - Do not claim completion until required tests or checks have actually passed.
 - If the same approach fails repeatedly, stop and choose a different approach or report the blocker.
-- Load only task-relevant files and summarize large error output.

rules/coding-style.md

@@ -1,7 +1,5 @@
 # Coding Style
 
-- Prefer immutable updates over mutating objects or arrays in place.
-- Keep files focused; split large modules when a real boundary appears.
 - Validate external input and handle errors with useful messages.
-- Keep functions small enough to review, with shallow control flow.
-- Avoid hardcoded values, stray debug output, and unrelated refactors.
+- Add abstraction only when it removes meaningful duplication or complexity.
+- Prefer structured parsers/APIs over ad-hoc string manipulation.

rules/permissions-guide.md

@@ -2,4 +2,3 @@
 
 - Keep allowlists narrow and tied to real workflows.
 - Do not bypass safety hooks by changing commands unless the user explicitly approves.
-- When adding hooks or permissions, add a regression test for the intended tool schema.