ブランチ名生成戦略を設定可能に

デフォルトをローマ字化(高速)に変更し、AI生成が必要な場合は config.yaml で branchNameStrategy: ai を設定可能にした。これによりブランチ名生成の待ち時間を削減し、LLMコストも削減できる。

また、coder エージェントに「根本原因修正後の安全機構迂回は禁止」ルールを追加した。

resources/global/en/agents/default/coder.md

@@ -24,6 +24,7 @@ You are the implementer. **Focus on implementation, not design decisions.**
 - Making design decisions arbitrarily → Report and ask for guidance
 - Dismissing reviewer feedback → Prohibited (your understanding is wrong)
 - **Adding backward compatibility or legacy support without being asked → Absolutely prohibited (fallbacks, old API maintenance, migration code, etc. are unnecessary unless explicitly instructed)**
+- **Layering workarounds that bypass safety mechanisms on top of a root cause fix → Prohibited (e.g., fixing path resolution AND adding `git add -f` to override `.gitignore`. If the root fix is correct, the bypass is unnecessary. Safety mechanisms exist for a reason)**
 
 ## Most Important Rule
 

resources/global/ja/agents/default/coder.md

@@ -24,6 +24,7 @@
 - 設計判断を勝手にする → 報告して判断を仰ぐ
 - レビュワーの指摘を軽視する → 禁止（あなたの認識が間違っている）
 - **後方互換・Legacy対応を勝手に追加する → 絶対禁止（フォールバック、古いAPI維持、移行期コードなど、明示的な指示がない限り不要）**
+- **根本原因を修正した上で安全機構を迂回するワークアラウンドを重ねる → 禁止（例: パス解決を直したのに `.gitignore` を無視する `git add -f` も追加する。根本修正が正しいなら追加の迂回は不要。安全機構は理由があって存在する）**
 
 ## 最重要ルール
 

src/__tests__/summarize.test.ts

@@ -44,6 +44,7 @@ beforeEach(() => {
     logLevel: 'info',
     provider: 'claude',
     model: undefined,
+    branchNameStrategy: 'ai',
   });
 });
 
@@ -162,13 +163,14 @@ describe('summarizeTaskName', () => {
   });
 
   it('should use provider from config.yaml', async () => {
-    // Given: config has codex provider
+    // Given: config has codex provider with branchNameStrategy: 'ai'
     mockLoadGlobalConfig.mockReturnValue({
       language: 'ja',
       defaultPiece: 'default',
       logLevel: 'info',
       provider: 'codex',
       model: 'gpt-4',
+      branchNameStrategy: 'ai',
     });
     mockProviderCall.mockResolvedValue({
       agent: 'summarizer',
@@ -252,19 +254,110 @@ describe('summarizeTaskName', () => {
     expect(result).not.toMatch(/^-|-$/); // No leading/trailing hyphens
   });
 
-  it('should use LLM by default', async () => {
-    // Given
+  it('should use romaji by default', async () => {
+    // Given: branchNameStrategy is not set (undefined)
+    mockLoadGlobalConfig.mockReturnValue({
+      language: 'ja',
+      defaultPiece: 'default',
+      logLevel: 'info',
+      provider: 'claude',
+      model: undefined,
+      branchNameStrategy: undefined,
+    });
+
+    // When: useLLM not specified, branchNameStrategy not set
+    const result = await summarizeTaskName('test task', { cwd: '/project' });
+
+    // Then: should NOT call provider, should return romaji
+    expect(mockProviderCall).not.toHaveBeenCalled();
+    expect(result).toMatch(/^[a-z0-9-]+$/);
+  });
+
+  it('should use AI when branchNameStrategy is ai', async () => {
+    // Given: branchNameStrategy is 'ai'
+    mockLoadGlobalConfig.mockReturnValue({
+      language: 'ja',
+      defaultPiece: 'default',
+      logLevel: 'info',
+      provider: 'claude',
+      model: undefined,
+      branchNameStrategy: 'ai',
+    });
     mockProviderCall.mockResolvedValue({
       agent: 'summarizer',
       status: 'done',
-      content: 'add-auth',
+      content: 'ai-generated-slug',
       timestamp: new Date(),
     });
 
-    // When: useLLM not specified (defaults to true)
-    await summarizeTaskName('test', { cwd: '/project' });
+    // When: useLLM not specified, branchNameStrategy is 'ai'
+    const result = await summarizeTaskName('test task', { cwd: '/project' });
 
     // Then: should call provider
     expect(mockProviderCall).toHaveBeenCalled();
+    expect(result).toBe('ai-generated-slug');
+  });
+
+  it('should use romaji when branchNameStrategy is romaji', async () => {
+    // Given: branchNameStrategy is 'romaji'
+    mockLoadGlobalConfig.mockReturnValue({
+      language: 'ja',
+      defaultPiece: 'default',
+      logLevel: 'info',
+      provider: 'claude',
+      model: undefined,
+      branchNameStrategy: 'romaji',
+    });
+
+    // When
+    const result = await summarizeTaskName('test task', { cwd: '/project' });
+
+    // Then: should NOT call provider
+    expect(mockProviderCall).not.toHaveBeenCalled();
+    expect(result).toMatch(/^[a-z0-9-]+$/);
+  });
+
+  it('should respect explicit useLLM option over config', async () => {
+    // Given: branchNameStrategy is 'romaji' but useLLM is explicitly true
+    mockLoadGlobalConfig.mockReturnValue({
+      language: 'ja',
+      defaultPiece: 'default',
+      logLevel: 'info',
+      provider: 'claude',
+      model: undefined,
+      branchNameStrategy: 'romaji',
+    });
+    mockProviderCall.mockResolvedValue({
+      agent: 'summarizer',
+      status: 'done',
+      content: 'explicit-ai-slug',
+      timestamp: new Date(),
+    });
+
+    // When: useLLM is explicitly true
+    const result = await summarizeTaskName('test task', { cwd: '/project', useLLM: true });
+
+    // Then: should call provider (explicit option overrides config)
+    expect(mockProviderCall).toHaveBeenCalled();
+    expect(result).toBe('explicit-ai-slug');
+  });
+
+  it('should respect explicit useLLM false over config with ai strategy', async () => {
+    // Given: branchNameStrategy is 'ai' but useLLM is explicitly false
+    mockLoadGlobalConfig.mockReturnValue({
+      language: 'ja',
+      defaultPiece: 'default',
+      logLevel: 'info',
+      provider: 'claude',
+      model: undefined,
+      branchNameStrategy: 'ai',
+    });
+
+    // When: useLLM is explicitly false
+    const result = await summarizeTaskName('test task', { cwd: '/project', useLLM: false });
+
+    // Then: should NOT call provider (explicit option overrides config)
+    expect(mockProviderCall).not.toHaveBeenCalled();
+    expect(result).toMatch(/^[a-z0-9-]+$/);
   });
 });

src/core/models/global-config.ts

@@ -59,6 +59,8 @@ export interface GlobalConfig {
   bookmarksFile?: string;
   /** Path to piece categories file (default: ~/.takt/preferences/piece-categories.yaml) */
   pieceCategoriesFile?: string;
+  /** Branch name generation strategy: 'romaji' (fast, default) or 'ai' (slow) */
+  branchNameStrategy?: 'romaji' | 'ai';
 }
 
 /** Project-level configuration */

src/core/models/schemas.ts

@@ -268,6 +268,8 @@ export const GlobalConfigSchema = z.object({
   bookmarks_file: z.string().optional(),
   /** Path to piece categories file (default: ~/.takt/preferences/piece-categories.yaml) */
   piece_categories_file: z.string().optional(),
+  /** Branch name generation strategy: 'romaji' (fast, default) or 'ai' (slow) */
+  branch_name_strategy: z.enum(['romaji', 'ai']).optional(),
 });
 
 /** Project config schema */

src/infra/config/global/globalConfig.ts

@@ -87,6 +87,7 @@ export class GlobalConfigManager {
       minimalOutput: parsed.minimal_output,
       bookmarksFile: parsed.bookmarks_file,
       pieceCategoriesFile: parsed.piece_categories_file,
+      branchNameStrategy: parsed.branch_name_strategy,
     };
     this.cachedConfig = config;
     return config;
@@ -143,6 +144,9 @@ export class GlobalConfigManager {
     if (config.pieceCategoriesFile) {
       raw.piece_categories_file = config.pieceCategoriesFile;
     }
+    if (config.branchNameStrategy) {
+      raw.branch_name_strategy = config.branchNameStrategy;
+    }
     writeFileSync(configPath, stringifyYaml(raw), 'utf-8');
     this.invalidateCache();
   }

src/infra/task/summarize.ts

@@ -53,7 +53,8 @@ export class TaskSummarizer {
     taskName: string,
     options: SummarizeOptions,
   ): Promise<string> {
-    const useLLM = options.useLLM ?? true;
+    const globalConfig = loadGlobalConfig();
+    const useLLM = options.useLLM ?? (globalConfig.branchNameStrategy === 'ai');
     log.info('Summarizing task name', { taskName, useLLM });
 
     if (!useLLM) {
@@ -61,8 +62,6 @@ export class TaskSummarizer {
       log.info('Task name romanized', { original: taskName, slug });
       return slug || 'task';
     }
-
-    const globalConfig = loadGlobalConfig();
     const providerType = (globalConfig.provider as ProviderType) ?? 'claude';
     const model = options.model ?? globalConfig.model;
 

src/infra/task/types.ts

@@ -63,7 +63,7 @@ export interface SummarizeOptions {
   cwd: string;
   /** Model to use (optional, defaults to config or haiku) */
   model?: string;
-  /** Use LLM for summarization (default: true). If false, uses romanization. */
+  /** Use LLM for summarization. Defaults to config.branchNameStrategy === 'ai'. If false, uses romanization. */
   useLLM?: boolean;
 }