rafcode 3.0.0 → 3.8.0

package/tests/unit/dependency-integration.test.ts

@@ -1,3 +1,4 @@
+import { jest } from '@jest/globals';
 import * as fs from 'node:fs';
 import * as path from 'node:path';
 import * as os from 'node:os';
@@ -13,7 +14,25 @@ import {
   type DerivedTask,
 } from '../../src/core/state-derivation.js';
 import { getOutcomeFilePath, extractTaskNameFromPlanFile } from '../../src/utils/paths.js';
-import { getExecutionPrompt } from '../../src/prompts/execution.js';
+
+const suiteHomeDir = fs.mkdtempSync(path.join(os.tmpdir(), 'raf-dependency-home-'));
+let mockHomeDir = suiteHomeDir;
+
+jest.unstable_mockModule('node:os', () => ({
+  homedir: () => mockHomeDir,
+  tmpdir: () => os.tmpdir(),
+}));
+
+jest.unstable_mockModule('../../src/utils/config.js', () => ({
+  getCommitFormat: () => '{prefix}[{projectName}:{taskId}] {description}',
+  getCommitPrefix: () => 'RAF',
+  renderCommitMessage: (template: string, variables: Record<string, string>) =>
+    template.replace(/\{(\w+)\}/g, (match, key: string) => variables[key] ?? match),
+  resetConfigCache: jest.fn(),
+}));
+
+const { getExecutionPrompt } = await import('../../src/prompts/execution.js');
+const { resetConfigCache } = await import('../../src/utils/config.js');
 
 /**
  * Integration tests for the complete dependency blocking flow.
@@ -25,6 +44,8 @@ describe('Dependency Integration Flow', () => {
   let projectPath: string;
 
   beforeEach(() => {
+    fs.rmSync(path.join(mockHomeDir, '.raf'), { recursive: true, force: true });
+    resetConfigCache();
     tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'raf-dep-integration-'));
     projectPath = path.join(tempDir, '1-test-project');
     fs.mkdirSync(projectPath, { recursive: true });
@@ -34,9 +55,14 @@ describe('Dependency Integration Flow', () => {
   });
 
   afterEach(() => {
+    resetConfigCache();
     fs.rmSync(tempDir, { recursive: true, force: true });
   });
 
+  afterAll(() => {
+    fs.rmSync(suiteHomeDir, { recursive: true, force: true });
+  });
+
   describe('Complete dependency chain simulation', () => {
     /**
      * Simulates a real project with the following dependency structure:

package/tests/unit/do-model-display.test.ts

@@ -0,0 +1,35 @@
+import { formatResolvedTaskModel } from '../../src/commands/do.js';
+import type { ModelEntry } from '../../src/types/config.js';
+
+describe('do model display', () => {
+  it('should include effort and fast metadata in verbose model logs', () => {
+    const entry: ModelEntry = {
+      model: 'sonnet',
+      provider: 'claude',
+      reasoningEffort: 'low',
+      fast: true,
+    };
+
+    expect(formatResolvedTaskModel(entry)).toBe('claude-sonnet-4-5-20250929, low, fast');
+  });
+
+  it('should omit effort when unavailable', () => {
+    const entry: ModelEntry = {
+      model: 'sonnet',
+      provider: 'claude',
+    };
+
+    expect(formatResolvedTaskModel(entry)).toBe('claude-sonnet-4-5-20250929');
+  });
+
+  it('should omit fast when false', () => {
+    const entry: ModelEntry = {
+      model: 'sonnet',
+      provider: 'claude',
+      reasoningEffort: 'low',
+      fast: false,
+    };
+
+    expect(formatResolvedTaskModel(entry)).toBe('claude-sonnet-4-5-20250929, low');
+  });
+});

package/tests/unit/execution-prompt.test.ts

@@ -1,6 +1,34 @@
-import { getExecutionPrompt, ExecutionPromptParams, summarizeOutcome } from '../../src/prompts/execution.js';
+import { jest } from '@jest/globals';
+import * as fs from 'node:fs';
+import * as os from 'node:os';
+import * as path from 'node:path';
+import type { ExecutionPromptParams } from '../../src/prompts/execution.js';
+
+const suiteHomeDir = fs.mkdtempSync(path.join(os.tmpdir(), 'raf-execution-prompt-home-'));
+let mockHomeDir = suiteHomeDir;
+
+jest.unstable_mockModule('node:os', () => ({
+  homedir: () => mockHomeDir,
+  tmpdir: () => os.tmpdir(),
+}));
+
+const { getExecutionPrompt, summarizeOutcome } = await import('../../src/prompts/execution.js');
+const { resetConfigCache } = await import('../../src/utils/config.js');
 
 describe('Execution Prompt', () => {
+  beforeEach(() => {
+    fs.rmSync(path.join(mockHomeDir, '.raf'), { recursive: true, force: true });
+    resetConfigCache();
+  });
+
+  afterEach(() => {
+    resetConfigCache();
+  });
+
+  afterAll(() => {
+    fs.rmSync(suiteHomeDir, { recursive: true, force: true });
+  });
+
   const baseParams: ExecutionPromptParams = {
     projectPath: '/Users/test/RAF/42-task-naming-improvements',
     planPath: '/Users/test/RAF/42-task-naming-improvements/plans/1-enhance-identifier-resolution.md',
@@ -16,7 +44,7 @@ describe('Execution Prompt', () => {
   describe('Commit Message Format', () => {
     it('should include RAF commit schema format with description placeholder in prompt', () => {
       const prompt = getExecutionPrompt(baseParams);
-      expect(prompt).toContain('RAF[42:1] <description>');
+      expect(prompt).toContain('RAF[task-naming-improvements:1] <description>');
     });
 
     it('should instruct to write meaningful description', () => {
@@ -28,25 +56,25 @@ describe('Execution Prompt', () => {
     it('should use plain numeric task IDs', () => {
       const params = { ...baseParams, taskNumber: 1 };
       const prompt = getExecutionPrompt(params);
-      expect(prompt).toContain('RAF[42:1]');
+      expect(prompt).toContain('RAF[task-naming-improvements:1]');
     });
 
     it('should use numeric task IDs for double digits', () => {
       const params = { ...baseParams, taskNumber: 12 };
       const prompt = getExecutionPrompt(params);
-      expect(prompt).toContain('RAF[42:12]');
+      expect(prompt).toContain('RAF[task-naming-improvements:12]');
     });
 
     it('should use numeric task IDs for large numbers', () => {
       const params = { ...baseParams, taskNumber: 123 };
       const prompt = getExecutionPrompt(params);
-      expect(prompt).toContain('RAF[42:123]');
+      expect(prompt).toContain('RAF[task-naming-improvements:123]');
     });
 
-    it('should include project number from numeric prefix', () => {
-      const params = { ...baseParams, projectNumber: '7' };
+    it('should use project name from folder in commit format', () => {
+      const params = { ...baseParams, projectPath: '/Users/test/RAF/7-auth-fix', projectNumber: '7' };
       const prompt = getExecutionPrompt(params);
-      expect(prompt).toContain('RAF[7:1]');
+      expect(prompt).toContain('RAF[auth-fix:1]');
     });
 
     it('should not include commit instructions when autoCommit is false', () => {
@@ -67,7 +95,9 @@ describe('Execution Prompt', () => {
       const params = { ...baseParams, autoCommit: false };
       const prompt = getExecutionPrompt(params);
       expect(prompt).not.toContain('On Failure');
-      expect(prompt).not.toContain('do NOT commit');
+      // The general outcome section still mentions not committing on failure,
+      // but the git-specific instructions block should be absent
+      expect(prompt).not.toContain('Git Instructions');
     });
 
     it('should instruct not to add Co-Authored-By or other trailers', () => {
@@ -96,7 +126,7 @@ describe('Execution Prompt', () => {
         outcomeFilePath: '/Users/test/RAF/42-task-naming-improvements/outcomes/6-update-execution-prompt.md',
       };
       const prompt = getExecutionPrompt(params);
-      expect(prompt).toContain('RAF[42:6] <description>');
+      expect(prompt).toContain('RAF[task-naming-improvements:6] <description>');
     });
 
     it('should generate correct commit message format for first task', () => {
@@ -112,7 +142,7 @@ describe('Execution Prompt', () => {
         outcomeFilePath: '/Users/test/RAF/1-fix-bug/outcomes/1-identify-issue.md',
       };
       const prompt = getExecutionPrompt(params);
-      expect(prompt).toContain('RAF[1:1] <description>');
+      expect(prompt).toContain('RAF[fix-bug:1] <description>');
     });
 
     it('should generate correct commit message format for numeric project', () => {
@@ -128,7 +158,7 @@ describe('Execution Prompt', () => {
         outcomeFilePath: '/Users/test/RAF/7-feature-branch/outcomes/2-implement-feature.md',
       };
       const prompt = getExecutionPrompt(params);
-      expect(prompt).toContain('RAF[7:2] <description>');
+      expect(prompt).toContain('RAF[feature-branch:2] <description>');
     });
   });
 
@@ -163,7 +193,7 @@ describe('Execution Prompt', () => {
 
     it('should include instructions for writing outcome file', () => {
       const prompt = getExecutionPrompt(baseParams);
-      expect(prompt).toContain('You MUST write an outcome file');
+      expect(prompt).toContain('Outcome file path');
       expect(prompt).toContain('summary of what was done');
     });
 
@@ -176,25 +206,25 @@ describe('Execution Prompt', () => {
 
     it('should distinguish between code tasks and documentation tasks', () => {
       const prompt = getExecutionPrompt(baseParams);
-      expect(prompt).toContain('For code tasks');
-      expect(prompt).toContain('For documentation/report tasks');
+      expect(prompt).toContain('documentation/report tasks');
     });
 
     it('should instruct that marker is last line in outcome file', () => {
       const prompt = getExecutionPrompt(baseParams);
-      expect(prompt).toContain('completion marker MUST be the LAST line in the outcome file');
+      expect(prompt).toContain('completion marker as the LAST line');
     });
   });
 
   describe('Commit Workflow Rules', () => {
-    it('should include rule to commit code and outcome together on success', () => {
+    it('should include commit instructions when autoCommit is true', () => {
       const prompt = getExecutionPrompt(baseParams);
-      expect(prompt).toContain('On SUCCESS: Commit code changes AND outcome file together');
+      expect(prompt).toContain('Git Instructions');
+      expect(prompt).toContain('git add');
     });
 
     it('should include rule not to commit on failure', () => {
       const prompt = getExecutionPrompt(baseParams);
-      expect(prompt).toContain('On FAILURE: Do NOT commit');
+      expect(prompt).toContain('do NOT commit');
     });
 
     it('should specify that changes are preserved for debugging on failure', () => {

package/tests/unit/name-generator.test.ts

@@ -1,5 +1,11 @@
 import { jest } from '@jest/globals';
 import { EventEmitter } from 'node:events';
+import type { HarnessName } from '../../src/types/config.js';
+
+let currentNameGenerationModel = {
+  model: 'sonnet',
+  harness: 'claude' as HarnessName,
+};
 
 // Helper to create a mock spawn that returns a fake ChildProcess
 function createMockSpawn(stdoutData: string | null, exitCode: number = 0) {
@@ -28,6 +34,11 @@ jest.unstable_mockModule('node:child_process', () => ({
   execSync: jest.fn(), // keep available for transitive imports
 }));
 
+jest.unstable_mockModule('../../src/utils/config.js', () => ({
+  getModel: jest.fn(() => currentNameGenerationModel),
+  isValidModelName: jest.fn(() => true),
+}));
+
 // Import after mocking
 const { generateProjectName, generateProjectNames, sanitizeGeneratedName } =
   await import('../../src/utils/name-generator.js');
@@ -35,10 +46,14 @@ const { generateProjectName, generateProjectNames, sanitizeGeneratedName } =
 describe('Name Generator', () => {
   beforeEach(() => {
     jest.clearAllMocks();
+    currentNameGenerationModel = {
+      model: 'sonnet',
+      harness: 'claude',
+    };
   });
 
   describe('generateProjectName', () => {
-    it('should return sanitized name from Claude response', async () => {
+    it('should return sanitized name from CLI response', async () => {
       mockSpawn.mockReturnValue(createMockSpawn('user-auth-system\n'));
 
       const result = await generateProjectName('Build a user authentication system');
@@ -46,8 +61,8 @@ describe('Name Generator', () => {
       expect(result).toBe('user-auth-system');
       expect(mockSpawn).toHaveBeenCalledTimes(1);
       expect(mockSpawn).toHaveBeenCalledWith(
-        'claude',
-        expect.arrayContaining(['--model', 'haiku', '--no-session-persistence', '-p']),
+        currentNameGenerationModel.harness,
+        expect.arrayContaining(['--model', currentNameGenerationModel.model, '--no-session-persistence', '-p']),
         expect.any(Object)
       );
     });
@@ -77,7 +92,7 @@ describe('Name Generator', () => {
       expect(result).toBe('some-project-name');
     });
 
-    it('should fall back to word extraction when Claude fails', async () => {
+    it('should fall back to word extraction when CLI fails', async () => {
       mockSpawn.mockReturnValue(createMockSpawn(null, 1));
 
       const result = await generateProjectName('Build a user authentication system with OAuth');
@@ -85,7 +100,7 @@ describe('Name Generator', () => {
       expect(result).toBe('build-user-authentication');
     });
 
-    it('should fall back to word extraction when Claude returns empty', async () => {
+    it('should fall back to word extraction when CLI returns empty', async () => {
       mockSpawn.mockReturnValue(createMockSpawn(''));
 
       const result = await generateProjectName('Implement caching layer for database');
@@ -93,7 +108,7 @@ describe('Name Generator', () => {
       expect(result).toBe('implement-caching-layer');
     });
 
-    it('should fall back to word extraction when Claude returns single char', async () => {
+    it('should fall back to word extraction when CLI returns single char', async () => {
       mockSpawn.mockReturnValue(createMockSpawn('a'));
 
       const result = await generateProjectName('Add new logging functionality');
@@ -109,7 +124,7 @@ describe('Name Generator', () => {
       expect(result).toBe('project');
     });
 
-    it('should truncate long names from Claude', async () => {
+    it('should truncate long names from CLI', async () => {
       const longName =
         'this-is-a-very-long-project-name-that-exceeds-the-maximum-allowed-length-for-folder-names';
       mockSpawn.mockReturnValue(createMockSpawn(longName));
@@ -153,7 +168,7 @@ describe('Name Generator', () => {
   });
 
   describe('generateProjectNames', () => {
-    it('should return multiple sanitized names from Claude response', async () => {
+    it('should return multiple sanitized names from CLI response', async () => {
       mockSpawn.mockReturnValue(
         createMockSpawn('phoenix-rise\nturbo-boost\nbug-squasher\ncatalyst\nmerlin\n')
       );
@@ -227,7 +242,7 @@ describe('Name Generator', () => {
       expect(result.length).toBe(3);
     });
 
-    it('should fall back to single name when Claude fails', async () => {
+    it('should fall back to single name when CLI fails', async () => {
       mockSpawn.mockReturnValue(createMockSpawn(null, 1));
 
       const result = await generateProjectNames('Build a user authentication system with OAuth');
@@ -235,13 +250,12 @@ describe('Name Generator', () => {
       expect(result).toEqual(['build-user-authentication']);
     });
 
-    it('should fall back to single name when too few names returned', async () => {
+    it('should keep 1-2 valid names instead of falling back', async () => {
       mockSpawn.mockReturnValue(createMockSpawn('phoenix\nturbo\n'));
 
       const result = await generateProjectNames('Build something awesome');
 
-      // Only 2 names returned, so fallback kicks in
-      expect(result).toEqual(['build-something-awesome']);
+      expect(result).toEqual(['phoenix', 'turbo']);
     });
 
     it('should filter out invalid/short names', async () => {
@@ -271,6 +285,62 @@ describe('Name Generator', () => {
 
       expect(result).toEqual(['some-project']);
     });
+
+    it('should fall back when codex returns no usable suggestions after sanitization', async () => {
+      currentNameGenerationModel = {
+        model: 'gpt-5.4',
+        harness: 'codex',
+      };
+      mockSpawn.mockReturnValue(createMockSpawn('a\n!\n'));
+
+      const result = await generateProjectNames('Build something awesome');
+
+      expect(result).toEqual(['build-something-awesome']);
+    });
+
+    it('should no longer accept a harness argument (config-driven)', async () => {
+      // The harness is embedded in the config, not passed as a parameter
+      // This test verifies that generateProjectNames has a single-parameter signature
+      mockSpawn.mockReturnValue(
+        createMockSpawn('phoenix\nturbo-boost\ncatalyst\n')
+      );
+
+      const result = await generateProjectNames('Build something');
+
+      expect(result).toEqual(['phoenix', 'turbo-boost', 'catalyst']);
+      expect(mockSpawn).toHaveBeenCalledWith(
+        currentNameGenerationModel.harness,
+        expect.arrayContaining(['--model', currentNameGenerationModel.model]),
+        expect.any(Object)
+      );
+    });
+
+    it('should use codex exec invocation for name generation when configured', async () => {
+      currentNameGenerationModel = {
+        model: 'gpt-5.4',
+        harness: 'codex',
+      };
+      mockSpawn.mockReturnValue(createMockSpawn('phoenix-rise\nturbo-boost\nbug-squasher\n'));
+
+      const result = await generateProjectNames('Build something');
+
+      expect(result).toEqual(['phoenix-rise', 'turbo-boost', 'bug-squasher']);
+      expect(mockSpawn).toHaveBeenCalledWith(
+        'codex',
+        expect.arrayContaining([
+          'exec',
+          '--skip-git-repo-check',
+          '--ephemeral',
+          '--color',
+          'never',
+          '-m',
+          'gpt-5.4',
+        ]),
+        expect.any(Object)
+      );
+      const promptArg = (mockSpawn.mock.calls[0][1] as string[]).at(-1);
+      expect(promptArg).toContain('Output EXACTLY 5 project names');
+    });
   });
 
   describe('sanitizeGeneratedName', () => {

package/tests/unit/plan-command-auto-flag.test.ts

@@ -52,26 +52,23 @@ describe('Plan Command - Auto Flag', () => {
       expect(amendOption).toBeDefined();
     });
 
-    it('should allow --auto with --model', () => {
+    it('should not have --provider flag (removed)', () => {
       const command = createPlanCommand();
       const options = command.options;
 
-      const autoOption = options.find((opt) => opt.long === '--auto');
-      const modelOption = options.find((opt) => opt.long === '--model');
-
-      expect(autoOption).toBeDefined();
-      expect(modelOption).toBeDefined();
+      const providerOption = options.find((opt) => opt.long === '--provider');
+      expect(providerOption).toBeUndefined();
     });
 
-    it('should allow --auto with --sonnet', () => {
+    it('should not have removed --model or --sonnet flags', () => {
       const command = createPlanCommand();
       const options = command.options;
 
-      const autoOption = options.find((opt) => opt.long === '--auto');
+      const modelOption = options.find((opt) => opt.long === '--model');
       const sonnetOption = options.find((opt) => opt.long === '--sonnet');
 
-      expect(autoOption).toBeDefined();
-      expect(sonnetOption).toBeDefined();
+      expect(modelOption).toBeUndefined();
+      expect(sonnetOption).toBeUndefined();
     });
   });
 

package/tests/unit/plan-command.test.ts

@@ -344,10 +344,10 @@ describe('Plan Command - Amend Functionality', () => {
 
       const { systemPrompt } = getAmendPrompt(params);
 
-      expect(systemPrompt).toContain('AMENDMENT MODE');
+      expect(systemPrompt).toContain('Amendment Mode');
       expect(systemPrompt).toContain('Task 1: first [COMPLETED] [PROTECTED]');
       expect(systemPrompt).toContain('Task 2: second [PENDING] [MODIFIABLE]');
-      expect(systemPrompt).toContain('starting from number 3');
+      expect(systemPrompt).toContain('start from number 3');
       expect(systemPrompt).toContain('/test/project');
     });
 
@@ -377,11 +377,9 @@ describe('Plan Command - Amend Functionality', () => {
 
       const { systemPrompt } = getAmendPrompt(params);
 
-      expect(systemPrompt).toContain('NEVER modify [PROTECTED] tasks (completed)');
-      expect(systemPrompt).toContain('DO NOT renumber existing tasks');
-      expect(systemPrompt).toContain('You MAY modify [MODIFIABLE] tasks (pending/failed)');
-      expect(systemPrompt).toContain('NEVER modify COMPLETED task plans - they are [PROTECTED]');
-      expect(systemPrompt).toContain('You MAY modify non-completed task plans (pending/failed)');
+      expect(systemPrompt).toContain('[PROTECTED] tasks (completed): NEVER modify');
+      expect(systemPrompt).toContain('Do NOT renumber existing tasks');
+      expect(systemPrompt).toContain('[MODIFIABLE] tasks (pending/failed): MAY modify');
     });
 
     it('should include failed task status with MODIFIABLE indicator in system prompt', () => {
@@ -413,7 +411,7 @@ describe('Plan Command - Amend Functionality', () => {
 
       const { systemPrompt } = getAmendPrompt(params);
 
-      expect(systemPrompt).toContain('### Protected Tasks (COMPLETED - cannot be modified)');
+      expect(systemPrompt).toContain('### Protected (COMPLETED)');
       expect(systemPrompt).toContain('- Task 1: first');
       expect(systemPrompt).toContain('- Task 2: second');
     });
@@ -432,7 +430,7 @@ describe('Plan Command - Amend Functionality', () => {
 
       const { systemPrompt } = getAmendPrompt(params);
 
-      expect(systemPrompt).toContain('### Modifiable Tasks (PENDING/FAILED - can be modified if requested)');
+      expect(systemPrompt).toContain('### Modifiable (PENDING/FAILED)');
       expect(systemPrompt).toContain('- Task 2: second');
       expect(systemPrompt).toContain('- Task 3: third');
     });
@@ -449,7 +447,7 @@ describe('Plan Command - Amend Functionality', () => {
 
       const { systemPrompt } = getAmendPrompt(params);
 
-      expect(systemPrompt).toContain('### Protected Tasks (COMPLETED - cannot be modified)\n(none)');
+      expect(systemPrompt).toContain('### Protected (COMPLETED)\n(none)');
     });
 
     it('should show (none) when there are no modifiable tasks', () => {
@@ -464,7 +462,7 @@ describe('Plan Command - Amend Functionality', () => {
 
       const { systemPrompt } = getAmendPrompt(params);
 
-      expect(systemPrompt).toContain('### Modifiable Tasks (PENDING/FAILED - can be modified if requested)\n(none)');
+      expect(systemPrompt).toContain('### Modifiable (PENDING/FAILED)\n(none)');
     });
 
     it('should handle mixed task statuses correctly', () => {
@@ -489,8 +487,8 @@ describe('Plan Command - Amend Functionality', () => {
       expect(systemPrompt).toContain('Task 4: tests [PENDING] [MODIFIABLE]');
 
       // Check separate lists
-      expect(systemPrompt).toContain('### Protected Tasks (COMPLETED - cannot be modified)');
-      expect(systemPrompt).toContain('### Modifiable Tasks (PENDING/FAILED - can be modified if requested)');
+      expect(systemPrompt).toContain('### Protected (COMPLETED)');
+      expect(systemPrompt).toContain('### Modifiable (PENDING/FAILED)');
     });
 
     it('should include correct plans directory path', () => {
@@ -506,7 +504,6 @@ describe('Plan Command - Amend Functionality', () => {
       const { systemPrompt } = getAmendPrompt(params);
 
       expect(systemPrompt).toContain('/my/project/path/plans/2-task-name.md');
-      expect(systemPrompt).toContain('/my/project/path/plans/3-task-name.md');
     });
 
     it('should include new task description in user message', () => {
@@ -532,7 +529,7 @@ describe('Plan Command - Amend Functionality', () => {
 
       const { systemPrompt } = getAmendPrompt(params);
 
-      expect(systemPrompt).toContain('AMENDMENT MODE');
+      expect(systemPrompt).toContain('Amendment Mode');
       expect(systemPrompt).toContain('AskUserQuestion');
       expect(systemPrompt).toContain('Interview the User');
     });
@@ -597,9 +594,9 @@ describe('Plan Command - Amend Functionality', () => {
 
       const { systemPrompt } = getAmendPrompt(params);
 
-      expect(systemPrompt).toContain('Identifying Follow-up Tasks');
       expect(systemPrompt).toContain('follow-up to task NN');
-      expect(systemPrompt).toContain('outcome file paths for completed tasks are listed above');
+      expect(systemPrompt).toContain('follow-up');
+      expect(systemPrompt).toContain('outcome');
     });
   });
 

package/tests/unit/planning-prompt.test.ts

@@ -76,8 +76,8 @@ describe('Planning Prompt', () => {
 
       const { systemPrompt } = getPlanningPrompt(params);
 
-      expect(systemPrompt).toContain('/test/project/plans/1-task-name.md');
-      expect(systemPrompt).toContain('/test/project/plans/2-task-name.md');
+      expect(systemPrompt).toContain('/test/project/plans/');
+      expect(systemPrompt).toContain('1-task-name.md');
     });
 
     it('should include task guidelines in system prompt', () => {
@@ -88,7 +88,7 @@ describe('Planning Prompt', () => {
 
       const { systemPrompt } = getPlanningPrompt(params);
 
-      expect(systemPrompt).toContain('identify distinct');
+      expect(systemPrompt).toContain('Identify distinct tasks');
       expect(systemPrompt).toContain('independently completable');
       expect(systemPrompt).toContain('10-30 minutes');
     });
@@ -116,9 +116,9 @@ describe('Planning Prompt', () => {
 
       const { systemPrompt } = getPlanningPrompt(params);
 
-      expect(systemPrompt).toContain('ALWAYS interview the user before creating plans');
-      expect(systemPrompt).toContain('numbered order (1, 2, 3');
-      expect(systemPrompt).toContain('kebab-case names');
+      expect(systemPrompt).toContain('Do not skip this step');
+      expect(systemPrompt).toContain('execution order');
+      expect(systemPrompt).toContain('kebab-case');
     });
 
     it('should show raf do without --worktree when worktreeMode is false', () => {
@@ -146,7 +146,7 @@ describe('Planning Prompt', () => {
       expect(systemPrompt).not.toContain('--worktree');
     });
 
-    it('should show raf do with --worktree when worktreeMode is true', () => {
+    it('should show raf do without --worktree when worktreeMode is true', () => {
       const params: PlanningPromptParams = {
         projectPath: '/test/project',
         inputContent: 'Some project',
@@ -155,7 +155,8 @@ describe('Planning Prompt', () => {
 
       const { systemPrompt } = getPlanningPrompt(params);
 
-      expect(systemPrompt).toContain('raf do <project> --worktree');
+      expect(systemPrompt).toContain('raf do <project>');
+      expect(systemPrompt).not.toContain('--worktree');
     });
 
     it('should include project description in user message', () => {