Files
Claude-Code-Workflow/.claude/skills/skill-generator/phases/03-phase-generation.md
catlog22 be498acf59 feat: Add code analysis and LLM action templates with detailed configurations and examples
- Introduced a comprehensive code analysis action template for integrating code exploration and analysis capabilities.
- Added LLM action template for seamless integration of LLM calls with customizable prompts and tools.
- Implemented a benchmark search script to compare multiple search methods across various dimensions including speed, result quality, ranking stability, and coverage.
- Provided preset configurations for common analysis tasks and LLM actions, enhancing usability and flexibility.
2026-01-03 17:37:49 +08:00

22 KiB
Raw Blame History

Phase 3: Phase Generation

根据执行模式生成 Phase 文件,包含声明式工作流编排和上下文策略支持。

Objective

  • Sequential 模式:生成顺序 Phase 文件 + 声明式编排器
  • Autonomous 模式:生成编排器和动作文件
  • 支持 文件上下文内存上下文 两种策略

Input

  • 依赖: skill-config.json, SKILL.md (Phase 1-2 产出)
  • 模板: templates/sequential-phase.md, templates/autonomous-*.md

上下文策略 (P0 增强)

根据 config.context_strategy 生成不同的上下文管理代码:

策略 适用场景 优点 缺点
file 复杂多阶段任务 持久化、可调试、可恢复 IO 开销
memory 简单线性任务 速度快 无法恢复、调试困难
const CONTEXT_STRATEGIES = {
  file: {
    read: (key) => `JSON.parse(Read(\`\${workDir}/context/${key}.json\`))`,
    write: (key, data) => `Write(\`\${workDir}/context/${key}.json\`, JSON.stringify(${data}, null, 2))`,
    init: `Bash(\`mkdir -p "\${workDir}/context"\`)`
  },
  memory: {
    read: (key) => `state.context.${key}`,
    write: (key, data) => `state.context.${key} = ${data}`,
    init: `state.context = {}`
  }
};

Execution Steps

Step 1: 读取配置和模板

const config = JSON.parse(Read(`${workDir}/skill-config.json`));
const skillDir = `.claude/skills/${config.skill_name}`;
const contextStrategy = config.context_strategy || 'file'; // 默认文件策略

// 读取模板
const skillRoot = '.claude/skills/skill-generator';

Step 2: Sequential 模式 - 生成阶段文件 + 声明式编排器

if (config.execution_mode === 'sequential') {
  const phases = config.sequential_config.phases;
  
  // ========== P0 增强: 生成声明式编排器 ==========
  const workflowOrchestrator = generateSequentialOrchestrator(config, phases);
  Write(`${skillDir}/phases/_orchestrator.md`, workflowOrchestrator);
  
  // ========== P0 增强: 生成工作流定义 ==========
  const workflowDef = generateWorkflowDefinition(config, phases);
  Write(`${skillDir}/workflow.json`, JSON.stringify(workflowDef, null, 2));
  
  // 生成各阶段文件
  for (let i = 0; i < phases.length; i++) {
    const phase = phases[i];
    const prevPhase = i > 0 ? phases[i-1] : null;
    const nextPhase = i < phases.length - 1 ? phases[i+1] : null;
    
    const content = generateSequentialPhase({
      phaseNumber: i + 1,
      phaseId: phase.id,
      phaseName: phase.name,
      phaseDescription: phase.description || `Execute ${phase.name}`,
      input: prevPhase ? prevPhase.output : "user input",
      output: phase.output,
      nextPhase: nextPhase ? nextPhase.id : null,
      config: config,
      contextStrategy: contextStrategy
    });
    
    Write(`${skillDir}/phases/${phase.id}.md`, content);
  }
}

// ========== P0 增强: 声明式工作流定义 ==========
function generateWorkflowDefinition(config, phases) {
  return {
    skill_name: config.skill_name,
    version: "1.0.0",
    execution_mode: "sequential",
    context_strategy: config.context_strategy || "file",
    
    // 声明式阶段列表 (类似 software-manual 的 agents_to_run)
    phases_to_run: phases.map(p => p.id),
    
    // 阶段配置
    phases: phases.map((p, i) => ({
      id: p.id,
      name: p.name,
      order: i + 1,
      input: i > 0 ? phases[i-1].output : null,
      output: p.output,
      // 可选的并行配置
      parallel: p.parallel || false,
      // 可选的条件执行
      condition: p.condition || null,
      // Agent 配置
      agent: p.agent || {
        type: "universal-executor",
        run_in_background: false
      }
    })),
    
    // 终止条件
    termination: {
      on_success: "all_phases_completed",
      on_error: "stop_and_report",
      max_retries: 3
    }
  };
}

// ========== P0 增强: 声明式编排器 ==========
function generateSequentialOrchestrator(config, phases) {
  return `# Sequential Orchestrator

声明式工作流编排器,按 \`workflow.json\` 定义顺序执行阶段。

## 工作流定义

\`\`\`javascript
const workflow = JSON.parse(Read(\`\${skillDir}/workflow.json\`));
\`\`\`

## 编排逻辑

\`\`\`javascript
async function runSequentialWorkflow(workDir) {
  const workflow = JSON.parse(Read(\`\${skillDir}/workflow.json\`));
  const contextStrategy = workflow.context_strategy;
  
  // 初始化上下文
  ${config.context_strategy === 'file' ? 
    `Bash(\`mkdir -p "\${workDir}/context"\`);` :
    `const state = { context: {} };`}
  
  // 执行状态追踪
  const execution = {
    started_at: new Date().toISOString(),
    phases_completed: [],
    current_phase: null,
    errors: []
  };
  
  Write(\`\${workDir}/execution-state.json\`, JSON.stringify(execution, null, 2));
  
  // 按声明顺序执行阶段
  for (const phaseId of workflow.phases_to_run) {
    const phaseConfig = workflow.phases.find(p => p.id === phaseId);
    
    // 更新执行状态
    execution.current_phase = phaseId;
    Write(\`\${workDir}/execution-state.json\`, JSON.stringify(execution, null, 2));
    
    console.log(\`[Orchestrator] Executing: \${phaseId}\`);
    
    try {
      // 检查条件执行
      if (phaseConfig.condition) {
        const shouldRun = evaluateCondition(phaseConfig.condition, execution);
        if (!shouldRun) {
          console.log(\`[Orchestrator] Skipping \${phaseId} (condition not met)\`);
          continue;
        }
      }
      
      // 执行阶段
      const result = await executePhase(phaseId, phaseConfig, workDir);
      
      // 记录完成
      execution.phases_completed.push({
        id: phaseId,
        completed_at: new Date().toISOString(),
        output: phaseConfig.output
      });
      
    } catch (error) {
      execution.errors.push({
        phase: phaseId,
        message: error.message,
        timestamp: new Date().toISOString()
      });
      
      // 错误处理策略
      if (workflow.termination.on_error === 'stop_and_report') {
        console.error(\`[Orchestrator] Failed at \${phaseId}: \${error.message}\`);
        break;
      }
    }
    
    Write(\`\${workDir}/execution-state.json\`, JSON.stringify(execution, null, 2));
  }
  
  // 完成
  execution.current_phase = null;
  execution.completed_at = new Date().toISOString();
  Write(\`\${workDir}/execution-state.json\`, JSON.stringify(execution, null, 2));
  
  return execution;
}

async function executePhase(phaseId, phaseConfig, workDir) {
  const phasePrompt = Read(\`\${skillDir}/phases/\${phaseId}.md\`);
  
  // 使用 Task 调用 Agent
  const result = await Task({
    subagent_type: phaseConfig.agent?.type || 'universal-executor',
    run_in_background: phaseConfig.agent?.run_in_background || false,
    prompt: \`
[PHASE] \${phaseId}
[WORK_DIR] \${workDir}
[INPUT] \${phaseConfig.input ? \`\${workDir}/\${phaseConfig.input}\` : 'None'}
[OUTPUT] \${workDir}/\${phaseConfig.output}

\${phasePrompt}
\`
  });
  
  return JSON.parse(result);
}
\`\`\`

## 阶段执行计划

| Order | Phase | Input | Output | Agent |
|-------|-------|-------|--------|-------|
${phases.map((p, i) => 
  `| ${i+1} | ${p.id} | ${i > 0 ? phases[i-1].output : '-'} | ${p.output} | ${p.agent?.type || 'universal-executor'} |`
).join('\n')}

## 错误恢复

\`\`\`javascript
// 从指定阶段恢复执行
async function resumeFromPhase(phaseId, workDir) {
  const workflow = JSON.parse(Read(\`\${skillDir}/workflow.json\`));
  const startIndex = workflow.phases_to_run.indexOf(phaseId);
  
  if (startIndex === -1) {
    throw new Error(\`Phase not found: \${phaseId}\`);
  }
  
  // 从指定阶段开始执行
  const remainingPhases = workflow.phases_to_run.slice(startIndex);
  // ...继续执行
}
\`\`\`
`;
}

// 生成阶段文件(增强上下文策略支持)
function generateSequentialPhase(params) {
  const contextCode = params.contextStrategy === 'file' ? {
    readPrev: `const prevOutput = JSON.parse(Read(\`\${workDir}/${params.input}\`));`,
    writeResult: `Write(\`\${workDir}/${params.output}\`, JSON.stringify(result, null, 2));`,
    readContext: (key) => `JSON.parse(Read(\`\${workDir}/context/${key}.json\`))`,
    writeContext: (key) => `Write(\`\${workDir}/context/${key}.json\`, JSON.stringify(data, null, 2))`
  } : {
    readPrev: `const prevOutput = state.context.prevPhaseOutput;`,
    writeResult: `state.context.${params.phaseId.replace(/-/g, '_')}_output = result;`,
    readContext: (key) => `state.context.${key}`,
    writeContext: (key) => `state.context.${key} = data`
  };

  return `# Phase ${params.phaseNumber}: ${params.phaseName}

${params.phaseDescription}

## Objective

- 主要目标描述
- 具体任务列表

## Input

- 依赖: \`${params.input}\`
- 配置: \`{workDir}/skill-config.json\`
- 上下文策略: \`${params.contextStrategy}\`

## Execution Steps

### Step 1: 读取输入

\`\`\`javascript
// 上下文策略: ${params.contextStrategy}
${params.phaseNumber > 1 ? contextCode.readPrev : '// 首阶段,直接从配置开始'}
\`\`\`

### Step 2: 核心处理

\`\`\`javascript
// TODO: 实现核心逻辑
const result = {
  status: 'completed',
  data: {
    // 处理结果
  },
  metadata: {
    phase: '${params.phaseId}',
    timestamp: new Date().toISOString()
  }
};
\`\`\`

### Step 3: 输出结果

\`\`\`javascript
// 写入阶段产出 (上下文策略: ${params.contextStrategy})
${contextCode.writeResult}

// 返回简要信息给编排器
return {
  status: 'completed',
  output_file: '${params.output}',
  summary: '阶段 ${params.phaseNumber} 完成'
};
\`\`\`

## Output

- **File**: \`${params.output}\`
- **Format**: ${params.output.endsWith('.json') ? 'JSON' : 'Markdown'}
- **Context Strategy**: ${params.contextStrategy}

## Quality Checklist

- [ ] 输入数据验证通过
- [ ] 核心逻辑执行成功
- [ ] 输出格式正确
- [ ] 上下文正确保存

${params.nextPhase ? 
  `## Next Phase\n\n→ [Phase ${params.phaseNumber + 1}: ${params.nextPhase}](${params.nextPhase}.md)` : 
  `## Completion\n\n此为最后阶段输出最终产物。`}
`;
}

Step 3: Autonomous 模式 - 生成编排器 (增强版)

if (config.execution_mode === 'autonomous' || config.execution_mode === 'hybrid') {
  const contextStrategy = config.context_strategy || 'file';
  
  // 生成状态 Schema (增强文件策略支持)
  const stateSchema = generateStateSchema(config, contextStrategy);
  Write(`${skillDir}/phases/state-schema.md`, stateSchema);
  
  // 生成编排器 (增强版)
  const orchestrator = generateEnhancedOrchestrator(config, contextStrategy);
  Write(`${skillDir}/phases/orchestrator.md`, orchestrator);
  
  // 生成动作目录
  const actionCatalog = generateActionCatalog(config);
  Write(`${skillDir}/specs/action-catalog.md`, actionCatalog);
  
  // 生成动作文件
  for (const action of config.autonomous_config.actions) {
    const actionContent = generateEnhancedAction(action, config, contextStrategy);
    Write(`${skillDir}/phases/actions/${action.id}.md`, actionContent);
  }
}

// 增强版编排器生成
function generateEnhancedOrchestrator(config, contextStrategy) {
  const actions = config.autonomous_config.actions;
  
  return `# Orchestrator (Enhanced)

增强版编排器,支持声明式动作调度和文件上下文策略。

## 配置

- **上下文策略**: ${contextStrategy}
- **终止条件**: ${config.autonomous_config.termination_conditions?.join(', ') || 'task_completed'}

## 声明式动作目录

\`\`\`javascript
const ACTION_CATALOG = ${JSON.stringify(actions.map(a => ({
  id: a.id,
  name: a.name,
  preconditions: a.preconditions || [],
  effects: a.effects || [],
  priority: a.priority || 0
})), null, 2)};
\`\`\`

## 上下文管理 (${contextStrategy} 策略)

\`\`\`javascript
const ContextManager = {
  ${contextStrategy === 'file' ? `
  // 文件策略: 持久化到 .scratchpad
  init: (workDir) => {
    Bash(\`mkdir -p "\${workDir}/context"\`);
    Write(\`\${workDir}/state.json\`, JSON.stringify(initialState, null, 2));
  },
  
  readState: (workDir) => JSON.parse(Read(\`\${workDir}/state.json\`)),
  
  writeState: (workDir, state) => {
    state.updated_at = new Date().toISOString();
    Write(\`\${workDir}/state.json\`, JSON.stringify(state, null, 2));
  },
  
  readContext: (workDir, key) => {
    try {
      return JSON.parse(Read(\`\${workDir}/context/\${key}.json\`));
    } catch { return null; }
  },
  
  writeContext: (workDir, key, data) => {
    Write(\`\${workDir}/context/\${key}.json\`, JSON.stringify(data, null, 2));
  }` : `
  // 内存策略: 仅在运行时保持
  state: null,
  context: {},
  
  init: (workDir) => {
    ContextManager.state = { ...initialState };
    ContextManager.context = {};
  },
  
  readState: () => ContextManager.state,
  
  writeState: (workDir, state) => {
    state.updated_at = new Date().toISOString();
    ContextManager.state = state;
  },
  
  readContext: (workDir, key) => ContextManager.context[key],
  
  writeContext: (workDir, key, data) => {
    ContextManager.context[key] = data;
  }`}
};
\`\`\`

## 决策逻辑

\`\`\`javascript
function selectNextAction(state) {
  // 1. 终止条件检查
${config.autonomous_config.termination_conditions?.map(c => 
  `  if (${getTerminationCheck(c)}) return null;`
).join('\n') || '  if (state.status === "completed") return null;'}
  
  // 2. 错误限制检查
  if (state.error_count >= 3) return 'action-abort';
  
  // 3. 按优先级选择满足前置条件的动作
  const availableActions = ACTION_CATALOG
    .filter(a => checkPreconditions(a.preconditions, state))
    .filter(a => !state.completed_actions.includes(a.id))
    .sort((a, b) => b.priority - a.priority);
  
  if (availableActions.length > 0) {
    return availableActions[0].id;
  }
  
  // 4. 默认完成
  return 'action-complete';
}

function checkPreconditions(conditions, state) {
  if (!conditions || conditions.length === 0) return true;
  return conditions.every(cond => {
    // 支持多种条件格式
    if (cond.includes('===')) {
      const [left, right] = cond.split('===').map(s => s.trim());
      return eval(\`state.\${left}\`) === eval(right);
    }
    return state[cond] === true;
  });
}
\`\`\`

## 执行循环 (增强版)

\`\`\`javascript
async function runOrchestrator(workDir) {
  console.log('=== Orchestrator Started ===');
  console.log(\`Context Strategy: ${contextStrategy}\`);
  
  // 初始化
  ContextManager.init(workDir);
  
  let iteration = 0;
  const MAX_ITERATIONS = 100;
  
  while (iteration < MAX_ITERATIONS) {
    iteration++;
    
    // 1. 读取状态
    const state = ContextManager.readState(workDir);
    console.log(\`[Iteration \${iteration}] Status: \${state.status}, Completed: \${state.completed_actions.length}\`);
    
    // 2. 选择动作
    const actionId = selectNextAction(state);
    
    if (!actionId) {
      console.log('=== All actions completed ===');
      state.status = 'completed';
      ContextManager.writeState(workDir, state);
      break;
    }
    
    console.log(\`[Iteration \${iteration}] Executing: \${actionId}\`);
    
    // 3. 更新当前动作
    state.current_action = actionId;
    ContextManager.writeState(workDir, state);
    
    // 4. 执行动作
    try {
      const actionPrompt = Read(\`\${skillDir}/phases/actions/\${actionId}.md\`);
      
      const result = await Task({
        subagent_type: 'universal-executor',
        run_in_background: false,
        prompt: \`
[STATE]
\${JSON.stringify(state, null, 2)}

[WORK_DIR]
\${workDir}

[CONTEXT_STRATEGY]
${contextStrategy}

[ACTION]
\${actionPrompt}

[RETURN FORMAT]
Return JSON: { "status": "completed"|"failed", "stateUpdates": {...}, "summary": "..." }
\`
      });
      
      const actionResult = JSON.parse(result);
      
      // 5. 更新状态
      state.completed_actions.push(actionId);
      state.current_action = null;
      Object.assign(state, actionResult.stateUpdates || {});
      
      console.log(\`[Iteration \${iteration}] Completed: \${actionResult.summary || actionId}\`);
      
    } catch (error) {
      console.error(\`[Iteration \${iteration}] Error: \${error.message}\`);
      state.errors.push({
        action: actionId,
        message: error.message,
        timestamp: new Date().toISOString()
      });
      state.error_count++;
      state.current_action = null;
    }
    
    ContextManager.writeState(workDir, state);
  }
  
  console.log('=== Orchestrator Finished ===');
  return ContextManager.readState(workDir);
}
\`\`\`

## 动作目录

| Action | Priority | Preconditions | Effects |
|--------|----------|---------------|---------|
${actions.map(a => 
  `| [${a.id}](actions/${a.id}.md) | ${a.priority || 0} | ${a.preconditions?.join(', ') || '-'} | ${a.effects?.join(', ') || '-'} |`
).join('\n')}

## 调试与恢复

\`\`\`javascript
// 从特定状态恢复
async function resumeFromState(workDir) {
  const state = ContextManager.readState(workDir);
  console.log(\`Resuming from: \${state.current_action || 'start'}\`);
  console.log(\`Completed actions: \${state.completed_actions.join(', ')}\`);
  return runOrchestrator(workDir);
}

// 重试失败的动作
async function retryFailedAction(workDir) {
  const state = ContextManager.readState(workDir);
  if (state.errors.length > 0) {
    const lastError = state.errors[state.errors.length - 1];
    console.log(\`Retrying: \${lastError.action}\`);
    state.error_count = Math.max(0, state.error_count - 1);
    ContextManager.writeState(workDir, state);
    return runOrchestrator(workDir);
  }
}
\`\`\`
`;
}

// 增强版动作生成
function generateEnhancedAction(action, config, contextStrategy) {
  return `# Action: ${action.name}

${action.description || '执行 ' + action.name + ' 操作'}

## Purpose

${action.description || 'TODO: 描述此动作的目的'}

## Preconditions

${action.preconditions?.map(p => `- [ ] \`${p}\``).join('\n') || '- [ ] 无特殊前置条件'}

## Context Access (${contextStrategy} 策略)

\`\`\`javascript
// 读取共享上下文
${contextStrategy === 'file' ?
  `const sharedData = JSON.parse(Read(\`\${workDir}/context/shared.json\`));` :
  `const sharedData = state.context.shared || {};`}

// 写入共享上下文
${contextStrategy === 'file' ?
  `Write(\`\${workDir}/context/shared.json\`, JSON.stringify(updatedData, null, 2));` :
  `state.context.shared = updatedData;`}
\`\`\`

## Execution

\`\`\`javascript
async function execute(state, workDir) {
  // 1. 读取必要数据
  ${contextStrategy === 'file' ?
    `const input = JSON.parse(Read(\`\${workDir}/context/input.json\`));` :
    `const input = state.context.input || {};`}
  
  // 2. 执行核心逻辑
  // TODO: 实现动作逻辑
  const result = {
    // 处理结果
  };
  
  // 3. 保存结果 (${contextStrategy} 策略)
  ${contextStrategy === 'file' ?
    `Write(\`\${workDir}/context/${action.id.replace(/-/g, '_')}_result.json\`, JSON.stringify(result, null, 2));` :
    `// 结果通过 stateUpdates 返回`}
  
  // 4. 返回状态更新
  return {
    status: 'completed',
    stateUpdates: {
      completed_actions: [...state.completed_actions, '${action.id}'],
      ${contextStrategy === 'memory' ? `context: { ...state.context, ${action.id.replace(/-/g, '_')}_result: result }` : '// 文件策略:结果已保存到文件'}
    },
    summary: '${action.name} 完成'
  };
}
\`\`\`

## State Updates

\`\`\`javascript
return {
  status: 'completed',
  stateUpdates: {
    completed_actions: [...state.completed_actions, '${action.id}'],
${action.effects?.map(e => `    // Effect: ${e}`).join('\n') || '    // 无额外效果'}
  }
};
\`\`\`

## Error Handling

| 错误类型 | 处理方式 |
|----------|----------|
| 数据验证失败 | 返回错误,不更新状态 |
| 执行异常 | 记录错误,增加 error_count |
| 上下文读取失败 | 使用默认值或跳过 |

## Next Actions (Hints)

- 成功: 由编排器根据 \`ACTION_CATALOG\` 优先级决定
- 失败: 重试或 \`action-abort\`
`;
}

// 生成动作目录
function generateActionCatalog(config) {
  const actions = config.autonomous_config.actions;
  
  return `# Action Catalog

${config.display_name} 的可用动作目录(声明式)。

## 动作定义

\`\`\`json
${JSON.stringify(actions.map(a => ({
  id: a.id,
  name: a.name,
  description: a.description,
  preconditions: a.preconditions || [],
  effects: a.effects || [],
  priority: a.priority || 0
})), null, 2)}
\`\`\`

## 动作依赖图

\`\`\`mermaid
graph TD
${actions.map((a, i) => {
  const deps = a.preconditions?.filter(p => p.startsWith('completed_actions.includes'))
    .map(p => p.match(/'([^']+)'/)?.[1])
    .filter(Boolean) || [];
  
  if (deps.length === 0 && i === 0) {
    return `    START((Start)) --> ${a.id.replace(/-/g, '_')}[${a.name}]`;
  } else if (deps.length > 0) {
    return deps.map(d => `    ${d.replace(/-/g, '_')} --> ${a.id.replace(/-/g, '_')}[${a.name}]`).join('\n');
  }
  return '';
}).filter(Boolean).join('\n')}
    ${actions[actions.length-1]?.id.replace(/-/g, '_') || 'last'} --> END((End))
\`\`\`

## 选择优先级

| Priority | Action | Description |
|----------|--------|-------------|
${actions.sort((a, b) => (b.priority || 0) - (a.priority || 0)).map(a => 
  `| ${a.priority || 0} | ${a.id} | ${a.description || a.name} |`
).join('\n')}
`;
}

Step 4: 辅助函数

function toPascalCase(str) {
  return str.split('-').map(s => s.charAt(0).toUpperCase() + s.slice(1)).join('');
}

function getDefaultValue(type) {
  if (type.endsWith('[]')) return '[]';
  if (type === 'number') return '0';
  if (type === 'boolean') return 'false';
  if (type === 'string') return '""';
  return '{}';
}

function getTerminationCheck(condition) {
  const checks = {
    'user_exit': 'state.status === "user_exit"',
    'error_limit': 'state.error_count >= 3',
    'task_completed': 'state.status === "completed"',
    'max_iterations': 'iteration >= MAX_ITERATIONS'
  };
  return checks[condition] || `state.${condition}`;
}

function getPreconditionCheck(action) {
  if (!action.preconditions?.length) return 'true';
  return action.preconditions.map(p => `state.${p}`).join(' && ');
}

Output

Sequential 模式

  • phases/_orchestrator.md (声明式编排器)
  • workflow.json (工作流定义)
  • phases/01-{step}.md, 02-{step}.md, ...

Autonomous 模式

  • phases/orchestrator.md (增强版编排器)
  • phases/state-schema.md
  • specs/action-catalog.md (声明式动作目录)
  • phases/actions/action-{name}.md (多个)

Next Phase

Phase 4: Specs & Templates