Refactor agent configurations and context gathering processes across multiple workflow tools

- Removed agent references in context-gather and test-context-gather commands for clarity. - Updated test-task-generate to streamline agent configuration references. - Enhanced action-planning-agent documentation by removing redundant examples. - Adjusted execute command to eliminate direct agent path references. - Improved context-gather documentation to clarify agent autonomy and project.json integration. - Revised test-context-gather to focus on agent delegation and coverage analysis. - Cleaned up test-task-generate to focus on task generation rules without direct agent references. - Implemented automatic agent assignment based on exploration file names in deep analysis phase. - Expanded project exploration phase to include intelligent angle selection based on software type. - Enhanced output schemas and success criteria for exploration tasks to ensure clarity and completeness.
2026-02-05 01:50:27 +08:00 · 2025-12-26 16:20:46 +08:00
parent a5ba7c0f6c
commit c8a914aeca
15 changed files with 751 additions and 106 deletions
--- a/.claude/commands/workflow/execute.md
+++ b/.claude/commands/workflow/execute.md
@@ -410,7 +410,6 @@ Task(subagent_type="{meta.agent}",
     1. Read complete task JSON: {session.task_json_path}
     2. Load context package: {session.context_package_path}

-     Follow complete execution guidelines in @.claude/agents/{meta.agent}.md

     **Session Paths**:
     - Workflow Dir: {session.workflow_dir}
--- a/.claude/commands/workflow/tools/context-gather.md
+++ b/.claude/commands/workflow/tools/context-gather.md
@@ -15,7 +15,6 @@ allowed-tools: Task(*), Read(*), Glob(*)

 Orchestrator command that invokes `context-search-agent` to gather comprehensive project context for implementation planning. Generates standardized `context-package.json` with codebase analysis, dependencies, and conflict detection.

-**Agent**: `context-search-agent` (`.claude/agents/context-search-agent.md`)

 ## Core Philosophy

@@ -431,6 +430,5 @@ if (historicalConflicts.length > 0 && currentRisk === "low") {

 - **Detection-first**: Always check for existing package before invoking agent
 - **Project.json integration**: Agent reads `.workflow/project.json` as primary source for project context, avoiding redundant analysis
- **Agent autonomy**: Agent handles all discovery logic per `.claude/agents/context-search-agent.md`
 - **No redundancy**: This command is a thin orchestrator, all logic in agent
 - **Plan-specific**: Use this for implementation planning; brainstorm mode uses direct agent call
--- a/.claude/commands/workflow/tools/task-generate-tdd.md
+++ b/.claude/commands/workflow/tools/task-generate-tdd.md
@@ -239,15 +239,6 @@ If conflict_risk was medium/high, modifications have been applied to:

 **Agent Configuration Reference**: All TDD task generation rules, quantification requirements, Red-Green-Refactor cycle structure, quality standards, and execution details are defined in action-planning-agent.

-Refer to: @.claude/agents/action-planning-agent.md for:
- TDD Task Decomposition Standards
- Red-Green-Refactor Cycle Requirements
- Quantification Requirements (MANDATORY)
- 5-Field Task JSON Schema
- IMPL_PLAN.md Structure (TDD variant)
- TODO_LIST.md Format
- TDD Execution Flow & Quality Validation
-
 ### TDD-Specific Requirements Summary

 #### Task Structure Philosophy
--- a/.claude/commands/workflow/tools/test-context-gather.md
+++ b/.claude/commands/workflow/tools/test-context-gather.md
@@ -14,7 +14,7 @@ allowed-tools: Task(*), Read(*), Glob(*)

 Orchestrator command that invokes `test-context-search-agent` to gather comprehensive test coverage context for test generation workflows. Generates standardized `test-context-package.json` with coverage analysis, framework detection, and source implementation context.

-**Agent**: `test-context-search-agent` (`.claude/agents/test-context-search-agent.md`)
+

 ## Core Philosophy

@@ -89,7 +89,6 @@ Task(
  run_in_background=false,
  description="Gather test coverage context",
  prompt=`
-You are executing as test-context-search-agent (.claude/agents/test-context-search-agent.md).

 ## Execution Mode
 **PLAN MODE** (Comprehensive) - Full Phase 1-3 execution
@@ -229,7 +228,7 @@ Refer to `test-context-search-agent.md` Phase 3.2 for complete `test-context-pac
 ## Notes

 - **Detection-first**: Always check for existing test-context-package before invoking agent
- **Agent autonomy**: Agent handles all coverage analysis logic per `.claude/agents/test-context-search-agent.md`
+
 - **No redundancy**: This command is a thin orchestrator, all logic in agent
 - **Framework agnostic**: Supports Jest, Mocha, pytest, RSpec, Go testing, etc.
 - **Coverage focus**: Primary goal is identifying implementation files without tests
--- a/.claude/commands/workflow/tools/test-task-generate.md
+++ b/.claude/commands/workflow/tools/test-task-generate.md
@@ -107,8 +107,6 @@ CRITICAL:
 - Follow the progressive loading strategy defined in your agent specification (load context incrementally from memory-first approach)

 ## AGENT CONFIGURATION REFERENCE
-All test task generation rules, schemas, and quality standards are defined in your agent specification:
-@.claude/agents/action-planning-agent.md

 Refer to your specification for:
 - Test Task JSON Schema (6-field structure with test-specific metadata)
--- a/.claude/skills/command-guide/reference/agents/action-planning-agent.md
+++ b/.claude/skills/command-guide/reference/agents/action-planning-agent.md
@@ -806,8 +806,6 @@ Use `analysis_results.complexity` or task count to determine structure:
 **Examples**:
 - GOOD: `"Implement 5 commands: [cmd1, cmd2, cmd3, cmd4, cmd5]"`
 - BAD: `"Implement new commands"`
- GOOD: `"5 files created: verify by ls .claude/commands/*.md | wc -l = 5"`
- BAD: `"All commands implemented successfully"`

 ### 3.2 Planning & Organization Standards

--- a/.claude/skills/command-guide/reference/commands/workflow/execute.md
+++ b/.claude/skills/command-guide/reference/commands/workflow/execute.md
@@ -400,7 +400,7 @@ Task(subagent_type="{meta.agent}",
     1. Read complete task JSON: {session.task_json_path}
     2. Load context package: {session.context_package_path}

-     Follow complete execution guidelines in @.claude/agents/{meta.agent}.md
+

     **Session Paths**:
     - Workflow Dir: {session.workflow_dir}
--- a/.claude/skills/command-guide/reference/commands/workflow/tools/context-gather.md
+++ b/.claude/skills/command-guide/reference/commands/workflow/tools/context-gather.md
@@ -15,7 +15,6 @@ allowed-tools: Task(*), Read(*), Glob(*)

 Orchestrator command that invokes `context-search-agent` to gather comprehensive project context for implementation planning. Generates standardized `context-package.json` with codebase analysis, dependencies, and conflict detection.

-**Agent**: `context-search-agent` (`.claude/agents/context-search-agent.md`)

 ## Core Philosophy

@@ -429,6 +428,6 @@ if (historicalConflicts.length > 0 && currentRisk === "low") {

 - **Detection-first**: Always check for existing package before invoking agent
 - **Project.json integration**: Agent reads `.workflow/project.json` as primary source for project context, avoiding redundant analysis
- **Agent autonomy**: Agent handles all discovery logic per `.claude/agents/context-search-agent.md`
+
 - **No redundancy**: This command is a thin orchestrator, all logic in agent
 - **Plan-specific**: Use this for implementation planning; brainstorm mode uses direct agent call
--- a/.claude/skills/command-guide/reference/commands/workflow/tools/task-generate-tdd.md
+++ b/.claude/skills/command-guide/reference/commands/workflow/tools/task-generate-tdd.md
@@ -238,14 +238,7 @@ If conflict_risk was medium/high, modifications have been applied to:

 **Agent Configuration Reference**: All TDD task generation rules, quantification requirements, Red-Green-Refactor cycle structure, quality standards, and execution details are defined in action-planning-agent.

-Refer to: @.claude/agents/action-planning-agent.md for:
- TDD Task Decomposition Standards
- Red-Green-Refactor Cycle Requirements
- Quantification Requirements (MANDATORY)
- 5-Field Task JSON Schema
- IMPL_PLAN.md Structure (TDD variant)
- TODO_LIST.md Format
- TDD Execution Flow & Quality Validation
+

 ### TDD-Specific Requirements Summary

--- a/.claude/skills/command-guide/reference/commands/workflow/tools/test-context-gather.md
+++ b/.claude/skills/command-guide/reference/commands/workflow/tools/test-context-gather.md
@@ -14,8 +14,6 @@ allowed-tools: Task(*), Read(*), Glob(*)

 Orchestrator command that invokes `test-context-search-agent` to gather comprehensive test coverage context for test generation workflows. Generates standardized `test-context-package.json` with coverage analysis, framework detection, and source implementation context.

-**Agent**: `test-context-search-agent` (`.claude/agents/test-context-search-agent.md`)
-
 ## Core Philosophy

 - **Agent Delegation**: Delegate all test coverage analysis to `test-context-search-agent` for autonomous execution
@@ -88,7 +86,6 @@ Task(
  subagent_type="test-context-search-agent",
  description="Gather test coverage context",
  prompt=`
-You are executing as test-context-search-agent (.claude/agents/test-context-search-agent.md).

 ## Execution Mode
 **PLAN MODE** (Comprehensive) - Full Phase 1-3 execution
@@ -228,7 +225,7 @@ Refer to `test-context-search-agent.md` Phase 3.2 for complete `test-context-pac
 ## Notes

 - **Detection-first**: Always check for existing test-context-package before invoking agent
- **Agent autonomy**: Agent handles all coverage analysis logic per `.claude/agents/test-context-search-agent.md`
+
 - **No redundancy**: This command is a thin orchestrator, all logic in agent
 - **Framework agnostic**: Supports Jest, Mocha, pytest, RSpec, Go testing, etc.
 - **Coverage focus**: Primary goal is identifying implementation files without tests
--- a/.claude/skills/command-guide/reference/commands/workflow/tools/test-task-generate.md
+++ b/.claude/skills/command-guide/reference/commands/workflow/tools/test-task-generate.md
@@ -106,8 +106,6 @@ CRITICAL:
 - Follow the progressive loading strategy defined in your agent specification (load context incrementally from memory-first approach)

 ## AGENT CONFIGURATION REFERENCE
-All test task generation rules, schemas, and quality standards are defined in your agent specification:
-@.claude/agents/action-planning-agent.md

 Refer to your specification for:
 - Test Task JSON Schema (6-field structure with test-specific metadata)
--- a/.claude/skills/copyright-docs/phases/01.5-project-exploration.md
+++ b/.claude/skills/copyright-docs/phases/01.5-project-exploration.md
@@ -0,0 +1,150 @@
+# Phase 1.5: Project Exploration
+
+基于元数据，启动并行探索 Agent 收集代码信息。
+
+## Execution
+
+### Step 1: Intelligent Angle Selection
+
+```javascript
+// 根据软件类型选择探索角度
+const ANGLE_PRESETS = {
+  'CLI': ['architecture', 'commands', 'algorithms', 'exceptions'],
+  'API': ['architecture', 'endpoints', 'data-structures', 'interfaces'],
+  'SDK': ['architecture', 'interfaces', 'data-structures', 'algorithms'],
+  'DataProcessing': ['architecture', 'algorithms', 'data-structures', 'dataflow'],
+  'Automation': ['architecture', 'algorithms', 'exceptions', 'dataflow']
+};
+
+// 从 metadata.category 映射到预设
+function getCategoryKey(category) {
+  if (category.includes('CLI') || category.includes('命令行')) return 'CLI';
+  if (category.includes('API') || category.includes('后端')) return 'API';
+  if (category.includes('SDK') || category.includes('库')) return 'SDK';
+  if (category.includes('数据处理')) return 'DataProcessing';
+  if (category.includes('自动化')) return 'Automation';
+  return 'API'; // default
+}
+
+const categoryKey = getCategoryKey(metadata.category);
+const selectedAngles = ANGLE_PRESETS[categoryKey];
+
+console.log(`
+## Exploration Plan
+
+Software: ${metadata.software_name}
+Category: ${metadata.category} → ${categoryKey}
+Selected Angles: ${selectedAngles.join(', ')}
+
+Launching ${selectedAngles.length} parallel explorations...
+`);
+```
+
+### Step 2: Launch Parallel Agents (Direct Output)
+
+**⚠️ CRITICAL**: Agents write output files directly.
+
+```javascript
+const explorationTasks = selectedAngles.map((angle, index) =>
+  Task({
+    subagent_type: "cli-explore-agent",
+    run_in_background: false,
+    description: `Explore: ${angle}`,
+    prompt: `
+## Exploration Objective
+为 CPCC 软著申请文档执行 **${angle}** 探索。
+
+## Assigned Context
+- **Exploration Angle**: ${angle}
+- **Software Name**: ${metadata.software_name}
+- **Scope Path**: ${metadata.scope_path}
+- **Category**: ${metadata.category}
+- **Exploration Index**: ${index + 1} of ${selectedAngles.length}
+- **Output File**: ${sessionFolder}/exploration-${angle}.json
+
+## MANDATORY FIRST STEPS
+1. Run: ccw tool exec get_modules_by_depth '{}' (project structure)
+2. Run: rg -l "{relevant_keyword}" --type ts (locate relevant files)
+3. Analyze from ${angle} perspective
+
+## Exploration Strategy (${angle} focus)
+
+**Step 1: Structural Scan**
+- 识别与 ${angle} 相关的模块和文件
+- 分析导入/导出关系
+
+**Step 2: Pattern Recognition**
+- ${angle} 相关的设计模式
+- 代码组织方式
+
+**Step 3: Write Output**
+- 输出 JSON 到指定路径
+
+## Expected Output Schema
+
+**File**: ${sessionFolder}/exploration-${angle}.json
+
+\`\`\`json
+{
+  "angle": "${angle}",
+  "findings": {
+    "structure": [
+      { "component": "...", "type": "module|layer|service", "path": "...", "description": "..." }
+    ],
+    "patterns": [
+      { "name": "...", "usage": "...", "files": ["path1", "path2"] }
+    ],
+    "key_files": [
+      { "path": "src/file.ts", "relevance": 0.85, "rationale": "Core ${angle} logic" }
+    ]
+  },
+  "insights": [
+    { "observation": "...", "cpcc_section": "2|3|4|5|6|7", "recommendation": "..." }
+  ],
+  "_metadata": {
+    "exploration_angle": "${angle}",
+    "exploration_index": ${index + 1},
+    "software_name": "${metadata.software_name}",
+    "timestamp": "ISO8601"
+  }
+}
+\`\`\`
+
+## Success Criteria
+- [ ] get_modules_by_depth 执行完成
+- [ ] 至少识别 3 个相关文件
+- [ ] patterns 包含具体代码示例
+- [ ] insights 关联到 CPCC 章节 (2-7)
+- [ ] JSON 输出到指定路径
+- [ ] Return: 2-3 句话总结 ${angle} 发现
+`
+  })
+);
+
+// Execute all exploration tasks in parallel
+```
+
+## Output
+
+Session folder structure after exploration:
+
+```
+${sessionFolder}/
+├── exploration-architecture.json
+├── exploration-{angle2}.json
+├── exploration-{angle3}.json
+└── exploration-{angle4}.json
+```
+
+## Downstream Usage (Phase 2 Analysis Input)
+
+Phase 2 agents read exploration files as context:
+
+```javascript
+// Discover exploration files by known angle pattern
+const explorationData = {};
+selectedAngles.forEach(angle => {
+  const filePath = `${sessionFolder}/exploration-${angle}.json`;
+  explorationData[angle] = JSON.parse(Read(filePath));
+});
+```
--- a/.claude/skills/copyright-docs/phases/02-deep-analysis.md
+++ b/.claude/skills/copyright-docs/phases/02-deep-analysis.md
@@ -5,15 +5,161 @@
 > **模板参考**: [../templates/agent-base.md](../templates/agent-base.md)
 > **规范参考**: [../specs/cpcc-requirements.md](../specs/cpcc-requirements.md)

-## Agent 执行前置条件
+## Exploration → Agent 自动分配

-**每个 Agent 必须首先读取以下规范文件**：
+根据 Phase 1.5 生成的 exploration 文件名自动分配对应的 analysis agent。
+
+### 映射规则

 ```javascript
-// Agent 启动时的第一步操作
-const specs = {
-  cpcc: Read(`${skillRoot}/specs/cpcc-requirements.md`)
+// Exploration 角度 → Agent 映射（基于文件名识别，不读取内容）
+const EXPLORATION_TO_AGENT = {
+  'architecture': 'architecture',
+  'commands': 'functions',      // CLI 命令 → 功能模块
+  'endpoints': 'interfaces',    // API 端点 → 接口设计
+  'algorithms': 'algorithms',
+  'data-structures': 'data_structures',
+  'dataflow': 'data_structures', // 数据流 → 数据结构
+  'interfaces': 'interfaces',
+  'exceptions': 'exceptions'
 };
+
+// 从文件名提取角度
+function extractAngle(filename) {
+  // exploration-architecture.json → architecture
+  const match = filename.match(/exploration-(.+)\.json$/);
+  return match ? match[1] : null;
+}
+
+// 分配 agent
+function assignAgent(explorationFile) {
+  const angle = extractAngle(path.basename(explorationFile));
+  return EXPLORATION_TO_AGENT[angle] || null;
+}
+
+// Agent 配置（用于 buildAgentPrompt）
+const AGENT_CONFIGS = {
+  architecture: {
+    role: '系统架构师，专注于分层设计和模块依赖',
+    section: '2',
+    output: 'section-2-architecture.md',
+    focus: '分层结构、模块依赖、数据流向'
+  },
+  functions: {
+    role: '功能分析师，专注于功能点识别和交互',
+    section: '3',
+    output: 'section-3-functions.md',
+    focus: '功能点枚举、模块分组、入口文件、功能交互'
+  },
+  algorithms: {
+    role: '算法工程师，专注于核心逻辑和复杂度分析',
+    section: '4',
+    output: 'section-4-algorithms.md',
+    focus: '核心算法、流程步骤、复杂度、输入输出'
+  },
+  data_structures: {
+    role: '数据建模师，专注于实体关系和类型定义',
+    section: '5',
+    output: 'section-5-data-structures.md',
+    focus: '实体定义、属性类型、关系映射、枚举'
+  },
+  interfaces: {
+    role: 'API设计师，专注于接口契约和协议',
+    section: '6',
+    output: 'section-6-interfaces.md',
+    focus: 'API端点、参数校验、响应格式、时序'
+  },
+  exceptions: {
+    role: '可靠性工程师，专注于异常处理和恢复策略',
+    section: '7',
+    output: 'section-7-exceptions.md',
+    focus: '异常类型、错误码、处理模式、恢复策略'
+  }
+};
+```
+
+### 自动发现与分配流程
+
+```javascript
+// 1. 发现所有 exploration 文件（仅看文件名）
+const explorationFiles = bash(`find ${sessionFolder} -name "exploration-*.json" -type f`)
+  .split('\n')
+  .filter(f => f.trim());
+
+// 2. 按文件名自动分配 agent
+const agentAssignments = explorationFiles.map(file => {
+  const angle = extractAngle(path.basename(file));
+  const agentName = EXPLORATION_TO_AGENT[angle];
+  return {
+    exploration_file: file,
+    angle: angle,
+    agent: agentName,
+    output_file: AGENT_CONFIGS[agentName]?.output
+  };
+}).filter(a => a.agent);
+
+// 3. 补充未被 exploration 覆盖的必需 agent（分配相关 exploration）
+const coveredAgents = new Set(agentAssignments.map(a => a.agent));
+const requiredAgents = ['architecture', 'functions', 'algorithms', 'data_structures', 'interfaces', 'exceptions'];
+const missingAgents = requiredAgents.filter(a => !coveredAgents.has(a));
+
+// 相关性映射：为缺失 agent 分配最相关的 exploration
+const RELATED_EXPLORATIONS = {
+  architecture: ['architecture', 'dataflow', 'interfaces'],
+  functions: ['commands', 'endpoints', 'architecture'],
+  algorithms: ['algorithms', 'dataflow', 'architecture'],
+  data_structures: ['data-structures', 'dataflow', 'architecture'],
+  interfaces: ['interfaces', 'endpoints', 'architecture'],
+  exceptions: ['exceptions', 'algorithms', 'architecture']
+};
+
+function findRelatedExploration(agent, availableFiles) {
+  const preferences = RELATED_EXPLORATIONS[agent] || ['architecture'];
+  for (const pref of preferences) {
+    const match = availableFiles.find(f => f.includes(`exploration-${pref}.json`));
+    if (match) return { file: match, angle: pref, isRelated: true };
+  }
+  // 最后兜底：任意 exploration 都比没有强
+  return availableFiles.length > 0
+    ? { file: availableFiles[0], angle: extractAngle(path.basename(availableFiles[0])), isRelated: true }
+    : { file: null, angle: null, isRelated: false };
+}
+
+missingAgents.forEach(agent => {
+  const related = findRelatedExploration(agent, explorationFiles);
+  agentAssignments.push({
+    exploration_file: related.file,
+    angle: related.angle,
+    agent: agent,
+    output_file: AGENT_CONFIGS[agent].output,
+    is_related: related.isRelated  // 标记为相关而非直接匹配
+  });
+});
+
+console.log(`
+## Agent Auto-Assignment
+
+Found ${explorationFiles.length} exploration files:
+${agentAssignments.map(a => {
+  if (!a.exploration_file) return `- ${a.agent} agent (no exploration)`;
+  if (a.is_related) return `- ${a.agent} agent ← ${a.angle} (related)`;
+  return `- ${a.agent} agent ← ${a.angle} (direct)`;
+}).join('\n')}
+`);
+```
+
+---
+
+## Agent 执行前置条件
+
+**每个 Agent 接收 exploration 文件路径，自行读取内容**：
+
+```javascript
+// Agent prompt 中包含文件路径
+// Agent 启动后的操作顺序：
+// 1. Read exploration 文件（如有）
+// 2. Read CPCC 规范文件
+// 3. 执行分析任务
 ```

 规范文件路径（相对于 skill 根目录）：
@@ -47,26 +193,90 @@ const specs = {
 ## 执行流程

 ```javascript
-// 1. 准备目录
+// 1. 发现 exploration 文件并自动分配 agent
+const explorationFiles = bash(`find ${sessionFolder} -name "exploration-*.json" -type f`)
+  .split('\n')
+  .filter(f => f.trim());
+
+const agentAssignments = explorationFiles.map(file => {
+  const angle = extractAngle(path.basename(file));
+  const agentName = EXPLORATION_TO_AGENT[angle];
+  return { exploration_file: file, angle, agent: agentName };
+}).filter(a => a.agent);
+
+// 补充必需 agent
+const coveredAgents = new Set(agentAssignments.map(a => a.agent));
+const requiredAgents = ['architecture', 'functions', 'algorithms', 'data_structures', 'interfaces', 'exceptions'];
+requiredAgents.filter(a => !coveredAgents.has(a)).forEach(agent => {
+  agentAssignments.push({ exploration_file: null, angle: null, agent });
+});
+
+// 2. 准备目录
 Bash(`mkdir -p ${outputDir}/sections`);

-// 2. 并行启动 6 个 Agent
-const results = await Promise.all([
-  launchAgent('architecture', metadata, outputDir),
-  launchAgent('functions', metadata, outputDir),
-  launchAgent('algorithms', metadata, outputDir),
-  launchAgent('data_structures', metadata, outputDir),
-  launchAgent('interfaces', metadata, outputDir),
-  launchAgent('exceptions', metadata, outputDir)
-]);
+// 3. 并行启动所有 Agent（传递 exploration 文件路径）
+const results = await Promise.all(
+  agentAssignments.map(assignment =>
+    Task({
+      subagent_type: "cli-explore-agent",
+      run_in_background: false,
+      description: `Analyze: ${assignment.agent}`,
+      prompt: buildAgentPrompt(assignment, metadata, outputDir)
+    })
+  )
+);

-// 3. 收集返回信息
+// 4. 收集返回信息
 const summaries = results.map(r => JSON.parse(r));

-// 4. 传递给 Phase 2.5
+// 5. 传递给 Phase 2.5
 return { summaries, cross_notes: summaries.flatMap(s => s.cross_module_notes) };
 ```

+### Agent Prompt 构建
+
+```javascript
+function buildAgentPrompt(assignment, metadata, outputDir) {
+  const config = AGENT_CONFIGS[assignment.agent];
+  let contextSection = '';
+
+  if (assignment.exploration_file) {
+    const matchType = assignment.is_related ? '相关' : '直接匹配';
+    contextSection = `[CONTEXT]
+**Exploration 文件**: ${assignment.exploration_file}
+**匹配类型**: ${matchType}
+首先读取此文件获取 ${assignment.angle} 探索结果作为分析上下文。
+${assignment.is_related ? `注意：这是相关探索结果（非直接匹配），请提取与 ${config.focus} 相关的信息。` : ''}
+`;
+  }
+
+  return `
+${contextSection}
+[SPEC]
+读取规范文件：
+- Read: ${skillRoot}/specs/cpcc-requirements.md
+
+[ROLE] ${config.role}
+
+[TASK]
+分析 ${metadata.scope_path}，生成 Section ${config.section}。
+输出: ${outputDir}/sections/${config.output}
+
+[CPCC_SPEC]
+- 内容基于代码分析，无臆测
+- 图表编号: 图${config.section}-1, 图${config.section}-2...
+- 每个子章节 ≥100字
+- 包含文件路径引用
+
+[FOCUS]
+${config.focus}
+
+[RETURN JSON]
+{"status":"completed","output_file":"${config.output}","summary":"<50字>","cross_module_notes":[],"stats":{}}
+`;
+}
+```
+
 ---

 ## Agent 提示词
--- a/.claude/skills/project-analyze/phases/02-project-exploration.md
+++ b/.claude/skills/project-analyze/phases/02-project-exploration.md
@@ -1,75 +1,176 @@
 # Phase 2: Project Exploration

-Launch parallel exploration agents based on report type.
+Launch parallel exploration agents based on report type and task context.

 ## Execution

-### Step 1: Map Exploration Angles
+### Step 1: Intelligent Angle Selection

 ```javascript
-const angleMapping = {
-  architecture: ["Layer Structure", "Module Dependencies", "Entry Points", "Data Flow"],
-  design: ["Design Patterns", "Class Relationships", "Interface Contracts", "State Management"],
-  methods: ["Core Algorithms", "Critical Paths", "Public APIs", "Complex Logic"],
-  comprehensive: ["Layer Structure", "Design Patterns", "Core Algorithms", "Data Flow"]
+// Angle presets based on report type (adapted from lite-plan.md)
+const ANGLE_PRESETS = {
+  architecture: ['layer-structure', 'module-dependencies', 'entry-points', 'data-flow'],
+  design: ['design-patterns', 'class-relationships', 'interface-contracts', 'state-management'],
+  methods: ['core-algorithms', 'critical-paths', 'public-apis', 'complex-logic'],
+  comprehensive: ['architecture', 'patterns', 'dependencies', 'integration-points']
 };

-const angles = angleMapping[config.type];
+// Depth-based angle count
+const angleCount = {
+  shallow: 2,
+  standard: 3,
+  deep: 4
+};
+
+function selectAngles(reportType, depth) {
+  const preset = ANGLE_PRESETS[reportType] || ANGLE_PRESETS.comprehensive;
+  const count = angleCount[depth] || 3;
+  return preset.slice(0, count);
+}
+
+const selectedAngles = selectAngles(config.type, config.depth);
+
+console.log(`
+## Exploration Plan
+
+Report Type: ${config.type}
+Depth: ${config.depth}
+Selected Angles: ${selectedAngles.join(', ')}
+
+Launching ${selectedAngles.length} parallel explorations...
+`);
 ```

-### Step 2: Launch Parallel Agents
+### Step 2: Launch Parallel Agents (Direct Output)

-For each angle, launch an exploration agent:
+**⚠️ CRITICAL**: Agents write output files directly. No aggregation needed.

 ```javascript
-Task({
-  subagent_type: "cli-explore-agent",
-  run_in_background: false,
-  description: `Explore: ${angle}`,
-  prompt: `
+// Launch agents with pre-assigned angles
+const explorationTasks = selectedAngles.map((angle, index) =>
+  Task({
+    subagent_type: "cli-explore-agent",
+    run_in_background: false,  // ⚠️ MANDATORY: Must wait for results
+    description: `Explore: ${angle}`,
+    prompt: `
 ## Exploration Objective
-Execute **${angle}** exploration for project analysis report.
+Execute **${angle}** exploration for ${config.type} project analysis report.

-## Context
- **Angle**: ${angle}
+## Assigned Context
+- **Exploration Angle**: ${angle}
 - **Report Type**: ${config.type}
 - **Depth**: ${config.depth}
 - **Scope**: ${config.scope}
+- **Exploration Index**: ${index + 1} of ${selectedAngles.length}
+- **Output File**: ${sessionFolder}/exploration-${angle}.json

-## Exploration Protocol
-1. Structural Discovery (get_modules_by_depth, rg, glob)
-2. Pattern Recognition (conventions, naming, organization)
-3. Relationship Mapping (dependencies, integration points)
+## MANDATORY FIRST STEPS (Execute by Agent)
+**You (cli-explore-agent) MUST execute these steps in order:**
+1. Run: ccw tool exec get_modules_by_depth '{}' (project structure)
+2. Run: rg -l "{relevant_keyword}" --type ts (locate relevant files)
+3. Analyze project from ${angle} perspective

-## Output Format
+## Exploration Strategy (${angle} focus)
+
+**Step 1: Structural Scan** (Bash)
+- get_modules_by_depth.sh → identify modules related to ${angle}
+- find/rg → locate files relevant to ${angle} aspect
+- Analyze imports/dependencies from ${angle} perspective
+
+**Step 2: Semantic Analysis** (Gemini/Qwen CLI)
+- How does existing code handle ${angle} concerns?
+- What patterns are used for ${angle}?
+- Identify key architectural decisions related to ${angle}
+
+**Step 3: Write Output Directly**
+- Consolidate ${angle} findings into JSON
+- Write to output file path specified above
+
+## Expected Output Schema
+
+**File**: ${sessionFolder}/exploration-${angle}.json
+
+\`\`\`json
 {
  "angle": "${angle}",
  "findings": {
-    "structure": [...],
-    "patterns": [...],
-    "relationships": [...],
-    "key_files": [{path, relevance, rationale}]
+    "structure": [
+      { "component": "...", "type": "module|layer|service", "description": "..." }
+    ],
+    "patterns": [
+      { "name": "...", "usage": "...", "files": ["path1", "path2"] }
+    ],
+    "relationships": [
+      { "from": "...", "to": "...", "type": "depends|imports|calls", "strength": "high|medium|low" }
+    ],
+    "key_files": [
+      { "path": "src/file.ts", "relevance": 0.85, "rationale": "Core ${angle} logic" }
+    ]
  },
-  "insights": [...]
+  "insights": [
+    { "observation": "...", "impact": "high|medium|low", "recommendation": "..." }
+  ],
+  "_metadata": {
+    "exploration_angle": "${angle}",
+    "exploration_index": ${index + 1},
+    "report_type": "${config.type}",
+    "timestamp": "ISO8601"
+  }
 }
+\`\`\`
+
+## Success Criteria
+- [ ] get_modules_by_depth.sh executed
+- [ ] At least 3 relevant files identified with ${angle} rationale
+- [ ] Patterns are actionable (code examples, not generic advice)
+- [ ] Relationships include concrete file references
+- [ ] JSON output written to ${sessionFolder}/exploration-${angle}.json
+- [ ] Return: 2-3 sentence summary of ${angle} findings
 `
-})
-```
+  })
+);

-### Step 3: Aggregate Results
-
-Merge all exploration results into unified findings:
-
-```javascript
-const aggregatedFindings = {
-  structure: [],      // from all angles
-  patterns: [],       // from all angles  
-  relationships: [],  // from all angles
-  key_files: [],      // deduplicated
-  insights: []        // prioritized
-};
+// Execute all exploration tasks in parallel
 ```

 ## Output

-Save exploration results to `exploration-{angle}.json` files.
+Session folder structure after exploration:
+
+```
+${sessionFolder}/
+├── exploration-{angle1}.json      # Agent 1 direct output
+├── exploration-{angle2}.json      # Agent 2 direct output
+├── exploration-{angle3}.json      # Agent 3 direct output (if applicable)
+└── exploration-{angle4}.json      # Agent 4 direct output (if applicable)
+```
+
+## Downstream Usage (Phase 3 Analysis Input)
+
+Subsequent analysis phases MUST read exploration outputs as input:
+
+```javascript
+// Discover exploration files by known angle pattern
+const explorationData = {};
+selectedAngles.forEach(angle => {
+  const filePath = `${sessionFolder}/exploration-${angle}.json`;
+  explorationData[angle] = JSON.parse(Read(filePath));
+});
+
+// Pass to analysis agent
+Task({
+  subagent_type: "analysis-agent",
+  prompt: `
+## Analysis Input
+
+### Exploration Data by Angle
+${Object.entries(explorationData).map(([angle, data]) => `
+#### ${angle}
+${JSON.stringify(data, null, 2)}
+`).join('\n')}
+
+## Analysis Task
+Synthesize findings from all exploration angles...
+`
+});
+```
--- a/.claude/skills/project-analyze/phases/03-deep-analysis.md
+++ b/.claude/skills/project-analyze/phases/03-deep-analysis.md
@@ -5,16 +5,176 @@
 > **规范参考**: [../specs/quality-standards.md](../specs/quality-standards.md)
 > **写作风格**: [../specs/writing-style.md](../specs/writing-style.md)

-## Agent 执行前置条件
+## Exploration → Agent 自动分配

-**每个 Agent 必须首先读取以下规范文件**：
+根据 Phase 2 生成的 exploration 文件名自动分配对应的 analysis agent。
+
+### 映射规则

 ```javascript
-// Agent 启动时的第一步操作
-const specs = {
-  quality: Read(`${skillRoot}/specs/quality-standards.md`),
-  style: Read(`${skillRoot}/specs/writing-style.md`)
+// Exploration 角度 → Agent 映射（基于文件名识别，不读取内容）
+const EXPLORATION_TO_AGENT = {
+  // Architecture Report 角度
+  'layer-structure': 'layers',
+  'module-dependencies': 'dependencies',
+  'entry-points': 'entrypoints',
+  'data-flow': 'dataflow',
+
+  // Design Report 角度
+  'design-patterns': 'patterns',
+  'class-relationships': 'classes',
+  'interface-contracts': 'interfaces',
+  'state-management': 'state',
+
+  // Methods Report 角度
+  'core-algorithms': 'algorithms',
+  'critical-paths': 'paths',
+  'public-apis': 'apis',
+  'complex-logic': 'logic',
+
+  // Comprehensive 角度
+  'architecture': 'overview',
+  'patterns': 'patterns',
+  'dependencies': 'dependencies',
+  'integration-points': 'entrypoints'
 };
+
+// 从文件名提取角度
+function extractAngle(filename) {
+  // exploration-layer-structure.json → layer-structure
+  const match = filename.match(/exploration-(.+)\.json$/);
+  return match ? match[1] : null;
+}
+
+// 分配 agent
+function assignAgent(explorationFile) {
+  const angle = extractAngle(path.basename(explorationFile));
+  return EXPLORATION_TO_AGENT[angle] || null;
+}
+
+// Agent 配置（用于 buildAgentPrompt）
+const AGENT_CONFIGS = {
+  overview: {
+    role: '首席系统架构师',
+    task: '基于代码库全貌，撰写"总体架构"章节，洞察核心价值主张和顶层技术决策',
+    focus: '领域边界与定位、架构范式、核心技术决策、顶层模块划分',
+    constraint: '避免罗列目录结构，重点阐述设计意图，包含至少1个 Mermaid 架构图'
+  },
+  layers: {
+    role: '资深软件设计师',
+    task: '分析系统逻辑分层结构，撰写"逻辑视点与分层架构"章节',
+    focus: '职责分配体系、数据流向与约束、边界隔离策略、异常处理流',
+    constraint: '不要列举具体文件名，关注层级间契约和隔离艺术'
+  },
+  dependencies: {
+    role: '集成架构专家',
+    task: '审视系统外部连接与内部耦合，撰写"依赖管理与生态集成"章节',
+    focus: '外部集成拓扑、核心依赖分析、依赖注入与控制反转、供应链安全',
+    constraint: '禁止简单列出依赖配置，必须分析集成策略和风险控制模型'
+  },
+  dataflow: {
+    role: '数据架构师',
+    task: '追踪系统数据流转机制，撰写"数据流与状态管理"章节',
+    focus: '数据入口与出口、数据转换管道、持久化策略、一致性保障',
+    constraint: '关注数据生命周期和形态演变，不要罗列数据库表结构'
+  },
+  entrypoints: {
+    role: '系统边界分析师',
+    task: '识别系统入口设计和关键路径，撰写"系统入口与调用链"章节',
+    focus: '入口类型与职责、请求处理管道、关键业务路径、异常与边界处理',
+    constraint: '关注入口设计哲学，不要逐个列举所有端点'
+  },
+  patterns: {
+    role: '核心开发规范制定者',
+    task: '挖掘代码中的复用机制和标准化实践，撰写"设计模式与工程规范"章节',
+    focus: '架构级模式、通信与并发模式、横切关注点实现、抽象与复用策略',
+    constraint: '避免教科书式解释，必须结合项目上下文说明应用场景'
+  },
+  classes: {
+    role: '领域模型设计师',
+    task: '分析系统类型体系和领域模型，撰写"类型体系与领域建模"章节',
+    focus: '领域模型设计、继承与组合策略、职责分配原则、类型安全与约束',
+    constraint: '关注建模思想，用 UML 类图辅助说明核心关系'
+  },
+  interfaces: {
+    role: '契约设计专家',
+    task: '分析系统接口设计和抽象层次，撰写"接口契约与抽象设计"章节',
+    focus: '抽象层次设计、契约与实现分离、扩展点设计、版本演进策略',
+    constraint: '关注接口设计哲学，不要逐个列举接口方法签名'
+  },
+  state: {
+    role: '状态管理架构师',
+    task: '分析系统状态管理机制，撰写"状态管理与生命周期"章节',
+    focus: '状态模型设计、状态生命周期、并发与一致性、状态恢复与容错',
+    constraint: '关注状态管理设计决策，不要列举具体变量名'
+  },
+  algorithms: {
+    role: '算法架构师',
+    task: '分析系统核心算法设计，撰写"核心算法与计算模型"章节',
+    focus: '算法选型与权衡、计算模型设计、性能与可扩展性、正确性保障',
+    constraint: '关注算法思想，用流程图辅助说明复杂逻辑'
+  },
+  paths: {
+    role: '性能架构师',
+    task: '分析系统关键执行路径，撰写"关键路径与性能设计"章节',
+    focus: '关键业务路径、性能敏感区域、瓶颈识别与缓解、降级与熔断',
+    constraint: '关注路径设计战略考量，不要罗列所有代码执行步骤'
+  },
+  apis: {
+    role: 'API 设计规范专家',
+    task: '分析系统对外接口设计规范，撰写"API 设计与规范"章节',
+    focus: 'API 设计风格、命名与结构规范、版本管理策略、错误处理规范',
+    constraint: '关注设计规范和一致性，不要逐个列举所有 API 端点'
+  },
+  logic: {
+    role: '业务逻辑架构师',
+    task: '分析系统业务逻辑建模，撰写"业务逻辑与规则引擎"章节',
+    focus: '业务规则建模、决策点设计、边界条件处理、业务流程编排',
+    constraint: '关注业务逻辑组织方式，不要逐行解释代码逻辑'
+  }
+};
+```
+
+### 自动发现与分配流程
+
+```javascript
+// 1. 发现所有 exploration 文件（仅看文件名）
+const explorationFiles = bash(`find ${sessionFolder} -name "exploration-*.json" -type f`)
+  .split('\n')
+  .filter(f => f.trim());
+
+// 2. 按文件名自动分配 agent
+const agentAssignments = explorationFiles.map(file => {
+  const angle = extractAngle(path.basename(file));
+  const agentName = EXPLORATION_TO_AGENT[angle];
+  return {
+    exploration_file: file,
+    angle: angle,
+    agent: agentName,
+    output_file: `section-${agentName}.md`
+  };
+}).filter(a => a.agent);  // 过滤未映射的角度
+
+console.log(`
+## Agent Auto-Assignment
+
+Found ${explorationFiles.length} exploration files:
+${agentAssignments.map(a => `- ${a.angle} → ${a.agent} agent`).join('\n')}
+`);
+```
+
+---
+
+## Agent 执行前置条件
+
+**每个 Agent 接收 exploration 文件路径，自行读取内容**：
+
+```javascript
+// Agent prompt 中包含文件路径
+// Agent 启动后的操作顺序：
+// 1. Read exploration 文件（上下文输入）
+// 2. Read 规范文件
+// 3. 执行分析任务
 ```

 规范文件路径（相对于 skill 根目录）：
@@ -617,15 +777,30 @@ Task({
 ## 执行流程

 ```javascript
-// 1. 根据报告类型选择 Agent 配置
-const agentConfigs = getAgentConfigs(config.type);
+// 1. 发现 exploration 文件并自动分配 agent
+const explorationFiles = bash(`find ${sessionFolder} -name "exploration-*.json" -type f`)
+  .split('\n')
+  .filter(f => f.trim());
+
+const agentAssignments = explorationFiles.map(file => {
+  const angle = extractAngle(path.basename(file));
+  const agentName = EXPLORATION_TO_AGENT[angle];
+  return { exploration_file: file, angle, agent: agentName };
+}).filter(a => a.agent);

 // 2. 准备目录
 Bash(`mkdir "${outputDir}\\sections"`);

-// 3. 并行启动所有 Agent
+// 3. 并行启动所有 Agent（传递 exploration 文件路径）
 const results = await Promise.all(
-  agentConfigs.map(agent => launchAgent(agent, config, outputDir))
+  agentAssignments.map(assignment =>
+    Task({
+      subagent_type: "cli-explore-agent",
+      run_in_background: false,
+      description: `Analyze: ${assignment.agent}`,
+      prompt: buildAgentPrompt(assignment, config, outputDir)
+    })
+  )
 );

 // 4. 收集简要返回信息
@@ -635,6 +810,45 @@ const summaries = results.map(r => JSON.parse(r));
 return { summaries, cross_notes: summaries.flatMap(s => s.cross_module_notes) };
 ```

+### Agent Prompt 构建
+
+```javascript
+function buildAgentPrompt(assignment, config, outputDir) {
+  const agentConfig = AGENT_CONFIGS[assignment.agent];
+  return `
+[CONTEXT]
+**Exploration 文件**: ${assignment.exploration_file}
+首先读取此文件获取 ${assignment.angle} 探索结果作为分析上下文。
+
+[SPEC]
+读取规范文件：
+- Read: ${skillRoot}/specs/quality-standards.md
+- Read: ${skillRoot}/specs/writing-style.md
+
+[ROLE] ${agentConfig.role}
+
+[TASK]
+${agentConfig.task}
+输出: ${outputDir}/sections/section-${assignment.agent}.md
+
+[STYLE]
+- 严谨专业的中文技术写作，专业术语保留英文
+- 完全客观的第三人称视角，严禁"我们"、"开发者"
+- 段落式叙述，采用"论点-论据-结论"结构
+- 善用逻辑连接词体现设计推演过程
+
+[FOCUS]
+${agentConfig.focus}
+
+[CONSTRAINT]
+${agentConfig.constraint}
+
+[RETURN JSON]
+{"status":"completed","output_file":"section-${assignment.agent}.md","summary":"<50字>","cross_module_notes":[],"stats":{}}
+`;
+}
+```
+
 ## Output

 各 Agent 写入 `sections/section-xxx.md`，返回简要 JSON 供 Phase 3.5 汇总。