# Phase 3: Task Generation

Generate implementation plan documents (IMPL_PLAN.md, task JSONs, TODO_LIST.md) using action-planning-agent - produces planning artifacts, does NOT execute code implementation.

## Auto Mode

When `--yes` or `-y`: Skip user questions, use defaults (no materials, Agent executor, Codex CLI tool).

## Core Philosophy

- **Planning Only**: Generate planning documents (IMPL_PLAN.md, task JSONs, TODO_LIST.md) - does NOT implement code
- **Agent-Driven Document Generation**: Delegate plan generation to action-planning-agent
- **NO Redundant Context Sorting**: Context priority sorting is ALREADY completed in context-gather Phase 2
  - Use `context-package.json.prioritized_context` directly
  - DO NOT re-sort files or re-compute priorities
  - `priority_tiers` and `dependency_order` are pre-computed and ready-to-use
- **N+1 Parallel Planning**: Auto-detect multi-module projects, enable parallel planning (2+1 or 3+1 mode)
- **Progressive Loading**: Load context incrementally (Core → Selective → On-Demand) due to analysis.md file size
- **Memory-First**: Reuse loaded documents from conversation memory
- **Smart Selection**: Load synthesis_output OR guidance + relevant role analyses, NOT all role analyses
- **MCP-Enhanced**: Use MCP tools for advanced code analysis and research
- **Path Clarity**: All `focus_paths` prefer absolute paths (e.g., `D:\\project\\src\\module`), or clear relative paths from project root (e.g., `./src/module`)
- **Explicit Lifecycle**: Manage subagent lifecycle with spawn_agent → wait → close_agent

## Execution Process

```
Input Parsing:
   ├─ Parse flags: --session
   └─ Validation: session_id REQUIRED

Phase 0: User Configuration (Interactive)
   ├─ Question 1: Supplementary materials/guidelines?
   ├─ Question 2: Execution method preference (Agent/CLI/Hybrid)
   ├─ Question 3: CLI tool preference (if CLI selected)
   └─ Store: userConfig for agent prompt

Phase 1: Context Preparation & Module Detection (Command)
   ├─ Assemble session paths (metadata, context package, output dirs)
   ├─ Provide metadata (session_id, execution_mode, mcp_capabilities)
   ├─ Auto-detect modules from context-package + directory structure
   └─ Decision:
      ├─ modules.length == 1 → Single Agent Mode (Phase 2A)
      └─ modules.length >= 2 → Parallel Mode (Phase 2B + Phase 3)

Phase 2A: Single Agent Planning (Original Flow)
   ├─ Spawn action-planning-agent
   ├─ Wait for completion
   ├─ Close agent
   ├─ Outputs: Task JSONs, IMPL_PLAN.md, TODO_LIST.md
   └─ Lifecycle: spawn_agent → wait → close_agent

Phase 2B: N Parallel Planning (Multi-Module)
   ├─ Spawn N action-planning-agents simultaneously (one per module)
   ├─ Wait for all agents (batch wait)
   ├─ Close all agents
   ├─ Each generates module-scoped tasks (IMPL-{prefix}{seq}.json)
   ├─ Task ID format: IMPL-A1, IMPL-A2... / IMPL-B1, IMPL-B2...
   └─ Each module limited to ≤9 tasks

Phase 3: Integration (+1 Coordinator, Multi-Module Only)
   ├─ Spawn coordinator agent
   ├─ Wait for completion
   ├─ Close agent
   ├─ Collect all module task JSONs
   ├─ Resolve cross-module dependencies (CROSS::{module}::{pattern} → actual ID)
   ├─ Generate unified IMPL_PLAN.md (grouped by module)
   └─ Generate TODO_LIST.md (hierarchical: module → tasks)
```

## Document Generation Lifecycle

### Phase 0: User Configuration (Interactive)

**Purpose**: Collect user preferences before task generation to ensure generated tasks match execution expectations.

**Auto Mode Check**:
```javascript
const autoYes = $ARGUMENTS.includes('--yes') || $ARGUMENTS.includes('-y')

// Check for prep-package auto-configuration (from /prompts:prep-plan)
const prepPath = `${projectRoot}/.workflow/.prep/plan-prep-package.json`
const prepExec = fs.existsSync(prepPath) ? JSON.parse(Read(prepPath))?.execution : null

if (autoYes || prepExec) {
  const source = prepExec ? 'prep-package' : '--yes flag'
  console.log(`[${source}] Using defaults: ${prepExec?.execution_method || 'agent'} executor, ${prepExec?.preferred_cli_tool || 'codex'} CLI`)
  userConfig = {
    supplementaryMaterials: prepExec?.supplementary_materials || { type: "none", content: [] },
    executionMethod: prepExec?.execution_method || "agent",
    preferredCliTool: prepExec?.preferred_cli_tool || "codex",
    enableResume: true
  }
  // Skip to Phase 1
}
```

**User Questions** (skipped if autoYes):
```javascript
if (!autoYes) ASK_USER([
  {
    id: "materials",
    type: "select",
    prompt: "Do you have supplementary materials or guidelines to include?",
    options: [
      { label: "No additional materials", description: "Use existing context only" },
      { label: "Provide file paths", description: "I'll specify paths to include" },
      { label: "Provide inline content", description: "I'll paste content directly" }
    ]
  },
  {
    id: "execution-method",
    type: "select",
    prompt: "Select execution method for generated tasks:",
    options: [
      { label: "Agent (Recommended)", description: "Claude agent executes tasks directly" },
      { label: "Hybrid", description: "Agent orchestrates, calls CLI for complex steps" },
      { label: "CLI Only", description: "All execution via CLI tools (codex/gemini/qwen)" }
    ]
  },
  {
    id: "cli-tool",
    type: "select",
    prompt: "If using CLI, which tool do you prefer?",
    options: [
      { label: "Codex (Recommended)", description: "Best for implementation tasks" },
      { label: "Gemini", description: "Best for analysis and large context" },
      { label: "Qwen", description: "Alternative analysis tool" },
      { label: "Auto", description: "Let agent decide per-task" }
    ]
  }
])  // BLOCKS (wait for user response)
```

**Handle Materials Response** (skipped if autoYes):
```javascript
if (!autoYes && userConfig.materials === "Provide file paths") {
  // Follow-up question for file paths
  const pathsResponse = ASK_USER([{
    id: "material-paths",
    type: "input",
    prompt: "Enter file paths to include (comma-separated or one per line):",
    options: [
      { label: "Enter paths", description: "Provide paths in text input" }
    ]
  }])  // BLOCKS (wait for user response)
  userConfig.supplementaryPaths = parseUserPaths(pathsResponse)
}
```

**Build userConfig**:
```javascript
const userConfig = {
  supplementaryMaterials: {
    type: "none|paths|inline",
    content: [...],  // Parsed paths or inline content
  },
  executionMethod: "agent|hybrid|cli",
  preferredCliTool: "codex|gemini|qwen|auto",
  enableResume: true  // Always enable resume for CLI executions
}
```

**Pass to Agent**: Include `userConfig` in agent prompt for Phase 2A/2B.

### Phase 1: Context Preparation & Module Detection (Command Responsibility)

**Command prepares session paths, metadata, detects module structure. Context priority sorting is NOT performed here - it's already completed in context-gather Phase 2.**

**Session Path Structure**:
```
{projectRoot}/.workflow/active/WFS-{session-id}/
├── workflow-session.json          # Session metadata
├── planning-notes.md              # Consolidated planning notes
├── .process/
│   └── context-package.json       # Context package with artifact catalog
├── .task/                         # Output: Task JSON files
│   ├── IMPL-A1.json               # Multi-module: prefixed by module
│   ├── IMPL-A2.json
│   ├── IMPL-B1.json
│   └── ...
├── IMPL_PLAN.md                   # Output: Implementation plan (grouped by module)
└── TODO_LIST.md                   # Output: TODO list (hierarchical)
```

**Command Preparation**:
1. **Assemble Session Paths** for agent prompt:
   - `session_metadata_path`
   - `context_package_path`
   - Output directory paths

2. **Provide Metadata** (simple values):
   - `session_id`
   - `mcp_capabilities` (available MCP tools)

3. **Auto Module Detection** (determines single vs parallel mode):
   ```javascript
   function autoDetectModules(contextPackage, projectRoot) {
     // === Complexity Gate: Only parallelize for High complexity ===
     const complexity = contextPackage.metadata?.complexity || 'Medium';
     if (complexity !== 'High') {
       // Force single agent mode for Low/Medium complexity
       // This maximizes agent context reuse for related tasks
       return [{ name: 'main', prefix: '', paths: ['.'] }];
     }

     // Priority 1: Explicit frontend/backend separation
     if (exists('src/frontend') && exists('src/backend')) {
       return [
         { name: 'frontend', prefix: 'A', paths: ['src/frontend'] },
         { name: 'backend', prefix: 'B', paths: ['src/backend'] }
       ];
     }

     // Priority 2: Monorepo structure
     if (exists('packages/*') || exists('apps/*')) {
       return detectMonorepoModules();  // Returns 2-3 main packages
     }

     // Priority 3: Context-package dependency clustering
     const modules = clusterByDependencies(contextPackage.dependencies?.internal);
     if (modules.length >= 2) return modules.slice(0, 3);

     // Default: Single module (original flow)
     return [{ name: 'main', prefix: '', paths: ['.'] }];
   }
   ```

**Decision Logic**:
- `complexity !== 'High'` → Force Phase 2A (Single Agent, maximize context reuse)
- `modules.length == 1` → Phase 2A (Single Agent, original flow)
- `modules.length >= 2 && complexity == 'High'` → Phase 2B + Phase 3 (N+1 Parallel)

**Note**: CLI tool usage is now determined semantically by action-planning-agent based on user's task description, not by flags.

### Phase 2A: Single Agent Planning (Original Flow)

**Condition**: `modules.length == 1` (no multi-module detected)

**Purpose**: Generate IMPL_PLAN.md, task JSONs, and TODO_LIST.md - planning documents only, NOT code implementation.

**Agent Invocation**:
```javascript
// Spawn action-planning-agent
const planningAgentId = spawn_agent({
  message: `
## TASK ASSIGNMENT

### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/action-planning-agent.md (MUST read first)
2. Read: ${projectRoot}/.workflow/project-tech.json
3. Read: ${projectRoot}/.workflow/project-guidelines.json

---

## TASK OBJECTIVE
Generate implementation planning documents (IMPL_PLAN.md, task JSONs, TODO_LIST.md) for workflow session

IMPORTANT: This is PLANNING ONLY - you are generating planning documents, NOT implementing code.

CRITICAL: Follow the progressive loading strategy defined in agent specification (load analysis.md files incrementally due to file size)

## PLANNING NOTES (PHASE 1-2 CONTEXT)
Load: ${projectRoot}/.workflow/active/${session_id}/planning-notes.md

This document contains:
- User Intent: Original GOAL and KEY_CONSTRAINTS from Phase 1
- Context Findings: Critical files, architecture, and constraints from Phase 2
- Conflict Decisions: Resolved conflicts and planning constraints from Phase 2
- Consolidated Constraints: All constraints from all phases

**USAGE**: Read planning-notes.md FIRST. Use Consolidated Constraints list to guide task sequencing and dependencies.

## SESSION PATHS
Input:
  - Session Metadata: ${projectRoot}/.workflow/active/${session_id}/workflow-session.json
  - Planning Notes: ${projectRoot}/.workflow/active/${session_id}/planning-notes.md
  - Context Package: ${projectRoot}/.workflow/active/${session_id}/.process/context-package.json

Output:
  - Task Dir: ${projectRoot}/.workflow/active/${session_id}/.task/
  - IMPL_PLAN: ${projectRoot}/.workflow/active/${session_id}/IMPL_PLAN.md
  - TODO_LIST: ${projectRoot}/.workflow/active/${session_id}/TODO_LIST.md

## CONTEXT METADATA
Session ID: ${session_id}
MCP Capabilities: {exa_code, exa_web, code_index}

## USER CONFIGURATION (from Phase 0)
Execution Method: ${userConfig.executionMethod}  // agent|hybrid|cli
Preferred CLI Tool: ${userConfig.preferredCliTool}  // codex|gemini|qwen|auto
Supplementary Materials: ${userConfig.supplementaryMaterials}

## EXECUTION METHOD MAPPING
Based on userConfig.executionMethod, set task-level meta.execution_config:

"agent" →
  meta.execution_config = { method: "agent", cli_tool: null, enable_resume: false }
  Agent executes implementation_approach steps directly

"cli" →
  meta.execution_config = { method: "cli", cli_tool: userConfig.preferredCliTool, enable_resume: true }
  Agent executes pre_analysis, then hands off full context to CLI via buildCliHandoffPrompt()

"hybrid" →
  Per-task decision: Analyze task complexity, set method to "agent" OR "cli" per task
  - Simple tasks (≤3 files, straightforward logic) → method: "agent"
  - Complex tasks (>3 files, complex logic, refactoring) → method: "cli"
  CLI tool: userConfig.preferredCliTool, enable_resume: true

IMPORTANT: Do NOT add command field to implementation_approach steps. Execution routing is controlled by task-level meta.execution_config.method only.

## PRIORITIZED CONTEXT (from context-package.prioritized_context) - ALREADY SORTED
Context sorting is ALREADY COMPLETED in context-gather Phase 2. DO NOT re-sort.
Direct usage:
- **user_intent**: Use goal/scope/key_constraints for task alignment
- **priority_tiers.critical**: These files are PRIMARY focus for task generation
- **priority_tiers.high**: These files are SECONDARY focus
- **dependency_order**: Use this for task sequencing - already computed
- **sorting_rationale**: Reference for understanding priority decisions

## EXPLORATION CONTEXT (from context-package.exploration_results) - SUPPLEMENT ONLY
If prioritized_context is incomplete, fall back to exploration_results:
- Load exploration_results from context-package.json
- Use aggregated_insights.critical_files for focus_paths generation
- Apply aggregated_insights.constraints to acceptance criteria
- Reference aggregated_insights.all_patterns for implementation approach
- Use aggregated_insights.all_integration_points for precise modification locations
- Use conflict_indicators for risk-aware task sequencing

## CONFLICT RESOLUTION CONTEXT (if exists)
- Check context-package.conflict_detection.resolution_file for conflict-resolution.json path
- If exists, load .process/conflict-resolution.json:
  - Apply planning_constraints as task constraints (for brainstorm-less workflows)
  - Reference resolved_conflicts for implementation approach alignment
  - Handle custom_conflicts with explicit task notes

## EXPECTED DELIVERABLES
1. Task JSON Files (.task/IMPL-*.json)
   - 6-field schema (id, title, status, context_package_path, meta, context, flow_control)
   - Quantified requirements with explicit counts
   - Artifacts integration from context package
   - **focus_paths generated directly from prioritized_context.priority_tiers (critical + high)**
     - NO re-sorting or re-prioritization - use pre-computed tiers as-is
     - Critical files are PRIMARY focus, High files are SECONDARY
   - Flow control with pre_analysis steps (use prioritized_context.dependency_order for task sequencing)
   - **CLI Execution IDs and strategies (MANDATORY)**

2. Implementation Plan (IMPL_PLAN.md)
   - Context analysis and artifact references
   - Task breakdown and execution strategy
   - Complete structure per agent definition

3. TODO List (TODO_LIST.md)
   - Hierarchical structure (containers, pending, completed markers)
   - Links to task JSONs and summaries
   - Matches task JSON hierarchy

## CLI EXECUTION ID REQUIREMENTS (MANDATORY)
Each task JSON MUST include:
- **cli_execution_id**: Unique ID for CLI execution (format: \`{session_id}-{task_id}\`)
- **cli_execution**: Strategy object based on depends_on:
  - No deps → \`{ "strategy": "new" }\`
  - 1 dep (single child) → \`{ "strategy": "resume", "resume_from": "parent-cli-id" }\`
  - 1 dep (multiple children) → \`{ "strategy": "fork", "resume_from": "parent-cli-id" }\`
  - N deps → \`{ "strategy": "merge_fork", "merge_from": ["id1", "id2", ...] }\`

**CLI Execution Strategy Rules**:
1. **new**: Task has no dependencies - starts fresh CLI conversation
2. **resume**: Task has 1 parent AND that parent has only this child - continues same conversation
3. **fork**: Task has 1 parent BUT parent has multiple children - creates new branch with parent context
4. **merge_fork**: Task has multiple parents - merges all parent contexts into new conversation

**Execution Command Patterns**:
- new: \`ccw cli -p "[prompt]" --tool [tool] --mode write --id [cli_execution_id]\`
- resume: \`ccw cli -p "[prompt]" --resume [resume_from] --tool [tool] --mode write\`
- fork: \`ccw cli -p "[prompt]" --resume [resume_from] --id [cli_execution_id] --tool [tool] --mode write\`
- merge_fork: \`ccw cli -p "[prompt]" --resume [merge_from.join(',')] --id [cli_execution_id] --tool [tool] --mode write\`

## QUALITY STANDARDS
Hard Constraints:
  - Task count <= 18 (hard limit - request re-scope if exceeded)
  - All requirements quantified (explicit counts and enumerated lists)
  - Acceptance criteria measurable (include verification commands)
  - Artifact references mapped from context package
  - All documents follow agent-defined structure

## SUCCESS CRITERIA
- All planning documents generated successfully:
  - Task JSONs valid and saved to .task/ directory
  - IMPL_PLAN.md created with complete structure
  - TODO_LIST.md generated matching task JSONs
- Return completion status with document count and task breakdown summary

## PLANNING NOTES RECORD (REQUIRED)
After completing, update planning-notes.md:

**File**: ${projectRoot}/.workflow/active/${session_id}/planning-notes.md

1. **Task Generation (Phase 3)**: Task count and key tasks
2. **N+1 Context**: Key decisions (with rationale) + deferred items

\`\`\`markdown
## Task Generation (Phase 3)
### [Action-Planning Agent] YYYY-MM-DD
- **Tasks**: [count] ([IDs])

## N+1 Context
### Decisions
| Decision | Rationale | Revisit? |
|----------|-----------|----------|
| [choice] | [why] | [Yes/No] |

### Deferred
- [ ] [item] - [reason]
\`\`\`
`
});

// Wait for planning agent to complete
const planningResult = wait({
  ids: [planningAgentId],
  timeout_ms: 900000  // 15 minutes
});

// Close planning agent
close_agent({ id: planningAgentId });
```

### Phase 2B: N Parallel Planning (Multi-Module)

**Condition**: `modules.length >= 2` (multi-module detected)

**Purpose**: Launch N action-planning-agents simultaneously, one per module, for parallel task JSON generation.

**Note**: Phase 2B agents generate Task JSONs ONLY. IMPL_PLAN.md and TODO_LIST.md are generated by Phase 3 Coordinator.

**Parallel Agent Invocation**:
```javascript
// Spawn N agents in parallel (one per module)
const moduleAgents = [];

modules.forEach(module => {
  const agentId = spawn_agent({
    message: `
## TASK ASSIGNMENT

### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/action-planning-agent.md (MUST read first)
2. Read: ${projectRoot}/.workflow/project-tech.json
3. Read: ${projectRoot}/.workflow/project-guidelines.json

---

## TASK OBJECTIVE
Generate task JSON files for ${module.name} module within workflow session

IMPORTANT: This is PLANNING ONLY - generate task JSONs, NOT implementing code.
IMPORTANT: Generate Task JSONs ONLY. IMPL_PLAN.md and TODO_LIST.md by Phase 3 Coordinator.

CRITICAL: Follow the progressive loading strategy defined in agent specification (load analysis.md files incrementally due to file size)

## PLANNING NOTES (PHASE 1-2 CONTEXT)
Load: ${projectRoot}/.workflow/active/${session_id}/planning-notes.md

This document contains consolidated constraints and user intent to guide module-scoped task generation.

## MODULE SCOPE
- Module: ${module.name} (${module.type})
- Focus Paths: ${module.paths.join(', ')}
- Task ID Prefix: IMPL-${module.prefix}
- Task Limit: ≤6 tasks (hard limit for this module)
- Other Modules: ${otherModules.join(', ')} (reference only, do NOT generate tasks for them)

## SESSION PATHS
Input:
  - Session Metadata: ${projectRoot}/.workflow/active/${session_id}/workflow-session.json
  - Planning Notes: ${projectRoot}/.workflow/active/${session_id}/planning-notes.md
  - Context Package: ${projectRoot}/.workflow/active/${session_id}/.process/context-package.json

Output:
  - Task Dir: ${projectRoot}/.workflow/active/${session_id}/.task/

## CONTEXT METADATA
Session ID: ${session_id}
MCP Capabilities: {exa_code, exa_web, code_index}

## USER CONFIGURATION (from Phase 0)
Execution Method: ${userConfig.executionMethod}  // agent|hybrid|cli
Preferred CLI Tool: ${userConfig.preferredCliTool}  // codex|gemini|qwen|auto
Supplementary Materials: ${userConfig.supplementaryMaterials}

## EXECUTION METHOD MAPPING
Based on userConfig.executionMethod, set task-level meta.execution_config:

"agent" →
  meta.execution_config = { method: "agent", cli_tool: null, enable_resume: false }
  Agent executes implementation_approach steps directly

"cli" →
  meta.execution_config = { method: "cli", cli_tool: userConfig.preferredCliTool, enable_resume: true }
  Agent executes pre_analysis, then hands off full context to CLI via buildCliHandoffPrompt()

"hybrid" →
  Per-task decision: Analyze task complexity, set method to "agent" OR "cli" per task
  - Simple tasks (≤3 files, straightforward logic) → method: "agent"
  - Complex tasks (>3 files, complex logic, refactoring) → method: "cli"
  CLI tool: userConfig.preferredCliTool, enable_resume: true

IMPORTANT: Do NOT add command field to implementation_approach steps. Execution routing is controlled by task-level meta.execution_config.method only.

## PRIORITIZED CONTEXT (from context-package.prioritized_context) - ALREADY SORTED
Context sorting is ALREADY COMPLETED in context-gather Phase 2. DO NOT re-sort.
Filter by module scope (${module.paths.join(', ')}):
- **user_intent**: Use for task alignment within module
- **priority_tiers.critical**: Filter for files in ${module.paths.join(', ')} → PRIMARY focus
- **priority_tiers.high**: Filter for files in ${module.paths.join(', ')} → SECONDARY focus
- **dependency_order**: Use module-relevant entries for task sequencing

## EXPLORATION CONTEXT (from context-package.exploration_results) - SUPPLEMENT ONLY
If prioritized_context is incomplete for this module, fall back to exploration_results:
- Load exploration_results from context-package.json
- Filter for ${module.name} module: Use aggregated_insights.critical_files matching ${module.paths.join(', ')}
- Apply module-relevant constraints from aggregated_insights.constraints
- Reference aggregated_insights.all_patterns applicable to ${module.name}
- Use aggregated_insights.all_integration_points for precise modification locations within module scope
- Use conflict_indicators for risk-aware task sequencing

## CONFLICT RESOLUTION CONTEXT (if exists)
- Check context-package.conflict_detection.resolution_file for conflict-resolution.json path
- If exists, load .process/conflict-resolution.json:
  - Apply planning_constraints relevant to ${module.name} as task constraints
  - Reference resolved_conflicts affecting ${module.name} for implementation approach alignment
  - Handle custom_conflicts with explicit task notes

## CROSS-MODULE DEPENDENCIES
- For dependencies ON other modules: Use placeholder depends_on: ["CROSS::{module}::{pattern}"]
- Example: depends_on: ["CROSS::B::api-endpoint"] (this module depends on B's api-endpoint task)
- Phase 3 Coordinator resolves to actual task IDs
- For dependencies FROM other modules: Document in task context as "provides_for" annotation

## EXPECTED DELIVERABLES
Task JSON Files (.task/IMPL-${module.prefix}*.json):
  - 6-field schema (id, title, status, context_package_path, meta, context, flow_control)
  - Task ID format: IMPL-${module.prefix}1, IMPL-${module.prefix}2, ...
  - Quantified requirements with explicit counts
  - Artifacts integration from context package (filtered for ${module.name})
  - **focus_paths generated directly from prioritized_context.priority_tiers filtered by ${module.paths.join(', ')}**
    - NO re-sorting - use pre-computed tiers filtered for this module
    - Critical files are PRIMARY focus, High files are SECONDARY
  - Flow control with pre_analysis steps (use prioritized_context.dependency_order for module task sequencing)
  - **CLI Execution IDs and strategies (MANDATORY)**
  - Focus ONLY on ${module.name} module scope

## CLI EXECUTION ID REQUIREMENTS (MANDATORY)
Each task JSON MUST include:
- **cli_execution_id**: Unique ID for CLI execution (format: \`{session_id}-IMPL-${module.prefix}{seq}\`)
- **cli_execution**: Strategy object based on depends_on:
  - No deps → \`{ "strategy": "new" }\`
  - 1 dep (single child) → \`{ "strategy": "resume", "resume_from": "parent-cli-id" }\`
  - 1 dep (multiple children) → \`{ "strategy": "fork", "resume_from": "parent-cli-id" }\`
  - N deps → \`{ "strategy": "merge_fork", "merge_from": ["id1", "id2", ...] }\`
  - Cross-module dep → \`{ "strategy": "cross_module_fork", "resume_from": "CROSS::{module}::{pattern}" }\`

**CLI Execution Strategy Rules**:
1. **new**: Task has no dependencies - starts fresh CLI conversation
2. **resume**: Task has 1 parent AND that parent has only this child - continues same conversation
3. **fork**: Task has 1 parent BUT parent has multiple children - creates new branch with parent context
4. **merge_fork**: Task has multiple parents - merges all parent contexts into new conversation
5. **cross_module_fork**: Task depends on task from another module - Phase 3 resolves placeholder

**Execution Command Patterns**:
- new: \`ccw cli -p "[prompt]" --tool [tool] --mode write --id [cli_execution_id]\`
- resume: \`ccw cli -p "[prompt]" --resume [resume_from] --tool [tool] --mode write\`
- fork: \`ccw cli -p "[prompt]" --resume [resume_from] --id [cli_execution_id] --tool [tool] --mode write\`
- merge_fork: \`ccw cli -p "[prompt]" --resume [merge_from.join(',')] --id [cli_execution_id] --tool [tool] --mode write\`
- cross_module_fork: (Phase 3 resolves placeholder, then uses fork pattern)

## QUALITY STANDARDS
Hard Constraints:
  - Task count <= 9 for this module (hard limit - coordinate with Phase 3 if exceeded)
  - All requirements quantified (explicit counts and enumerated lists)
  - Acceptance criteria measurable (include verification commands)
  - Artifact references mapped from context package (module-scoped filter)
  - Focus paths use absolute paths or clear relative paths from project root
  - Cross-module dependencies use CROSS:: placeholder format

## SUCCESS CRITERIA
- Task JSONs saved to .task/ with IMPL-${module.prefix}* naming
- All task JSONs include cli_execution_id and cli_execution strategy
- Cross-module dependencies use CROSS:: placeholder format consistently
- Focus paths scoped to ${module.paths.join(', ')} only
- Return: task count, task IDs, dependency summary (internal + cross-module)

## PLANNING NOTES RECORD (REQUIRED)
After completing, append to planning-notes.md:

\`\`\`markdown
### [${module.name}] YYYY-MM-DD
- **Tasks**: [count] ([IDs])
- **CROSS deps**: [placeholders used]
\`\`\`
    `
  });
  
  moduleAgents.push(agentId);
});

// Batch wait for all module agents
const moduleResults = wait({
  ids: moduleAgents,
  timeout_ms: 900000  // 15 minutes
});

// Close all module agents
moduleAgents.forEach(agentId => {
  close_agent({ id: agentId });
});
```

**Output Structure** (direct to .task/):
```
.task/
├── IMPL-A1.json      # Module A (e.g., frontend)
├── IMPL-A2.json
├── IMPL-B1.json      # Module B (e.g., backend)
├── IMPL-B2.json
└── IMPL-C1.json      # Module C (e.g., shared)
```

**Task ID Naming**:
- Format: `IMPL-{prefix}{seq}.json`
- Prefix: A, B, C... (assigned by detection order)
- Sequence: 1, 2, 3... (per-module increment)

### Phase 3: Integration (+1 Coordinator Agent, Multi-Module Only)

**Condition**: Only executed when `modules.length >= 2`

**Purpose**: Collect all module tasks, resolve cross-module dependencies, generate unified IMPL_PLAN.md and TODO_LIST.md documents.

**Coordinator Agent Invocation**:
```javascript
// Wait for all Phase 2B agents to complete (already done above)

// Spawn +1 Coordinator Agent
const coordinatorAgentId = spawn_agent({
  message: `
## TASK ASSIGNMENT

### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/action-planning-agent.md (MUST read first)
2. Read: ${projectRoot}/.workflow/project-tech.json
3. Read: ${projectRoot}/.workflow/project-guidelines.json

---

## TASK OBJECTIVE
Integrate all module task JSONs, resolve cross-module dependencies, and generate unified IMPL_PLAN.md and TODO_LIST.md

IMPORTANT: This is INTEGRATION ONLY - consolidate existing task JSONs, NOT creating new tasks.

## SESSION PATHS
Input:
  - Session Metadata: ${projectRoot}/.workflow/active/${session_id}/workflow-session.json
  - Context Package: ${projectRoot}/.workflow/active/${session_id}/.process/context-package.json
  - Task JSONs: ${projectRoot}/.workflow/active/${session_id}/.task/IMPL-*.json (from Phase 2B)
Output:
  - Updated Task JSONs: ${projectRoot}/.workflow/active/${session_id}/.task/IMPL-*.json (resolved dependencies)
  - IMPL_PLAN: ${projectRoot}/.workflow/active/${session_id}/IMPL_PLAN.md
  - TODO_LIST: ${projectRoot}/.workflow/active/${session_id}/TODO_LIST.md

## CONTEXT METADATA
Session ID: ${session_id}
Modules: ${modules.map(m => m.name + '(' + m.prefix + ')').join(', ')}
Module Count: ${modules.length}

## INTEGRATION STEPS
1. Collect all .task/IMPL-*.json, group by module prefix
2. Resolve CROSS:: dependencies → actual task IDs, update task JSONs
3. Generate IMPL_PLAN.md (multi-module format per agent specification)
4. Generate TODO_LIST.md (hierarchical format per agent specification)

## CROSS-MODULE DEPENDENCY RESOLUTION
- Pattern: CROSS::{module}::{pattern} → IMPL-{module}* matching title/context
- Example: CROSS::B::api-endpoint → IMPL-B1 (if B1 title contains "api-endpoint")
- Log unresolved as warnings

## EXPECTED DELIVERABLES
1. Updated Task JSONs with resolved dependency IDs
2. IMPL_PLAN.md - multi-module format with cross-dependency section
3. TODO_LIST.md - hierarchical by module with cross-dependency section

## SUCCESS CRITERIA
- No CROSS:: placeholders remaining in task JSONs
- IMPL_PLAN.md and TODO_LIST.md generated with multi-module structure
- Return: task count, per-module breakdown, resolved dependency count

## PLANNING NOTES RECORD (REQUIRED)
After integration, update planning-notes.md:

\`\`\`markdown
### [Coordinator] YYYY-MM-DD
- **Total**: [count] tasks
- **Resolved**: [CROSS:: resolutions]

## N+1 Context
### Decisions
| Decision | Rationale | Revisit? |
|----------|-----------|----------|
| CROSS::X → IMPL-Y | [why this resolution] | [Yes/No] |

### Deferred
- [ ] [unresolved CROSS or conflict] - [reason]
\`\`\`
  `
});

// Wait for coordinator agent to complete
const coordinatorResult = wait({
  ids: [coordinatorAgentId],
  timeout_ms: 600000  // 10 minutes
});

// Close coordinator agent
close_agent({ id: coordinatorAgentId });
```

**Dependency Resolution Algorithm**:
```javascript
function resolveCrossModuleDependency(placeholder, allTasks) {
  const [, targetModule, pattern] = placeholder.match(/CROSS::(\w+)::(.+)/);
  const candidates = allTasks.filter(t =>
    t.id.startsWith(`IMPL-${targetModule}`) &&
    (t.title.toLowerCase().includes(pattern.toLowerCase()) ||
     t.context?.description?.toLowerCase().includes(pattern.toLowerCase()))
  );
  return candidates.length > 0
    ? candidates.sort((a, b) => a.id.localeCompare(b.id))[0].id
    : placeholder; // Keep for manual resolution
}
```

## Output

- **Files**:
  - `{projectRoot}/.workflow/active/{sessionId}/IMPL_PLAN.md`
  - `{projectRoot}/.workflow/active/{sessionId}/.task/IMPL-*.json`
  - `{projectRoot}/.workflow/active/{sessionId}/TODO_LIST.md`
- **Updated**: `planning-notes.md` with task generation record and N+1 context

## Next Step

Return to orchestrator. Present user with action choices (or auto-continue if `--yes`):
1. Verify Plan Quality (Recommended) → `workflow:plan-verify`
2. Start Execution → Phase 4 (phases/04-execution.md)
3. Review Status Only → `workflow:status`