subagent-driven-development（子 Agent 驱动开发）

Subagent-Driven Development

Core Principle

Fresh context per task. Review between tasks.

Each task gets a clean subagent with no accumulated confusion. You review between tasks.

How It Works

1. Load the implementation plan
2. For each task:

- Dispatch fresh subagent - Subagent implements ONLY that task - Review the changes - Approve or request fixes - Move to next task

Benefits

• Clean Context: Each subagent starts fresh
• Focused Work: One task at a time
• Review Points: Catch issues early
• Controlled Progress: You stay in charge

Execution Flow

Step 1: Load the Plan

Loading plan from: docs/plans/YYYY-MM-DD-feature-name.md

Tasks identified:
1. [ ] Task 1: Description
2. [ ] Task 2: Description
3. [ ] Task 3: Description

Starting with Task 1...

Step 2: Dispatch Subagent

For each task, create a focused prompt:

## Task: [Task Name]

### Context
- Project: [brief description]
- Current branch: [branch name]
- Dependencies: [relevant info]

### Instructions
[Exact instructions from plan]

### Files to Modify
- `path/to/file.ts`

### Test to Write First
[Test code from plan]

### Implementation
[Implementation code from plan]

### Success Criteria
- [ ] Test passes
- [ ] No other tests broken
- [ ] Code follows project style

Step 3: Review Changes

After subagent completes:

## Task 1 Complete

### Changes Made:
- Modified: `path/to/file.ts` (+25/-3)
- Added: `path/to/file.test.ts` (+40)

### Test Results:
✅ All tests passing (47 total)

### Review Checklist:
- [ ] Test covers the requirement
- [ ] Implementation is correct
- [ ] No unnecessary changes
- [ ] Code style matches project

**Approve and continue to Task 2?**

Step 4: Handle Issues

If review finds problems:

## Issues Found in Task 1

1. Test doesn't cover edge case X
2. Missing error handling for Y

**Options:**
A) Request fixes from subagent
B) Fix manually
C) Skip and note for later

Which approach?

Subagent Guidelines

What Subagents Should Do

• Follow the plan exactly
• Write tests first
• Make minimal changes
• Report what was done

What Subagents Should NOT Do

• Make "improvements" outside scope
• Skip tests
• Refactor unrelated code
• Change the plan

Progress Tracking

Maintain task status:

## Progress: Feature Name

- [x] Task 1: Setup database schema ✅
- [x] Task 2: Create API endpoint ✅
- [ ] Task 3: Add validation (IN PROGRESS)
- [ ] Task 4: Write integration tests
- [ ] Task 5: Update documentation

Current: Task 3 of 5

Checkpoints

After Each Task

• Run all tests
• Review diff
• Commit if approved

After All Tasks

• Run integration tests
• Manual verification
• Final review

Rollback

If things go wrong:

# Revert last task
git revert HEAD

# Or reset to checkpoint
git reset --hard <commit-before-task>

Communication Pattern

Starting

"I'm using subagent-driven-development to implement [feature]. I'll dispatch a fresh subagent for each task and review between them."

Between Tasks

"Task [N] complete. Changes: [summary]. Ready to review before Task [N+1]?"

Completing

"All [N] tasks complete. Running final verification..."