agent-orchestration

提供实现代理与主代理协同且保护主上下文的编排规则与模式。

在需要实现功能或修改大量内容时，给出如何编排主代理与实现代理的规则，避免把所有文件读取和编辑都放在主会话里把上下文耗光。用在要拆分实现步骤、派生 agent 真正去读写文件并返回结果的场景。特色是强调用实现型 agent 完成繁重工作，只把必要输出回主控，从而保护主上下文和对话历史。

# Agent Orchestration Rules

When the user asks to implement something, use implementation agents to preserve main context.

## The Pattern

**Wrong - burns context:**
```
Main: Read files → Understand → Make edits → Report
      (2000+ tokens consumed in main context)
```

**Right - preserves context:**
```
Main: Spawn agent("implement X per plan")
      ↓
Agent: Reads files → Understands → Edits → Tests
      ↓
Main: Gets summary (~200 tokens)
```

## When to Use Agents

| Task Type | Use Agent? | Reason |
|-----------|------------|--------|
| Multi-file implementation | Yes | Agent handles complexity internally |
| Following a plan phase | Yes | Agent reads plan, implements |
| New featur