bookworm-smart-assistant/docs/agent-orchestration.md

Agent 编排规范 (v4.3)

调用拓扑 (DAG)

                    用户请求
                       │
                  ┌────▼────┐
                  │orchestrator│ (opus) ─ 复杂多技能任务入口
                  └─┬──┬──┬─┘
          ┌────────┘  │  └────────┐
          ▼           ▼           ▼
   ┌─────────┐  ┌──────────┐  ┌─────────────┐
   │full-stack│  │research- │  │canvas-ui-   │
   │-builder  │  │analyst   │  │designer     │
   └─────────┘  └──────────┘  └─────────────┘
   (sonnet)     (sonnet)      (sonnet)

  ── 质量链 (串行) ──────────────────────────
  test-writer → code-reviewer → quality-gate → pre-deploy-checker
  (sonnet)      (opus)          (sonnet)       (sonnet)

  ── 自进化链 (串行) ────────────────────────
  drift-detector(hook) → self-auditor → self-healer → evolution-log
                         (sonnet)       (sonnet)

触发条件

Agent	触发方式	触发词/条件
orchestrator	自动	"从零开发"、"全面优化"、需 3+ 技能协作
full-stack-builder	orchestrator 调度	全栈开发子任务
research-analyst	orchestrator/用户	"调研"、"分析竞品"、"技术选型"
canvas-ui-designer	用户	"画界面"、"UI 原型"、"线框图"
test-writer	用户/quality-gate	"写测试"、代码变更后建议
code-reviewer	用户/质量链	"代码审查"、PR 提交前
quality-gate	orchestrator	交付前综合质量检查
pre-deploy-checker	orchestrator/用户	"部署检查"、上线前
self-auditor	drift-detector/用户	"自检"、"审计"、配置漂移检测
self-healer	self-auditor	审计发现可自动修复的问题

并行/串行规则

• 可并行: orchestrator 调度的子 Agent (full-stack + research + canvas)
• 必须串行: 质量链 (test → review → gate → deploy)
• 必须串行: 自进化链 (detect → audit → heal)
• 隔离执行: self-healer 只修改元数据层，不触碰业务逻辑

模型分配 (2 opus + 8 sonnet)

• opus (深度推理): orchestrator (复杂任务分解)、code-reviewer (安全审查)
• sonnet (效率优先): 其余 8 个 Agent