Back to Details

delegate

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

/delegate

/delegate

You orchestrate. Specialists do the work.
Reference pattern for invoking multiple AI tools and synthesizing their outputs.
你负责编排，专家负责执行任务。
这是调用多个AI工具并合成其输出的参考模式。

Your Role

你的角色

You don't analyze/review/audit yourself. You:
  1. Route — Send work to appropriate specialists
  2. Collect — Gather their outputs
  3. Curate — Validate, filter, resolve conflicts
  4. Synthesize — Produce unified output
你无需亲自进行分析/审查/审计工作，你需要：
  1. 路由 — 将任务分配给合适的专家
  2. 收集 — 收集他们的输出结果
  3. 整理 — 验证、过滤、解决冲突
  4. 合成 — 生成统一的输出结果

Your Team

你的团队

Moonbridge MCP — Unified Agent Interface

Moonbridge MCP — 统一Agent接口

One interface, multiple backends. Moonbridge wraps both Codex and Kimi:
AdapterStrengthsWhen to Use
codex
Long-context, tool calling, securityRefactors, migrations, debugging, backend
kimi
Native vision, extended thinkingUI from designs, visual debugging, frontend
Single agent:
mcp__moonbridge__spawn_agent({
  "prompt": "...",
  "adapter": "codex",           // or "kimi"
  "reasoning_effort": "high"    // codex: low/medium/high/xhigh
  // OR
  "thinking": true              // kimi: extended reasoning
})
Timeout: Moonbridge defaults: Codex=30min, Kimi=10min. Override for edge cases:
TaskTimeout
Quick check
60
Large refactor
3600
mcp__moonbridge__spawn_agent({ ..., "timeout_seconds": 3600 })
Parallel agents (same or mixed adapters):
mcp__moonbridge__spawn_agents_parallel({
  "agents": [
    {"prompt": "Backend API", "adapter": "codex", "reasoning_effort": "high"},
    {"prompt": "Frontend UI", "adapter": "kimi", "thinking": true},
    {"prompt": "Tests", "adapter": "codex", "reasoning_effort": "medium"}
  ]
})
一个接口，多个后端。Moonbridge同时封装了Codex和Kimi：
适配器优势使用场景
codex
长上下文、工具调用、安全性重构、迁移、调试、后端开发
kimi
原生视觉能力、深度思考根据设计稿生成UI、视觉调试、前端开发
单个Agent调用：
mcp__moonbridge__spawn_agent({
  "prompt": "...",
  "adapter": "codex",           // 或 "kimi"
  "reasoning_effort": "high"    // codex可选：low/medium/high/xhigh
  // 或者
  "thinking": true              // kimi启用深度思考
})
超时设置： Moonbridge默认超时：Codex=30分钟，Kimi=10分钟。可针对特殊情况覆盖默认值：
任务类型超时时间
快速检查
60
大型重构
3600
mcp__moonbridge__spawn_agent({ ..., "timeout_seconds": 3600 })
并行调用多个Agent（相同或混合适配器）：
mcp__moonbridge__spawn_agents_parallel({
  "agents": [
    {"prompt": "Backend API", "adapter": "codex", "reasoning_effort": "high"},
    {"prompt": "Frontend UI", "adapter": "kimi", "thinking": true},
    {"prompt": "Tests", "adapter": "codex", "reasoning_effort": "medium"}
  ]
})

Gemini CLI — Researcher, deep reasoner

Gemini CLI — 研究员、深度推理工具

  • Web grounding, thinking_level control, agentic vision
  • Best at: current best practices, pattern validation, design research
  • Invocation: 
    gemini "..."
    (bash)
  • 网络 grounding、思考级别控制、Agent视觉能力
  • 最适合：当前最佳实践、模式验证、设计研究
  • 调用方式：
    gemini "..."
    （bash命令）

Non-Agentic (Opinions Only)

非Agent类工具（仅提供观点）

Thinktank CLI — Expert council
  • Multiple models respond in parallel, synthesis mode
  • Best at: consensus, architecture validation, second opinions
  • Invocation: 
    thinktank instructions.md ./files --synthesis
    (bash)
  • Note: Cannot take action. Use for validation, not investigation.
Thinktank CLI — 专家委员会
  • 多模型并行响应，支持合成模式
  • 最适合：达成共识、架构验证、二次意见
  • 调用方式：
    thinktank instructions.md ./files --synthesis
    （bash命令）
  • 注意：无法执行操作。仅用于验证，不用于调查。

Agent Teams — Full Claude Code Teammates

Agent团队 — 完整的Claude Code协作团队

When workers need to communicate, challenge each other, or coordinate across layers.
Start a team: Describe the task and team structure in natural language. Claude handles spawning.
Lead in delegate mode: Shift+Tab after team creation. Lead coordinates only.
Plan approval: For risky work, require teammates to plan before implementing. Lead reviews and approves/rejects plans.
When to use over Moonbridge:
SignalTeamsMoonbridge
Workers must discuss findingsYESno
Competing hypotheses / debateYESno
Cross-layer (FE+BE+tests)YESno
"Implement this spec"noYES
Result-only, no coordinationnoYES
当协作成员需要沟通、相互质疑或跨层协调时使用。
创建团队： 用自然语言描述任务和团队结构，Claude会负责创建Agent。
以委托模式担任负责人： 创建团队后按Shift+Tab切换。负责人仅负责协调工作。
计划审批： 对于高风险任务，要求团队成员在执行前提交计划。负责人负责审核并批准/拒绝计划。
何时选择团队而非Moonbridge：
信号团队模式Moonbridge
协作成员需要讨论发现结果是否
需要提出竞争假设/进行辩论是否
跨层协作（前端+后端+测试）是否
"实现这个规格"否是
仅需结果，无需协调否是

Internal Agents (Task tool)

内部Agent（任务工具）

Domain specialists for focused review:
  • go-concurrency-reviewer
    , 
    react-pitfalls
    , 
    security-sentinel
  • data-integrity-guardian
    , 
    architecture-guardian
    , 
    config-auditor
专注于特定领域审查的专家：
  • go-concurrency-reviewer
    , 
    react-pitfalls
    , 
    security-sentinel
  • data-integrity-guardian
    , 
    architecture-guardian
    , 
    config-auditor

How to Delegate

如何委托任务

Apply 
/llm-communication
principles — state goals, not steps:
遵循
/llm-communication
原则 — 明确目标，而非步骤：

To Moonbridge Agents (Codex, Kimi)

给Moonbridge Agent（Codex、Kimi）

Give them latitude to investigate:
"Investigate this stack trace. Find root cause. Propose fix with file:line."
NOT:
"Step 1: Read file X. Step 2: Check line Y. Step 3: ..."
给予他们自主调查的权限：
"Investigate this stack trace. Find root cause. Propose fix with file:line."
而非：
"Step 1: Read file X. Step 2: Check line Y. Step 3: ..."

To Thinktank (Non-Agentic)

给Thinktank（非Agent类）

Provide context, ask for judgment:
"Here's the code and proposed fix. Is this approach sound?
What are we missing? Consensus and dissent."
提供上下文，请求判断：
"Here's the code and proposed fix. Is this approach sound?
What are we missing? Consensus and dissent."

Parallel Execution

并行执行

Run independent reviews in parallel:
  • Multiple moonbridge agents in same call (
    spawn_agents_parallel
    )
  • Multiple Task tool calls in same message
  • Gemini + Thinktank can run concurrently (both bash)
并行运行独立的审查任务：
  • 单次调用中启动多个Moonbridge Agent（
    spawn_agents_parallel
    ）
  • 单次消息中调用多个任务工具
  • Gemini和Thinktank可同时运行（均为bash命令）

Dependency-Aware Orchestration

依赖感知型编排

For large work (10+ subtasks, multiple phases), use DAG-based scheduling:
对于大型任务（10+子任务、多阶段），使用基于DAG的调度：

The Pattern

模式

Phase 1 (no deps):    Task 01, 02, 03 → run in parallel
Phase 2 (deps on P1): Task 04, 05     → blocked until P1 complete
Phase 3 (deps on P2): Task 06, 07, 08 → blocked until P2 complete
Key principles:
  1. Task decomposition — Break feature into atomic subtasks
  2. Dependency graph — DAG defines execution order
  3. Parallel execution — Independent tasks run simultaneously
  4. Fresh context — Each subagent starts clean (~40-75k tokens)
Phase 1 (无依赖):    任务01、02、03 → 并行运行
Phase 2 (依赖Phase1): 任务04、05     → 需等待Phase1完成后执行
Phase 3 (依赖Phase2): 任务06、07、08 → 需等待Phase2完成后执行
核心原则：
  1. 任务分解 — 将功能拆分为原子子任务
  2. 依赖图 — DAG定义执行顺序
  3. 并行执行 — 独立任务同时运行
  4. 全新上下文 — 每个子Agent启动时使用全新上下文（约40-75k tokens）

Step 1: Decompose

步骤1：分解任务

Split feature into atomic tasks. Ask:
  • What can run independently? → Same phase
  • What requires prior output? → Blocked
将功能拆分为原子任务，思考：
  • 哪些任务可以独立运行？→ 同一阶段
  • 哪些任务需要依赖前置输出？→ 阻塞状态

Step 2: Declare Dependencies

步骤2：声明依赖关系

Use TaskCreate/TaskUpdate primitives:
TaskCreate({subject: "Install packages", activeForm: "Installing packages"})
TaskCreate({subject: "cRPC builder", activeForm: "Building cRPC"})
TaskUpdate({taskId: "2", addBlockedBy: ["1"]})  # Task 2 waits for Task 1
使用TaskCreate/TaskUpdate原语：
TaskCreate({subject: "Install packages", activeForm: "Installing packages"})
TaskCreate({subject: "cRPC builder", activeForm: "Building cRPC"})
TaskUpdate({taskId: "2", addBlockedBy: ["1"]})  # 任务2需等待任务1完成

Step 3: Execute Phases

步骤3：执行阶段

Spawn all unblocked tasks in single message:
undefined
在单条消息中启动所有未阻塞的任务：
undefined

Phase 1 - all parallel via moonbridge

Phase 1 - 通过moonbridge并行启动所有任务

mcp__moonbridge__spawn_agents_parallel({ agents: [ {prompt: "Task 1: ...", adapter: "codex"}, {prompt: "Task 2: ...", adapter: "codex"}, {prompt: "Task 3: ...", adapter: "kimi", thinking: true} ] })
undefined
mcp__moonbridge__spawn_agents_parallel({ agents: [ {prompt: "Task 1: ...", adapter: "codex"}, {prompt: "Task 2: ...", adapter: "codex"}, {prompt: "Task 3: ...", adapter: "kimi", thinking: true} ] })
undefined

Step 4: Progress

步骤4：进度跟踪

After each phase:
  1. Mark completed tasks: 
    TaskUpdate({taskId: "1", status: "completed"})
  2. Check newly-unblocked: 
    TaskList()
  3. Spawn next phase
每个阶段完成后：
  1. 标记已完成任务：
    TaskUpdate({taskId: "1", status: "completed"})
  2. 检查新解锁的任务：
    TaskList()
  3. 启动下一阶段

When to Use DAG Orchestration

何时使用DAG编排

ScenarioUse DAG?
Large migration (10+ files, phases)✅ Yes
Multi-feature release✅ Yes
Single feature (1-5 files)❌ Overkill
Quick fix❌ Overkill
For typical feature work, simple parallel spawning is sufficient.
场景是否使用DAG？
大型迁移（10+文件、多阶段）✅ 是
多功能版本发布✅ 是
单个功能（1-5个文件）❌ 过于复杂
快速修复❌ 过于复杂
对于典型的功能开发任务，简单的并行启动已足够。

Curation (Your Core Job)

整理（你的核心工作）

For each finding:
Validate: Real issue or false positive? Applies to our context? Filter: Generic advice, style preferences contradicting conventions Resolve Conflicts: When tools disagree, explain tradeoff, make recommendation
对于每个发现结果：
验证：是真实问题还是误报？是否适用于我们的场景？ 过滤：通用建议、与规范冲突的风格偏好 解决冲突：当工具意见不一致时，解释权衡方案并给出建议

Output Template

输出模板

markdown
undefined
markdown
undefined

[Task]: [subject]

[任务]: [主题]

Action Plan

行动计划

Critical

关键优先级

  • file:line
    — Issue — Fix: [action] (Source: [tool])
  • file:line
    — 问题 — 修复方案：[操作]（来源：[工具]）

Important

重要优先级

  • file:line
    — Issue — Fix: [action] (Source: [tool])
  • file:line
    — 问题 — 修复方案：[操作]（来源：[工具]）

Suggestions

建议项

  • [improvement] (Source: [tool])
  • [改进建议]（来源：[工具]）

Synthesis

合成结果

Agreements — Multiple tools flagged:
  • [issue]
Conflicts — Differing opinions:
  • [Tool A] vs [Tool B]: [your recommendation]
Research — From Gemini:
  • [finding with citation]
undefined
共识内容 — 多个工具共同标记的问题：
  • [问题]
冲突内容 — 不同工具的意见分歧：
  • [工具A] vs [工具B]：[你的建议]
研究结果 — 来自Gemini：
  • [带引用的发现结果]
undefined

When to Use

适用场景

  • Code review — Multiple perspectives on changes
  • Incident investigation — Agentic tools investigate, Thinktank validates fix
  • Architecture decisions — Thinktank for consensus
  • Audit/check tasks — Parallel investigation across domains
  • 代码审查 — 从多个视角审视代码变更
  • 事件调查 — Agent工具负责调查，Thinktank验证修复方案
  • 架构决策 — 使用Thinktank达成共识
  • 审计/检查任务 — 跨领域并行调查

Note

注意事项

All Codex delegation goes through Moonbridge MCP. Use 
mcp__moonbridge__spawn_agent
with 
adapter: "codex"
. This gives you:
  • Single interface for both Kimi and Codex
  • Mixed-adapter parallel spawning
  • Consistent parameter naming
所有Codex委托任务必须通过Moonbridge MCP进行。使用
mcp__moonbridge__spawn_agent
并设置
adapter: "codex"
。这样做的优势：
  • 统一的Kimi和Codex调用接口
  • 支持混合适配器的并行启动
  • 一致的参数命名规则

Related

相关链接

  • /llm-communication
    — Prompt writing principles
  • /review-branch
    — Example implementation
  • /thinktank
    — Multi-model synthesis
  • /codex-coworker
    — Codex delegation patterns
  • /llm-communication
    — 提示词编写原则
  • /review-branch
    — 示例实现
  • /thinktank
    — 多模型合成
  • /codex-coworker
    — Codex委托模式