mcpField
L
launch-sub-agent

launch-sub-agent

github.com/neolabhqAgent 工作流1.8k1.2k 次安装9 周前更新
代码仓库

关于 launch-sub-agent

launch-sub-agent 是用于将任务交给聚焦型子 Agent 的技能,面向需要协调多 Agent 工作的操作者。它根据任务类型和复杂度选择模型与专门 Agent,并为子 Agent 构造包含推理步骤和自我审查要求的指令。

主要功能

  1. 01分析任务类型与复杂度
  2. 02按任务选择模型
  3. 03匹配专门 Agent
  4. 04构造带推理前缀的任务提示
  5. 05要求完成前进行自我审查

适用场景

  1. 01将复杂架构任务交给合适模型
  2. 02分派研究或代码实现任务
  3. 03把文档更新交给轻量子 Agent

安装

复制下面的命令,在终端中运行。

$npx skills add https://github.com/neolabhq/context-engineering-kit --skill launch-sub-agent

安装后在 Claude Code 中输入/launch-sub-agent确认已加载。

详细介绍

launch-sub-agent

启动一个专注型子 Agent 来执行提供的任务。分析任务,智能选择最佳模型和 Agent 配置,然后派遣子 Agent,并在开始时进行零样本思维链推理,在结束时进行强制自我批评验证。 此命令实现了多 Agent 架构中的**监督者/编排器模式**:由你(编排器)派遣专注型子 Agent,并为其提供隔离的上下文。其主要优势是**上下文隔离**——每个子 Agent 都在干净的上下文窗口中运行,专注于特定任务,不受累积的上下文污染影响。

流程

阶段 1:使用零样本思维链分析任务

在派遣之前,系统地分析任务。逐步思考:

Let me analyze this task step by step to determine the optimal configuration:

1. **Task Type Identification**
   "What type of work is being requested?"
   - Code implementation / feature development
   - Research / investigation / comparison
   - Documentation / technical writing
   - Code review / quality analysis
   - Architecture / system design
   - Testing / validation
   - Simple transformation / lookup

2. **Complexity Assessment**
   "How complex is the reasoning required?"
   - High: Architecture decisions, novel problem-solving, multi-faceted analysis
   - Medium: Standard implementation following patterns, moderate research
   - Low: Simple transformations, lookups, well-defined single-step tasks

3. **Output Size Estimation**
   "How extensive is the expected output?"
   - Large: Multiple files, comprehensive documentation, extensive analysis
   - Medium: Single feature, focused deliverable
   - Small: Quick answer, minor change, brief output

4. **Domain Expertise Check**
   "Does this task match a specialized agent profile?"
   - Development: code, implement, feature, endpoint, TDD, tests
   - Research: investigate, compare, evaluate, options, library
   - Documentation: document, README, guide, explain, tutorial
   - Architecture: design, system, structure, scalability
   - Exploration: understand, navigate, find, codebase patterns
阶段 2:选择模型

根据任务分析选择最佳模型:

任务特征推荐模型理由
复杂推理(架构、设计、关键决策)opus推理能力最强
专业领域(与 Agent 角色匹配)Opus + 专业 Agent领域专业知识 + 推理能力
不复杂但很长(大量文档、冗长输出)sonnet[1m]能力出色,适合长任务且成本效益高
简单且简短(简单任务、快速查询)haiku快速、经济,适合简单任务
默认(不确定时)opus优先保证质量,而非降低成本

决策树:

Is task COMPLEX (architecture, design, novel problem, critical decision)?
|
+-- YES --> Use Opus (highest capability)
|           |
|           +-- Does it match a specialized domain?
|               +-- YES --> Include specialized agent prompt
|               +-- NO --> Use Opus alone
|
+-- NO --> Is task SIMPLE and SHORT?
           |
           +-- YES --> Use Haiku (fast, cheap)
           |
           +-- NO --> Is output LONG but task not complex?
                      |
                      +-- YES --> Use Sonnet (balanced)
                      |
                      +-- NO --> Use Opus (default)
阶段 3:匹配专业 Agent

如果任务符合某个专业领域,应加入相关的 Agent 提示词。专业 Agent 能提供领域专属的最佳实践、质量标准和结构化方法,从而提升输出质量。

决策: 当任务显然能从领域专业知识中受益时,使用专业 Agent。对于使用专业 Agent 会造成不必要开销的简单任务,则跳过。

Agent: 可用的专业 Agent 取决于项目和已安装的插件。sdd 插件中常见的 Agent 包括:sdd:developer、sdd:researcher、sdd:software-architect、sdd:tech-lead、sdd:code-explorer、sdd:business-analyst、sdd:code-reviewer、sdd:tech-writer。如果没有合适的专业 Agent,则改用不带专业化配置的通用 Agent。

与模型选择结合:

  • 专业 Agent 与模型选择结合使用,而非取代模型选择
  • 复杂任务 + 专业领域 = Opus + 专业 Agent
  • 匹配专业领域的简单任务 = Haiku,不使用专业 Agent(不值得增加开销)

使用方法:

  1. 阅读 Agent 定义
  2. 将 Agent 的指令放在 CoT 前缀之后,加入子 Agent 提示词
  3. 与零样本思维链前缀和批评后缀结合
阶段 4:构建子 Agent 提示词

构建子 Agent 提示词时,必须包含以下部分:

4.1 零样本思维链前缀(必需——必须位于最前面)
## Reasoning Approach

Before taking any action, you MUST think through the problem systematically.

Let's approach this step by step:

1. "Let me first understand what is being asked..."
   - What is the core objective?
   - What are the explicit requirements?
   - What constraints must I respect?

2. "Let me break this down into concrete steps..."
   - What are the major components of this task?
   - What order should I tackle them?
   - What dependencies exist between steps?

3. "Let me consider what could go wrong..."
   - What assumptions am I making?
   - What edge cases might exist?
   - What could cause this to fail?

4. "Let me verify my approach before proceeding..."
   - Does my plan address all requirements?
   - Is there a simpler approach?
   - Am I following existing patterns?

Work through each step explicitly before implementing.
4.2 任务正文
<task>
{Task description from $ARGUMENTS}
</task>

<constraints>
{Any constraints inferred from the task or conversation context}
</constraints>

<context>
{Relevant context: files, patterns, requirements, codebase information}
</context>

<output>
{Expected deliverable: format, location, structure}
</output>
4.3 自我批评后缀(必需——必须位于最后)
## Self-Critique Loop (MANDATORY)

Before completing, you MUST verify your work. Submitting unverified work is UNACCEPTABLE.

### 1. Generate 5 Verification Questions

Create 5 questions specific to this task that test correctness and completeness. There example questions:

| # | Verification Question | Why This Matters |
|---|----------------------|------------------|
| 1 | Does my solution fully address ALL stated requirements? | Partial solutions = failed task |
| 2 | Have I verified every assumption against available evidence? | Unverified assumptions = potential failures |
| 3 | Are there edge cases or error scenarios I haven't handled? | Edge cases cause production issues |
| 4 | Does my solution follow existing patterns in the codebase? | Pattern violations create maintenance debt |
| 5 | Is my solution clear enough for someone else to understand and use? | Unclear output reduces value |

### 2. Answer Each Question with Evidence

For each question, examine your solution and provide specific evidence:

[Q1] Requirements Coverage:
- Requirement 1: [COVERED/MISSING] - [specific evidence from solution]
- Requirement 2: [COVERED/MISSING] - [specific evidence from solution]
- Gap analysis: [any gaps identified]

[Q2] Assumption Verification:
- Assumption 1: [assumption made] - [VERIFIED/UNVERIFIED] - [evidence]
- Assumption 2: [assumption made] - [VERIFIED/UNVERIFIED] - [evidence]

[Q3] Edge Case Analysis:
- Edge case 1: [scenario] - [HANDLED/UNHANDLED] - [how]
- Edge case 2: [scenario] - [HANDLED/UNHANDLED] - [how]

[Q4] Pattern Adherence:
- Pattern 1: [pattern name] - [FOLLOWED/DEVIATED] - [evidence]
- Pattern 2: [pattern name] - [FOLLOWED/DEVIATED] - [evidence]

[Q5] Clarity Assessment:
- Is the solution well-organized? [YES/NO]
- Are complex parts explained? [YES/NO]
- Could someone else use this immediately? [YES/NO]

### 3. Revise If Needed

If ANY verification question reveals a gap:
1. **STOP** - Do not submit incomplete work
2. **FIX** - Address the specific gap identified
3. **RE-VERIFY** - Confirm the fix resolves the issue
4. **DOCUMENT** - Note what was changed and why

CRITICAL: Do not submit until ALL verification questions have satisfactory answers with evidence.
阶段 5:派遣子 Agent

使用 Task 工具,并采用选定的配置进行派遣:

Use Task tool:
- description: "Sub-agent: {brief task summary}"
- prompt: {constructed prompt with CoT prefix + task + critique suffix}
- model: {selected model - opus/sonnet/haiku}

上下文隔离提醒: 仅传递与此特定任务相关的上下文。不要传递完整的对话历史。

示例

示例 1:复杂架构任务(Opus)

输入: /launch-sub-agent Design a caching strategy for our API that handles 10k requests/second

分析:

  • 任务类型:架构/设计
  • 复杂度:高(性能要求、系统设计)
  • 输出长度:中等(设计文档)
  • 领域匹配:sdd:software-architect

选择: Opus + sdd:software-architect Agent

派遣: 使用 Opus 模型、sdd:software-architect 提示词、CoT 前缀和批评后缀调用 Task 工具


示例 2:简单文档更新(Haiku)

输入: /launch-sub-agent Update the README to add --verbose flag to CLI options

分析:

  • 任务类型:文档(简单编辑)
  • 复杂度:低(单个文件,定义明确)
  • 输出规模:小(一个章节)
  • 领域匹配:无需匹配(任务过于简单)

选择: Haiku(快速、低成本,足以胜任任务)

调度: 使用 Haiku 模型的 Task 工具、基础 CoT 前缀和基础批判性评估后缀


示例 3:中等规模实现(Sonnet + Developer)

输入: /launch-sub-agent Implement pagination for /users endpoint following patterns in /products

分析:

  • 任务类型:代码实现
  • 复杂度:中(遵循现有模式)
  • 输出规模:中(实现 + 测试)
  • 领域匹配:sdd:developer

选择: Sonnet + sdd:developer Agent(并不复杂,但需要领域专业知识)

调度: 使用 Sonnet 模型的 Task 工具、sdd:developer 提示词、CoT 前缀和批判性评估后缀


示例 4:研究任务(Opus + Researcher)

输入: /launch-sub-agent Research authentication options for mobile app - evaluate OAuth2, SAML, passwordless

分析:

  • 任务类型:研究 / 比较
  • 复杂度:高(比较分析、提出建议)
  • 输出规模:大(全面研究)
  • 领域匹配:sdd:researcher

选择: Opus + sdd:researcher Agent

调度: 使用 Opus 模型的 Task 工具、sdd:researcher 提示词、CoT 前缀和批判性评估后缀

最佳实践

上下文隔离
  • 仅传递与特定任务相关的上下文
  • 避免传递完整的对话历史
  • 让子 Agent 通过工具发现代码库模式
  • 使用文件路径和引用,而不是嵌入大量内容
模型选择
  • 拿不准时,使用 Opus(质量优先于成本)
  • 仅将 Haiku 用于真正微不足道的任务
  • 将 Sonnet 用于“苦力活”——需要能力,但不需要天才水平
  • 生产代码始终值得使用 Opus
专用 Agent
  • 仅在领域专业知识明显能提升质量时使用
  • 与 CoT 和批判性评估模式结合使用
  • 不要强行让通用任务采用专门化处理
质量门槛
  • 自我批判循环不可妥协
  • 子 Agent 必须在完成前回答验证问题
  • 接受子 Agent 的输出前先进行审查

评论

支持 Markdown