rlm-subagent

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Subagent-Driven Execution with Fallback

基于Subagent的执行机制及回退方案

This skill provides parallel subagent execution for RLM Phase 3 (Implementation), Phase 3.5 (Code Review), and Phase 4 (Testing) with automatic fallback to sequential mode when subagents are unavailable.

该技能为RLM第三阶段（开发实现）、3.5阶段（代码审查）和第四阶段（测试验证）提供Subagent并行执行能力，当Subagent不可用时会自动回退至顺序执行模式。

Quick Reference

快速参考

Scenario	Action
Multiple independent sub-phases	Use Parallel Mode (subagents)
Single sub-phase or subagents unavailable	Use Sequential Mode (fallback)
Code review needed	Use Phase 3.5 (subagent or self-review)
Parallel testing	Use Phase 4 parallel dispatch

场景	操作
存在多个独立子阶段	使用并行模式（Subagent）
单个子阶段或Subagent不可用	使用顺序模式（回退方案）
需要代码审查	使用3.5阶段（Subagent审查或自审查）
并行测试	使用第四阶段并行调度

TODO Discipline for Subagents

Subagent的TODO规范

The Iron Law: NO SUBAGENT COMPLETION REPORT WITHOUT ALL TODOS CHECKED.

铁律：未完成所有TODO项，不得提交Subagent完成报告。

Subagent TODO Requirements

Subagent的TODO要求

Each subagent (implementer, code-reviewer) MUST:

Include
```
## TODO
```
section in their work output
Check off items as work progresses
Verify ALL items checked before reporting completion

每个Subagent（开发实现者、代码审查者）必须：

在工作输出中包含
```
## TODO
```
章节
随着工作推进勾选已完成项
在报告完成前确认所有项已勾选

Implementer Subagent TODO Template

开发实现Subagent的TODO模板

markdown

undefined

markdown

undefined

TODO

Read and understand assigned SP specification
Ask clarifying questions (if any)
Write failing test (RED phase)
Run test and verify failure
Implement minimal code (GREEN phase)
Run test and verify pass
Refactor while keeping tests green
Run integration tests
Self-review against plan
Document changes made
Report completion to controller

undefined

阅读并理解分配的SP规格说明
提出澄清问题（如有）
编写失败测试用例（RED阶段）
运行测试并验证失败结果
编写最简实现代码（GREEN阶段）
运行测试并验证通过结果
在保持测试通过的前提下重构代码
运行集成测试
对照计划进行自审查
记录所做的变更
向控制代理报告完成情况

undefined

Code Reviewer Subagent TODO Template

代码审查Subagent的TODO模板

markdown

undefined

markdown

undefined

TODO

Read original plan (Phase 3)
Read implementation summary (Phase 3)
Review git diff (BASE_SHA..HEAD_SHA)
Verify plan alignment
Assess code quality
Check TDD compliance
Categorize issues (Critical/Important/Minor)
Document positive findings
Render verdict
Report review completion

undefined

阅读原始计划（第三阶段）
阅读开发实现总结（第三阶段）
审查git diff（BASE_SHA..HEAD_SHA）
验证与计划的一致性
评估代码质量
检查TDD合规性
对问题进行分类（严重/重要/次要）
记录正面发现
给出审查结论
报告审查完成情况

undefined

Controller TODO Management

控制代理的TODO管理

The controller MUST:

Verify subagent outputs include completed TODOs
NOT accept completion reports with unchecked items
Request subagent to complete remaining items

控制代理必须：

验证Subagent的输出中包含已完成的TODO项
不接受存在未勾选项的完成报告
要求Subagent完成剩余项

Capability Detection

能力检测

At start of Phase 3, detect subagent availability:

IF can invoke "agent" or "Task" tool → Use Parallel Mode
ELSE → Use Sequential Fallback Mode

Detection rule: If the platform provides a subagent/task primitive, use parallel mode; otherwise fallback to sequential.

在第三阶段开始时，检测Subagent的可用性：

IF 可调用"agent"或"Task"工具 → 使用并行模式
ELSE → 使用顺序回退模式

检测规则： 如果平台提供子代理/任务原语，则使用并行模式；否则回退至顺序模式。

Parallel Mode (Subagents Available)

并行模式（Subagent可用）

Phase 3: Parallel Sub-Phase Implementation

第三阶段：并行子阶段开发实现

Controller responsibilities:

Read locked Phase 3 plan once
Extract all sub-phases (SP1, SP2, SP3...)
Determine dependencies (independent vs sequential)
Dispatch implementer subagent per independent SP
Two-stage review after each SP completion
Integration testing after all approved

Dispatch pattern:

typescript

// Parallel dispatch for independent SPs
await Promise.all([
  Task({ description: "Implement SP1", prompt: implementerPrompt(SP1) }),
  Task({ description: "Implement SP2", prompt: implementerPrompt(SP2) }),
  Task({ description: "Implement SP3", prompt: implementerPrompt(SP3) })
])

Two-stage review:

Spec Review: Verify SP requirements met (plan alignment)
Code Quality Review: Assess code quality, TDD compliance, standards

控制代理职责：

一次性读取锁定的第三阶段计划
提取所有子阶段（SP1、SP2、SP3……）
确定依赖关系（独立或顺序）
为每个独立SP调度开发实现Subagent
每个SP完成后进行两阶段审查
所有SP通过审查后执行集成测试

调度模式：

typescript

// 为独立SP调度并行任务
await Promise.all([
  Task({ description: "Implement SP1", prompt: implementerPrompt(SP1) }),
  Task({ description: "Implement SP2", prompt: implementerPrompt(SP2) }),
  Task({ description: "Implement SP3", prompt: implementerPrompt(SP3) })
])

两阶段审查：

规格审查： 验证是否满足SP需求（与计划的一致性）
代码质量审查： 评估代码质量、TDD合规性及编码标准

Phase 3.5: Code Review Subagent

3.5阶段：代码审查Subagent

Trigger: After Phase 3 implementation Action: Dispatch

agents/code-reviewer.md

subagent with:

BASE_SHA and HEAD_SHA of changes
Phase 3 plan for alignment check
Severity classification (Critical/Important/Minor)

Review loop: Issues found → implementer fixes → re-review

触发条件： 第三阶段开发完成后 操作： 调度

agents/code-reviewer.md

子代理，并提供：

变更的BASE_SHA和HEAD_SHA
第三阶段计划（用于一致性检查）
问题严重程度分类（严重/重要/次要）

审查循环： 发现问题 → 开发实现者修复 → 重新审查

Phase 4: Parallel Testing

第四阶段：并行测试

Dispatch pattern:

typescript

// Parallel test execution
await Promise.all([
  Task({ description: "Run unit tests", prompt: testPrompt("unit") }),
  Task({ description: "Run integration tests", prompt: testPrompt("integration") }),
  Task({ description: "Run E2E tests", prompt: testPrompt("e2e") })
])

Result aggregation:

Collect results from all subagents
Summarize pass/fail counts
Identify any critical failures

调度模式：

typescript

// 并行执行测试任务
await Promise.all([
  Task({ description: "Run unit tests", prompt: testPrompt("unit") }),
  Task({ description: "Run integration tests", prompt: testPrompt("integration") }),
  Task({ description: "Run E2E tests", prompt: testPrompt("e2e") })
])

结果汇总：

收集所有Subagent的测试结果
汇总通过/失败数量
识别严重失败项

Sequential Fallback Mode (No Subagents)

顺序回退模式（无Subagent可用）

Trigger: Subagent capability check fails

Characteristics:

Execute sub-phases sequentially in main agent context
Extended self-review checklist per sub-phase
Integration testing between each sub-phase
Full context preservation

Fallback trigger flow:

markdown

**Subagent Check:** ❌ NOT AVAILABLE
**Reason:** [Tool not found / Platform limitation / User request]
**Action:** Using SEQUENTIAL fallback mode

触发条件： Subagent能力检测不通过

特性：

在主代理上下文中顺序执行子阶段
每个子阶段使用扩展自审查清单
子阶段之间执行集成测试
完整保留上下文信息

回退触发流程：

markdown

**Subagent检查：** ❌ 不可用
**原因：** [工具未找到 / 平台限制 / 用户要求]
**操作：** 使用顺序回退模式

Sequential Execution Pattern

顺序执行模式

For each SP in [SP1, SP2, SP3...]:
  1. Execute SP implementation
  2. Self-review against plan (extended checklist)
  3. Document in Phase 4 artifact
  4. Run integration tests
  5. Proceed to next SP

For each SP in [SP1, SP2, SP3...]:
  1. 执行SP开发实现
  2. 对照计划进行扩展自审查
  3. 在第四阶段工件中记录
  4. 运行集成测试
  5. 进入下一个SP

Sequential Review Pattern

顺序审查模式

Phase 3.5 equivalent:

Main agent performs extended self-review
Use comprehensive checklist from
```
agents/code-reviewer.md
```
Document findings in Phase 3.5 artifact

等效3.5阶段操作：

主代理执行扩展自审查
使用
```
agents/code-reviewer.md
```
中的全面审查清单
在3.5阶段工件中记录审查结果

Subagent Prompts

Subagent提示词

Implementer Subagent

开发实现Subagent

File:

agents/implementer.md

Key requirements:

Self-contained context (full SP text provided)
TDD discipline enforcement
Question-before-work protocol
Self-review checklist before completion

Usage:

markdown

You are an Implementer Agent. Implement this sub-phase:

**SP Text:** [full sub-phase text]
**BASE_SHA:** [commit SHA]
**Context:** [relevant files]

Follow the process in agents/implementer.md

文件：

agents/implementer.md

核心要求：

上下文自包含（提供完整SP文本）
强制执行TDD规范
先提问再工作的流程
完成前执行自审查清单

使用方式：

markdown

You are an Implementer Agent. Implement this sub-phase:

**SP Text:** [full sub-phase text]
**BASE_SHA:** [commit SHA]
**Context:** [relevant files]

Follow the process in agents/implementer.md

Code Reviewer Subagent

代码审查Subagent

File:

agents/code-reviewer.md

Key requirements:

Plan alignment verification
Code quality assessment
Severity classification (Critical/Important/Minor)
Clear verdict (Approved / Changes Required)

Usage:

markdown

You are a Code Reviewer Agent. Review this implementation:

**Plan:** [Phase 3 TO-BE]
**Git Range:** [BASE_SHA..HEAD_SHA]
**Implementation:** [Phase 4 summary]

Follow the process in agents/code-reviewer.md

文件：

agents/code-reviewer.md

核心要求：

验证与计划的一致性
评估代码质量
问题严重程度分类（严重/重要/次要）
明确给出审查结论（通过/需要修改）

使用方式：

markdown

You are a Code Reviewer Agent. Review this implementation:

**Plan:** [Phase 3 TO-BE]
**Git Range:** [BASE_SHA..HEAD_SHA]
**Implementation:** [Phase 4 summary]

Follow the process in agents/code-reviewer.md

Artifact Documentation

工件文档规范

Phase 4 artifact must include:

markdown

undefined

第四阶段工件必须包含：

markdown

undefined

Execution Mode

Mode: Parallel / Sequential
Subagents Used: [names and counts]
Fallback Reason: [if applicable]

Mode: Parallel / Sequential
Subagents Used: [names and counts]
Fallback Reason: [if applicable]

Sub-phase Results

SP1: [status] - [subagent name or "main agent"]
SP2: [status] - [subagent name or "main agent"] ...

SP1: [status] - [subagent name or "main agent"]
SP2: [status] - [subagent name or "main agent"] ...


**3.5阶段工件（若使用）：**
```markdown

Review Results

Review Scope

SP1 Review: [status] - [reviewer name or "self-review"]
SP2 Review: [status] - [reviewer name or "self-review"] ...


**Phase 3.5 artifact (if used):**
```markdown

Git range: [BASE_SHA..HEAD_SHA]
Execution Mode: Parallel (subagent) / Sequential (self-review)
Reviewer: [subagent name / self]

Review Scope

Issues Found

Git range: [BASE_SHA..HEAD_SHA]
Execution Mode: Parallel (subagent) / Sequential (self-review)
Reviewer: [subagent name / self]

Critical: [count]
Important: [count]
Minor: [count]

Issues Found

Verdict

Critical: [count]
Important: [count]
Minor: [count]

[APPROVED / APPROVED WITH NOTES / CHANGES REQUIRED]


**第四阶段工件必须包含：**
```markdown

Verdict

Execution Mode

[APPROVED / APPROVED WITH NOTES / CHANGES REQUIRED]


**Phase 4 artifact must include:**
```markdown

Mode: Parallel / Sequential
Test Suites:
- Unit: [subagent name] / Main agent
- Integration: [subagent name] / Main agent
- E2E: [subagent name] / Main agent

Execution Mode

Results Summary

Mode: Parallel / Sequential
Test Suites:
- Unit: [subagent name] / Main agent
- Integration: [subagent name] / Main agent
- E2E: [subagent name] / Main agent

Total execution time: [X] minutes
Estimated sequential time: [Y] minutes
Speedup: [Z]x

undefined

Results Summary

决策树

Total execution time: [X] minutes
Estimated sequential time: [Y] minutes
Speedup: [Z]x

undefined

Starting Phase 4
        │
        ▼
Subagent available? ──NO──► SEQUENTIAL MODE
        │                   - Execute SPs sequentially
        │                   - Extended self-review
        │                   - Document as sequential
        │
       YES
        │
        ▼
3+ independent SPs? ──NO──► SEQUENTIAL MODE
        │                   - Single sub-phase
        │                   - No parallelism benefit
        │
       YES
        │
        ▼
   PARALLEL MODE
   - Dispatch implementer per SP
   - Two-stage review
   - Integration after all done

Decision Tree

质量保障

Starting Phase 4
        │
        ▼
Subagent available? ──NO──► SEQUENTIAL MODE
        │                   - Execute SPs sequentially
        │                   - Extended self-review
        │                   - Document as sequential
        │
       YES
        │
        ▼
3+ independent SPs? ──NO──► SEQUENTIAL MODE
        │                   - Single sub-phase
        │                   - No parallelism benefit
        │
       YES
        │
        ▼
   PARALLEL MODE
   - Dispatch implementer per SP
   - Two-stage review
   - Integration after all done

两种模式均强制执行：

TDD规范
工件锁定
关卡合规性
集成测试

并行模式优势：

每个子阶段拥有独立新鲜上下文
独立验证机制
执行速度更快（3-5倍）

顺序模式优势：

不依赖Subagent
协调更简单
上下文完整累积

Quality Maintenance

常见模式

—

模式1：多领域功能开发

Both modes enforce:

TDD discipline
Artifact locking
Gate compliance
Integration testing

Parallel Mode advantages:

Fresh context per sub-phase
Independent verification
Faster execution (3-5x)

Sequential Mode advantages:

No subagent dependency
Simpler coordination
Full context accumulation

场景： 功能涉及API、UI和数据库

第三阶段计划：

markdown

undefined

Common Patterns

Sub-phases

Pattern 1: Multi-Domain Feature

—

Scenario: Feature touches API, UI, and database

Phase 3 plan:

markdown

undefined

SP1: API changes (backend/)
SP2: UI changes (frontend/)
SP3: Database migration (migrations/)


**第四阶段执行：**
- SP1、SP2、SP3并行调度（独立领域）
- 每个开发实现Subagent处理独立文件
- 所有完成后执行集成测试

Sub-phases

模式2：顺序依赖场景

SP1: API changes (backend/)
SP2: UI changes (frontend/)
SP3: Database migration (migrations/)


**Phase 4 execution:**
- SP1, SP2, SP3 dispatched in parallel (independent domains)
- Each implementer works on isolated files
- Integration testing after all complete

场景： SP2依赖SP1的输出

第四阶段执行：

调度SP1 → 等待完成 → 审查 → 通过
调度SP2（携带SP1上下文） → 等待完成 → 审查 → 通过
顺序执行，不使用并行

Pattern 2: Sequential Dependencies

模式3：高风险变更场景

Scenario: SP2 depends on SP1 output

Phase 4 execution:

Dispatch SP1 → wait → review → approve
Dispatch SP2 (with SP1 context) → wait → review → approve
Sequential, not parallel

场景： 关键认证系统变更

第三阶段： 采用TDD方式开发实现 3.5阶段： 强制代码审查 第四阶段： 并行测试（单元测试+集成测试+E2E测试）

Pattern 3: High-Risk Change

故障排查

—

Subagent执行失败

Scenario: Critical auth system change

Phase 4: Implement with TDD Phase 3.5: Mandatory code review Phase 4: Parallel testing (unit + integration + E2E)

检查提示词是否上下文自包含
确认已提供BASE_SHA
查看日志排查错误
使用更清晰的指令重试

Troubleshooting

并行SP执行后集成测试失败

Subagent fails to complete

—

Check if prompt was self-contained
Verify BASE_SHA was provided
Review logs for errors
Retry with clearer instructions

检查SP之间的文件冲突
验证是否存在共享可变状态
查看提交历史确认顺序问题

Integration tests fail after parallel SPs

顺序模式执行过慢

Check for file conflicts between SPs
Verify no shared mutable state
Review commit history for ordering issues

考虑将计划拆分为更细粒度的SP
检查平台是否支持Subagent
验证工具检测功能是否正常

Sequential mode too slow

参考资料

Consider breaking into more granular SPs
Check if platform supports subagents
Verify tool detection is working

主工作流：
```
SKILL.md
```
开发实现Subagent：
```
agents/implementer.md
```
代码审查Subagent：
```
agents/code-reviewer.md
```
工件模板：
```
references/artifact-template.md
```

References

—

Main workflow:
```
SKILL.md
```
Implementer subagent:
```
agents/implementer.md
```
Code reviewer subagent:
```
agents/code-reviewer.md
```
Artifact templates:
```
references/artifact-template.md
```

—