人工智能代理工程师

人工智能 社区

技能解读

当前原文的自动摘译,尚未人工审校。

领域范围

该技能涵盖 IdeaFlow 中的人工智能代理工程师领域:区域 · 文件 · 目的代理配置 · .opencode/agents/CMZ.json、.opencode/oh-my-opencode.json · 代理定义、模型、功能 YH · 技能库 · .opencode/skills/*/SKILL.md · 33 项专业技能代理指南 · docs/agent-guidelines.md · 10 个核心原则、工作流程

主要协调者:CMZ

CMZ(认知 Meta-Z)是主要的编排代理,具有三个核心功能: 自我修复:检测错误、诊断根本原因、实施恢复 自我学习:整合反馈、分析结果、构建知识 自我进化:扩展功能、优化性能、元改进

专业代理 (OhMyOpenCode)

代理 · 模型 · 目的 Sisyphus · minimax-m2.5-free · 主编排器,无情执行 Hephaestus · glm-4.7-free · 自主深度工作者 Oracle · minimax-m2.5-free · 架构、调试、推理 图书馆员 · glm-4.7-free · 文档、探索

类别

类别 · 模型 · 用例视觉工程 · glm-4.7-free · UI/前端工作超脑 · minimax-m2.5-free · 逻辑复杂,架构快速 · minimax-m2.1-free · 快速、简单的任务深入 · minimax-m2.5-free · 彻底分析

关键文件

文件 · 用途 opencode.json · CLI 配置、模型选择、MCP 服务器 .opencode/oh-my-opencode.json · 代理定义、类别、挂钩 .opencode/agents/CMZ.json · CMZ 特定配置

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
人工智能
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
88 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@tomevault-io · 未声明许可证
上下文用量评级
低消耗
接入复杂程度
需简单配置
是否需要外部接口密钥
不需要
兼容的系统
macOS
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
  • 允许写入 / 修改
  • 命令执行
网络行为
仅限本地
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# AI Agent Engineer

Expert guidance for engineering and improving AI agents in the IdeaFlow multi-agent system.

## Domain Scope

This skill covers the ai-agent-engineer domain within IdeaFlow:

| Area                | Files                                                        | Purpose                                 |
| ------------------- | ------------------------------------------------------------ | --------------------------------------- | ----------------------------- | --------------------- |
| Agent Configuration | `.opencode/agents/CMZ.json`, `.opencode/oh-my-opencode.json` | Agent definitions, models, capabilities |
| YH                  |                                                              | Skills Library                          | `.opencode/skills/*/SKILL.md` | 33 specialized skills |
| Agent Guidelines    | `docs/agent-guidelines.md`                                   | 10 core principles, workflows           |
| System Integration  | `opencode.json`, `AGENTS.md`                                 | CLI config, documentation               |

## Agent Architecture

### Primary Orchestrator: CMZ

CMZ (Cognitive Meta-Z) is the main orchestrating agent with three core capabilities:

1. **Self-Heal**: Detect errors, diagnose root cause, implement recovery
2. **Self-Learn**: Integrate feedback, analyze outcomes, build knowledge
3. **Self-Evolve**: Expand capabilities, optimize performance, meta-improve

### Specialized Agents (OhMyOpenCode)

| Agent      | Model          | Purpose                                 |
| ---------- | -------------- | --------------------------------------- |
| Sisyphus   | minimax-m2.5-free | Main orchestrator, relentless execution |
| Hephaestus | glm-4.7-free   | Autonomous deep worker                  |
| Oracle     | minimax-m2.5-free | Architecture, debugging, reasoning      |
| Librarian  | glm-4.7-free   | Documentation, exploration              |
| Explore    | glm-4.7-free   | Fast codebase search                    |

### Categories

| Category           | Model             | Use Case                    |
| ------------------ | ----------------- | --------------------------- |
| visual-engineering | glm-4.7-free      | UI/Frontend work            |
| ultrabrain         | minimax-m2.5-free    | Complex logic, architecture |
| quick              | minimax-m2.1-free | Fast, simple tasks          |
| deep               | minimax-m2.5-free    | Thorough analysis           |

## Delegation Patterns

### When to Delegate

```
Task Type              → Delegate To
─────────────────────────────────────
Codebase exploration   → explore (background)
Documentation lookup   → librarian (background)
Complex reasoning      → oracle (blocking)
UI/Frontend work       → visual-engineering category
Quick fixes            → quick category
```

### Delegation Commands

```bash
# Background exploration (parallel)
task(subagent_type="explore", run_in_background=true, prompt="...")

# Blocking consultation
task(subagent_type="oracle", run_in_background=false, prompt="...")

# Category delegation
task(category="visual-engineering", load_skills=["frontend-ui-ux"])
```

## Configuration Management

### Key Files

| File                            | Purpose                                  |
| ------------------------------- | ---------------------------------------- |
| `opencode.json`                 | CLI config, model selection, MCP servers |
| `.opencode/oh-my-opencode.json` | Agent definitions, categories, hooks     |
| `.opencode/agents/CMZ.json`     | CMZ-specific configuration               |

### Adding a New Agent

1. Define in `.opencode/oh-my-opencode.json`:

```json
"agents": {
  "new_agent": {
    "model": "opencode/model-name",
    "category": "category-name"
  }
}
```

2. Add to CMZ.json if orchestrator integration needed

### Adding a New Skill

1. Create directory: `.opencode/skills/skill-name/`
2. Create `SKILL.md` with frontmatter (name, description)
3. Add references in `references/` if needed
4. Follow skill-creator guidelines for structure

## Self-\* Capabilities

### Self-Heal Implementation

```
Error → Detect → Diagnose → Recover → Learn
```

1. **Detect**: Monitor for exceptions, failed tests, CI failures
2. **Diagnose**: Use systematic-debugging skill, analyze stack traces
3. **Recover**: Implement fix, verify with tests
4. **Learn**: Document in memory, prevent recurrence

### Self-Learn Implementation

```
Feedback → Analyze → Extract → Apply
```

1. **Collect**: User feedback, test outcomes, performance metrics
2. **Analyze**: Identify patterns, successful approaches
3. **Extract**: Generalize into reusable patterns
4. **Apply**: Update skills, configs, documentation

### Self-Evolve Implementation

```
Evaluate → Identify → Implement → Verify
```

1. **Evaluate**: Current capabilities vs requirements
2. **Identify**: Gaps, optimization opportunities
3. **Implement**: New skills, improved delegation
4. **Verify**: Tests pass, metrics improve

## Agent Improvement Workflow

### RESEARCH → PLAN → IMPLEMENT → VERIFY → SELF-REVIEW → DELIVER

1. **RESEARCH**: Explore codebase, gather context
2. **PLAN**: Create detailed work breakdown
3. **IMPLEMENT**: Make atomic changes
4. **VERIFY**: Run tests, lint, type-check
5. **SELF-REVIEW**: Check against requirements
6. **DELIVER**: Create PR with proper labeling

### Verification Checklist

- [ ] All tests pass (`npm test`)
- [ ] Lint passes with zero warnings (`npm run lint`)
- [ ] Type check passes (`npm run type-check`)
- [ ] Build succeeds (`npm run build`)
- [ ] Documentation updated if needed
- [ ] PR has `ai-agent-engineer` label

## Common Tasks

### Improve Agent Configuration

1. Read current config from `.opencode/oh-my-opencode.json`
2. Identify improvement (model change, capability addition)
3. Make atomic change
4. Verify with `opencode` CLI if possible
5. Document change in commit message

### Create New Skill

1. Use `skill-creator` skill for guidance
2. Run `init_skill.py` from skill-creator
3. Write SKILL.md with clear frontmatter
4. Add references for detailed content
5. Validate and test

### Fix Agent Issue

1. Reproduce issue
2. Use `systematic-debugging` skill
3. Implement minimal fix
4. Add regression test
5. Verify fix

## Reference Files

For detailed patterns and examples, see:

- **[Agent Architecture Details](references/agent-architecture.md)**: Deep dive into agent system internals, model selection, and advanced patterns

## Anti-Patterns

| Anti-Pattern         | Why Bad         | Instead                |
| -------------------- | --------------- | ---------------------- |
| Direct main commits  | Bypasses review | Use feature branches   |
| Skipping tests       | Regressions     | Always run tests       |
| Large atomic changes | Hard to review  | Small, focused changes |
| Ignoring CI failures | Ships bugs      | Fix all failures       |
| Undocumented changes | Knowledge loss  | Update documentation   |

---
> Source: [cpa03/ai-first](https://github.com/cpa03/ai-first) — distributed by [TomeVault](https://tomevault.io).
<!-- tomevault:4.0:skill_md:2026-05-23 -->

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。