4.3 KiB
4.3 KiB
Self-Evolving Skill - OpenClaw集成指南
安装完成 ✅
文件位置
| 位置 | 说明 |
|---|---|
~/.openclaw/skills/self-evolving-skill/ |
技能根目录 |
~/.openclaw/agents/main/agent/mcp_servers.json |
MCP服务器配置 |
~/.openclaw/skills/self-evolving-skill/storage/ |
数据存储 |
项目结构
~/.openclaw/skills/self-evolving-skill/
├── core/ # Python核心
│ ├── residual_pyramid.py # SVD分解
│ ├── reflection_trigger.py # 自适应触发
│ ├── experience_replay.py # 经验回放
│ ├── skill_engine.py # 核心引擎
│ ├── storage.py # 持久化
│ └── mcp_server.py # MCP服务器
├── src/ # TypeScript SDK
├── SKILL.md # 技能文档
├── package.json # npm配置
├── mcporter_adapter.py # mcporter适配器
└── venv/ # Python虚拟环境
使用方式
1. 直接调用（推荐）
# 激活虚拟环境
source ~/.openclaw/skills/self-evolving-skill/venv/bin/activate
# 列出所有Skill
python3 ~/.openclaw/skills/self-evolving-skill/mcporter_adapter.py skill_list '{}'
# 创建新Skill
python3 ~/.openclaw/skills/self-evolving-skill/mcporter_adapter.py skill_create '{"name":"MySkill"}'
# 分析嵌入
python3 ~/.openclaw/skills/self-evolving-skill/mcporter_adapter.py skill_analyze '{"embedding":[0.1,0.2,0.3]}'
# 系统统计
python3 ~/.openclaw/skills/self-evolving-skill/mcporter_adapter.py skill_stats '{}'
2. OpenClaw MCP调用
在OpenClaw中可直接调用MCP工具：
{
"tool": "skill_create",
"arguments": {
"name": "ProblemSolver",
"description": "问题解决技能"
}
}
3. 启动MCP服务器
# 前台运行
source ~/.openclaw/skills/self-evolving-skill/venv/bin/activate
python3 ~/.openclaw/skills/self-evolving-skill/mcp_server.py --storage ~/.openclaw/skills/self-evolving-skill/storage
# 或通过配置自动启动（已在mcp_servers.json中配置）
MCP工具
| 工具 | 描述 | 参数 |
|---|---|---|
skill_create |
创建新的自演化Skill | name, description |
skill_execute |
执行Skill并触发学习 | skill_id, context, success, value |
skill_analyze |
分析嵌入向量（不触发学习） | embedding |
skill_list |
列出所有已保存的Skill | - |
skill_stats |
获取系统统计信息 | - |
skill_save |
持久化保存Skill | skill_id |
skill_load |
加载已保存的Skill | skill_id |
skill_clear |
清空所有数据和缓存 | - |
测试结果
=== skill_list ===
Skills: 20
=== skill_create ===
{"skill_id":"1ac4a2cb3f79347f","name":"TestOpenClaw"}
=== skill_analyze ===
{
"total_energy": 0.55,
"residual_ratio": 0.086,
"suggested_abstraction": "POLICY",
"novelty_score": 0.657
}
核心算法
1. 残差金字塔分解
pyramid = ResidualPyramid(max_layers=5, use_pca=True)
decomposition = pyramid.decompose(embedding)
# 输出:
# - residual_ratio: 残差能量比率
# - suggested_abstraction: POLICY / SUB_SKILL / PREDICATE
# - novelty_score: 综合新颖性
2. 三层跃迁规则
| 覆盖率 | 抽象层级 | 操作 |
|---|---|---|
| >80% | POLICY | 调整策略权重 |
| 40-80% | SUB_SKILL | 生成子Skill |
| <40% | PREDICATE | 归纳新谓词 |
3. 自适应阈值
trigger = ReflectionTrigger(
min_energy_ratio=0.10, # 初始阈值
value_gain_threshold=0.20, # 触发阈值
target_trigger_rate=0.15 # 目标15%触发率
)
OpenClaw配置
MCP服务器已配置在：
// ~/.openclaw/agents/main/agent/mcp_servers.json
{
"servers": {
"self-evolving-skill": {
"name": "self-evolving-skill",
"type": "stdio",
"command": "/bin/bash",
"args": [
"-c",
"source ~/.openclaw/skills/self-evolving-skill/venv/bin/activate && python3 ~/.openclaw/skills/self-evolving-skill/mcp_server.py --storage ~/.openclaw/skills/self-evolving-skill/storage"
]
}
}
}
下一步
- 在OpenClaw中测试MCP工具调用
- 集成到Agent执行流程
- 添加强化学习策略优化