Expand description
Skill synthesis: auto-generate, refine, and retire skills (EVO-4).
- Generate: 既总结成功经验,也总结失败经验
- 成功驱动:高成功率重复模式 → SKILL.md + script
- 失败驱动:持续失败模式 → 补全能力缺口的 Skill
- Refine: failed skill → analyze error trace → LLM fix → retry (max 2 rounds)
- Retire: low success rate or unused skills → archive
§React / Check / Retry
重度依赖大模型能力,每个环节都有校验与重试:
- Check: L3 内容门禁、L4 安全扫描、test_skill_invoke 实测
- Retry: 任何 LLM 输出 JSON 解析失败时,将错误反馈给大模型并重试 1 次
- 代码修改/修复仅由大模型完成,不使用正则或模式匹配
All evolved skills live in chat/skills/_evolved/ with .meta.json metadata.
A10: Newly generated skills go to _evolved/_pending/ until user confirms.
Structs§
- Skill
Meta - Skill
Validation - 单个技能的验证结果
Functions§
- confirm_
pending_ skill - evolve_
skills - Run skill evolution: generate new skills or refine existing ones.
- list_
pending_ skills - list_
pending_ skills_ with_ review - reject_
pending_ skill - repair_
one_ skill - 修复单个技能:打包整个目录给模型,模型返回修复,应用后验证,最多 MAX_REFINE_ROUNDS 轮
- repair_
skills - 修复技能:先验证,再对失败的逐个打包修复
- track_
skill_ usage - Update .meta.json after a skill execution (called from agent_loop).
- validate_
skills - 验证技能:对每个 skill infer → test → doc check,返回结果列表。
若
skill_names_filter为Some且非空,仅验证该列表中的技能名(目录名),否则验证全部。