Self-Evolve Skill

Agent自进化机制。三层架构（记忆→技能→规范），三种触发（关键词+踩坑+复盘），用户审批+Darwin评分双重质控。用得越多越懂你。触发词：自进化、进化、evolve、记住、以后、永远、下次、踩坑经验。

Security Scan

The skill's stated purpose (agent self‑evolution) matches most of its instructions, but it instructs the agent to modify workspace files, run git operations, read other skills, and notify all bots without declaring those runtime dependencies or fully limiting scope — this combination is concerning and should be reviewed before install.

Purpose & Capability

The skill claims to implement a self‑evolution/persistence protocol and includes instructions and a helper script that perform keyword detection, failure detection, generate approval cards, write to MEMORY.md/SKILL.md/CLAUDE.md, update evolution logs, and commit via git — these capabilities are coherent with the stated purpose. Minor mismatch: SKILL.md expects git operations and workspace filesystem access but the skill's metadata does not declare required binaries (e.g., git) or runtime permissions.

Instruction Scope

Runtime instructions direct the agent to read and write files under /root/.openclaw/workspace (MEMORY.md, SKILL.md for other skills, CLAUDE.md), run git commands (git log/git revert/git commit), call/consume a 'darwin-skill' for scoring (reading another skill's SKILL.md), notify all bots when CLAUDE.md changes, and maintain counters and logs. While these actions implement persistence, they broaden scope to system/workspace modification and cross‑skill interactions — including the ability to change global behavior (CLAUDE.md) which affects other bots. The SKILL.md attempts to forbid writing certain sensitive items, but the mechanics still permit high‑impact writes if approvals are given.

Install Mechanism

Instruction-only with an optional helper Python script using only standard library; no install spec or external downloads. This is low risk for supply-chain code fetching. However, the script assumes a filesystem and Python runtime and will be executed only if the agent chooses to run it.

Credentials

The skill requests no environment variables or external credentials. It does, however, read other skill files (e.g., darwin-skill/SKILL.md) and workspace paths — those are plausible for its purpose but are powerful capabilities that let it inspect other skills' metadata and the agent workspace.

Persistence & Privilege

The skill writes persistent files (MEMORY.md, SKILL.md, CLAUDE.md), updates an evolution counter and log, and instructs git commits and potential reverts. Writing CLAUDE.md and 'notifying allBot' are system‑wide actions that can change behavior of other bots. Although L3 writes require double confirmation in the spec, the skill can still gain impactful persistence when users approve. The skill is not forced always:true, but it can be invoked autonomously (model invocation enabled), increasing blast radius if misused.

What to consider before installing

Before installing: 1) Understand this skill will persist rules into workspace files and run git commits (MEMORY.md, SKILL.md, CLAUDE.md, evolution logs). Ensure you want those persistent changes and have backups. 2) Confirm that git and Python are available in your agent environment and that you accept the skill reading other skills' files (it references darwin-skill). 3) Be cautious about allowing L3/global changes — although SKILL.md requires double confirmation, approved L3 writes can change behavior for all bots. 4) If you want lower risk, disable autonomous invocation or require manual invocation/explicit confirmations for any writes to CLAUDE.md or other skills. 5) Test in a sandbox workspace first and review commits created by the skill to confirm it only writes expected content.

self-evolve.skill — Agent自进化机制

教一遍，记一辈子。用得越多，越懂你。

---

核心心智模型

模型1：遗忘曲线对抗 — Agent每次对话从零开始，进化机制就是"复习系统"。识别值得长期记住的经验→写入持久化文件→下次自动加载。失效条件：用户随口一说不是长期意图。

模型2：三层沉淀漏斗 — L1轻量偏好放记忆层，L2方法论放技能层，L3全局规则放规范层。越往上审批越严。失效条件：层级判断错误，不确定时宁可放L1等自然升级。

模型3：审批即训练 — 所有进化需用户审批。批准=正样本，忽略=负样本。频率控制：每日≤5次，每周≤15次，重要内容不受限。

---

进化层级协议

L1：对话记忆层 → MEMORY.md

• 写入条件：用户偏好、行为模式、经验碎片
• 审批：重要性≥medium，简单确认
• 格式：### 🧬 自进化 [来源] > 经验内容（一句话）
• Git：🧬 L1进化: <摘要>

L2：技能优化层 → SKILL.md

• 写入条件：踩坑经验、方法论改进、边界补充
• 审批：需Darwin评分≥70分（调用darwin-skill评估）
• Fallback：无法调用Darwin时，Agent干跑3个测试prompt自评，≥75分才允许
• 格式：### 🧬 踩坑经验 [日期] > 失败现象/根因/方案/失效条件
• Git：🧬 L2进化: <skill名> - <摘要>

L3：行为规范层 → CLAUDE.md

• 写入条件：全局行为约束、开发规范、安全边界
• 审批：重要性high + 双重确认（先确认层级，再确认内容）
• 格式：### 🧬 规范进化 [来源] > 规则/范围/生效时间
• 额外：通知所有Bot「CLAUDE.md已更新」
• Git：🧬 L3进化: <摘要>

---

触发机制

触发器A：关键词识别

级别	关键词	行为
紧急	以后、永远、每次、必须、一定、禁止	立即弹出进化卡片
标准	记住、下次、别再、不要再、记得	提示进化机会
温和	倾向、偏好、喜欢、习惯	记录候选池，批量审批

误触发保护：引用中/否定句/疑问句不触发或降级。

混合意图协议：检测临时状态+长期意图混合时，降级处理，标注"⚠️ 受临时情绪影响"。

语义模糊追问：意图不够具体时追问澄清，如"少说点开心的"→具体指减少emoji/正能量/闲聊？

触发器B：踩坑检测

指标	阈值
工具调用次数	>5次/对话
同操作重试	>3次
用户纠正	>2次
Git回滚	任何1次
操作超时	任何1次

时机：超阈值时内部标记，对话自然结束后统一处理。根因需用户确认后写入。

触发器C：定期复盘

每日HEARTBEAT/22:00任务后/用户主动请求时触发。读取记忆→识别重复模式→评估价值→过滤噪音→批量审批。

---

进化请求卡片（唯一标准格式）

🧬 进化请求 #<编号>

触发：<A/B/C> | 级别：<紧急/标准/温和>
建议记忆：> <一句话精炼表述>
⚠️ <情绪影响/意图模糊/根因待确认>（如有）
目标：L<1/2/3> → <文件路径>
评估：重要性<H/M/L> 复用性<H/M/L> 风险<无/低/中>
→ 「进化」|「进化 L2」|「忽略」|「修改」|「暂缓」

---

审批命令

用户回复	行为
进化	按建议层级写入+Git提交
进化 L1/L2/L3	覆盖层级后写入
忽略	跳过（负样本）
修改 <新内容>	替换后重新请求
暂缓	保留到下次复盘
确认根因	L2踩坑根因确认
修正根因 <新>	替换根因后继续

---

冲突检测与解决

写入前搜索现有规则：矛盾→用户选择保留哪个；重复→合并；层级冲突→建议升级。

---

安全边界

永远不能进化写入：安全约束、系统规则、核心身份、其他Bot专属规则。 内容限制：每条≤100字，不含代码/敏感信息。

诚实边界

1. 不是真学习（写规则不是改权重），规则越多边际效益递减
2. 噪音无法完全消除，需定期清理
3. 冲突可能漏检（依赖文本搜索）
4. L3风险最高，务必双重确认
5. 过度进化→审批疲劳→质量下降
6. L1/L3效果只能靠主观感受
7. 语义模糊依赖用户澄清
8. 根因分析可能错误

---

回答工作流

Step 1: 识别 → 扫描关键词/踩坑指标/复盘触发
Step 2: 提炼 → 去触发词，评估层级(L1/L2/L3)，噪音过滤
Step 3: 请求审批 → 弹出卡片，L3双重确认
Step 4: 冲突检测 → 搜索现有规则
Step 5: 执行写入 → L2需Darwin评分，L3通知所有Bot
Step 6: Git提交 → "🧬 L<n>进化: <摘要>"
Step 7: 日志记录 → evolution-log.md + 计数器更新

---

版本：v2.1.0 | 作者：Shike | MIT-0