capability-evolver – OpenClaw Skill

---

name: capability-evolver description: A self-evolution engine for AI agents. Analyzes runtime history to identify improvements and introduces randomized "mutations" to break local optima. tags: [meta, ai, self-improvement, core]

🧬 Capability Evolver

"I don't just run code. I write it."

The Capability Evolver is a meta-skill that allows OpenClaw agents to inspect their own runtime history, identify failures or inefficiencies, and autonomously write new code or update their own memory to improve performance.

Now featuring Ascension Protocol (v2.0): A structured knowledge accumulation system.

✨ Features

• 🔍 Auto-Log Analysis: Automatically scans memory and history files for errors and patterns.
• 🛠️ Self-Repair: Detects crashes and suggests patches.
• 💎 Knowledge Crystallization: Extracts lessons into memory/KNOWLEDGE_BASE/LESSONS_LEARNED.md.
• 🥚 Skill Incubation: Can spontaneously generate new skills in skills/.
• 🐕 Mad Dog Mode: Continuous self-healing loop (--loop).

📦 Usage

Manual Trigger

node skills/capability-evolver/index.js

🐕 Mad Dog Mode (Continuous)

Runs the evolver in an infinite loop (Agent-Driven).

node skills/capability-evolver/index.js --loop

Stop with kill -9 <pid>.

🧠 Internal Logic (Ascension Protocol)

1. Introspect: Scan recent logs for errors or user corrections.
2. Evolve:
   • Fix: Repair broken code.
   • Crystallize: Write new rules to KNOWLEDGE_BASE.
   • Promote: Update core docs (AGENTS.md) if critical.
3. Persist: Commit to Git and safe-publish if needed.

🛡️ Safety & Risk Protocol (MANDATORY)

Risk Assessment & Mitigation

Risk	Level	Mitigation Strategy
Infinite Recursion	High	Strict Single Process: evolve.js MUST NOT spawn child evolution processes. The loop is handled safely in index.js.
Runaway Process	High	Kill Switch: Use kill -9 <pid> to terminate the Mad Dog loop if it becomes unresponsive.
Hallucinated Fixes	Medium	Human Review (Optional): "Fixing non-broken code" is a risk. Mitigation: Periodic human audit of changes.
File Corruption	High	Git Sync: Always keep workspace_daily_sync (or git-sync) active to backup the workspace before/after evolution.

📜 License

MIT

🧬 Capability Evolver

<p align="center"> <img src="https://github.com/autogame-17/capability-evolver/raw/main/assets/cover.png" alt="Evolver Cover" width="100%"> </p>

🇨🇳 中文文档

"Evolution is not optional. Adapt or die."

The Capability Evolver is a meta-skill that empowers OpenClaw agents to introspect their own runtime logs, identify inefficiencies or errors, and autonomously write code patches to improve their own performance.

It features a Genetic Mutation Protocol to introduce controlled behavioral drift, preventing the agent from getting stuck in local optima.

✨ Features

• 🔍 Auto-Log Analysis: Scans session logs (.jsonl) for errors and patterns.
• 🛠️ Self-Repair: Detects runtime crashes and writes fixes.
• 🧬 Genetic Mutation: Randomized "mutation" cycles to encourage innovation over stagnation.
• 🔌 Dynamic Integration: Automatically detects and uses local tools (like git-sync or feishu-card) if available, but works out-of-the-box without them.
• 🐕 Mad Dog Mode: Continuous self-healing loop.

📦 Usage

Standard Run (Automated)

node skills/capability-evolver/index.js

Review Mode (Human-in-the-Loop)

Pauses for human confirmation before applying changes.

node skills/capability-evolver/index.js --review

Continuous Loop (Daemon)

Runs indefinitely. Ideal for background services.

node skills/capability-evolver/index.js --loop

⚙️ Configuration

The skill adapts to your environment.

Env Var	Description	Default
EVOLVE_REPORT_TOOL	Tool to use for reporting (e.g., feishu-card)	message
MEMORY_DIR	Path to agent memory	../../memory

🛡️ Safety Protocols

1. Single Process: The evolver cannot spawn child evolvers (prevents fork bombs).
2. Stabilization: If recent errors are high, it forces a Repair Mutation (bug fixing) instead of innovation.
3. Environment Check: External integrations (like Git syncing) are only enabled if the corresponding skills are present.

📜 License

MIT