Skill

Self-Improvement System for AI agents. Features: feedback collection, hint extraction, rule learning. Use for: learning from user corrections.

Security Scan

The skill's declared purpose (agent self-improvement) matches the files and instructions, but it relies on installing an external Python package (claw-rl) and auto-injecting learned rules into sessions — both of which create non-trivial risk that isn't verifiable from the shipped files alone.

Purpose & Capability

Name, description, scripts, and docs consistently describe a self-improvement RL system that collects feedback, extracts hints, stores learned rules, and injects them into sessions. Required binary (python3) and default workspace paths align with this purpose.

Instruction Scope

SKILL.md and scripts instruct the agent to install an external package (pip install claw-rl), run CLI helpers, start/stop a learning daemon, queue feedback, and auto-inject rules into sessions via OpenClaw config. The instructions do not directly read unrelated system files or require secrets, but autoInject/autoLearn behavior means the skill can alter agent behavior globally if enabled.

Install Mechanism

There is no curated install spec; the docs ask you to pip install 'claw-rl' from PyPI (or another index). Relying on an external package means arbitrary code will run that is not present in the skill bundle. That is a moderate-to-high risk unless the package source is audited or pinned to a known good release.

Credentials

The skill requests no environment variables, credentials, or config paths beyond a workspace directory (~/.openclaw/workspace). The lack of secret requirements is appropriate for a local feedback-and-learning plugin.

Persistence & Privilege

The skill does not set always:true and is user-invocable only, but the recommended OpenClaw config enables autoInject and autoLearn which give it ongoing influence over sessions (it may inject learned rules automatically). Starting a background learning daemon implies persistent background activity and writes to user workspace (SQLite DBs).

What to consider before installing

This skill appears to do what it says (collect feedback, learn rules, inject them), but most core logic is provided by an external Python package named 'claw-rl' — not by the scripts included here. Before installing or enabling autoInject/autoLearn you should: (1) verify the provenance of the 'claw-rl' package (check the GitHub repo, release tags, and PyPI owner), (2) audit the package for network calls, telemetry, or credential access, (3) avoid enabling autoInject in production until you trust the rules it will inject, (4) run the daemon and package in an isolated or sandboxed environment and back up ~/.openclaw/workspace, and (5) prefer a pinned package version or vendored, reviewed code if possible. If you cannot audit the external package, treat enabling this skill as higher risk.

claw-rl: Self-Improvement System for AI agents

Self-improvement system for AI agents with reinforcement learning and continuous learning.

Prerequisites

pip install claw-rl

Quick Start

Collect Feedback

python3 {baseDir}/scripts/collect_feedback.py "Great job!" --action file_created
python3 {baseDir}/scripts/collect_feedback.py "Wrong, use Chinese instead" --action response --negative

Get Learned Rules

python3 {baseDir}/scripts/get_rules.py --top-k 10
python3 {baseDir}/scripts/get_rules.py --context "user preference"

Check Learning Status

python3 {baseDir}/scripts/status.py

Start Learning Daemon

python3 {baseDir}/scripts/daemon.py start
python3 {baseDir}/scripts/daemon.py stop
python3 {baseDir}/scripts/daemon.py status

Core Components

Component	Purpose	Example
Binary RL Judge	Evaluate satisfaction from feedback	👍 → positive, 👎 → negative
OPD Extractor	Extract improvement hints	"Use Chinese" → rule hint
Learning Loop	Continuous background learning	Process feedback queue
MAB Strategy	Strategy selection via bandits	Thompson Sampling, ε-greedy

Learning Modes

Mode	Description	Use Case
calibration	Calibration learning	User satisfaction calibration
strategy	Strategy learning	Action selection optimization
value	Value preference learning	User preference learning
context	Context-aware learning	Situational rules

Configuration

OpenClaw config (openclaw.config.json):

{
  "plugins": {
    "slots": {
      "context-engine": "claw-rl"
    },
    "claw-rl": {
      "config": {
        "workspaceDir": "~/.openclaw/workspace",
        "autoInject": true,
        "autoLearn": true,
        "topK": 10
      }
    }
  }
}

Performance

Operation	Latency
Initialize	~2ms
Collect feedback	~0.03ms
Extract hint	~1ms
Get rules	~0.7ms
Process learning	~0.5ms

Advanced

See references for detailed documentation:

• Architecture - System design
• Learning - Learning algorithms
• MAB - Multi-Armed Bandit strategies