rsi-loop

What This Skill Does

The rsi-loop (Recursive Self-Improvement) skill is the core evolutionary engine for EvoClaw agents, providing a robust, four-stage pipeline designed to automate the process of learning from past failures and optimizing future performance. By implementing a systematic Observe → Analyze → Synthesize → Deploy workflow, this skill allows an agent to move beyond static logic and into dynamic self-correction. It tracks task outcomes, identifies recurring performance bottlenecks, generates targeted improvement proposals, and applies these changes directly to the agent's infrastructure.

Installation

To integrate this essential evolution primitive into your OpenClaw environment, execute the following command in your terminal:

clawhub install openclaw/skills/skills/bowen31337/rsi-loop

Ensure that your environment supports the necessary Python dependencies managed via uv for optimal performance.

Use Cases

This skill is ideal for agents operating in long-running, complex environments where performance stability is critical. Primary use cases include:

1. Performance Debugging: Automatically logging failed tasks and identifying underlying systemic issues like context loss or tool errors.
2. Routine Health Checks: Scheduling periodic self-improvement analysis cycles to review agent health scores and optimize memory usage.
3. Continuous Deployment of Fixes: Automating the creation and deployment of SOUL.md updates, routing fixes, or new skill generation based on historical performance data.
4. Adaptive Resource Management: Detecting patterns of inefficiency, such as rate-limit hits, and adjusting model usage or task routing accordingly.

Example Prompts

1. "I just finished that complex refactoring task; log the outcome as a success with quality 5 and analyze if my current approach to file-ops can be optimized."
2. "Run the full rsi-loop cycle to check my health score and generate any pending improvement proposals based on the last 7 days of tasks."
3. "Review the latest improvement proposals and list those waiting for approval so I can deploy the ones that improve my code_debug performance."

Tips & Limitations

• Honesty is Key: When logging outcomes, provide accurate quality scores (1–5) to ensure the analyzer generates high-impact, meaningful patterns.
• Review Before Deploy: While the system supports full automation, always perform a --dry-run or manual review of proposals before pushing changes to your core system configuration.
• Data Integrity: Keep the skills/rsi-loop/data/ directory clean to ensure that the patterns.json output remains accurate and relevant to current agent behavior.