REFINE: The Self-Evolving Agent

REFINE is an adaptive skill engine for structured session diagnostics. Use this skill when a user explicitly requests: logging error patterns across sessions...

This skill appears internally consistent: it runs offline, uses only the standard library, and enforces code-level sanitization before writing to refine_memory.json. Before installing or using it, consider: 1) Do not pass raw prompts, secrets, PII, or credentials in the context fields — even though the sanitizer is robust, avoiding sensitive data is best practice. 2) PRO mode requires providing a raw SkillPay token to the skill caller (the example passes it via headers); the skill hashes and compares it offline and does not persist the token, but supplying secrets to any third-party code should be deliberate. 3) The skill synthesises system-prompt patches — ensure any patch is reviewed before being incorporated into your agent's system prompt or applied automatically. 4) The skill writes refine_memory.json to the working directory; confirm that location is acceptable for storing diagnostic metadata. If you want extra assurance, review the complete main.py (especially the PRO/patch synthesis functions) to confirm there are no hidden network calls or automatic patch-application behaviors.

Type: OpenClaw Skill Name: refine-agent Version: 1.0.3 The REFINE skill bundle is a diagnostic tool designed for local error logging and session memory management. The code in main.py implements robust security measures, including a comprehensive sanitization function (_sanitize_context) that redacts sensitive keys (e.g., tokens, passwords), rejects nested objects, and truncates strings before writing to the local refine_memory.json file. It performs no network calls, uses constant-time comparison for authentication (hmac.compare_digest), and limits data storage to scalar values, effectively preventing accidental data exfiltration or sensitive information leakage.