REFINE: The Self-Evolving Agent
v1.0.3REFINE is an adaptive skill engine for structured session diagnostics. Use this skill when a user explicitly requests: logging error patterns across sessions...
⭐ 0· 98·0 current·0 all-time
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
Name/description, SKILL.md, skill.yaml and main.py align: the skill captures sanitized feedback and errors locally and (optionally) runs an offline PRO-mode patch synthesis flow. The only optional secret (SKILLPAY_TOKEN_HASH) and REFINE_MODE map directly to the described PRO mode and are proportionate to the stated purpose.
Instruction Scope
SKILL.md and code limit operations to local disk writes (refine_memory.json) and sanitization. The skill can synthesise 'System Prompt Patches' from local analysis — this is consistent with its purpose but is a behavioral risk: any generated patch that an operator or agent applies could change agent behavior. SKILL.md advises activation only on explicit requests, which mitigates scope creep.
Install Mechanism
No install spec; main.py is standard-library-only and skill.yaml lists no external dependencies. No downloads or external package installs are required, which is low-risk and proportional.
Credentials
All environment variables are optional. SKILLPAY_TOKEN_HASH (secret) and REFINE_MODE are justified by the PRO mode offline verification flow. The skill does not require unrelated credentials or broad environment access.
Persistence & Privilege
The skill persists sanitized data to a local file (refine_memory.json) and is not marked always:true. That persistence is expected for a diagnostics tool, but users should be aware data is written to the agent's working directory. Also note the skill can produce system-prompt patches (stored locally) — review before applying to agent/system prompts.
Assessment
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.Like a lobster shell, security has layers — review code before you run it.
ai-learningvk970yfxh3kky1n3cgghts39ww183h3mblatestvk970yfxh3kky1n3cgghts39ww183h3mbproductivityvk970yfxh3kky1n3cgghts39ww183h3mbself-improvementvk970yfxh3kky1n3cgghts39ww183h3mb
License
MIT-0
Free to use, modify, and redistribute. No attribution required.