moonlight-lupin

BM25-based skill retrieval — injects top-K relevant skill descriptions per turn via a pre_llm_call hook, replacing the full skill list in the system prompt. Token optimization for large skill libraries. No API keys; runs locally, no network. Disclosure — wraps Hermes' build_skills_system_prompt and clear_skills_system_prompt_cache in-process at session start so the system prompt lists skill names only, and appends a short BM25-ranked list of skill descriptions to each user message via the pre_llm_call hook (the system prompt is built once per session, so prompt caching is unaffected). Reads SKILL.md frontmatter from the profile's skill roots; no network access, no files written, pyyaml is the only dependency.