Skill Retriever Brings 10K-Category Semantic Skill Discovery to Hermes

Skill Retriever is an AgentSkillOS-powered plugin that adds semantic skill discovery to Hermes Agent. It organizes over 1,200 skills — 998 from a community corpus plus 211 Hermes-native ones — into a 10,000-category capability taxonomy, pre-filtering each query down to the top five most relevant skills. Because it runs through Hermes's pre_llm_call hook, it requires no core modification and adds no extra API cost, reusing the existing Hermes LLM connection in borrow-mode.

The project's core argument is that pure embedding-based semantic search is narrow: it favors textual similarity and can miss skills that are functionally important but look unrelated in embedding space. Skill Retriever's LLM-guided tree navigation walks the capability hierarchy instead, surfacing non-obvious but relevant skills — a distinction that matters once the corpus grows large enough to bury the right answer in noise.

Compared to Hermes's built-in discovery, which lists every installed skill flatly in the system prompt every turn (and starts bloating past roughly 200 skills), Skill Retriever behaves like a search engine: it injects hints into the user message only when a match is found, avoiding constant system-prompt overhead and scaling to 10,000+ skills at the cost of a few extra cheap LLM calls per turn for tree traversal.

On the safety side, every skill carries a source tag (hermes/community) and a scan result; all 1,200 skills were checked for dangerous patterns with none flagged. The project is MIT-licensed, requires Hermes Agent v0.18+ and Python 3.10+, and ships a CLI for searching, rebuilding the tree, and inspecting system stats.