Skill Retriever

Skill Retriever is a semantic skill-retrieval plugin for Hermes Agent that pre-filters a corpus of 1,200+ skills (998 community-contributed from AgentSkillOS plus 211 Hermes-native) down to the top 5 most relevant per user query. Rather than using vector embeddings, it navigates a pre-built LLM-driven capability taxonomy — a tree of ~10,000 categories — selecting relevant branches via recursive LLM calls at each level. Results are injected as natural-language hints into the user message before it reaches the model.

The central architectural bet is LLM-as-classifier instead of embeddings. Vector similarity can only surface skills that are semantically close in embedding space; the tree-based approach applies reasoning at each decision point, finding connections that a pure similarity search would miss. This comes at the cost of latency (1–3 extra LLM calls per query) and staleness (the tree must be rebuilt to include new skills), but scales to 10K+ skills where a flat system prompt would be unusable.

The search algorithm is a recursive tree descent with three key optimizations: auto-expansion when a node has few children (avoiding unnecessary LLM calls), early-stopping when a single selected branch contains few enough skills to collect wholesale, and parallel search of sibling branches via a thread pool. A final pruning stage uses a workflow-stage model (upstream → production → downstream) to deduplicate and rank results — reasoning about what a complete task pipeline needs rather than just scoring individual skills for relevance.

Tree construction happens offline in two phases: an LLM assigns all skills to five hardcoded root categories (content-creation, data-processing, development, automation, domain-specific), then recursively splits oversized groups. The builder uses a queue-based parallel pattern — FIRST_COMPLETED with a thread pool — so new sub-groups enter the queue as soon as any split finishes. Graceful degradation handles LLM failures: if splitting fails, the node becomes a leaf with all its skills rather than crashing.

The plugin layer is minimal (185 lines). It runs as a Hermes pre_llm_call hook, lazy-loads a searcher singleton on first use, skips messages under 10 characters, and formats results with safety badges (🔒hermes for trusted installed skills, 🌐community for unreviewed AgentSkillOS skills, ⚠️ for safety-flagged ones). Credential configuration cascades from dedicated environment variables through OpenAI defaults, so Hermes users need zero additional setup.

The codebase is ~4,400 lines of Python 3.10+, MIT-licensed, with 40 tests across 4 files. It ships with a CLI offering search, build, list, and info subcommands.