Librarian — Skill Discovery & Routing

Routes user intent to the right skill via a six-tier waterfall: EXACT (1.00) → PREFIX (0.85) → SUBSTRING (0.65) → TAG (0.50) → SEMANTIC (Levenshtein, 0.35) → NONE. Never auto-loads when confidence < 0.7 — asks instead. Multi-skill tasks load in dependency order. Tags drawn from the closed taxonomy in docs/tags.md. The librarian has no side effects; it only reads and routes.

Discovery is layered, so token cost stays flat as the garden grows to thousands of skills. Never scan every skill. Walk: map → domain index → category index → skill. Load skills/_index/MAP.md (the map — a tiny list of domains only, so it never bloats) and pick the domain. Open that domain's index skills/_index/<domain>/INDEX.md: for a flat domain (e.g. foundation) it lists the skills directly; otherwise it lists the categories — pick one and load that category index skills/_index/<domain>/<category>.md. Run the waterfall against that small table, then read the 1–3 winning SKILL.md files. Every load is bounded; none grow with the total skill count.