Skill Router

Concept of the skill

What it is: skill-router is the request-time dispatch discipline for choosing which existing skill should handle an agent request, plus which adjacent skills should be co-loaded or excluded.

Mental model: Read the compiled manifest, score activation signals additively, filter by project fit, apply quality and specificity tiebreakers, exclude conflicting ownership, expand dependencies and verification partners, and report coverage gaps instead of guessing.

Why it exists: Skill libraries decay when ambiguous requests silently activate the wrong skill. Routing needs explicit evidence and evals so maintainers can see whether the library covers real requests.

What it is NOT: It is not a one-skill metadata audit, new-skill authoring guide, library-wide infrastructure health audit, or general debugging workflow. It is also not model routing (choosing which LLM handles the request) and not a workflow orchestrator (sequencing multi-step subagent runs) — it is the one-shot decision of which skill's methodology the agent should follow before execution begins.

Adjacent concepts: graph-audit checks one skill or manifest surface; skill-scaffold creates new skills; skill-infrastructure watches library-wide health and routing-miss patterns; eval-driven-development helps design routing evals; context-graph supplies the graph context the router consumes.

One-line analogy: The router is an air-traffic controller for skills: it assigns the request to the right runway, co-loads support, and says when no runway is fit.

Common misconception: A router is not a nearest-neighbor guesser; a no-match result is useful evidence, not a failure to be hidden.

Coverage