# Router benchmark arms

The router benchmark (`benchmark router`, `packages/edge-eval/src/router.js`) compares three ways an agent can
reach a skill on the same tasks: none, find-skills with keyword rewriting, and Edge.

- `find-skills/SKILL.md`: vercel-labs/skills `skills/find-skills/SKILL.md` at 773fb2c (2026-07-10), MIT, copied
  unchanged. The find-skills arm installs it with `keyword-rewriting.md` appended, so it gets the same
  "rewrite the task into search keywords" help Edge's `find_skill` asks for with its `keywords` field.
- Benchmark-specific skills are kept out of retrieval in both search arms: a local proxy in front of Edge drops them
  from `find_skill` results and refuses `use_skill` for them, and an `npx`/`skills` shim filters them from
  `skills find` output and refuses `skills add` for them. Every filtered or refused call is logged in the attempt's
  `retrieval.jsonl`.