Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
// Routing eval fixtures for skills/skillify. Each positive intent includes
// at least one frontmatter trigger string as substring (structural matcher
// requirement) while paraphrasing real user phrasing.
{"intent":"Skillify this workflow end to end — full checklist, eval contract, and resolver row","expected_skill":"skillify"}
{"intent":"Check skill completeness on this workflow before we call it done","expected_skill":"skillify"}
{"intent":"Did this skill regress after my edit? Compare the fresh score against the prior baseline","expected_skill":"skillify"}
// Ambiguous vs skill-creator: creating the skill is skill-creator's lane;
// taking a draft through the checklist + evals is skillify's. This phrasing
// legitimately trips both.
{"intent":"Skillify this new skill draft — take it through the whole checklist and evals","expected_skill":"skillify","ambiguous_with":["skill-creator"]}
// Negative: adjacent (skills as a topic) but out of scope — a bundling
// question, not a make-this-a-proper-skill request.
{"intent":"Which of our skills are bundled into the plugin for downstream installs?","expected_skill":null}