An agent skill (Claude Code + Codex, same SKILL.md) that audits a repo
read-only and finds the fuzzy decisions — routing, triage, severity,
escalation — that could become a typed Jev question: Choice, Score, or Noul.
Output is a ranked, source-grounded report: file/line, current behavior, a
concrete question contract, a fallback, and what stays in code (math, auth,
ranking). It's conservative — "keep as code" and "zero candidates" are valid
findings. Includes a worked example and a stdlib validator in CI. Feedback
welcome on the suitability gates.
Output is a ranked, source-grounded report: file/line, current behavior, a concrete question contract, a fallback, and what stays in code (math, auth, ranking). It's conservative — "keep as code" and "zero candidates" are valid findings. Includes a worked example and a stdlib validator in CI. Feedback welcome on the suitability gates.