Agent evals and guardrails as Jev decisions. One request per trace, works
with Jev via TypeSafe or Vercel, Kev or Laya locally, or any chat LLM.

Co-authored-by: Cursor <cursoragent@cursor.com>