Jev in production › Agent guardrails and approvals
Claude Code Stop hook that asks Jev whether a turn ended without acting, broke a declared promise, promised unwatched monitoring or used the wrong language, then blocks the exit with a nudge. 81.8% precision at full payload on labelled turns (author).
A Claude Code Stop hook for turns that stop half-done. It catches three premature exit patterns — stopping without taking action, declaring an action without calling the tools to back it up, and claiming to monitor background work while nothing is running — and intercepts the exit with a nudge so the turn does not close on an empty promise.
Stingrays sit motionless on sand until stepped on. This hook operates the same way: silent during normal operation, intervening only when a turn stops prematurely. Every judgement goes to Jev: whether the turn did nothing, broke a promise, promised to watch something, or answered in the wrong language. The one thing counted locally is what is running in the background.
For the project's own README, linking back here: