jev-belay

How we reviewed this: we drove the hook in Docker against its own fake backend, ran its tests (69 passing, 5 skipped for want of a live key) and inspected a captured request. We made no Jev calls.

What it does

jev-belay, by valentynkit, hooks the end of a Claude Code turn. It reads the transcript locally first, and only asks anything when the shape is suspicious — files were changed, no check passed afterwards, and the closing message claims completion. The author reports that this means most stops never reach the model at all.

When it does ask, four questions go out: does the message claim done, does it claim verification, does verification apply here, and what was the outcome.

The redaction is real — fourteen secret patterns, and $HOME rewritten to ~ — and there’s a corpus plus a measurement tool in the repository, along with a “known limits” section that says which numbers haven’t been measured yet.

One correction to the README

It says what leaves the machine is “your prompt, the closing message, two counts. No tool inputs, no diffs”. In fact the state includes the failed check commands themselves — redacted and truncated, but real command strings, not counts. Worth knowing if your test commands carry arguments you’d rather not send.

Verdict

The design is right: evidence gathered locally, a call only when it matters, a nudge rather than a block. For a fifteen-commit project the discipline is notable.

For related gates, see Abide and pi-warden.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 20, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.