Jev Belay
A Claude Code Stop hook that reads the turn's transcript locally and spends one Jev call only when files changed without a passing check, so an unverified done claim is sent back to run the suite.
A Claude Code Stop hook that reads the turn's transcript locally and spends one Jev call only when files changed without a passing check, so an unverified done claim is sent back to run the suite.
The hook spends a request only when a turn changed files and no check has passed since. The state carries the task from the user's prompt, the final assistant message and a run summary with the file-change count and the check commands that ran with their pass or fail result. Four questions answer together: whether the message claims the work is done, whether it claims checks passed, whether running checks would apply, and a choice of complete, partial, blocked or other. Code blocks only above its 0.7 threshold with a decisive outcome pick.
Does the final message present the requested work as finished or working?
The state is the user's task, the final assistant message, and a run summary with the number of file changes and the check commands that ran and whether each passed. The transcript itself stays on the machine.
Does the final message claim that tests, a build or other checks were run and passed?
The state is the user's task, the final assistant message, and a run summary with the number of file changes and the check commands that ran and whether each passed. The transcript itself stays on the machine.
Would running the project's tests, build or lint be a meaningful way to check the work the task asks for?
The state is the user's task, the final assistant message, and a run summary with the number of file changes and the check commands that ran and whether each passed. The transcript itself stays on the machine.
What does the final message report about the task?
The state is the user's task, the final assistant message, and a run summary with the number of file changes and the check commands that ran and whether each passed. The transcript itself stays on the machine.
The Stop hook runs when Claude Code is about to end a turn and reads the session transcript locally.
With no file changes, or with a passing check after the last change, the turn ends without any model call.
The task, final message and run summary become the state for the four questions, with a bounded timeout.
A confident done or verified claim without a passing check sends the turn back with the reason, and any error or low confidence lets it end.
A clean run costs nothing because the local transcript already proves it, so the paid question is spent only where a false completion claim is possible. Presence of a passing check is evidence and absence of one is not, which is why the hook asks whether checks would even apply instead of treating silence as failure.