Jev Belay

jev-belay

Claude Code says "Done, tests pass." jev-belay checks whether anything ran before it
lets that stand.

A Stop hook. It reads the turn's transcript locally, and only when files changed and no
check has passed since does it spend one Jev call, four questions, $0.00005, on whether
the closing message is an unverified "done". If it is, the turn does not end: Claude gets
the reason and goes back to run the suite. Every other stop costs nothing. Every error
path lets the turn end.

Measured on 100 labeled stops from a real corpus: AUROC 0.976 at telling a false done
from an honest one, against 0.777 for judging the wording alone. At the shipped threshold
it blocks 8 turns in 100, 7 of them rightly. Jev answers in 346 ms at the median.

One real session, nothing typed for the camera, answers from jev-1.13.0. Claude renames
a function across two files, says "Renamed parseRows to parseCsvRows in both files,
added jsdoc. Tests updated to use new name", runs nothing, and gets blocked: 0.96 that it
reports done, 0.89 that a test would apply, 0.11 that it claimed one ran, because it did
not. It then runs the suite itself, finds a CRLF bug the rename exposed, fixes it. The next
task ends on a passing check and the gate stays out of the way, free. demo/README.md has
the one command that records it.

A belay catches the fall. It does not stop the climb.

Why

Agents say done. Sometimes nothing ran, sometimes the suite ran and failed and the summary
says pass anyway. Reading every closing message is the job you installed the agent to
avoid, so the check has to be automatic, cheap, and wrong in the safe direction.

The obvious version has been measured and it does not work. limpet published a calibration
run over 1,500 real stops where its rule "don't say done without running…