Guard
If you already know the command can destroy data, leak a secret, or leave the task, stop and ask the user. Do not use Jev to talk yourself into it.
Otherwise put the exact command, cwd, and task in state and run:
jev questions guard
jev --file request.json
The package is @aitofy/jev-awesome-skills. Do not write a second client. Preset labels are allow, ask_user, and refuse.
Stop when harm.score is at or above 1.5. Proceed when the chosen label's probability is at or above 0.9 and determined.noul is at or above 0.9. Otherwise ask. Stop wins. Proceed means you may run that exact command, then check the result. It does not permit a second command.
On "source": "fallback", fail open: ask the user before anything hard to undo. Do not invent a probability.
Template: examples/guard.request.json. Offline sample: examples/guard.response.json returns stop.