The AI terminal that shows its work and lets you undo it

Something I keep seeing across every AI coding community: an agent gets told not to touch something, then touches it anyway. it just never had a wall there, only a request in its instructions it decided didn't apply this time. Prod env vars gone, a repo gone, a folder that took years to build, empty. The model agreed not to do it and did it, because agreeing and being unable to are different things, and only one of them survives a bad turn. I don't think the fix is a smarter model. I think it's a layer underneath the model that doesn't ask it nicely. That's what Verlox is https://www.verlox.app. A CLI and agentic AI terminal where every plan gets written out before it runs: the command, the reason , a risk level. Nothing executes until you approve it. Anything it deletes gets captured first, so it's one click to restore instead of a race against a redeploy timer. Permissions live in the terminal, not in a prompt the model can talk itself out of. Still Windows only, still solo. If you've had one of these moments, I'd genuinely like to hear it.