Coding & Vibe Coding

Vibe coding gets a feature onto the screen before you have a theory of the change. The model writes the route, the component, and the migration, and you are left holding a diff that looks finished. This shelf is the pause before you push. It is for people who already have Cursor, Claude, or ChatGPT open next to a branch, and who want a second pass that talks about behavior instead of vibes.

The agents here take a real artifact from the session: a unified diff, the prompt you used to generate the feature, or a .env.example beside the code that reads it. A risk review names the behavior that changed and the check that would catch a regression. A test plan lists the cases the original prompt forgot, including the empty state and the person who is not the author. An environment audit compares the keys the code expects with the keys the example file admits, and it flags names that look like live secrets.

None of these agents replace a reviewer who knows the system, and none of them teach you how to attack someone else's. The security checker on this shelf refuses exploit steps on purpose. Paste the smallest input that still shows the behavior. If the model invents a line that was not in your diff, tell it to quote the hunk or drop the finding.