"Looks right" isn't a test. Every change gets three checks, in this order, before it's done.
1. Automated checks
Run the test suite, the type checker and the linter. Ask for failures pasted in full, not summarized.
2. A real runtime
Exercise the change the way a user would: open the page, call the endpoint, run the CLI. Many bugs only appear at runtime: layout overflow, a missing env var, a CSP violation.
3. Review against the map
Have the agent reread its diff against the map from the start of the session:
- Did it cross any connection that wasn't on the map? Why?
- For each connection crossed, are both sides consistent?
- What could it not verify?
For UI changes
- No horizontal scroll at 375px.
- Everything reachable and usable with the keyboard.
- Works with
prefers-reduced-motion. - Readable in light and dark mode.