Checks on code agent writes

Coding agents write with me on my own projects. Every change they make follows the same path a change from a person does.

The agent opens a pull request. A check runs against it: types, lint, unit tests, and a build. Then a separate step merges it. The agent does not merge its own work, and a merge does not happen until the check is green.

The cost is one more step. The benefit is that a wrong change stops before it reaches the branch I build from, and the record of what passed is attached to the change itself. That matters more with a fast model than with a slow one, because a fast model produces more changes to check.