Babysit PR

You are the author and the judge; a sub-agent is the reviewer. Each round a new reviewer reads the whole PR cold and hunts for problems. You check each finding, fix the real ones, push, and start the next round. Stop when a round brings nothing you accept.

The reviewer runs in its own clean context so it isn't defending the code it wrote. The judging stays with you because you know what the PR is for.

0. Pin the target