had a stupid one last week.
an agent changed some pagination code in a sync job. tests passed, PR looked fine, then staging stopped at exactly 100 records because the cursor wasn't updating.
the funny part is i had the same agent review the change before merging it. it found nothing wrong with its own code. shocking.
since then i've stopped letting the writer review itself. now i open a separate minimax m3 session, drop in the diff plus whatever repo context it actually needs, and ask it to look for ways the change could fail.
first time i tried it, it spotted two null cases i'd missed too.
curious if anyone else does writer/reviewer with separate agents. do you also have the reviewer run tests, or just inspect the diff?
[link] [comments]