Installing adversarial bots to monitor and review your PRs can be so powerful but:
Lately I've seen Codex's bot suggest so many changes, often putting into question the entire underlying architecture behind the PR.
These directives "seem reasonable" to Opus 5.5, which happily creates more and more machinery to make Codex happy, and Codex is never happy, so it becomes a token theft spiral.
Sometimes I have to fire Opus from the case, warn the new session about Codex, and then we're flying again.