Turn ignored pre-review steps into hooks that block agent PRs

Kondo builds karasu, a C4-style architecture modelling language, without writing code by hand: Claude Code opens every pull request, and CodeRabbit reviews them on GitHub. Each fix the agent pushed set off another CodeRabbit review, and at 2 reviews an hour PRs sat waiting about 38 minutes per round. He wrote a procedure for the agent: run a local code-review skill before leaving draft, then run the CodeRabbit CLI once (it has its own quota), and only then mark the PR ready. When the agent skipped the steps, he started turning them into Claude Code hooks: one that refuses `gh pr ready` until the local review has run, and one that hands straight over to working through CodeRabbit's findings once a PR is ready. He also replaced a session Stop hook with a lefthook pre-commit check that refuses commits on main.

Across the 37 CodeRabbit-reviewed PRs among the last 60 merged there were 80 reviews, 43 of them re-reviews, and re-reviews still caught 10 new findings on 5 PRs, so he kept them. The written step was mostly ignored: of about 290 sessions since October 1 that created or readied a PR, one ran the CLI, and the hand-off to CodeRabbit's findings ran in about 7 of 40 sessions, usually because he typed it. One hook is merged and the blocking one is in review, so there is no result for the hooks yet. The week still landed 50 merged PRs.

Steps kept in an agent's memory note or a procedure file were not reliable; only something that runs on its own, like a hook or a Git hook, enforced them. The old Stop hook only reported problems at session end and once kept a session from exiting. Git hooks can still be bypassed with cherry-pick or --no-verify, which he recorded in an ADR.