Karasu tunes AI code review for agent PRs with risk tiers and a local-first gate

The third weekly update on karasu, a text modeling language for system architecture, covers 50 merged PRs, 11 of them from Dependabot, and six decisions recorded as ADRs. The author describes using path-based risk tiers to decide how deeply a change is reviewed, with ADR and test-only changes merging automatically after approval. Pull requests leaving draft must first pass a local review and a CodeRabbit CLI run, because written process steps were not reliably followed and were backed by hooks instead. The AI review's effectiveness is measured with an eval of 19 past bugs, which rose from 5 to 7 caught after guideline revisions, though the author notes this figure is generous since the same cases were used for tuning. The update also reports a fix to karasu serve, which previously could not render models that import other files, and a new file-serving approach that sends the whole project directory in one response.