I built this because every PR review starts with the same five comments. Missing null checks, off-by-one indexing, leaked secrets, untyped any. The boring stuff a tired human reviewer skims past at 5pm. PR Diff Review is a small Claude API service that reads the diff, leaves the obvious comments, and gets out of the way of anything actually subtle.
The bot is a GitHub App. It reads a diff on PR open and posts comments inline on the lines they apply to. No batched summary at the bottom, no separate dashboard.
If a comment is uncertain it gets a "maybe" tag. If it's high-confidence it's posted plain. You can ignore or react.
Architecture decisions, naming conventions, taste calls — those are left alone on purpose. Humans should make those calls. The bot's job is to catch the things humans aren't actually good at: tedium and pattern-matching at scale.
Static prompt + tool definitions are cached. A typical PR review is under a cent. Long PRs are chunked and reviewed in parallel so latency is in the seconds, not minutes.