AI coding agents are pumping out code faster than humans can review it according to recent analysis. Kun Cheng's new tool, no-mistakes, tries to fix that by putting an AI reviewer between you and your git remote. Instead of pushing to origin, you push to a local proxy. It spins up a disposable worktree, runs validation (rebase, AI review, tests, linting), and only forwards to upstream if everything passes. Clean PRs get opened automatically.
The tool is agent-agnostic. It supports Claude, Codex, Rovo Dev, and OpenCode. Developers stay in control: auto-fix issues or review them manually, an approach that prioritizes human oversight similar to Stage. Traditional AI code review happens after PR creation, in what Cheng calls the outer loop, and that introduces latency. Moving validation earlier, to the pre-push stage, shrinks the feedback cycle.
CodeRabbit and Graphite sit on the async side, analyzing diffs after you submit a PR. Cursor and GitHub Copilot catch issues inline while you write, but they can't run full test suites or simulate rebases. No-mistakes occupies the middle ground: local, isolated, and working with full repo context.
One Hacker News commenter suggested running such an agent as part of PR review with acceptance based on custom scoring. Another recommended a feedback loop from the review agent back to the coding agent, which makes sense if you want the whole pipeline to actually learn from its mistakes.