When AI Writes the Diff, Code Review Needs to Record Understanding
AI agents make it cheap to produce a large diff. They do not make it cheap for a team to understand that diff six months later.
A green check and an approval tell you that a change passed a process. They do not tell you whether anyone besides the author can explain the files it changed.
Closing the gap at the file level
A lightweight way to close that gap:
- Track at the file level. Ask which changed files have a current reviewer who can explain them, and which depend on one person.
- Tie evidence to a version. Understanding recorded before a file changes can go stale. Recheck after meaningful edits, and expire old evidence.
- Teach while reviewing. A short explanation of what changed, why, and a few questions can make a PR review a learning loop.
- Use the signal to prioritize. A file with no current understanding or only one knowledgeable teammate deserves more attention than a routine, well-understood change.
- Measure codebase risk, not people. The useful question is βwhich files are nobody ready to maintain?β not βwhich developer scored lowest?β
Privacy and review checks
For privacy, teams can keep source code on developersβ machines and sync only file paths and review results. A pull-request check can then flag changed files with no current understanding evidence.
Iβm building Zevqi around this file-level approach. Its CLI is open source under MIT: github.com/zevqi/zevqi.
Comments
No comments yet. Start the discussion.