When Claude Code and Codex disagree, do not ask which one is right. Ask the reviewer for a concrete failing case and run it. If it fails, the author fixes it. If it passes, the finding was wrong. Edit collisions are a different problem with a different fix: give every task one owner.
Two kinds of conflict
1. They disagree about the code
A reviewer says "this is a bug" and the author says "it is fine". A judgment call does not settle that; a test does.
- Ask the reviewer for an input that fails, with the expected and actual result.
- Run it yourself, or have the author run it.
- Failing: the author fixes and adds a test for that case. Passing: record why and move on.
In the run behind our tested setup, the Codex reviewer did exactly this. It named slugify("hello world", 6), which returned hello-. The case was checkable in a minute, and the author fixed it. Nobody had to decide whose opinion to trust.
2. Their edits collide
Two agents change the same files and git reports a conflict.
- Prevent it. One owner per task, split by task rather than by file, and separate checkouts. See how to split the work. In Crewly, per-task git worktrees are an opt-in setting per project.
- Resolve it. The owner of the task that landed second resolves the conflict, then reruns the tests. A conflict resolved without a test run has not been resolved.
A stopping rule
Our recommendation, not a Crewly feature: two rounds between author and reviewer, then stop. If they still disagree after two, the failing case has not been found, so you or your team lead decide. In our run, round one found the problem and round two approved the fix.
Limits
One small run on macOS, one planted bug, Claude Code writing and Codex reviewing. We have not seen the two tools disagree on a design question, so the advice above covers bugs and edit collisions only.
FAQ
What if Codex says there is a bug and Claude Code disagrees?
Ask the reviewer for an input that fails. If it fails when you run it, fix it. If it passes, the finding was wrong and you can merge. Do not let the two agents argue in circles.
What if their edits conflict in git?
Prevent it with one owner per task and separate checkouts. If it still happens, the task's owner resolves it and reruns the tests before the commit counts.
Who decides when they cannot agree?
You do, or the lead you assigned. This is our recommendation, not a Crewly rule: after two rounds, stop and decide.