I let Claude Code write the change and Codex review the diff before anything lands. Not because one model is
"better", but because a second model reading the diff cold catches things the first one talked itself past.
Here's what that review actually catches, and the few rules around it that made the setup hold up on Windows and WSL.
1. Bugs that only show up when two requests overlap
The clearest example: Claude Code added a PATCH endpoint and its tests passed. Codex read the diff and flagged that a
DELETE arriving during the PATCH could make the PATCH return 200 with a null body. The tests never ran the two
requests together, so nothing failed. Claude took the finding back and asked before fixing it.
That's the pattern worth paying for: not style nits, but the case the author's own tests didn't think of.
2. "Done" that isn't evidence
Agents are good at saying a task is finished. The review is where "done" gets checked against what was asked. A rule
that helped: the reviewer runs the commands the change documents, as written, and a command that fails is a finding.
If the diff says "run npm test", the review runs npm test.
3. Things the agent should never have touched
Some changes are fine code in the wrong place. We keep a short list that never gets auto-approved, however sure the
agent sounds: rm -rf, anything that sends, edits outside the repo, and anything touching credentials or ~/.ssh.
And one rule that sounds silly until it matters: an agent never reads the config file that holds a key, not even to
"just check a key name". That harmless read is exactly how a secret ends up in a transcript.
4. Approvals the agent gave itself
The quiet failure in a multi-agent setup is one agent approving another agent's prompt to keep things moving. In our
setup an agent can't answer its own permission prompts or switch auto-approve on for itself; a person decides.
Two Windows/WSL lessons that cost us work first
- On my WSL setup /tmp is emptied on every boot. Claude Code had put its git worktrees there, so uncommitted work vanished on restart. Worktrees now live inside the repo, in a folder listed in .git/info/exclude.
- A hook only blocks the tool it's attached to. A PreToolUse hook on Edit won't stop the same change made with echo >> through the shell. Check the shell commands too, and test your hooks like code.
The setup
I run both agents side by side in SoluTerminal, a Windows app I build for exactly this: Claude Code and Codex in their
own terminal panes next to PowerShell and WSL, and the panes can message each other, so Claude can ask Codex for a
review without me copy-pasting. It's open source (Apache-2.0) and free during beta; you bring your own Claude Code or
Codex login. Windows 10/11 only, and the installer isn't code-signed yet, so SmartScreen warns.
https://soluterminal.com/?utm_source=devto&utm_campaign=oss-launch&utm_content=codex-review-article
What does your second reviewer catch that your first agent misses?
Top comments (0)