You describe the bug to Claude. Codex turns it into a test that fails, and Claude then works on the code until that test goes green. Claude never wrote the test, so it has nothing to quietly bend toward its own fix.
Let one model write both the fix and the test and the test ends up checking what the fix does. The bug you reported can still be there after a green run.
Tell Claude in plain words, with the steps that reproduce it.
Claude starts it. Codex adds the test and reports that it fails for the right reason.
It changes the code and leaves the test alone until the suite passes.
Codex runs it one more time to close the loop, then Claude closes the helper.
Behind it, one handoff each way. The first command puts Codex on the test. Its reply lands in Claude's session. Read the CLI reference
$ overvibe new --attach --tool codex --cwd . --name "web: failing test" \ --prompt "Write a test that reproduces: totals ignore discounts. Confirm it fails. When done, run: overvibe send \"$OVERVIBE_SESSION\" \"<test and failure>\"" # Claude fixes the code, then asks for the confirming run $ overvibe send "web: failing test" "Fix is in. Run the suite again and report."
Your instruction, plus the fact that you'd see it. Tell Claude the test file is off limits; any edit to it shows in the terminal and in git.
It should. The agents run whatever your project already uses (bun test, pytest, go test) in their own terminals.
Sure. Claude Code, Codex and agy can each take either side.
Claude drafts the advertorial, Codex goes after the hook and every claim, and you choose which version goes live.
From your phoneEvery Claude session starts with Remote Control on. From the Claude app, ask your coordinator which sessions need you.
Parallel features in worktreesTwo or three helpers build separate features at once, each in its own git worktree and branch, merged when they're done.
Install Overvibe and ask Claude to get a failing test out of Codex before it touches any code.
