Claude writes. Your tests check.
Codex reviews.

One command runs the whole loop: Claude writes the change, your tests check it, Codex reviews it, and the writer gets a repair round if the review blocks. You get a verified patch. Your branch stays untouched until you apply it.

Quickstart

npx github:inseat-labs/fusion run "Fix slugify so all tests pass" --verify "npm test"
npx github:inseat-labs/fusion apply latest

View on GitHub

Real output from a run on 30 September 2026.

$ npx github:inseat-labs/fusion run "Fix slugify so all tests pass" --verify "npm test"
fusion [0:00] leg-0 running   writer claude: write
fusion [0:41] leg-0 succeeded writer claude succeeded in 41s
fusion [0:41] leg-1 running   verify: npm test
fusion [0:42] leg-1 succeeded verify PASS (exit 0)
fusion [0:42] leg-2 running   reviewer codex: review (read-only)
fusion [1:28] leg-2 succeeded reviewer codex succeeded in 46s
fusion [1:28] review: APPROVE (0 blocking, 0 suggestion(s))

Status    READY: verify passed, codex approved. Your branch is untouched.
Writer    claude  write succeeded 41s
Verify    PASS  npm test
Reviewer  codex  APPROVE  0 blocking, 0 suggestion(s)
Repairs   0 of 1 used
Changed   1 file(s), +6 -1
Duration  1m 28s  (CLI-reported cost estimate: $0.2302; other legs report no cost)

What happened: A small repo had a failing test: slugify didn’t strip accents. Claude fixed it in a separate copy of the repo, npm test passed, and Codex reviewed the diff and approved. Nothing touched the real branch until apply.

How it works

1. WriteFusion makes an isolated git worktree, a separate copy of your repo at the last commit. The writer, Claude or Codex, makes the change there.
2. VerifyYour check command runs on that copy, for example npm test. Fusion can pick npm test automatically.
3. ReviewThe other vendor’s agent reads the diff with read-only access and answers approve or changes, listing any blocking findings.
4. Repair, then applyBlocking findings go back to the writer for a repair round (up to two). When the patch is ready, apply puts it in your working tree. list and discard manage runs; stats shows which agent’s patches pass your tests and get approved in your repo.

How it compares

Fusion next to other ways of pairing coding agents
Tool	What it does	Who runs the checks
openai/codex-plugin-cc	Lets you call a Codex review from inside Claude Code	You, by hand, when you ask for it
claude-squad, vibe-kanban	Run several agent sessions side by side	You review each session’s result yourself
Fusion	Automates write, test, review and repair for one task, and keeps a per-repo record	Your tests and the other vendor’s agent, every run

Limitations

One writer at a time.
It starts from your last commit; uncommitted changes are not included.
No budget cap yet. The cost shown is whatever each CLI reports.
The reviewer’s verdict is parsed from the JSON it returns.
Early (0.x) and not on npm yet. Run it with npx from GitHub.

Help shape it

Issues, run reports from your own repos, and pull requests are welcome.

Fusion roadmap ↗Contribute to Fusion ↗Maintainer: abenezer88 ↗abenezer@inseat.app ↗

Changing AI models?

Switch runs your real prompts on both models and shows which tool-calling behavior changed.

Explore Switch