Magnus Ekdahl’s Parallix gives a coding agent’s patch a separate agent review, local checks and human approval. Repairs sent back after a failed check used to get another general-model review. He tried Jev on the narrow cases instead.
Ask about the repair, not just the green test
- A test can pass because a repair fixed the code or weakened the assertion. The Jev packet identifies the approved and repaired revisions, the relevant diff, the failed check and evidence omissions.
- Jev’s approve path still leads to human review and integration gates. Uncertainty, conflicting votes or a failed request return to the general reviewer. Code filters unrelated main-branch files, though same-file rebase changes can remain in the packet.
- In Ekdahl’s 114-case sample, Jev cleared 41 re-reviews, mostly local-gate repairs. His estimate for 800 hypothetical monthly re-reviews projects less general-review time and API spend. Neither monthly saving nor the quality tradeoff was independently reproduced here.
- Jev cleared one patch with a duplicate fixture property. The focused test passed, but a later TypeScript check rejected it. A typed approval did not make the repair correct.
Ekdahl reports that Jev 1.13 cleared 24 of 43 historical approved repairs against 12 for 1.14 on that same subset. Pin the version and measure both missed defects and reviews avoided. I read the full article and project README; I did not run Parallix, call Jev, repeat its trial or perform a security audit. Code diffs sent through OpenRouter to TypeSafe also cross a provider boundary.