Passing tests is not enough.

A coding agent can fix the reported bug and still produce a patch a maintainer would reject. SWE-Gate turns that familiar review gap into two executable checks: one for functional behavior, one for a review-derived constraint such as compatibility, exception semantics, ordering, or metadata. 12

Across 303 repair instances from 75 open-source Python repositories, four model backends produced 644 functionally successful repairs. 221 of those repairs failed the constraint test, a 34.3% hidden-failure rate. Giving the constraint to the agent raised GPT-5.5's joint success from 41.3% to 52.8% in the paper's controlled ablation, while its functional success moved from 75.6% to 74.9%. 2

For a first product trial, choose one repeated repair class, write separate behavioral and acceptance tests, and run both beside the route you already trust. Promote only after joint success and review rework improve together. Keep human review for requirements that cannot be expressed as executable tests.