CODEXA

ProoFix makes AI-generated bug fixes accountable to the failures they are meant to solve. It addresses a critical burden in AI-assisted development: developers must still investigate whether a convincing patch fixes the reported behavior, introduces regressions, or changes more than intended. Built for Python repositories, ProoFix coordinates specialized agents through a complete repair workflow. They analyze code and dependencies, attempt to reproduce the failure, investigate its cause, map the affected files, and create an ordered repair plan before generating a focused patch. Each candidate faces multiple checks: target and regression tests, mutation testing that evaluates whether tests detect injected faults, and security scans that compare findings before and after the change. Acceptance-test integrity checks detect whether the test changed after reproduction, keeping validation tied to the original failure. The evidence also determines what the system does next. A failed check can trigger a revised patch, a narrower repair scope, or renewed investigation. Recovery is bounded, missing checks remain visibly unverified, and unresolved diagnosis concerns require human review—even when tests pass. Powered by LangGraph orchestration, a FastAPI backend, Redis, and a live React workspace, ProoFix gives developers and open-source maintainers a traceable path from failure to proposed repair. They can inspect the diagnosis, patch, executed checks, and remaining concerns together, then explicitly choose whether to publish an eligible repair as a draft GitHub pull request. Final acceptance stays with the developer. 9:02 PM