Splitting Bug Reproduction Into Diagnose-Then-Write With a Handoff Protocol Hits 81.76% on SWT-Bench Verified

arXiv (2608.23341, submitted 2026-08-24)·medium signalXBlueskyLinkedInCopy link

DPIAgent (arXiv 2608.23341, submitted 2026-08-24) argues that reproduction test generation is two tasks wearing one loop, and that merging them causes goal drift because the intermediate goal is never specified. It Divides the run into defect exploration and test generation, enforces a Protocol that records the diagnosis and test plan across the handoff, and Isolates each phase's toolset so irrelevant tools cannot mislead the agent. On SWT-Bench Verified with GPT-5 it reaches 81.76%, the highest reported for an open-source method, up to 11.88 points over the strongest baseline on GPT-5-Mini, and 86.17% with test selection added.