Specification and verification sit inside the agent loop
Repository-scale coding work has stronger evidence when the system carries its own plan and checks its own output. ReCodeAgent splits translation into analysis, planning, translation, and validation, then reports 99.4% compilation success and 86.5% test pass rate across 118 projects and four language pairs. REAgent uses the same broad pattern for issue fixing: build a structured requirement first, score it with generated tests, then refine it until the requirement is good enough to drive a patch. The reported gain is 9.17% to 24.83% in resolved issues over baselines. The common point is simple: repository agents improve when specification and verification are first-class steps inside the loop, not cleanup added at the end.