Kozuchi Agent: A Language-Agnostic Open-Weight Agent for Software Repair

Details

Industrial software-repair agents need to work across heterogeneous codebases without sending proprietary code to cloud APIs. Kozuchi Agent is designed around four principles: explicit phases, persistent state, deterministic tools, and a model-independent action interface. A CI-integrated evaluation pipeline reduces operator touch-points from five to one across heterogeneous internal clusters. On SWE-bench Verified, Kozuchi resolves 374/500 instances (74.8%) using locally hosted Qwen3.5-27B, no fine-tuning, and test-time scaling at TTS@8. On Multi-SWE-bench Java it ranks first among strict open-weight submissions (41/128, 32.03%) and fourth of 42 overall. On Python it ranks 12th of 135 and first among open-weight systems. Per-phase accuracy stays within ±5 percentage points across both languages, substantiating the language-agnostic claim. Analysis shows remaining failures center on semantic correctness and selection errors rather than edit formatting or proprietary-model access — meaning the open-weight gap is closing on closed models from a different direction than raw benchmark scores suggest. Accepted at IEEE/ACM ASE 2026 Industry Showcase in Munich.