一句话：提出 Kozuchi Agent，一个语言无关、开源权重的 LLM 修复代理与 CI 评估管道，在 SWE-bench 等基准上达到开源权重领先水平。
工业软件工程团队需要将 bug 报告转化为正确补丁的 LLM 代理，但基准规模操作面临长视野、工具使用纪律、上下文持久性、异构集群和评估复用等挑战。本文提出 Kozuchi Agent，一个语言无关、开源权重的修复代理及 CI 操作评估管道，通过显式阶段、持久状态、确定性工具、模型无关动作接口和跨代理测试时选择，使运行可审计且可重复。使用本地托管的 Qwen3.5-27B、无微调、TTS@8，在官方评估器上解决 374/500 个 SWE-bench Verified 实例；在 Multi-SWE-bench Java 上解决 41/128（32.03%），在严格开源权重提交中排名第一，总榜第四；Python 上排名第 12/135，开源权重系统第一。跨语言各阶段行为差异在 5 个百分点内。剩余失败主要源于语义正确性、Java 特定测试框架问题和选择错误。可复用 CI 阶段将异构内部集群的操作接触点从五个减少到一个。
Industrial software-engineering teams increasingly need LLM agents that turn bug reports into correct patches, yet benchmark-scale operation adds long horizons, tool-use discipline, context persistence, heterogeneous clusters, and evaluation reuse. We present Kozuchi Agent, a language-agnostic open-weight repair agent and CI-operated evaluation pipeline. Explicit phases, persistent state, deterministic tools, a model-independent action interface, and cross-agent test-time selection make runs auditable and repeatable. With locally hosted Qwen3.5-27B, no fine-tuning, and TTS@8, Kozuchi resolves 374/500 SWE-bench Verified instances on the official evaluator. Unchanged on Multi-SWE-bench Java, the same 27-billion-parameter agent resolves 41/128 instances (32.03%), ranking first among strict open-weight submissions and fourth of 42 overall; on Python it ranks 12th of 135 and first among open-weight systems. Per-phase behavior remains within 5 percentage points across languages. Remaining failures mainly reflect semantic correctness, Java-specific harness issues, and selection errors. Across both tracks, results compare favorably with open/local peers by parameter count. Analysis of candidate diversity, selector regret, and patch reliability shows that the remaining gap is primarily semantic correctness and selection rather than edit formatting or proprietary-model access. Operationally, reusable CI stages reduce operator touch-points from five to one across heterogeneous internal clusters.
@inproceedings{paperbot12231,
title = {Kozuchi Agent: A Language-Agnostic Open-Weight Agent for Software Repair},
author = {Mehdi Bahrami and Kosaku Kimura and Satoshi Munakata and Satoshi Nakashima and Yu Ishikawa and Kosuke Maeda and Nao Soma and Kenichi Kobayashi and Keisuke Miyazaki and Keizo Kato and Shigeki Fukuta and Tatsuo Kumano and Nobutaka Imamura and Kevin Musgrave and Shahbaz Abdul Khader and Kwun Ho Ngan and Joe Townsend and Fayas Asharindavida and Matthieu Parizy and Akira Sakai and Yuma Ichikawa and Yang Zhao and Michiaki Takizawa and Taku Fukui and Hiroki Ohtsuji and Wei-Peng Chen and Hiromichi Kobashi},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834531}
}