Alibaba Cloud's Kitta code review agent reviews 70,000+ patches in OpenAnolis
At the 2026 Yunqi conference, Alibaba Cloud technical expert Zhou Yuanhang presented the practice behind Kitta, a code review agent that moved from generic review to domain expertise. According to InfoQ China, Kitta has reviewed more than 70,000 patches online in the OpenAnolis community, with 97% gate accuracy and a 62% merge rate; the team says its domain-customization method was validated across six areas including OS kernels, compilers and agent-platform applications, yielding 17.0%–31.6% review-quality gains (another passage in the article says 17%–30%). Kitta distills review principles and recurring defect patterns from project history, customizes both review flow and domain knowledge, and iterates via offline evaluation plus online behavioral feedback, while feeding data back into Qwen base models — the article says its data has been used in Qwen 3.8 flash training. All operating and improvement figures are self-reported by the team in the talk and were not independently measured.
The report argues that as AI-assisted coding raises output and defects and vulnerabilities surface faster, code review has shifted from a supporting step to the first key gate in quality assurance — with Kitta's deployment figures offered as evidence for that claim.
随着 AI 深度参与研发流程，代码审查正在从辅助环节演变为质量保障体系的核心节点。阿里云技术专家周远航在 2026 云栖大会上分享了《代码审查进化：从通用审查到领域专家，Kitta 代码审查智能体的实践》。她详细介绍了代码审查智能体 Kitta 逐步打造为「领域专家」的进化历程：先分析 Al Coding 浪潮下研发流程面临的质量问题与代码审查场景的核心挑战，再从流程、数据、模型三个层面，介绍 Kitta 的领域定制方法、在龙蜥社区的落地效果，以及与 Qwen 基座模型协同进化的机制和信任边界的设计共识。
Kitta 在龙蜥社区线上运行至今，累计完成超 7 万个补丁的审查，门禁准确率达 97%，合入率稳定在 62%。其领域定制方法论已在操作系统内核、编译器、Agent 平台型应用等六大领域验证，实现 17.0%–31.6% 的审查质量提升，成功将 Maintainer 的精力从逐行审查释放至关键争议裁决。