Sashiko: An agentic Linux kernel code review system

The Linux Foundation has unveiled Sashiko, an open-source, agentic system designed to automatically review code patches submitted to the Linux kernel. Funded by Google, which provides all compute resources and LLM tokens, Sashiko acts like a team of specialized reviewers by monitoring the public Linux Kernel Mailing List (LKML). It uses a custom multi-stage review protocol and a set of open-source prompts to evaluate proposed changes across domains like security, architecture, and concurrency. The goal is to reinforce the kernel's integrity by catching bugs early, a mission reflected in its name, which refers to a Japanese stitching technique for reinforcement.

Powered by Google's Gemini 3.1 Pro model, Sashiko has demonstrated significant potential in testing. On an unfiltered selection of the last 1,000 upstream commits that were later tagged as 'Fixed,' the system successfully identified 53.6% of the bugs. Crucially, 100% of these bugs had originally passed through the existing human-driven review process, highlighting a key gap the AI aims to fill. The project is licensed under Apache 2.0 and is explicitly designed as an augmentation tool for human maintainers, not a replacement, acknowledging the probabilistic nature of LLM outputs.

The system represents a major step in applying AI to the complex, high-stakes domain of systems programming. By automating the initial triage and review of patches, Sashiko could drastically reduce the workload on kernel maintainers and accelerate the development cycle. Its open-source nature and foundation backing suggest it could become a standard tool in the kernel development workflow, setting a precedent for AI-assisted review in other large-scale open-source projects.