Sashiko code review system aims to help Linux maintainers
Google Cloud Blog says Sashiko uses Gemini-powered agents to review Linux kernel patches. The open-source project is intended to ease pressure on maintainers and catch problems before code is merged.

Key takeaways
- Google Cloud Blog says Sashiko uses Gemini agents to review Linux kernel patches.
- The system can run as many as seven specialist sessions and verifies difficult cases separately.
- Google operates Sashiko at sashiko.dev and says developers do not pay for reviews.
- The project was given to the Linux Foundation and publishes prompts, tool activity and rejected findings.
In this article
Why kernel review is a bottleneck

Google Cloud Blog says Sashiko code review is an open-source system that uses Gemini-powered agents to inspect Linux kernel patches. The post, published on 18 November 2026, describes how the project aims to help maintainers handle a growing review workload.
Linux is the core of many systems, from Android phones to cloud services and supercomputers. According to Google Cloud Blog, the kernel contains over 35 million lines, and more than 85,000 changes arrive each year. That volume makes review a constant task for a relatively small group of maintainers.
Automated tools already catch some problems. Build bots and fuzzers such as Syzkaller can find build errors and crashes. But Google says review remains the main method for protecting stability and safety. Maintainers read many proposed patches by hand each day to look for memory leaks, locking faults, race conditions and security flaws.
The pressure is growing as coding agents produce more patches. Google says some generated changes are useful, while others look plausible but are sent by people who do not grasp the systems they are changing. It also says AI-generated security reports can add work when they are duplicates or inaccurate.
How Sashiko reviews patches
Sashiko is an open-source review system built by the Google AI & Infrastructure team. It uses Gemini, an AI system from Google, to examine proposed kernel changes. Google describes it as agentic, meaning it uses multiple AI agents to carry out review tasks. Google says the project is named after a Japanese craft that patches worn fabric. The company also says it gave the project to the Linux Foundation earlier in the year.
The system is designed to separate two jobs: finding possible bugs and proving whether those bugs are real. It first gathers context around a patch. This includes changed functions, related data structures, direct callers and subsystem guides. A planner then decides which review passes are needed and how much reasoning each pass should receive.
As many as seven focused sessions can run at the same time. Each session looks at one class of problem, such as commit intent or code match. Other sessions examine control flow, resource lifetimes, locking, security and hardware. Google says this focus helps stop one type of issue from hiding another.
Verification and public reports
Each session uses git tools to search the repository. Google says a review uses roughly 100 tool calls on average. The sessions produce two lists: concerns they believe are real, and concerns they investigated and rejected, with code evidence for each conclusion.
A lead verifier then audits both lists. Clear bugs supported by the code pass through. If a local check proves a dismissal, it is removed. Other cases receive more work. This includes disputed bugs and weak dismissals where a specialist assumed that an unseen caller or lock would prevent a problem.
For difficult cases, Sashiko gives each one a separate verification agent. Each agent answers one question and must follow a strict rule. A bug claim or a dismissal needs firm code evidence at every call site. Google says this can catch cases where a specialist dismisses a real bug by mistake.
The final report uses the style of a maintainer question on a Linux mailing list. Google operates the service at sashiko.dev and examines patches sent to public Linux kernel mailing lists. Developers do not pay for these reviews. Google says the project publishes its prompts, tool activity, reasoning steps and rejected findings.
Lessons from AI review limits
Google says general-purpose AI models can struggle with deep review. They are trained to use fewer tokens, so they may skip work or reject concerns without proper checks. They can also be biased by the way a task is framed. A model may perform better when asked to prove one thing, such as a memory leak, than when asked to explore many different concerns.
Google also notes that commit messages can affect review results in ways that are hard to predict or even malicious. On the positive side, a model without persistent memory can treat repeated or similar patches as fresh work. Google says this can help when the same change is applied to many slightly different branches.
The Sashiko design reflects these lessons. It uses focused passes, parallel checks and a verification stage that records what was ruled out. Google presents this as a way to reduce false alarms and missed bugs, but these are design claims rather than independent test results.
Why it matters for maintainers
The main problem is not a lack of generated code. It is the limited supply of trusted review. Google says extra generated patches may only add to the queue already facing maintainers unless reviewers have a way to trust proposed changes.
Sashiko is intended to catch bugs before they land and reduce the load on human reviewers. It also aims to make the review process open, so contributors can inspect how the system reaches its conclusions. Over 100 contributors, among them kernel maintainers and engineers, have helped improve the prompts and the system. The project started with open-source Linux review prompts created by Chris Mason and later moved to a prompt system that combines generated notes with maintainer guidance.
Google says success will depend on whether maintainers read and act on the reports. That matters because Linux supports billions of devices and much of the cloud. If automated review can earn maintainer trust, it could help large open-source projects handle the growing volume of AI-assisted code.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



