Sashiko AI code review has become part of daily Linux kernel work, according to the Google engineers who built it. The open-source system, powered by Gemini and donated to the Linux Foundation this year, completes roughly 1,200 reviews a day across more than 100 kernel mailing lists.
Roman Gushchin and Eric Brewer of Google’s AI and Infrastructure team described the project and its data. They say review, not code generation, limits how fast large open-source projects move.
The problem it targets
Linux runs billions of Android phones, most cloud providers and nearly every supercomputer. It has more than 35 million lines of code and over 85,000 changes land every year. Maintainers read hundreds of patches by hand to catch memory leaks, deadlocks, race conditions and security bugs.
AI makes it worse, the authors say. Some generated patches are solid, many are plausible-looking “AI slop” from contributors who don’t understand the code, and AI-written security reports are often duplicates or wrong.

How a review runs
Sashiko splits finding bugs from verifying them. It first pulls in code around the diff, including modified functions, struct definitions and direct callers, plus the subsystem’s guides. A planner picks which specialist passes are needed.
Up to seven specialist sessions then run in parallel, each on one class of problem: commit intent, implementation match, control flow, resource lifetimes, locking, security and hardware. Each averages about 100 tool calls with git tools and returns two lists, concerns it thinks are real and ones it investigated and ruled out.
A lead verifier audits both lists. Clear bugs pass through, and shaky dismissals get their own verification agent, which needs concrete code proof across all call sites. The final report reads like a human maintainer asking questions, in the style of a Linux mailing list.
The authors say LLM reviewers tend to cut corners, are easy to bias with commit-log text, and perform better on one narrow task than several. Lack of persistent memory helps, because the model can review near-identical patches repeatedly.

What the numbers show
- Review emails go only to lists that opt in, and over 70 have. More than 1,500 kernel developers have replied.
- Sashiko is cited in over 2,000 mainline commits and over 500 of this year’s official kernel CVEs.
- Across the 18 releases from Linux 6.2 through 6.19 (254,000 commits), 3.4% of merged commits needed a fix within 60 days. That fell to 2.8% in Linux 7.1 and 2.2% in 7.2, a 35% drop.
- Among 21,500 tracked commits, patches with a clean review needed a fix within 60 days 0.9% of the time. Patches merged despite a critical warning needed one 7.6% of the time, 8x higher.
The authors caution that other factors affect release-to-release numbers, so Sashiko cannot take all the credit. They also cite LWN.net reporting that it was the top bug reporter in the 7.2 cycle.

Open and free for developers
Google runs the service at sashiko.dev at no cost to contributors, since paying for deep reviews with their own API keys would push people away. Every prompt, tool call, reasoning step and ruled-out bug is public. More than 100 contributors have helped refine the prompts, which started from Chris Mason’s open-source Linux review prompts.
Why maintainers trust a narrow tool
The design choices respond to specific failure modes the authors list. A model told to look for many things tends to find one obvious issue and stop, so each specialist gets a single focus. Models also tend to dismiss concerns without investigation, so every ruled-out bug must be recorded with supporting code, and a verifier can promote it back to a finding.
Recent changes include a new prompt system that combines auto-generated notes on each tree or version with notes from maintainers that cannot be derived from the code. The authors say pushback on false alarms from kernel developers has directly improved the prompts and the verification engine.
What it means for maintainers
A clean pass gives maintainers more confidence, and a critical warning is a reason to look again. Next, watch whether the follow-up fix rate keeps falling in later releases, and whether more lists opt in to Sashiko AI code review.
Source: Google Cloud Blog