On Tuesday, the Sashiko project released version 0.9.4, adding an experimental assessment of patch series design to a tool that already screens code posted to the Linux kernel mailing list. Developed and maintained by Google engineers, Sashiko uses Gemini as its default model but is designed to be LLM-agnosticโ€”a detail that matters more than it might appear.

The shift from simply flagging bugs to evaluating patch design is not a marginal upgrade. It means the AI is no longer limited to catching syntax errors or obvious defects; it attempts to judge how a series of patches is structured, how the changes relate to one another, and whether the overall flow is coherent. For kernel maintainers, who face a huge volume of patches, this triage capability can reduce the time spent on initial screening and focus attention on the most problematic submissions.

Making the tool LLM-agnostic has structural implications. It does not lock adopters into Gemini or a specific cloud API. A team managing development infrastructure can integrate local or self-hosted models, building inference pipelines where patch code and metadata remain under its own control. For those evaluating on-premise deployment, there are trade-offs between the cost of running local models and independence from external services, but Sashiko's architecture removes one typical constraint: dependence on a single model vendor.

There is a downside, however. As AI code review moves toward design assessment, it becomes more opaque: it is not always easy to reconstruct why a patch series is judged problematic. This can create friction for contributors, especially if automated flags shape public discussion on the mailing list. Trust in the kernel review process, historically based on explicit argumentation and human review, cannot be delegated without clear auditability criteria.

Version 0.9.4 is an incremental step, but it signals a clear direction: automated patch screening is becoming a continuous component of kernel maintenance. The fact that it is built to work with different models makes this path less tied to vendor choices and better suited to contexts where control over code and data is a priority.