A review pipeline with separate responsibilities

Google describes specialist passes that look for different kinds of defects, followed by verification of difficult cases. Both a suspected bug and a decision to dismiss it need supporting code evidence. This aims to catch an agent that notices a problem but then excuses it using an unverified assumption.

The authors report a decline in the kernel’s 60-day follow-up-fix rate from a 3.4% historical baseline to 2.2% in Linux 7.2. They explicitly caution that other factors can influence the comparison. These observations do not establish that Sashiko caused the entire improvement.

Specialization comes from project context

The repository describes a three-layer prompt structure: source-verified code notes, maintainer guidance and general review procedures. Linux is the most mature target; GCC, QEMU and Zephyr have different deployment or pilot status. Those distinctions matter when assessing readiness outside the kernel.

Local review operates on commits in a git checkout and leaves the working tree unchanged, according to the README. This is a review workflow, not LoRA training or a newly released set of model weights. Teams can inspect its Apache-2.0 code before considering adoption.

Evaluate the evaluator on known failures

The benchmarking guide provides a full Linux suite with 999 entries and smaller suites for iteration. Entries link a buggy commit with a description of the known problem. A structured model judge compares review findings with that reference; it is therefore important to inspect scoring as well as the final percentage.

Versioned tests distinguish exact detections, partial detections and false positives, and test whether earlier human feedback receives proper attribution. Saved run records include configuration and per-entry outcomes. Our assessment: a useful internal trial should include clean patches and previously missed defects, rather than counting only successful demonstrations.

Local execution and local inference are different choices

The provider guide documents Gemini as the default and alternatives through other APIs or compatible endpoints. Moving the reviewer into a local checkout does not, by itself, keep inference on that machine. Provider configuration determines where model requests go.

Before using private code, a team should decide which repository context it permits the chosen service to receive and how it will track review cost. Compare the same patch set under a fixed model, prompt version and effort setting. A clean AI report should inform a maintainer’s decision, while the project’s own quality rules determine whether a change can merge.