Google-developed Sashiko, the agentic AI code review system originally built to help review Linux kernel patches, has gained a new "performance specialist" review stage — a capability aimed at judging whether a proposed change makes the kernel faster or slower before it is merged.
The addition was reported by Phoronix, which has followed Sashiko's steady expansion since the project began surfacing reviews on the kernel mailing list (LKML). Alongside the existing correctness-focused review, Sashiko recently gained an experimental stage that evaluates the design of a patch series as a whole, and now a third stage dedicated to performance and benchmark considerations.
What the new stage actually checks
In plain terms, the performance specialist addresses a question human reviewers often defer until after merge: does this patch help or hurt? Stages of this kind typically flag changes that may degrade hot paths, introduce lock contention, or add overhead that only surfaces under load — the kind of non-functional regressions that are expensive to catch once code has landed in the mainline tree.
That marks a progression in Sashiko's remit. Its first skill was narrow: is this code correct, and does it follow the conventions of the subsystem? The patch-design stage widened the lens to ask whether a change is well conceived and worth merging at all. The performance stage extends it further, to whether the change does what it is supposed to do efficiently — moving the system from "is this correct?" toward "is this a good change on the merits?"
Phoronix notes that Sashiko has continued to uncover bugs and issues both as new patches appear on LKML and when users run the software themselves, though the report frames the new stage as complementary rather than authoritative. Static AI reasoning about performance cannot substitute for actual benchmark runs; its value lies in surfacing specific, verifiable concerns — a contention point, a degrading branch — for human scrutiny ahead of merge. Performance regressions are among the most costly kernel bugs to detect post-merge, and review-time detection shifts that work earlier, into a stage where it is far cheaper to fix.
That matters against the backdrop of LKML's chronic review bandwidth bottleneck. Kernel maintainers routinely triage hundreds of patches with limited time, and non-functional quality — performance, locking behaviour, scalability — is often the first thing to slip when reviewers are stretched. An AI agent that flags plausible performance risks before a human ever opens the diff does not replace maintainer judgement, but it changes what that judgement is spent on.
Open questions and the broader trend
Key questions remain unanswered. Google has not disclosed how confident Sashiko is in its performance assessments, nor whether the stage will integrate with formal benchmarking infrastructure — details that will likely determine how seriously LKML maintainers treat its feedback. How the mailing list actually responds to performance-stage reviews remains the real credibility test.
The story also fits a wider pattern: AI agents are increasingly being pointed at non-functional quality — performance, security posture, maintainability — rather than just functional correctness. Sashiko is one of the clearest public test cases for that shift at kernel scale, and notably it is self-hostable: the project is distributed as open source, so developers and maintainers outside Google can run the review pipeline against their own trees.
For the wider open-source community, the practical takeaway is that AI review is evolving from a linting exercise into something closer to a first-pass maintainer — one whose performance opinions, at least for now, still need to be verified before they are believed.
由 Google 開發的 Sashiko,是一套 agentic AI 代碼審查系統,最初旨在協助審查 Linux kernel patches,現已新增「效能專家」審查階段——該功能旨在於 patch 合併(merge)之前,判斷相關改動會令 kernel 變快還是變慢。
Phoronix 報導了此項更新。自項目開始在 kernel 郵件列表(LKML)上發佈審查以來,Phoronix 一直跟蹤 Sashiko 的穩步擴展。除現有的以正確性為重點的審查外,Sashiko 最近已加入一個實驗性階段,用於評估整個 patch series 的設計,如今再加入第三個階段,專責處理效能及 benchmark 方面的考量。
新階段實際審查什麼
簡單而言,效能專家針對的是一個人類審查者往往要等到合併之後才會處理的問題:這個 patch 是有益還是有害?此類階段通常會標記可能劣化 hot path、引入 lock contention,或增加只有在負載下才浮現的 overhead 的改動——這類 non-functional regressions 一旦代碼已進入 mainline tree,事後補救的成本極高。
這標誌著 Sashiko 職責範圍的演進。它第一項技能範圍狹窄:這段代碼是否正確?是否符合該 subsystem 的慣例?patch 設計階段擴闊了視野,追問某項改動是否構思完善、是否值得合併。效能階段則進一步延伸,追問該改動能否有效地做好它應做之事——將系統從「這是否正確?」推向「從實質而言這是否是一項好的改動?」
Phoronix 指出,無論是 LKML 上出現新 patches 時,還是用戶自行執行該軟件時,Sashiko 持續發現 bugs 及問題,不過報道將新階段定位為補充性而非權威性。對效能的靜態 AI 推理無法取代實際的 benchmark 運行;其價值在於於合併之前,向人類審查者提出具體、可驗證的疑慮——一個 contention point、一個劣化的分支。效能 regressions 是合併後代價最高的 kernel bugs 之一,而在審查階段偵測可將該項工作提前,置於修復成本遠為低廉的階段。
在 LKML 長期存在的審查帶寬瓶頸背景下,這點尤為重要。Kernel maintainers 通常要在時間有限的情況下分類處理數以百計的 patches,而 non-functional quality——效能、locking 行為、可擴展性——往往是審查者分身乏術時最先被忽略的。一個能在人類審查者打開 diff 之前標記出合理效能風險的 AI agent,並不能取代 maintainer 的判斷,但會改變該判斷所花費的方向。
未解問題與更廣泛的趨勢
關鍵問題仍有待解答。Google 並未披露 Sashiko 對其效能評估有多大信心,也未說明該階段是否會與正式的 benchmarking infrastructure 整合——這些細節很可能決定 LKML maintainers 會多認真對待其反饋。郵件列表實際上如何回應效能階段的審查,才是真正的可信度考驗。
這則報道亦符合一個更廣泛的模式:AI agents 越來越被指向 non-functional quality——效能、安全態勢(security posture)、可維護性——而非僅限功能正確性。Sashiko 是在 kernel 規模上該轉變最清晰的公開測試案例之一,而且值得注意的是,它可自行託管(self-hostable):該項目以 open source 形式發佈,因此 Google 以外的開發者及 maintainers 亦可對自己的 trees 執行該審查 pipeline。
對更廣泛的 open-source 社群而言,實務上的要點是:AI 審查正由一項 linting 工作演變為較接近一位初審 maintainer 的角色——至少目前為止,其效能方面的意見仍需經核實方可相信。
