To make AI-written kernel code review-grade, we studied two sources that set the bar: Google's engineering practices (the human review guide plus its 2025 AI review tooling) and the Linux kernel community's own process docs. This page maps what they require onto what linux-kernel-dev actually enforces.
要让 AI 写的内核代码达到能过 review 的水准，我们研究了两个定标准的来源：Google 的工程实践(人工 review 指南 + 它 2025 年的 AI review 工具)，以及 Linux 内核社区自己的流程文档。这一页把它们的要求，对应到 linux-kernel-dev 实际强制了什么。
Google's open eng-practices guide says the point of review is that overall code health improves over time — "there is no perfect code, only better code." A reviewer covers eight things; in 2025 an AI does the mechanical first pass. Google 开源的 eng-practices 指南说，review 的目的是让代码整体健康度随时间变好——"没有完美代码，只有更好的代码"。reviewer 看八件事；2025 年起 AI 先做机械的第一遍。
Design (the most important), functionality, complexity, tests, naming, comments, style, documentation. Reviewers balance progress against perfection — don't block a CL over tiny nits.设计(最重要)、功能、复杂度、测试、命名、注释、风格、文档。reviewer 平衡推进与完美——别为小毛病卡住一个 CL。
Gemini Code Assist auto-reviews each PR (summary + bug/style flags); a 2025 "memory" feature accrues an evolving record of the team's standards from PR interactions. AutoCommenter enforces best practices for C/C++/Java/Go/Python.Gemini Code Assist 自动审每个 PR(摘要 + bug/风格标注);2025 年的"记忆"功能从 PR 互动里积累一份会演进的团队规范记录。AutoCommenter 为 C/C++/Java/Go/Python 执行最佳实践。
The kernel community enforces quality with documented tools and process — not taste. Most relevant to us: a 2025 doc on AI assistants that lines up exactly with our own no-AI-signature rule. 内核社区靠有文档的工具和流程把质量卡住——不靠口味。和我们最相关的：2025 年一份讲 AI 助手的文档，跟我们"不加 AI 署名"的规则完全对上。
checkpatch.pl for style; then sparse (types / endianness / __user), smatch (flow), coccinelle (semantic patches), clang-format. Each catches a different class — all have false positives.checkpatch.pl 查风格；再加 sparse(类型/大小端/__user)、smatch(流分析)、coccinelle(语义补丁)、clang-format。各抓一类，都有误报。
Every source file starts with an SPDX license identifier; every patch carries a Signed-off-by certifying the DCO (git commit -s). All code must be GPL-2.0 compatible.每个源文件首行带 SPDX 许可标识；每个补丁带 Signed-off-by 认证 DCO(git commit -s)。所有代码必须与 GPL-2.0 兼容。
The coding-assistants doc: an AI must not add Signed-off-by — only a human can certify the DCO. Disclose AI help with Assisted-by:; the human reviews all of it and takes full responsibility.coding-assistants 文档：AI 绝不加 Signed-off-by——只有人能认证 DCO。用 Assisted-by: 披露 AI 辅助；人类审查全部并负全责。
One patch does one thing; a big change is split into a series, each independently reviewable, compilable, and bisect-safe.一个补丁只做一件事；大改动拆成一系列，每个能独立 review、独立编译、不破坏二分。
Each answer is scored by seven fresh-context critics; design, testing and complexity were added to match Google's core concerns. Kernel work adds one dimension Google doesn't need: safety, with a veto — a sleep-in-spinlock is not a nit. 每个答案由七个独立上下文的打分员评；为对齐 Google 最看重的几项，新增了 设计、测试 和 复杂度。内核活儿多一个 Google 不需要的维度：safety，带一票否决——持 spinlock 睡眠不是小毛病。
| Google dimensionGoogle 维度 | Our coverage我们怎么覆盖 |
|---|---|
| Design | kernel-design-critic — fits the subsystem framework, no reinvented wheels (new)—— 融入子系统框架、不重造轮子(新增) |
| Functionality | kernel-correctness-critic + fact_gate — claims grep-verified against a real tree—— 断言对真树 grep 核过 |
| Complexity | kernel-complexity-critic — could it be simpler, can a maintainer follow it (new)—— 能否更简单、维护者能否看懂(新增) |
| Tests | kernel-testing-critic + regression_test — KUnit/kselftest coverage, error-path tests, testability (new)—— KUnit/kselftest 覆盖、错误路径测试、可测性(新增) |
| Naming · Style | kernel-coding-style-critic + checkpatch_gate — coding-style.rst + kernel's own checkpatch.pl—— coding-style.rst + 内核自带 checkpatch.pl |
| Comments · Docs | kernel-completeness-critic — error handling, cleanup, the claims block—— 错误处理、清理路径、断言块 |
| (kernel-only) Safety(内核独有)安全 | kernel-safety-critic — atomic-sleep / UAPI / hardware rules; an error here vetoes the answer—— 原子睡眠/UAPI/硬件规则；这一轴出 error 直接否决 |
Weights: correctness .28 · safety .22 · design .13 · testing .12 · complexity .09 · coding-style .08 · completeness .08. See the weighted bar on the capabilities page. 权重：correctness .28 · safety .22 · design .13 · testing .12 · complexity .09 · coding-style .08 · completeness .08。加权条见能力页。
Google treats design, complexity and tests as core review concerns; we had no critic for them. Added three, taking the panel from 4 to 7 axes.Google 把设计、复杂度、测试都当 review 核心关注，我们之前没有对应打分员。补了三个，面板从 4 轴到 7 轴。
The kernel's rule became Forbidden Action #11: an AI/tool never adds Signed-off-by; disclose with Assisted-by:; the human signs and owns it.内核这条成了 Forbidden Action 第 11 条：AI/工具绝不加 Signed-off-by；用 Assisted-by: 披露；人类签名并负责。
Third: the submission workflow (patch-workflow.md) gained the static-analysis tools (sparse / smatch / coccinelle / clang-format), SPDX, and the DCO flow — so review-grade output is documented end to end, from checkpatch to sign-off.
第三条：提交流程(patch-workflow.md)补上了静态分析工具(sparse / smatch / coccinelle / clang-format)、SPDX、DCO 流程——从 checkpatch 到 sign-off，能过 review 的产物全程有文档。