Re: How should we handle using AI for reviewing commits?

[Sasha Levin]

On Thu, Mar 12, 2026 at 12:45:09PM -0400, Theodore Tso wrote:
> We recently used an AI review bot while applying an LTS backport to an
> internal kernel tree at $WORK.  While doing the review, it flagged a
> set of concerns which resulted in my creating a patch[1] to address
> the issues that it found in the kernel commit.
>
> [1] https://lore.kernel.org/all/20260310122806.1277631-1-tytso@xxxxxxx/
>
> In this commit there is no LLM generated output in the code, but there
> *is* LLM generated output in the commit description, since I quoted
> the concerns raised by the LLM.  Per the our new coding-assistants
> process document[2], "When AI tools contribute to kernel development,
> proper attribution helps track the evolving role of AI in the
> development process. Contributions should include an Assisted-by tag..."
>
> [2] https://www.kernel.org/doc/html/latest/process/coding-assistants.html
>
> When I was considering whether I should add something like:
>
> Assisted-by: Gemini:Gemini-3.1-Pro [TOOL]
>
> There was a couple of things that came to mind.  First, should we make
> some kind of distintion between exactly how the AI tool assisted in
> the development of the commit?  There's a big difference between using
> an AI assistant to find a potential issue, to an AI assistant which
> created new code out of whole cloth, with a spectrum of changes in
> between.  Given that the stated code was to "track the evolving role
> of AI", it occured to me that perhaps we should add some indication
> about exactly what was the nature of assistance that was provided.

Why not use Reported-by for these? We already use it for tools (like syzbot),
so this would just be a natural extension.

	Reported-by: Gemini:Gemini-3.1-Pro

> The second observation that I had was that example set of tools for
> [TOOL] was "specialized analysis tools": coccinelle, sparse, smatch,
> clang-tidy.  I assume the intent was if an AI bot started using tools
> like sparse, coccinelle, as an agent?w

Yup, both that as well as just space to list additional tools that the human
used too without adding multiple Assisted-by: lines.

> If there is a set of LLM prompts which has a name, would that also be
> appropriate for TOOL?  Chris Mason's repo has a fairly non-descriptive
> name, "review-prompts", but in the future when companies start making
> their internal review prompts public, some of them may have more
> evocative names that might be more unique and more marketing friendly.  :-)

Something like:

	Reported-by: Gemini:Gemini-3.1-Pro # https://example.com/my-magical-prompt

?

--
Thanks,
Sasha