On Fri, 3 Jul 2026 07:42:35 -0400
"Theodore Tso" <[email protected]> wrote:
> On Fri, Jul 03, 2026 at 09:30:28AM -0500, Greg KH wrote:
> > > So clarifying there that a "Assisted-by: LLM" is also good enough would
> > > make
> > > sense if we agree on that.
> >
> > If we all agree on that, I'm saying that _I_ don't care, but others
> > might.
>
> I agree that requiring the model name is not a good idea, because some
> developers might be using unreleased products where the name itself
> might not be public. So allowing the version number of the LLM to be
> elided would be a good change.
>
> I'll also note that it's starting to be the case where tool is
> actually quite relevant. So it's not just about the LLM Model, but
> whether you're using Codex, Claude Code, OpenCode, or Shahiko. You
> can use the tool or the harness with different models.
>
> So whether the tag is:
>
> Assisted-By: OpenCode:Gemini 3.5 Pro
> Assisted-By: Claude Code
> Assisted-By: LLM
I don't think an assisted-by for LLM actually helps.
I can't see any difference on adding it or doing something like:
Assisted-By: smath
Assisted-By: sparse
Assisted-By: coverity
Assisted-By: gcc
...
Assisted-By: my-favorite-web-search-engine
From my side, I don't care about whatever tool/process a developer has
to produce a patch. All I want is a proper patch that passes my review
criteria.
Also, identifying bad vibe-coding patches is usually easy: their
descriptions look like written by a former literature professor that just
took his first job as junior programmer: they come with useless comments,
random variable renames, unneeded code rewrites, etc.
Btw, if one wants them to be tagged as AI-produced, perhaps we can teach
Sashiko to identify it and add a big warning on its review. This will
probably be more reliable than asking people to add a LLM tag.
What I think we should do, instead, is to let it clear that vibe coding
is forbidden - e.g. just asking AI to do something and sending it as-is
(*).
So, I would add something like this at
Documentation/process/coding-assistants.rst:
"All patches shall be written or reviewed and modified by a human.
You should be able to understand exactly what the patch is doing,
why it is taking such approach and how. Pure non-trivial
machine-generated patches will be rejected."
(*) There are of course exceptions for it: typo fix patches for instance
can be pure machine-generated. Even before AI, we have patches like
that where people used some spell checker to fix typos. Again, they
need human review before submission, as sometimes, spell checker fails
when, for instance, it tries to replace an acronym with resembles
a word.
--
Thanks,
Mauro
