Re: [PATCH v5 3/6] agents: add a skill to determine stable tags

[Andrew Morton]

On Wed, 11 Nov 2026 21:32:17 -0500 Sasha Levin <sashal@xxxxxxxxxx> wrote:

> On Wed, Nov 11, 2026 at 05:10:55PM -0800, Andrew Morton wrote:
> >Marco had some thoughts on this just today:
> >
> >https://lore.kernel.org/CANpmjNMvM_TCs7Fao9+gM-NNpQyBKYumSuWOt_03ZX+tLwbjWw@xxxxxxxxxxxxxx
> 
> This makes a lot of sense.
> 
> >And Documentation/process/stable-kernel-rules.rst does ask for real
> >things, not theoretical stuff.
> 
> We've actually dropped[1] the "theoretical" part.
> 
> The problem we were seeing is that bots were sending real fixes, with a solid
> changelog, but not adding the stable@ tag.
> 
> Apparently our wording is pretty strong: unless you can prove that the issue is
> real, it remains theoretical. You'd have to either show a real case where you
> hit the bug, or to have reproduced the issue and can provide a way for us to
> reproduce it.
> 
> Most bots can't actually do that for technical reasons, even if the bug is 100%
> affecting our users, so they don't add the stable@ tag.

OK, I can see that's a problem.  The theoretical userspace-effects
statement will help here.

> >Also, I would keenly like to see changelogs clearly explain *why* a
> >backport is proposed.  As I have said 1000 times, "when fixing a bug
> >please clearly describe the userspace-visible runtime effects of that
> >bug".  Can we pleeeeeeeeeeze get this requirement clearly enshrined in
> >this patcset somewhere where the LLMs will read it and comply?  It's
> >really simple for the LLMs to cook up a paragraph for this, if so
> >directed.
> 
> This is a great point, and would make our lives much easier too.

Oh bless you, yes!

> It sounds a lot like something that humans would benefit from hearing as well?

Humans don't read stuff ;) But it would give me something to ask them to go
look at.  And to make their LLM go look at.


I picture some poor downstream person looking at a screenful of AI
babble wondering "should I add this to our kernel and if so why".  We
should tell them!

otoh, downstream people have AI too, so they can do what I do 5-10
times per day:

describe the worst-case userspace-visible runtime effects of this bug
<paste>

> If so, I can send a patch tomorrow to add it to
> Documentation/process/submitting-patches.rst and refer to it explicitly in the
> agents file?

We can't restate this too often IMO.  It's maddening.  So many
changelogs follow this form:

- here's what the code does
- here's why that's wrong
- here's what I did

So I sit there staring at this thinking "OK, but why should I care?".

Also, what's wrong with today :)