Re: Sashiko review email for kvm@vger.kernel.org?

[Janosch Frank <frankja@linux.ibm.com>]

On 5/19/26 22:25, Sean Christopherson wrote:
> +KVM and LKML, which I stupidly forgot to Cc :-(
> 
> On Tue, May 19, 2026, Janosch Frank wrote:
>> On 5/19/26 02:18, Sean Christopherson wrote:
>>> Hello people with a MAINTAINERS entry that uses kvm@vger.kernel.org as a mailing list!
>>>
>>> How would y'all feel about enabling Sashiko[*] emails for kvm@vger.kernel.org?  For
>>> a variety of reasons, Sashiko currently only logs its reviews to https://sashiko.dev
>>> by default.  If subsystems want to receive on-list mails, they need to opt-in.
>>>
>>> I really, really want to enable email reviews for KVM x86, as Sashiko has found
>>> an almost embarassing number of bugs that (a) would have been painful to debug
>>> had they escaped review and (b) I'm not convinced would have been found by
>>> traditional review.  And for me at least the signal:noise ratio is more than good
>>> enough.  There are still hallucinations and false positives from time to time,
>>> but the good outweighs the bad by quite a bit, it's cathartic to flame AI when
>>> it's wrong, and worst case scenario I can simply ignore Sashiko.
>>>
>>> But, currently Sashiko is only capable of making decisions based on mailing list,
>>> i.e. as things stand, _all_ of kvm@ needs to opt-in.  Before I badger Roman to
>>> add bespoke logic to Sashiko's email flow, e.g. to special case KVM x86 in some
>>> way, I want to see if there's a general consensus in favor (or against) enabling
>>> review emails.
>>>
>>> Thanks!
>>>
>>> [*] https://github.com/sashiko-dev/sashiko
>>
>> Sure,
>    ^
>    |- This
> 
>> as long as there will be no mails from Sashiko or at most one mail per thread.
> 
> is in conflict with the caveats here, as per-patch email is exactly what I am
> proposing (the current state is that Sashiko never sends mails).>
>> I'd appreciate something like "I found problems in this thread, here's the
>> link, go look if you want to".
> 
> While this would be easier than manually searching through the web site, there
> would still be a sizeable barrier between contributors/maintainers and Sashiko
> feedback.  E.g. to click through an out-of-band source, copy+paste the response
> into the original thread, etc.
> 
> What are your concerns with per-patch email?  Hallucinations and noise are
> definitely a potential problem, but IMO that can be handled by per-subsystem
> policy.  E.g. if the signal:noise ratio is particularly poor for a subsystem,
> that subsystem can document/state that responding to, or even reading, Sashiko
> mails is completely optional.
> 
> Beyond that, I'm struggling to understand why folks are opposed to getting mails
> from Sashiko, especially for a list as large as kvm@.  As others pointed out, the
> volume for kvm@ is already high, and it's trivially easy to create a filter for
> mails from sashiko-bot@kernel.org.  I.e. it seems highly unlikely that mail from
> Sashiko will require anyone to significantly rethink and/or rework their workflow,
> whereas the web site approach is very disruptive to email-based review, at least
> for me.

In my mind there's a clear separation between human feedback and 
non-human feedback. So the disruption is something that I prefer. Maybe 
I just haven't yet fully evaluated how much I trust its review in 
contrast to my human coworkers.

That being said, I won't be the one holding this back. I see value in 
this and I'd be happy to test this.