Re: AI-assisted review

[Tom Rini <trini@konsulko.com>]

[-- Attachment #1: Type: text/plain, Size: 2690 bytes --]

On Mon, May 18, 2026 at 10:55:40AM +0200, Michal Simek wrote:
> 
> 
> On 5/16/26 00:07, Tom Rini wrote:
> > On Fri, May 15, 2026 at 03:03:21PM -0600, Simon Glass wrote:
> > 
> > > Hi,
> > > 
> > > There was a query on the call this week about whether I am doing
> > > AI-assisted code review. As I said on the call: yes. Here is a brief
> > > description of how it works.
> > > 
> > > It is built into Patman (on the Concept tree) with a new 'patman
> > > review' command. You give it the series name / number, or perhaps a
> > > patch name/number and it applies the patches to a new branch, does a
> > > review then adds its comments to its database.
> > > 
> > > A '-d' flag can be used to create draft emails in Gmail (sorry, it
> > > doesn't support other email programs yet). You then check and update
> > > the emails and send them (or delete them). I am not an expert in
> > > handling the 'user voice' part of AI, but have made an attempt to make
> > > it follow any provided configuration, as well as to scan recent
> > > reviews to actually create to create a voice.
> > > 
> > > Obviously this is very rudimentary and could be expanded considerably.
> > > But the mere fact that it creates draft emails is a win for me, even
> > > if I ultimately delete or rewrite most of the comments. I can imagine
> > > 10 different ways to improve it to be more useful.
> > > 
> > > I wrote a blog post about it if you want more details, or you can ask me here.
> > > 
> > > I am very interested in hearing how others are using these new tools
> > > for code review.
> > 
> > And the big thing for now is that since we as a project do not yet have
> > an AI policy aside from "please don't". One of the points I was making
> > on the call is that there's a difference in value between "Human
> > reviewed it, looks fine" and "Human spent some tokens, agent didn't see
> > any problems".
> > 
> > And I know several other people have been doing at least first pass
> > reviews with various agent-tools, it's just no one else has been posting
> > reviews at your scale. And lessons learned from other projects is that
> > the prompts are more important than whatever wrapper around the agent
> > one is using.
> > 
> 
> Don't think scale is the problem. Tool and integration is another topic.

Simon posted approximately 100 reviews in about 24 hours. That scale is
a problem, when most of them are just reviewed-by tags, from someone
that has a history of doing human reviews. Reputation is a factor here
I'm trying to figure out how best to articulate.

I have thoughts on the rest that I want to get back to later, thanks.

-- 
Tom

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]