Re: Recommended local LLM-assisted review for netdev submissions?

[Jakub Kicinski <kuba@kernel.org>]

On Tue, 28 Apr 2026 08:17:55 -0700 Satish Kharat wrote:
> I'm the author of the in-flight enic SR-IOV series. Before posting
> revisions I run checkpatch --strict, make W=1 C=1 on touched files,
> scripts/kernel-doc -none, and `b4 prep --check`.
> 
> I've read Jesse's Netdev 0x18 talk on AI-enhanced reviews [1], the
> LWN coverage [2][3], and looked at sashiko [4]. What I haven't been
> able to tell from the public material:
> 
>   - Is sashiko (or a successor) what's currently producing the
>     LLM-style review replies showing up on netdev, or is that a
>     different pipeline?
>   - For contributors who'd like to self-review with the same
>     prompt/ruleset before posting, is there a recommended setup,
>     or is sashiko-as-is the closest thing today?

Our current recommendation for "individual" contributors is to use
the review prompts directly:
https://netdev-ai.bots.linux.dev/ai-local.html

If you have / can get LLM API keys sashiko seems to be producing
better results (although not strictly better). Ideally you'd want
to run Sashiko on all three of the major LLM models and then another
sweep to coalesce the results. We can't do this at netdev today because
it's cost prohibitive. But Google's Sashiko runs on Gemini, ours runs
on Claude and they find different things..

I tried to improve Sashiko + CLI models but from my limited
experimentation it's worse than just using the review prompts.

> Happy with a one-line pointer if it's already documented somewhere
> I missed.

Sorry for not having the source of truth. The space is rapidly
evolving :(