> Sashiko sometimes find bugs in adjacent code, but submitters don't want to have to fix bugs they didn't create.

I had my fair share of patches reviewed by humans like Simon and others where they caught a bug in adjacent code as well.
While I understand the scope creep frustration, I'm convinced that it makes the code base better overall because it means patch authors have to think about how their contribution fits into the design of the component they are touching.

It is interesting that LLMs provide a channel for that frustration, but it is something that a diligent, high quality human review can trigger as well (and we are better off for it). :)


Am Fr., 24. Juli 2026 um 15:52 Uhr schrieb Benjamin Redelings via ghc-devs <ghc-devs@haskell.org>:
On 7/24/26 4:12 AM, Simon Jakobi via ghc-devs wrote:
> I'm not sure if there have been any problems with LLM-generated
> reviews so far? My suggestion would be to drop the responsibility
> requirement for clearly marked LLM-generated MR reviews.

The linux kernel has been using an LLM patch review system called
"sashiko".  It definitely finds bugs, although not everything it
highlights is a real bug. There has been some debate about who should
pay the cost of checking its output -- patch approvers (maintainers) or
patch submitters -- and whether a response to the review by submitters
is optional or required.  Sashiko sometimes find bugs in adjacent code,
but submitters don't want to have to fix bugs they didn't create.

See: "Debating the role of large language models in the kernel
community" -- there's a section called "Email from Sashiko"

      https://lwn.net/SubscriberLink/1083275/d5518174ec4dcbcb/

-BenRI

_______________________________________________
ghc-devs mailing list -- ghc-devs@haskell.org
To unsubscribe send an email to ghc-devs-leave@haskell.org