http://bugs.winehq.org/show_bug.cgi?id=60241 --- Comment #36 from Rafał Mikrut <mikrutrafal@protonmail.com> ---
Spending token to get an AI to review the code located at a project URL and then sending the result to the project team, is not ‘doing the work’.
I don't really agree that the amount of manual effort changes whether the report is useful. What matters to me is whether it helps developers find and fix actual problems and in this case I think it clearly does, because as of now, roughly 10-15% of the findings in the report have already resulted in code being fixed by Nikolay, Hans and a few other contributors, who have spent considerable effort looking into these issues.
Anyone interested in AI reports can run AI reports, they don't need to spam Bugzilla unless an actual bug has been found.
When I was more active in Godot, I spent a lot of time finding bugs through testing and static analysis and I regularly used Coverity, Cppcheck and other tools, which I also tried to encourage people to use. In practice, however, very few people used them, partly because the results lived behind a separate service, required additional effort to set up and partly because they often contained a significant number of false positives. That is also why I started to create these reports in the first place and in this particular format(as a simple HTML page - I started with md files, which were quite ugly and hard to read for some people), with a description, the exact location of the problem and possible fixes, to make the whole issue review process as easy as possible for developers - and for multiple projects this worked really well. Just because anyone can create such a report doesn't really mean much, because how many people have actually done it for Wine? This report, for example, identified an issue that was fixed in https://gitlab.winehq.org/wine/wine/-/merge_requests/11876, which had apparently existed since 2004 (git blame https://gitlab.winehq.org/wine/wine/-/blob/153ea52a1d39a3986348a55aff4a2283e...). As for CI, this was generated using Claude Sonnet and I have already seen a lot of projects experimenting with LLM based analysis in CI just like systemd, but this would be quite a big step to integrate it with GitLab. I think that additional static analyzers added to CI are always a good thing, but except PVS-Studio(which is commercial), I haven't really found any static analyzer with a sufficiently low number of false positives. LLM analyzers, despite generating less reproducible findings, have a pretty big advantage over static analyzers, because they can analyze the context of the code. Overall, a large part of the discussion was mainly about whether AI should be allowed to be used in the project or not and in my opinion it should be, as long as its use provides some benefits. If this issue remains open, I will update these reports from time to time, but if you decide that this kind of issue is not welcome in the repository, I will respect that decision and no longer create such reports. Regarding updates to the report, to avoid spamming this thread, I will only update it here when a significant number of the reported issues have been changed or fixed. In the meantime, I will probably make "silent" updates approximately once a week here - https://github.com/qarmin/ClaudeReports/issues/1, to make it easier for multiple people to coordinate fixing the issues(so nobody has to manually remember which issues have already been fixed). The updated report has already been added as a new attachment - "Updated report 12.09.2026" - https://bugs.winehq.org/attachment.cgi?id=82094 -- Do not reply to this email, post in Bugzilla using the above URL to reply. You are receiving this mail because: You are watching all bug changes.