Ideas for working with AI-generated PRs

[hippietrail]

One thing I've noticed that coding bots do, that humans do too but can learn not to, is to go for naïve solutions assuming there are no and can be no false positives or edge cases. And when one is pointed out they will fix just that one and proudly declare the code is now perfect and has no more edge cases. Until you find another one.

I was originally thinking it would be a good idea just to have a page for contributors to read with agent-coding guidelines that would have an emphasis on looking for false positives and edge cases and some tips on how to hunt for them.

But this morning another idea popped into my head.

Why not just do PRs adding unit tests for those edge cases and false positives - and false negatives - to the PRs?

I'm still bad at git but I believe this is possible and that I've done it before. But I don't remember how. So on the assumption that it's not obvious to others, a quick intro on how to do so would surefly be of benefit. I went ahead and asked an AI:

git fetch origin pull/<PR-number>/head:<local-branch-name>
git checkout <local-branch-name>

• Make your changes (e.g., add unit tests)
• Commit and push your changes:

git add <changed-files>
git commit -m "Add unit tests for edge cases in <feature>"
git push origin <local-branch-name>

Opinions?