It's because there are too many for me to go into, so I did a high level scrub with basically this procedure:

• Create a working directory for the issue, generate a spec from the description, run codegen
• Reproduce the reported problem (if it's not clear how to reproduce, don't touch the bug)
• If, given today's code, the bug is no longer reproduced, explain why
• If there is any doubt, leave it open.

I have a spreadsheet tracking all this work.

Sometimes, it's a little bit of a poor fit in terms of the message, but I think I got most of the issues correct. When I generated the sheet, I went over all the bugs, and I thought they were appropriate to close, but yeah, the messages are robotic. I erred on the side of closing a little too aggressively, since some things are questions, and I answered them manually where i could.

The remaining ~300 issues and ~80 PR still need my human attention - and it's a stupendous amount of work, so yeah, I'm using labor saving tech.

There's a huge bug fix release coming shortly, thanks to these tools.

To add a little bit more; I spent several hours building a taxonomy of all the issues, sorting them by category, and whether they had anything actionable for me.

Once I went through all of them, I made a list of ones which could be closed, and had the AI close it. I think I goofed on #2170, but I think I made the right call on the ~100 issues closed and ~30 PR that were superseded (most were ancient)

We have 7 years of issues and PR's. I've got some time to revisit them now and some resources to do it. My (unrealistic) goal is to address every PR and Issue and just have a working set of actively worked issues. Every PR which I made with the help of AI is annotated with the fact that a robot helped write it. I should have put that in the messages for the issues that I closed.