[Bug] Conversation compaction causes context loss and performance degradation

Bug Description
The compaction system is extremely problematic. Upon compaction it's like it's had a lobotomy—it comes back quite often not really knowing what on earth it's doing, a bit like me when I've been working in code and then I go to sleep and wake up the next morning. Cursor doesn't seem to have this problem.

My guess is that Cursor will compact part of the conversation asynchronously as it's going. So my assumption is that they will take, imagine you've got sort of 20 conversational turns or something, it would take like the back 10 and then it'll split that in half again and then aggressively compact the first 5 and then compact the next 5 and then put them back into the context sort of, you know, flip them out and replace them. That's my guess as to how it works; anyway that's how I would do it.

So yeah, you need to seriously look at the compaction. It's a really big pain.

Additional Feedback

I think you need to change the feedback system from posting on GitHub. Like, often I just want to shoot a message to the team and I assume it'd go into a pile and then your AI will coalesce and surface.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.49
• Feedback ID: 075baced-cc84-464a-89bc-a7e2896334c1