There were various cases where Cursor servers were overloaded historically or had other issues due to an update (in last 24h).
So likely to have been an effect from that.
Yes you are right on longer chats. For me also when context becomes long from the thread it gets easier to see hallucinations from AI or mistakes in Yolo!
Thats why usually i keep chat focused on either Planning only, or Implementation (Act) only for specific features.
Also when i go back and forth with corrections it can confuse AI (tried several models) as the context is polluted with contradicting code inside.
I use a task.md file that AI writes plan into and updates it. That way i can tell it to focus on specific steps even if i start new chat.
20 000s or 200 000s is like a full timeout, can also be network related.
(hard to tell sometimes but i saw users having issue with their routers which when restarted worked for example, or firewall, antivirus on windows which caused interruptions etc..)
Also here is a related case.