I feel like I am being rate limited, but don’t know how to “play well.” If I were to play product manager for a moment (which you are the canonical company that doesn’t have them) I would say that as a customer, I think that is a you problem, not a me problem. Note, I say this in a loving way as I really do like Cursor, and it is an incredible product. Specifically, I keep getting, “Connection failed. If the problem persists, please check your internet connection or VPN”.
I clear my chats and even purge my historical ones. From time to time, I restart cursor. Typically clearing the chats/history works, so much so that I created a restart_context.md with all of the rules that I teach cursor each time. That seems to solve the token buffer is too full problem. There was something implemented recently, where I think you all summarized the token window, and that helped, but it also lost some details. So, it feels like you may have rolled it back as I started to get more errors again.
There are times I suspect that it is related to deployments. So, I wait, and it seems to get past it. That too also seems to work when I am barely doing anything and there isn’t much token context.
However, I seem to be rate limited on top of this a lot more recently. In the last week or so, I switched to Run Everything mode. I know there are potential issues with it, but I got to the point where I was done clicking allow or approve exception. At that point, cursor was free and started moving faster through problems, yay! However, I started getting “connection error” messages more often. Now, I can get them within 2 minutes. It is getting super annoying and honestly makes the product unusable. If it is rate limiting, then I would prefer that Cursor just “thinks” longer to naturally slow it down, rather than making me figure out how to work around it. Clearing the buffer is a huge set back and I just burned through a lot of spend because of having to reset so often. Said differently, using cursor to do system administrator tasks is powerful and simplified my life, but they are shorter in duration. So, I think this is going to come up more often if it is what I suspect.
The only other hypothesis is that you are using MCP servers to instrument some of these commands and that is chewing through all of the token window. It is iterating through multiple different approaches quickly, and hence could be eating the token window really quickly.
Anyway, perhaps I am in the minority here, but I feel like this is likely a bug. Would appreciate feedback if it is just me though.