Summarizing chat context - why?

[mwjt42]

No idea what this is for, it never used to require this?

And is clearly broken:

Summarizing chat context eternal loop

[image]

"Summarizing chat context" lingering / weird behvaiour

I get this notification in the chat, although it was a short chat, no context issue yet and the model continues with its tasks, with the “Summarizing chat context” still appearing (as you can see the model continues, also after thinking) Model: Sonnet 4.0 Thinking Pro plan [image]

"Summarizing Chat Content"

Anyone else read this as “leaking context”? It looks like the user is engaged in a complex debugging operation, which I’ll now forget all about since this summary will reboot the context. You should probably copy and paste the last few responses verbatim so that the vague summary doesn’t cause the LLM to go way off scope. I’ll just edit these 3 files now component.tsx +123 -7 And simplify your UI navigation.tsx +17 -371 There, I fixed it.

Summarizing Context Resets context in chat

When using sonnet 3.7 max (which isn’t cheap) cursor attempts to summarize (I assume when it hits a context limit or some other trigger with embeddings). Well when it summarizes and suggests a new chat, most of the time then resets the context so the LLM forgets what I was doing. This is making it so that Cursor increases the work I have to do re-explaining and adding the context back. If this is a dark pattern, it is beyond insidious and if it isn’t then it is a bug that should have been fixe…

[danperks]

Summarising has always happening in Cursor for a very long time, but it’s only recently that we decided to actually show you when this is happening, as it can have a significant effect on your workflow.

When your conversation gets longer than the context window of the model you are using, we have two choices:

1. We cut the top off your chat, and just feed the latest messages and context into the model
2. We tell you you’ve reached the limit and force you to start again
3. We summarise the oldest messages so the model still has a high level understanding of what happened, while still being able to continue your chat

We’ve always chosen option 3, and our summarisation strategy has gotten better over time. With this latest release, summarisation should be our best yet, keeping the nuance of what you’ve worked on so far, but condensing it down for the model to use!

If you’re seeing any freezing, I’d encourage you to post a bug report, ideally checking the contents of the console (by running the Developer: Toggle Developer Tools command), as there will likely be an error we need to fix!

[mwjt42]

Thanks Dan, that information is much appreciated.

Is it possible to use a progress bar, or % count or something instead?
A spinner does not instil the user with confidence that the process will end!

Also, what’s best (user) practice here? If it’s to ideally start a new chat where possible, a little prompt next to ‘summarizing context’ informing us the chat has exceeded / may exceed the current context window, and to consider starting a new chat, might ensure some better behaviour on the user’s part.

[danperks]

It really depends on your usage of Cursor and the chat.

Many of us in the team frequently start new chats, any time we move to a new area of the codebase or start work on a new feature.

Others like to stick in one, huge chat that summarises often.

The formal is usually the best way to keep context concise and not too verbose, but either way can work if you lean into the behaviours.

I believe you will see a prompt to start a new chat if you’ve seen a lot of summarisation in the same chat (but don’t quote me on that!)