When doing a possibly longer-running task with Claude Sonnet 4 thinking, e.g. analysing and fixing test issues, which possibly have a lot of output and thus fill up the context window quickly, there is a an issue with the context summarization. I did also see this behaviour with GPT-5 so it doesn’t seem to be strictly connected to Sonnet 4 Thinking.
“Summarizing context” steps seems to run indefinitely, both indicated visually by the shimmering of the action in the chat as well as visible in the resource consumption as my GPU and CPU utilization stay at a constant ~35% even after the agent has moved on past the summarization step and even after the full response has been completed.
When there are a lot of summarization steps (e.g. in a very long running interaction where the context limits are hit multiple times) this seems to stack and lead to extreme performance issues (extreme lagging) that make Cursor basically unusable without moving to a new chat (which is not a valid option IMO since I want to review the changes made before them being accepted automatically by opening a new chat).
When creating a new chat (CMD + N) the utilization immediately drops to normal idle usage and also the text from the “Summarizing context” step disappears and only the magic wand icon is visible.
Steps to Reproduce
Start a conversation that probably hits the context window limit and leads to summarization.
Expected Behavior
The GPU and CPU usage should drop after the summarization step is done and it should not have this extreme impact on performance.
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
Still waiting for cursor to finish “Summarizing chat context“ - after waiting ~5 Minutes I have sent the same task again, it summarizes in ~15 seconds and worked fine. Even after the result has been returned completely, I still see the animated message “Summarizing chat context“ directly below my already answered chat input when I scroll up a little in the chat.:
Current token usage in that chat: “~37% of 200k tokens”
I noticed this problem frequently since a few days. Also started new chat, but after some work, the never-ending summarization starts again. I use claude-4-sonnet explicitly (in Agent mode) for my current tasks.
