From using other AI tools, I expect context management commands like /summarize to just summarize and then stop. (See: Claude Code’s /compact command.) In the cursor GUI, /summarize just kicks the agent off again to go do… whatever. By sending an empty message. Basically equivalent to “please continue”. The problem is now it doesn’t have any context, so how does it even know what to do?
Steps to Reproduce
Do some work, notice context window growing, find a good stopping point and then run /summarize. Observe that the agent now just continues where it left off but without any of its context.
Expected Behavior
This should do one of two things:
Summarize and stop, await user input.
/summarize “some new prompt” and start running the new prompt after summarization is complete.
Operating System
Windows 10/11
Linux
Current Cursor Version (Menu → About Cursor → Copy)
Sorry but I’m unable to turn off privacy mode. It’s an org-wide toggle and we have security requirements that it stay on. If there was a way to run a one-off non-private command, I could do that.
If you see this, then it’s working as intended. Summarization happens in the background, the whole context gets compressed, and the chat still has the previous context.
Oh, that’s intended? It’s kind of tricky UX. Usually I want to summarize followed by a new prompt. I’ve just been stopping it after the summary and adding the new prompt, lest it continue on doing whatever it was doing before.
I don’t know, this still feels like a bug to me. The models get confused about the empty query and just start doing random stuff that I don’t want them to do. This example is with Claude but GPT-5 is similar.
I want the ability to /summarize to manage token usage without sending a new query at all.
