I have been using Composer since it was Cheetah and I have been very impressed with the model, however today since the update I have noticed it is now asking the user to do things it has been prompted to do where previously it has just automatically done them?
Steps to Reproduce
For example I have a prompt that asks it to use the browser tool and a set of instructions on how to run the tests on the item it has implemented, the model will instead respond with something like:
Upload a PDF and check the OCR worker logs to see the streaming output, and verify that RAG processing completes successfully.
Which is basically verbatim what I have asked the model to do using the browser tool - it pushes this back to the user to complete.
Expected Behavior
It should do what it did the first time and run these tests itself!
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
So ive been using this for the last few hours - compared to yesterday there is something completely different with this model, it is hesistant to iterate, it keeps pushing things back on the user, its not checking its work as it previously did and its ability to follow instructions seems at a claude level (aka will do it if you specifically tell it to do it in its own message).
I also notice that tool calling is failing more often than it has done previously.
Nothing has changed! If you are using very long conversations, I might recommend starting a new one with fresh context, or disabling MCP servers if you have many (even > 20 could impact context usage)
I dont have any MCP servers enabled in Cursor.
Something has changed in the last update - if its not the model then its the interaction with the model.
I’ll keep going and see if I can give any more examples but it is definitely not feeling as smooth and efficient as the past few days!
So this looks entirely time based - For the past few hours there has been a noticeable improvement in how it operates - and now it is just starting to misbehave again.
These are relatively low context usage (under 50% when iterating towards the end).
It almost feels like it has an inbuilt limit of how many iterations it will go through by itself before it feels like it has to stop despite not being finished and give a summary of “What is next”.