I’m using Task-Master with Cursor in my project which is mostly working great. Unfortunately sometimes when I tell Cursor to begin a subtask, halfway through the task the Agent just loses all context and begins reviewing the task requirements again. It’s like it reaches a context threshold and just stops midway through writing the code.I am using Claude 3.7 Sonnet
This is problematic because it leaves the app’s codebase in a half-baked state that is challenging to recover from. I have to attach to the context all the files that have recently been written or give a list, explain again what we were working on, ask it to review the files and compare with the task in my task-list and continue where it left off. I have to ask the Agent to validate and tell me what it thinks is left to complete so I can confirm it didn’t miss anything.
I understand there’s a limited context, I wonder if something can be done to either warn that a task is greater than the context we have left or if cursor can break up the task into even smaller chunks. I did break up the work before into subtasks and have Cursor read the complexity report from task-master to break down large task.
I’ve attached an image of where you can see the Agent was in the middle of building out UI components for the task and suddenly it got lost and was beginning to start over. I have to watch closely for this as it can begin to demolish my project and I have to start over from my last Git commit. You can see my response at the bottom trying to rebuild context.
I really want to stay with Cursor Pro but I am strongly considering moving to Claude Code as this problem is causing me to lose lots of time to recover from.
In the meantime I’m going to spend more time asking the agent to analyze if it thinks the upcoming task will fit into it’s context limits and break up the tasks further.
I’m on an M1 Macbook Air
Mac OS 15.4.1
Cursor Version is 0.49.6
tonight!