Opus won't actually DO anything

[dogberry]

I switched Cursor to use claude-3-opus. I told it:

Common helper functions like respondWithError, createXResponse, etc. are reimplemented in multiple handlers. These could potentially be extracted to the util package to reduce duplication. Create a plan for a refactor to address this issue.

It came up with an excellent plan and asked me if I wanted it to do the work. I said yes. So it then started looking at the codebase and basically repeated the plan back to me and said it would start and give me checkpoints. It then repeated the plan to me and asked if I had any question or wanted to start. I said to start.

Then it just repeated the process again of giving me the summary of the plan and asking if I wanted it to start. “Let me know if you have any other suggestions or see any issues with this refactoring plan. Otherwise, I’ll go ahead and start implementing it incrementally. I’ll keep you posted on the progress and any decisions made along the way.”

Telling it, “Yes, go ahead and start implementing it” again just repeats the cycle.

[dogberry]

Interestingly, I asked ChatGPT about this, just for snits and grins, and it told me, point blank:

Agent loops fixed “soon.” Cursor engineers have acknowledged the Apply gap for Opus; a backend patch is reportedly in testing. Until it ships, the loop is normal.

Bottom line: Yes—Sonnet “just works” for code edits, Opus gives a richer high-level plan but can’t yet drive Cursor’s Apply stage, so it ends up talking in circles. That behaviour is expected given the current integration state.