If there are too many lints, Cursor adds all of them to the AI context. This fills the context window right away. Then Cursor summarizes the context, the AI reads the lints again, and the same thing repeats.
This used about 5% of my Ultra plan API usage in a few minutes.
Steps to Reproduce
Open a project with a lot of lint errors
Ask Cursor AI to look at or fix the errors
Cursor adds the lint text to the context
The context fills up
Cursor summarizes it
The AI reads the lints again
The loop repeats
Expected Behavior
I expect my context windows to to blow up instantly.
Hey, thanks for the detailed report. There’s a real gap here: on this path, the linter output isn’t capped, so in a project with lots of lint errors it can fill the whole context window in one go. Then auto-summarization kicks in. It’s designed to turn on as the window gets close to full. After that, the model asks to read the lints again with the same set of errors, and that’s why you get the token-burning loop.
I’ve logged this on our side. I can’t share an ETA for a fix yet, but these workarounds should help a lot:
Narrow the request to specific files instead of fix all lints across the whole project.
Fix fewer diagnostics at a time, in smaller batches.
If it starts looping, start a fresh chat. Don’t keep going in a long thread since the context is already near the limit and summarization will trigger again.
Also, if your repo has lots of rules or large AGENTS.md files, they can bloat the context too. Check Context Explorer to see what’s taking up the window.
Let me know if the loop happens again after trying these steps.
Quick update: I re-checked on 3.22.12. read_lints still returns every diagnostic with no cap. Offset and limit are accepted, but they are not applied, so the summarization loop is still possible on lint-heavy projects. I reopened the internal ticket. For now, the workarounds are to scope lint reads to specific files and fix issues in smaller batches. I’ll update here when a fix ships.