Scripts over scripts under scripts inline scripts non-reusable scripts, echo a simple text scripts, mass edit files with no recourse scripts.
This is a total breaker, nothing works ok, it uses tons of tokens, runs weird things and those are malformed and incomplete, files are changed with no control, management or recourse, and the Agent chat has no understanding of what changed, what was before and after, because it doesn’t hit the context.
THIS IS COMPLETELY AGAINST THE CONCEPT OF AN IDE EDITOR, WHICH CURSOR IS AND HAS STARTED PROMOTED AS SUCH.
Steps to Reproduce
Select Auto, ask for simple tasks
Expected Behavior
Reply with normal text like a normal LLM assistant, edit files precisely and punctually with Accept/Reject control.
Thanks @vibe-qa. In both these cases, privacy mode is still enabled. You can disable it (and turn it back on after) under Cursor Settings > General > Privacy.
7b71cc38-237b-4859-9883-77581a05579d - working example, 500k tokens, review available (0.25$)
8b1b573d-bdee-49e6-85fd-3f16e35fb955 - off the rails, 1.3M tokens, review missing (0.46$)
A single shell command poisoned the rest of the inference, making the approach unstable, error-prone, wasteful, expensive and unable to rollback or review the changes. This time it appears to have been successful but usually it stumbles scripting more and more until chat compaction happens then it’s never successful.
