I recently used Cursor 0.47.8 (Claude 3.7 model, without enabling “thinking” mode) to generate client-side interface classes based on existing server-side API logic. While it initially created the classes correctly, it automatically deleted all the newly generated interfaces immediately afterward without any prompt or warning. This behavior is baffling and counterproductive.
The recent updates to Cursor seem to have introduced more regressions than improvements. Instead of boosting productivity, it now wastes time by forcing me to manually recover deleted code or double-check every auto-generated result. The tool feels increasingly unreliable, and this specific issue completely defeats the purpose of using an AI-assisted coding assistant.
Has anyone else encountered similar problems? Are there workarounds or settings to prevent this auto-deletion madness?
yes. I was flowing with cursor about 2 months ago and now it’s a struggle all the time. Even though I have tried to use the newer models they seem worse.
Things I have noticed:
@codebase removal in exchange for manually tagging seems to be worse. I can literally included a couple classes into the context and it one or more of them often leading to random new variables or patterns vs what exists.
AUTO SELECT on by default is killing me - constantly charging me for higher levels of models than I want I have to remember every time I launch cursor to change it away from AUTO. This should be a preference setting not always auto. This is a money grab.
Auto by default seems like the most practical thing to do from a UX perspective, otherwise you’ll have the opposite problem of having to remember to change it for different tasks.
I have found auto-select to not be great lately e.g., auto choosing claude-3.7 for a simple question about context I provide in chat, or running basic terminal commands. It really should follow a bottom-up decision tree of small models > non-premium > premium based on the complexity of the task at hand and preferring models the user has an API key when possible.
If you have a preferential model that you like to default to, set up a custom mode that only uses that model and change the default keybind to open chat to that mode.