Since about 5 days ago, gpt-5 responses are regularly cut short mid response.
From Cursor’s perspective, this is a complete, successful response.
This is nasty because this might be at the end of a long agentic chain of tool calls and generations with non-idempontent and/or expensive tool calls. The only way forward in Cursor is to restart since the last user message, which might be long ago.
Steps to Reproduce
Use cursor IDE
Expected Behavior
Please add an option to restart generation since the last well-formed tool call, not just the last user message.
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
I’m pretty sure this is OpenAI’s inference problem Cursor can do nothing about, yet the fix that I’ve suggested would be useful anyway.
Besides, often the model goes off the rail with terminal command syntax somewhere mid-rollout and it would also be great to stop right there and intercept with a user message “call such and such tool with such and such syntax”, instead of passively observing the model stumbling. Again, “in the previous paradigm” when model’s rollout contained a single or a few tool calls it would be fine to edit my previous message, but in the new paradigm with models doing dozens of research tool calls before generating terminal commands and stumbling with their syntax (for example) that is too costly.