Hey, thanks for the detailed report. The Request IDs, error frequency, and stack trace really help.
This looks like an issue with how Cursor formats function calls for the Gemini 3 API. Google’s API rejects them as invalid. There’s a similar bug in the backlog (forum discussion), but that one is specific to the MCP context.
For now:
- Try other models (GPT-5.2, Claude Opus/Sonnet) if you need to work urgently
- If there’s a pattern for when it breaks (specific prompts, context size, task type), share it here. It’ll help us reproduce it
I’ll update the thread if there’s any news.