Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
This works, but boy is it slow. I'm talking several minutes with no response. We might want to adjust the thinking-level (like we did with gpt-5). |
|
@cursoragent can you update this so that we turn on medium reasoning for gemini-3.1-pro. High reasoning is too slow. You can see an example of how we do this for gpt-5.2 in the openai client or open ai client utils |
|
I've implemented medium reasoning for Changes Made:
All existing tests pass, and the new tests verify the functionality works correctly. This should reduce the response time significantly compared to the HIGH reasoning level that was causing multi-minute delays. |
- Add thinking_config with thinking_level='MEDIUM' for gemini-3.1-pro-preview in get_gemini_completion_function_params (similar to reasoning_effort for gpt-5.2) - Update GeminiClient to use ThinkingConfig in both request_completions and stream_completions methods - Update _prepare_gemini_request_data_and_headers to include thinking_config in server requests - Add tests for thinking_config functionality Co-authored-by: Aaron Diamond-Reivich <aarondr77@gmail.com>
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
Single source of truth - get_gemini_completion_function_params sets thinking_config, and it's passed through to server functions rather than duplicating the model check. Co-authored-by: Aaron Diamond-Reivich <aarondr77@gmail.com>
Match OpenAI pattern - only set thinking_config in get_gemini_completion_function_params. The Mito server will handle thinking_config based on the model it receives. Co-authored-by: Aaron Diamond-Reivich <aarondr77@gmail.com>
Remove thinking_config from get_gemini_completion_function_params entirely. Check model == GEMINI_3_1_PRO_MODEL directly when building GenerateContentConfig. Co-authored-by: Aaron Diamond-Reivich <aarondr77@gmail.com>
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 2 potential issues.
Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
| model=model, | ||
| contents=contents, | ||
| message_type=message_type, | ||
| ) |
There was a problem hiding this comment.
Unused provider_data variable in streaming method
Low Severity
The provider_data variable is computed via get_gemini_completion_function_params but never referenced anywhere in stream_completions. The comment says "Get provider data for thinking config," but the thinking config is handled inline via a direct model == GEMINI_3_1_PRO_MODEL check instead. This is dead code that adds unnecessary computation and confusion about intent.
| response_mime_type=provider_data.get("config", {}).get("response_mime_type"), | ||
| response_schema=provider_data.get("config", {}).get("response_schema") | ||
| response_schema=provider_data.get("config", {}).get("response_schema"), | ||
| thinking_config=ThinkingConfig(thinking_level="MEDIUM") if model == GEMINI_3_1_PRO_MODEL else None |
There was a problem hiding this comment.
Thinking config missing for Mito server fallback path
Medium Severity
The ThinkingConfig(thinking_level="MEDIUM") is only applied in the direct API key path (self.api_key). Users without a Gemini API key who fall back to the Mito server don't receive the thinking config—neither via provider_data config nor as a separate parameter. The OpenAI equivalent (reasoning_effort for GPT 5.2) is passed through get_open_ai_completion_function_params so it reaches the Mito server, but the Gemini thinking config has no such mechanism.
Additional Locations (1)
3 participants
Description
Gemini 3.1 pro support!
Testing
Try it out!
Documentation
Note if any new documentation needs to addressed or reviewed.
Note
Medium Risk
Moderate risk because it changes Gemini API request/stream configuration (adds
ThinkingConfig) and updates model identifiers across backend and frontend, which could affect runtime behavior if the SDK/model name expectations differ.Overview
Adds support for the
gemini-3.1-pro-previewmodel across the stack, replacing references to the oldergemini-3-pro-previewin model selection and UI constants.When calling Gemini directly (API key path), requests and streaming now conditionally include
ThinkingConfig(thinking_level="MEDIUM")forgemini-3.1-pro-preview, while other models remain unchanged. Tests are updated to reflect the new model name and fast-model fallback behavior.Written by Cursor Bugbot for commit 6a40b14. This will update automatically on new commits. Configure here.