add support for latest Gemini preview model ids#424
add support for latest Gemini preview model ids#424w0wl0lxd wants to merge 1 commit intoBeehiveInnovations:mainfrom
Conversation
Google's current Gemini API exposes 3.1 Pro and 3 Flash preview model IDs that PAL does not recognize yet, which causes valid allowlists to fail during startup. Add the new Gemini registry entries with regression coverage and align the declared Python floor with the mcp dependency requirements so local runs resolve cleanly.
There was a problem hiding this comment.
Code Review
This pull request adds support for new Gemini preview models (gemini-3.1-pro-preview, gemini-3-flash-preview, and gemini-3.1-flash-lite-preview) by updating the model configuration and adding corresponding tests for model restrictions and aliases. It also bumps the minimum Python version to 3.10. A potential issue was identified in the configuration for gemini-3.1-flash-lite-preview, which enables extended thinking but lacks the required max_thinking_tokens field, rendering the feature unusable.
| "description": "Latest Gemini 3.1 Flash-Lite preview for low-cost, high-volume multimodal workloads", | ||
| "context_window": 1048576, | ||
| "max_output_tokens": 65536, | ||
| "supports_extended_thinking": true, |
There was a problem hiding this comment.
The model gemini-3.1-flash-lite-preview has supports_extended_thinking set to true, but it is missing the max_thinking_tokens field. In providers/gemini.py, the thinking budget calculation (lines 196-202) requires max_thinking_tokens > 0. Without this field, the model will claim to support extended thinking but will never actually use it because the budget calculation will be skipped. Given that other 'Lite' models (like gemini-2.0-flash-lite on line 170) do not support this feature, please verify if this should be false or if a max_thinking_tokens value needs to be added.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d14c983829
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| "description": "Latest Gemini 3.1 Flash-Lite preview for low-cost, high-volume multimodal workloads", | ||
| "context_window": 1048576, | ||
| "max_output_tokens": 65536, | ||
| "supports_extended_thinking": true, |
There was a problem hiding this comment.
This new model is marked with supports_extended_thinking: true, but the same object omits max_thinking_tokens. In GeminiModelProvider.generate_content (providers/gemini.py), the thinking config is only sent when max_thinking_tokens > 0, so thinking_mode will be silently ignored for gemini-3.1-flash-lite-preview even though capability metadata advertises thinking support. That can cause inconsistent behavior in auto/reasoning flows that choose this model based on the thinking flag.
Useful? React with 👍 / 👎.
Summary
gemini-3.1-pro-preview,gemini-3-flash-preview, andgemini-3.1-flash-lite-previewGOOGLE_ALLOWED_MODELSallowlists for the new canonical model namesrequires-pythonwith the effective dependency floor so localuv runresolution succeeds consistentlyWhy
Google's current Gemini API docs and examples use the newer 3.1/3.0 preview model ids, but PAL's bundled Gemini registry still only recognizes older entries such as
gemini-3-pro-preview. That causes validGOOGLE_ALLOWED_MODELSconfigurations to fail startup in auto mode because PAL treats the latest official model ids as unknown.Validation
uv run --directory /home/w0w/dev/pal-mcp-server --with pytest pytest tests/test_supported_models_aliases.py tests/test_model_restrictions.pyenv PATH="/etc/profiles/per-user/w0w/bin:/usr/local/bin:/usr/bin:/bin" GEMINI_API_KEY=*** DEFAULT_MODEL=auto GOOGLE_ALLOWED_MODELS="gemini-3.1-pro-preview,gemini-3-flash-preview" LOG_LEVEL=INFO /etc/profiles/per-user/w0w/bin/uv run --directory /home/w0w/dev/pal-mcp-server pal-mcp-server --help