[upstream #17200] #17200: Explicit Gemini thinking levels are dropped on Gemini-compatible custom endpoints #391

On a custom endpoint with customParams.defaultParamsEndpoint: google, getGoogleConfig builds the selected thinkingConfig.thinkingLevel, but transformToOpenAIConfig drops it through googleExcludeParams (packages/api/src/endpoints/openai/transform.ts) and never sends reasoning_effort. The gateway then uses the model's default level, so a user's low or high choice has no effect.

This is on dev for every Gemini 3+ model on that path (3.5 to 3.8 Flash, 3.1 Pro), not only Gemini 4 Argon. Google's OpenAI compatibility maps reasoning_effort to thinking_level (Argon: minimal and low to LOW, medium to MEDIUM, high to HIGH; https://ai.google.dev/gemini-api/docs/openai), so the fix is to translate an explicit thinkingLevel into reasoning_effort on that path, and to send nothing when it is unset so the model default applies.

Raised by Codex on LibreChat-AI#17200 (the Gemini 4 Argon support PR); deferred because the code path predates that PR and affects every Gemini 3+ model.