gemini-openai-api_skill

Overview

This skill documents how to integrate Gemini models through an OpenAI-compatible API for use as an auxiliary model. It covers recommended models for common tasks, how to control Gemini 'thinking' with extra_body, and how to normalize Gemini responses that may be wrapped in Markdown code blocks. The guidance is focused on practical configuration and parsing steps for JavaScript or Python OpenAI-compatible clients.

How this skill works

The skill explains the OpenAI-compatible base URL and maps task types to Gemini models (summary/correction/vision vs. emotion). It shows how to pass Gemini-specific runtime controls by nesting a google object inside an outer extra_body parameter, and how to detect and strip Markdown code fences around JSON responses. It also points to where to add provider and extra_body mappings in application config files.

When to use it

• When you need to add Gemini as a secondary/assist model behind an OpenAI-compatible client.
• When you must control or disable the model's internal 'thinking' behaviors (Gemini 2.5+).
• When Gemini responses include JSON wrapped in Markdown code blocks and need parsing.
• When configuring provider entries and model mappings in application config files.
• When selecting a task-appropriate Gemini model (summary, emotion, vision).

Best practices

• Use the base URL https://generativelanguage.googleapis.com/v1beta/openai/ for OpenAI-compatible requests.
• Always wrap Gemini runtime settings as {"extra_body": {"google": {...}}} — the outer key is required by the client.
• Choose gemini-3-flash-preview for summary/correction/vision and gemini-2.5-flash for emotion analysis.
• Disable or lower 'thinking' for predictable, deterministic outputs (set thinking_budget=0 or thinking_level='low').
• Handle Markdown code fences before JSON parse: strip leading/trailing ``` blocks and then JSON.parse.

Example use cases

• Adding Gemini to a multi-provider assistant where Gemini handles summarization and vision tasks.
• Setting thinking_budget=0 for swift short responses from gemini-2.5-flash in realtime UIs.
• Lowering include_thoughts and thinking_level for gemini-3-flash-preview to avoid internal chains of thought.
• Preprocessing model output to remove ```json code fences before converting to structured data.
• Registering gemini in config/api_providers.json and mapping EXTRA_BODY_GEMINI in config/__init__.py.