[Feature Request]: Migrate to Gemini_v2

[priyansh4320]

Is your feature request related to a problem? Please describe.

Implement Gemini V2 Client (ModelClientV2)

Why Implement Gemini V2 Client?

The current Gemini client (api_type: "google") uses the legacy ModelClient interface, which returns flattened ChatCompletion responses. This limits access to rich multimodal content and advanced Gemini 3 features.

Current Limitations (V1 Client)

1. Lost Rich Content: Thinking tokens, multimodal content (images, audio, video), and structured data are flattened or lost
2. No Thinking Support: Gemini 3's thinking/reasoning tokens not fully accessible
3. Limited Multimodal: Images, audio, and video content not preserved as structured blocks
4. No Type Safety: Responses are untyped, making it easy to introduce bugs
5. Inconsistent API: Different response formats across OpenAI, Anthropic, Gemini, and Bedrock clients

Benefits of V2 Client

1. Rich Content Preservation

   • All content types (text, images, audio, video, reasoning) preserved as typed ContentBlock objects
   • Thinking/reasoning tokens from Gemini 3 models fully accessible
   • No data loss during response transformation

2. Better Developer Experiencen

   V1 - Manual parsing

   response = client.create(params)
   messages = client.message_retrieval(response)
   content = messages[0] if messages else ""

   V2 - Direct property access

   response = client.create(params)
   text = response.text # All text content
   reasoning = response.reasoning # Thinking tokens
   tool_calls = response.get_tool_calls() # Tool calls as objects
   images = response.get_content_by_type("image") # Image content blocks
   3. Type Safety

   • Pydantic models with automatic validation
   • Typed content blocks (TextContent, ImageContent, ReasoningContent, etc.)
   • IDE autocomplete and type checking support

3. Forward Compatibility

   • Unknown content types automatically handled via GenericContent
   • No code changes needed when Gemini adds new features
   • Future-proof architecture

4. Provider-Agnostic Format

   • Unified response format across OpenAI, Anthropic, Gemini, and Bedrock
   • Easier to switch between providers
   • Consistent developer experience

5. Backward Compatible

   • Can use create_v1_compatible() to get V1 format when needed
   • Works seamlessly with existing V1 clients in group chats
   • No breaking changes to existing code

Use Cases Enabled

• Thinking/Reasoning: Full access to Gemini 3's thinking tokens and reasoning capabilities
• Multimodal Applications: Proper handling of images, audio, and video content
• Structured Outputs: Better support for Pydantic models and JSON schemas
• Tool Calling: Rich access to tool call information with thought signatures
• Cost Tracking: Built-in cost calculation per response

Migration Path

Migration is simple - just change api_type:

V1

{"api_type": "google", "model": "gemini-2.5-pro", ...}

V2 (recommended)

{"api_type": "gemini_v2", "model": "gemini-2.5-pro", ...} # All other configuration parameters remain identical### Advanced Features (Gemini 3 Models)

{
"api_type": "gemini_v2",
"model": "gemini-3-pro-preview",
"thinking_level": "High", # Options: "High", "Medium", "Low", "Minimal"
"thinking_budget": 10000, # Token budget (0 = disabled, -1 = automatic)
"include_thoughts": True # Include thinking tokens in response
}## Implementation Status

✅ Completed: Gemini V2 client implementation with full ModelClientV2 protocol support
✅ Completed: Comprehensive unit tests
✅ Completed: Integration tests
✅ Completed: Documentation updates
✅ Completed: Backward compatibility layer

Related

• Bedrock V2 Client: Similar implementation pattern
• OpenAI V2 Client: Reference implementation
• ModelClientV2 Protocol: autogen/llm_clients/client_v2.py

Describe the solution you'd like

No response

Additional context

No response

[elCaptnCode]

Thanks for the detailed write-up. This is already being worked on in #2349 — tracking progress there.