support openai compatible model reasoning content in streaming response

[UnderTreeTech]

support reasoning content in streaming response, therefore user can custom AfterModelCallback to filter thought response during streaming llm call.

[sjy3]

When the thinking mode is activated, the web page will generate multiple messages. If you are willing to fix the above issue, please resubmit a pull request (PR).

[UnderTreeTech]

@sjy3 Add a standard AfterModelCallback implementation that filters thought content. This allows flexible control over thought visibility:

func ThoughtFilterCallback(ctx agent.CallbackContext, llmResponse *model.LLMResponse, llmResponseError error) (*model.LLMResponse, error) {
    if llmResponseError != nil || llmResponse == nil || llmResponse.Content == nil {
        return nil, nil
    }

    var filteredParts []*genai.Part
    hasThought := false

    for _, part := range llmResponse.Content.Parts {
        if !part.Thought {
            filteredParts = append(filteredParts, part)
        } else {
            hasThought = true
        }
    }

    if hasThought {
        newResponse := *llmResponse
        newResponse.Content = &genai.Content{
            Role:  llmResponse.Content.Role,
            Parts: filteredParts,
        }
        return &newResponse, nil
    }

    return nil, nil
}

Usage:

agent, err := llmagent.New(llmagent.Config{
    Name: "MyAgent",
    Model: model,
    AfterModelCallbacks: []llmagent.AfterModelCallback{
        ThoughtFilterCallback,  // Filter thought content
    },
})

Benefits

• Flexibility: Can be enabled/disabled per agent by including/excluding the callback
• Consistency: Provides a standard approach across different OpenAI-compatible providers
• No Breaking Changes: Fully opt-in via the existing callback mechanism
• Frontend/Backend Choice: Decision to filter can be made at either layer

1. Backend filtering: Add the callback to filter thought content before sending to frontend
2. Frontend filtering: Don't use the callback, let frontend decide whether to display thought content based on part.Thought flag