[FEATURE] Support enabling thinking for Sonnet 4/ 3.7 Through LiteLLMModel

[havishammah]

Problem Statement

I would like Strands to support enabling thinking mode for Anthropic models when using them through LiteLLMModel.

By now none of this configs are enabling the thinking mode:

model = LiteLLMModel(
        model_id="hosted_vllm/" + LITE_LLM_MODEL_ID,
        client_args=client_args
        params={
            "temperature": 1,
            "top_p": 0.8,
            "max_tokens": 2048,
        },
        additional_request_fields={
            # Configure reasoning parameters
            "reasoning_config": {
                "type": "enabled",  # Turn on thinking
                "budget_tokens": 3000  # Thinking token budget
            }
        },
        thinking={"type": "enabled", "budget_tokens": 2048}
    )

Proposed Solution

Make one of those config to work:

thinking={"type": "enabled", "budget_tokens": 2048}

or

additional_request_fields={
            # Configure reasoning parameters
            "reasoning_config": {
                "type": "enabled",  # Turn on thinking
                "budget_tokens": 3000  # Thinking token budget
            }
        }

Use Case

When we need Anthropic models to work on thinking mode, retrieving 'reasoningContent' / 'reasoningText'

Alternatives Solutions

No response

Additional Context

No response

[liorheber]

Hi, is there any plan to add support in the coming future?

[pgrayy]

#947

[havishammah]

thanks 🙌