[BUG] LiteLLM Moonshot AI Kimi K2 thinking fails on multi-turn conversations with 'Invalid part type: thinking'

[westonbrown]

Checks

• I have updated to the latest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched issues and there are no duplicates

Strands Version

1.15.0

Python Version

3.10.18

Operating System

macOS (Darwin 25.0.0)

Installation Method

pip install strands==1.15.0

Steps to Reproduce

1. Install Strands and configure Moonshot AI via LiteLLM:

from strands.models import LiteLLMModel

model = LiteLLMModel(
    model_id="moonshot/kimi-k2-thinking",
    client_args={}
)

2. Make initial request with tool calling:

messages = [{"role": "user", "content": [{"text": "Analyze this target"}]}]
response = await model.stream(messages=messages, tool_specs=tools)

3. Add response to conversation history and make second request:

# Response contains reasoning content
messages.append(assistant_message_with_reasoning)
messages.append({"role": "user", "content": [{"text": "Continue"}]})
response2 = await model.stream(messages=messages, tool_specs=tools)

4. See error from Moonshot API

Expected Behavior

Multi-turn conversations with reasoning content should work seamlessly. Reasoning content should be preserved in a format compatible with Moonshot's OpenAI-compatible API.

Actual Behavior

Second request fails with:

MoonshotException - Invalid request: the message at position 2 with role 'assistant' 
contains an invalid part type: thinking

Additional Context

Root cause: In src/strands/models/litellm.py lines 95-102, format_request_message_content() converts reasoning content to:

{
    "type": "thinking",
    "thinking": reasoning_text["text"],
}

Moonshot API only accepts OpenAI-compatible content types: "text" and "image_url". The "thinking" type works with Claude via Bedrock but is rejected by Moonshot and other OpenAI-compatible providers.

Possible Solution

Modify src/strands/models/litellm.py lines 95-102:

if "reasoningContent" in content:
    reasoning_text = content["reasoningContent"]["reasoningText"]
    # Use text type for OpenAI-compatible providers
    return {
        "type": "text",
        "text": reasoning_text["text"],
    }

This preserves reasoning information while maintaining compatibility with all OpenAI-compatible providers (Moonshot, DeepSeek R1, etc.).

Related Issues

Similar to #652 which handled filtering reasoningContent for DeepSeek with Bedrock.