Title

Add support for thinking blocks and redacted thinking blocks in Anthropic v1/messages API

Relevant issues

Fixes LIT-1260

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - see details
• I have added a screenshot of my new test passing locally
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem

Type

🆕 New Feature

Changes

This PR adds support for both regular thinking blocks and redacted thinking blocks in the Anthropic v1/messages API experimental pass-through adapter.

What was added:

1. Type Definitions (litellm/types/llms/anthropic.py):

   • Added AnthropicResponseContentBlockThinking for regular thinking blocks
   • Added AnthropicResponseContentBlockRedactedThinking for redacted thinking blocks
   • Updated content union types to include both thinking block types

2. Transformation Logic (litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py):

   • Enhanced translate_anthropic_messages_to_openai() to detect and convert thinking blocks from Anthropic's content array format to OpenAI's thinking_blocks field
   • Enhanced translate_openai_response_to_anthropic() to convert thinking blocks back to Anthropic's content array format
   • Added support for both thinking and redacted_thinking content types
   • Added proper imports for ChatCompletionThinkingBlock and ChatCompletionRedactedThinkingBlock

3. Test Coverage (tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py):

   • Added test_translate_anthropic_messages_with_thinking_blocks_to_openai() for regular thinking blocks
   • Added test_translate_anthropic_messages_with_redacted_thinking_blocks_to_openai() for redacted thinking blocks
   • Both tests verify proper conversion and preservation of all fields

How it works:

Anthropic → OpenAI:

// Input (Anthropic)
{
  "role": "assistant", 
  "content": [
    {"type": "thinking", "thinking": "...", "signature": "..."},
    {"type": "text", "text": "Response"}
  ]
}

// Output (OpenAI)
{
  "role": "assistant",
  "content": "Response",
  "thinking_blocks": [
    {"type": "thinking", "thinking": "...", "signature": "..."}
  ]
}

OpenAI → Anthropic:

• Converts thinking_blocks back to Anthropic's content array format
• Supports both thinking and redacted_thinking types
• Preserves all fields (thinking/data, signature, etc.)

This ensures that the v1/messages API properly retains thinking blocks as required in the OpenAI format while maintaining compatibility with Anthropic's native format.

Before:

After: