This directory contains examples demonstrating how to use the OpenAI-compatible API with Claude's thinking blocks feature through the proxy server.

Thinking blocks are a special feature in Claude that captures the AI's reasoning process. They are formatted as:

<thinking signature="cryptographic_signature">
  The AI's internal reasoning process...
</thinking>

The proxy server preserves these blocks in multi-turn conversations, allowing you to:

• See how the AI reasons through problems
• Verify the authenticity of thinking blocks via signatures
• Maintain context across conversation turns

1. Install dependencies:

   pip install openai

2. Set your API key:

   export ANTHROPIC_API_KEY="your-anthropic-api-key"

3. Start the proxy server:

   uv run python main.py

A minimal example showing basic thinking block functionality:

python examples/simple_thinking_demo.py

Features:

• Basic thinking block parsing
• Multi-turn conversation
• Token usage tracking

Comprehensive examples with multiple scenarios:

python examples/openai_thinking_demo.py

Features:

• Streaming mode: See responses in real-time
• Non-streaming mode: Get complete responses
• Tool use: Thinking blocks with function calling
• Multi-turn conversations: Preserved thinking context

Shows how to get structured JSON responses:

python examples/openai_structured_response_demo.py

The proxy automatically maps OpenAI models to Claude models with thinking enabled:

• o1-preview → Claude with high thinking budget
• o1-mini → Claude with medium thinking budget
• gpt-4 → Standard Claude model

import re

# Pattern to match thinking blocks
thinking_pattern = r'<thinking signature="([^"]*)">(.*?)</thinking>'
matches = re.findall(thinking_pattern, content, re.DOTALL)

# Remove thinking blocks to get visible content
visible_content = re.sub(thinking_pattern, '', content, flags=re.DOTALL)

Always include the full content (with thinking blocks) in the conversation history:

messages = [
    {"role": "user", "content": "First question"},
    {"role": "assistant", "content": full_response_with_thinking},  # Important!
    {"role": "user", "content": "Follow-up question"}
]

1. Streaming: When streaming, thinking blocks appear inline with the content
2. Signatures: Use signatures to verify thinking block authenticity
3. Context: Thinking blocks help maintain context across turns
4. Models: Use o1-* models to enable thinking features
5. Temperature: When using thinking mode, temperature must be set to 1.0

If you get connection errors:

1. Ensure the proxy server is running
2. Check the base URL (default: http://localhost:8000/openai/v1)
3. Verify your API key is set correctly

If thinking blocks don't appear:

1. Use an appropriate model (o1-preview, o1-mini)
2. Check that your prompt encourages reasoning
3. Ensure you're parsing the content correctly