Add thought signature support for Google Gemini OpenAI compatibility

[ericproulx]

Summary

This PR adds support for preserving and transmitting thought signatures when using Google Gemini models through the OpenAI-compatible API. This enables seamless integration with Gemini's extended reasoning capabilities while maintaining compatibility with the OpenAI API format.

Background

Google Gemini models support extended reasoning through "thinking" modes, which generate thought signatures that need to be preserved when working with tool calls. When using Gemini models through an OpenAI-compatible endpoint, these thought signatures are transmitted in the extra_content.google.thought_signature field, following the OpenAI Realtime API's extension pattern.

Changes

Provider Implementation

• lib/ruby_llm/providers/openai/tools.rb:
  • Added extract_tool_call_thought_signature method to extract signatures from incoming tool calls
  • Modified parse_tool_calls to preserve thought signatures in ToolCall objects
  • Updated format_tool_calls to include extra_content.google.thought_signature when present

Testing

• spec/ruby_llm/providers/open_ai/tools_spec.rb (new file):
  • Comprehensive test coverage for parsing tool calls with and without thought signatures
  • Tests for formatting tool calls with thought signature preservation
  • Tests for the signature extraction helper method
  • Edge case handling (nil, empty, missing fields)

Technical Details

Incoming Tool Calls (Parsing)

When receiving tool calls from Gemini through the OpenAI-compatible API:

{
  'id' => 'call_456',
  'function' => { 'name' => 'weather', 'arguments' => '{"location":"Paris"}' },
  'extra_content' => {
    'google' => { 'thought_signature' => 'sig_abc123' }
  }
}

The signature is now extracted and stored in the ToolCall object.

Outgoing Tool Calls (Formatting)

When sending tool calls back to the API, thought signatures are preserved:

{
  id: 'call_456',
  type: 'function',
  function: { name: 'weather', arguments: '{"location":"Paris"}' },
  extra_content: {
    google: { thought_signature: 'sig_abc123' }
  }
}

Why This Matters

1. Google Gemini Compatibility: Enables full support for Gemini's thinking modes when using OpenAI-compatible endpoints
2. Reasoning Preservation: Maintains the connection between reasoning steps and tool calls
3. API Standard Alignment: Follows the OpenAI Realtime API's extra_content extension pattern
4. Backward Compatibility: Tool calls without thought signatures continue to work exactly as before

Backward Compatibility

✅ Fully backward compatible:

• Tool calls without thought signatures are handled identically to before
• No changes to the public API
• All existing tests continue to pass
• Only adds optional signature handling

Testing

• ✅ New comprehensive test suite for OpenAI tools module
• ✅ Tests for both parsing and formatting with/without signatures
• ✅ Edge cases handled (nil, empty inputs, missing fields)
• ✅ Existing functionality preserved

Related Documentation

This change aligns with:

• OpenAI Realtime API extra_content pattern
• Google Gemini's thinking mode capabilities
• RubyLLM's existing tool call infrastructure