feat: add support for reasoning field in chat completions

[shitiandmw]

Background

As reported in #1070 by @LubyRuffy, different OpenAI-compatible providers use different field names for reasoning content:

• DeepSeek: uses reasoning_content
• LM Studio, Ollama, and other LLM providers: use reasoning

This PR provides transparent support for both formats without breaking existing implementations.

Changes

• Modified ChatCompletionMessage.UnmarshalJSON to support the reasoning field
• Added custom UnmarshalJSON method for ChatCompletionStreamChoiceDelta
• Both reasoning_content and reasoning fields map to the same ReasoningContent property
• Priority handling: reasoning_content takes precedence over reasoning (backward compatible)
• Updated documentation comments to reflect support for both styles

Implementation Details

The implementation uses custom unmarshaling with a fallback mechanism:

1. First attempts to unmarshal reasoning_content (existing behavior)
2. If empty, falls back to reasoning field
3. Always exposes data through the unified ReasoningContent field

This approach ensures:

• ✅ Zero breaking changes - existing code continues to work
• ✅ Transparent to users - no API changes required
• ✅ Performance optimized - single JSON unmarshal operation
• ✅ Compatible with all providers - works with DeepSeek, LM Studio, Ollama, etc.

Testing

Added comprehensive test coverage in chat_reasoning_test.go:

• ✅ DeepSeek style (reasoning_content) support
• ✅ OpenAI style (reasoning) support
• ✅ Priority handling when both fields are present
• ✅ MultiContent message support
• ✅ Real-world streaming response parsing
• ✅ Marshal/Unmarshal consistency

All existing tests pass without modification.

Example Usage

// Works with both DeepSeek and LM Studio responses
stream, err := client.CreateChatCompletionStream(ctx, req)
for {
response, err := stream.Recv()
// ReasoningContent will be populated regardless of provider
reasoning := response.Choices[0].Delta.ReasoningContent
}## Breaking Changes
None. This is a fully backward-compatible enhancement.

Checklist

• Tests added and all passing
• No linter errors
• Maintains backward compatibility
• Follows Go conventions and project style
• Documentation updated (code comments)
• Addresses reported issue Add support for reasoning field from gpt-oss models while maintaining backward compatibility #1070

[codecov]

Codecov Report

❌ Patch coverage is 95.83333% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 99.52%. Comparing base (5d7a276) to head (c980c5d).

Files with missing lines	Patch %	Lines
chat_stream.go	90.90%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1092      +/-   ##
==========================================
- Coverage   99.59%   99.52%   -0.07%     
==========================================
  Files          34       34              
  Lines        2209     1887     -322     
==========================================
- Hits         2200     1878     -322     
+ Misses          6        5       -1     
- Partials        3        4       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.