v0.3.7

Bug Fix

• Filter reasoning_content from request messages for provider compatibility
  • vLLM, GLM, and most OpenAI-compatible providers don't recognize reasoning_content field
  • This field was being passed through from Anthropic-format history messages
  • Caused INVALID_ARGUMENT: content is not repeating error when using tool calls with vLLM
  • Now filters unsupported fields before sending to upstream providers

Debug Logging

• Added debug/warning logs for message content conversion
• Helps diagnose "content is not repeating" errors

Technical Details

When preserve_thinking_blocks=True (from thinking.type=enabled), thinking blocks in assistant messages are converted to reasoning_content field. However, most OpenAI-compatible providers (vLLM, GLM, BigModel) don't support this field and reject requests.

The fix filters out reasoning_content from messages before sending to the provider, while still preserving thinking blocks in the response conversion for Anthropic clients.

Commit History

• 9602d0c fix: filter reasoning_content from messages for provider compatibility
• d8d6198 debug: add logging for message content structure

Full Changelog: v0.3.6...v0.3.7