Environment

• Hermes commit: e493b1c48
• Model: moonshotai/kimi-k2.6 (NVIDIA)
• Platform: Telegram
• Config:

custom_providers:
  - provider: nvidia
    name: "NVIDIA AI"
    api_key: "..."
    api_base: "https://integrate.api.nvidia.com/v1"
    models:
      "moonshotai/kimi-k2.6":
        max_tokens: 4096
        chat_template_kwargs:
          thinking: true

Problem Description

When using Kimi K2-6 with chat_template_kwargs.thinking: true on NVIDIA API:

• Telegram receives a response with content: null
• The thinking/reasoning content is not displayed
• Manually calling the API confirms reasoning_content IS returned by NVIDIA

Manual API Verification (curl)

curl -X POST "https://integrate.api.nvidia.com/v1/chat/completions" \
  -H "Authorization: Bearer <key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/kimi-k2.6",
    "messages": [{"role": "user", "content": "1+1=?"}],
    "chat_template_kwargs": {"thinking": true},
    "max_tokens": 50
  }'

Response includes:

{
  "choices": [{
    "message": {
      "role": "assistant",
      "content": null,
      "reasoning_content": "...thinking text...",
      "tool_calls": null
    }
  }],
  "usage": {
    "prompt_tokens": 17,
    "completion_tokens": 50,
    "reasoning_tokens": 50
  }
}

Root Cause Analysis (by line number reference)

Looking at gateway/run.py:

• Line 10478: reasoning_content=msg.get("reasoning_content") — extracted from message
• Lines 13765-13768: Comment notes "The agent's _build_api_kwargs converts these to the provider-specific format (reasoning_content, etc.)"
• Line 13770-13773: Preserves reasoning, reasoning_details, codex_reasoning_items on assistant messages for multi-turn context

However, when displaying the response to the user, only content seems to be used, resulting in an empty message when content: null.

Proposed Fix

When content is null but reasoning_content exists, display the reasoning content instead. Alternatively, add a display.show_reasoning config option to prepend or display reasoning alongside the final answer.

Workaround

Disable thinking mode (thinking: false) to get normal content responses.