The content field in a user message may be an array instead of a plain string. This case is handled properly by llama.cpp and all OpenAI-compatible public services I tried. More specifically, I encountered this issue when trying to use llamafile with Smart Composer Obsidian extension which formats messages in such an array.

  {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Some user prompt."
        }
      ]
    }

With these changes, both server versions (regular and --v2) work properly with the above format.