Bug: tool_calls arguments converted to dict breaks chat templates that expect string (str + dict TypeError)

[mauricepruna]

When a conversation history includes prior tool calls (assistant message with tool_calls), process_message_content() in server.py converts arguments from a JSON string
to a Python dict via json.loads(). However, some model chat templates (e.g. DepSeek-R1) perform direct string concatenation with arguments, causing a TypeError: can only concatenate str (not "dict") to str → HTTP 404.

Steps to Reproduce

curl -X POST http://localhost:8081/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mlx-community/DeepSeek-R1-0528-Qwen3-8B-8bit",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "run ls"},
      {"role": "assistant", "content": null, "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "exec", "arguments": "{\"command\": \"ls\"}"}}]},
      {"role": "tool", "tool_call_id": "call_1", "content": "file1.txt"}
    ],
    "tools": [{"type": "function", "function": {"name": "exec", "description": "Run shell", "parameters": {"type": "object", "properties": {"command": {"type": "string"}},
"required": ["command"]}}}],
    "max_tokens": 10
  }'

Root Cause

process_message_content() (server.py ~line 165):

  func["arguments"] = json.loads(args)  # converts string → dict

The DeepSeek-R1 Jinja2 template then does:

  '```json\n' + tool['function']['arguments'] + '\n```'

Fails because arguments is now a dict, not a string.

Fix

Re-serialize back to string instead of leaving as dict:

  if isinstance(args, str):
      func["arguments"] = json.dumps(json.loads(args), ensure_ascii=False)
  elif isinstance(args, dict):
      func["arguments"] = json.dumps(args, ensure_ascii=False)

Environment

• mlx-lm: 0.31.2
• Model: mlx-community/DeepSeek-R1-0528-Qwen3-8B-8bit
• macOS arm64