JSON mode and tools do not mix

The Ollama case

Ollama #18944 (Ollama 0.40.3): with "format": "json" on /api/chat, the model cannot call tools. It writes the call as JSON text in content instead:

{
  "message": {
    "role": "assistant",
    "content": "{\"name\": \"time\", \"parameters\": {}}"
  }
}

No tool_calls field. If your code only checks tool_calls, the tool never runs and the user sees raw JSON.

The llama.cpp case

llama.cpp #30321: when a JSON schema lists an optional field before a required one, the grammar only allows required fields first. Many models then skip the optional field. The JSON is still valid, so nothing warns you. llama.cpp #30256 (lesson 02) is another JSON schema problem: thinking leaks in.

The rule

• In a step that may call a tool, send tools and no format or response_format.
• In a step that must return data, send a schema and no tools.
• In schemas, list required fields first. Check optional fields after parsing and log when they are missing.
• If you need both, make the tool return the data. Then the tool call is the structured output.

Check it with a test

import json
import urllib.request

def chat(body):
    req = urllib.request.Request("http://127.0.0.1:11434/api/chat", json.dumps(body).encode(),
                                 {"Content-Type": "application/json"})
    return json.load(urllib.request.urlopen(req))

tool = {"type": "function", "function": {"name": "get_time", "description": "Get the current time.",
        "parameters": {"type": "object", "properties": {"zone": {"type": "string"}}, "required": ["zone"]}}}
base = {"model": "qwen3:1.7b", "stream": False, "tools": [tool],
        "messages": [{"role": "user", "content": "Use your tool to get the time in UTC."}]}

for fmt in (None, "json"):
    body = dict(base, **({"format": fmt} if fmt else {}))
    msg = chat(body)["message"]
    print("format =", fmt, "| tool_calls:", bool(msg.get("tool_calls")), "| content:", msg.get("content", "")[:60])

Output on the test box, Ollama 0.40.4, Qwen3 1.7B Q8_0:

format = None | tool_calls: True | content: 
format = json | tool_calls: False | content: {"name": "get_time", "arguments": {"zone": "UTC"}}

So the problem from #18944 is still there in 0.40.4. Pull the model first with ollama pull qwen3:1.7b. If tool_calls is false when format is json, your build has the problem from #18944. Keep this test in CI.