Audit tool usage as well

[SushantGautam]

choices[].message.tool_calls

For Chat Completions, tool calling is also visible through tool_calls, but it only shows calls the model requested, not every available tool.

For chat completions, mozilla-ai/any-llm returns OpenAI-style ChatCompletion objects. The response message type extends OpenAI’s ChatCompletionMessage, so tool calls are available in the normal OpenAI place:

from any_llm import completion

resp = completion(
    model="gpt-4.1-mini",
    provider="openai",
    messages=[{"role": "user", "content": "What is the weather in Oslo?"}],
    tools=[get_weather],
)

msg = resp.choices[0].message

print(msg.tool_calls)

Tool call names:

for call in msg.tool_calls or []:
    print(call.function.name)
    print(call.function.arguments)

This works because any-llm uses OpenAI-compatible response types for chat completions, including OpenAI tool-call message types.

tracing: https://github.com/arize-ai/phoenix