Async models should be able to execute tools

[simonw]

This is not quite the same thing as:

• Sync and async models should be able to call async def tool functions #987

This issue is about making it so you can use llm.get_async_model() and then await model.chain("do something", tools=[...]) - whether or not those tools= are sync or async functions.

(There's nothing wrong with a sync tool function if it runs fast enough not to block the event loop. I'm going to leave it to users to wrap their long-running functions in an async threadpool and pass them in as async def callables there if they need to.)

[simonw]

I think I'll need to imitate the prev_response.text_or_raise() pattern for tool_calls() here too:

llm/llm/default_plugins/openai_models.py

Lines 527 to 530 in e47bdfe

	messages.append(
	{"role": "assistant", "content": prev_response.text_or_raise()}
	)
	tool_calls = prev_response.tool_calls()

That pattern exists so that this non-async method can still access information that might require an await response.tool_calls() call otherwise. Here's that pattern:

llm/llm/models.py

Lines 816 to 819 in e47bdfe

	def text_or_raise(self) -> str:
	if not self._done:
	raise ValueError("Response not yet awaited")
	return "".join(self._chunks)

This then needs to be used by OpenAI and Anthropic and Gemini.