Python: [Bug]: AnthropicClient ignores allow_multiple_tool_calls=False when tool_choice is unset (incl. every loop iteration after tool_choice="required") #8979

Description

Anthropic only lets you turn off parallel tool use through tool_choice.disable_parallel_tool_use. RawAnthropicClient maps allow_multiple_tool_calls onto that field only inside the tool_choice branch. When tool_choice is None, _prepare_tools_for_anthropic returns early and the setting is dropped (it was already popped from the request in _prepare_options). There is no warning or error.

tool_choice is None in two common cases:

1. Calling the client directly with tools and allow_multiple_tool_calls=False without setting tool_choice.
2. Every model call after the first one in the function-invocation loop when tool_choice="required", because _reset_required_tool_choice sets options["tool_choice"] = None after the first iteration. The first request has {"type": "any", "disable_parallel_tool_use": true}. Later requests have no tool_choice at all, so Claude can emit parallel tool_use blocks again.

• Expected: every request that sends tools also sends disable_parallel_tool_use: true when allow_multiple_tool_calls=False (Anthropic's default mode is auto, so {"type": "auto", "disable_parallel_tool_use": true}).
• Actual: tool_choice is omitted and parallel tool use stays enabled.

The OpenAI clients map allow_multiple_tool_calls to parallel_tool_calls whenever tools are present, regardless of tool_choice.

Code Sample

from unittest.mock import MagicMock
from agent_framework import Message, tool
from agent_framework_anthropic import AnthropicClient

@tool(approval_mode="never_require")
def get_weather(city: str) -> str:
    """Get the weather."""
    return f"sunny in {city}"

client = AnthropicClient(api_key="x", model="claude-sonnet-4-5", anthropic_client=MagicMock())
run_options = client._prepare_options(
    [Message(role="user", contents=["Weather in Paris and Rome?"])],
    {"tools": [get_weather], "allow_multiple_tool_calls": False},
)
print(run_options.get("tool_choice"))  # None -> parallel tool use not disabled

For the loop case, run get_response(..., options={"tools": [get_weather], "tool_choice": "required", "allow_multiple_tool_calls": False}) against a mocked beta.messages.create that returns tool_use twice. Request 1 has tool_choice={"type": "any", "disable_parallel_tool_use": True}. Requests 2 and 3 have no tool_choice.

Root cause

python/packages/anthropic/agent_framework_anthropic/_chat_client.py:1146-1148:

# Process tool choice
if options.get("tool_choice") is None:
    return result or None

allow_multiple_tool_calls is only read after this early return. It is also removed from the request at :701, so nothing else forwards it. The loop reset happens in python/packages/core/agent_framework/_tools.py:4380 (_reset_required_tool_choice).

Proposed fix

When tool_choice is unset, allow_multiple_tool_calls is set and tools are present, send {"type": "auto", "disable_parallel_tool_use": not allow_multiple_tool_calls}. This is Anthropic's default mode, so nothing else changes. Requests without tools or without allow_multiple_tool_calls are unchanged. The fix stays in the Anthropic option mapping; the function-invocation loop is not touched. PR with regression tests is ready.

Code Sample

Error Messages / Stack Traces

Package Versions

agent-framework-anthropic: 1.0.0b261002, agent-framework-core: 1.20.0 (main @ e9857bd)

Python Version

Python 3.12

Additional Context

No response