Python: Handoff workflow + Ollama: allow_multiple_tool_calls passed to AsyncClient.chat()

[danielpcampagna]

Description

When using HandoffBuilder with OllamaChatClient, the workflow crashes because the handoff orchestrator injects allow_multiple_tool_calls=False into agent default options, and OllamaChatClient._prepare_options() passes it through to ollama.AsyncClient.chat(), which doesn't accept that keyword.

Error

agent_framework.exceptions.ChatClientException: ("Ollama streaming chat request failed : AsyncClient.chat() got an unexpected keyword argument 'allow_multiple_tool_calls'", TypeError("AsyncClient.chat() got an unexpected keyword argument 'allow_multiple_tool_calls'"))

Root Cause

Two locations contribute to this bug:

1. agent_framework_orchestrations/_handoff.py (orchestrations layer)

HandoffAgentExecutor._clone_chat_agent() unconditionally sets allow_multiple_tool_calls=False at line 286, assuming all clients support this option:

cloned_options["allow_multiple_tool_calls"] = False

2. agent_framework_ollama/_chat_client.py (Ollama connector)

OllamaChatOptions documents this option as "Not supported. Not configurable in Ollama." (lines 216-217), and _prepare_options() skips None values (line 409). However, False is a non-None boolean, so allow_multiple_tool_calls=False passes through to AsyncClient.chat().

The similar unsupported option tool_choice is correctly excluded:

exclude_keys = {"instructions", "tool_choice"}

Suggested Fixes

Fix A (Ollama connector — defense-in-depth)

Add "allow_multiple_tool_calls" to the exclude_keys set in _prepare_options(), the same way tool_choice is excluded. Also consider excluding all keys typed as None in OllamaChatOptions (user, store, logit_bias, metadata).

Fix B (Orchestrations layer — client-aware)

HandoffAgentExecutor._clone_chat_agent() should check whether the underlying client supports allow_multiple_tool_calls before injecting it.

Fix C (Runtime — middleware workaround)

from agent_framework import ChatMiddlewareLayer, ChatMiddlewareContext

class OllamaOptionsFilter(ChatMiddlewareLayer):
    _UNSUPPORTED = frozenset({"allow_multiple_tool_calls"})
    async def send(self, ctx: ChatMiddlewareContext) -> None:
        for key in self._UNSUPPORTED:
            ctx.options.pop(key, None)
        await super().send(ctx)

client = OllamaChatClient(middleware=[OllamaOptionsFilter()])

Environment

• agent-framework-ollama v1.0.0b260630
• agent-framework-orchestrations v1.0.0
• ollama Python client v0.5.3
• Python 3.12

[github-actions]

🤖 Automated triage reproduction notes (agent-authored — trust but verify)

Agent analysis

Repro: OllamaChatClient._prepare_options() in python/packages/ollama/agent_framework_ollama/_chat_client.py (line ~402) has exclude_keys = {"instructions", "tool_choice"} but does not include "allow_multiple_tool_calls"; when HandoffAgentExecutor._clone_chat_agent() in python/packages/orchestrations/agent_framework_orchestrations/_handoff.py (line ~286) sets allow_multiple_tool_calls=False, the non-None boolean passes the filter and is unpacked into AsyncClient.chat(**options_dict), causing TypeError. Minimal repro: construct OllamaChatClient, call client._prepare_options(messages, {"model": "llama3", "allow_multiple_tool_calls": False}) and observe the key in the returned dict.

• Failing test: python/packages/ollama/tests/test_issue_7054_allow_multiple_tool_calls.py
• Files examined: python/packages/ollama/agent_framework_ollama/_chat_client.py, python/packages/orchestrations/agent_framework_orchestrations/_handoff.py, python/packages/ollama/tests/test_ollama_chat_client.py
• Tests run: python/packages/ollama/tests/test_issue_7054_allow_multiple_tool_calls.py
• Reported version: 1.0.0b260630
• Current version: 1.0.0b260709