azure-ai-agentserver-langgraph: SSE stream hangs when LangGraph agent uses tool calls

[Bartcardi82]

Package

azure-ai-agentserver-langgraph version 1.0.0b12 (with azure-ai-agentserver-core 1.0.0b12)

Describe the bug

When a LangGraph agent hosted in Azure AI Foundry uses tools (via @tool decorated functions), the Foundry playground SSE stream never closes. The agent generates the correct response text, but the playground spinner keeps running and the user cannot send follow-up messages without clicking "Stop".

Root cause (from Application Insights traces)

The ResponseFunctionCallArgumentEventGenerator and ResponseOutputTextEventGenerator in the agentserver cannot process LangGraph's streaming AIMessageChunk objects that contain tool calls. Every chunk produces a warning and is skipped:

FunctionCallArgumentEventGenerator did not process message: content='' tool_calls=[{'name': 'check_hr_capacity', 'args': {}, 'id': 'call_p2WaEs1A...', 'type': 'tool_call'}]

Message can not be processed by current generator ResponseFunctionCallArgumentEventGenerator: <class 'langchain_core.messages.ai.AIMessageChunk'>

This happens for:

1. Each tool call chunk (tool_calls=[...])
2. The finish chunk (finish_reason='tool_calls')
3. The usage metadata chunk
4. The final response finish chunk (finish_reason='stop')

Because the generators skip these chunks, the SSE stream never emits proper completion events and the connection stays open.

Secondary bug

There is also an async/sync mismatch when fetching conversation history:

TypeError: 'async for' requires an object with __aiter__ method, got coroutine

File "azure/ai/agentserver/langgraph/models/response_api_default_converter.py", line 251, in _fetch_historical_items
    async for item in openai_client.conversations.items.list(conversation_id):
TypeError: 'async for' requires an object with __aiter__ method, got coroutine

To reproduce

1. Create a LangGraph agent with one or more @tool decorated functions
2. Wrap it with from_langgraph(agent, credentials=DefaultAzureCredential())
3. Deploy to Azure AI Foundry as a hosted container agent
4. Open the Foundry playground and send a message that triggers a tool call
5. The response text appears but the spinner never stops

Minimal agent code

from langchain.agents import create_agent
from langchain.chat_models import init_chat_model
from langchain_core.tools import tool
from langgraph.checkpoint.memory import MemorySaver

@tool
def get_info(query: str) -> str:
    """Get information."""
    return f"Result for: {query}"

def create_agent():
    model = init_chat_model("azure_openai:gpt-4.1", ...)
    return create_agent(model, tools=[get_info], checkpointer=MemorySaver())

app.py

from azure.ai.agentserver.langgraph import from_langgraph
from azure.identity import DefaultAzureCredential
from agent import create_agent

app = from_langgraph(create_agent(), credentials=DefaultAzureCredential())

Expected behavior

The SSE stream should properly emit function_call, function_call_output, and output_text events for LangGraph tool call chunks, and close the stream when the agent completes its response.

Actual behavior

• All tool call chunks are skipped by the event generators
• The SSE stream never closes
• The Foundry playground hangs with the spinner

Environment

• azure-ai-agentserver-langgraph==1.0.0b12
• azure-ai-agentserver-core==1.0.0b12
• langchain==1.2.10
• langgraph==1.0.9
• langchain-core==1.2.14
• Python 3.11
• Azure AI Foundry hosted container

Additional context

• Agents without tool calls (pure text responses) work correctly
• The tool functions themselves execute successfully — the issue is purely in the SSE stream conversion
• parallel_tool_calls=False binding on the model does not prevent the issue (model still returns multiple tool calls)
• This was tested with both single and multiple tool calls — the stream hangs in both cases