Python: [Feature]: DevUI doesn't accept function_invocation_kwargs

[arnabbiswas1]

Description

When an Agent is executed directly with agent.run(..., function_invocation_kwargs=...), tool functions can access those values through FunctionInvocationContext.

When the same agent is served through agent_framework.devui, there is no way to pass values through function_invocation_kwargs. DevUI calls agent.run(...) without forwarding function_invocation_kwargs. As a result, tools that rely on ctx.kwargs[...] fail.

This makes DevUI unusable for agents whose tools depend on request-scoped invocation kwargs.

I have tried to pass the following message through the DevUI chat window, but as expected the tool call failed.

{
    "messages": "Call echo_ctx with value=hello",
    "function_invocation_kwargs":
    {
        "tenantId": "abc123"
    }
}

Code Sample

python
from typing import Annotated
from pydantic import Field
from dotenv import load_dotenv

from agent_framework import Agent, tool, FunctionInvocationContext
from agent_framework.devui import serve
from agent_framework.foundry import FoundryChatClient
from azure.identity import DefaultAzureCredential
import os

load_dotenv()


@tool(name="echo_ctx", description="Reads a value from invocation kwargs")
def echo_ctx(
    value: Annotated[str, Field(description="Any value")],
    ctx: FunctionInvocationContext,
):
    tenant_id = ctx.kwargs["tenantId"]
    return f"tenant={tenant_id}, value={value}"


client = FoundryChatClient(
    project_endpoint=os.environ["FOUNDRY_PROJECT_ENDPOINT"],
    model=os.environ["AZURE_AI_MODEL_DEPLOYMENT_NAME"],
    credential=DefaultAzureCredential(),
)

agent = Agent(
    client=client,
    name="TestAgent",
    instructions="You are a test agent",
    tools=[echo_ctx],
)

serve(entities=[agent], auto_open=False, auth_enabled=False)

Error Messages / Stack Traces

Package Versions

agent-framework-devui :1.0.0b260721

Python Version

Python 3.12.3

Additional Context

No response

[ebarkhordar]

Confirmed on main (28e02d46), measured over HTTP rather than through Foundry, so no credentials are needed to see it.

The drop point is packages/devui/agent_framework_devui/_executor.py:375-379: run_kwargs is built from stream plus an optional session and nothing else, so anything else on the request never reaches agent.run().

run_kwargs: dict[str, Any] = {"stream": True}
if session:
    run_kwargs["session"] = session

stream = cast(Any, agent.run(user_message, **run_kwargs))

The reason it fails as a KeyError inside the tool instead of a 422 is AgentFrameworkRequest.model_config = ConfigDict(extra="allow") (models/_openai_custom.py:318). The field is accepted, stored on the model as an extra field, and then never read.

Repro with a recording agent, no chat client involved:

class RecordingAgent(BaseAgent):
    def run(self, messages=None, *, stream=False, session=None, **kwargs):
        RECORDED.append({"stream": stream, "session": session is not None, "other_kwargs": dict(kwargs)})
        ...

server = DevServer(auth_enabled=False, ui_enabled=False)
server.set_pending_entities([RecordingAgent(name="RecordingAgent")])
client.post("/v1/responses", json={
    "metadata": {"entity_id": entity_id},
    "input": "Call echo_ctx with value=hello",
    "stream": False,
    "function_invocation_kwargs": {"tenantId": "abc123"},
})

Output:

POST /v1/responses -> 200
POST /v1/responses (via extra_body) -> 200
agent.run() call count: 2
kwargs the agent actually received: [{"stream": true, "session": false, "other_kwargs": {}},
                                     {"stream": true, "session": false, "other_kwargs": {}}]

The second POST is worth calling out for anyone hitting this before it is fixed: extra_body, the field whose comment says "Optional extra_body for advanced use cases", is not a workaround either. It is read in exactly two places, _executor.py:296 for response_id and _executor.py:469-470 for a workflow checkpoint_id, so a function_invocation_kwargs placed there is dropped the same way, also with a 200.

What I did not check: whether the workflow path at _executor.py:512-564 has the same gap, and whether any request field other than function_invocation_kwargs is expected to reach run().

[eavanvalkenburg]

I consider this a feature request since adding this is net-new, not really a bug. We will have a look at if we want to support this, the devui is not meant to be a full-fledged client that always covers the whole surface, because there is just too much in code that we do not want to replicate across both devui front and back-ends.

[arnabbiswas1]

Makes sense.