Python: [Bug]: WorkflowAgent drops response metadata when forwarding AgentResponseUpdate

[junnhwan]

Description

When a workflow executor emits an AgentResponseUpdate, wrapping the workflow with WorkflowAgent and consuming it through run(stream=True) silently drops several response metadata fields.

WorkflowAgent._convert_workflow_event_to_agent_response_updates() creates a new AgentResponseUpdate instance, but currently only copies the content, role, author, IDs, timestamp, and raw representation. The following fields are not forwarded:

• finish_reason
• continuation_token
• additional_properties

The method documentation says that output and intermediate updates are forwarded "as-is", so this appears to be an unintended loss of streaming response semantics.

This is distinct from #1331 / #3146, which addressed author_name, and from #4261, which concerns user-role messages being emitted as agent output.

Actual Result

The metadata fields are returned as None even though they were present on the executor's original AgentResponseUpdate.

Expected Result

WorkflowAgent should preserve these fields when forwarding an AgentResponseUpdate, just as it preserves the content, role, author, response ID, message ID, and timestamp.

Code Sample

Tested on main at commit edfe115ea06bca57ae5a123d0fac5b3fdda13603.

import asyncio

from agent_framework import (
    AgentResponseUpdate,
    Content,
    Executor,
    Message,
    WorkflowAgent,
    WorkflowBuilder,
    WorkflowContext,
    handler,
)


class MetadataExecutor(Executor):
    @handler
    async def handle(
        self,
        messages: list[Message],
        ctx: WorkflowContext[list[Message], AgentResponseUpdate],
    ) -> None:
        await ctx.yield_output(
            AgentResponseUpdate(
                contents=[Content.from_text(text="payload")],
                role="assistant",
                author_name="source-agent",
                response_id="source-response",
                message_id="source-message",
                finish_reason="stop",
                continuation_token="resume-token",
                additional_properties={
                    "provider_marker": "preserve-me",
                },
            )
        )


async def main() -> None:
    source = MetadataExecutor(id="metadata-executor")
    workflow = WorkflowBuilder(start_executor=source).build()
    agent = WorkflowAgent(workflow=workflow, name="wrapped-agent")

    updates = [update async for update in agent.run("input", stream=True)]
    update = updates[0]

    print(
        {
            "finish_reason": update.finish_reason,
            "continuation_token": update.continuation_token,
            "additional_properties": update.additional_properties,
        }
    )


asyncio.run(main())

Run with:

uv run --no-sync python -

Error Messages / Stack Traces

No exception is raised. The metadata is silently dropped during the WorkflowAgent streaming conversion.

Package Versions

• agent-framework-core: 1.16.0 (local source from main at commit edfe115ea06bca57ae5a123d0fac5b3fdda13603)

Python Version

• Python 3.13.1

Additional Context

• OS: Windows
• The current implementation loses the metadata before final response construction, so downstream consumers cannot recover it from AgentResponse.from_updates().
• Long-running operations can lose their continuation_token and become impossible to resume.
• Consumers cannot observe the provider/runtime finish reason.
• Provider-specific metadata in additional_properties is lost.
• Streaming behavior differs between a direct agent and the same agent executed through a workflow wrapper.
• The existing test suite does not currently cover metadata preservation for this conversion path.

A focused implementation would preserve the missing fields when reconstructing the AgentResponseUpdate in _convert_workflow_event_to_agent_response_updates() and add regression coverage in python/packages/core/tests/workflow/test_workflow_agent.py.

I would like to work on this issue if the proposed behavior aligns with the project's expectations. I already have a deterministic reproduction and can prepare a focused fix with regression coverage. Please let me know whether this issue can be assigned to me.

[liyangbing]

This looks like a contract-preservation bug rather than a cosmetic omission: wrapping an executor changes the observable stream even though the documentation promises forwarding “as-is”. The focused fix should preserve metadata without aliasing mutable provider properties:

• Copy finish_reason, continuation_token, and additional_properties at the conversion boundary; define whether nested properties are immutable or require a defensive copy.
• Add a parameterized regression matrix for direct executor output versus WorkflowAgent output: intermediate update, terminal update, empty content, multiple updates with the same response ID, and a provider-specific metadata key.
• Assert the full semantic tuple, not just the rendered text: role, author, response/message IDs, timestamp, finish reason, continuation token, and additional properties.
• Test the resume path explicitly. A preserved continuation token must either resume the same logical response or be rejected with a typed, observable reason; silently dropping it is worse than a hard failure.
• Treat unknown future metadata as pass-through data where possible, while preventing a wrapper from fabricating finish state. Record package revision and adapter path in the test receipt so changes in streaming semantics remain attributable.

The independent handbook has related guidance on durable session evidence and stop/retry boundaries: https://sandbaseai.github.io/deepseek-harness-handbook/live-session-log-durability.html and https://sandbaseai.github.io/deepseek-harness-handbook/stuck-turn-stop-and-retry.html . It is not an official DeepSeek AI project.

[github-actions]

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

Agent analysis

The bug is in python/packages/core/agent_framework/_workflows/_agent.py::_convert_workflow_event_to_agent_response_updates at lines 609-621. It occurs when an output or intermediate workflow event contains an AgentResponseUpdate with finish_reason, continuation_token, or additional_properties populated. Minimal repro: emit one such update from an executor, wrap its workflow in WorkflowAgent, and consume run(stream=True); all three forwarded fields are None.

• Failing test: python/packages/core/tests/workflow/test_workflow_agent.py::test_agent_response_update_preserves_response_metadata
• Files examined: python/packages/core/agent_framework/_workflows/_agent.py, python/packages/core/agent_framework/_types.py, python/packages/core/tests/workflow/test_workflow_agent.py, python/packages/core/pyproject.toml
• Tests run: baseline test_workflow_agent.py (53 passed), test_agent_response_update_preserves_response_metadata (failed)
• Reported version: 1.16.0
• Current version: 1.16.0

[junnhwan]

Thanks for triaging and reproducing this issue. I’m happy to discuss the intended forwarding semantics and prepare a focused fix with a regression test if the maintainers are comfortable with that. Please let me know whether you would like me to take this on.

[HeZ2z]

I've opened PR #7999 with a fix and a streaming regression test. Happy to adjust the forwarding semantics for the other reconstruction branches (AgentResponse / Message / list[Message]) if maintainers prefer them covered here rather than in a follow-up.