[Somay kousis (Somay-kousis)]

Motivation & Context

A workflow executor that yields an AgentResponse with usage_details (for example response = await agent.run(...) then await ctx.yield_output(response)) reports that usage through workflow.as_agent().run(...), but run(..., stream=True) ends with usage_details=None and no update carries a usage content. Streaming consumers of a workflow agent (hosting, DevUI, telemetry, cost tracking) lose the token counts. Details and repro in #9564.

Description & Review Guide

• What are the major changes? In WorkflowAgent's conversion of an AgentResponse output into stream updates (_workflows/_agent.py), the usage is appended as Content.from_usage(...) to the last update, next to the other response-level fields copied there. That is how AgentResponse.to_updates() carries usage, and from_updates reads it back from contents. When none of the response's messages is shown to the caller (only internal function-call messages, or no messages), one update is emitted that carries the usage and the same response-level fields (continuation_token, finish_reason, agent_id, ...). It has no message_id or author_name, so it joins the message in progress instead of splitting it when it lands between chunks of another stream.
• What is the impact of these changes? Streamed and non-streamed runs report the same usage_details. Usage from several AgentResponse outputs adds up the same way in both modes. Responses without usage are unchanged, and the usage-only update adds no message to the final response.
• What do you want reviewers to focus on? The usage-only update for the no-caller-facing-message case: whether you'd rather attach that usage some other way.

Tests in tests/workflow/test_workflow_agent.py:

• test_workflow_as_agent_stream_preserves_agent_response_usage: usage only on the last update, final response matches non-streaming.
• test_workflow_as_agent_stream_sums_usage_of_multiple_agent_responses: two outputs, same total in both modes.
• test_workflow_as_agent_stream_keeps_usage_without_caller_facing_messages[only_internal_call|no_messages]: usage, text and message count match non-streaming.
• test_workflow_as_agent_stream_usage_only_update_does_not_split_a_message: a usage-only response between two chunks of one streamed message leaves it as one message.
• test_workflow_as_agent_stream_usage_only_update_keeps_continuation_token: the usage-only update carries the response's continuation_token.

With _agent.py from main, 6 of these 7 fail. Removing only the usage-only branch fails the 4 tests that need it. tests/workflow: 1464 passed, 1 skipped, 2 xfailed; packages/orchestrations/tests: 297 passed, 6 skipped. Ruff check is clean on the changed files, and the core typing and test-typing tasks (pyrefly, ty, zuban, pyright, mypy) pass.

Related Issue

Fixes #9564

AI Assistance

• No material AI assistance was used.
• This is an AI-assisted contribution. I reviewed, understood, and verified all submitted content and accept responsibility for it.

AI assistance details: AI tools helped find the bug and write the code, the tests and this description. I reviewed all of it and ran the repro, the workflow tests, and the lint and typing checks myself.

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR links to an agreed issue with no competing open PR, or the Related Issue section documents a trivial-change or repository-automation exception.
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

🟡 Changes recommended

Usage-only updates expose filtered internal messages through their raw representation.

1 open issue

• Prevent filtered messages from leaking through raw representations · New

What changed in this PR

This PR addresses #9564 by preserving token usage when workflow agents stream AgentResponse outputs, aligning streaming behavior with non-streaming runs.

Changes:

• Appends usage content to the final update for each response.
• Emits a usage-only update when no messages remain after filtering, preserving response metadata.
• Adds regression tests for usage totals, filtered responses, message continuity, and continuation tokens.

File	Description
python/​packages/​core/​tests/​workflow/​test_workflow_agent.py	Adds streaming usage and metadata regression tests.
python/​packages/​core/​agent_framework/​_workflows/​_agent.py	Preserves response usage in normal and usage-only stream updates.

_{Review effort: Balanced}

[Somay kousis (Somay-kousis)]

Right, fixed in 56d3841: the usage-only update no longer sets raw_representation, so the filtered messages don't reach the caller through it or the final response. The edge-case test now also asserts that no streamed update has a raw representation and that the final response's raw representation holds no AgentResponse; it fails on aa19c4f and passes now.