Python: Fix: Preserve text content when usage data is present in streaming chunks

[toshihiro-otani]

Fix: Preserve text content when usage data is present in streaming chunks

Summary

Fixes streaming response parsing bug where text content is lost when chunk.usage is present in the same chunk as delta.content. This affects providers like Gemini that include both usage and text in the same chunk.

Fixes #[ISSUE_NUMBER]

Problem

The current implementation in _parse_response_update_from_openai performs an early return when chunk.usage is present:

if chunk.usage:
    return ChatResponseUpdate(...)  # Early return skips text parsing!

Impact:

• ❌ Streaming text content is lost (always empty string)
• ❌ Tool calls are also skipped
• ❌ Final conversation contains empty messages
• ✅ Non-streaming responses work correctly

Affected providers: Gemini (confirmed), potentially others that follow OpenAI spec strictly

Solution

Remove the early return and process usage alongside text/tool calls:

1. Add usage to contents list (instead of early return)
2. Continue processing text and tool calls from choices
3. All content types are preserved in the final ChatResponseUpdate

Changes

Modified Files

• agent_framework/openai/_chat_client.py:
  • _parse_response_update_from_openai method (lines 307-348)
  • Removed early return for chunk.usage
  • Process usage, text, and tool calls in a single flow

Tests Added

• test_parse_response_update_with_usage_and_text(): Verifies text is preserved when usage is present
• test_parse_response_update_usage_only(): Verifies usage-only chunks work correctly
• test_parse_response_update_text_only(): Verifies text-only chunks (regression test)

Verification

Before Fix

# Streaming with Gemini
async for update in agent.run_stream("Count 1 to 3"):
    print(update.text)  # Output: '' (empty!)

After Fix

# Streaming with Gemini
async for update in agent.run_stream("Count 1 to 3"):
    print(update.text)  # Output: '1', ',', '2, 3' (correct!)

Backward Compatibility

✅ OpenAI: No impact (likely sends usage in separate chunks)
✅ Azure OpenAI: No impact (tested)
✅ Existing tests: All pass

Testing Performed

• ✅ Unit tests for new behavior
• ✅ Integration test with Gemini streaming
• ✅ Integration test with OpenAI streaming (regression)
• ✅ Multi-agent workflow with Gemini (end-to-end)
• ✅ Tool calling with Gemini streaming

Checklist

• Code follows project style guidelines
• Tests added for bug fix
• All existing tests pass
• Documentation updated (if applicable)
• Issue linked in PR description
• Verified backward compatibility

Additional Notes

This fix enables full Gemini support for:

• ✅ Streaming text responses
• ✅ Tool calling in streaming mode
• ✅ Multi-agent workflows
• ✅ Usage tracking alongside content

The change is minimal and focused on the root cause, ensuring no side effects on other providers.

[Copilot]

Pull request overview

This pull request fixes a critical bug in the OpenAI chat client's streaming response parser where text content was lost when usage data and text content appeared in the same chunk. This issue specifically affected providers like Gemini that include both types of data in a single streaming chunk.

Changes:

• Removed early return when chunk.usage is present in _parse_response_update_from_openai
• Usage data is now appended to the contents list instead of triggering an early return
• Text and tool calls are processed normally after usage data, allowing all content types to be preserved

[Copilot]

The comment format "BUGFIX:" is not consistent with the existing comment style in this file. Other explanatory comments in the codebase (e.g., line 414: "When you enable asynchronous content filtering...") use a more descriptive, natural language style without special prefixes. Consider rephrasing to match the existing style, such as: "Process usage alongside text/tool calls to preserve all content types. Some providers (e.g., Gemini) include both usage and content in the same chunk."

Suggested change

	# BUGFIX: Process usage alongside text/tool calls instead of early return
	# Gemini (and potentially other providers) include both usage and content in the same chunk
	# Process usage alongside text and tool calls to preserve all content types.
	# Some providers (for example Gemini) include both usage and content in the same chunk.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework/openai
_chat_client.py	268	22	91%	44, 178–179, 183, 269, 276, 356–363, 365–368, 378, 463, 500, 516
TOTAL	16246	2373	85%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
3480	221 💤	0 ❌	0 🔥	1m 6s ⏱️