Description

When using stream=True with FoundryAgent.run(), the chat and execute_tool spans appear as siblings of invoke_agent instead of children. The span tree is flat instead of nested.

Expected:

invoke_agent (WeatherAgent)          ← parent
  ├── chat (gpt-5.4)                ← child
  ├── execute_tool (get_weather)     ← child
  └── chat (gpt-5.4)                ← child

Actual (streaming):

invoke_agent (WeatherAgent)          ← sibling
chat (gpt-5.4)                      ← sibling
execute_tool (get_weather)           ← sibling
chat (gpt-5.4)                      ← sibling

Root Cause:

In AgentTelemetryLayer._trace_agent_invocation() (observability.py), the streaming path creates the invoke_agent span with:

span = get_tracer().start_span(f"{operation} {span_name}")

This creates the span but does not activate it as the current context. Child spans (chat, execute_tool) have no way to discover invoke_agent as their parent.

Compare with the non-streaming path which correctly uses _get_span():

with _get_span(attributes=attributes, span_name_attribute=OtelAttr.AGENT_NAME) as span:
    # ... execute() runs here, child spans auto-parent correctly

_get_span() calls start_span() + trace.use_span(), which activates the span as the current context. The streaming path skips use_span() entirely.

The same pattern exists in ChatTelemetryLayer.get_response() streaming path, with this comment:

# Create span directly without trace.use_span() context attachment.
# Streaming spans are closed asynchronously in cleanup hooks, which run
# in a different async context than creation — using use_span() would
# cause "Failed to detach context" errors from OpenTelemetry.

This is a known trade-off, but it results in broken span hierarchy for all streaming agent invocations.

Code Sample

agent = FoundryAgent(...)  # or Agent(...)
session = agent.create_session()

# Streaming — spans are siblings (broken)
async for update in agent.run("Hello", session=session, stream=True):
    pass

# Non-streaming — spans are correctly nested
response = await agent.run("Hello", session=session, stream=False)

Error Messages / Stack Traces

Package Versions

agent-framework-core: 1.0.0, agent-framework-foundry: 1.0.0

Python Version

3.10

Additional Context

Suggested Fix

Consider using trace.use_span() with end_on_exit=False during the synchronous setup phase to establish parent context, then end the span asynchronously in the cleanup hook. The context detach would happen before yielding the stream, avoiding the cross-async-context issue.

Environment

• agent-framework: 1.0.0
• Python: 3.10
• OS: Windows