Python: Skip get_final_response in OTel _finalize_stream when stream errored

[droideronline]

Closes #5231

Problem

AgentTelemetryLayer registers _finalize_stream as a cleanup hook on the ResponseStream. Cleanup hooks run on both normal completion and streaming errors. Inside _finalize_stream, get_final_response() was called unconditionally — and get_final_response() always runs all registered result hooks, including _post_hook in _agents.py which calls _run_after_providers.

This caused after_run context providers to fire on error paths, which is incorrect.

The call chain on streaming error:

stream raises Exception
  → ResponseStream.__anext__ runs _run_cleanup_hooks()
    → _finalize_stream() called
      → get_final_response() called unconditionally   ← bug
        → result hooks fire (e.g. _post_hook)
          → _run_after_providers called               ← wrong on error

Fix

In both _finalize_stream closures (chat client layer and agent layer), check result_stream._consumed before calling get_final_response(). _consumed is set to True only in the StopAsyncIteration branch of __anext__ (normal stream completion). On the error path it remains False, so we can use it to skip get_final_response() safely.

async def _finalize_stream() -> None:
    try:
        if not result_stream._consumed:
            # Stream did not complete normally (e.g., it errored). Skip
            # get_final_response() to avoid firing result hooks such as
            # after_run context providers on error paths. The span is
            # still closed in the finally block below.
            return
        response = await result_stream.get_final_response()
        ...
    except Exception as exception:
        capture_exception(span=span, exception=exception, timestamp=time_ns())
    finally:
        _close_span()

The OTel span is still closed via the finally block in both the error and normal paths.

Notes

Affects both _trace_chat_client_invocation (chat layer) and _trace_agent_invocation (agent layer) — both had the same pattern and are fixed here.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Fixes OTel streaming instrumentation so result hooks (e.g., after_run context providers) don’t fire when a stream errors, by skipping get_final_response() unless the stream completed normally.

Changes:

• Add a _consumed guard in the chat stream _finalize_stream to avoid calling get_final_response() on error paths.
• Add the same guard in the agent stream _finalize_stream.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.