fix(openai_agents): tolerate a ResponseSpanData with no response - #3326

Summary

update_trace_properties_from_span_data() in deepeval/openai_agents/extractors.py dereferences span_data.response unconditionally. The OpenAI Agents SDK only populates that attribute when sensitive data is included in the trace, so with RunConfig(trace_include_sensitive_data=False) every ResponseSpanData reaches the trace processor with response=None and the call raises AttributeError.

Why

Root cause, deepeval/openai_agents/extractors.py:431-440 (before this change):

def update_trace_properties_from_span_data(
    trace: Trace,
    span_data: Union["ResponseSpanData", "GenerationSpanData"],
):
    if isinstance(span_data, ResponseSpanData):
        if not trace.input:
            trace.input = parse_response_input(
                span_data.input, span_data.response.instructions   # <-- line 436
            )
        raw_output = parse_response_output(span_data.response.output)

The sibling extractor for the same span data already handles this case — update_span_properties_from_response_span_data() returns early when response is None (deepeval/openai_agents/extractors.py:95-98) — but the trace-level one does not.

The SDK side (agents/models/openai_responses.py:613-619, openai-agents 0.22.3) assigns the span data only behind if tracing.include_data()::

if tracing.include_data():
    span_response.span_data.response = response
    span_response.span_data.input = input
elif redacted_response_id_endpoint_is_trusted and self._uses_official_openai_endpoint():
    span_response.span_data._response_id = response.id

So with trace_include_sensitive_data=False (ModelTracing.ENABLED_WITHOUT_DATA) the span carries only _response_id; response and input stay None.

Impact: DeepEvalTracingProcessor.on_span_end() (deepeval/openai_agents/callback_handler.py:121) raises inside the SDK's processor loop, which logs and swallows it (agents/tracing/provider.py:162-173). Execution never reaches observer.__exit__ (callback_handler.py:136-138), so the LlmSpan stays open in current_span_context / trace_manager.active_spans, its Observer stays in span_observers, and the trace loses the span silently. Verified locally against openai-agents 0.22.3: before the fix the SDK logs Error in trace processor during on_span_end and current_span_context.get() is still non-None after the run; after the fix span_observers is empty and the current span is cleared.

Changes

• deepeval/openai_agents/extractors.py — return early from update_trace_properties_from_span_data() when a ResponseSpanData has response=None, mirroring the existing guard in update_span_properties_from_response_span_data().
• tests/test_integrations/test_openai_agents/test_extractors.py — regression test: build a ResponseSpanData(response=None, input=None) and call the extractor; it must not raise and must leave trace.input / trace.output unset.

No other files are touched; the GenerationSpanData branch is unchanged.

Testing

Regression test on the unpatched tree (test file present, extractors.py at HEAD~1):

$ PYTHONPATH=$PWD .../deepeval_lc_venv/bin/python -m pytest tests/test_integrations/test_openai_agents/test_extractors.py -p no:cacheprovider -q
                trace.input = parse_response_input(
>                   span_data.input, span_data.response.instructions
                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                )
E               AttributeError: 'NoneType' object has no attribute 'instructions'
deepeval/openai_agents/extractors.py:436: AttributeError
FAILED tests/test_integrations/test_openai_agents/test_extractors.py::test_update_trace_properties_without_response
1 failed in 0.13s

Same command on the patched tree:

1 passed in 0.02s

Whole openai_agents module, patched: 15 failed, 1 passed, 1 skipped; same tree with extractors.py reverted: 16 failed, 1 skipped. The failure sets are identical apart from the new test (red → green); the 15 remaining failures are pre-existing and environmental — openai.OpenAIError: Missing credentials ... set the OPENAI_API_KEY environment variable — since this sandbox has no OPENAI_API_KEY, and the CI job is gated on if: ${{ env.OPENAI_API_KEY != '' }}.

Format and lint:

$ .../bin/python -m black --check deepeval/openai_agents/extractors.py tests/test_integrations/test_openai_agents/test_extractors.py
2 files would be left unchanged.
$ .../bin/python -m ruff check deepeval/openai_agents/extractors.py tests/test_integrations/test_openai_agents/test_extractors.py
All checks passed!

(mypy is scoped to ^deepeval/telemetry/ by the repo's pre-commit config, so it does not cover this diff.)

Notes

Related, separate report: #3314 covers the same failure mode for GenerationSpanData with usage=None in update_span_properties_from_generation_span_data (a different function and a different span type); that is not addressed here. This change is limited to the ResponseSpanData branch of update_trace_properties_from_span_data.