fix(core): fix streamed response span losing response/input when consumer stops at completed event

[XuQuanxin04]

Summary

When streaming with OpenAIResponsesModel, the model response and input were attached to the response_span only after the yield loop finished (lines 769-771). A consumer that stops iterating at the terminal response.completed event closes the generator at the yield, so that post-loop assignment never runs. The span is then exported with response and input set to None, even when sensitive-data capture is enabled (ModelTracing.ENABLED).

The SDK's own run loop consumes the stream exactly this way — run_single_turn_streamed wraps the model stream in aclosing(...) and breaks once it sees ResponseCompletedEvent — so the common streamed path silently lost the model I/O on the span.

This is asymmetric with three places that already populate the span before yielding for exactly this reason:

• the non-streamed get_response path sets response/input inside the span (lines 612-614);
• the sibling Chat Completions adapter populates its generation span before yielding the terminal event;
• the failure path in this same method records the error before yielding (with a comment about the GeneratorExit hazard).

The usage field was already handled correctly before the yield; only response/input were deferred.

Fix

Attach response and input to the span inside the ResponseCompletedEvent branch, gated on tracing.include_data(), before the terminal event is yielded — mirroring the existing usage assignment right above it. Redaction when data capture is disabled is unchanged.

Test plan

Added test_stream_span_records_io_when_consumer_stops_at_completed, which mirrors the existing chatcompletions test test_stream_span_is_recorded_for_a_consumer_that_stops_at_the_terminal_event. It monkeypatches _fetch_response to emit a single response.completed, stops iterating at that event (as the run loop does), closes the generator, and asserts the response span carries non-None response, input, and usage.

• Without the fix: response/input are None (reproduced).
• With the fix: all three are populated.
• Also verified that with ModelTracing.ENABLED_WITHOUT_DATA, response/input remain None (redaction preserved).
• pytest tests/models/test_openai_responses.py — 156 passed
• pytest tests/models/test_openai_chatcompletions_stream.py tests/test_provider_span_errors.py tests/tracing/ tests/test_agent_runner_streamed.py — 389 passed, 7 skipped
• ruff format --check and ruff check clean on changed files.

No issue number; this is a self-contained observability correctness fix found while reading the span code.

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh — could not run the bash script end-to-end on this Windows/Git-Bash host (it manages POSIX process groups via setsid/kill -TERM -PID); I ran its constituent steps directly (ruff format, ruff check, and the full relevant pytest suites above).
• I've confirmed all verification steps pass