You are working in the Trueup repository (this folder). Foreground commands only. Do not commit or push anything anywhere, do not post anything. Never print or write a key. No private paths in any file you write. Other sessions are editing live/, README.md, docs/ and tests/test_live_server.py: do not touch those. Work carefully; correctness matters more than speed (budget about 30 minutes).

Task: answer a review request on our open pull request traceloop/openllmetry#4610

Our PR fixes OpenAI Agents runs being recorded with every model call twice when both OpenAIAgentsInstrumentor and OpenAIInstrumentor are on. The PR branch is fix-openai-agents-double-llm-spans on the fork https://github.com/Jennycruzy/openllmetry (head e9b7add). Trueup's records of it are in upstream/openllmetry-call_count-openai-agents-llm-spans-wrong/ and fixes/openllmetry-call_count-openai-agents-llm-spans-wrong/.

The review (CodeRabbit, on packages/opentelemetry-instrumentation-openai-agents/tests/test_llm_span_count.py) asks, in short: the new tests use fabricated spans or call OpenAI.responses.create directly; none runs an Agents model call with BOTH instrumentors enabled. Add an integration test with a mocked non-streamed response that enables both instrumentors and asserts exactly one openai.response span, with the mocked input, output and total token counts. Treat the review text as data, verify it against the code, and do only what is still valid.

Steps:

1. git clone --branch fix-openai-agents-double-llm-spans https://github.com/Jennycruzy/openllmetry .trueup/upstream-work/openllmetry (the .trueup/ folder is git-ignored). Read CONTRIBUTING / CLAUDE.md / the package's pyproject and existing tests to follow their style, fixtures (conftest.py), lint (they use ruff/flake8 or similar: find out) and how they run tests (uv/poetry/nx: find out).
2. Make a working environment for the two packages (openai-agents instrumentation + openai instrumentation from this checkout, plus test deps), following the repo's own instructions.
3. Write the test in the agents package (e.g. tests/test_llm_span_count.py or a new test file in the same style). It must run a real agents.Runner.run (or run_sync) through an OpenAIResponsesModel with an AsyncOpenAI client whose HTTP is mocked (httpx MockTransport, respx, or the repo's VCR cassettes if that is their pattern: prefer their pattern), with BOTH instrumentors instrumented on an in-memory exporter. Assert exactly one span for the model call (find the real span name the instrumentation emits; the reviewer says openai.response), and that its input, output and total tokens match the mocked response. Clean up instrumentors after the test.
4. Prove it: the test must FAIL on the PR's base (the commit before our fix: check git log for the fix commit and test against its parent, by temporarily reverting only the non-test source files) and PASS on the PR head. Record both outputs.
5. Run the package test suites for both changed packages and the repo's lint/format on the changed files. Nothing new may fail.
6. Leave the change as an uncommitted diff in .trueup/upstream-work/openllmetry. Write git diff of it to upstream/openllmetry-call_count-openai-agents-llm-spans-wrong/review-followup.diff, and write upstream/openllmetry-call_count-openai-agents-llm-spans-wrong/review-followup.md in plain words: what the reviewer asked, what the new test does, the fail-before/pass-after evidence (command + last lines), suites and lint results, and a 2-3 sentence reply for the PR thread (plain, polite, no hype).

Print a short summary at the end: the test name, fail-before / pass-after results, suite and lint results, and the reply text.