fix(openai): trace responses.retrieve so background responses get output and usage

[Kairick]

What does this PR do?

Instruments Responses.retrieve and AsyncResponses.retrieve, so background responses get their output, token counts and cost.

create and parse are instrumented, retrieve is not. With background=True the create call returns status="queued" with no usage, and the real result arrives through retrieve() — either directly, or via responses.stream(response_id=...), which delegates to retrieve(stream=True). The generation therefore stays in the trace with an input and a model but no output, no usage_details and no cost, which reads as a free call rather than a missing one.

The response id is recorded as the generation input, since retrieve() carries no prompt of its own. Model, output and usage come from the response, so both the streaming and non-streaming paths are handled by the existing extraction code.

A key-free reproduction of the routing is in the issue.

Open question: retrieving a response that was already traced at creation time (non-background usage) will now record its usage a second time. This PR implements the straightforward version — instrument every retrieve — but restricting it to stream=True, or skipping usage when the response was not queued, are both easy to switch to if you prefer one of those.

Fixes #1834

Type of change

• Bug fix
• New feature
• Breaking change
• Refactor
• Documentation update
• Tooling, CI, or repo maintenance

Verification

Both new tests fail on main and pass with the change.

uv run --frozen pytest tests/unit/test_openai.py -q -k retrieve   # 2 passed (2 failed before the fix)
uv run --frozen pytest tests/unit/test_openai.py -q               # 36 passed
bash scripts/codex/quick-check.sh                                 # ruff, mypy, 670 passed, 2 skipped, 18 errors
uv run --frozen ruff format --check langfuse/openai.py tests/unit/test_openai.py

The 18 errors in the unit run come from tests/unit/test_prompt.py, which needs Langfuse credentials in the environment. They reproduce unchanged on main.

Not run: tests/e2e and tests/live_provider, which need a Langfuse server and provider keys. Per AGENTS.md this change is covered by exporter-local unit assertions, so no e2e or live-provider coverage was added.

Checklist

• I self-reviewed the diff using code_review.md.
• I added or updated tests for behavior changes.
• I updated docs, examples, or .env.template if needed. (Not needed: no public API surface or env variable changes.)
• I did not hand-edit generated files; if generated files changed, I used the upstream regeneration path.
• I did not commit secrets or credentials.

Greptile Summary

Instruments synchronous and asynchronous OpenAI Responses retrieval so completed background responses produce Langfuse generation telemetry.

• Registers Responses.retrieve and AsyncResponses.retrieve for tracing.
• Records the response ID as retrieval input while extracting model, output, and usage from the returned response.
• Adds unit coverage for synchronous and asynchronous retrieval.

Confidence Score: 5/5

The PR appears safe to merge, with no unacknowledged blocking or independently actionable issues identified.

The new retrieval registrations use the existing synchronous, asynchronous, and Responses API extraction paths, and the added tests cover the primary non-streaming behavior.

_{Reviews (1): Last reviewed commit: "fix(openai): trace responses.retrieve so..." | Re-trigger Greptile}

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

All committers have signed the CLA.

[Kairick]

recheck