[nik13]

Summary

Keep traceAI-openai: OpenAI Python does not expose a native GenAI OTel exporter that replaces this request instrumentor. Fix missing audio transcription usage through the existing OpenAIInstrumentor entry point rather than adding or deleting a package.

• Extract API-reported input/output/total token usage from typed transcription JSON responses and terminal SSE events, including SDK-floor dict extras. Preserve zero counts; never convert duration billing or absent/malformed usage into token counts.
• Add gen_ai.operation.name=transcription before the request so failures retain their operation and requested model.
• Finish SSE spans on explicit sync/async close() as well as exhaustion and context exit. Retain only the terminal transcription event, not uploaded audio or a growing list of deltas. Existing context-manager error/cancellation behavior is preserved.
• Keep prompts out of duplicate invocation parameters so input masking also works with the older locked fi-core. Retain output masking and provider attribution.
• Document a precise boundary: Responses HTTP/SSE is supported; responses.connect() WebSockets is not instrumented, and the separate Realtime package is not a substitute. Raw streaming-response wrappers, binary speech-download lifetimes and text/SRT/VTT transcription formats are outside this audio extraction contract.

Evidence and design

On base d6a9966dd4d5d5fb344f7f5ab11b7483c2c2a6bc, a real OpenAI client over a fake HTTP transport returned transcription usage 5/2/7 but emitted no normalized usage attributes on both SDK 1.69.0 and 3.42.0. The initial audio suite failed 42 cases on current and 34 on floor (eight unsupported diarized-format skips).

Typed endpoint coverage follows the design demonstrated by Braintrust's explicit audio transcription integration; no vendor code was copied.

OpenAI's direct WebSocket path bypasses OpenAI.request. Two fake-WebSocket regressions exercise the real sync/async SDK connections and assert that the HTTP instrumentor emits no spans, matching the documented boundary rather than claiming a nonexistent adapter.

Verification

All local Python tests ran under the existing macOS sandbox denying outbound network except localhost. Vendor calls use httpx.MockTransport; audio tests export real OTLP/HTTP protobuf to a localhost receiver. No vendor/model API, real credentials or Cloud collector was used.

Full python/frameworks/openai/tests results:

OpenAI SDK	fi-core	Result
3.42.0	repository 1.2.2	114 passed
1.69.0 (package floor)	repository 1.2.2	104 passed, 10 skipped
1.93.1 (CI lock)	released 1.1.0; OTel 1.34.1	102 passed, 12 skipped

Floor skips cover unavailable diarized responses and Responses WebSockets. The locked-core run additionally skips the two echoed-error privacy cases because that core predates hide_error_details; input/output masking and all usage tests still run. The current-core tests verify echoed errors are hidden on the wire. Existing 62 chat, Responses, embeddings, provider and context-manager tests remain green.

Coverage includes sync/async, non-stream/stream, zero/missing/duration/malformed usage, early/empty context exit, explicit/repeated close, consumer errors/cancellation, HTTP errors, SSE API errors, and privacy checks against the serialized OTLP payload.

Re-run with dependencies installed, in a network-denied sandbox permitting localhost:

python -m pytest python/frameworks/openai/tests --noconftest -q -o addopts='' -p no:cacheprovider

Release and docs

No version or changelog bump; release-please owns the patch release. No package deletion, merge, tag or publish is part of this PR. A corrected website page is held until this package fix is on the registry; no docs PR claiming unreleased behavior has been opened. The package README includes the support boundary now. Python audio examples were executed offline at floor/current; the held MDX page compiles. No live screenshot or docs-site visual verification is claimed.

E2E coverage

not applicable (no e2e/ harness) · flows hit: n/a · marker: n/a · verdict: not applicable

Self-review: no introduced blocking finding. New tests and git diff --check pass; no Ruff configuration, migrations, frontend or contract-generation changes apply.

🤖 Generated with Claude Code