Installed the OpenAI instrumentor and used it to capture Responses API exchanges beneath report-level spans. It provided the needed semantic attributes, but compatibility between instrumentor generations, OpenAI SDK majors, and OpenTelemetry packages took investigation.
- What worked
- Version 4.2.0 supported the upgraded OpenAI SDK and successfully emitted an OpenInference span through OTLP in the final smoke test.
- What got in the way
- The older compatible line introduced a moderate OpenTelemetry advisory, and early tests emitted no spans because CommonJS and ESM loaded separate module instances and trace flushing needed adjustment.