Summary
The GenAISemanticProcessor in opentelemetry-instrumentation-openai-agents maps ResponseSpanData to GenAIOperationName.CHAT, causing its span name to be chat {model} (e.g. chat gpt-4o-mini). This is the same span name generated for GenerationSpanData, making them indistinguishable in traces.
Details
In span_processor.py line 1533-1534:
elif _is_instance_of(span_data, ResponseSpanData):
return GenAIOperationName.CHAT # Response typically from chatThe OpenAI Agents SDK emits both ResponseSpanData (high-level API response wrapper) and GenerationSpanData (underlying chat completion) for a single LLM call. When both map to chat, the trace shows two spans named chat gpt-4o-mini with no way to tell them apart without inspecting attributes or parent-child relationships.
Expected behavior
ResponseSpanData should use a distinct operation name (e.g. response) so that the span name differentiates it from the GenerationSpanData chat span. This is how the base A365 SDK (OpenAIAgentsTraceProcessor) handles it — ResponseSpanData gets span name response while GenerationSpanData gets chat {model}.
Reproduction
- Run any agent using
OpenAIAgentsInstrumentor with console exporter enabled
- Send a message that triggers an LLM call
- Observe two spans both named
chat gpt-4o-mini in the trace output
Summary
The
GenAISemanticProcessorinopentelemetry-instrumentation-openai-agentsmapsResponseSpanDatatoGenAIOperationName.CHAT, causing its span name to bechat {model}(e.g.chat gpt-4o-mini). This is the same span name generated forGenerationSpanData, making them indistinguishable in traces.Details
In
span_processor.pyline 1533-1534:The OpenAI Agents SDK emits both
ResponseSpanData(high-level API response wrapper) andGenerationSpanData(underlying chat completion) for a single LLM call. When both map tochat, the trace shows two spans namedchat gpt-4o-miniwith no way to tell them apart without inspecting attributes or parent-child relationships.Expected behavior
ResponseSpanDatashould use a distinct operation name (e.g.response) so that the span name differentiates it from theGenerationSpanDatachatspan. This is how the base A365 SDK (OpenAIAgentsTraceProcessor) handles it —ResponseSpanDatagets span nameresponsewhileGenerationSpanDatagetschat {model}.Reproduction
OpenAIAgentsInstrumentorwith console exporter enabledchat gpt-4o-miniin the trace output