Repository navigation
Improve native GenAI semantic convention coverage #6626
Description
While adding Agent Framework to the OpenTelemetry GenAI semantic-conventions reference scenarios, we switched the scenario to use native Agent Framework telemetry (no manual reference spans) so the coverage reflects what AF emits today.
Observed native coverage from agent-framework with OpenAIChatClient and sensitive telemetry enabled:
invoke_agent_internalinferenceexecute_tool
The native telemetry is already useful, but compared with the GenAI semantic-conventions reference coverage we expected, these pieces are missing or not currently mapped:
-
No GenAI
invoke_workflowspan- Workflow execution emits native workflow spans such as
workflow.build,workflow.run,executor.process,edge_group.process, andmessage.send. - These use
workflow.*,executor.*, and message attributes, but do not producegen_ai.operation.name=invoke_workflowwithgen_ai.workflow.name,gen_ai.input.messages, andgen_ai.output.messages.
- Workflow execution emits native workflow spans such as
-
No
gen_ai.client.inference.operation.detailsevent/log- The reference repo has event coverage for inference operation details, but native AF telemetry did not emit this event.
-
Some request attributes appear only on the agent span, not the inference span
gen_ai.request.max_tokens,gen_ai.request.temperature, andgen_ai.request.top_pare present on theinvoke_agentspan.- The child
chat/inference spans currently includegen_ai.request.choice.countand model/response/usage attributes, but not those request option attributes.
-
Server attributes are partial
- Native inference currently reports
server.addressasUnknownagainst the local mock Responses endpoint and does not includeserver.port.
- Native inference currently reports
This came from running the GenAI semconv reference scenario against AF native telemetry in semantic-conventions-genai, using the local mock OpenAI Responses endpoint and OpenAIChatClient.
Metadata
Metadata
Type
Projects
- Status
Activity
eavanvalkenburg commentedon Jun 19, 2026
Reference coverage PR in OTel semantic conventions: open-telemetry/semantic-conventions-genai#325
eavanvalkenburg commentedon Oct 28, 2026
/plan Refresh the remaining native Python GenAI semantic-convention gap matrix against current main, preserving this issue's vNext timing and existing owner. Plan only the residual work; do not assume the original four observations are all still current.
Current main maps max_tokens, temperature and top_p through the shared span attribute extractor. OpenAIChatClient.service_url() is implemented by merged #9225 (fixing #9224). Workflow spans still use workflow.run/workflow.* naming, and presence of a server.port enum or URL parser is not proof that the correct attributes are exported on the relevant inference span.
Use the issue's reference scenario and open-telemetry/semantic-conventions-genai#325 as context, but first establish the exact semantic-convention versions/modes MAF currently supports and what the reference expects. Respect the existing latest-experimental opt-in. Do not silently rename stable spans, duplicate operations, enable new telemetry payloads or treat every evolving experimental field as mandatory.
Produce a bounded matrix for the original workflow operation, inference-operation-details event/log, inference request options and server address/port cases. For each, identify implemented/current gap/not applicable, the owning layer, the emitted shape and parentage, sensitivity/configuration gates and a focused regression scenario. Include streamed/nonstreamed success, failure and cancellation where the lifecycle differs. Verify actual emitted attributes with deterministic exporters and a mock provider rather than inferring coverage from constant names.
For any remaining invoke_workflow/event surface, compare mapping an existing span/event versus adding a new one, public compatibility and telemetry volume/cost, arbitrary typed workflow input/output serialization, redaction and exporter setup. Do not require workflow values to be chat messages or emit raw inputs/outputs by default.
Coordinate the middleware-span proposal in #8559, which already has a maintainer-requested detailed human design pending; do not decide that separate proposal here or post another request to its volunteers.
Return the residual implementation/contract decisions and smallest coherent test/documentation scope before recommending readiness. Keep vNext and the current assignee unchanged. No contributor assignment, release acceleration, blanket conformance claim or implementation approval is implied.