Problem
Waza emits internal session events but doesn't expose them in any industry-standard shape. Authors and platform teams who already run an OTel-compatible backend (Aspire, Jaeger, Tempo, Application Insights, Honeycomb, Datadog) can't drop waza into their existing observability pipeline.
Result: a debugging session means staring at results.json and transcripts instead of using familiar trace UIs.
Proposal
Add an opt-in OpenTelemetry exporter for agent runs:
waza run --otel-exporter <otlp|stdout|file> plus --otel-endpoint, --otel-headers.- Emit spans following the GenAI semantic conventions:
gen_ai.system, gen_ai.request.model, gen_ai.usage.input_tokens, gen_ai.tool.name, etc.
- Span hierarchy:
eval → task → turn → tool_call / model_call.
- Default-off; default redaction policy strips prompt/tool-arg/output content unless
--otel-include-payloads is set.
- Best-effort trace ID propagation: preserve trace IDs if the engine surfaces them; do not require Copilot SDK changes.
Why this matters for agentic-first
Agentic runs are graph-shaped (turns × tools × sub-agents). Flat logs hide the structure. OTel traces are how the rest of the industry navigates this shape, and exporting them means waza fits into the observability stack teams already pay for.
Acceptance criteria
Non-goals (filed separately)
Related
Problem
Waza emits internal session events but doesn't expose them in any industry-standard shape. Authors and platform teams who already run an OTel-compatible backend (Aspire, Jaeger, Tempo, Application Insights, Honeycomb, Datadog) can't drop waza into their existing observability pipeline.
Result: a debugging session means staring at
results.jsonand transcripts instead of using familiar trace UIs.Proposal
Add an opt-in OpenTelemetry exporter for agent runs:
waza run --otel-exporter <otlp|stdout|file>plus--otel-endpoint,--otel-headers.gen_ai.system,gen_ai.request.model,gen_ai.usage.input_tokens,gen_ai.tool.name, etc.eval→task→turn→tool_call/model_call.--otel-include-payloadsis set.Why this matters for agentic-first
Agentic runs are graph-shaped (turns × tools × sub-agents). Flat logs hide the structure. OTel traces are how the rest of the industry navigates this shape, and exporting them means waza fits into the observability stack teams already pay for.
Acceptance criteria
gen_ai.*semantic conventions; schema-test verifies attribute names without requiring a live backend.site/with an example wiring to a local collector.Non-goals (filed separately)
Related