Observability¶
With Micrometer on the classpath, every turn and every tool call is an observation: a timer, and a span parented to whatever the caller was in.
spring.ai.acp.harness.turn{spring.ai.kind=acp_harness_turn, spring.ai.acp.harness.runtime=goose, gen_ai.response.model=gpt-5.6-terra, spring.ai.acp.harness.session.kind=named, spring.ai.acp.harness.outcome=END_TURN}
spring.ai.acp.harness.tool.call{spring.ai.kind=acp_harness_tool_call, spring.ai.acp.harness.runtime=goose, spring.ai.acp.harness.tool.kind=READ, spring.ai.acp.harness.tool.status=COMPLETED}
The names follow Spring AI's rule: an OpenTelemetry GenAI key where one means the same thing, and otherwise a key scoped under the observation's own name. So spring.ai.kind is the tag Spring AI puts on its chat client, advisor, tool call and vector store observations, and gen_ai.response.model the key its chat models tag the model with; a breakdown by kind or by model covers the harness beside them. Everything else is spring.ai.acp.harness.*, matching the configuration properties.
gen_ai.response.model is the model the session is really using, not the one requested, and none when the agent kept its own default and never said which, as Spring AI spells an absent value on that key. The outcome is the agent's stop reason, or the exception's simple name when the turn did not reach one; it stays a spring.ai.acp.harness.outcome tag rather than gen_ai.response.finish_reasons, which Spring AI records as a high-cardinality list on the span. Session names, tool titles, context usage and cost are span attributes, not meter tags; contribute an ObservationConvention to change that. Disable with spring.ai.acp.harness.observations.enabled=false.
| Observation | Low-cardinality tags (meter dimensions) | High-cardinality (span attributes) |
|---|---|---|
spring.ai.acp.harness.turn |
spring.ai.kind, spring.ai.acp.harness.runtime, gen_ai.response.model, spring.ai.acp.harness.session.kind, spring.ai.acp.harness.outcome |
spring.ai.acp.harness.session.name, .context.used, .context.size, .cost |
spring.ai.acp.harness.tool.call |
spring.ai.kind, spring.ai.acp.harness.runtime, spring.ai.acp.harness.tool.kind, spring.ai.acp.harness.tool.status |
spring.ai.acp.harness.tool.title, spring.ai.acp.harness.tool.id |
The ACP specification defines no telemetry of its own. It does reserve traceparent, tracestate and baggage inside _meta for W3C trace context, which this client does not yet send.