otel-emit-at-will

An agent often wants to record something while it works: a value it just computed, an occurrence it noticed. The trap is reaching for a span every time, because managing trace context (start, parent, end) from inside a reasoning step is error-prone and orphans spans, which pollute the trace and lose the parent relationship. The fix is to route by signal type so emitting is safe without threading context by hand. This complements agent-otel-trajectory (which emits per-tool-call spans from the harness); this skill is the at-will value/event path.

Route by signal type

• Metrics are the primary at-will path. Counters, gauges, and histograms are context-free: they carry no trace or parent, so an agent can emit one at any point, safely, with no setup. Use a metric for any ad-hoc value: a running total, a count, a measured size or duration, a score. This is the default; reach for it first.
• Events/logs are occurrence markers. For "X happened" with no value, emit a log record or span event. Attach it to the active span if one exists; otherwise emit it standalone. No new span is created just to hold an event.
• Spans only for bounded operations you can parent. Create a span only for a clear start-to-end operation, and auto-parent it to the active trajectory span via traceparent. Never orphan a span. If you cannot establish a parent, do not emit a span: record a metric or an event instead. An orphan span is worse than no span.
• State changes whose history must be reconstructable are events carrying the change, not gauges of the result. If anyone will ask "how did this value get here?" (quota spend, budget, membership), a latest-value metric destroys provenance — emit the change itself, with enough data to re-derive the running state. Compose with event-sourcing (https://github.com/selamy-labs/agent-skills/tree/main/skills/event-sourcing).

Conventions

• Vendor-neutral OTLP. Emit via the OTLP protocol to whatever collector is configured; do not hardcode a backend.
• Semantic conventions. Use the standard gen_ai attributes for agent/model signals so the data is portable and queryable, not bespoke keys.
• Where the emit path lives. Emitting is a write concern: keep it out of a read-only query/telemetry server. Put it in the runtime, a dedicated emit surface, or a thin wrapper, so the read and write paths stay separate.

DONE means

The agent records ad-hoc values as metrics and occurrences as events, reserves spans for parentable bounded operations, and never emits an orphan span. Signals reach the OTLP backend correctly attributed (gen_ai semconv), with no hand-threaded trace context in the agent's reasoning.

Sources: OpenTelemetry signal model (metrics vs logs/events vs traces); trace-context (traceparent) propagation; OpenTelemetry gen_ai semantic conventions.