OTLP trace export

One span per request, assembled when the request leaves the processor. The trace is what the request stats already measure: when it was received, scheduled, prefilled, first-tokened and finished, its tokens, its cache hit and how it ended. Nothing new is measured and the serving loop gains no work it did not already have — --enable-log-request-stats and tracing read the same RequestTimings, and either one switches the stamps on.

The rules it obeys

• Emission anchors to the request lifecycle, never the round. The span is built at the same release points the finish log line is (OutputProcessor's release paths, including stream's in-place release), from the state the processor already holds. There is no per-round trace work.
• The loop never blocks. The span crosses to the export thread over a bounded channel (try_send); a full channel drops the span and counts it — the metrics relay's drop policy. The export thread batches, and a POST that fails logs and drops its batch; traces never queue behind an unreachable collector.
• Sampling happens before the record is built. --trace-sample-ratio keeps that share of requests; 0 (the default) keeps none and the publisher is not even constructed without an endpoint.
• Ids derive from the request id. The trace id is a 128-bit mix of the rid, so the same request resolves to the same trace id across ranks, logs and re-derivations; span ids mix the rid with a process counter so every span differs. There is no global tracing context to maintain.

The export

OTLP/HTTP with a JSON body — every OTLP/HTTP receiver takes JSON, and the body encodes with serde_json and ships with ureq, so the engine carries no opentelemetry dependency stack. Integers are strings and ids hex, as protobuf's JSON mapping writes 64-bit and bytes fields.

A request that names a continual-prompting session (session_params.id on the wire) carries it as the GenAI session.id attribute — every request stays its own trace, and the tracing UI threads one conversation's requests under the session. Requests that name no session carry no attribute; conversation branches (session_params.rid) linking to the request they continue is a later refinement.

Attributes follow the GenAI semantic conventions exactly (gen_ai.system and gen_ai.request.model on the resource; gen_ai.response.id, gen_ai.usage.input_tokens and gen_ai.usage.output_tokens, and gen_ai.response.finish_reasons as an array, on the span), so Langfuse, Datadog and the Jaeger UIs read them natively; TokenSpeed-specific facts (cached prompt tokens, preemption count and seconds) carry tokenspeed. prefixes. The span is a SERVER span; an aborted or transfer-failed request carries the error status. The lifecycle points ride as time-stamped events (tokenspeed.request.scheduled, …prefill_done, …first_token) between the span's start (received) and end (finished).

What stays out

• Context propagation. The engine derives its own trace ids; a request joining an upstream trace (W3C traceparent from the agent framework through the gateway) needs a metadata field on the wire request and is a separate change.
• Stage spans. One span per request, not per round or per stage; PD roles each trace their own request lifetimes, as they each measure their own queue wait.