@irisrun/observe

OTel-shaped spans, derived from the journal — determinism intact. Your agent already recorded everything it did. This turns a recorded session's inspection into spans, so observability is a projection of the journal, not a side effect woven through the engine. A turn that was observed still replays byte-identically.

What it is

toSpans(inspection) takes a SessionInspection (from @irisrun/inspect) and returns OTel-shaped Span[]: a root turn span parenting one child per effect (intent→result, carrying its outcome) and one per marker. spanIds are DETERMINISTIC — sessionId#seq, no RNG — so re-spanning is byte-identical; timing reads the journal's recorded ts.

Emission goes through an injected Sink (export(spans)), decoupling spans from any backend: collectingSink() accumulates in memory (for tests), consoleSink() prints one JSON span per line. Real OTLP export is a gated manual smoke. Depends on @irisrun/core + @irisrun/inspect only.

Use it

import { toSpans, collectingSink } from "@irisrun/observe";
import { inspectSession } from "@irisrun/inspect";

const { sink, spans } = collectingSink();
await sink.export(toSpans(await inspectSession(store, "s1")));

See docs/Observability.

Part of Iris — own, portable, verifiable state.