ai_sdk_telemetry 3.0.0

Optional OTLP HTTP metrics exporter for AI SDK Dart.

ai_sdk_telemetry

Optional OTLP/HTTP JSON metrics export for ai_sdk_dart.

OtlpHttpMetricSink implements the core TelemetryMetricSink interface with a bounded in-memory queue. Call flush() at an application-defined boundary and dispose() when the exporter is no longer needed. Both accept an export deadline. Export failures are reported through onDiagnostic; failed flushes retain metrics for a later flush while the sink remains open. Disposal closes admission immediately and makes one bounded final flush attempt.

The adapter exports numeric TelemetryMetric values as independent OTLP gauge data points. It does not aggregate observations into counters or histograms. Metric attributes support strings, booleans, numbers, and lists of those values. Unsupported, cyclic, non-finite, or over-limit values are omitted. Payloads are bounded by default to 64 KiB and 64 attributes per data point. The package supports native Dart platforms through dart:io; browser builds need a separate HTTP implementation of TelemetryMetricSink. It supports metrics only; it does not claim OTLP tracing, logs, exemplars, or retry scheduling beyond the bounded in-memory queue.

Redaction runs once before metric/resource attributes enter retained state. Retries preserve the original observation time and sanitized attributes. Observations with the same name share one OTLP metric containing their data points. Hosts that need counters or histograms should aggregate through a different sink.

An injected HttpClient stays caller-owned; dispose it in the host's lifecycle. Without an injected client, the sink owns and closes its client, replacing it after a connection timeout so later flushes can recover. A timeout before a caller-owned client returns a request cannot cancel that pending connection; if the request arrives later, the sink aborts it before sending content.