Prism OpenTelemetry for TypeScript

GenAI-convention OpenTelemetry spans from Prism telemetry. The TypeScript port of particle-academy/prism-opentelemetry.

Zero runtime dependencies. Node 22+.

npm install @particle-academy/prism-opentelemetry

No dependency on an OTel SDK either: you implement a one-method Tracer (startSpan) against whatever you already export with, and the package never reaches for a global.

Usage

import { SpanStore, TelemetrySubscriber } from '@particle-academy/prism-opentelemetry';

const telemetry = new TelemetrySubscriber(myTracer);

telemetry.onGenerationStarted(
  {
    traceId: 'b7ad6b7169203331',
    operation: 'chat',
    provider: 'anthropic',
    model: 'claude-sonnet-5-5',
  },
  input,
  advertisedTools,
);

Span names follow the GenAI convention — {operation} {model} — and the attribute keys are in GenAi and OpenInference so you are not retyping string literals that have to match a collector's expectations exactly.

Content capture is OFF by default

This is the one setting in the package with a privacy consequence, so it is a single explicit switch rather than something to infer from three config keys:

new TelemetrySubscriber(myTracer, new SpanStore(), {
  captureContent: true,   // prompts, completions, tool arguments
  captureMedia: false,    // attachment BYTES — separate, and still off
  maxContentLength: 65_536,
});

| | | |---|---| | captureContent | Off. Prompts, completions and tool arguments are user content, and a span export leaves your application. | | captureMedia | Off, and meaningless unless captureContent is on. A serialized message carries each attachment's bytes; withoutMediaBytes() keeps a media part's kind and mime type while dropping the payload, so a span can say an image was sent without shipping the image. | | maxContentLength | 64 KiB. A hostile or merely high-volume payload must not be able to bloat a span or the OTLP export. 0 disables the cap. | | recordExceptions | On. |

Two switches rather than one because the answers genuinely differ: a team that wants to see prompts in traces very rarely wants megabytes of base64 alongside them.

Tool spans and digests

advertisedToolDigest() gives a stable digest of a tool as advertised to the model. That is what lets a trace show that the tool set changed between two runs, rather than only that a tool was called — the pinning idea from prism-mcp applied to telemetry.

Parity

Attribute names, span naming and the capture gates are compared against the PHP reference and the Python port by prism-parity's opentelemetry corpus. Worth knowing before you rely on cross-language trace shape: when that corpus was written, zero of its thirteen rows agreed across the three languages. The gaps are tracked as G-23 through G-28 in the envelope's port-gaps register; read it before assuming a span from this port is byte-comparable with one from PHP.