@alvin0/ai-agent-sdk-observability-fetch

Runtime: Universal (Edge/Worker, browser, Deno, Bun, and Node).

pnpm add @alvin0/ai-agent-sdk-core @alvin0/ai-agent-sdk-observability-fetch

Universal acknowledged HTTPS exporter for @alvin0/ai-agent-sdk-core/observability. It sends bounded JSON batches with an idempotency key and retries only the observation batch—not the model/provider operation that produced it.

import { fetchObservationExporter } from '@alvin0/ai-agent-sdk-observability-fetch'

const exporter = fetchObservationExporter({
  endpoint: 'https://telemetry.example.com/v1/observations',
  headers: { authorization: `Bearer ${telemetryToken}` },
})

Use flushObservabilityWithWaitUntil() to hand a flush promise to an Edge host's explicit waitUntil function. The package never assumes a platform global.

The exporter uses POST JSON with idempotency-key: <batchId>, a 10-second request timeout, and batches capped at 256 events/512 KiB. It accepts HTTP 204 or { "acceptedBatchId": "<batchId>" }. Network failures, 408, 425, 429, and 5xx responses retry the identical serialized batch with bounded full jitter; other responses fail without retry. HTTPS is required except for an explicitly enabled localhost/loopback test endpoint, and redirects or cross-origin responses are rejected.

Composition: runtime.observability.exporters. Lifecycle: explicit-owned-or-borrowed; every runtime registration states ownership, requirement, and remote-acknowledged boundary.