Server agents

A DocxEditor.createServer runtime reports document-open diagnostics and a diagnostic code for each refused batch. A refusal counts as tracked when the batch enables tracking before the refusal. Create one telemetry client per process and pass its module to every runtime:

import { DocxEditor } from '@docx-editor.dev/editor-api';
import { createTelemetry } from '@docx-editor.dev/telemetry';
import { otelExporter } from '@docx-editor.dev/telemetry/otel';

const telemetry = createTelemetry({ exporters: [otelExporter()] });

const runtime = await DocxEditor.createServer(bytes, {
  author: 'Review agent',
  modules: [telemetry.module],
});
try {
  // Run the agent here.
} finally {
  runtime.dispose();
}

In a long-running worker, call await telemetry.shutdown() once when the process exits. In a serverless function or single-run job, call await telemetry.flush() before returning to send buffered data. Do not flush after each job in a long-running worker. Each flush forces an export and closes the deduplication windows for all jobs.

Each runtime has a separate editorSessionId, which lets you follow one agent run. A DocxEditor.createBrowser runtime reports refused batches through its host editor with origin: 'automation'. The editor's telemetry module therefore covers browser agents too.

On a server, the Sentry Node SDK can attach the incoming request body to an issue. If an agent endpoint receives instructions or document bytes, remove event.request.data in beforeSend. For refusal details, read error.debugInfo.innerError. For more information, see Batching, loading, and errors.