fix(server-utils): Record Vercel AI experimental_telemetry.metadata on gen_ai spans

dev-packages/node-integration-tests/suites/tracing/vercelai/span-streaming-v6/scenario-telemetry-metadata.mjs

@@ -0,0 +1,82 @@
+import * as Sentry from '@sentry/node';
+import { embed, embedMany, generateText, streamText } from 'ai';
+import { MockEmbeddingModelV3, MockLanguageModelV3, simulateReadableStream } from 'ai/test';
+
+const usage = {
+  inputTokens: { total: 10, noCache: 10, cached: 0 },
+  outputTokens: { total: 20, noCache: 20, cached: 0 },
+  totalTokens: { total: 30, noCache: 30, cached: 0 },
+};
+
+async function run() {
+  await Sentry.startSpan({ op: 'function', name: 'main' }, async () => {
+    await generateText({
+      experimental_telemetry: {
+        isEnabled: true,
+        functionId: 'support-chat',
+        metadata: { requestId: 'req_generate', tenantId: 'acme' },
+      },
+      model: new MockLanguageModelV3({
+        doGenerate: async () => ({
+          finishReason: { unified: 'stop', raw: 'stop' },
+          usage,
+          content: [{ type: 'text', text: 'Hello!' }],
+          warnings: [],
+        }),
+      }),
+      prompt: 'Hi',
+    });
+
+    const result = streamText({
+      experimental_telemetry: {
+        isEnabled: true,
+        functionId: 'support-stream',
+        metadata: { requestId: 'req_stream', tenantId: 'acme' },
+      },
+      model: new MockLanguageModelV3({
+        doStream: async () => ({
+          stream: simulateReadableStream({
+            chunks: [
+              { type: 'stream-start', warnings: [] },
+              { type: 'text-start', id: '0' },
+              { type: 'text-delta', id: '0', delta: 'Hello!' },
+              { type: 'text-end', id: '0' },
+              { type: 'finish', finishReason: { unified: 'stop', raw: 'stop' }, usage },
+            ],
+          }),
+        }),
+      }),
+      prompt: 'Hi',
+    });
+    for await (const _part of result.fullStream) {
+      void _part;
+    }
+
+    await embed({
+      experimental_telemetry: { isEnabled: true, metadata: { requestId: 'req_embed', tenantId: 'acme' } },
+      model: new MockEmbeddingModelV3({
+        doEmbed: async () => ({ embeddings: [[0.1, 0.2, 0.3]], usage: { tokens: 10 } }),
+      }),
+      value: 'Embed me',
+    });
+
+    await embedMany({
+      experimental_telemetry: { isEnabled: true, metadata: { requestId: 'req_embed_many', tenantId: 'acme' } },
+      model: new MockEmbeddingModelV3({
+        maxEmbeddingsPerCall: 5,
+        doEmbed: async () => ({
+          embeddings: [
+            [0.1, 0.2, 0.3],
+            [0.4, 0.5, 0.6],
+          ],
+          usage: { tokens: 20 },
+        }),
+      }),
+      values: ['First', 'Second'],
+    });
+  });
+
+  await Sentry.flush(2000);
+}
+
+run();

dev-packages/node-integration-tests/suites/tracing/vercelai/span-streaming-v6/test.ts

@@ -304,6 +304,44 @@ describe('Vercel AI integration (streaming, v6)', () => {
     },
   );
 
+  function spanWithMetadata(name: string, op: string, requestId: string) {
+    return expect.objectContaining({
+      name,
+      attributes: expect.objectContaining({
+        'vercel.ai.telemetry.metadata.requestId': attr(requestId),
+        'vercel.ai.telemetry.metadata.tenantId': attr('acme'),
+        [SEMANTIC_ATTRIBUTE_SENTRY_OP]: attr(op),
+      }),
+    });
+  }
+
+  const EXPECTED_SPANS_TELEMETRY_METADATA = {
+    items: expect.arrayContaining([
+      spanWithMetadata('invoke_agent support-chat', 'gen_ai.invoke_agent', 'req_generate'),
+      spanWithMetadata('generate_content mock-model-id', 'gen_ai.generate_content', 'req_generate'),
+      spanWithMetadata('invoke_agent support-stream', 'gen_ai.invoke_agent', 'req_stream'),
+      spanWithMetadata('generate_content mock-model-id', 'gen_ai.generate_content', 'req_stream'),
+      spanWithMetadata('embeddings mock-model-id', 'gen_ai.embeddings', 'req_embed'),
+      spanWithMetadata('embeddings mock-model-id', 'gen_ai.embeddings', 'req_embed_many'),
+    ]),
+  };
+
+  createEsmAndCjsTests(
+    __dirname,
+    'scenario-telemetry-metadata.mjs',
+    'instrument.mjs',
+    (createRunner, test) => {
+      test('records experimental_telemetry.metadata as span attributes', async () => {
+        await createRunner().expect({ span: EXPECTED_SPANS_TELEMETRY_METADATA }).start().completed();
+      });
+    },
+    {
+      additionalDependencies: {
+        ai: '^6.0.0',
+      },
+    },
+  );
+
   createEsmAndCjsTests(
     __dirname,
     'scenario-error-in-tool.mjs',

packages/server-utils/src/integrations/vercel-ai/vercel-ai-dc-subscriber.ts

@@ -33,7 +33,7 @@ import {
   GEN_AI_INVOKE_AGENT,
   GEN_AI_RERANK,
 } from '@sentry/conventions/op';
-import type { Span, SpanAttributes } from '@sentry/core';
+import type { Span, SpanAttributes, SpanAttributeValue } from '@sentry/core';
 import {
   _INTERNAL_skipAiProviderWrapping,
   captureException,
@@ -82,6 +82,7 @@ type GenAiOperation = keyof typeof GEN_AI_OPERATION_SPAN_OPS;
 const VERCEL_AI_OPERATION_ID_ATTRIBUTE = 'vercel.ai.operationId';
 const VERCEL_AI_MODEL_PROVIDER_ATTRIBUTE = 'vercel.ai.model.provider';
 const VERCEL_AI_SETTINGS_MAX_RETRIES_ATTRIBUTE = 'vercel.ai.settings.maxRetries';
+const VERCEL_AI_TELEMETRY_METADATA_ATTRIBUTE_PREFIX = 'vercel.ai.telemetry.metadata.';
 
 // Tracks the top-level operationId (and whether it streams) per `callId` so a model-call span can
 // name its `doGenerate`/`doStream` operation the same way the OTel integration does. `isStream` is
@@ -412,8 +413,9 @@ export function createSpanFromMessage(
     recordToolDescriptions(callId, event.tools);
   }
 
-  const baseAttributes: Record<string, string | number | boolean> = {
+  const baseAttributes: SpanAttributes = {
     [SEMANTIC_ATTRIBUTE_SENTRY_ORIGIN]: ORIGIN,
+    ...telemetryMetadataAttributes(event.telemetryMetadata),
     ...(provider ? { [GEN_AI_PROVIDER_NAME]: provider, [VERCEL_AI_MODEL_PROVIDER_ATTRIBUTE]: provider } : {}),
     ...(modelId ? { [GEN_AI_REQUEST_MODEL]: modelId } : {}),
     ...(maxRetries !== undefined ? { [VERCEL_AI_SETTINGS_MAX_RETRIES_ATTRIBUTE]: maxRetries } : {}),
@@ -450,6 +452,21 @@ export function createSpanFromMessage(
   }
 }
 
+/**
+ * `experimental_telemetry.metadata` (`ai` <= 6) as `vercel.ai.telemetry.metadata.<key>`, the names the OTel
+ * integration produced from the SDK's `ai.telemetry.metadata.*`. Only the orchestrion adapter sets
+ * `event.telemetryMetadata`; `ai` 7 has no `telemetry.metadata`.
+ */
+function telemetryMetadataAttributes(metadata: unknown): SpanAttributes {
+  const attributes: SpanAttributes = {};
+  if (isObjectLike(metadata)) {
+    for (const [key, value] of Object.entries(metadata)) {
+      attributes[`${VERCEL_AI_TELEMETRY_METADATA_ATTRIBUTE_PREFIX}${key}`] = value as SpanAttributeValue;
+    }
+  }
+  return attributes;
+}
+
 /** Start a `gen_ai.<operation>` span named `<operation> <suffix>` (or just `<operation>` when no suffix). */
 function startGenAiSpan(operation: GenAiOperation, suffix: string | undefined, attributes: SpanAttributes): Span {
   return startInactiveSpan({

packages/server-utils/src/integrations/vercel-ai/vercel-ai-orchestrion-subscriber.ts

@@ -107,6 +107,9 @@ const callIdBySpan = new WeakMap<Span, string>();
 // child `generate_content` span (whose event would fall back to the global default). v7's channel forwards
 // these flags on every event, so this keeps v6 identical.
 const recordingBySpan = new WeakMap<Span, ReturnType<typeof recording>>();
+// The operation's `experimental_telemetry.metadata`, keyed by its span, so the model-call span records it
+// too. The OTel integration put it on both, but not on tool-call spans.
+const telemetryMetadataBySpan = new WeakMap<Span, unknown>();
 interface OperationErrorInfo {
   // The span active when the operation was invoked (its call site, e.g. the enclosing request/`main` span).
   // When an operation's error bubbles out unhandled it reaches the global handler outside any span, so — as
@@ -163,6 +166,7 @@ export function subscribeVercelAiOrchestrionChannels(
           ...modelFields(callOptions.model),
           maxRetries: callOptions.maxRetries,
           value: callOptions.value,
+          telemetryMetadata: telemetry.metadata,
           ...recording(telemetry),
         },
       }),
@@ -180,6 +184,7 @@ export function subscribeVercelAiOrchestrionChannels(
           ...modelFields(callOptions.model),
           maxRetries: callOptions.maxRetries,
           values: callOptions.values,
+          telemetryMetadata: telemetry.metadata,
           ...recording(telemetry),
         },
       }),
@@ -263,6 +268,7 @@ function bindOperation(
         callIdBySpan.set(span, callId);
       }
       recordingBySpan.set(span, recording(telemetry));
+      telemetryMetadataBySpan.set(span, telemetry.metadata);
       // v5 has no `executeToolCall` channel, so patch each tool's `execute` to emit the tool-call span.
       // Inert on v6 (guarded inside `patchToolExecute` when the parent is `executeToolCall`'s own span).
       if (isObjectLike(callOptions.tools)) {
@@ -490,6 +496,7 @@ function patchModelMethod(
         // top-level `tools` array. Reading both keeps `tool.definitions` populated on the model-call span.
         tools: callArgs.tools ?? (isObjectLike(callArgs.mode) ? callArgs.mode.tools : undefined),
         messages: callArgs.prompt,
+        telemetryMetadata: telemetryMetadataBySpan.get(parent),
         // Inherit the enclosing operation's per-call recording flags so inputs/tools/outputs are recorded on
         // the model-call span whenever they are on the parent `invoke_agent` span.
         ...recordingBySpan.get(parent),
@@ -669,6 +676,7 @@ function buildTextMessage(type: 'generateText' | 'streamText' | 'generateObject'
       // Normalize to the message-array shape the shared core (and v7's channel) expects: a bare string
       // `prompt` becomes a single user message, matching the SDK's own normalization.
       messages: normalizePromptMessages(options),
+      telemetryMetadata: telemetry.metadata,
       ...recording(telemetry),
     },
   });