fix(nextjs): Add edge no-op shim for vercelAIIntegration

Removing `vercelAIIntegration` from `@sentry/vercel-edge` dropped it from the
`@sentry/nextjs` edge build (which re-exports that package), so named imports
from `@sentry/nextjs` would fail when Next.js compiles instrumentation for the
edge runtime — while `index.types.ts` still declares the export.

Mirror the existing `pinoIntegration` pattern: a no-op edge shim plus an explicit
server re-export so the export is statically detectable from both builds. Extends
the `serverExports` regression test to cover it.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: mydea

MIGRATION.md

@@ -875,7 +875,7 @@ Sentry.init({
 - AI integrations are no longer available in the browser SDK. They remain available in the server-side SDKs.
 - The AI instrumentation code moved out of `@sentry/core` into `@sentry/server-utils`. If you imported any AI helper **directly from `@sentry/core`**, import it from `@sentry/server-utils` instead (or keep importing it from your platform SDK, e.g. `@sentry/node`, if it re-exported that helper before — platform SDK availability is unchanged from v10). Affected helpers: `instrumentOpenAiClient`, `instrumentAnthropicAiClient`, `instrumentGoogleGenAIClient`, `instrumentWorkersAiClient`, `createLangChainCallbackHandler`, `instrumentLangChainEmbeddings`, `instrumentStateGraph`, `instrumentStateGraphCompile`, `instrumentCreateReactAgent`.
 - The `addVercelAiProcessors` helper was removed. It was an internal building block for setting up Vercel AI span processing by hand; `vercelAIIntegration()` now wires this up on its own, so add that integration instead of calling `addVercelAiProcessors` directly.
-- (Vercel Edge) `vercelAIIntegration` was removed from `@sentry/vercel-edge` (and, with it, the Edge runtime of `@sentry/nextjs`). Vercel AI is no longer instrumented on the Edge runtime.
+- (Vercel Edge) `vercelAIIntegration` was removed from `@sentry/vercel-edge`; Vercel AI is not instrumented on the Edge runtime. `@sentry/nextjs` keeps the export on its Edge build as a no-op (so `import { vercelAIIntegration }` from `@sentry/nextjs` still resolves in edge-compiled instrumentation files), with the real instrumentation running only in the Node runtime.
 - (Cloudflare & Deno) `vercelAIIntegration` no longer post-processes the OpenTelemetry spans emitted by the `ai` SDK. Instrumentation now goes solely through the channel-based instrumentation, the same as the other server SDKs.
 - The following low-level AI exports are no longer part of the public API (they were provider-instrumentation internals exported from `@sentry/core`):
   - Attribute/stream/util helpers: `extractOpenAiRequestAttributes`, `addOpenAiRequestAttributes`, `addOpenAiResponseAttributes`, `extractOpenAiRequestParameters`, `instrumentOpenAiStream`, `extractAnthropicRequestAttributes`, `addAnthropicRequestAttributes`, `addAnthropicResponseAttributes`, `instrumentAsyncIterableStream`, `instrumentMessageStream`, `extractGoogleGenAIRequestAttributes`, `addGoogleGenAIRequestAttributes`, `addGoogleGenAIResponseAttributes`, `instrumentGoogleGenAIStream`, `getProviderMetadataAttributes`, `getTruncatedJsonString`, `shouldEnableTruncation`, `resolveAIRecordingOptions`, `wrapToolsWithSpans`, `extractLLMFromParams`, `extractAgentNameFromParams`, `instrumentCompiledGraphInvoke`.

packages/nextjs/src/common/vercelAIIntegrationShim.ts

@@ -0,0 +1,12 @@
+import { defineIntegration } from '@sentry/core';
+
+/**
+ * Shim for the edge build so named imports from `@sentry/nextjs` stay resolvable in
+ * edge-compiled instrumentation modules. The real implementation ships in the server build;
+ * Vercel AI is not instrumented on the edge runtime.
+ */
+export const vercelAIIntegration = defineIntegration(() => {
+  return {
+    name: 'VercelAI',
+  };
+});

packages/nextjs/src/edge/index.ts

@@ -42,6 +42,7 @@ export * from '../common';
 export { captureUnderscoreErrorException } from '../common/pages-router-instrumentation/_error';
 
 export { pinoIntegration } from '../common/pinoIntegrationShim';
+export { vercelAIIntegration } from '../common/vercelAIIntegrationShim';
 
 // Override core span methods with Next.js-specific implementations that support Cache Components
 export { startSpan, startSpanManual, startInactiveSpan } from '../common/utils/nextSpan';

packages/nextjs/src/server/index.ts

@@ -25,8 +25,8 @@ import { maybeCleanupQueueSpan } from './vercelQueuesMonitoring';
 
 export * from '@sentry/node';
 
-// Explicitly re-export so it is statically detectable by turbopack
-export { pinoIntegration } from '@sentry/node';
+// Explicitly re-export so these are statically detectable by turbopack
+export { pinoIntegration, vercelAIIntegration } from '@sentry/node';
 
 export { captureUnderscoreErrorException } from '../common/pages-router-instrumentation/_error';
 

packages/nextjs/test/serverExports.test.ts

@@ -6,17 +6,19 @@ import { beforeAll, describe, expect, it } from 'vitest';
 /**
  * Node-only exports must be statically resolvable from the server AND edge builds, since Next.js compiles
  * instrumentation modules for the edge runtime too. Otherwise named imports from `@sentry/nextjs` fail to compile
- * under Turbopack/webpack.
+ * under Turbopack/webpack. Exports that are Node-only get a no-op shim on the edge build (e.g. `vercelAIIntegration`).
  *
  *
  * Regression test for https://github.com/getsentry/sentry-javascript/issues/21317
  */
-describe('`pinoIntegration` is a statically detectable export from every runtime build', () => {
+describe('Node-only integrations are statically detectable exports from every runtime build', () => {
   const builds = {
     server: resolve(__dirname, '../build/cjs/index.server.js'),
     edge: resolve(__dirname, '../build/cjs/edge/index.js'),
   };
 
+  const dualRuntimeExports = ['pinoIntegration', 'vercelAIIntegration'];
+
   const staticExports: Record<string, string[]> = {};
 
   beforeAll(async () => {
@@ -26,7 +28,10 @@ describe('`pinoIntegration` is a statically detectable export from every runtime
     }
   });
 
-  it.each(Object.keys(builds))('statically exports `pinoIntegration` from the %s build', runtime => {
-    expect(staticExports[runtime]).toContain('pinoIntegration');
-  });
+  it.each(Object.keys(builds).flatMap(runtime => dualRuntimeExports.map(name => ({ runtime, name }))))(
+    'statically exports `$name` from the $runtime build',
+    ({ runtime, name }) => {
+      expect(staticExports[runtime]).toContain(name);
+    },
+  );
 });