vercelAIIntegration doesn't activate on Vercel deployments without force: true

Problem

The vercelAIIntegration in @sentry/nextjs doesn't activate on Vercel deployments, causing AI SDK spans to have incorrect names:

On Vercel: Raw span names like ai.toolCall, ai.streamText (WRONG)
Locally: Transformed names like gen_ai.execute_tool, gen_ai.stream_text (CORRECT)

Root Cause

The Node SDK's vercelAIIntegration has conditional activation logic in afterAllSetup:

// packages/node/src/integrations/tracing/vercelai/index.ts:28-37
afterAllSetup(client) {
  const shouldForce = options.force ?? shouldForceIntegration(client);

  if (shouldForce) {
    addVercelAiProcessors(client);  // ← Span transformation enabled
  } else {
    instrumentation?.callWhenPatched(() => addVercelAiProcessors(client));
  }
}

addVercelAiProcessors is only called if:

force: true is passed, OR
Module detection succeeds (CJS mode only), OR
OTEL successfully patches the ai module

Why It Fails on Vercel

The ai package is intentionally NOT externalized in Next.js builds:

// packages/nextjs/src/config/withSentryConfig/constants.ts:5-8
// NOTE: 'ai' (Vercel AI SDK) is intentionally NOT included in this list.
// When externalized, Next.js doesn't properly handle the package's conditional exports,
// specifically the "react-server" export condition.

Because ai is bundled (not externalized):

OTEL's require hooks can't intercept module loading
instrumentation?.callWhenPatched() callback never fires
Module detection via shouldForceIntegration fails (requires CJS mode)
Result: addVercelAiProcessors() is never called on Vercel

Comparison: Edge vs Node Implementation

Edge implementation (vercel-edge/src/integrations/tracing/vercelai.ts):

setup(client) {
  addVercelAiProcessors(client);  // ALWAYS called unconditionally
}

Node implementation - conditional, requires force: true to guarantee activation in bundled environments.

Current Workaround

Users must explicitly add force: true in their server config:

// sentry.server.config.ts
import * as Sentry from "@sentry/nextjs";

Sentry.init({
  dsn: "...",
  integrations: [Sentry.vercelAIIntegration({ force: true })],
});

Suggested Fix

Consider one of these approaches:

Auto-detect Vercel environment: Default force: true when VERCEL env var is present
Document the workaround: Add clear documentation that force: true is required for Vercel deployments
Change default behavior: Make the Node implementation behave like Edge (always register processors)

Environment

@sentry/nextjs: Latest
Deployment: Vercel (serverless)
AI SDK: ai package from Vercel

JS-1627

Not sure if the react-server issue was fixed, if so, we can probably add it back and then it should work without passing the integration explicitely

#16736

updates here: The root cause here is that shouldForceIntegration() checks for 'ai' package by reading package.json from the filesystem. In serverless cold starts, filesystem isn't ready yet → read fails silently → empty result cached. So the integration isn't enforced by default due to missing ai package in the package.json. We are investigating a proper solution.

fix(nextjs): Populate __SENTRY_SERVER_MODULES__ in Turbopack (#19231)

A PR closing this issue has just been released 🚀

This issue was referenced by PR #19231, which was included in the 10.39.0 release.

Reopening — this still reproduces on Next.js 16 + Turbopack production builds (e.g. Vercel) on @sentry/nextjs@10.56.0. The symptom: AI SDK spans stay as raw ai.* (op: default) instead of being transformed to gen_ai.*. vercelAIIntegration({ force: true }) is still required as a workaround.

#19231 fixed the injection side (Turbopack now emits the value-injection rule), but the consumer can't see it. Root cause:

packages/node-core/src/integrations/modules.ts captures the value into a module-level const at evaluation time:

const SERVER_MODULES = typeof __SENTRY_SERVER_MODULES__ === 'undefined' ? {} : __SENTRY_SERVER_MODULES__;

That's correct for webpack (DefinePlugin replaces the token with a literal at build time), but Turbopack assigns globalThis.__SENTRY_SERVER_MODULES__ at runtime from the instrumentation.* value-injection loader — and the instrumentation file's ESM imports are hoisted above that assignment. Verified in a real Turbopack build (.next/server/chunks/[root-of-the-server]__*.js):

769449, e=>{ "use strict";
  var r = e.i(298962);                                  // imports @sentry/* → evaluates modules.ts (SERVER_MODULES = {})
  async function s(){ await e.A(145684) }
  globalThis.__SENTRY_SERVER_MODULES__ = {/* …deps incl "ai"… */}  // runs AFTER the import
}

So modules.ts evaluates before the global is set, SERVER_MODULES freezes to {}, and the other collectModules() sources are also empty on a bundled server (no full-dependency package.json at process.cwd(); ai bundled so absent from require.cache; ESM not CJS). Net: getModules().ai is undefined → shouldForceIntegration is false → addVercelAiProcessors never attaches. This affects all module-detection-based auto integrations (Vercel AI, OpenAI, Anthropic, Google GenAI, LangChain, LangGraph), not just Vercel AI, and also leaves event.modules empty on Turbopack.

Why CI didn't catch it: #19231's test asserts the config rule is generated (config-layer unit test), and the nextjs-16 AI E2E passes via the process.cwd() package.json fallback under local next start — which is unavailable on Vercel.

Fix in #21339: read __SENTRY_SERVER_MODULES__ lazily, supporting both the webpack token and the Turbopack runtime global, plus a regression test that fails on the old module-eval capture.

Thanks for re-opening and the detailed debugging description! We'll take a look.

fix(node-core): Read `__SENTRY_SERVER_MODULES__` lazily so Turbopack …

Navigation Menu

Repository navigation

vercelAIIntegration doesn't activate on Vercel deployments without force: true #19147

Problem

Root Cause

Why It Fails on Vercel

Comparison: Edge vs Node Implementation

Current Workaround

Suggested Fix

Environment

A PR closing this issue has just been released 🚀

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

vercelAIIntegration doesn't activate on Vercel deployments without force: true #19147

Description

Problem

Root Cause

Why It Fails on Vercel

Comparison: Edge vs Node Implementation

Current Workaround

Suggested Fix

Environment

Activity

linear commented on Feb 3, 2026

sergical commented on Feb 3, 2026

RulaKhaled commented on Feb 6, 2026

github-actions commented on Feb 16, 2026

A PR closing this issue has just been released 🚀

sergical commented on Jun 4, 2026

s1gr1d commented on Jun 8, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions