vercelAIIntegration: record AI SDK v7 runtimeContext (filtered by telemetry.includeRuntimeContext) as span attributes

[BrandonNoad]

Problem Statement

AI SDK v7 removed telemetry.metadata. Per-call metadata now goes in runtimeContext, and the app opts keys into telemetry with telemetry.includeRuntimeContext (telemetry docs, runtime context docs). @ai-sdk/otel records the included keys as ai.settings.runtimeContext.*. On v6 the equivalent data arrived through experimental_telemetry.metadata as vercel.ai.telemetry.metadata.*.

The v7 channel subscriber in @sentry/server-utils (vercel-ai-dc-subscriber, added in #21613) never reads runtimeContext, so on v7 there is no way to attach request-level metadata (user, tenant, request id, ...) to gen_ai spans. The integration docs say not to combine registerTelemetry with vercelAIIntegration, so registering @ai-sdk/otel alongside it is not a workaround either.

await streamText({
  model,
  messages,
  runtimeContext: { requestId: 'req_123', tenantId: 'acme', apiToken: 'secret' },
  telemetry: {
    isEnabled: true,
    functionId: 'support-chat',
    includeRuntimeContext: { requestId: true, tenantId: true },
  },
});

• Expected: the gen_ai.invoke_agent span (and ideally its gen_ai.generate_content / gen_ai.execute_tool children) carries requestId and tenantId, and not apiToken.
• Actual: none of the keys appear. The span has only vercel.ai.operationId, vercel.ai.model.provider, vercel.ai.settings.maxRetries and the gen_ai.* attributes.

Checked against @sentry/node 10.71.0 and the @sentry/server-utils subscriber in 10.75.3 and 11.0.0, with ai 7.0.100 / 7.0.113 on Node.

Solution Brainstorm

• In buildInvokeAgentSpan, set one attribute per event.runtimeContext key. vercel.ai.settings.runtimeContext.<key> would match @ai-sdk/otel's ai.settings.runtimeContext.* under the existing ai.* to vercel.ai.* renaming.
• Copy the same attributes onto child generate_content / execute_tool spans through the existing callId maps (like operationIdByCallId), since those channel events do not carry runtimeContext.
• Optionally, record event.toolContext on execute_tool spans, which the AI SDK filters by telemetry.includeToolsContext.

Dependency on the AI SDK. In ai <= 7.0.113 the channel publishes runtimeContext / toolsContext unfiltered, ignoring includeRuntimeContext (vercel/ai#21442). Reading event.runtimeContext on those versions would send keys the app never opted in to telemetry. The fix is open in vercel/ai#21447, which filters channel events the same way as registered integrations. So this should only read the field on AI SDK versions that include that fix, or otherwise guard against unfiltered context.

Additional Context

#24660 notes the same gap from the Cloudflare Think side ("neither subscriber reads either field").

[linear-code]

JS-3788

[isaacs]

Thanks for the detailed report and for tracking down the AI SDK side of this!

You're right. The v7 channel subscriber never reads runtimeContext, so there is currently no way to attach per-call metadata to gen_ai spans on ai 7. It's a gap in the feature: on v6 the metadata came from the AI SDK's own OTel attributes, and nothing replaced that when telemetry.metadata went away.

We don't want to read event.runtimeContext until the AI SDK filters it. As you noted, ai <= 7.0.113 publishes the full context to the channel. The channel event also doesn't carry the includeRuntimeContext allowlist, so we can't apply the filter ourselves. And a filtered context looks the same as an unfiltered one, so we can't detect which case we're in. If we read the field today, apps on those versions would send keys they never opted in, like the apiToken in your example.

vercel/ai#21447 merged today but isn't in a release yet. Once a release includes it, we can add support along the lines you suggested:

• vercel.ai.settings.runtimeContext.<key> on the invoke_agent span
• the same attributes copied to the child generate_content and execute_tool spans
• toolContext on execute_tool spans

We'll document the minimum ai version for this once it's added. Until then, as a workaround you can set attributes on the enclosing span yourself, for example Sentry.getActiveSpan()?.setAttribute('tenantId', 'acme') inside a Sentry.startSpan around the call.

While looking into this, I also found and confirmed a regression in SDK v11: experimental_telemetry.metadata on ai 6 no longer reaches the gen_ai spans. SDK v10 recorded it as vercel.ai.telemetry.metadata.*. I'll post a separate issue for that.

[BrandonNoad]

@isaacs the vercel/ai fix was published in v7.0.114