fix(server-utils): Keep OpenAI decision confidence on Vercel AI decide spans (#25457)

## What

Decide spans from Vercel AI `experimental_decide` with
`openai.decisionModel()` now keep the confidence of each answer.

## Why

The OpenAI provider moves confidence out of the answers into its
provider metadata, so before this change Sentry lost it. TypeSafe models
already kept it.

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

Author: andreiborza

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

@@ -701,13 +701,15 @@ function getOutputMessages(
 }
 
 /**
- * The AI SDK TypeSafe provider moves each answer's `confidence` out of the answers into
- * `providerMetadata.typesafe.confidence` (keyed by question id). Put it back so decide answers keep it.
+ * The AI SDK TypeSafe and OpenAI providers move each answer's `confidence` out of the answers into
+ * `providerMetadata.<provider>.confidence` (keyed by question id). Put it back so decide answers keep it.
  */
 function withProviderConfidence(result: Record<string, unknown>): unknown {
   const { answers, providerMetadata } = result;
-  const typesafe = isObjectLike(providerMetadata) ? providerMetadata.typesafe : undefined;
-  const confidence = isObjectLike(typesafe) && isObjectLike(typesafe.confidence) ? typesafe.confidence : undefined;
+  const metadata = isObjectLike(providerMetadata)
+    ? [providerMetadata.typesafe, providerMetadata.openai].find(m => isObjectLike(m) && isObjectLike(m.confidence))
+    : undefined;
+  const confidence = isObjectLike(metadata) && isObjectLike(metadata.confidence) ? metadata.confidence : undefined;
   if (!confidence || !isObjectLike(answers)) {
     return answers;
   }

packages/server-utils/test/integrations/vercel-ai/decide.test.ts

@@ -1,4 +1,5 @@
 import { afterEach, beforeEach, describe, expect, it } from 'vitest';
+import { GEN_AI_OUTPUT_MESSAGES } from '@sentry/conventions/attributes';
 import { GEN_AI_DECIDE } from '@sentry/conventions/op';
 import { getMainCarrier, setCurrentClient, spanToStaticSpanJSON } from '@sentry/core';
 import type { Span } from '@sentry/core';
@@ -41,7 +42,7 @@ describe('Vercel AI decide spans', () => {
         recordOutputs: true,
       },
       result: { answers: { wantsRefund: { type: 'boolean', probability: 0.99 } } },
-    } as unknown as Message;
+    } satisfies Message;
 
     const span = createSpanFromMessage(message, {} as Parameters<typeof createSpanFromMessage>[1])!;
     expect(span).toBeDefined();
@@ -51,8 +52,35 @@ describe('Vercel AI decide spans', () => {
     const json = spanToStaticSpanJSON(endedSpans[0]!);
     expect(json.description).toBe('decide typesafe-ai/jev');
     expect(json.op).toBe(GEN_AI_DECIDE);
-    expect(JSON.parse(json.data?.['gen_ai.output.messages'] as string)).toEqual([
+    expect(JSON.parse(json.data?.[GEN_AI_OUTPUT_MESSAGES] as string)).toEqual([
       { type: 'evaluation', answers: { wantsRefund: { type: 'boolean', probability: 0.99 } } },
     ]);
   });
+
+  it.each(['typesafe', 'openai'])('restores answer confidence from `providerMetadata.%s`', provider => {
+    const message = {
+      type: 'experimental_decide',
+      event: {
+        provider,
+        modelId: 'gpt-6-luna',
+        state: 'I was charged twice',
+        questions: { department: { type: 'choice', criteria: { billing: 'Charges', support: 'Other' } } },
+        recordInputs: true,
+        recordOutputs: true,
+      },
+      result: {
+        answers: { department: { type: 'choice', choice: 'billing' } },
+        providerMetadata: { [provider]: { confidence: { department: 0.91 } } },
+      },
+    } satisfies Message;
+
+    const span = createSpanFromMessage(message, {} as Parameters<typeof createSpanFromMessage>[1])!;
+    enrichSpanOnEnd(span, message, {} as Parameters<typeof enrichSpanOnEnd>[2]);
+    span.end();
+
+    const json = spanToStaticSpanJSON(endedSpans[0]!);
+    expect(JSON.parse(json.data?.[GEN_AI_OUTPUT_MESSAGES] as string)).toEqual([
+      { type: 'evaluation', answers: { department: { type: 'choice', choice: 'billing', confidence: 0.91 } } },
+    ]);
+  });
 });