fix(server-utils): Restore request config attributes on google-genai chat spans

dev-packages/cloudflare-integration-tests/suites/tracing/google-genai/index.ts

@@ -53,12 +53,35 @@ export default Sentry.withSentry(
       const client = Sentry.instrumentGoogleGenAIClient(new GoogleGenAI({ apiKey: 'mock-api-key' }));
 
       // Test 1: chats.create and sendMessage flow
+      // Every field here is read back off the chat instance for each message the chat sends, so
+      // the create config covers all of them at once.
       const chat = client.chats.create({
         model: 'gemini-1.5-pro',
         config: {
           temperature: 0.8,
           topP: 0.9,
+          topK: 40,
           maxOutputTokens: 150,
+          frequencyPenalty: 0.5,
+          presencePenalty: 0.3,
+          systemInstruction: 'You are a friendly robot.',
+          tools: [
+            {
+              functionDeclarations: [
+                {
+                  name: 'controlLight',
+                  parametersJsonSchema: {
+                    type: 'object',
+                    properties: {
+                      brightness: { type: 'number' },
+                      colorTemperature: { type: 'string' },
+                    },
+                    required: ['brightness', 'colorTemperature'],
+                  },
+                },
+              ],
+            },
+          ],
         },
         history: [
           {

dev-packages/cloudflare-integration-tests/suites/tracing/google-genai/test.ts

@@ -5,11 +5,16 @@ import {
   GEN_AI_INPUT_MESSAGES,
   GEN_AI_OPERATION_NAME,
   GEN_AI_PROVIDER_NAME,
+  GEN_AI_REQUEST_FREQUENCY_PENALTY,
   GEN_AI_REQUEST_MAX_TOKENS,
   GEN_AI_REQUEST_MODEL,
+  GEN_AI_REQUEST_PRESENCE_PENALTY,
   GEN_AI_REQUEST_TEMPERATURE,
+  GEN_AI_REQUEST_TOP_K,
   GEN_AI_REQUEST_TOP_P,
   GEN_AI_RESPONSE_TEXT,
+  GEN_AI_SYSTEM_INSTRUCTIONS,
+  GEN_AI_TOOL_DEFINITIONS,
   GEN_AI_USAGE_INPUT_TOKENS,
   GEN_AI_USAGE_OUTPUT_TOKENS,
   GEN_AI_USAGE_TOTAL_TOKENS,
@@ -49,7 +54,24 @@ it('traces Google GenAI chat, generateContent, and embedContent calls', async ({
           [GEN_AI_PROVIDER_NAME]: { value: 'google_genai', type: 'string' },
           [GEN_AI_OPERATION_NAME]: { value: 'chat', type: 'string' },
           [GEN_AI_REQUEST_MODEL]: { value: 'gemini-1.5-pro', type: 'string' },
+          // Set once on `chats.create()` and reused for every message the chat sends.
+          [GEN_AI_REQUEST_TEMPERATURE]: { value: 0.8, type: 'double' },
+          [GEN_AI_REQUEST_TOP_P]: { value: 0.9, type: 'double' },
+          [GEN_AI_REQUEST_TOP_K]: { value: 40, type: 'integer' },
+          [GEN_AI_REQUEST_MAX_TOKENS]: { value: 150, type: 'integer' },
+          [GEN_AI_REQUEST_FREQUENCY_PENALTY]: { value: 0.5, type: 'double' },
+          [GEN_AI_REQUEST_PRESENCE_PENALTY]: { value: 0.3, type: 'double' },
+          [GEN_AI_TOOL_DEFINITIONS]: {
+            value:
+              '[{"name":"controlLight","parametersJsonSchema":{"type":"object","properties":{"brightness":{"type":"number"},"colorTemperature":{"type":"string"}},"required":["brightness","colorTemperature"]}}]',
+            type: 'string',
+          },
           // collect LLM input and outputs (default true)
+          [GEN_AI_SYSTEM_INSTRUCTIONS]: {
+            value: '[{"type":"text","content":"You are a friendly robot."}]',
+            type: 'string',
+          },
+          // The create `history` stays off the span; only the message being sent is reported.
           [GEN_AI_INPUT_MESSAGES]: { value: '[{"role":"user","content":"Tell me a joke"}]', type: 'string' },
           [GEN_AI_RESPONSE_TEXT]: { value: 'Hello from Google GenAI!', type: 'string' },
           [GEN_AI_USAGE_INPUT_TOKENS]: { value: 8, type: 'integer' },

dev-packages/node-integration-tests/suites/tracing/google-genai-v2/scenario.mjs

@@ -54,7 +54,8 @@ async function run() {
     });
 
     // Test 1: chats.create and sendMessage flow
-    // This should generate two spans: one for chats.create and one for sendMessage
+    // `chats.create()` builds a local object and emits no span; only `sendMessage` does, and it
+    // reports the config given here.
     const chat = client.chats.create({
       model: 'gemini-1.5-pro',
       config: {

dev-packages/node-integration-tests/suites/tracing/google-genai/scenario.mjs

@@ -54,7 +54,8 @@ async function run() {
     });
 
     // Test 1: chats.create and sendMessage flow
-    // This should generate two spans: one for chats.create and one for sendMessage
+    // `chats.create()` builds a local object and emits no span; only `sendMessage` does, and it
+    // reports the config given here.
     const chat = client.chats.create({
       model: 'gemini-1.5-pro',
       config: {

dev-packages/node-integration-tests/suites/tracing/google-genai/test.ts

@@ -50,6 +50,11 @@ describe('Google GenAI integration', () => {
             expect(chatSpan!.attributes['sentry.origin'].value).toBe(EXPECTED_ORIGIN);
             expect(chatSpan!.attributes[GEN_AI_PROVIDER_NAME].value).toBe('google_genai');
             expect(chatSpan!.attributes[GEN_AI_REQUEST_MODEL].value).toBe('gemini-1.5-pro');
+            // Given once to `chats.create()` and reused for every message the chat sends.
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_TEMPERATURE].value).toBe(0.8);
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_TOP_P].value).toBe(0.9);
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_MAX_TOKENS].value).toBe(150);
+            expect(chatSpan!.attributes[GEN_AI_SYSTEM_INSTRUCTIONS]).toBeUndefined();
             expect(chatSpan!.attributes[GEN_AI_USAGE_INPUT_TOKENS].value).toBe(8);
             expect(chatSpan!.attributes[GEN_AI_USAGE_OUTPUT_TOKENS].value).toBe(12);
             expect(chatSpan!.attributes[GEN_AI_USAGE_TOTAL_TOKENS].value).toBe(20);
@@ -104,6 +109,13 @@ describe('Google GenAI integration', () => {
             expect(chatSpan!.attributes[GEN_AI_REQUEST_MODEL].value).toBe('gemini-1.5-pro');
             expect(chatSpan!.attributes[GEN_AI_INPUT_MESSAGES]).toBeDefined();
             expect(chatSpan!.attributes[GEN_AI_RESPONSE_TEXT]).toBeDefined();
+            // Given once to `chats.create()` and reused for every message the chat sends.
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_TEMPERATURE].value).toBe(0.8);
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_TOP_P].value).toBe(0.9);
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_MAX_TOKENS].value).toBe(150);
+            expect(chatSpan!.attributes[GEN_AI_SYSTEM_INSTRUCTIONS].value).toBe(
+              '[{"type":"text","content":"You are a friendly robot who likes to be funny."}]',
+            );
             expect(chatSpan!.attributes[GEN_AI_USAGE_INPUT_TOKENS].value).toBe(8);
             expect(chatSpan!.attributes[GEN_AI_USAGE_OUTPUT_TOKENS].value).toBe(12);
             expect(chatSpan!.attributes[GEN_AI_USAGE_TOTAL_TOKENS].value).toBe(20);
@@ -275,6 +287,10 @@ describe('Google GenAI integration', () => {
             expect(chatSpan!.status).toBe('ok');
             expect(chatSpan!.attributes[GEN_AI_OPERATION_NAME].value).toBe('chat');
             expect(chatSpan!.attributes[GEN_AI_RESPONSE_STREAMING].value).toBe(true);
+            // Given once to `chats.create()` and reused for every message the chat sends.
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_TEMPERATURE].value).toBe(0.8);
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_TOP_P].value).toBe(0.9);
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_MAX_TOKENS].value).toBe(150);
             expect(chatSpan!.attributes[GEN_AI_RESPONSE_ID].value).toBe('mock-response-streaming-id');
             expect(chatSpan!.attributes[GEN_AI_RESPONSE_MODEL].value).toBe('gemini-1.5-pro');
 
@@ -342,6 +358,10 @@ describe('Google GenAI integration', () => {
             expect(chatSpan!.attributes[GEN_AI_OPERATION_NAME].value).toBe('chat');
             expect(chatSpan!.attributes[GEN_AI_RESPONSE_STREAMING].value).toBe(true);
             expect(chatSpan!.attributes[GEN_AI_INPUT_MESSAGES]).toBeDefined();
+            // Given once to `chats.create()` and reused for every message the chat sends.
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_TEMPERATURE].value).toBe(0.8);
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_TOP_P].value).toBe(0.9);
+            expect(chatSpan!.attributes[GEN_AI_REQUEST_MAX_TOKENS].value).toBe(150);
             expect(chatSpan!.attributes[GEN_AI_RESPONSE_FINISH_REASONS].value).toBe('["STOP"]');
 
             const blockedSpan = container.items.find(span => span.name === 'generate_content blocked-model');

packages/server-utils/src/ai/google-genai/index.ts

@@ -251,6 +251,37 @@ export function addResponseAttributes(span: Span, response: GoogleGenAIResponse,
   }
 }
 
+/**
+ * Recover the config a chat message sends but does not carry in its own arguments.
+ *
+ * `chats.create()` takes a config that `@google/genai` reuses for every message, resolving each
+ * request as `params.config ?? chat.config`. A per-message config therefore replaces the chat
+ * config rather than merging into it, and the chat config applies only when the message omits one.
+ * That config lives on the chat instance, which both instrumentation paths already hold: the client
+ * proxy passes it as the method's `context`, the diagnostics-channel path as `data.self`.
+ *
+ * The chat `history` stays off the message spans. The SDK does send it, folded into `contents`, and
+ * the instance carries the whole transcript, but repeating every past turn on every message span
+ * duplicates what earlier spans already reported and grows without bound.
+ */
+export function resolveChatParams(
+  operationName: string,
+  params: Record<string, unknown> | undefined,
+  context: unknown,
+): Record<string, unknown> | undefined {
+  // `!= null` mirrors the SDK's `??`: an explicit `null` config falls back to the chat's.
+  if (operationName !== 'chat' || params?.config != null || !context || typeof context !== 'object') {
+    return params;
+  }
+
+  const chatConfig = (context as Record<string, unknown>).config;
+  if (!chatConfig || typeof chatConfig !== 'object') {
+    return params;
+  }
+
+  return { ...params, config: chatConfig };
+}
+
 /**
  * Instrument any async or synchronous genai method with Sentry spans
  * Handles operations like models.generateContent and chat.sendMessage and chats.create
@@ -269,7 +300,8 @@ function instrumentMethod<T extends unknown[], R>(
     apply(target, _, args: T): R | Promise<R> {
       const operationName = instrumentedMethod.operation || 'unknown';
       const params = args[0] as Record<string, unknown> | undefined;
-      const requestAttributes = extractRequestAttributes(operationName, params, context);
+      const attributeParams = resolveChatParams(operationName, params, context);
+      const requestAttributes = extractRequestAttributes(operationName, attributeParams, context);
       const model = requestAttributes[GEN_AI_REQUEST_MODEL] || 'unknown';
       const client = getClient();
       // With span streaming, omit the `'unknown'` model sentinel so the name stays low-cardinality.
@@ -289,8 +321,8 @@ function instrumentMethod<T extends unknown[], R>(
           },
           async (span: Span) => {
             try {
-              if (options.recordInputs && params) {
-                addPrivateRequestAttributes(span, params, operationName);
+              if (options.recordInputs && attributeParams) {
+                addPrivateRequestAttributes(span, attributeParams, operationName);
               }
               const stream = await target.apply(context, args);
               return instrumentStream(stream, span, Boolean(options.recordOutputs)) as R;
@@ -310,8 +342,8 @@ function instrumentMethod<T extends unknown[], R>(
           attributes: requestAttributes,
         },
         (span: Span) => {
-          if (options.recordInputs && params) {
-            addPrivateRequestAttributes(span, params, operationName);
+          if (options.recordInputs && attributeParams) {
+            addPrivateRequestAttributes(span, attributeParams, operationName);
           }
 
           // `onError` is a no-op because the rejection is rethrown to the caller and `startSpan` already

packages/server-utils/src/integrations/google-genai.ts

@@ -12,7 +12,12 @@ import {
   startInactiveSpan,
 } from '@sentry/core';
 import { getGenAiSpanOp, resolveAIRecordingOptions } from '../ai/core/utils';
-import { addPrivateRequestAttributes, addResponseAttributes, extractRequestAttributes } from '../ai/google-genai';
+import {
+  addPrivateRequestAttributes,
+  addResponseAttributes,
+  extractRequestAttributes,
+  resolveChatParams,
+} from '../ai/google-genai';
 import { instrumentStream } from '../ai/google-genai/streaming';
 import type { GoogleGenAIOptions, GoogleGenAIResponse } from '../ai/google-genai/types';
 import { CHANNELS } from '../orchestrion/channels';
@@ -103,7 +108,9 @@ function createGenAiSpan(
   }
 
   const args = data.arguments ?? [];
-  const params = args[0] as Record<string, unknown> | undefined;
+  // `chat.sendMessage()` carries no config of its own unless the caller passes one, so recover the
+  // one `chats.create()` left on the chat instance that the transform stashed in `data.self`.
+  const params = resolveChatParams(operation, args[0] as Record<string, unknown> | undefined, data.self);
 
   const { recordInputs } = resolveAIRecordingOptions(options);