[bot] Google GenAI `chats` API bypasses wrapper proxy — `sendMessage()` calls produce zero spans

[braintrust-bot]

Summary

The @google/genai SDK's multi-turn chat API (ai.chats.create() → chat.sendMessage() / chat.sendMessageStream()) bypasses the wrapGoogleGenAI proxy entirely. Calls through the chats interface produce zero Braintrust spans, even though the underlying models.generateContent() method IS instrumented.

This is a proxy architecture bug, not a missing-channel issue — the generation calls physically avoid the instrumented code path.

Root cause

The GoogleGenAI constructor stores a direct reference to this.models into the Chats object at construction time:

// Inside GoogleGenAI constructor (upstream SDK):
this.models = new Models(this.apiClient);
this.chats = new Chats(this.models, this.apiClient);  // stores original models reference

The Braintrust wrapper applies its proxy after construction completes:

// js/src/wrappers/google-genai.ts
function wrapGoogleGenAIClass(OriginalGoogleGenAI) {
  return new Proxy(OriginalGoogleGenAI, {
    construct(target, args) {
      const instance = Reflect.construct(target, args);  // constructor runs first
      return wrapGoogleGenAIInstance(instance);            // proxy applied second
    },
  });
}

wrapGoogleGenAIInstance only intercepts models property access (line 74–78), but the Chats object already holds a reference to the original, unproxied Models instance. When chat.sendMessage() internally calls this.modelsModule.generateContent(), it uses the stored reference — completely bypassing the proxy.

What instrumentation is missing

All calls through the chats API are untraced:

SDK Method	Description	Traced?
ai.models.generateContent()	Direct model call	Yes
ai.models.generateContentStream()	Direct streaming call	Yes
ai.models.embedContent()	Direct embedding call	Yes
chat.sendMessage()	Multi-turn chat message	No
chat.sendMessageStream()	Multi-turn chat streaming	No

The chats API is the recommended way to do multi-turn conversations in the @google/genai SDK. Users following official Google documentation would get zero observability.

Braintrust docs status

unclear — The Braintrust Gemini integration page at https://www.braintrust.dev/docs/integrations/ai-providers/gemini documents wrapGoogleGenAI as providing automatic tracing. The chats API (sendMessage / sendMessageStream) is not explicitly mentioned as supported or unsupported.

Upstream reference

• Google GenAI JS SDK chats API: https://ai.google.dev/gemini-api/docs/text-generation#multi-turn-conversations
• SDK source (chats.ts): sendMessage() delegates to this.modelsModule.generateContent() — the stored reference, not a dynamic property access
• SDK source (client.ts): this.chats = new Chats(this.models, ...) — models reference captured during construction
• @google/genai npm package — this is a stable, documented API surface

Local files inspected

• js/src/wrappers/google-genai.ts — wrapGoogleGenAIClass (line 62–67) applies proxy post-construction; wrapGoogleGenAIInstance (line 70–80) only intercepts models property
• js/src/vendor-sdk-types/google-genai.ts — GoogleGenAIClient only declares models (line 17–19), no chats property
• js/src/instrumentation/plugins/google-genai-channels.ts — channels exist for generateContent and generateContentStream but are never reached via chats path
• e2e/scenarios/google-genai-instrumentation/scenario.impl.mjs — no test cases using the chats API