[bot] Google GenAI: Chat session API (`chats.create`, `send_message`) not instrumented in Python (documented as traced in JS)

[braintrust-bot]

Summary

The google-genai Python SDK's Chat session API (client.chats.create(model=...) returning a Chat/AsyncChat with send_message()/send_message_stream()) has no dedicated instrumentation in this repo's Python integration. Braintrust's own docs confirm this capability is already designed and shipped for the legacy JS/TS SDK ("Chat spans (send_message, send_message_stream): message input and model as metadata") — it is simply missing from the Python integration.

What is missing

py/src/braintrust/integrations/google_genai/patchers.py and integration.py register patchers only for Models/AsyncModels (generate_content, generate_content_stream, embed_content, generate_images) and Interactions/GAOSInteractions (create/get/cancel/delete, sync and async). There is no patcher targeting google.genai.chats.Chats/AsyncChats/Chat/AsyncChat anywhere in the file, and no reference to chat, Chat, or send_message in tracing.py or patchers.py.

Verified against the pinned upstream version (google-genai==2.24.0, per py/uv.lock): Chat.send_message() does internally delegate to self._modules.generate_content(...) — i.e. it reuses the already-patched Models.generate_content path, so a generate_content span is still produced per turn. However, Chat/AsyncChat (subclass _BaseChat) separately maintains cross-turn conversation state — self._curated_history and self._comprehensive_history, updated via record_history() after every turn — that is invisible to any single generate_content span. There is no session-level span capturing the chat object's lifecycle, turn count, or curated-vs-comprehensive history divergence (e.g. when a turn is marked invalid and excluded from curated history sent to the model on subsequent turns).

Braintrust docs status: supported (JS/TS) / not_found (Python)

The Gemini integration page, Python section, documents only: content generation spans (generate_content/generate_content_stream), embedding spans (embed_content), image generation spans (generate_images), and interaction spans (interactions.create/get/cancel/delete). It contains zero mentions of chat, chats, send_message, or ChatSession in the Python section.

The same page's legacy-SDK section (for @google/generative-ai, JS/TS) explicitly documents: "Chat spans (send_message, send_message_stream): message input and model as metadata." This confirms Braintrust considers chat-session tracing a real, intended capability — it has just not been ported to the Python google-genai integration.

Upstream sources

• google-genai Python SDK, google/genai/chats.py (tag v2.24.0), classes Chat/AsyncChat/_BaseChat, factories Chats.create()/AsyncChats.create(): https://github.com/googleapis/python-genai/blob/v2.24.0/google/genai/chats.py
• Google GenAI Chat documentation: https://ai.google.dev/gemini-api/docs/text-generation#chat

Local repo files inspected

• py/src/braintrust/integrations/google_genai/integration.py (full file — patcher list)
• py/src/braintrust/integrations/google_genai/patchers.py (full file — no Chat/chats patcher)
• py/src/braintrust/integrations/google_genai/tracing.py (grepped for all top-level class/def names and chat — no matches)
• py/uv.lock (confirms pinned google-genai==2.24.0)