[BOT ISSUE] Google GenAI: Interactions API not instrumented

[braintrust-bot]

Summary

The Google GenAI Interactions API (client.interactions.*) is not instrumented. This is a publicly documented Beta API (available since google-genai SDK v1.55.0) that provides a unified interface for agentic applications with server-managed conversation state, background execution, and tool orchestration.

The Interactions API is distinct from the Live API tracked in #125. The Live API is WebSocket-based real-time bidirectional streaming; the Interactions API is a request/response API for server-managed agentic conversations where the server handles conversation history, tool execution, and multi-turn state.

What is missing

The following client.interactions.* methods produce zero Braintrust tracing:

Method	Description
client.interactions.create()	Create and execute an interaction (the primary inference call)
client.interactions.get()	Retrieve a completed interaction
client.interactions.cancel()	Cancel an in-progress interaction
client.interactions.delete()	Delete an interaction

At minimum, interactions.create() should be instrumented with:

• Interaction-level span: Capturing model config, input content, output content, token usage, and duration
• Tool use spans: The Interactions API supports function calling, Google Search grounding, code execution, and MCP tools — tool calls and results should appear as nested spans
• Multi-turn linkage: The API uses previous_interaction_id for server-side conversation state management; this relationship should be captured in span metadata to enable trace-level visibility into multi-turn conversations

Key distinguishing features from generate_content

• Server manages conversation state via previous_interaction_id (no client-side history resending)
• Supports long-running/background execution (e.g., Deep Research agent)
• Supports streaming via the same create() call
• Has its own content types (ToolCallContent, GoogleMapsContent, etc.)
• Returns interaction status (COMPLETE, INCOMPLETE, etc.)

Current instrumentation coverage

Google GenAI Method	Instrumented?
models.generate_content()	Yes
models.generate_content_stream()	Yes
models.embed_content()	Yes
models.generate_images()	Yes
aio.live.connect()	No (#125)
interactions.create()	No
interactions.get()	No

Braintrust docs status

not_found — the Gemini integration page documents generate_content, streaming, function calling, structured outputs, thinking tokens, context caching, and grounding, but does not mention the Interactions API.

Upstream sources

• Interactions API documentation: https://ai.google.dev/gemini-api/docs/interactions
• google-genai Python SDK on PyPI (v1.70.0): https://pypi.org/project/google-genai/
• SDK GitHub repository: https://github.com/googleapis/python-genai
• API entry points: client.interactions.create(), .get(), .cancel(), .delete()
• Available since SDK v1.55.0, currently labeled Beta

Local files inspected

• py/src/braintrust/integrations/google_genai/patchers.py — defines 8 patchers (4 sync + 4 async) for generate_content, generate_content_stream, embed_content, generate_images; zero references to interactions
• py/src/braintrust/integrations/google_genai/tracing.py — wrapper functions for the 4 patched method families; no interactions wrapper
• py/src/braintrust/integrations/google_genai/integration.py — integration class registers the 8 patchers; no interactions patcher
• py/src/braintrust/integrations/google_genai/test_google_genai.py — no interactions test cases
• py/noxfile.py — test_google_genai(latest) session exists but does not test interactions