/agent-observability-experiment-bootstrap

Provider: Google Gemini / Vertex AI

Triggered by introspection (Workflow step 2.5) when the call-site function imports google.generativeai and calls:

• google.generativeai.GenerativeModel(...).generate_content(...)
• genai.GenerativeModel(...).generate_content(...) (aliased import)

{{PROVIDER_ASSERTS}} substitution

Either GEMINI_API_KEY or GOOGLE_API_KEY works — both are valid env names per google-generativeai's SDK conventions. Emit a single combined assert:

assert os.getenv("GEMINI_API_KEY") or os.getenv("GOOGLE_API_KEY"), (
    "GEMINI_API_KEY or GOOGLE_API_KEY is required for the wired task_fn."
)

Vertex AI variant (separate authentication path)

If the call-site uses vertexai.generative_models.GenerativeModel (not google.generativeai), the auth path is Google Cloud Application Default Credentials, not an API key. Emit:

# Vertex AI uses Google Cloud ADC, not an API key.
# Run `gcloud auth application-default login` before running this file, or set
# GOOGLE_APPLICATION_CREDENTIALS to point at a service account JSON.
assert os.getenv("GOOGLE_APPLICATION_CREDENTIALS"), (
    "GOOGLE_APPLICATION_CREDENTIALS path is required for the wired task_fn (Vertex AI), "
    "or run `gcloud auth application-default login` before invoking."
)

Adapter notes

• GenerativeModel("gemini-pro").generate_content("prompt") returns a GenerateContentResponse. Extract via .text (single-candidate) or .candidates[0].content.parts[0].text.
• For chat: model.start_chat(history=[]).send_message("prompt") returns the same response shape.
• Async: generate_content_async(...) — wrap with asyncio.run(...).

Common gotchas

• Safety filters: Gemini may return an empty response if safety thresholds block it. .text raises ValueError in that case. If the user's function doesn't handle this, surface a WARNING:.
• Quota lives at the project level for Vertex AI, per-key for google.generativeai. Be aware of which path the user's function uses.