waken-gemini

A Gemini Target adapter for Waken — "nginx for AI agents." Routes waken Events to Google's google-genai SDK and returns the result as a waken Response.

Install

pip install waken-gemini

Usage

from waken import Runtime
from waken_gemini import GeminiAdapter

runtime = Runtime()
runtime.target("gemini", GeminiAdapter())
runtime.run()

GeminiAdapter reads GEMINI_API_KEY from the environment — that's google-genai's own convention (see genai.Client()), nothing waken-specific. Pass api_key=... (or any other genai.Client keyword) to override:

runtime.target("gemini", GeminiAdapter(model="gemini-2.5-pro", api_key="..."))

Sessions

Gemini's SDK, unlike some other providers, has no server-side resumable session id — the google-genai Chat object keeps conversation history client-side, in process memory. So GeminiAdapter maps a waken session_id to the SDK's Chat object itself (not a string id) and keeps reusing that same object turn over turn. A practical consequence: conversation history for a given waken session does not survive an adapter process restart, since there is nothing server-side to resume from. An Event with no session_id gets a fresh one-off Chat every call and is never stored.

Development

git clone https://github.com/WakenHQ/waken-gemini
cd waken-gemini
pip install -e ".[dev]"
pytest

Tests mock google.genai.Client entirely — no API key or network access needed to run the suite.

License

MIT