Python SDK

Both clients validate allowed, the returned event_id and the operation mode. A response for another event or an observation returned to a decision request raises QinzoraUnavailable; it never becomes an implicit allow. The event is serialized before sending so later caller mutation cannot change which response is expected.

qinzora.client exports QinzoraClient, AsyncQinzoraClient and QinzoraUnavailable. Both provide observe(event) and decide(event). Keep a client for the application lifecycle and close it on shutdown. No FastAPI dependency is imposed on the host; other languages can use HTTP.

from qinzora.client import QinzoraClient
from qinzora.models import Event


def is_allowed(url: str, key: str, event: Event) -> bool:
    with QinzoraClient(url, key) as guard:
        return guard.decide(event)["allowed"]

Construct Event from verified identity and resource ownership on the server. A false decision prevents the operation; true still requires host authorization. Handle QinzoraUnavailable as explicit unavailability, such as HTTP 503 without protected data.

from qinzora.client import AsyncQinzoraClient
from qinzora.models import Event


async def is_allowed(url: str, key: str, event: Event) -> bool:
    async with AsyncQinzoraClient(url, key) as guard:
        return (await guard.decide(event))["allowed"]

Default timeout is 2 seconds. Clients disable redirects/environment proxy inheritance and require a boolean allowed. Transport, status or invalid-response failures raise QinzoraUnavailable. There are no automatic retries or durable queues. An explicit retry after timeout must reuse the same Event/event_id: the server may already have committed. Async cancellation propagates and does not imply remote rollback. observe awaits the HTTP response; it is not a background queue.

See the executable documents API example and the AI adapter contract.