v2.7.0 closes the gap between “API exists” and “three lines in production.” Path activation means the default Lite path can enforce, alert, and attribute cost without Kafka — and the Python SDK can sit on the OpenAI client itself.
wrap(OpenAI()) — check and track on the hot path
pip install fluxmeter (1.4.0+), then FluxMeter(api_url="http://localhost:8000") for Lite. wrap(OpenAI(), meter, customer_id=...) runs a pre-call budget check and post-call track. BudgetExceededError means the provider was never called.
For streaming, wrap() reserves an estimate and can raise StreamKilledError mid-stream when estimated cost exceeds the reserve — so a long stream cannot silently overspend.
Historical note: older SDK modes could target Kafka directly; current repository SDK 2.0.0 is HTTP-only via Custody.
Hierarchy caps — hard max per agent run
POST /budget/{id}/cap with kind span or session sets a hard max spend. Pass parent_span_id or session_id on GET /budget/{id}/check. When spent + estimate exceeds the cap, the response returns reason=hierarchy_cap.
This is the ClipLive-style control: a single clip job cannot burn the whole creator wallet even if the monthly balance still has room.
Soft alert ladder on Lite ingest
POST /budget/{id}/webhook configures HTTPS delivery. On Lite, BUDGET_LOW, BUDGET_EXHAUSTED, and BUDGET_WARN fire after /ingest without Kafka. BUDGET_WARN fires at 70% and 90% of (initial_balance + topups), with warn_pct and spent_pct in the payload.
Full mode still uses the Kafka webhook-worker for LOW/EXHAUSTED; the WARN ladder is Lite-path today.
Try the path activation demo
Clone the engine repo and run demos/path_activation_demo.py (self-check, optional --live). Or make demo and exercise wrap(), /cap, and /webhook against localhost:8000.