Enforcement Modes

The enforcement controller has two modes: kill (deterministic, immediate termination) and alert (pause + human approval). The self-hosted, open-source SteerPlane API runs kill mode only — that's what's fully functional today, out of the box, for free.

Kill mode (self-hosted, free)

On any violation, execution terminates immediately and atomically with a structured diagnostic exception that is logged and persisted. This is the default, and the only enforcement path implemented in the open-source, self-hosted API.

@guard(agent_name="batch_job", max_cost_usd=20)
def run():
    agent.run()

Use kill mode for unattended jobs where any breach should stop the run outright — this is what the free, self-hosted SteerPlane API enforces for every limit type.

Alert mode (hosted / enterprise)

The SDK also accepts enforcement="alert", alert_threshold, and alert_timeout_sec — these parameters exist client-side, but the pause → notify → human-approve → resume workflow they describe requires the hosted/enterprise control plane, which isn't part of the open-source self-hosted stack.

@guard(
    agent_name="support_bot",
    max_cost_usd=10,
    enforcement="alert",       # requires a hosted/enterprise backend
    alert_threshold=0.8,
    alert_timeout_sec=1800,
)
def run():
    agent.run()

Pointed at a backend that implements it: approve resumes execution with extended limits; deny or timeout terminates as a safety net. Pointed at the free self-hosted API (which does not implement this workflow): the run fails closed and terminates immediately — see graceful degradation below.

Graceful degradation

SteerPlane is fault-tolerant by design.

• If the control plane (API server) becomes unreachable, the SDK keeps enforcing locally — loop detection, cost, step, and policy rules stay active regardless.
• If enforcement="alert" is set but the connected API doesn't implement the approval workflow (true of the free self-hosted API), the SDK fails closed: the run terminates safely with a clear "could not create an approval request" error, rather than continuing unprotected or hanging indefinitely.

The result: no agent ever runs without at least one active enforcement path.