Persistent Restart Loop Guard • AI Agent SDK

PersistentRestartLoopGuard: A :class:`RestartLoopGuard` whose events survive a process restart.

PersistentRestartLoopGuard

Defined in the protocols module.

A :class:RestartLoopGuard whose events survive a process restart.

The in-memory guard is reconstructed fresh on every boot, so a persisted turn that hard-crashes the whole process (OOM, segfault) can drive an infinite `supervisor → boot → auto-resume → crash loop the breaker never sees. This subclass loads its event history from an injected :class:RestartStoreProtocol` at construction and flushes back on every mutation, so the breaker observes the true cross-boot restart rate and can trip on a crash cycle slower than a single in-process window.

The window defaults to one hour (vs. the in-process 60s) because a cross-boot crash loop is paced by process (re)start latency, not by an in-loop retry. Pure decision preserved: all persistence is delegated to the injected store, so the trip logic remains provable in isolation.

Example::

guard = PersistentRestartLoopGuard(store, max_restarts=3,
                                   window_seconds=3600)
if guard.record(now=time.time()):
    # too many restart-interrupted boots — stop auto-resuming the turn
    ...

Constructor

Methods

Source