Failure recovery

Turn silent agent failures into bounded incident response.

Detect disconnected supervisors, repeated failed interventions, CI regressions, and exhausted budgets before they become runaway overnight jobs.

Without a supervisor

The problem

An unattended agent can look busy while disconnected, loop on the same failing action, or continue spending after the useful path is gone. A late log line should not make a stale process healthy again.

With Sleep Mode

The operating model

Sleep Mode uses durable heartbeats, failure counters, budgets, and explicit resume controls to turn uncertainty into a paused, reviewable state.

From vague autonomy to bounded operations.

01

Detect silence

A persisted watchdog pauses the run after four minutes without a newly accepted signed event.

02

Break loops

Three failed interventions trip a circuit breaker and stop new command leases.

03

Respect budgets

Cost and action deltas commit with a replay-proof signal ledger and pause at the armed ceiling.

04

Resume explicitly

Delayed journal replay never auto-resumes a paused run; an operator must understand the failure first.

Safety evidence built into the run
Server-receipt liveness Three-failure circuit breaker Immutable budgets Explicit-only resume