crash_recovery_default

Scope: domain • Default: prefer_resume

The fallback action when an agent crashes or is declared stuck by agent_health_strictness.

Positions

Position	Behavior
prefer_resume	Reload the latest session snapshot and continue. Safe because sessions are atomic.
prefer_restart_clean	Roll back to the snapshot before the crashed turn and retry. Useful when the last turn left state corrupted.
pause_for_review	Move the task to blocked and surface in the inbox. No automatic recovery.

Interaction with remediation_autonomy

crash_recovery_default chooses which recovery to attempt; remediation_autonomy decides whether the recovery needs approval. Example: crash_recovery_default: prefer_restart_clean + remediation_autonomy: confirm_destructive means: "try restart-clean, but ask me before rolling back destructive state."