Persistence and recovery
Preserve identity, homes, sessions, ownership and unfinished work across failure.
Persistence is the product boundary between a useful session and a durable organization. Pod loss must not imply work loss.
An Agent’s identity and Assignment live in PostgreSQL. Kubernetes owns the workload. Its PVC owns home state, private memory and unfinished worktrees. Herdr owns the native session. Git owns committed delivery. Recovery reconciles these authorities rather than pretending one healthy Pod proves the whole outcome is intact.
Recovery example
A ship Crewmate has uncommitted validation changes when its Pod is deliberately replaced. Supervision observes the missing or changed workload, verifies the intended home PVC, recreates or resumes the reviewed runtime shape and confirms the native harness is processing the existing Assignment. The preserved worktree continues; no duplicate Assignment is created.
The recovery report distinguishes detection time, return to useful work, preserved changes and any human repair. Missing telemetry is unknown, not zero. An Agent is not “recovered” merely because a terminal can be opened.
Last updated on