← All evidence
P-001 · Recovery

State survives the worker.

Live Reproducible Production

When a worker dies mid-task, AgentOS regenerates the resume packet directly from PostgreSQL state. The packet is the recovery state — not a memory of it. One real work order has been stood down and resumed 8 times.

Source of truth

Live work order in the API SOP project: build_rig. Resume packet: <wo>/standdown/checkpoint/RESUME.md, regenerated on demand from PG via agentos-hook standdown checkpoint --wo-dir <wo>.

Mechanism

The resume packet is not hand-authored. It is rendered by walking PG: open continuations, task states, work-package readiness, generated-artifact ledger (with SHA-256). Every field has a row in PG behind it.

Evidence
$ continue wo  wo:demo:platform-cleanup

  → re-rendering from PostgreSQL …

  Open continuations:        6
  Tasks:                    44
  Work packages:            33
  Resume generations:        8   (this WO has been stood down 8×)
  Generated artifacts:      56   (each with sha256 checksum)
  Stand-down state:         CHECKPOINTED · DORMANT
  Resume:                   pick up exactly where the worker died

  Same WO. Same PG state. Same checksum. Every time.
Reproduction
Source artifact
<work-order>/standdown/checkpoint/RESUME.md
Command
agentos-hook standdown checkpoint --wo-dir <wo>
Expected output
a regenerated RESUME packet enumerating every open continuation, task, work package, and generated artifact (each with sha256)
Verification
run the command twice at the same PG state; diff the two RESUME packets; expect identical bytes
Caveats
None. Mechanism is in production; numbers are from one live work order.