The kill is real. So is the resume.
Three historical runs, summarized in text. Each panel identifies the test environment and links a public summary with its date, result and scope. These summaries are editorial descriptions, not verbatim transcripts or current deployment assurances.
1The kill and the resume
The recorded gate-4 run: an AI coding agent mid-task, the microVM under it and the daemon supervising it hard-killed together, and the same session back at work with nothing lost.
Recorded on 2026-07-29 on the Apple Silicon evidence bench. A coding session was checkpointed, interrupted by a hard kill, and resumed. The recorded recovery completed with the workspace and session state preserved. 10 model calls were recorded; 29,871 replayed tokens were not purchased again. The recorded workload check passed. These are historical test results, not a recovery-time guarantee.
demo-report context · public result summary, with its own digest
what the agent did, from the record
Model calls recorded: 10
Replayed tokens not purchased again: 29,871
The recorded workload check passed.
Editorial summary: private identifiers and operating instructions are omitted. The linked summary retains the measured result, date and scope.
A whole-machine checkpoint under a live agent, then a kill that takes down the VM and its supervisor at once. Resume rebuilds the exact machine and the doctor checks it clean. The measured downtime is the honest headline; sub-second is roadmap and says so.
2One agent becomes three
Recorded on the production fleet host, 2026-09-03: a parent session is checkpointed once, then forked into three children. Same past, three futures, and a file written in branch A stays in branch A.
Historical team run · 2026-09-03 · fleet environment One parent checkpoint produced three child sessions. A file written in child A was present there and absent from children B and C. Each child retained a record of its parent checkpoint. This records branch isolation in the tested scenario; it is not a guarantee for every workload.
demo-fork-fleet context · public result summary, with its own digest
Editorial summary: private identifiers and operating instructions are omitted. The linked summary retains the measured result, date and scope.
Each child’s ledger begins with a fork-point naming the parent and its checkpoint, so every branch’s history is reconstructible to the divergence. This recording is operator-side on the fleet where customer sessions run, made before the fork verb shipped in the public client; the steered-divergence version under a live agent is the gate-5 bench record. phase-5
3The interrupted call is not bought twice
The gate-3 run: the VM is killed while a model call is in flight. On resume, the gateway serves the stored response instead of re-buying it, and the ledger shows the replayed tokens excluded from spend.
Historical replay test · laboratory environment A model call was interrupted and the session resumed. The client retried; the completed response was replayed without a second purchase. The session histories reconciled. Client retry behavior is a separate limitation; external side effects are not rolled back.
phase-3 context · public result summary, with its own digest
Editorial summary: private identifiers and operating instructions are omitted. The linked summary retains the measured result, date and scope.
The historical test replayed 32,291 tokens with no additional charge on the KeepState ledger. This is a completed-response replay result in the tested environment; it does not guarantee client retries, reverse external effects or settle the provider’s own accounting.
Try it yourself, free
Instant signup. Your first kill-and-resume is free, no card. Run the same kill on your own workload and watch it come back.
Try it yourself, freeType the kill yourself
The supervised interruption is a conversation, not a feature you can start here: a 30-minute call where you type the kill -9 on our bench and watch the session come back with its work intact.