Skip to content

Tutorials

Cold-mode recovery

Pause, save, restore, and wake sandboxes from backend-specific snapshot state.

Cold mode lets a sandbox stop consuming a running VM while keeping recoverable state.

Every verb on this page is an advanced single-VM op reached through machine. They work, but they are hidden from machine --help.

Terminal window
cargo run -- machine pause agent-sandbox
cargo run -- machine snapshot ls
cargo run -- machine resume agent-sandbox

Firecracker pause/resume writes sealed instance state and verifies it before resume. This is the local mvm primitive.

Terminal window
cargo run -- machine checkpoint create agent-sandbox --class vm-full
cargo run -- machine checkpoint restore <checkpoint-id>

checkpoint restore takes the checkpoint’s own id, not the VM name.

Full-VM checkpoints capture full machine state. Among the selectable workload runners only hvf and apple-container support save/restore; firecracker, libkrun, qemu, and wasm do not. Check mvmctl doctor for the backend and capability on your host before relying on one. mvm records the checkpoint content hash in the audit chain when the launch plan and host signer are available, and restore records whether the content matched the prior chain entry.

  • Snapshot artifacts may contain guest memory, files, credentials, tokens, and browser/session state.
  • Restore support is backend-specific.
  • A restore is not a security boundary; it is a lifecycle transition.
  • Deleting a snapshot is not the same as proving physical erasure.
  • Published latency numbers must state backend, artifact, and readiness boundary.