What persists, and what does not

What survives a pause

• Files and working directories, including uncommitted edits an agent made while you were away.
• Git state: branches, stashes and remotes are where you left them.
• Anything you installed, plus dotfiles and shell history.
• The workspace's event log and its snapshot history.

What does not survive is anything that only existed in memory: running processes, an agent mid-turn, a dev server you started, a shell's unsaved scrollback. When a workspace comes back it boots from its disk, the same way a machine boots after a clean shutdown.

When the disk is sealed

• After every agent turn.
• On reachpad pause, and on an idle pause.
• At most every five minutes while the workspace is active, without stopping the guest.

So the worst case for a workspace whose node dies mid-run is five minutes of lost writes, not the whole session. reachpad pause --wait returns only once the save is durable, which is the one to use before you shut a laptop on work you care about.

A workspace pauses itself

A workspace pauses itself after 10 minutes with no terminal input and no supervised agent running in it. A supervised agent that is grinding away with nobody typing keeps its workspace alive for as long as it runs.

An idle pause is an ordinary pause: the disk is sealed first, and the next run or attach boots the workspace again. It stops the meter, which is the point.

Going back to an earlier state

Every seal leaves a checkpoint you can fork or rewind to. Automatic recovery history is managed rather than kept forever; retained forks and rewinds pin the checkpoints they depend on. See forks and rewind.