Concurrent Writes

Several mounts can write to the same file system at once. Autosave sends each mount’s settled work through its session WAL. Before acknowledging a checkpoint, the server orders that update and applies its changed paths to the file system’s single shared timeline. A file system is a shared drive, not a branching repository — so reconciliation is simple and automatic, with one rule for the rare overlap.

​

Disjoint changes merge automatically

Replicated updates that touch different paths — or different files in the same directory — reconcile cleanly and never notice each other. The server rebases each update’s changed paths onto the current state of the timeline, so two agents writing to sessions/a/ and sessions/b/ both land, in full, with no coordination and no conflict. This is the common case, and it’s the case worth designing for: partition writers by directory or file and concurrent access is effortless. It’s exactly how one file system per user, mounted across many sandboxes, is meant to work — each session owns its own subtree.

​

Same-path writes are last-writer-wins

If two mounts change the same file concurrently, the later server-ordered checkpoint wins. The file on the timeline becomes the later writer’s version; the earlier writer’s change to that file is overwritten. There are no conflict markers, no merge to resolve, and no manual step — a file system behaves like a shared disk, where the last published update to a path is the one that sticks.

​

Making concurrency safe

Because same-path overwrites are silent, the way to keep concurrent agents from stepping on each other is to keep their writes disjoint:

• Give each agent or task its own subtree (sessions/<id>/, users/<id>/). This is the natural shape for agent workloads and eliminates same-path contention entirely.
• Treat shared files as append-only or single-owner. If one file must be written by many agents, funnel writes through one of them, or have each agent write its own file and merge at read time.
• Wait for settled-write replication, not an explicit snapshot. A small quiet edit normally reaches the shared head in about a second plus upload time; continuous writers flush at least every 5 seconds. Following mounts refresh after that checkpoint commits. They do not see bytes that are still in another mount’s open local batch.

​

Next Steps