Machine providers, server access, bootstrap

A machine provider lets a plugin create compute, and the division of labor with core is strict: “core owns enrollment, the identity files and the internals of installing the daemon”, and “machine plugins do not clone projects”.

Once a new machine is attached, core itself configures the project’s git remote and registers its source, and only then calls the environment provider that needs a projectCheckout. A machine is not scoped to a project: “nothing in its creation names a project, and projects reach the machine later through project sources”.

Order. validate/availability → create (idempotent by key) → await checkpoint(resource) → bootstrap({ key, executor, report, signal }) → return { status:"created", name, resource }.

Removal. remove({ hostId, resource, … }) when a checkpoint exists; reconcileCleanup({ key, … }) when it does not: “find and remove indeterminate allocations by the durable key; return failed until the allocation intent is resolved”. Removing a persistent machine cascades through the environment providers first; removing an ephemeral one skips teardown and never resumes suspended compute.

Idle. Idle timing is yours, not core’s: experimental_thread.events + experimental_terminal.input to extend the deadline, and bb.background.schedule plus a startup reconciliation to sweep. “The dispatch hook helps report status, but it is bypassable and does not protect terminal/file RPC.”

availability. Checks run in parallel with a 5 s deadline each; “malformed output, exceptions and timeouts all look like unavailable”, and “a registered provider is not by itself enough”.