Sandbox Lifecycle

Agent calls Sandbox.create() (e.g. Sandbox.create() / await Sandbox.create()) to spin up a fresh microVM. The call returns a sandbox handle carrying a sandbox_id before any code runs.

Persist the sandbox_id with the operator's task key BEFORE returning control to the caller, so a crash between create and first use does not orphan a running, billed microVM. Treat create as an allocation event: it consumes quota and starts the timeout clock. Always pair create with a guaranteed ki…

A long-running agent session persisted a sandbox_id. A later worker process calls Sandbox.connect(sandbox_id) to resume control of the same microVM rather than creating a new one.

Reconnect to a still-alive sandbox by its sandbox_id instead of creating a new one, preserving in-sandbox filesystem and process state. Handle the case where the sandbox has already timed out or been killed: connect should surface a not-found / expired error, and the agent must recreate rather than…

An interactive agent UI creates a sandbox on the user's first message. The create call adds noticeable startup latency before the first code runs.

Treat sandbox creation as a measurable cost on the critical path: create the sandbox ahead of the first interaction (pre-warm) when the workflow allows, or surface a loading state. Do not pin a hardcoded startup-time number into an SLA. [REQUIRES-VERIFICATION] for current microVM cold-start latency…