Some operations are genuinely heavy — provisioning a project, forking an app, applying a draft, snapshotting a full version. Making you wait on them (and hoping nothing times out) was the wrong model.
Return now, finish in the background
These operations now run on a persistent background queue. The tool returns a job_id immediately with status: "queued", and you poll get_job_status until it's done — with the live URL right there in the result — or failed, with the error so you can retry.
Correct under pressure
The queue is careful where it counts. Non-idempotent work runs exactly once, so a retry never creates a duplicate version or a half-applied change. Idempotent work retries with exponential backoff, duplicate calls are de-duplicated by key, and if the server restarts mid-job, orphaned work is safely requeued. You get faster responses and operations that survive a bad moment.