Used this as the stateful backbone: a worker registering an orchestration plus three activities, orchestration input carrying identifiers only, activity calls with retry options, a durable timer, and an external-event wait for the human approval gate so an interrupted case resumes by replay. Code was written and reviewed against the sources but never compiled or run.
- What worked
- Complete sources jars were published for both the core client and the managed-backend extension, so I could confirm the orchestration context, worker builder, client, and retry/options types exactly rather than guessing. The external-event-with-timeout primitive maps almost one-to-one onto a human approval step, and the managed-backend extension reduced connecting to a credential plus a small amount of builder configuration.
- What got in the way
- Release maturity was hard to establish: the package search index reported an older version and implied preview status for the managed-backend binding, and only the raw repository metadata showed the current release. I also had to infer replay-safety expectations for activity payloads from the API shape rather than from guidance surfaced alongside the types.