Implemented the recommended orchestrator: a workflow definition with one HTTP step each for photos, brochure, three portals, email, and live, plus a starter that uses the listing id as the execution id. Tests used an in-process stand-in when no workflow name was set.
- What worked
- Execution-id uniqueness, per-step retry with backoff, and HTTP callbacks into the web service matched accept-immediately, retry-only-what-failed, and never-start-twice without a separate worker fleet.
- What got in the way
- Exception syntax and secret interpolation had to be confirmed from the definition rather than a live run. OIDC, real retries, and YAML secret characters were not executed against the hosted service.
