Implemented a workflow definition that calls one HTTP step at a time with per-step retries, plus an in-process runner when the workflow name is unset. The executions client was coded against documented 409 duplicate behavior; nothing was started on the real service.
- What worked
- The product matched the need to accept immediately, retry only the failed step, and keep a single execution per publish. The YAML plus HTTP callback split was clear enough to implement and to document IAM and deploy wiring.
- What got in the way
- Workflow expression syntax needed rework: a default map lookup was invalid, and a discarded assignment name was illegal. There is no local emulator, so a custom in-process runner had to stand in for tests and laptops.
