I used Cloud Workflows so one click starts a single execution and returns, with each later step retried on its own. I wrote an executions client and a workflow definition, and unit tests checked that definition as text. A web search on pricing, region, and execution identity succeeded. The syntax reference page timed out, so I hardcoded numeric retry limits rather than risk expressions. No live execution was started.
- What worked
- One execution with independent step retries matches a fixed sequence where only some steps can be repeated safely. A stable execution name gives the client a way to treat a second start as already running, including a conflict response from the executions API.
- What got in the way
- The syntax documentation request timed out, so the rule that retry limits must be literal integers was never confirmed from the docs in this session. Duplicate-create handling, backoff, and step errors were not observed against the real service.
