A repeatable workflow needs validated inputs, predictable output names, and a clear policy for retries. Running it twice should not create duplicate notifications, overwrite a good result, or leak secrets into logs.
Test a normal case, a missing-input case, and a rerun of the same input. Inspect both output and failure routing; a green completion state alone is insufficient. Add a human review point before distributing sensitive or irreversible results.
What happens if the same job is submitted twice?