Most workflow demos stop at the happy path. I wanted importable examples for the part that usually breaks after launch: bad webhook input, duplicate work, untrusted AI or OCR output, partial failure, and retries that repeat side effects.
I built three inactive, credential-free workflows:
- Inquiry routing — validates webhook input, returns explicit HTTP status codes, retries 503s, and carries an idempotency key.
- Exact-match quote guard — treats AI or OCR extraction as untrusted, ignores claimed prices, matches only a fixed catalog reference, and sends unknown items to review without a total.
- Editorial approval and publishing — keeps a draft, waits for approval, records each channel separately, and retries only failed channels.
What is included:
- Importable JSON with no real customer data
- Sticky notes explaining setup and production gaps
- Synthetic success and failure fixtures
- Docker labs pinned to n8n 2.33.7
- Measured summaries and SHA-256 hashes
- CI checks for inactive workflows, public text, and evidence hashes
For the quote guard I ran five cases: exact single item, exact multi-item, unknown reference, malicious claimed price, and invalid quantity. Unknown references return 422 with no totals; malformed input returns 400.
Repository:
Start with the problem table in the README, import one workflow, and keep it inactive until authentication and approved data sources are added.
The evidence is deliberately narrow: it proves the listed synthetic cases, not production readiness. I would especially value feedback on failure cases that are still missing.