A successful node can still produce a broken conversation

An appointment parser can recognise a request correctly while qualification also tries to consume the same message. A reply can look fine while persistence leaves conflicting state. A proposed slot can be valid while an incompatible reminder remains pending.

Those are not node failures. They are failures in the relationship between routing, state, persistence, and scheduled work.

Test the sequence

The test framework I want is built around message sequences rather than isolated nodes. A case describes an initial lead state, runtime fixture, incoming messages, expected branch, expected state changes, expected reminder behaviour, and expected appointment outcome.

given: appointment selection is active
when:  a new requested time arrives
then:  appointment owns the turn
       qualification does not capture it
       no qualification reminder is scheduled

Useful variations include clean answers, slang, misspellings, incomplete messages, negative replies, unexpected questions, and appointment requests made during qualification.

What I learned

Testing an AI-assisted workflow cannot stop when a model returns a reasonable sentence. The surrounding state machine and irreversible actions are equally important.

Every real failure should become a regression case. That is how a vague lesson becomes a durable safety check.

What should happen next

I want a repeatable CI-style run that loads a known runtime fixture, executes the suite, and produces a concise report before workflow deployment. The goal is not to prove the system is perfect; it is to make regressions visible before they reach a customer.