Pitfalls every QA hits the first time — and how to avoid them.
| Mistake | Fix |
|---|---|
I used says for an API call | Use calls + target. says is for talking. |
sequence fails but the steps are there | sequence checks relative order, not adjacency. If A is before B, it passes. |
| I want an IF branch in the conversation | v0.1 has no branching. Two paths = two separate Sessions. |
llm_judge passes sometimes and fails others | Vague criteria ("be friendly") produce inconsistent results. Be specific. |
| I don't know which actor to use | user = the tester. assistant = the bot. tool = external API/system. human = real person. |
Next: Available tools →