# Smoke tests

Required tests:

1. simple chat without tools
2. task creates execution graph
3. weak model response success
4. weak model invalid → retry
5. weak invalid → strong fallback
6. external disabled by config → no external call
7. policy confirm → confirmation card
8. policy full_auto → execute immediately
9. MCP tool result stored
10. local worker command result stored
11. event stream emits task/node events
12. finalizer creates user result