Runs and Debugging
When a build fails, do not guess from the final answer alone. Inspect the run as a sequence:
- inputs
- retrieval behavior
- tool calls
- output
- persisted trace or event surface
What to verify on every run
- the run actually started with the input you expected
- retrieval inputs were attached when needed
- the intended tool was visible and callable
- the final output reflects the observed intermediate behavior
Fast triage order
- bad answer with no tool call: check tool availability and prompt shape
- bad answer with no useful context: check retrieval setup and corpus scope
- incomplete run visibility: check trace and event surfaces before changing the graph
Docs-driven debugging rule
If a tutorial step cannot be debugged from visible product surfaces, that is not only a product problem. It is also a docs gap and should be logged.