Testing and logs

Chat logs

The conversation as it happened, session by session, what the user sent, what the agent replied. Read them to answer questions about behaviour: did the agent understand, did it escalate when it should have, was the answer any good.

Compiled logs

What each node was actually given after its template expressions were resolved. This is the debugging tool, because templates are compiled at run time. The value a node receives depends on what the previous node produced on that particular run.

Read the compiled log when:

• An AI node gives a confident answer to something nobody asked, its input probably resolved to empty.
• An API call fails. The URL or body may not be what you think.
• A node did not run at all. it was probably on a Decision branch that was not taken.
• The workflow throws a compilation exception: the log names the expression that could not be compiled.

A debugging order that works

1. Did the trigger fire? No session at all means the channel, schedule or endpoint is the problem, not the workflow.
2. Did every node run? A missing node usually means an untaken Decision branch.
3. What did each node receive? Compiled log. Empty values point at a key path that does not match.
4. What did the AI node decide? Check whether it called the tools you expected, if not, the tool descriptions need work.
5. Did the response go to the right place? A mismatched trigger and response pair looks healthy in the logs and silent to the customer.

Testing while you build

Add a UI trigger even to a workflow destined for WhatsApp. It gives you a chat interface with no channel setup, so you are testing your logic rather than your integration. Remove it or leave it. an extra trigger costs nothing until it fires.

Common errors and what they mean