Skip to content
Language

Observability

Reading a trace

A trace is the list of every tool call in a session, in the order they happened. It fills in live while your agent runs.

What's recorded per call

One entry in the trace. Arguments and response are stored in full, not summarised.
Order
A running number per session. The order is often the answer already: did it search first, or create straight away?
Tool
The name your agent picked — out of the list it was shown.
Arguments
In full, exactly as sent.
Response
In full, exactly as returned.
Latency
The measured duration of the call.
Status
Success or error.
Scenario
If a failure was forced: which one.

Live, then readable later

The viewer gets new entries the moment they're written. Nothing is polled and nothing is buffered.

A call that fails is written too, so a failed attempt doesn't disappear from the record. The one exception is schema validation: when it rejects the arguments, there is no call to record in the first place.

What your agent sends ends up in the trace

What a trace doesn't hold

Prompts, system messages, model responses, token usage: none of it travels over this connection. Mockd sees the tool side, not the model side.