Skip to main content

​
Check these first

Most problems fall into a handful of buckets:

FAQ

The questions that come up most.

Tools not firing

Usually the description, not the wiring.

Prompt behaviour

How the fields become one system prompt.

Errors

Status codes and what they mean.

​
Include these when you ask

The difference between a same-day answer and three rounds of questions:
​
header
Every API response carries one. It identifies the exact request in our logs — by far the most useful single thing you can send.
​
string
For anything that happened during a conversation. It ties to the transcript, duration, and end reason.
​
string
So the configuration can be read rather than guessed at.
​
text
Especially the caller’s actual words if the agent behaved oddly — the transcript usually explains it.
Never send an API key or a credential value in a support request, a screenshot, or a shared log. If one has been exposed, revoke it in the dashboard and create a replacement — keys are independent, so nothing else breaks.

​
Narrowing it down yourself

Test the same agent in the browser and on the phone. If it behaves differently, the difference is in the call path — connection type, tools that do not exist on phone calls, or audio quality — rather than the prompt.
Invoke the tool directly from the dashboard’s test panel with sample parameters. If it works there but not in a call, the model is not deciding to call it — that is a description problem.
Check your own latency and status codes first. An endpoint that exceeds the tool’s timeout produces an agent that goes quiet mid-sentence, which looks like a platform fault and is not.
Agent version history shows what changed and when, with your change notes. Compare against the last version that behaved.

​
Status

Platform status and incident history: omnia-voice.com