Requests

Requests is the screen you open when cost looks odd, an error shows up, or someone reports a strange response. Every AI call that passed the gateway is kept here, so you narrow the list with filters and then open one call to find the cause. For the full process of narrowing a cost spike down to this screen, see Respond to a cost spike.

The list and the detail

Area	What you see
List	One call per row. Key fields include time, provider, model, user, tokens, cost, latency, and status
Detail: conversation	The actual system and user messages and the response body. Failed calls show the raw error message
Detail: breakdown	Prompt versus completion tokens, input versus output cost
Detail: events	What the user who received this response actually did

Failed rows get a red tint with the error message shown alongside. The token and cost breakdown makes anomalies like an abnormally large prompt stand out immediately. If the detail says “Logging is disabled or no messages were recorded”, no message bodies were kept for that call. Tokens, cost, and latency are still recorded all the same.

Tools for narrowing the cause

Tool	Use
Filters	Conditions such as provider, model, status (success or error), and range. If you arrived by clicking Overview’s error rate number, the list is already filtered to errors
Sorting	The time and latency columns. Sort by latency to find the slowest calls first
Open in Testing	Carries this call’s input straight into Testing. A problematic production call becomes a test case on the spot