The Chats section is the account’s activity in one list: conversations from every assistant, wherever they came from — the console, an embedded widget, the API, an inbound email or a trigger.
Finding a conversation
Filters narrow the list along the axes that matter in practice:
There is also full-text search inside a conversation’s messages, for when you remember what was said but not where.
Reading one
A conversation shows more than the transcript:
- Tool calls, expandable to their arguments and results — the fastest way to see why an answer was wrong.
- Recalled memories, when the assistant used memory: what it brought in and from where.
- Cost and tokens for the conversation.
- Feedback left on individual messages, thumbs and comments.
When an answer is wrong, the tool calls usually explain it before the prompt does. A tool that returned an error, or was never called at all, is a different fix from a prompt that asked for the wrong thing.
Conversations that were blocked
A conversation stopped by a usage limit is marked as such, with the rule that blocked it, the current value, the ceiling and when the window resets. The mark clears itself as soon as a later message goes through.
Agent runs
Agent threads have their own equivalent view, with the run’s state, its tasks, approvals and pauses. See Logs.
Through the API
Everything here is available programmatically — list conversations, read one, search inside it and submit feedback — which is how you build your own review workflow or export the lot. 
