Chats and context

One window holds as many conversations as you like, and the sidebar is the list of them. What is worth understanding is not the list — it is what each conversation can reach, which is decided by how you started it.

New chat, or new chat group

The New chat button starts a conversation that shares this one’s browser tabs. The caret beside it offers two other things: a new chat group, which gets its own new tabs, and a chat with one of your agents.

That is the whole of it, and it explains the single most common puzzle in the application: the agent says it cannot see a site you are certain is open, because the tab is open in a different group.

Chats in a group are shown nested in the sidebar, with children indented under their parent. A chat can also be given a child of its own from the row’s own menu, which is the tidy way to branch a long piece of work without losing the thread you were on.

Finding one again

Two controls sit above the list: a search control and a filter control. Both narrow what the list shows — pinned chats, running ones, ones with an agent — and neither asks the application for anything: they are a view over the list the window already has.

A row’s own menu carries the rest: pin it to the top, rename it, start a child chat, open its settings, or delete it.

Activity

Activity, pinned above the chats, is what has arrived in the conversations and threads you follow. It is not itself a chat — selecting it shows a panel instead of the transcript, and the conversation you were in goes on running behind it.

Whether a chat reaches Activity is the follow control in that chat’s header.

Which model a chat runs on

The model control in the chat header reads Default until you change it, and then names what you chose. It is permanent rather than tucked into a menu because, once several chats are running at once, the task and the model are what tell them apart at a glance.

The default itself is set in the AI Providers section of Preferences — see Getting started.

What the model is actually told

Before you type anything, the assistant has already been given a system prompt and a catalogue of the tools it may call. You can see both: the chat header’s ⋮ menu, then System prompt.

This is the place to look when the assistant behaves in a way no instruction of yours explains, and the place to look when you are wondering why a conversation is costing what it is. A tool catalogue is sent on every turn, and narrowing an agent’s Tools tab is the lever that shortens it.

Long conversations

A conversation cannot grow without limit — there is a ceiling on how much the model can be sent at once. A meter above the message box shows how full this conversation is, as a percentage and as a count against the ceiling, with a Compact button beside it.

Compacting folds the older part of what is sent to the model into a summary. The whole conversation stays on screen: nothing is deleted, and you can still scroll back and read it. What changes is what the model is handed on the next turn. It happens on its own when the conversation reaches the ceiling, and the button is there for when you would rather do it early.

The practical consequence: after a compaction, detail from early in a long conversation may need repeating. If something from the first hour matters for the next one, say it again rather than assuming it is still in view.

Bringing things into a chat

The paperclip control below the transcript attaches things to the conversation — files, and the contents of a folder you want the agent to work in. Attachments are listed above the message box, and each says whether the agent may write to it as well as read it.

Attaching a folder is the difference between an assistant that can talk about your work and one that can do it, and it is worth knowing about before you paste a long file into a message.