Chat settings and toggles

Where the controls live

There is no separate settings panel at the top of a chat. Everything sits in the composer, in two places:

• The settings panel(the sliders icon next to the message box) starts with Context. It holds the auto-compact switch, working-context target, and manual compact action. Tools and Autonomy follow.

• The composer toolbar holds the model selector and the Fast, Smart, Expert speed dial.

Model and speed

Pick the model in the composer toolbar, and set how hard it works with the Fast, Smart, Expert speed dial beside it. The speed dial is the one control for effort, so there is no separate reasoning-effort setting. See Choosing models.

Web search

The Search toggle lets the assistant look up current information and read specific pages from the web. Turn it off for conversations where you want answers from training data and your attached context only. See Web search.

Memory

The Memory toggle lets the agent read and write its own private memory: a persistent store that belongs to the agent, not to your workspace, so it can carry useful context from one chat to the next. See Agent memory.

Subagent

The Subagent toggle lets the agent spin up focused worker conversations to handle parts of a big request in parallel. The workers run as the same agent, not as some other agent in your workspace. See Subagent conversations.

Autonomy

The Autonomy level sets how freely the assistant acts on your behalf. It has four rungs, and the default is Auto:

• Read-only: the assistant can look but not touch. All tools that would change anything are removed from its reach.

• Ask: it asks you to confirm consequential actions before running them.

• Auto: it acts on its own but still pauses for actions that cross a boundary or cannot be undone.

• Full: it runs without pausing for confirmation.

See Autonomy and confirmations for how each level behaves and how confirmations appear.

Keep long conversations efficient

Very long chats can outgrow a model's context window. With Auto-compact on, idapt summarizes older messages as the working context fills. This limits repeated input costs while preserving the important state of the conversation.

Choose 64K, 128K, 256K, or Max. New chats inherit the agent's choice, then the 128K product default. Max uses as much of the active model's context as can be used safely. A smaller model can cap any choice to what it can serve.

A choice in the composer applies only to that chat. Use the reset icon to return to the agent or product default. The context indicator shows the effective target as its maximum while auto-compact is on. When you turn it off, the indicator uses the model's full context window.

You can also compact on demand at any time with the Compact conversation action in the settings panel, or the Compact button in the context indicator (the small bar chart button in the composer toolbar that shows how full the context is).

Set a spend limit

You can cap how much a single conversation spends from your Credit balance. Open the cost indicator(the dollar button in the composer toolbar), and the Usage Credits limit section sits at the bottom of the popover.

Type an amount or tap a preset, and it saves on its own with no Save button. A progress bar tracks spend against the limit, a Disable button removes it, and Reset spend window starts a fresh count without touching past history. The limit is enforced by idapt before any paid usage is reserved: it does not tell the assistant to budget or estimate costs.