Chat controls & modes

Every chat — and every individual request — can be tuned. The controls live in the chat's header menu and the composer bar. Most have a Default option that uses your preferences from Settings → Behavior; change one here and it sticks for this chat, so you can have a fast, autonomous scratch chat and a careful, approval-gated work chat side by side.

The per-chat pickers

Model

Pick any model your admin has enabled, grouped by provider, with a search box for long lists. Or choose auto — the assistant ranks the enabled models by capability (frontier / coding / reasoning / fast / vision / cheap) and picks the best fit within this chat's governance limits. If a model errors or is rate-limited, a fallback chain quietly cascades to the next compatible model, preferring a different provider, so a single overloaded account can't sink the turn.

Reasoning

On models that support it, toggle extended reasoning on or off and set the effort: Off, Low, Medium or High. More effort means more "thinking" before the answer — better on hard problems, slower and more tokens on easy ones.

Mode (collaboration style)

How much the agent plans versus just acts:

Mode	Behaviour
Default	Balanced — plans or acts as the task needs
Plan	Plans and confirms the approach with you before acting
Execute	Acts directly, with minimal back-and-forth
Pair	Small steps, thinks aloud, and checks in often

Note

This "mode" (collaboration style) is separate from the chat mode (standard / coding / custom) covered below.

Security mode

Controls how tool calls are gated in this chat:

Mode	Behaviour
Autonomous	Tools run without asking — fastest, most hands-off
Approve each	Every tool call waits for your approval, inline in the chat
Judge (LLM)	A safety model reviews each call before it runs

Approvals appear inline in the conversation; Always allow whitelists a specific command on a specific device so you're not asked again for it. Your default is set under Settings → Profile, and an admin can layer an instance-wide command policy on top (see Security & privacy).

Memory access

What the assistant may read from its long-term memory in this chat:

Access	Meaning
Full	Full access to everything remembered
None	No access — a private chat that nothing is learned from
Restricted	Only the areas and sources you tick

Restricted memory lets you pick areas — People & contacts, Work & projects, Places & devices, Notes & topics — and sources — Preferences & rules, Things you said, Agent inferences, Integration observations. An empty area or source means the agent sees nothing there, which is handy for a chat that should only draw on, say, your stated preferences and nothing inferred.

Data classification

Choose Standard (any configured model may be used) or Confidential (only local models, so the data stays on-prem). Confidential is fail-closed: there is no path to a provider that isn't cleared for it. Your organization may set a minimum that you can only make stricter, never looser.

Tools, integrations and devices

The composer's Integrations picker controls which tools the agent may use this turn — built-in tools, web tools, your configured integrations, and any online devices. A separate preview, "Available in this chat", lists exactly what's active, including anything hidden by governance so you understand why a tool might be missing.

Chat modes

Each chat runs in a chat mode that shapes its tools, workspace and side panel.

• Standard — the normal assistant.
• Coding — turns the chat into a workspace over a device's files, with an editor and terminal. See Coding & Cloud Tasks.
• Custom modes — your own modes, each with a name, an icon and an attached dashboard whose cards show in the chat's side panel. Create, edit and delete them from the mode switcher in the chat header. A "home" mode with your lights/scenes, or a "trip" mode with travel cards, puts the right controls beside the conversation.

When you start a fresh chat, a card grid lets you pick the mode up front. (The main chat is always standard.)

Slash commands

Type / at the start of the composer to open the command palette. Commands are either actions (they do something) or prompts (they send templated text). The palette is mode-aware — coding commands only show in coding chats.

Command	What it does
/new	Start a new chat
/rename <title>	Rename the current chat
/btw <question>	A quick research side question, answered without entering the history
/goal <objective>	Pursue a goal autonomously across turns (see Agents)
/main [message]	Jump to the main chat — or, with a message, send it there without leaving
/summarize	Summarize this conversation
/proofread <text>	Fix spelling and grammar

In coding chats you also get:

Command	What it does
/terminal	Show or hide the terminal panel
/review	Review the current workspace changes (git diff)
/explain <file>	Read and explain a file
/fix <problem>	Find and fix something in the workspace
/tests	Run the tests and report back
/init	Generate an AGENTS.md for the repo

Your own commands

Add custom commands under Settings → Commands. Give it a name, a template, and optionally a description and a mode. Use $ARGUMENTS or {args} in the template for the text after the command — for example, "Read $ARGUMENTS and explain it." Then type /yourcommand … in any chat and the template is sent (or, with no argument, dropped into the composer for you to finish).