Agents and sessions

Claude Code and Codex, models, permission modes, approvals and attachments.

Sessions

A session is one conversation with an agent in a workspace. A workspace can have several, as tabs above the chat: start another one with ⇧⌘N, for example to let a second agent review the first one's work. Each session keeps its own history, and picks up where it left off after a restart.

Claude Code and Codex

Both agents work the same way in Compagnion: the same chat, the same tools, the same approvals. Choose the agent and model per session, and set the defaults under Settings → Default models. Compagnion runs the claude and codex you already have, with your own subscription or API key; you can add extra subscriptions and keys under Settings → Account.

Permission modes

How much the agent may do without asking:

Mode What the agent may do
Ask Asks before running commands or editing files
Accept edits Edits files freely, asks before running commands
Auto Decides by itself what is safe, and asks for the rest
Plan Only reads and plans; changes nothing
Don't ask Everything, without asking

Pick a mode in Create or in the composer of a running session; the default for new sessions is under Settings → General.

Approvals

When the agent wants to do something its mode doesn't allow, the request appears in the chat with the command or the change, and the workspace's mark turns around in the sidebar. Allow lets it go ahead, Deny stops it, and Allow all (don't ask again) allows it and switches the session to Don't ask.

Talking while it works

Type a follow-up any time. While the agent is busy, the message waits and goes in as soon as the turn ends. Stop interrupts the turn.

Attachments

Paste images and files with ⌘V, drag them onto the chat, or pick them with the paperclip. They appear as chips where the cursor is, and the agent gets the file in the workspace. Pasted text over 5,000 characters becomes an attachment too; turn that off under Settings → General.

Secrets

Paste a token, a password or a private key and it becomes a lock chip, recognized on your Mac without anything going over the network. The secret is stored in the workspace where only you can read it, and the agent gets just the file to use in a command, so it doesn't end up in the chat, the history or the prompt by accident. How secrets work.

Context

A small ring in the composer fills up with the model's context window. It shows from 70%, orange, and red from 90%; hover it to see the number of tokens.