Basics

The message box

The message box is where you talk to a session's agent: + to attach, your message, and Send. Under it are where the session runs, what it runs with (the model, the effort and the mode) and how much of its plan and context it has used.

Sending

Type and press ↩ to send; ⇧↩ starts a new line. The square button at the right sends too (it lights up once there's something to send), and turns into Stop while the agent works.

Mentioning files

Type @ at the start of the message or after a space, and a list of the project's files opens, best matches first. ↑ and ↓ move, ↩ or Tab picks, Esc closes the list. The file goes into the message as @ and its path, quoted when the path has spaces. Claude Code reads a mentioned file itself.

The agent's commands

A / at the start of a message lists the agent's own commands as you type: Claude Code's, and Pi's (its commands, skills, prompt templates and extensions' commands). Codex has none here. ↑ and ↓ highlight one, Tab fills it in, ↩ runs it, Esc hides the list. In ⌘P, type / for the same commands.

In Claude Code, /model and a model's name switches the model, as picking it does. Magnum runs Pi's own commands, such as /model, /compact and /tree, as Pi's terminal does.

Images and files

Attach with the + button, by dropping files on the session, or by pasting an image or copied files (⌘V). Each shows as a chip above the box; ✕ takes it off. A message takes up to 5 attachments:

• Images (PNG, JPEG, GIF or WebP, up to 3.75 MB) go as images. Sent ones show in the conversation; click one for full size.
• Other files go as text, up to 200 KB.

Sending while the agent works

You can send while the agent works: press ↩, or the second round button that shows beside Stop once you type. The message waits above the box until the agent takes it.

• Claude Code reads it after the current step, or when the turn ends. Its ✕ (Don't send) takes it back, into the message box when that's empty.
• Codex reads it after its current step. It can't be taken back.
• Pi reads it after its current step. ⌥↩ sends a follow-up instead, for when Pi is done. ✕ gives all of Pi's waiting messages back to the message box.

Stopping the turn (Esc or ⌘.) leaves Claude Code's queued messages to run, and gives Pi's back.

Pi's shell commands

In a Pi session, a message starting with ! runs the rest as a shell command, and its output goes into Pi's context. With !!, the output stays out of it. Stop ends a command still running.

Under the message box

• The environment: where the session runs, named This Mac (This computer on Linux), a machine's name or a worktree's. Its menu lists where else it can run (Environments).
• The account, once you've added one in Settings → Accounts: the Claude account the project's sessions run on. Its menu picks another (Agents and models).
• The model: Magnum Auto (Magnum Router), then the agent's newest model of each family. Older models are under Older, Pi's and OpenCode's other providers each under its own name, the other agents under Agent (Agents and models), and Claude Code's advisor under Advisor.
• The effort (Thinking for Pi): Auto, which lets Magnum Router pick it, then the levels the model offers. The running agent takes a new effort at once, without starting again.
• The mode: how much the agent does without asking. ⇧Tab steps through Manual, Accept edits and Plan in Claude Code, and the first three of Codex's and Pi's. Every agent's modes are in Sessions.
• Remote Control, for Claude Code on your Anthropic sign-in: the phone button lets you follow and answer the session from claude.ai or the Claude app. Click it again for its link, or to turn it off.
• The meter (below).
• Detach, which turns tiles on with the session (Sessions).

Usage

The meter shows how full the agent's context window is (a ring, its colour changing from 80%), and a bar each for how much of your plan's 5-hour (5h) and weekly (Week) limits is used. Without plan limits, it shows what the session cost instead. Click it for Usage:

• Plan: each limit, how much is used and when it resets. Claude Code on your Anthropic account has 5-hour and weekly ones; Codex its usage limits; Pi its model's provider's.
• Context: a bar split by what fills the window, its size, and when Claude Code compacts the conversation. Each kind of content has a row; deferred tools show "—", as they aren't in the window until used. MCP tools (by server), custom agents and skills open to their items.
• This session: what it cost. On a plan, that's what its turns would cost at API prices; Pi shows dollars, dollars of extra usage, or what its plan includes.

Claude Code on another provider shows no cost: Claude Code prices every turn at Claude's rates, which hold only on Anthropic's own account.