Shell tab

The Shell tab is a real terminal with an assistant beside it. By default the assistant is Read-only: it reads and suggests. Switch it to Full access and it can also edit files and run commands in that terminal.

Open a terminal

It works on Linux, macOS and Windows. On Windows a picker in the tab lets you choose a PowerShell version or a WSL2 distro. Click + for another tab; each tab has its own terminal and its own assistant. A tab can be detached into its own window. Right-click the terminal for Copy, Paste and Restart shell.

Your $SHELL is used unless you set Settings → Shell → Shell binary. Command capture needs shell integration: bash and zsh get Haruspex's hooks, and fish 4 sends the needed markers itself. Other shells still work as terminals, but their commands are not captured, so copy the text you want into the assistant box instead. The badge in the sidebar header shows whether capture is working.

Tabs the Code tab opens

When the Code tab needs a command run in a terminal (a sudo, a password prompt), or you click Open in Shell on a command card, a new tab opens here at the Code session's folder with the command typed in. Nothing runs until you press Enter, and you can change the command first. When the agent asked for it, the Code session waits for the command to finish and reads its result; after that the tab is an ordinary terminal. See the code page.

Ask about what just happened

• F4, or the >_ button by the input box, sends your recent commands and their output to the assistant with no question.
• Anything you type is sent with your last 3 completed commands (command, output, exit code, folder). Change the number in Settings → Shell → Recent shell commands attached to each chat message; 0 sends only your question.
• The camera button captures a window and sends it.

Long output is cut in the middle before it is sent (Settings → Shell → Max output bytes per captured command, 8 KiB by default). By default the last 10 lines of your shell history file are also sent; turn that off in Settings → Shell → Include your shell history file in prompts.

Ctrl+Shift+A shows or hides the sidebar, and Ctrl+` moves focus between terminal and assistant.

Read-only (the default)

The lock in the sidebar header reads Read-only. The assistant can read config files and logs and suggest fixes, but it never runs anything. Each suggested command is a card with two buttons: Paste types it at your prompt without pressing Enter, and Run types it, presses Enter and sends the output back. Commands that match risky patterns, such as sudo, rm -rf, dd of=, a pipe to a shell or Remove-Item -Recurse -Force, get a red chip and ask again before they are typed. The check is a short pattern list: a command without a chip is not proven safe.

Full access: let the assistant edit files and run commands

Click the lock (Read-only) in the sidebar header to switch that tab to Full access; click it again to go back. Each tab has its own. New tabs start Read-only unless you turn on Settings → Shell → Start new shells with Full access. A tab moved to its own window keeps its setting.

With Full access the assistant runs commands in your live terminal, so they share your folder, environment and venv and show in your scrollback. It can edit and write files, search the project, drive a program waiting for input, and start servers or long builds in the background. Commands that look safe run on their own. A risky one opens Run this command? with Allow for this session, Allow once or Deny (the model is told it was denied). A command that reaches outside the project always asks. In a detached Shell window the question shows in that window.

Settings → Code → Auto-approve commands skips that prompt. It is off by default; only turn it on if you fully trust the model on this machine. Settings → Code also holds the command time limit (30 seconds by default) and the step limit per task (40); the Code tab uses them too.

Full access needs a capable model. Small local models make mistakes and get stuck; see the models page. Read the AI safety disclaimer in the README before using it.

Continue in the Code tab

Open in Code in the sidebar header starts a Code session in the folder the terminal is in now and switches to the Code tab (from a detached Shell window, the main window comes forward). The conversation here doesn't go with it. Use it for longer coding work: Code sessions are saved and can be reopened, forked and moved. See the code page.

The Shell assistant's conversation is not saved: it ends when you close the tab or quit, in both modes, and is not brought back when you return to a folder.

Stop runaway commands (memory limit)

Settings → Shell → Memory limit caps each command the assistant runs on its own, and each tab's terminal as a whole (your own commands too), at a share of your RAM: 50% by default, 0 turns it off. Over the limit, the system kills only the process using the memory; the shell and its scrollback stay. The assistant is told why, so it looks for the bug instead of re-running it. This needs Linux with a systemd user session; in WSL, lingering too. Tabs opened after a change use the new limit.

Repo instructions (AGENTS.md)

In a git repo with an AGENTS.md (or CLAUDE.md), the assistant reads it into every turn, in both modes. The first time, Haruspex asks whether to use the repo's instructions. While the file is in use, an AGENTS.md badge shows in the sidebar; click it to stop using them. With Full access, /init drafts an AGENTS.md for you. See the skills page for details.