Code tab

The Code tab is a coding agent that works in one project folder: it reads and edits files and runs commands there. On Windows it works on projects inside WSL; see the code-windows page.

Start a session

Click New session (or + next to the open sessions) and pick the project folder. Open in Code in a Shell tab's assistant starts one in that terminal's folder. The folder is fixed for the session's life: the agent can't read or write outside it.

Type what you want and press Enter. After the first reply the model names the session; a name you gave it is never replaced. Sessions are saved after every reply.

Find, fork and move sessions

The list on the left holds every saved session. To fork one or move it to its own window, see the code-sessions page.

Read what the agent did

• Edits show as a diff, red and green, with line numbers; past 40 lines Show all … lines expands it. Unlike in Chat, rewriting a file doesn't ask first: the diff shows what changed. A second whole rewrite in one turn needs a command (a test, say) run in between.
• Commands show as a card: command, exit code, time taken and output (the last 20 lines until Show all). Copy copies the command; Open in Shell types it into a new Shell tab at the folder without running it.
• Searches and file reads show as a short step list.
• Reasoning sits above its step, collapsed (live as Thinking…); what the agent says before a step sits above it in small grey text.
• A call still being written shows as a row: Writing, Editing or Preparing command….
• File names inside the folder, in diffs, reads, search results and answers (src/app.ts:42), open the file in an editor window, at its top.

Commands that need a terminal

Commands that ask for a password (sudo) or need a terminal can't run here, so the agent hands them to you: a new Shell tab opens at the folder with the command typed in, not run. Check or change it, then press Enter.

The conversation shows Waiting for you in Shell N — press Enter there, with Go to shell and Cancel. When it finishes, the agent gets the exit code, the output and the command as you ran it. Cancel stops the wait (the agent carries on without the result); Stop ends the turn; either way the Shell tab stays. Closing the Shell tab first gives no result. If the shell doesn't report finished commands (an old fish), the agent asks you.

Edit files

Files open in an editor window, one per folder, a tab per file. ⤢ on a tab moves the file to a window of its own. Ctrl / ⌘ + S saves; Ctrl / ⌘ + W closes the tab. The agent can open files there too but doesn't see your edits.

When something else changes a file, the editor reloads it and keeps your place. With unsaved edits it shows Changed on disk (Reload / Keep mine), and saving over a change you haven't seen asks Overwrite or Reload first. A deleted file shows Deleted on disk; saving recreates it. Live reload unavailable means outside changes won't show. Closing a window with unsaved edits asks first; closing the main window closes the editors too.

Steer or stop the agent

While the agent works you can keep typing: Enter queues your message (Queued), and the agent reads it at its next step (Delivered). × drops a queued message. Stop, or Esc, ends the turn; anything unread comes back into the input box. / opens the commands and skills list, as in Chat.

A dot on a session's tab shows it working (filled) or waiting for another turn (hollow): only one turn uses the model at a time. Stop on a waiting session gives your message back.

Pick the model and effort

The header above the conversation holds:

• the folder — click it to open it in your file manager;
• the git branch, with ● for uncommitted changes. Click it to switch branch, or to start one with Branch from current… or Branch from main… (shown when another branch is checked out). Switching waits until the agent is idle and changes are committed or stashed; it warns when another open session uses the repository;
• the model — click it to pick Settings model (follows Settings → Inference), Remote server (one saved in Settings → Inference: Probe, then pick a model) or OpenRouter (cloud) (Load models, then pick one), then Save. The local model is only reachable through Settings model, and picking one here never starts it;
• Effort — how hard the model thinks, where the model offers levels;
• the AGENTS.md badge when the repo's instructions are in use (see the skills page), and how full the context is.

Model and effort apply from the next message.

Background processes

Servers, watchers and long builds run in the background. Running: N in the header lists them: Output shows what each printed, Stop ends it. Closing the session's tab (it asks first) or quitting the app stops them.

Command approval and limits

Safe-looking commands run on their own. A risky one (sudo, rm -rf, a pipe to a shell) opens Run this command? first; one that reaches outside the folder always asks. Prompts from several sessions wait in line, each naming its session. Allow for this session lasts until its tab closes, even across windows. Settings → Code holds the command time limit, the step limit per task, the background log size and Auto-approve commands. See the settings page.

What it doesn't do yet

• Each command runs on its own, so cd and environment changes don't carry over.
• No undo or checkpoints: use git.
• Small local models make mistakes and get stuck; see the models page.