Shell commands

Type a shell command the way you'd type it in a terminal, and it runs in the tab's shell:

ls -la
git status
npm test

Output streams into the transcript line by line as it's produced, with ANSI colors and styling intact — a test suite's colored pass/fail summary looks the way it should. file.ts:42-style paths in the output are clickable and open the file in an editor tab at that line.

Plain commands like these are recognized as shell input automatically. When a line could be read more than one way — a shell command, a SQL query, a prompt for the agent — the app asks instead of guessing, floating a chooser above the command bar; pick a route with ↑/↓ and Return, or Escape to cancel. To skip recognition entirely, prefix the line with shell :

shell find . -name "*.ts"

The prefix is the deterministic escape hatch — whatever follows it goes straight to the shell.

One shell per tab, and it persists

Each tab has its own shell process that lives as long as the tab does. State accumulates the way it would in a terminal: cd somewhere and later commands in that tab run there; exported variables stick around. The working directory is also remembered per agent, so after janus --relaunch a restored tab's shell starts where it left off. If the shell process dies unexpectedly, a fresh one is spawned on your next command.

Closing a tab kills its shell; quitting the app kills them all.

Interactive programs take over the tab

Full-screen and interactive programs — htop, vim, less, man, python and other REPLs — can't run through the ordinary transcript. When you run one, the tab switches into a full-tab terminal: the transcript and command bar disappear and the program gets the whole tab, with every keystroke — including Ctrl+C, Ctrl+D, and Ctrl+Z — forwarded to it. Only Shift+←/Shift+→ still switch tabs, and you can keep several tabs' interactive programs running at once; each keeps its screen state while you're elsewhere.

Shift+Enter inserts a line continuation rather than submitting, which matters for programs (AI harnesses in particular) that accept multi-line input.

When the program exits, the transcript comes back exactly as it was — nothing about the takeover is logged.