Driving tty7 from the command line

tty7 is a thin, non-interactive client of the tty7 server. Every verb returns and exits; --json makes the output machine-readable. The GUI never has to be running — the server is what owns the panes.

First: where are you?

tty7 doctor

One table, and it answers everything you need before doing anything else: whether a server is reachable, whether the dialect matches, whether each agent's status hooks are installed, and whether TTY7_CONFIG_DIR / TTY7_WS / TTY7_PANE are set — i.e. whether you are running inside a tty7 pane.

Being inside a pane matters for two reasons: the address-taking verbs (split, send, capture, procs, wait, pane close) default to $TTY7_PANE, and run --keep files its pane into $TTY7_WS. Outside a tty7 shell you must name a target explicitly, and the error will say so rather than guessing.

The hooks row matters if you intend to delegate to another agent: without them an agent reports no status, so tty7 wait on it will only ever time out.

If tty7 doctor says the server is unreachable, stop and tell the user — do not run tty7 server start on your own initiative. Starting a server they didn't ask for changes what their GUI attaches to.

What are you here to do?

Four jobs, four shapes:

1. Run something that shouldn't block you or needs a real TTY — a dev server, a long test run, a TUI. Running a command.
2. Talk to something stateful over time — a REPL, ssh, a debugger. Same primitives: send, read, repeat.
3. Look at what this machine is doing — other panes, other agents, ports. Looking around, strictly read-only.
4. Hand work to another coding agent — one worker or a fan-out of several. Read references/delegation.md first; the short version is below.

The Bash tool remains right for anything that starts, does its job, and exits without needing a terminal or an audience. A pane earns its keep when the process outlives your turn, needs a real PTY, or should be visible to the user in their tty7 window — that last one is often the whole point.

Addresses

Re-resolve @N right before you use it; never cache one across a step that creates or removes a tab. Pane ids, tab ids and workspace ids are safe to remember — so when you create a tab and mean to address it again later, keep the id tty7 tab new --json hands back rather than counting @N a second time.

The sigils are optional wherever an address is expected: %42 and 42 are the same pane, @7 and 7 the same tab. Ids copied out of --json paste straight back in.

Omitting the address inside a tty7 shell means "this pane" / "this workspace". An explicit address always wins over the environment.

Running a command: two shapes

Blocking, with a real exit code

tty7 run -- cargo test          # streams to your stdout, exits with cargo's code
tty7 run --cwd /path -- make
tty7 run --keep -- cargo build  # leaves the pane as a new tab afterwards

The command's output streams to your stdout as it happens, and tty7 exits with the command's own exit code. This is the closest thing to a Bash call — the difference is the PTY and the fact that the user can see it.

Three things to know. --keep needs a workspace, so it only works inside a tty7 shell or with --ws <workspace>. With --json, the streamed output comes first and the JSON object last — the combined stream is not parseable as JSON, so read the last line. And the pane is 120 columns wide with no way to change it, so output that assumes a wider terminal wraps.

Non-blocking: a pane you talk to over time

This is the one that makes tty7 worth reaching for. Get a pane, send it work, come back later.

PANE=$(tty7 split --v)                  # or --h; splits $TTY7_PANE, prints "%83"
tty7 send "$PANE" 'npm run dev' --enter

split prints the new pane's address on stdout, which is what you capture into a variable. Without an axis it is a usage error — --v stacks the new pane below, --h puts it to the right.

Splitting $TTY7_PANE changes the user's visible layout, which is usually the point: they can watch the dev server you started. Say that you did it, and close the pane when you're done with it.

If you are not inside a tty7 pane there is nothing to split, so make your own place to work first. tty7 new --json /path/to/repo hands you both ids at once — don't go digging through ws tree for the pane:

read -r WS PANE < <(tty7 new --json /path/to/repo \
  | python3 -c 'import json,sys; d=json.load(sys.stdin); print(d["id"], "%%%d" % d["pane"])')

send types text into the pane exactly as a keyboard would; --enter appends the carriage return, or presses Enter on its own when you give it no text (tty7 send "$PANE" --enter runs what is already typed there). It does not wait and it does not tell you what happened — reading is a separate step, and waiting is tty7 wait.

For keystrokes rather than characters — Ctrl-C, Escape, the arrow keys — use --key: it takes enter escape tab backtab space backspace delete up down right left home end pageup pagedown, plus C-<char> for Ctrl and M-<char> for Alt. Repeat it for a sequence; text and keys compose, text first. Typing ^C as text does nothing — it arrives as two characters; --key C-c is the real interrupt.

A brand-new pane can swallow the Enter. A shell still working through its startup files — a prompt framework, fastfetch, anything that paints on login — takes the text you send but loses the carriage return that follows it, and the command just sits on the prompt line unexecuted. Nothing reports this: the send succeeded, and the pane looks like a worker that has not got going yet. So after sending the first command into a pane you just created, read the screen back and check it actually left the prompt:

tty7 capture "$PANE" --plain | tail -3   # command still sitting on the prompt?
tty7 send "$PANE" --enter                # then give it the Enter it lost

Cheaper than diagnosing it later, and only the first send into a fresh pane needs the check.

Reading a pane

If you want the screen, use --plain

tty7 capture %83 --plain

capture hands back what the daemon stored — the pane's bytes, escapes and all — and --plain replays them through a terminal grid and prints the resulting text instead. Not a stripper: colour and cursor escapes are gone, but also a line the shell wrapped at the pane's width comes back as one line, a progress bar that rewrote itself with \r reads as its final value, and a TUI's screen lands where it was drawn. Use it whenever a human would want to read the output.

Two details about what you get back either way: capture returns a snapshot, not a stream — call it again for a newer one. And by default it prints the newest scrollback segment (the ring splits on resize); --scrollback prints the whole ring, which for a pane that was never resized is the same thing.

If you want the result, redirect to a file

--plain gives you the screen, and a screen is a rectangle: whatever scrolled past the top of a long build log is gone, and the exit code was never on screen at all. So when what you want is the answer rather than the view, have the shell write it somewhere clean:

tty7 send "$PANE" 'cargo test > /tmp/t.log 2>&1; echo $? > /tmp/t.rc' --enter
# ...wait for it to finish (below), then:
cat /tmp/t.rc /tmp/t.log

Complete output, a real exit code, no terminal in the middle.

Knowing when a command has finished

Don't poll the screen and don't write your own loop — block on it:

tty7 wait "$PANE" --until free --changed --timeout 900

free means the foreground command has exited and the pane is back to its bare shell. --changed adds "and something actually ran while I watched", which is what you want on the line right after a send: without it, a command that has not started yet leaves the pane looking finished.

The whole shape, end to end:

tty7 send "$PANE" 'cargo test > /tmp/t.log 2>&1; echo $? > /tmp/t.rc' --enter
tty7 wait "$PANE" --until free --changed --timeout 900
cat /tmp/t.rc /tmp/t.log
tty7 pane close "$PANE"

Exit codes are built for this: 0 means a state you asked for was reached, 124 means the timeout ran out (the timeout(1) convention, so "not yet" is distinguishable from "broken"), 1 means the pane died first.

One trap in --changed: a command that finishes inside a single poll (500ms by default) is never seen running, so the wait keeps going until it times out. For something that quick, --interval 100, or drop --changed and read the .rc file. The timeout message says so when it happens.

If you want the process tree itself — "what is running in there", "which port is this pane serving" — that is tty7 procs %83: indented by depth, * on the foreground process, then the ports those processes are listening on. It is not the way to check on a coding agent, though: procs reports nothing running in this pane for a pane with a busy agent in it, so reading it as "the worker died" is wrong. Ask tty7 agents about those.

Handing work to another agent

A pane can hold another coding agent, and every primitive above works on it — plus one that only agents have: status hooks report working / waiting / done, so tty7 wait can block on the agent rather than its process tree.

Delegation has a playbook — references/delegation.md. Read it before you spawn a worker. It covers the whole arc: giving the worker its own git worktree and workspace, handing the task over with the delivery contract in the prompt, proving the command actually started, babysitting the states, collecting the result out of git, fanning out several workers, and cleaning up.

Four rules from it survive even if you read nothing else:

• Interactive mode, never -p. claude -p draws no TUI: the pane stays blank, capture --plain returns nothing, the user watches an empty rectangle, and the session dies after one turn so you cannot follow up. Hand the task as an argument to the interactive command instead.
• A worker that writes files gets its own git worktree. Two agents in one checkout trample each other and the user's working tree.
• Collect results from git, not from the screen. Tell the worker to commit; read the diff. A screen is a rectangle and the top of it is gone.
• After the first send into a new pane, confirm the command left the prompt — the swallowed-Enter check above.

Looking around

tty7 ls                    # every workspace: tabs, panes, who's attached
tty7 ws tree api           # one workspace as a tree — tabs, splits, panes, cwds
tty7 pane ls               # panes with their workspace, tab, cwd, live flag
tty7 pane ls --all         # + orphans: panes the server runs that no workspace holds
tty7 agents                # every coding agent on the machine and its status
tty7 status                # server pid, uptime, pane count, build, socket
tty7 machine ls            # this machine plus any linked remotes
tty7 events                # stream server events, one per line, until interrupted

tty7 agents is worth knowing about: it reports each pane running a recognised coding agent as idle / working / waiting / done, with the agent's own message beside it. If you are one of them, you are in that list too. It also prints a diagnostic — diagnostics in the JSON — when it can see an agent running whose status hooks are missing or outdated, which is the explanation for any agent that appears frozen.

Add --json to any of these to parse instead of eyeball. -q suppresses output on success but never suppresses errors.

Don't break the user's session

The panes on this machine are the user's real work, and some of them are other coding agents mid-task. Treat anything you did not create as read-only:

• Never send into a pane you didn't open. Keystrokes into another agent's pane, or into a shell the user is typing in, land in the middle of whatever is happening there. Check tty7 agents before you touch a pane. This goes double for --key: a stray C-c kills somebody's work.
• Never pane close / tab close / ws rm something you didn't create.
• Never pane close --orphans. It closes every abandoned pane on the machine, and an abandoned pane can still be running a real command. It is the user's broom; point them at it, don't swing it.
• Never server stop or server restart. Every pane on the machine dies with the server, including yours. If the server genuinely seems wedged, say so and let the user decide.
• Clean up what you did create. tty7 pane close %83 when you're done with a scratch pane; it takes several ids at once. ws rm hangs up the panes the workspace held, so removing a scratch workspace is enough on its own. What does leak is an interrupted tty7 run — that pane keeps running with nothing referencing it, and shows up under tty7 pane ls --all.

Remote machines

-m <machine> routes any command over a link the local server already holds:

tty7 -m devbox ls
tty7 -m devbox run -- cargo test

The name matches the full link key (me@devbox:22) or just the host. The CLI will not dial a fresh connection — if the link is down, or it's a jump/proxy chain, it says so and you should hand that back to the user, who can connect it from the GUI.

Not wired up yet

ws stop, machine connect and machine disconnect exit with a message saying they're not implemented. Don't build a plan around them.

References

• references/delegation.md — the delegation playbook: worktrees, handover, babysitting, collection, fan-out, cleanup. Read it whenever another agent is about to do the work.
• references/commands.md — every verb, subcommand and flag in one table, plus the JSON shape each one emits. Read it when you need a verb that isn't above, or when you're about to parse --json output and want to know the field names.