marver

A TUI workspace for AI agent sessions — tmux orchestration, git worktree management, and repo control in one place.

Start a task, describe it, and marver opens a tmux session with an agent already working in its own worktree. Inspired by k9s, cmux, and Warp.

Status: early development. Every subsystem exists and the loop closes — describe a task, an agent runs it, review the diff, commit — but nothing has been used in anger yet. The published crate is still a name reservation.

marver                  open the interface, starting a daemon if none is running
marver status           report whether a daemon is running, and what it holds
marver restart          stop the running daemon and start this version
marver upgrade          install the newest published marver
marver cleanup          remove the worktrees left by finished tasks
marver daemon           run the scheduler and hook receiver in the foreground
marver scan [root]      list the git repos it can see
marver --version        print the version (-v, -V)
marver --help           print usage (-h)

Every command takes --data-dir for where the database, socket, and log live, --scan-root for the directory searched for repos, --cap for how many agents may run at once, and --harness for which agent to run. The interface and the daemon find each other through those paths, so a --data-dir given to one must be given to the other.

Just run marver. If no daemon is listening it starts one, detached, the way a tmux client starts the server it needs — and that daemon keeps running after you close the interface. It owns the queue: it creates a worktree per repo, opens a tmux session, starts the agent with its reporting pointed back at itself, and notifies you when a task finishes or gets stuck. Its output goes to daemon.log beside the database.

Run marver daemon yourself to watch it in the foreground or to put it under a supervisor. Starting a second one is refused rather than allowed to displace the first.

Keys

Three rules hold everywhere:

• esc leaves the screen you are on. On the task list there is nothing underneath, so leaving is quitting.
• ctrl-c and ctrl-q also leave, from anywhere, including inside a text field — where ctrl-c drops what is being typed rather than the screen behind it, because the field is what the key is being pressed at. The agent's terminal is the exception for ctrl-c: there it goes to the agent, which is the one place that reflex means something else. ctrl-q leaves even there.
• x discards. A task on the list, a diff in review. It is never c, which means commit.

No other chord does anything. A modified key is not a letter binding, so ctrl-x in the review screen is not x and does not type an x into a commit message either.

The footer of every screen shows its own keys; the whole set is below.

Task list

Key
j k · ↓ ↑	move
g G · Home End	first, last
↵	open the agent's terminal
y	what it was asked to do — any key closes it
v	review the diff
n	new task
p	pause it, or resume it
f	say its agent has finished, when nothing else will
t	its todos — tab there reaches the global ones
x	cancel it; press again once it has ended to archive it
u	put an archived task back on the list
s	cycle order — newest, oldest, by repo
a	cycle what is shown — open, all, archived
r	refresh now
R	the repos, and where each one stands
esc q	quit

New task

Key
tab	next field: repos → prompt → title
space	select a repo
j k	move through repos
↵	queue it — from the repo list or the title
ctrl-s	queue it from anywhere, including inside the prompt
esc	discard

You land on the repos, because that is the order the decisions happen in: where the work goes, then what to do, then — only if you want to say — what to call it.

The title is optional. Left blank, the task is named after the first line of the prompt, cut at a word; the field shows the name it is about to take, dimmed, so it is an offer rather than a surprise found later in the list. Typing replaces it. A prompt with nothing nameable in it — a pasted stack trace, a line of punctuation — leaves the task named after its repos, which at least says where it is.

↵ inserts a newline in the prompt rather than queueing. A prompt is what the agent is told to do and is routinely several paragraphs, so the only way to type a newline is worth more there than the shortcut; ctrl-s is how you submit without leaving the field.

Agent terminal

Every key goes to the agent, which is the point of the screen, so leaving takes a chord.

Key
ctrl-q	back to the list
ctrl-esc	back to the list, where the terminal can say it — see below
↵	sent to the agent, and returns a blocked or reviewed task to running

esc is not a way out here: it is how Claude is interrupted, and the agent needs it. ctrl-esc is the binding this screen would like to have, but a terminal can only report it under the kitty keyboard protocol, which marver does not enable — turning it on changes how every key is reported, and this screen forwards every key through a hand-built encoding table. In Terminal.app and iTerm2 it will not arrive at all. ctrl-q is the one that always works, in every terminal, because raw mode clears IXON and so nothing upstream takes it for flow control.

Todos

Key
a	add one — ↵ to keep it, esc to drop it
↵	use it: sends it to the agent, or starts a task from it
space	tick it off, leaving it in the list
x	delete it
tab	switch between this task's todos and the global ones
esc q	back

Repos

Key
j k · ↓ ↑	move
g G · Home End	first, last
f	fetch it
p	fast-forward it
r	read them all again
esc q	back

Review

Key
j k · ↓ ↑	move between files
space s	stage or unstage the file
a	stage everything
tab	switch between the staged and working-tree diff
ctrl-d ctrl-u	scroll the diff
c	commit — type a message, ↵ to confirm, esc to cancel
x	reject — type instructions to the agent, ↵ to send
r	refresh
esc q	back

Rejecting types your instructions into the agent's live session and returns the task to running. Instructions containing control characters are refused: send-keys -l stops tmux reading them, but the agent still sees the bytes, and \r would submit half a sentence while \x03 would interrupt the very turn being asked for.

Upgrading

marver upgrade              # cargo install marver --force
marver upgrade --restart    # and swap the daemon too

marver upgrade drives cargo, because that is how the crate is distributed. If you installed marver some other way it says so rather than guessing, and you upgrade it however you installed it.

Installing a new marver does not replace a daemon that is already running: it keeps executing the old binary, and since it still holds the socket, nothing starts a replacement. This is deliberate — that daemon is supervising live agents, and restarting it underneath them would fail every task it was watching.

marver status reports both versions and says so when they differ, and marver restart does the swap when you are ready:

marver restart

It refuses while any task is running or blocked, because that is the one thing a restart can cost. marver hook always exits 0 — marver being down must never interfere with an agent — so a hook arriving while nothing is listening is simply lost, and the task is left running behind an agent that has already finished. Wait for those tasks to reach awaiting-review, or accept it with marver restart --force.

Nothing restarts the daemon on your behalf. It is supervising live agents, and the moment to interrupt that belongs to whoever knows what they are in the middle of. Agents already running in tmux survive a restart either way.

What an agent has spent

A task's screen shows what its agent is costing you, beside the repo and branch:

 running  api  marver/4-fix-auth  opus-5 · 310k ctx · 334k out

Three things, and the two numbers mean different things. ctx is a level — how full the context window was on the most recent turn — and out is a total, output tokens across every turn since the task started. Summing the first or replacing the second would both be wrong, and both would look plausible.

Nothing is shown until the agent has answered once. Unknown is not zero: a row of zeroes would claim the agent had run and produced nothing.

There is no cost figure, deliberately. None of this is in the hook payloads — the only source is the transcript JSONL that Claude Code writes, and it records no price of any kind. A number in dollars could only come from marver multiplying tokens by a price table it hardcoded, which would be confidently wrong the day prices moved. Tokens are what the transcript knows, so tokens are what marver reports.

That transcript is an internal format with no compatibility promise, which is why this whole feature is built to fail quietly: every field is optional, unparseable lines are skipped, and a transcript that stops making sense costs a task its numbers and nothing else. Reading is incremental — each hook reads only the bytes added since the last one — so watching a long task does not get more expensive as it runs.

The repos

R on the task list opens the other half of the picture. The task list says what the agents are doing; this says what they would be starting from:

 repo        branch              upstream       default          changes
 api         main                ↓4 origin/main                  clean
 marver      release-prep        ↑2 ↓1 origin/… ↓12 origin/main  3 files
 notes       detached                                            1 file
 old-thing   gone from disk

Four things per repo, and the two middle columns are different questions. upstream is how far the checked-out branch has drifted from the branch it tracks. default is how far it is from origin/main — the one that says whether starting a task here would branch from something stale. It is left blank when it would only repeat the upstream column, which is the usual case on main itself, because a number printed twice reads as two facts.

Ahead and behind are always both counted and only the non-zero one is shown. Behind is coloured and ahead is not: work you did is work you know about, and the number worth noticing is the one that appeared while you were not looking.

f fetches the repo under the cursor and p fast-forwards it. Both are as timid as they sound, deliberately, because they run on one keystroke against a repo you are not looking at and possibly one an agent is working in:

• f is a plain git fetch — no --prune. It adds remote-tracking refs and touches no branch and no working tree.
• p is git pull --ff-only. Where a merge would have left conflict markers in a working tree nobody asked to touch, and a rebase would have rewritten commits underneath an agent, this refuses to move and says so.

Neither blocks the interface. Reading one row costs several git subprocesses and a fetch costs a round trip to a host that may be slow, unreachable, or waiting for a passphrase, so the work happens on a thread and the rows fill in as they arrive — which is also why a screen of thirty repos is readable before it is finished. One job runs at a time; a second key press is told what is already running rather than queued behind it.

The list is the last scan's, not a fresh walk of the disk. The daemon owns scanning, and a screen that went looking on its own would disagree with the repo picker on the new-task screen. A repo that has moved since keeps its row and says gone from disk rather than quietly vanishing.

Other agents

marver --harness claude                   Claude Code (the default)
marver --harness codex                    OpenAI's codex
marver --harness opencode:opencode        anything else, by name and command
marver --harness aider:aider --yes        with arguments, if it needs them

Worktrees, a queue, a tmux session, a diff to review — none of that is specific to Claude Code. Being told what the agent did is, and it is what the states run on: without it a task sits in running behind an agent that stopped an hour ago, holding a slot.

So a harness is defined by what it can say, and marver does not pretend otherwise:

	how it reports	what marver learns
claude	a settings file of hooks	started, finished, blocked, and why
codex	its notify program	finished a turn
anything else	nothing	nothing

codex has exactly one event. agent-turn-complete is the whole of it, so a codex task moves to awaiting-review on its own but never enters blocked — when it stops to ask you something, marver still believes it is working. There are no token figures either: those come from Claude Code's transcript, and codex does not write one marver reads.

Anything else is started and then watched. You get the queue, the worktree, the session, the diff, the todos, and the terminal; what you do not get is a task that finishes by itself. f on the task list is how you say so.

Neither of the two has its config edited. Claude Code takes --settings and codex takes --config notify=[…], so the reporting is wired on the command line of that one task — marver never writes to ~/.claude or ~/.codex. A tool that edits your global config to run one task has to be trusted to undo it, and a crash in between leaves you notifying a daemon about tasks that no longer exist.

f is worth knowing about even on Claude Code. A Stop that arrives while the daemon is down is simply gone — marver hook always exits 0, deliberately — and f is how that task gets to awaiting-review without waiting for something that is never coming. Like pausing, it asserts rather than observes: nothing is sent to the agent and the agent is not asked.

Pausing

p on the task list sets a task aside, and p again picks it back up. One key in both directions, because from there you are deciding whether to look at something now or later, not which state it is in.

What it does depends on where the task is, and you should know which:

• A task that has not started is simply held out of the queue. The scheduler passes over it and starts the next one instead. Resuming puts it back in the queue, subject to the cap like anything else.
• A task with a live agent — running or blocked — is interrupted. marver sends the agent esc, which is how Claude is stopped mid-turn, and frees the concurrency slot so something else can run. The tmux session stays alive, so nothing is lost.

Which one you get is decided by whether the task has a session, and where resuming returns it to is decided the same way. Nothing records where it came from, so nothing can be wrong about it.

Two things worth knowing about the second case. Nothing confirms the interrupt — Claude Code's hooks report what an agent does, not what is done to it — so pausing asserts a state rather than observing one. And because of that, resuming types continue into the session rather than only setting the state back: an agent sitting idle behind a task marver believes is running would never produce the Stop that finishes it. It is a real prompt to a real agent, not an internal signal.

Pausing happens before the state moves, so a session that cannot be reached leaves the task exactly where it was, and says so.

Todos

t opens the todos for the task under the cursor; tab there switches to the global list. They are the same screen and the same keys, and they differ in what ↵ means — which is the only thing a todo is for:

• A task's todo is a note for an agent that already exists. ↵ types it into that agent's session and ticks it off. If it could not be delivered it is left undone, since losing the note as well as the attempt would be worse.
• A global todo is something you have not scoped yet. ↵ opens the new-task screen with the text as the prompt, leaving you to pick the repos — and since the title is written from the prompt, that is the only thing left to do. It is ticked off when the task is actually queued, not when the screen opens — a task you thought better of leaves the note outstanding.

space ticks one off and leaves it in the list; x deletes it. Two keys because they mean different things, and a list that hid the finished ones would make the first look like the second.

What a task was asked to do

y on the task list puts the brief over the row:

┌ task 1 ──────────────────────────────────────────────────────────┐
│ running  Make fetch work on every repo                           │
│api                                                               │
│described 2026-08-08 21:28 · 1h ago                               │
│started   2026-08-08 21:32 · 1h ago                               │
│                                                                  │
│Prompt                                                            │
│The repo screen fetches one repo at a time.                       │
│                                                                  │
│Add a key that fetches every repo at once, reporting each one as   │
│it lands rather than waiting for all of them.                     │
└ any key closes ──────────────────────────────────────────────────┘

Described and started are two different times, and the gap between them is the queue — which is the thing you are looking for when a task has not begun. Started is read from the event log rather than a column: a task that was paused and resumed started once, and updated_at moves with every hook.

Whatever you press next closes it and does nothing else. A popup you have to learn how to close is a popup that gets in the way, and one whose dismissing key also acted would make reading it a hazard — x is right there.

The same brief is written to task.md in the task's workspace when it launches, beside the worktrees, so the directory says what it is for. The agent can read it; so can you, a month later, having found a folder of changes with nothing to explain them. It is written once and never updated — it records what the task was asked to do, which does not change. What it went on to do is what the diff and the list are for, and a file that tried to track state would be wrong between every hook.

The popup reads the store rather than that file, since a task that has not launched has a prompt and no directory yet. Writing the file is best effort: a note about the work must never be the reason the work does not happen.

Archiving

x discards the task under the cursor, and what that means depends on whether it is still going:

• A task that is still going is cancelled. It stays on the list, because a task you just stopped is one you may still want to look at — the diff it left, the reason it went wrong.
• A task that has ended is archived: its tmux session is killed and the row leaves the list. a cycles round to the archive when you want it back.

Two presses of one key, in the order the thoughts occur: not this, then and take it off my screen. No single press does both, so nothing you press once can end a task and hide it in the same movement.

Archived is not a state. It is a date on the row, beside the state the task ended in, and the archive shows that state as the list does. Making it a state would answer "is it out of the way" by throwing away "how did it end", and a cancelled task and a committed one do not read the same six weeks later.

Killing the session is the point of doing this from here rather than waiting. The daemon reaps the sessions of finished tasks on its own pass — but only while it is running, and a row that has just left the list is the worst moment to leave an agent alive behind it. Archiving is therefore also the answer to "that session is still open and there is no daemon". If the session cannot be killed the task stays on the list and says why, because a row going quiet while its agent kept running is the one outcome nobody could have deduced.

Worktrees are not touched, here or by cancelling — see below. marver cleanup still finds archived tasks, since archiving changes nothing about the state it selects on.

Cleaning up

A task keeps its worktrees after it ends. Cancelling records that it is over and archiving kills its session, and both stop there deliberately — git worktree remove --force discards uncommitted changes, and losing unreviewed work as a side effect of tidying a list would be worse than a directory left on disk.

marver cleanup is where that gets undone, as a separate act:

marver cleanup              remove the worktrees of finished tasks
marver cleanup --dry-run    list what would go, change nothing
marver cleanup --branches   delete their branches too, where merged
marver cleanup --force      remove work that was never committed

Only committed, failed, and cancelled tasks are offered. Anything else is in use: a running agent is working in its worktree, and an awaiting-review task is what the review screen reads its diff from.

A worktree still holding changes is listed and left alone, because "finished" is a state marver recorded and not a promise about what is in the directory. It says which ones and what to type if you meant it.

Branches are kept by default. marver commits onto marver/<id>-<slug> and merges it nowhere, so for a committed task that branch holds the only copy of the work. --branches deletes with git branch -d, which refuses any branch whose commits exist nowhere else — so the branch of a task that did nothing is tidied away and the branch of one that did something survives, and cleanup says how many it kept. --force --branches deletes those too.

Nothing runs this on a schedule.

See ARCHITECTURE.md for the design and the reasoning behind it.

What it costs

Measured on an M-series Mac, marver 0.0.20, release build. Percentages are of one core.

	CPU	Memory
daemon, idle	0.03%	7.4 MB
daemon, five hours of real use	0.06%	7.2 MB
interface, task list open	0.01%	8.2 MB
interface, agent's terminal, quiet	0.2%
interface, agent's terminal, agent working	1.8%
interface, agent's terminal, 9 MB/s of output	31%

The daemon does nothing between events. It blocks on its socket with a 2s timeout, so an idle one is a few queries a minute; the repo scan it runs every 60 seconds takes 70µs over five repos, because the depth limit means it looks at tens of directories rather than thousands.

The interface costs what is moving. The task list re-reads the store four times a second — 0.03ms for ten tasks, 0.5ms for a thousand — and draws only when the answer differs from last time, which while nobody is typing it never does. The agent's terminal is the expensive screen, and the expense is the VT parser: at 9 MB/s it is 31% of a core, of which drawing is 0.7 and parsing is all the rest. Nothing an agent does approaches that rate — Claude Code repainting its interface as it works measured 1.8%, and most of that is the millisecond polling that makes typing feel direct rather than relayed.

The 1000-task figure is worth knowing rather than worrying about: it is 0.2% of a core, and a on the task list hides finished tasks by default precisely so the list stays a worklist.

Known gaps

Written down because finding them yourself is worse.

The whole loop has never been run end to end. Every part is tested — 570-odd tests, several driving real tmux servers, real git repos, and the real binary — and the library has been exercised from outside. But describe a task → agent works → review the diff → commit has not been completed in one pass by a person. Expect to be the first.

Nothing reclaims a worktree unless you ask. marver cleanup exists — see Cleaning up — but there is no key for it inside the interface and nothing runs it on a schedule, so a marver left alone accumulates directories and branches until you type the command.

Task ids name tmux sessions, and a tmux server is machine-wide. Two data directories that each hold a task 1 both want the session marver-1, so the second one to launch is refused with "session already exists" and its task fails. Cleanup no longer kills across that boundary — it checks the session's working directory first — but launching is still first-come-first-served. One data directory per machine is the supported arrangement.

A hook that arrives while the daemon is down is lost. marver hook always exits 0, deliberately, so marver being unavailable never interferes with an agent — which means a Stop landing in the gap is simply gone, and the task is left running behind an agent that has already finished. Startup reconciliation does not rescue it, because it only fails tasks whose tmux session has died and that session is still there. This is why marver restart refuses while anything is running or blocked.

An old daemon keeps serving after an upgrade. By design — it is supervising live agents. marver status warns and marver restart swaps it, but nothing forces the issue, and a future schema migration could have a new interface migrating a database an old daemon still has open.

A codex task never blocks, and shows no tokens. Codex reports one event — agent-turn-complete — so a codex agent that stops to ask you something leaves its task in running, and the token figures come from a transcript only Claude Code writes. f on the task list is how a task under any harness reaches awaiting-review when its agent cannot say so. See Other agents.

A harness marver does not know is started and then only watched. Everything except the reporting works. Nothing infers that an agent has finished from what its pane printed, and nothing will: guessing at that and being wrong would move tasks under you.

ctrl-esc does not arrive in most terminals. See Keys. ctrl-q always does.

A multi-repo commit is not atomic. One message becomes one commit per repo, and if the second fails the first still stands. The task then stays in awaiting-review with part of its work committed.

Status messages last 500ms, errors included, which is not long to read one.

Invalid option values are ignored silently. Unknown options are refused, but --cap notanumber falls back to the default without saying so.

list_events cannot reach events with no task. Scans log against no task id, and the only query takes one — so those rows are written and never read back.

Nothing confirms that a pause reached the agent. Claude Code reports what an agent does, not what is done to it, so paused is asserted rather than observed. If the interrupt did not land, the agent finishes its turn and the Stop hook arrives for a task marver has already moved — the transition is refused and the task stays paused with work quietly completed behind it. Resuming and reading the pane is the way out.

Upgrading past 0.0.16 changes the schema, and so does archiving. The first new marver to open the database migrates it, and an old daemon may still be running against it. 0.0.16 added paused, which an old daemon never asks for but cannot read in a hook; the archive adds a column, which an old daemon ignores entirely. marver restart after upgrading either way.

Token figures depend on a format nobody promised to keep. marver reads Claude Code's transcript JSONL for the model and token counts — see What an agent has spent. If that file changes shape the numbers stop appearing. Nothing else depends on them, and no cost is shown at all, because the transcript records no price to show.

Requirements

tmux 3.x, git, and Claude Code.

Name

A marver is the polished steel table a glassblower rolls hot glass across — shaping the piece and controlling its temperature between trips to the furnace. It is the surface where work takes form.

License

MIT OR Apache-2.0, at your option.