Skip to content

nomograph-ai/muxr

Repository files navigation

muxr hero

muxr

pipeline license built with GitLab crates.io

Harness session multiplexer for AI coding workflows. Owns the address <repo>/<campaign> across tmux, the filesystem, and your AI runtime, so a rename or move stays coherent in all three.

What it is

muxr is a tmux + AI runtime session manager. Sessions live as typed records on disk; tmux state and the runtime command stay in lockstep with that record. muxr save snapshots; muxr restore brings everything back.

The problem

You work across multiple projects. Each one needs its own terminal session with the right working directory and the right tool running. You open tabs, cd around, lose track of what is where. When you reboot, everything is gone.

muxr organizes tmux sessions into repos (local directory trees) and remotes (GCE instances). Each session knows where it lives and what tool to run. muxr save snapshots everything. muxr restore brings it back.

Install

cargo install nomograph-muxr

Pre-built binaries for macOS arm64 and Linux amd64 are available in releases.

Quick start

muxr init                                # create config
muxr work                                # open the work repo switchboard
muxr work retrieval-precision            # open work/retrieval-precision campaign session
muxr switch                              # interactive chooser to jump or launch
muxr save                                # snapshot before reboot
muxr restore                             # bring everything back

Sessions are addressed as <repo>/<campaign> -- two levels. Campaigns are kebab-case and name the initiative (e.g. cicd-stub-fix, retrieval-precision); they are not date-stamped. The per-repo switchboard is a singleton at <repo>/switchboard. When a topic crystallizes inside a broad campaign, muxr shard <new> spins it out into a sibling campaign instead of adding a third name level.

Config

One file: ~/.config/muxr/config.toml

default_tool = "claude"

[repos.work]
dir = "~/projects/work"
color = "#7aa2f7"

[repos.work.launch]
append_system_prompt_file = "HARNESS.md"   # optional repo-level prompt
add_dirs = ["~/docs/shared"]               # optional base --add-dir paths

[repos.personal]
dir = "~/projects/personal"
color = "#9ece6a"

[remotes.lab]
project = "my-gce-project"
zone = "us-central1-a"
user = "deploy"
color = "#d29922"
# connect = "mosh"       # default; set to "ssh" for gcloud compute ssh
# instance_prefix = ""   # optional prefix for GCE instance names

Repos are local directory trees; the table key is the name you pass as the first argument to muxr. Remotes are GCE instances resolved via gcloud. Each gets a color that shows up in the chooser and the tmux status bar. Remotes require the gcloud CLI.

Extensions

muxr has a small stable core and one extension mechanism for the fiddly bits that change. At an opinionated chokepoint muxr OPTIONALLY runs a configured command (sh -c <cmd>) with structured JSON on stdin and reads structured JSON from stdout; the point name is in MUXR_EXTENSION_POINT. With nothing configured, muxr runs its built-in default and behaves exactly as before -- so bare muxr is a fully usable launcher. The transport is a subprocess (it mirrors the pre_create hooks), not a plugin ABI.

[extensions]
# RESOLVER: intent {session,repo,campaign,resume_id,model} in -> layout facts
# {dir,campaign_md,log_path,runtime,add_dirs,resume_id} out. Any omitted field
# falls back to the built-in [layout]. Absent -> the built-in config-drive layout.
resolver = "my-resolver"
# MAKE-DURABLE: fired before a session is recycled/closed. Context in ->
# {message} out (the agent-facing flush prompt). Empty message -> just exit.
# Absent -> muxr's built-in recycle-flush prompt.
make_durable = "my-flush"

# Per-session tmux env (new-session -e, tmux 3.2+). Values templated with
# {session} {repo} {campaign} {session_slug}. This is how you couple a session
# to an external tool generically, e.g. binding synthesist:
[session_env]
SYNTHESIST_SESSION = "{session_slug}"

# Delegate the interactive picker to an external tool (e.g. sesh) for plain
# attach. Absent -> muxr's built-in campaign-aware TUI (the default).
[chooser]
command = "sesh connect \"$(sesh list | fzf)\""

A runtime that does not accept --add-dir (e.g. Pi) sets supports_add_dirs = false in its [tools.<name>] block; adding a runtime is otherwise pure config.

The full design -- the core, the contract, and what's deliberately not in core -- is in ARCHITECTURE.md.

How sessions work

muxr is a thin layer over tmux. Each session gets a named tmux session, the right working directory, and your default tool running.

muxr work retrieval-precision
  tmux new-session -s "work/retrieval-precision" -c ~/projects/work
  tmux send-keys "claude" Enter
  tmux attach -t "work/retrieval-precision"

Session names follow the pattern <repo>/<campaign>. The campaign is mandatory and validated as kebab-case. Sessions persist across terminal restarts because tmux keeps them alive.

The chooser

muxr switch opens an interactive chooser that merges everything you can act on into one list, grouped by repo:

  • live sessions -- Enter attaches.
  • dormant campaigns (on disk, not running) -- Enter launches them, so every campaign is visible at a glance, not just the running ones.
  • + new campaign… per repo -- Enter prompts for a slug and creates it.

Shards render indented under their hub. Remote sessions appear alongside local ones.

Key Action
j / k Navigate
Enter Switch / open dormant / create (context-sensitive)
n New campaign in the selected repo
r Rename a live session
d Kill a live session (with confirmation)
/ Fuzzy filter
q Quit

Bind it in tmux for instant access:

bind s display-popup -E -w '80%' -h '80%' "muxr switch"

Remote sessions

Remote sessions create a local tmux proxy that connects to a GCE instance and attaches to the remote tmux. The default connect method is mosh; set connect = "ssh" for gcloud compute ssh. Sessions appear in muxr ls and the switcher alongside local sessions.

muxr lab trustchain             # connect to remote, attach tmux
muxr lab ls                     # list remote sessions

Connections auto-reconnect on drops with exponential backoff.

The system prompt is a pointer

muxr composes the launch prompt as repo HARNESS rules + the campaign's what/how + a pointer: the one-line entrypoint plus the absolute paths of campaign.md/log.md and an instruction to re-read them. It does not inline the growing log body -- a fat prompt is resent every turn (burning the context window) and goes stale the moment you /serialize.

Because the system prompt survives /compact, the re-read instruction survives too, so the durable source of truth stays the on-disk files. After a compaction, re-anchor from disk instead of the lossy summary:

muxr reorient                   # nudge the current session to re-read its files
muxr reorient work/api          # or a named session

Keep log.md's entrypoint a tight "where we are / what's next" line -- that's the pointer you move as work advances.

Save and restore

muxr save                       # snapshot all sessions to JSON
muxr restore                    # recreate after reboot

Restore recreates local sessions with the correct directory and tool, rebuilding each session's full launch command (system prompt + working dirs + resume) so a restored session is identical to a freshly opened one. Remote sessions re-establish connections.

Upgrading running sessions

muxr upgrade                    # move every claude session onto the
                                # freshly installed binary, in place
muxr upgrade --dry-run          # show what would happen, touch nothing
muxr upgrade work/retrieval-precision  # upgrade just one session
muxr upgrade --model opus-4-8   # also switch model on relaunch

For each target muxr discovers the session id, sends a graceful /exit, waits for the harness to quit, then relaunches it with the full composed command and --resume. Because the binary name is resolved fresh, the relaunch picks up a newly installed harness version (e.g. a new Claude Code release) without losing the conversation, harness rules, or working directories. Run it from the muxr control shell, not from inside an agent session.

Commands

Command What it does
muxr Control plane (bare shell)
muxr <repo> Open the repo switchboard singleton
muxr <repo> <campaign> Open or attach to a campaign session
muxr <remote> [context...] Create or attach to a remote session
muxr switch Interactive chooser: switch / open dormant / create
muxr shard <new> Spin a topic out of the current campaign into a sibling (--repo, --from)
muxr reorient [name] Nudge a live session to re-read its campaign.md + log.md (use after /compact)
muxr ls List active sessions
muxr save Snapshot session state
muxr restore Recreate sessions after reboot
muxr kill <name> Kill a session
muxr kill all Kill all sessions
muxr retire <name>|all Graceful /exit + kill; drops from saved state
muxr upgrade [name] Relaunch sessions in place on the new binary (--dry-run, --tool, --model)
muxr broadcast [/cmd] Send a slash command to every harness session
muxr rename <name> Rename: tmux + session file on disk + runtime relink
muxr migrate-layout <repo> Migrate a repo's campaigns/ tree to the 2-level model (--dry-run, --keep-archives)
muxr init Create default config
muxr completions <shell> Shell completions (zsh, bash, fish)
muxr tmux-status tmux status bar integration

License

MIT


Part of Nomograph Labs.

About

Harness session multiplexer for AI coding workflows. Mirror of gitlab.com/nomograph/muxr.

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors