wtx

Every branch gets its own git worktree, ports, tmux session and coding agent.

wtx gives every branch its own checkout, its own ports, its own agent settings and its own tmux session. The agent works on that branch and nothing else: a pre-push hook stops it from pushing anywhere but its own branch, and landing the work is a human job from the main checkout.

It works with Claude Code and with opencode, and a repo picks which one, and which model, in one config file.

Install

uv tool install git+https://github.com/EPFL-ENAC/wtx
wtx install-machine --apply     # from a real terminal, not from an agent
wtx doctor

From a local checkout, install it editable so the tool follows the working tree (--force alone reuses uv's cached build and changes nothing):

uv tool install --reinstall --editable ~/code/wtx

Add it to a repo

cd my-repo
wtx init                # or, in a Claude Code session: /wtx-init

That writes wtx.toml (tracked, this is where your repo's choices live) and a three line .wt.toml, and prints the small app-side changes to make: every port becomes a variable.

Every day

wtgo feat/thing                     # branch, worktree, session, agent, attached
wtgo feat/thing --prompt brief.md   # same, and the agent starts on that brief
wtgo feat/thing --with k8s          # plus a paired worktree in the k8s repo, same branch name
wtgo all                            # bring every session back after a reboot
wtx curl backend /api/health        # this worktree's port, whatever it is
wtx monitor                         # every session on one screen
wtx land feat/thing                 # rebase, PR, merge, clean up
wtdone feat/thing                   # close the worktree, keep the branch

Moving between sessions

Inside tmux, with the prefix wtx install-machine sets up:

• prefix+s, the session list. One line per session, and a session waiting for you is coloured: yellow when it asks for a permission, cyan when it asks a question. The state is the same one the status line shows.
• prefix+g, every agent pane at once. One tile per session, waiting ones first, each showing the bottom of that session's agent pane. Enter or o on a tile goes to that session, q closes the tile.
• prefix+X, a menu on the session you are in: kill it, or close the worktree.

Outside tmux, a desktop notification tells you which session needs you. Clicking it brings the terminal to the front, on its workspace, on that session. The banner goes away on its own as soon as the agent starts working again.

Plan on one model, build on another

Turn [agent.orchestration] on in wtx.toml and a brief plans on the model you name, in plan mode. The plan ends with the effort it wants, and wtx carries that same conversation on to the implementation model at that effort. You read the plan and press accept. Nothing else.

[agent.orchestration]
enabled = true
plan_model = "fable"     # writes the plan
plan_effort = "low"      # planning is reading and thinking
build_model = "opus"     # implements it, at the effort the plan asked for

Per worktree, when you already know the plan is a big one:

wtx go feat/thing --prompt "..." --plan-effort high

Every key is in docs/schema.md.

Why

Read docs/traps.md. Every rule in this tool is there because something went wrong once.

Documentation

• docs/schema.md, every key in wtx.toml
• docs/traps.md, what breaks and what wtx does about it
• docs/migration.md, moving a repo off the old bash scripts