opencode4

Press SUPER + ] and four opencode panes open up: one orchestrator, three workers, a whole little team.

日本語

What this is

It's opencode running in a 2x2 grid of tmux panes. Top left is the orchestrator, the one who hands out tasks. The other three are workers; they do the work. Everyone talks through a plain text log on disk, so the orchestrator can read all the reports without tab-hunting.

Each pane carries a Todo sidebar on its right showing the agent's current task. When a task is done, oc-todo-clear wipes the entry, so the sidebar only ever shows what's still in flight.

I built this because I missed cmd + ] from my Mac days and wanted it on Omarchy. It's nothing fancy: a few bash scripts and one keybinding. If you've got tmux, opencode, and a Hyprland box, it'll run.

What you need

• Hyprland (omarchy ships with it; current versions work fine)
• tmux and opencode
• a terminal: ghostty, kitty, foot, alacritty, or xdg-terminal-exec
• uwsm-app, so the key can open a terminal inside the Wayland session. Optional — the installer just warns if it's missing.
• something to feed opencode models: a local Ollama instance or a cloud provider both work

Install

The easy way:

git clone https://github.com/drkai-lab/opencode4.git
cd opencode4
./install.sh

That puts opencode4, oc-send, oc-todo-view, and oc-todo-clear into ~/.local/bin, copies the two agent definitions into ~/.config/opencode/agents/ (your existing files are left alone), adds a binding to ~/.config/hypr/bindings.conf, and reloads Hyprland. Run it again and it just skips what's already done.

By hand:

install -m755 bin/opencode4 bin/oc-send bin/oc-todo-view bin/oc-todo-clear ~/.local/bin/
cp agents/orchestrator.md agents/worker.md ~/.config/opencode/agents/

then add one line to ~/.config/hypr/bindings.conf:

bindd = SUPER, bracketright, OpenCode 4-pane, exec, uwsm-app -- ghostty -e ~/.local/bin/opencode4

Using it

1. Hit SUPER + ]. The grid opens in whatever directory you're in — or reattaches to the session if it's already running.

2. Top left is the orchestrator. The rest are workers.

3. From the orchestrator pane, tell a worker what to do:

   oc-send WORKER-1 "Refactor the auth module and report back."

4. Workers write their replies to the wire log (~/.local/state/oc-grid/wire.log). The orchestrator reads it and folds the results together for you.

5. Watch the Todo sidebars: each one shows the agent's current task, and clears itself once the task is reported done.

6. Leaving? Ctrl-b d detaches. The session keeps running in the background, and SUPER + ] brings you back.

Two more keys control the grid from anywhere:

• SUPER + [ closes the terminal window but keeps the four agents running in the background. The session stays alive, so SUPER + ] returns you to it.
• SUPER + W shuts everything down: the four agents and the terminal window. Note that this replaces the usual "close window" meaning of SUPER + W. SUPER ALT + W still closes a window for good.

Everything the grid knows lives in ~/.local/state/oc-grid/ — the wire log, the role→pane map, and the todo/ files behind the sidebars. Remove that folder to start over.

Adjusting it

The agent definitions in ~/.config/opencode/agents/ decide what the orchestrator and workers look like. The installer never overwrites them, so edit freely. Usually all you change is the model: line to point at your backend. Restart the grid and you're set.

The Todo sidebars read from ~/.local/state/oc-grid/todo/<ROLE>. oc-send writes the task it sends there, and oc-todo-clear <ROLE> removes it (or swaps in a new one). The sidebar width is the OC_GRID_TODO_PCT environment variable, 23% of each pane by default.

Removing it

./uninstall.sh

Removes the scripts, the agent definitions (unless you changed them), and the binding, then reloads Hyprland.

License

MIT © drkai-lab