This fork replaces T3 Code's chat area with an Onirigiri
workspace: a niri-style canvas where every conversation, terminal, browser, file view and diff is a
pane you can arrange in rows and columns, pan across and zoom out to an overview. T3's sidebar,
settings and providers are unchanged. The spatial app lives in apps/spatial.
This is an alpha. Expect rough edges; the root README lists the known limits. To install a packaged build, download it from Releases.
You need Git, Node at the version pinned in package.json (engines) and
Corepack for pnpm. Node no longer bundles Corepack, so install it with
npm install --global corepack. The engine and the onirigiri-t3-adapter package are vendored as
tarballs in apps/spatial/vendor; each file name ends with the commit it was packed from, so
nothing else needs cloning.
git clone https://github.com/riteofstring/t3code-onirigiri.git
cd t3code-onirigiri
corepack pnpm install --frozen-lockfile --ignore-scripts
T3CODE_SINGLE_ORIGIN_DEV=1 npx vp run --filter @t3tools/spatial buildTo change the engine or adapter, build and pnpm pack it, replace the tarball in
apps/spatial/vendor, update the path in apps/spatial/package.json and reinstall.
npx vp run dist:desktop:dmg:arm64 on a Mac with Rust installed writes an unsigned disk image to
release/; dist:desktop:linux and dist:desktop:win:x64 build a Linux AppImage and a Windows
installer the same way. The Onirigiri Desktop Alpha workflow builds all three on GitHub and
attaches them to the run; for a tag named onirigiri-alpha-* it also collects them in a draft
prerelease. The app installs as "Onirigiri for T3 Code (Alpha)", while its own windows and menus
still say "T3 Code (Beta)". It reads and writes the same ~/.t3 data as T3 Code, so quit T3 Code
first and keep a copy of that folder. On Windows the installer uses T3 Code's installer identity
and install folder, so it replaces an installed T3 Code. The builds are unsigned: macOS asks you
to allow the app under System Settings → Privacy & Security, and Windows SmartScreen asks you to
choose "Run anyway". They carry no update feed, so the app does not update itself.
Start the backend in one terminal. It keeps its data in its own directory, separate from any T3 Code installation:
npx vp run dev:server --port 5734 --dev-url http://localhost:5743 --home-dir "$HOME/.t3-onirigiri"Serve the spatial app in a second terminal:
T3CODE_SINGLE_ORIGIN_DEV=1 T3CODE_PORT=5734 npx vp run --filter @t3tools/spatial preview --port 5743 --strictPortOpen the pairing URL the backend prints (http://localhost:5743/pair#token=…) in Chrome. Then add a
project and enable an installed, signed-in provider in Settings → Providers. Embedded browser panes
need the desktop app.
Pane pictures need WebGPU and Chrome's HTML-in-Canvas flag (chrome://flags/#canvas-draw-element);
the desktop app turns the feature on by itself. Without them every pane stays live and nothing else
changes.
Keep the second terminal above running and start the desktop client in place of the backend:
npx vp run dev:desktop --port 5734 --dev-url http://127.0.0.1:5743 --home-dir "$HOME/.t3-onirigiri"For hot reload while editing, replace the preview command with
T3CODE_SINGLE_ORIGIN_DEV=1 T3CODE_PORT=5734 npx vp run --filter @t3tools/spatial dev.
The controls at the bottom right switch between the two-dimensional layout and a single strip, open the overview, show the minimap and set row height and column width. The keyboard button in the advanced controls lists every shortcut. The main ones (Option on macOS, Alt elsewhere):
| Keys | Action |
|---|---|
| Option + arrows | Move between panes |
| Shift + Option + arrows | Move between panes while typing |
| Option + Control + arrows | Move the focused pane |
| Option + O | Overview |
| Option + N | New pane |
| Enter | Leave the overview at the selection, or fill an empty cell |
Click New pane in an empty cell to add a chat, terminal, browser, files, diff or other tool beside the conversation it belongs to.
The workspace has a readout that shows whether pictures are being taken: the browser version,
whether HTML-in-Canvas and WebGPU are available, and how many captures started, were saved or
failed. In the browser, add ?pictureStatus=1 to the address; ?pictureStatus=0 hides it again.
In the desktop app, open View → Toggle Developer Tools and run this in the Console:
sessionStorage.setItem("onirigiri.pictureStatus", "1");
location.reload();The Pane pictures button in the workspace controls turns pictures off and on, which reloads the window, so you can compare the two.
?otherPanes=pictures is an experiment that also pictures terminal, file and browser panes while
they move. It is not supported, and ?otherPanes=live switches it off again.
Design constraints that span files are in design.md.