LabWatch for VS Code

See your GPU state where you are already looking, and open the dashboard inside the editor when you need detail.

Install without the Marketplace, while the listing is in review:

code --install-extension https://github.com/Galaxy-Chjs/LabWatch/releases/latest/download/labwatch-gpu-status-1.4.0.vsix

• Status bar — GPU 3 busy / 8, or GPU 0 98% · 33GB/48GB on a single-GPU machine. Refreshes on an interval. Clicking starts the collector when nothing is running, and opens the dashboard when something is.
• LabWatch sidebar — one compact row per GPU: utilisation, VRAM, temperature. A hover tooltip adds power, process count and whether the card looks busy or free.
• Dashboard panel — the full dashboard (cards, history, filesystems, processes) opens as a webview panel in the editor area, sized like any editor. LabWatch: Open Dashboard in Browser opens the same page in a browser for a second screen.
• Commands — Open Dashboard, Open Dashboard in Browser, Close Dashboard Panel, Start in Background, Stop, Refresh, Run Doctor, Show GPU Summary, Set Up Collector, How to Connect.

The extension is a view: it runs labwatch status --json for the sidebar and status bar, and the panel renders the same dashboard the collector serves. The GPU collector lives in LabWatch, never in the editor, so the three can never disagree.

Why the panel rather than a browser tab: over Remote-SSH a browser needs the dashboard port forwarded, so the page only exists on the far side of a tunnel. The panel loads it from the extension and forwards its API calls to the collector over loopback, so there is nothing to forward and no context switch.

Requirements

The extension is the view; a small Python program — the collector — does the reading. You do not have to install it yourself. On first use, if no collector is found and Python 3.10 or newer is present, the extension offers to build a private environment inside its own storage folder and install the collector there. Nothing global is installed and PATH is not modified.

If you prefer to install it yourself, any of these works:

pipx install labwatch-lite     # a command on PATH: labwatch
uvx labwatch-lite              # run once, install nothing
pip install labwatch-lite      # then: python -m labwatch

The distribution is called labwatch-lite because labwatch on PyPI belongs to an unrelated project; what these commands install is the labwatch command used below. 发行名是 labwatch-lite（PyPI 上的 labwatch 属于别的项目），装出来的命令仍是 labwatch。

Already running the collector somewhere specific — a conda environment, say? You do not even have to configure it: interpreters conda knows about are searched before anything is installed. Setting labwatch.pythonPath simply puts one first.

Where the collector is looked for

1. labwatch.pythonPath, if you set it;
2. labwatch on PATH — pipx, uv tool, a distro package;
3. python3 -m labwatch — a plain pip install --user;
4. interpreters conda knows about, because that is where a deliberately managed environment usually lives and where PATH frequently does not point;
5. its own private environment, built inside the extension's storage folder only when nothing above exists and you agree to it.

So a machine that can already run labwatch never gets a second copy. When the private environment does have to be built, it installs from PyPI or from labwatch.pipIndexUrl (an internal mirror behaves exactly as it would by hand), and the whole log goes to the LabWatch output channel with the interpreter's own error message rather than "command failed".

When something is wrong

The sidebar never shows a blank pane, and it never reports a working collector as broken. Installation problems — needs setup, needs repair, no Python found, setup failed — get their own headline and the action that resolves them. A collector that is installed but simply not running is not an error: it offers Start, and the underlying command output goes to the LabWatch output channel rather than into the sidebar. LabWatch: Run Doctor prints the collector's own diagnostics in a tab.

Remote-SSH

This is the case the extension was designed around:

Laptop (VS Code UI)
   │  Remote-SSH
   ▼
Linux GPU server  ← extension host runs here, so it sees the server's GPUs

Two things make it work without extra configuration:

1. In a remote window the extension host runs on the server, so labwatch status reports the server's GPUs and processes, not the laptop's.
2. The dashboard panel needs no port forwarding at all: the page comes from the extension and its API calls are made by the extension host, over loopback. If you would rather use a browser, LabWatch: Open Dashboard in Browser still asks vscode.env.asExternalUri to forward the port for you. Nothing is exposed on the server's network either way.

If the collector is missing on the remote, the extension offers to set it up there — the private environment is created on the server, which is where it belongs.

Settings

Build and install from source

The panel shows the built dashboard, so sync it in before packaging:

python scripts/sync-extension-dashboard.py    # from the repository root
cd vscode-extension
npm install
npm run compile
npx --yes @vscode/vsce package --no-dependencies
code --install-extension labwatch-gpu-status-1.4.0.vsix --force

For development, open the repository in VS Code and press F5 — .vscode/launch.json starts an Extension Development Host.

Tests

npm run compile && npm test

The formatting, guidance text and Python-environment logic are deliberately kept free of vscode imports, so they run under plain Node: 34 tests cover the payload parsing, every setup state's user-facing text, and the resolution order between pythonPath, PATH, and the private environment. Nothing in the suite spawns Python or touches the network — processes are injected.

The real end-to-end path has its own check, which does build a temporary environment and install into it:

python scripts/verify-extension-setup.py     # from the repository root

Files