linux-cpu

The cheapest lane and the one most runs start from. A ComfyUI checkout in a venv on a GitHub-hosted Ubuntu runner -- and an attach lane, so it proves your pack runs, not that it installs.

OS / accelerator	Linux / CPU
Install method	manual -- venv + a ComfyUI checkout
Runner	ubuntu-latest (GitHub-hosted, 1x billing)
Install path	attach
Config key	[test.linux]
Also accepts	linux, linux_cpu

What a green cell proves

That your pack imports, registers its nodes, and executes its CPU workflows against a real ComfyUI server on Linux. For most packs this is the single most informative lane per minute spent.

What it does not prove

That your pack installs. The workflow builds the venv, clones ComfyUI and installs your requirements in YAML behind a cache, then hands comfy-test a running server via --server-url. The install level does almost nothing, and provenance.install_mode records attach.

The cache key is only (lane, Python version), so ComfyUI and the torch family stay frozen at whatever HEAD first populated it until GitHub evicts. This lane therefore does not exercise the pin from ADR-0005 -- see torch, torchvision and torchaudio.

For proof of installability, run linux-cuda, or run comfy-test locally: both take the fresh path.

Gotchas

• Case sensitivity. Linux is the lane that catches import MyNodes when the file is mynodes.py. Windows and macOS will not.
• No GPU. Anything gated on torch.cuda.is_available() silently does not run here. That is what syntax is for -- it fails hardcoded .cuda() before a server is ever started.

See also

• Lanes -- all ten, and what attach costs you
• linux-cuda -- the same lane, fresh, with a GPU