cyberhirsch-seb_nodes/tests

360 tests

Every test here drives a real headless Chromium (Playwright) against a running ComfyUI. The embedded browser pane in the desktop app does not composite -- requestAnimationFrame never fires -- so nothing WebGL-based can be checked through it. Headless Chromium composites, so the editor and the node-face preview actually paint and can be measured by pixel.

One-time setup

pip install playwright pillow numpy opencv-python
python -m playwright install chromium

Start a scratch ComfyUI

Never test against the instance you are working in; the tests clear the graph and commit into the panorama store. For the desktop install:

G:\AI\ComfyUI\.venv\Scripts\python.exe G:\AI\ComfyUI\Install\resources\ComfyUI\main.py ^
    --port 8189 --base-directory G:\AI\ComfyUI ^
    --front-end-root G:\AI\ComfyUI\Install\resources\ComfyUI\web_custom_versions\desktop_app ^
    --cpu --disable-auto-launch

--front-end-root matters: the desktop app ships the frontend in its Electron bundle, not as the pip package main.py looks for. --cpu is fine -- nothing here samples.

Run

python tests/run_all.py            # everything
python tests/run_all.py mask yaw   # by name

run_all.py writes the marker panoramas into the instance's input folder on first use (make_testpano.py), runs each test_*.py as its own process, and prints PASS / FAIL / SKIP per file with the failing VERDICT block. Exit code = number of failures.

Environment (all optional): SEB_PORT (8189), SEB_INPUT_DIR (the scratch instance's input/, default derived from where this pack is installed), SEB_WORKFLOW_DIR (G:\AI\_Workflows\Utility).

What each test proves

test	proves
test_yaw_truth	Python's yaw/pitch point where the panorama says: +90 is longitude +90, +pitch is up
test_write_roundtrip	Write 360 puts a view back exactly where Read 360 took it; the mirror longitude is untouched
test_core_mask	core mask keeps the margin as context; Write feathers the commit edge instead of stopping dead
test_store_migration	changing canvas_width resamples committed work instead of discarding it
test_face_layout	one preview on the node face, box follows the framing aspect, editor button visible, edits and editor sync both re-fit
test_save_reframes	editor Save re-frames the node preview, and onto the same marker Python extracts
test_editor_fresh_graph	the editor shows the panorama before any run, and follows an image swap
test_camera_edit_persists	a camera typed on the node survives the editor rewriting state_json
test_editor_mask_path	a mask painted in the editor reaches Python's MASK output (editor mode); core ignores it
test_workflows_load	every seb_360_*.json loads with no unknown seb node types; missing third-party packs are listed, not failed
test_folder_loader	Load Images From Folder (Seb). one per run: a single queued prompt processes the whole folder one image per run, sequentially, saving as it goes, stops after the last, and a failed run does not continue the chain; past-the-end is a clear error. all in one run: every image in one run, source names kept via Save Image (Seb), alpha → mask, non-images ignored, chunking, unchanged folder cached
test_save_source_modes	Save Image (Seb) source_mode: same folder as source lands next to the original, named after it + suffix, never over it; replace source file overwrites in place keeping the format (JPEG/PNG/WebP), atomically, no temp files left; ignore source unchanged
test_pano_render	(older) the editor paints the panorama after a run
test_pano_roundtrip	(older) editor camera -> node widgets, view_height follows the framed aspect

Screenshots the tests take land next to them as _*.png and are ignored by git.

Reading a failure

Each test prints one JSON object. VERDICT is the list of named checks; the keys above it are the measurements those checks were made from, so a false can be traced to a number without re-running.