check grounding --mark
A checkout button an agent would miss
For computer-use agents: an action map in screenshot pixels, and for each target whether a click there actually reaches it. --mark draws the numbered overlay an agent would be handed.
Open the page under test · fixtures/grounding/partly-covered.html
Run it
vlmkit check grounding fixtures/grounding/partly-covered.html --resolution 1280x720 --mark {out}/mark.png
--resolution 1280x720 keeps the frame at full size for this page; by default the map is given in the downscaled frame a model is sent (640x360 here).What to look at
A transparent veil covers the button's centre. The map moves t1's click point to a spot that does reach it, and still reports the defect, because anything aiming at the centre misses.
The fix
Remove the veil, or give it pointer-events: none.
What vlmkit printed
vlmkit check grounding fixtures/grounding/partly-covered.html --resolution 1280x720 --mark {out}/mark.png · exit 1 — the command fails, as a gate in CI would
vlmkit check grounding source: fixtures/grounding/partly-covered.html (1280x720) status: suspect frame: 1280x720 screenshot px — cap 1280x720, scale 1.00 (every coordinate below is in these 1280x720 pixels) screen: as first loaded — --after "click x,y" measures the screen an action leaves targets: 2 actionable in frame (0 disabled or out of the frame) Action map (click point in screenshot px): t1 button "Proceed to checkout" @ (208,58) 200x40 #cta [occluded-target] t2 button "Save" @ (100,124) 120x40 #save Issues: x t1 occluded-target: #cta (button "Proceed to checkout") is covered at its own centre (140,60): a click there goes to #veil. This map aims at (208,58) instead, which does reach it, with 18px of room before the nearest thing that is not this element — still a defect, because anything aiming at the centre misses.