Repository navigation
Description
Note
🤖 Claude Opus 5.5 responding on behalf of samanyudas
What happened
The web engine's observation leaves out a checkbox whose <input> has a 0x0 box, even when its visible label is right next to it. Element Plus draws el-checkbox this way (.el-checkbox__original { opacity: 0; width: 0; height: 0; position: absolute; z-index: -1 }). The agent sees only a box and the label text, with no checkbox and no checked state.
Observed on main 34ccc5f3 through e2e mcp:
Current screen (revision b4, path /, 3 nodes):
#root document "Element Plus checkbox"
#n2 box
#n3 text="Agree"
Tapping #n3 ticks the box, but the action result says nothing changed, and the next observation is identical:
Tapped #n3.
No listed node changed within the wait after this action (re-observed as revision b30, path /); ...
The deterministic path does see the checkbox, so the two paths read the same page differently:
locate {role: "checkbox", name: "Agree"}matchescheckbox "Agree" value "on" [hidden].await expect(screen.getByRole('checkbox', 'Agree')).toBeChecked();passes after the tap.- Chrome's accessibility tree (
Accessibility.getFullAXTree) listscheckbox "Agree", not ignored. - Playwright's
getByRole('checkbox', { name: 'Agree' })finds 1 element, withisVisible()false and a 0x0 bounding box.
What this does to an agent step:
agent.act('accept the terms')has no checkbox node tocheck. It can only tap the label text, and it cannot see whether that tap ticked or cleared the box.- A judgment such as
agent.assert('the terms are accepted')reads the same tree, so it has no checked state to judge.
Cause: isHidden in packages/web/src/in-page/read-semantics.ts (about line 1104) treats a box with no width or height as hidden, to match Playwright's toBeVisible. That rule is reasonable for visibility, but for a checkable control with a visible <label> it removes the only node that carries the state.
A possible direction, not tested: when a checkbox or radio is hidden only by its box size and one of its labels is visible, list it, or carry its role and checked state on the label's node. A fix for this belongs in a crosscheck fixture first (packages/web/tests/crosscheck/fixtures.ts). No current fixture covers this case. I did not run the crosscheck on this page.
Reproduction
No external repo. The page below is self-contained.
page.html:
<!doctype html><meta charset="utf-8"><title>Element Plus checkbox</title>
<style>
.el-checkbox{cursor:pointer;display:inline-flex;align-items:center;position:relative}
.el-checkbox__input{cursor:pointer;display:inline-flex;position:relative}
.el-checkbox__original{opacity:0;z-index:-1;outline:none;width:0;height:0;margin:0;position:absolute}
.el-checkbox__inner{display:inline-block;position:relative;border:1px solid #dcdfe6;width:14px;height:14px;background:#fff;z-index:1}
.el-checkbox__input.is-checked .el-checkbox__inner{background:#409eff}
</style>
<label class="el-checkbox"><span class="el-checkbox__input"><input class="el-checkbox__original" type="checkbox" id="agree"
onchange="this.parentElement.classList.toggle('is-checked', this.checked)"><span class="el-checkbox__inner"></span></span><span class="el-checkbox__label">Agree</span></label>Serve it on any port as text/html; charset=utf-8, and point a web() target's app.url at it.
Steps to reproduce
- Open an
e2e mcpsession on that target and callobserve. There is nocheckboxnode, onlyboxandtext="Agree". - Call
tapon thetext="Agree"node. The result reports no listed node changed. - Call
observeagain. The tree is the same as in step 1. - Call
run_codewithawait expect(screen.getByRole('checkbox', 'Agree')).toBeChecked();. It passes, so the tap did tick the box.
Expected: the observation lists the checkbox and its checked state, so the tap in step 2 reports a change.
Versions
e2e 0.23.0, @e2e-dev/web 0.15.5 (main 34ccc5f3), Node 26.8.2, Chromium from playwright-core, Linux.
Engine
@e2e-dev/web
Anything else
Related: #1276 (Element Plus radio-button refused by check/tap) and #1355 (click a checkbox or radio through its own label). #1355 changes how the click lands. It does not change what the observation lists, so this issue stays open after #1355.
Activity
okwasniewski commentedon Nov 10, 2026
[claude-opus-5-5] automated triage on behalf of @okwasniewski
Confirmed and tracked in #1595, which is stacked on #1355 (#1276, clicking a radio/checkbox through its label). #1355 has to merge first. Both are labeled Ready for Human Review.
fix(web): list a checkbox or radio its own label draws over no box (#…