HTML / ARIA · 8 OCT 2026 · MCP 0.0.83

Verification with MCP 0.0.83

This additional verification on 8 October 2026 confirms the clicks and the Tab-access barrier. It also compares page representations and two ways to test activation.

Clicks and keyboard

Variant	MCP click	Tab access
A — `div`	0 → 1	Skipped
B — `div role="button"`	0 → 1	Skipped
C — `button`	0 → 1	Reached

All six MCP clicks succeed: three variants, with and without cursor: pointer. The six keyboard tests on the original sources produce four failures (A/B) and two passes (C). Test errors are preserved.

All 21 lines in the pre-interaction YAML block are identical for B and C within each dataset. Page URLs in response metadata differ. Reference e14 is local to each session. The cursor: default override removes the pointer annotation without removing the reference or preventing the click.

MCP original : A · B · C
MCP cursor: default : A · B · C
Tab / Enter / Space : A · B · C

Representations

NOTES appears in the MCP YAML despite aria-hidden=true on its parent. It is absent from the default body.ariaSnapshot() with Playwright 1.64.0 and from non-ignored nodes in the Chromium tree retrieved through CDP. Files preserve the full getFullAXTree response, including ignored nodes; the finding concerns exposed nodes only.

Representation files : A · B · C

- generic [active] [ref=e1]:
  - banner [ref=e2]:
    - generic [ref=e3]:
      - generic [ref=e4]: ATELIER
      - link "Voir le produit" [ref=e5] [cursor=pointer]:
        - /url: "#product"
  - main [ref=e6]:
    - region [ref=e7]:
      - generic [ref=e9]: NOTES
      - generic [ref=e10]:
        - paragraph [ref=e11]: Papeterie · Collection essentielle
        - heading "Carnet de notes" [level=1] [ref=e12]
        - paragraph [ref=e13]: Un carnet pour les idées, les croquis et les choses à retenir. Format A5, papier ivoire, 160 pages.
        - button "Ajouter au panier" [ref=e14] [cursor=pointer]
    - generic [ref=e15]:
      - paragraph [ref=e16]:
        - text: "Articles dans le panier :"
        - status [ref=e17]: "0"
      - button "Vider le panier" [ref=e18] [cursor=pointer]
    - link "Retour à la présentation" [ref=e20] [cursor=pointer]:
      - /url: index.html

Focus and activation

Each case uses a fresh context. click() adds an item for B and C. press(Enter) leaves the count at zero for B. In a separate case, only tabindex=-1 is added in memory to C: press(Enter) adds an item, while two Tab presses reach Empty cart and add nothing.

Variant	press(Enter)	Two Tabs then Enter
B	0	0
C	1	1
C with tabindex=-1	1	0

measurements.json : activation

Reproduce

Chrome 155.0.8059.39 · darwin 25.3.0 arm64 · Node v24.19.0
Playwright Test 1.64.0 · MCP 0.0.83
MCP / Playwright 1.64.0-alpha-1790635538000
2026-10-08T19:22:24.745Z

The five files in each dataset are identical to those in the initial evidence pack. Their hashes are recorded in the summary.

git clone https://github.com/edikkaweb/html-aria-agent-demo.git
cd html-aria-agent-demo/preuves/2026-10-08/controle-v21
npm ci
npm run measure

Node.js 22.19 or later and Google Chrome. New output goes to runs/DATE; existing directories are rejected. The collector checks the expected experimental observations. A/B keyboard tests retain failures and exit code 1.

Reports and sources

Scope

One run per condition and version, Chrome on macOS. Deterministic clicks without autonomous-agent decisions; no screen-reader trial. Lighthouse was not rerun: the six initial reports remain the evidence for the scores. No overall conformance conclusion.