Skip to content

HUMAN_UX_AUDIT_2026-08-22.md

Latest commit

 

History

History
29 lines (17 loc) · 2.43 KB

HUMAN_UX_AUDIT_2026-08-22.md

File metadata and controls

29 lines (17 loc) · 2.43 KB
Raw

NiakGPT — Human UX audit correction (2026-08-22)

Finding

The existing CI labels used the phrase human UX too broadly. sidebar-human-ux-v123.spec.js is a valuable automated user-journey test running the real MV3 extension in Chromium/Brave, but the ChatGPT surface is a controlled fixture with mocked backend routes. It is not a literal human/manual test on an authenticated live ChatGPT session.

Visual inspection of the stored CI evidence also exposed lab-only quality problems that automated assertions did not report:

  • a fixture screenshot renders Récents (encoding defect in the lab surface);
  • human-nav-stress.png contains intentionally sparse/unrealistic native markup and a broken/empty media rectangle;
  • experience-v116/final.png is primarily raw fixture markup and is not representative enough to certify production visual UX.

These are not automatically production NiakGPT bugs, but they prove that a green DOM gate must not be presented as a complete human visual UX certification.

Regression found when the UX gate was actually rerun

The corrected audit immediately found a genuine interaction regression in the existing automated sidebar journey: after the chat action trigger was remounted/replaced, clicking the same semantic … action could reopen its menu instead of closing it. The menu owner compared the trigger by DOM-node identity (state.button === button) even though ChatGPT/NiakGPT remounts can replace that node between the two clicks.

The runtime fix now identifies an open action by stable semantic identity (kind + id) and updates the stored trigger to the current connected replacement before closing/focus restoration. A deterministic remount regression test was added so this behavior cannot silently return.

New evidence contract

From now on, the CI language must distinguish:

  1. Automated user journey — functional/DOM/accessibility assertions driven by Playwright.
  2. Visual evidence — screenshots captured at meaningful interaction states.
  3. Live/manual human UX — reserved for an actually authenticated live ChatGPT session reviewed by a human; CI cannot claim this by itself.

For issue #55 specifically, automated UX coverage includes rapid Project switching while backend requests overlap, a ChatGPT thinking transition during the bridge network gap, recovery after returning idle, preservation of focus/scroll/menu usability, and action-menu behavior across trigger remounts.