Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
16 · Accessibility and long sessions
Across the app · Harness npm run contrast, npm run measure, npm run luminance
Two audiences that overlap more than they look: people who need the interface to accommodate them, and people who will sit in it for six hours. Both are served by the same things — contrast, size, stillness, and not being shouted at.
What is today — verified, not assumed
| Claim | Evidence |
|---|---|
| Every text pairing meets WCAG AA | npm run contrast: 37 pairings pass, both themes, every interactive state |
| Contrast is measured on rendered pairings | Alpha is composited over ancestors; hover measured separately |
| Reading measure is 68 characters | npm run measure at 10, 16 and 22 px, normal and focus mode |
| Dark-mode brightness is budgeted | npm run luminance: no surface > 4,000 px² exceeds luminance 0.15 |
| Reduced motion substitutes, not silences | transform: none, opacity: 1 verified under the OS setting |
| Materials are gated | Reduced-transparency and a manual override produce identical geometry |
| Controls are capsules, surfaces are not | npm run audit |
| No glass samples glass | npm run luminance walks every route |
Where it falls short
The command palette does not focus its own search field, and nothing traps focus.Fixed.src/lib/focus.tsprovidesuseSurfaceFocus, which waits for the surface to actually reportvisiblebefore focusing, trapsTabinside while open, and restores the trigger on close. Measured after the fix: focus is on the palette input on open, is still inside after fourteenTabpresses, and returns to the trigger onEscape. Applied to the palette, the shortcuts dialog, the sampling drawer and the setup modal.- Closed overlays are correctly inert, open ones are not.
inert={!open}is applied on the palette, context rail, sampling drawer, shortcuts dialog and the serving sheet — so a closed surface is out of the tab order, which is right. An open one still lets focus walk out into the page behind it. - Landmarks are adequate; the outline is not. One
<main>, one<aside>, andSectionHeadcorrectly renders<h2 className="h3">. What is missing is the level beneath: card titles are styled<div>s, so<h2>is the deepest level in the document and a screen reader cannot navigate within a section. - Toast announcements exist; engine and stream announcements do not.
.st-toastscarriesaria-live="polite"and is correct. The engine phase and the arrival of a completed response change silently. - Contrast is verified for text but not for non-text UI — icon buttons, toggle tracks, borders. WCAG 1.4.11 asks for 3:1 on those and nothing measures them.
- The reading text size tops out at 22 px and the UI chrome does not scale with it — correct by design, but there is no way to scale the chrome for someone who needs a larger interface overall.
- No high-contrast mode beyond the OS media query, which cannot be verified on a machine where it is not enabled.
- Long-session fatigue is handled ad hoc — dark accent tinting, ambient rules, no idle loops — but nothing measures it.
- The tool-approval card is the highest-stakes interaction in the app and has the least generous target sizing and no visible focus order.
The plan
Substance
Fix the focus drop and add a trap.Done.Deepen the heading hierarchy.Done, and it uncovered a second defect: every screen had two<h1>elements, because the decorative compact title copy in the scroll collapse was also an<h1>. It is now a<span aria-hidden>. Verified per route: oneh1each, withh2andh3beneath.- Announce what changes without user action — engine phase and response
completion (
aria-live="polite"). Toasts already do this correctly; leave them alone and do not announce streaming tokens. - Extend the contrast probe to non-text UI — 3:1 for icon buttons, toggle tracks, input borders and focus indicators, measured the same way text is.
- An interface scale independent of the reading size: 90 / 100 / 110 / 125%,
driving a root
--st-ui-scalethat the type and spacing tokens multiply through. The reading size stays conversation-only, as designed.
Long sessions
- A fatigue budget, measured. Extend
npm run luminancewith a second assertion: in dark mode, the total area of saturated surface on screen must stay under a threshold. This is what the green-button complaint actually was, expressed as a number. - Contrast mode verification: the probe runs a pass with
prefers-contrast: moreemulated, so the stronger-border path is tested rather than assumed.
Motion
Nothing here is decorative:
| Concern | Rule |
|---|---|
| Reduced motion | Opacity substitutes for movement; loops that carry no state stop entirely |
| Focus movement | Never animated. Focus is instant, always |
| Live-region updates | No animation on the announced element, so the announcement is not delayed |
| Long-session calm | One loop per region; nothing animates while generating except the caret and the status dot |
Verification
npm run contrastextended: non-text UI at 3:1, and aprefers-contrastpass.- New
flow a11y: for each modal, assertdocument.activeElementis inside the surface on open (this fails today for the palette), then tab to the end and assert focus never leaves; assert every screen has a heading outline deeper than<h2>; assert the approval card is operable by keyboard alone. - New luminance assertion: saturated surface area under threshold in dark mode.
- Manual pass with VoiceOver on Chat: sending, tool approval, completion.