Add browser, forced-colors, reflow, focus-obscuration, and motion release gates

[mattheliu]

Outcome

Catch non-screen-reader accessibility regressions that DOM checks alone cannot detect.

Required scenarios

• Chromium, Firefox, and WebKit assembled application routes.
• 200% and 400% zoom or equivalent reflow.
• Windows forced colors and system color preservation.
• Focus visibility and focus not obscured by sticky or modal surfaces.
• Reduced motion and no information conveyed only through animation.
• Keyboard-only completion of the published P0 task set.

Acceptance evidence

• Versioned test fixtures and commands are reproducible in CI where possible.
• Manual-only checks have a structured result format and named owner.
• Failures block the appropriate release gate.
• The support matrix distinguishes automated evidence from human observation.

[mattheliu]

Implementation evidence update (2026-08-30)

The required cross-browser accessibility gate is now implemented on omdsh-dev/deepseek-harness PR #1 at head 76aa5dfd7ca98288f425857f1c9ce5da979692af.

• Full evidence comment: feat(web): harden keyboard and focus accessibility deepseek-harness#1 (comment)
• Successful CI run: https://github.com/omdsh-dev/deepseek-harness/actions/runs/33264686646
• Accessibility matrix: Chromium 4/4; Firefox 3 passed + 1 Chromium-only forced-colors skip; WebKit 3 passed + the same skip.
• Linux exhaustive coverage: 16,004 passed, 41 skipped; 100% statements, branches, functions, and lines. Windows exhaustive coverage also passed.

The automated gate covers assembled-app landmarks and controls, 200%/400%-equivalent reflow, horizontal overflow, keyboard navigation, visible and unobscured focus, Settings modal containment/restoration, forced colors, and reduced motion.

This issue remains open: browser automation is not a substitute for observation with real screen readers, alternate input, or disabled users. Those results must be recorded separately and remain release evidence, not inferred from this CI result.

[mattheliu]

Partial P0 evidence: versioned non-AT browser contract

PR #19 adds the draft dsh-non-at-browser/1.0.0-draft contract and applies it to the Accessible View consumer introduced by #18:

• PR: test: add versioned non-AT browser gates #19
• Exact evidence summary: test: add versioned non-AT browser gates #19 (comment)
• Green workflow: https://github.com/omdsh-dev/dsh-accessibility/actions/runs/33269053072
• Exact DSH assembled job: https://github.com/omdsh-dev/dsh-accessibility/actions/runs/33269053072/job/99144135675

The public CI run passed:

• macOS, Ubuntu, and Windows × Node 22.19.0/24.x: all six validation jobs;
• exact DSH 0.1.1-rc.2 baseline at 3064d99cdc9653327e774b7306b839395b24a272;
• Chromium 149.0.7827.55, Firefox 151.0, and WebKit 26.5;
• 640/320 CSS px page overflow: 0 and programmatic scrollX 0 in all engines;
• Clear, Copy, and Tool disclosure at 320 CSS px: focus-visible and 9/9 topmost samples in all engines;
• reduced-motion media active with no moving transitions, named animations, or running motion keyframes;
• Chromium forced-colors emulation active with zero visible forced-color-adjust: none opt-outs;
• 4/4 exact assembled scenarios, including the original Accessible View privacy/keyboard/axe protocol;
• aggregate CI gate green.

The focus gate found a real defect during development: Copy could receive focus while fully covered by DSH's sticky composer at 320 CSS px. PR #19 fixes it by including the shell's inherited composer height in the controls' focus scroll margin.

Why this issue remains open

This is partial evidence:automated for Accessible View, not coverage of the whole DSH application. It is also not real 200%/400% browser zoom, Windows High Contrast, macOS Increase Contrast, magnifier, screen-reader/braille, switch/voice input, low-vision task evidence, or disabled-user validation.

Before #9 can close, the same versioned contract must cover every published P0 Web task route; named owners/results are still required for real zoom, text scaling, Windows High Contrast, focus-indicator contrast, OS Reduce Motion, and keyboard-only task completion. Failures must block the applicable release. This evidence does not authorize “fully accessible,” certified, AT-tested, or user-validated claims.

[mattheliu]

Cross-engine automated browser evidence — ee2420bcf34a0932db682c3dc3d77fe126fe2358

Hosted run attempt 3: https://github.com/omdsh-dev/deepseek-harness/actions/runs/33274552384/attempts/3

• Chromium: 6/6 passed.
• Firefox: 5 passed; forced-colors skipped because that engine does not expose the required emulation.
• WebKit: 5 passed; forced-colors skipped for the same capability boundary.
• Linux coverage: 989 passing files, 7 skipped, 100%.
• Windows coverage: 956 passing files, 3 skipped, 100%.
• Windows native final attempt: 69/69 passed, 1 declared skip.

The browser scenarios cover tree hierarchy/roving/focus return, the model selector, and the command popup including its production composer-focus binder. Two earlier Windows-native attempts ended in a Vitest pool-worker exit without an assertion failure; the third attempt passed the same suite and no assertion was waived.

This is automated browser evidence, not screen-reader speech-output evidence, disabled-user evidence, mobile AT coverage or a whole-product conformance claim.

[mattheliu]

Final cross-engine automated gate — 0f65b7f13b343c096f3e901889753c81feaa8155

Run 33281137892 passed every required job on its first attempt.

• Accessibility browser matrix: Chromium 6/6; Firefox 5 passed + 1 forced-colors capability skip; WebKit 5 passed + the same capability skip.
• Linux coverage: 989/996 files, 16,016/16,057 tests, 100% statements/branches/functions/lines.
• Windows coverage: 956/959 files, 15,344/15,369 tests, 100% statements/branches/functions/lines.
• Windows native also passed on the first attempt.

The gate retains assembled-app keyboard, reflow-equivalent, focus geometry/obscuration, reduced-motion and Chromium forced-colors-emulation evidence, now alongside automated coverage of the rebuilt Session/Workspace/search trees, model/command composites, Chat/Trajectory composites, keyboard-adjustable separator, and user-question flows. The prior projection-cache ordering failures were resolved as a product race and the deterministic regression proof remains enabled; no assertion was waived.

The exact scope ledger is #21 at 05bc87d8996a58d82bf62fd8a5b909899afdb8bd, with a green six-platform/Node matrix and aggregate gate in run 33282049058.

This issue remains open. The automated gate is still not real 200%/400% browser zoom, Windows High Contrast, macOS Increase Contrast, magnifier observation, complete focus-indicator contrast review, OS Reduce Motion, alternate input, real screen-reader/braille output, keyboard-only published P0 completion across every route, or disabled-user validation. Those results remain release-blocking evidence rather than inferred passes.

[liyangbing]

This is a useful complement to DOM-only checks. I would make each manual result bind to an exact route, viewport/zoom, browser build, fixture state, and interaction trace so a later UI change cannot accidentally reuse a stale “pass.”

For DSH’s dynamic surfaces, add focused cases around a streaming turn, tool/approval card insertion, Session switch, reconnect, and mobile/sidebar expansion. The assertion should cover the active element before/after the update, whether focus is visibly obscured, and whether a keyboard user can reach cancel/approve/send without pointer recovery. A visually hidden or truncated status should also have an accessible name/live-region policy that does not announce every token delta.

For forced colors and 400% zoom, check semantic order and action reachability rather than only screenshots: no clipped composer, no horizontal trap, no icon-only state without a text/name, and no contrast-dependent distinction between pending/failed/approved. Reduced motion should suppress transitions without suppressing the state change or its announcement.

I’d keep automated, manual, blocked, and not-applicable as distinct result states, with a named owner and evidence link for manual checks. A failure should block only the affected release gate, but an unknown/manual-not-run result must not be reported as pass. The handbook’s release-gate framing may help structure the matrix: agent harness scorecard.

[mattheliu]

Cross-engine transcript gate — a734cec0a00efdd18e6bd5286c7641f608c6311c

The exact candidate's accessibility browser matrix passed:

• Chromium: 7/7.
• Firefox: 6 passed + 1 forced-colors capability skip.
• WebKit: 6 passed + the same capability skip.

The seventh scenario uses a dedicated assembled-browser page with a real seeded Assistant turn. It verifies a named conversation log, named user/Assistant articles and transcript isolation across engines. Focused component coverage additionally proves that settled initial history is silent, streaming chunks do not create token-by-token live announcements, a running-to-idle transition creates one bounded atomic completion status, and unrelated renders do not repeat it.

The complete snapshot/artifact job passed 91/91 Web files and 307 tests, with 13 conditional skips across 320. Linux and Windows remain at 100% statements/branches/functions/lines. Exact counts, the single Windows failed-job rerun history and remaining boundaries are recorded in #21 at 272f020d27dfc111dac2b9e3aded48bf41b47315; ledger CI is green in run 33287571430.

This issue remains open. The new scenario is not listener-verified screen-reader speech, real 200%/400% zoom, Windows High Contrast, macOS Increase Contrast, magnifier observation, OS Reduce Motion, alternate input, mobile AT, braille output or disabled-user task validation. Streaming tool/approval insertion, reconnect/session switching and complete focus-obscuration coverage also still need manual, version-bound evidence.