Port and re-audit the core accessibility candidate on DSH 0.1.2-alpha.1

[mattheliu]

Outcome

Produce a reviewable accessibility-core candidate for the current upstream development line without mechanically replaying the rc.2 patch.

Work

• Inventory every rc.2 accessibility change against 0.1.2-alpha.1.
• Identify upstream fixes, remaining gaps, conflicts, and changes that should be dropped.
• Port only the remaining core-owned semantics, focus, keyboard, announcement, reflow, contrast, and motion behavior.
• Preserve an immutable rc.2 maintenance reference and document rollback.

Acceptance evidence

• Exact upstream and candidate commits are recorded.
• The overlap inventory and rationale are public.
• Type, unit, build, assembled-browser, keyboard, and deterministic accessibility gates pass.
• The compatibility ledger and known limitations are updated.
• No host defect is represented as repaired only by the companion plugin.
• The candidate is ready for upstream review with a focused change set.

[mattheliu]

The alpha.1 accessibility-core candidate is now public for review and cross-device validation:

• PR: feat(web): harden keyboard and focus accessibility deepseek-harness#1
• Upstream/base commit: cd5ef8148158c3a752a658978873241fdf8e2bbc (dsh-v0.1.2-alpha.1)
• Candidate commit: 2998514ec183a6f0331e75a8f736795b91d2e8ad

Implemented in this candidate: modal focus containment/restoration and background inertness; complete menu-button keyboard behavior; Tooltip-wrapped trigger forwarding; named shell landmarks and application heading; keyboard-adjustable separators; closed-Details exclusion; and a non-modal Context Meter disclosure/region contract.

Automated evidence:

• focused component tests: 70 passed;
• complete GUI lane: 288 files / 3,830 tests passed, 1 skipped;
• production build: passed;
• assembled Web replay: 92 files passed, 1 conditionally skipped; 311 tests passed, 15 skipped;
• contract typecheck, contract lint, documentation 15/15 gates, and repository hooks: passed.

These results are not an assistive-technology certification. Community validation on physical systems is requested.

Run the candidate

git clone https://github.com/omdsh-dev/deepseek-harness.git
cd deepseek-harness
git checkout feat/a11y-core-0.1.2-alpha.1
pnpm install --frozen-lockfile
pnpm run build
pnpm dsh web --no-open

Open the printed local URL, then test with one of these combinations:

• macOS + Safari or Chrome + VoiceOver;
• Windows 11 + Firefox or Chrome + NVDA;
• Windows 11 + Edge or Chrome + JAWS;
• Windows 11 + Edge + Narrator;
• Linux + Firefox + Orca.

First validation slice

1. Navigate by landmarks/headings and confirm one named main area, one application H1, and named Session navigation.
2. Open Settings or onboarding dialogs entirely by keyboard; verify initial focus, repeated forward/reverse Tab containment, Escape dismissal, and focus return.
3. Open model/action menus with Enter, Space, Arrow Up, and Arrow Down; test arrows, Home/End, typeahead, submenus, Escape, and Tab exit.
4. Reach the Session-navigation separator, use arrows, Shift+arrows, Home, and End, and record whether role, orientation, and value changes are announced.
5. Open Context Meter and verify it is announced as an expanded control with a related information region, not as a modal dialog.
6. Confirm controls in a closed Details pane cannot be reached through Tab or the screen-reader cursor.

Please comment with OS, browser/version, assistive technology/version, language, scenario number, spoken output, focus result, and pass/fail. File a separate issue for each blocker and link it here. Do not convert an automated DOM pass into a manual AT pass.

[mattheliu]

The current-upstream and rc.2 overlap audit is now public as a bilingual compatibility ledger in #21 (head 949e4d6792291b30e1c8a289f6b3a93802897b7c). Its six-platform/engine validation matrix and aggregate gate are green: https://github.com/omdsh-dev/dsh-accessibility/actions/runs/33270124844

Pinned records:

• official rc.2 source: b150a551b8d465e31e418e1b2eaf5e79bbb7d28e;
• rc.2 full-patch reference: 3064d99cdc9653327e774b7306b839395b24a272;
• official/current upstream alpha.1 source: cd5ef8148158c3a752a658978873241fdf8e2bbc;
• fork portability base: 073359c4d5a5b284d60dbc052f5aa370b4639892;
• partial alpha.1 candidate: 76aa5dfd7ca98288f425857f1c9ce5da979692af.

The material audit finding is that the candidate is partial. It rebuilds shared Modal/Menu behavior, Tooltip trigger forwarding, shell landmarks/H1, closed-Details exclusion, keyboard separators, Context Meter, narrow Settings focus/reflow behavior, and the three-engine deterministic gate. It does not yet have recorded dispositions for session/workspace/search trees, model/command composites, Chat/Trajectory composites, question/feedback flows, conversation log/articles/completion announcement, tool disclosures/trees, modal consumers, or the complete theme contrast/non-color set.

Therefore #7 remains open. PR omdsh-dev/deepseek-harness#1 must not yet be represented as a replacement for the rc.2 maintenance reference, an expanded companion peer range, an alpha.1 npm release candidate, or complete accessibility. Current VoiceOver/NVDA task records and disabled-developer evidence also remain independent gates.

[mattheliu]

Community verification checkpoint — ee2420bcf34a0932db682c3dc3d77fe126fe2358

The hosted automated run is fully green on attempt 3: https://github.com/omdsh-dev/deepseek-harness/actions/runs/33274552384/attempts/3

Please help verify the following with real assistive technology and report device, OS, browser or shell, AT name/version, input method, locale, exact steps, expected result, observed result, severity and whether independent recovery was possible:

1. Tree hierarchy and focus: navigate grouped, flat and search-result Workspace/Session trees; expand/collapse nodes; confirm levels and state are spoken; confirm focus returns after leaving search results.
2. Model menu: open from its single trigger, move to first/last and adjacent options, select or dismiss, and confirm focus returns to the trigger.
3. /model command: type the command, browse suggestions, execute it, and confirm focus returns to the real message composer.

Automated evidence currently covers Chromium 6/6, Firefox 5 plus one forced-colors capability skip, WebKit 5 plus one forced-colors capability skip, 989 Linux coverage files and 956 Windows coverage files at 100%. The final Windows native attempt passed 69/69 tests with 1 declared skip. Attempts 1 and 2 had a Vitest worker-process exit without a product assertion failure; this remains recorded rather than hidden.

Passing automation is not a substitute for VoiceOver/NVDA/Narrator/TalkBack/Braille-display or disabled-developer evidence. Please do not report the whole product as accessible from this checkpoint. Versioned scope and gaps: #21.

[mattheliu]

Community real-AT verification checkpoint — 0f65b7f13b343c096f3e901889753c81feaa8155

The final automated candidate passed every required job on the first attempt: https://github.com/omdsh-dev/deepseek-harness/actions/runs/33281137892. The exact bilingual scope and remaining gaps are recorded in #21 at 05bc87d8996a58d82bf62fd8a5b909899afdb8bd; its six validation jobs and aggregate gate are green: https://github.com/omdsh-dev/dsh-accessibility/actions/runs/33282049058.

Please test this exact revision on a physical system with VoiceOver, NVDA, JAWS, Narrator, Orca, a braille display, switch input, voice input, magnification, or another assistive technology:

git clone https://github.com/omdsh-dev/deepseek-harness.git
cd deepseek-harness
git checkout 0f65b7f13b343c096f3e901889753c81feaa8155
pnpm install --frozen-lockfile
pnpm run build
pnpm dsh web --no-open

Use a disposable test project and test data. Do not paste API keys, tokens, proprietary source, private prompts, filesystem secrets, or personal data into reports.

Suggested task slices:

1. Navigate landmarks/headings, Session/Workspace/search trees, expand/collapse and focus return.
2. Open/select/dismiss model and command popups; run /model; verify exact composer focus return.
3. Operate Chat/Trajectory tabs, range/all selection, virtual ledger rows, and the keyboard-adjustable details separator.
4. Answer, cancel, fail, and recover from user questions, including a custom answer and IME composition.
5. Complete one real core workflow independently: start/open a session, understand an action or failure, decide safely, recover, and continue.

For each result, report: candidate SHA; device; OS/build; browser/shell/version; AT/version; braille/input device; locale; task and exact steps; expected and observed speech/braille/focus; completion time or blocker; severity; whether independent recovery was possible; and any safety/confidence problem. File one linked issue per blocker.

A DOM or keyboard automation pass is not a manual AT pass. Please state failures and partial results as plainly as passes; do not infer whole-product accessibility from one scenario.

[liyangbing]

这个 checkpoint 对“自动化全绿 ≠ 辅助技术验证”的界分做得很好，也明确了候选分支仍是 partial，而不是 alpha.1 的完整 replacement。

我维护的独立 DeepSeek Harness 社区手册 community plugin audit 建议把候选发布再补一层 provenance/rollback 记录：

• exact upstream base、candidate SHA、lockfile digest、Node/pnpm、browser/AT 版本和 build identity；
• 把 DOM/axe、keyboard automation、实际 VoiceOver/NVDA/Narrator 等 AT、以及 disabled-developer 验证分成独立 gate，不用 aggregate green 替代缺失的 gate；
• 每个剩余 surface（trees、menus、commands、Chat/Trajectory、question/feedback、tool disclosures、contrast/non-color）记录 covered/partial/missing disposition；
• 插件安装前保存 profile manifest/lockfile/patch，验证 plugin disable/remove、冷启动和 rc.2 maintenance reference 恢复，不把 companion plugin 的修复描述成 Host defect 已修复；
• 报告中继续使用 disposable workspace/test data，并排除 credentials、private prompts、绝对个人路径和 transcript。

这样维护者可以直接判断 candidate 是可审查的 upstream review 输入、兼容性实验还是可发布包，而不会把某个成功任务扩大成整体可访问性承诺。手册为独立社区资料，不代表官方 DeepSeek AI 项目。

[mattheliu]

Community real-AT request: quiet conversation transcript — a734cec0a00efdd18e6bd5286c7641f608c6311c

The exact candidate and automated evidence are green in deepseek-harness run 33285824826. The versioned bilingual ledger is #21 at 272f020d27dfc111dac2b9e3aded48bf41b47315, with its six OS/Node jobs and aggregate gate green in run 33287571430.

Please test this exact revision on a physical system with VoiceOver, NVDA, JAWS, Narrator, Orca, a braille display or another assistive technology:

git clone https://github.com/omdsh-dev/deepseek-harness.git
cd deepseek-harness
git checkout a734cec0a00efdd18e6bd5286c7641f608c6311c
pnpm install --frozen-lockfile
pnpm run build
pnpm dsh web --no-open

Use a disposable project and synthetic prompts. Do not report credentials, private prompts, proprietary code, personal paths or unredacted transcripts.

Focused transcript scenarios:

1. Open an already settled conversation. Verify existing history does not trigger a fresh completion announcement.
2. Navigate the named conversation log and move between “User message” / “Assistant response” articles (Chinese: “对话记录” / “用户消息” / “Assistant 回复”). Record whether structure and boundaries are understandable in browse/virtual-cursor and focus modes.
3. Send a prompt that streams for several seconds. Verify token chunks are not announced one by one. On the observed running-to-idle transition, verify exactly one completion announcement: “Assistant response finished” (Chinese: “Assistant 回复已结束”).
4. Continue navigating or trigger unrelated UI updates. Verify the completion announcement does not repeat.
5. Exercise cancel, error/retry and a long conversation. Record whether the user can independently understand state, locate the latest response, recover and continue without unsafe guesswork.

For each result, report candidate SHA; device; OS/build; browser/shell/version; AT/version; verbosity/punctuation/language; braille/input device; exact task and steps; expected and observed speech/braille/focus; completion time or blocker; severity; independent recovery; and any confidence or safety problem. Partial and failed results are as valuable as passes.

Automated results are Chromium 7/7, Firefox 6 + 1 capability skip, WebKit 6 + 1 capability skip, plus 91/91 assembled Web snapshot files. They are not manual AT certification and must not be generalized to whole-product accessibility.

[mattheliu]

Alpha.1 checkpoint update: candidate d809463d916b05acc3d6fc5821e83832e1d5e0c8 passed 19/19 CI jobs on attempt 1 without a rerun. The compatibility ledger is updated in #21, including WorkspacePicker recovery and complete failed-checkpoint history.

This issue remains open because the alpha.1 candidate is still partial: remaining modal consumers, contrast/low-vision review, current VoiceOver/NVDA/braille tasks, and disabled-developer effectiveness/safety evidence are not complete. Official alpha.2 was published after the checkpoint; its separate 1,604-path delta and 128-path candidate overlap are tracked in Issue #22.