Computer-Use Agent Taxonomy中文
R02.07.01.03 · TASK CLASS

Accessibility audit & remediation

Audit an app or site with screen readers and checkers, then fix the defects.

20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.

Typical applications
axe · Lighthouse · VoiceOver · Accessibility Inspector
Class GUI-need label
HY · Hybrid

Task proposals · 20

Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.

R02.07.01.03-T01 Clear all serious axe violations on a five-page site

Original instruction

The sandbox marketing site in site/ (static HTML/CSS/JS, served at localhost:8080) has axe-core violations on 5 pages. Run axe in the browser, fix every serious and critical violation in the source, and keep the visual design unchanged.

Platforms
mac · macOS win · Windows lin · Linux
Applications
axe DevTools · Chrome
Interfaces
brow · Browser term · Terminal
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
HY · Hybrid
Interaction primitives
code · Code & commands vis · Visual judgment
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
2 / 5

Proposed verifier

Grader runs axe-core 4.9 on all 5 pages: zero serious/critical violations; total violations ≤2 minor. Screenshot SSIM ≥0.97 vs baseline per page. Hidden checks for aria-hidden misuse to block 'hide the element' shortcuts.

Proposed partial credit

60% serious/critical cleared (per page); 20% design unchanged; 20% no aria-hidden/role abuse

R02.07.01.03-T02 Audit a checkout flow with VoiceOver only

Original instruction

On macOS, use VoiceOver (no mouse) to complete the sandbox checkout at localhost:3000 in Safari. Record every barrier you hit in output/vo_audit.csv (step, element, WCAG criterion, what VoiceOver announced, expected) and note whether you could finish.

Platforms
mac · macOS web · Any OS (web)
Applications
VoiceOver · Safari
Interfaces
brow · Browser
GUI need · task setting
G2 · GUI-essential
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
key · Shortcuts vis · Visual judgment
Work product
WP01 · Document
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

8 seeded barriers (unlabeled icon button, focus trap, unannounced error, wrong heading order…). Rows matched by step/element and criterion; recall and precision scored. VoiceOver speech log (captured by grader) must show the agent reached the confirmation page. Pointer devices are disabled and the input log must show only keyboard/VoiceOver commands.

Proposed partial credit

55% barriers found; 15% criteria correct; 15% precision; 15% completed with VO

R02.07.01.03-T03 Fix keyboard navigation, focus order and modal focus trapping

Original instruction

In the React app fixtures/booking, make every interactive element reachable by Tab in visual order, add a visible 3:1 focus indicator, a skip link to main, and make the date modal trap focus, close on Escape and return focus to its trigger.

Platforms
mac · macOS win · Windows lin · Linux
Applications
Chrome · axe DevTools
Interfaces
brow · Browser term · Terminal
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
HY · Hybrid
Interaction primitives
code · Code & commands key · Shortcuts
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Hidden Playwright keyboard tests: Tab sequence equals spec order (24 stops), focus ring contrast ≥3:1 measured on screenshots, skip link moves focus to main, modal trap/Escape/return-focus behavior. Mouse-only handlers on divs rejected by eslint-plugin-jsx-a11y.

Proposed partial credit

35% tab order; 20% focus indicator contrast; 15% skip link; 30% modal behavior

R02.07.01.03-T04 Repair brand color tokens to meet WCAG contrast with minimal change

Original instruction

input/tokens.json defines 14 brand colors used for text and UI components; 9 pairs listed in input/pairs.csv fail WCAG 2.2 AA. Adjust tokens so every listed pair meets its required ratio (4.5:1 text, 3:1 UI), changing each color as little as possible, and rebuild the style guide page.

Platforms
mac · macOS win · Windows lin · Linux
Applications
Colour Contrast Analyser · Chrome
Interfaces
brow · Browser term · Terminal
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
vis · Visual judgment code · Code & commands
Work product
WP03 · Visual layout
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Grader computes contrast for all pairs: all pass. Mean ΔE2000 between old and new tokens ≤ reference solution ×1.25; hue shift ≤10° per token. Style guide renders with new tokens (computed CSS check).

Proposed partial credit

50% all pairs pass; 30% minimal change (ΔE); 10% hue preserved; 10% style guide rebuilt

R02.07.01.03-T05 Remediate an iOS app for VoiceOver

Original instruction

fixtures/Transit (SwiftUI) fails Xcode Accessibility Inspector audits on 4 screens. Fix labels, traits, grouping of route rows, heading traits for section titles, and Dynamic Type clipping, then confirm with VoiceOver in the Simulator that a route can be found and favorited.

Platforms
mac · macOS ios · iOS/iPadOS
Applications
Xcode · Accessibility Inspector · iOS Simulator · VoiceOver
Interfaces
mob · Mobile touch desk · Desktop GUI
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
HY · Hybrid
Interaction primitives
code · Code & commands vis · Visual judgment touch · Touch gestures
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
4 / 5

Proposed verifier

Hidden XCUITest runs performAccessibilityAudit on 4 screens: zero issues. Element tree checks: each route row is one element with label 'Line 5, to Airport, 4 min'; headings have header trait; at AX5 text size no truncated labels.

Proposed partial credit

40% audit clean; 25% row grouping/labels; 15% heading traits; 20% Dynamic Type

R02.07.01.03-T06 Fix TalkBack labels and touch targets in an Android app

Original instruction

In fixtures/Pantry (Android Views), fix missing content descriptions, 48 dp minimum touch targets, focus order on the edit screen and announce list updates. Verify with TalkBack and Accessibility Scanner on the API 34 emulator.

Platforms
mac · macOS win · Windows lin · Linux and · Android
Applications
Android Studio · TalkBack · Accessibility Scanner
Interfaces
mob · Mobile touch desk · Desktop GUI
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
HY · Hybrid
Interaction primitives
code · Code & commands touch · Touch gestures
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Espresso tests with AccessibilityChecks.enable() over 5 screens: zero errors. Hidden UiAutomator traversal of the edit screen records focus order == spec. Live region announcement present on list change (accessibilityLiveRegion polite).

Proposed partial credit

40% ATF checks clean; 25% touch targets; 20% focus order; 15% live region

R02.07.01.03-T07 Make a WPF app usable with Narrator and high contrast

Original instruction

fixtures/Ledger (WPF) is unusable with Narrator: unnamed buttons, custom controls without automation peers, mouse-only grid editing, and invisible text in high-contrast themes. Fix it and verify with Accessibility Insights for Windows FastPass and Narrator.

Platforms
win · Windows
Applications
Visual Studio · Accessibility Insights · Narrator
Interfaces
desk · Desktop GUI
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
HY · Hybrid
Interaction primitives
code · Code & commands key · Shortcuts vis · Visual judgment
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
4 / 5

Proposed verifier

Accessibility Insights CLI scan of 4 windows: zero failures. FlaUI tests: all buttons have Name, custom gauge exposes RangeValue pattern, grid cells editable via keyboard. High-contrast screenshot: text/background contrast ≥7:1 at OCR boxes.

Proposed partial credit

35% scan clean; 25% automation peers; 20% keyboard grid; 20% high contrast

R02.07.01.03-T08 Make a dashboard's tables and charts accessible

Original instruction

The sandbox analytics page renders a div-based table and 3 SVG charts with no text alternatives. Convert the table to semantic markup with headers and caption, give each chart a concise description plus an accessible data table toggle, and ensure sort buttons announce state.

Platforms
mac · macOS win · Windows lin · Linux
Applications
axe DevTools · NVDA · Chrome
Interfaces
brow · Browser term · Terminal
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
code · Code & commands vis · Visual judgment
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Grader: table has th scope and caption; NVDA (via guidepup) reading a cell announces its row and column headers; each chart has role=img + aria-label or figure/figcaption and a data table whose values equal the chart's data; aria-sort updates on click. axe clean.

Proposed partial credit

30% table semantics; 35% chart alternatives with exact data; 20% sort state; 15% axe clean

R02.07.01.03-T09 Make form errors perceivable to screen readers

Original instruction

The sandbox insurance quote form shows errors only as red borders. Add programmatically associated error messages, an error summary that receives focus on submit, required field indication, and autocomplete attributes, without changing validation logic.

Platforms
mac · macOS win · Windows lin · Linux
Applications
Chrome · NVDA · axe DevTools
Interfaces
brow · Browser term · Terminal
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
code · Code & commands type · Text & forms
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Playwright submits invalid data: summary focused, links jump to fields; each invalid input has aria-invalid and aria-describedby pointing to text; NVDA speech log on field focus includes the error. Validation rules unchanged (unit tests).

Proposed partial credit

30% error association; 25% summary focus; 20% required/autocomplete; 25% NVDA announces errors

R02.07.01.03-T10 Remediate a tagged PDF in Acrobat Pro to pass PAC

Original instruction

input/annual_report.pdf (24 pages, untagged scan-free PDF) must pass PDF/UA. In Acrobat Pro, add tags, fix reading order of the 2-column pages, mark artifacts, add alt text to 9 figures (text in input/alt.csv), table headers on 4 tables and set document language and title.

Platforms
mac · macOS win · Windows
Applications
Adobe Acrobat Pro
Interfaces
desk · Desktop GUI
GUI need · task setting
G2 · GUI-essential
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
click · Point & click drag · Drag & drop vis · Visual judgment
Work product
WP01 · Document
Conditions
licensed · Licensed pro app long · Long-horizon
Proposed difficulty
4 / 5

Proposed verifier

PAC 2024 (CLI) on output/annual_report_ua.pdf: zero PDF/UA and WCAG failures. Tag-tree checks: figure alt equals CSV, tables have TH with scope, reading order of pages 4–9 matches reference. Visual content unchanged (page SSIM ≥0.99). Acrobat JavaScript and Action Wizard are disabled, no PDF libraries are installed, and the file must be saved by Acrobat Pro (Producer/XMP).

Proposed partial credit

35% PAC clean; 25% reading order; 20% alt text; 20% table headers

R02.07.01.03-T11 Write a WCAG 2.2 conformance report with evidence

Original instruction

Evaluate the sandbox web app 'Clinicly' (6 key flows) against the 50 WCAG 2.2 A/AA success criteria using automated tools plus manual keyboard and screen-reader checks. Fill output/acr.csv (criterion, conformance level: Supports/Partially/Does Not Support/N/A, evidence, location).

Platforms
mac · macOS win · Windows web · Any OS (web)
Applications
axe DevTools · NVDA · VoiceOver · Chrome
Interfaces
brow · Browser
GUI need · task setting
G2 · GUI-essential
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
key · Shortcuts vis · Visual judgment click · Point & click
Work product
WP01 · Document
Conditions
long · Long-horizon
Proposed difficulty
5 / 5

Proposed verifier

Grader compares per-criterion conformance to an expert key built from 14 seeded defects: accuracy ≥0.85 for pass; each 'Does Not Support' must cite a location matching a seeded defect. Evidence quality checked by LLM rubric (10%). axe-core CLI/npm is not installed on the VM, and NVDA/VoiceOver speech logs captured by the grader must show manual passes through all 6 flows.

Proposed partial credit

55% conformance accuracy; 25% defect locations cited; 10% N/A correctness; 10% evidence rubric

R02.07.01.03-T12 Add captions and accessible controls to a video player

Original instruction

The custom video player in fixtures/learn-player has icon-only controls and no captions. Add a WebVTT captions track from input/lecture.srt (convert timing), a captions toggle, keyboard shortcuts per the APG, accessible names and pressed states for all controls.

Platforms
mac · macOS win · Windows lin · Linux
Applications
Chrome · axe DevTools
Interfaces
brow · Browser term · Terminal
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
code · Code & commands av · Audio/video over time key · Shortcuts
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Grader: VTT cues match SRT text and times ±10 ms; toggling captions changes textTracks mode; all buttons named with aria-pressed where toggles; keyboard space/k/arrows behave; axe clean.

Proposed partial credit

30% captions correct; 20% toggle; 30% names/states; 20% keyboard

R02.07.01.03-T13 Make a news site reflow at 400% zoom

Original instruction

The sandbox news site needs horizontal scrolling at 320 CSS px width and breaks with WCAG text-spacing overrides. Fix the CSS so content reflows at 1280 px with 400% zoom, nothing is lost with text-spacing (line 1.5, paragraph 2, letter 0.12, word 0.16 em), and fixed headers don't cover focused content.

Platforms
mac · macOS win · Windows lin · Linux
Applications
Chrome
Interfaces
brow · Browser term · Terminal
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
HY · Hybrid
Interaction primitives
code · Code & commands vis · Visual judgment
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
3 / 5

Proposed verifier

Playwright at 320×256 viewport across 6 pages: scrollWidth ≤ 320 except data tables; text-spacing bookmarklet applied: no clipped text (scrollHeight vs clientHeight on text containers). Focused elements not obscured by the sticky header (bounding-box overlap).

Proposed partial credit

40% reflow; 30% text spacing; 20% focus not obscured; 10% tables scroll locally

R02.07.01.03-T14 Honor reduced-motion preferences across a landing page

Original instruction

The landing page in site/ has parallax, auto-playing carousel and scroll-triggered animations. With prefers-reduced-motion: reduce, disable parallax and non-essential animations, stop the carousel autoplay, and add a pause button for the carousel in all modes.

Platforms
mac · macOS win · Windows lin · Linux
Applications
Chrome
Interfaces
brow · Browser term · Terminal
GUI need · task setting
CL · CLI-sufficient
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
code · Code & commands
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
2 / 5

Proposed verifier

Playwright with reducedMotion 'reduce': no CSS animations/transitions longer than 0.01 s running, carousel index unchanged after 10 s, scroll position-linked transforms constant. Normal mode: pause button stops rotation and is labelled.

Proposed partial credit

40% reduced motion respected; 30% carousel autoplay stops; 30% pause button works and is named

R02.07.01.03-T15 Announce route changes and manage focus in a single-page app

Original instruction

In the React SPA fixtures/portal, client-side navigation leaves screen-reader users silent and focus on the old link. Add route-change announcements, move focus to each new page's H1, update document titles, and keep browser back/forward working.

Platforms
mac · macOS win · Windows web · Any OS (web)
Applications
NVDA · VoiceOver · Chrome
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
code · Code & commands key · Shortcuts
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
4 / 5

Proposed verifier

Grader drives NVDA via guidepup across 8 route changes: speech log contains the new page title within 1 s each time; document.activeElement is the H1; titles match spec; history navigation repeats the announcements.

Proposed partial credit

40% announcements; 30% focus on H1; 15% titles; 15% back/forward

R02.07.01.03-T16 Rebuild a combobox and date picker to ARIA APG patterns

Original instruction

Replace the custom autocomplete and date picker in fixtures/travel-form with implementations following the ARIA Authoring Practices combobox (list autocomplete) and date picker dialog patterns, including full keyboard support and screen-reader announcements, without third-party widget libraries.

Platforms
mac · macOS win · Windows web · Any OS (web)
Applications
Chrome · NVDA · VoiceOver
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
code · Code & commands key · Shortcuts
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
5 / 5

Proposed verifier

Hidden Playwright tests derived from the APG keyboard tables (31 interactions) plus NVDA speech-log checks for option count and selected date announcements. ARIA attributes validated (aria-activedescendant, aria-expanded, grid roles). No widget libraries in package.json.

Proposed partial credit

45% combobox interactions; 40% date picker interactions; 15% SR announcements

R02.07.01.03-T17 Raise a page's Lighthouse accessibility score to 100

Original instruction

The sandbox product page at localhost:4000/p/42 scores 71 in Lighthouse accessibility. Fix the source until Lighthouse reports 100 for accessibility on desktop and mobile, without removing content.

Platforms
mac · macOS win · Windows lin · Linux
Applications
Lighthouse · Chrome
Interfaces
brow · Browser term · Terminal
GUI need · task setting
CL · CLI-sufficient
GUI need · ordinary access
CL · CLI-sufficient
Interaction primitives
code · Code & commands
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
2 / 5

Proposed verifier

Grader runs Lighthouse CLI (desktop and mobile presets): accessibility 100 both. Content parity: visible text and image count equal to baseline; no aria-hidden on content regions.

Proposed partial credit

60% score 100 on both presets (partial by score); 40% content parity

R02.07.01.03-T18 Fix a PowerPoint deck's accessibility issues

Original instruction

input/onboarding.pptx (18 slides) fails PowerPoint's Accessibility Checker. Give every slide a unique title, add alt text from input/alt.csv, mark decorative images, fix reading order on 5 slides with the Selection/Reading Order pane, and correct 3 low-contrast text boxes using theme colors.

Platforms
mac · macOS win · Windows
Applications
PowerPoint
Interfaces
desk · Desktop GUI
GUI need · task setting
G2 · GUI-essential
GUI need · ordinary access
G2 · GUI-essential
Interaction primitives
click · Point & click drag · Drag & drop vis · Visual judgment
Work product
WP03 · Visual layout
Conditions
licensed · Licensed pro app
Proposed difficulty
3 / 5

Proposed verifier

Grader parses PPTX XML: unique titles; descr attributes equal CSV; decorative flag set; shape order on 5 slides equals reference reading order; text contrast ≥4.5:1 from theme colors. Slide renders SSIM ≥0.97 vs original except fixed text boxes. VBA macros and AppleScript are disabled, python-pptx is absent, and the file must be saved by PowerPoint (docProps/app.xml).

Proposed partial credit

20% titles; 25% alt/decorative; 30% reading order; 25% contrast and layout unchanged

R02.07.01.03-T19 Verify landmarks and headings with VoiceOver on iOS Safari

Original instruction

On the iOS Simulator, navigate the sandbox clinic site with VoiceOver using the rotor (headings, landmarks, form controls). Fix the site's markup so the rotor exposes the heading outline and landmarks in input/structure_spec.md and every form control has a label, then re-verify.

Platforms
mac · macOS ios · iOS/iPadOS
Applications
VoiceOver · iOS Simulator · Safari
Interfaces
mob · Mobile touch term · Terminal
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
HY · Hybrid
Interaction primitives
touch · Touch gestures code · Code & commands
Work product
WP06 · Code/engineering
Conditions
None declared
Proposed difficulty
4 / 5

Proposed verifier

Grader extracts the Safari accessibility tree via WebKit inspector protocol: heading levels/texts and landmarks match spec order exactly; form controls labelled. VoiceOver rotor traversal (XCUITest-driven) lists the same headings.

Proposed partial credit

40% heading outline; 30% landmarks; 20% labels; 10% rotor verification

R02.07.01.03-T20 Bring a legacy jQuery app to WCAG 2.2 AA across 12 pages

Original instruction

Remediate the legacy jQuery app fixtures/permits (12 pages: forms, tabs, accordions, data grid, file upload, modal workflows) to WCAG 2.2 AA. Use axe, keyboard testing and NVDA, and keep all existing Cypress functional tests green.

Platforms
win · Windows web · Any OS (web)
Applications
axe DevTools · NVDA · Chrome
Interfaces
brow · Browser
GUI need · task setting
HY · Hybrid
GUI need · ordinary access
HY · Hybrid
Interaction primitives
code · Code & commands key · Shortcuts vis · Visual judgment
Work product
WP06 · Code/engineering
Conditions
long · Long-horizon
Proposed difficulty
5 / 5

Proposed verifier

Grader: axe zero violations on all 12 pages; hidden keyboard and NVDA speech tests (60 checks: tab order, widget patterns, status messages, target size 24 px) pass rate; Cypress suite green. Visual diff SSIM ≥0.95 per page.

Proposed partial credit

30% axe clean; 45% keyboard/NVDA checks; 15% functional tests green; 10% visuals preserved

Mapped research references

No benchmark reference is assigned to this class in the current mapping.

In this workflow family