Audit and fix WCAG accessibility issues in web apps

Audits a codebase, page, or rendered DOM against WCAG 2.2 and either reports the findings or fixes them, depending on which mode the request implies. In report mode it sweeps the target, groups violations by rule ID and component family (rather than listing the same issue N times), prioritizes critical and serious issues first, and produces a structured written report with per-pattern fix guidance, WCAG rule IDs, and affected file:line locations - without touching any files. In fix mode it runs an audit -> edit -> verify loop: baseline audit, apply the engine's authoritative Fix: directive verbatim for mechanically fixable violations, leave a TODO with the rule ID for anything contextual or visual rather than inventing content, then re-run the audit and confirm the diff shows the targeted violations fixed (-fixed) and nothing new introduced (+new).

It picks the audit method in a fixed order of preference: audit_live first for any URL (auto-connects to or launches a Chrome debug session, no user setup needed), then an audit-live-page prompt via a connected browser MCP (chrome-devtools-mcp, playwright-mcp, or puppeteer-mcp) if the target needs an authenticated or specific browser state, and audit_html for raw HTML, files, or rendered JSX strings when there's no URL to audit live. Live-DOM audits against React dev builds attach a Source: <file>:<line> (Symbol) pointer per violation via React DevTools fibers; when that's absent it falls back to stable hooks, then visible text, then tree position to locate the offending element.

Use report mode for requests like "audit my codebase," "review src/components/," or "give me an a11y report" - it audits and writes the report but does not edit anything. Use fix mode for "fix the a11y issues in X," "audit and fix," "make this accessible," or when handed an existing violation report to apply - it audits, edits, and verifies. If the user's intent is unclear, it asks rather than defaulting to fix mode.

It stops and asks the user to narrow scope rather than sweeping an entire codebase with no arguments given, stops and asks if a single audit returns more than roughly 50 violations (not actionable as a report), and in fix mode confirms scope with the user before touching files outside the obvious target or before applying more than about 10 mechanical fixes. It also declines to guess: any violation without a Fix: directive gets a TODO, not an invented fix, and any failed verification is named and stopped on rather than iterated through silently.

Report-mode input is a scope: a directory path, a list of files (plus their imports), a URL, or nothing (which triggers a scope-narrowing question). Fix-mode input is a target plus, optionally, an existing violation report to apply. Output in report mode is a structured markdown report: a summary with critical/serious/moderate/minor counts after deduplication, per-pattern critical and serious sections (pattern, WCAG ID and name, affected files, fix, and user-impact rationale), a deduplicated bullet list for moderate/minor issues, architectural recommendations, what still needs manual review (screen reader, keyboard, low-vision testing), and positive findings. Output in fix mode is, per cycle: which audit flow was used, violations by impact, what was applied (file plus rule), what was deferred as TODOs with reasons, and the final diff.

Prefers direct-CDP live-DOM auditing via audit_live, with chrome-devtools-mcp, playwright-mcp, or puppeteer-mcp as a fallback browser-MCP composition for authenticated or stateful pages, and raw audit_html plus audit_diff for HTML strings or files with no live DOM available. Can be invoked as a Task (general-purpose agent) for context isolation on very large sweeps, and calls explain_rule({ id: "<rule-id>" }) for guidance when a rule is unclear.

Developers and teams who need WCAG 2.2 issues found and prioritized (report mode) or found, mechanically fixed, and verified (fix mode) - without a human having to manually trawl a page or codebase rule by rule, and without the fix loop silently guessing at content it shouldn't invent.