Accessibility Agents 5.4: Playwright Runtime Checks and CI Guard Rails

Overview

Version 5.4 shifts the reliability model from static analysis toward live runtime verification and tightens the CI safety net across the entire release lifecycle. Three independent guard rails now validate your workflows, config schemas, and documentation version pins on every push, and a new Playwright-powered runner catches high-severity accessibility failures at runtime before they reach users.

Highlights

Playwright High-Impact Checks

A new playwright-high-impact-check.mjs runner in mcp-server/scripts/ scans live pages at four viewport widths (320, 768, 1024, and 1440 px), applying full wcag2a/wcag2aa/wcag21aa/wcag22aa rule sets. Beyond axe-core, it adds:

• Keyboard trap heuristic: detects interactive elements that receive focus but cannot be exited with the keyboard.
• Overflow scan: flags content clipping that hides text or controls at small viewports.
• Touch target sweep: identifies tap targets smaller than 44x44 px across mobile viewports.
• Artifact output: emits playwright-a11y-results.json and playwright-a11y-report.md as workflow artifacts per run.

Triggered via .github/workflows/playwright-high-impact-check.yml on pushes to main and any feature/ branch, with manual dispatch for on-demand scans.

CI Integrity Guard Rails

Three validator scripts now run in CI on every push and PR via .github/workflows/ci-integrity-guards.yml:

All three are also bundled into scripts/release-readiness-check.mjs for a single-command pre-release sweep.

Office/PDF/EPUB JSON Schemas

Canonical JSON schemas are now checked into .github/schemas/ for all three scan config families. VS Code settings.json maps each schema to the corresponding config file type, providing red-squiggle validation and autocomplete in-editor without any extension install.

Branch Hygiene Reporting

.github/workflows/branch-hygiene-report.yml runs weekly and on manual dispatch to surface stale long-lived release branches, giving maintainers an at-a-glance view of what can be pruned.

New Documentation

• docs/guides/playwright-high-impact-checks.md - setup, configuration, and interpreting results.
• docs/guides/ci-integrity-guards.md - guard rail architecture, local commands, and failure remediation.
• docs/guides/release-communications-checklist.md - updated with CI integrity and branch hygiene verification steps.
• docs/tools/playwright-integration.md - full Playwright + axe-core integration reference.

Full Changelog

• CI integrity guard workflow (ci-integrity-guards.yml) with three independent validators.
• scripts/validate-workflow-invariants.mjs asserts CI job ordering and required steps.
• scripts/validate-config-integrity.mjs validates templates against local JSON schemas.
• scripts/validate-doc-version-pins.mjs detects stale version pins in documentation.
• scripts/release-readiness-check.mjs aggregates all three for pre-release checks.
• mcp-server/scripts/playwright-high-impact-check.mjs with multi-viewport, keyboard trap, overflow, and touch target scanning.
• .github/workflows/playwright-high-impact-check.yml CI trigger for Playwright checks.
• .github/workflows/branch-hygiene-report.yml weekly stale-branch reporting.
• .github/schemas/office-config.schema.json, pdf-config.schema.json, epub-config.schema.json.
• Office/PDF/EPUB template profiles updated with $schema references.
• .vscode/settings.json schema mappings for all three config file families.
• PR template release checklist expanded with version alignment and action tag freshness checks.
• docs/guides/playwright-high-impact-checks.md, ci-integrity-guards.md, release-communications-checklist.md.
• docs/tools/playwright-integration.md.
• Documentation refresh: README.md, docs/getting-started.md, docs/USER_GUIDE.md, prd.md.

Accessibility Agents 5.3: CI Reliability and Release Safety

Overview

Accessibility Agents 5.3 focuses on reliability hardening in CI and release operations for markdown accessibility workflows.

This release improves regression gating, schema validation visibility, release consistency enforcement, and documentation coverage for the new controls.

Highlights

Regression-Only Markdown Gating

markdown-a11y-lint.mjs now supports regression mode in CI and CLI:

• --regression
• --baseline-ref <git-ref>

When enabled, only markdown files changed since the baseline ref are scanned for gate decisions. If git diff is unavailable, the scanner safely falls back to full scan behavior.

Markdown Config Schema Validation

.a11y-markdown-config.json is now validated at load time.

Validation warnings are emitted for unknown keys and invalid field types, while keeping scans non-blocking.

Published Markdown Config JSON Schema

A dedicated schema is now included:

• .github/schemas/markdown-config.schema.json

This documents valid config structure and supports editor validation/intellisense.

CI and Code Scanning Improvements

a11y-check.yml now includes markdown SARIF upload to GitHub Code Scanning and supports reliable execution for clean/no-output scenarios.

The markdown lint job also supports:

• dispatch-time mode controls
• repository variable controls
• regression mode controls

Release Consistency Guard

New workflow:

• .github/workflows/release-consistency-guard.yml

It enforces:

• version alignment across release manifests
• required CHANGELOG.md entry for the current version

Missing CHANGELOG coverage now fails the workflow.

Test Coverage Expansion

Added and integrated:

• scripts/test-markdown-scanner.mjs (expanded coverage)
• scripts/test-orchestrator-validator.mjs

CI now runs these suites in validate-orchestrator-contracts.yml to catch regressions earlier.

Documentation and Editor Experience

Updated docs for 5.3 controls and behavior:

• docs/USER_GUIDE.md
• docs/getting-started.md
• prd.md
• ENHANCEMENTS.md

VS Code config now includes JSON schema mappings for markdown, office, and PDF scan config files.

Upgrade Notes

No migration is required for existing users.

Recommended follow-up:

1. Set repository variables for markdown gating defaults:
   • A11Y_MARKDOWN_FAIL_ON
   • A11Y_MARKDOWN_FORMAT
   • A11Y_REGRESSION_MODE
2. Add or refresh .a11y-markdown-config.json from templates/markdown-config-moderate.json
3. Verify release-consistency-guard.yml passes before tagging releases

Full Changelog

See CHANGELOG.md for full details of 5.3.0 changes.

Accessibility Agents 5.2: Calibrated Scoring, Configurable Gates, Stronger Contracts

Overview

Accessibility Agents 5.2 focuses on operational quality and reliability: configurable markdown accessibility gates, SARIF output for machine-readable CI, calibrated web severity scoring guidance, and contract validation between orchestrators and specialists.

This release intentionally excludes signing and key-management changes.

Highlights

Configurable Markdown Accessibility Scanning

markdown-a11y-lint.mjs now supports:

• repository config file: .a11y-markdown-config.json
• per-rule enabled and severity overrides
• custom ignored directories
• configurable per-rule output limits
• runtime gate mode controls (none, error, warning)

SARIF Output for CI Pipelines

Markdown accessibility findings can now be exported as SARIF:

node .github/scripts/markdown-a11y-lint.mjs . \
  --format both \
  --output artifacts/markdown-a11y.sarif

This enables downstream code-scanning and artifact workflows without custom parsers.

CI Gate Maturity Controls

a11y-check.yml now supports dispatch-time and variable-driven gate controls:

• workflow inputs: enforcement_mode, output_format
• repo variables: A11Y_MARKDOWN_FAIL_ON, A11Y_MARKDOWN_FORMAT

Teams can adopt advisory mode first and tighten enforcement later without editing workflow logic.

Orchestrator-Specialist Contract Validation

New validator:

• scripts/validate-orchestrator-dispatch.js

New workflow:

• .github/workflows/validate-orchestrator-contracts.yml

The validator enforces required dispatch sections, verifies Read(".claude/specialists/*.md") references, checks Task(...) usage, and confirms referenced specialists exist.

Web Severity Scoring v2 Guidance

web-severity-scoring now documents:

• profile-based scoring (balanced, strict, advisory)
• calibration coefficients by rule family
• confidence drift guardrails
• normalized trend scoring for cross-audit comparability
• recommended output metadata for reproducibility

New Metadata and Markup Conventions

A new guide standardizes metadata and markdown structure:

• docs/guides/metadata-markup-conventions.md

It includes recommended frontmatter metadata patterns for agents and skills and instruction markup conventions for automation-friendly docs.

Additional Docs and Templates

• Added template: templates/markdown-config-moderate.json
• Updated: docs/getting-started.md with markdown scanner and CI gate usage
• Updated: AGENTS.md with metadata conventions
• Updated: README.md docs index to include metadata and markup conventions guide

Why This Matters

5.2 improves how teams operate these agents at scale:

• less CI fragility through explicit configuration
• better interop through SARIF
• reduced orchestration drift via contract validation
• more reliable trend analysis through calibrated scoring guidance
• clearer machine-readable metadata standards for future automation

Full Changelog

See CHANGELOG.md for complete details.

Accessibility Agents 5.1: Leaner, Smarter, Faster on Claude Code

The Problem We Solved

If you were using Accessibility Agents on Claude Code, you were paying a hidden tax.

Every conversation turn, Claude Code serialized every registered agent definition into the schema it sends to the model. With 80 agents defined, that was 12,000 to 16,000 tokens per turn — before you typed a single word. On a typical audit session, that overhead added up fast.

Version 5.1 eliminates that tax entirely.

The Fix: On-Demand Specialists

The root cause was architectural: all 73 specialist agents were registered in the Claude Code schema, so they were always serialized — even when you never needed them.

The solution is the specialist dispatch pattern:

• 7 orchestrators remain registered in .claude/agents/ — they coordinate work and appear in Claude's schema (1,500 tokens, down from 16,000)
• 73 specialists move to .claude/specialists/ — they are loaded on demand, only when an orchestrator actually needs them

When an orchestrator needs a specialist, it does two things:

1. Read(".claude/specialists/contrast-master.md")    # load the specialist
2. Task(prompt="<specialist_body>\n\n<your context>") # invoke it as a subagent

That's it. The specialist runs with full capability, reads no more schema than it needs, and costs nothing until it is actually used.

Result: 89% reduction in per-turn schema token cost — from ~14,000 tokens to ~1,500 tokens.

What Else Is in 5.1

All 80 Copilot Agents Are Structurally Clean

Every .github/agents/ file now passes the full validation suite:

• All required tools: fields present
• No empty body sections
• No emoji in prose content
• All descriptions under 200 characters (per agentskills.io spec)
• All markdownlint rules satisfied

The pre-commit hook enforces this on every future commit — no regressions.

New Severity Mapping Skill

A new severity-mapping skill provides canonical severity level definitions shared across web, document, and markdown audits. Previously each audit domain defined its own severity thresholds. Now there is one source of truth, and cross-format severity normalization is consistent.

Gemini CLI Parity

All 25 GitHub Copilot skills are now synced to .gemini/extensions/a11y-agents/skills/. Gemini CLI users get the same knowledge modules as Copilot users, with no manual sync required.

New Agent Terminology Instruction File

agent-terminology.instructions.md applies to all .md and .agent.md files. It enforces consistent vocabulary — specialist vs. orchestrator, dispatch vs. invoke, loaded vs. registered — so agent documentation reads as a coherent whole rather than a patchwork of individual authors.

Upgrade Guide

For Claude Code Users

No action required. The specialist move is transparent — orchestrators handle dispatch automatically. If you have custom agents that import from .claude/agents/, update those paths to .claude/specialists/ for the moved files.

For Copilot Users

No changes. .github/agents/ is unchanged in functionality; only the internal structure was cleaned up.

For Gemini CLI Users

Pull the latest .gemini/extensions/ directory. The skills/ subfolder is new in this release.

Numbers

Full Changelog

See CHANGELOG.md for the complete list of changes in this release.

What's Next

The 5.2 roadmap focuses on:

• Playwright-based behavioral scanning fully wired into the web-accessibility-wizard audit loop
• ePub accessibility specialist reaching parity with Word/Excel/PowerPoint specialists
• Codex CLI role definitions aligned with the .claude/specialists/ on-demand pattern
• VS Code extension (planned in 5.0 roadmap) — scaffolding begins

See ROADMAP.md for details.

This release reworks how Accessibility Agents ships for Codex.

We spent time trying to make the Codex plugin and marketplace path work as the primary distribution model. That research was useful, but the end result was not good enough for a release we would want people to depend on. In testing, local plugin packaging could work on a developer machine and Codex could cache an installed plugin, but discovery and install behavior for third-party plugins was still less predictable than we wanted. Too often the experience still depended on the Codex app UI, extra confirmation steps, or behavior that felt different from the simpler install story people are used to on other platforms.

Because of that, this release moves Codex support to a direct skills-pack model. The repo now ships a full codex-skills/ catalog with all 80 Accessibility Agents skills, and the Codex installer now copies those skills directly into .codex/skills/ for project installs or ~/.codex/skills/ for global installs. That gives us a distribution path that is easier to explain, easier to verify, and much more consistent with how people actually use Codex day to day.

We also removed the old .codex/AGENTS.md baseline from the Codex path. Experimental TOML roles are still available for focused passes, but they are now clearly positioned as optional. The primary Codex story is straightforward: install the skills, restart Codex, and use the agents directly.

What changed in 4.5.1:

• Added a full direct Codex skills pack under codex-skills/ with all 80 Accessibility Agents skills.
• Updated install.sh and install.ps1 so Codex installs copy skills directly into .codex/skills/ or ~/.codex/skills/.
• Removed the old repo-local Codex plugin and marketplace path from the primary distribution flow.
• Removed the obsolete .codex/AGENTS.md baseline from the Codex setup.
• Updated Codex docs to explain the plugin research, what worked, what did not hold up, and why direct skills are now the recommended path.
• Bumped release metadata to 4.5.1 across the changelog and release manifests.

Verification:

• Release version alignment passes across CHANGELOG.md, plugin.yaml, mcp-server/package.json, and gemini-extension.json.
• install.sh passes shell validation.
• The codex-skills/ pack validates with 80 SKILL.md files and required description frontmatter.

If you are already using Accessibility Agents with Codex, the upgrade path is simple: reinstall with the Codex flag and restart Codex so it reloads the installed skills.

Accessibility Agents 4.5.0

Released: March 27, 2026 | 232 files changed | 96 agents updated across 3 platforms

What's New

4.5.0 is our biggest quality-of-life release yet. The installer now guides you through MCP server setup automatically. Document audit reports speak in plain language instead of XML schemas. The entire agent suite is aligned to VS Code 1.113. And the installation, update, and uninstall tooling has been comprehensively rebuilt for reliability.

Guided MCP Server Installation

Before 4.5.0: You installed the agents, then had to manually navigate to the MCP server directory, figure out which npm packages you needed, configure VS Code settings.json by hand, and hope it all worked.

Now: The installer walks you through everything.

Interactive Capability Planner

When the installer detects Node.js, it presents a guided setup menu:

1. Baseline scanning - Core accessibility scanning tools
2. Browser testing - Adds Playwright and axe-core for live page testing
3. PDF-heavy workflow - Adds pdf-lib for form conversion
4. Everything - All of the above plus deep PDF validation prerequisites
5. Custom - Pick exactly what you need

Automatic Dependency Management

• Node.js detection and installation - If Node.js 18+ is not found, the installer offers to install it via winget (Windows) or brew (macOS)
• npm dependencies - Core MCP server packages installed automatically (npm install --omit=dev)
• Playwright and Chromium - When browser testing is selected, Playwright is installed and Chromium is downloaded (npx playwright install chromium)
• pdf-lib - Installed when PDF form capabilities are selected
• Java and veraPDF - For deep PDF/UA validation, the installer offers to set up Java 21 JRE and veraPDF through platform package managers

VS Code Settings Auto-Configuration

The installer automatically merges the MCP server entry into your VS Code settings.json - for both Stable and Insiders editions if you have them. No manual JSON editing required.

Health Check and Readiness Dashboard

At the end of installation, you see a readiness report:

MCP Capability Readiness
  [x] Node.js 18+
  [x] npm
  [x] MCP core dependencies
  [x] Baseline PDF scan
  [!] Deep PDF validation (Java found, veraPDF missing)
  [x] Local health smoke test - READY
  [x] Playwright
  [x] Chromium
  [x] pdf-lib

The installer actually starts the MCP server on a test port and hits /health to confirm everything works before declaring success.

Completely Rebuilt Installation Tooling

The install, update, and uninstall scripts for both PowerShell and Bash have been rewritten from the ground up with a shared library architecture.

New Command-Line Flags

Every script now supports these flags (both install.ps1/install.sh, update.ps1/update.sh, and uninstall.ps1/uninstall.sh):

The following table lists flags available across all installer operations.

The installer adds additional selection flags:

The following table lists installer-specific selection flags.

Machine-Readable Output

Every operation now writes a structured JSON summary file (.a11y-agent-team-install-summary.json, -update-summary.json, or -uninstall-summary.json) containing the full record of what was installed, where, which VS Code profiles were targeted, and what the MCP configuration state is. CI pipelines can parse these for automated validation.

Shared Installer Libraries

Two new shared modules (scripts/Installer.Common.ps1 and scripts/installer-common.sh) provide consistent cross-platform behavior for VS Code profile detection, JSON summary writing, directory operations, and backup metadata. This eliminated duplicated logic and the class of bugs where PowerShell and Bash installers behaved differently.

Clean Uninstallation

The uninstaller now removes the MCP server entry from VS Code settings.json automatically - including cleanup of empty mcp.servers objects. Previously you had to edit settings.json by hand after uninstalling.

Document Audit Reports Now Speak Plain Language

This is a fundamental change in how document accessibility findings are presented.

Native-Tool-First Remediation

Every finding across all document agents - Word, Excel, PowerPoint, PDF, and the CSV export - now follows a new three-part structure:

Start Here - tells the user exactly what to do in the native application (Word ribbon path, Excel menu, PowerPoint selection pane, Acrobat Pro tool panel). This is the first thing readers see.

Why It Matters - explains the accessibility impact in plain language.

Advanced / Technical Follow-Up - XML structures, scripting approaches, and programmatic fixes for power users who need them.

What Changed

• document-accessibility-wizard - New Remediation Writing Standard, new Start Here and Native App Action Plan sections in the report template
• word-accessibility - All findings restructured with Start Here leading to Word ribbon paths
• excel-accessibility - All findings restructured with Start Here leading to Excel menu paths
• powerpoint-accessibility - All findings restructured with Start Here leading to Selection Pane and slide layout paths
• pdf-accessibility - All findings restructured with Start Here leading to Acrobat Pro tool paths, with specific note about Acrobat triage before source rebuilds
• document-csv-reporter - CSV fix_suggestion column now leads with native-app action in the first sentence

Why This Matters

Most people who receive a document accessibility audit report are content authors, not developers. They know how to use Word and PowerPoint. They should not have to parse XML schemas to fix a missing alt text description. Now they do not have to.

96 Agent Definitions Updated Across 3 Platforms

Every agent change was synchronized across GitHub Copilot (.github/agents/), Claude Code (.claude/agents/), and the Claude Code plugin (claude-code-plugin/agents/).

Coordinator Safety

All coordinator agents (accessibility-lead, web-accessibility-wizard, document-accessibility-wizard, and others) now declare an explicit agents: list in their YAML frontmatter specifying exactly which specialist agents they are allowed to invoke. The validator enforces this at commit time.

This prevents the "wrong agent called" problem and makes the orchestration architecture visible and auditable in code review.

Platform Focus: Windows and macOS

4.5.0 removes Linux desktop accessibility references across the agent suite. This affects:

• Desktop accessibility specialist - ATK/AT-SPI API section removed
• Desktop a11y testing coach - Orca screen reader section removed
• Testing coach - Orca references removed
• Screen reader lab - Orca simulation mode removed
• Python specialist - Linux column removed from cross-platform reference tables
• wxPython specialist - Linux-specific guidance removed
• All installer scripts - Linux cron scheduling removed; shell auto-update now macOS-only
• All documentation - "macOS/Linux" changed to "macOS" throughout

The agents, installers, and documentation now explicitly target Windows and macOS. Linux users can still install and use the agents for web and document accessibility, but desktop accessibility testing coverage (screen reader APIs, platform automation) is Windows and macOS only.

Markdown Rendering Quality

Approximately 40 agent files received systematic formatting fixes:

• Blank lines added before bulleted lists for correct markdown spec rendering
• Code fences changed from bare triple-backtick to language-tagged blocks (text, yaml, json) so output is not randomly syntax-highlighted
• Trailing newlines normalized

These changes improve rendering quality in GitHub, VS Code preview, and documentation site builds.

MCP Server Documentation Overhaul

New: PDF Quick Start Guide

New standalone guide (PDF-QUICKSTART.md) for users who only need PDF scanning. Six steps from zero to working PDF accessibility scan, without touching the full agent suite.

Expanded MCP README

The MCP server README was substantially rewritten:

• "What Works Out of the Box" matrix - shows which agents need which MCP capabilities
• Prerequisite matrix - 9-row table showing what is required vs. optional
• Local vs. shared server comparison - helps teams decide deployment topology
• Actionable veraPDF setup - platform-specific install commands for Windows (winget/choco), macOS (brew), and manual download
• Verification examples - how to confirm each prerequisite is working

VS Code 1.113 Alignment

MCP Across Agent Types

MCP servers registered in VS ...

v4.10.0 — Enhanced Validator, Pre-commit Hook & Agent Fixes

This release brings a comprehensive overhaul of the agent validation system, a pre-commit hook to catch issues before they land, and fixes for 38 issues found across the agent fleet.

Highlights

Enhanced Validator

The validate-agents.js script is now a full schema validator built from three official sources:

• Official tool alias table from the GitHub custom-agents configuration reference
• VS Code qualified tools from the Copilot features reference
• MCP namespace patterns for <server>/* and <server>/<tool> validation

New checks include:

• Frontmatter property validation against the official schema
• Deprecated infer: detection with migration guidance
• Prompt body size enforcement (30K GitHub.com limit)
• Duplicate tool detection
• Claude Code tool validation (MCP(...) syntax)
• --strict, --quiet, --files CLI flags

Pre-commit Hook

New scripts/pre-commit hook validates only staged agent/skill files before every commit. Runs in --strict mode to block commits with validation warnings.

�ash node scripts/install-hooks.js # one-time setup

38 Agent Fixes

Other Changes

• MCP server dependency auto-install in install.ps1 / install.sh
• MCP setup guide added to docs/getting-started.md
• CI workflow runs --strict and triggers on claude-code-plugin/agents/**
• Broken link repairs from v4.0.0 follow-up

Upgrade

�ash git pull node scripts/install-hooks.js # install pre-commit hook

Or use the installer:

`powershell

PowerShell

./update.ps1
`

`�ash

Bash

./update.sh
`

Full changelog: v4.0.0...v4.10.0

Accessibility Agents v4.0.0

80 agents. 5 platforms. 25 skills. 134 prompts. 24 MCP tools. One mission: accessible software for everyone.

What's New

21 New Specialist Agents

7 New Knowledge Skills (25 total)

• CI Integration - axe-core CLI reference, baseline schema, CI/CD templates
• Testing Strategy - Automated vs manual testing coverage matrix
• Legal Compliance Mapping - Section 508, ADA, EN 301 549, EAA, AODA
• Email Accessibility - Email client rendering constraints
• Media Accessibility - Caption formats, audio description, WCAG 1.2.x
• Data Visualization - Chart patterns, SVG ARIA, color-safe palettes
• Office Remediation - OOXML patterns, python-docx/openpyxl/python-pptx

3 New Always-On Instructions (9 total)

• CSS Accessibility - Focus visibility, motion safety, high contrast
• Testing Accessibility - Accessibility assertions in test files
• Document Generation - Enforces accessibility metadata in doc libraries

4 New MCP Tools (24 total)

• fix_document_metadata - Fix title, language, author in Office documents
• fix_document_headings - Report heading structure issues in .docx files
• check_audit_cache - Check for changed files since last scan
• update_audit_cache - Save scan results for incremental scanning

23 New Prompts (134 total)

Including: component library audit, training scenarios, native app audit, web CI/CD setup, PR accessibility check, team onboarding, email template audit, media content audit, accessibility dashboard, statement generator, and more.

Comprehensive User Guide

New instructor-style User Guide covering all 80 agents, 25 skills, 134 prompts, 9 instructions, and 24 MCP tools.

Cross-Platform Sync

All 80 agents available on every supported platform:

• GitHub Copilot (VS Code + CLI) - 80 agents, 25 skills, 134 prompts, 9 instructions
• Claude Code - 80 agents with 3-hook enforcement system
• Claude Code Plugin - 80 agents
• Gemini CLI - 98 skills (80 agent + 18 knowledge)
• Codex CLI - 11 roles

Infrastructure

• HTTP-based MCP Server with Streamable HTTP + stdio
• Enterprise configuration schema
• Accessibility PR Gate CI workflow
• 52-test MCP server test suite
• Anthropic directory manifest for Claude Desktop

Bug Fixes

• Fixed fix_document_headings reading .docx ZIP as raw UTF-8
• 52/52 MCP server tests passing
• Full markdownlint compliance
• Fixed installer manifest initialization order
• Fixed wxPython SetName() screen reader issues
• Fixed USER_GUIDE coverage gaps

Upgrade

cd accessibility-agents
git pull origin main
bash update.sh   # or: powershell -File update.ps1

Full changelog: https://github.com/Community-Access/accessibility-agents/blob/main/CHANGELOG.md

Accessibility Agents v3.2.0

Compare: v3.0.0...v3.2.0

Release Scope (v3.0.0 -> v3.2.0)

• 238 files changed
• 7,968 insertions, 1,473 deletions

Highlights

1) Playwright Integration

• Added 5 Playwright MCP tools for behavioral accessibility scanning (keyboard traversal, dynamic state, viewport, contrast, a11y tree)
• New playwright-scanner agent for runtime accessibility verification
• New playwright-verifier agent for post-fix confirmation scans
• New playwright-testing skill with reusable scanning patterns
• New generate-a11y-tests prompt with CI template for automated test generation
• Integrated Playwright workflows into web-accessibility-wizard and web-issue-fixer

2) veraPDF PDF/UA Validator

• Added veraPDF integration for PDF/UA-1 and PDF/UA-2 validation
• New MCP tools for PDF standards compliance checking
• Integrated into document-accessibility-wizard scanning pipeline

3) PDF Form-to-HTML Converter

• Added PDF form extraction and accessible HTML conversion tool
• Preserves form field labels, grouping, and ARIA attributes in output

4) askQuestions Integration

• Added askQuestions tool integration across all 59 agents
• Enables agents to request clarification from users during complex audits
• Consistent integration pattern across the entire agent ecosystem

5) Plugin Distribution Fixes

• Fixed plugin symlink drift for docs, templates, and example directories (#57)
• Added .gitattributes for consistent line endings and symlink handling
• Updated CONTRIBUTING.md with Windows symlink documentation

6) NVDA Addon Specialist Updates

• Updated NVDA 2026.1 architecture transition documentation (#62)
• Corrected minimum addon version guidance across all 4 platform files
• Fixed 12 files with inaccurate version references

7) Codex Multi-Agent Support

• Added experimental multi-agent role support for Codex CLI (#59)

8) Additional Improvements

• Automated agent manifest generation workflow (#70)
• Gemini CLI hooks support with five lifecycle hook scripts
• Improved Copilot agent configuration with missing agents and updated WCAG references
• Fixed broken URLs and migrated Deque help links to Accessibility Insights
• Repository-wide accuracy and authority documentation refresh

Version Updates

Asset Counts

• 59 specialist agents
• 106 custom prompts
• 18 reusable skills
• 6 workspace instructions
• 338 manifest entries

Thanks to all contributors helping make accessibility workflows more reliable and easier to adopt.

Accessibility Agents v3.0.0

This release captures all changes from v2.5 to v3.0.0.

Compare: v2.5...v3.0.0

Release Scope (v2.5 -> v3.0.0)

• 375 files changed
• 17,825 insertions, 283 deletions
• Major delivery areas: lifecycle hooks, VS Code 1.110 integration, citation authority, browser verification, plugin packaging, and release management docs

Highlights

1) Cross-Platform Lifecycle Hooks (Phase 4)

• Added Python hook scripts in .github/hooks/scripts/ for SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, and SessionEnd
• Added consolidated hook configs for both Copilot and Claude flows:
  • .github/hooks/hooks-consolidated.jsonn - .claude/hooks/hooks-consolidated.jsonn- Implemented edit-gate enforcement for web UI files until accessibility review is completed
• Added setup and troubleshooting docs:
  • docs/guides/hooks-guide.mdn - docs/guides/hooks-troubleshooting.mdn - docs/HOOKS-CROSS-PLATFORM-STRATEGY.md`n

2) VS Code 1.110 Integration (Phases 5 and 6)

• Added Agent Debug Panel guidance and troubleshooting:
  • docs/guides/agent-debug-panel.mdn - updates in README.md, docs/hooks-guide.md, and docs/getting-started.mdn- Added session forking (/fork) guidance to orchestrator workflows
• Added getDiagnostics integration to specialist agents for lint-aware accessibility audits
• Added docs for:
  • OS notifications during long audits
  • AI co-author attribution (git.addAICoAuthor)
  • inline chat continuity
  • collapsible terminal tool output
  • custom thinking phrases
• Added built-in VS Code accessibility skill comparison:
  • docs/guides/vscode-builtin-skill-comparison.md`n

3) Citation Authority and Source Currency (Phase 1B)

• Added authoritative source citations across the agent ecosystem
• Added source registry and automated verification workflow:
  • .github/agents/SOURCE_REGISTRY.jsonn - .github/workflows/verify-sources.ymln

4) Agentic Browser Verification (Phase 3)

• Added browser-assisted validation design and testing guides:
  • docs/AGENTIC-BROWSER-TOOLS.mdn - docs/BROWSER-TOOLS-TESTING.mdn- Expanded remediation workflows in web-accessibility-wizard and web-issue-fixer`n

5) Marketplace and Prompt Ecosystem Expansion

• Added marketplace packaging metadata and submission support
• Added large prompt coverage expansion under .github/prompts/ for agent-aligned workflows

Notable Documentation Updates

• CHANGELOG.md now includes complete phased v3.0 history
• prd.md updated with completed phase status and release management tracking
• README.md expanded with current system requirements and troubleshooting
• CONTRIBUTING.md updated with version-currency testing requirements

Migration Notes from v2.5

• If you are on v2.5, update to latest toolchain versions first, then pull v3.0.0
• Hook-based enforcement is now part of normal web UI workflows
• Use the Agent Debug Panel in VS Code 1.110+ for visibility into hooks, tools, and agent loading

Full Commit Range

Recent high-impact commits in this range include:

• de47245 v3.0.0 release integration
• 77a92e8 consolidation into prd.md and migration audit
• ce4179a marketplace packaging and submission prep
• 4dfbdd9 autonomous browser verification implementation
• �511e08 context compaction and custom-skills guidance
• ac9d49 authoritative source citations rollout

Thanks to all contributors helping make accessibility workflows more reliable and easier to adopt.