Skip to content

RELEASE 7.0.0

github-actions[bot] edited this page Oct 29, 2026 · 2 revisions

Accessibility Agents 7.0: One Package on the Open Agent Standards

Covers 7.0.0 and its three patch releases, 7.0.1, 7.0.2 and 7.0.3, all from September 2026. None of them was tagged as a GitHub release; 7.1.0 is the first tagged release of the 7.x line. For what to install today, see RELEASE-7.1.0.md.

Overview

Accessibility Agents 7.0 replaces six hand-maintained, per-client trees with one package of Agent Skills in skills/. Claude Code, Codex, Copilot, Gemini CLI and Antigravity all read that one copy, so a fix lands everywhere at once.

The package is 108 skills in four tiers. Six routers are the only skills the model sees until it needs more; they dispatch 57 specialists, 16 helpers and 29 reference skills by name, not by pasting instructions. Always-on context cost fell by 84 to 93 percent depending on the client, and dispatching a specialist fell from 26,929 tokens to 100.

Highlights

One package, read natively by every client

  • 108 skills conforming to the Agent Skills specification.
  • AGENTS.md is the only always-on instruction file; CLAUDE.md is one line that imports it.
  • 843 per-client copies removed, including .claude/agents, codex-plugin, .gemini, .github/agents and .github/prompts.

Findings mean the same thing whichever skill produced them

  • One findings schema (skills/a11y-core/schemas/findings.schema.json) for every scanning skill.
  • One report renderer (render-report.mjs) that writes all seven required sections, a score and grade, and a delta against the previous run.
  • A WCAG 2.2 criterion list that every finding is checked against, so an audit cannot cite a criterion that does not exist, claim the wrong level, or cite the obsolete 4.1.1 Parsing as a requirement.

One enforcement gate

  • hooks/guard.mjs replaced four scripts in three languages, with a manifest per client.
  • The edit gate opens when the lead completes, not when it starts, so a cancelled review no longer unlocks every edit.
  • The reminder is injected once per session, capped at sixty words.

An MCP server you can auto-approve

  • Annotations and output schemas on all 39 tools; the 29 read-only scanners can be auto-approved, so a bulk scan stops asking once per document.
  • A render_accessibility_report tool, so clients without a shell produce the same reports.

Verification you can run

  • npm run verify runs every gate, each naming the standard it checks and what a failure would cost a person. Context budgets are enforced in CI.
  • A new installer, scripts/install.mjs, 8 KB replacing 215 KB of shell and PowerShell. It detects installed clients and will not overwrite a hooks file or a skill you edited.

Patch releases

  • 7.0.1 closed 24 high-severity dependency advisories and added the dependency gate that should have caught them.
  • 7.0.2 came from the first end-to-end audit: findings from different specialists about the same defect now merge, so one problem is reported once instead of eight times.
  • 7.0.3 fixed every MCP tool failing on invocation with an output validation error (#205), and added an end-to-end tools/call gate.

Upgrading from 6.x

The 6.x installers, per-client agent trees and .a11y-agent-manifest are gone. Remove a 6.x install with the steps in UNINSTALL.md ("Pre-7.0 installs"), then install 7.x as described in RELEASE-7.1.0.md.

Known issues fixed later

7.0 shipped with problems that the October 2026 audit found and 7.1.0 fixed: plugin installs loaded no hooks or MCP server, the edit gate deadlocked on Claude Code, the Codex hooks file was rejected by Codex, and run_axe_scan never ran. Upgrade to 7.1.0.

Full changelog

See the 7.0.0 to 7.0.3 entries in CHANGELOG.md and the full account in the 7.0 modernization.

Clone this wiki locally