keyboard-a11y-tester

Projects that follow the best practices below can voluntarily self-certify and show that they've achieved an Open Source Security Foundation (OpenSSF) best practices badge.

There is no set of practices that can guarantee that software will never have defects or vulnerabilities; even formal methods can fail if the specifications or assumptions are wrong. Nor is there any set of practices that can guarantee that a project will sustain a healthy and well-functioning development community. However, following best practices can help improve the results of projects. For example, some practices enable multi-person review before release, which can both help find otherwise hard-to-find technical vulnerabilities and help build trust and a desire for repeated interaction among developers from different companies. To earn a badge, all MUST and MUST NOT criteria must be met, all SHOULD criteria must be met OR be unmet with justification, and all SUGGESTED criteria must be met OR unmet (we want them considered at least). If you want to enter justification text as a generic comment, instead of being a rationale that the situation is acceptable, start the text block with '//' followed by a space. Feedback is welcome via the GitHub site as issues or pull requests There is also a mailing list for general discussion.

We gladly provide the information in several locales, however, if there is any conflict or inconsistency between the translations, the English version is the authoritative version.
If this is your project, please show your badge status on your project page! The badge status looks like this: Badge level for project 13561 is passing Here is how to embed it:
You can show your badge status by embedding this in your markdown file:
[![OpenSSF Best Practices](https://www.bestpractices.dev/projects/13561/badge)](https://www.bestpractices.dev/projects/13561)
or by embedding this in your HTML:
<a href="https://www.bestpractices.dev/projects/13561"><img src="https://www.bestpractices.dev/projects/13561/badge"></a>


These are the Gold level criteria. You can also view the Passing or Silver level criteria.

Baseline Series: Baseline Level 1 Baseline Level 2 Baseline Level 3

        

 Basics 0/5 ●

  • General

    Note that other projects may use the same name.

    An AI-assisted web accessibility tester that behaves like two W3C personas at once: a keyboard-only user ("Ade") and a screen-reader user ("Lakshmi"). It drives a page keyboard-only, records what happens at every focus stop, and emits evidence-linked findings mapped to specific WCAG success criteria — against any website.

    Please use SPDX license expression format; examples include "Apache-2.0", "BSD-2-Clause", "BSD-3-Clause", "GPL-2.0+", "LGPL-3.0+", "MIT", and "(BSD-2-Clause OR Ruby)". Do not include single quotes or double quotes.
    If there is more than one language, list them as comma-separated values (spaces optional) and sort them from most to least used. If there is a long list, please list at least the first three most common ones. If there is no language (e.g., this is a documentation-only or test-only project), use the single character "-". Please use a conventional capitalization for each language, e.g., "JavaScript".
    The Common Platform Enumeration (CPE) is a structured naming scheme for information technology systems, software, and packages. It is used in a number of systems and databases when reporting vulnerabilities.

    keyboard-a11y-tester

    An AI-assisted web accessibility tester that behaves like two W3C personas at once: a
    keyboard-only user ("Ade") and a screen-reader user ("Lakshmi"). It drives a page
    keyboard-only, records what happens at every focus stop, and emits evidence-linked
    findings mapped to specific WCAG success criteria — against any website. Both
    personas run in the same pass by default; a --persona flag restricts to just one.

    It has two layers:

    • a deterministic runner (scripts/runner.mjs) that owns the mechanical, reproducible
      work — driving the page keyboard-only, capturing a per-step trace + screenshots, and
      computing the machine-decidable checks for both personas;
    • an AI-judgment layer — the invoking agent — that reads the trace/screenshots/census
      and judges what rules can't (task completion, logical focus/reading order, form
      quality, announcement quality). See SKILL.md for how an agent drives it.

    Standalone and portable: it depends only on playwright, yaml, pngjs, pixelmatch,
    and @guidepup/virtual-screen-reader, needs no bundled test cases, and writes all output
    to a per-user temp directory (never into this folder). The screen-reader persona
    never drives a real screen reader (NVDA/VoiceOver) — see "Screen-reader detection" below.

    Documentation: docs/usage.md (setup, dependencies, quick start,
    CAPTCHAs) · docs/interface.md (full CLI reference, output file
    schema, WCAG checks table).

    Quick start

    As a Claude Code plugin — register this repo as a plugin marketplace, then install it:

    /plugin marketplace add ezufelt/keyboard-a11y-tester
/plugin install keyboard-a11y-tester@ezufelt

    The first command registers this repo as a marketplace (named ezufelt, per
    .claude-plugin/marketplace.json); the second installs the plugin. Once installed, the
    skill in SKILL.md becomes available to the agent.

    As a standalone clone — clone the repo and install its dependencies:

    git clone https://github.com/ezufelt/keyboard-a11y-tester.git
cd keyboard-a11y-tester
npm install
npx playwright install chromium

    Then drive it directly (see Run against any URL).

    Requirements & dependencies

    Requires Node.js ≥ 20 and Chromium (via Playwright), plus five small npm dependencies —
    no build step. Run node scripts/setup-check.mjs to verify both before your first run.

    See docs/usage.md for the full dependency
    table, licensing credit for @guidepup/virtual-screen-reader, and setup instructions.

    Run against any URL (no test file needed)

    # quick unattended blind Tab-crawl of the start page, per viewport
node scripts/runner.mjs --url https://example.com

# a full scenario, driven live by the agent one keystroke at a time
node scripts/runner.mjs serve --url https://example.com --goal "find the pricing page" \

     --viewport desktop --port 9400
#   → prints:  READY <session-dir>   (under the system temp dir)
node scripts/runner.mjs observe <session-dir>
node scripts/runner.mjs step    <session-dir> --press Tab      # one keystroke; prints observation
node scripts/runner.mjs step    <session-dir> --press Enter
node scripts/runner.mjs step    <session-dir> --type "hello@example.com"
node scripts/runner.mjs finish  <session-dir>                  # writes trace + findings
node scripts/runner.mjs stop    <session-dir>

    See docs/usage.md for the full
    quick-start walkthrough, and docs/interface.md for every CLI flag and
    the complete output file schema.

    Authenticated runs

    Pages behind a login can't be tested with a fresh, logged-out browser. Pass a Playwright
    storageState JSON file with --storage-state <file> to start the browser with its cookies
    and localStorage already loaded (e.g. an already-logged-in session). Generate one with
    context.storageState({ path: 'auth.json' }) or npx playwright codegen --save-storage=auth.json <url>.
    The file is validated (exists, parses as JSON, and looks like a real storageState export —
    i.e. has cookies/origins arrays) before the browser launches — a missing or malformed file
    fails the run immediately rather than silently testing the logged-out site. In serve mode
    it's applied once at launch and the session browser keeps the state alive for every subsequent
    step.

    A storageState file holds live session cookies/tokens — treat it as a secret. Don't commit
    it; .gitignore already excludes auth.json, storageState.json, and *storage-state*.json,
    but a differently-named file won't be caught automatically.

    What the runner does (deterministic layer)

    Playwright (full Chromium, new-headless + SwiftShader for real pixels) drives the page with
    only the keyboard — it never calls .click() or .focus(); if a control is only
    reachable by pointer, that is itself a finding. It drops to a raw CDP session for the
    accessibility tree (Accessibility.getPartialAXTree), the ground truth for name/role/state.
    At startup it fails fast if :focus-visible does not fire on CDP-driven key events
    (every focus-indicator check would otherwise be invalid) — skipped entirely when
    --persona screen-reader is passed, since that persona has no pixel/focus-ring work.

    Checks are evaluated per focus stop the persona actually visits (keyboard persona) or
    against a page-wide structural census (screen-reader persona) — this is scenario
    testing, not an exhaustive page audit. Conformance target: AA is pass/fail, AAA is
    informative.

    WCAG Level Persona Check
    2.4.7 AA keyboard Focus indicator present
    2.4.13 AAA (informative) keyboard Focus indicator strength
    1.4.1 AA keyboard Indicator is not colour-only
    2.1.2 AA keyboard Keyboard trap
    2.4.1 AA keyboard No skip link
    2.4.3 AA keyboard Positive tabindex
    3.2.1 AA keyboard Context change from focus alone
    3.3.2 AA keyboard File input named only by the user-agent default ("Choose File")
    4.1.2 AA keyboard Focusable control with no accessible name
    1.1.1 AA screen-reader Missing alt text/aria-label
    1.3.1 AA screen-reader Heading level skip
    1.3.1 AA screen-reader Duplicate, unlabeled landmark roles
    4.1.2 AA screen-reader Interactive control announced as a bare role
    4.1.3 AA screen-reader Declared live region that never announced anything

    See docs/interface.md for the authoritative version of
    this table (full check descriptions) and the W3C persona references.

    Output

    Everything is written under a per-user temp dir (${TMPDIR}/keyboard-a11y-tester/…, or
    --out): a trace.json (per-step evidence), deterministic-findings.json (WCAG findings),
    screen-reader-census.json (screen-reader persona), and cropped screenshots/step_NNNN.png
    per viewport. See docs/interface.md for the
    complete directory layout and field-by-field schema of every output file.

    Focus-visible detection (2.4.7 AA presence + 2.4.13 AAA strength)

    Presence (AA) uses two independent signals, so a faint-but-real indicator is never
    missed:

    1. the focused element's computed style declares an outline or box-shadow (ground
      truth — recorded in the trace as computed_focus_style), or
    2. pixels change on focus (catches background/colour indicators with no outline).

    Either one means the indicator is present → AA pass. Pixel diffing compares the focused
    frame to a scroll-aligned baseline (the next step's frame, where the element is no longer
    focused — so focus is never manipulated programmatically), measuring ring slices at
    increasing offset (thin and offset outlines), the interior, and top/bottom edge bands.

    Strength (AAA, informative) measures whether the indicator meets 2.4.13 Focus
    Appearance — changed area ≥ a 2px-thick perimeter of the control, and ≥ 3:1 WCAG luminance
    contrast between focused and unfocused states. Advisory only. (This measure is unreliable
    on pages that mutate between steps — e.g. "load more" — because the neighbour-frame
    baseline then differs by content, not just the focus ring; treat AAA numbers on such pages
    with caution. AA presence is unaffected, being driven by the computed style.)

    So 2.4.7 (AA) requires only that an indicator is visible with no size/contrast minimum: a
    faint 1px or low-opacity ring passes AA and is flagged weak at AAA — rather than being
    falsely reported as "no focus indicator."

    Screen-reader detection (Lakshmi)

    The screen-reader persona is emulated, never driven for real: @guidepup/virtual-screen-reader
    builds an ARIA/ACCNAME-spec accessible tree over the live page and computes what a
    spec-compliant screen reader would announce, entirely in the browser's own JS engine — no
    NVDA/JAWS/VoiceOver is launched, and it works the same way on any OS the runner itself
    supports.

    Its self-contained browser bundle is injected via Playwright's context.addInitScript,
    which is not subject to the page's own CSP — verified against both a synthetic CSP-locked
    page and a real CSP-locked production site. Once injected, its virtual cursor tracks
    real keyboard focus automatically     (it listens for native focusin events), so every
    step you drive with real Tab/Enter/etc. produces a matching sr_announcement with no
    separate "chasing" logic and no drift between what's focused and what's reported as
    announced. The same mechanism also wires a MutationObserver that computes WAI-ARIA
    live-region semantics and captures "assertive: …"/"polite: …" announcements as they
    happen — this is what 4.1.3 (Status Messages) findings are derived from.

    Separately, once per newly-visited page URL, an ephemeral instance walks the entire page
    in reading order (never touching the live per-step monitor) to build
    screen-reader-census.json — the source for the heading-hierarchy, duplicate-landmark,
    missing-alt-text, and bare-role-control checks, since those need whole-page context rather
    than just the stops a keyboard user's Tab order happens to visit.

    This augments but does not replace testing with a real screen reader and real users —
    the upstream library's own README says exactly that, and it's worth repeating: this checks
    what a spec-compliant screen reader should announce given the page's ARIA/HTML, not the
    specific quirks of any one real screen reader implementation.

    CAPTCHAs

    CAPTCHAs detect automation and refuse to run; the runner has a page-scoped, human-approved
    compatibility workaround. See docs/usage.md for details.

    License

    MIT © Everett Zufelt. See LICENSE.

  • Prerequisites


    Not enough for a badge.

    The project MUST achieve a silver level badge. [achieve_silver]

  • Project oversight


    Unknown required information, not enough for a badge.

    The project MUST have a "bus factor" of 2 or more. (URL required) [bus_factor]
    A "bus factor" (aka "truck factor") is the minimum number of project members that have to suddenly disappear from a project ("hit by a bus") before the project stalls due to lack of knowledgeable or competent personnel. The truck-factor tool can estimate this for projects on GitHub. For more information, see Assessing the Bus Factor of Git Repositories by Cosentino et al.

    Unknown required information, not enough for a badge.

    The project MUST have at least two unassociated significant contributors. (URL required) [contributors_unassociated]
    Contributors are associated if they are paid to work by the same organization (as an employee or contractor) and the organization stands to benefit from the project's results. Financial grants do not count as being from the same organization if they pass through other organizations (e.g., science grants paid to different organizations from a common government or NGO source do not cause contributors to be associated). Someone is a significant contributor if they have made non-trivial contributions to the project in the past year. Examples of good indicators of a significant contributor are: written at least 1,000 lines of code, contributed 50 commits, or contributed at least 20 pages of documentation.

  • Other


    Unknown required information, not enough for a badge.

    The project MUST include a license statement in each source file. This MAY be done by including the following inside a comment near the beginning of each file: SPDX-License-Identifier: [SPDX license expression for project]. [license_per_file]
    This MAY also be done by including a statement in natural language identifying the license. The project MAY also include a stable URL pointing to the license text, or the full license text. Note that the criterion license_location requires the project license be in a standard location. See this SPDX tutorial for more information about SPDX license expressions. Note the relationship with copyright_per_file, whose content would typically precede the license information.

 Change Control 1/4 ●

  • Public version-controlled source repository


    Enough for a badge!

    The project's source repository MUST use a common distributed version control software (e.g., git or mercurial). [repo_distributed]
    Git is not specifically required and projects can use centralized version control software (such as subversion) with justification.

    Repository on GitHub, which uses git. git is distributed.


    Unknown required information, not enough for a badge.

    The project MUST clearly identify small tasks that can be performed by new or casual contributors. (URL required) [small_tasks]
    This identification is typically done by marking selected issues in an issue tracker with one or more tags the project uses for the purpose, e.g., up-for-grabs, first-timers-only, "Small fix", microtask, or IdealFirstBug. These new tasks need not involve adding functionality; they can be improving documentation, adding test cases, or anything else that aids the project and helps the contributor understand more about the project.

    Unknown required information, not enough for a badge.

    The project MUST require two-factor authentication (2FA) for developers for changing a central repository or accessing sensitive data (such as private vulnerability reports). This 2FA mechanism MAY use mechanisms without cryptographic mechanisms such as SMS, though that is not recommended. [require_2FA]

    Unknown required information, not enough for a badge.

    The project's two-factor authentication (2FA) SHOULD use cryptographic mechanisms to prevent impersonation. Short Message Service (SMS) based 2FA, by itself, does NOT meet this criterion, since it is not encrypted. [secure_2FA]
    A 2FA mechanism that meets this criterion would be a Time-based One-Time Password (TOTP) application that automatically generates an authentication code that changes after a certain period of time. Note that GitHub supports TOTP.

 Quality 1/7 ●

  • Coding standards


    Unknown required information, not enough for a badge.

    The project MUST document its code review requirements, including how code review is conducted, what must be checked, and what is required to be acceptable. (URL required) [code_review_standards]
    See also two_person_review and contribution_requirements.

    Unknown required information, not enough for a badge.

    The project MUST have at least 50% of all proposed modifications reviewed before release by a person other than the author, to determine if it is a worthwhile modification and free of known issues which would argue against its inclusion [two_person_review]

  • Working build system


    Unknown required information, not enough for a badge.

    The project MUST have a reproducible build. If no building occurs (e.g., scripting languages where the source code is used directly instead of being compiled), select "not applicable" (N/A). (URL required) [build_reproducible]
    A reproducible build means that multiple parties can independently redo the process of generating information from source files and get exactly the same bit-for-bit result. In some cases, this can be resolved by forcing some sort order. JavaScript developers may consider using npm shrinkwrap and webpack OccurrenceOrderPlugin. GCC and clang users may find the -frandom-seed option useful. The build environment (including the toolset) can often be defined for external parties by specifying the cryptographic hash of a specific container or virtual machine that they can use for rebuilding. The reproducible builds project has documentation on how to do this.

  • Automated test suite


    Enough for a badge!

    A test suite MUST be invocable in a standard way for that language. (URL required) [test_invocation]
    For example, "make check", "mvn test", or "rake test" (Ruby).

    https://github.com/ezufelt/keyboard-a11y-tester/blob/main/package.json


    Not enough for a badge.

    The project MUST implement continuous integration, where new or changed code is frequently integrated into a central code repository and automated tests are run on the result. (URL required) [test_continuous_integration]
    In most cases this means that each developer who works full-time on the project integrates at least daily.

    New project


    Unknown required information, not enough for a badge.

    The project MUST have FLOSS automated test suite(s) that provide at least 90% statement coverage if there is at least one FLOSS tool that can measure this criterion in the selected language. [test_statement_coverage90]

    Unknown required information, not enough for a badge.

    The project MUST have FLOSS automated test suite(s) that provide at least 80% branch coverage if there is at least one FLOSS tool that can measure this criterion in the selected language. [test_branch_coverage80]

 Security 0/5 ●

  • Use basic good cryptographic practices

    Note that some software does not need to use cryptographic mechanisms. If your project produces software that (1) includes, activates, or enables encryption functionality, and (2) might be released from the United States (US) to outside the US or to a non-US-citizen, you may be legally required to take a few extra steps. Typically this just involves sending an email. For more information, see the encryption section of Understanding Open Source Technology & US Export Controls.

    Unknown required information, not enough for a badge.

    The software produced by the project MUST support secure protocols for all of its network communications, such as SSHv2 or later, TLS1.2 or later (HTTPS), IPsec, SFTP, and SNMPv3. Insecure protocols such as FTP, HTTP, telnet, SSLv3 or earlier, and SSHv1 MUST be disabled by default, and only enabled if the user specifically configures it. If the software produced by the project does not support network communications, select "not applicable" (N/A). [crypto_used_network]

    Unknown required information, not enough for a badge.

    The software produced by the project MUST, if it supports or uses TLS, support at least TLS version 1.2. Note that the predecessor of TLS was called SSL. If the software does not use TLS, select "not applicable" (N/A). [crypto_tls12]

  • Secured delivery against man-in-the-middle (MITM) attacks


    Unknown required information, not enough for a badge.

    The project website, repository (if accessible via the web), and download site (if separate) MUST include key hardening headers with nonpermissive values. (URL required) [hardened_site]
    Note that GitHub and GitLab are known to meet this. Sites such as https://securityheaders.com/ can quickly check this. The key hardening headers are: Content Security Policy (CSP), HTTP Strict Transport Security (HSTS), X-Content-Type-Options (as "nosniff"), and X-Frame-Options. Fully static web sites with no ability to log in via the web pages could omit some hardening headers with less risk, but there's no reliable way to detect such sites, so we require these headers even if they are fully static sites.

  • Other security issues


    Unknown required information, not enough for a badge.

    The project MUST have performed a security review within the last 5 years. This review MUST consider the security requirements and security boundary. [security_review]
    This MAY be done by the project members and/or an independent evaluation. This evaluation MAY be supported by static and dynamic analysis tools, but there also must be human review to identify problems (particularly in design) that tools cannot detect.

    Unknown required information, not enough for a badge.

    Hardening mechanisms MUST be used in the software produced by the project so that software defects are less likely to result in security vulnerabilities. (URL required) [hardening]
    Hardening mechanisms may include HTTP headers like Content Security Policy (CSP), compiler flags to mitigate attacks (such as -fstack-protector), or compiler flags to eliminate undefined behavior. For our purposes least privilege is not considered a hardening mechanism (least privilege is important, but separate).

 Analysis 1/2 ●

  • Dynamic code analysis


    Not enough for a badge.

    The project MUST apply at least one dynamic analysis tool to any proposed major production release of the software produced by the project before its release. [dynamic_analysis]
    A dynamic analysis tool examines the software by executing it with specific inputs. For example, the project MAY use a fuzzing tool (e.g., American Fuzzy Lop) or a web application scanner (e.g., OWASP ZAP or w3af). In some cases the OSS-Fuzz project may be willing to apply fuzz testing to your project. For purposes of this criterion the dynamic analysis tool needs to vary the inputs in some way to look for various kinds of problems or be an automated test suite with at least 80% branch coverage. The Wikipedia page on dynamic analysis and the OWASP page on fuzzing identify some dynamic analysis tools. The analysis tool(s) MAY be focused on looking for security vulnerabilities, but this is not required.

    New project


    Barely enough for a badge.

    The project SHOULD include many run-time assertions in the software it produces and check those assertions during dynamic analysis. [dynamic_analysis_enable_assertions]
    This criterion does not suggest enabling assertions during production; that is entirely up to the project and its users to decide. This criterion's focus is instead to improve fault detection during dynamic analysis before deployment. Enabling assertions in production use is completely different from enabling assertions during dynamic analysis (such as testing). In some cases enabling assertions in production use is extremely unwise (especially in high-integrity components). There are many arguments against enabling assertions in production, e.g., libraries should not crash callers, their presence may cause rejection by app stores, and/or activating an assertion in production may expose private data such as private keys. Beware that in many Linux distributions NDEBUG is not defined, so C/C++ assert() will by default be enabled for production in those environments. It may be important to use a different assertion mechanism or defining NDEBUG for production in those environments.

    New project



This data is available under the Community Data License Agreement – Permissive, Version 2.0 (CDLA-Permissive-2.0). This means that a Data Recipient may share the Data, with or without modifications, so long as the Data Recipient makes available the text of this agreement with the shared Data. Please credit ezufelt and the OpenSSF Best Practices badge contributors.

Project badge entry owned by: ezufelt.
Entry created on 2026-07-10 18:22:26 UTC, last updated on 2026-07-10 21:04:01 UTC. Last achieved passing badge on 2026-07-10 21:04:01 UTC.