a11y-toolkit
AGENTS.md7 starsChanged 14 days agoMITView on GitHub
Rules for coding agents (and humans) contributing to this repository. A zero-dependency (stdlib-only) WCAG 2.2 accessibility engine + MCP server (37 tools) + CLI. Python 3.9+. The engine runs as an MCP server, a CLI (a11ytoolkit), and is vendored by a11y.es (production). Zero telemetry is a product promise — never add phone-home code. This is the PUBLIC layer of a 4-repo system (canonical map lives in the private a11y-toolkit-pro/ARCHITECTURE.md): - a11y-toolkit-pro (private) — the consulting playbook (508 GSA tree, tender…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,439 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository2 other files it gives its agents.
- llms.txtllms.txt
- a11y-toolkitSkill
The file
We split this AGENTS.md into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 78
If it alters architecture: docs/ARCHITECTURE.md (the module map is pinned by test_solido).
- Distinctive: line 80
If it touches deployment of the SaaS: deploy/README.md the same day (the tunnel-era lesson: a runbook describing a dead architecture cost …
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AGENTS.md — working on a11y-toolkit
Rules for coding agents (and humans) contributing to this repository.
What this codebase is
A zero-dependency (stdlib-only) WCAG 2.2 accessibility engine + MCP server
(37 tools) + CLI. Python 3.9+. The engine runs as an MCP server, a CLI
(a11ytoolkit), and is vendored by a11y.es (production). Zero telemetry is a
product promise — never add phone-home code.
The constellation (who consumes this engine)
This is the PUBLIC layer of a 4-repo system (canonical map lives in the
private a11y-toolkit-pro/ARCHITECTURE.md):
- a11y-toolkit-pro (private) — the consulting playbook (508 GSA tree, tender annexes, WCAG-EM sampling, remediation planning), extracted from this repo on 2026-11-04 by founder decision. Depends on this engine, never the reverse. Nothing from those modules comes back here.
- a11y.es (private product) — vendors this engine into its Docker image; changes here ship to production via their sync ritual.
- Lince/cpa11y (private product) — pins this engine by commit.
Implications for ANY change here: keep the engine runnable with the pro
package absent; the conformance DRAFT generators (a11yvpat,
a11yopenacr, a11ystatement) are a FROZEN layer — maintenance fixes
only, new conformance capability goes to the SaaS, not here (each carries
a "FROZEN LAYER" docstring stamp). Severities/messages are calibrated
against expert review — a change that moves a score is a public behavior
change and needs its regression test in the same commit.
Non-negotiable invariants (all pinned by tests — make test runs them)
python3 tests/test_solido.py— es/en message catalogs key-identical, every finding has_remremediation, version files aligned, tool/prompt counts consistent across every surface, hostile-HTML fuzz never crashes.python3 tests/test_naming.py— the public surface and the code are English-only: tool params, enums, dict keys, identifiers, string-key READS, phrase-table/JS placeholder consistency. Spanish lives ONLY inlang=esoutput content and comments.python3 tests/test_contract.py— tool schema ↔ dispatcher ↔ function signature must agree mechanically.- Signal→criterion map and Understanding-URL coverage are pinned in test_solido.
Conventions
- Release flow:
make release V=X.Y.Z(atomic bump, full-suite gate — everytests/test_*.pyby discovery, single v-prefixed commit, tag, push — PyPI publishes the tag). No intermediate work commits; the working tree is picked up whole. - Internal/strategy material NEVER goes in this repo (pre-commit guard blocks it; drafts live outside). Public docs are English; CHANGELOG keeps history.
- Bilingual output: every user-facing message exists in es AND en catalogs.
- Finding emission:
add(severity, 'X.Y.Z', 'signal_name')— a new signal automatically joins the signal→criterion map (test re-derives it); a new message key needs its_remtwin in both languages.
After changing anything
make test # all suites; red means the invariant you touched is load-bearingThe accessibility rule applies to this repo too: it must audit clean.
VITAL: publish every change (founder rule, 2026-11-01)
Any change to public capabilities MUST ship with its publication in the SAME commit — no exceptions, no "later":
- CHANGELOG.md — entry in the
pendientesection at the top (what it is, which pattern it comes from, what it does NOT do). - README.md AND README.es.md — both, always: if the change adds or alters a user-facing capability, it goes in the command examples block; if it changes numbers or coverage claims, those lines update too. The Spanish README is a full sibling, never a stale translation.
- If it alters architecture: docs/ARCHITECTURE.md (the module map is pinned by test_solido).
- If it touches deployment of the SaaS: deploy/README.md the same day (the tunnel-era lesson: a runbook describing a dead architecture cost an hour of diagnosis).
Red means the publication invariant you touched is load-bearing.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

