Accessibility

Run the Accessibility Test Framework (ATF) over each rendered preview and surface the findings, the semantic hierarchy, the annotated overlay PNG, and a 48 dp touch-target audit as four related kinds.

At a glance


Kinds	`a11y/atf`, `a11y/hierarchy`, `a11y/overlay`, `a11y/touchTargets`
Schema version	1
Modules	`:data-a11y-core` (published) · `:data-a11y-connector` · `:data-a11y-hierarchy-android`
Render mode	a11y (daemon-only — opt in via the VS Code chip toggle or `compose-preview a11y`; standalone `composePreviewRender` Gradle task does not produce a11y data)
Cost	low
Token usage	~100 tok inline (e.g. `a11y/atf` ~370 chars); +~1.5 k per `a11y/overlay` PNG read. See token usage.
Transport	inline (JSON) · path (overlay PNG)
Platforms	Android

What it answers

a11y/atf — does this preview violate WCAG / Android a11y best practices? Each AccessibilityFinding carries a check id, severity, source, and the node it implicates. The renderer keeps LocalInspectionMode = false so Compose populates real semantics; infinite animations tick through rather than parking on the paused frame clock (true regardless of the a11y opt-in).
a11y/hierarchy — what does an assistive technology see? Per-node label, role, states, merged, boundsInScreen. Read it the same way TalkBack would.
a11y/overlay — the rendered PNG with finding bounding boxes baked in. Pure-image; transport=path.
a11y/touchTargets — every interactive node with its measured bounds vs. the 48 dp Material minimum, with overlap detection between adjacent targets.

What it does NOT answer

It is not a substitute for testing with real assistive technology — TalkBack on a real device may surface gestures or focus order issues that a single static frame can’t.
It does not opine on copy quality (placeholder text, jargon). Use text/strings for that.
Findings are advisory, not pass/fail; severity is the agent’s signal.

Use cases

PR review: gate UI PRs on no new ERROR-severity findings; reviewers can paste the overlay PNG inline.
Refactor confidence: diff a11y/hierarchy before/after a state-hoisting change to confirm semantics didn’t regress.
Density audit: sweep all previews for a11y/touchTargets violations on a Wear watch face where small targets matter most.

Payload shape

AccessibilityFindingsPayload, AccessibilityHierarchyPayload, AccessibilityTouchTargetsPayload, AccessibilityOverlayArtifact (path-only) — defined in :data-a11y-core.

// a11y/atf
{
  "findings": [
    { "checkId": "TouchTargetSize", "severity": "ERROR",
      "label": "Submit", "boundsInScreen": "48,200,144,232" }
  ]
}

// a11y/hierarchy
{
  "nodes": [
    { "label": "Submit", "role": "Button", "states": ["Enabled"],
      "merged": true, "boundsInScreen": "48,200,144,232" }
  ]
}

Enabling

A11y data products are produced exclusively by the preview daemon. There is no Gradle DSL or property knob — opt in per invocation through one of:

VS Code chip toggle — clicking an a11y/* chip on a preview card routes a per-preview data/subscribe through the daemon. The next render attaches the post-capture walk and stamps a11y-atf.json, a11y-hierarchy.json, a11y-overlay.png, and a11y-touchTargets.json under build/compose-previews/data/<id>/.
compose-preview a11y — the CLI command spins up a temporary daemon, renders with the a11y subscription on, drains the findings, and shuts down.
MCP / agent flows — the daemon’s data/subscribe request enables any subset of a11y/* kinds for a given preview.

The standalone :composePreviewRenderAll Gradle task does NOT produce a11y sidecars — it’s a normal-render-only path.

Companion products

Layout inspector — compose/semantics for testTag / role / mergeMode without the ATF audit.
Strings — text/strings for drawn-text quality next to assistive labels.

See also the agent-facing review guidance in skills/compose-preview/references/a11y.md.