| name | adaptive-cards-accessibility |
| description | Semantics contract for Adaptive Card elements โ how labels, headings, live regions, and decorative-vs-meaningful images are exposed to screen readers, and how to test them with `tester.ensureSemantics()`. Use when adding or changing any element, input, or action widget, and at code review.
|
Adaptive Cards Accessibility Skill
Overview
A rendered card is mostly author-supplied content, so accessibility here is
not "add a label" โ it is deciding whose text becomes the accessible name,
and making sure the widget tree does not announce that text twice or not at all.
The library already has helpers for this. Use them rather than hand-rolling
Semantics widgets; a bespoke Semantics wrapper next to a visible label is the
most common way to introduce a double announcement.
The four rules
1. Author text is the accessible name โ never substitute a placeholder
altText from card JSON is the accessible name for images, icons, and media.
When the author omits it, the element is decorative and must be excluded
from the semantics tree.
Pass null through. Do not invent a fallback string.
// packages/flutter_adaptive_cards_fs/lib/src/utils/adaptive_image_utils.dart
AdaptiveImageUtils.getImage(url, semanticsLabel: adaptiveMap['altText']?.toString());
A null label makes Flutter drop the image from the semantics tree entirely,
which is the correct behavior for decoration. A fallback like
'alt text not set' or 'image' is worse than nothing: it forces a screen
reader to read noise on every decorative image. There is a regression test
pinning this (accessibility_semantics_test.dart โ "Image without altText is
not labeled 'alt text not set'"), so this is not a preference.
2. Inputs: link the visible label to the control, and exclude the duplicate
The visible label built by loadLabel() is a sibling widget. Assistive
technology reads it as an unrelated node, so focusing the field announces the
value with no idea what it is for. Two helpers in
packages/flutter_adaptive_cards_fs/lib/src/utils/utils.dart fix this:
inputSemanticsLabel({label, isRequired}) โ builds the spoken name and
appends ", required". The visible label marks required fields with a *
glyph, which screen readers do not reliably convey.labelInputSemantics({field, label, isRequired}) โ wraps the control in
MergeSemantics(Semantics(label: ...)) so the label and the field's own value
are announced as one node. Returns field unchanged when there is no label.
The visible label must then be wrapped in ExcludeSemantics, or the name is
announced twice. The pair is load-bearing โ always apply both:
// see text.dart, number.dart, date.dart, time.dart for the canonical shape
ExcludeSemantics(child: loadLabel(context: context, label: label, isRequired: isRequired)),
labelInputSemantics(field: myTextField, label: label, isRequired: isRequired),
3. Validation errors are live regions
A validation message that appears after a failed submit is invisible to a screen
reader user who has already moved focus past the field. The error builder in
utils.dart wraps the message in Semantics(liveRegion: true) so assistive
technology announces it on appearance without re-focusing the control. Any new
transient, user-visible status message should do the same.
4. Headings are structural, not stylistic
A TextBlock with "style": "heading" sets header: true and a
headingLevel from HostConfig (defaulting to 2). Screen reader users navigate by
heading, so a heading that only looks big is not a heading. Never fake one with
font size alone.
Reference: what is already wired
Read these before adding semantics to a new element โ the pattern you need almost
certainly exists.
| Concern | Where |
|---|
| Input label + required, exclude duplicate | utils.dart โ inputSemanticsLabel, labelInputSemantics |
| Error message live region | utils.dart (error builder) |
| Decorative vs meaningful image (null label) | utils/adaptive_image_utils.dart |
| Heading role + level | cards/elements/text_block.dart, rich_text_block.dart |
Value semantics + ExcludeSemantics on the visual | widgets/rating_stars.dart |
| Merged control semantics | cards/inputs/toggle.dart, rating.dart, choice_set.dart |
| Progress announcement | cards/elements/progress_bar.dart, progress_ring.dart |
| Table header role + header-to-cell association | cards/containers/table.dart |
Why Table annotates instead of merging
Flutter validates table roles strictly โ a SemanticsRole.table node's children
must all be rows, and a row's children must be cell/columnHeader. But
TableRow is a data holder with no render object, so there is no node to hang
a row role on; using the table roles would trip Flutter's own assertion, and
faking one row-node-per-cell would misreport the table's geometry to assistive
technology.
So table.dart does the thing that works on every platform today: header cells
get Semantics(header: true), and each body cell is annotated with its column
header (Semantics(label: 'Status')), which Flutter absorbs into the same node as
the cell's text โ announcing "Status Delayed" instead of a context-free "Delayed".
Annotate, do not MergeSemantics. Merging would flatten a cell's
selectAction or nested Input.* into the cell node and destroy its independent
focusability. Annotation leaves interactive descendants as their own nodes; there
is a regression test pinning exactly this (table_semantics_test.dart โ "a
selectAction cell stays an independently focusable button").
Residual limitation: a column whose header cell has no text (image-only) is
left unlabeled โ correctly, since there is nothing to speak. And because Flutter
exposes no true table geometry, screen readers still cannot report "row 3 of 12"
or navigate cell-by-cell as a grid.
Testing
Semantics are not built unless a test asks for them. Every accessibility test
must open a handle and dispose it:
testWidgets('heading-styled TextBlock exposes header + level', (tester) async {
final handle = tester.ensureSemantics(); // REQUIRED
await tester.pumpWidget(buildCard({'type': 'TextBlock', 'text': 'H', 'style': 'heading'}));
await tester.pumpAndSettle();
final data = tester.getSemantics(find.text('H')).getSemanticsData();
expect(data.flagsCollection.isHeader, isTrue);
expect(data.headingLevel, 2);
handle.dispose(); // REQUIRED
});
Assertion styles, by what you are proving:
- A name is announced โ
find.bySemanticsLabel('...').
- A name is not announced (decorative, or no double-read) โ
expect(find.bySemanticsLabel('...'), findsNothing).
- A role or flag (
isHeader, isButton, isTextField, headingLevel,
isLiveRegion) โ tester.getSemantics(finder).getSemanticsData().
Existing suites to extend rather than duplicate:
test/accessibility_semantics_test.dart โ decorative images, selectAction
button naming, rating values, carousel dots.test/input_label_semantics_test.dart โ label + required, no double-read.test/input_error_and_progress_a11y_test.dart โ live regions, progress.test/textblock_heading_and_factset_semantics_test.dart โ heading roles.test/inputs/choice_set_filtered_semantics_test.dart โ filtered choice sets.test/containers/table_semantics_test.dart โ header cells, header-to-cell
association, interactive cells staying focusable.
A new element with a user-visible name, role, or value needs a semantics test.
It is the only way these regressions get caught โ they are invisible in goldens.
Known gaps
These are the known holes at the time of writing โ confirm each against the
current widget before acting, since one may already be fixed. Do not treat the
list as a backlog you must clear in passing, but do not add to it, and prefer
fixing the one you are already touching.
flutter_adaptive_charts_fs โ no semantics anywhere. A chart is a picture
of data with no text alternative. It needs at least a summary label
(title + series + notable values).CompoundButton โ correctly an ElevatedButton (button role) with a
correctly-unlabeled decorative icon, but title and description are separate
Text nodes inside it. Verify they merge into one announcement; if they read as
two nodes, wrap in MergeSemantics.CodeBlock โ plain Text; language is parsed but never announced.
Accordion (ExpansionTile) and TabSet (TabBar/Tab) inherit Material's
built-in expand/collapse and tab semantics, so they are not gaps โ do not
"fix" them by wrapping in redundant Semantics.
Review checklist
