fix: merge HeZeBang's fork fixes + accessibility/i18n audit items (#128)

demo/lib/main.dart

@@ -881,7 +881,8 @@ class GalleryScreen extends StatelessWidget {
               style: const TextStyle(
                 fontSize: 12,
                 fontWeight: FontWeight.w600,
-                color: Color(0xFF9893A8),
+                // #9893A8 on white was ~3:1 at 12px; this passes AA (~7:1).
+                color: Color(0xFF5B5570),
                 letterSpacing: 0.2,
               ),
             ),

packages/flutter_ai_client/CHANGELOG.md

@@ -1,5 +1,16 @@
 # Changelog
 
+## 0.3.1
+
+- Fix: settle dangling tool calls before `submit`/`regenerate`/`editMessage`. A
+  turn interrupted mid agent-loop (`stop()`, a `maxSteps` cutoff, or a
+  transcript rehydrated from storage) left a `ToolCallPart` with no matching
+  `ToolResultPart`; providers reject such a history, so every later `submit`
+  failed with a request error. The controller now inserts a synthesized error
+  result immediately after each affected assistant message — in place, not
+  appended, since providers require the result in the following turn.
+  Thanks to @HeZeBang for the report and the fix. (#149)
+
 ## 0.3.0
 
 - Add `UseChatController.load(AiConversation)` — swaps the transcript in place

packages/flutter_ai_client/lib/src/use_chat_controller.dart

@@ -244,7 +244,9 @@ class UseChatController extends ChangeNotifier {
     _capture = _Capture.reset; // a new user turn starts a fresh branch set
     _step = 0;
     _toolCallCounts.clear();
-    _processor.reset(_processor.conversation.append(userMessage));
+    _processor.reset(
+      _settleDanglingToolCalls(_processor.conversation).append(userMessage),
+    );
     _status = ChatStatus.submitted;
     _scheduleNotify();
     return _beginTurn();
@@ -263,7 +265,10 @@ class UseChatController extends ChangeNotifier {
     _step = 0;
     _toolCallCounts.clear();
     _processor.reset(
-      _processor.conversation.copyWith(messages: all.sublist(0, lastUser + 1)),
+      _settleDanglingToolCalls(
+        _processor.conversation
+            .copyWith(messages: all.sublist(0, lastUser + 1)),
+      ),
     );
     _status = ChatStatus.submitted;
     _scheduleNotify();
@@ -308,11 +313,13 @@ class UseChatController extends ChangeNotifier {
     _step = 0;
     _toolCallCounts.clear();
     _processor.reset(
-      _processor.conversation.copyWith(
-        messages: [
-          ...all.sublist(0, index),
-          original.copyWith(parts: parts, status: AiMessageStatus.complete),
-        ],
+      _settleDanglingToolCalls(
+        _processor.conversation.copyWith(
+          messages: [
+            ...all.sublist(0, index),
+            original.copyWith(parts: parts, status: AiMessageStatus.complete),
+          ],
+        ),
       ),
     );
     _status = ChatStatus.submitted;
@@ -655,6 +662,51 @@ class UseChatController extends ChangeNotifier {
     _dispatch();
   }
 
+  /// Settles tool calls left unanswered anywhere in the transcript — a turn
+  /// stopped/replaced mid agent-loop, `maxSteps` cutting a loop short, or a
+  /// dirty transcript rehydrated from storage — by inserting a synthesized
+  /// error [ToolResultPart] message directly after each affected assistant
+  /// message. Providers reject a history containing a tool call with no
+  /// result in the immediately-following turn (so the fix-up must be inserted
+  /// in place, not appended at the end), and without it every subsequent
+  /// [submit] on the conversation fails with a request error.
+  AiConversation _settleDanglingToolCalls(AiConversation conversation) {
+    final msgs = conversation.messages;
+    final answered = <String>{
+      for (final m in msgs)
+        for (final p in m.parts)
+          if (p is ToolResultPart) p.toolCallId,
+    };
+    var changed = false;
+    final out = <AiMessage>[];
+    for (final m in msgs) {
+      out.add(m);
+      if (m.role != AiRole.assistant) continue;
+      final dangling = m.parts
+          .whereType<ToolCallPart>()
+          .where((c) => !answered.contains(c.toolCallId))
+          .toList();
+      if (dangling.isEmpty) continue;
+      changed = true;
+      out.add(
+        AiMessage(
+          id: _newId(),
+          role: AiRole.tool,
+          parts: <AiPart>[
+            for (final call in dangling)
+              ToolResultPart(
+                toolCallId: call.toolCallId,
+                isError: true,
+                result: 'Cancelled: the turn was interrupted before this tool '
+                    'call produced a result.',
+              ),
+          ],
+        ),
+      );
+    }
+    return changed ? conversation.copyWith(messages: out) : conversation;
+  }
+
   /// Tool calls in the latest assistant message that have no matching
   /// [ToolResultPart] anywhere in the transcript yet.
   /// A stable identity for a tool call — name plus JSON-encoded args — used to

packages/flutter_ai_client/pubspec.yaml

@@ -1,6 +1,6 @@
 name: flutter_ai_client
 description: "Provider-agnostic chat controller for flutter_ai: wraps any LlmProvider with optimistic send, cancellation, regeneration, and frame-batched streaming."
-version: 0.3.0
+version: 0.3.1
 repository: https://github.com/ananmouaz/flutter_ai/tree/main/packages/flutter_ai_client
 issue_tracker: https://github.com/ananmouaz/flutter_ai/issues
 homepage: https://github.com/ananmouaz/flutter_ai

packages/flutter_ai_client/test/use_chat_controller_test.dart

@@ -953,6 +953,103 @@ void main() {
         hasLength(1),
       );
     });
+
+    test('submit settles dangling tool calls with synthesized error results',
+        () async {
+      // Turn 1 ends with an unanswered tool call (manual mode, never
+      // executed). Submitting a new user message must first append error
+      // results for it — providers reject a history containing a tool call
+      // with no following result.
+      final provider = _ToolThenTextProvider();
+      final controller = UseChatController(
+        provider: provider,
+        scheduler: syncScheduler,
+        idGenerator: seqIds(),
+      );
+      addTearDown(controller.dispose);
+
+      await controller.sendText('weather?');
+      await controller.sendText('never mind');
+
+      final messages = controller.messages;
+      // user, assistant(tool call), tool(synthesized error), user, assistant.
+      expect(messages.map((m) => m.role).toList(), [
+        AiRole.user,
+        AiRole.assistant,
+        AiRole.tool,
+        AiRole.user,
+        AiRole.assistant,
+      ]);
+      final settled = messages[2].parts.whereType<ToolResultPart>().single;
+      expect(settled.toolCallId, 'c1');
+      expect(settled.isError, isTrue);
+    });
+
+    test('submit settles dangling tool calls buried mid-history', () async {
+      // A rehydrated transcript where an interrupted agent loop left an
+      // unanswered tool call in the MIDDLE of the history (later turns
+      // completed normally). The settle must insert the synthesized result
+      // directly after the affected assistant message — providers require the
+      // result in the immediately-following turn, so appending at the end
+      // would not fix the request.
+      const dirty = AiConversation(
+        id: 'thread-dirty',
+        messages: [
+          AiMessage(id: 'u1', role: AiRole.user, parts: [TextPart('q1')]),
+          AiMessage(
+            id: 'a1',
+            role: AiRole.assistant,
+            parts: [
+              ToolCallPart(
+                toolCallId: 'c9',
+                toolName: 'get_week_schedule',
+                args: {},
+              ),
+            ],
+            status: AiMessageStatus.complete,
+          ),
+          AiMessage(id: 'u2', role: AiRole.user, parts: [TextPart('q2')]),
+          AiMessage(
+            id: 'a2',
+            role: AiRole.assistant,
+            parts: [TextPart('answer 2')],
+            status: AiMessageStatus.complete,
+          ),
+        ],
+      );
+      final provider = ScriptedProvider(const [
+        MessageStarted(messageId: 'a3', role: AiRole.assistant),
+        TextDelta(messageId: 'a3', delta: 'answer 3'),
+        MessageFinished(messageId: 'a3', reason: FinishReason.stop),
+      ]);
+      final controller = UseChatController(
+        provider: provider,
+        scheduler: syncScheduler,
+        idGenerator: seqIds(),
+        initial: dirty,
+      );
+      addTearDown(controller.dispose);
+
+      await controller.sendText('q3');
+
+      final messages = controller.messages;
+      // u1, a1(tool call), tool(synthesized), u2, a2, u3, a3.
+      expect(messages.map((m) => m.role).toList(), [
+        AiRole.user,
+        AiRole.assistant,
+        AiRole.tool,
+        AiRole.user,
+        AiRole.assistant,
+        AiRole.user,
+        AiRole.assistant,
+      ]);
+      final settled = messages[2].parts.whereType<ToolResultPart>().single;
+      expect(settled.toolCallId, 'c9');
+      expect(settled.isError, isTrue);
+      // The provider must have been sent the settled history too.
+      final sent = provider.lastConversation!.messages;
+      expect(sent[2].role, AiRole.tool);
+    });
   });
 
   group('threads', () {

packages/flutter_ai_elements/CHANGELOG.md

@@ -1,5 +1,37 @@
 # Changelog
 
+## 0.3.0
+
+Accessibility, internationalization, and cross-fork compile fixes.
+
+- Fix: `aiLightHaptic` compiles against both upstream Flutter and the OHOS
+  Flutter fork. The exhaustive `TargetPlatform` switch could not cover both
+  enums at once (the fork adds `TargetPlatform.ohos`), so it is now an
+  iOS/Android allowlist. Thanks to @HeZeBang. (#149)
+- Fix: strings that bypassed `AiLocalizations` are now localizable — tool-call
+  section labels (`toolArguments`/`toolResult`/`toolError`), the context meter
+  (`contextLabel`), source expand/collapse (`moreSources`/`showLess`), live
+  voice statuses (`liveConnecting`/`liveListening`/`liveThinking`/
+  `liveSpeaking`), the empty-state headline (`emptyStateTitle`), the inline
+  citation and model-selector semantics labels (`citation`/`selectModel`), and
+  the attachment remove badge (`removeAttachment`). (#128)
+- Fix: touch targets below the 44px platform minimum — the attachment remove
+  badge and `AiBranch`'s navigation arrows now use `InkResponse` over a 44px
+  box, which also gives them focus traversal and keyboard activation.
+  `AiModelSelector`'s trigger moves from `GestureDetector` to `InkWell` for the
+  same reason. (#128)
+- Fix: the caution-tone confirm button painted white on amber (~2.2:1). The
+  label color is now derived from the fill's brightness, and the buttons use a
+  `minHeight` instead of a fixed 40px height so labels are not clipped at large
+  text scales. (#128)
+- Fix: RTL gaps — the streaming loader alignment, the "New chat" label, and the
+  Markdown blockquote's border/padding are now directional. (#128)
+- **Potentially source-breaking:** `AiChainOfThought.title`,
+  `AiContextMeter.label`, `AiEmptyState.title`, and `AiComposer.hintText` are
+  now `String?` and default to the matching `AiLocalizations` value. Passing a
+  string still works; only reading one of these fields as a non-nullable
+  `String` needs a change.
+
 ## 0.2.0
 
 Extensibility release — driven by dogfooding a full Gemini-clone app on the

packages/flutter_ai_elements/lib/src/l10n/ai_localizations.dart

@@ -49,6 +49,20 @@ class AiLocalizations {
     this.loading = 'Loading',
     this.you = 'You',
     this.assistant = 'Assistant',
+    this.toolArguments = 'Arguments',
+    this.toolResult = 'Result',
+    this.toolError = 'Error',
+    this.contextLabel = 'Context',
+    this.showLess = 'Show less',
+    this.moreSources = _defaultMoreSources,
+    this.removeAttachment = 'Remove attachment',
+    this.liveConnecting = 'Connecting…',
+    this.liveListening = 'Listening',
+    this.liveThinking = 'Thinking…',
+    this.liveSpeaking = 'Speaking',
+    this.emptyStateTitle = 'Start the conversation',
+    this.citation = _defaultCitation,
+    this.selectModel = _defaultSelectModel,
   });
 
   /// Copy-to-clipboard action.
@@ -138,6 +152,56 @@ class AiLocalizations {
   /// Avatar accessibility label for the assistant.
   final String assistant;
 
+  /// Section title for a tool call's arguments.
+  final String toolArguments;
+
+  /// Section title for a tool call's successful result.
+  final String toolResult;
+
+  /// Section title for a tool call's error result.
+  final String toolError;
+
+  /// Leading label on the context-window meter.
+  final String contextLabel;
+
+  /// Collapse action on a list that was expanded (e.g. sources).
+  final String showLess;
+
+  /// Expand action on a collapsed source list, given the hidden chip count —
+  /// a callback rather than a string so translations can decline the number.
+  final String Function(int hiddenCount) moreSources;
+
+  /// Accessibility label on the remove badge of a staged attachment.
+  final String removeAttachment;
+
+  /// Live-voice status while the session is connecting.
+  final String liveConnecting;
+
+  /// Live-voice status while the mic is open.
+  final String liveListening;
+
+  /// Live-voice status while the model is generating.
+  final String liveThinking;
+
+  /// Live-voice status while the model is talking.
+  final String liveSpeaking;
+
+  /// Default headline on the empty conversation state.
+  final String emptyStateTitle;
+
+  /// Accessibility label for an inline citation badge, given its number.
+  final String Function(int number) citation;
+
+  /// Accessibility label for the model selector, given the current model.
+  final String Function(String modelLabel) selectModel;
+
+  static String _defaultMoreSources(int hiddenCount) => '+$hiddenCount more';
+
+  static String _defaultCitation(int number) => 'Citation $number';
+
+  static String _defaultSelectModel(String modelLabel) =>
+      'Select model, $modelLabel';
+
   /// The nearest [AiLocalizations]. Resolution order: an [AiLocalizationsScope]
   /// in the tree (the simplest way to override — no delegate wiring), then a
   /// `Localizations` delegate, then the English defaults.

packages/flutter_ai_elements/lib/src/widgets/ai_branch.dart

@@ -83,14 +83,21 @@ class _Arrow extends StatelessWidget {
       button: true,
       enabled: onTap != null,
       label: label,
-      child: GestureDetector(
+      // InkResponse over a 44px box: the arrow paints at 18px, but branch
+      // navigation needs a platform-minimum touch target — plus focus
+      // traversal and keyboard activation, which a GestureDetector lacks.
+      child: InkResponse(
         onTap: onTap,
-        child: Padding(
-          padding: const EdgeInsets.all(4),
-          child: Icon(
-            icon,
-            size: 18,
-            color: onTap == null ? color?.withValues(alpha: 0.3) : color,
+        radius: 22,
+        child: SizedBox(
+          width: 44,
+          height: 44,
+          child: Center(
+            child: Icon(
+              icon,
+              size: 18,
+              color: onTap == null ? color?.withValues(alpha: 0.3) : color,
+            ),
           ),
         ),
       ),