feat(mobile): render messages with gpt_markdown 1.3.4 plusparse

mobile/lib/features/channels/message_content.dart

@@ -1,8 +1,11 @@
 import 'message_reply_drag_target.dart';
 import 'package:buzz/shared/widgets/feedback_snackbar.dart';
+
 import 'dart:async';
+
 import '../../shared/widgets/ios_media_carousel.dart';
 import '../../shared/widgets/message_attachment_corners.dart';
+
 import 'dart:io';
 import 'dart:math' as math;
 
@@ -37,6 +40,7 @@ import 'local_send_bubble_surface.dart';
 import 'media_viewer_page.dart';
 import 'message_content/link_normalizer.dart';
 import 'message_content/markdown_escapes.dart';
+import 'message_content/markdown_tables.dart';
 import 'message_content/preview_text.dart';
 import 'message_media.dart';
 import 'message_content/bubble_segments.dart';
@@ -299,10 +303,8 @@ class MessageContent extends HookConsumerWidget {
         ? extractTrailingImageGallery(normalizedContent, imetaByUrl)
         : null;
     final galleryText = trailingGallery?.content ?? normalizedContent;
-    // gpt_markdown has no backslash escapes, so hold escaped characters as
-    // stand-ins that _MarkdownEscapeMd and the link builders restore.
     final markdownContent = useMemoized(
-      () => shieldMarkdownProseEscapes(galleryText),
+      () => separateMarkdownTables(escapeMarkdownMathDelimiters(galleryText)),
       [galleryText],
     );
     final customEmoji = _mergeCustomEmoji(
@@ -340,8 +342,8 @@ class MessageContent extends HookConsumerWidget {
 
     final finalContent = useMemoized(() {
       // Replace spaces with non-breaking spaces inside known mention names
-      // so the gpt_markdown combined regex can match multi-word names
-      // even when caseSensitive is not preserved.
+      // so _MentionMd's pattern matches multi-word names as one token, in
+      // the bound name's case.
       // Skip content inside backticks to avoid altering inline code.
       final mentionParts = markdownContent.split('`');
       final mentionBuf = StringBuffer();
@@ -363,17 +365,10 @@ class MessageContent extends HookConsumerWidget {
           mentionBuf.write(segment);
         }
       }
-      var result = mentionBuf.toString();
-
-      // Ensure channel links at the very start of content don't get
-      // swallowed by markdown processing.
-      if (RegExp(r'^#[A-Za-z0-9_]').hasMatch(result)) {
-        result = '\u200B$result';
-      }
-      return result;
+      return mentionBuf.toString();
     }, [markdownContent, mentionPresentationKey]);
 
-    final inlineComponents = _useMessageInlineComponents(
+    final inlinePatterns = _useMessageInlinePatterns(
       content: content,
       finalContent: finalContent,
       mentionNames: resolvedMentionNames,
@@ -410,41 +405,41 @@ class MessageContent extends HookConsumerWidget {
             : null,
         followLinkColor: false,
         // normalizeBareLinks() already turns bare URLs into Markdown links;
-        // gpt_markdown 1.2.0 autolinks by default, so both would run.
+        // gpt_markdown autolinks by default, so both would run.
         autolink: false,
+        // Desktop has no GitHub alerts: a `> [!NOTE]` quote stays a quote.
+        alertBuilder: (_, alert) => alert.asBlockQuote(),
         codeBuilder: (context, name, code, closed) => _MessageCodeBlock(
           name: name,
           code: code,
           bubble: bubble,
           ink: bubble && ownMessage ? ownTextColor : null,
         ),
-        linkBuilder: (context, linkText, shieldedUrl, linkStyle) {
-          final url = unshieldMarkdownEscapes(shieldedUrl);
-          return _buildLink(
-            context,
+        // Same widget-shaped link as before: chips, media links and taps
+        // are handled in _buildLink, so the package's own tap is not used.
+        inlineLinkBuilder: (link) => link.asWidgetSpan(
+          _buildLink(
+            link.context,
             ref,
-            linkText,
-            url,
-            imetaByUrl[url],
-            linkStyle,
+            TextSpan(children: link.labelSpans),
+            link.url,
+            imetaByUrl[link.url],
+            link.style,
             style,
             resolvedChannelTap,
             resolvedChannelNames,
-          );
-        },
-        imageBuilder: (context, shieldedUrl, _, _) {
-          final imageUrl = unshieldMarkdownEscapes(shieldedUrl);
-          return _buildMedia(
-            context,
-            imageUrl,
-            imetaByUrl[imageUrl],
-            onReply: onMediaReply == null ? null : mediaReply,
-            onMore: onMediaMore == null ? null : mediaMore,
-          );
-        },
+          ),
+        ),
+        imageBuilder: (context, imageUrl, _, _) => _buildMedia(
+          context,
+          imageUrl,
+          imetaByUrl[imageUrl],
+          onReply: onMediaReply == null ? null : mediaReply,
+          onMore: onMediaMore == null ? null : mediaMore,
+        ),
         textAlign: textAlign ?? (bubble ? TextAlign.start : null),
         maxLines: maxLines,
-        inlineComponents: inlineComponents,
+        inlinePatterns: inlinePatterns,
       ),
     );
     final segments = bubble
@@ -687,8 +682,13 @@ class MessageContent extends HookConsumerWidget {
 
     // Mobile has no repo/PR/issue destination yet. Keep these presentation-only
     // instead of exposing a control whose tap cannot do anything.
+    // The container keeps the chip's label its own node; without it, the
+    // non-interactive chip merges into the surrounding paragraph's label.
     if (buzzLink is EntityDeepLink) {
-      return IgnorePointer(child: renderedLink);
+      return Semantics(
+        container: true,
+        child: IgnorePointer(child: renderedLink),
+      );
     }
 
     return GestureDetector(
@@ -1036,7 +1036,7 @@ class _MessageCodeBlock extends HookWidget {
   }
 }
 
-class _MentionMd extends InlineMd {
+class _MentionMd {
   final Map<String, Set<String>> bindings;
   final Map<String, String> displayLabels;
   final Map<String, String> mentionNames;
@@ -1062,28 +1062,24 @@ class _MentionMd extends InlineMd {
   /// placeholder nested inside the link's own placeholder does not paint on
   /// iOS — an authored `[@mention](url)` renders as nothing. Link resolution
   /// wins over token detection inside a label.
-  @override
-  Set<MarkdownScope> get scopes => MarkdownComponent.allScopesExceptLinkLabel;
+  late final InlinePattern pattern = InlinePattern(
+    pattern: _exp,
+    builder: (context, match, style) => span(context, match[0]!, style),
+  );
 
-  @override
   RegExp get exp => _exp;
 
-  @override
-  InlineSpan span(
-    BuildContext context,
-    String text,
-    final GptMarkdownConfig config,
-  ) {
+  InlineSpan span(BuildContext context, String text, TextStyle style) {
     final raw = exp.firstMatch(text.trim())?.group(0);
     if (raw == null) {
-      return TextSpan(text: text, style: config.style);
+      return TextSpan(text: text, style: style);
     }
 
     final name = raw.substring(1).replaceAll('\u00A0', ' ').toLowerCase();
     final matches = bindings[name] ?? const <String>{};
     final pubkey = matches.length == 1 ? matches.single : null;
     if (bindings.containsKey(name) && matches.length != 1) {
-      return TextSpan(text: text, style: config.style);
+      return TextSpan(text: text, style: style);
     }
     final displayName = name.contains(RegExp(r'\([0-9a-f]{64}\)'))
         ? displayLabels[name]
@@ -1100,7 +1096,7 @@ class _MentionMd extends InlineMd {
       label: visibleLabel,
       semanticsLabel: fullLabel,
       isAgent: isAgent,
-      textStyle: config.style,
+      textStyle: style,
     );
 
     return WidgetSpan(

mobile/lib/features/channels/message_content/bubble_segments.dart

@@ -46,8 +46,10 @@ List<({String text, bool attachment})> messageBubbleSegments(
     final isAttachment =
         match != null &&
         (line.trim().startsWith('!') ||
-            // Content arrives with escapes shielded; imeta keys are plain.
-            imeta.containsKey(unshieldMarkdownEscapes(match.group(1)!)));
+            // CommonMark unescapes link destinations; imeta keys are plain.
+            imeta.containsKey(
+              unshieldMarkdownEscapes(shieldMarkdownEscapes(match.group(1)!)),
+            ));
     if (attachment != null &&
         isAttachment != attachment &&
         line.trim().isNotEmpty) {

mobile/lib/features/channels/message_content/inline_components.dart

@@ -1,6 +1,6 @@
 part of '../message_content.dart';
 
-List<MarkdownComponent> _useMessageInlineComponents({
+List<InlinePattern> _useMessageInlinePatterns({
   required String content,
   required String finalContent,
   required Map<String, String> mentionNames,
@@ -46,7 +46,7 @@ List<MarkdownComponent> _useMessageInlineComponents({
     hasMentionHandler: onMentionTap != null,
   );
 
-  // gpt_markdown compares components by identity. Recreating them on every
+  // gpt_markdown compares patterns by identity. Recreating them on every
   // parent rebuild reparses unchanged message bodies on the UI isolate.
   return useMemoized(
     () => [
@@ -60,16 +60,16 @@ List<MarkdownComponent> _useMessageInlineComponents({
         },
         agentMentionPubkeys: inputs.agentPubkeys,
         onMentionTap: inputs.hasMentionHandler ? mentionTap : null,
-      ),
-      CustomEmojiMd(inputs.customEmoji, content: finalContent, size: emojiSize),
+      ).pattern,
+      CustomEmojiMd(
+        inputs.customEmoji,
+        content: finalContent,
+        size: emojiSize,
+      ).pattern,
       _ChannelLinkMd(
         channelNames: inputs.channelNames,
         onChannelTap: channelTap,
-      ),
-      _MarkdownLabelEntityMd(),
-      _MarkdownEscapeMd(),
-      for (final component in MarkdownComponent.inlineComponents)
-        component is ATagMd ? _MessageLinkMd() : component,
+      ).pattern,
     ],
     [inputs],
   );
@@ -144,41 +144,3 @@ class _InlineComponentInputs {
   int get hashCode =>
       Object.hash(content, finalContent, emojiSize, hasMentionHandler);
 }
-
-// Encoded label punctuation must remain literal even inside nested formatting.
-class _MarkdownLabelEntityMd extends InlineMd {
-  @override
-  RegExp get exp => RegExp(r'&#(?:91|92|93);');
-
-  @override
-  Set<MarkdownScope> get scopes => const {MarkdownScope.linkLabel};
-
-  @override
-  InlineSpan span(
-    BuildContext context,
-    String text,
-    GptMarkdownConfig config,
-  ) => TextSpan(text: decodeMarkdownLabelSyntax(text), style: config.style);
-}
-
-// Backslash-escaped characters, held by shieldMarkdownEscapes so no other
-// syntax matches them, shown as themselves.
-class _MarkdownEscapeMd extends InlineMd {
-  @override
-  RegExp get exp => RegExp('(?:${shieldedMarkdownEscape.pattern})+');
-
-  @override
-  InlineSpan span(
-    BuildContext context,
-    String text,
-    GptMarkdownConfig config,
-  ) => TextSpan(text: unshieldMarkdownEscapes(text), style: config.style);
-}
-
-// Labels are already normalized, including literal/nested brackets. The
-// package's permissive .*? label matcher can otherwise start at a prose '['
-// and swallow a later image or link before its own component gets to render.
-class _MessageLinkMd extends ATagMd {
-  @override
-  RegExp get exp => RegExp(r'(?<!!)\[[^\[\]\n]*\]\([^\s]*\)');
-}

mobile/lib/features/channels/message_content/link_normalizer.dart

@@ -186,15 +186,45 @@ bool _hasInlineCloserOnLine(String content, int start, int delimiterLength) {
   return false;
 }
 
+/// Whether a closing run follows before the paragraph ends: a code span
+/// cannot continue past a blank line, a list item, heading, fence or a deeper
+/// block quote.
 bool _hasInlineCloser(String content, int start, int delimiterLength) {
+  final end = _paragraphEnd(content, start);
   for (final run in _backtickRunPattern.allMatches(content, start)) {
+    if (run.start >= end) return false;
     if (run.end - run.start == delimiterLength) return true;
   }
   return false;
 }
 
-// Numeric entities shield literal label punctuation from gpt_markdown's
-// bracket-blind image/link matchers and its earlier \[...\] math pass.
+int _paragraphEnd(String content, int start) {
+  // A block quote can interrupt a paragraph, so a line opening a deeper quote
+  // than the paragraph's own line ends it. Equal or shallower markers continue
+  // it, as in a multi-line code span inside a quote.
+  final openerDepth = _lineQuoteDepth(
+    content,
+    start == 0 ? 0 : content.lastIndexOf('\n', start - 1) + 1,
+  );
+  var newline = content.indexOf('\n', start);
+  while (newline >= 0) {
+    if (_lineQuoteDepth(content, newline + 1) > openerDepth ||
+        _paragraphBreak.matchAsPrefix(content, newline + 1) != null) {
+      return newline;
+    }
+    newline = content.indexOf('\n', newline + 1);
+  }
+  return content.length;
+}
+
+final _paragraphBreak = RegExp(
+  r'[ \t]*(?:$|[-*+][ \t]|\d{1,9}[.)][ \t]|#{1,6}(?:[ \t]|$)|`{3}|~{3})',
+  multiLine: true,
+);
+
+// Numeric entities keep literal label punctuation out of the line-based
+// attachment, gallery and preview matchers, which do not balance brackets.
+// The renderer decodes them.
 String _normalizeMarkdownLabel(String label) => label.replaceAllMapped(
   RegExp(r'\\([\\\[\]])|[\[\]]'),
   (match) => '&#${(match[1] ?? match[0]!).codeUnitAt(0)};',
@@ -349,10 +379,11 @@ String _normalizeProseLinks(String segment) {
   return normalized;
 }
 
-// An autolink keeps its backslashes literally. Encode them so none is read
-// as an escape of the synthesized `]` or `)`.
+// An autolink keeps its backslashes literally. Escape them so none is read
+// as an escape of the synthesized `]` or `)`. A numeric entity would show
+// verbatim: gpt_markdown leaves a URL-shaped label undecoded.
 String _autolink(String url) =>
-    '[${url.replaceAll(r'\', '\')}](${url.replaceAll(r'\', '%5C')})';
+    '[${url.replaceAll(r'\', r'\\')}](${url.replaceAll(r'\', '%5C')})';
 
 String _normalizeBareLink(String segment, Match match) {
   final matched = match[0]!;

mobile/lib/features/channels/message_content/markdown_escapes.dart

@@ -52,3 +52,30 @@ const _marker = '\u2060';
 const _fullwidthOffset = 0xFEE0;
 // CommonMark: a backslash escapes any ASCII punctuation character.
 final _escape = RegExp(r'\\([!-/:-@\[-`{-~])|\u2060');
+
+/// [content] with `\[`, `\]`, `\(` and `\)` outside code written as numeric
+/// entities.
+///
+/// gpt_markdown reads those pairs as LaTeX delimiters, but in CommonMark, and
+/// on desktop, they are escapes for literal brackets and parentheses. The
+/// renderer decodes the entities as text, so `\[not a link\](x)` shows
+/// `[not a link](x)` instead of a formula. `\\` stays an escaped backslash.
+String escapeMarkdownMathDelimiters(String content) =>
+    _mathDelimiter.hasMatch(content)
+    ? mapMarkdownProse(
+        content,
+        (prose) => prose.replaceAllMapped(
+          _backslashPair,
+          (m) => switch (m[1]) {
+            '[' => '[',
+            ']' => ']',
+            '(' => '(',
+            ')' => ')',
+            _ => m[0]!,
+          },
+        ),
+      )
+    : content;
+
+final _mathDelimiter = RegExp(r'\\[\[\]()]');
+final _backslashPair = RegExp(r'\\(.)', dotAll: true);