@bendyline/squisq-editor-react 2.1.0 → 2.2.0

package/src/tiptapBridge.ts

@@ -1,986 +0,0 @@
-/**
- * Tiptap Bridge
- *
- * Conversion utilities between raw markdown source and Tiptap's JSON/HTML
- * content format. Uses a lightweight HTML-based approach: we convert markdown
- * to a simple HTML representation that Tiptap can consume, and parse
- * Tiptap's HTML output back to markdown.
- *
- * This bridge preserves markdown semantics much better than going through
- * Tiptap's native markdown extension, since we control the conversion
- * using squisq's own parser.
- */
-
-import { resolveIcon } from '@bendyline/squisq/icons';
-import {
-  matchTrailingTemplateAnnotation,
-  matchTrailingPandocAttr,
-  tokenizeAttrTokens,
-} from '@bendyline/squisq/markdown';
-
-// Hoisted regex patterns for inline markdown ↔ HTML conversion
-const RE_BOLD_STAR = /\*\*(.+?)\*\*/g;
-const RE_BOLD_UNDER = /__(.+?)__/g;
-const RE_ITALIC_STAR = /\*(.+?)\*/g;
-const RE_ITALIC_UNDER = /_(.+?)_/g;
-const RE_STRIKETHROUGH = /~~(.+?)~~/g;
-const RE_INLINE_CODE = /`(.+?)`/g;
-const RE_LINK = /\[(.+?)\]\((.+?)\)/g;
-// `*?` on the alt — an empty alt (`![](foo.png)`) is valid markdown and
-// the most common shape for pasted/uploaded images that don't yet have
-// a human-picked caption. Previously required at least one alt char,
-// which dropped those images on the floor during markdown→HTML.
-const RE_IMAGE = /!\[(.*?)\]\((.+?)\)/g;
-// Mentions: `@[Display](scheme:id)` — scheme-part must start with a letter
-// so plain `$100` or price-style parentheticals don't accidentally match.
-// remark-stringify may round-trip the colon as `\:` — tolerate either.
-const RE_MENTION = /@\[([^\]]+?)\]\(([a-z][a-z0-9+.-]*)\\?:([^)\s]+)\)/gi;
-const RE_MENTION_TAG = /<span\b[^>]*?\bdata-mention\b[^>]*?>(?:<[^>]+>)*([^<]*)<\/span>/gi;
-
-// Inline FontAwesome icon. Markdown form: `{[github]}` (bare) or
-// `{[fa-solid:user]}` (qualified). HTML form (produced by the
-// `InlineIcon` Tiptap node and consumed back by `htmlToInline`):
-// `<i data-icon="github" data-family="brands" data-name="github"
-//     class="fa-brands fa-github" contenteditable="false"></i>`.
-const RE_ICON_MD = /\{\[([a-zA-Z0-9_:-]+)\]\}/g;
-const RE_ICON_TAG = /<i\b[^>]*?\bdata-icon="([^"]*)"[^>]*?><\/i>/gi;
-const RE_STRONG_TAG = /<strong>(.*?)<\/strong>/g;
-const RE_B_TAG = /<b>(.*?)<\/b>/g;
-const RE_EM_TAG = /<em>(.*?)<\/em>/g;
-const RE_I_TAG = /<i>(.*?)<\/i>/g;
-const RE_S_TAG = /<s>(.*?)<\/s>/g;
-const RE_DEL_TAG = /<del>(.*?)<\/del>/g;
-const RE_CODE_TAG = /<code>(.*?)<\/code>/g;
-const RE_A_TAG = /<a[^>]+href="([^"]*)"[^>]*>(.*?)<\/a>/g;
-// Matches any `<img>` tag and captures its `src` + `alt` regardless of
-// attribute order. TipTap's Image extension renders `<img src="..."
-// alt="...">` (src first), while some other producers — including our
-// own `markdownToTiptap` conversion — emit alt-first. The previous
-// regex required alt-before-src and silently dropped every src-first
-// image; `RE_STRIP_TAGS` below would then delete the unmatched tag,
-// so the outgoing markdown had no image reference at all.
-const RE_IMG_TAG = /<img\b([^>]*)>/g;
-const RE_STRIP_TAGS = /<[^>]+>/g;
-
-/**
- * Convert raw markdown source to Tiptap-consumable HTML content.
- * Uses a simple markdown-to-HTML conversion that maps cleanly to
- * Tiptap's ProseMirror schema.
- */
-export function markdownToTiptap(markdown: string): string {
-  if (!markdown.trim()) return '<p></p>';
-
-  // Normalize line endings — content from zip archives may use \r\n
-  const html = markdown.replace(/\r\n/g, '\n').replace(/\r/g, '\n');
-
-  // Process blocks line by line for accurate conversion
-  const lines = html.split('\n');
-  const outputBlocks: string[] = [];
-  let inCodeBlock = false;
-  let codeBlockLang = '';
-  let codeBlockLines: string[] = [];
-  let inList = false;
-  let listItems: string[] = [];
-  let listType: 'ul' | 'ol' | 'task' = 'ul';
-  let inTable = false;
-  let tableLines: string[] = [];
-  let pendingBlankLines = 0;
-
-  const flushPendingBlankParagraphs = () => {
-    if (pendingBlankLines === 0) return;
-    const emptyParagraphs =
-      outputBlocks.length === 0 ? pendingBlankLines : Math.max(0, pendingBlankLines - 1);
-    for (let i = 0; i < emptyParagraphs; i++) {
-      outputBlocks.push('<p></p>');
-    }
-    pendingBlankLines = 0;
-  };
-
-  const pushBlock = (block: string) => {
-    flushPendingBlankParagraphs();
-    outputBlocks.push(block);
-  };
-
-  const flushList = () => {
-    if (inList && listItems.length > 0) {
-      const tag = listType === 'ol' ? 'ol' : 'ul';
-      const attr = listType === 'task' ? ' data-type="taskList"' : '';
-      pushBlock(`<${tag}${attr}>${listItems.join('')}</${tag}>`);
-      listItems = [];
-      inList = false;
-    }
-  };
-
-  const flushTable = () => {
-    if (!inTable || tableLines.length === 0) {
-      inTable = false;
-      tableLines = [];
-      return;
-    }
-
-    // Validate: need at least 2 lines and second must be a separator
-    const separatorCells = tableLines.length >= 2 ? parseTableCells(tableLines[1]) : [];
-    const isSeparator =
-      separatorCells.length > 0 && separatorCells.every((cell) => /^:?-+:?$/.test(cell.trim()));
-
-    if (tableLines.length < 2 || !isSeparator) {
-      // Not a valid table — render accumulated lines as paragraphs
-      for (const tl of tableLines) {
-        pushBlock(`<p>${inlineToHtml(tl)}</p>`);
-      }
-      inTable = false;
-      tableLines = [];
-      return;
-    }
-
-    const alignments = parseAlignments(tableLines[1]);
-    const headerCells = parseTableCells(tableLines[0]);
-
-    // Build header row
-    const thHtml = headerCells
-      .map((cell, i) => {
-        const align = alignments[i];
-        const style = align ? ` style="text-align: ${align}"` : '';
-        return `<th${style}>${inlineToHtml(cell)}</th>`;
-      })
-      .join('');
-
-    // Build body rows
-    const bodyHtml = tableLines
-      .slice(2)
-      .map((rowLine) => {
-        const cells = parseTableCells(rowLine);
-        const tdHtml = cells
-          .map((cell, i) => {
-            const align = alignments[i];
-            const style = align ? ` style="text-align: ${align}"` : '';
-            return `<td${style}>${inlineToHtml(cell)}</td>`;
-          })
-          .join('');
-        return `<tr>${tdHtml}</tr>`;
-      })
-      .join('');
-
-    pushBlock(`<table><thead><tr>${thHtml}</tr></thead><tbody>${bodyHtml}</tbody></table>`);
-
-    inTable = false;
-    tableLines = [];
-  };
-
-  for (let i = 0; i < lines.length; i++) {
-    const line = lines[i];
-
-    // Code fence handling
-    if (line.startsWith('```')) {
-      if (!inCodeBlock) {
-        flushList();
-        flushPendingBlankParagraphs();
-        inCodeBlock = true;
-        codeBlockLang = line.slice(3).trim();
-        codeBlockLines = [];
-        continue;
-      } else {
-        const langAttr = codeBlockLang ? ` class="language-${escapeHtml(codeBlockLang)}"` : '';
-        pushBlock(`<pre><code${langAttr}>${escapeHtml(codeBlockLines.join('\n'))}</code></pre>`);
-        inCodeBlock = false;
-        codeBlockLang = '';
-        codeBlockLines = [];
-        continue;
-      }
-    }
-
-    if (inCodeBlock) {
-      codeBlockLines.push(line);
-      continue;
-    }
-
-    // If in table and current line is not a table row, flush
-    if (inTable && !/^\|.*\|$/.test(line.trim())) {
-      flushTable();
-    }
-
-    // Blank line flushes list
-    if (line.trim() === '') {
-      flushList();
-      pendingBlankLines++;
-      continue;
-    }
-
-    flushPendingBlankParagraphs();
-
-    // Headings
-    const headingMatch = line.match(/^(#{1,6})\s+(.+)$/);
-    if (headingMatch) {
-      flushList();
-      const level = headingMatch[1].length;
-      let text = headingMatch[2];
-      let attrs = '';
-
-      // Peel off trailing brace-blocks in any order: the squisq-native
-      // `{[template …]}` annotation and the Pandoc `{#id .class key=value}`
-      // attribute block may both appear at the end of the heading line.
-      // Loop until neither matches. The matchers are imported from core so
-      // this stays in sync with packages/core/src/markdown/convert.ts.
-      let templateInner: string | null = null;
-      let pandocInner: string | null = null;
-      for (let pass = 0; pass < 4; pass++) {
-        let matched = false;
-        if (templateInner == null) {
-          const m = matchTrailingTemplateAnnotation(text);
-          if (m) {
-            templateInner = m.inner.trim();
-            text = text.slice(0, m.index).trimEnd();
-            matched = true;
-          }
-        }
-        if (pandocInner == null) {
-          const m = matchTrailingPandocAttr(text);
-          if (m) {
-            pandocInner = m.inner.trim();
-            text = text.slice(0, m.index).trimEnd();
-            matched = true;
-          }
-        }
-        if (!matched) break;
-      }
-
-      if (templateInner != null) {
-        // Quote-aware tokenization (shared with the core parser) keeps a
-        // quoted value like caption="Beach at sunset" as one token. The
-        // tokens are stored raw — quotes included — so tiptapToMarkdown
-        // can re-join them into the annotation verbatim.
-        const tokens = tokenizeAttrTokens(templateInner);
-        if (tokens.length === 0) {
-          // Preserve an authored `{[]}` through the editable Tiptap document
-          // while keeping the raw marker out of the visible heading text.
-          attrs += ' data-template-empty="true"';
-        }
-        const firstIsParam = tokens.length > 0 && tokens[0].indexOf('=') > 0;
-        if (!firstIsParam && tokens[0]) {
-          attrs += ` data-template="${escapeHtml(tokens[0])}"`;
-        }
-        const params = tokens.slice(firstIsParam ? 0 : 1).filter((t) => t.includes('='));
-        if (params.length > 0) {
-          attrs += ` data-template-params="${escapeHtml(params.join(' '))}"`;
-        }
-      }
-      if (pandocInner != null) {
-        attrs += ` data-block-attrs="${escapeHtml(pandocInner)}"`;
-      }
-
-      pushBlock(`<h${level}${attrs}>${inlineToHtml(text)}</h${level}>`);
-      continue;
-    }
-
-    // Thematic break
-    if (/^(---|\*\*\*|___)(\s*)$/.test(line.trim())) {
-      flushList();
-      pushBlock('<hr>');
-      continue;
-    }
-
-    // Blockquote
-    if (line.startsWith('> ')) {
-      flushList();
-      const quoteLines = [line.slice(2)];
-      while (i + 1 < lines.length && lines[i + 1].startsWith('> ')) {
-        i++;
-        quoteLines.push(lines[i].slice(2));
-      }
-      pushBlock(
-        `<blockquote>${quoteLines.map((quoteLine) => `<p>${inlineToHtml(quoteLine)}</p>`).join('')}</blockquote>`,
-      );
-      continue;
-    }
-
-    // Task list item. Text is optional so freshly-typed, still-empty tasks
-    // survive a round-trip. The checkbox itself is drawn by TaskItem's node
-    // view from the data-checked attribute, so we only emit the text content
-    // (in a paragraph, matching Tiptap's own node structure).
-    const taskMatch = line.match(/^[-*+]\s+\[([xX ])\]\s*(.*)$/);
-    if (taskMatch) {
-      if (!inList || listType !== 'task') {
-        flushList();
-        inList = true;
-        listType = 'task';
-      }
-      const checkedAttr = taskMatch[1].toLowerCase() === 'x' ? ' data-checked="true"' : '';
-      listItems.push(
-        `<li data-type="taskItem"${checkedAttr}><p>${inlineToHtml(taskMatch[2])}</p></li>`,
-      );
-      continue;
-    }
-
-    // Unordered list item
-    const ulMatch = line.match(/^[-*+]\s+(.+)$/);
-    if (ulMatch) {
-      if (!inList || listType !== 'ul') {
-        flushList();
-        inList = true;
-        listType = 'ul';
-      }
-      listItems.push(`<li><p>${inlineToHtml(ulMatch[1])}</p></li>`);
-      continue;
-    }
-
-    // Ordered list item
-    const olMatch = line.match(/^\d+\.\s+(.+)$/);
-    if (olMatch) {
-      if (!inList || listType !== 'ol') {
-        flushList();
-        inList = true;
-        listType = 'ol';
-      }
-      listItems.push(`<li><p>${inlineToHtml(olMatch[1])}</p></li>`);
-      continue;
-    }
-
-    // Table row
-    if (/^\|.*\|$/.test(line.trim())) {
-      if (!inTable) {
-        flushList();
-        inTable = true;
-        tableLines = [];
-      }
-      tableLines.push(line);
-      continue;
-    }
-
-    // Standalone image — emit as a top-level block `<img>` instead of
-    // wrapping in `<p>`. The Tiptap Image extension is configured with
-    // `inline: false`, so `<p><img></p>` parses to an empty paragraph
-    // (the block image can't live inside the paragraph). That bug
-    // manifested as a broken-image glyph after a markdown → WYSIWYG
-    // round-trip for any dropped/pasted image.
-    const standaloneImageMatch = line.trim().match(/^!\[(.*?)\]\((.+?)\)$/);
-    if (standaloneImageMatch) {
-      flushList();
-      const alt = escapeHtml(standaloneImageMatch[1] ?? '');
-      const src = escapeHtml(standaloneImageMatch[2] ?? '');
-      pushBlock(`<img alt="${alt}" src="${src}">`);
-      continue;
-    }
-
-    // Standalone raw HTML `<img>` line — emitted by `tiptapToMarkdown`
-    // when the user has resized an image (width/height attrs are
-    // serialized as HTML rather than markdown shorthand so the
-    // dimensions survive round-trip). Pass the tag through unchanged
-    // so Tiptap's Image extension parses width/height attributes.
-    const trimmed = line.trim();
-    if (/^<img\b[^>]*>$/i.test(trimmed)) {
-      flushList();
-      pushBlock(trimmed);
-      continue;
-    }
-
-    // Standalone `<video>` / `<audio>` line — emitted by the recorder
-    // (RecorderEntry) when it saves a clip, and by `tiptapToMarkdown`
-    // when the WYSIWYG editor's TiptapVideo/TiptapAudio nodes serialize
-    // back to markdown. Pass through unchanged so the editor's parseHTML
-    // picks up the tag attributes (`src`, `controls`, …).
-    if (/^<(?:video|audio)\b[^>]*>(?:[\s\S]*?<\/(?:video|audio)>)?$/i.test(trimmed)) {
-      flushList();
-      pushBlock(trimmed);
-      continue;
-    }
-
-    // Regular paragraph
-    flushList();
-    pushBlock(`<p>${inlineToHtml(line)}</p>`);
-  }
-
-  // Close any remaining open blocks
-  if (inCodeBlock) {
-    const langAttr = codeBlockLang ? ` class="language-${escapeHtml(codeBlockLang)}"` : '';
-    pushBlock(`<pre><code${langAttr}>${escapeHtml(codeBlockLines.join('\n'))}</code></pre>`);
-  }
-  flushList();
-  flushTable();
-  flushPendingBlankParagraphs();
-
-  return outputBlocks.join('') || '<p></p>';
-}
-
-/**
- * Convert Tiptap HTML output back to markdown source.
- * Extracts semantic structure from HTML and produces clean markdown.
- */
-export function tiptapToMarkdown(html: string): string {
-  if (!html || html === '<p></p>') return '';
-
-  const lines: string[] = [];
-
-  // Simple regex-based HTML to markdown conversion
-  // This works because Tiptap produces clean, predictable HTML
-  let remaining = html;
-
-  while (remaining.length > 0) {
-    // Headings
-    const headingMatch = remaining.match(/^<h([1-6])([^>]*)>(.*?)<\/h\1>/s);
-    if (headingMatch) {
-      const level = parseInt(headingMatch[1], 10);
-      const attrs = headingMatch[2];
-      const headingHtml = headingMatch[3];
-      // Template/property badges are editor chrome, not heading content. Strip
-      // them structurally before converting the inline HTML. Older builds
-      // briefly rendered the template label as real badge text; using the
-      // markup boundary preserves that cleanup without mistaking a legitimate
-      // heading suffix (for example "A Famous Quote") for the badge label.
-      const chromeStart = headingHtml.search(
-        /<span\b[^>]*\bclass="[^"]*\bsquisq-(?:template|props)-badge\b[^"]*"[^>]*>/i,
-      );
-      let text = htmlToInline(chromeStart >= 0 ? headingHtml.slice(0, chromeStart) : headingHtml);
-
-      // Re-inject heading annotations from data attributes. Canonical
-      // emit order: Pandoc `{#…}` first, then squisq `{[…]}` annotation
-      // (matches blockToMdast in core/markdown/convert.ts).
-      const blockAttrsMatch = attrs.match(/data-block-attrs="([^"]*)"/);
-      const tmplMatch = attrs.match(/data-template="([^"]+)"/);
-      const paramsMatch = attrs.match(/data-template-params="([^"]+)"/);
-      const hasEmptyTemplateAnnotation = /\sdata-template-empty(?:="[^"]*")?/.test(attrs);
-      if (blockAttrsMatch) {
-        const inner = unescapeHtml(blockAttrsMatch[1]);
-        text += ` {${inner}}`;
-      }
-      if (tmplMatch || paramsMatch) {
-        let annotation = tmplMatch ? tmplMatch[1] : '';
-        if (paramsMatch) {
-          annotation += (annotation ? ' ' : '') + unescapeHtml(paramsMatch[1]);
-        }
-        text += ` {[${annotation}]}`;
-      } else if (hasEmptyTemplateAnnotation) {
-        text += ' {[]}';
-      }
-
-      lines.push('#'.repeat(level) + ' ' + text);
-      lines.push('');
-      remaining = remaining.slice(headingMatch[0].length);
-      continue;
-    }
-
-    // Code blocks — tolerate attributes on <pre> and <code>: real
-    // editor.getHTML() output carries class="squisq-code-block" on the
-    // <pre> (StarterKit codeBlock HTMLAttributes), which a bare
-    // `<pre><code` anchor never matches, silently dropping the fence.
-    const codeMatch = remaining.match(/^<pre\b[^>]*><code\b([^>]*)>(.*?)<\/code><\/pre>/s);
-    if (codeMatch) {
-      const lang = /\bclass="language-([^"]*)"/.exec(codeMatch[1] ?? '')?.[1] ?? '';
-      const code = unescapeHtml(codeMatch[2]);
-      lines.push('```' + lang);
-      lines.push(code);
-      lines.push('```');
-      lines.push('');
-      remaining = remaining.slice(codeMatch[0].length);
-      continue;
-    }
-
-    // Blockquote
-    const bqMatch = remaining.match(/^<blockquote>(.*?)<\/blockquote>/s);
-    if (bqMatch) {
-      const paragraphs = bqMatch[1]
-        .split(/<\/p>\s*<p[^>]*>/i)
-        .map((paragraph) => paragraph.replace(/^<p[^>]*>/i, '').replace(/<\/p>\s*$/i, ''));
-      for (const paragraph of paragraphs) {
-        for (const quoteLine of htmlToInline(paragraph).split('\n')) {
-          lines.push('> ' + quoteLine);
-        }
-      }
-      const next = remaining.slice(bqMatch[0].length);
-      // Older bridge output and direct Tiptap edits can still leave adjacent
-      // blockquote nodes. Keep those nodes adjacent when serializing;
-      // inserting the normal block separator here turns one quote into two
-      // paragraphs every time the document passes through WYSIWYG mode.
-      if (!/^\s*<blockquote>/.test(next)) lines.push('');
-      remaining = next;
-      continue;
-    }
-
-    // Horizontal rule
-    if (
-      remaining.startsWith('<hr>') ||
-      remaining.startsWith('<hr/>') ||
-      remaining.startsWith('<hr />')
-    ) {
-      const hrMatch = remaining.match(/^<hr\s*\/?>/);
-      lines.push('---');
-      lines.push('');
-      remaining = remaining.slice(hrMatch![0].length);
-      continue;
-    }
-
-    // Table (with optional Tiptap tableWrapper div; table tag may have style attrs)
-    const tableMatch =
-      remaining.match(
-        /^<div[^>]*class="[^"]*tableWrapper[^"]*"[^>]*><table[^>]*>(.*?)<\/table>\s*<\/div>/s,
-      ) || remaining.match(/^<table[^>]*>(.*?)<\/table>/s);
-    if (tableMatch) {
-      const tableContent = tableMatch[1];
-
-      // Extract all rows with their cells
-      const rows: { content: string; align: string | null; isHeader: boolean }[][] = [];
-      const rowRegex = /<tr[^>]*>(.*?)<\/tr>/gs;
-      let rowExec;
-      while ((rowExec = rowRegex.exec(tableContent)) !== null) {
-        const rowHtml = rowExec[1];
-        const cells: { content: string; align: string | null; isHeader: boolean }[] = [];
-        const cellRegex = /<(th|td)([^>]*)>(.*?)<\/\1>/gs;
-        let cellExec;
-        while ((cellExec = cellRegex.exec(rowHtml)) !== null) {
-          const tag = cellExec[1];
-          const attrs = cellExec[2];
-          const content = htmlToInline(cellExec[3].replace(/<\/?p>/g, ''));
-          const alignExec = attrs.match(/text-align:\s*(left|center|right)/);
-          cells.push({
-            content,
-            align: alignExec ? alignExec[1] : null,
-            isHeader: tag === 'th',
-          });
-        }
-        if (cells.length > 0) {
-          rows.push(cells);
-        }
-      }
-
-      if (rows.length > 0) {
-        // Header row = first row with th cells, or just the first row
-        const headerIdx = rows.findIndex((r) => r.some((c) => c.isHeader));
-        const hIdx = headerIdx >= 0 ? headerIdx : 0;
-        const headerRow = rows[hIdx];
-        const dataRows = rows.filter((_, i) => i !== hIdx);
-
-        const aligns = headerRow.map((c) => c.align);
-        lines.push('| ' + headerRow.map((c) => c.content || ' ').join(' | ') + ' |');
-        lines.push(
-          '| ' +
-            aligns
-              .map((a) => {
-                if (a === 'center') return ':---:';
-                if (a === 'right') return '---:';
-                return '---';
-              })
-              .join(' | ') +
-            ' |',
-        );
-        for (const row of dataRows) {
-          lines.push('| ' + row.map((c) => c.content || ' ').join(' | ') + ' |');
-        }
-        lines.push('');
-      }
-
-      remaining = remaining.slice(tableMatch[0].length);
-      continue;
-    }
-
-    // Task list
-    const taskListMatch = remaining.match(/^<ul[^>]*data-type="taskList"[^>]*>(.*?)<\/ul>/s);
-    if (taskListMatch) {
-      const items = taskListMatch[1].matchAll(/<li[^>]*data-type="taskItem"[^>]*>.*?<\/li>/gs);
-      for (const item of items) {
-        // Read checked state from the attribute value only — a bare
-        // includes('checked') matches the substring in data-checked="false".
-        const checked = /data-checked="true"/.test(item[0]);
-        // Drop the checkbox chrome (<label>…</label>) before reading text;
-        // Tiptap puts the task text in the trailing content node, not the label.
-        const body = item[0].replace(/<label\b[^>]*>.*?<\/label>/s, '');
-        const text = htmlToInline(body.replace(/<[^>]+>/g, '').trim());
-        lines.push(`- [${checked ? 'x' : ' '}] ${text}`.trimEnd());
-      }
-      lines.push('');
-      remaining = remaining.slice(taskListMatch[0].length);
-      continue;
-    }
-
-    // Unordered list
-    const ulMatch = remaining.match(/^<ul>(.*?)<\/ul>/s);
-    if (ulMatch) {
-      const items = ulMatch[1].matchAll(/<li>(.*?)<\/li>/gs);
-      for (const item of items) {
-        lines.push(...renderListItem('- ', item[1]));
-      }
-      lines.push('');
-      remaining = remaining.slice(ulMatch[0].length);
-      continue;
-    }
-
-    // Ordered list
-    const olMatch = remaining.match(/^<ol[^>]*>(.*?)<\/ol>/s);
-    if (olMatch) {
-      const items = [...olMatch[1].matchAll(/<li>(.*?)<\/li>/gs)];
-      items.forEach((item, idx) => {
-        lines.push(...renderListItem(`${idx + 1}. `, item[1]));
-      });
-      lines.push('');
-      remaining = remaining.slice(olMatch[0].length);
-      continue;
-    }
-
-    // Paragraph
-    const pMatch = remaining.match(/^<p>(.*?)<\/p>/s);
-    if (pMatch) {
-      const text = htmlToInline(pMatch[1]);
-      if (text.trim()) {
-        lines.push(text);
-        lines.push('');
-      } else {
-        lines.push('');
-      }
-      remaining = remaining.slice(pMatch[0].length);
-      continue;
-    }
-
-    // Block-level image. TipTap's Image extension with `inline: false`
-    // emits `<img src alt>` as a bare top-level element (no wrapping
-    // `<p>`). Without this handler the skip-unknown-tags catch-all
-    // below silently drops the image from the outgoing markdown —
-    // the bug that made the chat composer ship image-less messages
-    // even though the editor showed the picture. Handled here,
-    // before the inline walker ever sees it.
-    const imgMatch = remaining.match(/^<img\b([^>]*)>/);
-    if (imgMatch) {
-      const attrs = imgMatch[1] ?? '';
-      const src = /\bsrc="([^"]*)"/i.exec(attrs)?.[1];
-      if (src) {
-        const alt = /\balt="([^"]*)"/i.exec(attrs)?.[1] ?? '';
-        lines.push(serializeImage(src, alt, attrs));
-        lines.push('');
-      }
-      remaining = remaining.slice(imgMatch[0].length);
-      continue;
-    }
-
-    // Block-level `<video>` / `<audio>` — emitted by our TiptapVideo /
-    // TiptapAudio atom nodes (block group). Serialize the whole tag
-    // (opening + closing) back to markdown unchanged; CommonMark allows
-    // inline HTML, and our renderer plus the InlinePreviewGutter both
-    // know how to parse the htmlElement back out.
-    const mediaMatch = remaining.match(/^<(video|audio)\b([^>]*)>(?:[\s\S]*?<\/\1>)?/);
-    if (mediaMatch) {
-      const tag = mediaMatch[1] === 'video' ? 'video' : 'audio';
-      lines.push(serializeMediaTag(tag, mediaMatch[2] ?? ''));
-      lines.push('');
-      remaining = remaining.slice(mediaMatch[0].length);
-      continue;
-    }
-
-    // Skip unknown tags or whitespace
-    const skipMatch = remaining.match(/^(<[^>]+>|\s+)/);
-    if (skipMatch) {
-      remaining = remaining.slice(skipMatch[0].length);
-      continue;
-    }
-
-    // Plain text (shouldn't happen in well-formed Tiptap output)
-    const textMatch = remaining.match(/^([^<]+)/);
-    if (textMatch) {
-      lines.push(unescapeHtml(textMatch[1]));
-      remaining = remaining.slice(textMatch[0].length);
-      continue;
-    }
-
-    // Safety: skip one character to avoid infinite loop
-    remaining = remaining.slice(1);
-  }
-
-  return lines.join('\n');
-}
-
-/**
- * Render a list item's HTML content as one or more markdown lines.
- * Handles `<p>` paragraph breaks (blank line) and `<br>` hard breaks
- * (two trailing spaces). Continuation lines are indented to keep them
- * inside the list item.
- */
-function renderListItem(prefix: string, html: string): string[] {
-  const indent = ' '.repeat(prefix.length);
-
-  // Pull out any block-level media nested in the item (e.g. a recording
-  // dropped onto a list bullet, or a clip dragged into one). The inline
-  // paragraph walk below ignores `<video>` / `<audio>` tags, so without
-  // this they'd be silently dropped on serialize and lost from the
-  // markdown source. We collect them and re-emit each as an indented
-  // continuation line so they stay inside the list item and round-trip
-  // through the markdown parser's block-media handler.
-  const media: string[] = [];
-  const htmlWithoutMedia = html.replace(
-    /<(video|audio)\b([^>]*)>(?:[\s\S]*?<\/\1>)?/gi,
-    (_full, tag: string, attrs: string) => {
-      media.push(serializeMediaTag(tag.toLowerCase() === 'video' ? 'video' : 'audio', attrs ?? ''));
-      return '';
-    },
-  );
-
-  // Split on </p><p> to detect paragraph breaks within the item
-  const paragraphs = htmlWithoutMedia
-    .split(/<\/p>\s*<p[^>]*>/i)
-    .map((p) => p.replace(/^<p[^>]*>/i, '').replace(/<\/p>\s*$/i, ''));
-
-  const textLines: string[] = [];
-  paragraphs.forEach((paragraph, pIdx) => {
-    const inline = htmlToInline(paragraph).trim();
-    if (!inline) return;
-
-    // Each <br> already became "  \n" in htmlToInline; split on it now.
-    const subLines = inline.split('\n');
-    subLines.forEach((sub, sIdx) => {
-      if (pIdx === 0 && sIdx === 0) {
-        textLines.push(prefix + sub);
-      } else {
-        // Blank line separator between paragraphs (sIdx === 0 means new paragraph)
-        if (sIdx === 0) textLines.push('');
-        textLines.push(indent + sub);
-      }
-    });
-  });
-
-  if (textLines.length === 0 && media.length === 0) return [prefix];
-
-  // Text first (its first line carries the bullet marker), then each media
-  // tag as its own indented continuation block. If the item is media-only,
-  // the first tag takes the bullet so the item isn't emitted empty.
-  const result: string[] = [];
-  if (textLines.length > 0) {
-    result.push(...textLines);
-    for (const tag of media) {
-      result.push('');
-      result.push(indent + tag);
-    }
-  } else {
-    result.push(prefix + media[0]);
-    for (const tag of media.slice(1)) {
-      result.push('');
-      result.push(indent + tag);
-    }
-  }
-  return result;
-}
-
-// ─── Table helpers ───────────────────────────────────────
-
-/** Split a GFM table row into trimmed cell strings (strips outer pipes). */
-function parseTableCells(line: string): string[] {
-  let inner = line.trim();
-  if (inner.startsWith('|')) inner = inner.slice(1);
-  if (inner.endsWith('|')) inner = inner.slice(0, -1);
-  return inner.split('|').map((cell) => cell.trim());
-}
-
-/** Parse a GFM separator line into column alignments. */
-function parseAlignments(separatorLine: string): (string | null)[] {
-  return parseTableCells(separatorLine).map((cell) => {
-    const s = cell.replace(/\s/g, '');
-    if (s.startsWith(':') && s.endsWith(':')) return 'center';
-    if (s.endsWith(':')) return 'right';
-    if (s.startsWith(':')) return 'left';
-    return null;
-  });
-}
-
-// ─── Helpers ─────────────────────────────────────────────
-
-/**
- * Serialize a parsed `<img>` tag back to markdown. When the tag carries
- * an explicit `width` and/or `height` we emit a raw HTML `<img>` (the
- * markdown shorthand `![alt](src)` has no syntax for dimensions);
- * otherwise the friendlier shorthand is used. Markdown allows inline
- * HTML, so the HTML form parses and renders identically in any
- * CommonMark/GFM viewer.
- */
-/**
- * Serialize a `<video>` or `<audio>` tag (from the Tiptap atom node's
- * `renderHTML`) back to markdown. We re-emit only the attributes the
- * recorder + the renderer care about, in a stable order, so the
- * round-tripped markdown stays deterministic regardless of how Tiptap
- * decided to order them on its output.
- */
-function serializeMediaTag(tag: 'video' | 'audio', attrs: string): string {
-  const src = /\bsrc="([^"]*)"/i.exec(attrs)?.[1] ?? '';
-  const controls = /\bcontrols\b/i.test(attrs);
-  const width = /\bwidth="([^"]*)"/i.exec(attrs)?.[1];
-  const height = /\bheight="([^"]*)"/i.exec(attrs)?.[1];
-  const poster = tag === 'video' ? /\bposter="([^"]*)"/i.exec(attrs)?.[1] : undefined;
-  const parts = [`<${tag} src="${src}"`];
-  if (controls) parts.push(' controls');
-  if (width) parts.push(` width="${width}"`);
-  if (height) parts.push(` height="${height}"`);
-  if (poster) parts.push(` poster="${poster}"`);
-  parts.push(`></${tag}>`);
-  return parts.join('');
-}
-
-function serializeImage(src: string, alt: string, attrs: string): string {
-  const width = /\bwidth="([^"]*)"/i.exec(attrs)?.[1];
-  const height = /\bheight="([^"]*)"/i.exec(attrs)?.[1];
-  const title = /\btitle="([^"]*)"/i.exec(attrs)?.[1];
-  if (!width && !height) {
-    return `![${alt}](${src})`;
-  }
-  const parts = [`<img alt="${alt}" src="${src}"`];
-  if (width) parts.push(` width="${width}"`);
-  if (height) parts.push(` height="${height}"`);
-  if (title) parts.push(` title="${title}"`);
-  parts.push('>');
-  return parts.join('');
-}
-
-function escapeHtml(text: string): string {
-  return text
-    .replace(/&/g, '&amp;')
-    .replace(/</g, '&lt;')
-    .replace(/>/g, '&gt;')
-    .replace(/"/g, '&quot;');
-}
-
-function unescapeHtml(text: string): string {
-  return text
-    .replace(/&lt;/g, '<')
-    .replace(/&gt;/g, '>')
-    .replace(/&quot;/g, '"')
-    .replace(/&nbsp;|&#160;|&#xA0;/gi, '\u00a0')
-    .replace(/&amp;/g, '&');
-}
-
-/** Convert inline markdown to HTML for Tiptap consumption */
-function inlineToHtml(text: string): string {
-  // Extract images/mentions/links to opaque placeholders BEFORE running
-  // any inline-formatting regexes. Otherwise `_` characters inside a URL
-  // (e.g. `mikehome_files/IMG_6829.JPEG`) get turned into `<em>` tags by
-  // the underscore-italic rule, mangling the src so the image renders
-  // broken after a markdown ↔ WYSIWYG round-trip.
-  const placeholders: string[] = [];
-  const stash = (html: string): string => {
-    const token = `\u0000PH${placeholders.length}\u0000`;
-    placeholders.push(html);
-    return token;
-  };
-
-  // We run extraction on the RAW text (before escapeHtml) so the
-  // captured groups are the literal markdown contents; then we
-  // selectively escape the parts of each placeholder that need it.
-
-  // Images first: ![alt](src) — must be before links so the `!` prefix is consumed
-  let staged = text.replace(RE_IMAGE, (_m, alt, src) =>
-    stash(`<img alt="${escapeHtml(alt)}" src="${escapeHtml(src)}">`),
-  );
-
-  // Mentions: @[Display](scheme:id) — must run before links so the
-  // bracket+paren isn't consumed as a regular link.
-  staged = staged.replace(RE_MENTION, (_m, label, kind, id) =>
-    stash(
-      `<span data-mention="true" data-kind="${escapeHtml(kind)}" data-id="${escapeHtml(id)}" data-label="${escapeHtml(label)}" class="mention">@${escapeHtml(label)}</span>`,
-    ),
-  );
-
-  // FontAwesome inline icons: {[github]} / {[fa-solid:user]} — resolve
-  // against the FA catalog so unknown / ambiguous tokens stay as
-  // literal text (preserved through escapeHtml later). Stashing here
-  // alongside images/mentions keeps the `_` and other punctuation in
-  // the class attribute from being mangled by the underscore-italic
-  // regex below.
-  staged = staged.replace(RE_ICON_MD, (full, token) => {
-    const icon = resolveIcon(token);
-    if (!icon) return full; // leave literal text — author may have meant it
-    return stash(
-      `<i class="fa-${icon.family} fa-${icon.name}" data-icon="${escapeHtml(token)}" data-family="${icon.family}" data-name="${icon.name}" contenteditable="false"></i>`,
-    );
-  });
-
-  // Links: [text](url) — stash but keep the link text available to
-  // inline formatting by recursing.
-  staged = staged.replace(RE_LINK, (_m, linkText, href) =>
-    stash(`<a href="${escapeHtml(href)}">${inlineToHtml(linkText)}</a>`),
-  );
-
-  let result = escapeHtml(staged);
-
-  // Bold: **text** or __text__
-  result = result.replace(RE_BOLD_STAR, '<strong>$1</strong>');
-  result = result.replace(RE_BOLD_UNDER, '<strong>$1</strong>');
-
-  // Italic: *text* or _text_
-  result = result.replace(RE_ITALIC_STAR, '<em>$1</em>');
-  result = result.replace(RE_ITALIC_UNDER, '<em>$1</em>');
-
-  // Strikethrough: ~~text~~
-  result = result.replace(RE_STRIKETHROUGH, '<s>$1</s>');
-
-  // Inline code: `text`
-  result = result.replace(RE_INLINE_CODE, '<code>$1</code>');
-
-  // Restore placeholders. escapeHtml turned each `\u0000PHn\u0000` into
-  // the same string (the NULs and digits are escape-safe), so the
-  // restoration regex still matches.
-  // eslint-disable-next-line no-control-regex
-  result = result.replace(/\u0000PH(\d+)\u0000/g, (_m, idx) => placeholders[Number(idx)] ?? '');
-
-  return preserveLeadingSpaces(result);
-}
-
-/** Convert inline HTML back to markdown */
-function htmlToInline(html: string): string {
-  let result = html;
-
-  // Soft line breaks — convert <br> to GFM hard-break syntax (two trailing
-  // spaces + newline) before stripping tags so the newline survives.
-  result = result.replace(/<br\s*\/?>/gi, '  \n');
-
-  // FontAwesome inline icons — emit the original token. Must run before
-  // RE_I_TAG (which matches a bare `<i>` and would otherwise eat icon
-  // tags too). The token captured via data-icon already carries the
-  // qualified form when needed, so source round-trips exactly.
-  result = result.replace(RE_ICON_TAG, (_m, token) => `{[${token}]}`);
-
-  // Strong
-  result = result.replace(RE_STRONG_TAG, '**$1**');
-  result = result.replace(RE_B_TAG, '**$1**');
-
-  // Em
-  result = result.replace(RE_EM_TAG, '*$1*');
-  result = result.replace(RE_I_TAG, '*$1*');
-
-  // Strikethrough
-  result = result.replace(RE_S_TAG, '~~$1~~');
-  result = result.replace(RE_DEL_TAG, '~~$1~~');
-
-  // Code
-  result = result.replace(RE_CODE_TAG, '`$1`');
-
-  // Mentions — match before the link handler so the span isn't stripped
-  // out as an unknown tag. Pull kind + id out of the data attributes.
-  result = result.replace(RE_MENTION_TAG, (match, _inner) => {
-    const kind = /data-kind="([^"]*)"/i.exec(match)?.[1] ?? '';
-    const id = /data-id="([^"]*)"/i.exec(match)?.[1] ?? '';
-    const label = /data-label="([^"]*)"/i.exec(match)?.[1] ?? '';
-    if (!kind || !id || !label) return match;
-    return `@[${label}](${kind}:${id})`;
-  });
-
-  // Links
-  result = result.replace(RE_A_TAG, '[$2]($1)');
-
-  // Images — order-agnostic attribute parsing (tiptap emits src-first,
-  // our markdown-to-html emits alt-first; either must serialize back).
-  // When a width/height is present we serialize as raw HTML `<img>` so
-  // the dimensions survive the round-trip; otherwise the markdown
-  // shorthand `![alt](src)` is used.
-  result = result.replace(RE_IMG_TAG, (match, attrs: string) => {
-    const src = /\bsrc="([^"]*)"/i.exec(attrs)?.[1];
-    if (!src) return match;
-    const alt = /\balt="([^"]*)"/i.exec(attrs)?.[1] ?? '';
-    return serializeImage(src, alt, attrs);
-  });
-
-  // Strip remaining tags
-  result = result.replace(RE_STRIP_TAGS, '');
-
-  return restoreLeadingSpaces(unescapeHtml(result));
-}
-
-function preserveLeadingSpaces(html: string): string {
-  return html.replace(/^ +/, (spaces) => '&nbsp;'.repeat(spaces.length));
-}
-
-function restoreLeadingSpaces(text: string): string {
-  return text.replace(/(^|\n)(\u00a0+)/g, (_match, prefix: string, spaces: string) => {
-    return prefix + ' '.repeat(spaces.length);
-  });
-}