|
| 1 | +/** |
| 2 | + * Rehype plugin to restore limited HTML elements inside Markdown table cells. |
| 3 | + * |
| 4 | + * ## Problem |
| 5 | + * The remark/rehype pipeline neutralizes inline HTML as literal text |
| 6 | + * (remarkLiteralHtml) so that XML/HTML snippets in LLM responses display |
| 7 | + * as-is instead of being rendered. This causes <br> and <ul> markup in |
| 8 | + * table cells to show as plain text. |
| 9 | + * |
| 10 | + * ## Solution |
| 11 | + * This plugin traverses the HAST post-conversion, parses whitelisted HTML |
| 12 | + * patterns from text nodes, and replaces them with actual HAST element nodes |
| 13 | + * that will be rendered as real HTML. |
| 14 | + * |
| 15 | + * ## Supported HTML |
| 16 | + * - `<br>` / `<br/>` / `<br />` - Line breaks (inline) |
| 17 | + * - `<ul><li>...</li></ul>` - Unordered lists (block) |
| 18 | + * |
| 19 | + * ## Key Implementation Details |
| 20 | + * |
| 21 | + * ### 1. Sibling Combination (Critical) |
| 22 | + * The Markdown pipeline may fragment content across multiple text nodes and `<br>` |
| 23 | + * elements. For example, `<ul><li>a</li></ul>` might arrive as: |
| 24 | + * - Text: `"<ul>"` |
| 25 | + * - Element: `<br>` |
| 26 | + * - Text: `"<li>a</li></ul>"` |
| 27 | + * |
| 28 | + * We must combine consecutive text nodes and `<br>` elements into a single string |
| 29 | + * before attempting to parse list markup. Without this, list detection fails. |
| 30 | + * |
| 31 | + * ### 2. visitParents for Deep Traversal |
| 32 | + * Table cell content may be wrapped in intermediate elements (e.g., `<p>` tags). |
| 33 | + * Using `visitParents` instead of direct child iteration ensures we find text |
| 34 | + * nodes at any depth within the cell. |
| 35 | + * |
| 36 | + * ### 3. Reference Comparison for No-Op Detection |
| 37 | + * When checking if `<br>` expansion changed anything, we compare: |
| 38 | + * `expanded.length !== 1 || expanded[0] !== textNode` |
| 39 | + * |
| 40 | + * This catches both cases: |
| 41 | + * - Multiple nodes created (text was split) |
| 42 | + * - Single NEW node created (original had only `<br>`, now it's an element) |
| 43 | + * |
| 44 | + * A simple `length > 1` check would miss the single `<br>` case. |
| 45 | + * |
| 46 | + * ### 4. Strict List Validation |
| 47 | + * `parseList()` rejects malformed markup by checking for garbage text between |
| 48 | + * `<li>` elements. This prevents creating broken DOM from partial matches like |
| 49 | + * `<ul>garbage<li>a</li></ul>`. |
| 50 | + * |
| 51 | + * ### 5. Newline Substitution for `<br>` in Combined String |
| 52 | + * When combining siblings, existing `<br>` elements become `\n` in the combined |
| 53 | + * string. This allows list content to span visual lines while still being parsed |
| 54 | + * as a single unit. |
| 55 | + * |
| 56 | + * @example |
| 57 | + * // Input Markdown: |
| 58 | + * // | Feature | Notes | |
| 59 | + * // |---------|-------| |
| 60 | + * // | Multi-line | First<br>Second | |
| 61 | + * // | List | <ul><li>A</li><li>B</li></ul> | |
| 62 | + * // |
| 63 | + * // Without this plugin: <br> and <ul> render as literal text |
| 64 | + * // With this plugin: <br> becomes line break, <ul> becomes actual list |
| 65 | + */ |
| 66 | + |
| 67 | +import type { Plugin } from 'unified'; |
| 68 | +import type { Element, ElementContent, Root, Text } from 'hast'; |
| 69 | +import { visit } from 'unist-util-visit'; |
| 70 | +import { visitParents } from 'unist-util-visit-parents'; |
| 71 | +import { BR_PATTERN, LIST_PATTERN, LI_PATTERN } from '$lib/constants/table-html-restorer'; |
| 72 | + |
| 73 | +/** |
| 74 | + * Expands text containing `<br>` tags into an array of text nodes and br elements. |
| 75 | + */ |
| 76 | +function expandBrTags(value: string): ElementContent[] { |
| 77 | + const matches = [...value.matchAll(BR_PATTERN)]; |
| 78 | + if (!matches.length) return [{ type: 'text', value } as Text]; |
| 79 | + |
| 80 | + const result: ElementContent[] = []; |
| 81 | + let cursor = 0; |
| 82 | + |
| 83 | + for (const m of matches) { |
| 84 | + if (m.index! > cursor) { |
| 85 | + result.push({ type: 'text', value: value.slice(cursor, m.index) } as Text); |
| 86 | + } |
| 87 | + result.push({ type: 'element', tagName: 'br', properties: {}, children: [] } as Element); |
| 88 | + cursor = m.index! + m[0].length; |
| 89 | + } |
| 90 | + |
| 91 | + if (cursor < value.length) { |
| 92 | + result.push({ type: 'text', value: value.slice(cursor) } as Text); |
| 93 | + } |
| 94 | + |
| 95 | + return result; |
| 96 | +} |
| 97 | + |
| 98 | +/** |
| 99 | + * Parses a `<ul><li>...</li></ul>` string into a HAST element. |
| 100 | + * Returns null if the markup is malformed or contains unexpected content. |
| 101 | + */ |
| 102 | +function parseList(value: string): Element | null { |
| 103 | + const match = value.trim().match(LIST_PATTERN); |
| 104 | + if (!match) return null; |
| 105 | + |
| 106 | + const body = match[1]; |
| 107 | + const items: ElementContent[] = []; |
| 108 | + let cursor = 0; |
| 109 | + |
| 110 | + for (const liMatch of body.matchAll(LI_PATTERN)) { |
| 111 | + // Reject if there's non-whitespace between list items |
| 112 | + if (body.slice(cursor, liMatch.index!).trim()) return null; |
| 113 | + |
| 114 | + items.push({ |
| 115 | + type: 'element', |
| 116 | + tagName: 'li', |
| 117 | + properties: {}, |
| 118 | + children: expandBrTags(liMatch[1] ?? '') |
| 119 | + } as Element); |
| 120 | + |
| 121 | + cursor = liMatch.index! + liMatch[0].length; |
| 122 | + } |
| 123 | + |
| 124 | + // Reject if no items found or trailing garbage exists |
| 125 | + if (!items.length || body.slice(cursor).trim()) return null; |
| 126 | + |
| 127 | + return { type: 'element', tagName: 'ul', properties: {}, children: items } as Element; |
| 128 | +} |
| 129 | + |
| 130 | +/** |
| 131 | + * Processes a single table cell, restoring HTML elements from text content. |
| 132 | + */ |
| 133 | +function processCell(cell: Element) { |
| 134 | + visitParents(cell, 'text', (textNode: Text, ancestors) => { |
| 135 | + const parent = ancestors[ancestors.length - 1]; |
| 136 | + if (!parent || parent.type !== 'element') return; |
| 137 | + |
| 138 | + const parentEl = parent as Element; |
| 139 | + const siblings = parentEl.children as ElementContent[]; |
| 140 | + const startIndex = siblings.indexOf(textNode as ElementContent); |
| 141 | + if (startIndex === -1) return; |
| 142 | + |
| 143 | + // Combine consecutive text nodes and <br> elements into one string |
| 144 | + let combined = ''; |
| 145 | + let endIndex = startIndex; |
| 146 | + |
| 147 | + for (let i = startIndex; i < siblings.length; i++) { |
| 148 | + const sib = siblings[i]; |
| 149 | + if (sib.type === 'text') { |
| 150 | + combined += (sib as Text).value; |
| 151 | + endIndex = i; |
| 152 | + } else if (sib.type === 'element' && (sib as Element).tagName === 'br') { |
| 153 | + combined += '\n'; |
| 154 | + endIndex = i; |
| 155 | + } else { |
| 156 | + break; |
| 157 | + } |
| 158 | + } |
| 159 | + |
| 160 | + // Try parsing as list first (replaces entire combined range) |
| 161 | + const list = parseList(combined); |
| 162 | + if (list) { |
| 163 | + siblings.splice(startIndex, endIndex - startIndex + 1, list); |
| 164 | + return; |
| 165 | + } |
| 166 | + |
| 167 | + // Otherwise, just expand <br> tags in this text node |
| 168 | + const expanded = expandBrTags(textNode.value); |
| 169 | + if (expanded.length !== 1 || expanded[0] !== textNode) { |
| 170 | + siblings.splice(startIndex, 1, ...expanded); |
| 171 | + } |
| 172 | + }); |
| 173 | +} |
| 174 | + |
| 175 | +export const rehypeRestoreTableHtml: Plugin<[], Root> = () => (tree) => { |
| 176 | + visit(tree, 'element', (node: Element) => { |
| 177 | + if (node.tagName === 'td' || node.tagName === 'th') { |
| 178 | + processCell(node); |
| 179 | + } |
| 180 | + }); |
| 181 | +}; |
0 commit comments