The best way to control Claude Code, Codex and more from any computer or phone
| 1 | import { isLocalEditorFileUrl, LOCAL_FILE_LINK_SCHEMES } from "@t3tools/contracts"; |
| 2 | import type { |
| 3 | Heading, |
| 4 | Link, |
| 5 | Paragraph, |
| 6 | PhrasingContent, |
| 7 | Root, |
| 8 | RootContent, |
| 9 | TableCell, |
| 10 | Text, |
| 11 | } from "mdast"; |
| 12 | import { remarkKeepWindowsPaths } from "./markdownWindowsPaths.ts"; |
| 13 | import { type PluggableList, type Processor, type Transformer, unified } from "unified"; |
| 14 | import remarkParse from "remark-parse"; |
| 15 | import remarkRehype from "remark-rehype"; |
| 16 | import { find, html } from "property-information"; |
| 17 | import rehypeRaw from "rehype-raw"; |
| 18 | import rehypeSanitize, { defaultSchema } from "rehype-sanitize"; |
| 19 | import remarkBreaks from "remark-breaks"; |
| 20 | import remarkGfm from "remark-gfm"; |
| 21 | import { remarkGithubAlerts } from "./markdownGithubAlerts.ts"; |
| 22 | import { remarkNormalizeListItemIndentation } from "./markdownListIndentation.ts"; |
| 23 | import { remarkKeepOrderedListNumbers } from "./markdownOrderedListNumbers.ts"; |
| 24 | import { |
| 25 | remarkEnableLeadingWhitespace, |
| 26 | remarkKeepLeadingWhitespace, |
| 27 | } from "./markdownLeadingWhitespace.ts"; |
| 28 | import { |
| 29 | CODEX_ARTIFACT_TEMPLATE_HAST_PROPERTIES, |
| 30 | remarkCodexDirectives, |
| 31 | } from "./codexMarkdownDirectives.ts"; |
| 32 | import { |
| 33 | absoluteFilePathMatches, |
| 34 | isWindowsDrivePathHref, |
| 35 | normalizeMarkdownLinkDestination, |
| 36 | } from "./markdownLinks.ts"; |
| 37 | import { isWindowsAbsolutePath } from "./path.ts"; |
| 38 | import { THREAD_LINK_PROTOCOL } from "./threadLinks.ts"; |
| 39 | |
| 40 | export { remarkEnableOrderedListNumbers } from "./markdownOrderedListNumbers.ts"; |
| 41 | |
| 42 | type MarkdownImageHastNode = { |
| 43 | type?: string; |
| 44 | value?: string; |
| 45 | tagName?: string; |
| 46 | properties?: Record<string, unknown>; |
| 47 | children?: MarkdownImageHastNode[]; |
| 48 | }; |
| 49 | |
| 50 | function meaningfulHastChildren(node: MarkdownImageHastNode): MarkdownImageHastNode[] { |
| 51 | return (node.children ?? []).filter( |
| 52 | (child) => !(child.type === "text" && (child as { value?: string }).value?.trim() === ""), |
| 53 | ); |
| 54 | } |
| 55 | |
| 56 | const STANDALONE_IMAGE_BLOCKS = new Set([ |
| 57 | "p", |
| 58 | "div", |
| 59 | "li", |
| 60 | "td", |
| 61 | "th", |
| 62 | "figure", |
| 63 | "center", |
| 64 | "blockquote", |
| 65 | ]); |
| 66 | |
| 67 | function soleImageDescendant(node: MarkdownImageHastNode): MarkdownImageHastNode | undefined { |
| 68 | const children = meaningfulHastChildren(node); |
| 69 | if (children.length !== 1) return undefined; |
| 70 | const only = children[0]; |
| 71 | if (only?.type !== "element") return undefined; |
| 72 | if (only.tagName === "img") return only; |
| 73 | return only.tagName === "a" || only.tagName === "strong" || only.tagName === "em" |
| 74 | ? soleImageDescendant(only) |
| 75 | : undefined; |
| 76 | } |
| 77 | |
| 78 | function markStandaloneImages(node: MarkdownImageHastNode) { |
| 79 | if (node.type === "root" || (node.tagName && STANDALONE_IMAGE_BLOCKS.has(node.tagName))) { |
| 80 | const image = soleImageDescendant(node); |
| 81 | if (image) image.properties = { ...image.properties, dataStandalone: true }; |
| 82 | } |
| 83 | node.children?.forEach((child) => { |
| 84 | if (child.type === "element") markStandaloneImages(child); |
| 85 | }); |
| 86 | } |
| 87 | |
| 88 | // Whole tags, attributes, and comments, so closing-anchor text inside them never counts as a tag. |
| 89 | const RAW_HTML_TAG_PATTERN = /<!--[\s\S]*?(?:-->|$)|<\/?[A-Za-z](?:[^"'<>]|"[^"]*"|'[^']*')*>/g; |
| 90 | |
| 91 | function rehypePreserveBareAnchorPlaceholders() { |
| 92 | return (tree: MarkdownImageHastNode) => { |
| 93 | const anchors: Array<MarkdownImageHastNode | null> = []; |
| 94 | let rawTextTag: string | undefined; |
| 95 | const visit = (node: MarkdownImageHastNode) => { |
| 96 | if (node.type === "raw" && typeof node.value === "string") { |
| 97 | const tags = new RegExp(RAW_HTML_TAG_PATTERN); |
| 98 | let offset = 0; |
| 99 | while (rawTextTag !== "plaintext") { |
| 100 | // Raw text ends at its closing tag even inside comment-looking text. |
| 101 | const matcher = rawTextTag ? new RegExp(`</${rawTextTag}\\s*>`, "gi") : tags; |
| 102 | matcher.lastIndex = offset; |
| 103 | const match = matcher.exec(node.value); |
| 104 | if (!match) break; |
| 105 | const [tag] = match; |
| 106 | offset = matcher.lastIndex; |
| 107 | if (rawTextTag) { |
| 108 | rawTextTag = undefined; |
| 109 | continue; |
| 110 | } |
| 111 | if (tag.startsWith("<!--")) continue; |
| 112 | const closing = /^<\/([a-z]+)\s*>$/i.exec(tag)?.[1]?.toLowerCase(); |
| 113 | const opening = /^<([a-z]+)(?:\s|\/?>)/i.exec(tag)?.[1]?.toLowerCase(); |
| 114 | if ( |
| 115 | opening && |
| 116 | /^(?:script|style|textarea|title|xmp|iframe|noembed|noframes|plaintext)$/.test(opening) |
| 117 | ) { |
| 118 | rawTextTag = opening; |
| 119 | } else if (opening === "a") { |
| 120 | anchors.push(node.value === tag && /^<a\s*\/?>$/i.test(tag) ? node : null); |
| 121 | } else if (closing === "a") { |
| 122 | anchors.pop(); |
| 123 | } |
| 124 | } |
| 125 | } |
| 126 | node.children?.forEach(visit); |
| 127 | }; |
| 128 | |
| 129 | visit(tree); |
| 130 | for (const anchor of anchors) { |
| 131 | if (anchor) anchor.type = "text"; |
| 132 | } |
| 133 | }; |
| 134 | } |
| 135 | |
| 136 | const COLOR_SCHEME_MEDIA = /^\(\s*prefers-color-scheme\s*:\s*(dark|light)\s*\)$/i; |
| 137 | |
| 138 | /** |
| 139 | * The renderer resolves an `<img>`'s `src` (workspace files, host media) but never a `<source>`'s |
| 140 | * `srcset`, which would load unresolved whenever its media matched. A color-scheme source moves |
| 141 | * onto the `<img>` for the renderer to pick by theme; every source is then dropped. |
| 142 | */ |
| 143 | function moveColorSchemeSourcesToImage(picture: MarkdownImageHastNode) { |
| 144 | if (!picture.children) return; |
| 145 | const image = picture.children.find((child) => child.tagName === "img"); |
| 146 | for (const source of picture.children) { |
| 147 | if (source.tagName !== "source" || !image) continue; |
| 148 | const media = source.properties?.media; |
| 149 | const scheme = typeof media === "string" ? COLOR_SCHEME_MEDIA.exec(media)?.[1] : undefined; |
| 150 | const srcSet = source.properties?.srcSet; |
| 151 | const candidate = Array.isArray(srcSet) ? srcSet[0] : srcSet; |
| 152 | // A candidate's URL runs to the first space (a descriptor such as `2x` may follow). |
| 153 | const url = |
| 154 | typeof candidate === "string" ? candidate.trim().split(/\s/, 1)[0]?.replace(/,+$/, "") : ""; |
| 155 | const key = scheme?.toLowerCase() === "dark" ? "dataDarkSrc" : "dataLightSrc"; |
| 156 | // As in a browser, the first source that matches a scheme is the one shown. |
| 157 | if (scheme && url && image.properties?.[key] === undefined) { |
| 158 | // The sanitizer reads a drive letter as a protocol; as a file URL it resolves to the same path. |
| 159 | const value = isWindowsDrivePathHref(url) ? `file:///${url.replaceAll("\\", "/")}` : url; |
| 160 | image.properties = { ...image.properties, [key]: value }; |
| 161 | } |
| 162 | } |
| 163 | picture.children = picture.children.filter((child) => child.tagName !== "source"); |
| 164 | } |
| 165 | |
| 166 | function rehypePreserveImageSourceMeta() { |
| 167 | return (tree: MarkdownImageHastNode) => { |
| 168 | const visit = (node: MarkdownImageHastNode) => { |
| 169 | const src = node.properties?.src; |
| 170 | const title = node.properties?.title; |
| 171 | if (node.type === "element" && node.tagName === "picture") { |
| 172 | moveColorSchemeSourcesToImage(node); |
| 173 | } |
| 174 | if (node.type === "element" && node.tagName === "img") { |
| 175 | node.properties = { |
| 176 | ...node.properties, |
| 177 | ...(typeof src === "string" && isWindowsDrivePathHref(src) ? { dataLocalSrc: src } : {}), |
| 178 | ...(typeof title === "string" ? { dataMarkdownTitle: title } : {}), |
| 179 | }; |
| 180 | } |
| 181 | node.children?.forEach(visit); |
| 182 | }; |
| 183 | |
| 184 | visit(tree); |
| 185 | markStandaloneImages(tree); |
| 186 | }; |
| 187 | } |
| 188 | |
| 189 | const IMAGE_SRC_PROTOCOLS = [...(defaultSchema.protocols?.src ?? []), "file", "t3-context"]; |
| 190 | |
| 191 | // The sanitizer compares schemes case-sensitively, so lowercase accepted editor links first. |
| 192 | function rehypeNormalizeLocalEditorLinks() { |
| 193 | return (tree: MarkdownImageHastNode) => { |
| 194 | const visit = (node: MarkdownImageHastNode) => { |
| 195 | const href = node.properties?.href; |
| 196 | if (node.type === "element" && node.tagName === "a" && typeof href === "string") { |
| 197 | try { |
| 198 | const url = new URL(href); |
| 199 | if (isLocalEditorFileUrl(url)) node.properties = { ...node.properties, href: url.href }; |
| 200 | } catch { |
| 201 | // Not an absolute URL; the sanitizer decides. |
| 202 | } |
| 203 | } |
| 204 | node.children?.forEach(visit); |
| 205 | }; |
| 206 | |
| 207 | visit(tree); |
| 208 | }; |
| 209 | } |
| 210 | |
| 211 | const CHAT_MARKDOWN_SANITIZE_SCHEMA = { |
| 212 | ...defaultSchema, |
| 213 | attributes: { |
| 214 | ...defaultSchema.attributes, |
| 215 | "*": (defaultSchema.attributes?.["*"] ?? []).filter((attribute) => attribute !== "title"), |
| 216 | // Fence languages plus the two math classes the web chat's math plugin emits. |
| 217 | code: [ |
| 218 | ["className", /^language-./, "math-inline", "math-display"], |
| 219 | "dataCodeMeta", |
| 220 | "dataInlineCode", |
| 221 | ], |
| 222 | blockquote: [...(defaultSchema.attributes?.blockquote ?? []), "dataAlert"], |
| 223 | div: [...(defaultSchema.attributes?.div ?? []), ...CODEX_ARTIFACT_TEMPLATE_HAST_PROPERTIES], |
| 224 | a: [...(defaultSchema.attributes?.a ?? []), "dataPullRequestAutolink"], |
| 225 | img: [ |
| 226 | ...(defaultSchema.attributes?.img ?? []), |
| 227 | "dataDarkSrc", |
| 228 | "dataLightSrc", |
| 229 | "dataLocalSrc", |
| 230 | "dataMarkdownTitle", |
| 231 | "dataStandalone", |
| 232 | ], |
| 233 | }, |
| 234 | protocols: { |
| 235 | ...defaultSchema.protocols, |
| 236 | href: [ |
| 237 | ...(defaultSchema.protocols?.href ?? []), |
| 238 | "file", |
| 239 | "t3-citation", |
| 240 | "t3-context", |
| 241 | THREAD_LINK_PROTOCOL, |
| 242 | // Scheme only: the renderer's URL transform checks the rest of the link shape. |
| 243 | ...LOCAL_FILE_LINK_SCHEMES, |
| 244 | ], |
| 245 | src: IMAGE_SRC_PROTOCOLS, |
| 246 | dataDarkSrc: IMAGE_SRC_PROTOCOLS, |
| 247 | dataLightSrc: IMAGE_SRC_PROTOCOLS, |
| 248 | }, |
| 249 | } satisfies Parameters<typeof rehypeSanitize>[0]; |
| 250 | |
| 251 | export const CHAT_MARKDOWN_REMARK_PLUGINS: PluggableList = [ |
| 252 | remarkGfm, |
| 253 | remarkEndAutolinksAtScriptPunctuation, |
| 254 | remarkKeepAutolinkSuffixes, |
| 255 | remarkLinkAbsoluteFilePaths, |
| 256 | remarkKeepWindowsPaths, |
| 257 | remarkGithubAlerts, |
| 258 | remarkKeepOrderedListNumbers, |
| 259 | remarkNormalizeListItemIndentation, |
| 260 | remarkCodexDirectives, |
| 261 | remarkBreakAroundImageLines, |
| 262 | remarkPreserveCodeMeta, |
| 263 | remarkNormalizeLinksAndTagInlineCode, |
| 264 | ]; |
| 265 | |
| 266 | // Leading whitespace is kept only where each line renders as a line. |
| 267 | export const CHAT_MARKDOWN_REMARK_PLUGINS_WITH_BREAKS: PluggableList = [ |
| 268 | remarkGfm, |
| 269 | remarkEndAutolinksAtScriptPunctuation, |
| 270 | remarkKeepAutolinkSuffixes, |
| 271 | remarkKeepLeadingWhitespace, |
| 272 | remarkLinkAbsoluteFilePaths, |
| 273 | remarkKeepWindowsPaths, |
| 274 | remarkGithubAlerts, |
| 275 | remarkKeepOrderedListNumbers, |
| 276 | remarkNormalizeListItemIndentation, |
| 277 | remarkCodexDirectives, |
| 278 | remarkBreakAroundImageLines, |
| 279 | remarkBreaks, |
| 280 | remarkPreserveCodeMeta, |
| 281 | remarkNormalizeLinksAndTagInlineCode, |
| 282 | ]; |
| 283 | |
| 284 | // Users type paths like ~/a and (~/b, which GFM's single-tilde strikethrough hides as <del>. |
| 285 | // Appended after the base list, this reconfigures its remark-gfm entry; ~~double~~ still strikes. |
| 286 | // Users also indent lines with spaces, which CommonMark drops from a paragraph. |
| 287 | export const CHAT_USER_MARKDOWN_REMARK_PLUGINS: PluggableList = [ |
| 288 | [remarkGfm, { singleTilde: false }], |
| 289 | remarkEnableLeadingWhitespace, |
| 290 | ]; |
| 291 | |
| 292 | // Start tags an HTML parser closes an open `<p>` for. |
| 293 | const CLOSES_PARAGRAPH = new Set( |
| 294 | "address article aside blockquote center details dialog dir div dl fieldset figcaption figure footer form h1 h2 h3 h4 h5 h6 header hgroup hr li listing main menu nav ol p plaintext pre search section summary table ul xmp dd dt".split( |
| 295 | " ", |
| 296 | ), |
| 297 | ); |
| 298 | const HEADING = /^h[1-6]$/; |
| 299 | |
| 300 | interface HtmlScope { |
| 301 | readonly anchor: boolean; |
| 302 | readonly paragraph: boolean; |
| 303 | readonly heading: boolean; |
| 304 | } |
| 305 | |
| 306 | /** |
| 307 | * Whether the tree needs the HTML parser: raw HTML, or nesting it repairs (a footnote call in a |
| 308 | * link label nests `<a>`). Otherwise it only rewrites properties as hast-util-to-parse5 does and |
| 309 | * drops `data`, both repeated here, and moves blank text that renders the same. |
| 310 | */ |
| 311 | function needsHtmlParser(node: MarkdownImageHastNode, scope: HtmlScope): boolean { |
| 312 | if (node.type === "raw") return true; |
| 313 | let inner = scope; |
| 314 | if (node.type === "element" && node.tagName) { |
| 315 | const { tagName } = node; |
| 316 | const heading = HEADING.test(tagName); |
| 317 | if ( |
| 318 | (tagName === "a" && scope.anchor) || |
| 319 | (scope.paragraph && CLOSES_PARAGRAPH.has(tagName)) || |
| 320 | (heading && scope.heading) |
| 321 | ) { |
| 322 | return true; |
| 323 | } |
| 324 | inner = { |
| 325 | anchor: scope.anchor || tagName === "a", |
| 326 | paragraph: scope.paragraph || tagName === "p", |
| 327 | heading: scope.heading || heading, |
| 328 | }; |
| 329 | const properties = node.properties ?? {}; |
| 330 | for (const key of Object.keys(properties)) { |
| 331 | const value = properties[key]; |
| 332 | const info = find(html, key); |
| 333 | if ( |
| 334 | value === false || |
| 335 | value === null || |
| 336 | value === undefined || |
| 337 | Number.isNaN(value) || |
| 338 | (!value && info.boolean) |
| 339 | ) { |
| 340 | delete properties[key]; |
| 341 | } else if (value === true && !info.boolean && !info.overloadedBoolean) { |
| 342 | properties[key] = ""; |
| 343 | } |
| 344 | } |
| 345 | delete (node as { data?: unknown }).data; |
| 346 | } |
| 347 | return (node.children ?? []).some((child) => needsHtmlParser(child, inner)); |
| 348 | } |
| 349 | |
| 350 | const TOP_SCOPE: HtmlScope = { anchor: false, paragraph: false, heading: false }; |
| 351 | |
| 352 | /** rehype-raw reparses the whole tree through an HTML parser, which most trees can skip. */ |
| 353 | function rehypeRawWhenPresent(this: Processor) { |
| 354 | const transform = rehypeRaw.call(this) as unknown as ( |
| 355 | tree: MarkdownImageHastNode, |
| 356 | file: unknown, |
| 357 | ) => unknown; |
| 358 | return (tree: MarkdownImageHastNode, file: unknown) => |
| 359 | needsHtmlParser(tree, TOP_SCOPE) ? transform(tree, file) : undefined; |
| 360 | } |
| 361 | |
| 362 | export const CHAT_MARKDOWN_REHYPE_PLUGINS: PluggableList = [ |
| 363 | rehypePreserveBareAnchorPlaceholders, |
| 364 | rehypeRawWhenPresent, |
| 365 | rehypePreserveImageSourceMeta, |
| 366 | rehypeNormalizeLocalEditorLinks, |
| 367 | [rehypeSanitize, CHAT_MARKDOWN_SANITIZE_SCHEMA], |
| 368 | ]; |
| 369 | |
| 370 | const isImageNode = (node: PhrasingContent | undefined): boolean => |
| 371 | node?.type === "image" || |
| 372 | node?.type === "imageReference" || |
| 373 | ((node?.type === "link" || node?.type === "linkReference") && |
| 374 | node.children.length === 1 && |
| 375 | isImageNode(node.children[0])); |
| 376 | const endsLine = (node: PhrasingContent | undefined) => |
| 377 | node === undefined || |
| 378 | node.type === "break" || |
| 379 | (node.type === "text" && /\n[ \t]*$/.test(node.value)); |
| 380 | const startsLine = (node: PhrasingContent | undefined) => |
| 381 | node === undefined || |
| 382 | node.type === "break" || |
| 383 | (node.type === "text" && /^[ \t]*\n/.test(node.value)); |
| 384 | |
| 385 | function breakLinesAroundImages(paragraph: Paragraph) { |
| 386 | const children = paragraph.children; |
| 387 | if (!children.some(isImageNode)) return; |
| 388 | const ownsLine = (index: number) => |
| 389 | isImageNode(children[index]) && |
| 390 | endsLine(children[index - 1]) && |
| 391 | startsLine(children[index + 1]); |
| 392 | paragraph.children = children.flatMap((child, index): PhrasingContent[] => { |
| 393 | if (child.type !== "text") return [child]; |
| 394 | const afterImage = ownsLine(index - 1); |
| 395 | const beforeImage = ownsLine(index + 1); |
| 396 | if (!afterImage && !beforeImage) return [child]; |
| 397 | const parts: PhrasingContent[] = []; |
| 398 | let value = child.value; |
| 399 | if (afterImage) { |
| 400 | value = value.replace(/^[ \t]*\n/, ""); |
| 401 | parts.push({ type: "break" }); |
| 402 | } |
| 403 | const breakAfter = beforeImage && /\n[ \t]*$/.test(value); |
| 404 | if (breakAfter) value = value.replace(/[ \t]*\n[ \t]*$/, ""); |
| 405 | if (value) parts.push({ type: "text", value }); |
| 406 | if (breakAfter) parts.push({ type: "break" }); |
| 407 | return parts; |
| 408 | }); |
| 409 | } |
| 410 | |
| 411 | /** |
| 412 | * Agents label screenshots with a line like `**Palette**` right above the image. CommonMark |
| 413 | * joins those lines into one paragraph, which puts the image inline beside its label. An |
| 414 | * image alone on its source line keeps that line; prose and inline badges still wrap. |
| 415 | */ |
| 416 | function remarkBreakAroundImageLines() { |
| 417 | return (tree: Root) => { |
| 418 | const visit = (node: Root | RootContent) => { |
| 419 | if (node.type === "paragraph") breakLinesAroundImages(node); |
| 420 | else if ("children" in node) node.children.forEach(visit); |
| 421 | }; |
| 422 | visit(tree); |
| 423 | }; |
| 424 | } |
| 425 | |
| 426 | type MarkdownAstNode = { |
| 427 | type?: string; |
| 428 | meta?: unknown; |
| 429 | url?: string; |
| 430 | data?: { |
| 431 | hProperties?: Record<string, unknown>; |
| 432 | }; |
| 433 | children?: MarkdownAstNode[]; |
| 434 | }; |
| 435 | |
| 436 | function remarkKeepAutolinkSuffixes(): Transformer<Root> { |
| 437 | return (tree, file) => { |
| 438 | const source = String(file); |
| 439 | const visit = (node: Root | RootContent) => { |
| 440 | if (!("children" in node) || node.type === "link" || node.type === "linkReference") return; |
| 441 | for (let index = 0; index < node.children.length; index++) { |
| 442 | const child = node.children[index]; |
| 443 | const next = node.children[index + 1]; |
| 444 | if (child?.type === "link" && /^https?:\/\//i.test(child.url) && next?.type === "text") { |
| 445 | const suffix = /^[_~]+/.exec(next.value)?.[0]; |
| 446 | if (!suffix) continue; |
| 447 | const label = child.children.length === 1 ? child.children[0] : undefined; |
| 448 | if (label?.type !== "text") continue; |
| 449 | if (label.value !== child.url && `http://${label.value}` !== child.url) continue; |
| 450 | const start = child.position?.start.offset; |
| 451 | const end = child.position?.end.offset; |
| 452 | const bare = |
| 453 | !child.position || |
| 454 | (start !== undefined && |
| 455 | end !== undefined && |
| 456 | source.slice(start, end) === label.value && |
| 457 | source.slice(end, end + suffix.length) === suffix); |
| 458 | if (bare) { |
| 459 | child.url += suffix; |
| 460 | label.value += suffix; |
| 461 | next.value = next.value.slice(suffix.length); |
| 462 | if (child.position && end !== undefined) { |
| 463 | const boundary = { |
| 464 | ...child.position.end, |
| 465 | column: child.position.end.column + suffix.length, |
| 466 | offset: end + suffix.length, |
| 467 | }; |
| 468 | child.position.end = boundary; |
| 469 | if (label.position) label.position.end = boundary; |
| 470 | if (next.position) next.position.start = boundary; |
| 471 | } |
| 472 | if (!next.value) node.children.splice(index + 1, 1); |
| 473 | } |
| 474 | } |
| 475 | if (child) visit(child); |
| 476 | } |
| 477 | }; |
| 478 | visit(tree); |
| 479 | }; |
| 480 | } |
| 481 | |
| 482 | // Punctuation and symbols from CJK Symbols and Punctuation and from Halfwidth and Fullwidth Forms, |
| 483 | // plus the Arabic-script comma, semicolon, question mark, and full stop. |
| 484 | const SCRIPT_PUNCTUATION = /(?=[ -〿-،؛؟۔])[\p{P}\p{S}]/u; |
| 485 | |
| 486 | /** |
| 487 | * GFM runs a bare URL on until whitespace, so Chinese, Japanese, and Arabic punctuation joins it, |
| 488 | * along with the rest of the sentence and any `**` that should have closed around it. A URL that is |
| 489 | * ASCII up to such a mark ends there. Unicode earlier in the URL means the mark may belong to it, so it stays. |
| 490 | */ |
| 491 | function scriptPunctuationIndex(link: Link): number { |
| 492 | const label = link.children.length === 1 ? link.children[0] : undefined; |
| 493 | if (label?.type !== "text" || !/^https?:\/\//i.test(link.url)) return -1; |
| 494 | if (link.url !== label.value && link.url !== `http://${label.value}`) return -1; |
| 495 | const index = label.value.search(/[^!-~]/u); |
| 496 | return index > 0 && SCRIPT_PUNCTUATION.test(label.value.charAt(index)) ? index : -1; |
| 497 | } |
| 498 | |
| 499 | /** GFM's text pass links bare URLs without positions; its tokenizer's links span their URL. */ |
| 500 | function isBareLink(link: Link, source: string): boolean { |
| 501 | const start = link.position?.start.offset; |
| 502 | const end = link.position?.end.offset; |
| 503 | const label = link.children[0]; |
| 504 | return ( |
| 505 | !link.position || |
| 506 | (start !== undefined && |
| 507 | end !== undefined && |
| 508 | label?.type === "text" && |
| 509 | source.slice(start, end) === label.value) |
| 510 | ); |
| 511 | } |
| 512 | |
| 513 | /** Source offsets of the marks that tokenized bare URLs run past. */ |
| 514 | function scriptPunctuationOffsets(node: RootContent, source: string): number[] { |
| 515 | if (node.type === "link") { |
| 516 | const index = scriptPunctuationIndex(node); |
| 517 | const start = node.position?.start.offset; |
| 518 | return index > 0 && start !== undefined && isBareLink(node, source) ? [start + index] : []; |
| 519 | } |
| 520 | return "children" in node && node.type !== "linkReference" |
| 521 | ? (node.children as RootContent[]).flatMap((child) => scriptPunctuationOffsets(child, source)) |
| 522 | : []; |
| 523 | } |
| 524 | |
| 525 | const hasReference = (node: RootContent): boolean => |
| 526 | node.type === "linkReference" || |
| 527 | node.type === "imageReference" || |
| 528 | node.type === "footnoteReference" || |
| 529 | ("children" in node && (node.children as RootContent[]).some(hasReference)); |
| 530 | |
| 531 | /** How many of the ascending numbers are below `value`. */ |
| 532 | function countBelow(sorted: readonly number[], value: number): number { |
| 533 | let low = 0; |
| 534 | let high = sorted.length; |
| 535 | while (low < high) { |
| 536 | const middle = (low + high) >> 1; |
| 537 | if (sorted[middle]! < value) low = middle + 1; |
| 538 | else high = middle; |
| 539 | } |
| 540 | return low; |
| 541 | } |
| 542 | |
| 543 | /** |
| 544 | * Parses the block's text again with a tab before each mark. The tab ends the URL as a space would, |
| 545 | * so emphasis around the URL closes and later URLs and code get their own nodes. The tabs are then |
| 546 | * dropped and positions mapped back to the source. Returns false and leaves the block alone when its |
| 547 | * text would parse differently on its own, as a quote's continuation lines do. |
| 548 | */ |
| 549 | function reparseAtScriptPunctuation( |
| 550 | processor: Processor, |
| 551 | block: Paragraph | Heading | TableCell, |
| 552 | source: string, |
| 553 | offsets: number[], |
| 554 | ): boolean { |
| 555 | // A task item's paragraph starts at its checkbox; its first child starts after it. |
| 556 | const from = block.children[0]?.position?.start; |
| 557 | const to = block.children.at(-1)?.position?.end; |
| 558 | if (from?.offset === undefined || to?.offset === undefined || hasReference(block)) return false; |
| 559 | const base = from.offset; |
| 560 | const text = source.slice(base, to.offset); |
| 561 | // A tab is the marker, a quote's `>` on a later line would read as a new quote, a list's |
| 562 | // indentation would join a code span, and only a table unescapes `\|` in code. |
| 563 | if ( |
| 564 | text.includes("\t") || |
| 565 | /[\r\n][ ]*>/.test(text) || |
| 566 | (/[\r\n] /.test(text) && text.includes("`")) || |
| 567 | (block.type === "tableCell" && text.includes("\\|")) |
| 568 | ) { |
| 569 | return false; |
| 570 | } |
| 571 | const marks = offsets.map((offset) => offset - base); |
| 572 | // Each pass finds the next URL in a run with no spaces, like `http://a,http://b,http://c`. |
| 573 | for (let attempt = 0; attempt < 16; attempt++) { |
| 574 | marks.sort((a, b) => a - b); |
| 575 | const tabs = marks.map((mark, index) => mark + index); |
| 576 | const unmarked = (offset: number) => offset - countBelow(tabs, offset); |
| 577 | const marked = [0, ...marks] |
| 578 | .map((start, index) => text.slice(start, marks[index] ?? text.length)) |
| 579 | .join("\t"); |
| 580 | const [paragraph, ...rest] = (processor.parse(marked) as Root).children; |
| 581 | if (paragraph?.type !== "paragraph" || rest.length > 0) return false; |
| 582 | const more = scriptPunctuationOffsets(paragraph, marked); |
| 583 | if (more.length > 0) { |
| 584 | marks.push(...more.map(unmarked)); |
| 585 | continue; |
| 586 | } |
| 587 | |
| 588 | let tabCount = 0; |
| 589 | const moved = new Set<object>(); |
| 590 | const restore = (node: RootContent): boolean => { |
| 591 | for (const point of node.position ? [node.position.start, node.position.end] : []) { |
| 592 | const offset = point.offset; |
| 593 | if (offset === undefined || moved.has(point)) continue; |
| 594 | moved.add(point); |
| 595 | const tabsOnLine = countBelow(tabs, offset) - countBelow(tabs, offset - point.column + 1); |
| 596 | point.column = (point.line === 1 ? from.column - 1 : 0) + point.column - tabsOnLine; |
| 597 | point.line += from.line - 1; |
| 598 | point.offset = base + unmarked(offset); |
| 599 | } |
| 600 | if (node.type === "text") { |
| 601 | tabCount += node.value.split("\t").length - 1; |
| 602 | node.value = node.value.replaceAll("\t", ""); |
| 603 | } else if ( |
| 604 | ("value" in node && node.value.includes("\t")) || |
| 605 | ("url" in node && node.url.includes("\t")) |
| 606 | ) { |
| 607 | return false; |
| 608 | } |
| 609 | if (!("children" in node)) return true; |
| 610 | const children = node.children as RootContent[]; |
| 611 | if (!children.every(restore)) return false; |
| 612 | node.children = children.filter( |
| 613 | (child) => child.type !== "text" || child.value !== "", |
| 614 | ) as typeof node.children; |
| 615 | return true; |
| 616 | }; |
| 617 | if (!restore(paragraph) || tabCount !== tabs.length) return false; |
| 618 | block.children = paragraph.children; |
| 619 | return true; |
| 620 | } |
| 621 | return false; |
| 622 | } |
| 623 | |
| 624 | /** Ends the bare URLs a reparse could not, and those GFM's text pass linked after emphasis. */ |
| 625 | function splitAtScriptPunctuation(node: RootContent, source: string) { |
| 626 | if (!("children" in node) || node.type === "link" || node.type === "linkReference") return; |
| 627 | const children = node.children as RootContent[]; |
| 628 | for (let index = 0; index < children.length; index++) { |
| 629 | const child = children[index]!; |
| 630 | if (child.type !== "link") { |
| 631 | splitAtScriptPunctuation(child, source); |
| 632 | continue; |
| 633 | } |
| 634 | const at = isBareLink(child, source) ? scriptPunctuationIndex(child) : -1; |
| 635 | const label = child.children[0]; |
| 636 | if (at < 0 || label?.type !== "text") continue; |
| 637 | // GFM drops this punctuation from the end of a URL; remarkKeepAutolinkSuffixes restores `_` and `~`. |
| 638 | const length = label.value.slice(0, at).replace(/[!"'*,.:;?_~]+$/, "").length; |
| 639 | const tail: Text = { type: "text", value: label.value.slice(length) }; |
| 640 | const start = child.position?.start; |
| 641 | if (child.position && start?.offset !== undefined) { |
| 642 | const end = { |
| 643 | line: start.line, |
| 644 | column: start.column + length, |
| 645 | offset: start.offset + length, |
| 646 | }; |
| 647 | tail.position = { start: end, end: child.position.end }; |
| 648 | child.position = { start, end }; |
| 649 | label.position = { start, end }; |
| 650 | } |
| 651 | child.url = child.url.slice(0, child.url.length - label.value.length + length); |
| 652 | label.value = label.value.slice(0, length); |
| 653 | children.splice(index + 1, 0, tail); |
| 654 | index++; |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | function remarkEndAutolinksAtScriptPunctuation(this: Processor): Transformer<Root> { |
| 659 | return (tree, file) => { |
| 660 | const source = String(file); |
| 661 | if (!SCRIPT_PUNCTUATION.test(source)) return; |
| 662 | const visit = (node: RootContent) => { |
| 663 | if (node.type === "paragraph" || node.type === "heading" || node.type === "tableCell") { |
| 664 | const offsets = scriptPunctuationOffsets(node, source); |
| 665 | if (offsets.length > 0) reparseAtScriptPunctuation(this, node, source, offsets); |
| 666 | splitAtScriptPunctuation(node, source); |
| 667 | } else if ("children" in node) { |
| 668 | (node.children as RootContent[]).forEach(visit); |
| 669 | } |
| 670 | }; |
| 671 | tree.children.forEach(visit); |
| 672 | }; |
| 673 | } |
| 674 | |
| 675 | function remarkLinkAbsoluteFilePaths(): Transformer<Root> { |
| 676 | return (tree) => { |
| 677 | // Raw `<a>` tags are sibling html nodes, so text inside one is not under a link node. |
| 678 | let rawAnchorDepth = 0; |
| 679 | const visit = (node: Root | RootContent) => { |
| 680 | if (!("children" in node) || node.type === "link" || node.type === "linkReference") return; |
| 681 | node.children = node.children.flatMap((child): typeof node.children => { |
| 682 | if (child.type === "html") { |
| 683 | for (const [tag] of child.value.matchAll(RAW_HTML_TAG_PATTERN)) { |
| 684 | if (/^<a[\s/>]/i.test(tag)) rawAnchorDepth += 1; |
| 685 | else if (/^<\/a[\s>]/i.test(tag)) rawAnchorDepth = Math.max(0, rawAnchorDepth - 1); |
| 686 | } |
| 687 | } |
| 688 | if (child.type !== "text" || rawAnchorDepth > 0) { |
| 689 | visit(child); |
| 690 | return [child]; |
| 691 | } |
| 692 | const parts: PhrasingContent[] = []; |
| 693 | let offset = 0; |
| 694 | for (const { index, path } of absoluteFilePathMatches(child.value)) { |
| 695 | if (index > offset) parts.push({ type: "text", value: child.value.slice(offset, index) }); |
| 696 | parts.push({ type: "link", url: path, children: [{ type: "text", value: path }] }); |
| 697 | offset = index + path.length; |
| 698 | } |
| 699 | if (offset === 0) return [child]; |
| 700 | if (offset < child.value.length) |
| 701 | parts.push({ type: "text", value: child.value.slice(offset) }); |
| 702 | return parts; |
| 703 | }); |
| 704 | }; |
| 705 | visit(tree); |
| 706 | }; |
| 707 | } |
| 708 | |
| 709 | function remarkPreserveCodeMeta() { |
| 710 | return (tree: MarkdownAstNode) => { |
| 711 | const visit = (node: MarkdownAstNode) => { |
| 712 | if (node.type === "code" && typeof node.meta === "string" && node.meta.trim().length > 0) { |
| 713 | node.data = { |
| 714 | ...node.data, |
| 715 | hProperties: { |
| 716 | ...node.data?.hProperties, |
| 717 | dataCodeMeta: node.meta.trim(), |
| 718 | }, |
| 719 | }; |
| 720 | } |
| 721 | node.children?.forEach(visit); |
| 722 | }; |
| 723 | |
| 724 | visit(tree); |
| 725 | }; |
| 726 | } |
| 727 | |
| 728 | function remarkNormalizeLinksAndTagInlineCode() { |
| 729 | return (tree: MarkdownAstNode) => { |
| 730 | const visit = (node: MarkdownAstNode, insideLink: boolean) => { |
| 731 | if ( |
| 732 | (node.type === "link" || node.type === "definition") && |
| 733 | typeof node.url === "string" && |
| 734 | /^[A-Za-z]:[\\/]/.test(node.url) |
| 735 | ) { |
| 736 | node.url = `file:///${node.url.replaceAll("\\", "/")}`; |
| 737 | } |
| 738 | if (node.type === "inlineCode" && !insideLink) { |
| 739 | node.data = { |
| 740 | ...node.data, |
| 741 | hProperties: { |
| 742 | ...node.data?.hProperties, |
| 743 | dataInlineCode: "", |
| 744 | }, |
| 745 | }; |
| 746 | } |
| 747 | const childInsideLink = insideLink || node.type === "link" || node.type === "linkReference"; |
| 748 | node.children?.forEach((child) => visit(child, childInsideLink)); |
| 749 | }; |
| 750 | |
| 751 | visit(tree, false); |
| 752 | }; |
| 753 | } |
| 754 | |
| 755 | export function shouldPreserveAssistantLineBreaks(text: string): boolean { |
| 756 | return /^★ Insight(?:\s|─)/mu.test(text); |
| 757 | } |
| 758 | |
| 759 | export function remarkAssistantInsights(this: Processor) { |
| 760 | return (tree: Root, file: { value: unknown }) => { |
| 761 | const source = String(file.value); |
| 762 | if (!source.includes("★ Insight")) return; |
| 763 | const protectedRanges: { start: number; end: number; code: boolean }[] = []; |
| 764 | const inspect = (node: RootContent | PhrasingContent) => { |
| 765 | const start = node.position?.start.offset; |
| 766 | const end = node.position?.end.offset; |
| 767 | if (start !== undefined && end !== undefined) { |
| 768 | const code = |
| 769 | node.type === "code" || |
| 770 | node.type === "html" || |
| 771 | (node.type === "inlineCode" && node.position?.start.line !== node.position?.end.line); |
| 772 | if (code || node.type === "list" || node.type === "blockquote") |
| 773 | protectedRanges.push({ start, end, code }); |
| 774 | } |
| 775 | if ("children" in node) node.children.forEach(inspect); |
| 776 | }; |
| 777 | tree.children.forEach(inspect); |
| 778 | const starts = new Set<number>(); |
| 779 | const offsets: { original: number; rewritten: number; prefix: number }[] = []; |
| 780 | let indent: string | undefined; |
| 781 | let original = 0; |
| 782 | let rewritten = 0; |
| 783 | const lines = source.split(/(?<=\n)|(?<=\r)(?!\n)/u).map((line, index) => { |
| 784 | const value = line.replace(/[\r\n]+$/u, ""); |
| 785 | const blocked = protectedRanges.some( |
| 786 | (range) => |
| 787 | original >= range.start && original < range.end && (indent === undefined || range.code), |
| 788 | ); |
| 789 | const opening = |
| 790 | indent === undefined && !blocked |
| 791 | ? /^( {0,3})(`?)★ Insight[ \t]+─{3,}\2[ \t]*$/u.exec(value) |
| 792 | : null; |
| 793 | const closing = |
| 794 | indent !== undefined && !blocked ? /^( {0,3})(`?)─{3,}\2[ \t]*$/u.exec(value) : null; |
| 795 | let prefix = 0; |
| 796 | let output = line; |
| 797 | if (opening) { |
| 798 | indent = opening[1] ?? ""; |
| 799 | starts.add(index + 1); |
| 800 | output = ">" + line.slice(value.length); |
| 801 | } else if (closing && closing[1] === indent) { |
| 802 | indent = undefined; |
| 803 | output = line.slice(value.length); |
| 804 | } else if (indent !== undefined) { |
| 805 | prefix = 2; |
| 806 | output = "> " + line; |
| 807 | } |
| 808 | offsets.push({ original, rewritten, prefix }); |
| 809 | original += line.length; |
| 810 | rewritten += output.length; |
| 811 | return output; |
| 812 | }); |
| 813 | if (starts.size === 0) return; |
| 814 | const parsed = this.parse(lines.join("")) as Root; |
| 815 | const restorePosition = (node: RootContent | PhrasingContent) => { |
| 816 | if (node.position) { |
| 817 | for (const point of [node.position.start, node.position.end]) { |
| 818 | const offset = offsets[point.line - 1]; |
| 819 | if (!offset) continue; |
| 820 | point.column = Math.max(1, point.column - offset.prefix); |
| 821 | if (point.offset !== undefined) |
| 822 | point.offset += offset.original - offset.rewritten - offset.prefix; |
| 823 | } |
| 824 | } |
| 825 | if ("children" in node) node.children.forEach(restorePosition); |
| 826 | }; |
| 827 | for (const node of parsed.children) { |
| 828 | if (node.type === "blockquote" && starts.has(node.position?.start.line ?? 0)) { |
| 829 | node.data = { ...node.data, hProperties: { dataAlert: "insight" } }; |
| 830 | } |
| 831 | restorePosition(node); |
| 832 | } |
| 833 | tree.children = parsed.children; |
| 834 | }; |
| 835 | } |
| 836 | |
| 837 | const imageSourceProcessor = unified() |
| 838 | .use(remarkParse) |
| 839 | .use(CHAT_MARKDOWN_REMARK_PLUGINS) |
| 840 | .use(remarkRehype, { allowDangerousHtml: true }) |
| 841 | .use(CHAT_MARKDOWN_REHYPE_PLUGINS); |
| 842 | |
| 843 | export type ChatMarkdownImageProperties = { |
| 844 | readonly src?: unknown; |
| 845 | readonly dataLocalSrc?: unknown; |
| 846 | readonly dataDarkSrc?: unknown; |
| 847 | readonly dataLightSrc?: unknown; |
| 848 | }; |
| 849 | |
| 850 | /** Keep the authored path for file links and normalize Windows separators for classification. */ |
| 851 | export function chatMarkdownImageSource( |
| 852 | properties?: ChatMarkdownImageProperties, |
| 853 | theme?: "dark" | "light", |
| 854 | ) { |
| 855 | const themeSrc = |
| 856 | theme === undefined |
| 857 | ? undefined |
| 858 | : properties?.[theme === "dark" ? "dataDarkSrc" : "dataLightSrc"]; |
| 859 | const localSrc = typeof themeSrc === "string" ? undefined : properties?.dataLocalSrc; |
| 860 | const source = |
| 861 | typeof themeSrc === "string" |
| 862 | ? themeSrc |
| 863 | : typeof localSrc === "string" |
| 864 | ? localSrc |
| 865 | : properties?.src; |
| 866 | if (typeof source !== "string") return null; |
| 867 | const authoredSrc = normalizeMarkdownLinkDestination(source); |
| 868 | return { |
| 869 | authoredSrc, |
| 870 | classifiedSrc: typeof localSrc === "string" ? authoredSrc.replaceAll("\\", "/") : authoredSrc, |
| 871 | }; |
| 872 | } |
| 873 | |
| 874 | interface ImageSourceNode { |
| 875 | readonly type: string; |
| 876 | readonly tagName?: string; |
| 877 | readonly properties?: ChatMarkdownImageProperties; |
| 878 | readonly children?: ReadonlyArray<ImageSourceNode>; |
| 879 | } |
| 880 | |
| 881 | /** |
| 882 | * Collect the fallback and every theme variant so images are available to all viewers. |
| 883 | */ |
| 884 | export function chatMarkdownImageSources(markdown: string): Array<string> { |
| 885 | const sources: Array<string> = []; |
| 886 | const visit = (node: ImageSourceNode) => { |
| 887 | if (node.type === "element" && node.tagName === "img") { |
| 888 | const properties = node.properties; |
| 889 | const fallback = chatMarkdownImageSource(properties); |
| 890 | if (fallback) sources.push(fallback.classifiedSrc); |
| 891 | for (const [theme, variant] of [ |
| 892 | ["dark", properties?.dataDarkSrc], |
| 893 | ["light", properties?.dataLightSrc], |
| 894 | ] as const) { |
| 895 | if (typeof variant !== "string") continue; |
| 896 | const source = chatMarkdownImageSource(properties, theme); |
| 897 | if (source) sources.push(source.classifiedSrc); |
| 898 | } |
| 899 | } |
| 900 | node.children?.forEach(visit); |
| 901 | }; |
| 902 | visit(imageSourceProcessor.runSync(imageSourceProcessor.parse(markdown)) as ImageSourceNode); |
| 903 | return sources; |
| 904 | } |
| 905 | |
| 906 | const linkDestinationProcessor = unified().use(remarkParse).use(remarkKeepWindowsPaths); |
| 907 | |
| 908 | // An angle destination cannot hold an unescaped `<` or `>`, or end on a lone backslash. |
| 909 | const ANGLE_DESTINATION_CONTENT = /^(?:[^\\<>]|\\.)*$/; |
| 910 | |
| 911 | /** Reads a destination from `extractMarkdownLinkHrefs` as the renderer's parser does. */ |
| 912 | export function decodeMarkdownLinkEscapes(href: string): string { |
| 913 | // Only the parser knows character references, so destinations that may have one go through it. |
| 914 | if (href.includes("&")) { |
| 915 | // The extractor drops angle brackets, and only that form keeps an unmatched `)` in the path. |
| 916 | const destination = ANGLE_DESTINATION_CONTENT.test(href) ? `<${href}>` : href; |
| 917 | const paragraph = linkDestinationProcessor.parse(`[](${destination})`).children[0]; |
| 918 | const link = paragraph?.type === "paragraph" ? paragraph.children[0] : undefined; |
| 919 | if (link?.type === "link") return link.url; |
| 920 | } |
| 921 | return isWindowsAbsolutePath(href) ? href : href.replace(/\\([!-/:-@[-`{-~])/g, "$1"); |
| 922 | } |