diff options
| -rw-r--r-- | castle-search/README.md | 57 | ||||
| -rw-r--r-- | castle-search/src/cdp.ts | 28 | ||||
| -rw-r--r-- | castle-search/src/http.ts | 291 | ||||
| -rw-r--r-- | castle-search/src/imap.ts | 148 | ||||
| -rw-r--r-- | castle-search/src/search.ts | 21 | ||||
| -rw-r--r-- | castle-search/src/totp.ts | 39 | ||||
| -rw-r--r-- | castle-search/src/x-limits.ts | 109 | ||||
| -rw-r--r-- | castle-search/src/x-login.ts | 259 | ||||
| -rw-r--r-- | castle-search/src/x-search.ts | 467 | ||||
| -rw-r--r-- | castle-search/src/youtube.ts | 303 | ||||
| -rw-r--r-- | castle-search/test/http.test.ts | 70 | ||||
| -rw-r--r-- | castle-search/test/x-limits.test.ts | 98 | ||||
| -rw-r--r-- | castle-search/test/x-login.test.ts | 91 | ||||
| -rw-r--r-- | castle-search/test/youtube.test.ts | 72 |
14 files changed, 1900 insertions, 153 deletions
diff --git a/castle-search/README.md b/castle-search/README.md index 0e45aaf..1621fe3 100644 --- a/castle-search/README.md +++ b/castle-search/README.md @@ -47,18 +47,57 @@ switch engines, because this protocol has no field for one. There is no authentication: bind it to loopback (the default) or a trusted address. `GET /health` answers `{"status": true}`. -### X search +### X.com: search, posts, profiles, trends -`POST /x-search` searches X (Twitter) instead: `{"query": "from:nasa", "count": -10, "sort": "latest"}` in, `{"query", "sort", "posts": [...]}` out, each post -with its link, author, time, text, counts and any quoted post. The query is X's -own search syntax. It needs the browser signed in to X — X serves search to -nobody else — and it says so when it is not, rather than returning nothing. -`GET /openapi.json` describes it, so Open WebUI can offer it as a tool. Free per -search, where xAI's own X search charges per post returned. +Four reads of X (Twitter), each `POST` a JSON object and get JSON back: + +| Route | In | Out | +|---|---|---| +| `/x-search` | `query` (X's own search syntax), `count`, `sort` (`latest` or `top`) | the posts | +| `/x-post` | `url` (a post link or its id), `replies` | the post whole, the `thread` above it when it is a reply, its top `replies` | +| `/x-user` | `handle`, `count` | the `profile` (bio, location, website, joined, exact follower, following and post counts) and its latest posts | +| `/x-trends` | `tab`: `trending`, `news`, `sports`, `entertainment`, `for_you` | the trends, each with a `query` for `/x-search` | + +Each post comes with its link, author, time, text, counts, photos (X's image +link and alt text), videos (a still), a link card and any quoted post. A search +result X cut short is marked `truncated`; `/x-post` reads it whole. It needs the +browser signed in to X, since X serves none of this to anybody else, and it +says so when it is not, rather than returning nothing. `GET /openapi.json` +describes all four, as "X.com", so Open WebUI can offer them as tools. Free, +where xAI's own X search charges per post returned. Trends are for the +signed-in account's region. X keeps only the posts near the screen in the page, so it scrolls and collects -by link, up to 20 posts. A search takes 4–7 s. +by link, up to 20 posts (or replies). A read takes 4–7 s. + +All four share one set of limits, so the account reads X like a person, not a +scraper (`src/x-limits.ts`): one read at a time, 30 s ± a random 0–25 s between them, +drawn afresh each time (a read that comes sooner waits), and at most 30 an hour and 150 a day, beyond which a +read fails saying when the next may run. The counts live in memory, so a +restart starts them afresh. + +Signed out, it signs back in by itself when `CCS_X_LOGIN_FILE` names a JSON +file of the account's details (`src/x-login.ts`): `username`, `password`, +`totpSecret` (the authenticator app's base32 secret) and, optionally, `imap` +(`host`, `user`, `password`) for the code X emails when a sign-in looks +unusual, read from the subject over IMAP (`src/imap.ts`). Then it searches +again, once. It tries at most once every 30 minutes, never sends the password +twice in one try, and stops at any step it does not recognise, such as a +captcha, saying a person has to sign in. A sign-in takes about 6 s. + +### YouTube transcripts + +`POST /youtube-transcript` `{"url": "https://youtu.be/...", "language": "en"}` +files a YouTube video's transcript as plain text in `CCS_TRANSCRIPT_DIR`, named +`<upload date>-<Title-With-Dashes>.txt`, and returns it with the title, +channel, upload date and length (`src/youtube.ts`). Captions only, never the +video: human-written when the video has them, YouTube's automatic ones +otherwise, with the rolling repeats of automatic captions taken out. A video +filed before is read back from the folder. yt-dlp (`CCS_YTDLP`) updates +itself at most daily, and again when a fetch fails; when YouTube rate-limits +captions (HTTP 429), the text comes from youtube-transcript-api +(`CCS_YT_PYTHON`) instead. `GET /youtube/openapi.json` describes it apart from +X.com, as "YouTube". The recipe is Hermes's, from bambos. ## Engines diff --git a/castle-search/src/cdp.ts b/castle-search/src/cdp.ts index aa01b56..64c382e 100644 --- a/castle-search/src/cdp.ts +++ b/castle-search/src/cdp.ts @@ -242,7 +242,12 @@ export class PageSession { this.#sessionId = sessionId; } - static async open(conn: BrowserConnection, signal?: AbortSignal): Promise<PageSession> { + /** + * `browserContextId`, from Target.createBrowserContext, opens the page in a + * context of its own, with no cookies: used to try a sign-in without + * touching the browser's own signed-in session. + */ + static async open(conn: BrowserConnection, signal?: AbortSignal, browserContextId?: string): Promise<PageSession> { if (signal?.aborted) throw abortError(signal); // Deliberately NOT raced against the abort signal. raceAbort abandons the @@ -251,7 +256,11 @@ export class PageSession { // using, with nothing left holding its targetId to close it. It is a fast // command already bounded by COMMAND_TIMEOUT_MS, so waiting it out costs // little and guarantees we own whatever it created. - const created = await conn.send("Target.createTarget", { url: "about:blank", background: true }); + const created = await conn.send("Target.createTarget", { + url: "about:blank", + background: true, + ...(browserContextId ? { browserContextId } : {}), + }); const targetId = created["targetId"]; if (typeof targetId !== "string") { throw new BrowserUnavailableError("Target.createTarget did not return a targetId"); @@ -331,6 +340,21 @@ export class PageSession { } /** + * Types `text` into whatever has focus, as a paste would: one input event, so + * a framework-controlled field (React's) sees it. Never logged. + */ + async insertText(text: string, signal?: AbortSignal): Promise<void> { + await raceAbort(this.#conn.send("Input.insertText", { text }, this.#sessionId), signal); + } + + /** Presses and releases Enter in whatever has focus. */ + async pressEnter(signal?: AbortSignal): Promise<void> { + const key = { key: "Enter", code: "Enter", windowsVirtualKeyCode: 13 }; + await raceAbort(this.#conn.send("Input.dispatchKeyEvent", { type: "keyDown", ...key, text: "\r" }, this.#sessionId), signal); + await raceAbort(this.#conn.send("Input.dispatchKeyEvent", { type: "keyUp", ...key }, this.#sessionId), signal); + } + + /** * Close the tab. Idempotent and never throws: it runs on error paths, where a * second failure would mask the one worth reporting. */ diff --git a/castle-search/src/http.ts b/castle-search/src/http.ts index 17c8c70..076c624 100644 --- a/castle-search/src/http.ts +++ b/castle-search/src/http.ts @@ -20,17 +20,36 @@ * an address only trusted peers reach. Each search opens and closes one * background tab in a browser someone may be looking at. * - * POST /x-search searches X instead (x-search.ts), through the browser's - * signed-in X session, and GET /openapi.json describes it, so that Open WebUI - * can offer it as a tool: `{ "query", "count", "sort" }` in, `{ "query", "sort", - * "posts": [...] }` out, 502 with `{ "error" }` on failure. + * POST /x-search, /x-post, /x-user and /x-trends read X instead (x-search.ts), + * through the browser's signed-in X session, and GET /openapi.json describes + * them, so that Open WebUI can offer them as tools ("X.com"). POST + * /youtube-transcript files a YouTube video's transcript and returns it + * (youtube.ts), described at GET /youtube/openapi.json ("YouTube"). Each + * answers 502 with `{ "error" }` on failure. */ import { createServer, type IncomingMessage, type ServerResponse } from "node:http"; import { messageOf } from "./errors.ts"; import { disconnect, search, type SearchOutcome, type SearchRequest } from "./search.ts"; -import { xSearch, type XSearchOutcome, type XSearchRequest } from "./x-search.ts"; +import { + handleOf, + postIdOf, + TREND_TABS, + type XPostOutcome, + type XPostRequest, + xPost, + xSearch, + type XSearchOutcome, + type XSearchRequest, + type XTrendsOutcome, + type XTrendTab, + xTrends, + xUser, + type XUserOutcome, + type XUserRequest, +} from "./x-search.ts"; +import { DEFAULT_MAX_CHARS, type TranscriptOutcome, type TranscriptRequest, videoIdOf, youtubeTranscript } from "./youtube.ts"; export const DEFAULT_LISTEN = "127.0.0.1:11158"; const MAX_BODY_BYTES = 64 * 1024; @@ -75,47 +94,124 @@ export const handleSearch = async (raw: string, run: Searcher, signal: AbortSign }; export type XSearcher = (request: XSearchRequest, signal: AbortSignal) => Promise<XSearchOutcome>; +export type XPostReader = (request: XPostRequest, signal: AbortSignal) => Promise<XPostOutcome>; +export type XUserReader = (request: XUserRequest, signal: AbortSignal) => Promise<XUserOutcome>; +export type XTrendsReader = (tab: XTrendTab, signal: AbortSignal) => Promise<XTrendsOutcome>; const X_DEFAULT_COUNT = 10; -/** POST /x-search: `{ query, count?, sort? }` in, `{ query, sort, posts }` out. */ -export const handleXSearch = async (raw: string, run: XSearcher, signal: AbortSignal): Promise<Reply> => { - let parsed: unknown; +const jsonBody = (raw: string): Record<string, unknown> | Reply => { try { - parsed = JSON.parse(raw); + const parsed: unknown = JSON.parse(raw || "{}"); + if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) return parsed as Record<string, unknown>; } catch { - return { status: 400, body: { error: "request body is not JSON" } }; + // fall through } - const { query, count, sort } = (parsed ?? {}) as { query?: unknown; count?: unknown; sort?: unknown }; + return { status: 400, body: { error: "request body is not a JSON object" } }; +}; +const isReply = (v: unknown): v is Reply => typeof (v as Reply).status === "number" && "body" in (v as Reply); + +const countOf = (v: unknown, fallback: number, min = 1): number => { + const n = typeof v === "number" && Number.isFinite(v) ? Math.trunc(v) : fallback; + return Math.min(MAX_COUNT, Math.max(min, n)); +}; + +const answer = async (work: () => Promise<unknown>): Promise<Reply> => { + try { + return { status: 200, body: await work() }; + } catch (e) { + return { status: 502, body: { error: messageOf(e) } }; + } +}; + +/** POST /x-search: `{ query, count?, sort? }` in, `{ query, sort, posts }` out. */ +export const handleXSearch = async (raw: string, run: XSearcher, signal: AbortSignal): Promise<Reply> => { + const body = jsonBody(raw); + if (isReply(body)) return body; + const { query, count, sort } = body; if (typeof query !== "string" || query.trim() === "") { return { status: 400, body: { error: "query must be a non-empty string" } }; } if (sort !== undefined && sort !== null && sort !== "latest" && sort !== "top") { return { status: 400, body: { error: 'sort must be "latest" or "top"' } }; } - const n = typeof count === "number" && Number.isFinite(count) ? Math.trunc(count) : X_DEFAULT_COUNT; - try { - const outcome = await run( - { query: query.trim(), count: Math.min(MAX_COUNT, Math.max(1, n)), sort: sort ?? "latest" }, - signal, - ); - return { status: 200, body: { query: outcome.query, sort: outcome.sort, posts: outcome.posts } }; - } catch (e) { - return { status: 502, body: { error: messageOf(e) } }; + return answer(async () => { + const outcome = await run({ query: query.trim(), count: countOf(count, X_DEFAULT_COUNT), sort: sort ?? "latest" }, signal); + return { query: outcome.query, sort: outcome.sort, posts: outcome.posts }; + }); +}; + +/** POST /x-post: `{ url, replies? }` in, `{ post, thread, replies }` out. */ +export const handleXPost = async (raw: string, run: XPostReader, signal: AbortSignal): Promise<Reply> => { + const body = jsonBody(raw); + if (isReply(body)) return body; + if (typeof body["url"] !== "string" || !postIdOf(body["url"])) { + return { status: 400, body: { error: "url must be a link to a post on X (x.com/<handle>/status/<id>) or its id" } }; + } + const url = body["url"]; + return answer(() => run({ url, replies: countOf(body["replies"], X_DEFAULT_COUNT, 0) }, signal)); +}; + +/** POST /x-user: `{ handle, count? }` in, `{ profile, posts }` out. */ +export const handleXUser = async (raw: string, run: XUserReader, signal: AbortSignal): Promise<Reply> => { + const body = jsonBody(raw); + if (isReply(body)) return body; + if (typeof body["handle"] !== "string" || !handleOf(body["handle"])) { + return { status: 400, body: { error: "handle must be an X handle, such as nasa or @nasa" } }; + } + const handle = body["handle"]; + return answer(() => run({ handle, count: countOf(body["count"], X_DEFAULT_COUNT, 0) }, signal)); +}; + +/** POST /x-trends: `{ tab? }` in, `{ tab, trends }` out. */ +export const handleXTrends = async (raw: string, run: XTrendsReader, signal: AbortSignal): Promise<Reply> => { + const body = jsonBody(raw); + if (isReply(body)) return body; + const tab = body["tab"] ?? "trending"; + if (!TREND_TABS.includes(tab as XTrendTab)) { + return { status: 400, body: { error: `tab must be one of ${TREND_TABS.join(", ")}` } }; + } + return answer(() => run(tab as XTrendTab, signal)); +}; + +export type TranscriptFetcher = (request: TranscriptRequest, signal: AbortSignal) => Promise<TranscriptOutcome>; + +/** POST /youtube-transcript: `{ url, language?, max_chars? }` in, the transcript and where it was filed out. */ +export const handleTranscript = async (raw: string, run: TranscriptFetcher, signal: AbortSignal): Promise<Reply> => { + const body = jsonBody(raw); + if (isReply(body)) return body; + if (typeof body["url"] !== "string" || !videoIdOf(body["url"])) { + return { status: 400, body: { error: "url must be a YouTube video link (youtube.com/watch?v=..., youtu.be/..., shorts/...)" } }; + } + const language = body["language"]; + if (language !== undefined && (typeof language !== "string" || !/^[A-Za-z]{2,3}(-[A-Za-z0-9]+)?$/.test(language))) { + return { status: 400, body: { error: "language must be a language code such as en, pt or en-GB" } }; } + const max = body["max_chars"]; + const maxChars = typeof max === "number" && Number.isFinite(max) ? Math.max(1000, Math.trunc(max)) : DEFAULT_MAX_CHARS; + const url = body["url"]; + return answer(() => run({ url, language: language as string | undefined, maxChars }, signal)); +}; + +const jsonSchema = (properties: Record<string, unknown>, required: string[]) => ({ + required: true, + content: { "application/json": { schema: { type: "object", required, properties } } }, +}); +const RESPONSES = { + "200": { description: "What X shows", content: { "application/json": { schema: { type: "object" } } } }, + "502": { description: "X could not be read; error says why", content: { "application/json": { schema: { type: "object" } } } }, }; /** - * The OpenAPI description Open WebUI reads to offer /x-search as a tool. Its - * descriptions are what the model sees, so they say what the tool is for and - * how to write a query. + * The OpenAPI description Open WebUI reads to offer the X reads as tools. Its + * descriptions are what the model sees, so they say what each is for. */ export const OPENAPI = { openapi: "3.1.0", info: { - title: "castle X search", - version: "1.0.0", - description: "Searches posts on X (Twitter) through a signed-in browser. Free per search.", + title: "X.com", + version: "2.0.0", + description: "Reads X (Twitter) through a signed-in browser: search, posts with replies, profiles, trends. Free.", }, paths: { "/x-search": { @@ -127,26 +223,103 @@ export const OPENAPI = { "reactions to news, posts about a topic. The query uses X's search syntax: from:handle, to:handle, " + '"exact phrase", OR, -word, since:YYYY-MM-DD, until:YYYY-MM-DD, min_faves:N, filter:media, ' + "-filter:replies, lang:en. A handle is written without @ after from:. Each post comes with its " + - "link: cite it. Text cut short by X is marked truncated.", - requestBody: { - required: true, - content: { - "application/json": { - schema: { - type: "object", - required: ["query"], - properties: { - query: { type: "string", description: "X search query, e.g. from:nasa since:2026-11-01" }, - count: { type: "integer", minimum: 1, maximum: MAX_COUNT, default: X_DEFAULT_COUNT, description: "How many posts" }, - sort: { type: "string", enum: ["latest", "top"], default: "latest", description: "latest: newest first; top: most engaged" }, - }, - }, - }, + "link: cite it. A post marked truncated was cut short by X: read_x_post gives it whole. Photos " + + "come with their image link and alt text.", + requestBody: jsonSchema( + { + query: { type: "string", description: "X search query, e.g. from:nasa since:2026-11-01" }, + count: { type: "integer", minimum: 1, maximum: MAX_COUNT, default: X_DEFAULT_COUNT, description: "How many posts" }, + sort: { type: "string", enum: ["latest", "top"], default: "latest", description: "latest: newest first; top: most engaged" }, }, - }, + ["query"], + ), + responses: RESPONSES, + }, + }, + "/x-post": { + post: { + operationId: "read_x_post", + summary: "Read one post on X, whole, with its thread and replies", + description: + "Read a post on X from its link: its full text (never truncated), photos with their image link " + + "and alt text, its link card, the posts above it when it is a reply (thread, oldest first), and " + + "the top replies under it. Use it to read a long post, a thread, or what people answered.", + requestBody: jsonSchema( + { + url: { type: "string", description: "The post's link, x.com/<handle>/status/<id>, or the id" }, + replies: { type: "integer", minimum: 0, maximum: MAX_COUNT, default: X_DEFAULT_COUNT, description: "How many replies" }, + }, + ["url"], + ), + responses: RESPONSES, + }, + }, + "/x-user": { + post: { + operationId: "read_x_profile", + summary: "Read an X profile and its latest posts", + description: + "Read an X account's profile: name, bio, location, website, when it joined, exact follower, " + + "following and post counts, then its latest posts as its profile shows them (a pinned post and " + + "reposts included, marked in context).", + requestBody: jsonSchema( + { + handle: { type: "string", description: "The account's handle, such as nasa or @nasa" }, + count: { type: "integer", minimum: 0, maximum: MAX_COUNT, default: X_DEFAULT_COUNT, description: "How many of its posts" }, + }, + ["handle"], + ), + responses: RESPONSES, + }, + }, + "/x-trends": { + post: { + operationId: "x_trends", + summary: "What is trending on X now", + description: + "List what X shows as trending now, for the signed-in account's region (the United Kingdom). " + + "Each trend has a query: pass it to search_x for the posts.", + requestBody: jsonSchema( + { tab: { type: "string", enum: [...TREND_TABS], default: "trending", description: "Which explore tab" } }, + [], + ), + responses: RESPONSES, + }, + }, + }, +} as const; + +/** The YouTube tool's own description, so Open WebUI lists it apart from X.com. */ +export const YOUTUBE_OPENAPI = { + openapi: "3.1.0", + info: { + title: "YouTube", + version: "1.0.0", + description: "A YouTube video's transcript, filed in Igor's Nextcloud transcript folder and returned.", + }, + paths: { + "/youtube-transcript": { + post: { + operationId: "youtube_transcript", + summary: "Get a YouTube video's transcript", + description: + "Fetch a YouTube video's transcript from its link: the video's captions (human-written when it " + + "has them, YouTube's automatic ones otherwise) as clean prose, with its title, channel, upload " + + "date and length. It is also filed in Igor's Nextcloud, in the transcript folder, as " + + "<upload date>-<Title>.txt; a video filed before is read from there. Use it whenever a YouTube " + + "link comes up and its content matters: to summarise it, answer questions about it, or quote " + + "it. Automatic captions garble names and numbers: say so when one matters.", + requestBody: jsonSchema( + { + url: { type: "string", description: "The video's link: youtube.com/watch?v=..., youtu.be/..., shorts/..." }, + language: { type: "string", default: "en", description: "Caption language code, such as en or pt" }, + max_chars: { type: "integer", minimum: 1000, default: DEFAULT_MAX_CHARS, description: "Most characters of text to return; the file always has all of it" }, + }, + ["url"], + ), responses: { - "200": { description: "Matching posts", content: { "application/json": { schema: { type: "object" } } } }, - "502": { description: "The search failed; error says why", content: { "application/json": { schema: { type: "object" } } } }, + "200": { description: "The transcript and where it was filed", content: { "application/json": { schema: { type: "object" } } } }, + "502": { description: "No transcript; error says why", content: { "application/json": { schema: { type: "object" } } } }, }, }, }, @@ -199,9 +372,20 @@ const main = (): void => { send(res, { status: 200, body: OPENAPI }); return; } - const isX = path === "/x-search"; - if (req.method !== "POST" || (path !== "/" && path !== "/search" && !isX)) { - send(res, { status: 404, body: { error: "POST a search to / or /search, or an X search to /x-search" } }); + if (req.method === "GET" && path === "/youtube/openapi.json") { + send(res, { status: 200, body: YOUTUBE_OPENAPI }); + return; + } + const xRoutes: Record<string, (raw: string, signal: AbortSignal) => Promise<Reply>> = { + "/x-search": (raw, signal) => handleXSearch(raw, xSearch, signal), + "/x-post": (raw, signal) => handleXPost(raw, xPost, signal), + "/x-user": (raw, signal) => handleXUser(raw, xUser, signal), + "/x-trends": (raw, signal) => handleXTrends(raw, xTrends, signal), + "/youtube-transcript": (raw, signal) => handleTranscript(raw, youtubeTranscript, signal), + }; + const xRoute = xRoutes[path ?? ""]; + if (req.method !== "POST" || (path !== "/" && path !== "/search" && !xRoute)) { + send(res, { status: 404, body: { error: "POST a search to / or /search, an X read to /x-search, /x-post, /x-user or /x-trends, or a YouTube link to /youtube-transcript" } }); return; } // Stop the search, and free its tab, if the client goes away first. @@ -217,16 +401,9 @@ const main = (): void => { return; } const started = Date.now(); - const reply = isX - ? await handleXSearch(raw, xSearch, controller.signal) - : await handleSearch(raw, runSearch, controller.signal); - const outcome = - reply.status !== 200 - ? `error ${reply.status}` - : isX - ? `${(reply.body as { posts: unknown[] }).posts.length} posts` - : `${(reply.body as unknown[]).length} results`; - console.log(`${isX ? "x-search" : "search"} ${outcome} in ${Date.now() - started} ms`); + const reply = xRoute ? await xRoute(raw, controller.signal) : await handleSearch(raw, runSearch, controller.signal); + const outcome = reply.status !== 200 ? `error ${reply.status}` : xRoute ? "ok" : `${(reply.body as unknown[]).length} results`; + console.log(`${xRoute ? path!.slice(1) : "search"} ${outcome} in ${Date.now() - started} ms`); if (reply.status !== 200) console.log(` ${(reply.body as { error: string }).error}`); send(res, reply); }); diff --git a/castle-search/src/imap.ts b/castle-search/src/imap.ts new file mode 100644 index 0000000..ef02979 --- /dev/null +++ b/castle-search/src/imap.ts @@ -0,0 +1,148 @@ +/** + * Just enough IMAP to read a login code X emailed: log in over TLS, look at the + * inbox read-only, read the subjects of X's recent mail. No dependency, so + * castle-search keeps having none. + * + * X puts the code in the subject ("Your X confirmation code is abc12def"), so + * only headers are fetched, and nothing is marked read or changed. + */ + +import { connect, type TLSSocket } from "node:tls"; + +export type ImapAccount = { host: string; port?: number; user: string; password: string }; + +export type MailHeader = { received: number; from: string; subject: string }; + +/** RFC 2047 encoded words, the B and Q forms, as subjects use them. */ +export const decodeWords = (s: string): string => + s + .replace(/\?=\s+=\?/g, "?==?") + .replace(/=\?([^?]+)\?([BbQq])\?([^?]*)\?=/g, (_, charset: string, enc: string, text: string) => { + const bytes = + enc.toUpperCase() === "B" + ? Buffer.from(text, "base64") + : Buffer.from( + text.replace(/_/g, " ").replace(/=([0-9A-Fa-f]{2})/g, (_m, h: string) => String.fromCharCode(parseInt(h, 16))), + "latin1", + ); + try { + return new TextDecoder(charset).decode(bytes); + } catch { + return bytes.toString("utf8"); + } + }); + +/** The code in one of X's subjects, or undefined when the mail is something else. */ +export const xCodeFromSubject = (subject: string): string | undefined => { + const m = + /\bcode is:?\s*([A-Za-z0-9]{5,10})\b/i.exec(subject) ?? + /^\s*([A-Za-z0-9]{5,10}) is your X (?:verification|confirmation|login) code\b/i.exec(subject); + return m?.[1]; +}; + +/** "18-Nov-2026 10:16:23 +0000" -> epoch ms. */ +export const parseInternalDate = (s: string): number => Date.parse(s.replace(/^(\d{1,2})-(\w{3})-(\d{4})/, "$1 $2 $3")); + +/** Splits a FETCH response into one header set per message. */ +export const parseFetch = (raw: string): MailHeader[] => { + const out: MailHeader[] = []; + for (const part of raw.split(/\r\n(?=\* \d+ FETCH )/)) { + const date = /INTERNALDATE "([^"]+)"/.exec(part)?.[1]; + if (!date) continue; + const unfolded = part.replace(/\r\n[ \t]+/g, " "); + const header = (name: string) => new RegExp(`^${name}:\\s*(.*)$`, "im").exec(unfolded)?.[1]?.trim() ?? ""; + out.push({ received: parseInternalDate(date), from: decodeWords(header("From")), subject: decodeWords(header("Subject")) }); + } + return out; +}; + +const quote = (s: string): string => `"${s.replace(/\\/g, "\\\\").replace(/"/g, '\\"')}"`; + +const MONTHS = ["Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct", "Nov", "Dec"]; +const imapDate = (ms: number): string => { + const d = new Date(ms); + return `${d.getUTCDate()}-${MONTHS[d.getUTCMonth()]}-${d.getUTCFullYear()}`; +}; + +class Session { + readonly #socket: TLSSocket; + #buffer = ""; + #tag = 0; + #waiter: (() => void) | undefined; + #error: Error | undefined; + + constructor(socket: TLSSocket) { + this.#socket = socket; + socket.setEncoding("latin1"); + socket.on("data", (chunk: string) => { + this.#buffer += chunk; + this.#waiter?.(); + }); + const fail = (e: Error) => { + this.#error ??= e; + this.#waiter?.(); + }; + socket.on("error", fail); + socket.on("close", () => fail(new Error("the IMAP server closed the connection"))); + } + + async #until(done: (buffer: string) => boolean, signal: AbortSignal): Promise<string> { + while (!done(this.#buffer)) { + if (this.#error) throw this.#error; + if (signal.aborted) throw signal.reason; + await new Promise<void>((resolve) => { + this.#waiter = resolve; + signal.addEventListener("abort", () => resolve(), { once: true }); + }); + this.#waiter = undefined; + } + const out = this.#buffer; + this.#buffer = ""; + return out; + } + + greeting(signal: AbortSignal): Promise<string> { + return this.#until((b) => b.includes("\r\n"), signal); + } + + /** Sends one command; `shown` is what an error may say was sent, so a password never is. */ + async command(line: string, signal: AbortSignal, shown = line): Promise<string> { + const tag = `c${++this.#tag}`; + this.#socket.write(`${tag} ${line}\r\n`); + const reply = await this.#until((b) => new RegExp(`(^|\\r\\n)${tag} (OK|NO|BAD)[^\\r\\n]*\\r\\n`).test(b), signal); + const status = new RegExp(`(?:^|\\r\\n)${tag} (OK|NO|BAD)([^\\r\\n]*)`).exec(reply); + if (status?.[1] !== "OK") throw new Error(`IMAP ${shown.split(" ")[0]} failed: ${status?.[1]}${status?.[2]}`); + return reply; + } + + close(): void { + this.#socket.end(); + this.#socket.destroy(); + } +} + +/** Headers of mail from `fromDomain` received at or after `sinceMs`, newest first. */ +export const recentMail = async ( + account: ImapAccount, + fromDomain: string, + sinceMs: number, + signal: AbortSignal, +): Promise<MailHeader[]> => { + const socket = connect({ host: account.host, port: account.port ?? 993, servername: account.host }); + const s = new Session(socket); + try { + await s.greeting(signal); + await s.command(`LOGIN ${quote(account.user)} ${quote(account.password)}`, signal, "LOGIN"); + await s.command("EXAMINE INBOX", signal); + const found = await s.command(`SEARCH SINCE ${imapDate(sinceMs - 86_400_000)} FROM ${quote(fromDomain)}`, signal); + const ids = (/\* SEARCH([\d ]*)/.exec(found)?.[1] ?? "").trim().split(/\s+/).filter(Boolean).slice(-10); + if (ids.length === 0) return []; + const raw = await s.command(`FETCH ${ids.join(",")} (INTERNALDATE BODY.PEEK[HEADER.FIELDS (FROM SUBJECT)])`, signal); + await s.command("LOGOUT", signal).catch(() => ""); + return parseFetch(raw) + .filter((m) => m.received >= sinceMs) + .sort((a, b) => b.received - a.received); + } finally { + s.close(); + } +}; diff --git a/castle-search/src/search.ts b/castle-search/src/search.ts index 0178349..1ddd523 100644 --- a/castle-search/src/search.ts +++ b/castle-search/src/search.ts @@ -402,16 +402,31 @@ const isGoogleHost = (url: string): boolean => /^https?:\/\/([a-z0-9-]+\.)*googl export const withPage = async <T>( signal: AbortSignal, fn: (page: PageSession, endpoint: string) => Promise<T>, + options: { isolated?: boolean } = {}, ): Promise<T> => { const release = await acquire(signal); try { const conn = await ensureConnection(signal); const endpoint = describeCdpTarget(conn.target); - const page = await PageSession.open(conn, signal); + // isolated: a browser context of its own, with none of the profile's + // cookies, thrown away afterwards with everything in it. + let context: string | undefined; + if (options.isolated) { + const created = await conn.send("Target.createBrowserContext", { disposeOnDetach: true }); + if (typeof created["browserContextId"] !== "string") { + throw new BrowserUnavailableError("Target.createBrowserContext did not return a browserContextId"); + } + context = created["browserContextId"]; + } try { - return await fn(page, endpoint); + const page = await PageSession.open(conn, signal, context); + try { + return await fn(page, endpoint); + } finally { + await page.close(); + } } finally { - await page.close(); + if (context) await conn.send("Target.disposeBrowserContext", { browserContextId: context }).catch(() => {}); } } finally { release(); diff --git a/castle-search/src/totp.ts b/castle-search/src/totp.ts new file mode 100644 index 0000000..c14a906 --- /dev/null +++ b/castle-search/src/totp.ts @@ -0,0 +1,39 @@ +/** + * TOTP (RFC 6238): the six-digit code an authenticator app shows, from the + * base32 secret behind its QR code. SHA-1, 30-second steps, as X uses. + */ + +import { createHmac } from "node:crypto"; + +const B32 = "ABCDEFGHIJKLMNOPQRSTUVWXYZ234567"; + +export const base32Decode = (input: string): Buffer => { + const clean = input.toUpperCase().replace(/[\s=-]/g, ""); + let bits = 0; + let value = 0; + const out: number[] = []; + for (const ch of clean) { + const i = B32.indexOf(ch); + if (i < 0) throw new Error("the TOTP secret is not base32"); + value = (value << 5) | i; + bits += 5; + if (bits >= 8) { + out.push((value >>> (bits - 8)) & 0xff); + bits -= 8; + } + } + return Buffer.from(out); +}; + +export const totp = (secret: string, nowMs: number = Date.now(), digits = 6, stepS = 30): string => { + const counter = Buffer.alloc(8); + counter.writeBigUInt64BE(BigInt(Math.floor(nowMs / 1000 / stepS))); + const h = createHmac("sha1", base32Decode(secret)).update(counter).digest(); + const o = h[h.length - 1]! & 0x0f; + const n = (h.readUInt32BE(o) & 0x7fffffff) % 10 ** digits; + return String(n).padStart(digits, "0"); +}; + +/** Seconds the code for `nowMs` stays valid. */ +export const totpRemaining = (nowMs: number = Date.now(), stepS = 30): number => + stepS - (Math.floor(nowMs / 1000) % stepS); diff --git a/castle-search/src/x-limits.ts b/castle-search/src/x-limits.ts new file mode 100644 index 0000000..bdc4229 --- /dev/null +++ b/castle-search/src/x-limits.ts @@ -0,0 +1,109 @@ +/** + * Limits on X searches, so the signed-in account reads X like a person rather + * than a scraper, and X has no reason to lock it. + * + * - one search at a time; + * - GAP_MS ± up to JITTER_MS between the starts of two searches, drawn + * afresh each time (5-55 s), so they do not come at a machine's even + * beat: one that comes sooner waits its turn rather than failing; + * - at most PER_HOUR in any rolling hour and PER_DAY in any rolling day: + * beyond that a search fails at once, saying when the next one may run. + * + * Counts are kept in memory: a restart of the service starts them afresh. + */ + +/** Owner's choice, 2026-11-18: 30 s plus or minus a random 0-25 s. */ +export const GAP_MS = 30_000; +export const JITTER_MS = 25_000; +export const PER_HOUR = 30; +export const PER_DAY = 150; +const HOUR_MS = 3_600_000; +const DAY_MS = 24 * HOUR_MS; + +export class XRateLimitError extends Error { + override readonly name = "XRateLimitError"; +} + +export type Clock = { + now: () => number; + sleep: (ms: number, signal: AbortSignal) => Promise<void>; + /** Uniform in [0, 1), as Math.random. */ + random: () => number; +}; + +const realSleep = (ms: number, signal: AbortSignal): Promise<void> => + new Promise((resolve, reject) => { + if (signal.aborted) return reject(signal.reason); + const t = setTimeout(resolve, ms); + signal.addEventListener("abort", () => { + clearTimeout(t); + reject(signal.reason); + }, { once: true }); + }); + +export const realClock: Clock = { now: Date.now, sleep: realSleep, random: Math.random }; + +export type Limits = { gap: number; jitter: number; hour: number; day: number }; + +const at = (ms: number): string => new Date(ms).toISOString().replace("T", " ").slice(0, 16) + " UTC"; + +export class XLimiter { + readonly #clock: Clock; + readonly #limits: Limits; + /** Start times of the searches in the last day, oldest first. */ + #starts: number[] = []; + /** The tail of the queue: each search waits for the one before it. */ + #tail: Promise<void> = Promise.resolve(); + + constructor(clock: Clock = realClock, limits: Limits = { gap: GAP_MS, jitter: JITTER_MS, hour: PER_HOUR, day: PER_DAY }) { + this.#clock = clock; + this.#limits = limits; + } + + /** Runs `fn` once the limits allow, one at a time; throws XRateLimitError when over a cap. */ + async run<T>(signal: AbortSignal, fn: () => Promise<T>): Promise<T> { + const previous = this.#tail; + let done!: () => void; + this.#tail = new Promise<void>((resolve) => (done = resolve)); + try { + await Promise.race([ + previous, + new Promise<never>((_, reject) => { + if (signal.aborted) reject(signal.reason); + signal.addEventListener("abort", () => reject(signal.reason), { once: true }); + }), + ]); + // An abandoned search ahead of us must not hold the queue: `previous` + // settles when it finishes or is aborted, whichever comes first. + await this.#admit(signal); + return await fn(); + } finally { + done(); + } + } + + async #admit(signal: AbortSignal): Promise<void> { + const now = this.#clock.now(); + this.#starts = this.#starts.filter((t) => t > now - DAY_MS); + const lastHour = this.#starts.filter((t) => t > now - HOUR_MS); + if (this.#starts.length >= this.#limits.day) { + throw new XRateLimitError( + `X search limit reached: ${this.#limits.day} searches in the last 24 hours, to keep the account ` + + `from being locked. The next one can run after ${at(this.#starts[0]! + DAY_MS)}. Tell the user.`, + ); + } + if (lastHour.length >= this.#limits.hour) { + throw new XRateLimitError( + `X search limit reached: ${this.#limits.hour} searches in the last hour, to keep the account ` + + `from being locked. The next one can run after ${at(lastHour[0]! + HOUR_MS)}. Tell the user, ` + + `and do not retry before then.`, + ); + } + const last = this.#starts[this.#starts.length - 1]; + const gap = Math.max(0, this.#limits.gap + (this.#clock.random() * 2 - 1) * this.#limits.jitter); + if (last !== undefined && now - last < gap) { + await this.#clock.sleep(gap - (now - last), signal); + } + this.#starts.push(this.#clock.now()); + } +} diff --git a/castle-search/src/x-login.ts b/castle-search/src/x-login.ts new file mode 100644 index 0000000..d4d33d1 --- /dev/null +++ b/castle-search/src/x-login.ts @@ -0,0 +1,259 @@ +/** + * Signing castle Chrome back in to X, when an X search finds it signed out. + * + * X's sign-in is a series of steps on x.com/i/jf/onboarding/web (2026-11), each + * named in the URL's hash: the username, `login_enter_password`, + * `two_factor_code` (the authenticator app's code), and sometimes a code X + * emails when the sign-in looks unusual. Each step is recognised by its + * fields and words rather than its name alone, since X renames them, and + * answered from the credentials file: + * + * {"username": "...", "password": "...", "totpSecret": "BASE32...", + * "imap": {"host": "imap.gmail.com", "user": "...", "password": "..."}} + * + * `imap` is optional: without it an emailed code fails the sign-in, saying so. + * The file is named by CCS_X_LOGIN_FILE; cloud-setup's macos/castle-search.sh + * writes it from szecret, readable by its owner only. + * + * A step it does not recognise (a captcha, "verify your phone") fails the + * sign-in with what the page said: that needs a person. Nothing secret is + * ever put in an error or a log line. + */ + +import { readFileSync } from "node:fs"; + +import type { PageSession } from "./cdp.ts"; +import { BrowserUnavailableError } from "./errors.ts"; +import { type ImapAccount, recentMail, xCodeFromSubject } from "./imap.ts"; +import { totp, totpRemaining } from "./totp.ts"; + +export type XCredentials = { username: string; password: string; totpSecret?: string; imap?: ImapAccount }; + +export const loadCredentials = (path: string | undefined): XCredentials | undefined => { + if (!path) return undefined; + let parsed: Partial<XCredentials>; + try { + parsed = JSON.parse(readFileSync(path, "utf8")) as Partial<XCredentials>; + } catch (e) { + throw new BrowserUnavailableError(`cannot read the X sign-in file ${path}: ${(e as Error).message}`); + } + if (typeof parsed.username !== "string" || typeof parsed.password !== "string") { + throw new BrowserUnavailableError(`the X sign-in file ${path} needs "username" and "password"`); + } + return parsed as XCredentials; +}; + +export const LOGIN_URL = "https://x.com/i/flow/login"; +/** Whole sign-in, emailed code included: X's mail can take a minute. */ +export const X_LOGIN_TIMEOUT_MS = 150_000; +const MAX_STEPS = 10; +const EMAIL_WAIT_MS = 90_000; +const EMAIL_POLL_MS = 5_000; + +export type LoginStep = + | { kind: "done" } + | { kind: "username" } + | { kind: "password" } + | { kind: "totp" } + | { kind: "email-code" } + | { kind: "wrong-password" } + | { kind: "loading" } + | { kind: "unknown" }; + +export type LoginProbe = { + url: string; + path: string; + hash: string; + text: string; + /** Visible inputs: name, type, autocomplete. */ + inputs: { name: string; type: string; autocomplete: string }[]; + signedIn: boolean; +}; + +/** + * An input a person could see. X keeps a password field on its first page for + * autofill, laid out but at opacity 0 under aria-hidden: not a step to answer. + */ +const SHOWN = `((i) => i.type !== "hidden" && i.offsetParent !== null && + !i.closest('[aria-hidden="true"], [inert]') && getComputedStyle(i).opacity !== "0")`; + +/** Reads the sign-in page as it stands. Runs in the page; self-contained. */ +export const LOGIN_PROBE = `(() => { + const scope = document.querySelector('[role="dialog"]') || document.querySelector("main") || document.body; + const inputs = [...document.querySelectorAll("input")] + .filter(${SHOWN}) + .map((i) => ({ name: i.name || "", type: i.type || "", autocomplete: i.autocomplete || "" })); + return { + url: location.href, + path: location.pathname, + hash: location.hash, + text: (scope ? scope.innerText : "").trim().slice(0, 600), + inputs, + signedIn: !!document.querySelector('[data-testid="SideNav_AccountSwitcher_Button"], [data-testid="AppTabBar_Profile_Link"]'), + }; +})()`; + +const onSignInPage = (path: string): boolean => /^\/(login|i\/flow\/login|i\/jf\/onboarding|account\/access)/.test(path); + +/** Which step the page is on. Pure, so it can be tested on recorded probes. */ +export const classify = (p: LoginProbe): LoginStep => { + if (p.signedIn && !onSignInPage(p.path)) return { kind: "done" }; + const has = (f: (i: LoginProbe["inputs"][number]) => boolean) => p.inputs.some(f); + const text = p.text; + if (/wrong password|incorrect password/i.test(text)) return { kind: "wrong-password" }; + if (has((i) => i.type === "password")) return { kind: "password" }; + const codeField = has((i) => i.autocomplete.includes("one-time-code") || i.name === "text" || i.name === "code" || i.name === "verfication_code"); + if (codeField && /authenticat(or|ion) app|two[- ]factor/i.test(text) && !/email/i.test(text)) return { kind: "totp" }; + if (codeField && /email/i.test(text) && /code/i.test(text)) return { kind: "email-code" }; + if (has((i) => i.name === "username_or_email" || i.autocomplete.startsWith("username"))) return { kind: "username" }; + // "Enter your phone number or username": X asks again when it is unsure. + if (has((i) => i.name === "text") && /phone number or username|username/i.test(text)) return { kind: "username" }; + if (p.inputs.length === 0 && text === "") return { kind: "loading" }; + return { kind: "unknown" }; +}; + +const sleep = (ms: number, signal: AbortSignal): Promise<void> => + new Promise((resolve, reject) => { + if (signal.aborted) return reject(signal.reason); + const t = setTimeout(resolve, ms); + signal.addEventListener("abort", () => { + clearTimeout(t); + reject(signal.reason); + }, { once: true }); + }); + +/** Focuses the step's text field: the password field on the password step, else the first one. */ +const FOCUS = (password: boolean) => `(() => { + const inputs = [...document.querySelectorAll("input")].filter(${SHOWN}); + const el = ${password ? `inputs.find((i) => i.type === "password")` : `inputs.find((i) => i.type !== "password" && i.type !== "checkbox")`}; + if (!el) return false; + el.focus(); + if (el.value) { el.select(); } + return document.activeElement === el; +})()`; + +/** False when the field is gone, the page moving between steps: look again. */ +const answer = async (page: PageSession, text: () => string, password: boolean, signal: AbortSignal): Promise<boolean> => { + if (!(await page.evaluate<boolean>(FOCUS(password), signal).catch(() => false))) return false; + await page.insertText(text(), signal); + await sleep(300, signal); + await page.pressEnter(signal); + return true; +}; + +/** What makes one step differ from the next: not its text, which changes as it is typed into. */ +const shape = (p: LoginProbe): string => `${p.path}${p.hash}|${p.inputs.map((i) => `${i.name}:${i.type}`).join(",")}`; + +/** Waits for X's emailed code, sent at or after `since`. */ +const emailedCode = async (imap: ImapAccount, since: number, signal: AbortSignal): Promise<string> => { + const deadline = Date.now() + EMAIL_WAIT_MS; + let seen = 0; + while (Date.now() < deadline) { + // A minute's slack for the two clocks; an older code is from an earlier sign-in. + const mail = await recentMail(imap, "x.com", since - 60_000, signal); + seen = mail.length; + for (const m of mail) { + const code = xCodeFromSubject(m.subject); + if (code) return code; + } + await sleep(EMAIL_POLL_MS, signal); + } + throw new BrowserUnavailableError( + `X sign-in: X asked for the code it emailed, and none arrived at ${imap.user} within ` + + `${EMAIL_WAIT_MS / 1000} s (${seen} recent mails from x.com, none with a code in the subject). ` + + `The account's email may not be that mailbox.`, + ); +}; + +/** + * Signs in on `page`. Resolves when X shows the signed-in app; throws + * BrowserUnavailableError naming the step it could not get past. + */ +export const xLogin = async ( + page: PageSession, + creds: XCredentials, + signal: AbortSignal, + log: (line: string) => void = () => {}, +): Promise<void> => { + const started = Date.now(); + await page.navigate(LOGIN_URL, signal); + const answered = new Map<string, number>(); + let last: LoginProbe | undefined; + + for (let steps = 0; steps < MAX_STEPS; ) { + // A step's Enter navigates, and a probe caught mid-navigation fails: retry. + const probe = await page.evaluate<LoginProbe>(LOGIN_PROBE, signal).catch((e: unknown) => { + if (signal.aborted) throw e; + return undefined; + }); + if (!probe) { + await sleep(400, signal); + continue; + } + last = probe; + const step = classify(probe); + if (step.kind === "done") { + log(`x-login signed in as ${creds.username}`); + return; + } + if (step.kind === "loading") { + await sleep(400, signal); + continue; + } + // Coming back to a step means the answer was refused: stop rather than + // hammer X with a wrong password or a stale code. The password is never + // sent twice; a code gets one more go, in case it expired on the way. + const key = step.kind; + if (step.kind === "wrong-password" || (answered.get(key) ?? 0) >= (key === "password" ? 1 : 2)) { + throw new BrowserUnavailableError( + `X sign-in refused at the ${step.kind} step (${probe.url}): "${probe.text.slice(0, 200)}". ` + + `Tell the user: someone has to sign in to X in castle Chrome by hand.`, + ); + } + if (step.kind === "unknown") { + throw new BrowserUnavailableError( + `X sign-in stopped at a step this tool does not handle (${probe.url}): "${probe.text.slice(0, 300)}". ` + + `Tell the user: someone has to sign in to X in castle Chrome by hand.`, + ); + } + + let code: (() => string) | undefined; + if (step.kind === "username") code = () => creds.username; + else if (step.kind === "password") code = () => creds.password; + else if (step.kind === "totp") { + const secret = creds.totpSecret; + if (!secret) throw new BrowserUnavailableError("X sign-in wants an authenticator code and the sign-in file has no totpSecret"); + // A code about to expire may be stale by the time X checks it. + if (totpRemaining() < 5) await sleep(totpRemaining() * 1000 + 500, signal); + code = () => totp(secret); + } else if (step.kind === "email-code") { + if (!creds.imap) { + throw new BrowserUnavailableError( + `X sign-in wants a code it emailed, and the sign-in file has no mailbox to read it from ` + + `(${probe.url}). Tell the user: someone has to sign in to X in castle Chrome by hand.`, + ); + } + const emailed = await emailedCode(creds.imap, started, signal); + code = () => emailed; + } + if (!code || !(await answer(page, code, step.kind === "password", signal))) { + await sleep(400, signal); + continue; + } + log(`x-login answered ${step.kind} at ${probe.hash || probe.path}`); + answered.set(key, (answered.get(key) ?? 0) + 1); + steps++; + + // Wait for the next step, up to ~10 s; a refused answer stays on this one. + const before = shape(probe); + for (let w = 0; w < 25; w++) { + await sleep(400, signal); + const now = await page.evaluate<LoginProbe>(LOGIN_PROBE, signal).catch(() => undefined); + if (!now || shape(now) !== before || classify(now).kind !== step.kind) break; + } + } + throw new BrowserUnavailableError( + `X sign-in did not finish within ${MAX_STEPS} steps (last at ${last?.url}). Tell the user: someone has ` + + `to sign in to X in castle Chrome by hand.`, + ); +}; diff --git a/castle-search/src/x-search.ts b/castle-search/src/x-search.ts index 6ea086c..93e8b56 100644 --- a/castle-search/src/x-search.ts +++ b/castle-search/src/x-search.ts @@ -1,26 +1,54 @@ /** - * Search X (x.com) through the browser's own signed-in session. + * Read X (x.com) through the browser's own signed-in session: search posts, + * read one post with its thread and replies, read a profile, list trends. * - * X serves search only to a signed-in session and renders it from JavaScript, - * so the page is the only interface there is: open x.com/search in a background - * tab, wait for posts to render, read them, scroll for more, close the tab. - * Free per search, unlike xAI's x_search ($5 per 1,000 posts returned). + * X serves all of this only to a signed-in session and renders it from + * JavaScript, so the page is the only interface there is: open the page in a + * background tab, wait for it to render, read it, scroll for more, close the + * tab. Free, unlike xAI's x_search ($5 per 1,000 posts returned). * - * The query is X's own search syntax, passed through untouched: from:handle, - * to:handle, since:2026-11-01, until:..., min_faves:100, filter:media, - * -filter:replies, lang:en, "exact phrase", OR. + * A search query is X's own search syntax, passed through untouched: + * from:handle, to:handle, since:2026-11-01, until:..., min_faves:100, + * filter:media, -filter:replies, lang:en, "exact phrase", OR. * - * Every failure throws, as for the engine searches (errors.ts): signed out, an - * X error page, or nothing rendering are reported as such, never as "no posts". - * Only X's own "No results" page is an empty answer, and even that is reported - * as an error naming the query, since an over-narrow query is the usual cause. + * Every read shares one set of limits (x-limits.ts) and, signed out, signs the + * browser back in once (x-login.ts). Every failure throws, as for the engine + * searches (errors.ts): signed out, an X error page, or nothing rendering are + * reported as such, never as an empty answer. X's own "No results" or "doesn't + * exist" page is reported as an error naming what was asked for. */ +import type { PageSession } from "./cdp.ts"; import { BrowserUnavailableError, NoResultsError } from "./errors.ts"; import { withPage } from "./search.ts"; +import { XLimiter } from "./x-limits.ts"; +import { loadCredentials, X_LOGIN_TIMEOUT_MS, xLogin } from "./x-login.ts"; + +/** Shared by every X read in this process: see x-limits.ts. */ +const limiter = new XLimiter(); + +/** + * Signed out, a read signs the browser back in (x-login.ts) when + * CCS_X_LOGIN_FILE names the credentials, then reads again, once. At most one + * sign-in attempt per LOGIN_COOLDOWN_MS: a sign-in that keeps failing is what + * gets an account locked, so until the cooldown has passed every read fails + * with the reason the last attempt gave. + */ +export const LOGIN_COOLDOWN_MS = 30 * 60_000; +let lastLogin: { at: number; error?: string } | undefined; + +class SignedOut extends Error { + readonly at: string; + constructor(at: string) { + super(`signed out at ${at}`); + this.at = at; + } +} export type XSort = "latest" | "top"; +export type XMedia = { type: "photo" | "video"; url?: string; alt?: string }; + export type XPost = { url: string; author: string; @@ -28,7 +56,7 @@ export type XPost = { /** ISO 8601, from the post's <time datetime>. */ time: string; text: string; - /** X cut the text short behind "Show more"; the full post is at url. */ + /** X cut the text short behind "Show more"; x_post reads it whole. */ truncated?: true; /** "@someone reposted", "Pinned" and the like. */ context?: string; @@ -36,18 +64,48 @@ export type XPost = { reposts?: number; likes?: number; views?: number; - /** Count of photos and videos attached. */ - media?: number; - quoted?: { author: string; handle: string; text: string }; + /** Photos (with X's image URL and alt text) and videos (with a still). */ + media?: XMedia[]; + /** A link card: the page the post links to. */ + card?: { url: string; title?: string }; + quoted?: { author: string; handle: string; text: string; time?: string; url?: string }; }; export type XSearchRequest = { query: string; count: number; sort?: XSort | undefined }; - export type XSearchOutcome = { query: string; sort: XSort; searchUrl: string; posts: XPost[] }; +export type XPostRequest = { url: string; replies: number }; +export type XPostOutcome = { + post: XPost; + /** The posts above it in its thread, oldest first, when it is a reply. */ + thread: XPost[]; + replies: XPost[]; +}; + +export type XUserRequest = { handle: string; count: number }; +export type XProfile = { + name: string; + handle: string; + url: string; + bio?: string; + location?: string; + website?: string; + joined?: string; + followers?: number; + following?: number; + posts?: number; + avatar?: string; +}; +export type XUserOutcome = { profile: XProfile; posts: XPost[] }; + +export const TREND_TABS = ["trending", "news", "sports", "entertainment", "for_you"] as const; +export type XTrendTab = (typeof TREND_TABS)[number]; +export type XTrend = { rank?: number; name: string; context?: string; detail?: string; query: string }; +export type XTrendsOutcome = { tab: XTrendTab; trends: XTrend[] }; + /** X shows ~5-10 posts per screen; scroll this many times at most for more. */ const MAX_SCROLLS = 6; -/** Wall clock for one search: render, scrolls and all. */ +/** Wall clock for one read: render, scrolls and all. */ export const X_SEARCH_TIMEOUT_MS = 25_000; const POLL_MS = 400; @@ -58,6 +116,16 @@ export const buildXSearchUrl = (query: string, sort: XSort): string => { return `https://x.com/search?${params.toString()}`; }; +/** The post id in an X or Twitter post link, or a bare id. */ +export const postIdOf = (url: string): string | undefined => + /^\d{5,25}$/.test(url.trim()) ? url.trim() : /(?:x|twitter)\.com\/(?:[^/]+|i(?:\/web)?)\/status(?:es)?\/(\d+)/i.exec(url)?.[1]; + +/** A handle as X writes it, from "@nasa", "nasa" or a profile link. */ +export const handleOf = (s: string): string | undefined => { + const m = /^(?:https?:\/\/(?:www\.)?(?:x|twitter)\.com\/)?@?([A-Za-z0-9_]{1,15})\/?$/.exec(s.trim()); + return m?.[1]; +}; + /** Parses X's action-bar label: "441 replies, 971 reposts, 8852 likes, 1247 bookmarks, 16029782 views". */ export const parseMetrics = (label: string | null | undefined): Partial<Record<"replies" | "reposts" | "likes" | "views", number>> => { const out: Partial<Record<"replies" | "reposts" | "likes" | "views", number>> = {}; @@ -72,6 +140,21 @@ export const parseMetrics = (label: string | null | undefined): Partial<Record<" return out; }; +/** One trend cell's lines: "2", "·", "Gaming · Trending", "GTA 6", "Trending with #GTAVI". */ +export const parseTrend = (text: string): XTrend | undefined => { + const lines = text.split("\n").map((l) => l.trim()).filter((l) => l && l !== "·"); + let rank: number | undefined; + if (lines[0] && /^\d+$/.test(lines[0])) rank = Number(lines.shift()); + const context = lines.shift(); + const name = lines.shift(); + if (!name) return undefined; + const trend: XTrend = { name, query: name }; + if (rank !== undefined) trend.rank = rank; + if (context) trend.context = context; + if (lines.length) trend.detail = lines.join("; "); + return trend; +}; + type RawPost = Omit<XPost, "replies" | "reposts" | "likes" | "views"> & { metrics: string | null }; type XProbe = { @@ -79,32 +162,41 @@ type XProbe = { state: "signed-out" | "posts" | "empty" | "error" | "loading"; posts: RawPost[]; note: string; + /** The profile's JSON-LD, on a profile page. */ + profile: string | null; }; /** * Reads the page as it stands. Runs in the page; must stay self-contained. * Ads are skipped by having no <time>: every real post has one. */ -const PROBE = `(() => { +export const PROBE = `(() => { const url = location.href; if (/^\\/(login|i\\/flow\\/login|i\\/jf\\/onboarding|account\\/access)/.test(location.pathname)) { - return { url, state: "signed-out", posts: [], note: location.pathname }; + return { url, state: "signed-out", posts: [], note: location.pathname, profile: null }; } const txt = (el) => (el ? el.innerText.trim() : ""); const nameOf = (scope) => { const lines = txt(scope.querySelector('[data-testid="User-Name"]')).split("\\n"); return { author: lines[0] || "", handle: lines.find((l) => l.startsWith("@")) || "" }; }; + const clean = (href) => href.split("?")[0].replace(/\\/(analytics|photo\\/\\d+|video\\/\\d+)$/, ""); const posts = []; for (const a of document.querySelectorAll('article[data-testid="tweet"]')) { - const time = a.querySelector("time"); + // The quoted post: the clickable box holding post text. Not simply the + // first div[role=link], which can be something else in the post. + const quote = [...a.querySelectorAll('div[role="link"]')].find((d) => d.querySelector('[data-testid="tweetText"], [data-testid="User-Name"]')) || null; + const mine = (el) => !(quote && quote.contains(el)); + // The post's own time, a link to it. On a post's own page that sits below + // the text, after the quoted post's unlinked one. + const time = [...a.querySelectorAll("time")].find((t) => mine(t) && t.closest("a")); const link = time && time.closest("a"); if (!time || !link) continue; - const quote = a.querySelector('div[role="link"]'); - const own = [...a.querySelectorAll('[data-testid="tweetText"]')].filter((t) => !(quote && quote.contains(t))); + const own = [...a.querySelectorAll('[data-testid="tweetText"]')].filter(mine); + const byline = [...a.querySelectorAll('[data-testid="User-Name"]')].find(mine); const post = { - url: link.href.split("?")[0].replace(/\\/(analytics|photo\\/\\d+)$/, ""), - ...nameOf(a), + url: clean(link.href), + ...nameOf(byline ? byline.parentElement : a), time: time.getAttribute("datetime") || "", text: own.map(txt).join("\\n"), metrics: null, @@ -114,23 +206,57 @@ const PROBE = `(() => { if (a.querySelector('[data-testid="tweet-text-show-more-link"]')) post.truncated = true; const ctx = txt(a.querySelector('[data-testid="socialContext"]')); if (ctx) post.context = ctx; - const media = [...a.querySelectorAll('[data-testid="tweetPhoto"], video')].filter((m) => !(quote && quote.contains(m))).length; - if (media) post.media = media; + const media = []; + // A video's still is an <img> in the same box: listed with the video, not as a photo. + for (const img of [...a.querySelectorAll('[data-testid="tweetPhoto"] img')].filter((i) => mine(i) && !i.closest('[data-testid="tweetPhoto"]').querySelector("video"))) { + const m = { type: "photo", url: img.src.replace(/name=\\w+/, "name=large") }; + if (img.alt && img.alt !== "Image") m.alt = img.alt; + media.push(m); + } + for (const v of [...a.querySelectorAll("video")].filter(mine)) { + const m = { type: "video" }; + if (v.poster) m.url = v.poster; + const label = v.getAttribute("aria-label"); + if (label && label !== "Embedded video") m.alt = label; + media.push(m); + } + if (media.length) post.media = media; + const card = [...a.querySelectorAll('[data-testid="card.wrapper"] a[href]')].find(mine); + if (card) { + post.card = { url: card.href }; + const title = txt(card).split("\\n").filter(Boolean).pop(); + if (title) post.card.title = title.slice(0, 200); + } if (quote) { const qt = txt(quote.querySelector('[data-testid="tweetText"]')); - if (qt) post.quoted = { ...nameOf(quote), text: qt }; + if (qt) { + post.quoted = { ...nameOf(quote), text: qt }; + const qtime = quote.querySelector("time"); + if (qtime) post.quoted.time = qtime.getAttribute("datetime") || ""; + const qlink = quote.querySelector('a[href*="/status/"]'); + if (qlink) post.quoted.url = clean(qlink.href); + } } posts.push(post); } - if (posts.length) return { url, state: "posts", posts, note: "" }; + const schema = document.querySelector('[data-testid="UserProfileSchema-test"]'); + const profile = schema ? schema.textContent : null; + if (posts.length) return { url, state: "posts", posts, note: "", profile }; const empty = document.querySelector('[data-testid="emptyState"]'); - if (empty) return { url, state: "empty", posts, note: txt(empty).slice(0, 300) }; - const retry = [...document.querySelectorAll('button, [role="button"]')].find((b) => /^Retry$/.test(txt(b))); - if (retry) { - const col = document.querySelector('[data-testid="primaryColumn"]'); - return { url, state: "error", posts, note: txt(col).slice(0, 300) }; + if (empty) return { url, state: "empty", posts, note: txt(empty).slice(0, 300), profile }; + const col = document.querySelector('[data-testid="primaryColumn"]'); + const said = txt(col); + if (/doesn.t exist|was deleted|account is suspended|these posts are protected|you.re unable to view this post/i.test(said)) { + return { url, state: "empty", posts, note: said.slice(0, 300), profile }; } - return { url, state: "loading", posts, note: "" }; + const retry = [...document.querySelectorAll('button, [role="button"]')].find((b) => /^Retry$/.test(txt(b))); + if (retry) return { url, state: "error", posts, note: said.slice(0, 300), profile }; + return { url, state: "loading", posts, note: "", profile }; +})()`; + +const TRENDS = `(() => { + if (/^\\/(login|i\\/flow\\/login|i\\/jf\\/onboarding|account\\/access)/.test(location.pathname)) return null; + return [...document.querySelectorAll('[data-testid="trend"]')].map((t) => t.innerText); })()`; const SCROLL = `window.scrollBy(0, window.innerHeight * 2), true`; @@ -150,18 +276,41 @@ const finish = (raw: RawPost): XPost => { return { ...rest, ...parseMetrics(metrics) }; }; -export const xSearch = async (request: XSearchRequest, outer: AbortSignal): Promise<XSearchOutcome> => { - const sort: XSort = request.sort ?? "latest"; - const searchUrl = buildXSearchUrl(request.query, sort); - const signal = AbortSignal.any([outer, AbortSignal.timeout(X_SEARCH_TIMEOUT_MS)]); +type Saw = (state: string) => void; + +// --- Shared: limits, sign-in, deadline ----------------------------------- + +/** + * Runs one X read: in turn with every other (the limits), within its own + * deadline, signing in and reading again once if X has signed the browser out. + */ +const xRead = <T>(outer: AbortSignal, what: string, read: (page: PageSession, signal: AbortSignal, saw: Saw) => Promise<T>): Promise<T> => + // The deadline starts once the limiter lets the read through, so waiting + // its turn does not eat into it. + limiter.run(outer, async () => { + try { + return await attempt(outer, what, read); + } catch (e) { + if (!(e instanceof SignedOut)) throw e; + // The sign-in has its own deadline, and the read after it a fresh one. + await signIn(e.at, outer); + return await attempt(outer, what, read); + } + }); +const attempt = async <T>( + outer: AbortSignal, + what: string, + read: (page: PageSession, signal: AbortSignal, saw: Saw) => Promise<T>, +): Promise<T> => { + const signal = AbortSignal.any([outer, AbortSignal.timeout(X_SEARCH_TIMEOUT_MS)]); let last = "nothing yet"; try { - return await run(request, sort, searchUrl, signal, (state) => (last = state)); + return await withPage(signal, (page) => read(page, signal, (state) => (last = state))); } catch (e) { if (signal.aborted && !outer.aborted) { throw new BrowserUnavailableError( - `X search did not finish within ${X_SEARCH_TIMEOUT_MS / 1000} s (${searchUrl}; last seen: ${last}). ` + + `X did not finish ${what} within ${X_SEARCH_TIMEOUT_MS / 1000} s (last seen: ${last}). ` + `X may be slow or stuck on a page this tool does not recognise.`, ); } @@ -169,38 +318,110 @@ export const xSearch = async (request: XSearchRequest, outer: AbortSignal): Prom } }; -const run = ( - request: XSearchRequest, - sort: XSort, - searchUrl: string, - signal: AbortSignal, - saw: (state: string) => void, -): Promise<XSearchOutcome> => - withPage(signal, async (page, endpoint) => { - await page.navigate(searchUrl, signal); - - // X renders after the load event: poll until posts, an empty or error - // page, or a sign-in redirect shows up. - let probe: XProbe; - for (;;) { - probe = await page.evaluate<XProbe>(PROBE, signal); - saw(`${probe.state} at ${probe.url}`); - if (probe.state !== "loading") break; - await sleep(POLL_MS, signal); - } +const signedOutError = (at: string, why: string): BrowserUnavailableError => + new BrowserUnavailableError( + `X needs a signed-in X session in castle Chrome, and it is signed out (X sent it to ${at}). ` + + `${why} Tell the user: someone has to sign in to X in that browser again.`, + ); - if (probe.state === "signed-out") { - throw new BrowserUnavailableError( - `X search needs a signed-in X session in the browser (${endpoint}), and it is signed out ` + - `(X sent it to ${probe.url}). Tell the user: someone has to sign in to X in that browser again.`, - ); - } +/** Signs the browser in, within its own deadline, or throws saying why not. */ +const signIn = async (at: string, outer: AbortSignal): Promise<void> => { + const creds = loadCredentials(process.env["CCS_X_LOGIN_FILE"]); + if (!creds) throw signedOutError(at, "No sign-in details are configured (CCS_X_LOGIN_FILE)."); + if (lastLogin && Date.now() - lastLogin.at < LOGIN_COOLDOWN_MS) { + const next = new Date(lastLogin.at + LOGIN_COOLDOWN_MS).toISOString().slice(11, 16); + throw signedOutError( + at, + `Signing in again was tried at ${new Date(lastLogin.at).toISOString().slice(11, 16)} UTC` + + (lastLogin.error ? ` and failed: ${lastLogin.error}` : ", yet it is signed out again.") + + ` The next try is after ${next} UTC.`, + ); + } + lastLogin = { at: Date.now() }; + const signal = AbortSignal.any([outer, AbortSignal.timeout(X_LOGIN_TIMEOUT_MS)]); + try { + await withPage(signal, (page) => xLogin(page, creds, signal, (line) => console.log(line))); + } catch (e) { + const why = signal.aborted && !outer.aborted ? `it took over ${X_LOGIN_TIMEOUT_MS / 1000} s` : (e as Error).message; + lastLogin.error = why; + console.log(`x-login failed: ${why}`); + throw signedOutError(at, `Signing in again failed: ${why}`); + } +}; + +/** + * Opens `url` and waits for X to render: posts, an empty or error page, or a + * sign-in redirect. `ready` can ask for more before posts count, such as a + * profile's header. + */ +const open = async ( + page: PageSession, + url: string, + signal: AbortSignal, + saw: Saw, + ready: (p: XProbe) => boolean = (p) => p.state !== "loading", +): Promise<XProbe> => { + await page.navigate(url, signal); + for (;;) { + const probe = await page.evaluate<XProbe>(PROBE, signal); + saw(`${probe.state} at ${probe.url}`); + if (probe.state === "signed-out") throw new SignedOut(probe.url); if (probe.state === "error") { throw new BrowserUnavailableError( - `X answered with an error page instead of results (${probe.url}): "${probe.note}". This is ` + - `usually X rate-limiting the account; wait a few minutes rather than retrying in a loop.`, + `X answered with an error page (${probe.url}): "${probe.note}". This is usually X rate-limiting ` + + `the account; wait a few minutes rather than retrying in a loop.`, ); } + if (probe.state === "empty" || ready(probe)) return probe; + await sleep(POLL_MS, signal); + } +}; + +/** + * X keeps only the posts near the viewport in the page, so collect by URL + * while scrolling rather than reading once at the end. Stops at `enough`, or + * when scrolling brings nothing new. + */ +const collect = async (page: PageSession, first: XProbe, enough: (seen: Map<string, RawPost>) => boolean, signal: AbortSignal): Promise<RawPost[]> => { + // The newest reading of a post wins: photos and cards render after the + // text, so the first reading of a post can lack them. A Map keeps the order + // posts were first seen in. + const seen = new Map<string, RawPost>(); + const take = (p: XProbe) => p.posts.forEach((post) => seen.set(post.url, post)); + take(first); + // Let the first screen settle: photos, cards and a post's replies arrive + // after the post itself. Settled means two readings in a row agree, and + // there are enough posts or ~3 s have passed. + const shape = (p: XProbe) => p.posts.map((x) => `${x.url}:${x.media?.length ?? 0}:${x.card ? 1 : 0}`).join(" "); + let previous = shape(first); + for (let w = 0; w < 8; w++) { + await sleep(POLL_MS, signal); + const now = await page.evaluate<XProbe>(PROBE, signal); + take(now); + const current = shape(now); + if (current === previous && (enough(seen) || w >= 3)) break; + previous = current; + } + for (let i = 0; i < MAX_SCROLLS && !enough(seen); i++) { + const before = seen.size; + await page.evaluate(SCROLL, signal); + // New posts arrive over the next second or so. + for (let w = 0; w < 4 && seen.size === before; w++) { + await sleep(POLL_MS, signal); + take(await page.evaluate<XProbe>(PROBE, signal)); + } + if (seen.size === before) break; // the end of the page + } + return [...seen.values()]; +}; + +// --- The reads --------------------------------------------------------------- + +export const xSearch = (request: XSearchRequest, outer: AbortSignal): Promise<XSearchOutcome> => { + const sort: XSort = request.sort ?? "latest"; + const searchUrl = buildXSearchUrl(request.query, sort); + return xRead(outer, `the search ${searchUrl}`, async (page, signal, saw) => { + const probe = await open(page, searchUrl, signal, saw); if (probe.state === "empty") { throw new NoResultsError( `X found no posts for ${JSON.stringify(request.query)} (${sort}). X said: "${probe.note}". ` + @@ -208,22 +429,104 @@ const run = ( `or try sort "top" or a wider date range.`, ); } + const posts = await collect(page, probe, (s) => s.size >= request.count, signal); + return { query: request.query, sort, searchUrl, posts: posts.slice(0, request.count).map(finish) }; + }); +}; + +export const xPost = (request: XPostRequest, outer: AbortSignal): Promise<XPostOutcome> => { + const id = postIdOf(request.url); + if (!id) return Promise.reject(new BrowserUnavailableError(`not an X post link: ${request.url}`)); + // /i/status/<id> redirects to the post under its author's handle. + const url = `https://x.com/i/status/${id}`; + const isMain = (p: RawPost) => p.url.endsWith(`/status/${id}`); + return xRead(outer, `reading ${url}`, async (page, signal, saw) => { + const probe = await open(page, url, signal, saw, (p) => p.posts.some(isMain)); + if (probe.state === "empty") { + throw new NoResultsError(`X shows no post at ${url}: "${probe.note}". It may be deleted, or its account protected or suspended.`); + } + const posts = await collect(page, probe, (s) => s.size >= request.replies + 1 + [...s.values()].findIndex(isMain), signal); + const at = posts.findIndex(isMain); + if (at < 0) throw new BrowserUnavailableError(`X's page for ${url} did not show the post itself (${probe.url})`); + return { + post: finish(posts[at]!), + thread: posts.slice(0, at).map(finish), + replies: posts.slice(at + 1, at + 1 + request.replies).map(finish), + }; + }); +}; + +type ProfileSchema = { + dateCreated?: string; + relatedLink?: string[]; + mainEntity?: { + additionalName?: string; + name?: string; + description?: string; + homeLocation?: { name?: string }; + image?: { contentUrl?: string }; + url?: string; + interactionStatistic?: { name?: string; userInteractionCount?: number }[]; + }; +}; - // X keeps only the posts near the viewport in the page, so collect by URL - // while scrolling rather than reading once at the end. - const seen = new Map<string, RawPost>(); - const take = (p: XProbe) => p.posts.forEach((post) => seen.has(post.url) || seen.set(post.url, post)); - take(probe); - for (let i = 0; i < MAX_SCROLLS && seen.size < request.count; i++) { - const before = seen.size; - await page.evaluate(SCROLL, signal); - // New posts arrive over the next second or so. - for (let w = 0; w < 4 && seen.size === before; w++) { +/** The profile from X's JSON-LD block: exact counts, where the page shows "92.2M". */ +export const parseProfile = (json: string, handle: string): XProfile => { + const s = JSON.parse(json) as ProfileSchema; + const e = s.mainEntity ?? {}; + const count = (name: string) => e.interactionStatistic?.find((i) => i.name === name)?.userInteractionCount; + const profile: XProfile = { name: e.name ?? handle, handle: `@${e.additionalName ?? handle}`, url: e.url ?? `https://x.com/${handle}` }; + if (e.description) profile.bio = e.description; + if (e.homeLocation?.name) profile.location = e.homeLocation.name; + const site = s.relatedLink?.find((l) => !/^https?:\/\/t\.co\//.test(l)); + if (site) profile.website = site; + if (s.dateCreated) profile.joined = s.dateCreated; + const followers = count("Follows"); + const following = count("Friends"); + const posts = count("Tweets"); + if (followers !== undefined) profile.followers = followers; + if (following !== undefined) profile.following = following; + if (posts !== undefined) profile.posts = posts; + if (e.image?.contentUrl) profile.avatar = e.image.contentUrl; + return profile; +}; + +export const xUser = (request: XUserRequest, outer: AbortSignal): Promise<XUserOutcome> => { + const handle = handleOf(request.handle); + if (!handle) return Promise.reject(new BrowserUnavailableError(`not an X handle: ${request.handle}`)); + const url = `https://x.com/${handle}`; + return xRead(outer, `reading ${url}`, async (page, signal, saw) => { + // The header renders before the posts; wait for both, or for X to say + // the account has none to show. + const probe = await open(page, url, signal, saw, (p) => p.profile !== null && p.state !== "loading"); + if (!probe.profile) { + throw new NoResultsError(`X shows no account @${handle}: "${probe.note}". Check the handle.`); + } + const profile = parseProfile(probe.profile, handle); + const posts = probe.state === "posts" && request.count > 0 ? await collect(page, probe, (s) => s.size >= request.count, signal) : []; + return { profile, posts: posts.slice(0, request.count).map(finish) }; + }); +}; + +export const xTrends = (tab: XTrendTab, outer: AbortSignal): Promise<XTrendsOutcome> => { + const url = `https://x.com/explore/tabs/${tab}`; + return xRead(outer, `reading ${url}`, async (page, signal, saw) => { + await page.navigate(url, signal); + for (;;) { + const cells = await page.evaluate<string[] | null>(TRENDS, signal); + if (cells === null) throw new SignedOut(url); + saw(`${cells.length} trends at ${url}`); + if (cells.length) { + // The list renders in one go; a moment more lets the last cells fill. await sleep(POLL_MS, signal); - take(await page.evaluate<XProbe>(PROBE, signal)); + const all = (await page.evaluate<string[] | null>(TRENDS, signal)) ?? cells; + return { tab, trends: all.map(parseTrend).filter((t): t is XTrend => t !== undefined) }; + } + const probe = await page.evaluate<XProbe>(PROBE, signal); + if (probe.state === "error") { + throw new BrowserUnavailableError(`X answered with an error page (${url}): "${probe.note}".`); } - if (seen.size === before) break; // the end of the results + await sleep(POLL_MS, signal); } - - return { query: request.query, sort, searchUrl, posts: [...seen.values()].slice(0, request.count).map(finish) }; }); +}; diff --git a/castle-search/src/youtube.ts b/castle-search/src/youtube.ts new file mode 100644 index 0000000..18b1d49 --- /dev/null +++ b/castle-search/src/youtube.ts @@ -0,0 +1,303 @@ +/** + * A YouTube video's transcript, filed as plain text in a folder (castle's + * synced Nextcloud "transcript" folder) and handed back to the caller. + * + * The recipe is the one Hermes worked out on bambos (its skill + * youtube-transcript-to-nextcloud), so files from either read the same: + * + * 1. the bare video link: share-link junk (?si=, ?t=, &list=) dropped; + * 2. title, channel, upload date and length from yt-dlp; + * 3. subtitles only, never the video: human-written ones when the video has + * them, YouTube's automatic ones otherwise, as VTT; + * 4. YouTube's rolling captions repeat the end of the previous cue, so each + * cue keeps only the words the text so far does not already end with; + * 5. filed as <upload date>-<Title-With-Dashes>.txt, paragraphs of 90 words. + * + * yt-dlp breaks whenever YouTube changes, so it updates itself (`yt-dlp -U`) + * at most once a day, and once more when a fetch fails. YouTube rate-limits + * its caption downloads (HTTP 429); then the transcript comes from + * youtube-transcript-api instead, which uses another endpoint. + * + * Configured by environment: CCS_YTDLP (the yt-dlp binary, one it can update), + * CCS_YT_PYTHON (a python with youtube_transcript_api, optional), + * CCS_TRANSCRIPT_DIR (where files go). One transcript at a time. + */ + +import { execFile } from "node:child_process"; +import { existsSync, mkdtempSync, readdirSync, readFileSync, rmSync, statSync, writeFileSync } from "node:fs"; +import { tmpdir } from "node:os"; +import { join } from "node:path"; + +import { BrowserUnavailableError, NoResultsError } from "./errors.ts"; + +export type TranscriptRequest = { url: string; language?: string | undefined; maxChars?: number | undefined }; + +export type TranscriptOutcome = { + file: string; + /** Already in the folder from an earlier run, so not fetched again. */ + existing: boolean; + url: string; + title: string; + channel: string; + /** YYYY-MM-DD */ + uploaded: string; + /** Seconds. */ + duration?: number; + /** "manual", "automatic", or "youtube-transcript-api". */ + source: string; + words: number; + text: string; + /** The text was longer than maxChars; the file has all of it. */ + truncated?: true; +}; + +const FETCH_TIMEOUT_MS = 120_000; +const UPDATE_EVERY_MS = 24 * 3_600_000; +export const DEFAULT_MAX_CHARS = 60_000; + +// --- Pure parts, tested -------------------------------------------------- + +/** The video id in any YouTube link: watch?v=, youtu.be/, shorts/, live/, embed/. */ +export const videoIdOf = (url: string): string | undefined => { + const s = url.trim(); + if (/^[A-Za-z0-9_-]{11}$/.test(s)) return s; + const m = + /(?:youtube\.com\/(?:watch\?(?:.*&)?v=|shorts\/|live\/|embed\/|v\/)|youtu\.be\/|youtube-nocookie\.com\/embed\/)([A-Za-z0-9_-]{11})/i.exec(s); + return m?.[1]; +}; + +/** "Dr. Becky: Parenting 101 | Call Her Daddy" -> "Dr-Becky-Parenting-101-Call-Her-Daddy". */ +export const titleSlug = (title: string): string => { + const words = title + .normalize("NFKD") + .replace(/[\u0300-\u036f]/g, "") + .replace(/['’]/g, "") + .match(/[A-Za-z0-9]+/g); + return (words ?? []).join("-").slice(0, 100).replace(/-+$/, "") || "transcript"; +}; + +/** "20051024" -> "2005-10-24". */ +export const isoDate = (yyyymmdd: string): string => + /^\d{8}$/.test(yyyymmdd) ? `${yyyymmdd.slice(0, 4)}-${yyyymmdd.slice(4, 6)}-${yyyymmdd.slice(6, 8)}` : new Date().toISOString().slice(0, 10); + +const SKIP = ["WEBVTT", "NOTE", "Kind:", "Language:", "STYLE", "REGION", "X-TIMESTAMP"]; + +/** Each cue's text, timings and markup removed. */ +export const cueTexts = (vtt: string): string[] => { + const out: string[] = []; + for (const block of vtt.replace(/\r/g, "").split(/\n\s*\n/)) { + const lines = block + .split("\n") + .map((l) => l.trim()) + .filter((l) => l && !l.includes("-->") && !SKIP.some((p) => l.startsWith(p)) && !/^\d+$/.test(l)); + if (!lines.length) continue; + const text = lines + .join(" ") + .replace(/<[^>]*>/g, "") + .replace(/\[[^\]]*\]/g, " ") // [Music], [Applause] + .replace(/>/g, " ") + .replace(/</g, "<") + .replace(/&/g, "&") + .replace(/'/g, "'") + .replace(/"/g, '"') + .replace(/ /g, " ") + .replace(/\s+/g, " ") + .trim(); + if (text) out.push(text); + } + return out; +}; + +/** + * Rolling captions: each cue starts with the end of what came before. Keep + * only the words past the longest overlap with the text so far. + */ +export const dedupeRolling = (cues: string[], window = 80): string[] => { + const emitted: string[] = []; + for (const cue of cues) { + const words = cue.split(" ").filter(Boolean); + const tail = emitted.slice(-window); + let overlap = 0; + for (let k = Math.min(tail.length, words.length); k > 0; k--) { + if (tail.slice(-k).join(" ") === words.slice(0, k).join(" ")) { + overlap = k; + break; + } + } + emitted.push(...words.slice(overlap)); + } + return emitted; +}; + +export const paragraphs = (words: string[], per = 90): string => { + const out: string[] = []; + for (let i = 0; i < words.length; i += per) out.push(words.slice(i, i + per).join(" ")); + return out.join("\n\n"); +}; + +// --- Running yt-dlp ------------------------------------------------------ + +const run = (cmd: string, args: string[], signal: AbortSignal, timeout = FETCH_TIMEOUT_MS): Promise<{ code: number; stdout: string; stderr: string }> => + new Promise((resolve) => { + execFile(cmd, args, { signal, timeout, maxBuffer: 64 * 1024 * 1024 }, (error, stdout, stderr) => { + const code = error ? (typeof (error as { code?: unknown }).code === "number" ? (error as { code: number }).code : 1) : 0; + resolve({ code, stdout: String(stdout), stderr: String(stderr) }); + }); + }); + +const ytdlp = (): string => process.env["CCS_YTDLP"] ?? "yt-dlp"; + +let lastUpdate = 0; +const update = async (signal: AbortSignal, force = false): Promise<void> => { + if (!force && Date.now() - lastUpdate < UPDATE_EVERY_MS) return; + lastUpdate = Date.now(); + const r = await run(ytdlp(), ["-U"], signal, 60_000); + console.log(`yt-dlp -U: ${(r.stdout + r.stderr).trim().split("\n").pop() ?? ""}`); +}; + +type Meta = { title: string; channel: string; uploaded: string; duration?: number; id: string }; + +const metadata = async (url: string, signal: AbortSignal): Promise<Meta | { error: string }> => { + const r = await run(ytdlp(), ["--no-warnings", "--skip-download", "--no-playlist", "--print", "%(title)s\t%(uploader)s\t%(upload_date)s\t%(duration)s\t%(id)s", url], signal); + const line = r.stdout.trim().split("\n").pop() ?? ""; + const [title, channel, date, duration, id] = line.split("\t"); + if (r.code !== 0 || !title || !id) return { error: (r.stderr.trim().split("\n").pop() ?? `yt-dlp exited ${r.code}`).slice(0, 300) }; + const meta: Meta = { title, channel: channel && channel !== "NA" ? channel : "", uploaded: isoDate(date ?? ""), id }; + const d = Number(duration); + if (Number.isFinite(d) && d > 0) meta.duration = d; + return meta; +}; + +/** The subtitle file yt-dlp wrote: human-written beats automatic (".en-orig" is automatic). */ +const pickSubs = (dir: string, lang: string): { path: string; source: string } | undefined => { + const files = readdirSync(dir).filter((f) => f.endsWith(".vtt") && statSync(join(dir, f)).size > 0); + const auto = files.find((f) => f.includes(`.${lang}-orig.`)); + const manual = files.find((f) => f !== auto); + if (manual) { + const path = join(dir, manual); + const text = readFileSync(path, "utf8"); + // Beside an -orig track, the plain one is human-written unless it is the + // same file. Alone, its word-timing tags say whether it is automatic. + const automatic = auto !== undefined ? text === readFileSync(join(dir, auto), "utf8") : /<\d\d:\d\d:\d\d\.\d{3}><c>/.test(text); + return { path, source: automatic ? "automatic" : "manual" }; + } + return auto ? { path: join(dir, auto), source: "automatic" } : undefined; +}; + +const subtitles = async (id: string, lang: string, dir: string, signal: AbortSignal): Promise<{ words: string[]; source: string } | { error: string; rateLimited: boolean }> => { + // Only the language itself and its regional variants: a pattern like + // "en-.*" also takes YouTube's machine translations ("en-de"), which are + // what its rate limit falls on first. + const langs = [lang, `${lang}-orig`, ...(lang === "en" ? ["en-GB", "en-US"] : [])].join(","); + const r = await run( + ytdlp(), + ["--no-warnings", "--skip-download", "--no-playlist", "--write-subs", "--write-auto-subs", "--sub-langs", langs, "--sub-format", "vtt", "-o", join(dir, "%(id)s"), `https://www.youtube.com/watch?v=${id}`], + signal, + ); + // yt-dlp fails the whole run when one of several tracks fails; a track it + // did write is still good. + const picked = pickSubs(dir, lang); + if (picked) return { words: dedupeRolling(cueTexts(readFileSync(picked.path, "utf8"))), source: picked.source }; + const err = r.stderr.trim().split("\n").filter((l) => l.startsWith("ERROR")).pop() ?? ""; + return { error: err || `no ${lang} subtitles`, rateLimited: /429|Too Many Requests/.test(err) }; +}; + +const TRANSCRIPT_API = ` +import json, sys +from youtube_transcript_api import YouTubeTranscriptApi +vid, lang = sys.argv[1], sys.argv[2] +langs = [lang] + ([f"{lang}-US", f"{lang}-GB"] if lang == "en" else []) +f = YouTubeTranscriptApi().fetch(vid, languages=langs) +print(json.dumps([s.text for s in f])) +`; + +const transcriptApi = async (id: string, lang: string, signal: AbortSignal): Promise<string[] | { error: string }> => { + const py = process.env["CCS_YT_PYTHON"]; + if (!py) return { error: "no youtube-transcript-api configured (CCS_YT_PYTHON)" }; + const r = await run(py, ["-I", "-c", TRANSCRIPT_API, id, lang], signal); + if (r.code !== 0) return { error: (r.stderr.trim().split("\n").pop() ?? "").slice(0, 300) }; + // Discrete snippets, not rolling cues: no dedupe needed. + return (JSON.parse(r.stdout) as string[]).join(" ").replace(/\[[^\]]*\]/g, " ").split(/\s+/).filter(Boolean); +}; + +// --- The tool ------------------------------------------------------------ + +let tail: Promise<unknown> = Promise.resolve(); + +/** One at a time: yt-dlp updating itself under a running fetch breaks both. */ +export const youtubeTranscript = (request: TranscriptRequest, signal: AbortSignal): Promise<TranscriptOutcome> => { + const next = tail.then(() => fetchTranscript(request, signal)); + tail = next.catch(() => {}); + return next; +}; + +const fetchTranscript = async (request: TranscriptRequest, signal: AbortSignal): Promise<TranscriptOutcome> => { + const id = videoIdOf(request.url); + if (!id) throw new NoResultsError(`not a YouTube video link: ${request.url}`); + const lang = (request.language ?? "en").trim() || "en"; + const maxChars = request.maxChars ?? DEFAULT_MAX_CHARS; + const url = `https://www.youtube.com/watch?v=${id}`; + const folder = process.env["CCS_TRANSCRIPT_DIR"]; + if (!folder || !existsSync(folder)) { + throw new BrowserUnavailableError(`the transcript folder (CCS_TRANSCRIPT_DIR=${folder ?? "unset"}) does not exist`); + } + + await update(signal); + let meta = await metadata(url, signal); + if ("error" in meta) { + // An extractor broken by a YouTube change: a newer yt-dlp is nearly always the fix. + await update(signal, true); + meta = await metadata(url, signal); + if ("error" in meta) throw new BrowserUnavailableError(`yt-dlp could not read ${url}: ${meta.error}`); + } + + const name = `${meta.uploaded}-${titleSlug(meta.title)}.txt`; + const file = join(folder, name); + const shape = (text: string, source: string, existing: boolean): TranscriptOutcome => { + const out: TranscriptOutcome = { + file: `transcript/${name}`, + existing, + url, + title: meta.title, + channel: meta.channel, + uploaded: meta.uploaded, + source, + words: text.split(/\s+/).filter(Boolean).length, + text: text.length > maxChars ? text.slice(0, maxChars) : text, + }; + if ("duration" in meta && meta.duration !== undefined) out.duration = meta.duration; + if (text.length > maxChars) out.truncated = true; + return out; + }; + if (existsSync(file)) return shape(readFileSync(file, "utf8").trim(), "earlier run", true); + + const scratch = mkdtempSync(join(tmpdir(), "yt-")); + try { + const subs = await subtitles(id, lang, scratch, signal); + let words: string[]; + let source: string; + if ("words" in subs) { + ({ words, source } = subs); + } else { + // A 429 is YouTube limiting the caption endpoint: retrying yt-dlp gets + // another. The transcript API is a different endpoint. + const alt = await transcriptApi(id, lang, signal); + if (!Array.isArray(alt)) { + throw new NoResultsError( + `no ${lang} transcript for ${url}: yt-dlp said "${subs.error}"` + + `${subs.rateLimited ? " (YouTube rate-limiting captions)" : ""}; youtube-transcript-api said "${alt.error}". ` + + `The video may have no captions in that language: try another language code.`, + ); + } + words = alt; + source = "youtube-transcript-api"; + } + if (words.length === 0) throw new NoResultsError(`the ${lang} captions of ${url} are empty`); + const text = paragraphs(words); + writeFileSync(file, text + "\n", "utf8"); + console.log(`youtube ${id} -> ${name} (${words.length} words, ${source})`); + return shape(text, source, false); + } finally { + rmSync(scratch, { recursive: true, force: true }); + } +}; diff --git a/castle-search/test/http.test.ts b/castle-search/test/http.test.ts index b2c90ab..c076863 100644 --- a/castle-search/test/http.test.ts +++ b/castle-search/test/http.test.ts @@ -116,3 +116,73 @@ test("X's action-bar label parses, singular and plural", () => { test("the OpenAPI description names the operation Open WebUI will offer", () => { assert.equal(OPENAPI.paths["/x-search"].post.operationId, "search_x"); }); + +import { handleXPost, handleXTrends, handleXUser } from "../src/http.ts"; +import { handleOf, parseProfile, parseTrend, postIdOf } from "../src/x-search.ts"; + +test("post links of every shape give the post id", () => { + assert.equal(postIdOf("https://x.com/NASA/status/2122740072386609201"), "2122740072386609201"); + assert.equal(postIdOf("https://twitter.com/nasa/status/2122740072386609201?s=20"), "2122740072386609201"); + assert.equal(postIdOf("https://x.com/i/web/status/2122740072386609201"), "2122740072386609201"); + assert.equal(postIdOf("2122740072386609201"), "2122740072386609201"); + assert.equal(postIdOf("https://x.com/NASA"), undefined); +}); + +test("handles are read from @name, name or a profile link", () => { + assert.equal(handleOf("@NASA"), "NASA"); + assert.equal(handleOf("https://x.com/nasa/"), "nasa"); + assert.equal(handleOf("not a handle"), undefined); + assert.equal(handleOf("waytoolonghandle_123"), undefined); +}); + +test("a trend cell's lines become a trend", () => { + assert.deepEqual(parseTrend("2\n·\nVideo game genres · Trending\nGTA 6\nTrending with #GTAVI, Metacritic"), { + rank: 2, context: "Video game genres · Trending", name: "GTA 6", query: "GTA 6", detail: "Trending with #GTAVI, Metacritic", + }); + assert.deepEqual(parseTrend("Trending in United Kingdom\nGlastonbury"), { context: "Trending in United Kingdom", name: "Glastonbury", query: "Glastonbury" }); + assert.equal(parseTrend("1"), undefined); +}); + +test("a profile's JSON-LD gives exact counts", () => { + const p = parseProfile( + JSON.stringify({ + dateCreated: "2007-12-19T20:20:32.000Z", + relatedLink: ["https://t.co/9NkQJKAnuU", "http://www.nasa.gov/"], + mainEntity: { + additionalName: "NASA", name: "NASA", description: "Making the seemingly impossible, possible.", + homeLocation: { name: "Pale Blue Dot" }, url: "https://x.com/NASA", + interactionStatistic: [ + { name: "Follows", userInteractionCount: 92225170 }, + { name: "Friends", userInteractionCount: 115 }, + { name: "Tweets", userInteractionCount: 74553 }, + ], + }, + }), + "nasa", + ); + assert.equal(p.handle, "@NASA"); + assert.equal(p.followers, 92225170); + assert.equal(p.following, 115); + assert.equal(p.posts, 74553); + assert.equal(p.website, "http://www.nasa.gov/"); + assert.equal(p.location, "Pale Blue Dot"); +}); + +test("the X reads check what they are given before touching the browser", async () => { + const never = async (): Promise<never> => { + throw new Error("must not run"); + }; + assert.equal((await handleXPost(JSON.stringify({ url: "https://x.com/NASA" }), never, signal)).status, 400); + assert.equal((await handleXUser(JSON.stringify({ handle: "two words" }), never, signal)).status, 400); + assert.equal((await handleXTrends(JSON.stringify({ tab: "music" }), never, signal)).status, 400); + assert.equal((await handleXPost("[1]", never, signal)).status, 400); + const seen: unknown[] = []; + await handleXPost(JSON.stringify({ url: "https://x.com/a/status/12345678", replies: 99 }), async (r) => (seen.push(r), { post: {}, thread: [], replies: [] }) as never, signal); + await handleXTrends("", async (t) => (seen.push(t), { tab: t, trends: [] }), signal); + assert.deepEqual(seen, [{ url: "https://x.com/a/status/12345678", replies: 20 }, "trending"]); +}); + +test("every X read is in the OpenAPI description", () => { + assert.deepEqual(Object.keys(OPENAPI.paths), ["/x-search", "/x-post", "/x-user", "/x-trends"]); + assert.equal(OPENAPI.info.title, "X.com"); +}); diff --git a/castle-search/test/x-limits.test.ts b/castle-search/test/x-limits.test.ts new file mode 100644 index 0000000..6be5f7d --- /dev/null +++ b/castle-search/test/x-limits.test.ts @@ -0,0 +1,98 @@ +import assert from "node:assert/strict"; +import { test } from "node:test"; + +import { type Clock, XLimiter, XRateLimitError } from "../src/x-limits.ts"; + +/** A clock that only moves when told to, recording the waits asked of it. */ +const fakeClock = () => { + let now = 1_000_000; + const waits: number[] = []; + let draw = 0.5; + const clock: Clock = { + random: () => draw, + now: () => now, + sleep: async (ms) => { + waits.push(ms); + now += ms; + }, + }; + return { clock, waits, advance: (ms: number) => (now += ms), setDraw: (r: number) => (draw = r) }; +}; + +const signal = new AbortController().signal; +const ok = async () => "ok"; + +test("a search sooner than the gap waits out the rest of it", async () => { + const { clock, waits, advance } = fakeClock(); + const l = new XLimiter(clock, { gap: 10_000, jitter: 0, hour: 30, day: 150 }); + await l.run(signal, ok); + advance(3_000); + await l.run(signal, ok); + assert.deepEqual(waits, [7_000]); +}); + +test("over the hourly cap a search fails at once, naming when it can run", async () => { + const { clock, advance } = fakeClock(); + const l = new XLimiter(clock, { gap: 0, jitter: 0, hour: 2, day: 150 }); + await l.run(signal, ok); + await l.run(signal, ok); + await assert.rejects(l.run(signal, ok), (e) => e instanceof XRateLimitError && /last hour/.test(e.message)); + advance(3_600_001); + assert.equal(await l.run(signal, ok), "ok"); +}); + +test("the daily cap holds across hours", async () => { + const { clock, advance } = fakeClock(); + const l = new XLimiter(clock, { gap: 0, jitter: 0, hour: 100, day: 3 }); + for (let i = 0; i < 3; i++) { + await l.run(signal, ok); + advance(3_600_001); + } + await assert.rejects(l.run(signal, ok), (e) => e instanceof XRateLimitError && /24 hours/.test(e.message)); +}); + +test("one at a time: the second starts only after the first finishes", async () => { + const { clock } = fakeClock(); + const l = new XLimiter(clock, { gap: 0, jitter: 0, hour: 30, day: 150 }); + const order: string[] = []; + let release!: () => void; + const first = l.run(signal, async () => { + order.push("first start"); + await new Promise<void>((r) => (release = r)); + order.push("first end"); + }); + const second = l.run(signal, async () => { + order.push("second start"); + }); + await new Promise((r) => setTimeout(r, 10)); + release(); + await Promise.all([first, second]); + assert.deepEqual(order, ["first start", "first end", "second start"]); +}); + +test("a failed search does not block the queue", async () => { + const { clock } = fakeClock(); + const l = new XLimiter(clock, { gap: 0, jitter: 0, hour: 30, day: 150 }); + await assert.rejects(l.run(signal, async () => { + throw new Error("boom"); + })); + assert.equal(await l.run(signal, ok), "ok"); +}); + +test("the gap is drawn afresh each time, within gap ± jitter", async () => { + const { clock, waits, setDraw } = fakeClock(); + const l = new XLimiter(clock, { gap: 30_000, jitter: 25_000, hour: 30, day: 150 }); + await l.run(signal, ok); + setDraw(0); // the shortest: 30 - 25 s + await l.run(signal, ok); + setDraw(0.999999); // close to the longest: 30 + 25 s + await l.run(signal, ok); + assert.equal(waits[0], 5_000); + assert.ok(waits[1]! > 54_999 && waits[1]! <= 55_000); +}); + +test("the defaults are 30 s ± 25 s", async () => { + const { GAP_MS, JITTER_MS } = await import("../src/x-limits.ts"); + assert.equal(GAP_MS, 30_000); + assert.equal(JITTER_MS, 25_000); +}); diff --git a/castle-search/test/x-login.test.ts b/castle-search/test/x-login.test.ts new file mode 100644 index 0000000..11468d5 --- /dev/null +++ b/castle-search/test/x-login.test.ts @@ -0,0 +1,91 @@ +import assert from "node:assert/strict"; +import { test } from "node:test"; + +import { decodeWords, parseFetch, parseInternalDate, xCodeFromSubject } from "../src/imap.ts"; +import { classify, type LoginProbe } from "../src/x-login.ts"; +import { base32Decode, totp, totpRemaining } from "../src/totp.ts"; + +// RFC 6238 appendix B, SHA-1: the secret "12345678901234567890". +const RFC_SECRET = "GEZDGNBVGY3TQOJQGEZDGNBVGY3TQOJQ"; + +test("totp matches RFC 6238's SHA-1 vectors", () => { + assert.equal(totp(RFC_SECRET, 59_000, 8), "94287082"); + assert.equal(totp(RFC_SECRET, 1_111_111_109_000, 8), "07081804"); + assert.equal(totp(RFC_SECRET, 1_234_567_890_000, 8), "89005924"); + assert.equal(totp(RFC_SECRET, 59_000), "287082"); +}); + +test("base32 ignores spaces and case, and refuses other characters", () => { + assert.deepEqual(base32Decode("gezd gnbv"), base32Decode("GEZDGNBV")); + assert.throws(() => base32Decode("GEZD1")); +}); + +test("totpRemaining counts down within the step", () => { + assert.equal(totpRemaining(0), 30); + assert.equal(totpRemaining(29_000), 1); +}); + +test("X's code is read from the subjects X uses, and nothing else", () => { + assert.equal(xCodeFromSubject("Your X confirmation code is abc12def"), "abc12def"); + assert.equal(xCodeFromSubject("Your X verification code is: 123456"), "123456"); + assert.equal(xCodeFromSubject("123456 is your X verification code"), "123456"); + assert.equal(xCodeFromSubject("New login to X from Chrome on Mac"), undefined); + assert.equal(xCodeFromSubject("Security alert"), undefined); +}); + +test("encoded subjects are decoded", () => { + assert.equal(decodeWords("=?UTF-8?B?WW91ciBYIGNvbmZpcm1hdGlvbiBjb2RlIGlzIGFiYzEyZGVm?="), "Your X confirmation code is abc12def"); + assert.equal(decodeWords("=?UTF-8?Q?Caf=C3=A9_menu?="), "Café menu"); +}); + +test("a FETCH reply splits into one header set per message", () => { + const raw = + '* 1 FETCH (INTERNALDATE "18-Nov-2026 10:16:23 +0000" BODY[HEADER.FIELDS (FROM SUBJECT)] {60}\r\n' + + "From: X <verify@x.com>\r\nSubject: Your X confirmation code is\r\n abc12def\r\n\r\n)\r\n" + + '* 2 FETCH (INTERNALDATE "18-Nov-2026 10:20:00 +0000" BODY[HEADER.FIELDS (FROM SUBJECT)] {40}\r\n' + + "From: X <info@x.com>\r\nSubject: New login\r\n\r\n)\r\nc4 OK Success\r\n"; + const mails = parseFetch(raw); + assert.equal(mails.length, 2); + assert.equal(mails[0]!.subject, "Your X confirmation code is abc12def"); + assert.equal(mails[0]!.received, parseInternalDate("18-Nov-2026 10:16:23 +0000")); + assert.equal(mails[0]!.received, Date.UTC(2026, 10, 18, 10, 16, 23)); + assert.equal(mails[1]!.from, "X <info@x.com>"); +}); + +const probe = (over: Partial<LoginProbe>): LoginProbe => ({ + url: "https://x.com/i/jf/onboarding/web?mode=login", + path: "/i/jf/onboarding/web", + hash: "", + text: "", + inputs: [], + signedIn: false, + ...over, +}); + +// The steps as castle Chrome saw them on 2026-11-18. +test("each step of X's sign-in is recognised", () => { + assert.equal( + classify(probe({ text: "See what's happening\n\nEmail or username\n\nContinue", inputs: [{ name: "username_or_email", type: "text", autocomplete: "username webauthn" }] })).kind, + "username", + ); + assert.equal( + classify(probe({ hash: "#/s/login_enter_password/r-1", text: "Login\n\nUsername\nPassword", inputs: [{ name: "username", type: "text", autocomplete: "" }, { name: "password", type: "password", autocomplete: "" }] })).kind, + "password", + ); + assert.equal( + classify(probe({ hash: "#/s/two_factor_code/r-2", text: "Enter your two factor code\n\nUse your authenticator app to generate the code.", inputs: [{ name: "", type: "text", autocomplete: "one-time-code" }] })).kind, + "totp", + ); + assert.equal( + classify(probe({ text: "Check your email\n\nYou'll receive a code to verify here so you can sign in.", inputs: [{ name: "text", type: "text", autocomplete: "" }] })).kind, + "email-code", + ); + assert.equal( + classify(probe({ text: "Enter your phone number or username\n\nThere was unusual login activity on your account.", inputs: [{ name: "text", type: "text", autocomplete: "" }] })).kind, + "username", + ); + assert.equal(classify(probe({ url: "https://x.com/home", path: "/home", signedIn: true })).kind, "done"); + assert.equal(classify(probe({})).kind, "loading"); + assert.equal(classify(probe({ text: "Authenticate your account\n\nSolve this puzzle" })).kind, "unknown"); + assert.equal(classify(probe({ text: "Wrong password!", inputs: [{ name: "password", type: "password", autocomplete: "" }] })).kind, "wrong-password"); +}); diff --git a/castle-search/test/youtube.test.ts b/castle-search/test/youtube.test.ts new file mode 100644 index 0000000..c5530fb --- /dev/null +++ b/castle-search/test/youtube.test.ts @@ -0,0 +1,72 @@ +import assert from "node:assert/strict"; +import { test } from "node:test"; + +import { handleTranscript } from "../src/http.ts"; +import { cueTexts, dedupeRolling, isoDate, paragraphs, titleSlug, videoIdOf } from "../src/youtube.ts"; + +test("every shape of YouTube link gives the video id, share junk ignored", () => { + for (const url of [ + "https://www.youtube.com/watch?v=jNQXAC9IVRw", + "https://youtube.com/watch?feature=share&v=jNQXAC9IVRw&t=30s", + "https://youtu.be/jNQXAC9IVRw?si=T-d5TUXNFjmR1Uad", + "https://www.youtube.com/shorts/jNQXAC9IVRw", + "https://www.youtube.com/live/jNQXAC9IVRw?feature=shared", + "https://m.youtube.com/watch?v=jNQXAC9IVRw&list=PL123", + "jNQXAC9IVRw", + ]) { + assert.equal(videoIdOf(url), "jNQXAC9IVRw", url); + } + assert.equal(videoIdOf("https://vimeo.com/123"), undefined); +}); + +test("titles become the archive's dashed names", () => { + assert.equal(titleSlug("How to stay calm when you know you'll be stressed | Daniel Levitin | TED"), "How-to-stay-calm-when-you-know-youll-be-stressed-Daniel-Levitin-TED"); + assert.equal(titleSlug("Anthropic Files for IPO: Nasdaq!"), "Anthropic-Files-for-IPO-Nasdaq"); + assert.equal(titleSlug("Café Müller"), "Cafe-Muller"); + assert.equal(titleSlug("!!!"), "transcript"); + assert.equal(isoDate("20051024"), "2005-10-24"); +}); + +// The shape of YouTube's automatic captions: each cue repeats the last one's words. +const ROLLING = `WEBVTT +Kind: captions +Language: en + +00:00:00.000 --> 00:00:02.000 align:start position:0% +the filing is out<00:00:01.000><c> ENT</c><00:00:01.500><c> is</c> + +00:00:02.000 --> 00:00:02.010 align:start position:0% +the filing is out ENT is + +00:00:02.010 --> 00:00:04.000 align:start position:0% +the filing is out ENT is +going to be [Music] big & bold + +00:00:04.000 --> 00:00:06.000 +going to be big & bold for everyone +`; + +test("rolling captions come out once, markup and noise gone", () => { + const cues = cueTexts(ROLLING); + assert.equal(cues[0], "the filing is out ENT is"); + assert.equal(dedupeRolling(cues).join(" "), "the filing is out ENT is going to be big & bold for everyone"); +}); + +test("words are set in paragraphs of 90", () => { + const words = Array.from({ length: 200 }, (_, i) => `w${i}`); + const text = paragraphs(words); + assert.equal(text.split("\n\n").length, 3); + assert.equal(text.split("\n\n")[0]!.split(" ").length, 90); +}); + +test("the transcript route checks the link and language first", async () => { + const signal = new AbortController().signal; + const never = async (): Promise<never> => { + throw new Error("must not run"); + }; + assert.equal((await handleTranscript(JSON.stringify({ url: "https://vimeo.com/1" }), never, signal)).status, 400); + assert.equal((await handleTranscript(JSON.stringify({ url: "https://youtu.be/jNQXAC9IVRw", language: "english!" }), never, signal)).status, 400); + const seen: unknown[] = []; + await handleTranscript(JSON.stringify({ url: "https://youtu.be/jNQXAC9IVRw", language: "pt", max_chars: 10 }), async (r) => (seen.push(r), {}) as never, signal); + assert.deepEqual(seen, [{ url: "https://youtu.be/jNQXAC9IVRw", language: "pt", maxChars: 1000 }]); +}); |
