diff options
| author | Igor Soarez <igor@soarez.org> | 2026-11-18 04:55:45 +0000 |
|---|---|---|
| committer | Igor Soarez <igor@soarez.org> | 2026-11-18 04:55:45 +0000 |
| commit | b03415aafeea985561f3980f8b484173530a05f8 (patch) | |
| tree | 3748f62ae52a5c4d7906ec3268bb8b2863a5114a /castle-search | |
| parent | 4c1f9b4ea943d03e9610f13344c372e630db5b9d (diff) | |
castle-search: search X over HTTP, through the browser's signed-in session
POST /x-search reads x.com/search in a background tab: posts with their link,
author, time, text, counts and quoted post, scrolling for up to 20. GET
/openapi.json describes it for Open WebUI's tool servers. Signed out, an X
error page or a stuck page are errors, never an empty list.
Diffstat (limited to 'castle-search')
| -rw-r--r-- | castle-search/README.md | 13 | ||||
| -rw-r--r-- | castle-search/src/http.ts | 107 | ||||
| -rw-r--r-- | castle-search/src/search.ts | 24 | ||||
| -rw-r--r-- | castle-search/src/x-search.ts | 229 | ||||
| -rw-r--r-- | castle-search/test/http.test.ts | 51 |
5 files changed, 419 insertions, 5 deletions
diff --git a/castle-search/README.md b/castle-search/README.md index 7e4fd5c..0e45aaf 100644 --- a/castle-search/README.md +++ b/castle-search/README.md @@ -47,6 +47,19 @@ 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 + +`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. + +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. + ## Engines Four, with independent indexes and independent rate limits. That last part is diff --git a/castle-search/src/http.ts b/castle-search/src/http.ts index 31b1eaa..17c8c70 100644 --- a/castle-search/src/http.ts +++ b/castle-search/src/http.ts @@ -19,12 +19,18 @@ * It has no authentication, so bind it to loopback, as the default does, or to * 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. */ 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"; export const DEFAULT_LISTEN = "127.0.0.1:11158"; const MAX_BODY_BYTES = 64 * 1024; @@ -68,6 +74,85 @@ export const handleSearch = async (raw: string, run: Searcher, signal: AbortSign } }; +export type XSearcher = (request: XSearchRequest, signal: AbortSignal) => Promise<XSearchOutcome>; + +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; + try { + parsed = JSON.parse(raw); + } catch { + return { status: 400, body: { error: "request body is not JSON" } }; + } + const { query, count, sort } = (parsed ?? {}) as { query?: unknown; count?: unknown; sort?: unknown }; + 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) } }; + } +}; + +/** + * 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. + */ +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.", + }, + paths: { + "/x-search": { + post: { + operationId: "search_x", + summary: "Search posts on X (Twitter)", + description: + "Search X for posts, newest first by default. Use it for anything on X: what someone posted, " + + "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" }, + }, + }, + }, + }, + }, + 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" } } } }, + }, + }, + }, + }, +} as const; + const readBody = (req: IncomingMessage): Promise<string> => new Promise((resolve, reject) => { let size = 0; @@ -110,8 +195,13 @@ const main = (): void => { send(res, { status: 200, body: { status: true } }); return; } - if (req.method !== "POST" || (path !== "/" && path !== "/search")) { - send(res, { status: 404, body: { error: "POST a search to / or /search" } }); + if (req.method === "GET" && path === "/openapi.json") { + 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" } }); return; } // Stop the search, and free its tab, if the client goes away first. @@ -127,9 +217,16 @@ const main = (): void => { return; } const started = Date.now(); - const reply = await handleSearch(raw, runSearch, controller.signal); - const outcome = reply.status === 200 ? `${(reply.body as unknown[]).length} results` : `error ${reply.status}`; - console.log(`search ${outcome} in ${Date.now() - started} ms`); + 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`); if (reply.status !== 200) console.log(` ${(reply.body as { error: string }).error}`); send(res, reply); }); diff --git a/castle-search/src/search.ts b/castle-search/src/search.ts index b5effb8..0178349 100644 --- a/castle-search/src/search.ts +++ b/castle-search/src/search.ts @@ -393,3 +393,27 @@ export const resolveLink = async (url: string, signal: AbortSignal): Promise<Res }; const isGoogleHost = (url: string): boolean => /^https?:\/\/([a-z0-9-]+\.)*google\.(?:com?\.)?[a-z]{2,3}\//.test(url); + +/** + * Runs `fn` on a fresh background tab of the shared browser, under the same + * concurrency cap and connection as a search, and closes the tab on every path. + * For page work that is not an engine search, such as x-search.ts. + */ +export const withPage = async <T>( + signal: AbortSignal, + fn: (page: PageSession, endpoint: string) => Promise<T>, +): 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); + try { + return await fn(page, endpoint); + } finally { + await page.close(); + } + } finally { + release(); + } +}; diff --git a/castle-search/src/x-search.ts b/castle-search/src/x-search.ts new file mode 100644 index 0000000..6ea086c --- /dev/null +++ b/castle-search/src/x-search.ts @@ -0,0 +1,229 @@ +/** + * Search X (x.com) through the browser's own signed-in session. + * + * 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 ($5 per 1,000 posts returned). + * + * The query is X's own search syntax, passed through untouched: from:handle, + * to:handle, since:2026-11-01, until:..., min_faves:100, filter:media, + * -filter:replies, lang:en, "exact phrase", OR. + * + * Every failure throws, as for the engine searches (errors.ts): signed out, an + * X error page, or nothing rendering are reported as such, never as "no posts". + * Only X's own "No results" page is an empty answer, and even that is reported + * as an error naming the query, since an over-narrow query is the usual cause. + */ + +import { BrowserUnavailableError, NoResultsError } from "./errors.ts"; +import { withPage } from "./search.ts"; + +export type XSort = "latest" | "top"; + +export type XPost = { + url: string; + author: string; + handle: string; + /** ISO 8601, from the post's <time datetime>. */ + time: string; + text: string; + /** X cut the text short behind "Show more"; the full post is at url. */ + truncated?: true; + /** "@someone reposted", "Pinned" and the like. */ + context?: string; + replies?: number; + reposts?: number; + likes?: number; + views?: number; + /** Count of photos and videos attached. */ + media?: number; + quoted?: { author: string; handle: string; text: string }; +}; + +export type XSearchRequest = { query: string; count: number; sort?: XSort | undefined }; + +export type XSearchOutcome = { query: string; sort: XSort; searchUrl: string; posts: XPost[] }; + +/** X shows ~5-10 posts per screen; scroll this many times at most for more. */ +const MAX_SCROLLS = 6; +/** Wall clock for one search: render, scrolls and all. */ +export const X_SEARCH_TIMEOUT_MS = 25_000; +const POLL_MS = 400; + +export const buildXSearchUrl = (query: string, sort: XSort): string => { + const params = new URLSearchParams({ q: query, src: "typed_query" }); + // f=live is the Latest tab. Top is the default tab, no parameter. + if (sort === "latest") params.set("f", "live"); + return `https://x.com/search?${params.toString()}`; +}; + +/** 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>> = {}; + for (const m of (label ?? "").matchAll(/(\d+)\s+(repl(?:y|ies)|reposts?|likes?|views?)/g)) { + const n = Number(m[1]); + const word = m[2] ?? ""; + if (word.startsWith("repl")) out.replies = n; + else if (word.startsWith("repost")) out.reposts = n; + else if (word.startsWith("like")) out.likes = n; + else if (word.startsWith("view")) out.views = n; + } + return out; +}; + +type RawPost = Omit<XPost, "replies" | "reposts" | "likes" | "views"> & { metrics: string | null }; + +type XProbe = { + url: string; + state: "signed-out" | "posts" | "empty" | "error" | "loading"; + posts: RawPost[]; + note: string; +}; + +/** + * 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 = `(() => { + 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 }; + } + 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 posts = []; + for (const a of document.querySelectorAll('article[data-testid="tweet"]')) { + const time = a.querySelector("time"); + 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 post = { + url: link.href.split("?")[0].replace(/\\/(analytics|photo\\/\\d+)$/, ""), + ...nameOf(a), + time: time.getAttribute("datetime") || "", + text: own.map(txt).join("\\n"), + metrics: null, + }; + const bar = a.querySelector('[role="group"][aria-label]'); + if (bar) post.metrics = bar.getAttribute("aria-label"); + if (a.querySelector('[data-testid="tweet-text-show-more-link"]')) post.truncated = true; + const ctx = txt(a.querySelector('[data-testid="socialContext"]')); + if (ctx) post.context = ctx; + const media = [...a.querySelectorAll('[data-testid="tweetPhoto"], video')].filter((m) => !(quote && quote.contains(m))).length; + if (media) post.media = media; + if (quote) { + const qt = txt(quote.querySelector('[data-testid="tweetText"]')); + if (qt) post.quoted = { ...nameOf(quote), text: qt }; + } + posts.push(post); + } + if (posts.length) return { url, state: "posts", posts, note: "" }; + const empty = document.querySelector('[data-testid="emptyState"]'); + if (empty) return { url, state: "empty", posts, note: txt(empty).slice(0, 300) }; + const retry = [...document.querySelectorAll('button, [role="button"]')].find((b) => /^Retry$/.test(txt(b))); + if (retry) { + const col = document.querySelector('[data-testid="primaryColumn"]'); + return { url, state: "error", posts, note: txt(col).slice(0, 300) }; + } + return { url, state: "loading", posts, note: "" }; +})()`; + +const SCROLL = `window.scrollBy(0, window.innerHeight * 2), true`; + +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 }); + }); + +const finish = (raw: RawPost): XPost => { + const { metrics, ...rest } = raw; + 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)]); + + let last = "nothing yet"; + try { + return await run(request, sort, searchUrl, 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: ${last}). ` + + `X may be slow or stuck on a page this tool does not recognise.`, + ); + } + throw e; + } +}; + +const run = ( + request: XSearchRequest, + sort: XSort, + searchUrl: string, + signal: AbortSignal, + saw: (state: string) => void, +): Promise<XSearchOutcome> => + withPage(signal, async (page, endpoint) => { + await page.navigate(searchUrl, signal); + + // X renders after the load event: poll until posts, an empty or error + // page, or a sign-in redirect shows up. + let probe: XProbe; + for (;;) { + probe = await page.evaluate<XProbe>(PROBE, signal); + saw(`${probe.state} at ${probe.url}`); + if (probe.state !== "loading") break; + await sleep(POLL_MS, signal); + } + + if (probe.state === "signed-out") { + throw new BrowserUnavailableError( + `X search needs a signed-in X session in the browser (${endpoint}), and it is signed out ` + + `(X sent it to ${probe.url}). Tell the user: someone has to sign in to X in that browser again.`, + ); + } + if (probe.state === "error") { + throw new BrowserUnavailableError( + `X answered with an error page instead of results (${probe.url}): "${probe.note}". This is ` + + `usually X rate-limiting the account; wait a few minutes rather than retrying in a loop.`, + ); + } + if (probe.state === "empty") { + throw new NoResultsError( + `X found no posts for ${JSON.stringify(request.query)} (${sort}). X said: "${probe.note}". ` + + `The page loaded normally, so the query is the cause: check the handle, loosen the terms, ` + + `or try sort "top" or a wider date range.`, + ); + } + + // 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++) { + await sleep(POLL_MS, signal); + take(await page.evaluate<XProbe>(PROBE, signal)); + } + if (seen.size === before) break; // the end of the results + } + + return { query: request.query, sort, searchUrl, posts: [...seen.values()].slice(0, request.count).map(finish) }; + }); diff --git a/castle-search/test/http.test.ts b/castle-search/test/http.test.ts index 6b3045b..b2c90ab 100644 --- a/castle-search/test/http.test.ts +++ b/castle-search/test/http.test.ts @@ -65,3 +65,54 @@ test("listen addresses parse", () => { assert.throws(() => parseListen("11158")); assert.throws(() => parseListen("h:99999")); }); + +import { buildXSearchUrl, parseMetrics } from "../src/x-search.ts"; +import { handleXSearch, OPENAPI, type XSearcher } from "../src/http.ts"; + +const xReturning = (seen: unknown[] = []): XSearcher => async (request) => { + seen.push(request); + return { query: request.query, sort: request.sort ?? "latest", searchUrl: "", posts: [] }; +}; + +test("x-search defaults to the 10 latest posts and clamps count", async () => { + const seen: unknown[] = []; + await handleXSearch(JSON.stringify({ query: " from:nasa " }), xReturning(seen), signal); + await handleXSearch(JSON.stringify({ query: "x", count: 99, sort: "top" }), xReturning(seen), signal); + assert.deepEqual(seen, [ + { query: "from:nasa", count: 10, sort: "latest" }, + { query: "x", count: 20, sort: "top" }, + ]); +}); + +test("x-search refuses a bad sort or query rather than guessing", async () => { + assert.equal((await handleXSearch(JSON.stringify({ query: "x", sort: "new" }), xReturning(), signal)).status, 400); + assert.equal((await handleXSearch(JSON.stringify({ query: "" }), xReturning(), signal)).status, 400); +}); + +test("a failed x-search is a 502 with its reason", async () => { + const reply = await handleXSearch( + JSON.stringify({ query: "x" }), + async () => { + throw new Error("signed out"); + }, + signal, + ); + assert.equal(reply.status, 502); + assert.deepEqual(reply.body, { error: "signed out" }); +}); + +test("x search URLs: latest is f=live, top has no tab parameter", () => { + assert.equal(buildXSearchUrl("from:a b", "latest"), "https://x.com/search?q=from%3Aa+b&src=typed_query&f=live"); + assert.equal(buildXSearchUrl("from:a b", "top"), "https://x.com/search?q=from%3Aa+b&src=typed_query"); +}); + +test("X's action-bar label parses, singular and plural", () => { + assert.deepEqual(parseMetrics("1 reply, 971 reposts, 8852 likes, 1247 bookmarks, 16029782 views"), { + replies: 1, reposts: 971, likes: 8852, views: 16029782, + }); + assert.deepEqual(parseMetrics(null), {}); +}); + +test("the OpenAPI description names the operation Open WebUI will offer", () => { + assert.equal(OPENAPI.paths["/x-search"].post.operationId, "search_x"); +}); |
