@chatpanel/events 0.102.0 → 0.104.0

package/media-transcript.js

@@ -1,535 +1,12 @@
-// Video transcripts, as a contract rather than as a scraper.
-//
-// "Summarise this video" is the same question as "summarise this page" — the only
-// difference is where the words live. On a video page the words are NOT in the DOM: the
-// article body a reader would summarise is a caption track, fetched separately, published
-// in four different serialisations, and duplicated once per language plus once more for
-// the machine-generated version.
-//
-// So the CHOOSING and the PARSING live here, not in a client:
-//   • which of eleven caption tracks is the one the user meant (manual over ASR, their
-//     language over the uploader's, an explicit ask over both);
-//   • json3 / srv3 / srv1 XML / WebVTT / SRT → one segment list;
-//   • segments → readable paragraphs with timestamps you can cite and click.
-//
-// None of that needs a browser, and every client will need all of it: the extension reads
-// the tab, the bridge may be handed a URL by a CLI agent, the gateway may be asked to
-// summarise one server-side, and a mobile app has no DOM to scrape at all. Written inside
-// the extension it would be copied three times and would disagree three ways about which
-// track is "the" transcript.
-//
-// WHAT IS NOT HERE: fetching. Every platform gates caption URLs on the session that asked
-// (cookies, origin, a consent cookie, a per-load token), so the FETCH has to happen where
-// that session is — in the page for the extension, behind the user's own credentials for a
-// CLI. Callers pass a `fetchText` in; this module never reaches the network, which is also
-// what keeps it testable and dependency-free.
+const y=Object.freeze(["youtube.com","www.youtube.com","m.youtube.com","music.youtube.com","youtube-nocookie.com","www.youtube-nocookie.com","youtu.be"]),N=/^[A-Za-z0-9_-]{11}$/;function T(e){const t=String(e||"").trim();if(!t)return 0;if(/^\d+$/.test(t))return Number(t);const n=/^(?:(\d+)h)?(?:(\d+)m)?(?:(\d+)s)?$/.exec(t);return!n||!(n[1]||n[2]||n[3])?0:Number(n[1]||0)*3600+Number(n[2]||0)*60+Number(n[3]||0)}function $(e){let t;try{t=new URL(String(e||"").trim())}catch{return null}if(t.protocol!=="http:"&&t.protocol!=="https:")return null;const n=t.hostname.toLowerCase().replace(/^www\./,"");if(!y.some(c=>n===c.replace(/^www\./,"")))return null;let o="";if(n==="youtu.be")o=t.pathname.split("/").filter(Boolean)[0]||"";else{const c=t.pathname.split("/").filter(Boolean);c[0]==="watch"?o=t.searchParams.get("v")||"":["shorts","embed","live","v"].includes(c[0])&&(o=c[1]||"")}if(!N.test(o))return null;const i=T(t.searchParams.get("t")||t.searchParams.get("start"));return{videoId:o,start:i,url:`https://www.youtube.com/watch?v=${o}`}}function L(e){return $(e)!==null}function g(e){if(!e||typeof e!="object")return null;const t=String(e.baseUrl||e.url||"");if(!t)return null;const n=e.name?.simpleText||e.name?.runs?.map(r=>r.text).join("")||String(e.label||"");return{baseUrl:t,lang:String(e.languageCode||e.lang||"").toLowerCase(),name:n||"",generated:String(e.kind||"")==="asr"||/auto-generated/i.test(n),translatable:e.isTranslatable!==!1}}function D(e){const t=e?.captions?.playerCaptionsTracklistRenderer?.captionTracks;return Array.isArray(t)?t.map(g).filter(Boolean):[]}function B(e){const t=e?.videoDetails||{},n=String(t.videoId||""),r=Number(t.lengthSeconds||0)||0;return{videoId:n,title:String(t.title||""),author:String(t.author||""),durationSec:r,url:n?`https://www.youtube.com/watch?v=${n}`:"",description:String(t.shortDescription||""),live:!!t.isLiveContent}}function m(e){return String(e||"").toLowerCase().replace("_","-").split("-")[0]}function A(e,{language:t="",languages:n=["en"]}={}){const r=(Array.isArray(e)?e:[]).map(g).filter(Boolean);if(!r.length)return null;const o=String(t||"").toLowerCase(),i=(o?[o]:n||[]).map(s=>String(s).toLowerCase()),c=(s,u)=>{let a=0;const l=i.indexOf(s.lang),p=i.findIndex(f=>m(f)===m(s.lang));return l>=0?a+=1e3-l*10:p>=0&&(a+=800-p*10),s.generated||(a+=100),a-u};return r.map((s,u)=>({t:s,s:c(s,u)})).sort((s,u)=>u.s-s.s)[0].t}const I=Object.freeze({clientName:"ANDROID",clientVersion:"20.10.38"});function z(e){const t=/"INNERTUBE_API_KEY":\s*"([^"]+)"/.exec(String(e||""))||/INNERTUBE_API_KEY\\":\\"([^\\"]+)/.exec(String(e||""));return t?t[1]:""}function Y(e,{apiKey:t="",client:n=I}={}){return e?{url:`https://www.youtube.com/youtubei/v1/player${t?`?key=${encodeURIComponent(t)}`:""}`,method:"POST",headers:{"content-type":"application/json"},body:JSON.stringify({context:{client:{...n}},videoId:e})}:null}function b(e,{fmt:t="json3",language:n=""}={}){let r;try{r=new URL(String(e||""))}catch{return""}return t&&r.searchParams.set("fmt",t),n&&r.searchParams.set("tlang",n),r.toString()}const v={amp:"&",lt:"<",gt:">",quot:'"',apos:"'","#39":"'",nbsp:" "};function R(e){let t=String(e||"");for(let n=0;n<2&&(t=t.replace(/&(#x?[0-9a-fA-F]+|[a-zA-Z]+);/g,(r,o)=>{if(o[0]==="#"){const c=o[1]==="x"||o[1]==="X"?parseInt(o.slice(2),16):parseInt(o.slice(1),10);return Number.isFinite(c)&&c>0?String.fromCodePoint(c):r}const i=v[o.toLowerCase()];return i===void 0?r:i}),!!/&[a-zA-Z#]/.test(t));n++);return t}const x=/<\/?(?:i|b|u|s|br|font|v|c|ruby|rt|rp)\b[^>]*>/gi;function d(e){return R(String(e||"").replace(x,"")).replace(x,"").replace(/\s+/g," ").trim()}function S(e){const n=String(e||"").trim().replace(",",".").split(":").map(Number);if(n.some(o=>!Number.isFinite(o)))return NaN;let r=0;for(const o of n)r=r*60+o;return Math.round(r*1e3)}function k(e){let t;try{t=typeof e=="string"?JSON.parse(e):e}catch{return null}if(!Array.isArray(t?.events))return null;const n=[];for(const r of t.events){const o=d((r?.segs||[]).map(i=>i?.utf8||"").join(""));o&&n.push({start:Number(r.tStartMs)||0,dur:Number(r.dDurationMs)||0,text:o})}return n}function E(e){const t=String(e||"");if(!/<(transcript|timedtext|text|p)\b/i.test(t))return null;const n=[],r=/<(?:text|p)\b([^>]*)>([\s\S]*?)<\/(?:text|p)>/gi;let o;for(;o=r.exec(t);){const i=o[1],c=f=>{const h=new RegExp(`${f}="([^"]*)"`,"i").exec(i);return h?Number(h[1]):NaN},s=c("start"),u=Number.isFinite(s)?s*1e3:c("t"),a=c("dur"),l=Number.isFinite(a)?a*1e3:c("d"),p=d(o[2]);p&&n.push({start:Math.round(Number.isFinite(u)?u:0),dur:Math.round(Number.isFinite(l)?l:0),text:p})}return n.length?n:null}function M(e,t){const n=String(e||"").replace(/\r\n?/g,`
+`).split(/\n{2,}/),r=[];for(const o of n){const i=o.split(`
+`).filter(f=>f.trim()),c=i.findIndex(f=>t.test(f));if(c<0)continue;const[s,u]=i[c].split(t),a=S(s),l=S(String(u||"").split(/\s+/)[0]),p=d(i.slice(c+1).join(" "));!p||!Number.isFinite(a)||r.push({start:a,dur:Number.isFinite(l)?Math.max(0,l-a):0,text:p})}return r.length?r:null}const U=e=>M(e,/\s*-->\s*/);function C(e){const t=typeof e=="string"?e:JSON.stringify(e??"");return t.trim()?k(t)||E(t)||U(t)||[]:[]}function w(e){const t=Math.max(0,Math.round(Number(e)||0)/1e3),n=Math.floor(t/3600),r=Math.floor(t%3600/60),o=Math.floor(t%60),i=c=>String(c).padStart(2,"0");return n?`${n}:${i(r)}:${i(o)}`:`${r}:${i(o)}`}function P(e,{windowMs:t=3e4,gapMs:n=2500}={}){const r=(Array.isArray(e)?e:[]).filter(s=>s&&s.text),o=[];let i=null,c=0;for(const s of r){const u=Number(s.start)||0,a=u-c,l=i&&u-i.start>=t,p=i&&c>0&&a>=n;!i||l||p?(i={start:u,text:s.text},o.push(i)):i.text+=" "+s.text,c=u+(Number(s.dur)||0)}for(const s of o)s.text=_(s.text);return o}function _(e){const t=String(e||"").split(" ");for(let n=Math.min(12,t.length>>1);n>=3;n--)for(let r=0;r+2*n<=t.length;r++){let o=!0;for(let i=0;i<n&&o;i++)o=t[r+i]===t[r+n+i];o&&(t.splice(r+n,n),r--)}return t.join(" ")}const j=12e4;function F(e,{timestamps:t=!0,windowMs:n=3e4,gapMs:r=2500,maxChars:o=j}={}){const s=P(e,{windowMs:n,gapMs:r}).map(u=>t?`[${w(u.start)}] ${u.text}`:u.text).join(`
 
-/** Hostnames that serve YouTube watch pages. */
-export const YOUTUBE_HOSTS = Object.freeze([
-  'youtube.com', 'www.youtube.com', 'm.youtube.com', 'music.youtube.com',
-  'youtube-nocookie.com', 'www.youtube-nocookie.com', 'youtu.be',
-]);
+`);return s.length<=o?s:`${s.slice(0,o)}
 
-const VIDEO_ID_RE = /^[A-Za-z0-9_-]{11}$/;
+…[transcript truncated at ${o} characters]`}function O({meta:e={},segments:t=[],language:n="",generated:r=!1,source:o="",...i}={}){const c=[];e.title&&c.push(`# ${e.title}`);const s=[];e.author&&s.push(`Channel: ${e.author}`),e.durationSec&&s.push(`Duration: ${w(e.durationSec*1e3)}`),n&&s.push(`Captions: ${n}${r?" (auto-generated)":""}`),e.url&&s.push(`URL: ${e.url}`),s.length&&c.push(s.join(" · "));const u=String(e.description||"").trim();u&&c.push(`## Description
+${u.slice(0,2e3)}`),c.push("## Transcript");const a=F(t,i),l=`${c.join(`
 
-/** `1h2m3s`, `90s`, `90` → seconds. Returns 0 for anything unparseable. */
-function parseTimeParam(raw) {
-  const s = String(raw || '').trim();
-  if (!s) return 0;
-  if (/^\d+$/.test(s)) return Number(s);
-  const m = /^(?:(\d+)h)?(?:(\d+)m)?(?:(\d+)s)?$/.exec(s);
-  if (!m || !(m[1] || m[2] || m[3])) return 0;
-  return Number(m[1] || 0) * 3600 + Number(m[2] || 0) * 60 + Number(m[3] || 0);
-}
+`)}
 
-/**
- * Recognise a YouTube video URL in any of the shapes people actually paste.
- *
- * Every surface has its own: /watch?v=, youtu.be/, /shorts/, /embed/, /live/, /v/, and the
- * mobile and music hosts on top. Matching only /watch?v= — the one everybody writes first —
- * silently drops Shorts, which is most of what gets pasted into a chat.
- *
- * @returns {{videoId: string, start: number, url: string} | null}
- */
-export function parseYouTubeUrl(input) {
-  let u;
-  try {
-    u = new URL(String(input || '').trim());
-  } catch {
-    return null;
-  }
-  if (u.protocol !== 'http:' && u.protocol !== 'https:') return null;
-  const host = u.hostname.toLowerCase().replace(/^www\./, '');
-  const known = YOUTUBE_HOSTS.some((h) => host === h.replace(/^www\./, ''));
-  if (!known) return null;
-
-  let id = '';
-  if (host === 'youtu.be') {
-    id = u.pathname.split('/').filter(Boolean)[0] || '';
-  } else {
-    const parts = u.pathname.split('/').filter(Boolean);
-    if (parts[0] === 'watch') id = u.searchParams.get('v') || '';
-    else if (['shorts', 'embed', 'live', 'v'].includes(parts[0])) id = parts[1] || '';
-  }
-  if (!VIDEO_ID_RE.test(id)) return null;
-  const start = parseTimeParam(u.searchParams.get('t') || u.searchParams.get('start'));
-  return { videoId: id, start, url: `https://www.youtube.com/watch?v=${id}` };
-}
-
-export function isYouTubeUrl(input) {
-  return parseYouTubeUrl(input) !== null;
-}
-
-// --------------------------------------------------------------------------
-// Track selection
-// --------------------------------------------------------------------------
-
-/** Normalise one YouTube captionTracks entry into the shape the rest of this file uses. */
-function normalizeTrack(t) {
-  if (!t || typeof t !== 'object') return null;
-  const baseUrl = String(t.baseUrl || t.url || '');
-  if (!baseUrl) return null;
-  const name = t.name?.simpleText
-    || t.name?.runs?.map((r) => r.text).join('')
-    || String(t.label || '');
-  return {
-    baseUrl,
-    lang: String(t.languageCode || t.lang || '').toLowerCase(),
-    name: name || '',
-    // `asr` is YouTube's marker for the machine-generated track. It is usually the ONLY
-    // track on a video, so it must never be filtered out — only ranked below a human one.
-    generated: String(t.kind || '') === 'asr' || /auto-generated/i.test(name),
-    translatable: t.isTranslatable !== false,
-  };
-}
-
-/**
- * Pull the caption tracks (and the video's own metadata) out of a YouTube player response.
- *
- * The player response is the JSON blob the watch page hands its player. Reading captions
- * from it is what every open-source transcript library does, because it is the only place
- * the *signed* caption URLs exist — they carry an expiring signature, so a URL guessed
- * from the video id is rejected, and one copied from a previous load has expired.
- */
-export function captionTracksFromPlayerResponse(pr) {
-  const list = pr?.captions?.playerCaptionsTracklistRenderer?.captionTracks;
-  if (!Array.isArray(list)) return [];
-  return list.map(normalizeTrack).filter(Boolean);
-}
-
-export function videoMetaFromPlayerResponse(pr) {
-  const d = pr?.videoDetails || {};
-  const videoId = String(d.videoId || '');
-  const seconds = Number(d.lengthSeconds || 0) || 0;
-  return {
-    videoId,
-    title: String(d.title || ''),
-    author: String(d.author || ''),
-    durationSec: seconds,
-    url: videoId ? `https://www.youtube.com/watch?v=${videoId}` : '',
-    // The uploader's own description is frequently where the links, chapters and
-    // corrections live — a summary that ignores it misses what the video points at.
-    description: String(d.shortDescription || ''),
-    live: !!d.isLiveContent,
-  };
-}
-
-/** `en-GB` and `en_gb` both mean the `en` family. */
-function langFamily(code) {
-  return String(code || '').toLowerCase().replace('_', '-').split('-')[0];
-}
-
-/**
- * Choose the track to read.
- *
- * Ranked, not filtered: a video with only an auto-generated Hindi track must still return
- * that track for an English-preferring user, because the alternative is telling them there
- * is no transcript when there plainly is one. Preference order:
- *
- *   1. an explicitly requested language (exact code beats family)
- *   2. one of the user's preferred languages
- *   3. a human-written track over the machine one
- *   4. the order YouTube itself listed them (its own default is first)
- */
-export function pickCaptionTrack(tracks, { language = '', languages = ['en'] } = {}) {
-  const list = (Array.isArray(tracks) ? tracks : []).map(normalizeTrack).filter(Boolean);
-  if (!list.length) return null;
-  const wanted = String(language || '').toLowerCase();
-  const prefs = (wanted ? [wanted] : languages || []).map((l) => String(l).toLowerCase());
-  const score = (t, i) => {
-    let s = 0;
-    const exact = prefs.indexOf(t.lang);
-    const family = prefs.findIndex((p) => langFamily(p) === langFamily(t.lang));
-    if (exact >= 0) s += 1000 - exact * 10;
-    else if (family >= 0) s += 800 - family * 10;
-    if (!t.generated) s += 100;
-    return s - i; // stable: YouTube's own ordering breaks every remaining tie
-  };
-  return list
-    .map((t, i) => ({ t, s: score(t, i) }))
-    .sort((a, b) => b.s - a.s)[0].t;
-}
-
-// --------------------------------------------------------------------------
-// The InnerTube player request — how a caption URL that WORKS is obtained
-// --------------------------------------------------------------------------
-//
-// The caption URLs printed into the watch page's HTML answer HTTP 200 with an EMPTY BODY —
-// measured on every video tried, with and without session cookies, Referer and Origin. The
-// ones returned by the InnerTube player endpoint for the ANDROID client do not.
-//
-// AND THE CLIENT VERSION IS THE WHOLE DIFFERENCE, which is worth stating because it is
-// invisible and it will go stale:
-//
-//   clientVersion 20.10.38  -> 1 track,  60,441 bytes of captions
-//   clientVersion 19.09.37  -> no captionTracks at all
-//   clientVersion 17.31.35  -> no captionTracks at all
-//
-// A stale version does not error. It returns a well-formed player response with the
-// `captions` block missing, which reads exactly like "this video has no subtitles" — so the
-// failure mode of letting this rot is a feature that quietly claims videos have no captions.
-// tests/media-transcript.test.js pins the shape; a live check is the client's job.
-
-/** The InnerTube client whose player response carries usable caption URLs. */
-export const INNERTUBE_ANDROID = Object.freeze({ clientName: 'ANDROID', clientVersion: '20.10.38' });
-
-/** The public InnerTube key is printed into every watch page; it is not a secret. */
-export function innertubeApiKeyFromHtml(html) {
-  const m = /"INNERTUBE_API_KEY":\s*"([^"]+)"/.exec(String(html || ''))
-    || /INNERTUBE_API_KEY\\":\\"([^\\"]+)/.exec(String(html || ''));
-  return m ? m[1] : '';
-}
-
-/**
- * The request to make, as data — so the caller performs it wherever its network is.
- *
- * Returned rather than sent for the same reason nothing else here fetches: the extension, the
- * bridge and a mobile client each have their own idea of what "fetch" means, and this file
- * has to run in all three.
- */
-export function innertubePlayerRequest(videoId, { apiKey = '', client = INNERTUBE_ANDROID } = {}) {
-  if (!videoId) return null;
-  const query = apiKey ? `?key=${encodeURIComponent(apiKey)}` : '';
-  return {
-    url: `https://www.youtube.com/youtubei/v1/player${query}`,
-    method: 'POST',
-    headers: { 'content-type': 'application/json' },
-    body: JSON.stringify({ context: { client: { ...client } }, videoId }),
-  };
-}
-
-/**
- * Ask a caption URL for a specific serialisation.
- *
- * json3 is the one to want: it is the only format that reports per-segment durations
- * reliably, and it needs no XML parser — which matters because a DOMParser does not exist
- * in a service worker, in Node, or in a mobile JS runtime.
- */
-export function timedTextUrl(baseUrl, { fmt = 'json3', language = '' } = {}) {
-  let u;
-  try {
-    u = new URL(String(baseUrl || ''));
-  } catch {
-    return '';
-  }
-  if (fmt) u.searchParams.set('fmt', fmt);
-  // Ask YouTube to translate only when the track we found is not already the language
-  // asked for; `tlang` on a matching track returns a machine round-trip of itself.
-  if (language) u.searchParams.set('tlang', language);
-  return u.toString();
-}
-
-// --------------------------------------------------------------------------
-// Parsing — four serialisations, one segment list
-// --------------------------------------------------------------------------
-
-const XML_ENTITIES = {
-  amp: '&', lt: '<', gt: '>', quot: '"', apos: "'", '#39': "'", nbsp: ' ',
-};
-
-/** Caption text is double-escaped on the XML endpoints (`&amp;#39;` for an apostrophe). */
-function decodeEntities(s) {
-  let out = String(s || '');
-  for (let pass = 0; pass < 2; pass++) {
-    out = out.replace(/&(#x?[0-9a-fA-F]+|[a-zA-Z]+);/g, (m, name) => {
-      if (name[0] === '#') {
-        const code = name[1] === 'x' || name[1] === 'X'
-          ? parseInt(name.slice(2), 16)
-          : parseInt(name.slice(1), 10);
-        return Number.isFinite(code) && code > 0 ? String.fromCodePoint(code) : m;
-      }
-      const hit = XML_ENTITIES[name.toLowerCase()];
-      return hit === undefined ? m : hit;
-    });
-    if (!/&[a-zA-Z#]/.test(out)) break;
-  }
-  return out;
-}
-
-// Only the tags a caption track actually carries. A blanket `<[^>]*>` strip would also
-// eat `<tag>` where the speaker said "angle bracket tag" and the track escaped it.
-const CAPTION_MARKUP_RE = /<\/?(?:i|b|u|s|br|font|v|c|ruby|rt|rp)\b[^>]*>/gi;
-
-/**
- * Caption text, as text.
- *
- * ORDER MATTERS TWICE. Literal markup is stripped BEFORE decoding, because srv3 wraps every
- * word in an `<s>` span for karaoke timing. Then entities are decoded — twice, since the XML
- * endpoints double-escape (`&amp;#39;` for an apostrophe) — and markup is stripped once more,
- * because `<i>` reaches us escaped rather than literal on some tracks. Decoding first would
- * turn an escaped `&lt;tag&gt;` the speaker actually said into markup and delete it.
- */
-function cleanText(s) {
-  return decodeEntities(String(s || '').replace(CAPTION_MARKUP_RE, ''))
-    .replace(CAPTION_MARKUP_RE, '')
-    .replace(/\s+/g, ' ')
-    .trim();
-}
-
-/** `00:01:02.500` / `01:02,500` / `62.5` → milliseconds. */
-function parseClock(raw) {
-  const s = String(raw || '').trim().replace(',', '.');
-  const parts = s.split(':').map(Number);
-  if (parts.some((n) => !Number.isFinite(n))) return NaN;
-  let sec = 0;
-  for (const p of parts) sec = sec * 60 + p;
-  return Math.round(sec * 1000);
-}
-
-function parseJson3(body) {
-  let doc;
-  try {
-    doc = typeof body === 'string' ? JSON.parse(body) : body;
-  } catch {
-    return null;
-  }
-  if (!Array.isArray(doc?.events)) return null;
-  const out = [];
-  for (const ev of doc.events) {
-    const text = cleanText((ev?.segs || []).map((s) => s?.utf8 || '').join(''));
-    if (!text) continue; // json3 emits empty timing-only events between cues
-    out.push({
-      start: Number(ev.tStartMs) || 0,
-      dur: Number(ev.dDurationMs) || 0,
-      text,
-    });
-  }
-  return out;
-}
-
-function parseTimedTextXml(body) {
-  const src = String(body || '');
-  if (!/<(transcript|timedtext|text|p)\b/i.test(src)) return null;
-  const out = [];
-  // srv1 uses <text start dur>, srv3 uses <p t d>. One regex over both beats requiring a
-  // DOMParser that half our runtimes do not have.
-  const re = /<(?:text|p)\b([^>]*)>([\s\S]*?)<\/(?:text|p)>/gi;
-  let m;
-  while ((m = re.exec(src))) {
-    const attrs = m[1];
-    const num = (name) => {
-      const a = new RegExp(`${name}="([^"]*)"`, 'i').exec(attrs);
-      return a ? Number(a[1]) : NaN;
-    };
-    const startSec = num('start');
-    const startMs = Number.isFinite(startSec) ? startSec * 1000 : num('t');
-    const durSec = num('dur');
-    const durMs = Number.isFinite(durSec) ? durSec * 1000 : num('d');
-    const text = cleanText(m[2]);
-    if (!text) continue;
-    out.push({
-      start: Math.round(Number.isFinite(startMs) ? startMs : 0),
-      dur: Math.round(Number.isFinite(durMs) ? durMs : 0),
-      text,
-    });
-  }
-  return out.length ? out : null;
-}
-
-function parseCueList(body, sepRe) {
-  const blocks = String(body || '').replace(/\r\n?/g, '\n').split(/\n{2,}/);
-  const out = [];
-  for (const block of blocks) {
-    const lines = block.split('\n').filter((l) => l.trim());
-    const timeIdx = lines.findIndex((l) => sepRe.test(l));
-    if (timeIdx < 0) continue;
-    const [rawStart, rawEnd] = lines[timeIdx].split(sepRe);
-    const start = parseClock(rawStart);
-    const end = parseClock(String(rawEnd || '').split(/\s+/)[0]);
-    const text = cleanText(lines.slice(timeIdx + 1).join(' '));
-    if (!text || !Number.isFinite(start)) continue;
-    out.push({ start, dur: Number.isFinite(end) ? Math.max(0, end - start) : 0, text });
-  }
-  return out.length ? out : null;
-}
-
-const parseVtt = (body) => parseCueList(body, /\s*-->\s*/);
-
-/**
- * Any caption serialisation → `[{ start, dur, text }]` (milliseconds), sniffed by content.
- *
- * Sniffing rather than trusting a declared format: the same `fmt=json3` URL answers with
- * XML when the parameter is dropped by a proxy or the track predates json3, and a parser
- * chosen from the request instead of the response fails on exactly those.
- */
-export function parseTimedText(body) {
-  const src = typeof body === 'string' ? body : JSON.stringify(body ?? '');
-  if (!src.trim()) return [];
-  return parseJson3(src) || parseTimedTextXml(src) || parseVtt(src) || [];
-}
-
-// --------------------------------------------------------------------------
-// Formatting
-// --------------------------------------------------------------------------
-
-/** ms → `m:ss`, or `h:mm:ss` once it earns the hour. */
-export function formatTimestamp(ms) {
-  const total = Math.max(0, Math.round(Number(ms) || 0) / 1000);
-  const h = Math.floor(total / 3600);
-  const m = Math.floor((total % 3600) / 60);
-  const s = Math.floor(total % 60);
-  const pad = (n) => String(n).padStart(2, '0');
-  return h ? `${h}:${pad(m)}:${pad(s)}` : `${m}:${pad(s)}`;
-}
-
-/**
- * Caption cues → paragraphs.
- *
- * Cues are 2-6 words long and arrive several per second; handed over raw, a 40-minute talk
- * is 8,000 lines of fragments in which nothing is a sentence. The model then spends its
- * attention on reassembly rather than on the content, and every quote it produces is a
- * fragment. Grouping by a time budget (and by a real pause) restores paragraphs and cuts
- * the token count roughly in half by removing the line breaks alone.
- */
-export function groupSegments(segments, { windowMs = 30000, gapMs = 2500 } = {}) {
-  const list = (Array.isArray(segments) ? segments : []).filter((s) => s && s.text);
-  const out = [];
-  let cur = null;
-  let prevEnd = 0;
-  for (const seg of list) {
-    const start = Number(seg.start) || 0;
-    const gap = start - prevEnd;
-    const tooLong = cur && start - cur.start >= windowMs;
-    const paused = cur && prevEnd > 0 && gap >= gapMs;
-    if (!cur || tooLong || paused) {
-      cur = { start, text: seg.text };
-      out.push(cur);
-    } else {
-      cur.text += ' ' + seg.text;
-    }
-    prevEnd = start + (Number(seg.dur) || 0);
-  }
-  // Auto-generated tracks repeat the tail of each cue as the head of the next (a rolling
-  // two-line caption). Left in, roughly a third of the transcript is duplicated text.
-  for (const p of out) p.text = dedupeOverlap(p.text);
-  return out;
-}
-
-/**
- * Collapse an immediately-repeated run of words: `a b c b c d` → `a b c d`.
- *
- * The rolling two-line caption an auto-generated track emits means each cue re-states the
- * tail of the one before it, so a naive join duplicates roughly a third of the transcript.
- * Longest run first, so `b c b c` collapses as one four-word repeat rather than twice.
- */
-function dedupeOverlap(text) {
-  const words = String(text || '').split(' ');
-  for (let k = Math.min(12, words.length >> 1); k >= 3; k--) {
-    for (let i = 0; i + 2 * k <= words.length; i++) {
-      let same = true;
-      for (let j = 0; j < k && same; j++) same = words[i + j] === words[i + k + j];
-      if (same) { words.splice(i + k, k); i--; }
-    }
-  }
-  return words.join(' ');
-}
-
-export const TRANSCRIPT_MAX_CHARS = 120_000;
-
-/**
- * Segments → the text a model reads.
- *
- * Timestamps are kept by default and are not decoration: they are what lets an answer say
- * "at 12:04 they say…", and what lets the panel turn that into a link that seeks the video.
- */
-export function formatTranscript(segments, {
-  timestamps = true, windowMs = 30000, gapMs = 2500, maxChars = TRANSCRIPT_MAX_CHARS,
-} = {}) {
-  const groups = groupSegments(segments, { windowMs, gapMs });
-  const lines = groups.map((g) => (timestamps ? `[${formatTimestamp(g.start)}] ${g.text}` : g.text));
-  const text = lines.join('\n\n');
-  if (text.length <= maxChars) return text;
-  return `${text.slice(0, maxChars)}\n\n…[transcript truncated at ${maxChars} characters]`;
-}
-
-/**
- * The whole thing a caller attaches or hands a model: metadata header + transcript body.
- *
- * The header exists because a transcript alone is anonymous. "Summarise this" over bare
- * caption text produces a summary that cannot say what it summarised, and a model that
- * cannot see the duration guesses at the shape of what it is reading.
- */
-export function buildTranscriptDocument({
-  meta = {}, segments = [], language = '', generated = false, source = '', ...opts
-} = {}) {
-  const head = [];
-  if (meta.title) head.push(`# ${meta.title}`);
-  const facts = [];
-  if (meta.author) facts.push(`Channel: ${meta.author}`);
-  if (meta.durationSec) facts.push(`Duration: ${formatTimestamp(meta.durationSec * 1000)}`);
-  if (language) facts.push(`Captions: ${language}${generated ? ' (auto-generated)' : ''}`);
-  if (meta.url) facts.push(`URL: ${meta.url}`);
-  if (facts.length) head.push(facts.join(' · '));
-  const desc = String(meta.description || '').trim();
-  if (desc) head.push(`## Description\n${desc.slice(0, 2000)}`);
-  head.push('## Transcript');
-  const body = formatTranscript(segments, opts);
-  const text = `${head.join('\n\n')}\n\n${body}`;
-  return {
-    title: meta.title || 'Video transcript',
-    url: meta.url || '',
-    language,
-    generated,
-    source,
-    segments: segments.length,
-    durationSec: meta.durationSec || 0,
-    text,
-    chars: text.length,
-  };
-}
-
-/**
- * The one orchestration worth sharing: tracks → chosen track → fetched body → document.
- *
- * `fetchText(url)` is injected, because WHERE the fetch runs is the entire reliability
- * story and it differs per client (see the header). Everything either side of it is
- * identical everywhere, so it lives here and gets tested with a fake fetch.
- */
-export async function transcriptFromTracks({
-  tracks, meta = {}, fetchText, language = '', languages = ['en'], source = '', ...opts
-} = {}) {
-  const track = pickCaptionTrack(tracks, { language, languages });
-  if (!track) return null;
-  // Only ask for a translation when the track genuinely is not the language wanted.
-  const wantTranslation = !!language && langFamily(track.lang) !== langFamily(language);
-  const attempts = [
-    timedTextUrl(track.baseUrl, { fmt: 'json3', language: wantTranslation ? language : '' }),
-    timedTextUrl(track.baseUrl, { fmt: 'srv1', language: wantTranslation ? language : '' }),
-    track.baseUrl,
-  ].filter(Boolean);
-  for (const url of attempts) {
-    let body;
-    try {
-      body = await fetchText(url);
-    } catch {
-      continue; // a format the endpoint refuses is a reason to try the next, not to fail
-    }
-    const segments = parseTimedText(body);
-    if (segments.length) {
-      return buildTranscriptDocument({
-        meta, segments, language: track.lang, generated: track.generated, source, ...opts,
-      });
-    }
-  }
-  return null;
-}
+${a}`;return{title:e.title||"Video transcript",url:e.url||"",language:n,generated:r,source:o,segments:t.length,durationSec:e.durationSec||0,text:l,chars:l.length}}async function q({tracks:e,meta:t={},fetchText:n,language:r="",languages:o=["en"],source:i="",...c}={}){const s=A(e,{language:r,languages:o});if(!s)return null;const u=!!r&&m(s.lang)!==m(r),a=[b(s.baseUrl,{fmt:"json3",language:u?r:""}),b(s.baseUrl,{fmt:"srv1",language:u?r:""}),s.baseUrl].filter(Boolean);for(const l of a){let p;try{p=await n(l)}catch{continue}const f=C(p);if(f.length)return O({meta:t,segments:f,language:s.lang,generated:s.generated,source:i,...c})}return null}export{I as INNERTUBE_ANDROID,j as TRANSCRIPT_MAX_CHARS,y as YOUTUBE_HOSTS,O as buildTranscriptDocument,D as captionTracksFromPlayerResponse,w as formatTimestamp,F as formatTranscript,P as groupSegments,z as innertubeApiKeyFromHtml,Y as innertubePlayerRequest,L as isYouTubeUrl,C as parseTimedText,$ as parseYouTubeUrl,A as pickCaptionTrack,b as timedTextUrl,q as transcriptFromTracks,B as videoMetaFromPlayerResponse};

package/meeting-analyzers.js

@@ -1,83 +1 @@
-// What a meeting produces, as declarations.
-//
-// A meeting already generates several kinds of derived thing: a running summary, insight
-// sections (decisions, action items, risks), and live monitors that watch for an answer to
-// a standing question. Each is written separately — its own prompt shape, its own cadence,
-// its own storage — so adding a fifth means touching several files and nothing outside the
-// extension can offer one.
-//
-// They are the same shape underneath: run over the transcript so far, on some trigger,
-// producing a typed result that is stored and shown. Declaring that shape gives three
-// things at once — a Plugins entry the user can switch off, a cadence the runtime can honour
-// without each analyzer implementing its own timer, and a contract the gateway could later
-// run server-side without rewriting callers.
-//
-// WHAT THIS IS NOT: a scheduler. Declaring "every 90 seconds" does not start a timer here;
-// the host decides when to run and this says what running means. Putting the clock in the
-// contract would make it untestable and unrunnable off a browser.
-
-export class AnalyzerError extends Error {
-  constructor(code, message) { super(message); this.name = 'AnalyzerError'; this.code = code; }
-}
-
-/** When an analyzer wants to run. The host maps these to its own timers and events. */
-export const CADENCES = Object.freeze([
-  'periodic',   // every `everyMs` while the meeting is live
-  'on-demand',  // only when the user asks
-  'on-end',     // once, when the meeting finishes
-]);
-
-/**
- * @param produces what the result IS — 'summary' | 'sections' | 'answer' | 'text'. The host
- *        uses it to decide where the output goes, so an analyzer never has to know about
- *        storage.
- * @param run   async ({ transcript, summary, previous, meeting, ask }) => result. `ask` is
- *        the model call, injected: an analyzer that imported one could not run in the
- *        gateway, and could not be tested without a network.
- */
-export function defineMeetingAnalyzer({
-  id, label, produces = 'text', cadence = 'on-demand', everyMs = 0,
-  minTranscriptChars = 0, description = '', run,
-}) {
-  if (!id) throw new AnalyzerError('BAD_ANALYZER', 'analyzer.id required');
-  if (typeof run !== 'function') throw new AnalyzerError('BAD_ANALYZER', `analyzer '${id}': run required`);
-  if (!CADENCES.includes(cadence)) throw new AnalyzerError('BAD_ANALYZER', `analyzer '${id}': unknown cadence '${cadence}'`);
-  if (cadence === 'periodic' && !(everyMs > 0)) {
-    // A periodic analyzer with no interval would either never run or run every tick, and
-    // both look like a bug in the analyzer rather than in its declaration.
-    throw new AnalyzerError('BAD_ANALYZER', `analyzer '${id}': periodic cadence needs everyMs`);
-  }
-  return Object.freeze({ id, label: label || id, produces, cadence, everyMs, minTranscriptChars, description, run });
-}
-
-export function createAnalyzerRegistry() {
-  const analyzers = [];
-  return {
-    add(a) {
-      analyzers.push(a);
-      return () => { const i = analyzers.indexOf(a); if (i >= 0) analyzers.splice(i, 1); };
-    },
-    list: () => [...analyzers],
-    get: (id) => analyzers.find((a) => a.id === id) || null,
-
-    /**
-     * Which analyzers are due right now.
-     *
-     * `lastRunAt` is passed in rather than held here: the registry is a declaration, and a
-     * registry that remembered when things ran would be a second place for that truth to
-     * live — beside the meeting record that already has to store it.
-     */
-    due({ now, cadence = 'periodic', lastRunAt = {}, transcriptChars = 0, admit = null } = {}) {
-      return analyzers.filter((a) => {
-        if (a.cadence !== cadence) return false;
-        if (admit && !admit(a)) return false;
-        // Below the threshold there is nothing worth spending a model call on — an empty
-        // transcript summarised is a paragraph of apology.
-        if (transcriptChars < a.minTranscriptChars) return false;
-        if (a.cadence !== 'periodic') return true;
-        const last = lastRunAt[a.id] || 0;
-        return !last || (now - last) >= a.everyMs;
-      });
-    },
-  };
-}
+class l extends Error{constructor(r,n){super(n),this.name="AnalyzerError",this.code=r}}const u=Object.freeze(["periodic","on-demand","on-end"]);function f({id:e,label:r,produces:n="text",cadence:i="on-demand",everyMs:s=0,minTranscriptChars:c=0,description:t="",run:o}){if(!e)throw new l("BAD_ANALYZER","analyzer.id required");if(typeof o!="function")throw new l("BAD_ANALYZER",`analyzer '${e}': run required`);if(!u.includes(i))throw new l("BAD_ANALYZER",`analyzer '${e}': unknown cadence '${i}'`);if(i==="periodic"&&!(s>0))throw new l("BAD_ANALYZER",`analyzer '${e}': periodic cadence needs everyMs`);return Object.freeze({id:e,label:r||e,produces:n,cadence:i,everyMs:s,minTranscriptChars:c,description:t,run:o})}function a(){const e=[];return{add(r){return e.push(r),()=>{const n=e.indexOf(r);n>=0&&e.splice(n,1)}},list:()=>[...e],get:r=>e.find(n=>n.id===r)||null,due({now:r,cadence:n="periodic",lastRunAt:i={},transcriptChars:s=0,admit:c=null}={}){return e.filter(t=>{if(t.cadence!==n||c&&!c(t)||s<t.minTranscriptChars)return!1;if(t.cadence!=="periodic")return!0;const o=i[t.id]||0;return!o||r-o>=t.everyMs})}}}export{l as AnalyzerError,u as CADENCES,a as createAnalyzerRegistry,f as defineMeetingAnalyzer};