Keysat
0ae59f3550
Introduces RECAP_MODE=multi alongside single-mode self-host: - Tenant auth + accounts (magic-link via System SMTP), per-tenant credit pool, anonymous trial minting with per-IP/-64 caps - Self-serve Pro/Max purchase: inline Lightning (BTCPay) + card (Zaprite), prepaid 30-day periods, expiry-reminder emails - Core-decoupling: relay owns cloud tier/expiry keyed by Recaps user-id - SQLite (better-sqlite3) schema for multi-mode; filesystem unchanged for single - StartOS actions/versions through 0.2.155
4352 lines
182 KiB
JavaScript
4352 lines
182 KiB
JavaScript
import express from "express";
|
||
import cors from "cors";
|
||
import { execFile } from "child_process";
|
||
import { promisify } from "util";
|
||
import fs from "fs/promises";
|
||
import { createWriteStream } from "fs";
|
||
import path from "path";
|
||
import os from "os";
|
||
import https from "https";
|
||
import http from "http";
|
||
import { randomUUID } from "crypto";
|
||
import * as license from "./license.js";
|
||
import {
|
||
sendEvent,
|
||
extractVideoId,
|
||
formatTime,
|
||
parseTimestampedTranscript,
|
||
synthesizeEntriesFromText,
|
||
safeText,
|
||
retryGemini,
|
||
fetchUrl,
|
||
} from "./util.js";
|
||
import { buildAnalysisPrompt } from "./gemini-helpers.js";
|
||
import { runChunkedAnalysis, CHUNKING_CUTOFF_SECONDS } from "./chunked-analyze.js";
|
||
import { getProvider, resolveProviderOpts, PROVIDER_NAMES } from "./providers/index.js";
|
||
import {
|
||
getAudioDuration,
|
||
splitAudioFile,
|
||
downloadPodcastAudio,
|
||
} from "./audio.js";
|
||
import { checkYtdlp, autoUpdateYtdlp } from "./ytdlp.js";
|
||
import {
|
||
initCookies,
|
||
ytCookieArgs,
|
||
ytExtraArgs,
|
||
ytCookieMethod,
|
||
setupCookieRoutes,
|
||
getCookieFilePath,
|
||
} from "./cookies.js";
|
||
import * as config from "./config.js";
|
||
import { initInstallId, getInstallId } from "./install-id.js";
|
||
import * as relayState from "./relay-state.js";
|
||
import * as relayDefault from "./relay-default.js";
|
||
import {
|
||
startRelayCapabilitiesRefresh,
|
||
getRelayCapabilities,
|
||
refreshRelayCapabilities,
|
||
} from "./relay-capabilities.js";
|
||
import {
|
||
isResolvableShareUrl,
|
||
isApplePodcastUrl,
|
||
isSpotifyUrl,
|
||
resolveShareUrl,
|
||
URLResolveError,
|
||
} from "./url-resolver.js";
|
||
import { resolveApiKey, getConfigSnapshot } from "./config.js";
|
||
import * as licenseMW from "./license-middleware.js";
|
||
import { setupLicensePurchaseRoutes } from "./license-purchase.js";
|
||
import {
|
||
setupCreditsPurchaseRoutes,
|
||
sweepUnappliedPurchases,
|
||
} from "./credits-purchase.js";
|
||
import {
|
||
setupLicenseMiddleware,
|
||
setupLicenseRoutes,
|
||
startLicenseRefresh,
|
||
refreshLicenseOnline,
|
||
isFreeUser,
|
||
tryAcquireFreeSlot,
|
||
releaseFreeSlot,
|
||
getCurrentFreeJob,
|
||
abortCurrentFreeJob,
|
||
isFreeJobAborted,
|
||
appendCurrentJobLog,
|
||
} from "./license-middleware.js";
|
||
import {
|
||
initHistory,
|
||
saveToHistory,
|
||
loadMeta,
|
||
saveMeta,
|
||
setupHistoryRoutes,
|
||
getHistoryDir,
|
||
scopeForRequest,
|
||
migrateLegacyLibrary,
|
||
} from "./history.js";
|
||
import {
|
||
getProcessedVideoIds,
|
||
isKnownVideo,
|
||
loadSubscriptions,
|
||
saveSubscriptions,
|
||
loadSkipList,
|
||
addToSkipList,
|
||
loadSeenList,
|
||
addToSeenList,
|
||
loadAutoQueue,
|
||
saveAutoQueue,
|
||
mutateAutoQueue,
|
||
listSubscriptionScopes,
|
||
listAutoQueueScopes,
|
||
migrateGlobalSubscriptionsToOwner,
|
||
} from "./subscriptions.js";
|
||
import {
|
||
mintInternalSession,
|
||
deleteInternalSession,
|
||
adminUserId,
|
||
SESSION_COOKIE,
|
||
} from "./auth-routes.js";
|
||
import { setupLibraryRoutes } from "./library.js";
|
||
import { setupTtsRoutes } from "./tts-routes.js";
|
||
import {
|
||
initAdminAuth,
|
||
setupAdminAuthMiddleware,
|
||
setupAdminAuthRoutes,
|
||
} from "./admin-auth.js";
|
||
import { buildTenantAuthMiddleware } from "./tenant-auth.js";
|
||
|
||
const execFileAsync = promisify(execFile);
|
||
const app = express();
|
||
const PORT = process.env.PORT || 3001;
|
||
|
||
// ── Multi-tenant mode toggle ────────────────────────────────────────────
|
||
// RECAP_MODE is set by startos/main.ts based on the recap_mode field of
|
||
// the config. "single" (default) preserves the original self-hosted
|
||
// behavior; "multi" enables email + magic-link auth, per-user libraries,
|
||
// per-user keysat licenses, and BTCPay subscriptions.
|
||
//
|
||
// Anything outside "multi" is treated as single — defensive parsing so
|
||
// a stray env value can't accidentally enable multi-tenant routing.
|
||
const RECAP_MODE = process.env.RECAP_MODE === "multi" ? "multi" : "single";
|
||
console.log(`[boot] RECAP_MODE=${RECAP_MODE}`);
|
||
|
||
// ── Data directory (configurable for StartOS or local dev) ────────────────
|
||
// On StartOS: DATA_DIR=/data (persistent volume)
|
||
// On local dev: defaults to project root (parent of server/)
|
||
const __dirname = path.dirname(new URL(import.meta.url).pathname);
|
||
const DATA_DIR = process.env.DATA_DIR || path.join(__dirname, "..");
|
||
const historyDir = path.join(DATA_DIR, "history");
|
||
const configDir = path.join(DATA_DIR, "config");
|
||
await fs.mkdir(historyDir, { recursive: true }).catch(() => {});
|
||
await fs.mkdir(configDir, { recursive: true }).catch(() => {});
|
||
|
||
await initHistory({ dataDir: DATA_DIR, mode: RECAP_MODE });
|
||
|
||
// Per-install identity for the upcoming relay backend. Generated once
|
||
// on first boot, persisted in DATA_DIR. Surfaced via /api/health and
|
||
// /api/install-id (read-only). See ./install-id.js for details.
|
||
await initInstallId({ dataDir: DATA_DIR });
|
||
|
||
// API key + live reload moved to ./config.js
|
||
await config.initConfig({ dataDir: DATA_DIR });
|
||
const envPath = config.getEnvPath();
|
||
|
||
// Cookies state + helpers + routes moved to ./cookies.js
|
||
await initCookies({ dataDir: DATA_DIR, envPath });
|
||
|
||
// Admin login gate state. Reads username + scrypt hash + session secret
|
||
// from /data/config/startos-config.json (set via the "Set Admin
|
||
// Password" StartOS action) and refreshes on the same poll cadence as
|
||
// config.js. When no hash is set, the gate is a no-op.
|
||
//
|
||
// Single-mode only. In multi-tenant mode the operator authenticates as
|
||
// a regular user (is_admin = 1) via magic-link, so the legacy admin
|
||
// gate is bypassed entirely.
|
||
if (RECAP_MODE === "single") {
|
||
await initAdminAuth({ dataDir: DATA_DIR });
|
||
}
|
||
|
||
// ── Multi-tenant store + outbound mail ─────────────────────────────────
|
||
// Only initialized when RECAP_MODE === 'multi'. Lazy-imported so the
|
||
// single-mode boot never touches the native better-sqlite3 binding or
|
||
// loads nodemailer.
|
||
if (RECAP_MODE === "multi") {
|
||
const { initDb } = await import("./db.js");
|
||
const { initSmtp } = await import("./smtp.js");
|
||
await initDb({ dataDir: DATA_DIR });
|
||
initSmtp();
|
||
// One-time migration: pre-0.2.77 single-mode installs wrote
|
||
// /data/history/*.json flat. Move them into /data/history/owner/ so
|
||
// they're addressable under a scope. Idempotent (sentinel-guarded).
|
||
try {
|
||
await migrateLegacyLibrary();
|
||
} catch (err) {
|
||
console.warn("[boot] migrateLegacyLibrary failed:", err);
|
||
}
|
||
|
||
// Fixup for installs upgraded from <0.2.91: the first admin signup
|
||
// used to rename /data/history/owner/ → /data/history/<admin_user_id>/.
|
||
// The new convention is admin → always "owner" so multi↔single mode
|
||
// switching is reversible. If the legacy rename happened, reverse it
|
||
// here. Idempotent — runs only when the data layout requires it.
|
||
try {
|
||
const { reclaimAdminLibraryToOwner } = await import("./history.js");
|
||
const { getDb } = await import("./db.js");
|
||
await reclaimAdminLibraryToOwner({ db: getDb() });
|
||
} catch (err) {
|
||
console.warn("[boot] reclaimAdminLibraryToOwner failed:", err);
|
||
}
|
||
}
|
||
|
||
// resolveApiKey moved to ./config.js
|
||
|
||
app.use(cors());
|
||
app.use(express.json({ limit: "100mb" }));
|
||
|
||
// ── Tenant auth (multi-mode) / synthetic owner (single-mode) ───────────
|
||
// MUST run before every other /api/* middleware so downstream handlers
|
||
// can rely on req.userId being populated. In single mode the middleware
|
||
// is a no-op shim that stamps req.userId="owner". In multi mode it
|
||
// validates the recap_session cookie and 401s non-public paths.
|
||
app.use(buildTenantAuthMiddleware({ mode: RECAP_MODE }));
|
||
|
||
// Single-mode whoami stub. Lets the frontend probe what mode it's
|
||
// running in WITHOUT branching on a separate config endpoint. In
|
||
// single mode every "user" is the operator, so always state:signed_in
|
||
// with is_admin:true — the existing settings UI is the right thing to
|
||
// render for them.
|
||
if (RECAP_MODE === "single") {
|
||
app.get("/api/account/whoami", (req, res) => {
|
||
res.json({
|
||
recap_mode: "single",
|
||
state: "signed_in",
|
||
user: { id: "owner", is_admin: true, has_license: true },
|
||
});
|
||
});
|
||
}
|
||
|
||
// ── Magic-link auth + whoami (multi mode only) ──────────────────────────
|
||
// /auth/request-link, /auth/verify, /auth/signout, /api/account/whoami.
|
||
// These have to be registered AFTER the tenant-auth middleware (so the
|
||
// public-path bypass routes them past the 401 gate) but BEFORE any
|
||
// /api/* route that assumes req.userId is set.
|
||
if (RECAP_MODE === "multi") {
|
||
const { setupAuthRoutes } = await import("./auth-routes.js");
|
||
const { setupAdminRoutes } = await import("./admin-routes.js");
|
||
const { setupAccountRoutes } = await import("./account-routes.js");
|
||
const { setupBillingRoutes } = await import("./billing-routes.js");
|
||
setupAuthRoutes(app);
|
||
// Account routes (per-user session management). Mount before admin
|
||
// routes so request-handler order is predictable; both use different
|
||
// paths so no conflict.
|
||
setupAccountRoutes(app);
|
||
// Admin routes (tenant list, grant credits, recent-signups). Every
|
||
// route inside is wrapped in requireOperator — non-admins hit 403.
|
||
setupAdminRoutes(app);
|
||
// Self-serve purchase: POST /api/billing/buy + GET /api/billing/status.
|
||
// A signed-in Core user buys their own Pro/Max period via the relay's
|
||
// BTCPay invoice. Multi-mode only (single-mode operators carry a
|
||
// license, not a relay-owned cloud tier).
|
||
setupBillingRoutes(app);
|
||
// Expiry-reminder emails: a daily scan asks the relay who's expiring and
|
||
// emails them via the recaps.cc SMTP transport. Self-gates on SMTP +
|
||
// public URL + relay being configured, so it's a safe no-op until then.
|
||
const { startReminderScheduler } = await import("./subscription-reminders.js");
|
||
startReminderScheduler();
|
||
|
||
// /api/account/whoami — frontend hits this on every page load to
|
||
// determine which UI state to render:
|
||
// - signed-in user → full app
|
||
// - active trial w/budget → app + "N trials left" badge
|
||
// - trial exhausted → sign-up screen
|
||
// - no cookies → landing page (paste-and-try)
|
||
app.get("/api/account/whoami", async (req, res) => {
|
||
const base = { recap_mode: "multi" };
|
||
if (req.user) {
|
||
// Self-healing sweep: if the user paid for credits but the
|
||
// BTCPay redirect killed the frontend poller before /api/credits/
|
||
// invoice/:id could fire, the pending_purchases row stays
|
||
// unapplied. Catch it up here so the very next whoami call
|
||
// (every page load) reflects the correct balance.
|
||
try {
|
||
await sweepUnappliedPurchases({
|
||
buyerType: "user",
|
||
buyerId: req.user.id,
|
||
req,
|
||
});
|
||
} catch (err) {
|
||
console.warn(
|
||
`[whoami] purchase sweep failed for user ${req.user.id}: ${err?.message || err}`,
|
||
);
|
||
}
|
||
return res.json({
|
||
...base,
|
||
state: "signed_in",
|
||
user: {
|
||
id: req.user.id,
|
||
email: req.user.email,
|
||
display_name: req.user.display_name,
|
||
is_admin: !!req.user.is_admin,
|
||
has_license: !!req.user.keysat_license,
|
||
has_password: !!req.user.password_hash,
|
||
created_at: req.user.created_at,
|
||
},
|
||
});
|
||
}
|
||
if (req.trial) {
|
||
// Same sweep for anon visitors who bought a credit pack — apply
|
||
// any settled invoices to anon_trials.credits_total before we
|
||
// compute credits_remaining for the response.
|
||
try {
|
||
await sweepUnappliedPurchases({
|
||
buyerType: "anon",
|
||
buyerId: req.trial.cookie_id,
|
||
req,
|
||
});
|
||
// Re-read the trial row so the response reflects any credits
|
||
// just applied by the sweep.
|
||
const { lookupTrial } = await import("./anon-trial.js");
|
||
const refreshed = lookupTrial(req.trial.cookie_id);
|
||
if (refreshed) req.trial = refreshed;
|
||
} catch (err) {
|
||
console.warn(
|
||
`[whoami] purchase sweep failed for trial ${req.trial.cookie_id}: ${err?.message || err}`,
|
||
);
|
||
}
|
||
return res.json({
|
||
...base,
|
||
state: "trial",
|
||
trial: {
|
||
credits_total: req.trial.credits_total,
|
||
credits_used: req.trial.credits_used,
|
||
credits_remaining:
|
||
req.trial.credits_total - req.trial.credits_used,
|
||
},
|
||
});
|
||
}
|
||
// Anonymous — include the operator-configured trial allowance AND
|
||
// the post-signup grant so the UI can compose dynamic copy on the
|
||
// tier signup modal ("3 trial credits, +5 more on signup" vs
|
||
// "your trial credits transfer over"). Without exposing both, the
|
||
// Free card has to hardcode a number that goes stale the moment
|
||
// an operator tunes their config.
|
||
//
|
||
// ALSO check whether this visitor's IP is at the lifetime
|
||
// trials_per_ip_lifetime cap. If yes, force available_trial_credits
|
||
// to 0 and stamp a trial_blocked_reason so the UI can swap the
|
||
// misleading "N free credits ready" pill for "you've used your
|
||
// free trial — sign up or buy credits" copy. Without this gate
|
||
// the pill advertises credits the visitor literally cannot mint
|
||
// (anon-trial.js's issueIfEligible() will refuse), the user
|
||
// submits a video, the relay rejects, and they see a confusing
|
||
// "Processing failed" with the player half-loaded.
|
||
let availableTrialCredits = 0;
|
||
let signupGrantCredits = 0;
|
||
let trialBlockedReason = null;
|
||
try {
|
||
const snap = await getConfigSnapshot();
|
||
availableTrialCredits = Math.max(
|
||
0,
|
||
parseInt(snap?.trial_credits_per_visitor ?? 1, 10) || 0,
|
||
);
|
||
signupGrantCredits = Math.max(
|
||
0,
|
||
parseInt(snap?.tenant_default_credits ?? 0, 10) || 0,
|
||
);
|
||
// IP cap check — only meaningful in multi mode where the
|
||
// anon-trial system exists. Lazy-import so single-mode doesn't
|
||
// pull the trial DB in unnecessarily.
|
||
if (availableTrialCredits > 0 && RECAP_MODE === "multi") {
|
||
try {
|
||
const { getClientIp, ipTrialsLifetime } =
|
||
await import("./anon-trial.js");
|
||
const ip = getClientIp(req);
|
||
const perIpLifetime = Math.max(
|
||
1,
|
||
parseInt(
|
||
snap?.trials_per_ip_lifetime ?? snap?.trials_per_ip_per_day ?? 5,
|
||
10,
|
||
) || 5,
|
||
);
|
||
if (ip && ipTrialsLifetime(ip) >= perIpLifetime) {
|
||
availableTrialCredits = 0;
|
||
trialBlockedReason = "ip_cap_reached";
|
||
}
|
||
} catch (err) {
|
||
// best-effort — if the cap check fails we don't want to
|
||
// suddenly downgrade a legitimate visitor's pill to zero.
|
||
// Leave the configured value and log for diagnosis.
|
||
console.warn(
|
||
`[whoami] IP cap check failed: ${err?.message || err}`,
|
||
);
|
||
}
|
||
}
|
||
} catch {
|
||
// best-effort; fall through with 0 so UI degrades to "Sign in"
|
||
// CTA rather than a misleading "N free credits" badge.
|
||
}
|
||
res.json({
|
||
...base,
|
||
state: "anonymous",
|
||
available_trial_credits: availableTrialCredits,
|
||
signup_grant_credits: signupGrantCredits,
|
||
// Set only when the visitor's IP can't mint a new trial cookie.
|
||
// The frontend uses this to swap the "N free credits ready"
|
||
// pill for an honest "trial used up — sign up / buy" CTA AND
|
||
// to refuse the optimistic submit-flow (no video player
|
||
// render, no fake processing status).
|
||
trial_blocked_reason: trialBlockedReason,
|
||
});
|
||
});
|
||
}
|
||
|
||
// ── Admin login gate (single mode only) ─────────────────────────────────
|
||
// MUST run before the license middleware: if an admin password is set,
|
||
// nobody (licensed or not) reaches the activation flow without first
|
||
// passing the login. Endpoints needed by the login UI itself
|
||
// (/api/admin/status, /api/admin/login, /api/admin/logout) and
|
||
// /api/health stay open.
|
||
if (RECAP_MODE === "single") {
|
||
setupAdminAuthMiddleware(app);
|
||
setupAdminAuthRoutes(app);
|
||
}
|
||
|
||
// ── Keysat licensing ────────────────────────────────────────────────────────
|
||
// All license-aware request handling (gate, Pro feature gates, /api/license
|
||
// routes, free-tier slot management, periodic online refresh) lives in
|
||
// ./license-middleware.js. Importers read the current state via
|
||
// licenseMW.LIC (a live binding).
|
||
setupLicenseMiddleware(app);
|
||
setupLicenseRoutes(app);
|
||
// In-app purchase flow: proxies Keysat's public policies + purchase
|
||
// + poll endpoints so the buy page renders in Recap's own visual
|
||
// style instead of being redirected to Keysat's hosted /buy/<slug>
|
||
// page. On a settled invoice the issued license is written to disk
|
||
// and licenseMW.refreshLicenseOnline picks it up immediately.
|
||
setupLicensePurchaseRoutes(app, {
|
||
onLicenseActivated: () => licenseMW.refreshLicenseOnline("post-purchase"),
|
||
});
|
||
// Credit-purchase proxy: lets the buyer top-up relay credits via the
|
||
// operator's BTCPay store. All BTCPay credentials live on the relay
|
||
// — Recap just forwards the buyer's pick and proxies the polling.
|
||
setupCreditsPurchaseRoutes(app);
|
||
startLicenseRefresh();
|
||
|
||
// Boot-time fetch of the relay's /relay/capabilities so the chunking
|
||
// decision in /api/process can honor the operator's current backend
|
||
// preference. Refreshes hourly. Safe defaults are used until the
|
||
// first successful fetch lands.
|
||
startRelayCapabilitiesRefresh();
|
||
|
||
// History storage + routes moved to ./history.js
|
||
// (saveToHistory, loadMeta, saveMeta are imported above)
|
||
// addToSkipList is scope-keyed (./subscriptions.js): deleting a session
|
||
// suppresses re-queueing that video in the SAME scope's subscriptions.
|
||
setupHistoryRoutes(app, { addToSkipList: (scope, id) => addToSkipList(scope, id) });
|
||
|
||
// Audio-first ("walking mode") TTS routes — synthesize + cache + serve
|
||
// per-topic summary clips. Self-gates access (Max in multi mode); the
|
||
// /api/tts prefix is license-exempt so the route's own gate decides.
|
||
setupTtsRoutes(app);
|
||
|
||
// Serve the frontend from ../public
|
||
app.use(express.static(path.join(__dirname, "..", "public")));
|
||
app.use("/assets", express.static(path.join(__dirname, "..", "assets")));
|
||
|
||
// checkYtdlp + autoUpdateYtdlp moved to ./ytdlp.js
|
||
|
||
// PRICING + calcCost + buildAnalysisPrompt moved to ./gemini-helpers.js
|
||
// safeText + retryGemini moved to ./util.js
|
||
|
||
// ── Health check ───────────────────────────────────────────────────────────
|
||
|
||
app.get("/api/health", async (req, res) => {
|
||
const info = await checkYtdlp();
|
||
// Check cookies.txt freshness
|
||
const cookieMethod = ytCookieMethod();
|
||
let cookieInfo = { method: cookieMethod };
|
||
if (cookieMethod === "cookies.txt") {
|
||
try {
|
||
const stat = await fs.stat(getCookieFilePath());
|
||
const ageMs = Date.now() - stat.mtimeMs;
|
||
const ageDays = Math.floor(ageMs / (1000 * 60 * 60 * 24));
|
||
cookieInfo.fileAgeDays = ageDays;
|
||
cookieInfo.fileExpiring = ageDays > 12; // cookies typically expire after ~14 days
|
||
} catch {}
|
||
}
|
||
res.json({ ok: true, ytdlp: info.installed, hasServerKey: !!config.serverApiKey, cookies: cookieInfo, installId: getInstallId(), ...info });
|
||
});
|
||
|
||
// Read-only install identity. Used by the UI's settings panel so the
|
||
// operator can verify the install has been provisioned, and by the
|
||
// future relay client to attach the X-Recap-Install-Id header. Open
|
||
// path — license gate doesn't apply (the relay needs this ID to be
|
||
// reachable before any credits have been granted).
|
||
app.get("/api/install-id", (_req, res) => {
|
||
res.json({ installId: getInstallId() });
|
||
});
|
||
|
||
// Proxy through to the relay's /relay/policy endpoint. Used by the UI
|
||
// to render dynamic copy (e.g. "N relay credits" in the activation
|
||
// screen reflects whatever the operator currently has the relay
|
||
// configured to give Core users — no Recap update needed when the
|
||
// operator tunes tier quotas). Cached in-memory for a short window
|
||
// so a busy UI doesn't hammer the relay.
|
||
let __cachedRelayPolicy = { at: 0, body: null };
|
||
app.get("/api/relay/policy", async (_req, res) => {
|
||
const base = relayDefault.getRelayBaseURL();
|
||
if (!base) return res.json({ configured: false, tiers: null });
|
||
const ttl = 5 * 60 * 1000;
|
||
if (__cachedRelayPolicy.body && Date.now() - __cachedRelayPolicy.at < ttl) {
|
||
return res.json({ configured: true, ...__cachedRelayPolicy.body });
|
||
}
|
||
try {
|
||
const r = await fetch(`${base.replace(/\/$/, "")}/relay/policy`, {
|
||
signal: AbortSignal.timeout(5000),
|
||
});
|
||
if (!r.ok) {
|
||
return res.json({
|
||
configured: true,
|
||
tiers: null,
|
||
error: `HTTP ${r.status}`,
|
||
});
|
||
}
|
||
const body = await r.json();
|
||
__cachedRelayPolicy = { at: Date.now(), body };
|
||
res.json({ configured: true, ...body });
|
||
} catch (err) {
|
||
res.json({
|
||
configured: true,
|
||
tiers: null,
|
||
error: err?.message || String(err),
|
||
});
|
||
}
|
||
});
|
||
|
||
// Last-known relay state (credits + tier) cached in-process. The UI
|
||
// polls this for the "N credits remaining · Tier: X" banner; the
|
||
// underlying numbers are refreshed every time a relay provider call
|
||
// lands.
|
||
//
|
||
// First-paint UX: when the cache is empty (no relay calls yet AND no
|
||
// prior ping has populated it), opportunistically hit the relay's
|
||
// /balance endpoint with a short timeout so the banner shows real
|
||
// numbers on first page load instead of "balance unknown". Best
|
||
// effort — if the relay is unreachable the cache stays null and the
|
||
// UI falls back to its "unknown" copy. `?refresh=1` forces a ping
|
||
// even when the cache already has a value.
|
||
//
|
||
// `configured` is true when relay-default.js has a non-empty URL —
|
||
// operator-controlled at build time, never user-configurable.
|
||
app.get("/api/relay/status", async (req, res) => {
|
||
const configured = !!relayDefault.getRelayBaseURL();
|
||
|
||
// ── Multi-mode: return per-user credit view ────────────────────────────
|
||
// Each user role sees a different "credits remaining" number:
|
||
// - trial: trial.credits_total - trial.credits_used (local)
|
||
// - free tenant: tenant_credits.balance (local, set by operator)
|
||
// - paid user: their relay-pool balance (relay-side, license-keyed)
|
||
// - admin: the operator's relay pool (current single-mode behavior)
|
||
// - anonymous: no credits to show; configured + null balance
|
||
//
|
||
// The frontend reads `creditsRemaining` and `tier` and renders a
|
||
// single badge. We construct the right shape for the requester here.
|
||
if (RECAP_MODE === "multi" && !(req.user && req.user.is_admin)) {
|
||
// Anonymous + no trial cookie at all.
|
||
if (!req.user && !req.trial) {
|
||
return res.json({
|
||
configured,
|
||
creditsRemaining: null,
|
||
tier: null,
|
||
scope: "anonymous",
|
||
lastUpdated: Date.now(),
|
||
lastError: null,
|
||
});
|
||
}
|
||
// Self-healing sweep — catch any settled-but-unapplied credit
|
||
// purchases for this buyer before we report their balance. Without
|
||
// it, an anon or signed-in user who pays via BTCPay but loses the
|
||
// poll loop on redirect sees a stale balance until they manually
|
||
// re-trigger the apply (which there's no UI for). Cheap: bounded
|
||
// query (5 invoices, 30-min lookback) + short-timeout relay calls.
|
||
try {
|
||
if (req.user) {
|
||
await sweepUnappliedPurchases({
|
||
buyerType: "user",
|
||
buyerId: req.user.id,
|
||
req,
|
||
});
|
||
} else if (req.trial) {
|
||
await sweepUnappliedPurchases({
|
||
buyerType: "anon",
|
||
buyerId: req.trial.cookie_id,
|
||
req,
|
||
});
|
||
// Re-read trial row so the response uses post-sweep credits.
|
||
const { lookupTrial } = await import("./anon-trial.js");
|
||
const refreshed = lookupTrial(req.trial.cookie_id);
|
||
if (refreshed) req.trial = refreshed;
|
||
}
|
||
} catch (err) {
|
||
console.warn(
|
||
`[relay/status] purchase sweep failed: ${err?.message || err}`,
|
||
);
|
||
}
|
||
// Trial visitor.
|
||
if (req.trial) {
|
||
const remaining = Math.max(
|
||
0,
|
||
(req.trial.credits_total || 0) - (req.trial.credits_used || 0),
|
||
);
|
||
return res.json({
|
||
configured,
|
||
creditsRemaining: remaining,
|
||
tier: "trial",
|
||
scope: "trial",
|
||
lastUpdated: Date.now(),
|
||
lastError: null,
|
||
});
|
||
}
|
||
// Core-decoupling: a user's paid status is their relay-owned tier
|
||
// (req.user.tier), NOT any leftover Keysat license. A PAID cloud user
|
||
// (tier pro/max) falls through to the real relay ping below, which
|
||
// pickRelayIdentity routes to their cloud user:<id> pool. Everyone
|
||
// else signed in is a free tenant on the operator pool — regardless of
|
||
// whether they still carry a legacy license (which is now ignored).
|
||
const isPaidCloudUser =
|
||
req.user && (req.user.tier === "pro" || req.user.tier === "max");
|
||
// Free tenant. Lazy replenish-check happens inside getOrInit, so the
|
||
// displayed balance reflects any due refill.
|
||
if (req.user && !isPaidCloudUser) {
|
||
try {
|
||
const { getOrInit } = await import("./tenant-credits.js");
|
||
const credits = await getOrInit(req.user.id);
|
||
return res.json({
|
||
configured,
|
||
creditsRemaining: credits?.total ?? 0,
|
||
// Surface the bucket split so the UI can show "5 free + 10
|
||
// purchased = 15 total" if it wants. Frontend currently
|
||
// just reads creditsRemaining; the extra fields are
|
||
// additive.
|
||
purchasedBalance: credits?.purchased ?? 0,
|
||
replenishBalance: credits?.replenish ?? 0,
|
||
replenishPeriod: credits?.period ?? "off",
|
||
tier: "free",
|
||
scope: "free_tenant",
|
||
lastUpdated: Date.now(),
|
||
lastError: null,
|
||
});
|
||
} catch (err) {
|
||
console.warn("[relay/status] tenant_credits lookup failed:", err);
|
||
return res.json({
|
||
configured,
|
||
creditsRemaining: null,
|
||
tier: "free",
|
||
scope: "free_tenant",
|
||
lastUpdated: Date.now(),
|
||
lastError: "balance_unknown",
|
||
});
|
||
}
|
||
}
|
||
// Paid cloud user (tier pro/max) — fall through to the real relay
|
||
// ping. pickRelayIdentity routes it to their cloud user:<id> identity
|
||
// (or the operator pool if the operator key isn't configured), so the
|
||
// response reflects the pool their calls actually bill.
|
||
}
|
||
|
||
// Determine which credit-key this request would use against the relay.
|
||
// The relay-state cache is keyed by the same value so each (operator
|
||
// / paid user) gets its own snapshot — no more clobbering when an
|
||
// admin and a tenant hit the endpoint in the same Node process.
|
||
// Constructs a temporary provider just to inherit its creditKey
|
||
// logic; cheap (no network).
|
||
const identityCfg = await getConfigSnapshot();
|
||
const probeOpts = resolveProviderOpts("relay", {
|
||
config: identityCfg,
|
||
clientOpts: {},
|
||
req,
|
||
});
|
||
const requestCreditKey = relayState.computeCreditKey({
|
||
// Cloud (core-decoupling) identity keys by userId; license/install
|
||
// identities key by their respective fields. probeOpts carries
|
||
// exactly one of these shapes depending on pickRelayIdentity, so
|
||
// pass all three — computeCreditKey picks userId → license → install
|
||
// in priority order, matching what the provider records under.
|
||
installId: probeOpts.installId,
|
||
licenseKey: probeOpts.licenseKey,
|
||
userId: probeOpts.userId,
|
||
});
|
||
|
||
let snapshot = relayState.getRelayState(requestCreditKey);
|
||
const wantsRefresh = req.query.refresh === "1" || req.query.refresh === "true";
|
||
const cacheEmpty = snapshot.creditsRemaining === null && snapshot.tier === null;
|
||
// Auto-refresh stale cache. Without this, /api/relay/status returns
|
||
// whatever was last cached forever, missing balance changes that
|
||
// happen outside Recap's request flow — e.g. an operator-side
|
||
// BTCPay webhook crediting purchased invoices, or the rescan tool
|
||
// in the relay dashboard. 10-second TTL is short enough that
|
||
// out-of-band changes show up promptly while still avoiding a
|
||
// ping on every dashboard render burst.
|
||
const STALE_THRESHOLD_MS = 10_000;
|
||
const cacheStale =
|
||
!cacheEmpty &&
|
||
typeof snapshot.lastUpdated === "number" &&
|
||
Date.now() - snapshot.lastUpdated > STALE_THRESHOLD_MS;
|
||
if (configured && (cacheEmpty || wantsRefresh || cacheStale)) {
|
||
try {
|
||
const provider = getProvider("relay", probeOpts);
|
||
await provider.pingBalance({ timeoutMs: 5000 });
|
||
snapshot = relayState.getRelayState(requestCreditKey);
|
||
} catch (err) {
|
||
// Log + record so the UI shows a real error instead of a silent
|
||
// "balance unknown". The provider call itself records when the
|
||
// ping reaches the network; this catches the earlier failures
|
||
// (e.g. createRelayProvider throwing because install-id wasn't
|
||
// initialized yet) where lastError wouldn't otherwise be set.
|
||
console.error(
|
||
`[relay/status] ping failed: ${err?.message || err} (stack: ${err?.stack || "n/a"})`
|
||
);
|
||
relayState.recordRelayError(err?.message || String(err), requestCreditKey);
|
||
snapshot = relayState.getRelayState(requestCreditKey);
|
||
}
|
||
}
|
||
res.json({ ...snapshot, configured });
|
||
});
|
||
|
||
// ── Status endpoints ───────────────────────────────────────────────────────
|
||
|
||
app.post("/api/heartbeat", (req, res) => {
|
||
res.json({ ok: true, sleeping: false });
|
||
});
|
||
|
||
app.get("/api/status", (req, res) => {
|
||
res.json({ ok: true, sleeping: false, uptime: process.uptime() });
|
||
});
|
||
|
||
// Shutdown: used by the macOS .app launcher to stop the server cleanly.
|
||
// On StartOS this endpoint is unused (StartOS manages the container lifecycle).
|
||
app.post("/api/shutdown", (req, res) => {
|
||
res.json({ ok: true, message: "Server shutting down..." });
|
||
console.log("\n Server shutdown requested from browser. Goodbye!\n");
|
||
setTimeout(() => process.exit(0), 300);
|
||
});
|
||
|
||
// ── Manual update endpoint ─────────────────────────────────────────────────
|
||
|
||
app.post("/api/update-ytdlp", async (req, res) => {
|
||
const result = await autoUpdateYtdlp(DATA_DIR);
|
||
const info = await checkYtdlp();
|
||
res.json({ ...result, ...info });
|
||
});
|
||
|
||
// ── Cookie management endpoints ───────────────────────────────────────────
|
||
|
||
// /api/cookies/* routes registered via setupCookieRoutes (./cookies.js)
|
||
setupCookieRoutes(app);
|
||
|
||
|
||
// ── Library export/import ──── moved to ./library.js ─────────
|
||
setupLibraryRoutes(app);
|
||
|
||
|
||
// ── Subscriptions ─────────────────────────────────────────────────────────
|
||
// Storage (subscriptions / skip / seen / auto-queue) + dedup live in
|
||
// ./subscriptions.js, keyed by scope. Endpoints pass scopeForRequest(req);
|
||
// the check loop + processor operate on SUBSCRIPTIONS_SCOPE ("owner") while
|
||
// the feature is operator-only in multi mode.
|
||
|
||
// List recent videos from a channel/playlist via yt-dlp (no download)
|
||
// Uses --flat-playlist for speed; returns id + title (no upload_date in flat mode)
|
||
async function listChannelVideosFast(url, limit = 15) {
|
||
const { stdout } = await execFileAsync("yt-dlp", [
|
||
"--print", "%(id)s|%(title)s",
|
||
"--no-download",
|
||
"--playlist-end", String(limit),
|
||
"--flat-playlist",
|
||
...ytCookieArgs(),
|
||
...ytExtraArgs(),
|
||
url,
|
||
], { timeout: 60000 });
|
||
return stdout.trim().split("\n").filter(Boolean).map(line => {
|
||
const idx = line.indexOf("|");
|
||
return { id: line.slice(0, idx), title: line.slice(idx + 1) };
|
||
});
|
||
}
|
||
|
||
// Fetch upload_date for a batch of video IDs (processes in batches of 3)
|
||
// Bails after 2 consecutive failures to avoid grinding through blocked requests
|
||
async function fetchUploadDates(videoIds) {
|
||
if (videoIds.length === 0) return {};
|
||
const dateMap = {};
|
||
const batchSize = 50;
|
||
let consecutiveFails = 0;
|
||
for (let i = 0; i < videoIds.length; i += batchSize) {
|
||
const batch = videoIds.slice(i, i + batchSize);
|
||
const urls = batch.map(id => `https://www.youtube.com/watch?v=${id}`);
|
||
try {
|
||
const { stdout } = await execFileAsync("yt-dlp", [
|
||
"--print", "%(id)s|%(upload_date)s",
|
||
"--no-download",
|
||
...ytCookieArgs(),
|
||
...urls,
|
||
], { timeout: 45000 });
|
||
for (const line of stdout.trim().split("\n").filter(Boolean)) {
|
||
const [id, date] = line.split("|");
|
||
if (id && date && date !== "NA") dateMap[id] = date;
|
||
}
|
||
consecutiveFails = 0;
|
||
subLog(` Batch ${Math.floor(i/batchSize)+1}: got dates for ${batch.length} video(s)`);
|
||
} catch (err) {
|
||
consecutiveFails++;
|
||
subLog(` ⚠ Batch ${Math.floor(i/batchSize)+1} failed: ${err.message.slice(0, 80)}`);
|
||
if (consecutiveFails >= 2) {
|
||
subLog(` ⚠ 2 consecutive failures — aborting yt-dlp date fetch (bot detection likely)`);
|
||
break;
|
||
}
|
||
}
|
||
}
|
||
return dateMap;
|
||
}
|
||
|
||
// ── RSS-based date fetching (bypasses bot detection) ─────────────────────
|
||
|
||
// Fetch a URL and return the response body as a string
|
||
// fetchUrl moved to ./util.js
|
||
|
||
// Get channel_id from a YouTube channel/playlist URL using yt-dlp
|
||
async function getChannelId(url) {
|
||
// Method 1: flat-playlist channel_id (fast, may return NA)
|
||
try {
|
||
const { stdout } = await execFileAsync("yt-dlp", [
|
||
"--print", "%(channel_id)s",
|
||
"--no-download",
|
||
"--playlist-items", "1",
|
||
"--flat-playlist",
|
||
...ytCookieArgs(),
|
||
url,
|
||
], { timeout: 15000 });
|
||
const id = stdout.trim().split("\n")[0];
|
||
if (id && id !== "NA" && id.length > 5) return id;
|
||
} catch {}
|
||
// Method 2: non-flat single video (slower but gets full metadata)
|
||
try {
|
||
const { stdout } = await execFileAsync("yt-dlp", [
|
||
"--print", "%(channel_id)s",
|
||
"--no-download",
|
||
"--playlist-items", "1",
|
||
...ytCookieArgs(),
|
||
url,
|
||
], { timeout: 30000 });
|
||
const id = stdout.trim().split("\n")[0];
|
||
if (id && id !== "NA" && id.length > 5) return id;
|
||
} catch {}
|
||
return null;
|
||
}
|
||
|
||
// Fetch video dates from YouTube RSS feed (no auth, no bot detection)
|
||
// Returns { videoId: "YYYYMMDD", ... } for up to 15 most recent videos
|
||
async function fetchDatesFromRSS(channelId) {
|
||
const dateMap = {};
|
||
if (!channelId) return dateMap;
|
||
try {
|
||
const rssUrl = `https://www.youtube.com/feeds/videos.xml?channel_id=${channelId}`;
|
||
const xml = await fetchUrl(rssUrl);
|
||
const entryRegex = /<entry>[\s\S]*?<yt:videoId>([^<]+)<\/yt:videoId>[\s\S]*?<published>([^<]+)<\/published>[\s\S]*?<\/entry>/g;
|
||
let match;
|
||
while ((match = entryRegex.exec(xml)) !== null) {
|
||
const videoId = match[1];
|
||
const published = match[2]; // e.g. "2025-12-01T18:00:00+00:00"
|
||
const date = published.slice(0, 10).replace(/-/g, ""); // "20251201"
|
||
dateMap[videoId] = date;
|
||
}
|
||
} catch (err) {
|
||
subLog(` ⚠ RSS feed fetch failed: ${err.message}`);
|
||
}
|
||
return dateMap;
|
||
}
|
||
|
||
// ── Podcast RSS feed parsing ────────────────────────────────────────────────
|
||
|
||
// Detect if a URL looks like a podcast RSS feed
|
||
function isPodcastFeedUrl(url) {
|
||
if (!url) return false;
|
||
const u = url.trim().toLowerCase();
|
||
// Common podcast RSS feed patterns
|
||
if (u.includes("/feed") || u.includes("/rss") || u.includes("feeds.") || u.includes(".xml")) return true;
|
||
if (u.includes("anchor.fm") || u.includes("feeds.buzzsprout") || u.includes("feeds.simplecast")) return true;
|
||
if (u.includes("feeds.megaphone") || u.includes("feeds.transistor") || u.includes("feeds.libsyn")) return true;
|
||
if (u.includes("feeds.podcastmirror") || u.includes("feeds.acast") || u.includes("feeds.fireside")) return true;
|
||
if (u.includes("rss.art19") || u.includes("podbean.com/feed")) return true;
|
||
return false;
|
||
}
|
||
|
||
// Apple Podcasts SHOW URL — the whole-podcast page, no `?i=<id>`
|
||
// episode parameter. Format:
|
||
// https://podcasts.apple.com/<country>/podcast/<slug>/id<num>
|
||
// We do prefix-match via regex, then USE URL PARSING to confirm no
|
||
// `i` query param (the episode discriminator). The earlier all-regex
|
||
// version had a misplaced negative lookahead that matched episode
|
||
// URLs too, sending episode links into the subscribe flow by mistake.
|
||
const APPLE_PODCAST_PREFIX_RE =
|
||
/^https?:\/\/(?:www\.)?podcasts\.apple\.com\/[^/]+\/podcast\/[^/]+\/id(\d+)/i;
|
||
|
||
function isAppleShowUrl(url) {
|
||
if (!url) return false;
|
||
const trimmed = url.trim();
|
||
if (!APPLE_PODCAST_PREFIX_RE.test(trimmed)) return false;
|
||
try {
|
||
const parsed = new URL(trimmed);
|
||
// Episode URLs have `?i=<id>` (or `&i=<id>`). Anything else is
|
||
// the show landing page.
|
||
return !parsed.searchParams.has("i");
|
||
} catch {
|
||
return false;
|
||
}
|
||
}
|
||
|
||
// Legacy alias — the regex's first capture group is still used to
|
||
// pull the podcast id in resolveAppleShowToFeed below. Keep it named
|
||
// to avoid splitting that helper into two regex+parse blocks.
|
||
const APPLE_SHOW_URL_RE = APPLE_PODCAST_PREFIX_RE;
|
||
|
||
// Resolve an Apple Podcasts show URL to its RSS feed URL via the
|
||
// public iTunes Lookup API. No auth required. Returns the feedUrl
|
||
// string or null if Apple doesn't have a usable RSS URL on file
|
||
// (rare — shows distributed only via Apple-exclusive feeds, mostly
|
||
// Apple Podcast Subscriptions paid content).
|
||
async function resolveAppleShowToFeed(url) {
|
||
const m = url.match(APPLE_SHOW_URL_RE);
|
||
if (!m) return null;
|
||
const podcastId = m[1];
|
||
try {
|
||
const raw = await fetchUrl(
|
||
`https://itunes.apple.com/lookup?id=${encodeURIComponent(podcastId)}`,
|
||
);
|
||
const parsed = JSON.parse(raw);
|
||
const show = (parsed?.results || []).find(
|
||
(r) => r && r.wrapperType === "track" && r.kind === "podcast",
|
||
);
|
||
return show?.feedUrl || null;
|
||
} catch {
|
||
return null;
|
||
}
|
||
}
|
||
|
||
// Fetch and parse a podcast RSS feed → returns { title, episodes: [{ id, title, date, audioUrl, duration }] }
|
||
async function parsePodcastRSS(feedUrl, limit = 200) {
|
||
const xml = await fetchUrl(feedUrl);
|
||
|
||
// Extract podcast title
|
||
const titleMatch = xml.match(/<channel>[\s\S]*?<title>(?:<!\[CDATA\[)?(.*?)(?:\]\]>)?<\/title>/);
|
||
const podcastTitle = titleMatch ? titleMatch[1].trim() : "Unknown Podcast";
|
||
|
||
// Extract episodes from <item> elements
|
||
const episodes = [];
|
||
const itemRegex = /<item>([\s\S]*?)<\/item>/g;
|
||
let match;
|
||
while ((match = itemRegex.exec(xml)) !== null && episodes.length < limit) {
|
||
const item = match[1];
|
||
|
||
// GUID (unique episode identifier)
|
||
const guidMatch = item.match(/<guid[^>]*>(?:<!\[CDATA\[)?(.*?)(?:\]\]>)?<\/guid>/);
|
||
const guid = guidMatch ? guidMatch[1].trim() : null;
|
||
|
||
// Title
|
||
const epTitleMatch = item.match(/<title>(?:<!\[CDATA\[)?(.*?)(?:\]\]>)?<\/title>/);
|
||
const epTitle = epTitleMatch ? epTitleMatch[1].trim() : "Untitled Episode";
|
||
|
||
// Publish date
|
||
const pubDateMatch = item.match(/<pubDate>([^<]+)<\/pubDate>/);
|
||
const pubDate = pubDateMatch ? pubDateMatch[1].trim() : null;
|
||
let dateStr = ""; // YYYYMMDD
|
||
if (pubDate) {
|
||
try {
|
||
const d = new Date(pubDate);
|
||
if (!isNaN(d.getTime())) {
|
||
dateStr = d.toISOString().slice(0, 10).replace(/-/g, "");
|
||
}
|
||
} catch {}
|
||
}
|
||
|
||
// Audio enclosure URL
|
||
const enclosureMatch = item.match(/<enclosure[^>]+url=["']([^"']+)["']/);
|
||
const audioUrl = enclosureMatch ? enclosureMatch[1].trim() : null;
|
||
|
||
// Duration (itunes:duration)
|
||
const durMatch = item.match(/<itunes:duration>([^<]+)<\/itunes:duration>/);
|
||
const duration = durMatch ? durMatch[1].trim() : "";
|
||
|
||
if (!audioUrl) continue; // Skip episodes without audio
|
||
|
||
// Use guid, or fall back to audioUrl as unique ID
|
||
const id = guid || audioUrl;
|
||
|
||
episodes.push({ id, title: epTitle, date: dateStr, audioUrl, duration });
|
||
}
|
||
|
||
return { title: podcastTitle, episodes };
|
||
}
|
||
|
||
// Download a podcast episode audio file via HTTP(S) to a local path
|
||
// downloadPodcastAudio moved to ./audio.js
|
||
|
||
// Get channel name from URL
|
||
async function fetchChannelName(url) {
|
||
// Try fast flat-playlist approach first
|
||
try {
|
||
const { stdout } = await execFileAsync("yt-dlp", [
|
||
"--print", "%(channel)s",
|
||
"--no-download",
|
||
"--playlist-end", "1",
|
||
"--flat-playlist",
|
||
...ytCookieArgs(),
|
||
url,
|
||
], { timeout: 15000 });
|
||
const name = stdout.trim().split("\n")[0];
|
||
if (name && name !== "NA") return name;
|
||
} catch {}
|
||
// Fallback: fetch without flat-playlist (slower but gets channel from video metadata)
|
||
try {
|
||
const { stdout } = await execFileAsync("yt-dlp", [
|
||
"--print", "%(channel)s",
|
||
"--no-download",
|
||
"--playlist-end", "1",
|
||
...ytCookieArgs(),
|
||
url,
|
||
], { timeout: 30000 });
|
||
const name = stdout.trim().split("\n")[0];
|
||
if (name && name !== "NA") return name;
|
||
} catch {}
|
||
// Last resort: extract handle from URL
|
||
try {
|
||
const u = new URL(url);
|
||
const handleMatch = u.pathname.match(/\/@([^/]+)/);
|
||
if (handleMatch) return "@" + handleMatch[1];
|
||
} catch {}
|
||
return "Unknown Channel";
|
||
}
|
||
|
||
// Subscription storage (subscriptions / skip / seen / auto-queue), the
|
||
// dedup, scope enumeration, and the file-locking that serializes
|
||
// read-modify-writes all live in ./subscriptions.js now — scope-keyed and
|
||
// unit-tested. getProcessedVideoIds(scope) + isKnownVideo() are the dedup
|
||
// that keeps the auto-queue from re-offering already-handled videos.
|
||
//
|
||
// While subscriptions are operator-only in multi mode, the check loop +
|
||
// background processor operate on this scope. Endpoints use
|
||
// scopeForRequest(req) (= "owner" for the operator). Per-tenant flips these
|
||
// to each user's scope — see docs/per-tenant-subscriptions-plan.md.
|
||
const SUBSCRIPTIONS_SCOPE = "owner";
|
||
|
||
// Prune auto-queue items that have since been summarized (e.g. cleared from
|
||
// the queue then re-discovered by a check). Runs per scope at boot, after
|
||
// the migration. The scope-aware getProcessedVideoIds is what makes this
|
||
// work — it used to scan the wrong directory and prune nothing.
|
||
async function reconcileAutoQueueWithLibrary(scope) {
|
||
const processed = await getProcessedVideoIds(scope);
|
||
let removed = 0;
|
||
await mutateAutoQueue(scope, (items) => {
|
||
const before = items.length;
|
||
const kept = items.filter((q) => !processed.has(q.videoId));
|
||
removed = before - kept.length;
|
||
return kept;
|
||
});
|
||
if (removed > 0) {
|
||
console.log(
|
||
` Auto-queue (${scope}): removed ${removed} already-processed item(s)`,
|
||
);
|
||
}
|
||
}
|
||
|
||
// Which user's identity the processor assumes to run a scope's items over
|
||
// the authenticated loopback /api/process call. Single mode → null (no auth;
|
||
// the loopback resolves to "owner"). Multi mode: "owner" → the operator;
|
||
// a tenant scope IS that user's id.
|
||
function userIdForScope(scope) {
|
||
if (RECAP_MODE !== "multi") return null;
|
||
if (scope === "owner") return adminUserId();
|
||
return scope;
|
||
}
|
||
|
||
// ── Background processing queue ──────────────────────────────────────────
|
||
// Processes "approved" auto-queue items sequentially with configurable delay
|
||
// between items to avoid hammering YouTube with rapid-fire downloads.
|
||
const processingConfigPath = path.join(configDir, "processing-config.json");
|
||
let processingConfig = {
|
||
delaySeconds: 300, // Default delay between processing queue items (5 minutes)
|
||
enabled: true, // Whether background processing is active
|
||
};
|
||
let processingState = {
|
||
running: false, // Is the processor loop currently active?
|
||
currentItem: null, // The item currently being processed (or null)
|
||
lastCompleted: null, // Timestamp of last completed item
|
||
rush: false, // If true, skip delay before next item
|
||
log: [], // Recent processing log entries
|
||
};
|
||
|
||
function procLog(msg) {
|
||
console.log(` [processor] ${msg}`);
|
||
processingState.log.push({ t: new Date().toISOString(), msg });
|
||
if (processingState.log.length > 100) processingState.log.shift();
|
||
}
|
||
|
||
async function loadProcessingConfig() {
|
||
try {
|
||
const data = JSON.parse(await fs.readFile(processingConfigPath, "utf-8"));
|
||
if (typeof data.delaySeconds === "number") processingConfig.delaySeconds = data.delaySeconds;
|
||
if (typeof data.enabled === "boolean") processingConfig.enabled = data.enabled;
|
||
} catch {}
|
||
}
|
||
async function saveProcessingConfig() {
|
||
await fs.writeFile(processingConfigPath, JSON.stringify(processingConfig, null, 2));
|
||
}
|
||
await loadProcessingConfig();
|
||
|
||
// Find the next approved item across every scope's auto-queue. Returns
|
||
// { scope, item } (owner scope first, then tenants) or null when none.
|
||
async function findNextApprovedItem() {
|
||
const scopes = await listAutoQueueScopes();
|
||
for (const scope of scopes) {
|
||
const queue = await loadAutoQueue(scope);
|
||
const item = queue.find((q) => q.status === "approved");
|
||
if (item) return { scope, item };
|
||
}
|
||
return null;
|
||
}
|
||
|
||
// The background processor: picks "approved" items, processes via internal HTTP,
|
||
// waits the configured delay, then picks the next one. Runs continuously.
|
||
async function backgroundProcessor() {
|
||
if (processingState.running) return; // Already running
|
||
processingState.running = true;
|
||
procLog("Background processor started");
|
||
|
||
while (true) {
|
||
// Find the next approved item across every owner's queue (a scope can
|
||
// hold queued items even with no current subscriptions). Each item is
|
||
// processed AS its owner — see processItemInternally(item, scope).
|
||
const next = await findNextApprovedItem();
|
||
if (!next) {
|
||
procLog("No approved items in any queue — processor sleeping");
|
||
processingState.running = false;
|
||
processingState.currentItem = null;
|
||
return;
|
||
}
|
||
const { scope, item } = next;
|
||
|
||
if (!processingConfig.enabled) {
|
||
procLog("Processing is paused — processor sleeping");
|
||
processingState.running = false;
|
||
processingState.currentItem = null;
|
||
return;
|
||
}
|
||
|
||
// Wait the configured delay (unless rush mode or first item)
|
||
if (processingState.lastCompleted && !processingState.rush) {
|
||
const delaySec = processingConfig.delaySeconds;
|
||
procLog(`Waiting ${delaySec}s before next item...`);
|
||
await new Promise(r => setTimeout(r, delaySec * 1000));
|
||
}
|
||
processingState.rush = false;
|
||
|
||
// Atomically claim the item: re-check it's still approved (it may have
|
||
// been removed/changed while we waited) and flip it to "processing".
|
||
let claimed = null;
|
||
await mutateAutoQueue(scope, (items) => {
|
||
const it = items.find((q) => q.id === item.id);
|
||
if (it && it.status === "approved") {
|
||
it.status = "processing";
|
||
claimed = { ...it };
|
||
}
|
||
});
|
||
if (!claimed) {
|
||
procLog(`Item ${item.id} was removed or status changed — skipping`);
|
||
continue;
|
||
}
|
||
processingState.currentItem = claimed;
|
||
procLog(`Processing: ${claimed.title} (${claimed.url})`);
|
||
|
||
let result = null;
|
||
let errMsg = null;
|
||
try {
|
||
// Call /api/process via loopback, AS the scope's owner.
|
||
result = await processItemInternally(claimed, scope);
|
||
processingState.lastCompleted = new Date().toISOString();
|
||
procLog(`✓ Completed: ${claimed.title}`);
|
||
} catch (err) {
|
||
errMsg = err.message || String(err);
|
||
procLog(`✗ Failed: ${claimed.title} — ${errMsg.slice(0, 200)}`);
|
||
}
|
||
|
||
// Write the terminal status back atomically.
|
||
await mutateAutoQueue(scope, (items) => {
|
||
const it = items.find((q) => q.id === item.id);
|
||
if (!it) return;
|
||
if (errMsg) {
|
||
it.status = "failed";
|
||
it.error = errMsg;
|
||
it.failedAt = new Date().toISOString();
|
||
} else {
|
||
it.status = "completed";
|
||
it.completedAt = new Date().toISOString();
|
||
it.historyId = (result && result.historyId) || null;
|
||
}
|
||
});
|
||
processingState.currentItem = null;
|
||
}
|
||
}
|
||
|
||
// Internal HTTP request to /api/process — consumes the SSE stream and
|
||
// waits for the "result" or "error" event. This reuses the entire existing
|
||
// pipeline without any code duplication.
|
||
function processItemInternally(item, scope) {
|
||
return new Promise((resolve, reject) => {
|
||
// Pick a provider for the auto-queue run. Historically this
|
||
// function hard-required a server-side Gemini API key (legacy
|
||
// pre-relay-as-provider design), so a user with the relay set
|
||
// up but no Gemini key would silently fail here — the request
|
||
// never even reached /api/process and the failure only landed
|
||
// in the in-memory processingState.log, invisible to the
|
||
// dashboard. That broke subscription auto-processing for every
|
||
// relay-only user.
|
||
//
|
||
// New behavior: prefer the relay (the modern default for fresh
|
||
// installs and the most common config), fall back to Gemini
|
||
// when only a local key is configured, and fail with a clear,
|
||
// user-visible error when neither is set. The frontend stores
|
||
// the user's actual preference in localStorage which the server
|
||
// can't read — picking based on what's CONFIGURED here mirrors
|
||
// the practical reality (the relay is the only modern option
|
||
// that doesn't require a key) and matches the interactive
|
||
// experience for relay users.
|
||
const relayConfigured = !!relayDefault.getRelayBaseURL();
|
||
const geminiKey = resolveApiKey(null);
|
||
let transcriptionProvider, analysisProvider;
|
||
if (relayConfigured) {
|
||
transcriptionProvider = "relay";
|
||
analysisProvider = "relay";
|
||
} else if (geminiKey) {
|
||
transcriptionProvider = "gemini";
|
||
analysisProvider = "gemini";
|
||
} else {
|
||
return reject(
|
||
new Error(
|
||
"No transcribe/analyze provider is configured. Set the Recap Relay URL in Settings, or paste a Gemini API key.",
|
||
),
|
||
);
|
||
}
|
||
|
||
const body = JSON.stringify({
|
||
url: item.url,
|
||
// Only used by the Gemini provider; harmless for the relay
|
||
// path (relay provider ignores this field).
|
||
apiKey: "USE_SERVER_KEY",
|
||
type: item.type || undefined,
|
||
title: item.title || undefined,
|
||
uploadDate: item.uploadDate || undefined,
|
||
episodeId: item.videoId || undefined,
|
||
transcriptionProvider,
|
||
transcriptionModel:
|
||
transcriptionProvider === "relay" ? "relay-default" : undefined,
|
||
analysisProvider,
|
||
analysisModel:
|
||
analysisProvider === "relay" ? "relay-default" : undefined,
|
||
});
|
||
|
||
// Per-tenant: assume the owning user's identity for this loopback call
|
||
// so /api/process gates, bills, and saves to the right account. Single
|
||
// mode needs no cookie (the loopback resolves to "owner"). A short-lived
|
||
// real session is minted here and deleted on every exit path below.
|
||
const headers = {
|
||
"Content-Type": "application/json",
|
||
"Content-Length": Buffer.byteLength(body),
|
||
};
|
||
let internalSession = null;
|
||
const ownerUserId = userIdForScope(scope);
|
||
if (ownerUserId) {
|
||
try {
|
||
internalSession = mintInternalSession(ownerUserId);
|
||
headers["Cookie"] = `${SESSION_COOKIE}=${internalSession}`;
|
||
} catch (err) {
|
||
return reject(
|
||
new Error(`Could not establish processing identity: ${err?.message || err}`),
|
||
);
|
||
}
|
||
}
|
||
const cleanup = () => deleteInternalSession(internalSession);
|
||
|
||
const req = http.request({
|
||
hostname: "127.0.0.1",
|
||
port: PORT,
|
||
path: "/api/process",
|
||
method: "POST",
|
||
headers,
|
||
timeout: 1800000, // 30 minutes max for very long videos
|
||
}, (res) => {
|
||
let buffer = "";
|
||
let lastResult = null;
|
||
let lastError = null;
|
||
|
||
res.on("data", (chunk) => {
|
||
buffer += chunk.toString();
|
||
|
||
// Parse SSE events from the buffer
|
||
const lines = buffer.split("\n");
|
||
buffer = lines.pop() || ""; // Keep incomplete line in buffer
|
||
|
||
let currentEvent = null;
|
||
for (const line of lines) {
|
||
if (line.startsWith("event: ")) {
|
||
currentEvent = line.slice(7).trim();
|
||
} else if (line.startsWith("data: ") && currentEvent) {
|
||
try {
|
||
const data = JSON.parse(line.slice(6));
|
||
if (currentEvent === "result") lastResult = data;
|
||
if (currentEvent === "error") lastError = data;
|
||
if (currentEvent === "log") {
|
||
procLog(` [${data.elapsed || "?"}s] ${data.message}`);
|
||
}
|
||
} catch {}
|
||
currentEvent = null;
|
||
} else if (line === "") {
|
||
currentEvent = null;
|
||
}
|
||
}
|
||
});
|
||
|
||
res.on("end", () => {
|
||
cleanup();
|
||
if (lastError) {
|
||
reject(new Error(lastError.message || "Processing failed"));
|
||
} else if (lastResult) {
|
||
resolve(lastResult);
|
||
} else {
|
||
reject(new Error("Processing ended without result"));
|
||
}
|
||
});
|
||
|
||
res.on("error", (err) => {
|
||
cleanup();
|
||
reject(err);
|
||
});
|
||
});
|
||
|
||
req.on("error", (err) => {
|
||
cleanup();
|
||
reject(err);
|
||
});
|
||
req.on("timeout", () => {
|
||
req.destroy();
|
||
cleanup();
|
||
reject(new Error("Processing timed out after 30 minutes"));
|
||
});
|
||
|
||
req.write(body);
|
||
req.end();
|
||
});
|
||
}
|
||
|
||
// Wake up the processor whenever there are approved items
|
||
function kickProcessor() {
|
||
// Just start the loop if idle — it loads the queue itself and sleeps
|
||
// immediately if there's nothing approved, so no pre-check needed.
|
||
if (!processingState.running && processingConfig.enabled) {
|
||
backgroundProcessor().catch((err) => procLog(`Processor error: ${err.message}`));
|
||
}
|
||
}
|
||
|
||
let subCheckRunning = false;
|
||
let subCheckPromise = null;
|
||
let subCheckLog = []; // Stores recent check logs for debug endpoint
|
||
function subLog(msg) { console.log(msg); subCheckLog.push({ t: new Date().toISOString(), msg }); if (subCheckLog.length > 200) subCheckLog.shift(); }
|
||
|
||
async function checkSubscriptions() {
|
||
if (subCheckRunning) {
|
||
// Wait for current check to finish, then run again
|
||
if (subCheckPromise) await subCheckPromise;
|
||
return checkSubscriptions();
|
||
}
|
||
subCheckRunning = true;
|
||
subCheckPromise = _checkSubscriptionsInner().finally(() => {
|
||
subCheckRunning = false;
|
||
subCheckPromise = null;
|
||
});
|
||
return subCheckPromise;
|
||
}
|
||
|
||
async function _checkSubscriptionsInner() {
|
||
// Pro-tier feature: skip silently when not entitled. The HTTP gate above
|
||
// returns 402 to callers; this guards the background timer + manual paths.
|
||
if (!licenseMW.LIC.entitlements.has("subscriptions")) {
|
||
subCheckLog = [];
|
||
subLog("Skipped: subscriptions require a Pro license.");
|
||
return;
|
||
}
|
||
subCheckLog = []; // Clear logs for fresh check
|
||
// Operator-only in multi mode, so this resolves to ["owner"]. When
|
||
// per-tenant subscriptions ship it fans out to every subscribing tenant.
|
||
const scopes = await listSubscriptionScopes();
|
||
for (const scope of scopes) {
|
||
try {
|
||
await checkScopeSubscriptions(scope);
|
||
} catch (err) {
|
||
subLog(`⚠ Subscription check failed for scope ${scope}: ${err.message}`);
|
||
}
|
||
}
|
||
}
|
||
|
||
// Discover + queue new videos for ONE scope's subscriptions. New items are
|
||
// collected locally and appended to the scope's auto-queue atomically at the
|
||
// end (mutateAutoQueue) so a concurrent endpoint mutation can't be lost.
|
||
async function checkScopeSubscriptions(scope) {
|
||
const subs = await loadSubscriptions(scope);
|
||
if (subs.length === 0) { return; }
|
||
|
||
const processedIds = await getProcessedVideoIds(scope);
|
||
const skippedIds = await loadSkipList(scope);
|
||
const seenIds = await loadSeenList(scope);
|
||
const queue = await loadAutoQueue(scope);
|
||
// Dedup also against items already in the queue. queuedIds grows as we add
|
||
// within this run so the same video isn't queued twice across subs.
|
||
const queuedIds = new Set(queue.map(q => {
|
||
// For YouTube: extract video ID from URL; for podcasts: use stored videoId (GUID)
|
||
if (q.videoId) return q.videoId;
|
||
const m = q.url.match(/[?&]v=([a-zA-Z0-9_-]{11})/);
|
||
return m ? m[1] : null;
|
||
}).filter(Boolean));
|
||
|
||
// Split queue counts into active (pending/approved/processing — what the
|
||
// UI shows) vs done (completed/failed — terminal, hidden by default). All
|
||
// four buckets feed dedup against newly discovered videos below.
|
||
const qActive = queue.filter(q => ["pending", "approved", "processing"].includes(q.status)).length;
|
||
const qCompleted = queue.filter(q => q.status === "completed").length;
|
||
const qFailed = queue.filter(q => q.status === "failed").length;
|
||
const qDone = qCompleted + qFailed;
|
||
subLog(`[${scope}] ${processedIds.size} in library, ${skippedIds.size} skipped, ${seenIds.size} seen — queue: ${qActive} active, ${qDone} done (${qCompleted} completed, ${qFailed} failed)`);
|
||
if (qFailed > 0) {
|
||
subLog(` (queue has ${qFailed} failed item${qFailed === 1 ? "" : "s"} — open the Queue panel and toggle "Show all" to view + retry)`);
|
||
}
|
||
|
||
let changed = false;
|
||
const newItems = []; // appended to the scope's queue atomically at the end
|
||
|
||
for (const sub of subs) {
|
||
if (sub.paused) { subLog(`⏸ ${sub.name} — paused, skipping`); continue; }
|
||
|
||
try {
|
||
const icon = sub.type === "podcast" ? "🎙" : "📡";
|
||
subLog(`${icon} Checking: ${sub.name} (${sub.url})`);
|
||
|
||
const cutoffDate = sub.createdAt.replace(/[-T:\.Z]/g, "").slice(0, 8);
|
||
|
||
if (sub.type === "podcast") {
|
||
// ── Podcast subscription: discover episodes from RSS feed ──
|
||
const { episodes } = await parsePodcastRSS(sub.url, 200);
|
||
subLog(` Found ${episodes.length} episode(s) in RSS feed`);
|
||
|
||
if (episodes.length === 0) {
|
||
sub.lastChecked = new Date().toISOString();
|
||
changed = true;
|
||
continue;
|
||
}
|
||
|
||
// Filter out already-known episodes (in library / queue / skip / seen)
|
||
const dedupSets = { processedIds, queuedIds, skippedIds, seenIds };
|
||
const unknowns = episodes.filter((ep) => !isKnownVideo(ep.id, dedupSets));
|
||
const filtered = episodes.length - unknowns.length;
|
||
subLog(` ${unknowns.length} to check, ${filtered} already known`);
|
||
|
||
if (unknowns.length === 0) {
|
||
sub.lastChecked = new Date().toISOString();
|
||
changed = true;
|
||
continue;
|
||
}
|
||
|
||
subLog(` Cutoff date: ${cutoffDate}`);
|
||
|
||
let newCount = 0;
|
||
const seenNow = [];
|
||
for (const ep of unknowns) {
|
||
if (!ep.date || ep.date.length !== 8) continue; // skip undated
|
||
if (ep.date < cutoffDate) {
|
||
seenNow.push(ep.id);
|
||
continue; // before cutoff
|
||
}
|
||
subLog(` ✅ ${ep.date} — ${ep.title.slice(0, 60)}`);
|
||
|
||
const itemStatus = sub.autoDownload ? "approved" : "pending";
|
||
newItems.push({
|
||
id: `auto-${Date.now()}-${Buffer.from(ep.id).toString("base64url").slice(0, 16)}`,
|
||
videoId: ep.id, // episode GUID
|
||
url: ep.audioUrl,
|
||
title: ep.title,
|
||
uploadDate: ep.date,
|
||
subscriptionId: sub.id,
|
||
subscriptionName: sub.name,
|
||
status: itemStatus,
|
||
type: "podcast",
|
||
duration: ep.duration || "",
|
||
});
|
||
queuedIds.add(ep.id);
|
||
newCount++;
|
||
}
|
||
|
||
if (seenNow.length > 0) await addToSeenList(scope, seenNow);
|
||
|
||
sub.lastChecked = new Date().toISOString();
|
||
subLog(` → ${newCount} episode(s) queued for approval from ${sub.name}`);
|
||
changed = true;
|
||
|
||
} else {
|
||
// ── YouTube subscription: discover videos via yt-dlp + RSS dates ──
|
||
|
||
// Scale fetch limit based on how far back the subscription date goes
|
||
const daysSinceSub = Math.max(1, Math.ceil((Date.now() - new Date(sub.createdAt).getTime()) / 86400000));
|
||
const estimatedVideos = Math.ceil(daysSinceSub / 7) * 4;
|
||
const fetchLimit = Math.min(Math.max(estimatedVideos, 15), 200);
|
||
subLog(` Subscription age: ${daysSinceSub}d → fetching up to ${fetchLimit} videos`);
|
||
|
||
const candidates = await listChannelVideosFast(sub.url, fetchLimit);
|
||
subLog(` Found ${candidates.length} recent video(s)`);
|
||
|
||
if (candidates.length === 0) {
|
||
subLog(` ⚠ No videos returned from yt-dlp for this channel`);
|
||
sub.lastChecked = new Date().toISOString();
|
||
changed = true;
|
||
continue;
|
||
}
|
||
|
||
const unknowns = candidates.filter(
|
||
(v) => !isKnownVideo(v.id, { processedIds, queuedIds, skippedIds, seenIds }),
|
||
);
|
||
const filtered = candidates.length - unknowns.length;
|
||
subLog(` ${unknowns.length} to check, ${filtered} already known`);
|
||
|
||
if (unknowns.length === 0) {
|
||
sub.lastChecked = new Date().toISOString();
|
||
changed = true;
|
||
continue;
|
||
}
|
||
|
||
subLog(` Cutoff date: ${cutoffDate} — fetching upload dates...`);
|
||
|
||
// Step 1: Get channel ID (cached on subscription, or fetch once)
|
||
if (!sub.channelId) {
|
||
subLog(` Resolving channel ID...`);
|
||
sub.channelId = await getChannelId(sub.url);
|
||
if (sub.channelId) {
|
||
subLog(` Resolved channel ID: ${sub.channelId}`);
|
||
changed = true;
|
||
} else {
|
||
subLog(` ⚠ Could not resolve channel ID — RSS will be skipped`);
|
||
}
|
||
}
|
||
|
||
// Step 2: RSS feed — fast, no bot detection, covers ~15 most recent
|
||
const dateMap = sub.channelId ? await fetchDatesFromRSS(sub.channelId) : {};
|
||
const rssCount = Object.keys(dateMap).length;
|
||
if (sub.channelId) subLog(` RSS feed: got dates for ${rssCount} videos`);
|
||
|
||
// Step 3: Smart early termination
|
||
const needDates = unknowns.filter(v => !dateMap[v.id]);
|
||
let earlyTermination = false;
|
||
if (needDates.length > 0 && rssCount > 0) {
|
||
const rssDates = Object.values(dateMap).sort();
|
||
const oldestRssDate = rssDates[0];
|
||
if (oldestRssDate < cutoffDate) {
|
||
earlyTermination = true;
|
||
subLog(` Oldest RSS video (${oldestRssDate}) is before cutoff — ${needDates.length} older video(s) are definitely too old, skipping yt-dlp`);
|
||
} else {
|
||
subLog(` ${needDates.length} video(s) not in RSS — trying yt-dlp for dates...`);
|
||
const ytDates = await fetchUploadDates(needDates.map(v => v.id));
|
||
Object.assign(dateMap, ytDates);
|
||
const ytCount = Object.keys(ytDates).length;
|
||
if (ytCount > 0) subLog(` yt-dlp added dates for ${ytCount} more video(s)`);
|
||
}
|
||
} else if (needDates.length > 0 && rssCount === 0) {
|
||
subLog(` ${needDates.length} video(s) need dates — trying yt-dlp...`);
|
||
const ytDates = await fetchUploadDates(needDates.map(v => v.id));
|
||
Object.assign(dateMap, ytDates);
|
||
const ytCount = Object.keys(ytDates).length;
|
||
if (ytCount > 0) subLog(` yt-dlp added dates for ${ytCount} more video(s)`);
|
||
}
|
||
|
||
const gotDates = Object.keys(dateMap).length;
|
||
if (gotDates > 0 || needDates.length === 0) {
|
||
subLog(` Total dates: ${gotDates} of ${unknowns.length} videos`);
|
||
} else {
|
||
subLog(` ⚠ No dates available — skipping. Try setting YT_COOKIES_FROM in .env`);
|
||
sub.lastChecked = new Date().toISOString();
|
||
changed = true;
|
||
continue;
|
||
}
|
||
|
||
let newCount = 0;
|
||
for (const video of unknowns) {
|
||
const uploadDate = dateMap[video.id];
|
||
if (!uploadDate || uploadDate.length !== 8) {
|
||
continue;
|
||
}
|
||
if (uploadDate < cutoffDate) {
|
||
subLog(` ⏭ ${video.id} (${uploadDate}) — before cutoff`);
|
||
continue;
|
||
}
|
||
subLog(` ✅ ${video.id}${uploadDate ? ` (${uploadDate})` : ""} — ${video.title.slice(0,50)}`);
|
||
|
||
const ytItemStatus = sub.autoDownload ? "approved" : "pending";
|
||
newItems.push({
|
||
id: `auto-${Date.now()}-${video.id}`,
|
||
videoId: video.id,
|
||
url: `https://www.youtube.com/watch?v=${video.id}`,
|
||
title: video.title,
|
||
uploadDate: uploadDate || null,
|
||
subscriptionId: sub.id,
|
||
subscriptionName: sub.name,
|
||
status: ytItemStatus,
|
||
});
|
||
queuedIds.add(video.id);
|
||
newCount++;
|
||
}
|
||
|
||
// Only add to seen list when we can PROVE a video is too old
|
||
const seenNow = unknowns.filter(v => {
|
||
if (queuedIds.has(v.id)) return false;
|
||
const d = dateMap[v.id];
|
||
if (d && d.length === 8 && d < cutoffDate) return true;
|
||
if (d && d.length === 8 && d >= cutoffDate) return false;
|
||
return earlyTermination;
|
||
}).map(v => v.id);
|
||
if (seenNow.length > 0) await addToSeenList(scope, seenNow);
|
||
|
||
sub.lastChecked = new Date().toISOString();
|
||
subLog(` → ${newCount} video(s) queued for approval from ${sub.name}`);
|
||
changed = true;
|
||
}
|
||
} catch (err) {
|
||
subLog(` ⚠ FAILED for ${sub.name}: ${err.message}`);
|
||
}
|
||
}
|
||
|
||
if (changed) await saveSubscriptions(scope, subs);
|
||
// Append everything discovered this run in one atomic mutation.
|
||
if (newItems.length > 0) {
|
||
await mutateAutoQueue(scope, (items) => {
|
||
items.push(...newItems);
|
||
});
|
||
}
|
||
|
||
const finalQueue = await loadAutoQueue(scope);
|
||
const pendingCount = finalQueue.filter(q => q.status === "pending").length;
|
||
const approvedCount = finalQueue.filter(q => q.status === "approved").length;
|
||
subLog(`[${scope}] Done. ${pendingCount} pending, ${approvedCount} approved in auto-queue.`);
|
||
|
||
// Wake up the background processor if there are approved items. (Behind
|
||
// the operator-only gate the processor works the "owner" queue, which is
|
||
// the only scope that reaches here.)
|
||
if (approvedCount > 0) {
|
||
subLog(`Kicking background processor for ${approvedCount} approved item(s)...`);
|
||
kickProcessor();
|
||
}
|
||
}
|
||
|
||
// Scope for a subscription / auto-queue request. Behind the operator-only
|
||
// gate (license-middleware) this is always "owner" (admin in multi mode, or
|
||
// single mode). Forward-compatible with per-tenant — it's just the signed-in
|
||
// user's scope. Falls back to "owner" if scope resolution somehow fails.
|
||
function subScope(req) {
|
||
try {
|
||
return scopeForRequest(req);
|
||
} catch {
|
||
return SUBSCRIPTIONS_SCOPE;
|
||
}
|
||
}
|
||
|
||
// CRUD endpoints
|
||
app.get("/api/subscriptions", async (req, res) => {
|
||
const subs = await loadSubscriptions(subScope(req));
|
||
res.json({ subscriptions: subs });
|
||
});
|
||
|
||
// Extract a normalized channel key from a YouTube URL for dedup
|
||
function channelKeyFromUrl(url) {
|
||
try {
|
||
const u = new URL(url);
|
||
const path = u.pathname.toLowerCase().replace(/\/+$/, "");
|
||
// /@handle/videos → @handle, /@handle/streams → @handle
|
||
const handleMatch = path.match(/\/(@[^/]+)/);
|
||
if (handleMatch) return handleMatch[1];
|
||
// /channel/UCxxx → channel/UCxxx
|
||
const channelMatch = path.match(/\/(channel\/[^/]+)/);
|
||
if (channelMatch) return channelMatch[1];
|
||
// /c/name or /user/name
|
||
const cMatch = path.match(/\/(c|user)\/([^/]+)/);
|
||
if (cMatch) return cMatch[0];
|
||
// /playlist?list=PLxxx
|
||
const list = u.searchParams.get("list");
|
||
if (list) return `playlist/${list}`;
|
||
return path; // fallback
|
||
} catch {
|
||
return url.toLowerCase().replace(/\/+$/, "");
|
||
}
|
||
}
|
||
|
||
app.post("/api/subscriptions", async (req, res) => {
|
||
let { url, since, type, autoDownload } = req.body;
|
||
if (!url) return res.status(400).json({ error: "Missing url" });
|
||
|
||
// Apple Podcasts SHOW URLs (`/podcast/<slug>/id<num>` with NO
|
||
// `?i=<id>` episode param) get resolved to the show's RSS feed
|
||
// before we persist the subscription. The feed URL is what the
|
||
// poll loop fetches to discover new episodes, so storing the
|
||
// landing-page URL would be a dead end.
|
||
if (isAppleShowUrl(url)) {
|
||
try {
|
||
const feed = await resolveAppleShowToFeed(url);
|
||
if (!feed) {
|
||
return res.status(400).json({
|
||
error: "apple_show_no_feed",
|
||
message:
|
||
"Apple doesn't list an RSS feed for that show — it may be an Apple Podcasts Subscriptions exclusive. Paste the show's direct RSS URL instead.",
|
||
});
|
||
}
|
||
console.log(`[subscribe] Apple show → resolved feed: ${feed}`);
|
||
url = feed;
|
||
type = "podcast";
|
||
} catch (err) {
|
||
return res.status(502).json({
|
||
error: "apple_show_lookup_failed",
|
||
message: `Couldn't reach Apple to resolve that show URL: ${err?.message || err}`,
|
||
});
|
||
}
|
||
}
|
||
|
||
const isPodcast = type === "podcast" || isPodcastFeedUrl(url);
|
||
|
||
const scope = subScope(req);
|
||
const subs = await loadSubscriptions(scope);
|
||
|
||
// Prevent duplicates
|
||
if (isPodcast) {
|
||
const normalizedUrl = url.trim().toLowerCase().replace(/\/+$/, "");
|
||
if (subs.find(s => s.url.trim().toLowerCase().replace(/\/+$/, "") === normalizedUrl)) {
|
||
return res.status(409).json({ error: "Already subscribed to this podcast" });
|
||
}
|
||
} else {
|
||
const newKey = channelKeyFromUrl(url);
|
||
if (subs.find(s => channelKeyFromUrl(s.url) === newKey)) {
|
||
return res.status(409).json({ error: "Already subscribed to this channel" });
|
||
}
|
||
}
|
||
|
||
// Use provided cutoff date, or default to right now
|
||
const cutoff = since ? new Date(since).toISOString() : new Date().toISOString();
|
||
const sub = {
|
||
id: `sub-${Date.now()}`,
|
||
url,
|
||
name: "Loading...",
|
||
type: isPodcast ? "podcast" : "youtube",
|
||
channelId: null,
|
||
createdAt: cutoff,
|
||
lastChecked: null,
|
||
paused: false,
|
||
autoDownload: autoDownload === true,
|
||
};
|
||
subs.push(sub);
|
||
await saveSubscriptions(scope, subs);
|
||
|
||
// Respond immediately so the UI isn't blocked
|
||
res.json(sub);
|
||
|
||
// Background: resolve name and check for new content
|
||
(async () => {
|
||
try {
|
||
if (isPodcast) {
|
||
// Fetch podcast title from RSS feed
|
||
const { title } = await parsePodcastRSS(url, 1);
|
||
const freshSubs = await loadSubscriptions(scope);
|
||
const s = freshSubs.find(x => x.id === sub.id);
|
||
if (s) {
|
||
s.name = title || url;
|
||
await saveSubscriptions(scope, freshSubs);
|
||
}
|
||
console.log(` 🎙 New podcast subscription: ${title} — checking for episodes...`);
|
||
} else {
|
||
const [name, channelId] = await Promise.all([
|
||
fetchChannelName(url),
|
||
getChannelId(url),
|
||
]);
|
||
const freshSubs = await loadSubscriptions(scope);
|
||
const s = freshSubs.find(x => x.id === sub.id);
|
||
if (s) {
|
||
s.name = name;
|
||
if (channelId) s.channelId = channelId;
|
||
await saveSubscriptions(scope, freshSubs);
|
||
}
|
||
console.log(` 📡 New subscription: ${name} — checking for recent videos...`);
|
||
}
|
||
await checkSubscriptions();
|
||
} catch (err) {
|
||
console.error(" ⚠ Background subscription setup error:", err.message);
|
||
}
|
||
})();
|
||
});
|
||
|
||
app.delete("/api/subscriptions/:id", async (req, res) => {
|
||
const scope = subScope(req);
|
||
let subs = await loadSubscriptions(scope);
|
||
subs = subs.filter(s => s.id !== req.params.id);
|
||
await saveSubscriptions(scope, subs);
|
||
// Also remove any auto-queue items from this subscription.
|
||
await mutateAutoQueue(scope, (items) =>
|
||
items.filter((q) => q.subscriptionId !== req.params.id),
|
||
);
|
||
res.json({ ok: true });
|
||
});
|
||
|
||
app.put("/api/subscriptions/:id/pause", async (req, res) => {
|
||
const scope = subScope(req);
|
||
const subs = await loadSubscriptions(scope);
|
||
const sub = subs.find(s => s.id === req.params.id);
|
||
if (!sub) return res.status(404).json({ error: "Subscription not found" });
|
||
sub.paused = !sub.paused;
|
||
await saveSubscriptions(scope, subs);
|
||
res.json(sub);
|
||
});
|
||
|
||
app.put("/api/subscriptions/:id/since", async (req, res) => {
|
||
const { since } = req.body;
|
||
if (!since) return res.status(400).json({ error: "Missing 'since' date" });
|
||
const scope = subScope(req);
|
||
const subs = await loadSubscriptions(scope);
|
||
const sub = subs.find(s => s.id === req.params.id);
|
||
if (!sub) return res.status(404).json({ error: "Subscription not found" });
|
||
sub.createdAt = new Date(since).toISOString();
|
||
await saveSubscriptions(scope, subs);
|
||
res.json(sub);
|
||
});
|
||
|
||
// Debug: subscription check logs (viewable in-app)
|
||
app.get("/api/sub-check-log", async (req, res) => {
|
||
const q = await loadAutoQueue(subScope(req));
|
||
res.json({
|
||
log: subCheckLog,
|
||
autoQueueCount: q.length,
|
||
autoQueue: q.map((x) => ({ id: x.id, videoId: x.videoId, title: x.title, status: x.status, sub: x.subscriptionName })),
|
||
});
|
||
});
|
||
|
||
// Auto-queue endpoints (frontend polls these)
|
||
app.get("/api/auto-queue", async (req, res) => {
|
||
const queue = await loadAutoQueue(subScope(req));
|
||
// Return all items grouped by status for the frontend
|
||
const showAll = req.query.all === "true";
|
||
const items = showAll
|
||
? queue
|
||
: queue.filter(q => ["pending", "approved", "processing"].includes(q.status));
|
||
res.json({
|
||
items,
|
||
checkRunning: subCheckRunning,
|
||
counts: {
|
||
pending: queue.filter(q => q.status === "pending").length,
|
||
approved: queue.filter(q => q.status === "approved").length,
|
||
processing: queue.filter(q => q.status === "processing").length,
|
||
completed: queue.filter(q => q.status === "completed").length,
|
||
failed: queue.filter(q => q.status === "failed").length,
|
||
},
|
||
});
|
||
});
|
||
|
||
app.delete("/api/auto-queue/:id", async (req, res) => {
|
||
await mutateAutoQueue(subScope(req), (items) =>
|
||
items.filter((q) => q.id !== req.params.id),
|
||
);
|
||
res.json({ ok: true });
|
||
});
|
||
|
||
app.post("/api/auto-queue/:id/skip", async (req, res) => {
|
||
const scope = subScope(req);
|
||
let videoId = null;
|
||
await mutateAutoQueue(scope, (items) => {
|
||
const it = items.find((q) => q.id === req.params.id);
|
||
if (it && it.videoId) videoId = it.videoId;
|
||
return items.filter((q) => q.id !== req.params.id);
|
||
});
|
||
if (videoId) await addToSkipList(scope, videoId);
|
||
res.json({ ok: true });
|
||
});
|
||
|
||
// Approve a single auto-queue item for background processing
|
||
app.post("/api/auto-queue/:id/approve", async (req, res) => {
|
||
let item = null;
|
||
let badStatus = null;
|
||
await mutateAutoQueue(subScope(req), (items) => {
|
||
const it = items.find((q) => q.id === req.params.id);
|
||
if (!it) return;
|
||
if (it.status !== "pending") { badStatus = it.status; return; }
|
||
it.status = "approved";
|
||
item = { ...it };
|
||
});
|
||
if (badStatus) return res.status(400).json({ error: `Cannot approve item with status '${badStatus}'` });
|
||
if (!item) return res.status(404).json({ error: "Item not found" });
|
||
kickProcessor();
|
||
res.json({ ok: true, item });
|
||
});
|
||
|
||
// Approve all pending items at once
|
||
app.post("/api/auto-queue/approve-all", async (req, res) => {
|
||
let count = 0;
|
||
await mutateAutoQueue(subScope(req), (items) => {
|
||
for (const it of items) {
|
||
if (it.status === "pending") { it.status = "approved"; count++; }
|
||
}
|
||
});
|
||
if (count > 0) kickProcessor();
|
||
res.json({ ok: true, approved: count });
|
||
});
|
||
|
||
// Retry a failed item
|
||
app.post("/api/auto-queue/:id/retry", async (req, res) => {
|
||
let item = null;
|
||
let badStatus = null;
|
||
await mutateAutoQueue(subScope(req), (items) => {
|
||
const it = items.find((q) => q.id === req.params.id);
|
||
if (!it) return;
|
||
if (it.status !== "failed") { badStatus = it.status; return; }
|
||
it.status = "approved";
|
||
it.error = undefined;
|
||
it.failedAt = undefined;
|
||
item = { ...it };
|
||
});
|
||
if (badStatus) return res.status(400).json({ error: `Cannot retry item with status '${badStatus}'` });
|
||
if (!item) return res.status(404).json({ error: "Item not found" });
|
||
kickProcessor();
|
||
res.json({ ok: true, item });
|
||
});
|
||
|
||
// Clear completed and failed items from the queue
|
||
app.post("/api/auto-queue/clear-finished", async (req, res) => {
|
||
let removed = 0;
|
||
await mutateAutoQueue(subScope(req), (items) => {
|
||
const before = items.length;
|
||
const kept = items.filter((q) => !["completed", "failed"].includes(q.status));
|
||
removed = before - kept.length;
|
||
return kept;
|
||
});
|
||
res.json({ ok: true, removed });
|
||
});
|
||
|
||
app.post("/api/subscriptions/check-now", async (req, res) => {
|
||
if (subCheckRunning) return res.json({ ok: true, message: "Already checking" });
|
||
checkSubscriptions().catch(err => console.error(" ⚠ Manual subscription check error:", err.message));
|
||
res.json({ ok: true, message: "Check started" });
|
||
});
|
||
|
||
// ── Auto-download toggle per subscription ──────────────────────────────────
|
||
app.put("/api/subscriptions/:id/auto-download", async (req, res) => {
|
||
const scope = subScope(req);
|
||
const subs = await loadSubscriptions(scope);
|
||
const sub = subs.find(s => s.id === req.params.id);
|
||
if (!sub) return res.status(404).json({ error: "Subscription not found" });
|
||
sub.autoDownload = req.body.enabled === true;
|
||
await saveSubscriptions(scope, subs);
|
||
res.json({ ok: true, subscription: sub });
|
||
});
|
||
|
||
// ── Background processing status & configuration ───────────────────────────
|
||
|
||
// Get current processing status (what's running, queue depth, config)
|
||
app.get("/api/processing/status", async (req, res) => {
|
||
const queue = await loadAutoQueue(subScope(req));
|
||
res.json({
|
||
running: processingState.running,
|
||
currentItem: processingState.currentItem
|
||
? { id: processingState.currentItem.id, title: processingState.currentItem.title, url: processingState.currentItem.url }
|
||
: null,
|
||
lastCompleted: processingState.lastCompleted,
|
||
config: processingConfig,
|
||
counts: {
|
||
approved: queue.filter(q => q.status === "approved").length,
|
||
processing: queue.filter(q => q.status === "processing").length,
|
||
pending: queue.filter(q => q.status === "pending").length,
|
||
},
|
||
log: processingState.log.slice(-20),
|
||
});
|
||
});
|
||
|
||
// Update processing configuration (delay between items, enable/disable)
|
||
app.put("/api/processing/config", async (req, res) => {
|
||
if (typeof req.body.delaySeconds === "number" && req.body.delaySeconds >= 0) {
|
||
processingConfig.delaySeconds = Math.max(0, Math.min(3600, req.body.delaySeconds));
|
||
}
|
||
if (typeof req.body.enabled === "boolean") {
|
||
processingConfig.enabled = req.body.enabled;
|
||
// If enabling and there are approved items, kick the processor
|
||
if (req.body.enabled) kickProcessor();
|
||
}
|
||
await saveProcessingConfig();
|
||
res.json({ ok: true, config: processingConfig });
|
||
});
|
||
|
||
// Rush: skip the delay before the next item in the queue
|
||
app.post("/api/processing/rush", (req, res) => {
|
||
processingState.rush = true;
|
||
// If not currently running but there are approved items, start processing
|
||
kickProcessor();
|
||
res.json({ ok: true, message: "Rush mode enabled — next item will process immediately" });
|
||
});
|
||
|
||
// Processing log (for debug/monitoring)
|
||
app.get("/api/processing/log", (req, res) => {
|
||
res.json({ log: processingState.log });
|
||
});
|
||
|
||
// ── Full pipeline: URL → audio → transcript → topic analysis ──────────────
|
||
|
||
// Single yt-dlp call that returns the full info-dict as JSON. One shot
|
||
// gives us title, channel, description, chapters, duration, and the
|
||
// per-language caption availability — used both for richer transcription
|
||
// prompts and for the captions-first path (skip audio + transcription
|
||
// entirely when YouTube has usable captions for this video).
|
||
async function fetchYouTubeMetadata(videoId) {
|
||
try {
|
||
const { stdout } = await execFileAsync(
|
||
"yt-dlp",
|
||
["-j", "--no-warnings", "--skip-download", `https://www.youtube.com/watch?v=${videoId}`],
|
||
{ timeout: 30000, maxBuffer: 10 * 1024 * 1024 }
|
||
);
|
||
const info = JSON.parse(stdout);
|
||
return {
|
||
title: info.title || "",
|
||
uploadDate: info.upload_date || "",
|
||
channel: info.channel || info.uploader || "",
|
||
// Truncate to keep prompt size sane — descriptions can be huge
|
||
// (release-notes-stuffed Lex Fridman podcasts run thousands of chars).
|
||
description: (info.description || "").slice(0, 2000),
|
||
chapters: Array.isArray(info.chapters) ? info.chapters : [],
|
||
duration: typeof info.duration === "number" ? info.duration : 0,
|
||
hasManualCaptions: !!(info.subtitles && Object.keys(info.subtitles).length > 0),
|
||
hasAutoCaptions: !!(info.automatic_captions && Object.keys(info.automatic_captions).length > 0),
|
||
};
|
||
} catch {
|
||
return null;
|
||
}
|
||
}
|
||
|
||
// Pulls YouTube captions for the given video and parses them into the
|
||
// app's standard { offset, text, duration } entries shape. Prefers
|
||
// manual captions over auto-generated. Returns null if no captions are
|
||
// available or parsing produces too few segments to be useful.
|
||
//
|
||
// Output is the same shape that `parseTimestampedTranscript()` produces
|
||
// from a Gemini transcription, so downstream code (the analysis step,
|
||
// the chunk renderer) doesn't care how the transcript got here.
|
||
async function tryFetchYouTubeCaptions(videoId, tmpDir, opts, log) {
|
||
const { hasManual, hasAuto } = opts;
|
||
if (!hasManual && !hasAuto) return null;
|
||
const url = `https://www.youtube.com/watch?v=${videoId}`;
|
||
const captionsBase = path.join(tmpDir, "captions");
|
||
|
||
// Try manual first (cleaner — punctuated, no fragmented words);
|
||
// fall back to auto-generated if no manual subs.
|
||
const langs = "en.*,en";
|
||
try {
|
||
if (hasManual) {
|
||
await execFileAsync(
|
||
"yt-dlp",
|
||
["--skip-download", "--write-subs", "--sub-langs", langs,
|
||
"--sub-format", "json3", "--no-warnings", "-o", captionsBase, url],
|
||
{ timeout: 30000 }
|
||
);
|
||
} else {
|
||
await execFileAsync(
|
||
"yt-dlp",
|
||
["--skip-download", "--write-auto-subs", "--sub-langs", langs,
|
||
"--sub-format", "json3", "--no-warnings", "-o", captionsBase, url],
|
||
{ timeout: 30000 }
|
||
);
|
||
}
|
||
} catch (err) {
|
||
log(1, `⚠ Caption download failed: ${(err.message || "").slice(0, 200)}`);
|
||
return null;
|
||
}
|
||
|
||
// yt-dlp names the file like `captions.en.json3` or `captions.en-US.json3`.
|
||
// Some videos report subtitles as "available" via the metadata API but
|
||
// yt-dlp can't actually pull a usable JSON3 file (age-gate, members-only,
|
||
// wrong language match against our `en.*,en` pattern, etc.). Surface
|
||
// exactly which failure mode hit so the operator can decide whether
|
||
// to disable captions toggle for this video or pursue a workaround.
|
||
let captionPath = null;
|
||
let availableFiles = [];
|
||
try {
|
||
availableFiles = await fs.readdir(tmpDir);
|
||
const match = availableFiles.find(
|
||
(f) => f.startsWith("captions.") && f.endsWith(".json3")
|
||
);
|
||
if (match) captionPath = path.join(tmpDir, match);
|
||
} catch (err) {
|
||
log(1, `⚠ Captions tmpDir read failed: ${(err.message || "").slice(0, 200)}`);
|
||
}
|
||
if (!captionPath) {
|
||
// Did yt-dlp produce ANY caption file in another format?
|
||
const otherCaps = availableFiles.filter((f) => f.startsWith("captions."));
|
||
if (otherCaps.length > 0) {
|
||
log(
|
||
1,
|
||
`⚠ Captions: yt-dlp wrote ${otherCaps.join(", ")} but not the expected .json3 — likely subtitle-language mismatch (we asked for en.*,en). Video may only have captions in another language.`
|
||
);
|
||
} else {
|
||
log(
|
||
1,
|
||
`⚠ Captions: yt-dlp produced no caption file despite metadata reporting them available. Possibly age-gated, members-only, or yt-dlp couldn't decrypt them.`
|
||
);
|
||
}
|
||
return null;
|
||
}
|
||
|
||
let parsed;
|
||
try {
|
||
const content = await fs.readFile(captionPath, "utf-8");
|
||
parsed = JSON.parse(content);
|
||
} catch (err) {
|
||
log(1, `⚠ Caption parse failed: ${(err.message || "").slice(0, 200)}`);
|
||
return null;
|
||
}
|
||
|
||
const rawEvents = Array.isArray(parsed.events) ? parsed.events : [];
|
||
const entries = [];
|
||
for (const evt of rawEvents) {
|
||
const text = (Array.isArray(evt.segs) ? evt.segs : [])
|
||
.map((s) => s.utf8 || "")
|
||
.join("")
|
||
.replace(/\n+/g, " ")
|
||
.trim();
|
||
if (!text) continue;
|
||
entries.push({
|
||
offset: (evt.tStartMs || 0) / 1000,
|
||
text,
|
||
duration: (evt.dDurationMs || 0) / 1000,
|
||
});
|
||
}
|
||
if (entries.length < 5) {
|
||
log(
|
||
1,
|
||
`⚠ Captions parsed but only ${entries.length} non-empty entries (raw events: ${rawEvents.length}). Captions look empty/malformed — falling back.`
|
||
);
|
||
return null;
|
||
}
|
||
return { entries, source: hasManual ? "manual" : "auto" };
|
||
}
|
||
|
||
// Coalesce a list of timestamped entries into coarser chunks. Used
|
||
// after pulling auto-captions (which are typically 1–3-word fragments
|
||
// every 1–3 seconds) so the downstream analysis prompt has ~100
|
||
// medium-sized segments instead of ~900 tiny ones. Manual captions and
|
||
// Gemini-transcribed entries already average 10–30s and pass through
|
||
// unchanged when their median duration exceeds the threshold.
|
||
function coalesceTranscriptEntries(entries, targetSeconds = 15) {
|
||
if (!Array.isArray(entries) || entries.length === 0) return entries;
|
||
// Sample median duration; only coalesce if entries are too fine.
|
||
const durations = entries.map((e) => e.duration || 0).sort((a, b) => a - b);
|
||
const median = durations[Math.floor(durations.length / 2)];
|
||
if (median >= 8) return entries; // already coarse enough
|
||
|
||
const merged = [];
|
||
let current = null;
|
||
for (const e of entries) {
|
||
const text = (e.text || "").trim();
|
||
if (!text) continue;
|
||
if (!current) {
|
||
current = {
|
||
offset: e.offset,
|
||
text,
|
||
duration: e.duration || 0,
|
||
};
|
||
continue;
|
||
}
|
||
const elapsedFromStart = (e.offset || 0) - current.offset;
|
||
if (elapsedFromStart >= targetSeconds) {
|
||
merged.push(current);
|
||
current = {
|
||
offset: e.offset,
|
||
text,
|
||
duration: e.duration || 0,
|
||
};
|
||
} else {
|
||
current.text = `${current.text} ${text}`.replace(/\s+/g, " ").trim();
|
||
current.duration = ((e.offset || 0) + (e.duration || 0)) - current.offset;
|
||
}
|
||
}
|
||
if (current) merged.push(current);
|
||
return merged;
|
||
}
|
||
|
||
// Reduce an over-long transcript down to roughly `maxEntries` segments
|
||
// so the analysis prompt fits in smaller model contexts (typical local
|
||
// models are 16k–32k tokens — a 2-hour podcast at Parakeet's ~5s
|
||
// granularity easily blows past that). Bucket size is computed from
|
||
// total audio duration so we end up with at most `maxEntries` segments
|
||
// regardless of source granularity. Unlike coalesceTranscriptEntries
|
||
// above, this one is purely count-driven — no median-duration guard,
|
||
// because the failure mode is "context exceeded" not "bad UX from
|
||
// over-fragmented entries".
|
||
//
|
||
// Returns { coalesced, indexMap } where indexMap[i] = { startOrig,
|
||
// endOrig } maps coalesced-entry i back to a range of original-entry
|
||
// indices. The caller uses this to translate section start/end indices
|
||
// returned by the analyzer back into the original entries array, so
|
||
// the final transcript displayed to the user keeps its full granularity
|
||
// — only the analyzer sees the coarser view.
|
||
//
|
||
// FLAGGED TO WATCH: ship 0.2.28. Auto-coalesce may degrade analysis
|
||
// quality on borderline content (the LLM sees fewer, longer segments
|
||
// — section boundaries get blurrier). If users report missed topics or
|
||
// imprecise section starts on long content, the alternative is real
|
||
// chunked analysis (split into overlapping windows, analyze each,
|
||
// stitch sections at boundaries) — significantly more involved.
|
||
function coalesceForAnalysis(entries, maxEntries = 400) {
|
||
if (!Array.isArray(entries) || entries.length <= maxEntries) {
|
||
return { coalesced: entries, indexMap: null };
|
||
}
|
||
const lastEntry = entries[entries.length - 1];
|
||
const totalDuration = (lastEntry.offset || 0) + (lastEntry.duration || 0);
|
||
if (totalDuration <= 0) {
|
||
return { coalesced: entries, indexMap: null };
|
||
}
|
||
// Bucket size: roughly total / maxEntries, but never tighter than 15s
|
||
// (a typical "natural pause" gap — going much smaller would defeat
|
||
// the point of coalescing).
|
||
const targetSeconds = Math.max(15, Math.ceil(totalDuration / maxEntries));
|
||
|
||
const merged = [];
|
||
const indexMap = [];
|
||
let current = null;
|
||
let currentStartOrig = 0;
|
||
for (let i = 0; i < entries.length; i++) {
|
||
const e = entries[i];
|
||
const text = (e.text || "").trim();
|
||
if (!text) continue;
|
||
if (!current) {
|
||
current = { offset: e.offset, text, duration: e.duration || 0 };
|
||
currentStartOrig = i;
|
||
continue;
|
||
}
|
||
const elapsedFromStart = (e.offset || 0) - current.offset;
|
||
if (elapsedFromStart >= targetSeconds) {
|
||
merged.push(current);
|
||
indexMap.push({ startOrig: currentStartOrig, endOrig: i - 1 });
|
||
current = { offset: e.offset, text, duration: e.duration || 0 };
|
||
currentStartOrig = i;
|
||
} else {
|
||
current.text = `${current.text} ${text}`.replace(/\s+/g, " ").trim();
|
||
current.duration = ((e.offset || 0) + (e.duration || 0)) - current.offset;
|
||
}
|
||
}
|
||
if (current) {
|
||
merged.push(current);
|
||
indexMap.push({ startOrig: currentStartOrig, endOrig: entries.length - 1 });
|
||
}
|
||
return { coalesced: merged, indexMap };
|
||
}
|
||
|
||
// ── In-flight free-tier job status + cancel ─────────────────────────────────
|
||
// Lets the web UI render a "Currently processing X — Cancel" banner after
|
||
// a browser refresh, when the SSE stream from the original /api/process
|
||
// call is no longer attached. Only the free-tier slot is tracked today;
|
||
// paid-tier batch queueing happens client-side.
|
||
app.get("/api/process/current", (req, res) => {
|
||
// ?logs=1 returns the in-flight job's accumulated log buffer so a
|
||
// browser refresh mid-pipeline can repopulate the activity log
|
||
// instead of starting blank. Default is the lightweight header-only
|
||
// shape used by the 5s banner poll.
|
||
const includeLogs = req.query.logs === "1" || req.query.logs === "true";
|
||
res.json({ job: getCurrentFreeJob({ includeLogs }) });
|
||
});
|
||
|
||
app.post("/api/process/cancel", (_req, res) => {
|
||
const had = abortCurrentFreeJob();
|
||
if (!had) return res.json({ ok: true, cancelled: false });
|
||
// We don't kill the in-flight yt-dlp / model API call here — the
|
||
// pipeline polls isFreeJobAborted() at each major step and throws
|
||
// early, which lands in the request handler's finally block where the
|
||
// slot is released. So cancellation latency is bounded by the time
|
||
// until the next checkpoint (a few seconds in practice, up to the
|
||
// length of one outstanding model call).
|
||
res.json({ ok: true, cancelled: true });
|
||
});
|
||
|
||
// ── Auto-discovery of provider connection info ──────────────────────────────
|
||
// The picker UI hits this on boot to pre-fill placeholders for providers
|
||
// that have a server-detectable default — most notably Ollama on
|
||
// StartOS, reachable at the documented `http://<package-id>.startos:<port>`
|
||
// internal hostname when installed alongside Recap.
|
||
app.get("/api/providers/discover", async (_req, res) => {
|
||
const out = {};
|
||
|
||
// Ollama: prefer the URL the operator set via the StartOS action,
|
||
// then try the canonical service-discovery hostname, finally fall
|
||
// back to localhost (useful for dev outside StartOS).
|
||
try {
|
||
const cfg = await config.getConfigSnapshot();
|
||
const fromConfig = (cfg.ollama_base_url || "").trim();
|
||
let ollamaUrl = null;
|
||
let source = null;
|
||
if (fromConfig) {
|
||
ollamaUrl = fromConfig;
|
||
source = "config";
|
||
} else {
|
||
const candidate = "http://ollama.startos:11434";
|
||
const ok = await fetch(`${candidate}/api/tags`, {
|
||
signal: AbortSignal.timeout(1500),
|
||
}).then((r) => r.ok).catch(() => false);
|
||
if (ok) {
|
||
ollamaUrl = candidate;
|
||
source = "startos-dep";
|
||
}
|
||
}
|
||
if (ollamaUrl) {
|
||
// Probe /api/tags to list installed models — picker UI surfaces
|
||
// them as dropdown options so users don't have to type them.
|
||
let models = [];
|
||
try {
|
||
const tagsRes = await fetch(`${ollamaUrl.replace(/\/$/, "")}/api/tags`, {
|
||
signal: AbortSignal.timeout(2000),
|
||
});
|
||
if (tagsRes.ok) {
|
||
const data = await tagsRes.json();
|
||
models = (data.models || []).map((m) => m.name).filter(Boolean);
|
||
}
|
||
} catch {}
|
||
out.ollama = { baseURL: ollamaUrl, source, models };
|
||
}
|
||
} catch {}
|
||
|
||
res.json(out);
|
||
});
|
||
|
||
// Quick connection-test endpoint. The picker UI Test button hits this
|
||
// to verify a provider+model+opts combo actually works before the user
|
||
// commits to using it. Sends a tiny prompt and returns the model's
|
||
// 3-word answer + latency, or a clear error string.
|
||
app.post("/api/providers/test", async (req, res) => {
|
||
const { providerId, model, opts: clientOpts } = req.body || {};
|
||
if (!providerId || typeof providerId !== "string") {
|
||
return res.status(400).json({ ok: false, error: "missing providerId" });
|
||
}
|
||
if (!model || typeof model !== "string") {
|
||
return res.status(400).json({ ok: false, error: "missing model" });
|
||
}
|
||
if (!PROVIDER_NAMES.includes(providerId)) {
|
||
return res.status(400).json({ ok: false, error: `unknown provider: ${providerId}` });
|
||
}
|
||
let provider;
|
||
try {
|
||
const cfg = await config.getConfigSnapshot();
|
||
const resolvedOpts = resolveProviderOpts(providerId, {
|
||
config: cfg,
|
||
clientOpts: (clientOpts && typeof clientOpts === "object") ? clientOpts : {},
|
||
req,
|
||
});
|
||
provider = getProvider(providerId, resolvedOpts);
|
||
} catch (err) {
|
||
return res.status(400).json({ ok: false, error: err.message });
|
||
}
|
||
if (!provider.capabilities.analyze) {
|
||
return res.status(400).json({ ok: false, error: `${providerId} cannot analyze text` });
|
||
}
|
||
const t0 = Date.now();
|
||
// Special-case the relay: a real analyzeText round-trip would burn
|
||
// one of the user's lifetime credits per click. Use pingBalance
|
||
// instead — verifies connectivity + auth, returns the current
|
||
// balance, no charge.
|
||
if (providerId === "relay") {
|
||
try {
|
||
const env = await provider.pingBalance({ timeoutMs: 5000 });
|
||
const latencyMs = Date.now() - t0;
|
||
const credits = env?.credits_remaining;
|
||
const tier = env?.tier || "core";
|
||
const summary =
|
||
credits == null
|
||
? `Connected · Tier: ${tier}`
|
||
: `Connected · Tier: ${tier} · ${credits} credit${credits === 1 ? "" : "s"} remaining`;
|
||
return res.json({
|
||
ok: true,
|
||
text: summary,
|
||
latencyMs,
|
||
provider: providerId,
|
||
model,
|
||
});
|
||
} catch (err) {
|
||
return res.json({
|
||
ok: false,
|
||
error: (err?.message || String(err)).slice(0, 300),
|
||
latencyMs: Date.now() - t0,
|
||
});
|
||
}
|
||
}
|
||
try {
|
||
const result = await provider.analyzeText({
|
||
prompt: "Respond with exactly three words confirming you received this prompt.",
|
||
model,
|
||
retries: 0,
|
||
});
|
||
const latencyMs = Date.now() - t0;
|
||
return res.json({
|
||
ok: true,
|
||
text: (result.text || "").trim().slice(0, 200),
|
||
latencyMs,
|
||
provider: providerId,
|
||
model,
|
||
});
|
||
} catch (err) {
|
||
return res.json({
|
||
ok: false,
|
||
error: (err?.message || String(err)).slice(0, 300),
|
||
latencyMs: Date.now() - t0,
|
||
});
|
||
}
|
||
});
|
||
|
||
// Per-provider, per-field boolean indicating whether the StartOS
|
||
// config has a non-empty value for each PROVIDER_KEY_FIELDS slot.
|
||
// The picker UI uses this to (a) show a "✓ Server-configured" hint
|
||
// under empty inputs, and (b) decide whether the Delete button is
|
||
// visible when localStorage is empty but the server has a value.
|
||
//
|
||
// Never returns the actual values — only booleans. Anything stored in
|
||
// startos-config.json could be secret and shouldn't surface in any
|
||
// response that could end up in a screenshot.
|
||
app.get("/api/providers/credentials-status", async (_req, res) => {
|
||
const { PROVIDER_KEY_FIELDS } = await import("./providers/index.js");
|
||
const cfg = await getConfigSnapshot();
|
||
const status = {};
|
||
for (const [providerId, fields] of Object.entries(PROVIDER_KEY_FIELDS)) {
|
||
status[providerId] = {};
|
||
for (const [optName, cfgKey] of Object.entries(fields)) {
|
||
const v = cfg[cfgKey];
|
||
status[providerId][optName] =
|
||
typeof v === "string" && v.trim().length > 0;
|
||
}
|
||
}
|
||
res.json({ status });
|
||
});
|
||
|
||
// Clear all server-side config fields for a provider (the StartOS
|
||
// action-set values). The picker UI's Delete button calls this in
|
||
// addition to wiping localStorage so credentials are gone from BOTH
|
||
// storage paths. Returns the list of fields that were cleared.
|
||
app.post("/api/providers/:id/clear", async (req, res) => {
|
||
const providerId = req.params.id;
|
||
if (!PROVIDER_NAMES.includes(providerId)) {
|
||
return res.status(400).json({ ok: false, error: `unknown provider: ${providerId}` });
|
||
}
|
||
// Build a patch of { config_field: "" } for every PROVIDER_KEY_FIELDS
|
||
// entry mapped to this provider. Unknown providers or providers with
|
||
// no server-side fields (e.g. relay, whose URL is hardcoded) end
|
||
// up clearing nothing — still a 200, just with empty `cleared`.
|
||
const { PROVIDER_KEY_FIELDS } = await import("./providers/index.js");
|
||
const fields = PROVIDER_KEY_FIELDS[providerId] || {};
|
||
const patch = {};
|
||
for (const cfgKey of Object.values(fields)) {
|
||
if (typeof cfgKey === "string") patch[cfgKey] = "";
|
||
}
|
||
try {
|
||
await config.mergeConfig(patch);
|
||
return res.json({ ok: true, cleared: Object.keys(patch) });
|
||
} catch (err) {
|
||
return res.status(500).json({
|
||
ok: false,
|
||
error: (err?.message || String(err)).slice(0, 300),
|
||
});
|
||
}
|
||
});
|
||
|
||
app.post("/api/process", async (req, res) => {
|
||
// ── Multi-mode credit gate ───────────────────────────────────────────────
|
||
// Four reachable states arriving here in multi-mode:
|
||
// (a) req.user is admin → no local gate (operator pool)
|
||
// (b) req.user with license → license tier handles the gate (license-mw)
|
||
// (b2) paid cloud user → relay-side user:<id> quota gates it; NO
|
||
// local tenant_credits gate (core-decoupling)
|
||
// (c) req.user free tenant → tenant_credits.balance gate + debit
|
||
// (d) req.trial holder → trial cookie gate + debit
|
||
// (e) anonymous, no trial → try issueIfEligible, then re-check
|
||
//
|
||
// The signed-in-free-tenant case (c) was unhandled pre-0.2.92: their
|
||
// tenant_credits.balance was DISPLAYED but never gated or debited,
|
||
// so they could summarize indefinitely. Now we check + debit
|
||
// alongside the trial flow.
|
||
//
|
||
// Core-decoupling: paid status is the relay-owned tier (pro/max), NOT a
|
||
// Keysat license. Paid cloud users are metered by the relay's user:<id>
|
||
// monthly quota (its own 402 envelope), so they skip this local gate.
|
||
// Everyone else signed in (tier core) is a free tenant gated on
|
||
// tenant_credits — including accounts that still carry a now-ignored
|
||
// legacy license.
|
||
const isPaidCloudUser =
|
||
req.user && (req.user.tier === "pro" || req.user.tier === "max");
|
||
if (
|
||
RECAP_MODE === "multi" &&
|
||
req.user &&
|
||
!req.user.is_admin &&
|
||
!isPaidCloudUser
|
||
) {
|
||
// Free signed-in tenant — gate on tenant_credits.balance. We
|
||
// CHECK here at request start; the actual debit happens after
|
||
// saveToHistory succeeds (so a failed pipeline doesn't burn a
|
||
// credit). Same pattern as the trial flow below.
|
||
try {
|
||
const { getOrInit } = await import("./tenant-credits.js");
|
||
const credits = await getOrInit(req.user.id);
|
||
if (!credits || credits.total <= 0) {
|
||
return res.status(402).json({
|
||
error: "credits_exhausted",
|
||
message:
|
||
"You're out of credits. Buy more or wait for your free credits to refresh.",
|
||
});
|
||
}
|
||
// Stash for the post-save debit step
|
||
req.creditsToDebit = "tenant";
|
||
} catch (err) {
|
||
console.error("[/api/process] tenant_credits gate failed:", err);
|
||
return res.status(500).json({ error: "internal_error" });
|
||
}
|
||
}
|
||
|
||
if (RECAP_MODE === "multi" && !req.user) {
|
||
if (!req.trial && typeof req.userId === "undefined") {
|
||
// Pre-trial visitor on their first POST. Try to mint a cookie.
|
||
try {
|
||
const { issueIfEligible } = await import("./anon-trial.js");
|
||
const trial = await issueIfEligible({ req, res });
|
||
if (trial) {
|
||
req.userId = `anon:${trial.cookie_id}`;
|
||
req.trial = trial;
|
||
}
|
||
} catch (err) {
|
||
console.warn("[/api/process] anon-trial mint failed:", err);
|
||
}
|
||
}
|
||
if (!req.trial) {
|
||
// No session, no trial — either trials disabled, IP-capped, or
|
||
// their previous trial is exhausted (middleware fell through).
|
||
return res.status(401).json({
|
||
error: "trial_unavailable",
|
||
message:
|
||
"Sign in to keep using Recaps. New here? Create an account in a minute — no payment required.",
|
||
});
|
||
}
|
||
// Verify trial still has budget at request time. Cheap to check,
|
||
// catches the case where a parallel request just spent the last
|
||
// credit between middleware lookup and this handler firing.
|
||
const { lookupTrial, hasTrialBudget } = await import("./anon-trial.js");
|
||
const fresh = lookupTrial(req.trial.cookie_id);
|
||
if (!fresh || !hasTrialBudget(fresh)) {
|
||
return res.status(402).json({
|
||
error: "trial_exhausted",
|
||
message:
|
||
"You've used all your free trials. Sign up for an account to keep going.",
|
||
});
|
||
}
|
||
req.trial = fresh;
|
||
}
|
||
|
||
let {
|
||
url, apiKey: clientKey, model, type: itemType, title: itemTitle, uploadDate: itemUploadDate, episodeId,
|
||
transcriptionProvider: reqTransProvider,
|
||
transcriptionModel: reqTransModel,
|
||
analysisProvider: reqAnaProvider,
|
||
analysisModel: reqAnaModel,
|
||
providerOpts: reqProviderOpts,
|
||
useYouTubeCaptions: reqUseYTCaptions,
|
||
} = req.body;
|
||
// Defensive: a stale browser cache or older client might submit
|
||
// the literal string "Untitled" as the title (a sentinel that
|
||
// earlier Recap versions used as a falsy placeholder but which is
|
||
// actually truthy and pollutes every downstream gate — see relay
|
||
// 0.2.57's same normalization). Strip it here at the request
|
||
// boundary so all subsequent code can rely on "itemTitle is either
|
||
// a real title or empty". Also trim whitespace + drop "untitled"
|
||
// case-insensitively in case some other client variant emits it.
|
||
if (typeof itemTitle === "string") {
|
||
const trimmed = itemTitle.trim();
|
||
if (trimmed === "" || trimmed.toLowerCase() === "untitled") {
|
||
itemTitle = "";
|
||
} else {
|
||
itemTitle = trimmed;
|
||
}
|
||
}
|
||
// Default: use captions when available (huge speed/cost win). The
|
||
// picker-UI toggle lets users force a full transcription pass when
|
||
// they want speaker labels (captions don't have them) or when the
|
||
// auto-captions quality is too low.
|
||
const useYouTubeCaptions = reqUseYTCaptions !== false;
|
||
// Sentinel error message thrown by checkCancelled() when the user
|
||
// hits the in-flight banner's Cancel button. Declared at handler
|
||
// scope (not inside the try) so the catch block can compare against
|
||
// it — see the catch a few hundred lines below.
|
||
const CANCELLED_MARK = "__recap_cancelled__";
|
||
// Per-provider client-side opts: { gemini: {apiKey}, anthropic: {apiKey},
|
||
// openai: {apiKey}, "openai-compatible": {apiKey, baseURL}, ollama: {baseURL} }.
|
||
// Each provider's opts override that provider's config-stored values
|
||
// (set via the StartOS actions). Used by the picker UI to BYO keys
|
||
// per provider without round-tripping the StartOS dashboard.
|
||
const providerOpts = (reqProviderOpts && typeof reqProviderOpts === "object") ? reqProviderOpts : {};
|
||
|
||
// Provider selection: each pipeline step (transcribe + analyze) can
|
||
// independently target any registered provider. Both default to gemini
|
||
// so existing clients (which don't send provider fields) keep working.
|
||
const transcriptionProvider = reqTransProvider || "gemini";
|
||
const analysisProvider = reqAnaProvider || "gemini";
|
||
|
||
// Free tier: unlicensed users can summarize one video at a time. They
|
||
// still bring their own key — same as paid users today; the key can
|
||
// come from either the StartOS config action (server-side) or the
|
||
// web UI Settings panel (client-side). The future "bundled key" relay
|
||
// (paid users' requests proxied through the operator's service) isn't
|
||
// built yet, so there's nothing here that gates key sourcing by tier.
|
||
// AbortController for this request. Fired by abortCurrentFreeJob()
|
||
// when the user hits Cancel — passed through to every provider SDK
|
||
// call (transcription + analysis) so in-flight network requests
|
||
// reject immediately instead of running to completion.
|
||
const abortController = new AbortController();
|
||
|
||
// Stable identifier for this summarize job. Sent to the relay
|
||
// (when used) as `X-Recap-Job-Id`. The relay charges 1 credit on
|
||
// the first call with a given jobId and treats subsequent calls
|
||
// with the same id as free — so a full summary (transcribe +
|
||
// analyze) costs one credit regardless of which steps route
|
||
// through the relay. Non-relay providers ignore this opt.
|
||
const jobId = randomUUID();
|
||
|
||
const isFree = isFreeUser();
|
||
if (isFree) {
|
||
if (!tryAcquireFreeSlot({ url, title: itemTitle, abortController })) {
|
||
const current = getCurrentFreeJob();
|
||
const elapsedSec = current ? Math.round(current.elapsedMs / 1000) : 0;
|
||
const what = current?.title || current?.url || "another video";
|
||
return res.status(409).json({
|
||
error: "processing_in_progress",
|
||
message:
|
||
`A summary is already being processed (${what}, started ${elapsedSec}s ago). Free mode handles one video at a time — wait for it to finish, or cancel it from the status bar at the top of the app.`,
|
||
currentJob: current,
|
||
});
|
||
}
|
||
}
|
||
|
||
if (!url) {
|
||
if (isFree) releaseFreeSlot();
|
||
return res.status(400).json({ error: "Missing url" });
|
||
}
|
||
if (!PROVIDER_NAMES.includes(transcriptionProvider)) {
|
||
if (isFree) releaseFreeSlot();
|
||
return res.status(400).json({ error: "unknown_provider", message: `Unknown transcription provider: ${transcriptionProvider}` });
|
||
}
|
||
if (!PROVIDER_NAMES.includes(analysisProvider)) {
|
||
if (isFree) releaseFreeSlot();
|
||
return res.status(400).json({ error: "unknown_provider", message: `Unknown analysis provider: ${analysisProvider}` });
|
||
}
|
||
|
||
// Resolve per-provider construction opts from the StartOS config blob,
|
||
// overlaying any client-supplied opts (req.body.providerOpts[name]).
|
||
// For Gemini, the legacy single "apiKey" field on the request body
|
||
// also flows through as a fallback when providerOpts.gemini.apiKey
|
||
// isn't set — keeps pre-picker-UI clients working.
|
||
const cfg = await getConfigSnapshot();
|
||
function clientOptsFor(name) {
|
||
const fromBody = (providerOpts[name] && typeof providerOpts[name] === "object") ? providerOpts[name] : {};
|
||
if (name === "gemini" && !fromBody.apiKey) {
|
||
const legacy = resolveApiKey(clientKey);
|
||
if (legacy) return { ...fromBody, apiKey: legacy };
|
||
}
|
||
return fromBody;
|
||
}
|
||
const transcriptionOpts = resolveProviderOpts(transcriptionProvider, {
|
||
config: cfg,
|
||
clientOpts: clientOptsFor(transcriptionProvider),
|
||
req,
|
||
});
|
||
const analysisOpts = resolveProviderOpts(analysisProvider, {
|
||
config: cfg,
|
||
clientOpts: clientOptsFor(analysisProvider),
|
||
req,
|
||
});
|
||
|
||
let transcriber;
|
||
try {
|
||
transcriber = getProvider(transcriptionProvider, transcriptionOpts);
|
||
} catch (err) {
|
||
if (isFree) releaseFreeSlot();
|
||
return res.status(400).json({
|
||
error: "transcription_provider_not_configured",
|
||
message: `Transcription provider ${transcriptionProvider} is not configured: ${err.message}`,
|
||
});
|
||
}
|
||
if (!transcriber.capabilities.transcribe) {
|
||
if (isFree) releaseFreeSlot();
|
||
return res.status(400).json({
|
||
error: "transcription_unsupported",
|
||
message: `Provider ${transcriptionProvider} does not support audio transcription. Pick a different transcription provider (gemini or openai).`,
|
||
});
|
||
}
|
||
|
||
let analyzer;
|
||
try {
|
||
analyzer = transcriptionProvider === analysisProvider
|
||
? transcriber
|
||
: getProvider(analysisProvider, analysisOpts);
|
||
} catch (err) {
|
||
if (isFree) releaseFreeSlot();
|
||
return res.status(400).json({
|
||
error: "analysis_provider_not_configured",
|
||
message: `Analysis provider ${analysisProvider} is not configured: ${err.message}`,
|
||
});
|
||
}
|
||
if (!analyzer.capabilities.analyze) {
|
||
if (isFree) releaseFreeSlot();
|
||
return res.status(400).json({
|
||
error: "analysis_unsupported",
|
||
message: `Provider ${analysisProvider} does not support text analysis.`,
|
||
});
|
||
}
|
||
|
||
// ── Resolve Apple Podcasts / Spotify share URLs ──
|
||
// Most users paste apple/spotify share links rather than RSS feeds.
|
||
// Resolve those to a direct audio enclosure URL up-front so the rest
|
||
// of the handler runs the existing podcast pipeline unchanged. On
|
||
// resolver failure we surface a clear error rather than letting the
|
||
// downstream code stumble on a URL it can't process.
|
||
let resolvedFrom = null; // "apple" | "spotify" | null — for the result event
|
||
if (isResolvableShareUrl(url)) {
|
||
try {
|
||
const resolved = await resolveShareUrl(url, {
|
||
podcastIndexKey: cfg.podcastindex_api_key,
|
||
podcastIndexSecret: cfg.podcastindex_api_secret,
|
||
});
|
||
if (resolved) {
|
||
console.log(
|
||
` [url-resolver] ${resolved.source} → ${resolved.audioUrl.slice(0, 80)} ("${resolved.title}")`
|
||
);
|
||
url = resolved.audioUrl;
|
||
if (!itemTitle) itemTitle = resolved.title;
|
||
if (!itemUploadDate) itemUploadDate = resolved.uploadDate;
|
||
if (!episodeId) episodeId = resolved.episodeId;
|
||
itemType = "podcast";
|
||
resolvedFrom = resolved.source;
|
||
}
|
||
} catch (err) {
|
||
if (isFree) releaseFreeSlot();
|
||
const code = err instanceof URLResolveError ? err.code : "resolve_failed";
|
||
return res.status(400).json({
|
||
error: code,
|
||
message: err?.message || "Failed to resolve share URL",
|
||
});
|
||
}
|
||
}
|
||
|
||
// Determine if this is a podcast episode or YouTube video
|
||
const isPodcast = itemType === "podcast" || /\.(mp3|m4a|ogg|opus|wav|aac)(\?|$)/i.test(url);
|
||
const videoId = isPodcast ? (episodeId || url) : extractVideoId(url);
|
||
|
||
if (!isPodcast && !videoId) {
|
||
if (isFree) releaseFreeSlot();
|
||
return res.status(400).json({ error: "Invalid YouTube URL" });
|
||
}
|
||
|
||
const tmpDir = await fs.mkdtemp(path.join(os.tmpdir(), "yts-"));
|
||
const audioExt = isPodcast ? (url.match(/\.(mp3|m4a|ogg|opus|wav|aac)/i)?.[1] || "mp3") : "mp3";
|
||
const audioPath = path.join(tmpDir, `audio.${audioExt}`);
|
||
const mimeType = { mp3: "audio/mp3", m4a: "audio/mp4", ogg: "audio/ogg", opus: "audio/opus", wav: "audio/wav", aac: "audio/aac" }[audioExt] || "audio/mp3";
|
||
|
||
try {
|
||
const pipelineStart = Date.now();
|
||
|
||
// Set up SSE
|
||
res.writeHead(200, {
|
||
"Content-Type": "text/event-stream",
|
||
"Cache-Control": "no-cache",
|
||
Connection: "keep-alive",
|
||
});
|
||
|
||
// Helper to send log entries with elapsed time. Also pushes onto the
|
||
// free-tier in-flight job's server-side buffer (no-op when there is
|
||
// no free job, e.g. licensed user) so a browser refresh during a
|
||
// long pipeline can re-hydrate the activity log from the server
|
||
// instead of starting blank.
|
||
const logHistory = [];
|
||
function log(step, message, detail) {
|
||
const elapsed = ((Date.now() - pipelineStart) / 1000).toFixed(1);
|
||
const logMsg = `[${elapsed}s] ${message}`;
|
||
console.log(` ${logMsg}`);
|
||
const entry = { elapsed, message, detail: detail || null };
|
||
logHistory.push(entry);
|
||
appendCurrentJobLog(entry);
|
||
sendEvent(res, "status", { step, message });
|
||
sendEvent(res, "log", entry);
|
||
}
|
||
|
||
// Cancellation checkpoint. Polled at each major pipeline boundary;
|
||
// throws CANCELLED_MARK (declared at handler scope above) when the
|
||
// user has hit Cancel on the status banner so the request bails
|
||
// out cleanly via the catch path. Also catches abort signals fired
|
||
// by abortCurrentFreeJob() in case an SDK call swallowed the abort
|
||
// and returned cleanly.
|
||
function checkCancelled() {
|
||
if (abortController.signal.aborted) {
|
||
throw new Error(CANCELLED_MARK);
|
||
}
|
||
if (isFree && isFreeJobAborted()) {
|
||
throw new Error(CANCELLED_MARK);
|
||
}
|
||
}
|
||
|
||
// ── Recap Relay mode (unified pipeline) ──────────────────────
|
||
// When the user picks "Recap Relay" in the picker, both providers
|
||
// are set to "relay". The relay's /relay/summarize-url endpoint
|
||
// does the WHOLE pipeline (download + transcribe + chunked-analyze)
|
||
// server-side and streams per-window section results back to us
|
||
// over SSE. We forward those events to the browser as the existing
|
||
// event types (transcript_ready / sections_partial / result) so
|
||
// the front-end's incremental rendering keeps working unchanged.
|
||
//
|
||
// Why this branch exists instead of just letting the per-step
|
||
// pickers route to relay independently:
|
||
// 1. Latency — saves 12+ round-trips (one per analyze window)
|
||
// and the bandwidth of sending the full transcript back
|
||
// to Recap just for Recap to ship it back as 12 prompts.
|
||
// 2. Settings respect — the operator's Settings-tab knobs
|
||
// (window body, overlap, concurrency, etc.) only take
|
||
// effect when the relay drives the analyze loop. Per-step
|
||
// relay use would leave Recap's hardcoded constants in
|
||
// charge of windowing.
|
||
// 3. Credits clarity — one summarize = one credit, charged
|
||
// atomically on the relay side. No "what if transcribe
|
||
// succeeded but analyze fell back to Anthropic" accounting.
|
||
const relayMode =
|
||
transcriptionProvider === "relay" && analysisProvider === "relay";
|
||
if (relayMode) {
|
||
// Title surrogate — used for the activity log display + as a
|
||
// fallback if the relay doesn't echo a real title back (older
|
||
// relay, or a podcast/non-YouTube URL where yt-dlp metadata
|
||
// isn't available). Defaults to "Untitled" so log messages
|
||
// remain readable.
|
||
const titleSurrogate = itemTitle || "Untitled";
|
||
// titleHint sent to the relay — must NOT be the surrogate.
|
||
// Passing "Untitled" as a hint causes a subtle bug: the relay's
|
||
// download step uses `if (!title && audio.title) title = audio.title;`
|
||
// which treats any truthy operator-supplied title (including
|
||
// "Untitled") as a real hint and skips the yt-dlp-extracted
|
||
// title fallback. Net effect: the library entry stays
|
||
// "Untitled" forever even though yt-dlp got the real title.
|
||
// Pass only the actual operator/subscription-supplied title (or
|
||
// empty string, which the provider serializes as omitted JSON).
|
||
const titleHintRaw = itemTitle ? String(itemTitle).trim() : "";
|
||
|
||
// For YouTube URLs, fetch the metadata block (channel, description,
|
||
// chapters) BEFORE handing off to the relay. The relay can extract
|
||
// these fields itself via yt-dlp (relay 0.2.63+ does so) but
|
||
// fetching here gives the relay a head-start AND lets older relays
|
||
// that don't yet do their own extraction still get the speaker-ID
|
||
// metadata. The fetch costs ~1-2 seconds — small relative to the
|
||
// 1-3 minute pipeline. Defense-in-depth: belt-and-suspenders with
|
||
// the relay-side fallback.
|
||
//
|
||
// For podcasts, skip — fetchYouTubeMetadata is YouTube-specific
|
||
// and would return null anyway. The podcast resolver above
|
||
// already populated itemTitle when relevant; channel/description/
|
||
// chapters aren't available for podcast feeds.
|
||
let relayChannelHint = "";
|
||
let relayDescriptionHint = "";
|
||
let relayChaptersHint = [];
|
||
if (!isPodcast && videoId) {
|
||
try {
|
||
const meta = await fetchYouTubeMetadata(videoId);
|
||
if (meta) {
|
||
relayChannelHint = meta.channel || "";
|
||
relayDescriptionHint = meta.description || "";
|
||
relayChaptersHint = Array.isArray(meta.chapters) ? meta.chapters : [];
|
||
log(
|
||
1,
|
||
`YouTube metadata fetched: channel="${relayChannelHint.slice(0, 60)}", ${relayChaptersHint.length} chapter(s), description ${relayDescriptionHint.length} chars`
|
||
);
|
||
}
|
||
} catch (err) {
|
||
// Non-fatal — the relay's own yt-dlp extraction will fill these
|
||
// fields when our request body sends them as empty strings.
|
||
log(1, `YouTube metadata fetch failed: ${err?.message || err} — relay will extract via its own yt-dlp`);
|
||
}
|
||
}
|
||
|
||
log(1, `Recap Relay mode — full pipeline on operator's relay`);
|
||
// Cache of parsed transcript entries — populated by the
|
||
// onTranscribeComplete provider callback below as soon as the
|
||
// relay finishes the transcribe phase. Used by the per-window
|
||
// streaming callback to build full { title, summary, entries,
|
||
// startTime } chunk objects, which is the shape the browser's
|
||
// sections_partial handler reads via `data.chunks`. Without
|
||
// this cache, per-window streaming sent just titles + summaries
|
||
// and the browser had no entries to render under each section.
|
||
let streamedRelayEntries = [];
|
||
let relayResult;
|
||
// Track whether we've emitted the early transcript_ready event
|
||
// yet. In sequential mode this fires from onTranscribeComplete
|
||
// (when relay's transcribe finishes); in pipelined mode it can
|
||
// fire even earlier from the FIRST onWindowComplete (when an
|
||
// early window's sections arrive before transcribe is done).
|
||
// Either way: we want the browser to flip from loading view to
|
||
// results view ASAP so it can render the streaming sections.
|
||
let earlyTranscriptReadyEmitted = false;
|
||
const emitEarlyTranscriptReady = (extra = {}) => {
|
||
if (earlyTranscriptReadyEmitted) return;
|
||
earlyTranscriptReadyEmitted = true;
|
||
sendEvent(res, "transcript_ready", {
|
||
videoId,
|
||
// Best-known title at this point — the surrogate from the
|
||
// user-supplied title hint or "Untitled". The final result
|
||
// event also carries videoTitle and the browser updates
|
||
// it then, so any final title mismatch self-heals.
|
||
videoTitle: titleSurrogate,
|
||
videoChannel: relayChannelHint || "",
|
||
videoChapters: relayChaptersHint || [],
|
||
videoUploadDate: itemUploadDate || "",
|
||
videoDescription: relayDescriptionHint || "",
|
||
entries: streamedRelayEntries || [],
|
||
type: isPodcast ? "podcast" : "youtube",
|
||
willChunk: true, // pessimistic; corrected at result time
|
||
...extra,
|
||
});
|
||
};
|
||
try {
|
||
relayResult = await transcriber.summarizeUrl({
|
||
mediaUrl: url,
|
||
mediaType: isPodcast ? "podcast" : "youtube",
|
||
mimeType: undefined,
|
||
titleHint: titleHintRaw,
|
||
channelHint: relayChannelHint,
|
||
descriptionHint: relayDescriptionHint,
|
||
chaptersHint: relayChaptersHint,
|
||
onProgress: (msg) => log(2, msg),
|
||
// Capture transcript entries the moment the relay finishes
|
||
// transcribing. Analyze runs serially AFTER transcribe in
|
||
// the relay's pipeline — so by the time the FIRST
|
||
// onWindowComplete fires, this callback has already populated
|
||
// the entries cache below. That lets the onWindowComplete
|
||
// callback build proper { title, summary, entries, startTime }
|
||
// chunk objects (the shape the browser's sections_partial
|
||
// handler expects under .chunks) rather than the raw
|
||
// ownedSections-without-entries we sent pre-v0.2.75.
|
||
onTranscribeComplete: ({ transcript }) => {
|
||
streamedRelayEntries = parseTimestampedTranscript(transcript || "");
|
||
log(2, `Relay transcript parsed — ${streamedRelayEntries.length} segments cached for streaming`);
|
||
// Sequential-mode UI flip: emit transcript_ready as soon
|
||
// as transcribe is done so the browser switches from
|
||
// loading view to results view BEFORE the analyze
|
||
// sections start streaming in. Pipelined mode beats us to
|
||
// this from onWindowComplete (the first window may
|
||
// complete before transcribe is fully done) — either way
|
||
// the emitter is idempotent.
|
||
emitEarlyTranscriptReady();
|
||
},
|
||
// Per-window streaming: forward each window's owned sections
|
||
// to the browser as sections_partial events with the chunk
|
||
// SHAPE the browser handler expects (.chunks not
|
||
// .partialSections). Without this, the browser's
|
||
// `data.chunks || []` read at the sections_partial handler
|
||
// resolved to an empty array and the operator saw zero
|
||
// sections appear during the analyze phase — they only
|
||
// landed when the terminal "result" event arrived at the
|
||
// very end of the pipeline.
|
||
onWindowComplete: ({ windowIdx, totalWindows, ownedSections, windowEntries }) => {
|
||
// Pipelined-mode UI flip: if the first window completes
|
||
// before transcribe-done (very possible — first window
|
||
// only needs ~20min of audio fully transcribed), emit
|
||
// transcript_ready NOW so the browser flips to results
|
||
// view in time for these partial sections to render.
|
||
emitEarlyTranscriptReady();
|
||
// Pipelined mode (relay v0.2.89+): each window arrives
|
||
// with its OWN entries embedded, before transcribe has
|
||
// fully completed. Section indices are LOCAL to those
|
||
// window entries, so we slice them directly.
|
||
//
|
||
// Sequential mode (older relays, or Gemini-transcribe
|
||
// path): windowEntries is null. The relay's transcribe
|
||
// phase has already finished by the time these arrive,
|
||
// so streamedRelayEntries (populated by
|
||
// onTranscribeComplete) is non-empty and section
|
||
// indices are global into it.
|
||
const entries = Array.isArray(windowEntries) && windowEntries.length > 0
|
||
? windowEntries
|
||
: (streamedRelayEntries || []);
|
||
const partialChunks = (ownedSections || []).map((s) => {
|
||
const start = Math.max(0, s.startIndex ?? 0);
|
||
const end = entries.length > 0
|
||
? Math.min(entries.length - 1, s.endIndex ?? start)
|
||
: (s.endIndex ?? start);
|
||
const slice = entries.length > 0 ? entries.slice(start, end + 1) : [];
|
||
return {
|
||
title: s.title,
|
||
summary: s.summary,
|
||
entries: slice,
|
||
startTime: slice[0]?.offset || 0,
|
||
};
|
||
}).filter((c) => c.entries.length > 0 || !entries.length);
|
||
sendEvent(res, "sections_partial", {
|
||
windowIdx,
|
||
totalWindows,
|
||
chunks: partialChunks,
|
||
});
|
||
},
|
||
signal: abortController.signal,
|
||
jobId,
|
||
});
|
||
} catch (err) {
|
||
if (abortController.signal.aborted) throw new Error(CANCELLED_MARK);
|
||
sendEvent(res, "error", { message: err?.message || String(err) });
|
||
sendEvent(res, "result", { videoId, entries: [], chunks: [], rawTranscript: "" });
|
||
if (isFree) releaseFreeSlot();
|
||
res.end();
|
||
return;
|
||
}
|
||
checkCancelled();
|
||
|
||
// Backfill the real title now that the relay has finished. The
|
||
// relay extracts it via yt-dlp during the download step (or
|
||
// echoes back the client-supplied hint) and returns it in the
|
||
// result envelope's `title` field — see relay 0.2.53+. For older
|
||
// relays the field is null, so we keep the original surrogate.
|
||
// Trim + drop empty strings to avoid storing whitespace-only
|
||
// titles that would render as a blank row in the library list.
|
||
const relayTitle = typeof relayResult.title === "string"
|
||
? relayResult.title.trim()
|
||
: "";
|
||
const resolvedTitle = relayTitle || titleSurrogate;
|
||
|
||
// Parse the bracketed transcript into the entries array shape
|
||
// Recap's UI / history layer expects.
|
||
const relayEntries = parseTimestampedTranscript(relayResult.transcript || "");
|
||
|
||
// Phase 1E — attach speaker labels to each parsed entry by
|
||
// time-matching against the raw per-segment diarization data
|
||
// the relay returned. The relay's `transcript_segments` are the
|
||
// fine-grained Parakeet segments (often 1-3s each) with a
|
||
// `speaker` field per segment. Recap's `relayEntries` are the
|
||
// MERGED readable lines (sortAndDedupeTranscript + mergeShort-
|
||
// Entries collapsed multiple Parakeet segments into one
|
||
// sentence-sized thought). To attach speakers to merged
|
||
// entries: for each entry, find the raw segment whose [start,
|
||
// end] contains the entry's offset (start time). The merged
|
||
// entry's start time always corresponds to the start of one of
|
||
// the underlying segments — that segment's speaker is the
|
||
// entry's speaker. Speaker stays null when:
|
||
// - relay version < 0.2.88 (no transcript_segments field)
|
||
// - diarization was off on the operator's relay
|
||
// - relay couldn't match the entry's time to any segment
|
||
const relaySegments = Array.isArray(relayResult.transcript_segments)
|
||
? relayResult.transcript_segments
|
||
: null;
|
||
if (relaySegments && relaySegments.length > 0) {
|
||
// Build a sorted-by-start list for an O(N log N + M log N) match.
|
||
// Linear scan would be O(N*M) which is fine for typical sizes
|
||
// but the sorted form also makes the data debuggable.
|
||
const sortedSegs = relaySegments
|
||
.slice()
|
||
.sort((a, b) => (a.start || 0) - (b.start || 0));
|
||
for (const entry of relayEntries) {
|
||
const t = entry.offset || 0;
|
||
// Find the segment whose [start, end] contains t. Use
|
||
// binary search for efficiency on long transcripts.
|
||
let lo = 0;
|
||
let hi = sortedSegs.length - 1;
|
||
let found = null;
|
||
while (lo <= hi) {
|
||
const mid = (lo + hi) >> 1;
|
||
const seg = sortedSegs[mid];
|
||
if (t < (seg.start || 0)) {
|
||
hi = mid - 1;
|
||
} else if (t > (seg.end || 0)) {
|
||
lo = mid + 1;
|
||
} else {
|
||
found = seg;
|
||
break;
|
||
}
|
||
}
|
||
// If no segment contains t (gaps in diarization output),
|
||
// fall back to the segment that starts CLOSEST to t.
|
||
if (!found && sortedSegs.length > 0) {
|
||
const idx = Math.min(Math.max(lo, 0), sortedSegs.length - 1);
|
||
const candidates = [
|
||
sortedSegs[idx],
|
||
sortedSegs[idx - 1],
|
||
sortedSegs[idx + 1],
|
||
].filter(Boolean);
|
||
let bestDist = Infinity;
|
||
for (const seg of candidates) {
|
||
const segMid = ((seg.start || 0) + (seg.end || 0)) / 2;
|
||
const d = Math.abs(segMid - t);
|
||
if (d < bestDist) {
|
||
bestDist = d;
|
||
found = seg;
|
||
}
|
||
}
|
||
// Cap the fallback at 5s — beyond that the assignment is
|
||
// probably bogus, leave it null so the UI can render
|
||
// "(speaker unknown)" or just drop the chip.
|
||
if (bestDist > 5) found = null;
|
||
}
|
||
if (found && found.speaker) {
|
||
entry.speaker = found.speaker;
|
||
entry.speaker_confidence =
|
||
typeof found.speaker_confidence === "number"
|
||
? found.speaker_confidence
|
||
: null;
|
||
// Phase 2 — surface the suppression-pass uncertainty
|
||
// flag so the chip renders with a "?" suffix for
|
||
// best-guess attribution. Absent on older relays.
|
||
entry.speaker_uncertain = !!found.speaker_uncertain;
|
||
}
|
||
}
|
||
}
|
||
|
||
const contentType = isPodcast ? "podcast" : "youtube";
|
||
// Note: transcript_ready was emitted EARLIER by
|
||
// emitEarlyTranscriptReady() — either from the first
|
||
// onWindowComplete (pipelined mode) or from
|
||
// onTranscribeComplete (sequential mode). Re-emitting here
|
||
// would either (a) be a no-op since the browser handler is
|
||
// already in results view, or (b) double-trigger the chunk
|
||
// wipe / streamWindowsTotal reset. The real title arrives
|
||
// via the result event below — browser's result handler
|
||
// updates state.videoTitle on receipt — so dropping the
|
||
// duplicate emit is safe.
|
||
|
||
// Build chunks from the relay's stitched analysis sections.
|
||
const sections = Array.isArray(relayResult.analysis?.sections)
|
||
? relayResult.analysis.sections
|
||
: [];
|
||
const chunks = sections
|
||
.map((s) => {
|
||
const start = Math.max(0, s.startIndex ?? 0);
|
||
const end = Math.min(relayEntries.length - 1, s.endIndex ?? start);
|
||
const slice = relayEntries.slice(start, end + 1);
|
||
return {
|
||
title: s.title,
|
||
summary: s.summary,
|
||
entries: slice,
|
||
startTime: slice[0]?.offset || 0,
|
||
};
|
||
})
|
||
.filter((c) => c.entries.length > 0);
|
||
|
||
const historyId = await saveToHistory(
|
||
scopeForRequest(req),
|
||
videoId,
|
||
url,
|
||
resolvedTitle,
|
||
chunks,
|
||
relayEntries,
|
||
logHistory,
|
||
itemUploadDate || "",
|
||
contentType,
|
||
relayResult.speakers || null,
|
||
relayResult.speaker_names || null,
|
||
).catch(() => null);
|
||
// Debit local credit on success — trial cookie OR free signed-in
|
||
// tenant, depending on who's making the request. AFTER
|
||
// saveToHistory rather than at request-start so a failed pipeline
|
||
// doesn't burn a credit.
|
||
if (RECAP_MODE === "multi" && req.trial) {
|
||
try {
|
||
const { debitOne } = await import("./anon-trial.js");
|
||
debitOne(req.trial.cookie_id);
|
||
} catch (err) {
|
||
console.warn("[/api/process] trial debit failed:", err);
|
||
}
|
||
} else if (RECAP_MODE === "multi" && req.creditsToDebit === "tenant") {
|
||
try {
|
||
const { gateAndDebit } = await import("./tenant-credits.js");
|
||
await gateAndDebit(req.user.id);
|
||
} catch (err) {
|
||
console.warn("[/api/process] tenant_credits debit failed:", err);
|
||
}
|
||
}
|
||
sendEvent(res, "result", {
|
||
videoId,
|
||
videoTitle: resolvedTitle,
|
||
entries: relayEntries,
|
||
chunks,
|
||
historyId,
|
||
type: contentType,
|
||
// Phase 1E — speaker legend + per-segment confidence summary.
|
||
// Null when diarization wasn't available (older relay, or
|
||
// operator disabled it). Frontend renders a colored chip per
|
||
// speaker above the topic list; each transcript line uses
|
||
// the entry.speaker field (already attached above) to pick
|
||
// the matching chip color.
|
||
speakers: relayResult.speakers || null,
|
||
// Phase 2 — inferred speaker names from the relay's post-
|
||
// cluster polish pass. Map { Speaker_A: "Matt Hill", ... }
|
||
// with null values for speakers the LLM couldn't confidently
|
||
// name. Recap legend prefers the inferred name when present.
|
||
speaker_names: relayResult.speaker_names || null,
|
||
});
|
||
if (isFree) releaseFreeSlot();
|
||
res.end();
|
||
return;
|
||
}
|
||
|
||
let videoTitle = itemTitle || "Untitled";
|
||
let videoUploadDate = itemUploadDate || "";
|
||
// Rich metadata used to ground the transcription prompt (speaker
|
||
// names from the channel/description/chapters) and surfaced to
|
||
// downstream UI. Populated for YouTube videos only; podcasts have
|
||
// their own (lighter) metadata path.
|
||
let videoChannel = "";
|
||
let videoDescription = "";
|
||
let videoChapters = [];
|
||
// If captions came from YouTube we skip Step 1 (download audio) +
|
||
// Step 2 (transcribe) and jump straight to analysis. `entries` is
|
||
// also the post-Step-2 output of the transcription path, so this
|
||
// value just flows through.
|
||
let entries = null;
|
||
let captionSource = null; // "manual" | "auto" | null
|
||
let transcriptText = "";
|
||
let txCost = { inputTokens: 0, outputTokens: 0, thinkingTokens: 0, totalTokens: 0, totalCost: "0", totalCostDisplay: "$0.00" };
|
||
// Populated from YouTube metadata or (later) Apple/Spotify resolution.
|
||
// Used by the relay-URL fast-path to decide if the audio fits the
|
||
// relay's reported capability. null = unknown → fast-path attempts
|
||
// anyway and lets the relay reject if the file is over its cap.
|
||
let knownDuration = null;
|
||
// Hoisted out of the audio-path block so Step 3 (analysis) can
|
||
// reference analysisModel even when the captions fast-path skips
|
||
// the entire audio + transcription block. transcriptionModel is
|
||
// only used inside the audio block but lives here for symmetry.
|
||
// Per-provider model defaults: caller's request → provider's first
|
||
// listed model → Gemini fallback (preserves the prior default for
|
||
// legacy clients that don't send model fields).
|
||
const transcriptionModel =
|
||
reqTransModel ||
|
||
transcriber.listTranscriptionModels()[0] ||
|
||
"gemini-3-flash-preview";
|
||
const analysisModel =
|
||
reqAnaModel ||
|
||
model ||
|
||
analyzer.listAnalysisModels()[0] ||
|
||
"gemini-3.1-pro-preview";
|
||
|
||
if (resolvedFrom) {
|
||
log(
|
||
0,
|
||
`Resolved ${
|
||
resolvedFrom === "apple" ? "Apple Podcasts" : "Spotify"
|
||
} share URL → ${url.slice(0, 80)}${url.length > 80 ? "…" : ""}`
|
||
);
|
||
}
|
||
|
||
// ── Step 0 (YouTube only): metadata + captions fast path ──
|
||
if (!isPodcast && videoId) {
|
||
log(1, "Fetching video metadata...");
|
||
const meta = await fetchYouTubeMetadata(videoId);
|
||
if (meta) {
|
||
if (meta.title) videoTitle = meta.title;
|
||
if (meta.uploadDate) videoUploadDate = meta.uploadDate;
|
||
videoChannel = meta.channel;
|
||
videoDescription = meta.description;
|
||
videoChapters = meta.chapters;
|
||
if (typeof meta.duration === "number" && meta.duration > 0) {
|
||
knownDuration = meta.duration;
|
||
}
|
||
log(1, `Video title: ${videoTitle}${videoChannel ? ` (${videoChannel})` : ""}`);
|
||
if ((meta.hasManualCaptions || meta.hasAutoCaptions) && !useYouTubeCaptions) {
|
||
log(1, `YouTube captions available but the user has the captions fast-path disabled — will transcribe audio directly.`);
|
||
} else if (meta.hasManualCaptions || meta.hasAutoCaptions) {
|
||
log(1, `YouTube captions available (${meta.hasManualCaptions ? "manual" : "auto"}) — attempting fast-path…`);
|
||
const cap = await tryFetchYouTubeCaptions(
|
||
videoId,
|
||
tmpDir,
|
||
{ hasManual: meta.hasManualCaptions, hasAuto: meta.hasAutoCaptions },
|
||
log
|
||
);
|
||
if (cap && cap.entries && cap.entries.length > 0) {
|
||
// Auto-captions fragment audio every 1–3 seconds (often only
|
||
// 1–3 words per entry). Sending hundreds of micro-segments
|
||
// to the analyzer balloons the prompt and overloads the
|
||
// model's index-tracking ("Section N must start at index
|
||
// K+1 of section M…"). Coalesce into ~15s blocks for a
|
||
// saner analysis prompt while keeping timestamps accurate.
|
||
const rawCount = cap.entries.length;
|
||
entries = coalesceTranscriptEntries(cap.entries, 15);
|
||
captionSource = cap.source;
|
||
log(
|
||
1,
|
||
`Using YouTube ${cap.source === "manual" ? "manual" : "auto-generated"} captions — ${rawCount} segments coalesced to ${entries.length}, skipping audio download + transcription`
|
||
);
|
||
} else {
|
||
log(1, "Captions unusable or failed to parse — falling back to audio transcription");
|
||
}
|
||
} else {
|
||
log(1, "No captions available — will download audio and transcribe");
|
||
}
|
||
}
|
||
}
|
||
|
||
checkCancelled();
|
||
|
||
// ── Step 0.5: Relay-URL fast-path ──
|
||
// When transcription routes through the operator's relay AND we
|
||
// have a public source URL (YouTube watch URL, or direct podcast
|
||
// .mp3/.m4a enclosure), hand the URL off to /relay/transcribe-url
|
||
// so the relay does the download itself. This saves the buyer's
|
||
// home upload bandwidth — often the slowest leg of the pipeline,
|
||
// since residential connections rarely match the relay's symmetric
|
||
// datacenter link. Only fires if:
|
||
// 1. captions fast-path didn't already populate entries
|
||
// 2. caller asked to use the relay for transcription
|
||
// 3. duration (when known) fits the relay's reported capability
|
||
// On any failure we fall through to the legacy local-download
|
||
// path, which has its own retry/fallback machinery.
|
||
if (!entries && transcriptionProvider === "relay") {
|
||
// Fresh fetch so the fast-path decision matches the routing
|
||
// decision the relay will make for THIS install — same
|
||
// rationale as the Step 2 refresh below.
|
||
await refreshRelayCapabilities().catch(() => {});
|
||
const caps = getRelayCapabilities();
|
||
const withinCapability =
|
||
typeof caps.max_audio_minutes !== "number" ||
|
||
knownDuration === null ||
|
||
knownDuration <= caps.max_audio_minutes * 60;
|
||
if (!withinCapability) {
|
||
log(
|
||
1,
|
||
`Relay-URL fast-path skipped — known duration ${(knownDuration / 60).toFixed(1)}min exceeds relay capability ${caps.max_audio_minutes}min`
|
||
);
|
||
} else {
|
||
const mediaType = isPodcast ? "podcast" : "youtube";
|
||
const handoffUrl = isPodcast
|
||
? url
|
||
: `https://www.youtube.com/watch?v=${videoId}`;
|
||
log(
|
||
1,
|
||
`Relay-URL fast-path: handing ${mediaType} URL to relay (saves your upload bandwidth)…`
|
||
);
|
||
try {
|
||
const relayResult = await transcriber.transcribeUrl({
|
||
mediaUrl: handoffUrl,
|
||
mediaType,
|
||
mimeType,
|
||
titleHint: videoTitle,
|
||
channelHint: videoChannel,
|
||
descriptionHint: videoDescription,
|
||
chaptersHint: videoChapters,
|
||
onProgress: (msg) => log(2, msg),
|
||
signal: abortController.signal,
|
||
jobId,
|
||
});
|
||
if (relayResult && relayResult.text) {
|
||
transcriptText = relayResult.text;
|
||
const parsed = parseTimestampedTranscript(transcriptText);
|
||
if (parsed.length === 0) {
|
||
log(
|
||
1,
|
||
`⚠ Relay-URL transcript parsed to zero segments — falling back to local download`
|
||
);
|
||
transcriptText = "";
|
||
entries = null;
|
||
} else {
|
||
entries = parsed;
|
||
}
|
||
if (!entries) {
|
||
// fall through to local-download flow
|
||
} else {
|
||
// Single-segment expansion — same as the local path below.
|
||
// Parakeet-style backends return one big text blob; spread
|
||
// it into per-sentence entries with interpolated timestamps
|
||
// so the analyzer sees more than one section's worth of
|
||
// structure.
|
||
if (
|
||
entries.length === 1 &&
|
||
knownDuration &&
|
||
knownDuration > 30 &&
|
||
(entries[0].text || "").length > 100
|
||
) {
|
||
const synthesized = synthesizeEntriesFromText(
|
||
entries[0].text,
|
||
knownDuration
|
||
);
|
||
if (synthesized.length > 1) {
|
||
log(
|
||
2,
|
||
`Single-segment relay transcript expanded into ${synthesized.length} synthetic entries`
|
||
);
|
||
entries = synthesized;
|
||
}
|
||
}
|
||
txCost = relayResult.cost || txCost;
|
||
log(
|
||
2,
|
||
`Relay-URL transcription complete — ${entries.length} segments, skipping local download`
|
||
);
|
||
}
|
||
} else {
|
||
log(
|
||
1,
|
||
`⚠ Relay-URL returned empty — falling back to local download`
|
||
);
|
||
}
|
||
} catch (err) {
|
||
if (err?.message === CANCELLED_MARK || abortController.signal.aborted) {
|
||
throw err;
|
||
}
|
||
const msg = err?.message || String(err);
|
||
// Surface err.cause too — Node's fetch() wraps low-level
|
||
// failures (DNS, TLS, dead-socket) in a bare "fetch failed"
|
||
// message and stashes the real reason in .cause. Without it
|
||
// the activity log is uninformative on transport errors.
|
||
const cause = err?.cause?.message || err?.cause?.code || err?.cause;
|
||
const detail = cause ? ` (cause: ${String(cause).slice(0, 200)})` : "";
|
||
// Did the relay accept the request and then fail at its
|
||
// backend (Parakeet/Gemma/Gemini), or did we fail before
|
||
// the relay even processed it (network blip, TLS issue,
|
||
// relay unreachable)?
|
||
//
|
||
// Backend failures → DON'T fall back to local download.
|
||
// The relay's /relay/transcribe endpoint would route to
|
||
// the same broken backend, costing 60+ seconds of audio
|
||
// upload bandwidth for the same error.
|
||
//
|
||
// Transport failures → DO fall back. The relay may be
|
||
// intermittently unreachable; upload-path might catch a
|
||
// moment when it works.
|
||
const isBackendFailure =
|
||
/CUDA error|Resource exhausted|Parakeet transcribe|Gemma analyze|model does not exist|model.*not found|backend_error|all attempts failed|HTTP 5\d\d/i.test(
|
||
msg
|
||
);
|
||
if (isBackendFailure) {
|
||
log(
|
||
1,
|
||
`⚠ Relay-URL fast-path failed: ${msg.slice(0, 300)}${detail}`
|
||
);
|
||
log(
|
||
1,
|
||
`Skipping local-download fallback — same relay backend would just fail again. ` +
|
||
`Operator hardware (or Gemini quota) needs attention.`
|
||
);
|
||
sendEvent(res, "error", {
|
||
message: msg.slice(0, 500),
|
||
});
|
||
res.end();
|
||
if (isFree) releaseFreeSlot();
|
||
return;
|
||
}
|
||
log(
|
||
1,
|
||
`⚠ Relay-URL fast-path failed: ${msg.slice(0, 200)}${detail} — falling back to local download (transport-level error, relay may be intermittent)`
|
||
);
|
||
}
|
||
}
|
||
}
|
||
|
||
checkCancelled();
|
||
|
||
// ── Step 1: Download audio (skipped when captions populated entries) ──
|
||
if (!entries) {
|
||
const dlStart = Date.now();
|
||
|
||
if (isPodcast) {
|
||
log(1, "Downloading podcast episode...");
|
||
await downloadPodcastAudio(url, audioPath);
|
||
|
||
const stats = await fs.stat(audioPath);
|
||
const sizeMB = (stats.size / (1024 * 1024)).toFixed(1);
|
||
const dlTime = ((Date.now() - dlStart) / 1000).toFixed(1);
|
||
log(1, `Episode downloaded (${sizeMB} MB) in ${dlTime}s`, `File size: ${sizeMB} MB`);
|
||
log(1, `Episode: ${videoTitle}`);
|
||
|
||
} else {
|
||
log(1, "Downloading audio from YouTube...");
|
||
|
||
const dlBaseArgs = [
|
||
"-x",
|
||
"--audio-format", "mp3",
|
||
"--audio-quality", "5",
|
||
"-o", audioPath,
|
||
"--no-playlist",
|
||
"--print", "%(title)s|||%(upload_date)s",
|
||
"--no-simulate",
|
||
];
|
||
const dlUrl = `https://www.youtube.com/watch?v=${videoId}`;
|
||
const cookieArgs = ytCookieArgs();
|
||
const hasCookies = cookieArgs.length > 0;
|
||
let usedCookies = false;
|
||
let dlStdout = "";
|
||
|
||
// Helper: attempt a single yt-dlp download
|
||
async function attemptDownload(args, label) {
|
||
const result = await execFileAsync("yt-dlp", args, { timeout: 600000 });
|
||
return result.stdout || "";
|
||
}
|
||
|
||
// Helper: check if error is a bot detection / rate limit block
|
||
function isBotBlock(errText) {
|
||
return /Sign in|confirm you're not a bot|bot detection|JsChallengeProvider|js.*challenge|HTTP Error 403|Too Many Requests|429/i.test(errText);
|
||
}
|
||
|
||
// ── Smart download with retry ──
|
||
// Strategy: cookies → no-cookies → wait & retry (up to 3 attempts with increasing delays)
|
||
const MAX_RETRIES = 3;
|
||
const RETRY_DELAYS = [30, 60, 120]; // seconds — escalating backoff
|
||
let downloaded = false;
|
||
let lastError = "";
|
||
|
||
for (let attempt = 0; attempt <= MAX_RETRIES && !downloaded; attempt++) {
|
||
// On retry attempts, wait before trying again
|
||
if (attempt > 0) {
|
||
const waitSec = RETRY_DELAYS[Math.min(attempt - 1, RETRY_DELAYS.length - 1)];
|
||
log(1, `⏳ YouTube is rate-limiting. Waiting ${waitSec}s before retry ${attempt}/${MAX_RETRIES}...`);
|
||
sendEvent(res, "status", { step: 1, message: `Rate limited — retrying in ${waitSec}s (attempt ${attempt}/${MAX_RETRIES})` });
|
||
await new Promise(r => setTimeout(r, waitSec * 1000));
|
||
log(1, `Retrying download (attempt ${attempt}/${MAX_RETRIES})...`);
|
||
// Clean up any partial file from previous attempt
|
||
await fs.unlink(audioPath).catch(() => {});
|
||
}
|
||
|
||
// Try with cookies first
|
||
if (hasCookies && !usedCookies) {
|
||
try {
|
||
log(1, attempt === 0 ? "Trying download with browser cookies (ad-free)..." : "Retrying with cookies...");
|
||
dlStdout = await attemptDownload([...dlBaseArgs, ...cookieArgs, dlUrl], "cookies");
|
||
usedCookies = true;
|
||
downloaded = true;
|
||
break;
|
||
} catch (cookieErr) {
|
||
const cookieMsg = (cookieErr.stderr || "") + " " + (cookieErr.message || "");
|
||
if (attempt === 0) log(1, `⚠ Cookie download failed: ${cookieMsg.trim().slice(0, 200)}`);
|
||
log(1, "Retrying without cookies...");
|
||
await fs.unlink(audioPath).catch(() => {});
|
||
}
|
||
}
|
||
|
||
// Try without cookies
|
||
if (!downloaded) {
|
||
try {
|
||
dlStdout = await attemptDownload([...dlBaseArgs, dlUrl], "no-cookies");
|
||
downloaded = true;
|
||
break;
|
||
} catch (dlErr) {
|
||
lastError = (dlErr.stderr || "") + " " + (dlErr.stdout || "") + " " + (dlErr.message || "");
|
||
const blocked = isBotBlock(lastError);
|
||
|
||
if (blocked && attempt < MAX_RETRIES) {
|
||
log(1, `⚠ YouTube bot detection triggered`);
|
||
// Will loop back and wait
|
||
continue;
|
||
}
|
||
|
||
if (blocked && attempt === MAX_RETRIES) {
|
||
// Last resort: try yt-dlp auto-update in case there's a newer version that handles this
|
||
log(1, "All retries exhausted — attempting yt-dlp auto-update as last resort...");
|
||
const updateResult = await autoUpdateYtdlp(DATA_DIR);
|
||
if (updateResult.success) {
|
||
log(1, "yt-dlp updated! Final retry...");
|
||
try {
|
||
const retryResult = await attemptDownload([...dlBaseArgs, dlUrl], "post-update");
|
||
dlStdout = retryResult;
|
||
downloaded = true;
|
||
break;
|
||
} catch { /* fall through to error */ }
|
||
}
|
||
}
|
||
|
||
// Non-bot error or exhausted retries
|
||
if (!downloaded) {
|
||
log(1, `⚠ yt-dlp error: ${lastError.trim().slice(0, 300)}`);
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
if (!downloaded) {
|
||
const blocked = isBotBlock(lastError);
|
||
let hint = "";
|
||
if (blocked) {
|
||
hint = "\n\nYouTube is temporarily blocking downloads from your IP address. " +
|
||
"This is usually caused by:\n" +
|
||
"• Recent VPN use (YouTube flags VPN IPs)\n" +
|
||
"• Too many downloads in a short period\n" +
|
||
"• YouTube's general anti-bot measures\n\n" +
|
||
"What to try:\n" +
|
||
"• Wait 10-30 minutes and try again\n" +
|
||
"• Disconnect any VPN/proxy\n" +
|
||
"• Upload fresh cookies.txt via Settings\n" +
|
||
"• Try a different network (mobile hotspot, etc.)";
|
||
}
|
||
throw new Error(`Download failed after ${MAX_RETRIES} retries.${hint}\n\nLast error: ${lastError.trim().slice(0, 300)}`);
|
||
}
|
||
|
||
if (!usedCookies && hasCookies) {
|
||
log(1, "⚠ Downloaded without cookies — audio may contain ads");
|
||
}
|
||
|
||
const stats = await fs.stat(audioPath);
|
||
const sizeMB = (stats.size / (1024 * 1024)).toFixed(1);
|
||
const dlTime = ((Date.now() - dlStart) / 1000).toFixed(1);
|
||
log(1, `Audio downloaded (${sizeMB} MB) in ${dlTime}s`, `File size: ${sizeMB} MB`);
|
||
|
||
// Extract title from the --print output of the download command
|
||
const fallbackTitle = videoTitle !== "Untitled" ? videoTitle : null;
|
||
let gotTitle = false;
|
||
|
||
// First try: parse title from the download stdout (most reliable — same call that succeeded)
|
||
if (dlStdout) {
|
||
const printLines = dlStdout.split("\n").map(l => l.trim()).filter(Boolean);
|
||
for (const line of printLines) {
|
||
if (line.includes("|||")) {
|
||
const sep = line.indexOf("|||");
|
||
const t = line.slice(0, sep).trim();
|
||
const d = line.slice(sep + 3).trim();
|
||
if (t && t !== "NA") {
|
||
videoTitle = t;
|
||
if (d && d !== "NA") videoUploadDate = d;
|
||
gotTitle = true;
|
||
log(1, `Video title: ${videoTitle}`);
|
||
break;
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
// Second try: separate yt-dlp call (no cookies needed for public metadata)
|
||
if (!gotTitle) {
|
||
try {
|
||
const { stdout } = await execFileAsync("yt-dlp", [
|
||
"--print", "%(title)s|||%(upload_date)s",
|
||
"--no-download",
|
||
`https://www.youtube.com/watch?v=${videoId}`,
|
||
], { timeout: 15000 });
|
||
const raw = stdout.trim();
|
||
const sep = raw.indexOf("|||");
|
||
if (sep > 0) {
|
||
videoTitle = raw.slice(0, sep).trim() || fallbackTitle || "Untitled";
|
||
const d = raw.slice(sep + 3).trim();
|
||
if (d && d !== "NA") videoUploadDate = d || videoUploadDate;
|
||
} else {
|
||
videoTitle = raw || fallbackTitle || "Untitled";
|
||
}
|
||
gotTitle = videoTitle !== "Untitled";
|
||
if (gotTitle) log(1, `Video title: ${videoTitle}`);
|
||
} catch {
|
||
// Title fetch failed
|
||
}
|
||
}
|
||
|
||
// Third try: use the queue-provided title
|
||
if (!gotTitle && fallbackTitle) {
|
||
videoTitle = fallbackTitle;
|
||
log(1, `Using queue title: ${fallbackTitle}`);
|
||
gotTitle = true;
|
||
}
|
||
|
||
if (!gotTitle) {
|
||
log(1, "⚠ Could not fetch video title");
|
||
}
|
||
}
|
||
|
||
checkCancelled();
|
||
|
||
// ── Step 2: Transcribe audio ──
|
||
// Detect audio duration to choose strategy
|
||
const audioDuration = await getAudioDuration(audioPath);
|
||
const audioDurMin = audioDuration ? (audioDuration / 60).toFixed(1) : "unknown";
|
||
log(2, `Audio duration: ${audioDuration ? formatTime(Math.floor(audioDuration)) : "unknown"} (${audioDurMin} min)`);
|
||
|
||
// Strategy:
|
||
// < threshold → full file (fast, cheap, reliable on capable backends)
|
||
// ≥ threshold → chunked transcription
|
||
// If full-file transcription is truncated or empty → fall back to chunks
|
||
//
|
||
// Thresholds are provider-aware. For the relay provider, we ask
|
||
// the relay (via /relay/capabilities) what limits it can accept —
|
||
// if the operator has routed transcription to their own Parakeet
|
||
// hardware (which handles 2+ hour podcasts in one shot), the
|
||
// relay reports very large limits and we skip chunking entirely.
|
||
// For other providers, we use the legacy Gemini-safe defaults.
|
||
let chunkTimeThresholdSec = 60 * 60; // 60 min
|
||
let chunkSizeThresholdBytes = 30 * 1024 * 1024; // 30 MB
|
||
let preferredChunkSeconds = 2700; // 45 min
|
||
if (transcriptionProvider === "relay") {
|
||
// Force a fresh fetch so the chunking decision matches the
|
||
// routing decision the relay will make for THIS install RIGHT
|
||
// NOW. The hourly background refresh is a fallback for when
|
||
// this inline fetch fails (relay unreachable, etc.).
|
||
await refreshRelayCapabilities().catch(() => {});
|
||
const caps = getRelayCapabilities();
|
||
if (typeof caps.max_audio_minutes === "number") {
|
||
chunkTimeThresholdSec = caps.max_audio_minutes * 60;
|
||
}
|
||
if (typeof caps.max_audio_mb === "number") {
|
||
chunkSizeThresholdBytes = caps.max_audio_mb * 1024 * 1024;
|
||
}
|
||
if (caps.preferred_chunk_seconds === null) {
|
||
// Hardware-capable: never chunk. Set both thresholds to
|
||
// Infinity so the needsChunking check stays simple.
|
||
chunkTimeThresholdSec = Infinity;
|
||
chunkSizeThresholdBytes = Infinity;
|
||
preferredChunkSeconds = null;
|
||
} else if (typeof caps.preferred_chunk_seconds === "number") {
|
||
preferredChunkSeconds = caps.preferred_chunk_seconds;
|
||
}
|
||
// Plain-language line for the activity log. The full caps
|
||
// payload (tier, backend pref, chunking thresholds) is verbose
|
||
// and internal — surface only what a non-developer reader
|
||
// benefits from: the max audio length the relay will accept,
|
||
// since that's what dictates the chunking behavior they see.
|
||
if (caps.max_audio_minutes) {
|
||
log(2, `Relay will transcribe up to ${caps.max_audio_minutes} minutes per upload.`);
|
||
} else {
|
||
log(2, `Relay ready to transcribe.`);
|
||
}
|
||
}
|
||
let audioFileSize = 0;
|
||
try { audioFileSize = (await fs.stat(audioPath)).size; } catch {}
|
||
const audioSizeMB = (audioFileSize / (1024 * 1024)).toFixed(1);
|
||
const needsChunking =
|
||
(audioDuration && audioDuration >= chunkTimeThresholdSec) ||
|
||
audioFileSize >= chunkSizeThresholdBytes;
|
||
if (needsChunking) {
|
||
const reason = audioDuration >= chunkTimeThresholdSec ? `${audioDurMin} min` : `${audioSizeMB} MB`;
|
||
log(2, `Large audio (${reason}) — will use chunked transcription with ${transcriber.name}/${transcriptionModel}`);
|
||
} else if (transcriptionProvider === "relay" && preferredChunkSeconds === null) {
|
||
log(2, `Sending full file to relay (relay will chunk server-side if needed for its backend)`);
|
||
}
|
||
|
||
// Transcription model fallback chain: user's chosen model first,
|
||
// then the rest of the provider's list. If Gemini 3 Flash hits a
|
||
// 503 capacity error, the wrapper transparently retries with
|
||
// 2.5 Flash. Matches the analysis fallback pattern below.
|
||
const transcriptionFallbacks = [
|
||
transcriptionModel,
|
||
...transcriber.listTranscriptionModels().filter((m) => m !== transcriptionModel),
|
||
];
|
||
|
||
// Thin wrapper: keeps the call-site shape the chunking + main
|
||
// pipeline already use. `transcriber.transcribeAudio` returns
|
||
// { text, usage, cost, finishReason, blockReason, raw } — callers
|
||
// read off that normalized shape regardless of which provider is
|
||
// doing the transcription. On hard failure (after the provider's
|
||
// own retry loop), walks the fallback chain to the next model.
|
||
async function transcribeSingleFile(filePath, mType, titleHint, modelName, offsetSeconds = 0) {
|
||
// Build the per-call fallback list: caller's chosen model
|
||
// first, then the others. (Same as transcriptionFallbacks but
|
||
// honors per-chunk overrides like chunkResult retries.)
|
||
const chain = [modelName, ...transcriptionFallbacks.filter((m) => m !== modelName)];
|
||
let lastErr;
|
||
for (const tryModel of chain) {
|
||
try {
|
||
return await transcriber.transcribeAudio({
|
||
filePath,
|
||
mimeType: mType,
|
||
titleHint,
|
||
// Rich-context hints used by the provider's prompt builder
|
||
// to anchor speaker-name extraction. Provider implementations
|
||
// that don't care (e.g. Whisper) ignore these.
|
||
channelHint: videoChannel,
|
||
descriptionHint: videoDescription,
|
||
chaptersHint: videoChapters,
|
||
model: tryModel,
|
||
offsetSeconds,
|
||
onProgress: (msg) => log(2, msg),
|
||
signal: abortController.signal,
|
||
// Shared with the analyze call below so the relay bundles
|
||
// both into one credit charge. Non-relay providers ignore it.
|
||
jobId,
|
||
});
|
||
} catch (err) {
|
||
// Cancellation: bail immediately, don't try the next model.
|
||
if (abortController.signal.aborted || err?.name === "AbortError") {
|
||
throw new Error(CANCELLED_MARK);
|
||
}
|
||
lastErr = err;
|
||
const msg = err?.message || String(err);
|
||
log(2, `⚠ Transcription with ${tryModel} failed: ${msg.slice(0, 150)}`);
|
||
if (tryModel !== chain[chain.length - 1]) {
|
||
log(2, `Falling back to next transcription model...`);
|
||
}
|
||
}
|
||
}
|
||
throw lastErr || new Error("All transcription models failed");
|
||
}
|
||
|
||
// ── Helper: chunked transcription for long audio ──
|
||
async function transcribeChunked(srcPath, srcMime, title, modelName, logFn) {
|
||
const chunkDir = path.join(os.tmpdir(), `yt-chunks-${Date.now()}`);
|
||
await fs.mkdir(chunkDir, { recursive: true });
|
||
|
||
try {
|
||
// Honor the provider's preferred chunk size (relay reports
|
||
// this via /relay/capabilities; legacy defaults to 45 min).
|
||
const chunkSec = preferredChunkSeconds || 2700;
|
||
const audioChunks = await splitAudioFile(srcPath, chunkDir, chunkSec);
|
||
if (!audioChunks || audioChunks.length <= 1) return null; // splitting not needed
|
||
|
||
logFn(`Split audio into ${audioChunks.length} chunks for transcription (parallel, up to ${Math.min(audioChunks.length, 6)} in flight)`);
|
||
let totalIn = 0, totalOut = 0;
|
||
const perChunkEntries = new Array(audioChunks.length).fill(null);
|
||
|
||
// Fire chunks in parallel with a bounded in-flight count.
|
||
// 6 is well under Gemini's paid-tier RPM (1k for flash, 150
|
||
// for pro) and stays comfortable on the relay (which serves
|
||
// each upload independently). Out-of-order completion is
|
||
// fine — we sort by chunk index before merging at the end.
|
||
const CHUNK_CONCURRENCY = 6;
|
||
let nextChunkIdx = 0;
|
||
const transcribeOne = async () => {
|
||
while (true) {
|
||
const my = nextChunkIdx++;
|
||
if (my >= audioChunks.length) return;
|
||
const chunk = audioChunks[my];
|
||
logFn(`Transcribing chunk ${chunk.index + 1}/${audioChunks.length} (starts at ${formatTime(chunk.startOffset)})...`);
|
||
try {
|
||
const chunkResult = await transcribeSingleFile(
|
||
chunk.path, "audio/mpeg", title,
|
||
modelName,
|
||
chunk.startOffset
|
||
);
|
||
totalIn += chunkResult.cost.inputTokens;
|
||
totalOut += chunkResult.cost.outputTokens;
|
||
const chunkText = chunkResult.text;
|
||
if (!chunkText) {
|
||
logFn(`⚠ Chunk ${chunk.index + 1} returned empty response — skipping`);
|
||
perChunkEntries[my] = [];
|
||
continue;
|
||
}
|
||
const chunkEntries = parseTimestampedTranscript(chunkText);
|
||
if (chunk.startOffset > 0) {
|
||
const firstOffset = chunkEntries.length > 0 ? chunkEntries[0].offset : 0;
|
||
const alreadyAdjusted = firstOffset >= chunk.startOffset * 0.8;
|
||
if (!alreadyAdjusted) {
|
||
for (const e of chunkEntries) {
|
||
e.offset += chunk.startOffset;
|
||
}
|
||
logFn(`Adjusted chunk ${chunk.index + 1} timestamps by +${formatTime(chunk.startOffset)}`);
|
||
}
|
||
}
|
||
// Sanity cap: drop any entries whose absolute offset
|
||
// exceeds the chunk's true time window. Observed bug
|
||
// (May 2026): gemini-3.1-flash-lite emitted timestamps
|
||
// like [10:12:44] on a 45-min chunk, which then survived
|
||
// the merge step and caused the stitched transcript to
|
||
// claim 10:12:44 total length AND wiped out subsequent
|
||
// chunks' entries (the merge dedupe rule dropped anything
|
||
// <= the running max offset). Small tolerance (10s) for
|
||
// models that mark the END of a phrase rather than the
|
||
// start, where the actual emission can slightly exceed
|
||
// the chunk boundary.
|
||
const chunkEndAbsolute = chunk.startOffset + chunk.durationSec;
|
||
const TOLERANCE_SEC = 10;
|
||
const validUpper = chunkEndAbsolute + TOLERANCE_SEC;
|
||
const filtered = chunkEntries.filter((e) => e.offset <= validUpper);
|
||
const dropped = chunkEntries.length - filtered.length;
|
||
if (dropped > 0) {
|
||
const worstOffset = Math.max(...chunkEntries.map((e) => e.offset || 0));
|
||
logFn(
|
||
`⚠ Chunk ${chunk.index + 1}: dropped ${dropped} segment(s) with bogus timestamps past chunk end ${formatTime(chunkEndAbsolute)} (worst: ${formatTime(worstOffset)}). Model: ${modelName}`
|
||
);
|
||
}
|
||
// Sort by offset before handing off to the merge step.
|
||
// Some models emit segments out of chronological order
|
||
// (observed on noisy transcribe outputs); the merge step
|
||
// relies on monotonically-increasing offsets to dedupe
|
||
// boundary overlaps. Sorting here guarantees that
|
||
// invariant per chunk.
|
||
filtered.sort((a, b) => (a.offset || 0) - (b.offset || 0));
|
||
// Detect truncated transcripts. If the last entry's
|
||
// absolute offset is significantly less than the chunk's
|
||
// expected end, the model probably hit its output-token
|
||
// cap mid-transcript and lost the tail. Surface as a
|
||
// loud warning so the operator knows which chunk + model
|
||
// misbehaved and can switch to a model with bigger
|
||
// output capacity. We keep the partial entries because
|
||
// they're better than nothing, but the operator should
|
||
// know they're looking at incomplete content.
|
||
const lastAbsolute = filtered.length > 0
|
||
? filtered[filtered.length - 1].offset
|
||
: chunk.startOffset;
|
||
const expectedEnd = chunk.startOffset + chunk.durationSec;
|
||
const coverageRatio = chunk.durationSec > 0
|
||
? (lastAbsolute - chunk.startOffset) / chunk.durationSec
|
||
: 1;
|
||
if (coverageRatio < 0.8 && chunk.durationSec > 60) {
|
||
const missingSec = expectedEnd - lastAbsolute;
|
||
logFn(
|
||
`⚠ Chunk ${chunk.index + 1}: transcript appears TRUNCATED — last entry at ${formatTime(lastAbsolute)} but chunk ends at ${formatTime(expectedEnd)} (missing ${formatTime(missingSec)} of speech). Model: ${modelName}. Likely cause: model hit maxOutputTokens. Consider a model with larger output capacity, or shrink chunk size.`
|
||
);
|
||
}
|
||
logFn(`Chunk ${chunk.index + 1}: ${filtered.length} segments, last timestamp ${filtered.length > 0 ? formatTime(filtered[filtered.length - 1].offset) : "N/A"} (coverage ${(coverageRatio * 100).toFixed(0)}%)`);
|
||
perChunkEntries[my] = filtered;
|
||
} catch (err) {
|
||
logFn(`⚠ Chunk ${chunk.index + 1} failed: ${err?.message?.slice(0, 120) || err}`);
|
||
perChunkEntries[my] = [];
|
||
}
|
||
}
|
||
};
|
||
const workers = Array.from(
|
||
{ length: Math.min(audioChunks.length, CHUNK_CONCURRENCY) },
|
||
transcribeOne
|
||
);
|
||
await Promise.all(workers);
|
||
|
||
// Merge chunks in chronological order. Per-chunk timestamps
|
||
// are absolute (we adjusted by startOffset above), so merging
|
||
// in index order is the same as in time order. The overlap-
|
||
// skip rule preserves the original behavior: any entry whose
|
||
// offset has already been covered by a prior chunk is dropped.
|
||
const allEntries = [];
|
||
for (const chunkEntries of perChunkEntries) {
|
||
if (!chunkEntries) continue;
|
||
const lastExistingTime = allEntries.length > 0 ? allEntries[allEntries.length - 1].offset : -1;
|
||
for (const e of chunkEntries) {
|
||
if (e.offset > lastExistingTime) allEntries.push(e);
|
||
}
|
||
}
|
||
|
||
// Recalculate durations
|
||
for (let i = 0; i < allEntries.length - 1; i++) {
|
||
allEntries[i].duration = allEntries[i + 1].offset - allEntries[i].offset;
|
||
}
|
||
if (allEntries.length > 0) allEntries[allEntries.length - 1].duration = 15;
|
||
|
||
logFn(`Chunked transcription complete: ${allEntries.length} total segments`);
|
||
return {
|
||
entries: allEntries,
|
||
cost: {
|
||
inputTokens: totalIn, outputTokens: totalOut, thinkingTokens: 0,
|
||
totalTokens: totalIn + totalOut,
|
||
totalCost: "0", totalCostDisplay: "",
|
||
},
|
||
};
|
||
} finally {
|
||
try { await fs.rm(chunkDir, { recursive: true, force: true }); } catch {}
|
||
}
|
||
}
|
||
|
||
// entries / transcriptText / txCost are declared earlier (top of
|
||
// the request handler) since the captions-fast-path needs to
|
||
// populate them before this audio-transcription block runs.
|
||
const txStart = Date.now();
|
||
|
||
if (needsChunking) {
|
||
// ── Very long audio: go straight to chunked transcription ──
|
||
log(2, `Skipping full-file attempt — using chunked transcription for ${audioDurMin} min audio`);
|
||
const chunkedResult = await transcribeChunked(audioPath, mimeType, videoTitle, transcriptionModel, (msg) => log(2, msg));
|
||
if (chunkedResult && chunkedResult.entries.length > 0) {
|
||
entries = chunkedResult.entries;
|
||
txCost = chunkedResult.cost;
|
||
} else {
|
||
log(2, `⚠ Chunked transcription returned no entries, trying full file as last resort...`);
|
||
entries = null; // fall through to full-file attempt below
|
||
}
|
||
}
|
||
|
||
if (!entries) {
|
||
// ── Normal: transcribe full file ──
|
||
const transcriptResult = await transcribeSingleFile(audioPath, mimeType, videoTitle, transcriptionModel);
|
||
|
||
transcriptText = transcriptResult.text;
|
||
if (!transcriptText) {
|
||
log(2, `⚠ Full-file transcription returned empty — falling back to chunked transcription...`);
|
||
const chunkedResult = await transcribeChunked(audioPath, mimeType, videoTitle, transcriptionModel, (msg) => log(2, msg));
|
||
if (chunkedResult && chunkedResult.entries.length > 0) {
|
||
entries = chunkedResult.entries;
|
||
txCost = chunkedResult.cost;
|
||
} else {
|
||
throw new Error(`${transcriber.name} returned empty transcription for both full file and chunked attempts. Try again or use a shorter video.`);
|
||
}
|
||
} else {
|
||
txCost = transcriptResult.cost;
|
||
const txTime = ((Date.now() - txStart) / 1000).toFixed(1);
|
||
log(2, `Transcription complete in ${txTime}s`, `${transcriptText.length} chars received`);
|
||
|
||
entries = parseTimestampedTranscript(transcriptText);
|
||
log(2, `Parsed ${entries.length} transcript segments`);
|
||
}
|
||
|
||
// ── Single-segment expansion ──
|
||
// Whisper-API endpoints that don't return per-segment timestamps
|
||
// (e.g. NVIDIA Parakeet, some bare-Whisper wrappers) hand back a
|
||
// single text blob. parseTimestampedTranscript dumps that into
|
||
// one entry at [0:00], which (a) trips the truncation detector
|
||
// below and (b) leaves the analyzer with one giant segment so
|
||
// it can only produce one section. Synthesize sentence-based
|
||
// entries with interpolated timestamps so both code paths
|
||
// downstream work like they do for Gemini/Whisper-1.
|
||
if (entries.length === 1 && audioDuration && audioDuration > 30 && (entries[0].text || "").length > 100) {
|
||
const synthesized = synthesizeEntriesFromText(entries[0].text, audioDuration);
|
||
if (synthesized.length > 1) {
|
||
log(2, `Single-segment transcript expanded into ${synthesized.length} synthetic sentence-based entries with interpolated timestamps`);
|
||
entries = synthesized;
|
||
}
|
||
}
|
||
|
||
// ── Truncation detection → fall back to chunks ──
|
||
if (audioDuration && entries.length > 0) {
|
||
const lastEntryTime = entries[entries.length - 1].offset;
|
||
const coverageRatio = lastEntryTime / audioDuration;
|
||
const missingSeconds = audioDuration - lastEntryTime;
|
||
|
||
if (coverageRatio < 0.90 && missingSeconds > 120) {
|
||
log(2, `⚠ Transcript truncated — covers ${formatTime(lastEntryTime)} of ${formatTime(Math.floor(audioDuration))} (${(coverageRatio * 100).toFixed(0)}%)`);
|
||
log(2, `Falling back to chunked transcription...`);
|
||
const chunkedResult = await transcribeChunked(audioPath, mimeType, videoTitle, transcriptionModel, (msg) => log(2, msg));
|
||
if (chunkedResult && chunkedResult.entries.length > 0) {
|
||
entries = chunkedResult.entries;
|
||
txCost = chunkedResult.cost;
|
||
const finalCoverage = entries[entries.length - 1].offset;
|
||
log(2, `Coverage after chunking: ${formatTime(finalCoverage)} of ${formatTime(Math.floor(audioDuration))}`);
|
||
}
|
||
} else {
|
||
log(2, `Transcript coverage: ${formatTime(lastEntryTime)} of ${formatTime(Math.floor(audioDuration))} (${(coverageRatio * 100).toFixed(0)}%) — OK`);
|
||
}
|
||
}
|
||
}
|
||
|
||
const txTotalTime = ((Date.now() - txStart) / 1000).toFixed(1);
|
||
log(2, `Total transcription time: ${txTotalTime}s — ${entries.length} segments`);
|
||
|
||
if (!entries || entries.length === 0) {
|
||
const preview = (transcriptText || "").slice(0, 500).replace(/\n/g, " ↵ ");
|
||
log(2, `⚠ Transcript parse failed. Preview: ${preview}`);
|
||
sendEvent(res, "error", { message: "Failed to parse transcript." });
|
||
sendEvent(res, "result", { videoId, entries: [], chunks: [], rawTranscript: transcriptText });
|
||
res.end();
|
||
return;
|
||
}
|
||
|
||
} // end if (!entries) — close the audio + transcription block
|
||
|
||
if (!entries || entries.length === 0) {
|
||
throw new Error("No transcript available — neither captions nor audio transcription produced segments.");
|
||
}
|
||
|
||
checkCancelled();
|
||
|
||
// ── Step 3: Topic analysis with model fallback ──
|
||
// Fallback order: caller's chosen model first, then the rest of the
|
||
// analyzer provider's analysis-model list.
|
||
const providerModels = analyzer.listAnalysisModels();
|
||
const analysisFallbacks = [
|
||
analysisModel,
|
||
...providerModels.filter((m) => m !== analysisModel),
|
||
];
|
||
|
||
// ── Auto-coalesce input for the analysis LLM (single-shot path only) ──
|
||
// Long content uses the chunked path (each ~18-min window only sees
|
||
// a small slice of entries, so context isn't a concern) — coalescing
|
||
// there would just degrade section-boundary precision unnecessarily.
|
||
// Short content still goes through a single analyze call; for tiny
|
||
// local models with 8-16K context windows, a dense Parakeet
|
||
// transcript at full granularity can still overflow, so we keep
|
||
// the safety-net coalesce for that path.
|
||
const lastEntry = entries[entries.length - 1];
|
||
const totalSec = (lastEntry?.offset || 0) + (lastEntry?.duration || 0);
|
||
const willChunkAnalysis = totalSec > CHUNKING_CUTOFF_SECONDS;
|
||
const { coalesced: analysisEntries, indexMap: analysisIndexMap } =
|
||
willChunkAnalysis
|
||
? { coalesced: entries, indexMap: null }
|
||
: coalesceForAnalysis(entries, 400);
|
||
if (analysisEntries !== entries) {
|
||
log(
|
||
3,
|
||
`⚠ Single-shot path: coalesced ${entries.length} → ${analysisEntries.length} segments for analyzer (transcript display keeps full granularity).`
|
||
);
|
||
} else if (willChunkAnalysis) {
|
||
log(
|
||
3,
|
||
`Chunked path: keeping full segment granularity (${entries.length} entries) — chunker handles per-window context.`
|
||
);
|
||
}
|
||
|
||
// Helper: convert sections (in analysisEntries index space) into
|
||
// Recap "chunks" (full UI shape with entries slice + startTime).
|
||
// Used by both the per-window streaming callback AND the final
|
||
// assembly below — keeps the two paths in lockstep so what
|
||
// streams matches what gets saved.
|
||
const sectionsToChunks = (sections) =>
|
||
sections
|
||
.map((section) => {
|
||
let start = Math.max(0, section.startIndex);
|
||
let end = section.endIndex;
|
||
if (analysisIndexMap) {
|
||
const mappedStart = analysisIndexMap[Math.max(0, Math.min(start, analysisIndexMap.length - 1))];
|
||
const mappedEnd = analysisIndexMap[Math.max(0, Math.min(end, analysisIndexMap.length - 1))];
|
||
start = mappedStart ? mappedStart.startOrig : 0;
|
||
end = mappedEnd ? mappedEnd.endOrig : entries.length - 1;
|
||
}
|
||
start = Math.max(0, start);
|
||
end = Math.min(entries.length - 1, end);
|
||
const sectionEntries = entries.slice(start, end + 1);
|
||
return {
|
||
title: section.title,
|
||
summary: section.summary,
|
||
entries: sectionEntries,
|
||
startTime: sectionEntries[0]?.offset || 0,
|
||
};
|
||
})
|
||
.filter((c) => c.entries.length > 0);
|
||
|
||
// Emit transcript_ready so the browser can switch from the
|
||
// loading screen to the results view immediately — transcript
|
||
// pane populated, sections pane in "analyzing…" state. Sections
|
||
// then stream in via sections_partial events as analyze windows
|
||
// complete (only the chunked path streams; short content gets
|
||
// its sections in the final result event).
|
||
const contentType = isPodcast ? "podcast" : "youtube";
|
||
sendEvent(res, "transcript_ready", {
|
||
videoId,
|
||
videoTitle,
|
||
videoChannel,
|
||
videoChapters,
|
||
videoUploadDate,
|
||
videoDescription,
|
||
entries,
|
||
type: contentType,
|
||
willChunk: willChunkAnalysis,
|
||
});
|
||
|
||
// Single entry point: runChunkedAnalysis decides per-content
|
||
// whether to single-shot (≤25 min) or window-and-stitch (>25 min).
|
||
// Per-window concurrency is bounded inside; model fallback walks
|
||
// the analysisFallbacks list within each window independently.
|
||
let analysisResult = null;
|
||
let usedAnalysisModel = analysisModel;
|
||
const anaStart = Date.now();
|
||
log(3, `Analyzing topics across ${analysisEntries.length} segments with ${analyzer.name}...`);
|
||
try {
|
||
analysisResult = await runChunkedAnalysis({
|
||
entries: analysisEntries,
|
||
analyzer,
|
||
fallbackModels: analysisFallbacks,
|
||
onProgress: (msg) => log(3, msg),
|
||
onWindowComplete: ({ windowIdx, totalWindows, ownedSections }) => {
|
||
// Convert this window's owned sections into the final
|
||
// chunk shape and stream them to the browser. The stitcher's
|
||
// "trust the next window for the overlap region" rule was
|
||
// applied per-window inside runChunkedAnalysis, so what we
|
||
// emit here is final — the UI won't have to revise it.
|
||
const partialChunks = sectionsToChunks(ownedSections);
|
||
if (partialChunks.length > 0) {
|
||
sendEvent(res, "sections_partial", {
|
||
windowIdx,
|
||
totalWindows,
|
||
chunks: partialChunks,
|
||
});
|
||
}
|
||
},
|
||
signal: abortController.signal,
|
||
jobId,
|
||
});
|
||
usedAnalysisModel = analysisResult.model || analysisModel;
|
||
} catch (err) {
|
||
if (abortController.signal.aborted || err?.name === "AbortError") {
|
||
throw new Error(CANCELLED_MARK);
|
||
}
|
||
throw err;
|
||
}
|
||
|
||
if (!analysisResult) {
|
||
throw new Error("All analysis models failed. Please try again later.");
|
||
}
|
||
|
||
const analysisText = analysisResult.text;
|
||
if (!analysisText) {
|
||
throw new Error(`${analyzer.name} returned an empty analysis. The transcript may be too long for the model. Try again.`);
|
||
}
|
||
const anaTime = ((Date.now() - anaStart) / 1000).toFixed(1);
|
||
const anaCost = analysisResult.cost;
|
||
|
||
// Parse the analysis JSON
|
||
let analysisJson;
|
||
try {
|
||
let jsonStr = analysisText.trim();
|
||
const codeBlockMatch = jsonStr.match(/```(?:json)?\s*([\s\S]*?)```/);
|
||
if (codeBlockMatch) jsonStr = codeBlockMatch[1].trim();
|
||
analysisJson = JSON.parse(jsonStr);
|
||
} catch (e) {
|
||
console.error("Failed to parse analysis JSON:", e.message);
|
||
sendEvent(res, "error", { message: "Topic analysis returned invalid JSON. Returning raw transcript." });
|
||
sendEvent(res, "result", { videoId, entries, chunks: [], rawTranscript: transcriptText });
|
||
res.end();
|
||
return;
|
||
}
|
||
|
||
// Build final chunks via the same helper used for streaming —
|
||
// guarantees what the browser accumulated from sections_partial
|
||
// events matches what gets saved to history and sent in the
|
||
// canonical result event.
|
||
const chunks = sectionsToChunks(analysisJson.sections);
|
||
|
||
const totalTime = ((Date.now() - pipelineStart) / 1000).toFixed(1);
|
||
log(3, `Topic analysis complete in ${anaTime}s — found ${chunks.length} topics`);
|
||
log(3, `Pipeline finished in ${totalTime}s`);
|
||
|
||
// Save to history. As of the library-for-everyone change, free
|
||
// users save too — the app feels broken otherwise (summarize a
|
||
// video, never find it again).
|
||
const historyId = await saveToHistory(scopeForRequest(req), videoId, url, videoTitle, chunks, entries, logHistory, videoUploadDate, contentType).catch(() => null);
|
||
|
||
// Debit local credit on success — trial cookie OR free signed-in tenant.
|
||
if (RECAP_MODE === "multi" && req.trial) {
|
||
try {
|
||
const { debitOne } = await import("./anon-trial.js");
|
||
debitOne(req.trial.cookie_id);
|
||
} catch (err) {
|
||
console.warn("[/api/process] trial debit failed:", err);
|
||
}
|
||
} else if (RECAP_MODE === "multi" && req.creditsToDebit === "tenant") {
|
||
try {
|
||
const { gateAndDebit } = await import("./tenant-credits.js");
|
||
await gateAndDebit(req.user.id);
|
||
} catch (err) {
|
||
console.warn("[/api/process] tenant_credits debit failed:", err);
|
||
}
|
||
}
|
||
|
||
sendEvent(res, "result", { videoId, videoTitle, entries, chunks, historyId, type: contentType });
|
||
res.end();
|
||
|
||
} catch (err) {
|
||
// Treat any AbortError or aborted-signal state as a user cancellation
|
||
// even if it didn't surface as CANCELLED_MARK upstream (e.g. an SDK
|
||
// throws AbortError before checkCancelled() runs).
|
||
const cancelled =
|
||
err?.message === CANCELLED_MARK ||
|
||
err?.name === "AbortError" ||
|
||
abortController.signal.aborted;
|
||
if (!cancelled) {
|
||
// Dump as much context as the runtime gives us. Generic
|
||
// messages like "Error in input stream" are usually wrappers
|
||
// around an underlying SDK / Node stream error; the cause +
|
||
// stack are what actually tells us what happened.
|
||
console.error("Pipeline error:", {
|
||
name: err?.name,
|
||
message: err?.message,
|
||
code: err?.code,
|
||
status: err?.status || err?.statusCode,
|
||
cause: err?.cause,
|
||
stack: err?.stack,
|
||
transcriptionProvider,
|
||
analysisProvider,
|
||
});
|
||
}
|
||
if (!res.headersSent) {
|
||
res.status(cancelled ? 499 : 500).json({
|
||
error: cancelled ? "cancelled" : err.message,
|
||
});
|
||
} else if (cancelled) {
|
||
sendEvent(res, "cancelled", { message: "Cancelled by user" });
|
||
res.end();
|
||
} else {
|
||
sendEvent(res, "error", { message: err.message });
|
||
res.end();
|
||
}
|
||
} finally {
|
||
if (isFree) releaseFreeSlot();
|
||
// Clean up temp files
|
||
try {
|
||
await fs.rm(tmpDir, { recursive: true, force: true });
|
||
} catch {}
|
||
}
|
||
});
|
||
|
||
// ── Helpers ────────────────────────────────────────────────────────────────
|
||
|
||
// getAudioDuration + splitAudioFile moved to ./audio.js
|
||
|
||
// sendEvent / extractVideoId / formatTime / parseTimestampedTranscript moved to ./util.js
|
||
|
||
// buildAnalysisPrompt moved to ./gemini-helpers.js
|
||
|
||
// ── Network mode ──────────────────────────────────────────────────────────
|
||
// On StartOS (DATA_DIR=/data): always bind to 0.0.0.0 (container networking)
|
||
// On local Mac dev: default to localhost (safe on public Wi-Fi)
|
||
// - Your .app launcher sets LAN_MODE=true (Home) or false (Traveling)
|
||
// - Running "npm start" directly defaults to localhost
|
||
const isStartOS = process.env.DATA_DIR && process.env.DATA_DIR !== path.join(__dirname, "..");
|
||
const lanMode = isStartOS ? true : process.env.LAN_MODE === "true";
|
||
const BIND_HOST = lanMode ? "0.0.0.0" : "127.0.0.1";
|
||
|
||
app.get("/api/network-mode", (req, res) => {
|
||
res.json({ lan: lanMode });
|
||
});
|
||
|
||
// ── Start server ───────────────────────────────────────────────────────────
|
||
|
||
app.listen(PORT, BIND_HOST, async () => {
|
||
console.log(`\n Recap API running on http://${BIND_HOST}:${PORT}`);
|
||
console.log(` Data directory: ${DATA_DIR}`);
|
||
console.log(` Recap mode: ${RECAP_MODE}`);
|
||
console.log(` Checking yt-dlp...`);
|
||
|
||
const info = await checkYtdlp();
|
||
if (!info.installed) {
|
||
console.log(` ⚠ yt-dlp not found. Install it: pip install yt-dlp\n`);
|
||
} else if (info.updateAvailable) {
|
||
console.log(` ✓ yt-dlp ${info.version} found`);
|
||
console.log(` ↑ Update available: ${info.latestVersion}`);
|
||
console.log(` Auto-updating...`);
|
||
const result = await autoUpdateYtdlp(DATA_DIR);
|
||
if (result.success) {
|
||
const refreshed = await checkYtdlp();
|
||
console.log(` ✓ yt-dlp updated to ${refreshed.version}\n`);
|
||
} else {
|
||
console.log(` ⚠ Auto-update failed. Run manually: yt-dlp -U\n`);
|
||
}
|
||
} else {
|
||
console.log(` ✓ yt-dlp ${info.version} (up to date)\n`);
|
||
}
|
||
|
||
// One-time, idempotent: migrate any legacy install-wide subscription
|
||
// files into the operator's own scope so storage is uniformly per-scope.
|
||
try {
|
||
const moved = await migrateGlobalSubscriptionsToOwner();
|
||
if (moved > 0) {
|
||
console.log(` 📦 Migrated ${moved} subscription file(s) into the operator scope`);
|
||
}
|
||
} catch (err) {
|
||
console.warn(" ⚠ Subscription migration skipped:", err.message);
|
||
}
|
||
// Prune queued items that have since been summarized, across every scope
|
||
// that has a queue (owner + each subscribing tenant).
|
||
for (const s of await listAutoQueueScopes()) {
|
||
await reconcileAutoQueueWithLibrary(s).catch(() => {});
|
||
}
|
||
|
||
// Check subscriptions on startup (fans out over all subscribing scopes).
|
||
console.log(` 📡 Checking subscriptions for new videos...`);
|
||
await checkSubscriptions().catch(err => console.error(" ⚠ Subscription check error:", err.message));
|
||
|
||
// Recover items stuck in "processing" from a crash → re-queue as approved
|
||
// (across all scopes), then start the processor if anything is ready.
|
||
let anyApproved = false;
|
||
for (const s of await listAutoQueueScopes()) {
|
||
await mutateAutoQueue(s, (items) => {
|
||
for (const item of items) {
|
||
if (item.status === "processing") {
|
||
console.log(` ⚠ Recovering stuck item (${s}): ${item.title}`);
|
||
item.status = "approved";
|
||
}
|
||
}
|
||
if (items.some((q) => q.status === "approved")) anyApproved = true;
|
||
});
|
||
}
|
||
if (anyApproved) {
|
||
console.log(` 🔄 Starting background processor...`);
|
||
// Delay slightly so the server is fully ready before internal HTTP calls
|
||
setTimeout(() => kickProcessor(), 2000);
|
||
}
|
||
|
||
console.log(` ⚙ Processing config: ${processingConfig.delaySeconds}s delay, ${processingConfig.enabled ? "enabled" : "paused"}`);
|
||
|
||
// Check subscriptions every hour (runs continuously on StartOS)
|
||
setInterval(() => {
|
||
checkSubscriptions().catch(err => console.error(" ⚠ Subscription check error:", err.message));
|
||
}, 60 * 60 * 1000);
|
||
});
|