diff --git a/apps/api-daemon/src/youtubePolling.ts b/apps/api-daemon/src/youtubePolling.ts
index ea4b914..77a82cf 100644
--- a/apps/api-daemon/src/youtubePolling.ts
+++ b/apps/api-daemon/src/youtubePolling.ts
@@ -1,12 +1,10 @@
-import { execFile } from "node:child_process";
-import { promisify } from "node:util";
 import { prisma } from "@streamclipper/db";
 import { env } from "./env";
 import { streamIngestQueue } from "./queues";
 import { sendTelegramMessage } from "./telegram";
-import { ytdlpAntiBotArgs } from "./ytdlpCookies";
 
-const execFileAsync = promisify(execFile);
+const BROWSER_USER_AGENT =
+  "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0 Safari/537.36";
 
 /**
  * Live detection failures are kept per channel and surfaced via GET /status
@@ -15,28 +13,45 @@ const execFileAsync = promisify(execFile);
 export const lastPollErrors = new Map<string, string>();
 
 /**
- * A channel's /live page HTML embeds dozens of unrelated `videoId`s (channel
- * video list, recommendations, etc.) — a plain regex grabbed whichever one
- * happened to appear first, which had nothing to do with the actual live
- * broadcast (confirmed: captured recordings didn't match the real stream).
- * yt-dlp's extractor correctly resolves the channel's *current* live video,
- * so the check goes back through it. Crucially, no `-f` format selector is
- * passed here: format resolution is what triggers YouTube's PO-token wall,
- * and this call only needs the video id / live flag, not downloadable
- * formats — those are only resolved later, in the actual capture.
+ * Two earlier approaches both had real problems:
+ *  - A naive regex over the channel /live page HTML grabbed the FIRST
+ *    "videoId" anywhere on the page — which is one of dozens of unrelated
+ *    ids (channel video list, recommendations), not the live broadcast.
+ *    Confirmed: captured recordings didn't match the real stream.
+ *  - Routing the check through yt-dlp resolved the correct video but is
+ *    increasingly blocked by YouTube's bot-check/proof-of-origin-token wall
+ *    on datacenter IPs — even a plain --simulate call, no cookies fix this
+ *    reliably.
+ *
+ * A plain page fetch was never blocked all day — only yt-dlp invocations
+ * were. The fix is extracting the *right* field instead of avoiding the
+ * fetch: YouTube embeds a single `videoDetails` object holding the id/live
+ * flag/channelId of whichever video the page is actually showing (present
+ * only when a channel really is live — verified empty for an offline
+ * channel). Scoping to that one object (and cross-checking channelId)
+ * avoids both the wrong-video bug and yt-dlp's bot-check entirely.
  */
 async function findLiveVideoId(channelId: string): Promise<string | null> {
   const liveUrl = `https://www.youtube.com/channel/${channelId}/live`;
 
   try {
-    const { stdout } = await execFileAsync(
-      "yt-dlp",
-      ["--simulate", "--no-warnings", ...(await ytdlpAntiBotArgs()), "--print", "%(id)s", liveUrl],
-      { timeout: 20_000 },
-    );
+    const res = await fetch(liveUrl, { headers: { "User-Agent": BROWSER_USER_AGENT } });
+    if (!res.ok) {
+      lastPollErrors.set(channelId, `HTTP ${res.status} fetching ${liveUrl}`);
+      return null;
+    }
+
+    const html = await res.text();
     lastPollErrors.delete(channelId);
-    const videoId = stdout.trim().split("\n")[0];
-    return videoId || null;
+
+    const match = html.match(/"videoDetails":\{"videoId":"([^"]+)"/);
+    if (!match) return null;
+
+    const window = html.slice(match.index!, match.index! + 600);
+    if (!/"isLive":true/.test(window)) return null;
+    if (!window.includes(`"channelId":"${channelId}"`)) return null;
+
+    return match[1];
   } catch (err) {
     const message = err instanceof Error ? err.message : String(err);
     lastPollErrors.set(channelId, message.slice(0, 500));