fix(youtube): stop cookie write-back poisoning (netcup bot-detection) + PO token provider + cache dir

CHANGELOG.md

@@ -5,6 +5,35 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.25.0] - 2026-08-22
+
+### Fixed
+- YouTube downloads intermittently failing with "Sign in to confirm you're not a
+  bot" after the move to the netcup datacenter box. Root cause: yt-dlp saves the
+  rotated cookie jar back to the shared `cookiefile` after every run; on the
+  flagged datacenter IP YouTube *rejects* the rotation ("cookies no longer valid …
+  rotated in the browser as a security measure"), so the saved-back cookies are
+  invalid and poison the keeper-managed file for the next download. (On the old
+  GCE IP the rotation was accepted, so the write-back was harmless — which is why
+  it worked for months and only broke after migration.) Each yt-dlp run now
+  operates on a throwaway copy of the cookie file (`isolated_cookiefile` /
+  `ytdlp_opts_isolated`), so the canonical file — written only by the credential
+  keeper — stays pristine and concurrent downloads no longer clobber each other.
+  Applied to the download, availability-check, search, credential-check, and
+  deep-health paths.
+- yt-dlp cache was silently disabled on the server: it runs with
+  `HOME=/opt/flacfetch` where `~/.cache` is the Spotify OAuth token *file*, so
+  yt-dlp's default `~/.cache/yt-dlp` path died with `NotADirectoryError` (forcing
+  a fresh JS-challenge solve every request). A new `FLACFETCH_YTDLP_CACHE_DIR`
+  override gives yt-dlp its own collision-free cache directory.
+
+### Added
+- Proof-of-Origin (PO) token provider for YouTube. YouTube now binds a GVS PO
+  token to downloads from datacenter IPs; flacfetch now runs the
+  `bgutil-ytdlp-pot-provider` HTTP server (Node, localhost) plus the matching
+  yt-dlp plugin, provisioned and kept current by `deploy/provision.sh`. The host
+  nftables firewall keeps the token server localhost-only.
+
 ## [0.23.0] - 2026-08-14
 
 ### Fixed

deploy/provision.sh

@@ -68,6 +68,18 @@ BROWSER_PROFILE_DIR="$DATA_MOUNT/browser-profiles"
 PLAYWRIGHT_BROWSERS_PATH="$APP_DIR/ms-playwright"  # shared, service-user-owned (not /root/.cache)
 DENO_INSTALL=/opt/deno
 LIBRESPOT_BIN=/usr/local/bin/librespot
+# yt-dlp Proof-of-Origin token provider (bgutil). YouTube increasingly binds a
+# GVS PO token to downloads from datacenter IPs; without a provider, authenticated
+# (cookie) downloads intermittently fail with "Sign in to confirm you're not a
+# bot". The HTTP-server variant runs on localhost and the matching yt-dlp plugin
+# (installed into the venv) auto-detects it on the default port.
+BGUTIL_POT_DIR=/opt/bgutil-pot
+BGUTIL_POT_VERSION="${BGUTIL_POT_VERSION:-1.3.2}"
+BGUTIL_POT_PORT="${BGUTIL_POT_PORT:-4416}"
+# Dedicated yt-dlp cache dir. HOME/.cache on this box is the Spotify OAuth token
+# *file*, so yt-dlp's default ~/.cache/yt-dlp path dies with NotADirectoryError
+# and silently disables player/signature + PO-token caching. Give it its own dir.
+YTDLP_CACHE_DIR="$APP_DIR/ytdlp-cache"
 HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 
 log()  { echo -e ">>> $*"; }
@@ -377,6 +389,47 @@ FLACFETCH_VERSION="$(python -c 'import flacfetch; print(flacfetch.__version__)'
 log "flacfetch version: $FLACFETCH_VERSION"
 python -c 'import yt_dlp_ejs' 2>/dev/null && log "yt-dlp-ejs available" || warn "yt-dlp-ejs not available"
 
+# Dedicated yt-dlp cache dir (see YTDLP_CACHE_DIR note above). Under $APP_DIR so
+# Stage 8b's chown to the service user covers it.
+mkdir -p "$YTDLP_CACHE_DIR"
+
+# =============================================================================
+log "Stage 5b — yt-dlp PO Token provider (bgutil, HTTP server)"
+# =============================================================================
+# The venv plugin (installed here) talks to a local bgutil HTTP server (built
+# below) that mints Proof-of-Origin tokens via YouTube's BotGuard challenge. The
+# plugin auto-detects the server on the default port, so no --extractor-args are
+# needed. Node.js is the server runtime (deno alone can't populate node_modules).
+# Pin the plugin to the SERVER version — bgutil requires matching plugin/server
+# versions; bump BGUTIL_POT_VERSION to upgrade both together (re-provision rebuilds
+# the server + re-pins the plugin). update-ytdlp.sh deliberately does NOT bump it.
+pip install --upgrade "bgutil-ytdlp-pot-provider==${BGUTIL_POT_VERSION}" --quiet \
+  || warn "bgutil PO-token plugin install failed"
+if ! command -v node >/dev/null 2>&1; then
+  log "installing Node.js (bgutil PO server runtime)"
+  apt-get install -y nodejs npm >/dev/null 2>&1 \
+    || warn "nodejs/npm install failed — PO token provider server will be unavailable"
+fi
+if command -v node >/dev/null 2>&1; then
+  if [ ! -d "$BGUTIL_POT_DIR/.git" ]; then
+    git clone --quiet "https://github.com/Brainicism/bgutil-ytdlp-pot-provider.git" "$BGUTIL_POT_DIR" \
+      || warn "bgutil clone failed"
+  fi
+  if [ -d "$BGUTIL_POT_DIR/.git" ]; then
+    git config --global --add safe.directory "$BGUTIL_POT_DIR"
+    git -C "$BGUTIL_POT_DIR" fetch --tags --quiet || warn "bgutil fetch failed"
+    git -C "$BGUTIL_POT_DIR" checkout --quiet "$BGUTIL_POT_VERSION" \
+      || warn "bgutil checkout $BGUTIL_POT_VERSION failed"
+    if ( cd "$BGUTIL_POT_DIR/server" && npm ci --silent && npx --yes tsc ) >/dev/null 2>&1; then
+      log "bgutil PO server built ($BGUTIL_POT_VERSION)"
+    else
+      warn "bgutil PO server build failed (PO token provider will be unavailable)"
+    fi
+  fi
+else
+  warn "node unavailable — skipping bgutil PO server build"
+fi
+
 # =============================================================================
 log "Stage 6 — secrets"
 # =============================================================================
@@ -593,6 +646,7 @@ Environment="FLACFETCH_DOWNLOAD_DIR=${DOWNLOAD_DIR}"
 Environment="TRANSMISSION_HOST=localhost"
 Environment="TRANSMISSION_PORT=9091"
 Environment="YOUTUBE_COOKIES_FILE=${YOUTUBE_COOKIES_FILE}"
+Environment="FLACFETCH_YTDLP_CACHE_DIR=${YTDLP_CACHE_DIR}"
 ExecStart=$VENV/bin/flacfetch serve --host 0.0.0.0 --port 8080
 Restart=always
 RestartSec=10
@@ -617,6 +671,40 @@ RestartSec=5
 WantedBy=multi-user.target
 XVFB_SERVICE
 
+# yt-dlp PO Token provider — bgutil HTTP server.
+# The server binary has no host-bind flag (binds :: then 0.0.0.0), but the host
+# nftables firewall already has `policy drop` on input with no accept rule for
+# $BGUTIL_POT_PORT, so it's unreachable externally while `iif lo accept` keeps it
+# available to local yt-dlp. Do NOT use systemd IPAddressDeny/Allow here: those
+# filter EGRESS too, and the server must reach Google's BotGuard/WAA endpoints to
+# mint tokens (blocking egress makes every mint fail with getaddrinfo EAI_AGAIN).
+if [ -f "$BGUTIL_POT_DIR/server/build/main.js" ]; then
+cat > /etc/systemd/system/bgutil-pot.service <<POT_SERVICE
+[Unit]
+Description=bgutil yt-dlp PO Token Provider (HTTP server)
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=$FF_SERVICE_USER
+Group=$FF_SERVICE_USER
+ExecStart=/usr/bin/node $BGUTIL_POT_DIR/server/build/main.js --port $BGUTIL_POT_PORT
+Restart=always
+RestartSec=5
+Environment=NODE_ENV=production
+NoNewPrivileges=true
+ProtectSystem=strict
+ProtectHome=true
+PrivateTmp=true
+
+[Install]
+WantedBy=multi-user.target
+POT_SERVICE
+else
+  warn "bgutil PO server not built — skipping bgutil-pot.service"
+fi
+
 # yt-dlp auto-update (daily 04:00 UTC)
 cat > "$APP_DIR/update-ytdlp.sh" <<'UPDATE_SCRIPT'
 #!/bin/bash
@@ -625,6 +713,9 @@ exec >> /var/log/ytdlp-update.log 2>&1
 echo "yt-dlp update started at $(date)"
 cd /opt/flacfetch && source venv/bin/activate
 OLD=$(python -c "import yt_dlp; print(yt_dlp.version.__version__)" 2>/dev/null || echo unknown)
+# NOTE: the bgutil PO-token plugin is intentionally NOT upgraded here — it is
+# pinned to the bgutil server version by provision.sh (mismatched plugin/server
+# versions can break POT minting). Bump BGUTIL_POT_VERSION + re-provision instead.
 pip install --upgrade yt-dlp yt-dlp-ejs --quiet
 NEW=$(python -c "import yt_dlp; print(yt_dlp.version.__version__)" 2>/dev/null || echo unknown)
 if [ "$OLD" != "$NEW" ]; then echo "yt-dlp $OLD -> $NEW"; systemctl restart flacfetch; fi
@@ -854,6 +945,10 @@ fi
 # ---- enable + (re)start ----------------------------------------------------
 systemctl daemon-reload
 systemctl enable --now xvfb >/dev/null 2>&1 || true
+if [ -f /etc/systemd/system/bgutil-pot.service ]; then
+  systemctl enable --now bgutil-pot >/dev/null 2>&1 || true
+  for _ in $(seq 1 10); do curl -s "http://127.0.0.1:$BGUTIL_POT_PORT/ping" 2>/dev/null | grep -q version && { log "bgutil PO server healthy"; break; }; sleep 1; done
+fi
 systemctl enable --now ytdlp-update.timer >/dev/null 2>&1 || true
 systemctl enable --now transmission-maintenance.timer >/dev/null 2>&1 || true
 systemctl enable flacfetch >/dev/null 2>&1 || true

flacfetch/__init__.py

@@ -1,6 +1,6 @@
 """flacfetch - Search and download high-quality audio from multiple sources."""
 
-__version__ = "0.24.0"
+__version__ = "0.25.0"
 __author__ = "Andrew Beveridge"
 __email__ = "andrew@beveridge.uk"
 

flacfetch/api/services/credential_check.py

@@ -243,7 +243,12 @@ def check_youtube_credentials() -> CredentialCheckResult:
             'skip_download': True,
         }
 
-        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
+        # Use a throwaway cookie copy so this validation check never writes yt-dlp's
+        # rotated (and, on a flagged IP, rejected) cookies back over the canonical
+        # keeper-managed file. See downloaders.youtube.isolated_cookiefile.
+        from ...downloaders.youtube import ytdlp_opts_isolated
+
+        with ytdlp_opts_isolated(ydl_opts) as opts, yt_dlp.YoutubeDL(opts) as ydl:
             info = ydl.extract_info(test_url, download=False)
 
         if info and info.get('title'):

flacfetch/api/services/health_check.py

@@ -357,8 +357,13 @@ def _extract_youtube_info(self, ydl_opts: Dict[str, Any]) -> Optional[Dict[str,
         """Extract info from YouTube test video (runs in thread pool)."""
         import yt_dlp
 
+        # Use a throwaway cookie copy so this periodic health probe never writes
+        # yt-dlp's rotated (and, on a flagged IP, rejected) cookies back over the
+        # canonical keeper-managed file. See downloaders.youtube.isolated_cookiefile.
+        from ...downloaders.youtube import ytdlp_opts_isolated
+
         try:
-            with yt_dlp.YoutubeDL(ydl_opts) as ydl:
+            with ytdlp_opts_isolated(ydl_opts) as opts, yt_dlp.YoutubeDL(opts) as ydl:
                 return ydl.extract_info(YOUTUBE_TEST_VIDEO, download=False)
         except Exception as e:
             logger.warning(f"YouTube info extraction failed: {e}")

flacfetch/downloaders/youtube.py

@@ -1,6 +1,9 @@
 import logging
 import os
+import shutil
+import tempfile
 import time
+from contextlib import contextmanager
 from dataclasses import dataclass
 from typing import Optional
 
@@ -62,6 +65,40 @@ def get_cookies_file() -> Optional[str]:
     return None
 
 
+def get_ytdlp_cache_dir() -> Optional[str]:
+    """
+    Resolve an explicit yt-dlp cache directory, if one is configured.
+
+    yt-dlp defaults its cache to ``$XDG_CACHE_HOME`` / ``~/.cache/yt-dlp``. On the
+    flacfetch server the service runs with ``HOME=/opt/flacfetch`` where
+    ``~/.cache`` is a *file* (the Spotify OAuth token cache), so yt-dlp's default
+    path resolves to ``/opt/flacfetch/.cache/yt-dlp`` and every cache write dies
+    with ``NotADirectoryError``. That silently disables player/signature and
+    PO-token caching, forcing a fresh JS-challenge solve on every request (slower
+    and more bot-detectable).
+
+    Set ``FLACFETCH_YTDLP_CACHE_DIR`` to a real directory to give yt-dlp its own
+    cache location that can't collide. When unset (e.g. library/CLI use on a dev
+    machine with a normal ``~/.cache``), we leave yt-dlp's default untouched.
+
+    Returns:
+        The cache directory (created if needed) or None to use yt-dlp's default.
+    """
+    cache_dir = os.environ.get("FLACFETCH_YTDLP_CACHE_DIR")
+    if not cache_dir:
+        return None
+
+    try:
+        os.makedirs(cache_dir, exist_ok=True)
+    except OSError as e:
+        # Don't let a misconfigured cache path break downloads; yt-dlp treats a
+        # missing/False cachedir as "caching disabled" and still works.
+        logger.warning(f"Could not create yt-dlp cache dir {cache_dir!r}: {e}")
+        return None
+
+    return cache_dir
+
+
 def get_ytdlp_base_opts(cookies_file: Optional[str] = None) -> dict:
     """
     Get base yt-dlp options with common settings including cookies if available.
@@ -81,9 +118,81 @@ def get_ytdlp_base_opts(cookies_file: Optional[str] = None) -> dict:
     if cookies_file:
         opts["cookiefile"] = cookies_file
 
+    # Route yt-dlp's cache to a dedicated dir when configured, so it doesn't
+    # collide with the Spotify token file at HOME/.cache on the server.
+    cache_dir = get_ytdlp_cache_dir()
+    if cache_dir:
+        opts["cachedir"] = cache_dir
+
     return opts
 
 
+@contextmanager
+def isolated_cookiefile(cookies_file: Optional[str]):
+    """Yield a throwaway copy of the cookie file for a single yt-dlp run.
+
+    yt-dlp saves the (possibly rotated) cookie jar back to ``cookiefile`` after
+    every run. On a flagged datacenter IP YouTube *rejects* the rotation
+    ("The provided YouTube account cookies are no longer valid. They have likely
+    been rotated in the browser as a security measure."), so the cookies yt-dlp
+    saves back are invalid and poison the shared file for the next download —
+    cookies that worked seconds ago start returning "Sign in to confirm you're
+    not a bot" within minutes. (On a trusted IP the rotation is accepted, so the
+    write-back is harmless: this only bites since the move to the datacenter box.)
+
+    Handing each run its own copy keeps the canonical, keeper-managed cookie file
+    pristine (the credential keeper is its only writer) and isolates concurrent
+    downloads from one another. Yields the temp copy path, or the original value
+    when there's nothing to copy (no cookies / file missing).
+    """
+    if not cookies_file or not os.path.exists(cookies_file):
+        yield cookies_file
+        return
+
+    tmp_path = None
+    try:
+        fd, tmp_path = tempfile.mkstemp(prefix="ytdlp-cookies-", suffix=".txt")
+        os.close(fd)
+        shutil.copyfile(cookies_file, tmp_path)
+    except OSError as e:
+        # Copy failed (disk full, perms, ...). Fall back to the canonical file so
+        # downloads still work — we lose write-back protection for this one run,
+        # which is strictly better than failing the download outright.
+        logger.warning(f"Could not create temp cookie copy, using canonical file: {e}")
+        if tmp_path and os.path.exists(tmp_path):
+            try:
+                os.unlink(tmp_path)
+            except OSError:
+                pass
+        yield cookies_file
+        return
+
+    try:
+        yield tmp_path
+    finally:
+        if os.path.exists(tmp_path):
+            try:
+                os.unlink(tmp_path)
+            except OSError as e:
+                logger.warning(f"Could not remove temp cookie copy {tmp_path!r}: {e}")
+
+
+@contextmanager
+def ytdlp_opts_isolated(ydl_opts: dict):
+    """Run yt-dlp with a throwaway copy of the cookie file (see isolated_cookiefile).
+
+    Swaps ``cookiefile`` in a shallow copy of ``ydl_opts`` for a per-run temp copy
+    so the canonical keeper-managed cookies are never written back to. A no-op when
+    no ``cookiefile`` is configured.
+    """
+    original = ydl_opts.get("cookiefile")
+    with isolated_cookiefile(original) as cf:
+        if original:
+            yield {**ydl_opts, "cookiefile": cf}
+        else:
+            yield ydl_opts
+
+
 @dataclass
 class YoutubeAvailability:
     """Result of a YouTube video availability check."""
@@ -147,7 +256,7 @@ def check_youtube_availability(
     })
 
     try:
-        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
+        with ytdlp_opts_isolated(ydl_opts) as opts, yt_dlp.YoutubeDL(opts) as ydl:
             info = ydl.extract_info(url, download=False)
             if info:
                 return YoutubeAvailability(
@@ -353,7 +462,7 @@ def _timeout_hook(d):
 
         downloaded_file: str | None = None
         try:
-            with yt_dlp.YoutubeDL(ydl_opts) as ydl:
+            with ytdlp_opts_isolated(ydl_opts) as opts, yt_dlp.YoutubeDL(opts) as ydl:
                 info = ydl.extract_info(release.download_url, download=True)
                 # extract_info returns None when match_filter rejected the media.
                 if info is None:

flacfetch/providers/youtube.py

@@ -4,7 +4,7 @@
 
 from ..core.interfaces import Provider
 from ..core.models import AudioFormat, MediaSource, Quality, Release, TrackQuery
-from ..downloaders.youtube import get_ytdlp_base_opts
+from ..downloaders.youtube import get_ytdlp_base_opts, ytdlp_opts_isolated
 
 
 class YoutubeProvider(Provider):
@@ -46,7 +46,7 @@ def search(self, query: TrackQuery) -> list[Release]:
 
         releases = []
         try:
-            with yt_dlp.YoutubeDL(ydl_opts) as ydl:
+            with ytdlp_opts_isolated(ydl_opts) as opts, yt_dlp.YoutubeDL(opts) as ydl:
                 info = ydl.extract_info(search_query, download=False)
                 if info and 'entries' in info:
                     for entry in info['entries']:

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "flacfetch"
-version = "0.24.0"
+version = "0.25.0"
 description = "Search and download high-quality audio from multiple sources"
 readme = "README.md"
 requires-python = ">=3.10,<4.0"