Use YouTube player clients that are not bot-checked

services/media.py

@@ -30,20 +30,29 @@
 
 log = logging.getLogger("loopify.media")
 
-# The player_client fallback chain, as a CLI value for the streaming subprocess.
-_PLAYER_CLIENTS = "default,android_vr,tv_embedded"
+# YouTube player clients, tried in order. ONE definition — both the metadata
+# options below and the streaming subprocess derive from this, because two
+# separate lists silently drift and then playback and search disagree about
+# which client to use.
+#
+# Ordered by measured behaviour, not by theory. Every client yt-dlp offers was
+# tested against the same video from a residential IP:
+#
+#   web_embedded  works, ~3.2s    <- fastest that works
+#   mweb          works, ~9.3s    <- reliable but slow, kept as a fallback
+#   tv_embedded   bot-checked     <- kept last: needs no JS runtime, so it is
+#                                    the only option on a host without Deno
+#   default / web / android_vr / tv / ios / android_music  all bot-checked
+#
+# The previous chain was "default,android_vr,tv_embedded" — every entry of which
+# is now bot-checked, so yt-dlp exhausted it and 5 of 6 tracks failed to load.
+#
+# A residential IP reduces YouTube's bot-checking but does NOT remove it. Valid
+# cookies (COOKIES_PATH) still help and remain supported; getting the client
+# chain right avoids needing them at all.
+_PLAYER_CLIENTS = ("web_embedded", "mweb", "tv_embedded")
 
 # Base yt-dlp config shared by every call.
-#
-# player_client fallback chain, tried in order:
-#   * "default" (web) gives clean audio-only formats and works best WITH cookies,
-#     but needs a JS runtime (Deno/Node) to solve YouTube's signature challenge —
-#     deploy installs Deno for exactly this (see deploy/setup.sh).
-#   * "android_vr"/"tv_embedded" need no JS runtime and are the fallback for
-#     environments without one (e.g. a dev box). yt-dlp automatically advances to
-#     the next client if one yields no usable formats.
-# From a datacenter IP, valid cookies (COOKIES_PATH) are what defeats YouTube's
-# "confirm you're not a bot" check — see deploy/README.md.
 YTDL_OPTIONS = {
     "format": "bestaudio/best",
     "noplaylist": True,
@@ -53,7 +62,7 @@
     "source_address": "0.0.0.0",   # bind to IPv4; avoids some 403s
     "skip_download": True,
     "extractor_args": {
-        "youtube": {"player_client": ["default", "android_vr", "tv_embedded"]},
+        "youtube": {"player_client": list(_PLAYER_CLIENTS)},
     },
 }
 
@@ -321,7 +330,7 @@ def spawn_stream(track: dict) -> AudioStream:
         "-f", "bestaudio/best",
         "-o", "-",                       # write audio to stdout
         "-q", "--no-warnings", "--no-playlist",
-        "--extractor-args", f"youtube:player_client={_PLAYER_CLIENTS}",
+        "--extractor-args", f"youtube:player_client={','.join(_PLAYER_CLIENTS)}",
         "--source-address", "0.0.0.0",
     ]
     cookies = YTDL_OPTIONS.get("cookiefile")

tests/test_media_helpers.py

@@ -113,3 +113,57 @@ def test_first_entry_passes_through_a_single_result():
 # Process teardown and failure classification moved to AudioStream when the
 # stderr-deadlock fix landed; they are covered against real subprocesses in
 # tests/test_audio_stream.py.
+
+
+# -- YouTube player clients --------------------------------------------
+
+def test_the_two_client_usages_cannot_drift():
+    """
+    The chain is used twice: as a list for metadata extraction and as a
+    comma-joined string for the streaming subprocess. When those were separate
+    literals they drifted, and search and playback disagreed about which client
+    to use.
+    """
+    from_options = media.YTDL_OPTIONS["extractor_args"]["youtube"]["player_client"]
+    assert list(media._PLAYER_CLIENTS) == list(from_options)
+
+
+def test_spawn_stream_passes_the_same_chain(monkeypatch):
+    captured = {}
+    monkeypatch.setattr(media.AudioStream, "launch", classmethod(
+        lambda cls, cmd: captured.setdefault("cmd", cmd)
+    ))
+    media.spawn_stream({"url": "https://example.invalid/x", "title": "T"})
+    cmd = captured["cmd"]
+    arg = cmd[cmd.index("--extractor-args") + 1]
+    assert arg == f"youtube:player_client={','.join(media._PLAYER_CLIENTS)}"
+
+
+def test_a_working_client_comes_first():
+    """
+    Order is latency: yt-dlp tries each client in turn, so a blocked client at
+    the front costs a full round trip before anything can play. web_embedded
+    measured 3.2s against mweb's 9.3s, both succeeding 6/6.
+    """
+    assert media._PLAYER_CLIENTS[0] == "web_embedded"
+
+
+@pytest.mark.parametrize("blocked", [
+    "default", "web", "android_vr", "tv", "ios", "android_music",
+])
+def test_clients_known_to_be_bot_checked_are_not_in_the_chain(blocked):
+    """
+    Every one of these was measured returning "sign in to confirm you're not a
+    bot". The previous chain consisted entirely of such clients, which is why
+    5 of 6 tracks failed to load.
+    """
+    assert blocked not in media._PLAYER_CLIENTS
+
+
+def test_a_client_needing_no_js_runtime_remains_as_a_last_resort():
+    """
+    web_embedded and mweb need a JS runtime (Deno) to solve the signature
+    challenge. tv_embedded does not, so it is the only thing that can work on a
+    host where Deno failed to install.
+    """
+    assert "tv_embedded" in media._PLAYER_CLIENTS