The old jdymusic CDN (e.g. m701/m801.music.126.net/.../jdymusic/obj/...)
intermittently drops connections from Node's HTTP stack — even with a
browser UA — while the same URL fetched via WinHTTP works. Newer paths
(/jd-musicrep-ts/, /ymusic/) do not have this restriction.
On Windows, dispatch /jdymusic/ URLs through `System.Net.WebClient` in
PowerShell into a temp file, then run ffmpeg against the file. Other
URLs and non-Windows platforms keep the direct ffmpeg + browser-UA path
from the previous commit. Also fix buildFfmpegArgs to omit HTTP-only
flags (-reconnect_*, -headers) when the input is a local file path —
new ffmpeg builds reject these as "Option not found".
stop() now also kills any in-flight PowerShell downloader and removes
the temp dir, so cancellation is clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Netease's m701/m801.music.126.net CDN RST connections from FFmpeg's
default Lavf UA. The bot's frame loop treats an early ffmpeg exit as
trackEnd and advances the queue, which surfaced as songs auto-skipping
mid-playlist. Add a browser UA + music.163.com Referer for these URLs,
extract args into a tested buildFfmpegArgs(), and harden reconnect
flags (delay_max 5 -> 30, plus reconnect_on_network_error /
reconnect_on_http_error). Verified by A/B running ffmpeg against the
same fresh CDN URL: legacy args got 0 bytes + "End of file" reading
HTTP response; fixed args streamed the full track to completion.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Final review of caeef65 caught a stale-currentIndex bug in
/play-next-song and cmdPlayNext: when the player is idle but
queue.currentIndex >= 0 (natural end-of-track, or playNext gave
up after retries without queue.clear), addNext splices mid-queue
and playAt(size-1) jumps PAST the inserted song to whatever
was last. Capture the insertion slot before calling addNext and
promote that exact index instead.
Add a regression test covering the [a,b,c,d] queue with
currentIndex=1 case -- splice at 2 yields x at index 2, but
size-1 would point at d (index 4).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Splices into currentIndex+1 and shifts both playedIndices and the
history back-stack to keep references valid. Falls through to plain
push when nothing is playing so the idle-bot "add → start playing"
flow is unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Code review of 390d3fa flagged that dropping playedIndices.clear() from
playAt is an unnecessary behavior change. The two operations (push
history, reset random pool) are not in tension — restoring the clear
preserves the original 'explicit pick restarts shuffle' semantic that
Random mode users rely on, while still tracking the prev-history stack.
Also add a regression test that the 50-entry HISTORY_LIMIT cap works.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
In random modes, prev was just doing currentIndex-1 in the array, which
has no relationship to what the user actually played before. Add a
50-entry back-stack that's pushed by next/playAt, popped by prev, reset
on play/clear/setMode, and shifted by remove. Sequential and Loop modes
keep their old fallback for the case where history is empty.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Inert in this commit — no callers yet. Sets up the back-stack used
by the upcoming history-aware prev rewrite.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
RandomLoop randomly selects from queue, causing the same song to play
multiple times when the queue is small (~3 songs from personal_fm API).
Random mode ensures each song plays once before repeating, relying on
refillFm() to fetch fresh songs before exhaustion.
Also fix the proactive refill condition to use unplayedCount() instead
of size() - getCurrentIndex(), which was meaningless for non-sequential
play modes.
Three corner cases fixed:
1. Removing the currently-playing song caused the next song in the array
to be silently marked as "played" and skipped. Root cause: playedIndices
was updated in next() by marking currentIndex, but after remove() shifts
currentIndex, it pointed to the wrong song. Fix: mark songs as played
at play-time (in play/playAt/next/prev) instead of at next-request-time.
2. Using prev() in Random mode could cause a song to play twice — the
song navigated to via prev() was not recorded in playedIndices, so
next() could randomly select it again. Fix: prev() now marks the
returned song as played.
3. Switching to Random mode mid-playback could cause the current song to
repeat because setMode() cleared playedIndices without preserving the
currently-playing song. Fix: setMode() now re-adds currentIndex after
clearing.
https://claude.ai/code/session_01W3ZncxL5VfdZeB4qqYWDmY
In Random (shuffle) mode, the queue's next() method would return the same
song indefinitely when only one song was in the playlist, and never terminate
even with multiple songs. This happened because played songs were not tracked.
Added a playedIndices Set to track which songs have already been played in
Random mode. Once all songs have been played once, next() returns null to
stop playback — matching the expected behavior where Random plays each song
once in random order, while RandomLoop is the mode for infinite shuffling.
https://claude.ai/code/session_01W3ZncxL5VfdZeB4qqYWDmY
The ffmpeg-static npm package bundles a pre-compiled binary that passes
`ffmpeg -version` but crashes with SIGSEGV during actual audio processing
inside Docker containers (incompatible glibc/missing shared libraries).
Changes:
- Install system FFmpeg via apt-get in Docker production stage, which is
always compatible with the container runtime
- Reverse FFmpeg resolution priority in player.ts: prefer system FFmpeg,
fall back to ffmpeg-static (for non-Docker environments like Windows)
- Optimize Dockerfile multi-stage build: compile native modules (opus,
better-sqlite3) in builder stage and copy to production, eliminating
the need for build tools (python3, make, g++) in the production image
- Fix qq-music-api dependency from file:../qq-music-api (path outside
Docker build context, breaks npm ci) to npm registry ^2.2.10
https://claude.ai/code/session_01JAV8sBokoifKh4Hc8XbJws
Major bug fixes and corner-case hardening across the backend, plus a
comprehensive feature test suite. All 94 unit tests + 51 integration
tests pass against a local TS3 server.
Lifecycle & state consistency
-----------------------------
- Bug A: startBot() now wraps connect() in a 15s deadline. A hung TS
handshake no longer blocks the /start HTTP call forever; the failing
instance is torn down and the caller gets a clean 500.
- Bug B: executeCommand rejects audio-dispatching commands (play, add,
next, skip, prev, playlist, album, fm) when the bot is disconnected.
Config-only commands (vol, mode, clear, stop, queue, now, lyrics)
still work so the UI stays usable while offline.
- Bug C: the tsClient 'disconnected' handler always clears player state
now, even when connect() never completed. A separate disconnectEmitted
flag guards duplicate external event emission. Previously an orphaned
connect attempt that idle-timed-out would leave playing=true forever.
- resolveAndPlay re-checks this.connected AFTER the URL-resolve await so
a stop() during the network call can't spawn ffmpeg on a disconnected
bot.
- connect() throws if disconnect() fired during the handshake await,
preventing a concurrent stop from being overwritten by a late connected
flag flip.
- startBot always disconnects the outgoing BotInstance before creating
a replacement, covering the mid-handshake case where isConnected()
still returned false but the library client was live.
- startBot now reuses the stored identity so server groups granted to
the bot survive restarts (was regenerating a fresh UID each time).
WebSocket reliability
---------------------
- BotManager extends EventEmitter and emits 'botInstance' whenever a
new instance is created. websocket.ts listens and re-attaches its
stateChange / connected / disconnected listeners immediately, fixing
the bug where player-bar UI never updated until manual refresh.
- attachedBots map now stores the BotInstance reference and detaches
stale listeners when the instance is replaced. Safety-net interval
(5s) also reconciles to catch anything missed.
- removeBot emits 'botInstanceRemoved' -> WS broadcasts a new
{type:"botRemoved", botId} message. Client drops the bot from its
local store instead of showing it as permanently offline.
HTTP input validation
---------------------
- /volume rejects non-number, NaN, Infinity, and out-of-range values
with a proper 400 instead of a 200 OK wrapping a usage-text string.
- /mode rejects anything not in {seq, loop, random, rloop} with 400.
- /seek rejects NaN / Infinity / negative (previously NaN slipped
through typeof==="number" and poisoned seekOffset).
- /play-at validates index < queue.size() BEFORE stopping current
playback (was silently killing the current song on invalid input).
- /play, /add, /playlist, /play-by-id, /add-by-id, /play-playlist
all honour platform=youtube now (previously fell through to netease
and silently played the wrong platform).
YouTube made truly optional
---------------------------
- Lazy checkYtDlpAvailable() runs `yt-dlp --version` once, caches only
positive results so users can install yt-dlp mid-run and have it
picked up without a restart.
- getAuthStatus() returns loggedIn=false with nickname
"YouTube (yt-dlp not installed)" when the binary is missing. UI can
grey out YouTube instead of silently returning empty searches.
- findYtDlp() picks .exe on win32 and bare binary elsewhere.
- /auth/status?platform=youtube now routes to the YouTube provider
instead of falling through to NetEase and leaking the NetEase
user's nickname + avatar.
- /auth/cookie rejects platform=youtube with 400 instead of clobbering
the NetEase cookie entry.
- README documents yt-dlp install paths (bin/ local vs PATH) and adds
a dedicated "Optional: YouTube source" section.
Bot Selector UI
---------------
- New power button in each row of the dropdown with play-state-aware
styling: disabled + wait-cursor during API call, green highlight when
connected, greys out when the bot is offline.
- Dropdown always visible when >=1 bot exists, bigger font + padding.
Queue correctness
-----------------
- PlayQueue.remove(current) now decrements currentIndex so next() in
sequential mode advances to the shifted song. Previously removing
the currently-playing track silently skipped the next track because
current() falsely reported it as active and next() then incremented
past it.
Vote-skip hardening
-------------------
- cmdVote: needed threshold is Math.max(1, ceil(users/2)) so a single
voter in an empty channel can't unanimously pass a vote with
needed=0.
- resolveAndPlay clears voteSkipUsers on every new track load so votes
can't leak across songs via cmdPlay/cmdPlaylist/cmdAlbum/cmdFm paths.
cmdAdd parity
-------------
- cmdAdd auto-plays the newly-added song if the player was idle,
matching /api/player/:id/add-by-id behaviour. Previously add'ing to
an empty queue on a connected+idle bot silently enqueued without
starting playback.
Test suite
----------
- scripts/test_full_feature.py — 51 tests across 10 groups exercising
every HTTP endpoint, WebSocket broadcasts, all music providers, bot
lifecycle, disconnected-state corners, seek validation, input
validation, and the main race conditions. Captures and restores the
target bot's initial state. Resilient to TS3 anti-flood via retry
with exponential backoff. Runs against a real local TS3 server.
- scripts/test_rapid_cycle.py — Bugs A/B/C regressions
- scripts/test_corner_cases.py — disconnect-during-connect race, config
commands while disconnected, etc.
- scripts/test_more_corners.py — resolveAndPlay race, seek NaN
- scripts/test_power_button.py — E2E for the new power button
- scripts/test_bot_remove.py — E2E for WS botRemoved broadcast
- scripts/test_playbar.py — player bar auto-show regression (updated
to restore bot state on exit)
- scripts/test_multibot.py — two-bot concurrent playback monitor
- src/audio/queue.test.ts — 4 new vitest cases for remove() edge cases
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two issues fixed:
1. Cross-platform ffmpeg-static resolution: skip Windows .exe paths on Linux
and always fall back to "ffmpeg" instead of a known-bad path
2. Prevent trackEnd cascade when ffmpeg spawn fails — track consecutive
failures and stop after 3, suppressing trackEnd on spawn errors
https://claude.ai/code/session_013vHRF8BbDGjZLqheFS85Q6
- Global uncaughtException/unhandledRejection handlers in index.ts
- FFmpeg stdout/stderr stream error handlers in player.ts
- HTTP server and WebSocket server error handlers in server.ts
- Safe WebSocket broadcast with try-catch in websocket.ts
- Catch async errors from textMessage handler in instance.ts
- Reset voiceFramesSent counter on reconnect in client.ts
https://claude.ai/code/session_01EjpEsC2GCsvwbu4n3XC8EE
The ffmpeg-static bundled binary was crashing immediately (exitCode: null,
no data produced). Now we run `ffmpeg -version` to verify the binary works
before selecting it, with automatic fallback to system ffmpeg.
Also logs ffmpeg binary path at info level and captures signal in close event.
https://claude.ai/code/session_01EjpEsC2GCsvwbu4n3XC8EE
Adds info-level logs at each stage of the audio pipeline to help
diagnose why playback produces no audible output:
- FFmpeg first PCM data received
- FFmpeg process exit code
- FFmpeg stderr (errors/HTTP/stream info at info level)
- First opus frame encoded and emitted
- First voice packet sent to TeamSpeak
- Error catching in frame encoding loop
https://claude.ai/code/session_01EjpEsC2GCsvwbu4n3XC8EE
Avoid repeated filesystem checks on every play() call by resolving
the ffmpeg binary path once at module load. Also verify execute
permission after chmod to handle noexec mounts.
https://claude.ai/code/session_01CqfKgV8GuCmWNpfx86H62X
The bundled ffmpeg-static binary may lose execute permission after npm
install on some platforms, causing EACCES errors during playback. Now
checks and fixes the permission automatically before spawning ffmpeg.
https://claude.ai/code/session_01CqfKgV8GuCmWNpfx86H62X
Implement BiliBiliProvider for video-as-audio playback using direct
BiliBili API calls (search, video info, DASH audio URL extraction).
Add QR code login support and cookie persistence. Update FFmpeg to
send Referer header for BiliBili CDN URLs. Extend platform union type
to "netease" | "qq" | "bilibili" across all interfaces. Add -b flag
for chat commands and B站 badge in web UI.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add isAdvancing guard to playNext() to prevent concurrent calls from
skipping songs (trackEnd + user !next race condition)
- Replace removeAllListeners in WebSocket setup with tracked named
listeners, attach only once per bot instead of every 5 seconds
- Add .catch() to unhandled playNext() in cmdVote
- Add sessionId counter to AudioPlayer to discard stale setTimeout
callbacks after stop()+play() transitions
- Defer nulling TS3 client until disconnect() promise resolves
- Add null checks for providers in play-by-id and add-by-id endpoints
- Add PCM buffer backpressure (pause FFmpeg at 960KB, resume at 480KB)
- Change volume slider from @input to @change to avoid flooding server
- Wrap error-reporting sendTextMessage in try/catch to prevent
unhandled rejection in error handler
- Clamp elapsed getter to song duration
- Log FFmpeg stderr at debug level instead of swallowing
- Return null from prev() in Sequential mode instead of wrapping
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two bugs prevented auto-advance:
1. FFmpeg close handler was empty — ChildProcess object persisted after
close, so frame loop check `!this.ffmpeg` never triggered trackEnd
2. playNext() is async (resolves URL) but trackEnd handler didn't catch
rejected promises, causing silent failures
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Users no longer need to install FFmpeg separately — it is now bundled as
an npm dependency (ffmpeg-static). player.ts resolves the bundled binary
automatically and falls back to system PATH. start.bat and a new
setup.bat script handle dependency installation and project building.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Progress/lyrics sync:
- AudioPlayer tracks framesPlayed (ground truth: frames * 20ms + seekOffset)
- BotStatus.elapsed is the real elapsed from server
- Frontend polls GET /elapsed every 3s for ground truth
- Client interpolates between syncs (serverElapsed + timeSinceSync)
- Pause freezes at interpolated value, resume resyncs from server
- No more client-side timestamp drift
Playlist play all:
- Stops current playback before loading
- Respects play mode: random/rloop picks random first song, seq plays first
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- AudioPlayer.seek() restarts FFmpeg with -ss parameter for actual seeking
- New POST /api/player/:botId/seek endpoint
- BotStatus includes seekOffset and playStartTime from server
- Frontend elapsed = seekOffset + (now - playStartedAt)
- Progress bar click now seeks to clicked position
- All timing actions use _resetTiming() helper with seekOffset
- Lyrics sync uses store.elapsed which includes seekOffset
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Follows TS3AudioBot-NetEaseCloudmusic-plugin pattern:
- QueuedSong.url is now optional (metadata-only queue entries)
- resolveAndPlay() fetches URL on-demand right before playback
- Playlist/album/FM load instantly (only metadata stored)
- playNext() resolves URL lazily, skips up to 3 songs on failure
- Avoids expired URLs for songs deep in the queue
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- High-precision frame loop using performance.now() tracking
- PCM-level volume control (live adjustable, no FFmpeg restart)
- Proper buffer drain on track end
- Fixes audio quality and playback speed issues
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
