Root cause of 'tile appears but content empty': demux() spooled the live
NUT/H264 input to a temp file and awaited stream 'finish' — but the merge
ffmpeg output never ends during playback, so demux deadlocked, no probe,
no transcode, 0 frames sent.
- Demuxer: pipe input straight into ffmpeg stdin (-i pipe:0), parse NAL
frames live from stdout; parse video metadata from ffmpeg stderr with a
1.5s race (fall back to H264 defaults). No spool, no await-end.
- screenShareController: pass width/height/frameRate (1280x720@30) to
playStream — matches the prepareStream encode settings, so setVideoAttributes
gets real dimensions even when ffmpeg can't report metadata on an open pipe.
- Add tests/golive-demux-live-e2e.ts: proves frames flow while input is
still open (regression test for the deadlock).
- signalStream: flip voice state to self_video:true/self_deaf:false before
STREAM_CREATE (Discord silently ignores the request while video disabled)
- createStream: attach dispatch listeners before first signal (race), clean
up listeners on timeout, retry STREAM_CREATE every 3s up to 4 attempts
(upstream issue #217/#219 — Discord randomly drops the request)
- sendOpcode: direct [goLive:Streamer] log bypassing bootstrap debug filter
(proves op 18 is actually broadcast)
binding.gyp appended '/libdatachannel.so.0.24.0' to LDC_LIB; nixpkgs output
layout is <out>/lib/libdatachannel.so.0.24.1. Make LDC_LIB the complete
library path (env or default) and drop the append.
libdatachannel-src fetchFromGitHub + manual cmake build fails: GitHub tarball
does not include git submodules (deps/plog, libjuice, libsrtp, usrsctp) →
CMake 'source directory does not contain CMakeLists.txt'.
Switch to pkgs.libdatachannel (0.24.1): nixpkgs builds submodules + ships
lib/dev outputs. In the Nix sandbox everything is consistent (store glibc),
so the GLIBC_ABI_GNU2_TLS issue that blocks host-local use of 0.24.1 does
not apply to the Nix build. binding.gyp defaults stay on local 0.24.0 for
dev; Nix sets LDC_INCLUDE/LDC_LIB to the store paths.
- Replace the per-package rebuild loop (node-datachannel cmake-js, zeromq)
with: opus build + libdatachannel-min N-API binding build (fetchFromGitHub
libdatachannel v0.24.0 — pinned because nixpkgs 0.24.1 is glibc-incompatible
with this host; sha256 1jk53qs…).
- Removes ~760MB of node-datachannel build/cleanup cruft from the build
phase; node_modules now 423MB (was 1.5GB).
Track.setPacketizer(kind, ssrc, pt, clockRate, ...) builds the same
media-handler chain node-datachannel does for @dank074:
RtpPacketizer (Opus | H264 | H265 | AV1) → RtcpSrReporter →
RtcpNackResponder → PacingHandler(25Mbps, 1ms) for video
Track.sendFrame(encodedFrame) packetizes into RTP; addTimestamp(delta)
advances the RTP timestamp (node-datachannel contract).
Verified test-packetizer.js: two peers connected over tracks, real opus
frames + AnnexB H264 (SPS/PPS/IDR) flow through the chain without crash.
This removes the need for a JS RTP packetizer entirely — libdatachannel
0.24 has the full media stack built in.
Expose rtc::Track with send(binary) for raw RTP — verified:
- SDP from addTrack(audio)+addTrack(video) has m=audio (opus 120)
and m=video (H264 101 + H265/VP8/VP9/AV1 + RTX)
- libdatachannel Track::send() sends RAW RTP/RTCP when no media
handler is set (verified in src/track.cpp impl::Track::outgoing) —
so RTP packetization can live in pure JS, keeping the binding minimal
Also fix: Track class was missing from InitAll exports (crash on
TrackWrap::NewInstance — null FunctionReference).
Phase 0 of GoLive rewrite (drop @dank074/node-datachannel 771MB):
minimal N-API binding exposing PeerConnection/DataChannel/ICE/SDP,
built against libdatachannel 0.24.0 (from node-datachannel _deps source).
Verified: offer/answer/ICE/DataChannel roundtrip between two local
peers (test-handshake.js). Key findings:
- callbacks must be registered in ctor BEFORE createDataChannel
- SDP with candidates comes from localDescription() at gathering Complete
- answer auto-generates on setRemoteDescription(offer); do NOT call
setLocalDescription() after or role=actpass breaks the peer
prepareStream (from @dank074/discord-video-stream) unconditionally appends
audio filters 'volume@internal_lib' + 'azmq' that exist ONLY in its custom
node-av jellyfin-ffmpeg build. The Nix deployment runs plain ffmpeg-headless
on PATH, so fluent-ffmpeg died instantly with 'Filter not found' (exit 8),
the NUT output stream stayed empty, and playStream's node-av demux failed
with 'Failed to open input from Readable stream: Invalid data found when
processing input' — every screen share failed ~100ms after start.
Fix: pass customFfmpegFlags ['-filter:a','anull'] — ffmpeg applies the LAST
-filter:a for a stream, so the trailing no-op filter overrides the custom
chain (verified: command ends with '-filter:a anull', transcode runs, node-av
demux finds video+audio). Realtime volume control was already removed from
GMW (a690e5b), so dropping the filters is lossless.
Verified end-to-end with the failing URL (youtu.be/fONoh7Pc6VU, AV1+Opus
DASH): getDirectScreenInput → NUT merge → patched prepareStream → node-av
demux finds H264 video + Opus audio streams.
Root cause (3rd layer after 50371bd + 4f4c435): a vision model run
(2026-08-10) returned 'Maaf, saya tidak melihat gambar apapun yang terlampir...'
and that text was cached as a VALID vision_llm result (image + phash keys,
24h/7d TTL). Every subsequent analysis of the same image (same hash/phash)
hit the poisoned cache, so image analysis looked broken forever even though
9router responded fine — the moderation LLM wrote 'lampiran yang gagal
terbaca' from a cache hit.
Also: mimo via 9router streams reasoning in delta.reasoning +
delta.reasoning_details[].text (content:"") — extractChunkText only read
delta.reasoning_content, so those runs aggregated empty → 'Vision API null
response' (observed 08:54/09:07/09:38).
Fixes:
- llmClient.extractChunkText: fall back to delta.reasoning and
reasoning_details[].text (mimo), on top of reasoning_content (gemma).
- visionAnalyzer: isNoImageSeenText() detects 'no image' style outputs;
such results are NEVER cached, and poisoned entries are purged when hit
(LRU/DB/phash) so re-analysis actually re-runs vision.
- Tests: reasoning/reasoning_details extraction + isNoImageSeenText
(Indonesian + English, no false positives on real descriptions).
Root cause (2nd layer after 50371bd): the analysis worker could pick up an
image message while its attachment upload was still in flight
(upload_status='pending'). downloadAndExtractFrame then fell back to the
Discord CDN URL (cdn.discordapp.com), which often 404s for old/purged links,
and 'if (!res.ok) return' silently dropped the image — no log, no vision
call, empty image map, and the LLM produced a text-only verdict like
'lampiran yang gagal terbaca oleh sistem'.
Fixes:
- ai-analysis-worker: skip targets whose attachment upload is still pending
(both batch + individual paths) — they stay ai_status='pending' and the
next 15s cycle analyzes them after the upload lands.
- mediaDownloader.downloadAndExtractFrame: try uploaded_url first, then
discord_url as fallback; log non-OK responses (status + host) instead of
silently returning; log when all candidate URLs fail.
Root cause: 9router combo 'multimodal' routes to cloudflare-ai/@cf/google/
gemma-4-26b-a4b-it which streams ALL output in delta.reasoning_content
(content:"") and finishes with 'length' at max_tokens. llmClient only read
delta.content, so llmVision returned empty → every image moderation fell back
to text-only analysis ('Meskipun analisis gambar gagal' in every ai_analysis).
Fix: extractChunkText() prefers delta.content then falls back to
delta.reasoning_content (also handles message/text/response fields), with
unit tests for the exact 9router chunk shape. Verified live against a real
DB image: oc/mimo-v2.5-free (new first model in the multimodal combo) returns
a proper description in delta.content.
gateway output 1.4G -> 424M (-70%), backend 198M -> 60M (-70%).
- pruneProd: delete every .pnpm dir not in 'pnpm list --prod' graph
(biome/typescript/esbuild/drizzle-kit/vitest/tsx ~150MB+) then drop
dangling symlinks (top-level, scoped dirs, hoist, .bin) so stdenv
noBrokenSymlinks fixup passes.
- NOT using 'pnpm install --prod': it collapses the public-hoist dir
(.pnpm/node_modules) that peer resolution relies on for
@lng2004/node-datachannel + @seydx/node-av-linux-x64 (voice breaks).
- node-datachannel: strip build/_deps (cmake FetchContent ~380MB) +
nested node_modules (nw-gyp/typescript/puppeteer ~380MB) after
compile; runtime needs only build/Release/node_datachannel.node.
- verified: native binaries (datachannel/opus/zeromq) intact, all 27
runtime modules resolve, dev tools 0.
- VPS-hop attic push now runs via sudo so attic reads root's config
(~/.config/attic) which has the imrnes-ts server (Tailscale). Without
it the push ran as the CI user whose config only has pub ->
'Server imrnes-ts does not exist', silently skipping the cache upload.
- Direct push retried 3x (attic push is idempotent): a transient 502
(e.g. atticd restart mid-push, Traefik blip) no longer aborts the
whole closure upload before falling back to VPS-hop.
- Restore $VPS_USER in the 3 ssh:// nix copy fallbacks (was committed
as masked '***' -> nix copy would ssh as user '***' and fail).
Root causes found by reproducing the 2026-08-10 run:
- attic 'gmw' cache was created private -> every narinfo/nix-cache-info
read returned 401, so the VPS could never actually substitute from
attic ('Substituted from Attic cache' was a false positive when the
store path happened to be already present locally).
- VPS nix.conf used extra-trusted-substituters, which Determinate Nix
never merges for nix-store CLI clients; extra-substituters (all
users, no trust gate) fixes substitution (verified end-to-end:
delete path -> nix-store --realise pulls from attic over HTTPS).
- runner bootstrap of the attic client depended on nix copy --from
ssh:// (fragile, failed on runner); now the prebuilt attic client
closure lives in the attic cache itself and the runner pulls it over
HTTPS via extra-substituters configured in the Install Nix step.
Also surfaces bootstrap stderr on fallback for future debugging.
The old Push-to-Attic step SSH-copied the full closure (~794MB gateway) to
the VPS on every new store path before attic push — at ~500KB/s that took
25+ minutes (observed 40min+ in-flight run). The runner can now push
straight to the public attic endpoint (https://attic.asepharyana.my.id,
token auth validated) after pulling the prebuilt attic client closure
(52MB) from the VPS via nix copy --from. Falls back to the VPS-hop flow
whenever the direct path fails.
- <message> targets now carry time (ISO), repetitions (N identical short texts = spam signal), bot and edited flags; escape id/user XML
- rich <user_reputation>: total_infractions, clean_streak, last_offense_days_ago, repeat_offender (7-day window)
- <user_history> with last flagged messages for repeat offenders (wires dead getUserRecentInfractions)
- <user_profile as_of> staleness signal; <location_context topic> from captured channel topic
- prompt framing + output instructions teach the LLM to use the new signals without treating history as proof
- tests: contextEnrichment.test.ts (13) + topic cases in conversationContext.test.ts
When the ONLY violation is offensive_username (message content clean):
- Message is NOT deleted (nickname-only violation bypasses auto-delete)
- Member's server nickname is reset to default username via
setNickname(null) (Discord shows the global username again)
- Action 'reset_nickname' logged to moderation_actions; cooldown
10min per guild:user (LRU) so repeated messages by same member
don't hammer the Discord PATCH
- Config: AUTO_NICKNAME_RESET_ENABLED / AUTO_NICKNAME_RESET_COOLDOWN_MS
- resolveDisplayName(): member.displayName from captured metadata,
falls back to global username
- Applied to context lines, target message blocks, and media message
blocks — LLM sees the name the channel actually sees (nickname can
carry moderation signal itself)
- Conversation context recency gates (GAP_MS/MAX_AGE_MS): drop stale
messages before silence gaps; cold_start anchor + flow descriptor
tells LLM whether conversation is ongoing or restarted
- [location] block: channel name, thread name, nsfw/age flags from
captured metadata (thread names instead of bare IDs)
- Link media -> multimodal: text-batch URL fetches that resolve to
images now run vision analysis (bounded 15s) and switch prompt to
mixed mode; <web_content> gains og:title for page context
- pnpm-workspace.yaml: approve sharp build script (unblocks install)
A single long guild name was clipped (whitespace-nowrap + narrow min-width),
so the dropdown rendered as a tiny 144px box with truncated text. Size the
popup to fit-content up to the available width and let option text wrap.
Select dropdowns (voice tab, guild selector) rendered with a transparent
background because --color-popover/--color-input were undefined, and the
popup overlapped the trigger due to alignItemWithTrigger. Define the missing
theme tokens (popover, popover-foreground, input, secondary) and default the
select popup to open below the trigger.
- Implemented RadioGroup and RadioGroupItem for radio button functionality.
- Created ResizablePanelGroup, ResizablePanel, and ResizableHandle for resizable panels.
- Developed Sidebar component with context for state management and various subcomponents (SidebarTrigger, SidebarMenu, etc.).
- Added Spinner component for loading indicators.
- Introduced Table component with TableHeader, TableBody, TableFooter, and related subcomponents for structured data display.
- Built Toast component for notifications with customizable actions and icons.
- Implemented ToggleGroup and ToggleGroupItem for toggle button functionality with context support.
- Loop: toggle via POST /api/media/loop → COMMAND_MEDIA_LOOP; gateway
replays finished music track on natural end (queue untouched); status
payload exposes loop flag; FE tombol Loop di music-player + mini-player.
- Kualitas suara: music playback sekarang di-transcode sekali via ffmpeg ke
OggOpus 48kHz stereo 192kbps dengan volume di-bake ke encode — menghindari
double lossy encode (inlineVolume) yang bikin suara buram. Screen share
tetap pakai jalur lama.
- Backend: MediaState.loop, setLoop service, route + schema validation.
Volume sudah di-set default 0.3 di gateway (suara kecil saat play), dan
user bisa naikin sendiri di Discord (command media:volume) — jadi kontrol
volume lewat dashboard tak perlu. Hapus:
- BE: POST /api/media/volume route, mediaVolumeSchema, setVolume service
- FE: useMediaVolume hook, mediaApi.volume, slider volume di music-player
+ mini-player, field volume/setVolume di MediaPlayerProvider
Pertahankan COMMAND_MEDIA_VOLUME di gateway (masih dipakai command DC)
dan mic volume (terpisah, tetap di voice page).
Volume play music terlalu besar buat user — default sekarang 0.3 (30%)
di semua layer: player gateway (musicVolume=0.3), backend state/schema
(default 0.3), dan UI slider (fallback 0.3). User tetap bisa naikin
manual via slider volume di dashboard.
is-active after sleep 3 false-fails when the unit is still activating
(e.g. Next standalone boot >3s) — exit 3 flagged the deploy red even though
the service came up fine. Poll is-active up to 30s and only fail if it never
reaches 'active'.
noBrokenSymlinks fails frontend build: .next/standalone/node_modules/.pnpm/
node_modules/semver -> missing target. The standalone server never resolves
pnpm's hoisted .pnpm dir (it bundles its own node_modules) — delete broken
symlinks before install so stdenv check passes.
Sebelumnya frontend hanya static export yang di-serve nginx di dalam package
proxy. Sekarang frontend = runtime mandiri (Next.js standalone :4017) yang
nginx proxikan ('/' -> Next server, '/api' + '/ws' -> backend :4001). Tambah
'frontend' ke matrix deploy agar dideploy + memulai unit gmw-frontend.
Rombak total alur data frontend: dari static-export CSR (tiap browser
fetch sendiri + akumulasi state voice per-tab) jadi server-side rendering.
Frontend (Next.js):
- next.config: output export -> standalone; halaman jadi server components
- server data layer baru src/lib/api/server.ts (GMW_BACKEND_URL, no window)
- dashboard/media/messages/moderation/recordings/voice page -> RSC yang
fetch backend di render-time, seed ke client view (SWR fallbackData)
- hook-hook utama terima initialData -> first paint data server, revalidate
SWR setelahnya, tanpa spinner-blank-load
- messages: guild/channel/tab/selected dibaca dari URL di server, page awal
di-fetch server-side
Shared realtime state (voice) server-authoritative:
- backend src/modules/voice/live-speaker.ts: agregat voice_active_user dari
gateway jadi snapshot authoritatif (single source of truth semua browser)
- GET /api/voice/status kini include activeSpeakers
- WS initial states kirim voice_state snapshot saat connect (late join
langsung dapat state yang sama, bukan daftar kosong)
- useSpeakers seed dari server snapshot + voice_state full-replace +
voice_active_user delta upsert
Deploy:
- flake.nix: frontend package build SSR standalone (server.js wrapper,
GMW_FRONTEND_PORT=4017); proxy nginx template proxy / -> Next server,
/api + /ws tetap ke backend :4001
Music playback produced no audio: with `-o -` yt-dlp streams media on
stdout and emits its `--print` title/duration headers on stderr, but
resolveMediaUrl read them from stdout — stripping two binary 'lines' off
the WebM container and corrupting the stream (player 'playing' but silent).
Now headers are read from stderr and the stdout media stream is returned
untouched.
Screenshare was failing with EACCES: getDirectScreenInput used --no-simulate,
making yt-dlp write .f*.part files into the read-only Nix store CWD. Dropped
it — simulate mode still returns requested_formats[].url in the JSON.
Adds tests/mediaResolve.test.ts (stderr-header + untouched-stream regression).
shutdown() awaited httpServer.close(), which waits for ALL open
connections. A lingering WS/keep-alive socket left the process zombie
forever after an uncaughtException (e.g. pg 'Connection terminated
unexpectedly' to imrnes) — no exit, so systemd Restart=always could
never revive it; /api/guilds returned 502 until manual restart.
Add 10s force-exit timer in shutdown(); clear it on clean completion.
asepharyana