The Redis media:status key was only rewritten after a command received via
Redis. When the last track ended naturally (AudioPlayer Idle -> advanceQueue
with an empty queue), currentTrackItem was cleared but the status key was not
persisted — so the backend's cached status and the frontend's 10s polling
stayed stuck showing the finished track as 'playing' forever.
Wire a media-status sink (commandHandler provides the real redisPub to
MediaHandler) and re-publish status after auto-advance, so natural track end
updates the UI.
VoiceActivityTimeline was never given a data prop — the Activity tab always
showed an empty Recharts bar chart while the connection tab already had live
speaker state. Replace the dead chart with a live speaker/activity list fed
from the same WebSocket data, so the tab reflects real state instead of
misleading empty bars.
The voice_recordings table has no duration column, but the backend aliased
duration_bytes = size_bytes (file size in bytes) and RecordingCard divided it
by 60 as if it were seconds — a 3MB MP3 rendered as a nonsensical '55924:3'
fake timestamp. Drop the fabricated field and show real file size instead.
Also render edited_content fallback in the analysis search results for
consistency with message cards/detail.
message_updated broadcasts only {id, edited_content, edited_at} (+ reset
ai_* fields), but the frontend replaced the whole cached record, wiping
username/content/channel_id/created_at -> blank cards and the
'the channel_id of undefined' crash on /messages. Merge partials over the
existing record (list + detail), make list-patching channel-filter aware,
show edited content/badge, and fix the message_updated WS type.
Also broadcast type:'edited' + ai reset in message_updated so the live UI
matches the DB update.
Backend:
- GET /api/dashboard/activity?days=1..90 — daily buckets (messages,
flagged, active_users) + hourly distribution last 24h
- clamped days param, reuses existing indexes (idx_messages_created,
ai_status_created)
Frontend:
- ActivityChart: area chart messages+flagged per day, 7/14/30d range
- HourlyActivityChart: 24h bars with peak highlight
- ModerationDonut: clean/flagged/warned/error breakdown with live
summary line (server X% clean)
- Dashboard layout: activity 2/3 + donut 1/3, hourly + top channels
Verified: endpoint returns real data from prod DB (784/1897/8 msgs
per day), tsc clean backend+frontend.
Sebelumnya CI hanya build nix -> deploy tanpa verifikasi — placeholder
tests sempat rusak berbulan-bulan tanpa terdeteksi. Job 'test' baru:
- pnpm install + tsc --noEmit + vitest run untuk backend & discord-gateway
- biome check (errors fail, warnings pass)
- build-and-deploy now needs: test
packages/shared dihapus (5802d02), modul pindah ke src/shared/. Update
import placeholder.test.ts (gateway + backend) ke path lokal + aktifkan
tests/ di backend vitest config dengan alias @. Sebelumnya vitest run
gateway selalu gagal; sekarang 60+36 tests pass.
- recording-card: kartu aktif di-highlight (ring primary + glow pulse),
badge 'Now Playing'/'Loading'/'Paused', tombol play berubah jadi Pause
saat playing dan spinner saat loading, waveform equalizer beranimasi
(animate-eq, delay per bar) saat playing / pulse saat loading.
- recording-player: jadi now-playing panel — tombol play/pause + spinner
loading, progress bar + waktu (current/duration), status 'loading…',
audio element pindah ke sini + event onPlay/onPause/onWaiting/onCanPlay/
onPlaying/onError naik ke page.
- recordings/page: state isPlaying/isLoadingAudio + audioRef, togglePlay
(klik card lain = ganti track, klik card sama = pause/resume).
- globals.css: keyframes eq-bounce + card-glow.
Verified: FE tsc0, next build 10/10 static pages.
Voice page (connection tab):
- ListenControl baru: toggle Listen (Headphones) — mulai PcmPlayer dari
user gesture, subscribe onPcm WS, volume slider, bar level per-user
REAL dari PCM (bukan random).
- lib/audio/pcm-player.ts (baru): ScriptProcessorNode mixer — ring buffer
2s per user (hash FNV-1a sama dengan gateway), upsampling 24k→48k
linear, mix semua user ke mono, gain volume, cleanup ring diam 5s.
- useVoiceListen + hashUserId di hooks; auto-stop saat disconnect.
Recordings:
- recording-player: reset src+load+play() eksplisit (bukan autoPlay doang),
tampilkan filename + error state 'playback failed' kalau file rusak.
- recording-card: tombol Download fetch blob (CORS tele open) → objectURL
→ force download dengan nama asli; fallback buka tab baru kalau fetch
gagal; spinner saat mendownload.
Verified: FE tsc 0, next build 10/10 static pages.
Audit lanjutan: 6x 'LLM API request failed: Request was aborted' per jam.
Root cause: 9router/omniroute SELALU balas SSE (data: chunks) walau request
tanpa stream:true — SDK OpenAI non-stream menunggu FULL body sebelum parse,
jadi batch moderasi besar yang upstream-nya lambat kena timeout 30-60s dan
di-abort. llmClient sudah punya agregasi streaming (chunks → ChatCompletion).
Fix: stream:true di llmCaller (moderasi batch/individual), llmVision,
cultureLearner, userProfileLearner. Verified: SDK stream test 806ms vs
sebelumnya abort. Caller lain (recovery worker dll) lewat llmCaller sama.
Audit log produksi (sejak deploy13:38) menemukan 3 isu:
1. mediaDownloader.ts spawn /usr/bin/ffprobe + /usr/bin/ffmpeg (path keras) —
ENOENT di Nix karena binary cuma di ffmpeg-headless closure. Pakai
PATH-resolved ('ffprobe'/'ffmpeg') seperti voice-recording module
(ffmpegProcess.ts/transmitter.ts) — 5 media warning hilang.
2. individualFallbackProcessor log error 'Success' di level50 tiap fallback
BERHASIL (logModerationError dengan new Error('Success')) — ganti
logger.info dengan verdict yang sama; error log cuma untuk error asli.
3. moderationResponseParser: strip frasa penutup generik ('Tidak ada
indikasi pelanggaran.') yang masih sering dikeluarkan LLM walau prompt
melarang (277/1486 analisis mengandung frasa, termasuk hari ini).
sanitizeGenericCleanCloser hanya mencocok frasa di AKHIR, teks substantif
tetap utuh. Unit test: 6/6 pass.
Migration 0011 (add voice_recordings.transcription) when=1781388000000
lebih kecil dari 0010 (1781390000000) yang sudah ter-apply — drizzle
skip diam-diam (folderMillis <= max(created_at)), kolom transcription
tidak pernah dibuat. Recording OGG sukses tapi INSERT voice_recordings
gagal 42703 di produksi.
Fix: when=1785600000000 (> max applied 1785551832190) + apply manual
ALTER TABLE + insert row __drizzle_migrations dengan hash file yang
sama (c368acb0...) supaya gateway restart berikutnya skip (idempotent).
pnpm 11.17 mengabaikan field pnpm.onlyBuiltDependencies di package.json.
Native deps voice (@discordjs/opus, @lng2004/node-datachannel, zeromq, dll)
tidak pernah kebangun di Nix store karena flake pnpmInstall pakai
--ignore-scripts dan pnpm rebuild tanpa approval. Hasil: receiver/rekaman/
GoLive diam-diam tanpa decoder/encoder native.
pnpm approve-builds --all menulis allowBuilds:true per package di
pnpm-workspace.yaml (harus tracked — flake source cuma ikut file git).
node-crc tetap gagal build (MSRV cargo:: check) tapi tidak pernah
di-import di source — harmless.
Root cause voice tidak berfungsi di produksi: ffmpeg & yt-dlp cuma ada
di devShell, bukan di package discord-gateway. Bukti dari log gateway:
'FFmpeg/avconv not found!' saat voice:transmit:start (mic -> Discord),
yang juga mematikan music playback (StreamType.Arbitrary butuh ffmpeg)
dan segment muxing rekaman.
- buildInputs: pkgs.ffmpeg-headless + pkgs.yt-dlp
- wrapper export PATH ke keduanya sebelum exec node
Verifikasi: nix build PASS; closure berisi ffmpeg-8.1.2 + yt-dlp-2026.07.04;
wrapper PATH mengarah ke keduanya; ffmpeg/yt-dlp jalan.
Hapus tab Live + komponennya (LiveStream, ModQueue) dari halaman
dashboard — tab Stats/Users/Channels tetap. useReview tetap dipakai
messages page (review tab), jadi hook tidak dihapus.
Verifikasi: tsc PASS, biome 0 error, next build PASS.
msg.username bisa undefined saat pesan live masuk lewat WS
(message_updated Partial payload / capture tidak lengkap) ->
msg.username.charAt(0) TypeError, halaman /messages mati.
- message-card + search-panel: username?.charAt(0) ?? '?'
- created_at di-guard juga biar tidak render 'Invalid Date'
Verifikasi: tsc PASS, build PASS, biome 0 error. DB saat ini 0 row
null username (1176 total) — crash murni dari jalur WS live.
- useMessageDetail: guard attachments fetcher — revalidate bisa race
detail load, detail.data undefined saat fetcher jalan ->
TypeError 'Cannot read properties of undefined (reading channel_id)'
yang bikin halaman /messages mati (Next error boundary). Sekarang
fetcher balikin [] kalau channel_id belum ada; hapus non-null
assertion. Diverifikasi: /messages sebelumnya crash, sekarang render
dengan data asli (list, verdict, confidence, sticker, emoji).
- ResponsiveContainer (recharts 3.8): initialDimension -1 di render
pertama -> warning 'width(-1) and height(-1)'. Pakai height numerik
tetap (192/160/48) + minWidth/minHeight 0 -> calculatedHeight >0,
warning hilang; width tetap responsif via ResizeObserver. Chart
baru dirender setelah mount (useMounted) biar container punya ukuran.
Diverifikasi console: 0 warning, 0 error di dashboard & voice.
/api/chat/history 500 (relation "chatbot_messages" does not exist):
codebase di-rename mascot->chatbot (977a6f9) tapi tabel DB tidak
pernah di-migrate — backend SELECT dari chatbot_messages, DB masih
mascot_chat_messages dengan kolom mascot_response.
- Migration 0013 (idempotent): ALTER TABLE mascot_chat_messages RENAME
TO chatbot_messages, RENAME COLUMN mascot_response -> bot_response,
RENAME INDEX -> idx_chatbot_messages_user_created; journal when >
max(created_at) di __drizzle_migrations (0012)
- Diterapkan live via psql; riwayat chat lama tetap ada
- Verifikasi: GET /api/chat/history 200 + data, POST /api/chat 200
+ tersimpan (total history 1 -> 2)
Rombak data layer frontend:
- Hapus @tanstack/react-query (package.json, lockfile, provider di
dashboard layout) — ganti SWR 2.4.2 + SWRConfig (revalidateOnFocus
false, deduping 10s, no retry on 404)
- Semua hooks data ditulis ulang ke useSWR; useAction() helper baru
pengganti useMutation dengan surface kompatibel (mutate/mutateAsync/
isPending/error)
- useMessages + useMessagesHasMore share satu SWR key — probe cursor
yang tadinya dobel fetch API sekarang deduped
- WS sync (messages/media/recordings) pindah dari queryClient ke
SWR mutate dengan filter key + revalidate:false
- useMessageSearch() dipakai search-panel & search-overlay; search
overlay backdrop div -> button (fix a11y lint)
Rapikan UI + isi data:
- Tab stats recordings: placeholder 'coming soon' diganti stat asli
(total, ukuran, speaker unik, top speakers)
- Empty states konsisten via EmptyState (images/review/recordings),
EmptyState terima className
- biome check --write: 0 error, 8 warning pre-existing
- Verifikasi: tsc --noEmit PASS, next build PASS (11 halaman static),
API live dicek — semua endpoint dashboard/messages/guilds/config/
voice/media/recordings/review balikin data
QoL lanjutan dari fix60084b3: content pesan mentah masih nampilin
snowflake (<@&roleid>, <@userid>, <:emoji:id>) di log moderasi dan
prompt LLM. Sekarang dirender ke nama yang bisa dibaca:
- Gateway capture: metadata menyimpan mentionedRoles + mentionedUsers
(id+name) dari message.mentions, disimpan ke metadata JSON
- renderDiscordMentions(): <@&id> -> @RoleName, <@id> -> @Username,
<:name:id> -> :name:, fallback @role/@user — dipakai di
conversationContext (konteks LLM) dan moderationBuilders
(getAnalysisContent) sehingga LLM lihat nama role/user beneran,
bukan placeholder generik
- Frontend renderMessageContent() (mirror gateway) dipasang di semua
tempat nampilin content: message-card, message-detail(-view),
search-overlay, search-panel, users/channels section, live-stream,
mod-queue, review list; sticker-only message tetap [Sticker: name],
pesan teks+sticker kini ikut nampilin nama sticker
- tsc --noEmit PASS di gateway & frontend; renderDiscordMentions
diverifikasi manual (6 kasus: role/user/emoji/unknown/plain)
- Deleted the text analysis prompt constants and helpers as they are no longer needed.
- Added batch search functionality for Qdrant to optimize vector searches.
- Implemented methods for deleting expired Qdrant points and invalidating cache based on content hash.
- Updated text batch processor to use new timeout configurations and modified content building for moderation prompts.
- Enhanced text cache store to support new Qdrant integration and improved cache invalidation logic.
- Introduced a new user reputation model with a more nuanced trust scoring system, including penalties and rewards for user behavior.
- Added unit tests for the new trust model to ensure correctness of penalty and trust gain calculations.
- Updated configuration schema to reflect new timeout settings and removed deprecated OpenAI moderation keys.
New qdrantClient.ts (zero-dep fetch REST): ensure collection with cosine
distance (auto-recreate on vector-size change), upsert point w/ verdict
payload, search w/ expires_at filter + score threshold.
textCacheStore: when QDRANT_URL set, embeddings are upserted to Qdrant
(primary) and searched there first; Postgres embedding column remains as
legacy fallback for pre-Qdrant rows. Config: QDRANT_URL/COLLECTION/API_KEY.
QDRANT_URL already in repo .env; added to VPS env + GATEWAY_ENV secret.
Drizzle 0.45 migrate() runs only migrations with folderMillis > the
latest created_at in __drizzle_migrations. Hand-written when was
1781580000000 (June 16) < stored 1781672400000 (June 17), so 0012 was
silently skipped and the embedding column never created. Bumped when to
now; migrator will apply it on next deploy.
OpenAI SDK v6 defaults to encoding_format=base64; llama-nemotron-embed
(Nvidia-backed) returns 400 'do not support base64'. Semantic cache was
silently disabled in prod. encoding_format: 'float' fixes it.
Previously /etc/gmw/*.env was managed by hand on the VPS; AI_LLM_* and
other vars drifted silently (AI_LLM_EMBEDDING_MODEL was missing until
added manually today). Now:
- BACKEND_ENV / GATEWAY_ENV secrets hold the full env block per service
- deploy.yml streams the secret to the VPS via stdin before the deploy
(never through argv — no shell escaping issues, no secret exposure)
- env file chown'd gmw:gmw, chmod 600; empty secret = skip (no clobber)
- .env.example documents the declarative workflow
Update production env = edit the Gitea secret + push, never SSH by hand.
TS compiled this fine, but the JS spec forbids mixing || and ??
without explicit parens; Node threw 'Unexpected token ??' at startup,
crash-looping gmw-discord-gateway (restart counter 250). Wrap the
fallback chain in parens so the expression is valid.
Root cause: ai-analysis-worker read llmResult.explanation and
llmResult.toxicityScore — fields the LLM pipeline never produces
(canonical AnalysisResult uses analysis/score). Every message fell back
to the bare template "Tidak ada indikasi pelanggaran." and the stored
score was always 0.
- Map analysis/score correctly; fallback now quotes the message content
- Prompt: ban generic analysis phrasing, require reply context
- LLM context: include replied-to message content (metadata.reference)
so the model can explain what the user is replying to
- Frontend: show thread/channel names from metadata instead of raw IDs
(message card, detail views, search overlay); detail panel now
displays the ai_analysis text
- Auto-delete log/DM include the descriptive analysis as the reason
Delete fastClassifier.ts (manual regex patterns for phone/email/IP/crypto/
spam/toxicity) and simpleFallback.ts. These hardcoded patterns were the
source of false positives (Discord emoji snowflakes matched phone_number,
URL digits matched phone, etc.) and produced heuristic verdicts whenever
the LLM failed.
New flow: Message → LLM (with conversation context, media evidence, user
reputation) → verdict. On LLM failure the message is marked 'error' and
retried by the recovery worker — no heuristic verdicts, ever.
Discord markdown tokens (custom emoji/mentions/timestamps) are normalized
to readable placeholders ([emoji:name], @user, @role, #channel, [time])
before reaching the LLM via discordTokens.ts.
Custom emoji (<:name:id>), user/role/channel mentions and timestamps embed
long numeric snowflakes that tripped the phone_number / personal_info /
ip_address_sharing patterns — e.g. <:mambotongue:1463255254220148939> was
flagged as phone_number. Strip Discord markdown tokens before pattern
matching and require phone matches to not sit inside a longer digit run.
- API/WS clients: same-origin by default, drop dead imphnen hardcode
- chatbot history: map BE rows {user_message,bot_response,created_at}
- message_deleted WS payload: object {id,deleted_at}, not bare string
- dashboard: wire top-channels chart + live mod queue from /api/review,
add Users & Channels tabs consuming /api/dashboard/users|channels
- recordings: live WS sync via voice_recording_uploaded; duration_bytes optional
- remove dead widgets with no BE data source (trend chart, heatmap)
- Deleted the frontend refactor design document to streamline project scope.
- Removed the services refactoring design document to eliminate redundancy.
- Eliminated the visual redesign document as part of the cleanup process.
- Purged the Discord Automod redesign document to focus on current objectives.
127.0.0.1 from inside Docker container is the container's own
loopback, not host's. Changed to listen 8080 (all interfaces)
so Traefik can reach it via host.docker.internal:8080.
asepharyana