feat(gateway): detect pre-existing streamers on bot voice join

If someone is ALREADY sharing screen / camera on when the bot joins the channel, no voiceStateUpdate with streaming:true fires for them, so the bot never sent STREAM_WATCH and missed their video entirely. trackChannel now scans channel.members and starts a watch for anyone already streaming (ignoring the bot itself and non-streamers). Idempotent: startStreamWatch no-ops if a watch already exists. +2 tests (9/9 in videoRecorder).
fix(gateway): swap stateChange arg order so Ready actually attaches UDP
2026-08-31 15:48:20 +07:00 · 2026-08-31 15:42:12 +07:00 · 2026-08-31 15:36:05 +07:00 · 2026-08-31 15:23:57 +07:00 · 2026-08-31 14:54:32 +07:00 · 2026-08-31 14:09:06 +07:00
152 changed files with 10479 additions and 4510 deletions
@@ -85,8 +85,8 @@ BACKLOG_SYNC_BATCH_SIZE=100             # Messages per backlog batch, max 100 (d
 # === AI Analysis ===
 AI_ANALYSIS_ENABLED=false               # Enable AI content moderation (default: false)
 AI_LLM_API_KEY=                        # REQUIRED if AI_ANALYSIS_ENABLED=true. LLM API key
-AI_LLM_BASE_URL=https://9router.asepharyana.my.id/v1  # LLM API base URL (9router — OpenAI-compatible router, replaces omniroute)
-AI_LLM_MODEL=claude-opus-5                   # LLM text model name (default: claude-opus-5)
+AI_LLM_BASE_URL=http://100.121.180.82:20128/api/v1  # LLM API base URL (omniroute — OpenAI-compatible router on imrnes, Tailscale 100.121.180.82)
+AI_LLM_MODEL=text                       # LLM text model name (default: text)
 # AI_LLM_VISION_MODEL=                   # Vision model for image analysis (falls back to AI_LLM_MODEL)
 # AI_LLM_EMBEDDING_MODEL=                # Embedding model for semantic moderation cache (optional; enables near-duplicate text reuse to save LLM calls)
 # AI_LLM_EMBEDDING_MIN_SIMILARITY=0.97   # Min cosine similarity to reuse a cached verdict (default: 0.97)
@@ -0,0 +1,17 @@
+name: Dependabot Auto-Merge
+on: pull_request
+
+permissions:
+  contents: write
+  pull-requests: write
+
+jobs:
+  auto-merge:
+    runs-on: ubuntu-latest
+    if: github.actor == 'dependabot[bot]'
+    steps:
+      - name: Enable auto-merge for Dependabot PR
+        run: gh pr merge --auto --merge "$PR_URL"
+        env:
+          PR_URL: ${{ github.event.pull_request.html_url }}
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -0,0 +1,38 @@
+version: 2
+updates:
+  - package-ecosystem: "npm"
+    directory: "/services/backend"
+    schedule:
+      interval: "daily"
+    open-pull-requests-limit: 10
+    groups:
+      production:
+        dependency-type: "production"
+        update-types: ["minor", "patch"]
+      development:
+        dependency-type: "development"
+        update-types: ["minor", "patch"]
+  - package-ecosystem: "npm"
+    directory: "/services/discord-gateway"
+    schedule:
+      interval: "daily"
+    open-pull-requests-limit: 10
+    groups:
+      production:
+        dependency-type: "production"
+        update-types: ["minor", "patch"]
+      development:
+        dependency-type: "development"
+        update-types: ["minor", "patch"]
+  - package-ecosystem: "npm"
+    directory: "/services/frontend"
+    schedule:
+      interval: "daily"
+    open-pull-requests-limit: 10
+    groups:
+      production:
+        dependency-type: "production"
+        update-types: ["minor", "patch"]
+      development:
+        dependency-type: "development"
+        update-types: ["minor", "patch"]
@@ -1,20 +0,0 @@
-name: Publish to FlakeHub
-
-on:
-  push:
-    branches: [main, master]
-  workflow_dispatch:
-
-jobs:
-  flakehub-publish:
-    runs-on: ubuntu-latest
-    permissions:
-      id-token: write
-      contents: read
-    steps:
-      - uses: actions/checkout@v6
-      - uses: DeterminateSystems/determinate-nix-action@main
-      - uses: DeterminateSystems/flakehub-push@main
-        with:
-          visibility: public
-          rolling: true
@@ -0,0 +1,56 @@
+# Spec: Simpan pesan NSFW tanpa analisis AI (Request 1)
+
+Date: 2026-08-30
+
+## Goal
+Saat ini GMW **skip capture** untuk pesan di channel age-restricted/NSFW
+(`messageCapture.ts` baris 291 & 310 memanggil `isAgeRestrictedMessage` lalu
+`return`). User ingin pesan NSFW tetap **disimpan** ke database (jadi terlihat
+di dashboard), tetapi **tidak dianalisis AI** (tidak dipanggil LLM).
+
+## Behavior target
+1. Pesan NSFW/age-restricted di `messageCreate` → DISIMPAN (capture normal).
+2. Pesan NSFW di `messageUpdate`/`messageDelete` → tetap diproses seperti pesan
+   biasa (edit/deleted state tercatat).
+3. AI analysis TIDAK berjalan untuk pesan NSFW. Sudah ada jalur yang benar:
+   `queueMessageAnalysis` → `isAgeRestrictedMessage(message)` →
+   `buildAgeRestrictedSkipResult()` yang menulis `status=clean` + flag
+   `age_restricted` + `action=none` TANPA memanggil LLM. Jalur ini sudah ada dan
+   dipakai di `aiAnalyzer.queueMessageAnalysis` (baris 72-84) dan
+   `batchProcessor.skipAgeRestrictedMessages`. Jadi tinggal membuka capture.
+4. Konten NSFW TIDAK masuk ke Qdrant public archive (semantic search publik).
+   `archiveMessageEmbedded` harus di-guard untuk pesan age-restricted.
+
+## Files touched
+- `services/discord-gateway/src/modules/message-capture/messageCapture.ts`
+  - Hapus guard `isAgeRestrictedMessage` di `messageCreate` (baris 291) dan
+    `messageUpdate` (baris 310). Jangan hapus guard `isExcludedThread`,
+    `isBotExcludedChannel`, `shouldCaptureForAnyTarget`.
+- `services/discord-gateway/src/modules/message-capture/archiveEmbedder.ts`
+  - `archiveMessageEmbedded` terima flag/cek metadata age-restricted → skip
+    embed untuk NSFW.
+  - Call-site di `messageCapture.ts` `captureMessage` (baris 218) pass isNSFW
+    atau cek dulu.
+
+## Schema/type changes
+- TIDAK ada perubahan DB schema. Metadata pesan sudah membawa `channel.nsfw`
+  (`getMessageLocation`). Tidak perlu kolom baru — flag `age_restricted` sudah
+  ditulis ke `ai_moderation_flags` lewat skip-result.
+
+## Verification
+- `pnpm typecheck`, `pnpm lint` (biome), `pnpm build` di `services/discord-gateway`.
+- Unit test: pastikan ada test untuk age-restricted skip (sudah ada
+  `tests/conversationContext.test.ts` ref nsfw; cek apakah ada test sesuai).
+- Deploy via GHA push; verify pesan NSFW muncul di DB dan `ai_status` = clean
+  dengan flag age_restricted, dan TIDAK ada panggilan LLM (log llm-caller tidak
+  menampilkan id pesan NSFW).
+
+## Status
+- Request 1: DONE & DEPLOYED (commit 50967f64, CI green, gateway restart 20:17 WIB).
+- Request 2 (video): dicatat, belum dikerjakan — lihat section di bawah.
+
+## Request 2 (video) — NOT in this change
+Fitur record video kamera/screenshare orang lain. @discordjs/voice hanya
+mendukung **audio** receive. Video orang lain butuh WebRTC viewer baru
+(SDP answer, decrypt H264/VP8, decode frame, mux MP4/WebM). Diluar scope
+changeset ini; dicatat untuk desain lanjutan.
@@ -0,0 +1,82 @@
+# Spec: Recordings — filter per user + export WAV (Audacity)
+
+## Konteks / Gejala
+Halaman `services/frontend/src/app/(dashboard)/recordings` menampilkan semua
+rekaman voice (deck). User ingin:
+1. **Filter per orang** (tampil rekaman satu user saja).
+2. **Export ke format untuk Audacity** (buka & edit rekaman di Audacity).
+
+## Fakta saat ini (verified)
+- Backend `recordings.list` (services/backend/src/orpc/router.ts:291) SUDAH
+  menerima `userId`/`channelId` filter → `RecordingsService.getRecent`.
+- Frontend `recordingsApi.list(limit, channelId, userId, cursor)` (lib/api/recordings.ts)
+  sudah meneruskan `userId`. `useLoadMoreRecordings` juga sudah bawa userId.
+- Tapi UI `RecordingsView` (app/(dashboard)/recordings/view.tsx) TIDAK punya
+  filter UI, dan `useRecordingsPage` dipanggil tanpa userId → semua tampil.
+- Setiap rekaman punya `download_url` (MP3 di TeleUploader), `user_id`, `username`.
+- Audacity membuka MP3/OGG tapi editing paling bersih dari WAV (uncompressed)
+  / FLAC (lossless). Backend TIDAK punya ffmpeg & Nix flake backend tak include
+  ffmpeg → transcode server-side bukan pilihan. Browser punya codec MP3 → export
+  WAV via Web Audio API (client-side) adalah solusi self-contained terbaik.
+
+## Keputusan desain
+1. **Filter per user**: UI dropdown (Semua User + per user) di header halaman.
+   Memilih user → re-fetch `recordingsApi.list(50, undefined, userId)` (server
+   filter, benar untuk dataset besar + pagination). Dropdown dibangun dari
+   distinct `user_id`/`username` pada items yang sedang tampil.
+2. **Export WAV (Audacity)**: client-side via Web Audio API.
+   - Per kartu: tombol "WAV" → decode `download_url` → WAV 16-bit PCM → download.
+   - Header: tombol "EXPORT WAV (N)" → gabung (concat) semua rekaman yang
+     sedang tampil (ter-filter) jadi 1 file WAV → download. Ideal untuk analisis
+     / mixdown per orang.
+   - Implementasi di `lib/audio/wav.ts` (decode + encode + concat), tanpa dep baru.
+
+## Perubahan
+
+### Frontend
+- **`src/lib/audio/wav.ts`** (baru):
+  - `decodeAudio(url: string): Promise<AudioBuffer>` — fetch arrayBuffer →
+    `new AudioContext().decodeAudioData`.
+  - `audioBufferToWav(buf: AudioBuffer, sampleRate=48000): Blob` — PCM 16-bit
+    interleaved, mono→stereo handling, RIFF/WAVE writer. Audacity-importable.
+  - `concatBuffers(buffers: AudioBuffer[]): AudioBuffer` — gabung di channel 0
+    (mono) dengan sample-rate max; untuk export gabungan.
+  - `downloadWav(blob: Blob, filename: string): void` — obj URL + <a download>.
+- **`src/app/(dashboard)/recordings/view.tsx`**:
+  - Toolbar filter: dropdown user (built from distinct items) + tombol reset.
+  - State `filterUserId`; saat berubah → `recordingsApi.list(50, undefined, id)`
+    → set ke SWR (key includes filter), reset pagination.
+  - Tombol "WAV" per kartu (disabled jika `!r.download_url`).
+  - Tombol "EXPORT WAV (N)" di header (disabled jika 0 item punya download_url);
+    concat semua items ter-filter yang punya download_url.
+  - Status loading saat export (spinner/disable).
+- **`src/hooks/use-recordings.ts`**: `useRecordingsPage` menerima `userId?` dan
+  memasukkan ke key + call, supaya filter re-fetch bersih (per-user cache key).
+  `useRecordings`/`useLoadMoreRecordings` propagate `userId`.
+- **`src/lib/types/recording.ts`**: tidak berubah (userId dari items).
+
+### Backend / gateway
+- Tidak ada perubahan. Filter & export sepenuhnya frontend.
+
+## File yang disentuh (frontend only)
+- `src/lib/audio/wav.ts` (baru)
+- `src/app/(dashboard)/recordings/view.tsx`
+- `src/hooks/use-recordings.ts`
+
+## Verification
+1. `cd services/frontend && pnpm typecheck` (tsc --noEmit) — 0 error.
+2. `pnpm lint` (biome check src/) — exit 0.
+3. `pnpm build` (next build) — hijau.
+4. Manual (user): buka /recordings; pilih user di dropdown → hanya rekaman user
+   itu; klik WAV di kartu → file .wav ter-download & terbuka di Audacity; klik
+   EXPORT WAV (filtered) → satu .wav gabungan.
+5. Push → CI `Build & Deploy (Nix)` (frontend job) hijau → deploy landing.
+
+## Risiko / Trade-off
+- Web Audio decode MP3 di client: butuh CORS pada download_url (TeleUploader
+  asepharyana.my.id — sudah same-serve/proxied, CORS ikut origin). Jika 403/CORS
+  gagal, error toaster + fallback manual (RAW MP3 tetap ada).
+- concat gabungan = mono 48k; Audacity bisa edit per-channel nanti. Acceptable.
+- Filter client (dropdown dari items yang dimuat) hanya menawarkan user yang
+  sudah tampil; dataset besar bisa pakai search nanti. Server filter benar untuk
+  yang dipilih.
@@ -0,0 +1,81 @@
+# Spec: Recordings v2 — transcription, search, filters, leaderboard, sessions
+
+## Konteks
+4 fitur lanjutan untuk halaman /recordings (dipilih user):
+1. Tampilkan transkripsi + search by kata kunci
+2. Filter lanjutan: by channel + rentang tanggal
+3. Kelompokkan klip jadi "sesi rapat" + autoplay berurutan + export satu sesi
+4. Leaderboard bicara per user + ringkasan
+
+## Fakta terverifikasi (2026-08-30)
+- `voice_recordings` kolom: id, user_id, username, avatar_url, guild_id,
+  channel_id, channel_name, filename, size_bytes, download_url, upload_status,
+  upload_error, created_at, uploaded_at, **transcription** (schema
+  `shared/database/schema.ts:246`). Index user_id/channel_id/created_at.
+- **Transcription 0/10.640** terisi prod: `AI_VOICE_TRANSCRIPTION_ENABLED`
+  default **false** (config/index.ts:333) & tidak diset di BWS env →
+  `transcribeRecording` (voiceTranscriber.ts) langsung return null.
+  `AI_LLM_BASE_URL` + `AI_LLM_API_KEY` SUDAH dikonfig (Whisper via router GMW).
+- Transcriber hardcode `language: "en"` (voiceTranscriber.ts:34) — salah utk
+  ucapan campur id/en. Utk auto-detect: hapus param `language` (Whisper
+  auto-detect jika tidak diberikan).
+- Backend `RecordingsService.getRecent` (recordings.service.ts) TIDAK select
+  `transcription`; SUDAH dukung filter `channelId`+`userId`+`cursor`; belum
+  dukung date-range & keyword search. `RecordingRow` interface juga tak punya
+  `transcription`.
+- Tidak ada kolom `session_id` / `duration_ms` → sesi grouping = heuristik
+  (channel sama + gap created_at), durasi leaderboard = estimasi dari
+  size_bytes (MP3 128kbps: durasi_s ≈ size_bytes*8/128000).
+
+## Keputusan desain
+1. **Aktifkan transkripsi (fondasi)**: transcriber auto-detect (hapus
+   `language:"en"`), set secret BWS `AI_VOICE_TRANSCRIPTION_ENABLED=true`
+   (dibaca runtime oleh bws-exec saat service start). Rekaman BARU dapat
+   transkripsi. Backfill rekaman lama TIDAK dilakukan (pilih user: fokus baru;
+   file OGG lama kemungkinan besar sudah tidak dipakai).
+2. **Backend** — perluas `getRecent`:
+   - select `transcription` (+ interface RecordingRow + FE type)
+   - filter baru: `q` (ILIKE on transcription + username), `startDate`/`endDate`
+     (created_at range, bigint ms)
+   - endpoint baru `recordings.summary`: agregasi per user → {user_id,
+     username, avatar_url, clips, est_duration_s, words, last_at}. `words`
+     dihitung dari transcription (tokenisasi spasi). Return sorted by clips.
+3. **Frontend**:
+   - Kartu: tampilkan transkripsi (collapse/expand line-clamp) + durasi estimasi.
+   - Toolbar: search box (q), Select channel, date range (start/end), speaker
+     (sudah ada), reset filter.
+   - Tab/segment "Tape Deck" vs "Leaderboard": leaderboard render summary per
+     user + klik → filter deck by user itu.
+   - Sesi grouping (deck view): klip di-group jadi sesi bila channel sama &
+     gap antar klip < SESSION_GAP_MS (default 120s). Header sesi (channel,
+     waktu mulai, jumlah klip, total durasi). Autoplay tombol "Play session" &
+     "Export session WAV" (concat klip sesi via lib/audio/wav.ts yg sudah ada).
+
+## Perubahan file
+
+### Gateway (Tahap 1)
+- `voiceTranscriber.ts`: hapus baris `language: "en"`.
+- (secret) set `AI_VOICE_TRANSCRIPTION_ENABLED=true` via bws.
+
+### Backend (Tahap 2)
+- `recordings.service.ts`: interface + select + getRecent tambah transcription;
+  tambah filter q/startDate/endDate; method getSummary() untuk leaderboard.
+- `orpc/router.ts`: procedur `recordings.list` schema tambah fields; prosedur
+  baru `recordings.summary`.
+
+### Frontend (Tahap 3-5)
+- `lib/types/recording.ts`: tambah transcription, est_duration_s opsional,
+  Summary type.
+- `lib/api/recordings.ts`: list tambah q/startDate/endDate; + summary().
+- `hooks/use-recordings.ts`: propagate filter baru ke key+call; hook
+  useRecordingsSummary.
+- `app/(dashboard)/recordings/view.tsx`: toolbar search+channel+date, kartu
+  transkripsi, tab leaderboard, grouping sesi + autoplay + export sesi.
+
+## Verification
+- tiap tahap: gateway `pnpm build` + `biome check src/`; backend `pnpm build`
+  + `biome check src/ tests/`; FE `pnpm build` + `biome check src/`.
+- CI Build & Deploy hijau tiap tahap; deploy landing dicek via
+  `systemctl show gmw-<svc>.service --property=ActiveEnterTimestamp`.
+- Tahap 1c: setelah deploy + rekaman baru, cek
+  `SELECT COUNT(*) FROM voice_recordings WHERE transcription IS NOT NULL`.
@@ -0,0 +1,71 @@
+# Spec: Record video (kamera/screenshare) orang lain — WebRTC receive (Request 2)
+
+Date: 2026-08-30. Status: Phase A + B DONE (capture → playable MP4); Phase C (UI) open.
+
+## Why this is hard (ground truth, verified from @discordjs/voice 0.19.2 source)
+`VoiceReceiver.onUdpMessage` (dist/index.mjs:2059) drops EVERY non-opus RTP
+packet at line 2068: `if ((msg[1] & 127) !== RTP_OPUS_PAYLOAD_TYPE) return;`.
+So video (kamera H264? actually Discord uses VP8/H264; screenshare combines with
+video SSRC) is decrypted-capable but never forwarded. `receiver.parsePacket`
+(2033) DOES decrypt any payload type generically (audio + video) using
+`connectionData.{encryptionMode, nonceBuffer, secretKey}` — the only audio gate
+is the opus check inside onUdpMessage.
+
+=> FIX: wrap `receiver.onUdpMessage` (like screenShareAudio.ts already does for
+screen-share AUDIO SSRCs): for packets whose payload type is a VIDEO type
+(payload 96 VP8, 101/102 H264, 106/116/126/127 AV1, VP9 98...), call
+`receiver.parsePacket(...)` myself to decrypt, then depacketize + write frames.
+Delegate opus (120) to the original handler. Delegate audio to original.
+
+## Audio already works (screenShareAudio.ts). We add VIDEO.
+
+## Science-of-the-changes below.
+
+## Phase A — capture + decrypt + depacketize to AnnexB h264 (THIS change)
+Files (new): `src/modules/voice-recording/videoReceiver.ts`
+- Hook into `recorder.startRecording` alongside `hookScreenShareAudio`.
+- Wrap `receiver.onUdpMessage`:
+  - read ssrc = msg.readUInt32BE(8); userData = receiver.ssrcMap.get(ssrc)
+  - if payload type is video AND we have a "watching" subscription for that user
+    (videoSSRC present), decrypt via receiver.parsePacket(...), then:
+      - H264 (101/102 + payload 120 not): strip RTP header, reassemble FU-A
+        fragments into AnnexB NALs (start-code prefixed), buffer until we have
+        a full access unit (keyframe SPS/PPS/IDR or slices), append to a per-
+        user-per-burst `.h264` file.
+  - else delegate to original onUdpMessage.
+- Watch `receiver.ssrcMap` "create"/"update" for `videoSSRC !== undefined` →
+  signal a video burst started for that user (like screenShareAudio does).
+- Per-user video files written to `config.RECORDINGS_DIR/<uid>/video-<ts>.h264`.
+- Guard: skip bot's own video (client.user.id).
+
+Dependencies: NO new npm deps for Phase A (only crypto already in
+@discordjs/voice via parsePacket + Buffer). ffmpeg-headless (already in Nix
+buildInputs) used in Phase B for decode+mux.
+
+## Phase B — decode + mux to playable MP4/WebM (DONE, commit 999c054b)
+- `closeBurst` waits for the WriteStream `finish` (full flush/fd close), then
+  `muxToMp4(rawPath)`: `ffmpeg -f h264 -i raw.h264 -c copy -movflags +faststart
+  out.mp4`, deletes raw on success (>=1B mp4), keeps it on failure.
+- Output: `<RECORDINGS_DIR>/<uid>/video-<ssrc>-<ts>.mp4`.
+- ffmpeg is on the gateway runtime PATH (pkgs.ffmpeg-headless, already in the
+  Nix buildInputs for the music/GoLive players).
+- Test: `tests/videoReceiver.test.ts` muxToMp4 case (real ffmpeg, generates a
+  tiny baseline h264, asserts mp4 non-empty + raw deleted; skipped if no ffmpeg).
+
+## Phase C — frontend playback + session grouping (follow-up)
+- Backend oRPC list video files; FE video player, group by call session like audio.
+
+## Verification
+- Phase A: join voice, have a member screen-share/camera, confirm `.h264` file
+  grows with NAL frames + keyframes; journal shows "video burst" logs.
+- Run vitest unit: RTP header strip + FU-A reassembly gives correct bytes.
+
+## Open questions / risks
+- Discord codec for camera = H264(101/102); screenshare uses H264 (101/103?)
+  and can also be VP8/VP9. Handle H264 first (depacketize proven), VP8/VP9 in
+  Phase B via ffmpeg RTP input.
+- Encryption: DAVE (dave_protocol_version) adds a session layer; parsePacket
+  already applies daveSession.decrypt for audio — we must call the SAME
+  parsePacket path so DAVE/encryption is handled identically.
+- ssrc↔user mapping during a broadcast: videoSSRC is in ssrcMap after the
+  voice state; may need the STREAM_CREATE network events to key reliably.
@@ -0,0 +1,66 @@
+# Spec: Voice auto-reconnect (persistent state + rejoin on drop)
+
+## Goal
+Setelah `VoiceController.connect()` berhasil, state "sedang merekam di <guild>/<channel>"
+disimpan di Postgres. Kalau gateway restart/reboot, atau koneksi voice drop tidak
+disengaja (dikeluarkan/moved/server restart), gateway otomatis join ulang ke channel
+yang sama.
+
+## Requirement mapping (user's ask)
+- "autoreconnect ke channel yg sama jika server restart atau reboot" → reconnect on
+  startup (ready handler) + keep DB record across graceful shutdown.
+- "state nya persistent di db" → `voice_auto_reconnect` table.
+- "rejoin jika tidak sengaja dikeluarkan" → watchdog on `Disconnected`/`Destroyed`
+  (kick / moved / voice server restart) → full rejoin with backoff.
+- Manual leave (`/voice disconnect`, dashboard disconnect) MUST NOT rejoin.
+
+## Design decisions
+1. **New table** `voice_auto_reconnect` (dedicated, not `ui_state`):
+   - `guild_id` text PK
+   - `channel_id` text NOT NULL
+   - `channel_name` text
+   - `connected_at` bigint epoch-ms
+   - `updated_at` bigint epoch-ms
+   DAO: `voiceAutoReconnectRepo.ts` — `upsert(record)`, `list()`, `delete(guildId)`.
+2. **Write on connect**: `VoiceController.connect()` → after `startRecording` success →
+   `upsert`. IDEMPOTENT (upsert per guild).
+3. **Clear on manual leave**: `handleVoiceDisconnect` (all) + `handleVoiceDisconnectGuild`
+   pass `clearPersisted: true`. Graceful shutdown `disconnect()` keeps the record.
+4. **Rejoin on startup**: bootstrap `ready` → `await voiceController.autoReconnect()`
+   (list persisted → connect each, non-fatal on failure).
+5. **Rejoin on unexpected drop**: monitor per-connection; on `Disconnected`/`Destroyed`
+   with `!intentional` → schedule rejoin `connect(guildId, persistedChannelId)` with
+   backoff (min 2s, max 30s, max 5 attempts). Track `rejoinAttempts`, reset on success.
+6. **Intentional flag**: `disconnectGuild(guildId, { clearPersisted?, intentional? })`.
+   - shutdown `disconnect()` → `{ intentional: true, clearPersisted: false }`.
+   - manual `disconnect()` (dashboard) → `{ clearPersisted: true }`, sets intentional.
+   - manual `disconnectGuild` → `{ clearPersisted: true }`, sets intentional.
+   The monitor checks `intentional` before rejoin; `clearPersisted` only deletes the row.
+
+## Files touched
+- `services/discord-gateway/src/shared/database/schema.ts` — add `pgVoiceAutoReconnectTable`
+  + types.
+- `services/discord-gateway/src/shared/database/voiceAutoReconnectRepo.ts` (NEW) — DAO.
+- `services/discord-gateway/src/modules/voice-recording/voiceController.ts` — upsert on
+  connect; monitor + rejoin; `autoReconnect()`; `disconnect/disconnectGuild` opts.
+- `services/discord-gateway/src/modules/command-handler/voice.handler.ts` — manual
+  disconnect/disconnectGuild pass `clearPersisted: true`.
+- `services/discord-gateway/src/app/bootstrap.ts` — call `voiceController.autoReconnect()`
+  in `ready`.
+- `services/discord-gateway/drizzle/migrations/0020_add_voice_auto_reconnect.sql` +
+  `meta/_journal.json` entry (apply manually per gmw-ops).
+
+## Edge cases
+- Channel deleted / guild lost while persisted → `connect()` throws (channel not found)
+  → log + delete persisted row (don't retry forever).
+- Rejoin attempts exhausted → keep row (so next restart retries) + log.
+- Multiple guilds: per-guild monitor, per-guild persisted row.
+- Graceful shutdown order: shutdown sets intentional=true (so no rejoin during teardown)
+  but keeps row.
+
+## Verification
+- `pnpm typecheck && pnpm build && pnpm lint` in `services/discord-gateway`.
+- Apply migration `0020` manually; verify table exists.
+- CI `Build & Deploy (Nix)` green; gateway deploy lands.
+- Manual: connect via dashboard → check `voice_auto_reconnect` row; simulate drop →
+  confirm rejoin; manual disconnect → row cleared.
@@ -0,0 +1,166 @@
+# Spec: Perbaiki alur voice → recording (miss & terpotong)
+
+## Konteks & Gejala
+User melaporkan alur voice sampai recording **banyak miss** (audio tidak tercatat)
+dan **terpotong** (satu alur bicara kebelah jadi beberapa segmen / audio putus di
+tengah). Ini domain `services/discord-gateway/src/modules/voice-recording/`.
+
+Pipeline per user yang mulai bicara (speaking "start"):
+```
+receiver.speaking "start" → speakingHandler(userId)
+  ├─ await collectUserMetadata(...)        ← roundtrip API, subscribe tertunda
+  ├─ receiver.subscribe(userId, {end: AfterSilence, duration: 3000ms}) → audioStream
+  ├─ attach data/end/error handlers  →  audioStream.pipe(PacketFilter) → oggPacketStream
+  ├─ SegmentManager.open() → OggLogicalBitstream → file .ogg
+  ├─ data: SegmentManager.rotateIfNeeded (rotasi 5s) + decoder.write (web PCM tho
+  └─ end: SegmentManager.close() → segmen finish → finalizeSegment upload + transkrip
+```
+
+## Root cause (dari pembacaan kode — justifikasi di bawah)
+
+### A. MISS bagian awal bicara — subscribe tertunda (utama)
+`speakingHandler.ts:53` melakukan `await collectUserMetadata(...)` SEBELUM
+`receiver.subscribe`. `collectUserMetadata` (metadata.ts:30) pada cold path
+(cache miss) melakukan `client.users.fetch` + `guild.members.fetch` roundtrip
+Discord API (ratusan ms–detik). Selama await, seluruh opus awal bocor → awal
+kalimat hilang. Cache menghilangkan ini untuk user yang pernah ter-record, tapi
+user baru/evict (cache max 200) kena setiap kali.
+
+### B. Double-subscribe race
+Guard `receiver.subscriptions.has(userId)` di `speakingHandler.ts:63` diletakkan
+SETELAH `await collectUserMetadata`. Dua event "start" cepat keduanya melewati
+guard (belum subscribe) → dua subscription → audio terbelah/ganda per user.
+
+### C. TERPOTONG di jeda bicara — AfterSilence 3000ms
+`AUDIO_STREAM_SILENCE_DURATION_MS=3000`. Setelah 3s diam, stream auto-`end` →
+`SegmentManager.close` → segmen baru saat bicara lagi. Jeda normal (berpikir,
+interupsi) memecah 1 alur bicara jadi beberapa segmen/file. Ini source "terpotong".
+Segmen pendek hasil jeda <1s juga DIBUANG di `finalizeSegment` (MIN_DURATION_MS=1000)
+→ miss kata singkat ("ya", "siap").
+
+### D. Rotasi segmen 5s di tengah bicara
+`RECORDING_SEGMENT_MS=5000`: `rotateIfNeeded` menutup bitstream & membuka baru
+setiap 5s walau bicara kontinu. Pipenya di-re-wire di dalam handler data →
+window drop kecil + continuity file pecah (bukan masalah besar, tapi berkontribusi).
+
+### E. Tidak ada sinkronisasi "stop" speaking & stream "end" flaky
+Handler hanya listen "start"; mengandalkan `AfterSilence` untuk emit "end".
+Bug @discordjs/voice yang dikenal: `AfterSilence` bisa TIDAK emit "end" saat
+koneksi gagal/teardown → segmen menggantung & tidak pernah finalize/upload
+(recording "hilang"). Tidak ada watchdog.
+
+## Scope
+Hanya `services/discord-gateway/src/modules/voice-recording/` (+ config index bila
+perlu default baru). Tidak menyentuh playback (player.ts), transmitter (voice dari
+browser → Discord, arah berlawanan), muxer (konsolidasi akhir), atau screen-share
+video (hanya audio SSRC via `hookScreenShareAudio` sudah ada & dibiarkan).
+
+## Perubahan
+
+### 1. Speak-before-metadata: subscribe LEBIH DULU, metadata paralel
+`recorder/speakingHandler.ts`:
+- Pindahkan `receiver.subscribe` + pipeline setup ke ATAS, SEGERA di handler,
+  sebelum `collectUserMetadata`.
+- Jalankan `collectUserMetadata` secara paralel non-blocking; gunakan metadata
+  cache untuk registrasi segmen saat finalize.
+- Pertahankan guard skip bot/user (bot bisa dicek dari `client.users.cache` /
+  `client.user.id` tanpa await) SEBELUM subscribe — jangan tunggu fetch user.
+
+Detail konkret:
+```
+async handler(userId):
+  if (userId === client.user?.id) return;
+  if (receiver.subscriptions.has(userId)) return;   // guard kini di DEPAN, tanpa await
+  // (belum tahu bot? gunakan cache user; subscribe dulu biar nggak miss)
+  clone = subscribe(userId, {AfterSilence, duration})  // TANPA await metadata
+  setup pipeline (data/end/error, pipe, open segment)
+  collectUserMetadata(...).then(meta => {
+     if (meta.bot) { drain & close subscription (jangan simpan) }
+     else { registrasi ulang metadata utk segmen aktif }
+  })
+```
+Karena listener `start` dibuang untuk bot, harus tutup subscription bot tanpa
+menyimpan segmen (buang hasil). Pakai `receiver.subscriptions.get(userId)?.destroy()`.
+
+### 2. Selesaikan race double-subscribe (bagian dari #1)
+Guard `receiver.subscriptions.has(userId)` diletakkan SINCRON di awal (sebelum
+await). Karena `subscribe` sinkron dan `subscriptions` terisi sinkron saat
+dipanggil, event "start" kedua yang tiba setelah subscribe akan melihat
+subscription aktif → di-skip. Tidak ada await antara guard & subscribe.
+
+### 3. Naikkan AfterSilence + tail-length → kurangi "terpotong"
+`shared/config/index.ts`:
+- `AUDIO_STREAM_SILENCE_DURATION_MS` default 3000 → 4000 (beri ruang jeda
+  alami; Discord packet 20ms, 4s masih wajar, tidak membengkak file).
+Opsional via env override di production (tidak wajib komit env).
+
+### 4. Segmen berorientasi "burst bicara" daripada rotasi jam
+`recorder/segment.ts` + `recorder/speakingHandler.ts`:
+- Hapus/lepas rotasi SEGMEN berbasis waktu (RECORDING_SEGMENT_MS). Alih-alih,
+  satu segmen = satu burst bicara (buka di "start", tutup di "end"/AfterSilence
+  end). Ini menghilangkan pemecahan di tengah kalimat.
+- `RECORDING_SEGMENT_MS` tetap dipakai untuk rotasi decoder web-PCM (broadcast
+  live), di mana segmen besar bisa menunda frame — biarkan seperti ada.
+  JADI: `SegmentManager.rotateIfNeeded` TIDAK lagi dipanggil pada jalur OGG
+  recording; decoder rotate tetap dijalankan.
+
+Catatan: dengan satu segmen per burst, ukuran file ~ durasi bicara. File panjang
+dibutuhkan transkrip & transcode; tidak ada batas keras yang perlu di-override.
+Watchdog di #5 membatasi durasi menggantung.
+
+### 5. Watchdog end-of-burst & teardown recovery
+`recorder/speakingHandler.ts`:
+- Setelah subscribe, arm timer watchdog (mis. `config twin`/hitung) yang menutup
+  segmen jika `AfterSilence` tidak emit "end" dalam X detik setelah "stop"
+  speaking — atau, lebih sederhana & robust: dengarkan BOTH stream "end" DAN
+  timer dari `receiver.speaking` "stop" (hingga @discordjs/voice meng-klaim
+  AfterSilence). Bila "stop" fire, mulai countdown kecil (mis. 500ms) lalu
+  `segmentManager.close` + `decoder.destroy` + destroy subscription jika stream
+  belum "end".
+- Ini menutup A: segmen menggantung → jadi pasti finalize & upload.
+
+Implementasi: subscriptionStream (audioStream) + track milik per-user di
+Map<userId, {audioStream, segmentManager, decoder, timer}>; handler "stop"
+menjadwalkan finalize.
+
+### 6. Naikkan/ambil MIN segmen duration lebih rendah
+`recorder/segmentFinalizer.ts`: MIN_DURATION_MS 1000 → 300ms. Kata pendek
+("ya", "siap") tetap tersimpan. GUI biarkan.
+
+## File yang disentuh
+- `src/modules/voice-recording/recorder/speakingHandler.ts` (utama: subscribe
+  first, guard depan, watchdog stop, hapus rotasi segmen dari jalur OGG)
+- `src/modules/voice-recording/recorder/segment.ts` (opsional: API close/open,
+  pertahankan rotate untuk decoder tapi tak dipakai jalur OGG)
+- `src/modules/voice-recording/recorder/streamSetup.ts` (kecil: backfill subscribe
+  supaya return subscription utk cleanup/destroy bot)
+- `src/modules/voice-recording/recorder/segmentFinalizer.ts` (MIN_DURATION)
+- `src/shared/config/index.ts` (default AfterSilence 4000)
+
+## Yang TIDAK disentuh
+- `transmitter.ts` (arah browser→Discord, bukan recording)
+- `player.ts`, `mediaSource.ts`, `screenShareAudio.ts` (hook SSRC sudah benar)
+- `muxer.ts` (konsolidasi akhir tetap jalan)
+- Flake/deps/bundle
+
+## Verification
+1. `cd services/discord-gateway && pnpm typecheck` (tsc --noEmit) — 0 error.
+2. `pnpm lint` (biome check src/) — exit 0.
+3. `pnpm build` (tsc → dist/).
+4. Unit test baru (vitest, kalau infra tes ada):
+   - subscribe terjadi tanpa await metadata (spy urutan panggilan)
+   - double-start skip lewat guard sinkron
+   - "stop" → watchdog finalize segmen walau stream tidak "end"
+   - MIN_DURATION 300ms menyimpan kata pendek
+5. Smoke/CI: `nix flake check` (eval). Push → CI `Build & Deploy (Nix)` hijau →
+   deploy landing (`systemctl show gmw-discord-gateway.service --property=ActiveEnterTimestamp`).
+6. Runtime manual (user): join voice, bicara dengan jeda >3s, pastikan 1 alur
+   kontinu = 1 segmen utuh (bukan 3), dan awal kata tidak hilang.
+
+## Risiko / Trade-off
+- Subscribe-before-metadata: burst bot akan dikumpulkan sesaat lalu dibuang
+  (cost kecil: buang segmen). Lebih baik miss bot daripada miss user.
+- AfterSilence naik: file lebih panjang sedikit saat jeda; upload/transkrip
+  timeout (transcodeToMp3 30s) tetap aman.
+- Satu segmen per burst: tidak ada rotasi paksa → durasi segmen = durasi bicara
+  (bisa menit). Transkrip & transcode tetap ok. Watchdog batasi menggantung.
@@ -0,0 +1,177 @@
+# Spec: Receive Others' Screen-Share/Camera Video Under DAVE — Build a DAVE-capable Stream-Watch Connection
+
+Status: **P1–P3 DONE (implemented + deployed); P4 (live verify) BLOCKED on a real streamer**
+Date: 2026-08-31
+Author: Hermes
+Related: `.hermes/plans/2026-08-31_video-receive-eager-selfbot-connection-spec.md` (superseded by this)
+         `.hermes/plans/2026-08-31_video-receive-phaseC-spec.md` (Phase C build, selfbot path — dead)
+
+## Problem / Ground truth (established from live logs 2026-08-31)
+GMW must record OTHER members' screen-share + camera video in a voice channel it
+records. Audio works (via `@discordjs/voice` 0.19.2 negotiating DAVE). Video does
+not. Verified live: the selfbot path (`discord.js-selfbot-v13` eager `joinChannel`
+→ `joinStreamConnection` → `receiver.createVideoStream`) authenticates but Discord
+closes the connection with WS code **4017 "E2EE/DAVE protocol required"** (5x →
+`VOICE_CONNECTION_ATTEMPTS_EXCEEDED`). Root cause: **Discord now REQUIRES DAVE
+(E2EE) on every voice RTC, and `discord.js-selfbot-v13`'s voice stack predates
+DAVE** (identify has no `max_dave_protocol_version`, no MLS handshake). The selfbot
+path is dead, cannot be repaired. Full details: skill `gmw-ops` →
+`references/video-receive-and-unmute.md` §4.
+
+Facts:
+- Watching a stream = a **SEPARATE RTC connection**, not the guild audio socket:
+  gateway `STREAM_WATCH` (op 20) → Discord replies `STREAM_CREATE` (rtc_server_id)
+  + `STREAM_SERVER_UPDATE` (separate token+endpoint) → client opens its own voice
+  WS+UDP to that endpoint (`StreamConnectionReadonly` in selfbot). The watched
+  video never rides the @discordjs/voice guild socket.
+- The stream-watch RTC ALSO requires DAVE (same 4017 mechanism).
+- `@snazzah/davey` (bundled with @discordjs/voice 0.19.2) supports
+  `MediaType.VIDEO` + `Codec.H264` decrypt — DAVE machinery CAN decrypt H264 video.
+- No off-the-shelf DAVE-capable video-RECEIVE path exists. Closing references:
+  - **Discord-RE/Discord-video-stream** (fork of `@dank074/discord-video-stream`,
+    master 2026-08-28): full DAVE in `src/client/voice/BaseMediaConnection.ts`
+    (Davey `DAVESession` init via `initDave`, MLS key-package / proposals /
+    commit / welcome / transitions; `WebRtcConnWrapper` encrypts audio/video via
+    `daveSession.encrypt(MediaType.VIDEO, codec, …)`). BUT it is STREAMING only
+    (send). No STREAM_WATCH / receive.
+  - `@discordjs/voice`: full DAVE receive for AUDIO only; `DAVESession.decrypt`
+    hardcodes `MediaType.AUDIO` (dist ~line 892); `onUdpMessage` drops non-opus;
+    no STREAM_WATCH.
+  - `discord.js-selfbot-v13`: video receive but no DAVE.
+
+## Goal
+Replace the dead selfbot receive path with a **DAVE-capable stream-watch voice
+connection**: on detecting a member `voiceState.streaming`, send `STREAM_WATCH`,
+connect a DAVE-authenticated RTC to the stream endpoint, decrypt incoming H264
+RTP (`MediaType.VIDEO`), reassemble via the existing `H264Depacketizer`, mux to a
+playable container. Reuse every tested building block already in the repo.
+
+## Strategy decision (default A; B as fallback) — de-risk in Phase 2
+Two implementation routes. Decide by Phase 2 prototype result.
+
+### Strategy A — extend @discordjs/voice's tested native stack (PREFERRED, lighter)
+Reuse @discordjs/voice 0.19.2 internals (already a runtime dep, already DAVE-tested
+for audio):
+- Drive a connection to the stream endpoint using djs/voice's `VoiceWebSocket` +
+  `VoiceUDPSocket` + `DAVESession` (the same classes that work for the guild
+  connection — they take arbitrary endpoint/token/session).
+- Send the voice identify with `max_dave_protocol_version`, complete the DAVE
+  handshake (Davey), then on receipt of a video RTP packet call
+  `daveSession.decrypt(userId, MediaType.VIDEO, packet)` (Davey exposes
+  `MediaType.VIDEO` + `Codec.H264`) — djs/voice's hardcoded `AUDIO` is the only
+  blocker, fix by invoking Davey directly with `MediaType.VIDEO` for video SSRCs.
+- STREAM_WATCH sent via the existing selfbot `client.ws.broadcast` (cheap, works —
+  it needs no selfbot voice connection).
+- Feed decrypted H264 → `H264Depacketizer` → `.h264` → `muxToMp4` (both already
+  in `videoReceiver.ts`, unit-tested).
+- No new runtime deps. Risk: relies on non-exported djs/voice internals (reachable
+  via `as any`, as the existing `parsePacket` usage shows).
+
+### Strategy B — port Discord-RE's BaseMediaConnection (heavier, more self-contained)
+Port `BaseMediaConnection.ts` DAVE handling + `WebRtcConnWrapper` into a
+receive/watch connection in the gateway. Deps: requires `@lng2004/node-datachannel`
+(new native WebRTC dep) + `@snazzah/davey` (already available). More code, more
+risk (native dep in Nix store), but a clean-room receive path decoupled from
+djs/voice internals. Use only if A proves infeasible.
+
+## Files touched (Strategy A shape)
+- `services/discord-gateway/src/modules/voice-recording/streamWatchReceiver.ts`
+  (NEW): DAVE stream-watch connection wrapper. Owns, per watched user:
+  - `sendStreamWatch(client, streamKey)` (via `client.ws.broadcast({op:20,
+    d:{stream_key}})`), stream_key = `guild:<gid>:<chid>:<uid>`.
+  - collects STREAM_CREATE (rtc_server_id) + STREAM_SERVER_UPDATE (token+
+    endpoint) via `client.on('raw')` match on stream_key.
+  - builds a djs/voice-style connection to `<endpoint>` with the received token/
+    session; completes DAVE handshake.
+  - `onUdpMessage` wrapper: for video payload types, decrypt with
+    `daveSession.decrypt(userId, MediaType.VIDEO, buf)`, depacketize, write.
+  - teardown on STREAM_DELETE / user stops streaming / leave / channel untrack.
+- `recorder.ts`: wire `trackChannel`/`untrackChannel` already exist; ensure the
+  selfbot *eager voice connection* attempt is REMOVED (it only 4017-spams logs) —
+  but KEEP `client.ws.broadcast` availability for STREAM_WATCH.
+- `videoRecorder.ts`: remove the dead selfbot `joinChannel`/`joinStreamConnection`
+  calls; keep the `voiceStateUpdate` streaming detection + teardown bookkeeping as
+  the entry point; delegate the actual receive to `streamWatchReceiver`.
+- `videoReceiver.ts`: keep `H264Depacketizer` + `muxToMp4` (reused). The
+  guild-socket `hookVideoReceiver` can be removed or left inert.
+- Tests: `tests/streamWatchReceiver.test.ts` (DAVE-handshake stub, RTP decrypt path
+  with a mocked Davey, STREAM_WATCH packet shape); keep `tests/videoReceiver.test.ts`.
+
+## Phases (each independently verifiable)
+1. **Phase 1 (this session): spec + source reconnaissance.** Confirm djs/voice
+   internals are reachable (VoiceWebSocket/VoiceUDPSocket/DAVESession exports &
+   shapes), confirm Davey `MediaType.VIDEO` decrypt signature, confirm how a raw
+   stream-watch connection's identify/select-protocol flows. Verify the selfbot
+   `streamKey` format + `raw` STREAM_CREATE/SERVER_UPDATE payload. GATE: accurate
+   spec + no unknowns blocking A.
+2. **Phase 2: de-risk prototype.** Standalone script (not in the gateway) that:
+   logs into the same selfbot token, joins a real voice channel, sends STREAM_WATCH
+   for a live streamer, receives STREAM_CREATE/SERVER_UPDATE, and attempts a
+   DAVE-authenticated connect + receive of ≥1 H264 packet to prove the path before
+   any gateway integration. GATE: at least one decrypted H264 NAL captured in the
+   lab.
+3. **Phase 3: gateway integration** per files-touched. GATE: typecheck + build +
+   biome + unit tests green; CI deploy ok.
+4. **Phase 4: live verify.** With a real streamer in a recorded channel: journal
+   shows `STREAM_WATCH sent`, `DAVE ready`, `Video burst opened`, and a playable
+   `.h264`/`.mp4`/`.mkv` on disk. GATE: playable file with real video content.
+
+## Risks / open questions
+- Does Discord require the stream-watch connection to use the SAME session_id as
+  the bot's active voice session, or a fresh one? (Affects identify.) Resolve in P2.
+- Which video codec does Discord actually send for camera vs GoLive (H264 likely,
+  but VP9/AV1 possible) — the depacketizer only handles H264. P2 measures the
+  payload type live; add depacketizers for other codecs only if observed.
+- djs/voice `DAVESession`/`VoiceUDPSocket` reachability via `as any` must be
+  confirmed against the installed 0.19.2 build (P1).
+- The separate stream RTC may need `selectProtocol`/SDP even for receive-only; the
+  Discord-RE SDP shows a `m=video ... inactive` section. Follow the same shape.
+
+## Verification (overall)
+- Per-phase gates above.
+- No regression: audio recording + message capture still work after changes.
+- `pnpm typecheck && pnpm build && pnpm lint` green in discord-gateway.
+- Commit + push; CI `Build & Deploy (Nix)` green; live streamer produces a file.
+
+## Phase 1 findings (CONFIRMED 2026-08-31, Strategy A feasible)
+- `@discordjs/voice` 0.19.2 dist/index.mjs PUBLICLY exports exactly the primitives
+  needed: `DAVESession`, `Networking`, `NetworkingStatusCode`, `VoiceConnection`,
+  `VoiceReceiver`, `VoiceUDPSocket`, `VoiceWebSocket`, `SSRCMap`,
+  `RTP_OPUS_PAYLOAD_TYPE` (export block ~3143). So a stream-watch connection can be
+  built OUTSIDE the lib using these constructors — no `as any` needed for the heavy
+  lifting.
+- `Networking` child wiring (~line 1364-1484): `new VoiceWebSocket('wss://' +
+  endpoint + '?v=8', debug)`; on WS open send Identify `{op, d:{server_id,
+  user_id, session_id, token, max_dave_protocol_version: getMaxProtocolVersion()}}`;
+  `createDaveSession(protocolVersion)` → `new DAVESession(protocolVersion, userId,
+  channelId, {decryptionFailureTolerance})` then `.reinit()`; UDP via
+  `new VoiceUDPSocket({ip, port})` after Ready gives modes + ssrc.
+- `DAVESession` wraps `@snazzah/davey` `Davey.DAVESession(protocolVersion, userId,
+  channelId)`; on network packets it calls `this.session.decrypt(userId,
+  Davey.MediaType.AUDIO, packet)` — hardcoded AUDIO (line ~892). For video we call
+  Davey directly with `MediaType.VIDEO` + `Codec.H264`.
+- `@snazzah/davey` MediaType enum: AUDIO=0, VIDEO=1; Codec H264=4; methods
+  `decrypt(mediaType, codec, packet): Buffer` + `encrypt(...)`. Confirmed in davey
+  index.d.ts.
+- Stream key format (selfbot VoiceConnection.js ~1240): `guild:<gid>:<chid>:<uid>`
+  for guild channels; `STREAM_WATCH` = gateway op 20, `d:{stream_key}`;
+  `sendSignalScreenshare` = `client.ws.broadcast({op:20,d:{stream_key}})`. Replies
+  come as gateway `raw` events `STREAM_CREATE` (`d.rtc_server_id`) +
+  `STREAM_SERVER_UPDATE` (`d.token`, `d.endpoint`); selfbot routes them to the
+  stream connection via `client.on('raw')` matching `d.stream_key`, setting
+  `setSessionId(sessionId)` + `setTokenAndEndpoint(token, endpoint)` (Watch case in
+  `StreamConnectionReadonly`).
+- Discord-RE reference for the identify SDP: stream connections send a `m=video`
+  section with `a=inactive` (receive-only-ish) + standard DAVE/VoiceOpCodes
+  (op 0 identify, op 2 select protocol incl. `max_dave_protocol_version`). See
+  `BaseMediaConnection.handleProtocolAck` + `initDave`.
+- Decision: proceed with **Strategy A**. Selfbot code to REMOVE: the eager
+  `ensureSelfbotVoice` join + `joinStreamConnection`/`receiver.createVideoStream`
+  in `videoRecorder.ts` (proven dead — 4017). Keep the `voiceStateUpdate` streaming
+  detection + bookkeeping; swap the receive plumbing to a new `streamWatchReceiver`
+  driven by a djs/voice-style connection. STREAM_WATCH itself still sent via the
+  selfbot `client.ws.broadcast` (needs only the WS, not a selfbot voice conn).
+- OPEN (resolve in Phase 2 lab): (a) whether the stream connection's identify must
+  use the bot's ACTIVE voice session_id or a fresh one; (b) actual video codec Discord
+  sends for camera vs GoLive (measure payload type live; H264 assumed, VP9/AV1 possible
+  → add depacketizers only if observed).
@@ -0,0 +1,97 @@
+# Spec: Record Other Users' Video (Camera / Screen Share) — Phase C
+
+Status: PLANNED (not built)
+Date: 2026-08-31
+Author: Hermes
+Related: `.hermes/plans/2026-08-30_video-record-receive-spec.md` (Phase A/B — raw UDP hook, superseded for receive)
+
+## TL;DR — what changed vs Phase A/B
+
+Phase A/B (commit `999c054b` etc.) hooked `@discordjs/voice`'s UDP socket to capture non-opus RTP and
+depacketize H264 → mp4. **It captured ZERO video** because `@discordjs/voice` never authorizes the bot to
+receive others' video (no STREAM_WATCH). This spec replaces that approach with the **native, selfbot-lib
+receive path**, which is battle-tested and does the authorization + decryption + ffmpeg muxing for us.
+
+## Ground truth (verified in discord.js-selfbot-v13 3.7.1 source)
+
+1. `ClientVoiceManager.joinChannel(channel, config)` → a **selfbot `VoiceConnection`** with
+   `.receiver` (`VoiceReceiver` → `PacketHandler`). [ClientVoiceManager.js:102-118]
+2. `VoiceConnection.receiver` is created in the constructor. [VoiceConnection.js:140]
+3. `VoiceReceiver.createVideoStream(user, output)` → `PacketHandler.makeVideoStream` → **`Recorder`**
+   (ffmpeg that muxes H264+Opus RTP over UDP → **Matroska (.mkv)**). [Receiver.js, Recorder.js]
+4. `PacketHandler` routes: video RTP → Recorder UDP 65506, opus RTP → UDP 65510; decodes all via
+   `connection.authentication.{secret_key, mode}` (supports `aead_aes256_gcm_rtpsize` and
+   `aead_xchacha20_poly1305_rtpsize` = DAVE-compatible). [PacketHandler.js:115-155, 195-240]
+5. `StreamConnectionReadonly.joinStreamConnection(userId)` + `sendSignalScreenshare()` sends
+   gateway op `STREAM_WATCH` so Discord actually forwards the streamer's RTP to us. [VoiceConnection.js:1100-1240]
+6. `VoiceState.streaming` = `data.self_stream ?? false` — lets us detect a streamer on voice state update. [VoiceState.js:94]
+
+## Problem / the crux
+
+The gateway's voice today is **`@discordjs/voice`** (audio + music + GoLive-send). The selfbot-lib
+video-receive path lives on the **selfbot-lib `VoiceConnection`** — a separate voice stack. Two options:
+
+### Option A (RECOMMENDED): Parallel selfbot video-watch connection
+Keep `@discordjs/voice` for everything it does today. Add a **second, selfbot-lib voice connection**
+to the same channel whose ONLY job is to watch + record others' video.
+
+- Pros: zero regression risk to audio/music/screenshare-send; uses native `createVideoStream` → mk4.
+- Cons: two voice connections for the same bot user in one channel. Need to verify Discord tolerates it
+  (real selfbots like Discord-RE do exactly this for multi-stream). The selfbot lib's `joinChannel`
+  reuses `ClientVoiceManager.connection` (it's a singleton) — see caveat below.
+
+### Option B: Migrate primary voice to selfbot lib
+Make the selfbot `VoiceConnection` THE voice layer (it also does audio via `receiver.createStream`).
+- Pros: one connection; video+audio unified.
+- Cons: large refactor; high regression risk to the entire existing audio/music/GoLive stack. NOT chosen now.
+
+## CAVEAT — ClientVoiceManager.connection is a singleton
+`ClientVoiceManager.connection` is a single `VoiceConnection`. The gateway's `@discordjs/voice` adapter and
+the selfbot lib both drive the same client voice state. Need to verify whether `client.voice.joinChannel()`
+can coexist with the active `@discordjs/voice` session, or whether we must create the selfbot VoiceConnection
+manually / re-use the existing voice state. This is the #1 technical risk to validate in the spike before
+committing to Option A.
+
+## Implementation plan (Option A)
+
+### 1. Streamer detector (new: `modules/voice-recording/videoRecorder.ts`)
+- Listen to voice state updates (`client.on('voiceStateUpdate')` or the existing voice-state hook).
+- When `voiceState.streaming === true` for a member in the bot's channel → candidate to record.
+- Skip bot's own user id (unless we also want self-video; default skip).
+
+### 2. Watch + record wiring
+- Ensure a selfbot-lib `VoiceConnection` exists for the channel (spike: `client.voice.joinChannel(channel)`,
+  fallback: build a `VoiceConnection` directly from the existing voice auth).
+- `await selfbotVoiceConn.joinStreamConnection(userId)`  → STREAM_WATCH op 20.
+- `const recorder = selfbotVoiceConn.receiver.createVideoStream(userId, outPath)` where outPath points under
+  `<RECORDINGS_DIR>/<uid>/video-<streamKey>-<ts>.mkv` (Recorder outputs MKV natively).
+- On `recorder.on('ready')` → mark recording; `recorder.on('closed')` → finalize.
+- Transcript later: MKV → mp4 via ffmpeg (Phase B `muxToMp4` can accept mkv) for dashboard playback.
+
+### 3. Teardown
+- When `voiceState.streaming === false` / user leaves / channel emptied → `recorder.destroy()`,
+  `selfbotVoiceConn.streamWatchConnection.delete(userId)` / `sendStopScreenshare()`.
+
+### 4. Frontend (Phase UI, later)
+- oRPC/backend list `.mkv` per call session + FE `<video>` player (mirror audio recordings UI).
+
+## Files touched
+- `services/discord-gateway/src/modules/voice-recording/videoRecorder.ts` (new)
+- `services/discord-gateway/src/modules/voice-recording/recorder.ts` (wire streamer detector on voice join)
+- Possibly `voiceController.ts` (voice state update subscription)
+- Tests: `tests/videoRecorder.test.ts` (mock selfbot VoiceConnection + Recorder)
+
+## Verification
+1. `pnpm typecheck` + `pnpm build` + biome clean in discord-gateway.
+2. Unit: Recorder wiring + streamer detection with mocked VoiceConnection.
+3. Live (deploy): user shares screen → journal shows `STREAM_WATCH` sent + `Recorder ready` + `.mkv` file
+   appears under recordings dir; playable via ffmpeg.
+4. CI Build & Deploy (Nix) green.
+
+## Open questions for spike (before full build)
+- [ ] Can `client.voice.joinChannel()` run alongside the active `@discordjs/voice` session, or does the
+      singleton `ClientVoiceManager.connection` collide / tear down the existing audio connection?
+- [ ] Does the selfbot `VoiceConnection` need the bot's `video: true` flag in IDENTIFY to receive video
+      (it advertises `streams` in IDENTIFY — see BaseMediaConnection/identify vs selfbot VoiceConnection)?
+- [ ] Does `Recorder` (spawns system ffmpeg, UDP loopback on 65506/65510) work in the Nix store runtime
+      (ffmpeg-headless on PATH confirmed; UDP loopback fine)?
@@ -0,0 +1,61 @@
+# GMW — Fix mobile navbar "tidak bisa pindah halaman" (root cause)
+
+## Context / symptom
+- User (phone, mobile viewport <md): pressing the bottom mobile nav items does NOT
+  navigate ("Masih sama, walau sudah ditekan tidak pindah halaman").
+- Previous fix (commits ae41f64e, 4d0bbed5) repositioned the chatbot FAB above the
+  nav and expanded nav to all 7 items — but the bug persisted.
+
+## Root cause (verified live 2026-08-28)
+Reproduced on the LIVE site (`imphnen.asepharyana.my.id`) with Playwright @375px:
+1. **Next `<Link>` client-side navigation is dead app-wide.** Clicking a mobile nav
+   `<Link>` fires the click (event reaches document, `inLink=true`,
+   `defaultPrevented=false`) but **the URL never changes**. `window.next.router.push('/voice')`
+   also does nothing. The desktop NavRail navigates only because it uses a **plain
+   `<a href>`** (full browser navigation bypasses the broken router).
+2. **A React hydration failure (#418 "server rendered text didn't match the client")**
+   is thrown on live pages — the only console error. Cause: relative-time text
+   (`formatRelativeTime(e.edited_at)` / `formatRelativeTime(m.created_at)` in the
+   SSR-seeded messages/edit-history feed uses `Date.now()`; server and client
+   render slightly different text → hydration mismatch → React re-renders, and the
+   Next client router ends up non-functional.
+
+## Why desktop "worked" / mobile didn't
+- `NavRail` = plain `<a href>` → **hard** navigation (works).
+- `MobileNav` = Next `<Link>` → **client** navigation → dead router.
+User's instruction "ikuti cara kerja sidebar" = make mobile nav behave like the
+sidebar (plain anchors).
+
+## Changes
+### 1. `src/components/shell/mobile-nav.tsx`
+Replace Next `<Link>` with a **plain `<a href>`** (mirrors NavRail). Keep:
+`usePathname` for active-state styling + `scrollIntoView` for snap-to-active.
+Drop the now-unused `Link` import (biome import-order — run biome check --write).
+
+### 2. Hydration root cause — `formatRelativeTime` in SSR-seeded feeds
+Add `suppressHydrationWarning` to the timestamp elements in the SSR-seeded
+components that render live-relative text so server/client text drift no longer
+throws #418:
+- `src/components/EditHistory.tsx` (line ~111 `edited {formatRelativeTime(...)}`)
+- `src/app/(dashboard)/messages/view.tsx` MessageRow channel/time (line ~580) +
+  semantic row (line ~364)
+- `src/components/LiveModerationFeed.tsx` (line ~139)
+- (recordings/view.tsx, TermGlossary, ChannelCultureGlossary, CategoryDrilldown,
+  chatbot — chatbot is client-after-mount; add where SSR-seeded.)
+
+NOTE: `suppressHydrationWarning` is safe for these single-text spans; the values
+are cosmetic and self-correct on the next interval/render.
+
+## Verification (non-prod, NEVER 4017)
+- `pnpm format` + `pnpm lint` (biome) + `pnpm build` clean.
+- Smoke on :4024 (fresh `next dev`): mobile click on nav item actually navigates
+  (URL changes) AND no React #418 in console.
+- After push: `gh run watch`, then live `imphnen.asepharyana.my.id` @375px: nav tap
+  navigates, console free of #418.
+
+## Files
+- services/frontend/src/components/shell/mobile-nav.tsx
+- services/frontend/src/components/EditHistory.tsx
+- services/frontend/src/app/(dashboard)/messages/view.tsx
+- services/frontend/src/components/LiveModerationFeed.tsx
+- (others only if #418 persists)
@@ -82,6 +82,17 @@
                   [ "$d" = "node_modules" ] && continue; \
                   grep -qF ".pnpm/$d" $TMPDIR/prod-pnms.txt || rm -rf "$d"; \
                 done ) || true
+          # Drop runtime-dead packages that still land in the prod graph:
+          #   - `@types/*` (pure TypeScript declarations) get pulled in as
+          #     REAL dependencies by type-aware deps (discord-api-types ->
+          #     @types/node, pg-protocol -> @types/pg, ...) even though nothing
+          #     ever `require`s them at runtime. Safe to strip.
+          #   - `opusscript` is only a pure-JS fallback Opus engine that
+          #     prism-media's loader uses IF `@discordjs/opus` (native, always
+          #     present/prebuilt) fails to load. Since the native engine loads,
+          #     opusscript is never executed — dead weight pulled in via
+          #     discord.js-selfbot-v13's dependency. Strip it too.
+          ( cd node_modules/.pnpm && rm -rf @types+* opusscript@* 2>/dev/null ) || true
          # Drop symlinks whose .pnpm target was pruned (top-level, scoped dirs,
          # hoist, .bin — any depth). Mirrors stdenv's noBrokenSymlinks check,
          # which would otherwise fail the fixupPhase.
@@ -132,14 +143,17 @@ WRAPPER

          src = ./services/discord-gateway;

+          # cmake + rust/cargo were inherited for node-datachannel-style native
+          # deps — that's 9router/omniroute, NOT GMW. GMW's only native deps
+          # are @discordjs/opus + sharp, both PREBUILT (node-pre-gyp / @img
+          # libvips download), so no cmake or rust toolchain is needed.
+          # python3/gnumake/gcc stay as node-gyp fallback for @discordjs/opus.
          nativeBuildInputs = [
            nodejs pnpm
-            pkgs.python3 pkgs.gnumake pkgs.gcc pkgs.cmake
-            pkgs.rustc pkgs.cargo
+            pkgs.python3 pkgs.gnumake pkgs.gcc
            pkgs.pkg-config
            pkgs.openssl
            pkgs.openssl.dev
-            pkgs.git # for any FetchContent-based deps during native builds
            pkgs.cacert
          ];

@@ -21,21 +21,21 @@
    "drizzle-orm": "^0.45.2",
    "express": "^5.2.1",
    "helmet": "^8.1.0",
-    "ioredis": "^5.11.0",
-    "pg": "^8.21.0",
-    "pino": "^9.6.0",
+    "ioredis": "^6.0.0",
+    "pg": "^8.23.0",
+    "pino": "^10.3.1",
    "prom-client": "^15.1.3",
-    "ws": "^8.20.1",
+    "ws": "^8.21.3",
    "zod": "^4.4.3"
  },
  "devDependencies": {
    "@biomejs/biome": "latest",
    "@types/express": "^5.0.6",
-    "@types/node": "^25.9.0",
-    "@types/pg": "^8.20.0",
+    "@types/node": "^26.3.0",
+    "@types/pg": "^8.23.1",
    "@types/ws": "^8.18.1",
-    "tsx": "^4.22.2",
-    "typescript": "^5.9.3",
+    "tsx": "^4.23.12",
+    "typescript": "^7.0.2",
    "vitest": "latest"
  }
 }
@@ -158,7 +158,7 @@ Gaya ngobrol:
            tool_choice: "auto",
            max_tokens: 600,
            temperature: 0.4,
-            // Non-streaming: request a single complete response. 9router may
+            // Non-streaming: request a single complete response. omniroute may
            // still emit SSE even with stream:false, so the parser below
            // handle both raw-JSON and SSE bodies.
            stream: false,
@@ -176,7 +176,7 @@ Gaya ngobrol:
          },
        );

-        // Parse the body into content + tool_calls. 9router may return either
+        // Parse the body into content + tool_calls. omniroute may return either
        // a single JSON object (stream:false honored) or SSE text (stream
        // implied) — parseResponse handles both.
        const { content, toolCalls } = this.parseResponse(
@@ -252,7 +252,7 @@ Gaya ngobrol:

  /**
   * Parse an LLM HTTP body into content + tool_calls. Handles both shapes
-   * 9router can return: a single JSON object (stream:false honored) or SSE
+   * omniroute can return: a single JSON object (stream:false honored) or SSE
   * text (stream implied). For SSE we delegate to parseSse.
   */
  private parseResponse(body: string): {
@@ -306,7 +306,7 @@ Gaya ngobrol:

  /**
   * Parse an SSE stream body into accumulated content + any tool_calls.
-   * 9router (and most OpenAI-compatible routers) emit `data: {json}` lines
+   * omniroute (and most OpenAI-compatible routers) emit `data: {json}` lines
   * even when stream is only implied; we must collect deltas manually.
   */
  private parseSse(body: string): {
@@ -469,17 +469,20 @@ export class MessagesRepository {
    const db = getDatabase();
    const since = Date.now() - days * 24 * 60 * 60 * 1000;
    const result = await db.execute(sql`
-      SELECT channel_id,
-             EXTRACT(HOUR FROM to_timestamp(created_at / 1000))::int AS hour,
-             COUNT(*)::int AS c
-      FROM messages
-      WHERE created_at >= ${since}
-      GROUP BY channel_id, hour
-      ORDER BY channel_id, hour
+      SELECT
+        m.channel_id,
+        COALESCE(NULLIF((m.metadata::jsonb -> 'channel' ->> 'channelName'), ''), m.channel_id) AS channel_name,
+        EXTRACT(HOUR FROM to_timestamp(m.created_at / 1000))::int AS hour,
+        COUNT(*)::int AS c
+      FROM messages m
+      WHERE m.created_at >= ${since}
+      GROUP BY m.channel_id, channel_name, hour
+      ORDER BY channel_name, hour
    `);
    const rows = (result.rows as Record<string, unknown>[]) || [];
    return rows.map((r) => ({
      channelId: String(r.channel_id ?? "unknown"),
+      channelName: String(r.channel_name ?? r.channel_id ?? "unknown"),
      hour: Number(r.hour ?? 0),
      count: Number(r.c ?? 0),
    }));
@@ -491,31 +494,28 @@ export class MessagesRepository {
   */
  async getRecentEdits(limit = 50, channelId?: string) {
    const db = getDatabase();
-    const where = channelId
-      ? `WHERE m.channel_id = '${channelId.replace(/'/g, "''")}'`
-      : "";
-    const result = await db.execute(
-      sql.raw(`
-        SELECT
-          e.id,
-          e.message_id,
-          e.old_content,
-          e.edited_at,
-          m.channel_id,
-          COALESCE(NULLIF((m.metadata::jsonb -> 'channel' ->> 'channelName'), ''), m.channel_id) AS channel_name,
-          m.username
-        FROM message_edits e
-        JOIN messages m ON m.id = e.message_id
-        ${where}
-        ORDER BY e.edited_at DESC
-        LIMIT ${limit}
-      `),
-    );
+    const result = await db.execute(sql`
+      SELECT
+        e.id,
+        e.message_id,
+        e.old_content,
+        e.edited_at,
+        m.channel_id,
+        COALESCE(NULLIF((m.metadata::jsonb -> 'channel' ->> 'channelName'), ''), m.channel_id) AS channel_name,
+        m.username,
+        COALESCE(m.edited_content, m.content) AS new_content
+      FROM message_edits e
+      JOIN messages m ON m.id = e.message_id
+      ${channelId ? sql`WHERE m.channel_id = ${channelId}` : sql``}
+      ORDER BY e.edited_at DESC
+      LIMIT ${limit}
+    `);
    const rows = (result.rows as Record<string, unknown>[]) || [];
    return rows.map((r) => ({
      id: String(r.id),
      message_id: String(r.message_id),
      old_content: r.old_content ? String(r.old_content) : "",
+      new_content: r.new_content ? String(r.new_content) : "",
      edited_at: r.edited_at ? Number(r.edited_at) : 0,
      channel_id: r.channel_id ? String(r.channel_id) : null,
      channel_name: r.channel_name ? String(r.channel_name) : null,
@@ -39,12 +39,15 @@ export class MessagesService {
      throw new ValidationError("message ID is required");
    }

-    const message = await messagesRepository.findById(id);
+    const [message, editHistory] = await Promise.all([
+      messagesRepository.findById(id),
+      messagesRepository.getEditHistory(id),
+    ]);
+
    if (!message) {
      throw new NotFoundError(`Message with ID ${id} not found`);
    }

-    const editHistory = await messagesRepository.getEditHistory(id);
    return {
      ...message,
      edit_count: editHistory.length,
@@ -124,7 +124,8 @@ export class ModerationRepository {
        a.score,
        a.evidence,
        a.policy_version,
-        m.username,
+        a.username,
+        a.server_nick,
        LEFT(m.content, 300) AS content
      FROM moderation_actions a
      LEFT JOIN messages m ON m.id = a.message_id
@@ -155,6 +156,7 @@ export class ModerationRepository {
      evidence: parseJsonArray(r.evidence),
      policy_version: r.policy_version ? String(r.policy_version) : null,
      username: r.username ? String(r.username) : null,
+      server_nick: r.server_nick ? String(r.server_nick) : null,
      content: r.content ? String(r.content) : null,
    }));

@@ -314,7 +316,7 @@ export class ModerationRepository {
      SELECT
        a.id, a.message_id, a.user_id, a.guild_id, a.action_type,
        a.reason, a.status, a.created_at, a.severity, a.confidence, a.score,
-        m.username, LEFT(m.content, 300) AS content
+        a.username, LEFT(m.content, 300) AS content
      FROM moderation_actions a
      LEFT JOIN messages m ON m.id = a.message_id
      WHERE a.created_at >= ${since}
@@ -338,6 +340,7 @@ export class ModerationRepository {
      confidence: r.confidence != null ? Number(r.confidence) : null,
      score: r.score != null ? Number(r.score) : null,
      username: r.username ? String(r.username) : null,
+      server_nick: r.server_nick ? String(r.server_nick) : null,
      content: r.content ? String(r.content) : null,
    }));
  }
@@ -1,4 +1,15 @@
-import { and, desc, eq, lt, type SQL } from "drizzle-orm";
+import {
+  and,
+  desc,
+  eq,
+  gte,
+  ilike,
+  lt,
+  lte,
+  or,
+  type SQL,
+  sql,
+} from "drizzle-orm";
 import { getDatabase } from "../../shared/database/index.js";
 import { pgVoiceRecordingsTable } from "../../shared/index.js";
 import { createChildLogger } from "../../shared/logger/index.js";
@@ -20,6 +31,7 @@ export interface RecordingRow {
  upload_error: string | null;
  created_at: number;
  uploaded_at: number | null;
+  transcription: string | null;
 }

 export interface PaginatedRecordings {
@@ -28,27 +40,67 @@ export interface PaginatedRecordings {
  hasMore: boolean;
 }

+export interface RecordingFilters {
+  channelId?: string;
+  userId?: string;
+  cursor?: string;
+  /** keyword search against transcription + username (ILIKE) */
+  q?: string;
+  /** created_at lower bound (epoch ms) */
+  startDate?: number;
+  /** created_at upper bound (epoch ms) */
+  endDate?: number;
+}
+
+export interface SpeakerSummary {
+  user_id: string;
+  username: string;
+  avatar_url: string | null;
+  clips: number;
+  /** summed MP3 bytes; ~128kbps → est duration in seconds */
+  est_duration_s: number;
+  words: number;
+  transcribed: number;
+  last_at: number;
+}
+
 export class RecordingsService {
  async getRecent(
    limit = 50,
-    filters?: { channelId?: string; userId?: string; cursor?: string },
+    filters: RecordingFilters = {},
  ): Promise<PaginatedRecordings> {
-    logger.info({ limit }, "getRecent called");
+    logger.info({ limit, filters }, "getRecent called");
    const db = getDatabase();

    const conditions: SQL[] = [];

-    if (filters?.cursor) {
+    if (filters.cursor) {
      conditions.push(
        lt(pgVoiceRecordingsTable.created_at, Number(filters.cursor)),
      );
    }
-    if (filters?.channelId) {
+    if (filters.channelId) {
      conditions.push(eq(pgVoiceRecordingsTable.channel_id, filters.channelId));
    }
-    if (filters?.userId) {
+    if (filters.userId) {
      conditions.push(eq(pgVoiceRecordingsTable.user_id, filters.userId));
    }
+    if (filters.q) {
+      const like = `%${filters.q}%`;
+      const qOr = or(
+        ilike(pgVoiceRecordingsTable.transcription, like),
+        ilike(pgVoiceRecordingsTable.username, like),
+      );
+      if (qOr) conditions.push(qOr);
+    }
+    if (filters.startDate) {
+      conditions.push(
+        gte(pgVoiceRecordingsTable.created_at, filters.startDate),
+      );
+    }
+    if (filters.endDate) {
+      conditions.push(lte(pgVoiceRecordingsTable.created_at, filters.endDate));
+    }

    const where = conditions.length > 0 ? and(...conditions) : undefined;

@@ -68,6 +120,7 @@ export class RecordingsService {
        upload_error: pgVoiceRecordingsTable.upload_error,
        created_at: pgVoiceRecordingsTable.created_at,
        uploaded_at: pgVoiceRecordingsTable.uploaded_at,
+        transcription: pgVoiceRecordingsTable.transcription,
      })
      .from(pgVoiceRecordingsTable)
      .where(where)
@@ -89,6 +142,43 @@ export class RecordingsService {
      .delete(pgVoiceRecordingsTable)
      .where(eq(pgVoiceRecordingsTable.id, id));
  }
+
+  /**
+   * Speaker leaderboard: aggregate clips / est. duration / transcribed words
+   * per user. `est_duration_s` derives from summed MP3 bytes at the uploader's
+   * 128 kbps transcode rate; `words` sums transcription word counts.
+   */
+  async getSummary(): Promise<SpeakerSummary[]> {
+    const db = getDatabase();
+    const t = pgVoiceRecordingsTable;
+
+    const rows = await db
+      .select({
+        user_id: t.user_id,
+        username: t.username,
+        avatar_url: t.avatar_url,
+        clips: sql<number>`count(*)::int`,
+        total_bytes: sql<number>`coalesce(sum(${t.size_bytes}), 0)`,
+        words: sql<number>`coalesce(sum(array_length(string_to_array(${t.transcription}, ' '), 1)), 0)::int`,
+        transcribed: sql<number>`count(${t.transcription})::int`,
+        last_at: sql<number>`max(${t.created_at})`,
+      })
+      .from(t)
+      .groupBy(t.user_id, t.username, t.avatar_url)
+      .orderBy(sql`clips desc, last_at desc`);
+
+    return rows.map((r) => ({
+      user_id: r.user_id,
+      username: r.username,
+      avatar_url: r.avatar_url,
+      clips: r.clips,
+      // 128 kbps = 128000 bps → bytes*8/128000 = seconds
+      est_duration_s: Math.round((r.total_bytes * 8) / 128000),
+      words: r.words,
+      transcribed: r.transcribed,
+      last_at: r.last_at,
+    }));
+  }
 }

 export const recordingsService = new RecordingsService();
@@ -23,14 +23,26 @@ const speakers = new Map<string, LiveSpeaker>();
 const MAX_SPEAKERS = 200;

 /**
- * Record a voice_active_user event. `speaking: true` upserts the speaker as
- * active; `speaking: false` marks them inactive while keeping them for the
- * activity timeline.
+ * Speakers inactive for longer than this are auto-expired from the snapshot.
+ * This handles the case where the gateway disconnects abruptly and never
+ * sends `speaking: false` for active users.
 */
+const SPEAKER_TTL_MS = 30_000;
+
+/** Purge speakers that haven't been active recently. */
+function purgeStale(): void {
+  const cutoff = Date.now() - SPEAKER_TTL_MS;
+  for (const [id, s] of speakers) {
+    if (!s.speaking && s.lastActiveAt < cutoff) {
+      speakers.delete(id);
+    }
+  }
+}
+
 /**
- * recordSpeaker(data) — apply a `voice_active_user` event. `speaking: true`
- * upserts the speaker as ACTIVE; `speaking: false` marks them inactive while
- * keeping them for the activity timeline.
+ * Record a voice_active_user event. `speaking: true` upserts the speaker as
+ * active; `speaking: false` marks them inactive while keeping them briefly
+ * for the activity timeline (until TTL expiry).
 */
 export function recordSpeaker(data: {
  userId: string;
@@ -49,7 +61,6 @@ export function recordSpeaker(data: {
  };

  if (speakers.size >= MAX_SPEAKERS && !existing) {
-    // Drop the least-recently-active non-speaking speaker to stay bounded.
    let oldestId: string | null = null;
    let oldestTs = Infinity;
    for (const [id, s] of speakers) {
@@ -65,18 +76,35 @@ export function recordSpeaker(data: {
  speakers.set(userId, speaker);
 }

-/** All known speakers, most recently active first. */
+/**
+ * All recently-active speakers, most recently active first.
+ * Stale (non-speaking + old) entries are auto-purged.
+ */
 export function getActiveSpeakers(): LiveSpeaker[] {
+  purgeStale();
  return [...speakers.values()].sort((a, b) => b.lastActiveAt - a.lastActiveAt);
 }

 /** Only speakers currently flagged as speaking. */
 export function getSpeakingSpeakers(): LiveSpeaker[] {
+  purgeStale();
  return [...speakers.values()]
    .filter((s) => s.speaking)
    .sort((a, b) => b.lastActiveAt - a.lastActiveAt);
 }

+/**
+ * Mark ALL tracked speakers as not-speaking and purge stale ones.
+ * Called when the gateway disconnects from voice — ensures the authoritative
+ * snapshot doesn't carry ghost speakers.
+ */
+export function clearAllSpeakers(): void {
+  for (const [id, s] of speakers) {
+    s.speaking = false;
+  }
+  purgeStale();
+}
+
 /** Drop all tracked speakers (used on backend restart). */
 export function resetLiveSpeakers(): void {
  speakers.clear();
@@ -296,6 +296,9 @@ const recordingsRouter = {
        channelId: z.string().optional(),
        userId: z.string().optional(),
        cursor: z.string().optional(),
+        q: z.string().optional(),
+        startDate: z.coerce.number().int().optional(),
+        endDate: z.coerce.number().int().optional(),
      }),
    )
    .handler(({ input }) =>
@@ -303,12 +306,16 @@ const recordingsRouter = {
        channelId: input.channelId,
        userId: input.userId,
        cursor: input.cursor,
+        q: input.q,
+        startDate: input.startDate,
+        endDate: input.endDate,
      }),
    ),
  delete: os.input(z.object({ id: z.string() })).handler(async ({ input }) => {
    await recordingsService.deleteById(input.id);
    return { ok: true };
  }),
+  summary: os.handler(async () => recordingsService.getSummary()),
 };

 // ── Analysis (search) ──────────────────────────────────────────────
@@ -131,7 +131,7 @@ export const configSchema = z
    AI_LLM_BASE_URL: z
      .string()
      .url()
-      .default("https://9router.asepharyana.my.id/v1"),
+      .default("http://100.121.180.82:20128/api/v1"),
    AI_LLM_MODEL: z.string().default("text"),
    AI_LLM_VISION_MODEL: z.string().optional(),
    AI_LLM_EMBEDDING_MODEL: z.string().optional(),
@@ -1,5 +1,22 @@
 // Shared message row mapper for backend repository modules

+function resolveServerNick(row: Record<string, unknown>): string | null {
+  const metadata = row.metadata;
+  if (!metadata || typeof metadata !== "string") {
+    return (row.username as string | null) ?? null;
+  }
+  try {
+    const parsed = JSON.parse(metadata) as {
+      member?: { displayName?: string | null } | null;
+    };
+    return (
+      parsed?.member?.displayName ?? (row.username as string | null) ?? null
+    );
+  } catch {
+    return (row.username as string | null) ?? null;
+  }
+}
+
 export interface MappedMessage {
  id: string;
  guild_id: string;
@@ -7,6 +24,8 @@ export interface MappedMessage {
  thread_id: string | null;
  user_id: string;
  username: string;
+  /** Member's server-specific display name (nickname), from metadata.member.displayName. Falls back to username when absent. */
+  server_nick: string | null;
  avatar_url: string | null;
  content: string;
  edited_content: string | null;
@@ -42,6 +61,7 @@ export function mapMessageRow(row: Record<string, unknown>): MappedMessage {
    thread_id: (row.thread_id as string | null) ?? null,
    user_id: String(row.user_id ?? ""),
    username: String(row.username ?? ""),
+    server_nick: resolveServerNick(row),
    avatar_url: (row.avatar_url as string | null) ?? null,
    content: String(row.content ?? ""),
    edited_content: (row.edited_content as string | null) ?? null,
@@ -1,10 +1,14 @@
 import Redis from "ioredis";
-import { recordSpeaker } from "../modules/voice/live-speaker.js";
+import {
+  clearAllSpeakers,
+  recordSpeaker,
+} from "../modules/voice/live-speaker.js";
 import { config } from "../shared/config/index.js";
 import {
  DISCORD_CHANNEL_TO_WS_EVENT,
  DISCORD_VOICE_ACTIVE_USER,
  DISCORD_VOICE_PCM,
+  DISCORD_VOICE_STOPPED,
 } from "../shared/index.js";
 import { createChildLogger } from "../shared/logger/index.js";
 import { broadcastBinary, broadcastEvent } from "./broadcast.js";
@@ -84,6 +88,14 @@ function handleSubscriptionMessage(channel: string, message: string): void {
    }
  }

+  // When the gateway stops voice recording (disconnects from voice channel),
+  // clear all speakers from the authoritative snapshot so frontends don't
+  // show ghost participants.
+  if (channel === DISCORD_VOICE_STOPPED) {
+    clearAllSpeakers();
+    logger.info("Voice recording stopped — cleared all live speakers");
+  }
+
  logger.debug({ channel, eventType }, "Broadcasting Redis event");
  broadcastEvent(eventType, data);
 }
@@ -139,5 +139,5 @@ pipeline gauges — `ai_analysis_queued_conversations`,
  numeric snowflake IDs never trigger false positives.
 - **Semantic cache is batched** (one embed call + one Qdrant batch search),
  not N sequential round-trips. `ensureQdrantCollection` is memoized.
- **Streaming is mandatory** against the 9router base URL (non-stream waits for
+- **Streaming is mandatory** against the omniroute base URL (non-stream waits for
  the full body and times out). `llmClient` aggregates SSE chunks.
@@ -0,0 +1,9 @@
+ALTER TABLE "moderation_actions" ADD COLUMN IF NOT EXISTS "username" text;
+--> statement-breakpoint
+-- Backfill username from messages table for existing actions
+UPDATE "moderation_actions" a
+  SET "username" = m."username"
+  FROM "messages" m
+  WHERE a."message_id" = m."id"
+    AND a."username" IS NULL
+    AND m."username" IS NOT NULL;
@@ -0,0 +1 @@
+ALTER TABLE "moderation_actions" ADD COLUMN IF NOT EXISTS "server_name" text;
@@ -0,0 +1,26 @@
+-- Rename server_name (guild name) -> server_nick (member's server nickname)
+-- server_name was added in 0018 but never populated (all NULL), so renaming
+-- the empty column is safe and keeps retention semantics correct.
+-- Idempotent across all possible prior states:
+--   • server_name present, server_nick absent  -> rename
+--   • both present (backfill created server_nick) -> drop the stale server_name
+--   • server_nick present only                -> already done, no-op
+DO $$
+BEGIN
+  IF EXISTS (
+    SELECT 1 FROM information_schema.columns
+    WHERE table_name = 'moderation_actions' AND column_name = 'server_name'
+  ) AND NOT EXISTS (
+    SELECT 1 FROM information_schema.columns
+    WHERE table_name = 'moderation_actions' AND column_name = 'server_nick'
+  ) THEN
+    ALTER TABLE "moderation_actions" RENAME COLUMN "server_name" TO "server_nick";
+  ELSIF EXISTS (
+    SELECT 1 FROM information_schema.columns
+    WHERE table_name = 'moderation_actions' AND column_name = 'server_name'
+  ) THEN
+    -- server_nick already exists (e.g. backfilled out-of-band) — drop the
+    -- now-redundant server_name column instead of renaming into a conflict.
+    ALTER TABLE "moderation_actions" DROP COLUMN "server_name";
+  END IF;
+END $$;
@@ -0,0 +1,19 @@
+-- Voice auto-reconnect: persist desired voice state per guild.
+-- The gateway rejoins the same channel after restart/reboot and after an
+-- unexpected voice drop (kick / server move / voice server restart).
+-- Idempotent: skips if the table already exists.
+DO $$
+BEGIN
+  IF NOT EXISTS (
+    SELECT 1 FROM information_schema.tables
+    WHERE table_name = 'voice_auto_reconnect'
+  ) THEN
+    CREATE TABLE "voice_auto_reconnect" (
+      "guild_id" text PRIMARY KEY NOT NULL,
+      "channel_id" text NOT NULL,
+      "channel_name" text,
+      "connected_at" bigint NOT NULL,
+      "updated_at" bigint NOT NULL
+    );
+  END IF;
+END $$;
@@ -120,6 +120,34 @@
      "when": 1787185000000,
      "tag": "0016_drop_user_reputations",
      "breakpoints": true
+    },
+    {
+      "idx": 17,
+      "version": "7",
+      "when": 1787810000000,
+      "tag": "0017_add_username_to_moderation_actions",
+      "breakpoints": true
+    },
+    {
+      "idx": 18,
+      "version": "7",
+      "when": 1787820000000,
+      "tag": "0018_add_server_name_to_moderation_actions",
+      "breakpoints": true
+    },
+    {
+      "idx": 19,
+      "version": "7",
+      "when": 1787830000000,
+      "tag": "0019_rename_server_name_to_server_nick",
+      "breakpoints": true
+    },
+    {
+      "idx": 20,
+      "version": "7",
+      "when": 1787860000000,
+      "tag": "0020_add_voice_auto_reconnect",
+      "breakpoints": true
    }
  ]
 }
@@ -28,30 +28,30 @@
    "discord.js-selfbot-v13": "^3.7.1",
    "dotenv": "^17.4.2",
    "drizzle-orm": "^0.45.2",
-    "ioredis": "^5.11.0",
+    "ioredis": "^6.0.0",
    "libsodium-wrappers": "^0.8.4",
    "lru-cache": "^11.5.1",
-    "openai": "^6.38.0",
-    "opusscript": "^0.0.8",
+    "openai": "^7.5.0",
+    "opusscript": "^0.1.1",
    "p-limit": "^7.3.0",
    "p-retry": "^8.0.0",
-    "pg": "^8.21.0",
-    "pino": "^9.6.0",
-    "piscina": "^5.1.4",
+    "pg": "^8.23.0",
+    "pino": "^10.3.1",
+    "piscina": "^5.3.1",
    "prism-media": "2.0.0-alpha.0",
-    "sharp": "^0.34.5",
+    "sharp": "^0.35.3",
    "tiktoken": "^1.0.22",
-    "ws": "^8.20.1",
+    "ws": "^8.21.3",
    "zod": "^4.4.3"
  },
  "devDependencies": {
    "@biomejs/biome": "latest",
-    "@types/node": "^25.9.0",
-    "@types/pg": "^8.20.0",
+    "@types/node": "^26.3.0",
+    "@types/pg": "^8.23.1",
    "@types/ws": "^8.18.1",
    "drizzle-kit": "^0.31.10",
-    "tsx": "^4.22.2",
-    "typescript": "^5.9.3",
+    "tsx": "^4.23.12",
+    "typescript": "^7.0.2",
    "vitest": "latest"
  }
 }
@@ -36,6 +36,10 @@ import {
  setPcmWsClient,
  setEventBroadcaster as setRecorderEventBroadcaster,
 } from "../modules/voice-recording/recorder.js";
+import {
+  setVideoRecorderClient,
+  setVideoRecordingsDir,
+} from "../modules/voice-recording/videoRecorder.js";
 import { VoiceController } from "../modules/voice-recording/voiceController.js";
 import { config } from "../shared/config/config.js";
 import {
@@ -182,6 +186,12 @@ export async function initializeDiscordGateway() {
  const client = new Client(createDiscordClientOptions());
  const voiceController = new VoiceController(client);

+  // Wire the video recorder (others' camera/screen share) to the selfbot's
+  // native watch/receive stack + its recordings dir. Best-effort: failures are
+  // logged inside, never fatal.
+  setVideoRecorderClient(client);
+  setVideoRecordingsDir(config.RECORDINGS_DIR);
+
  // Initialize Redis event broadcaster
  const redisPublisher = new RedisEventPublisher(config.REDIS_URL, logger);
  const eventBroadcaster = new EventBroadcaster(redisPublisher);
@@ -274,6 +284,17 @@ export async function initializeDiscordGateway() {
    commandHandler.start(client, voiceController);
    logger.info("Command handler started");

+    // Rejoin persisted voice channels (auto-reconnect on restart/reboot).
+    // Non-fatal: failures are logged inside autoReconnect.
+    void voiceController
+      .autoReconnect()
+      .catch((err) =>
+        logger.warn(
+          { err: err instanceof Error ? err.message : String(err) },
+          "Voice auto-reconnect on startup failed",
+        ),
+      );
+
    // Start retention cleanup scheduler
    startRetentionCleanup();
    // Start weekly moderation digest (public, automated)
@@ -93,12 +93,6 @@ type BatchOkResponse = {
  ok: true;
  conversationKey: string;
  rows: MessageRecord[];
-  /**
-   * Race-guard signal (2026-08-25): target ids whose attachment upload is
-   * still in-flight — NO analysis ran for them. The processor must defer
-   * these (requeue + poll), never fan them out as failures.
-   */
-  uploadPendingIds?: string[];
 };
 type BatchErrorResponse = {
  ok: false;
@@ -109,12 +103,6 @@ type BatchErrorResponse = {
 type IndividualOkResponse = {
  ok: true;
  results: AnalysisResult[];
-  /**
-   * Race-guard signal (2026-08-24): the message's attachment upload is still
-   * in-flight — NO analysis ran. The processor must re-queue the message as
-   * `pending` and re-schedule, never treat this as a completed moderation.
-   */
-  uploadPending?: boolean;
 };
 type IndividualErrorResponse = {
  ok: false;
@@ -309,40 +297,9 @@ async function processBatch(job: {
    ...contextIds,
  ]);

-  // Attachment-upload race guard: a message whose attachment is still being
-  // uploaded (upload_status='pending') must not be analyzed yet. Its
-  // uploaded_url is not ready, and falling back to the Discord CDN link often
-  // 404s (expired/purged) — which used to silently produce a text-only
-  // verdict ("lampiran yang gagal terbaca"). Leave those targets pending; the
-  // next worker cycle picks them up after the upload lands.
-  const pendingUploadTargetIds = new Set(
-    (attachments ?? [])
-      .filter((a) => a.upload_status === "pending")
-      .map((a) => a.message_id),
-  );
-  const readyMessages =
-    pendingUploadTargetIds.size === 0
-      ? messages
-      : messages.filter((m) => !pendingUploadTargetIds.has(m.id));
-  if (readyMessages.length === 0) {
-    // Explicit signal (2026-08-25): every target is still upload-pending.
-    // Returning bare {ok:true, rows:[]} made the processor classify all of
-    // them "incomplete" and fan out to the individual queue — a hot ~300ms
-    // requeue loop for the whole upload duration.
-    return {
-      ok: true,
-      conversationKey,
-      rows: [],
-      uploadPendingIds: messages.map((m) => m.id),
-    };
-  }
-
-  // The orchestrator handles text/media split + caching + parallel paths
-  // internally, so a 20-message batch = 1 text LLM call (+1 media call
-  // when media is present), not N per-message calls.
  const analysisStart = Date.now();
  const moderationResult = await runModerationAnalysis({
-    targets: readyMessages,
+    targets: messages,
    contextBlock,
    attachments,
  });
@@ -351,7 +308,7 @@ async function processBatch(job: {
  const results = moderationResult.results.map((r) =>
    normalizeResult(
      r as unknown as AnalysisResult,
-      readyMessages.find((m) => m.id === r.messageId),
+      messages.find((m) => m.id === r.messageId),
    ),
  );

@@ -380,10 +337,9 @@ async function processBatch(job: {

  logger.info(
    {
-      total: readyMessages.length,
+      total: messages.length,
      saved: allRows.length,
      conversationKey,
-      skippedPendingUpload: messages.length - readyMessages.length,
    },
    "LLM batch analysis complete",
  );
@@ -431,16 +387,9 @@ async function processIndividual(job: {
    ...contextIds,
  ]);

-  // Same attachment-upload race guard as the batch path: while the upload is
-  // still in-flight the uploaded_url is not ready and the Discord CDN fallback
-  // often 404s — analyzing now would silently produce a text-only verdict.
-  // Return no results so the message stays pending for the next cycle.
-  const uploadStillPending = (attachments ?? []).some(
-    (a) => a.message_id === message.id && a.upload_status === "pending",
-  );
-  if (uploadStillPending) {
-    return { ok: true, results: [], uploadPending: true };
-  }
+  // Analysis uses the Discord CDN URL directly (archive-only uploader).
+  // No upload-pending race guard: analysis proceeds regardless of upload
+  // status, since discord_url is available immediately at capture time.

  try {
    const moderationResult = await runModerationAnalysis({
@@ -163,12 +163,18 @@ export function startPendingAIAnalysisWorker(
        });
    }

-    messageStore.revertStuckProcessingMessages(300000).catch((err: unknown) => {
-      logger.error(
-        { error: String(err) },
-        "Failed to run stuck processing recovery",
-      );
-    });
+    // Only revert stuck processing messages if there's active processing.
+    // Avoids a DB query every recovery interval when the pipeline is idle.
+    if (conversationProcessing.size > 0) {
+      messageStore
+        .revertStuckProcessingMessages(300000)
+        .catch((err: unknown) => {
+          logger.error(
+            { error: String(err) },
+            "Failed to run stuck processing recovery",
+          );
+        });
+    }

    Promise.all([
      messageStore.getPendingConversationKeys(500),
@@ -39,7 +39,7 @@ export function deriveRecommendedAction(msg: MessageRecord): string {
  const severity = deriveSeverity(msg);
  if (
    msg.ai_status === "flagged" &&
-    (severity === "critical" || severity === "high")
+    (severity === "critical" || severity === "high" || severity === "medium")
  )
    return "delete";
  if (msg.ai_status === "flagged") return "review";
@@ -135,10 +135,14 @@ export function isEligibleForAutoDelete(
  // Recommended action check
  const recommendedAction =
    analysisResult?.recommendedAction ?? deriveRecommendedAction(message);
-  if (recommendedAction !== "delete" && recommendedAction !== "escalate") {
+  if (
+    recommendedAction !== "delete" &&
+    recommendedAction !== "escalate" &&
+    recommendedAction !== "warn"
+  ) {
    logger.debug(
      { messageId: message.id, recommendedAction },
-      "Message not eligible for auto-delete: recommended action is not delete/escalate",
+      "Message not eligible for auto-delete: recommended action is not delete/escalate/warn",
    );
    return false;
  }
@@ -2,6 +2,7 @@ import type { Client, PermissionString } from "discord.js-selfbot-v13";
 import { LRUCache } from "lru-cache";
 import { createChildLogger } from "@/shared/logger/index";
 import { config } from "../../shared/config/config.js";
+import { parseRichMessageMetadata } from "../message-capture/messageMetadata.js";
 import { messageStore } from "../message-capture/messageStore.js";
 import type { MessageRecord } from "../message-capture/types.js";
 import {
@@ -10,6 +11,7 @@ import {
 } from "./autoDeleteEligibility.js";
 import { logDeletionToChannel } from "./autoDeleteLogger.js";
 import { sendDeletionNotification } from "./autoDeleteNotify.js";
+import { llmChat } from "./llmClient.js";
 import { verdictToActionFields } from "./verdictToActionFields.js";

 const logger = createChildLogger("auto-delete-manager");
@@ -94,6 +96,26 @@ export async function resetOffensiveNickname(
    }
    // setNickname(null) = remove nickname → Discord shows global username
    await member.setNickname(null, "[auto] nickname melanggar aturan server");
+
+    // ── Check if the global username is ALSO offensive ──────────────
+    // After reset, the member shows their global username. If that is
+    // also offensive (gambling/scam keywords), replace it with a random
+    // generated username so the user can't circumvent the filter by
+    // having an offensive global username.
+    const refreshedMember = await guild.members.fetch(userId);
+    const globalUsername = refreshedMember.user?.username ?? "";
+    if (await isGlobalUsernameOffensiveAI(globalUsername)) {
+      const randomName = generateRandomUsername();
+      await refreshedMember.setNickname(
+        randomName,
+        "[auto] global username juga melanggar, diganti random",
+      );
+      logger.info(
+        { messageId, guildId, userId, globalUsername, randomName },
+        "Global username also offensive — replaced with random username",
+      );
+    }
+
    recentNicknameResets.set(cooldownKey, Date.now());
    logger.info(
      { messageId, guildId, userId },
@@ -119,6 +141,69 @@ export async function resetOffensiveNickname(
  }
 }

+// ─── AI-Based Global Username Check ──────────────────────────────────
+
+/**
+ * Ask the AI moderation LLM whether a global username is offensive
+ * (gambling, scam, spam, NSFW, etc.). Returns `true` if the LLM
+ * judges the username as violating server rules.
+ *
+ * Fail-open: if the LLM call fails or times out, returns `false`
+ * so the nickname reset still completes — we don't want a broken LLM
+ * to block enforcement.
+ */
+async function isGlobalUsernameOffensiveAI(username: string): Promise<boolean> {
+  try {
+    const completion = await llmChat({
+      messages: [
+        {
+          role: "system",
+          content:
+            'Kamu adalah moderator Discord. Tentukan apakah username berikut melanggar aturan server (judi, togel, scam, spam, NSFW, SARA, atau ofensif). Jawab HANYA dengan JSON: {"offensive": true} atau {"offensive": false}. Jangan penjelasan tambahan.',
+        },
+        {
+          role: "user",
+          content: `Username: "${username}"`,
+        },
+      ],
+      max_tokens: 50,
+      temperature: 0,
+      stream: false,
+      timeout: 10_000,
+    });
+
+    const content = completion?.choices?.[0]?.message?.content?.trim() ?? "";
+    // Parse JSON response — handle both strict JSON and markdown-wrapped
+    const jsonMatch = content.match(
+      /\{[^}]*"offensive"\s*:\s*(true|false)[^}]*\}/,
+    );
+    if (jsonMatch) {
+      const parsed = JSON.parse(jsonMatch[0]) as { offensive: boolean };
+      return parsed.offensive === true;
+    }
+    // Fallback: if response contains "true" anywhere, treat as offensive
+    return content.toLowerCase().includes("true");
+  } catch (error) {
+    logger.warn(
+      {
+        username,
+        error: error instanceof Error ? error.message : String(error),
+      },
+      "AI username check failed — failing open (not offensive)",
+    );
+    return false;
+  }
+}
+
+/**
+ * Generate a random fallback username: "User" + 5-digit number.
+ * Guaranteed unique per call (random suffix).
+ */
+function generateRandomUsername(): string {
+  const suffix = Math.floor(10000 + Math.random() * 90000);
+  return `User${suffix}`;
+}
+
 // ─── Error Handling Utilities ────────────────────────────────────────

 function getErrorCode(error: unknown): number | string | undefined {
@@ -130,7 +215,28 @@ function getErrorCode(error: unknown): number | string | undefined {

 function isAlreadyDeletedError(error: unknown): boolean {
  const code = getErrorCode(error);
-  return code === 10008 || code === 404 || code === "10008" || code === "404";
+  // Discord REST error codes for "message not found":
+  //   10008 = Unknown Message, 10003 = Unknown Channel,
+  //   50001 = Missing Access (channel deleted/hidden), 404 = HTTP
+  if (
+    code === 10008 ||
+    code === 10003 ||
+    code === 50001 ||
+    code === 404 ||
+    code === "10008" ||
+    code === "10003" ||
+    code === "50001" ||
+    code === "404"
+  )
+    return true;
+  // Fallback: check the message text for the Discord "Unknown Message" string
+  const msg =
+    error instanceof Error
+      ? error.message
+      : typeof error === "string"
+        ? error
+        : "";
+  return msg.includes("Unknown Message") || msg.includes("Unknown Channel");
 }

 function hasChannelMessagesApi(channel: unknown): channel is {
@@ -164,6 +270,21 @@ function hasPermissionApi(channel: unknown): channel is {

 // ─── Database Action Log ─────────────────────────────────────────────

+/**
+ * Extract the member's server-specific display name (nickname) from a
+ * captured message's metadata. Falls back to the global username when no
+ * server nickname exists (member.displayName defaults to the guild nickname
+ * when set, otherwise the global username).
+ */
+function resolveServerNick(message: MessageRecord): string | null {
+  try {
+    const parsed = parseRichMessageMetadata(message.metadata);
+    return parsed?.member?.displayName ?? message.username ?? null;
+  } catch {
+    return message.username ?? null;
+  }
+}
+
 async function logAutoDeleteAttempt(
  message: MessageRecord,
  result: AutoDeleteResult,
@@ -176,6 +297,8 @@ async function logAutoDeleteAttempt(
      action_type: "delete_message",
      reason: result.reason,
      ...verdictToActionFields(message),
+      username: message.username,
+      server_nick: resolveServerNick(message),
      executed_by: "auto-delete-manager",
      status: result.deleted
        ? "executed"
@@ -241,6 +364,8 @@ export async function attemptAutoDeleteFlaggedMessage(
            reason:
              "nickname melanggar aturan server (offensive_username); pesan dibiarkan",
            ...verdictToActionFields(message),
+            username: message.username,
+            server_nick: resolveServerNick(message),
            executed_by: "auto-delete-manager",
            status: resetOk ? "executed" : "failed",
            error: resetOk ? null : "nickname_reset_failed",
@@ -1,15 +1,7 @@
 /**
 * batchOutcomeClassifier.ts
 *
- * Pure partitioner of the batch worker response (2026-08-25).
- *
- * Bug history: the batch race guard returned `{ok:true, rows:[]}` when every
- * target's attachment upload was still in-flight. The processor classified all
- * of them as "incomplete" and fanned out to the individual queue, where the
- * guard there requeued + rescheduled at the 250ms debounce — a hot ~300ms loop
- * for the entire upload duration (~10 cycles in 3s in prod logs). Root fix:
- * the worker now reports `uploadPendingIds` explicitly and this pure function
- * partitions the outcome so upload-pending targets NEVER enter the fanout.
+ * Pure partitioner of the batch worker response.
 */

 export interface BatchRowLike {
@@ -21,15 +13,12 @@ export interface BatchRowLike {
 export interface BatchWorkerResponseLike {
  ok?: boolean;
  rows?: BatchRowLike[];
-  /** Explicit race-guard signal from the worker (2026-08-25). */
-  uploadPendingIds?: string[];
  error?: string;
 }

 /** One target's per-message disposition after a batch attempt. */
 export type BatchTargetKind =
  | "completed"
-  | "upload_pending"
  | "incomplete"
  | "parse_failed"
  | "api_failed";
@@ -40,7 +29,7 @@ function flagsOf(row: { ai_moderation_flags?: string | null }): string[] {
    const parsed = JSON.parse(row.ai_moderation_flags) as unknown;
    return Array.isArray(parsed) ? (parsed as string[]) : [];
  } catch {
-    return [] as string[];
+    return [];
  }
 }

@@ -48,14 +37,13 @@ function flagsOf(row: { ai_moderation_flags?: string | null }): string[] {
 * Partition the input message ids into per-message dispositions for one batch
 * worker response. Pure: no DB/Piscina/logger — unit-testable directly.
 *
- * Priority per id: explicit uploadPendingIds → completed row → flag-based
- * failure kinds → unexplained missing (treated like incomplete).
+ * Priority per id: completed row → flag-based failure kinds → unexplained
+ * missing (treated like incomplete).
 */
 export function partitionBatchOutcome(
  messages: ReadonlyArray<{ id: string }>,
  response: BatchWorkerResponseLike,
 ): Map<string, BatchTargetKind> {
-  const pendingSet = new Set(response.uploadPendingIds ?? []);
  const rowsById = new Map(
    (response.rows ?? [])
      .filter((r): r is BatchRowLike & { id: string } => Boolean(r?.id))
@@ -64,10 +52,6 @@ export function partitionBatchOutcome(

  const out = new Map<string, BatchTargetKind>();
  for (const msg of messages) {
-    if (pendingSet.has(msg.id)) {
-      out.set(msg.id, "upload_pending");
-      continue;
-    }
    const row = rowsById.get(msg.id);
    if (!row) {
      // Unexplained drop: LLM silently omitted it. Same retryable bucket as
@@ -92,17 +76,3 @@ export function partitionBatchOutcome(
  }
  return out;
 }
-
-/**
- * Linear backoff ramp for consecutive upload-pending polls:
- * poll N (1-based) waits min(base × N, cap). Keeps latency low for fast
- * uploads while bounding total polling cost for long uploads.
- */
-export function computeUploadPollDelayMs(
-  consecutivePolls: number,
-  baseMs: number,
-  capMs: number,
-): number {
-  const n = Math.max(1, Math.floor(consecutivePolls));
-  return Math.min(Math.round(baseMs * n), Math.round(capMs));
-}
@@ -4,10 +4,7 @@ import { isAgeRestrictedMetadata } from "../message-capture/messageMetadata.js";
 import { messageStore } from "../message-capture/messageStore.js";
 import type { MessageRecord } from "../message-capture/types.js";
 import { pickBatchWithinBudget as pickBatchWithinBudgetPure } from "./batchBudget.js";
-import {
-  computeUploadPollDelayMs,
-  partitionBatchOutcome,
-} from "./batchOutcomeClassifier.js";
+import { partitionBatchOutcome } from "./batchOutcomeClassifier.js";
 import { workerPool } from "./circuitBreaker.js";
 import { estimateTokens } from "./conversationContext.js";
 import {
@@ -25,20 +22,11 @@ import {

 const logger = createChildLogger("batch-processor");

-/**
- * Consecutive upload-pending poll counter per conversation (2026-08-25).
- * Drives the linear backoff ramp while attachments are still uploading;
- * cleared as soon as a batch comes back with no upload-pending targets.
- */
-const conversationUploadPolls = new Map<string, number>();
-
 export interface AnalysisWorkerResponse {
  ok: boolean;
  conversationKey: string;
  rows: MessageRecord[];
  error?: string;
-  /** Explicit upload-in-flight signal from the batch race guard (2026-08-25). */
-  uploadPendingIds?: string[];
 }

 // ---------------------------------------------------------------------------
@@ -154,8 +142,6 @@ export async function processBatch(

  activeRequests++;
  let shouldScheduleNext = false;
-  /** Set when upload-pending targets defer the next cycle by this many ms. */
-  let deferredUploadRescheduleMs: number | null = null;
  try {
    const result = (await workerPool.run({
      type: "batch",
@@ -163,20 +149,14 @@ export async function processBatch(
      messages,
    })) as AnalysisWorkerResponse;

-    // Do not broadcast or auto-delete if it's an API failure that will be reverted.
+    // Broadcast + auto-delete only for successfully analyzed rows.
+    // Error rows (API failures, parse failures, incomplete) will be
+    // retried by the individual fallback queue — do NOT schedule
+    // auto-delete for them (they'd be logged as not_eligible anyway).
    for (const row of result.rows) {
-      let isApiFailure = false;
-      if (row.ai_status === "error") {
-        try {
-          const flags = JSON.parse(row.ai_moderation_flags ?? "[]") as string[];
-          isApiFailure = flags.includes("analysis_api_failed");
-        } catch {}
-      }
-
-      if (!isApiFailure) {
-        broadcastAnalysisCompleted(row);
-        scheduleAutoDelete(row);
-      }
+      if (row.ai_status === "error") continue;
+      broadcastAnalysisCompleted(row);
+      scheduleAutoDelete(row);
    }

    if (!result.ok) {
@@ -214,19 +194,16 @@ export async function processBatch(
      return;
    }

-    // Batch succeeded -- partition per-message outcome explicitly (2026-08-25).
-    // upload_pending targets are DEFERRED (never fanned out): the old code
-    // treated them as incomplete -> individual queue -> requeue+250ms
-    // reschedule -> hot ~300ms loop for the whole upload duration.
+    // Batch succeeded -- partition per-message outcome (2026-08-25).
    const outcomeById = partitionBatchOutcome(messages, result);
    const messagesForIndividualQueue: MessageRecord[] = [];
    const apiFailedMessages: MessageRecord[] = [];
-    const uploadPendingMessages: MessageRecord[] = [];

    for (const msg of messages) {
      switch (outcomeById.get(msg.id)) {
-        case "upload_pending":
-          uploadPendingMessages.push(msg);
+        case "completed":
+          // Successfully analyzed — already broadcast + auto-delete scheduled
+          // above. Do NOT re-enqueue for individual fallback.
          break;
        case "api_failed":
          // Preserve the dedicated api-failure semantics below: revert +
@@ -241,59 +218,6 @@ export async function processBatch(
      }
    }

-    if (uploadPendingMessages.length > 0) {
-      const polls = (conversationUploadPolls.get(conversationKey) ?? 0) + 1;
-      conversationUploadPolls.set(conversationKey, polls);
-      const delayMs = computeUploadPollDelayMs(
-        polls,
-        config.AI_ANALYSIS_UPLOAD_POLL_MS,
-        config.AI_ANALYSIS_MAX_UPLOAD_POLL_MS,
-      );
-      logger.debug(
-        {
-          conversationKey,
-          count: uploadPendingMessages.length,
-          ids: uploadPendingMessages.map((m) => m.id),
-          pollAttempt: polls,
-          delayMs,
-        },
-        "Attachment upload in-flight for batch targets — deferring with poll backoff",
-      );
-
-      // Put the rows back to `pending` so the scheduler owns them again.
-      await messageStore
-        .updateMessagesAIAnalysisBulk(
-          uploadPendingMessages.map((msg) => ({
-            messageId: msg.id,
-            result: {
-              status: "pending",
-              flags: null,
-              score: null,
-              analysis: null,
-              categories: null,
-              severity: null,
-              confidence: null,
-              recommendedAction: null,
-              analyzedAt: null,
-              error: null,
-            },
-          })),
-        )
-        .catch((err: unknown) => {
-          logger.error(
-            { error: String(err), ids: uploadPendingMessages.map((m) => m.id) },
-            "Failed to revert upload-pending batch targets to pending",
-          );
-          return [] as MessageRecord[];
-        });
-
-      // Poll backoff instead of the 250ms debounce: the finally-block
-      // schedules the next cycle after this delay instead of immediately.
-      deferredUploadRescheduleMs = delayMs;
-    } else {
-      conversationUploadPolls.delete(conversationKey);
-    }
-
    if (messagesForIndividualQueue.length > 0) {
      logger.warn(
        {
@@ -370,11 +294,7 @@ export async function processBatch(
      resetConversationBatchFailures(conversationKey);
      conversationErrorCooldown.delete(conversationKey);
    }
-    // Upload-pending defer owns the next-cycle timing; don't let the default
-    // immediate schedule override it.
-    if (deferredUploadRescheduleMs === null) {
-      shouldScheduleNext = true;
-    }
+    shouldScheduleNext = true;
  } catch (error) {
    recordConversationBatchFailure(conversationKey);

@@ -411,17 +331,7 @@ export async function processBatch(
    if (conversationProcessing.get(conversationKey) === processingStartedAt) {
      conversationProcessing.delete(conversationKey);
    }
-    if (deferredUploadRescheduleMs !== null) {
-      // Upload still in-flight: re-schedule after the backoff delay instead of
-      // immediately (the old path hot-looped at ~250-300ms per cycle).
-      const delayMs = deferredUploadRescheduleMs;
-      setTimeout(() => {
-        // Dynamic import to avoid circular dependency at module scope
-        import("./batchScheduler.js").then((m) =>
-          m.scheduleConversationAnalysis(conversationKey),
-        );
-      }, delayMs).unref();
-    } else if (shouldScheduleNext) {
+    if (shouldScheduleNext) {
      setImmediate(() => {
        // Dynamic import to avoid circular dependency at module scope
        import("./batchScheduler.js").then((m) =>
@@ -136,6 +136,12 @@ export function formatMessageForPrompt(
  return `[${label}] id=${msg.id} time=${timestamp} user=${resolveDisplayName(msg)}: ${content}${mediaSuffix}${refInfo}`;
 }

+/** Estimate tokens for a batch of strings — single encode call, ~5x faster than per-line. */
+export function estimateTokensBatch(texts: string[]): number {
+  const combined = texts.join("\n");
+  return getEncoder().encode(combined).length + texts.length * 15;
+}
+
 /** Max content chars per context line — a single huge paste (log dump,
 *  copypasta) must not eat the whole conversation budget. */
 const CONTEXT_LINE_CONTENT_MAX_CHARS = 1500;
@@ -264,10 +270,7 @@ export function buildConversationContext(
  const targetLines = targets.map((msg) =>
    formatMessageForPrompt(msg, "target"),
  );
-  let usedTokens = targetLines.reduce(
-    (sum, line) => sum + estimateTokens(line),
-    0,
-  );
+  let usedTokens = estimateTokensBatch(targetLines);

  const contextLines = gatedNewestFirst.map((msg) =>
    formatMessageForPrompt(msg, "context"),
@@ -13,16 +13,10 @@
 * has an explicit, testable owner.
 */

-export type WorkerResultKind =
-  | "success"
-  | "upload_pending"
-  | "incomplete"
-  | "error";
+export type WorkerResultKind = "success" | "incomplete" | "error";

 export interface ClassifiableWorkerResult {
  ok?: boolean;
-  /** Upload-pending marker set by ai-analysis-worker's race guard. */
-  uploadPending?: boolean;
  results?: Array<{ status?: string; flags?: string[] | string } | undefined>;
  error?: string;
 }
@@ -40,7 +34,6 @@ function flagsOf(r: { flags?: string[] | string }): string[] {

 /**
 * Classify an individual-fallback worker response:
- * - "upload_pending": explicit race-guard signal — retry shortly, NOT an error.
 * - "success": at least one result and none is analysis_incomplete.
 * - "incomplete": LLM ran but dropped/failed this message after retries
 *   (analysis_incomplete flag) — terminal exhausted path.
@@ -50,7 +43,6 @@ function flagsOf(r: { flags?: string[] | string }): string[] {
 export function classifyIndividualWorkerResult(
  result: ClassifiableWorkerResult,
 ): WorkerResultKind {
-  if (result.uploadPending === true) return "upload_pending";
  const results = (result.results ?? []).filter(
    (r): r is NonNullable<typeof r> => Boolean(r),
  );
@@ -79,75 +79,21 @@ async function processIndividualFallback(
      message,
      skipNormalAnalysis: false,
    } as unknown)) as
-      | { ok: true; results: AnalysisResult[]; uploadPending?: boolean }
+      | { ok: true; results: AnalysisResult[] }
      | { ok: false; results: AnalysisResult[]; error: string };

-    // Explicit outcome classification (2026-08-24): the old code treated any
-    // ok:true as a completed moderation, so the upload-pending race guard's
-    // empty results left messages stuck in `processing` until the 300s
-    // cleanup reverted them — the root cause of the ~330s attachment delays.
+    // Explicit outcome classification: the old code treated any
+    // ok:true as a completed moderation, so empty or unexplainable
+    // results left messages stuck in `processing` until the 300s
+    // cleanup reverted them.
    const kind = classifyIndividualWorkerResult(workerResult);

-    if (kind === "upload_pending") {
-      // Attachment still uploading — put the row back to `pending` and
-      // re-schedule this conversation immediately. The next scheduler cycle
-      // (~debounce 250ms) re-fetches; once upload_status flips to done the
-      // race guard passes and analysis proceeds. NOT an error: never touches
-      // the circuit breaker counters.
-      const revertedRows = await messageStore
-        .updateMessagesAIAnalysisBulk([
-          {
-            messageId,
-            result: {
-              status: "pending",
-              flags: null,
-              score: null,
-              analysis: null,
-              categories: null,
-              severity: null,
-              confidence: null,
-              recommendedAction: null,
-              analyzedAt: null,
-              error: null,
-            },
-          },
-        ])
-        .catch((dbErr: unknown) => {
-          logger.error(
-            { messageId, error: String(dbErr) },
-            "Failed to revert upload-pending message to pending",
-          );
-          return [] as MessageRecord[];
-        });
-      for (const row of revertedRows) {
-        broadcastAnalysisCompleted(row);
-      }
-      logger.debug(
-        { messageId, conversationKey },
-        "Individual fallback: attachment upload in-flight — requeued as pending + rescheduled",
-      );
-      setImmediate(() => {
-        import("./batchScheduler.js")
-          .then((m) => m.scheduleConversationAnalysis(conversationKey))
-          .catch(() => {});
-      });
-      return;
-    }
-
    let analysisResult: { results: AnalysisResult[] } | null = null;
-
    if (kind === "success") {
      analysisResult = workerResult;
-    } else if (kind === "incomplete") {
-      exhaustedOnIncomplete = true;
-      analysisResult = null;
    } else {
-      // "error" — includes ok:true with unexplainable empty results (the old
-      // silent-success bug). Throw so it is treated as a transient failure.
-      throw new Error(
-        (workerResult as { error?: string }).error ??
-          "Individual worker returned no explainable results",
-      );
+      exhaustedOnIncomplete = kind === "incomplete";
+      analysisResult = null;
    }

    // No heuristic fallback: an incomplete/errored LLM result stays a
@@ -16,6 +16,7 @@ import { delay, retryWithBackoff } from "@/shared/utils/index";
 import { config } from "../../shared/config/config.js";
 import type { AnalysisResult } from "../message-capture/types.js";
 import { llmChat } from "./llmClient.js";
+import { parseModerationResponse } from "./moderationResponseParser.js";
 import { logModerationError } from "./responseLogger.js";

 const log = createChildLogger("llm-caller");
@@ -79,11 +80,12 @@ export async function callModerationLLM(
                ];
          const completion = await llmChat({
            messages,
-            max_tokens: maxTokens ?? 16384,
+            max_tokens:
+              maxTokens ?? config.AI_LLM_MAX_COMPLETION_TOKENS ?? 16384,
            jsonResponse: { type: "json_object" },
            retries: 0,
            signal,
-            // Router (9router / formerly omniroute) always streams SSE even when the
+            // Router (omniroute) always streams SSE even when the
            // request omits `stream`. In non-stream mode the OpenAI SDK waits
            // for the FULL body before parsing, so slow/long upstream streams
            // hit the 30s/60s timeout and abort mid-generation. Streaming mode
@@ -106,9 +108,6 @@ export async function callModerationLLM(
          if (!rawContent) throw new Error("No content in LLM response");

          try {
-            const { parseModerationResponse } = await import(
-              "./moderationResponseParser.js"
-            );
            return {
              parsed: parseModerationResponse(rawContent, targetIds),
              result: completion,
@@ -94,7 +94,7 @@ type LLMResponseChunk = {
 * `delta.content`; falls back to reasoning fields so reasoning-only models
 * still produce usable aggregated text. Providers differ in the field name:
 * - DeepSeek-style / Cloudflare gemma → `delta.reasoning_content`
- * - mimo (via 9router) streams reasoning in `delta.reasoning` +
+ * - mimo (via omniroute) streams reasoning in `delta.reasoning` +
 *   `delta.reasoning_details[].text` (content:"") — without these fallbacks
 *   vision aggregation came back empty ("Vision API null response").
 * Exported for unit tests.
@@ -225,7 +225,7 @@ export function buildLlmParams(
      reasoning: { enabled: false },
      // vLLM / Qwen / litellm
      chat_template_kwargs: { enable_thinking: false },
-      // Anthropic / Claude-format (9router exposes thinkingFormat
+      // Anthropic / Claude-format (omniroute exposes thinkingFormat
      // "claude-adaptive" / "claude-budget" on its reasoning models)
      thinking: { type: "disabled" },
    } as Record<string, unknown>);
@@ -296,14 +296,13 @@ export async function downloadAndExtractFrame(
  imageMap: Map<string, MessageImagePart[]>,
 ): Promise<void> {
  const log = createChildLogger("mediaAnalysis");
-  // Prefer the upload proxy (uploaded_url); the Discord CDN link can expire
-  // or be purged (404), and a non-OK response used to silently drop the image
-  // from vision analysis (no log, empty image map → text-only verdict). Try
-  // each candidate URL in order and surface failures.
+  // Analysis now uses the Discord CDN URL directly (uploaded_url is archive-only).
+  // Try discord_url first; fall back to uploaded_url (Tele proxy) if the CDN
+  // link returns a non-OK response (expired/purged).
  const urlCandidates = [
-    att.uploaded_url,
-    att.discord_url && att.discord_url !== att.uploaded_url
-      ? att.discord_url
+    att.discord_url,
+    att.uploaded_url && att.uploaded_url !== att.discord_url
+      ? att.uploaded_url
      : null,
  ].filter((u): u is string => Boolean(u));
  if (urlCandidates.length === 0) return;
@@ -461,7 +461,9 @@ export async function runModerationAnalysis(
      cacheKey,
      stored,
      embeddingsByKey.get(cacheKey),
-    ).catch(() => {});
+    ).catch((err: unknown) => {
+      log.warn({ cacheKey, error: String(err) }, "Cache write failed");
+    });

    // Dual-key write-back (2026-08-24): the FIRST analysis of a message runs
    // WITH conversation context (accurate), but its verdict is also stored
@@ -69,6 +69,26 @@ export async function buildCorrectedFewShotExamples(): Promise<string> {
  }
 }

+// ---------------------------------------------------------------------------
+// Few-shot correction cache (refreshes hourly)
+// ---------------------------------------------------------------------------
+let _correctedExamplesCache: string | null = null;
+let _correctedExamplesCacheAt = 0;
+const CORRECTED_CACHE_TTL_MS = 60 * 60 * 1000;
+
+async function getCachedCorrectedExamples(): Promise<string> {
+  const now = Date.now();
+  if (
+    _correctedExamplesCache !== null &&
+    now - _correctedExamplesCacheAt < CORRECTED_CACHE_TTL_MS
+  ) {
+    return _correctedExamplesCache;
+  }
+  _correctedExamplesCache = await buildCorrectedFewShotExamples();
+  _correctedExamplesCacheAt = now;
+  return _correctedExamplesCache;
+}
+
 // ---------------------------------------------------------------------------
 // Text-only batch
 // ---------------------------------------------------------------------------
@@ -88,7 +108,22 @@ export async function runTextOnlyBatch(
      for (const url of extractUrlsFromText(msg.edited_content ?? msg.content))
        allUrls.add(url);
    }
-    const urlArr = Array.from(allUrls).slice(0, 10);
+    // Domain dedup: max 3 URLs per domain to avoid rate-limiting
+    const domainCounts = new Map<string, number>();
+    const urlArr: string[] = [];
+    for (const url of allUrls) {
+      try {
+        const domain = new URL(url).hostname;
+        const count = domainCounts.get(domain) ?? 0;
+        if (count >= 3) continue;
+        domainCounts.set(domain, count + 1);
+      } catch {
+        /* invalid URL, skip */
+        continue;
+      }
+      urlArr.push(url);
+      if (urlArr.length >= 10) break;
+    }
    if (urlArr.length === 0) {
      return {
        text: new Map<string, string>(),
@@ -192,7 +227,7 @@ export async function runTextOnlyBatch(
  // Corrected false-positive examples are static per batch — fetch ONCE
  // here instead of inside the per-sub-batch retry closure (which would
  // re-query the DB on every sub-batch and every parse-error retry).
-  const correctedExamples = await buildCorrectedFewShotExamples();
+  const correctedExamples = await getCachedCorrectedExamples();

  // ── URL images → multimodal vision evidence (hoisted out of the sub-batch
  //     loop) ───────────────────────────────────────────────────────────
@@ -368,9 +403,13 @@ export async function runTextOnlyBatch(
      );
    } catch (err: any) {
      if (err.name === "AbortError" || abortController.signal.aborted) {
-        throw new Error(
-          `Text-only batch sub-batch ${i + 1} timed out for messages ${targetIds.join(", ")}`,
+        // Sub-batch timed out — log but DO NOT throw. Previous sub-batches'
+        // results are already in allResults; throwing would discard them.
+        log.warn(
+          { subBatch: i + 1, targetIds, timeoutMs },
+          "Sub-batch timed out — preserving partial results from prior sub-batches",
        );
+        continue;
      }
      throw err;
    } finally {
@@ -103,6 +103,8 @@ export class ModerationHandler {
          | "kick_user"
          | "ban_user",
        reason: payload.reason ?? null,
+        username: null,
+        server_nick: null,
        executed_by: payload.executed_by ?? "command-handler",
        status: "executed",
        error: null,
				`@@ -0,0 +1 @@`
				`ALTER TABLE "moderation_actions" ADD COLUMN IF NOT EXISTS "server_name" text;`