fix: inline imports

* feat: make video recording optional, deleting video tracks

.gitignore

@@ -3,6 +3,8 @@ server/.env
 server/.env.production
 .env
 Caddyfile
+livekit.yaml
+egress.yaml
 .env.hatchet
 server/exportdanswer
 .vercel
@@ -24,5 +26,10 @@ www/.env.production
 .secrets
 opencode.json
 
+certs/
+docker-compose.ca.yml
+docker-compose.gpu-ca.yml
+Caddyfile.gpu-host
+.env.gpu-host
 vibedocs/
 server/tests/integration/logs/

CHANGELOG.md

@@ -1,5 +1,63 @@
 # Changelog
 
+## [0.44.0](https://github.com/GreyhavenHQ/reflector/compare/v0.43.0...v0.44.0) (2026-04-07)
+
+
+### Features
+
+* allow livekit to work with tunnels udp and tcp with quickguide ([#948](https://github.com/GreyhavenHQ/reflector/issues/948)) ([f4f94a0](https://github.com/GreyhavenHQ/reflector/commit/f4f94a0d9998030e5ef7f01935d99722045165ac))
+* Livekit - Selfhost video room solution ([#946](https://github.com/GreyhavenHQ/reflector/issues/946)) ([bc8338f](https://github.com/GreyhavenHQ/reflector/commit/bc8338fa4f136534f5f27784f5dd10d47cecf412))
+
+
+### Bug Fixes
+
+* allow anonymous recording start and add light theme email icon ([#949](https://github.com/GreyhavenHQ/reflector/issues/949)) ([08c276e](https://github.com/GreyhavenHQ/reflector/commit/08c276e4f493c7960158a44e77062af1b30c3352))
+* deactivate meeting button and better deactivation heuristics ([#950](https://github.com/GreyhavenHQ/reflector/issues/950)) ([26239f0](https://github.com/GreyhavenHQ/reflector/commit/26239f05a34af07ebba764d669343c32e40e63bf))
+
+## [0.43.0](https://github.com/GreyhavenHQ/reflector/compare/v0.42.0...v0.43.0) (2026-03-31)
+
+
+### Features
+
+* show trash for soft deleted transcripts and hard delete option ([#942](https://github.com/GreyhavenHQ/reflector/issues/942)) ([ec8b497](https://github.com/GreyhavenHQ/reflector/commit/ec8b49738e8e76f6e5d2496a42cb454ef6c2d7c7))
+
+
+### Bug Fixes
+
+* add parakeet as default transcriber and fix diarizer image ([#939](https://github.com/GreyhavenHQ/reflector/issues/939)) ([cc9c5cd](https://github.com/GreyhavenHQ/reflector/commit/cc9c5cd4a5f4123ef957ad82461ca37a727d1ba6))
+* cpu usage + email improvements ([#944](https://github.com/GreyhavenHQ/reflector/issues/944)) ([8c4f5e9](https://github.com/GreyhavenHQ/reflector/commit/8c4f5e9c0f893f4cb029595505b53136f04760f4))
+
+## [0.42.0](https://github.com/GreyhavenHQ/reflector/compare/v0.41.0...v0.42.0) (2026-03-30)
+
+
+### Features
+
+* custom ca for caddy ([#931](https://github.com/GreyhavenHQ/reflector/issues/931)) ([12bf0c2](https://github.com/GreyhavenHQ/reflector/commit/12bf0c2d77f9915b79b1eb1decd77ed2dadbb31d))
+* mixdown modal services + processor pattern ([#936](https://github.com/GreyhavenHQ/reflector/issues/936)) ([d164e48](https://github.com/GreyhavenHQ/reflector/commit/d164e486cc33ff8babf6cff6c163893cfc56fd76))
+
+
+### Bug Fixes
+
+* grpc tls for local hatchet ([#937](https://github.com/GreyhavenHQ/reflector/issues/937)) ([a22789d](https://github.com/GreyhavenHQ/reflector/commit/a22789d5486bf8b83e33ab2fb5eb3ee9799c6d47))
+* remove share public from integration tests ([#938](https://github.com/GreyhavenHQ/reflector/issues/938)) ([7b3b5b9](https://github.com/GreyhavenHQ/reflector/commit/7b3b5b98586449afd0b6996ba9fd7aec8308bbc6))
+
+## [0.41.0](https://github.com/GreyhavenHQ/reflector/compare/v0.40.0...v0.41.0) (2026-03-25)
+
+
+### Features
+
+* add auto-generated captions, speaker-colored progress bar with sync controls, and speaker tooltip to cloud video player ([#926](https://github.com/GreyhavenHQ/reflector/issues/926)) ([f19113a](https://github.com/GreyhavenHQ/reflector/commit/f19113a3cfa27797a70b9496bfcf1baff9d89f0d))
+* send email in share transcript and add email sending in room ([#924](https://github.com/GreyhavenHQ/reflector/issues/924)) ([e2ba502](https://github.com/GreyhavenHQ/reflector/commit/e2ba502697ce331c4d87fb019648fcbe4e7cca73))
+* zulip dag monitor for failed runs ([#928](https://github.com/GreyhavenHQ/reflector/issues/928)) ([1f98790](https://github.com/GreyhavenHQ/reflector/commit/1f98790e7bc58013690ec81aefa051da5e36e93e))
+
+## [0.40.0](https://github.com/GreyhavenHQ/reflector/compare/v0.39.0...v0.40.0) (2026-03-20)
+
+
+### Features
+
+* allow participants to ask for email transcript ([#923](https://github.com/GreyhavenHQ/reflector/issues/923)) ([55222ec](https://github.com/GreyhavenHQ/reflector/commit/55222ecc4736f99ad461f03a006c8d97b5876142))
+* download files, show cloud video, solf deletion with no reprocessing ([#920](https://github.com/GreyhavenHQ/reflector/issues/920)) ([a76f114](https://github.com/GreyhavenHQ/reflector/commit/a76f1143783d3cf137a8847a851b72302e04445b))
+
 ## [0.39.0](https://github.com/GreyhavenHQ/reflector/compare/v0.38.2...v0.39.0) (2026-03-18)
 
 

CLAUDE.md

@@ -41,14 +41,14 @@ uv run celery -A reflector.worker.app beat
 
 **Testing:**
 ```bash
-# Run all tests with coverage
-uv run pytest
+# Run all tests with coverage (requires Redis on localhost)
+REDIS_HOST=localhost REDIS_PORT=6379 uv run pytest
 
 # Run specific test file
-uv run pytest tests/test_transcripts.py
+REDIS_HOST=localhost REDIS_PORT=6379 uv run pytest tests/test_transcripts.py
 
 # Run tests with verbose output
-uv run pytest -v
+REDIS_HOST=localhost REDIS_PORT=6379 uv run pytest -v
 ```
 
 **Process Audio Files:**
@@ -192,3 +192,13 @@ Modal.com integration for scalable ML processing:
 ## Pipeline/worker related info
 
 If you need to do any worker/pipeline related work, search for "Pipeline" classes and their "create" or "build" methods to find the main processor sequence. Look for task orchestration patterns (like "chord", "group", or "chain") to identify the post-processing flow with parallel execution chains. This will give you abstract vision on how processing pipeling is organized.
+
+## Documentation
+
+- New documentation files go in `docsv2/`, not in `docs/docs/`.
+- Existing `docs/` directory contains legacy Docusaurus docs.
+
+## Code Style
+
+- Always put imports at the top of the file. Let ruff/pre-commit handle sorting and formatting of imports.
+- The **only** imports allowed to remain inline are from `reflector.db.*` modules (e.g., `reflector.db.transcripts`, `reflector.db.meetings`, `reflector.db.recordings`, `reflector.db.rooms`). These stay as deferred/inline imports inside `fresh_db_connection()` blocks in Hatchet pipeline task functions — this is intentional to avoid sharing DB connections across forked processes. All other imports (utilities, services, processors, storage, third-party libs) **must** go at the top of the file, even in Hatchet workflows.

docker-compose.gpu-host.yml

@@ -0,0 +1,106 @@
+# Standalone GPU host for Reflector — transcription, diarization, translation.
+#
+# Usage: ./scripts/setup-gpu-host.sh [--domain DOMAIN] [--custom-ca PATH] [--api-key KEY] [--cpu]
+#   or:  docker compose -f docker-compose.gpu-host.yml --profile gpu [--profile caddy] up -d
+#
+# Processing mode (pick ONE — mutually exclusive, both bind port 8000):
+#   --profile gpu      NVIDIA GPU container (requires nvidia-container-toolkit)
+#   --profile cpu      CPU-only container (no GPU required, slower)
+#
+# Optional:
+#   --profile caddy    Caddy reverse proxy with HTTPS
+#
+# This file is checked into the repo. The setup script generates:
+#   - .env.gpu-host            (HF_TOKEN, API key, port config)
+#   - Caddyfile.gpu-host       (Caddy config, only with --domain)
+#   - docker-compose.gpu-ca.yml (CA cert mounts, only with --custom-ca)
+
+services:
+  # ===========================================================
+  # GPU service — NVIDIA GPU accelerated
+  # Activated with: --profile gpu
+  # ===========================================================
+
+  gpu:
+    build:
+      context: ./gpu/self_hosted
+      dockerfile: Dockerfile
+    profiles: [gpu]
+    restart: unless-stopped
+    ports:
+      - "${GPU_HOST_PORT:-8000}:8000"
+    environment:
+      HF_TOKEN: ${HF_TOKEN:-}
+      REFLECTOR_GPU_APIKEY: ${REFLECTOR_GPU_APIKEY:-}
+    volumes:
+      - gpu_cache:/root/.cache
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8000/docs"]
+      interval: 15s
+      timeout: 5s
+      retries: 10
+      start_period: 120s
+    networks:
+      default:
+        aliases:
+          - transcription
+
+  # ===========================================================
+  # CPU service — no GPU required, uses Dockerfile.cpu
+  # Activated with: --profile cpu
+  # Mutually exclusive with gpu (both bind port 8000)
+  # ===========================================================
+
+  cpu:
+    build:
+      context: ./gpu/self_hosted
+      dockerfile: Dockerfile.cpu
+    profiles: [cpu]
+    restart: unless-stopped
+    ports:
+      - "${GPU_HOST_PORT:-8000}:8000"
+    environment:
+      HF_TOKEN: ${HF_TOKEN:-}
+      REFLECTOR_GPU_APIKEY: ${REFLECTOR_GPU_APIKEY:-}
+    volumes:
+      - gpu_cache:/root/.cache
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8000/docs"]
+      interval: 15s
+      timeout: 5s
+      retries: 10
+      start_period: 120s
+    networks:
+      default:
+        aliases:
+          - transcription
+
+  # ===========================================================
+  # Caddy — reverse proxy with HTTPS (optional)
+  # Activated with: --profile caddy
+  # Proxies to "transcription" network alias (works for both gpu and cpu)
+  # ===========================================================
+
+  caddy:
+    image: caddy:2-alpine
+    profiles: [caddy]
+    restart: unless-stopped
+    ports:
+      - "80:80"
+      - "${CADDY_HTTPS_PORT:-443}:443"
+    volumes:
+      - ./Caddyfile.gpu-host:/etc/caddy/Caddyfile:ro
+      - caddy_data:/data
+      - caddy_config:/config
+
+volumes:
+  gpu_cache:
+  caddy_data:
+  caddy_config:

docker-compose.selfhosted.yml

@@ -35,8 +35,8 @@ services:
     image: monadicalsas/reflector-backend:latest
     restart: unless-stopped
     ports:
-      - "127.0.0.1:1250:1250"
-      - "51000-51100:51000-51100/udp"
+      - "${BIND_HOST:-127.0.0.1}:1250:1250"
+      - "40000-40100:40000-40100/udp"
     env_file:
       - ./server/.env
     environment:
@@ -50,7 +50,7 @@ services:
       # HF_TOKEN needed for in-process pyannote diarization (--cpu mode)
       HF_TOKEN: ${HF_TOKEN:-}
       # WebRTC: fixed UDP port range for ICE candidates (mapped above)
-      WEBRTC_PORT_RANGE: "51000-51100"
+      WEBRTC_PORT_RANGE: "40000-40100"
       # Hatchet workflow engine (always-on for processing pipelines)
       HATCHET_CLIENT_SERVER_URL: ${HATCHET_CLIENT_SERVER_URL:-http://hatchet:8888}
       HATCHET_CLIENT_HOST_PORT: ${HATCHET_CLIENT_HOST_PORT:-hatchet:7077}
@@ -116,7 +116,7 @@ services:
     image: monadicalsas/reflector-frontend:latest
     restart: unless-stopped
     ports:
-      - "127.0.0.1:3000:3000"
+      - "${BIND_HOST:-127.0.0.1}:3000:3000"
     env_file:
       - ./www/.env
     environment:
@@ -308,6 +308,24 @@ services:
       - web
       - server
 
+  # ===========================================================
+  # Mailpit — local SMTP sink for testing email transcript notifications
+  # Start with: --profile mailpit
+  # Web UI at http://localhost:8025
+  # ===========================================================
+
+  mailpit:
+    image: axllent/mailpit:latest
+    profiles: [mailpit]
+    restart: unless-stopped
+    ports:
+      - "127.0.0.1:8025:8025"   # Web UI
+    healthcheck:
+      test: ["CMD", "wget", "-q", "--spider", "http://localhost:8025/api/v1/messages"]
+      interval: 10s
+      timeout: 3s
+      retries: 5
+
   # ===========================================================
   # Hatchet workflow engine + workers
   # Required for all processing pipelines (file, live, Daily.co multitrack).
@@ -321,7 +339,7 @@ services:
       postgres:
         condition: service_healthy
     ports:
-      - "127.0.0.1:8888:8888"
+      - "0.0.0.0:8888:8888"   # Hatchet dashboard (plain HTTP — no TLS)
       - "127.0.0.1:7078:7077"
     env_file:
       - ./.env.hatchet
@@ -348,7 +366,7 @@ services:
       context: ./server
       dockerfile: Dockerfile
     image: monadicalsas/reflector-backend:latest
-    profiles: [dailyco]
+    profiles: [dailyco, livekit]
     restart: unless-stopped
     env_file:
       - ./server/.env
@@ -388,6 +406,40 @@ services:
     volumes:
       - server_data:/app/data
 
+  # ===========================================================
+  # LiveKit — self-hosted open-source video platform
+  # Activated via --profile livekit (auto-detected from LIVEKIT_API_KEY in server/.env)
+  # ===========================================================
+
+  livekit-server:
+    image: livekit/livekit-server:v1.10.1
+    profiles: [livekit]
+    restart: unless-stopped
+    ports:
+      - "7880:7880"                     # Signaling (HTTP/WS)
+      - "7881:7881"                     # WebRTC over TCP
+      - "${LIVEKIT_UDP_PORTS:-44200-44300:44200-44300}/udp"   # WebRTC ICE (range or single port for tunnels)
+    volumes:
+      - ./livekit.yaml:/etc/livekit.yaml:ro
+    command: ["--config", "/etc/livekit.yaml"]
+    depends_on:
+      redis:
+        condition: service_started
+
+  livekit-egress:
+    image: livekit/egress:v1.12.0
+    profiles: [livekit]
+    restart: unless-stopped
+    environment:
+      EGRESS_CONFIG_FILE: /etc/egress.yaml
+    volumes:
+      - ./egress.yaml:/etc/egress.yaml:ro
+    depends_on:
+      redis:
+        condition: service_started
+      livekit-server:
+        condition: service_started
+
 volumes:
   postgres_data:
   redis_data:

docs/docs/installation/daily-setup.md

@@ -95,6 +95,12 @@ DAILYCO_STORAGE_AWS_BUCKET_NAME=<your-bucket-from-daily-setup>
 DAILYCO_STORAGE_AWS_REGION=us-east-1
 DAILYCO_STORAGE_AWS_ROLE_ARN=<your-role-arn-from-daily-setup>
 
+# Worker credentials for reading/deleting recordings from Daily's S3 bucket.
+# Required when transcript storage uses a different bucket or credentials
+# (e.g., selfhosted with Garage or a separate S3 account).
+DAILYCO_STORAGE_AWS_ACCESS_KEY_ID=<your-aws-access-key>
+DAILYCO_STORAGE_AWS_SECRET_ACCESS_KEY=<your-aws-secret-key>
+
 # Transcript storage (should already be configured from main setup)
 # TRANSCRIPT_STORAGE_BACKEND=aws
 # TRANSCRIPT_STORAGE_AWS_ACCESS_KEY_ID=<your-key>
@@ -103,6 +109,19 @@ DAILYCO_STORAGE_AWS_ROLE_ARN=<your-role-arn-from-daily-setup>
 # TRANSCRIPT_STORAGE_AWS_REGION=<your-bucket-region>
 ```
 
+:::info Two separate credential sets for Daily.co
+
+- **`ROLE_ARN`** — Used by Daily's API to *write* recordings into your S3 bucket (configured via Daily dashboard).
+- **`ACCESS_KEY_ID` / `SECRET_ACCESS_KEY`** — Used by Reflector workers to *read* recordings for transcription and *delete* them on consent denial or permanent transcript deletion.
+
+Required IAM permissions for the worker key on the Daily recordings bucket:
+- `s3:GetObject` — Download recording files for processing
+- `s3:DeleteObject` — Remove files on consent denial, trash destroy, or data retention cleanup
+- `s3:ListBucket` — Scan for recordings needing reprocessing
+
+If the worker keys are not set, Reflector falls back to the transcript storage master key, which then needs cross-bucket access to the Daily bucket.
+:::
+
 ---
 
 ## Restart Services

docs/pnpm-lock.yaml

@@ -701,6 +701,10 @@ packages:
     resolution: {integrity: sha512-05WQkdpL9COIMz4LjTxGpPNCdlpyimKppYNoJ5Di5EUObifl8t4tuLuUBBZEpoLYOmfvIWrsp9fCl0HoPRVTdA==}
     engines: {node: '>=6.9.0'}
 
+  '@babel/runtime@7.29.2':
+    resolution: {integrity: sha512-JiDShH45zKHWyGe4ZNVRrCjBz8Nh9TMmZG1kh4QTK8hCBTWBi8Da+i7s1fJw7/lYpM4ccepSNfqzZ/QvABBi5g==}
+    engines: {node: '>=6.9.0'}
+
   '@babel/template@7.28.6':
     resolution: {integrity: sha512-YA6Ma2KsCdGb+WC6UpBVFJGXL58MDA6oyONbjyF/+5sBgxY/dwkhLogbMT2GXXyU84/IhRw/2D1Os1B/giz+BQ==}
     engines: {node: '>=6.9.0'}
@@ -1490,42 +1494,36 @@ packages:
     engines: {node: '>= 10.0.0'}
     cpu: [arm]
     os: [linux]
-    libc: [glibc]
 
   '@parcel/watcher-linux-arm-musl@2.5.6':
     resolution: {integrity: sha512-Ve3gUCG57nuUUSyjBq/MAM0CzArtuIOxsBdQ+ftz6ho8n7s1i9E1Nmk/xmP323r2YL0SONs1EuwqBp2u1k5fxg==}
     engines: {node: '>= 10.0.0'}
     cpu: [arm]
     os: [linux]
-    libc: [musl]
 
   '@parcel/watcher-linux-arm64-glibc@2.5.6':
     resolution: {integrity: sha512-f2g/DT3NhGPdBmMWYoxixqYr3v/UXcmLOYy16Bx0TM20Tchduwr4EaCbmxh1321TABqPGDpS8D/ggOTaljijOA==}
     engines: {node: '>= 10.0.0'}
     cpu: [arm64]
     os: [linux]
-    libc: [glibc]
 
   '@parcel/watcher-linux-arm64-musl@2.5.6':
     resolution: {integrity: sha512-qb6naMDGlbCwdhLj6hgoVKJl2odL34z2sqkC7Z6kzir8b5W65WYDpLB6R06KabvZdgoHI/zxke4b3zR0wAbDTA==}
     engines: {node: '>= 10.0.0'}
     cpu: [arm64]
     os: [linux]
-    libc: [musl]
 
   '@parcel/watcher-linux-x64-glibc@2.5.6':
     resolution: {integrity: sha512-kbT5wvNQlx7NaGjzPFu8nVIW1rWqV780O7ZtkjuWaPUgpv2NMFpjYERVi0UYj1msZNyCzGlaCWEtzc+exjMGbQ==}
     engines: {node: '>= 10.0.0'}
     cpu: [x64]
     os: [linux]
-    libc: [glibc]
 
   '@parcel/watcher-linux-x64-musl@2.5.6':
     resolution: {integrity: sha512-1JRFeC+h7RdXwldHzTsmdtYR/Ku8SylLgTU/reMuqdVD7CtLwf0VR1FqeprZ0eHQkO0vqsbvFLXUmYm/uNKJBg==}
     engines: {node: '>= 10.0.0'}
     cpu: [x64]
     os: [linux]
-    libc: [musl]
 
   '@parcel/watcher-win32-arm64@2.5.6':
     resolution: {integrity: sha512-3ukyebjc6eGlw9yRt678DxVF7rjXatWiHvTXqphZLvo7aC5NdEgFufVwjFfY51ijYEWpXbqF5jtrK275z52D4Q==}
@@ -2254,11 +2252,11 @@ packages:
     resolution: {integrity: sha512-2hCgjEmP8YLWQ130n2FerGv7rYpfBmnmp9Uy2Le1vge6X3gZIfSmEzP5QTDElFxcvVcXlEn8Aq6MU/PZygIOog==}
     engines: {node: '>=14.16'}
 
-  brace-expansion@1.1.12:
-    resolution: {integrity: sha512-9T9UjW3r0UW5c1Q7GTwllptXwhvYmEzFhzMfZ9H7FQWt+uZePjZPjBP/W1ZEyZ1twGWom5/56TF4lPcqjnDHcg==}
+  brace-expansion@1.1.13:
+    resolution: {integrity: sha512-9ZLprWS6EENmhEOpjCYW2c8VkmOvckIJZfkr7rBW6dObmfgJ/L1GpSYW5Hpo9lDz4D1+n0Ckz8rU7FwHDQiG/w==}
 
-  brace-expansion@2.0.2:
-    resolution: {integrity: sha512-Jt0vHyM+jmUBqojB7E1NIYadt0vI0Qxjxd2TErW94wDz+E2LAm5vKMXXwg6ZZBTHPuUlDgQHKXvjGBdfcF1ZDQ==}
+  brace-expansion@2.0.3:
+    resolution: {integrity: sha512-MCV/fYJEbqx68aE58kv2cA/kiky1G8vux3OR6/jbS+jIMe/6fJWa0DTzJU7dqijOWYwHi1t29FlfYI9uytqlpA==}
 
   braces@3.0.3:
     resolution: {integrity: sha512-yQbXgO/OSZVD2IsiLlro+7Hf6Q18EJrKSEsdoMzKePKXct3gvD8oLcOQdIzGupr5Fj+EDe8gO/lxc1BzfMpxvA==}
@@ -3410,8 +3408,8 @@ packages:
   graphlib@2.1.8:
     resolution: {integrity: sha512-jcLLfkpoVGmH7/InMC/1hIvOPSUh38oJtGhvrOFGzioE1DZ+0YW16RgmOJhHiuWTvGiJQ9Z1Ik43JvkRPRvE+A==}
 
-  gray-matter@https://codeload.github.com/jonschlinkert/gray-matter/tar.gz/234163e317c87fe031e9368ffabde9c9149ce3ec:
-    resolution: {tarball: https://codeload.github.com/jonschlinkert/gray-matter/tar.gz/234163e317c87fe031e9368ffabde9c9149ce3ec}
+  gray-matter@https://codeload.github.com/jonschlinkert/gray-matter/tar.gz/234163e:
+    resolution: {tarball: https://codeload.github.com/jonschlinkert/gray-matter/tar.gz/234163e}
     version: 4.0.3
     engines: {node: '>=6.0'}
 
@@ -4533,8 +4531,8 @@ packages:
   path-parse@1.0.7:
     resolution: {integrity: sha512-LDJzPVEEEPR+y48z93A0Ed0yXb8pAByGWo/k5YYdYgpY2/2EsOsksJrq7lOHxryrVOn1ejG6oAp8ahvOIQD8sw==}
 
-  path-to-regexp@0.1.12:
-    resolution: {integrity: sha512-RA1GjUVMnvYFxuqovrEqZoxxW5NUZqbwKtYz/Tt7nXerk0LbLblQmrsgdeOxV5SFHf0UDggjS/bSeOZwt1pmEQ==}
+  path-to-regexp@0.1.13:
+    resolution: {integrity: sha512-A/AGNMFN3c8bOlvV9RreMdrv7jsmF9XIfDeCd87+I8RNg6s78BhJxMu69NEMHBSJFxKidViTEdruRwEk/WIKqA==}
 
   path-to-regexp@1.9.0:
     resolution: {integrity: sha512-xIp7/apCFJuUHdDLWe8O1HIkb0kQrOMb/0u6FXQjemHn/ii5LrIzU6bdECnsiTF/GjZkMEKg1xdiZwNqDYlZ6g==}
@@ -4555,12 +4553,12 @@ packages:
   picocolors@1.1.1:
     resolution: {integrity: sha512-xceH2snhtb5M9liqDsmEw56le376mTZkEX/jEb/RxNFyegNul7eNslCXP9FDj/Lcu0X8KEyMceP2ntpaHrDEVA==}
 
-  picomatch@2.3.1:
-    resolution: {integrity: sha512-JU3teHTNjmE2VCGFzuY8EXzCDVwEqB2a8fsIvwaStHhAWJEeVd1o1QD80CU6+ZdEXXSLbSsuLwJjkCBWqRQUVA==}
+  picomatch@2.3.2:
+    resolution: {integrity: sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA==}
     engines: {node: '>=8.6'}
 
-  picomatch@4.0.3:
-    resolution: {integrity: sha512-5gTmgEY/sqK6gFXLIsQNH19lWb4ebPDLA4SdLP7dsWkIXHWlG66oPuVvXSGFPppYZz8ZDZq0dYYrbHfBCVUb1Q==}
+  picomatch@4.0.4:
+    resolution: {integrity: sha512-QP88BAKvMam/3NxH6vj2o21R6MjxZUAd6nlwAS/pnGvN9IVLocLHxGYIzFhg6fUQ+5th6P4dv4eW9jX3DSIj7A==}
     engines: {node: '>=12'}
 
   pirates@4.0.7:
@@ -7024,6 +7022,8 @@ snapshots:
 
   '@babel/runtime@7.28.6': {}
 
+  '@babel/runtime@7.29.2': {}
+
   '@babel/template@7.28.6':
     dependencies:
       '@babel/code-frame': 7.29.0
@@ -8162,7 +8162,7 @@ snapshots:
       fs-extra: 11.3.3
       github-slugger: 1.5.0
       globby: 11.1.0
-      gray-matter: https://codeload.github.com/jonschlinkert/gray-matter/tar.gz/234163e317c87fe031e9368ffabde9c9149ce3ec
+      gray-matter: https://codeload.github.com/jonschlinkert/gray-matter/tar.gz/234163e
       jiti: 1.21.7
       js-yaml: 4.1.1
       lodash: 4.17.23
@@ -8473,7 +8473,7 @@ snapshots:
       detect-libc: 2.1.2
       is-glob: 4.0.3
       node-addon-api: 7.1.1
-      picomatch: 4.0.3
+      picomatch: 4.0.4
     optionalDependencies:
       '@parcel/watcher-android-arm64': 2.5.6
       '@parcel/watcher-darwin-arm64': 2.5.6
@@ -8645,7 +8645,7 @@ snapshots:
 
   '@slorber/react-helmet-async@1.3.0(react-dom@19.2.4(react@19.2.4))(react@19.2.4)':
     dependencies:
-      '@babel/runtime': 7.28.6
+      '@babel/runtime': 7.29.2
       invariant: 2.2.4
       prop-types: 15.8.1
       react: 19.2.4
@@ -9244,7 +9244,7 @@ snapshots:
   anymatch@3.1.3:
     dependencies:
       normalize-path: 3.0.0
-      picomatch: 2.3.1
+      picomatch: 2.3.2
 
   arg@5.0.2: {}
 
@@ -9378,12 +9378,12 @@ snapshots:
       widest-line: 4.0.1
       wrap-ansi: 8.1.0
 
-  brace-expansion@1.1.12:
+  brace-expansion@1.1.13:
     dependencies:
       balanced-match: 1.0.2
       concat-map: 0.0.1
 
-  brace-expansion@2.0.2:
+  brace-expansion@2.0.3:
     dependencies:
       balanced-match: 1.0.2
 
@@ -10436,7 +10436,7 @@ snapshots:
       methods: 1.1.2
       on-finished: 2.4.1
       parseurl: 1.3.3
-      path-to-regexp: 0.1.12
+      path-to-regexp: 0.1.13
       proxy-addr: 2.0.7
       qs: 6.14.2
       range-parser: 1.2.1
@@ -10485,9 +10485,9 @@ snapshots:
     dependencies:
       websocket-driver: 0.7.4
 
-  fdir@6.5.0(picomatch@4.0.3):
+  fdir@6.5.0(picomatch@4.0.4):
     optionalDependencies:
-      picomatch: 4.0.3
+      picomatch: 4.0.4
 
   feed@4.2.2:
     dependencies:
@@ -10658,7 +10658,7 @@ snapshots:
     dependencies:
       lodash: 4.17.23
 
-  gray-matter@https://codeload.github.com/jonschlinkert/gray-matter/tar.gz/234163e317c87fe031e9368ffabde9c9149ce3ec:
+  gray-matter@https://codeload.github.com/jonschlinkert/gray-matter/tar.gz/234163e:
     dependencies:
       js-yaml: 4.1.1
       kind-of: 6.0.3
@@ -11080,7 +11080,7 @@ snapshots:
       chalk: 4.1.2
       ci-info: 3.9.0
       graceful-fs: 4.2.11
-      picomatch: 2.3.1
+      picomatch: 2.3.2
 
   jest-worker@27.5.1:
     dependencies:
@@ -11780,7 +11780,7 @@ snapshots:
   micromatch@4.0.8:
     dependencies:
       braces: 3.0.3
-      picomatch: 2.3.1
+      picomatch: 2.3.2
 
   mime-db@1.33.0: {}
 
@@ -11824,11 +11824,11 @@ snapshots:
 
   minimatch@3.1.5:
     dependencies:
-      brace-expansion: 1.1.12
+      brace-expansion: 1.1.13
 
   minimatch@5.1.8:
     dependencies:
-      brace-expansion: 2.0.2
+      brace-expansion: 2.0.3
 
   minimist@1.2.8: {}
 
@@ -12127,7 +12127,7 @@ snapshots:
 
   path-parse@1.0.7: {}
 
-  path-to-regexp@0.1.12: {}
+  path-to-regexp@0.1.13: {}
 
   path-to-regexp@1.9.0:
     dependencies:
@@ -12146,9 +12146,9 @@ snapshots:
 
   picocolors@1.1.1: {}
 
-  picomatch@2.3.1: {}
+  picomatch@2.3.2: {}
 
-  picomatch@4.0.3: {}
+  picomatch@4.0.4: {}
 
   pirates@4.0.7: {}
 
@@ -12852,7 +12852,7 @@ snapshots:
 
   readdirp@3.6.0:
     dependencies:
-      picomatch: 2.3.1
+      picomatch: 2.3.2
 
   readdirp@4.1.2: {}
 
@@ -13510,8 +13510,8 @@ snapshots:
 
   tinyglobby@0.2.15:
     dependencies:
-      fdir: 6.5.0(picomatch@4.0.3)
-      picomatch: 4.0.3
+      fdir: 6.5.0(picomatch@4.0.4)
+      picomatch: 4.0.4
 
   tinypool@1.1.1: {}
 

docsv2/custom-ca-setup.md

@@ -0,0 +1,338 @@
+# Custom CA Certificate Setup
+
+Use a private Certificate Authority (CA) with Reflector self-hosted deployments. This covers two scenarios:
+
+1. **Custom local domain** — Serve Reflector over HTTPS on an internal domain (e.g., `reflector.local`) using certs signed by your own CA
+2. **Backend CA trust** — Let Reflector's backend services (server, workers, GPU) make HTTPS calls to GPU, LLM, or other internal services behind your private CA
+
+Both can be used independently or together.
+
+## Quick Start
+
+### Generate test certificates
+
+```bash
+./scripts/generate-certs.sh reflector.local
+```
+
+This creates `certs/` with:
+- `ca.key` + `ca.crt` — Root CA (10-year validity)
+- `server-key.pem` + `server.pem` — Server certificate (1-year, SAN: domain + localhost + 127.0.0.1)
+
+### Deploy with custom CA + domain
+
+```bash
+# Add domain to /etc/hosts on the server (use 127.0.0.1 for local, or server LAN IP for network access)
+echo "127.0.0.1 reflector.local" | sudo tee -a /etc/hosts
+
+# Run setup — pass the certs directory
+./scripts/setup-selfhosted.sh --gpu --caddy --domain reflector.local --custom-ca certs/
+
+# Trust the CA on your machine (see "Trust the CA" section below)
+```
+
+### Deploy with CA trust only (GPU/LLM behind private CA)
+
+```bash
+# Only need the CA cert file — no Caddy TLS certs needed
+./scripts/setup-selfhosted.sh --hosted --custom-ca /path/to/corporate-ca.crt
+```
+
+## How `--custom-ca` Works
+
+The flag accepts a **directory** or a **single file**:
+
+### Directory mode
+
+```bash
+--custom-ca certs/
+```
+
+Looks for these files by convention:
+- `ca.crt` (required) — CA certificate to trust
+- `server.pem` + `server-key.pem` (optional) — TLS certificate/key for Caddy
+
+If `server.pem` + `server-key.pem` are found AND `--domain` is provided:
+- Caddy serves HTTPS using those certs
+- Backend containers trust the CA for outbound calls
+
+If only `ca.crt` is found:
+- Backend containers trust the CA for outbound calls
+- Caddy is unaffected (uses Let's Encrypt, self-signed, or no Caddy)
+
+### Single file mode
+
+```bash
+--custom-ca /path/to/corporate-ca.crt
+```
+
+Only injects CA trust into backend containers. No Caddy TLS changes.
+
+## Scenarios
+
+### Scenario 1: Custom local domain
+
+Your Reflector instance runs on an internal network. You want `https://reflector.local` with proper TLS (no browser warnings).
+
+```bash
+# 1. Generate certs
+./scripts/generate-certs.sh reflector.local
+
+# 2. Add to /etc/hosts on the server
+echo "127.0.0.1 reflector.local" | sudo tee -a /etc/hosts
+
+# 3. Deploy
+./scripts/setup-selfhosted.sh --gpu --garage --caddy --domain reflector.local --custom-ca certs/
+
+# 4. Trust the CA on your machine (see "Trust the CA" section below)
+```
+
+If other machines on the network need to access it, add the server's LAN IP to `/etc/hosts` on those machines instead:
+```bash
+echo "192.168.1.100 reflector.local" | sudo tee -a /etc/hosts
+```
+
+And include that IP as an extra SAN when generating certs:
+```bash
+./scripts/generate-certs.sh reflector.local "IP:192.168.1.100"
+```
+
+### Scenario 2: GPU/LLM behind corporate CA
+
+Your GPU or LLM server (e.g., `https://gpu.internal.corp`) uses certificates signed by your corporate CA. Reflector's backend needs to trust that CA for outbound HTTPS calls.
+
+```bash
+# Get the CA certificate from your IT team (PEM format)
+# Then deploy — Caddy can still use Let's Encrypt or self-signed
+./scripts/setup-selfhosted.sh --hosted --garage --caddy --custom-ca /path/to/corporate-ca.crt
+```
+
+This works because:
+- **TLS cert/key** = "this is my identity" — for Caddy to serve HTTPS to browsers
+- **CA cert** = "I trust this authority" — for backend containers to verify outbound connections
+
+Your Reflector frontend can use Let's Encrypt (public domain) or self-signed certs, while the backend trusts a completely different CA for GPU/LLM calls.
+
+### Scenario 3: Both combined (same CA)
+
+Custom domain + GPU/LLM all behind the same CA:
+
+```bash
+./scripts/generate-certs.sh reflector.local "DNS:gpu.local"
+./scripts/setup-selfhosted.sh --gpu --garage --caddy --domain reflector.local --custom-ca certs/
+```
+
+### Scenario 4: Multiple CAs (local domain + remote GPU on different CA)
+
+Your Reflector uses one CA for `reflector.local`, but the GPU host uses a different CA:
+
+```bash
+# Your local domain setup
+./scripts/generate-certs.sh reflector.local
+
+# Deploy with your CA + trust the GPU host's CA too
+./scripts/setup-selfhosted.sh --hosted --garage --caddy \
+    --domain reflector.local \
+    --custom-ca certs/ \
+    --extra-ca /path/to/gpu-machine-ca.crt
+```
+
+`--extra-ca` appends additional CA certs to the trust bundle. Backend containers trust ALL CAs — your local domain AND the GPU host's certs both work.
+
+You can repeat `--extra-ca` for multiple remote services:
+```bash
+--extra-ca /path/to/gpu-ca.crt --extra-ca /path/to/llm-ca.crt
+```
+
+For setting up a dedicated GPU host, see [Standalone GPU Host Setup](gpu-host-setup.md).
+
+## Trust the CA on Client Machines
+
+After deploying, clients need to trust the CA to avoid browser warnings.
+
+### macOS
+
+```bash
+sudo security add-trusted-cert -d -r trustRoot \
+    -k /Library/Keychains/System.keychain certs/ca.crt
+```
+
+### Linux (Ubuntu/Debian)
+
+```bash
+sudo cp certs/ca.crt /usr/local/share/ca-certificates/reflector-ca.crt
+sudo update-ca-certificates
+```
+
+### Linux (RHEL/Fedora)
+
+```bash
+sudo cp certs/ca.crt /etc/pki/ca-trust/source/anchors/reflector-ca.crt
+sudo update-ca-trust
+```
+
+### Windows (PowerShell as admin)
+
+```powershell
+Import-Certificate -FilePath .\certs\ca.crt -CertStoreLocation Cert:\LocalMachine\Root
+```
+
+### Firefox (all platforms)
+
+Firefox uses its own certificate store:
+1. Settings > Privacy & Security > View Certificates
+2. Authorities tab > Import
+3. Select `ca.crt` and check "Trust this CA to identify websites"
+
+## How It Works Internally
+
+### Docker entrypoint CA injection
+
+Each backend container (server, worker, beat, hatchet workers, GPU) has an entrypoint script (`docker-entrypoint.sh`) that:
+
+1. Checks if a CA cert is mounted at `/usr/local/share/ca-certificates/custom-ca.crt`
+2. If present, runs `update-ca-certificates` to create a **combined bundle** (system CAs + custom CA)
+3. Sets environment variables so all Python/gRPC libraries use the combined bundle:
+
+| Env var | Covers |
+|---------|--------|
+| `SSL_CERT_FILE` | httpx, OpenAI SDK, llama-index, Python ssl module |
+| `REQUESTS_CA_BUNDLE` | requests library (transitive dependencies) |
+| `CURL_CA_BUNDLE` | curl CLI (container healthchecks) |
+
+Note: `GRPC_DEFAULT_SSL_ROOTS_FILE_PATH` is intentionally NOT set. Setting it causes grpcio to attempt TLS on internal Hatchet gRPC connections that run without TLS, resulting in handshake failures. The internal Hatchet connection uses `HATCHET_CLIENT_TLS_STRATEGY=none` (plaintext).
+
+When no CA cert is mounted, the entrypoint is a no-op — containers behave exactly as before.
+
+### Why this replaces manual certifi patching
+
+Previously, the workaround for trusting a private CA in Python was to patch certifi's bundle directly:
+
+```bash
+# OLD approach — fragile, do NOT use
+cat custom-ca.crt >> $(python -c "import certifi; print(certifi.where())")
+```
+
+This breaks whenever certifi is updated (any `pip install`/`uv sync` overwrites the bundle and the CA is lost).
+
+Our entrypoint approach is permanent because:
+
+1. `SSL_CERT_FILE` is checked by Python's `ssl.create_default_context()` **before** falling back to `certifi.where()`. When set, certifi's bundle is never read.
+2. `REQUESTS_CA_BUNDLE` similarly overrides certifi for the `requests` library.
+3. The CA is injected at container startup (runtime), not baked into the Python environment. It survives image rebuilds, dependency updates, and `uv sync`.
+
+```
+Python SSL lookup chain:
+  ssl.create_default_context()
+    → SSL_CERT_FILE env var? → YES → use combined bundle (system + custom CA) ✓
+    → (certifi.where() is never reached)
+```
+
+This covers all outbound HTTPS calls: httpx (transcription, diarization, translation, webhooks), OpenAI SDK (transcription), llama-index (LLM/summarization), and requests (transitive dependencies).
+
+### Compose override
+
+The setup script generates `docker-compose.ca.yml` which mounts the CA cert into every backend container as a read-only bind mount. This file is:
+- Only generated when `--custom-ca` is passed
+- Deleted on re-runs without `--custom-ca` (prevents stale overrides)
+- Added to `.gitignore`
+
+### Node.js (frontend)
+
+The web container uses `NODE_EXTRA_CA_CERTS` which **adds** to Node's trust store (unlike Python's `SSL_CERT_FILE` which replaces it). This is set via the compose override.
+
+## Generate Your Own CA (Manual)
+
+If you prefer not to use `generate-certs.sh`:
+
+```bash
+# 1. Create CA
+openssl genrsa -out ca.key 4096
+openssl req -x509 -new -nodes -key ca.key -sha256 -days 3650 \
+    -out ca.crt -subj "/CN=My CA/O=My Organization"
+
+# 2. Create server key
+openssl genrsa -out server-key.pem 2048
+
+# 3. Create CSR with SANs
+openssl req -new -key server-key.pem -out server.csr \
+    -subj "/CN=reflector.local" \
+    -addext "subjectAltName=DNS:reflector.local,DNS:localhost,IP:127.0.0.1"
+
+# 4. Sign with CA
+openssl x509 -req -in server.csr -CA ca.crt -CAkey ca.key \
+    -CAcreateserial -out server.pem -days 365 -sha256 \
+    -copy_extensions copyall
+
+# 5. Clean up
+rm server.csr ca.srl
+```
+
+## Using Existing Corporate Certificates
+
+If your organization already has a CA:
+
+1. Get the CA certificate in PEM format from your IT team
+2. If you have a PKCS#12 (.p12/.pfx) bundle, extract the CA cert:
+   ```bash
+   openssl pkcs12 -in bundle.p12 -cacerts -nokeys -out ca.crt
+   ```
+3. If you have multiple intermediate CAs, concatenate them into one PEM file:
+   ```bash
+   cat intermediate-ca.crt root-ca.crt > ca.crt
+   ```
+
+## Troubleshooting
+
+### Browser: "Your connection is not private"
+
+The CA is not trusted on the client machine. See "Trust the CA" section above.
+
+Check certificate expiry:
+```bash
+openssl x509 -noout -dates -in certs/server.pem
+```
+
+### Backend: `SSL: CERTIFICATE_VERIFY_FAILED`
+
+CA cert not mounted or not loaded. Check inside the container:
+```bash
+docker compose exec server env | grep SSL_CERT_FILE
+docker compose exec server python -c "
+import ssl, os
+print('SSL_CERT_FILE:', os.environ.get('SSL_CERT_FILE', 'not set'))
+ctx = ssl.create_default_context()
+print('CA certs loaded:', ctx.cert_store_stats())
+"
+```
+
+### Caddy: "certificate is not valid for any names"
+
+Domain in Caddyfile doesn't match the certificate's SAN/CN. Check:
+```bash
+openssl x509 -noout -text -in certs/server.pem | grep -A1 "Subject Alternative Name"
+```
+
+### Certificate chain issues
+
+If you have intermediate CAs, concatenate them into `server.pem`:
+```bash
+cat server-cert.pem intermediate-ca.pem > certs/server.pem
+```
+
+Verify the chain:
+```bash
+openssl verify -CAfile certs/ca.crt certs/server.pem
+```
+
+### Certificate renewal
+
+Custom CA certs are NOT auto-renewed (unlike Let's Encrypt). Replace cert files and restart:
+```bash
+# Replace certs
+cp new-server.pem certs/server.pem
+cp new-server-key.pem certs/server-key.pem
+
+# Restart Caddy to pick up new certs
+docker compose restart caddy
+```

docsv2/firewall-ports.md

@@ -0,0 +1,97 @@
+# Firewall & Port Requirements
+
+Ports that need to be open on your server firewall, organized by deployment mode.
+
+## With Caddy (--caddy or --ip or --domain)
+
+Caddy acts as the reverse proxy. Most services are only accessible through Caddy on port 443.
+
+| Port | Protocol | Direction | Service | Required? |
+|------|----------|-----------|---------|-----------|
+| 443 | TCP | Inbound | Caddy HTTPS — web app, API, LiveKit signaling (`/lk-ws`) | Yes |
+| 80 | TCP | Inbound | Caddy HTTP — redirects to HTTPS | Yes |
+| 44200-44300 | UDP | Inbound | LiveKit WebRTC media (audio/video) | Yes (if LiveKit enabled) |
+| 7881 | TCP | Inbound | LiveKit TCP media fallback (when UDP is blocked by client network) | Recommended |
+| 8888 | TCP | Inbound | Hatchet dashboard (plain HTTP, no TLS) | Optional (admin only) |
+
+Ports that do NOT need to be open (proxied through Caddy):
+- 1250 (backend API)
+- 3000 (frontend)
+- 7880 (LiveKit signaling — proxied via `/lk-ws`)
+- 3900 (Garage S3)
+
+## Without Caddy (direct access)
+
+All services need direct port access. Use this only for local development or trusted networks.
+
+| Port | Protocol | Direction | Service | Required? |
+|------|----------|-----------|---------|-----------|
+| 3000 | TCP | Inbound | Frontend (Next.js) | Yes |
+| 1250 | TCP | Inbound | Backend API (FastAPI) | Yes |
+| 7880 | TCP | Inbound | LiveKit signaling (WebSocket) | Yes (if LiveKit enabled) |
+| 7881 | TCP | Inbound | LiveKit TCP media fallback | Recommended |
+| 44200-44300 | UDP | Inbound | LiveKit WebRTC media | Yes (if LiveKit enabled) |
+| 40000-40100 | UDP | Inbound | Reflector WebRTC (browser recording) | Yes (if using browser WebRTC) |
+| 3900 | TCP | Inbound | Garage S3 (for presigned URLs in browser) | Yes (if using Garage) |
+| 8888 | TCP | Inbound | Hatchet dashboard | Optional |
+
+> **Important:** Without Caddy, all traffic is plain HTTP. Browsers block microphone/camera access on non-HTTPS pages (except `localhost`). Use `--ip` (which implies Caddy) for any non-localhost deployment.
+
+## Internal-Only Ports (never expose)
+
+These ports are used between Docker containers and should NOT be open on the firewall:
+
+| Port | Service | Purpose |
+|------|---------|---------|
+| 5432 | PostgreSQL | Database |
+| 6379 | Redis | Cache + message broker |
+| 7077 | Hatchet gRPC | Worker communication |
+
+## Cloud Provider Firewall Examples
+
+### DigitalOcean (with Caddy + LiveKit)
+
+```bash
+# Create firewall
+doctl compute firewall create \
+  --name reflector \
+  --inbound-rules "protocol:tcp,ports:443,address:0.0.0.0/0 protocol:tcp,ports:80,address:0.0.0.0/0 protocol:udp,ports:44200-44300,address:0.0.0.0/0 protocol:tcp,ports:7881,address:0.0.0.0/0 protocol:tcp,ports:22,address:0.0.0.0/0" \
+  --outbound-rules "protocol:tcp,ports:all,address:0.0.0.0/0 protocol:udp,ports:all,address:0.0.0.0/0" \
+  --droplet-ids <DROPLET_ID>
+```
+
+### AWS Security Group (with Caddy + LiveKit)
+
+| Type | Port Range | Source | Description |
+|------|-----------|--------|-------------|
+| HTTPS | 443 | 0.0.0.0/0 | Web app + API + LiveKit signaling |
+| HTTP | 80 | 0.0.0.0/0 | Redirect to HTTPS |
+| Custom UDP | 44200-44300 | 0.0.0.0/0 | LiveKit WebRTC media |
+| Custom TCP | 7881 | 0.0.0.0/0 | LiveKit TCP fallback |
+| SSH | 22 | Your IP | Admin access |
+
+### Ubuntu UFW (with Caddy + LiveKit)
+
+```bash
+sudo ufw allow 443/tcp    # Caddy HTTPS
+sudo ufw allow 80/tcp     # HTTP redirect
+sudo ufw allow 7881/tcp   # LiveKit TCP fallback
+sudo ufw allow 44200:44300/udp  # LiveKit WebRTC media
+sudo ufw allow 22/tcp     # SSH
+sudo ufw enable
+```
+
+## Port Ranges Explained
+
+### Why 44200-44300 for LiveKit?
+
+LiveKit's WebRTC ICE candidates use UDP. The port range was chosen to avoid collisions:
+- **40000-40100** — Reflector's own WebRTC (browser recording)
+- **44200-44300** — LiveKit WebRTC
+- **49152-65535** — macOS ephemeral ports (reserved by OS)
+
+The range is configurable in `livekit.yaml` under `rtc.port_range_start` / `rtc.port_range_end`. If changed, update `docker-compose.selfhosted.yml` port mapping to match.
+
+### Why 101 ports?
+
+100 UDP ports support ~100 concurrent WebRTC connections (roughly 50 participants with audio + video). For larger deployments, increase the range in both `livekit.yaml` and `docker-compose.selfhosted.yml`.

docsv2/gpu-host-setup.md

@@ -0,0 +1,294 @@
+# Standalone GPU Host Setup
+
+Deploy Reflector's GPU transcription/diarization/translation service on a dedicated machine, separate from the main Reflector instance. Useful when:
+
+- Your GPU machine is on a different network than the Reflector server
+- You want to share one GPU service across multiple Reflector instances
+- The GPU machine has special hardware/drivers that can't run the full stack
+- You need to scale GPU processing independently
+
+## Architecture
+
+```
+┌─────────────────────┐         HTTPS          ┌────────────────────┐
+│  Reflector Server    │ ────────────────────── │  GPU Host          │
+│  (server, worker,    │  TRANSCRIPT_URL        │  (transcription,   │
+│   web, postgres,     │  DIARIZATION_URL       │   diarization,     │
+│   redis, hatchet)    │  TRANSLATE_URL         │   translation)     │
+│                      │                        │                    │
+│  setup-selfhosted.sh │                        │  setup-gpu-host.sh │
+│  --hosted            │                        │                    │
+└─────────────────────┘                        └────────────────────┘
+```
+
+The GPU service is a standalone FastAPI app that exposes transcription, diarization, translation, and audio padding endpoints. It has **no dependencies** on PostgreSQL, Redis, Hatchet, or any other Reflector service.
+
+## Quick Start
+
+### On the GPU machine
+
+```bash
+git clone <reflector-repo>
+cd reflector
+
+# Set HuggingFace token (required for diarization models)
+export HF_TOKEN=your-huggingface-token
+
+# Deploy with HTTPS (Let's Encrypt)
+./scripts/setup-gpu-host.sh --domain gpu.example.com --api-key my-secret-key
+
+# Or deploy with custom CA
+./scripts/generate-certs.sh gpu.local
+./scripts/setup-gpu-host.sh --domain gpu.local --custom-ca certs/ --api-key my-secret-key
+```
+
+### On the Reflector machine
+
+```bash
+# If the GPU host uses a custom CA, trust it
+./scripts/setup-selfhosted.sh --hosted --garage --caddy \
+    --extra-ca /path/to/gpu-machine-ca.crt
+
+# Or if you already have --custom-ca for your local domain
+./scripts/setup-selfhosted.sh --hosted --garage --caddy \
+    --domain reflector.local --custom-ca certs/ \
+    --extra-ca /path/to/gpu-machine-ca.crt
+```
+
+Then configure `server/.env` to point to the GPU host:
+
+```bash
+TRANSCRIPT_BACKEND=modal
+TRANSCRIPT_URL=https://gpu.example.com
+TRANSCRIPT_MODAL_API_KEY=my-secret-key
+
+DIARIZATION_BACKEND=modal
+DIARIZATION_URL=https://gpu.example.com
+DIARIZATION_MODAL_API_KEY=my-secret-key
+
+TRANSLATION_BACKEND=modal
+TRANSLATE_URL=https://gpu.example.com
+TRANSLATION_MODAL_API_KEY=my-secret-key
+```
+
+## Script Options
+
+```
+./scripts/setup-gpu-host.sh [OPTIONS]
+
+Options:
+  --domain DOMAIN    Domain name for HTTPS (Let's Encrypt or custom cert)
+  --custom-ca PATH   Custom CA (directory or single PEM file)
+  --extra-ca FILE    Additional CA cert to trust (repeatable)
+  --api-key KEY      API key to protect the service (strongly recommended)
+  --cpu              CPU-only mode (no NVIDIA GPU required)
+  --port PORT        Host port (default: 443 with Caddy, 8000 without)
+```
+
+## Deployment Scenarios
+
+### Public internet with Let's Encrypt
+
+GPU machine has a public IP and domain:
+
+```bash
+./scripts/setup-gpu-host.sh --domain gpu.example.com --api-key my-secret-key
+```
+
+Requirements:
+- DNS A record: `gpu.example.com` → GPU machine's public IP
+- Ports 80 and 443 open
+- Caddy auto-provisions Let's Encrypt certificate
+
+### Internal network with custom CA
+
+GPU machine on a private network:
+
+```bash
+# Generate certs on the GPU machine
+./scripts/generate-certs.sh gpu.internal "IP:192.168.1.200"
+
+# Deploy
+./scripts/setup-gpu-host.sh --domain gpu.internal --custom-ca certs/ --api-key my-secret-key
+```
+
+On each machine that connects (including the Reflector server), add DNS:
+```bash
+echo "192.168.1.200 gpu.internal" | sudo tee -a /etc/hosts
+```
+
+### IP-only (no domain)
+
+No domain needed — just use the machine's IP:
+
+```bash
+./scripts/setup-gpu-host.sh --api-key my-secret-key
+```
+
+Caddy is not used; the GPU service runs directly on port 8000 (HTTP). For HTTPS without a domain, the Reflector machine connects via `http://<GPU_IP>:8000`.
+
+### CPU-only (no NVIDIA GPU)
+
+Works on any machine — transcription will be slower:
+
+```bash
+./scripts/setup-gpu-host.sh --cpu --domain gpu.example.com --api-key my-secret-key
+```
+
+## DNS Resolution
+
+The Reflector server must be able to reach the GPU host by name or IP.
+
+| Setup | DNS Method | TRANSCRIPT_URL example |
+|-------|------------|----------------------|
+| Public domain | DNS A record | `https://gpu.example.com` |
+| Internal domain | `/etc/hosts` on both machines | `https://gpu.internal` |
+| IP only | No DNS needed | `http://192.168.1.200:8000` |
+
+For internal domains, add the GPU machine's IP to `/etc/hosts` on the Reflector machine:
+```bash
+echo "192.168.1.200 gpu.internal" | sudo tee -a /etc/hosts
+```
+
+If the Reflector server runs in Docker, the containers resolve DNS from the host (Docker's default DNS behavior). So adding to the host's `/etc/hosts` is sufficient.
+
+## Multi-CA Setup
+
+When your Reflector instance has its own CA (for `reflector.local`) and the GPU host has a different CA:
+
+**On the GPU machine:**
+```bash
+./scripts/generate-certs.sh gpu.local
+./scripts/setup-gpu-host.sh --domain gpu.local --custom-ca certs/ --api-key my-key
+```
+
+**On the Reflector machine:**
+```bash
+# Your local CA for reflector.local + the GPU host's CA
+./scripts/setup-selfhosted.sh --hosted --garage --caddy \
+    --domain reflector.local \
+    --custom-ca certs/ \
+    --extra-ca /path/to/gpu-machine-ca.crt
+```
+
+The `--extra-ca` flag appends the GPU host's CA to the trust bundle. Backend containers trust both CAs — your local domain works AND outbound calls to the GPU host succeed.
+
+You can repeat `--extra-ca` for multiple remote services:
+```bash
+--extra-ca /path/to/gpu-ca.crt --extra-ca /path/to/llm-ca.crt
+```
+
+## API Key Authentication
+
+The GPU service uses Bearer token authentication via `REFLECTOR_GPU_APIKEY`:
+
+```bash
+# Test from the Reflector machine
+curl -s https://gpu.example.com/docs                              # No auth needed for docs
+curl -s -X POST https://gpu.example.com/v1/audio/transcriptions \
+    -H "Authorization: Bearer <my-secret-key>" \                    #gitleaks:allow
+    -F "file=@audio.wav"
+```
+
+If `REFLECTOR_GPU_APIKEY` is not set, the service accepts all requests (open access). Always use `--api-key` for internet-facing deployments.
+
+The same key goes in Reflector's `server/.env` as `TRANSCRIPT_MODAL_API_KEY` and `DIARIZATION_MODAL_API_KEY`.
+
+## Files
+
+| File | Checked in? | Purpose |
+|------|-------------|---------|
+| `docker-compose.gpu-host.yml` | Yes | Static compose file with profiles (`gpu`, `cpu`, `caddy`) |
+| `.env.gpu-host` | No (generated) | Environment variables (HF_TOKEN, API key, ports) |
+| `Caddyfile.gpu-host` | No (generated) | Caddy config (only when using HTTPS) |
+| `docker-compose.gpu-ca.yml` | No (generated) | CA cert mounts override (only with --custom-ca) |
+| `certs/` | No (generated) | Staged certificates (when using --custom-ca) |
+
+The compose file is checked into the repo — you can read it to understand exactly what runs. The script only generates env vars, Caddyfile, and CA overrides. Profiles control which service starts:
+
+```bash
+# What the script does under the hood:
+docker compose -f docker-compose.gpu-host.yml --profile gpu --profile caddy \
+    --env-file .env.gpu-host up -d
+
+# CPU mode:
+docker compose -f docker-compose.gpu-host.yml --profile cpu --profile caddy \
+    --env-file .env.gpu-host up -d
+```
+
+Both `gpu` and `cpu` services get the network alias `transcription`, so Caddy's config works with either.
+
+## Management
+
+```bash
+# View logs
+docker compose -f docker-compose.gpu-host.yml --profile gpu logs -f gpu
+
+# Restart
+docker compose -f docker-compose.gpu-host.yml --profile gpu restart gpu
+
+# Stop
+docker compose -f docker-compose.gpu-host.yml --profile gpu --profile caddy down
+
+# Re-run setup
+./scripts/setup-gpu-host.sh [same flags]
+
+# Rebuild after code changes
+docker compose -f docker-compose.gpu-host.yml --profile gpu build gpu
+docker compose -f docker-compose.gpu-host.yml --profile gpu up -d gpu
+```
+
+If you deployed with `--custom-ca`, include the CA override in manual commands:
+```bash
+docker compose -f docker-compose.gpu-host.yml -f docker-compose.gpu-ca.yml \
+    --profile gpu logs -f gpu
+```
+
+## Troubleshooting
+
+### GPU service won't start
+
+Check logs:
+```bash
+docker compose -f docker-compose.gpu-host.yml logs gpu
+```
+
+Common causes:
+- NVIDIA driver not installed or `nvidia-container-toolkit` missing
+- `HF_TOKEN` not set (diarization model download fails)
+- Port already in use
+
+### Reflector can't connect to GPU host
+
+From the Reflector machine:
+```bash
+# Test HTTPS connectivity
+curl -v https://gpu.example.com/docs
+
+# If using custom CA, test with explicit CA
+curl --cacert /path/to/gpu-ca.crt https://gpu.internal/docs
+```
+
+From inside the Reflector container:
+```bash
+docker compose exec server python -c "
+import httpx
+r = httpx.get('https://gpu.internal/docs')
+print(r.status_code)
+"
+```
+
+### SSL: CERTIFICATE_VERIFY_FAILED
+
+The Reflector backend doesn't trust the GPU host's CA. Fix:
+```bash
+# Re-run Reflector setup with the GPU host's CA
+./scripts/setup-selfhosted.sh --hosted --extra-ca /path/to/gpu-ca.crt
+```
+
+### Diarization returns errors
+
+- Accept pyannote model licenses on HuggingFace:
+  - https://huggingface.co/pyannote/speaker-diarization-3.1
+  - https://huggingface.co/pyannote/segmentation-3.0
+- Verify `HF_TOKEN` is set in `.env.gpu-host`

docsv2/livekit-setup.md

@@ -0,0 +1,297 @@
+# LiveKit Setup (Self-Hosted Video Platform)
+
+LiveKit is the recommended open-source, self-hosted video platform for Reflector. It replaces Daily.co for deployments that need free, fully self-hosted video rooms with per-participant audio recording.
+
+> LiveKit runs alongside Daily.co and Whereby — you choose the platform per room. Existing Daily/Whereby setups are not affected.
+
+## What LiveKit Provides
+
+- **Video/audio rooms** — WebRTC-based conferencing via `livekit-server` (Go SFU)
+- **Per-participant audio recording** — Track Egress writes each participant's audio to S3 as a separate OGG/Opus file (no composite video, no Chrome dependency)
+- **S3-compatible storage** — works with Garage, MinIO, AWS S3, or any S3-compatible provider via `force_path_style`
+- **Webhook events** — participant join/leave, egress start/end, room lifecycle
+- **JWT access tokens** — per-participant tokens with granular permissions
+
+## Architecture
+
+```
+                    ┌─────────────────┐
+  Participants ────>│  livekit-server  │ :7880 (WS signaling)
+   (browser)        │   (Go SFU)      │ :7881 (TCP RTC)
+                    │                  │ :44200-44300/udp (ICE)
+                    └────────┬────────┘
+                             │ media forwarding
+                    ┌────────┴────────┐
+                    │  livekit-egress  │  Track Egress
+                    │  (per-track OGG) │  writes to S3
+                    └────────┬────────┘
+                             │
+                    ┌────────┴────────┐
+                    │   S3 Storage    │  Garage / MinIO / AWS
+                    │ (audio tracks)  │
+                    └─────────────────┘
+```
+
+Both services share Redis with the existing Reflector stack (same instance, same db).
+
+## Quick Start
+
+### Option 1: Via Setup Script (Recommended)
+
+Pass `--livekit` to the setup script. It generates all credentials and config automatically:
+
+```bash
+# First run — --livekit generates credentials and config files
+./scripts/setup-selfhosted.sh --gpu --ollama-gpu --livekit --garage --caddy
+
+# Re-runs — LiveKit is auto-detected from existing LIVEKIT_API_KEY in server/.env
+./scripts/setup-selfhosted.sh
+```
+
+The `--livekit` flag will:
+1. Generate `LIVEKIT_API_KEY` and `LIVEKIT_API_SECRET` (random credentials)
+2. Set `LIVEKIT_URL`, `LIVEKIT_PUBLIC_URL`, and storage credentials in `server/.env`
+3. Generate `livekit.yaml` and `egress.yaml` config files
+4. Set `DEFAULT_VIDEO_PLATFORM=livekit`
+5. Enable the `livekit` Docker Compose profile
+6. Start `livekit-server` and `livekit-egress` containers
+
+On subsequent re-runs (without flags), the script detects the existing `LIVEKIT_API_KEY` in `server/.env` and re-enables the profile automatically.
+
+### Option 2: Manual Setup
+
+If you prefer manual configuration:
+
+1. **Generate credentials:**
+
+```bash
+export LK_KEY="reflector_$(openssl rand -hex 8)"
+export LK_SECRET="$(openssl rand -hex 32)"
+```
+
+2. **Add to `server/.env`:**
+
+```env
+# LiveKit connection
+LIVEKIT_URL=ws://livekit-server:7880
+LIVEKIT_API_KEY=$LK_KEY
+LIVEKIT_API_SECRET=$LK_SECRET
+LIVEKIT_PUBLIC_URL=wss://your-domain:7880    # or ws://your-ip:7880
+
+# LiveKit egress S3 storage (reuse transcript storage or configure separately)
+LIVEKIT_STORAGE_AWS_BUCKET_NAME=reflector-bucket
+LIVEKIT_STORAGE_AWS_REGION=us-east-1
+LIVEKIT_STORAGE_AWS_ACCESS_KEY_ID=your-key
+LIVEKIT_STORAGE_AWS_SECRET_ACCESS_KEY=your-secret
+LIVEKIT_STORAGE_AWS_ENDPOINT_URL=http://garage:3900   # for Garage/MinIO
+
+# Set LiveKit as default platform for new rooms
+DEFAULT_VIDEO_PLATFORM=livekit
+```
+
+3. **Create `livekit.yaml`:**
+
+```yaml
+port: 7880
+rtc:
+  tcp_port: 7881
+  port_range_start: 44200
+  port_range_end: 44300
+redis:
+  address: redis:6379
+keys:
+  your_api_key: your_api_secret
+webhook:
+  urls:
+    - http://server:1250/v1/livekit/webhook
+  api_key: your_api_key
+logging:
+  level: info
+room:
+  empty_timeout: 300
+  max_participants: 0
+```
+
+4. **Create `egress.yaml`:**
+
+```yaml
+api_key: your_api_key
+api_secret: your_api_secret
+ws_url: ws://livekit-server:7880
+health_port: 7082
+log_level: info
+session_limits:
+  file_output_max_duration: 4h
+```
+
+5. **Start with the livekit profile:**
+
+```bash
+docker compose -f docker-compose.selfhosted.yml --profile livekit up -d livekit-server livekit-egress
+```
+
+## Environment Variables Reference
+
+### Required
+
+| Variable | Description | Example |
+|----------|-------------|---------|
+| `LIVEKIT_URL` | Internal WebSocket URL (server -> LiveKit) | `ws://livekit-server:7880` |
+| `LIVEKIT_API_KEY` | API key for authentication | `reflector_a1b2c3d4e5f6` |
+| `LIVEKIT_API_SECRET` | API secret for token signing and webhooks | `64-char hex string` |
+
+### Recommended
+
+| Variable | Description | Example |
+|----------|-------------|---------|
+| `LIVEKIT_PUBLIC_URL` | Public WebSocket URL (browser -> LiveKit). **Must be reachable from participants' browsers**, not a Docker-internal address. Without `--domain`, set to `ws://<server-ip>:7880`. With `--domain`, set to `wss://<domain>:7880`. | `wss://reflector.example.com:7880` |
+| `LIVEKIT_WEBHOOK_SECRET` | Webhook verification secret. Defaults to `LIVEKIT_API_SECRET` if not set. Only needed if you want a separate secret for webhooks. | (same as API secret) |
+| `DEFAULT_VIDEO_PLATFORM` | Default platform for new rooms | `livekit` |
+
+### Storage (for Track Egress)
+
+Track Egress writes per-participant audio files to S3. If not configured, falls back to the transcript storage credentials.
+
+| Variable | Description | Example |
+|----------|-------------|---------|
+| `LIVEKIT_STORAGE_AWS_BUCKET_NAME` | S3 bucket for egress output | `reflector-bucket` |
+| `LIVEKIT_STORAGE_AWS_REGION` | S3 region | `us-east-1` |
+| `LIVEKIT_STORAGE_AWS_ACCESS_KEY_ID` | S3 access key | `GK...` |
+| `LIVEKIT_STORAGE_AWS_SECRET_ACCESS_KEY` | S3 secret key | `...` |
+| `LIVEKIT_STORAGE_AWS_ENDPOINT_URL` | S3 endpoint (for Garage/MinIO) | `http://garage:3900` |
+
+## Docker Compose Services
+
+Two services are added under the `livekit` profile in `docker-compose.selfhosted.yml`:
+
+### livekit-server
+
+| Setting | Value |
+|---------|-------|
+| Image | `livekit/livekit-server:v1.10.1` |
+| Ports | 7880 (signaling), 7881 (TCP RTC), 44200-44300/udp (ICE) |
+| Config | `./livekit.yaml` mounted at `/etc/livekit.yaml` |
+| Depends on | Redis |
+
+### livekit-egress
+
+| Setting | Value |
+|---------|-------|
+| Image | `livekit/egress:v1.10.1` |
+| Config | `./egress.yaml` mounted at `/etc/egress.yaml` |
+| Depends on | Redis, livekit-server |
+
+No `--cap-add=SYS_ADMIN` is needed because Track Egress does not use Chrome (that's only for Room Composite video recording, which we don't use).
+
+## Port Ranges
+
+| Range | Protocol | Service | Notes |
+|-------|----------|---------|-------|
+| 7880 | TCP | LiveKit signaling | WebSocket connections from browsers (direct, no Caddy) |
+| 7881 | TCP | LiveKit RTC over TCP | Fallback when UDP is blocked |
+| 44200-44300 | UDP | LiveKit ICE | WebRTC media. Avoids collision with Reflector WebRTC (40000-40100) and macOS ephemeral ports (49152-65535) |
+
+### TLS / Caddy Integration
+
+When `--caddy` is enabled (HTTPS), the setup script automatically:
+
+1. Adds a `/lk-ws` reverse proxy route to the Caddyfile that proxies `wss://domain/lk-ws` → `ws://livekit-server:7880`
+2. Sets `LIVEKIT_PUBLIC_URL` to `wss://<domain>/lk-ws` (or `wss://<ip>/lk-ws`)
+
+This avoids mixed-content blocking (browsers reject `ws://` connections on `https://` pages). Caddy handles TLS termination; LiveKit server itself runs plain WebSocket internally.
+
+Without `--caddy`, browsers connect directly to LiveKit on port 7880 via `ws://`.
+
+### Security Note: on_demand TLS
+
+When using `--ip` (Caddy with self-signed certs), the Caddyfile uses `tls internal { on_demand }`. This generates certificates dynamically for any hostname/IP on first TLS request.
+
+**Risk:** An attacker can trigger certificate generation for arbitrary hostnames by sending TLS requests with spoofed SNI values, causing disk and CPU usage. This is a low-severity resource exhaustion risk, not a data theft risk.
+
+**Mitigations:**
+- For LAN/development use: not a concern (not internet-exposed)
+- For cloud VMs: restrict port 443 access via firewall to trusted IPs
+- For production: use `--domain` with a real domain name instead of `--ip` — Caddy uses Let's Encrypt (no `on_demand` needed)
+
+| Deployment | `LIVEKIT_PUBLIC_URL` | How it works |
+|---|---|---|
+| localhost, no Caddy | `ws://localhost:7880` | Direct connection |
+| LAN IP, no Caddy | `ws://192.168.1.x:7880` | Direct connection |
+| IP + Caddy | `wss://192.168.1.x/lk-ws` | Caddy terminates TLS, proxies to LiveKit |
+| Domain + Caddy | `wss://example.com/lk-ws` | Caddy terminates TLS, proxies to LiveKit |
+
+## Webhook Endpoint
+
+LiveKit sends webhook events to `POST /v1/livekit/webhook`. Events handled:
+
+| Event | Action |
+|-------|--------|
+| `participant_joined` | Logs participant join, updates meeting state |
+| `participant_left` | Logs participant leave |
+| `egress_started` | Logs recording start |
+| `egress_ended` | Logs recording completion with output file info |
+| `room_started` / `room_finished` | Logs room lifecycle |
+
+Webhooks are authenticated via JWT in the `Authorization` header, verified using the API secret.
+
+## Frontend
+
+The LiveKit room component uses `@livekit/components-react` with the prebuilt `<VideoConference>` UI. It includes:
+
+- Recording consent dialog (same as Daily/Whereby)
+- Email transcript button (feature-gated)
+- Extensible overlay buttons for custom actions
+
+When a user joins a LiveKit room, the backend generates a JWT access token and returns it in the `room_url` query parameter. The frontend parses this and passes it to the LiveKit React SDK.
+
+## Separate Server Deployment
+
+For larger deployments (15+ participants, multiple simultaneous rooms), LiveKit can run on a dedicated server:
+
+1. Run `livekit-server` and `livekit-egress` on a separate machine
+2. Point `LIVEKIT_URL` to the remote LiveKit server (e.g., `ws://livekit-host:7880`)
+3. Set `LIVEKIT_PUBLIC_URL` to the public-facing URL (e.g., `wss://livekit.example.com`)
+4. Configure the remote LiveKit's `webhook.urls` to point back to the Reflector server
+5. Both need access to the same Redis (or configure LiveKit's own Redis)
+6. Both need access to the same S3 storage
+
+## Troubleshooting
+
+### LiveKit server not starting
+
+```bash
+# Check logs
+docker compose -f docker-compose.selfhosted.yml logs livekit-server --tail 30
+
+# Verify config
+cat livekit.yaml
+
+# Common issues:
+# - Redis not reachable (check redis service is running)
+# - Port 7880 already in use
+# - Invalid API key format in livekit.yaml
+```
+
+### Participants can't connect
+
+```bash
+# Check that LIVEKIT_PUBLIC_URL is accessible from the browser
+# It must be the URL the browser can reach, not the Docker-internal URL
+
+# Check firewall allows ports 7880, 7881, and 44200-44300/udp
+sudo ufw status  # or iptables -L
+
+# Verify the access token is being generated
+docker compose -f docker-compose.selfhosted.yml logs server | grep livekit
+```
+
+### Track Egress not writing files
+
+```bash
+# Check egress logs
+docker compose -f docker-compose.selfhosted.yml logs livekit-egress --tail 30
+
+# Verify S3 credentials
+# Egress receives S3 config per-request from the server, so check server/.env:
+grep LIVEKIT_STORAGE server/.env
+```

docsv2/migrate-daily-to-livekit.md

@@ -0,0 +1,73 @@
+# Migrating from Daily.co to LiveKit
+
+This guide covers running LiveKit alongside Daily.co or fully replacing it.
+
+## Both Platforms Run Simultaneously
+
+LiveKit and Daily.co coexist — the platform is selected **per room**. You don't need to migrate all rooms at once.
+
+- Existing Daily rooms continue to work as-is
+- New rooms can use LiveKit
+- Each room's `platform` field determines which video service is used
+- Transcripts, topics, summaries work identically regardless of platform
+
+## Step 1: Enable LiveKit
+
+Add `--livekit` to your setup command:
+
+```bash
+# If currently running:
+./scripts/setup-selfhosted.sh --gpu --ollama-gpu --garage --caddy
+
+# Add --livekit:
+./scripts/setup-selfhosted.sh --gpu --ollama-gpu --livekit --garage --caddy
+```
+
+This starts `livekit-server` + `livekit-egress` containers alongside your existing stack.
+
+## Step 2: Set Default Platform
+
+The setup script automatically sets `DEFAULT_VIDEO_PLATFORM=livekit` in `server/.env`. This means **new rooms** default to LiveKit. Existing rooms keep their current platform.
+
+To keep Daily as the default for new rooms:
+```bash
+# In server/.env, change:
+DEFAULT_VIDEO_PLATFORM=daily
+```
+
+## Step 3: Switch Individual Rooms
+
+In the Rooms admin page, edit any room and change the **Platform** dropdown from "Daily" to "LiveKit". The next meeting in that room will use LiveKit.
+
+Previously recorded Daily transcripts for that room are unaffected.
+
+## Step 4: (Optional) Remove Daily.co
+
+Once all rooms use LiveKit and you no longer need Daily.co:
+
+1. Remove `DAILY_API_KEY` and related Daily settings from `server/.env`
+2. Re-run the setup script — it won't activate the `dailyco` profile
+3. Hatchet workers are shared between Daily and LiveKit, so they continue running
+
+Daily-specific services that stop:
+- `hatchet-worker-cpu` with `dailyco` profile (but continues if `livekit` profile is active)
+- Daily webhook polling tasks (`poll_daily_recordings`, etc.)
+
+## What Changes for Users
+
+| Feature | Daily.co | LiveKit |
+|---------|---------|---------|
+| Video/audio quality | Daily.co SFU | LiveKit SFU (comparable) |
+| Pre-join screen | Daily's built-in iframe | LiveKit PreJoin component (name + device selection) |
+| Recording | Starts via REST API from frontend | Auto Track Egress (automatic, no user action) |
+| Multitrack audio | Per-participant WebM tracks | Per-participant OGG tracks |
+| Transcript quality | Same pipeline | Same pipeline |
+| Self-hosted | No (SaaS only) | Yes (fully self-hosted) |
+
+## Database Changes
+
+None required. The `platform` field on rooms and meetings already supports `"livekit"`. LiveKit recordings use recording IDs prefixed with `lk-` to distinguish them from Daily recordings.
+
+## Rollback
+
+To revert a room back to Daily, just change the Platform dropdown back to "Daily" in the Rooms admin page. No data migration needed.

docsv2/selfhosted-architecture.md

@@ -24,6 +24,8 @@ This document explains the internals of the self-hosted deployment: how the setu
 The self-hosted deployment runs the entire Reflector platform on a single server using Docker Compose. A single bash script (`scripts/setup-selfhosted.sh`) handles all configuration and orchestration. The key design principles are:
 
 - **One command to deploy** — flags select which features to enable
+- **Config memory** — CLI args are saved to `data/.selfhosted-last-args`; re-run with no flags to replay
+- **Per-service overrides** — individual ML backends (transcript, diarization, translation, padding, mixdown) can be overridden independently from the base mode
 - **Idempotent** — safe to re-run without losing existing configuration
 - **Profile-based composition** — Docker Compose profiles activate optional services
 - **No external dependencies required** — with `--garage` and `--ollama-*`, everything runs locally
@@ -61,8 +63,9 @@ Creates or updates the backend environment file from `server/.env.selfhosted.exa
 - **Infrastructure** — PostgreSQL URL, Redis host, Celery broker (all pointing to Docker-internal hostnames)
 - **Public URLs** — `BASE_URL` and `CORS_ORIGIN` computed from the domain (if `--domain`), IP (if detected on Linux), or `localhost`
 - **WebRTC** — `WEBRTC_HOST` set to the server's LAN IP so browsers can reach UDP ICE candidates
-- **Specialized models** — always points to `http://transcription:8000` (the Docker network alias shared by GPU and CPU containers)
-- **HuggingFace token** — prompts interactively for pyannote model access; writes to root `.env` so Docker Compose can inject it into GPU/CPU containers
+- **ML backends (per-service)** — Each ML service (transcript, diarization, translation, padding, mixdown) is configured independently using "effective backends" (`EFF_TRANSCRIPT`, `EFF_DIARIZATION`, `EFF_TRANSLATION`, `EFF_PADDING`, `EFF_MIXDOWN`). These are resolved from the base mode default + any `--transcript`/`--diarization`/`--translation`/`--padding`/`--mixdown` overrides. For `modal` backends, the URL is `http://transcription:8000` (GPU mode), user-provided (hosted mode), or read from existing env (CPU mode with override). For CPU backends, no URL is needed (in-process). If a service is overridden to `modal` in CPU mode without a URL configured, the script warns the user to set `TRANSCRIPT_URL` in `server/.env`
+- **CPU timeouts** — `TRANSCRIPT_FILE_TIMEOUT` and `DIARIZATION_FILE_TIMEOUT` are increased to 3600s only for services actually using CPU backends (whisper/pyannote), not blanket for the whole mode
+- **HuggingFace token** — prompted when diarization uses `pyannote` (in-process) or when GPU mode is active (GPU container needs it). Writes to root `.env` so Docker Compose can inject it into GPU/CPU containers
 - **LLM** — if `--ollama-*` is used, configures `LLM_URL` pointing to the Ollama container. Otherwise, warns that the user needs to configure an external LLM
 - **Public mode** — sets `PUBLIC_MODE=true` so the app is accessible without authentication by default
 - **Password auth** — if `--password` is passed, sets `AUTH_BACKEND=password`, `PUBLIC_MODE=false`, `ADMIN_EMAIL=admin@localhost`, and `ADMIN_PASSWORD_HASH` (the hash generated in Step 1). The admin user is provisioned in the database on container startup via `runserver.sh`
@@ -167,6 +170,8 @@ These start regardless of which flags you pass:
 | `ollama-cpu` | `ollama-cpu` | Local Ollama LLM on CPU |
 | `garage` | `garage` | Local S3-compatible object storage |
 | `caddy` | `caddy` | Reverse proxy with SSL |
+| `dailyco` | `hatchet-worker-cpu` | Hatchet workflow workers for Daily.co multitrack processing |
+| `livekit` | `livekit-server`, `livekit-egress` | Self-hosted video platform + per-participant audio recording |
 
 ### The "transcription" Alias
 
@@ -203,11 +208,17 @@ Both the `gpu` and `cpu` services define a Docker network alias of `transcriptio
       │ :8000     │  └─────────┘  └─────────┘
       └───────────┘
             │
-      ┌─────┴─────┐     ┌─────────┐
-      │  ollama   │     │ garage  │
-      │(optional) │     │(optional│
-      │ :11435    │     │  S3)    │
-      └───────────┘     └─────────┘
+      ┌─────┴─────┐     ┌─────────┐     ┌──────────────┐
+      │  ollama   │     │ garage  │     │livekit-server│
+      │(optional) │     │(optional│     │  (optional)  │
+      │ :11435    │     │  S3)    │     │  :7880       │
+      └───────────┘     └─────────┘     └──────┬───────┘
+                                               │
+                                        ┌──────┴───────┐
+                                        │livekit-egress│
+                                        │ (Track Egress│
+                                        │  to S3)      │
+                                        └──────────────┘
 ```
 
 ### How Services Interact
@@ -228,11 +239,19 @@ Both the `gpu` and `cpu` services define a Docker network alias of `transcriptio
 Environment variables flow through multiple layers. Understanding this prevents confusion when debugging:
 
 ```
-Flags (--gpu, --garage, etc.)
+CLI args (--gpu, --garage, --padding modal, --mixdown modal, etc.)
   │
-  ├── setup-selfhosted.sh interprets flags
+  ├── Config memory: saved to data/.selfhosted-last-args
+  │   (replayed on next run if no args provided)
+  │
+  ├── setup-selfhosted.sh resolves effective backends:
+  │     EFF_TRANSCRIPT  = override or base mode default
+  │     EFF_DIARIZATION = override or base mode default
+  │     EFF_TRANSLATION = override or base mode default
+  │     EFF_PADDING     = override or base mode default
+  │     EFF_MIXDOWN     = override or base mode default
   │     │
-  │     ├── Writes server/.env (backend config)
+  │     ├── Writes server/.env (backend config, per-service backends)
   │     ├── Writes www/.env (frontend config)
   │     ├── Writes .env (HF_TOKEN for compose interpolation)
   │     └── Writes Caddyfile (proxy routes)
@@ -309,7 +328,9 @@ You can point your own reverse proxy (nginx, Traefik, etc.) at these ports.
 
 ### WebRTC and UDP
 
-The server exposes UDP ports 50000-50100 for WebRTC ICE candidates. The `WEBRTC_HOST` variable tells the server which IP to advertise in ICE candidates — this must be the server's actual IP address (not a domain), because WebRTC uses UDP which doesn't go through the HTTP reverse proxy.
+The server exposes UDP ports 40000-40100 for Reflector's own WebRTC ICE candidates. When LiveKit is enabled, it additionally uses ports 44200-44300/udp for its WebRTC ICE candidates. The `WEBRTC_HOST` variable tells the server which IP to advertise in ICE candidates — this must be the server's actual IP address (not a domain), because WebRTC uses UDP which doesn't go through the HTTP reverse proxy.
+
+Port ranges are chosen to avoid collision with macOS ephemeral ports (49152-65535).
 
 ---
 
@@ -415,7 +436,10 @@ All services communicate over Docker's default bridge network. Only specific por
 | 3903 | Garage | `0.0.0.0:3903` | Garage admin API |
 | 8000 | GPU/CPU | `127.0.0.1:8000` | ML model API (localhost only) |
 | 11435 | Ollama | `127.0.0.1:11435` | Ollama API (localhost only) |
-| 50000-50100/udp | Server | `0.0.0.0:50000-50100` | WebRTC ICE candidates |
+| 40000-40100/udp | Server | `0.0.0.0:40000-40100` | Reflector WebRTC ICE candidates |
+| 7880 | LiveKit | `0.0.0.0:7880` | LiveKit signaling (WS) |
+| 7881 | LiveKit | `0.0.0.0:7881` | LiveKit RTC over TCP |
+| 44200-44300/udp | LiveKit | `0.0.0.0:44200-44300` | LiveKit WebRTC ICE candidates |
 
 Services bound to `127.0.0.1` are only accessible from the host itself (not from the network). Caddy is the only service exposed to the internet on standard HTTP/HTTPS ports.
 
@@ -432,6 +456,8 @@ Inside the Docker network, services reach each other by their compose service na
 | `transcription` | GPU or CPU container (network alias) |
 | `ollama` / `ollama-cpu` | Ollama container |
 | `garage` | Garage S3 container |
+| `livekit-server` | LiveKit SFU server |
+| `livekit-egress` | LiveKit Track Egress service |
 
 ---
 

docsv2/selfhosted-production.md

@@ -70,7 +70,7 @@ That's it. The script generates env files, secrets, starts all containers, waits
 
 ## ML Processing Modes (Required)
 
-Pick `--gpu`, `--cpu`, or `--hosted`. This determines how **transcription, diarization, translation, and audio padding** run:
+Pick `--gpu`, `--cpu`, or `--hosted`. This determines how **transcription, diarization, translation, audio padding, and audio mixdown** run:
 
 | Flag | What it does | Requires |
 |------|-------------|----------|
@@ -144,6 +144,7 @@ Browse all available models at https://ollama.com/library.
 
 | Flag | What it does |
 |------|-------------|
+| `--livekit` | Enables LiveKit self-hosted video platform. Generates API credentials, starts `livekit-server` + `livekit-egress`. See [LiveKit Setup](livekit-setup.md). |
 | `--garage` | Starts Garage (local S3-compatible storage). Auto-configures bucket, keys, and env vars. |
 | `--caddy` | Starts Caddy reverse proxy on ports 80/443 with self-signed cert. |
 | `--domain DOMAIN` | Use a real domain with Let's Encrypt auto-HTTPS (implies `--caddy`). Requires DNS A record pointing to this server and ports 80/443 open. |
@@ -154,10 +155,74 @@ Without `--garage`, you **must** provide S3-compatible credentials (the script w
 
 Without `--caddy` or `--domain`, no ports are exposed. Point your own reverse proxy at `web:3000` (frontend) and `server:1250` (API).
 
+## Video Platform (LiveKit)
+
+For self-hosted video rooms with per-participant audio recording, add `--livekit` to your setup command:
+
+```bash
+./scripts/setup-selfhosted.sh --gpu --ollama-gpu --livekit --garage --caddy
+```
+
+This generates LiveKit API credentials, creates config files (`livekit.yaml`, `egress.yaml`), and starts `livekit-server` (WebRTC SFU) + `livekit-egress` (per-participant audio recording to S3). LiveKit reuses the same Redis and S3 storage as the rest of the stack.
+
+New rooms default to LiveKit when `DEFAULT_VIDEO_PLATFORM=livekit` is set (done automatically by the setup script). Existing Daily.co and Whereby rooms continue to work. On re-runs, the script detects the existing `LIVEKIT_API_KEY` in `server/.env` automatically.
+
+> For detailed configuration, environment variables, ports, and troubleshooting, see [LiveKit Setup](livekit-setup.md).
+
 **Using a domain (recommended for production):** Point a DNS A record at your server's IP, then pass `--domain your.domain.com`. Caddy will automatically obtain and renew a Let's Encrypt certificate. Ports 80 and 443 must be open.
 
 **Without a domain:** `--caddy` alone uses a self-signed certificate. Browsers will show a security warning that must be accepted.
 
+## Per-Service Backend Overrides
+
+Override individual ML services without changing the base mode. Useful when you want most services on one backend but need specific services on another.
+
+| Flag | Valid backends | Default (`--gpu`/`--hosted`) | Default (`--cpu`) |
+|------|---------------|------------------------------|-------------------|
+| `--transcript BACKEND` | `whisper`, `modal` | `modal` | `whisper` |
+| `--diarization BACKEND` | `pyannote`, `modal` | `modal` | `pyannote` |
+| `--translation BACKEND` | `marian`, `modal`, `passthrough` | `modal` | `marian` |
+| `--padding BACKEND` | `pyav`, `modal` | `modal` | `pyav` |
+| `--mixdown BACKEND` | `pyav`, `modal` | `modal` | `pyav` |
+
+**Examples:**
+
+```bash
+# CPU base, but use a remote modal service for padding only
+./scripts/setup-selfhosted.sh --cpu --padding modal --garage --caddy
+
+# GPU base, but skip translation entirely (passthrough)
+./scripts/setup-selfhosted.sh --gpu --translation passthrough --garage --caddy
+
+# CPU base with remote modal diarization and translation
+./scripts/setup-selfhosted.sh --cpu --diarization modal --translation modal --garage
+```
+
+When overriding a service to `modal` in `--cpu` mode, the script will warn you to configure the service URL (`TRANSCRIPT_URL` etc.) in `server/.env` to point to your GPU service, then re-run.
+
+When overriding a service to a CPU backend (e.g., `--transcript whisper`) in `--gpu` mode, that service runs in-process on the server/worker containers while the GPU container still serves the remaining `modal` services.
+
+## Config Memory (No-Flag Re-run)
+
+After a successful run, the script saves your CLI arguments to `data/.selfhosted-last-args`. On subsequent runs with no arguments, the saved configuration is automatically replayed:
+
+```bash
+# First run — saves the config
+./scripts/setup-selfhosted.sh --gpu --ollama-gpu --garage --caddy
+
+# Later re-runs — same config, no flags needed
+./scripts/setup-selfhosted.sh
+# => "No flags provided — replaying saved configuration:"
+# => "  --gpu --ollama-gpu --garage --caddy"
+```
+
+To change the configuration, pass new flags — they override and replace the saved config:
+
+```bash
+# Switch to CPU mode with overrides — this becomes the new saved config
+./scripts/setup-selfhosted.sh --cpu --padding modal --garage --caddy
+```
+
 ## What the Script Does
 
 1. **Prerequisites check** — Docker, NVIDIA GPU (if needed), compose file exists
@@ -189,6 +254,8 @@ Without `--caddy` or `--domain`, no ports are exposed. Point your own reverse pr
 | `TRANSCRIPT_URL` | Specialized model endpoint | `http://transcription:8000` |
 | `PADDING_BACKEND` | Audio padding backend (`pyav` or `modal`) | `modal` (selfhosted), `pyav` (default) |
 | `PADDING_URL` | Audio padding endpoint (when `PADDING_BACKEND=modal`) | `http://transcription:8000` |
+| `MIXDOWN_BACKEND` | Audio mixdown backend (`pyav` or `modal`) | `modal` (selfhosted), `pyav` (default) |
+| `MIXDOWN_URL` | Audio mixdown endpoint (when `MIXDOWN_BACKEND=modal`) | `http://transcription:8000` |
 | `LLM_URL` | OpenAI-compatible LLM endpoint | Auto-set for Ollama modes |
 | `LLM_API_KEY` | LLM API key | `not-needed` for Ollama |
 | `LLM_MODEL` | LLM model name | `qwen2.5:14b` for Ollama (override with `--llm-model`) |
@@ -199,6 +266,11 @@ Without `--caddy` or `--domain`, no ports are exposed. Point your own reverse pr
 | `DAILY_SUBDOMAIN` | Daily.co subdomain | *(unset)* |
 | `DAILYCO_STORAGE_AWS_ACCESS_KEY_ID` | AWS access key for reading Daily's recording bucket | *(unset)* |
 | `DAILYCO_STORAGE_AWS_SECRET_ACCESS_KEY` | AWS secret key for reading Daily's recording bucket | *(unset)* |
+| `ZULIP_REALM` | Zulip server hostname (e.g. `zulip.example.com`) | *(unset)* |
+| `ZULIP_API_KEY` | Zulip bot API key | *(unset)* |
+| `ZULIP_BOT_EMAIL` | Zulip bot email address | *(unset)* |
+| `ZULIP_DAG_STREAM` | Zulip stream for pipeline failure alerts | *(unset)* |
+| `ZULIP_DAG_TOPIC` | Zulip topic for pipeline failure alerts | *(unset)* |
 | `HATCHET_CLIENT_TOKEN` | Hatchet API token (auto-generated) | *(unset)* |
 | `HATCHET_CLIENT_SERVER_URL` | Hatchet server URL | Auto-set when Daily.co configured |
 | `HATCHET_CLIENT_HOST_PORT` | Hatchet gRPC address | Auto-set when Daily.co configured |
@@ -248,6 +320,48 @@ TRANSCRIPT_STORAGE_AWS_REGION=us-east-1
 TRANSCRIPT_STORAGE_AWS_ENDPOINT_URL=http://minio:9000
 ```
 
+### S3 IAM Permissions Reference
+
+Reflector uses up to 3 separate S3 credential sets, each scoped to a specific bucket. When using AWS IAM in production, each key should have only the permissions it needs.
+
+**Transcript storage key** (`TRANSCRIPT_STORAGE_AWS_*`) — the main bucket for processed files:
+
+```json
+{
+  "Effect": "Allow",
+  "Action": ["s3:PutObject", "s3:GetObject", "s3:DeleteObject", "s3:ListBucket"],
+  "Resource": ["arn:aws:s3:::reflector-media/*", "arn:aws:s3:::reflector-media"]
+}
+```
+
+Used for: processed MP3 audio, waveform JSON, temporary pipeline files. Deletions happen during trash "Destroy", consent-denied cleanup, and public mode data retention.
+
+**Daily.co worker key** (`DAILYCO_STORAGE_AWS_ACCESS_KEY_ID/SECRET_ACCESS_KEY`) — for reading and cleaning up Daily recordings:
+
+```json
+{
+  "Effect": "Allow",
+  "Action": ["s3:GetObject", "s3:DeleteObject", "s3:ListBucket"],
+  "Resource": ["arn:aws:s3:::your-daily-bucket/*", "arn:aws:s3:::your-daily-bucket"]
+}
+```
+
+Used for: downloading multitrack recording files for processing, deleting track files and composed video on consent denial or trash destroy. No `s3:PutObject` needed — Daily's own API writes via the Role ARN.
+
+**Whereby worker key** (`WHEREBY_STORAGE_AWS_ACCESS_KEY_ID/SECRET_ACCESS_KEY`) — same pattern as Daily:
+
+```json
+{
+  "Effect": "Allow",
+  "Action": ["s3:GetObject", "s3:DeleteObject", "s3:ListBucket"],
+  "Resource": ["arn:aws:s3:::your-whereby-bucket/*", "arn:aws:s3:::your-whereby-bucket"]
+}
+```
+
+> **Fallback behavior:** If platform-specific worker keys are not set, Reflector falls back to the transcript storage master key with a bucket override. This means the master key would need cross-bucket access to the Daily/Whereby buckets. For least-privilege, configure platform-specific keys so each only accesses its own bucket.
+
+> **Garage / single-bucket setups:** When using Garage or a single S3 bucket for everything, one master key with full permissions on that bucket is sufficient. The IAM scoping above only matters when using separate buckets per platform (typical in AWS production).
+
 ## What Authentication Enables
 
 By default, Reflector runs in **public mode** (`AUTH_BACKEND=none`, `PUBLIC_MODE=true`) — anyone can create and view transcripts without logging in. Transcripts are anonymous (not linked to any user) and cannot be edited or deleted after creation.
@@ -571,9 +685,9 @@ docker compose -f docker-compose.selfhosted.yml exec gpu curl http://localhost:8
 ## Updating
 
 ```bash
-# Option A: Pull latest prebuilt images and restart
+# Option A: Pull latest prebuilt images and restart (replays saved config automatically)
 docker compose -f docker-compose.selfhosted.yml down
-./scripts/setup-selfhosted.sh <same-flags-as-before>
+./scripts/setup-selfhosted.sh
 
 # Option B: Build from source (after git pull) and restart
 git pull
@@ -584,6 +698,8 @@ docker compose -f docker-compose.selfhosted.yml down
 docker compose -f docker-compose.selfhosted.yml build gpu  # or cpu
 ```
 
+> **Note on config memory:** Running with no flags replays the saved config from your last run. Running with *any* flags replaces the saved config entirely — the script always saves the complete set of flags you provide. See [Config Memory](#config-memory-no-flag-re-run).
+
 The setup script is idempotent — it won't overwrite existing secrets or env vars that are already set.
 
 ## Architecture Overview

docsv2/tunnel-setup.md

@@ -0,0 +1,144 @@
+# Tunnel Setup (Self-Hosting Behind NAT)
+
+Expose your self-hosted Reflector + LiveKit stack to the internet without port forwarding, static IPs, or cloud VMs using tunneling services.
+
+## Requirements
+
+You need **two tunnels**:
+
+| Tunnel | Protocol | What it carries | Local port | Examples |
+|--------|----------|----------------|------------|----------|
+| **TCP tunnel** | TCP | Web app, API, LiveKit signaling (WebSocket) | 443 (Caddy) | playit.gg, ngrok, Cloudflare Tunnel, bore, frp |
+| **UDP tunnel** | UDP | WebRTC audio/video media | Assigned by tunnel service | playit.gg, frp |
+
+> **Important:** Most tunneling services only support TCP. WebRTC media requires UDP. Make sure your chosen service supports UDP tunnels. As of writing, [playit.gg](https://playit.gg) is one of the few that supports both TCP and UDP (premium $3/mo).
+
+## Architecture
+
+```
+Internet participants
+    │
+    ├── TCP tunnel (HTTPS)
+    │   └── tunnel service → your machine port 443 (Caddy)
+    │       ├── /v1/*          → server:1250 (API)
+    │       ├── /lk-ws/*       → livekit-server:7880 (signaling)
+    │       └── /*             → web:3000 (frontend)
+    │
+    └── UDP tunnel
+        └── tunnel service → your machine port N (LiveKit ICE)
+```
+
+## Setup
+
+### Step 1: Create tunnels with your chosen service
+
+Create two tunnels and note the public addresses:
+
+- **TCP tunnel**: Points to your local port `443`
+  - You'll get an address like `your-tunnel.example.com:PORT`
+- **UDP tunnel**: Points to a local port (e.g., `14139`)
+  - You'll get an address like `udp-host.example.com:PORT`
+  - **The local port must match the public port** (or LiveKit ICE candidates won't match). Set the local port to the same number as the public port assigned by the tunnel service.
+
+### Step 2: Run the setup script
+
+```bash
+./scripts/setup-selfhosted.sh <mode> --livekit --garage \
+  --tunnels <TCP_ADDRESS>,<UDP_ADDRESS>
+```
+
+Example:
+```bash
+./scripts/setup-selfhosted.sh --cpu --livekit --garage \
+  --tunnels my-tunnel.example.com:9055,udp-host.example.com:14139
+```
+
+Or use separate flags:
+```bash
+./scripts/setup-selfhosted.sh --cpu --livekit --garage \
+  --tunnel-tcp my-tunnel.example.com:9055 \
+  --tunnel-udp udp-host.example.com:14139
+```
+
+The script automatically:
+- Sets all URLs (API, frontend, LiveKit signaling) to the TCP tunnel address
+- Configures LiveKit with the UDP tunnel port and resolved IP for ICE candidates
+- Enables Caddy with self-signed TLS (catch-all on port 443)
+- Saves tunnel config for re-runs
+
+### Step 3: Start the tunnel agent
+
+Run your tunneling service's agent/client on the same machine. It must be running whenever you want external access.
+
+### Step 4: Access
+
+Share `https://<TCP_TUNNEL_ADDRESS>` with participants. They'll need to accept the self-signed certificate warning in their browser.
+
+## Flag Reference
+
+| Flag | Description |
+|------|-------------|
+| `--tunnels TCP,UDP` | Both tunnel addresses comma-separated (e.g., `host:9055,host:14139`) |
+| `--tunnel-tcp ADDR` | TCP tunnel address only (e.g., `host.example.com:9055`) |
+| `--tunnel-udp ADDR` | UDP tunnel address only (e.g., `host.example.com:14139`) |
+
+Tunnel flags:
+- Imply `--caddy` (HTTPS required for browser mic/camera access)
+- Are mutually exclusive with `--ip` and `--domain`
+- Are saved to config memory (re-run without flags replays saved config)
+
+## UDP Port Matching
+
+LiveKit advertises ICE candidates with a specific IP and port. The browser connects to that exact address. If the tunnel's public port differs from the local port, ICE will fail.
+
+**Correct setup:** Set the tunnel's local port to match its public port.
+
+```
+Tunnel assigns public port 14139
+  → Set local port to 14139
+  → LiveKit listens on 14139 (udp_port in livekit.yaml)
+  → Docker maps 14139:14139/udp
+  → ICE candidates advertise tunnel_ip:14139
+  → Browser connects to tunnel_ip:14139 → tunnel → local:14139 → LiveKit
+```
+
+If your tunneling service doesn't let you choose the local port, you'll need to update `livekit.yaml` manually with the assigned ports.
+
+## TLS Certificate Warning
+
+With tunnel services on non-standard ports (e.g., `:9055`), Let's Encrypt can't auto-provision certificates (it requires ports 80/443). Caddy uses `tls internal` which generates a self-signed certificate. Participants will see a browser warning they must accept.
+
+**To avoid the warning:**
+- Use a tunnel service that provides port 443 for TCP
+- Or use a real domain with `--domain` on a server with a public IP
+
+## Compatible Tunnel Services
+
+| Service | TCP | UDP | Free tier | Notes |
+|---------|-----|-----|-----------|-------|
+| [playit.gg](https://playit.gg) | Yes (premium) | Yes (premium) | Limited | $3/mo premium. Supports both TCP + UDP. |
+| [ngrok](https://ngrok.com) | Yes | No | Limited | TCP only — needs a separate UDP tunnel for media |
+| [Cloudflare Tunnel](https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/) | Yes | No | Yes | TCP only — needs a separate UDP tunnel for media |
+| [bore](https://github.com/ekzhang/bore) | Yes | No | Self-hosted | TCP only |
+| [frp](https://github.com/fatedier/frp) | Yes | Yes | Self-hosted | Requires your own VPS to run the frp server |
+| [Tailscale Funnel](https://tailscale.com/kb/1223/funnel) | Yes | No | Free (3 nodes) | TCP only, requires Tailscale account |
+
+For a full self-contained setup without a VPS, playit.gg (TCP + UDP) is currently the simplest option.
+
+## Limitations
+
+- **Latency**: Adds a hop through the tunnel service's relay servers
+- **Bandwidth**: Tunnel services may have bandwidth limits on free/cheap tiers
+- **Reliability**: Depends on the tunnel service's uptime
+- **Certificate warning**: Unavoidable with non-standard ports (see above)
+- **Single UDP port**: Tunnel mode uses a single UDP port instead of a range, which limits concurrent WebRTC connections (~50 participants max)
+- **Not production-grade**: Suitable for demos, small teams, development, and privacy-first setups. For production, use a server with a public IP.
+
+## Comparison
+
+| Approach | Cost | Setup | Data location | Port forwarding needed |
+|----------|------|-------|---------------|----------------------|
+| **Tunnel (this guide)** | $0-3/mo | Low | Your machine | No |
+| **Cloud VM** | $5-20/mo | Low | Cloud provider | No |
+| **Port forwarding** | $0 | Medium | Your machine | Yes (router config) |
+| **VPN mesh (Tailscale)** | $0 | Low | Your machine | No (VPN peers only) |

egress.yaml.example

@@ -0,0 +1,26 @@
+# LiveKit Egress configuration
+# Generated by setup-selfhosted.sh — do not edit manually.
+# See: https://docs.livekit.io/self-hosting/egress/
+
+api_key: __LIVEKIT_API_KEY__
+api_secret: __LIVEKIT_API_SECRET__
+ws_url: ws://livekit-server:7880
+redis:
+  address: redis:6379
+
+# Health check
+health_port: 7082
+
+# Logging
+log_level: info
+
+# CPU cost limits (Track Egress only — no composite video)
+# Track Egress costs 1.0 CPU unit per track; hundreds can run on one instance.
+# Default max_cpu_utilization is 0.8 (80% of available cores).
+
+# Session limits
+session_limits:
+  file_output_max_duration: 4h   # Max 4 hours per recording
+
+# S3 storage is configured per-request via the API (not here).
+# The server passes S3 credentials when starting each Track Egress.

gpu/modal_deployments/deploy-all.sh

@@ -114,8 +114,8 @@ modal secret create reflector-gpu REFLECTOR_GPU_APIKEY="$API_KEY"
 
 # --- Deploy Functions ---
 echo ""
-echo "Deploying transcriber (Whisper)..."
-TRANSCRIBER_URL=$(modal deploy reflector_transcriber.py 2>&1 | grep -o 'https://[^ ]*web.modal.run' | head -1)
+echo "Deploying transcriber (Parakeet)..."
+TRANSCRIBER_URL=$(modal deploy reflector_transcriber_parakeet.py 2>&1 | grep -o 'https://[^ ]*web.modal.run' | head -1)
 if [ -z "$TRANSCRIBER_URL" ]; then
     echo "Error: Failed to deploy transcriber. Check Modal dashboard for details."
     exit 1
@@ -132,13 +132,22 @@ fi
 echo "  -> $DIARIZER_URL"
 
 echo ""
-echo "Deploying padding (CPU audio processing via Modal SDK)..."
-modal deploy reflector_padding.py
-if [ $? -ne 0 ]; then
+echo "Deploying padding (CPU audio processing)..."
+PADDING_URL=$(modal deploy reflector_padding.py 2>&1 | grep -o 'https://[^ ]*web.modal.run' | head -1)
+if [ -z "$PADDING_URL" ]; then
     echo "Error: Failed to deploy padding. Check Modal dashboard for details."
     exit 1
 fi
-echo "  -> reflector-padding.pad_track (Modal SDK function)"
+echo "  -> $PADDING_URL"
+
+echo ""
+echo "Deploying mixdown (CPU multi-track audio mixing)..."
+MIXDOWN_URL=$(modal deploy reflector_mixdown.py 2>&1 | grep -o 'https://[^ ]*web.modal.run' | head -1)
+if [ -z "$MIXDOWN_URL" ]; then
+    echo "Error: Failed to deploy mixdown. Check Modal dashboard for details."
+    exit 1
+fi
+echo "  -> $MIXDOWN_URL"
 
 # --- Output Configuration ---
 echo ""
@@ -157,5 +166,11 @@ echo "DIARIZATION_BACKEND=modal"
 echo "DIARIZATION_URL=$DIARIZER_URL"
 echo "DIARIZATION_MODAL_API_KEY=$API_KEY"
 echo ""
-echo "# Padding uses Modal SDK (requires MODAL_TOKEN_ID/SECRET in worker containers)"
+echo "PADDING_BACKEND=modal"
+echo "PADDING_URL=$PADDING_URL"
+echo "PADDING_MODAL_API_KEY=$API_KEY"
+echo ""
+echo "MIXDOWN_BACKEND=modal"
+echo "MIXDOWN_URL=$MIXDOWN_URL"
+echo "MIXDOWN_MODAL_API_KEY=$API_KEY"
 echo "# --- End Modal Configuration ---"

gpu/modal_deployments/reflector_diarizer.py

@@ -113,12 +113,14 @@ def download_pyannote_audio():
 
 
 diarizer_image = (
-    modal.Image.debian_slim(python_version="3.10")
+    modal.Image.from_registry(
+        "nvidia/cuda:11.8.0-cudnn8-devel-ubuntu22.04", add_python="3.10"
+    )
     .pip_install(
         "pyannote.audio==3.1.0",
         "requests",
         "onnx",
-        "torchaudio",
+        "torchaudio==2.0.1",
         "onnxruntime-gpu",
         "torch==2.0.0",
         "transformers==4.34.0",
@@ -133,14 +135,6 @@ diarizer_image = (
         secrets=[modal.Secret.from_name("hf_token")],
     )
     .run_function(migrate_cache_llm)
-    .env(
-        {
-            "LD_LIBRARY_PATH": (
-                "/usr/local/lib/python3.10/site-packages/nvidia/cudnn/lib/:"
-                "/opt/conda/lib/python3.10/site-packages/nvidia/cublas/lib/"
-            )
-        }
-    )
 )
 
 

gpu/modal_deployments/reflector_mixdown.py

@@ -0,0 +1,385 @@
+"""
+Reflector GPU backend - audio mixdown
+=====================================
+
+CPU-intensive multi-track audio mixdown service.
+Mixes N audio tracks into a single MP3 using PyAV amix filter graph.
+
+IMPORTANT: This mixdown logic is duplicated from server/reflector/utils/audio_mixdown.py
+for Modal deployment isolation (Modal can't import from server/reflector/). If you modify
+the PyAV filter graph or mixdown algorithm, you MUST update both:
+  - gpu/modal_deployments/reflector_mixdown.py (this file)
+  - server/reflector/utils/audio_mixdown.py
+
+Constants duplicated from server/reflector/utils/audio_constants.py for same reason.
+"""
+
+import os
+import tempfile
+from fractions import Fraction
+import asyncio
+
+import modal
+
+S3_TIMEOUT = 120  # Higher than padding (60s) — multiple track downloads
+MIXDOWN_TIMEOUT = 1200 + (S3_TIMEOUT * 2)  # 1440s total
+SCALEDOWN_WINDOW = 60
+DISCONNECT_CHECK_INTERVAL = 2
+
+app = modal.App("reflector-mixdown")
+
+# CPU-based image (mixdown is CPU-bound, no GPU needed)
+image = (
+    modal.Image.debian_slim(python_version="3.12")
+    .apt_install("ffmpeg")  # Required by PyAV
+    .pip_install(
+        "av==13.1.0",  # PyAV for audio processing
+        "requests==2.32.3",  # HTTP for presigned URL downloads/uploads
+        "fastapi==0.115.12",  # API framework
+    )
+)
+
+
+@app.function(
+    cpu=4.0,  # Higher than padding (2.0) for multi-track mixing
+    timeout=MIXDOWN_TIMEOUT,
+    scaledown_window=SCALEDOWN_WINDOW,
+    image=image,
+    secrets=[modal.Secret.from_name("reflector-gpu")],
+)
+@modal.asgi_app()
+def web():
+    from fastapi import Depends, FastAPI, HTTPException, Request, status
+    from fastapi.security import OAuth2PasswordBearer
+    from pydantic import BaseModel
+
+    class MixdownRequest(BaseModel):
+        track_urls: list[str]
+        output_url: str
+        target_sample_rate: int | None = None
+        offsets_seconds: list[float] | None = None
+
+    class MixdownResponse(BaseModel):
+        size: int
+        duration_ms: float = 0.0
+        cancelled: bool = False
+
+    web_app = FastAPI()
+
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+        if apikey == os.environ["REFLECTOR_GPU_APIKEY"]:
+            return
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail="Invalid API key",
+            headers={"WWW-Authenticate": "Bearer"},
+        )
+
+    @web_app.post("/mixdown", dependencies=[Depends(apikey_auth)])
+    async def mixdown_endpoint(request: Request, req: MixdownRequest) -> MixdownResponse:
+        """Modal web endpoint for mixing audio tracks with disconnect detection."""
+        import logging
+        import threading
+
+        logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
+        logger = logging.getLogger(__name__)
+
+        valid_urls = [u for u in req.track_urls if u]
+        if not valid_urls:
+            raise HTTPException(status_code=400, detail="No valid track URLs provided")
+        if req.offsets_seconds is not None:
+            if len(req.offsets_seconds) != len(req.track_urls):
+                raise HTTPException(
+                    status_code=400,
+                    detail=f"offsets_seconds length ({len(req.offsets_seconds)}) "
+                    f"must match track_urls ({len(req.track_urls)})",
+                )
+            if any(o > 18000 for o in req.offsets_seconds):
+                raise HTTPException(status_code=400, detail="offsets_seconds exceeds maximum 18000s (5 hours)")
+        if not req.output_url:
+            raise HTTPException(status_code=400, detail="output_url cannot be empty")
+
+        logger.info(f"Mixdown request: {len(valid_urls)} tracks")
+
+        # Thread-safe cancellation flag
+        cancelled = threading.Event()
+
+        async def check_disconnect():
+            """Background task to check for client disconnect."""
+            while not cancelled.is_set():
+                await asyncio.sleep(DISCONNECT_CHECK_INTERVAL)
+                if await request.is_disconnected():
+                    logger.warning("Client disconnected, setting cancellation flag")
+                    cancelled.set()
+                    break
+
+        disconnect_task = asyncio.create_task(check_disconnect())
+
+        try:
+            result = await asyncio.get_event_loop().run_in_executor(
+                None, _mixdown_tracks_blocking, req, cancelled, logger
+            )
+            return MixdownResponse(**result)
+        finally:
+            cancelled.set()
+            disconnect_task.cancel()
+            try:
+                await disconnect_task
+            except asyncio.CancelledError:
+                pass
+
+    def _mixdown_tracks_blocking(req, cancelled, logger) -> dict:
+        """Blocking CPU-bound mixdown work with periodic cancellation checks.
+
+        Downloads all tracks, builds PyAV amix filter graph, encodes to MP3,
+        and uploads the result to the presigned output URL.
+        """
+        import av
+        import requests
+        from av.audio.resampler import AudioResampler
+        import time
+
+        temp_dir = tempfile.mkdtemp()
+        track_paths = []
+        output_path = None
+        last_check = time.time()
+
+        try:
+            # --- Download all tracks ---
+            valid_urls = [u for u in req.track_urls if u]
+            for i, url in enumerate(valid_urls):
+                if cancelled.is_set():
+                    logger.info("Cancelled during download phase")
+                    return {"size": 0, "duration_ms": 0.0, "cancelled": True}
+
+                logger.info(f"Downloading track {i}")
+                response = requests.get(url, stream=True, timeout=S3_TIMEOUT)
+                response.raise_for_status()
+
+                track_path = os.path.join(temp_dir, f"track_{i}.webm")
+                total_bytes = 0
+                chunk_count = 0
+                with open(track_path, "wb") as f:
+                    for chunk in response.iter_content(chunk_size=8192):
+                        if chunk:
+                            f.write(chunk)
+                            total_bytes += len(chunk)
+                            chunk_count += 1
+                            if chunk_count % 12 == 0:
+                                now = time.time()
+                                if now - last_check >= DISCONNECT_CHECK_INTERVAL:
+                                    if cancelled.is_set():
+                                        logger.info(f"Cancelled during track {i} download")
+                                        return {"size": 0, "duration_ms": 0.0, "cancelled": True}
+                                    last_check = now
+
+                track_paths.append(track_path)
+                logger.info(f"Track {i} downloaded: {total_bytes} bytes")
+
+            if not track_paths:
+                raise ValueError("No tracks downloaded")
+
+            # --- Detect sample rate ---
+            target_sample_rate = req.target_sample_rate
+            if target_sample_rate is None:
+                for path in track_paths:
+                    try:
+                        container = av.open(path)
+                        for frame in container.decode(audio=0):
+                            target_sample_rate = frame.sample_rate
+                            container.close()
+                            break
+                        else:
+                            container.close()
+                            continue
+                        break
+                    except Exception:
+                        continue
+            if target_sample_rate is None:
+                raise ValueError("Could not detect sample rate from any track")
+
+            logger.info(f"Target sample rate: {target_sample_rate}")
+
+            # --- Calculate per-input delays ---
+            input_offsets_seconds = None
+            if req.offsets_seconds is not None:
+                input_offsets_seconds = [
+                    req.offsets_seconds[i] for i, url in enumerate(req.track_urls) if url
+                ]
+
+            delays_ms = []
+            if input_offsets_seconds is not None:
+                base = min(input_offsets_seconds) if input_offsets_seconds else 0.0
+                delays_ms = [max(0, int(round((o - base) * 1000))) for o in input_offsets_seconds]
+            else:
+                delays_ms = [0 for _ in track_paths]
+
+            # --- Build filter graph ---
+            # N abuffer -> optional adelay -> amix -> aformat -> abuffersink
+            graph = av.filter.Graph()
+            inputs = []
+
+            for idx in range(len(track_paths)):
+                args = (
+                    f"time_base=1/{target_sample_rate}:"
+                    f"sample_rate={target_sample_rate}:"
+                    f"sample_fmt=s32:"
+                    f"channel_layout=stereo"
+                )
+                in_ctx = graph.add("abuffer", args=args, name=f"in{idx}")
+                inputs.append(in_ctx)
+
+            mixer = graph.add("amix", args=f"inputs={len(inputs)}:normalize=0", name="mix")
+            fmt = graph.add(
+                "aformat",
+                args=f"sample_fmts=s32:channel_layouts=stereo:sample_rates={target_sample_rate}",
+                name="fmt",
+            )
+            sink = graph.add("abuffersink", name="out")
+
+            for idx, in_ctx in enumerate(inputs):
+                delay_ms = delays_ms[idx] if idx < len(delays_ms) else 0
+                if delay_ms > 0:
+                    adelay = graph.add(
+                        "adelay",
+                        args=f"delays={delay_ms}|{delay_ms}:all=1",
+                        name=f"delay{idx}",
+                    )
+                    in_ctx.link_to(adelay)
+                    adelay.link_to(mixer, 0, idx)
+                else:
+                    in_ctx.link_to(mixer, 0, idx)
+
+            mixer.link_to(fmt)
+            fmt.link_to(sink)
+            graph.configure()
+
+            # --- Open all containers and decode ---
+            containers = []
+            output_path = os.path.join(temp_dir, "mixed.mp3")
+
+            try:
+                for path in track_paths:
+                    containers.append(av.open(path))
+
+                decoders = [c.decode(audio=0) for c in containers]
+                active = [True] * len(decoders)
+                resamplers = [
+                    AudioResampler(format="s32", layout="stereo", rate=target_sample_rate)
+                    for _ in decoders
+                ]
+
+                # Open output MP3
+                out_container = av.open(output_path, "w", format="mp3")
+                out_stream = out_container.add_stream("libmp3lame", rate=target_sample_rate)
+                total_duration = 0
+
+                while any(active):
+                    # Check cancellation periodically
+                    now = time.time()
+                    if now - last_check >= DISCONNECT_CHECK_INTERVAL:
+                        if cancelled.is_set():
+                            logger.info("Cancelled during mixing")
+                            out_container.close()
+                            return {"size": 0, "duration_ms": 0.0, "cancelled": True}
+                        last_check = now
+
+                    for i, (dec, is_active) in enumerate(zip(decoders, active)):
+                        if not is_active:
+                            continue
+                        try:
+                            frame = next(dec)
+                        except StopIteration:
+                            active[i] = False
+                            inputs[i].push(None)
+                            continue
+
+                        if frame.sample_rate != target_sample_rate:
+                            continue
+
+                        out_frames = resamplers[i].resample(frame) or []
+                        for rf in out_frames:
+                            rf.sample_rate = target_sample_rate
+                            rf.time_base = Fraction(1, target_sample_rate)
+                            inputs[i].push(rf)
+
+                        while True:
+                            try:
+                                mixed = sink.pull()
+                            except Exception:
+                                break
+                            mixed.sample_rate = target_sample_rate
+                            mixed.time_base = Fraction(1, target_sample_rate)
+                            for packet in out_stream.encode(mixed):
+                                out_container.mux(packet)
+                                total_duration += packet.duration
+
+                # Flush filter graph
+                while True:
+                    try:
+                        mixed = sink.pull()
+                    except Exception:
+                        break
+                    mixed.sample_rate = target_sample_rate
+                    mixed.time_base = Fraction(1, target_sample_rate)
+                    for packet in out_stream.encode(mixed):
+                        out_container.mux(packet)
+                        total_duration += packet.duration
+
+                # Flush encoder
+                for packet in out_stream.encode(None):
+                    out_container.mux(packet)
+                    total_duration += packet.duration
+
+                # Calculate duration in ms
+                last_tb = out_stream.time_base
+                duration_ms = 0.0
+                if last_tb and total_duration > 0:
+                    duration_ms = round(float(total_duration * last_tb * 1000), 2)
+
+                out_container.close()
+
+            finally:
+                for c in containers:
+                    try:
+                        c.close()
+                    except Exception:
+                        pass
+
+            file_size = os.path.getsize(output_path)
+            logger.info(f"Mixdown complete: {file_size} bytes, {duration_ms}ms")
+
+            if cancelled.is_set():
+                logger.info("Cancelled after mixing, before upload")
+                return {"size": 0, "duration_ms": 0.0, "cancelled": True}
+
+            # --- Upload result ---
+            logger.info("Uploading mixed audio to S3")
+            with open(output_path, "rb") as f:
+                upload_response = requests.put(req.output_url, data=f, timeout=S3_TIMEOUT)
+            upload_response.raise_for_status()
+            logger.info(f"Upload complete: {file_size} bytes")
+
+            return {"size": file_size, "duration_ms": duration_ms}
+
+        finally:
+            # Cleanup all temp files
+            for path in track_paths:
+                if os.path.exists(path):
+                    try:
+                        os.unlink(path)
+                    except Exception as e:
+                        logger.warning(f"Failed to cleanup track file: {e}")
+            if output_path and os.path.exists(output_path):
+                try:
+                    os.unlink(output_path)
+                except Exception as e:
+                    logger.warning(f"Failed to cleanup output file: {e}")
+            try:
+                os.rmdir(temp_dir)
+            except Exception as e:
+                logger.warning(f"Failed to cleanup temp directory: {e}")
+
+    return web_app

gpu/modal_deployments/reflector_padding.py

@@ -52,10 +52,12 @@ OPUS_DEFAULT_BIT_RATE = 128000
     timeout=PADDING_TIMEOUT,
     scaledown_window=SCALEDOWN_WINDOW,
     image=image,
+    secrets=[modal.Secret.from_name("reflector-gpu")],
 )
 @modal.asgi_app()
 def web():
-    from fastapi import FastAPI, Request, HTTPException
+    from fastapi import Depends, FastAPI, HTTPException, Request, status
+    from fastapi.security import OAuth2PasswordBearer
     from pydantic import BaseModel
 
     class PaddingRequest(BaseModel):
@@ -70,7 +72,18 @@ def web():
 
     web_app = FastAPI()
 
-    @web_app.post("/pad")
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+        if apikey == os.environ["REFLECTOR_GPU_APIKEY"]:
+            return
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail="Invalid API key",
+            headers={"WWW-Authenticate": "Bearer"},
+        )
+
+    @web_app.post("/pad", dependencies=[Depends(apikey_auth)])
     async def pad_track_endpoint(request: Request, req: PaddingRequest) -> PaddingResponse:
         """Modal web endpoint for padding audio tracks with disconnect detection.
         """

gpu/self_hosted/Dockerfile

@@ -42,6 +42,7 @@ COPY pyproject.toml uv.lock /app/
 COPY ./app /app/app
 COPY ./main.py /app/
 COPY ./runserver.sh /app/
+COPY ./docker-entrypoint.sh /app/
 
 # prevent uv failing with too many open files on big cpus
 ENV UV_CONCURRENT_INSTALLS=16
@@ -52,6 +53,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 EXPOSE 8000
 
-CMD ["sh", "/app/runserver.sh"]
+RUN chmod +x /app/docker-entrypoint.sh
+
+CMD ["sh", "/app/docker-entrypoint.sh"]
 
 

gpu/self_hosted/Dockerfile.cpu

@@ -26,6 +26,7 @@ COPY pyproject.toml uv.lock /app/
 COPY ./app /app/app
 COPY ./main.py /app/
 COPY ./runserver.sh /app/
+COPY ./docker-entrypoint.sh /app/
 
 # prevent uv failing with too many open files on big cpus
 ENV UV_CONCURRENT_INSTALLS=16
@@ -36,4 +37,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 EXPOSE 8000
 
-CMD ["sh", "/app/runserver.sh"]
+RUN chmod +x /app/docker-entrypoint.sh
+
+CMD ["sh", "/app/docker-entrypoint.sh"]

gpu/self_hosted/app/factory.py

@@ -3,6 +3,7 @@ from contextlib import asynccontextmanager
 from fastapi import FastAPI
 
 from .routers.diarization import router as diarization_router
+from .routers.mixdown import router as mixdown_router
 from .routers.padding import router as padding_router
 from .routers.transcription import router as transcription_router
 from .routers.translation import router as translation_router
@@ -29,4 +30,5 @@ def create_app() -> FastAPI:
     app.include_router(translation_router)
     app.include_router(diarization_router)
     app.include_router(padding_router)
+    app.include_router(mixdown_router)
     return app

gpu/self_hosted/app/routers/mixdown.py

@@ -0,0 +1,288 @@
+"""
+Audio mixdown endpoint for selfhosted GPU service.
+
+CPU-intensive multi-track audio mixing service for combining N audio tracks
+into a single MP3 using PyAV amix filter graph.
+
+IMPORTANT: This mixdown logic is duplicated from server/reflector/utils/audio_mixdown.py
+for deployment isolation (self_hosted can't import from server/reflector/). If you modify
+the PyAV filter graph or mixdown algorithm, you MUST update both:
+  - gpu/self_hosted/app/routers/mixdown.py (this file)
+  - server/reflector/utils/audio_mixdown.py
+
+Constants duplicated from server/reflector/utils/audio_constants.py for same reason.
+"""
+
+import logging
+import os
+import tempfile
+from fractions import Fraction
+
+import av
+import requests
+from av.audio.resampler import AudioResampler
+from fastapi import APIRouter, Depends, HTTPException
+from pydantic import BaseModel
+
+from ..auth import apikey_auth
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(tags=["mixdown"])
+
+S3_TIMEOUT = 120
+
+
+class MixdownRequest(BaseModel):
+    track_urls: list[str]
+    output_url: str
+    target_sample_rate: int | None = None
+    offsets_seconds: list[float] | None = None
+
+
+class MixdownResponse(BaseModel):
+    size: int
+    duration_ms: float = 0.0
+    cancelled: bool = False
+
+
+@router.post("/mixdown", dependencies=[Depends(apikey_auth)], response_model=MixdownResponse)
+def mixdown_tracks(req: MixdownRequest):
+    """Mix multiple audio tracks into single MP3 using PyAV amix filter graph."""
+    valid_urls = [u for u in req.track_urls if u]
+    if not valid_urls:
+        raise HTTPException(status_code=400, detail="No valid track URLs provided")
+    if req.offsets_seconds is not None:
+        if len(req.offsets_seconds) != len(req.track_urls):
+            raise HTTPException(
+                status_code=400,
+                detail=f"offsets_seconds length ({len(req.offsets_seconds)}) "
+                f"must match track_urls ({len(req.track_urls)})",
+            )
+        if any(o > 18000 for o in req.offsets_seconds):
+            raise HTTPException(
+                status_code=400, detail="offsets_seconds exceeds maximum 18000s (5 hours)"
+            )
+    if not req.output_url:
+        raise HTTPException(status_code=400, detail="output_url cannot be empty")
+
+    logger.info("Mixdown request: %d tracks", len(valid_urls))
+
+    temp_dir = tempfile.mkdtemp()
+    track_paths = []
+    output_path = None
+
+    try:
+        # --- Download all tracks ---
+        for i, url in enumerate(valid_urls):
+            logger.info("Downloading track %d", i)
+            response = requests.get(url, stream=True, timeout=S3_TIMEOUT)
+            response.raise_for_status()
+
+            track_path = os.path.join(temp_dir, f"track_{i}.webm")
+            total_bytes = 0
+            with open(track_path, "wb") as f:
+                for chunk in response.iter_content(chunk_size=8192):
+                    if chunk:
+                        f.write(chunk)
+                        total_bytes += len(chunk)
+
+            track_paths.append(track_path)
+            logger.info("Track %d downloaded: %d bytes", i, total_bytes)
+
+        if not track_paths:
+            raise HTTPException(status_code=400, detail="No tracks could be downloaded")
+
+        # --- Detect sample rate ---
+        target_sample_rate = req.target_sample_rate
+        if target_sample_rate is None:
+            for path in track_paths:
+                try:
+                    container = av.open(path)
+                    for frame in container.decode(audio=0):
+                        target_sample_rate = frame.sample_rate
+                        container.close()
+                        break
+                    else:
+                        container.close()
+                        continue
+                    break
+                except Exception:
+                    continue
+        if target_sample_rate is None:
+            raise HTTPException(
+                status_code=400, detail="Could not detect sample rate from any track"
+            )
+
+        logger.info("Target sample rate: %d", target_sample_rate)
+
+        # --- Calculate per-input delays ---
+        input_offsets_seconds = None
+        if req.offsets_seconds is not None:
+            input_offsets_seconds = [
+                req.offsets_seconds[i] for i, url in enumerate(req.track_urls) if url
+            ]
+
+        delays_ms = []
+        if input_offsets_seconds is not None:
+            base = min(input_offsets_seconds) if input_offsets_seconds else 0.0
+            delays_ms = [max(0, int(round((o - base) * 1000))) for o in input_offsets_seconds]
+        else:
+            delays_ms = [0 for _ in track_paths]
+
+        # --- Build filter graph ---
+        # N abuffer -> optional adelay -> amix -> aformat -> abuffersink
+        graph = av.filter.Graph()
+        inputs = []
+
+        for idx in range(len(track_paths)):
+            args = (
+                f"time_base=1/{target_sample_rate}:"
+                f"sample_rate={target_sample_rate}:"
+                f"sample_fmt=s32:"
+                f"channel_layout=stereo"
+            )
+            in_ctx = graph.add("abuffer", args=args, name=f"in{idx}")
+            inputs.append(in_ctx)
+
+        mixer = graph.add("amix", args=f"inputs={len(inputs)}:normalize=0", name="mix")
+        fmt = graph.add(
+            "aformat",
+            args=f"sample_fmts=s32:channel_layouts=stereo:sample_rates={target_sample_rate}",
+            name="fmt",
+        )
+        sink = graph.add("abuffersink", name="out")
+
+        for idx, in_ctx in enumerate(inputs):
+            delay_ms = delays_ms[idx] if idx < len(delays_ms) else 0
+            if delay_ms > 0:
+                adelay = graph.add(
+                    "adelay",
+                    args=f"delays={delay_ms}|{delay_ms}:all=1",
+                    name=f"delay{idx}",
+                )
+                in_ctx.link_to(adelay)
+                adelay.link_to(mixer, 0, idx)
+            else:
+                in_ctx.link_to(mixer, 0, idx)
+
+        mixer.link_to(fmt)
+        fmt.link_to(sink)
+        graph.configure()
+
+        # --- Open all containers and decode ---
+        containers = []
+        output_path = os.path.join(temp_dir, "mixed.mp3")
+
+        try:
+            for path in track_paths:
+                containers.append(av.open(path))
+
+            decoders = [c.decode(audio=0) for c in containers]
+            active = [True] * len(decoders)
+            resamplers = [
+                AudioResampler(format="s32", layout="stereo", rate=target_sample_rate)
+                for _ in decoders
+            ]
+
+            # Open output MP3
+            out_container = av.open(output_path, "w", format="mp3")
+            out_stream = out_container.add_stream("libmp3lame", rate=target_sample_rate)
+            total_duration = 0
+
+            while any(active):
+                for i, (dec, is_active) in enumerate(zip(decoders, active)):
+                    if not is_active:
+                        continue
+                    try:
+                        frame = next(dec)
+                    except StopIteration:
+                        active[i] = False
+                        inputs[i].push(None)
+                        continue
+
+                    if frame.sample_rate != target_sample_rate:
+                        continue
+
+                    out_frames = resamplers[i].resample(frame) or []
+                    for rf in out_frames:
+                        rf.sample_rate = target_sample_rate
+                        rf.time_base = Fraction(1, target_sample_rate)
+                        inputs[i].push(rf)
+
+                    while True:
+                        try:
+                            mixed = sink.pull()
+                        except Exception:
+                            break
+                        mixed.sample_rate = target_sample_rate
+                        mixed.time_base = Fraction(1, target_sample_rate)
+                        for packet in out_stream.encode(mixed):
+                            out_container.mux(packet)
+                            total_duration += packet.duration
+
+            # Flush filter graph
+            while True:
+                try:
+                    mixed = sink.pull()
+                except Exception:
+                    break
+                mixed.sample_rate = target_sample_rate
+                mixed.time_base = Fraction(1, target_sample_rate)
+                for packet in out_stream.encode(mixed):
+                    out_container.mux(packet)
+                    total_duration += packet.duration
+
+            # Flush encoder
+            for packet in out_stream.encode(None):
+                out_container.mux(packet)
+                total_duration += packet.duration
+
+            # Calculate duration in ms
+            last_tb = out_stream.time_base
+            duration_ms = 0.0
+            if last_tb and total_duration > 0:
+                duration_ms = round(float(total_duration * last_tb * 1000), 2)
+
+            out_container.close()
+
+        finally:
+            for c in containers:
+                try:
+                    c.close()
+                except Exception:
+                    pass
+
+        file_size = os.path.getsize(output_path)
+        logger.info("Mixdown complete: %d bytes, %.2fms", file_size, duration_ms)
+
+        # --- Upload result ---
+        logger.info("Uploading mixed audio to S3")
+        with open(output_path, "rb") as f:
+            upload_response = requests.put(req.output_url, data=f, timeout=S3_TIMEOUT)
+        upload_response.raise_for_status()
+        logger.info("Upload complete: %d bytes", file_size)
+
+        return MixdownResponse(size=file_size, duration_ms=duration_ms)
+
+    except HTTPException:
+        raise
+    except Exception as e:
+        logger.error("Mixdown failed: %s", e, exc_info=True)
+        raise HTTPException(status_code=500, detail=f"Mixdown failed: {e}") from e
+    finally:
+        for path in track_paths:
+            if os.path.exists(path):
+                try:
+                    os.unlink(path)
+                except Exception as e:
+                    logger.warning("Failed to cleanup track file: %s", e)
+        if output_path and os.path.exists(output_path):
+            try:
+                os.unlink(output_path)
+            except Exception as e:
+                logger.warning("Failed to cleanup output file: %s", e)
+        try:
+            os.rmdir(temp_dir)
+        except Exception as e:
+            logger.warning("Failed to cleanup temp directory: %s", e)

gpu/self_hosted/docker-entrypoint.sh

@@ -0,0 +1,23 @@
+#!/bin/sh
+set -e
+
+# Custom CA certificate injection
+# If a CA cert is mounted at this path (via docker-compose.ca.yml),
+# add it to the system trust store and configure all Python SSL libraries.
+CUSTOM_CA_PATH="/usr/local/share/ca-certificates/custom-ca.crt"
+
+if [ -s "$CUSTOM_CA_PATH" ]; then
+    echo "[entrypoint] Custom CA certificate detected, updating trust store..."
+    update-ca-certificates 2>/dev/null
+
+    # update-ca-certificates creates a combined bundle (system + custom CAs)
+    COMBINED_BUNDLE="/etc/ssl/certs/ca-certificates.crt"
+    export SSL_CERT_FILE="$COMBINED_BUNDLE"
+    export REQUESTS_CA_BUNDLE="$COMBINED_BUNDLE"
+    export CURL_CA_BUNDLE="$COMBINED_BUNDLE"
+    # Note: GRPC_DEFAULT_SSL_ROOTS_FILE_PATH is intentionally NOT set here.
+    # Setting it causes grpcio to attempt TLS on connections that may be plaintext.
+    echo "[entrypoint] CA trust store updated (SSL_CERT_FILE=$COMBINED_BUNDLE)"
+fi
+
+exec sh /app/runserver.sh

gpu/self_hosted/uv.lock

@@ -13,7 +13,7 @@ wheels = [
 
 [[package]]
 name = "aiohttp"
-version = "3.13.3"
+version = "3.13.4"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "aiohappyeyeballs" },
@@ -24,76 +24,76 @@ dependencies = [
     { name = "propcache" },
     { name = "yarl" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/50/42/32cf8e7704ceb4481406eb87161349abb46a57fee3f008ba9cb610968646/aiohttp-3.13.3.tar.gz", hash = "sha256:a949eee43d3782f2daae4f4a2819b2cb9b0c5d3b7f7a927067cc84dafdbb9f88", size = 7844556, upload-time = "2026-01-03T17:33:05.204Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/45/4a/064321452809dae953c1ed6e017504e72551a26b6f5708a5a80e4bf556ff/aiohttp-3.13.4.tar.gz", hash = "sha256:d97a6d09c66087890c2ab5d49069e1e570583f7ac0314ecf98294c1b6aaebd38", size = 7859748, upload-time = "2026-03-28T17:19:40.6Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/a0/be/4fc11f202955a69e0db803a12a062b8379c970c7c84f4882b6da17337cc1/aiohttp-3.13.3-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:b903a4dfee7d347e2d87697d0713be59e0b87925be030c9178c5faa58ea58d5c", size = 739732, upload-time = "2026-01-03T17:30:14.23Z" },
-    { url = "https://files.pythonhosted.org/packages/97/2c/621d5b851f94fa0bb7430d6089b3aa970a9d9b75196bc93bb624b0db237a/aiohttp-3.13.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:a45530014d7a1e09f4a55f4f43097ba0fd155089372e105e4bff4ca76cb1b168", size = 494293, upload-time = "2026-01-03T17:30:15.96Z" },
-    { url = "https://files.pythonhosted.org/packages/5d/43/4be01406b78e1be8320bb8316dc9c42dbab553d281c40364e0f862d5661c/aiohttp-3.13.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:27234ef6d85c914f9efeb77ff616dbf4ad2380be0cda40b4db086ffc7ddd1b7d", size = 493533, upload-time = "2026-01-03T17:30:17.431Z" },
-    { url = "https://files.pythonhosted.org/packages/8d/a8/5a35dc56a06a2c90d4742cbf35294396907027f80eea696637945a106f25/aiohttp-3.13.3-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d32764c6c9aafb7fb55366a224756387cd50bfa720f32b88e0e6fa45b27dcf29", size = 1737839, upload-time = "2026-01-03T17:30:19.422Z" },
-    { url = "https://files.pythonhosted.org/packages/bf/62/4b9eeb331da56530bf2e198a297e5303e1c1ebdceeb00fe9b568a65c5a0c/aiohttp-3.13.3-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:b1a6102b4d3ebc07dad44fbf07b45bb600300f15b552ddf1851b5390202ea2e3", size = 1703932, upload-time = "2026-01-03T17:30:21.756Z" },
-    { url = "https://files.pythonhosted.org/packages/7c/f6/af16887b5d419e6a367095994c0b1332d154f647e7dc2bd50e61876e8e3d/aiohttp-3.13.3-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:c014c7ea7fb775dd015b2d3137378b7be0249a448a1612268b5a90c2d81de04d", size = 1771906, upload-time = "2026-01-03T17:30:23.932Z" },
-    { url = "https://files.pythonhosted.org/packages/ce/83/397c634b1bcc24292fa1e0c7822800f9f6569e32934bdeef09dae7992dfb/aiohttp-3.13.3-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2b8d8ddba8f95ba17582226f80e2de99c7a7948e66490ef8d947e272a93e9463", size = 1871020, upload-time = "2026-01-03T17:30:26Z" },
-    { url = "https://files.pythonhosted.org/packages/86/f6/a62cbbf13f0ac80a70f71b1672feba90fdb21fd7abd8dbf25c0105fb6fa3/aiohttp-3.13.3-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:9ae8dd55c8e6c4257eae3a20fd2c8f41edaea5992ed67156642493b8daf3cecc", size = 1755181, upload-time = "2026-01-03T17:30:27.554Z" },
-    { url = "https://files.pythonhosted.org/packages/0a/87/20a35ad487efdd3fba93d5843efdfaa62d2f1479eaafa7453398a44faf13/aiohttp-3.13.3-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:01ad2529d4b5035578f5081606a465f3b814c542882804e2e8cda61adf5c71bf", size = 1561794, upload-time = "2026-01-03T17:30:29.254Z" },
-    { url = "https://files.pythonhosted.org/packages/de/95/8fd69a66682012f6716e1bc09ef8a1a2a91922c5725cb904689f112309c4/aiohttp-3.13.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:bb4f7475e359992b580559e008c598091c45b5088f28614e855e42d39c2f1033", size = 1697900, upload-time = "2026-01-03T17:30:31.033Z" },
-    { url = "https://files.pythonhosted.org/packages/e5/66/7b94b3b5ba70e955ff597672dad1691333080e37f50280178967aff68657/aiohttp-3.13.3-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:c19b90316ad3b24c69cd78d5c9b4f3aa4497643685901185b65166293d36a00f", size = 1728239, upload-time = "2026-01-03T17:30:32.703Z" },
-    { url = "https://files.pythonhosted.org/packages/47/71/6f72f77f9f7d74719692ab65a2a0252584bf8d5f301e2ecb4c0da734530a/aiohttp-3.13.3-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:96d604498a7c782cb15a51c406acaea70d8c027ee6b90c569baa6e7b93073679", size = 1740527, upload-time = "2026-01-03T17:30:34.695Z" },
-    { url = "https://files.pythonhosted.org/packages/fa/b4/75ec16cbbd5c01bdaf4a05b19e103e78d7ce1ef7c80867eb0ace42ff4488/aiohttp-3.13.3-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:084911a532763e9d3dd95adf78a78f4096cd5f58cdc18e6fdbc1b58417a45423", size = 1554489, upload-time = "2026-01-03T17:30:36.864Z" },
-    { url = "https://files.pythonhosted.org/packages/52/8f/bc518c0eea29f8406dcf7ed1f96c9b48e3bc3995a96159b3fc11f9e08321/aiohttp-3.13.3-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:7a4a94eb787e606d0a09404b9c38c113d3b099d508021faa615d70a0131907ce", size = 1767852, upload-time = "2026-01-03T17:30:39.433Z" },
-    { url = "https://files.pythonhosted.org/packages/9d/f2/a07a75173124f31f11ea6f863dc44e6f09afe2bca45dd4e64979490deab1/aiohttp-3.13.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:87797e645d9d8e222e04160ee32aa06bc5c163e8499f24db719e7852ec23093a", size = 1722379, upload-time = "2026-01-03T17:30:41.081Z" },
-    { url = "https://files.pythonhosted.org/packages/3c/4a/1a3fee7c21350cac78e5c5cef711bac1b94feca07399f3d406972e2d8fcd/aiohttp-3.13.3-cp312-cp312-win32.whl", hash = "sha256:b04be762396457bef43f3597c991e192ee7da460a4953d7e647ee4b1c28e7046", size = 428253, upload-time = "2026-01-03T17:30:42.644Z" },
-    { url = "https://files.pythonhosted.org/packages/d9/b7/76175c7cb4eb73d91ad63c34e29fc4f77c9386bba4a65b53ba8e05ee3c39/aiohttp-3.13.3-cp312-cp312-win_amd64.whl", hash = "sha256:e3531d63d3bdfa7e3ac5e9b27b2dd7ec9df3206a98e0b3445fa906f233264c57", size = 455407, upload-time = "2026-01-03T17:30:44.195Z" },
-    { url = "https://files.pythonhosted.org/packages/97/8a/12ca489246ca1faaf5432844adbfce7ff2cc4997733e0af120869345643a/aiohttp-3.13.3-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:5dff64413671b0d3e7d5918ea490bdccb97a4ad29b3f311ed423200b2203e01c", size = 734190, upload-time = "2026-01-03T17:30:45.832Z" },
-    { url = "https://files.pythonhosted.org/packages/32/08/de43984c74ed1fca5c014808963cc83cb00d7bb06af228f132d33862ca76/aiohttp-3.13.3-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:87b9aab6d6ed88235aa2970294f496ff1a1f9adcd724d800e9b952395a80ffd9", size = 491783, upload-time = "2026-01-03T17:30:47.466Z" },
-    { url = "https://files.pythonhosted.org/packages/17/f8/8dd2cf6112a5a76f81f81a5130c57ca829d101ad583ce57f889179accdda/aiohttp-3.13.3-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:425c126c0dc43861e22cb1c14ba4c8e45d09516d0a3ae0a3f7494b79f5f233a3", size = 490704, upload-time = "2026-01-03T17:30:49.373Z" },
-    { url = "https://files.pythonhosted.org/packages/6d/40/a46b03ca03936f832bc7eaa47cfbb1ad012ba1be4790122ee4f4f8cba074/aiohttp-3.13.3-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:7f9120f7093c2a32d9647abcaf21e6ad275b4fbec5b55969f978b1a97c7c86bf", size = 1720652, upload-time = "2026-01-03T17:30:50.974Z" },
-    { url = "https://files.pythonhosted.org/packages/f7/7e/917fe18e3607af92657e4285498f500dca797ff8c918bd7d90b05abf6c2a/aiohttp-3.13.3-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:697753042d57f4bf7122cab985bf15d0cef23c770864580f5af4f52023a56bd6", size = 1692014, upload-time = "2026-01-03T17:30:52.729Z" },
-    { url = "https://files.pythonhosted.org/packages/71/b6/cefa4cbc00d315d68973b671cf105b21a609c12b82d52e5d0c9ae61d2a09/aiohttp-3.13.3-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:6de499a1a44e7de70735d0b39f67c8f25eb3d91eb3103be99ca0fa882cdd987d", size = 1759777, upload-time = "2026-01-03T17:30:54.537Z" },
-    { url = "https://files.pythonhosted.org/packages/fb/e3/e06ee07b45e59e6d81498b591fc589629be1553abb2a82ce33efe2a7b068/aiohttp-3.13.3-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:37239e9f9a7ea9ac5bf6b92b0260b01f8a22281996da609206a84df860bc1261", size = 1861276, upload-time = "2026-01-03T17:30:56.512Z" },
-    { url = "https://files.pythonhosted.org/packages/7c/24/75d274228acf35ceeb2850b8ce04de9dd7355ff7a0b49d607ee60c29c518/aiohttp-3.13.3-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f76c1e3fe7d7c8afad7ed193f89a292e1999608170dcc9751a7462a87dfd5bc0", size = 1743131, upload-time = "2026-01-03T17:30:58.256Z" },
-    { url = "https://files.pythonhosted.org/packages/04/98/3d21dde21889b17ca2eea54fdcff21b27b93f45b7bb94ca029c31ab59dc3/aiohttp-3.13.3-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:fc290605db2a917f6e81b0e1e0796469871f5af381ce15c604a3c5c7e51cb730", size = 1556863, upload-time = "2026-01-03T17:31:00.445Z" },
-    { url = "https://files.pythonhosted.org/packages/9e/84/da0c3ab1192eaf64782b03971ab4055b475d0db07b17eff925e8c93b3aa5/aiohttp-3.13.3-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:4021b51936308aeea0367b8f006dc999ca02bc118a0cc78c303f50a2ff6afb91", size = 1682793, upload-time = "2026-01-03T17:31:03.024Z" },
-    { url = "https://files.pythonhosted.org/packages/ff/0f/5802ada182f575afa02cbd0ec5180d7e13a402afb7c2c03a9aa5e5d49060/aiohttp-3.13.3-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:49a03727c1bba9a97d3e93c9f93ca03a57300f484b6e935463099841261195d3", size = 1716676, upload-time = "2026-01-03T17:31:04.842Z" },
-    { url = "https://files.pythonhosted.org/packages/3f/8c/714d53bd8b5a4560667f7bbbb06b20c2382f9c7847d198370ec6526af39c/aiohttp-3.13.3-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:3d9908a48eb7416dc1f4524e69f1d32e5d90e3981e4e37eb0aa1cd18f9cfa2a4", size = 1733217, upload-time = "2026-01-03T17:31:06.868Z" },
-    { url = "https://files.pythonhosted.org/packages/7d/79/e2176f46d2e963facea939f5be2d26368ce543622be6f00a12844d3c991f/aiohttp-3.13.3-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:2712039939ec963c237286113c68dbad80a82a4281543f3abf766d9d73228998", size = 1552303, upload-time = "2026-01-03T17:31:08.958Z" },
-    { url = "https://files.pythonhosted.org/packages/ab/6a/28ed4dea1759916090587d1fe57087b03e6c784a642b85ef48217b0277ae/aiohttp-3.13.3-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:7bfdc049127717581866fa4708791220970ce291c23e28ccf3922c700740fdc0", size = 1763673, upload-time = "2026-01-03T17:31:10.676Z" },
-    { url = "https://files.pythonhosted.org/packages/e8/35/4a3daeb8b9fab49240d21c04d50732313295e4bd813a465d840236dd0ce1/aiohttp-3.13.3-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:8057c98e0c8472d8846b9c79f56766bcc57e3e8ac7bfd510482332366c56c591", size = 1721120, upload-time = "2026-01-03T17:31:12.575Z" },
-    { url = "https://files.pythonhosted.org/packages/bc/9f/d643bb3c5fb99547323e635e251c609fbbc660d983144cfebec529e09264/aiohttp-3.13.3-cp313-cp313-win32.whl", hash = "sha256:1449ceddcdbcf2e0446957863af03ebaaa03f94c090f945411b61269e2cb5daf", size = 427383, upload-time = "2026-01-03T17:31:14.382Z" },
-    { url = "https://files.pythonhosted.org/packages/4e/f1/ab0395f8a79933577cdd996dd2f9aa6014af9535f65dddcf88204682fe62/aiohttp-3.13.3-cp313-cp313-win_amd64.whl", hash = "sha256:693781c45a4033d31d4187d2436f5ac701e7bbfe5df40d917736108c1cc7436e", size = 453899, upload-time = "2026-01-03T17:31:15.958Z" },
-    { url = "https://files.pythonhosted.org/packages/99/36/5b6514a9f5d66f4e2597e40dea2e3db271e023eb7a5d22defe96ba560996/aiohttp-3.13.3-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:ea37047c6b367fd4bd632bff8077449b8fa034b69e812a18e0132a00fae6e808", size = 737238, upload-time = "2026-01-03T17:31:17.909Z" },
-    { url = "https://files.pythonhosted.org/packages/f7/49/459327f0d5bcd8c6c9ca69e60fdeebc3622861e696490d8674a6d0cb90a6/aiohttp-3.13.3-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:6fc0e2337d1a4c3e6acafda6a78a39d4c14caea625124817420abceed36e2415", size = 492292, upload-time = "2026-01-03T17:31:19.919Z" },
-    { url = "https://files.pythonhosted.org/packages/e8/0b/b97660c5fd05d3495b4eb27f2d0ef18dc1dc4eff7511a9bf371397ff0264/aiohttp-3.13.3-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:c685f2d80bb67ca8c3837823ad76196b3694b0159d232206d1e461d3d434666f", size = 493021, upload-time = "2026-01-03T17:31:21.636Z" },
-    { url = "https://files.pythonhosted.org/packages/54/d4/438efabdf74e30aeceb890c3290bbaa449780583b1270b00661126b8aae4/aiohttp-3.13.3-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:48e377758516d262bde50c2584fc6c578af272559c409eecbdd2bae1601184d6", size = 1717263, upload-time = "2026-01-03T17:31:23.296Z" },
-    { url = "https://files.pythonhosted.org/packages/71/f2/7bddc7fd612367d1459c5bcf598a9e8f7092d6580d98de0e057eb42697ad/aiohttp-3.13.3-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:34749271508078b261c4abb1767d42b8d0c0cc9449c73a4df494777dc55f0687", size = 1669107, upload-time = "2026-01-03T17:31:25.334Z" },
-    { url = "https://files.pythonhosted.org/packages/00/5a/1aeaecca40e22560f97610a329e0e5efef5e0b5afdf9f857f0d93839ab2e/aiohttp-3.13.3-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:82611aeec80eb144416956ec85b6ca45a64d76429c1ed46ae1b5f86c6e0c9a26", size = 1760196, upload-time = "2026-01-03T17:31:27.394Z" },
-    { url = "https://files.pythonhosted.org/packages/f8/f8/0ff6992bea7bd560fc510ea1c815f87eedd745fe035589c71ce05612a19a/aiohttp-3.13.3-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:2fff83cfc93f18f215896e3a190e8e5cb413ce01553901aca925176e7568963a", size = 1843591, upload-time = "2026-01-03T17:31:29.238Z" },
-    { url = "https://files.pythonhosted.org/packages/e3/d1/e30e537a15f53485b61f5be525f2157da719819e8377298502aebac45536/aiohttp-3.13.3-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bbe7d4cecacb439e2e2a8a1a7b935c25b812af7a5fd26503a66dadf428e79ec1", size = 1720277, upload-time = "2026-01-03T17:31:31.053Z" },
-    { url = "https://files.pythonhosted.org/packages/84/45/23f4c451d8192f553d38d838831ebbc156907ea6e05557f39563101b7717/aiohttp-3.13.3-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:b928f30fe49574253644b1ca44b1b8adbd903aa0da4b9054a6c20fc7f4092a25", size = 1548575, upload-time = "2026-01-03T17:31:32.87Z" },
-    { url = "https://files.pythonhosted.org/packages/6a/ed/0a42b127a43712eda7807e7892c083eadfaf8429ca8fb619662a530a3aab/aiohttp-3.13.3-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:7b5e8fe4de30df199155baaf64f2fcd604f4c678ed20910db8e2c66dc4b11603", size = 1679455, upload-time = "2026-01-03T17:31:34.76Z" },
-    { url = "https://files.pythonhosted.org/packages/2e/b5/c05f0c2b4b4fe2c9d55e73b6d3ed4fd6c9dc2684b1d81cbdf77e7fad9adb/aiohttp-3.13.3-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:8542f41a62bcc58fc7f11cf7c90e0ec324ce44950003feb70640fc2a9092c32a", size = 1687417, upload-time = "2026-01-03T17:31:36.699Z" },
-    { url = "https://files.pythonhosted.org/packages/c9/6b/915bc5dad66aef602b9e459b5a973529304d4e89ca86999d9d75d80cbd0b/aiohttp-3.13.3-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:5e1d8c8b8f1d91cd08d8f4a3c2b067bfca6ec043d3ff36de0f3a715feeedf926", size = 1729968, upload-time = "2026-01-03T17:31:38.622Z" },
-    { url = "https://files.pythonhosted.org/packages/11/3b/e84581290a9520024a08640b63d07673057aec5ca548177a82026187ba73/aiohttp-3.13.3-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:90455115e5da1c3c51ab619ac57f877da8fd6d73c05aacd125c5ae9819582aba", size = 1545690, upload-time = "2026-01-03T17:31:40.57Z" },
-    { url = "https://files.pythonhosted.org/packages/f5/04/0c3655a566c43fd647c81b895dfe361b9f9ad6d58c19309d45cff52d6c3b/aiohttp-3.13.3-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:042e9e0bcb5fba81886c8b4fbb9a09d6b8a00245fd8d88e4d989c1f96c74164c", size = 1746390, upload-time = "2026-01-03T17:31:42.857Z" },
-    { url = "https://files.pythonhosted.org/packages/1f/53/71165b26978f719c3419381514c9690bd5980e764a09440a10bb816ea4ab/aiohttp-3.13.3-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:2eb752b102b12a76ca02dff751a801f028b4ffbbc478840b473597fc91a9ed43", size = 1702188, upload-time = "2026-01-03T17:31:44.984Z" },
-    { url = "https://files.pythonhosted.org/packages/29/a7/cbe6c9e8e136314fa1980da388a59d2f35f35395948a08b6747baebb6aa6/aiohttp-3.13.3-cp314-cp314-win32.whl", hash = "sha256:b556c85915d8efaed322bf1bdae9486aa0f3f764195a0fb6ee962e5c71ef5ce1", size = 433126, upload-time = "2026-01-03T17:31:47.463Z" },
-    { url = "https://files.pythonhosted.org/packages/de/56/982704adea7d3b16614fc5936014e9af85c0e34b58f9046655817f04306e/aiohttp-3.13.3-cp314-cp314-win_amd64.whl", hash = "sha256:9bf9f7a65e7aa20dd764151fb3d616c81088f91f8df39c3893a536e279b4b984", size = 459128, upload-time = "2026-01-03T17:31:49.2Z" },
-    { url = "https://files.pythonhosted.org/packages/6c/2a/3c79b638a9c3d4658d345339d22070241ea341ed4e07b5ac60fb0f418003/aiohttp-3.13.3-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:05861afbbec40650d8a07ea324367cb93e9e8cc7762e04dd4405df99fa65159c", size = 769512, upload-time = "2026-01-03T17:31:51.134Z" },
-    { url = "https://files.pythonhosted.org/packages/29/b9/3e5014d46c0ab0db8707e0ac2711ed28c4da0218c358a4e7c17bae0d8722/aiohttp-3.13.3-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:2fc82186fadc4a8316768d61f3722c230e2c1dcab4200d52d2ebdf2482e47592", size = 506444, upload-time = "2026-01-03T17:31:52.85Z" },
-    { url = "https://files.pythonhosted.org/packages/90/03/c1d4ef9a054e151cd7839cdc497f2638f00b93cbe8043983986630d7a80c/aiohttp-3.13.3-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:0add0900ff220d1d5c5ebbf99ed88b0c1bbf87aa7e4262300ed1376a6b13414f", size = 510798, upload-time = "2026-01-03T17:31:54.91Z" },
-    { url = "https://files.pythonhosted.org/packages/ea/76/8c1e5abbfe8e127c893fe7ead569148a4d5a799f7cf958d8c09f3eedf097/aiohttp-3.13.3-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:568f416a4072fbfae453dcf9a99194bbb8bdeab718e08ee13dfa2ba0e4bebf29", size = 1868835, upload-time = "2026-01-03T17:31:56.733Z" },
-    { url = "https://files.pythonhosted.org/packages/8e/ac/984c5a6f74c363b01ff97adc96a3976d9c98940b8969a1881575b279ac5d/aiohttp-3.13.3-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:add1da70de90a2569c5e15249ff76a631ccacfe198375eead4aadf3b8dc849dc", size = 1720486, upload-time = "2026-01-03T17:31:58.65Z" },
-    { url = "https://files.pythonhosted.org/packages/b2/9a/b7039c5f099c4eb632138728828b33428585031a1e658d693d41d07d89d1/aiohttp-3.13.3-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:10b47b7ba335d2e9b1239fa571131a87e2d8ec96b333e68b2a305e7a98b0bae2", size = 1847951, upload-time = "2026-01-03T17:32:00.989Z" },
-    { url = "https://files.pythonhosted.org/packages/3c/02/3bec2b9a1ba3c19ff89a43a19324202b8eb187ca1e928d8bdac9bbdddebd/aiohttp-3.13.3-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:3dd4dce1c718e38081c8f35f323209d4c1df7d4db4bab1b5c88a6b4d12b74587", size = 1941001, upload-time = "2026-01-03T17:32:03.122Z" },
-    { url = "https://files.pythonhosted.org/packages/37/df/d879401cedeef27ac4717f6426c8c36c3091c6e9f08a9178cc87549c537f/aiohttp-3.13.3-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:34bac00a67a812570d4a460447e1e9e06fae622946955f939051e7cc895cfab8", size = 1797246, upload-time = "2026-01-03T17:32:05.255Z" },
-    { url = "https://files.pythonhosted.org/packages/8d/15/be122de1f67e6953add23335c8ece6d314ab67c8bebb3f181063010795a7/aiohttp-3.13.3-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:a19884d2ee70b06d9204b2727a7b9f983d0c684c650254679e716b0b77920632", size = 1627131, upload-time = "2026-01-03T17:32:07.607Z" },
-    { url = "https://files.pythonhosted.org/packages/12/12/70eedcac9134cfa3219ab7af31ea56bc877395b1ac30d65b1bc4b27d0438/aiohttp-3.13.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:5f8ca7f2bb6ba8348a3614c7918cc4bb73268c5ac2a207576b7afea19d3d9f64", size = 1795196, upload-time = "2026-01-03T17:32:09.59Z" },
-    { url = "https://files.pythonhosted.org/packages/32/11/b30e1b1cd1f3054af86ebe60df96989c6a414dd87e27ad16950eee420bea/aiohttp-3.13.3-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:b0d95340658b9d2f11d9697f59b3814a9d3bb4b7a7c20b131df4bcef464037c0", size = 1782841, upload-time = "2026-01-03T17:32:11.445Z" },
-    { url = "https://files.pythonhosted.org/packages/88/0d/d98a9367b38912384a17e287850f5695c528cff0f14f791ce8ee2e4f7796/aiohttp-3.13.3-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:a1e53262fd202e4b40b70c3aff944a8155059beedc8a89bba9dc1f9ef06a1b56", size = 1795193, upload-time = "2026-01-03T17:32:13.705Z" },
-    { url = "https://files.pythonhosted.org/packages/43/a5/a2dfd1f5ff5581632c7f6a30e1744deda03808974f94f6534241ef60c751/aiohttp-3.13.3-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:d60ac9663f44168038586cab2157e122e46bdef09e9368b37f2d82d354c23f72", size = 1621979, upload-time = "2026-01-03T17:32:15.965Z" },
-    { url = "https://files.pythonhosted.org/packages/fa/f0/12973c382ae7c1cccbc4417e129c5bf54c374dfb85af70893646e1f0e749/aiohttp-3.13.3-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:90751b8eed69435bac9ff4e3d2f6b3af1f57e37ecb0fbeee59c0174c9e2d41df", size = 1822193, upload-time = "2026-01-03T17:32:18.219Z" },
-    { url = "https://files.pythonhosted.org/packages/3c/5f/24155e30ba7f8c96918af1350eb0663e2430aad9e001c0489d89cd708ab1/aiohttp-3.13.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:fc353029f176fd2b3ec6cfc71be166aba1936fe5d73dd1992ce289ca6647a9aa", size = 1769801, upload-time = "2026-01-03T17:32:20.25Z" },
-    { url = "https://files.pythonhosted.org/packages/eb/f8/7314031ff5c10e6ece114da79b338ec17eeff3a079e53151f7e9f43c4723/aiohttp-3.13.3-cp314-cp314t-win32.whl", hash = "sha256:2e41b18a58da1e474a057b3d35248d8320029f61d70a37629535b16a0c8f3767", size = 466523, upload-time = "2026-01-03T17:32:22.215Z" },
-    { url = "https://files.pythonhosted.org/packages/b4/63/278a98c715ae467624eafe375542d8ba9b4383a016df8fdefe0ae28382a7/aiohttp-3.13.3-cp314-cp314t-win_amd64.whl", hash = "sha256:44531a36aa2264a1860089ffd4dce7baf875ee5a6079d5fb42e261c704ef7344", size = 499694, upload-time = "2026-01-03T17:32:24.546Z" },
+    { url = "https://files.pythonhosted.org/packages/1e/bd/ede278648914cabbabfdf95e436679b5d4156e417896a9b9f4587169e376/aiohttp-3.13.4-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:ee62d4471ce86b108b19c3364db4b91180d13fe3510144872d6bad5401957360", size = 752158, upload-time = "2026-03-28T17:16:06.901Z" },
+    { url = "https://files.pythonhosted.org/packages/90/de/581c053253c07b480b03785196ca5335e3c606a37dc73e95f6527f1591fe/aiohttp-3.13.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:c0fd8f41b54b58636402eb493afd512c23580456f022c1ba2db0f810c959ed0d", size = 501037, upload-time = "2026-03-28T17:16:08.82Z" },
+    { url = "https://files.pythonhosted.org/packages/fa/f9/a5ede193c08f13cc42c0a5b50d1e246ecee9115e4cf6e900d8dbd8fd6acb/aiohttp-3.13.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:4baa48ce49efd82d6b1a0be12d6a36b35e5594d1dd42f8bfba96ea9f8678b88c", size = 501556, upload-time = "2026-03-28T17:16:10.63Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/10/88ff67cd48a6ec36335b63a640abe86135791544863e0cfe1f065d6cef7a/aiohttp-3.13.4-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d738ebab9f71ee652d9dbd0211057690022201b11197f9a7324fd4dba128aa97", size = 1757314, upload-time = "2026-03-28T17:16:12.498Z" },
+    { url = "https://files.pythonhosted.org/packages/8b/15/fdb90a5cf5a1f52845c276e76298c75fbbcc0ac2b4a86551906d54529965/aiohttp-3.13.4-cp312-cp312-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:0ce692c3468fa831af7dceed52edf51ac348cebfc8d3feb935927b63bd3e8576", size = 1731819, upload-time = "2026-03-28T17:16:14.558Z" },
+    { url = "https://files.pythonhosted.org/packages/ec/df/28146785a007f7820416be05d4f28cc207493efd1e8c6c1068e9bdc29198/aiohttp-3.13.4-cp312-cp312-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:8e08abcfe752a454d2cb89ff0c08f2d1ecd057ae3e8cc6d84638de853530ebab", size = 1793279, upload-time = "2026-03-28T17:16:16.594Z" },
+    { url = "https://files.pythonhosted.org/packages/10/47/689c743abf62ea7a77774d5722f220e2c912a77d65d368b884d9779ef41b/aiohttp-3.13.4-cp312-cp312-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:5977f701b3fff36367a11087f30ea73c212e686d41cd363c50c022d48b011d8d", size = 1891082, upload-time = "2026-03-28T17:16:18.71Z" },
+    { url = "https://files.pythonhosted.org/packages/b0/b6/f7f4f318c7e58c23b761c9b13b9a3c9b394e0f9d5d76fbc6622fa98509f6/aiohttp-3.13.4-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:54203e10405c06f8b6020bd1e076ae0fe6c194adcee12a5a78af3ffa3c57025e", size = 1773938, upload-time = "2026-03-28T17:16:21.125Z" },
+    { url = "https://files.pythonhosted.org/packages/aa/06/f207cb3121852c989586a6fc16ff854c4fcc8651b86c5d3bd1fc83057650/aiohttp-3.13.4-cp312-cp312-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:358a6af0145bc4dda037f13167bef3cce54b132087acc4c295c739d05d16b1c3", size = 1579548, upload-time = "2026-03-28T17:16:23.588Z" },
+    { url = "https://files.pythonhosted.org/packages/6c/58/e1289661a32161e24c1fe479711d783067210d266842523752869cc1d9c2/aiohttp-3.13.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:898ea1850656d7d61832ef06aa9846ab3ddb1621b74f46de78fbc5e1a586ba83", size = 1714669, upload-time = "2026-03-28T17:16:25.713Z" },
+    { url = "https://files.pythonhosted.org/packages/96/0a/3e86d039438a74a86e6a948a9119b22540bae037d6ba317a042ae3c22711/aiohttp-3.13.4-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:7bc30cceb710cf6a44e9617e43eebb6e3e43ad855a34da7b4b6a73537d8a6763", size = 1754175, upload-time = "2026-03-28T17:16:28.18Z" },
+    { url = "https://files.pythonhosted.org/packages/f4/30/e717fc5df83133ba467a560b6d8ef20197037b4bb5d7075b90037de1018e/aiohttp-3.13.4-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:4a31c0c587a8a038f19a4c7e60654a6c899c9de9174593a13e7cc6e15ff271f9", size = 1762049, upload-time = "2026-03-28T17:16:30.941Z" },
+    { url = "https://files.pythonhosted.org/packages/e4/28/8f7a2d4492e336e40005151bdd94baf344880a4707573378579f833a64c1/aiohttp-3.13.4-cp312-cp312-musllinux_1_2_riscv64.whl", hash = "sha256:2062f675f3fe6e06d6113eb74a157fb9df58953ffed0cdb4182554b116545758", size = 1570861, upload-time = "2026-03-28T17:16:32.953Z" },
+    { url = "https://files.pythonhosted.org/packages/78/45/12e1a3d0645968b1c38de4b23fdf270b8637735ea057d4f84482ff918ad9/aiohttp-3.13.4-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:3d1ba8afb847ff80626d5e408c1fdc99f942acc877d0702fe137015903a220a9", size = 1790003, upload-time = "2026-03-28T17:16:35.468Z" },
+    { url = "https://files.pythonhosted.org/packages/eb/0f/60374e18d590de16dcb39d6ff62f39c096c1b958e6f37727b5870026ea30/aiohttp-3.13.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:b08149419994cdd4d5eecf7fd4bc5986b5a9380285bcd01ab4c0d6bfca47b79d", size = 1737289, upload-time = "2026-03-28T17:16:38.187Z" },
+    { url = "https://files.pythonhosted.org/packages/02/bf/535e58d886cfbc40a8b0013c974afad24ef7632d645bca0b678b70033a60/aiohttp-3.13.4-cp312-cp312-win32.whl", hash = "sha256:fc432f6a2c4f720180959bc19aa37259651c1a4ed8af8afc84dd41c60f15f791", size = 434185, upload-time = "2026-03-28T17:16:40.735Z" },
+    { url = "https://files.pythonhosted.org/packages/1e/1a/d92e3325134ebfff6f4069f270d3aac770d63320bd1fcd0eca023e74d9a8/aiohttp-3.13.4-cp312-cp312-win_amd64.whl", hash = "sha256:6148c9ae97a3e8bff9a1fc9c757fa164116f86c100468339730e717590a3fb77", size = 461285, upload-time = "2026-03-28T17:16:42.713Z" },
+    { url = "https://files.pythonhosted.org/packages/e3/ac/892f4162df9b115b4758d615f32ec63d00f3084c705ff5526630887b9b42/aiohttp-3.13.4-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:63dd5e5b1e43b8fb1e91b79b7ceba1feba588b317d1edff385084fcc7a0a4538", size = 745744, upload-time = "2026-03-28T17:16:44.67Z" },
+    { url = "https://files.pythonhosted.org/packages/97/a9/c5b87e4443a2f0ea88cb3000c93a8fdad1ee63bffc9ded8d8c8e0d66efc6/aiohttp-3.13.4-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:746ac3cc00b5baea424dacddea3ec2c2702f9590de27d837aa67004db1eebc6e", size = 498178, upload-time = "2026-03-28T17:16:46.766Z" },
+    { url = "https://files.pythonhosted.org/packages/94/42/07e1b543a61250783650df13da8ddcdc0d0a5538b2bd15cef6e042aefc61/aiohttp-3.13.4-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:bda8f16ea99d6a6705e5946732e48487a448be874e54a4f73d514660ff7c05d3", size = 498331, upload-time = "2026-03-28T17:16:48.9Z" },
+    { url = "https://files.pythonhosted.org/packages/20/d6/492f46bf0328534124772d0cf58570acae5b286ea25006900650f69dae0e/aiohttp-3.13.4-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:4b061e7b5f840391e3f64d0ddf672973e45c4cfff7a0feea425ea24e51530fc2", size = 1744414, upload-time = "2026-03-28T17:16:50.968Z" },
+    { url = "https://files.pythonhosted.org/packages/e2/4d/e02627b2683f68051246215d2d62b2d2f249ff7a285e7a858dc47d6b6a14/aiohttp-3.13.4-cp313-cp313-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:b252e8d5cd66184b570d0d010de742736e8a4fab22c58299772b0c5a466d4b21", size = 1719226, upload-time = "2026-03-28T17:16:53.173Z" },
+    { url = "https://files.pythonhosted.org/packages/7b/6c/5d0a3394dd2b9f9aeba6e1b6065d0439e4b75d41f1fb09a3ec010b43552b/aiohttp-3.13.4-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:20af8aad61d1803ff11152a26146d8d81c266aa8c5aa9b4504432abb965c36a0", size = 1782110, upload-time = "2026-03-28T17:16:55.362Z" },
+    { url = "https://files.pythonhosted.org/packages/0d/2d/c20791e3437700a7441a7edfb59731150322424f5aadf635602d1d326101/aiohttp-3.13.4-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:13a5cc924b59859ad2adb1478e31f410a7ed46e92a2a619d6d1dd1a63c1a855e", size = 1884809, upload-time = "2026-03-28T17:16:57.734Z" },
+    { url = "https://files.pythonhosted.org/packages/c8/94/d99dbfbd1924a87ef643833932eb2a3d9e5eee87656efea7d78058539eff/aiohttp-3.13.4-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:534913dfb0a644d537aebb4123e7d466d94e3be5549205e6a31f72368980a81a", size = 1764938, upload-time = "2026-03-28T17:17:00.221Z" },
+    { url = "https://files.pythonhosted.org/packages/49/61/3ce326a1538781deb89f6cf5e094e2029cd308ed1e21b2ba2278b08426f6/aiohttp-3.13.4-cp313-cp313-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:320e40192a2dcc1cf4b5576936e9652981ab596bf81eb309535db7e2f5b5672f", size = 1570697, upload-time = "2026-03-28T17:17:02.985Z" },
+    { url = "https://files.pythonhosted.org/packages/b6/77/4ab5a546857bb3028fbaf34d6eea180267bdab022ee8b1168b1fcde4bfdd/aiohttp-3.13.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:9e587fcfce2bcf06526a43cb705bdee21ac089096f2e271d75de9c339db3100c", size = 1702258, upload-time = "2026-03-28T17:17:05.28Z" },
+    { url = "https://files.pythonhosted.org/packages/79/63/d8f29021e39bc5af8e5d5e9da1b07976fb9846487a784e11e4f4eeda4666/aiohttp-3.13.4-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:9eb9c2eea7278206b5c6c1441fdd9dc420c278ead3f3b2cc87f9b693698cc500", size = 1740287, upload-time = "2026-03-28T17:17:07.712Z" },
+    { url = "https://files.pythonhosted.org/packages/55/3a/cbc6b3b124859a11bc8055d3682c26999b393531ef926754a3445b99dfef/aiohttp-3.13.4-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:29be00c51972b04bf9d5c8f2d7f7314f48f96070ca40a873a53056e652e805f7", size = 1753011, upload-time = "2026-03-28T17:17:10.053Z" },
+    { url = "https://files.pythonhosted.org/packages/e0/30/836278675205d58c1368b21520eab9572457cf19afd23759216c04483048/aiohttp-3.13.4-cp313-cp313-musllinux_1_2_riscv64.whl", hash = "sha256:90c06228a6c3a7c9f776fe4fc0b7ff647fffd3bed93779a6913c804ae00c1073", size = 1566359, upload-time = "2026-03-28T17:17:12.433Z" },
+    { url = "https://files.pythonhosted.org/packages/50/b4/8032cc9b82d17e4277704ba30509eaccb39329dc18d6a35f05e424439e32/aiohttp-3.13.4-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:a533ec132f05fd9a1d959e7f34184cd7d5e8511584848dab85faefbaac573069", size = 1785537, upload-time = "2026-03-28T17:17:14.721Z" },
+    { url = "https://files.pythonhosted.org/packages/17/7d/5873e98230bde59f493bf1f7c3e327486a4b5653fa401144704df5d00211/aiohttp-3.13.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:1c946f10f413836f82ea4cfb90200d2a59578c549f00857e03111cf45ad01ca5", size = 1740752, upload-time = "2026-03-28T17:17:17.387Z" },
+    { url = "https://files.pythonhosted.org/packages/7b/f2/13e46e0df051494d7d3c68b7f72d071f48c384c12716fc294f75d5b1a064/aiohttp-3.13.4-cp313-cp313-win32.whl", hash = "sha256:48708e2706106da6967eff5908c78ca3943f005ed6bcb75da2a7e4da94ef8c70", size = 433187, upload-time = "2026-03-28T17:17:19.523Z" },
+    { url = "https://files.pythonhosted.org/packages/ea/c0/649856ee655a843c8f8664592cfccb73ac80ede6a8c8db33a25d810c12db/aiohttp-3.13.4-cp313-cp313-win_amd64.whl", hash = "sha256:74a2eb058da44fa3a877a49e2095b591d4913308bb424c418b77beb160c55ce3", size = 459778, upload-time = "2026-03-28T17:17:21.964Z" },
+    { url = "https://files.pythonhosted.org/packages/6d/29/6657cc37ae04cacc2dbf53fb730a06b6091cc4cbe745028e047c53e6d840/aiohttp-3.13.4-cp314-cp314-macosx_10_13_universal2.whl", hash = "sha256:e0a2c961fc92abeff61d6444f2ce6ad35bb982db9fc8ff8a47455beacf454a57", size = 749363, upload-time = "2026-03-28T17:17:24.044Z" },
+    { url = "https://files.pythonhosted.org/packages/90/7f/30ccdf67ca3d24b610067dc63d64dcb91e5d88e27667811640644aa4a85d/aiohttp-3.13.4-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:153274535985a0ff2bff1fb6c104ed547cec898a09213d21b0f791a44b14d933", size = 499317, upload-time = "2026-03-28T17:17:26.199Z" },
+    { url = "https://files.pythonhosted.org/packages/93/13/e372dd4e68ad04ee25dafb050c7f98b0d91ea643f7352757e87231102555/aiohttp-3.13.4-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:351f3171e2458da3d731ce83f9e6b9619e325c45cbd534c7759750cabf453ad7", size = 500477, upload-time = "2026-03-28T17:17:28.279Z" },
+    { url = "https://files.pythonhosted.org/packages/e5/fe/ee6298e8e586096fb6f5eddd31393d8544f33ae0792c71ecbb4c2bef98ac/aiohttp-3.13.4-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:f989ac8bc5595ff761a5ccd32bdb0768a117f36dd1504b1c2c074ed5d3f4df9c", size = 1737227, upload-time = "2026-03-28T17:17:30.587Z" },
+    { url = "https://files.pythonhosted.org/packages/b0/b9/a7a0463a09e1a3fe35100f74324f23644bfc3383ac5fd5effe0722a5f0b7/aiohttp-3.13.4-cp314-cp314-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:d36fc1709110ec1e87a229b201dd3ddc32aa01e98e7868083a794609b081c349", size = 1694036, upload-time = "2026-03-28T17:17:33.29Z" },
+    { url = "https://files.pythonhosted.org/packages/57/7c/8972ae3fb7be00a91aee6b644b2a6a909aedb2c425269a3bfd90115e6f8f/aiohttp-3.13.4-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:42adaeea83cbdf069ab94f5103ce0787c21fb1a0153270da76b59d5578302329", size = 1786814, upload-time = "2026-03-28T17:17:36.035Z" },
+    { url = "https://files.pythonhosted.org/packages/93/01/c81e97e85c774decbaf0d577de7d848934e8166a3a14ad9f8aa5be329d28/aiohttp-3.13.4-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:92deb95469928cc41fd4b42a95d8012fa6df93f6b1c0a83af0ffbc4a5e218cde", size = 1866676, upload-time = "2026-03-28T17:17:38.441Z" },
+    { url = "https://files.pythonhosted.org/packages/5a/5f/5b46fe8694a639ddea2cd035bf5729e4677ea882cb251396637e2ef1590d/aiohttp-3.13.4-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0c0c7c07c4257ef3a1df355f840bc62d133bcdef5c1c5ba75add3c08553e2eed", size = 1740842, upload-time = "2026-03-28T17:17:40.783Z" },
+    { url = "https://files.pythonhosted.org/packages/20/a2/0d4b03d011cca6b6b0acba8433193c1e484efa8d705ea58295590fe24203/aiohttp-3.13.4-cp314-cp314-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:f062c45de8a1098cb137a1898819796a2491aec4e637a06b03f149315dff4d8f", size = 1566508, upload-time = "2026-03-28T17:17:43.235Z" },
+    { url = "https://files.pythonhosted.org/packages/98/17/e689fd500da52488ec5f889effd6404dece6a59de301e380f3c64f167beb/aiohttp-3.13.4-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:76093107c531517001114f0ebdb4f46858ce818590363e3e99a4a2280334454a", size = 1700569, upload-time = "2026-03-28T17:17:46.165Z" },
+    { url = "https://files.pythonhosted.org/packages/d8/0d/66402894dbcf470ef7db99449e436105ea862c24f7ea4c95c683e635af35/aiohttp-3.13.4-cp314-cp314-musllinux_1_2_armv7l.whl", hash = "sha256:6f6ec32162d293b82f8b63a16edc80769662fbd5ae6fbd4936d3206a2c2cc63b", size = 1707407, upload-time = "2026-03-28T17:17:48.825Z" },
+    { url = "https://files.pythonhosted.org/packages/2f/eb/af0ab1a3650092cbd8e14ef29e4ab0209e1460e1c299996c3f8288b3f1ff/aiohttp-3.13.4-cp314-cp314-musllinux_1_2_ppc64le.whl", hash = "sha256:5903e2db3d202a00ad9f0ec35a122c005e85d90c9836ab4cda628f01edf425e2", size = 1752214, upload-time = "2026-03-28T17:17:51.206Z" },
+    { url = "https://files.pythonhosted.org/packages/5a/bf/72326f8a98e4c666f292f03c385545963cc65e358835d2a7375037a97b57/aiohttp-3.13.4-cp314-cp314-musllinux_1_2_riscv64.whl", hash = "sha256:2d5bea57be7aca98dbbac8da046d99b5557c5cf4e28538c4c786313078aca09e", size = 1562162, upload-time = "2026-03-28T17:17:53.634Z" },
+    { url = "https://files.pythonhosted.org/packages/67/9f/13b72435f99151dd9a5469c96b3b5f86aa29b7e785ca7f35cf5e538f74c0/aiohttp-3.13.4-cp314-cp314-musllinux_1_2_s390x.whl", hash = "sha256:bcf0c9902085976edc0232b75006ef38f89686901249ce14226b6877f88464fb", size = 1768904, upload-time = "2026-03-28T17:17:55.991Z" },
+    { url = "https://files.pythonhosted.org/packages/18/bc/28d4970e7d5452ac7776cdb5431a1164a0d9cf8bd2fffd67b4fb463aa56d/aiohttp-3.13.4-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:c3295f98bfeed2e867cab588f2a146a9db37a85e3ae9062abf46ba062bd29165", size = 1723378, upload-time = "2026-03-28T17:17:58.348Z" },
+    { url = "https://files.pythonhosted.org/packages/53/74/b32458ca1a7f34d65bdee7aef2036adbe0438123d3d53e2b083c453c24dd/aiohttp-3.13.4-cp314-cp314-win32.whl", hash = "sha256:a598a5c5767e1369d8f5b08695cab1d8160040f796c4416af76fd773d229b3c9", size = 438711, upload-time = "2026-03-28T17:18:00.728Z" },
+    { url = "https://files.pythonhosted.org/packages/40/b2/54b487316c2df3e03a8f3435e9636f8a81a42a69d942164830d193beb56a/aiohttp-3.13.4-cp314-cp314-win_amd64.whl", hash = "sha256:c555db4bc7a264bead5a7d63d92d41a1122fcd39cc62a4db815f45ad46f9c2c8", size = 464977, upload-time = "2026-03-28T17:18:03.367Z" },
+    { url = "https://files.pythonhosted.org/packages/47/fb/e41b63c6ce71b07a59243bb8f3b457ee0c3402a619acb9d2c0d21ef0e647/aiohttp-3.13.4-cp314-cp314t-macosx_10_13_universal2.whl", hash = "sha256:45abbbf09a129825d13c18c7d3182fecd46d9da3cfc383756145394013604ac1", size = 781549, upload-time = "2026-03-28T17:18:05.779Z" },
+    { url = "https://files.pythonhosted.org/packages/97/53/532b8d28df1e17e44c4d9a9368b78dcb6bf0b51037522136eced13afa9e8/aiohttp-3.13.4-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:74c80b2bc2c2adb7b3d1941b2b60701ee2af8296fc8aad8b8bc48bc25767266c", size = 514383, upload-time = "2026-03-28T17:18:08.096Z" },
+    { url = "https://files.pythonhosted.org/packages/1b/1f/62e5d400603e8468cd635812d99cb81cfdc08127a3dc474c647615f31339/aiohttp-3.13.4-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:c97989ae40a9746650fa196894f317dafc12227c808c774929dda0ff873a5954", size = 518304, upload-time = "2026-03-28T17:18:10.642Z" },
+    { url = "https://files.pythonhosted.org/packages/90/57/2326b37b10896447e3c6e0cbef4fe2486d30913639a5cfd1332b5d870f82/aiohttp-3.13.4-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:dae86be9811493f9990ef44fff1685f5c1a3192e9061a71a109d527944eed551", size = 1893433, upload-time = "2026-03-28T17:18:13.121Z" },
+    { url = "https://files.pythonhosted.org/packages/d2/b4/a24d82112c304afdb650167ef2fe190957d81cbddac7460bedd245f765aa/aiohttp-3.13.4-cp314-cp314t-manylinux2014_armv7l.manylinux_2_17_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:1db491abe852ca2fa6cc48a3341985b0174b3741838e1341b82ac82c8bd9e871", size = 1755901, upload-time = "2026-03-28T17:18:16.21Z" },
+    { url = "https://files.pythonhosted.org/packages/9e/2d/0883ef9d878d7846287f036c162a951968f22aabeef3ac97b0bea6f76d5d/aiohttp-3.13.4-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.manylinux_2_28_ppc64le.whl", hash = "sha256:0e5d701c0aad02a7dce72eef6b93226cf3734330f1a31d69ebbf69f33b86666e", size = 1876093, upload-time = "2026-03-28T17:18:18.703Z" },
+    { url = "https://files.pythonhosted.org/packages/ad/52/9204bb59c014869b71971addad6778f005daa72a96eed652c496789d7468/aiohttp-3.13.4-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.manylinux_2_28_s390x.whl", hash = "sha256:8ac32a189081ae0a10ba18993f10f338ec94341f0d5df8fff348043962f3c6f8", size = 1970815, upload-time = "2026-03-28T17:18:21.858Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/b5/e4eb20275a866dde0f570f411b36c6b48f7b53edfe4f4071aa1b0728098a/aiohttp-3.13.4-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:98e968cdaba43e45c73c3f306fca418c8009a957733bac85937c9f9cf3f4de27", size = 1816223, upload-time = "2026-03-28T17:18:24.729Z" },
+    { url = "https://files.pythonhosted.org/packages/d8/23/e98075c5bb146aa61a1239ee1ac7714c85e814838d6cebbe37d3fe19214a/aiohttp-3.13.4-cp314-cp314t-manylinux_2_31_riscv64.manylinux_2_39_riscv64.whl", hash = "sha256:ca114790c9144c335d538852612d3e43ea0f075288f4849cf4b05d6cd2238ce7", size = 1649145, upload-time = "2026-03-28T17:18:27.269Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/c1/7bad8be33bb06c2bb224b6468874346026092762cbec388c3bdb65a368ee/aiohttp-3.13.4-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:ea2e071661ba9cfe11eabbc81ac5376eaeb3061f6e72ec4cc86d7cdd1ffbdbbb", size = 1816562, upload-time = "2026-03-28T17:18:29.847Z" },
+    { url = "https://files.pythonhosted.org/packages/5c/10/c00323348695e9a5e316825969c88463dcc24c7e9d443244b8a2c9cf2eae/aiohttp-3.13.4-cp314-cp314t-musllinux_1_2_armv7l.whl", hash = "sha256:34e89912b6c20e0fd80e07fa401fd218a410aa1ce9f1c2f1dad6db1bd0ce0927", size = 1800333, upload-time = "2026-03-28T17:18:32.269Z" },
+    { url = "https://files.pythonhosted.org/packages/84/43/9b2147a1df3559f49bd723e22905b46a46c068a53adb54abdca32c4de180/aiohttp-3.13.4-cp314-cp314t-musllinux_1_2_ppc64le.whl", hash = "sha256:0e217cf9f6a42908c52b46e42c568bd57adc39c9286ced31aaace614b6087965", size = 1820617, upload-time = "2026-03-28T17:18:35.238Z" },
+    { url = "https://files.pythonhosted.org/packages/a9/7f/b3481a81e7a586d02e99387b18c6dafff41285f6efd3daa2124c01f87eae/aiohttp-3.13.4-cp314-cp314t-musllinux_1_2_riscv64.whl", hash = "sha256:0c296f1221e21ba979f5ac1964c3b78cfde15c5c5f855ffd2caab337e9cd9182", size = 1643417, upload-time = "2026-03-28T17:18:37.949Z" },
+    { url = "https://files.pythonhosted.org/packages/8f/72/07181226bc99ce1124e0f89280f5221a82d3ae6a6d9d1973ce429d48e52b/aiohttp-3.13.4-cp314-cp314t-musllinux_1_2_s390x.whl", hash = "sha256:d99a9d168ebaffb74f36d011750e490085ac418f4db926cce3989c8fe6cb6b1b", size = 1849286, upload-time = "2026-03-28T17:18:40.534Z" },
+    { url = "https://files.pythonhosted.org/packages/1a/e6/1b3566e103eca6da5be4ae6713e112a053725c584e96574caf117568ffef/aiohttp-3.13.4-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:cb19177205d93b881f3f89e6081593676043a6828f59c78c17a0fd6c1fbed2ba", size = 1782635, upload-time = "2026-03-28T17:18:43.073Z" },
+    { url = "https://files.pythonhosted.org/packages/37/58/1b11c71904b8d079eb0c39fe664180dd1e14bebe5608e235d8bfbadc8929/aiohttp-3.13.4-cp314-cp314t-win32.whl", hash = "sha256:c606aa5656dab6552e52ca368e43869c916338346bfaf6304e15c58fb113ea30", size = 472537, upload-time = "2026-03-28T17:18:46.286Z" },
+    { url = "https://files.pythonhosted.org/packages/bc/8f/87c56a1a1977d7dddea5b31e12189665a140fdb48a71e9038ff90bb564ec/aiohttp-3.13.4-cp314-cp314t-win_amd64.whl", hash = "sha256:014dcc10ec8ab8db681f0d68e939d1e9286a5aa2b993cbbdb0db130853e02144", size = 506381, upload-time = "2026-03-28T17:18:48.74Z" },
 ]
 
 [[package]]
@@ -2153,7 +2153,7 @@ wheels = [
 
 [[package]]
 name = "requests"
-version = "2.32.5"
+version = "2.33.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "certifi" },
@@ -2161,9 +2161,9 @@ dependencies = [
     { name = "idna" },
     { name = "urllib3" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/c9/74/b3ff8e6c8446842c3f5c837e9c3dfcfe2018ea6ecef224c710c85ef728f4/requests-2.32.5.tar.gz", hash = "sha256:dbba0bac56e100853db0ea71b82b4dfd5fe2bf6d3754a8893c3af500cec7d7cf", size = 134517, upload-time = "2025-08-18T20:46:02.573Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/34/64/8860370b167a9721e8956ae116825caff829224fbca0ca6e7bf8ddef8430/requests-2.33.0.tar.gz", hash = "sha256:c7ebc5e8b0f21837386ad0e1c8fe8b829fa5f544d8df3b2253bff14ef29d7652", size = 134232, upload-time = "2026-03-25T15:10:41.586Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/1e/db/4254e3eabe8020b458f1a747140d32277ec7a271daf1d235b70dc0b4e6e3/requests-2.32.5-py3-none-any.whl", hash = "sha256:2462f94637a34fd532264295e186976db0f5d453d1cdd31473c85a6a161affb6", size = 64738, upload-time = "2025-08-18T20:46:00.542Z" },
+    { url = "https://files.pythonhosted.org/packages/56/5d/c814546c2333ceea4ba42262d8c4d55763003e767fa169adc693bd524478/requests-2.33.0-py3-none-any.whl", hash = "sha256:3324635456fa185245e24865e810cecec7b4caf933d7eb133dcde67d48cee69b", size = 65017, upload-time = "2026-03-25T15:10:40.382Z" },
 ]
 
 [[package]]

livekit.yaml.example

@@ -0,0 +1,34 @@
+# LiveKit server configuration
+# Generated by setup-selfhosted.sh — do not edit manually.
+# See: https://docs.livekit.io/self-hosting/deployment/
+
+port: 7880
+rtc:
+  tcp_port: 7881
+  port_range_start: 44200
+  port_range_end: 44300
+  # use_external_ip: true  # Uncomment for production with public IP
+
+redis:
+  address: redis:6379
+
+keys:
+  # API key : API secret (generated by setup script)
+  # devkey: secret
+  __LIVEKIT_API_KEY__: __LIVEKIT_API_SECRET__
+
+webhook:
+  urls:
+    - http://server:1250/v1/livekit/webhook
+  api_key: __LIVEKIT_API_KEY__
+
+logging:
+  level: info
+
+# Room settings
+room:
+  empty_timeout: 300     # 5 minutes after last participant leaves
+  max_participants: 0    # 0 = unlimited
+
+# Track Egress only (no composite video)
+# Egress is configured via egress.yaml on the egress service

scripts/generate-certs.sh

@@ -0,0 +1,130 @@
+#!/usr/bin/env bash
+#
+# Generate a local CA and server certificate for Reflector self-hosted deployments.
+#
+# Usage:
+#   ./scripts/generate-certs.sh DOMAIN [EXTRA_SANS...]
+#
+# Examples:
+#   ./scripts/generate-certs.sh reflector.local
+#   ./scripts/generate-certs.sh reflector.local "DNS:gpu.local,IP:192.168.1.100"
+#
+# Generates in certs/:
+#   ca.key           — CA private key (keep secret)
+#   ca.crt           — CA certificate (distribute to clients)
+#   server-key.pem   — Server private key
+#   server.pem       — Server certificate (signed by CA)
+#
+# Then use with setup-selfhosted.sh:
+#   ./scripts/setup-selfhosted.sh --gpu --caddy --domain DOMAIN --custom-ca certs/
+#
+set -euo pipefail
+
+DOMAIN="${1:?Usage: $0 DOMAIN [EXTRA_SANS...]}"
+EXTRA_SANS="${2:-}"
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+CERTS_DIR="$(cd "$SCRIPT_DIR/.." && pwd)/certs"
+
+# Colors
+GREEN='\033[0;32m'
+CYAN='\033[0;36m'
+NC='\033[0m'
+info()  { echo -e "${CYAN}==>${NC} $*"; }
+ok()    { echo -e "${GREEN}  ✓${NC} $*"; }
+
+# Check for openssl
+if ! command -v openssl &>/dev/null; then
+    echo "Error: openssl is required but not found. Install it first." >&2
+    exit 1
+fi
+
+mkdir -p "$CERTS_DIR"
+
+# Build SAN list
+SAN_LIST="DNS:$DOMAIN,DNS:localhost,IP:127.0.0.1"
+if [[ -n "$EXTRA_SANS" ]]; then
+    SAN_LIST="$SAN_LIST,$EXTRA_SANS"
+fi
+
+info "Generating CA and server certificate for: $DOMAIN"
+echo "  SANs: $SAN_LIST"
+echo ""
+
+# --- Step 1: Generate CA ---
+if [[ -f "$CERTS_DIR/ca.key" ]] && [[ -f "$CERTS_DIR/ca.crt" ]]; then
+    ok "CA already exists at certs/ca.key + certs/ca.crt — reusing"
+else
+    info "Generating CA key and certificate..."
+    openssl genrsa -out "$CERTS_DIR/ca.key" 4096 2>/dev/null
+    openssl req -x509 -new -nodes \
+        -key "$CERTS_DIR/ca.key" \
+        -sha256 -days 3650 \
+        -out "$CERTS_DIR/ca.crt" \
+        -subj "/CN=Reflector Local CA/O=Reflector Self-Hosted"
+    ok "CA certificate generated (valid for 10 years)"
+fi
+
+# --- Step 2: Generate server key ---
+info "Generating server key..."
+openssl genrsa -out "$CERTS_DIR/server-key.pem" 2048 2>/dev/null
+ok "Server key generated"
+
+# --- Step 3: Create CSR with SANs ---
+info "Creating certificate signing request..."
+openssl req -new \
+    -key "$CERTS_DIR/server-key.pem" \
+    -out "$CERTS_DIR/server.csr" \
+    -subj "/CN=$DOMAIN" \
+    -addext "subjectAltName=$SAN_LIST"
+ok "CSR created"
+
+# --- Step 4: Sign with CA ---
+info "Signing server certificate with CA..."
+openssl x509 -req \
+    -in "$CERTS_DIR/server.csr" \
+    -CA "$CERTS_DIR/ca.crt" \
+    -CAkey "$CERTS_DIR/ca.key" \
+    -CAcreateserial \
+    -out "$CERTS_DIR/server.pem" \
+    -days 365 -sha256 \
+    -copy_extensions copyall \
+    2>/dev/null
+ok "Server certificate signed (valid for 1 year)"
+
+# --- Cleanup ---
+rm -f "$CERTS_DIR/server.csr" "$CERTS_DIR/ca.srl"
+
+# --- Set permissions ---
+chmod 644 "$CERTS_DIR/ca.crt" "$CERTS_DIR/server.pem"
+chmod 600 "$CERTS_DIR/ca.key" "$CERTS_DIR/server-key.pem"
+
+echo ""
+echo "=========================================="
+echo -e " ${GREEN}Certificates generated in certs/${NC}"
+echo "=========================================="
+echo ""
+echo "  certs/ca.key           CA private key (keep secret)"
+echo "  certs/ca.crt           CA certificate (distribute to clients)"
+echo "  certs/server-key.pem   Server private key"
+echo "  certs/server.pem       Server certificate for $DOMAIN"
+echo ""
+echo "  SANs: $SAN_LIST"
+echo ""
+echo "Use with setup-selfhosted.sh:"
+echo "  ./scripts/setup-selfhosted.sh --gpu --caddy --domain $DOMAIN --custom-ca certs/"
+echo ""
+echo "Trust the CA on your machine:"
+case "$(uname -s)" in
+    Darwin)
+        echo "  sudo security add-trusted-cert -d -r trustRoot -k /Library/Keychains/System.keychain certs/ca.crt"
+        ;;
+    Linux)
+        echo "  sudo cp certs/ca.crt /usr/local/share/ca-certificates/reflector-ca.crt"
+        echo "  sudo update-ca-certificates"
+        ;;
+    *)
+        echo "  See docsv2/custom-ca-setup.md for your platform"
+        ;;
+esac
+echo ""

scripts/run-integration-tests.sh

@@ -13,14 +13,25 @@
 # Optional:
 #   LLM_MODEL    — Model name (default: qwen2.5:14b)
 #
+# Flags:
+#   --build      — Rebuild backend Docker images (server, workers, test-runner)
+#
 # Usage:
 #   export LLM_URL="https://api.openai.com/v1"
 #   export LLM_API_KEY="sk-..."
 #   export HF_TOKEN="hf_..."
 #   ./scripts/run-integration-tests.sh
+#   ./scripts/run-integration-tests.sh --build   # rebuild backend images
 #
 set -euo pipefail
 
+BUILD_FLAG=""
+for arg in "$@"; do
+    case "$arg" in
+        --build) BUILD_FLAG="--build" ;;
+    esac
+done
+
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 REPO_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
 COMPOSE_DIR="$REPO_ROOT/server/tests"
@@ -66,7 +77,7 @@ trap cleanup EXIT
 
 # ── Step 1: Build and start infrastructure ──────────────────────────────────
 info "Building and starting infrastructure services..."
-$COMPOSE up -d --build postgres redis garage hatchet mock-daily
+$COMPOSE up -d --build postgres redis garage hatchet mock-daily mailpit
 
 # ── Step 2: Set up Garage (S3 bucket + keys) ───────────────────────────────
 wait_for "Garage" "$COMPOSE exec -T garage /garage stats" 60
@@ -116,7 +127,7 @@ ok "Hatchet token generated"
 
 # ── Step 4: Start backend services ──────────────────────────────────────────
 info "Starting backend services..."
-$COMPOSE up -d server worker hatchet-worker-cpu hatchet-worker-llm test-runner
+$COMPOSE up -d $BUILD_FLAG server worker hatchet-worker-cpu hatchet-worker-llm test-runner
 
 # ── Step 5: Wait for server + run migrations ────────────────────────────────
 wait_for "Server" "$COMPOSE exec -T test-runner curl -sf http://server:1250/health" 60

scripts/setup-gpu-host.sh

@@ -0,0 +1,496 @@
+#!/usr/bin/env bash
+#
+# Standalone GPU service setup for Reflector.
+# Deploys ONLY the GPU transcription/diarization/translation service on a dedicated machine.
+# The main Reflector instance connects to this machine over HTTPS.
+#
+# Usage:
+#   ./scripts/setup-gpu-host.sh [--domain DOMAIN] [--custom-ca PATH] [--extra-ca FILE] [--api-key KEY] [--cpu] [--build]
+#
+# Options:
+#   --domain DOMAIN    Domain name for this GPU host (e.g., gpu.example.com)
+#                      With --custom-ca: uses custom TLS cert. Without: uses Let's Encrypt.
+#   --custom-ca PATH   Custom CA certificate (dir with ca.crt + server.pem + server-key.pem, or single PEM file)
+#   --extra-ca FILE    Additional CA cert to trust (repeatable)
+#   --api-key KEY      API key to protect the GPU service (recommended for internet-facing deployments)
+#   --cpu              Use CPU-only Dockerfile (no NVIDIA GPU required)
+#   --build            Build image from source (default: build, since no pre-built GPU image is published)
+#   --port PORT        Host port to expose (default: 443 with Caddy, 8000 without)
+#
+# Examples:
+#   # GPU on LAN with custom CA
+#   ./scripts/generate-certs.sh gpu.local
+#   ./scripts/setup-gpu-host.sh --domain gpu.local --custom-ca certs/ --api-key my-secret-key
+#
+#   # GPU on public internet with Let's Encrypt
+#   ./scripts/setup-gpu-host.sh --domain gpu.example.com --api-key my-secret-key
+#
+#   # GPU on LAN, IP access only (self-signed cert)
+#   ./scripts/setup-gpu-host.sh --api-key my-secret-key
+#
+#   # CPU-only mode (no NVIDIA GPU)
+#   ./scripts/setup-gpu-host.sh --cpu --api-key my-secret-key
+#
+# After setup, configure the main Reflector instance to use this GPU:
+#   In server/.env on the Reflector machine:
+#     TRANSCRIPT_BACKEND=modal
+#     TRANSCRIPT_URL=https://gpu.example.com
+#     TRANSCRIPT_MODAL_API_KEY=my-secret-key
+#     DIARIZATION_BACKEND=modal
+#     DIARIZATION_URL=https://gpu.example.com
+#     DIARIZATION_MODAL_API_KEY=my-secret-key
+#     TRANSLATION_BACKEND=modal
+#     TRANSLATE_URL=https://gpu.example.com
+#     TRANSLATION_MODAL_API_KEY=my-secret-key
+#
+# DNS Resolution:
+#   - Public domain: Create a DNS A record pointing to this machine's public IP.
+#   - Internal domain (e.g., gpu.local): Add to /etc/hosts on both machines:
+#       <GPU_MACHINE_IP> gpu.local
+#   - IP-only: Use the machine's IP directly in TRANSCRIPT_URL/DIARIZATION_URL.
+#     The Reflector backend must trust the CA or accept self-signed certs.
+#
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ROOT_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
+
+GPU_DIR="$ROOT_DIR/gpu/self_hosted"
+OS="$(uname -s)"
+
+# --- Colors ---
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+CYAN='\033[0;36m'
+NC='\033[0m'
+
+info()  { echo -e "${CYAN}==>${NC} $*"; }
+ok()    { echo -e "${GREEN}  ✓${NC} $*"; }
+warn()  { echo -e "${YELLOW}  !${NC} $*"; }
+err()   { echo -e "${RED}  ✗${NC} $*" >&2; }
+
+# --- Parse arguments ---
+CUSTOM_DOMAIN=""
+CUSTOM_CA=""
+EXTRA_CA_FILES=()
+API_KEY=""
+USE_CPU=false
+HOST_PORT=""
+
+SKIP_NEXT=false
+ARGS=("$@")
+for i in "${!ARGS[@]}"; do
+    if [[ "$SKIP_NEXT" == "true" ]]; then
+        SKIP_NEXT=false
+        continue
+    fi
+    arg="${ARGS[$i]}"
+    case "$arg" in
+        --domain)
+            next_i=$((i + 1))
+            if [[ $next_i -ge ${#ARGS[@]} ]] || [[ "${ARGS[$next_i]}" == --* ]]; then
+                err "--domain requires a domain name"
+                exit 1
+            fi
+            CUSTOM_DOMAIN="${ARGS[$next_i]}"
+            SKIP_NEXT=true ;;
+        --custom-ca)
+            next_i=$((i + 1))
+            if [[ $next_i -ge ${#ARGS[@]} ]] || [[ "${ARGS[$next_i]}" == --* ]]; then
+                err "--custom-ca requires a path to a directory or PEM certificate file"
+                exit 1
+            fi
+            CUSTOM_CA="${ARGS[$next_i]}"
+            SKIP_NEXT=true ;;
+        --extra-ca)
+            next_i=$((i + 1))
+            if [[ $next_i -ge ${#ARGS[@]} ]] || [[ "${ARGS[$next_i]}" == --* ]]; then
+                err "--extra-ca requires a path to a PEM certificate file"
+                exit 1
+            fi
+            if [[ ! -f "${ARGS[$next_i]}" ]]; then
+                err "--extra-ca file not found: ${ARGS[$next_i]}"
+                exit 1
+            fi
+            EXTRA_CA_FILES+=("${ARGS[$next_i]}")
+            SKIP_NEXT=true ;;
+        --api-key)
+            next_i=$((i + 1))
+            if [[ $next_i -ge ${#ARGS[@]} ]] || [[ "${ARGS[$next_i]}" == --* ]]; then
+                err "--api-key requires a key value"
+                exit 1
+            fi
+            API_KEY="${ARGS[$next_i]}"
+            SKIP_NEXT=true ;;
+        --cpu)
+            USE_CPU=true ;;
+        --port)
+            next_i=$((i + 1))
+            if [[ $next_i -ge ${#ARGS[@]} ]] || [[ "${ARGS[$next_i]}" == --* ]]; then
+                err "--port requires a port number"
+                exit 1
+            fi
+            HOST_PORT="${ARGS[$next_i]}"
+            SKIP_NEXT=true ;;
+        --build)
+            ;; # Always build from source for GPU, flag accepted for compatibility
+        *)
+            err "Unknown argument: $arg"
+            err "Usage: $0 [--domain DOMAIN] [--custom-ca PATH] [--extra-ca FILE] [--api-key KEY] [--cpu] [--port PORT]"
+            exit 1
+            ;;
+    esac
+done
+
+# --- Resolve CA paths ---
+CA_CERT_PATH=""
+TLS_CERT_PATH=""
+TLS_KEY_PATH=""
+USE_CUSTOM_CA=false
+USE_CADDY=false
+
+if [[ -n "$CUSTOM_CA" ]] || [[ -n "${EXTRA_CA_FILES[0]+x}" ]]; then
+    USE_CUSTOM_CA=true
+fi
+
+if [[ -n "$CUSTOM_CA" ]]; then
+    CUSTOM_CA="${CUSTOM_CA%/}"
+    if [[ -d "$CUSTOM_CA" ]]; then
+        [[ -f "$CUSTOM_CA/ca.crt" ]] || { err "$CUSTOM_CA/ca.crt not found"; exit 1; }
+        CA_CERT_PATH="$CUSTOM_CA/ca.crt"
+        if [[ -f "$CUSTOM_CA/server.pem" ]] && [[ -f "$CUSTOM_CA/server-key.pem" ]]; then
+            TLS_CERT_PATH="$CUSTOM_CA/server.pem"
+            TLS_KEY_PATH="$CUSTOM_CA/server-key.pem"
+        elif [[ -f "$CUSTOM_CA/server.pem" ]] || [[ -f "$CUSTOM_CA/server-key.pem" ]]; then
+            warn "Found only one of server.pem/server-key.pem — both needed for TLS. Skipping."
+        fi
+    elif [[ -f "$CUSTOM_CA" ]]; then
+        CA_CERT_PATH="$CUSTOM_CA"
+    else
+        err "--custom-ca path not found: $CUSTOM_CA"
+        exit 1
+    fi
+elif [[ -n "${EXTRA_CA_FILES[0]+x}" ]]; then
+    CA_CERT_PATH="${EXTRA_CA_FILES[0]}"
+    unset 'EXTRA_CA_FILES[0]'
+    EXTRA_CA_FILES=("${EXTRA_CA_FILES[@]+"${EXTRA_CA_FILES[@]}"}")
+fi
+
+# Caddy if we have a domain or TLS certs
+if [[ -n "$CUSTOM_DOMAIN" ]] || [[ -n "$TLS_CERT_PATH" ]]; then
+    USE_CADDY=true
+fi
+
+# Default port
+if [[ -z "$HOST_PORT" ]]; then
+    if [[ "$USE_CADDY" == "true" ]]; then
+        HOST_PORT="443"
+    else
+        HOST_PORT="8000"
+    fi
+fi
+
+# Detect primary IP
+PRIMARY_IP=""
+if [[ "$OS" == "Linux" ]]; then
+    PRIMARY_IP=$(hostname -I 2>/dev/null | awk '{print $1}' || true)
+    if [[ "$PRIMARY_IP" == "127."* ]] || [[ -z "$PRIMARY_IP" ]]; then
+        PRIMARY_IP=$(ip -4 route get 1 2>/dev/null | sed -n 's/.*src \([0-9.]*\).*/\1/p' || true)
+    fi
+fi
+
+# --- Display config ---
+echo ""
+echo "=========================================="
+echo " Reflector — Standalone GPU Host Setup"
+echo "=========================================="
+echo ""
+echo "  Mode:    $(if [[ "$USE_CPU" == "true" ]]; then echo "CPU-only"; else echo "NVIDIA GPU"; fi)"
+echo "  Caddy:   $USE_CADDY"
+[[ -n "$CUSTOM_DOMAIN" ]] && echo "  Domain:  $CUSTOM_DOMAIN"
+[[ "$USE_CUSTOM_CA" == "true" ]] && echo "  CA:      Custom"
+[[ -n "$TLS_CERT_PATH" ]] && echo "  TLS:     Custom cert"
+[[ -n "$API_KEY" ]] && echo "  Auth:    API key protected"
+[[ -z "$API_KEY" ]] && echo "  Auth:    NONE (open access — use --api-key for production!)"
+echo "  Port:    $HOST_PORT"
+echo ""
+
+# --- Prerequisites ---
+info "Checking prerequisites"
+
+if ! command -v docker &>/dev/null; then
+    err "Docker not found. Install Docker first."
+    exit 1
+fi
+ok "Docker available"
+
+if ! docker compose version &>/dev/null; then
+    err "Docker Compose V2 not found."
+    exit 1
+fi
+ok "Docker Compose V2 available"
+
+if [[ "$USE_CPU" != "true" ]]; then
+    if ! docker info 2>/dev/null | grep -qi nvidia; then
+        warn "NVIDIA runtime not detected in Docker. GPU mode may fail."
+        warn "Install nvidia-container-toolkit if you have an NVIDIA GPU."
+    else
+        ok "NVIDIA Docker runtime available"
+    fi
+fi
+
+# --- Stage certificates ---
+CERTS_DIR="$ROOT_DIR/certs"
+if [[ "$USE_CUSTOM_CA" == "true" ]]; then
+    info "Staging certificates"
+    mkdir -p "$CERTS_DIR"
+
+    if [[ -n "$CA_CERT_PATH" ]]; then
+        local_ca_dest="$CERTS_DIR/ca.crt"
+        src_id=$(ls -i "$CA_CERT_PATH" 2>/dev/null | awk '{print $1}')
+        dst_id=$(ls -i "$local_ca_dest" 2>/dev/null | awk '{print $1}')
+        if [[ "$src_id" != "$dst_id" ]] || [[ -z "$dst_id" ]]; then
+            cp "$CA_CERT_PATH" "$local_ca_dest"
+        fi
+        chmod 644 "$local_ca_dest"
+        ok "CA certificate staged"
+
+        # Append extra CAs
+        for extra_ca in "${EXTRA_CA_FILES[@]+"${EXTRA_CA_FILES[@]}"}"; do
+            echo "" >> "$local_ca_dest"
+            cat "$extra_ca" >> "$local_ca_dest"
+            ok "Appended extra CA: $extra_ca"
+        done
+    fi
+
+    if [[ -n "$TLS_CERT_PATH" ]]; then
+        cert_dest="$CERTS_DIR/server.pem"
+        key_dest="$CERTS_DIR/server-key.pem"
+        src_id=$(ls -i "$TLS_CERT_PATH" 2>/dev/null | awk '{print $1}')
+        dst_id=$(ls -i "$cert_dest" 2>/dev/null | awk '{print $1}')
+        if [[ "$src_id" != "$dst_id" ]] || [[ -z "$dst_id" ]]; then
+            cp "$TLS_CERT_PATH" "$cert_dest"
+            cp "$TLS_KEY_PATH" "$key_dest"
+        fi
+        chmod 644 "$cert_dest"
+        chmod 600 "$key_dest"
+        ok "TLS cert/key staged"
+    fi
+fi
+
+# --- Build profiles and compose command ---
+COMPOSE_FILE="$ROOT_DIR/docker-compose.gpu-host.yml"
+COMPOSE_PROFILES=()
+GPU_SERVICE="gpu"
+
+if [[ "$USE_CPU" == "true" ]]; then
+    COMPOSE_PROFILES+=("cpu")
+    GPU_SERVICE="cpu"
+else
+    COMPOSE_PROFILES+=("gpu")
+fi
+if [[ "$USE_CADDY" == "true" ]]; then
+    COMPOSE_PROFILES+=("caddy")
+fi
+
+# Compose command helper
+compose_cmd() {
+    local profiles="" files="-f $COMPOSE_FILE"
+    if [[ "$USE_CUSTOM_CA" == "true" ]] && [[ -f "$ROOT_DIR/docker-compose.gpu-ca.yml" ]]; then
+        files="$files -f $ROOT_DIR/docker-compose.gpu-ca.yml"
+    fi
+    for p in "${COMPOSE_PROFILES[@]}"; do
+        profiles="$profiles --profile $p"
+    done
+    docker compose $files $profiles "$@"
+}
+
+# Generate CA compose override if needed (mounts certs into containers)
+if [[ "$USE_CUSTOM_CA" == "true" ]]; then
+    info "Generating docker-compose.gpu-ca.yml override"
+    ca_override="$ROOT_DIR/docker-compose.gpu-ca.yml"
+    cat > "$ca_override" << 'CAEOF'
+# Generated by setup-gpu-host.sh — custom CA trust.
+# Do not edit manually; re-run setup-gpu-host.sh with --custom-ca to regenerate.
+services:
+  gpu:
+    volumes:
+      - ./certs/ca.crt:/usr/local/share/ca-certificates/custom-ca.crt:ro
+  cpu:
+    volumes:
+      - ./certs/ca.crt:/usr/local/share/ca-certificates/custom-ca.crt:ro
+CAEOF
+
+    if [[ -n "$TLS_CERT_PATH" ]]; then
+        cat >> "$ca_override" << 'CADDYCAEOF'
+  caddy:
+    volumes:
+      - ./certs:/etc/caddy/certs:ro
+CADDYCAEOF
+    fi
+    ok "Generated docker-compose.gpu-ca.yml"
+else
+    rm -f "$ROOT_DIR/docker-compose.gpu-ca.yml"
+fi
+
+# --- Generate Caddyfile ---
+if [[ "$USE_CADDY" == "true" ]]; then
+    info "Generating Caddyfile.gpu-host"
+
+    CADDYFILE="$ROOT_DIR/Caddyfile.gpu-host"
+
+    if [[ -n "$TLS_CERT_PATH" ]] && [[ -n "$CUSTOM_DOMAIN" ]]; then
+        cat > "$CADDYFILE" << CADDYEOF
+# Generated by setup-gpu-host.sh — Custom TLS cert for $CUSTOM_DOMAIN
+$CUSTOM_DOMAIN {
+    tls /etc/caddy/certs/server.pem /etc/caddy/certs/server-key.pem
+    reverse_proxy transcription:8000
+}
+CADDYEOF
+        ok "Caddyfile: custom TLS for $CUSTOM_DOMAIN"
+    elif [[ -n "$CUSTOM_DOMAIN" ]]; then
+        cat > "$CADDYFILE" << CADDYEOF
+# Generated by setup-gpu-host.sh — Let's Encrypt for $CUSTOM_DOMAIN
+$CUSTOM_DOMAIN {
+    reverse_proxy transcription:8000
+}
+CADDYEOF
+        ok "Caddyfile: Let's Encrypt for $CUSTOM_DOMAIN"
+    else
+        cat > "$CADDYFILE" << 'CADDYEOF'
+# Generated by setup-gpu-host.sh — self-signed cert for IP access
+:443 {
+    tls internal
+    reverse_proxy transcription:8000
+}
+CADDYEOF
+        ok "Caddyfile: self-signed cert for IP access"
+    fi
+fi
+
+# --- Generate .env ---
+info "Generating GPU service .env"
+
+GPU_ENV="$ROOT_DIR/.env.gpu-host"
+cat > "$GPU_ENV" << EOF
+# Generated by setup-gpu-host.sh
+# HuggingFace token for pyannote diarization models
+HF_TOKEN=${HF_TOKEN:-}
+# API key to protect the GPU service (set via --api-key)
+REFLECTOR_GPU_APIKEY=${API_KEY:-}
+# Port configuration
+GPU_HOST_PORT=${HOST_PORT}
+CADDY_HTTPS_PORT=${HOST_PORT}
+EOF
+
+if [[ -z "${HF_TOKEN:-}" ]]; then
+    warn "HF_TOKEN not set. Diarization requires a HuggingFace token."
+    warn "Set it: export HF_TOKEN=your-token-here and re-run, or edit .env.gpu-host"
+fi
+
+ok "Generated .env.gpu-host"
+
+# --- Build and start ---
+info "Building $GPU_SERVICE image (first build downloads ML models — may take a while)..."
+compose_cmd --env-file "$GPU_ENV" build "$GPU_SERVICE"
+ok "$GPU_SERVICE image built"
+
+info "Starting services..."
+compose_cmd --env-file "$GPU_ENV" up -d
+ok "Services started"
+
+# --- Wait for health ---
+info "Waiting for GPU service to be healthy (model loading takes 1-2 minutes)..."
+local_url="http://localhost:8000"
+for i in $(seq 1 40); do
+    if curl -sf "$local_url/docs" >/dev/null 2>&1; then
+        ok "GPU service is healthy!"
+        break
+    fi
+    if [[ $i -eq 40 ]]; then
+        err "GPU service did not become healthy after 5 minutes."
+        err "Check logs: docker compose -f docker-compose.gpu-host.yml logs gpu"
+        exit 1
+    fi
+    sleep 8
+done
+
+# --- Summary ---
+echo ""
+echo "=========================================="
+echo -e " ${GREEN}GPU service is running!${NC}"
+echo "=========================================="
+echo ""
+
+if [[ "$USE_CADDY" == "true" ]]; then
+    if [[ -n "$CUSTOM_DOMAIN" ]]; then
+        echo "  URL:     https://$CUSTOM_DOMAIN"
+    elif [[ -n "$PRIMARY_IP" ]]; then
+        echo "  URL:     https://$PRIMARY_IP"
+    else
+        echo "  URL:     https://localhost"
+    fi
+else
+    if [[ -n "$PRIMARY_IP" ]]; then
+        echo "  URL:     http://$PRIMARY_IP:$HOST_PORT"
+    else
+        echo "  URL:     http://localhost:$HOST_PORT"
+    fi
+fi
+
+echo "  Health:  curl \$(URL)/docs"
+[[ -n "$API_KEY" ]] && echo "  API key: $API_KEY"
+echo ""
+echo "  Configure the main Reflector instance (in server/.env):"
+echo ""
+
+local_gpu_url=""
+if [[ "$USE_CADDY" == "true" ]]; then
+    if [[ -n "$CUSTOM_DOMAIN" ]]; then
+        local_gpu_url="https://$CUSTOM_DOMAIN"
+    elif [[ -n "$PRIMARY_IP" ]]; then
+        local_gpu_url="https://$PRIMARY_IP"
+    else
+        local_gpu_url="https://localhost"
+    fi
+else
+    if [[ -n "$PRIMARY_IP" ]]; then
+        local_gpu_url="http://$PRIMARY_IP:$HOST_PORT"
+    else
+        local_gpu_url="http://localhost:$HOST_PORT"
+    fi
+fi
+
+echo "    TRANSCRIPT_BACKEND=modal"
+echo "    TRANSCRIPT_URL=$local_gpu_url"
+[[ -n "$API_KEY" ]] && echo "    TRANSCRIPT_MODAL_API_KEY=$API_KEY"
+echo "    DIARIZATION_BACKEND=modal"
+echo "    DIARIZATION_URL=$local_gpu_url"
+[[ -n "$API_KEY" ]] && echo "    DIARIZATION_MODAL_API_KEY=$API_KEY"
+echo "    TRANSLATION_BACKEND=modal"
+echo "    TRANSLATE_URL=$local_gpu_url"
+[[ -n "$API_KEY" ]] && echo "    TRANSLATION_MODAL_API_KEY=$API_KEY"
+echo ""
+
+if [[ "$USE_CUSTOM_CA" == "true" ]]; then
+    echo "  The Reflector instance must also trust this CA."
+    echo "  On the Reflector machine, run setup-selfhosted.sh with:"
+    echo "    --extra-ca /path/to/this-machines-ca.crt"
+    echo ""
+fi
+
+echo "  DNS Resolution:"
+if [[ -n "$CUSTOM_DOMAIN" ]]; then
+    echo "    Ensure '$CUSTOM_DOMAIN' resolves to this machine's IP."
+    echo "    Public: Create a DNS A record."
+    echo "    Internal: Add to /etc/hosts on the Reflector machine:"
+    echo "      ${PRIMARY_IP:-<GPU_IP>} $CUSTOM_DOMAIN"
+else
+    echo "    Use this machine's IP directly in TRANSCRIPT_URL/DIARIZATION_URL."
+fi
+echo ""
+echo "  To stop:   docker compose -f docker-compose.gpu-host.yml down"
+echo "  To re-run: ./scripts/setup-gpu-host.sh $*"
+echo "  Logs:      docker compose -f docker-compose.gpu-host.yml logs -f gpu"
+echo ""

server/Dockerfile

@@ -6,7 +6,7 @@ ENV PYTHONUNBUFFERED=1 \
 
 # builder install base dependencies
 WORKDIR /tmp
-RUN apt-get update && apt-get install -y curl ffmpeg && apt-get clean
+RUN apt-get update && apt-get install -y curl ffmpeg ca-certificates && apt-get clean
 ADD https://astral.sh/uv/install.sh /uv-installer.sh
 RUN sh /uv-installer.sh && rm /uv-installer.sh
 ENV PATH="/root/.local/bin/:$PATH"
@@ -18,7 +18,7 @@ COPY pyproject.toml uv.lock README.md /app/
 RUN uv sync --compile-bytecode --locked
 
 # bootstrap
-COPY alembic.ini runserver.sh /app/
+COPY alembic.ini docker-entrypoint.sh runserver.sh /app/
 COPY images /app/images
 COPY migrations /app/migrations
 COPY reflector /app/reflector
@@ -35,4 +35,6 @@ RUN if [ "$(uname -m)" = "aarch64" ] && [ ! -f /usr/lib/libgomp.so.1 ]; then \
 # Pre-check just to make sure the image will not fail
 RUN uv run python -c "import silero_vad.model"
 
-CMD ["./runserver.sh"]
+RUN chmod +x /app/docker-entrypoint.sh
+
+CMD ["./docker-entrypoint.sh"]

server/docker-entrypoint.sh

@@ -0,0 +1,25 @@
+#!/bin/bash
+set -e
+
+# Custom CA certificate injection
+# If a CA cert is mounted at this path (via docker-compose.ca.yml),
+# add it to the system trust store and configure all Python SSL libraries.
+CUSTOM_CA_PATH="/usr/local/share/ca-certificates/custom-ca.crt"
+
+if [ -s "$CUSTOM_CA_PATH" ]; then
+    echo "[entrypoint] Custom CA certificate detected, updating trust store..."
+    update-ca-certificates 2>/dev/null
+
+    # update-ca-certificates creates a combined bundle (system + custom CAs)
+    COMBINED_BUNDLE="/etc/ssl/certs/ca-certificates.crt"
+    export SSL_CERT_FILE="$COMBINED_BUNDLE"
+    export REQUESTS_CA_BUNDLE="$COMBINED_BUNDLE"
+    export CURL_CA_BUNDLE="$COMBINED_BUNDLE"
+    # Note: GRPC_DEFAULT_SSL_ROOTS_FILE_PATH is intentionally NOT set here.
+    # Setting it causes grpcio to attempt TLS on internal Hatchet connections
+    # that run without TLS (SERVER_GRPC_INSECURE=t), resulting in handshake failures.
+    # If you need gRPC with custom CA, set GRPC_DEFAULT_SSL_ROOTS_FILE_PATH explicitly.
+    echo "[entrypoint] CA trust store updated (SSL_CERT_FILE=$COMBINED_BUNDLE)"
+fi
+
+exec ./runserver.sh

server/docs/daily_pipeline.md

@@ -419,3 +419,18 @@ User-room broadcasts to `user:{user_id}`:
 - `TRANSCRIPT_STATUS`
 - `TRANSCRIPT_FINAL_TITLE`
 - `TRANSCRIPT_DURATION`
+
+## Failed Runs Monitor (Hatchet Cron)
+
+A `FailedRunsMonitor` Hatchet cron workflow runs hourly (`0 * * * *`) and checks for failed pipeline runs
+(DiarizationPipeline, FilePipeline, LivePostProcessingPipeline) in the last hour. For each failed run,
+it renders a DAG status overview and posts it to Zulip.
+
+**Required env vars** (all must be set to enable):
+- `ZULIP_REALM` — Zulip server hostname
+- `ZULIP_API_KEY` — Zulip bot API key
+- `ZULIP_BOT_EMAIL` — Zulip bot email
+- `ZULIP_DAG_STREAM` — Zulip stream for alerts
+- `ZULIP_DAG_TOPIC` — Zulip topic for alerts
+
+If any of these are unset, the monitor workflow is not registered with the Hatchet worker.

server/migrations/versions/501c73a6b0d5_add_soft_delete_fields.py

@@ -0,0 +1,47 @@
+"""add soft delete fields to transcript and recording
+
+Revision ID: 501c73a6b0d5
+Revises: e1f093f7f124
+Create Date: 2026-03-19 00:00:00.000000
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "501c73a6b0d5"
+down_revision: Union[str, None] = "e1f093f7f124"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "transcript",
+        sa.Column("deleted_at", sa.DateTime(timezone=True), nullable=True),
+    )
+    op.add_column(
+        "recording",
+        sa.Column("deleted_at", sa.DateTime(timezone=True), nullable=True),
+    )
+    op.create_index(
+        "idx_transcript_not_deleted",
+        "transcript",
+        ["id"],
+        postgresql_where=sa.text("deleted_at IS NULL"),
+    )
+    op.create_index(
+        "idx_recording_not_deleted",
+        "recording",
+        ["id"],
+        postgresql_where=sa.text("deleted_at IS NULL"),
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("idx_recording_not_deleted", table_name="recording")
+    op.drop_index("idx_transcript_not_deleted", table_name="transcript")
+    op.drop_column("recording", "deleted_at")
+    op.drop_column("transcript", "deleted_at")

server/migrations/versions/a2b3c4d5e6f7_add_email_recipients_to_meeting.py

@@ -0,0 +1,29 @@
+"""add email_recipients to meeting
+
+Revision ID: a2b3c4d5e6f7
+Revises: 501c73a6b0d5
+Create Date: 2026-03-20 00:00:00.000000
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects.postgresql import JSONB
+
+revision: str = "a2b3c4d5e6f7"
+down_revision: Union[str, None] = "501c73a6b0d5"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "meeting",
+        sa.Column("email_recipients", JSONB, nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("meeting", "email_recipients")

server/migrations/versions/b4c7e8f9a012_add_email_transcript_to_to_room.py

@@ -0,0 +1,28 @@
+"""add email_transcript_to to room
+
+Revision ID: b4c7e8f9a012
+Revises: a2b3c4d5e6f7
+Create Date: 2026-03-24 00:00:00.000000
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "b4c7e8f9a012"
+down_revision: Union[str, None] = "a2b3c4d5e6f7"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "room",
+        sa.Column("email_transcript_to", sa.String(), nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("room", "email_transcript_to")

server/migrations/versions/c1d2e3f4a5b6_add_store_video_to_room_and_meeting.py

@@ -0,0 +1,43 @@
+"""add store_video to room and meeting
+
+Revision ID: c1d2e3f4a5b6
+Revises: b4c7e8f9a012
+Create Date: 2026-04-08 00:00:00.000000
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "c1d2e3f4a5b6"
+down_revision: Union[str, None] = "b4c7e8f9a012"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "room",
+        sa.Column(
+            "store_video",
+            sa.Boolean(),
+            nullable=False,
+            server_default=sa.false(),
+        ),
+    )
+    op.add_column(
+        "meeting",
+        sa.Column(
+            "store_video",
+            sa.Boolean(),
+            nullable=False,
+            server_default=sa.false(),
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("meeting", "store_video")
+    op.drop_column("room", "store_video")

server/pyproject.toml

@@ -40,6 +40,9 @@ dependencies = [
     "icalendar>=6.0.0",
     "hatchet-sdk==1.22.16",
     "pydantic>=2.12.5",
+    "aiosmtplib>=3.0.0",
+    "email-validator>=2.0.0",
+    "livekit-api>=1.1.0",
 ]
 
 [dependency-groups]

server/reflector/app.py

@@ -13,18 +13,22 @@ from reflector.events import subscribers_shutdown, subscribers_startup
 from reflector.logger import logger
 from reflector.metrics import metrics_init
 from reflector.settings import settings
+from reflector.views.config import router as config_router
 from reflector.views.daily import router as daily_router
+from reflector.views.livekit import router as livekit_router
 from reflector.views.meetings import router as meetings_router
 from reflector.views.rooms import router as rooms_router
 from reflector.views.rtc_offer import router as rtc_offer_router
 from reflector.views.transcripts import router as transcripts_router
 from reflector.views.transcripts_audio import router as transcripts_audio_router
+from reflector.views.transcripts_download import router as transcripts_download_router
 from reflector.views.transcripts_participants import (
     router as transcripts_participants_router,
 )
 from reflector.views.transcripts_process import router as transcripts_process_router
 from reflector.views.transcripts_speaker import router as transcripts_speaker_router
 from reflector.views.transcripts_upload import router as transcripts_upload_router
+from reflector.views.transcripts_video import router as transcripts_video_router
 from reflector.views.transcripts_webrtc import router as transcripts_webrtc_router
 from reflector.views.transcripts_websocket import router as transcripts_websocket_router
 from reflector.views.user import router as user_router
@@ -97,15 +101,19 @@ app.include_router(transcripts_audio_router, prefix="/v1")
 app.include_router(transcripts_participants_router, prefix="/v1")
 app.include_router(transcripts_speaker_router, prefix="/v1")
 app.include_router(transcripts_upload_router, prefix="/v1")
+app.include_router(transcripts_download_router, prefix="/v1")
+app.include_router(transcripts_video_router, prefix="/v1")
 app.include_router(transcripts_websocket_router, prefix="/v1")
 app.include_router(transcripts_webrtc_router, prefix="/v1")
 app.include_router(transcripts_process_router, prefix="/v1")
 app.include_router(user_router, prefix="/v1")
 app.include_router(user_api_keys_router, prefix="/v1")
 app.include_router(user_ws_router, prefix="/v1")
+app.include_router(config_router, prefix="/v1")
 app.include_router(zulip_router, prefix="/v1")
 app.include_router(whereby_router, prefix="/v1")
 app.include_router(daily_router, prefix="/v1/daily")
+app.include_router(livekit_router, prefix="/v1/livekit")
 if auth_router:
     app.include_router(auth_router, prefix="/v1")
 add_pagination(app)

server/reflector/db/meetings.py

@@ -1,3 +1,4 @@
+from contextlib import asynccontextmanager
 from datetime import datetime, timedelta
 from typing import Any, Literal
 
@@ -66,6 +67,9 @@ meetings = sa.Table(
     # Daily.co composed video (Brady Bunch grid layout) - Daily.co only, not Whereby
     sa.Column("daily_composed_video_s3_key", sa.String, nullable=True),
     sa.Column("daily_composed_video_duration", sa.Integer, nullable=True),
+    # Email recipients for transcript notification
+    sa.Column("email_recipients", JSONB, nullable=True),
+    sa.Column("store_video", sa.Boolean, nullable=False, server_default=sa.false()),
     sa.Index("idx_meeting_room_id", "room_id"),
     sa.Index("idx_meeting_calendar_event", "calendar_event_id"),
 )
@@ -116,6 +120,10 @@ class Meeting(BaseModel):
     # Daily.co composed video (Brady Bunch grid) - Daily.co only
     daily_composed_video_s3_key: str | None = None
     daily_composed_video_duration: int | None = None
+    # Email recipients for transcript notification
+    # Each entry is {"email": str, "include_link": bool} or a legacy plain str
+    email_recipients: list[dict | str] | None = None
+    store_video: bool = False
 
 
 class MeetingController:
@@ -146,6 +154,7 @@ class MeetingController:
             calendar_event_id=calendar_event_id,
             calendar_metadata=calendar_metadata,
             platform=room.platform,
+            store_video=room.store_video,
         )
         query = meetings.insert().values(**meeting.model_dump())
         await get_database().execute(query)
@@ -159,6 +168,17 @@ class MeetingController:
         results = await get_database().fetch_all(query)
         return [Meeting(**result) for result in results]
 
+    async def get_all_inactive_livekit(self) -> list[Meeting]:
+        """Get inactive LiveKit meetings (for multitrack processing discovery)."""
+        query = meetings.select().where(
+            sa.and_(
+                meetings.c.is_active == sa.false(),
+                meetings.c.platform == "livekit",
+            )
+        )
+        results = await get_database().fetch_all(query)
+        return [Meeting(**result) for result in results]
+
     async def get_by_room_name(
         self,
         room_name: str,
@@ -388,6 +408,36 @@ class MeetingController:
         # If was_null=False, the WHERE clause prevented the update
         return was_null
 
+    @asynccontextmanager
+    async def transaction(self):
+        """A context manager for database transaction."""
+        async with get_database().transaction(isolation="serializable"):
+            yield
+
+    async def add_email_recipient(
+        self, meeting_id: str, email: str, *, include_link: bool = True
+    ) -> list[dict]:
+        """Add an email to the meeting's email_recipients list (no duplicates).
+
+        Each entry is stored as {"email": str, "include_link": bool}.
+        Legacy plain-string entries are normalised on read.
+        """
+        async with self.transaction():
+            meeting = await self.get_by_id(meeting_id)
+            if not meeting:
+                raise ValueError(f"Meeting {meeting_id} not found")
+            # Normalise legacy string entries
+            current: list[dict] = [
+                entry
+                if isinstance(entry, dict)
+                else {"email": entry, "include_link": True}
+                for entry in (meeting.email_recipients or [])
+            ]
+            if not any(r["email"] == email for r in current):
+                current.append({"email": email, "include_link": include_link})
+                await self.update_meeting(meeting_id, email_recipients=current)
+        return current
+
     async def increment_num_clients(self, meeting_id: str) -> None:
         """Atomically increment participant count."""
         query = (

server/reflector/db/recordings.py

@@ -1,4 +1,4 @@
-from datetime import datetime
+from datetime import datetime, timezone
 from typing import Literal
 
 import sqlalchemy as sa
@@ -24,6 +24,7 @@ recordings = sa.Table(
     ),
     sa.Column("meeting_id", sa.String),
     sa.Column("track_keys", sa.JSON, nullable=True),
+    sa.Column("deleted_at", sa.DateTime(timezone=True), nullable=True),
     sa.Index("idx_recording_meeting_id", "meeting_id"),
 )
 
@@ -40,6 +41,7 @@ class Recording(BaseModel):
     # track_keys can be empty list [] if recording finished but no audio was captured (silence/muted)
     # None means not a multitrack recording, [] means multitrack with no tracks
     track_keys: list[str] | None = None
+    deleted_at: datetime | None = None
 
     @property
     def is_multitrack(self) -> bool:
@@ -69,6 +71,18 @@ class RecordingController:
         return Recording(**result) if result else None
 
     async def remove_by_id(self, id: str) -> None:
+        query = (
+            recordings.update()
+            .where(recordings.c.id == id)
+            .values(deleted_at=datetime.now(timezone.utc))
+        )
+        await get_database().execute(query)
+
+    async def restore_by_id(self, id: str) -> None:
+        query = recordings.update().where(recordings.c.id == id).values(deleted_at=None)
+        await get_database().execute(query)
+
+    async def hard_delete_by_id(self, id: str) -> None:
         query = recordings.delete().where(recordings.c.id == id)
         await get_database().execute(query)
 
@@ -114,6 +128,7 @@ class RecordingController:
             .where(
                 recordings.c.bucket_name == bucket_name,
                 recordings.c.track_keys.isnot(None),
+                recordings.c.deleted_at.is_(None),
                 or_(
                     transcripts.c.id.is_(None),
                     transcripts.c.status == "error",

server/reflector/db/rooms.py

@@ -63,6 +63,10 @@ rooms = sqlalchemy.Table(
         nullable=False,
         server_default=sqlalchemy.sql.false(),
     ),
+    sqlalchemy.Column("email_transcript_to", sqlalchemy.String, nullable=True),
+    sqlalchemy.Column(
+        "store_video", sqlalchemy.Boolean, nullable=False, server_default=false()
+    ),
     sqlalchemy.Index("idx_room_is_shared", "is_shared"),
     sqlalchemy.Index("idx_room_ics_enabled", "ics_enabled"),
 )
@@ -92,6 +96,8 @@ class Room(BaseModel):
     ics_last_etag: str | None = None
     platform: Platform = Field(default_factory=lambda: settings.DEFAULT_VIDEO_PLATFORM)
     skip_consent: bool = False
+    email_transcript_to: str | None = None
+    store_video: bool = False
 
 
 class RoomController:
@@ -147,6 +153,8 @@ class RoomController:
         ics_enabled: bool = False,
         platform: Platform = settings.DEFAULT_VIDEO_PLATFORM,
         skip_consent: bool = False,
+        email_transcript_to: str | None = None,
+        store_video: bool = False,
     ):
         """
         Add a new room
@@ -172,6 +180,8 @@ class RoomController:
             "ics_enabled": ics_enabled,
             "platform": platform,
             "skip_consent": skip_consent,
+            "email_transcript_to": email_transcript_to,
+            "store_video": store_video,
         }
 
         room = Room(**room_data)

server/reflector/db/search.py

@@ -138,6 +138,7 @@ class SearchParameters(BaseModel):
     source_kind: SourceKind | None = None
     from_datetime: datetime | None = None
     to_datetime: datetime | None = None
+    include_deleted: bool = False
 
 
 class SearchResultDB(BaseModel):
@@ -387,6 +388,11 @@ class SearchController:
             transcripts.join(rooms, transcripts.c.room_id == rooms.c.id, isouter=True)
         )
 
+        if params.include_deleted:
+            base_query = base_query.where(transcripts.c.deleted_at.isnot(None))
+        else:
+            base_query = base_query.where(transcripts.c.deleted_at.is_(None))
+
         if params.query_text is not None:
             # because already initialized based on params.query_text presence above
             assert search_query is not None
@@ -394,7 +400,13 @@ class SearchController:
                 transcripts.c.search_vector_en.op("@@")(search_query)
             )
 
-        if params.user_id:
+        if params.include_deleted:
+            # Trash view: only show user's own deleted transcripts.
+            # Defense-in-depth: require user_id to prevent leaking all users' trash.
+            if not params.user_id:
+                return [], 0
+            base_query = base_query.where(transcripts.c.user_id == params.user_id)
+        elif params.user_id:
             base_query = base_query.where(
                 sqlalchemy.or_(
                     transcripts.c.user_id == params.user_id, rooms.c.is_shared
@@ -419,6 +431,8 @@ class SearchController:
 
         if params.query_text is not None:
             order_by = sqlalchemy.desc(sqlalchemy.text("rank"))
+        elif params.include_deleted:
+            order_by = sqlalchemy.desc(transcripts.c.deleted_at)
         else:
             order_by = sqlalchemy.desc(transcripts.c.created_at)
 

server/reflector/db/transcripts.py

@@ -24,7 +24,7 @@ from reflector.db.utils import is_postgresql
 from reflector.logger import logger
 from reflector.processors.types import Word as ProcessorWord
 from reflector.settings import settings
-from reflector.storage import get_transcripts_storage
+from reflector.storage import get_source_storage, get_transcripts_storage
 from reflector.utils import generate_uuid4
 from reflector.utils.webvtt import topics_to_webvtt
 
@@ -91,6 +91,7 @@ transcripts = sqlalchemy.Table(
     sqlalchemy.Column("webvtt", sqlalchemy.Text),
     # Hatchet workflow run ID for resumption of failed workflows
     sqlalchemy.Column("workflow_run_id", sqlalchemy.String),
+    sqlalchemy.Column("deleted_at", sqlalchemy.DateTime(timezone=True), nullable=True),
     sqlalchemy.Column(
         "change_seq",
         sqlalchemy.BigInteger,
@@ -238,6 +239,7 @@ class Transcript(BaseModel):
     webvtt: str | None = None
     workflow_run_id: str | None = None  # Hatchet workflow run ID for resumption
     change_seq: int | None = None
+    deleted_at: datetime | None = None
 
     @field_serializer("created_at", when_used="json")
     def serialize_datetime(self, dt: datetime) -> str:
@@ -418,6 +420,8 @@ class TranscriptController:
             rooms, transcripts.c.room_id == rooms.c.id, isouter=True
         )
 
+        query = query.where(transcripts.c.deleted_at.is_(None))
+
         if user_id:
             query = query.where(
                 or_(transcripts.c.user_id == user_id, rooms.c.is_shared)
@@ -482,6 +486,14 @@ class TranscriptController:
             return None
         return Transcript(**result)
 
+    async def get_by_meeting_id(self, meeting_id: str) -> Transcript | None:
+        """Get a transcript by meeting_id (first match)."""
+        query = transcripts.select().where(transcripts.c.meeting_id == meeting_id)
+        result = await get_database().fetch_one(query)
+        if not result:
+            return None
+        return Transcript(**result)
+
     async def get_by_recording_id(
         self, recording_id: str, **kwargs
     ) -> Transcript | None:
@@ -500,7 +512,10 @@ class TranscriptController:
         """
         Get transcripts by room_id (direct access without joins)
         """
-        query = transcripts.select().where(transcripts.c.room_id == room_id)
+        query = transcripts.select().where(
+            transcripts.c.room_id == room_id,
+            transcripts.c.deleted_at.is_(None),
+        )
         if "user_id" in kwargs:
             query = query.where(transcripts.c.user_id == kwargs["user_id"])
         if "order_by" in kwargs:
@@ -531,8 +546,11 @@ class TranscriptController:
         if not result:
             raise HTTPException(status_code=404, detail="Transcript not found")
 
-        # if the transcript is anonymous, share mode is not checked
         transcript = Transcript(**result)
+        if transcript.deleted_at is not None:
+            raise HTTPException(status_code=404, detail="Transcript not found")
+
+        # if the transcript is anonymous, share mode is not checked
         if transcript.user_id is None:
             return transcript
 
@@ -632,56 +650,169 @@ class TranscriptController:
         user_id: str | None = None,
     ) -> None:
         """
-        Remove a transcript by id
+        Soft-delete a transcript by id.
+
+        Sets deleted_at on the transcript and its associated recording.
+        All files (S3 and local) are preserved for later retrieval.
         """
         transcript = await self.get_by_id(transcript_id)
         if not transcript:
             return
         if user_id is not None and transcript.user_id != user_id:
             return
+        if transcript.deleted_at is not None:
+            return
+
+        now = datetime.now(timezone.utc)
+
+        # Soft-delete the associated recording (keeps S3 files intact)
+        if transcript.recording_id:
+            try:
+                await recordings_controller.remove_by_id(transcript.recording_id)
+            except Exception as e:
+                logger.warning(
+                    "Failed to soft-delete recording",
+                    exc_info=e,
+                    recording_id=transcript.recording_id,
+                )
+
+        # Soft-delete the transcript (keeps all files intact)
+        query = (
+            transcripts.update()
+            .where(transcripts.c.id == transcript_id)
+            .values(deleted_at=now)
+        )
+        await get_database().execute(query)
+
+    async def restore_by_id(
+        self,
+        transcript_id: str,
+        user_id: str | None = None,
+    ) -> bool:
+        """
+        Restore a soft-deleted transcript by clearing deleted_at.
+
+        Also restores the associated recording if present.
+        Returns True if the transcript was restored, False otherwise.
+        """
+        transcript = await self.get_by_id(transcript_id)
+        if not transcript:
+            return False
+        if transcript.deleted_at is None:
+            return False
+        if user_id is not None and transcript.user_id != user_id:
+            return False
+
+        query = (
+            transcripts.update()
+            .where(transcripts.c.id == transcript_id)
+            .values(deleted_at=None)
+        )
+        await get_database().execute(query)
+
+        if transcript.recording_id:
+            try:
+                await recordings_controller.restore_by_id(transcript.recording_id)
+            except Exception as e:
+                logger.warning(
+                    "Failed to restore recording",
+                    exc_info=e,
+                    recording_id=transcript.recording_id,
+                )
+
+        return True
+
+    async def hard_delete(self, transcript_id: str) -> None:
+        """
+        Permanently delete a transcript, its recording, and all associated files.
+
+        Only deletes transcript-owned resources:
+        - Transcript row and recording row from DB (first, to make data inaccessible)
+        - Transcript audio in S3 storage
+        - Recording files in S3 (both object_key and track_keys, since a recording can have both)
+        - Local files (data_path directory)
+
+        Does NOT delete: meetings, consent records, rooms, or any shared entity.
+        Requires the transcript to be soft-deleted first (deleted_at must be set).
+        """
+        transcript = await self.get_by_id(transcript_id)
+        if not transcript:
+            return
+        if transcript.deleted_at is None:
+            return
+
+        # Collect file references before deleting DB rows
+        recording = None
+        recording_storage = None
+        if transcript.recording_id:
+            recording = await recordings_controller.get_by_id(transcript.recording_id)
+            # Determine the correct storage backend for recording files.
+            # Recordings from different platforms (daily, whereby) live in
+            # platform-specific buckets with separate credentials.
+            if recording and recording.meeting_id:
+                from reflector.db.meetings import meetings_controller  # noqa: PLC0415
+
+                meeting = await meetings_controller.get_by_id(recording.meeting_id)
+                if meeting:
+                    recording_storage = get_source_storage(meeting.platform)
+            if recording_storage is None:
+                recording_storage = get_transcripts_storage()
+
+        # 1. Hard-delete DB rows first (makes data inaccessible immediately)
+        if recording:
+            await recordings_controller.hard_delete_by_id(recording.id)
+        await get_database().execute(
+            transcripts.delete().where(transcripts.c.id == transcript_id)
+        )
+
+        # 2. Delete transcript audio from S3 (always uses transcript storage)
+        transcript_storage = get_transcripts_storage()
         if transcript.audio_location == "storage" and not transcript.audio_deleted:
             try:
-                await get_transcripts_storage().delete_file(
-                    transcript.storage_audio_path
-                )
+                await transcript_storage.delete_file(transcript.storage_audio_path)
             except Exception as e:
                 logger.warning(
                     "Failed to delete transcript audio from storage",
                     exc_info=e,
-                    transcript_id=transcript.id,
+                    transcript_id=transcript_id,
+                    path=transcript.storage_audio_path,
                 )
+
+        # 3. Delete recording files from S3 (both object_key and track_keys —
+        # a recording can have both, unlike consent cleanup which uses elif).
+        # Uses platform-specific storage resolved above.
+        if recording and recording.bucket_name and recording_storage:
+            keys_to_delete = []
+            if recording.track_keys:
+                keys_to_delete = recording.track_keys
+            if recording.object_key:
+                keys_to_delete.append(recording.object_key)
+
+            for key in keys_to_delete:
+                try:
+                    await recording_storage.delete_file(
+                        key, bucket=recording.bucket_name
+                    )
+                except Exception as e:
+                    logger.warning(
+                        "Failed to delete recording file",
+                        exc_info=e,
+                        key=key,
+                        bucket=recording.bucket_name,
+                    )
+
+        # 4. Delete local files
         transcript.unlink()
-        if transcript.recording_id:
-            try:
-                recording = await recordings_controller.get_by_id(
-                    transcript.recording_id
-                )
-                if recording:
-                    try:
-                        await get_transcripts_storage().delete_file(
-                            recording.object_key, bucket=recording.bucket_name
-                        )
-                    except Exception as e:
-                        logger.warning(
-                            "Failed to delete recording object from S3",
-                            exc_info=e,
-                            recording_id=transcript.recording_id,
-                        )
-                    await recordings_controller.remove_by_id(transcript.recording_id)
-            except Exception as e:
-                logger.warning(
-                    "Failed to delete recording row",
-                    exc_info=e,
-                    recording_id=transcript.recording_id,
-                )
-        query = transcripts.delete().where(transcripts.c.id == transcript_id)
-        await get_database().execute(query)
 
     async def remove_by_recording_id(self, recording_id: str):
         """
-        Remove a transcript by recording_id
+        Soft-delete a transcript by recording_id
         """
-        query = transcripts.delete().where(transcripts.c.recording_id == recording_id)
+        query = (
+            transcripts.update()
+            .where(transcripts.c.recording_id == recording_id)
+            .values(deleted_at=datetime.now(timezone.utc))
+        )
         await get_database().execute(query)
 
     @staticmethod

server/reflector/email.py

@@ -0,0 +1,162 @@
+from email.mime.multipart import MIMEMultipart
+from email.mime.text import MIMEText
+from html import escape
+
+import aiosmtplib
+import structlog
+
+from reflector.db.transcripts import SourceKind, Transcript
+from reflector.settings import settings
+from reflector.utils.transcript_formats import transcript_to_text_timestamped
+
+logger = structlog.get_logger(__name__)
+
+
+def is_email_configured() -> bool:
+    return bool(settings.SMTP_HOST and settings.SMTP_FROM_EMAIL)
+
+
+def get_transcript_url(transcript: Transcript) -> str:
+    return f"{settings.UI_BASE_URL}/transcripts/{transcript.id}"
+
+
+def _get_timestamped_text(transcript: Transcript) -> str:
+    """Build the full timestamped transcript text using existing utility."""
+    if not transcript.topics:
+        return ""
+    is_multitrack = transcript.source_kind == SourceKind.ROOM
+    return transcript_to_text_timestamped(
+        transcript.topics, transcript.participants, is_multitrack=is_multitrack
+    )
+
+
+def _build_plain_text(transcript: Transcript, url: str, include_link: bool) -> str:
+    title = transcript.title or "Unnamed recording"
+    lines = [f"Reflector: {title}", ""]
+
+    if transcript.short_summary:
+        lines.extend(["Summary:", transcript.short_summary, ""])
+
+    timestamped = _get_timestamped_text(transcript)
+    if timestamped:
+        lines.extend(["Transcript:", timestamped, ""])
+
+    if include_link:
+        lines.append(f"View transcript: {url}")
+        lines.append("")
+
+    lines.append(
+        "This email was sent because you requested to receive "
+        "the transcript from a meeting."
+    )
+    return "\n".join(lines)
+
+
+def _build_html(transcript: Transcript, url: str, include_link: bool) -> str:
+    title = escape(transcript.title or "Unnamed recording")
+
+    summary_html = ""
+    if transcript.short_summary:
+        summary_html = (
+            f'<p style="color:#555;margin-bottom:16px;">'
+            f"{escape(transcript.short_summary)}</p>"
+        )
+
+    transcript_html = ""
+    timestamped = _get_timestamped_text(transcript)
+    if timestamped:
+        # Build styled transcript lines
+        styled_lines = []
+        for line in timestamped.split("\n"):
+            if not line.strip():
+                continue
+            # Lines are formatted as "[MM:SS] Speaker: text"
+            if line.startswith("[") and "] " in line:
+                bracket_end = line.index("] ")
+                timestamp = escape(line[: bracket_end + 1])
+                rest = line[bracket_end + 2 :]
+                if ": " in rest:
+                    colon_pos = rest.index(": ")
+                    speaker = escape(rest[:colon_pos])
+                    text = escape(rest[colon_pos + 2 :])
+                    styled_lines.append(
+                        f'<div style="margin-bottom:4px;">'
+                        f'<span style="color:#888;font-size:12px;">{timestamp}</span> '
+                        f"<strong>{speaker}:</strong> {text}</div>"
+                    )
+                else:
+                    styled_lines.append(
+                        f'<div style="margin-bottom:4px;">{escape(line)}</div>'
+                    )
+            else:
+                styled_lines.append(
+                    f'<div style="margin-bottom:4px;">{escape(line)}</div>'
+                )
+
+        transcript_html = (
+            '<h3 style="margin-top:20px;margin-bottom:8px;">Transcript</h3>'
+            '<div style="background:#f7f7f7;padding:16px;border-radius:6px;'
+            'font-size:13px;line-height:1.6;max-height:600px;overflow-y:auto;">'
+            f"{''.join(styled_lines)}</div>"
+        )
+
+    link_html = ""
+    if include_link:
+        link_html = (
+            '<p style="margin-top:20px;">'
+            f'<a href="{url}" style="display:inline-block;padding:10px 20px;'
+            "background:#4A90D9;color:#fff;text-decoration:none;"
+            'border-radius:4px;">View Transcript</a></p>'
+        )
+
+    return f"""\
+<div style="font-family:sans-serif;max-width:600px;margin:0 auto;">
+  <h2 style="margin-bottom:4px;">{title}</h2>
+  {summary_html}
+  {transcript_html}
+  {link_html}
+  <p style="color:#999;font-size:12px;margin-top:20px;">This email was sent because you requested to receive the transcript from a meeting.</p>
+</div>"""
+
+
+async def send_transcript_email(
+    to_emails: list[str],
+    transcript: Transcript,
+    *,
+    include_link: bool = True,
+) -> int:
+    """Send transcript notification to all emails. Returns count sent."""
+    if not is_email_configured() or not to_emails:
+        return 0
+
+    url = get_transcript_url(transcript)
+    title = transcript.title or "Unnamed recording"
+    sent = 0
+
+    for email_addr in to_emails:
+        msg = MIMEMultipart("alternative")
+        msg["Subject"] = f"Reflector: {title}"
+        msg["From"] = settings.SMTP_FROM_EMAIL
+        msg["To"] = email_addr
+
+        msg.attach(MIMEText(_build_plain_text(transcript, url, include_link), "plain"))
+        msg.attach(MIMEText(_build_html(transcript, url, include_link), "html"))
+
+        try:
+            await aiosmtplib.send(
+                msg,
+                hostname=settings.SMTP_HOST,
+                port=settings.SMTP_PORT,
+                username=settings.SMTP_USERNAME,
+                password=settings.SMTP_PASSWORD,
+                start_tls=settings.SMTP_USE_TLS,
+            )
+            sent += 1
+        except Exception:
+            logger.exception(
+                "Failed to send transcript email",
+                to=email_addr,
+                transcript_id=transcript.id,
+            )
+
+    return sent

server/reflector/hatchet/constants.py

@@ -21,6 +21,7 @@ class TaskName(StrEnum):
     CLEANUP_CONSENT = "cleanup_consent"
     POST_ZULIP = "post_zulip"
     SEND_WEBHOOK = "send_webhook"
+    SEND_EMAIL = "send_email"
     PAD_TRACK = "pad_track"
     TRANSCRIBE_TRACK = "transcribe_track"
     DETECT_CHUNK_TOPIC = "detect_chunk_topic"
@@ -59,7 +60,13 @@ TIMEOUT_AUDIO = 720  # Audio processing: padding, mixdown (Hatchet execution_tim
 TIMEOUT_AUDIO_HTTP = (
     660  # httpx timeout for pad_track — below 720 so Hatchet doesn't race
 )
-TIMEOUT_HEAVY = 600  # Transcription, fan-out LLM tasks (Hatchet execution_timeout)
+TIMEOUT_HEAVY = 1200  # Transcription, fan-out LLM tasks (Hatchet execution_timeout)
 TIMEOUT_HEAVY_HTTP = (
-    540  # httpx timeout for transcribe_track — below 600 so Hatchet doesn't race
+    1150  # httpx timeout for transcribe_track — below 1200 so Hatchet doesn't race
+)
+TIMEOUT_EXTRA_HEAVY = (
+    3600  # Detect Topics, fan-out LLM tasks (Hatchet execution_timeout)
+)
+TIMEOUT_EXTRA_HEAVY_HTTP = (
+    3400  # httpx timeout for detect_topics — below 3600 so Hatchet doesn't race
 )

server/reflector/hatchet/run_workers_llm.py

@@ -10,12 +10,14 @@ from reflector.hatchet.client import HatchetClientManager
 from reflector.hatchet.workflows.daily_multitrack_pipeline import (
     daily_multitrack_pipeline,
 )
+from reflector.hatchet.workflows.failed_runs_monitor import failed_runs_monitor
 from reflector.hatchet.workflows.file_pipeline import file_pipeline
 from reflector.hatchet.workflows.live_post_pipeline import live_post_pipeline
 from reflector.hatchet.workflows.subject_processing import subject_workflow
 from reflector.hatchet.workflows.topic_chunk_processing import topic_chunk_workflow
 from reflector.hatchet.workflows.track_processing import track_workflow
 from reflector.logger import logger
+from reflector.settings import settings
 
 SLOTS = 10
 WORKER_NAME = "llm-worker-pool"
@@ -34,6 +36,34 @@ def main():
             error=str(e),
         )
 
+    workflows = [
+        daily_multitrack_pipeline,
+        file_pipeline,
+        live_post_pipeline,
+        topic_chunk_workflow,
+        subject_workflow,
+        track_workflow,
+    ]
+
+    _zulip_dag_enabled = all(
+        [
+            settings.ZULIP_REALM,
+            settings.ZULIP_API_KEY,
+            settings.ZULIP_BOT_EMAIL,
+            settings.ZULIP_DAG_STREAM,
+            settings.ZULIP_DAG_TOPIC,
+        ]
+    )
+    if _zulip_dag_enabled:
+        workflows.append(failed_runs_monitor)
+        logger.info(
+            "FailedRunsMonitor cron enabled",
+            stream=settings.ZULIP_DAG_STREAM,
+            topic=settings.ZULIP_DAG_TOPIC,
+        )
+    else:
+        logger.info("FailedRunsMonitor cron disabled (Zulip DAG not configured)")
+
     logger.info(
         "Starting Hatchet LLM worker pool (all tasks except mixdown)",
         worker_name=WORKER_NAME,
@@ -47,14 +77,7 @@ def main():
         labels={
             "pool": POOL,
         },
-        workflows=[
-            daily_multitrack_pipeline,
-            file_pipeline,
-            live_post_pipeline,
-            topic_chunk_workflow,
-            subject_workflow,
-            track_workflow,
-        ],
+        workflows=workflows,
     )
 
     try:

server/reflector/hatchet/workflows/daily_multitrack_pipeline.py

@@ -18,10 +18,11 @@ import json
 import tempfile
 import time
 from contextlib import asynccontextmanager
-from datetime import timedelta
+from datetime import datetime, timedelta
 from pathlib import Path
 from typing import Any, Callable, Coroutine, Protocol, TypeVar
 
+import databases
 import httpx
 from hatchet_sdk import (
     ConcurrencyExpression,
@@ -33,6 +34,7 @@ from hatchet_sdk.labels import DesiredWorkerLabel
 from pydantic import BaseModel
 
 from reflector.dailyco_api.client import DailyApiClient
+from reflector.email import is_email_configured, send_transcript_email
 from reflector.hatchet.broadcast import (
     append_event_and_broadcast,
     set_status_and_broadcast,
@@ -40,6 +42,7 @@ from reflector.hatchet.broadcast import (
 from reflector.hatchet.client import HatchetClientManager
 from reflector.hatchet.constants import (
     TIMEOUT_AUDIO,
+    TIMEOUT_EXTRA_HEAVY,
     TIMEOUT_HEAVY,
     TIMEOUT_LONG,
     TIMEOUT_MEDIUM,
@@ -51,6 +54,7 @@ from reflector.hatchet.error_classification import is_non_retryable
 from reflector.hatchet.workflows.models import (
     ActionItemsResult,
     ConsentResult,
+    EmailResult,
     FinalizeResult,
     MixdownResult,
     PaddedTrackInfo,
@@ -80,9 +84,10 @@ from reflector.hatchet.workflows.topic_chunk_processing import (
     topic_chunk_workflow,
 )
 from reflector.hatchet.workflows.track_processing import TrackInput, track_workflow
+from reflector.llm import LLM
 from reflector.logger import logger
 from reflector.pipelines import topic_processing
-from reflector.processors import AudioFileWriterProcessor
+from reflector.processors.audio_mixdown_auto import AudioMixdownAutoProcessor
 from reflector.processors.summary.models import ActionItemsResponse
 from reflector.processors.summary.prompts import (
     RECAP_PROMPT,
@@ -92,22 +97,28 @@ from reflector.processors.summary.prompts import (
 from reflector.processors.summary.summary_builder import SummaryBuilder
 from reflector.processors.types import TitleSummary, Word
 from reflector.processors.types import Transcript as TranscriptType
+from reflector.redis_cache import get_async_redis_client
 from reflector.settings import settings
+from reflector.storage import get_source_storage, get_transcripts_storage
 from reflector.utils.audio_constants import (
     PRESIGNED_URL_EXPIRATION_SECONDS,
     WAVEFORM_SEGMENTS,
 )
-from reflector.utils.audio_mixdown import (
-    detect_sample_rate_from_tracks,
-    mixdown_tracks_pyav,
-)
 from reflector.utils.audio_waveform import get_audio_waveform
 from reflector.utils.daily import (
     filter_cam_audio_tracks,
     parse_daily_recording_filename,
 )
+from reflector.utils.livekit import parse_livekit_track_filepath
 from reflector.utils.string import NonEmptyString, assert_non_none_and_non_empty
-from reflector.utils.transcript_constants import TOPIC_CHUNK_WORD_COUNT
+from reflector.utils.transcript_constants import (
+    compute_max_subjects,
+    compute_topic_chunk_size,
+)
+from reflector.utils.webhook import (
+    fetch_transcript_webhook_payload,
+    send_webhook_request,
+)
 from reflector.zulip import post_transcript_notification
 
 
@@ -136,8 +147,6 @@ async def fresh_db_connection():
     The real fix would be making the db module fork-aware instead of bypassing it.
     Current pattern is acceptable given Hatchet's process model.
     """
-    import databases  # noqa: PLC0415
-
     from reflector.db import _database_context  # noqa: PLC0415
 
     _database_context.set(None)
@@ -174,8 +183,6 @@ async def set_workflow_error_status(transcript_id: NonEmptyString) -> bool:
 
 def _spawn_storage():
     """Create fresh storage instance for writing to our transcript bucket."""
-    from reflector.storage import get_transcripts_storage  # noqa: PLC0415
-
     return get_transcripts_storage()
 
 
@@ -274,8 +281,10 @@ def with_error_handling(
 )
 @with_error_handling(TaskName.GET_RECORDING)
 async def get_recording(input: PipelineInput, ctx: Context) -> RecordingResult:
-    """Fetch recording metadata from Daily.co API."""
-    ctx.log(f"get_recording: starting for recording_id={input.recording_id}")
+    """Fetch recording metadata. Platform-aware: Daily calls API, LiveKit skips."""
+    ctx.log(
+        f"get_recording: starting for recording_id={input.recording_id}, platform={input.source_platform}"
+    )
     ctx.log(
         f"get_recording: transcript_id={input.transcript_id}, room_id={input.room_id}"
     )
@@ -300,6 +309,18 @@ async def get_recording(input: PipelineInput, ctx: Context) -> RecordingResult:
             )
             ctx.log(f"get_recording: status set to 'processing' and broadcasted")
 
+    # LiveKit: no external API call needed — metadata comes from S3 track listing
+    if input.source_platform == "livekit":
+        ctx.log(
+            "get_recording: LiveKit platform — skipping API call (metadata from S3)"
+        )
+        return RecordingResult(
+            id=input.recording_id,
+            mtg_session_id=None,
+            duration=0,  # Duration calculated from tracks later
+        )
+
+    # Daily.co: fetch recording metadata from API
     if not settings.DAILY_API_KEY:
         ctx.log("get_recording: ERROR - DAILY_API_KEY not configured")
         raise ValueError("DAILY_API_KEY not configured")
@@ -333,11 +354,12 @@ async def get_recording(input: PipelineInput, ctx: Context) -> RecordingResult:
 )
 @with_error_handling(TaskName.GET_PARTICIPANTS)
 async def get_participants(input: PipelineInput, ctx: Context) -> ParticipantsResult:
-    """Fetch participant list from Daily.co API and update transcript in database."""
-    ctx.log(f"get_participants: transcript_id={input.transcript_id}")
+    """Fetch participant list and update transcript. Platform-aware."""
+    ctx.log(
+        f"get_participants: transcript_id={input.transcript_id}, platform={input.source_platform}"
+    )
 
     recording = ctx.task_output(get_recording)
-    mtg_session_id = recording.mtg_session_id
     async with fresh_db_connection():
         from reflector.db.transcripts import (  # noqa: PLC0415
             TranscriptDuration,
@@ -348,8 +370,8 @@ async def get_participants(input: PipelineInput, ctx: Context) -> ParticipantsRe
         transcript = await transcripts_controller.get_by_id(input.transcript_id)
         if not transcript:
             raise ValueError(f"Transcript {input.transcript_id} not found")
-        # Note: title NOT cleared - preserves existing titles
-        # Duration from Daily API (seconds -> milliseconds) - master source
+
+        # Duration from recording metadata (seconds -> milliseconds)
         duration_ms = recording.duration * 1000 if recording.duration else 0
         await transcripts_controller.update(
             transcript,
@@ -361,65 +383,134 @@ async def get_participants(input: PipelineInput, ctx: Context) -> ParticipantsRe
             },
         )
 
-        await append_event_and_broadcast(
-            input.transcript_id,
-            transcript,
-            "DURATION",
-            TranscriptDuration(duration=duration_ms),
-            logger=logger,
-        )
-
-        mtg_session_id = assert_non_none_and_non_empty(
-            mtg_session_id, "mtg_session_id is required"
-        )
-        daily_api_key = assert_non_none_and_non_empty(
-            settings.DAILY_API_KEY, "DAILY_API_KEY is required"
-        )
-
-        async with DailyApiClient(
-            api_key=daily_api_key, base_url=settings.DAILY_API_URL
-        ) as client:
-            participants = await client.get_meeting_participants(mtg_session_id)
-
-        id_to_name = {}
-        id_to_user_id = {}
-        for p in participants.data:
-            if p.user_name:
-                id_to_name[p.participant_id] = p.user_name
-            if p.user_id:
-                id_to_user_id[p.participant_id] = p.user_id
-
-        track_keys = [t["s3_key"] for t in input.tracks]
-        cam_audio_keys = filter_cam_audio_tracks(track_keys)
+        if duration_ms:
+            await append_event_and_broadcast(
+                input.transcript_id,
+                transcript,
+                "DURATION",
+                TranscriptDuration(duration=duration_ms),
+                logger=logger,
+            )
 
         participants_list: list[ParticipantInfo] = []
-        for idx, key in enumerate(cam_audio_keys):
+
+        if input.source_platform == "livekit":
+            # LiveKit: participant identity is in the track dict or can be parsed from filepath
+            # Look up identity → Reflector user_id mapping from Redis
+            # (stored at join time in rooms.py)
+            identity_to_user_id: dict[str, str] = {}
             try:
-                parsed = parse_daily_recording_filename(key)
-                participant_id = parsed.participant_id
-            except ValueError as e:
-                logger.error(
-                    "Failed to parse Daily recording filename",
-                    error=str(e),
-                    key=key,
+                from reflector.db.meetings import (
+                    meetings_controller as mc,  # noqa: PLC0415
                 )
-                continue
 
-            default_name = f"Speaker {idx}"
-            name = id_to_name.get(participant_id, default_name)
-            user_id = id_to_user_id.get(participant_id)
+                meeting = (
+                    await mc.get_by_id(transcript.meeting_id)
+                    if transcript.meeting_id
+                    else None
+                )
+                if meeting:
+                    redis_client = await get_async_redis_client()
+                    mapping_key = f"livekit:participant_map:{meeting.room_name}"
+                    raw_map = await redis_client.hgetall(mapping_key)
+                    identity_to_user_id = {
+                        k.decode() if isinstance(k, bytes) else k: v.decode()
+                        if isinstance(v, bytes)
+                        else v
+                        for k, v in raw_map.items()
+                    }
+                    ctx.log(
+                        f"get_participants: loaded {len(identity_to_user_id)} identity→user_id mappings from Redis"
+                    )
+            except Exception as e:
+                ctx.log(
+                    f"get_participants: could not load identity map from Redis: {e}"
+                )
 
-            participant = TranscriptParticipant(
-                id=participant_id, speaker=idx, name=name, user_id=user_id
-            )
-            await transcripts_controller.upsert_participant(transcript, participant)
-            participants_list.append(
-                ParticipantInfo(
-                    participant_id=participant_id,
-                    user_name=name,
+            for idx, track in enumerate(input.tracks):
+                identity = track.get("participant_identity")
+                if not identity:
+                    # Reprocess path: parse from S3 key
+                    try:
+                        parsed = parse_livekit_track_filepath(track["s3_key"])
+                        identity = parsed.participant_identity
+                    except (ValueError, KeyError):
+                        identity = f"speaker-{idx}"
+
+                # Strip the uuid suffix from identity for display name
+                # e.g., "Juan-2bcea0" → "Juan"
+                display_name = (
+                    identity.rsplit("-", 1)[0] if "-" in identity else identity
+                )
+                reflector_user_id = identity_to_user_id.get(identity)
+
+                participant = TranscriptParticipant(
+                    id=identity,
                     speaker=idx,
+                    name=display_name,
+                    user_id=reflector_user_id,
                 )
+                await transcripts_controller.upsert_participant(transcript, participant)
+                participants_list.append(
+                    ParticipantInfo(
+                        participant_id=identity,
+                        user_name=display_name,
+                        speaker=idx,
+                    )
+                )
+        else:
+            # Daily.co: fetch participant names from API
+            mtg_session_id = recording.mtg_session_id
+            mtg_session_id = assert_non_none_and_non_empty(
+                mtg_session_id, "mtg_session_id is required"
             )
+            daily_api_key = assert_non_none_and_non_empty(
+                settings.DAILY_API_KEY, "DAILY_API_KEY is required"
+            )
+
+            async with DailyApiClient(
+                api_key=daily_api_key, base_url=settings.DAILY_API_URL
+            ) as client:
+                participants = await client.get_meeting_participants(mtg_session_id)
+
+            id_to_name = {}
+            id_to_user_id = {}
+            for p in participants.data:
+                if p.user_name:
+                    id_to_name[p.participant_id] = p.user_name
+                if p.user_id:
+                    id_to_user_id[p.participant_id] = p.user_id
+
+            track_keys = [t["s3_key"] for t in input.tracks]
+            cam_audio_keys = filter_cam_audio_tracks(track_keys)
+
+            for idx, key in enumerate(cam_audio_keys):
+                try:
+                    parsed = parse_daily_recording_filename(key)
+                    participant_id = parsed.participant_id
+                except ValueError as e:
+                    logger.error(
+                        "Failed to parse Daily recording filename",
+                        error=str(e),
+                        key=key,
+                    )
+                    continue
+
+                default_name = f"Speaker {idx}"
+                name = id_to_name.get(participant_id, default_name)
+                user_id = id_to_user_id.get(participant_id)
+
+                participant = TranscriptParticipant(
+                    id=participant_id, speaker=idx, name=name, user_id=user_id
+                )
+                await transcripts_controller.upsert_participant(transcript, participant)
+                participants_list.append(
+                    ParticipantInfo(
+                        participant_id=participant_id,
+                        user_name=name,
+                        speaker=idx,
+                    )
+                )
 
         ctx.log(f"get_participants complete: {len(participants_list)} participants")
 
@@ -434,6 +525,7 @@ async def get_participants(input: PipelineInput, ctx: Context) -> ParticipantsRe
 @daily_multitrack_pipeline.task(
     parents=[get_participants],
     execution_timeout=timedelta(seconds=TIMEOUT_HEAVY),
+    schedule_timeout=timedelta(seconds=TIMEOUT_HEAVY),
     retries=3,
     backoff_factor=2.0,
     backoff_max_seconds=30,
@@ -441,11 +533,60 @@ async def get_participants(input: PipelineInput, ctx: Context) -> ParticipantsRe
 @with_error_handling(TaskName.PROCESS_TRACKS)
 async def process_tracks(input: PipelineInput, ctx: Context) -> ProcessTracksResult:
     """Spawn child workflows for each track (dynamic fan-out)."""
-    ctx.log(f"process_tracks: spawning {len(input.tracks)} track workflows")
+    ctx.log(
+        f"process_tracks: spawning {len(input.tracks)} track workflows, platform={input.source_platform}"
+    )
 
     participants_result = ctx.task_output(get_participants)
     source_language = participants_result.source_language
 
+    # For LiveKit: calculate padding offsets from filename timestamps.
+    # OGG files don't have embedded start_time metadata, so we pre-calculate.
+    track_padding: dict[int, float] = {}
+    if input.source_platform == "livekit":
+        timestamps = []
+        for i, track in enumerate(input.tracks):
+            ts_str = track.get("timestamp")
+            if ts_str:
+                try:
+                    ts = datetime.fromisoformat(ts_str)
+                    timestamps.append((i, ts))
+                except (ValueError, TypeError):
+                    ctx.log(
+                        f"process_tracks: could not parse timestamp for track {i}: {ts_str}"
+                    )
+                    timestamps.append((i, None))
+            else:
+                # Reprocess path: parse timestamp from S3 key
+                try:
+                    parsed = parse_livekit_track_filepath(track["s3_key"])
+                    timestamps.append((i, parsed.timestamp))
+                    ctx.log(
+                        f"process_tracks: parsed timestamp from S3 key for track {i}: {parsed.timestamp}"
+                    )
+                except (ValueError, KeyError):
+                    timestamps.append((i, None))
+
+        valid_timestamps = [(i, ts) for i, ts in timestamps if ts is not None]
+        if valid_timestamps:
+            earliest = min(ts for _, ts in valid_timestamps)
+            # LiveKit Track Egress outputs OGG/Opus files, but the transcription
+            # service only accepts WebM. The padding step converts OGG→WebM as a
+            # side effect of applying the adelay filter. For the earliest track
+            # (offset=0), we use a minimal padding to force this conversion.
+            LIVEKIT_MIN_PADDING_SECONDS = (
+                0.001  # 1ms — inaudible, forces OGG→WebM conversion
+            )
+
+            for i, ts in valid_timestamps:
+                offset = (ts - earliest).total_seconds()
+                if offset == 0.0:
+                    offset = LIVEKIT_MIN_PADDING_SECONDS
+                track_padding[i] = offset
+                ctx.log(
+                    f"process_tracks: track {i} padding={offset}s (from filename timestamp)"
+                )
+
     bulk_runs = [
         track_workflow.create_bulk_run_item(
             input=TrackInput(
@@ -455,6 +596,7 @@ async def process_tracks(input: PipelineInput, ctx: Context) -> ProcessTracksRes
                 transcript_id=input.transcript_id,
                 language=source_language,
                 source_platform=input.source_platform,
+                padding_seconds=track_padding.get(i),
             )
         )
         for i, track in enumerate(input.tracks)
@@ -517,6 +659,7 @@ async def process_tracks(input: PipelineInput, ctx: Context) -> ProcessTracksRes
 @daily_multitrack_pipeline.task(
     parents=[process_tracks],
     execution_timeout=timedelta(seconds=TIMEOUT_AUDIO),
+    schedule_timeout=timedelta(seconds=TIMEOUT_HEAVY),
     retries=2,
     backoff_factor=2.0,
     backoff_max_seconds=15,
@@ -537,7 +680,7 @@ async def process_tracks(input: PipelineInput, ctx: Context) -> ProcessTracksRes
 )
 @with_error_handling(TaskName.MIXDOWN_TRACKS)
 async def mixdown_tracks(input: PipelineInput, ctx: Context) -> MixdownResult:
-    """Mix all padded tracks into single audio file using PyAV (same as Celery)."""
+    """Mix all padded tracks into single audio file via configured backend."""
     ctx.log("mixdown_tracks: mixing padded tracks into single audio file")
 
     track_result = ctx.task_output(process_tracks)
@@ -577,52 +720,66 @@ async def mixdown_tracks(input: PipelineInput, ctx: Context) -> MixdownResult:
     if not valid_urls:
         raise ValueError("No valid padded tracks to mixdown")
 
-    target_sample_rate = detect_sample_rate_from_tracks(valid_urls, logger=logger)
-    if not target_sample_rate:
-        logger.error("Mixdown failed - no decodable audio frames found")
-        raise ValueError("No decodable audio frames in any track")
-
-    output_path = tempfile.mktemp(suffix=".mp3")
-    duration_ms_callback_capture_container = [0.0]
-
-    async def capture_duration(d):
-        duration_ms_callback_capture_container[0] = d
-
-    writer = AudioFileWriterProcessor(path=output_path, on_duration=capture_duration)
-
-    await mixdown_tracks_pyav(
-        valid_urls,
-        writer,
-        target_sample_rate,
-        offsets_seconds=None,
-        logger=logger,
-        progress_callback=make_audio_progress_logger(ctx, TaskName.MIXDOWN_TRACKS),
-        expected_duration_sec=recording_duration if recording_duration > 0 else None,
-    )
-    await writer.flush()
-
-    file_size = Path(output_path).stat().st_size
     storage_path = f"{input.transcript_id}/audio.mp3"
 
-    with open(output_path, "rb") as mixed_file:
-        await storage.put_file(storage_path, mixed_file)
+    # Generate presigned PUT URL for the output (used by modal backend;
+    # pyav backend ignores it and writes locally instead)
+    output_url = await storage.get_file_url(
+        storage_path,
+        operation="put_object",
+        expires_in=PRESIGNED_URL_EXPIRATION_SECONDS,
+    )
 
-    Path(output_path).unlink(missing_ok=True)
+    processor = AudioMixdownAutoProcessor()
+    result = await processor.mixdown_tracks(
+        valid_urls, output_url, offsets_seconds=None
+    )
+
+    if result.output_path:
+        # Pyav backend wrote locally — upload to storage ourselves
+        output_file = Path(result.output_path)
+        with open(output_file, "rb") as mixed_file:
+            await storage.put_file(storage_path, mixed_file)
+        output_file.unlink(missing_ok=True)
+        # Clean up the temp directory the pyav processor created
+        try:
+            output_file.parent.rmdir()
+        except OSError:
+            pass
+    # else: modal backend already uploaded to output_url
 
     async with fresh_db_connection():
-        from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
+        from reflector.db.transcripts import (  # noqa: PLC0415
+            TranscriptDuration,
+            transcripts_controller,
+        )
 
         transcript = await transcripts_controller.get_by_id(input.transcript_id)
         if transcript:
-            await transcripts_controller.update(
-                transcript, {"audio_location": "storage"}
-            )
+            update_data = {"audio_location": "storage"}
+            # Set duration from mixdown if not already set (LiveKit: duration starts at 0)
+            if not transcript.duration or transcript.duration == 0:
+                update_data["duration"] = result.duration_ms
+            await transcripts_controller.update(transcript, update_data)
 
-    ctx.log(f"mixdown_tracks complete: uploaded {file_size} bytes to {storage_path}")
+            # Broadcast duration update if it was missing
+            if not transcript.duration or transcript.duration == 0:
+                await append_event_and_broadcast(
+                    input.transcript_id,
+                    transcript,
+                    "DURATION",
+                    TranscriptDuration(duration=result.duration_ms),
+                    logger=logger,
+                )
+                ctx.log(
+                    f"mixdown_tracks: set duration={result.duration_ms}ms from mixdown"
+                )
+
+    ctx.log(f"mixdown_tracks complete: {result.size} bytes to {storage_path}")
 
     return MixdownResult(
         audio_key=storage_path,
-        duration=duration_ms_callback_capture_container[0],
+        duration=result.duration_ms,
         tracks_mixed=len(valid_urls),
     )
 
@@ -699,7 +856,8 @@ async def generate_waveform(input: PipelineInput, ctx: Context) -> WaveformResul
 
 @daily_multitrack_pipeline.task(
     parents=[process_tracks],
-    execution_timeout=timedelta(seconds=TIMEOUT_HEAVY),
+    execution_timeout=timedelta(seconds=TIMEOUT_EXTRA_HEAVY),
+    schedule_timeout=timedelta(seconds=TIMEOUT_HEAVY),
     retries=3,
     backoff_factor=2.0,
     backoff_max_seconds=30,
@@ -722,7 +880,8 @@ async def detect_topics(input: PipelineInput, ctx: Context) -> TopicsResult:
         transcripts_controller,
     )
 
-    chunk_size = TOPIC_CHUNK_WORD_COUNT
+    duration_seconds = words[-1].end - words[0].start if words else 0
+    chunk_size = compute_topic_chunk_size(duration_seconds, len(words))
     chunks = []
     for i in range(0, len(words), chunk_size):
         chunk_words = words[i : i + chunk_size]
@@ -812,7 +971,7 @@ async def detect_topics(input: PipelineInput, ctx: Context) -> TopicsResult:
 
     ctx.log(f"detect_topics complete: found {len(topics_list)} topics")
 
-    return TopicsResult(topics=topics_list)
+    return TopicsResult(topics=topics_list, duration_seconds=duration_seconds)
 
 
 @daily_multitrack_pipeline.task(
@@ -906,10 +1065,9 @@ async def extract_subjects(input: PipelineInput, ctx: Context) -> SubjectsResult
             participant_name_to_id={},
         )
 
-    # Deferred imports: Hatchet workers fork processes, fresh imports avoid
-    # sharing DB connections and LLM HTTP pools across forks
+    # Deferred DB import: Hatchet workers fork processes, fresh imports avoid
+    # sharing DB connections across forks
     from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
-    from reflector.llm import LLM  # noqa: PLC0415
 
     async with fresh_db_connection():
         transcript = await transcripts_controller.get_by_id(input.transcript_id)
@@ -949,8 +1107,14 @@ async def extract_subjects(input: PipelineInput, ctx: Context) -> SubjectsResult
                 participant_names, participant_name_to_id=participant_name_to_id
             )
 
+        max_subjects = compute_max_subjects(topics_result.duration_seconds)
+        ctx.log(
+            f"extract_subjects: duration={topics_result.duration_seconds:.0f}s, "
+            f"max_subjects={max_subjects}"
+        )
+
         ctx.log("extract_subjects: calling LLM to extract subjects")
-        await builder.extract_subjects()
+        await builder.extract_subjects(max_subjects=max_subjects)
 
         ctx.log(f"extract_subjects complete: {len(builder.subjects)} subjects")
 
@@ -1033,14 +1197,13 @@ async def generate_recap(input: PipelineInput, ctx: Context) -> RecapResult:
     subjects_result = ctx.task_output(extract_subjects)
     process_result = ctx.task_output(process_subjects)
 
-    # Deferred imports: Hatchet workers fork processes, fresh imports avoid
-    # sharing DB connections and LLM HTTP pools across forks
+    # Deferred DB import: Hatchet workers fork processes, fresh imports avoid
+    # sharing DB connections across forks
     from reflector.db.transcripts import (  # noqa: PLC0415
         TranscriptFinalLongSummary,
         TranscriptFinalShortSummary,
         transcripts_controller,
     )
-    from reflector.llm import LLM  # noqa: PLC0415
 
     subject_summaries = process_result.subject_summaries
 
@@ -1129,13 +1292,12 @@ async def identify_action_items(
         ctx.log("identify_action_items: no transcript text, returning empty")
         return ActionItemsResult(action_items=ActionItemsResponse())
 
-    # Deferred imports: Hatchet workers fork processes, fresh imports avoid
-    # sharing DB connections and LLM HTTP pools across forks
+    # Deferred DB import: Hatchet workers fork processes, fresh imports avoid
+    # sharing DB connections across forks
     from reflector.db.transcripts import (  # noqa: PLC0415
         TranscriptActionItems,
         transcripts_controller,
     )
-    from reflector.llm import LLM  # noqa: PLC0415
 
     # TODO: refactor SummaryBuilder methods into standalone functions
     llm = LLM(settings=settings)
@@ -1272,10 +1434,6 @@ async def cleanup_consent(input: PipelineInput, ctx: Context) -> ConsentResult:
         )
         from reflector.db.recordings import recordings_controller  # noqa: PLC0415
         from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
-        from reflector.storage import (  # noqa: PLC0415
-            get_source_storage,
-            get_transcripts_storage,
-        )
 
         transcript = await transcripts_controller.get_by_id(input.transcript_id)
         if not transcript:
@@ -1283,6 +1441,7 @@ async def cleanup_consent(input: PipelineInput, ctx: Context) -> ConsentResult:
             return ConsentResult()
 
         consent_denied = False
+        meeting = None
         if transcript.meeting_id:
             meeting = await meetings_controller.get_by_id(transcript.meeting_id)
             if meeting:
@@ -1345,6 +1504,22 @@ async def cleanup_consent(input: PipelineInput, ctx: Context) -> ConsentResult:
                 logger.error(error_msg, exc_info=True)
                 deletion_errors.append(error_msg)
 
+        # Delete cloud video if present
+        if meeting and meeting.daily_composed_video_s3_key:
+            try:
+                source_storage = get_source_storage("daily")
+                await source_storage.delete_file(meeting.daily_composed_video_s3_key)
+                await meetings_controller.update_meeting(
+                    meeting.id,
+                    daily_composed_video_s3_key=None,
+                    daily_composed_video_duration=None,
+                )
+                ctx.log(f"Deleted cloud video: {meeting.daily_composed_video_s3_key}")
+            except Exception as e:
+                error_msg = f"Failed to delete cloud video: {e}"
+                logger.error(error_msg, exc_info=True)
+                deletion_errors.append(error_msg)
+
         if deletion_errors:
             logger.warning(
                 "[Hatchet] cleanup_consent completed with errors",
@@ -1355,7 +1530,7 @@ async def cleanup_consent(input: PipelineInput, ctx: Context) -> ConsentResult:
             ctx.log(f"cleanup_consent completed with {len(deletion_errors)} errors")
         else:
             await transcripts_controller.update(transcript, {"audio_deleted": True})
-            ctx.log("cleanup_consent: all audio deleted successfully")
+            ctx.log("cleanup_consent: all audio and video deleted successfully")
 
     return ConsentResult()
 
@@ -1407,10 +1582,6 @@ async def send_webhook(input: PipelineInput, ctx: Context) -> WebhookResult:
 
     async with fresh_db_connection():
         from reflector.db.rooms import rooms_controller  # noqa: PLC0415
-        from reflector.utils.webhook import (  # noqa: PLC0415
-            fetch_transcript_webhook_payload,
-            send_webhook_request,
-        )
 
         room = await rooms_controller.get_by_id(input.room_id)
         if not room or not room.webhook_url:
@@ -1465,6 +1636,96 @@ async def send_webhook(input: PipelineInput, ctx: Context) -> WebhookResult:
             return WebhookResult(webhook_sent=False)
 
 
+@daily_multitrack_pipeline.task(
+    parents=[cleanup_consent],
+    execution_timeout=timedelta(seconds=TIMEOUT_SHORT),
+    retries=5,
+    backoff_factor=2.0,
+    backoff_max_seconds=15,
+)
+@with_error_handling(TaskName.SEND_EMAIL, set_error_status=False)
+async def send_email(input: PipelineInput, ctx: Context) -> EmailResult:
+    """Send transcript email to collected recipients."""
+    ctx.log(f"send_email: transcript_id={input.transcript_id}")
+
+    if not is_email_configured():
+        ctx.log("send_email skipped (SMTP not configured)")
+        return EmailResult(skipped=True)
+
+    async with fresh_db_connection():
+        from reflector.db.meetings import meetings_controller  # noqa: PLC0415
+        from reflector.db.recordings import recordings_controller  # noqa: PLC0415
+        from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
+
+        transcript = await transcripts_controller.get_by_id(input.transcript_id)
+        if not transcript:
+            ctx.log("send_email skipped (transcript not found)")
+            return EmailResult(skipped=True)
+
+        meeting = None
+        if transcript.meeting_id:
+            meeting = await meetings_controller.get_by_id(transcript.meeting_id)
+        if not meeting and transcript.recording_id:
+            recording = await recordings_controller.get_by_id(transcript.recording_id)
+            if recording and recording.meeting_id:
+                meeting = await meetings_controller.get_by_id(recording.meeting_id)
+
+        # Normalise meeting recipients (legacy strings → dicts)
+        meeting_recipients: list[dict] = (
+            [
+                entry
+                if isinstance(entry, dict)
+                else {"email": entry, "include_link": True}
+                for entry in (meeting.email_recipients or [])
+            ]
+            if meeting and meeting.email_recipients
+            else []
+        )
+
+        # Room-level email always gets a link (room owner)
+        from reflector.db.rooms import rooms_controller  # noqa: PLC0415
+
+        room_email = None
+        if transcript.room_id:
+            room = await rooms_controller.get_by_id(transcript.room_id)
+            if room and room.email_transcript_to:
+                room_email = room.email_transcript_to
+
+        # Build two groups: with link and without link
+        with_link = [
+            r["email"] for r in meeting_recipients if r.get("include_link", True)
+        ]
+        without_link = [
+            r["email"] for r in meeting_recipients if not r.get("include_link", True)
+        ]
+
+        if room_email:
+            if room_email not in with_link:
+                with_link.append(room_email)
+            without_link = [e for e in without_link if e != room_email]
+
+        if not with_link and not without_link:
+            ctx.log("send_email skipped (no email recipients)")
+            return EmailResult(skipped=True)
+
+        # For room-level emails, do NOT change share_mode (only set public if meeting had recipients)
+        if meeting and meeting.email_recipients:
+            await transcripts_controller.update(transcript, {"share_mode": "public"})
+
+        count = 0
+        if with_link:
+            count += await send_transcript_email(
+                with_link, transcript, include_link=True
+            )
+        if without_link:
+            count += await send_transcript_email(
+                without_link, transcript, include_link=False
+            )
+        ctx.log(f"send_email complete: sent {count} emails")
+
+    return EmailResult(emails_sent=count)
+
+
 async def on_workflow_failure(input: PipelineInput, ctx: Context) -> None:
     """Run when the workflow is truly dead (all retries exhausted).
 

server/reflector/hatchet/workflows/failed_runs_monitor.py

@@ -0,0 +1,109 @@
+"""
+Hatchet cron workflow: FailedRunsMonitor
+
+Runs hourly, queries Hatchet for failed pipeline runs in the last hour,
+and posts details to Zulip for visibility.
+
+Only registered with the worker when Zulip DAG settings are configured.
+"""
+
+from datetime import datetime, timedelta, timezone
+
+from hatchet_sdk import Context
+from hatchet_sdk.clients.rest.models import V1TaskStatus
+
+from reflector.hatchet.client import HatchetClientManager
+from reflector.logger import logger
+from reflector.settings import settings
+from reflector.tools.render_hatchet_run import render_run_detail
+from reflector.zulip import send_message_to_zulip
+
+MONITORED_PIPELINES = {
+    "DiarizationPipeline",
+    "FilePipeline",
+    "LivePostProcessingPipeline",
+}
+
+LOOKBACK_HOURS = 1
+
+hatchet = HatchetClientManager.get_client()
+
+failed_runs_monitor = hatchet.workflow(
+    name="FailedRunsMonitor",
+    on_crons=["0 * * * *"],
+)
+
+
+async def _check_failed_runs() -> dict:
+    """Core logic: query for failed pipeline runs and post each to Zulip.
+
+    Extracted from the Hatchet task for testability.
+    """
+    now = datetime.now(tz=timezone.utc)
+    since = now - timedelta(hours=LOOKBACK_HOURS)
+
+    client = HatchetClientManager.get_client()
+
+    try:
+        result = await client.runs.aio_list(
+            statuses=[V1TaskStatus.FAILED],
+            since=since,
+            until=now,
+            limit=200,
+        )
+    except Exception:
+        logger.exception("[FailedRunsMonitor] Failed to list runs from Hatchet")
+        return {"checked": 0, "reported": 0, "error": "failed to list runs"}
+
+    rows = result.rows or []
+
+    # Filter to main pipelines only (skip child workflows like TrackProcessing, etc.)
+    failed_main_runs = [run for run in rows if run.workflow_name in MONITORED_PIPELINES]
+
+    if not failed_main_runs:
+        logger.info(
+            "[FailedRunsMonitor] No failed pipeline runs in the last hour",
+            total_failed=len(rows),
+            since=since.isoformat(),
+        )
+        return {"checked": len(rows), "reported": 0}
+
+    logger.info(
+        "[FailedRunsMonitor] Found failed pipeline runs",
+        count=len(failed_main_runs),
+        since=since.isoformat(),
+    )
+
+    reported = 0
+    for run in failed_main_runs:
+        try:
+            details = await client.runs.aio_get(run.workflow_run_external_id)
+            content = render_run_detail(details)
+            await send_message_to_zulip(
+                settings.ZULIP_DAG_STREAM,
+                settings.ZULIP_DAG_TOPIC,
+                content,
+            )
+            reported += 1
+        except Exception:
+            logger.exception(
+                "[FailedRunsMonitor] Failed to report run",
+                workflow_run_id=run.workflow_run_external_id,
+                workflow_name=run.workflow_name,
+            )
+
+    logger.info(
+        "[FailedRunsMonitor] Finished reporting",
+        reported=reported,
+        total_failed_main=len(failed_main_runs),
+    )
+    return {"checked": len(rows), "reported": reported}
+
+
+@failed_runs_monitor.task(
+    execution_timeout=timedelta(seconds=120),
+    retries=1,
+)
+async def check_failed_runs(input, ctx: Context) -> dict:
+    """Hatchet task entry point — delegates to _check_failed_runs."""
+    return await _check_failed_runs()

server/reflector/hatchet/workflows/file_pipeline.py

@@ -15,9 +15,12 @@ import json
 from datetime import timedelta
 from pathlib import Path
 
+import av
+import httpx
 from hatchet_sdk import Context
 from pydantic import BaseModel
 
+from reflector.email import is_email_configured, send_transcript_email
 from reflector.hatchet.broadcast import (
     append_event_and_broadcast,
     set_status_and_broadcast,
@@ -37,6 +40,7 @@ from reflector.hatchet.workflows.daily_multitrack_pipeline import (
 )
 from reflector.hatchet.workflows.models import (
     ConsentResult,
+    EmailResult,
     TitleResult,
     TopicsResult,
     WaveformResult,
@@ -45,9 +49,30 @@ from reflector.hatchet.workflows.models import (
 )
 from reflector.logger import logger
 from reflector.pipelines import topic_processing
+from reflector.pipelines.transcription_helpers import transcribe_file_with_processor
+from reflector.processors import AudioFileWriterProcessor
+from reflector.processors.file_diarization import FileDiarizationInput
+from reflector.processors.file_diarization_auto import FileDiarizationAutoProcessor
+from reflector.processors.transcript_diarization_assembler import (
+    TranscriptDiarizationAssemblerInput,
+    TranscriptDiarizationAssemblerProcessor,
+)
+from reflector.processors.types import (
+    DiarizationSegment,
+    Word,
+)
+from reflector.processors.types import (
+    Transcript as TranscriptType,
+)
 from reflector.settings import settings
+from reflector.storage import get_source_storage, get_transcripts_storage
 from reflector.utils.audio_constants import WAVEFORM_SEGMENTS
 from reflector.utils.audio_waveform import get_audio_waveform
+from reflector.utils.webhook import (
+    fetch_transcript_webhook_payload,
+    send_webhook_request,
+)
+from reflector.zulip import post_transcript_notification
 
 
 class FilePipelineInput(BaseModel):
@@ -133,10 +158,6 @@ async def extract_audio(input: FilePipelineInput, ctx: Context) -> ExtractAudioR
         ctx.log(f"extract_audio: processing {audio_file}")
 
         # Extract audio and write as MP3
-        import av  # noqa: PLC0415
-
-        from reflector.processors import AudioFileWriterProcessor  # noqa: PLC0415
-
         duration_ms_container = [0.0]
 
         async def capture_duration(d):
@@ -187,8 +208,6 @@ async def upload_audio(input: FilePipelineInput, ctx: Context) -> UploadAudioRes
     extract_result = ctx.task_output(extract_audio)
     audio_path = extract_result.audio_path
 
-    from reflector.storage import get_transcripts_storage  # noqa: PLC0415
-
     storage = get_transcripts_storage()
     if not storage:
         raise ValueError(
@@ -230,10 +249,6 @@ async def transcribe(input: FilePipelineInput, ctx: Context) -> TranscribeResult
             raise ValueError(f"Transcript {input.transcript_id} not found")
         source_language = transcript.source_language
 
-    from reflector.pipelines.transcription_helpers import (  # noqa: PLC0415
-        transcribe_file_with_processor,
-    )
-
     result = await transcribe_file_with_processor(audio_url, source_language)
 
     ctx.log(f"transcribe complete: {len(result.words)} words")
@@ -262,13 +277,6 @@ async def diarize(input: FilePipelineInput, ctx: Context) -> DiarizeResult:
     upload_result = ctx.task_output(upload_audio)
     audio_url = upload_result.audio_url
 
-    from reflector.processors.file_diarization import (  # noqa: PLC0415
-        FileDiarizationInput,
-    )
-    from reflector.processors.file_diarization_auto import (  # noqa: PLC0415
-        FileDiarizationAutoProcessor,
-    )
-
     processor = FileDiarizationAutoProcessor()
     input_data = FileDiarizationInput(audio_url=audio_url)
 
@@ -351,18 +359,6 @@ async def assemble_transcript(
     transcribe_result = ctx.task_output(transcribe)
     diarize_result = ctx.task_output(diarize)
 
-    from reflector.processors.transcript_diarization_assembler import (  # noqa: PLC0415
-        TranscriptDiarizationAssemblerInput,
-        TranscriptDiarizationAssemblerProcessor,
-    )
-    from reflector.processors.types import (  # noqa: PLC0415
-        DiarizationSegment,
-        Word,
-    )
-    from reflector.processors.types import (  # noqa: PLC0415
-        Transcript as TranscriptType,
-    )
-
     words = [Word(**w) for w in transcribe_result.words]
     transcript_data = TranscriptType(
         words=words, translation=transcribe_result.translation
@@ -435,17 +431,6 @@ async def detect_topics(input: FilePipelineInput, ctx: Context) -> TopicsResult:
         TranscriptTopic,
         transcripts_controller,
     )
-    from reflector.processors.transcript_diarization_assembler import (  # noqa: PLC0415
-        TranscriptDiarizationAssemblerInput,
-        TranscriptDiarizationAssemblerProcessor,
-    )
-    from reflector.processors.types import (  # noqa: PLC0415
-        DiarizationSegment,
-        Word,
-    )
-    from reflector.processors.types import (  # noqa: PLC0415
-        Transcript as TranscriptType,
-    )
 
     words = [Word(**w) for w in transcribe_result.words]
     transcript_data = TranscriptType(
@@ -686,7 +671,6 @@ async def cleanup_consent(input: FilePipelineInput, ctx: Context) -> ConsentResu
         )
         from reflector.db.recordings import recordings_controller  # noqa: PLC0415
         from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
-        from reflector.storage import get_transcripts_storage  # noqa: PLC0415
 
         transcript = await transcripts_controller.get_by_id(input.transcript_id)
         if not transcript:
@@ -695,6 +679,7 @@ async def cleanup_consent(input: FilePipelineInput, ctx: Context) -> ConsentResu
 
         consent_denied = False
         recording = None
+        meeting = None
         if transcript.recording_id:
             recording = await recordings_controller.get_by_id(transcript.recording_id)
             if recording and recording.meeting_id:
@@ -754,6 +739,22 @@ async def cleanup_consent(input: FilePipelineInput, ctx: Context) -> ConsentResu
             logger.error(error_msg, exc_info=True)
             deletion_errors.append(error_msg)
 
+        # Delete cloud video if present
+        if meeting and meeting.daily_composed_video_s3_key:
+            try:
+                source_storage = get_source_storage("daily")
+                await source_storage.delete_file(meeting.daily_composed_video_s3_key)
+                await meetings_controller.update_meeting(
+                    meeting.id,
+                    daily_composed_video_s3_key=None,
+                    daily_composed_video_duration=None,
+                )
+                ctx.log(f"Deleted cloud video: {meeting.daily_composed_video_s3_key}")
+            except Exception as e:
+                error_msg = f"Failed to delete cloud video: {e}"
+                logger.error(error_msg, exc_info=True)
+                deletion_errors.append(error_msg)
+
         if deletion_errors:
             logger.warning(
                 "[Hatchet] cleanup_consent completed with errors",
@@ -762,7 +763,7 @@ async def cleanup_consent(input: FilePipelineInput, ctx: Context) -> ConsentResu
             )
         else:
             await transcripts_controller.update(transcript, {"audio_deleted": True})
-            ctx.log("cleanup_consent: all audio deleted successfully")
+            ctx.log("cleanup_consent: all audio and video deleted successfully")
 
     return ConsentResult()
 
@@ -785,7 +786,6 @@ async def post_zulip(input: FilePipelineInput, ctx: Context) -> ZulipResult:
 
     async with fresh_db_connection():
         from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
-        from reflector.zulip import post_transcript_notification  # noqa: PLC0415
 
         transcript = await transcripts_controller.get_by_id(input.transcript_id)
         if transcript:
@@ -815,10 +815,6 @@ async def send_webhook(input: FilePipelineInput, ctx: Context) -> WebhookResult:
 
     async with fresh_db_connection():
         from reflector.db.rooms import rooms_controller  # noqa: PLC0415
-        from reflector.utils.webhook import (  # noqa: PLC0415
-            fetch_transcript_webhook_payload,
-            send_webhook_request,
-        )
 
         room = await rooms_controller.get_by_id(input.room_id)
         if not room or not room.webhook_url:
@@ -834,8 +830,6 @@ async def send_webhook(input: FilePipelineInput, ctx: Context) -> WebhookResult:
             ctx.log(f"send_webhook skipped (could not build payload): {payload}")
             return WebhookResult(webhook_sent=False, skipped=True)
 
-        import httpx  # noqa: PLC0415
-
         try:
             response = await send_webhook_request(
                 url=room.webhook_url,
@@ -859,6 +853,97 @@ async def send_webhook(input: FilePipelineInput, ctx: Context) -> WebhookResult:
             return WebhookResult(webhook_sent=False)
 
 
+@file_pipeline.task(
+    parents=[cleanup_consent],
+    execution_timeout=timedelta(seconds=TIMEOUT_SHORT),
+    retries=5,
+    backoff_factor=2.0,
+    backoff_max_seconds=15,
+)
+@with_error_handling(TaskName.SEND_EMAIL, set_error_status=False)
+async def send_email(input: FilePipelineInput, ctx: Context) -> EmailResult:
+    """Send transcript email to collected recipients."""
+    ctx.log(f"send_email: transcript_id={input.transcript_id}")
+
+    if not is_email_configured():
+        ctx.log("send_email skipped (SMTP not configured)")
+        return EmailResult(skipped=True)
+
+    async with fresh_db_connection():
+        from reflector.db.meetings import meetings_controller  # noqa: PLC0415
+        from reflector.db.recordings import recordings_controller  # noqa: PLC0415
+        from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
+
+        transcript = await transcripts_controller.get_by_id(input.transcript_id)
+        if not transcript:
+            ctx.log("send_email skipped (transcript not found)")
+            return EmailResult(skipped=True)
+
+        # Try transcript.meeting_id first, then fall back to recording.meeting_id
+        meeting = None
+        if transcript.meeting_id:
+            meeting = await meetings_controller.get_by_id(transcript.meeting_id)
+        if not meeting and transcript.recording_id:
+            recording = await recordings_controller.get_by_id(transcript.recording_id)
+            if recording and recording.meeting_id:
+                meeting = await meetings_controller.get_by_id(recording.meeting_id)
+
+        # Normalise meeting recipients (legacy strings → dicts)
+        meeting_recipients: list[dict] = (
+            [
+                entry
+                if isinstance(entry, dict)
+                else {"email": entry, "include_link": True}
+                for entry in (meeting.email_recipients or [])
+            ]
+            if meeting and meeting.email_recipients
+            else []
+        )
+
+        # Room-level email always gets a link (room owner)
+        from reflector.db.rooms import rooms_controller  # noqa: PLC0415
+
+        room_email = None
+        if transcript.room_id:
+            room = await rooms_controller.get_by_id(transcript.room_id)
+            if room and room.email_transcript_to:
+                room_email = room.email_transcript_to
+
+        # Build two groups: with link and without link
+        with_link = [
+            r["email"] for r in meeting_recipients if r.get("include_link", True)
+        ]
+        without_link = [
+            r["email"] for r in meeting_recipients if not r.get("include_link", True)
+        ]
+
+        if room_email:
+            if room_email not in with_link:
+                with_link.append(room_email)
+            without_link = [e for e in without_link if e != room_email]
+
+        if not with_link and not without_link:
+            ctx.log("send_email skipped (no email recipients)")
+            return EmailResult(skipped=True)
+
+        # For room-level emails, do NOT change share_mode (only set public if meeting had recipients)
+        if meeting and meeting.email_recipients:
+            await transcripts_controller.update(transcript, {"share_mode": "public"})
+
+        count = 0
+        if with_link:
+            count += await send_transcript_email(
+                with_link, transcript, include_link=True
+            )
+        if without_link:
+            count += await send_transcript_email(
+                without_link, transcript, include_link=False
+            )
+        ctx.log(f"send_email complete: sent {count} emails")
+
+    return EmailResult(emails_sent=count)
+
+
 # --- On failure handler ---
 
 

server/reflector/hatchet/workflows/live_post_pipeline.py

@@ -14,9 +14,11 @@ are not shared across forks, avoiding connection pooling issues.
 
 from datetime import timedelta
 
+import httpx
 from hatchet_sdk import Context
 from pydantic import BaseModel
 
+from reflector.email import is_email_configured, send_transcript_email
 from reflector.hatchet.client import HatchetClientManager
 from reflector.hatchet.constants import (
     TIMEOUT_HEAVY,
@@ -32,13 +34,31 @@ from reflector.hatchet.workflows.daily_multitrack_pipeline import (
 )
 from reflector.hatchet.workflows.models import (
     ConsentResult,
+    EmailResult,
     TitleResult,
     WaveformResult,
     WebhookResult,
     ZulipResult,
 )
 from reflector.logger import logger
+from reflector.pipelines.main_live_pipeline import (
+    PipelineMainTitle,
+    PipelineMainWaveform,
+    pipeline_convert_to_mp3,
+    pipeline_diarization,
+    pipeline_remove_upload,
+    pipeline_summaries,
+    pipeline_upload_mp3,
+)
+from reflector.pipelines.main_live_pipeline import (
+    cleanup_consent as _cleanup_consent,
+)
 from reflector.settings import settings
+from reflector.utils.webhook import (
+    fetch_transcript_webhook_payload,
+    send_webhook_request,
+)
+from reflector.zulip import post_transcript_notification
 
 
 class LivePostPipelineInput(BaseModel):
@@ -89,9 +109,6 @@ async def waveform(input: LivePostPipelineInput, ctx: Context) -> WaveformResult
 
     async with fresh_db_connection():
         from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
-        from reflector.pipelines.main_live_pipeline import (  # noqa: PLC0415
-            PipelineMainWaveform,
-        )
 
         transcript = await transcripts_controller.get_by_id(input.transcript_id)
         if not transcript:
@@ -116,10 +133,6 @@ async def generate_title(input: LivePostPipelineInput, ctx: Context) -> TitleRes
     ctx.log(f"generate_title: starting for transcript_id={input.transcript_id}")
 
     async with fresh_db_connection():
-        from reflector.pipelines.main_live_pipeline import (  # noqa: PLC0415
-            PipelineMainTitle,
-        )
-
         runner = PipelineMainTitle(transcript_id=input.transcript_id)
         await runner.run()
 
@@ -140,10 +153,6 @@ async def convert_mp3(input: LivePostPipelineInput, ctx: Context) -> ConvertMp3R
     ctx.log(f"convert_mp3: starting for transcript_id={input.transcript_id}")
 
     async with fresh_db_connection():
-        from reflector.pipelines.main_live_pipeline import (  # noqa: PLC0415
-            pipeline_convert_to_mp3,
-        )
-
         await pipeline_convert_to_mp3(transcript_id=input.transcript_id)
 
     ctx.log("convert_mp3 complete")
@@ -163,10 +172,6 @@ async def upload_mp3(input: LivePostPipelineInput, ctx: Context) -> UploadMp3Res
     ctx.log(f"upload_mp3: starting for transcript_id={input.transcript_id}")
 
     async with fresh_db_connection():
-        from reflector.pipelines.main_live_pipeline import (  # noqa: PLC0415
-            pipeline_upload_mp3,
-        )
-
         await pipeline_upload_mp3(transcript_id=input.transcript_id)
 
     ctx.log("upload_mp3 complete")
@@ -188,10 +193,6 @@ async def remove_upload(
     ctx.log(f"remove_upload: starting for transcript_id={input.transcript_id}")
 
     async with fresh_db_connection():
-        from reflector.pipelines.main_live_pipeline import (  # noqa: PLC0415
-            pipeline_remove_upload,
-        )
-
         await pipeline_remove_upload(transcript_id=input.transcript_id)
 
     ctx.log("remove_upload complete")
@@ -211,10 +212,6 @@ async def diarize(input: LivePostPipelineInput, ctx: Context) -> DiarizeResult:
     ctx.log(f"diarize: starting for transcript_id={input.transcript_id}")
 
     async with fresh_db_connection():
-        from reflector.pipelines.main_live_pipeline import (  # noqa: PLC0415
-            pipeline_diarization,
-        )
-
         await pipeline_diarization(transcript_id=input.transcript_id)
 
     ctx.log("diarize complete")
@@ -234,10 +231,6 @@ async def cleanup_consent(input: LivePostPipelineInput, ctx: Context) -> Consent
     ctx.log(f"cleanup_consent: transcript_id={input.transcript_id}")
 
     async with fresh_db_connection():
-        from reflector.pipelines.main_live_pipeline import (  # noqa: PLC0415
-            cleanup_consent as _cleanup_consent,
-        )
-
         await _cleanup_consent(transcript_id=input.transcript_id)
 
     ctx.log("cleanup_consent complete")
@@ -259,10 +252,6 @@ async def final_summaries(
     ctx.log(f"final_summaries: starting for transcript_id={input.transcript_id}")
 
     async with fresh_db_connection():
-        from reflector.pipelines.main_live_pipeline import (  # noqa: PLC0415
-            pipeline_summaries,
-        )
-
         await pipeline_summaries(transcript_id=input.transcript_id)
 
     ctx.log("final_summaries complete")
@@ -287,7 +276,6 @@ async def post_zulip(input: LivePostPipelineInput, ctx: Context) -> ZulipResult:
 
     async with fresh_db_connection():
         from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
-        from reflector.zulip import post_transcript_notification  # noqa: PLC0415
 
         transcript = await transcripts_controller.get_by_id(input.transcript_id)
         if transcript:
@@ -317,10 +305,6 @@ async def send_webhook(input: LivePostPipelineInput, ctx: Context) -> WebhookRes
 
     async with fresh_db_connection():
         from reflector.db.rooms import rooms_controller  # noqa: PLC0415
-        from reflector.utils.webhook import (  # noqa: PLC0415
-            fetch_transcript_webhook_payload,
-            send_webhook_request,
-        )
 
         room = await rooms_controller.get_by_id(input.room_id)
         if not room or not room.webhook_url:
@@ -336,8 +320,6 @@ async def send_webhook(input: LivePostPipelineInput, ctx: Context) -> WebhookRes
             ctx.log(f"send_webhook skipped (could not build payload): {payload}")
             return WebhookResult(webhook_sent=False, skipped=True)
 
-        import httpx  # noqa: PLC0415
-
         try:
             response = await send_webhook_request(
                 url=room.webhook_url,
@@ -361,6 +343,96 @@ async def send_webhook(input: LivePostPipelineInput, ctx: Context) -> WebhookRes
             return WebhookResult(webhook_sent=False)
 
 
+@live_post_pipeline.task(
+    parents=[final_summaries],
+    execution_timeout=timedelta(seconds=TIMEOUT_SHORT),
+    retries=5,
+    backoff_factor=2.0,
+    backoff_max_seconds=15,
+)
+@with_error_handling(TaskName.SEND_EMAIL, set_error_status=False)
+async def send_email(input: LivePostPipelineInput, ctx: Context) -> EmailResult:
+    """Send transcript email to collected recipients."""
+    ctx.log(f"send_email: transcript_id={input.transcript_id}")
+
+    if not is_email_configured():
+        ctx.log("send_email skipped (SMTP not configured)")
+        return EmailResult(skipped=True)
+
+    async with fresh_db_connection():
+        from reflector.db.meetings import meetings_controller  # noqa: PLC0415
+        from reflector.db.recordings import recordings_controller  # noqa: PLC0415
+        from reflector.db.transcripts import transcripts_controller  # noqa: PLC0415
+
+        transcript = await transcripts_controller.get_by_id(input.transcript_id)
+        if not transcript:
+            ctx.log("send_email skipped (transcript not found)")
+            return EmailResult(skipped=True)
+
+        meeting = None
+        if transcript.meeting_id:
+            meeting = await meetings_controller.get_by_id(transcript.meeting_id)
+        if not meeting and transcript.recording_id:
+            recording = await recordings_controller.get_by_id(transcript.recording_id)
+            if recording and recording.meeting_id:
+                meeting = await meetings_controller.get_by_id(recording.meeting_id)
+
+        # Normalise meeting recipients (legacy strings → dicts)
+        meeting_recipients: list[dict] = (
+            [
+                entry
+                if isinstance(entry, dict)
+                else {"email": entry, "include_link": True}
+                for entry in (meeting.email_recipients or [])
+            ]
+            if meeting and meeting.email_recipients
+            else []
+        )
+
+        # Room-level email always gets a link (room owner)
+        from reflector.db.rooms import rooms_controller  # noqa: PLC0415
+
+        room_email = None
+        if transcript.room_id:
+            room = await rooms_controller.get_by_id(transcript.room_id)
+            if room and room.email_transcript_to:
+                room_email = room.email_transcript_to
+
+        # Build two groups: with link and without link
+        with_link = [
+            r["email"] for r in meeting_recipients if r.get("include_link", True)
+        ]
+        without_link = [
+            r["email"] for r in meeting_recipients if not r.get("include_link", True)
+        ]
+
+        if room_email:
+            if room_email not in with_link:
+                with_link.append(room_email)
+            without_link = [e for e in without_link if e != room_email]
+
+        if not with_link and not without_link:
+            ctx.log("send_email skipped (no email recipients)")
+            return EmailResult(skipped=True)
+
+        # For room-level emails, do NOT change share_mode (only set public if meeting had recipients)
+        if meeting and meeting.email_recipients:
+            await transcripts_controller.update(transcript, {"share_mode": "public"})
+
+        count = 0
+        if with_link:
+            count += await send_transcript_email(
+                with_link, transcript, include_link=True
+            )
+        if without_link:
+            count += await send_transcript_email(
+                without_link, transcript, include_link=False
+            )
+        ctx.log(f"send_email complete: sent {count} emails")
+
+    return EmailResult(emails_sent=count)
+
+
 # --- On failure handler ---
 
 

server/reflector/hatchet/workflows/models.py

@@ -102,6 +102,7 @@ class TopicsResult(BaseModel):
     """Result from detect_topics task."""
 
     topics: list[TitleSummary]
+    duration_seconds: float = 0
 
 
 class TitleResult(BaseModel):
@@ -170,3 +171,10 @@ class WebhookResult(BaseModel):
     webhook_sent: bool
     skipped: bool = False
     response_code: int | None = None
+
+
+class EmailResult(BaseModel):
+    """Result from send_email task."""
+
+    emails_sent: int = 0
+    skipped: bool = False

server/reflector/hatchet/workflows/padding_workflow.py

@@ -13,6 +13,8 @@ from reflector.hatchet.client import HatchetClientManager
 from reflector.hatchet.constants import TIMEOUT_AUDIO
 from reflector.hatchet.workflows.models import PadTrackResult
 from reflector.logger import logger
+from reflector.processors.audio_padding_auto import AudioPaddingAutoProcessor
+from reflector.storage import get_source_storage, get_transcripts_storage
 from reflector.utils.audio_constants import PRESIGNED_URL_EXPIRATION_SECONDS
 from reflector.utils.audio_padding import extract_stream_start_time_from_container
 
@@ -51,11 +53,6 @@ async def pad_track(input: PaddingInput, ctx: Context) -> PadTrackResult:
     )
 
     try:
-        from reflector.storage import (  # noqa: PLC0415
-            get_source_storage,
-            get_transcripts_storage,
-        )
-
         # Source reads: use platform-specific credentials
         source_storage = get_source_storage(input.source_platform)
         source_url = await source_storage.get_file_url(
@@ -104,10 +101,6 @@ async def pad_track(input: PaddingInput, ctx: Context) -> PadTrackResult:
             expires_in=PRESIGNED_URL_EXPIRATION_SECONDS,
         )
 
-        from reflector.processors.audio_padding_auto import (  # noqa: PLC0415
-            AudioPaddingAutoProcessor,
-        )
-
         processor = AudioPaddingAutoProcessor()
         result = await processor.pad_track(
             track_url=source_url,

server/reflector/hatchet/workflows/subject_processing.py

@@ -15,12 +15,14 @@ from pydantic import BaseModel
 from reflector.hatchet.client import HatchetClientManager
 from reflector.hatchet.constants import LLM_RATE_LIMIT_KEY, TIMEOUT_HEAVY
 from reflector.hatchet.workflows.models import SubjectSummaryResult
+from reflector.llm import LLM
 from reflector.logger import logger
 from reflector.processors.summary.prompts import (
     DETAILED_SUBJECT_PROMPT_TEMPLATE,
     PARAGRAPH_SUMMARY_PROMPT,
     build_participant_instructions,
 )
+from reflector.settings import settings
 
 
 class SubjectInput(BaseModel):
@@ -60,11 +62,6 @@ async def generate_detailed_summary(
         subject_index=input.subject_index,
     )
 
-    # Deferred imports: Hatchet workers fork processes, fresh imports ensure
-    # LLM HTTP connection pools aren't shared across forks
-    from reflector.llm import LLM  # noqa: PLC0415
-    from reflector.settings import settings  # noqa: PLC0415
-
     llm = LLM(settings=settings)
 
     participant_instructions = build_participant_instructions(input.participant_names)

server/reflector/hatchet/workflows/topic_chunk_processing.py

@@ -18,9 +18,13 @@ from pydantic import BaseModel
 from reflector.hatchet.client import HatchetClientManager
 from reflector.hatchet.constants import LLM_RATE_LIMIT_KEY, TIMEOUT_MEDIUM
 from reflector.hatchet.workflows.models import TopicChunkResult
+from reflector.llm import LLM
 from reflector.logger import logger
 from reflector.processors.prompts import TOPIC_PROMPT
+from reflector.processors.transcript_topic_detector import TopicResponse
 from reflector.processors.types import Word
+from reflector.settings import settings
+from reflector.utils.text import clean_title
 
 
 class TopicChunkInput(BaseModel):
@@ -64,15 +68,6 @@ async def detect_chunk_topic(input: TopicChunkInput, ctx: Context) -> TopicChunk
         text_length=len(input.chunk_text),
     )
 
-    # Deferred imports: Hatchet workers fork processes, fresh imports avoid
-    # sharing LLM HTTP connection pools across forks
-    from reflector.llm import LLM  # noqa: PLC0415
-    from reflector.processors.transcript_topic_detector import (  # noqa: PLC0415
-        TopicResponse,
-    )
-    from reflector.settings import settings  # noqa: PLC0415
-    from reflector.utils.text import clean_title  # noqa: PLC0415
-
     llm = LLM(settings=settings, temperature=0.9)
 
     prompt = TOPIC_PROMPT.format(text=input.chunk_text)

server/reflector/hatchet/workflows/track_processing.py

@@ -9,9 +9,9 @@ because Hatchet workflow DAGs are defined statically, but the number of tracks v
 at runtime. Child workflow spawning via `aio_run()` + `asyncio.gather()` is the
 standard pattern for dynamic fan-out. See `process_tracks` in daily_multitrack_pipeline.py.
 
-Note: This file uses deferred imports (inside tasks) intentionally.
+Note: DB imports (reflector.db.*) are kept inline (deferred) intentionally.
 Hatchet workers run in forked processes; fresh imports per task ensure
-storage/DB connections are not shared across forks.
+DB connections are not shared across forks.
 """
 
 from datetime import timedelta
@@ -24,6 +24,9 @@ from reflector.hatchet.client import HatchetClientManager
 from reflector.hatchet.constants import TIMEOUT_AUDIO, TIMEOUT_HEAVY
 from reflector.hatchet.workflows.models import PadTrackResult, TranscribeTrackResult
 from reflector.logger import logger
+from reflector.pipelines.transcription_helpers import transcribe_file_with_processor
+from reflector.processors.audio_padding_auto import AudioPaddingAutoProcessor
+from reflector.storage import get_source_storage, get_transcripts_storage
 from reflector.utils.audio_constants import PRESIGNED_URL_EXPIRATION_SECONDS
 from reflector.utils.audio_padding import extract_stream_start_time_from_container
 
@@ -37,6 +40,9 @@ class TrackInput(BaseModel):
     transcript_id: str
     language: str = "en"
     source_platform: str = "daily"
+    # Pre-calculated padding in seconds (from filename timestamps for LiveKit).
+    # When set, overrides container metadata extraction for start_time.
+    padding_seconds: float | None = None
 
 
 hatchet = HatchetClientManager.get_client()
@@ -53,23 +59,22 @@ track_workflow = hatchet.workflow(name="TrackProcessing", input_validator=TrackI
 async def pad_track(input: TrackInput, ctx: Context) -> PadTrackResult:
     """Pad single audio track with silence for alignment.
 
-    Extracts stream.start_time from WebM container metadata and applies
-    silence padding using PyAV filter graph (adelay).
+    For Daily: extracts stream.start_time from WebM container metadata.
+    For LiveKit: uses pre-calculated padding_seconds from filename timestamps
+    (OGG files don't have embedded start_time metadata).
     """
-    ctx.log(f"pad_track: track {input.track_index}, s3_key={input.s3_key}")
+    ctx.log(
+        f"pad_track: track {input.track_index}, s3_key={input.s3_key}, padding_seconds={input.padding_seconds}"
+    )
     logger.info(
         "[Hatchet] pad_track",
         track_index=input.track_index,
         s3_key=input.s3_key,
         transcript_id=input.transcript_id,
+        padding_seconds=input.padding_seconds,
     )
 
     try:
-        from reflector.storage import (  # noqa: PLC0415
-            get_source_storage,
-            get_transcripts_storage,
-        )
-
         # Source reads: use platform-specific credentials
         source_storage = get_source_storage(input.source_platform)
         source_url = await source_storage.get_file_url(
@@ -79,10 +84,16 @@ async def pad_track(input: TrackInput, ctx: Context) -> PadTrackResult:
             bucket=input.bucket_name,
         )
 
-        with av.open(source_url) as in_container:
-            start_time_seconds = extract_stream_start_time_from_container(
-                in_container, input.track_index, logger=logger
-            )
+        if input.padding_seconds is not None:
+            # Pre-calculated offset (LiveKit: from filename timestamps)
+            start_time_seconds = input.padding_seconds
+            ctx.log(f"pad_track: using pre-calculated padding={start_time_seconds}s")
+        else:
+            # Extract from container metadata (Daily: WebM start_time)
+            with av.open(source_url) as in_container:
+                start_time_seconds = extract_stream_start_time_from_container(
+                    in_container, input.track_index, logger=logger
+                )
 
         # If no padding needed, return original S3 key
         if start_time_seconds <= 0:
@@ -107,10 +118,6 @@ async def pad_track(input: TrackInput, ctx: Context) -> PadTrackResult:
             expires_in=PRESIGNED_URL_EXPIRATION_SECONDS,
         )
 
-        from reflector.processors.audio_padding_auto import (  # noqa: PLC0415
-            AudioPaddingAutoProcessor,
-        )
-
         processor = AudioPaddingAutoProcessor()
         result = await processor.pad_track(
             track_url=source_url,
@@ -166,11 +173,6 @@ async def transcribe_track(input: TrackInput, ctx: Context) -> TranscribeTrackRe
             raise ValueError("Missing padded_key from pad_track")
 
         # Presign URL on demand (avoids stale URLs on workflow replay)
-        from reflector.storage import (  # noqa: PLC0415
-            get_source_storage,
-            get_transcripts_storage,
-        )
-
         # If bucket_name is set, file is still in the platform's source bucket (no padding applied).
         # If bucket_name is None, padded file was written to our transcript storage.
         if bucket_name:
@@ -185,10 +187,6 @@ async def transcribe_track(input: TrackInput, ctx: Context) -> TranscribeTrackRe
             bucket=bucket_name,
         )
 
-        from reflector.pipelines.transcription_helpers import (  # noqa: PLC0415
-            transcribe_file_with_processor,
-        )
-
         transcript = await transcribe_file_with_processor(audio_url, input.language)
 
         # Tag all words with speaker index

server/reflector/livekit_api/__init__.py

@@ -0,0 +1,12 @@
+"""
+LiveKit API Module — thin wrapper around the livekit-api SDK.
+"""
+
+from .client import LiveKitApiClient
+from .webhooks import create_webhook_receiver, verify_webhook
+
+__all__ = [
+    "LiveKitApiClient",
+    "create_webhook_receiver",
+    "verify_webhook",
+]

server/reflector/livekit_api/client.py

@@ -0,0 +1,195 @@
+"""
+LiveKit API client wrapping the official livekit-api Python SDK.
+
+Handles room management, access tokens, and Track Egress for
+per-participant audio recording to S3-compatible storage.
+"""
+
+from datetime import timedelta
+
+from livekit.api import (
+    AccessToken,
+    AutoTrackEgress,
+    CreateRoomRequest,
+    DeleteRoomRequest,
+    DirectFileOutput,
+    EgressInfo,
+    ListEgressRequest,
+    ListParticipantsRequest,
+    LiveKitAPI,
+    Room,
+    RoomEgress,
+    S3Upload,
+    StopEgressRequest,
+    TrackEgressRequest,
+    VideoGrants,
+)
+
+
+class LiveKitApiClient:
+    """Thin wrapper around LiveKitAPI for Reflector's needs."""
+
+    def __init__(
+        self,
+        url: str,
+        api_key: str,
+        api_secret: str,
+        s3_bucket: str | None = None,
+        s3_region: str | None = None,
+        s3_access_key: str | None = None,
+        s3_secret_key: str | None = None,
+        s3_endpoint: str | None = None,
+    ):
+        self._url = url
+        self._api_key = api_key
+        self._api_secret = api_secret
+        self._s3_bucket = s3_bucket
+        self._s3_region = s3_region or "us-east-1"
+        self._s3_access_key = s3_access_key
+        self._s3_secret_key = s3_secret_key
+        self._s3_endpoint = s3_endpoint
+        self._api = LiveKitAPI(url=url, api_key=api_key, api_secret=api_secret)
+
+    # ── Room management ──────────────────────────────────────────
+
+    async def create_room(
+        self,
+        name: str,
+        empty_timeout: int = 300,
+        max_participants: int = 0,
+        enable_auto_track_egress: bool = False,
+        track_egress_filepath: str = "livekit/{room_name}/{publisher_identity}-{time}",
+    ) -> Room:
+        """Create a LiveKit room.
+
+        Args:
+            name: Room name (unique identifier).
+            empty_timeout: Seconds to keep room alive after last participant leaves.
+            max_participants: 0 = unlimited.
+            enable_auto_track_egress: If True, automatically record each participant's
+                audio track to S3 as a separate file (OGG/Opus).
+            track_egress_filepath: S3 filepath template for auto track egress.
+                Supports {room_name}, {publisher_identity}, {time}.
+        """
+        egress = None
+        if enable_auto_track_egress:
+            egress = RoomEgress(
+                tracks=AutoTrackEgress(
+                    filepath=track_egress_filepath,
+                    s3=self._build_s3_upload(),
+                ),
+            )
+
+        req = CreateRoomRequest(
+            name=name,
+            empty_timeout=empty_timeout,
+            max_participants=max_participants,
+            egress=egress,
+        )
+        return await self._api.room.create_room(req)
+
+    async def delete_room(self, room_name: str) -> None:
+        await self._api.room.delete_room(DeleteRoomRequest(room=room_name))
+
+    async def list_participants(self, room_name: str):
+        resp = await self._api.room.list_participants(
+            ListParticipantsRequest(room=room_name)
+        )
+        return resp.participants
+
+    # ── Access tokens ────────────────────────────────────────────
+
+    def create_access_token(
+        self,
+        room_name: str,
+        participant_identity: str,
+        participant_name: str | None = None,
+        can_publish: bool = True,
+        can_subscribe: bool = True,
+        room_admin: bool = False,
+        ttl_seconds: int = 86400,
+    ) -> str:
+        """Generate a JWT access token for a participant."""
+        token = AccessToken(
+            api_key=self._api_key,
+            api_secret=self._api_secret,
+        )
+        token.identity = participant_identity
+        token.name = participant_name or participant_identity
+        token.ttl = timedelta(seconds=ttl_seconds)
+        token.with_grants(
+            VideoGrants(
+                room_join=True,
+                room=room_name,
+                can_publish=can_publish,
+                can_subscribe=can_subscribe,
+                room_admin=room_admin,
+            )
+        )
+        return token.to_jwt()
+
+    # ── Track Egress (per-participant audio recording) ───────────
+
+    def _build_s3_upload(self) -> S3Upload:
+        """Build S3Upload config for egress output."""
+        if not all([self._s3_bucket, self._s3_access_key, self._s3_secret_key]):
+            raise ValueError(
+                "S3 storage not configured for LiveKit egress. "
+                "Set LIVEKIT_STORAGE_AWS_* environment variables."
+            )
+        kwargs = {
+            "access_key": self._s3_access_key,
+            "secret": self._s3_secret_key,
+            "bucket": self._s3_bucket,
+            "region": self._s3_region,
+            "force_path_style": True,  # Required for Garage/MinIO
+        }
+        if self._s3_endpoint:
+            kwargs["endpoint"] = self._s3_endpoint
+        return S3Upload(**kwargs)
+
+    async def start_track_egress(
+        self,
+        room_name: str,
+        track_sid: str,
+        s3_filepath: str,
+    ) -> EgressInfo:
+        """Start Track Egress for a single audio track (writes OGG/Opus to S3).
+
+        Args:
+            room_name: LiveKit room name.
+            track_sid: Track SID to record.
+            s3_filepath: S3 key path for the output file.
+        """
+        req = TrackEgressRequest(
+            room_name=room_name,
+            track_id=track_sid,
+            file=DirectFileOutput(
+                filepath=s3_filepath,
+                s3=self._build_s3_upload(),
+            ),
+        )
+        return await self._api.egress.start_track_egress(req)
+
+    async def list_egress(self, room_name: str | None = None) -> list[EgressInfo]:
+        req = ListEgressRequest()
+        if room_name:
+            req.room_name = room_name
+        resp = await self._api.egress.list_egress(req)
+        return list(resp.items)
+
+    async def stop_egress(self, egress_id: str) -> EgressInfo:
+        return await self._api.egress.stop_egress(
+            StopEgressRequest(egress_id=egress_id)
+        )
+
+    # ── Cleanup ──────────────────────────────────────────────────
+
+    async def close(self):
+        await self._api.aclose()
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        await self.close()

server/reflector/livekit_api/webhooks.py

@@ -0,0 +1,52 @@
+"""
+LiveKit webhook verification and event parsing.
+
+LiveKit signs webhooks using the API secret as a JWT.
+The WebhookReceiver from the SDK handles verification.
+"""
+
+from livekit.api import TokenVerifier, WebhookEvent, WebhookReceiver
+
+from reflector.logger import logger
+
+
+def create_webhook_receiver(api_key: str, api_secret: str) -> WebhookReceiver:
+    """Create a WebhookReceiver for verifying LiveKit webhook signatures."""
+    return WebhookReceiver(
+        token_verifier=TokenVerifier(api_key=api_key, api_secret=api_secret)
+    )
+
+
+def verify_webhook(
+    receiver: WebhookReceiver,
+    body: str | bytes,
+    auth_header: str,
+) -> WebhookEvent | None:
+    """Verify and parse a LiveKit webhook event.
+
+    Returns the parsed WebhookEvent if valid, None if verification fails.
+    Logs at different levels depending on failure type:
+    - WARNING: invalid signature, expired token, malformed JWT (expected rejections)
+    - ERROR: unexpected exceptions (potential bugs or attacks)
+    """
+    if isinstance(body, bytes):
+        body = body.decode("utf-8")
+    try:
+        return receiver.receive(body, auth_header)
+    except (ValueError, KeyError) as e:
+        # Expected verification failures (bad JWT, wrong key, expired, malformed)
+        logger.warning(
+            "LiveKit webhook verification failed",
+            error=str(e),
+            error_type=type(e).__name__,
+        )
+        return None
+    except Exception as e:
+        # Unexpected errors — log at ERROR for visibility (potential attack or SDK bug)
+        logger.error(
+            "Unexpected error during LiveKit webhook verification",
+            error=str(e),
+            error_type=type(e).__name__,
+            exc_info=True,
+        )
+        return None

server/reflector/pipelines/main_live_pipeline.py

@@ -38,6 +38,7 @@ from reflector.db.transcripts import (
     TranscriptWaveform,
     transcripts_controller,
 )
+from reflector.hatchet.client import HatchetClientManager
 from reflector.logger import logger
 from reflector.pipelines.runner import PipelineMessage, PipelineRunner
 from reflector.processors import (
@@ -61,7 +62,7 @@ from reflector.processors.types import (
 )
 from reflector.processors.types import Transcript as TranscriptProcessorType
 from reflector.settings import settings
-from reflector.storage import get_transcripts_storage
+from reflector.storage import get_source_storage, get_transcripts_storage
 from reflector.views.transcripts import GetTranscriptTopic
 from reflector.ws_events import TranscriptEventName
 from reflector.ws_manager import WebsocketManager, get_ws_manager
@@ -671,6 +672,22 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
         logger.error(error_msg, exc_info=e)
         deletion_errors.append(error_msg)
 
+    # Delete cloud video if present
+    if meeting and meeting.daily_composed_video_s3_key:
+        try:
+            source_storage = get_source_storage("daily")
+            await source_storage.delete_file(meeting.daily_composed_video_s3_key)
+            await meetings_controller.update_meeting(
+                meeting.id,
+                daily_composed_video_s3_key=None,
+                daily_composed_video_duration=None,
+            )
+            logger.info(f"Deleted cloud video: {meeting.daily_composed_video_s3_key}")
+        except Exception as e:
+            error_msg = f"Failed to delete cloud video: {e}"
+            logger.error(error_msg, exc_info=e)
+            deletion_errors.append(error_msg)
+
     if deletion_errors:
         logger.warning(
             f"Consent cleanup completed with {len(deletion_errors)} errors",
@@ -678,7 +695,7 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
         )
     else:
         await transcripts_controller.update(transcript, {"audio_deleted": True})
-        logger.info("Consent cleanup done - all audio deleted")
+        logger.info("Consent cleanup done - all audio and video deleted")
 
 
 @get_transcript
@@ -798,8 +815,6 @@ async def pipeline_post(*, transcript_id: str, room_id: str | None = None):
     """
     Run the post pipeline via Hatchet.
     """
-    from reflector.hatchet.client import HatchetClientManager  # noqa: PLC0415
-
     await HatchetClientManager.start_workflow(
         "LivePostProcessingPipeline",
         {

server/reflector/pipelines/topic_processing.py

@@ -18,7 +18,7 @@ from reflector.processors import (
 )
 from reflector.processors.types import TitleSummary
 from reflector.processors.types import Transcript as TranscriptType
-from reflector.utils.transcript_constants import TOPIC_CHUNK_WORD_COUNT
+from reflector.utils.transcript_constants import compute_topic_chunk_size
 
 
 class EmptyPipeline:
@@ -39,7 +39,10 @@ async def detect_topics(
     on_topic_callback: Callable,
     empty_pipeline: EmptyPipeline,
 ) -> list[TitleSummary]:
-    chunk_size = TOPIC_CHUNK_WORD_COUNT
+    duration_seconds = (
+        transcript.words[-1].end - transcript.words[0].start if transcript.words else 0
+    )
+    chunk_size = compute_topic_chunk_size(duration_seconds, len(transcript.words))
     topics: list[TitleSummary] = []
 
     async def on_topic(topic: TitleSummary):

server/reflector/processors/__init__.py

@@ -4,6 +4,8 @@ from .audio_diarization_auto import AudioDiarizationAutoProcessor  # noqa: F401
 from .audio_downscale import AudioDownscaleProcessor  # noqa: F401
 from .audio_file_writer import AudioFileWriterProcessor  # noqa: F401
 from .audio_merge import AudioMergeProcessor  # noqa: F401
+from .audio_mixdown import AudioMixdownProcessor  # noqa: F401
+from .audio_mixdown_auto import AudioMixdownAutoProcessor  # noqa: F401
 from .audio_padding import AudioPaddingProcessor  # noqa: F401
 from .audio_padding_auto import AudioPaddingAutoProcessor  # noqa: F401
 from .audio_transcript import AudioTranscriptProcessor  # noqa: F401

server/reflector/processors/audio_mixdown.py

@@ -0,0 +1,27 @@
+"""
+Base class for audio mixdown processors.
+"""
+
+from pydantic import BaseModel
+
+
+class MixdownResponse(BaseModel):
+    size: int
+    duration_ms: float = 0.0
+    cancelled: bool = False
+    output_path: str | None = (
+        None  # Local file path (pyav sets this; modal leaves None)
+    )
+
+
+class AudioMixdownProcessor:
+    """Base class for audio mixdown processors."""
+
+    async def mixdown_tracks(
+        self,
+        track_urls: list[str],
+        output_url: str,
+        target_sample_rate: int | None = None,
+        offsets_seconds: list[float] | None = None,
+    ) -> MixdownResponse:
+        raise NotImplementedError

server/reflector/processors/audio_mixdown_auto.py

@@ -0,0 +1,32 @@
+import importlib
+
+from reflector.processors.audio_mixdown import AudioMixdownProcessor
+from reflector.settings import settings
+
+
+class AudioMixdownAutoProcessor(AudioMixdownProcessor):
+    _registry = {}
+
+    @classmethod
+    def register(cls, name, kclass):
+        cls._registry[name] = kclass
+
+    def __new__(cls, name: str | None = None, **kwargs):
+        if name is None:
+            name = settings.MIXDOWN_BACKEND
+        if name not in cls._registry:
+            module_name = f"reflector.processors.audio_mixdown_{name}"
+            importlib.import_module(module_name)
+
+        # gather specific configuration for the processor
+        # search `MIXDOWN_XXX_YYY`, push to constructor as `xxx_yyy`
+        config = {}
+        name_upper = name.upper()
+        settings_prefix = "MIXDOWN_"
+        config_prefix = f"{settings_prefix}{name_upper}_"
+        for key, value in settings:
+            if key.startswith(config_prefix):
+                config_name = key[len(settings_prefix) :].lower()
+                config[config_name] = value
+
+        return cls._registry[name](**config | kwargs)

server/reflector/processors/audio_mixdown_modal.py

@@ -0,0 +1,110 @@
+"""
+Modal.com backend for audio mixdown.
+"""
+
+import asyncio
+import os
+
+import httpx
+
+from reflector.hatchet.constants import TIMEOUT_HEAVY_HTTP
+from reflector.logger import logger
+from reflector.processors.audio_mixdown import AudioMixdownProcessor, MixdownResponse
+from reflector.processors.audio_mixdown_auto import AudioMixdownAutoProcessor
+
+
+class AudioMixdownModalProcessor(AudioMixdownProcessor):
+    """Audio mixdown processor using Modal.com/self-hosted backend via HTTP."""
+
+    def __init__(
+        self, mixdown_url: str | None = None, modal_api_key: str | None = None
+    ):
+        self.mixdown_url = mixdown_url or os.getenv("MIXDOWN_URL")
+        if not self.mixdown_url:
+            raise ValueError(
+                "MIXDOWN_URL required to use AudioMixdownModalProcessor. "
+                "Set MIXDOWN_URL environment variable or pass mixdown_url parameter."
+            )
+
+        self.modal_api_key = modal_api_key or os.getenv("MODAL_API_KEY")
+
+    async def mixdown_tracks(
+        self,
+        track_urls: list[str],
+        output_url: str,
+        target_sample_rate: int | None = None,
+        offsets_seconds: list[float] | None = None,
+    ) -> MixdownResponse:
+        """Mix audio tracks via remote Modal/self-hosted backend.
+
+        Args:
+            track_urls: Presigned GET URLs for source audio tracks
+            output_url: Presigned PUT URL for output MP3
+            target_sample_rate: Sample rate for output (Hz), auto-detected if None
+            offsets_seconds: Optional per-track delays in seconds for alignment
+        """
+        valid_count = len([u for u in track_urls if u])
+        log = logger.bind(track_count=valid_count)
+        log.info("Sending Modal mixdown HTTP request")
+
+        url = f"{self.mixdown_url}/mixdown"
+
+        headers = {}
+        if self.modal_api_key:
+            headers["Authorization"] = f"Bearer {self.modal_api_key}"
+
+        # Scale timeout with track count: base TIMEOUT_HEAVY_HTTP + 60s per track beyond 2
+        extra_timeout = max(0, (valid_count - 2)) * 60
+        timeout = TIMEOUT_HEAVY_HTTP + extra_timeout
+
+        try:
+            async with httpx.AsyncClient(timeout=timeout) as client:
+                response = await client.post(
+                    url,
+                    headers=headers,
+                    json={
+                        "track_urls": track_urls,
+                        "output_url": output_url,
+                        "target_sample_rate": target_sample_rate,
+                        "offsets_seconds": offsets_seconds,
+                    },
+                    follow_redirects=True,
+                )
+
+                if response.status_code != 200:
+                    error_body = response.text
+                    log.error(
+                        "Modal mixdown API error",
+                        status_code=response.status_code,
+                        error_body=error_body,
+                    )
+
+                response.raise_for_status()
+                result = response.json()
+
+            # Check if work was cancelled
+            if result.get("cancelled"):
+                log.warning("Modal mixdown was cancelled by disconnect detection")
+                raise asyncio.CancelledError(
+                    "Mixdown cancelled due to client disconnect"
+                )
+
+            log.info("Modal mixdown complete", size=result["size"])
+            return MixdownResponse(**result)
+        except asyncio.CancelledError:
+            log.warning(
+                "Modal mixdown cancelled (Hatchet timeout, disconnect detected on Modal side)"
+            )
+            raise
+        except httpx.TimeoutException as e:
+            log.error("Modal mixdown timeout", error=str(e), exc_info=True)
+            raise Exception(f"Modal mixdown timeout: {e}") from e
+        except httpx.HTTPStatusError as e:
+            log.error("Modal mixdown HTTP error", error=str(e), exc_info=True)
+            raise Exception(f"Modal mixdown HTTP error: {e}") from e
+        except Exception as e:
+            log.error("Modal mixdown unexpected error", error=str(e), exc_info=True)
+            raise
+
+
+AudioMixdownAutoProcessor.register("modal", AudioMixdownModalProcessor)

server/reflector/processors/audio_mixdown_pyav.py

@@ -0,0 +1,101 @@
+"""
+PyAV audio mixdown processor.
+
+Mixes N tracks in-process using the existing utility from reflector.utils.audio_mixdown.
+Writes to a local temp file (does NOT upload to S3 — the pipeline handles upload).
+"""
+
+import os
+import tempfile
+
+from reflector.logger import logger
+from reflector.processors.audio_file_writer import AudioFileWriterProcessor
+from reflector.processors.audio_mixdown import AudioMixdownProcessor, MixdownResponse
+from reflector.processors.audio_mixdown_auto import AudioMixdownAutoProcessor
+from reflector.utils.audio_mixdown import (
+    detect_sample_rate_from_tracks,
+    mixdown_tracks_pyav,
+)
+
+
+class AudioMixdownPyavProcessor(AudioMixdownProcessor):
+    """Audio mixdown processor using PyAV (no HTTP backend).
+
+    Writes the mixed output to a local temp file and returns its path
+    in MixdownResponse.output_path. The caller is responsible for
+    uploading the file and cleaning it up.
+    """
+
+    async def mixdown_tracks(
+        self,
+        track_urls: list[str],
+        output_url: str,
+        target_sample_rate: int | None = None,
+        offsets_seconds: list[float] | None = None,
+    ) -> MixdownResponse:
+        log = logger.bind(track_count=len(track_urls))
+        log.info("Starting local PyAV mixdown")
+
+        valid_urls = [url for url in track_urls if url]
+        if not valid_urls:
+            raise ValueError("No valid track URLs provided")
+
+        # Auto-detect sample rate if not provided
+        if target_sample_rate is None:
+            target_sample_rate = detect_sample_rate_from_tracks(
+                valid_urls, logger=logger
+            )
+            if not target_sample_rate:
+                raise ValueError("No decodable audio frames in any track")
+
+        # Write to temp MP3 file
+        temp_dir = tempfile.mkdtemp()
+        output_path = os.path.join(temp_dir, "mixed.mp3")
+        duration_ms_container = [0.0]
+
+        async def capture_duration(d):
+            duration_ms_container[0] = d
+
+        writer = AudioFileWriterProcessor(
+            path=output_path, on_duration=capture_duration
+        )
+
+        try:
+            await mixdown_tracks_pyav(
+                valid_urls,
+                writer,
+                target_sample_rate,
+                offsets_seconds=offsets_seconds,
+                logger=logger,
+            )
+            await writer.flush()
+
+            file_size = os.path.getsize(output_path)
+            log.info(
+                "Local mixdown complete",
+                size=file_size,
+                duration_ms=duration_ms_container[0],
+            )
+
+            return MixdownResponse(
+                size=file_size,
+                duration_ms=duration_ms_container[0],
+                output_path=output_path,
+            )
+
+        except Exception as e:
+            # Cleanup on failure
+            if os.path.exists(output_path):
+                try:
+                    os.unlink(output_path)
+                except Exception:
+                    pass
+            try:
+                os.rmdir(temp_dir)
+            except Exception:
+                pass
+            log.error("Local mixdown failed", error=str(e), exc_info=True)
+            raise
+
+
+AudioMixdownAutoProcessor.register("pyav", AudioMixdownPyavProcessor)

server/reflector/processors/audio_padding_pyav.py

@@ -10,6 +10,7 @@ import os
 import tempfile
 
 import av
+import requests
 
 from reflector.logger import logger
 from reflector.processors.audio_padding import AudioPaddingProcessor, PaddingResponse
@@ -65,8 +66,6 @@ class AudioPaddingPyavProcessor(AudioPaddingProcessor):
         track_index: int,
     ) -> PaddingResponse:
         """Blocking padding work: download, pad with PyAV, upload."""
-        import requests
-
         log = logger.bind(track_index=track_index, padding_seconds=start_time_seconds)
         temp_dir = tempfile.mkdtemp()
         input_path = None

server/reflector/processors/summary/prompts.py

@@ -43,7 +43,8 @@ DETAILED_SUBJECT_PROMPT_TEMPLATE = dedent(
       include any deadlines or timeframes discussed for completion or follow-up.
     - Mention unresolved issues or topics needing further discussion, aiding in
       planning future meetings or follow-up actions.
-    - Do not include topic unrelated to {subject}.
+    - Be specific and cite participant names when attributing statements or actions.
+    - Do not include topics unrelated to {subject}.
 
     # OUTPUT
     Your summary should be clear, concise, and structured, covering all major
@@ -58,6 +59,7 @@ PARAGRAPH_SUMMARY_PROMPT = dedent(
     """
     Summarize the mentioned topic in 1 paragraph.
     It will be integrated into the final summary, so just for this topic.
+    Preserve key decisions and action items. Do not introduce new information.
     """
 ).strip()
 

server/reflector/processors/summary/summary_builder.py

@@ -48,17 +48,24 @@ TRANSCRIPTION_TYPE_PROMPT = dedent(
     """
 ).strip()
 
-SUBJECTS_PROMPT = dedent(
-    """
-    What are the main / high level topic of the meeting.
-    Do not include direct quotes or unnecessary details.
-    Be concise and focused on the main ideas.
-    A subject briefly mentioned should not be included.
-    There should be maximum 6 subjects.
-    Do not write complete narrative sentences for the subject,
-    you must write a concise subject using noun phrases.
-    """
-).strip()
+_DEFAULT_MAX_SUBJECTS = 6
+
+
+def build_subjects_prompt(max_subjects: int = _DEFAULT_MAX_SUBJECTS) -> str:
+    """Build subjects extraction prompt with a dynamic subject cap."""
+    subject_word = "subject" if max_subjects == 1 else "subjects"
+    return dedent(
+        f"""
+        What are the main / high level topics of the meeting.
+        Do not include direct quotes or unnecessary details.
+        Be concise and focused on the main ideas.
+        A subject briefly mentioned should not be included.
+        There should be maximum {max_subjects} {subject_word}.
+        Do not write complete narrative sentences for the subject,
+        you must write a concise subject using noun phrases.
+        """
+    ).strip()
+
 
 ACTION_ITEMS_PROMPT = dedent(
     """
@@ -145,7 +152,7 @@ class SubjectsResponse(BaseModel):
     """Pydantic model for extracted subjects/topics"""
 
     subjects: list[str] = Field(
-        description="List of main subjects/topics discussed, maximum 6 items",
+        description="List of main subjects/topics discussed",
     )
 
 
@@ -345,11 +352,14 @@ class SummaryBuilder:
     # Summary
     # ----------------------------------------------------------------------------
 
-    async def extract_subjects(self) -> None:
+    async def extract_subjects(self, max_subjects: int = _DEFAULT_MAX_SUBJECTS) -> None:
         """Extract main subjects/topics from the transcript."""
-        self.logger.info("--- extract main subjects using TreeSummarize")
+        self.logger.info(
+            "--- extract main subjects using TreeSummarize",
+            max_subjects=max_subjects,
+        )
 
-        subjects_prompt = SUBJECTS_PROMPT
+        subjects_prompt = build_subjects_prompt(max_subjects)
 
         try:
             response = await self._get_structured_response(
@@ -358,7 +368,7 @@ class SummaryBuilder:
                 tone_name="Meeting assistant that talk only as list item",
             )
 
-            self.subjects = response.subjects
+            self.subjects = response.subjects[:max_subjects]
             self.logger.info(f"Extracted subjects: {self.subjects}")
 
         except Exception as e:

server/reflector/schemas/platform.py

@@ -1,5 +1,6 @@
 from typing import Literal
 
-Platform = Literal["whereby", "daily"]
+Platform = Literal["whereby", "daily", "livekit"]
 WHEREBY_PLATFORM: Platform = "whereby"
 DAILY_PLATFORM: Platform = "daily"
+LIVEKIT_PLATFORM: Platform = "livekit"

server/reflector/services/transcript_process.py

@@ -155,12 +155,17 @@ async def prepare_transcript_processing(validation: ValidationOk) -> PrepareResu
                 )
 
     if track_keys:
+        # Detect platform from recording ID prefix
+        source_platform = (
+            "livekit" if recording_id and recording_id.startswith("lk-") else "daily"
+        )
         return MultitrackProcessingConfig(
             bucket_name=bucket_name,  # type: ignore (validated above)
             track_keys=track_keys,
             transcript_id=validation.transcript_id,
             recording_id=recording_id,
             room_id=validation.room_id,
+            source_platform=source_platform,
         )
 
     return FileProcessingConfig(

server/reflector/settings.py

@@ -127,6 +127,14 @@ class Settings(BaseSettings):
     PADDING_URL: str | None = None
     PADDING_MODAL_API_KEY: str | None = None
 
+    # Audio Mixdown
+    # backends:
+    #   - pyav: in-process PyAV mixdown (no HTTP, runs in same process)
+    #   - modal: HTTP API client (works with Modal.com OR self-hosted gpu/self_hosted/)
+    MIXDOWN_BACKEND: str = "pyav"
+    MIXDOWN_URL: str | None = None
+    MIXDOWN_MODAL_API_KEY: str | None = None
+
     # Sentry
     SENTRY_DSN: str | None = None
 
@@ -187,6 +195,23 @@ class Settings(BaseSettings):
     DAILY_WEBHOOK_UUID: str | None = (
         None  # Webhook UUID for this environment. Not used by production code
     )
+
+    # LiveKit integration (self-hosted open-source video platform)
+    LIVEKIT_URL: str | None = (
+        None  # e.g. ws://livekit:7880 (internal) or wss://livekit.example.com
+    )
+    LIVEKIT_API_KEY: str | None = None
+    LIVEKIT_API_SECRET: str | None = None
+    LIVEKIT_WEBHOOK_SECRET: str | None = None  # Defaults to API_SECRET if not set
+    # LiveKit egress S3 storage (Track Egress writes per-participant audio here)
+    LIVEKIT_STORAGE_AWS_BUCKET_NAME: str | None = None
+    LIVEKIT_STORAGE_AWS_REGION: str | None = None
+    LIVEKIT_STORAGE_AWS_ACCESS_KEY_ID: str | None = None
+    LIVEKIT_STORAGE_AWS_SECRET_ACCESS_KEY: str | None = None
+    LIVEKIT_STORAGE_AWS_ENDPOINT_URL: str | None = None  # For Garage/MinIO
+    # Public URL for LiveKit (used in frontend room_url, e.g. wss://livekit.example.com)
+    LIVEKIT_PUBLIC_URL: str | None = None
+
     # Platform Configuration
     DEFAULT_VIDEO_PLATFORM: Platform = DAILY_PLATFORM
 
@@ -194,6 +219,16 @@ class Settings(BaseSettings):
     ZULIP_REALM: str | None = None
     ZULIP_API_KEY: str | None = None
     ZULIP_BOT_EMAIL: str | None = None
+    ZULIP_DAG_STREAM: str | None = None
+    ZULIP_DAG_TOPIC: str | None = None
+
+    # Email / SMTP integration (for transcript email notifications)
+    SMTP_HOST: str | None = None
+    SMTP_PORT: int = 587
+    SMTP_USERNAME: str | None = None
+    SMTP_PASSWORD: str | None = None
+    SMTP_FROM_EMAIL: str | None = None
+    SMTP_USE_TLS: bool = True
 
     # Hatchet workflow orchestration (always enabled for multitrack processing)
     HATCHET_CLIENT_TOKEN: str | None = None

server/reflector/storage/__init__.py

@@ -57,6 +57,22 @@ def get_source_storage(platform: str) -> Storage:
                 aws_secret_access_key=settings.WHEREBY_STORAGE_AWS_SECRET_ACCESS_KEY,
             )
 
+    elif platform == "livekit":
+        if (
+            settings.LIVEKIT_STORAGE_AWS_ACCESS_KEY_ID
+            and settings.LIVEKIT_STORAGE_AWS_SECRET_ACCESS_KEY
+            and settings.LIVEKIT_STORAGE_AWS_BUCKET_NAME
+        ):
+            from reflector.storage.storage_aws import AwsStorage
+
+            return AwsStorage(
+                aws_bucket_name=settings.LIVEKIT_STORAGE_AWS_BUCKET_NAME,
+                aws_region=settings.LIVEKIT_STORAGE_AWS_REGION or "us-east-1",
+                aws_access_key_id=settings.LIVEKIT_STORAGE_AWS_ACCESS_KEY_ID,
+                aws_secret_access_key=settings.LIVEKIT_STORAGE_AWS_SECRET_ACCESS_KEY,
+                aws_endpoint_url=settings.LIVEKIT_STORAGE_AWS_ENDPOINT_URL,
+            )
+
     return get_transcripts_storage()
 
 

server/reflector/storage/base.py

@@ -116,9 +116,12 @@ class Storage:
         expires_in: int = 3600,
         *,
         bucket: str | None = None,
+        extra_params: dict | None = None,
     ) -> str:
         """Generate presigned URL. bucket: override instance default if provided."""
-        return await self._get_file_url(filename, operation, expires_in, bucket=bucket)
+        return await self._get_file_url(
+            filename, operation, expires_in, bucket=bucket, extra_params=extra_params
+        )
 
     async def _get_file_url(
         self,
@@ -127,6 +130,7 @@ class Storage:
         expires_in: int = 3600,
         *,
         bucket: str | None = None,
+        extra_params: dict | None = None,
     ) -> str:
         raise NotImplementedError
 

server/reflector/storage/storage_aws.py

@@ -170,16 +170,23 @@ class AwsStorage(Storage):
         expires_in: int = 3600,
         *,
         bucket: str | None = None,
+        extra_params: dict | None = None,
     ) -> str:
         actual_bucket = bucket or self._bucket_name
         folder = self.aws_folder
         s3filename = f"{folder}/{filename}" if folder else filename
+        params = {}
+        if extra_params:
+            params.update(extra_params)
+        # Always set Bucket/Key after extra_params to prevent overrides
+        params["Bucket"] = actual_bucket
+        params["Key"] = s3filename
         async with self.session.client(
             "s3", config=self.boto_config, endpoint_url=self._endpoint_url
         ) as client:
             presigned_url = await client.generate_presigned_url(
                 operation,
-                Params={"Bucket": actual_bucket, "Key": s3filename},
+                Params=params,
                 ExpiresIn=expires_in,
             )
 

server/reflector/tools/deleted_transcripts.py

@@ -0,0 +1,257 @@
+#!/usr/bin/env python
+"""
+CLI tool for managing soft-deleted transcripts.
+
+Usage:
+    uv run python -m reflector.tools.deleted_transcripts list
+    uv run python -m reflector.tools.deleted_transcripts files <transcript_id>
+    uv run python -m reflector.tools.deleted_transcripts download <transcript_id> [--output-dir ./]
+"""
+
+import argparse
+import asyncio
+import json
+import os
+
+import structlog
+
+from reflector.db import get_database
+from reflector.db.meetings import meetings_controller
+from reflector.db.recordings import recordings_controller
+from reflector.db.transcripts import Transcript, transcripts
+from reflector.storage import get_source_storage, get_transcripts_storage
+
+logger = structlog.get_logger(__name__)
+
+
+async def list_deleted():
+    """List all soft-deleted transcripts."""
+    database = get_database()
+    await database.connect()
+    try:
+        query = (
+            transcripts.select()
+            .where(transcripts.c.deleted_at.isnot(None))
+            .order_by(transcripts.c.deleted_at.desc())
+        )
+        results = await database.fetch_all(query)
+
+        if not results:
+            print("No deleted transcripts found.")
+            return
+
+        print(
+            f"{'ID':<40} {'Title':<40} {'Deleted At':<28} {'Recording ID':<40} {'Meeting ID'}"
+        )
+        print("-" * 180)
+        for row in results:
+            t = Transcript(**row)
+            title = (t.title or "")[:38]
+            deleted = t.deleted_at.isoformat() if t.deleted_at else ""
+            print(
+                f"{t.id:<40} {title:<40} {deleted:<28} {t.recording_id or '':<40} {t.meeting_id or ''}"
+            )
+
+        print(f"\nTotal: {len(results)} deleted transcript(s)")
+    finally:
+        await database.disconnect()
+
+
+async def list_files(transcript_id: str):
+    """List all S3 keys associated with a deleted transcript."""
+    database = get_database()
+    await database.connect()
+    try:
+        query = transcripts.select().where(transcripts.c.id == transcript_id)
+        result = await database.fetch_one(query)
+        if not result:
+            print(f"Transcript {transcript_id} not found.")
+            return
+
+        t = Transcript(**result)
+        if t.deleted_at is None:
+            print(f"Transcript {transcript_id} is not deleted.")
+            return
+
+        print(f"Transcript: {t.id}")
+        print(f"Title: {t.title}")
+        print(f"Deleted at: {t.deleted_at}")
+        print()
+
+        files = []
+
+        # Transcript audio
+        if t.audio_location == "storage" and not t.audio_deleted:
+            files.append(("Transcript audio", t.storage_audio_path, None))
+
+        # Recording files
+        if t.recording_id:
+            recording = await recordings_controller.get_by_id(t.recording_id)
+            if recording:
+                if recording.object_key:
+                    files.append(
+                        (
+                            "Recording object_key",
+                            recording.object_key,
+                            recording.bucket_name,
+                        )
+                    )
+                if recording.track_keys:
+                    for i, key in enumerate(recording.track_keys):
+                        files.append((f"Track {i}", key, recording.bucket_name))
+
+        # Cloud video
+        if t.meeting_id:
+            meeting = await meetings_controller.get_by_id(t.meeting_id)
+            if meeting and meeting.daily_composed_video_s3_key:
+                files.append(("Cloud video", meeting.daily_composed_video_s3_key, None))
+
+        if not files:
+            print("No associated files found.")
+            return
+
+        print(f"{'Type':<25} {'Bucket':<30} {'S3 Key'}")
+        print("-" * 120)
+        for label, key, bucket in files:
+            print(f"{label:<25} {bucket or '(default)':<30} {key}")
+
+        # Generate presigned URLs
+        print("\nPresigned URLs (valid for 1 hour):")
+        print("-" * 120)
+        storage = get_transcripts_storage()
+        for label, key, bucket in files:
+            try:
+                url = await storage.get_file_url(key, bucket=bucket, expires_in=3600)
+                print(f"{label}: {url}")
+            except Exception as e:
+                print(f"{label}: ERROR - {e}")
+    finally:
+        await database.disconnect()
+
+
+async def download_files(transcript_id: str, output_dir: str):
+    """Download all files associated with a deleted transcript."""
+    database = get_database()
+    await database.connect()
+    try:
+        query = transcripts.select().where(transcripts.c.id == transcript_id)
+        result = await database.fetch_one(query)
+        if not result:
+            print(f"Transcript {transcript_id} not found.")
+            return
+
+        t = Transcript(**result)
+        if t.deleted_at is None:
+            print(f"Transcript {transcript_id} is not deleted.")
+            return
+
+        dest = os.path.join(output_dir, t.id)
+        os.makedirs(dest, exist_ok=True)
+
+        storage = get_transcripts_storage()
+
+        # Download transcript audio
+        if t.audio_location == "storage" and not t.audio_deleted:
+            try:
+                data = await storage.get_file(t.storage_audio_path)
+                path = os.path.join(dest, "audio.mp3")
+                with open(path, "wb") as f:
+                    f.write(data)
+                print(f"Downloaded: {path}")
+            except Exception as e:
+                print(f"Failed to download audio: {e}")
+
+        # Download recording files
+        if t.recording_id:
+            recording = await recordings_controller.get_by_id(t.recording_id)
+            if recording and recording.track_keys:
+                tracks_dir = os.path.join(dest, "tracks")
+                os.makedirs(tracks_dir, exist_ok=True)
+                for i, key in enumerate(recording.track_keys):
+                    try:
+                        data = await storage.get_file(key, bucket=recording.bucket_name)
+                        filename = os.path.basename(key) or f"track_{i}"
+                        path = os.path.join(tracks_dir, filename)
+                        with open(path, "wb") as f:
+                            f.write(data)
+                        print(f"Downloaded: {path}")
+                    except Exception as e:
+                        print(f"Failed to download track {i}: {e}")
+
+        # Download cloud video
+        if t.meeting_id:
+            meeting = await meetings_controller.get_by_id(t.meeting_id)
+            if meeting and meeting.daily_composed_video_s3_key:
+                try:
+                    source_storage = get_source_storage("daily")
+                    data = await source_storage.get_file(
+                        meeting.daily_composed_video_s3_key
+                    )
+                    path = os.path.join(dest, "cloud_video.mp4")
+                    with open(path, "wb") as f:
+                        f.write(data)
+                    print(f"Downloaded: {path}")
+                except Exception as e:
+                    print(f"Failed to download cloud video: {e}")
+
+        # Write metadata
+        metadata = {
+            "id": t.id,
+            "title": t.title,
+            "created_at": t.created_at.isoformat() if t.created_at else None,
+            "deleted_at": t.deleted_at.isoformat() if t.deleted_at else None,
+            "duration": t.duration,
+            "source_language": t.source_language,
+            "target_language": t.target_language,
+            "short_summary": t.short_summary,
+            "long_summary": t.long_summary,
+            "topics": [topic.model_dump() for topic in t.topics] if t.topics else [],
+            "participants": [p.model_dump() for p in t.participants]
+            if t.participants
+            else [],
+            "action_items": t.action_items,
+            "webvtt": t.webvtt,
+            "recording_id": t.recording_id,
+            "meeting_id": t.meeting_id,
+        }
+        path = os.path.join(dest, "metadata.json")
+        with open(path, "w") as f:
+            json.dump(metadata, f, indent=2, default=str)
+        print(f"Downloaded: {path}")
+
+        print(f"\nAll files saved to: {dest}")
+    finally:
+        await database.disconnect()
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Manage soft-deleted transcripts")
+    subparsers = parser.add_subparsers(dest="command", required=True)
+
+    subparsers.add_parser("list", help="List all deleted transcripts")
+
+    files_parser = subparsers.add_parser(
+        "files", help="List S3 keys for a deleted transcript"
+    )
+    files_parser.add_argument("transcript_id", help="Transcript ID")
+
+    download_parser = subparsers.add_parser(
+        "download", help="Download files for a deleted transcript"
+    )
+    download_parser.add_argument("transcript_id", help="Transcript ID")
+    download_parser.add_argument(
+        "--output-dir", default=".", help="Output directory (default: .)"
+    )
+
+    args = parser.parse_args()
+
+    if args.command == "list":
+        asyncio.run(list_deleted())
+    elif args.command == "files":
+        asyncio.run(list_files(args.transcript_id))
+    elif args.command == "download":
+        asyncio.run(download_files(args.transcript_id, args.output_dir))
+
+
+if __name__ == "__main__":
+    main()

server/reflector/tools/process.py

@@ -333,7 +333,9 @@ if __name__ == "__main__":
         if not s3_urls:
             parser.error("At least one S3 URL required for multitrack processing")
 
-        from reflector.tools.cli_multitrack import process_multitrack_cli
+        from reflector.tools.cli_multitrack import (
+            process_multitrack_cli,  # circular import
+        )
 
         asyncio.run(
             process_multitrack_cli(

server/reflector/tools/render_hatchet_run.py

@@ -0,0 +1,412 @@
+"""
+Render Hatchet workflow runs as text DAG.
+
+Usage:
+    # Show latest 5 runs (summary table)
+    uv run -m reflector.tools.render_hatchet_run
+
+    # Show specific run with full DAG + task details
+    uv run -m reflector.tools.render_hatchet_run <workflow_run_id>
+
+    # Drill into Nth run from the list (1-indexed)
+    uv run -m reflector.tools.render_hatchet_run --show 1
+
+    # Show latest N runs
+    uv run -m reflector.tools.render_hatchet_run --last 10
+
+    # Filter by status
+    uv run -m reflector.tools.render_hatchet_run --status FAILED
+    uv run -m reflector.tools.render_hatchet_run --status RUNNING
+"""
+
+import argparse
+import asyncio
+from collections import defaultdict
+from datetime import datetime, timedelta, timezone
+
+from hatchet_sdk.clients.rest.models import (
+    V1TaskEvent,
+    V1TaskStatus,
+    V1TaskSummary,
+    V1WorkflowRunDetails,
+    WorkflowRunShapeItemForWorkflowRunDetails,
+)
+
+from reflector.hatchet.client import HatchetClientManager
+
+STATUS_ICON = {
+    V1TaskStatus.COMPLETED: "\u2705",
+    V1TaskStatus.RUNNING: "\u23f3",
+    V1TaskStatus.FAILED: "\u274c",
+    V1TaskStatus.QUEUED: "\u23f8\ufe0f",
+    V1TaskStatus.CANCELLED: "\u26a0\ufe0f",
+}
+
+STATUS_LABEL = {
+    V1TaskStatus.COMPLETED: "Complete",
+    V1TaskStatus.RUNNING: "Running",
+    V1TaskStatus.FAILED: "FAILED",
+    V1TaskStatus.QUEUED: "Queued",
+    V1TaskStatus.CANCELLED: "Cancelled",
+}
+
+
+def _fmt_time(dt: datetime | None) -> str:
+    if dt is None:
+        return "-"
+    return dt.strftime("%H:%M:%S")
+
+
+def _fmt_duration(ms: int | None) -> str:
+    if ms is None:
+        return "-"
+    secs = ms / 1000
+    if secs < 60:
+        return f"{secs:.1f}s"
+    mins = secs / 60
+    return f"{mins:.1f}m"
+
+
+def _fmt_status_line(task: V1TaskSummary) -> str:
+    """Format a status line like: Complete (finished 20:31:44)"""
+    label = STATUS_LABEL.get(task.status, task.status.value)
+    icon = STATUS_ICON.get(task.status, "?")
+
+    if task.status == V1TaskStatus.COMPLETED and task.finished_at:
+        return f"{icon} {label} (finished {_fmt_time(task.finished_at)})"
+    elif task.status == V1TaskStatus.RUNNING and task.started_at:
+        parts = [f"started {_fmt_time(task.started_at)}"]
+        if task.duration:
+            parts.append(f"{_fmt_duration(task.duration)} elapsed")
+        return f"{icon} {label} ({', '.join(parts)})"
+    elif task.status == V1TaskStatus.FAILED and task.finished_at:
+        return f"{icon} {label} (failed {_fmt_time(task.finished_at)})"
+    elif task.status == V1TaskStatus.CANCELLED:
+        return f"{icon} {label}"
+    elif task.status == V1TaskStatus.QUEUED:
+        return f"{icon} {label}"
+    return f"{icon} {label}"
+
+
+def _topo_sort(
+    shape: list[WorkflowRunShapeItemForWorkflowRunDetails],
+) -> list[str]:
+    """Topological sort of step_ids from shape DAG."""
+    step_ids = {s.step_id for s in shape}
+    children_map: dict[str, list[str]] = {}
+    in_degree: dict[str, int] = {sid: 0 for sid in step_ids}
+
+    for s in shape:
+        children = [c for c in (s.children_step_ids or []) if c in step_ids]
+        children_map[s.step_id] = children
+        for c in children:
+            in_degree[c] += 1
+
+    queue = sorted(sid for sid, deg in in_degree.items() if deg == 0)
+    result: list[str] = []
+    while queue:
+        node = queue.pop(0)
+        result.append(node)
+        for c in children_map.get(node, []):
+            in_degree[c] -= 1
+            if in_degree[c] == 0:
+                queue.append(c)
+                queue.sort()
+
+    return result
+
+
+def render_run_detail(details: V1WorkflowRunDetails) -> str:
+    """Render a single workflow run as markdown DAG with task details."""
+    shape = details.shape or []
+    tasks = details.tasks or []
+    events = details.task_events or []
+    run = details.run
+
+    if not shape:
+        return f"Run {run.metadata.id}: {run.status.value} (no shape data)"
+
+    # Build lookups
+    step_to_shape: dict[str, WorkflowRunShapeItemForWorkflowRunDetails] = {
+        s.step_id: s for s in shape
+    }
+    step_to_name: dict[str, str] = {s.step_id: s.task_name for s in shape}
+
+    # Reverse edges (parents)
+    parents: dict[str, list[str]] = {s.step_id: [] for s in shape}
+    for s in shape:
+        for child_id in s.children_step_ids or []:
+            if child_id in parents:
+                parents[child_id].append(s.step_id)
+
+    # Join tasks by step_id
+    task_by_step: dict[str, V1TaskSummary] = {}
+    for t in tasks:
+        if t.step_id and t.step_id in step_to_name:
+            task_by_step[t.step_id] = t
+
+    # Events indexed by task_external_id
+    events_by_task: dict[str, list[V1TaskEvent]] = defaultdict(list)
+    for ev in events:
+        events_by_task[ev.task_id].append(ev)
+
+    ordered = _topo_sort(shape)
+
+    lines: list[str] = []
+
+    # Run header
+    run_icon = STATUS_ICON.get(run.status, "?")
+    run_name = run.display_name or run.workflow_id
+    dur = _fmt_duration(run.duration)
+    lines.append(f"**{run_name}** {run_icon} {dur}")
+    lines.append(f"ID: `{run.metadata.id}`")
+    if run.additional_metadata:
+        meta_parts = [f"{k}=`{v}`" for k, v in run.additional_metadata.items()]
+        lines.append(f"Meta: {', '.join(meta_parts)}")
+    if run.error_message:
+        # Take first line of error only for header
+        first_line = run.error_message.split("\n")[0]
+        lines.append(f"Error: {first_line}")
+    lines.append("")
+
+    # DAG Status Overview table (collapsible)
+    lines.append("```spoiler DAG Status Overview")
+    lines.append("| Node | Status | Duration | Dependencies |")
+    lines.append("|------|--------|----------|--------------|")
+
+    for step_id in ordered:
+        s = step_to_shape[step_id]
+        t = task_by_step.get(step_id)
+        name = step_to_name[step_id]
+        icon = STATUS_ICON.get(t.status, "?") if t else "?"
+        dur = _fmt_duration(t.duration) if t else "-"
+
+        parent_names = [step_to_name[p] for p in parents[step_id]]
+        child_names = [
+            step_to_name[c] for c in (s.children_step_ids or []) if c in step_to_name
+        ]
+        deps_left = ", ".join(parent_names) if parent_names else ""
+        deps_right = ", ".join(child_names) if child_names else ""
+        if deps_left and deps_right:
+            deps = f"{deps_left} \u2192 {deps_right}"
+        elif deps_right:
+            deps = f"\u2192 {deps_right}"
+        elif deps_left:
+            deps = f"{deps_left} \u2192"
+        else:
+            deps = "-"
+
+        lines.append(f"| {name} | {icon} | {dur} | {deps} |")
+
+    lines.append("```")
+    lines.append("")
+
+    # Node details (collapsible)
+    lines.append("```spoiler Node Details")
+    for step_id in ordered:
+        t = task_by_step.get(step_id)
+        name = step_to_name[step_id]
+
+        if not t:
+            lines.append(f"**\U0001f4e6 {name}**")
+            lines.append("Status: no task data")
+            lines.append("")
+            continue
+
+        lines.append(f"**\U0001f4e6 {name}**")
+        lines.append(f"Status: {_fmt_status_line(t)}")
+
+        if t.duration:
+            lines.append(f"Duration: {_fmt_duration(t.duration)}")
+        if t.retry_count and t.retry_count > 0:
+            lines.append(f"Retries: {t.retry_count}")
+
+        # Fan-out children
+        if t.num_spawned_children and t.num_spawned_children > 0:
+            children = t.children or []
+            completed = sum(1 for c in children if c.status == V1TaskStatus.COMPLETED)
+            failed = sum(1 for c in children if c.status == V1TaskStatus.FAILED)
+            running = sum(1 for c in children if c.status == V1TaskStatus.RUNNING)
+            lines.append(
+                f"Spawned children: {completed}/{t.num_spawned_children} done"
+                f"{f', {running} running' if running else ''}"
+                f"{f', {failed} failed' if failed else ''}"
+            )
+
+        # Error message (first meaningful line only, full trace in events)
+        if t.error_message:
+            err_lines = t.error_message.strip().split("\n")
+            # Find first non-empty, non-traceback line
+            err_summary = err_lines[0]
+            for line in err_lines:
+                stripped = line.strip()
+                if stripped and not stripped.startswith(
+                    ("Traceback", "File ", "{", ")")
+                ):
+                    err_summary = stripped
+                    break
+            lines.append(f"Error: `{err_summary}`")
+
+        # Events log
+        task_events = sorted(
+            events_by_task.get(t.task_external_id, []),
+            key=lambda e: e.timestamp,
+        )
+        if task_events:
+            lines.append("Events:")
+            for ev in task_events:
+                ts = ev.timestamp.strftime("%H:%M:%S")
+                ev_icon = ""
+                if ev.event_type.value == "FINISHED":
+                    ev_icon = "\u2705 "
+                elif ev.event_type.value in ("FAILED", "TIMED_OUT"):
+                    ev_icon = "\u274c "
+                elif ev.event_type.value == "STARTED":
+                    ev_icon = "\u25b6\ufe0f "
+                elif ev.event_type.value == "RETRYING":
+                    ev_icon = "\U0001f504 "
+                elif ev.event_type.value == "CANCELLED":
+                    ev_icon = "\u26a0\ufe0f "
+
+                msg = ev.message.strip()
+                if ev.error_message:
+                    # Just first line of error in event log
+                    err_first = ev.error_message.strip().split("\n")[0]
+                    if msg:
+                        msg += f" | {err_first}"
+                    else:
+                        msg = err_first
+
+                if msg:
+                    lines.append(f"  `{ts}` {ev_icon}{ev.event_type.value}: {msg}")
+                else:
+                    lines.append(f"  `{ts}` {ev_icon}{ev.event_type.value}")
+
+        lines.append("")
+
+    lines.append("```")
+    return "\n".join(lines)
+
+
+def render_run_summary(idx: int, run: V1TaskSummary) -> str:
+    """One-line summary for a run in the list view."""
+    icon = STATUS_ICON.get(run.status, "?")
+    name = run.display_name or run.workflow_name or "?"
+    run_id = run.workflow_run_external_id or "?"
+    dur = _fmt_duration(run.duration)
+    started = _fmt_time(run.started_at)
+    meta = ""
+    if run.additional_metadata:
+        meta_parts = [f"{k}=`{v}`" for k, v in run.additional_metadata.items()]
+        meta = f"  ({', '.join(meta_parts)})"
+    return (
+        f"  {idx}. {icon} **{name}** started={started} dur={dur}{meta}\n"
+        f"     `{run_id}`"
+    )
+
+
+async def _fetch_run_list(
+    count: int = 5,
+    statuses: list[V1TaskStatus] | None = None,
+) -> list[V1TaskSummary]:
+    client = HatchetClientManager.get_client()
+    since = datetime.now(timezone.utc) - timedelta(days=7)
+    runs = await client.runs.aio_list(
+        since=since,
+        statuses=statuses,
+        limit=count,
+    )
+    return runs.rows or []
+
+
+async def list_recent_runs(
+    count: int = 5,
+    statuses: list[V1TaskStatus] | None = None,
+) -> str:
+    """List recent workflow runs as text."""
+    rows = await _fetch_run_list(count, statuses)
+
+    if not rows:
+        return "No runs found in the last 7 days."
+
+    lines = [f"Recent runs ({len(rows)}):", ""]
+    for i, run in enumerate(rows, 1):
+        lines.append(render_run_summary(i, run))
+
+    lines.append("")
+    lines.append("Use `--show N` to see full DAG for run N")
+    return "\n".join(lines)
+
+
+async def show_run(workflow_run_id: str) -> str:
+    """Fetch and render a single run."""
+    client = HatchetClientManager.get_client()
+    details = await client.runs.aio_get(workflow_run_id)
+    return render_run_detail(details)
+
+
+async def show_nth_run(
+    n: int,
+    count: int = 5,
+    statuses: list[V1TaskStatus] | None = None,
+) -> str:
+    """Fetch list, then drill into Nth run."""
+    rows = await _fetch_run_list(count, statuses)
+
+    if not rows:
+        return "No runs found in the last 7 days."
+    if n < 1 or n > len(rows):
+        return f"Invalid index {n}. Have {len(rows)} runs (1-{len(rows)})."
+
+    run = rows[n - 1]
+    return await show_run(run.workflow_run_external_id)
+
+
+async def main_async(args: argparse.Namespace) -> None:
+    statuses = [V1TaskStatus(args.status)] if args.status else None
+
+    if args.run_id:
+        output = await show_run(args.run_id)
+    elif args.show is not None:
+        output = await show_nth_run(args.show, count=args.last, statuses=statuses)
+    else:
+        output = await list_recent_runs(count=args.last, statuses=statuses)
+
+    print(output)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Render Hatchet workflow runs as text DAG"
+    )
+    parser.add_argument(
+        "run_id",
+        nargs="?",
+        default=None,
+        help="Workflow run ID to show in detail. If omitted, lists recent runs.",
+    )
+    parser.add_argument(
+        "--show",
+        type=int,
+        default=None,
+        metavar="N",
+        help="Show full DAG for the Nth run in the list (1-indexed)",
+    )
+    parser.add_argument(
+        "--last",
+        type=int,
+        default=5,
+        help="Number of recent runs to list (default: 5)",
+    )
+    parser.add_argument(
+        "--status",
+        choices=["QUEUED", "RUNNING", "COMPLETED", "FAILED", "CANCELLED"],
+        help="Filter by status",
+    )
+
+    args = parser.parse_args()
+    asyncio.run(main_async(args))
+
+
+if __name__ == "__main__":
+    main()

server/reflector/tools/runpipeline.py

@@ -5,6 +5,7 @@ This tools help to either create a pipeline from command line,
 or read a yaml description of a pipeline and run it.
 """
 
+import importlib
 import json
 
 from reflector.logger import logger
@@ -37,8 +38,6 @@ def get_jsonl(filename, filter_processor_name=None):
 
 
 def get_processor(name):
-    import importlib
-
     module_name = f"reflector.processors.{name}"
     class_name = snake_to_camel(name) + "Processor"
     module = importlib.import_module(module_name)

server/reflector/utils/livekit.py

@@ -0,0 +1,112 @@
+"""
+LiveKit track file utilities.
+
+Parse participant identity and timing from Auto Track Egress S3 filepaths.
+
+Actual filepath format from LiveKit Auto Track Egress:
+  livekit/{room_name}/{publisher_identity}-{ISO_timestamp}-{track_id}.{ext}
+
+Examples:
+  livekit/myroom-20260401172036/juan-4b82ed-2026-04-01T195758-TR_AMR3SWs74Divho.ogg
+  livekit/myroom-20260401172036/juan2-63abcf-2026-04-01T195847-TR_AMyoSbM7tAQbYj.ogg
+  livekit/myroom-20260401172036/EG_K5sipvfB5fTM.json  (manifest, skip)
+  livekit/myroom-20260401172036/juan-4b82ed-2026-04-01T195727-TR_VC679dgMQBdfhT.webm  (video, skip)
+"""
+
+import re
+from dataclasses import dataclass
+from datetime import datetime, timezone
+
+from reflector.utils.string import NonEmptyString
+
+
+@dataclass
+class LiveKitTrackFile:
+    """Parsed info from a LiveKit track egress filepath."""
+
+    s3_key: str
+    room_name: str
+    participant_identity: str
+    timestamp: datetime  # Parsed from ISO timestamp in filename
+    track_id: str  # LiveKit track ID (e.g., TR_AMR3SWs74Divho)
+
+
+# Pattern: livekit/{room_name}/{identity}-{ISO_date}T{time}-{track_id}.{ext}
+# The identity can contain alphanumeric, hyphens, underscores
+# ISO timestamp is like 2026-04-01T195758
+# Track ID starts with TR_
+_TRACK_FILENAME_PATTERN = re.compile(
+    r"^livekit/(?P<room_name>[^/]+)/(?P<identity>.+?)-(?P<timestamp>\d{4}-\d{2}-\d{2}T\d{6})-(?P<track_id>TR_\w+)\.(?P<ext>\w+)$"
+)
+
+
+def parse_livekit_track_filepath(s3_key: str) -> LiveKitTrackFile:
+    """Parse a LiveKit track egress filepath into components.
+
+    Args:
+        s3_key: S3 key like 'livekit/myroom-20260401/juan-4b82ed-2026-04-01T195758-TR_AMR3SWs74Divho.ogg'
+
+    Returns:
+        LiveKitTrackFile with parsed components.
+
+    Raises:
+        ValueError: If the filepath doesn't match the expected format.
+    """
+    match = _TRACK_FILENAME_PATTERN.match(s3_key)
+    if not match:
+        raise ValueError(
+            f"LiveKit track filepath doesn't match expected format: {s3_key}"
+        )
+
+    # Parse ISO-ish timestamp (e.g., 2026-04-01T195758 → datetime)
+    ts_str = match.group("timestamp")
+    try:
+        ts = datetime.strptime(ts_str, "%Y-%m-%dT%H%M%S").replace(tzinfo=timezone.utc)
+    except ValueError:
+        raise ValueError(f"Cannot parse timestamp '{ts_str}' from: {s3_key}")
+
+    return LiveKitTrackFile(
+        s3_key=s3_key,
+        room_name=match.group("room_name"),
+        participant_identity=match.group("identity"),
+        timestamp=ts,
+        track_id=match.group("track_id"),
+    )
+
+
+def filter_audio_tracks(s3_keys: list[str]) -> list[str]:
+    """Filter S3 keys to only audio tracks (.ogg), excluding manifests and video."""
+    return [k for k in s3_keys if k.endswith(".ogg")]
+
+
+def calculate_track_offsets(
+    tracks: list[LiveKitTrackFile],
+) -> list[tuple[LiveKitTrackFile, float]]:
+    """Calculate silence padding offset for each track.
+
+    The earliest track starts at time zero. Each subsequent track
+    gets (track_timestamp - earliest_timestamp) seconds of silence prepended.
+
+    Returns:
+        List of (track, offset_seconds) tuples.
+    """
+    if not tracks:
+        return []
+
+    earliest = min(t.timestamp for t in tracks)
+    return [(t, (t.timestamp - earliest).total_seconds()) for t in tracks]
+
+
+def extract_livekit_base_room_name(livekit_room_name: str) -> NonEmptyString:
+    """Extract base room name from LiveKit timestamped room name.
+
+    LiveKit rooms use the same naming as Daily: {base_name}-YYYYMMDDHHMMSS
+    """
+    base_name = livekit_room_name.rsplit("-", 1)[0]
+    assert base_name, f"Extracted base name is empty from: {livekit_room_name}"
+    return NonEmptyString(base_name)
+
+
+def recording_lock_key(room_name: str) -> str:
+    """Redis lock key for preventing duplicate processing."""
+    return f"livekit:processing:{room_name}"

server/reflector/utils/transcript_constants.py

@@ -4,5 +4,67 @@ Shared transcript processing constants.
 Used by both Hatchet workflows and Celery pipelines for consistent processing.
 """
 
-# Topic detection: number of words per chunk for topic extraction
+import math
+
+# Topic detection: legacy static chunk size, used as fallback
 TOPIC_CHUNK_WORD_COUNT = 300
+
+# Dynamic chunking curve parameters
+# Formula: target_topics = _COEFFICIENT * duration_minutes ^ _EXPONENT
+# Derived from anchors: 5 min -> 3 topics, 180 min -> 40 topics
+_TOPIC_CURVE_COEFFICIENT = 0.833
+_TOPIC_CURVE_EXPONENT = 0.723
+_MIN_TOPICS = 2
+_MAX_TOPICS = 50
+_MIN_CHUNK_WORDS = 375
+_MAX_CHUNK_WORDS = 1500
+
+
+def compute_topic_chunk_size(duration_seconds: float, total_words: int) -> int:
+    """Calculate optimal chunk size for topic detection based on recording duration.
+
+    Uses a power-curve function to scale topic count sublinearly with duration,
+    producing fewer LLM calls for longer recordings while maintaining topic quality.
+
+    Returns the number of words per chunk.
+    """
+    if total_words <= 0 or duration_seconds <= 0:
+        return _MIN_CHUNK_WORDS
+
+    duration_minutes = duration_seconds / 60.0
+    target_topics = _TOPIC_CURVE_COEFFICIENT * math.pow(
+        duration_minutes, _TOPIC_CURVE_EXPONENT
+    )
+    target_topics = int(round(max(_MIN_TOPICS, min(_MAX_TOPICS, target_topics))))
+
+    chunk_size = total_words // target_topics
+    chunk_size = max(_MIN_CHUNK_WORDS, min(_MAX_CHUNK_WORDS, chunk_size))
+    return chunk_size
+
+
+# Subject extraction: scale max subjects with recording duration
+# Short calls get fewer subjects to avoid over-analyzing trivial content
+_SUBJECT_DURATION_THRESHOLDS = [
+    (5 * 60, 1),  # ≤ 5 min  → 1 subject
+    (15 * 60, 2),  # ≤ 15 min → 2 subjects
+    (30 * 60, 3),  # ≤ 30 min → 3 subjects
+    (45 * 60, 4),  # ≤ 45 min → 4 subjects
+    (60 * 60, 5),  # ≤ 60 min → 5 subjects
+]
+_MAX_SUBJECTS = 6
+
+
+def compute_max_subjects(duration_seconds: float) -> int:
+    """Calculate maximum number of subjects to extract based on recording duration.
+
+    Uses a step function: short recordings get fewer subjects to avoid
+    generating excessive detail for trivial content.
+    """
+    if duration_seconds <= 0:
+        return 1
+
+    for threshold, max_subjects in _SUBJECT_DURATION_THRESHOLDS:
+        if duration_seconds <= threshold:
+            return max_subjects
+
+    return _MAX_SUBJECTS

server/reflector/video_platforms/factory.py

@@ -1,7 +1,7 @@
 from reflector.settings import settings
 from reflector.storage import get_dailyco_storage, get_whereby_storage
 
-from ..schemas.platform import WHEREBY_PLATFORM, Platform
+from ..schemas.platform import LIVEKIT_PLATFORM, WHEREBY_PLATFORM, Platform
 from .base import VideoPlatformClient, VideoPlatformConfig
 from .registry import get_platform_client
 
@@ -44,6 +44,27 @@ def get_platform_config(platform: Platform) -> VideoPlatformConfig:
             s3_region=daily_storage.region,
             aws_role_arn=daily_storage.role_credential,
         )
+    elif platform == LIVEKIT_PLATFORM:
+        if not settings.LIVEKIT_URL:
+            raise ValueError(
+                "LIVEKIT_URL is required when platform='livekit'. "
+                "Set LIVEKIT_URL environment variable."
+            )
+        if not settings.LIVEKIT_API_KEY or not settings.LIVEKIT_API_SECRET:
+            raise ValueError(
+                "LIVEKIT_API_KEY and LIVEKIT_API_SECRET are required when platform='livekit'. "
+                "Set LIVEKIT_API_KEY and LIVEKIT_API_SECRET environment variables."
+            )
+        return VideoPlatformConfig(
+            api_key=settings.LIVEKIT_API_KEY,
+            webhook_secret=settings.LIVEKIT_WEBHOOK_SECRET
+            or settings.LIVEKIT_API_SECRET,
+            api_url=settings.LIVEKIT_URL,
+            s3_bucket=settings.LIVEKIT_STORAGE_AWS_BUCKET_NAME,
+            s3_region=settings.LIVEKIT_STORAGE_AWS_REGION,
+            aws_access_key_id=settings.LIVEKIT_STORAGE_AWS_ACCESS_KEY_ID,
+            aws_access_key_secret=settings.LIVEKIT_STORAGE_AWS_SECRET_ACCESS_KEY,
+        )
     else:
         raise ValueError(f"Unknown platform: {platform}")
 

server/reflector/video_platforms/livekit.py

@@ -0,0 +1,192 @@
+"""
+LiveKit video platform client for Reflector.
+
+Self-hosted, open-source alternative to Daily.co.
+Uses Track Egress for per-participant audio recording (no composite video).
+"""
+
+from datetime import datetime, timezone
+from urllib.parse import urlencode
+from uuid import uuid4
+
+from reflector.db.rooms import Room
+from reflector.livekit_api.client import LiveKitApiClient
+from reflector.livekit_api.webhooks import create_webhook_receiver, verify_webhook
+from reflector.logger import logger
+from reflector.settings import settings
+
+from ..schemas.platform import Platform
+from ..utils.string import NonEmptyString
+from .base import ROOM_PREFIX_SEPARATOR, VideoPlatformClient
+from .models import MeetingData, SessionData, VideoPlatformConfig
+
+
+class LiveKitClient(VideoPlatformClient):
+    PLATFORM_NAME: Platform = "livekit"
+    TIMESTAMP_FORMAT = "%Y%m%d%H%M%S"
+
+    def __init__(self, config: VideoPlatformConfig):
+        super().__init__(config)
+        self._api_client = LiveKitApiClient(
+            url=config.api_url or "",
+            api_key=config.api_key,
+            api_secret=config.webhook_secret,  # LiveKit uses API secret for both auth and webhooks
+            s3_bucket=config.s3_bucket,
+            s3_region=config.s3_region,
+            s3_access_key=config.aws_access_key_id,
+            s3_secret_key=config.aws_access_key_secret,
+            s3_endpoint=settings.LIVEKIT_STORAGE_AWS_ENDPOINT_URL,
+        )
+        self._webhook_receiver = create_webhook_receiver(
+            api_key=config.api_key,
+            api_secret=config.webhook_secret,
+        )
+
+    async def create_meeting(
+        self, room_name_prefix: NonEmptyString, end_date: datetime, room: Room
+    ) -> MeetingData:
+        """Create a LiveKit room for this meeting.
+
+        LiveKit rooms are created explicitly via API. A new room is created
+        for each Reflector meeting (same pattern as Daily.co).
+        """
+        now = datetime.now(timezone.utc)
+        timestamp = now.strftime(self.TIMESTAMP_FORMAT)
+        room_name = f"{room_name_prefix}{ROOM_PREFIX_SEPARATOR}{timestamp}"
+
+        # Calculate empty_timeout from end_date (seconds until expiry)
+        # Ensure end_date is timezone-aware for subtraction
+        end_date_aware = (
+            end_date if end_date.tzinfo else end_date.replace(tzinfo=timezone.utc)
+        )
+        remaining = int((end_date_aware - now).total_seconds())
+        empty_timeout = max(300, min(remaining, 86400))  # 5 min to 24 hours
+
+        # Enable auto track egress for cloud recording (per-participant audio to S3).
+        # Gracefully degrade if S3 credentials are missing — room still works, just no recording.
+        enable_recording = room.recording_type == "cloud"
+        egress_enabled = False
+        if enable_recording:
+            try:
+                self._api_client._build_s3_upload()  # Validate credentials exist
+                egress_enabled = True
+            except ValueError:
+                logger.warning(
+                    "S3 credentials not configured — room created without auto track egress. "
+                    "Set LIVEKIT_STORAGE_AWS_* to enable recording.",
+                    room_name=room_name,
+                )
+
+        lk_room = await self._api_client.create_room(
+            name=room_name,
+            empty_timeout=empty_timeout,
+            enable_auto_track_egress=egress_enabled,
+        )
+
+        logger.info(
+            "LiveKit room created",
+            room_name=lk_room.name,
+            room_sid=lk_room.sid,
+            empty_timeout=empty_timeout,
+            auto_track_egress=egress_enabled,
+        )
+
+        # room_url includes the server URL + room name as query param.
+        # The join endpoint in rooms.py appends the token as another query param.
+        # Frontend parses: ws://host:7880?room=<name>&token=<jwt>
+        public_url = settings.LIVEKIT_PUBLIC_URL or settings.LIVEKIT_URL or ""
+        room_url = f"{public_url}?{urlencode({'room': lk_room.name})}"
+
+        return MeetingData(
+            meeting_id=lk_room.sid or str(uuid4()),
+            room_name=lk_room.name,
+            room_url=room_url,
+            host_room_url=room_url,
+            platform=self.PLATFORM_NAME,
+            extra_data={"livekit_room_sid": lk_room.sid},
+        )
+
+    async def get_room_sessions(self, room_name: str) -> list[SessionData]:
+        """Get current participants in a LiveKit room.
+
+        For historical sessions, we rely on webhook-stored data (same as Daily).
+        This returns currently-connected participants.
+        """
+        try:
+            participants = await self._api_client.list_participants(room_name)
+            return [
+                SessionData(
+                    session_id=p.sid,
+                    started_at=datetime.fromtimestamp(
+                        p.joined_at if p.joined_at else 0, tz=timezone.utc
+                    ),
+                    ended_at=None,  # Still active
+                )
+                for p in participants
+                if p.sid  # Skip empty entries
+            ]
+        except Exception as e:
+            logger.debug(
+                "Could not list LiveKit participants (room may not exist)",
+                room_name=room_name,
+                error=str(e),
+            )
+            return []
+
+    async def upload_logo(self, room_name: str, logo_path: str) -> bool:
+        # LiveKit doesn't have a logo upload concept; handled in frontend theming
+        return True
+
+    def verify_webhook_signature(
+        self, body: bytes, signature: str, timestamp: str | None = None
+    ) -> bool:
+        """Verify LiveKit webhook signature.
+
+        LiveKit sends the JWT in the Authorization header. The `signature`
+        param here receives the Authorization header value.
+        """
+        event = verify_webhook(self._webhook_receiver, body, signature)
+        return event is not None
+
+    def create_access_token(
+        self,
+        room_name: str,
+        participant_identity: str,
+        participant_name: str | None = None,
+        is_admin: bool = False,
+    ) -> str:
+        """Generate a LiveKit access token for a participant."""
+        return self._api_client.create_access_token(
+            room_name=room_name,
+            participant_identity=participant_identity,
+            participant_name=participant_name,
+            room_admin=is_admin,
+        )
+
+    async def start_track_egress(
+        self,
+        room_name: str,
+        track_sid: str,
+        s3_filepath: str,
+    ):
+        """Start Track Egress for a single audio track."""
+        return await self._api_client.start_track_egress(
+            room_name=room_name,
+            track_sid=track_sid,
+            s3_filepath=s3_filepath,
+        )
+
+    async def list_egress(self, room_name: str | None = None):
+        return await self._api_client.list_egress(room_name=room_name)
+
+    async def stop_egress(self, egress_id: str):
+        return await self._api_client.stop_egress(egress_id=egress_id)
+
+    async def close(self):
+        await self._api_client.close()
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        await self.close()

server/reflector/video_platforms/registry.py

@@ -1,6 +1,11 @@
 from typing import Dict, Type
 
-from ..schemas.platform import DAILY_PLATFORM, WHEREBY_PLATFORM, Platform
+from ..schemas.platform import (
+    DAILY_PLATFORM,
+    LIVEKIT_PLATFORM,
+    WHEREBY_PLATFORM,
+    Platform,
+)
 from .base import VideoPlatformClient, VideoPlatformConfig
 
 _PLATFORMS: Dict[Platform, Type[VideoPlatformClient]] = {}
@@ -26,10 +31,12 @@ def get_available_platforms() -> list[Platform]:
 
 def _register_builtin_platforms():
     from .daily import DailyClient  # noqa: PLC0415
+    from .livekit import LiveKitClient  # noqa: PLC0415
     from .whereby import WherebyClient  # noqa: PLC0415
 
     register_platform(WHEREBY_PLATFORM, WherebyClient)
     register_platform(DAILY_PLATFORM, DailyClient)
+    register_platform(LIVEKIT_PLATFORM, LiveKitClient)
 
 
 _register_builtin_platforms()

server/reflector/views/config.py

@@ -0,0 +1,20 @@
+from fastapi import APIRouter
+from pydantic import BaseModel
+
+from reflector.email import is_email_configured
+from reflector.settings import settings
+
+router = APIRouter()
+
+
+class ConfigResponse(BaseModel):
+    zulip_enabled: bool
+    email_enabled: bool
+
+
+@router.get("/config", response_model=ConfigResponse)
+async def get_config():
+    return ConfigResponse(
+        zulip_enabled=bool(settings.ZULIP_REALM),
+        email_enabled=is_email_configured(),
+    )

server/reflector/views/daily.py

@@ -15,6 +15,7 @@ from reflector.dailyco_api import (
 from reflector.db.meetings import meetings_controller
 from reflector.logger import logger as _logger
 from reflector.settings import settings
+from reflector.storage import get_source_storage
 from reflector.video_platforms.factory import create_platform_client
 from reflector.worker.process import (
     poll_daily_room_presence_task,
@@ -219,6 +220,30 @@ async def _handle_recording_ready(event: RecordingReadyEvent):
 
         track_keys = [t.s3Key for t in tracks if t.type == "audio"]
 
+        # Delete video tracks when store_video is disabled (same pattern as LiveKit).
+        # Only delete if we have a meeting AND store_video is explicitly false.
+        # If no meeting found, leave files alone (can't confirm user intent).
+        video_track_keys = [t.s3Key for t in tracks if t.type == "video"]
+        if video_track_keys:
+            meeting = await meetings_controller.get_by_room_name(room_name)
+            if meeting is not None and not meeting.store_video:
+                storage = get_source_storage("daily")
+                for video_key in video_track_keys:
+                    try:
+                        await storage.delete_file(video_key)
+                        logger.info(
+                            "Deleted video track from raw-tracks recording",
+                            s3_key=video_key,
+                            room_name=room_name,
+                        )
+                    except Exception as e:
+                        # Non-critical — pipeline filters these out anyway
+                        logger.warning(
+                            "Failed to delete video track from raw-tracks recording",
+                            s3_key=video_key,
+                            error=str(e),
+                        )
+
         logger.info(
             "Raw-tracks recording queuing processing",
             recording_id=recording_id,

server/reflector/views/livekit.py

@@ -0,0 +1,245 @@
+"""LiveKit webhook handler.
+
+Processes LiveKit webhook events for participant tracking and
+Track Egress recording completion.
+
+LiveKit sends webhooks as POST requests with JWT authentication
+in the Authorization header.
+
+Webhooks are used as fast-path triggers and logging. Track discovery
+for the multitrack pipeline uses S3 listing (source of truth), not
+webhook data.
+"""
+
+from fastapi import APIRouter, HTTPException, Request
+
+from reflector.db.meetings import meetings_controller
+from reflector.livekit_api.webhooks import create_webhook_receiver, verify_webhook
+from reflector.logger import logger as _logger
+from reflector.settings import settings
+from reflector.storage import get_source_storage
+
+router = APIRouter()
+
+logger = _logger.bind(platform="livekit")
+
+# Module-level receiver, lazily initialized on first webhook
+_webhook_receiver = None
+
+
+def _get_webhook_receiver():
+    global _webhook_receiver
+    if _webhook_receiver is None:
+        if not settings.LIVEKIT_API_KEY or not settings.LIVEKIT_API_SECRET:
+            raise ValueError("LiveKit not configured")
+        _webhook_receiver = create_webhook_receiver(
+            api_key=settings.LIVEKIT_API_KEY,
+            api_secret=settings.LIVEKIT_WEBHOOK_SECRET or settings.LIVEKIT_API_SECRET,
+        )
+    return _webhook_receiver
+
+
+@router.post("/webhook")
+async def livekit_webhook(request: Request):
+    """Handle LiveKit webhook events.
+
+    LiveKit webhook events include:
+    - participant_joined / participant_left
+    - egress_started / egress_updated / egress_ended
+    - room_started / room_finished
+    - track_published / track_unpublished
+    """
+    if not settings.LIVEKIT_API_KEY or not settings.LIVEKIT_API_SECRET:
+        raise HTTPException(status_code=500, detail="LiveKit not configured")
+
+    body = await request.body()
+    auth_header = request.headers.get("Authorization", "")
+
+    receiver = _get_webhook_receiver()
+    event = verify_webhook(receiver, body, auth_header)
+    if event is None:
+        logger.warning(
+            "Invalid LiveKit webhook signature",
+            has_auth=bool(auth_header),
+            has_body=bool(body),
+        )
+        raise HTTPException(status_code=401, detail="Invalid webhook signature")
+
+    event_type = event.event
+
+    match event_type:
+        case "participant_joined":
+            await _handle_participant_joined(event)
+        case "participant_left":
+            await _handle_participant_left(event)
+        case "egress_started":
+            await _handle_egress_started(event)
+        case "egress_ended":
+            await _handle_egress_ended(event)
+        case "room_started":
+            logger.info(
+                "Room started",
+                room_name=event.room.name if event.room else None,
+            )
+        case "room_finished":
+            await _handle_room_finished(event)
+        case "track_published" | "track_unpublished":
+            logger.debug(
+                f"Track event: {event_type}",
+                room_name=event.room.name if event.room else None,
+                participant=event.participant.identity if event.participant else None,
+            )
+        case _:
+            logger.debug(
+                "Unhandled LiveKit webhook event",
+                event_type=event_type,
+            )
+
+    return {"status": "ok"}
+
+
+async def _handle_participant_joined(event):
+    room_name = event.room.name if event.room else None
+    participant = event.participant
+
+    if not room_name or not participant:
+        logger.warning("participant_joined: missing room or participant data")
+        return
+
+    meeting = await meetings_controller.get_by_room_name(room_name)
+    if not meeting:
+        logger.warning("participant_joined: meeting not found", room_name=room_name)
+        return
+
+    logger.info(
+        "Participant joined",
+        meeting_id=meeting.id,
+        room_name=room_name,
+        participant_identity=participant.identity,
+        participant_sid=participant.sid,
+    )
+
+
+async def _handle_participant_left(event):
+    room_name = event.room.name if event.room else None
+    participant = event.participant
+
+    if not room_name or not participant:
+        logger.warning("participant_left: missing room or participant data")
+        return
+
+    meeting = await meetings_controller.get_by_room_name(room_name)
+    if not meeting:
+        logger.warning("participant_left: meeting not found", room_name=room_name)
+        return
+
+    logger.info(
+        "Participant left",
+        meeting_id=meeting.id,
+        room_name=room_name,
+        participant_identity=participant.identity,
+        participant_sid=participant.sid,
+    )
+
+
+async def _handle_egress_started(event):
+    egress = event.egress_info
+    logger.info(
+        "Egress started",
+        room_name=egress.room_name if egress else None,
+        egress_id=egress.egress_id if egress else None,
+    )
+
+
+async def _handle_egress_ended(event):
+    """Handle Track Egress completion. Delete video files immediately to save storage.
+
+    AutoTrackEgress records ALL tracks (audio + video). Audio is kept for the
+    transcription pipeline. Video files are unused and deleted on completion.
+    This saves ~50x storage (video is 98% of egress output for HD cameras).
+    """
+    egress = event.egress_info
+    if not egress:
+        logger.warning("egress_ended: no egress info in payload")
+        return
+
+    # EGRESS_FAILED = 4
+    if egress.status == 4:
+        logger.error(
+            "Egress failed",
+            room_name=egress.room_name,
+            egress_id=egress.egress_id,
+            error=egress.error,
+        )
+        return
+
+    file_results = list(egress.file_results)
+    logger.info(
+        "Egress ended",
+        room_name=egress.room_name,
+        egress_id=egress.egress_id,
+        status=egress.status,
+        num_files=len(file_results),
+        filenames=[f.filename for f in file_results] if file_results else [],
+    )
+
+    # Delete video files (.webm) immediately — only audio (.ogg) is needed for transcription.
+    # Video tracks are 50-90x larger than audio and unused by the pipeline.
+    # JSON manifests are kept (lightweight metadata, ~430 bytes each).
+    for file_result in file_results:
+        filename = file_result.filename
+        if filename and filename.endswith(".webm"):
+            try:
+                storage = get_source_storage("livekit")
+                await storage.delete_file(filename)
+                logger.info(
+                    "Deleted video egress file",
+                    filename=filename,
+                    room_name=egress.room_name,
+                )
+            except Exception as e:
+                # Non-critical — pipeline filters these out anyway
+                logger.warning(
+                    "Failed to delete video egress file",
+                    filename=filename,
+                    error=str(e),
+                )
+
+
+async def _handle_room_finished(event):
+    """Fast-path: trigger multitrack processing when room closes.
+
+    This is an optimization — if missed, the process_livekit_ended_meetings
+    beat task catches it within ~2 minutes.
+    """
+    room_name = event.room.name if event.room else None
+    if not room_name:
+        logger.warning("room_finished: no room name in payload")
+        return
+
+    logger.info("Room finished", room_name=room_name)
+
+    meeting = await meetings_controller.get_by_room_name(room_name)
+    if not meeting:
+        logger.warning("room_finished: meeting not found", room_name=room_name)
+        return
+
+    # Deactivate the meeting — LiveKit room is destroyed, so process_meetings
+    # can't detect this via API (list_participants returns empty for deleted rooms).
+    if meeting.is_active:
+        await meetings_controller.update_meeting(meeting.id, is_active=False)
+        logger.info("room_finished: meeting deactivated", meeting_id=meeting.id)
+
+    # Import here to avoid circular imports (worker imports views)
+    from reflector.worker.process import process_livekit_multitrack
+
+    process_livekit_multitrack.delay(
+        room_name=room_name,
+        meeting_id=meeting.id,
+    )
+
+    logger.info(
+        "room_finished: queued multitrack processing",
+        meeting_id=meeting.id,
+        room_name=room_name,
+    )

server/reflector/views/meetings.py

@@ -4,7 +4,7 @@ from typing import Annotated, Any, Optional
 from uuid import UUID
 
 from fastapi import APIRouter, Depends, HTTPException, Request
-from pydantic import BaseModel
+from pydantic import BaseModel, EmailStr
 
 import reflector.auth as auth
 from reflector.dailyco_api import RecordingType
@@ -91,9 +91,7 @@ class StartRecordingRequest(BaseModel):
 async def start_recording(
     meeting_id: NonEmptyString,
     body: StartRecordingRequest,
-    user: Annotated[
-        Optional[auth.UserInfo], Depends(auth.current_user_optional_if_public_mode)
-    ],
+    user: Annotated[Optional[auth.UserInfo], Depends(auth.current_user_optional)],
 ) -> dict[str, Any]:
     """Start cloud or raw-tracks recording via Daily.co REST API.
 
@@ -151,3 +149,26 @@ async def start_recording(
         raise HTTPException(
             status_code=500, detail=f"Failed to start recording: {str(e)}"
         )
+
+
+class AddEmailRecipientRequest(BaseModel):
+    email: EmailStr
+
+
+@router.post("/meetings/{meeting_id}/email-recipient")
+async def add_email_recipient(
+    meeting_id: str,
+    request: AddEmailRecipientRequest,
+    user: Annotated[Optional[auth.UserInfo], Depends(auth.current_user_optional)],
+):
+    """Add an email address to receive the transcript link when processing completes."""
+    meeting = await meetings_controller.get_by_id(meeting_id)
+    if not meeting:
+        raise HTTPException(status_code=404, detail="Meeting not found")
+
+    include_link = user is not None
+    recipients = await meetings_controller.add_email_recipient(
+        meeting_id, request.email, include_link=include_link
+    )
+
+    return {"status": "success", "email_recipients": recipients}

server/reflector/views/rooms.py

@@ -1,4 +1,6 @@
 import logging
+import re
+import uuid
 from datetime import datetime, timedelta, timezone
 from enum import Enum
 from typing import Annotated, Any, Literal, Optional
@@ -14,7 +16,7 @@ from reflector.db import get_database
 from reflector.db.calendar_events import calendar_events_controller
 from reflector.db.meetings import meetings_controller
 from reflector.db.rooms import rooms_controller
-from reflector.redis_cache import RedisAsyncLock
+from reflector.redis_cache import RedisAsyncLock, get_async_redis_client
 from reflector.schemas.platform import Platform
 from reflector.services.ics_sync import ics_sync_service
 from reflector.utils.url import add_query_param
@@ -44,6 +46,8 @@ class Room(BaseModel):
     ics_last_etag: Optional[str] = None
     platform: Platform
     skip_consent: bool = False
+    email_transcript_to: str | None = None
+    store_video: bool = False
 
 
 class RoomDetails(Room):
@@ -74,6 +78,7 @@ class Meeting(BaseModel):
     platform: Platform
     daily_composed_video_s3_key: str | None = None
     daily_composed_video_duration: int | None = None
+    store_video: bool = False
 
 
 class CreateRoom(BaseModel):
@@ -93,6 +98,8 @@ class CreateRoom(BaseModel):
     ics_enabled: bool = False
     platform: Platform
     skip_consent: bool = False
+    email_transcript_to: str | None = None
+    store_video: bool = False
 
 
 class UpdateRoom(BaseModel):
@@ -112,6 +119,8 @@ class UpdateRoom(BaseModel):
     ics_enabled: Optional[bool] = None
     platform: Optional[Platform] = None
     skip_consent: Optional[bool] = None
+    email_transcript_to: Optional[str] = None
+    store_video: Optional[bool] = None
 
 
 class CreateRoomMeeting(BaseModel):
@@ -253,6 +262,8 @@ async def rooms_create(
         ics_enabled=room.ics_enabled,
         platform=room.platform,
         skip_consent=room.skip_consent,
+        email_transcript_to=room.email_transcript_to,
+        store_video=room.store_video,
     )
 
 
@@ -321,6 +332,7 @@ async def rooms_create_meeting(
                         and meeting.recording_type == room.recording_type
                         and meeting.recording_trigger == room.recording_trigger
                         and meeting.platform == room.platform
+                        and meeting.store_video == room.store_video
                     )
                     if not settings_match:
                         logger.info(
@@ -550,6 +562,7 @@ async def rooms_join_meeting(
     room_name: str,
     meeting_id: str,
     user: Annotated[Optional[auth.UserInfo], Depends(auth.current_user_optional)],
+    display_name: str | None = None,
 ):
     user_id = user["sub"] if user else None
     room = await rooms_controller.get_by_name(room_name)
@@ -594,4 +607,46 @@ async def rooms_join_meeting(
         meeting = meeting.model_copy()
         meeting.room_url = add_query_param(meeting.room_url, "t", token)
 
+    elif meeting.platform == "livekit":
+        client = create_platform_client(meeting.platform)
+        # Identity must be unique per participant to avoid S3 key collisions.
+        # Format: {readable_name}-{short_uuid} ensures uniqueness even for same names.
+        uid_suffix = uuid.uuid4().hex[:6]
+        if display_name:
+            safe_name = re.sub(r"[^a-zA-Z0-9_-]", "_", display_name.strip())[:40]
+            participant_identity = (
+                f"{safe_name}-{uid_suffix}" if safe_name else f"anon-{uid_suffix}"
+            )
+        elif user_id:
+            email = getattr(user, "email", None)
+            if email and "@" in email:
+                participant_identity = f"{email.split('@')[0]}-{uid_suffix}"
+            else:
+                participant_identity = f"{user_id[:12]}-{uid_suffix}"
+        else:
+            participant_identity = f"anon-{uid_suffix}"
+        participant_name = display_name or participant_identity
+
+        # Store identity → Reflector user_id mapping for the pipeline
+        # (so TranscriptParticipant.user_id can be set correctly)
+        if user_id:
+            redis_client = await get_async_redis_client()
+            mapping_key = f"livekit:participant_map:{meeting.room_name}"
+            await redis_client.hset(mapping_key, participant_identity, user_id)
+            await redis_client.expire(mapping_key, 7 * 86400)  # 7 day TTL
+
+        token = client.create_access_token(
+            room_name=meeting.room_name,
+            participant_identity=participant_identity,
+            participant_name=participant_name,
+            is_admin=user_id == room.user_id if user_id else False,
+        )
+        # Close the platform client to release aiohttp session
+        if hasattr(client, "close"):
+            await client.close()
+
+        meeting = meeting.model_copy()
+        # For LiveKit, room_url is the WS URL; token goes as a query param
+        meeting.room_url = add_query_param(meeting.room_url, "token", token)
+
     return meeting

server/reflector/views/rtc_offer.py

@@ -11,6 +11,7 @@ from reflector.events import subscribers_shutdown
 from reflector.logger import logger
 from reflector.pipelines.runner import PipelineRunner
 from reflector.settings import settings
+from reflector.webrtc_ports import resolve_webrtc_host, rewrite_sdp_host
 
 sessions = []
 router = APIRouter()
@@ -128,8 +129,6 @@ async def rtc_offer_base(
 
     # Rewrite ICE candidate IPs when running behind Docker bridge networking
     if settings.WEBRTC_HOST:
-        from reflector.webrtc_ports import resolve_webrtc_host, rewrite_sdp_host
-
         host_ip = resolve_webrtc_host(settings.WEBRTC_HOST)
         sdp = rewrite_sdp_host(sdp, host_ip)
 

server/reflector/views/transcripts.py

@@ -16,6 +16,7 @@ from pydantic import (
 
 import reflector.auth as auth
 from reflector.db import get_database
+from reflector.db.meetings import meetings_controller
 from reflector.db.recordings import recordings_controller
 from reflector.db.rooms import rooms_controller
 from reflector.db.search import (
@@ -39,6 +40,7 @@ from reflector.db.transcripts import (
     transcripts_controller,
 )
 from reflector.db.users import user_controller
+from reflector.email import is_email_configured, send_transcript_email
 from reflector.processors.types import Transcript as ProcessorTranscript
 from reflector.processors.types import Word
 from reflector.schemas.transcript_formats import TranscriptFormat, TranscriptSegment
@@ -112,6 +114,8 @@ class GetTranscriptMinimal(BaseModel):
     room_name: str | None = None
     audio_deleted: bool | None = None
     change_seq: int | None = None
+    has_cloud_video: bool = False
+    cloud_video_duration: int | None = None
 
 
 class TranscriptParticipantWithEmail(TranscriptParticipant):
@@ -305,6 +309,7 @@ async def transcripts_search(
     source_kind: Optional[SourceKind] = None,
     from_datetime: SearchFromDatetimeParam = None,
     to_datetime: SearchToDatetimeParam = None,
+    include_deleted: bool = False,
     user: Annotated[
         Optional[auth.UserInfo], Depends(auth.current_user_optional_if_public_mode)
     ] = None,
@@ -312,6 +317,12 @@ async def transcripts_search(
     """Full-text search across transcript titles and content."""
     user_id = user["sub"] if user else None
 
+    if include_deleted and not user_id:
+        raise HTTPException(
+            status_code=401,
+            detail="Authentication required to view deleted transcripts",
+        )
+
     if from_datetime and to_datetime and from_datetime > to_datetime:
         raise HTTPException(
             status_code=400, detail="'from' must be less than or equal to 'to'"
@@ -326,6 +337,7 @@ async def transcripts_search(
         source_kind=source_kind,
         from_datetime=from_datetime,
         to_datetime=to_datetime,
+        include_deleted=include_deleted,
     )
 
     results, total = await search_controller.search_transcripts(search_params)
@@ -501,6 +513,14 @@ async def transcript_get(
                 )
             )
 
+    has_cloud_video = False
+    cloud_video_duration = None
+    if transcript.meeting_id:
+        meeting = await meetings_controller.get_by_id(transcript.meeting_id)
+        if meeting and meeting.daily_composed_video_s3_key:
+            has_cloud_video = True
+            cloud_video_duration = meeting.daily_composed_video_duration
+
     base_data = {
         "id": transcript.id,
         "user_id": transcript.user_id,
@@ -524,6 +544,8 @@ async def transcript_get(
         "audio_deleted": transcript.audio_deleted,
         "change_seq": transcript.change_seq,
         "participants": participants,
+        "has_cloud_video": has_cloud_video,
+        "cloud_video_duration": cloud_video_duration,
     }
 
     if transcript_format == "text":
@@ -601,6 +623,54 @@ async def transcript_delete(
     return DeletionStatus(status="ok")
 
 
+@router.post("/transcripts/{transcript_id}/restore", response_model=DeletionStatus)
+async def transcript_restore(
+    transcript_id: str,
+    user: Annotated[auth.UserInfo, Depends(auth.current_user)],
+):
+    """Restore a soft-deleted transcript."""
+    user_id = user["sub"]
+    transcript = await transcripts_controller.get_by_id(transcript_id)
+    if not transcript:
+        raise HTTPException(status_code=404, detail="Transcript not found")
+    if transcript.deleted_at is None:
+        raise HTTPException(status_code=400, detail="Transcript is not deleted")
+    if not transcripts_controller.user_can_mutate(transcript, user_id):
+        raise HTTPException(status_code=403, detail="Not authorized")
+
+    await transcripts_controller.restore_by_id(transcript.id, user_id=user_id)
+    await get_ws_manager().send_json(
+        room_id=f"user:{user_id}",
+        message={"event": "TRANSCRIPT_RESTORED", "data": {"id": transcript.id}},
+    )
+    return DeletionStatus(status="ok")
+
+
+@router.delete("/transcripts/{transcript_id}/destroy", response_model=DeletionStatus)
+async def transcript_destroy(
+    transcript_id: str,
+    user: Annotated[auth.UserInfo, Depends(auth.current_user)],
+):
+    """Permanently delete a transcript and all associated files."""
+    user_id = user["sub"]
+    transcript = await transcripts_controller.get_by_id(transcript_id)
+    if not transcript:
+        raise HTTPException(status_code=404, detail="Transcript not found")
+    if transcript.deleted_at is None:
+        raise HTTPException(
+            status_code=400, detail="Transcript must be soft-deleted first"
+        )
+    if not transcripts_controller.user_can_mutate(transcript, user_id):
+        raise HTTPException(status_code=403, detail="Not authorized")
+
+    await transcripts_controller.hard_delete(transcript.id)
+    await get_ws_manager().send_json(
+        room_id=f"user:{user_id}",
+        message={"event": "TRANSCRIPT_DELETED", "data": {"id": transcript.id}},
+    )
+    return DeletionStatus(status="ok")
+
+
 @router.get(
     "/transcripts/{transcript_id}/topics",
     response_model=list[GetTranscriptTopic],
@@ -686,8 +756,6 @@ async def transcript_post_to_zulip(
     )
     if not transcript:
         raise HTTPException(status_code=404, detail="Transcript not found")
-    if not transcripts_controller.user_can_mutate(transcript, user_id):
-        raise HTTPException(status_code=403, detail="Not authorized")
     content = get_zulip_message(transcript, include_topics)
 
     message_updated = False
@@ -705,3 +773,31 @@ async def transcript_post_to_zulip(
         await transcripts_controller.update(
             transcript, {"zulip_message_id": response["id"]}
         )
+
+
+class SendEmailRequest(BaseModel):
+    email: str
+
+
+class SendEmailResponse(BaseModel):
+    sent: int
+
+
+@router.post("/transcripts/{transcript_id}/email", response_model=SendEmailResponse)
+async def transcript_send_email(
+    transcript_id: str,
+    request: SendEmailRequest,
+    user: Annotated[Optional[auth.UserInfo], Depends(auth.current_user_optional)],
+):
+    if not is_email_configured():
+        raise HTTPException(status_code=400, detail="Email not configured")
+    user_id = user["sub"] if user else None
+    transcript = await transcripts_controller.get_by_id_for_http(
+        transcript_id, user_id=user_id
+    )
+    if not transcript:
+        raise HTTPException(status_code=404, detail="Transcript not found")
+    sent = await send_transcript_email(
+        [request.email], transcript, include_link=(transcript.share_mode == "public")
+    )
+    return SendEmailResponse(sent=sent)

server/reflector/views/transcripts_audio.py

@@ -53,9 +53,22 @@ async def transcript_get_audio_mp3(
         else:
             user_id = token_user["sub"]
 
-    transcript = await transcripts_controller.get_by_id_for_http(
-        transcript_id, user_id=user_id
-    )
+    if not user_id and not token:
+        # No authentication provided at all. Only anonymous transcripts
+        # (user_id=None) are accessible without auth, to preserve
+        # pipeline access via _generate_local_audio_link().
+        transcript = await transcripts_controller.get_by_id(transcript_id)
+        if not transcript or transcript.deleted_at is not None:
+            raise HTTPException(status_code=404, detail="Transcript not found")
+        if transcript.user_id is not None:
+            raise HTTPException(
+                status_code=status.HTTP_401_UNAUTHORIZED,
+                detail="Authentication required",
+            )
+    else:
+        transcript = await transcripts_controller.get_by_id_for_http(
+            transcript_id, user_id=user_id
+        )
 
     if transcript.audio_location == "storage":
         # proxy S3 file, to prevent issue with CORS
@@ -94,16 +107,16 @@ async def transcript_get_audio_mp3(
         request,
         transcript.audio_mp3_filename,
         content_type="audio/mpeg",
-        content_disposition=f"attachment; filename={filename}",
+        content_disposition=f"inline; filename={filename}",
     )
 
 
 @router.get("/transcripts/{transcript_id}/audio/waveform")
 async def transcript_get_audio_waveform(
     transcript_id: str,
-    user: Annotated[Optional[auth.UserInfo], Depends(auth.current_user_optional)],
+    user: Annotated[auth.UserInfo, Depends(auth.current_user)],
 ) -> AudioWaveform:
-    user_id = user["sub"] if user else None
+    user_id = user["sub"]
     transcript = await transcripts_controller.get_by_id_for_http(
         transcript_id, user_id=user_id
     )

server/reflector/views/transcripts_download.py

@@ -0,0 +1,169 @@
+"""
+Transcript download endpoint — generates a zip archive with all transcript files.
+"""
+
+import json
+import os
+import tempfile
+import zipfile
+from typing import Annotated
+
+from fastapi import APIRouter, Depends, HTTPException
+from fastapi.responses import StreamingResponse
+
+import reflector.auth as auth
+from reflector.db.meetings import meetings_controller
+from reflector.db.recordings import recordings_controller
+from reflector.db.transcripts import transcripts_controller
+from reflector.logger import logger
+from reflector.storage import get_source_storage, get_transcripts_storage
+
+router = APIRouter()
+
+
+@router.get(
+    "/transcripts/{transcript_id}/download/zip",
+    operation_id="transcript_download_zip",
+)
+async def transcript_download_zip(
+    transcript_id: str,
+    user: Annotated[auth.UserInfo, Depends(auth.current_user)],
+):
+    user_id = user["sub"]
+    transcript = await transcripts_controller.get_by_id_for_http(
+        transcript_id, user_id=user_id
+    )
+    if not transcripts_controller.user_can_mutate(transcript, user_id):
+        raise HTTPException(status_code=403, detail="Not authorized")
+
+    recording = None
+    if transcript.recording_id:
+        recording = await recordings_controller.get_by_id(transcript.recording_id)
+
+    meeting = None
+    if transcript.meeting_id:
+        meeting = await meetings_controller.get_by_id(transcript.meeting_id)
+
+    truncated_id = str(transcript.id).split("-")[0]
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        zip_path = os.path.join(tmpdir, f"transcript_{truncated_id}.zip")
+
+        with zipfile.ZipFile(zip_path, "w", zipfile.ZIP_DEFLATED) as zf:
+            # Transcript audio
+            if transcript.audio_location == "storage" and not transcript.audio_deleted:
+                try:
+                    storage = get_transcripts_storage()
+                    data = await storage.get_file(transcript.storage_audio_path)
+                    audio_path = os.path.join(tmpdir, "audio.mp3")
+                    with open(audio_path, "wb") as f:
+                        f.write(data)
+                    zf.write(audio_path, "audio.mp3")
+                except Exception as e:
+                    logger.warning(
+                        "Failed to download transcript audio for zip",
+                        exc_info=e,
+                        transcript_id=transcript.id,
+                    )
+            elif (
+                not transcript.audio_deleted
+                and hasattr(transcript, "audio_mp3_filename")
+                and transcript.audio_mp3_filename
+                and transcript.audio_mp3_filename.exists()
+            ):
+                zf.write(str(transcript.audio_mp3_filename), "audio.mp3")
+
+            # Recording tracks (multitrack)
+            if recording and recording.track_keys:
+                try:
+                    source_storage = get_source_storage(
+                        "daily" if recording.track_keys else None
+                    )
+                except Exception:
+                    source_storage = get_transcripts_storage()
+
+                for i, key in enumerate(recording.track_keys):
+                    try:
+                        data = await source_storage.get_file(
+                            key, bucket=recording.bucket_name
+                        )
+                        filename = os.path.basename(key) or f"track_{i}"
+                        track_path = os.path.join(tmpdir, f"track_{i}")
+                        with open(track_path, "wb") as f:
+                            f.write(data)
+                        zf.write(track_path, f"tracks/{filename}")
+                    except Exception as e:
+                        logger.warning(
+                            "Failed to download track for zip",
+                            exc_info=e,
+                            track_key=key,
+                        )
+
+            # Cloud video
+            if meeting and meeting.daily_composed_video_s3_key:
+                try:
+                    source_storage = get_source_storage("daily")
+                    data = await source_storage.get_file(
+                        meeting.daily_composed_video_s3_key
+                    )
+                    video_path = os.path.join(tmpdir, "cloud_video.mp4")
+                    with open(video_path, "wb") as f:
+                        f.write(data)
+                    zf.write(video_path, "cloud_video.mp4")
+                except Exception as e:
+                    logger.warning(
+                        "Failed to download cloud video for zip",
+                        exc_info=e,
+                        s3_key=meeting.daily_composed_video_s3_key,
+                    )
+
+            # Metadata JSON
+            metadata = {
+                "id": transcript.id,
+                "title": transcript.title,
+                "created_at": (
+                    transcript.created_at.isoformat() if transcript.created_at else None
+                ),
+                "duration": transcript.duration,
+                "source_language": transcript.source_language,
+                "target_language": transcript.target_language,
+                "short_summary": transcript.short_summary,
+                "long_summary": transcript.long_summary,
+                "topics": (
+                    [t.model_dump() for t in transcript.topics]
+                    if transcript.topics
+                    else []
+                ),
+                "participants": (
+                    [p.model_dump() for p in transcript.participants]
+                    if transcript.participants
+                    else []
+                ),
+                "action_items": transcript.action_items,
+                "webvtt": transcript.webvtt,
+                "recording_id": transcript.recording_id,
+                "meeting_id": transcript.meeting_id,
+            }
+            meta_path = os.path.join(tmpdir, "metadata.json")
+            with open(meta_path, "w") as f:
+                json.dump(metadata, f, indent=2, default=str)
+            zf.write(meta_path, "metadata.json")
+
+        # Read zip into memory before tmpdir is cleaned up
+        with open(zip_path, "rb") as f:
+            zip_bytes = f.read()
+
+    def iter_zip():
+        offset = 0
+        chunk_size = 64 * 1024
+        while offset < len(zip_bytes):
+            yield zip_bytes[offset : offset + chunk_size]
+            offset += chunk_size
+
+    return StreamingResponse(
+        iter_zip(),
+        media_type="application/zip",
+        headers={
+            "Content-Disposition": f"attachment; filename=transcript_{truncated_id}.zip"
+        },
+    )

server/reflector/views/transcripts_video.py

@@ -0,0 +1,60 @@
+"""
+Transcript cloud video endpoint — returns a presigned URL for streaming playback.
+"""
+
+from typing import Annotated
+
+from fastapi import APIRouter, Depends, HTTPException
+from pydantic import BaseModel
+
+import reflector.auth as auth
+from reflector.db.meetings import meetings_controller
+from reflector.db.transcripts import transcripts_controller
+from reflector.storage import get_source_storage
+
+router = APIRouter()
+
+
+class VideoUrlResponse(BaseModel):
+    url: str
+    duration: int | None = None
+    content_type: str = "video/mp4"
+
+
+@router.get(
+    "/transcripts/{transcript_id}/video/url",
+    operation_id="transcript_get_video_url",
+    response_model=VideoUrlResponse,
+)
+async def transcript_get_video_url(
+    transcript_id: str,
+    user: Annotated[auth.UserInfo, Depends(auth.current_user)],
+):
+    user_id = user["sub"]
+
+    transcript = await transcripts_controller.get_by_id_for_http(
+        transcript_id, user_id=user_id
+    )
+
+    if not transcript.meeting_id:
+        raise HTTPException(status_code=404, detail="No video available")
+
+    meeting = await meetings_controller.get_by_id(transcript.meeting_id)
+    if not meeting or not meeting.daily_composed_video_s3_key:
+        raise HTTPException(status_code=404, detail="No video available")
+
+    source_storage = get_source_storage("daily")
+    url = await source_storage.get_file_url(
+        meeting.daily_composed_video_s3_key,
+        operation="get_object",
+        expires_in=900,
+        extra_params={
+            "ResponseContentDisposition": "inline",
+            "ResponseContentType": "video/mp4",
+        },
+    )
+
+    return VideoUrlResponse(
+        url=url,
+        duration=meeting.daily_composed_video_duration,
+    )

server/reflector/views/transcripts_webrtc.py

@@ -4,6 +4,7 @@ from fastapi import APIRouter, Depends, HTTPException, Request
 
 import reflector.auth as auth
 from reflector.db.transcripts import transcripts_controller
+from reflector.pipelines.main_live_pipeline import PipelineMainLive
 
 from .rtc_offer import RtcOffer, rtc_offer_base
 
@@ -28,8 +29,6 @@ async def transcript_record_webrtc(
         raise HTTPException(status_code=400, detail="Transcript is locked")
 
     # create a pipeline runner
-    from reflector.pipelines.main_live_pipeline import PipelineMainLive  # noqa: PLC0415
-
     pipeline_runner = PipelineMainLive(transcript_id=transcript_id)
 
     # FIXME do not allow multiple recording at the same time

server/reflector/webrtc_ports.py

@@ -11,6 +11,8 @@ This allows running the server in Docker with bridge networking
 import asyncio
 import socket
 
+import aioice.ice
+
 from reflector.logger import logger
 
 
@@ -36,9 +38,7 @@ def patch_aioice_port_range(min_port: int, max_port: int) -> None:
     Works by temporarily wrapping loop.create_datagram_endpoint() during
     aioice's get_component_candidates() to intercept bind(addr, 0) calls.
     """
-    import aioice.ice as _ice
-
-    _original = _ice.Connection.get_component_candidates
+    _original = aioice.ice.Connection.get_component_candidates
     _state = {"next_port": min_port}
 
     async def _patched_get_component_candidates(self, component, addresses, timeout=5):
@@ -78,7 +78,7 @@ def patch_aioice_port_range(min_port: int, max_port: int) -> None:
         finally:
             loop.create_datagram_endpoint = _orig_create
 
-    _ice.Connection.get_component_candidates = _patched_get_component_candidates
+    aioice.ice.Connection.get_component_candidates = _patched_get_component_candidates
     logger.info(
         "aioice patched for WebRTC port range",
         min_port=min_port,
@@ -102,8 +102,6 @@ def rewrite_sdp_host(sdp: str, target_ip: str) -> str:
     Replace container-internal IPs in SDP with target_ip so that
     ICE candidates advertise a routable address.
     """
-    import aioice.ice
-
     container_ips = aioice.ice.get_host_addresses(use_ipv4=True, use_ipv6=False)
     for ip in container_ips:
         if ip != "127.0.0.1" and ip != target_ip:

server/reflector/worker/app.py

@@ -30,6 +30,8 @@ def build_beat_schedule(
     whereby_api_key=None,
     aws_process_recording_queue_url=None,
     daily_api_key=None,
+    livekit_api_key=None,
+    livekit_url=None,
     public_mode=False,
     public_data_retention_days=None,
     healthcheck_url=None,
@@ -83,7 +85,25 @@ def build_beat_schedule(
     else:
         logger.info("Daily.co beat tasks disabled (no DAILY_API_KEY)")
 
-    _any_platform = _whereby_enabled or _daily_enabled
+    _livekit_enabled = bool(livekit_api_key and livekit_url)
+    if _livekit_enabled:
+        beat_schedule["process_livekit_ended_meetings"] = {
+            "task": "reflector.worker.process.process_livekit_ended_meetings",
+            "schedule": 120,  # Every 2 minutes
+        }
+        beat_schedule["reprocess_failed_livekit_recordings"] = {
+            "task": "reflector.worker.process.reprocess_failed_livekit_recordings",
+            "schedule": crontab(hour=5, minute=0),
+        }
+        logger.info(
+            "LiveKit beat tasks enabled",
+            tasks=[
+                "process_livekit_ended_meetings",
+                "reprocess_failed_livekit_recordings",
+            ],
+        )
+
+    _any_platform = _whereby_enabled or _daily_enabled or _livekit_enabled
     if _any_platform:
         beat_schedule["process_meetings"] = {
             "task": "reflector.worker.process.process_meetings",
@@ -146,7 +166,6 @@ else:
     app.conf.broker_connection_retry_on_startup = True
     app.autodiscover_tasks(
         [
-            "reflector.pipelines.main_live_pipeline",
             "reflector.worker.healthcheck",
             "reflector.worker.process",
             "reflector.worker.cleanup",
@@ -158,6 +177,8 @@ else:
         whereby_api_key=settings.WHEREBY_API_KEY,
         aws_process_recording_queue_url=settings.AWS_PROCESS_RECORDING_QUEUE_URL,
         daily_api_key=settings.DAILY_API_KEY,
+        livekit_api_key=settings.LIVEKIT_API_KEY,
+        livekit_url=settings.LIVEKIT_URL,
         public_mode=settings.PUBLIC_MODE,
         public_data_retention_days=settings.PUBLIC_DATA_RETENTION_DAYS,
         healthcheck_url=settings.HEALTHCHECK_URL,

server/reflector/worker/cleanup.py

@@ -90,7 +90,9 @@ async def cleanup_old_transcripts(
 ):
     """Delete old anonymous transcripts and their associated recordings/meetings."""
     query = transcripts.select().where(
-        (transcripts.c.created_at < cutoff_date) & (transcripts.c.user_id.is_(None))
+        (transcripts.c.created_at < cutoff_date)
+        & (transcripts.c.user_id.is_(None))
+        & (transcripts.c.deleted_at.is_(None))
     )
     old_transcripts = await db.fetch_all(query)
 

server/reflector/worker/process.py

@@ -1,3 +1,4 @@
+import asyncio
 import json
 import os
 import re
@@ -12,6 +13,7 @@ from celery import shared_task
 from celery.utils.log import get_task_logger
 from pydantic import ValidationError
 
+from reflector.asynctask import asynctask
 from reflector.dailyco_api import FinishedRecordingResponse, RecordingResponse
 from reflector.db.daily_participant_sessions import (
     DailyParticipantSession,
@@ -25,19 +27,26 @@ from reflector.db.transcripts import (
     transcripts_controller,
 )
 from reflector.hatchet.client import HatchetClientManager
-from reflector.pipelines.main_live_pipeline import asynctask
 from reflector.pipelines.topic_processing import EmptyPipeline
-from reflector.processors import AudioFileWriterProcessor
+from reflector.processors.audio_file_writer import AudioFileWriterProcessor
 from reflector.processors.audio_waveform_processor import AudioWaveformProcessor
 from reflector.redis_cache import RedisAsyncLock
 from reflector.settings import settings
-from reflector.storage import get_transcripts_storage
+from reflector.storage import get_source_storage, get_transcripts_storage
 from reflector.utils.daily import (
     DailyRoomName,
     extract_base_room_name,
     filter_cam_audio_tracks,
     recording_lock_key,
 )
+from reflector.utils.livekit import (
+    extract_livekit_base_room_name,
+    filter_audio_tracks,
+    parse_livekit_track_filepath,
+)
+from reflector.utils.livekit import (
+    recording_lock_key as livekit_recording_lock_key,
+)
 from reflector.utils.string import NonEmptyString
 from reflector.video_platforms.factory import create_platform_client
 from reflector.video_platforms.whereby_utils import (
@@ -104,6 +113,12 @@ async def process_recording(bucket_name: str, object_key: str):
     room = await rooms_controller.get_by_id(meeting.room_id)
 
     recording = await recordings_controller.get_by_object_key(bucket_name, object_key)
+    if recording and recording.deleted_at is not None:
+        logger.info(
+            "Skipping soft-deleted recording",
+            recording_id=recording.id,
+        )
+        return
     if not recording:
         recording = await recordings_controller.create(
             Recording(
@@ -115,6 +130,13 @@ async def process_recording(bucket_name: str, object_key: str):
         )
 
     transcript = await transcripts_controller.get_by_recording_id(recording.id)
+    if transcript and transcript.deleted_at is not None:
+        logger.info(
+            "Skipping soft-deleted transcript for recording",
+            recording_id=recording.id,
+            transcript_id=transcript.id,
+        )
+        return
     if transcript:
         await transcripts_controller.update(
             transcript,
@@ -262,6 +284,13 @@ async def _process_multitrack_recording_inner(
     # Check if recording already exists (reprocessing path)
     recording = await recordings_controller.get_by_id(recording_id)
 
+    if recording and recording.deleted_at is not None:
+        logger.info(
+            "Skipping soft-deleted recording",
+            recording_id=recording_id,
+        )
+        return
+
     if recording and recording.meeting_id:
         # Reprocessing: recording exists with meeting already linked
         meeting = await meetings_controller.get_by_id(recording.meeting_id)
@@ -341,6 +370,13 @@ async def _process_multitrack_recording_inner(
         )
 
     transcript = await transcripts_controller.get_by_recording_id(recording.id)
+    if transcript and transcript.deleted_at is not None:
+        logger.info(
+            "Skipping soft-deleted transcript for recording",
+            recording_id=recording.id,
+            transcript_id=transcript.id,
+        )
+        return
     if not transcript:
         transcript = await transcripts_controller.add(
             "",
@@ -537,6 +573,15 @@ async def store_cloud_recording(
         )
         return False
 
+    if not meeting.store_video:
+        logger.info(
+            f"Cloud recording ({source}): skipped, store_video=false",
+            recording_id=recording_id,
+            room_name=room_name,
+            meeting_id=meeting.id,
+        )
+        return False
+
     success = await meetings_controller.set_cloud_recording_if_missing(
         meeting_id=meeting.id,
         s3_key=s3_key,
@@ -844,13 +889,30 @@ async def process_meetings():
                 elif has_had_sessions:
                     should_deactivate = True
                     logger_.info("Meeting ended - all participants left")
-                elif current_time > end_date:
-                    should_deactivate = True
-                    logger_.info(
-                        "Meeting deactivated - scheduled time ended with no participants",
-                    )
-                else:
-                    logger_.debug("Meeting not yet started, keep it")
+                elif not has_had_sessions:
+                    # No sessions recorded — either no one joined, or webhooks
+                    # didn't arrive (e.g. local dev without tunnel).
+                    meeting_start = meeting.start_date
+                    if meeting_start.tzinfo is None:
+                        meeting_start = meeting_start.replace(tzinfo=timezone.utc)
+                    age_minutes = (current_time - meeting_start).total_seconds() / 60
+                    is_scheduled = bool(meeting.calendar_event_id)
+
+                    if is_scheduled and current_time > end_date:
+                        # Scheduled meeting past its end time with no participants
+                        should_deactivate = True
+                        logger_.info(
+                            "Meeting deactivated - scheduled time ended with no participants",
+                        )
+                    elif not is_scheduled and age_minutes > 30:
+                        # On-the-fly meeting with no sessions after 30 min
+                        should_deactivate = True
+                        logger_.info(
+                            "Meeting deactivated - no sessions after 30 min",
+                            age_minutes=round(age_minutes, 1),
+                        )
+                    else:
+                        logger_.debug("Meeting not yet started, keep it")
 
                 if should_deactivate:
                     await meetings_controller.update_meeting(
@@ -1140,3 +1202,298 @@ async def trigger_daily_reconciliation() -> None:
 
     except Exception as e:
         logger.error("Reconciliation trigger failed", error=str(e), exc_info=True)
+
+
+# ============================================================
+# LiveKit multitrack recording tasks
+# ============================================================
+
+
+@shared_task
+@asynctask
+async def process_livekit_multitrack(
+    room_name: str,
+    meeting_id: str,
+):
+    """
+    Process LiveKit multitrack recording by discovering tracks on S3.
+
+    Tracks are discovered via S3 listing (source of truth), not webhooks.
+    Called from room_finished webhook (fast-path) or beat task (fallback).
+    """
+    logger.info(
+        "Processing LiveKit multitrack recording",
+        room_name=room_name,
+        meeting_id=meeting_id,
+    )
+
+    lock_key = livekit_recording_lock_key(room_name)
+    async with RedisAsyncLock(
+        key=lock_key,
+        timeout=600,
+        extend_interval=60,
+        skip_if_locked=True,
+        blocking=False,
+    ) as lock:
+        if not lock.acquired:
+            logger.warning(
+                "LiveKit processing skipped - lock already held",
+                room_name=room_name,
+                lock_key=lock_key,
+            )
+            return
+
+        await _process_livekit_multitrack_inner(room_name, meeting_id)
+
+
+async def _process_livekit_multitrack_inner(
+    room_name: str,
+    meeting_id: str,
+):
+    """Inner processing logic for LiveKit multitrack recording."""
+    # 1. Discover tracks by listing S3 prefix.
+    # Wait briefly for egress files to finish flushing to S3 — the room_finished
+    # webhook fires after empty_timeout, but egress finalization may still be in progress.
+    EGRESS_FLUSH_DELAY = 10  # seconds — egress typically flushes within a few seconds
+    EGRESS_RETRY_DELAY = 30  # seconds — retry if first listing finds nothing
+
+    await asyncio.sleep(EGRESS_FLUSH_DELAY)
+
+    storage = get_source_storage("livekit")
+    s3_prefix = f"livekit/{room_name}/"
+    all_keys = await storage.list_objects(prefix=s3_prefix)
+
+    # Filter to audio tracks only (.ogg) — skip .json manifests and .webm video
+    audio_keys = filter_audio_tracks(all_keys) if all_keys else []
+
+    if not audio_keys:
+        # Retry once after a longer delay — egress may still be flushing
+        logger.info(
+            "No audio tracks found yet, retrying after delay",
+            room_name=room_name,
+            retry_delay=EGRESS_RETRY_DELAY,
+        )
+        await asyncio.sleep(EGRESS_RETRY_DELAY)
+        all_keys = await storage.list_objects(prefix=s3_prefix)
+        audio_keys = filter_audio_tracks(all_keys) if all_keys else []
+
+    # Sanity check: compare audio tracks against egress manifests.
+    # Each Track Egress (audio or video) produces a .json manifest.
+    # Video tracks produce .webm files. So expected audio count ≈ manifests - video files.
+    if all_keys:
+        manifest_count = sum(1 for k in all_keys if k.endswith(".json"))
+        video_count = sum(1 for k in all_keys if k.endswith(".webm"))
+        expected_audio = manifest_count - video_count
+        if expected_audio > len(audio_keys) and expected_audio > 0:
+            # Some audio tracks may still be flushing — wait and retry
+            logger.info(
+                "Expected more audio tracks based on manifests, waiting for late flushes",
+                room_name=room_name,
+                expected=expected_audio,
+                found=len(audio_keys),
+            )
+            await asyncio.sleep(EGRESS_RETRY_DELAY)
+            all_keys = await storage.list_objects(prefix=s3_prefix)
+            audio_keys = filter_audio_tracks(all_keys) if all_keys else []
+
+    logger.info(
+        "S3 track discovery complete",
+        room_name=room_name,
+        total_files=len(all_keys) if all_keys else 0,
+        audio_files=len(audio_keys),
+    )
+
+    if not audio_keys:
+        logger.warning(
+            "No audio track files found on S3 after retries",
+            room_name=room_name,
+            s3_prefix=s3_prefix,
+        )
+        return
+
+    # 2. Parse track info from filenames
+    parsed_tracks = []
+    for key in audio_keys:
+        try:
+            parsed = parse_livekit_track_filepath(key)
+            parsed_tracks.append(parsed)
+        except ValueError as e:
+            logger.warning("Skipping unparseable track file", s3_key=key, error=str(e))
+
+    if not parsed_tracks:
+        logger.warning(
+            "No valid track files found after parsing",
+            room_name=room_name,
+            raw_keys=all_keys,
+        )
+        return
+
+    track_keys = [t.s3_key for t in parsed_tracks]
+
+    # 3. Find meeting and room
+    meeting = await meetings_controller.get_by_id(meeting_id)
+    if not meeting:
+        logger.error(
+            "Meeting not found for LiveKit recording",
+            meeting_id=meeting_id,
+            room_name=room_name,
+        )
+        return
+
+    base_room_name = extract_livekit_base_room_name(room_name)
+    room = await rooms_controller.get_by_name(base_room_name)
+    if not room:
+        logger.error("Room not found", room_name=base_room_name)
+        return
+
+    # 4. Create recording
+    recording_id = f"lk-{room_name}"
+    bucket_name = settings.LIVEKIT_STORAGE_AWS_BUCKET_NAME or ""
+
+    existing_recording = await recordings_controller.get_by_id(recording_id)
+    if existing_recording and existing_recording.deleted_at is not None:
+        logger.info("Skipping soft-deleted recording", recording_id=recording_id)
+        return
+
+    if not existing_recording:
+        recording = await recordings_controller.create(
+            Recording(
+                id=recording_id,
+                bucket_name=bucket_name,
+                object_key=s3_prefix,
+                recorded_at=datetime.now(timezone.utc),
+                meeting_id=meeting.id,
+                track_keys=track_keys,
+            )
+        )
+    else:
+        recording = existing_recording
+
+    # 5. Create or get transcript
+    transcript = await transcripts_controller.get_by_recording_id(recording.id)
+    if transcript and transcript.deleted_at is not None:
+        logger.info("Skipping soft-deleted transcript", recording_id=recording.id)
+        return
+    if not transcript:
+        transcript = await transcripts_controller.add(
+            "",
+            source_kind=SourceKind.ROOM,
+            source_language="en",
+            target_language="en",
+            user_id=room.user_id,
+            recording_id=recording.id,
+            share_mode="semi-private",
+            meeting_id=meeting.id,
+            room_id=room.id,
+        )
+
+    # 6. Start Hatchet pipeline (reuses DiarizationPipeline with source_platform="livekit")
+    workflow_id = await HatchetClientManager.start_workflow(
+        workflow_name="DiarizationPipeline",
+        input_data={
+            "recording_id": recording_id,
+            "tracks": [
+                {
+                    "s3_key": t.s3_key,
+                    "participant_identity": t.participant_identity,
+                    "timestamp": t.timestamp.isoformat(),
+                }
+                for t in parsed_tracks
+            ],
+            "bucket_name": bucket_name,
+            "transcript_id": transcript.id,
+            "room_id": room.id,
+            "source_platform": "livekit",
+        },
+        additional_metadata={
+            "transcript_id": transcript.id,
+            "recording_id": recording_id,
+        },
+    )
+    logger.info(
+        "Started LiveKit Hatchet workflow",
+        workflow_id=workflow_id,
+        transcript_id=transcript.id,
+        room_name=room_name,
+        num_tracks=len(parsed_tracks),
+    )
+
+    await transcripts_controller.update(transcript, {"workflow_run_id": workflow_id})
+
+
+@shared_task
+@asynctask
+async def process_livekit_ended_meetings():
+    """Check for inactive LiveKit meetings that need multitrack processing.
+
+    Runs on a beat schedule. Catches cases where room_finished webhook was missed.
+    Only processes meetings that:
+    - Platform is "livekit"
+    - is_active=False (already deactivated by process_meetings)
+    - No associated transcript yet
+    """
+    from reflector.db.transcripts import transcripts_controller as tc  # noqa: PLC0415
+
+    all_livekit = await meetings_controller.get_all_inactive_livekit()
+
+    queued = 0
+    for meeting in all_livekit:
+        # Skip if already has a transcript
+        existing = await tc.get_by_meeting_id(meeting.id)
+        if existing:
+            continue
+
+        logger.info(
+            "Found unprocessed inactive LiveKit meeting",
+            meeting_id=meeting.id,
+            room_name=meeting.room_name,
+        )
+
+        process_livekit_multitrack.delay(
+            room_name=meeting.room_name,
+            meeting_id=meeting.id,
+        )
+        queued += 1
+
+    if queued > 0:
+        logger.info("Queued LiveKit multitrack processing", count=queued)
+
+
+@shared_task
+@asynctask
+async def reprocess_failed_livekit_recordings():
+    """Reprocess LiveKit recordings that failed.
+
+    Runs daily at 5 AM. Finds recordings with livekit prefix and error status.
+    """
+    bucket_name = settings.LIVEKIT_STORAGE_AWS_BUCKET_NAME
+    if not bucket_name:
+        return
+
+    failed = await recordings_controller.get_multitrack_needing_reprocessing(
+        bucket_name
+    )
+    livekit_failed = [r for r in failed if r.id.startswith("lk-")]
+
+    for recording in livekit_failed:
+        if not recording.meeting_id:
+            logger.warning(
+                "Skipping reprocess — no meeting_id",
+                recording_id=recording.id,
+            )
+            continue
+
+        meeting = await meetings_controller.get_by_id(recording.meeting_id)
+        if not meeting:
+            continue
+
+        logger.info(
+            "Reprocessing failed LiveKit recording",
+            recording_id=recording.id,
+            meeting_id=meeting.id,
+        )
+
+        process_livekit_multitrack.delay(
+            room_name=meeting.room_name,
+            meeting_id=meeting.id,
+        )