Media-Server-Setup/README.md

vanguard-stream — SYS-1 (Xeon E-2136, 32 GB ECC, 2×4 TB SATA soft RAID)

STATUS.md = what runs, who does what, what's unfinished (kept current). CommunityDiscordMedia.md = design + phases.

Ingest from OBS, browse/scrub/download for editors, automatic normalization to one editing standard, dump button, 15-day archive pull to a home HDD. Everything is plain files under /srv/media.

1. Reinstall from the OVH manager

Server → Reinstall → Debian 13 (Trixie). Tick custom partitioning and use distribution kernel. Layout (RAID 1 across both HDDs):

mount	size	fs
/boot	1 GB	ext4
/	250 GB	ext4
swap	8 GB per disk	swap
/srv	rest (~3.4 TiB)	xfs

Add your SSH key. OVH's template creates a debian user with sudo; root SSH is disabled. Wait for the mail.

2. Install

scp -r vanguard-stream debian@SERVER:~/
ssh debian@SERVER
cd ~/vanguard-stream && sudo bash install.sh

install.sh installs docker, ffmpeg, rclone; creates /srv/media/{streams,clips,finals,trash,.markers} owned by uid 1000 (the debian user); copies the systemd units; adds debian to the docker group; starts the compose stack. It prints what you still have to edit.

The folder stays where it is: compose bind-mounts the configs and the dock page from it. After install, log out and back in once so the docker group applies, then plain docker compose works from ~/vanguard-stream without sudo.

3. Configure (domain, admin, streamers, editors)

mediamtx.yml, copyparty.conf, Caddyfile, dump-users.txt and .env are generated on the server by configure.py and gitignored. The *.example.* files in the repo are reference only. Never copy them to the server: doing so once replaced the live streamer passwords with CHANGEME and the next apply wiped the streamers. To update the server, copy scripts (configure.py, bin/, docker-compose.yml, …) and re-run ./configure.py.

cd ~/vanguard-stream
./configure.py --full      # first time: asks everything, then applies
./configure.py             # later: add/remove/reset streamers and editors, then apply
./configure.py --show      # print current users and the OBS settings to hand out

Stateless: it reads the current values out of Caddyfile, mediamtx.yml and .env, regenerates Caddyfile, mediamtx.yml, copyparty.conf and streamers.json (→ /etc/vanguard), creates streams/<name> and clips/<name> (with .stignore) for each streamer, restarts dumpd and the containers. Editors and admins are Authentik groups (editors, media-admins, admins), not local accounts. Removing a streamer revokes publishing only; their files stay on disk. Passwords may not contain : & ? # / \ quotes or spaces (they ride inside the stream key).

4. Firewall (host nftables; OVH edge firewall is optional on top)

Applied automatically at the end of install.sh. To re-apply by hand:

cd ~/vanguard-stream && sudo bash firewall/firewall.sh

Installs firewall/nftables.conf as /etc/nftables.conf, syntax-checks, applies, enables at boot. It uses its own table (inet vanguard) so applying or reloading never disturbs Docker's rules, and filters input only so container networking keeps working. Keep your SSH session open and test a new one before closing it.

port	proto	who
22	tcp	anywhere, rate limited (edit ssh_allow in /etc/nftables.conf to restrict to your IPs)
80, 443	tcp	world (Caddy)
443	udp	world (Caddy HTTP/3)
1935	tcp	world (RTMP ingest)
8890	udp	world (SRT ingest)
22000	tcp+udp	world (Syncthing)
3306	tcp	world, rate limited (MariaDB for the bot dev; restrict via db_allow in /etc/nftables.conf)
2222	tcp	world (Gitea ssh)
8189	tcp+udp	world (WebRTC media for the live portal)

Everything else is dropped and a trickle is logged (journalctl -k | grep vg-drop). Authentik (9000), Jellyfin (8096), copyparty, the Syncthing UI and dumpd bind to 127.0.0.1 only and are reached via Caddy or an ssh tunnel, so they are deliberately not opened. MariaDB runs on the host network (not a Docker published port) precisely so the firewall can govern it. Syncthing's admin UI is proxied at https://HOST/sync/ behind Syncthing's own GUI login (set it in Actions → Settings → GUI before relying on this). The ssh tunnel to 8384 still works as a fallback.

To change rules: edit /etc/nftables.conf, then sudo nft -f /etc/nftables.conf or sudo systemctl reload nftables. Never systemctl stop|restart nftables — Debian's unit flushes the whole ruleset on stop, which also wipes Docker's NAT rules and takes the containers offline until Docker is restarted.

5. Streamer setup (once)

Being a streamer = being in the Authentik streamers group (a Discord role via the roles→groups policy). There are no stream passwords or keys to hand out: the dock mints a 12-hour publish token for the logged-in Discord user each time Go Live is pressed and sets it as the OBS stream key itself.

OBS → Tools → WebSocket Server Settings → enable, set a password. OBS → Docks → Custom Browser Docks → https://media.example.com/dock/ → log in with Discord once. In the dock's Settings: ingest host (prefilled), OBS WebSocket password/port, optional title prefix/suffix ({date} → YYYY-MM-DD, e.g. prefix HelmetCam- or suffix -{date}). Per stream: type a title, click Go Live. The dock shows the final stream name (<you>/<Title>) as you type and enforces the naming rules the metadata relies on: letters, digits, _ . - only, no double underscore (it separates title and timestamp in filenames), no leading dot, max 64 chars. After a bad night: Dump tonight (dumps the night of the account you are logged in as; admins/media-admins may dump anyone's).

Manual OBS/other software: dock → Manual OBS setup shows the RTMP server, the stream key and the SRT URL with the token hidden (Show / Copy / Fresh token). Tokens expire after 12 h; a running stream is not cut when its token expires, but a reconnect needs a fresh one.

Browser streaming without OBS: the live portal's Go Live (Windows Chrome 141+ for window audio + GPU encode).

5b. Syncthing hardening (do once, in the server's Syncthing UI)

Syncthing receives straight into clips/<user> (Syncthing UI path /var/syncthing/clips/<user>), receive-only, with a .stignore (from syncthing/stignore) that pulls video files only. The gamer's folder structure is mirrored as-is. One copy per clip on the server: compliant files are left alone; a non-compliant file (HEVC/MKV/VFR) is converted to name.mp4 (name.h264.mp4 if the raw was already .mp4), the raw's name is appended to .stlocal (an #included ignore list, so Syncthing drops it without flagging or re-pulling), and the raw is deleted locally — it still exists on the gamer's PC. configure.py creates the folder + ignore file for each streamer.

In the UI (https://HOST/sync/):

• each clips folder → path /var/syncthing/clips/<user>, Folder Type: Receive Only, Edit → Advanced → Minimum Free Disk Space: 300 GB (so clips can never fill the mirror and stall the recordings).
• Actions → Settings → Connections → Incoming Rate Limit: 12500 KiB/s (~100 Mbps) so a gamer sharing a year of old clips cannot starve the ingest during stream hours.

5c. Authentik (Discord login) and Jellyfin

DNS: auth and watch as CNAMEs to the main host. Then run ./addCommunityMediaServer.sh — it asks for both hostnames, checks DNS, generates .env (Authentik secret + DB password; never commit it), adds the Caddy site blocks, starts the containers, waits for them and prints the steps below with your hostnames filled in. (./configure.py --full asks for the same hostnames if you prefer.) Then:

Authentik first run — https://auth.HOST/if/flow/initial-setup/ → set the akadmin password.

1. Discord developer portal → New Application → OAuth2: add redirect https://auth.HOST/source/oauth/callback/discord/ (last segment = the source slug below), copy Client ID + Secret.
2. Authentik → Directory → Federation & Social login → Create → Discord OAuth Source: slug discord, consumer key/secret = Client ID/Secret, user matching "link on unique identifier", scopes empty.
3. Flows & Stages → Stages → default-authentication-identification → Sources: select Discord, tick "Show sources' labels". Without this the Discord button never appears on the login page.
4. Flows & Stages → Stages → default-source-enrollment-write → User type: Internal. Otherwise Discord users are External and get "Interface can only be accessed by internal users".
5. Directory → Groups: create streamers, editors, admins. Add people after their first login.
6. (optional) Restrict login to members of your Discord server: add scope guilds to the source; Customization → Policies → Create Expression Policy from authentik/discord-guild-policy.py (fill in your server ID); bind it to both default-source-enrollment and default-source-authentication (flow → Policy / Group / User Bindings) and set each flow's Policy engine mode to all. Server ID: Discord Developer Mode → right-click server → Copy ID.

Jellyfin first run — https://watch.HOST/ → wizard: admin user, add libraries: /media/streams and /media/clips as Shows with only "Vanguard Metadata" as metadata downloader (Series/Season/Episode) and Screen Grabber + Vanguard Metadata as image fetchers; /media/finals as Home Videos with fetchers off. The Vanguard Metadata plugin (jellyfin-plugin-vanguard/, built with its package.sh) is installed by addCommunityMediaServer.sh / installJellyfinPlugin.sh from jellyfin-plugin-vanguard/dist/.

• Dashboard → Playback → Transcoding: disable hardware acceleration. Per-user "allow transcoding" is switched off automatically for everyone (SSO users included) by the Vanguard plugin's "enforce playback policy" task; turn that toggle off in the plugin config once the server has Quick Sync (SYS-3).
• Dashboard → Plugins → Repositories → add https://raw.githubusercontent.com/9p4/jellyfin-plugin-sso/manifest-release/manifest.json, then Catalog → SSO Authentication → install → restart Jellyfin.
• Authentik → Applications → Create with provider → OAuth2/OpenID: redirect https://watch.HOST/sso/OID/redirect/authentik, note client ID/secret.
• Jellyfin → Plugins → SSO-Auth → add OID provider "authentik": endpoint https://auth.HOST/application/o/<slug>/, client ID/secret, enable, roles claim groups, admin roles admins, enabled folders = the three libraries, Scheme Override https, Port Override empty (a value appends :443 and Authentik rejects the redirect). Save.
• The plugin adds no button. Test with https://watch.HOST/sso/OID/start/authentik, then Dashboard → General → Branding: Login disclaimer = jellyfin/branding-disclaimer.html (set the hostname), Custom CSS = jellyfin/branding.css. The page then shows only "Sign in with Discord"; a small "Admin login" link reveals the local username/password form. Jellyfin creates the account on first SSO login.

Phase 4 is the Vanguard Metadata Jellyfin plugin; Phase 5 puts copyparty, /sync and /dump behind Authentik forward-auth.

5d. MariaDB + Discord bot (for the bot dev)

mariadb (public on 3306, host network, root local-only, 32-char generated passwords) with a database/user for the bot, and discord-bot, a Python 3.12 container that runs bot/app/main.py and idles until that file exists. Credentials live in .env (generated by configure.py; bot token + guild ID are asked in --full or the [d] menu item; ./configure.py --show prints them).

The bot also serves a web UI on 127.0.0.1:8000, published as https://bot.HOST/ behind Authentik forward-auth: Caddy asks Authentik, then injects X-Authentik-Username/-Groups/-Email. Create the matching Proxy Provider in Authentik (forward auth, single application, external host https://bot.HOST), add it to the embedded outpost, and bind a group policy. Until then the site returns an error, never an open page.

The build brief for whoever writes the bot is bot/AGENT.md; bot/README.md is the ops cheat-sheet.

docker compose logs -f discord-bot                     # watch it
docker compose restart discord-bot                     # after code changes
docker compose build discord-bot && docker compose up -d discord-bot   # after requirements.txt changes

5e. Gitea + Actions runner

gitea (MariaDB-backed, git.HOST via Caddy, git-over-ssh on 2222, registration closed) and act_runner (jobs run as Docker containers; labels ubuntu-latest, ubuntu-22.04, dotnet, python). The gitea db/user are created by mariadb/init/10-gitea.sh on MariaDB's first start; for a MariaDB that already existed run ./createGiteaDb.sh once.

1. ./configure.py --full asks the Gitea hostname (CNAME git); apply starts both containers.
2. https://git.HOST/ → install page (db pre-filled) → create the admin account → Install.
3. ./registerGiteaRunner.sh → generates a registration token, stores it in .env, starts the runner.
4. Remotes: ssh://git@git.HOST:2222/<org>/<repo>.git or https://git.HOST/<org>/<repo>.git.

Runner caveat: it holds the Docker socket, so anyone who can push a workflow can run as root on the box. Keep repo write access to trusted people.

5f. Live portal (live.HOST)

React + dockview page (live/) behind Authentik forward-auth: every active stream is a dockable window, audio follows the focused window, per-window mute + volume. WebRTC (WHEP) with HLS fallback, served by MediaMTX (HLS 8888, WebRTC 8889, API 9997 on loopback; WebRTC media on 8189 udp/tcp, firewall opened). MediaMTX runs the -ffmpeg image so the optional "is live" Discord hook (LIVE_WEBHOOK_URL in .env, asked by configure.py --full) can post when a stream starts.

1. CNAME live → main host; ./configure.py --full asks the hostname (also --hosts … LIVE).
2. Authentik: Proxy Provider (forward auth, single application, external host https://live.HOST) → embedded outpost → group policy (e.g. everyone in the server, or streamers+editors+admins).
3. Dev box: ./deployLive.sh builds live/ and copies dist/ to /srv/live (Caddy serves it directly).

Go Live button (window audio: Windows + Chrome/Edge 141+; Linux/Mac get video only): broadcast a window from the browser (no OBS) as <user>/HelmetCam-<date> — a fresh publish token comes from /me/stream-token for the logged-in Discord user (member of streamers).

Bandwidth: no transcoding, every viewer gets the source bitrate (~15 Mbps) → ~30 concurrent viewers on the 500 Mbps line. For a crowd, restream to Twitch from OBS.

5g. Authentik forward-auth for the Caddy-gated hosts (main host, live., bot.)

Design: Authentik authenticates once for the whole domain; Caddy authorizes per host from the X-Authentik-Groups header the outpost returns; Discord roles fill the Authentik groups at login.

1. Applications → Outposts → edit "authentik Embedded Outpost" → config authentik_host: https://auth.HOST.
2. Applications → Create with provider → Application Vanguard sites; Proxy Provider, authorization flow implicit-consent, mode Forward auth (domain level), authentication URL https://auth.HOST, cookie domain = the parent domain (e.g. dumpstervigilantes.rip). Edit the Embedded Outpost → Applications → add it → Update. One login now covers every gated host.
3. Per-host group requirement = .env: BOT_REQUIRE_GROUP=bot-users (default), LIVE_REQUIRE_GROUP= (empty = any logged-in member), SYNC_REQUIRE_GROUP=media-admins (the /sync Syncthing UI on the main host). On the main host copyparty itself authorizes by group (editors, media-admins); the dump endpoint lets admins/media-admins dump anyone's night. Leave "Intercept header authentication" off: login is Discord only, no app passwords, no WebDAV. ./configure.py apply regenerates the Caddy blocks; a user outside the group gets a 403 from Caddy before the app sees the request.
4. Discord roles → groups: Directory → Groups: create bot-users (and keep streamers, editors, admins). Discord source → Scopes: guilds guilds.members.read. Customization → Policies → Expression Policy discord-roles-to-groups from authentik/discord-roles-to-groups.py (fill in the server ID and the role-id → group table). Bind it to default-source-enrollment and default-source-authentication (engine mode all, after discord-guild-member). From then on giving someone a Discord role gives them the matching access on next login; removing it removes it. copyparty additionally requires a header with a secret name that only Caddy adds (idp-h-key, COPYPARTY_IDP_KEY in .env), so identity headers cannot be forged even if copyparty were reached directly. Logout: copyparty's logout link, the live portal's "Sign out" and the dock's "sign out" go through /outpost.goauthentik.io/sign_out. Set the provider's Invalidation flow to default-provider-invalidation-flow so that also ends the Authentik login: every logout then forces a full Discord re-login on the next visit (and picks up group changes). Discord itself may approve silently if you are still logged into Discord in that browser.
5. Debug: https://HOST/whoami prints the username/groups headers exactly as the apps receive them. Group changes in Authentik apply on the next login — log out at https://auth.HOST/ and back in.
6. Test in a private window: https://live.HOST/ → Discord login → portal; https://bot.HOST/ → 403 unless you hold the bot role; https://live.HOST/me/stream-key shows your username.

Jellyfin and Gitea use OIDC providers instead, not this.

6. Editors

Log in with Discord at https://media.example.com/ (copyparty behind Caddy + Authentik). Access is by Authentik group: editors read streams and clips and write /finals; media-admins have full rights everywhere including /trash. Scrub any .mp4 in the browser, download what you want.

No WebDAV: login is Discord only (no passwords of any kind), and a mounted drive cannot do a browser login. Download from the browser: single files, or select files/folders and use copyparty's zip/tar download.

Available but not set up. WebDAV mounts can be enabled later without code changes if we decide we want them: in Authentik, edit the Vanguard sites proxy provider and tick Intercept header authentication; each editor then creates an app password under their Authentik user settings and mounts the same URL with their Authentik username + that app password. Caddy and copyparty already pass the identity through, so nothing else changes. It is off because it introduces a password-like credential alongside Discord login.

Every .mp4 an editor sees is the same kind of file (see "Editing standard"). Raw clipper output that needed converting is kept in a hidden .src/ folder next to it (media-admins can show dotfiles in copyparty).

7. Home archive box

Copy home/rclone-archive.sh to the home machine, set the vars, run from cron nightly. It connects over SFTP as debian (all media files are owned by uid 1000 so it can delete them). Pulls everything older than 15 days off the server, including the hidden .src/ raw files.

Publish authentication (no passwords)

MediaMTX asks dumpd (POST /mtx-auth on 127.0.0.1:8081, reachable only from Docker's bridge) for every connection. Reads/playback are allowed (viewers are gated by Caddy + Authentik). Publishing needs a token minted by GET /me/stream-token behind Authentik for a member of streamers/admins/media-admins; dumpd checks the HMAC signature (/etc/vanguard/stream-token.secret, generated into .env as STREAM_TOKEN_SECRET), the expiry, and that the path starts with that user's name. MediaMTX hot-reloads paths/recording settings but not the auth section: after changing auth, docker compose restart mediamtx between streams. configure.py reloads Caddy in place and restarts only copyparty.

Timecode for multicam editing (frame accurate)

Every published stream carries a QuickTime timecode track (tmcd, time of day in Eastern, at the file's frame rate) and a UTC creation_time, written by the sweep during the lossless remux, plus a sidecar <Title>__<ts>.json with the exact capture time (the Jellyfin plugin reads it). Resolve and Premiere read the track natively: drop clips from several streamers on a multicam timeline, sync by timecode.

Where the time comes from, best first:

1. Overscan band — a thin strip of extra rows below the picture carrying the server clock, encoded (black/white cells, format in bin/vgstrip.py) plus a small readable clock. The server decodes it (bin/vgstrip.py) at the start, middle and end of each recording → the timecode is frame-accurate at the start and the sidecar records drift over the file. Game image untouched. OBS: dock → Timecode band → "Add band + clock source" (canvas grows by W/160 rows, a locked browser source Vanguard Clock → /dock/clock/ sits in the band; only while not streaming). Go Live: the "Timecode band" checkbox (default on).
2. Server arrival time (no band) — microseconds in the recording filename. Within ~0.1–0.3 s per streamer (encoder + network delay); coarse but always available.

Viewing: the live portal crops the band; Jellyfin shows it as a thin strip at the bottom (VHS-style); editors export at the game resolution (e.g. 2560×1440) which drops the band — finals carry nothing. Cross-check when it matters: everyone is in the same voice chat, so audio-waveform sync is sample-accurate. Why continuous and not a start burst: a PC clock drifts vs the server (10 ppm ≈ 4 frames over 2 h).

Dropped frames / gaps: OBS keeps stamping frames from its clock, so after a drop the file has a hole and every frame that exists is still at the right time — the timecode stays correct, and players (Jellyfin, the portal) hold the last frame across the hole with audio in sync. By default the sweep does not re-encode a stream to fill holes (hours of CPU on this box); the sidecar lists them (gaps.count, total_missing_ms, first 50 positions). Premiere preserves timing on such files; in Resolve check the clip's frame-rate interpretation. Gap filling (available, off by default). The sweep can re-encode a stream whose holes total at least FILL_GAPS_MS milliseconds to constant frame rate with duplicated frames (raw kept in .src/, timecode carried over). It costs ~2× the stream length in CPU on the E-2136, so it is off until the box has a hardware encoder. Turn it on (threshold 1 s) / off, on the server:

sudo sed -i 's|^#* *Environment=FILL_GAPS_MS=.*|Environment=FILL_GAPS_MS=1000|' /etc/systemd/system/normalize-sweep.service && sudo systemctl daemon-reload
sudo sed -i 's|^Environment=FILL_GAPS_MS=.*|# Environment=FILL_GAPS_MS=1000|' /etc/systemd/system/normalize-sweep.service && sudo systemctl daemon-reload

Check: systemctl show normalize-sweep -p Environment. Takes effect on the next sweep run; already-published files are not revisited. Clips are always converted to CFR regardless.

Layout on the server

/srv/media/streams/<user>/<Title>/.live/<ts>.mp4             recording in progress (hidden from Jellyfin/copyparty)
/srv/media/streams/<user>/<Title>__<ts>.json                 capture time (UTC ms), timecode, source (strip|arrival), drift
/srv/media/streams/<user>/<Title>__<ts>.mp4                  finished stream, one flat MP4 per recording, directly in
                                                             the streamer's folder (several files inside one title
                                                             folder would be ONE Jellyfin episode with "versions")
/srv/media/clips/<user>/clip.mp4                             clip (synced as-is, or converted; raw removed)
/srv/media/clips/<user>/.stlocal                             Syncthing ignore list maintained by the sweep
/srv/media/clips/<user>/.src/clip.mkv                        raw source, only for clips that did NOT arrive via Syncthing
/srv/media/clips/<user>/clip.normalize-failed.log            only if normalization failed; delete to retry
/srv/media/finals/                                           editors upload here
/srv/media/trash/<night>/...                                 dumped, purged after 3 days
/srv/media/.markers/                                         dump markers
/srv/media/.syncthing/                                       Syncthing config

Editing standard (what normalize-sweep enforces)

MP4 · H.264 High, 8-bit 4:2:0 · constant frame rate · AAC audio (all tracks kept) · index at the front (instant seek in browser and NLE) · at most 2560x1440 and 60 fps (downscaled to fit, never upscaled; aspect kept).

input	action
already compliant	nothing
MediaMTX recording in <Title>/.live/ (fragmented MP4)	once idle 3 min: decode the timecode band (or use arrival time), lossless remux to <user>/<Title>__<ts>.mp4 with a tmcd track + creation_time, write the .json sidecar, remove .live and the empty title folder
in a Syncthing folder (.stfolder marker)	compliant / index-only: untouched; otherwise convert → name.mp4, add raw to .stlocal, delete raw (one copy)
HEVC / 10-bit / VFR / MKV / >1440p / >60fps / non-AAC	full-res re-encode (x264 veryfast CRF 18) → name.mp4; source kept in .src/
ffmpeg fails or duration mismatch	name.normalize-failed.log written, file skipped until that log is deleted

Runs every 5 min on files idle ≥3 min and not open. Software x264 on the E-2136 does ~0.5x realtime at 1440p60, so a 2 h clip batch takes ~4 h; streams only remux (minutes). All at idle priority.

Reboot behaviour

Everything comes back on its own, in this order:

1. mdadm assembles the mirror, systemd mounts /srv (local-fs). A degraded mirror still boots.
2. docker.service has RequiresMountsFor=/srv/media (drop-in installed by install.sh), so it never starts with the media tree missing. All containers are restart: unless-stopped.
3. dumpd.service starts; normalize-sweep.timer fires 2 min after boot and every 5 min after; trash-purge.timer is Persistent=true so a missed 05:00 run fires at boot.
4. A stream cut by a reboot is playable up to the last second (fMP4 parts are flushed every 1 s); OBS reconnects and MediaMTX opens a new file in the same title folder.
5. Half-written *.part files are overwritten by the next sweep. Dumps are idempotent.

Time: install.sh sets the server clock to America/New_York so journal, timers, file timestamps and the dump window all read in Eastern. Dump "night" = 17:00 → 02:00 Eastern (pinned in dumpd via DUMP_TZ, independent of the host zone). Between 02:00 and 17:00 the default night is the previous evening; after 17:00 it is tonight. The 05:00 trash purge is 05:00 Eastern.

Timers

unit	when	does
normalize-sweep.timer	every 5 min	remux/re-encode finished files to the editing standard
trash-purge.timer	daily 05:00 Eastern	deletes trash older than 3 days
dumpd.service	always	the dump endpoint on 127.0.0.1:8080, behind Caddy at /dump