obs

The OBS container that streams the A Dana Life dashcam slow-TV broadcast. It runs a headless OBS Studio that composites the VLC-served dashcam video plus the onscreen overlays and pushes the result to Twitch / YouTube.

This image was extracted from the tripbot monorepo (with full git history) and published as ghcr.io/adanalife/obs. The Go side of the system — the chat bot, the vlc-server, the onscreens-server, and the OBS watchdog/websocket client (pkg/obs in tripbot) — stays in tripbot. This repo owns only the OBS image and its deployment.

What's in the image

A headless OBS Studio, driven entirely by baked-in config and supervised helper processes:

• OBS Studio — PPA build on amd64 (CEF-bundled, for browser_source); compiled from source against the aarch64 CEF tarball on arm64.
• Display stack — sway (headless Wayland compositor) + wayvnc + noVNC, so OBS's OpenGL composite hits the host iGPU (VAAPI encode) and the desktop is reachable in a browser.
• supervisor manages OBS, sway, wayvnc, noVNC, the hourly browser-source refresh (a workaround for CEF's per-frame memory leak), and a small Flask obs-server exposing /health/ready, /version, and POST /admin/shutdown.
• Scene/profile templates (config/) rendered at startup from env vars.
• Background audio beds — one Background Audio scene source, playing one of three beds: the SomaFM stream, a license-clean car-interior drone rendered at build time (carhum/), or an album from the mounted music share. Selected at startup by OBS_BACKGROUND_AUDIO and switchable live from the admin console. The source is audio-only in intent but OBS renders any video its file carries (album tracks embed cover art), so its scene item is parked off-canvas — hiding it would stop the audio too.

Layout

Path	What
Dockerfile	amd64 image (OBS from the obsproject PPA)
Dockerfile.arm64	arm64 image (OBS from source, via the CEF base below)
Dockerfile.arm64-base	the arm64 CEF compile base → ghcr.io/adanalife/obs-cef-base
config/	OBS scene + profile templates (*.tmpl rendered by entrypoint.sh)
script/	in-image startup scripts (sway, wayvnc, noVNC, obs-server)
scripts/	repo tooling, not baked into the image (check-changelog-fragment.sh, the pre-push changelog guard) — note the near-identical name to script/ above
supervisor/	per-process supervisord configs
bin/	obs-browser-refresh, obs-media-restart (host/in-image Python helpers)
carhum/	car-hum FLAC generator (build-time only)
assets/	Twitch overlay PNGs baked into the image
desktop-profiles/	reference OBS Studio profiles for local desktop (macOS/Windows)
cdk8s/	the Kubernetes deployment (synthesized into cdk8s/dist/)

Build & smoke-test locally

Needs Docker running.

task image:build          # amd64 image → obs:dev
task image:smoke          # build, run (no STREAM_KEY), wait for the healthcheck
task image:build:arm64    # arm64 image (pulls ghcr.io/adanalife/obs-cef-base)
task carhum:render        # render the FLAC variants into carhum/out/ (numpy/scipy/ffmpeg)

Configuration

The deployment sets these at runtime (see cdk8s/); the image runs headless without them (the healthcheck only needs OBS + the Wayland session up):

Env var	Purpose
STREAM_KEY	Twitch/YouTube ingest key (per env + platform)
STREAM_PLATFORM	twitch (default), youtube, facebook, tiktok, or instagram — selects the ingest service, the default background-audio bed, and the canvas orientation (tiktok/instagram are portrait); see entrypoint.sh
OBS_BACKGROUND_AUDIO	starting background-audio bed: somafm, carhum, or album. Unset → somafm on twitch, album on tiktok, carhum elsewhere. Only the starting bed — tripbot rewrites the source live
OBS_VERTICAL	force portrait output (true/false); overrides the per-platform default so any platform can be tested vertical. Portrait renders a generated Vertical scene — `Main$ \text{rotated} 90° \text{CW} \text{into} \text{a} 1080 \times 1920 \text{canvas}
$OBS_WEBSOCKET_PASSWD`	obs-websocket auth (tripbot's watchdog connects with it)
OBS_QUALITY_PRESET	encoder quality preset (low on stage)
OBS_STREAM_ENCODER	encoder selection (e.g. VAAPI vs x264)
DASHCAM_RTSP_URL	the VLC-served dashcam RTSP source
VLC_URL_BASE / ONSCREENS_URL_BASE	the VLC + onscreens HTTP bases for browser sources

The tripbot contract (the one coupling that survives the split)

tripbot drives the Background Audio source over the OBS WebSocket — the audio watchdog swaps it to a local bed when SomaFM drops, !carsound picks a drone, and the console's bed selector switches between all three. Three hand-maintained contracts hold that together; change one side → update the other:

Contract	Here	In tripbot
Source name	Background Audio in config/Tripbot.json.tmpl	BackgroundAudioInputName
Car-hum variants	carhum/render-variants.sh + the Dockerfiles' COPY	the carSound list in pkg/chatbot/carsound.go
Bed names + paths	set_background_audio in entrypoint.sh	the bed registry in pkg/obs/beds

(Same shape as the eventbus contracts shared with tripbot-console.)

Releasing

Trunk-based main + release-please, with towncrier changelog fragments:

1. Feature PRs target main (squash-merge, conventional title); each adds a fragment (task changelog:add TYPE=<type> — no PR number needed, CI fills it in on push) or carries the skip-changelog label.
2. dev-image.yml floats ghcr.io/adanalife/obs:main (amd64) on every main push — what stage deploys.
3. release-please.yml maintains a standing release PR that bumps the version + the prod pin (cdk8s/versions.yaml) from the conventional commits, collates the changelog.d/ fragments into CHANGELOG.md, and re-synths cdk8s/dist/ on the PR branch.
4. To ship: squash-merge the release PR. That tags vX.Y.Z, creates the GitHub Release, and dispatches release.yml to build the multi-arch image to GHCR. No manual version/changelog steps — the version follows from the commit types (feat: → minor, fix: → patch, feat!:/BREAKING CHANGE → major).

The arm64 CEF base (Dockerfile.arm64-base) is rebuilt only when it changes, by obs-base.yml (a ~90-min compile) — see that workflow's header.