version: "3.8"

services:
  fade-out:
    build:
      context: .
      args:
        # Set by scripts/deploy.sh so the running image can be identified.
        BUILD_COMMIT: ${BUILD_COMMIT:-unknown}
        BUILD_TIME: ${BUILD_TIME:-unknown}
        BUILD_SOURCE: ${BUILD_SOURCE:-local}
    # Or use pre-built image:
    # image: ghcr.io/rancur/fade-out:latest
    container_name: fade-out
    restart: unless-stopped
    ports:
      - "8500:8000"
    volumes:
      # Host paths come from .env so this file works unmodified on any machine.
      # Defaults drop everything under ./media/ next to the compose file; point
      # them at your NAS shares (e.g. /volume1/music/mixes) by setting the
      # matching variables in .env.
      #
      # Audio + video watch folders are READ-WRITE so the optional "Rename
      # source files to match titles" setting (Settings -> Advanced, OFF by
      # default) can rename finished recordings in place. Change them to :ro if
      # you never want fade-out to touch the source files — with the setting on
      # and a :ro mount every rename is simply skipped and logged, and nothing
      # else about the pipeline changes.
      - ${WATCH_AUDIO_DIR:-./media/audio}:/watch/audio:rw
      - ${WATCH_VIDEO_DIR:-./media/video}:/watch/video:rw
      - ${WATCH_CUE_DIR:-./media/cue-sheets}:/watch/djctl-cue:ro
      # Vertical clips for the Shorts auto-uploader (READ-ONLY — the pipeline
      # only ever reads the source recordings)
      - ${WATCH_SHORTS_DIR:-./media/shorts}:/watch/shorts:ro
      # Output folders
      - ${OUTPUT_THUMBNAILS_DIR:-./media/thumbnails}:/output/thumbnails
      - ${OUTPUT_COVER_ART_DIR:-./media/cover-art}:/output/cover-art
      # Persistent data (SQLite DB, state, tokens)
      - ./data:/data
    environment:
      - TZ=${TZ:-UTC}
      - OPENAI_API_KEY=${OPENAI_API_KEY}
      - FAL_API_KEY=${FAL_API_KEY}
      - SOUNDCLOUD_CLIENT_ID=${SOUNDCLOUD_CLIENT_ID}
      - SOUNDCLOUD_CLIENT_SECRET=${SOUNDCLOUD_CLIENT_SECRET}
      - SOUNDCLOUD_ACCESS_TOKEN=${SOUNDCLOUD_ACCESS_TOKEN}
      - SOUNDCLOUD_REFRESH_TOKEN=${SOUNDCLOUD_REFRESH_TOKEN}
      - SOUNDCLOUD_EMAIL=${SOUNDCLOUD_EMAIL}
      - SOUNDCLOUD_PASSWORD=${SOUNDCLOUD_PASSWORD}
      - YOUTUBE_API_KEY=${YOUTUBE_API_KEY}
      - YOUTUBE_CLIENT_ID=${YOUTUBE_CLIENT_ID}
      - YOUTUBE_CLIENT_SECRET=${YOUTUBE_CLIENT_SECRET}
      - YOUTUBE_REFRESH_TOKEN=${YOUTUBE_REFRESH_TOKEN}
      - PUBLIC_URL=${PUBLIC_URL}
      - NOTIFICATION_DISCORD_WEBHOOK_URL=${NOTIFICATION_DISCORD_WEBHOOK_URL}
      # Read-only GitHub credential for the deployment freshness check. The
      # setting existed but was never passed through to the container, so on a
      # private repo the check could only ever answer "cannot determine".
      # Optional: unset, the check stays honestly unknown rather than green.
      - GITHUB_TOKEN=${GITHUB_TOKEN:-}
    # Run unprivileged if your bind mounts allow it. Left commented because the
    # container must be able to read the watch folders and write the output
    # folders, and the right UID/GID depends on your NAS share permissions.
    # See SECURITY.md.
    # user: "1000:1000"
    # Analysis of a long DJ set (beat/energy/track detection over a multi-hour
    # WAV) is the memory high-water mark. 4g was not enough: on 2026-08-21 the
    # container was OOM-killed at 4.19 GB mid-analysis, the boot sweep then
    # marked the mix failed, and the set silently never published. Keep real
    # headroom here — an OOM kill in the middle of the pipeline is data loss,
    # not a slowdown.
    mem_limit: 12g
    logging:
      driver: json-file
      options:
        max-size: "50m"
        max-file: "3"