# Albert on the server. One volume (`albert-data` -> /data) holds all runtime state:
# the subscription tokens, the kaeru memory vault, per-channel history, and the
# scheduler alarms — so restarts and image rebuilds never lose anything.
#
#   docker compose build                       # build the image (see README: add swap on 2 GB)
#   docker compose run --rm albert login       # one-time: sign in with the ChatGPT subscription
#   docker compose up -d                        # run the service
#   docker compose logs -f                      # live logs
#
# Voice (hearing / speaking, long recordings included) and image generation run natively
# on the ChatGPT subscription through Albert's own connectors, reading /data/auth.json as
# the service user — nothing to set up beyond `login`. No sandboxed script reads the
# token, so it stays 0600.

services:
  albert:
    build:
      context: .
      dockerfile: docker/Dockerfile
    image: albert:latest
    container_name: albert
    restart: unless-stopped
    env_file: contrib/deploy/.env
    # The container runs as the unprivileged `albert` user (see docker/Dockerfile).
    # forkd drops skill scripts to `albert-scripts`, which from a non-root process needs
    # these two capabilities — the same ones the systemd unit grants as AmbientCapabilities.
    cap_add:
      - SETUID
      - SETGID
    volumes:
      - albert-data:/data
      # Connector manifests (owner_chat, which calendars) — mounted so you can edit
      # them without rebuilding the image. Read-write so the telegram ACL can persist.
      - ./config:/opt/albert/config
      # Persona + operating instructions, mounted so they can be edited live —
      # they're hot-reloaded by mtime (no rebuild, no restart).
      - ./soul.md:/opt/albert/soul.md:ro
      - ./system.md:/opt/albert/system.md:ro
      # Live config — an OPERATOR-OWNED, git-ignored file, NOT a tracked one.
      # Seed it once before the first run:  cp contrib/deploy/albert.toml ./albert.local.toml
      # Mounted read-write: the self-config skill edits this exact file, so its changes
      # must land on a real persistent file, not the image's ephemeral layer. The tracked
      # contrib/deploy/albert.toml stays a pristine default. (Create it first — Docker
      # would otherwise make an empty directory in its place.)
      - ./albert.local.toml:/opt/albert/albert.toml
    # `albert login` needs an interactive terminal for the paste fallback.
    stdin_open: true
    tty: true

volumes:
  albert-data: