[Unit]
Description=Sentinel agent — AI/GPU server monitoring collector
Documentation=file:///usr/share/doc/sentinel-agent/README.md
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
User=sentinel
Group=sentinel
ExecStart=/usr/local/bin/sentinel-agent run --config /etc/sentinel/agent.yaml
Restart=on-failure
RestartSec=5s
# Unbuffered stdout so journald sees lines immediately.
Environment=PYTHONUNBUFFERED=1

# --- hardening ---
# v2 update commands deliberately DROP NoNewPrivileges (blocks sudo's setuid
# transition) and ProtectSystem=strict (its read-only /usr namespace is
# inherited by sudo'd children, so apt could never install). Compensating
# controls: the root-owned sentinel-update-helper's fixed case arms + the 4
# exact-argv sudoers lines (project/v2/protocol.md §7).
ProtectHome=true
PrivateTmp=true
# NVML needs the GPU character devices; keep device access on.
PrivateDevices=false
# DeviceAllow globs do not cross "/" (and on some systemd versions do not match
# at all), so "/dev/nvidia*" silently blocked /dev/nvidia-caps/* and NVML init
# failed with "Unknown Error". List literal nodes instead and adjust to the
# host's GPUs:  ls /dev/nvidia* /dev/nvidia-caps/*
DeviceAllow=/dev/nvidiactl
DeviceAllow=/dev/nvidia-modeset
DeviceAllow=/dev/nvidia-uvm
DeviceAllow=/dev/nvidia-uvm-tools
DeviceAllow=/dev/nvidia0
DeviceAllow=/dev/nvidia1
DeviceAllow=/dev/nvidia-caps/nvidia-cap1
DeviceAllow=/dev/nvidia-caps/nvidia-cap2
# The buffer lives under /var/lib/sentinel. ReadWritePaths stays listed so the
# unit keeps working unchanged if ProtectSystem is ever re-added.
ReadWritePaths=/var/lib/sentinel
StandardOutput=journal
StandardError=journal

[Install]
WantedBy=multi-user.target