Sermo is a portable, safety-first service supervisor for Linux hosts. It sits above systemd and OpenRC, validates a service before it ever acts on it, understands the service's real processes, and applies guarded remediation rules — it never restarts blindly and never kills the wrong process.

Where an init system answers "is the unit active?", Sermo answers the harder operational questions: is the service actually healthy, is it safe to touch right now, and if it isn't healthy, what is the safe thing to do? It monitors, diagnoses, remediates under strict invariants, keeps availability (SLA) history, watches host-level resources, sends notifications, and serves a live web dashboard — all from a single daemon.

A plain "restart on failure" supervisor is dangerous on a real host: it restarts during a backup, kills a process that happens to share a binary name, or acts on a service whose config is broken and makes the outage worse. Sermo is built around the opposite principle — prove it is safe, then act:

• Guarded, never blind. Every start/stop/restart/reload/resume, plus the manual stale-pidfile repair path, runs through one operation engine that checks preflight, guards and runtime locks first. A blocked action does not run.
• Knows the real processes. It discovers a service's actual PIDs from /proc, so health, residual detection and (rare, opt-in) kills are based on the resolved executable and UID — never on a process name.
• Validates before it touches anything. Broken config, a failing required preflight, or an active named lock blocks remediation instead of amplifying an outage.
• Portable. The same configuration and behaviour run over systemd and OpenRC; the init backend is auto-detected.
• Honest availability. SLA counts only cycles it actually observed — time before any evidence could exist is a gap, not counted as downtime.

Monitoring & health

• A fleet of independent workers — one per service — each running its own checks, evaluating rules and driving remediation on its own schedule; a panic in one worker never takes down the daemon.
• A broad check catalog: service state/version/config, TCP ports, HTTP(S) and WebSocket endpoints, TLS certificate expiry, database connectivity and queries (MySQL/MariaDB, MongoDB, InfluxDB, SQLite integrity, arbitrary SQL), egress interface, default route, firewall rules, clock drift, file/dir size growth, disk throughput (hdparm), hardware sensors, and process/count metrics.
• Check interdependencies (requires / skip_when_changed) so a probe is only run when its prerequisites hold.
• Host watches — monitor resources that aren't services (mounts, RAID arrays, network uplinks, certificates, …), firing hook commands and/or notifications.

Safe remediation

• A single operation engine shared by the CLI and the daemon: operation lock → named runtime locks → required preflight → guards → staged residual-aware stop/start or explicit atomic init restart → verify + postflight.
• Named runtime locks to fence maintenance windows (backups, migrations): sermoctl lock … -- COMMAND holds a TTL'd lock for the duration of a command.
• Guards, windows and remediation policy to express when and how often an action may run, with escalation.
• Hard safety invariants that YAML cannot disable (see below).

Availability & history

• Per-service SLA over rolling windows (hour → year), stored as one archive per graph window so a year of history stays bounded without losing an incident.
• An event/activity log with retention and compaction of the state store.

Operate

• A focused operator CLI (sermoctl) for status, safe lifecycle actions, config validation, locks, processes, preflight, inventory, SLA and events.
• Notifications to email, Slack, Teams and webhook sinks (ntfy/Telegram/ Gotify) with a templated default message.
• An optional interactive Telegram report bot (read-only): ask it /status, /services, /sla and it replies with live reports — long polling only, no inbound port, answering allow-listed chats.
• A daemon-wide panic switch to pause all automatic remediation instantly.
• Guided wizards for common setups (service, docker, vm, mount, volume, net, uplink).

Web dashboard (optional)

• A live, self-contained dashboard: per-service checks, SLA history, latency graphs, an event feed, and the full inventory (services / apps / libraries) and host watches.
• Push-driven via server-sent events with a polling fallback, and built to WCAG 2.2 AA accessibility.
• Loopback HTTP with optional auth — expose it only behind a TLS reverse proxy.

A single daemon (sermod) loads the configuration and the packaged catalog, resolves them into a service tree, and builds a fleet: one Worker per service and one Watch per host resource or app. A scheduler runs them in a loop. The CLI and the web UI talk to the daemon over HTTP and signals. Every action on a service — manual or automatic — goes through operation.Engine, which coordinates locks, preflight, guards and the init backend, so the CLI and the daemon can never diverge in how they act.

clients ── sermoctl (CLI) ─┐                    ┌── operation.Engine ── systemd / OpenRC
           browser (Web UI) ┤                    ├── named locks (oplock + scanner)
                            │   sermod (daemon)  │
      signals ── SIGHUP ────┼── Monitor ─ Scheduler ─ Fleet ┤── state store (SLA · events · metrics)
                SIGTERM ────┘        │   (Worker per service │── notifiers (email/slack/teams/webhook)
                                     │    Watch per resource)│
      config + packaged catalog ────┘                       └── web.Server (dashboard + /api)

See docs/architecture.md for the faithful, code-anchored diagrams (operation pipeline, lock states, monitoring cycle).

• sermoctl — the operator CLI: status, safe start/stop/restart/reload/ resume/repair, config validate, locks, processes, preflight, per-service availability/SLA, inventory and events. Read-only commands do not need root.
• sermod — the daemon: one independent worker per service runs checks, evaluates rules and drives remediation through the same safe operation engine sermoctl uses. It also runs host watches that fire hook commands and/or notifications (email, Slack, Teams, webhooks), and can serve the web dashboard (set web.port, recommended 9797) — loopback HTTP with optional auth; expose it only behind a TLS reverse proxy (how).

These cannot be turned off in YAML — validation rejects any security: toggle that tries. In full in docs/safety.md:

1. No action on a failed required preflight — blocked with preflight_failed.
2. No action a guard blocks — guards are evaluated before remediation.
3. Active named runtime locks always block service actions — checked automatically, no rule needed.
4. Never SIGKILL by default — force_kill is false unless explicitly enabled.
5. Never kill by process name — a kill requires an exact match on the resolved /proc/<pid>/exe path and the real UID against an explicit kill_only_if selector; anything it cannot positively identify is reported, not killed.
6. Never send terminating signals to PID 1 or kernel threads — blocked centrally, not configurable.
7. force_kill: true requires kill_only_if with non-empty users and exe_any selectors.

make build      # produces bin/sermoctl and bin/sermod
make test       # run the test suite

Requires Go 1.26.7+. Runtime dependencies: systemctl or rc-service on the host.

sermod runs as root. It manages services owned by different users and accesses privileged areas (service control, signalling other users' processes, cross-user /proc inspection including per-process IO, raw ICMP sockets), so the packaged units run it as root; it warns at startup if it is not. The config is therefore trusted, root-owned input — command checks and hooks run as root (never via a shell), so keep /etc/sermo root-only and put secrets in the environment (${env:NAME}). See safety. Read-only sermoctl commands (status, config, etc.) do not need root.

make install honors the standard GNU directory variables and DESTDIR staging, and installs the binaries, the full catalog (keeping the services/apps/libs/patterns layout), a sample sermo.yml, the default notification template, the tmpfiles.d config, and both the systemd unit and the OpenRC init script (with their binary/config paths rewritten to match):

sudo make install PREFIX=/usr                 # /usr/bin, /usr/sbin or merged-/usr /usr/bin, /etc/sermo, ...
make install DESTDIR=/tmp/stage PREFIX=/usr    # stage for packaging

Key variables (override on the command line): DESTDIR, PREFIX/prefix, bindir, sbindir, datadir, sysconfdir, TMPFILESDIR, SYSTEMD_UNITDIR, OPENRC_INITDIR. Granular targets are available too: install-bin, install-catalog, install-config, install-templates, install-tmpfiles, install-systemd, install-openrc (and uninstall). An existing sermo.yml is never overwritten. make install does not create /var/lib/sermo; the installed tmpfiles.d config owns that directory creation.

On merged-/usr hosts where /usr/sbin is a symlink to /usr/bin, the default sbindir collapses to $(bindir) so DESTDIR packages do not materialize a real usr/sbin directory and replace the host symlink when extracted. Pass an explicit sbindir=... only when the target really has a distinct sbin directory.

Do not deploy a DESTDIR tree by extracting a tar archive directly into / with preserved directory metadata. A staged tree contains directory entries such as ./, etc/ and usr/; plain tar -xpf can apply those modes to existing system directories. Use the package manager, copy the installed files directly, or extract ad-hoc test archives with:

sudo tar --no-overwrite-dir -C / -xpf sermo-stage.tar

# Inspect a unit (no config needed)
sermoctl backend
sermoctl status nginx
sermoctl is-active nginx

# List catalog inventory, not configured runtime targets
sermoctl services      # packaged catalog service profiles (nginx, mariadb, ...)
sermoctl services all  # include profiles not installed on this host
sermoctl services --notify ops-email  # email a services inventory report
sermoctl apps          # tools/runtimes (only installed)
sermoctl apps all      # include not-installed
sermoctl libs          # shared libraries (restart triggers)

# Validate configuration
sermoctl config validate

# Operate a configured service through the safe engine
sermoctl restart apache-main

# Pause / resume monitoring of a service (e.g. for maintenance)
sermoctl unmonitor apache-main   # daemon stops checking it
sermoctl monitor apache-main     # resume
sermoctl daemon reload           # ask sermod to re-read its config

# Fence a maintenance window with a named runtime lock
sermoctl lock apache-main --reason backup --ttl 1h -- /usr/local/bin/backup.sh

# Availability (SLA) per service over rolling windows (hour..year)
sermoctl sla                     # all services
sermoctl sla apache-main         # one service
sermoctl sla --series apache-main --since 168h  # per-minute series (graph data)

# Run the daemon
sermod run --config /etc/sermo/sermo.yml

Packaged definitions live under catalog/, sample configs under examples/, packaging units under packaging/. The on-host file layout is in configuration → layout. Daemon flags (--verbose) are in CLI → sermod daemon flags.

• Configuration — global config, catalog services, services, merge and variables; docs/sermo-all.yml is the complete annotated example.
• Rules — checks, conditions, windows, guards, remediation policy.
• Services — writing and overriding services.
• CLI — commands, flags and exit codes.
• Safety — the invariants that cannot be disabled: no unguarded actions, no SIGKILL by default, never kill by name (exact resolved-exe + UID match only).
• Web dashboard — what every panel, badge and graph on the dashboard means.
• Architecture — end-to-end diagrams of the daemon, the operation pipeline, lock states and the monitoring cycle.