Ainb Fleet:daemons

ainb fleet:daemons — daemon health view

The runtime-health answer for the fleet's four background daemons. Reads the on-disk liveness signals each daemon already maintains and renders one typed row per daemon. The same aggregator (fleet::daemons::collect) backs the TUI Daemons screen, so the CLI and the TUI can never drift.

This is the question ainb fleet bridge status could never answer: that verb reports only install state. daemons reports whether the process is actually running and serving right now — a crashed-but-installed bridge shows stopped with a stale-heartbeat reason here.

The four daemons

Rows are emitted in this exact order (bridge, notifyd, ATC, then the fleet daemon when it has a row), even when every daemon is stopped — so "everything stopped" is a legible table, never an empty one.

The fleet daemon is a conditional row. ATC is the fleet's supervisor and the standalone daemon is what it replaced, so a permanently-stopped row for it presented the two as a choice — the competing control path the supervisor exists to remove, still on the screen you read to see who is running. It is therefore listed only when it is NOT stopped.

That is deliberately not "always hidden". A daemon that is up got there through --force-race, which means two controllers really are auto-continuing the same panes and ATC's per-session retry cap cannot hold. That row appears, reports degraded, and its reason opens with RACING ATC. Hiding it would make the one state you most need to see the only one the screen does not show.

The row is only a display decision: ainb daemon fleet-daemon start|stop|restart keeps working whether or not a row is drawn.

Run

ainb fleet daemons                       # fixed-width text table (default)
ainb --format json fleet daemons         # typed rows for scripting

--format is a global flag — it goes before fleet. There are no other flags or subcommands; daemons is a pure read.

Text columns

DAEMON         STATE      PID      UPTIME     LAST ACTIVITY  ERRORS  HEALTH

• STATE — ● running · ○ stopped · ? unknown.
• PID — the OS pid when known (heartbeat / PID file); - for ATC (timer-driven).
• UPTIME — compact since-start (5s, 12m, 3h, 2d); - when unknown.
• LAST ACTIVITY — relative time of the last real activity (5s ago, just now).
• ERRORS — error count observed this run.
• HEALTH — for a running+connected daemon, the connection channel + reason (e.g. Telegram (@bot) — running + connected); otherwise the reason string, which is the load-bearing field for telling a clean stop from a crash.

States and what they mean (JSON state + reason)

JSON fields (per row)

The pid-identity recycle guard

A daemon's heartbeat records both its pid and its process started_at. Bare liveness (kill(pid, 0)) is not trusted: when a daemon dies and the OS hands its pid to a different, unrelated process, that pid is alive but is no longer our daemon. The probe cross-checks process identity — the heartbeat's started_at against the live process's real OS start time — so a recycled pid classifies as stopped (reason: pid … recycled (different process, crashed)), never a false running. A dead pid is caught immediately regardless of how recent the heartbeat looks; a live-but-wedged pid whose beat went quiet past the 90s staleness window reports stopped with a wedged? reason.

For notifyd specifically, "connected" requires an actually-bound Unix socket (a non-blocking connect() that a listener accepts) — not merely a notify.sock file on disk, which a crashed daemon can leave behind. The DB is reported as "file present", honestly not "reachable", since exists() does not prove the sqlite file opens and accepts writes.

Composition patterns

# Any daemon not running?
ainb --format json fleet daemons | jq '[.[] | select(.state != "running")]'

# Is the bridge connected to a chat channel right now?
ainb --format json fleet daemons | jq '.[] | select(.kind=="bridge") | {state, connected, channel, reason}'

# Surface any crash reasons in one line each
ainb --format json fleet daemons | jq -r '.[] | select(.reason | test("crash|stale|wedged")) | "\(.kind): \(.reason)"'

Caveats

• Read-only. daemons never starts, stops, or restarts anything — it only reports. To manage a daemon, use its own verb (ainb fleet bridge install/uninstall, ainb fleet atc setup/teardown, ainb fleet daemon).
• notifyd ignores $AINB_HOME. It always uses ~/.agents-in-a-box; the probe mirrors that so it reads the same files notifyd writes (a set $AINB_HOME is only honoured for an isolated test run).
• ATC surfaces as one row. v1 reports the most-recently-beating provisioned instance; the instance name is in the channel label. Heartbeat-disabled instances are never counted as running, even with a recent leftover beat.
• The ATC row's channel names the supervisor mode — tower · FULL(claude) · every 15m or tower · LITE · no LLM. A lite instance has no scheduler and no brain session, so it is probed by its scan loop instead; one provisioned in lite with no scanner running reads degraded (the fleet has a designated owner and that owner is absent), not stopped.
• The mode is switchable from the row. Enter on the ATC row offers the mode it is not in, and the screen prints what each mode does inline. It is the same transition ainb fleet atc mode <name> --set <mode> performs — the screen shells that verb rather than reimplementing the stop-then-start ordering.