███████╗████████╗ █████╗ ███████╗███████╗
██╔════╝╚══██╔══╝██╔══██╗██╔════╝██╔════╝
███████╗   ██║   ███████║█████╗  █████╗
╚════██║   ██║   ██╔══██║██╔══╝  ██╔══╝
███████║   ██║   ██║  ██║██║     ██║
╚══════╝   ╚═╝   ╚═╝  ╚═╝╚═╝     ╚═╝
   P R O D U C T I O N   S U P P O R T

Agentic radio staff for a creator who never stops.

kannaka-staff is the constellation's watcher. Kannaka writes, broadcasts, dreams, remembers — the staff keeps the lights on. Health probes across every node, alert routing for production incidents, deploy assistance for the operator-of-one running the whole show.

[]() []()

What It Watches

┌───────────────────────────────────────────────────────────┐
│                  constellation health                     │
├─────────────────┬─────────────────────┬───────────────────┤
│  systemd        │  HTTP probes        │  NATS subjects    │
│  · kannaka-*    │  · /api/state       │  · RADIO.alert.*  │
│  · nats         │  · /api/swarm       │  · QUEEN.event.*  │
│  · icecast      │  · /api/hrm/status  │  · KANNAKA.*      │
├─────────────────┼─────────────────────┼───────────────────┤
│  Disk           │  Process            │  Substrate        │
│  · root usage   │  · pgrep kannaka    │  · phi publish?   │
│  · prune-cron   │  · zombie detect    │  · 65s cadence?   │
│  · snapshot dir │  · binary inode     │  · clusters=96?   │
└─────────────────┴─────────────────────┴───────────────────┘

When something silently fails — a service deactivating, the disk filling, a zombie binary holding a stale inode — the staff publishes an alert to the bus and (optionally) opens an issue, posts to Bluesky, or wakes a maintainer.

Run

git clone https://github.com/NickFlach/kannaka-staff.git
cd kannaka-staff
npm install

# Defaults to port 8889; --watch reruns on file change
node src/index.js --port 8889

Designed to live on the same box as kannaka-radio + kannaka-memory so the watcher can see systemd directly. The Oracle production box runs it as kannaka-staff.service.

Operate (ADR-004 — truthful operations)

The staff holds itself to the standard it holds the constellation to: every surface can say "no", and says it when the "no" is knowable.

• Action results are honest. /action/* returns ok:true only for completed work (HTTP 200). Queued/launched work returns { accepted: true, jobId } (HTTP 202) — outcome lands in role history, alerts.jsonl, and the bus. Preflight failures refuse immediately (500), nothing queued.
• Per-action kill switches. Every write-action honors STAFF_ACTION_<NAME>=0 (e.g. STAFF_ACTION_RESTART_RADIO=0) and refuses visibly when disabled. The authoritative action registry is the governance table in ADR-004 — code, registry, and table are sync-tested.
• Audit provenance. Every write-action leaves an alerts.jsonl row with an actor (operator:local, operator:hmac, or staff:<loop>) — nobody guesses who restarted the radio at 03:11.
• GET /api/health — per-subsystem tick freshness (probe loop, all nine roles, the heartbeat). Anything whose last tick is older than 2× its interval flips the endpoint to 503: process-alive-but-loop-wedged, visible to one curl.
• KANNAKA.staff.heartbeat — published to the swarm NATS every probe tick so an outside watcher can see the staff die. STAFF_HEARTBEAT=0 disables.

Constellation

License

Space Child License v1.0.