Hermes Health.ts degrades a healthy, quiet gateway: staleness is judged from the event-driven gateway_state.json instead of the 30s loop heartbeat

[lgertel]

What the check does

LifeOS/install/LIFEOS/HERMES/Health.ts at main (ce046f2, release 7.40.4), lines 40 and 207–208:

const STATE_STALE_MS = 15 * 60 * 1000;
…
if (pidAlive && stateAgeMs !== null && stateAgeMs > STATE_STALE_MS) {
  problems.push(`gateway_state.json last written ${Math.round(stateAgeMs / 60000)}m ago`);
}

Any problem pushes status to degraded, so the menu bar row (modules/menubar.ts → hermesHealth()) and the /assistant card read gateway up · N/N channels · 1 issue for as long as the state file stays old.

What the gateway actually writes

gateway_state.json is event-driven. In NousResearch/hermes-agent (41447a6, 2026-08-25) it is written through gateway/status.py's n(...) writer, called from gateway/run.py on platform state changes and active-agent count changes — not on a timer. On a quiet channel it legitimately sits untouched for hours, and that is health, not staleness.

Liveness is proven elsewhere: gateway/shutdown_watchdog.py rewrites <HERMES_HOME>/state/gateway.heartbeat every 30 s:

47:DEFAULT_HEARTBEAT_INTERVAL_S = 30.0
58:_HEARTBEAT_RELATIVE = ("state", "gateway.heartbeat")

(the file holds {"pid", "updated_at", "monotonic", "start_time"}). Not to be confused with the LifeOS Heartbeat cron feature in HermesSidecar.md; this is the gateway main loop's own liveness stamp.

Observed

On a low-traffic install: pid alive, one channel connected, state/gateway.heartbeat written seconds before the check — and Health.ts reported 🟡 gateway up · 1/1 channels · 1 issue with gateway_state.json last written 881m ago, continuously for ~14 h. The gateway was fine the whole time; the instrument was reading the wrong file.

Suggested fix

When state/gateway.heartbeat exists, judge staleness from it (we used 5 min — ten missed beats — with the message loop heartbeat last written Nm ago — main loop hung), and fall back to the gateway_state.json age check only when the heartbeat file is absent (a gateway predating it). Add heartbeatUpdatedAt to HermesHealth; keep stateUpdatedAt as context. Down / flapping / fatal-platform detection unchanged.

Verified on a fork with three fixtures (stale heartbeat → degraded; dead pid → down; no heartbeat file → old fallback message) plus the live case above now reading 🟢 gateway up · 1/1 channels. Happy to open a PR if wanted.