Both sides now actively detect half-dead WS connections instead of
waiting for kernel TCP keepalive (~2hrs default on Linux). Bug user
reported: "claudemesh peer list" shows zero peers despite running
sessions, because NAT/CGNAT silently dropped the WS flow but neither
side noticed.
Broker (apps/broker/src/index.ts):
- Add lastPongAt to PeerConn, populate at connections.set sites,
bump in ws.on("pong").
- 30s ping loop now also terminates conns whose pong is >75s stale.
ws.terminate() fires the close handler → existing peer_left path.
Daemon (apps/cli/src/daemon/ws-lifecycle.ts):
- Add idle watchdog at 30s cadence, started after hello-ack.
- Bumps lastActivity on incoming message, ping, and pong frames.
- Sends sock.ping() if recent activity, terminates if idle >75s.
- Watchdog cleared on close handler + explicit close().
CLI 1.34.15 → 1.34.16. Broker stays 0.1.0 (deploys from main).
Spec: .artifacts/specs/2026-05-05-continuous-presence.md (full lease
model + resume token, this commit ships only the watchdogs — first
of four progressive layers).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ffd0621ccc