box/bin/agent-health.sh

#!/bin/bash
# GOLDEN PATH: container -> VM (34.139.37.135) -> bl (100.123.153.75) -> netns -> browser -> agent
# This script is the operator's heartbeat. If it stops, agents go dark.
# When debugging: trace each hop. Don't assume -- verify.

# Operator health monitor for Muse agents on bl.
# Checks each agent via API every 5 minutes. If unresponsive:
# 1. Restart the browser
# 2. Re-check
# 3. Log failure if still down
#
# Run via systemd timer or cron: */5 * * * * /home/super/Projects/NetVM/bin/agent-health.sh
#
# Agents are defined in ACCOUNTS.md. This script reads the active ones.

NETVM_BIN="$(cd "$(dirname "$0")" && pwd)"
LOG="/tmp/agent-health.log"

check_cdp_port() {
    # Verify CDP port is actually listening in the netns.
    # A browser can be running but not bound to CDP (zombie state).
    local node=$1
    local cdp_port=$2
    if sudo ip netns exec "warp-$node" ss -tln 2>/dev/null | grep -q ":$cdp_port "; then
        return 0
    else
        return 1
    fi
}

check_agent() {
    local agent=$1
    local node=$2
    local cdp_port=$3

    # First: verify CDP port is listening (catches zombie browsers)
    if ! check_cdp_port "$node" "$cdp_port"; then
        echo "$(date -Iseconds) $agent: FAIL (cdp port $cdp_port not listening)" >> "$LOG"
        return 1
    fi

    # Then: try API messages command (lightweight check)
    if timeout 30 "$NETVM_BIN/netvm-exec.sh" "$node" -- python3 "$NETVM_BIN/muse-chat-api.py" --account "$agent" messages 1 > /dev/null 2>&1; then
        echo "$(date -Iseconds) $agent: OK" >> "$LOG"
        return 0
    else
        echo "$(date -Iseconds) $agent: FAIL (api timeout)" >> "$LOG"
        return 1
    fi
}

restart_browser() {
    local agent=$1
    local cdp_port=$2
    
    echo "$(date -Iseconds) $agent: restarting browser..." >> "$LOG"
    # Kill existing by exact PIDs (pkill patterns are unreliable)
    # Find chromium processes for this profile
    for pid in $(pgrep -f "chromium.*profiles/$agent" 2>/dev/null); do
        kill -9 "$pid" 2>/dev/null
    done
    sleep 3
    # Verify port is free before restart
    if sudo ip netns exec "warp-$agent" ss -tln 2>/dev/null | grep -q ":$cdp_port "; then
        echo "$(date -Iseconds) $agent: WARNING - port $cdp_port still bound after kill" >> "$LOG"
    fi
    # Restart via netvm-chrome.sh in its own systemd scope.
    # This oneshot service runs with KillMode=control-group, so anything
    # spawned directly under it (nohup AND setsid both stay in the cgroup)
    # is SIGKILLed when the service exits — observed 2026-10-03: every
    # restart "recovered" then died at service teardown, looping forever.
    # A transient scope escapes the service cgroup and survives.
    # NOTE: systemd-run --scope WAITS for the scope's processes (even with
    # --no-block, verified 2026-10-03), so background it — the scope itself
    # is an independent unit and outlives the wrapper.
    cd "$NETVM_BIN"
    systemd-run --user --scope --unit="netvm-chrome-$agent-$(date +%s)" \
        ./netvm-chrome.sh --headless --cdp-port "$cdp_port" "$agent" https://muse.ai \
        > "/tmp/bl-$agent.log" 2>&1 &
    sleep 15
    echo "$(date -Iseconds) $agent: browser restarted" >> "$LOG"
}

# Main
echo "=== Health check $(date -Iseconds) ===" >> "$LOG"

check_one() {
    local agent=$1
    local cdp_port=$2
    # node == agent == profile (unified naming)
    if ! check_agent "$agent" "$agent" "$cdp_port"; then
        restart_browser "$agent" "$cdp_port"
        sleep 5
        if ! check_agent "$agent" "$agent" "$cdp_port"; then
            echo "$(date -Iseconds) $agent: CRITICAL - still down after restart" >> "$LOG"
            # TODO: Alert operator (e.g., via board post or email)
        else
            echo "$(date -Iseconds) $agent: RECOVERED after restart" >> "$LOG"
        fi
    fi
}

# Every active node in the NODES.md registry gets checked — new nodes
# propagate automatically, no per-node blocks to add.
"$NETVM_BIN/netvm-registry.py" 2>/dev/null | while IFS=: read -r node port; do
    [ -n "$node" ] && [ -n "$port" ] && check_one "$node" "$port"
done

# Trim log (keep last 1000 lines)
tail -1000 "$LOG" > "$LOG.tmp" && mv "$LOG.tmp" "$LOG"