#!/usr/bin/env bash
# Health watch for the AVC phone agent. Run from cron every minute:
#   * * * * * /home/tocmo0nlord/avc-phone/scripts/healthwatch.sh
#
# Curls /health; after FAIL_THRESHOLD consecutive failures it sends ONE Twilio SMS to
# ALERT_SMS_TO (re-alerts at most every ALERT_COOLDOWN_MIN), and sends a recovery SMS
# when the service comes back. systemd already auto-restarts the process — this exists
# so a human finds out when restarts aren't enough (crash loop, GPU wedged, box issues).
# State lives in $STATE_DIR; activity is appended to healthwatch.log (gitignored).
set -u

PROJ="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
# shellcheck disable=SC1091
set -a; . "$PROJ/.env" 2>/dev/null; set +a

HEALTH_URL="${HEALTH_URL:-http://127.0.0.1:8200/health}"
FAIL_THRESHOLD="${FAIL_THRESHOLD:-3}"
ALERT_COOLDOWN_MIN="${ALERT_COOLDOWN_MIN:-30}"
STATE_DIR="${STATE_DIR:-$HOME/.local/state/avc-healthwatch}"
LOG="$PROJ/healthwatch.log"
mkdir -p "$STATE_DIR"
FAILS_F="$STATE_DIR/consecutive_fails"
ALERTED_F="$STATE_DIR/alerted_at_epoch"

log() { echo "$(date '+%F %T') $*" >> "$LOG"; }

send_sms() {
    local body="$1"
    if [ -z "${TWILIO_ACCOUNT_SID:-}" ] || [ -z "${TWILIO_AUTH_TOKEN:-}" ] \
       || [ -z "${TWILIO_PHONE_NUMBER:-}" ] || [ -z "${ALERT_SMS_TO:-}" ]; then
        log "ALERT (sms unconfigured): $body"
        return 1
    fi
    curl -s -m 15 -X POST \
        "https://api.twilio.com/2010-04-01/Accounts/$TWILIO_ACCOUNT_SID/Messages.json" \
        --data-urlencode "From=$TWILIO_PHONE_NUMBER" \
        --data-urlencode "To=$ALERT_SMS_TO" \
        --data-urlencode "Body=$body" \
        -u "$TWILIO_ACCOUNT_SID:$TWILIO_AUTH_TOKEN" > /dev/null \
        && log "SMS sent: $body" || log "SMS FAILED: $body"
}

fails=$(cat "$FAILS_F" 2>/dev/null || echo 0)

if curl -sf -m 5 "$HEALTH_URL" > /dev/null 2>&1; then
    if [ -f "$ALERTED_F" ]; then
        send_sms "AVC phone agent RECOVERED ($(hostname), $(date '+%H:%M'))."
        rm -f "$ALERTED_F"
    fi
    [ "$fails" != "0" ] && log "healthy again after $fails failure(s)"
    echo 0 > "$FAILS_F"
    exit 0
fi

fails=$((fails + 1))
echo "$fails" > "$FAILS_F"
log "health check FAILED ($fails consecutive)"

if [ "$fails" -ge "$FAIL_THRESHOLD" ]; then
    now=$(date +%s)
    last=$(cat "$ALERTED_F" 2>/dev/null || echo 0)
    if [ $((now - last)) -ge $((ALERT_COOLDOWN_MIN * 60)) ]; then
        state=$(systemctl is-active avc-phone 2>/dev/null || echo unknown)
        send_sms "ALERT: AVC phone agent DOWN — $fails failed health checks ($(hostname), systemd: $state, $(date '+%H:%M')). Callers may be getting errors."
        echo "$now" > "$ALERTED_F"
    fi
fi