Skip to content

Business operating system for Claude Code — 57 skills, 21 agents, smart daemon. Unified inbox (WhatsApp/Email/Slack/Telegram), autonomous PR merge, full-AWS monitoring, revenue (Stripe+RevenueCat), e-commerce (Shopify), marketing (Klaviyo/Meta/GA4), voice (Bland/ElevenLabs), APM (Datadog/NewRelic/OTEL), YOLO mode.

feat(daemon): fires-watcher + push notification sinks (Telegram/Discord/ntfy/Pushover) (#21)

Aurora Capitalca18de8b05af
diff --git a/claude-ops/.claude-plugin/plugin.json b/claude-ops/.claude-plugin/plugin.json
index 56d645bd..b46d5d3b 100644
--- a/claude-ops/.claude-plugin/plugin.json
+++ b/claude-ops/.claude-plugin/plugin.json
@@ -224,6 +224,34 @@
"type": "string",
"sensitive": false,
"default": ""
+ },
+ "telegram_notify_chat_id": {
+ "title": "Telegram Notification Chat ID",
+ "description": "Chat ID to receive fire alerts. Defaults to telegram_phone's DM if unset.",
+ "type": "string",
+ "sensitive": false,
+ "default": ""
+ },
+ "ntfy_topic": {
+ "title": "ntfy.sh Topic",
+ "description": "Topic name (e.g., claude-ops-fires-<random>) for ntfy.sh pushes. Free, no account required.",
+ "type": "string",
+ "sensitive": true,
+ "default": ""
+ },
+ "pushover_user_key": {
+ "title": "Pushover User Key",
+ "description": "From https://pushover.net/ — paired with pushover_app_token.",
+ "type": "string",
+ "sensitive": true,
+ "default": ""
+ },
+ "pushover_app_token": {
+ "title": "Pushover App Token",
+ "description": "App-specific token from https://pushover.net/apps/build",
+ "type": "string",
+ "sensitive": true,
+ "default": ""
}
}
}
diff --git a/claude-ops/docs/notifications.md b/claude-ops/docs/notifications.md
new file mode 100644
index 00000000..2ca474f3
--- /dev/null
+++ b/claude-ops/docs/notifications.md
@@ -0,0 +1,151 @@
+# Push Notifications (`fires-watcher`)
+
+`claude-ops` ships a background service called `fires-watcher` that replaces
+the old polling workflow of `/ops:fires`. Instead of you having to remember to
+ask for fires, the daemon polls your fire sources every 60 seconds and pushes
+a notification the moment a new CRITICAL or HIGH incident appears.
+
+> Related issue: [#21 — push notifications for P0s](https://github.com/Lifecycle-Innovations-Limited/claude-ops/issues/21).
+
+## What `fires-watcher` does
+
+Every 60 seconds (configurable via `FIRES_WATCHER_INTERVAL`), the watcher:
+
+1. Runs `bin/ops-infra` and inspects ECS cluster health. A cluster in
+ `degraded` state produces a **HIGH** event; a cluster that is fully
+ `down` produces a **CRITICAL** event.
+2. Queries the Sentry API (`organizations/<org>/issues/?query=is:unresolved`)
+ for unresolved issues at `level=error` (**HIGH**) or `level=fatal`
+ (**CRITICAL**) — only when `$SENTRY_AUTH_TOKEN` and `sentry_org` are both
+ configured.
+3. Diffs the result against the previous state file at
+ `~/.claude/plugins/data/ops-ops-marketplace/fires-watcher.state.json`.
+4. For every new or escalated incident, calls
+ `scripts/ops-notify.sh <severity> <title> <body> [--link <url>]`, which
+ fans out to every configured sink.
+5. Updates a health file at
+ `~/.claude/plugins/data/ops-ops-marketplace/fires-watcher.health` so the
+ main daemon can monitor it.
+
+The watcher is **read-only** — it never mutates infrastructure (Rule 5).
+
+## The six sink options
+
+`ops-notify.sh` tries sinks in this order and fans out to **all** that are
+configured (there is no "pick one" — you get alerts everywhere you want them):
+
+| Priority | Sink | Trigger env / pref |
+| -------- | ----------------- | ----------------------------------------------------------------------------------- |
+| 1 | Telegram bot | `$TELEGRAM_BOT_TOKEN` + `$TELEGRAM_NOTIFY_CHAT_ID` (falls back to `$TELEGRAM_OWNER_ID`) |
+| 2 | Discord webhook | `$DISCORD_WEBHOOK_URL` (shared with `/ops:comms discord send`) |
+| 3 | ntfy.sh | `$NTFY_TOPIC` (optionally `$NTFY_SERVER` for a self-hosted instance) |
+| 4 | Pushover | `$PUSHOVER_USER` + `$PUSHOVER_TOKEN` |
+| 5 | macOS `osascript` | Local desktop notification (macOS only — guarded behind a Darwin check) |
+| 6 | stderr log | Always runs if nothing else is configured — written to `logs/ops-notify.log` |
+
+Every env var above also has a `$PREFS_PATH` fallback (the dispatch script
+reads `~/.claude/plugins/data/ops-ops-marketplace/preferences.json` when the
+env var isn't set), so you don't have to export secrets globally.
+
+### Severity-to-emoji mapping
+
+| Severity | Emoji | Notes |
+| -------- | ----- | -------------------------------------- |
+| CRITICAL | 🔴 | Pushover priority `1`, ntfy priority `5` |
+| HIGH | 🟠 | Pushover priority `1`, ntfy priority `4` |
+| MEDIUM | 🟡 | Pushover priority `0`, ntfy priority `3` |
+| LOW | 🟢 | Pushover priority `-1`, ntfy priority `2` |
+
+## How to choose a sink
+
+- **Most users →** Telegram. You already set up a Telegram bot to use
+ `/ops:comms`, so this is zero extra work.
+- **No account, free →** ntfy.sh. Pick a random topic name like
+ `claude-ops-fires-<random>`, subscribe from the ntfy mobile app, done.
+- **Premium iOS/Android →** Pushover. ~$5 one-time per platform, fastest and
+ most reliable delivery with priority bypass for CRITICAL.
+- **Team channel →** Discord webhook. Drops the alert into an existing
+ `#incidents` channel where your team already lives.
+- **Desktop only →** macOS `osascript`. Silent fallback; fine for a dev
+ laptop but don't rely on it as your only sink.
+
+Configure multiple — if your laptop is offline, a Pushover push still reaches
+you; if Telegram's API is flaky, Discord still works.
+
+## Debounce rules
+
+To avoid waking you up every minute about the same fire:
+
+- **New incident** — notify immediately, record fingerprint + severity +
+ timestamp.
+- **Same fingerprint, same severity** — suppress for **30 minutes** from the
+ last notification. Override with `FIRES_WATCHER_DEBOUNCE` (seconds).
+- **Same fingerprint, higher severity** — re-notify immediately (e.g., ECS
+ cluster went from `degraded` → `down`). The stored severity is bumped so
+ the next HIGH event doesn't re-trigger until the 30-minute window elapses.
+- **Incident resolves** — the fingerprint simply stops appearing in the probe
+ output; no "all clear" ping is sent (noise vs signal tradeoff).
+
+Fingerprints are stable across ticks:
+- Infra: `infra:<cluster>:<status>`
+- Sentry: `sentry:<issue_id>`
+
+State lives at `~/.claude/plugins/data/ops-ops-marketplace/fires-watcher.state.json`
+and survives daemon restarts.
+
+## Testing your setup
+
+Send a test event through every configured sink:
+
+```bash
+scripts/ops-notify.sh CRITICAL "test" "is this thing on"
+```
+
+You should see a red-dot alert on every sink you've wired up. Check
+`logs/ops-notify.log` under your data dir for a per-sink status line.
+
+To verify the watcher loop, run it in the foreground for 2-3 ticks:
+
+```bash
+FIRES_WATCHER_INTERVAL=10 scripts/ops-fires-watcher.sh
+```
+
+Tail the log with `tail -f ~/.claude/plugins/data/ops-ops-marketplace/logs/fires-watcher.log`.
+
+## Enabling / disabling the watcher
+
+`fires-watcher` is **disabled by default**. Opt in either way:
+
+1. **Guided:** `/ops:setup notifications` — walks you through sink selection
+ and enables the service.
+2. **Manual:** edit
+ `~/.claude/plugins/data/ops-ops-marketplace/daemon-services.json` and flip
+ `services.fires-watcher.enabled` to `true`, then restart the daemon.
+
+To disable it again without losing config:
+
+- Run `/ops:setup notifications --skip`, **or**
+- Flip `enabled` back to `false` in `daemon-services.json` and restart the
+ daemon (`scripts/ops-daemon.sh restart`).
+
+## Polling vs event-driven — tradeoffs
+
+This design uses 60-second polling rather than webhooks. Rationale:
+
+- Sentry + AWS both support webhooks, but receiving them would require the
+ user to run (and expose) an HTTP listener — a hard ask for a CLI plugin
+ that runs on random laptops behind NAT.
+- 60 s worst-case latency is acceptable for CRITICAL P0s; most on-call
+ rotations already tolerate ≥ 1 minute of alert lag.
+- Polling works uniformly across any fire source (Sentry, AWS, Datadog, New
+ Relic, custom CloudWatch) without per-source webhook handlers.
+- If sub-minute latency becomes important later, the watcher can be swapped
+ for a webhook listener without changing the notification surface —
+ `ops-notify.sh` is the stable boundary.
+
+## Files
+
+- `scripts/ops-fires-watcher.sh` — the daemon poll loop
+- `scripts/ops-notify.sh` — the sink dispatcher
+- `scripts/daemon-services.default.json` — service registration
+- `skills/setup/SKILL.md` §3m — guided setup flow
diff --git a/claude-ops/scripts/daemon-services.default.json b/claude-ops/scripts/daemon-services.default.json
index d0e77735..b9875fe4 100644
--- a/claude-ops/scripts/daemon-services.default.json
+++ b/claude-ops/scripts/daemon-services.default.json
@@ -19,6 +19,14 @@
"health_file": "~/.claude/plugins/data/ops-ops-marketplace/memories/.health",
"restart_delay": 3600,
"cron": "*/30 * * * *"
+ },
+ "fires-watcher": {
+ "enabled": false,
+ "command": "__SCRIPTS_DIR__/ops-fires-watcher.sh",
+ "health_file": "~/.claude/plugins/data/ops-ops-marketplace/fires-watcher.health",
+ "restart_delay": 60,
+ "max_restarts": 20,
+ "_note": "Polls ops-infra + Sentry every 60s; pushes notifications on new CRITICAL/HIGH incidents. Requires at least one notify sink (Telegram bot, Discord webhook, ntfy, Pushover, or macOS). Run /ops:setup notifications to enable."
}
}
}
diff --git a/claude-ops/scripts/ops-fires-watcher.sh b/claude-ops/scripts/ops-fires-watcher.sh
new file mode 100755
index 00000000..bb9d9f1f
--- /dev/null
+++ b/claude-ops/scripts/ops-fires-watcher.sh
@@ -0,0 +1,237 @@
+#!/usr/bin/env bash
+# ops-fires-watcher.sh — Daemon loop that polls fire sources every 60s and
+# pushes notifications for new CRITICAL/HIGH incidents.
+#
+# Sources polled each tick:
+# - bin/ops-infra (ECS cluster health — 'down' / 'degraded' → HIGH/CRITICAL)
+# - Sentry API (unresolved issues at level:error/fatal when
+# $SENTRY_AUTH_TOKEN + sentry_org are configured)
+#
+# State file: ${DATA_DIR}/fires-watcher.state.json — keyed by incident
+# fingerprint, tracks { severity, last_notified_at, first_seen_at }.
+#
+# Debounce policy:
+# - Don't re-notify for the same fingerprint within 30 minutes
+# - UNLESS the severity *rises* (e.g. degraded → down), in which case we
+# re-notify immediately and update the stored severity.
+#
+# Rule 5: read-only polling — this watcher never mutates infra.
+
+set -euo pipefail
+
+# ─── Paths ───────────────────────────────────────────────────────────────────
+PLUGIN_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+# shellcheck disable=SC1091
+[ -r "$PLUGIN_ROOT/lib/os-detect.sh" ] && . "$PLUGIN_ROOT/lib/os-detect.sh"
+# shellcheck disable=SC1091
+[ -r "$PLUGIN_ROOT/lib/credential-store.sh" ] && . "$PLUGIN_ROOT/lib/credential-store.sh"
+
+DATA_DIR="${OPS_DATA_DIR:-${CLAUDE_PLUGIN_DATA_DIR:-$HOME/.claude/plugins/data/ops-ops-marketplace}}"
+LOG_DIR="$DATA_DIR/logs"
+LOG="$LOG_DIR/fires-watcher.log"
+STATE_FILE="${FIRES_WATCHER_STATE:-$DATA_DIR/fires-watcher.state.json}"
+HEALTH_FILE="${FIRES_WATCHER_HEALTH:-$DATA_DIR/fires-watcher.health}"
+NOTIFY="$PLUGIN_ROOT/scripts/ops-notify.sh"
+OPS_INFRA="$PLUGIN_ROOT/bin/ops-infra"
+PREFS_PATH="${PREFS_PATH:-$DATA_DIR/preferences.json}"
+
+POLL_INTERVAL="${FIRES_WATCHER_INTERVAL:-60}"
+DEBOUNCE_SECS="${FIRES_WATCHER_DEBOUNCE:-1800}" # 30 minutes
+
+mkdir -p "$LOG_DIR"
+mkdir -p "$(dirname "$STATE_FILE")"
+
+# ─── Logging ────────────────────────────────────────────────────────────────
+log() {
+ printf '%s [fires-watcher] %s\n' "$(date -u +%Y-%m-%dT%H:%M:%SZ)" "$1" >> "$LOG"
+}
+
+# ─── Severity rank (higher = worse) ─────────────────────────────────────────
+sev_rank() {
+ case "$1" in
+ CRITICAL) echo 4 ;;
+ HIGH) echo 3 ;;
+ MEDIUM) echo 2 ;;
+ LOW) echo 1 ;;
+ *) echo 0 ;;
+ esac
+}
+
+# ─── State I/O ──────────────────────────────────────────────────────────────
+ensure_state() {
+ if [ ! -f "$STATE_FILE" ]; then
+ echo '{"incidents":{}}' > "$STATE_FILE"
+ fi
+}
+
+# Decide if we should notify given a fingerprint + new severity.
+# Echoes "notify" or "skip"; on notify, updates the state file.
+should_notify() {
+ local fp="$1" new_sev="$2" now_epoch
+ now_epoch="$(date +%s)"
+ ensure_state
+
+ command -v jq >/dev/null 2>&1 || { echo "notify"; return 0; }
+
+ local prev_sev prev_notified
+ prev_sev="$(jq -r --arg fp "$fp" '.incidents[$fp].severity // empty' "$STATE_FILE")"
+ prev_notified="$(jq -r --arg fp "$fp" '.incidents[$fp].last_notified_at // 0' "$STATE_FILE")"
+
+ local decision="skip"
+ if [ -z "$prev_sev" ]; then
+ decision="notify" # brand new incident
+ elif [ "$(sev_rank "$new_sev")" -gt "$(sev_rank "$prev_sev")" ]; then
+ decision="notify" # severity escalated
+ else
+ local age=$(( now_epoch - prev_notified ))
+ if [ "$age" -ge "$DEBOUNCE_SECS" ]; then
+ decision="notify" # debounce window elapsed
+ fi
+ fi
+
+ if [ "$decision" = "notify" ]; then
+ local tmp
+ tmp="$(mktemp)"
+ jq --arg fp "$fp" \
+ --arg sev "$new_sev" \
+ --argjson now "$now_epoch" \
+ '.incidents[$fp] = {
+ severity: $sev,
+ last_notified_at: $now,
+ first_seen_at: (.incidents[$fp].first_seen_at // $now)
+ }' "$STATE_FILE" > "$tmp" && mv "$tmp" "$STATE_FILE"
+ fi
+
+ echo "$decision"
+}
+
+# ─── Notify wrapper ─────────────────────────────────────────────────────────
+dispatch() {
+ local severity="$1" title="$2" body="$3" link="${4:-}"
+ if [ ! -x "$NOTIFY" ]; then
+ log "WARN: $NOTIFY not executable — skipping dispatch"
+ return 0
+ fi
+ if [ -n "$link" ]; then
+ "$NOTIFY" "$severity" "$title" "$body" --link "$link" || \
+ log "dispatch failed for title=$title"
+ else
+ "$NOTIFY" "$severity" "$title" "$body" || \
+ log "dispatch failed for title=$title"
+ fi
+}
+
+# ─── Source: ops-infra ──────────────────────────────────────────────────────
+probe_infra() {
+ [ -x "$OPS_INFRA" ] || return 0
+ command -v jq >/dev/null 2>&1 || return 0
+
+ local output
+ output="$("$OPS_INFRA" 2>/dev/null || echo '{}')"
+ [ -z "$output" ] && return 0
+
+ # Emit lines: "SEV<TAB>fingerprint<TAB>title<TAB>body"
+ echo "$output" | jq -r '
+ (.clusters // [])[]
+ | select(.status == "down" or .status == "degraded")
+ | "\(if .status == "down" then "CRITICAL" else "HIGH" end)\t" +
+ "infra:\(.cluster):\(.status)\t" +
+ "ECS \(.cluster) \(.status)\t" +
+ "Cluster \(.cluster) is \(.status). Services: \((.services // []) | map(select(.running != .desired)) | map("\(.name) \(.running)/\(.desired)") | join(", "))"
+ ' 2>/dev/null || true
+}
+
+# ─── Source: Sentry ─────────────────────────────────────────────────────────
+probe_sentry() {
+ local token="${SENTRY_AUTH_TOKEN:-}"
+ local org="${SENTRY_ORG:-}"
+ if [ -z "$token" ] && [ -f "$PREFS_PATH" ] && command -v jq >/dev/null 2>&1; then
+ token="$(jq -r '.sentry_auth_token // empty' "$PREFS_PATH" 2>/dev/null || true)"
+ fi
+ if [ -z "$org" ] && [ -f "$PREFS_PATH" ] && command -v jq >/dev/null 2>&1; then
+ org="$(jq -r '.sentry_org // empty' "$PREFS_PATH" 2>/dev/null || true)"
+ fi
+
+ if [ -z "$token" ] || [ -z "$org" ]; then
+ return 0
+ fi
+ command -v curl >/dev/null 2>&1 || return 0
+ command -v jq >/dev/null 2>&1 || return 0
+
+ local response
+ response="$(curl -s --max-time 15 \
+ -H "Authorization: Bearer ${token}" \
+ "https://sentry.io/api/0/organizations/${org}/issues/?query=is:unresolved+level:[error,fatal]&limit=25" \
+ 2>/dev/null || echo '[]')"
+
+ # Defensively fall back if response is not valid JSON array
+ if ! echo "$response" | jq -e 'type == "array"' >/dev/null 2>&1; then
+ return 0
+ fi
+
+ echo "$response" | jq -r '
+ .[]
+ | select(.level == "error" or .level == "fatal")
+ | "\(if .level == "fatal" then "CRITICAL" else "HIGH" end)\t" +
+ "sentry:\(.id)\t" +
+ "Sentry: \((.title // "issue") | .[0:80])\t" +
+ "Project: \(.project.slug // "unknown"). Events: \(.count // "?"). Users: \(.userCount // "?"). URL: \(.permalink // "")"
+ ' 2>/dev/null || true
+}
+
+# ─── Health writer ──────────────────────────────────────────────────────────
+write_health() {
+ local status="$1" msg="${2:-}"
+ cat > "$HEALTH_FILE" <<EOF
+{
+ "status": "${status}",
+ "message": "${msg}",
+ "last_tick": "$(date -u +%Y-%m-%dT%H:%M:%SZ)",
+ "poll_interval": ${POLL_INTERVAL},
+ "state_file": "${STATE_FILE}"
+}
+EOF
+}
+
+# ─── Main tick ──────────────────────────────────────────────────────────────
+tick() {
+ local events new_count=0 seen_count=0
+ events="$( (probe_infra; probe_sentry) 2>/dev/null || true )"
+ if [ -z "$events" ]; then
+ write_health "ok" "no events"
+ return 0
+ fi
+
+ while IFS=$'\t' read -r sev fp title body; do
+ [ -z "${fp:-}" ] && continue
+ seen_count=$((seen_count + 1))
+ local decision
+ decision="$(should_notify "$fp" "$sev")"
+ if [ "$decision" = "notify" ]; then
+ local link=""
+ # Prefer an explicit URL embedded in the body (e.g., Sentry permalink)
+ link="$(printf '%s' "$body" | sed -n 's#.*URL: \(https\?://[^ ]*\).*#\1#p' | head -1)"
+ if [ -n "$link" ]; then
+ dispatch "$sev" "$title" "$body" "$link"
+ else
+ dispatch "$sev" "$title" "$body"
+ fi
+ new_count=$((new_count + 1))
+ log "NEW $sev $fp :: $title"
+ fi
+ done <<< "$events"
+
+ write_health "ok" "seen=${seen_count} notified=${new_count}"
+}
+
+# ─── Loop ───────────────────────────────────────────────────────────────────
+log "START fires-watcher (interval=${POLL_INTERVAL}s, debounce=${DEBOUNCE_SECS}s)"
+write_health "starting" "booting fires-watcher"
+
+trap 'log "STOP SIGTERM"; write_health "stopped" "received SIGTERM"; exit 0' TERM
+trap 'log "STOP SIGINT"; write_health "stopped" "received SIGINT"; exit 0' INT
+
+while true; do
+ tick || log "TICK ERROR (continuing)"
+ sleep "$POLL_INTERVAL"
+done
diff --git a/claude-ops/scripts/ops-notify.sh b/claude-ops/scripts/ops-notify.sh
new file mode 100755
index 00000000..59f921c6
--- /dev/null
+++ b/claude-ops/scripts/ops-notify.sh
@@ -0,0 +1,273 @@
+#!/usr/bin/env bash
+# ops-notify.sh — Fan-out notification dispatcher for claude-ops.
+#
+# Usage:
+# scripts/ops-notify.sh <severity> <title> <body> [--link <url>]
+#
+# <severity> is one of: CRITICAL | HIGH | MEDIUM | LOW (case-insensitive).
+# Delivery sinks (all optional, all fanned out when configured):
+# 1. Telegram bot — $TELEGRAM_BOT_TOKEN + $TELEGRAM_NOTIFY_CHAT_ID (or $TELEGRAM_OWNER_ID)
+# 2. Discord webhook — $DISCORD_WEBHOOK_URL
+# 3. ntfy.sh — $NTFY_TOPIC (optionally $NTFY_SERVER, default https://ntfy.sh)
+# 4. Pushover — $PUSHOVER_USER + $PUSHOVER_TOKEN
+# 5. macOS osascript — local desktop notification (only on macOS)
+# 6. stderr log — always runs if nothing else is configured
+#
+# Rule 0: no hardcoded tokens, chat IDs, or webhooks live in this file — every
+# credential is read from the environment or $PREFS_PATH.
+
+set -euo pipefail
+
+# ─── Paths + library sourcing ───────────────────────────────────────────────
+SCRIPT_DIR="$(cd "$(dirname "$0")/.." && pwd)"
+# shellcheck disable=SC1091
+[ -r "$SCRIPT_DIR/lib/os-detect.sh" ] && . "$SCRIPT_DIR/lib/os-detect.sh"
+# shellcheck disable=SC1091
+[ -r "$SCRIPT_DIR/lib/credential-store.sh" ] && . "$SCRIPT_DIR/lib/credential-store.sh"
+
+DATA_DIR="${OPS_DATA_DIR:-${CLAUDE_PLUGIN_DATA_DIR:-$HOME/.claude/plugins/data/ops-ops-marketplace}}"
+LOG_DIR="$DATA_DIR/logs"
+LOG="$LOG_DIR/ops-notify.log"
+PREFS_PATH="${PREFS_PATH:-$DATA_DIR/preferences.json}"
+
+mkdir -p "$LOG_DIR"
+
+# ─── Portable OS check (matches bin/ops-autofix pattern) ────────────────────
+IS_MACOS=false
+[ "$(uname -s)" = "Darwin" ] && IS_MACOS=true
+
+# ─── Logging ────────────────────────────────────────────────────────────────
+log() {
+ printf '%s [ops-notify] %s\n' "$(date -u +%Y-%m-%dT%H:%M:%SZ)" "$1" >> "$LOG"
+}
+
+# ─── Argument parsing ───────────────────────────────────────────────────────
+if [[ $# -lt 3 ]]; then
+ echo "Usage: $0 <CRITICAL|HIGH|MEDIUM|LOW> <title> <body> [--link <url>]" >&2
+ exit 2
+fi
+
+SEVERITY="$(echo "$1" | tr '[:lower:]' '[:upper:]')"
+TITLE="$2"
+BODY="$3"
+shift 3
+LINK=""
+while [[ $# -gt 0 ]]; do
+ case "$1" in
+ --link)
+ LINK="${2:-}"; shift 2 ;;
+ *)
+ shift ;;
+ esac
+done
+
+case "$SEVERITY" in
+ CRITICAL|HIGH|MEDIUM|LOW) : ;;
+ *)
+ echo "ops-notify: invalid severity '$SEVERITY' (expected CRITICAL|HIGH|MEDIUM|LOW)" >&2
+ exit 2
+ ;;
+esac
+
+# ─── Severity emoji mapping ─────────────────────────────────────────────────
+severity_emoji() {
+ case "$1" in
+ CRITICAL) echo "🔴" ;;
+ HIGH) echo "🟠" ;;
+ MEDIUM) echo "🟡" ;;
+ LOW) echo "🟢" ;;
+ *) echo "⚪" ;;
+ esac
+}
+
+severity_color_hex() {
+ # Decimal RGB values for Discord embed color field
+ case "$1" in
+ CRITICAL) echo "15158332" ;; # red
+ HIGH) echo "15105570" ;; # orange
+ MEDIUM) echo "15844367" ;; # yellow
+ LOW) echo "3066993" ;; # green
+ *) echo "9807270" ;; # gray
+ esac
+}
+
+EMOJI="$(severity_emoji "$SEVERITY")"
+FULL_TITLE="${EMOJI} [${SEVERITY}] ${TITLE}"
+
+# ─── Pref loading helper ────────────────────────────────────────────────────
+# Pulls a top-level key from $PREFS_PATH if the file exists. Caller may
+# use this to provide fallbacks when env vars aren't set.
+pref_get() {
+ local key="$1"
+ [ -f "$PREFS_PATH" ] || return 0
+ command -v jq >/dev/null 2>&1 || return 0
+ jq -r --arg k "$key" '.[$k] // empty' "$PREFS_PATH" 2>/dev/null || true
+}
+
+# ─── Sink: Telegram bot ─────────────────────────────────────────────────────
+send_telegram() {
+ local token="${TELEGRAM_BOT_TOKEN:-$(pref_get telegram_bot_token)}"
+ local chat_id="${TELEGRAM_NOTIFY_CHAT_ID:-${TELEGRAM_OWNER_ID:-$(pref_get telegram_notify_chat_id)}}"
+ if [ -z "$token" ] || [ -z "$chat_id" ]; then
+ return 1
+ fi
+ local text
+ if [ -n "$LINK" ]; then
+ text="${FULL_TITLE}"$'\n\n'"${BODY}"$'\n\n'"${LINK}"
+ else
+ text="${FULL_TITLE}"$'\n\n'"${BODY}"
+ fi
+ curl -s -o /dev/null -w '%{http_code}' \
+ -X POST "https://api.telegram.org/bot${token}/sendMessage" \
+ --data-urlencode "chat_id=${chat_id}" \
+ --data-urlencode "text=${text}" \
+ --max-time 10 || echo "000"
+}
+
+# ─── Sink: Discord webhook ──────────────────────────────────────────────────
+send_discord() {
+ local url="${DISCORD_WEBHOOK_URL:-$(pref_get discord_webhook_url)}"
+ if [ -z "$url" ]; then
+ return 1
+ fi
+ local color payload
+ color="$(severity_color_hex "$SEVERITY")"
+ # Build JSON via jq to handle special characters safely
+ if command -v jq >/dev/null 2>&1; then
+ payload="$(jq -n \
+ --arg title "$FULL_TITLE" \
+ --arg body "$BODY" \
+ --arg link "$LINK" \
+ --argjson color "$color" \
+ '{
+ embeds: [{
+ title: $title,
+ description: $body,
+ color: $color,
+ url: (if $link == "" then null else $link end)
+ }]
+ }')"
+ else
+ # Fallback: minimal escape of double quotes + newlines
+ local esc_title esc_body
+ esc_title="$(printf '%s' "$FULL_TITLE" | sed 's/"/\\"/g')"
+ esc_body="$(printf '%s' "$BODY" | sed 's/"/\\"/g' | tr '\n' ' ')"
+ payload="{\"embeds\":[{\"title\":\"${esc_title}\",\"description\":\"${esc_body}\",\"color\":${color}}]}"
+ fi
+ curl -s -o /dev/null -w '%{http_code}' \
+ -X POST "$url" \
+ -H 'Content-Type: application/json' \
+ --data "$payload" \
+ --max-time 10 || echo "000"
+}
+
+# ─── Sink: ntfy.sh ──────────────────────────────────────────────────────────
+send_ntfy() {
+ local topic="${NTFY_TOPIC:-$(pref_get ntfy_topic)}"
+ if [ -z "$topic" ]; then
+ return 1
+ fi
+ local server="${NTFY_SERVER:-https://ntfy.sh}"
+ local priority
+ case "$SEVERITY" in
+ CRITICAL) priority="5" ;;
+ HIGH) priority="4" ;;
+ MEDIUM) priority="3" ;;
+ LOW) priority="2" ;;
+ *) priority="3" ;;
+ esac
+ local body_with_link="$BODY"
+ [ -n "$LINK" ] && body_with_link="${BODY}"$'\n\n'"${LINK}"
+ local sev_lc
+ sev_lc="$(echo "$SEVERITY" | tr '[:upper:]' '[:lower:]')"
+ curl -s -o /dev/null -w '%{http_code}' \
+ -X POST "${server}/${topic}" \
+ -H "Title: ${FULL_TITLE}" \
+ -H "Priority: ${priority}" \
+ -H "Tags: warning,${sev_lc}" \
+ --data "$body_with_link" \
+ --max-time 10 || echo "000"
+}
+
+# ─── Sink: Pushover ─────────────────────────────────────────────────────────
+send_pushover() {
+ local user="${PUSHOVER_USER:-$(pref_get pushover_user_key)}"
+ local token="${PUSHOVER_TOKEN:-$(pref_get pushover_app_token)}"
+ if [ -z "$user" ] || [ -z "$token" ]; then
+ return 1
+ fi
+ local priority
+ case "$SEVERITY" in
+ CRITICAL) priority="1" ;; # high-priority, bypass quiet hours
+ HIGH) priority="1" ;;
+ MEDIUM) priority="0" ;;
+ LOW) priority="-1" ;;
+ *) priority="0" ;;
+ esac
+ local url_args=()
+ if [ -n "$LINK" ]; then
+ url_args+=(--data-urlencode "url=${LINK}")
+ fi
+ curl -s -o /dev/null -w '%{http_code}' \
+ -X POST "https://api.pushover.net/1/messages.json" \
+ --data-urlencode "token=${token}" \
+ --data-urlencode "user=${user}" \
+ --data-urlencode "title=${FULL_TITLE}" \
+ --data-urlencode "message=${BODY}" \
+ --data-urlencode "priority=${priority}" \
+ "${url_args[@]}" \
+ --max-time 10 || echo "000"
+}
+
+# ─── Sink: macOS osascript ──────────────────────────────────────────────────
+send_macos() {
+ $IS_MACOS || return 1
+ command -v osascript >/dev/null 2>&1 || return 1
+ # Escape double quotes for AppleScript literal
+ local esc_title esc_body
+ esc_title="$(printf '%s' "$FULL_TITLE" | sed 's/"/\\"/g')"
+ esc_body="$(printf '%s' "$BODY" | sed 's/"/\\"/g')"
+ osascript -e "display notification \"${esc_body}\" with title \"${esc_title}\"" >/dev/null 2>&1 || true
+ echo "200"
+}
+
+# ─── Sink: stderr ───────────────────────────────────────────────────────────
+send_stderr() {
+ {
+ printf '%s\n' "$FULL_TITLE"
+ printf '%s\n' "$BODY"
+ [ -n "$LINK" ] && printf 'link: %s\n' "$LINK"
+ } >&2
+ echo "200"
+}
+
+# ─── Dispatch ───────────────────────────────────────────────────────────────
+delivered=0
+
+run_sink() {
+ local name="$1" fn="$2"
+ local result
+ result="$($fn 2>/dev/null || true)"
+ if [ -n "$result" ] && [ "$result" != "000" ]; then
+ log "SINK ${name}: status=${result}"
+ delivered=$((delivered + 1))
+ else
+ log "SINK ${name}: skipped (not configured or failed)"
+ fi
+}
+
+run_sink telegram send_telegram
+run_sink discord send_discord
+run_sink ntfy send_ntfy
+run_sink pushover send_pushover
+run_sink macos send_macos
+
+if [ "$delivered" -eq 0 ]; then
+ log "SINK stderr: fallback (no sinks configured)"
+ send_stderr >/dev/null 2>&1 || true
+ printf '%s [ops-notify] %s | %s\n' \
+ "$(date -u +%Y-%m-%dT%H:%M:%SZ)" "$FULL_TITLE" "$BODY" >> "$LOG"
+fi
+
+log "DISPATCHED severity=${SEVERITY} delivered=${delivered} title=\"${TITLE}\""
+exit 0
diff --git a/claude-ops/skills/setup/SKILL.md b/claude-ops/skills/setup/SKILL.md
index c74940d1..6232a5da 100644
--- a/claude-ops/skills/setup/SKILL.md
+++ b/claude-ops/skills/setup/SKILL.md
@@ -2055,6 +2055,114 @@ Prefer a Doppler reference (`doppler:STRIPE_SECRET_KEY`, `doppler:REVENUECAT_API
---
+### 3m — Notifications (fires-watcher sinks)
+
+Sets up push notifications for CRITICAL/HIGH fires so the user stops having to poll `/ops:fires` manually. Gated behind the `fires-watcher` daemon service (disabled by default).
+
+**Before prompting**, run the sink auto-scan **in background** (Rule 4 — all parallel):
+
+```bash
+# Already-configured sinks (env + $PREFS_PATH)
+printenv TELEGRAM_BOT_TOKEN TELEGRAM_NOTIFY_CHAT_ID TELEGRAM_OWNER_ID \
+ DISCORD_WEBHOOK_URL NTFY_TOPIC PUSHOVER_USER PUSHOVER_TOKEN 2>/dev/null
+
+jq -r 'to_entries
+ | map(select(.key | test("telegram_bot_token|telegram_notify_chat_id|discord_webhook_url|ntfy_topic|pushover_user_key|pushover_app_token")))
+ | .[] | "\(.key)=\(.value | if type == "string" and length > 6 then .[0:6] + "***" else . end)"' \
+ "$PREFS_PATH" 2>/dev/null
+
+# Macro: is the Telegram bot already configured via /ops:setup telegram?
+jq -r '.telegram.bot_token // empty, .telegram.owner_id // empty' "$PREFS_PATH" 2>/dev/null
+```
+
+Cache the results. Then **filter the option list to what isn't already configured** (Rule 1 — max 4).
+
+Present via `AskUserQuestion` (≤4 options). Typical first batch:
+
+| Option | Header | Description |
+| ------------------------------------- | ---------- | ------------------------------------------------------------------------------ |
+| Use Telegram (recommended) | telegram | Reuse the bot you already configured in 3a. Zero extra setup. |
+| Configure ntfy.sh | ntfy | Free, no account. Pick a random topic, install the app on your phone. |
+| Configure Pushover | pushover | ~$5 one-time. Highest-reliability mobile delivery with priority bypass for P0. |
+| Skip — poll with /ops:fires instead | skip | Leaves `fires-watcher` disabled. You'll keep asking for fires manually. |
+
+If Telegram's bot token + owner ID aren't already in `$PREFS_PATH`, swap `[Use Telegram]` for `[Configure Telegram bot — jump to 3a]` and re-enter 3a before returning here. Per Rule 3, don't silently skip — always offer an explicit `[Skip]`.
+
+For anyone who wants Discord on top (many users want both team + personal delivery), run a second `AskUserQuestion`:
+
+| Option | Header | Description |
+| --------------------------------- | ---------- | --------------------------------------------------- |
+| Also add Discord webhook | discord | Fan out to a #incidents channel in addition to X. |
+| No — just the sink I picked | done | Single-sink setup. |
+
+#### Per-sink capture
+
+**Telegram** — if `$TELEGRAM_NOTIFY_CHAT_ID` is empty, default it to `$TELEGRAM_OWNER_ID` (the bot will DM the owner) and save `{"telegram_notify_chat_id": "<owner_id>"}` to `$PREFS_PATH`. Smoke-test in background (Rule 4):
+
+```bash
+scripts/ops-notify.sh LOW "setup test" "fires-watcher Telegram sink is live" &
+```
+
+**ntfy.sh** — generate a random topic if the user has none:
+
+```bash
+TOPIC="claude-ops-fires-$(openssl rand -hex 6 2>/dev/null || head -c 12 /dev/urandom | base64 | tr -dc 'a-z0-9' | head -c 12)"
+echo "$TOPIC"
+```
+
+Show the topic and install instructions (`open https://ntfy.sh/<topic>` in a browser, or scan the QR in the ntfy mobile app). Save to `$PREFS_PATH` under `ntfy_topic`.
+
+**Pushover** — per Rule 3 offer: `[Paste user key + app token]` / `[Deep hunt — spawn agent]` / `[Skip]`. On paste, validate with a smoke-test in background (Rule 4):
+
+```bash
+curl -s -X POST https://api.pushover.net/1/messages.json \
+ --data-urlencode "token=$PUSHOVER_TOKEN" \
+ --data-urlencode "user=$PUSHOVER_USER" \
+ --data-urlencode "title=claude-ops setup test" \
+ --data-urlencode "message=fires-watcher Pushover sink is live" &
+```
+
+Expect `{"status":1,...}` within 5 s.
+
+**Discord** — if the user already set `discord_default_webhook_url` via a prior run (or issue #20's work), reuse it; otherwise Rule-3-prompt for a webhook URL (`https://discord.com/api/webhooks/...`).
+
+#### Save + enable daemon service
+
+Write to `$PREFS_PATH` (merge, never overwrite unrelated keys):
+
+```json
+{
+ "notifications": {
+ "configured_sinks": ["telegram", "ntfy"],
+ "configured_at": "<ISO timestamp>"
+ },
+ "telegram_notify_chat_id": "<chat_id>",
+ "ntfy_topic": "<topic>"
+}
+```
+
+Then **enable the daemon service** in background (Rule 4):
+
+```bash
+jq '.services["fires-watcher"].enabled = true' \
+ ~/.claude/plugins/data/ops-ops-marketplace/daemon-services.json \
+ > /tmp/daemon-services.json.new && \
+ mv /tmp/daemon-services.json.new ~/.claude/plugins/data/ops-ops-marketplace/daemon-services.json && \
+ scripts/ops-daemon.sh restart &
+```
+
+Confirm the watcher is alive after restart:
+
+```bash
+cat ~/.claude/plugins/data/ops-ops-marketplace/fires-watcher.health 2>/dev/null
+```
+
+Expect `{"status": "ok", ...}` within 60 seconds.
+
+> **Deep-dive:** see `${CLAUDE_PLUGIN_ROOT}/docs/notifications.md`, `${CLAUDE_PLUGIN_ROOT}/scripts/ops-fires-watcher.sh`, and `${CLAUDE_PLUGIN_ROOT}/scripts/ops-notify.sh` for sink priority rationale, debounce rules, and a troubleshooting walk-through.
+
+---
+
## Step 4 — Configure MCPs (if selected)
For each MCP that isn't in `mcp_configured`, offer bulk setup first: