paging

name	paging
description	Routes HIGH and CRITICAL audit-trail events to an operator-configured webhook (PagerDuty, Opsgenie, Slack, generic). Reads new rows from state/audit.jsonl since the last cursor and POSTs a generic event payload. Use when the developer asks about on-call paging, alert routing, webhook setup, audit-event escalation, or wants to verify F-011 closure. Auto-triggers on: "page on-call", "wire pager", "paging config", "alert routing", "F-011", "send HIGH events to", "configure webhook". Do not use for crafting the audit log itself (see audit-awareness) or for verifying chain integrity (see audit-verify).
allowed-tools	["Read","Bash","Edit"]

Operator-facing skill for the audit-trail pager. Helps wire the webhook, verify config, run the pager manually, and inspect undelivered events. - state/audit.jsonl exists (created on first PostToolUse event). - Operator has a JSON-POST webhook endpoint ready (PagerDuty, Opsgenie, Slack, or generic).

Step 1 — Initial setup

cd ${CLAUDE_PLUGIN_ROOT}
cp state/paging-config.example.json state/paging-config.json
# Edit state/paging-config.json:
#   - enabled: true
#   - webhook_url: <your endpoint>
#   - min_severity: HIGH (or CRITICAL for stricter routing)

paging-config.json is gitignored — keep your webhook URL local.

Step 2 — Smoke test (dry-run)

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/pager.py --dry-run

Expected output: JSON summary with dry_run: true. No webhook POST. The cursor advances even on dry-run — re-run will report scanned: 0 unless new audit rows have landed.

Step 3 — Live run

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/pager.py

Output JSON includes delivered, undelivered, rate_limited. If undelivered > 0, the rows are queued in state/paging-undelivered.jsonl with a retry counter.

Step 4 — Retry queued failures

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/pager.py --retry

Successfully retried events are dropped from the queue. Events past max_retries are dropped and logged as exhausted in state/paging-runs.log.

Step 5 — Periodic runner

Wire pager.py into a cron/systemd timer. Suggested cadence:

Time window	Cadence
Business hours	every 30s
Off-hours	every 5m
Retry sweep	every 15m, post --retry

<config_reference>

Field	Default	Meaning
enabled	false	Master switch. False = classify only, never POST.
webhook_url	""	JSON-POST endpoint.
min_severity	"HIGH"	Floor: DEBUG < INFO < LOW < MEDIUM < HIGH < CRITICAL.
rate_limit_per_min	5	Token bucket. Suppressed pages do NOT enter the retry queue.
timeout_seconds	5	Per-POST timeout.
max_retries	3	Drop undelivered after this many retries.

</config_reference>

<failure_modes>

• F08 tool-mis-invocation: don't shell out to curl — pager.py uses urllib so it ships stdlib-only. If you want curl, write a wrapper, but keep the canonical entry point as pager.py.
• F10 destructive without confirmation: pager.py never deletes audit.jsonl rows. The cursor is the only forward-moving state.
• F09 parallel race: do not run two pager.py instances concurrently. The cursor write is atomic but two runners will both POST the same rows. Cron-style serialization is sufficient — one runner per minute.

</failure_modes>

Advisory only. Paging is observability, not a gate. Even when the webhook is unreachable, pager.py exits 0 and writes to `state/paging-undelivered.jsonl`. F-011 closure: HIGH+ events are routed to operator-configured paging. Operator wires the actual endpoint; this skill ships the routing.