linecanary-monitor

LineCanary Monitor

Use this skill when the user cares about a phone line staying healthy: an IVR menu, an AI receptionist, a front-desk line — anything customers dial.

It drives the runnable linecanary app, which places at most one CALL-E call per check per invocation, validates the structured result against operator-written assertions, compares timing and answers against the line's own history, and exits 0/1/2 for automation.

When to use

• "Is our phone line / voice agent still working?" — run the checks live and interpret the report.
• "Watch this line" / "monitor our IVR" — set up config, verification and a host schedule (cron or GitHub Actions; the app never self-schedules).
• "Did the voice-agent deploy break anything?" — run the smoke check (--only <check-id>) after a deploy, gate on the exit code.
• "Why did the canary page?" — read the JSON report and the baseline history, explain the regression kinds in plain words.

When not to use

• The line belongs to someone else and the user cannot verify ownership or produce a written authorization. LineCanary refuses unverified lines; do not help work around that — it is the product's compliance boundary.
• The user wants outbound calls to customers, leads or arbitrary businesses. That is not monitoring; decline and point at the safety notes.
• Sub-minute check frequency or bulk parallel probing. See references/safety.md — keep schedules proportionate (15–60 minutes is the intended shape).

How it works

1. Config as code: linecanary.config.json declares lines (with an ownership block), checks (task + strict resultSchema + assertions + timing bounds + confidence floor) and alerting. Full semantics in references/config-reference.md.
2. Ownership verification: a greeting_code line is verified by one call that must hear the operator's code in the line's own greeting; client lines under written authority use attestation. Verification is pinned to the phone number — a changed number re-verifies.
3. Every run is dry-run by default and prints the plan without dialing. --live places the calls, evaluates, diffs against the baseline history and appends to it. Every call opens with an AI disclosure.
4. Exit codes: 0 healthy · 1 regressions or failing checks · 2 the run itself broke (config, credentials, API). Treat 1 as "page a human", 2 as "the monitoring is broken, not the line".