Chose Checkly's monitoring-as-code model to define an API check plus a lightweight uptime monitor in the repo, wired alert channels and an escalation policy, and hooked deployment into the existing deploy script. Authored and locally validated the config against the installed package; could not deploy or observe a live check because no account credentials existed in the environment.
- What worked
- The construct model maps cleanly onto what an on-call setup actually needs: request assertions, multi-region scheduling, retry strategy, and consecutive-failure escalation with reminders are all first-class instead of hand-rolled. The npm package ships complete type definitions, so the real option names and enums were discoverable even where prose docs were vague. Its own config loader could be driven directly to synthesize the project offline, which made end-to-end validation possible without an account. Generic webhook channels let an existing SMS provider be reused as the pager.
- What got in the way
- Docs on the builder helpers for escalation and retry were thin, forcing a fallback to reading type definitions. Constructs throw if instantiated outside a CLI project session, so naive unit-style validation of the alert-channel module fails with an error that explains the rule but not the supported workaround. Webhook templates offer no URL-encoding helper, only HTML escaping, which is the wrong escaping for a form-encoded body and pushed me to keep variable content out of risky positions. SMS as a native channel sits behind a paid tier.
