A returned 200 is the proof

The chapter ends when a public request returns a 200. The first probe is the liveness stub: curl -fsS https://echo-portal.fly.dev/health. It hits the :health route in PortalWeb.Router — get("/health", PortalWeb.CourseController, :health) — whose action is one line: def health(conn, _params), do: send_resp(conn, 200, "ok"). The response is a 200 with the literal body ok and no domain call; it reaches no context, no store, no engine. That is deliberate — the stub proves the endpoint is up and answering HTTP, and it answers even while a domain dependency is still settling, so it isolates “the node accepts connections” from “the application works.”

The second probe is the route smoke: curl -fsS https://echo-portal.fly.dev/courses. It returns a 200 and the live catalog markup — proving the endpoint serves the application, not only the liveness stub. A stub that answers while every real route 500s is a half-truth; the smoke closes that gap by exercising a path that reaches the running catalog. Two probes, two proofs: /health says “the node is up,” /courses says “the application serves.”

Select a probe below to read what it returns, what it proves, and why both are worth running.

An operator probes the endpoint with a bare curl — no cookie jar, no token, no login step. That works because PortalWeb.Router declares get("/health", PortalWeb.CourseController, :health) outside the :browser pipeline (the router’s own note records the route as the retained domain-free liveness probe). A route inside :browser runs the session plug and the CSRF protection, so a write to it would need a token minted from a prior page load; a route outside that pipeline carries neither. The liveness probe is deliberately placed where a credential-free curl reaches it.

That is why /health — not a page route — is the liveness probe. A page that lives under the :browser pipeline would still render for an operator’s GET, but coupling the liveness contract to a session-bearing surface invites a future CSRF or auth change to silently break the probe. The stub stays small, domain-free, and pipeline-free on purpose: it answers the one question a probe must answer — is the node up? — with the fewest moving parts that can fail.

Select a route placement below to read whether the operator’s bare curl reaches it, and which plugs sit in the way.

The -fsS idiom is what turns a curl into a gate. -f (--fail) makes curl exit non-zero on any non-2xx response — and on a connection failure — instead of printing an error page with exit 0. -s (--silent) hides the progress meter, and -S (--show-error) keeps a real error message visible even while silent. Put together: the command stays quiet on success, prints a genuine error on failure, and — the load-bearing part — its exit code reports the result.

That makes the gate a shell one-liner whose branch is decided by the exit status, not by parsing output:

Why this is the gate and not a config inspection: a configuration that “looks right” never fails a build — a wrong value fails a boot. The release could carry a host typo, a stale secret, or an adapter that cannot reach its database, and every one of those reads as plausible text on disk. Only a live 200 returned over the public internet distinguishes “deployed correctly” from “deployed wrong,” because only the boot — not the file — exercises the values against reality. The exit code is the one signal that has actually been across the network and back.