Self-host preflight

Before starting the first assessment on a new or changed deployment, run the API preflight diagnostic. It checks dependencies that the normal readiness probe does not gate: S3/MinIO, Flagship/flagd, runner gRPC, Polar billing when configured, and required API secrets.

docker compose run --rm api aisafe-api preflight

The command prints one line per check:

• PASS means the dependency responded correctly.
• FAIL means the deployment needs attention before running assessments.
• SKIP means the dependency is optional or checked by another service.

The process exits 0 only when every required check passes. Use that exit code as a deploy gate in CI or automation.

Organization administrators can also read named check results at GET /health/diagnostics through an authorized session. That response uses generic status descriptions. Run the command inside the deployment to see dependency details.

Host and container monitoring

The compose stack ships two exporters as ordinary containers with read-only access to the host: node_exporter (host memory, memory pressure, CPU, disks and kernel OOM kills) and cAdvisor (each container's memory, memory limit, CPU, restarts and OOM kills). The bundled Prometheus scrapes them and alerts when the host runs low on memory or disk or a container keeps hitting its memory limit. Both serve standard Prometheus metrics on the stack's internal network (node-exporter:9100 and cadvisor:8080), so your own monitoring can scrape them too.

Every service in the stack has a memory limit, so a service that grows out of bounds is restarted on its own instead of exhausting the host. Give the host at least the memory the stack was sized for (32 GB) and keep the limits in the compose file unless you have measured that a service needs more.