Validate model availability and configuration in the startup probe
We currently have configuration that tells the application where to reach Vertex and Anthropic.
The startup probe currently uses the /monitoring/healthz
endpoint which does not include any probes. This means that an AI-gateway instance that cannot reach Vertex or Anthropic would be considered ready, while it shouldn't.
To fix this, we should decouple the startup check (readiness) from the health checks (liveness), because we don't want to send a bogus prompt to the providers every time we perform a health check. If a model becomes unavailable, we should catch it through our SLIs (#441 (closed))
This is using fastapi-health
for a list of probes. We should create a new /monitoring/ready
endpoint that includes checks for all models.