LlmHealthIndicator

Reports what was actually checked. By default (llm4s.health.probe=false) it makes no provider call: it is UP with probe=disabled and the provider and model names, meaning "configured", not "reachable". With llm4s.health.probe=true it sends a one-token completion on executor, waits at most llm4s.health.probe-timeout, and caches the outcome for llm4s.health.probe-ttl. A failed or timed-out probe is DOWN with a redacted message.