Editing Openai/694057b6-101c-8007-9a65-c40578c7252d (section)

===== When the worker starts llama-server, it needs a cheap, reliable way to decide “the server is ready to accept requests” (and later, “it’s still alive”). =====

So the missing decision is: what exact HTTP request should the worker use as its readiness/health probe?

Recommended default for llama-server:
* Ready check: GET /v1/models - If it returns HTTP 200 + parses as JSON, consider the worker READY.
* If that endpoint ever proves unreliable in your build: fall back to a tiny POST /v1/chat/completions with max_tokens=1 (warmup-style probe), but only if needed.

This probe is separate from “slow prefill” timeouts—health probes should answer “is the server reachable and responding at all?” not “is this particular request producing tokens yet?”

I’ll update the doc wording to make that explicit.