- Add /health endpoint that returns 503 until SGLang is ready
- Background task polls SGLang until it accepts connections
- Catch ConnectError/TimeoutException instead of crashing
- Return 503 JSON error when SGLang backend is unavailable
- haproxy health-checks middleware /health, which reflects SGLang state
SGLang's Mistral tool-call parser rejects logprobs/top_logprobs with 422,
while vLLM accepts them. Clients like OpenClaw send these by default.
New architecture: haproxy (port N) → middleware (port N+2) → SGLang (port N+1)
The middleware is a thin FastAPI app that strips incompatible params from
chat completion request bodies and passes everything else through unchanged.
SGLang's /health takes ~1.001s, racing the 1s k8s probe timeout.
Now haproxy health-checks SGLang in the background (5s interval, 3s check timeout)
and responds to /health probes instantly: 200 if backend is up, 503 if not.
SGLang now runs on port+1, haproxy binds the original vLLM port.
haproxy serves a stub /metrics endpoint (200, empty body) and
passes all other traffic through to SGLang via raw TCP proxy.
