fix: Supervise both children instead of gating startup on the model

Dockerfile

@@ -41,7 +41,7 @@ RUN chmod +x /entrypoint.sh
 
 ENV PORT=8090
 
-HEALTHCHECK --interval=5s --timeout=3s --start-period=60s \
+HEALTHCHECK --interval=5s --timeout=3s --start-period=180s \
   CMD curl -sf http://127.0.0.1:${PORT:-8090}/health || exit 1
 
 EXPOSE 8090

README.md

@@ -42,6 +42,8 @@ SDK requests hit the localaik proxy, which translates Gemini, OpenAI, or Anthrop
 docker run -d -p 8090:8090 gokhalh/localaik
 ```
 
+The port starts accepting connections before the model has loaded, so wait for `GET /health` to return 200 rather than for the port to open. A TCP liveness check will let requests through too early. Docker Compose users want `condition: service_healthy`.
+
 Or with Docker Compose:
 
 ```yaml
@@ -207,8 +209,9 @@ clients send to localaik are still discarded and never forwarded. It is attached
 only to requests whose host matches `LK_UPSTREAM`, and while it is set a
 redirect from your upstream is returned to the caller rather than followed.
 
-`/health` returns 503 until your upstream answers, so existing healthchecks and
-CI wait loops work unchanged.
+`/health` returns 503 until your upstream answers, so any wait loop that polls it
+for a 200 works unchanged. As with the model-bundled tags, the port opens before
+the upstream is reachable, so a TCP liveness check is not enough.
 
 ### Security
 
@@ -406,5 +409,6 @@ docker build --target proxy -t gokhalh/localaik:proxy .
 - Intended for tests and development, not production
 - Image size is dominated by model weights (not applicable to the `proxy` tag, which ships none)
 - Cold starts can take tens of seconds while the model loads
+- `/health` reports 503 during that window, and the container keeps running rather than exiting if the model never finishes loading
 - PDF rendering adds latency per page
 

entrypoint.sh

@@ -1,13 +1,6 @@
 #!/bin/sh
 set -eu
 
-cleanup() {
-  jobs -p | xargs -r kill 2>/dev/null || true
-  wait || true
-}
-
-trap 'cleanup; exit 0' INT TERM
-
 LLAMA_SERVER_BIN="${LLAMA_SERVER_BIN:-/app/llama-server}"
 if [ ! -x "${LLAMA_SERVER_BIN}" ]; then
   if command -v llama-server >/dev/null 2>&1; then
@@ -35,23 +28,45 @@ SERVER_ARGS="${SERVER_ARGS} --ctx-size ${LK_CTX_SIZE:-8192}"
 [ "${LK_CONT_BATCHING:-0}" = "1" ] && SERVER_ARGS="${SERVER_ARGS} --cont-batching"
 [ "${LK_MLOCK:-0}" = "1" ]         && SERVER_ARGS="${SERVER_ARGS} --mlock"
 
+LLAMA_PID=""
+PROXY_PID=""
+
+# Armed before either child exists, and iterating so an unset PID cannot make
+# kill swallow the other one.
+trap 'for p in ${LLAMA_PID} ${PROXY_PID}; do kill "${p}" 2>/dev/null || true; done; exit 0' INT TERM
+
 # shellcheck disable=SC2086
 "${LLAMA_SERVER_BIN}" ${SERVER_ARGS} &
+LLAMA_PID=$!
 
-echo "localaik: loading model..."
-tries=0
-until curl -sf http://127.0.0.1:8080/health >/dev/null 2>&1; do
-  tries=$((tries + 1))
-  if [ "${tries}" -ge 120 ]; then
-    echo "localaik: model failed to load after 120s" >&2
-    cleanup
-    exit 1
-  fi
+localaik --port "${PORT:-8090}" --upstream "http://127.0.0.1:8080/v1" &
+PROXY_PID=$!
+
+echo "localaik: supervising; /health reports 503 until the model is ready"
+
+while kill -0 "${LLAMA_PID}" 2>/dev/null && kill -0 "${PROXY_PID}" 2>/dev/null; do
   sleep 1
 done
 
-echo "localaik: model ready"
-echo "localaik: listening on port ${PORT:-8090}"
-exec localaik \
-  --port "${PORT:-8090}" \
-  --upstream "http://127.0.0.1:8080/v1"
+if kill -0 "${LLAMA_PID}" 2>/dev/null; then
+  died="localaik"
+  dead_pid="${PROXY_PID}"
+  survivor="${LLAMA_PID}"
+else
+  died="llama-server"
+  dead_pid="${LLAMA_PID}"
+  survivor="${PROXY_PID}"
+fi
+
+kill "${survivor}" 2>/dev/null || true
+# Waited so its last log lines reach docker logs before the container tears down.
+wait "${survivor}" 2>/dev/null || true
+
+status=0
+wait "${dead_pid}" || status=$?
+if [ "${status}" -eq 0 ]; then
+  status=1
+fi
+
+echo "localaik: ${died} exited with status ${status}, stopping the container" >&2
+exit "${status}"