diff --git a/apps/server/src/main/kotlin/com/correx/apps/server/Main.kt b/apps/server/src/main/kotlin/com/correx/apps/server/Main.kt
index 97ba6a80..e65e3b87 100644
--- a/apps/server/src/main/kotlin/com/correx/apps/server/Main.kt
+++ b/apps/server/src/main/kotlin/com/correx/apps/server/Main.kt
@@ -441,10 +441,15 @@ fun main() {
             val seeded = DefaultEventReplayer(eventStore, com.correx.apps.server.health.HealthProjection())
                 .rebuild(com.correx.apps.server.health.SYSTEM_SESSION)
                 .subjects.mapValues { it.value.status }
-            // Monitor the llama server only when a llamacpp provider URL is configured — otherwise the
-            // LLAMA_SERVER subject would report a spurious DEGRADED on a box that runs no local model.
+            // Monitor the llama server only when one is configured — otherwise the LLAMA_SERVER subject
+            // would report a spurious DEGRADED on a box that runs no local model. Covers BOTH paths: an
+            // explicit static [[providers]] url, or the managed [[models]] server at [models].host:port.
             // A dedicated short-timeout client keeps a hung server from stalling the probe loop.
-            val llamaProbe = correxConfig.providers.firstOrNull { it.type == "llamacpp" }?.url?.let { url ->
+            val llamaUrl = correxConfig.providers.firstOrNull { it.type == "llamacpp" }?.url
+                ?: correxConfig.models.firstOrNull()?.let {
+                    "http://${correxConfig.modelsSettings.host}:${correxConfig.modelsSettings.port}"
+                }
+            val llamaProbe = llamaUrl?.let { url ->
                 val healthHttpClient = io.ktor.client.HttpClient(io.ktor.client.engine.cio.CIO) {
                     engine { requestTimeout = LLAMA_HEALTH_TIMEOUT_MS }
                 }