LM Studio: send completions to /v1, keep discovery on the raw address

embabel-agent-autoconfigure/models/embabel-agent-lmstudio-autoconfigure/src/main/kotlin/com/embabel/agent/config/models/lmstudio/LmStudioModelsConfig.kt

@@ -45,7 +45,7 @@ class LmStudioProperties : RetryProperties {
     /**
      * Base URL for LM Studio endpoint
      */
-    var baseUrl: String = "http://127.0.0.1:1234"
+    var baseUrl: String = DEFAULT_BASE_URL
 
     /**
      * API key for LM Studio. Defaults to null as apiKey isn't supported yet.
@@ -74,7 +74,10 @@ class LmStudioProperties : RetryProperties {
 
     override val propertyPrefix: String = PREFIX
     companion object {
-        const val PREFIX  = "embabel.agent.platform.models.lmstudio"
+        const val PREFIX = "embabel.agent.platform.models.lmstudio"
+        const val DEFAULT_HOST = "localhost"
+        const val DEFAULT_PORT = 1234
+        const val DEFAULT_BASE_URL = "http://$DEFAULT_HOST:$DEFAULT_PORT"
     }
 }
 
@@ -94,7 +97,15 @@ class LmStudioModelsConfig(
     @Qualifier("aiModelWebClientBuilder")
     webClientBuilder: ObjectProvider<WebClient.Builder>,
 ) : OpenAiCompatibleModelFactory(
-    baseUrl = lmStudioProperties.baseUrl,
+    // The SDK appends `/chat/completions` to whatever it is given — the factory's
+    // own docs say to bake the full path into the base URL — while discovery
+    // below appends `/v1/models` to the SAME property. Both cannot be satisfied
+    // by one configured value, and the mismatch is silent: models are listed
+    // (discovery is right) and then every completion 404s with LM Studio's
+    // "Unexpected endpoint or method", which surfaces to the caller as the
+    // baffling "`choices` is not set". Normalise here so the configured value
+    // stays the plain server address a user would write.
+    baseUrl = chatBaseUrl(lmStudioProperties.baseUrl),
     apiKey = lmStudioProperties.apiKey,
     completionsPath = null,
     embeddingsPath = null,
@@ -105,6 +116,18 @@ class LmStudioModelsConfig(
 
     private val log = LoggerFactory.getLogger(LmStudioModelsConfig::class.java)
 
+    companion object {
+        /**
+         * The URL the OpenAI SDK should treat as its API root: LM Studio serves
+         * the OpenAI surface under `/v1`, so append it unless the operator has
+         * already done so.
+         */
+        internal fun chatBaseUrl(baseUrl: String): String {
+            val clean = baseUrl.trimEnd('/')
+            return if (clean.endsWith("/v1")) clean else "$clean/v1"
+        }
+    }
+
     // OpenAI-compatible models response
     @JsonIgnoreProperties(ignoreUnknown = true)
     private data class ModelResponse(
@@ -197,7 +220,11 @@ class LmStudioModelsConfig(
                 .requestFactory(requestFactory)
                 .build()
 
-            val cleanBaseUrl = baseUrl?.trimEnd('/') ?: "http://127.0.0.1:1234"
+            // The RAW configured address, deliberately — NOT the inherited `baseUrl`,
+            // which is normalised to end in /v1 for the SDK. Reading that here
+            // built `…/v1/api/v1/models`, discovered nothing, and left the
+            // appliance unable to start with a local default model.
+            val cleanBaseUrl = lmStudioProperties.baseUrl.trimEnd('/')
             val apiUrl = if (cleanBaseUrl.contains("/api")) {
                 cleanBaseUrl
             } else {
@@ -230,7 +257,7 @@ class LmStudioModelsConfig(
 
             response.models?: emptyList()
         } catch (e: Exception) {
-            log.warn("Failed to load models from {}: {}", baseUrl, e.message)
+            log.warn("Failed to load models from {}: {}", lmStudioProperties.baseUrl, e.message)
             emptyList()
         }
     }

embabel-agent-autoconfigure/models/embabel-agent-lmstudio-autoconfigure/src/test/kotlin/com/embabel/agent/config/models/lmstudio/LmStudioModelsConfigTest.kt

@@ -16,6 +16,7 @@
 package com.embabel.agent.config.models.lmstudio
 
 import io.micrometer.observation.ObservationRegistry
+import kotlin.test.assertEquals
 import io.mockk.*
 import org.junit.jupiter.api.AfterEach
 import org.junit.jupiter.api.BeforeEach
@@ -191,4 +192,26 @@ class LmStudioModelsConfigTest {
         }
     }
 
+
+    // -----------------------------------------------------------------------
+    // The chat base URL. Discovery appends `/v1/models` to the configured value
+    // while the OpenAI SDK appends `/chat/completions` to it, so one of the two
+    // has to normalise — and getting this wrong is silent: models list fine and
+    // every completion 404s as "`choices` is not set".
+    // -----------------------------------------------------------------------
+
+    @Test
+    fun `chat base url gains the v1 the SDK expects`() {
+        assertEquals("${LmStudioProperties.DEFAULT_BASE_URL}/v1", LmStudioModelsConfig.chatBaseUrl(LmStudioProperties.DEFAULT_BASE_URL))
+    }
+
+    @Test
+    fun `a trailing slash does not produce a double slash`() {
+        assertEquals("${LmStudioProperties.DEFAULT_BASE_URL}/v1", LmStudioModelsConfig.chatBaseUrl("${LmStudioProperties.DEFAULT_BASE_URL}/"))
+    }
+
+    @Test
+    fun `an operator who already wrote v1 is not given two`() {
+        assertEquals("http://host.docker.internal:${LmStudioProperties.DEFAULT_PORT}/v1", LmStudioModelsConfig.chatBaseUrl("http://host.docker.internal:${LmStudioProperties.DEFAULT_PORT}/v1"))
+    }
 }