google/gemini provider cleanup. replaced 3.0 pro preview with 3.1 pro preview. fixed issues with 2.5 models

AlexanderLuck · AlexanderLuck · commit f5a2a167843d · 2026-02-27T14:43:50.000+01:00
diff --git a/src/main/java/ee/carlrobert/codegpt/completions/CompletionRequestService.java b/src/main/java/ee/carlrobert/codegpt/completions/CompletionRequestService.java
@@ -1,5 +1,7 @@
 package ee.carlrobert.codegpt.completions;
 
+import com.fasterxml.jackson.core.JsonProcessingException;
+import com.fasterxml.jackson.databind.node.ObjectNode;
 import com.intellij.openapi.application.ApplicationManager;
 import com.intellij.openapi.components.Service;
 import ee.carlrobert.codegpt.completions.factory.CustomOpenAIRequest;
@@ -12,7 +14,11 @@
 import ee.carlrobert.llm.client.anthropic.completion.ClaudeCompletionRequest;
 import ee.carlrobert.llm.client.codegpt.request.InlineEditRequest;
 import ee.carlrobert.llm.client.codegpt.request.chat.ChatCompletionRequest;
+import ee.carlrobert.llm.client.google.completion.ApiResponseError;
 import ee.carlrobert.llm.client.google.completion.GoogleCompletionRequest;
+import ee.carlrobert.llm.client.google.completion.GoogleCompletionResponse;
+import ee.carlrobert.llm.client.google.completion.GoogleContentPart;
+import ee.carlrobert.llm.client.google.models.GoogleModel;
 import ee.carlrobert.llm.client.openai.completion.ErrorDetails;
 import ee.carlrobert.llm.client.openai.completion.OpenAIChatCompletionEventSourceListener;
 import ee.carlrobert.llm.client.openai.completion.OpenAITextCompletionEventSourceListener;
@@ -21,6 +27,7 @@
 import ee.carlrobert.llm.client.openai.completion.response.OpenAIChatCompletionResponseChoice;
 import ee.carlrobert.llm.client.openai.completion.response.OpenAIChatCompletionResponseChoiceDelta;
 import ee.carlrobert.llm.completion.CompletionEventListener;
+import ee.carlrobert.llm.completion.CompletionEventSourceListener;
 import ee.carlrobert.llm.completion.CompletionRequest;
 import java.io.IOException;
 import java.util.Collection;
@@ -30,6 +37,8 @@
 import java.util.stream.Stream;
 import okhttp3.Call;
 import okhttp3.Callback;
+import okhttp3.HttpUrl;
+import okhttp3.MediaType;
 import okhttp3.Request;
 import okhttp3.RequestBody;
 import okhttp3.Response;
@@ -41,6 +50,10 @@
 @Service
 public final class CompletionRequestService {
 
+  private static final String GOOGLE_BASE_URL =
+      "https://generativelanguage.googleapis.com";
+  private static final MediaType JSON_MEDIA_TYPE = MediaType.parse("application/json");
+
   private CompletionRequestService() {
   }
 
@@ -240,10 +253,12 @@ public EventSource getChatCompletionAsync(
           eventListener);
     }
     if (request instanceof GoogleCompletionRequest completionRequest) {
+      var model = ModelSelectionService.getInstance().getModelForFeature(featureType, null);
+      if (model != null && GoogleModel.findByCode(model) == null) {
+        return getGoogleNonEnumModelCompletionAsync(completionRequest, model, eventListener);
+      }
       return CompletionClientProvider.getGoogleClient().getChatCompletionAsync(
-          completionRequest,
-          ModelSelectionService.getInstance().getModelForFeature(featureType, null),
-          eventListener);
+          completionRequest, model, eventListener);
     }
 
     throw new IllegalStateException("Unknown request type: " + request.getClass());
@@ -293,11 +308,14 @@ public String getChatCompletion(CompletionRequest request, ServiceType serviceTy
           .getText();
     }
     if (request instanceof GoogleCompletionRequest completionRequest) {
+      var model = ApplicationManager.getApplication()
+          .getService(ModelSelectionService.class)
+          .getModelForFeature(featureType, null);
+      if (model != null && GoogleModel.findByCode(model) == null) {
+        return getGoogleNonEnumModelCompletion(completionRequest, model);
+      }
       return CompletionClientProvider.getGoogleClient().getChatCompletion(
-              completionRequest,
-              ApplicationManager.getApplication()
-                  .getService(ModelSelectionService.class)
-                  .getModelForFeature(featureType, null))
+              completionRequest, model)
           .getCandidates().get(0)
           .getContent().getParts().get(0)
           .getText();
@@ -333,6 +351,117 @@ public static boolean isRequestAllowed(ServiceType serviceType) {
     };
   }
 
+  private EventSource getGoogleNonEnumModelCompletionAsync(
+      GoogleCompletionRequest request,
+      String model,
+      CompletionEventListener<String> eventListener) {
+    try {
+      var httpRequest = buildGoogleNonEnumRequest(model, "streamGenerateContent", request, true);
+      var httpClient = CompletionClientProvider.getDefaultClientBuilder().build();
+      return EventSources.createFactory(httpClient).newEventSource(
+          httpRequest, createGoogleEventSourceListener(eventListener));
+    } catch (JsonProcessingException e) {
+      throw new RuntimeException("Failed to serialize Google completion request", e);
+    }
+  }
+
+  private String getGoogleNonEnumModelCompletion(
+      GoogleCompletionRequest request,
+      String model) {
+    try {
+      var httpRequest = buildGoogleNonEnumRequest(model, "generateContent", request, false);
+      var httpClient = CompletionClientProvider.getDefaultClientBuilder().build();
+      try (var response = httpClient.newCall(httpRequest).execute()) {
+        return DeserializationUtil.mapResponse(response, GoogleCompletionResponse.class)
+            .getCandidates().get(0)
+            .getContent().getParts().get(0)
+            .getText();
+      }
+    } catch (IOException e) {
+      throw new RuntimeException("Failed to get Google completion", e);
+    }
+  }
+
+  private Request buildGoogleNonEnumRequest(
+      String model, String action, Object requestBody, boolean stream)
+      throws JsonProcessingException {
+    var apiKey = CredentialsStore.INSTANCE.getCredential(CredentialKey.GoogleApiKey.INSTANCE);
+    var urlBuilder = HttpUrl.parse(
+        GOOGLE_BASE_URL + "/v1beta/models/" + model + ":" + action).newBuilder();
+    if (apiKey != null && !apiKey.isEmpty()) {
+      urlBuilder.addQueryParameter("key", apiKey);
+    }
+    if (stream) {
+      urlBuilder.addQueryParameter("alt", "sse");
+    }
+
+    var mapper = DeserializationUtil.OBJECT_MAPPER;
+    var jsonNode = (ObjectNode) mapper.valueToTree(requestBody);
+
+    // Inject thinkingConfig for models that support thinking (3.x+)
+    var genConfig = jsonNode.has("generationConfig")
+        ? (ObjectNode) jsonNode.get("generationConfig")
+        : mapper.createObjectNode();
+    if (!genConfig.has("thinkingConfig")) {
+      var thinkingConfig = mapper.createObjectNode();
+      thinkingConfig.put("thinkingLevel", "low");
+      genConfig.set("thinkingConfig", thinkingConfig);
+    }
+    if (!jsonNode.has("generationConfig")) {
+      jsonNode.set("generationConfig", genConfig);
+    }
+
+    return new Request.Builder()
+        .url(urlBuilder.build())
+        .header("Cache-Control", "no-cache")
+        .header("Content-Type", "application/json")
+        .header("Accept", stream ? "text/event-stream" : "text/json")
+        .post(RequestBody.create(mapper.writeValueAsString(jsonNode), JSON_MEDIA_TYPE))
+        .build();
+  }
+
+  private CompletionEventSourceListener<String> createGoogleEventSourceListener(
+      CompletionEventListener<String> eventListener) {
+    return new CompletionEventSourceListener<>(eventListener) {
+      @Override
+      protected String getMessage(String data) {
+        try {
+          var candidates = DeserializationUtil.OBJECT_MAPPER
+              .readValue(data, GoogleCompletionResponse.class)
+              .getCandidates();
+          return (candidates == null
+              ? Stream.<GoogleCompletionResponse.Candidate>empty()
+              : candidates.stream())
+              .filter(Objects::nonNull)
+              .flatMap(candidate -> {
+                if (candidate.getContent() != null
+                    && candidate.getContent().getParts() != null) {
+                  return candidate.getContent().getParts().stream();
+                }
+                return Stream.empty();
+              })
+              .filter(Objects::nonNull)
+              .filter(part -> part.getThought() == null || !part.getThought())
+              .findFirst()
+              .map(GoogleContentPart::getText)
+              .orElse("");
+        } catch (JsonProcessingException e) {
+          // ignore
+        }
+        return "";
+      }
+
+      @Override
+      protected ErrorDetails getErrorDetails(String data) throws JsonProcessingException {
+        var googleError = DeserializationUtil.OBJECT_MAPPER
+            .readValue(data, ApiResponseError.class).getError();
+        return googleError == null ? null
+            : new ErrorDetails(googleError.getMessage(), googleError.getStatus(), null,
+            googleError.getCode());
+      }
+    };
+  }
+
   /**
    * Content of the first choice.
    * <ul>
diff --git a/src/main/kotlin/ee/carlrobert/codegpt/completions/factory/GoogleRequestFactory.kt b/src/main/kotlin/ee/carlrobert/codegpt/completions/factory/GoogleRequestFactory.kt
@@ -27,17 +27,30 @@ class GoogleRequestFactory : BaseRequestFactory() {
     override fun createChatRequest(params: ChatCompletionParameters): GoogleCompletionRequest {
         val configuration = service<ConfigurationSettings>().state
         val selectedModel = ModelSelectionService.getInstance().getModelForFeature(FeatureType.CHAT)
-        val messages = buildGoogleMessages(selectedModel, params)
+        val systemInstruction = buildSystemInstruction(params)
+
+        val messages = if (!systemInstruction.isNullOrBlank()) {
+            listOf(
+                GoogleCompletionContent("user", listOf(systemInstruction)),
+                GoogleCompletionContent("model", listOf("Understood."))
+            ) + buildGoogleMessages(selectedModel, params)
+        } else {
+            buildGoogleMessages(selectedModel, params)
+        }
+
         return GoogleCompletionRequest.Builder(messages)
             .generationConfig(
                 GoogleGenerationConfig.Builder()
                     .maxOutputTokens(configuration.maxTokens)
                     .temperature(configuration.temperature.toDouble()).build()
             )
-            .systemInstruction(buildSystemInstruction(params))
             .build()
     }
 
+    private fun isNonEnumModel(model: String?): Boolean {
+        return model != null && GoogleModel.findByCode(model) == null
+    }
+
     override fun createBasicCompletionRequest(
         systemPrompt: String,
         userPrompt: String,
@@ -104,6 +117,8 @@ class GoogleRequestFactory : BaseRequestFactory() {
                 break
             }
 
+            if (prevMessage.response.isNullOrBlank()) continue
+
             prevMessage.imageFilePath?.takeIf { it.isNotEmpty() }?.let { imagePath ->
                 try {
                     val imageData = Files.readAllBytes(Path.of(imagePath))
diff --git a/src/main/kotlin/ee/carlrobert/codegpt/settings/models/ModelRegistry.kt b/src/main/kotlin/ee/carlrobert/codegpt/settings/models/ModelRegistry.kt
@@ -5,7 +5,6 @@ import ai.koog.prompt.executor.clients.anthropic.AnthropicModels.Opus_4_5
 import ai.koog.prompt.executor.clients.anthropic.AnthropicModels.Sonnet_4_5
 import ai.koog.prompt.executor.clients.google.GoogleModels.Gemini2_5Flash
 import ai.koog.prompt.executor.clients.google.GoogleModels.Gemini2_5Pro
-import ai.koog.prompt.executor.clients.google.GoogleModels.Gemini3_Pro_Preview
 import ai.koog.prompt.executor.clients.mistralai.MistralAIModels.Chat.DevstralMedium
 import ai.koog.prompt.executor.clients.openai.OpenAIModels.Chat.GPT4_1
 import ai.koog.prompt.executor.clients.openai.OpenAIModels.Chat.GPT4_1Mini
@@ -458,7 +457,7 @@ class ModelRegistry {
                 )
             },
             ServiceType.GOOGLE to listOf(
-                LLMModelWrapper(Gemini3_Pro_Preview, name = "Gemini 3 Pro Preview"),
+                LLMModelWrapper(Gemini3_1_Pro_Preview, name = "Gemini 3.1 Pro Preview"),
                 LLMModelWrapper(Gemini3_Flash_Preview, name = "Gemini 3 Flash Preview"),
                 LLMModelWrapper(Gemini2_5Pro, name = "Gemini 2.5 Pro"),
                 LLMModelWrapper(Gemini2_5Flash, name = "Gemini 2.5 Flash")
@@ -564,8 +563,8 @@ class ModelRegistry {
         return listOf(
             ModelSelection(
                 ServiceType.GOOGLE,
-                Gemini3_Pro_Preview.id,
-                "Gemini 3 Pro Preview",
+                Gemini3_1_Pro_Preview.id,
+                "Gemini 3.1 Pro Preview",
                 Icons.Google
             ),
             ModelSelection(
@@ -831,6 +830,16 @@ class ModelRegistry {
 
     private fun getGoogleModels(): List<ModelSelection> {
         return listOf(
+            ModelSelection(
+                ServiceType.GOOGLE,
+                Gemini3_1_Pro_Preview.id,
+                "Gemini 3.1 Pro Preview"
+            ),
+            ModelSelection(
+                ServiceType.GOOGLE,
+                Gemini3_Flash_Preview.id,
+                "Gemini 3 Flash Preview"
+            ),
             ModelSelection(
                 ServiceType.GOOGLE,
                 GoogleModel.GEMINI_2_5_PRO_PREVIEW.code,
@@ -996,6 +1005,25 @@ class ModelRegistry {
     }
 }
 
+public val Gemini3_1_Pro_Preview: LLModel = LLModel(
+    provider = LLMProvider.Google,
+    id = "gemini-3.1-pro-preview",
+    capabilities = listOf(
+        LLMCapability.Temperature,
+        LLMCapability.Completion,
+        LLMCapability.MultipleChoices,
+        LLMCapability.Vision.Image,
+        LLMCapability.Vision.Video,
+        LLMCapability.Audio,
+        LLMCapability.Tools,
+        LLMCapability.ToolChoice,
+        LLMCapability.Schema.JSON.Basic,
+        LLMCapability.Schema.JSON.Standard,
+    ),
+    contextLength = 1_048_576,
+    maxOutputTokens = 65_536,
+)
+
 public val Gemini3_Flash_Preview: LLModel = LLModel(
     provider = LLMProvider.Google,
     id = "gemini-3-flash-preview",
