🔭 feat: Support Gemini 4 Argon Input and History Rules

src/llm/google/index.ts

@@ -16,14 +16,15 @@
 import type { BaseMessage, UsageMetadata } from '@langchain/core/messages';
 import type { GeminiApiUsageMetadata, InputTokenDetails } from './types';
 import type { GoogleClientOptions, GoogleThinkingConfig } from '@/types';
-import { smoothGenerationChunks } from '@/llm/stream/chunkAdapters';
-import { resolveStreamDelay } from '@/llm/stream/smoother';
 import {
   convertResponseContentToChatGenerationChunk,
   convertBaseMessagesToContent,
+  isGemini3PlusModel,
   dropUnsupportedModelTurnPrefill,
   mapGenerateContentResultToChatResult,
 } from './utils/common';
+import { smoothGenerationChunks } from '@/llm/stream/chunkAdapters';
+import { resolveStreamDelay } from '@/llm/stream/smoother';
 
 type GoogleToolConfigWithServerSideInvocations = ToolConfig & {
   includeServerSideToolInvocations?: boolean;
@@ -45,13 +46,14 @@
   /**
    * Override to add gemini-3 model support for multimodal and function calling thought signatures
    */
+  /** Gemini 3 and later accept image, audio, video and PDF input. */
   get _isMultimodalModel(): boolean {
     return (
       this.model.startsWith('gemini-1.5') ||
       this.model.startsWith('gemini-2') ||
       (this.model.startsWith('gemma-3-') &&
         !this.model.startsWith('gemma-3-1b')) ||
-      this.model.startsWith('gemini-3')
+      isGemini3PlusModel(this.model)
     );
   }
 
@@ -170,7 +172,7 @@
       total_tokens: usageMetadata.totalTokenCount ?? 0,
     };
 
     if (usageMetadata.cachedContentTokenCount) {
       output.input_token_details ??= {};
       output.input_token_details.cache_read =
         usageMetadata.cachedContentTokenCount;

src/llm/google/multimodal.test.ts

@@ -0,0 +1,25 @@
+import { describe, expect, test } from '@jest/globals';
+import { CustomChatGoogleGenerativeAI } from './index';
+
+const isMultimodal = (model: string): boolean =>
+  new CustomChatGoogleGenerativeAI({ model, apiKey: 'test-key' })
+    ._isMultimodalModel;
+
+describe('CustomChatGoogleGenerativeAI._isMultimodalModel', () => {
+  test('accepts Gemini 3 and later, including Gemini 4 Argon', () => {
+    expect(isMultimodal('gemini-3-pro-preview')).toBe(true);
+    expect(isMultimodal('gemini-3.8-flash')).toBe(true);
+    expect(isMultimodal('gemini-4-argon')).toBe(true);
+    expect(isMultimodal('models/gemini-4-argon')).toBe(true);
+    expect(isMultimodal('gemini-4-argon-cyber-permissive')).toBe(true);
+    expect(isMultimodal('gemini-10-pro')).toBe(true);
+  });
+
+  test('keeps the earlier generations and text-only models as they were', () => {
+    expect(isMultimodal('gemini-1.5-pro')).toBe(true);
+    expect(isMultimodal('gemini-2.5-flash')).toBe(true);
+    expect(isMultimodal('gemma-3-27b-it')).toBe(true);
+    expect(isMultimodal('gemma-3-1b-it')).toBe(false);
+    expect(isMultimodal('gemini-pro')).toBe(false);
+  });
+});

src/llm/google/utils/common.test.ts

@@ -4,20 +4,40 @@ import type {
   Content,
   EnhancedGenerateContentResponse,
 } from '@google/generative-ai';
-import {
-  STREAMED_TOOL_CALL_SEAL_METADATA_KEY,
-  STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY,
-  GOOGLE_STREAMED_TOOL_CALL_ADAPTER,
-} from '@/tools/streamedToolCallSeals';
 import {
   convertResponseContentToChatGenerationChunk,
   convertMessageContentToParts,
   dropUnsupportedModelTurnPrefill,
   rejectsModelTurnPrefill,
+  isGemini3PlusModel,
 } from './common';
+import {
+  STREAMED_TOOL_CALL_SEAL_METADATA_KEY,
+  STREAMED_TOOL_CALL_ADAPTER_METADATA_KEY,
+  GOOGLE_STREAMED_TOOL_CALL_ADAPTER,
+} from '@/tools/streamedToolCallSeals';
 import { toLangChainContent } from '@/messages/langchain';
 
 describe('convertMessageContentToParts cross-provider tool calls', () => {
+  test.each([
+    ['gemini-3.5-flash', true],
+    ['gemini-4-argon', true],
+    ['models/gemini-4-argon-cyber-permissive', true],
+    ['gemini-2.5-pro', false],
+  ])(
+    'adds a placeholder thought signature to unsigned tool calls on %s: %s',
+    (model, expected) => {
+      const message = new AIMessage({
+        content: '',
+        tool_calls: [{ id: 'call_weather', name: 'weather', args: {} }],
+      });
+
+      const [part] = convertMessageContentToParts(message, false, [], model);
+
+      expect('thoughtSignature' in part).toBe(expected);
+    }
+  );
+
   test('deduplicates Anthropic tool_use content mirrored in parsed tool_calls', () => {
     const message = new AIMessage({
       content: toLangChainContent([
@@ -147,6 +167,14 @@ describe('rejectsModelTurnPrefill', () => {
     expect(rejectsModelTurnPrefill('google/gemini-3.8-flash-lite')).toBe(true);
   });
 
+  test('covers Gemini 4 Argon and its variants', () => {
+    expect(rejectsModelTurnPrefill('gemini-4-argon')).toBe(true);
+    expect(rejectsModelTurnPrefill('models/gemini-4-argon')).toBe(true);
+    expect(rejectsModelTurnPrefill('gemini-4-argon-cyber-permissive')).toBe(
+      true
+    );
+  });
+
   test('splits the 3.5 generation, where only Flash-Lite rejects prefill', () => {
     expect(rejectsModelTurnPrefill('gemini-3.5-flash')).toBe(false);
     expect(rejectsModelTurnPrefill('gemini-3.5-flash-latest')).toBe(false);
@@ -227,3 +255,19 @@ describe('dropUnsupportedModelTurnPrefill', () => {
     ).toBeUndefined();
   });
 });
+
+describe('isGemini3PlusModel', () => {
+  test('matches Gemini 3 and later, with or without a prefix', () => {
+    expect(isGemini3PlusModel('gemini-3-pro-preview')).toBe(true);
+    expect(isGemini3PlusModel('gemini-4-argon')).toBe(true);
+    expect(isGemini3PlusModel('models/gemini-4-argon')).toBe(true);
+    expect(isGemini3PlusModel('google/gemini-10-pro')).toBe(true);
+  });
+
+  test('rejects earlier generations and missing ids', () => {
+    expect(isGemini3PlusModel('gemini-2.5-flash')).toBe(false);
+    expect(isGemini3PlusModel('gemini-pro')).toBe(false);
+    expect(isGemini3PlusModel('gemma-3-27b-it')).toBe(false);
+    expect(isGemini3PlusModel(undefined)).toBe(false);
+  });
+});

src/llm/google/utils/common.ts

@@ -104,6 +104,15 @@ function createGoogleFunctionResponsePart({
  */
 export const iife = <T>(fn: () => T): T => fn();
 
+/** `gemini-3`, `gemini-3.8-flash`, `gemini-4-argon`, and any later major version. */
+const GEMINI_3_PLUS_PATTERN = /^gemini-(?:[3-9]|[1-9]\d+)(?:[.-]|$)/;
+
+/** Whether `model` (optionally `models/` or provider prefixed) is Gemini 3 or later. */
+export function isGemini3PlusModel(model?: string): boolean {
+  const modelId = model?.toLowerCase().split('/').pop() ?? '';
+  return GEMINI_3_PLUS_PATTERN.test(modelId);
+}
+
 export function getMessageAuthor(message: BaseMessage): string {
   const type = message._getType();
   if (ChatMessage.isInstance(message)) {
@@ -540,7 +549,7 @@ export function convertMessageContentToParts(
             return signature;
           }
         }
-        if (model?.includes('gemini-3') === true) {
+        if (isGemini3PlusModel(model)) {
           return DUMMY_SIGNATURE;
         }
         return '';
@@ -681,12 +690,17 @@ const NO_PREFILL_FLASH_MIN_VERSION = { major: 3, minor: 6 } as const;
 const GEMINI_FLASH_VERSION_PATTERN = /^gemini-(\d+)(?:\.(\d+))?-flash(?:$|-)/;
 
 /**
- * Models that reject prefill despite predating
- * {@link NO_PREFILL_FLASH_MIN_VERSION}. Gemini 3.5 Flash-Lite enforces the
+ * Models that reject prefill outside the Flash version cutoff
+ * ({@link NO_PREFILL_FLASH_MIN_VERSION}). Gemini 3.5 Flash-Lite enforces the
  * restriction while its sibling Gemini 3.5 Flash still accepts a trailing model
- * turn, so the 3.5 generation cannot be expressed as a version cutoff.
+ * turn, so the 3.5 generation cannot be expressed as a version cutoff. Gemini 4
+ * Argon is not a Flash model; its histories cannot end in a `model` turn.
+ * @see https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-4-argon
  */
-const NO_PREFILL_GEMINI_MODELS = ['gemini-3.5-flash-lite'] as const;
+const NO_PREFILL_GEMINI_MODELS = [
+  'gemini-3.5-flash-lite',
+  'gemini-4-argon',
+] as const;
 
 export function rejectsModelTurnPrefill(model?: string): boolean {
   if (model == null || model === '') {