feat: optional vision flag in OpenAI LLM config

packages/api/src/endpoints/openai/config.ts

@@ -97,6 +97,7 @@ export function getOpenAIConfig(
 ): t.OpenAIConfigResult {
   const {
     proxy,
+    vision,
     addParams,
     dropParams,
     defaultQuery,
@@ -202,6 +203,7 @@ export function getOpenAIConfig(
       defaultParams,
       modelOptions,
       useOpenRouter,
+      vision,
       reasoningFormat: getReasoningFormat({
         customFormat: options.customParams?.reasoningFormat,
         isVercel: Boolean(isVercel),

packages/api/src/endpoints/openai/llm.spec.ts

@@ -27,6 +27,38 @@ describe('getOpenAILLMConfig', () => {
       expect(result.tools).toEqual([]);
     });
 
+    it('should forward vision=false to llmConfig', () => {
+      const result = getOpenAILLMConfig({
+        apiKey: 'test-api-key',
+        streaming: true,
+        vision: false,
+        modelOptions: { model: 'gpt-4' },
+      });
+
+      expect(result.llmConfig).toHaveProperty('vision', false);
+    });
+
+    it('should forward vision=true to llmConfig', () => {
+      const result = getOpenAILLMConfig({
+        apiKey: 'test-api-key',
+        streaming: true,
+        vision: true,
+        modelOptions: { model: 'gpt-4' },
+      });
+
+      expect(result.llmConfig).toHaveProperty('vision', true);
+    });
+
+    it('should omit vision from llmConfig when not provided', () => {
+      const result = getOpenAILLMConfig({
+        apiKey: 'test-api-key',
+        streaming: true,
+        modelOptions: { model: 'gpt-4' },
+      });
+
+      expect(result.llmConfig).not.toHaveProperty('vision');
+    });
+
     it('should handle model options including temperature and penalties', () => {
       const result = getOpenAILLMConfig({
         apiKey: 'test-api-key',

packages/api/src/endpoints/openai/llm.ts

@@ -508,6 +508,7 @@ export function getOpenAILLMConfig({
   dropParams,
   defaultParams,
   useOpenRouter,
+  vision,
   reasoningFormat = ReasoningParameterFormat.reasoningEffort,
   modelOptions: _modelOptions,
 }: {
@@ -520,6 +521,7 @@ export function getOpenAILLMConfig({
   dropParams?: string[];
   defaultParams?: Record<string, unknown>;
   useOpenRouter?: boolean;
+  vision?: boolean;
   reasoningFormat?: ReasoningParameterFormat;
   azure?: false | t.AzureOptions;
 }): Pick<t.LLMConfigResult, 'llmConfig' | 'tools'> & {
@@ -555,6 +557,10 @@ export function getOpenAILLMConfig({
     modelOptions,
   ) as OpenAILLMConfig;
 
+  if (vision != null) {
+    llmConfig.vision = vision;
+  }
+
   if (frequency_penalty != null) {
     llmConfig.frequencyPenalty = frequency_penalty;
   }

packages/api/src/types/openai.ts

@@ -14,6 +14,8 @@ export type OpenAIModelOptions = Partial<OpenAIParameters>;
  */
 export interface OpenAIConfigOptions {
   modelOptions?: OpenAIModelOptions;
+  /** When false, the client strips image content for models without vision support. */
+  vision?: boolean;
   directEndpoint?: boolean;
   reverseProxyUrl?: string | null;
   baseURLIsUserProvided?: boolean;
@@ -34,6 +36,8 @@ export type OAIClientOptions = Omit<OpenAIClientOptions, 'verbosity'> & {
   include_reasoning?: boolean;
   /** Replays `reasoning_content` on tool-bearing turns (DeepSeek thinking-mode, #13366). */
   includeReasoningContent?: boolean;
+  /** When false, the client strips image content for models without vision support. */
+  vision?: boolean;
   promptCache?: boolean;
   promptCacheTtl?: '5m' | '1h';
   _lc_stream_delay?: number;