fix(openai): default splitToolMedia so tool-returned images reach strict OpenAI-compatible backends

packages/cli/src/config/settingsSchema.ts

@@ -1259,9 +1259,9 @@ const SETTINGS_SCHEMA = {
             label: 'Split Tool Result Media',
             category: 'Generation Configuration',
             requiresRestart: false,
-            default: false,
+            default: true,
             description:
-              'When true, media (images / audio / video / files) returned by MCP tool calls is split into a follow-up user message instead of being embedded in the tool message. Required for strict OpenAI-compatible servers (e.g., LM Studio) that reject non-text content on `role: "tool"` messages with HTTP 400 "Invalid \'messages\' in payload". Default false preserves the prior behavior for permissive providers. See QwenLM/qwen-code#3616.',
+              'When true, media (images / audio / video / files) returned by tool calls — including the built-in read_file and MCP tools — is split into a follow-up user message instead of being embedded in the `role: "tool"` message. The OpenAI Chat Completions spec only permits text on tool messages, so strict OpenAI-compatible servers (e.g., doubao / new-api / LM Studio) silently drop or reject embedded media and the model never sees an image read via read_file (QwenLM/qwen-code#4876, #3616). Default true is spec-compliant and safe for permissive providers; set false only to restore the legacy embed-in-tool-message behavior.',
             parentKey: 'generationConfig',
             showInDialog: false,
           },

packages/core/src/core/contentGenerator.ts

@@ -125,14 +125,15 @@ export type ContentGeneratorConfig = {
   // Supported input modalities. Unsupported media types are replaced with text
   // placeholders. Leave undefined to use automatic detection from model name.
   modalities?: InputModalities;
-  // When true, media parts in MCP tool responses are split into a follow-up
-  // `role: "user"` message instead of being embedded inside the `role: "tool"`
-  // message. The OpenAI Chat Completions spec only permits string / text-part
-  // content on tool messages; strict OpenAI-compatible servers (notably
-  // LM Studio) reject anything else with HTTP 400 "Invalid 'messages' in
-  // payload". Enable this for any provider that strictly validates tool
-  // message content. Default: false (preserves prior behavior for permissive
-  // providers). See QwenLM/qwen-code#3616.
+  // When true, media parts in tool responses (including the built-in read_file
+  // and MCP tools) are split into a follow-up `role: "user"` message instead of
+  // being embedded inside the `role: "tool"` message. The OpenAI Chat
+  // Completions spec only permits string / text-part content on tool messages;
+  // strict OpenAI-compatible servers (e.g. doubao / new-api / LM Studio) drop or
+  // reject anything else (HTTP 400 "Invalid 'messages' in payload"), so an image
+  // read via read_file never reaches the model. Default: true (spec-compliant
+  // and safe for permissive providers); set false to restore the legacy
+  // embed-in-tool-message behavior. See QwenLM/qwen-code#4876, #3616.
   splitToolMedia?: boolean;
 };
 

packages/core/src/core/loggingContentGenerator/loggingContentGenerator.ts

@@ -105,6 +105,7 @@ export class LoggingContentGenerator implements ContentGenerator {
   private openaiLogger?: OpenAILogger;
   private schemaCompliance?: 'auto' | 'openapi_30';
   private modalities?: InputModalities;
+  private splitToolMedia?: boolean;
   private readonly generatorAuthType: ContentGeneratorConfig['authType'];
 
   constructor(
@@ -113,6 +114,7 @@ export class LoggingContentGenerator implements ContentGenerator {
     generatorConfig: ContentGeneratorConfig,
   ) {
     this.modalities = generatorConfig.modalities;
+    this.splitToolMedia = generatorConfig.splitToolMedia;
     this.generatorAuthType = generatorConfig.authType;
 
     // Extract fields needed for initialization from passed config
@@ -736,6 +738,10 @@ export class LoggingContentGenerator implements ContentGenerator {
     return {
       model,
       modalities: this.modalities ?? {},
+      // Mirror the pipeline default (see pipeline.ts createRequestContext) so the
+      // --openai-logging fallback reconstruction reflects the same split as the
+      // request actually sent. Opt out via generationConfig.splitToolMedia = false.
+      splitToolMedia: this.splitToolMedia ?? true,
       startTime: 0,
     };
   }

packages/core/src/core/modalityDefaults.test.ts

@@ -218,6 +218,43 @@ describe('defaultModalities', () => {
     });
   });
 
+  describe('ByteDance Doubao', () => {
+    it('returns image for doubao-seed-2.0-pro (issue #4876)', () => {
+      const m = defaultModalities('doubao-seed-2.0-pro');
+      expect(m.image).toBe(true);
+      expect(m.video).toBeUndefined();
+      expect(m.audio).toBeUndefined();
+    });
+
+    it('returns image for doubao-seed-1.6', () => {
+      expect(defaultModalities('doubao-seed-1.6').image).toBe(true);
+    });
+
+    it('returns image for doubao-1.5-vision-pro', () => {
+      expect(defaultModalities('doubao-1.5-vision-pro').image).toBe(true);
+    });
+
+    it('returns image for doubao-vision', () => {
+      expect(defaultModalities('doubao-vision').image).toBe(true);
+    });
+
+    it('returns text-only for doubao-seedance (text→video generation model)', () => {
+      expect(defaultModalities('doubao-seedance-1.0-pro')).toEqual({});
+    });
+
+    it('returns text-only for doubao-seedream (text→image generation model)', () => {
+      expect(defaultModalities('doubao-seedream-3.0')).toEqual({});
+    });
+
+    it('returns text-only for doubao-pro-32k', () => {
+      expect(defaultModalities('doubao-pro-32k')).toEqual({});
+    });
+
+    it('returns text-only for doubao-lite-4k', () => {
+      expect(defaultModalities('doubao-lite-4k')).toEqual({});
+    });
+  });
+
   describe('unknown models', () => {
     it('returns text-only for unrecognized models', () => {
       expect(defaultModalities('some-random-model-xyz')).toEqual({});

packages/core/src/core/modalityDefaults.ts

@@ -79,6 +79,18 @@ const MODALITY_PATTERNS: Array<[RegExp, InputModalities]> = [
   // -------------------
   [/^kimi-k2\.5/, { image: true, video: true }],
   [/^kimi-/, {}],
+
+  // -------------------
+  // ByteDance Doubao — Seed-series and *-vision / *-vl models accept image
+  // input; other Doubao models (pro / lite / text) are text-only.
+  // (QwenLM/qwen-code#4876)
+  // -------------------
+  // seedance (text→video) and seedream (text→image) are generation models with
+  // text-only input — exclude them before the multimodal Seed chat series.
+  [/^doubao-seed(ance|ream)/, {}],
+  [/^doubao-seed/, { image: true }],
+  [/^doubao-.*(vision|vl)/, { image: true }],
+  [/^doubao/, {}],
 ];
 
 /**

packages/core/src/core/openaiContentGenerator/converter.test.ts

@@ -787,10 +787,11 @@ describe('OpenAIContentConverter', () => {
       expect(img?.image_url?.url).toBe('data:image/png;base64,xxx');
     });
 
-    it('should preserve prior embedded-media behavior when splitToolMedia is false (default) on parallel tool calls (issue #3616)', () => {
+    it('should preserve embedded-media behavior when splitToolMedia is explicitly false (opt-out) on parallel tool calls (issue #3616, #4876)', () => {
       // Same input as the parallel-tool-calls split test, but with the flag
-      // off. Asserts that the opt-in is actually opt-in: media stays embedded
-      // in the tool message and no follow-up user message is synthesised.
+      // explicitly off. Since #4876 the default is true (spec-compliant), so
+      // this asserts the opt-out path: media stays embedded in the tool
+      // message and no follow-up user message is synthesised.
       const request: GenerateContentParameters = {
         model: 'models/test',
         contents: [
@@ -826,11 +827,10 @@ describe('OpenAIContentConverter', () => {
         ],
       };
 
-      // requestContext default has splitToolMedia undefined / false
-      const messages = converter.convertGeminiRequestToOpenAI(
-        request,
-        requestContext,
-      );
+      const messages = converter.convertGeminiRequestToOpenAI(request, {
+        ...requestContext,
+        splitToolMedia: false,
+      });
 
       const toolMessages = messages.filter((m) => m.role === 'tool');
       const userMessages = messages.filter((m) => m.role === 'user');

packages/core/src/core/openaiContentGenerator/converter.ts

@@ -594,18 +594,19 @@ function processContent(
         requestContext,
       );
       if (toolMessage) {
-        // Opt-in only (ContentGeneratorConfig.splitToolMedia). OpenAI spec
-        // only permits string / text-part content on `role: "tool"` messages.
-        // Strict OpenAI-compatible servers (e.g. LM Studio) reject tool
-        // messages containing image_url / input_audio / video_url / file
-        // parts with HTTP 400 "Invalid 'messages' in payload". When the flag
-        // is set, strip non-text media from this tool message and accumulate
-        // it; the combined media is emitted as a single follow-up user
-        // message after the parts loop completes — preserving the
-        // "all tool responses contiguous" requirement for parallel tool
-        // calls. Default (flag false) preserves prior behavior: media is
-        // embedded in the tool message and permissive providers continue
-        // to receive it that way. See #3616.
+        // Controlled by ContentGeneratorConfig.splitToolMedia (default true;
+        // resolved in pipeline.ts). OpenAI spec only permits string / text-part
+        // content on `role: "tool"` messages. Strict OpenAI-compatible servers
+        // (e.g. doubao / new-api / LM Studio) silently drop or reject tool
+        // messages containing image_url / input_audio / video_url / file parts
+        // (HTTP 400 "Invalid 'messages' in payload"), so an image read via
+        // read_file never reaches the model. When the flag is set, strip
+        // non-text media from this tool message and accumulate it; the combined
+        // media is emitted as a single follow-up user message after the parts
+        // loop completes — preserving the "all tool responses contiguous"
+        // requirement for parallel tool calls. Opt out (flag false) to restore
+        // the legacy behavior: media embedded in the tool message, which only
+        // permissive providers accept. See #4876, #3616.
         if (
           requestContext.splitToolMedia &&
           Array.isArray(toolMessage.content)

packages/core/src/core/openaiContentGenerator/pipeline.test.ts

@@ -301,6 +301,50 @@ describe('ContentGenerationPipeline', () => {
       );
     });
 
+    it('should default splitToolMedia to true when neither provider override nor content generator config sets it (issue #4876)', async () => {
+      const request: GenerateContentParameters = {
+        model: 'test-model',
+        contents: [{ parts: [{ text: 'Hello' }], role: 'user' }],
+      };
+      const userPromptId = 'test-prompt-id';
+      const mockMessages = [
+        { role: 'user', content: 'Hello' },
+      ] as OpenAI.Chat.ChatCompletionMessageParam[];
+      const mockOpenAIResponse = {
+        id: 'response-id',
+        choices: [
+          { message: { content: 'Hello response' }, finish_reason: 'stop' },
+        ],
+        created: Date.now(),
+        model: 'test-model',
+      } as OpenAI.Chat.ChatCompletion;
+      const mockGeminiResponse = new GenerateContentResponse();
+
+      // Neither the provider nor the content generator config sets
+      // splitToolMedia — it must default to true so tool-returned images are
+      // moved out of the spec-violating `role: "tool"` message (#4876).
+      mockProvider.getRequestContextOverrides = vi.fn().mockReturnValue({});
+      mockContentGeneratorConfig.splitToolMedia = undefined;
+      (mockConverter.convertGeminiRequestToOpenAI as Mock).mockReturnValue(
+        mockMessages,
+      );
+      (mockConverter.convertOpenAIResponseToGemini as Mock).mockReturnValue(
+        mockGeminiResponse,
+      );
+      (mockClient.chat.completions.create as Mock).mockResolvedValue(
+        mockOpenAIResponse,
+      );
+
+      await pipeline.execute(request, userPromptId);
+
+      expect(mockConverter.convertGeminiRequestToOpenAI).toHaveBeenCalledWith(
+        request,
+        expect.objectContaining({
+          splitToolMedia: true,
+        }),
+      );
+    });
+
     it('should fall back to configured model when request.model is empty', async () => {
       // Arrange — empty model string is falsy, should fall back to contentGeneratorConfig.model
       const request: GenerateContentParameters = {

packages/core/src/core/openaiContentGenerator/pipeline.ts

@@ -618,7 +618,14 @@ export class ContentGenerationPipeline {
       splitToolMedia:
         providerOverrides.splitToolMedia ??
         this.contentGeneratorConfig.splitToolMedia ??
-        false,
+        // Default true: the OpenAI Chat Completions spec only permits text on
+        // `role: "tool"` messages, so tool-returned media (e.g. an image read
+        // by read_file) embedded there is silently dropped or rejected by
+        // strict providers (doubao / new-api / LM Studio) and the model never
+        // sees it (QwenLM/qwen-code#4876). Splitting it into a follow-up user
+        // message is spec-compliant and safe for permissive providers too.
+        // Opt out via generationConfig.splitToolMedia = false.
+        true,
       ...(toolCallParser ? { toolCallParser } : {}),
       ...(responseParsingOptions ? { responseParsingOptions } : {}),
       ...(taggedThinkingParser ? { taggedThinkingParser } : {}),

packages/vscode-ide-companion/schemas/settings.schema.json

@@ -540,9 +540,9 @@
               "default": true
             },
             "splitToolMedia": {
-              "description": "When true, media (images / audio / video / files) returned by MCP tool calls is split into a follow-up user message instead of being embedded in the tool message. Required for strict OpenAI-compatible servers (e.g., LM Studio) that reject non-text content on `role: \"tool\"` messages with HTTP 400 \"Invalid 'messages' in payload\". Default false preserves the prior behavior for permissive providers. See QwenLM/qwen-code#3616.",
+              "description": "When true, media (images / audio / video / files) returned by tool calls — including the built-in read_file and MCP tools — is split into a follow-up user message instead of being embedded in the `role: \"tool\"` message. The OpenAI Chat Completions spec only permits text on tool messages, so strict OpenAI-compatible servers (e.g., doubao / new-api / LM Studio) silently drop or reject embedded media and the model never sees an image read via read_file (QwenLM/qwen-code#4876, #3616). Default true is spec-compliant and safe for permissive providers; set false only to restore the legacy embed-in-tool-message behavior.",
               "type": "boolean",
-              "default": false
+              "default": true
             },
             "schemaCompliance": {
               "description": "The compliance mode for tool schemas sent to the model. Use \"openapi_30\" for strict OpenAPI 3.0 compatibility (e.g., for Gemini). Options: auto, openapi_30",