fix: stop silently dropping unsupported multimodal content in kosong providers (#632)

* fix: stop silently dropping unsupported media in provider conversions

* fix: keep non-standard audio/video parts off the chat completions wire

.changeset/fix-provider-media-placeholders.md

@@ -0,0 +1,6 @@
+---
+"@moonshot-ai/kosong": patch
+"@moonshot-ai/kimi-code": patch
+---
+
+Degrade unsupported audio/video to placeholder text and reattach tool result media instead of silently dropping them.

packages/kosong/src/providers/anthropic.ts

@@ -412,6 +412,15 @@ interface AnthropicImageBlock {
   cache_control?: { type: 'ephemeral' };
 }
 
+// The Messages API has no representation for audio or video input. Instead of
+// silently dropping such parts (the model would not even know an attachment
+// existed), emit a placeholder text block so it can acknowledge the gap.
+// Consecutive parts of the same kind collapse into a single placeholder.
+const OMITTED_MEDIA_PLACEHOLDER = {
+  audio_url: '(audio omitted: not supported by this provider)',
+  video_url: '(video omitted: not supported by this provider)',
+} as const;
+
 const SUPPORTED_B64_MEDIA_TYPES = new Set(['image/png', 'image/jpeg', 'image/gif', 'image/webp']);
 
 function imageUrlPartToAnthropic(url: string): AnthropicImageBlock {
@@ -458,8 +467,13 @@ function toolResultToBlock(toolCallId: string, content: ContentPart[]): ToolResu
       }
     } else if (part.type === 'image_url') {
       blocks.push(imageUrlPartToAnthropic(part.imageUrl.url));
+    } else if (part.type === 'audio_url' || part.type === 'video_url') {
+      const placeholder = OMITTED_MEDIA_PLACEHOLDER[part.type];
+      const last = blocks.at(-1);
+      if (!(last?.type === 'text' && last.text === placeholder)) {
+        blocks.push({ type: 'text', text: placeholder });
+      }
     }
-    // Other types not supported by Anthropic in tool results
   }
   return {
     type: 'tool_result',
@@ -522,8 +536,13 @@ function convertMessage(message: Message, model: string): MessageParam {
       } else if (part.think !== '' && shouldPreserveUnsignedThinking(model)) {
         blocks.push({ type: 'thinking', thinking: part.think } as unknown as ThinkingBlockParam);
       }
+    } else if (part.type === 'audio_url' || part.type === 'video_url') {
+      const placeholder = OMITTED_MEDIA_PLACEHOLDER[part.type];
+      const last = blocks.at(-1);
+      if (!(last?.type === 'text' && last.text === placeholder)) {
+        blocks.push({ type: 'text', text: placeholder } satisfies TextBlockParam);
+      }
     }
-    // audio_url, video_url: not supported by Anthropic, skip
   }
 
   // Tool calls -> ToolUseBlockParam

packages/kosong/src/providers/openai-common.ts

@@ -286,6 +286,18 @@ export function normalizeOpenAIFinishReason(raw: string | null | undefined): {
  */
 export type ToolMessageConversion = 'extract_text' | null;
 
+/**
+ * Shared wording for tool-result media that cannot live inside the tool
+ * message itself and is reattached as a follow-up user message instead.
+ */
+export const TOOL_RESULT_MEDIA_PROMPT = 'Attached media from tool result:';
+export const TOOL_RESULT_MEDIA_PLACEHOLDER = '(see attached media)';
+
+/** A content part that is neither plain text nor reasoning. */
+export function isMediaPart(part: ContentPart): boolean {
+  return part.type !== 'text' && part.type !== 'think';
+}
+
 /**
  * Convert tool-role message content according to the chosen strategy.
  */

packages/kosong/src/providers/openai-legacy.ts

@@ -21,6 +21,8 @@ import {
   isFunctionToolCall,
   normalizeOpenAIFinishReason,
   type OpenAIContentPart,
+  TOOL_RESULT_MEDIA_PLACEHOLDER,
+  TOOL_RESULT_MEDIA_PROMPT,
   type ToolMessageConversion,
   reasoningEffortToThinkingEffort,
   thinkingEffortToReasoningEffort,
@@ -50,8 +52,6 @@ const OPENAI_CHAT_TOOL_CALL_ID_POLICY: ToolCallIdPolicy = {
   normalize: (id) => sanitizeToolCallId(id, 64),
   maxLength: 64,
 };
-const TOOL_RESULT_IMAGE_PROMPT = 'Attached image(s) from tool result:';
-const TOOL_RESULT_IMAGE_PLACEHOLDER = '(see attached image)';
 
 function extractReasoningContent(
   source: unknown,
@@ -220,46 +220,55 @@ function convertMessage(
   return result;
 }
 
+// Chat Completions has no url-based audio/video content part (only base64
+// `input_audio`), so unlike images these cannot be reattached as user input.
+// Note the omission inline in the tool message text instead.
+const OMITTED_AUDIO_PLACEHOLDER = '(audio omitted: not supported by this provider)';
+const OMITTED_VIDEO_PLACEHOLDER = '(video omitted: not supported by this provider)';
+
 function convertToolMessageContentForChat(
   message: Message,
   conversion: ToolMessageConversion,
 ): string | OpenAIContentPart[] {
   const content = convertToolMessageContent(message, conversion);
-  if (
-    typeof content === 'string' &&
-    content.length === 0 &&
-    message.content.some((part) => part.type === 'image_url')
-  ) {
-    return TOOL_RESULT_IMAGE_PLACEHOLDER;
+  if (typeof content !== 'string') {
+    return content;
   }
-  return content;
+  const lines: string[] = content.length > 0 ? [content] : [];
+  if (message.content.some((part) => part.type === 'audio_url')) {
+    lines.push(OMITTED_AUDIO_PLACEHOLDER);
+  }
+  if (message.content.some((part) => part.type === 'video_url')) {
+    lines.push(OMITTED_VIDEO_PLACEHOLDER);
+  }
+  if (lines.length === 0 && message.content.some((part) => part.type === 'image_url')) {
+    return TOOL_RESULT_MEDIA_PLACEHOLDER;
+  }
+  return lines.join('\n');
 }
 
 function toolResultImageParts(message: Message): OpenAIContentPart[] {
   const images: OpenAIContentPart[] = [];
   for (const part of message.content) {
     if (part.type !== 'image_url') continue;
-    images.push({
-      type: 'image_url',
-      image_url:
-        part.imageUrl.id === undefined
-          ? { url: part.imageUrl.url }
-          : { url: part.imageUrl.url, id: part.imageUrl.id },
-    });
+    const converted = convertContentPart(part);
+    if (converted !== null) {
+      images.push(converted);
+    }
   }
   return images;
 }
 
-function appendToolResultImagesMessage(
+function appendToolResultMediaMessage(
   messages: OpenAIMessage[],
-  pendingToolResultImages: OpenAIContentPart[],
+  pendingToolResultMedia: OpenAIContentPart[],
 ): void {
-  if (pendingToolResultImages.length === 0) return;
+  if (pendingToolResultMedia.length === 0) return;
   messages.push({
     role: 'user',
-    content: [{ type: 'text', text: TOOL_RESULT_IMAGE_PROMPT }, ...pendingToolResultImages],
+    content: [{ type: 'text', text: TOOL_RESULT_MEDIA_PROMPT }, ...pendingToolResultMedia],
   });
-  pendingToolResultImages.length = 0;
+  pendingToolResultMedia.length = 0;
 }
 
 function convertHistoryMessages(
@@ -268,19 +277,19 @@ function convertHistoryMessages(
   toolMessageConversion: ToolMessageConversion,
 ): OpenAIMessage[] {
   const messages: OpenAIMessage[] = [];
-  const pendingToolResultImages: OpenAIContentPart[] = [];
+  const pendingToolResultMedia: OpenAIContentPart[] = [];
 
   for (const msg of history) {
     if (msg.role !== 'tool') {
-      appendToolResultImagesMessage(messages, pendingToolResultImages);
+      appendToolResultMediaMessage(messages, pendingToolResultMedia);
     }
     messages.push(convertMessage(msg, reasoningKey, toolMessageConversion));
     if (msg.role === 'tool') {
-      pendingToolResultImages.push(...toolResultImageParts(msg));
+      pendingToolResultMedia.push(...toolResultImageParts(msg));
     }
   }
 
-  appendToolResultImagesMessage(messages, pendingToolResultImages);
+  appendToolResultMediaMessage(messages, pendingToolResultMedia);
   return messages;
 }
 export class OpenAILegacyStreamedMessage implements StreamedMessage {

packages/kosong/src/providers/openai-responses.ts

@@ -25,6 +25,9 @@ import {
 } from './capability-registry';
 import {
   convertOpenAIError,
+  isMediaPart,
+  TOOL_RESULT_MEDIA_PLACEHOLDER,
+  TOOL_RESULT_MEDIA_PROMPT,
   type ToolMessageConversion,
   reasoningEffortToThinkingEffort,
   thinkingEffortToReasoningEffort,
@@ -363,6 +366,12 @@ interface ResponseToolParam {
   parameters: Record<string, unknown>;
   strict: boolean;
 }
+// The Responses API has no input type for video, and only mp3/wav audio can
+// be inlined as input_file data. Degrade such parts to placeholder text so
+// the model still learns an attachment existed instead of silently losing it.
+const OMITTED_AUDIO_PLACEHOLDER = '(audio omitted: unsupported audio format)';
+const OMITTED_VIDEO_PLACEHOLDER = '(video omitted: not supported by this provider)';
+
 function contentPartsToInputItems(parts: ContentPart[]): unknown[] {
   const items: unknown[] = [];
   for (const part of parts) {
@@ -381,14 +390,14 @@ function contentPartsToInputItems(parts: ContentPart[]): unknown[] {
         break;
       case 'audio_url': {
         const mapped = mapAudioUrlToInputItem(part.audioUrl.url);
-        if (mapped !== null) {
-          items.push(mapped);
-        }
+        items.push(mapped ?? { type: 'input_text', text: OMITTED_AUDIO_PLACEHOLDER });
         break;
       }
-      case 'think':
       case 'video_url':
-        // think: handled separately. video_url: not supported by Responses API.
+        items.push({ type: 'input_text', text: OMITTED_VIDEO_PLACEHOLDER });
+        break;
+      case 'think':
+        // Handled separately as reasoning items.
         break;
     }
   }
@@ -405,7 +414,7 @@ function contentPartsToOutputItems(parts: ContentPart[]): unknown[] {
   return items;
 }
 
-function messageContentToFunctionOutputItems(content: ContentPart[]): string | unknown[] {
+function messageContentToFunctionOutputItems(content: ContentPart[]): unknown[] {
   const items: unknown[] = [];
   for (const part of content) {
     switch (part.type) {
@@ -424,15 +433,14 @@ function messageContentToFunctionOutputItems(content: ContentPart[]): string | u
         // branch here, audio returned by a tool would be dropped on the
         // next turn.
         const mapped = mapAudioUrlToInputItem(part.audioUrl.url);
-        if (mapped !== null) {
-          items.push(mapped);
-        }
+        items.push(mapped ?? { type: 'input_text', text: OMITTED_AUDIO_PLACEHOLDER });
         break;
       }
-      case 'think':
       case 'video_url':
-        // think / video_url still intentionally skipped: the Responses
-        // API has no representation for them inside a function_call_output.
+        items.push({ type: 'input_text', text: OMITTED_VIDEO_PLACEHOLDER });
+        break;
+      case 'think':
+        // Handled separately as reasoning items.
         break;
     }
   }
@@ -477,10 +485,20 @@ function convertMessage(
   // tool role -> function_call_output
   if (role === 'tool') {
     const callId = message.toolCallId ?? '';
-    const output: string | unknown[] =
-      toolMessageConversion === 'extract_text'
-        ? extractText(message)
-        : messageContentToFunctionOutputItems(message.content);
+    let output: string | unknown[];
+    if (toolMessageConversion === 'extract_text') {
+      // Plain-string output for backends that reject structured
+      // function_call_output. Media parts are reattached as a user message
+      // by `convertHistoryMessages`; when the result carries no text at
+      // all, point the model at that follow-up message.
+      const text = extractText(message);
+      output =
+        text.length === 0 && message.content.some(isMediaPart)
+          ? TOOL_RESULT_MEDIA_PLACEHOLDER
+          : text;
+    } else {
+      output = messageContentToFunctionOutputItems(message.content);
+    }
     return [
       {
         call_id: callId,
@@ -571,6 +589,49 @@ function convertTool(tool: Tool): ResponseToolParam {
     strict: false,
   };
 }
+
+/**
+ * Convert the history, buffering tool-result media when `extract_text`
+ * flattens tool outputs to plain strings. The buffered media items are
+ * reattached as a single user message after each run of consecutive tool
+ * messages — mirroring the OpenAI Chat Completions provider.
+ */
+function convertHistoryMessages(
+  history: readonly Message[],
+  modelName: string,
+  toolMessageConversion: ToolMessageConversion,
+): unknown[] {
+  const input: unknown[] = [];
+  const pendingToolResultMedia: unknown[] = [];
+
+  const flushPendingMedia = (): void => {
+    if (pendingToolResultMedia.length === 0) return;
+    input.push({
+      type: 'message',
+      role: 'user',
+      content: [
+        { type: 'input_text', text: TOOL_RESULT_MEDIA_PROMPT },
+        ...pendingToolResultMedia,
+      ],
+    });
+    pendingToolResultMedia.length = 0;
+  };
+
+  for (const msg of history) {
+    if (msg.role !== 'tool') {
+      flushPendingMedia();
+    }
+    input.push(...convertMessage(msg, modelName, toolMessageConversion));
+    if (msg.role === 'tool' && toolMessageConversion === 'extract_text') {
+      pendingToolResultMedia.push(
+        ...messageContentToFunctionOutputItems(msg.content.filter(isMediaPart)),
+      );
+    }
+  }
+
+  flushPendingMedia();
+  return input;
+}
 export class OpenAIResponsesStreamedMessage implements StreamedMessage {
   private _id: string | null = null;
   private _usage: TokenUsage | null = null;
@@ -991,9 +1052,9 @@ export class OpenAIResponsesChatProvider implements ChatProvider {
       history,
       OPENAI_RESPONSES_TOOL_CALL_ID_POLICY,
     );
-    for (const msg of normalizedHistory) {
-      input.push(...convertMessage(msg, this._model, this._toolMessageConversion));
-    }
+    input.push(
+      ...convertHistoryMessages(normalizedHistory, this._model, this._toolMessageConversion),
+    );
 
     const kwargs: Record<string, unknown> = { ...this._generationKwargs };
     const reasoningEffort = kwargs['reasoning_effort'] as string | undefined;

packages/kosong/test/anthropic.test.ts

@@ -447,6 +447,73 @@ describe('AnthropicChatProvider', () => {
       });
     });
 
+    it('user audio/video parts degrade to placeholder text, consecutive same-kind collapse', async () => {
+      // The Messages API cannot carry audio or video. Dropping the parts
+      // silently would leave the model unaware an attachment ever existed,
+      // so each unsupported part degrades to a placeholder text block.
+      const provider = createProvider();
+      const history: Message[] = [
+        {
+          role: 'user',
+          content: [
+            { type: 'text', text: 'Listen and watch:' },
+            { type: 'audio_url', audioUrl: { url: 'https://example.com/a.mp3' } },
+            { type: 'audio_url', audioUrl: { url: 'https://example.com/b.mp3' } },
+            { type: 'video_url', videoUrl: { url: 'https://example.com/c.mp4' } },
+          ] satisfies ContentPart[],
+          toolCalls: [],
+        },
+      ];
+      const body = await captureRequestBody(provider, '', [], history);
+      const messages = body['messages'] as Record<string, unknown>[];
+
+      expect(messages[0]?.['content']).toEqual([
+        { type: 'text', text: 'Listen and watch:' },
+        { type: 'text', text: '(audio omitted: not supported by this provider)' },
+        {
+          type: 'text',
+          text: '(video omitted: not supported by this provider)',
+          cache_control: { type: 'ephemeral' },
+        },
+      ]);
+    });
+
+    it('tool result audio degrades to placeholder text inside tool_result', async () => {
+      const provider = createProvider();
+      const history: Message[] = [
+        { role: 'user', content: [{ type: 'text', text: 'Say hi' }], toolCalls: [] },
+        {
+          role: 'assistant',
+          content: [],
+          toolCalls: [{ type: 'function', id: 'call_tts', name: 'tts', arguments: '{}' }],
+        },
+        {
+          role: 'tool',
+          content: [
+            { type: 'audio_url', audioUrl: { url: 'https://example.com/hi.mp3' } },
+          ] satisfies ContentPart[],
+          toolCallId: 'call_tts',
+          toolCalls: [],
+        },
+      ];
+      const body = await captureRequestBody(provider, '', [], history);
+      const messages = body['messages'] as Record<string, unknown>[];
+
+      expect(messages[2]).toEqual({
+        role: 'user',
+        content: [
+          {
+            type: 'tool_result',
+            tool_use_id: 'call_tts',
+            content: [
+              { type: 'text', text: '(audio omitted: not supported by this provider)' },
+            ],
+            cache_control: { type: 'ephemeral' },
+          },
+        ],
+      });
+    });
+
     it('parallel tool calls and tool results (request body capture)', async () => {
       const provider = createProvider();
       const tcAdd: ToolCall = {

packages/kosong/test/openai-legacy.test.ts

@@ -332,12 +332,77 @@ describe('OpenAILegacyChatProvider', () => {
       expect(messages[3]).toEqual({
         role: 'user',
         content: [
-          { type: 'text', text: 'Attached image(s) from tool result:' },
+          { type: 'text', text: 'Attached media from tool result:' },
           { type: 'image_url', image_url: { url: 'https://example.com/image.png' } },
         ],
       });
     });
 
+    it('tool call with audio result notes the omission inline without reattaching', async () => {
+      // Chat Completions has no url-based audio/video content part (only
+      // base64 input_audio), so unlike images these cannot be reattached as
+      // a user message — a standard OpenAI endpoint would reject the request
+      // with a 400. The tool message notes the omission inline instead.
+      const provider = createProvider();
+      const history: Message[] = [
+        { role: 'user', content: [{ type: 'text', text: 'Say hi' }], toolCalls: [] },
+        {
+          role: 'assistant',
+          content: [],
+          toolCalls: [{ type: 'function', id: 'call_tts', name: 'tts', arguments: '{}' }],
+        },
+        {
+          role: 'tool',
+          content: [
+            { type: 'audio_url', audioUrl: { url: 'https://example.com/hi.mp3' } },
+          ] satisfies ContentPart[],
+          toolCallId: 'call_tts',
+          toolCalls: [],
+        },
+      ];
+      const body = await captureRequestBody(provider, '', [], history);
+
+      const messages = body['messages'] as Record<string, unknown>[];
+      expect(messages[2]).toEqual({
+        role: 'tool',
+        content: '(audio omitted: not supported by this provider)',
+        tool_call_id: 'call_tts',
+      });
+      // No follow-up user message: audio_url is not a standard Chat
+      // Completions content part and must not reach the wire.
+      expect(messages).toHaveLength(3);
+    });
+
+    it('tool call with text and video result appends the omission note to the text', async () => {
+      const provider = createProvider();
+      const history: Message[] = [
+        { role: 'user', content: [{ type: 'text', text: 'Record it' }], toolCalls: [] },
+        {
+          role: 'assistant',
+          content: [],
+          toolCalls: [{ type: 'function', id: 'call_rec', name: 'record', arguments: '{}' }],
+        },
+        {
+          role: 'tool',
+          content: [
+            { type: 'text', text: 'recorded 5s clip' },
+            { type: 'video_url', videoUrl: { url: 'https://example.com/rec.mp4' } },
+          ] satisfies ContentPart[],
+          toolCallId: 'call_rec',
+          toolCalls: [],
+        },
+      ];
+      const body = await captureRequestBody(provider, '', [], history);
+
+      const messages = body['messages'] as Record<string, unknown>[];
+      expect(messages[2]).toEqual({
+        role: 'tool',
+        content: 'recorded 5s clip\n(video omitted: not supported by this provider)',
+        tool_call_id: 'call_rec',
+      });
+      expect(messages).toHaveLength(3);
+    });
+
     it('groups consecutive tool result images after all matching tool messages', async () => {
       const provider = createProvider();
       const history: Message[] = [
@@ -388,12 +453,12 @@ describe('OpenAILegacyChatProvider', () => {
             },
           ],
         },
-        { role: 'tool', content: '(see attached image)', tool_call_id: 'call_first' },
+        { role: 'tool', content: '(see attached media)', tool_call_id: 'call_first' },
         { role: 'tool', content: 'second', tool_call_id: 'call_second' },
         {
           role: 'user',
           content: [
-            { type: 'text', text: 'Attached image(s) from tool result:' },
+            { type: 'text', text: 'Attached media from tool result:' },
             { type: 'image_url', image_url: { url: 'https://example.com/first.png' } },
             { type: 'image_url', image_url: { url: 'https://example.com/second.png' } },
           ],

packages/kosong/test/openai-responses.test.ts

@@ -292,7 +292,7 @@ describe('OpenAIResponsesChatProvider', () => {
       ]);
     });
 
-    it('user message with unsupported audio_url format drops the audio part silently', async () => {
+    it('user message with unsupported audio_url format degrades to placeholder text', async () => {
       const provider = createProvider();
       const history: Message[] = [
         {
@@ -307,8 +307,12 @@ describe('OpenAIResponsesChatProvider', () => {
       const body = await captureRequestBody(provider, '', [], history);
 
       const input = body['input'] as Array<{ content: unknown[] }>;
-      // Only the text part survives; the unsupported ogg audio is dropped.
-      expect(input[0]!.content).toEqual([{ type: 'input_text', text: 'Bare text' }]);
+      // The unsupported ogg audio degrades to a placeholder instead of
+      // silently vanishing, so the model knows an attachment existed.
+      expect(input[0]!.content).toEqual([
+        { type: 'input_text', text: 'Bare text' },
+        { type: 'input_text', text: '(audio omitted: unsupported audio format)' },
+      ]);
     });
 
     it('multiple consecutive ThinkParts with the same encrypted value aggregate into one reasoning item with multiple summaries', async () => {
@@ -407,6 +411,102 @@ describe('OpenAIResponsesChatProvider', () => {
       });
     });
 
+    it('toolMessageConversion=extract_text reattaches tool result media as a user message', async () => {
+      // extract_text flattens function_call_output to a plain string for
+      // backends that reject structured output. Media must not vanish with
+      // the flattening: the image-only result gets a placeholder string and
+      // the media items are reattached as a follow-up user message after
+      // the run of consecutive tool messages.
+      const provider = new OpenAIResponsesChatProvider({
+        model: 'gpt-4.1',
+        apiKey: 'test-key',
+        toolMessageConversion: 'extract_text',
+      });
+      const history: Message[] = [
+        { role: 'user', content: [{ type: 'text', text: 'Q' }], toolCalls: [] },
+        {
+          role: 'assistant',
+          content: [],
+          toolCalls: [
+            { type: 'function', id: 'call_shot', name: 'screenshot', arguments: '{}' },
+            { type: 'function', id: 'call_read', name: 'read', arguments: '{}' },
+          ],
+        },
+        {
+          role: 'tool',
+          content: [
+            { type: 'image_url', imageUrl: { url: 'https://example.com/shot.png' } },
+          ],
+          toolCallId: 'call_shot',
+          toolCalls: [],
+        },
+        {
+          role: 'tool',
+          content: [{ type: 'text', text: 'file body' }],
+          toolCallId: 'call_read',
+          toolCalls: [],
+        },
+      ];
+      const body = await captureRequestBody(provider, '', [], history);
+
+      const input = body['input'] as Array<Record<string, unknown>>;
+      expect(input).toEqual([
+        { type: 'message', role: 'user', content: [{ type: 'input_text', text: 'Q' }] },
+        {
+          type: 'function_call',
+          call_id: 'call_shot',
+          name: 'screenshot',
+          arguments: '{}',
+        },
+        { type: 'function_call', call_id: 'call_read', name: 'read', arguments: '{}' },
+        {
+          type: 'function_call_output',
+          call_id: 'call_shot',
+          output: '(see attached media)',
+        },
+        { type: 'function_call_output', call_id: 'call_read', output: 'file body' },
+        {
+          type: 'message',
+          role: 'user',
+          content: [
+            { type: 'input_text', text: 'Attached media from tool result:' },
+            { type: 'input_image', image_url: 'https://example.com/shot.png' },
+          ],
+        },
+      ]);
+    });
+
+    it('video_url in tool result degrades to placeholder text in function_call_output', async () => {
+      const provider = createProvider();
+      const history: Message[] = [
+        { role: 'user', content: [{ type: 'text', text: 'Record it' }], toolCalls: [] },
+        {
+          role: 'assistant',
+          content: [],
+          toolCalls: [{ type: 'function', id: 'call_rec', name: 'record', arguments: '{}' }],
+        },
+        {
+          role: 'tool',
+          content: [
+            { type: 'video_url', videoUrl: { url: 'https://example.com/rec.mp4' } },
+          ],
+          toolCallId: 'call_rec',
+          toolCalls: [],
+        },
+      ];
+      const body = await captureRequestBody(provider, '', [], history);
+
+      const input = body['input'] as Array<Record<string, unknown>>;
+      const fnOutput = input.find((item) => item['type'] === 'function_call_output');
+      expect(fnOutput).toEqual({
+        type: 'function_call_output',
+        call_id: 'call_rec',
+        output: [
+          { type: 'input_text', text: '(video omitted: not supported by this provider)' },
+        ],
+      });
+    });
+
     it('parallel tool calls produce multiple function_call and function_call_output items', async () => {
       const provider = createProvider();
       const history: Message[] = [
@@ -1120,7 +1220,7 @@ describe('OpenAIResponsesChatProvider', () => {
       expect(body['max_output_tokens']).toBe(512);
     });
 
-    it('video_url in user content is silently skipped (Responses API has no video representation)', async () => {
+    it('video_url in user content degrades to placeholder text (no video input type)', async () => {
       const provider = createProvider();
       const history: Message[] = [
         {
@@ -1134,12 +1234,14 @@ describe('OpenAIResponsesChatProvider', () => {
       ];
       const body = await captureRequestBody(provider, '', [], history);
 
-      // Only the text part survives; video is dropped.
       const input = body['input'] as Array<{ content: unknown[] }>;
-      expect(input[0]!.content).toEqual([{ type: 'input_text', text: 'Watch this:' }]);
+      expect(input[0]!.content).toEqual([
+        { type: 'input_text', text: 'Watch this:' },
+        { type: 'input_text', text: '(video omitted: not supported by this provider)' },
+      ]);
     });
 
-    it('audio_url with unsupported scheme is silently dropped from user content', async () => {
+    it('audio_url with unsupported scheme degrades to placeholder text', async () => {
       const provider = createProvider();
       const history: Message[] = [
         {
@@ -1154,11 +1256,14 @@ describe('OpenAIResponsesChatProvider', () => {
       const body = await captureRequestBody(provider, '', [], history);
 
       const input = body['input'] as Array<{ content: unknown[] }>;
-      // file:// URL is unsupported → drop
-      expect(input[0]!.content).toEqual([{ type: 'input_text', text: 'Hear:' }]);
+      // file:// URL cannot be encoded as input_file → placeholder
+      expect(input[0]!.content).toEqual([
+        { type: 'input_text', text: 'Hear:' },
+        { type: 'input_text', text: '(audio omitted: unsupported audio format)' },
+      ]);
     });
 
-    it('audio_url data URL with unknown subtype is silently dropped', async () => {
+    it('audio_url data URL with unknown subtype degrades to placeholder text', async () => {
       const provider = createProvider();
       const history: Message[] = [
         {
@@ -1173,8 +1278,11 @@ describe('OpenAIResponsesChatProvider', () => {
       const body = await captureRequestBody(provider, '', [], history);
 
       const input = body['input'] as Array<{ content: unknown[] }>;
-      // ogg subtype is not mp3/wav → drop
-      expect(input[0]!.content).toEqual([{ type: 'input_text', text: 'OGG:' }]);
+      // ogg subtype is not mp3/wav → placeholder
+      expect(input[0]!.content).toEqual([
+        { type: 'input_text', text: 'OGG:' },
+        { type: 'input_text', text: '(audio omitted: unsupported audio format)' },
+      ]);
     });
   });