fix(agent): allow tool responses to pass image content to next-round vision input

packages/global/core/ai/llm/type.ts

@@ -118,11 +118,12 @@ export type ChatCompletionUserMessageParam = z.infer<typeof ChatCompletionUserMe
 
 /**
  * Tool message: 对齐 openai SDK 的 ChatCompletionToolMessageParam，新增可选 `name`
- * SDK content 仅允许纯文本 part
+ * content 允许 string 或 content part 数组（含 image_url），
+ * 以便工具返回的图片 URL 作为图像进入下一轮请求
  */
 export const ChatCompletionToolMessageParamSchema = z.object({
   role: z.literal('tool'),
-  content: z.union([z.string(), z.array(ChatCompletionContentPartTextSchema)]),
+  content: z.union([z.string(), z.array(ChatCompletionContentPartSchema)]),
   tool_call_id: z.string(),
   name: z.string().optional()
 });

packages/service/core/ai/llm/agentLoop/provider/piAgent/message.ts

@@ -78,15 +78,13 @@ const parseToolArguments = (value: string) => {
 };
 
 const getContentText = (
-  content:
-    | string
-    | null
-    | Array<{ type: 'text'; text: string } | { type: 'refusal'; refusal: string }>
-    | undefined
+  content: string | null | Array<{ type: string; text?: string; refusal?: string }> | undefined
 ) => {
   if (!content) return '';
   if (typeof content === 'string') return content;
-  return content.map((item) => (item.type === 'text' ? item.text : item.refusal)).join('');
+  return content
+    .map((item) => (item.type === 'text' ? (item.text ?? '') : (item.refusal ?? '')))
+    .join('');
 };
 
 /**

packages/service/core/ai/llm/promptCall/index.ts

@@ -56,9 +56,15 @@ export const promptToolCallMessageRewrite = (
         delete message.tool_calls;
       }
     } else if (message.role === 'tool') {
+      const toolContent =
+        typeof message.content === 'string'
+          ? message.content
+          : Array.isArray(message.content)
+            ? message.content.map((item) => (item.type === 'text' ? item.text : '')).join('')
+            : '';
       cloneMessages.splice(i, 1, {
         role: 'user',
-        content: `<ToolResponse>\n${message.content}\n</ToolResponse>`
+        content: `<ToolResponse>\n${toolContent}\n</ToolResponse>`
       });
     }
   }

packages/service/core/ai/llm/utils.ts

@@ -747,6 +747,23 @@ export const loadRequestMessages = async ({
             ...formatAssistantItem(item, supportReason),
             content: formatContent
           };
+        } else if (item.role === ChatCompletionRequestMessageRoleEnum.Tool) {
+          // 工具返回的文本里可能包含图片 URL，提取为 image_url 内容块进入下一轮
+          if (typeof item.content === 'string' && item.content !== '') {
+            const parts = parseTextUserContentParts(item.content);
+            // 只有提取到媒体 part（存在非 text part）时才结构化，否则保持 string 兼容
+            const hasMedia = parts.some((part) => part.type !== 'text');
+            if (hasMedia) {
+              const normalizedContent = await normalizeMediaContentParts(parts);
+              if (normalizedContent.length > 0) {
+                return {
+                  ...item,
+                  content: normalizedContent
+                };
+              }
+            }
+          }
+          return item;
         } else {
           return item;
         }

packages/service/test/core/ai/llm/utils.test.ts

@@ -1154,6 +1154,65 @@ describe('loadRequestMessages function tests', () => {
     });
   });
 
+  describe('Tool message image extraction', () => {
+    it('should extract image URL from tool message into image_url part when useVision is true', async () => {
+      serviceEnv.MULTIPLE_DATA_TO_BASE64 = false;
+      const imageUrl = 'https://example.com/screenshot.png';
+      const messages: ChatCompletionMessageParam[] = [
+        { role: ChatCompletionRequestMessageRoleEnum.User, content: 'Generate a chart' },
+        {
+          role: ChatCompletionRequestMessageRoleEnum.Tool,
+          tool_call_id: 'call1',
+          content: `Generated chart: ${imageUrl}`
+        }
+      ];
+
+      const result = await loadRequestMessages({ messages, useVision: true });
+
+      const toolMessage = result.find((m) => m.role === ChatCompletionRequestMessageRoleEnum.Tool);
+      expect(toolMessage).toBeDefined();
+      const content = toolMessage!.content as any[];
+      expect(Array.isArray(content)).toBe(true);
+      expect(
+        content.some((item: any) => item.type === 'image_url' && item.image_url.url === imageUrl)
+      ).toBe(true);
+      expect(content.some((item: any) => item.type === 'text')).toBe(true);
+    });
+
+    it('should keep tool message as string when useVision is false', async () => {
+      const imageUrl = 'https://example.com/screenshot.png';
+      const messages: ChatCompletionMessageParam[] = [
+        { role: ChatCompletionRequestMessageRoleEnum.User, content: 'Generate a chart' },
+        {
+          role: ChatCompletionRequestMessageRoleEnum.Tool,
+          tool_call_id: 'call1',
+          content: `Generated chart: ${imageUrl}`
+        }
+      ];
+
+      const result = await loadRequestMessages({ messages, useVision: false });
+
+      const toolMessage = result.find((m) => m.role === ChatCompletionRequestMessageRoleEnum.Tool);
+      expect(toolMessage!.content).toBe(`Generated chart: ${imageUrl}`);
+    });
+
+    it('should keep tool message as string when it contains no image URL', async () => {
+      const messages: ChatCompletionMessageParam[] = [
+        { role: ChatCompletionRequestMessageRoleEnum.User, content: 'Hello' },
+        {
+          role: ChatCompletionRequestMessageRoleEnum.Tool,
+          tool_call_id: 'call1',
+          content: 'plain text result'
+        }
+      ];
+
+      const result = await loadRequestMessages({ messages, useVision: true });
+
+      const toolMessage = result.find((m) => m.role === ChatCompletionRequestMessageRoleEnum.Tool);
+      expect(toolMessage!.content).toBe('plain text result');
+    });
+  });
+
   describe('Complex scenarios', () => {
     it('should handle mixed conversation with all message types', async () => {
       const messages: ChatCompletionMessageParam[] = [