fix: strip multimodal parts for text-only models

test/test_api_client_multimodal_sanitize.py

@@ -0,0 +1,54 @@
+from __future__ import annotations
+
+import unittest
+
+from utils.api_client import DeepSeekClient
+
+
+class APIClientMultimodalSanitizeTest(unittest.TestCase):
+    def test_text_only_model_strips_image_parts_to_text(self):
+        client = DeepSeekClient(web_mode=True)
+        client.model_key = "DeepSeek-V4-Pro High"  # custom_models.json: multimodal=none
+
+        messages = [
+            {
+                "role": "tool",
+                "name": "mcp__playwright__browser_take_screenshot",
+                "content": [
+                    {"type": "text", "text": "截图成功"},
+                    {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc"}},
+                ],
+            }
+        ]
+
+        sanitized = client._sanitize_messages_for_model_capability(messages)
+        self.assertEqual(len(sanitized), 1)
+        self.assertIsInstance(sanitized[0].get("content"), str, sanitized[0])
+        self.assertIn("截图成功", sanitized[0].get("content") or "", sanitized[0])
+
+    def test_multimodal_model_keeps_image_parts(self):
+        client = DeepSeekClient(web_mode=True)
+        client.model_key = "qwen3-vl-plus"  # 内置多模态模型
+
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "看图"},
+                    {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc"}},
+                ],
+            }
+        ]
+
+        sanitized = client._sanitize_messages_for_model_capability(messages)
+        self.assertEqual(len(sanitized), 1)
+        self.assertIsInstance(sanitized[0].get("content"), list, sanitized[0])
+        self.assertEqual(
+            [part.get("type") for part in sanitized[0].get("content") or []],
+            ["text", "image_url"],
+            sanitized[0],
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

utils/api_client.py

@@ -316,6 +316,101 @@ class DeepSeekClient:
             "content": "\n\n".join(c for c in merged_contents if c)
         }
         return [merged] + messages[idx:]
+
+    @staticmethod
+    def _sanitize_content_for_capability(
+        content: Any,
+        *,
+        supports_image: bool,
+        supports_video: bool,
+    ) -> Any:
+        """根据模型多模态能力裁剪 content，避免向纯文本模型发送 image_url/video_url。"""
+        if not isinstance(content, list):
+            return content
+
+        kept_parts: List[Dict[str, Any]] = []
+        text_segments: List[str] = []
+
+        for part in content:
+            if not isinstance(part, dict):
+                text = str(part).strip()
+                if text:
+                    text_segments.append(text)
+                continue
+
+            ctype = str(part.get("type") or "").strip().lower()
+            if ctype == "text":
+                text = str(part.get("text") or "")
+                if text:
+                    text_segments.append(text)
+                    kept_parts.append({"type": "text", "text": text})
+                continue
+            if ctype == "image_url":
+                if supports_image:
+                    kept_parts.append(part)
+                continue
+            if ctype == "video_url":
+                if supports_video:
+                    kept_parts.append(part)
+                continue
+
+            # 其他未知/暂不支持类型统一忽略（纯文本模型下只保留 text）
+            continue
+
+        # 纯文本模型：统一回退为字符串，规避供应商对 content part 类型的严格校验
+        if not supports_image and not supports_video:
+            return "\n".join(seg for seg in text_segments if seg).strip()
+
+        # 半多模态模型（例如只支持图片）：尽量保留可支持的 part
+        if kept_parts:
+            if len(kept_parts) == 1 and kept_parts[0].get("type") == "text":
+                return kept_parts[0].get("text", "")
+            return kept_parts
+        return "\n".join(seg for seg in text_segments if seg).strip()
+
+    def _sanitize_messages_for_model_capability(self, messages: List[Dict]) -> List[Dict]:
+        model_key = str(self.model_key or "").strip()
+        if not model_key:
+            return messages
+
+        try:
+            from config.model_profiles import model_supports_image, model_supports_video
+        except Exception:
+            return messages
+
+        supports_image = bool(model_supports_image(model_key))
+        supports_video = bool(model_supports_video(model_key))
+        if supports_image and supports_video:
+            return messages
+
+        changed = 0
+        sanitized: List[Dict[str, Any]] = []
+        for message in messages or []:
+            if not isinstance(message, dict):
+                continue
+            msg_copy = dict(message)
+            original_content = msg_copy.get("content")
+            new_content = self._sanitize_content_for_capability(
+                original_content,
+                supports_image=supports_image,
+                supports_video=supports_video,
+            )
+            if new_content != original_content:
+                changed += 1
+            msg_copy["content"] = new_content
+            sanitized.append(msg_copy)
+
+        if changed:
+            self._debug_log(
+                {
+                    "event": "sanitize_messages_for_model_capability",
+                    "model_key": model_key,
+                    "supports_image": supports_image,
+                    "supports_video": supports_video,
+                    "changed_messages": changed,
+                }
+            )
+        return sanitized
     
     def set_deep_thinking_mode(self, enabled: bool):
         """配置深度思考模式（持续使用思考模型）。"""
@@ -596,6 +691,7 @@ class DeepSeekClient:
         is_minimax = self.model_key == "minimax-m2.5" or "minimax" in lower_base_url
 
         final_messages = self._merge_system_messages(messages)
+        final_messages = self._sanitize_messages_for_model_capability(final_messages)
 
         payload = {
             "model": api_config["model_id"],