myagent-ai 1.16.12 → 1.16.14

package/agents/base.py

@@ -242,7 +242,9 @@ class BaseAgent(ABC):
                     if role == "system":
                         system_msg = content
                         continue
-                    anth_messages.append({"role": role, "content": content})
+                    # 转换 OpenAI Vision 格式为 Anthropic 格式
+                    anth_content = self.llm._convert_to_anthropic_content(content)
+                    anth_messages.append({"role": role, "content": anth_content})
 
                 create_kwargs = {
                     "model": self.llm.model,

package/agents/main_agent.py

@@ -693,6 +693,41 @@ class MainAgent(BaseAgent):
                         )
                     break
 
+                # [v1.16.13] 特殊处理模型不支持图片输入 — 去掉图片用纯文本重试
+                _vision_keywords = ["doesn't support image", "does not support image", "model_incompatible", "image input", "not support vision", "unsupported multimodal", "image capability"]
+                if any(kw in _llm_error.lower() for kw in _vision_keywords) and context.metadata.get("user_images"):
+                    logger.warning(f"[{task_id}] 模型不支持图片输入，去掉图片用纯文本重试")
+                    context.metadata["user_images"] = []
+                    # 用纯文本消息替换最后一条多模态消息
+                    _text_only_msg = context.user_message or "请处理上述上下文。"
+                    if len(messages) > 0 and isinstance(messages[-1].content, list):
+                        messages[-1] = Message(role="user", content=_text_only_msg)
+                    # 重试 LLM 调用
+                    if stream_response and self.llm:
+                        response = await self._call_llm_stream(
+                            messages, text_delta_callback=text_delta_callback,
+                            stream_response=stream_response,
+                        )
+                    else:
+                        response = await self._call_llm(messages)
+                    if response.success:
+                        # 纯文本重试成功，给回复加上提示前缀
+                        _vision_prefix = "⚠️ 当前模型不支持图片识别，已自动使用纯文本模式处理（图片未发送给模型）。\n\n"
+                        llm_raw = _vision_prefix + response.content
+                        context.working_memory["final_response"] = llm_raw
+                        await self._emit_v2_event("v2_reasoning", {"content": llm_raw}, stream_callback)
+                        if self.memory:
+                            self.memory.add_session(
+                                session_id=context.session_id,
+                                role="assistant",
+                                content=llm_raw,
+                            )
+                        break
+                    else:
+                        # 纯文本也失败了，走下面的通用错误处理
+                        _llm_error = response.error or ""
+                        logger.error(f"[{task_id}] 纯文本重试也失败: {_llm_error}")
+
                 # 其他 LLM 错误
                 error_msg = f"LLM 调用失败: {response.error}"
                 context.working_memory["final_response"] = error_msg

package/core/llm.py

@@ -238,6 +238,41 @@ class LLMClient:
     # 所有使用 OpenAI 兼容接口的提供商
     _OPENAI_COMPATIBLE_PROVIDERS = ("openai", "custom", "modelscope", "deepseek", "moonshot", "qwen", "dashscope")
 
+    @staticmethod
+    def _convert_to_anthropic_content(content):
+        """将 OpenAI Vision 格式的 content 转换为 Anthropic 格式
+
+        OpenAI 格式: [{"type": "text", "text": "..."}, {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}]
+        Anthropic 格式: [{"type": "text", "text": "..."}, {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": "..."}}]
+        """
+        if not isinstance(content, list):
+            return content
+
+        import re
+        anth_content = []
+        for item in content:
+            if isinstance(item, dict) and item.get("type") == "image_url":
+                url = item.get("image_url", {}).get("url", "")
+                # 解析 data URI: data:image/png;base64,xxxxx
+                match = re.match(r'^data:([^;]+);base64,(.+)$', url, re.DOTALL)
+                if match:
+                    media_type = match.group(1)
+                    b64_data = match.group(2)
+                    anth_content.append({
+                        "type": "image",
+                        "source": {
+                            "type": "base64",
+                            "media_type": media_type,
+                            "data": b64_data,
+                        }
+                    })
+                else:
+                    # 非 data URI 格式（如 http URL），Anthropic 也支持但方式不同，暂保留原格式
+                    anth_content.append(item)
+            else:
+                anth_content.append(item)
+        return anth_content
+
     def _ensure_client(self):
         """延迟初始化 LLM 客户端"""
         if self._client is not None:
@@ -552,7 +587,9 @@ class LLMClient:
             if m.role == "system":
                 system_msg = m.content
                 continue
-            anth_messages.append({"role": m.role, "content": m.content})
+            # 转换 OpenAI Vision 格式为 Anthropic 格式
+            anth_content = self._convert_to_anthropic_content(m.content)
+            anth_messages.append({"role": m.role, "content": anth_content})
 
         create_kwargs = {
             "model": self.model,
@@ -669,6 +706,7 @@ class LLMClient:
                 logger.error(f"流式调用不支持提供商: {self.provider}")
         except Exception as e:
             logger.error(f"流式 LLM 调用失败: {e}")
+            raise
 
     async def _stream_openai(self, kwargs: dict) -> AsyncGenerator[str, None]:
         """OpenAI / 兼容接口 (含 Zhipu) 流式调用
@@ -715,7 +753,9 @@ class LLMClient:
             if m.role == "system":
                 system_msg = m.content
                 continue
-            anth_messages.append({"role": m.role, "content": m.content})
+            # 转换 OpenAI Vision 格式为 Anthropic 格式
+            anth_content = self._convert_to_anthropic_content(m.content)
+            anth_messages.append({"role": m.role, "content": anth_content})
 
         create_kwargs = {
             "model": self.model,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "myagent-ai",
-  "version": "1.16.12",
+  "version": "1.16.14",
   "description": "本地桌面端执行型AI助手 - Open Interpreter 风格 | Local Desktop Execution-Oriented AI Assistant",
   "main": "main.py",
   "bin": {
@@ -43,4 +43,4 @@
     "python": ">=3.10",
     "node": ">=18"
   }
-}
+}

package/web/api_server.py

@@ -852,6 +852,7 @@ class ApiServer:
                         agent_path=agent_path, agent_system_prompt=agent_system_prompt,
                         chat_mode=chat_mode, stream_response=proxy,
                         voice_text=voice_text,
+                        user_images=user_images, user_files=user_files,
                     )
                 elif self.core.main_agent and self.core.llm:
                     full_response = await self._stream_process_message(
@@ -3700,7 +3701,7 @@ class ApiServer:
     async def _try_model_chain_stream(self, model_chain, message, session_id,
                                        agent_path=None, agent_system_prompt=None,
                                        chat_mode="", stream_response=None,
-                                       voice_text=""):
+                                       voice_text="", user_images=None, user_files=None):
         """流式版本的模型链调用，逐token输出到SSE
         
         使用 asyncio.Lock 保护共享的 self.core.llm，防止并发请求互相干扰。
@@ -3716,12 +3717,13 @@ class ApiServer:
                 agent_path=agent_path, agent_system_prompt=agent_system_prompt,
                 chat_mode=chat_mode, stream_response=stream_response,
                 voice_text=voice_text,
+                user_images=user_images, user_files=user_files,
             )
 
     async def _try_model_chain_stream_inner(self, model_chain, message, session_id,
                                              agent_path=None, agent_system_prompt=None,
                                              chat_mode="", stream_response=None,
-                                             voice_text=""):
+                                             voice_text="", user_images=None, user_files=None):
         """_try_model_chain_stream 的实际执行体（已在 _model_chain_lock 保护下）"""
         llm = self.core.llm
         full_text = ""
@@ -3753,6 +3755,7 @@ class ApiServer:
                     message, session_id, stream_response,
                     agent_path=agent_path, agent_system_prompt=agent_system_prompt,
                     chat_mode=chat_mode, voice_text=voice_text,
+                    user_images=user_images, user_files=user_files,
                 )
                 if result and not result.startswith("⚠️") and not result.startswith("❌"):
                     return result

package/web/ui/chat/chat_main.js

@@ -2776,6 +2776,15 @@ function initAttachmentUI() {
   window._pendingImages = [];
   window._pendingFiles = [];
   
+  // [v1.16.13] 图片压缩配置
+  window._IMAGE_COMPRESS = {
+    maxWidth: 2048,          // 最大宽度
+    maxHeight: 2048,         // 最大高度
+    quality: 0.85,           // JPEG 压缩质量
+    maxSizeBytes: 4 * 1024 * 1024,  // 单张图片最大 4MB (base64后约 5.3MB)
+    skipCompression: false,  // 跳过压缩（如果图片已经足够小）
+  };
+  
   // 创建隐藏的文件输入
   if (!document.getElementById('imageFileInput')) {
     var imgInput = document.createElement('input');
@@ -2799,34 +2808,90 @@ function initAttachmentUI() {
   }
 }
 
+// [v1.16.13] 图片压缩函数 — 缩放尺寸 + JPEG 压缩，返回 Promise<{base64, width, height}>
+function compressImage(file) {
+  return new Promise(function(resolve, reject) {
+    var cfg = window._IMAGE_COMPRESS || {};
+    var reader = new FileReader();
+    reader.onerror = function() { reject(new Error('读取文件失败')); };
+    reader.onload = function(e) {
+      var img = new Image();
+      img.onerror = function() { reject(new Error('图片加载失败')); };
+      img.onload = function() {
+        var w = img.naturalWidth;
+        var h = img.naturalHeight;
+        // 如果图片已足够小且体积未超限，直接返回原始数据
+        var rawBase64 = e.target.result.split(',')[1];
+        if (cfg.skipCompression && file.size <= (cfg.maxSizeBytes || 4194304) && w <= (cfg.maxWidth || 2048) && h <= (cfg.maxHeight || 2048)) {
+          resolve({ base64: rawBase64, width: w, height: h, type: file.type });
+          return;
+        }
+        // 计算缩放比例
+        var maxW = cfg.maxWidth || 2048;
+        var maxH = cfg.maxHeight || 2048;
+        var ratio = Math.min(maxW / w, maxH / h, 1);
+        var newW = Math.round(w * ratio);
+        var newH = Math.round(h * ratio);
+        // Canvas 绘制并压缩
+        var canvas = document.createElement('canvas');
+        canvas.width = newW;
+        canvas.height = newH;
+        var ctx = canvas.getContext('2d');
+        ctx.drawImage(img, 0, 0, newW, newH);
+        // 如果是 PNG 透明图且不大，保留 PNG；否则转 JPEG
+        var useJpeg = file.type !== 'image/png' || file.size > 500000;
+        var mimeType = useJpeg ? 'image/jpeg' : file.type;
+        var quality = useJpeg ? (cfg.quality || 0.85) : undefined;
+        var dataUrl = canvas.toDataURL(mimeType, quality);
+        var b64 = dataUrl.split(',')[1];
+        // 如果压缩后反而更大（罕见），用原始数据
+        if (b64.length > rawBase64.length) {
+          b64 = rawBase64;
+          mimeType = file.type;
+        }
+        resolve({ base64: b64, width: newW, height: newH, type: mimeType });
+      };
+      img.src = e.target.result;
+    };
+    reader.readAsDataURL(file);
+  });
+}
+
 function handleFileSelect(input, type) {
   var files = input.files;
   if (!files || files.length === 0) return;
   
   for (var i = 0; i < files.length; i++) {
     (function(file) {
-      var reader = new FileReader();
-      reader.onload = function(e) {
-        var base64 = e.target.result.split(',')[1];
-        if (type === 'image' && file.type.startsWith('image/')) {
+      if (type === 'image' && file.type.startsWith('image/')) {
+        // [v1.16.13] 图片文件使用压缩
+        compressImage(file).then(function(result) {
           window._pendingImages.push({
-            type: file.type,
-            data: base64,
+            type: result.type,
+            data: result.base64,
             name: file.name,
             size: file.size,
           });
-        } else {
+          renderAttachmentPreview();
+          updateSendBtnState();
+        }).catch(function(err) {
+          console.error('图片压缩失败:', err);
+        });
+      } else {
+        var reader = new FileReader();
+        reader.onload = function(e) {
+          var base64 = e.target.result.split(',')[1];
           window._pendingFiles.push({
             type: file.type,
             data: base64,
             name: file.name,
             size: file.size,
           });
-        }
-        renderAttachmentPreview();
-        updateSendBtnState();
-      };
-      reader.readAsDataURL(file);
+          renderAttachmentPreview();
+          updateSendBtnState();
+        };
+        reader.readAsDataURL(file);
+      }
     })(files[i]);
   }
 }
@@ -2842,19 +2907,19 @@ function handlePasteEvent(e) {
       var file = items[i].getAsFile();
       if (file) {
         (function(f) {
-          var reader = new FileReader();
-          reader.onload = function(ev) {
-            var base64 = ev.target.result.split(',')[1];
+          // [v1.16.13] 粘贴图片使用压缩
+          compressImage(f).then(function(result) {
             window._pendingImages.push({
-              type: f.type,
-              data: base64,
+              type: result.type,
+              data: result.base64,
               name: 'paste-' + Date.now() + '.png',
               size: f.size,
             });
             renderAttachmentPreview();
             updateSendBtnState();
-          };
-          reader.readAsDataURL(f);
+          }).catch(function(err) {
+            console.error('粘贴图片压缩失败:', err);
+          });
         })(file);
       }
     }
@@ -2868,18 +2933,25 @@ function handleDropEvent(e) {
   if (!dt || !dt.files || dt.files.length === 0) return;
   for (var i = 0; i < dt.files.length; i++) {
     (function(file) {
-      var reader = new FileReader();
-      reader.onload = function(ev) {
-        var base64 = ev.target.result.split(',')[1];
-        if (file.type.startsWith('image/')) {
-          window._pendingImages.push({ type: file.type, data: base64, name: file.name, size: file.size });
-        } else {
+      if (file.type.startsWith('image/')) {
+        // [v1.16.13] 拖放图片使用压缩
+        compressImage(file).then(function(result) {
+          window._pendingImages.push({ type: result.type, data: result.base64, name: file.name, size: file.size });
+          renderAttachmentPreview();
+          updateSendBtnState();
+        }).catch(function(err) {
+          console.error('拖放图片压缩失败:', err);
+        });
+      } else {
+        var reader = new FileReader();
+        reader.onload = function(ev) {
+          var base64 = ev.target.result.split(',')[1];
           window._pendingFiles.push({ type: file.type, data: base64, name: file.name, size: file.size });
-        }
-        renderAttachmentPreview();
-        updateSendBtnState();
-      };
-      reader.readAsDataURL(file);
+          renderAttachmentPreview();
+          updateSendBtnState();
+        };
+        reader.readAsDataURL(file);
+      }
     })(dt.files[i]);
   }
 }