myagent-ai 1.6.5 → 1.6.7

package/agents/base.py

@@ -145,6 +145,7 @@ class BaseAgent(ABC):
         request_kwargs.update(kwargs)
 
         full_text = ""
+        full_reasoning = ""  # Track reasoning tokens (for o1/o3/DeepSeek-R1 etc.)
         tool_calls_acc: Dict[int, Dict] = {}  # index -> {id, name, arguments_str}
         finish_reason = ""
 
@@ -166,6 +167,12 @@ class BaseAgent(ABC):
             else:
                 await _write_sse({"type": "text_delta", "content": delta_text})
 
+        async def _emit_reasoning_delta(delta_text: str):
+            """处理推理 token：发送 reasoning_delta SSE 事件"""
+            nonlocal full_reasoning
+            full_reasoning += delta_text
+            await _write_sse({"type": "reasoning_delta", "content": delta_text})
+
         try:
             if self.llm.provider in self.llm._OPENAI_COMPATIBLE_PROVIDERS or self.llm.provider == "zhipu":
                 # 使用异步客户端流式
@@ -190,6 +197,11 @@ class BaseAgent(ABC):
                     if delta.content:
                         await _emit_text_delta(delta.content)
 
+                    # Handle reasoning_content (OpenAI o1/o3, DeepSeek-R1, Qwen-QwQ etc.)
+                    reasoning_content = getattr(delta, 'reasoning_content', None) or getattr(delta, 'reasoning', None)
+                    if reasoning_content:
+                        await _emit_reasoning_delta(reasoning_content)
+
                     # Handle tool_call deltas (accumulate)
                     if hasattr(delta, 'tool_calls') and delta.tool_calls:
                         for tc_delta in delta.tool_calls:
@@ -252,8 +264,17 @@ class BaseAgent(ABC):
                     if event is None:
                         break
                     if event.type == "content_block_delta":
-                        if hasattr(event.delta, "text"):
+                        # Handle text content
+                        if hasattr(event.delta, "text") and event.delta.text:
                             await _emit_text_delta(event.delta.text)
+                        # Handle extended thinking (Anthropic reasoning)
+                        if hasattr(event.delta, "thinking") and event.delta.thinking:
+                            await _emit_reasoning_delta(event.delta.thinking)
+                        # Handle thinking delta via type
+                        if hasattr(event.delta, "type") and event.delta.type == "thinking_delta":
+                            thinking_text = getattr(event.delta, "thinking", "")
+                            if thinking_text:
+                                await _emit_reasoning_delta(thinking_text)
                     elif event.type == "message_stop":
                         finish_reason = "stop"
 
@@ -292,9 +313,14 @@ class BaseAgent(ABC):
                         break
                     try:
                         data = json.loads(line.decode('utf-8') if isinstance(line, bytes) else line)
-                        content = data.get("message", {}).get("content", "")
+                        message = data.get("message", {})
+                        content = message.get("content", "")
                         if content:
                             await _emit_text_delta(content)
+                        # Handle Ollama thinking/reasoning (e.g., DeepSeek-R1 via Ollama)
+                        thinking = message.get("thinking", "")
+                        if thinking:
+                            await _emit_reasoning_delta(thinking)
                         if data.get("done"):
                             finish_reason = "stop"
                             # Record usage from Ollama

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "myagent-ai",
-  "version": "1.6.5",
+  "version": "1.6.7",
   "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": {

package/web/api_server.py

@@ -592,7 +592,16 @@ class ApiServer:
                     agent_path=agent_path, agent_system_prompt=agent_system_prompt,
                     chat_mode=chat_mode, stream_response=response,
                 )
+            elif self.core.main_agent and self.core.llm:
+                # model_chain is empty (no model_id configured), but LLM is available
+                # Use _stream_process_message for true token-by-token streaming
+                full_response = await self._stream_process_message(
+                    clean_message, session_id, response,
+                    agent_path=agent_path, agent_system_prompt=agent_system_prompt,
+                    chat_mode=chat_mode,
+                )
             else:
+                # No LLM at all — non-streaming fallback
                 full_response = await self.core.process_message(clean_message, session_id)
                 await response.write(("data: " + json.dumps({"type": "text", "content": full_response}) + "\n\n").encode())
 
@@ -640,6 +649,12 @@ class ApiServer:
                         agent_path=agent_path, agent_system_prompt=agent_system_prompt_q,
                         chat_mode=chat_mode, stream_response=response,
                     )
+                elif self.core.main_agent and self.core.llm:
+                    full_response = await self._stream_process_message(
+                        clean_message_q, session_id, response,
+                        agent_path=agent_path, agent_system_prompt=agent_system_prompt_q,
+                        chat_mode=chat_mode,
+                    )
                 else:
                     full_response = await self.core.process_message(clean_message_q, session_id)
                     await response.write(("data: " + json.dumps({"type": "text", "content": full_response}) + "\n\n").encode())
@@ -3870,6 +3885,7 @@ class ApiServer:
                 # 获取agent的显示信息
                 avatar = "🤖"
                 display_name = agent_path
+                agent_color = "var(--accent)"
                 if agent_cfg:
                     avatar = agent_cfg.get("avatar_emoji", "🤖") or "🤖"
                     display_name = agent_cfg.get("name", agent_path)
@@ -3879,6 +3895,7 @@ class ApiServer:
                     "agent_path": agent_path,
                     "name": display_name,
                     "avatar": avatar,
+                    "agent_color": agent_color,
                     "response": response,
                 }
             except Exception as e:

package/web/ui/chat.html

@@ -3351,6 +3351,17 @@ function renderMessages() {
         <div class="thought-content">${renderMarkdown(msg.thought)}</div>
       </details>`;
     })() : '';
+    const reasoningHtml = msg.reasoning ? (() => {
+      const isStreaming = !!msg.streaming;
+      return `<details class="thought-block ${isStreaming ? 'streaming' : ''}" ${isStreaming ? 'open' : ''}>
+        <summary>
+          <span class="thought-icon">💡</span>
+          <span class="thought-label">模型推理过程</span>
+          ${isStreaming ? '<span class="thought-badge">推理中...</span>' : '<span class="thought-badge">已完成</span>'}
+        </summary>
+        <div class="thought-content">${renderMarkdown(msg.reasoning)}</div>
+      </details>`;
+    })() : '';
     const actionBtns = (!isUser && msg.content) ? `
       <div class="msg-actions">
         <button class="msg-action-btn" onclick="copyMessage(${i})" title="复制">
@@ -3376,6 +3387,7 @@ function renderMessages() {
       <div class="message-row ${msg.role}">
         <div class="message-avatar">${avatar}</div>
         <div style="flex:1;min-width:0">
+          ${reasoningHtml}
           ${thoughtHtml}
           ${content || streamingIndicator ? `<div class="message-bubble">${content}${ttsIndicator}</div>` : ''}
           ${streamingIndicator}
@@ -3393,6 +3405,165 @@ function renderMessages() {
   scrollToBottom();
 }
 
+// ── Incremental Streaming Update (avoid full DOM rebuild on every token) ──
+function updateStreamingMessage(msgIdx) {
+  // Only update the last streaming message bubble without rebuilding the entire DOM
+  const msg = state.messages[msgIdx];
+  if (!msg || msg.role !== 'assistant') return;
+
+  const container = document.getElementById('messagesInner');
+  if (!container) return;
+
+  // Find or create the streaming message row
+  const rows = container.querySelectorAll('.message-row.assistant');
+  let targetRow = null;
+  // Count assistant rows to find the right one
+  let assistantCount = 0;
+  for (const row of rows) {
+    assistantCount++;
+    if (assistantCount === msgIdx + 1) {
+      targetRow = row;
+      break;
+    }
+  }
+
+  // Fallback: if we can't find by count, use last assistant row
+  if (!targetRow && rows.length > 0) {
+    targetRow = rows[rows.length - 1];
+  }
+
+  if (!targetRow) {
+    // Message row doesn't exist yet, fall back to full render
+    renderMessages();
+    return;
+  }
+
+  const contentArea = targetRow.querySelector(':scope > div');
+  if (!contentArea) return;
+
+  // Update reasoning block (model inference/reasoning - e.g. o1, DeepSeek-R1)
+  let reasoningBlock = contentArea.querySelector('.thought-block .thought-label');
+  // Find reasoning block by looking for "模型推理过程" label
+  let reasoningDetails = null;
+  const allThoughtBlocks = contentArea.querySelectorAll('.thought-block');
+  for (const tb of allThoughtBlocks) {
+    const label = tb.querySelector('.thought-label');
+    if (label && label.textContent.includes('模型推理过程')) {
+      reasoningDetails = tb;
+      break;
+    }
+  }
+  if (msg.reasoning) {
+    const reasoningHtml = `<details class="thought-block ${msg.streaming ? 'streaming' : ''}" ${msg.streaming ? 'open' : ''}>
+        <summary>
+          <span class="thought-icon">💡</span>
+          <span class="thought-label">模型推理过程</span>
+          ${msg.streaming ? '<span class="thought-badge">推理中...</span>' : '<span class="thought-badge">已完成</span>'}
+        </summary>
+        <div class="thought-content">${renderMarkdown(msg.reasoning)}</div>
+      </details>`;
+    if (reasoningDetails) {
+      reasoningDetails.outerHTML = reasoningHtml;
+    } else {
+      contentArea.insertAdjacentHTML('afterbegin', reasoningHtml);
+    }
+  }
+
+  // Update thought block (agent thinking)
+  let thoughtBlock = null;
+  const updatedBlocks = contentArea.querySelectorAll('.thought-block');
+  for (const tb of updatedBlocks) {
+    const label = tb.querySelector('.thought-label');
+    if (label && label.textContent.includes('Agent 思考过程')) {
+      thoughtBlock = tb;
+      break;
+    }
+  }
+  if (msg.thought) {
+    const thoughtHtml = `<details class="thought-block ${msg.streaming ? 'streaming' : ''}" ${msg.streaming ? 'open' : ''}>
+        <summary>
+          <span class="thought-icon">🧠</span>
+          <span class="thought-label">Agent 思考过程</span>
+          ${msg.streaming ? '<span class="thought-badge">思考中...</span>' : '<span class="thought-badge">已完成</span>'}
+        </summary>
+        <div class="thought-content">${renderMarkdown(msg.thought)}</div>
+      </details>`;
+    if (thoughtBlock) {
+      thoughtBlock.outerHTML = thoughtHtml;
+    } else {
+      contentArea.insertAdjacentHTML('afterbegin', thoughtHtml);
+    }
+  } else if (thoughtBlock && !msg.thought) {
+    thoughtBlock.remove();
+  }
+
+  // Update message bubble content
+  const bubble = contentArea.querySelector('.message-bubble');
+  const content = msg.content ? renderMarkdown(msg.content) : '';
+  const streamingIndicator = msg.streaming && !msg.content && !msg.thought ? `
+      <div class="streaming-indicator">
+        <div class="spinner"></div>
+        <div class="streaming-dots">
+          <span class="dot"></span><span class="dot"></span><span class="dot"></span>
+        </div>
+        <span style="font-weight:500">Agent 正在思考...</span>
+      </div>` : '';
+
+  if (bubble) {
+    if (content) {
+      bubble.innerHTML = content;
+      bubble.style.display = '';
+    } else {
+      bubble.style.display = 'none';
+    }
+  } else if (content) {
+    contentArea.insertAdjacentHTML('afterbegin', `<div class="message-bubble">${content}</div>`);
+  }
+
+  // Update streaming indicator
+  let indicator = contentArea.querySelector('.streaming-indicator');
+  if (streamingIndicator) {
+    if (!indicator) {
+      contentArea.insertAdjacentHTML('beforeend', streamingIndicator);
+    }
+  } else if (indicator) {
+    indicator.remove();
+  }
+
+  // Update exec events
+  const existingPanel = contentArea.querySelector('.exec-events-panel');
+  if (msg.exec_events && msg.exec_events.length > 0) {
+    const execHtml = renderExecEvents(msg.exec_events, msgIdx);
+    if (existingPanel) {
+      existingPanel.outerHTML = execHtml;
+    } else {
+      contentArea.insertAdjacentHTML('beforeend', execHtml);
+    }
+  }
+
+  // Auto-scroll
+  scrollToBottom();
+}
+
+// Throttle streaming updates (max ~20fps to avoid excessive reflows)
+const _streamThrottle = { last: 0, pending: false, idx: -1 };
+function throttledStreamUpdate(msgIdx) {
+  _streamThrottle.idx = msgIdx;
+  if (_streamThrottle.pending) return;
+  const now = performance.now();
+  if (now - _streamThrottle.last < 50) {
+    _streamThrottle.pending = true;
+    requestAnimationFrame(() => {
+      _streamThrottle.pending = false;
+      _streamThrottle.last = performance.now();
+      updateStreamingMessage(_streamThrottle.idx);
+    });
+    return;
+  }
+  _streamThrottle.last = now;
+  updateStreamingMessage(msgIdx);
+}
+
 // ── Execution Events Rendering ──
 function renderExecEvents(events, msgIndex) {
   // 过滤出有意义的事件（仅展示调用/结果对，跳过中间状态）
@@ -3789,6 +3960,8 @@ async function sendMessage() {
           
           if (evt.type === 'session') {
             sessionIdReceived = evt.session_id;
+            // Sync the actual session ID (backend may prefix with agent_path)
+            state.activeSessionId = evt.session_id;
           } else if (evt.type === 'text') {
             fullResponse = evt.content;
             state.messages[msgIdx].content = evt.content;
@@ -3797,12 +3970,12 @@ async function sendMessage() {
             // Incremental streaming token
             fullResponse += evt.content;
             state.messages[msgIdx].content = fullResponse;
-            renderMessages();
+            throttledStreamUpdate(msgIdx);
           } else if (evt.type === 'thought_delta') {
             // Agent 思考过程增量文本（流式推送，单独显示）
             fullThought += evt.content;
             state.messages[msgIdx].thought = fullThought;
-            renderMessages();
+            throttledStreamUpdate(msgIdx);
           } else if (evt.type === 'thought') {
             // Agent 思考过程文本
             if (fullThought.trim() === '') {
@@ -3811,7 +3984,7 @@ async function sendMessage() {
               fullThought += '\n\n' + evt.content;
             }
             state.messages[msgIdx].thought = fullThought;
-            renderMessages();
+            throttledStreamUpdate(msgIdx);
           } else if (evt.type === 'queue_start') {
             // New message starting from queue
             if (state.messages[msgIdx]) {
@@ -3829,13 +4002,13 @@ async function sendMessage() {
             // Clear intermediate text from previous agent loop iterations
             fullResponse = '';
             state.messages[msgIdx].content = '';
-            renderMessages();
+            throttledStreamUpdate(msgIdx);
           } else if (evt.type === 'exec_event') {
             // Real-time execution event (tool call, code exec, skill result, etc.)
             execEventsReceived.push(evt.data);
             // 立即更新消息的 exec_events 并渲染
             state.messages[msgIdx].exec_events = [...execEventsReceived];
-            renderMessages();
+            throttledStreamUpdate(msgIdx);
           } else if (evt.type === 'task_plan_updated') {
             // 任务计划已更新，刷新侧边栏任务列表
             loadTaskPlan();
@@ -3845,6 +4018,15 @@ async function sendMessage() {
               execEventsReceived = evt.exec_events;
               state.messages[msgIdx].exec_events = [...execEventsReceived];
             }
+          } else if (evt.type === 'reasoning_delta') {
+            // 模型推理过程增量文本（OpenAI o1/o3/DeepSeek-R1 等推理模型）
+            if (!state.messages[msgIdx].reasoning) state.messages[msgIdx].reasoning = '';
+            state.messages[msgIdx].reasoning += evt.content;
+            throttledStreamUpdate(msgIdx);
+          } else if (evt.type === 'reasoning') {
+            // 模型推理完整文本
+            state.messages[msgIdx].reasoning = evt.content;
+            throttledStreamUpdate(msgIdx);
           } else if (evt.type === 'error') {
             fullResponse = '❌ ' + evt.error;
             state.messages[msgIdx].content = fullResponse;
@@ -4770,7 +4952,22 @@ async function selectGroup(gid) {
 
     // Load messages
     var msgsData = await getGroupMessages(gid);
-    groupMessages = msgsData || [];
+    // Normalize backend message format: map sender->role, sender_name->agent_name, sender_avatar->agent_emoji
+    groupMessages = (msgsData || []).map(function(m) {
+      if (m.sender) {
+        return {
+          role: m.sender === 'user' ? 'user' : 'assistant',
+          agent: m.agent_path || '',
+          agent_name: m.sender_name || '',
+          agent_emoji: m.sender_avatar || '🤖',
+          agent_color: m.agent_color || '',
+          content: m.content || '',
+          time: m.timestamp ? new Date(m.timestamp * 1000).toISOString() : (m.time || ''),
+          type: m.msg_type === 'system' ? 'system' : '',
+        };
+      }
+      return m;
+    });
     renderGroupMessages();
   } catch (e) {
     toast('加载群聊失败: ' + e.message, 'error');
@@ -4847,8 +5044,9 @@ function renderGroupMessages() {
       // Multiple agent responses (broadcast response)
       for (var j = 0; j < msg.responses.length; j++) {
         var r = msg.responses[j];
-        var rName = r.agent_name || r.agent || 'Agent';
-        var rEmoji = r.agent_emoji || '🤖';
+        // Backend returns "name" and "avatar" fields, with "agent_name"/"agent_emoji" as fallback
+        var rName = r.name || r.agent_name || r.agent || 'Agent';
+        var rEmoji = r.avatar || r.agent_emoji || '🤖';
         var rColor = r.agent_color || 'var(--accent)';
         html += '<div class="group-msg-row">'
           + '<div class="group-msg-avatar" style="background:' + rColor + ';color:#fff">' + rEmoji + '</div>'