myagent-ai 1.10.4 → 1.10.6

package/agents/main_agent.py

@@ -53,6 +53,8 @@ class MainAgent(BaseAgent):
 <get_knowledge>下一轮执行时需要从知识库搜索获得的知识，填写检索关键词或描述。如context中已包含充足的<knowledge>内容，则为空。如需更多专业知识支撑，则填写相关搜索词。</get_knowledge>
 <askuser>需要询问用户的内容，如无，则为空</askuser>
 <finish>true/false，是否结束循环调用llm。如"askuser"为非空，则"finish"输出true。否则，根据"context"判断任务是否已完成，是否结束llm回调</finish>
+<finish_reason>当 finish=true 时必填，详细说明为什么现在结束任务（如：任务已完成/需要用户补充信息/信息不足无法继续等）。finish=false 时为空。</finish_reason>
+<next_step>当 finish=false 时必填，描述下一步计划做什么（简洁明了，1-2句话）。finish=true 时为空。</next_step>
 
 </output>
 
@@ -70,7 +72,9 @@ class MainAgent(BaseAgent):
 11. <get_knowledge>: 如果当前 <knowledge> 内容不足以完成任务，填写需要从知识库搜索的关键词；否则为空
 12. <askuser>: 当信息不足需要用户补充时，在此填写要问的问题
 13. <finish>: 当任务已完成或需要等待用户回应时为 true；否则为 false 继续执行
-14. 使用中文输出所有内容
+14. <finish_reason>: **finish=true 时必须填写**，详细说明结束原因（任务完成/等待用户/信息不足/无法处理等）
+15. <next_step>: **finish=false 时必须填写**，描述下一步计划做什么，要求简洁明确（1-2句话）
+16. 使用中文输出所有内容
 
 ## 工具选择指南
 - **搜索信息**: 用 `web_search`（返回标题+URL+摘要），不要用 browser_open
@@ -420,7 +424,7 @@ class MainAgent(BaseAgent):
             if all_tool_outputs:
                 messages.append(Message(
                     role="user",
-                    content=f"[上一轮工具执行结果汇总]\n{truncate_str(all_tool_outputs, 15000)}"
+                    content=f"[上一轮工具执行结果汇总]\n{truncate_str(all_tool_outputs, 30000)}"
                 ))
                 all_tool_outputs = ""
 
@@ -464,6 +468,8 @@ class MainAgent(BaseAgent):
                     "remember": truncate_str(parsed.remember, 200),
                     "ask_user": truncate_str(parsed.ask_user, 200),
                     "finish": parsed.finish,
+                    "finish_reason": truncate_str(parsed.finish_reason, 200),
+                    "next_step": truncate_str(parsed.next_step, 200),
                     "parse_success": parsed.parse_success,
                 }},
                 stream_callback,
@@ -700,20 +706,59 @@ class MainAgent(BaseAgent):
 
                 # 发送工具结果事件
                 # 提取实际输出：SkillResult 有 output/message/data，ExecResult 有 stdout/stderr
+                def _format_data_for_llm(data):
+                    """将结构化 data 格式化为 LLM 可读的文本"""
+                    if data is None:
+                        return ""
+                    if isinstance(data, str):
+                        return data
+                    if isinstance(data, dict):
+                        # 搜索结果列表格式 (web_search)
+                        results = data.get("results")
+                        if isinstance(results, list):
+                            lines = []
+                            for i, r in enumerate(results, 1):
+                                title = r.get("title", "")
+                                url = r.get("url", "")
+                                snippet = r.get("snippet", "")
+                                lines.append(f"{i}. {title}\n   URL: {url}\n   {snippet}")
+                            return "\n".join(lines)
+                        # 网页内容格式 (web_read)
+                        if "url" in data and "content" in data:
+                            title = data.get("title", "")
+                            content = data.get("content", "")
+                            lines = [f"标题: {title}", f"URL: {data['url']}", f"内容:\n{content}"]
+                            return "\n".join(lines)
+                        # 通用 dict: key-value 格式
+                        parts = []
+                        for k, v in data.items():
+                            if k == "results":
+                                continue  # 已在上面处理
+                            parts.append(f"{k}: {v}")
+                        return "\n".join(parts) if parts else str(data)
+                    if isinstance(data, list):
+                        return "\n".join(str(item) for item in data)
+                    return str(data)
+
                 def _extract_tool_output(tr):
                     """从工具结果中提取实际输出文本"""
+                    # 优先使用 output 字段 (技能明确设置的输出)
                     out = tr.get("output", "")
                     if out:
                         return out
+                    # 如果 output 为空，尝试智能格式化 data 字段
+                    data = tr.get("data")
+                    if data is not None:
+                        formatted = _format_data_for_llm(data)
+                        if formatted:
+                            return formatted
+                    # 降级到 message / stdout / error
                     out = tr.get("message", "")
                     if out:
                         return out
                     out = tr.get("stdout", "")
                     if out:
                         return out
-                    data = tr.get("data")
-                    if data is not None:
-                        return str(data) if not isinstance(data, str) else data
                     return tr.get("error", "")
 
                 tool_output_text = _extract_tool_output(tool_result)
@@ -745,16 +790,18 @@ class MainAgent(BaseAgent):
                     need_callback = True
 
                 output_str = tool_output_text
+                # 搜索和网页读取类工具允许更长的输出
+                _max_output = 6000 if tool_name in ("web_search", "web_read", "url_read") else 3000
                 tool_outputs_parts.append(
                     f"### {before_call}\n"
                     f"**工具**: {tool_name}\n"
                     f"**结果**: {'成功' if tool_result.get('success') else '失败'}\n"
-                    f"{truncate_str(output_str, 2000)}\n"
+                    f"{truncate_str(output_str, _max_output)}\n"
                 )
 
                 conversation_history.append(Message(
                     role="assistant",
-                    content=f"执行工具 {tool_name}:\n{truncate_str(output_str, 3000)}",
+                    content=f"执行工具 {tool_name}:\n{truncate_str(output_str, _max_output)}",
                 ))
                 conversation_history.append(Message(
                     role="user",

package/core/output_parser.py

@@ -86,6 +86,8 @@ class ParsedOutput:
         get_knowledge: Knowledge search keywords for the next loop iteration.
             The ContextBuilder will use this to perform RAG retrieval.
         finish: When ``True`` the execution loop should terminate.
+        finish_reason: When finish=True, explains why the task is ending.
+        next_step: When finish=False, describes what to do next.
         raw_text: The verbatim raw text returned by the LLM.
         parse_success: Whether the XML was parsed successfully (``True``)
             or the regex fallback was used (``False``).
@@ -99,6 +101,8 @@ class ParsedOutput:
     ask_user: str = ""
     get_knowledge: str = ""
     finish: bool = False
+    finish_reason: str = ""
+    next_step: str = ""
     response: str = ""  # 模型对用户的直接回复（友好自然的话语）
     raw_text: str = ""
     parse_success: bool = False
@@ -363,6 +367,8 @@ def _parse_xml_content(xml_content: str) -> ParsedOutput:
     parsed.ask_user = _safe_strip(root.findtext("askuser"))
     parsed.get_knowledge = _safe_strip(root.findtext("get_knowledge"))
     parsed.finish = _parse_bool(root.findtext("finish"), False)
+    parsed.finish_reason = _safe_strip(root.findtext("finish_reason"))
+    parsed.next_step = _safe_strip(root.findtext("next_step"))
     parsed.response = _safe_strip(root.findtext("response"))
 
     return parsed
@@ -393,6 +399,8 @@ def _fallback_regex_parse(raw_text: str) -> ParsedOutput:
     parsed.ask_user = _safe_strip(tag_map.get("askuser"))
     parsed.get_knowledge = _safe_strip(tag_map.get("get_knowledge"))
     parsed.finish = _parse_bool(tag_map.get("finish"), False)
+    parsed.finish_reason = _safe_strip(tag_map.get("finish_reason"))
+    parsed.next_step = _safe_strip(tag_map.get("next_step"))
     parsed.response = _safe_strip(tag_map.get("response"))
 
     # For toolstocal we attempt to find individual <tool> blocks.
@@ -482,6 +490,8 @@ def parse_output(raw_text: str) -> ParsedOutput:
             parsed.ask_user = _safe_strip(root.findtext("askuser"))
             parsed.get_knowledge = _safe_strip(root.findtext("get_knowledge"))
             parsed.finish = _parse_bool(root.findtext("finish"), False)
+            parsed.finish_reason = _safe_strip(root.findtext("finish_reason"))
+            parsed.next_step = _safe_strip(root.findtext("next_step"))
             parsed.response = _safe_strip(root.findtext("response"))
             return parsed
     except ET.ParseError:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "myagent-ai",
-  "version": "1.10.4",
+  "version": "1.10.6",
   "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

@@ -2357,6 +2357,9 @@ class ApiServer:
                 pass
         return web.json_response({**agent_info, "sessions": sessions})
 
+    # Internal keys that should not appear in chat history UI
+    _HIDDEN_KEYS = {"llm_output", "tool_call", "tool_result"}
+
     async def handle_get_messages(self, request):
         sid = request.match_info["sid"]
         if not self.core.memory: return web.json_response([])
@@ -2364,6 +2367,8 @@ class ApiServer:
         offset = int(request.query.get("offset", 0))
         entries = self.core.memory.get_conversation(sid, limit=limit + offset)
         entries = entries[offset:]
+        # Filter out internal entries (LLM raw output, tool calls/results)
+        entries = [e for e in entries if (e.key or "") not in self._HIDDEN_KEYS]
         return web.json_response([{"role": e.role, "content": e.content, "time": e.created_at, "key": e.key or ""} for e in entries])
 
     async def handle_get_messages_query(self, request):
@@ -2376,6 +2381,8 @@ class ApiServer:
         offset = int(request.query.get("offset", 0))
         entries = self.core.memory.get_conversation(sid, limit=limit + offset)
         entries = entries[offset:]
+        # Filter out internal entries (LLM raw output, tool calls/results)
+        entries = [e for e in entries if (e.key or "") not in self._HIDDEN_KEYS]
         return web.json_response([{"role": e.role, "content": e.content, "time": e.created_at, "key": e.key or ""} for e in entries])
 
     async def handle_delete_session(self, request):

package/web/ui/chat/chat_main.js

@@ -1765,9 +1765,14 @@ async function selectSession(id) {
     const loaded = (Array.isArray(data) ? data : []).filter(function(m) {
       return m && (m.role === 'user' || m.role === 'assistant' || m.role === 'tool');
     }).map(function(m) {
+      var content = (m.content != null) ? String(m.content) : '';
+      // Strip XML tags from assistant messages (backend may store raw LLM XML output)
+      if (m.role === 'assistant' && content && content.trim().startsWith('<')) {
+        content = (typeof _stripXmlTags === 'function') ? _stripXmlTags(content) : content;
+      }
       return {
         role: m.role || 'assistant',
-        content: (m.content != null) ? String(m.content) : '',
+        content: content,
         time: m.time || m.created_at || '',
         key: m.key || '',
       };
@@ -1824,9 +1829,14 @@ async function loadMoreMessages() {
     const loaded = data.filter(function(m) {
       return m && (m.role === 'user' || m.role === 'assistant' || m.role === 'tool');
     }).map(function(m) {
+      var content = (m.content != null) ? String(m.content) : '';
+      // Strip XML tags from assistant messages (backend may store raw LLM XML output)
+      if (m.role === 'assistant' && content && content.trim().startsWith('<')) {
+        content = (typeof _stripXmlTags === 'function') ? _stripXmlTags(content) : content;
+      }
       return {
         role: m.role || 'assistant',
-        content: (m.content != null) ? String(m.content) : '',
+        content: content,
         time: m.time || m.created_at || '',
       };
     });

package/web/ui/chat/flow_engine.js

@@ -305,22 +305,25 @@ function updateStreamingMessage(msgIdx) {
   const container = document.getElementById('messagesInner');
   if (!container) return;
 
-  // Find or create the streaming message row
-  const rows = container.querySelectorAll('.message-row.assistant');
+  // Find the streaming message row by counting ALL message rows (not just assistant)
+  // msgIdx is the global index in state.messages, so we count all rows to match
+  const allRows = container.querySelectorAll('.message-row');
   let targetRow = null;
-  // Count assistant rows to find the right one
-  let assistantCount = 0;
-  for (const row of rows) {
-    assistantCount++;
-    if (assistantCount === msgIdx + 1) {
+  let rowCount = 0;
+  for (const row of allRows) {
+    if (rowCount === msgIdx) {
       targetRow = row;
       break;
     }
+    rowCount++;
   }
 
   // Fallback: if we can't find by count, use last assistant row
-  if (!targetRow && rows.length > 0) {
-    targetRow = rows[rows.length - 1];
+  if (!targetRow) {
+    const assistantRows = container.querySelectorAll('.message-row.assistant');
+    if (assistantRows.length > 0) {
+      targetRow = assistantRows[assistantRows.length - 1];
+    }
   }
 
   if (!targetRow) {
@@ -343,12 +346,32 @@ function updateStreamingMessage(msgIdx) {
     }
   }
   if (msg.reasoning) {
-    // Count words/chars during streaming for live counter
     const reasoningLen = msg.reasoning.length;
     const reasoningWordCount = msg.streaming
       ? '<span class="thought-word-count">' + reasoningLen + ' 字</span>'
       : '';
-    const reasoningHtml = `<details class="thought-block ${msg.streaming ? 'streaming' : ''}" ${msg.streaming ? 'open' : ''}>
+    if (reasoningDetails) {
+      // Incremental update: only update word count and badge, append new text to content
+      const label = reasoningDetails.querySelector('.thought-label');
+      if (label) label.innerHTML = '模型推理过程' + reasoningWordCount;
+      const badge = reasoningDetails.querySelector('.thought-badge');
+      if (badge) badge.textContent = msg.streaming ? '推理中...' : '已完成';
+      // Incremental text append for streaming (avoid full markdown rebuild)
+      const thoughtContent = reasoningDetails.querySelector('.thought-content');
+      if (thoughtContent && msg.streaming) {
+        const prevLen = reasoningDetails._lastReasoningLen || 0;
+        if (msg.reasoning.length > prevLen) {
+          const newText = msg.reasoning.substring(prevLen);
+          thoughtContent.insertAdjacentHTML('beforeend', renderMarkdown(newText));
+          reasoningDetails._lastReasoningLen = msg.reasoning.length;
+        }
+      } else if (thoughtContent && !msg.streaming) {
+        // Final render once streaming stops
+        thoughtContent.innerHTML = renderMarkdown(msg.reasoning);
+        reasoningDetails._lastReasoningLen = msg.reasoning.length;
+      }
+    } else {
+      const reasoningHtml = `<details class="thought-block ${msg.streaming ? 'streaming' : ''}" ${msg.streaming ? 'open' : ''}>
         <summary>
           <span class="thought-icon">💡</span>
           <span class="thought-label">模型推理过程${reasoningWordCount}</span>
@@ -356,10 +379,10 @@ function updateStreamingMessage(msgIdx) {
         </summary>
         <div class="thought-content">${renderMarkdown(msg.reasoning)}</div>
       </details>`;
-    if (reasoningDetails) {
-      reasoningDetails.outerHTML = reasoningHtml;
-    } else {
       contentArea.insertAdjacentHTML('afterbegin', reasoningHtml);
+      // Set initial length tracking
+      const newBlock = contentArea.querySelector(':scope > .thought-block');
+      if (newBlock) newBlock._lastReasoningLen = msg.reasoning.length;
     }
   }
 
@@ -378,7 +401,26 @@ function updateStreamingMessage(msgIdx) {
     const thoughtWordCount = msg.streaming
       ? '<span class="thought-word-count">' + thoughtLen + ' 字</span>'
       : '';
-    const thoughtHtml = `<details class="thought-block ${msg.streaming ? 'streaming' : ''}" ${msg.streaming ? 'open' : ''}>
+    if (thoughtBlock) {
+      // Incremental update for thought block too
+      const label = thoughtBlock.querySelector('.thought-label');
+      if (label) label.innerHTML = 'Agent 思考过程' + thoughtWordCount;
+      const badge = thoughtBlock.querySelector('.thought-badge');
+      if (badge) badge.textContent = msg.streaming ? '思考中...' : '已完成';
+      const thoughtContent = thoughtBlock.querySelector('.thought-content');
+      if (thoughtContent && msg.streaming) {
+        const prevLen = thoughtBlock._lastThoughtLen || 0;
+        if (msg.thought.length > prevLen) {
+          const newText = msg.thought.substring(prevLen);
+          thoughtContent.insertAdjacentHTML('beforeend', renderMarkdown(newText));
+          thoughtBlock._lastThoughtLen = msg.thought.length;
+        }
+      } else if (thoughtContent && !msg.streaming) {
+        thoughtContent.innerHTML = renderMarkdown(msg.thought);
+        thoughtBlock._lastThoughtLen = msg.thought.length;
+      }
+    } else {
+      const thoughtHtml = `<details class="thought-block ${msg.streaming ? 'streaming' : ''}" ${msg.streaming ? 'open' : ''}>
         <summary>
           <span class="thought-icon">💭</span>
           <span class="thought-label">Agent 思考过程${thoughtWordCount}</span>
@@ -386,16 +428,14 @@ function updateStreamingMessage(msgIdx) {
         </summary>
         <div class="thought-content">${renderMarkdown(msg.thought)}</div>
       </details>`;
-    if (thoughtBlock) {
-      thoughtBlock.outerHTML = thoughtHtml;
-    } else {
-      // Insert after reasoning block if exists, otherwise at beginning
       const existingReasoning = contentArea.querySelectorAll('.thought-block');
       if (existingReasoning.length > 0) {
         existingReasoning[existingReasoning.length - 1].insertAdjacentHTML('afterend', thoughtHtml);
       } else {
         contentArea.insertAdjacentHTML('afterbegin', thoughtHtml);
       }
+      const newBlock = contentArea.querySelectorAll('.thought-block');
+      if (newBlock.length > 0) newBlock[newBlock.length - 1]._lastThoughtLen = msg.thought.length;
     }
   }
 
@@ -417,7 +457,26 @@ function updateStreamingMessage(msgIdx) {
       const v2WordCount = msg.streaming
         ? '<span class="thought-word-count">' + v2Len + ' 字</span>'
         : '';
-      const v2Html = `<details class="thought-block ${msg.streaming ? 'streaming' : ''}" ${msg.streaming ? 'open' : ''}>
+      if (v2ReasoningBlock) {
+        // Incremental update for V2 reasoning block
+        const label = v2ReasoningBlock.querySelector('.thought-label');
+        if (label) label.innerHTML = 'V2 推理过程' + v2WordCount;
+        const badge = v2ReasoningBlock.querySelector('.thought-badge');
+        if (badge) badge.textContent = msg.streaming ? '推理中...' : '已完成';
+        const thoughtContent = v2ReasoningBlock.querySelector('.thought-content');
+        if (thoughtContent && msg.streaming) {
+          const prevLen = v2ReasoningBlock._lastV2Len || 0;
+          if (msg._v2Reasoning.length > prevLen) {
+            const newText = msg._v2Reasoning.substring(prevLen);
+            thoughtContent.insertAdjacentHTML('beforeend', renderMarkdown(newText));
+            v2ReasoningBlock._lastV2Len = msg._v2Reasoning.length;
+          }
+        } else if (thoughtContent && !msg.streaming) {
+          thoughtContent.innerHTML = renderMarkdown(msg._v2Reasoning);
+          v2ReasoningBlock._lastV2Len = msg._v2Reasoning.length;
+        }
+      } else if (!msg.thought) {
+        const v2Html = `<details class="thought-block ${msg.streaming ? 'streaming' : ''}" ${msg.streaming ? 'open' : ''}>
           <summary>
             <span class="thought-icon">🧠</span>
             <span class="thought-label">V2 推理过程${v2WordCount}</span>
@@ -425,10 +484,9 @@ function updateStreamingMessage(msgIdx) {
           </summary>
           <div class="thought-content">${renderMarkdown(msg._v2Reasoning)}</div>
         </details>`;
-      if (v2ReasoningBlock) {
-        v2ReasoningBlock.outerHTML = v2Html;
-      } else if (!msg.thought) {
         contentArea.insertAdjacentHTML('afterbegin', v2Html);
+        const newBlock = contentArea.querySelector(':scope > .thought-block');
+        if (newBlock) newBlock._lastV2Len = msg._v2Reasoning.length;
       }
     }
   }