entari-plugin-hyw 0.3.5__py3-none-any.whl → 4.0.0rc14__py3-none-any.whl

hyw_core/agent.py

@@ -0,0 +1,768 @@
+"""
+Agent Pipeline
+
+Tool-calling agent that can autonomously use web_tool to search/screenshot.
+Maximum 2 rounds of tool calls, up to 3 parallel calls per round.
+"""
+
+import asyncio
+import json
+import re
+import time
+from dataclasses import dataclass, field
+from typing import Any, Callable, Awaitable, Dict, List, Optional
+
+from loguru import logger
+from openai import AsyncOpenAI
+
+from .definitions import get_web_tool, get_refuse_answer_tool, AGENT_SYSTEM_PROMPT
+from .stages.base import StageContext, StageResult
+from .search import SearchService
+
+
+@dataclass
+class AgentSession:
+    """Agent session with tool call tracking."""
+    session_id: str
+    user_query: str
+    tool_calls: List[Dict[str, Any]] = field(default_factory=list)
+    tool_results: List[Dict[str, Any]] = field(default_factory=list)
+    conversation_history: List[Dict] = field(default_factory=list)
+    messages: List[Dict] = field(default_factory=list)  # LLM conversation
+    created_at: float = field(default_factory=time.time)
+    
+    # Round tracking (each round can have up to 3 parallel tool calls)
+    round_count: int = 0
+    
+    # Image tracking
+    user_image_count: int = 0  # Number of images from user input
+    total_image_count: int = 0  # Total images including web screenshots
+    
+    # Time tracking
+    search_time: float = 0.0  # Total time spent on search/screenshot
+    llm_time: float = 0.0  # Total time spent on LLM calls
+    first_llm_time: float = 0.0  # Time for first LLM call (understanding intent)
+    
+    # Usage tracking
+    usage_totals: Dict[str, int] = field(default_factory=lambda: {"input_tokens": 0, "output_tokens": 0})
+    
+    @property
+    def call_count(self) -> int:
+        """Total number of individual tool calls."""
+        return len(self.tool_calls)
+    
+    @property
+    def should_force_summary(self) -> bool:
+        """Force summary after 2 rounds of tool calls."""
+        return self.round_count >= 2
+
+
+def parse_filter_syntax(query: str, max_count: int = 3):
+    """
+    Parse enhanced filter syntax supporting:
+    - Chinese/English colons (: ：) and commas (, ，)
+    - Multiple filters: "mcmod=2, github=1 : xxx" 
+    - Index lists: "1, 2, 3 : xxx"
+    - Max total selections
+    
+    Returns:
+        filters: list of (filter_type, filter_value, count) tuples
+                 filter_type: 'index' or 'link'
+                 filter_value: int (for index) or str (for link match term)
+                 count: how many to get (default 1)
+        search_query: the actual search query
+        error_msg: error message if exceeded max
+    """
+    import re
+    
+    # Skip filter parsing if query contains URL (has :// pattern)
+    if re.search(r'https?://', query):
+        return [], query.strip(), None
+    
+    # Normalize colons
+    query = query.replace('：', ':')
+    
+    if ':' not in query:
+        return [], query.strip(), None
+    
+    parts = query.split(':', 1)
+    if len(parts) != 2:
+        return [], query.strip(), None
+    
+    filter_part = parts[0].strip()
+    search_query = parts[1].strip()
+    
+    if not filter_part or not search_query:
+        return [], query.strip(), None
+    
+    # Parse filter expressions
+    filters = []
+    total_count = 0
+    
+    # Normalize commas
+    filter_part = filter_part.replace('，', ',').replace('、', ',')
+    filter_items = [f.strip() for f in filter_part.split(',') if f.strip()]
+    
+    for item in filter_items:
+        # Check for "term=count" format (link filter)
+        if '=' in item:
+            term, count_str = item.split('=', 1)
+            term = term.strip().lower()
+            try:
+                count = int(count_str.strip())
+            except ValueError:
+                count = 1
+            if term and count > 0:
+                filters.append(('link', term, count))
+                total_count += count
+        # Check for pure number (index filter)
+        elif item.isdigit():
+            idx = int(item)
+            if 1 <= idx <= 10:
+                filters.append(('index', idx, 1))
+                total_count += 1
+    
+    if total_count > max_count:
+        return None, search_query, f"⚠️ 最多选择{max_count}个结果"
+    
+    return filters, search_query, None
+
+
+class AgentPipeline:
+    """
+    Tool-calling agent pipeline.
+    
+    Flow:
+    1. 用户输入 → LLM (with tools)
+    2. If tool_call: execute all tools in parallel → notify user with batched message → loop
+    3. If call_count >= 2 rounds: force summary on next call
+    4. Return final content
+    """
+    
+    MAX_TOOL_ROUNDS = 2  # Maximum rounds of tool calls
+    MAX_PARALLEL_TOOLS = 3  # Maximum parallel tool calls per round
+    MAX_LLM_RETRIES = 3  # Maximum retries for empty API responses
+    LLM_RETRY_DELAY = 1.0  # Delay between retries in seconds
+    
+    def __init__(
+        self, 
+        config: Any, 
+        search_service: SearchService,
+        send_func: Optional[Callable[[str], Awaitable[None]]] = None
+    ):
+        self.config = config
+        self.search_service = search_service
+        self.send_func = send_func
+        self.client = AsyncOpenAI(base_url=config.base_url, api_key=config.api_key)
+    
+    async def execute(
+        self,
+        user_input: str,
+        conversation_history: List[Dict],
+        images: List[str] = None,
+        model_name: str = None,
+    ) -> Dict[str, Any]:
+        """Execute agent with tool-calling loop."""
+        start_time = time.time()
+        
+        # Get model config
+        model_cfg = self.config.get_model_config("main")
+        model = model_name or model_cfg.model_name or self.config.model_name
+        
+        client = AsyncOpenAI(
+            base_url=model_cfg.base_url or self.config.base_url,
+            api_key=model_cfg.api_key or self.config.api_key
+        )
+        
+        # Create session
+        session = AgentSession(
+            session_id=str(time.time()),
+            user_query=user_input,
+            conversation_history=conversation_history.copy()
+        )
+        
+        # Create context for results
+        context = StageContext(
+            user_input=user_input,
+            images=images or [],
+            conversation_history=conversation_history,
+        )
+        
+        # Build initial messages
+        language = getattr(self.config, "language", "Simplified Chinese")
+        from datetime import datetime
+        current_time = datetime.now().strftime("%Y-%m-%d %H:%M")
+        system_prompt = AGENT_SYSTEM_PROMPT + f"\n\n用户要求的语言: {language}\n当前时间: {current_time}"
+        
+        # Build user content with images if provided
+        user_image_count = len(images) if images else 0
+        session.user_image_count = user_image_count
+        session.total_image_count = user_image_count
+        
+        if images:
+            user_content: List[Dict[str, Any]] = [{"type": "text", "text": user_input}]
+            for img_b64 in images:
+                url = f"data:image/jpeg;base64,{img_b64}" if not img_b64.startswith("data:") else img_b64
+                user_content.append({"type": "image_url", "image_url": {"url": url}})
+        else:
+            user_content = user_input
+        
+        session.messages = [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_content}
+        ]
+        
+        # Add image source hint for user images
+        if user_image_count > 0:
+            if user_image_count == 1:
+                hint = "第1张图片来自用户输入，请将这张图片作为用户输入的参考"
+            else:
+                hint = f"第1-{user_image_count}张图片来自用户输入，请将这{user_image_count}张图片作为用户输入的参考"
+            session.messages.append({"role": "system", "content": hint})
+        
+        # Tool definitions
+        web_tool = get_web_tool()
+        refuse_tool = get_refuse_answer_tool()
+        tools = [web_tool, refuse_tool]
+        
+        usage_totals = {"input_tokens": 0, "output_tokens": 0}
+        final_content = ""
+        
+        # Send initial status notification
+        if self.send_func:
+            try:
+                await self.send_func("💭 正在理解用户意图...")
+            except Exception as e:
+                logger.warning(f"AgentPipeline: Failed to send initial notification: {e}")
+        
+        # Agent loop
+        while True:
+            # Check if we need to force summary (no tools)
+            if session.should_force_summary:
+                logger.info(f"AgentPipeline: Max tool rounds ({self.MAX_TOOL_ROUNDS}) reached, forcing summary")
+                # Add context message about collected info
+                if context.web_results:
+                    context_msg = self._format_web_context(context)
+                    session.messages.append({
+                        "role": "system", 
+                        "content": f"你已经完成了{session.call_count}次工具调用。请基于已收集的信息给出最终回答。\n\n{context_msg}"
+                    })
+                
+                
+                # Final call without tools (with retry)
+                response = None
+                for retry in range(self.MAX_LLM_RETRIES):
+                    try:
+                        response = await client.chat.completions.create(
+                            model=model,
+                            messages=session.messages,
+                            temperature=self.config.temperature,
+                        )
+                        
+                        if response.usage:
+                            usage_totals["input_tokens"] += response.usage.prompt_tokens or 0
+                            usage_totals["output_tokens"] += response.usage.completion_tokens or 0
+                        
+                        # Check for valid response
+                        if response.choices:
+                            break  # Success, exit retry loop
+                        
+                        # Empty choices, retry
+                        logger.warning(f"AgentPipeline: Empty choices in force-summary (attempt {retry + 1}/{self.MAX_LLM_RETRIES}): {response}")
+                        if retry < self.MAX_LLM_RETRIES - 1:
+                            await asyncio.sleep(self.LLM_RETRY_DELAY)
+                    except Exception as e:
+                        logger.warning(f"AgentPipeline: LLM error (attempt {retry + 1}/{self.MAX_LLM_RETRIES}): {e}")
+                        if retry < self.MAX_LLM_RETRIES - 1:
+                            await asyncio.sleep(self.LLM_RETRY_DELAY)
+                        else:
+                            return {
+                                "llm_response": f"Error: {e}",
+                                "success": False,
+                                "error": str(e),
+                                "stats": {"total_time": time.time() - start_time}
+                            }
+                
+                # Final check after all retries
+                if not response or not response.choices:
+                    logger.error(f"AgentPipeline: All retries failed for force-summary")
+                    return {
+                        "llm_response": "抱歉，AI 服务返回了空响应，请稍后重试。",
+                        "success": False,
+                        "error": "Empty response from API after retries",
+                        "stats": {"total_time": time.time() - start_time},
+                        "usage": usage_totals,
+                    }
+                
+                final_content = response.choices[0].message.content or ""
+                break
+            
+            # Normal call with tools (with retry)
+            llm_start = time.time()
+            response = None
+            
+            for retry in range(self.MAX_LLM_RETRIES):
+                try:
+                    response = await client.chat.completions.create(
+                        model=model,
+                        messages=session.messages,
+                        temperature=self.config.temperature,
+                        tools=tools,
+                        tool_choice="auto",
+                    )
+                    
+                    # Check for valid response
+                    if response.choices:
+                        break  # Success, exit retry loop
+                    
+                    # Empty choices, retry
+                    logger.warning(f"AgentPipeline: Empty choices (attempt {retry + 1}/{self.MAX_LLM_RETRIES}): {response}")
+                    if retry < self.MAX_LLM_RETRIES - 1:
+                        await asyncio.sleep(self.LLM_RETRY_DELAY)
+                except Exception as e:
+                    logger.warning(f"AgentPipeline: LLM error (attempt {retry + 1}/{self.MAX_LLM_RETRIES}): {e}")
+                    if retry < self.MAX_LLM_RETRIES - 1:
+                        await asyncio.sleep(self.LLM_RETRY_DELAY)
+                    else:
+                        logger.error(f"AgentPipeline: All retries failed: {e}")
+                        return {
+                            "llm_response": f"Error: {e}",
+                            "success": False,
+                            "error": str(e),
+                            "stats": {"total_time": time.time() - start_time}
+                        }
+            
+            llm_duration = time.time() - llm_start
+            session.llm_time += llm_duration
+            
+            # Track first LLM call time (理解用户意图)
+            if session.call_count == 0 and session.first_llm_time == 0:
+                session.first_llm_time = llm_duration
+            
+            # Final check after all retries
+            if not response or not response.choices:
+                logger.error(f"AgentPipeline: All retries failed, empty choices")
+                return {
+                    "llm_response": "抱歉，AI 服务返回了空响应，请稍后重试。",
+                    "success": False,
+                    "error": "Empty response from API after retries",
+                    "stats": {"total_time": time.time() - start_time},
+                    "usage": usage_totals,
+                }
+            
+            if response.usage:
+                usage_totals["input_tokens"] += response.usage.prompt_tokens or 0
+                usage_totals["output_tokens"] += response.usage.completion_tokens or 0
+            
+            message = response.choices[0].message
+            
+            # Check for tool calls
+            if not message.tool_calls:
+                # Model chose to answer directly
+                final_content = message.content or ""
+                logger.info(f"AgentPipeline: Model answered directly after {session.call_count} tool calls")
+                break
+            
+            # Add assistant message with tool calls
+            session.messages.append({
+                "role": "assistant",
+                "content": message.content,
+                "tool_calls": [
+                    {
+                        "id": tc.id,
+                        "type": "function",
+                        "function": {"name": tc.function.name, "arguments": tc.function.arguments}
+                    }
+                    for tc in message.tool_calls
+                ]
+            })
+            
+            # Execute all tool calls in parallel
+            tool_tasks = []
+            tool_call_ids = []
+            tool_call_names = []
+            tool_call_args_list = []
+            
+            for tool_call in message.tool_calls:
+                tc_id = tool_call.id
+                func_name = tool_call.function.name
+                
+                try:
+                    args = json.loads(tool_call.function.arguments)
+                except json.JSONDecodeError:
+                    args = {}
+                
+                tool_call_ids.append(tc_id)
+                tool_call_names.append(func_name)
+                tool_call_args_list.append(args)
+                logger.info(f"AgentPipeline: Queueing tool '{func_name}' with args: {args}")
+            
+            # Check for refuse_answer first (handle immediately)
+            for idx, func_name in enumerate(tool_call_names):
+                if func_name == "refuse_answer":
+                    args = tool_call_args_list[idx]
+                    reason = args.get("reason", "Refused")
+                    context.should_refuse = True
+                    context.refuse_reason = reason
+                    
+                    session.messages.append({
+                        "role": "tool",
+                        "tool_call_id": tool_call_ids[idx],
+                        "content": f"已拒绝回答: {reason}"
+                    })
+                    
+                    return {
+                        "llm_response": "",
+                        "success": True,
+                        "refuse_answer": True,
+                        "refuse_reason": reason,
+                        "stats": {"total_time": time.time() - start_time},
+                        "usage": usage_totals,
+                    }
+            
+            # Execute web_tool calls in parallel
+            search_start = time.time()
+            tasks_to_run = []
+            task_indices = []
+            
+            for idx, func_name in enumerate(tool_call_names):
+                if func_name == "web_tool":
+                    tasks_to_run.append(self._execute_web_tool(tool_call_args_list[idx], context))
+                    task_indices.append(idx)
+            
+            # Run all web_tool calls in parallel
+            if tasks_to_run:
+                results = await asyncio.gather(*tasks_to_run, return_exceptions=True)
+            else:
+                results = []
+            
+            session.search_time += time.time() - search_start
+            
+            # Process results and collect notifications
+            notifications = []
+            result_map = {}  # Map task index to result
+            
+            for i, result in enumerate(results):
+                task_idx = task_indices[i]
+                if isinstance(result, Exception):
+                    result_map[task_idx] = {"summary": f"执行失败: {result}", "results": []}
+                else:
+                    result_map[task_idx] = result
+            
+            # Add all tool results to messages and collect notifications
+            for idx, func_name in enumerate(tool_call_names):
+                tc_id = tool_call_ids[idx]
+                args = tool_call_args_list[idx]
+                
+                if func_name == "web_tool":
+                    result = result_map.get(idx, {"summary": "未执行", "results": []})
+                    
+                    # Track tool call
+                    session.tool_calls.append({"name": func_name, "args": args})
+                    session.tool_results.append(result)
+                    
+                    # Collect notification
+                    notifications.append(f"🔍 {result['summary']}")
+                    
+                    # Add tool result to messages
+                    result_content = f"搜索完成: {result['summary']}\n\n找到 {len(result.get('results', []))} 个结果"
+                    session.messages.append({
+                        "role": "tool",
+                        "tool_call_id": tc_id,
+                        "content": result_content
+                    })
+                    
+                    # Add image source hint for web screenshots
+                    screenshot_count = result.get("screenshot_count", 0)
+                    if screenshot_count > 0:
+                        start_idx_img = session.total_image_count + 1
+                        end_idx_img = session.total_image_count + screenshot_count
+                        session.total_image_count = end_idx_img
+                        
+                        source_desc = result.get("source_desc", "网页截图")
+                        if start_idx_img == end_idx_img:
+                            hint = f"第{start_idx_img}张图片来自{source_desc}，作为查询的参考资料"
+                        else:
+                            hint = f"第{start_idx_img}-{end_idx_img}张图片来自{source_desc}，作为查询的参考资料"
+                        session.messages.append({"role": "system", "content": hint})
+                else:
+                    # Unknown tool
+                    session.messages.append({
+                        "role": "tool",
+                        "tool_call_id": tc_id,
+                        "content": f"Unknown tool: {func_name}"
+                    })
+            
+            # Send batched notification (up to 3 lines)
+            if self.send_func and notifications:
+                try:
+                    # Join notifications with newlines, max 3 lines
+                    notification_msg = "\n".join(notifications[:3])
+                    await self.send_func(notification_msg)
+                except Exception as e:
+                    logger.warning(f"AgentPipeline: Failed to send notification: {e}")
+            
+            # Increment round count after processing all tool calls in this round
+            if tasks_to_run:
+                session.round_count += 1
+        
+        # Build final response
+        total_time = time.time() - start_time
+        stats = {"total_time": total_time}
+        
+        # Update conversation history
+        conversation_history.append({"role": "user", "content": user_input})
+        conversation_history.append({"role": "assistant", "content": final_content})
+        
+        stages_used = self._build_stages_ui(session, context, usage_totals, total_time)
+        logger.info(f"AgentPipeline: Built stages_used = {stages_used}")
+        
+        return {
+            "llm_response": final_content,
+            "success": True,
+            "stats": stats,
+            "model_used": model,
+            "conversation_history": conversation_history,
+            "usage": usage_totals,
+            "web_results": context.web_results,
+            "tool_calls_count": session.call_count,
+            "stages_used": stages_used,
+        }
+    
+    async def _execute_web_tool(self, args: Dict, context: StageContext) -> Dict[str, Any]:
+        """执行 web_tool - 复用 /w 逻辑，支持过滤器语法"""
+        query = args.get("query", "")
+        
+        # 1. URL 截图模式 - 检测 query 中是否包含 URL
+        url_match = re.search(r'https?://\S+', query)
+        if url_match:
+            url = url_match.group(0)
+            # Send URL screenshot notification
+            if self.send_func:
+                try:
+                    short_url = url[:40] + "..." if len(url) > 40 else url
+                    await self.send_func(f"📸 正在截图: {short_url}")
+                except Exception:
+                    pass
+            
+            logger.info(f"AgentPipeline: Screenshot URL with content: {url}")
+            # Use screenshot_with_content to get both screenshot and text
+            result = await self.search_service.screenshot_with_content(url)
+            screenshot_b64 = result.get("screenshot_b64")
+            content = result.get("content", "")
+            title = result.get("title", "")
+            
+            if screenshot_b64:
+                context.web_results.append({
+                    "_id": context.next_id(),
+                    "_type": "page",
+                    "url": url,
+                    "title": title or "Screenshot",
+                    "screenshot_b64": screenshot_b64,
+                    "content": content,  # Text content for LLM
+                })
+                return {
+                    "summary": f"已截图: {url[:50]}{'...' if len(url) > 50 else ''}",
+                    "results": [{"_type": "screenshot", "url": url}],
+                    "screenshot_count": 1,
+                    "source_desc": f"URL截图 ({url[:30]}...)"
+                }
+            return {
+                "summary": f"截图失败: {url[:50]}",
+                "results": [],
+                "screenshot_count": 0
+            }
+        
+        # 2. 解析过滤器语法
+        filters, search_query, error = parse_filter_syntax(query, max_count=3)
+        
+        if error:
+            return {"summary": error, "results": []}
+        
+        # 3. 如果有过滤器，发送搜索+截图预告
+        if filters and self.send_func:
+            try:
+                # Build filter description
+                filter_desc_parts = []
+                for f_type, f_val, f_count in filters:
+                    if f_type == 'index':
+                        filter_desc_parts.append(f"第{f_val}个")
+                    else:
+                        filter_desc_parts.append(f"{f_val}={f_count}")
+                filter_desc = ", ".join(filter_desc_parts)
+                await self.send_func(f"🔍 正在搜索 \"{search_query}\" 并匹配 [{filter_desc}]...")
+            except Exception:
+                pass
+        
+        logger.info(f"AgentPipeline: Searching for: {search_query}")
+        results = await self.search_service.search(search_query)
+        visible = [r for r in results if not r.get("_hidden")]
+        
+        # Add search results to context
+        for r in results:
+            r["_id"] = context.next_id()
+            if "_type" not in r:
+                r["_type"] = "search"
+            r["query"] = search_query
+            context.web_results.append(r)
+        
+        # 4. 如果有过滤器，截图匹配的链接
+        if filters:
+            urls = self._collect_filter_urls(filters, visible)
+            if urls:
+                logger.info(f"AgentPipeline: Taking screenshots with content of {len(urls)} URLs")
+                # Use screenshot_with_content to get both screenshot and text
+                screenshot_tasks = [self.search_service.screenshot_with_content(u) for u in urls]
+                results = await asyncio.gather(*screenshot_tasks)
+                
+                # Add screenshots and content to context
+                successful_count = 0
+                for url, result in zip(urls, results):
+                    screenshot_b64 = result.get("screenshot_b64") if isinstance(result, dict) else None
+                    content = result.get("content", "") if isinstance(result, dict) else ""
+                    title = result.get("title", "") if isinstance(result, dict) else ""
+                    
+                    if screenshot_b64:
+                        successful_count += 1
+                        # Find and update the matching result
+                        for r in context.web_results:
+                            if r.get("url") == url:
+                                r["screenshot_b64"] = screenshot_b64
+                                r["content"] = content  # Text content for LLM
+                                r["title"] = title or r.get("title", "")
+                                r["_type"] = "page"
+                                break
+                
+                return {
+                    "summary": f"搜索 \"{search_query}\" 并截图 {successful_count} 个匹配结果",
+                    "results": [{"url": u, "_type": "page"} for u in urls],
+                    "screenshot_count": successful_count,
+                    "source_desc": f"搜索 \"{search_query}\" 的网页截图"
+                }
+        
+        # 5. 普通搜索模式 (无截图)
+        return {
+            "summary": f"搜索 \"{search_query}\" 找到 {len(visible)} 条结果",
+            "results": visible,
+            "screenshot_count": 0
+        }
+    
+    def _collect_filter_urls(self, filters: List, visible: List[Dict]) -> List[str]:
+        """Collect URLs based on filter specifications."""
+        urls = []
+        
+        for filter_type, filter_value, count in filters:
+            if filter_type == 'index':
+                idx = filter_value - 1  # Convert to 0-based
+                if 0 <= idx < len(visible):
+                    url = visible[idx].get("url", "")
+                    if url and url not in urls:
+                        urls.append(url)
+            else:
+                # Link filter
+                found_count = 0
+                for res in visible:
+                    url = res.get("url", "")
+                    title = res.get("title", "")
+                    # Match filter against both URL and title
+                    if (filter_value in url.lower() or filter_value in title.lower()) and url not in urls:
+                        urls.append(url)
+                        found_count += 1
+                        if found_count >= count:
+                            break
+        
+        return urls
+    
+    def _format_web_context(self, context: StageContext) -> str:
+        """Format web results for summary context."""
+        if not context.web_results:
+            return ""
+        
+        lines = ["## 已收集的信息\n"]
+        for r in context.web_results:
+            idx = r.get("_id", "?")
+            title = r.get("title", "Untitled")
+            url = r.get("url", "")
+            content = r.get("content", "")[:500] if r.get("content") else ""
+            has_screenshot = "有截图" if r.get("screenshot_b64") else ""
+            
+            lines.append(f"[{idx}] {title}")
+            if url:
+                lines.append(f"    URL: {url}")
+            if has_screenshot:
+                lines.append(f"    {has_screenshot}")
+            if content:
+                lines.append(f"    摘要: {content[:200]}...")
+            lines.append("")
+        
+        return "\n".join(lines)
+    
+    def _build_stages_ui(self, session: AgentSession, context: StageContext, usage_totals: Dict, total_time: float) -> List[Dict[str, Any]]:
+        """Build stages UI for rendering - compatible with App.vue flow section.
+        
+        Flow: Instruct (意图) → Search (搜索) → Summary (总结)
+        """
+        stages = []
+        
+        # Get model config for pricing
+        model_cfg = self.config.get_model_config("main")
+        model_name = model_cfg.model_name or self.config.model_name
+        input_price = getattr(model_cfg, "input_price", 0) or 0
+        output_price = getattr(model_cfg, "output_price", 0) or 0
+        
+        # 1. Instruct Stage (理解用户意图 - 第一次LLM调用)
+        if session.first_llm_time > 0:
+            # Estimate tokens for first call (rough split based on proportion)
+            # Since we track total usage, we approximate first call as ~40% of total
+            first_call_ratio = 0.4 if session.call_count > 0 else 1.0
+            instruct_input = int(usage_totals.get("input_tokens", 0) * first_call_ratio)
+            instruct_output = int(usage_totals.get("output_tokens", 0) * first_call_ratio)
+            instruct_cost = (instruct_input * input_price + instruct_output * output_price) / 1_000_000
+            
+            stages.append({
+                "name": "Instruct",
+                "model": model_name,
+                "provider": model_cfg.model_provider or "OpenRouter",
+                "description": "理解用户意图",
+                "time": session.first_llm_time,
+                "usage": {"input_tokens": instruct_input, "output_tokens": instruct_output},
+                "cost": instruct_cost,
+            })
+        
+        # 2. Search Stage (搜索)
+        if session.tool_calls:
+            # Collect all search descriptions
+            search_descriptions = []
+            for tc, result in zip(session.tool_calls, session.tool_results):
+                desc = result.get("summary", "")
+                if desc:
+                    search_descriptions.append(desc)
+            
+            stages.append({
+                "name": "Search",
+                "model": "",
+                "provider": "Web",
+                "description": " → ".join(search_descriptions) if search_descriptions else "Web Search",
+                "time": session.search_time,
+            })
+        
+        # 3. Summary Stage (总结)
+        # Calculate remaining tokens after instruct
+        summary_ratio = 0.6 if session.call_count > 0 else 0.0
+        summary_input = int(usage_totals.get("input_tokens", 0) * summary_ratio)
+        summary_output = int(usage_totals.get("output_tokens", 0) * summary_ratio)
+        summary_cost = (summary_input * input_price + summary_output * output_price) / 1_000_000
+        summary_time = session.llm_time - session.first_llm_time
+        
+        if summary_time > 0 or session.call_count > 0:
+            stages.append({
+                "name": "Summary",
+                "model": model_name,
+                "provider": model_cfg.model_provider or "OpenRouter",
+                "description": f"生成回答 ({session.call_count} 次工具调用)",
+                "time": max(0, summary_time),
+                "usage": {"input_tokens": summary_input, "output_tokens": summary_output},
+                "cost": summary_cost,
+            })
+        
+        return stages