entari-plugin-hyw 4.0.0rc10__py3-none-any.whl → 4.0.0rc12__py3-none-any.whl

hyw_core/core.py

@@ -10,10 +10,10 @@ from dataclasses import dataclass, field
 from typing import Dict, List, Any, Optional, Callable, Awaitable
 
 from loguru import logger
-from openai import AsyncOpenAI
 
 from .config import HywCoreConfig, ModelConfig
 from .pipeline import ModularPipeline
+from .agent import AgentPipeline
 from .search import SearchService
 from .stages.base import StageContext
 
@@ -95,22 +95,19 @@ class HywCore:
         self.config = config
         self._send_func = send_func
         
-        # Create OpenAI client
-        self._client = AsyncOpenAI(
-            api_key=config.api_key,
-            base_url=config.base_url if config.base_url else None
-        )
-        
         # Create search service
         self._search_service = SearchService(config)
         
-        # Create pipeline
+        # Create pipeline (for non-agent mode)
         self._pipeline = ModularPipeline(
             config=config,
             search_service=self._search_service,
             send_func=send_func
         )
         
+        # Agent pipeline (lazy init)
+        self._agent_pipeline = None
+        
         # Create renderer (lazy init)
         self._renderer = None
         
@@ -225,6 +222,140 @@ class HywCore:
             
         except Exception as e:
             logger.error(f"HywCore query failed: {e}")
+            logger.exception("Query error details:")
+            return QueryResponse(
+                success=False,
+                content="",
+                error=str(e),
+                total_time=time.time() - start_time
+            )
+    
+    async def query_agent(
+        self, 
+        request: QueryRequest,
+        output_path: Optional[str] = None
+    ) -> QueryResponse:
+        """
+        Agent-mode query with tool-calling capability.
+        
+        Uses AgentPipeline which can autonomously call web_tool up to 2 times.
+        Each tool call triggers an IM notification via send_notification callback.
+        
+        Args:
+            request: QueryRequest with user input, images, history
+            output_path: Optional path to save rendered image
+            
+        Returns:
+            QueryResponse with content, rendered image path, and metadata
+        """
+        start_time = time.time()
+        
+        try:
+            # Get or create agent pipeline with current send_func
+            send_func = request.send_notification or self._send_func
+            
+            if self._agent_pipeline is None or self._agent_pipeline.send_func != send_func:
+                self._agent_pipeline = AgentPipeline(
+                    config=self.config,
+                    search_service=self._search_service,
+                    send_func=send_func
+                )
+            
+            # Execute agent pipeline
+            result = await self._agent_pipeline.execute(
+                user_input=request.user_input,
+                conversation_history=request.conversation_history,
+                images=request.images if request.images else None,
+                model_name=request.model_name
+            )
+            
+            total_time = time.time() - start_time
+            
+            # Check for refusal
+            if result.get("refuse_answer"):
+                return QueryResponse(
+                    success=True,
+                    content="",
+                    should_refuse=True,
+                    refuse_reason=result.get("refuse_reason", ""),
+                    total_time=total_time
+                )
+            
+            # Check for error
+            if not result.get("success", True):
+                return QueryResponse(
+                    success=False,
+                    content="",
+                    error=result.get("error", "Unknown error"),
+                    total_time=total_time
+                )
+            
+            # Extract response data
+            content = result.get("llm_response", "")
+            usage = result.get("usage", {})
+            
+            # Convert web_results to references format for frontend
+            # Only include references that are actually cited in the markdown
+            import re
+            web_results = result.get("web_results", [])
+            
+            # Build visible results list (excluding hidden items)
+            visible_results = [r for r in web_results if not r.get("_hidden")]
+            
+            # Parse markdown to find which citations are used (pattern: [number])
+            citation_pattern = re.compile(r'\[(\d+)\]')
+            cited_ids = set()
+            for match in citation_pattern.finditer(content):
+                cited_ids.add(int(match.group(1)))
+            
+            # Only include cited references, in order of first appearance
+            references = []
+            for idx in sorted(cited_ids):
+                # idx is 1-based in markdown
+                if 1 <= idx <= len(visible_results):
+                    r = visible_results[idx - 1]
+                    references.append({
+                        "title": r.get("title", ""),
+                        "url": r.get("url", ""),
+                        "snippet": r.get("content", "")[:300] if r.get("content") else "",
+                        "images": r.get("images", []),
+                        "is_fetched": r.get("_type") == "page",
+                        "raw_screenshot_b64": r.get("screenshot_b64"),
+                    })
+            
+            # Build response
+            response = QueryResponse(
+                success=True,
+                content=content,
+                usage=usage,
+                total_time=total_time,
+                references=references,
+                web_results=web_results,
+                stages_used=result.get("stages_used", [])
+            )
+            
+            # Render image if output path provided
+            if output_path and content:
+                await self._ensure_renderer()
+                
+                render_success = await self._renderer.render(
+                    markdown_content=content,
+                    output_path=output_path,
+                    stats=result.get("stats", {}),
+                    references=references,
+                    page_references=[],
+                    stages_used=result.get("stages_used", []),
+                    theme_color=self.config.theme_color
+                )
+                
+                if render_success:
+                    response.image_path = output_path
+            
+            return response
+            
+        except Exception as e:
+            logger.error(f"HywCore query_agent failed: {e}")
+            logger.exception("Agent query error details:")
             return QueryResponse(
                 success=False,
                 content="",
@@ -261,6 +392,15 @@ class HywCore:
         """
         # Default to full_page=True as requested for /w command
         return await self._search_service.screenshot_url(url, full_page=True)
+
+    async def screenshot_with_content(self, url: str, max_content_length: int = 8000) -> Dict[str, Any]:
+        """
+        Capture screenshot and extract page content.
+        
+        Returns:
+            Dict with screenshot_b64, content (truncated), title, url
+        """
+        return await self._search_service.screenshot_with_content(url, max_content_length=max_content_length)
     
     async def screenshot_batch(self, urls: List[str]) -> List[Optional[str]]:
         """

hyw_core/crawling/completeness.py

@@ -46,13 +46,39 @@ IMAGE_CHECK_JS = """
         results.total++;
         
         const src = img.src || img.getAttribute('data-src') || '';
+        const dataSrc = img.getAttribute('data-src') || img.getAttribute('data-original') || 
+                        img.getAttribute('data-lazy-src') || img.getAttribute('data-lazy') || '';
+        const className = (typeof img.className === 'string' ? img.className : '').toLowerCase();
+        const loadingAttr = img.getAttribute('loading') || '';
+        
+        // Enhanced placeholder detection for blurred preview images (like mcmod.cn)
         const isPlaceholder = (
-            (img.getAttribute('data-src') && img.src !== img.getAttribute('data-src')) ||
+            // 1. data-src exists but not yet loaded into src
+            (dataSrc && img.src !== dataSrc) ||
+            // 2. Natural size much smaller than display size (blurred placeholder)
             (img.naturalWidth < 50 && img.clientWidth > 100) ||
+            (img.naturalWidth < 100 && img.clientWidth > 200 && img.naturalWidth * 4 < img.clientWidth) ||
+            // 3. Common placeholder keywords in src
             src.includes('placeholder') ||
             src.includes('loading') ||
+            src.includes('blank') ||
+            // 4. SVG placeholder or 1x1 tracking pixel
             src.startsWith('data:image/svg+xml') ||
-            (img.naturalWidth === 1 && img.naturalHeight === 1)
+            (img.naturalWidth === 1 && img.naturalHeight === 1) ||
+            // 5. Lazy-loading class indicators (common patterns)
+            className.includes('lazy') ||
+            className.includes('lazyload') ||
+            className.includes('lozad') ||
+            className.includes('b-lazy') ||
+            // 6. Blur indicators (common for LQIP - Low Quality Image Placeholder)
+            className.includes('blur') ||
+            src.includes('blur') ||
+            src.includes('thumb') ||
+            src.includes('thumbnail') ||
+            // 7. loading="lazy" + not complete (browser native lazy loading)
+            (loadingAttr === 'lazy' && !img.complete) ||
+            // 8. CSS blur filter applied (visual blurring)
+            (window.getComputedStyle(img).filter || '').includes('blur')
         );
         
         if (isPlaceholder) {
@@ -304,6 +330,8 @@ def trigger_lazy_load(tab: Any, config: Optional[CrawlConfig] = None) -> None:
     Scroll through page to trigger lazy-loaded images.
     
     Implements Crawl4AI's scan_full_page behavior.
+    Strategy: Fast scroll with minimal delay (0.2s) per step to trigger network requests,
+    then wait at the bottom for all images to settle.
     
     Args:
         tab: DrissionPage tab object
@@ -316,33 +344,94 @@ def trigger_lazy_load(tab: Any, config: Optional[CrawlConfig] = None) -> None:
     start = time.time()
     current_pos = 0
     
-    logger.debug("CompletenessChecker: Starting lazy load trigger scroll")
+    logger.info(f"CompletenessChecker: Starting lazy load scroll (fast scroll + final wait)")
     
     try:
-        while time.time() - start < config.scroll_timeout:
-            # Scroll down
+        max_scroll_steps = 100
+        step_count = 0
+        
+        # 1. Fast Scroll Phase
+        while step_count < max_scroll_steps:
+            step_count += 1
             current_pos += config.scroll_step
             tab.run_js(f"window.scrollTo(0, {current_pos});")
             
-            # Wait for lazy load triggers
-            time.sleep(config.scroll_delay)
+            # Simple fixed delay per step (0.2s) as requested
+            time.sleep(0.2)
             
             # Check if reached bottom
             height = tab.run_js("""
-                return Math.max(
+                Math.max(
                     document.body.scrollHeight || 0,
                     document.documentElement.scrollHeight || 0
-                );
+                )
             """, as_expr=True) or 0
             
             if current_pos >= height:
+                logger.debug(f"CompletenessChecker: Reached bottom at position {current_pos}")
                 break
+
+        # 2. Wait Phase at Bottom (Wait for images to settle - reduced timeout)
+        logger.debug("CompletenessChecker: Reached bottom, waiting for images to settle (max 2s)...")
+        wait_start = time.time()
+        max_wait_at_bottom = 2.0  # Reduced from 8s to 2s - scroll usually triggers loading quickly
+        
+        # Quick check: just verify images are not placeholders (simplified check)
+        check_all_images_js = """
+        (() => {
+            const imgs = Array.from(document.querySelectorAll('img'));
+            if (imgs.length === 0) return true;
+            
+            // Quick check: count non-placeholder images that are loaded
+            let loaded_count = 0;
+            let total_count = 0;
+            
+            for (const img of imgs) {
+                // Skip tiny images
+                if (img.clientWidth < 50 && img.clientHeight < 50) continue;
+                
+                total_count++;
+                const dataSrc = img.getAttribute('data-src') || img.getAttribute('data-original') || '';
+                const src = img.src || '';
+                
+                // Check if placeholder
+                const isPlaceholder = (
+                    (dataSrc && img.src !== dataSrc) ||
+                    (img.naturalWidth < 50 && img.clientWidth > 100) ||
+                    src.includes('placeholder') || src.includes('loading')
+                );
+                
+                // If not placeholder and loaded, count it
+                if (!isPlaceholder && img.complete && img.naturalWidth > 0) {
+                    loaded_count++;
+                }
+            }
+            
+            // If most images are loaded (80%+), consider it done
+            return total_count === 0 || (loaded_count / total_count) >= 0.8;
+        })()
+        """
+
+        # Quick check loop with shorter interval
+        check_count = 0
+        max_checks = 4  # 2s / 0.5s = 4 checks max
+        while check_count < max_checks:
+            try:
+                all_loaded = tab.run_js(check_all_images_js, as_expr=True)
+                if all_loaded:
+                    elapsed_wait = time.time() - wait_start
+                    logger.debug(f"CompletenessChecker: Images settled at bottom in {elapsed_wait:.1f}s")
+                    break
+            except:
+                pass
+            time.sleep(0.5)
+            check_count += 1
         
         # Scroll back to top
         tab.run_js("window.scrollTo(0, 0);")
         
         elapsed = time.time() - start
-        logger.debug(f"CompletenessChecker: Lazy load scroll complete in {elapsed:.1f}s")
+        logger.info(f"CompletenessChecker: Lazy load scroll complete - {step_count} steps in {elapsed:.1f}s")
         
     except Exception as e:
         logger.warning(f"CompletenessChecker: Lazy load scroll failed: {e}")

hyw_core/definitions.py

@@ -6,66 +6,28 @@ All global prompts and tool definitions for the pipeline stages.
 
 from typing import Dict, Any
 
-# =============================================================================
-# PROMPTS
-# =============================================================================
-
-
-
-
-
-SUMMARY_REPORT_SP = """# 你是一个信息整合专家 (Summary Agent).
-你需要根据用户问题、搜索结果和网页详细内容，生成最终的回答.
-如果用户发送你好或空内容回应你好即可.
-
-## 内容审查
-以下情况调用 `refuse_answer`:
-- 内容涉及隐喻政治事件、任务、现役国家领导人
-- r18+、r18g(但不包含正常galgame、科普等)
-
-## 过程要求
-- 用户要求的回复语言(包裹在 language 标签内)
-```language
-{language}
-```
-- 字数控制在600字以内, 百科式风格, 语言严谨不啰嗦.
-- 视觉信息: 输入中如果包含自动获取的网页截图，请分析图片中的信息作为参考.
-- 注意分辨搜索内容是否和用户问题有直接关系, 避免盲目相信混为一谈.
-- 正文格式: 
-  - 先给出一个 `# `大标题约 8-10 个字, 不要有多余废话, 不要直接回答用户的提问.
-  - 然后紧接着给出一个 <summary>...</summary>, 除了给出一个约 100 字的纯文本简介, 介绍本次输出的长文的清晰、重点概括.
-  - 随后开始详细二级标题 + markdown 正文, 语言描绘格式丰富多样, 简洁准确可信.
-  - 请不要给出过长的代码、表格列数等, 只讲重点和准确的数据.
-  - 不支持渲染: 链接, 图片链接, mermaid
-  - 支持渲染: 公式, 代码高亮, 只在需要的时候给出.
-  - 图片链接、链接框架会自动渲染出, 你无需显式给出.
-- 引用:
-  > 重要: 所有正文内容必须基于实际信息, 保证百分百真实度
-  - 信息来源已按获取顺序编号为 [1], [2], [3]...
-  - 优先引用优质 fetch 抓取的页面的资源, 但如果抓取到需要登录、需要验证码、需要跳转到其他网站等无法获取的资源, 则不引用此资源
-  - 正文中直接使用 [1] 格式引用, 只引用对回答有帮助的来源, 只使用官方性较强的 wiki、官方网站、资源站等等, 不使用第三方转载新闻网站.
-  - 无需给出参考文献列表, 系统会自动生成
+# Used by SummaryStage - language appended at runtime
+SUMMARY_REPORT_SP = """# 你是一个总结助手 (Agent), 你的职责是基于搜索工具给出的信息，回答用户的问题或解释用户问题中的关键词。
+## 核心原则
+最小限度使用自身知识, 尽可能使用 web_tool 获取信息.
+
+## 工具使用指南
+- 适当时候调用 `refuse_answer`
+
+## 回答格式
+- `# ` 大标题约 8-10 个字
+- <summary>...</summary> 约 100 字的概括
+- 二级标题 + markdown 正文
+- 正文使用 [1] 格式引用信息来源, 无需写出源, 系统自动渲染
 """
 
-
-# =============================================================================
-# VISION DESCRIPTION PROMPT
-# =============================================================================
-
-
-
-
-# =============================================================================
-# TOOL DEFINITIONS
-# =============================================================================
-
 def get_refuse_answer_tool() -> Dict[str, Any]:
     """Tool for refusing to answer inappropriate content."""
     return {
         "type": "function",
         "function": {
             "name": "refuse_answer",
-            "description": "违规内容拒绝回答",
+            "description": "违规内容拒绝回答，内容涉及隐喻政治事件、任务、现役国家领导人、r18+、r18g(但不包含正常galgame、科普等)",
             "parameters": {
                 "type": "object",
                 "properties": {
@@ -77,7 +39,63 @@ def get_refuse_answer_tool() -> Dict[str, Any]:
     }
 
 
+def get_web_tool() -> Dict[str, Any]:
+    """Tool for web search with filter syntax and URL screenshot."""
+    return {
+        "type": "function",
+        "function": {
+            "name": "web_tool",
+            "description": """搜索网页或截图指定URL。用于获取最新信息、查找资料。
+网页搜索(大部分问题优先使用此方法): 
+直接传入搜索词如 "python async" 会返回搜索结果列表
+
+网页截图(当用户明确要求截图时使用):
+传入完整URL如 "https://example.com" 会直接截图该页面
+
+网页搜索 + 网页截图(可以预测能直接搜到什么样的结果时使用): (最终截图最多3张)
+- 域名过滤: "github=2: python async" → 会搜索 "python async github" 并截图 链接/标题包含 "github" 的前2个结果
+- 序号选择: "1,2: minecraft mods" → 会搜索 "minecraft mods" 并截图第1、2个结果
+- 多域名: "mcmod=1, github=1: forge mod" → 会搜索 "forge mod mcmod github" 并截图 链接/标题包含 "mcmod" 的前1个结果和 链接/标题包含 "github" 的前1个结果
+""",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "query": {
+                        "type": "string",
+                        "description": "搜索查询或URL。支持过滤器语法(见描述)"
+                    }
+                },
+                "required": ["query"]
+            }
+        }
+    }
+
 
+# =============================================================================
+# AGENT PROMPTS
+# =============================================================================
 
+AGENT_SYSTEM_PROMPT = """# 你是一个智能助手 (Agent), 你的职责是使用 `web_tool` 工具来帮助用户搜索网页或截图URL, 同时完成用户分配给你的任务.
+## 任务
+理解用户意图分配给你的任务.
+如果用户没有明确分配任务, 则默认任务为解释用户问题中的关键词.
+
+## 核心原则
+最小限度使用自身知识, 尽可能使用 web_tool 获取信息.
+
+## 工具使用指南
+- 积极使用 web_tool 获取信息
+    - 搜索时, 关键词保证简单、指向准确、利于传统搜索引擎.
+    - 获取页面截图时, 只使用官方性较强的 wiki、官方网站、资源站等等, 不使用第三方转载新闻网站.
+- 最多可调用2次工具, 之后必须给出最终回答
+- 适当时候调用 `refuse_answer`
+- 对于具体任务, 如果是转述、格式化、翻译等, 请直接给出最终回答, 不再调用工具
+
+## 回答格式
+- `# ` 大标题约 8-10 个字
+- <summary>...</summary> 约 100 字的概括
+- 二级标题 + markdown 正文
+- 正文使用 [1] 格式引用信息来源, 无需写出源, 系统自动渲染
+"""
 
 

hyw_core/search.py

@@ -159,3 +159,13 @@ class SearchService:
         """
         service = get_screenshot_service(headless=self._headless)
         return await service.screenshot_url(url, full_page=full_page)
+
+    async def screenshot_with_content(self, url: str, max_content_length: int = 8000) -> Dict[str, Any]:
+        """
+        Capture screenshot and extract page content.
+        
+        Returns:
+            Dict with screenshot_b64, content (truncated), title, url
+        """
+        service = get_screenshot_service(headless=self._headless)
+        return await service.screenshot_with_content(url, max_content_length=max_content_length)

hyw_core/stages/summary.py

@@ -43,9 +43,7 @@ class SummaryStage(BaseStage):
         # Select prompt
         language = getattr(self.config, "language", "Simplified Chinese")
         
-        system_prompt = SUMMARY_REPORT_SP.format(
-            language=language
-        )
+        system_prompt = SUMMARY_REPORT_SP + f"\n\n用户要求的语言: {language}"
         
         # Build Context Message
         context_message = f"## Web Search & Page Content\n\n```context\n{full_context}\n```"