speccore 8.3.87 → 8.3.97

package/dist/core/prompt-builder.js

@@ -6,6 +6,7 @@ exports.buildPrompt = buildPrompt;
 exports.formatPrompt = formatPrompt;
 exports.parseAiResponse = parseAiResponse;
 exports.outputNeedsInfo = outputNeedsInfo;
+exports.processMarkdownContent = processMarkdownContent;
 /**
  * prompt-builder — 统一的 Spec → AI Prompt 构建引擎
  *
@@ -21,7 +22,10 @@ const logger_1 = require("../utils/logger");
 const knowledge_graph_1 = require("./knowledge-graph");
 const context_builder_1 = require("./context-builder");
 const spec_paths_1 = require("./spec-paths");
+const rag_engine_1 = require("./rag-engine");
 const unified_retrieval_1 = require("./unified-retrieval");
+// v8.3.94+: 视觉模型引擎（ specs 图片理解）
+const vision_engine_1 = require("./vision-engine");
 // v6.93.0+: Prompt 插件系统
 const prompt_plugins_1 = require("./prompt-plugins");
 const techStackCache = new Map();
@@ -118,7 +122,7 @@ async function loadApiSpecs(cwd, taskDir, reqContent) {
     if (tableMatch) {
         const startIdx = content.indexOf(tableMatch[0]);
         const afterTable = content.substring(startIdx);
-        const lines = afterTable.split('\n');
+        const lines = afterTable.split(/\r?\n/);
         for (let i = 2; i < lines.length; i++) { // 跳过表头和分隔线
             const line = lines[i].trim();
             if (!line.startsWith('|'))
@@ -151,7 +155,7 @@ async function loadDataModels(cwd, taskDir, reqContent) {
         const fieldTableRegex = /\|.*字段.*\|.*类型.*\|.*说明.*\|/gi;
         let match;
         let currentModel = null;
-        const lines = section.split('\n');
+        const lines = section.split(/\r?\n/);
         for (let i = 0; i < lines.length; i++) {
             // 检测模型名称（### 标题 或 **粗体**）
             const headingMatch = lines[i].match(/^#{2,4}\s*(.+)/);
@@ -192,7 +196,7 @@ async function loadBusinessRules(cwd, taskDir, reqContent) {
         const content = await cachedRead(constitutionCache, constitutionPath, async () => await (0, fs_extra_1.readFile)(constitutionPath, 'utf-8'));
         const namingSection = content.match(/##\s*命名规范[\s\S]*?(?=## |$)/i);
         if (namingSection) {
-            const lines = namingSection[0].split('\n');
+            const lines = namingSection[0].split(/\r?\n/);
             for (const line of lines) {
                 if (line.match(/^[-*]\s+(.+)/)) {
                     rules.push({ rule: RegExp.$1.trim() });
@@ -205,7 +209,7 @@ async function loadBusinessRules(cwd, taskDir, reqContent) {
         if (content) {
             const ruleSection = content.match(/(?:业务规则|约束条件|Constraint)[\s\S]*?(?=## |\n##|$)/i);
             if (ruleSection) {
-                const lines = ruleSection[0].split('\n');
+                const lines = ruleSection[0].split(/\r?\n/);
                 for (const line of lines) {
                     if (line.match(/^[-*]\s+(.+)/)) {
                         rules.push({ rule: RegExp.$1.trim() });
@@ -227,6 +231,8 @@ async function loadExtraSpecs(cwd, taskDir, platform, iteration, options) {
     const MAX_TOTAL = options?.maxTotalChars ?? 8000;
     let totalChars = 0;
     const seenPaths = new Set();
+    // v8.3.94+: 读取视觉模型配置（用于 specs 图片理解）
+    const visionConfig = await (0, vision_engine_1.loadVisionConfig)(cwd);
     const files = [
         { name: '开发指南', path: '00-specs/DEV_GUIDE.md' },
         { name: '任务上下文', path: '_shared/CONTEXT.md' },
@@ -258,7 +264,9 @@ async function loadExtraSpecs(cwd, taskDir, platform, iteration, options) {
                     if (await (0, fs_extra_1.pathExists)(featurePlatDir)) {
                         const entries = await (0, fs_extra_1.readdir)(featurePlatDir, { withFileTypes: true });
                         for (const entry of entries) {
-                            if (entry.name.endsWith('.md') && !(0, task_utils_1.isTimestampBackup)(entry.name)) {
+                            const isMd = entry.name.endsWith('.md') && !(0, task_utils_1.isTimestampBackup)(entry.name);
+                            const isHtml = (entry.name.endsWith('.html') || entry.name.endsWith('.htm')) && !(0, task_utils_1.isTimestampBackup)(entry.name);
+                            if (isMd || isHtml) {
                                 files.push({
                                     name: `${feature}-${platform}端规格`,
                                     path: (0, path_1.join)(featurePlatDir, entry.name),
@@ -314,14 +322,23 @@ async function loadExtraSpecs(cwd, taskDir, platform, iteration, options) {
         const fullPath = (0, path_1.join)(cwd, taskDir, f.path);
         seenPaths.add(fullPath);
         if (await (0, fs_extra_1.pathExists)(fullPath)) {
-            let content = await (0, fs_extra_1.readFile)(fullPath, 'utf-8');
+            let rawContent = await (0, fs_extra_1.readFile)(fullPath, 'utf-8');
+            // v8.3.92+: HTML 原型文件提取文本内容
+            const isHtml = fullPath.endsWith('.html') || fullPath.endsWith('.htm');
+            let content = isHtml ? (0, rag_engine_1.extractHtmlText)(rawContent) : rawContent;
+            // v8.3.93+: Markdown 链接自动展开 + 图片提取
+            if (fullPath.endsWith('.md')) {
+                content = await processMarkdownContent(content, fullPath, seenPaths, visionConfig, {
+                    maxLinkDepth: 2, maxLinkChars: 1200, maxSvgChars: 1500,
+                });
+            }
             // 跳过空文件或纯占位符文件
             if (content.trim().length <= 50 || content.trim().match(/^#+\s*待填充|^<!--\s*AI-FILL\s*-->$/m)) {
                 continue;
             }
             // 单文件大小限制
             if (content.length > MAX_PER_FILE) {
-                content = content.slice(0, MAX_PER_FILE) + `\n\n> ... (已截断，原文件 ${content.length} 字)`;
+                content = content.slice(0, MAX_PER_FILE) + `\n\n> ... (已截断，原文件 ${rawContent.length} 字，完整内容请 Read: ${fullPath})`;
             }
             // 总大小限制
             if (totalChars + content.length > MAX_TOTAL) {
@@ -345,7 +362,16 @@ async function loadExtraSpecs(cwd, taskDir, platform, iteration, options) {
             continue;
         seenPaths.add(fullPath);
         try {
-            let content = await (0, fs_extra_1.readFile)(fullPath, 'utf-8');
+            let rawContent = await (0, fs_extra_1.readFile)(fullPath, 'utf-8');
+            // v8.3.92+: HTML 原型文件提取文本内容
+            const isHtml = fullPath.endsWith('.html') || fullPath.endsWith('.htm');
+            let content = isHtml ? (0, rag_engine_1.extractHtmlText)(rawContent) : rawContent;
+            // v8.3.93+: Markdown 链接自动展开 + 图片提取
+            if (fullPath.endsWith('.md')) {
+                content = await processMarkdownContent(content, fullPath, seenPaths, visionConfig, {
+                    maxLinkDepth: 2, maxLinkChars: 1200, maxSvgChars: 1500,
+                });
+            }
             if (content.trim().length <= 50 || content.trim().match(/^#+\s*待填充|^<!--\s*AI-FILL\s*-->$/m)) {
                 continue;
             }
@@ -374,7 +400,7 @@ async function loadExtraSpecs(cwd, taskDir, platform, iteration, options) {
  */
 async function scanUserCustomFiles(cwd, taskDir, platform, seenPaths) {
     const results = [];
-    const validExts = ['.md', '.yaml', '.yml', '.json'];
+    const validExts = ['.md', '.yaml', '.yml', '.json', '.html', '.htm'];
     const skipDirs = new Set(['.meta', '.git', 'node_modules', 'tests', 'src', 'dist', 'build']);
     // 1. 扫描 00-specs/ 下所有文件（排除已在白名单中的）
     const specsDir = (0, path_1.join)(cwd, taskDir, '00-specs');
@@ -448,17 +474,28 @@ async function loadAllTaskContext(cwd, taskDir, platform, iteration, graph) {
     const MAX_TOTAL = 20000;
     let totalChars = 0;
     const seen = new Set();
+    // v8.3.94+: 读取视觉模型配置（用于 specs 图片理解）
+    const visionConfig = await (0, vision_engine_1.loadVisionConfig)(cwd);
     const addFile = async (fullPath, name, relPath) => {
         if (seen.has(fullPath))
             return;
         seen.add(fullPath);
         if (!(await (0, fs_extra_1.pathExists)(fullPath)))
             return;
-        let content = await (0, fs_extra_1.readFile)(fullPath, 'utf-8');
+        let rawContent = await (0, fs_extra_1.readFile)(fullPath, 'utf-8');
+        // v8.3.92+: HTML 原型文件提取文本内容
+        const isHtml = fullPath.endsWith('.html') || fullPath.endsWith('.htm');
+        let content = isHtml ? (0, rag_engine_1.extractHtmlText)(rawContent) : rawContent;
+        // v8.3.93+: Markdown 链接自动展开 + 图片提取
+        if (fullPath.endsWith('.md')) {
+            content = await processMarkdownContent(content, fullPath, seen, visionConfig, {
+                maxLinkDepth: 2, maxLinkChars: 1500, maxSvgChars: 2000,
+            });
+        }
         if (content.trim().length <= 50 || content.trim().match(/^#+\s*待填充|^<!--\s*AI-FILL\s*-->$/m))
             return;
         if (content.length > MAX_PER_FILE) {
-            content = content.slice(0, MAX_PER_FILE) + `\n\n> ... (已截断，原文件 ${content.length} 字)`;
+            content = content.slice(0, MAX_PER_FILE) + `\n\n> ... (已截断，原文件 ${rawContent.length} 字，完整内容请 Read: ${fullPath})`;
         }
         if (totalChars + content.length > MAX_TOTAL)
             return;
@@ -484,7 +521,7 @@ async function loadAllTaskContext(cwd, taskDir, platform, iteration, graph) {
                         continue;
                     await scanTaskDir(fullPath, `${prefix}${item.name}/`);
                 }
-                else if (/\.(md|yaml|yml)$/i.test(item.name)) {
+                else if (/\.(md|yaml|yml|html|htm)$/i.test(item.name)) {
                     // 排除自检阶段文件（TEST.md / SCHEMA.md / REVIEW.md 等）
                     if (CODEGEN_EXCLUDE_FILES.has(item.name.toLowerCase()))
                         continue;
@@ -593,7 +630,7 @@ const RULES_DESC = {
  */
 function extractHeadings(content) {
     const headings = [];
-    for (const line of content.split('\n')) {
+    for (const line of content.split(/\r?\n/)) {
         if (line.startsWith('## ')) {
             headings.push(line.replace(/^##\s+/, '').trim());
         }
@@ -604,7 +641,7 @@ function extractHeadings(content) {
  * 提取首段摘要：# 标题后的第一段非空内容，≤ 200 字
  */
 function extractSummary(content) {
-    const lines = content.split('\n');
+    const lines = content.split(/\r?\n/);
     let foundTitle = false;
     let summaryLines = [];
     let totalLen = 0;
@@ -698,7 +735,7 @@ function buildTOCEntry(path, description, content, maxSections) {
         sections: maxSections ? sections.slice(0, maxSections) : sections,
         summary: extractSummary(content) || undefined,
         platforms: extractPlatforms(path, content) || undefined,
-        lineCount: content.split('\n').length,
+        lineCount: content.split(/\r?\n/).length,
         tags: extractTags(sections) || undefined,
     };
 }
@@ -889,8 +926,8 @@ function formatGlobalContext(ctx, platform) {
                 for (const e of entries) {
                     // PATTERNS:architecture/x.md → architecture; platforms/admin/x.md → admin
                     const sub = group.prefix === 'PATTERNS:'
-                        ? e.path.replace('PATTERNS:', '').split('/')[0]
-                        : e.path.split('/')[1];
+                        ? e.path.replace('PATTERNS:', '').split(/[\\/]/)[0]
+                        : e.path.split(/[\\/]/)[1];
                     // v8.2.0+: PATTERNS 按端过滤 — 只保留通用分类 + 当前端相关模式
                     if (group.prefix === 'PATTERNS:' && platform) {
                         const isCommon = PATTERN_COMMON_CATEGORIES.has(sub);
@@ -1291,7 +1328,7 @@ async function buildPrompt(command, options) {
                             path: ccPath,
                             content: `## 前后端一致性校验报告\n\n${ccContent}`,
                         });
-                        logger_1.logger?.info?.(`   📋 已注入一致性校验报告: ${ccPath.replace(cwd + '/', '')}`);
+                        logger_1.logger?.info?.(`   📋 已注入一致性校验报告: ${(0, path_1.relative)(cwd, ccPath)}`);
                         break;
                     }
                 }
@@ -1721,6 +1758,164 @@ function outputNeedsInfo(req) {
     process.stdout.write(lines.join('\n'));
     process.exitCode = 11;
 }
+// ═══════════════════════════════════════════════════════════
+// Markdown 链接自动展开 + 图片提取（v8.3.93+）
+// ═══════════════════════════════════════════════════════════
+/** 提取 Markdown 文本链接 [text](path)，排除图片链接 */
+function extractMarkdownLinks(content) {
+    const links = [];
+    const regex = /\[([^\]]+)\]\(([^)]+)\)/g;
+    let match;
+    while ((match = regex.exec(content)) !== null) {
+        const text = match[1];
+        const path = match[2].trim();
+        // 跳过图片链接 ![alt](path) — 检查前一个字符
+        if (match.index > 0 && content.charAt(match.index - 1) === '!')
+            continue;
+        // 跳过外部链接、锚点、协议链接
+        if (/^(https?:|mailto:|ftp:|#|javascript:)/i.test(path))
+            continue;
+        links.push({ text, path });
+    }
+    return links;
+}
+/** 提取 Markdown 图片链接 ![alt](path) */
+function extractMarkdownImages(content) {
+    const images = [];
+    const regex = /!\[([^\]]*)\]\(([^)]+)\)/g;
+    let match;
+    while ((match = regex.exec(content)) !== null) {
+        images.push({ alt: match[1], path: match[2].trim() });
+    }
+    return images;
+}
+/** 判断是否为外部链接 */
+function isExternalLink(path) {
+    return /^(https?:|mailto:|ftp:|javascript:)/i.test(path);
+}
+/** 解析链接相对路径 */
+function resolveLinkPath(baseDir, linkPath) {
+    if (linkPath.startsWith('/'))
+        return linkPath;
+    return (0, path_1.join)(baseDir, linkPath);
+}
+/** 递归展开 Markdown 链接指向的文件内容（防循环，限深度） */
+async function expandMarkdownLinks(content, baseDir, seenPaths, maxDepth = 2, currentDepth = 0, maxChars = 1500) {
+    if (currentDepth >= maxDepth)
+        return content;
+    const links = extractMarkdownLinks(content);
+    if (links.length === 0)
+        return content;
+    const expansions = [];
+    for (const link of links) {
+        if (isExternalLink(link.path))
+            continue;
+        const resolved = resolveLinkPath(baseDir, link.path);
+        if (seenPaths.has(resolved))
+            continue;
+        seenPaths.add(resolved);
+        try {
+            if (!await (0, fs_extra_1.pathExists)(resolved))
+                continue;
+            const st = await (0, fs_extra_1.stat)(resolved);
+            if (!st.isFile())
+                continue;
+            let linkContent = await (0, fs_extra_1.readFile)(resolved, 'utf-8');
+            // HTML 文件提取文本
+            if (resolved.endsWith('.html') || resolved.endsWith('.htm')) {
+                linkContent = (0, rag_engine_1.extractHtmlText)(linkContent);
+            }
+            if (linkContent.trim().length <= 30)
+                continue;
+            const isTruncated = linkContent.length > maxChars;
+            if (isTruncated) {
+                linkContent = linkContent.slice(0, maxChars) + `\n\n> ... (已截断，完整内容请 Read: ${resolved})`;
+            }
+            expansions.push(`\n\n<!-- 展开链接: ${link.path} -->\n**[链接展开] ${link.text}** (${link.path}):\n\n${linkContent}`);
+        }
+        catch { /* 忽略读取失败的链接文件 */ }
+    }
+    const expanded = content + expansions.join('');
+    // 递归展开新内容中的链接（深度 + 1）
+    if (currentDepth + 1 < maxDepth) {
+        return expandMarkdownLinks(expanded, baseDir, seenPaths, maxDepth, currentDepth + 1, maxChars);
+    }
+    return expanded;
+}
+/** inline Markdown 图片信息（SVG 直接读取，启用视觉模型时描述位图，其他记录元信息） */
+async function inlineMarkdownImages(content, baseDir, seenPaths, visionConfig, maxSvgChars = 2000) {
+    const images = extractMarkdownImages(content);
+    if (images.length === 0)
+        return content;
+    const inlines = [];
+    let visionCallCount = 0;
+    const maxVisionCalls = visionConfig?.maxImagesPerPrompt ?? 10;
+    for (const img of images) {
+        if (isExternalLink(img.path)) {
+            inlines.push(`\n<!-- 图片: ${img.alt || '无描述'} | URL: ${img.path} -->`);
+            continue;
+        }
+        const resolved = resolveLinkPath(baseDir, img.path);
+        if (seenPaths.has(resolved))
+            continue;
+        seenPaths.add(resolved);
+        // SVG 直接 inline 其文本内容
+        if (resolved.endsWith('.svg')) {
+            try {
+                if (await (0, fs_extra_1.pathExists)(resolved)) {
+                    const svgContent = await (0, fs_extra_1.readFile)(resolved, 'utf-8');
+                    if (svgContent.length <= maxSvgChars) {
+                        inlines.push(`\n<!-- SVG 图片: ${img.path} -->\n${svgContent}`);
+                    }
+                    else {
+                        inlines.push(`\n<!-- SVG 图片: ${img.path} (内容过长已省略，大小: ${svgContent.length} 字符) -->`);
+                    }
+                }
+            }
+            catch { /* ignore */ }
+        }
+        else if ((0, vision_engine_1.isVisionEnabled)(visionConfig) && visionCallCount < maxVisionCalls) {
+            // v8.3.94+: 调用视觉模型描述图片
+            visionCallCount++;
+            try {
+                const result = await (0, vision_engine_1.describeImage)(resolved, visionConfig);
+                if (result.success) {
+                    inlines.push(`\n<!-- 图片描述 (视觉模型: ${visionConfig.provider}): ${img.path} -->\n**[图片内容]** ${img.alt || '无描述'}\n\n${result.description}`);
+                }
+                else {
+                    inlines.push(`\n<!-- 图片: ${img.alt || '无描述'} | 路径: ${img.path} | 描述失败: ${result.error} -->`);
+                }
+            }
+            catch (e) {
+                inlines.push(`\n<!-- 图片: ${img.alt || '无描述'} | 路径: ${img.path} | 描述失败: ${e.message || e} -->`);
+            }
+        }
+        else {
+            // 其他图片（PNG/JPG/GIF/WebP 等）：提取 alt + 路径 + 文件大小信息
+            inlines.push(`\n<!-- 图片: ${img.alt || '无描述'} | 路径: ${img.path} -->`);
+            try {
+                if (await (0, fs_extra_1.pathExists)(resolved)) {
+                    const st = await (0, fs_extra_1.stat)(resolved);
+                    inlines.push(`<!-- 图片大小: ${(st.size / 1024).toFixed(1)} KB | 完整路径: ${resolved} -->`);
+                }
+            }
+            catch { /* ignore */ }
+        }
+    }
+    return content + inlines.join('');
+}
+/**
+ * 统一处理 Markdown 内容：展开链接 + 提取图片
+ * 在 loadExtraSpecs / loadAllTaskContext 的文件加载后调用
+ * v8.3.97+: 导出供 analyze-engine / analyze 命令使用
+ */
+async function processMarkdownContent(content, filePath, seenPaths, visionConfig, options) {
+    const baseDir = (0, path_1.dirname)(filePath);
+    let processed = content;
+    processed = await expandMarkdownLinks(processed, baseDir, seenPaths, options?.maxLinkDepth ?? 2, 0, options?.maxLinkChars ?? 1500);
+    processed = await inlineMarkdownImages(processed, baseDir, seenPaths, visionConfig, options?.maxSvgChars ?? 2000);
+    return processed;
+}
 function generateExamples(command, req) {
     const iter = req.provided.iteration || 'Q1';
     const task = req.availableOptions?.tasks?.[0] || 'Task-001';