@tencentcloud/sdk-mcp 1.7.0 → 1.7.2

package/streamable-http.js

@@ -3,66 +3,31 @@ import { StreamableHTTPServerTransport } from '@modelcontextprotocol/sdk/server/
 import { createServer } from './server.js';
 const app = express();
 app.use(express.json());
-const server = createServer();
-const transport = new StreamableHTTPServerTransport({
-    sessionIdGenerator: undefined, // set to undefined for stateless servers
-});
-// Setup routes for the server
-const setupServer = async () => {
+// 无状态 Streamable HTTP：每个请求都新建独立 McpServer 实例和 transport，
+// 避免复用已 connect 的实例导致 "Already connected"。
+async function handleMcpRequest(req, res) {
+    const server = createServer();
+    const transport = new StreamableHTTPServerTransport({
+        sessionIdGenerator: undefined, // 无状态服务：不生成持久 session
+    });
     await server.connect(transport);
-};
+    await transport.handleRequest(req, res, req.body);
+}
 app.post('/mcp', async (req, res) => {
-    console.log('Received MCP request:', req.body);
-    try {
-        await transport.handleRequest(req, res, req.body);
-    }
-    catch (error) {
-        console.error('Error handling MCP request:', error);
-        if (!res.headersSent) {
-            res.status(500).json({
-                jsonrpc: '2.0',
-                error: {
-                    code: -32603,
-                    message: 'Internal server error',
-                },
-                id: null,
-            });
-        }
-    }
+    await handleMcpRequest(req, res);
 });
 app.get('/mcp', async (req, res) => {
-    console.log('Received GET MCP request');
-    res.writeHead(405).end(JSON.stringify({
-        jsonrpc: "2.0",
-        error: {
-            code: -32000,
-            message: "Method not allowed."
-        },
-        id: null
-    }));
+    await handleMcpRequest(req, res);
 });
 app.delete('/mcp', async (req, res) => {
-    console.log('Received DELETE MCP request');
-    res.writeHead(405).end(JSON.stringify({
-        jsonrpc: "2.0",
-        error: {
-            code: -32000,
-            message: "Method not allowed."
-        },
-        id: null
-    }));
+    await handleMcpRequest(req, res);
 });
 // Start the server
-const PORT = process.env.PORT || 3088;
-setupServer().then(() => {
-    app.listen(PORT, (error) => {
-        if (error) {
-            console.error('Failed to start server:', error);
-            process.exit(1);
-        }
-        console.log(`Weather MCP Server listening on port ${PORT}`);
-    });
-}).catch(error => {
-    console.error('Failed to set up the server:', error);
-    process.exit(1);
+const PORT = process.env.PORT || 8081;
+app.listen(PORT, (error) => {
+    if (error) {
+        console.error('Failed to start server:', error);
+        process.exit(1);
+    }
+    console.log(`RTC MCP Server (streamable HTTP) listening on port ${PORT}`);
 });

package/tools/common/finalize_answer.js

@@ -1,14 +1,6 @@
 import { z } from 'zod';
 import { FINALIZE_ANSWER_DEFINITION, FOLLOWUP_QUESTION } from '../definitions/finalize_answer_def.js';
 import { reportCLSClient } from '../../utils/report-cls-client.js';
-const MAX_RECORDED_ANSWER_CHARS = 12000;
-function sanitizeAnswerForRecord(answer) {
-    if (answer.length <= MAX_RECORDED_ANSWER_CHARS)
-        return answer;
-    const head = answer.slice(0, 9000);
-    const tail = answer.slice(-2500);
-    return `${head}\n\n[...truncated_for_record:${answer.length - (9000 + 2500)} chars omitted...]\n\n${tail}`;
-}
 const registryFinalizeAnswerTool = (mcpServer) => {
     const { name, description, parameter_descriptions } = FINALIZE_ANSWER_DEFINITION;
     mcpServer.registerTool(name, {
@@ -24,7 +16,7 @@ const registryFinalizeAnswerTool = (mcpServer) => {
         reportCLSClient({
             method: 'record_answer',
             queryID: params.queryID,
-            answer: sanitizeAnswerForRecord(params.answer),
+            answer: params.answer,
             prompt: params.prompt,
             product: params.product ? params.product.join(',') : '',
             framework: params.framework ?? '',

package/tools/common/search_trtc_knowledge.js

@@ -2,7 +2,9 @@ import { randomUUID } from 'node:crypto';
 import { z } from 'zod';
 import { SEARCH_KNOWLEDGE_DEFINITION } from '../definitions/search_trtc_knowledge_def.js';
 import { searchKnowledge } from '../../knowledge/search/search-knowledge.js';
+import { enrichL1Fragments } from '../../knowledge/search/l1-enrich.js';
 import { reportCLSClient } from '../../utils/report-cls-client.js';
+import { preflightRoute } from '../../utils/preflight-route.js';
 const registrySearchTRTCKnowledgeTool = (mcpServer) => {
     const { name, description, parameter_descriptions } = SEARCH_KNOWLEDGE_DEFINITION;
     mcpServer.registerTool(name, {
@@ -21,18 +23,11 @@ const registrySearchTRTCKnowledgeTool = (mcpServer) => {
             limit: z.number().min(1).max(16).default(8).describe(parameter_descriptions.limit),
             from: z.string().default('').describe(parameter_descriptions.from),
         },
-    }, (params) => {
+    }, async (params) => {
         const qid = params.queryID || randomUUID();
-        reportCLSClient({
-            method: 'search_trtc_knowledge',
-            queryID: qid,
-            prompt: params.prompt,
-            framework: params.frameworks?.join(',') ?? '',
-            product: params.product ? params.product.join(',') : '',
-            ide: params.ide,
-            from: params.from,
-        });
-        const response = searchKnowledge({
+        // preflight 只做归一化（归类、矛盾校正），不做 multi-path 编排。
+        // multi-platform 由 search core 的 filterPool（按 frameworks 交集）原生处理。
+        const routePlan = preflightRoute({
             prompt: params.prompt,
             product: params.product,
             frameworks: params.frameworks,
@@ -40,6 +35,25 @@ const registrySearchTRTCKnowledgeTool = (mcpServer) => {
             scenario: params.scenario,
             limit: params.limit,
         });
+        reportCLSClient({
+            method: 'search_trtc_knowledge',
+            queryID: qid,
+            prompt: params.prompt,
+            framework: routePlan.primary.frameworks?.join(',') ?? '',
+            product: routePlan.primary.product?.join(',') ?? '',
+            ide: params.ide,
+            from: params.from,
+            info: JSON.stringify({
+                preflight: routePlan.meta,
+                raw_product: params.product ?? [],
+                raw_frameworks: params.frameworks ?? [],
+                normalized_product: routePlan.primary.product,
+                normalized_frameworks: routePlan.primary.frameworks,
+            }),
+        });
+        // 单次检索：multi-platform same-product 由 frameworks 数组承接；TRC web+native 由 product 数组承接。
+        const result = searchKnowledge(routePlan.primary);
+        const response = await enrichL1Fragments(result);
         return {
             content: [{ type: 'text', text: JSON.stringify({ ...response, queryID: qid }) }],
         };

package/tools/definitions/finalize_answer_def.js

@@ -9,7 +9,7 @@ When to call:
 
 Input:
 - prompt: Original user question. Must match the search_trtc_knowledge prompt.
-- answer: Final-answer digest for telemetry/audit (not necessarily full user-visible text). Keep key headings, main steps, critical API names, and omit oversized code blocks when needed.
+- answer: Full synthesized answer (same text sent to the user).
 - queryID: REQUIRED. Must match the non-empty queryID returned by search_trtc_knowledge.
 - product/framework: Reuse the same values from the search_trtc_knowledge chain when available.
 
@@ -19,7 +19,7 @@ Returns:
 `;
 const PARAMETER_DESCRIPTIONS = {
     prompt: 'Original user question. Must match search_trtc_knowledge.',
-    answer: 'Final-answer digest for telemetry/audit. Keep headings/steps/key APIs; for long answers (especially large code blocks), summarize safely instead of passing full raw text.',
+    answer: 'Full synthesized answer (same text sent to the user).',
     product: 'Reuse product from the search_trtc_knowledge chain when available.',
     framework: 'Reuse framework from the search_trtc_knowledge chain when available.',
     queryID: 'REQUIRED. Non-empty queryID returned by search_trtc_knowledge.',

package/tools/definitions/search_trtc_knowledge_def.js

@@ -1,6 +1,8 @@
 const TOOL_DESCRIPTION = `
 凡 Tencent RTC 文档咨询（Chat/Call/Live/Room/Push/Web TRTC/Native TRTC）：FAQ、错误码、API、组件、feature、插件、集成查阅、排障 → MUST 先本工具。禁止 web_search。写接入代码 → get_*_integration。
 
+**prompt 必须原样保留用户问题**：禁止扩写、发散、改写、补充 API 用法/场景描述。用户问什么就传什么原文。若需补充语义（如 API/事件/具体实现），用 types/frameworks/product 结构化参数表达，或另行搜索。
+
 **product**（能推断 MUST 传；不能确定先问用户）：
 push:TIMPush/推送/厂商/FCM/华为/小米/OPPO（≠chat）| chat:会话/TUIKit/IM/主题 | call:CallKit | live:LiveKit | room:RoomKit/会议
 rtcengine:Web TRTC | native_trtc_sdk:TRTCCloud | 跨域:["push","chat"]
@@ -19,10 +21,10 @@ types含api/component/feature/plugin/sdk/integration→MUST | API/组件/enterRo
 
 **L0/L1**：本地fragments仅基于片段答；chat/urls/须fetch官网URL，禁止臆造。
 
-**返回**：status、confidence、fragments(doc_urls)、meta.references/exact_hits、top_score/second_score、queryID、message、meta.answer_contract。答案须严格遵守 answer_contract（结论先行、用户可执行、未确认显式化）；正文不展开证据链，且禁止输出“命中片段/检索结果/置信度”等系统话术，但末尾需保留“参考文档”并列出 references/doc_urls。最终回复前 MUST 调用 finalize_answer；明确 follow-up 反馈后调用 submit_feedback。
+**返回**：status、confidence、fragments(doc_urls)、meta.references、queryID、message、meta.answer_contract。答案遵守 answer_contract(完整输出、用户可执行、未确认显式化),禁止精简,禁止输出"命中片段/检索结果/置信度/分数"等系统话术,但末尾需保留"参考文档"并列出 references/doc_urls。最终回复前 MUST 调用 finalize_answer;明确 follow-up 反馈后调用 submit_feedback。
 `;
 const PARAMETER_DESCRIPTIONS = {
-    prompt: '用户原始问题全文，保留完整上下文，不要压缩成关键词摘要。',
+    prompt: '用户原始问题原文，MUST 原样透传，禁止扩写/发散/补充（如补 API 用法、场景、追问）。用户没说的不写进去；需补充语义用 types/frameworks 等结构化参数表达。',
     product: '产品数组：chat|call|live|room|push|rtcengine|native_trtc_sdk。能推断 MUST 传；push优先(TIMPush/推送→push，禁止误归chat)。live 常见关键词：LiveKit/直播/语聊房/TUILiveKit/AtomicXCore/开播/观看。不能确定先问用户。',
     frameworks: '平台数组：react|vue|web|miniprogram|android|ios|flutter|c++|uni-app|harmonyos|react-native|unity|unreal-engine|donut。支持传多个平台用于跨端检索；CallKit 中 TUICallKit 建议按 react/vue 区分检索，TUICallEngine/TUICallEvent 仍归 web；小程序用 miniprogram；harmonyos/react-native/unity/unreal/donut仅push适用。若未传且问题依赖具体客户端平台，必须先调用 present_framework_choice 询问后重搜。',
     types: '默认不传；开放["faq"]。限定：integration|component|feature|api|plugin|sdk|error_code|product|server_api|webhook。component 与 feature 等价（同时匹配）；发现型/罗列型问题（"有哪些组件"）MUST 省略 types。控制台/计费/产品能力配置→product（非feature）。勿主动窄传。',

package/utils/fetch-doc-content.js

@@ -0,0 +1,198 @@
+/**
+ * fetch-doc-content.ts
+ *
+ * 在命中 chat/urls L1 URL 映射行时，由 server 内部 fetch 官网目标文档并清洗成
+ * markdown 返回给 AI，避免 AI 自行联网抓取（不可靠、拿到的还是整页 HTML）。
+ *
+ * 设计要点：
+ * - 白名单域名：仅允许 fetch 我们自己的 urls 映射里出现的官方域名，防 SSRF。
+ * - Node 22 原生 fetch 会自动解 gzip（解决 cloud.tencent.com 返回 gzip 导致乱码的问题）。
+ * - 内部分流函数 fetch 按 hostname 选策略：
+ *     - trtc.io（en 端 urls 主域名）：`url + '.md'` 直拿 text/markdown，跳过 cheerio 清洗
+ *     - 其他：cheerio + turndown 清洗 HTML（cloud.tencent.com / web.sdk.qcloud.com）
+ * - 内存缓存 + 超时 + 失败兜底（返回 null，不阻断原有检索）。
+ * - 对外函数 fetchDocContent 签名不变（l1-enrich 调用方无感知）。
+ * - 缓存 key 始终用原 URL（不带 .md 后缀），避免缓存 key 污染。
+ */
+import * as cheerio from 'cheerio';
+import TurndownService from 'turndown';
+/**
+ * 白名单域名：只允许这些域名的文档页被 fetch。
+ * 与 zh/resource/knowledge/chat/urls/ 和 en/resource/knowledge/chat/urls/ 下的映射保持一致。
+ */
+const ALLOWED_HOSTS = new Set([
+    'cloud.tencent.com',
+    'web.sdk.qcloud.com',
+    'trtc.io',
+]);
+/**
+ * 走"url + '.md' 直拿 markdown"策略的域名。
+ * 这类域名提供原生 markdown 视图（content-type=text/markdown），
+ * 比 cheerio+turndown 清洗 HTML 更干净、token 更省。
+ */
+const DIRECT_MD_HOSTS = new Set([
+    'trtc.io',
+]);
+const REQUEST_TIMEOUT_MS = 15000;
+/** 默认截断长度（字符），超过则按段落截断 */
+const DEFAULT_MAX_CHARS = 50000;
+/** 内存缓存：url -> markdown，避免同一咨询/相近请求重复打官网 */
+const cache = new Map();
+/** 判断 URL 是否命中白名单域名 */
+function isAllowedUrl(url) {
+    try {
+        const { hostname } = new URL(url);
+        return ALLOWED_HOSTS.has(hostname);
+    }
+    catch {
+        return false;
+    }
+}
+/**
+ * 从 HTML 中提取正文容器（cheerio 按域名选择稳定容器）。
+ * 找不到时回退到更宽泛的选择器；仍找不到返回 null，交由调用方退化为整页清洗。
+ */
+function extractArticleSelector(hostname) {
+    if (hostname === 'cloud.tencent.com') {
+        // 腾讯云文档：正文在 #docArticleContent 内（SSR 直出），依次回退
+        return ['#docArticleContent', '#doc-slate-root', '.J-markdown-box'];
+    }
+    if (hostname === 'web.sdk.qcloud.com') {
+        // JSDoc 静态文档：正文在 #main 的 article 内
+        return ['#main article', '#main', 'article'];
+    }
+    return [];
+}
+/** 使用 turndown 将选中的 HTML 节点转成 markdown */
+function convertToMarkdown(html) {
+    const turndown = new TurndownService({
+        headingStyle: 'atx', // ## 风格标题
+        codeBlockStyle: 'fenced', // ``` 围栏代码块
+        bulletListMarker: '-',
+    });
+    turndown.remove(['script', 'style', 'nav', 'noscript', 'iframe']);
+    let md = turndown.turndown(html);
+    md = md
+        .replace(/[ \t]+\n/g, '\n')
+        .replace(/\n{3,}/g, '\n\n')
+        .trim();
+    return md;
+}
+/** 按段落截断到 maxChars */
+function truncateMarkdown(md, maxChars) {
+    if (md.length <= maxChars)
+        return md;
+    const head = md.slice(0, maxChars);
+    const lastBreak = head.lastIndexOf('\n\n');
+    return `${head.slice(0, lastBreak > 0 ? lastBreak : maxChars)}\n…[已截断]`;
+}
+/**
+ * 公共 fetch helper：带 timeout + 错误吞咽，把 Response 交给 handle 解析。
+ * 用 globalThis.fetch 避免与下方内部分流函数 fetch 同名冲突。
+ */
+async function fetchWithTimeout(url, handle) {
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(), REQUEST_TIMEOUT_MS);
+    try {
+        const res = await globalThis.fetch(url, {
+            signal: controller.signal,
+            headers: {
+                'User-Agent': 'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0 Safari/537.36',
+                // 加 text/markdown 偏好，让 trtc.io 等支持 .md 后缀的站点优先返回 markdown
+                'Accept': 'text/markdown,text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8',
+                'Accept-Language': 'zh-CN,zh;q=0.9,en;q=0.8',
+            },
+            // Node fetch 自动处理 gzip/deflate/br
+            redirect: 'follow',
+        });
+        if (!res.ok)
+            return null;
+        return await handle(res);
+    }
+    catch {
+        return null;
+    }
+    finally {
+        clearTimeout(timer);
+    }
+}
+/** cheerio+turndown 路径：处理 cloud.tencent.com / web.sdk.qcloud.com 等 HTML 文档 */
+async function fetchAndConvert(url, maxChars) {
+    return fetchWithTimeout(url, async (res) => {
+        const html = await res.text();
+        const { hostname } = new URL(url);
+        // cheerio 定位正文容器
+        const $ = cheerio.load(html);
+        const selectors = extractArticleSelector(hostname);
+        let article = selectors.map((sel) => $(sel)).find((node) => node.length > 0);
+        if (!article || article.length === 0) {
+            // 找不到容器时退化为 body 清洗（会偏大，但不丢正文）
+            article = $('body');
+        }
+        const markdown = convertToMarkdown(article.html() ?? '');
+        if (!markdown)
+            return null;
+        const truncated = markdown.length > maxChars;
+        const final = truncated ? truncateMarkdown(markdown, maxChars) : markdown;
+        return {
+            markdown: final,
+            truncated,
+            htmlBytes: Buffer.byteLength(html, 'utf-8'),
+        };
+    });
+}
+/**
+ * direct-md 路径：trtc.io 提供 `url + '.md'` 视图，
+ * content-type=text/markdown，可直接消费，跳过 cheerio 清洗。
+ * 失败时（content-type 不对 / 404）返回 null，让调用方走 fallback 路径。
+ */
+async function fetchDirectMarkdown(url, maxChars) {
+    const mdUrl = `${url}.md`;
+    return fetchWithTimeout(mdUrl, async (res) => {
+        const ct = res.headers.get('content-type') ?? '';
+        if (!/text\/markdown/i.test(ct))
+            return null;
+        const md = await res.text();
+        const truncated = md.length > maxChars;
+        return {
+            markdown: truncated ? truncateMarkdown(md, maxChars) : md,
+            truncated,
+            htmlBytes: 0,
+        };
+    });
+}
+/**
+ * 内部分流 fetch：按 hostname 选择策略。
+ * - DIRECT_MD_HOSTS 内的域名走 fetchDirectMarkdown；失败 fallback 到 fetchAndConvert。
+ * - 其他域名走 fetchAndConvert（cheerio+turndown）。
+ */
+async function fetch(url, hostname, maxChars) {
+    if (DIRECT_MD_HOSTS.has(hostname)) {
+        const direct = await fetchDirectMarkdown(url, maxChars);
+        // 防御：.md 视图失败时（如 content-type 不对、404）fallback 到 cheerio 路径
+        return direct ?? fetchAndConvert(url, maxChars);
+    }
+    return fetchAndConvert(url, maxChars);
+}
+/**
+ * fetch 官网文档并清洗为 markdown。
+ * 命中白名单才请求；失败/超时/非白名单一律返回 null，不抛错。
+ */
+export async function fetchDocContent(url, opts) {
+    if (!isAllowedUrl(url))
+        return null;
+    // 缓存 key 用原 URL（不带 .md），让 fetchDocContent 调用方透明
+    const cached = cache.get(url);
+    if (cached)
+        return { markdown: cached, truncated: false, htmlBytes: 0 };
+    const { hostname } = new URL(url);
+    const maxChars = opts?.maxChars ?? DEFAULT_MAX_CHARS;
+    const result = await fetch(url, hostname, maxChars);
+    if (result)
+        cache.set(url, result.markdown);
+    return result;
+}
+/** 清空内存缓存（测试/热更时用） */
+export function clearFetchDocCache() {
+    cache.clear();
+}

package/utils/preflight-route.js

@@ -0,0 +1,197 @@
+// ========== 信号分层（方案 §3，中英双语） ==========
+// 产品词：产品线路标识
+const PRODUCT_SIGNALS = {
+    call: /CallKit|TUICallKit|TUICallEngine|语音通话|视频通话|来电|去电|拨打|\b(?:voice|video)\s+calls?\b|incoming\s+calls?|outgoing\s+calls?|make\s+a\s+call|\bdial(?:ing)?\b/i,
+    live: /LiveKit|TUILiveKit|直播|语聊房|连麦|开播|观看直播|\blive\s+stream(?:ing)?\b|\bvoice\s+room\b|co-?host(?:ing)?|go\s+live|start\s+stream(?:ing)?|watch(?:ing)?\s+live|\bbroadcast(?:ing)?\b/i,
+    room: /RoomKit|会议|会中|会议室|屏幕分享|白板|\bconference\b|\bmeetings?\b|in-?meeting|meeting\s+room|screen\s+shar(?:e|ing)|white\s*board/i,
+    chat: /\bchat\b|\bim\b|会话|消息|TUIKit|ChatKit|\bconversation\b|\bmessaging\b|\binstant\s+messaging\b/i,
+    push: /TIMPush|推送|离线推送|FCM|APNs|华为|小米|OPPO|vivo|\bpush(?:\s+notification)?s?\b|offline\s+push|vendor\s+channel|\bHuawei\b|\bXiaomi\b|\bHonor\b/i,
+};
+// 平台词（纯运行形态，不含任何产品词）
+const PLATFORM_NATIVE_SIGNAL = /\b(android|ios|flutter|c\+\+|native)\b|安卓\s*(sdk|端|原生)|iOS\s*(SDK|端|原生)|原生\s*(sdk|端)?|native\s+(?:sdk|client|app)|mobile\s+(?:sdk|native)/i;
+const PLATFORM_WEB_SIGNAL = /\b(web|h5|html5|browser|webview)\b|浏览器|网页|Web端|H5端|网页端|web\s+(?:app|client|page)|in-?browser|mobile\s+browser|安卓\s*(浏览器|网页|H5|Web|webview)|iOS\s*(浏览器|网页|H5|Web|webview)|Android\s*(?:浏览器|网页|H5|Web|webview|browser|web)|iPad\s*(?:浏览器|网页|H5|Web|browser|web)|\b(Safari|Chrome|Edge|UC|QQ浏览器|QQ\s+Browser)\b/i;
+// 平台+产品复合词（运行形态 + 产品线路标识）
+const WEB_PRODUCT_PAIR_SIGNAL = /\bweb\s+(trtc|sdk|im|chat|roomkit)\b|\b(trtc|sdk|im|chat|roomkit)\s+web\b|trtc[-\s]?sdk|trtc\.js/i;
+// API 词（客户端调用形态）
+const NATIVE_TRTC_CLASS_SIGNAL = /\b(TRTCCloud|TRTCCloudListener|TRTCCloudDelegate|ITRTCCloud|TRTCTypeDef|TRTCCloudCallback|TXAudioEffectManager|TXBeautyManager|TXDeviceManager|TXLiteAVCode|TXLiveSDKTypeDef)\b/i;
+const NATIVE_TRTC_METHOD_SIGNAL = /\b(?:TRTCCloud|ITRTCCloud|TX[A-Za-z0-9_]+)\s*(?:\.|::|->)\s*(?:set|get|start|stop|enable|disable)[A-Z][A-Za-z0-9_]+\b/i;
+const WEB_TRTC_API_RE = /TRTC\.(setCurrentSpeaker|getMicrophoneList|getSpeakerList|getCameraList|isSupported|create)|trtc\.(startLocalAudio|startLocalVideo|updateLocalAudio|updateLocalVideo)/i;
+const WEB_TRTC_TOKEN_RE = /^TRTC\.(setCurrentSpeaker|getMicrophoneList|getSpeakerList|getCameraList|isSupported|create)$/i;
+// 协议级 TRTC 信号
+const TRTC_SIGNAL = /\bTRTC\b|\bWebRTC\b|rtcengine|TRTCCloud|腾讯实时音视频/i;
+// ========== 工具函数 ==========
+function uniq(list) {
+    return [...new Set(list)];
+}
+function extractQueryApiTokens(prompt) {
+    const matches = prompt.match(/\b(?:TRTC|trtc)\.[A-Za-z_][A-Za-z0-9_]*\b/g) ?? [];
+    return uniq(matches.map((m) => m.replace(/^trtc\./i, 'TRTC.')));
+}
+function collectPromptSignals(prompt) {
+    const products = new Set();
+    if (PRODUCT_SIGNALS.call.test(prompt))
+        products.add('call');
+    if (PRODUCT_SIGNALS.live.test(prompt))
+        products.add('live');
+    if (PRODUCT_SIGNALS.room.test(prompt))
+        products.add('room');
+    if (PRODUCT_SIGNALS.chat.test(prompt))
+        products.add('chat');
+    if (PRODUCT_SIGNALS.push.test(prompt))
+        products.add('push');
+    if (TRTC_SIGNAL.test(prompt))
+        products.add('trtc');
+    // 平台 tokens：web signal 锁定 → 压制 native 单字(安卓/iOS 仅归 hostEnv);
+    // 真复合词(unreal/unity/harmonyos/react-native/flutter/c++)不受影响。
+    const platforms = [];
+    const webLocked = PLATFORM_WEB_SIGNAL.test(prompt) || WEB_PRODUCT_PAIR_SIGNAL.test(prompt);
+    if (webLocked) {
+        platforms.push('web');
+    }
+    const nativeSpecificHits = [];
+    if (/\b(?:unreal|UE[45]?|unreal-engine)\b/i.test(prompt))
+        nativeSpecificHits.push('unreal-engine');
+    if (/\b(?:unity)\b/i.test(prompt))
+        nativeSpecificHits.push('unity');
+    if (/\b(?:harmonyos|鸿蒙|HarmonyOS)\b/i.test(prompt))
+        nativeSpecificHits.push('harmonyos');
+    if (/\b(?:react-native)\b/i.test(prompt))
+        nativeSpecificHits.push('react-native');
+    if (/\b(?:flutter)\b/i.test(prompt))
+        nativeSpecificHits.push('flutter');
+    if (/\b(?:c\+\+)\b/i.test(prompt))
+        nativeSpecificHits.push('c++');
+    // 安卓/iOS 单字只在非 web-locked 时进 platforms(web-locked 时仅归 hostEnv)
+    if (!webLocked && /\b(?:ios|iphone|ipad|苹果)\b/i.test(prompt))
+        nativeSpecificHits.push('ios');
+    if (!webLocked && /\b(?:android|安卓)\b/i.test(prompt))
+        nativeSpecificHits.push('android');
+    // 真复合词(如 unity/unreal/flutter/harmonyos/react-native/c++)无条件 push;单字在 webLocked 时被压制
+    if (nativeSpecificHits.length > 0) {
+        platforms.push(...nativeSpecificHits);
+    }
+    else if (!webLocked && PLATFORM_NATIVE_SIGNAL.test(prompt)) {
+        // 仅 "native" 词,未指定细分 → 默认 android(webLocked 时不触发 fallback,避免被 \b(android|ios|...) 反向命中把 web 路径覆盖成 native)
+        platforms.push('android');
+    }
+    // Host env：仅用于标记，不进 platforms 归一化
+    let hostEnv = null;
+    if (/\bios\b|iphone|ipad|苹果/i.test(prompt))
+        hostEnv = 'ios';
+    else if (/\bandroid\b|安卓/i.test(prompt))
+        hostEnv = 'android';
+    const queryApiTokens = extractQueryApiTokens(prompt);
+    const hasWebTrtcApi = PLATFORM_WEB_SIGNAL.test(prompt)
+        || WEB_TRTC_API_RE.test(prompt)
+        || queryApiTokens.some((t) => WEB_TRTC_TOKEN_RE.test(t));
+    const hasNativeTrtcApi = NATIVE_TRTC_CLASS_SIGNAL.test(prompt) || NATIVE_TRTC_METHOD_SIGNAL.test(prompt);
+    return { products, platforms, hostEnv, hasWebTrtcApi, hasNativeTrtcApi, queryApiTokens };
+}
+function reconcileWithPrompt(promptValues, explicit) {
+    // 核心规则：prompt 是真值源，explicit 只在 prompt 为空时启用。
+    // 这条规则必须严格遵守——
+    //   「prompt 是用户问题表达出来的真实意图」，explicit 只在外部上下文（IDE/文件/历史）有补充价值，
+    //   但 prompt 既然明示了「Web」，就要按 Web 来，不要因为 explicit 多了 'android' 就把它"补"进来当成双路。
+    // 情形 1：prompt 真值存在 → 信任 prompt 真值
+    //   - 即便 explicit 扩张（prompt 外多列了平台）→ 也只输出 prompt 真值（prompt_override=true 表示 agent 推理被忽略）
+    //   - 即便 explicit 收缩（prompt 真值的子集）→ 也只输出 prompt 真值（避免丢失用户意图）
+    if (promptValues.length > 0) {
+        const promptSet = new Set(promptValues);
+        const sameAsExplicit = !!explicit?.length
+            && explicit.length === promptValues.length
+            && explicit.every((v) => promptSet.has(v));
+        return { values: uniq(promptValues), corrected: !sameAsExplicit };
+    }
+    // 情形 2：prompt 真值无平台信息 → 信任上游 agent 基于上下文（IDE/文件/历史）的推理
+    if (!promptValues.length) {
+        return { values: uniq(explicit ?? []), corrected: false };
+    }
+    return { values: [], corrected: false };
+}
+// ========== Phase 3: 单 SearchParams 构造 ==========
+//
+// 关键约束：**永远只返回单 primary**，consumer 只跑一次 searchKnowledge。
+// multi-platform / multi-product 通过把 frameworks / product 设成数组交给 search core 处理。
+// search core 的 filterPool.applyRankMultipliers 已原生支持多值归并。
+function passthroughPlan(params, reason, env) {
+    return { primary: params, meta: { route_reason: reason, environment_downgraded: env } };
+}
+/** 单产品路径：product 是单值；frameworks 是数组（可能多平台共享一个 product） */
+function singleProductPlan(params, product, frameworks, hostEnv, reason) {
+    const envDiff = hostEnv != null && !frameworks.includes(hostEnv);
+    return {
+        primary: { ...params, product: [product], frameworks },
+        meta: { route_reason: reason, environment_downgraded: envDiff },
+    };
+}
+/** TRC 路径：product 是数组（rtcengine/native_trtc_sdk 两个产品码，按 platforms 各命中其一）；frameworks 也是数组 */
+function trcMultiProductPlan(params, products, frameworks, reason, env) {
+    return {
+        primary: { ...params, product: products, frameworks },
+        meta: { route_reason: reason, environment_downgraded: env },
+    };
+}
+// ========== 主入口（基于 prompt 真值归一化） ==========
+export function preflightRoute(params) {
+    const prompt = params.prompt ?? '';
+    const explicitFrameworks = params.frameworks?.length ? uniq(params.frameworks) : undefined;
+    const explicitProducts = params.product?.length ? uniq(params.product) : undefined;
+    // Phase 1
+    const signals = collectPromptSignals(prompt);
+    // —— 真跨主题（多产品信号）直接 passthrough
+    const nonTrtcProducts = [...signals.products].filter((p) => p !== 'trtc');
+    if (nonTrtcProducts.length > 1) {
+        return passthroughPlan(params, 'cross_product_passthrough', false);
+    }
+    // Phase 2: frameworks 调和
+    const fwRecon = reconcileWithPrompt(signals.platforms, explicitFrameworks);
+    const frameworks = fwRecon.values;
+    const envDowngraded = signals.hostEnv != null && !frameworks.includes(signals.hostEnv);
+    // Phase 2.5: products 调和（仅当 prompt 有非 TRTC 单产品时取此产品；TRTC 由 platform 推断 rtcengine/native_trtc_sdk）
+    const explicitTrtcBoth = explicitProducts?.includes('rtcengine') && explicitProducts?.includes('native_trtc_sdk');
+    const explicitTrtcAny = explicitProducts?.some((p) => p === 'rtcengine' || p === 'native_trtc_sdk') ?? false;
+    const isTrtcFamily = signals.products.has('trtc') || explicitTrtcBoth || (explicitTrtcAny && !nonTrtcProducts[0]);
+    // —— TRTC 家族：产物随平台分流
+    if (isTrtcFamily) {
+        const hasWeb = frameworks.includes('web');
+        const natives = frameworks.filter((f) => f !== 'web');
+        // TRTC + 多平台（web + native）→ product 数组，search core filterPool 原生接受多 product 交集
+        if (hasWeb && natives.length > 0) {
+            const reason = fwRecon.corrected ? 'trtc_multi_platform_prompt_override' : 'trtc_multi_platform';
+            return trcMultiProductPlan(params, ['rtcengine', 'native_trtc_sdk'], frameworks, reason, envDowngraded);
+        }
+        // TRTC + 仅 web
+        if (hasWeb) {
+            const reason = fwRecon.corrected ? 'trtc_web_prompt_override' : 'trtc_web';
+            return singleProductPlan(params, 'rtcengine', ['web'], signals.hostEnv, reason);
+        }
+        // TRTC + 仅 native
+        if (natives.length > 0) {
+            const reason = fwRecon.corrected ? 'trtc_native_prompt_override' : 'trtc_native';
+            return singleProductPlan(params, 'native_trtc_sdk', natives, signals.hostEnv, reason);
+        }
+        // TRTC 但 prompt / explicit 都没有可用平台 → 透传
+        return passthroughPlan(params, 'trtc_no_platform_resolved', false);
+    }
+    // —— 单非 TRTC 产品
+    if (nonTrtcProducts.length === 1) {
+        const product = nonTrtcProducts[0];
+        // 多平台 → product 单值，frameworks 多值（search core filterPool 原生按 framework 交集过滤）
+        if (frameworks.length >= 2) {
+            const reason = fwRecon.corrected ? 'multi_platform_prompt_override' : 'multi_platform';
+            return singleProductPlan(params, product, frameworks, signals.hostEnv, reason);
+        }
+        // 单平台
+        if (frameworks.length === 1) {
+            const reason = fwRecon.corrected ? 'single_platform_prompt_override' : 'single_platform';
+            return singleProductPlan(params, product, frameworks, signals.hostEnv, reason);
+        }
+        // 无平台信号：透传（保留 explicit frameworks）
+        if (explicitFrameworks?.length) {
+            return singleProductPlan(params, product, uniq(explicitFrameworks), signals.hostEnv, 'product_no_platform_explicit');
+        }
+        return passthroughPlan(params, 'product_no_platform_signal', false);
+    }
+    // —— 无产品信号：纯透传
+    return passthroughPlan(params, 'no_product_signal', false);
+}

package/utils/report-cls-client.js

@@ -32,7 +32,7 @@ function sendCLSReport(reportData) {
     const time = Date.now();
     const data = {
         platform: 'mcp',
-        verison: '1.7.0',
+        verison: '1.7.2',
         sdkappid: `${SDKAppID}`,
         method: method,
         userid: queryID || userID,

package/utils/report-es-client.js

@@ -7,7 +7,7 @@ function reportESClient(reportData) {
         report: [
             {
                 platform: 'mcp',
-                version: '1.7.0',
+                version: '1.7.2',
                 sdkappid: `${SDKAppID}`,
                 method: method,
                 userID: `${userID}`,