@cc-monitoring/proxy-core 0.2.3 → 0.3.0

package/dist/utils/pi-runtime.js

@@ -2,16 +2,16 @@
 // - ModelRuntime 懒加载，全局单例
 // - registerProvider 注册一个 "ccm" provider，走 anthropic-messages 协议到本地 proxy
 // - createPiSession() 返回一个新 AgentSession，cwd 钉死在 docsDir
-// - 按 chat / learn 两种模式挂载不同工具集
+// - 按 chat / outline / refine / tag 四种模式挂载不同工具集 + system prompt
 import fs from 'node:fs';
 import path from 'node:path';
 import os from 'node:os';
 import { Type } from 'typebox';
-import { createAgentSession, DefaultResourceLoader, ModelRuntime, createWriteToolDefinition, createEditToolDefinition, } from '@earendil-works/pi-coding-agent';
+import { createAgentSession, DefaultResourceLoader, ModelRuntime, createWriteToolDefinition, } from '@earendil-works/pi-coding-agent';
 import { dataDir } from './paths.js';
-import { sandboxWriteOps, sandboxEditOps, } from './pi-sandbox-ops.js';
-import { browserTools } from './pi-browser-tools.js';
+import { sandboxWriteOps } from './pi-sandbox-ops.js';
 import { knowledgeStore } from './knowledge-store.js';
+import { createCrawlTools } from './pi-crawl-tools.js';
 // 知识文档目录：所有文件工具的 cwd 和沙箱白名单根
 const docsDir = path.join(dataDir, 'docs');
 // pi agent 目录（auth / models / system.md）
@@ -59,70 +59,196 @@ async function getRuntime(port) {
     }
     return runtimePromise;
 }
-// 默认 system prompt 兜底（system.md 不存在时使用）
-const DEFAULT_SYSTEM_PROMPT = `You are a knowledge management AI operating inside a sandbox.
-Your knowledge directory (~/.cc-monitoring/docs) stores a **dictionary**: concise, distilled entries a future reader can look up. You do NOT copy source text — you digest it.
+const KB_STRUCTURE = `Your knowledge directory (~/.cc-monitoring/docs) stores a **dictionary**: concise, distilled entries a future reader can look up. You do NOT copy source text — you digest it.
 
-STRUCTURE:
-- Each topic is a subfolder: docs/<topic-slug>/
-- Each topic has docs/<slug>/index.json (metadata + chunk list, NO body)
-- Bodies live in docs/<topic-slug>/chunks/chunk-NNN.md
-- Root docs/index.json lists all topics as JSON
+STRUCTURE (all paths below are RELATIVE TO the knowledge root ~/.cc-monitoring/docs — the write tool's working directory IS this root, so NEVER prefix a path with "docs/"):
+- Each topic is a subfolder: <topic-slug>/
+- Each topic has <slug>/index.json (metadata + chunk list, NO body)
+- Bodies live in <topic-slug>/chunks/chunk-NNN.md
+- Root index.json lists all topics as JSON
+- Fetched source pages for a topic are already prepared in <slug>/sources/ — you only READ them, never write or modify them`;
+export const CHAT_SYSTEM_PROMPT = `You are a knowledge management AI operating inside a sandbox.
+${KB_STRUCTURE}
 
-CHAT MODE (answer questions):
+ANSWERING QUESTIONS:
 1. Call list_topics to see available topics
 2. If topic is unclear, call search_chunks(query) to find relevant chunks across topics
 3. Call get_topic(topic) to see that topic's chunk list
 4. Call read_chunk(topic, chunk_id) for relevant chunks — only read chunks you need
 5. Prefer answering from knowledge base chunks when relevant content exists; cite which chunks you used
-6. If no relevant chunk exists, answer from your general knowledge; if you truly don't know, say so
+6. If no relevant chunk exists, answer from your general knowledge; if you truly don't know, say so`;
+export const OUTLINE_SYSTEM_PROMPT = `You are an outline-phase worker. The relevant web pages have ALREADY been fetched and saved to sources/ for your assigned topic. Your job: read them and propose an outline of dictionary entries. Do NOT write chunks.
 
-LEARN MODE (distill material into a dictionary — NOT a copy of the source):
-0. First READ all input material end-to-end. Do NOT start writing chunk-by-chunk as you read. Before writing, internally outline the distinct concepts you extracted. Only then begin.
-1. Call list_topics first to check for existing similar topics — reuse, don't duplicate
-2. Use browser tools to read web pages if the input is a URL
-3. For each concept, produce a dictionary entry. An entry is a STAND-ALONE, retrievable note — the future reader has NO access to the source document. It must make sense alone.
-4. Write each chunk as markdown with this structure (write in the source document's language):
+${KB_STRUCTURE}
 
-   \`\`\`
-   # <term or concept name>
+WORKFLOW:
+1. Call read_source(topic) to load ALL fetched source pages (multiple pages are separated by ---).
+2. Outline the DISTINCT concepts actually present across these pages. One concept = one future chunk. Do not bundle unrelated ideas.
+3. Call add_outline(slug, title, items) ONCE at the end:
+   - slug: the EXACT slug given in your task
+   - title: human-readable topic title
+   - items: array of { id, title, oneLine } — id is a short stable slug per item (e.g. "use-state"), oneLine is a single-sentence definition
 
-   - 定义 / Definition: <one sentence, in your own words>
-   - 是什么 / What it is: <2-4 bullets, the core idea>
-   - 不是什么 / What it is NOT: <common confusions, if any>
-   - 关键点 / Key points: <3-6 bullets, the load-bearing facts>
-   - 适用场景 / When to use: <where this applies>
-   - 误区 / Pitfalls: <common mistakes, edge cases>
-   - 来源 / Source: <url>
-   \`\`\`
+HARD RULES:
+- Use ONLY content present in the fetched sources. Do NOT add concepts from training-data memory. If the sources are thin, outline only what is there.
+- read_source MUST be your first call. If it returns an error, STOP and report it immediately.
+- Do NOT write chunk files, source files, or index.json.
+- Do NOT summarize prematurely — outline only; distillation happens later.
+- After add_outline succeeds, print the returned slug and item count. Stop.`;
+export const CRAWL_SYSTEM_PROMPT = `You are a crawl-phase worker. You drive the browser YOURSELF to incrementally fetch web pages for your assigned topic and save every useful page into sources/. Do NOT outline or write chunks.
 
-5. RULES — these are hard constraints:
-   - NO large quotes. A short quote (≤2 sentences) is allowed ONLY as evidence under a bullet you already wrote in your own words.
-   - Every bullet must be YOU re-expressing the idea. If you find yourself copying sentences, stop and rephrase.
-   - One concept per chunk. Do not bundle unrelated ideas into one entry.
-   - If the material is thin on a concept, write what exists; do NOT invent content to fill sections. Omit empty sections.
-6. Write chunks to docs/<topic-slug>/chunks/chunk-NNN.md (number sequentially from 001)
-7. Call add_topic(slug, title, summary, tags) to register/update the topic in the index
-8. Call add_chunk(topic, chunk_id, title, summary) for each chunk you wrote
-9. Do NOT write or edit index.json files directly — use add_topic/add_chunk tools
+WORKFLOW:
+1. Call crawl_open to open the entry URL given in your task.
+2. Call save_source(url) to save the CURRENT page.
+3. From the page's links, pick same-site, valuable CONTENT pages — skip login/signup, navigation chrome, language switchers, pagination, and pure anchors. Call crawl_visit(url), then save_source(url) for each page.
+4. When there are no more useful links, or the page/depth limit is reached, call crawl_finish, print the returned stats, and stop.
 
-FINISHING:
-- After all chunks are written, output a plain summary: the topic created/updated, and the list of chunk titles. Do NOT ask the user what to deepen next. The learning task is complete.
-`;
-/** 从 ~/.cc-monitoring/pi/system.md 读取自定义 system prompt，不存在则用兜底文本 */
-function loadSystemPrompt() {
+HARD RULES:
+- Write ONLY into sources/ via save_source. Never write other files.
+- Never visit cross-site URLs.
+- If a tool returns an error, do NOT retry the same URL; pick another link or finish.
+- At most ~20 pages. Prefer a few dense pages over many thin ones.
+- After crawl_finish succeeds, stop.`;
+export const REFINE_SYSTEM_PROMPT = `You are a refine-phase worker. For ONE item from the topic's outline, produce a single dictionary chunk.
+
+${KB_STRUCTURE}
+
+WORKFLOW:
+1. Call read_source(topic) to load ALL fetched source pages (multiple pages separated by ---).
+2. Focus on your assigned item. Do NOT distill other items.
+3. Write ONE chunk file <topic>/chunks/chunk-NNN.md (number sequentially from 001; path RELATIVE TO the knowledge root, no "docs/" prefix) using the write tool, in the source document's language, with this structure:
+
+\`\`\`
+# <term or concept name>
+
+- 定义 / Definition: <one sentence, in your own words>
+- 是什么 / What it is: <2-4 bullets, the core idea>
+- 不是什么 / What it is NOT: <common confusions, if any>
+- 关键点 / Key points: <3-6 bullets, the load-bearing facts>
+- 适用场景 / When to use: <where this applies>
+- 误区 / Pitfalls: <common mistakes, edge cases>
+- 来源 / Source: <url(s) of the page(s) you actually used — copy them from the "Source:" headers at the top of each fetched page; list several URLs if the chunk draws on several pages>
+\`\`\`
+
+4. Call add_chunk(topic, chunk_id, title, summary, tags) to register it.
+   - tags: 3-6 lowercase hyphenated tags describing the concept (synonyms, key terms).
+
+HARD RULES:
+- read_source MUST be your first call. If it returns an error, STOP and report it immediately. Do NOT write a placeholder source.md yourself, and do NOT write the chunk from memory.
+- Every fact must come from the fetched sources. If a claim is not supported by them, omit it. Never fill gaps from memory.
+- NO large quotes. A short quote (≤2 sentences) is allowed ONLY as evidence under a bullet you already wrote in your own words.
+- Every bullet must be YOU re-expressing the idea. If you find yourself copying sentences, stop and rephrase.
+- One concept per chunk.
+- If the material is thin on a concept, write what exists; do NOT invent content. Omit empty sections.
+- Do NOT call add_topic. Do NOT write index.json.`;
+export const TAG_SYSTEM_PROMPT = `You are a tag-phase worker. Enrich an existing topic with rich, retrieval-friendly tags.
+
+${KB_STRUCTURE}
+
+WORKFLOW:
+1. Call get_topic(slug) to list this topic's chunks.
+2. Call read_chunk for the chunks to understand the actual content.
+3. Generate 5 categories of tags (10-20 total), all lowercase hyphenated:
+   - synonyms: alternate names / aliases
+   - hypernym: broader parent concept (e.g. "state-management" for redux)
+   - trigger-scenarios: situations where a reader would reach for this (e.g. "form-validation")
+   - confusable: things this is often mixed up with
+   - typical-questions: questions a reader might search (e.g. "how-does-usestate-work")
+4. Call add_topic(slug, title, summary, tags) to OVERWRITE the topic's tags + metadata. Reuse the existing title and summary from get_topic; do not change them.
+
+HARD RULES:
+- Do NOT write chunk files. Do NOT call add_chunk.
+- Do NOT invent tags the content does not support.`;
+/** 从 ~/.cc-monitoring/pi/system.md 读取自定义 system prompt（仅 chat 模式覆盖） */
+function loadChatSystemPrompt() {
     const mdPath = path.join(agentDir, 'system.md');
     try {
         return fs.readFileSync(mdPath, 'utf8');
     }
     catch {
-        return DEFAULT_SYSTEM_PROMPT;
+        return CHAT_SYSTEM_PROMPT;
+    }
+}
+function systemPromptForMode(mode) {
+    switch (mode) {
+        case 'chat':
+            return loadChatSystemPrompt();
+        case 'outline':
+            return OUTLINE_SYSTEM_PROMPT;
+        case 'crawl':
+            return CRAWL_SYSTEM_PROMPT;
+        case 'refine':
+            return REFINE_SYSTEM_PROMPT;
+        case 'tag':
+            return TAG_SYSTEM_PROMPT;
+        case 'reconcile':
+            return 'Return ONLY a single minified JSON object, no prose, no code fence.';
+    }
+}
+// ============================================================
+// 工具构造
+const ok = (text) => ({ content: [{ type: 'text', text }], details: undefined });
+const err = (e) => ({ content: [{ type: 'text', text: `error: ${e instanceof Error ? e.message : String(e)}` }], details: undefined });
+// 与 knowledge-store TOPIC_RE 保持一致：防路径穿越（../、绝对路径、反斜杠等）
+const SLUG_RE = /^[a-z0-9][a-z0-9-]*$/;
+function assertSlug(slug, field) {
+    if (!SLUG_RE.test(slug)) {
+        throw new Error(`invalid ${field}: ${slug}`);
     }
 }
-/** 构造知识检索领域工具（chat / learn 共用） */
+/** add_outline：对已存在的 slug 目录登记大纲（写 outline.json + seed topic） */
+function buildAddOutlineTool() {
+    return {
+        name: 'add_outline',
+        label: 'add_outline',
+        description: '在已由抓取阶段准备好的 <slug> 目录登记大纲（写 outline.json + seed topic），只调用一次。',
+        parameters: Type.Object({
+            slug: Type.String({ description: '任务指定的小写连字符 slug，如 react-hooks' }),
+            title: Type.String(),
+            items: Type.Array(Type.Object({
+                id: Type.String({ description: '条目稳定 id，如 use-state' }),
+                title: Type.String(),
+                oneLine: Type.String({ description: '一句话定义' }),
+            })),
+        }),
+        async execute(_id, params) {
+            assertSlug(params.slug, 'slug');
+            const topicDir = path.join(docsDir, params.slug);
+            if (!fs.existsSync(topicDir))
+                throw new Error(`topic dir not prepared: ${params.slug}`);
+            if (!fs.existsSync(path.join(topicDir, 'sources')))
+                throw new Error(`sources missing for topic: ${params.slug}`);
+            fs.writeFileSync(path.join(topicDir, 'outline.json'), JSON.stringify({ slug: params.slug, source: 'sources/', items: params.items }, null, 2));
+            await knowledgeStore.upsertTopic({
+                slug: params.slug,
+                title: params.title,
+                summary: '(refining)',
+                tags: [],
+            });
+            return ok(JSON.stringify({ slug: params.slug, itemCount: params.items.length }));
+        },
+    };
+}
+/** read_source：读取该主题 sources/ 目录下全部已抓取网页 */
+function buildReadSourceTool() {
+    return {
+        name: 'read_source',
+        label: 'read_source',
+        description: '读取该主题 sources/ 目录下全部已抓取网页（按文件名排序、--- 分隔）。',
+        parameters: Type.Object({ topic: Type.String({ description: '主题 slug' }) }),
+        async execute(_id, params) {
+            assertSlug(params.topic, 'topic');
+            const dir = path.join(docsDir, params.topic, 'sources');
+            const files = fs.readdirSync(dir).filter((f) => f.endsWith('.md')).sort();
+            if (files.length === 0)
+                throw new Error(`no sources for topic: ${params.topic}`);
+            const merged = files.map((f) => fs.readFileSync(path.join(dir, f), 'utf8')).join('\n\n---\n\n');
+            return ok(merged);
+        },
+    };
+}
+/** 知识检索领域工具 */
 function buildKnowledgeTools() {
-    const ok = (text) => ({ content: [{ type: 'text', text }], details: undefined });
-    const err = (e) => ({ content: [{ type: 'text', text: `error: ${e instanceof Error ? e.message : String(e)}` }], details: undefined });
     return [
         {
             name: 'list_topics',
@@ -212,6 +338,7 @@ function buildKnowledgeTools() {
                 chunk_id: Type.String({ description: '如 chunk-001' }),
                 title: Type.String(),
                 summary: Type.String(),
+                tags: Type.Array(Type.String({ description: '小写连字符标签，3-6 个' })),
             }),
             async execute(_id, params) {
                 try {
@@ -220,6 +347,7 @@ function buildKnowledgeTools() {
                         chunkId: params.chunk_id,
                         title: params.title,
                         summary: params.summary,
+                        tags: params.tags,
                     });
                     return ok('chunk ref added');
                 }
@@ -230,30 +358,45 @@ function buildKnowledgeTools() {
         },
     ];
 }
+const KNOWLEDGE_TOOLS = buildKnowledgeTools();
+const KT = (name) => KNOWLEDGE_TOOLS.find((t) => t.name === name);
+const WRITE_TOOL = createWriteToolDefinition(docsDir, { operations: sandboxWriteOps });
 /** 按模式构建工具集（ToolDefinition 数组，用于 customTools） */
-function buildModeTools(mode) {
-    const knowledgeTools = buildKnowledgeTools();
-    if (mode === 'chat') {
-        // chat：只用领域工具，移除通用 read/grep/find/ls（防全量灌入）
-        return knowledgeTools;
+function buildModeTools(mode, crawlGuard) {
+    switch (mode) {
+        case 'chat':
+            return [KT('list_topics'), KT('get_topic'), KT('search_chunks'), KT('read_chunk')];
+        case 'outline':
+            return [buildReadSourceTool(), buildAddOutlineTool()];
+        case 'crawl':
+            if (!crawlGuard)
+                throw new Error('crawl 模式必须传入 crawlGuard');
+            return createCrawlTools(crawlGuard);
+        case 'refine':
+            return [buildReadSourceTool(), WRITE_TOOL, KT('add_chunk')];
+        case 'tag':
+            return [KT('list_topics'), KT('get_topic'), KT('read_chunk'), KT('add_topic')];
+        // reconcile：无工具纯 JSON 调用
+        case 'reconcile':
+            return [];
     }
-    // learn：领域工具（查重/读现有）+ write/edit + browser
-    // 注意：createXxxToolDefinition 返回带具体 schema 的 ToolDefinition<TParams, ...>，
-    // 与无泛型的 ToolDefinition 在 renderCall 上不协变，需要 as ToolDefinition 抹平。
-    return [
-        ...knowledgeTools,
-        createWriteToolDefinition(docsDir, { operations: sandboxWriteOps }),
-        createEditToolDefinition(docsDir, { operations: sandboxEditOps }),
-        ...browserTools,
-    ];
 }
 /** 按模式构建工具白名单（tools 字段，全小写） */
 function buildToolAllowlist(mode) {
-    const base = ['list_topics', 'get_topic', 'search_chunks', 'read_chunk'];
-    if (mode === 'chat') {
-        return base;
+    switch (mode) {
+        case 'chat':
+            return ['list_topics', 'get_topic', 'search_chunks', 'read_chunk'];
+        case 'outline':
+            return ['read_source', 'add_outline'];
+        case 'crawl':
+            return ['crawl_open', 'save_source', 'crawl_visit', 'crawl_finish'];
+        case 'refine':
+            return ['read_source', 'write', 'add_chunk'];
+        case 'tag':
+            return ['list_topics', 'get_topic', 'read_chunk', 'add_topic'];
+        case 'reconcile':
+            return [];
     }
-    return [...base, 'add_topic', 'add_chunk', 'write', 'edit', ...browserTools.map((t) => t.name)];
 }
 export async function createPiSession(options) {
     const runtime = await getRuntime(options.port);
@@ -267,8 +410,8 @@ export async function createPiSession(options) {
     // 但 before_provider_headers handler 是每次上游请求时才触发，
     // 所以先放一个可变变量，返回 session 后再填值。
     let currentSessionId = '';
-    const systemPrompt = loadSystemPrompt();
-    const modeTools = buildModeTools(options.mode);
+    const systemPrompt = systemPromptForMode(options.mode);
+    const modeTools = buildModeTools(options.mode, options.crawlGuard);
     const toolAllowlist = buildToolAllowlist(options.mode);
     const resourceLoader = new DefaultResourceLoader({
         cwd,
@@ -303,3 +446,47 @@ export async function createPiSession(options) {
     currentSessionId = session.sessionId;
     return session;
 }
+// ============================================================
+// 无工具 JSON 调用（reconcile 模式）
+/**
+ * 从任意文本截取第一个 `{` 到最后一个 `}` 之间的子串并解析为 JSON 对象。
+ * 容忍 prose / 代码围栏包裹。无 `{` 或解析失败时抛错。
+ */
+export function extractJsonObject(raw) {
+    const start = raw.indexOf('{');
+    const end = raw.lastIndexOf('}');
+    if (start === -1 || end === -1 || end < start) {
+        throw new Error('no json object in response');
+    }
+    try {
+        return JSON.parse(raw.slice(start, end + 1));
+    }
+    catch {
+        throw new Error('no json object in response');
+    }
+}
+/**
+ * 建 reconcile（无工具）session，跑一次纯 JSON 调用并返回解析结果。
+ * instructions 与 prompt 合并为单条用户消息：规则在前，JSON 数据在后。
+ */
+export async function runJsonPrompt(opts) {
+    const session = await createPiSession({ port: opts.port, mode: 'reconcile' });
+    // 订阅累积 assistant 文本 chunk
+    let accumulated = '';
+    const unsub = session.subscribe((ev) => {
+        if (ev.type === 'message_update') {
+            const ae = ev.assistantMessageEvent;
+            if (ae?.type === 'text_delta' && typeof ae.delta === 'string') {
+                accumulated += ae.delta;
+            }
+        }
+    });
+    try {
+        await session.prompt(`${opts.instructions}\n\n${opts.prompt}`);
+        await session.waitForIdle();
+    }
+    finally {
+        unsub();
+    }
+    return extractJsonObject(accumulated);
+}

package/dist/utils/web-crawler.d.ts

@@ -0,0 +1,13 @@
+export declare const CRAWL_MAX_PAGES = 20;
+export declare const CRAWL_MAX_DEPTH = 1;
+/**
+ * 从 URL 生成 slug。
+ * host 取去 www. 后的首个 DNS label（pi.dev → pi），pathname 最多取 3 个非泛段。
+ */
+export declare function slugFromUrl(raw: string): string;
+/** 从文件名/标签生成 slug：取 basename、去扩展名后清洗，空则 'topic' */
+export declare function slugFromLabel(label: string): string;
+/** 同站判定：去 www.、小写后相等，或 h 是 base 的子域 */
+export declare function isSameSite(hostname: string, baseHostname: string): boolean;
+/** 解析候选链接：仅 http(s)、清 hash、排除二进制/文档资源；无法解析返回 null */
+export declare function normalizeCandidate(href: string, entry: URL): URL | null;

package/dist/utils/web-crawler.js

@@ -0,0 +1,64 @@
+// web-crawler：抓取相关纯函数（slug 生成 / 同站判定 / 候选链接归一）与抓取上限常量。
+// 渐进抓取的有状态逻辑见 pi-crawl-tools.ts（CrawlGuard）。
+export const CRAWL_MAX_PAGES = 20;
+export const CRAWL_MAX_DEPTH = 1;
+// ---------- 纯函数辅助 ----------
+const GENERIC_SEGMENTS = new Set(['docs', 'doc']);
+/** 单段清洗：小写 + 非 a-z0-9 字符折叠成 -（不做首尾裁剪，留给上层统一处理） */
+function cleanSegment(seg) {
+    return seg.toLowerCase().replace(/[^a-z0-9]+/g, '-');
+}
+/** 校验 slug 合法性，非法兜底 'topic' */
+function finalizeSlug(slug) {
+    const s = slug
+        .replace(/-+/g, '-')
+        .replace(/^-+|-+$/g, '')
+        .slice(0, 60)
+        .replace(/-+$/, '');
+    return /^[a-z0-9][a-z0-9-]*$/.test(s) ? s : 'topic';
+}
+/**
+ * 从 URL 生成 slug。
+ * host 取去 www. 后的首个 DNS label（pi.dev → pi），pathname 最多取 3 个非泛段。
+ */
+export function slugFromUrl(raw) {
+    const u = new URL(raw);
+    const hostLabel = u.hostname.replace(/^www\./, '').split('.')[0];
+    const segments = [cleanSegment(hostLabel)];
+    for (const part of u.pathname.split('/')) {
+        const seg = cleanSegment(part);
+        if (seg && !GENERIC_SEGMENTS.has(seg))
+            segments.push(seg);
+    }
+    return finalizeSlug(segments.slice(0, 4).join('-'));
+}
+/** 从文件名/标签生成 slug：取 basename、去扩展名后清洗，空则 'topic' */
+export function slugFromLabel(label) {
+    const base = label.split(/[\\/]/).pop() ?? label;
+    const stem = base.replace(/\.[^.]+$/, '');
+    return finalizeSlug(cleanSegment(stem));
+}
+/** 同站判定：去 www.、小写后相等，或 h 是 base 的子域 */
+export function isSameSite(hostname, baseHostname) {
+    const norm = (h) => h.toLowerCase().replace(/^www\./, '');
+    const h = norm(hostname);
+    const base = norm(baseHostname);
+    return h === base || h.endsWith(`.${base}`);
+}
+const BINARY_EXTENSIONS = /\.(pdf|zip|rar|7z|exe|dmg|pkg|png|jpg|jpeg|gif|svg|webp|mp3|mp4|mov|avi|doc|docx|xls|xlsx|ppt|pptx)$/i;
+/** 解析候选链接：仅 http(s)、清 hash、排除二进制/文档资源；无法解析返回 null */
+export function normalizeCandidate(href, entry) {
+    let u;
+    try {
+        u = new URL(href, entry);
+    }
+    catch {
+        return null;
+    }
+    if (u.protocol !== 'http:' && u.protocol !== 'https:')
+        return null;
+    u.hash = '';
+    if (BINARY_EXTENSIONS.test(u.pathname))
+        return null;
+    return u;
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cc-monitoring/proxy-core",
-  "version": "0.2.3",
+  "version": "0.3.0",
   "description": "Claude Code 代理工具 + Token 消耗监控面板",
   "keywords": [
     "claude-code-proxy",

package/dist/ui/assets/channel-BdcV3Ykv.js

@@ -1 +0,0 @@
-import{aj as o,ak as n}from"./index-BVM6YOjH.js";const t=(a,r)=>o.lang.round(n.parse(a)[r]);export{t as c};

package/dist/ui/assets/classDiagram-OUVF2IWQ-CAZR5rAv.js

@@ -1 +0,0 @@
-import{s as a,c as s,a as e,C as t}from"./chunk-V7JOEXUC-Cbhq5qgO.js";import{_ as i}from"./index-BVM6YOjH.js";import"./chunk-5VM5RSS4-BQ0aVGl-.js";import"./chunk-XXDRQBXY-CKrih4pC.js";import"./chunk-VR4S4FIN-D6WJuaS1.js";import"./chunk-32BRIVSS-C7oMpoeu.js";var u={parser:e,get db(){return new t},renderer:s,styles:a,init:i(r=>{r.class||(r.class={}),r.class.arrowMarkerAbsolute=r.arrowMarkerAbsolute},"init")};export{u as diagram};

package/dist/ui/assets/classDiagram-v2-EOCWNBFH-CAZR5rAv.js

@@ -1 +0,0 @@
-import{s as a,c as s,a as e,C as t}from"./chunk-V7JOEXUC-Cbhq5qgO.js";import{_ as i}from"./index-BVM6YOjH.js";import"./chunk-5VM5RSS4-BQ0aVGl-.js";import"./chunk-XXDRQBXY-CKrih4pC.js";import"./chunk-VR4S4FIN-D6WJuaS1.js";import"./chunk-32BRIVSS-C7oMpoeu.js";var u={parser:e,get db(){return new t},renderer:s,styles:a,init:i(r=>{r.class||(r.class={}),r.class.arrowMarkerAbsolute=r.arrowMarkerAbsolute},"init")};export{u as diagram};