autosnippet
Version:
Extract code patterns into a knowledge base for AI coding assistants
208 lines (207 loc) • 7.33 kB
JavaScript
/**
* AgentRouter — Intent → Preset 解析器
*
* 在统一架构下，Router 的职责简化为:
* - 理解用户意图 (intent classification)
* - 映射到正确的 Preset (配置组合)
* - 创建 AgentRuntime 并执行
*
* 路由策略 (优先级递减):
* 1. 手动指定 (API `preset` 参数)
* 2. 渠道特征 (如飞书终端命令以 > 开头 → remote-exec)
* 3. 关键词匹配 (零延迟正则)
* 4. LLM 意图分类 (精确但需 AI 调用)
* 5. 默认 → chat
*
* 关键区别 (vs 旧 AgentRouter):
* - 旧版路由到 AgentMode (chat/bootstrap/scan/lark_bridge) → 4 种 Agent 类型
* - 新版路由到 Preset (chat/insight/remote-exec) → 同一 Runtime 的不同配置
* - 不再存在 "LarkBridge" 模式 — 飞书和前端共享同一 Preset
*
* @module AgentRouter
*/
import Logger from '#infra/logging/Logger.js';
import { AgentEventBus } from './AgentEventBus.js';
import { Channel } from './AgentMessage.js';
/** Preset 名称枚举 */
export const PresetName = Object.freeze({
CHAT: 'chat',
INSIGHT: 'insight',
LARK: 'lark',
REMOTE_EXEC: 'remote-exec',
});
/** 关键词 → Preset 映射 (零延迟快速路径) */
const KEYWORD_ROUTES = [
{
preset: PresetName.INSIGHT,
keywords: [
/冷启动|cold[\s-]?start|初始化知识库|bootstrap/i,
/重建知识库|rebuild.*knowledge/i,
/全项目.*分析|analyze.*entire.*project/i,
/扫描|scan|分析.*目录|分析.*文件夹|analyze.*folder/i,
/检查.*target|审计.*模块|audit.*module/i,
/深度分析.*路径|知识提取/i,
],
},
{
preset: PresetName.REMOTE_EXEC,
keywords: [
/^[>$]\s*/, // 以 > 或 $ 开头 = 终端命令
/运行命令|执行命令|run.*command|exec.*command/i,
],
},
];
/** LLM 路由分类 Schema */
const ROUTE_CLASSIFICATION_SCHEMA = {
name: 'classify_intent',
description: 'Classify user intent to determine which agent preset to use',
parameters: {
type: 'object',
properties: {
preset: {
type: 'string',
enum: ['chat', 'insight', 'lark', 'remote-exec'],
description: 'The preset to route to. chat=conversation, insight=code analysis and knowledge extraction, lark=Lark knowledge management, remote-exec=terminal commands',
},
confidence: {
type: 'number',
description: 'Confidence score 0-1',
},
reasoning: {
type: 'string',
description: 'Brief reasoning for the classification',
},
},
required: ['preset', 'confidence'],
},
};
export class AgentRouter {
#logger;
#bus;
#aiProvider = null;
#executor = null;
constructor() {
this.#logger = Logger.getInstance();
this.#bus = AgentEventBus.getInstance();
}
/** 设置 AI Provider (用于 LLM 路由) */
setAiProvider(provider) {
this.#aiProvider = provider;
}
/** 设置执行器 — Factory 提供的 (presetName, message, opts) => AgentResult */
setExecutor(executor) {
this.#executor = executor;
}
/**
* 路由并执行
*
* @param message 统一消息
* @param [opts.preset] 手动指定 Preset (跳过路由)
* @param [opts.strategyOpts] 策略特定选项 (如 FanOut 的 items)
* @returns >}
*/
async route(message, opts = {}) {
// 1. 手动指定
let preset = opts.preset || message.metadata?.mode;
// 2. 渠道特征路由
if (!preset) {
preset = this.#matchChannel(message);
if (preset) {
this.#logger.info(`[AgentRouter] Channel match → ${preset}`);
}
}
// 3. 关键词匹配
if (!preset) {
preset = this.#matchKeyword(message.content);
if (preset) {
this.#logger.info(`[AgentRouter] Keyword match → ${preset}`);
}
}
// 4. LLM 分类
if (!preset && this.#aiProvider) {
preset = await this.#classifyWithLLM(message.content);
if (preset) {
this.#logger.info(`[AgentRouter] LLM classification → ${preset}`);
}
}
// 5. 默认 → chat
if (!preset) {
preset = PresetName.CHAT;
}
// 执行
if (!this.#executor) {
throw new Error('[AgentRouter] No executor set. Call setExecutor() first.');
}
this.#logger.info(`[AgentRouter] Dispatching → preset="${preset}" channel=${message.channel}`);
const result = await this.#executor(preset, message, opts);
return { preset, result };
}
/**
* 仅分类意图，不执行
* @returns preset name
*/
async classify(message) {
let preset = message.metadata?.mode;
if (!preset) {
preset = this.#matchChannel(message);
}
if (!preset) {
preset = this.#matchKeyword(message.content);
}
if (!preset && this.#aiProvider) {
preset = await this.#classifyWithLLM(message.content);
}
return preset || PresetName.CHAT;
}
// ─── 私有方法 ────────────────────────────────
/**
* 渠道特征路由
* 飞书消息默认 → lark preset (知识管理对话)
* 飞书终端命令 (> 开头) → remote-exec
*/
#matchChannel(message) {
if (message.channel === Channel.LARK) {
const text = message.content.trim();
// 飞书消息以 > 或 $ 开头 → 终端命令
if (/^[>$]\s*/.test(text)) {
return PresetName.REMOTE_EXEC;
}
// 默认走 lark preset (知识管理对话)
return PresetName.LARK;
}
return null;
}
#matchKeyword(prompt) {
for (const route of KEYWORD_ROUTES) {
for (const re of route.keywords) {
if (re.test(prompt)) {
return route.preset;
}
}
}
return null;
}
async #classifyWithLLM(prompt) {
try {
const result = await this.#aiProvider.chatWithTools(`Classify this user message into the correct preset: "${prompt.slice(0, 300)}"`, {
messages: [],
toolSchemas: [ROUTE_CLASSIFICATION_SCHEMA],
toolChoice: 'required',
systemPrompt: 'You classify user intents for an AI coding assistant. Respond by calling the classify_intent function.',
temperature: 0,
maxTokens: 200,
});
if (result.functionCalls?.[0]?.args?.preset) {
const classified = result.functionCalls[0].args;
if (classified.confidence > 0.6) {
return classified.preset;
}
}
}
catch (err) {
this.#logger.warn(`[AgentRouter] LLM classification failed: ${err.message}`);
}
return null;
}
}
export default AgentRouter;