mirror of
https://github.com/hellodigua/ChatLab.git
synced 2026-05-06 13:06:09 +08:00
digua
1016 lines
35 KiB
TypeScript
1016 lines
35 KiB
TypeScript
/**
|
||
* AI Agent 执行器
|
||
* 处理 Function Calling 循环，支持多轮工具调用
|
||
*/
|
||
|
||
import type { ChatMessage, ChatOptions, ChatStreamChunk, ToolCall } from './llm/types'
|
||
import { chatStream, chat } from './llm'
|
||
import { getAllToolDefinitions, executeToolCalls } from './tools'
|
||
import type { ToolContext, OwnerInfo } from './tools/types'
|
||
import { aiLogger } from './logger'
|
||
import { randomUUID } from 'crypto'
|
||
|
||
// 思考类标签列表（可按需扩展）
|
||
const THINK_TAGS = ['think', 'analysis', 'reasoning', 'reflection', 'thought', 'thinking']
|
||
const THINK_START_TAGS = THINK_TAGS.map((tag) => `<${tag}>`)
|
||
const TOOL_CALL_START_TAG = '<tool_call>'
|
||
const TOOL_CALL_END_TAG = '</tool_call>'
|
||
|
||
// ==================== Fallback 解析器 ====================
|
||
|
||
/**
|
||
* 从文本内容中提取思考类标签内容
|
||
*/
|
||
function extractThinkingContent(content: string): { thinking: string; cleanContent: string } {
|
||
if (!content) {
|
||
return { thinking: '', cleanContent: '' }
|
||
}
|
||
|
||
const tagPattern = THINK_TAGS.join('|')
|
||
const thinkRegex = new RegExp(`<(${tagPattern})>([\\s\\S]*?)<\\/\\1>`, 'gi')
|
||
const thinkingParts: string[] = []
|
||
let cleanContent = content
|
||
|
||
const matches = content.matchAll(thinkRegex)
|
||
for (const match of matches) {
|
||
const thinkText = match[2].trim()
|
||
if (thinkText) {
|
||
thinkingParts.push(thinkText)
|
||
}
|
||
cleanContent = cleanContent.replace(match[0], '')
|
||
}
|
||
|
||
return { thinking: thinkingParts.join('\n').trim(), cleanContent: cleanContent.trim() }
|
||
}
|
||
|
||
/**
|
||
* 从文本内容中解析 <tool_call> 标签并转换为标准 ToolCall 格式
|
||
*/
|
||
function parseToolCallTags(content: string): ToolCall[] | null {
|
||
const toolCallRegex = /<tool_call>\s*([\s\S]*?)\s*<\/tool_call>/gi
|
||
const toolCalls: ToolCall[] = []
|
||
|
||
const matches = content.matchAll(toolCallRegex)
|
||
for (const match of matches) {
|
||
try {
|
||
const jsonStr = match[1].trim()
|
||
const parsed = JSON.parse(jsonStr)
|
||
|
||
if (parsed.name && parsed.arguments) {
|
||
toolCalls.push({
|
||
id: `fallback-${randomUUID()}`,
|
||
type: 'function',
|
||
function: {
|
||
name: parsed.name,
|
||
arguments: typeof parsed.arguments === 'string' ? parsed.arguments : JSON.stringify(parsed.arguments),
|
||
},
|
||
})
|
||
}
|
||
} catch (e) {
|
||
aiLogger.warn('Agent', 'Failed to parse tool_call tag', { content: match[1], error: String(e) })
|
||
}
|
||
}
|
||
|
||
return toolCalls.length > 0 ? toolCalls : null
|
||
}
|
||
|
||
/**
|
||
* 检测内容是否包含工具调用标签（用于判断是否需要 fallback 解析）
|
||
*/
|
||
function hasToolCallTags(content: string): boolean {
|
||
return /<tool_call>/i.test(content)
|
||
}
|
||
|
||
/**
|
||
* 清理 <tool_call> 标签内容，避免将工具调用文本展示给用户
|
||
*/
|
||
function stripToolCallTags(content: string): string {
|
||
return content.replace(/<tool_call>[\s\S]*?<\/tool_call>/gi, '').trim()
|
||
}
|
||
|
||
type StreamMode = 'text' | 'think' | 'tool_call'
|
||
|
||
/**
|
||
* 创建流式解析器：将文本按 content/think/tool_call 分流
|
||
*/
|
||
function createStreamParser(handlers: {
|
||
onText: (text: string) => void
|
||
onThink: (text: string, tag: string) => void
|
||
onThinkStart?: (tag: string) => void
|
||
onThinkEnd?: (tag: string) => void
|
||
}): { push: (text: string) => void; flush: () => void } {
|
||
let buffer = ''
|
||
let mode: StreamMode = 'text'
|
||
let currentThinkTag = ''
|
||
|
||
const startTags = [...THINK_START_TAGS, TOOL_CALL_START_TAG]
|
||
const startTagsLower = startTags.map((tag) => tag.toLowerCase())
|
||
const toolCallStartLower = TOOL_CALL_START_TAG.toLowerCase()
|
||
const toolCallEndLower = TOOL_CALL_END_TAG.toLowerCase()
|
||
const maxStartTagLength = Math.max(...startTags.map((tag) => tag.length))
|
||
|
||
const findNextTagIndex = (lowerBuffer: string): { index: number; tag: string } | null => {
|
||
let hitIndex = -1
|
||
let hitTag = ''
|
||
for (const tag of startTagsLower) {
|
||
const index = lowerBuffer.indexOf(tag)
|
||
if (index !== -1 && (hitIndex === -1 || index < hitIndex)) {
|
||
hitIndex = index
|
||
hitTag = tag
|
||
}
|
||
}
|
||
return hitIndex === -1 ? null : { index: hitIndex, tag: hitTag }
|
||
}
|
||
|
||
const emitText = (text: string) => {
|
||
if (text) {
|
||
handlers.onText(text)
|
||
}
|
||
}
|
||
|
||
const emitThink = (text: string) => {
|
||
if (text) {
|
||
handlers.onThink(text, currentThinkTag || 'think')
|
||
}
|
||
}
|
||
|
||
const processBuffer = () => {
|
||
let safety = 0
|
||
while (buffer && safety < 10000) {
|
||
safety += 1
|
||
if (mode === 'text') {
|
||
const lowerBuffer = buffer.toLowerCase()
|
||
const hit = findNextTagIndex(lowerBuffer)
|
||
if (!hit) {
|
||
// 保留一段尾部，避免标签被截断
|
||
const keepLength = Math.max(1, maxStartTagLength - 1)
|
||
if (buffer.length > keepLength) {
|
||
emitText(buffer.slice(0, buffer.length - keepLength))
|
||
buffer = buffer.slice(buffer.length - keepLength)
|
||
}
|
||
break
|
||
}
|
||
|
||
if (hit.index > 0) {
|
||
emitText(buffer.slice(0, hit.index))
|
||
buffer = buffer.slice(hit.index)
|
||
}
|
||
|
||
const lowerHead = buffer.toLowerCase()
|
||
if (lowerHead.startsWith(hit.tag)) {
|
||
if (hit.tag === toolCallStartLower) {
|
||
mode = 'tool_call'
|
||
buffer = buffer.slice(TOOL_CALL_START_TAG.length)
|
||
continue
|
||
}
|
||
|
||
// 进入思考模式
|
||
currentThinkTag = hit.tag.slice(1, -1)
|
||
mode = 'think'
|
||
handlers.onThinkStart?.(currentThinkTag)
|
||
buffer = buffer.slice(startTags[startTagsLower.indexOf(hit.tag)].length)
|
||
continue
|
||
}
|
||
|
||
// 未识别的 < 视为普通文本
|
||
emitText(buffer.slice(0, 1))
|
||
buffer = buffer.slice(1)
|
||
continue
|
||
}
|
||
|
||
if (mode === 'think') {
|
||
const endTag = `</${currentThinkTag}>`
|
||
const lowerBuffer = buffer.toLowerCase()
|
||
const endIndex = lowerBuffer.indexOf(endTag)
|
||
if (endIndex === -1) {
|
||
const keepLength = Math.max(1, endTag.length - 1)
|
||
if (buffer.length > keepLength) {
|
||
emitThink(buffer.slice(0, buffer.length - keepLength))
|
||
buffer = buffer.slice(buffer.length - keepLength)
|
||
}
|
||
break
|
||
}
|
||
|
||
if (endIndex > 0) {
|
||
emitThink(buffer.slice(0, endIndex))
|
||
}
|
||
|
||
buffer = buffer.slice(endIndex + endTag.length)
|
||
mode = 'text'
|
||
handlers.onThinkEnd?.(currentThinkTag)
|
||
currentThinkTag = ''
|
||
continue
|
||
}
|
||
|
||
if (mode === 'tool_call') {
|
||
const lowerBuffer = buffer.toLowerCase()
|
||
const endIndex = lowerBuffer.indexOf(toolCallEndLower)
|
||
if (endIndex === -1) {
|
||
const keepLength = Math.max(1, TOOL_CALL_END_TAG.length - 1)
|
||
if (buffer.length > keepLength) {
|
||
buffer = buffer.slice(buffer.length - keepLength)
|
||
}
|
||
break
|
||
}
|
||
|
||
buffer = buffer.slice(endIndex + TOOL_CALL_END_TAG.length)
|
||
mode = 'text'
|
||
continue
|
||
}
|
||
}
|
||
}
|
||
|
||
return {
|
||
push(text: string) {
|
||
if (!text) return
|
||
buffer += text
|
||
processBuffer()
|
||
},
|
||
flush() {
|
||
if (!buffer) return
|
||
if (mode === 'text') {
|
||
emitText(buffer)
|
||
} else if (mode === 'think') {
|
||
emitThink(buffer)
|
||
}
|
||
buffer = ''
|
||
},
|
||
}
|
||
}
|
||
|
||
/**
|
||
* Agent 配置
|
||
*/
|
||
export interface AgentConfig {
|
||
/** 最大工具调用轮数（防止无限循环） */
|
||
maxToolRounds?: number
|
||
/** LLM 选项 */
|
||
llmOptions?: ChatOptions
|
||
/** 中止信号，用于取消执行 */
|
||
abortSignal?: AbortSignal
|
||
}
|
||
|
||
/**
|
||
* Token 使用量
|
||
*/
|
||
export interface TokenUsage {
|
||
promptTokens: number
|
||
completionTokens: number
|
||
totalTokens: number
|
||
}
|
||
|
||
/**
|
||
* Agent 流式响应 chunk
|
||
*/
|
||
export interface AgentStreamChunk {
|
||
/** chunk 类型 */
|
||
type: 'content' | 'think' | 'tool_start' | 'tool_result' | 'done' | 'error'
|
||
/** 文本内容（type=content 时） */
|
||
content?: string
|
||
/** 思考标签名称（type=think 时） */
|
||
thinkTag?: string
|
||
/** 思考耗时（毫秒，type=think 时可选） */
|
||
thinkDurationMs?: number
|
||
/** 工具名称（type=tool_start/tool_result 时） */
|
||
toolName?: string
|
||
/** 工具调用参数（type=tool_start 时） */
|
||
toolParams?: Record<string, unknown>
|
||
/** 工具执行结果（type=tool_result 时） */
|
||
toolResult?: unknown
|
||
/** 错误信息（type=error 时） */
|
||
error?: string
|
||
/** 是否完成 */
|
||
isFinished?: boolean
|
||
/** Token 使用量（type=done 时返回累计值） */
|
||
usage?: TokenUsage
|
||
}
|
||
|
||
/**
|
||
* Agent 执行结果
|
||
*/
|
||
export interface AgentResult {
|
||
/** 最终文本响应 */
|
||
content: string
|
||
/** 使用的工具列表 */
|
||
toolsUsed: string[]
|
||
/** 工具调用轮数 */
|
||
toolRounds: number
|
||
/** 总 Token 使用量（累计所有 LLM 调用） */
|
||
totalUsage?: TokenUsage
|
||
}
|
||
|
||
// ==================== 提示词配置类型 ====================
|
||
|
||
/**
|
||
* 用户自定义提示词配置
|
||
*/
|
||
export interface PromptConfig {
|
||
/** 角色定义（可编辑区） */
|
||
roleDefinition: string
|
||
/** 回答要求（可编辑区） */
|
||
responseRules: string
|
||
}
|
||
|
||
// 国际化内容
|
||
const i18nContent = {
|
||
'zh-CN': {
|
||
currentDateIs: '当前日期是',
|
||
chatTypeDesc: { private: '私聊记录', group: '群聊记录' },
|
||
chatContext: { private: '对话', group: '群聊' },
|
||
ownerNote: (displayName: string, platformId: string, chatContext: string) => `当前用户身份：
|
||
- 用户在${chatContext}中的身份是「${displayName}」（platformId: ${platformId}）
|
||
- 当用户提到"我"、"我的"时，指的就是「${displayName}」
|
||
- 查询"我"的发言时，使用 sender_id 参数筛选该成员
|
||
`,
|
||
memberNotePrivate: `成员查询策略：
|
||
- 私聊只有两个人，可以直接获取成员列表
|
||
- 当用户提到"对方"、"他/她"时，通过 get_group_members 获取另一方信息
|
||
`,
|
||
memberNoteGroup: `成员查询策略：
|
||
- 当用户提到特定群成员（如"张三说过什么"、"小明的发言"等）时，应先调用 get_group_members 获取成员列表
|
||
- 群成员有三种名称：accountName（原始昵称）、groupNickname（群昵称）、aliases（用户自定义别名）
|
||
- 通过 get_group_members 的 search 参数可以模糊搜索这三种名称
|
||
- 找到成员后，使用其 id 字段作为 search_messages 的 sender_id 参数来获取该成员的发言
|
||
`,
|
||
timeParamsIntro: '时间参数：按用户提到的精度组合 year/month/day/hour',
|
||
timeParamExample1: (year: number) => `"10月" → year: ${year}, month: 10`,
|
||
timeParamExample2: (year: number) => `"10月1号" → year: ${year}, month: 10, day: 1`,
|
||
timeParamExample3: (year: number) => `"10月1号下午3点" → year: ${year}, month: 10, day: 1, hour: 15`,
|
||
defaultYearNote: (year: number, prevYear: number) => `未指定年份默认${year}年，若该月份未到则用${prevYear}年`,
|
||
responseInstruction: '根据用户的问题，选择合适的工具获取数据，然后基于数据给出回答。',
|
||
responseRulesTitle: '回答要求：',
|
||
fallbackRoleDefinition: (chatType: string) => `你是一个专业的${chatType}记录分析助手。
|
||
你的任务是帮助用户理解和分析他们的${chatType}记录数据。`,
|
||
fallbackResponseRules: `1. 基于工具返回的数据回答，不要编造信息
|
||
2. 如果数据不足以回答问题，请说明
|
||
3. 回答要简洁明了，使用 Markdown 格式`,
|
||
},
|
||
'en-US': {
|
||
currentDateIs: 'Current date is',
|
||
chatTypeDesc: { private: 'private chat records', group: 'group chat records' },
|
||
chatContext: { private: 'conversation', group: 'group chat' },
|
||
ownerNote: (displayName: string, platformId: string, chatContext: string) => `Current user identity:
|
||
- The user's identity in this ${chatContext} is "${displayName}" (platformId: ${platformId})
|
||
- When the user refers to "I" or "my", it refers to "${displayName}"
|
||
- When querying "my" messages, use the sender_id parameter to filter for this member
|
||
`,
|
||
memberNotePrivate: `Member query strategy:
|
||
- Private chats only have two participants, so the member list can be directly obtained
|
||
- When the user refers to "the other party" or "he/she", get the other participant's information via get_group_members
|
||
`,
|
||
memberNoteGroup: `Member query strategy:
|
||
- When the user refers to specific group members (e.g., "what did John say", "Mary's messages"), first call get_group_members to get the member list
|
||
- Group members have three names: accountName (original nickname), groupNickname (group nickname), aliases (user-defined aliases)
|
||
- The search parameter of get_group_members can be used for fuzzy searching these three names
|
||
- Once a member is found, use their id field as the sender_id parameter for search_messages to retrieve their messages
|
||
`,
|
||
timeParamsIntro: 'Time parameters: combine year/month/day/hour based on user mention',
|
||
timeParamExample1: (year: number) => `"October" → year: ${year}, month: 10`,
|
||
timeParamExample2: (year: number) => `"October 1st" → year: ${year}, month: 10, day: 1`,
|
||
timeParamExample3: (year: number) => `"October 1st 3 PM" → year: ${year}, month: 10, day: 1, hour: 15`,
|
||
defaultYearNote: (year: number, prevYear: number) =>
|
||
`If year is not specified, defaults to ${year}. If the month has not yet occurred, ${prevYear} is used.`,
|
||
responseInstruction:
|
||
"Based on the user's question, select appropriate tools to retrieve data, then provide an answer based on the data.",
|
||
responseRulesTitle: 'Response requirements:',
|
||
fallbackRoleDefinition: (chatType: string) => `You are a professional ${chatType} analysis assistant.
|
||
Your task is to help users understand and analyze their ${chatType} data.`,
|
||
fallbackResponseRules: `1. Answer based on data returned by tools, do not fabricate information
|
||
2. If data is insufficient to answer, please state so
|
||
3. Keep answers concise and clear, use Markdown format`,
|
||
},
|
||
}
|
||
|
||
/**
|
||
* 获取系统锁定部分的提示词（策略说明、时间处理等）
|
||
*
|
||
* 注意：工具定义通过 Function Calling 的 tools 参数传递给 LLM，
|
||
* 无需在 System Prompt 中重复描述，以节省 Token。
|
||
*
|
||
* @param chatType 聊天类型 ('group' | 'private')
|
||
* @param ownerInfo Owner 信息（当前用户在对话中的身份）
|
||
* @param locale 语言设置
|
||
*/
|
||
function getLockedPromptSection(
|
||
chatType: 'group' | 'private',
|
||
ownerInfo?: OwnerInfo,
|
||
locale: string = 'zh-CN'
|
||
): string {
|
||
const content = i18nContent[locale as keyof typeof i18nContent] || i18nContent['zh-CN']
|
||
const now = new Date()
|
||
const dateLocale = locale === 'zh-CN' ? 'zh-CN' : 'en-US'
|
||
const currentDate = now.toLocaleDateString(dateLocale, {
|
||
year: 'numeric',
|
||
month: 'long',
|
||
day: 'numeric',
|
||
weekday: 'long',
|
||
})
|
||
|
||
const isPrivate = chatType === 'private'
|
||
const chatContext = content.chatContext[chatType]
|
||
|
||
// Owner 说明（当用户设置了"我是谁"时）
|
||
const ownerNote = ownerInfo ? content.ownerNote(ownerInfo.displayName, ownerInfo.platformId, chatContext) : ''
|
||
|
||
// 成员说明（私聊只有2人）
|
||
const memberNote = isPrivate ? content.memberNotePrivate : content.memberNoteGroup
|
||
|
||
const year = now.getFullYear()
|
||
const prevYear = year - 1
|
||
|
||
return `${content.currentDateIs} ${currentDate}。
|
||
${ownerNote}
|
||
${memberNote}
|
||
${content.timeParamsIntro}
|
||
- ${content.timeParamExample1(year)}
|
||
- ${content.timeParamExample2(year)}
|
||
- ${content.timeParamExample3(year)}
|
||
${content.defaultYearNote(year, prevYear)}
|
||
|
||
${content.responseInstruction}`
|
||
}
|
||
|
||
/**
|
||
* 获取 Fallback 角色定义（主要配置来自前端 src/config/prompts.ts）
|
||
* 仅在前端未传递 promptConfig 时使用
|
||
*/
|
||
function getFallbackRoleDefinition(chatType: 'group' | 'private', locale: string = 'zh-CN'): string {
|
||
const content = i18nContent[locale as keyof typeof i18nContent] || i18nContent['zh-CN']
|
||
const chatTypeDesc = chatType === 'private' ? (locale === 'zh-CN' ? '私聊' : 'private chat') : (locale === 'zh-CN' ? '群聊' : 'group chat')
|
||
return content.fallbackRoleDefinition(chatTypeDesc)
|
||
}
|
||
|
||
/**
|
||
* 获取 Fallback 回答要求（主要配置来自前端 src/config/prompts.ts）
|
||
* 仅在前端未传递 promptConfig 时使用
|
||
*/
|
||
function getFallbackResponseRules(locale: string = 'zh-CN'): string {
|
||
const content = i18nContent[locale as keyof typeof i18nContent] || i18nContent['zh-CN']
|
||
return content.fallbackResponseRules
|
||
}
|
||
|
||
/**
|
||
* 构建完整的系统提示词
|
||
*
|
||
* 提示词配置主要来自前端 src/config/prompts.ts，通过 promptConfig 参数传递。
|
||
* Fallback 仅在前端未传递配置时使用（一般不会发生）。
|
||
*
|
||
* @param chatType 聊天类型 ('group' | 'private')
|
||
* @param promptConfig 用户自定义提示词配置（来自前端激活的预设）
|
||
* @param ownerInfo Owner 信息（当前用户在对话中的身份）
|
||
* @param locale 语言设置
|
||
*/
|
||
function buildSystemPrompt(
|
||
chatType: 'group' | 'private' = 'group',
|
||
promptConfig?: PromptConfig,
|
||
ownerInfo?: OwnerInfo,
|
||
locale: string = 'zh-CN'
|
||
): string {
|
||
const content = i18nContent[locale as keyof typeof i18nContent] || i18nContent['zh-CN']
|
||
|
||
// 使用用户配置或 fallback
|
||
const roleDefinition = promptConfig?.roleDefinition || getFallbackRoleDefinition(chatType, locale)
|
||
const responseRules = promptConfig?.responseRules || getFallbackResponseRules(locale)
|
||
|
||
// 获取锁定的系统部分（包含动态日期、工具说明和 Owner 信息）
|
||
const lockedSection = getLockedPromptSection(chatType, ownerInfo, locale)
|
||
|
||
// 组合完整提示词
|
||
return `${roleDefinition}
|
||
|
||
${lockedSection}
|
||
|
||
${content.responseRulesTitle}
|
||
${responseRules}`
|
||
}
|
||
|
||
/**
|
||
* Agent 执行器类
|
||
* 处理带 Function Calling 的对话流程
|
||
*/
|
||
export class Agent {
|
||
private context: ToolContext
|
||
private config: AgentConfig
|
||
private messages: ChatMessage[] = []
|
||
private toolsUsed: string[] = []
|
||
private toolRounds: number = 0
|
||
private abortSignal?: AbortSignal
|
||
private historyMessages: ChatMessage[] = []
|
||
private chatType: 'group' | 'private' = 'group'
|
||
private promptConfig?: PromptConfig
|
||
private locale: string = 'zh-CN'
|
||
/** 累计 Token 使用量 */
|
||
private totalUsage: TokenUsage = { promptTokens: 0, completionTokens: 0, totalTokens: 0 }
|
||
|
||
constructor(
|
||
context: ToolContext,
|
||
config: AgentConfig = {},
|
||
historyMessages: ChatMessage[] = [],
|
||
chatType: 'group' | 'private' = 'group',
|
||
promptConfig?: PromptConfig,
|
||
locale: string = 'zh-CN'
|
||
) {
|
||
this.context = context
|
||
this.abortSignal = config.abortSignal
|
||
this.historyMessages = historyMessages
|
||
this.chatType = chatType
|
||
this.promptConfig = promptConfig
|
||
this.locale = locale
|
||
this.config = {
|
||
maxToolRounds: config.maxToolRounds ?? 5,
|
||
llmOptions: config.llmOptions ?? { temperature: 0.7, maxTokens: 2048 },
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 检查是否已中止
|
||
*/
|
||
private isAborted(): boolean {
|
||
return this.abortSignal?.aborted ?? false
|
||
}
|
||
|
||
/**
|
||
* 累加 Token 使用量
|
||
*/
|
||
private addUsage(usage?: { promptTokens: number; completionTokens: number; totalTokens: number }): void {
|
||
if (usage) {
|
||
this.totalUsage.promptTokens += usage.promptTokens
|
||
this.totalUsage.completionTokens += usage.completionTokens
|
||
this.totalUsage.totalTokens += usage.totalTokens
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 执行对话（非流式）
|
||
* @param userMessage 用户消息
|
||
*/
|
||
async execute(userMessage: string): Promise<AgentResult> {
|
||
aiLogger.info('Agent', '用户问题', userMessage)
|
||
|
||
// 检查是否已中止
|
||
if (this.isAborted()) {
|
||
return { content: '', toolsUsed: [], toolRounds: 0, totalUsage: this.totalUsage }
|
||
}
|
||
|
||
// 初始化消息（包含历史记录）
|
||
const systemPrompt = buildSystemPrompt(this.chatType, this.promptConfig, this.context.ownerInfo, this.locale)
|
||
this.messages = [
|
||
{ role: 'system', content: systemPrompt },
|
||
...this.historyMessages, // 插入历史对话
|
||
{ role: 'user', content: userMessage },
|
||
]
|
||
this.toolsUsed = []
|
||
this.toolRounds = 0
|
||
|
||
// 获取所有工具定义
|
||
const tools = await getAllToolDefinitions()
|
||
|
||
// 执行循环
|
||
while (this.toolRounds < this.config.maxToolRounds!) {
|
||
// 每轮开始时检查是否中止
|
||
if (this.isAborted()) {
|
||
return {
|
||
content: '',
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
|
||
const response = await chat(this.messages, {
|
||
...this.config.llmOptions,
|
||
tools,
|
||
abortSignal: this.abortSignal,
|
||
})
|
||
|
||
// 累加 Token 使用量
|
||
this.addUsage(response.usage)
|
||
|
||
const { cleanContent } = extractThinkingContent(response.content)
|
||
let toolCallsToProcess = response.tool_calls
|
||
|
||
// 如果没有标准 tool_calls，尝试 fallback 解析
|
||
if (response.finishReason !== 'tool_calls' || !response.tool_calls) {
|
||
// Fallback: 检查内容中是否有 <tool_call> 标签
|
||
if (hasToolCallTags(response.content)) {
|
||
// 解析 tool_call 标签
|
||
const fallbackToolCalls = parseToolCallTags(response.content)
|
||
if (fallbackToolCalls && fallbackToolCalls.length > 0) {
|
||
toolCallsToProcess = fallbackToolCalls
|
||
} else {
|
||
// 解析失败，返回清理后的内容
|
||
const sanitizedContent = stripToolCallTags(cleanContent)
|
||
aiLogger.info('Agent', 'AI 回复', sanitizedContent)
|
||
return {
|
||
content: sanitizedContent,
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
} else {
|
||
// 没有 tool_call 标签，正常完成
|
||
aiLogger.info('Agent', 'AI 回复', cleanContent)
|
||
return {
|
||
content: cleanContent,
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
}
|
||
|
||
// 处理工具调用
|
||
await this.handleToolCalls(toolCallsToProcess!)
|
||
this.toolRounds++
|
||
}
|
||
|
||
// 超过最大轮数，强制让 LLM 总结
|
||
aiLogger.warn('Agent', '达到最大工具调用轮数', { maxRounds: this.config.maxToolRounds })
|
||
this.messages.push({
|
||
role: 'user',
|
||
content: '请根据已获取的信息给出回答，不要再调用工具。',
|
||
})
|
||
|
||
const finalResponse = await chat(this.messages, this.config.llmOptions)
|
||
this.addUsage(finalResponse.usage)
|
||
const finalCleanContent = stripToolCallTags(extractThinkingContent(finalResponse.content).cleanContent)
|
||
return {
|
||
content: finalCleanContent,
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 执行对话（流式）
|
||
* @param userMessage 用户消息
|
||
* @param onChunk 流式回调
|
||
*/
|
||
async executeStream(userMessage: string, onChunk: (chunk: AgentStreamChunk) => void): Promise<AgentResult> {
|
||
aiLogger.info('Agent', '用户问题', userMessage)
|
||
|
||
// 检查是否已中止
|
||
if (this.isAborted()) {
|
||
onChunk({ type: 'done', isFinished: true, usage: this.totalUsage })
|
||
return { content: '', toolsUsed: [], toolRounds: 0, totalUsage: this.totalUsage }
|
||
}
|
||
|
||
// 初始化消息（包含历史记录）
|
||
const systemPrompt = buildSystemPrompt(this.chatType, this.promptConfig, this.context.ownerInfo, this.locale)
|
||
this.messages = [
|
||
{ role: 'system', content: systemPrompt },
|
||
...this.historyMessages, // 插入历史对话
|
||
{ role: 'user', content: userMessage },
|
||
]
|
||
this.toolsUsed = []
|
||
this.toolRounds = 0
|
||
|
||
const tools = await getAllToolDefinitions()
|
||
let finalContent = ''
|
||
|
||
// 执行循环
|
||
while (this.toolRounds < this.config.maxToolRounds!) {
|
||
// 每轮开始时检查是否中止
|
||
if (this.isAborted()) {
|
||
onChunk({ type: 'done', isFinished: true, usage: this.totalUsage })
|
||
return {
|
||
content: finalContent,
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
|
||
let accumulatedContent = ''
|
||
let roundContent = ''
|
||
let toolCalls: ToolCall[] | undefined
|
||
let thinkStartAt: number | null = null // 记录思考开始时间
|
||
const parser = createStreamParser({
|
||
onText: (text) => {
|
||
roundContent += text
|
||
onChunk({ type: 'content', content: text })
|
||
},
|
||
onThinkStart: () => {
|
||
thinkStartAt = Date.now()
|
||
},
|
||
onThink: (text, tag) => {
|
||
onChunk({ type: 'think', content: text, thinkTag: tag })
|
||
},
|
||
onThinkEnd: (tag) => {
|
||
if (thinkStartAt === null) return
|
||
const durationMs = Date.now() - thinkStartAt
|
||
thinkStartAt = null
|
||
onChunk({ type: 'think', content: '', thinkTag: tag, thinkDurationMs: durationMs })
|
||
},
|
||
})
|
||
|
||
// 流式调用 LLM（传入 abortSignal）
|
||
for await (const chunk of chatStream(this.messages, {
|
||
...this.config.llmOptions,
|
||
tools,
|
||
abortSignal: this.abortSignal,
|
||
})) {
|
||
// 每个 chunk 时检查是否中止
|
||
if (this.isAborted()) {
|
||
onChunk({ type: 'done', isFinished: true, usage: this.totalUsage })
|
||
return {
|
||
content: finalContent + roundContent,
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
if (chunk.content) {
|
||
accumulatedContent += chunk.content
|
||
// 按标签切分后输出到内容/思考区
|
||
parser.push(chunk.content)
|
||
}
|
||
|
||
if (chunk.tool_calls) {
|
||
toolCalls = chunk.tool_calls
|
||
}
|
||
|
||
// 累加 Token 使用量（流式响应在最后一个 chunk 返回 usage）
|
||
if (chunk.usage) {
|
||
this.addUsage(chunk.usage)
|
||
}
|
||
|
||
if (chunk.isFinished) {
|
||
// 收尾：清空解析器缓冲
|
||
parser.flush()
|
||
|
||
// 如果没有标准 tool_calls，尝试 fallback 解析
|
||
if (chunk.finishReason !== 'tool_calls' || !toolCalls) {
|
||
// Fallback: 检查内容中是否有 <tool_call> 标签
|
||
if (hasToolCallTags(accumulatedContent)) {
|
||
// 提取 thinking 内容
|
||
const { cleanContent } = extractThinkingContent(accumulatedContent)
|
||
|
||
// 解析 tool_call 标签
|
||
const fallbackToolCalls = parseToolCallTags(accumulatedContent)
|
||
if (fallbackToolCalls && fallbackToolCalls.length > 0) {
|
||
toolCalls = fallbackToolCalls
|
||
// 不返回，继续执行工具调用
|
||
} else {
|
||
// 解析失败，作为普通响应处理
|
||
const sanitizedContent = stripToolCallTags(cleanContent)
|
||
if (sanitizedContent.startsWith(roundContent)) {
|
||
const remainingContent = sanitizedContent.slice(roundContent.length)
|
||
if (remainingContent) {
|
||
onChunk({ type: 'content', content: remainingContent })
|
||
}
|
||
} else if (sanitizedContent) {
|
||
aiLogger.warn('Agent', '流式内容与清理结果不一致，跳过补发', {
|
||
roundContentLength: roundContent.length,
|
||
sanitizedLength: sanitizedContent.length,
|
||
})
|
||
}
|
||
finalContent = sanitizedContent
|
||
aiLogger.info('Agent', 'AI 回复', finalContent)
|
||
onChunk({ type: 'done', isFinished: true, usage: this.totalUsage })
|
||
return {
|
||
content: finalContent,
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
} else {
|
||
// 没有 tool_call 标签，正常完成
|
||
const sanitizedContent = stripToolCallTags(extractThinkingContent(accumulatedContent).cleanContent)
|
||
if (sanitizedContent.startsWith(roundContent)) {
|
||
const remainingContent = sanitizedContent.slice(roundContent.length)
|
||
if (remainingContent) {
|
||
onChunk({ type: 'content', content: remainingContent })
|
||
}
|
||
} else if (sanitizedContent) {
|
||
aiLogger.warn('Agent', '流式内容与清理结果不一致，跳过补发', {
|
||
roundContentLength: roundContent.length,
|
||
sanitizedLength: sanitizedContent.length,
|
||
})
|
||
}
|
||
finalContent = sanitizedContent
|
||
aiLogger.info('Agent', 'AI 回复', finalContent)
|
||
onChunk({ type: 'done', isFinished: true, usage: this.totalUsage })
|
||
return {
|
||
content: finalContent,
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
// 兜底收尾：防止未收到 isFinished
|
||
parser.flush()
|
||
|
||
// 处理工具调用
|
||
if (toolCalls && toolCalls.length > 0) {
|
||
// 通知前端开始执行工具（包含参数和时间范围）
|
||
for (const tc of toolCalls) {
|
||
let toolParams: Record<string, unknown> | undefined
|
||
try {
|
||
toolParams = JSON.parse(tc.function.arguments || '{}')
|
||
|
||
// 对于消息获取类工具，用用户配置的 limit 覆盖 LLM 传递的值（保持显示一致）
|
||
const toolsWithLimit = ['search_messages', 'get_recent_messages', 'get_conversation_between']
|
||
if (this.context.maxMessagesLimit && toolsWithLimit.includes(tc.function.name)) {
|
||
toolParams = {
|
||
...toolParams,
|
||
limit: this.context.maxMessagesLimit, // 用户配置优先
|
||
}
|
||
}
|
||
|
||
// 对于搜索类工具，添加时间范围信息
|
||
if (
|
||
this.context.timeFilter &&
|
||
(tc.function.name === 'search_messages' || tc.function.name === 'get_recent_messages')
|
||
) {
|
||
toolParams = {
|
||
...toolParams,
|
||
_timeFilter: this.context.timeFilter,
|
||
}
|
||
}
|
||
} catch {
|
||
toolParams = undefined
|
||
}
|
||
onChunk({ type: 'tool_start', toolName: tc.function.name, toolParams })
|
||
}
|
||
|
||
await this.handleToolCalls(toolCalls, onChunk)
|
||
this.toolRounds++
|
||
}
|
||
}
|
||
|
||
// 超过最大轮数
|
||
aiLogger.warn('Agent', '达到最大工具调用轮数', { maxRounds: this.config.maxToolRounds })
|
||
|
||
// 检查是否已中止
|
||
if (this.isAborted()) {
|
||
onChunk({ type: 'done', isFinished: true, usage: this.totalUsage })
|
||
return {
|
||
content: finalContent,
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
|
||
this.messages.push({
|
||
role: 'user',
|
||
content: '请根据已获取的信息给出回答，不要再调用工具。',
|
||
})
|
||
|
||
// 最后一轮不带 tools（传入 abortSignal）
|
||
let finalRawContent = ''
|
||
let finalThinkStartAt: number | null = null // 记录最终思考开始时间
|
||
const finalParser = createStreamParser({
|
||
onText: (text) => {
|
||
finalContent += text
|
||
onChunk({ type: 'content', content: text })
|
||
},
|
||
onThinkStart: () => {
|
||
finalThinkStartAt = Date.now()
|
||
},
|
||
onThink: (text, tag) => {
|
||
onChunk({ type: 'think', content: text, thinkTag: tag })
|
||
},
|
||
onThinkEnd: (tag) => {
|
||
if (finalThinkStartAt === null) return
|
||
const durationMs = Date.now() - finalThinkStartAt
|
||
finalThinkStartAt = null
|
||
onChunk({ type: 'think', content: '', thinkTag: tag, thinkDurationMs: durationMs })
|
||
},
|
||
})
|
||
for await (const chunk of chatStream(this.messages, {
|
||
...this.config.llmOptions,
|
||
abortSignal: this.abortSignal,
|
||
})) {
|
||
if (this.isAborted()) {
|
||
onChunk({ type: 'done', isFinished: true, usage: this.totalUsage })
|
||
break
|
||
}
|
||
if (chunk.content) {
|
||
finalRawContent += chunk.content
|
||
finalParser.push(chunk.content)
|
||
}
|
||
// 累加 Token 使用量
|
||
if (chunk.usage) {
|
||
this.addUsage(chunk.usage)
|
||
}
|
||
if (chunk.isFinished) {
|
||
finalParser.flush()
|
||
const sanitizedContent = stripToolCallTags(extractThinkingContent(finalRawContent).cleanContent)
|
||
if (sanitizedContent.startsWith(finalContent)) {
|
||
const remainingContent = sanitizedContent.slice(finalContent.length)
|
||
if (remainingContent) {
|
||
finalContent += remainingContent
|
||
onChunk({ type: 'content', content: remainingContent })
|
||
}
|
||
} else if (sanitizedContent) {
|
||
aiLogger.warn('Agent', '最终内容与清理结果不一致，跳过补发', {
|
||
finalContentLength: finalContent.length,
|
||
sanitizedLength: sanitizedContent.length,
|
||
})
|
||
finalContent = sanitizedContent
|
||
}
|
||
onChunk({ type: 'done', isFinished: true, usage: this.totalUsage })
|
||
}
|
||
}
|
||
|
||
// 兜底收尾：防止未收到 isFinished
|
||
finalParser.flush()
|
||
|
||
return {
|
||
content: finalContent,
|
||
toolsUsed: this.toolsUsed,
|
||
toolRounds: this.toolRounds,
|
||
totalUsage: this.totalUsage,
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 处理工具调用
|
||
*/
|
||
private async handleToolCalls(toolCalls: ToolCall[], onChunk?: (chunk: AgentStreamChunk) => void): Promise<void> {
|
||
// 记录调用的工具及参数
|
||
for (const tc of toolCalls) {
|
||
aiLogger.info('Agent', `工具调用: ${tc.function.name}`, tc.function.arguments)
|
||
}
|
||
|
||
// 添加 assistant 消息（包含 tool_calls）
|
||
this.messages.push({
|
||
role: 'assistant',
|
||
content: '',
|
||
tool_calls: toolCalls,
|
||
})
|
||
|
||
// 执行工具（传递 locale 用于工具返回结果的国际化）
|
||
const results = await executeToolCalls(toolCalls, { ...this.context, locale: this.locale })
|
||
|
||
// 添加 tool 消息
|
||
for (let i = 0; i < toolCalls.length; i++) {
|
||
const tc = toolCalls[i]
|
||
const result = results[i]
|
||
|
||
this.toolsUsed.push(tc.function.name)
|
||
|
||
// 通知前端工具执行结果
|
||
if (onChunk) {
|
||
onChunk({
|
||
type: 'tool_result',
|
||
toolName: tc.function.name,
|
||
toolResult: result.success ? result.result : result.error,
|
||
})
|
||
}
|
||
|
||
// 记录工具执行结果
|
||
if (result.success) {
|
||
aiLogger.info('Agent', `工具结果: ${tc.function.name}`, result.result)
|
||
} else {
|
||
aiLogger.warn('Agent', `工具失败: ${tc.function.name}`, result.error)
|
||
}
|
||
|
||
// 添加工具结果消息
|
||
this.messages.push({
|
||
role: 'tool',
|
||
content: result.success ? JSON.stringify(result.result) : `错误: ${result.error}`,
|
||
tool_call_id: tc.id,
|
||
})
|
||
}
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 创建 Agent 并执行对话（便捷函数）
|
||
*/
|
||
export async function runAgent(
|
||
userMessage: string,
|
||
context: ToolContext,
|
||
config?: AgentConfig,
|
||
historyMessages?: ChatMessage[],
|
||
chatType?: 'group' | 'private'
|
||
): Promise<AgentResult> {
|
||
const agent = new Agent(context, config, historyMessages, chatType)
|
||
return agent.execute(userMessage)
|
||
}
|
||
|
||
/**
|
||
* 创建 Agent 并流式执行对话（便捷函数）
|
||
*/
|
||
export async function runAgentStream(
|
||
userMessage: string,
|
||
context: ToolContext,
|
||
onChunk: (chunk: AgentStreamChunk) => void,
|
||
config?: AgentConfig,
|
||
historyMessages?: ChatMessage[],
|
||
chatType?: 'group' | 'private'
|
||
): Promise<AgentResult> {
|
||
const agent = new Agent(context, config, historyMessages, chatType)
|
||
return agent.executeStream(userMessage, onChunk)
|
||
}
|