feat: 完善群聊日报模板与图片理解

src/main/services/ai-provider-service.ts

@@ -69,7 +69,8 @@ export class AIProviderService {
       configured: Boolean(
         provider && provider.models.length && (provider.hasApiKey || !needsApiKey(provider))
       ),
-      status: provider?.status || 'untested'
+      status: provider?.status || 'untested',
+      timeoutMs: provider?.advanced.timeoutMs
     }
   }
 
@@ -169,6 +170,37 @@ export class AIProviderService {
     }
   }
 
+  /**
+   * 多模态图片理解。
+   * 输入:text + image parts 的 messages,返回 AI 文本响应。
+   * 与 testVision 区别:不校验 prompt,不写入 capability marker(供 ImageInsightService 复用)。
+   */
+  async analyzeImage(
+    messages: Array<{
+      role: string
+      content: string | Array<{ type: 'text'; text: string } | { type: 'image'; dataUrl: string }>
+    }>,
+    options?: AIChatRequestOptions
+  ): Promise<{
+    success: boolean
+    data?: string
+    usage?: { input?: number; output?: number; total?: number; estimated?: boolean }
+    error?: string
+  }> {
+    try {
+      const imagePart = messages
+        .flatMap((message) => (typeof message.content === 'string' ? [] : message.content))
+        .find((part) => part.type === 'image')
+      if (!imagePart || imagePart.type !== 'image') throw new Error('图片识别请求缺少图片数据')
+      const imageError = validateVisionImage(imagePart.dataUrl)
+      if (imageError) throw new Error(imageError)
+      const result = await this.request(messages as AIMessage[], options)
+      return { success: true, ...result }
+    } catch (error) {
+      return { success: false, error: safeAIError(error) }
+    }
+  }
+
   async testVision(request: AIVisionTestRequest): Promise<AIVisionTestResult> {
     const startedAt = Date.now()
     const imageError = validateVisionImage(request.imageDataUrl)
@@ -215,7 +247,13 @@ export class AIProviderService {
   }> {
     if (options?.apiKey) return this.requestLegacy(messages, options)
     const resolved = this.resolveProvider(options)
-    return requestProvider(resolved.provider, resolved.key, resolved.model, messages, testing)
+    const provider = options?.timeoutMs
+      ? {
+          ...resolved.provider,
+          advanced: { ...resolved.provider.advanced, timeoutMs: options.timeoutMs }
+        }
+      : resolved.provider
+    return requestProvider(provider, resolved.key, resolved.model, messages, testing)
   }
 
   private resolveProvider(options?: { providerId?: string; modelId?: string }): {
@@ -263,12 +301,38 @@ export class AIProviderService {
   }
 
   private markVisionCapability(providerId: string, modelId: string): void {
+    this.markCapabilities(providerId, modelId, { vision: true, ocr: true })
+  }
+
+  /**
+   * 标记模型已验证的 capabilities(已存在则跳过)。
+   * OCR 跟随 vision:几乎所有 vision 模型都能 OCR,标记 vision 时同步标记 ocr。
+   */
+  private markCapabilities(
+    providerId: string,
+    modelId: string,
+    caps: { vision?: boolean; ocr?: boolean }
+  ): void {
     const data = this.readMetadata()
     const provider = data.providers.find((item) => item.id === providerId)
     const model = provider?.models.find((item) => item.id === modelId)
-    if (!provider || !model || model.capabilities.vision) return
-    model.capabilities.vision = true
-    this.writeMetadata(data)
+    if (!provider || !model) return
+    // 老配置可能没有 ocr 字段,补默认 false
+    if (typeof model.capabilities.ocr !== 'boolean') model.capabilities.ocr = false
+    let changed = false
+    if (caps.vision === true && !model.capabilities.vision) {
+      model.capabilities.vision = true
+      // vision 开启默认带 ocr(派生能力)
+      if (!model.capabilities.ocr) {
+        model.capabilities.ocr = true
+      }
+      changed = true
+    }
+    if (caps.ocr === true && !model.capabilities.ocr) {
+      model.capabilities.ocr = true
+      changed = true
+    }
+    if (changed) this.writeMetadata(data)
   }
 
   private ensureEnvironmentMigration(): void {
@@ -300,6 +364,14 @@ export class AIProviderService {
     const data = fs.readJsonSync(filePath) as AIProviderMetadataFile
     if (data.version !== 1 || !Array.isArray(data.providers))
       throw new Error('invalid provider metadata')
+    // 老配置兼容:补 capabilities.ocr 默认值(vision 派生 OCR)
+    for (const provider of data.providers) {
+      for (const model of provider.models) {
+        if (typeof model.capabilities.ocr !== 'boolean') {
+          model.capabilities.ocr = model.capabilities.vision === true
+        }
+      }
+    }
     return data
   }
 
@@ -325,7 +397,7 @@ function deepSeekProvider(baseUrl?: string, model?: string): AIProviderSummary {
       {
         name: modelId === 'deepseek-chat' ? 'DeepSeek Chat' : modelId,
         id: modelId,
-        capabilities: { chat: true, vision: false, longContext: true }
+        capabilities: { chat: true, vision: false, ocr: false, longContext: true }
       }
     ],
     defaultModel: modelId,
@@ -454,7 +526,7 @@ async function requestOpenAICompatible(
     },
     provider.advanced.timeoutMs
   )
-  const payload = (await response.json()) as OpenAIResponsePayload
+  const payload = await parseJsonResponse<OpenAIResponsePayload>(response)
   if (!response.ok) throw new Error(payload.error?.message || `AI 请求失败 (${response.status})`)
   return {
     data: String(payload.choices?.[0]?.message?.content || ''),
@@ -508,7 +580,7 @@ async function requestAnthropic(
     },
     provider.advanced.timeoutMs
   )
-  const payload = (await response.json()) as AnthropicResponsePayload
+  const payload = await parseJsonResponse<AnthropicResponsePayload>(response)
   if (!response.ok)
     throw new Error(payload.error?.message || `Anthropic 请求失败 (${response.status})`)
   return {
@@ -543,6 +615,20 @@ async function fetchWithTimeout(
   }
 }
 
+async function parseJsonResponse<T>(response: Response): Promise<T> {
+  const body = await response.text()
+  try {
+    return JSON.parse(body) as T
+  } catch {
+    const looksLikeHtml = /^\s*(?:<!doctype\s+html|<html\b)/i.test(body)
+    const status = `${response.status}${response.statusText ? ` ${response.statusText}` : ''}`
+    if (looksLikeHtml) {
+      throw new Error(`模型服务返回了网页而不是 JSON（HTTP ${status}），请稍后重试或检查中转服务`)
+    }
+    throw new Error(`模型服务返回格式异常（HTTP ${status}）`)
+  }
+}
+
 function safeAIError(error: unknown): string {
   if (error instanceof DOMException && error.name === 'AbortError') return 'AI 请求超时'
   const message = error instanceof Error ? error.message : String(error)

src/main/services/chat-service.ts

@@ -56,7 +56,14 @@ export interface FormattedMessage {
 export interface GroupSnapshot {
   roomId: string
   memberCount: number
-  members: { wxid: string; nickname: string; avatar: string }[]
+  members: {
+    wxid: string
+    nickname: string
+    groupNickname: string
+    wechatNickname: string
+    remark: string
+    avatar: string
+  }[]
 }
 
 const MSG_TYPE_DICT: Record<number, string> = {
@@ -292,6 +299,9 @@ export function getGroupSnapshot(userMd5: string): GroupSnapshot | null {
     .map((member) => ({
       wxid: member.m_nsUsrName,
       nickname: member.nickname || '',
+      groupNickname: member.groupNickname || '',
+      wechatNickname: member.wechatNickname || '',
+      remark: member.remark || '',
       avatar: member.m_nsHeadImgUrl || ''
     }))
 

src/main/services/image-insight-prompt.ts

@@ -0,0 +1,94 @@
+// src/main/services/image-insight-prompt.ts
+// 图片理解 prompt 模板 — 输出严格的 JSON,便于程序化解析
+
+export const IMAGE_ANALYSIS_SYSTEM_PROMPT = `你是微信群聊的图片分析助手。
+请根据用户提供的图片和图片前后的聊天上下文,生成对该图片的结构化理解。
+
+输出要求(严格遵守):
+1. 必须输出 JSON,不要用 markdown 代码块包裹
+2. description:1-2 句中文,30-80 字,描述图片核心内容
+3. ocrText:如果图片含文字(截图、文档、票据等),提取出来;纯风景/表情包可填空字符串
+4. tags:3-6 个中文关键词标签
+5. category:screenshot / photo / meme / document / chart / other 之一
+6. importance:low / medium / high — 根据图片的信息密度和后续讨论热度判断
+
+禁止:
+- 不要猜测图片中未明确可见的内容
+- 不要复述聊天上下文本身(那是 description 之外的事)
+- 不要输出 markdown 标记`
+
+export interface ImageAnalysisContext {
+  sender: string
+  sentAt: number
+  contextBefore: string[] // 图片前 1-3 条消息
+  contextAfter: string[] // 图片后 1-3 条消息
+}
+
+export function buildImageAnalysisUserText(ctx: ImageAnalysisContext): string {
+  const before = ctx.contextBefore.length
+    ? ctx.contextBefore.map((m, i) => `  ${i + 1}. ${m}`).join('\n')
+    : '  (无前文)'
+  const after = ctx.contextAfter.length
+    ? ctx.contextAfter.map((m, i) => `  ${i + 1}. ${m}`).join('\n')
+    : '  (无后续讨论)'
+  const time = new Date(ctx.sentAt * 1000).toLocaleString('zh-CN', { hour12: false })
+
+  return `发送者:${ctx.sender}
+时间:${time}
+
+图片前的聊天:
+${before}
+
+图片后的聊天:
+${after}
+
+请输出 JSON(严格遵守 system 要求):
+{"description":"...","ocrText":"...","tags":["..."],"category":"...","importance":"..."}`
+}
+
+/**
+ * 把 AI 文本响应解析成结构化字段。
+ * 容忍:无 markdown 包裹、有 markdown 包裹、尾部有杂质等。
+ */
+export function parseImageAnalysisResponse(raw: string): {
+  description: string
+  ocrText: string
+  tags: string[]
+  category: 'screenshot' | 'photo' | 'meme' | 'document' | 'chart' | 'other'
+  importance: 'low' | 'medium' | 'high'
+} {
+  const text = raw.trim()
+  // 提取 JSON 段
+  const jsonMatch = text.match(/\{[\s\S]*\}/)
+  if (!jsonMatch) {
+    throw new Error('AI 未返回合法 JSON')
+  }
+  let parsed: Record<string, unknown>
+  try {
+    parsed = JSON.parse(jsonMatch[0])
+  } catch {
+    throw new Error('AI 返回的 JSON 无法解析')
+  }
+
+  const description = String(parsed.description || '').trim()
+  if (!description) throw new Error('AI 未返回 description')
+
+  const ocrText = String(parsed.ocrText || '').trim()
+  const tagsRaw = parsed.tags
+  const tags = Array.isArray(tagsRaw)
+    ? tagsRaw.map((t) => String(t).trim()).filter(Boolean).slice(0, 8)
+    : []
+
+  const categoryRaw = String(parsed.category || 'other').toLowerCase()
+  const category: 'screenshot' | 'photo' | 'meme' | 'document' | 'chart' | 'other' =
+    ['screenshot', 'photo', 'meme', 'document', 'chart'].includes(categoryRaw)
+      ? (categoryRaw as 'screenshot' | 'photo' | 'meme' | 'document' | 'chart')
+      : 'other'
+
+  const importanceRaw = String(parsed.importance || 'medium').toLowerCase()
+  const importance: 'low' | 'medium' | 'high' = ['low', 'medium', 'high'].includes(importanceRaw)
+    ? (importanceRaw as 'low' | 'medium' | 'high')
+    : 'medium'
+
+  return { description, ocrText, tags, category, importance }
+}

src/main/services/image-insight-service.ts

@@ -0,0 +1,279 @@
+// src/main/services/image-insight-service.ts
+// WechatExplorer AI 图片理解基础设施
+//
+// 设计原则:
+// 1. base64 不走 IPC,只在 main 内部流转(renderer 只看到 ImageInsight 结构化结果)
+// 2. 同图(imageHash)走缓存,绝不重复调 AI
+// 3. 失败不抛,日志记录 + 返回原状(不阻塞日报)
+// 4. 第一阶段:Top 3 热点图 + 缓存命中即返回,未命中并发调 AI
+
+import crypto from 'crypto'
+import { randomUUID } from 'crypto'
+import { imageInsightsStore } from '../db/image-insights-store'
+import {
+  buildImageAnalysisUserText,
+  IMAGE_ANALYSIS_SYSTEM_PROMPT,
+  parseImageAnalysisResponse
+} from './image-insight-prompt'
+import type {
+  ImageAnalysisRequest,
+  ImageAnalysisResponse,
+  ImageCandidate,
+  ImageCandidateQuery,
+  ImageInsight
+} from '../../shared/image-insight'
+
+/**
+ * 单张图片的最小信息(由 renderer 从已加载的 messages 中提取并传入 main)。
+ * 这样可以避免 ImageInsightService 自己重新查询消息,且参数语义清晰。
+ */
+export interface ImageCandidateInput {
+  messageId: string
+  md5?: string
+  datName?: string
+  sessionId: string
+  sender: string
+  sentAt: number
+  /** 图片发出后 8 条消息内、不同发言人的回复数(由 renderer 计算) */
+  responseCount: number
+  /** 表情/语音互动条数 */
+  interactionCount: number
+}
+
+interface ProviderServiceLike {
+  list(): ProviderSummaryLike
+  analyzeImage(
+    messages: Array<{
+      role: string
+      content: string | Array<{ type: 'text'; text: string } | { type: 'image'; dataUrl: string }>
+    }>,
+    options?: { providerId?: string; modelId?: string }
+  ): Promise<{
+    success: boolean
+    data?: string
+    error?: string
+  }>
+}
+
+interface DecryptServiceLike {
+  findImageFile(
+    md5?: string,
+    imageDatName?: string,
+    options?: { allowThumbnail?: boolean }
+  ): string | null
+  decryptImageToBase64(datPath: string): string | null
+}
+
+interface ProviderSummaryLike {
+  providers: Array<{
+    id: string
+    isDefault: boolean
+    defaultModel: string
+    models: Array<{ id: string; capabilities: { vision: boolean; ocr: boolean } }>
+  }>
+  defaultProviderId?: string
+}
+
+class ImageInsightService {
+  private providerService: ProviderServiceLike | null = null
+  private decryptService: DecryptServiceLike | null = null
+  /** 最近一次实际使用的默认 AI provider/model，仅用于写入分析元数据 */
+  private runtimeProviderId: string | undefined = undefined
+  private runtimeModelId: string | undefined = undefined
+
+  /** 注入依赖(由 main/index.ts 在 app ready 后调用) */
+  bind(deps: {
+    providerService: ProviderServiceLike & { list(): ProviderSummaryLike }
+    decryptService: DecryptServiceLike
+  }): void {
+    this.providerService = deps.providerService
+    this.decryptService = deps.decryptService
+    console.log(
+      '[ImageInsightService] bind ok, default provider=%s model=%s',
+      this.runtimeProviderId,
+      this.runtimeModelId
+    )
+    // 读取默认 provider/model(后续 analyze 时使用)
+    try {
+      const list = deps.providerService.list()
+      const provider =
+        list.providers.find((p) => p.id === list.defaultProviderId) || list.providers[0]
+      this.runtimeProviderId = provider?.id
+      this.runtimeModelId = provider?.defaultModel
+      console.log(
+        '[ImageInsightService] bind loaded default provider=%s model=%s',
+        this.runtimeProviderId,
+        this.runtimeModelId
+      )
+    } catch (error) {
+      console.warn('[ImageInsightService] bind list failed:', error)
+    }
+  }
+
+  /**
+   * 计算图片缓存 key:imageHash。
+   * 策略:优先微信原始 md5,无 md5 才用 sha256(rawBytes).slice(0, 32)
+   */
+  private async computeImageHash(
+    md5: string | undefined,
+    datName: string | undefined
+  ): Promise<string | null> {
+    if (md5 && md5.trim()) return md5.trim().toLowerCase()
+    if (!this.decryptService) return null
+    const filePath = this.decryptService.findImageFile(undefined, datName, { allowThumbnail: true })
+    if (!filePath) return null
+    // 一次性读盘 + sha256(只在没有 md5 时才付出 IO)
+    try {
+      const fs = await import('fs-extra')
+      const buf = await fs.readFile(filePath)
+      const sha = crypto.createHash('sha256').update(buf).digest('hex').slice(0, 32)
+      return `sha256:${sha}`
+    } catch (error) {
+      console.warn('[ImageInsightService] computeImageHash failed:', error)
+      return null
+    }
+  }
+
+  /** 通过 hash 拿 Insight(只读缓存,无 AI 调用) */
+  getInsight(imageHash: string): ImageInsight | null {
+    return imageInsightsStore.getByHash(imageHash)
+  }
+
+  /**
+   * 主入口:分析一张图片。
+   * 1. 通过 imageHash 查缓存,命中即返回
+   * 2. 未命中:解密图片 → 调 AI → 解析响应 → 落库 → 返回
+   * 3. 任意步骤失败:记录日志,返回 success=false,**不抛**
+   */
+  async analyze(request: ImageAnalysisRequest): Promise<ImageAnalysisResponse> {
+    try {
+      if (!request.force) {
+        const cached = imageInsightsStore.getByHash(request.imageHash)
+        if (cached) {
+          return { success: true, insight: cached, fromCache: true }
+        }
+      }
+
+      if (!this.providerService) {
+        return { success: false, error: 'AI Provider 未初始化' }
+      }
+
+      const messages = [
+        {
+          role: 'system',
+          content: IMAGE_ANALYSIS_SYSTEM_PROMPT
+        },
+        {
+          role: 'user',
+          content: [
+            {
+              type: 'text' as const,
+              text: buildImageAnalysisUserText({
+                sender: request.sender,
+                sentAt: request.sentAt,
+                contextBefore: [],
+                contextAfter: []
+              })
+            },
+            { type: 'image' as const, dataUrl: request.imageDataUrl }
+          ]
+        }
+      ]
+
+      const list = this.providerService.list()
+      const provider =
+        list.providers.find((item) => item.id === list.defaultProviderId) || list.providers[0]
+      this.runtimeProviderId = provider?.id
+      this.runtimeModelId = provider?.defaultModel
+      const result = await this.providerService.analyzeImage(messages, {
+        providerId: this.runtimeProviderId,
+        modelId: this.runtimeModelId
+      })
+      if (!result.success || !result.data) {
+        console.warn('[ImageInsightService] analyze vision failed: %s', result.error || 'no data')
+        return { success: false, error: result.error || 'AI 未返回内容' }
+      }
+      console.log('[ImageInsightService] analyze ok, description=%s', result.data.slice(0, 80))
+
+      const parsed = parseImageAnalysisResponse(result.data)
+      const insight: ImageInsight = {
+        id: randomUUID(),
+        messageId: request.messageId,
+        imageHash: request.imageHash,
+        md5: undefined,
+        datName: undefined,
+        description: parsed.description,
+        ocrText: parsed.ocrText || undefined,
+        tags: parsed.tags,
+        category: parsed.category,
+        importance: parsed.importance,
+        provider: this.runtimeProviderId || '',
+        model: this.runtimeModelId || '',
+        createdAt: Date.now(),
+        updatedAt: Date.now(),
+        sender: request.sender,
+        sentAt: request.sentAt,
+        sessionId: request.sessionId
+      }
+      imageInsightsStore.upsert(insight)
+      return { success: true, insight, fromCache: false }
+    } catch (error) {
+      const message = error instanceof Error ? error.message : String(error)
+      console.warn('[ImageInsightService] analyze failed:', message)
+      return { success: false, error: message }
+    }
+  }
+
+  /**
+   * 日报入口:从 renderer 传入的图片消息候选中挑 Top N + 命中缓存的 Insight。
+   *
+   * 设计:不自己查 chat-service(参数语义不清),而是由 renderer 从已加载的 messages 中
+   * 提取图片消息 + 计算热度后传入。这样既复用现有数据,又避免 userMd5/sessionId 混淆。
+   */
+  async listTopHotImages(
+    query: ImageCandidateQuery,
+    inputs: ImageCandidateInput[] = []
+  ): Promise<ImageCandidate[]> {
+    const limit = query.limit ?? 3
+    const candidates: ImageCandidate[] = []
+    console.log('[ImageInsightService] listTopHotImages received %d inputs', inputs.length)
+    for (const input of inputs) {
+      const hash = await this.computeImageHash(input.md5, input.datName)
+      if (!hash) {
+        console.log(
+          '[ImageInsightService] skip %s: hash empty (md5=%s datName=%s)',
+          input.messageId,
+          input.md5,
+          input.datName
+        )
+        continue
+      }
+      const heatScore = input.responseCount * 3 + input.interactionCount * 2 + 1
+      const candidate: ImageCandidate = {
+        messageId: input.messageId,
+        imageHash: hash,
+        md5: input.md5,
+        datName: input.datName,
+        sessionId: input.sessionId,
+        sender: input.sender,
+        sentAt: input.sentAt,
+        heatScore
+      }
+      const cached = imageInsightsStore.getByHash(hash)
+      if (cached) candidate.insight = cached
+      candidates.push(candidate)
+    }
+    candidates.sort((a, b) => b.heatScore - a.heatScore)
+    return candidates.slice(0, limit)
+  }
+
+  /**
+   * 列出会话所有 insights(按时间倒序,供未来 UI 复用)
+   */
+  listBySession(sessionId: string, limit?: number): ImageInsight[] {
+    return imageInsightsStore.listBySession(sessionId, limit)
+  }
+
+}
+
+export const imageInsightService = new ImageInsightService()