shenjiecode · midmirror · Aug 2, 2026
diff --git a/README.md b/README.md
@@ -148,7 +148,7 @@ WeChat  ⇄  pi TUI session  ⇄  AI model + tools
 - WeChat messages are fetched via iLink Bot API long polling
 - Incoming messages are injected into the active pi session via `pi.sendUserMessage()`
 - When you type in TUI, a preview is sent to WeChat
-- AI replies are delivered incrementally (per `message_end`) and finalized on `agent_end`
+- AI replies are delivered incrementally (per `message_end`); the `agent_end` catch-up replay was removed to prevent duplicate historical messages after session restore
 - Only the TUI session that runs `/wechat start` holds the connection
 
 ## FAQ
@@ -304,7 +304,7 @@ pi install git:github.com/shenjiecode/pi-wechat-assistant
 - 微信消息通过 iLink Bot API 长轮询获取
 - 收到的消息通过 `pi.sendUserMessage()` 注入当前 pi 会话
 - TUI 输入时微信端会收到预览
-- AI 回复增量发送（每条 `message_end` 即发），`agent_end` 时补发遗漏
+- AI 回复增量发送（每条 `message_end` 即发），不依赖 `agent_end` 补发（防止恢复会话后历史回复重发）
 - 只有执行 `/wechat start` 的 TUI 会话持有连接
 
 ## 常见问题
@@ -460,7 +460,7 @@ WeChat  ⇄  pi TUIセッション  ⇄  AIモデル + ツール
 - WeChatメッセージはiLink Bot APIのロングポーリングで取得
 - 受信メッセージは `pi.sendUserMessage()` で現在のpiセッションに注入
 - TUIで入力するとWeChat側にプレビューが送信
-- AI返信は増分的に送信（`message_end` ごと）、`agent_end` で残りを補完
+- AI返信は増分的に送信（`message_end` ごと）、`agent_end` での補完は行わない（セッション復元後の過去返信重複送信を防止）
 - `/wechat start` を実行したTUIセッションのみが接続を保持
 
 ## FAQ
@@ -616,7 +616,7 @@ WeChat  ⇄  pi TUI 세션  ⇄  AI 모델 + 도구
 - WeChat 메시지는 iLink Bot API 롱 폴링으로 가져옴
 - 수신 메시지는 `pi.sendUserMessage()` 로 현재 pi 세션에 주입
 - TUI에서 입력하면 WeChat 측에 미리보기 전송
-- AI 응답은 증분 전송 (`message_end` 마다), `agent_end` 에서 나머지 보완
+- AI 응답은 증분 전송 (`message_end` 마다), `agent_end` 보완 없음 (세션 복원 후 과거 응답 중복 방지)
 - `/wechat start` 를 실행한 TUI 세션만 연결 유지
 
 ## FAQ

diff --git a/src/auth.ts b/src/auth.ts
@@ -15,6 +15,8 @@ const CREDS_FILE = path.join(STATE_DIR, 'credentials.json')
 const CONFIG_FILE = path.join(STATE_DIR, 'config.json')
 const LOCK_FILE = path.join(STATE_DIR, 'session.lock')
 const CONTEXT_TOKENS_FILE = path.join(STATE_DIR, 'context-tokens.json')
+const CURSOR_FILE = path.join(STATE_DIR, 'cursor.json')
+const SEEN_IDS_FILE = path.join(STATE_DIR, 'seen-ids.json')
 
 export function getStateDir(): string {
   return STATE_DIR
@@ -52,6 +54,35 @@ async function deleteFile(filePath: string): Promise<void> {
   }
 }
 
+// --- 消息游标（跨重启持久化，防止重放历史消息导致重复回复） ---
+
+export async function loadCursor(): Promise<string> {
+  const data = await readJsonFile<{ cursor: string }>(CURSOR_FILE)
+  return data?.cursor ?? ''
+}
+
+export async function saveCursor(cursor: string): Promise<void> {
+  if (!cursor) return
+  await writeJsonFile(CURSOR_FILE, { cursor, updatedAt: new Date().toISOString() })
+}
+
+// --- 已见消息 id（持久化去重，游标丢失/过期时兜底防重发） ---
+
+const MAX_SEEN_IDS = 5_000
+
+export async function loadSeenIds(): Promise<Set<string>> {
+  const data = await readJsonFile<{ ids: string[] }>(SEEN_IDS_FILE)
+  return new Set(data?.ids ?? [])
+}
+
+export async function saveSeenIds(ids: Set<string>): Promise<void> {
+  const arr = Array.from(ids)
+  if (arr.length === 0) return
+  // 只保留最近 N 个，防止文件无限增长
+  const trimmed = arr.length > MAX_SEEN_IDS ? arr.slice(arr.length - MAX_SEEN_IDS) : arr
+  await writeJsonFile(SEEN_IDS_FILE, { ids: trimmed, updatedAt: new Date().toISOString() })
+}
+
 // --- 凭证 ---
 
 export async function loadCredentials(): Promise<Credentials | null> {

diff --git a/src/client.ts b/src/client.ts
@@ -22,6 +22,10 @@ import {
   loadContextTokens,
   saveContextTokensThrottled,
   flushContextTokens,
+  loadCursor,
+  saveCursor,
+  loadSeenIds,
+  saveSeenIds,
 } from './auth.js'
 import {
   CDN_BASE,
@@ -130,6 +134,7 @@ export class WeixinClient {
   private cursor = ''
   private readonly typingTickets = new Map<string, string>()
   private readonly contextTokens = new Map<string, string>()
+  private _seenMessageIds = new Set<string>()
   private _lastActiveUserId: string | null = null
   private _contextTokensDirty = false
   private _disposed = false
@@ -152,6 +157,10 @@ export class WeixinClient {
     for (const [userId, token] of Object.entries(persisted.tokens)) {
       this.contextTokens.set(userId, token)
     }
+    // 恢复上次游标，避免重启后服务端重放历史消息导致重复回复
+    this.cursor = await loadCursor()
+    // 恢复已见消息 id（游标丢失/过期时兜底去重）
+    this._seenMessageIds = await loadSeenIds()
   }
 
   get accountId(): string { return this.credentials.accountId }
@@ -187,12 +196,35 @@ export class WeixinClient {
     }
 
     this.cursor = response.get_updates_buf || this.cursor
+    // 持久化游标（写失败不致命，下次可能重放；由 _seenMessageIds 兜底去重）
+    if (this.cursor) {
+      saveCursor(this.cursor).catch(() => {})
+    }
+
     const incoming: IncomingMessage[] = []
 
     for (const raw of response.msgs ?? []) {
       this.rememberContext(raw)
       const normalized = this.normalizeIncomingMessage(raw)
-      if (normalized) incoming.push(normalized)
+      if (!normalized) continue
+      const id = normalized.messageId
+      if (id && this._seenMessageIds.has(id)) {
+        debugLog(`[DEDUP] skip already-seen message ${id}`)
+        continue
+      }
+      if (id) {
+        this._seenMessageIds.add(id)
+        // 防止 Set 无限增长：超限重建（极端情况下去重失效，游标仍是主防线）
+        if (this._seenMessageIds.size > 10_000) {
+          this._seenMessageIds.clear()
+          this._seenMessageIds.add(id)
+        }
+      }
+      incoming.push(normalized)
+    }
+    // 持久化已见 id（fire-and-forget，节流由文件大小控制）
+    if (incoming.length > 0) {
+      saveSeenIds(this._seenMessageIds).catch(() => {})
     }
     return incoming
   }

diff --git a/src/index.ts b/src/index.ts
@@ -15,7 +15,7 @@ import { splitAndFilterMarkdown } from './message.js'
 import { MessageQueue } from './queue.js'
 import { handleRemoteCommand, type RemoteCommandDeps } from './remote-commands.js'
 import { registerCommands, type CommandDeps } from './commands.js'
-import { ok, fail, formatError, isAbortError, extractAllAssistantReplies, extractTextFromMessageContent } from './utils.js'
+import { ok, fail, formatError, isAbortError, extractTextFromMessageContent } from './utils.js'
 import {
   POLL_RETRY_BASE_MS,
   POLL_RETRY_MAX_MS,
@@ -480,7 +480,9 @@ export default function wechatAssistant(pi: ExtensionAPI) {
     }
   })
 
-  // agent 结束 → 补发遗漏 + 收尾
+  // agent 结束 → 收尾（补发已移除：message_end 已逐条增量发送，
+  // 补发依赖 event.messages 为“本次 run”的范围假设，omp -c 恢复会话后
+  // messages 含全量历史，slice(sentCount) 会把历史回复全部重发到微信）
   pi.on('agent_end', async (event, ctx) => {
     latestCtx = ctx
     agentIdle = true
@@ -491,24 +493,6 @@ export default function wechatAssistant(pi: ExtensionAPI) {
     const assistantMsgs = turn.messages.filter(m => m?.role === 'assistant').length
     log(`[AGENT-END] turn#${turn.seq} source=${turn.wechatConversationActive ? 'WECHAT' : 'TUI'} targetUser=${turn.targetUser} messages=${msgCount} assistant=${assistantMsgs} sentCount=${turn.sentCount}`)
 
-    const allReplies = extractAllAssistantReplies(turn.messages)
-    const newReplies = allReplies.slice(turn.sentCount)
-    log(`[AGENT-END-REPLIES] all=${allReplies.length} sent=${turn.sentCount} new=${newReplies.length}`)
-
-    if (turn.wechatConversationActive && newReplies.length > 0 && client && turn.targetUser) {
-      try {
-        await queue.sendRepliesToWechat(newReplies, turn.targetUser)
-        log(`[AGENT-END-DONE] sent ${newReplies.length} remaining replies`)
-      } catch (err) {
-        log(`[AGENT-END-ERROR] ${formatError(err)}`)
-        notify(`发送微信回复失败: ${formatError(err)}`, 'error')
-      }
-    } else if (allReplies.length === 0) {
-      log(`[AGENT-END-NOREPLY] no assistant text`)
-    } else {
-      log(`[AGENT-END-SAFE] all replies already sent incrementally`)
-    }
-
     if (queue.activeRequest) {
       await client?.stopTyping(queue.activeRequest.userId).catch(() => {})
       queue.activeRequest = null

diff --git a/src/queue.ts b/src/queue.ts
@@ -6,7 +6,6 @@ import { randomUUID } from 'node:crypto'
 import type { ExtensionContext, ExtensionCommandContext } from '@mariozechner/pi-coding-agent'
 import { debugLog } from './logger.js'
 import { WeixinClient } from './client.js'
-import { splitAndFilterMarkdown } from './message.js'
 import { fetchImageAsBase64, fetchFile, saveFileToDisk, type ImageData } from './media.js'
 import {
   ACK_TEXT,
@@ -374,19 +373,6 @@ export class MessageQueue {
     }
   }
 
-  // --- 统一发送到微信 ---
-
-  async sendRepliesToWechat(replies: string[], targetUserId: string): Promise<void> {
-    const client = this.getClient()
-    if (!client) return
-    for (const reply of replies) {
-      const chunks = splitAndFilterMarkdown(reply)
-      for (const chunk of chunks) {
-        await client.sendText(targetUserId, chunk)
-      }
-    }
-  }
-
   // --- 重置 ---
 
   reset(): void {

diff --git a/src/utils.ts b/src/utils.ts
@@ -19,33 +19,6 @@ export async function renderQrCode(url: string): Promise<string> {
   })
 }
 
-export function extractAllAssistantReplies(
-  messages: Array<{ role?: string; content?: unknown }>,
-): string[] {
-  const replies: string[] = []
-  for (let i = 0; i < messages.length; i++) {
-    const message = messages[i]
-    if (message?.role !== 'assistant') continue
-    if (typeof message.content === 'string') {
-      const text = message.content.trim()
-      if (text) replies.push(text)
-      continue
-    }
-    if (!Array.isArray(message.content)) continue
-    const text = message.content
-      .filter(
-        (part): part is { type: 'text'; text: string } =>
-          typeof part === 'object' && part !== null && (part as { type?: string }).type === 'text',
-      )
-      .map((part) => part.text.trim())
-      .filter(Boolean)
-      .join('\n')
-      .trim()
-    if (text) replies.push(text)
-  }
-  return replies
-}
-
 export function extractTextFromMessageContent(content: unknown): string | null {
   if (typeof content === 'string') return content.trim() || null
   if (!Array.isArray(content)) return null

diff --git a/tests/utils.test.ts b/tests/utils.test.ts
@@ -6,7 +6,6 @@ import { describe, it, expect } from 'vitest'
 import {
   ok,
   fail,
-  extractAllAssistantReplies,
   extractTextFromMessageContent,
   summarizePreview,
   formatError,
@@ -26,50 +25,6 @@ describe('ok / fail', () => {
   })
 })
 
-describe('extractAllAssistantReplies', () => {
-  it('extracts string content from assistant messages', () => {
-    const messages = [
-      { role: 'user', content: 'hello' },
-      { role: 'assistant', content: 'Hi there' },
-      { role: 'assistant', content: '  How can I help?  ' },
-    ]
-    expect(extractAllAssistantReplies(messages)).toEqual(['Hi there', 'How can I help?'])
-  })
-
-  it('extracts array content with text parts', () => {
-    const messages = [
-      {
-        role: 'assistant',
-        content: [
-          { type: 'text', text: 'Part 1' },
-          { type: 'toolCall', name: 'send_file' },
-          { type: 'text', text: 'Part 2' },
-        ],
-      },
-    ]
-    expect(extractAllAssistantReplies(messages)).toEqual(['Part 1\nPart 2'])
-  })
-
-  it('skips non-assistant messages', () => {
-    const messages = [
-      { role: 'user', content: 'hello' },
-      { role: 'toolResult', content: 'result' },
-    ]
-    expect(extractAllAssistantReplies(messages)).toEqual([])
-  })
-
-  it('skips empty assistant messages', () => {
-    const messages = [
-      { role: 'assistant', content: '' },
-      { role: 'assistant', content: '  ' },
-    ]
-    expect(extractAllAssistantReplies(messages)).toEqual([])
-  })
-
-  it('handles empty array', () => {
-    expect(extractAllAssistantReplies([])).toEqual([])
-  })
-})
 
 describe('extractTextFromMessageContent', () => {
   it('extracts from string', () => {