feat: enforce full AI outbound generation and reduce template replies

core/agent_pre_rules.py

@@ -70,10 +70,9 @@ class AgentPreRuleService:
         )
 
     async def _rule_pred_meaningless_short_text(self, ctx: RuleContext) -> bool:
-        from core.pydantic_ai_agent import _is_meaningless_short_text
-
         message = ctx.get("message")
-        return _is_meaningless_short_text(message.msg)
+        state = ctx.get("state")
+        return self.agent._should_handle_as_meaningless_short_text(state, message.msg)
 
     async def _rule_act_meaningless_short_text(self, ctx: RuleContext) -> RuleResult:
         from core.pydantic_ai_agent import AgentResponse

core/ai_reply_flow.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 import logging
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING, Any, Optional, Tuple
 from core.post_ops import negotiation_strategy_reply
 
 logger = logging.getLogger("cs_agent")
@@ -10,7 +10,43 @@ if TYPE_CHECKING:
     from core.pydantic_ai_agent import AgentDeps, ConversationState, CustomerMessage, CustomerServiceAgent
 
 
-def select_target_agent(agent: "CustomerServiceAgent", message: "CustomerMessage", state: "ConversationState"):
+def _select_agent_by_intent(
+    agent: "CustomerServiceAgent",
+    message: "CustomerMessage",
+    state: "ConversationState",
+) -> Tuple[Optional[Any], str]:
+    """
+    AI 意图优先路由；识别不到时返回 (None, "intent:none")，由关键词兜底。
+    """
+    try:
+        from utils.intent_analyzer import detect_intent
+
+        decision = detect_intent(message.msg or "")
+        intent = (decision.intent or "").strip()
+        source = decision.source or "none"
+        score = float(decision.score or 0.0)
+    except Exception:
+        intent, source, score = "", "error", 0.0
+
+    if not intent:
+        return None, "intent:none"
+
+    if intent in ("询价", "砍价"):
+        return agent.agent_pricing, f"intent:{intent}|src:{source}|score:{score:.3f}"
+    if intent in ("修改", "加急"):
+        return agent.agent_processing, f"intent:{intent}|src:{source}|score:{score:.3f}"
+    if intent == "售后":
+        return agent.agent_after_sale, f"intent:{intent}|src:{source}|score:{score:.3f}"
+    if intent == "转接":
+        return agent.agent_after_sale, f"intent:{intent}|src:{source}|score:{score:.3f}"
+    if intent in ("打招呼", "批量", "发图"):
+        target = agent.agent_after_sale if state.stage == "售后" else agent.agent
+        return target, f"intent:{intent}|src:{source}|score:{score:.3f}"
+
+    return None, f"intent:unmapped:{intent}|src:{source}|score:{score:.3f}"
+
+
+def select_target_agent(agent: "CustomerServiceAgent", message: "CustomerMessage", state: "ConversationState") -> Tuple[Any, str]:
     msg_lower = message.msg.lower()
     pricing_kw = ["多少钱", "多少一张", "报价", "给个价", "几块", "价位", "能便宜点吗"]
     processing_kw = ["安排", "处理一下", "开始做", "做一下", "尽快", "加急", "付款了", "已付款"]
@@ -45,18 +81,23 @@ def select_target_agent(agent: "CustomerServiceAgent", message: "CustomerMessage
         "卫星地图",
     ]
     target_agent = agent.agent_after_sale if state.stage == "售后" else agent.agent
+
+    ai_target, ai_reason = _select_agent_by_intent(agent, message, state)
+    if ai_target is not None:
+        return ai_target, ai_reason
+
     risk_hit = any(k in msg_lower for k in risk_kw) or agent._is_political_inquiry(message.msg) or agent._is_map_inquiry(message.msg)
     if risk_hit:
-        return agent.agent_risk
+        return agent.agent_risk, "keyword:risk"
     if any(k in message.msg for k in order_markers):
-        return agent.agent_order
+        return agent.agent_order, "keyword:order"
     if any(k in msg_lower for k in processing_kw):
-        return agent.agent_processing
+        return agent.agent_processing, "keyword:processing"
     if any(k in msg_lower for k in pricing_kw):
-        return agent.agent_pricing
+        return agent.agent_pricing, "keyword:pricing"
     if any(k in msg_lower for k in similar_kw):
-        return agent.agent_similar
-    return target_agent
+        return agent.agent_similar, "keyword:similar"
+    return target_agent, "fallback:default"
 
 
 async def execute_ai_turn(
@@ -68,7 +109,8 @@ async def execute_ai_turn(
     deps: "AgentDeps",
     history: list,
 ) -> str:
-    target_agent = select_target_agent(agent, message, state)
+    target_agent, route_reason = select_target_agent(agent, message, state)
+    logger.info("[路由] %s", route_reason)
     result = await target_agent.run(user_prompt, deps=deps, message_history=history)
     agent.message_histories[message.from_id] = result.all_messages()[-30:]
     reply_text = agent._colloquialize_reply(agent._normalize_reply_text(result.output))

core/pydantic_ai_agent.py

@@ -490,23 +490,9 @@ class CustomerServiceAgent:
         """
         收图阶段回复默认走 AI 改写，失败时回退到固定模板。
         """
-        # 首张收图先承接“我看一下”，避免机械地立刻催“发完统一报价”。
+        first_image_ack = "收到，我先看一下哈，稍等哈。"
         if scene == "collect_ack" and len(state.pending_image_urls) == 1:
-            first_ack = [
-                "收到了，我先看一下哈，稍等哈",
-                "这张我收到了，我先看下，稍等我一下哈",
-                "收到这张了，我先过一眼，稍等哈",
-                "我先看这张哈，稍等我一下",
-                "图我收到了，我先看一眼，稍等我回你哈",
-                "这张先记上了，我先看下细节，稍等哈",
-                "收到哈，我先过一遍这张，稍等我会儿",
-                "我先看这张效果，稍等一下哈",
-                "图到了，我先看下清晰度，稍等哈",
-                "这张我先看着，稍等我一下就回你",
-                "收到这张了，我先核一下细节，稍等哈",
-                "我先把这张看完，稍等我一会儿哈",
-            ]
-            return random.choice(first_ack)
+            fallback = first_image_ack
         if not self.dynamic_collection_replies:
             return fallback
         try:
@@ -525,7 +511,7 @@ class CustomerServiceAgent:
                 f"客户原话: {message.msg}\n"
                 f"当前已收图片数: {len(state.pending_image_urls)}\n"
                 f"当前需求摘要: {pending_req}\n"
-                "输出要求: 不超过2句话，像真人店主聊天。"
+                "输出要求: 不超过2句话，像真人店主聊天；避免复用固定模板句。"
             )
             result = await self.agent_natural_reply.run(user_prompt, deps=deps, message_history=history)
             self.message_histories[message.from_id] = result.all_messages()[-30:]
@@ -695,6 +681,44 @@ class CustomerServiceAgent:
         clean = msg.strip().rstrip("！!？?。.～~")
         return clean in self._COOLDOWN_PATTERNS
 
+    def _should_handle_as_meaningless_short_text(self, state: ConversationState, msg: str) -> bool:
+        """
+        无意义短句仅在“非业务处理中”生效，避免误拦截真实推进消息。
+        例如：已在收图/待报价阶段时，客户发“好的/在吗”不应直接 ping。
+        """
+        customer_text, _ = self._split_customer_text(msg or "")
+        text = (customer_text or "").strip()
+        if not _is_meaningless_short_text(text):
+            return False
+        if self._extract_image_urls(text):
+            return False
+        if (getattr(state, "pending_image_urls", None) or []):
+            return False
+        if getattr(state, "quote_phase", "idle") in {"collecting", "ready_to_quote", "waiting_result"}:
+            return False
+        return True
+
+    async def build_auto_quote_reply(self, state: ConversationState, message: CustomerMessage) -> AgentResponse:
+        """
+        自动报价内部入口：不走 process_message，避免伪造客户语句污染上下文。
+        """
+        quote_res = await self._quote_pending_images(state, message)
+        reply_text = self._colloquialize_reply(quote_res.get("reply", ""))
+        reply_text = await self._rewrite_reply_with_ai(
+            message=message,
+            state=state,
+            reply=reply_text,
+            scene="batch_quote_reply",
+        )
+        need_transfer = bool(quote_res.get("need_transfer"))
+        state.last_reply_at = datetime.now()
+        return AgentResponse(
+            reply=reply_text,
+            should_reply=not need_transfer,
+            need_transfer=need_transfer,
+            transfer_msg=TRANSFER_MESSAGE if need_transfer else "",
+        )
+
     async def process_message(self, message: CustomerMessage) -> AgentResponse:
         """处理客户消息并生成回复。"""
         return await process_incoming_message(self, message)

core/websocket_client.py

@@ -8,7 +8,7 @@ import secrets
 import time
 import hashlib
 from collections import deque
-from datetime import datetime
+from datetime import datetime, timedelta
 from pathlib import Path
 from typing import Optional, Dict, Any, List
 from utils.observability import emit_activity, build_trace_id
@@ -170,6 +170,8 @@ class QingjianAPIClient:
         self._last_reply_sent_at: dict = {}  # customer_key -> monotonic ts
         self._outbound_semantic_seen: dict = {}  # customer_key -> {semantic_key: ts}
         self._outbound_class_seen: dict = {}     # customer_key -> {reply_class: ts}
+        self._outbound_template_seen: dict = {}  # customer_key -> {template_family: ts}
+        self._unreplied_followup_sent: dict = {}  # customer_key -> monotonic ts（补偿消息节流）
         self._inbound_log_seen: dict = {}    # signature -> monotonic ts（防重复写入）
         self._outbound_log_seen: dict = {}   # signature -> monotonic ts（防重复写入）
         self._tianwang_callback_url = (
@@ -179,6 +181,7 @@ class QingjianAPIClient:
         self._tianwang_agent_name = os.getenv("TIANWANG_AGENT_NAME", "终结者").strip() or "终结者"
         self._reply_guard_enabled = os.getenv("AI_REPLY_GUARD_ENABLED", "true").lower() in ("1", "true", "yes")
         self._reply_guard_verbose = os.getenv("AI_REPLY_GUARD_VERBOSE", "false").lower() in ("1", "true", "yes")
+        self._force_ai_generate_reply = os.getenv("FORCE_AI_GENERATE_ALL_REPLIES", "true").lower() in ("1", "true", "yes")
         
         # 延迟加载任务模块（避免循环导入）
         self.task_scheduler = None
@@ -429,6 +432,22 @@ class QingjianAPIClient:
             return "ack"
         return "general"
 
+    @staticmethod
+    def _template_family(reply: str) -> str:
+        """识别高频模板家族，做疲劳抑制。"""
+        s = (reply or "").strip()
+        if not s:
+            return ""
+        if "需求我记上了" in s and "继续发图" in s:
+            return "collect_remind"
+        if ("这批图过一遍" in s or "收齐了" in s or "收好了" in s) and ("总价" in s or "报价" in s):
+            return "quote_defer"
+        if "图片收到了" in s and "继续发" in s:
+            return "collect_ack"
+        if "好嘞，你稍等下，我这边看一下" in s:
+            return "fallback_ack"
+        return ""
+
     def _outbound_arbiter(self, original_msg: dict, reply_content: str, trace_id: str) -> tuple[bool, str]:
         """
         统一出站裁决层：
@@ -447,11 +466,17 @@ class QingjianAPIClient:
             class_window = max(20, int(os.getenv("AI_OUTBOUND_CLASS_DEDUPE_SECONDS", "90")))
         except Exception:
             class_window = 90
+        try:
+            template_window = max(120, int(os.getenv("AI_OUTBOUND_TEMPLATE_FATIGUE_SECONDS", "600")))
+        except Exception:
+            template_window = 600
 
         sem_bucket = self._outbound_semantic_seen.setdefault(key, {})
         cls_bucket = self._outbound_class_seen.setdefault(key, {})
+        tpl_bucket = self._outbound_template_seen.setdefault(key, {})
         self._prune_seen(sem_bucket, now_mono, ttl_sec=max(sem_window * 2, 240))
         self._prune_seen(cls_bucket, now_mono, ttl_sec=max(class_window * 2, 180))
+        self._prune_seen(tpl_bucket, now_mono, ttl_sec=max(template_window * 2, 1200))
 
         if sem_key and (now_mono - sem_bucket.get(sem_key, 0.0)) < sem_window:
             self._activity_log(
@@ -466,6 +491,19 @@ class QingjianAPIClient:
             )
             return False, "semantic_duplicate"
 
+        template_family = self._template_family(reply_content)
+        if template_family and (now_mono - tpl_bucket.get(template_family, 0.0)) < template_window:
+            self._activity_log(
+                "outbound_arbiter_block",
+                trace_id=trace_id,
+                acc_id=original_msg.get("acc_id", ""),
+                customer_id=original_msg.get("from_id", ""),
+                reason="template_fatigue",
+                template_family=template_family,
+                msg=reply_content,
+            )
+            return False, "template_fatigue"
+
         if reply_class in {"quote", "collect", "ack"} and (now_mono - cls_bucket.get(reply_class, 0.0)) < class_window:
             self._activity_log(
                 "outbound_arbiter_block",
@@ -481,16 +519,199 @@ class QingjianAPIClient:
         if sem_key:
             sem_bucket[sem_key] = now_mono
         cls_bucket[reply_class] = now_mono
+        if template_family:
+            tpl_bucket[template_family] = now_mono
         self._activity_log(
             "outbound_arbiter_pass",
             trace_id=trace_id,
             acc_id=original_msg.get("acc_id", ""),
             customer_id=original_msg.get("from_id", ""),
             reply_class=reply_class,
+            template_family=template_family,
             semantic_key=sem_key[:80] if sem_key else "",
         )
         return True, "pass"
 
+    async def _unreplied_followup_loop(self):
+        """
+        定时补偿：对“最后一条是客户消息且长时间未回复”的会话，补发一次自然跟进。
+        """
+        if not self.enable_agent or not self.agent:
+            return
+        while self.running:
+            try:
+                await asyncio.sleep(max(30, int(os.getenv("UNREPLIED_FOLLOWUP_SCAN_SECONDS", "90"))))
+                await self._scan_and_send_unreplied_followups()
+            except asyncio.CancelledError:
+                break
+            except Exception as e:
+                self._activity_log("unreplied_followup_loop_error", error=str(e))
+
+    async def _scan_and_send_unreplied_followups(self):
+        from db import chat_log_db as cdb
+        try:
+            idle_minutes = max(5, int(os.getenv("UNREPLIED_FOLLOWUP_IDLE_MINUTES", "12")))
+            max_age_minutes = max(idle_minutes, int(os.getenv("UNREPLIED_FOLLOWUP_MAX_AGE_MINUTES", "180")))
+            followup_cd = max(300, int(os.getenv("UNREPLIED_FOLLOWUP_COOLDOWN_SECONDS", "3600")))
+            limit = max(10, int(os.getenv("UNREPLIED_FOLLOWUP_LIMIT", "40")))
+        except Exception:
+            idle_minutes, max_age_minutes, followup_cd, limit = 12, 180, 3600, 40
+
+        now = datetime.now()
+        window_start = (now - timedelta(minutes=max_age_minutes)).strftime("%Y-%m-%d %H:%M:%S")
+        conn = None
+        try:
+            conn = cdb._get_conn()
+            rows = conn.execute(
+                cdb._sql(
+                    """
+                    SELECT acc_id, customer_id, MAX(id) AS last_id
+                    FROM chat_logs
+                    WHERE timestamp >= ?
+                    GROUP BY acc_id, customer_id
+                    ORDER BY MAX(id) DESC
+                    LIMIT ?
+                    """
+                ),
+                (window_start, limit * 6),
+            ).fetchall()
+            sessions = [dict(r) for r in rows]
+            sent = 0
+            for s in sessions:
+                if sent >= limit:
+                    break
+                acc_id = str(s.get("acc_id", "") or "")
+                cid = str(s.get("customer_id", "") or "")
+                if not acc_id or not cid:
+                    continue
+                ckey = f"{acc_id}:{cid}"
+                if not self._is_owned_by_this_worker(ckey):
+                    continue
+                last = conn.execute(
+                    cdb._sql(
+                        """
+                        SELECT id, direction, message, timestamp, customer_name, acc_id, platform
+                        FROM chat_logs
+                        WHERE acc_id = ? AND customer_id = ?
+                        ORDER BY id DESC
+                        LIMIT 1
+                        """
+                    ),
+                    (acc_id, cid),
+                ).fetchone()
+                if not last:
+                    continue
+                last = dict(last)
+                if str(last.get("direction", "")) != "in":
+                    continue
+                last_ts = last.get("timestamp")
+                if isinstance(last_ts, datetime):
+                    last_dt = last_ts
+                else:
+                    last_dt = datetime.strptime(str(last_ts)[:19], "%Y-%m-%d %H:%M:%S")
+                idle_s = (now - last_dt).total_seconds()
+                if idle_s < idle_minutes * 60 or idle_s > max_age_minutes * 60:
+                    continue
+                now_mono = time.monotonic()
+                if (now_mono - self._unreplied_followup_sent.get(ckey, 0.0)) < followup_cd:
+                    continue
+                # 避免对明显结束语/确认语再次打扰
+                last_msg = str(last.get("message", "") or "").strip().lower()
+                if last_msg in {"好的", "好", "ok", "收到", "嗯", "哦"}:
+                    continue
+
+                followup = await self._compose_ai_scene_reply(
+                    original_msg={
+                        "acc_id": acc_id,
+                        "from_id": cid,
+                        "from_name": self.to_chinese(last.get("customer_name", "") or cid),
+                        "acc_type": str(last.get("platform", "") or "AliWorkbench"),
+                        "msg": str(last.get("message", "") or ""),
+                    },
+                    scene="unreplied_followup",
+                    intent_hint="客户上一条消息还没接上，先自然承接并请对方补一句当前要处理的图或要求。",
+                    fallback="刚看到你消息了，我在的。你把要处理的图或要求再发我一下，我马上接着看。",
+                )
+                fake = {
+                    "acc_id": acc_id,
+                    "from_id": cid,
+                    "from_name": self.to_chinese(last.get("customer_name", "") or cid),
+                    "cy_id": cid,
+                    "cy_name": self.to_chinese(last.get("customer_name", "") or cid),
+                    "acc_type": str(last.get("platform", "") or "AliWorkbench"),
+                    "msg": str(last.get("message", "") or ""),
+                    "msg_type": 0,
+                }
+                await self.send_reply(fake, followup)
+                self._unreplied_followup_sent[ckey] = now_mono
+                sent += 1
+                self._activity_log(
+                    "unreplied_followup_sent",
+                    acc_id=acc_id,
+                    customer_id=cid,
+                    idle_seconds=int(idle_s),
+                    last_msg=str(last.get("message", "") or "")[:120],
+                    reply=followup,
+                )
+        finally:
+            try:
+                if conn:
+                    conn.close()
+            except Exception:
+                pass
+
+    async def _compose_ai_scene_reply(
+        self,
+        *,
+        original_msg: dict,
+        scene: str,
+        intent_hint: str,
+        fallback: str,
+    ) -> str:
+        """
+        场景化 AI 直接生成回复（不依赖固定模板）。
+        """
+        if not self.enable_agent or not self.agent or not AgentDeps:
+            return fallback
+        try:
+            deps = AgentDeps(
+                msg_id=str(original_msg.get("msg_id", "") or f"{scene}_gen"),
+                acc_id=str(original_msg.get("acc_id", "") or ""),
+                from_id=str(original_msg.get("from_id", "") or ""),
+                platform=str(original_msg.get("acc_type", "") or ""),
+            )
+            customer_msg = self.to_chinese(str(original_msg.get("msg", "") or ""))
+            prompt = (
+                "你是淘宝客服，直接生成一条发给客户的话。\n"
+                f"场景: {scene}\n"
+                f"意图: {intent_hint}\n"
+                f"客户原话: {customer_msg}\n"
+                "要求: 1-2句，自然口语，不要模板腔，不要新增价格/承诺；只输出最终回复。\n"
+            )
+            result = await self.agent.agent_natural_reply.run(prompt, deps=deps, message_history=[])
+            out = str(getattr(result, "output", "") or "").strip()
+            if not out:
+                return fallback
+            if out.startswith("话术|") or "[转移会话]" in out or "TRANSFER_REQUESTED" in out:
+                return fallback
+            self._activity_log(
+                "ai_scene_reply_generated",
+                acc_id=str(original_msg.get("acc_id", "") or ""),
+                customer_id=str(original_msg.get("from_id", "") or ""),
+                scene=scene,
+                generated=out[:160],
+            )
+            return out
+        except Exception as e:
+            self._activity_log(
+                "ai_scene_reply_error",
+                acc_id=str(original_msg.get("acc_id", "") or ""),
+                customer_id=str(original_msg.get("from_id", "") or ""),
+                scene=scene,
+                error=str(e),
+            )
+            return fallback
+
     async def receive_messages(self):
         """持续接收消息"""
         try:
@@ -1296,7 +1517,7 @@ class QingjianAPIClient:
                         return
                     # 标记本批次已自动触发，避免同内容循环“马上报价”。
                     self._auto_quote_done_sig[capture_key] = capture_sig
-                    # 直接置为可报价，然后走“发完了，报价吧”触发既有报价链路
+                    # 直接置为可报价，走内部自动报价入口（不伪造客户语句）。
                     self.agent._mark_quote_ready(st)
                     self.agent._sync_pending_quote_state(capture_cid, st)
                     self._activity_log(
@@ -1308,7 +1529,7 @@ class QingjianAPIClient:
                     notify_msg = CustomerMessage(
                         msg_id="auto_quote_idle_trigger",
                         acc_id=capture_data.get('acc_id', ''),
-                        msg="发完了，报价吧",
+                        msg="__AUTO_QUOTE_INTERNAL_TRIGGER__",
                         from_id=capture_cid,
                         from_name=self.to_chinese(capture_data.get('from_name', '') or capture_data.get('cy_name', '')),
                         cy_id=capture_data.get('cy_id', ''),
@@ -1318,7 +1539,7 @@ class QingjianAPIClient:
                         goods_name=self.to_chinese(capture_data.get('goods_name', '')) if capture_data.get('goods_name') else None,
                         goods_order=self.to_chinese(capture_data.get('goods_order', '')) if capture_data.get('goods_order') else None,
                     )
-                    response = await self.agent.process_message(notify_msg)
+                    response = await self.agent.build_auto_quote_reply(st, notify_msg)
                     if response.should_reply and response.reply and not response.need_transfer:
                         await self.send_reply(capture_data, response.reply)
                         self._activity_log(
@@ -1636,7 +1857,12 @@ class QingjianAPIClient:
         logger.info(f"系统询单命中 | 店铺:{acc_id} | 客户:{customer_id} | action:{action}")
 
         if action == "reply":
-            reply = policy.get("reply") or "您好，这边已收到询单消息，稍后由人工客服跟进处理。"
+            reply = await self._compose_ai_scene_reply(
+                original_msg=data,
+                scene="system_inquiry_reply",
+                intent_hint="这是系统客服询单消息，简短确认已收到并说明会跟进即可。",
+                fallback=(policy.get("reply") or "您好，这边已收到询单消息，稍后由人工客服跟进处理。"),
+            )
             await self.send_reply(data, reply)
             metrics_emit("system_inquiry_auto_reply", customer_id=customer_id, acc_id=acc_id)
             return True
@@ -1976,6 +2202,10 @@ class QingjianAPIClient:
             return
 
         reply_content = self._colloquialize_outbound_reply(reply_content)
+        reply_content = await self._ai_generate_outbound_reply(
+            original_msg=original_msg,
+            reply_content=str(reply_content or ""),
+        )
 
         # 同一客户外发限流：N 秒内最多 1 条
         try:
@@ -2058,6 +2288,62 @@ class QingjianAPIClient:
         reply["_trace_id"] = trace_id
         await self.send_message(reply)
 
+    async def _ai_generate_outbound_reply(self, original_msg: dict, reply_content: str) -> str:
+        """
+        强制全量 AI 出站生成层：
+        - 所有普通文本外发先由 AI 生成最终话术；
+        - 控制命令/纯链接/转接指令直接绕过。
+        """
+        text = (reply_content or "").strip()
+        if not text:
+            return text
+        if text.startswith("话术|") or "[转移会话]" in text or "TRANSFER_REQUESTED" in text:
+            return text
+        if re.fullmatch(r"https?://\S+", text):
+            return text
+        if not self._force_ai_generate_reply or not self.enable_agent or not self.agent or not AgentDeps:
+            return text
+        try:
+            deps = AgentDeps(
+                msg_id=str(original_msg.get("msg_id", "") or "outbound_generate"),
+                acc_id=str(original_msg.get("acc_id", "") or ""),
+                from_id=str(original_msg.get("from_id", "") or ""),
+                platform=str(original_msg.get("acc_type", "") or ""),
+            )
+            customer_msg = self.to_chinese(str(original_msg.get("msg", "") or ""))
+            prompt = (
+                "你是淘宝客服外发文案生成器。请根据“回复意图草稿”生成最终发给客户的话。\n"
+                "要求：\n"
+                "1) 保留原意，不新增价格/承诺/流程；\n"
+                "2) 自然像真人聊天，不用固定模板句；\n"
+                "3) 1-2句；\n"
+                "4) 只输出最终回复文本。\n\n"
+                f"客户原话: {customer_msg}\n"
+                f"回复意图草稿: {text}\n"
+            )
+            result = await self.agent.agent_natural_reply.run(prompt, deps=deps, message_history=[])
+            out = str(getattr(result, "output", "") or "").strip()
+            if not out:
+                return text
+            if out.startswith("话术|") or "[转移会话]" in out:
+                return text
+            self._activity_log(
+                "ai_generate_reply",
+                acc_id=str(original_msg.get("acc_id", "") or ""),
+                customer_id=str(original_msg.get("from_id", "") or ""),
+                draft=text[:160],
+                generated=out[:160],
+            )
+            return out
+        except Exception as e:
+            self._activity_log(
+                "ai_generate_reply_error",
+                acc_id=str(original_msg.get("acc_id", "") or ""),
+                customer_id=str(original_msg.get("from_id", "") or ""),
+                error=str(e),
+            )
+            return text
+
     def _colloquialize_outbound_reply(self, text: Any) -> Any:
         """统一外发口语化处理，避免机械话术。"""
         if not isinstance(text, str):
@@ -2393,6 +2679,11 @@ class QingjianAPIClient:
         except Exception as e:
             logger.info(f"[{self.get_time()}] 企微启动消息模块加载失败: {e}")
 
+        # 未回复会话补偿（可关闭）
+        if os.getenv("UNREPLIED_FOLLOWUP_ENABLED", "true").lower() in ("1", "true", "yes"):
+            tasks.append(self._unreplied_followup_loop())
+            logger.info(f"[{self.get_time()}] 未回复会话补偿任务已启动")
+
         await asyncio.gather(*tasks)