perf: fast-route orchestration and short-reply guard for qingjian

qingjian_cs/app/agents.py

@@ -15,7 +15,7 @@ from .agent_tools import (
     tool_extract_size_pairs,
     tool_is_meaningless_short,
 )
-from .config import OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_MODEL_NAME
+from .config import AGENT_MAX_ITERS, OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_MODEL_NAME
 from .models import Decision, DecisionModel, RouteModel
 from .rules import rules_prompt
 
@@ -66,7 +66,7 @@ class _AgentRuntime:
             formatter=OpenAIChatFormatter(),
             toolkit=toolkit,
             memory=InMemoryMemory(),
-            max_iters=8,
+            max_iters=max(1, AGENT_MAX_ITERS),
         )
 
     @staticmethod

qingjian_cs/app/client.py

@@ -1,12 +1,18 @@
 import asyncio
 import json
+import re
 import time
 from collections import defaultdict
 
 import websockets
 
 from .callbacks import post_tianwang_callback
-from .config import AUTO_QUOTE_WAIT_SECONDS, MESSAGE_DEBOUNCE_SECONDS, QINGJIAN_WS_URI
+from .config import (
+    AUTO_QUOTE_WAIT_SECONDS,
+    MESSAGE_DEBOUNCE_SECONDS,
+    QINGJIAN_WS_URI,
+    SHORT_REPLY_MAX_CHARS,
+)
 from .logger import setup_logger
 from .observability import activity_event, build_trace_id
 from .orchestrator import Orchestrator
@@ -27,7 +33,7 @@ class QingjianClient:
         self.pending_images: dict[str, list[str]] = defaultdict(list)
         self.auto_quote_tasks: dict[str, asyncio.Task] = {}
         self.last_reply_key: dict[str, str] = {}
-        self.recent_outbound: dict[str, tuple[str, float]] = {}
+        self.recent_outbound: list[tuple[str, str, str, float]] = []
 
     @staticmethod
     def _customer_key(data: dict) -> str:
@@ -57,6 +63,7 @@ class QingjianClient:
         text = str(text or "").strip()
         if not text:
             return
+        text = self._shorten_reply(text)
         msg = {
             "msg_id": "",
             "acc_id": data.get("acc_id", ""),
@@ -70,22 +77,47 @@ class QingjianClient:
         }
         activity_event(self.logger, "send_reply_attempt", trace_id=trace_id, customer_id=data.get("from_id", "-"), msg=text)
         await self.send_message(msg)
-        self.recent_outbound[self._customer_key(data)] = (text, time.monotonic())
+        self.recent_outbound.append((str(data.get("acc_id", "")), str(data.get("from_id", "")), text, time.monotonic()))
+        if len(self.recent_outbound) > 200:
+            self.recent_outbound = self.recent_outbound[-200:]
         activity_event(self.logger, "send_reply_success", trace_id=trace_id, customer_id=data.get("from_id", "-"), msg=text)
 
+    @staticmethod
+    def _clean_text(text: str) -> str:
+        t = str(text or "").strip()
+        t = re.sub(r"\s+", "", t)
+        return t
+
+    def _shorten_reply(self, text: str) -> str:
+        max_len = max(8, int(SHORT_REPLY_MAX_CHARS))
+        t = str(text or "").strip()
+        if len(t) <= max_len:
+            return t
+        parts = re.split(r"[。！？!?]", t)
+        head = next((p.strip() for p in parts if p and p.strip()), t)
+        if len(head) > max_len:
+            head = head[:max_len].rstrip("，,;；:： ")
+        return head or t[:max_len]
+
     def _is_outbound_echo(self, data: dict, msg: str) -> bool:
         """
         轻简可能会把我方刚发送文本回推为“收到消息”。
-        对同 customer_key 的“短时间完全相同文本”做回环拦截，避免无限对话。
+        对“短时间完全相同文本”做回环拦截，兼容 acc/from 对调回推，避免无限对话。
         """
-        key = self._customer_key(data)
+        in_acc = str(data.get("acc_id", ""))
-        last = self.recent_outbound.get(key)
+        in_from = str(data.get("from_id", ""))
-        if not last:
+        in_msg = self._clean_text(msg)
+        now = time.monotonic()
+        if not in_msg:
             return False
-        last_msg, ts = last
+        for out_acc, out_to, out_msg, ts in reversed(self.recent_outbound):
-        if (time.monotonic() - ts) > 120:
+            if (now - ts) > 120:
-            return False
+                break
-        return str(msg or "").strip() == last_msg
+            if self._clean_text(out_msg) != in_msg:
+                continue
+            if (out_acc == in_acc and out_to == in_from) or (out_acc == in_from and out_to == in_acc):
+                return True
+        return False
 
     async def _handle_decision(self, data: dict, merged_msg: str, *, auto_quote: bool = False) -> None:
         key = self._customer_key(data)

qingjian_cs/app/config.py

@@ -14,6 +14,9 @@ OPENAI_MODEL_NAME = os.getenv("OPENAI_MODEL_NAME", "doubao-seed-2-0-pro-260215")
 
 MESSAGE_DEBOUNCE_SECONDS = int(os.getenv("MESSAGE_DEBOUNCE_SECONDS", "6"))
 AUTO_QUOTE_WAIT_SECONDS = int(os.getenv("AUTO_QUOTE_WAIT_SECONDS", "18"))
+AGENT_MAX_ITERS = int(os.getenv("AGENT_MAX_ITERS", "3"))
+FAST_ROUTE_ENABLED = os.getenv("FAST_ROUTE_ENABLED", "1").strip() in {"1", "true", "True", "yes", "on"}
+SHORT_REPLY_MAX_CHARS = int(os.getenv("SHORT_REPLY_MAX_CHARS", "28"))
 
 STORE_BACKEND = os.getenv("STORE_BACKEND", "sqlite").strip().lower()
 STORE_SQLITE_PATH = os.getenv("STORE_SQLITE_PATH", "").strip()

qingjian_cs/app/logger.py

@@ -16,5 +16,7 @@ def setup_logger() -> logging.Logger:
     logging.getLogger("agentscope").setLevel(logging.ERROR)
     logging.getLogger("agentscope.formatter").setLevel(logging.ERROR)
     logging.getLogger("agentscope.agent").setLevel(logging.ERROR)
+    logging.getLogger("_openai_formatter").setLevel(logging.ERROR)
+    logging.getLogger("_react_agent").setLevel(logging.ERROR)
 
     return logger

qingjian_cs/app/orchestrator.py

@@ -3,8 +3,15 @@ from __future__ import annotations
 from typing import Any
 
 from .agents import AfterSalesAgent, PreSalesAgent, QuoteAgent, RiskAgent, RouterAgent
+from .config import FAST_ROUTE_ENABLED
 from .models import Decision
-from .rules import detect_intent, detect_order_status
+from .rules import (
+    detect_intent,
+    detect_order_status,
+    has_map_or_political_risk,
+    has_porn_risk,
+    requests_external_contact,
+)
 from .state_machine import evolve_after_sales_state, migrate_state_schema
 from .store import ConversationStore
 
@@ -35,9 +42,13 @@ class Orchestrator:
             "order_status": order_status,
         }
 
-        # 先风控
+        msg = str(context.get("msg", "") or "")
-        risk_decision = await self.risk.decide(merged_ctx)
+        goods_name = str(context.get("goods_name", "") or "")
-        if risk_decision.action in {"transfer"}:
+        risk_hit = has_map_or_political_risk(msg, goods_name) or has_porn_risk(msg) or requests_external_contact(msg)
+
+        # 命中硬风控才调用 RiskAgent，避免每条消息都先走一轮模型。
+        if risk_hit:
+            risk_decision = await self.risk.decide(merged_ctx)
             route = "risk"
             new_state = evolve_after_sales_state(
                 {**prev_state, **(risk_decision.state_patch or {})},
@@ -51,7 +62,23 @@ class Orchestrator:
             self.store.append_event(customer_key, "decision", {"route": route, "action": risk_decision.action, "reason": risk_decision.reason})
             return route, risk_decision, new_state
 
-        route, route_reason = await self.router.route(merged_ctx)
+        route = ""
+        route_reason = ""
+        if FAST_ROUTE_ENABLED:
+            pending_images = int(context.get("pending_images", 0) or 0)
+            auto_quote_trigger = bool(context.get("auto_quote_trigger", False))
+            if intent in {"pricing", "finish_or_quote_trigger"} and (pending_images > 0 or auto_quote_trigger):
+                route = "quote"
+                route_reason = "fast_route_quote_with_pending_images"
+            elif order_status == "refund":
+                route = "after_sales"
+                route_reason = "fast_route_refund"
+            elif intent in {"image", "greeting", "nonsense", "pricing", "finish_or_quote_trigger", "unknown"}:
+                route = "pre_sales"
+                route_reason = "fast_route_common_presales"
+
+        if not route:
+            route, route_reason = await self.router.route(merged_ctx)
 
         if route == "quote":
             decision = await self.quote.decide(merged_ctx)

qingjian_cs/app/rules.py

@@ -127,6 +127,7 @@ def rules_prompt() -> str:
         "10) 尺寸明显超大(如>=2m*2m): 提示需补图/重做边缘, 不要直接承诺一模一样。\n"
         "11) 店铺差异化: 按 acc_id/persona 口吻回复, 保持真人聊天。\n"
         "12) 最终输出只允许一个动作, 不能混合。\n"
+        "13) reply 必须简短: 优先 1 句, 一般不超过 28 个汉字, 禁止长段解释。\n"
         "输出格式:\n"
         '{"action":"reply|quote|transfer|noop","reply":"","transfer_msg":"","quote_mode":"flush_pending|analyze_current_or_recent|collect_only","reason":""}'
     )