v1.7: session-id mapping + conversation summary + contact profiling + exponential backoff reconnect

CHANGELOG.md

@@ -1,5 +1,32 @@
 # Changelog
 
+## v1.7 (2026-06-05) — 对话上下文 + 客户画像 + 指数退避重连
+
+### 新增
+- **`--session-id` 对话上下文** — WS Agent 维护 `conv_id → session_id` 映射，每次调 `qwenpaw agents chat` 时传入 `--session-id`，AI 获得完整对话历史
+  - 同一会话的连续消息不再断开上下文，AI 知道"刚才说过什么"
+  - 持久化到状态文件，重启不丢失
+  - 自动清理超过 10000 条的大映射表
+- **对话摘要** — 每 15 轮 AI 回复后自动调用 AI 压缩历史，生成 1-2 句话摘要
+  - 下次请求时将摘要注入 prompt 开头，减少 token 消耗
+  - 长对话 AI 仍能记住核心信息（客户需求、讨论过的产品）
+- **客户画像** — 维护 `contact_id → 画像` 映射（姓名、最近 3 次交互记录）
+  - 每次 AI 回复后自动更新画像
+  - 下次对话自动注入客户历史上下文
+  - 持久化到状态文件，重启不丢失
+- **WebSocket 指数退避重连** — 断线重连从固定 5s 改为指数退避
+  - 初始 5s → 10s → 20s → 40s → 最大 60s
+  - 重连前自动续期 Chatwoot session（防止长时间断线后 token 过期）
+  - 失败时继续尝试，不会停止
+
+### 文件
+- `chatwoot_ws_agent.py` 从 1294 行增至 1459 行（+165 行）
+- 新增 9 个函数：`_get_or_create_session`、`_prune_sessions`、`_summarize_conversation`、`_get_conversation_context`、`_update_contact_profile`、`_get_contact_context`
+- 修改 4 个函数：`call_qwenpaw_ai`（+session_id）、`generate_ai_reply`（+session_id）、`handle_incoming_message`（+3 层 context）、`save_state/load_state`（+3 个持久化字段）
+- 重写 1 个函数：`WSAgent.start`（指数退避重连）
+
+---
+
 ## v1.6 (2026-06-05) — Platform Gateway + 5 平台 API 集成
 
 ### 新增

README.md

@@ -241,6 +241,7 @@ chatwoot-ai-agent/
 | v1.4 | 多租户架构，Provision Server，状态持久化，安全性重构 |
 | v1.5 | 消息防抖（5s 累积合并），AI 错误重试（指数退避）|
 | v1.6 | Platform Gateway 库——Amazon/JD/Taobao/PDD/TikTok 5 平台统一 API 集成 |
+| v1.7 | 对话上下文（`--session-id`）+ 对话摘要 + 客户画像 + WebSocket 指数退避重连 |
 
 ## 许可证
 

chatwoot_ws_agent.py

@@ -484,6 +484,16 @@ def save_state():
                 "ai_pending_convs": list(_ai_pending_convs),
                 "saved_at": time.time()
             }
+        # Add session map
+        with _conv_session_lock:
+            state["conv_session_map"] = _conv_session_map.copy()
+        # Add summaries & rounds
+        with _conv_conv_lock:
+            state["conv_summaries"] = _conv_summaries.copy()
+            state["conv_rounds"] = _conv_rounds.copy()
+        # Add contact profiles
+        with _contact_profile_lock:
+            state["contact_profiles"] = _contact_profiles.copy()
         STATE_FILE.write_text(json.dumps(state, ensure_ascii=False), encoding="utf-8")
     except Exception as e:
         log(f"save_state error: {e}", "WARN")
@@ -511,8 +521,20 @@ def load_state():
             human_active_convs = state.get("human_active_convs", {})
         with _ai_pending_lock:
             _ai_pending_convs = set(state.get("ai_pending_convs", []))
+        # Restore session map
+        with _conv_session_lock:
+            _conv_session_map.update(state.get("conv_session_map", {}))
+        # Restore summaries & rounds
+        with _conv_conv_lock:
+            _conv_summaries.update(state.get("conv_summaries", {}))
+            _conv_rounds.update(state.get("conv_rounds", {}))
+        # Restore contact profiles
+        with _contact_profile_lock:
+            _contact_profiles.update(state.get("contact_profiles", {}))
 
-        log(f"State restored: {len(ai_sent_msg_ids)} msg_ids, {len(human_active_convs)} active convs, {len(_ai_pending_convs)} pending")
+        log(f"State restored: {len(ai_sent_msg_ids)} msg_ids, {len(human_active_convs)} active convs, "
+            f"{len(_conv_session_map)} sessions, {len(_conv_summaries)} summaries, "
+            f"{len(_contact_profiles)} profiles")
     except Exception as e:
         log(f"load_state error: {e} (starting fresh)", "WARN")
         STATE_FILE.unlink(missing_ok=True)
@@ -527,13 +549,105 @@ _debounce_msgs: dict[int, list[str]] = {}  # conv_id → accumulated messages
 _debounce_lock = Lock()
 
 
+# ===== SESSION-ID MAPPING (per-conversation QwenPaw context) =====
+_conv_session_map: dict[int, str] = {}  # conv_id → qwenpaw session_id
+_conv_session_lock = Lock()
+
+
+def _get_or_create_session(conv_id):
+    """Return existing session_id for a conversation, or create a new one."""
+    with _conv_session_lock:
+        sid = _conv_session_map.get(conv_id)
+        if not sid:
+            sid = f"conv_{conv_id}_{int(time.time())}"
+            _conv_session_map[conv_id] = sid
+        return sid
+
+
+def _prune_sessions():
+    """Trim session map when it grows too large."""
+    with _conv_session_lock:
+        if len(_conv_session_map) > 10000:
+            _conv_session_map.clear()
+            log("🧹 Session map too large, cleared", "WARN")
+
+
+# ===== CONVERSATION SUMMARIZATION =====
+_conv_rounds: dict[int, int] = {}  # conv_id → AI reply count
+_conv_summaries: dict[int, str] = {}  # conv_id → latest AI-generated summary
+_conv_conv_lock = Lock()
+SUMMARY_THRESHOLD = 15  # generate summary after every N AI replies
+
+
+def _summarize_conversation(conv_id, prompt, reply, target_agent):
+    """Ask AI to produce a short summary of this conversation so far."""
+    with _conv_conv_lock:
+        old_summary = _conv_summaries.get(conv_id, "")
+    summary_prompt = (
+        f"Existing summary: {old_summary}\n\n"
+        f"Latest customer message: {prompt[:200]}\n"
+        f"Your reply: {reply[:500]}\n\n"
+        "Produce a concise 1-2 sentence summary (in English) of the "
+        "entire conversation so far. Focus on customer needs, products "
+        "discussed, and any decisions made. Keep it under 100 words."
+    )
+    summary = call_qwenpaw_ai(summary_prompt, target_agent=target_agent)
+    if summary:
+        with _conv_conv_lock:
+            _conv_summaries[conv_id] = summary
+
+
+def _get_conversation_context(conv_id):
+    """Return summary string for this conv, or empty string."""
+    with _conv_conv_lock:
+        return _conv_summaries.get(conv_id, "")
+
+
+# ===== CONTACT PROFILING =====
+_contact_profiles: dict[int, dict] = {}  # contact_id → profile dict
+_contact_profile_lock = Lock()
+
+
+def _update_contact_profile(contact_id, sender_name, msg, reply):
+    """Extract relevant info from conversation and update contact profile."""
+    if not contact_id:
+        return
+    with _contact_profile_lock:
+        profile = _contact_profiles.get(contact_id, {})
+        profile["name"] = sender_name
+        profile["last_seen"] = time.time()
+        # Keep a rolling log of last 3 interactions
+        interactions = profile.get("interactions", [])
+        interactions.append({"msg": msg[:100], "reply": reply[:100], "ts": time.time()})
+        profile["interactions"] = interactions[-3:]
+        _contact_profiles[contact_id] = profile
+
+
+def _get_contact_context(contact_id):
+    """Return a prompt-friendly profile string for this contact."""
+    if not contact_id:
+        return ""
+    with _contact_profile_lock:
+        profile = _contact_profiles.get(contact_id)
+    if not profile:
+        return ""
+    parts = [f"Contact name: {profile.get('name', 'Unknown')}"]
+    interactions = profile.get("interactions", [])
+    if interactions:
+        parts.append(f"Last {len(interactions)} interaction(s):")
+        for i, ix in enumerate(interactions):
+            parts.append(f"  [{i+1}] \"{ix['msg']}\" → \"{ix['reply'][:80]}\"")
+    return "\n".join(parts)
+
+
 # ===== AI HELPERS =====
 
 import re as _re  # used for session-id stripping
 
 GATEWAY_ENABLED = os.environ.get("GATEWAY_ENABLED", "1") == "1"
 
-def call_qwenpaw_ai(prompt, system_prompt=None, target_agent="sourcing-agent", retries=2):
+def call_qwenpaw_ai(prompt, system_prompt=None, target_agent="sourcing-agent",
+                    retries=2, session_id=None):
     """Call QwenPaw AI via agents chat, with retries on failure.
 
     Args:
@@ -541,18 +655,19 @@ def call_qwenpaw_ai(prompt, system_prompt=None, target_agent="sourcing-agent", r
         system_prompt: Optional system instructions (appended before prompt)
         target_agent: Which agent to call (default:sourcing-agent, translation:default)
         retries: Number of retries on failure (exponential backoff)
+        session_id: Optional QwenPaw session ID for conversation continuity
     """
     full_prompt = (system_prompt + "\n\n" + prompt) if system_prompt else prompt
+    cmd = ["qwenpaw", "agents", "chat",
+           "--from-agent", "wordpress",
+           "--to-agent", target_agent,
+           "--text", full_prompt]
+    if session_id:
+        cmd.extend(["--session-id", session_id])
     last_error = None
     for attempt in range(1 + retries):
         try:
-            result = subprocess.run(
+            result = subprocess.run(cmd, capture_output=True, text=True, timeout=90)
-                ["qwenpaw", "agents", "chat",
-                 "--from-agent", "wordpress",
-                 "--to-agent", target_agent,
-                 "--text", full_prompt],
-                capture_output=True, text=True, timeout=90
-            )
             if result.returncode == 0 and result.stdout.strip():
                 # Clean output: remove INFO/SESSION/Agent lines
                 raw = result.stdout.strip()
@@ -594,7 +709,7 @@ def translate_to_chinese(text):
     result = call_qwenpaw_ai(prompt, target_agent="default")
     return result if result else text
 
-def generate_ai_reply(customer_msg, sender_name, inbox_id, context=""):
+def generate_ai_reply(customer_msg, sender_name, inbox_id, context="", session_id=None):
     config = INBOX_CONFIG.get(inbox_id)
     if not config:
         log(f"No config for inbox #{inbox_id}, skipping", "WARN")
@@ -606,7 +721,8 @@ def generate_ai_reply(customer_msg, sender_name, inbox_id, context=""):
         customer_msg=body
     )
     reply = call_qwenpaw_ai(prompt, config["system_prompt"],
-                            target_agent=config["target_agent"])
+                            target_agent=config["target_agent"],
+                            session_id=session_id)
 
     if not reply:
         return reply
@@ -826,11 +942,30 @@ def handle_incoming_message(msg_data, headers):
 
     log(f"📩 [{config['name']}] New msg #{msg_id} from '{sender_name}' in conv #{conv_id}: {content[:100]}", inbox_name=config['name'])
 
+    # ===== SESSION-ID: maintain conversation continuity =====
+    session_id = _get_or_create_session(conv_id)
+    # ===== CONVERSATION SUMMARY: inject past context =====
+    conv_context = _get_conversation_context(conv_id)
+    # ===== CONTACT PROFILE: inject customer history =====
+    contact_id = msg_data.get("sender", {}).get("id")
+    contact_context = _get_contact_context(contact_id)
+
+    # Build enriched context
+    context_parts = []
+    gateway_ctx = _enrich_context(inbox_id, content, config)
+    if gateway_ctx:
+        context_parts.append(gateway_ctx)
+    if conv_context:
+        context_parts.append(f"[Conversation so far: {conv_context}]")
+    if contact_context:
+        context_parts.append(f"[Customer history: {contact_context}]")
+    all_context = "\n".join(context_parts)
+
     # Generate AI reply with inbox-specific config
     log(f"🤖 [{config['name']}] Generating AI reply...", inbox_name=config['name'])
     start_time = time.time()
-    context = _enrich_context(inbox_id, content, config)
+    reply = generate_ai_reply(content, sender_name, inbox_id,
-    reply = generate_ai_reply(content, sender_name, inbox_id, context=context)
+                              context=all_context, session_id=session_id)
     duration_ms = (time.time() - start_time) * 1000
 
     if not reply:
@@ -840,6 +975,22 @@ def handle_incoming_message(msg_data, headers):
 
     log(f"💬 [{config['name']}] AI reply ({duration_ms:.0f}ms): {reply[:150]}", inbox_name=config['name'])
 
+    # ===== POST-REPLY: update conversation summary & contact profile =====
+    with _conv_conv_lock:
+        rounds = _conv_rounds.get(conv_id, 0) + 1
+        _conv_rounds[conv_id] = rounds
+    if rounds > 0 and rounds % SUMMARY_THRESHOLD == 0:
+        # Generate summary every SUMMARY_THRESHOLD AI replies
+        _summarize_conversation(conv_id, content, reply, config["target_agent"])
+    _update_contact_profile(contact_id, sender_name, content, reply)
+    
+    if not reply:
+        log(f"⚠️ [{config['name']}] Empty AI reply, skipping", "WARN", config['name'])
+        metrics.record_reply(inbox_id, config['name'], False, duration_ms)
+        return
+
+    log(f"💬 [{config['name']}] AI reply ({duration_ms:.0f}ms): {reply[:150]}", inbox_name=config['name'])
+
     # Check if AI thinks this needs human handoff
     needs_handoff = False
     if "[HANDOFF]" in reply:
@@ -1021,10 +1172,25 @@ class WSAgent:
         metrics.ws_disconnected(f"status={status}, msg={msg}")
 
     def start(self):
-        """Start WebSocket connection (blocking)."""
+        """Start WebSocket connection with exponential backoff reconnect."""
         self.headers = ensure_session()
-        log(f"Connecting to {CW_WS_URL}")
+        delay = 5  # start at 5s
+        max_delay = 60
+        attempt = 0
 
+        while self.running.is_set():
+            if attempt > 0:
+                log(f"🔄 Reconnecting in {delay}s (attempt #{attempt})...")
+                time.sleep(delay)
+                delay = min(delay * 2, max_delay)  # 5 → 10 → 20 → 40 → 60
+                # Re-fetch session headers in case they expired during downtime
+                try:
+                    self.headers = ensure_session()
+                except Exception as e:
+                    log(f"Session renewal failed before reconnect: {e}", "WARN")
+                    continue
+
+            log(f"Connecting to {CW_WS_URL}")
             self.ws = websocket.WebSocketApp(
                 CW_WS_URL,
                 on_open=self.on_open,
@@ -1032,14 +1198,13 @@ class WSAgent:
                 on_error=self.on_error,
                 on_close=self.on_close
             )
-
-        # Run forever (blocks)
             self.ws.run_forever(
                 sslopt={"cert_reqs": ssl.CERT_NONE},
                 ping_interval=15,
                 ping_timeout=5,
-            reconnect=5  # built-in reconnect
+                reconnect=0  # disable internal reconnect; we handle it
             )
+            attempt += 1
 
     def stop(self):
         """Stop the WebSocket connection."""