feat(telegram): ADR-019 Phase 4 - conversation memory + chat_id propagation

ADR-019 Phase 4：新增 services/openclaw_session.py 管理 chat_id 級別的多輪
對話歷史與 carry-over slot。In-memory，30 分鐘 TTL，重啟清空（臨時對話 state
不該污染 ai_insights 永久記憶）。

openclaw_answer 簽章加 chat_id=None 可選參數：
- 傳入時 agent 會看到該 chat 最近 5 輪對話歷史，注入 system prompt
- Ollama / Gemini FC 兩條路徑都會在生成成功後 append_turn 寫回 session
- system prompt 加決策規則：「若歷史顯示用戶剛被問參數，優先用該答案接續執行」

Caller 全部更新傳 chat_id：
- routes/openclaw_bot_routes.py:5479 (handle_cmd 不認識指令 fallback)
- routes/openclaw_bot_routes.py:5916 (webhook NL 路徑)
- routes/openclaw_bot_routes.py:_agent_dispatch_cmd (Phase 3 hook)
- services/telegram_bot_service.py:934 (polling NL fallback)

向下相容：chat_id=None 時行為與舊版完全相同（無 multi-turn 記憶）。

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

routes/openclaw_bot_routes.py

@@ -4303,13 +4303,20 @@ def _execute_tool(name: str, args: dict) -> dict:
     return {"error": f"unknown tool: {name}"}
 
 
-def openclaw_answer(question: str):
+def openclaw_answer(question: str, chat_id: int = None):
     """
     Function Calling 架構 — AI 自主決定查什麼、怎麼回答
     不再靠 if/else 規則判斷意圖
+
+    ADR-019 Phase 4：可選 chat_id 啟用對話 state。傳入後 agent 會看到該 chat 的
+    最近對話歷史，並把本輪 (question, answer) 寫回 session 供下輪使用。
+    chat_id 為 None 時行為與舊版完全相同（無 multi-turn 記憶）。
     """
+    from services import openclaw_session
+
     now       = datetime.now(TAIPEI_TZ)
     today_str = now.strftime("%Y/%m/%d")
+    history_ctx = openclaw_session.history_as_prompt(chat_id) if chat_id else ""
 
     # ── 功能說明直接導 help ───────────────────────────────────
     if _is_help_question(question):
@@ -4363,13 +4370,16 @@ def openclaw_answer(question: str):
 
             sys_prompt = (
                 f"你是 OpenClaw（小O），電商智能助理。今天{today}。\n"
+                + (f"【最近對話】\n{history_ctx}\n\n" if history_ctx else "")
                 + (f"【業績資料】{db_ctx}\n" if db_ctx else "")
                 + (f"【市場情報】{mcp_ctx[:400]}\n" if mcp_ctx else "")
                 + "請用繁體中文直接回答，不要開場白，300字以內。"
             )
-            
+
             resp = ollama_service.generate(question, system_prompt=sys_prompt, timeout=180)
             if resp.success and resp.content:
+                if chat_id:
+                    openclaw_session.append_turn(chat_id, question, resp.content)
                 if _LEARNING_ENABLED:
                     import threading as _thr
                     _thr.Thread(target=store_conversation,
@@ -4390,7 +4400,8 @@ def openclaw_answer(question: str):
             sys_msg = (
                 f"你是 OpenClaw（小O），服務「小龍蝦」電商業務團隊的 AI 助理。\n"
                 f"今天是 {today_str}。\n"
-                "你有四個工具可以使用：\n"
+                + (f"\n【最近對話歷史】\n{history_ctx}\n" if history_ctx else "")
+                + "你有四個工具可以使用：\n"
                 "1. query_sales — 查自家業績資料庫\n"
                 "2. get_market_intel — 取得外部市場情報（新聞/熱搜/PTT口碑/匯率/天氣/節慶）\n"
                 "3. get_knowledge — 查歷史分析知識庫\n"
@@ -4400,7 +4411,8 @@ def openclaw_answer(question: str):
                 "- 若 data_freshness_warning 顯示當月無資料，禁止編造『業績為零』，"
                 "改主動問用戶是否要改看上月（並附 latest_data_date）\n"
                 "- 工具結果回來後用繁體中文自然回答，不要開場白，不要多餘客套話\n"
-                "- 同時呼叫多個工具時請平行送出"
+                "- 同時呼叫多個工具時請平行送出\n"
+                "- 若【最近對話歷史】顯示用戶剛被詢問某參數（如月份），優先用該答案接續執行"
             )
 
             # 第一輪：讓 Gemini 判斷需要呼叫哪些工具
@@ -4428,6 +4440,8 @@ def openclaw_answer(question: str):
             if not tool_calls:
                 text = "".join(p.get("text", "") for p in parts if "text" in p).strip()
                 if text:
+                    if chat_id:
+                        openclaw_session.append_turn(chat_id, question, text)
                     if _LEARNING_ENABLED:
                         import threading as _thr
                         _thr.Thread(target=store_conversation,
@@ -4476,6 +4490,8 @@ def openclaw_answer(question: str):
 
             if final:
                 sys_log.info(f"[FC] done tools={used_sources} reply={len(final)}chars")
+                if chat_id:
+                    openclaw_session.append_turn(chat_id, question, final)
                 if _LEARNING_ENABLED:
                     import threading as _thr
                     _thr.Thread(target=store_conversation,
@@ -4568,7 +4584,7 @@ def _agent_dispatch_cmd(cmd, arg, chat_id, reply_to) -> bool:
     nl_question = _CMD_TO_NL[cmd](arg)
     try:
         sys_log.info(f"[AgentDispatch] cmd:{cmd}:{arg or ''} → NL: {nl_question}")
-        txt, kb = openclaw_answer(nl_question)
+        txt, kb = openclaw_answer(nl_question, chat_id=chat_id)
         send_message(chat_id, txt, reply_to, keyboard=kb)
         return True
     except Exception as e:
@@ -5460,7 +5476,7 @@ def handle_cmd(cmd, arg, chat_id, reply_to):
 
     else:
         # 不認識的指令 → 自然語言
-        txt, kb = openclaw_answer(cmd + (' ' + arg if arg else ''))
+        txt, kb = openclaw_answer(cmd + (' ' + arg if arg else ''), chat_id=chat_id)
         send_message(chat_id, txt, reply_to, kb)
 
 
@@ -5897,7 +5913,7 @@ def telegram_webhook():
         if question.startswith('/') or cmd in KNOWN:
             handle_cmd(cmd, arg, chat_id, msg_id)
         else:
-            txt, kb = openclaw_answer(question)
+            txt, kb = openclaw_answer(question, chat_id=chat_id)
             send_message(chat_id, txt, msg_id, kb)
 
         sys_log.info(f"[OpenClawBot] → replied chat={chat_id}")

services/openclaw_session.py

@@ -0,0 +1,133 @@
+"""ADR-019 Phase 4：OpenClaw 對話 state 管理（in-memory）。
+
+提供 chat_id 級別的多輪對話歷史 + carry-over slot：
+- history：最近 N 輪 (user, bot) pair，給 agent 注入 prompt 用
+- slots：跨訊息 carry-over（如「剛問月份 → 用戶答 4 → 接續執行」）
+
+設計取捨：
+- in-memory（不寫 DB），TTL 30 分鐘清掃，重啟後清空 — 對話 state 為臨時資料，
+  不該污染 ai_insights；要長期記憶請走 store_conversation 既有路徑。
+- 單進程：本檔不處理跨 worker 一致性。多 worker 部署時 session 可能落在不同
+  worker，但 OpenClaw bot 容器目前單 process（python-telegram-bot polling），
+  尚無此問題。若未來 webhook 要多 worker，再考慮搬 Redis。
+"""
+
+from __future__ import annotations
+
+from collections import deque
+from datetime import datetime, timedelta
+from threading import Lock
+from typing import Any, Optional
+
+
+_SESSION_TTL_MIN = 30
+_HISTORY_MAX = 5
+
+_sessions: dict[int, dict] = {}  # chat_id → {history, slots, updated_at}
+_lock = Lock()
+
+
+def _cleanup_expired_locked(now: datetime) -> None:
+    """假設 caller 已持鎖。移除超過 TTL 的 session。"""
+    cutoff = now - timedelta(minutes=_SESSION_TTL_MIN)
+    expired = [cid for cid, s in _sessions.items()
+               if s.get('updated_at', cutoff) < cutoff]
+    for cid in expired:
+        _sessions.pop(cid, None)
+
+
+def _get_or_create_locked(chat_id: int, now: datetime) -> dict:
+    """假設 caller 已持鎖。"""
+    if chat_id not in _sessions:
+        _sessions[chat_id] = {
+            'history': deque(maxlen=_HISTORY_MAX),
+            'slots': {},
+            'updated_at': now,
+        }
+    s = _sessions[chat_id]
+    s['updated_at'] = now
+    return s
+
+
+def append_turn(chat_id: int, user_msg: str, bot_reply: str) -> None:
+    """記錄一輪對話。content 會被截斷到 500 字以避免 prompt 爆量。"""
+    if not chat_id:
+        return
+    with _lock:
+        now = datetime.now()
+        s = _get_or_create_locked(chat_id, now)
+        s['history'].append({
+            'user': (user_msg or '')[:500],
+            'bot': (bot_reply or '')[:500],
+            'ts': now.isoformat(timespec='seconds'),
+        })
+        _cleanup_expired_locked(now)
+
+
+def set_slot(chat_id: int, key: str, value: Any) -> None:
+    """設定 carry-over slot。"""
+    if not chat_id:
+        return
+    with _lock:
+        now = datetime.now()
+        s = _get_or_create_locked(chat_id, now)
+        s['slots'][key] = value
+
+
+def get_slot(chat_id: int, key: str, default: Any = None) -> Any:
+    if not chat_id:
+        return default
+    with _lock:
+        s = _sessions.get(chat_id)
+        if not s:
+            return default
+        return s['slots'].get(key, default)
+
+
+def pop_slot(chat_id: int, key: str, default: Any = None) -> Any:
+    if not chat_id:
+        return default
+    with _lock:
+        s = _sessions.get(chat_id)
+        if not s:
+            return default
+        return s['slots'].pop(key, default)
+
+
+def history_as_prompt(chat_id: int, max_chars: int = 800) -> str:
+    """組合歷史對話成 system prompt 用的 context；無歷史回空字串。"""
+    if not chat_id:
+        return ""
+    with _lock:
+        s = _sessions.get(chat_id)
+        if not s or not s['history']:
+            return ""
+        lines = []
+        for turn in s['history']:
+            lines.append(f"[{turn['ts']}]\n用戶：{turn['user']}\n小O：{turn['bot']}")
+        text = "\n---\n".join(lines)
+    if len(text) > max_chars:
+        text = "...(較舊內容已截斷)\n" + text[-max_chars:]
+    return text
+
+
+def session_stats() -> dict:
+    """測試/觀察用：回傳目前 session 數量、各 chat 歷史長度。"""
+    with _lock:
+        return {
+            'active_sessions': len(_sessions),
+            'detail': {
+                cid: {'history_len': len(s['history']),
+                      'slot_keys': list(s['slots'].keys()),
+                      'updated_at': s['updated_at'].isoformat(timespec='seconds')}
+                for cid, s in _sessions.items()
+            }
+        }
+
+
+def reset_session(chat_id: int) -> None:
+    """測試/手動清除 session。"""
+    if not chat_id:
+        return
+    with _lock:
+        _sessions.pop(chat_id, None)

services/telegram_bot_service.py

@@ -931,8 +931,9 @@ class TrendTelegramBot:
             await update.message.reply_chat_action(action='typing')
             try:
                 from routes.openclaw_bot_routes import openclaw_answer
-                txt, kb = openclaw_answer(text)
-                
+                _chat_id = getattr(update.effective_chat, 'id', None)
+                txt, kb = openclaw_answer(text, chat_id=_chat_id)
+
                 reply_markup = None
                 if kb:
                     inline_kb = []