強化 EA JSON fallback 與 EDM cache 自癒

config.py

@@ -325,7 +325,7 @@ YOUTUBE_API_KEY = os.getenv('YOUTUBE_API_KEY', '')
 # ==========================================
 # 系統版本與路徑
 # ==========================================
-SYSTEM_VERSION = "V10.382"
+SYSTEM_VERSION = "V10.383"
 LOG_FILE_PATH = os.path.join(BASE_DIR, 'logs/system.log')
 public_url = PUBLIC_URL  # 用於模板顯示
 

docs/AI_INTELLIGENCE_MODULE_SOT.md

@@ -117,6 +117,7 @@ SQL漏斗(~300筆)
 - ElephantAlpha 使用 NVIDIA NIM hosted API；production 預設模型為 `nvidia/llama-3.3-nemotron-super-49b-v1.5`，`ELEPHANT_ALPHA_FALLBACK_MODELS` 需保留至少一個可呼叫備援；403/404、408/409/425/429、5xx、timeout 與 connection error 必須嘗試下一個模型。
 - ElephantAlpha L3 HITL 只允許發送有實證、可審核、可行動的升級告警；價格類 trigger 無 Hermes 具體威脅時，只記錄 suppressed escalation telemetry 與 cooldown，不寫 pending `human_review`，不發 Telegram 空告警。
 - ElephantAlpha 價格類 trigger 的 HITL / 決策 prefetch 必須先使用觸發 SQL 與 `competitor_prices` / `price_records` 的 DB 實證生成 SKU、MOMO / PChome 價差與建議 action lines；完整 Hermes LLM prefetch 預設關閉（`ELEPHANT_ALPHA_HERMES_LLM_PREFETCH_ENABLED=false`），避免 5s timeout 後落入無實證摘要或雲端備援。若無 DB 實證，只記錄 suppressed telemetry / cooldown，不發 Telegram 空告警。
+- ElephantAlpha 協調器收到非純 JSON、fenced JSON 或混文字 JSON 時，必須先做容錯抽取；仍無法解析時，只能使用 DB/Hermes 實證生成保守 HITL fallback。fallback 不得放入 OpenClaw `generate_*` 類舊策略步驟，也不得暗示已自動調價。
 - `resource_optimization` 不再交給 LLM 生成「預期效益 / 已執行」敘事，顯示名稱統一為「資源壓力治理」。此 trigger 必須先由程式量測 `action_plans` backlog、P1/P2 數、pending_review、逾時項目與 CPU load；只有 CPU 達門檻、P1/P2 積壓或逾時積壓才發 Telegram「資源壓力告警」。單純 queue 大但 CPU 正常只記錄 telemetry，不派發 Hermes/NemoTron、不宣稱 48 小時效益；Telegram 段落使用「系統處置紀錄」而非泛稱「已執行」，避免暗示 AI 已完成未經驗證的外部動作。
 - `resource_optimization` 會先執行 `ActionPlanHygieneService` 清理過期噪音：只關閉超過 72 小時的 `code_review_fix` / `openclaw_recommendation` 類 advisory action_plans，以及 NemoTron `direct_response/reply_simple` 舊聊天回覆計畫；將狀態改為 `auto_disabled` 或 `rejected` 並寫入 `metadata_json.hygiene_history`。不刪資料，也不碰 NemoTron human_review / pricing / tool action 類業務行動。
 - `momo-scheduler` 每 6 小時固定執行 `run_action_plan_hygiene_task()`，讓過期 advisory action_plans 的關閉不再依賴 `resource_optimization` 告警觸發；排程失敗會經 EventRouter 發送 `action_plan_hygiene_failure`。
@@ -617,5 +618,6 @@ POSTGRES_HOST=momo-db
 | 2026-05-20 | 指定日期競品簡報可能混用目前 `competitor_prices` 快取價 | V10.315 起 `fetch_competitor_comparison_results()` 有 start/end date 時改用 `competitor_price_history` 期間快照，MOMO 價格取報表結束日前最新價；即時報表才使用目前有效 `competitor_prices` |
 | 2026-05-20 | PChome 覆蓋率分子可能被非活躍或無 MOMO 現價 SKU 膨脹 | V10.317 起 `fetch_competitor_coverage()` 的 `valid_matches` 改為 active MOMO latest price 與有效 PChome `identity_v2` 價格交集，確保 daily/growth/PPT/AI 看到的比價資料品質不被舊快取列高估 |
 | 2026-05-20 | EA HITL 告警可能把非 SKU 診斷誤排成待審 SKU，或在缺少 DB/Hermes 實證時打擾人工 | V10.318 起 `ea_escalation` 僅對含 SKU/價格比較的 actions 使用競價卡片；非 SKU 診斷改為「待確認事項」。價格類低信心事件若無 DB/Hermes 實證，測試鎖定只 suppress、不寫 human_review、不發 Telegram |
+| 2026-05-21 | ElephantAlpha NIM/LLM 回應偶爾不是純 JSON，會觸發 `json.loads()` 失敗並落入舊式空泛策略 fallback | V10.383 起協調器容忍 fenced/混文字 JSON；無法解析時改用 DB/Hermes 實證 fallback，且 fallback 不再包含 OpenClaw `generate_*` 舊步驟或自動調價暗示 |
 | 2026-05-20 | Telegram HTML parse mode 不支援 `<br>`，可能導致告警或報告送出 400 | V10.321 起 Telegram template 發送前會把 `<br>` / `<br/>` / `<BR />` 轉為換行；保留其他 HTML 標籤，非 HTML parse mode 不改寫 |
 | 2026-05-20 | 部分舊 Telegram 入口繞過中央 sanitizer，且 RAG awaiting review 使用錯誤 `chat_id=` 參數會讓人工審核推播失敗 | V10.322 起 Bot API price decision 走 `send_telegram_with_result()`；`price_decision()` 補 `report_url` 相容並 escape 動態欄位；RAG awaiting review 改用 `chat_ids=[...]` 呼叫 `_send_telegram_raw()` |

docs/memory/history_logs.md

@@ -13,6 +13,7 @@
 ## 📅 詳細更新日誌 (考古存檔)
 
 ### 2026-05-21：瀏覽器測試守門與 PChome 熱路徑優化
+- **V10.383 EA JSON fallback / EDM cache 自癒 / 比對別名補強**: Elephant Alpha 協調器現在可容忍 fenced JSON 與混文字 JSON；若仍無法解析，會改用 DB/Hermes 實證產生保守人工覆核決策，不再輸出舊式 OpenClaw 策略 plan 或自動調價暗示。EDM promo dashboard shared cache 遇到損毀 pickle 會自動刪除並重建，避免每個 worker 重複噴 `UnicodeDecodeError`。Marketplace matcher 補上 Curel/珂潤、Karadium 與兩個強 identity anchor 測試，降低真同款漏報。
 - **V10.382 唇膏 exact identity 寬價差豁免**: marketplace matcher 對「同品牌 + 共享唇膏 identity anchor + 規格完全一致 + 無色號/變體衝突」的唇膏類商品，允許 sequence score 略低時仍套用 `price_penalty_suppressed_wide_exact_identity`；這只處理 PChome/MOMO 標題順序與行銷字差異造成的真同款漏報，不放寬顯性色號不同的 hard veto。
 - **V10.381 browse.sh 比價診斷計畫**: PChome feeder 在 `no_result`、`no_match`、低信心、單位價覆核、既有配對保護與爬蟲錯誤時，會把 read-only `browse_diagnostic_json` 寫入 `competitor_match_attempts`，內含 PChome search URL 與建議 `browse get/open` 命令；正式排程仍 API-first，`PCHOME_FEEDER_BROWSE_SH_EXECUTE_ENABLED=false` 預設不自動開瀏覽器，避免瀏覽器彈窗、登入或密碼提示干擾。
 - **V10.380 111 Ollama final fallback 收斂**: 111 Mac fallback 從救急路徑改成更短的保護路徑，`OLLAMA_111_MAX_TIMEOUT` 預設由 45s 收緊到 20s，並新增 `OLLAMA_111_NUM_PREDICT=512` 輸出上限；落到 111 時仍會降級重模型到 `llama3.2:latest`、縮 `num_ctx=4096`、`keep_alive=5m`，避免 GCP-A/GCP-B 短暫 timeout 後把長篇 Hermes/OpenClaw 工作轉嫁到 111 造成 swap 與 load 飆高。

routes/edm_routes.py

@@ -120,11 +120,22 @@ def _load_shared_promo_dashboard_cache(cache_key):
         if not isinstance(entries, dict):
             return None
         return entries.get(cache_key)
-    except Exception:
-        sys_log.debug("promo dashboard shared cache load failed", exc_info=True)
+    except Exception as exc:
+        _discard_shared_promo_dashboard_cache("load_corrupt", exc)
         return None
 
 
+def _discard_shared_promo_dashboard_cache(reason, exc=None):
+    """刪除損毀的跨 worker 快取，避免每個 worker 重複噴 traceback。"""
+    try:
+        os.remove(_PROMO_SHARED_CACHE_FILE)
+        sys_log.info("promo dashboard shared cache discarded | reason=%s | error=%s", reason, exc)
+    except FileNotFoundError:
+        return
+    except OSError:
+        sys_log.debug("promo dashboard shared cache discard failed", exc_info=True)
+
+
 def _write_shared_promo_dashboard_cache(cache_key, data):
     """原子寫入促銷 dashboard 共享快取；失敗不阻斷頁面回應。"""
     cache_file = str(_PROMO_SHARED_CACHE_FILE)
@@ -132,10 +143,14 @@ def _write_shared_promo_dashboard_cache(cache_key, data):
     entries = {}
     try:
         if os.path.exists(_PROMO_SHARED_CACHE_FILE):
-            with open(_PROMO_SHARED_CACHE_FILE, 'rb') as f:
-                payload = pickle.load(f)
-            if payload.get('version') == SYSTEM_VERSION and isinstance(payload.get('entries'), dict):
-                entries = payload['entries']
+            try:
+                with open(_PROMO_SHARED_CACHE_FILE, 'rb') as f:
+                    payload = pickle.load(f)
+                if payload.get('version') == SYSTEM_VERSION and isinstance(payload.get('entries'), dict):
+                    entries = payload['entries']
+            except Exception as exc:
+                _discard_shared_promo_dashboard_cache("write_corrupt", exc)
+                entries = {}
         entries[cache_key] = data
         while len(entries) > _PROMO_DASHBOARD_CACHE_MAX:
             entries.pop(next(iter(entries)))

services/elephant_alpha_orchestrator.py

@@ -15,6 +15,7 @@ Position: Super Orchestrator above Hermes/NemoTron/OpenClaw
 import os
 import json
 import asyncio
+from json import JSONDecodeError
 from datetime import datetime, timedelta
 from typing import Dict, List, Any, Optional
 from dataclasses import dataclass
@@ -241,8 +242,22 @@ BUSINESS OBJECTIVE: Optimize e-commerce performance through intelligent automati
             if not response.success:
                 raise RuntimeError(response.error)
 
-            # Parse and validate response
-            decision_data = json.loads(response.content)
+            # Parse and validate response. Some NIM-compatible models still wrap
+            # JSON in fenced blocks or prepend reasoning text even with json_mode.
+            try:
+                decision_data = self._extract_json_object(response.content)
+            except ValueError as parse_error:
+                logger.warning(
+                    "[ElephantAlpha] Coordination JSON parse failed; using evidence fallback. "
+                    "model=%s error=%s preview=%r",
+                    response.model,
+                    parse_error,
+                    (response.content or "")[:240],
+                )
+                return self._fallback_decision(
+                    business_context,
+                    reason="Elephant Alpha 回應不是可解析 JSON，已改用實證 fallback。",
+                )
             decision = self._parse_strategic_decision(decision_data)
             
             # Log decision for learning
@@ -253,7 +268,46 @@ BUSINESS OBJECTIVE: Optimize e-commerce performance through intelligent automati
         except Exception as e:
             logger.error(f"[ElephantAlpha] Coordination failed: {e}")
             # Fallback to conservative decision
-            return self._fallback_decision(business_context)
+            return self._fallback_decision(
+                business_context,
+                reason=f"Elephant Alpha 協調失敗：{type(e).__name__}",
+            )
+
+    @staticmethod
+    def _extract_json_object(raw: str) -> Dict[str, Any]:
+        """Extract one JSON object from tolerant LLM output."""
+        text_value = (raw or "").strip()
+        if not text_value:
+            raise ValueError("empty response")
+
+        if text_value.startswith("```"):
+            lines = text_value.splitlines()
+            if lines and lines[0].strip().startswith("```"):
+                lines = lines[1:]
+            if lines and lines[-1].strip().startswith("```"):
+                lines = lines[:-1]
+            text_value = "\n".join(lines).strip()
+
+        try:
+            parsed = json.loads(text_value)
+            if isinstance(parsed, dict):
+                return parsed
+            raise ValueError("JSON root is not an object")
+        except JSONDecodeError:
+            pass
+
+        decoder = json.JSONDecoder()
+        for idx, char in enumerate(text_value):
+            if char != "{":
+                continue
+            try:
+                parsed, _end = decoder.raw_decode(text_value[idx:])
+            except JSONDecodeError:
+                continue
+            if isinstance(parsed, dict):
+                return parsed
+
+        raise ValueError("no JSON object found")
     
     def _build_coordination_prompt(self, context: Dict[str, Any]) -> str:
         """Build detailed coordination prompt for Elephant Alpha"""
@@ -431,22 +485,56 @@ Provide your strategic decision in the specified JSON format.
         finally:
             session.close()
     
-    def _fallback_decision(self, context: Dict[str, Any]) -> StrategicDecision:
+    @staticmethod
+    def _context_concrete_actions(context: Dict[str, Any]) -> List[str]:
+        conditions = context.get("conditions") if isinstance(context, dict) else {}
+        if not isinstance(conditions, dict):
+            return []
+        for key in ("_prefetched_hermes_threats", "_db_evidence_actions"):
+            actions = conditions.get(key)
+            if isinstance(actions, list):
+                cleaned = [str(action).strip() for action in actions if str(action).strip()]
+                if cleaned:
+                    return cleaned[:5]
+        return []
+
+    def _fallback_decision(self, context: Dict[str, Any], *, reason: str = "Elephant Alpha unavailable") -> StrategicDecision:
         """Fallback decision if Elephant Alpha fails"""
+        concrete_actions = self._context_concrete_actions(context)
+        trigger_type = str((context or {}).get("trigger_type") or "unknown")
+        if concrete_actions:
+            return StrategicDecision(
+                priority="high",
+                agents_required=["hermes", "elephant_alpha"],
+                reasoning=(
+                    f"{reason} 已保留 {len(concrete_actions)} 筆 DB/Hermes 價格比對實證；"
+                    "僅送人工覆核，不執行自動調價。"
+                ),
+                expected_outcome="產生可稽核的人工覆核告警，避免使用無法解析的 LLM 推論文字。",
+                confidence=0.74,
+                execution_plan=[],
+                resource_requirements={
+                    "compute_cost": "$0.00",
+                    "time_estimate": "人工覆核",
+                    "human_oversight": "required",
+                },
+            )
+
         return StrategicDecision(
             priority="medium",
-            agents_required=["openclaw"],
-            reasoning="Elephant Alpha unavailable, using conservative OpenClaw strategy",
-            expected_outcome="Basic strategic analysis",
+            agents_required=["elephant_alpha"],
+            reasoning=(
+                f"{reason} trigger={trigger_type}，且沒有可稽核的 DB/Hermes 實證；"
+                "不產生策略型行動計畫，避免把推測當成事實。"
+            ),
+            expected_outcome="不執行自動動作；需要先確認資料來源或等待下一輪具體實證。",
             confidence=0.6,
-            execution_plan=[{
-                "step": 1,
-                "agent": "openclaw",
-                "action": "generate_market_analysis",
-                "parameters": context,
-                "expected_duration": "2-3 minutes"
-            }],
-            resource_requirements={"compute_cost": "$0.00", "time_estimate": "5 minutes"}
+            execution_plan=[],
+            resource_requirements={
+                "compute_cost": "$0.00",
+                "time_estimate": "等待實證",
+                "human_oversight": "required",
+            },
         )
 
 # Singleton instance

services/marketplace_product_matcher.py

@@ -304,6 +304,8 @@ SEARCH_NOISE_TOKENS = {
 }
 
 SEARCH_IDENTITY_ANCHORS = (
+    "潤浸保濕清爽身體乳液",
+    "閃亮珍珠眼影棒",
     "智能光感應無線自動除臭芳香噴霧機",
     "usb精油薰香機",
     "超音波水氧機",
@@ -520,6 +522,8 @@ BRAND_ALIAS_OVERRIDES = {
     "xiaomi": ("小米有品", "小米", "xiaomi"),
     "mac": ("m.a.c", "mac", "m a c"),
     "opi": ("o.p.i", "opi", "o p i"),
+    "curel": ("curel", "珂潤"),
+    "karadium": ("karadium",),
     "st雞仔牌": ("日本雞仔牌st", "日本st雞仔牌", "st雞仔牌", "雞仔牌st", "雞仔牌"),
 }
 

tests/test_cache_manager.py

@@ -316,6 +316,31 @@ def test_promo_dashboard_shared_cache_ignores_other_versions(tmp_path, monkeypat
     assert edm_routes._load_shared_promo_dashboard_cache(cache_key) is None
 
 
+def test_promo_dashboard_shared_cache_discards_corrupt_payload(tmp_path, monkeypatch):
+    from routes import edm_routes
+
+    shared_cache = tmp_path / "promo_dashboard_cache.pkl"
+    shared_cache.write_bytes(b"\x96not-a-pickle")
+    monkeypatch.setattr(edm_routes, "_PROMO_SHARED_CACHE_FILE", shared_cache)
+
+    assert edm_routes._load_shared_promo_dashboard_cache(("edm",)) is None
+    assert not shared_cache.exists()
+
+
+def test_promo_dashboard_shared_cache_write_recovers_from_corrupt_payload(tmp_path, monkeypatch):
+    from routes import edm_routes
+
+    shared_cache = tmp_path / "promo_dashboard_cache.pkl"
+    shared_cache.write_bytes(b"\x96not-a-pickle")
+    cache_key = ("edm", "default", "desc", "", "bucket", (1, "ts", 1))
+    data = {"items_in_batch": [{"sku": "SKU-1"}]}
+    monkeypatch.setattr(edm_routes, "_PROMO_SHARED_CACHE_FILE", shared_cache)
+
+    edm_routes._write_shared_promo_dashboard_cache(cache_key, data)
+
+    assert edm_routes._load_shared_promo_dashboard_cache(cache_key) == data
+
+
 def test_sales_analysis_preview_context_cache_avoids_reloading_options(tmp_path, monkeypatch):
     from routes import sales_routes
 

tests/test_elephant_alpha_orchestrator.py

@@ -0,0 +1,99 @@
+import asyncio
+
+
+def test_elephant_orchestrator_extracts_fenced_json_object():
+    from services.elephant_alpha_orchestrator import ElephantAlphaOrchestrator
+
+    payload = ElephantAlphaOrchestrator._extract_json_object(
+        """```json
+{"priority":"high","confidence":0.91}
+```"""
+    )
+
+    assert payload == {"priority": "high", "confidence": 0.91}
+
+
+def test_elephant_orchestrator_extracts_json_object_from_prefaced_text():
+    from services.elephant_alpha_orchestrator import ElephantAlphaOrchestrator
+
+    payload = ElephantAlphaOrchestrator._extract_json_object(
+        '先做分析：\n{"priority":"medium","agents_required":["hermes"]}\n以上。'
+    )
+
+    assert payload["priority"] == "medium"
+    assert payload["agents_required"] == ["hermes"]
+
+
+def test_elephant_orchestrator_fallback_uses_concrete_evidence_without_plan():
+    from services.elephant_alpha_orchestrator import ElephantAlphaOrchestrator
+
+    orchestrator = ElephantAlphaOrchestrator()
+    decision = orchestrator._fallback_decision(
+        {
+            "trigger_type": "market_opportunity",
+            "conditions": {
+                "_prefetched_hermes_threats": [
+                    "[SKU-1] 測試商品｜MOMO $1,200 vs PChome $990｜每件價差 NT$ 210"
+                ]
+            },
+        },
+        reason="測試解析失敗",
+    )
+
+    assert decision.priority == "high"
+    assert decision.confidence == 0.74
+    assert decision.execution_plan == []
+    assert "1 筆 DB/Hermes 價格比對實證" in decision.reasoning
+    assert "自動調價" in decision.reasoning
+
+
+def test_elephant_orchestrator_fallback_without_evidence_has_no_openclaw_action():
+    from services.elephant_alpha_orchestrator import ElephantAlphaOrchestrator
+
+    orchestrator = ElephantAlphaOrchestrator()
+    decision = orchestrator._fallback_decision(
+        {"trigger_type": "market_opportunity", "conditions": {}},
+        reason="測試解析失敗",
+    )
+
+    assert decision.confidence == 0.6
+    assert decision.execution_plan == []
+    assert decision.agents_required == ["elephant_alpha"]
+    assert "沒有可稽核的 DB/Hermes 實證" in decision.reasoning
+
+
+def test_elephant_orchestrator_uses_evidence_fallback_on_non_json_response(monkeypatch):
+    from services.elephant_service import ElephantResponse
+    from services.elephant_alpha_orchestrator import ElephantAlphaOrchestrator
+
+    class FakeElephant:
+        def generate(self, **_kwargs):
+            return ElephantResponse(
+                success=True,
+                content="\n我無法輸出 JSON，但建議先處理。",
+                model="fake-model",
+            )
+
+    async def _noop_log(*_args, **_kwargs):
+        raise AssertionError("parse fallback should not log an LLM decision")
+
+    orchestrator = ElephantAlphaOrchestrator()
+    orchestrator.elephant = FakeElephant()
+    monkeypatch.setattr(orchestrator, "_get_recent_performance_metrics", lambda: {})
+    monkeypatch.setattr(orchestrator, "_get_agent_status", lambda: {})
+    monkeypatch.setattr(orchestrator, "_get_system_load", lambda: "normal")
+    monkeypatch.setattr(orchestrator, "_get_pending_actions_count", lambda: 0)
+    monkeypatch.setattr(orchestrator, "_log_decision", _noop_log)
+
+    decision = asyncio.run(orchestrator.analyze_and_coordinate({
+        "trigger_type": "price_drop_alert",
+        "conditions": {
+            "_prefetched_hermes_threats": [
+                "[SKU-9] 實證商品｜MOMO $1,000 vs PChome $800｜每件價差 NT$ 200"
+            ]
+        },
+    }))
+
+    assert decision.priority == "high"
+    assert decision.execution_plan == []
+    assert "1 筆 DB/Hermes 價格比對實證" in decision.reasoning

tests/test_marketplace_product_matcher.py

@@ -531,6 +531,38 @@ def test_marketplace_matcher_promotes_nivea_dry_lotion_with_long_shared_anchor()
     assert "shared_identity_anchor_nivea_dry_lotion" in diagnostics.reasons
 
 
+def test_marketplace_matcher_promotes_curel_body_lotion_with_brand_alias():
+    from services.marketplace_product_matcher import score_marketplace_match
+
+    diagnostics = score_marketplace_match(
+        "【Curel 珂潤】潤浸保濕清爽身體乳液 220ml",
+        "珂潤 Curel 潤浸保濕清爽身體乳液220ml",
+        momo_price=399,
+        competitor_price=399,
+    )
+
+    assert diagnostics.score >= 0.76
+    assert diagnostics.hard_veto is False
+    assert "brand_match" in diagnostics.tags
+    assert "shared_identity_anchor" in diagnostics.reasons
+
+
+def test_marketplace_matcher_promotes_karadium_pearl_shadow_stick_anchor():
+    from services.marketplace_product_matcher import score_marketplace_match
+
+    diagnostics = score_marketplace_match(
+        "【KARADIUM】閃亮珍珠眼影棒 1.4g",
+        "karadium 閃亮珍珠眼影棒 1.4g",
+        momo_price=259,
+        competitor_price=259,
+    )
+
+    assert diagnostics.score >= 0.76
+    assert diagnostics.hard_veto is False
+    assert "brand_match" in diagnostics.tags
+    assert "shared_identity_anchor" in diagnostics.reasons
+
+
 def test_marketplace_matcher_rejects_refill_core_vs_case_only_pack():
     from services.marketplace_product_matcher import score_marketplace_match