diff --git a/config.py b/config.py index 4e944d9..2c3113d 100644 --- a/config.py +++ b/config.py @@ -402,7 +402,7 @@ YOUTUBE_API_KEY = os.getenv('YOUTUBE_API_KEY', '') # ========================================== # 系統版本與路徑 # ========================================== -SYSTEM_VERSION = "V10.623" +SYSTEM_VERSION = "V10.624" LOG_FILE_PATH = os.path.join(BASE_DIR, 'logs/system.log') public_url = PUBLIC_URL # 用於模板顯示 diff --git a/docs/AI_INTELLIGENCE_MODULE_SOT.md b/docs/AI_INTELLIGENCE_MODULE_SOT.md index 8e0951a..8ea5734 100644 --- a/docs/AI_INTELLIGENCE_MODULE_SOT.md +++ b/docs/AI_INTELLIGENCE_MODULE_SOT.md @@ -2,7 +2,7 @@ > **最後更新**: 2026-06-16 (台北時間) > **狀態**: 🟢 四 AI Agent 自動化閉環已落地；LLM 路由紅線升級為 Ollama-first 三主機級聯；PChome 後台業績匯入韌性已補強；產品定位正名為「PChome 業績成長自動化作戰系統」；外部市場來源正規化層、自動同步、作戰清單與價格參考表優先讀取、CSV 備援預檢、前台操作入口、高可見頁面繁中化守門與比價/作戰 UI 工作台化已建立 -> **適用版本**: V10.623 +> **適用版本**: V10.624 --- @@ -197,6 +197,7 @@ SQL漏斗(~300筆) - 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 `price_drop_alert` / `market_opportunity` trigger 不得對整張 `price_records` 做全表最新價聚合；必須先篩最近有效 `identity_v2` PChome 候選，再用 per-SKU `JOIN LATERAL` 讀最新 MOMO 價格，並把 `match_score`、`tags`、`match_diagnostic_json` 帶入 evidence。 - 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。 +- V10.624 起 ElephantAlpha 價格類 trigger 即使信心度達自主門檻，也只能發送 HITL 價格覆核通知；必須跳過 orchestrator `execution_plan` 內的 Hermes/NemoTron/OpenClaw 長任務 step。這是價格決策護欄：避免 60 秒 execution timeout 卡住 scheduler，也避免把價格策略誤描述為已自動執行。 - ElephantAlpha 執行器若遇到舊版 OpenClaw strategy 類步驟（含 `generate_market_strategy` / `generate_dynamic_pricing_strategy` / `generate_resource_optimization_strategy`），只能記錄為 advisory skipped，不得觸發 circuit breaker，也不得轉成實際排程、外部呼叫或價格行動。 - `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` 的 Telegram 必須包含 `decision_envelope` 區塊，標明 `source_agent=elephant_alpha`、資料品質、量測證據、`can_auto_execute=false` 與 deterministic trace；此路徑不呼叫 Gemini、不呼叫 Hermes/NemoTron，也不得把 queue backlog 翻譯成主機資源耗盡。 diff --git a/docs/memory/current_execution_queue_20260524.md b/docs/memory/current_execution_queue_20260524.md index db0ee00..329ac16 100644 --- a/docs/memory/current_execution_queue_20260524.md +++ b/docs/memory/current_execution_queue_20260524.md @@ -303,3 +303,10 @@ - `/price_comparison` 的結果區新增決策摘要，先顯示「需檢查售價或活動 / 可主推曝光 / 觀察賣點」三類數字與建議，再往下看明細表。 - `/ai_intelligence` 第一屏新增今日任務摘要，直接顯示今日任務、可立即處理、待補比價與最新業績日；資料來自現有 PChome growth API stats。 - 測試守門新增 `priceDecisionGrid`、`price-workflow-strip`、`price-result-summary-grid`、`growth-executive-strip` 與 `renderGrowthExecutiveSummary`，避免頁面退回只有文字說明的狀態。 + +## 27. 2026-06-18 V10.624 ElephantAlpha 價格決策只進 HITL + +- 部署後觀察到 `price_drop_alert` 進入 ElephantAlpha execution plan 後可能卡在 Hermes/NemoTron step，最後以 60 秒 timeout 污染 scheduler log。 +- V10.624 將價格類 trigger 的高信心路徑改為「有實證就發 L3 HITL 價格覆核通知」，不再執行 orchestrator `execution_plan`，避免長任務 timeout 與自動調價誤解。 +- 新增 `price_decision_review` 決策信封，固定標示 `can_auto_execute=false`、`requires_hitl=true`、`execution_plan skipped`；通知只呈現 DB/Hermes 具體價差實證。 +- 測試新增高信心價格決策不執行長任務 step 的守門，避免未來又把價格告警回退成自主執行。 diff --git a/services/elephant_alpha_autonomous_engine.py b/services/elephant_alpha_autonomous_engine.py index e2bbf74..c7435d8 100644 --- a/services/elephant_alpha_autonomous_engine.py +++ b/services/elephant_alpha_autonomous_engine.py @@ -716,6 +716,18 @@ class ElephantAlphaAutonomousEngine: return if decision.confidence >= (0.85 if trigger.trigger_type in {"price_drop_alert", "market_opportunity"} else self.confidence_threshold): + if trigger.trigger_type in _PRICE_RELATED_TRIGGERS: + # 價格類決策即使信心高，也只進 HITL 覆核通知；不得執行 + # orchestrator 給出的 Hermes/NemoTron/OpenClaw 長任務 step。 + # 這避免 scheduler 被 60s execution timeout 卡住，也避免自動調價。 + await self._notify_telegram_executed(decision, trigger) + self._store_escalation(trigger.trigger_type) + self._log.info( + "Price decision queued for HITL review; execution plan skipped: %s", + trigger.trigger_type, + ) + self._circuit_reset() + return try: await self._run_with_timeout( self._execute_decision, @@ -1770,6 +1782,10 @@ class ElephantAlphaAutonomousEngine: decision: StrategicDecision, trigger: AutonomousTrigger, ) -> None: + if trigger.trigger_type in _PRICE_RELATED_TRIGGERS: + await self._notify_price_decision_review(decision, trigger) + return + if trigger.trigger_type == "resource_optimization": self._log.warning( "Suppressed legacy autonomous execution Telegram template for resource_optimization; " @@ -1800,6 +1816,123 @@ class ElephantAlphaAutonomousEngine: except Exception as e: self._log.error("Telegram audit failed (non-blocking): %s", e) + @staticmethod + def _build_price_review_decision_envelope( + decision: StrategicDecision, + trigger: AutonomousTrigger, + *, + concrete_actions: Optional[List[str]] = None, + ) -> Dict[str, Any]: + concrete_actions = [str(item).strip() for item in (concrete_actions or []) if str(item).strip()] + trigger_label = _zh_trigger(trigger.trigger_type) + evidence = [ + { + "type": "confidence", + "metric": "decision_confidence", + "value": f"{float(decision.confidence or 0):.2f}", + "basis": "ElephantAlpha high-confidence price signal; HITL still required", + "confidence": float(decision.confidence or 0), + }, + { + "type": "trigger", + "metric": "trigger_type", + "value": trigger.trigger_type, + "basis": trigger_label, + }, + ] + for idx, action in enumerate(concrete_actions[:3], start=1): + evidence.append({ + "type": "price_evidence", + "metric": f"candidate_{idx}", + "value": action[:120], + "basis": "DB/Hermes concrete price evidence", + }) + + return { + "decision_id": f"ea_price_review_{trigger.trigger_type}_{int(datetime.now().timestamp())}", + "decision_type": "price_decision_review", + "source_agent": "elephant_alpha", + "severity": "P2", + "confidence": float(decision.confidence or 0), + "analysis": "已找到價格比對實證，轉人工覆核；未批准前不執行調價或外部修復。", + "subject": { + "sku": trigger.trigger_type, + "name": f"Elephant Alpha · {trigger_label}", + }, + "evidence": evidence, + "recommended_action": { + "action": "review_price_or_promo", + "owner": "ops", + "requires_hitl": True, + }, + "expected_impact": { + "risk_reduction": "prevent unverified automated price action while preserving actionable evidence", + }, + "guardrails": { + "can_auto_execute": False, + "blocked_reason": "price decisions require HITL; execution_plan skipped", + "data_quality": "complete" if concrete_actions else "missing", + }, + "trace": { + "model": "deterministic_price_hitl_gate", + "provider": "elephant_alpha", + "trigger": trigger.trigger_type, + }, + } + + async def _notify_price_decision_review( + self, + decision: StrategicDecision, + trigger: AutonomousTrigger, + ) -> None: + concrete_actions = ( + self._get_prefetched_concrete_actions(trigger) + or self._get_trigger_db_concrete_actions(trigger) + ) + if not concrete_actions: + self._record_suppressed_escalation(decision, trigger, "no_concrete_price_evidence") + return + + try: + from services.telegram_templates import triaged_alert, _send_telegram_raw + + decision_envelope = self._build_price_review_decision_envelope( + decision, + trigger, + concrete_actions=concrete_actions, + ) + msg, keyboard = triaged_alert( + base_event={ + "event_type": "ea_price_review", + "title": f"🐘 EA 價格覆核 · {_zh_trigger(trigger.trigger_type)}", + "summary": ( + f"找到 {len(concrete_actions)} 筆價格比對實證，已轉人工覆核；" + "未批准前不自動調價。" + ), + "id": decision_envelope.get("decision_id"), + "decision_envelope": decision_envelope, + }, + tier_label="🐘 Elephant Alpha · L3 HITL", + ai_summary=( + f"已保留 {len(concrete_actions)} 筆 DB/Hermes 價格比對實證；" + "本通知只要求人工覆核，不執行外部修復或調價。" + ), + ai_cause=( + f"觸發類型：{_zh_trigger(trigger.trigger_type)} | " + f"信心度：{decision.confidence:.2f} | " + "高信心價格訊號仍需 HITL" + ), + ai_actions=concrete_actions, + ) + await self._run_with_timeout(_send_telegram_raw, msg, timeout=10, reply_markup=keyboard) + self._log.info( + "Price HITL review Telegram sent: %s concrete=%d", + trigger.trigger_type, + len(concrete_actions), + ) + except Exception as e: + self._log.error("Price HITL review Telegram failed (non-blocking): %s", e) + @staticmethod def _get_prefetched_concrete_actions(trigger: AutonomousTrigger) -> Optional[List[str]]: actions = (trigger.conditions or {}).get("_prefetched_hermes_threats") diff --git a/tests/test_elephant_alpha_engine.py b/tests/test_elephant_alpha_engine.py index be8965f..f55327f 100644 --- a/tests/test_elephant_alpha_engine.py +++ b/tests/test_elephant_alpha_engine.py @@ -198,6 +198,93 @@ def test_execute_autonomous_decision_uses_db_evidence_without_hermes_prefetch(mo assert notified == ["price_drop_alert"] +def test_high_confidence_price_decision_skips_execution_plan_and_goes_hitl(monkeypatch): + import services.elephant_alpha_autonomous_engine as engine_module + from services.elephant_alpha_autonomous_engine import ( + AutonomousTrigger, + ElephantAlphaAutonomousEngine, + ) + from services.elephant_alpha_orchestrator import StrategicDecision + + engine = ElephantAlphaAutonomousEngine() + notified = [] + stored = [] + + async def _capture_context(context): + return StrategicDecision( + priority="high", + agents_required=["hermes", "nemotron"], + reasoning="已有 DB 價格比對實證，轉人工覆核。", + expected_outcome="人工確認價格策略", + confidence=0.96, + execution_plan=[ + {"agent": "hermes", "action": "analyze_price_competition"}, + {"agent": "nemotron", "action": "dispatch_alert", "parameters": {"threats": []}}, + ], + resource_requirements={}, + ) + + async def _execution_should_not_run(decision): + raise AssertionError("price decisions must not execute long-running autonomous steps") + + async def _capture_notify(decision, trigger): + notified.append((trigger.trigger_type, len(decision.execution_plan))) + + monkeypatch.setattr(engine_module.elephant_orchestrator, "analyze_and_coordinate", _capture_context) + monkeypatch.setattr(engine, "_execute_decision", _execution_should_not_run) + monkeypatch.setattr(engine, "_notify_telegram_executed", _capture_notify) + monkeypatch.setattr(engine, "_store_escalation", lambda trigger_type: stored.append(trigger_type)) + + trigger = AutonomousTrigger( + trigger_type="price_drop_alert", + conditions={"_db_evidence_actions": ["[SKU-1] DB 實證價差"]}, + threshold=0.8, + enabled=True, + ) + + asyncio.run(engine._execute_autonomous_decision(trigger)) + + assert notified == [("price_drop_alert", 2)] + assert stored == ["price_drop_alert"] + + +def test_price_review_decision_envelope_blocks_auto_execution(): + from services.elephant_alpha_autonomous_engine import ( + AutonomousTrigger, + ElephantAlphaAutonomousEngine, + ) + from services.elephant_alpha_orchestrator import StrategicDecision + + decision = StrategicDecision( + priority="high", + agents_required=["elephant_alpha"], + reasoning="已有價格比對實證。", + expected_outcome="人工確認價格策略。", + confidence=0.96, + execution_plan=[{"agent": "hermes", "action": "analyze_price_competition"}], + resource_requirements={}, + ) + trigger = AutonomousTrigger( + trigger_type="price_drop_alert", + conditions={}, + threshold=0.8, + enabled=True, + ) + + envelope = ElephantAlphaAutonomousEngine._build_price_review_decision_envelope( + decision, + trigger, + concrete_actions=["[SKU-1] MOMO $1,200 vs PChome $990"], + ) + + assert envelope["decision_type"] == "price_decision_review" + assert envelope["source_agent"] == "elephant_alpha" + assert envelope["recommended_action"]["requires_hitl"] is True + assert envelope["guardrails"]["can_auto_execute"] is False + assert envelope["guardrails"]["blocked_reason"] == "price decisions require HITL; execution_plan skipped" + assert envelope["guardrails"]["data_quality"] == "complete" + + def test_price_trigger_queries_use_lateral_latest_price_lookup(monkeypatch): import services.elephant_alpha_autonomous_engine as engine_module from services.elephant_alpha_autonomous_engine import (