落地 L2 安全記憶動作

services/agent_actions.py

@@ -6,7 +6,7 @@ L2 NemoTron 可安全呼叫的動作集合。嚴格限制：
 - 不可動 prod 資料表 / 容器 / 外部系統
 - 所有 action 必須 dual-write 審計軌跡
 
-現階段為 **stub + 完整 interface**，供 event_router 串接。真實執行邏輯將於 Phase 3 填入。
+EventRouter 只會執行本檔 SAFE_ACTIONS；所有動作必須保守、可審計、可回放。
 """
 
 from __future__ import annotations
@@ -51,6 +51,29 @@ def _audit(action: str, params: dict, result: dict, latency_ms: float) -> int |
         return None
 
 
+def _store_action_memory(
+    insight_type: str,
+    content: str,
+    *,
+    product_sku: str | None = None,
+    metadata: dict | None = None,
+    status: str = "approved",
+) -> int | None:
+    """Write a concrete L2 action outcome into OpenClaw memory."""
+    from services.openclaw_learning_service import store_insight
+    return store_insight(
+        insight_type=insight_type,
+        content=content,
+        period=datetime.now().strftime("%Y-%m-%d"),
+        product_sku=product_sku,
+        metadata=metadata or {},
+        ai_model="agent_actions",
+        confidence=0.8,
+        created_by="agent_actions",
+        status=status,
+    )
+
+
 ALLOWED_RETRY_TASKS = {
     "run_auto_import_task", "run_momo_task", "run_edm_task",
     "run_competitor_price_feeder_task", "run_backup_monitor_task",
@@ -205,28 +228,46 @@ def is_silenced(event_key: str) -> bool:
 # 🏷️ 三個既有 NemoTron tool 的 wrapper（供 event_router 統一調用）
 # =====================================================================
 def flag_for_human_review(sku: str, concern: str) -> dict:
-    """升級到 L3 HITL（包裝 NemoTron 既有 tool，保持呼叫介面一致）"""
+    """升級到 L3 HITL：寫入 human_review 記憶，等待人工後續處理。"""
     t0 = time.time()
-    # TODO Phase 3: 接入 nemoton_dispatcher_service._exec_flag_for_human_review
-    result = {"status": "stub", "sku": sku, "concern": concern,
-              "note": "Phase 1 stub，Phase 3 接 NemoTron"}
+    insight_id = _store_action_memory(
+        "human_review",
+        f"SKU={sku} 需要人工覆核：{concern}",
+        product_sku=sku,
+        metadata={"sku": sku, "concern": concern, "source": "flag_for_human_review"},
+        status="pending",
+    )
+    result = {"status": "pending_review", "sku": sku, "concern": concern, "insight_id": insight_id}
     _audit("flag_for_human_review", {"sku": sku, "concern": concern},
            result, (time.time() - t0) * 1000)
     return result
 
 
 def route_to_km(sku: str, domain: str, summary: str) -> dict:
-    """KM 歸檔（Phase 3 接 NemoTron）"""
+    """KM 歸檔：將 NemoTron/Hermes 判斷沉澱為可檢索知識。"""
     t0 = time.time()
-    result = {"status": "stub", "note": "Phase 3 接 NemoTron"}
+    insight_id = _store_action_memory(
+        "km_entry",
+        f"[{domain}] SKU={sku}：{summary}",
+        product_sku=sku,
+        metadata={"sku": sku, "domain": domain, "summary": summary, "source": "route_to_km"},
+    )
+    result = {"status": "archived", "sku": sku, "domain": domain, "insight_id": insight_id}
     _audit("route_to_km", {"sku": sku, "domain": domain}, result, (time.time() - t0) * 1000)
     return result
 
 
 def mark_for_relearn(sku: str, reason: str) -> dict:
-    """標記重新訓練（Phase 3 接 NemoTron）"""
+    """標記重新訓練：寫入 relearn_marker 供 OpenClaw/品質批次使用。"""
     t0 = time.time()
-    result = {"status": "stub", "note": "Phase 3 接 NemoTron"}
+    insight_id = _store_action_memory(
+        "relearn_marker",
+        f"SKU={sku} 需要重新學習：{reason}",
+        product_sku=sku,
+        metadata={"sku": sku, "reason": reason, "source": "mark_for_relearn"},
+        status="pending",
+    )
+    result = {"status": "marked", "sku": sku, "reason": reason, "insight_id": insight_id}
     _audit("mark_for_relearn", {"sku": sku, "reason": reason}, result, (time.time() - t0) * 1000)
     return result
 

tests/test_agent_actions.py

@@ -0,0 +1,50 @@
+def test_flag_for_human_review_writes_pending_memory(monkeypatch):
+    import services.agent_actions as actions
+    import services.openclaw_learning_service as learning
+
+    calls = []
+    monkeypatch.setattr(actions, "_audit", lambda *args, **kwargs: 999)
+    monkeypatch.setattr(
+        learning,
+        "store_insight",
+        lambda **kwargs: calls.append(kwargs) or 123,
+    )
+
+    result = actions.flag_for_human_review("SKU-1", "銷量斷崖，請人工確認")
+
+    assert result["status"] == "pending_review"
+    assert result["insight_id"] == 123
+    assert calls[0]["insight_type"] == "human_review"
+    assert calls[0]["status"] == "pending"
+    assert calls[0]["product_sku"] == "SKU-1"
+
+
+def test_route_to_km_writes_archived_memory(monkeypatch):
+    import services.agent_actions as actions
+    import services.openclaw_learning_service as learning
+
+    calls = []
+    monkeypatch.setattr(actions, "_audit", lambda *args, **kwargs: 999)
+    monkeypatch.setattr(learning, "store_insight", lambda **kwargs: calls.append(kwargs) or 456)
+
+    result = actions.route_to_km("SKU-2", "pricing", "競品價差擴大")
+
+    assert result == {"status": "archived", "sku": "SKU-2", "domain": "pricing", "insight_id": 456}
+    assert calls[0]["insight_type"] == "km_entry"
+    assert calls[0]["metadata"]["domain"] == "pricing"
+
+
+def test_mark_for_relearn_writes_pending_marker(monkeypatch):
+    import services.agent_actions as actions
+    import services.openclaw_learning_service as learning
+
+    calls = []
+    monkeypatch.setattr(actions, "_audit", lambda *args, **kwargs: 999)
+    monkeypatch.setattr(learning, "store_insight", lambda **kwargs: calls.append(kwargs) or 789)
+
+    result = actions.mark_for_relearn("SKU-3", "NemoTron 信心不足")
+
+    assert result["status"] == "marked"
+    assert result["insight_id"] == 789
+    assert calls[0]["insight_type"] == "relearn_marker"
+    assert calls[0]["status"] == "pending"