refactor(backend): update and add service layers

- services/ontology.py: update ontology service
- services/orchestrator.py: update orchestrator service
- services/user_agent.py: update user agent service
- services/settings.py: update settings service
- services/expense_claims.py: update expense claims service
- services/agent_conversations.py: add new agent conversations service

server/src/app/services/ontology.py

@@ -120,6 +120,24 @@ EXPLAIN_KEYWORDS = ("为什么", "依据", "原因", "怎么处理", "是否可
 COMPARE_KEYWORDS = ("对比", "比较", "相比", "差异", "变化")
 RISK_KEYWORDS = ("风险", "异常", "重复", "超标", "超预算", "逾期", "验真", "巡检")
 DRAFT_KEYWORDS = ("生成", "草稿", "起草", "拟一份", "创建", "发起", "准备")
+DRAFT_FOLLOW_UP_KEYWORDS = (
+    "继续",
+    "补充",
+    "补一下",
+    "修改",
+    "改成",
+    "改为",
+    "换成",
+    "更新",
+    "确认",
+    "提交",
+    "保存",
+    "客户是",
+    "地点是",
+    "金额是",
+    "日期是",
+    "时间是",
+)
 OPERATE_KEYWORDS = (
     "直接付款",
     "帮我付款",
@@ -200,6 +218,7 @@ STATUS_KEYWORDS = {
 }
 
 PRIVILEGED_ROLE_CODES = {"manager", "finance", "approver", "executive"}
+CONTEXTUAL_SCENARIOS = {"expense", "accounts_receivable", "accounts_payable", "knowledge"}
 
 
 @dataclass(slots=True)
@@ -289,12 +308,17 @@ class SemanticOntologyService:
             raise ValueError("query 不能为空。")
 
         AgentFoundationService(self.db).ensure_foundation_ready()
+        context_json = payload.context_json or {}
         reference = self._load_reference_catalog()
         compact_query = self._compact(query)
 
         entities = self._extract_entities(query, compact_query, reference)
         rule_scenario, scenario_score = self._detect_scenario(compact_query)
         time_range, _time_score = self._extract_time_range(query, compact_query)
+        context_scenario = self._resolve_context_scenario(context_json)
+        if rule_scenario == "unknown" and context_scenario is not None:
+            rule_scenario = context_scenario
+            scenario_score = max(scenario_score, 0.14)
         if rule_scenario == "unknown":
             inferred_scenario = self._infer_scenario_from_entities(entities)
             if inferred_scenario is not None:
@@ -316,6 +340,17 @@ class SemanticOntologyService:
             entities=entities,
             time_range=time_range,
         )
+        if self._should_inherit_expense_draft(
+            compact_query,
+            scenario=rule_scenario,
+            entities=entities,
+            time_range=time_range,
+            context_json=context_json,
+        ):
+            rule_scenario = "expense"
+            rule_intent = "draft"
+            scenario_score = max(scenario_score, 0.18)
+            intent_score = max(intent_score, 0.18)
         metrics = self._extract_metrics(compact_query)
         constraints = self._extract_constraints(compact_query, entities)
         model_parse = self._parse_with_model(
@@ -353,7 +388,7 @@ class SemanticOntologyService:
                 intent=intent,
                 entities=entities,
                 time_range=time_range,
-                context_json=payload.context_json or {},
+                context_json=context_json,
             )
         )
         ambiguity = self._normalize_short_text_list(
@@ -362,7 +397,7 @@ class SemanticOntologyService:
         risk_flags = self._extract_risk_flags(compact_query, scenario)
         permission = self._resolve_permission(
             compact_query,
-            payload.context_json or {},
+            context_json,
             intent,
         )
 
@@ -524,6 +559,13 @@ class SemanticOntologyService:
     def _compact(text: str) -> str:
         return re.sub(r"\s+", "", text).lower()
 
+    @staticmethod
+    def _resolve_context_scenario(context_json: dict[str, Any]) -> str | None:
+        value = str(context_json.get("conversation_scenario") or "").strip()
+        if value in CONTEXTUAL_SCENARIOS:
+            return value
+        return None
+
     def _detect_scenario(self, compact_query: str) -> tuple[str, float]:
         scores = {key: 0.0 for key in SCENARIO_KEYWORDS}
         for scenario, keywords in SCENARIO_KEYWORDS.items():
@@ -581,6 +623,68 @@ class SemanticOntologyService:
             return "draft", 0.22
         return "query", 0.10
 
+    @staticmethod
+    def _looks_like_follow_up_message(compact_query: str) -> bool:
+        if not compact_query:
+            return False
+        if any(keyword in compact_query for keyword in DRAFT_FOLLOW_UP_KEYWORDS):
+            return True
+        if compact_query.startswith(("那", "这", "它", "这个", "那个")):
+            return True
+
+        has_domain_keyword = any(
+            keyword in compact_query
+            for keyword, _weight in (
+                *SCENARIO_KEYWORDS["expense"],
+                *SCENARIO_KEYWORDS["accounts_receivable"],
+                *SCENARIO_KEYWORDS["accounts_payable"],
+                *SCENARIO_KEYWORDS["knowledge"],
+            )
+        )
+        return len(compact_query) <= 12 and not has_domain_keyword
+
+    def _should_inherit_expense_draft(
+        self,
+        compact_query: str,
+        *,
+        scenario: str,
+        entities: list[OntologyEntity],
+        time_range: OntologyTimeRange,
+        context_json: dict[str, Any],
+    ) -> bool:
+        context_scenario = self._resolve_context_scenario(context_json)
+        draft_claim_id = str(context_json.get("draft_claim_id") or "").strip()
+        if context_scenario != "expense" and not draft_claim_id:
+            return False
+
+        if any(keyword in compact_query for keyword in DRAFT_FOLLOW_UP_KEYWORDS):
+            return True
+        if self._looks_like_expense_narrative(
+            compact_query,
+            scenario="expense",
+            entities=entities,
+            time_range=time_range,
+        ):
+            return True
+        if self._looks_like_follow_up_message(compact_query):
+            return True
+
+        if any(keyword in compact_query for keyword in OPERATE_KEYWORDS):
+            return False
+        if any(keyword in compact_query for keyword in COMPARE_KEYWORDS + RISK_KEYWORDS):
+            return False
+        if any(keyword in compact_query for keyword in QUERY_KEYWORDS):
+            return False
+
+        return bool(
+            draft_claim_id
+            and any(
+                item.type
+                in {"amount", "customer", "employee", "expense_type", "project", "invoice"}
+                for item in entities
+            )
+        )
+
     @staticmethod
     def _is_generic_expense_prompt(compact_query: str) -> bool:
         return compact_query in GENERIC_EXPENSE_PROMPTS
@@ -670,6 +774,11 @@ class SemanticOntologyService:
                 "ocr_documents": payload.context_json.get("ocr_documents", []),
                 "request_context": payload.context_json.get("request_context"),
                 "role_codes": payload.context_json.get("role_codes", []),
+                "conversation_id": payload.context_json.get("conversation_id"),
+                "conversation_scenario": payload.context_json.get("conversation_scenario"),
+                "conversation_intent": payload.context_json.get("conversation_intent"),
+                "draft_claim_id": payload.context_json.get("draft_claim_id"),
+                "conversation_history": payload.context_json.get("conversation_history", []),
             },
             "rule_candidates": {
                 "scenario": fallback_scenario,
@@ -690,6 +799,8 @@ class SemanticOntologyService:
             "意图 intent 只能是：query, explain, compare, risk_check, draft, operate。"
             "如果用户是在描述一笔待处理费用、待报销事项、上传票据或希望整理报销，"
             "即使没有明确说“生成草稿”，也优先使用 expense + draft。"
+            "如果提供了 conversation_history，必须把最近轮次作为当前追问的上下文，"
+            "正确理解“这个”“那笔”“改成 800”“继续补充”这类省略表达。"
             "出现“客户”不等于应收，出现“供应商”不等于应付，必须结合动作词和业务目标判断。"
             "只有明确查询、统计、列出、多少、明细、对比时才优先使用 query 或 compare。"
             "附件名称和 OCR 摘要只作为辅助证据，不能编造未出现的事实。"