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/agent_conversations.py

@@ -0,0 +1,398 @@
+from __future__ import annotations
+
+import uuid
+from datetime import UTC, datetime, timedelta
+from typing import Any
+
+from sqlalchemy import select
+from sqlalchemy.orm import Session
+
+from app.models.agent_conversation import AgentConversation, AgentConversationMessage
+from app.services.settings import SettingsService
+
+STATEFUL_CONTEXT_KEYS = (
+    "entry_source",
+    "request_context",
+    "attachment_names",
+    "attachment_count",
+    "ocr_summary",
+    "ocr_documents",
+)
+DEFAULT_CONVERSATION_RETENTION_DAYS = 3
+
+
+class AgentConversationService:
+    def __init__(self, db: Session) -> None:
+        self.db = db
+
+    def get_or_create_conversation(
+        self,
+        *,
+        conversation_id: str | None,
+        user_id: str | None,
+        source: str,
+        context_json: dict[str, Any],
+    ) -> AgentConversation:
+        self.prune_expired_conversations()
+
+        normalized_id = str(conversation_id or "").strip()
+        normalized_user_id = str(user_id or "").strip() or None
+        conversation = self.get_conversation(normalized_id) if normalized_id else None
+        if conversation is not None and conversation.user_id != normalized_user_id:
+            normalized_id = ""
+            conversation = None
+
+        if conversation is None:
+            conversation = AgentConversation(
+                conversation_id=normalized_id or f"conv_{uuid.uuid4().hex[:16]}",
+                user_id=normalized_user_id,
+                source=source,
+                entry_source=str(context_json.get("entry_source") or "").strip() or None,
+                title=self._resolve_title(context_json),
+                state_json=self._extract_state_json(context_json),
+            )
+            self.db.add(conversation)
+            self.db.commit()
+            self.db.refresh(conversation)
+            return conversation
+
+        if not conversation.user_id and normalized_user_id:
+            conversation.user_id = normalized_user_id
+        if not conversation.entry_source:
+            conversation.entry_source = str(context_json.get("entry_source") or "").strip() or None
+        if not conversation.title:
+            conversation.title = self._resolve_title(context_json)
+        conversation.state_json = self._merge_state_json(
+            conversation.state_json,
+            self._extract_state_json(context_json),
+        )
+        self.db.add(conversation)
+        self.db.commit()
+        self.db.refresh(conversation)
+        return conversation
+
+    def prune_expired_conversations(
+        self,
+        *,
+        retention_days: int | None = None,
+    ) -> int:
+        resolved_retention_days = retention_days or self._resolve_retention_days()
+        cutoff = datetime.now(UTC) - timedelta(days=max(1, resolved_retention_days))
+        stmt = select(AgentConversation).where(AgentConversation.updated_at < cutoff)
+        expired_conversations = list(self.db.scalars(stmt).all())
+        if not expired_conversations:
+            return 0
+
+        for conversation in expired_conversations:
+            self.db.delete(conversation)
+
+        self.db.commit()
+        return len(expired_conversations)
+
+    def _resolve_retention_days(self) -> int:
+        try:
+            settings_row, _ = SettingsService(self.db).ensure_settings_ready()
+            configured_days = int(
+                getattr(
+                    settings_row,
+                    "conversation_retention_days",
+                    DEFAULT_CONVERSATION_RETENTION_DAYS,
+                )
+                or DEFAULT_CONVERSATION_RETENTION_DAYS
+            )
+            return max(1, min(configured_days, 10))
+        except Exception:
+            self.db.rollback()
+            return DEFAULT_CONVERSATION_RETENTION_DAYS
+
+    def get_conversation(self, conversation_id: str) -> AgentConversation | None:
+        normalized_id = str(conversation_id or "").strip()
+        if not normalized_id:
+            return None
+        stmt = select(AgentConversation).where(AgentConversation.conversation_id == normalized_id)
+        return self.db.scalar(stmt)
+
+    def get_latest_conversation_for_user(
+        self,
+        *,
+        user_id: str | None,
+        source: str | None = "user_message",
+    ) -> AgentConversation | None:
+        self.prune_expired_conversations()
+
+        normalized_user_id = str(user_id or "").strip()
+        if not normalized_user_id:
+            return None
+
+        stmt = select(AgentConversation).where(AgentConversation.user_id == normalized_user_id)
+        if source:
+            stmt = stmt.where(AgentConversation.source == source)
+        stmt = stmt.order_by(AgentConversation.updated_at.desc(), AgentConversation.created_at.desc())
+        return self.db.scalar(stmt.limit(1))
+
+    def hydrate_context_json(
+        self,
+        *,
+        conversation: AgentConversation,
+        context_json: dict[str, Any],
+        history_limit: int = 8,
+    ) -> dict[str, Any]:
+        merged = dict(context_json or {})
+        state_json = dict(conversation.state_json or {})
+
+        merged["conversation_id"] = conversation.conversation_id
+        merged["conversation_history"] = self.list_message_history(
+            conversation.conversation_id,
+            limit=history_limit,
+        )
+        if conversation.last_scenario:
+            merged.setdefault("conversation_scenario", conversation.last_scenario)
+        if conversation.last_intent:
+            merged.setdefault("conversation_intent", conversation.last_intent)
+        if conversation.draft_claim_id and not str(merged.get("draft_claim_id") or "").strip():
+            merged["draft_claim_id"] = conversation.draft_claim_id
+        merged["conversation_state"] = state_json
+
+        for key in STATEFUL_CONTEXT_KEYS:
+            if self._is_empty_value(merged.get(key)) and not self._is_empty_value(state_json.get(key)):
+                merged[key] = state_json.get(key)
+
+        return merged
+
+    def append_message(
+        self,
+        *,
+        conversation_id: str,
+        role: str,
+        content: str,
+        run_id: str | None = None,
+        message_json: dict[str, Any] | None = None,
+    ) -> AgentConversationMessage | None:
+        normalized_content = str(content or "").strip()
+        if not normalized_content:
+            return None
+
+        conversation = self.get_conversation(conversation_id)
+        if conversation is None:
+            return None
+
+        message = AgentConversationMessage(
+            conversation_id=conversation_id,
+            run_id=run_id,
+            role=str(role or "user").strip() or "user",
+            content=normalized_content,
+            message_json=message_json or {},
+            created_at=datetime.now(UTC),
+        )
+        conversation.message_count = int(conversation.message_count or 0) + 1
+        if role == "user" and not conversation.title:
+            conversation.title = normalized_content[:48]
+        conversation.updated_at = datetime.now(UTC)
+        self.db.add(message)
+        self.db.add(conversation)
+        self.db.commit()
+        self.db.refresh(message)
+        return message
+
+    def list_message_history(
+        self,
+        conversation_id: str,
+        *,
+        limit: int = 8,
+    ) -> list[dict[str, Any]]:
+        normalized_id = str(conversation_id or "").strip()
+        if not normalized_id or limit <= 0:
+            return []
+
+        stmt = (
+            select(AgentConversationMessage)
+            .where(AgentConversationMessage.conversation_id == normalized_id)
+            .order_by(AgentConversationMessage.created_at.desc())
+            .limit(limit)
+        )
+        messages = list(self.db.scalars(stmt).all())
+        messages.reverse()
+        return [
+            {
+                "role": item.role,
+                "content": item.content,
+                "run_id": item.run_id,
+                "created_at": item.created_at.isoformat() if item.created_at else None,
+            }
+            for item in messages
+        ]
+
+    def list_messages(
+        self,
+        conversation_id: str,
+        *,
+        limit: int | None = None,
+    ) -> list[AgentConversationMessage]:
+        normalized_id = str(conversation_id or "").strip()
+        if not normalized_id:
+            return []
+
+        stmt = (
+            select(AgentConversationMessage)
+            .where(AgentConversationMessage.conversation_id == normalized_id)
+            .order_by(AgentConversationMessage.created_at.asc(), AgentConversationMessage.id.asc())
+        )
+        if limit and limit > 0:
+            stmt = stmt.limit(limit)
+        return list(self.db.scalars(stmt).all())
+
+    def update_state(
+        self,
+        *,
+        conversation_id: str,
+        run_id: str | None,
+        scenario: str | None,
+        intent: str | None,
+        context_json: dict[str, Any],
+        draft_payload: dict[str, Any] | None = None,
+    ) -> AgentConversation | None:
+        conversation = self.get_conversation(conversation_id)
+        if conversation is None:
+            return None
+
+        conversation.last_run_id = str(run_id or "").strip() or conversation.last_run_id
+        conversation.last_scenario = str(scenario or "").strip() or conversation.last_scenario
+        conversation.last_intent = str(intent or "").strip() or conversation.last_intent
+        if draft_payload and str(draft_payload.get("claim_id") or "").strip():
+            conversation.draft_claim_id = str(draft_payload["claim_id"]).strip()
+
+        next_state = self._merge_state_json(
+            conversation.state_json,
+            self._extract_state_json(context_json),
+        )
+        if draft_payload:
+            if str(draft_payload.get("claim_id") or "").strip():
+                next_state["draft_claim_id"] = str(draft_payload["claim_id"]).strip()
+            if str(draft_payload.get("claim_no") or "").strip():
+                next_state["draft_claim_no"] = str(draft_payload["claim_no"]).strip()
+            if str(draft_payload.get("status") or "").strip():
+                next_state["draft_status"] = str(draft_payload["status"]).strip()
+        conversation.state_json = next_state
+        conversation.updated_at = datetime.now(UTC)
+
+        self.db.add(conversation)
+        self.db.commit()
+        self.db.refresh(conversation)
+        return conversation
+
+    def delete_user_conversations(
+        self,
+        *,
+        user_id: str | None,
+        source: str | None = "user_message",
+    ) -> int:
+        normalized_user_id = str(user_id or "").strip()
+        if not normalized_user_id:
+            return 0
+
+        stmt = select(AgentConversation).where(AgentConversation.user_id == normalized_user_id)
+        if source:
+            stmt = stmt.where(AgentConversation.source == source)
+        conversations = list(self.db.scalars(stmt).all())
+        if not conversations:
+            return 0
+
+        for conversation in conversations:
+            self.db.delete(conversation)
+
+        self.db.commit()
+        return len(conversations)
+
+    def serialize_conversation(
+        self,
+        conversation: AgentConversation,
+        *,
+        include_messages: bool = True,
+        message_limit: int | None = None,
+    ) -> dict[str, Any]:
+        payload = {
+            "conversation_id": conversation.conversation_id,
+            "user_id": conversation.user_id,
+            "source": conversation.source,
+            "entry_source": conversation.entry_source,
+            "title": conversation.title,
+            "last_run_id": conversation.last_run_id,
+            "last_scenario": conversation.last_scenario,
+            "last_intent": conversation.last_intent,
+            "draft_claim_id": conversation.draft_claim_id,
+            "state_json": dict(conversation.state_json or {}),
+            "message_count": int(conversation.message_count or 0),
+            "updated_at": conversation.updated_at,
+            "messages": [],
+        }
+        if include_messages:
+            payload["messages"] = [
+                self.serialize_message(item)
+                for item in self.list_messages(conversation.conversation_id, limit=message_limit)
+            ]
+        return payload
+
+    @staticmethod
+    def serialize_message(message: AgentConversationMessage) -> dict[str, Any]:
+        return {
+            "id": message.id,
+            "role": message.role,
+            "content": message.content,
+            "run_id": message.run_id,
+            "message_json": dict(message.message_json or {}),
+            "created_at": message.created_at,
+        }
+
+    @staticmethod
+    def _is_empty_value(value: Any) -> bool:
+        if value is None:
+            return True
+        if isinstance(value, str):
+            return not value.strip()
+        if isinstance(value, (list, tuple, set, dict)):
+            return len(value) == 0
+        return False
+
+    @staticmethod
+    def _resolve_title(context_json: dict[str, Any]) -> str | None:
+        request_context = context_json.get("request_context")
+        if isinstance(request_context, dict):
+            for key in ("reason", "title", "id"):
+                value = str(request_context.get(key) or "").strip()
+                if value:
+                    return value[:200]
+        return None
+
+    @staticmethod
+    def _extract_state_json(context_json: dict[str, Any]) -> dict[str, Any]:
+        state_json: dict[str, Any] = {}
+        for key in STATEFUL_CONTEXT_KEYS:
+            value = context_json.get(key)
+            if value is None:
+                continue
+            if isinstance(value, str) and not value.strip():
+                continue
+            if isinstance(value, (list, dict)) and not value:
+                continue
+            state_json[key] = value
+
+        draft_claim_id = str(context_json.get("draft_claim_id") or "").strip()
+        if draft_claim_id:
+            state_json["draft_claim_id"] = draft_claim_id
+        return state_json
+
+    @staticmethod
+    def _merge_state_json(
+        current_state: dict[str, Any] | None,
+        incoming_state: dict[str, Any] | None,
+    ) -> dict[str, Any]:
+        merged = dict(current_state or {})
+        for key, value in (incoming_state or {}).items():
+            if value is None:
+                continue
+            if isinstance(value, str) and not value.strip():
+                continue
+            if isinstance(value, (list, dict)) and not value:
+                continue
+            merged[key] = value
+        return merged

server/src/app/services/expense_claims.py

@@ -40,6 +40,7 @@ class ExpenseClaimService:
         self._ensure_ready()
 
         claim = self._find_target_claim(ontology=ontology, context_json=context_json)
+        is_new_claim = claim is None
         before_json = self._serialize_claim(claim) if claim is not None else None
 
         employee = self._resolve_employee(ontology=ontology, context_json=context_json)
@@ -47,12 +48,30 @@ class ExpenseClaimService:
         occurred_at = self._resolve_occurred_at(ontology)
         expense_type = self._resolve_expense_type(ontology.entities)
         location = self._resolve_location(message=message, context_json=context_json)
-        reason = self._resolve_reason(message=message, context_json=context_json)
+        reason = self._resolve_reason(
+            message=message,
+            context_json=context_json,
+            allow_message_fallback=is_new_claim,
+        )
         attachment_count = self._resolve_attachment_count(context_json)
 
+        final_amount = amount if amount is not None else (claim.amount if claim is not None else Decimal("0.00"))
+        final_occurred_at = (
+            occurred_at if occurred_at is not None else (claim.occurred_at if claim is not None else datetime.now(UTC))
+        )
+        final_expense_type = expense_type or (claim.expense_type if claim is not None else "other")
+        final_location = location or (claim.location if claim is not None else "待补充")
+        final_reason = reason or (claim.reason if claim is not None else "待补充")
+        final_attachment_count = (
+            attachment_count if attachment_count > 0 else int(claim.invoice_count or 0) if claim is not None else 0
+        )
+        final_risk_flags = list(ontology.risk_flags) or (
+            list(claim.risk_flags_json or []) if claim is not None else []
+        )
+
         if claim is None:
             claim = ExpenseClaim(
-                claim_no=self._generate_claim_no(occurred_at),
+                claim_no=self._generate_claim_no(final_occurred_at),
                 employee_id=employee.id if employee is not None else None,
                 employee_name=employee.name if employee is not None else self._resolve_employee_name(
                     ontology=ontology,
@@ -65,16 +84,16 @@ class ExpenseClaimService:
                     context_json=context_json,
                 ),
                 project_code=self._resolve_project_code(ontology.entities),
-                expense_type=expense_type,
-                reason=reason,
-                location=location,
-                amount=amount,
+                expense_type=final_expense_type,
+                reason=final_reason,
+                location=final_location,
+                amount=final_amount,
                 currency="CNY",
-                invoice_count=attachment_count,
-                occurred_at=occurred_at,
+                invoice_count=final_attachment_count,
+                occurred_at=final_occurred_at,
                 status="draft",
                 approval_stage="待补充",
-                risk_flags_json=list(ontology.risk_flags),
+                risk_flags_json=final_risk_flags,
             )
             self.db.add(claim)
         else:
@@ -86,6 +105,7 @@ class ExpenseClaimService:
                     ontology=ontology,
                     context_json=context_json,
                     user_id=user_id,
+                    fallback=claim.employee_name,
                 )
             )
             claim.department_id = employee.organization_unit_id if employee is not None else claim.department_id
@@ -95,24 +115,24 @@ class ExpenseClaimService:
                 fallback=claim.department_name,
             )
             claim.project_code = self._resolve_project_code(ontology.entities) or claim.project_code
-            claim.expense_type = expense_type or claim.expense_type
-            claim.reason = reason
-            claim.location = location
-            claim.amount = amount
-            claim.invoice_count = attachment_count
-            claim.occurred_at = occurred_at
+            claim.expense_type = final_expense_type
+            claim.reason = final_reason
+            claim.location = final_location
+            claim.amount = final_amount
+            claim.invoice_count = final_attachment_count
+            claim.occurred_at = final_occurred_at
             claim.status = "draft"
             claim.approval_stage = "待补充"
-            claim.risk_flags_json = list(ontology.risk_flags)
+            claim.risk_flags_json = final_risk_flags
 
         self.db.flush()
         self._upsert_primary_item(
             claim=claim,
-            occurred_at=occurred_at,
-            expense_type=expense_type,
-            amount=amount,
-            reason=reason,
-            location=location,
+            occurred_at=final_occurred_at,
+            expense_type=final_expense_type,
+            amount=final_amount,
+            reason=final_reason,
+            location=final_location,
             attachment_names=self._resolve_attachment_names(context_json),
         )
         self.db.commit()
@@ -130,7 +150,7 @@ class ExpenseClaimService:
 
         return {
             "message": (
-                f"已创建报销草稿 {claim.claim_no}，当前状态为 draft。"
+                f"已{'创建' if is_new_claim else '更新'}报销草稿 {claim.claim_no}，当前状态为 draft。"
                 "你可以继续补充费用明细、客户单位和票据附件。"
             ),
             "draft_only": True,
@@ -229,6 +249,7 @@ class ExpenseClaimService:
         ontology: OntologyParseResult,
         context_json: dict[str, Any],
         user_id: str | None,
+        fallback: str = "待补充",
     ) -> str:
         for item in ontology.entities:
             if item.type == "employee" and item.value.strip():
@@ -237,7 +258,7 @@ class ExpenseClaimService:
             value = str(context_json.get(key) or "").strip()
             if value:
                 return value
-        return str(user_id or "待补充").strip() or "待补充"
+        return str(user_id or fallback).strip() or fallback
 
     @staticmethod
     def _resolve_department_name(
@@ -270,26 +291,33 @@ class ExpenseClaimService:
         return None
 
     @staticmethod
-    def _resolve_expense_type(entities: list[OntologyEntity]) -> str:
+    def _resolve_expense_type(entities: list[OntologyEntity]) -> str | None:
         for item in entities:
             if item.type == "expense_type":
                 normalized = item.normalized_value.strip()
                 if normalized:
                     return normalized
-        return "other"
+        return None
 
     @staticmethod
-    def _resolve_reason(*, message: str, context_json: dict[str, Any]) -> str:
+    def _resolve_reason(
+        *,
+        message: str,
+        context_json: dict[str, Any],
+        allow_message_fallback: bool,
+    ) -> str | None:
         request_context = context_json.get("request_context")
         if isinstance(request_context, dict):
             for key in ("reason", "title"):
                 value = str(request_context.get(key) or "").strip()
                 if value:
                     return value
-        return str(message or "").strip()[:500] or "待补充"
+        if not allow_message_fallback:
+            return None
+        return str(message or "").strip()[:500] or None
 
     @staticmethod
-    def _resolve_location(*, message: str, context_json: dict[str, Any]) -> str:
+    def _resolve_location(*, message: str, context_json: dict[str, Any]) -> str | None:
         request_context = context_json.get("request_context")
         if isinstance(request_context, dict):
             for key in ("city", "location"):
@@ -299,10 +327,10 @@ class ExpenseClaimService:
         compact = str(message or "").replace(" ", "")
         if "客户现场" in compact:
             return "客户现场"
-        return "待补充"
+        return None
 
     @staticmethod
-    def _resolve_occurred_at(ontology: OntologyParseResult) -> datetime:
+    def _resolve_occurred_at(ontology: OntologyParseResult) -> datetime | None:
         start_date = ontology.time_range.start_date
         if start_date:
             try:
@@ -310,10 +338,10 @@ class ExpenseClaimService:
                 return datetime(parsed.year, parsed.month, parsed.day, tzinfo=UTC)
             except ValueError:
                 pass
-        return datetime.now(UTC)
+        return None
 
     @staticmethod
-    def _resolve_amount(entities: list[OntologyEntity]) -> Decimal:
+    def _resolve_amount(entities: list[OntologyEntity]) -> Decimal | None:
         for item in entities:
             if item.type != "amount" or item.role == "threshold":
                 continue
@@ -321,7 +349,7 @@ class ExpenseClaimService:
                 return Decimal(item.normalized_value).quantize(Decimal("0.01"))
             except (InvalidOperation, ValueError):
                 continue
-        return Decimal("0.00")
+        return None
 
     @staticmethod
     def _resolve_attachment_names(context_json: dict[str, Any]) -> list[str]:

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。"

server/src/app/services/orchestrator.py

@@ -32,6 +32,7 @@ from app.schemas.orchestrator import (
 )
 from app.schemas.user_agent import UserAgentRequest, UserAgentResponse
 from app.services.agent_assets import AgentAssetService
+from app.services.agent_conversations import AgentConversationService
 from app.services.expense_claims import ExpenseClaimService
 from app.services.agent_foundation import AgentFoundationService
 from app.services.agent_runs import AgentRunService
@@ -62,6 +63,7 @@ class OrchestratorService:
     def __init__(self, db: Session) -> None:
         self.db = db
         self.asset_service = AgentAssetService(db)
+        self.conversation_service = AgentConversationService(db)
         self.expense_claim_service = ExpenseClaimService(db)
         self.run_service = AgentRunService(db)
         self.ontology_service = SemanticOntologyService(db)
@@ -69,10 +71,28 @@ class OrchestratorService:
 
     def run(self, payload: OrchestratorRequest) -> OrchestratorResponse:
         AgentFoundationService(self.db).ensure_foundation_ready()
+        context_json = dict(payload.context_json or {})
+        conversation_id = str(payload.conversation_id or "").strip() or None
+        conversation = None
+        if payload.source == AgentRunSource.USER_MESSAGE.value:
+            conversation = self.conversation_service.get_or_create_conversation(
+                conversation_id=conversation_id,
+                user_id=payload.user_id,
+                source=payload.source,
+                context_json=context_json,
+            )
+            conversation_id = conversation.conversation_id
+            context_json = self.conversation_service.hydrate_context_json(
+                conversation=conversation,
+                context_json=context_json,
+            )
+
         route_json: dict[str, Any] = {
             "orchestrated_by": AgentName.ORCHESTRATOR.value,
             "stage": "created",
         }
+        if conversation_id:
+            route_json["conversation_id"] = conversation_id
         run = self.run_service.create_run(
             agent=AgentName.ORCHESTRATOR.value,
             source=payload.source,
@@ -87,15 +107,27 @@ class OrchestratorService:
 
         try:
             message, task_asset = self._resolve_message(payload)
+            if conversation is not None:
+                self.conversation_service.append_message(
+                    conversation_id=conversation.conversation_id,
+                    role="user",
+                    content=message,
+                    run_id=run.run_id,
+                    message_json={
+                        "attachment_names": context_json.get("attachment_names", []),
+                        "attachment_count": context_json.get("attachment_count", 0),
+                        "ocr_summary": context_json.get("ocr_summary", ""),
+                    },
+                )
             ontology = self.ontology_service.parse_for_run(
                 OntologyParseRequest(
                     query=message,
                     user_id=payload.user_id,
-                    context_json=payload.context_json,
+                    context_json=context_json,
                 ),
                 run_id=run.run_id,
             )
-            if payload.context_json.get("simulate_orchestrator_exception"):
+            if context_json.get("simulate_orchestrator_exception"):
                 raise RuntimeError("simulated orchestrator exception")
             selected_agent, route_reason = self._select_agent(payload, ontology)
             capabilities = self._select_capabilities(
@@ -159,6 +191,7 @@ class OrchestratorService:
                     capabilities=capabilities,
                     requires_confirmation=requires_confirmation,
                     task_asset=task_asset,
+                    context_json=context_json,
                 )
             else:
                 outcome = self._execute_user_agent(
@@ -167,6 +200,7 @@ class OrchestratorService:
                     ontology=ontology,
                     capabilities=capabilities,
                     requires_confirmation=requires_confirmation,
+                    context_json=context_json,
                 )
 
             final_status = (
@@ -176,10 +210,19 @@ class OrchestratorService:
                 and ontology.permission.level == AgentPermissionLevel.APPROVAL_REQUIRED.value
                 else outcome.status
             )
+            response_status = self._normalize_response_status(final_status)
             result_message = (
                 str(outcome.result.get("message", "")).strip()
                 or "Orchestrator 执行完成。"
             )
+            trace_summary = OrchestratorTraceSummary(
+                scenario=ontology.scenario,
+                intent=ontology.intent,
+                tool_count=outcome.tool_count,
+                failed_tool_count=outcome.failed_tool_count,
+                selected_capability_codes=selected_capability_codes,
+                degraded=outcome.degraded,
+            )
             self.run_service.update_run(
                 run.run_id,
                 agent=selected_agent or AgentName.ORCHESTRATOR.value,
@@ -195,22 +238,51 @@ class OrchestratorService:
                 error_message=None,
                 finished_at=datetime.now(UTC),
             )
+            if conversation is not None and conversation_id:
+                draft_payload = outcome.result.get("draft_payload")
+                self.conversation_service.update_state(
+                    conversation_id=conversation_id,
+                    run_id=run.run_id,
+                    scenario=ontology.scenario,
+                    intent=ontology.intent,
+                    context_json=context_json,
+                    draft_payload=draft_payload if isinstance(draft_payload, dict) else None,
+                )
+                self.conversation_service.append_message(
+                    conversation_id=conversation_id,
+                    role="assistant",
+                    content=result_message,
+                    run_id=run.run_id,
+                    message_json={
+                        "status": final_status,
+                        "scenario": ontology.scenario,
+                        "intent": ontology.intent,
+                        "attachment_names": context_json.get("attachment_names", []),
+                        "attachment_count": context_json.get("attachment_count", 0),
+                        "draft_payload": draft_payload if isinstance(draft_payload, dict) else None,
+                        "orchestrator_payload": {
+                            "run_id": run.run_id,
+                            "conversation_id": conversation_id,
+                            "selected_agent": selected_agent,
+                            "route_reason": route_reason,
+                            "permission_level": ontology.permission.level,
+                            "status": response_status,
+                            "requires_confirmation": requires_confirmation,
+                            "trace_summary": trace_summary.model_dump(),
+                            "result": outcome.result,
+                        },
+                    },
+                )
             return OrchestratorResponse(
                 run_id=run.run_id,
+                conversation_id=conversation_id,
                 selected_agent=selected_agent,
                 route_reason=route_reason,
                 permission_level=ontology.permission.level,
-                status=self._normalize_response_status(final_status),
+                status=response_status,
                 result=outcome.result,
                 requires_confirmation=requires_confirmation,
-                trace_summary=OrchestratorTraceSummary(
-                    scenario=ontology.scenario,
-                    intent=ontology.intent,
-                    tool_count=outcome.tool_count,
-                    failed_tool_count=outcome.failed_tool_count,
-                    selected_capability_codes=selected_capability_codes,
-                    degraded=outcome.degraded,
-                ),
+                trace_summary=trace_summary,
             )
         except Exception as exc:
             logger.exception("Orchestrator run failed run_id=%s", run.run_id)
@@ -223,8 +295,25 @@ class OrchestratorService:
                 error_message=str(exc),
                 finished_at=datetime.now(UTC),
             )
+            if conversation is not None and conversation_id:
+                self.conversation_service.update_state(
+                    conversation_id=conversation_id,
+                    run_id=run.run_id,
+                    scenario=None,
+                    intent=None,
+                    context_json=context_json,
+                    draft_payload=None,
+                )
+                self.conversation_service.append_message(
+                    conversation_id=conversation_id,
+                    role="assistant",
+                    content=f"Orchestrator 执行失败：{exc}",
+                    run_id=run.run_id,
+                    message_json={"status": AgentRunStatus.FAILED.value},
+                )
             return OrchestratorResponse(
                 run_id=run.run_id,
+                conversation_id=conversation_id,
                 selected_agent=None,
                 route_reason="orchestrator_exception",
                 permission_level=AgentPermissionLevel.READ.value,
@@ -336,6 +425,7 @@ class OrchestratorService:
         ontology: OntologyParseResult,
         capabilities: dict[str, list[AgentAssetListItem | AgentAssetRead]],
         requires_confirmation: bool,
+        context_json: dict[str, Any],
     ) -> ExecutionOutcome:
         selected_capability_codes = self._flatten_capability_codes(capabilities)
         if requires_confirmation:
@@ -347,7 +437,7 @@ class OrchestratorService:
                     "message": payload.message,
                     "permission_level": ontology.permission.level,
                 },
-                context_json=payload.context_json,
+                context_json=context_json,
                 executor=lambda: {
                     "confirmation_title": "操作需要确认",
                     "message": f"{ontology.permission.reason} 当前仅返回确认摘要，不直接执行动作。",
@@ -372,7 +462,7 @@ class OrchestratorService:
                 tool_type=AgentToolType.DATABASE.value,
                 tool_name=self._database_tool_name(ontology.scenario),
                 request_json=self._build_ontology_json(ontology),
-                context_json=payload.context_json,
+                context_json=context_json,
                 executor=lambda: self._build_database_answer(ontology),
                 fallback_factory=lambda exc: {
                     "message": f"数据库查询暂时不可用，已返回降级说明：{exc}",
@@ -386,7 +476,7 @@ class OrchestratorService:
                         user_id=payload.user_id,
                         message=payload.message or "",
                         ontology=ontology,
-                        context_json=payload.context_json,
+                        context_json=context_json,
                         tool_payload=tool_payload,
                         selected_capability_codes=selected_capability_codes,
                         degraded=degraded,
@@ -409,7 +499,7 @@ class OrchestratorService:
                 tool_type=AgentToolType.DATABASE.value,
                 tool_name="knowledge.search",
                 request_json=self._build_ontology_json(ontology),
-                context_json=payload.context_json,
+                context_json=context_json,
                 executor=lambda: self._build_knowledge_answer(ontology, capabilities),
                 fallback_factory=lambda exc: {
                     "message": f"知识检索暂时不可用，建议稍后重试：{exc}",
@@ -423,7 +513,7 @@ class OrchestratorService:
                         user_id=payload.user_id,
                         message=payload.message or "",
                         ontology=ontology,
-                        context_json=payload.context_json,
+                        context_json=context_json,
                         tool_payload=tool_payload,
                         selected_capability_codes=selected_capability_codes,
                         degraded=degraded,
@@ -446,7 +536,7 @@ class OrchestratorService:
                 tool_type=AgentToolType.RULE_ENGINE.value,
                 tool_name=self._rule_tool_name(capabilities),
                 request_json=self._build_ontology_json(ontology),
-                context_json=payload.context_json,
+                context_json=context_json,
                 executor=lambda: self._build_rule_answer(ontology),
                 fallback_factory=lambda exc: {
                     "message": f"规则检查暂时不可用，已返回人工复核建议：{exc}",
@@ -460,7 +550,7 @@ class OrchestratorService:
                         user_id=payload.user_id,
                         message=payload.message or "",
                         ontology=ontology,
-                        context_json=payload.context_json,
+                        context_json=context_json,
                         tool_payload=tool_payload,
                         selected_capability_codes=selected_capability_codes,
                         degraded=degraded,
@@ -499,7 +589,7 @@ class OrchestratorService:
                 user_id=payload.user_id,
                 message=payload.message or "",
                 ontology=ontology,
-                context_json=payload.context_json,
+                context_json=context_json,
             )
             fallback_factory = lambda exc: {
                 "message": f"报销草稿落库失败，请稍后再试：{exc}",
@@ -511,7 +601,7 @@ class OrchestratorService:
             tool_type=tool_type,
             tool_name=tool_name,
             request_json=self._build_ontology_json(ontology),
-            context_json=payload.context_json,
+            context_json=context_json,
             executor=executor,
             fallback_factory=fallback_factory,
         )
@@ -522,7 +612,7 @@ class OrchestratorService:
                     user_id=payload.user_id,
                     message=payload.message or "",
                     ontology=ontology,
-                    context_json=payload.context_json,
+                    context_json=context_json,
                     tool_payload=tool_payload,
                     selected_capability_codes=selected_capability_codes,
                     degraded=degraded,
@@ -548,6 +638,7 @@ class OrchestratorService:
         capabilities: dict[str, list[AgentAssetListItem | AgentAssetRead]],
         requires_confirmation: bool,
         task_asset: AgentAssetRead | None,
+        context_json: dict[str, Any],
     ) -> ExecutionOutcome:
         if requires_confirmation:
             return ExecutionOutcome(
@@ -566,7 +657,7 @@ class OrchestratorService:
             tool_type=AgentToolType.RULE_ENGINE.value,
             tool_name=self._rule_tool_name(capabilities),
             request_json=self._build_ontology_json(ontology),
-            context_json=payload.context_json,
+            context_json=context_json,
             executor=lambda: self._build_rule_answer(ontology),
             fallback_factory=lambda exc: {
                 "message": f"规则巡检失败，已降级为待人工复核：{exc}",
@@ -581,7 +672,7 @@ class OrchestratorService:
                 "task_code": task_asset.code if task_asset is not None else "",
                 "scenario": ontology.scenario,
             },
-            context_json=payload.context_json,
+            context_json=context_json,
             executor=lambda: self._build_mcp_answer(task_asset, ontology),
             fallback_factory=lambda exc: {
                 "message": f"MCP 调用失败，已使用缓存快照降级：{exc}",
@@ -806,6 +897,8 @@ class OrchestratorService:
         }
         if response.draft_payload is not None:
             result["draft_payload"] = response.draft_payload.model_dump()
+        if response.review_payload is not None:
+            result["review_payload"] = response.review_payload.model_dump()
         return result
 
     @staticmethod

server/src/app/services/settings.py

@@ -204,6 +204,7 @@ class SettingsService:
         settings_row.admin_account = payload.adminForm.adminAccount
         settings_row.admin_email = payload.adminForm.adminEmail
         settings_row.session_timeout = payload.adminForm.sessionTimeout
+        settings_row.conversation_retention_days = payload.sessionForm.conversationRetentionDays
         settings_row.notice_email = payload.adminForm.noticeEmail
         settings_row.mfa_enabled = payload.adminForm.mfaEnabled
         settings_row.strong_password = payload.adminForm.strongPassword
@@ -429,6 +430,7 @@ class SettingsService:
             admin_account=admin_account,
             admin_email=admin_email,
             session_timeout=30,
+            conversation_retention_days=3,
             notice_email=admin_email,
             mfa_enabled=True,
             strong_password=True,
@@ -520,6 +522,10 @@ class SettingsService:
 
         if "system_settings" in table_names:
             settings_columns = {column["name"] for column in inspector.get_columns("system_settings")}
+            if "conversation_retention_days" not in settings_columns:
+                migration_statements.append(
+                    "ALTER TABLE system_settings ADD COLUMN conversation_retention_days INTEGER DEFAULT 3"
+                )
             if "onlyoffice_enabled" not in settings_columns:
                 migration_statements.append(
                     "ALTER TABLE system_settings ADD COLUMN onlyoffice_enabled BOOLEAN DEFAULT FALSE"
@@ -600,6 +606,9 @@ class SettingsService:
                 "loginAlertEnabled": settings_row.login_alert_enabled,
                 "adminPasswordConfigured": bool(secrets_row.admin_password_hash),
             },
+            sessionForm={
+                "conversationRetentionDays": settings_row.conversation_retention_days,
+            },
             llmForm={
                 "mainProvider": main_model.provider,
                 "mainModel": main_model.model_name,

server/src/app/services/user_agent.py

@@ -2,14 +2,24 @@ from __future__ import annotations
 
 import json
 import re
+from datetime import UTC, datetime, timedelta
 
+from sqlalchemy import select
 from sqlalchemy.orm import Session
 
 from app.core.agent_enums import AgentAssetStatus, AgentAssetType
+from app.models.financial_record import ExpenseClaim
 from app.schemas.agent_asset import AgentAssetListItem
 from app.schemas.user_agent import (
     UserAgentCitation,
     UserAgentDraftPayload,
+    UserAgentReviewAction,
+    UserAgentReviewClaimGroup,
+    UserAgentReviewDocumentCard,
+    UserAgentReviewDocumentField,
+    UserAgentReviewPayload,
+    UserAgentReviewRiskBrief,
+    UserAgentReviewSlotCard,
     UserAgentRequest,
     UserAgentResponse,
     UserAgentSuggestedAction,
@@ -53,8 +63,32 @@ EXPENSE_TYPE_LABELS = {
     "meal": "餐费",
     "meeting": "会务",
     "entertainment": "招待",
+    "other": "其他",
 }
 
+GROUP_SCENE_LABELS = {
+    "travel": "差旅费",
+    "entertainment": "业务招待费",
+    "meal": "伙食费",
+    "transport": "交通费",
+    "hotel": "住宿费",
+    "other": "其他费用",
+}
+
+SLOT_LABELS = {
+    "expense_type": "报销类型",
+    "customer_name": "客户名称",
+    "time_range": "发生时间",
+    "location": "地点",
+    "merchant_name": "酒店/商户",
+    "amount": "金额",
+    "participants": "参与人员",
+    "attachments": "票据附件",
+}
+
+DATE_TEXT_PATTERN = re.compile(r"(\d{4}[年/-]\d{1,2}[月/-]\d{1,2}日?)")
+AMOUNT_TEXT_PATTERN = re.compile(r"(\d+(?:\.\d+)?)\s*(?:元|万元|万)")
+
 
 class UserAgentService:
     def __init__(self, db: Session) -> None:
@@ -72,23 +106,32 @@ class UserAgentService:
             if payload.ontology.intent == "draft"
             else None
         )
+        review_payload = self._build_review_payload(
+            payload,
+            citations=citations,
+            draft_payload=draft_payload,
+        )
 
         if payload.degraded and payload.tool_payload.get("message"):
             return UserAgentResponse(
                 answer=str(payload.tool_payload["message"]),
                 citations=citations,
                 suggested_actions=suggested_actions,
+                review_payload=review_payload,
                 risk_flags=risk_flags,
                 requires_confirmation=payload.requires_confirmation,
             )
 
-        guided_answer = self._build_guided_answer(payload)
+        guided_answer = None
+        if draft_payload is None or draft_payload.claim_id is None:
+            guided_answer = self._build_guided_answer(payload)
         if guided_answer:
             return UserAgentResponse(
                 answer=guided_answer,
                 citations=citations,
                 suggested_actions=suggested_actions,
                 draft_payload=draft_payload,
+                review_payload=review_payload,
                 risk_flags=risk_flags,
                 requires_confirmation=payload.requires_confirmation,
             )
@@ -98,20 +141,23 @@ class UserAgentService:
             citations=citations,
             draft_payload=draft_payload,
         )
-        answer = self._generate_answer_with_model(
-            payload,
-            citations=citations,
-            suggested_actions=suggested_actions,
-            risk_flags=risk_flags,
-            draft_payload=draft_payload,
-            fallback_answer=fallback_answer,
-        )
+        answer = None
+        if not self._should_skip_model_answer(payload, review_payload):
+            answer = self._generate_answer_with_model(
+                payload,
+                citations=citations,
+                suggested_actions=suggested_actions,
+                risk_flags=risk_flags,
+                draft_payload=draft_payload,
+                fallback_answer=fallback_answer,
+            )
 
         return UserAgentResponse(
             answer=answer or fallback_answer,
             citations=citations,
             suggested_actions=suggested_actions,
             draft_payload=draft_payload,
+            review_payload=review_payload,
             risk_flags=risk_flags,
             requires_confirmation=payload.requires_confirmation,
         )
@@ -129,6 +175,13 @@ class UserAgentService:
         if payload.ontology.intent == "risk_check":
             return self._build_risk_answer(payload, citations)
 
+        if payload.ontology.intent == "draft":
+            tool_message = str(payload.tool_payload.get("message") or "").strip()
+            if tool_message and (
+                str(payload.tool_payload.get("claim_id") or "").strip()
+                or str(payload.tool_payload.get("claim_no") or "").strip()
+            ):
+                return tool_message
         if payload.ontology.intent == "draft" and draft_payload is not None:
             return (
                 f"已生成 {draft_payload.title}，当前仅返回待人工确认的草稿内容，"
@@ -243,6 +296,11 @@ class UserAgentService:
                 "attachment_names": self._resolve_attachment_names(payload),
                 "ocr_summary": payload.context_json.get("ocr_summary", ""),
                 "ocr_documents": payload.context_json.get("ocr_documents", []),
+                "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": self._resolve_conversation_history(payload),
             },
             "tool_payload": payload.tool_payload,
             "citations": [item.model_dump(mode="json") for item in citations],
@@ -267,6 +325,7 @@ class UserAgentService:
             "并明确要求补充费用类型、金额、时间、事由、参与对象或上传票据。"
             "如果上下文里只有附件名称，必须明确说明你只拿到了附件名称，"
             "不能假装已看过图片、PDF 或发票内容。"
+            "如果提供了 conversation_history，必须结合最近轮次理解追问、代词、省略字段和补充信息。"
             "不要声称已经提交、审批、付款、入账或真正执行了任何动作；如果只是建议、草稿或待确认，要明确说清楚。"
             "若给出了风险标签、制度引用或建议动作，可以简洁吸收进回答，但不要新增未提供的事实。"
             "只输出最终给用户看的自然语言，不要输出 JSON、Markdown、标题、"
@@ -447,6 +506,424 @@ class UserAgentService:
             ),
         ]
 
+    def _build_review_payload(
+        self,
+        payload: UserAgentRequest,
+        *,
+        citations: list[UserAgentCitation],
+        draft_payload: UserAgentDraftPayload | None,
+    ) -> UserAgentReviewPayload | None:
+        attachment_count = self._resolve_attachment_count(payload)
+        ocr_documents = self._resolve_ocr_documents(payload)
+        if payload.ontology.scenario != "expense":
+            return None
+        if payload.ontology.intent not in {"draft", "operate"} and attachment_count <= 0 and not ocr_documents:
+            return None
+
+        slot_cards = self._build_review_slot_cards(payload, ocr_documents=ocr_documents)
+        document_cards = self._build_review_document_cards(payload, ocr_documents=ocr_documents)
+        claim_groups = self._build_review_claim_groups(
+            payload,
+            document_cards=document_cards,
+        )
+        risk_briefs = self._build_review_risk_briefs(
+            payload,
+            citations=citations,
+            document_cards=document_cards,
+            claim_groups=claim_groups,
+        )
+        confirmation_actions = self._build_review_confirmation_actions(
+            payload,
+            claim_groups=claim_groups,
+            draft_payload=draft_payload,
+        )
+        intent_summary = self._build_review_intent_summary(
+            payload,
+            slot_cards=slot_cards,
+            claim_groups=claim_groups,
+        )
+
+        return UserAgentReviewPayload(
+            intent_summary=intent_summary,
+            scenario=payload.ontology.scenario,
+            intent=payload.ontology.intent,
+            missing_slots=list(payload.ontology.missing_slots),
+            risk_briefs=risk_briefs,
+            slot_cards=slot_cards,
+            document_cards=document_cards,
+            claim_groups=claim_groups,
+            confirmation_actions=confirmation_actions,
+        )
+
+    def _build_review_slot_cards(
+        self,
+        payload: UserAgentRequest,
+        *,
+        ocr_documents: list[dict[str, object]],
+    ) -> list[UserAgentReviewSlotCard]:
+        first_doc_fields = self._extract_document_fields(ocr_documents[0]) if ocr_documents else {}
+        missing_slots = set(payload.ontology.missing_slots)
+        entity_map = self._collect_entity_values(payload)
+
+        time_value = self._format_time_range(payload)
+        location_value = self._resolve_location_value(payload)
+        merchant_value = self._extract_document_merchant_name(ocr_documents[0]) if ocr_documents else ""
+        customer_value = entity_map.get("customer", "")
+        participants_value = entity_map.get("participants", "")
+        amount_value = entity_map.get("amount")
+        if not amount_value:
+            ocr_total_amount = self._sum_ocr_amounts(ocr_documents)
+            amount_value = f"{ocr_total_amount:.2f}元" if ocr_total_amount > 0 else ""
+        expense_type_code = entity_map.get("expense_type_code", "")
+        expense_type_value = EXPENSE_TYPE_LABELS.get(expense_type_code, entity_map.get("expense_type", ""))
+        if not expense_type_value and ocr_documents:
+            expense_type_value = self._infer_expense_type_from_documents(payload, ocr_documents)
+        attachment_value = (
+            f"{self._resolve_attachment_count(payload)} 份附件"
+            if self._resolve_attachment_count(payload)
+            else ""
+        )
+
+        cards = [
+            self._make_slot_card(
+                key="expense_type",
+                value=expense_type_value,
+                source="user_text" if expense_type_value else "system",
+                confidence=0.9 if expense_type_value else 0.0,
+                missing_slots=missing_slots,
+            ),
+            self._make_slot_card(
+                key="customer_name",
+                value=customer_value,
+                source="user_text" if customer_value else "system",
+                confidence=0.88 if customer_value else 0.0,
+                missing_slots=missing_slots,
+            ),
+            self._make_slot_card(
+                key="time_range",
+                value=time_value,
+                source="user_text" if time_value else "system",
+                confidence=0.9 if time_value else 0.0,
+                missing_slots=missing_slots,
+            ),
+            self._make_slot_card(
+                key="location",
+                value=location_value,
+                source="page_context" if location_value and location_value != "客户现场" else "user_text",
+                confidence=0.82 if location_value else 0.0,
+                required=False,
+                missing_slots=missing_slots,
+            ),
+            self._make_slot_card(
+                key="merchant_name",
+                value=merchant_value,
+                source="ocr" if merchant_value else "system",
+                confidence=0.72 if merchant_value else 0.0,
+                required=False,
+                missing_slots=missing_slots,
+            ),
+            self._make_slot_card(
+                key="amount",
+                value=amount_value,
+                source="user_text" if entity_map.get("amount") else "ocr" if amount_value else "system",
+                confidence=0.92 if amount_value else 0.0,
+                missing_slots=missing_slots,
+            ),
+            self._make_slot_card(
+                key="participants",
+                value=participants_value,
+                source="user_text" if participants_value else "system",
+                confidence=0.8 if participants_value else 0.0,
+                missing_slots=missing_slots,
+            ),
+            self._make_slot_card(
+                key="attachments",
+                value=attachment_value,
+                source="upload" if attachment_value else "system",
+                confidence=1.0 if attachment_value else 0.0,
+                missing_slots=missing_slots,
+            ),
+        ]
+        return cards
+
+    def _build_review_document_cards(
+        self,
+        payload: UserAgentRequest,
+        *,
+        ocr_documents: list[dict[str, object]],
+    ) -> list[UserAgentReviewDocumentCard]:
+        cards: list[UserAgentReviewDocumentCard] = []
+        for index, item in enumerate(ocr_documents, start=1):
+            classified = self._classify_document(item, payload)
+            fields = self._extract_document_fields(item)
+            cards.append(
+                UserAgentReviewDocumentCard(
+                    index=index,
+                    filename=str(item.get("filename") or f"document-{index}"),
+                    document_type=classified["document_type"],
+                    suggested_expense_type=classified["expense_type"],
+                    scene_label=GROUP_SCENE_LABELS.get(
+                        classified["group_code"],
+                        classified["scene_label"],
+                    ),
+                    summary=str(item.get("summary") or item.get("text") or "").strip(),
+                    avg_score=float(item.get("avg_score") or 0.0),
+                    warnings=[str(warning) for warning in item.get("warnings", []) if str(warning).strip()],
+                    fields=[
+                        UserAgentReviewDocumentField(
+                            label=label,
+                            value=value,
+                            source="ocr",
+                        )
+                        for label, value in fields.items()
+                        if str(value).strip()
+                    ],
+                )
+            )
+        return cards
+
+    def _build_review_claim_groups(
+        self,
+        payload: UserAgentRequest,
+        *,
+        document_cards: list[UserAgentReviewDocumentCard],
+    ) -> list[UserAgentReviewClaimGroup]:
+        groups: dict[str, dict[str, object]] = {}
+        for card in document_cards:
+            group_code = self._normalize_group_code(card.suggested_expense_type)
+            bucket = groups.setdefault(
+                group_code,
+                {
+                    "document_indexes": [],
+                    "amount_total": 0.0,
+                    "expense_type": group_code,
+                    "scene_label": GROUP_SCENE_LABELS.get(group_code, "其他费用"),
+                    "reasons": [],
+                },
+            )
+            bucket["document_indexes"].append(card.index)
+            bucket["amount_total"] = float(bucket["amount_total"]) + self._extract_amount_from_card(card)
+            bucket["reasons"].append(f"{card.filename} 识别为 {card.scene_label}")
+
+        if not groups:
+            expense_type_code = self._collect_entity_values(payload).get("expense_type_code", "other")
+            group_code = self._normalize_group_code(expense_type_code)
+            groups[group_code] = {
+                "document_indexes": [],
+                "amount_total": self._resolve_amount_value(payload),
+                "expense_type": expense_type_code or "other",
+                "scene_label": GROUP_SCENE_LABELS.get(group_code, "其他费用"),
+                "reasons": ["当前主要依据用户文本和页面上下文进行分单建议。"],
+            }
+
+        claim_groups: list[UserAgentReviewClaimGroup] = []
+        for index, (group_code, bucket) in enumerate(groups.items(), start=1):
+            title = f"建议报销单 {index}：{bucket['scene_label']}"
+            rationale = (
+                "；".join(dict.fromkeys(str(item) for item in bucket["reasons"]))
+                if bucket["reasons"]
+                else "当前仅有单一场景，无需拆单。"
+            )
+            claim_groups.append(
+                UserAgentReviewClaimGroup(
+                    group_code=group_code,
+                    title=title,
+                    expense_type=str(bucket["expense_type"]),
+                    scene_label=str(bucket["scene_label"]),
+                    document_indexes=list(bucket["document_indexes"]),
+                    amount_total=round(float(bucket["amount_total"]), 2),
+                    rationale=rationale,
+                )
+            )
+        return claim_groups
+
+    def _build_review_risk_briefs(
+        self,
+        payload: UserAgentRequest,
+        *,
+        citations: list[UserAgentCitation],
+        document_cards: list[UserAgentReviewDocumentCard],
+        claim_groups: list[UserAgentReviewClaimGroup],
+    ) -> list[UserAgentReviewRiskBrief]:
+        briefs: list[UserAgentReviewRiskBrief] = []
+        employee_name = self._collect_entity_values(payload).get("employee_name") or str(
+            payload.context_json.get("name") or ""
+        ).strip()
+        if employee_name:
+            since = datetime.now(UTC) - timedelta(days=90)
+            stmt = select(ExpenseClaim).where(
+                ExpenseClaim.employee_name == employee_name,
+                ExpenseClaim.occurred_at >= since,
+            )
+            recent_claims = list(self.db.scalars(stmt).all())
+            if recent_claims:
+                risky_count = sum(1 for item in recent_claims if item.risk_flags_json)
+                draft_count = sum(1 for item in recent_claims if item.status == "draft")
+                briefs.append(
+                    UserAgentReviewRiskBrief(
+                        title="历史报销画像",
+                        level="info",
+                        content=(
+                            f"{employee_name} 最近 90 天共有 {len(recent_claims)} 笔报销，"
+                            f"其中 {risky_count} 笔带风险标记，{draft_count} 笔仍处于草稿态。"
+                        ),
+                    )
+                )
+                current_amount = self._resolve_amount_value(payload)
+                if current_amount > 0:
+                    duplicate_count = sum(
+                        1
+                        for item in recent_claims
+                        if abs(float(item.amount) - current_amount) < 0.01
+                    )
+                    if duplicate_count:
+                        briefs.append(
+                            UserAgentReviewRiskBrief(
+                                title="金额重复预警",
+                                level="warning",
+                                content=(
+                                    f"近 90 天发现 {duplicate_count} 笔金额相同的报销记录，"
+                                    "提交前建议核对是否为重复报销或拆分不当。"
+                                ),
+                            )
+                        )
+
+        if citations:
+            briefs.append(
+                UserAgentReviewRiskBrief(
+                    title="制度注意事项",
+                    level="info",
+                    content=citations[0].excerpt or f"请先核对 {citations[0].title} 的制度要求。",
+                )
+            )
+
+        warning_count = sum(len(item.warnings) for item in document_cards)
+        if warning_count:
+            briefs.append(
+                UserAgentReviewRiskBrief(
+                    title="票据识别提醒",
+                    level="warning",
+                    content=f"当前共有 {warning_count} 条票据识别提示，建议逐张确认 OCR 识别字段。",
+                )
+            )
+
+        if len(claim_groups) > 1:
+            briefs.append(
+                UserAgentReviewRiskBrief(
+                    title="建议拆单",
+                    level="high",
+                    content=f"系统检测到 {len(claim_groups)} 类费用场景，建议拆成多张报销单后再提交。",
+                )
+            )
+
+        return briefs[:4]
+
+    def _build_review_confirmation_actions(
+        self,
+        payload: UserAgentRequest,
+        *,
+        claim_groups: list[UserAgentReviewClaimGroup],
+        draft_payload: UserAgentDraftPayload | None,
+    ) -> list[UserAgentReviewAction]:
+        actions: list[UserAgentReviewAction] = []
+
+        if claim_groups:
+            if len(claim_groups) > 1:
+                actions.append(
+                    UserAgentReviewAction(
+                        label=f"按 {len(claim_groups)} 张报销单生成",
+                        action_type="split_claims",
+                        description="保留当前识别结果，并按费用场景拆分生成多张报销草稿。",
+                        emphasis="primary",
+                    )
+                )
+            else:
+                actions.append(
+                    UserAgentReviewAction(
+                        label="确认并继续生成草稿",
+                        action_type="confirm_review",
+                        description="确认当前识别字段无误后，继续生成或覆盖当前报销草稿。",
+                        emphasis="primary",
+                    )
+                )
+
+        for slot in payload.ontology.missing_slots[:3]:
+            label = SLOT_LABELS.get(slot, slot)
+            actions.append(
+                UserAgentReviewAction(
+                    label=f"补充{label}",
+                    action_type="fill_slot",
+                    description=f"当前还缺少 {label}，补充后可提升分单和建单准确度。",
+                    emphasis="secondary",
+                )
+            )
+
+        if self._resolve_attachment_count(payload) <= 0:
+            actions.append(
+                UserAgentReviewAction(
+                    label="继续上传票据",
+                    action_type="upload_more",
+                    description="上传发票、行程单或电子票据后，系统会重新识别并完善报销分组。",
+                    emphasis="secondary",
+                )
+            )
+
+        if draft_payload is not None and draft_payload.claim_no:
+            actions.append(
+                UserAgentReviewAction(
+                    label=f"查看草稿 {draft_payload.claim_no}",
+                    action_type="open_claim",
+                    description="查看当前已创建的报销草稿，并继续补充字段或附件。",
+                    emphasis="secondary",
+                )
+            )
+
+        return actions[:5]
+
+    def _build_review_intent_summary(
+        self,
+        payload: UserAgentRequest,
+        *,
+        slot_cards: list[UserAgentReviewSlotCard],
+        claim_groups: list[UserAgentReviewClaimGroup],
+    ) -> str:
+        slots = {item.key: item for item in slot_cards}
+        expense_type = slots.get("expense_type")
+        amount = slots.get("amount")
+        time_range = slots.get("time_range")
+        location = slots.get("location")
+        customer = slots.get("customer_name")
+
+        summary = "系统识别出您想要发起一笔报销。"
+        if expense_type and expense_type.value:
+            summary = f"系统识别出您想要报销{expense_type.value}。"
+        details: list[str] = []
+        if customer and customer.value:
+            details.append(f"客户名称：{customer.value}")
+        if time_range and time_range.value:
+            details.append(f"时间：{time_range.value}")
+        if location and location.value:
+            details.append(f"地点：{location.value}")
+        if amount and amount.value:
+            details.append(f"金额：{amount.value}")
+        if claim_groups and len(claim_groups) > 1:
+            details.append(f"建议拆分为 {len(claim_groups)} 张报销单")
+        if details:
+            return f"{summary} {'；'.join(details)}。"
+        return summary
+
+    @staticmethod
+    def _should_skip_model_answer(
+        payload: UserAgentRequest,
+        review_payload: UserAgentReviewPayload | None,
+    ) -> bool:
+        if review_payload is None:
+            return False
+        return payload.ontology.scenario == "expense" and (
+            payload.ontology.intent == "draft"
+            or int(payload.context_json.get("attachment_count") or 0) > 0
+        )
+
     def _build_rule_citations(self, payload: UserAgentRequest) -> list[UserAgentCitation]:
         domain = self._resolve_domain(payload.ontology.scenario)
         items = self.asset_service.list_assets(
@@ -516,6 +993,45 @@ class UserAgentService:
             return []
         return [str(name) for name in names if str(name).strip()]
 
+    @staticmethod
+    def _resolve_attachment_count(payload: UserAgentRequest) -> int:
+        names = UserAgentService._resolve_attachment_names(payload)
+        if names:
+            return len(names)
+        try:
+            return max(0, int(payload.context_json.get("attachment_count") or 0))
+        except (TypeError, ValueError):
+            return 0
+
+    @staticmethod
+    def _resolve_ocr_documents(payload: UserAgentRequest) -> list[dict[str, object]]:
+        documents = payload.context_json.get("ocr_documents")
+        if not isinstance(documents, list):
+            return []
+        normalized: list[dict[str, object]] = []
+        for item in documents[:8]:
+            if not isinstance(item, dict):
+                continue
+            normalized.append(item)
+        return normalized
+
+    @staticmethod
+    def _resolve_conversation_history(payload: UserAgentRequest) -> list[dict[str, object]]:
+        history = payload.context_json.get("conversation_history")
+        if not isinstance(history, list):
+            return []
+
+        normalized: list[dict[str, object]] = []
+        for item in history[-8:]:
+            if not isinstance(item, dict):
+                continue
+            role = str(item.get("role") or "").strip()
+            content = str(item.get("content") or "").strip()
+            if not role or not content:
+                continue
+            normalized.append({"role": role, "content": content})
+        return normalized
+
     @staticmethod
     def _resolve_domain(scenario: str) -> str | None:
         if scenario == "expense":
@@ -557,3 +1073,210 @@ class UserAgentService:
             if len(cleaned) >= 2:
                 break
         return "；".join(cleaned[:2])
+
+    def _collect_entity_values(self, payload: UserAgentRequest) -> dict[str, str]:
+        values = {
+            "employee_name": "",
+            "customer": "",
+            "participants": "",
+            "amount": "",
+            "expense_type": "",
+            "expense_type_code": "",
+        }
+        participants: list[str] = []
+        for item in payload.ontology.entities:
+            if item.type == "employee" and not values["employee_name"]:
+                values["employee_name"] = item.value
+            elif item.type == "customer" and not values["customer"]:
+                values["customer"] = item.value
+            elif item.type == "amount" and item.role != "threshold" and not values["amount"]:
+                values["amount"] = f"{item.value}元" if "元" not in item.value else item.value
+            elif item.type == "expense_type" and not values["expense_type_code"]:
+                values["expense_type_code"] = item.normalized_value
+                values["expense_type"] = EXPENSE_TYPE_LABELS.get(
+                    item.normalized_value,
+                    item.value,
+                )
+            elif item.type in {"participant", "person"} and item.value.strip():
+                participants.append(item.value.strip())
+        if participants:
+            values["participants"] = "、".join(dict.fromkeys(participants))
+        return values
+
+    def _format_time_range(self, payload: UserAgentRequest) -> str:
+        time_range = payload.ontology.time_range
+        if time_range.raw:
+            return time_range.raw
+        if time_range.start_date and time_range.end_date:
+            if time_range.start_date == time_range.end_date:
+                return time_range.start_date
+            return f"{time_range.start_date} 至 {time_range.end_date}"
+        return ""
+
+    def _resolve_location_value(self, payload: UserAgentRequest) -> str:
+        request_context = payload.context_json.get("request_context")
+        if isinstance(request_context, dict):
+            for key in ("city", "location"):
+                value = str(request_context.get(key) or "").strip()
+                if value:
+                    return value
+        city_match = re.search(r"去(?P<city>[\u4e00-\u9fa5]{2,8})(?:出差|拜访|参会|见客户|客户现场)", payload.message)
+        if city_match:
+            return city_match.group("city").strip()
+        if "客户现场" in payload.message.replace(" ", ""):
+            return "客户现场"
+        return ""
+
+    def _make_slot_card(
+        self,
+        *,
+        key: str,
+        value: str,
+        source: str,
+        confidence: float,
+        missing_slots: set[str],
+        required: bool = True,
+    ) -> UserAgentReviewSlotCard:
+        is_missing = key in missing_slots or not str(value).strip()
+        return UserAgentReviewSlotCard(
+            key=key,
+            label=SLOT_LABELS.get(key, key),
+            value=str(value or "").strip(),
+            source=source,
+            confidence=confidence,
+            required=required,
+            confirmed=not is_missing and source in {"user_text", "page_context", "upload"},
+            status="missing" if is_missing else "identified" if source == "user_text" else "inferred",
+            hint=f"建议补充 {SLOT_LABELS.get(key, key)}。"
+            if is_missing and required
+            else "",
+        )
+
+    def _classify_document(
+        self,
+        item: dict[str, object],
+        payload: UserAgentRequest,
+    ) -> dict[str, str]:
+        text = " ".join(
+            [
+                str(item.get("filename") or ""),
+                str(item.get("summary") or ""),
+                str(item.get("text") or ""),
+            ]
+        ).lower()
+        compact = text.replace(" ", "")
+        expense_type_code = self._collect_entity_values(payload).get("expense_type_code", "")
+        has_customer = bool(self._collect_entity_values(payload).get("customer"))
+
+        if any(keyword in compact for keyword in ("机票", "航班", "火车", "高铁", "行程单")):
+            return {
+                "document_type": "travel_ticket",
+                "expense_type": "travel",
+                "group_code": "travel",
+                "scene_label": "差旅票据",
+            }
+        if any(keyword in compact for keyword in ("酒店", "住宿", "宾馆")):
+            return {
+                "document_type": "hotel_invoice",
+                "expense_type": "hotel",
+                "group_code": "travel",
+                "scene_label": "住宿票据",
+            }
+        if any(keyword in compact for keyword in ("打车", "出租车", "滴滴", "网约车", "过路费", "停车")):
+            return {
+                "document_type": "transport_receipt",
+                "expense_type": "transport",
+                "group_code": "travel",
+                "scene_label": "交通票据",
+            }
+        if any(keyword in compact for keyword in ("餐", "饭店", "酒楼", "酒家", "餐饮", "meal")):
+            group_code = "entertainment" if expense_type_code == "entertainment" or has_customer else "meal"
+            return {
+                "document_type": "meal_receipt",
+                "expense_type": group_code,
+                "group_code": group_code,
+                "scene_label": "餐饮票据",
+            }
+        return {
+            "document_type": "other",
+            "expense_type": expense_type_code or "other",
+            "group_code": self._normalize_group_code(expense_type_code or "other"),
+            "scene_label": "其他票据",
+        }
+
+    @staticmethod
+    def _normalize_group_code(expense_type_code: str) -> str:
+        if expense_type_code in {"travel", "hotel", "transport"}:
+            return "travel"
+        if expense_type_code in {"entertainment", "meal"}:
+            return expense_type_code
+        return "other"
+
+    def _extract_document_fields(self, item: dict[str, object]) -> dict[str, str]:
+        text = " ".join([str(item.get("summary") or ""), str(item.get("text") or "")]).strip()
+        fields: dict[str, str] = {}
+        amount_match = AMOUNT_TEXT_PATTERN.search(text)
+        if amount_match:
+            fields["金额"] = f"{amount_match.group(1)}元"
+        date_match = DATE_TEXT_PATTERN.search(text)
+        if date_match:
+            fields["时间"] = date_match.group(1)
+
+        merchant = self._extract_document_merchant_name(item)
+        if merchant:
+            fields["商户/酒店"] = merchant
+        return fields
+
+    @staticmethod
+    def _extract_document_merchant_name(item: dict[str, object]) -> str:
+        text = " ".join([str(item.get("summary") or ""), str(item.get("text") or "")]).strip()
+        for keyword in ("酒店", "宾馆", "饭店", "酒楼", "餐厅", "航空", "铁路", "滴滴"):
+            if keyword in text:
+                return keyword
+        return ""
+
+    @staticmethod
+    def _extract_amount_from_card(card: UserAgentReviewDocumentCard) -> float:
+        for item in card.fields:
+            if item.label != "金额":
+                continue
+            try:
+                return float(str(item.value).replace("元", "").strip())
+            except ValueError:
+                return 0.0
+        return 0.0
+
+    def _resolve_amount_value(self, payload: UserAgentRequest) -> float:
+        for item in payload.ontology.entities:
+            if item.type == "amount" and item.role != "threshold":
+                try:
+                    return float(item.normalized_value)
+                except ValueError:
+                    return 0.0
+        return 0.0
+
+    def _sum_ocr_amounts(self, ocr_documents: list[dict[str, object]]) -> float:
+        total = 0.0
+        for item in ocr_documents:
+            fields = self._extract_document_fields(item)
+            amount_text = str(fields.get("金额") or "").replace("元", "").strip()
+            if not amount_text:
+                continue
+            try:
+                total += float(amount_text)
+            except ValueError:
+                continue
+        return total
+
+    def _infer_expense_type_from_documents(
+        self,
+        payload: UserAgentRequest,
+        ocr_documents: list[dict[str, object]],
+    ) -> str:
+        labels: list[str] = []
+        for item in ocr_documents:
+            classified = self._classify_document(item, payload)
+            label = GROUP_SCENE_LABELS.get(classified["group_code"], "")
+            if label and label not in labels:
+                labels.append(label)
+        return " + ".join(labels[:3])