feat: 报销预审会话状态管理与工作台交互增强

- 新增差旅报销会话状态管理与对话模型重构
- 增强风险观测服务与运行时聊天上下文作用域
- 优化工作台图标资源、助理意图识别与摘要工具
- 完善报销创建视图样式与差旅详情页标准调整交互
- 补充风险观测、运行时聊天与报销端点测试覆盖

server/src/app/api/v1/endpoints/steward.py

@@ -0,0 +1,78 @@
+from __future__ import annotations
+
+import asyncio
+import json
+from collections.abc import AsyncIterator
+from typing import Annotated, Any
+
+from fastapi import APIRouter, Depends, HTTPException, status
+from fastapi.responses import StreamingResponse
+from sqlalchemy.orm import Session
+
+from app.api.deps import get_db
+from app.schemas.common import ErrorResponse
+from app.schemas.steward import StewardPlanRequest, StewardPlanResponse
+from app.services.runtime_chat import RuntimeChatService
+from app.services.steward_intent_agent import StewardIntentAgent
+from app.services.steward_planner import StewardPlannerService
+
+router = APIRouter(prefix="/steward")
+DbSession = Annotated[Session, Depends(get_db)]
+
+
+@router.post(
+    "/plans",
+    response_model=StewardPlanResponse,
+    summary="生成小财管家任务计划",
+    description="把首页自然语言和附件元信息拆解为可确认、可追踪、可分派的财务任务计划。",
+    responses={
+        status.HTTP_400_BAD_REQUEST: {
+            "model": ErrorResponse,
+            "description": "请求缺少任务描述，无法生成小财管家计划。",
+        }
+    },
+)
+def create_steward_plan(payload: StewardPlanRequest, db: DbSession) -> StewardPlanResponse:
+    try:
+        return _build_steward_planner(db).build_plan(payload)
+    except ValueError as exc:
+        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail=str(exc)) from exc
+
+
+@router.post(
+    "/plans/stream",
+    summary="流式生成小财管家任务计划",
+    description="以 NDJSON 逐条返回小财管家的过程摘要事件，最后返回完整任务计划。",
+)
+async def stream_steward_plan(payload: StewardPlanRequest, db: DbSession) -> StreamingResponse:
+    return StreamingResponse(
+        _iter_steward_plan_events(payload, _build_steward_planner(db)),
+        media_type="application/x-ndjson",
+    )
+
+
+async def _iter_steward_plan_events(
+    payload: StewardPlanRequest,
+    planner: StewardPlannerService,
+) -> AsyncIterator[str]:
+    try:
+        plan = planner.build_plan(payload)
+    except ValueError as exc:
+        yield _encode_stream_event("error", {"message": str(exc)})
+        return
+
+    for event in plan.thinking_events:
+        yield _encode_stream_event("thinking", event.model_dump(mode="json"))
+        await asyncio.sleep(0.18)
+
+    yield _encode_stream_event("plan", plan.model_dump(mode="json"))
+
+
+def _encode_stream_event(event: str, data: dict[str, Any]) -> str:
+    return json.dumps({"event": event, "data": data}, ensure_ascii=False) + "\n"
+
+
+def _build_steward_planner(db: Session) -> StewardPlannerService:
+    return StewardPlannerService(
+        intent_agent=StewardIntentAgent(RuntimeChatService(db)),
+    )

server/src/app/api/v1/router.py

@@ -22,6 +22,7 @@ from app.api.v1.endpoints.receipt_folder import router as receipt_folder_router
 from app.api.v1.endpoints.reimbursements import router as reimbursements_router
 from app.api.v1.endpoints.risk_observations import router as risk_observations_router
 from app.api.v1.endpoints.settings import router as settings_router
+from app.api.v1.endpoints.steward import router as steward_router
 from app.api.v1.endpoints.system_logs import router as system_logs_router
 
 router = APIRouter()
@@ -47,4 +48,5 @@ router.include_router(employee_profiles_router, tags=["employee-profiles"])
 router.include_router(reimbursements_router, prefix="/reimbursements", tags=["reimbursements"])
 router.include_router(risk_observations_router, tags=["risk-observations"])
 router.include_router(settings_router, tags=["settings"])
+router.include_router(steward_router, tags=["steward"])
 router.include_router(system_logs_router, tags=["system-logs"])

server/src/app/schemas/reimbursement.py

@@ -126,6 +126,7 @@ class ExpenseClaimStandardAdjustmentRisk(BaseModel):
     item_id: str | None = Field(default=None, max_length=120)
     title: str | None = Field(default=None, max_length=120)
     risk: str | None = Field(default=None, max_length=500)
+    application_days: int | None = Field(default=None, ge=1, le=365)
     original_amount: Decimal | None = None
     reimbursable_amount: Decimal | None = None
 

server/src/app/schemas/steward.py

@@ -0,0 +1,90 @@
+from __future__ import annotations
+
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+
+StewardTaskType = Literal["expense_application", "reimbursement"]
+StewardAssignedAgent = Literal["application_assistant", "reimbursement_assistant"]
+StewardPlanningSource = Literal["llm_function_call", "rule_fallback"]
+StewardTaskStatus = Literal[
+    "planned",
+    "needs_confirmation",
+    "ready_to_delegate",
+    "delegated",
+    "completed",
+    "blocked",
+]
+StewardConfirmationStatus = Literal["pending", "confirmed", "rejected"]
+
+
+class StewardAttachmentInput(BaseModel):
+    name: str = Field(description="附件原始文件名。")
+    media_type: str = Field(default="", description="附件 MIME 类型。")
+    ocr_summary: str = Field(default="", description="可选 OCR 摘要。")
+    ocr_fields: dict[str, Any] = Field(default_factory=dict, description="可选 OCR 结构化字段。")
+
+
+class StewardPlanRequest(BaseModel):
+    message: str = Field(description="用户在首页输入的自然语言任务。")
+    user_id: str | None = Field(default=None, description="当前用户 ID。")
+    client_now_iso: str | None = Field(default=None, description="客户端当前时间 ISO 字符串。")
+    attachments: list[StewardAttachmentInput] = Field(default_factory=list, description="随本次输入上传的附件。")
+    context_json: dict[str, Any] = Field(default_factory=dict, description="调用方上下文。")
+
+
+class StewardThinkingEvent(BaseModel):
+    event_id: str = Field(description="过程摘要事件 ID。")
+    stage: str = Field(description="阶段编码。")
+    title: str = Field(description="面向用户展示的阶段标题。")
+    content: str = Field(description="面向用户展示的过程摘要。")
+    status: str = Field(default="completed", description="事件状态。")
+
+
+class StewardTask(BaseModel):
+    task_id: str = Field(description="小财管家任务 ID。")
+    task_type: StewardTaskType = Field(description="任务类型。")
+    assigned_agent: StewardAssignedAgent = Field(description="建议分派的下游助手。")
+    title: str = Field(description="任务标题。")
+    summary: str = Field(description="任务摘要。")
+    status: StewardTaskStatus = Field(default="needs_confirmation", description="任务状态。")
+    confidence: float = Field(default=0.0, ge=0.0, le=1.0, description="识别置信度。")
+    ontology_fields: dict[str, str] = Field(default_factory=dict, description="归一化后的业务本体字段。")
+    missing_fields: list[str] = Field(default_factory=list, description="仍缺失的本体字段。")
+    confirmation_required: bool = Field(default=True, description="执行前是否需要用户确认。")
+
+
+class StewardAttachmentGroup(BaseModel):
+    group_id: str = Field(description="附件归集组 ID。")
+    target_task_id: str | None = Field(default=None, description="建议归属的任务 ID。")
+    scene: str = Field(description="归集场景编码。")
+    scene_label: str = Field(description="归集场景展示名。")
+    attachment_names: list[str] = Field(default_factory=list, description="建议纳入的附件名称。")
+    excluded_attachment_names: list[str] = Field(default_factory=list, description="建议排除或单独处理的附件名称。")
+    confidence: float = Field(default=0.0, ge=0.0, le=1.0, description="归集置信度。")
+    rationale: str = Field(default="", description="归集依据。")
+    confirmation_required: bool = Field(default=True, description="归集前是否需要用户确认。")
+
+
+class StewardConfirmationAction(BaseModel):
+    confirmation_id: str = Field(description="确认动作 ID。")
+    action_type: str = Field(description="确认动作类型。")
+    label: str = Field(description="确认按钮文案。")
+    description: str = Field(default="", description="确认动作说明。")
+    target_task_id: str | None = Field(default=None, description="关联任务 ID。")
+    attachment_group_id: str | None = Field(default=None, description="关联附件归集组 ID。")
+    status: StewardConfirmationStatus = Field(default="pending", description="确认状态。")
+    payload: dict[str, Any] = Field(default_factory=dict, description="确认后继续执行所需载荷。")
+
+
+class StewardPlanResponse(BaseModel):
+    plan_id: str = Field(description="小财管家计划 ID。")
+    plan_status: str = Field(default="needs_confirmation", description="计划状态。")
+    planning_source: StewardPlanningSource = Field(default="rule_fallback", description="计划生成来源。")
+    summary: str = Field(description="计划摘要。")
+    thinking_events: list[StewardThinkingEvent] = Field(default_factory=list, description="过程摘要事件。")
+    tasks: list[StewardTask] = Field(default_factory=list, description="拆解后的任务。")
+    attachment_groups: list[StewardAttachmentGroup] = Field(default_factory=list, description="附件归集建议。")
+    confirmation_groups: list[StewardConfirmationAction] = Field(default_factory=list, description="等待用户确认的动作。")
+    model_call_traces: list[dict[str, Any]] = Field(default_factory=list, description="模型工具调用轨迹。")

server/src/app/services/expense_claim_attachment_operations.py

@@ -254,6 +254,7 @@ class ExpenseClaimAttachmentOperationsMixin:
         )
 
         self._sync_claim_from_items(claim)
+        self._refresh_claim_pre_review_flags(claim, is_application_claim=False)
         self.db.commit()
         self.db.refresh(claim)
 
@@ -356,6 +357,7 @@ class ExpenseClaimAttachmentOperationsMixin:
         item.invoice_id = None
 
         self._sync_claim_from_items(claim)
+        self._refresh_claim_pre_review_flags(claim, is_application_claim=False)
         self.db.commit()
         self.db.refresh(claim)
 

server/src/app/services/expense_claim_pre_review.py

@@ -139,3 +139,61 @@ class ExpenseClaimPreReviewMixin:
             )
         ]
         return [*preserved_flags, next_flag]
+
+    def _refresh_claim_pre_review_flags(
+        self,
+        claim: ExpenseClaim,
+        *,
+        is_application_claim: bool | None = None,
+        reviewed_at: datetime | None = None,
+    ) -> bool:
+        if claim is None:
+            return False
+
+        if is_application_claim is None:
+            is_application_claim = self._is_expense_application_claim(claim)
+        reviewed_at = reviewed_at or datetime.now(UTC)
+
+        if is_application_claim:
+            preserved_flags = [
+                flag
+                for flag in list(claim.risk_flags_json or [])
+                if not (
+                    isinstance(flag, dict)
+                    and str(flag.get("source") or "").strip() == "submission_review"
+                    and str(flag.get("hit_source") or "").strip() == "rule_center"
+                )
+            ]
+            application_review = self.evaluate_platform_risk_rules(
+                claim,
+                business_stage="expense_application",
+            )
+            review_flags = [*preserved_flags, *list(application_review.get("flags") or [])]
+        else:
+            review_result = self._run_ai_submission_review(claim)
+            review_flags = list(review_result.get("risk_flags") or [])
+
+        blocking_count = self._count_ai_pre_review_blocking_risks(review_flags)
+        claim.risk_flags_json = self._replace_ai_pre_review_flag(
+            review_flags,
+            self._build_ai_pre_review_flag(
+                passed=blocking_count <= 0,
+                blocking_count=blocking_count,
+                reviewed_at=reviewed_at,
+                business_stage=risk_business_stage_for_claim(
+                    is_application_claim=is_application_claim,
+                ),
+            ),
+        )
+        if not is_application_claim:
+            claim.approval_stage = "\u5f85\u63d0\u4ea4"
+        claim.submitted_at = None
+        return True
+
+    @staticmethod
+    def _has_ai_pre_review_flag(claim: ExpenseClaim) -> bool:
+        return any(
+            isinstance(flag, dict)
+            and str(flag.get("source") or "").strip() == "ai_pre_review"
+            for flag in list(claim.risk_flags_json or [])
+        )

server/src/app/services/expense_claims.py

@@ -48,6 +48,7 @@ from app.services.expense_claim_attachment_analysis import ExpenseClaimAttachmen
 from app.services.expense_claim_attachment_document import ExpenseClaimAttachmentDocumentMixin
 from app.services.expense_claim_attachment_operations import ExpenseClaimAttachmentOperationsMixin
 from app.services.expense_claim_budget_flow import ExpenseClaimBudgetFlowMixin
+from app.services.expense_claim_workflow_constants import DIRECT_MANAGER_APPROVAL_STAGE
 from app.services.expense_claim_document_item_builder import ExpenseClaimDocumentItemBuilderMixin
 from app.services.expense_claim_document_parsing import ExpenseClaimDocumentParsingMixin
 from app.services.expense_claim_draft_flow import ExpenseClaimDraftFlowMixin
@@ -278,6 +279,9 @@ class ExpenseClaimService(
         if payload.reason is not None:
             claim.reason = self._normalize_optional_text(payload.reason, allow_empty=True) or "待补充"
 
+        if not self._is_expense_application_claim(claim):
+            self._refresh_claim_pre_review_flags(claim, is_application_claim=False)
+
         self.db.commit()
         self.db.refresh(claim)
 
@@ -306,6 +310,146 @@ class ExpenseClaimService(
         normalized = Decimal(value or Decimal("0.00")).quantize(Decimal("0.01"))
         return f"{normalized:.2f}"
 
+    @staticmethod
+    def _normalize_standard_adjustment_days(value: Any) -> int | None:
+        if value is None:
+            return None
+        if isinstance(value, int):
+            return value if 1 <= value <= 365 else None
+        text = str(value or "").strip()
+        if not text:
+            return None
+        match = re.search(r"\d{1,3}", text)
+        if not match:
+            return None
+        days = int(match.group(0))
+        return days if 1 <= days <= 365 else None
+
+    @staticmethod
+    def _normalize_standard_adjustment_text(value: Any) -> str:
+        text = str(value or "").strip()
+        if not text or text in {"-", "N/A", "n/a"}:
+            return ""
+        if text in {"待补充", "未知", "暂无", "非必填"}:
+            return ""
+        return text
+
+    def _iter_standard_adjustment_application_details(self, claim: ExpenseClaim) -> list[dict[str, Any]]:
+        details: list[dict[str, Any]] = []
+        for flag in list(claim.risk_flags_json or []):
+            if not isinstance(flag, dict):
+                continue
+            detail = flag.get("application_detail") or flag.get("applicationDetail")
+            if isinstance(detail, dict):
+                details.append(detail)
+            related = flag.get("related_application") or flag.get("relatedApplication")
+            if isinstance(related, dict):
+                details.append(related)
+        return details
+
+    def _resolve_standard_adjustment_days(
+        self,
+        claim: ExpenseClaim,
+        item: ExpenseClaimItem,
+        entry: Any,
+    ) -> int:
+        direct_days = self._normalize_standard_adjustment_days(getattr(entry, "application_days", None))
+        if direct_days is not None:
+            return direct_days
+
+        for detail in self._iter_standard_adjustment_application_details(claim):
+            for key in ("application_days", "applicationDays", "days"):
+                detail_days = self._normalize_standard_adjustment_days(detail.get(key))
+                if detail_days is not None:
+                    return detail_days
+
+        candidates = [
+            getattr(entry, "risk", None),
+            getattr(entry, "title", None),
+            item.item_reason,
+            claim.reason,
+        ]
+        for text in candidates:
+            match = re.search(r"(\d{1,3})\s*(?:天|晚|夜)", str(text or ""))
+            if match:
+                days = self._normalize_standard_adjustment_days(match.group(1))
+                if days is not None:
+                    return days
+        return 1
+
+    def _resolve_standard_adjustment_location(
+        self,
+        claim: ExpenseClaim,
+        item: ExpenseClaimItem,
+    ) -> str:
+        for value in (item.item_location, claim.location):
+            text = self._normalize_standard_adjustment_text(value)
+            if text:
+                return text
+
+        for detail in self._iter_standard_adjustment_application_details(claim):
+            for key in ("application_location", "applicationLocation", "location", "city"):
+                text = self._normalize_standard_adjustment_text(detail.get(key))
+                if text:
+                    return text
+        return ""
+
+    def _resolve_policy_standard_reimbursable_amount(
+        self,
+        *,
+        claim: ExpenseClaim,
+        item: ExpenseClaimItem,
+        entry: Any,
+        current_user: CurrentUserContext,
+    ) -> Decimal | None:
+        item_type = str(item.item_type or "").strip().lower()
+        if item_type not in {"hotel", "hotel_ticket"}:
+            return None
+
+        location = self._resolve_standard_adjustment_location(claim, item)
+        grade = str(claim.employee_grade or current_user.grade or "").strip()
+        if not location or not grade:
+            return None
+
+        try:
+            from app.services.travel_reimbursement_calculator import TravelReimbursementCalculatorService
+
+            result = TravelReimbursementCalculatorService(self.db).calculate(
+                TravelReimbursementCalculatorRequest(
+                    days=self._resolve_standard_adjustment_days(claim, item, entry),
+                    location=location,
+                    grade=grade,
+                ),
+                current_user,
+            )
+        except Exception:
+            return None
+
+        return self._normalize_standard_adjustment_amount(result.hotel_amount)
+
+    def _resolve_standard_adjustment_reimbursable_amount(
+        self,
+        *,
+        claim: ExpenseClaim,
+        item: ExpenseClaimItem,
+        entry: Any,
+        original_amount: Decimal,
+        current_user: CurrentUserContext,
+    ) -> Decimal:
+        policy_amount = self._resolve_policy_standard_reimbursable_amount(
+            claim=claim,
+            item=item,
+            entry=entry,
+            current_user=current_user,
+        )
+        if policy_amount is not None:
+            return min(max(policy_amount, Decimal("0.00")), original_amount)
+
+        entry_amount = self._normalize_standard_adjustment_amount(entry.reimbursable_amount)
+        if entry_amount is not None:
+            return min(max(entry_amount, Decimal("0.00")), original_amount)
+        return original_amount
+
     def accept_standard_adjustment(
         self,
         *,
@@ -340,11 +484,13 @@ class ExpenseClaimService(
                 self._normalize_standard_adjustment_amount(entry.original_amount)
                 or Decimal(item.item_amount or Decimal("0.00")).quantize(Decimal("0.01"))
             )
-            reimbursable_amount = (
-                self._normalize_standard_adjustment_amount(entry.reimbursable_amount)
-                or original_amount
+            reimbursable_amount = self._resolve_standard_adjustment_reimbursable_amount(
+                claim=claim,
+                item=item,
+                entry=entry,
+                original_amount=original_amount,
+                current_user=current_user,
             )
-            reimbursable_amount = min(max(reimbursable_amount, Decimal("0.00")), original_amount)
             employee_absorbed_amount = (original_amount - reimbursable_amount).quantize(Decimal("0.01"))
             item_label = (
                 str(item.item_reason or "").strip()
@@ -456,6 +602,7 @@ class ExpenseClaimService(
 
         self._refresh_item_attachment_analysis(item)
         self._sync_claim_from_items(claim)
+        self._refresh_claim_pre_review_flags(claim, is_application_claim=False)
         self.db.commit()
         self.db.refresh(claim)
 
@@ -510,6 +657,7 @@ class ExpenseClaimService(
         self.db.add(item)
 
         self._sync_claim_from_items(claim)
+        self._refresh_claim_pre_review_flags(claim, is_application_claim=False)
         self.db.commit()
         self.db.refresh(claim)
 
@@ -548,6 +696,7 @@ class ExpenseClaimService(
         self.db.delete(item)
 
         self._sync_claim_from_items(claim)
+        self._refresh_claim_pre_review_flags(claim, is_application_claim=False)
         self.db.commit()
         self.db.refresh(claim)
 
@@ -645,12 +794,13 @@ class ExpenseClaimService(
                 budget_flags,
                 business_stage="reimbursement",
             )
-            review_result = self._run_ai_submission_review(claim)
+            if not self._has_ai_pre_review_flag(claim):
+                self._refresh_claim_pre_review_flags(claim, is_application_claim=False)
+
+            claim.status = "submitted"
+            claim.approval_stage = DIRECT_MANAGER_APPROVAL_STAGE
+            claim.submitted_at = datetime.now(UTC)
 
-            claim.status = str(review_result.get("status") or "supplement")
-            claim.approval_stage = str(review_result.get("approval_stage") or "待补充")
-            claim.risk_flags_json = list(review_result.get("risk_flags") or [])
-            claim.submitted_at = datetime.now(UTC) if claim.status == "submitted" else None
 
         self.db.commit()
         self.db.refresh(claim)
@@ -872,11 +1022,3 @@ class ExpenseClaimService(
 
 
 
-
-
-
-
-
-
-
-

server/src/app/services/risk_observations.py

@@ -33,17 +33,25 @@ FEEDBACK_STATUS_MAP = {
 
 
 class RiskObservationService:
+    _storage_ready_cache: set[str] = set()
+
     def __init__(self, db: Session) -> None:
         self.db = db
 
     def ensure_storage_ready(self) -> None:
+        bind = self.db.get_bind()
+        cache_key = str(getattr(bind, "url", "") or id(bind))
+        if cache_key in self._storage_ready_cache:
+            return
+
         Base.metadata.create_all(
-            bind=self.db.get_bind(),
+            bind=bind,
             tables=[
                 RiskObservation.__table__,
                 RiskObservationFeedback.__table__,
             ],
         )
+        self._storage_ready_cache.add(cache_key)
 
     def upsert_observation(
         self,

server/src/app/services/runtime_chat.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import json
 from dataclasses import dataclass
 from http import HTTPStatus
 from time import monotonic, sleep
@@ -61,6 +62,23 @@ class RuntimeChatResult:
         return [item.model_dump() for item in self.calls]
 
 
+@dataclass(slots=True)
+class RuntimeChatToolCall:
+    name: str
+    arguments: dict[str, Any]
+    call_id: str | None = None
+    raw_arguments: str = ""
+
+
+@dataclass(slots=True)
+class RuntimeToolCallResult:
+    tool_call: RuntimeChatToolCall | None
+    calls: list[RuntimeChatCallTrace]
+
+    def calls_as_dicts(self) -> list[dict[str, Any]]:
+        return [item.model_dump() for item in self.calls]
+
+
 class RuntimeChatService:
     def __init__(self, db: Session) -> None:
         self.db = db
@@ -208,6 +226,131 @@ class RuntimeChatService:
 
         return RuntimeChatResult(None, calls)
 
+    def complete_with_tool_call(
+        self,
+        messages: list[dict[str, Any]],
+        *,
+        tools: list[dict[str, Any]],
+        tool_choice: dict[str, Any] | str | None = None,
+        slot_priority: tuple[str, ...] = ("main", "backup"),
+        max_tokens: int = 1200,
+        temperature: float = 0.1,
+        timeout_seconds: int | None = None,
+        slot_timeouts: dict[str, int] | None = None,
+        max_attempts: int | None = None,
+    ) -> RuntimeToolCallResult:
+        configs: list[dict[str, str]] = []
+        calls: list[RuntimeChatCallTrace] = []
+        for slot in slot_priority:
+            config = self._load_chat_slot(slot)
+            if config is None:
+                calls.append(
+                    RuntimeChatCallTrace(
+                        slot=slot,
+                        provider="",
+                        model="",
+                        attempt=0,
+                        status="skipped",
+                        skipped_reason="not_configured",
+                    )
+                )
+                continue
+            configs.append(config)
+        if not configs:
+            return RuntimeToolCallResult(None, calls)
+
+        resolved_timeout_seconds = timeout_seconds or DEFAULT_RUNTIME_CHAT_TIMEOUT_SECONDS
+        resolved_slot_timeouts = dict(slot_timeouts or {})
+        resolved_max_attempts = max_attempts or DEFAULT_RUNTIME_CHAT_RETRY_ATTEMPTS
+
+        for attempt in range(1, resolved_max_attempts + 1):
+            for config in configs:
+                cache_key = self._build_slot_cache_key(config)
+                if _slot_failure_until.get(cache_key, 0.0) > monotonic():
+                    logger.info(
+                        "Skip runtime chat tool slot=%s provider=%s because it is in cooldown",
+                        config["slot"],
+                        config["provider"],
+                    )
+                    calls.append(
+                        RuntimeChatCallTrace(
+                            slot=config["slot"],
+                            provider=config["provider"],
+                            model=config["model"],
+                            attempt=attempt,
+                            status="skipped",
+                            skipped_reason="cooldown",
+                        )
+                    )
+                    continue
+                started = monotonic()
+                try:
+                    tool_call = self._request_chat_tool_call(
+                        config,
+                        messages,
+                        tools=tools,
+                        tool_choice=tool_choice,
+                        max_tokens=max_tokens,
+                        temperature=temperature,
+                        timeout_seconds=resolved_slot_timeouts.get(
+                            config["slot"],
+                            resolved_timeout_seconds,
+                        ),
+                    )
+                    duration_ms = int((monotonic() - started) * 1000)
+                    if tool_call is not None:
+                        _slot_failure_until.pop(cache_key, None)
+                        calls.append(
+                            RuntimeChatCallTrace(
+                                slot=config["slot"],
+                                provider=config["provider"],
+                                model=config["model"],
+                                attempt=attempt,
+                                status="succeeded",
+                                duration_ms=duration_ms,
+                            )
+                        )
+                        return RuntimeToolCallResult(tool_call, calls)
+                    calls.append(
+                        RuntimeChatCallTrace(
+                            slot=config["slot"],
+                            provider=config["provider"],
+                            model=config["model"],
+                            attempt=attempt,
+                            status="empty",
+                            duration_ms=duration_ms,
+                            error_message="模型未返回工具调用。",
+                        )
+                    )
+                except Exception as exc:
+                    duration_ms = int((monotonic() - started) * 1000)
+                    _slot_failure_until[cache_key] = (
+                        monotonic() + DEFAULT_RUNTIME_CHAT_FAILURE_COOLDOWN_SECONDS
+                    )
+                    calls.append(
+                        RuntimeChatCallTrace(
+                            slot=config["slot"],
+                            provider=config["provider"],
+                            model=config["model"],
+                            attempt=attempt,
+                            status="failed",
+                            duration_ms=duration_ms,
+                            error_message=str(exc),
+                        )
+                    )
+                    logger.warning(
+                        "Runtime chat tool request failed slot=%s provider=%s attempt=%s/%s: %s",
+                        config["slot"],
+                        config["provider"],
+                        attempt,
+                        resolved_max_attempts,
+                        exc,
+                    )
+            if attempt < resolved_max_attempts:
+                sleep(DEFAULT_RUNTIME_CHAT_RETRY_DELAY_SECONDS)
+
+        return RuntimeToolCallResult(None, calls)
+
     @staticmethod
     def _build_slot_cache_key(config: dict[str, str]) -> str:
         return "|".join(
@@ -295,6 +438,51 @@ class RuntimeChatService:
             timeout_seconds=timeout_seconds,
         )
 
+    def _request_chat_tool_call(
+        self,
+        config: dict[str, str],
+        messages: list[dict[str, Any]],
+        *,
+        tools: list[dict[str, Any]],
+        tool_choice: dict[str, Any] | str | None,
+        max_tokens: int,
+        temperature: float,
+        timeout_seconds: int,
+    ) -> RuntimeChatToolCall | None:
+        provider = config["provider"]
+        endpoint = config["endpoint"]
+        model = config["model"]
+        api_key = config["apiKey"]
+
+        if provider == "Azure OpenAI":
+            return self._request_azure_openai_tool_call(
+                endpoint=endpoint,
+                model=model,
+                api_key=api_key,
+                messages=messages,
+                tools=tools,
+                tool_choice=tool_choice,
+                max_tokens=max_tokens,
+                temperature=temperature,
+                timeout_seconds=timeout_seconds,
+            )
+
+        if provider == "Ollama":
+            raise ConnectivityCheckError("Ollama 暂不支持小财管家 function calling。")
+
+        return self._request_openai_compatible_tool_call(
+            provider=provider,
+            endpoint=endpoint,
+            model=model,
+            api_key=api_key,
+            messages=messages,
+            tools=tools,
+            tool_choice=tool_choice,
+            max_tokens=max_tokens,
+            temperature=temperature,
+            timeout_seconds=timeout_seconds,
+        )
+
     def _request_openai_compatible(
         self,
         *,
@@ -331,6 +519,46 @@ class RuntimeChatService:
             )
         return self._extract_openai_text(payload)
 
+    def _request_openai_compatible_tool_call(
+        self,
+        *,
+        provider: str,
+        endpoint: str,
+        model: str,
+        api_key: str,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        tool_choice: dict[str, Any] | str | None,
+        max_tokens: int,
+        temperature: float,
+        timeout_seconds: int,
+    ) -> RuntimeChatToolCall | None:
+        url = _ensure_path(_normalize_endpoint(endpoint), "chat/completions")
+        request_payload: dict[str, Any] = {
+            "model": model,
+            "messages": messages,
+            "tools": tools,
+            "tool_choice": tool_choice or "auto",
+            "max_tokens": max_tokens,
+            "temperature": temperature,
+        }
+        if provider == "GLM":
+            request_payload["thinking"] = {"type": "disabled"}
+
+        status_code, payload = _send_json_request(
+            "POST",
+            url,
+            headers=_build_headers(api_key=api_key, use_bearer=True),
+            payload=request_payload,
+            timeout_seconds=timeout_seconds,
+        )
+        if status_code >= HTTPStatus.BAD_REQUEST:
+            raise ConnectivityCheckError(
+                f"模型接口返回异常状态 {status_code}。",
+                status_code=status_code,
+            )
+        return self._extract_openai_tool_call(payload)
+
     def _request_ollama(
         self,
         *,
@@ -396,6 +624,41 @@ class RuntimeChatService:
             )
         return self._extract_openai_text(payload)
 
+    def _request_azure_openai_tool_call(
+        self,
+        *,
+        endpoint: str,
+        model: str,
+        api_key: str,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        tool_choice: dict[str, Any] | str | None,
+        max_tokens: int,
+        temperature: float,
+        timeout_seconds: int,
+    ) -> RuntimeChatToolCall | None:
+        deployment_base = _build_azure_deployment_base(endpoint, model)
+        url = f"{deployment_base}/chat/completions?api-version={AZURE_API_VERSION}"
+        status_code, payload = _send_json_request(
+            "POST",
+            url,
+            headers=_build_headers(api_key=api_key, use_bearer=False, use_api_key=True),
+            payload={
+                "messages": messages,
+                "tools": tools,
+                "tool_choice": tool_choice or "auto",
+                "max_tokens": max_tokens,
+                "temperature": temperature,
+            },
+            timeout_seconds=timeout_seconds,
+        )
+        if status_code >= HTTPStatus.BAD_REQUEST:
+            raise ConnectivityCheckError(
+                f"Azure OpenAI 返回异常状态 {status_code}。",
+                status_code=status_code,
+            )
+        return self._extract_openai_tool_call(payload)
+
     @staticmethod
     def _extract_openai_text(payload: Any) -> str:
         if not isinstance(payload, dict):
@@ -426,3 +689,74 @@ class RuntimeChatService:
             return text.strip()
 
         return ""
+
+    @staticmethod
+    def _extract_openai_tool_call(payload: Any) -> RuntimeChatToolCall | None:
+        if not isinstance(payload, dict):
+            return None
+
+        choices = payload.get("choices")
+        if not isinstance(choices, list) or not choices:
+            return None
+
+        first_choice = choices[0]
+        if not isinstance(first_choice, dict):
+            return None
+
+        message = first_choice.get("message")
+        if not isinstance(message, dict):
+            return None
+
+        tool_calls = message.get("tool_calls")
+        if isinstance(tool_calls, list) and tool_calls:
+            first_tool = tool_calls[0]
+            if isinstance(first_tool, dict):
+                function_payload = first_tool.get("function")
+                if isinstance(function_payload, dict):
+                    return RuntimeChatService._build_runtime_tool_call(
+                        name=function_payload.get("name"),
+                        arguments=function_payload.get("arguments"),
+                        call_id=first_tool.get("id"),
+                    )
+
+        function_call = message.get("function_call")
+        if isinstance(function_call, dict):
+            return RuntimeChatService._build_runtime_tool_call(
+                name=function_call.get("name"),
+                arguments=function_call.get("arguments"),
+                call_id=None,
+            )
+
+        return None
+
+    @staticmethod
+    def _build_runtime_tool_call(
+        *,
+        name: Any,
+        arguments: Any,
+        call_id: Any,
+    ) -> RuntimeChatToolCall | None:
+        tool_name = str(name or "").strip()
+        if not tool_name:
+            return None
+
+        raw_arguments = ""
+        if isinstance(arguments, dict):
+            parsed_arguments = arguments
+            raw_arguments = json.dumps(arguments, ensure_ascii=False)
+        else:
+            raw_arguments = str(arguments or "").strip()
+            if not raw_arguments:
+                parsed_arguments = {}
+            else:
+                parsed = json.loads(raw_arguments)
+                if not isinstance(parsed, dict):
+                    raise ValueError("工具调用参数必须是 JSON object。")
+                parsed_arguments = parsed
+
+        return RuntimeChatToolCall(
+            name=tool_name,
+            arguments=parsed_arguments,
+            call_id=str(call_id).strip() if call_id else None,
+            raw_arguments=raw_arguments,
+        )

server/src/app/services/steward_constants.py

@@ -0,0 +1,18 @@
+from __future__ import annotations
+
+
+BUSINESS_CANONICAL_FIELD_ORDER = (
+    "expense_type",
+    "time_range",
+    "location",
+    "reason",
+    "amount",
+    "transport_mode",
+    "attachments",
+    "customer_name",
+    "merchant_name",
+    "department_name",
+    "employee_name",
+    "employee_no",
+)
+BUSINESS_CANONICAL_FIELDS = frozenset(BUSINESS_CANONICAL_FIELD_ORDER)

server/src/app/services/steward_intent_agent.py

@@ -0,0 +1,220 @@
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from datetime import date
+from typing import Any
+
+from app.schemas.steward import StewardPlanRequest
+from app.services.ontology_field_registry import normalize_ontology_form_values
+from app.services.runtime_chat import RuntimeChatService
+
+
+STEWARD_INTENT_FUNCTION_NAME = "submit_steward_intent_plan"
+
+
+@dataclass(frozen=True, slots=True)
+class StewardIntentAgentResult:
+    payload: dict[str, Any]
+    model_call_traces: list[dict[str, Any]]
+
+
+class StewardIntentAgent:
+    """使用大模型 function calling 识别小财管家的复合财务意图。"""
+
+    def __init__(self, runtime_chat_service: RuntimeChatService) -> None:
+        self.runtime_chat_service = runtime_chat_service
+        self.last_call_traces: list[dict[str, Any]] = []
+
+    def detect(
+        self,
+        request: StewardPlanRequest,
+        *,
+        base_date: date,
+        canonical_fields: list[str],
+    ) -> StewardIntentAgentResult | None:
+        result = self.runtime_chat_service.complete_with_tool_call(
+            self._build_messages(request, base_date=base_date, canonical_fields=canonical_fields),
+            tools=[self._build_intent_tool_schema(canonical_fields)],
+            tool_choice={
+                "type": "function",
+                "function": {"name": STEWARD_INTENT_FUNCTION_NAME},
+            },
+            max_tokens=1800,
+            temperature=0.1,
+            timeout_seconds=18,
+            max_attempts=1,
+        )
+        self.last_call_traces = result.calls_as_dicts()
+        if result.tool_call is None or result.tool_call.name != STEWARD_INTENT_FUNCTION_NAME:
+            return None
+        return StewardIntentAgentResult(
+            payload=result.tool_call.arguments,
+            model_call_traces=self.last_call_traces,
+        )
+
+    @staticmethod
+    def _build_messages(
+        request: StewardPlanRequest,
+        *,
+        base_date: date,
+        canonical_fields: list[str],
+    ) -> list[dict[str, Any]]:
+        context_payload = {
+            "message": request.message,
+            "base_date": base_date.isoformat(),
+            "client_now_iso": request.client_now_iso,
+            "user_id": request.user_id,
+            "canonical_ontology_fields": canonical_fields,
+            "review_form_values": normalize_ontology_form_values(
+                request.context_json.get("review_form_values")
+            ),
+            "context_json": {
+                key: value
+                for key, value in request.context_json.items()
+                if key
+                in {
+                    "entry_source",
+                    "session_type",
+                    "role_codes",
+                    "username",
+                    "name",
+                    "department_name",
+                    "employee_grade",
+                    "employee_no",
+                    "client_timezone_offset_minutes",
+                }
+            },
+            "attachments": [
+                {
+                    "index": index + 1,
+                    "name": item.name,
+                    "media_type": item.media_type,
+                    "ocr_summary": item.ocr_summary,
+                    "ocr_fields": item.ocr_fields,
+                }
+                for index, item in enumerate(request.attachments)
+                if item.name
+            ],
+        }
+        return [
+            {
+                "role": "system",
+                "content": (
+                    "你是 X-Financial 的小财管家意图识别智能体。"
+                    "你必须通过 function calling 输出结构化计划，不能只返回普通文本。"
+                    "当前版本只支持 expense_application 和 reimbursement 两类任务；"
+                    "你只做识别、拆解、归集和确认点规划，不能执行入库、绑定附件或提交审批。"
+                    "所有 ontology_fields 只能使用调用方给出的 canonical_ontology_fields；"
+                    "如果输入里出现 occurred_date、transport_type、reason_value 等别名，必须映射为 canonical 字段。"
+                    "相对日期必须以 base_date 为准转换为明确日期。"
+                    "thinking_events 只能是面向用户的过程摘要，不能暴露内部推理链。"
+                ),
+            },
+            {
+                "role": "user",
+                "content": json.dumps(context_payload, ensure_ascii=False),
+            },
+        ]
+
+    @staticmethod
+    def _build_intent_tool_schema(canonical_fields: list[str]) -> dict[str, Any]:
+        return {
+            "type": "function",
+            "function": {
+                "name": STEWARD_INTENT_FUNCTION_NAME,
+                "description": "提交小财管家的复合财务意图识别结果。",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "thinking_events": {
+                            "type": "array",
+                            "items": {
+                                "type": "object",
+                                "properties": {
+                                    "stage": {"type": "string"},
+                                    "title": {"type": "string"},
+                                    "content": {"type": "string"},
+                                },
+                                "required": ["stage", "title", "content"],
+                            },
+                        },
+                        "tasks": {
+                            "type": "array",
+                            "items": {
+                                "type": "object",
+                                "properties": {
+                                    "task_type": {
+                                        "type": "string",
+                                        "enum": ["expense_application", "reimbursement"],
+                                    },
+                                    "title": {"type": "string"},
+                                    "summary": {"type": "string"},
+                                    "confidence": {
+                                        "type": "number",
+                                        "minimum": 0,
+                                        "maximum": 1,
+                                    },
+                                    "ontology_fields": {
+                                        "type": "object",
+                                        "additionalProperties": {"type": "string"},
+                                    },
+                                    "missing_fields": {
+                                        "type": "array",
+                                        "items": {
+                                            "type": "string",
+                                            "enum": canonical_fields,
+                                        },
+                                    },
+                                },
+                                "required": [
+                                    "task_type",
+                                    "title",
+                                    "summary",
+                                    "confidence",
+                                    "ontology_fields",
+                                    "missing_fields",
+                                ],
+                            },
+                        },
+                        "attachment_groups": {
+                            "type": "array",
+                            "items": {
+                                "type": "object",
+                                "properties": {
+                                    "target_task_index": {
+                                        "type": "integer",
+                                        "minimum": 1,
+                                    },
+                                    "scene": {"type": "string"},
+                                    "scene_label": {"type": "string"},
+                                    "attachment_names": {
+                                        "type": "array",
+                                        "items": {"type": "string"},
+                                    },
+                                    "excluded_attachment_names": {
+                                        "type": "array",
+                                        "items": {"type": "string"},
+                                    },
+                                    "confidence": {
+                                        "type": "number",
+                                        "minimum": 0,
+                                        "maximum": 1,
+                                    },
+                                    "rationale": {"type": "string"},
+                                },
+                                "required": [
+                                    "scene",
+                                    "scene_label",
+                                    "attachment_names",
+                                    "excluded_attachment_names",
+                                    "confidence",
+                                    "rationale",
+                                ],
+                            },
+                        },
+                    },
+                    "required": ["thinking_events", "tasks", "attachment_groups"],
+                },
+            },
+        }

server/src/app/services/steward_model_plan_builder.py

@@ -0,0 +1,365 @@
+from __future__ import annotations
+
+import re
+import uuid
+from datetime import date
+from typing import Any
+
+from app.schemas.steward import (
+    StewardAttachmentGroup,
+    StewardAttachmentInput,
+    StewardPlanRequest,
+    StewardPlanResponse,
+    StewardTask,
+    StewardThinkingEvent,
+)
+from app.services.ontology_field_registry import normalize_ontology_form_values
+from app.services.steward_constants import BUSINESS_CANONICAL_FIELDS
+from app.services.steward_intent_agent import StewardIntentAgentResult
+
+
+class StewardModelPlanBuilder:
+    """把模型 function calling 返回值转换为小财管家的服务端计划。"""
+
+    def __init__(self, planner: Any) -> None:
+        self.planner = planner
+
+    def build(
+        self,
+        intent_result: StewardIntentAgentResult,
+        *,
+        request: StewardPlanRequest,
+        base_date: date,
+    ) -> StewardPlanResponse | None:
+        tasks = self._build_tasks_from_model_payload(intent_result.payload, request, base_date)
+        if not tasks:
+            return None
+
+        attachment_groups = self._build_attachment_groups_from_model_payload(
+            intent_result.payload,
+            request.attachments,
+            tasks,
+        )
+        if request.attachments and not attachment_groups:
+            attachment_groups = self.planner._build_attachment_groups(request.attachments, tasks)
+        confirmation_groups = self.planner._build_confirmation_actions(tasks, attachment_groups)
+        thinking_events = self._build_llm_thinking_events(
+            intent_result.payload,
+            tasks=tasks,
+            attachment_groups=attachment_groups,
+            attachments=request.attachments,
+        )
+
+        return StewardPlanResponse(
+            plan_id=f"steward_plan_{uuid.uuid4().hex[:12]}",
+            plan_status="needs_confirmation" if confirmation_groups else "ready_to_delegate",
+            planning_source="llm_function_call",
+            summary=self.planner._build_summary(tasks, attachment_groups),
+            thinking_events=thinking_events,
+            tasks=tasks,
+            attachment_groups=attachment_groups,
+            confirmation_groups=confirmation_groups,
+            model_call_traces=intent_result.model_call_traces,
+        )
+
+    def _build_tasks_from_model_payload(
+        self,
+        payload: dict[str, Any],
+        request: StewardPlanRequest,
+        base_date: date,
+    ) -> list[StewardTask]:
+        raw_tasks = payload.get("tasks")
+        if not isinstance(raw_tasks, list):
+            return []
+
+        tasks: list[StewardTask] = []
+        for raw_task in raw_tasks:
+            if not isinstance(raw_task, dict):
+                continue
+            task_type = str(raw_task.get("task_type") or "").strip()
+            if task_type not in {"expense_application", "reimbursement"}:
+                continue
+
+            task_index = len(tasks) + 1
+            fields = self._sanitize_model_ontology_fields(
+                raw_task.get("ontology_fields"),
+                request=request,
+                base_date=base_date,
+            )
+            supplement_segment = " ".join(
+                [
+                    str(raw_task.get("title") or ""),
+                    str(raw_task.get("summary") or ""),
+                ]
+            )
+            supplement_fields = self.planner._extract_ontology_fields(
+                supplement_segment,
+                task_type,
+                base_date,
+                request,
+            )
+            for key, value in supplement_fields.items():
+                fields.setdefault(key, value)
+
+            assigned_agent = (
+                "application_assistant"
+                if task_type == "expense_application"
+                else "reimbursement_assistant"
+            )
+            task_id = f"task_{'app' if task_type == 'expense_application' else 'reim'}_{task_index:03d}"
+            title_prefix = "费用申请" if task_type == "expense_application" else "费用报销"
+            title = self.planner._clean_text(raw_task.get("title")) or self.planner._build_task_title(
+                title_prefix,
+                fields,
+                task_index,
+            )
+            summary = self.planner._clean_text(raw_task.get("summary")) or self.planner._build_task_summary(
+                supplement_segment,
+                fields,
+            )
+            missing_fields = self._sanitize_model_missing_fields(
+                raw_task.get("missing_fields"),
+                task_type=task_type,
+                fields=fields,
+            )
+            tasks.append(
+                StewardTask(
+                    task_id=task_id,
+                    task_type=task_type,  # type: ignore[arg-type]
+                    assigned_agent=assigned_agent,  # type: ignore[arg-type]
+                    title=title,
+                    summary=summary,
+                    status="needs_confirmation",
+                    confidence=self._resolve_model_confidence(
+                        raw_task.get("confidence"),
+                        segment=supplement_segment,
+                        fields=fields,
+                        task_type=task_type,
+                    ),
+                    ontology_fields=fields,
+                    missing_fields=missing_fields,
+                    confirmation_required=True,
+                )
+            )
+
+        return tasks
+
+    def _sanitize_model_ontology_fields(
+        self,
+        raw_fields: Any,
+        *,
+        request: StewardPlanRequest,
+        base_date: date,
+    ) -> dict[str, str]:
+        normalized_context = normalize_ontology_form_values(request.context_json.get("review_form_values"))
+        fields: dict[str, str] = {
+            key: value
+            for key, value in normalized_context.items()
+            if key in BUSINESS_CANONICAL_FIELDS and str(value or "").strip()
+        }
+        if not isinstance(raw_fields, dict):
+            return fields
+
+        normalized_model_fields = normalize_ontology_form_values(raw_fields)
+        for key, value in normalized_model_fields.items():
+            if key not in BUSINESS_CANONICAL_FIELDS:
+                continue
+            normalized_value = self._normalize_model_field_value(key, value, base_date)
+            if normalized_value:
+                fields[key] = normalized_value
+        if request.attachments and not fields.get("attachments"):
+            fields["attachments"] = "、".join(item.name for item in request.attachments if item.name)
+        return {key: value for key, value in fields.items() if key in BUSINESS_CANONICAL_FIELDS and value}
+
+    def _build_attachment_groups_from_model_payload(
+        self,
+        payload: dict[str, Any],
+        attachments: list[StewardAttachmentInput],
+        tasks: list[StewardTask],
+    ) -> list[StewardAttachmentGroup]:
+        raw_groups = payload.get("attachment_groups")
+        if not isinstance(raw_groups, list) or not attachments:
+            return []
+
+        uploaded_names = {item.name for item in attachments if item.name}
+        groups: list[StewardAttachmentGroup] = []
+        for raw_group in raw_groups:
+            if not isinstance(raw_group, dict):
+                continue
+            attachment_names = self._filter_uploaded_attachment_names(
+                raw_group.get("attachment_names"),
+                uploaded_names,
+            )
+            excluded_names = self._filter_uploaded_attachment_names(
+                raw_group.get("excluded_attachment_names"),
+                uploaded_names,
+            )
+            if not attachment_names and not excluded_names:
+                continue
+
+            scene = self.planner._clean_text(raw_group.get("scene")) or "other"
+            groups.append(
+                StewardAttachmentGroup(
+                    group_id=f"ag_{self._slug_scene(scene)}_{len(groups) + 1:03d}",
+                    target_task_id=self._resolve_model_group_target_task_id(raw_group, tasks),
+                    scene=scene,
+                    scene_label=self.planner._clean_text(raw_group.get("scene_label")) or "待确认费用",
+                    attachment_names=attachment_names,
+                    excluded_attachment_names=excluded_names,
+                    confidence=self._clamp_confidence(raw_group.get("confidence"), default=0.68),
+                    rationale=(
+                        self.planner._clean_text(raw_group.get("rationale"))
+                        or "模型根据附件线索生成归集建议。"
+                    ),
+                    confirmation_required=True,
+                )
+            )
+
+        return groups
+
+    def _build_llm_thinking_events(
+        self,
+        payload: dict[str, Any],
+        *,
+        tasks: list[StewardTask],
+        attachment_groups: list[StewardAttachmentGroup],
+        attachments: list[StewardAttachmentInput],
+    ) -> list[StewardThinkingEvent]:
+        events = [
+            StewardThinkingEvent(
+                event_id="intent_agent_function_call",
+                stage="llm_function_call",
+                title="意图识别智能体接管",
+                content=(
+                    "已调用系统主模型的 submit_steward_intent_plan 工具，"
+                    "把用户话术转换为可校验的结构化财务任务计划。"
+                ),
+            )
+        ]
+        raw_events = payload.get("thinking_events")
+        if isinstance(raw_events, list):
+            for raw_event in raw_events[:4]:
+                if not isinstance(raw_event, dict):
+                    continue
+                title = self.planner._clean_text(raw_event.get("title"))
+                content = self.planner._clean_text(raw_event.get("content"))
+                if not title or not content:
+                    continue
+                events.append(
+                    StewardThinkingEvent(
+                        event_id=f"intent_agent_model_{len(events):03d}",
+                        stage=self.planner._clean_text(raw_event.get("stage")) or "model_summary",
+                        title=title,
+                        content=content,
+                    )
+                )
+        if len(events) == 1:
+            events.extend(self.planner._build_thinking_events(tasks, attachment_groups, attachments)[1:])
+        return events
+
+    def _sanitize_model_missing_fields(
+        self,
+        raw_missing_fields: Any,
+        *,
+        task_type: str,
+        fields: dict[str, str],
+    ) -> list[str]:
+        missing_fields: list[str] = []
+        if isinstance(raw_missing_fields, list):
+            for item in raw_missing_fields:
+                key = str(item or "").strip()
+                if key in BUSINESS_CANONICAL_FIELDS and key not in missing_fields and not fields.get(key):
+                    missing_fields.append(key)
+        for key in self.planner._resolve_missing_fields(task_type, fields):
+            if key not in missing_fields:
+                missing_fields.append(key)
+        return missing_fields
+
+    def _resolve_model_confidence(
+        self,
+        value: Any,
+        *,
+        segment: str,
+        fields: dict[str, str],
+        task_type: str,
+    ) -> float:
+        return self._clamp_confidence(
+            value,
+            default=self.planner._resolve_task_confidence(segment, fields, task_type),
+        )
+
+    def _normalize_model_field_value(self, key: str, value: Any, base_date: date) -> str:
+        cleaned = self.planner._clean_text(value)
+        if not cleaned:
+            return ""
+        if key == "time_range":
+            return self.planner._extract_time_range(cleaned, base_date) or cleaned
+        if key == "expense_type":
+            return self._normalize_expense_type_value(cleaned)
+        if key == "transport_mode":
+            return self._normalize_transport_mode_value(cleaned)
+        return cleaned
+
+    @staticmethod
+    def _normalize_expense_type_value(value: str) -> str:
+        normalized = str(value or "").strip().lower()
+        if normalized in {"travel", "travel_application", "差旅", "差旅费", "出差"}:
+            return "travel"
+        if normalized in {"transport", "traffic", "交通", "交通费", "打车", "出租车"}:
+            return "transport"
+        if normalized in {"entertainment", "meal", "招待", "接待", "餐饮", "业务招待"}:
+            return "entertainment"
+        if normalized in {"office", "办公", "办公用品"}:
+            return "office"
+        return normalized
+
+    @staticmethod
+    def _normalize_transport_mode_value(value: str) -> str:
+        normalized = str(value or "").strip().lower()
+        if normalized in {"train", "高铁", "动车", "火车"}:
+            return "train"
+        if normalized in {"flight", "air", "飞机", "机票", "航班"}:
+            return "flight"
+        if normalized in {"taxi", "出租车", "的士", "网约车", "打车"}:
+            return "taxi"
+        if normalized in {"subway", "地铁"}:
+            return "subway"
+        return normalized
+
+    @staticmethod
+    def _filter_uploaded_attachment_names(raw_names: Any, uploaded_names: set[str]) -> list[str]:
+        if not isinstance(raw_names, list):
+            return []
+        names: list[str] = []
+        for raw_name in raw_names:
+            name = str(raw_name or "").strip()
+            if name in uploaded_names and name not in names:
+                names.append(name)
+        return names
+
+    @staticmethod
+    def _resolve_model_group_target_task_id(raw_group: dict[str, Any], tasks: list[StewardTask]) -> str | None:
+        try:
+            target_index = int(raw_group.get("target_task_index") or 0)
+        except (TypeError, ValueError):
+            target_index = 0
+        if target_index > 0 and target_index <= len(tasks):
+            return tasks[target_index - 1].task_id
+
+        target_task_id = str(raw_group.get("target_task_id") or "").strip()
+        if target_task_id and any(task.task_id == target_task_id for task in tasks):
+            return target_task_id
+        return None
+
+    @staticmethod
+    def _slug_scene(value: str) -> str:
+        normalized = re.sub(r"[^a-zA-Z0-9_]+", "_", str(value or "").strip().lower()).strip("_")
+        return normalized or "other"
+
+    @staticmethod
+    def _clamp_confidence(value: Any, *, default: float) -> float:
+        try:
+            parsed = float(value)
+        except (TypeError, ValueError):
+            parsed = default
+        return round(min(1.0, max(0.0, parsed)), 2)

server/src/app/services/steward_planner.py

@@ -0,0 +1,645 @@
+from __future__ import annotations
+
+import re
+import uuid
+from dataclasses import dataclass
+from datetime import UTC, date, datetime, timedelta
+from typing import Any
+
+from app.schemas.steward import (
+    StewardAttachmentGroup,
+    StewardAttachmentInput,
+    StewardConfirmationAction,
+    StewardPlanRequest,
+    StewardPlanResponse,
+    StewardTask,
+    StewardThinkingEvent,
+)
+from app.services.steward_constants import BUSINESS_CANONICAL_FIELD_ORDER, BUSINESS_CANONICAL_FIELDS
+from app.services.ontology_field_registry import normalize_ontology_form_values
+from app.services.steward_intent_agent import StewardIntentAgent
+from app.services.steward_model_plan_builder import StewardModelPlanBuilder
+
+
+CITY_NAMES = (
+    "北京",
+    "上海",
+    "广州",
+    "深圳",
+    "杭州",
+    "南京",
+    "苏州",
+    "成都",
+    "重庆",
+    "天津",
+    "武汉",
+    "西安",
+    "长沙",
+    "郑州",
+    "青岛",
+    "厦门",
+    "福州",
+    "合肥",
+    "济南",
+    "沈阳",
+    "大连",
+    "宁波",
+    "无锡",
+)
+
+APPLICATION_SPLIT_PATTERN = re.compile(r"(?:^|[，,。；;])[^，,。；;]*?(?:申请|出差申请|差旅申请)[^，,。；;]*")
+REIMBURSEMENT_PATTERN = re.compile(r"(?:我要报销|还需要报销|需要报销|报销)([^，,。；;！？?!\n]+)")
+MONTH_DAY_PATTERN = re.compile(r"(?P<month>\d{1,2})\s*月\s*(?P<day>\d{1,2})\s*(?:日|号)?")
+ISO_DATE_PATTERN = re.compile(r"(?P<year>\d{4})[-/年](?P<month>\d{1,2})[-/月](?P<day>\d{1,2})(?:日)?")
+
+
+@dataclass(frozen=True)
+class PlannedTaskDraft:
+    task_type: str
+    segment: str
+    index: int
+
+
+class StewardPlannerService:
+    """小财管家第一版规划服务：只生成计划，不执行入库类动作。"""
+
+    def __init__(self, intent_agent: StewardIntentAgent | None = None) -> None:
+        self.intent_agent = intent_agent
+
+    def build_plan(self, request: StewardPlanRequest) -> StewardPlanResponse:
+        message = self._clean_text(request.message)
+        if not message:
+            raise ValueError("小财管家需要一段任务描述。")
+
+        base_date = self._resolve_base_date(request.client_now_iso, request.context_json)
+        model_call_traces: list[dict[str, Any]] = []
+        fallback_reason = ""
+        if self.intent_agent is not None:
+            try:
+                intent_result = self.intent_agent.detect(
+                    request,
+                    base_date=base_date,
+                    canonical_fields=list(BUSINESS_CANONICAL_FIELD_ORDER),
+                )
+                if intent_result is not None:
+                    model_call_traces = intent_result.model_call_traces
+                    llm_plan = StewardModelPlanBuilder(self).build(
+                        intent_result,
+                        request=request,
+                        base_date=base_date,
+                    )
+                    if llm_plan is not None:
+                        return llm_plan
+                model_call_traces = getattr(self.intent_agent, "last_call_traces", []) or model_call_traces
+                fallback_reason = "主模型未返回可用的 function calling 计划，已切换到规则兜底。"
+            except Exception as exc:
+                model_call_traces = getattr(self.intent_agent, "last_call_traces", []) or model_call_traces
+                fallback_reason = f"主模型 function calling 调用失败，已切换到规则兜底：{exc}"
+
+        return self._build_rule_fallback_plan(
+            request,
+            base_date=base_date,
+            model_call_traces=model_call_traces,
+            fallback_reason=fallback_reason,
+        )
+
+    def _build_rule_fallback_plan(
+        self,
+        request: StewardPlanRequest,
+        *,
+        base_date: date,
+        model_call_traces: list[dict[str, Any]] | None = None,
+        fallback_reason: str = "",
+    ) -> StewardPlanResponse:
+        message = self._clean_text(request.message)
+        task_drafts = self._extract_task_drafts(message)
+        tasks = [self._build_task(draft, base_date, request) for draft in task_drafts]
+        if not tasks:
+            tasks = [self._build_fallback_task(message, base_date, request)]
+
+        attachment_groups = self._build_attachment_groups(request.attachments, tasks)
+        confirmation_groups = self._build_confirmation_actions(tasks, attachment_groups)
+        thinking_events = self._build_thinking_events(tasks, attachment_groups, request.attachments)
+        if fallback_reason:
+            thinking_events.insert(
+                0,
+                StewardThinkingEvent(
+                    event_id="intent_agent_rule_fallback",
+                    stage="rule_fallback",
+                    title="意图识别智能体进入兜底模式",
+                    content=fallback_reason,
+                ),
+            )
+        plan_id = f"steward_plan_{uuid.uuid4().hex[:12]}"
+
+        return StewardPlanResponse(
+            plan_id=plan_id,
+            plan_status="needs_confirmation" if confirmation_groups else "ready_to_delegate",
+            planning_source="rule_fallback",
+            summary=self._build_summary(tasks, attachment_groups),
+            thinking_events=thinking_events,
+            tasks=tasks,
+            attachment_groups=attachment_groups,
+            confirmation_groups=confirmation_groups,
+            model_call_traces=model_call_traces or [],
+        )
+
+    def _extract_task_drafts(self, message: str) -> list[PlannedTaskDraft]:
+        drafts: list[PlannedTaskDraft] = []
+        first_reimbursement = self._find_first_reimbursement_index(message)
+        application_source = message[:first_reimbursement] if first_reimbursement >= 0 else message
+        if self._looks_like_application(application_source):
+            drafts.append(
+                PlannedTaskDraft(
+                    task_type="expense_application",
+                    segment=application_source.strip("，,。；; "),
+                    index=len(drafts) + 1,
+                )
+            )
+
+        for match in REIMBURSEMENT_PATTERN.finditer(message):
+            segment = f"报销{match.group(1)}"
+            drafts.append(
+                PlannedTaskDraft(
+                    task_type="reimbursement",
+                    segment=segment.strip("，,。；; "),
+                    index=len(drafts) + 1,
+                )
+            )
+
+        return drafts
+
+    @staticmethod
+    def _find_first_reimbursement_index(message: str) -> int:
+        candidates = [message.find(item) for item in ("我要报销", "还需要报销", "需要报销", "报销")]
+        positives = [item for item in candidates if item >= 0]
+        return min(positives) if positives else -1
+
+    @staticmethod
+    def _looks_like_application(text: str) -> bool:
+        compact = re.sub(r"\s+", "", text)
+        return bool(compact) and "申请" in compact and bool(re.search(r"出差|差旅|费用|交通|住宿|采购|会务|会议", compact))
+
+    def _build_task(
+        self,
+        draft: PlannedTaskDraft,
+        base_date: date,
+        request: StewardPlanRequest,
+    ) -> StewardTask:
+        fields = self._extract_ontology_fields(draft.segment, draft.task_type, base_date, request)
+        missing_fields = self._resolve_missing_fields(draft.task_type, fields)
+        task_id = f"task_{'app' if draft.task_type == 'expense_application' else 'reim'}_{draft.index:03d}"
+        assigned_agent = (
+            "application_assistant"
+            if draft.task_type == "expense_application"
+            else "reimbursement_assistant"
+        )
+        title_prefix = "费用申请" if draft.task_type == "expense_application" else "费用报销"
+        title = self._build_task_title(title_prefix, fields, draft.index)
+        return StewardTask(
+            task_id=task_id,
+            task_type=draft.task_type,  # type: ignore[arg-type]
+            assigned_agent=assigned_agent,  # type: ignore[arg-type]
+            title=title,
+            summary=self._build_task_summary(draft.segment, fields),
+            status="needs_confirmation",
+            confidence=self._resolve_task_confidence(draft.segment, fields, draft.task_type),
+            ontology_fields=fields,
+            missing_fields=missing_fields,
+            confirmation_required=True,
+        )
+
+    def _build_fallback_task(
+        self,
+        message: str,
+        base_date: date,
+        request: StewardPlanRequest,
+    ) -> StewardTask:
+        task_type = "reimbursement" if "报销" in message or request.attachments else "expense_application"
+        draft = PlannedTaskDraft(task_type=task_type, segment=message, index=1)
+        task = self._build_task(draft, base_date, request)
+        return task.model_copy(update={"confidence": min(task.confidence, 0.58)})
+
+    def _extract_ontology_fields(
+        self,
+        segment: str,
+        task_type: str,
+        base_date: date,
+        request: StewardPlanRequest,
+    ) -> dict[str, str]:
+        normalized_context = normalize_ontology_form_values(request.context_json.get("review_form_values"))
+        fields: dict[str, str] = {
+            key: value
+            for key, value in normalized_context.items()
+            if key in BUSINESS_CANONICAL_FIELDS and str(value or "").strip()
+        }
+        expense_type = self._infer_expense_type(segment, task_type)
+        if expense_type and not fields.get("expense_type"):
+            fields["expense_type"] = expense_type
+        time_range = self._extract_time_range(segment, base_date)
+        if time_range and not fields.get("time_range"):
+            fields["time_range"] = time_range
+        location = self._extract_location(segment)
+        if location and not fields.get("location"):
+            fields["location"] = location
+        reason = self._extract_reason(segment, task_type)
+        if reason and not fields.get("reason"):
+            fields["reason"] = reason
+        transport_mode = self._extract_transport_mode(segment)
+        if transport_mode and not fields.get("transport_mode"):
+            fields["transport_mode"] = transport_mode
+        if request.attachments:
+            fields["attachments"] = "、".join(item.name for item in request.attachments if item.name)
+
+        return {key: value for key, value in fields.items() if key in BUSINESS_CANONICAL_FIELDS and value}
+
+    @staticmethod
+    def _infer_expense_type(segment: str, task_type: str) -> str:
+        compact = re.sub(r"\s+", "", segment)
+        if re.search(r"招待|接待|餐饮|宴请|客户吃饭|业务餐", compact):
+            return "entertainment"
+        if re.search(r"出差|差旅|住宿|酒店|机票|航班|高铁|火车", compact):
+            return "travel"
+        if re.search(r"交通|出租车|的士|网约车|打车|地铁|公交", compact):
+            return "transport" if task_type == "reimbursement" else "travel"
+        return "travel" if task_type == "expense_application" else "other"
+
+    def _extract_time_range(self, segment: str, base_date: date) -> str:
+        compact = re.sub(r"\s+", "", segment)
+        if "昨天" in compact:
+            return (base_date - timedelta(days=1)).isoformat()
+        if "前天" in compact:
+            return (base_date - timedelta(days=2)).isoformat()
+        if "明天" in compact:
+            return (base_date + timedelta(days=1)).isoformat()
+        if "后天" in compact:
+            return (base_date + timedelta(days=2)).isoformat()
+
+        iso_match = ISO_DATE_PATTERN.search(compact)
+        if iso_match:
+            return self._safe_date(
+                int(iso_match.group("year")),
+                int(iso_match.group("month")),
+                int(iso_match.group("day")),
+            )
+
+        month_day = MONTH_DAY_PATTERN.search(compact)
+        if month_day:
+            return self._safe_date(
+                base_date.year,
+                int(month_day.group("month")),
+                int(month_day.group("day")),
+            )
+        return ""
+
+    @staticmethod
+    def _safe_date(year: int, month: int, day: int) -> str:
+        try:
+            return date(year, month, day).isoformat()
+        except ValueError:
+            return ""
+
+    @staticmethod
+    def _extract_location(segment: str) -> str:
+        compact = re.sub(r"\s+", "", segment)
+        for prefix in ("去", "到", "赴", "前往"):
+            match = re.search(fr"{prefix}({'|'.join(CITY_NAMES)})", compact)
+            if match:
+                return match.group(1)
+        for city in CITY_NAMES:
+            if city in compact:
+                return city
+        return ""
+
+    @staticmethod
+    def _extract_reason(segment: str, task_type: str) -> str:
+        cleaned = re.sub(r"\s+", "", segment).strip("，,。；; ")
+        if task_type == "expense_application":
+            match = re.search(r"(辅助|支持|协助|支撑|参加|拜访|调研|实施|部署|审核).+", cleaned)
+            if match:
+                return StewardPlannerService._strip_trailing_connectors(match.group(0))
+            reason = re.sub(r"^.*?(?:出差|差旅)", "", cleaned).strip("，,。；;的费用")
+            return StewardPlannerService._strip_trailing_connectors(reason) or cleaned
+        cleaned = re.sub(r"^报销", "", cleaned)
+        cleaned = re.sub(r"^(?:昨天|前天|明天|后天|\d{1,2}月\d{1,2}(?:日|号)?)的?", "", cleaned)
+        return cleaned.strip("，,。；; ") or segment.strip()
+
+    @staticmethod
+    def _strip_trailing_connectors(value: str) -> str:
+        cleaned = str(value or "").strip("，,。；; ")
+        return re.sub(r"(?:并且|而且|同时|另外|还需要|需要)$", "", cleaned).strip("，,。；; ")
+
+    @staticmethod
+    def _extract_transport_mode(segment: str) -> str:
+        compact = re.sub(r"\s+", "", segment)
+        if re.search(r"高铁|动车|火车", compact):
+            return "train"
+        if re.search(r"飞机|机票|航班", compact):
+            return "flight"
+        if re.search(r"出租车|的士|网约车|打车", compact):
+            return "taxi"
+        if "交通" in compact:
+            return "other"
+        return ""
+
+    @staticmethod
+    def _resolve_missing_fields(task_type: str, fields: dict[str, str]) -> list[str]:
+        required = ["expense_type", "time_range", "reason"]
+        if task_type == "expense_application":
+            required.append("location")
+        return [key for key in required if not str(fields.get(key) or "").strip()]
+
+    @staticmethod
+    def _resolve_task_confidence(segment: str, fields: dict[str, str], task_type: str) -> float:
+        compact = re.sub(r"\s+", "", segment)
+        intent_score = 1.0 if ("申请" in compact if task_type == "expense_application" else "报销" in compact) else 0.45
+        time_score = 1.0 if fields.get("time_range") else 0.0
+        location_score = 1.0 if fields.get("location") else 0.2
+        scene_score = 1.0 if fields.get("expense_type") and fields["expense_type"] != "other" else 0.35
+        confidence = min(1.0, 0.35 * intent_score + 0.25 * time_score + 0.2 * location_score + 0.2 * scene_score)
+        return round(max(0.45, confidence), 2)
+
+    def _build_attachment_groups(
+        self,
+        attachments: list[StewardAttachmentInput],
+        tasks: list[StewardTask],
+    ) -> list[StewardAttachmentGroup]:
+        if not attachments:
+            return []
+
+        classified = [(item, self._classify_attachment(item)) for item in attachments if item.name]
+        travel_related = [item.name for item, scene in classified if scene in {"travel", "transport"}]
+        excluded = [item.name for item, scene in classified if scene not in {"travel", "transport"}]
+        target_task = self._resolve_attachment_target_task(tasks)
+
+        groups: list[StewardAttachmentGroup] = []
+        if travel_related:
+            confidence = 0.72 + min(0.18, len(travel_related) * 0.04)
+            groups.append(
+                StewardAttachmentGroup(
+                    group_id="ag_travel_001",
+                    target_task_id=target_task.task_id if target_task else None,
+                    scene="travel",
+                    scene_label="差旅相关费用",
+                    attachment_names=travel_related,
+                    excluded_attachment_names=excluded,
+                    confidence=round(confidence, 2),
+                    rationale="附件名称或 OCR 摘要中包含差旅、交通、住宿、火车、机票等线索。",
+                    confirmation_required=True,
+                )
+            )
+        elif excluded:
+            groups.append(
+                StewardAttachmentGroup(
+                    group_id="ag_other_001",
+                    target_task_id=None,
+                    scene="other",
+                    scene_label="待人工确认费用",
+                    attachment_names=excluded,
+                    excluded_attachment_names=[],
+                    confidence=0.5,
+                    rationale="当前附件缺少可稳定归属到申请或报销任务的差旅线索。",
+                    confirmation_required=True,
+                )
+            )
+        return groups
+
+    @staticmethod
+    def _resolve_attachment_target_task(tasks: list[StewardTask]) -> StewardTask | None:
+        reimbursement_tasks = [item for item in tasks if item.task_type == "reimbursement"]
+        for task in reimbursement_tasks:
+            if task.ontology_fields.get("expense_type") == "travel":
+                return task
+        return reimbursement_tasks[0] if reimbursement_tasks else None
+
+    @staticmethod
+    def _classify_attachment(attachment: StewardAttachmentInput) -> str:
+        text = " ".join(
+            [
+                attachment.name,
+                attachment.media_type,
+                attachment.ocr_summary,
+                " ".join(f"{key}:{value}" for key, value in attachment.ocr_fields.items()),
+            ]
+        )
+        compact = re.sub(r"\s+", "", text).lower()
+        if re.search(r"招待|接待|餐饮|宴请|客户|meal|entertainment", compact):
+            return "entertainment"
+        if re.search(r"酒店|住宿|差旅|出差|高铁|火车|动车|机票|航班|train|flight|hotel|travel", compact):
+            return "travel"
+        if re.search(r"出租车|的士|网约车|打车|交通|taxi|transport", compact):
+            return "transport"
+        return "other"
+
+    def _build_confirmation_actions(
+        self,
+        tasks: list[StewardTask],
+        attachment_groups: list[StewardAttachmentGroup],
+    ) -> list[StewardConfirmationAction]:
+        actions: list[StewardConfirmationAction] = []
+        for task in tasks:
+            if task.task_type == "expense_application":
+                action_type = "confirm_create_application"
+                label = "确认创建申请单"
+            else:
+                action_type = "confirm_create_reimbursement_draft"
+                label = "确认创建报销草稿"
+            actions.append(
+                StewardConfirmationAction(
+                    confirmation_id=f"confirm_{task.task_id}",
+                    action_type=action_type,
+                    label=label,
+                    description=f"确认后把“{task.title}”交给{self._agent_label(task.assigned_agent)}继续核对。",
+                    target_task_id=task.task_id,
+                    payload={
+                        "task_id": task.task_id,
+                        "task_type": task.task_type,
+                        "assigned_agent": task.assigned_agent,
+                        "ontology_fields": task.ontology_fields,
+                    },
+                )
+            )
+
+        for group in attachment_groups:
+            actions.append(
+                StewardConfirmationAction(
+                    confirmation_id=f"confirm_{group.group_id}",
+                    action_type="confirm_attachment_group",
+                    label="确认附件归集",
+                    description=f"确认后将 {len(group.attachment_names)} 份附件按“{group.scene_label}”归集。",
+                    target_task_id=group.target_task_id,
+                    attachment_group_id=group.group_id,
+                    payload={
+                        "attachment_group_id": group.group_id,
+                        "target_task_id": group.target_task_id,
+                        "attachment_names": group.attachment_names,
+                        "excluded_attachment_names": group.excluded_attachment_names,
+                    },
+                )
+            )
+        return actions
+
+    @staticmethod
+    def _agent_label(assigned_agent: str) -> str:
+        return "申请助手" if assigned_agent == "application_assistant" else "报销助手"
+
+    def _build_thinking_events(
+        self,
+        tasks: list[StewardTask],
+        attachment_groups: list[StewardAttachmentGroup],
+        attachments: list[StewardAttachmentInput],
+    ) -> list[StewardThinkingEvent]:
+        application_count = sum(1 for item in tasks if item.task_type == "expense_application")
+        reimbursement_count = sum(1 for item in tasks if item.task_type == "reimbursement")
+        task_intent_summary = self._summarize_task_intents(tasks)
+        ontology_summary = self._summarize_ontology_coverage(tasks)
+        delegation_summary = self._summarize_delegation_targets(tasks)
+        events = [
+            StewardThinkingEvent(
+                event_id="intent_agent_entry",
+                stage="intent_agent",
+                title="意图识别智能体接管",
+                content=(
+                    f"检测到复合财务话术，当前不是单一助手会话；"
+                    f"已进入小财管家编排模式，候选任务共 {len(tasks)} 个。"
+                ),
+            ),
+            StewardThinkingEvent(
+                event_id="intent_task_split",
+                stage="task_split",
+                title=f"拆分申请 {application_count} 个、报销 {reimbursement_count} 个",
+                content=task_intent_summary,
+            ),
+            StewardThinkingEvent(
+                event_id="intent_ontology_mapping",
+                stage="ontology_mapping",
+                title="映射业务本体字段",
+                content=ontology_summary,
+            ),
+        ]
+        if attachments:
+            events.append(
+                StewardThinkingEvent(
+                    event_id="intent_attachment_correlation",
+                    stage="attachment_correlation",
+                    title="关联附件与任务线索",
+                    content=self._summarize_attachment_correlation(attachment_groups, len(attachments)),
+                )
+            )
+        events.append(
+            StewardThinkingEvent(
+                event_id="intent_delegation_gate",
+                stage="delegation_gate",
+                title="生成确认点并准备分派",
+                content=f"{delegation_summary} 创建单据、生成草稿、绑定附件和提交审批都会等待用户确认。",
+            )
+        )
+        return events
+
+    @staticmethod
+    def _summarize_task_intents(tasks: list[StewardTask]) -> str:
+        if not tasks:
+            return "当前输入尚未形成稳定任务，先保留为待确认财务事项。"
+        parts = []
+        for task in tasks:
+            task_label = "申请" if task.task_type == "expense_application" else "报销"
+            fields = task.ontology_fields
+            anchors = []
+            if fields.get("time_range"):
+                anchors.append(fields["time_range"])
+            if fields.get("location"):
+                anchors.append(fields["location"])
+            if fields.get("expense_type"):
+                anchors.append(fields["expense_type"])
+            anchor_text = "、".join(anchors) if anchors else "待补充关键字段"
+            parts.append(f"{task_label}：{task.title}（{anchor_text}）")
+        return "；".join(parts)
+
+    @staticmethod
+    def _summarize_ontology_coverage(tasks: list[StewardTask]) -> str:
+        canonical_keys = []
+        missing_keys = []
+        for task in tasks:
+            canonical_keys.extend(task.ontology_fields.keys())
+            missing_keys.extend(task.missing_fields)
+        unique_keys = sorted({item for item in canonical_keys if item})
+        unique_missing = sorted({item for item in missing_keys if item})
+        mapped = "、".join(unique_keys) if unique_keys else "暂无稳定字段"
+        missing = "；缺失字段：" + "、".join(unique_missing) if unique_missing else ""
+        return f"已使用 canonical ontology fields：{mapped}{missing}。兼容字段只作为输入别名，不直接进入业务逻辑。"
+
+    @staticmethod
+    def _summarize_attachment_correlation(
+        attachment_groups: list[StewardAttachmentGroup],
+        total_attachment_count: int,
+    ) -> str:
+        grouped_names = []
+        excluded_names = []
+        for group in attachment_groups:
+            grouped_names.extend(group.attachment_names)
+            excluded_names.extend(group.excluded_attachment_names)
+        grouped_text = "、".join(grouped_names) if grouped_names else "暂无可稳定归集附件"
+        excluded_text = "；排除或单独确认：" + "、".join(excluded_names) if excluded_names else ""
+        return f"已核对 {total_attachment_count} 份附件，建议归集：{grouped_text}{excluded_text}。"
+
+    @staticmethod
+    def _summarize_delegation_targets(tasks: list[StewardTask]) -> str:
+        application_count = sum(1 for item in tasks if item.assigned_agent == "application_assistant")
+        reimbursement_count = sum(1 for item in tasks if item.assigned_agent == "reimbursement_assistant")
+        parts = []
+        if application_count:
+            parts.append(f"{application_count} 个申请任务交给申请助手")
+        if reimbursement_count:
+            parts.append(f"{reimbursement_count} 个报销任务交给报销助手")
+        return "；".join(parts) + "。" if parts else "尚无可分派任务。"
+
+    @staticmethod
+    def _build_summary(tasks: list[StewardTask], attachment_groups: list[StewardAttachmentGroup]) -> str:
+        parts = [f"我识别到 {len(tasks)} 个待处理任务"]
+        if attachment_groups:
+            grouped = sum(len(item.attachment_names) for item in attachment_groups)
+            parts.append(f"并形成 {grouped} 份附件的归集建议")
+        parts.append("，请确认后我再分派给对应助手执行。")
+        return "".join(parts)
+
+    @staticmethod
+    def _build_task_title(prefix: str, fields: dict[str, str], index: int) -> str:
+        location = fields.get("location", "")
+        time_range = fields.get("time_range", "")
+        expense_type = fields.get("expense_type", "")
+        subject = location or {"travel": "差旅", "transport": "交通", "entertainment": "招待"}.get(expense_type, "")
+        if subject and time_range:
+            return f"{prefix} {time_range} {subject}"
+        if subject:
+            return f"{prefix} {subject}"
+        return f"{prefix} {index}"
+
+    @staticmethod
+    def _build_task_summary(segment: str, fields: dict[str, str]) -> str:
+        field_parts = []
+        for key, label in (
+            ("time_range", "时间"),
+            ("location", "地点"),
+            ("expense_type", "费用类型"),
+            ("reason", "事由"),
+            ("transport_mode", "交通方式"),
+        ):
+            value = fields.get(key)
+            if value:
+                field_parts.append(f"{label}：{value}")
+        return "；".join(field_parts) or segment
+
+    @staticmethod
+    def _resolve_base_date(client_now_iso: str | None, context_json: dict[str, Any]) -> date:
+        raw_value = client_now_iso or str(context_json.get("client_now_iso") or "").strip()
+        if raw_value:
+            try:
+                parsed = datetime.fromisoformat(raw_value.replace("Z", "+00:00"))
+                return parsed.date()
+            except ValueError:
+                pass
+        return datetime.now(UTC).date()
+
+    @staticmethod
+    def _clean_text(value: Any) -> str:
+        return re.sub(r"\s+", " ", str(value or "")).strip()