feat(server): 会话上下文保留（LLM 历史 + 确定性兜底双保险）

解决用户删除草稿后说'再提交'丢失上下文的问题：

- steward.py 新增 _inject_recent_conversation_history：build_plan 前读最近 10 条对话注入 context_json
- steward_intent_agent.py 的 _build_messages 把 recent_history 暴露给模型，system prompt 加确认类话术引导
- 新建 steward_context_resume.py：should_resume_recent_task 检测'再提交'类话术 + state 有可恢复 flow，attach_resumed_task 从 state 恢复 task
- 两个 plan 入口（/plans 和 /plans/stream）都已接入双保险
- 后端 67 passed，端到端验证'上海出差→再提交'成功恢复 task

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

@@ -29,6 +29,10 @@ from app.services.agent_conversations import AgentConversationService
 from app.services.expense_claim_draft_flow import APPROVED_APPLICATION_LINK_STATUSES
 from app.services.expense_claims import ExpenseClaimService
 from app.services.runtime_chat import RuntimeChatService
+from app.services.steward_context_resume import (
+    attach_resumed_task,
+    should_resume_recent_task,
+)
 from app.services.steward_flow_state import StewardFlowStateService
 from app.services.steward_graph_action_runtime import StewardGraphActionRuntime
 from app.services.steward_graph_planner import StewardGraphPlannerService
@@ -61,7 +65,9 @@ def create_steward_plan(payload: StewardPlanRequest, db: DbSession) -> StewardPl
     try:
         planner = _build_steward_planner(db)
         hydrated_payload = _hydrate_required_application_gate(db, payload, planner)
+        hydrated_payload = _inject_recent_conversation_history(db, hydrated_payload)
         plan = planner.build_plan(hydrated_payload)
+        plan = _apply_context_resume(db, hydrated_payload, plan)
         return _attach_conversation_state(db, hydrated_payload, plan)
     except ValueError as exc:
         raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail=str(exc)) from exc
@@ -143,7 +149,9 @@ async def _iter_steward_plan_events(
 
     try:
         hydrated_payload = _hydrate_required_application_gate(db, payload, planner)
+        hydrated_payload = _inject_recent_conversation_history(db, hydrated_payload)
         plan = planner.build_plan(hydrated_payload)
+        plan = _apply_context_resume(db, hydrated_payload, plan)
         plan = _attach_conversation_state(db, hydrated_payload, plan)
     except ValueError as exc:
         yield _encode_stream_event("error", {"message": str(exc)})
@@ -495,3 +503,66 @@ def _resolve_current_steward_state(
         return stored_state
     incoming_state = context_json.get("steward_state") or context_json.get("stewardState")
     return incoming_state if isinstance(incoming_state, dict) else {}
+
+
+def _inject_recent_conversation_history(
+    db: Session,
+    payload: StewardPlanRequest,
+) -> StewardPlanRequest:
+    """读取会话最近 10 条对话历史，注入 context_json.recent_history 供 LLM 关联上下文。
+
+    历史只给模型用，不返回前端。在 get_or_create_conversation 之前读取，
+    使用前端传入的 conversation_id，避免把本轮消息算进历史。
+    """
+    context_json = dict(payload.context_json or {})
+    conversation_id = _resolve_conversation_id(context_json)
+    if not conversation_id:
+        return payload
+    try:
+        recent_history = AgentConversationService(db).list_message_history(
+            conversation_id,
+            limit=10,
+        )
+    except Exception:
+        recent_history = []
+    if not recent_history:
+        return payload
+    return payload.model_copy(
+        update={
+            "context_json": {
+                **context_json,
+                "recent_history": recent_history,
+            }
+        }
+    )
+
+
+def _apply_context_resume(
+    db: Session,
+    payload: StewardPlanRequest,
+    plan: StewardPlanResponse,
+) -> StewardPlanResponse:
+    """确定性兜底：若 plan 无 task 且用户说"再提交"类话术，从会话状态恢复最近 task。
+
+    不依赖 LLM 理解力，100% 可靠地恢复上下文。LLM 注入历史（保险②）覆盖更模糊话术。
+    """
+    if plan.tasks or plan.candidate_flows:
+        return plan
+    context_json = dict(payload.context_json or {})
+    conversation_id = _resolve_conversation_id(context_json)
+    if not conversation_id:
+        return plan
+    try:
+        conversation = AgentConversationService(db).get_conversation(conversation_id)
+    except Exception:
+        conversation = None
+    if conversation is None:
+        return plan
+    current_state = _resolve_current_steward_state(
+        conversation.state_json if isinstance(conversation.state_json, dict) else {},
+        context_json,
+    )
+    resume_flow_id = should_resume_recent_task(payload.message, current_state)
+    if not resume_flow_id:
+        return plan
+    return attach_resumed_task(plan, current_state, resume_flow_id)