feat(ai): issue verified application preview decisions

2026-07-14 14:37:53 +08:00
parent a662cfe6c3
commit 5b24630710
32 changed files with 1976 additions and 217 deletions
--- a/document/development/2026-07-13/feature/ai-expense-closed-loop-and-value-proof/CONCEPT.md
+++ b/document/development/2026-07-13/feature/ai-expense-closed-loop-and-value-proof/CONCEPT.md
@@ -209,6 +209,7 @@
 - `expense_cases`：统一费用事件和当前阶段。
 - `expense_case_links`：连接申请、报销、票据、审批、付款、凭证和外部对象。
 - `business_events`：append-only 业务事件账本。
+- `ai_application_preview_decisions`：申请落单前由服务端重新规范化并短期签发的预览建议；不创建空 Expense Case，不保存建议明文，动作成功后一次消费。
 - `ai_decisions`：结构化 AI 建议与版本证据。
 - `ai_decision_feedback`：接受、修改、拒绝和忽略。
 - `workflow_outcomes`：退回、补件、审批、付款、审计和最终结果。
@@ -250,6 +251,8 @@
 - 旧 `ReimbursementRequest` 进入只读兼容和迁移状态，停止新增第二套业务编排。
 - 现有 `risk_flags_json` 保持读取兼容，新审批、付款、归档和关系事件写入结构化表。
 - API 新字段优先追加，不在同一阶段破坏现有前端契约。
+- 从未成功签发 `decision_id` 的旧申请预览，或签发接口失败且服务端没有当前会话有效决策时，保存/提交继续走现有 `client_observed` 路径；当前会话一旦存在有效服务端决策，省略 `decision_id` 必须拒绝，不能主动降级绕过核验。兼容路径不得在用户已经编辑后把最终值反向登记成“原始建议”。
+- 新预览使用 30 分钟有效的服务端 `decision_id`。保存草稿或提交成功后一次消费；保存草稿会尽力签发基于已保存事实的新 `decision_id`，供后续提交继续核验，但续签属于业务提交后的派生能力，续签失败不能把已成功保存返回为失败。相同动作请求 ID、动作和最终指纹允许安全重放，其他重放拒绝。
 - 迁移桥接期集中维护 migration-owned 表集合；所有 legacy bootstrap 只能创建集合之外的表。标准启动在 `alembic upgrade head` 前只读核对 revision 与自有表集合，发现漂移时拒绝启动，不自动猜测、stamp 或修复。
 - 所有表结构通过 Alembic 迁移，不继续在请求路径执行 DDL。

@@ -268,6 +271,8 @@
 - 已初始化系统的 Bootstrap 状态只返回脱敏信息，Bootstrap 写入、系统设置、模型连通性、缓存、审计与系统日志等敏感管理面由平台管理员权限保护；Vite 本地 Setup 桥在初始化完成后锁定重新配置入口。
 - P0 数据契约即引入最小 `tenant_id`、数据库约束、行级过滤、向量库命名空间和对象存储前缀隔离；删除传播、数据导出和私有部署加固可在商业化阶段继续完善。
 - Expense Case 用户态查询只返回流程摘要：Case 基础状态、关系类型、事件类型、操作人、发生时间及白名单业务载荷；不返回关联资源 ID、幂等键、correlation、causation、聚合标识或 Outbox 投递状态。有权查看入口单据的本人、当前审批人、财务和管理员可查看整 Case 的安全摘要，无权限与跨租户查询继续以 404 隐藏资源存在性。
+- 申请预览签发与消费同时绑定 `tenant_id`、actor、Bearer 登录会话和会话 ID；动作入口按这些服务端事实锁行校验，不能只凭随机 UUID 授权。浏览器提供的模型来源、`finalValue`、用户和角色声明均不作为可信事实。
+- 预览与学习账本中的字段值使用独立、可版本化的费用申请密钥计算 HMAC-SHA256 指纹，密钥目录和文件权限分别为 `0700`、`0600`；核验旧决策必须使用其签发版本，缺失版本直接拒绝，不静默生成替代密钥。字段指纹同时编码“字段是否存在”，避免新增或删除字段被误判为采纳；训练资格仍保持关闭，直至审批、付款或人工核验闭环完成。
 - 自动化权限按动作、金额、场景、风险和有效期授予，不使用全局“允许 Agent 自动执行”开关。
 - 收款账户变更、资金支付、制度发布、高风险驳回和敏感主数据变更执行双人或更高等级复核。

@@ -278,6 +283,7 @@
 - Qdrant/few-shot 不可用：使用 stable Prompt 和基础规则，不阻塞主流程。
 - 外部支付/ERP/税务连接器不可用：事件进入 retryable 状态，保留人工处理入口和幂等键。
 - 记忆冲突或可信度不足：只展示建议，不自动应用。
+- 预览签发超时或不可用：仍展示本地预览并允许编辑、保存和提交，内部标记为降级；此时只记录 `client_observed` 行为证据，不升级为服务端核验或训练样本。
 - Savings 证据不足：保留机会状态，不进入已实现节省。

 ## 算法与公式
@@ -515,7 +521,10 @@ docker exec -w /app -e SERVER_VENV_DIR=/tmp/x-financial-server-venv \
 - 2026-07-14（克隆库联调）：将持久开发库以只读 `pg_dump` 流式恢复到一次性 tmpfs PostgreSQL 17，原 40 张表、4 张费用单、105 名员工、248 条预算和 62 个 Agent 资产完整保留。升级到 `20260713_0002` 后首次 dry-run 识别 4 张旧单，apply 创建 4 组 Case/Link/Event，重复 apply 创建 0 条；隔离后端完成登录、`/auth/me`、旧单时间线 200 和登出，内部回填字段未出现在 API。临时数据库和隔离进程已清理，持久库复查仍为原数据签名且没有 migration-owned 表。
 - 2026-07-14（安全复核与质量验证）：交叉审查后补齐超长租户幂等键稳定哈希、Link/Case 租户一致性、孤立 Case 冲突、URL 路由参数覆盖防护和部分批次失败进度摘要；前端将快照语义明确为“纳入时状态/节点”。容器内 65 项后端定向测试、11 项前端测试、Ruff 和 Vite 生产构建通过。全库代码体积门禁仍被本次未修改的 `RiskRuleGenerationService` 807 行存量问题阻断，未混入当前功能提交。
 - 2026-07-14（AI 行为闭环首个切片）：新增 `AIDecision`、`AIDecisionFeedback` 和 `WorkflowOutcome` 三类独立事实，把申请预填建议、用户确认或显式字段纠正、草稿保存或申请提交结果通过稳定 `decision_id` 关联。只有已认证的申请预览快速入口显式传入服务端 `CurrentUserContext` 时才写账本；通用 User Agent、模板预览和单据详情编辑路径均不写入，避免伪造身份和非 AI 操作污染样本。
- 2026-07-14（字段纠正与隐私边界）：申请表编辑器在当前会话中维护首次建议值与最终显式编辑值，多次修改保留最初建议，改回原值会移除差异；日期调整同时记录 `time` 与派生 `days`。服务端最终值始终从解析后的申请 facts 重建，不信任客户端 `finalValue`；持久化层只保存字段名和 SHA-256 值指纹，不复制事由、地点、人员、金额等原始值，原始对话、Prompt、模型全文、附件和支付敏感信息也不进入学习账本。
- 2026-07-14（可信度边界）：浏览器编辑轨迹统一标记为 `verification_status=client_observed`、`trust_level=behavioral_analytics_only` 且 `training_eligible=false`，只能用于受限的产品行为统计和后续人工核验，不能直接作为模型训练、审计结论、风险自动放行或企业记忆激活依据。后续需增加服务端预览登记并返回不可伪造的 `decision_id`，再实现拒绝/忽略、审批/退回/付款结果追踪、记忆候选与下次建议变化。
+- 2026-07-14（字段纠正与隐私边界）：申请表编辑器在当前会话中维护首次建议值与最终显式编辑值，多次修改保留最初建议，改回原值会移除差异；日期调整同时记录 `time` 与派生 `days`。服务端最终值始终从解析后的申请 facts 重建，不信任客户端 `finalValue`；持久化层只保存字段名和带密钥版本的 HMAC-SHA256 值指纹，不复制事由、地点、人员、金额等原始值，原始对话、Prompt、模型全文、附件和支付敏感信息也不进入学习账本。
+- 2026-07-14（可信度边界）：浏览器编辑轨迹仍标记为 `verification_status=client_observed`、`trust_level=behavioral_analytics_only` 且 `training_eligible=false`，只能用于受限产品统计；服务端签发预览消费后提升为 `server_verified`，仅证明“服务端签发快照与最终业务事实的差异已核验”，不证明浏览器声称的模型来源，也仍不直接训练模型、自动放行风险或激活企业记忆。
 - 2026-07-14（事务与一致性）：学习三表与申请、预算及 Business Event 使用同一 Session，在最终 commit 前 `flush`，失败整体回滚；提交场景通过提交服务的事务内回调关联真实 `application_submitted` 事件。Case、Business Event 和 Decision 使用包含 `tenant_id + expense_case_id` 的复合外键，既拒绝跨租户，也拒绝同租户跨 Case 串联。前端估算使用消息级请求版本和输入指纹，只合并估算派生字段，旧响应不会覆盖后续编辑。
 - 2026-07-14（0003 迁移与验证）：新增 `20260714_0003_ai_learning_loop.py`，三张学习账本表纳入集中迁移所有权；AgentRun 与旧 ExpenseClaim 保持带索引软引用以兼容空库迁移顺序。一次性 tmpfs PostgreSQL 17 的升级、重复升级、跨租户/同租户跨 Case 外键拒绝、降级和再次升级 4 项通过；容器内学习账本、入口边界和迁移所有权定向 29 项，日期联动及异步估算关键场景 5 项通过。持久开发库只读复查仍为 `40|4|105|248|62`，7 张 migration-owned 表数量为 0。
+- 2026-07-14（服务端预览决策）：新增认证的 `POST /api/v1/reimbursements/application-previews`。服务端从原始申请文本和当前登录身份重新解析字段、重跑规则测算，再签发 30 分钟有效的 `decision_id`；浏览器返回的模型来源和建议值不被直接认证。预览绑定租户、actor、登录会话和 conversation，字段只保存独立版本密钥生成的 HMAC 指纹；字段存在性也纳入指纹，新增、删除和改值都能由服务端识别。
+- 2026-07-14（消费与续签）：快速保存/提交显式携带动作类型、`decision_id` 和稳定请求 ID。服务端锁行校验后，在 Claim、Expense Case、Business Event、AIDecision、Feedback、Outcome 的同一事务中一次消费；当前会话已有有效决策时禁止省略 ID 降级。保存草稿提交成功后再尽力签发基于服务端事实的新 `decision_id`，续签失败只返回无新 ID，不反转已成功业务动作；相同请求安全重放不会重复建单或重复写学习账本。
+- 2026-07-14（0004 迁移与验证）：新增 `20260714_0004_ai_application_preview_decisions.py`，预览决策表和正式 Decision 复合租户外键纳入集中迁移所有权。一次性隔离 PostgreSQL 17 完整 upgrade、重复 upgrade、约束/索引、downgrade 和再次 upgrade 4 项通过；容器内新增决策安全用例 9 项、旧快速保存/提交 4 项、申请学习账本 9 项、迁移与所有权 26 项通过且条件型 PostgreSQL 用例 1 项跳过，3 组前端定向测试及 Vite 生产构建通过。持久开发库只读确认 8 张 migration-owned 表数量仍为 0。
--- a/document/development/2026-07-13/feature/ai-expense-closed-loop-and-value-proof/TODO.md
+++ b/document/development/2026-07-13/feature/ai-expense-closed-loop-and-value-proof/TODO.md
@@ -39,6 +39,8 @@
 - [ ] [CONCEPT: 数据与契约] 定义 `business_events` 事件信封、事件词典、correlation ID、幂等键和版本策略。
 - [x] [CONCEPT: 业务事件与 AI 决策] 定义 `ai_decisions`、`ai_decision_feedback` 和 `workflow_outcomes` 契约。
  证据：`ai_learning.py`、`expense_application_learning.py`、`20260714_0003_ai_learning_loop.py`；三类事实分别表达 AI 建议、用户采纳/显式字段纠正和业务结果，技术执行成功、用户反馈与工作流结果不复用同一状态。
+- [x] [CONCEPT: 业务事件与 AI 决策] 定义落单前服务端预览决策的签发、授权、过期、一次消费、重放和续签契约。
+  证据：`ai_application_preview.py`、`expense_application_preview_decisions.py`、`expense_application_snapshot.py`；预览不创建空 Case，绑定租户、actor、登录会话与 conversation，30 分钟过期，保存/提交成功后一次消费；保存草稿在业务提交后尽力返回基于服务端事实的新 `decision_id`，续签失败不反转已成功动作。
 - [ ] [CONCEPT: 记忆与学习] 定义 `memory_entries`、证据链接、优先级、有效期、敏感等级、撤销和遗忘契约。
 - [ ] [CONCEPT: 自动化决策] 定义动作风险、金额阈值、置信度、证据完整度、可逆性、抽检率和企业授权策略。
 - [ ] [CONCEPT: 节省与价值] 定义 Savings Ledger 的机会、执行、实现、确认、去重和归因状态。
@@ -67,6 +69,8 @@
  证据：`20260713_0001_expense_case_business_events.py`；一次性 PostgreSQL 17 已验证表、唯一约束、复合索引、外键级联、降级和再次升级。
 - [x] [CONCEPT: 业务事件与 AI 决策] 新增 `ai_decisions`、`ai_decision_feedback`、`workflow_outcomes` 表及迁移。
  证据：`20260714_0003_ai_learning_loop.py`、`schema_ownership.py`、`migration_preflight.py`；三表具备租户幂等约束，Case/Event/Decision 使用包含费用 Case 的复合租户外键。一次性 PostgreSQL 17 完整升级、重复升级、跨租户及同租户跨 Case 外键拒绝、降级和再次升级 4 项通过，持久开发库未迁移。
+- [x] [CONCEPT: 业务事件与 AI 决策] 新增服务端预览决策表、认证签发接口和正式 Decision 关联迁移。
+  证据：`20260714_0004_ai_application_preview_decisions.py`、`expense_application_previews.py`、`reimbursements.py`；动作消费与 Claim、Case、Business Event 和学习三表同事务，跨登录会话重放被拒绝；独立版本密钥生成的 HMAC 指纹不保存低熵字段明文，并编码字段存在性。只有从未成功签发的旧调用可无 ID 降级，当前会话已有有效决策时省略 ID 会返回 409。
 - [ ] [CONCEPT: 数据与契约] 为申请、票据、草稿、提交、退回、审批、付款和归档接入统一 correlation ID。
 - [ ] [CONCEPT: 业务事件与 AI 决策] 建立事务 Outbox：申请、提交、退回、审批、支付和入账状态与事件同事务提交，消费端按事件 ID 幂等处理。
 - [x] [CONCEPT: 业务事件与 AI 决策] 完成首批草稿、提交、退回、审批、申请转报销、付款和申请归档事件旁写，具备 correlation、causation、幂等键及事务回滚。
@@ -112,7 +116,8 @@
 - [ ] [CONCEPT: 记忆激活] 实现 candidate/active/suppressed/expired/revoked 记忆状态机。
 - [ ] [CONCEPT: 记忆激活] 实现用户、部门、企业记忆优先级、冲突解释、时间衰减和最小样本要求。
 - [x] [CONCEPT: 记忆与学习] 为 AI 申请预填记录用户原样采纳、显式字段修改和草稿/提交结果证据。
-  证据：`expenseApplicationDecisionFeedback.js`、`useApplicationPreviewEditor.js`、`expense_application_learning.py`；改回原建议会清除字段差异，日期联动同步记录天数，最终值由服务端 facts 重建，客户端 `finalValue` 不作为事实源。账本只保存字段名和值指纹，证据标记为 `client_observed` 且 `training_eligible=false`，尚不直接训练模型或激活记忆。
+  证据：`expenseApplicationDecisionFeedback.js`、`useApplicationPreviewEditor.js`、`expense_application_learning.py`、`expense_application_preview_decisions.py`；改回原建议会清除字段差异，日期联动同步记录天数，最终值由服务端 facts 重建。旧预览保持 `client_observed`；服务端签发预览由版本化 HMAC 快照与最终 facts 逐字段比对，字段新增、删除或改值都标记为 `server_verified` 编辑。两者均保持 `training_eligible=false`，尚不直接训练模型或激活记忆。
+- [ ] [CONCEPT: 记忆与学习] 将小财管家/通用 Orchestrator 的申请预览切换到认证签发与消费链路；在此之前不得把其客户端 preview 升级为服务端核验证据。
 - [ ] [CONCEPT: 记忆与学习] 从字段接受/修改/拒绝、退回、审批覆盖、付款和审计结果生成记忆证据。
 - [ ] [CONCEPT: 记忆与学习] 将已确认 few-shot 扩展到报销预审和审批辅助，并按租户、场景、制度版本过滤。
 - [ ] [CONCEPT: 风险与预审] 完成 golden case、Prompt/规则版本、Canary、回归门禁和自动回滚。
@@ -180,6 +185,8 @@
 - [x] [CONCEPT: 测试方案] 在现有开发数据的只读一次性克隆上验证迁移、历史回填和真实认证时间线链路。
  证据：源库与克隆初始签名均为 40 张表、4 张费用单、105 名员工、248 条预算、62 个 Agent 资产；克隆升级后 dry-run 为 eligible=4，apply created=4，重复 apply created=0，四条事件均为 system/suppressed；登录、`/auth/me`、旧单时间线和登出均返回 200，持久库复查不变且仍无 migration-owned 表。
 - [ ] [CONCEPT: 测试方案] 补充其余前端组件、键盘操作、移动真实接口和完整浏览器关键流程验证。
+- [x] [CONCEPT: 测试方案] 验证服务端预览决策签发、字段新增/改值差异判定、跨会话拒绝、无 ID 降级防绕过、安全重放、独立密钥权限、续签失败容错、旧路径兼容、会话恢复和真实 PostgreSQL 迁移。
+  证据：容器内新增决策安全用例 9 项、旧快速保存/提交 4 项、申请学习账本 9 项、迁移与所有权 26 项通过且条件型 PostgreSQL 用例 1 项跳过；一次性 PostgreSQL 17 迁移 4 项、前端定向 3 组及 Vite 生产构建已通过。临时 PostgreSQL 已清理，持久开发库 8 张 migration-owned 表数量仍为 0。
 - [ ] [CONCEPT: 测试方案] 所有后端、集成和迁移测试在当前主应用容器内执行，单条命令最大超时 60s。
 - [ ] [CONCEPT: 指标与验收] 记录测试、lint、typecheck、构建、端到端和未覆盖风险证据。

--- a/server/alembic/versions/20260714_0004_ai_application_preview_decisions.py
+++ b/server/alembic/versions/20260714_0004_ai_application_preview_decisions.py
@@ -0,0 +1,138 @@
+"""add server-issued expense application preview decisions
+
+Revision ID: 20260714_0004
+Revises: 20260714_0003
+Create Date: 2026-07-14 15:30:00
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision: str = "20260714_0004"
+down_revision: str | None = "20260714_0003"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "ai_application_preview_decisions",
+        sa.Column("id", sa.String(length=36), nullable=False),
+        sa.Column("tenant_id", sa.String(length=64), nullable=False),
+        sa.Column("actor_id", sa.String(length=120), nullable=False),
+        sa.Column("auth_session_id", sa.String(length=120), nullable=False),
+        sa.Column("conversation_id", sa.String(length=120), nullable=False),
+        sa.Column("decision_type", sa.String(length=80), nullable=False),
+        sa.Column("decision_source", sa.String(length=20), nullable=False),
+        sa.Column("field_keys_json", sa.JSON(), nullable=False),
+        sa.Column("field_fingerprints_json", sa.JSON(), nullable=False),
+        sa.Column("snapshot_fingerprint", sa.String(length=80), nullable=False),
+        sa.Column("fingerprint_key_version", sa.String(length=40), nullable=False),
+        sa.Column("source_version_json", sa.JSON(), nullable=False),
+        sa.Column("status", sa.String(length=20), nullable=False),
+        sa.Column("issue_request_id", sa.String(length=120), nullable=False),
+        sa.Column("consumed_action", sa.String(length=30), nullable=True),
+        sa.Column("consumed_request_id", sa.String(length=120), nullable=True),
+        sa.Column("consumed_claim_id", sa.String(length=36), nullable=True),
+        sa.Column("consumed_business_event_id", sa.String(length=36), nullable=True),
+        sa.Column("consumed_final_fingerprint", sa.String(length=80), nullable=True),
+        sa.Column("expires_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("consumed_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column(
+            "created_at",
+            sa.DateTime(timezone=True),
+            server_default=sa.func.now(),
+            nullable=False,
+        ),
+        sa.CheckConstraint(
+            "status IN ('issued', 'consumed', 'expired', 'revoked')",
+            name="ck_ai_application_preview_decisions_status",
+        ),
+        sa.CheckConstraint(
+            "decision_source IN ('heuristic', 'rule', 'hybrid', 'server_draft')",
+            name="ck_ai_application_preview_decisions_source",
+        ),
+        sa.CheckConstraint(
+            "consumed_action IS NULL OR consumed_action IN ('save_draft', 'submit')",
+            name="ck_ai_application_preview_decisions_consumed_action",
+        ),
+        sa.CheckConstraint(
+            "status != 'consumed' OR ("
+            "consumed_action IS NOT NULL AND consumed_request_id IS NOT NULL AND "
+            "consumed_claim_id IS NOT NULL AND consumed_business_event_id IS NOT NULL AND "
+            "consumed_final_fingerprint IS NOT NULL AND consumed_at IS NOT NULL)",
+            name="ck_ai_application_preview_decisions_consumed_fields",
+        ),
+        sa.CheckConstraint(
+            "expires_at > created_at",
+            name="ck_ai_application_preview_decisions_expiry",
+        ),
+        sa.PrimaryKeyConstraint("id"),
+        sa.UniqueConstraint(
+            "tenant_id",
+            "id",
+            name="uq_ai_application_preview_decisions_tenant_id",
+        ),
+        sa.UniqueConstraint(
+            "tenant_id",
+            "actor_id",
+            "auth_session_id",
+            "issue_request_id",
+            name="uq_ai_application_preview_decisions_issue_request",
+        ),
+    )
+    op.create_index(
+        "ix_ai_application_preview_decisions_actor_status_expiry",
+        "ai_application_preview_decisions",
+        ["tenant_id", "actor_id", "status", "expires_at"],
+    )
+    op.create_index(
+        "ix_ai_application_preview_decisions_conversation",
+        "ai_application_preview_decisions",
+        ["tenant_id", "conversation_id", "created_at"],
+    )
+
+    op.add_column(
+        "ai_decisions",
+        sa.Column("preview_decision_id", sa.String(length=36), nullable=True),
+    )
+    op.create_foreign_key(
+        "fk_ai_decisions_tenant_preview_decision",
+        "ai_decisions",
+        "ai_application_preview_decisions",
+        ["tenant_id", "preview_decision_id"],
+        ["tenant_id", "id"],
+        ondelete="RESTRICT",
+    )
+    op.create_unique_constraint(
+        "uq_ai_decisions_tenant_preview_decision",
+        "ai_decisions",
+        ["tenant_id", "preview_decision_id"],
+    )
+
+
+def downgrade() -> None:
+    op.drop_constraint(
+        "uq_ai_decisions_tenant_preview_decision",
+        "ai_decisions",
+        type_="unique",
+    )
+    op.drop_constraint(
+        "fk_ai_decisions_tenant_preview_decision",
+        "ai_decisions",
+        type_="foreignkey",
+    )
+    op.drop_column("ai_decisions", "preview_decision_id")
+
+    op.drop_index(
+        "ix_ai_application_preview_decisions_conversation",
+        table_name="ai_application_preview_decisions",
+    )
+    op.drop_index(
+        "ix_ai_application_preview_decisions_actor_status_expiry",
+        table_name="ai_application_preview_decisions",
+    )
+    op.drop_table("ai_application_preview_decisions")
--- a/server/src/app/api/v1/endpoints/expense_application_previews.py
+++ b/server/src/app/api/v1/endpoints/expense_application_previews.py
@@ -0,0 +1,230 @@
+from __future__ import annotations
+
+import logging
+from typing import Annotated
+
+from fastapi import APIRouter, Depends, HTTPException, status
+from sqlalchemy.orm import Session
+
+from app.api.deps import CurrentUserContext, get_current_user, get_db
+from app.schemas.expense_application_preview import (
+    ExpenseApplicationPreviewDecisionCreate,
+    ExpenseApplicationPreviewDecisionRead,
+)
+from app.schemas.ontology import OntologyParseResult, OntologyPermission
+from app.schemas.reimbursement import (
+    ExpenseApplicationPreviewActionPayload,
+    ExpenseApplicationPreviewActionResponse,
+    ExpenseApplicationPreviewActionResult,
+)
+from app.schemas.user_agent import UserAgentRequest
+from app.services.expense_application_preview_decisions import (
+    ExpenseApplicationPreviewDecisionService,
+    PreviewDecisionConflictError,
+)
+from app.services.expense_application_request_context import (
+    build_trusted_expense_application_context,
+)
+from app.services.user_agent import UserAgentService
+
+router = APIRouter(prefix="/reimbursements")
+logger = logging.getLogger(__name__)
+DbSession = Annotated[Session, Depends(get_db)]
+CurrentUser = Annotated[CurrentUserContext, Depends(get_current_user)]
+
+
+@router.post(
+    "/application-previews",
+    response_model=ExpenseApplicationPreviewDecisionRead,
+    status_code=status.HTTP_201_CREATED,
+    summary="签发费用申请预览决策",
+    description=(
+        "服务端重新解析申请原文、重算规则字段并签发短期 decision_id；"
+        "不接受客户端来源声明作为可信 AI 证据。"
+    ),
+)
+def issue_expense_application_preview_decision(
+    payload: ExpenseApplicationPreviewDecisionCreate,
+    db: DbSession,
+    current_user: CurrentUser,
+) -> ExpenseApplicationPreviewDecisionRead:
+    run_id = f"application-preview-decision:{payload.request_id}"
+    context_json = build_trusted_expense_application_context(current_user)
+    request = UserAgentRequest(
+        run_id=run_id,
+        user_id=current_user.username or current_user.name,
+        message=payload.message,
+        ontology=OntologyParseResult(run_id=run_id),
+        context_json=context_json,
+        tool_payload={},
+        selected_capability_codes=[],
+        degraded=False,
+        requires_confirmation=False,
+    )
+    try:
+        facts = UserAgentService(db)._resolve_expense_application_facts(request)
+        issued = ExpenseApplicationPreviewDecisionService(db).issue(
+            facts,
+            current_user,
+            conversation_id=payload.conversation_id or "",
+            request_id=payload.request_id,
+        )
+        db.commit()
+        db.refresh(issued.decision)
+    except PreviewDecisionConflictError as error:
+        db.rollback()
+        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail=str(error)) from error
+    except ValueError as error:
+        db.rollback()
+        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail=str(error)) from error
+
+    return ExpenseApplicationPreviewDecisionRead(
+        decision_id=issued.decision.id,
+        decision_source=issued.decision.decision_source,
+        expires_at=issued.decision.expires_at,
+        application_preview={
+            "fields": issued.fields,
+            "decisionId": issued.decision.id,
+            "decisionSource": issued.decision.decision_source,
+            "decisionExpiresAt": issued.decision.expires_at.isoformat(),
+        },
+    )
+
+
+@router.post(
+    "/application-preview-action",
+    response_model=ExpenseApplicationPreviewActionResponse,
+    summary="按申请核对预览快速保存或提交申请单",
+    description=(
+        "用于 AI 工作台已完成表格核对后的轻量建单/提交流程，避免重复进入通用 Orchestrator 编排。"
+    ),
+)
+def run_application_preview_action(
+    payload: ExpenseApplicationPreviewActionPayload,
+    db: DbSession,
+    current_user: CurrentUser,
+) -> ExpenseApplicationPreviewActionResponse:
+    context_json = build_trusted_expense_application_context(
+        current_user,
+        payload.context_json,
+    )
+    if payload.action_type == "save_draft":
+        context_json["application_action"] = "save_draft"
+        context_json["application_save_mode"] = True
+    elif payload.action_type == "submit":
+        context_json.pop("application_action", None)
+        context_json.pop("application_save_mode", None)
+    run_id = f"application-preview-action:{payload.conversation_id or current_user.username}"
+    request = UserAgentRequest(
+        run_id=run_id,
+        user_id=current_user.username or current_user.name,
+        message=payload.message,
+        ontology=OntologyParseResult(
+            scenario="expense",
+            intent="operate",
+            permission=OntologyPermission(
+                level="approval_required",
+                allowed=True,
+                reason="application preview fast action",
+            ),
+            confidence=1.0,
+            run_id=run_id,
+        ),
+        context_json=context_json,
+        tool_payload={},
+        selected_capability_codes=[],
+        degraded=False,
+        requires_confirmation=False,
+    )
+    try:
+        user_agent_service = UserAgentService(db)
+        facts = user_agent_service._resolve_expense_application_facts(request)
+        resolved_step = user_agent_service._resolve_expense_application_step(request, facts)
+        resolved_action = "save_draft" if resolved_step == "draft" else "submit"
+        if payload.action_type and payload.action_type != resolved_action:
+            raise ValueError("动作类型与申请内容不一致。")
+        preview_decision_service = ExpenseApplicationPreviewDecisionService(db)
+        preview_decision = None
+        if payload.decision_id:
+            preview_decision = preview_decision_service.require_for_action(
+                payload.decision_id,
+                current_user,
+                conversation_id=payload.conversation_id or "",
+                request_id=payload.request_id or "",
+                action_type=resolved_action,
+                final_values=facts,
+            )
+        else:
+            preview_decision_service.reject_active_decision_bypass(
+                current_user,
+                conversation_id=payload.conversation_id or "",
+            )
+        consumed_preview_decision_id = preview_decision.id if preview_decision is not None else ""
+        user_agent_response = user_agent_service._build_expense_application_response(
+            request,
+            risk_flags=[],
+            learning_current_user=current_user,
+            learning_preview_decision=preview_decision,
+            learning_action_request_id=payload.request_id or "",
+        )
+    except PreviewDecisionConflictError as error:
+        db.rollback()
+        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail=str(error)) from error
+    except ValueError as error:
+        db.rollback()
+        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail=str(error)) from error
+
+    next_preview_decision = None
+    if (
+        preview_decision is not None
+        and resolved_action == "save_draft"
+        and user_agent_response.draft_payload is not None
+    ):
+        try:
+            next_preview_decision = (
+                ExpenseApplicationPreviewDecisionService(db)
+                .issue(
+                    facts,
+                    current_user,
+                    conversation_id=payload.conversation_id or "",
+                    request_id=(f"next:{consumed_preview_decision_id}:{payload.request_id or ''}")[
+                        :120
+                    ],
+                    decision_source="server_draft",
+                )
+                .decision
+            )
+            db.commit()
+            db.refresh(next_preview_decision)
+        except Exception as error:
+            # 业务动作已经在 UserAgent 内提交；续签仅是派生能力，失败不能反转成功结果。
+            db.rollback()
+            next_preview_decision = None
+            logger.warning(
+                "费用申请草稿已保存，但后续预览决策签发失败：decision_id=%s error_type=%s",
+                consumed_preview_decision_id,
+                type(error).__name__,
+            )
+
+    return ExpenseApplicationPreviewActionResponse(
+        status="succeeded",
+        conversation_id=payload.conversation_id,
+        result=ExpenseApplicationPreviewActionResult(
+            message=user_agent_response.answer,
+            answer=user_agent_response.answer,
+            suggested_actions=[
+                action.model_dump(mode="json") for action in user_agent_response.suggested_actions
+            ],
+            risk_flags=user_agent_response.risk_flags,
+            requires_confirmation=user_agent_response.requires_confirmation,
+            draft_payload=(
+                user_agent_response.draft_payload.model_dump(mode="json")
+                if user_agent_response.draft_payload is not None
+                else None
+            ),
+            decision_id=(next_preview_decision.id if next_preview_decision is not None else None),
+            decision_expires_at=(
+                next_preview_decision.expires_at if next_preview_decision is not None else None
+            ),
+        ),
+    )
--- a/server/src/app/api/v1/endpoints/reimbursements.py
+++ b/server/src/app/api/v1/endpoints/reimbursements.py
@@ -10,11 +10,7 @@ from app.api.deps import CurrentUserContext, get_current_user, get_db
 from app.api.pagination import PageNumber, PageSize, page_payload, wants_page
 from app.schemas.budget import BudgetClaimAnalysisRead
 from app.schemas.common import ErrorResponse, PaginatedResponse
-from app.schemas.ontology import OntologyParseResult, OntologyPermission
 from app.schemas.reimbursement import (
-    ExpenseApplicationPreviewActionPayload,
-    ExpenseApplicationPreviewActionResponse,
-    ExpenseApplicationPreviewActionResult,
    ExpenseClaimActionResponse,
    ExpenseClaimApprovalPayload,
    ExpenseClaimAttachmentActionResponse,
@@ -31,13 +27,11 @@ from app.schemas.reimbursement import (
    TravelReimbursementCalculatorRequest,
    TravelReimbursementCalculatorResponse,
 )
-from app.schemas.user_agent import UserAgentRequest
 from app.services.budget import BudgetService
 from app.services.document_numbering import is_application_claim_no
 from app.services.expense_claims import ExpenseClaimService
 from app.services.reimbursement import ReimbursementService
 from app.services.travel_reimbursement_calculator import TravelReimbursementCalculatorService
-from app.services.user_agent import UserAgentService

 router = APIRouter()
 DbSession = Annotated[Session, Depends(get_db)]
@@ -95,100 +89,6 @@ def calculate_travel_reimbursement(
        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail=str(error)) from error


-def _build_application_preview_action_context(
-    payload: ExpenseApplicationPreviewActionPayload,
-    current_user: CurrentUserContext,
-) -> dict[str, object]:
-    context_json = dict(payload.context_json or {})
-    context_json.setdefault("session_type", "application")
-    context_json.setdefault("entry_source", "workbench_ai_inline")
-    context_json.setdefault("document_type", "expense_application")
-    context_json.setdefault("application_stage", "expense_application")
-    # 身份与权限字段只能来自服务端会话，不能保留请求体中的同名值。
-    context_json.update(
-        {
-            "tenant_id": current_user.tenant_id,
-            "role_codes": current_user.role_codes,
-            "is_admin": current_user.is_admin,
-            "username": current_user.username,
-            "name": current_user.name,
-            "department_name": current_user.department_name,
-            "position": current_user.position,
-            "grade": current_user.grade,
-            "employee_no": current_user.employee_no,
-            "manager_name": current_user.manager_name,
-        }
-    )
-    return context_json
-
-
-@router.post(
-    "/application-preview-action",
-    response_model=ExpenseApplicationPreviewActionResponse,
-    summary="按申请核对预览快速保存或提交申请单",
-    description=(
-        "用于 AI 工作台已完成表格核对后的轻量建单/提交流程，"
-        "避免重复进入通用 Orchestrator 编排。"
-    ),
-)
-def run_application_preview_action(
-    payload: ExpenseApplicationPreviewActionPayload,
-    db: DbSession,
-    current_user: CurrentUser,
-) -> ExpenseApplicationPreviewActionResponse:
-    context_json = _build_application_preview_action_context(payload, current_user)
-    run_id = f"application-preview-action:{payload.conversation_id or current_user.username}"
-    request = UserAgentRequest(
-        run_id=run_id,
-        user_id=current_user.username or current_user.name,
-        message=payload.message,
-        ontology=OntologyParseResult(
-            scenario="expense",
-            intent="operate",
-            permission=OntologyPermission(
-                level="approval_required",
-                allowed=True,
-                reason="application preview fast action",
-            ),
-            confidence=1.0,
-            run_id=run_id,
-        ),
-        context_json=context_json,
-        tool_payload={},
-        selected_capability_codes=[],
-        degraded=False,
-        requires_confirmation=False,
-    )
-    try:
-        user_agent_response = UserAgentService(db)._build_expense_application_response(
-            request,
-            risk_flags=[],
-            learning_current_user=current_user,
-        )
-    except ValueError as error:
-        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail=str(error)) from error
-
-    return ExpenseApplicationPreviewActionResponse(
-        status="succeeded",
-        conversation_id=payload.conversation_id,
-        result=ExpenseApplicationPreviewActionResult(
-            message=user_agent_response.answer,
-            answer=user_agent_response.answer,
-            suggested_actions=[
-                action.model_dump(mode="json")
-                for action in user_agent_response.suggested_actions
-            ],
-            risk_flags=user_agent_response.risk_flags,
-            requires_confirmation=user_agent_response.requires_confirmation,
-            draft_payload=(
-                user_agent_response.draft_payload.model_dump(mode="json")
-                if user_agent_response.draft_payload is not None
-                else None
-            ),
-        ),
-    )
-
-
@router.get(
    "/claims",
    response_model=list[ExpenseClaimRead] | PaginatedResponse[ExpenseClaimRead],
--- a/server/src/app/api/v1/router.py
+++ b/server/src/app/api/v1/router.py
@@ -6,17 +6,24 @@ from app.api.v1.endpoints.agent_feedback import router as agent_feedback_router
 from app.api.v1.endpoints.agent_runs import router as agent_runs_router
 from app.api.v1.endpoints.agent_traces import router as agent_traces_router
 from app.api.v1.endpoints.analytics import router as analytics_router
-from app.api.v1.endpoints.attachment_association_jobs import router as attachment_association_jobs_router
+from app.api.v1.endpoints.attachment_association_jobs import (
+    router as attachment_association_jobs_router,
+)
 from app.api.v1.endpoints.audit_logs import router as audit_logs_router
 from app.api.v1.endpoints.auth import router as auth_router
 from app.api.v1.endpoints.bootstrap import router as bootstrap_router
 from app.api.v1.endpoints.budgets import router as budgets_router
-from app.api.v1.endpoints.employees import router as employees_router
-from app.api.v1.endpoints.expense_cases import router as expense_cases_router
 from app.api.v1.endpoints.employee_profiles import router as employee_profiles_router
+from app.api.v1.endpoints.employees import router as employees_router
+from app.api.v1.endpoints.expense_application_previews import (
+    router as expense_application_previews_router,
+)
+from app.api.v1.endpoints.expense_cases import router as expense_cases_router
 from app.api.v1.endpoints.health import router as health_router
 from app.api.v1.endpoints.knowledge import router as knowledge_router
-from app.api.v1.endpoints.linked_reimbursement_draft_jobs import router as linked_reimbursement_draft_jobs_router
+from app.api.v1.endpoints.linked_reimbursement_draft_jobs import (
+    router as linked_reimbursement_draft_jobs_router,
+)
 from app.api.v1.endpoints.notification_states import router as notification_states_router
 from app.api.v1.endpoints.ocr import router as ocr_router
 from app.api.v1.endpoints.ontology import router as ontology_router
@@ -50,6 +57,7 @@ router.include_router(orchestrator_router, tags=["orchestrator"])
 router.include_router(receipt_folder_router, tags=["receipt-folder"])
 router.include_router(employees_router, prefix="/employees", tags=["employees"])
 router.include_router(expense_cases_router, tags=["expense-cases"])
+router.include_router(expense_application_previews_router, tags=["reimbursements"])
 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"])
--- a/server/src/app/core/expense_application_fingerprint_keys.py
+++ b/server/src/app/core/expense_application_fingerprint_keys.py
@@ -0,0 +1,56 @@
+from __future__ import annotations
+
+import base64
+import binascii
+import os
+import re
+import secrets
+from pathlib import Path
+
+from app.core.config import SERVER_DIR
+
+FINGERPRINT_KEY_DIRECTORY = SERVER_DIR / ".secrets" / "expense-application-fingerprints"
+ACTIVE_FINGERPRINT_KEY_VERSION = "v1"
+KEY_BYTES = 32
+_VERSION_PATTERN = re.compile(r"^[a-zA-Z0-9._-]{1,32}$")
+
+
+def get_expense_application_fingerprint_key(
+    version: str,
+    *,
+    create: bool,
+) -> bytes:
+    normalized_version = str(version or "").strip()
+    if not _VERSION_PATTERN.fullmatch(normalized_version):
+        raise ValueError("费用申请指纹密钥版本无效。")
+    key_path = FINGERPRINT_KEY_DIRECTORY / f"{normalized_version}.key"
+    if not key_path.exists():
+        if not create:
+            raise ValueError("费用申请指纹密钥版本不可用，不能核验旧预览。")
+        _create_key_atomically(key_path)
+    if key_path.is_symlink() or not key_path.is_file():
+        raise ValueError("费用申请指纹密钥文件无效。")
+    os.chmod(key_path, 0o600)
+    encoded = key_path.read_text(encoding="utf-8").strip()
+    try:
+        key = base64.urlsafe_b64decode(encoded.encode("ascii"))
+    except (binascii.Error, ValueError, UnicodeError) as error:
+        raise ValueError("费用申请指纹密钥内容无效。") from error
+    if len(key) != KEY_BYTES:
+        raise ValueError("费用申请指纹密钥长度无效。")
+    return key
+
+
+def _create_key_atomically(key_path: Path) -> None:
+    key_path.parent.mkdir(parents=True, exist_ok=True)
+    os.chmod(key_path.parent, 0o700)
+    encoded = base64.urlsafe_b64encode(secrets.token_bytes(KEY_BYTES))
+    flags = os.O_WRONLY | os.O_CREAT | os.O_EXCL
+    try:
+        descriptor = os.open(key_path, flags, 0o600)
+    except FileExistsError:
+        return
+    with os.fdopen(descriptor, "wb") as stream:
+        stream.write(encoded)
+        stream.flush()
+        os.fsync(stream.fileno())
--- a/server/src/app/db/base.py
+++ b/server/src/app/db/base.py
@@ -9,6 +9,7 @@ from app.models.agent_asset import (
 from app.models.agent_conversation import AgentConversation, AgentConversationMessage
 from app.models.agent_feedback import AgentOperationFeedback
 from app.models.agent_run import AgentRun, AgentToolCall, AgentTraceEvent, SemanticParseLog
+from app.models.ai_application_preview import AIApplicationPreviewDecision
 from app.models.ai_learning import AIDecision, AIDecisionFeedback, WorkflowOutcome
 from app.models.approval import ApprovalRecord
 from app.models.audit_log import AuditLog
@@ -53,6 +54,7 @@ __all__ = [
    "AgentRun",
    "AgentToolCall",
    "AgentTraceEvent",
+    "AIApplicationPreviewDecision",
    "AIDecision",
    "AIDecisionFeedback",
    "ApprovalRecord",
--- a/server/src/app/db/migration_preflight.py
+++ b/server/src/app/db/migration_preflight.py
@@ -38,8 +38,20 @@ MIGRATION_OWNED_TABLES_BY_REVISION: dict[str, frozenset[str]] = {
            "workflow_outcomes",
        }
    ),
+    "20260714_0004": frozenset(
+        {
+            "expense_cases",
+            "expense_case_links",
+            "business_events",
+            "auth_sessions",
+            "ai_application_preview_decisions",
+            "ai_decisions",
+            "ai_decision_feedback",
+            "workflow_outcomes",
+        }
+    ),
 }
-if MIGRATION_OWNED_TABLES_BY_REVISION["20260714_0003"] != MIGRATION_OWNED_TABLES:
+if MIGRATION_OWNED_TABLES_BY_REVISION["20260714_0004"] != MIGRATION_OWNED_TABLES:
    raise RuntimeError("latest Alembic revision must own the centralized migration table set")


--- a/server/src/app/db/schema_ownership.py
+++ b/server/src/app/db/schema_ownership.py
@@ -7,6 +7,7 @@ from app.db.base import Base
 MIGRATION_OWNED_TABLES: frozenset[str] = frozenset(
    {
        "auth_sessions",
+        "ai_application_preview_decisions",
        "ai_decisions",
        "ai_decision_feedback",
        "expense_cases",
--- a/server/src/app/models/init.py
+++ b/server/src/app/models/init.py
@@ -7,6 +7,7 @@ from app.models.agent_asset import (
 from app.models.agent_conversation import AgentConversation, AgentConversationMessage
 from app.models.agent_feedback import AgentOperationFeedback
 from app.models.agent_run import AgentRun, AgentToolCall, AgentTraceEvent, SemanticParseLog
+from app.models.ai_application_preview import AIApplicationPreviewDecision
 from app.models.ai_learning import AIDecision, AIDecisionFeedback, WorkflowOutcome
 from app.models.approval import ApprovalRecord
 from app.models.audit_log import AuditLog
@@ -49,6 +50,7 @@ __all__ = [
    "AgentRun",
    "AgentToolCall",
    "AgentTraceEvent",
+    "AIApplicationPreviewDecision",
    "ApprovalRecord",
    "AuditLog",
    "AuthSession",
--- a/server/src/app/models/ai_application_preview.py
+++ b/server/src/app/models/ai_application_preview.py
@@ -0,0 +1,105 @@
+from __future__ import annotations
+
+import uuid
+from datetime import datetime
+from typing import Any
+
+from sqlalchemy import CheckConstraint, DateTime, Index, String, UniqueConstraint, func
+from sqlalchemy.orm import Mapped, mapped_column
+from sqlalchemy.types import JSON
+
+from app.db.base_class import Base
+
+
+class AIApplicationPreviewDecision(Base):
+    """费用申请落单前，由服务端签发并短期持有的不可变建议快照。"""
+
+    __tablename__ = "ai_application_preview_decisions"
+    __table_args__ = (
+        UniqueConstraint(
+            "tenant_id",
+            "id",
+            name="uq_ai_application_preview_decisions_tenant_id",
+        ),
+        UniqueConstraint(
+            "tenant_id",
+            "actor_id",
+            "auth_session_id",
+            "issue_request_id",
+            name="uq_ai_application_preview_decisions_issue_request",
+        ),
+        CheckConstraint(
+            "status IN ('issued', 'consumed', 'expired', 'revoked')",
+            name="ck_ai_application_preview_decisions_status",
+        ),
+        CheckConstraint(
+            "decision_source IN ('heuristic', 'rule', 'hybrid', 'server_draft')",
+            name="ck_ai_application_preview_decisions_source",
+        ),
+        CheckConstraint(
+            "consumed_action IS NULL OR consumed_action IN ('save_draft', 'submit')",
+            name="ck_ai_application_preview_decisions_consumed_action",
+        ),
+        CheckConstraint(
+            "status != 'consumed' OR ("
+            "consumed_action IS NOT NULL AND consumed_request_id IS NOT NULL AND "
+            "consumed_claim_id IS NOT NULL AND consumed_business_event_id IS NOT NULL AND "
+            "consumed_final_fingerprint IS NOT NULL AND consumed_at IS NOT NULL)",
+            name="ck_ai_application_preview_decisions_consumed_fields",
+        ),
+        CheckConstraint(
+            "expires_at > created_at",
+            name="ck_ai_application_preview_decisions_expiry",
+        ),
+        Index(
+            "ix_ai_application_preview_decisions_actor_status_expiry",
+            "tenant_id",
+            "actor_id",
+            "status",
+            "expires_at",
+        ),
+        Index(
+            "ix_ai_application_preview_decisions_conversation",
+            "tenant_id",
+            "conversation_id",
+            "created_at",
+        ),
+    )
+
+    id: Mapped[str] = mapped_column(
+        String(36),
+        primary_key=True,
+        default=lambda: str(uuid.uuid4()),
+    )
+    tenant_id: Mapped[str] = mapped_column(String(64), nullable=False)
+    actor_id: Mapped[str] = mapped_column(String(120), nullable=False)
+    auth_session_id: Mapped[str] = mapped_column(String(120), nullable=False)
+    conversation_id: Mapped[str] = mapped_column(String(120), nullable=False, default="")
+    decision_type: Mapped[str] = mapped_column(String(80), nullable=False)
+    decision_source: Mapped[str] = mapped_column(String(20), nullable=False)
+    field_keys_json: Mapped[list[str]] = mapped_column(JSON, nullable=False, default=list)
+    field_fingerprints_json: Mapped[dict[str, str]] = mapped_column(
+        JSON,
+        nullable=False,
+        default=dict,
+    )
+    snapshot_fingerprint: Mapped[str] = mapped_column(String(80), nullable=False)
+    fingerprint_key_version: Mapped[str] = mapped_column(String(40), nullable=False)
+    source_version_json: Mapped[dict[str, Any]] = mapped_column(
+        JSON,
+        nullable=False,
+        default=dict,
+    )
+    status: Mapped[str] = mapped_column(String(20), nullable=False, default="issued")
+    issue_request_id: Mapped[str] = mapped_column(String(120), nullable=False)
+    consumed_action: Mapped[str | None] = mapped_column(String(30), nullable=True)
+    consumed_request_id: Mapped[str | None] = mapped_column(String(120), nullable=True)
+    consumed_claim_id: Mapped[str | None] = mapped_column(String(36), nullable=True)
+    consumed_business_event_id: Mapped[str | None] = mapped_column(String(36), nullable=True)
+    consumed_final_fingerprint: Mapped[str | None] = mapped_column(String(80), nullable=True)
+    expires_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), nullable=False)
+    consumed_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True),
+        server_default=func.now(),
+    )
--- a/server/src/app/models/ai_learning.py
+++ b/server/src/app/models/ai_learning.py
@@ -44,6 +44,20 @@ class AIDecision(Base):
            "idempotency_key",
            name="uq_ai_decisions_tenant_idempotency",
        ),
+        UniqueConstraint(
+            "tenant_id",
+            "preview_decision_id",
+            name="uq_ai_decisions_tenant_preview_decision",
+        ),
+        ForeignKeyConstraint(
+            ["tenant_id", "preview_decision_id"],
+            [
+                "ai_application_preview_decisions.tenant_id",
+                "ai_application_preview_decisions.id",
+            ],
+            ondelete="RESTRICT",
+            name="fk_ai_decisions_tenant_preview_decision",
+        ),
        ForeignKeyConstraint(
            ["tenant_id", "expense_case_id"],
            ["expense_cases.tenant_id", "expense_cases.id"],
@@ -90,6 +104,7 @@ class AIDecision(Base):
    expense_case_id: Mapped[str] = mapped_column(String(36), nullable=False)
    business_event_id: Mapped[str | None] = mapped_column(String(36), nullable=True)
    expense_claim_id: Mapped[str] = mapped_column(String(36), nullable=False, index=True)
+    preview_decision_id: Mapped[str | None] = mapped_column(String(36), nullable=True)
    agent_run_id: Mapped[str | None] = mapped_column(String(120), nullable=True)
    correlation_id: Mapped[str] = mapped_column(String(64), nullable=False)
    subject_type: Mapped[str] = mapped_column(String(50), nullable=False)
--- a/server/src/app/schemas/expense_application_preview.py
+++ b/server/src/app/schemas/expense_application_preview.py
@@ -0,0 +1,19 @@
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Any
+
+from pydantic import BaseModel, Field
+
+
+class ExpenseApplicationPreviewDecisionCreate(BaseModel):
+    message: str = Field(min_length=1, max_length=4000)
+    conversation_id: str | None = Field(default=None, max_length=120)
+    request_id: str = Field(min_length=1, max_length=120)
+
+
+class ExpenseApplicationPreviewDecisionRead(BaseModel):
+    decision_id: str
+    decision_source: str
+    expires_at: datetime
+    application_preview: dict[str, Any] = Field(default_factory=dict)
--- a/server/src/app/schemas/reimbursement.py
+++ b/server/src/app/schemas/reimbursement.py
@@ -191,6 +191,9 @@ class ExpenseApplicationPreviewActionPayload(BaseModel):
    source: str = Field(default="user_message", max_length=80)
    user_id: str | None = Field(default=None, max_length=120)
    conversation_id: str | None = Field(default=None, max_length=120)
+    action_type: str | None = Field(default=None, pattern="^(save_draft|submit)$")
+    decision_id: str | None = Field(default=None, max_length=36)
+    request_id: str | None = Field(default=None, max_length=120)
    message: str = Field(min_length=1, max_length=4000)
    context_json: dict[str, Any] = Field(default_factory=dict)

@@ -202,6 +205,8 @@ class ExpenseApplicationPreviewActionResult(BaseModel):
    risk_flags: list[str] = Field(default_factory=list)
    requires_confirmation: bool = False
    draft_payload: dict[str, Any] | None = None
+    decision_id: str | None = None
+    decision_expires_at: datetime | None = None


 class ExpenseApplicationPreviewActionResponse(BaseModel):
--- a/server/src/app/services/expense_application_learning.py
+++ b/server/src/app/services/expense_application_learning.py
@@ -1,7 +1,5 @@
 from __future__ import annotations

-import hashlib
-import json
 import uuid
 from dataclasses import dataclass
 from decimal import Decimal
@@ -11,10 +9,22 @@ from sqlalchemy import select
 from sqlalchemy.orm import Session

 from app.api.deps import CurrentUserContext
+from app.models.ai_application_preview import AIApplicationPreviewDecision
 from app.models.ai_learning import AIDecision, AIDecisionFeedback, WorkflowOutcome
 from app.models.expense_case import BusinessEvent
 from app.models.financial_record import ExpenseClaim
 from app.schemas.user_agent import UserAgentRequest
+from app.services.expense_application_preview_decisions import (
+    ExpenseApplicationPreviewDecisionService,
+)
+from app.services.expense_application_snapshot import (
+    PREVIEW_FIELD_ALIASES,
+    content_fingerprint,
+    field_value_fingerprint,
+    hmac_fingerprint,
+    safe_fact_snapshot,
+    snapshot_reference,
+)
 from app.services.expense_cases import ExpenseCaseService


@@ -28,50 +38,6 @@ class ExpenseApplicationLearningRecords:
 class ExpenseApplicationLearningService:
    """把 AI 申请预填、显式用户编辑和工作流结果原子写入学习账本。"""

-    _FACT_KEYS = (
-        "application_type",
-        "time",
-        "location",
-        "reason",
-        "days",
-        "transport_mode",
-        "amount",
-        "grade",
-        "department",
-        "position",
-        "manager_name",
-        "lodging_daily_cap",
-        "subsidy_daily_cap",
-        "transport_policy",
-        "policy_estimate",
-        "matched_city",
-        "rule_name",
-        "rule_version",
-        "hotel_amount",
-        "allowance_amount",
-        "transport_estimated_amount",
-        "transport_estimate_source",
-        "transport_estimate_confidence",
-        "policy_total_amount",
-    )
-    _FIELD_ALIASES = {
-        "applicationType": "application_type",
-        "application_type": "application_type",
-        "time": "time",
-        "time_return": "time",
-        "location": "location",
-        "reason": "reason",
-        "days": "days",
-        "transportMode": "transport_mode",
-        "transport_mode": "transport_mode",
-        "amount": "amount",
-        "grade": "grade",
-        "department": "department",
-        "position": "position",
-        "managerName": "manager_name",
-        "manager_name": "manager_name",
-    }
-
    def __init__(self, db: Session) -> None:
        self.db = db

@@ -84,10 +50,14 @@ class ExpenseApplicationLearningService:
        *,
        action_type: str,
        business_event: BusinessEvent,
+        preview_decision: AIApplicationPreviewDecision | None = None,
+        action_request_id: str = "",
    ) -> ExpenseApplicationLearningRecords | None:
        normalized_action = self._normalize_action_type(action_type)
        preview = self._application_preview(payload)
        if not self._is_learning_eligible_preview(preview):
+            if preview_decision is not None:
+                raise ValueError("服务端预览决策缺少可核验的申请字段。")
            return None
        tenant_id = ExpenseCaseService.normalize_tenant_id(current_user.tenant_id)
        expense_case = ExpenseCaseService(self.db).ensure_case_for_claim(
@@ -101,20 +71,48 @@ class ExpenseApplicationLearningService:
            raise PermissionError("AI 学习记录不能关联其他租户或费用事件的业务事件。")

        final_values = self._safe_fact_snapshot(facts)
-        changed_fields = self._safe_changed_fields(payload, final_values)
-        suggested_values = dict(final_values)
-        for item in changed_fields:
-            suggested_values[item["field_key"]] = item["suggested_value"]
+        if preview_decision is not None:
+            final_reference = self._snapshot_reference(
+                final_values,
+                key_version=preview_decision.fingerprint_key_version,
+                allow_key_creation=False,
+            )
+            changed_field_references = self._verified_changed_field_references(
+                preview_decision,
+                final_values,
+            )
+            suggestion_reference = {
+                "field_keys": list(preview_decision.field_keys_json or []),
+                "value_fingerprint": preview_decision.snapshot_fingerprint,
+                "fingerprint_key_version": preview_decision.fingerprint_key_version,
+            }
+            evidence_source = "server_preview_decision"
+            verification_status = "server_verified"
+        else:
+            final_reference = self._snapshot_reference(final_values)
+            changed_fields = self._safe_changed_fields(payload, final_values)
+            suggested_values = dict(final_values)
+            for item in changed_fields:
+                suggested_values[item["field_key"]] = item["suggested_value"]
+            changed_field_references = self._changed_field_references(changed_fields)
+            suggestion_reference = self._snapshot_reference(suggested_values)
+            evidence_source = (
+                "client_explicit_field_edit"
+                if changed_field_references
+                else "client_action_confirmation"
+            )
+            verification_status = "client_observed"

-        feedback_type = "edited" if changed_fields else "accepted"
+        feedback_type = "edited" if changed_field_references else "accepted"
        correlation_id = ExpenseCaseService.normalize_correlation_id(payload.run_id)
        fingerprint_payload = {
            "tenant_id": tenant_id,
            "expense_claim_id": claim.id,
            "action_type": normalized_action,
-            "suggested_values": suggested_values,
-            "final_values": final_values,
-            "changed_fields": changed_fields,
+            "preview_decision_id": preview_decision.id if preview_decision is not None else None,
+            "suggestion_reference": suggestion_reference,
+            "final_reference": final_reference,
+            "changed_fields": changed_field_references,
        }
        content_fingerprint = self._fingerprint(fingerprint_payload)
        idempotency_key = self._idempotency_key(
@@ -133,24 +131,38 @@ class ExpenseApplicationLearningService:
            expense_case_id=expense_case.id,
            business_event_id=business_event.id,
            expense_claim_id=claim.id,
+            preview_decision_id=(preview_decision.id if preview_decision is not None else None),
            agent_run_id=str(payload.run_id or "").strip() or None,
            correlation_id=correlation_id,
            subject_type="expense_claim",
            subject_id=claim.id,
            decision_type="expense_application_prefill",
-            decision_source=self._decision_source(preview, facts),
+            decision_source=(
+                preview_decision.decision_source
+                if preview_decision is not None
+                else self._decision_source(preview, facts)
+            ),
            status="executed",
            automation_mode="prefill",
            confidence=self._confidence(payload),
-            suggestion_json=self._snapshot_reference(suggested_values),
+            suggestion_json=suggestion_reference,
            evidence_json={
-                "evidence_source": "client_preview_edit_trace",
-                "trust_level": "behavioral_analytics_only",
+                "evidence_source": (
+                    "server_preview_decision"
+                    if preview_decision is not None
+                    else "client_preview_edit_trace"
+                ),
+                "trust_level": (
+                    "server_snapshot_verified"
+                    if preview_decision is not None
+                    else "behavioral_analytics_only"
+                ),
                "model_refined": bool(preview.get("modelRefined")),
                "model_review_status": self._safe_text(preview.get("modelReviewStatus"), 40),
            },
            version_json={
                "schema_version": 1,
+                "fingerprint_key_version": suggestion_reference.get("fingerprint_key_version"),
                "parse_strategy": self._safe_text(preview.get("parseStrategy"), 60),
                "rule_name": self._safe_text(facts.get("rule_name"), 120),
                "rule_version": self._safe_text(facts.get("rule_version"), 80),
@@ -170,28 +182,22 @@ class ExpenseApplicationLearningService:
            action_type=normalized_action,
            actor_id=str(current_user.username or "anonymous").strip() or "anonymous",
            actor_type="user",
-            evidence_source=(
-                "client_explicit_field_edit"
-                if changed_fields
-                else "client_action_confirmation"
-            ),
-            verification_status="client_observed",
+            evidence_source=evidence_source,
+            verification_status=verification_status,
            training_eligible=False,
-            final_value_json=self._snapshot_reference(final_values),
-            changed_fields_json=self._changed_field_references(changed_fields),
+            final_value_json=final_reference,
+            changed_fields_json=changed_field_references,
            idempotency_key=f"feedback:{idempotency_key}"[:120],
            content_fingerprint=self._fingerprint(
                {
                    "decision_id": decision_id,
                    "feedback_type": feedback_type,
-                    "final_values": final_values,
-                    "changed_fields": changed_fields,
+                    "final_reference": final_reference,
+                    "changed_fields": changed_field_references,
                }
            ),
        )
-        outcome_type = (
-            "application_submitted" if normalized_action == "submit" else "draft_saved"
-        )
+        outcome_type = "application_submitted" if normalized_action == "submit" else "draft_saved"
        outcome = WorkflowOutcome(
            id=self._stable_id("outcome", tenant_id, idempotency_key),
            tenant_id=tenant_id,
@@ -206,9 +212,7 @@ class ExpenseApplicationLearningService:
            actor_type="user",
            result_json={
                "claim_status": "submitted" if normalized_action == "submit" else "draft",
-                "approval_stage": (
-                    "直属领导审批" if normalized_action == "submit" else "待提交"
-                ),
+                "approval_stage": ("直属领导审批" if normalized_action == "submit" else "待提交"),
                "feedback_type": feedback_type,
            },
            idempotency_key=f"outcome:{idempotency_key}"[:120],
@@ -222,6 +226,15 @@ class ExpenseApplicationLearningService:
        )
        self.db.add_all([decision, feedback, outcome])
        self.db.flush()
+        if preview_decision is not None:
+            ExpenseApplicationPreviewDecisionService(self.db).consume(
+                preview_decision,
+                request_id=action_request_id,
+                action_type=normalized_action,
+                final_values=final_values,
+                claim=claim,
+                business_event=business_event,
+            )
        return ExpenseApplicationLearningRecords(decision, feedback, outcome)

    def _find_existing(
@@ -255,11 +268,7 @@ class ExpenseApplicationLearningService:

    @classmethod
    def _safe_fact_snapshot(cls, facts: dict[str, str]) -> dict[str, str]:
-        return {
-            key: cls._safe_text(facts.get(key), 500)
-            for key in cls._FACT_KEYS
-            if cls._safe_text(facts.get(key), 500)
-        }
+        return safe_fact_snapshot(facts)

    @classmethod
    def _safe_changed_fields(
@@ -279,7 +288,7 @@ class ExpenseApplicationLearningService:
            if not isinstance(raw_item, dict):
                continue
            alias = cls._safe_text(raw_item.get("fieldKey"), 60)
-            field_key = cls._FIELD_ALIASES.get(alias)
+            field_key = PREVIEW_FIELD_ALIASES.get(alias)
            if not field_key or field_key not in final_values:
                continue
            suggested_value = cls._safe_text(raw_item.get("suggestedValue"), 500)
@@ -309,11 +318,55 @@ class ExpenseApplicationLearningService:
        return str(preview.get("modelReviewStatus") or "").strip().lower() != "template"

    @classmethod
-    def _snapshot_reference(cls, values: dict[str, str]) -> dict[str, Any]:
-        return {
-            "field_keys": sorted(values),
-            "value_fingerprint": cls._fingerprint(values),
-        }
+    def _snapshot_reference(
+        cls,
+        values: dict[str, str],
+        *,
+        key_version: str | None = None,
+        allow_key_creation: bool = True,
+    ) -> dict[str, Any]:
+        if key_version is None:
+            return snapshot_reference(values)
+        return snapshot_reference(
+            values,
+            key_version=key_version,
+            allow_key_creation=allow_key_creation,
+        )
+
+    @staticmethod
+    def _verified_changed_field_references(
+        decision: AIApplicationPreviewDecision,
+        final_values: dict[str, str],
+    ) -> list[dict[str, str]]:
+        fingerprints = decision.field_fingerprints_json or {}
+        changes: list[dict[str, str]] = []
+        suggested_field_keys = set(decision.field_keys_json or [])
+        final_field_keys = set(final_values)
+        for field_key in sorted(suggested_field_keys | final_field_keys):
+            suggested_fingerprint = str(fingerprints.get(field_key) or "").strip()
+            if not suggested_fingerprint:
+                suggested_fingerprint = field_value_fingerprint(
+                    "",
+                    present=False,
+                    key_version=decision.fingerprint_key_version,
+                    allow_key_creation=False,
+                )
+            final_fingerprint = field_value_fingerprint(
+                final_values.get(field_key, ""),
+                present=field_key in final_values,
+                key_version=decision.fingerprint_key_version,
+                allow_key_creation=False,
+            )
+            if suggested_fingerprint == final_fingerprint:
+                continue
+            changes.append(
+                {
+                    "field_key": field_key,
+                    "suggested_value_fingerprint": suggested_fingerprint,
+                    "final_value_fingerprint": final_fingerprint,
+                }
+            )
+        return changes

    @classmethod
    def _changed_field_references(
@@ -323,8 +376,8 @@ class ExpenseApplicationLearningService:
        return [
            {
                "field_key": item["field_key"],
-                "suggested_value_fingerprint": cls._fingerprint(item["suggested_value"]),
-                "final_value_fingerprint": cls._fingerprint(item["final_value"]),
+                "suggested_value_fingerprint": hmac_fingerprint(item["suggested_value"]),
+                "final_value_fingerprint": hmac_fingerprint(item["final_value"]),
            }
            for item in changes
        ]
@@ -363,14 +416,7 @@ class ExpenseApplicationLearningService:

    @staticmethod
    def _fingerprint(payload: object) -> str:
-        serialized = json.dumps(
-            payload,
-            ensure_ascii=False,
-            sort_keys=True,
-            separators=(",", ":"),
-            default=str,
-        )
-        return f"sha256:{hashlib.sha256(serialized.encode('utf-8')).hexdigest()}"
+        return content_fingerprint(payload)

    @staticmethod
    def _idempotency_key(action_type: str, claim_id: str, fingerprint: str) -> str:
--- a/server/src/app/services/expense_application_preview_decisions.py
+++ b/server/src/app/services/expense_application_preview_decisions.py
@@ -0,0 +1,276 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import UTC, datetime, timedelta
+from typing import Any
+
+from sqlalchemy import select
+from sqlalchemy.orm import Session
+
+from app.api.deps import CurrentUserContext
+from app.models.ai_application_preview import AIApplicationPreviewDecision
+from app.models.expense_case import BusinessEvent
+from app.models.financial_record import ExpenseClaim
+from app.services.expense_application_snapshot import (
+    FINGERPRINT_KEY_VERSION,
+    field_fingerprints,
+    hmac_fingerprint,
+    preview_response_fields,
+    safe_fact_snapshot,
+)
+from app.services.expense_cases import ExpenseCaseService
+
+PREVIEW_DECISION_TTL = timedelta(minutes=30)
+
+
+class PreviewDecisionConflictError(ValueError):
+    """预览决策已过期、已消费或与当前动作不一致。"""
+
+
+@dataclass(frozen=True)
+class IssuedPreviewDecision:
+    decision: AIApplicationPreviewDecision
+    fields: dict[str, str]
+
+
+class ExpenseApplicationPreviewDecisionService:
+    """签发、授权和消费费用申请预览决策，不保存明文建议值。"""
+
+    def __init__(self, db: Session) -> None:
+        self.db = db
+
+    def issue(
+        self,
+        facts: dict[str, Any],
+        current_user: CurrentUserContext,
+        *,
+        conversation_id: str,
+        request_id: str,
+        decision_source: str | None = None,
+    ) -> IssuedPreviewDecision:
+        tenant_id = ExpenseCaseService.normalize_tenant_id(current_user.tenant_id)
+        actor_id = self._actor_id(current_user)
+        auth_session_id = self._auth_session_id(current_user)
+        normalized_request_id = self._required_text(request_id, "预览签发请求 ID", 120)
+        values = safe_fact_snapshot(facts)
+        if not values.get("time") and not values.get("location") and not values.get("reason"):
+            raise ValueError("未识别到可签发的费用申请字段。")
+
+        existing = self.db.scalar(
+            select(AIApplicationPreviewDecision).where(
+                AIApplicationPreviewDecision.tenant_id == tenant_id,
+                AIApplicationPreviewDecision.actor_id == actor_id,
+                AIApplicationPreviewDecision.auth_session_id == auth_session_id,
+                AIApplicationPreviewDecision.issue_request_id == normalized_request_id,
+            )
+        )
+        if existing is not None:
+            snapshot_fingerprint = hmac_fingerprint(
+                values,
+                key_version=existing.fingerprint_key_version,
+                allow_key_creation=False,
+            )
+            if existing.snapshot_fingerprint != snapshot_fingerprint:
+                raise PreviewDecisionConflictError("同一预览签发请求不能对应不同内容。")
+            return IssuedPreviewDecision(existing, preview_response_fields(facts))
+
+        now = datetime.now(UTC)
+        source = decision_source or self._decision_source(facts)
+        snapshot_fingerprint = hmac_fingerprint(
+            values,
+            key_version=FINGERPRINT_KEY_VERSION,
+        )
+        decision = AIApplicationPreviewDecision(
+            tenant_id=tenant_id,
+            actor_id=actor_id,
+            auth_session_id=auth_session_id,
+            conversation_id=str(conversation_id or "").strip()[:120],
+            decision_type="expense_application_prefill",
+            decision_source=source,
+            field_keys_json=sorted(values),
+            field_fingerprints_json=field_fingerprints(
+                values,
+                key_version=FINGERPRINT_KEY_VERSION,
+            ),
+            snapshot_fingerprint=snapshot_fingerprint,
+            fingerprint_key_version=FINGERPRINT_KEY_VERSION,
+            source_version_json={
+                "schema_version": 1,
+                "rule_name_fingerprint": hmac_fingerprint(
+                    values.get("rule_name", ""),
+                    key_version=FINGERPRINT_KEY_VERSION,
+                ),
+                "rule_version_fingerprint": hmac_fingerprint(
+                    values.get("rule_version", ""),
+                    key_version=FINGERPRINT_KEY_VERSION,
+                ),
+            },
+            status="issued",
+            issue_request_id=normalized_request_id,
+            expires_at=now + PREVIEW_DECISION_TTL,
+        )
+        self.db.add(decision)
+        self.db.flush()
+        return IssuedPreviewDecision(decision, preview_response_fields(facts))
+
+    def require_for_action(
+        self,
+        decision_id: str,
+        current_user: CurrentUserContext,
+        *,
+        conversation_id: str,
+        request_id: str,
+        action_type: str,
+        final_values: dict[str, Any],
+    ) -> AIApplicationPreviewDecision:
+        tenant_id = ExpenseCaseService.normalize_tenant_id(current_user.tenant_id)
+        actor_id = self._actor_id(current_user)
+        auth_session_id = self._auth_session_id(current_user)
+        normalized_id = self._required_text(decision_id, "预览决策 ID", 36)
+        normalized_request_id = self._required_text(request_id, "动作请求 ID", 120)
+        normalized_action = self._normalize_action(action_type)
+        decision = self.db.scalar(
+            select(AIApplicationPreviewDecision)
+            .where(
+                AIApplicationPreviewDecision.id == normalized_id,
+                AIApplicationPreviewDecision.tenant_id == tenant_id,
+                AIApplicationPreviewDecision.actor_id == actor_id,
+                AIApplicationPreviewDecision.auth_session_id == auth_session_id,
+            )
+            .with_for_update()
+        )
+        if decision is None:
+            raise ValueError("预览决策不存在或不属于当前登录会话。")
+        if decision.conversation_id != str(conversation_id or "").strip()[:120]:
+            raise ValueError("预览决策不属于当前会话。")
+
+        final_fingerprint = self._final_fingerprint(decision, final_values)
+        if decision.status == "consumed":
+            if (
+                decision.consumed_action == normalized_action
+                and decision.consumed_request_id == normalized_request_id
+                and decision.consumed_final_fingerprint == final_fingerprint
+            ):
+                return decision
+            raise PreviewDecisionConflictError("该预览决策已被其他动作消费，请重新生成预览。")
+        if decision.status != "issued":
+            raise PreviewDecisionConflictError("该预览决策当前不可使用，请重新生成预览。")
+        if self._as_utc(decision.expires_at) <= datetime.now(UTC):
+            raise PreviewDecisionConflictError("该预览决策已过期，请重新生成预览。")
+        return decision
+
+    def reject_active_decision_bypass(
+        self,
+        current_user: CurrentUserContext,
+        *,
+        conversation_id: str,
+    ) -> None:
+        """同一登录会话已有有效签发时，不允许通过省略 decision_id 降级。"""
+        tenant_id = ExpenseCaseService.normalize_tenant_id(current_user.tenant_id)
+        actor_id = self._actor_id(current_user)
+        auth_session_id = str(current_user.auth_session_id or "").strip()[:120]
+        # 旧调用没有会话标识，也不具备签发 decision_id 的前提，继续走非可信兼容路径。
+        if not auth_session_id:
+            return
+        normalized_conversation_id = str(conversation_id or "").strip()[:120]
+        active_decision = self.db.scalar(
+            select(AIApplicationPreviewDecision)
+            .where(
+                AIApplicationPreviewDecision.tenant_id == tenant_id,
+                AIApplicationPreviewDecision.actor_id == actor_id,
+                AIApplicationPreviewDecision.auth_session_id == auth_session_id,
+                AIApplicationPreviewDecision.conversation_id == normalized_conversation_id,
+                AIApplicationPreviewDecision.status == "issued",
+                AIApplicationPreviewDecision.expires_at > datetime.now(UTC),
+            )
+            .order_by(AIApplicationPreviewDecision.created_at.desc())
+            .with_for_update()
+        )
+        if active_decision is not None:
+            raise PreviewDecisionConflictError(
+                "当前会话已有服务端签发预览，动作必须携带 decision_id。"
+            )
+
+    def consume(
+        self,
+        decision: AIApplicationPreviewDecision,
+        *,
+        request_id: str,
+        action_type: str,
+        final_values: dict[str, Any],
+        claim: ExpenseClaim,
+        business_event: BusinessEvent,
+    ) -> None:
+        normalized_request_id = self._required_text(request_id, "动作请求 ID", 120)
+        normalized_action = self._normalize_action(action_type)
+        final_fingerprint = self._final_fingerprint(decision, final_values)
+        if decision.status == "consumed":
+            if (
+                decision.consumed_action == normalized_action
+                and decision.consumed_request_id == normalized_request_id
+                and decision.consumed_claim_id == claim.id
+                and decision.consumed_business_event_id == business_event.id
+                and decision.consumed_final_fingerprint == final_fingerprint
+            ):
+                return
+            raise PreviewDecisionConflictError("该预览决策已被消费。")
+        if decision.status != "issued":
+            raise PreviewDecisionConflictError("该预览决策当前不可消费。")
+        decision.status = "consumed"
+        decision.consumed_action = normalized_action
+        decision.consumed_request_id = normalized_request_id
+        decision.consumed_claim_id = claim.id
+        decision.consumed_business_event_id = business_event.id
+        decision.consumed_final_fingerprint = final_fingerprint
+        decision.consumed_at = datetime.now(UTC)
+        self.db.flush()
+
+    @staticmethod
+    def _final_fingerprint(
+        decision: AIApplicationPreviewDecision,
+        final_values: dict[str, Any],
+    ) -> str:
+        return hmac_fingerprint(
+            safe_fact_snapshot(final_values),
+            key_version=decision.fingerprint_key_version,
+            allow_key_creation=False,
+        )
+
+    @staticmethod
+    def _decision_source(facts: dict[str, Any]) -> str:
+        has_rule = bool(str(facts.get("rule_version") or "").strip())
+        has_transport_estimate = bool(str(facts.get("transport_estimate_source") or "").strip())
+        if has_rule and has_transport_estimate:
+            return "hybrid"
+        if has_rule:
+            return "rule"
+        return "heuristic"
+
+    @staticmethod
+    def _actor_id(current_user: CurrentUserContext) -> str:
+        return str(current_user.employee_id or current_user.username or "").strip()[:120]
+
+    @staticmethod
+    def _auth_session_id(current_user: CurrentUserContext) -> str:
+        session_id = str(current_user.auth_session_id or "").strip()[:120]
+        if not session_id:
+            raise ValueError("当前登录会话缺少可验证的会话标识，请重新登录。")
+        return session_id
+
+    @staticmethod
+    def _required_text(value: object, label: str, limit: int) -> str:
+        normalized = str(value or "").strip()[:limit]
+        if not normalized:
+            raise ValueError(f"{label}不能为空。")
+        return normalized
+
+    @staticmethod
+    def _normalize_action(action_type: str) -> str:
+        normalized = str(action_type or "").strip().lower()
+        if normalized not in {"save_draft", "submit"}:
+            raise ValueError("不支持的预览决策动作。")
+        return normalized
+
+    @staticmethod
+    def _as_utc(value: datetime) -> datetime:
+        return value if value.tzinfo is not None else value.replace(tzinfo=UTC)
--- a/server/src/app/services/expense_application_request_context.py
+++ b/server/src/app/services/expense_application_request_context.py
@@ -0,0 +1,32 @@
+from __future__ import annotations
+
+from typing import Any
+
+from app.api.deps import CurrentUserContext
+
+
+def build_trusted_expense_application_context(
+    current_user: CurrentUserContext,
+    base_context: dict[str, Any] | None = None,
+) -> dict[str, Any]:
+    context_json = dict(base_context or {})
+    context_json.setdefault("session_type", "application")
+    context_json.setdefault("entry_source", "workbench_ai_inline")
+    context_json.setdefault("document_type", "expense_application")
+    context_json.setdefault("application_stage", "expense_application")
+    # 身份与权限字段必须覆盖请求体中的同名值。
+    context_json.update(
+        {
+            "tenant_id": current_user.tenant_id,
+            "role_codes": current_user.role_codes,
+            "is_admin": current_user.is_admin,
+            "username": current_user.username,
+            "name": current_user.name,
+            "department_name": current_user.department_name,
+            "position": current_user.position,
+            "grade": current_user.grade,
+            "employee_no": current_user.employee_no,
+            "manager_name": current_user.manager_name,
+        }
+    )
+    return context_json
--- a/server/src/app/services/expense_application_snapshot.py
+++ b/server/src/app/services/expense_application_snapshot.py
@@ -0,0 +1,195 @@
+from __future__ import annotations
+
+import hashlib
+import hmac
+import json
+from typing import Any
+
+from app.core.expense_application_fingerprint_keys import (
+    ACTIVE_FINGERPRINT_KEY_VERSION,
+    get_expense_application_fingerprint_key,
+)
+
+FINGERPRINT_KEY_VERSION = ACTIVE_FINGERPRINT_KEY_VERSION
+
+FACT_KEYS = (
+    "application_type",
+    "time",
+    "location",
+    "reason",
+    "days",
+    "transport_mode",
+    "amount",
+    "grade",
+    "department",
+    "position",
+    "manager_name",
+    "lodging_daily_cap",
+    "subsidy_daily_cap",
+    "transport_policy",
+    "policy_estimate",
+    "matched_city",
+    "rule_name",
+    "rule_version",
+    "hotel_amount",
+    "allowance_amount",
+    "transport_estimated_amount",
+    "transport_estimate_source",
+    "transport_estimate_confidence",
+    "policy_total_amount",
+)
+
+PREVIEW_FIELD_ALIASES = {
+    "applicationType": "application_type",
+    "application_type": "application_type",
+    "time": "time",
+    "time_return": "time",
+    "location": "location",
+    "reason": "reason",
+    "days": "days",
+    "transportMode": "transport_mode",
+    "transport_mode": "transport_mode",
+    "amount": "amount",
+    "grade": "grade",
+    "department": "department",
+    "position": "position",
+    "managerName": "manager_name",
+    "manager_name": "manager_name",
+}
+
+PREVIEW_RESPONSE_KEYS = {
+    "application_type": "applicationType",
+    "time": "time",
+    "location": "location",
+    "reason": "reason",
+    "days": "days",
+    "transport_mode": "transportMode",
+    "amount": "amount",
+    "grade": "grade",
+    "department": "department",
+    "position": "position",
+    "manager_name": "managerName",
+    "lodging_daily_cap": "lodgingDailyCap",
+    "subsidy_daily_cap": "subsidyDailyCap",
+    "transport_policy": "transportPolicy",
+    "policy_estimate": "policyEstimate",
+    "matched_city": "matchedCity",
+    "rule_name": "ruleName",
+    "rule_version": "ruleVersion",
+    "hotel_amount": "hotelAmount",
+    "allowance_amount": "allowanceAmount",
+    "transport_estimated_amount": "transportEstimatedAmount",
+    "transport_estimate_source": "transportEstimateSource",
+    "transport_estimate_confidence": "transportEstimateConfidence",
+    "policy_total_amount": "policyTotalAmount",
+}
+
+
+def safe_text(value: object, limit: int = 500) -> str:
+    return str(value or "").strip()[:limit]
+
+
+def safe_fact_snapshot(facts: dict[str, Any]) -> dict[str, str]:
+    return {key: value for key in FACT_KEYS if (value := safe_text(facts.get(key)))}
+
+
+def canonical_preview_fields(preview: dict[str, Any]) -> dict[str, str]:
+    fields = preview.get("fields") if isinstance(preview, dict) else None
+    if not isinstance(fields, dict):
+        return {}
+    values: dict[str, str] = {}
+    for alias, field_key in PREVIEW_FIELD_ALIASES.items():
+        value = safe_text(fields.get(alias))
+        if value:
+            values[field_key] = value
+    return values
+
+
+def preview_response_fields(facts: dict[str, Any]) -> dict[str, str]:
+    snapshot = safe_fact_snapshot(facts)
+    return {
+        response_key: snapshot[field_key]
+        for field_key, response_key in PREVIEW_RESPONSE_KEYS.items()
+        if field_key in snapshot
+    }
+
+
+def hmac_fingerprint(
+    payload: object,
+    *,
+    key_version: str = FINGERPRINT_KEY_VERSION,
+    allow_key_creation: bool = True,
+) -> str:
+    serialized = _serialize(payload)
+    secret_key = get_expense_application_fingerprint_key(
+        key_version,
+        create=allow_key_creation,
+    )
+    digest = hmac.new(
+        secret_key,
+        b"expense-application-preview:v1:" + serialized,
+        hashlib.sha256,
+    ).hexdigest()
+    return f"hmac-sha256:{digest}"
+
+
+def content_fingerprint(payload: object) -> str:
+    return f"sha256:{hashlib.sha256(_serialize(payload)).hexdigest()}"
+
+
+def field_value_fingerprint(
+    value: str,
+    *,
+    present: bool,
+    key_version: str = FINGERPRINT_KEY_VERSION,
+    allow_key_creation: bool = True,
+) -> str:
+    return hmac_fingerprint(
+        {"present": present, "value": value if present else ""},
+        key_version=key_version,
+        allow_key_creation=allow_key_creation,
+    )
+
+
+def field_fingerprints(
+    values: dict[str, str],
+    *,
+    key_version: str = FINGERPRINT_KEY_VERSION,
+    allow_key_creation: bool = True,
+) -> dict[str, str]:
+    return {
+        key: field_value_fingerprint(
+            value,
+            present=True,
+            key_version=key_version,
+            allow_key_creation=allow_key_creation,
+        )
+        for key, value in sorted(values.items())
+    }
+
+
+def snapshot_reference(
+    values: dict[str, str],
+    *,
+    key_version: str = FINGERPRINT_KEY_VERSION,
+    allow_key_creation: bool = True,
+) -> dict[str, Any]:
+    return {
+        "field_keys": sorted(values),
+        "value_fingerprint": hmac_fingerprint(
+            values,
+            key_version=key_version,
+            allow_key_creation=allow_key_creation,
+        ),
+        "fingerprint_key_version": key_version,
+    }
+
+
+def _serialize(payload: object) -> bytes:
+    return json.dumps(
+        payload,
+        ensure_ascii=False,
+        sort_keys=True,
+        separators=(",", ":"),
+        default=str,
+    ).encode("utf-8")
--- a/server/src/app/services/user_agent_application.py
+++ b/server/src/app/services/user_agent_application.py
@@ -8,6 +8,7 @@ from sqlalchemy import or_, select
 from sqlalchemy.exc import IntegrityError

 from app.api.deps import CurrentUserContext
+from app.models.ai_application_preview import AIApplicationPreviewDecision
 from app.models.expense_case import BusinessEvent
 from app.models.financial_record import ExpenseClaim
 from app.schemas.reimbursement import TravelReimbursementCalculatorRequest
@@ -646,6 +647,8 @@ class UserAgentApplicationPersistenceMixin:
        *,
        action_type: str,
        business_event: BusinessEvent,
+        preview_decision: AIApplicationPreviewDecision | None = None,
+        action_request_id: str = "",
    ) -> None:
        if trusted_current_user is None:
            return
@@ -656,6 +659,8 @@ class UserAgentApplicationPersistenceMixin:
            trusted_current_user,
            action_type=action_type,
            business_event=business_event,
+            preview_decision=preview_decision,
+            action_request_id=action_request_id,
        )

    @staticmethod
@@ -727,6 +732,8 @@ class UserAgentApplicationPersistenceMixin:
        *,
        submit: bool,
        learning_current_user: CurrentUserContext | None = None,
+        learning_preview_decision: AIApplicationPreviewDecision | None = None,
+        learning_action_request_id: str = "",
    ) -> ExpenseClaim:
        current_user = self._build_application_current_user(payload)
        previous_status = str(claim.status or "").strip()
@@ -771,6 +778,8 @@ class UserAgentApplicationPersistenceMixin:
                    learning_current_user,
                    action_type="save_draft",
                    business_event=draft_event,
+                    preview_decision=learning_preview_decision,
+                    action_request_id=learning_action_request_id,
                )
                self.db.commit()
                self.db.refresh(claim)
@@ -794,6 +803,8 @@ class UserAgentApplicationPersistenceMixin:
                        learning_current_user,
                        action_type="submit",
                        business_event=event,
+                        preview_decision=learning_preview_decision,
+                        action_request_id=learning_action_request_id,
                    )
                ),
            )
@@ -811,6 +822,8 @@ class UserAgentApplicationPersistenceMixin:
        *,
        submit: bool,
        learning_current_user: CurrentUserContext | None = None,
+        learning_preview_decision: AIApplicationPreviewDecision | None = None,
+        learning_action_request_id: str = "",
    ) -> ExpenseClaim:
        current_user = self._build_application_current_user(payload)
        access_policy = ExpenseClaimAccessPolicy(self.db)
@@ -885,6 +898,8 @@ class UserAgentApplicationPersistenceMixin:
                    learning_current_user,
                    action_type="save_draft",
                    business_event=draft_event,
+                    preview_decision=learning_preview_decision,
+                    action_request_id=learning_action_request_id,
                )
                self.db.commit()
                self.db.refresh(claim)
@@ -904,6 +919,8 @@ class UserAgentApplicationPersistenceMixin:
                        learning_current_user,
                        action_type="submit",
                        business_event=event,
+                        preview_decision=learning_preview_decision,
+                        action_request_id=learning_action_request_id,
                    )
                ),
            )
@@ -1286,6 +1303,8 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
        *,
        risk_flags: list[str],
        learning_current_user: CurrentUserContext | None = None,
+        learning_preview_decision: AIApplicationPreviewDecision | None = None,
+        learning_action_request_id: str = "",
    ) -> UserAgentResponse:
        facts = self._resolve_expense_application_facts(payload)
        step = self._resolve_expense_application_step(payload, facts)
@@ -1299,6 +1318,8 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
                    editable_claim,
                    submit=step == "submitted",
                    learning_current_user=learning_current_user,
+                    learning_preview_decision=learning_preview_decision,
+                    learning_action_request_id=learning_action_request_id,
                )
                facts["application_edit_mode"] = "true"
            elif step == "submitted":
@@ -1312,6 +1333,8 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
                        facts,
                        submit=True,
                        learning_current_user=learning_current_user,
+                        learning_preview_decision=learning_preview_decision,
+                        learning_action_request_id=learning_action_request_id,
                    )
            else:
                application_claim = self._create_expense_application_record(
@@ -1319,6 +1342,8 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
                    facts,
                    submit=False,
                    learning_current_user=learning_current_user,
+                    learning_preview_decision=learning_preview_decision,
+                    learning_action_request_id=learning_action_request_id,
                )
            if application_claim is not None:
                facts["application_no"] = application_claim.claim_no
--- a/server/tests/auth_helpers.py
+++ b/server/tests/auth_helpers.py
@@ -24,6 +24,8 @@ def _read_test_user_headers(
    employee_no: Annotated[str | None, Header(alias="X-Auth-Employee-No")] = None,
    manager_name: Annotated[str | None, Header(alias="X-Auth-Manager-Name")] = None,
    tenant_id: Annotated[str | None, Header(alias="X-Auth-Tenant-Id")] = None,
+    employee_id: Annotated[str | None, Header(alias="X-Auth-Employee-Id")] = None,
+    auth_session_id: Annotated[str | None, Header(alias="X-Auth-Session-Id")] = None,
 ) -> CurrentUserContext:
    normalized_username = str(username or "").strip()
    normalized_name = str(name or normalized_username).strip()
@@ -52,4 +54,6 @@ def _read_test_user_headers(
        grade=str(grade or "").strip(),
        employee_no=str(employee_no or "").strip(),
        manager_name=str(manager_name or "").strip(),
+        employee_id=str(employee_id or "").strip(),
+        auth_session_id=str(auth_session_id or "").strip(),
    )
--- a/server/tests/test_alembic_migrations.py
+++ b/server/tests/test_alembic_migrations.py
@@ -19,7 +19,7 @@ from app.db.schema_ownership import MIGRATION_OWNED_TABLES, create_legacy_schema

 MIGRATION_TEST_DATABASE_URL = os.getenv("MIGRATION_TEST_DATABASE_URL", "").strip()
 LEGACY_PROBE_TABLE = "legacy_migration_probe_records"
-HEAD_REVISION = "20260714_0003"
+HEAD_REVISION = "20260714_0004"
 SERVER_DIR = Path(__file__).resolve().parents[1]
 ALEMBIC_INI_PATH = SERVER_DIR / "alembic.ini"

@@ -214,6 +214,18 @@ def _assert_head_schema(engine: Engine) -> None:
        "uq_ai_decisions_tenant_idempotency",
        ("tenant_id", "idempotency_key"),
    )
+    _assert_unique_constraint(
+        engine,
+        "ai_decisions",
+        "uq_ai_decisions_tenant_preview_decision",
+        ("tenant_id", "preview_decision_id"),
+    )
+    _assert_unique_constraint(
+        engine,
+        "ai_application_preview_decisions",
+        "uq_ai_application_preview_decisions_issue_request",
+        ("tenant_id", "actor_id", "auth_session_id", "issue_request_id"),
+    )
    _assert_unique_constraint(
        engine,
        "ai_decision_feedback",
@@ -265,6 +277,23 @@ def _assert_head_schema(engine: Engine) -> None:
            "ix_auth_sessions_tenant_username": ("tenant_id", "username"),
        },
    )
+    _assert_indexes(
+        engine,
+        "ai_application_preview_decisions",
+        {
+            "ix_ai_application_preview_decisions_actor_status_expiry": (
+                "tenant_id",
+                "actor_id",
+                "status",
+                "expires_at",
+            ),
+            "ix_ai_application_preview_decisions_conversation": (
+                "tenant_id",
+                "conversation_id",
+                "created_at",
+            ),
+        },
+    )
    _assert_indexes(
        engine,
        "ai_decisions",
@@ -305,6 +334,12 @@ def _assert_head_schema(engine: Engine) -> None:
    )
    _assert_cascade_foreign_key(engine, "expense_case_links")
    _assert_cascade_foreign_key(engine, "business_events")
+    _assert_composite_foreign_key(
+        engine,
+        "ai_decisions",
+        ("tenant_id", "preview_decision_id"),
+        "ai_application_preview_decisions",
+    )
    _assert_composite_foreign_key(
        engine,
        "ai_decisions",
--- a/server/tests/test_expense_application_preview_decisions.py
+++ b/server/tests/test_expense_application_preview_decisions.py
@@ -0,0 +1,420 @@
+from __future__ import annotations
+
+import stat
+from collections.abc import Generator
+from pathlib import Path
+
+import pytest
+from auth_helpers import install_legacy_header_auth_override
+from fastapi.testclient import TestClient
+from sqlalchemy import create_engine, select
+from sqlalchemy.orm import Session, sessionmaker
+from sqlalchemy.pool import StaticPool
+
+from app.api.deps import get_db
+from app.core import expense_application_fingerprint_keys as fingerprint_keys
+from app.db.base import Base
+from app.main import create_app
+from app.models.ai_application_preview import AIApplicationPreviewDecision
+from app.models.ai_learning import AIDecision, AIDecisionFeedback
+from app.models.employee import Employee
+from app.models.financial_record import ExpenseClaim
+from app.models.role import Role
+from app.services.expense_application_preview_decisions import (
+    ExpenseApplicationPreviewDecisionService,
+)
+
+
+def build_client() -> tuple[TestClient, sessionmaker[Session]]:
+    engine = create_engine(
+        "sqlite+pysqlite:///:memory:",
+        connect_args={"check_same_thread": False},
+        poolclass=StaticPool,
+    )
+    session_factory = sessionmaker(bind=engine, autoflush=False, autocommit=False)
+    Base.metadata.create_all(bind=engine)
+    app = create_app()
+    install_legacy_header_auth_override(app)
+
+    def override_db() -> Generator[Session, None, None]:
+        db = session_factory()
+        try:
+            yield db
+        finally:
+            db.close()
+
+    app.dependency_overrides[get_db] = override_db
+    return TestClient(app), session_factory
+
+
+def seed_employee(db: Session) -> None:
+    role = Role(id="role-preview-user", role_code="user", name="员工")
+    employee = Employee(
+        id="employee-preview-owner",
+        employee_no="E-PREVIEW-001",
+        name="张三",
+        email="preview-owner@example.com",
+        position="实施顾问",
+        grade="P4",
+        roles=[role],
+    )
+    db.add_all([role, employee])
+    db.commit()
+
+
+def auth_headers(*, session_id: str = "session-preview-owner") -> dict[str, str]:
+    return {
+        "X-Auth-Username": "preview-owner@example.com",
+        "X-Auth-Name": "Zhang San",
+        "X-Auth-Employee-No": "E-PREVIEW-001",
+        "X-Auth-Employee-Id": "employee-preview-owner",
+        "X-Auth-Session-Id": session_id,
+        "X-Auth-Tenant-Id": "tenant-preview",
+        "X-Auth-Grade": "P4",
+        "X-Auth-Role-Codes": "user",
+    }
+
+
+def issue_preview(
+    client: TestClient,
+    *,
+    request_id: str = "issue-preview-1",
+    include_transport: bool = True,
+) -> dict:
+    transport_line = "出行方式：火车\n" if include_transport else ""
+    response = client.post(
+        "/api/v1/reimbursements/application-previews",
+        headers=auth_headers(),
+        json={
+            "message": (
+                "申请时间：2026-07-20 至 2026-07-22\n"
+                "地点：上海\n事由：客户现场实施\n天数：3天\n"
+                f"{transport_line}申请金额：1800元"
+            ),
+            "conversation_id": "conversation-preview-1",
+            "request_id": request_id,
+        },
+    )
+    assert response.status_code == 201, response.text
+    return response.json()
+
+
+def build_action_payload(issued: dict, *, reason: str = "客户现场实施") -> dict:
+    fields = {
+        **issued["application_preview"]["fields"],
+        "reason": reason,
+    }
+    return {
+        "source": "user_message",
+        "user_id": "forged-user@example.com",
+        "conversation_id": "conversation-preview-1",
+        "action_type": "save_draft",
+        "decision_id": issued["decision_id"],
+        "request_id": "action-preview-save-1",
+        "message": (
+            "费用申请保存草稿\n申请时间：2026-07-20 至 2026-07-22\n"
+            f"地点：上海\n事由：{reason}\n申请金额：1800元\n保存草稿"
+        ),
+        "context_json": {
+            "application_action": "save_draft",
+            "application_save_mode": True,
+            "application_preview": {
+                "modelReviewStatus": "server_registered",
+                "fields": fields,
+            },
+        },
+    }
+
+
+def build_legacy_action_payload() -> dict:
+    return build_action_payload(
+        {
+            "decision_id": "legacy-placeholder",
+            "application_preview": {
+                "fields": {
+                    "time": "2026-07-20 至 2026-07-22",
+                    "location": "上海",
+                    "reason": "客户现场实施",
+                    "days": "3天",
+                    "transportMode": "火车",
+                    "amount": "1800元",
+                    "grade": "P4",
+                }
+            },
+        }
+    )
+
+
+def test_server_preview_decision_is_consumed_with_verified_feedback() -> None:
+    client, session_factory = build_client()
+    with session_factory() as db:
+        seed_employee(db)
+
+    issued = issue_preview(client)
+    action_payload = build_action_payload(issued)
+    response = client.post(
+        "/api/v1/reimbursements/application-preview-action",
+        headers=auth_headers(),
+        json=action_payload,
+    )
+
+    assert response.status_code == 200, response.text
+    result = response.json()["result"]
+    assert result["draft_payload"]["claim_id"]
+    assert result["decision_id"]
+    assert result["decision_id"] != issued["decision_id"]
+    retry_response = client.post(
+        "/api/v1/reimbursements/application-preview-action",
+        headers=auth_headers(),
+        json=action_payload,
+    )
+    assert retry_response.status_code == 200, retry_response.text
+    assert (
+        retry_response.json()["result"]["draft_payload"]["claim_id"]
+        == result["draft_payload"]["claim_id"]
+    )
+    assert retry_response.json()["result"]["decision_id"] == result["decision_id"]
+
+    with session_factory() as db:
+        consumed = db.get(AIApplicationPreviewDecision, issued["decision_id"])
+        next_decision = db.get(AIApplicationPreviewDecision, result["decision_id"])
+        decision = db.scalar(select(AIDecision))
+        feedback = db.scalar(select(AIDecisionFeedback))
+        assert consumed is not None
+        assert consumed.status == "consumed"
+        assert consumed.consumed_action == "save_draft"
+        assert consumed.field_fingerprints_json
+        assert "客户现场实施" not in str(consumed.field_fingerprints_json)
+        assert next_decision is not None
+        assert next_decision.status == "issued"
+        assert next_decision.decision_source == "server_draft"
+        assert decision is not None
+        assert decision.preview_decision_id == consumed.id
+        assert decision.suggestion_json["value_fingerprint"].startswith("hmac-sha256:")
+        assert feedback is not None
+        assert feedback.verification_status == "server_verified"
+        assert feedback.feedback_type == "accepted"
+        assert feedback.training_eligible is False
+        assert len(list(db.scalars(select(AIDecision)).all())) == 1
+
+
+def test_server_preview_decision_detects_server_side_field_edit() -> None:
+    client, session_factory = build_client()
+    with session_factory() as db:
+        seed_employee(db)
+
+    issued = issue_preview(client, request_id="issue-preview-edited")
+    response = client.post(
+        "/api/v1/reimbursements/application-preview-action",
+        headers=auth_headers(),
+        json=build_action_payload(issued, reason="客户现场验收"),
+    )
+
+    assert response.status_code == 200, response.text
+    with session_factory() as db:
+        feedback = db.scalar(select(AIDecisionFeedback))
+        assert feedback is not None
+        assert feedback.feedback_type == "edited"
+        assert feedback.verification_status == "server_verified"
+        assert feedback.changed_fields_json == [
+            {
+                "field_key": "reason",
+                "suggested_value_fingerprint": feedback.changed_fields_json[0][
+                    "suggested_value_fingerprint"
+                ],
+                "final_value_fingerprint": feedback.changed_fields_json[0][
+                    "final_value_fingerprint"
+                ],
+            }
+        ]
+        assert "客户现场" not in str(feedback.changed_fields_json)
+
+
+def test_server_preview_decision_detects_field_added_after_issuance() -> None:
+    client, session_factory = build_client()
+    with session_factory() as db:
+        seed_employee(db)
+
+    issued = issue_preview(
+        client,
+        request_id="issue-preview-added-field",
+        include_transport=False,
+    )
+    payload = build_action_payload(issued)
+    payload["message"] = payload["message"].replace(
+        "申请金额：1800元",
+        "出行方式：飞机\n申请金额：1800元",
+    )
+    payload["context_json"]["application_preview"]["fields"]["transportMode"] = "飞机"
+    response = client.post(
+        "/api/v1/reimbursements/application-preview-action",
+        headers=auth_headers(),
+        json=payload,
+    )
+
+    assert response.status_code == 200, response.text
+    with session_factory() as db:
+        feedback = db.scalar(select(AIDecisionFeedback))
+        assert feedback is not None
+        assert feedback.feedback_type == "edited"
+        assert {item["field_key"] for item in feedback.changed_fields_json} >= {"transport_mode"}
+
+
+def test_server_preview_decision_rejects_cross_session_replay_without_writes() -> None:
+    client, session_factory = build_client()
+    with session_factory() as db:
+        seed_employee(db)
+
+    issued = issue_preview(client, request_id="issue-preview-cross-session")
+    response = client.post(
+        "/api/v1/reimbursements/application-preview-action",
+        headers=auth_headers(session_id="another-session"),
+        json=build_action_payload(issued),
+    )
+
+    assert response.status_code == 400
+    assert "不属于当前登录会话" in response.json()["detail"]
+    with session_factory() as db:
+        decision = db.get(AIApplicationPreviewDecision, issued["decision_id"])
+        assert decision is not None
+        assert decision.status == "issued"
+        assert list(db.scalars(select(ExpenseClaim)).all()) == []
+        assert list(db.scalars(select(AIDecision)).all()) == []
+
+
+def test_active_server_preview_cannot_downgrade_by_omitting_decision_id() -> None:
+    client, session_factory = build_client()
+    with session_factory() as db:
+        seed_employee(db)
+
+    issued = issue_preview(client, request_id="issue-preview-downgrade")
+    payload = build_action_payload(issued)
+    payload.pop("decision_id")
+    payload.pop("request_id")
+    response = client.post(
+        "/api/v1/reimbursements/application-preview-action",
+        headers=auth_headers(),
+        json=payload,
+    )
+
+    assert response.status_code == 409, response.text
+    assert "必须携带 decision_id" in response.json()["detail"]
+    with session_factory() as db:
+        decision = db.get(AIApplicationPreviewDecision, issued["decision_id"])
+        assert decision is not None
+        assert decision.status == "issued"
+        assert list(db.scalars(select(ExpenseClaim)).all()) == []
+        assert list(db.scalars(select(AIDecision)).all()) == []
+
+
+def test_legacy_action_without_active_server_decision_keeps_client_observed_path() -> None:
+    client, session_factory = build_client()
+    with session_factory() as db:
+        seed_employee(db)
+
+    payload = build_legacy_action_payload()
+    payload.pop("decision_id")
+    payload.pop("request_id")
+    response = client.post(
+        "/api/v1/reimbursements/application-preview-action",
+        headers=auth_headers(),
+        json=payload,
+    )
+
+    assert response.status_code == 200, response.text
+    with session_factory() as db:
+        feedback = db.scalar(select(AIDecisionFeedback))
+        assert feedback is not None
+        assert feedback.verification_status == "client_observed"
+        assert feedback.training_eligible is False
+
+
+def test_missing_fingerprint_key_rejects_action_without_business_writes() -> None:
+    client, session_factory = build_client()
+    with session_factory() as db:
+        seed_employee(db)
+
+    issued = issue_preview(client, request_id="issue-preview-missing-key")
+    with session_factory() as db:
+        decision = db.get(AIApplicationPreviewDecision, issued["decision_id"])
+        assert decision is not None
+        decision.fingerprint_key_version = "missing-test-key"
+        db.commit()
+
+    response = client.post(
+        "/api/v1/reimbursements/application-preview-action",
+        headers=auth_headers(),
+        json=build_action_payload(issued),
+    )
+
+    assert response.status_code == 400, response.text
+    assert "密钥版本不可用" in response.json()["detail"]
+    with session_factory() as db:
+        decision = db.get(AIApplicationPreviewDecision, issued["decision_id"])
+        assert decision is not None
+        assert decision.status == "issued"
+        assert list(db.scalars(select(ExpenseClaim)).all()) == []
+        assert list(db.scalars(select(AIDecision)).all()) == []
+
+
+def test_draft_save_stays_successful_when_next_decision_issuance_fails(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    client, session_factory = build_client()
+    with session_factory() as db:
+        seed_employee(db)
+
+    original_issue = ExpenseApplicationPreviewDecisionService.issue
+
+    def fail_server_draft_issue(
+        service: ExpenseApplicationPreviewDecisionService,
+        *args: object,
+        **kwargs: object,
+    ) -> object:
+        if kwargs.get("decision_source") == "server_draft":
+            raise RuntimeError("simulated next decision failure")
+        return original_issue(service, *args, **kwargs)
+
+    monkeypatch.setattr(
+        ExpenseApplicationPreviewDecisionService,
+        "issue",
+        fail_server_draft_issue,
+    )
+    issued = issue_preview(client, request_id="issue-preview-next-failure")
+    response = client.post(
+        "/api/v1/reimbursements/application-preview-action",
+        headers=auth_headers(),
+        json=build_action_payload(issued),
+    )
+
+    assert response.status_code == 200, response.text
+    assert response.json()["result"]["draft_payload"]["claim_id"]
+    assert response.json()["result"]["decision_id"] is None
+    with session_factory() as db:
+        consumed = db.get(AIApplicationPreviewDecision, issued["decision_id"])
+        assert consumed is not None
+        assert consumed.status == "consumed"
+        assert len(list(db.scalars(select(ExpenseClaim)).all())) == 1
+        assert len(list(db.scalars(select(AIDecision)).all())) == 1
+
+
+def test_fingerprint_key_is_created_with_private_permissions(
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    key_directory = tmp_path / "fingerprint-keys"
+    monkeypatch.setattr(
+        fingerprint_keys,
+        "FINGERPRINT_KEY_DIRECTORY",
+        key_directory,
+    )
+
+    key = fingerprint_keys.get_expense_application_fingerprint_key(
+        "permission-test",
+        create=True,
+    )
+
+    key_path = key_directory / "permission-test.key"
+    assert len(key) == fingerprint_keys.KEY_BYTES
+    assert stat.S_IMODE(key_directory.stat().st_mode) == 0o700
+    assert stat.S_IMODE(key_path.stat().st_mode) == 0o600
--- a/server/tests/test_migration_preflight.py
+++ b/server/tests/test_migration_preflight.py
@@ -48,7 +48,7 @@ def test_unversioned_database_without_migration_owned_tables_is_safe(engine: Eng

@pytest.mark.parametrize(
    "owned_table",
-    sorted(MIGRATION_OWNED_TABLES_BY_REVISION["20260714_0003"]),
+    sorted(MIGRATION_OWNED_TABLES_BY_REVISION["20260714_0004"]),
 )
 def test_unversioned_database_with_any_migration_owned_table_is_rejected(
    engine: Engine,
@@ -91,6 +91,11 @@ def test_known_revision_requires_and_accepts_its_exact_owned_table_set(
            "20260714_0003",
            MIGRATION_OWNED_TABLES_BY_REVISION["20260714_0003"] - {"ai_decisions"},
        ),
+        (
+            "20260714_0004",
+            MIGRATION_OWNED_TABLES_BY_REVISION["20260714_0004"]
+            - {"ai_application_preview_decisions"},
+        ),
    ],
 )
 def test_known_revision_with_missing_or_unexpected_owned_tables_is_rejected(
--- a/server/tests/test_schema_ownership.py
+++ b/server/tests/test_schema_ownership.py
@@ -16,6 +16,7 @@ def test_create_legacy_schema_never_creates_migration_owned_tables() -> None:

    assert MIGRATION_OWNED_TABLES == frozenset(
        {
+            "ai_application_preview_decisions",
            "ai_decision_feedback",
            "ai_decisions",
            "auth_sessions",
--- a/server/tests/test_user_agent_application_draft_events.py
+++ b/server/tests/test_user_agent_application_draft_events.py
@@ -441,14 +441,14 @@ def test_ai_application_draft_records_explicit_field_correction() -> None:
        assert decision is not None
        assert decision.training_eligible is False
        assert decision.suggestion_json["field_keys"]
-        assert decision.suggestion_json["value_fingerprint"].startswith("sha256:")
+        assert decision.suggestion_json["value_fingerprint"].startswith("hmac-sha256:")
        assert feedback is not None
        assert feedback.feedback_type == "edited"
        assert feedback.verification_status == "client_observed"
        assert feedback.training_eligible is False
        assert feedback.changed_fields_json[0]["field_key"] == "reason"
        assert feedback.changed_fields_json[0]["suggested_value_fingerprint"].startswith(
-            "sha256:"
+            "hmac-sha256:"
        )
        assert "客户拜访" not in str(feedback.changed_fields_json)
        assert "客户现场实施" not in str(feedback.final_value_json)
--- a/web/src/composables/workbenchAiMode/useWorkbenchAiApplicationPreviewFlow.js
+++ b/web/src/composables/workbenchAiMode/useWorkbenchAiApplicationPreviewFlow.js
@@ -13,6 +13,8 @@ import { fetchExpenseClaims } from '../../services/reimbursements.js'
 import {
  AI_APPLICATION_ACTION_SAVE_DRAFT,
  AI_APPLICATION_ACTION_SUBMIT,
+  createAiApplicationRequestId,
+  registerAiApplicationPreviewDecision,
  runAiApplicationPreviewAction
 } from '../../services/aiApplicationPreviewActions.js'
 import {
@@ -374,9 +376,20 @@ export function useWorkbenchAiApplicationPreviewFlow({
      return false
    }

-    const normalizedPreview = normalizeApplicationPreview(targetMessage.applicationPreview)
+    let normalizedPreview = normalizeApplicationPreview(targetMessage.applicationPreview)
    const isSubmit = actionType === AI_APPLICATION_ACTION_SUBMIT
    const userText = String(options.userText || (isSubmit ? '直接提交' : '保存草稿')).trim()
+    const actionRequestId = String(
+      normalizedPreview.decisionActionRequestId || createAiApplicationRequestId('action')
+    ).trim()
+    if (normalizedPreview.decisionId) {
+      normalizedPreview = normalizeApplicationPreview({
+        ...normalizedPreview,
+        decisionActionRequestId: actionRequestId
+      })
+      targetMessage.applicationPreview = normalizedPreview
+      persistCurrentConversation()
+    }

    if (isSubmit && !normalizedPreview.readyToSubmit) {
      if (!options.skipUserMessage) {
@@ -454,12 +467,27 @@ export function useWorkbenchAiApplicationPreviewFlow({
        applicationPreview: normalizedPreview,
        currentUser: currentUser.value || {},
        conversationId: conversationId.value,
-        draftPayload: targetMessage.draftPayload || options.draftPayload || null
+        draftPayload: targetMessage.draftPayload || options.draftPayload || null,
+        requestId: actionRequestId
      })
      const draftPayload = extractInlineApplicationDraftPayload(payload)
      if (draftPayload) {
        targetMessage.draftPayload = draftPayload
      }
+      const nextDecisionId = String(payload?.result?.decision_id || '').trim()
+      targetMessage.applicationPreview = normalizeApplicationPreview({
+        ...(targetMessage.applicationPreview || normalizedPreview),
+        ...(nextDecisionId
+          ? {
+              decisionId: nextDecisionId,
+              decisionExpiresAt: String(payload?.result?.decision_expires_at || '').trim(),
+              decisionTrackingStatus: 'registered'
+            }
+          : {
+              decisionTrackingStatus: normalizedPreview.decisionId ? 'consumed' : 'legacy'
+            }),
+        decisionActionRequestId: ''
+      })
      targetMessage.suggestedActions = []
      const detailActions = buildInlineApplicationDetailAction(draftPayload)
      const nextTaskAction = buildApplicationPreviewNextTaskAction(targetMessage)
@@ -598,11 +626,26 @@ export function useWorkbenchAiApplicationPreviewFlow({
          ontologyFields: options.ontologyFields
        })
      )
-      const content = buildLocalApplicationPreviewMessage(preview)
+      let registeredPreview = preview
+      try {
+        registeredPreview = await registerAiApplicationPreviewDecision({
+          applicationPreview: preview,
+          conversationId: conversationId.value,
+          message: previewSourceText,
+          requestId: createAiApplicationRequestId('preview')
+        })
+      } catch (error) {
+        console.warn('Application preview decision registration failed:', error)
+        registeredPreview = normalizeApplicationPreview({
+          ...preview,
+          decisionTrackingStatus: 'degraded'
+        })
+      }
+      const content = buildLocalApplicationPreviewMessage(registeredPreview)
      const previewMessage = createInlineMessage('assistant', content, {
        id: pendingMessage.id,
-        applicationPreview: preview,
-        suggestedActions: buildInlineApplicationPreviewSuggestedActions(preview),
+        applicationPreview: registeredPreview,
+        suggestedActions: buildInlineApplicationPreviewSuggestedActions(registeredPreview),
        requestedSubmit: Boolean(options.requestedSubmit),
        submitRequiresConfirmation: Boolean(options.submitRequiresConfirmation),
        stewardRemainingTasks: Array.isArray(options.stewardRemainingTasks) ? options.stewardRemainingTasks : [],
--- a/web/src/services/aiApplicationPreviewActions.js
+++ b/web/src/services/aiApplicationPreviewActions.js
@@ -8,6 +8,54 @@ import {
 export const AI_APPLICATION_ACTION_SAVE_DRAFT = 'ai_application_save_draft'
 export const AI_APPLICATION_ACTION_SUBMIT = 'ai_application_submit'

+export function createAiApplicationRequestId(prefix = 'application') {
+  const normalizedPrefix = normalizeText(prefix) || 'application'
+  if (typeof globalThis.crypto?.randomUUID === 'function') {
+    return `${normalizedPrefix}:${globalThis.crypto.randomUUID()}`
+  }
+  return `${normalizedPrefix}:${Date.now()}:${Math.random().toString(16).slice(2)}`
+}
+
+export function resolveAiApplicationPreviewDecisionId(preview = {}) {
+  return normalizeText(preview?.decisionId || preview?.decision_id)
+}
+
+export async function registerAiApplicationPreviewDecision({
+  applicationPreview,
+  conversationId = '',
+  message = '',
+  requestId = ''
+} = {}, options = {}) {
+  const normalizedPreview = normalizeApplicationPreview(applicationPreview || {})
+  const issueRequestId = normalizeText(requestId) || createAiApplicationRequestId('preview')
+  const payload = await apiRequest('/reimbursements/application-previews', {
+    method: 'POST',
+    body: JSON.stringify({
+      message: normalizeText(message),
+      conversation_id: normalizeText(conversationId) || null,
+      request_id: issueRequestId
+    }),
+    timeoutMs: 30000,
+    timeoutMessage: '申请预览校验超时，已切换为普通核对模式。',
+    ...options
+  })
+  const serverPreview = payload?.application_preview && typeof payload.application_preview === 'object'
+    ? payload.application_preview
+    : {}
+  return normalizeApplicationPreview({
+    ...normalizedPreview,
+    ...serverPreview,
+    fields: {
+      ...(normalizedPreview.fields || {}),
+      ...(serverPreview.fields || {})
+    },
+    decisionId: normalizeText(payload?.decision_id || serverPreview.decisionId),
+    decisionSource: normalizeText(payload?.decision_source || serverPreview.decisionSource),
+    decisionExpiresAt: normalizeText(payload?.expires_at || serverPreview.decisionExpiresAt),
+    decisionTrackingStatus: 'registered'
+  })
+}
+
 function normalizeText(value) {
  return String(value || '').trim()
 }
@@ -56,7 +104,8 @@ export function buildAiApplicationPreviewActionPayload({
  applicationPreview,
  currentUser = {},
  conversationId = '',
-  draftPayload = null
+  draftPayload = null,
+  requestId = ''
 } = {}) {
  const normalizedPreview = normalizeApplicationPreview(applicationPreview || {})
  const message = buildAiApplicationPreviewActionText(actionType, normalizedPreview)
@@ -72,6 +121,8 @@ export function buildAiApplicationPreviewActionPayload({
    : []
  const draftClaimId = normalizeText(draftPayload?.claim_id || draftPayload?.claimId)
  const isSubmit = actionType === AI_APPLICATION_ACTION_SUBMIT
+  const decisionId = resolveAiApplicationPreviewDecisionId(normalizedPreview)
+  const actionRequestId = normalizeText(requestId || normalizedPreview.decisionActionRequestId)
  const applicationEditableFields = Array.isArray(normalizedPreview.editableFields)
    ? normalizedPreview.editableFields.map((field) => normalizeText(field)).filter(Boolean)
    : []
@@ -80,6 +131,9 @@ export function buildAiApplicationPreviewActionPayload({
    source: 'user_message',
    user_id: username,
    conversation_id: normalizeText(conversationId) || null,
+    action_type: isSubmit ? 'submit' : 'save_draft',
+    ...(decisionId ? { decision_id: decisionId } : {}),
+    ...(decisionId && actionRequestId ? { request_id: actionRequestId } : {}),
    message,
    context_json: {
      role_codes: roleCodes,
--- a/web/src/utils/aiWorkbenchConversationStore.js
+++ b/web/src/utils/aiWorkbenchConversationStore.js
@@ -257,6 +257,15 @@ function normalizeMessage(message = {}) {
    content: safeString(message.content),
    pending: false,
    feedback: safeString(message.feedback),
+    text: safeString(message.text || message.content),
+    applicationPreview: message.applicationPreview && typeof message.applicationPreview === 'object'
+      ? message.applicationPreview
+      : null,
+    draftPayload: message.draftPayload && typeof message.draftPayload === 'object'
+      ? message.draftPayload
+      : null,
+    requestedSubmit: Boolean(message.requestedSubmit),
+    submitRequiresConfirmation: Boolean(message.submitRequiresConfirmation),
    stewardPlan: message.stewardPlan && typeof message.stewardPlan === 'object'
      ? {
          ...message.stewardPlan,
--- a/web/tests/ai-application-preview-actions.test.mjs
+++ b/web/tests/ai-application-preview-actions.test.mjs
@@ -3,6 +3,7 @@ import assert from 'node:assert/strict'
 import {
  AI_APPLICATION_ACTION_SAVE_DRAFT,
  AI_APPLICATION_ACTION_SUBMIT,
+  registerAiApplicationPreviewDecision,
  runAiApplicationPreviewAction
 } from '../src/services/aiApplicationPreviewActions.js'

@@ -34,6 +35,7 @@ async function testSubmitActionUsesFastPreviewEndpoint() {
  await runAiApplicationPreviewAction({
    actionType: AI_APPLICATION_ACTION_SUBMIT,
    applicationPreview: {
+      decisionId: 'decision-fast-submit',
      fields: {
        applicationType: '差旅费用申请',
        time: '2026-07-01 至 2026-07-03',
@@ -55,7 +57,8 @@ async function testSubmitActionUsesFastPreviewEndpoint() {
      }
    },
    currentUser: { username: 'zhangsan@example.com', name: '张三' },
-    conversationId: 'conversation-fast-submit'
+    conversationId: 'conversation-fast-submit',
+    requestId: 'action-fast-submit'
  })

  assert.equal(capturedUrl, '/api/v1/reimbursements/application-preview-action')
@@ -63,6 +66,9 @@ async function testSubmitActionUsesFastPreviewEndpoint() {
  const body = JSON.parse(capturedOptions.body)
  assert.equal(body.context_json.session_type, 'application')
  assert.equal(body.context_json.application_stage, 'expense_application')
+  assert.equal(body.action_type, 'submit')
+  assert.equal(body.decision_id, 'decision-fast-submit')
+  assert.equal(body.request_id, 'action-fast-submit')
  assert.equal(body.context_json.application_preview.fields.transportMode, '火车')
  assert.deepEqual(body.context_json.application_preview.aiDecisionFeedback.changedFields, [
    {
@@ -110,6 +116,59 @@ async function testSaveDraftActionUsesFastPreviewEndpoint() {
  assert.equal(body.context_json.application_action, 'save_draft')
  assert.equal(body.context_json.application_save_mode, true)
  assert.equal(body.context_json.application_stage, 'expense_application')
+  assert.equal(body.action_type, 'save_draft')
+  assert.equal('decision_id' in body, false)
+  assert.equal('request_id' in body, false)
+}
+
+async function testRegistrationUsesServerCanonicalPreview() {
+  let capturedUrl = ''
+  let capturedOptions = null
+  global.fetch = async (url, options) => {
+    capturedUrl = String(url)
+    capturedOptions = options
+    return {
+      ok: true,
+      async json() {
+        return {
+          decision_id: 'decision-issued-1',
+          decision_source: 'rule',
+          expires_at: '2026-07-14T10:30:00Z',
+          application_preview: {
+            fields: {
+              reason: '服务端规范事由',
+              amount: '1800元'
+            }
+          }
+        }
+      }
+    }
+  }
+
+  const preview = await registerAiApplicationPreviewDecision({
+    applicationPreview: {
+      fields: {
+        reason: '客户端事由',
+        location: '上海'
+      }
+    },
+    conversationId: 'conversation-issued-1',
+    message: '去上海做客户现场实施',
+    requestId: 'preview-request-1'
+  })
+
+  assert.equal(capturedUrl, '/api/v1/reimbursements/application-previews')
+  assert.equal(capturedOptions.method, 'POST')
+  assert.deepEqual(JSON.parse(capturedOptions.body), {
+    message: '去上海做客户现场实施',
+    conversation_id: 'conversation-issued-1',
+    request_id: 'preview-request-1'
+  })
+  assert.equal(preview.decisionId, 'decision-issued-1')
+  assert.equal(preview.decisionSource, 'rule')
+  assert.equal(preview.fields.reason, '服务端规范事由')
+  assert.equal(preview.fields.location, '上海')
+  assert.equal(preview.decisionTrackingStatus, 'registered')
 }

 async function testEditDraftActionCarriesClaimAndEditableFields() {
@@ -167,6 +226,7 @@ async function testEditDraftActionCarriesClaimAndEditableFields() {
 async function run() {
  await testSubmitActionUsesFastPreviewEndpoint()
  await testSaveDraftActionUsesFastPreviewEndpoint()
+  await testRegistrationUsesServerCanonicalPreview()
  await testEditDraftActionCarriesClaimAndEditableFields()
  console.log('ai-application-preview-actions tests passed')
 }
--- a/web/tests/ai-workbench-conversation-store.test.mjs
+++ b/web/tests/ai-workbench-conversation-store.test.mjs
@@ -101,3 +101,45 @@ test('AI workbench conversation store preserves stewardRemainingTasks on message
  const persistedMessage = conversation.messages.find((m) => m.id === 'a1')
  assert.deepEqual(persistedMessage.stewardRemainingTasks, remainingTasks)
 })
+
+test('AI workbench conversation store preserves preview decision and retry context', () => {
+  installLocalStorageMock()
+  const user = { username: 'preview-user' }
+  const applicationPreview = {
+    decisionId: 'decision-preview-1',
+    decisionActionRequestId: 'action-preview-1',
+    fields: { reason: '客户现场实施' },
+    aiDecisionFeedback: {
+      schemaVersion: 1,
+      revision: 1,
+      changedFields: [
+        { fieldKey: 'reason', suggestedValue: '客户拜访', finalValue: '客户现场实施' }
+      ]
+    }
+  }
+
+  saveAiWorkbenchConversation(user, {
+    id: 'conv-preview-decision',
+    updatedAt: Date.now(),
+    messages: [
+      { id: 'u1', role: 'user', content: '发起出差申请' },
+      {
+        id: 'a1',
+        role: 'assistant',
+        content: '申请核对表',
+        text: '申请核对表',
+        applicationPreview,
+        draftPayload: { claim_id: 'claim-preview-1' },
+        requestedSubmit: true,
+        submitRequiresConfirmation: true
+      }
+    ]
+  })
+
+  const history = loadAiWorkbenchConversationHistory(user)
+  const restored = history[0].messages.find((message) => message.id === 'a1')
+  assert.deepEqual(restored.applicationPreview, applicationPreview)
+  assert.equal(restored.draftPayload.claim_id, 'claim-preview-1')
+  assert.equal(restored.requestedSubmit, true)
+  assert.equal(restored.submitRequiresConfirmation, true)
+})
--- a/web/tests/expense-application-decision-feedback.test.mjs
+++ b/web/tests/expense-application-decision-feedback.test.mjs
@@ -4,6 +4,7 @@ import { trackApplicationPreviewFieldEdit } from '../src/utils/expenseApplicatio

 function testTracksFirstExplicitEdit() {
  const previous = {
+    decisionId: 'decision-feedback-1',
    fields: {
      reason: '客户拜访',
      location: '上海'
@@ -27,12 +28,13 @@ function testTracksFirstExplicitEdit() {
    }
  ])
  assert.equal(tracked.aiDecisionFeedback.revision, 1)
+  assert.equal(tracked.decisionId, 'decision-feedback-1')
 }

 function testKeepsOriginalSuggestionAcrossRepeatedEdits() {
  const first = trackApplicationPreviewFieldEdit(
-    { fields: { transportMode: '火车' } },
-    { fields: { transportMode: '飞机' } },
+    { decisionId: 'decision-feedback-2', fields: { transportMode: '火车' } },
+    { decisionId: 'decision-feedback-2', fields: { transportMode: '飞机' } },
    'transportMode'
  )
  const second = trackApplicationPreviewFieldEdit(
@@ -48,6 +50,7 @@ function testKeepsOriginalSuggestionAcrossRepeatedEdits() {
      finalValue: '自驾'
    }
  ])
+  assert.equal(second.decisionId, 'decision-feedback-2')
 }

 function testRemovesDiffWhenUserRestoresSuggestion() {