feat(ai): add expense application feedback ledger

document/development/2026-07-13/feature/ai-expense-closed-loop-and-value-proof/CONCEPT.md

@@ -514,3 +514,8 @@ docker exec -w /app -e SERVER_VENV_DIR=/tmp/x-financial-server-venv \
 - 2026-07-14（历史旧单接入）：新增独立 `ExpenseCaseLegacyBackfillService` 和 `backfill_legacy_expense_claim_cases.py`。命令只读取显式 `DATABASE_URL`，默认 dry-run；apply 强制目标核验、精确确认、迁移 head、advisory lock 和批次事务。每张旧单只创建一条 `historical_claim_imported` 系统快照，保留源时间和指纹、明确不重建历史，并以 `suppressed` 阻止实时投递。
 - 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（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。

document/development/2026-07-13/feature/ai-expense-closed-loop-and-value-proof/TODO.md

@@ -37,7 +37,8 @@
 - [ ] [CONCEPT: 费用领域与编排] 定义 `Expense Case`、申请、票据、报销、审批、付款、凭证和归档的领域边界与迁移关系。
 - [ ] [CONCEPT: 数据与契约] 定义 `expense_cases`、`expense_case_links` 和最小状态机，明确非法状态跃迁。
 - [ ] [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。
-- [ ] [CONCEPT: 业务事件与 AI 决策] 定义 `ai_decisions`、`ai_decision_feedback` 和 `workflow_outcomes` 契约。
+- [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。
 - [ ] [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。
@@ -64,7 +65,8 @@
   证据：`server/src/app/services/expense_cases.py`、`server/src/app/api/v1/endpoints/expense_cases.py`、`GET /api/v1/expense-cases/by-claim/{claim_id}`；容器 OpenAPI 校验通过。
 - [x] [CONCEPT: 数据与契约] 新增 `expense_cases`、`expense_case_links` 和 `business_events` 表及迁移。
   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。
-- [ ] [CONCEPT: 业务事件与 AI 决策] 新增 `ai_decisions`、`ai_decision_feedback`、`workflow_outcomes` 表及迁移。
+- [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。
 - [ ] [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。
@@ -109,6 +111,8 @@
 - [ ] [CONCEPT: 风险与预审] 统一风险输出为事实、规则、证据、判断、建议动作和降级原因。
 - [ ] [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。
 - [ ] [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。
@@ -165,6 +169,8 @@
 - [ ] [CONCEPT: 测试方案] 为连接器幂等、重试、回执、失败恢复、重复付款和对账补充测试。
 - [ ] [CONCEPT: 测试方案] 跑通申请 → 票据 → 报销 → 预审 → 审批 → 付款 → 入账 → 归档端到端。
 - [ ] [CONCEPT: 测试方案] 跑通 AI 建议 → 用户修改 → 工作流结果 → 记忆激活 → 下次建议变化闭环。
+- [x] [CONCEPT: 测试方案] 跑通首个行为采集切片：AI 申请预填 → 用户接受/显式修改 → 草稿或提交结果同事务落账。
+  证据：`test_user_agent_application_draft_events.py`、`test_reimbursement_endpoints.py`、`expense-application-decision-feedback.test.mjs`、`expense-application-fast-preview.test.mjs`；覆盖可信入口、模板/详情排除、隐私指纹、同事务事件关联、日期联动及异步乱序响应。容器内学习账本与迁移所有权定向 29 项、一次性 PostgreSQL 迁移 4 项和前端关键场景 5 项通过。
 - [ ] [CONCEPT: 测试方案] 跑通风险反馈 → few-shot → golden case → Canary → 回滚闭环。
 - [ ] [CONCEPT: 测试方案] 跑通节省机会 → 执行 → 实现 → 财务确认 → ROI 看板闭环。
 - [x] [CONCEPT: 测试方案] 为已有 Expense Case 事件时间线补充视图模型、404 降级、详情页接入及相关响应式回归，并完成前端生产构建。

server/alembic/versions/20260714_0003_ai_learning_loop.py

@@ -0,0 +1,309 @@
+"""add AI decision feedback and workflow outcome ledger
+
+Revision ID: 20260714_0003
+Revises: 20260713_0002
+Create Date: 2026-07-14 10:30:00
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision: str = "20260714_0003"
+down_revision: str | None = "20260713_0002"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    # 为新账本的复合外键补充租户级唯一目标，阻止跨租户关联。
+    op.create_unique_constraint(
+        "uq_expense_cases_tenant_id",
+        "expense_cases",
+        ["tenant_id", "id"],
+    )
+    op.create_unique_constraint(
+        "uq_business_events_tenant_case_id",
+        "business_events",
+        ["tenant_id", "expense_case_id", "id"],
+    )
+
+    op.create_table(
+        "ai_decisions",
+        sa.Column("id", sa.String(length=36), nullable=False),
+        sa.Column("tenant_id", sa.String(length=64), nullable=False),
+        sa.Column("expense_case_id", sa.String(length=36), nullable=False),
+        sa.Column("business_event_id", sa.String(length=36), nullable=True),
+        sa.Column("expense_claim_id", sa.String(length=36), nullable=False),
+        sa.Column("agent_run_id", sa.String(length=120), nullable=True),
+        sa.Column("correlation_id", sa.String(length=64), nullable=False),
+        sa.Column("subject_type", sa.String(length=50), nullable=False),
+        sa.Column("subject_id", sa.String(length=100), nullable=False),
+        sa.Column("decision_type", sa.String(length=80), nullable=False),
+        sa.Column("decision_source", sa.String(length=20), nullable=False),
+        sa.Column("status", sa.String(length=20), nullable=False),
+        sa.Column("automation_mode", sa.String(length=40), nullable=False),
+        sa.Column("confidence", sa.Numeric(precision=5, scale=4), nullable=True),
+        sa.Column("suggestion_json", sa.JSON(), nullable=False),
+        sa.Column("evidence_json", sa.JSON(), nullable=False),
+        sa.Column("version_json", sa.JSON(), nullable=False),
+        sa.Column("schema_version", sa.Integer(), nullable=False),
+        sa.Column(
+            "training_eligible",
+            sa.Boolean(),
+            server_default=sa.false(),
+            nullable=False,
+        ),
+        sa.Column("idempotency_key", sa.String(length=120), nullable=False),
+        sa.Column("content_fingerprint", sa.String(length=71), nullable=False),
+        sa.Column(
+            "created_at",
+            sa.DateTime(timezone=True),
+            server_default=sa.func.now(),
+            nullable=False,
+        ),
+        sa.CheckConstraint(
+            "confidence IS NULL OR (confidence >= 0 AND confidence <= 1)",
+            name="ck_ai_decisions_confidence",
+        ),
+        sa.CheckConstraint(
+            "status IN ('suggested', 'accepted', 'edited', 'rejected', "
+            "'ignored', 'executed', 'rolled_back')",
+            name="ck_ai_decisions_status",
+        ),
+        sa.ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id", "business_event_id"],
+            [
+                "business_events.tenant_id",
+                "business_events.expense_case_id",
+                "business_events.id",
+            ],
+            name="fk_ai_decisions_tenant_event",
+            ondelete="RESTRICT",
+        ),
+        sa.ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id"],
+            ["expense_cases.tenant_id", "expense_cases.id"],
+            name="fk_ai_decisions_tenant_case",
+            ondelete="RESTRICT",
+        ),
+        sa.PrimaryKeyConstraint("id"),
+        sa.UniqueConstraint("tenant_id", "id", name="uq_ai_decisions_tenant_id"),
+        sa.UniqueConstraint(
+            "tenant_id",
+            "expense_case_id",
+            "id",
+            name="uq_ai_decisions_tenant_case_id",
+        ),
+        sa.UniqueConstraint(
+            "tenant_id",
+            "idempotency_key",
+            name="uq_ai_decisions_tenant_idempotency",
+        ),
+    )
+    op.create_index(
+        "ix_ai_decisions_tenant_case_time",
+        "ai_decisions",
+        ["tenant_id", "expense_case_id", "created_at"],
+    )
+    op.create_index(
+        "ix_ai_decisions_tenant_subject",
+        "ai_decisions",
+        ["tenant_id", "subject_type", "subject_id"],
+    )
+    op.create_index("ix_ai_decisions_expense_claim_id", "ai_decisions", ["expense_claim_id"])
+    op.create_index("ix_ai_decisions_agent_run_id", "ai_decisions", ["agent_run_id"])
+    op.create_index("ix_ai_decisions_correlation_id", "ai_decisions", ["correlation_id"])
+
+    op.create_table(
+        "ai_decision_feedback",
+        sa.Column("id", sa.String(length=36), nullable=False),
+        sa.Column("tenant_id", sa.String(length=64), nullable=False),
+        sa.Column("decision_id", sa.String(length=36), nullable=False),
+        sa.Column("expense_claim_id", sa.String(length=36), nullable=False),
+        sa.Column("correlation_id", sa.String(length=64), nullable=False),
+        sa.Column("feedback_type", sa.String(length=20), nullable=False),
+        sa.Column("action_type", sa.String(length=30), nullable=False),
+        sa.Column("actor_id", sa.String(length=120), nullable=False),
+        sa.Column("actor_type", sa.String(length=30), nullable=False),
+        sa.Column("evidence_source", sa.String(length=40), nullable=False),
+        sa.Column("verification_status", sa.String(length=30), nullable=False),
+        sa.Column(
+            "training_eligible",
+            sa.Boolean(),
+            server_default=sa.false(),
+            nullable=False,
+        ),
+        sa.Column("final_value_json", sa.JSON(), nullable=False),
+        sa.Column("changed_fields_json", sa.JSON(), nullable=False),
+        sa.Column("idempotency_key", sa.String(length=120), nullable=False),
+        sa.Column("content_fingerprint", sa.String(length=71), nullable=False),
+        sa.Column(
+            "created_at",
+            sa.DateTime(timezone=True),
+            server_default=sa.func.now(),
+            nullable=False,
+        ),
+        sa.CheckConstraint(
+            "feedback_type IN ('accepted', 'edited', 'rejected', 'ignored')",
+            name="ck_ai_decision_feedback_type",
+        ),
+        sa.CheckConstraint(
+            "verification_status IN ('client_observed', 'server_verified', "
+            "'human_verified', 'invalidated')",
+            name="ck_ai_decision_feedback_verification",
+        ),
+        sa.CheckConstraint(
+            "NOT training_eligible OR verification_status IN "
+            "('server_verified', 'human_verified')",
+            name="ck_ai_decision_feedback_training_eligibility",
+        ),
+        sa.ForeignKeyConstraint(
+            ["tenant_id", "decision_id"],
+            ["ai_decisions.tenant_id", "ai_decisions.id"],
+            name="fk_ai_decision_feedback_tenant_decision",
+            ondelete="RESTRICT",
+        ),
+        sa.PrimaryKeyConstraint("id"),
+        sa.UniqueConstraint(
+            "tenant_id",
+            "idempotency_key",
+            name="uq_ai_decision_feedback_tenant_idempotency",
+        ),
+    )
+    op.create_index(
+        "ix_ai_decision_feedback_tenant_decision_time",
+        "ai_decision_feedback",
+        ["tenant_id", "decision_id", "created_at"],
+    )
+    op.create_index(
+        "ix_ai_decision_feedback_tenant_type_time",
+        "ai_decision_feedback",
+        ["tenant_id", "feedback_type", "created_at"],
+    )
+    op.create_index(
+        "ix_ai_decision_feedback_expense_claim_id",
+        "ai_decision_feedback",
+        ["expense_claim_id"],
+    )
+    op.create_index(
+        "ix_ai_decision_feedback_correlation_id",
+        "ai_decision_feedback",
+        ["correlation_id"],
+    )
+
+    op.create_table(
+        "workflow_outcomes",
+        sa.Column("id", sa.String(length=36), nullable=False),
+        sa.Column("tenant_id", sa.String(length=64), nullable=False),
+        sa.Column("expense_case_id", sa.String(length=36), nullable=False),
+        sa.Column("decision_id", sa.String(length=36), nullable=False),
+        sa.Column("business_event_id", sa.String(length=36), nullable=True),
+        sa.Column("expense_claim_id", sa.String(length=36), nullable=False),
+        sa.Column("correlation_id", sa.String(length=64), nullable=False),
+        sa.Column("outcome_type", sa.String(length=50), nullable=False),
+        sa.Column("outcome_status", sa.String(length=20), nullable=False),
+        sa.Column("actor_id", sa.String(length=120), nullable=False),
+        sa.Column("actor_type", sa.String(length=30), nullable=False),
+        sa.Column("result_json", sa.JSON(), nullable=False),
+        sa.Column("idempotency_key", sa.String(length=120), nullable=False),
+        sa.Column("content_fingerprint", sa.String(length=71), nullable=False),
+        sa.Column(
+            "effective_at",
+            sa.DateTime(timezone=True),
+            server_default=sa.func.now(),
+            nullable=False,
+        ),
+        sa.Column(
+            "created_at",
+            sa.DateTime(timezone=True),
+            server_default=sa.func.now(),
+            nullable=False,
+        ),
+        sa.CheckConstraint(
+            "outcome_status IN ('recorded', 'verified', 'reversed')",
+            name="ck_workflow_outcomes_status",
+        ),
+        sa.ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id", "business_event_id"],
+            [
+                "business_events.tenant_id",
+                "business_events.expense_case_id",
+                "business_events.id",
+            ],
+            name="fk_workflow_outcomes_tenant_event",
+            ondelete="RESTRICT",
+        ),
+        sa.ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id", "decision_id"],
+            [
+                "ai_decisions.tenant_id",
+                "ai_decisions.expense_case_id",
+                "ai_decisions.id",
+            ],
+            name="fk_workflow_outcomes_tenant_decision",
+            ondelete="RESTRICT",
+        ),
+        sa.ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id"],
+            ["expense_cases.tenant_id", "expense_cases.id"],
+            name="fk_workflow_outcomes_tenant_case",
+            ondelete="RESTRICT",
+        ),
+        sa.PrimaryKeyConstraint("id"),
+        sa.UniqueConstraint(
+            "tenant_id",
+            "idempotency_key",
+            name="uq_workflow_outcomes_tenant_idempotency",
+        ),
+    )
+    op.create_index(
+        "ix_workflow_outcomes_tenant_case_time",
+        "workflow_outcomes",
+        ["tenant_id", "expense_case_id", "effective_at"],
+    )
+    op.create_index(
+        "ix_workflow_outcomes_tenant_type_time",
+        "workflow_outcomes",
+        ["tenant_id", "outcome_type", "effective_at"],
+    )
+    op.create_index(
+        "ix_workflow_outcomes_expense_claim_id",
+        "workflow_outcomes",
+        ["expense_claim_id"],
+    )
+    op.create_index(
+        "ix_workflow_outcomes_correlation_id",
+        "workflow_outcomes",
+        ["correlation_id"],
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("ix_workflow_outcomes_correlation_id", table_name="workflow_outcomes")
+    op.drop_index("ix_workflow_outcomes_expense_claim_id", table_name="workflow_outcomes")
+    op.drop_index("ix_workflow_outcomes_tenant_type_time", table_name="workflow_outcomes")
+    op.drop_index("ix_workflow_outcomes_tenant_case_time", table_name="workflow_outcomes")
+    op.drop_table("workflow_outcomes")
+
+    op.drop_index("ix_ai_decision_feedback_correlation_id", table_name="ai_decision_feedback")
+    op.drop_index("ix_ai_decision_feedback_expense_claim_id", table_name="ai_decision_feedback")
+    op.drop_index("ix_ai_decision_feedback_tenant_type_time", table_name="ai_decision_feedback")
+    op.drop_index("ix_ai_decision_feedback_tenant_decision_time", table_name="ai_decision_feedback")
+    op.drop_table("ai_decision_feedback")
+
+    op.drop_index("ix_ai_decisions_correlation_id", table_name="ai_decisions")
+    op.drop_index("ix_ai_decisions_agent_run_id", table_name="ai_decisions")
+    op.drop_index("ix_ai_decisions_expense_claim_id", table_name="ai_decisions")
+    op.drop_index("ix_ai_decisions_tenant_subject", table_name="ai_decisions")
+    op.drop_index("ix_ai_decisions_tenant_case_time", table_name="ai_decisions")
+    op.drop_table("ai_decisions")
+
+    op.drop_constraint(
+        "uq_business_events_tenant_case_id",
+        "business_events",
+        type_="unique",
+    )
+    op.drop_constraint("uq_expense_cases_tenant_id", "expense_cases", type_="unique")

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

@@ -163,6 +163,7 @@ def run_application_preview_action(
         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

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_learning import AIDecision, AIDecisionFeedback, WorkflowOutcome
 from app.models.approval import ApprovalRecord
 from app.models.audit_log import AuditLog
 from app.models.auth_session import AuthSession
@@ -52,6 +53,8 @@ __all__ = [
     "AgentRun",
     "AgentToolCall",
     "AgentTraceEvent",
+    "AIDecision",
+    "AIDecisionFeedback",
     "ApprovalRecord",
     "AuditLog",
     "AuthSession",
@@ -82,4 +85,5 @@ __all__ = [
     "SystemSetting",
     "SystemSettingSecret",
     "UserSessionMetric",
+    "WorkflowOutcome",
 ]

server/src/app/db/migration_preflight.py

@@ -27,8 +27,19 @@ MIGRATION_OWNED_TABLES_BY_REVISION: dict[str, frozenset[str]] = {
             "auth_sessions",
         }
     ),
+    "20260714_0003": frozenset(
+        {
+            "expense_cases",
+            "expense_case_links",
+            "business_events",
+            "auth_sessions",
+            "ai_decisions",
+            "ai_decision_feedback",
+            "workflow_outcomes",
+        }
+    ),
 }
-if MIGRATION_OWNED_TABLES_BY_REVISION["20260713_0002"] != MIGRATION_OWNED_TABLES:
+if MIGRATION_OWNED_TABLES_BY_REVISION["20260714_0003"] != MIGRATION_OWNED_TABLES:
     raise RuntimeError("latest Alembic revision must own the centralized migration table set")
 
 

server/src/app/db/schema_ownership.py

@@ -7,9 +7,12 @@ from app.db.base import Base
 MIGRATION_OWNED_TABLES: frozenset[str] = frozenset(
     {
         "auth_sessions",
+        "ai_decisions",
+        "ai_decision_feedback",
         "expense_cases",
         "expense_case_links",
         "business_events",
+        "workflow_outcomes",
     }
 )
 

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_learning import AIDecision, AIDecisionFeedback, WorkflowOutcome
 from app.models.approval import ApprovalRecord
 from app.models.audit_log import AuditLog
 from app.models.auth_session import AuthSession
@@ -51,6 +52,8 @@ __all__ = [
     "ApprovalRecord",
     "AuditLog",
     "AuthSession",
+    "AIDecision",
+    "AIDecisionFeedback",
     "BudgetAllocation",
     "BudgetReservation",
     "BudgetTransaction",
@@ -78,4 +81,5 @@ __all__ = [
     "SystemSetting",
     "SystemSettingSecret",
     "UserSessionMetric",
+    "WorkflowOutcome",
 ]

server/src/app/models/ai_learning.py

@@ -0,0 +1,260 @@
+from __future__ import annotations
+
+import uuid
+from datetime import datetime
+from decimal import Decimal
+from typing import Any
+
+from sqlalchemy import (
+    Boolean,
+    CheckConstraint,
+    DateTime,
+    ForeignKeyConstraint,
+    Index,
+    Integer,
+    Numeric,
+    String,
+    UniqueConstraint,
+    func,
+)
+from sqlalchemy.orm import Mapped, mapped_column
+from sqlalchemy.types import JSON
+
+from app.db.base_class import Base
+
+
+def _new_id() -> str:
+    return str(uuid.uuid4())
+
+
+class AIDecision(Base):
+    """AI 在费用流程中给出的结构化建议事实。"""
+
+    __tablename__ = "ai_decisions"
+    __table_args__ = (
+        UniqueConstraint("tenant_id", "id", name="uq_ai_decisions_tenant_id"),
+        UniqueConstraint(
+            "tenant_id",
+            "expense_case_id",
+            "id",
+            name="uq_ai_decisions_tenant_case_id",
+        ),
+        UniqueConstraint(
+            "tenant_id",
+            "idempotency_key",
+            name="uq_ai_decisions_tenant_idempotency",
+        ),
+        ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id"],
+            ["expense_cases.tenant_id", "expense_cases.id"],
+            ondelete="RESTRICT",
+            name="fk_ai_decisions_tenant_case",
+        ),
+        ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id", "business_event_id"],
+            [
+                "business_events.tenant_id",
+                "business_events.expense_case_id",
+                "business_events.id",
+            ],
+            ondelete="RESTRICT",
+            name="fk_ai_decisions_tenant_event",
+        ),
+        CheckConstraint(
+            "confidence IS NULL OR (confidence >= 0 AND confidence <= 1)",
+            name="ck_ai_decisions_confidence",
+        ),
+        CheckConstraint(
+            "status IN ('suggested', 'accepted', 'edited', 'rejected', "
+            "'ignored', 'executed', 'rolled_back')",
+            name="ck_ai_decisions_status",
+        ),
+        Index(
+            "ix_ai_decisions_tenant_case_time",
+            "tenant_id",
+            "expense_case_id",
+            "created_at",
+        ),
+        Index(
+            "ix_ai_decisions_tenant_subject",
+            "tenant_id",
+            "subject_type",
+            "subject_id",
+        ),
+        Index("ix_ai_decisions_agent_run_id", "agent_run_id"),
+        Index("ix_ai_decisions_correlation_id", "correlation_id"),
+    )
+
+    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=_new_id)
+    tenant_id: Mapped[str] = mapped_column(String(64), nullable=False)
+    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)
+    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)
+    subject_id: Mapped[str] = mapped_column(String(100), nullable=False)
+    decision_type: Mapped[str] = mapped_column(String(80), nullable=False)
+    decision_source: Mapped[str] = mapped_column(String(20), nullable=False)
+    status: Mapped[str] = mapped_column(String(20), nullable=False)
+    automation_mode: Mapped[str] = mapped_column(String(40), nullable=False)
+    confidence: Mapped[Decimal | None] = mapped_column(Numeric(5, 4), nullable=True)
+    suggestion_json: Mapped[dict[str, Any]] = mapped_column(JSON, nullable=False, default=dict)
+    evidence_json: Mapped[dict[str, Any]] = mapped_column(JSON, nullable=False, default=dict)
+    version_json: Mapped[dict[str, Any]] = mapped_column(JSON, nullable=False, default=dict)
+    schema_version: Mapped[int] = mapped_column(Integer, nullable=False, default=1)
+    training_eligible: Mapped[bool] = mapped_column(
+        Boolean, nullable=False, default=False, server_default="false"
+    )
+    idempotency_key: Mapped[str] = mapped_column(String(120), nullable=False)
+    content_fingerprint: Mapped[str] = mapped_column(String(71), nullable=False)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now()
+    )
+
+
+class AIDecisionFeedback(Base):
+    """用户对 AI 建议的采纳或显式字段纠正事实。"""
+
+    __tablename__ = "ai_decision_feedback"
+    __table_args__ = (
+        UniqueConstraint(
+            "tenant_id",
+            "idempotency_key",
+            name="uq_ai_decision_feedback_tenant_idempotency",
+        ),
+        ForeignKeyConstraint(
+            ["tenant_id", "decision_id"],
+            ["ai_decisions.tenant_id", "ai_decisions.id"],
+            ondelete="RESTRICT",
+            name="fk_ai_decision_feedback_tenant_decision",
+        ),
+        CheckConstraint(
+            "feedback_type IN ('accepted', 'edited', 'rejected', 'ignored')",
+            name="ck_ai_decision_feedback_type",
+        ),
+        CheckConstraint(
+            "verification_status IN ('client_observed', 'server_verified', "
+            "'human_verified', 'invalidated')",
+            name="ck_ai_decision_feedback_verification",
+        ),
+        CheckConstraint(
+            "NOT training_eligible OR verification_status IN "
+            "('server_verified', 'human_verified')",
+            name="ck_ai_decision_feedback_training_eligibility",
+        ),
+        Index(
+            "ix_ai_decision_feedback_tenant_decision_time",
+            "tenant_id",
+            "decision_id",
+            "created_at",
+        ),
+        Index(
+            "ix_ai_decision_feedback_tenant_type_time",
+            "tenant_id",
+            "feedback_type",
+            "created_at",
+        ),
+        Index("ix_ai_decision_feedback_correlation_id", "correlation_id"),
+    )
+
+    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=_new_id)
+    tenant_id: Mapped[str] = mapped_column(String(64), nullable=False)
+    decision_id: Mapped[str] = mapped_column(String(36), nullable=False)
+    expense_claim_id: Mapped[str] = mapped_column(String(36), nullable=False, index=True)
+    correlation_id: Mapped[str] = mapped_column(String(64), nullable=False)
+    feedback_type: Mapped[str] = mapped_column(String(20), nullable=False)
+    action_type: Mapped[str] = mapped_column(String(30), nullable=False)
+    actor_id: Mapped[str] = mapped_column(String(120), nullable=False)
+    actor_type: Mapped[str] = mapped_column(String(30), nullable=False, default="user")
+    evidence_source: Mapped[str] = mapped_column(String(40), nullable=False)
+    verification_status: Mapped[str] = mapped_column(String(30), nullable=False)
+    training_eligible: Mapped[bool] = mapped_column(
+        Boolean, nullable=False, default=False, server_default="false"
+    )
+    final_value_json: Mapped[dict[str, Any]] = mapped_column(JSON, nullable=False, default=dict)
+    changed_fields_json: Mapped[list[dict[str, Any]]] = mapped_column(
+        JSON, nullable=False, default=list
+    )
+    idempotency_key: Mapped[str] = mapped_column(String(120), nullable=False)
+    content_fingerprint: Mapped[str] = mapped_column(String(71), nullable=False)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now()
+    )
+
+
+class WorkflowOutcome(Base):
+    """与 AI 决策关联、但独立于技术执行状态的业务结果。"""
+
+    __tablename__ = "workflow_outcomes"
+    __table_args__ = (
+        UniqueConstraint(
+            "tenant_id",
+            "idempotency_key",
+            name="uq_workflow_outcomes_tenant_idempotency",
+        ),
+        ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id"],
+            ["expense_cases.tenant_id", "expense_cases.id"],
+            ondelete="RESTRICT",
+            name="fk_workflow_outcomes_tenant_case",
+        ),
+        ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id", "decision_id"],
+            [
+                "ai_decisions.tenant_id",
+                "ai_decisions.expense_case_id",
+                "ai_decisions.id",
+            ],
+            ondelete="RESTRICT",
+            name="fk_workflow_outcomes_tenant_decision",
+        ),
+        ForeignKeyConstraint(
+            ["tenant_id", "expense_case_id", "business_event_id"],
+            [
+                "business_events.tenant_id",
+                "business_events.expense_case_id",
+                "business_events.id",
+            ],
+            ondelete="RESTRICT",
+            name="fk_workflow_outcomes_tenant_event",
+        ),
+        CheckConstraint(
+            "outcome_status IN ('recorded', 'verified', 'reversed')",
+            name="ck_workflow_outcomes_status",
+        ),
+        Index(
+            "ix_workflow_outcomes_tenant_case_time",
+            "tenant_id",
+            "expense_case_id",
+            "effective_at",
+        ),
+        Index(
+            "ix_workflow_outcomes_tenant_type_time",
+            "tenant_id",
+            "outcome_type",
+            "effective_at",
+        ),
+        Index("ix_workflow_outcomes_correlation_id", "correlation_id"),
+    )
+
+    id: Mapped[str] = mapped_column(String(36), primary_key=True, default=_new_id)
+    tenant_id: Mapped[str] = mapped_column(String(64), nullable=False)
+    expense_case_id: Mapped[str] = mapped_column(String(36), nullable=False)
+    decision_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)
+    correlation_id: Mapped[str] = mapped_column(String(64), nullable=False)
+    outcome_type: Mapped[str] = mapped_column(String(50), nullable=False)
+    outcome_status: Mapped[str] = mapped_column(String(20), nullable=False)
+    actor_id: Mapped[str] = mapped_column(String(120), nullable=False)
+    actor_type: Mapped[str] = mapped_column(String(30), nullable=False, default="user")
+    result_json: Mapped[dict[str, Any]] = mapped_column(JSON, nullable=False, default=dict)
+    idempotency_key: Mapped[str] = mapped_column(String(120), nullable=False)
+    content_fingerprint: Mapped[str] = mapped_column(String(71), nullable=False)
+    effective_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now()
+    )
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now()
+    )

server/src/app/models/expense_case.py

@@ -18,6 +18,7 @@ def _new_id() -> str:
 class ExpenseCase(Base):
     __tablename__ = "expense_cases"
     __table_args__ = (
+        UniqueConstraint("tenant_id", "id", name="uq_expense_cases_tenant_id"),
         UniqueConstraint("tenant_id", "case_no", name="uq_expense_cases_tenant_case_no"),
         Index("ix_expense_cases_tenant_stage", "tenant_id", "current_stage"),
         Index("ix_expense_cases_tenant_status", "tenant_id", "status"),
@@ -73,6 +74,12 @@ class ExpenseCaseLink(Base):
 class BusinessEvent(Base):
     __tablename__ = "business_events"
     __table_args__ = (
+        UniqueConstraint(
+            "tenant_id",
+            "expense_case_id",
+            "id",
+            name="uq_business_events_tenant_case_id",
+        ),
         UniqueConstraint(
             "tenant_id",
             "aggregate_type",

server/src/app/services/expense_application_draft_events.py

@@ -9,7 +9,7 @@ from sqlalchemy import select
 from sqlalchemy.orm import Session
 
 from app.api.deps import CurrentUserContext
-from app.models.expense_case import BusinessEvent
+from app.models.expense_case import BusinessEvent, ExpenseCase
 from app.models.financial_record import ExpenseClaim
 from app.schemas.user_agent import UserAgentRequest
 from app.services.expense_cases import ExpenseCaseService
@@ -30,8 +30,8 @@ class ExpenseApplicationDraftEventService:
         event_type: str,
         previous_status: str = "",
         previous_approval_stage: str = "",
-    ) -> None:
+    ) -> tuple[ExpenseCase, BusinessEvent]:
-        ExpenseCaseService(self.db).record_claim_event(
+        return ExpenseCaseService(self.db).record_claim_event(
             claim,
             event_type=event_type,
             actor_id=current_user.username,

server/src/app/services/expense_application_learning.py

@@ -0,0 +1,382 @@
+from __future__ import annotations
+
+import hashlib
+import json
+import uuid
+from dataclasses import dataclass
+from decimal import Decimal
+from typing import Any
+
+from sqlalchemy import select
+from sqlalchemy.orm import Session
+
+from app.api.deps import CurrentUserContext
+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_cases import ExpenseCaseService
+
+
+@dataclass(frozen=True)
+class ExpenseApplicationLearningRecords:
+    decision: AIDecision
+    feedback: AIDecisionFeedback
+    outcome: WorkflowOutcome
+
+
+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
+
+    def record_action(
+        self,
+        payload: UserAgentRequest,
+        facts: dict[str, str],
+        claim: ExpenseClaim,
+        current_user: CurrentUserContext,
+        *,
+        action_type: str,
+        business_event: BusinessEvent,
+    ) -> ExpenseApplicationLearningRecords | None:
+        normalized_action = self._normalize_action_type(action_type)
+        preview = self._application_preview(payload)
+        if not self._is_learning_eligible_preview(preview):
+            return None
+        tenant_id = ExpenseCaseService.normalize_tenant_id(current_user.tenant_id)
+        expense_case = ExpenseCaseService(self.db).ensure_case_for_claim(
+            claim,
+            tenant_id=tenant_id,
+        )
+        if (
+            business_event.tenant_id != tenant_id
+            or business_event.expense_case_id != expense_case.id
+        ):
+            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"]
+
+        feedback_type = "edited" if changed_fields 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,
+        }
+        content_fingerprint = self._fingerprint(fingerprint_payload)
+        idempotency_key = self._idempotency_key(
+            normalized_action,
+            claim.id,
+            content_fingerprint,
+        )
+        existing = self._find_existing(tenant_id, idempotency_key)
+        if existing is not None:
+            return existing
+
+        decision_id = self._stable_id("decision", tenant_id, idempotency_key)
+        decision = AIDecision(
+            id=decision_id,
+            tenant_id=tenant_id,
+            expense_case_id=expense_case.id,
+            business_event_id=business_event.id,
+            expense_claim_id=claim.id,
+            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),
+            status="executed",
+            automation_mode="prefill",
+            confidence=self._confidence(payload),
+            suggestion_json=self._snapshot_reference(suggested_values),
+            evidence_json={
+                "evidence_source": "client_preview_edit_trace",
+                "trust_level": "behavioral_analytics_only",
+                "model_refined": bool(preview.get("modelRefined")),
+                "model_review_status": self._safe_text(preview.get("modelReviewStatus"), 40),
+            },
+            version_json={
+                "schema_version": 1,
+                "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),
+            },
+            schema_version=1,
+            training_eligible=False,
+            idempotency_key=idempotency_key,
+            content_fingerprint=content_fingerprint,
+        )
+        feedback = AIDecisionFeedback(
+            id=self._stable_id("feedback", tenant_id, idempotency_key),
+            tenant_id=tenant_id,
+            decision_id=decision_id,
+            expense_claim_id=claim.id,
+            correlation_id=correlation_id,
+            feedback_type=feedback_type,
+            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",
+            training_eligible=False,
+            final_value_json=self._snapshot_reference(final_values),
+            changed_fields_json=self._changed_field_references(changed_fields),
+            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,
+                }
+            ),
+        )
+        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,
+            expense_case_id=expense_case.id,
+            decision_id=decision_id,
+            business_event_id=business_event.id,
+            expense_claim_id=claim.id,
+            correlation_id=correlation_id,
+            outcome_type=outcome_type,
+            outcome_status="recorded",
+            actor_id=str(current_user.username or "anonymous").strip() or "anonymous",
+            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"
+                ),
+                "feedback_type": feedback_type,
+            },
+            idempotency_key=f"outcome:{idempotency_key}"[:120],
+            content_fingerprint=self._fingerprint(
+                {
+                    "decision_id": decision_id,
+                    "outcome_type": outcome_type,
+                    "feedback_type": feedback_type,
+                }
+            ),
+        )
+        self.db.add_all([decision, feedback, outcome])
+        self.db.flush()
+        return ExpenseApplicationLearningRecords(decision, feedback, outcome)
+
+    def _find_existing(
+        self,
+        tenant_id: str,
+        idempotency_key: str,
+    ) -> ExpenseApplicationLearningRecords | None:
+        decision = self.db.scalar(
+            select(AIDecision).where(
+                AIDecision.tenant_id == tenant_id,
+                AIDecision.idempotency_key == idempotency_key,
+            )
+        )
+        if decision is None:
+            return None
+        feedback = self.db.scalar(
+            select(AIDecisionFeedback).where(
+                AIDecisionFeedback.tenant_id == tenant_id,
+                AIDecisionFeedback.decision_id == decision.id,
+            )
+        )
+        outcome = self.db.scalar(
+            select(WorkflowOutcome).where(
+                WorkflowOutcome.tenant_id == tenant_id,
+                WorkflowOutcome.decision_id == decision.id,
+            )
+        )
+        if feedback is None or outcome is None:
+            raise RuntimeError("AI 学习账本幂等状态不完整，请重试。")
+        return ExpenseApplicationLearningRecords(decision, feedback, outcome)
+
+    @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)
+        }
+
+    @classmethod
+    def _safe_changed_fields(
+        cls,
+        payload: UserAgentRequest,
+        final_values: dict[str, str],
+    ) -> list[dict[str, str]]:
+        raw_feedback = cls._application_preview(payload).get("aiDecisionFeedback")
+        if not isinstance(raw_feedback, dict):
+            return []
+        raw_changes = raw_feedback.get("changedFields")
+        if not isinstance(raw_changes, list):
+            return []
+
+        changes: dict[str, dict[str, str]] = {}
+        for raw_item in raw_changes[:30]:
+            if not isinstance(raw_item, dict):
+                continue
+            alias = cls._safe_text(raw_item.get("fieldKey"), 60)
+            field_key = cls._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)
+            final_value = cls._safe_text(final_values.get(field_key), 500)
+            if suggested_value == final_value:
+                continue
+            changes[field_key] = {
+                "field_key": field_key,
+                "suggested_value": suggested_value,
+                "final_value": final_value,
+            }
+        return [changes[key] for key in sorted(changes)]
+
+    @staticmethod
+    def _application_preview(payload: UserAgentRequest) -> dict[str, Any]:
+        context = payload.context_json if isinstance(payload.context_json, dict) else {}
+        preview = context.get("application_preview")
+        return preview if isinstance(preview, dict) else {}
+
+    @staticmethod
+    def _is_learning_eligible_preview(preview: dict[str, Any]) -> bool:
+        fields = preview.get("fields")
+        if not isinstance(fields, dict) or not fields:
+            return False
+        if bool(preview.get("applicationEditMode") or preview.get("application_edit_mode")):
+            return False
+        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),
+        }
+
+    @classmethod
+    def _changed_field_references(
+        cls,
+        changes: list[dict[str, str]],
+    ) -> list[dict[str, str]]:
+        return [
+            {
+                "field_key": item["field_key"],
+                "suggested_value_fingerprint": cls._fingerprint(item["suggested_value"]),
+                "final_value_fingerprint": cls._fingerprint(item["final_value"]),
+            }
+            for item in changes
+        ]
+
+    @staticmethod
+    def _decision_source(preview: dict[str, Any], facts: dict[str, str]) -> str:
+        if preview.get("modelRefined") and facts.get("rule_version"):
+            return "hybrid"
+        if preview.get("modelRefined"):
+            return "model"
+        if facts.get("rule_version"):
+            return "rule"
+        return "heuristic"
+
+    @staticmethod
+    def _confidence(payload: UserAgentRequest) -> Decimal | None:
+        raw_value = getattr(payload.ontology, "confidence", None)
+        if raw_value is None:
+            return None
+        try:
+            value = Decimal(str(raw_value))
+        except Exception:
+            return None
+        return max(Decimal("0"), min(Decimal("1"), value)).quantize(Decimal("0.0001"))
+
+    @staticmethod
+    def _normalize_action_type(action_type: str) -> str:
+        normalized = str(action_type or "").strip().lower()
+        if normalized not in {"save_draft", "submit"}:
+            raise ValueError("不支持的 AI 申请动作。")
+        return normalized
+
+    @staticmethod
+    def _safe_text(value: object, limit: int) -> str:
+        return str(value or "").strip()[:limit]
+
+    @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()}"
+
+    @staticmethod
+    def _idempotency_key(action_type: str, claim_id: str, fingerprint: str) -> str:
+        digest = fingerprint.removeprefix("sha256:")
+        return f"ai-application:{action_type}:{claim_id}:{digest}"[:120]
+
+    @staticmethod
+    def _stable_id(record_type: str, tenant_id: str, idempotency_key: str) -> str:
+        return str(uuid.uuid5(uuid.NAMESPACE_URL, f"{record_type}:{tenant_id}:{idempotency_key}"))

server/src/app/services/expense_claims.py

@@ -4,6 +4,7 @@ import json
 import re
 import shutil
 import uuid
+from collections.abc import Callable
 from collections import defaultdict
 from datetime import UTC, date, datetime, timedelta
 from decimal import Decimal, InvalidOperation
@@ -20,6 +21,7 @@ from app.api.deps import CurrentUserContext
 from app.core.agent_enums import AgentAssetDomain, AgentAssetStatus, AgentAssetType
 from app.models.agent_asset import AgentAsset
 from app.models.employee import Employee
+from app.models.expense_case import BusinessEvent
 from app.models.financial_record import ExpenseClaim, ExpenseClaimItem
 from app.models.hermes_report import HermesRiskReport
 from app.models.risk_observation import RiskObservation, RiskObservationFeedback
@@ -563,7 +565,14 @@ class ExpenseClaimItemActionMixin:
             "item_id": item.id,
         }
 
-    def submit_claim(self, claim_id: str, current_user: CurrentUserContext) -> ExpenseClaim | None:
+    def submit_claim(
+        self,
+        claim_id: str,
+        current_user: CurrentUserContext,
+        *,
+        correlation_id: str | None = None,
+        before_commit: Callable[[BusinessEvent], None] | None = None,
+    ) -> ExpenseClaim | None:
         claim = self.get_claim(claim_id, current_user)
         if claim is None:
             return None
@@ -646,11 +655,12 @@ class ExpenseClaimItemActionMixin:
 
         claim.risk_flags_json = dedupe_claim_risk_flags(claim.risk_flags_json)
 
-        self._expense_cases.record_claim_event(
+        _, submission_event = self._expense_cases.record_claim_event(
             claim,
             event_type=("application_submitted" if is_application_claim else "claim_submitted"),
             actor_id=current_user.username,
             tenant_id=getattr(current_user, "tenant_id", None),
+            correlation_id=correlation_id,
             idempotency_key=(
                 f"submit:{claim.id}:{claim.submitted_at.isoformat()}"
                 if claim.submitted_at is not None
@@ -659,6 +669,8 @@ class ExpenseClaimItemActionMixin:
             previous_status=str(before_json.get("status") or ""),
             previous_approval_stage=str(before_json.get("approval_stage") or ""),
         )
+        if before_commit is not None:
+            before_commit(submission_event)
 
         self.db.commit()
         self.db.refresh(claim)

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.expense_case import BusinessEvent
 from app.models.financial_record import ExpenseClaim
 from app.schemas.reimbursement import TravelReimbursementCalculatorRequest
 from app.schemas.user_agent import (
@@ -28,6 +29,7 @@ from app.services.document_numbering import (
 from app.services.expense_application_draft_events import (
     ExpenseApplicationDraftEventService,
 )
+from app.services.expense_application_learning import ExpenseApplicationLearningService
 from app.services.expense_claim_access_policy import ExpenseClaimAccessPolicy
 from app.services.expense_claim_risk_stage import with_risk_business_stage
 from app.services.travel_reimbursement_calculator import TravelReimbursementCalculatorService
@@ -635,6 +637,27 @@ class UserAgentApplicationSlotMixin:
 
 
 class UserAgentApplicationPersistenceMixin:
+    def _record_expense_application_learning(
+        self,
+        payload: UserAgentRequest,
+        facts: dict[str, str],
+        claim: ExpenseClaim,
+        trusted_current_user: CurrentUserContext | None,
+        *,
+        action_type: str,
+        business_event: BusinessEvent,
+    ) -> None:
+        if trusted_current_user is None:
+            return
+        ExpenseApplicationLearningService(self.db).record_action(
+            payload,
+            facts,
+            claim,
+            trusted_current_user,
+            action_type=action_type,
+            business_event=business_event,
+        )
+
     @staticmethod
     def _resolve_application_edit_claim_id(context_json: dict[str, object]) -> str:
         if not isinstance(context_json, dict):
@@ -703,6 +726,7 @@ class UserAgentApplicationPersistenceMixin:
         claim: ExpenseClaim,
         *,
         submit: bool,
+        learning_current_user: CurrentUserContext | None = None,
     ) -> ExpenseClaim:
         current_user = self._build_application_current_user(payload)
         previous_status = str(claim.status or "").strip()
@@ -732,7 +756,7 @@ class UserAgentApplicationPersistenceMixin:
             claim.approval_stage = "待提交"
             claim.submitted_at = None
             try:
-                ExpenseApplicationDraftEventService(self.db).record(
+                _, draft_event = ExpenseApplicationDraftEventService(self.db).record(
                     payload,
                     claim,
                     current_user,
@@ -740,6 +764,14 @@ class UserAgentApplicationPersistenceMixin:
                     previous_status=previous_status,
                     previous_approval_stage=previous_approval_stage,
                 )
+                self._record_expense_application_learning(
+                    payload,
+                    facts,
+                    claim,
+                    learning_current_user,
+                    action_type="save_draft",
+                    business_event=draft_event,
+                )
                 self.db.commit()
                 self.db.refresh(claim)
                 return claim
@@ -749,10 +781,28 @@ class UserAgentApplicationPersistenceMixin:
 
         from app.services.expense_claims import ExpenseClaimService
 
-        submitted = ExpenseClaimService(self.db).submit_claim(claim.id, current_user)
+        try:
-        if submitted is None:
+            submitted = ExpenseClaimService(self.db).submit_claim(
-            raise ValueError("未找到可修改的申请单。")
+                claim.id,
-        return submitted
+                current_user,
+                correlation_id=payload.run_id,
+                before_commit=(
+                    lambda event: self._record_expense_application_learning(
+                        payload,
+                        facts,
+                        claim,
+                        learning_current_user,
+                        action_type="submit",
+                        business_event=event,
+                    )
+                ),
+            )
+            if submitted is None:
+                raise ValueError("未找到可修改的申请单。")
+            return submitted
+        except Exception:
+            self.db.rollback()
+            raise
 
     def _create_expense_application_record(
         self,
@@ -760,6 +810,7 @@ class UserAgentApplicationPersistenceMixin:
         facts: dict[str, str],
         *,
         submit: bool,
+        learning_current_user: CurrentUserContext | None = None,
     ) -> ExpenseClaim:
         current_user = self._build_application_current_user(payload)
         access_policy = ExpenseClaimAccessPolicy(self.db)
@@ -821,19 +872,41 @@ class UserAgentApplicationPersistenceMixin:
                         return existing
                 raise
             if not submit:
-                draft_event_service.record(
+                _, draft_event = draft_event_service.record(
                     payload,
                     claim,
                     current_user,
                     event_type="claim_draft_created",
                 )
+                self._record_expense_application_learning(
+                    payload,
+                    facts,
+                    claim,
+                    learning_current_user,
+                    action_type="save_draft",
+                    business_event=draft_event,
+                )
                 self.db.commit()
                 self.db.refresh(claim)
                 return claim
 
             from app.services.expense_claims import ExpenseClaimService
 
-            submitted = ExpenseClaimService(self.db).submit_claim(claim.id, current_user)
+            submitted = ExpenseClaimService(self.db).submit_claim(
+                claim.id,
+                current_user,
+                correlation_id=payload.run_id,
+                before_commit=(
+                    lambda event: self._record_expense_application_learning(
+                        payload,
+                        facts,
+                        claim,
+                        learning_current_user,
+                        action_type="submit",
+                        business_event=event,
+                    )
+                ),
+            )
             if submitted is None:
                 raise ValueError("未找到可提交的申请单。")
             return submitted
@@ -1212,6 +1285,7 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
         payload: UserAgentRequest,
         *,
         risk_flags: list[str],
+        learning_current_user: CurrentUserContext | None = None,
     ) -> UserAgentResponse:
         facts = self._resolve_expense_application_facts(payload)
         step = self._resolve_expense_application_step(payload, facts)
@@ -1224,6 +1298,7 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
                     facts,
                     editable_claim,
                     submit=step == "submitted",
+                    learning_current_user=learning_current_user,
                 )
                 facts["application_edit_mode"] = "true"
             elif step == "submitted":
@@ -1236,12 +1311,14 @@ class UserAgentApplicationMixin(UserAgentApplicationSlotMixin, UserAgentApplicat
                         payload,
                         facts,
                         submit=True,
+                        learning_current_user=learning_current_user,
                     )
             else:
                 application_claim = self._create_expense_application_record(
                     payload,
                     facts,
                     submit=False,
+                    learning_current_user=learning_current_user,
                 )
             if application_claim is not None:
                 facts["application_no"] = application_claim.claim_no

server/tests/test_alembic_migrations.py

@@ -9,19 +9,21 @@ import pytest
 from alembic.config import Config
 from sqlalchemy import create_engine, inspect, text
 from sqlalchemy.engine import Engine, make_url
+from sqlalchemy.exc import IntegrityError
 from sqlalchemy.pool import NullPool
 
 from alembic import command
 from app.core.config import get_settings
 from app.db.migration_preflight import MigrationPreflightError, validate_migration_state
-from app.db.schema_ownership import MIGRATION_OWNED_TABLES
+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 = "20260713_0002"
+HEAD_REVISION = "20260714_0003"
 SERVER_DIR = Path(__file__).resolve().parents[1]
 ALEMBIC_INI_PATH = SERVER_DIR / "alembic.ini"
 
+
 def _normalize_probe_component(value: str) -> str:
     return re.sub(r"[^a-z0-9]+", "-", value.lower()).strip("-")
 
@@ -143,6 +145,25 @@ def _assert_cascade_foreign_key(engine: Engine, table_name: str) -> None:
     assert str(matching[0].get("options", {}).get("ondelete", "")).upper() == "CASCADE"
 
 
+def _assert_composite_foreign_key(
+    engine: Engine,
+    table_name: str,
+    constrained_columns: tuple[str, ...],
+    referred_table: str,
+    referred_columns: tuple[str, ...] = ("tenant_id", "id"),
+) -> None:
+    foreign_keys = inspect(engine).get_foreign_keys(table_name, schema="public")
+    matching = [
+        item
+        for item in foreign_keys
+        if tuple(item["constrained_columns"]) == constrained_columns
+        and item["referred_table"] == referred_table
+        and tuple(item["referred_columns"]) == referred_columns
+    ]
+    assert len(matching) == 1
+    assert str(matching[0].get("options", {}).get("ondelete", "")).upper() == "RESTRICT"
+
+
 def _assert_head_schema(engine: Engine) -> None:
     names = _table_names(engine)
     assert MIGRATION_OWNED_TABLES.issubset(names)
@@ -163,6 +184,12 @@ def _assert_head_schema(engine: Engine) -> None:
         "uq_expense_case_links_resource",
         ("resource_type", "resource_id"),
     )
+    _assert_unique_constraint(
+        engine,
+        "business_events",
+        "uq_business_events_tenant_case_id",
+        ("tenant_id", "expense_case_id", "id"),
+    )
     _assert_unique_constraint(
         engine,
         "business_events",
@@ -175,6 +202,30 @@ def _assert_head_schema(engine: Engine) -> None:
         "uq_auth_sessions_token_hash",
         ("token_hash",),
     )
+    _assert_unique_constraint(
+        engine,
+        "ai_decisions",
+        "uq_ai_decisions_tenant_case_id",
+        ("tenant_id", "expense_case_id", "id"),
+    )
+    _assert_unique_constraint(
+        engine,
+        "ai_decisions",
+        "uq_ai_decisions_tenant_idempotency",
+        ("tenant_id", "idempotency_key"),
+    )
+    _assert_unique_constraint(
+        engine,
+        "ai_decision_feedback",
+        "uq_ai_decision_feedback_tenant_idempotency",
+        ("tenant_id", "idempotency_key"),
+    )
+    _assert_unique_constraint(
+        engine,
+        "workflow_outcomes",
+        "uq_workflow_outcomes_tenant_idempotency",
+        ("tenant_id", "idempotency_key"),
+    )
 
     _assert_indexes(
         engine,
@@ -214,8 +265,85 @@ def _assert_head_schema(engine: Engine) -> None:
             "ix_auth_sessions_tenant_username": ("tenant_id", "username"),
         },
     )
+    _assert_indexes(
+        engine,
+        "ai_decisions",
+        {
+            "ix_ai_decisions_tenant_case_time": (
+                "tenant_id",
+                "expense_case_id",
+                "created_at",
+            ),
+            "ix_ai_decisions_tenant_subject": (
+                "tenant_id",
+                "subject_type",
+                "subject_id",
+            ),
+        },
+    )
+    _assert_indexes(
+        engine,
+        "ai_decision_feedback",
+        {
+            "ix_ai_decision_feedback_tenant_decision_time": (
+                "tenant_id",
+                "decision_id",
+                "created_at",
+            ),
+        },
+    )
+    _assert_indexes(
+        engine,
+        "workflow_outcomes",
+        {
+            "ix_workflow_outcomes_tenant_case_time": (
+                "tenant_id",
+                "expense_case_id",
+                "effective_at",
+            ),
+        },
+    )
     _assert_cascade_foreign_key(engine, "expense_case_links")
     _assert_cascade_foreign_key(engine, "business_events")
+    _assert_composite_foreign_key(
+        engine,
+        "ai_decisions",
+        ("tenant_id", "expense_case_id"),
+        "expense_cases",
+    )
+    _assert_composite_foreign_key(
+        engine,
+        "ai_decisions",
+        ("tenant_id", "expense_case_id", "business_event_id"),
+        "business_events",
+        ("tenant_id", "expense_case_id", "id"),
+    )
+    _assert_composite_foreign_key(
+        engine,
+        "ai_decision_feedback",
+        ("tenant_id", "decision_id"),
+        "ai_decisions",
+    )
+    _assert_composite_foreign_key(
+        engine,
+        "workflow_outcomes",
+        ("tenant_id", "expense_case_id"),
+        "expense_cases",
+    )
+    _assert_composite_foreign_key(
+        engine,
+        "workflow_outcomes",
+        ("tenant_id", "expense_case_id", "decision_id"),
+        "ai_decisions",
+        ("tenant_id", "expense_case_id", "id"),
+    )
+    _assert_composite_foreign_key(
+        engine,
+        "workflow_outcomes",
+        ("tenant_id", "expense_case_id", "business_event_id"),
+        "business_events",
+        ("tenant_id", "expense_case_id", "id"),
+    )
 
 
 def _assert_runtime_cascade(engine: Engine) -> None:
@@ -271,6 +399,194 @@ def _assert_runtime_cascade(engine: Engine) -> None:
         ) == 0
 
 
+def _assert_learning_ledger_tenant_boundary(engine: Engine) -> None:
+    with engine.begin() as connection:
+        connection.execute(
+            text(
+                """
+                INSERT INTO expense_cases (
+                    id, tenant_id, case_no, scene_code, title, current_stage, status
+                ) VALUES
+                (
+                    'learning-probe-case', 'learning-probe', 'CASE-LEARNING-PROBE',
+                    'travel', '学习闭环迁移验证', 'claiming', 'active'
+                ),
+                (
+                    'learning-probe-case-b', 'learning-probe', 'CASE-LEARNING-PROBE-B',
+                    'travel', '学习闭环同租户第二费用单', 'claiming', 'active'
+                )
+                """
+            )
+        )
+        connection.execute(
+            text(
+                """
+                INSERT INTO business_events (
+                    id, tenant_id, expense_case_id, aggregate_type, aggregate_id,
+                    event_type, event_version, idempotency_key, correlation_id,
+                    actor_id, actor_type, payload_json, delivery_status, delivery_attempts
+                ) VALUES (
+                    'learning-probe-event', 'learning-probe', 'learning-probe-case',
+                    'expense_claim', 'learning-probe-claim', 'claim_draft_created', 1,
+                    'learning-probe-event-key', 'learning-probe-correlation',
+                    'learning-probe-user', 'user', '{}', 'pending', 0
+                )
+                """
+            )
+        )
+        connection.execute(
+            text(
+                """
+                INSERT INTO ai_decisions (
+                    id, tenant_id, expense_case_id, business_event_id,
+                    expense_claim_id, agent_run_id, correlation_id, subject_type,
+                    subject_id, decision_type, decision_source, status,
+                    automation_mode, confidence, suggestion_json, evidence_json,
+                    version_json, schema_version, idempotency_key, content_fingerprint
+                ) VALUES (
+                    'learning-probe-decision', 'learning-probe', 'learning-probe-case',
+                    'learning-probe-event', 'learning-probe-claim', NULL,
+                    'learning-probe-correlation', 'expense_claim', 'learning-probe-claim',
+                    'expense_application_prefill', 'hybrid', 'executed', 'prefill', 0.9,
+                    '{}', '{}', '{}', 1, 'learning-probe-decision-key',
+                    'sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa'
+                )
+                """
+            )
+        )
+        connection.execute(
+            text(
+                """
+                INSERT INTO ai_decision_feedback (
+                    id, tenant_id, decision_id, expense_claim_id, correlation_id,
+                    feedback_type, action_type, actor_id, actor_type, evidence_source,
+                    verification_status, final_value_json, changed_fields_json, idempotency_key,
+                    content_fingerprint
+                ) VALUES (
+                    'learning-probe-feedback', 'learning-probe', 'learning-probe-decision',
+                    'learning-probe-claim', 'learning-probe-correlation', 'accepted',
+                    'save_draft', 'learning-probe-user', 'user',
+                    'client_action_confirmation', 'client_observed', '{}', '[]',
+                    'learning-probe-feedback-key',
+                    'sha256:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb'
+                )
+                """
+            )
+        )
+        connection.execute(
+            text(
+                """
+                INSERT INTO workflow_outcomes (
+                    id, tenant_id, expense_case_id, decision_id, business_event_id,
+                    expense_claim_id, correlation_id, outcome_type, outcome_status,
+                    actor_id, actor_type, result_json, idempotency_key,
+                    content_fingerprint
+                ) VALUES (
+                    'learning-probe-outcome', 'learning-probe', 'learning-probe-case',
+                    'learning-probe-decision', 'learning-probe-event',
+                    'learning-probe-claim', 'learning-probe-correlation', 'draft_saved',
+                    'recorded', 'learning-probe-user', 'user', '{}',
+                    'learning-probe-outcome-key',
+                    'sha256:cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc'
+                )
+                """
+            )
+        )
+
+    with pytest.raises(IntegrityError):
+        with engine.begin() as connection:
+            connection.execute(
+                text(
+                    """
+                    INSERT INTO ai_decisions (
+                        id, tenant_id, expense_case_id, expense_claim_id,
+                        correlation_id, subject_type, subject_id, decision_type,
+                        decision_source, status, automation_mode, suggestion_json,
+                        evidence_json, version_json, schema_version, idempotency_key,
+                        content_fingerprint
+                    ) VALUES (
+                        'cross-tenant-decision', 'other-tenant', 'learning-probe-case',
+                        'cross-tenant-claim', 'cross-tenant-correlation', 'expense_claim',
+                        'cross-tenant-claim', 'expense_application_prefill', 'heuristic',
+                        'executed', 'prefill', '{}', '{}', '{}', 1,
+                        'cross-tenant-key',
+                        'sha256:dddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddd'
+                    )
+                    """
+                )
+            )
+
+    with pytest.raises(IntegrityError):
+        with engine.begin() as connection:
+            connection.execute(
+                text(
+                    """
+                    INSERT INTO ai_decision_feedback (
+                        id, tenant_id, decision_id, expense_claim_id, correlation_id,
+                        feedback_type, action_type, actor_id, actor_type, evidence_source,
+                        verification_status, training_eligible, final_value_json,
+                        changed_fields_json, idempotency_key, content_fingerprint
+                    ) VALUES (
+                        'unverified-training-feedback', 'learning-probe',
+                        'learning-probe-decision', 'learning-probe-claim',
+                        'unverified-training-correlation', 'accepted', 'save_draft',
+                        'learning-probe-user', 'user', 'client_action_confirmation',
+                        'client_observed', TRUE, '{}', '[]',
+                        'unverified-training-feedback-key',
+                        'sha256:9999999999999999999999999999999999999999999999999999999999999999'
+                    )
+                    """
+                )
+            )
+
+    with pytest.raises(IntegrityError):
+        with engine.begin() as connection:
+            connection.execute(
+                text(
+                    """
+                    INSERT INTO ai_decisions (
+                        id, tenant_id, expense_case_id, business_event_id,
+                        expense_claim_id, correlation_id, subject_type, subject_id,
+                        decision_type, decision_source, status, automation_mode,
+                        suggestion_json, evidence_json, version_json, schema_version,
+                        idempotency_key, content_fingerprint
+                    ) VALUES (
+                        'cross-case-event-decision', 'learning-probe',
+                        'learning-probe-case-b', 'learning-probe-event',
+                        'cross-case-event-claim', 'cross-case-event-correlation',
+                        'expense_claim', 'cross-case-event-claim',
+                        'expense_application_prefill', 'heuristic', 'executed',
+                        'prefill', '{}', '{}', '{}', 1,
+                        'cross-case-event-key',
+                        'sha256:eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee'
+                    )
+                    """
+                )
+            )
+
+    with pytest.raises(IntegrityError):
+        with engine.begin() as connection:
+            connection.execute(
+                text(
+                    """
+                    INSERT INTO workflow_outcomes (
+                        id, tenant_id, expense_case_id, decision_id,
+                        expense_claim_id, correlation_id, outcome_type, outcome_status,
+                        actor_id, actor_type, result_json, idempotency_key,
+                        content_fingerprint
+                    ) VALUES (
+                        'cross-case-decision-outcome', 'learning-probe',
+                        'learning-probe-case-b', 'learning-probe-decision',
+                        'cross-case-decision-claim', 'cross-case-decision-correlation',
+                        'draft_saved', 'recorded', 'learning-probe-user', 'user', '{}',
+                        'cross-case-decision-key',
+                        'sha256:ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff'
+                    )
+                    """
+                )
+            )
+
+
 def _create_legacy_sentinel(engine: Engine) -> None:
     with engine.begin() as connection:
         connection.execute(
@@ -337,9 +653,12 @@ def test_alembic_migration_cycle_on_disposable_postgres(
         _upgrade_head(migration_database_url)
         _assert_head_schema(engine)
         assert validate_migration_state(engine).revision == HEAD_REVISION
+        create_legacy_schema(engine)
+        assert "expense_claims" in _table_names(engine)
 
         _upgrade_head(migration_database_url)
         _assert_head_schema(engine)
+        _assert_learning_ledger_tenant_boundary(engine)
         _assert_runtime_cascade(engine)
 
         _create_legacy_sentinel(engine)

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["20260713_0002"]),
+    sorted(MIGRATION_OWNED_TABLES_BY_REVISION["20260714_0003"]),
 )
 def test_unversioned_database_with_any_migration_owned_table_is_rejected(
     engine: Engine,
@@ -87,6 +87,10 @@ def test_known_revision_requires_and_accepts_its_exact_owned_table_set(
             "20260713_0002",
             MIGRATION_OWNED_TABLES_BY_REVISION["20260713_0002"] - {"auth_sessions"},
         ),
+        (
+            "20260714_0003",
+            MIGRATION_OWNED_TABLES_BY_REVISION["20260714_0003"] - {"ai_decisions"},
+        ),
     ],
 )
 def test_known_revision_with_missing_or_unexpected_owned_tables_is_rejected(

server/tests/test_reimbursement_endpoints.py

@@ -16,6 +16,7 @@ from sqlalchemy.pool import StaticPool
 from app.api.deps import get_db
 from app.db.base import Base
 from app.main import create_app
+from app.models.ai_learning import AIDecision, AIDecisionFeedback, WorkflowOutcome
 from app.models.budget import BudgetAllocation, BudgetReservation, BudgetTransaction
 from app.models.employee import Employee
 from app.models.expense_case import BusinessEvent, ExpenseCaseLink
@@ -982,6 +983,23 @@ def test_application_preview_action_submits_without_orchestrator_run(monkeypatch
         assert event.expense_case_id == case_link.expense_case_id
         assert event.payload_json["previous_status"] == "draft"
         assert event.payload_json["next_status"] == "submitted"
+        decision = db.scalar(select(AIDecision).where(AIDecision.expense_claim_id == claim.id))
+        feedback = db.scalar(
+            select(AIDecisionFeedback).where(AIDecisionFeedback.expense_claim_id == claim.id)
+        )
+        outcome = db.scalar(
+            select(WorkflowOutcome).where(WorkflowOutcome.expense_claim_id == claim.id)
+        )
+        assert decision is not None
+        assert decision.expense_case_id == case_link.expense_case_id
+        assert decision.correlation_id == event.correlation_id
+        assert decision.business_event_id == event.id
+        assert feedback is not None
+        assert feedback.feedback_type == "accepted"
+        assert feedback.action_type == "submit"
+        assert outcome is not None
+        assert outcome.outcome_type == "application_submitted"
+        assert outcome.business_event_id == event.id
         assert any(
             isinstance(flag, dict)
             and flag.get("event_type") == "expense_application_submission"
@@ -1033,6 +1051,9 @@ def test_application_direct_submit_rolls_back_budget_when_case_event_fails(
         assert list(db.scalars(select(BudgetTransaction)).all()) == []
         assert list(db.scalars(select(ExpenseCaseLink)).all()) == []
         assert list(db.scalars(select(BusinessEvent)).all()) == []
+        assert list(db.scalars(select(AIDecision)).all()) == []
+        assert list(db.scalars(select(AIDecisionFeedback)).all()) == []
+        assert list(db.scalars(select(WorkflowOutcome)).all()) == []
 
 
 def test_application_preview_action_saves_draft_with_detail_reference(monkeypatch, tmp_path) -> None:

server/tests/test_schema_ownership.py

@@ -16,10 +16,13 @@ def test_create_legacy_schema_never_creates_migration_owned_tables() -> None:
 
     assert MIGRATION_OWNED_TABLES == frozenset(
         {
+            "ai_decision_feedback",
+            "ai_decisions",
             "auth_sessions",
             "business_events",
             "expense_case_links",
             "expense_cases",
+            "workflow_outcomes",
         }
     )
     assert table_names

server/tests/test_user_agent_application_draft_events.py

@@ -9,6 +9,7 @@ from sqlalchemy.orm import Session, sessionmaker
 from sqlalchemy.pool import StaticPool
 
 from app.db.base import Base
+from app.models.ai_learning import AIDecision, AIDecisionFeedback, WorkflowOutcome
 from app.models.employee import Employee
 from app.models.expense_case import BusinessEvent, ExpenseCaseLink
 from app.models.financial_record import ExpenseClaim
@@ -28,7 +29,12 @@ def build_session_factory() -> sessionmaker[Session]:
     return sessionmaker(bind=engine, autoflush=False, autocommit=False)
 
 
-def build_request(*, run_id: str, tenant_id: str = "tenant-a") -> UserAgentRequest:
+def build_request(
+    *,
+    run_id: str,
+    tenant_id: str = "tenant-a",
+    changed_fields: list[dict[str, str]] | None = None,
+) -> UserAgentRequest:
     return UserAgentRequest(
         run_id=run_id,
         user_id="owner@example.com",
@@ -40,6 +46,20 @@ def build_request(*, run_id: str, tenant_id: str = "tenant-a") -> UserAgentReque
             "name": "张三",
             "employee_no": "E001",
             "department_name": "市场部",
+            "application_preview": {
+                "modelReviewStatus": "completed",
+                "fields": {
+                    "applicationType": "差旅费用申请",
+                    "time": "2026-07-13 至 2026-07-14",
+                    "location": "上海",
+                    "reason": "客户现场差旅",
+                    "amount": "880元",
+                },
+                "aiDecisionFeedback": {
+                    "schemaVersion": 1,
+                    "changedFields": changed_fields or [],
+                },
+            },
         },
     )
 
@@ -84,11 +104,14 @@ def test_ai_application_draft_update_writes_tenant_scoped_event() -> None:
         db.add_all([owner, claim])
         db.commit()
 
-        updated = UserAgentService(db)._update_expense_application_record(
+        service = UserAgentService(db)
-            build_request(run_id="application-draft-update"),
+        request = build_request(run_id="application-draft-update")
+        updated = service._update_expense_application_record(
+            request,
             build_facts(reason="更新后的申请事由"),
             claim,
             submit=False,
+            learning_current_user=service._build_application_current_user(request),
         )
 
         assert updated.status == "draft"
@@ -103,6 +126,65 @@ def test_ai_application_draft_update_writes_tenant_scoped_event() -> None:
         link = db.scalar(select(ExpenseCaseLink).where(ExpenseCaseLink.resource_id == claim.id))
         assert link is not None
         assert link.tenant_id == "tenant-a"
+        decision = db.scalar(select(AIDecision).where(AIDecision.expense_claim_id == claim.id))
+        feedback = db.scalar(
+            select(AIDecisionFeedback).where(AIDecisionFeedback.expense_claim_id == claim.id)
+        )
+        outcome = db.scalar(
+            select(WorkflowOutcome).where(WorkflowOutcome.expense_claim_id == claim.id)
+        )
+        assert decision is not None
+        assert decision.business_event_id == event.id
+        assert feedback is not None
+        assert feedback.feedback_type == "accepted"
+        assert outcome is not None
+        assert outcome.outcome_type == "draft_saved"
+
+
+def test_untrusted_application_path_does_not_write_learning_ledger() -> None:
+    session_factory = build_session_factory()
+    with session_factory() as db:
+        claim = UserAgentService(db)._create_expense_application_record(
+            build_request(run_id="application-untrusted-path"),
+            build_facts(),
+            submit=False,
+        )
+
+        assert db.get(ExpenseClaim, claim.id) is not None
+        assert db.scalar(
+            select(BusinessEvent).where(BusinessEvent.aggregate_id == claim.id)
+        ) is not None
+        assert list(db.scalars(select(AIDecision)).all()) == []
+        assert list(db.scalars(select(AIDecisionFeedback)).all()) == []
+        assert list(db.scalars(select(WorkflowOutcome)).all()) == []
+
+
+@pytest.mark.parametrize(
+    "preview_patch",
+    [
+        {"modelReviewStatus": "template"},
+        {"applicationEditMode": True},
+    ],
+)
+def test_non_ai_or_detail_preview_is_not_recorded_as_ai_feedback(
+    preview_patch: dict[str, object],
+) -> None:
+    session_factory = build_session_factory()
+    with session_factory() as db:
+        request = build_request(run_id="application-ineligible-preview")
+        request.context_json["application_preview"].update(preview_patch)
+        service = UserAgentService(db)
+        claim = service._create_expense_application_record(
+            request,
+            build_facts(),
+            submit=False,
+            learning_current_user=service._build_application_current_user(request),
+        )
+
+        assert db.get(ExpenseClaim, claim.id) is not None
+        assert list(db.scalars(select(AIDecision)).all()) == []
+        assert list(db.scalars(select(AIDecisionFeedback)).all()) == []
+        assert list(db.scalars(select(WorkflowOutcome)).all()) == []
 
 
 def test_ai_application_draft_creation_rolls_back_when_event_write_fails(
@@ -217,8 +299,21 @@ def test_ai_application_draft_update_deduplicates_identical_snapshot() -> None:
         request = build_request(run_id="application-draft-idempotent")
         facts = build_facts(reason="同一份草稿")
         service = UserAgentService(db)
-        service._update_expense_application_record(request, facts, claim, submit=False)
+        learning_current_user = service._build_application_current_user(request)
-        service._update_expense_application_record(request, facts, claim, submit=False)
+        service._update_expense_application_record(
+            request,
+            facts,
+            claim,
+            submit=False,
+            learning_current_user=learning_current_user,
+        )
+        service._update_expense_application_record(
+            request,
+            facts,
+            claim,
+            submit=False,
+            learning_current_user=learning_current_user,
+        )
 
         events = list(
             db.scalars(
@@ -228,6 +323,9 @@ def test_ai_application_draft_update_deduplicates_identical_snapshot() -> None:
         assert len(events) == 1
         assert events[0].event_type == "claim_draft_updated"
         assert events[0].payload_json["previous_status"] == "returned"
+        assert len(list(db.scalars(select(AIDecision)).all())) == 1
+        assert len(list(db.scalars(select(AIDecisionFeedback)).all())) == 1
+        assert len(list(db.scalars(select(WorkflowOutcome)).all())) == 1
 
 
 def test_ai_application_draft_update_keeps_distinct_snapshots_in_same_run() -> None:
@@ -262,17 +360,20 @@ def test_ai_application_draft_update_keeps_distinct_snapshots_in_same_run() -> N
 
         request = build_request(run_id="application-draft-multi-version")
         service = UserAgentService(db)
+        learning_current_user = service._build_application_current_user(request)
         service._update_expense_application_record(
             request,
             build_facts(reason="第一版草稿"),
             claim,
             submit=False,
+            learning_current_user=learning_current_user,
         )
         service._update_expense_application_record(
             request,
             build_facts(reason="第二版草稿"),
             claim,
             submit=False,
+            learning_current_user=learning_current_user,
         )
 
         events = list(
@@ -283,3 +384,71 @@ def test_ai_application_draft_update_keeps_distinct_snapshots_in_same_run() -> N
         assert len(events) == 2
         assert {event.event_type for event in events} == {"claim_draft_updated"}
         assert len({event.idempotency_key for event in events}) == 2
+        assert len(list(db.scalars(select(AIDecision)).all())) == 2
+
+
+def test_ai_application_draft_records_explicit_field_correction() -> None:
+    session_factory = build_session_factory()
+    with session_factory() as db:
+        owner = Employee(
+            id="owner-1",
+            employee_no="E001",
+            name="张三",
+            email="owner@example.com",
+        )
+        claim = ExpenseClaim(
+            id="application-1",
+            claim_no="AP-DRAFT-FEEDBACK-001",
+            employee_id=owner.id,
+            employee_name=owner.name,
+            department_name="市场部",
+            expense_type="travel_application",
+            reason="原申请事由",
+            location="北京",
+            amount=Decimal("500.00"),
+            currency="CNY",
+            invoice_count=0,
+            occurred_at=datetime(2026, 7, 10, tzinfo=UTC),
+            submitted_at=None,
+            status="returned",
+            approval_stage="退回补充",
+            risk_flags_json=[],
+        )
+        db.add_all([owner, claim])
+        db.commit()
+
+        service = UserAgentService(db)
+        request = build_request(
+            run_id="application-draft-feedback",
+            changed_fields=[
+                {
+                    "fieldKey": "reason",
+                    "suggestedValue": "客户拜访",
+                    "finalValue": "客户端伪造值不会直接入账",
+                }
+            ],
+        )
+        service._update_expense_application_record(
+            request,
+            build_facts(reason="客户现场实施"),
+            claim,
+            submit=False,
+            learning_current_user=service._build_application_current_user(request),
+        )
+
+        decision = db.scalar(select(AIDecision))
+        feedback = db.scalar(select(AIDecisionFeedback))
+        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 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:"
+        )
+        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)

web/src/utils/expenseApplicationDecisionFeedback.js

@@ -0,0 +1,67 @@
+function normalizeText(value) {
+  return String(value || '').trim()
+}
+
+function resolveTrackedFieldKey(fieldKey = '') {
+  const normalized = normalizeText(fieldKey)
+  return normalized === 'time_return' ? 'time' : normalized
+}
+
+function resolveExistingChanges(preview = {}) {
+  const feedback = preview?.aiDecisionFeedback
+  return Array.isArray(feedback?.changedFields)
+    ? feedback.changedFields.filter((item) => item && typeof item === 'object')
+    : []
+}
+
+function resolveTrackedFieldKeys(fieldKeyOrKeys = '') {
+  const fieldKeys = Array.isArray(fieldKeyOrKeys) ? fieldKeyOrKeys : [fieldKeyOrKeys]
+  return [...new Set(fieldKeys.map(resolveTrackedFieldKey).filter(Boolean))]
+}
+
+function resolveFeedbackRevision(preview = {}) {
+  const revision = Number(preview?.aiDecisionFeedback?.revision)
+  return Number.isSafeInteger(revision) && revision >= 0 ? revision : 0
+}
+
+export function trackApplicationPreviewFieldEdit(
+  previousPreview = {},
+  nextPreview = {},
+  fieldKeyOrKeys = ''
+) {
+  const trackedFieldKeys = resolveTrackedFieldKeys(fieldKeyOrKeys)
+  if (!trackedFieldKeys.length) return nextPreview
+
+  const existingChanges = resolveExistingChanges(previousPreview)
+  const retainedChanges = existingChanges.filter(
+    (item) => !trackedFieldKeys.includes(resolveTrackedFieldKey(item.fieldKey))
+  )
+
+  for (const trackedFieldKey of trackedFieldKeys) {
+    const existingChange = existingChanges.find(
+      (item) => resolveTrackedFieldKey(item.fieldKey) === trackedFieldKey
+    )
+    const suggestedValue = existingChange
+      ? normalizeText(existingChange.suggestedValue)
+      : normalizeText(previousPreview?.fields?.[trackedFieldKey])
+    const finalValue = normalizeText(nextPreview?.fields?.[trackedFieldKey])
+    if (suggestedValue !== finalValue) {
+      retainedChanges.push({
+        fieldKey: trackedFieldKey,
+        suggestedValue,
+        finalValue
+      })
+    }
+  }
+
+  return {
+    ...nextPreview,
+    aiDecisionFeedback: {
+      schemaVersion: 1,
+      revision: resolveFeedbackRevision(previousPreview) + 1,
+      changedFields: retainedChanges.sort((left, right) => (
+        normalizeText(left.fieldKey).localeCompare(normalizeText(right.fieldKey))
+      ))
+    }
+  }
+}

web/src/views/scripts/useApplicationPreviewEditor.js

@@ -11,12 +11,66 @@ import {
   resolveApplicationDaysFromDateRange,
   refreshApplicationPreviewTransportEstimate
 } from '../../utils/expenseApplicationPreview.js'
+import { trackApplicationPreviewFieldEdit } from '../../utils/expenseApplicationDecisionFeedback.js'
 import {
   buildWorkbenchDateLabel,
   canApplyWorkbenchDateSelection,
   getTodayDateValue
 } from '../../utils/workbenchComposerDate.js'
 
+const APPLICATION_ESTIMATE_INPUT_FIELD_KEYS = [
+  'applicationType',
+  'time',
+  'location',
+  'days',
+  'transportMode',
+  'grade'
+]
+
+const APPLICATION_ESTIMATE_RESULT_FIELD_KEYS = [
+  'lodgingDailyCap',
+  'subsidyDailyCap',
+  'transportPolicy',
+  'policyEstimate',
+  'amount',
+  'matchedCity',
+  'ruleName',
+  'ruleVersion',
+  'hotelAmount',
+  'allowanceAmount',
+  'transportEstimatedAmount',
+  'transportEstimateDate',
+  'transportQueryLatencyMs',
+  'transportEstimateSource',
+  'transportEstimateConfidence',
+  'policyTotalAmount'
+]
+
+function buildApplicationEstimateInputFingerprint(preview = {}) {
+  const fields = preview?.fields || {}
+  return JSON.stringify(
+    APPLICATION_ESTIMATE_INPUT_FIELD_KEYS.map((fieldKey) => String(fields[fieldKey] || '').trim())
+  )
+}
+
+function mergeApplicationEstimateResult(currentPreview = {}, refreshedPreview = {}) {
+  const refreshedFields = refreshedPreview?.fields || {}
+  const estimateFields = Object.fromEntries(
+    APPLICATION_ESTIMATE_RESULT_FIELD_KEYS
+      .filter((fieldKey) => Object.hasOwn(refreshedFields, fieldKey))
+      .map((fieldKey) => [fieldKey, refreshedFields[fieldKey]])
+  )
+  return normalizeApplicationPreview({
+    ...currentPreview,
+    fields: {
+      ...(currentPreview?.fields || {}),
+      ...estimateFields
+    },
+    policyEstimate: refreshedPreview?.policyEstimate,
+    policyEstimateStatus: refreshedPreview?.policyEstimateStatus
+  })
+}
+
 function parseEditorDateMatches(value) {
   const text = String(value || '').trim()
   return [...text.matchAll(/20\d{2}[-/.]\d{1,2}[-/.]\d{1,2}/g)]
@@ -186,7 +240,14 @@ function validateApplicationPreviewDateRange(value = '') {
 }
 
 function shouldRefreshTransportEstimate(fieldKey) {
-  return ['transportMode', 'time', 'time_return', 'location', 'days'].includes(fieldKey)
+  return [
+    'applicationType',
+    'transportMode',
+    'time',
+    'time_return',
+    'location',
+    'days'
+  ].includes(fieldKey)
 }
 
 function resolveEditorCurrentUser(currentUser) {
@@ -270,6 +331,7 @@ export function useApplicationPreviewEditor({
   currentUser
 } = {}) {
   const applicationPreviewEditor = ref(buildEmptyEditor())
+  const estimateRequestVersions = new WeakMap()
 
   async function refreshApplicationPreviewEstimate(preview = {}) {
     const user = resolveEditorCurrentUser(currentUser)
@@ -406,14 +468,19 @@ export function useApplicationPreviewEditor({
       }
       return false
     }
-    const nextPreview = normalizeApplicationPreview({
+    const isDateField = isApplicationPreviewDateField(editor.fieldKey)
-      ...message.applicationPreview,
+    const nextPreview = trackApplicationPreviewFieldEdit(
-      fields: buildEditedApplicationPreviewFields(
+      message.applicationPreview,
-        message.applicationPreview.fields || {},
+      normalizeApplicationPreview({
-        editor,
+        ...message.applicationPreview,
-        nextValue
+        fields: buildEditedApplicationPreviewFields(
-      )
+          message.applicationPreview.fields || {},
-    })
+          editor,
+          nextValue
+        )
+      }),
+      isDateField ? [editor.fieldKey, 'days'] : editor.fieldKey
+    )
     const needRefreshEstimate = shouldRefreshTransportEstimate(editor.fieldKey)
     message.applicationPreview = needRefreshEstimate
       ? buildTransportEstimatePendingPreview(nextPreview)
@@ -422,9 +489,23 @@ export function useApplicationPreviewEditor({
     cancelApplicationPreviewEditor()
     persistSessionState?.()
     if (needRefreshEstimate) {
+      const requestVersion = (estimateRequestVersions.get(message) || 0) + 1
+      estimateRequestVersions.set(message, requestVersion)
+      const estimateInputFingerprint = buildApplicationEstimateInputFingerprint(nextPreview)
       const refreshedPreview = await refreshApplicationPreviewEstimate(nextPreview)
-      message.applicationPreview = refreshedPreview
+      const isLatestRequest = estimateRequestVersions.get(message) === requestVersion
-      message.text = buildLocalApplicationPreviewMessage(refreshedPreview)
+      const currentFingerprint = buildApplicationEstimateInputFingerprint(
+        message.applicationPreview
+      )
+      if (!isLatestRequest || currentFingerprint !== estimateInputFingerprint) {
+        return true
+      }
+      const mergedPreview = mergeApplicationEstimateResult(
+        message.applicationPreview,
+        refreshedPreview
+      )
+      message.applicationPreview = mergedPreview
+      message.text = buildLocalApplicationPreviewMessage(mergedPreview)
       persistSessionState?.()
       toast?.('已更新出行方式和费用测算。')
       return true

web/tests/ai-application-preview-actions.test.mjs

@@ -42,6 +42,16 @@ async function testSubmitActionUsesFastPreviewEndpoint() {
         days: '3天',
         transportMode: '火车',
         amount: '1000元'
+      },
+      aiDecisionFeedback: {
+        schemaVersion: 1,
+        changedFields: [
+          {
+            fieldKey: 'transportMode',
+            suggestedValue: '飞机',
+            finalValue: '火车'
+          }
+        ]
       }
     },
     currentUser: { username: 'zhangsan@example.com', name: '张三' },
@@ -54,6 +64,13 @@ async function testSubmitActionUsesFastPreviewEndpoint() {
   assert.equal(body.context_json.session_type, 'application')
   assert.equal(body.context_json.application_stage, 'expense_application')
   assert.equal(body.context_json.application_preview.fields.transportMode, '火车')
+  assert.deepEqual(body.context_json.application_preview.aiDecisionFeedback.changedFields, [
+    {
+      fieldKey: 'transportMode',
+      suggestedValue: '飞机',
+      finalValue: '火车'
+    }
+  ])
 }
 
 async function testSaveDraftActionUsesFastPreviewEndpoint() {

web/tests/expense-application-decision-feedback.test.mjs

@@ -0,0 +1,140 @@
+import assert from 'node:assert/strict'
+
+import { trackApplicationPreviewFieldEdit } from '../src/utils/expenseApplicationDecisionFeedback.js'
+
+function testTracksFirstExplicitEdit() {
+  const previous = {
+    fields: {
+      reason: '客户拜访',
+      location: '上海'
+    }
+  }
+  const next = {
+    ...previous,
+    fields: {
+      ...previous.fields,
+      reason: '客户现场实施'
+    }
+  }
+
+  const tracked = trackApplicationPreviewFieldEdit(previous, next, 'reason')
+
+  assert.deepEqual(tracked.aiDecisionFeedback.changedFields, [
+    {
+      fieldKey: 'reason',
+      suggestedValue: '客户拜访',
+      finalValue: '客户现场实施'
+    }
+  ])
+  assert.equal(tracked.aiDecisionFeedback.revision, 1)
+}
+
+function testKeepsOriginalSuggestionAcrossRepeatedEdits() {
+  const first = trackApplicationPreviewFieldEdit(
+    { fields: { transportMode: '火车' } },
+    { fields: { transportMode: '飞机' } },
+    'transportMode'
+  )
+  const second = trackApplicationPreviewFieldEdit(
+    first,
+    { ...first, fields: { transportMode: '自驾' } },
+    'transportMode'
+  )
+
+  assert.deepEqual(second.aiDecisionFeedback.changedFields, [
+    {
+      fieldKey: 'transportMode',
+      suggestedValue: '火车',
+      finalValue: '自驾'
+    }
+  ])
+}
+
+function testRemovesDiffWhenUserRestoresSuggestion() {
+  const edited = trackApplicationPreviewFieldEdit(
+    { fields: { location: '上海' } },
+    { fields: { location: '北京' } },
+    'location'
+  )
+  const restored = trackApplicationPreviewFieldEdit(
+    edited,
+    { ...edited, fields: { location: '上海' } },
+    'location'
+  )
+
+  assert.deepEqual(restored.aiDecisionFeedback.changedFields, [])
+}
+
+function testReturnDateUsesCanonicalTimeField() {
+  const tracked = trackApplicationPreviewFieldEdit(
+    { fields: { time: '2026-07-20 至 2026-07-22' } },
+    { fields: { time: '2026-07-20 至 2026-07-24' } },
+    'time_return'
+  )
+
+  assert.equal(tracked.aiDecisionFeedback.changedFields[0].fieldKey, 'time')
+}
+
+function testTracksDateAndDerivedDaysTogether() {
+  const previous = {
+    fields: {
+      time: '2026-02-20 至 2026-02-23',
+      days: '4天'
+    }
+  }
+  const next = {
+    fields: {
+      time: '2026-02-20 至 2026-02-24',
+      days: '5天'
+    }
+  }
+
+  const tracked = trackApplicationPreviewFieldEdit(
+    previous,
+    next,
+    ['time_return', 'days']
+  )
+
+  assert.deepEqual(tracked.aiDecisionFeedback.changedFields, [
+    { fieldKey: 'days', suggestedValue: '4天', finalValue: '5天' },
+    {
+      fieldKey: 'time',
+      suggestedValue: '2026-02-20 至 2026-02-23',
+      finalValue: '2026-02-20 至 2026-02-24'
+    }
+  ])
+  assert.equal(tracked.aiDecisionFeedback.revision, 1)
+
+  const editedAgain = trackApplicationPreviewFieldEdit(
+    tracked,
+    {
+      ...tracked,
+      fields: {
+        time: '2026-02-20 至 2026-02-25',
+        days: '6天'
+      }
+    },
+    ['time', 'days']
+  )
+  assert.equal(editedAgain.aiDecisionFeedback.changedFields[0].suggestedValue, '4天')
+  assert.equal(
+    editedAgain.aiDecisionFeedback.changedFields[1].suggestedValue,
+    '2026-02-20 至 2026-02-23'
+  )
+  assert.equal(editedAgain.aiDecisionFeedback.revision, 2)
+
+  const restored = trackApplicationPreviewFieldEdit(
+    editedAgain,
+    { ...editedAgain, fields: previous.fields },
+    ['time', 'days']
+  )
+  assert.deepEqual(restored.aiDecisionFeedback.changedFields, [])
+  assert.equal(restored.aiDecisionFeedback.revision, 3)
+}
+
+testTracksFirstExplicitEdit()
+testKeepsOriginalSuggestionAcrossRepeatedEdits()
+testRemovesDiffWhenUserRestoresSuggestion()
+testReturnDateUsesCanonicalTimeField()
+testTracksDateAndDerivedDaysTogether()
+console.log('expense-application-decision-feedback tests passed')

web/tests/expense-application-fast-preview.test.mjs

@@ -1916,6 +1916,14 @@ test('application preview editor recalculates days and subsidy after date range
   assert.equal(message.applicationPreview.fields.lodgingDailyCap, '450\u5143/\u5929')
   assert.equal(message.applicationPreview.fields.subsidyDailyCap, '100\u5143/\u5929')
   assert.match(message.applicationPreview.fields.policyEstimate, /\u8865\u8d34 400\u5143/)
+  assert.deepEqual(message.applicationPreview.aiDecisionFeedback.changedFields, [
+    { fieldKey: 'days', suggestedValue: '1\u5929', finalValue: '4\u5929' },
+    {
+      fieldKey: 'time',
+      suggestedValue: '2026-05-25',
+      finalValue: '2026-02-20 \u81f3 2026-02-23'
+    }
+  ])
 })
 
 test('application preview editor can edit return date from table row', async () => {
@@ -1981,6 +1989,14 @@ test('application preview editor can edit return date from table row', async ()
   })
   assert.equal(message.applicationPreview.fields.time, '2026-02-20 \u81f3 2026-02-24')
   assert.equal(message.applicationPreview.fields.days, '5\u5929')
+  assert.deepEqual(message.applicationPreview.aiDecisionFeedback.changedFields, [
+    { fieldKey: 'days', suggestedValue: '4\u5929', finalValue: '5\u5929' },
+    {
+      fieldKey: 'time',
+      suggestedValue: '2026-02-20 \u81f3 2026-02-23',
+      finalValue: '2026-02-20 \u81f3 2026-02-24'
+    }
+  ])
 })
 
 test('application preview editor can edit return date from inline table input', async () => {
@@ -2185,3 +2201,210 @@ test('application preview editor estimates after shorthand return date input', a
   assert.equal(message.applicationPreview.fields.amount, '1,650\u5143\uff08\u4e0d\u542b\u4ea4\u901a\uff09')
   assert.match(message.applicationPreview.fields.policyEstimate, /\u4ea4\u901a\u5f85\u8865\u5145/)
 })
+
+test('application preview editor ignores stale out-of-order estimate responses', async () => {
+  const message = {
+    id: 'application-preview-stale-estimate-message',
+    applicationPreview: normalizeApplicationPreview({
+      fields: {
+        applicationType: '差旅费用申请',
+        time: '2026-07-20 至 2026-07-21',
+        location: '上海',
+        reason: '客户现场实施',
+        days: '2天',
+        transportMode: '火车',
+        grade: 'P5'
+      }
+    }),
+    text: ''
+  }
+  const pendingEstimates = []
+  const editor = useApplicationPreviewEditor({
+    persistSessionState: () => {},
+    toast: () => {},
+    currentUser: ref({ grade: 'P5' }),
+    calculateTravelReimbursement: (payload) => new Promise((resolve) => {
+      pendingEstimates.push({ payload, resolve })
+    })
+  })
+
+  editor.openApplicationPreviewEditor(message, 'location', '上海')
+  editor.applicationPreviewEditor.value.draftValue = '北京'
+  const firstCommit = editor.commitApplicationPreviewEditor(message)
+
+  editor.openApplicationPreviewEditor(message, 'location', '北京')
+  editor.applicationPreviewEditor.value.draftValue = '广州'
+  const secondCommit = editor.commitApplicationPreviewEditor(message)
+
+  assert.equal(pendingEstimates.length, 2)
+  pendingEstimates[1].resolve({
+    days: 2,
+    location: '广州',
+    matched_city: '广州',
+    grade: 'P5',
+    hotel_rate: 600,
+    hotel_amount: 1200,
+    total_allowance_rate: 100,
+    allowance_amount: 200,
+    total_amount: 1400,
+    rule_name: '第二次估算规则',
+    rule_version: 'v2'
+  })
+  await secondCommit
+  pendingEstimates[0].resolve({
+    days: 2,
+    location: '北京',
+    matched_city: '北京',
+    grade: 'P5',
+    hotel_rate: 300,
+    hotel_amount: 600,
+    total_allowance_rate: 80,
+    allowance_amount: 160,
+    total_amount: 760,
+    rule_name: '过期估算规则',
+    rule_version: 'v1'
+  })
+  await firstCommit
+
+  assert.equal(message.applicationPreview.fields.location, '广州')
+  assert.equal(message.applicationPreview.fields.ruleName, '第二次估算规则')
+  assert.deepEqual(message.applicationPreview.aiDecisionFeedback.changedFields, [
+    { fieldKey: 'location', suggestedValue: '上海', finalValue: '广州' }
+  ])
+})
+
+test('application type edit starts a fresh estimate instead of leaving pending state', async () => {
+  const message = {
+    id: 'application-preview-type-estimate-message',
+    applicationPreview: normalizeApplicationPreview({
+      fields: {
+        applicationType: '差旅费用申请',
+        time: '2026-07-20 至 2026-07-21',
+        location: '上海',
+        reason: '客户现场实施',
+        days: '2天',
+        transportMode: '火车',
+        grade: 'P5'
+      }
+    }),
+    text: ''
+  }
+  const pendingEstimates = []
+  const editor = useApplicationPreviewEditor({
+    persistSessionState: () => {},
+    toast: () => {},
+    currentUser: ref({ grade: 'P5' }),
+    calculateTravelReimbursement: (payload) => new Promise((resolve) => {
+      pendingEstimates.push({ payload, resolve })
+    })
+  })
+
+  editor.openApplicationPreviewEditor(message, 'location', '上海')
+  editor.applicationPreviewEditor.value.draftValue = '北京'
+  const firstCommit = editor.commitApplicationPreviewEditor(message)
+
+  editor.openApplicationPreviewEditor(message, 'applicationType', '差旅费用申请')
+  editor.applicationPreviewEditor.value.draftValue = '出差费用申请'
+  const secondCommit = editor.commitApplicationPreviewEditor(message)
+
+  assert.equal(pendingEstimates.length, 2)
+  pendingEstimates[1].resolve({
+    days: 2,
+    location: '北京',
+    matched_city: '北京',
+    grade: 'P5',
+    hotel_rate: 600,
+    hotel_amount: 1200,
+    total_allowance_rate: 100,
+    allowance_amount: 200,
+    total_amount: 1400,
+    rule_name: '申请类型更新规则',
+    rule_version: 'v2'
+  })
+  await secondCommit
+  pendingEstimates[0].resolve({
+    days: 2,
+    location: '北京',
+    matched_city: '北京',
+    grade: 'P5',
+    hotel_rate: 300,
+    hotel_amount: 600,
+    total_allowance_rate: 80,
+    allowance_amount: 160,
+    total_amount: 760,
+    rule_name: '过期申请类型规则',
+    rule_version: 'v1'
+  })
+  await firstCommit
+
+  assert.equal(message.applicationPreview.fields.applicationType, '出差费用申请')
+  assert.equal(message.applicationPreview.fields.ruleName, '申请类型更新规则')
+  assert.equal(message.applicationPreview.policyEstimateStatus, 'completed')
+  assert.doesNotMatch(message.applicationPreview.fields.policyEstimate, /正在/)
+  assert.deepEqual(message.applicationPreview.aiDecisionFeedback.changedFields, [
+    {
+      fieldKey: 'applicationType',
+      suggestedValue: '差旅费用申请',
+      finalValue: '出差费用申请'
+    },
+    { fieldKey: 'location', suggestedValue: '上海', finalValue: '北京' }
+  ])
+})
+
+test('application preview editor merges estimate without dropping later edits', async () => {
+  const message = {
+    id: 'application-preview-merge-estimate-message',
+    applicationPreview: normalizeApplicationPreview({
+      fields: {
+        applicationType: '差旅费用申请',
+        time: '2026-07-20 至 2026-07-21',
+        location: '上海',
+        reason: '客户拜访',
+        days: '2天',
+        transportMode: '火车',
+        grade: 'P5'
+      }
+    }),
+    text: ''
+  }
+  let resolveEstimate
+  const editor = useApplicationPreviewEditor({
+    persistSessionState: () => {},
+    toast: () => {},
+    currentUser: ref({ grade: 'P5' }),
+    calculateTravelReimbursement: () => new Promise((resolve) => {
+      resolveEstimate = resolve
+    })
+  })
+
+  editor.openApplicationPreviewEditor(message, 'location', '上海')
+  editor.applicationPreviewEditor.value.draftValue = '北京'
+  const estimateCommit = editor.commitApplicationPreviewEditor(message)
+
+  editor.openApplicationPreviewEditor(message, 'reason', '客户拜访')
+  editor.applicationPreviewEditor.value.draftValue = '客户现场实施'
+  await editor.commitApplicationPreviewEditor(message)
+
+  resolveEstimate({
+    days: 2,
+    location: '北京',
+    matched_city: '北京',
+    grade: 'P5',
+    hotel_rate: 500,
+    hotel_amount: 1000,
+    total_allowance_rate: 100,
+    allowance_amount: 200,
+    total_amount: 1200,
+    rule_name: '并发合并规则',
+    rule_version: 'v1'
+  })
+  await estimateCommit
+
+  assert.equal(message.applicationPreview.fields.location, '北京')
+  assert.equal(message.applicationPreview.fields.reason, '客户现场实施')
+  assert.equal(message.applicationPreview.fields.ruleName, '并发合并规则')
+  assert.deepEqual(message.applicationPreview.aiDecisionFeedback.changedFields, [
+    { fieldKey: 'location', suggestedValue: '上海', finalValue: '北京' },
+    { fieldKey: 'reason', suggestedValue: '客户拜访', finalValue: '客户现场实施' }
+  ])
+})