问题摘要与归属

Runtime Compose 的派生 .query() 中声明 calculatedFields，再在 columns 选择该计算列，校验却把它当作“来源未输出字段”。同一基础字段与同一算式改成列内表达式即可验证、预览并执行。

本单先请求明确能力合同并改善诊断：当前 Skill 明确列出基础 dsl 的 calculatedFields，但没有明确承诺派生层等价支持。因此不把“必须支持此能力”预设成已确认合同 BUG；如果派生层本来不支持，也应指出不支持的参数，而非引导 AI 去源 QM 找该计算字段。

归属：Java Runtime 的 compose.validate / 派生 schema 推导与可修复诊断。关联 #160 的结构化诊断方向，但这里是一个独立的 calculatedFields 入口差异场景。

环境与版本核验

实测日期：2026-10-24，Windows、Java 17.0.12、Python 3.11.5。
采用已发布的 Launcher 0.1.39，按发行包原始 PowerShell 脚本启动；当前监听 PID 133260 的 JAR 路径已核对为 foggy-runtime-launcher-0.1.39.jar。
manifest source commit：7b6d06987e3ec90f05143a0552791f80381dcc4b，Java projectVersion：9.3.0-SNAPSHOT；不是本地源码编译结果。
JAR SHA256：8228d31c104d58b1b2f1d1c5bc856a25531e4d5af06e1e9eefed2be576f20bf9。
项目固定 CLI：foggy-runtime 0.1.25；CLI 子进程使用 UTF-8 输出。
Runtime：http://127.0.0.1:18067；Namespace：chinook-composition；命名数据源：chinook-composition-sqlite（SQLite，公开 Chinook v1.4.5 副本）。
securityMode=none-dev-test-only，匿名本地教学环境，Console/FAP 未启用；未使用生产数据、身份、私有接口或 SQL 直通。

capabilities 实际响应节选：

{
  "engine": "java",
  "runtimeApiVersion": "foggy-runtime-api/v1",
  "schemaVersion": "2026-06-06",
  "securityMode": "none-dev-test-only",
  "capabilities": {
    "compose.validate": "supported",
    "compose.preview": "supported",
    "compose.execute": "supported",
    "models.describe": "supported"
  }
}

模型前提（已实际 describe 核对）

ChinookInvoiceQueryModel 对应 Chinook 的 Invoice 表，必要公开字段如下：

字段	实际物理映射 / 角色
customerId	Invoice.CustomerId，客户键
invoiceDate	Invoice.InvoiceDate，账单日期
invoiceAmount	Invoice.Total，MONEY、sum；沿用样例原单位

下面只引用已公开字段和前一计划的输出别名。年度基线为 47 位有账单客户、83 张账单、477.53；原始数据与模型在复现前后均未修改。

复现步骤与实际请求

启动上述 0.1.39，绑定 SQLite，并加载 / describe 上述账单 QM。
将下列完整脚本保存为 01-derived-calculated-fields.fsscript。
使用项目固定 CLI 校验：

foggy-runtime --output json --base-url http://127.0.0.1:18067 --namespace chinook-composition compose validate --script ./01-derived-calculated-fields.fsscript

实际请求脚本：

const bills = dsl({
    model: 'ChinookInvoiceQueryModel',
    columns: ['customerId as customer_id', 'sum(invoiceAmount) as bill_amount'],
    slice: [{field: 'invoiceDate', op: '[)', value: ['2012-01-01', '2013-01-01']}],
    groupBy: ['customerId']
});
return {plans: bills.query({
    columns: ['customer_id', 'bill_amount', 'double_amount'],
    calculatedFields: [{name: 'double_amount', expression: 'bill_amount * 2'}],
    orderBy: ['customer_id'], limit: 100
})};

预期

若派生 query 承诺支持 calculatedFields，应从来源输出的 bill_amount 推导 double_amount，校验通过并正确计算。
若该派生入口不支持，应在校验阶段明确指出 calculatedFields 不适用 / 未支持，给出参数路径及建议写法（例如列内表达式）；不要仅报新声明的计算列“不在 source schema”。
不要求放宽字段权限或执行被拒绝的请求。

实际响应与关键诊断

CLI exitCode 2；success=false、field=double_amount、phase=compose.validate、safeToAutoRepair=true。完整实际响应：

{
  "diagnostics": {
    "attributes": {
      "composeErrorCode": "compose-schema-error/derived-query/unknown-field",
      "composeErrorPhase": "schema-derive",
      "datasourceType": "sqlite",
      "defaultDialect": "sqlite",
      "dialectSource": "namespace-datasource-type",
      "namespace": "chinook-composition",
      "namespaceDatasourceId": "chinook-composition-sqlite",
      "namespaceDatasourceStatus": "active",
      "resolvedDialect": "sqlite"
    },
    "warnings": []
  },
  "engine": "java",
  "error": {
    "code": "COMPOSE_SCRIPT_INVALID",
    "field": "double_amount",
    "message": "derived query references unknown field 'double_amount' not present in source output schema",
    "phase": "compose.validate",
    "safeToAutoRepair": true,
    "suggestedNextAction": "Inspect compose fields/schema and retry."
  },
  "runtimeApiVersion": "foggy-runtime-api/v1",
  "success": false
}

实际 CLI / 诊断记录：

exitCode=2, success=false
error.code=COMPOSE_SCRIPT_INVALID
error.field=double_amount
error.phase=compose.validate
diagnostics.attributes.composeErrorCode=compose-schema-error/derived-query/unknown-field
diagnostics.attributes.composeErrorPhase=schema-derive
safeToAutoRepair=true
message=derived query references unknown field 'double_amount' not present in source output schema

本例是结构化校验拒绝，本轮服务端增量日志未见对应的独立 Java 异常堆栈；没有为它拼接另一个请求的 ClassCastException 日志。失败请求没有 execute。

正常对照（同一模型、期间、字段与算式）

仅改为列内表达式，实际完整请求：

const bills = dsl({
    model: 'ChinookInvoiceQueryModel',
    columns: ['customerId as customer_id', 'sum(invoiceAmount) as bill_amount'],
    slice: [{field: 'invoiceDate', op: '[)', value: ['2012-01-01', '2013-01-01']}],
    groupBy: ['customerId']
});
return {plans: bills.query({
    columns: ['customer_id', 'bill_amount', 'bill_amount * 2 as double_amount'],
    orderBy: ['customer_id'], limit: 100
})};

该请求依次 compose validate → preview → execute，退出码均为 0，实际返回 47 行，每行 double_amount 等于 bill_amount × 2。执行响应节选（items 对应 Compose 的 data.value.plans）：

{
  "success": true,
  "data": {
    "valid": true,
    "value": {
      "plans": [
        {
          "bill_amount": 15.84,
          "customer_id": 1,
          "double_amount": 31.68
        },
        {
          "bill_amount": 0.99,
          "customer_id": 2,
          "double_amount": 1.98
        }
      ]
    }
  }
}

原始账单合计仍为 477.53，独立 Decimal 核对通过。这说明不是基础字段缺失或乘法本身不可用。

最初发现的 JOIN 形状

问题最初发生在两个 QM 按客户聚合并 JOIN 后，以 ROUND(bill_amount - line_amount, 2) 生成 amount_difference。加入本次单 QM 最小复现，是为了排除 JOIN 和小数函数对定位的干扰。原请求及实际响应如下，明细 QM 同样为公开 Chinook 的 InvoiceLine，lineAmount=UnitPrice×Quantity，通过 Invoice 取客户及日期。

最初的完整 JOIN 请求与实际错误

const bills = dsl({
    model: 'ChinookInvoiceQueryModel',
    columns: ['customerId as bill_customer_id', 'sum(invoiceAmount) as bill_amount'],
    slice: [{ field: 'invoiceDate', op: '[)', value: ['2012-01-01', '2013-01-01'] }],
    groupBy: ['customerId']
});

const lines = dsl({
    model: 'ChinookInvoiceLineQueryModel',
    columns: ['invoice$customerId as line_customer_id', 'sum(lineAmount) as line_amount'],
    slice: [{ field: 'invoice$invoiceDate', op: '[)', value: ['2012-01-01', '2013-01-01'] }],
    groupBy: ['invoice$customerId']
});

const joined = bills.join(lines, 'left', [
    { left: 'bill_customer_id', op: '=', right: 'line_customer_id' }
]);

const compared = joined.query({
    columns: ['bill_customer_id', 'line_customer_id', 'bill_amount', 'line_amount', 'amount_difference'],
    calculatedFields: [{ name: 'amount_difference', expression: 'ROUND(bill_amount - line_amount, 2)' }],
    orderBy: ['bill_customer_id'], limit: 100
});

return { plans: compared };

{
  "diagnostics": {
    "attributes": {
      "composeErrorCode": "compose-schema-error/derived-query/unknown-field",
      "composeErrorPhase": "schema-derive",
      "datasourceType": "sqlite",
      "defaultDialect": "sqlite",
      "dialectSource": "namespace-datasource-type",
      "namespace": "chinook-composition",
      "namespaceDatasourceId": "chinook-composition-sqlite",
      "namespaceDatasourceStatus": "active",
      "resolvedDialect": "sqlite"
    },
    "warnings": []
  },
  "engine": "java",
  "error": {
    "code": "COMPOSE_SCRIPT_INVALID",
    "field": "amount_difference",
    "message": "derived query references unknown field 'amount_difference' not present in source output schema",
    "phase": "compose.validate",
    "safeToAutoRepair": true,
    "suggestedNextAction": "Inspect compose fields/schema and retry."
  },
  "runtimeApiVersion": "foggy-runtime-api/v1",
  "success": false
}

建议验收

明确派生 query 的支持矩阵；对上述一种支持实现或明确拒绝实现增加测试。错误应能让 AI 知道是参数能力边界，还是引用了真正未输出的字段；若支持计算列，覆盖投影与后续筛选对该列的引用。保留现有字段治理边界。