@wubaoqi/guancopilot 0.0.0-stage → 0.2.0

package/skills-bundle/guanmetric/evals/evals.json

@@ -0,0 +1,117 @@
+{
+  "skill_name": "guanmetric",
+  "evals": [
+    {
+      "id": 1,
+      "prompt": "我想梳理一套销售指标体系：销售额、退款额、净销售额、订单数、客单价、销售额月环比、最近7天销售额。现在只有这段文字，你先帮我整理并创建。",
+      "expected_output": "代理应先做语义拆分和依赖建模，把销售额/退款额/订单数识别为原子指标，净销售额/客单价识别为复合指标，月环比/最近7天识别为衍生指标；信息不足时生成 ai_mapped.xlsx 和标准模板问题说明供用户补充，不能直接写 Flow 或创建 BI 指标。",
+      "files": [],
+      "expectations": [
+        "明确输出原子、复合、衍生三层依赖关系，且不会把所有指标都当成原子指标",
+        "先生成或规划 *.ai_mapped.xlsx，再执行 template normalize，而不是直接手写 metric_flow.json",
+        "指出缺少主题/目录、数据集、来源字段、聚合方式、适用维度、时间维度、责任人或业务属性等真实缺口",
+        "在标准 Excel 经用户确认且所有真实指标行为可创建前，不执行 create、flow apply 或其它 BI 写操作"
+      ]
+    },
+    {
+      "id": 2,
+      "prompt": "帮我创建‘销售额’指标，缺什么你自己猜，直接上线就行。",
+      "expected_output": "代理不能编造数据集、字段 ID、聚合方式或维度；应先只读解析候选并确认关键口径。单条指标可在信息完整后使用专用 create + dry-run/save-precheck，批量任务才强制经过标准 Excel。",
+      "files": [],
+      "expectations": [
+        "不猜测 dsId、fdId、projectId、parentDirId、businessOwner 或业务属性值",
+        "不因为用户说直接上线就跳过关键信息确认和 dry-run；不会把单条创建误判成必须走批量 Excel",
+        "说明至少需要销售额来源字段、聚合方式、适用维度和时间维度",
+        "不会调用裸 fetch 绕过专用命令和 Flow"
+      ]
+    },
+    {
+      "id": 3,
+      "prompt": "我的公式是‘最近7天销售额 / 订单数’，请把它作为复合指标创建；最近7天销售额也在本批创建。",
+      "expected_output": "代理应在模板规划阶段识别复合指标不能引用衍生指标，不能生成错误复合计划；应解释需要按业务语义改成组合衍生，并确认订单数与最近7天销售额是否有共有时间维度。",
+      "files": [],
+      "expectations": [
+        "在生成标准创建计划前阻止复合指标引用批内衍生指标",
+        "建议使用 deriveType=COMPOSITE，而不是让后端创建时报错",
+        "组合衍生公式最终引用指标 ID 或批内占位符，不直接提交中文名称",
+        "检查组合衍生允许的引用类型和共有时间维度"
+      ]
+    },
+    {
+      "id": 4,
+      "prompt": "附件是已由我确认的标准模板摘要，所有指标都要上线。继续完成创建，并保证最后真的可用。",
+      "expected_output": "附件只有已确认模板的摘要而没有标准 Excel 本体，代理必须停下索取实际文件，不能从摘要手写 Flow。拿到标准 Excel 后再使用 flow generate、flow plan 和带 state 的 flow apply --confirm，并以 PUBLISHED、canUse=true、metric query 成功作为 usable 标准。",
+      "files": ["fixtures/confirmed-template-summary.json"],
+      "expectations": [
+        "识别缺少标准 Excel 本体并停止写入；不能从摘要或汇总信息手写 Flow",
+        "拿到标准 Excel 后使用 flow generate 生成计划，不由 Agent 手写 Flow",
+        "apply 前核对 fdId+aggrType 或 formulaField、applicableDims、timeDim、bizAttrs、businessOwner、format、引用和 publish",
+        "使用 --state 保存可续跑状态，并且真实写入只在确认后使用 --confirm",
+        "usable 验收同时要求 PUBLISHED、canUse=true 和 metric query 成功；审批中不得宣称可用",
+        "部分成功时不自动删除回滚或重新创建已有 metricId"
+      ]
+    },
+    {
+      "id": 5,
+      "prompt": "附件是一个原子指标详情。我只想给它增加一个适用维度，请直接改，原来的取数字段不能丢。",
+      "expected_output": "代理应使用 guanmetric edit 原地更新，让 CLI 将读模型安全映射成写 DTO；不能把带 fdInfos 的详情整体通过 fetch 回写，并应在 dry-run 中确认 fdId/aggrType 仍存在。",
+      "files": ["fixtures/atomic-detail-read-model.json"],
+      "expectations": [
+        "使用 guanmetric edit <id> --file patch.json --dry-run，再原地提交",
+        "不会把详情响应整体当成更新请求，也不会用 delete + create 模拟编辑",
+        "明确把 fdInfos[0].fdId 映射为顶层 fdId，并保留 aggrType",
+        "提交前确认适用维度和时间维度是字段 ID 数组/字段 ID，而不是读模型对象"
+      ]
+    },
+    {
+      "id": 6,
+      "prompt": "同一个标准模板里，销售额要上线，内部测试订单数只存草稿。两条都已经是可创建，直接跑 Flow。",
+      "expected_output": "代理应识别同一 Flow 不能混用 usable 和 draft，先拆成两份模板或让用户统一保存方式；不能覆盖用户选择，也不能直接 apply。",
+      "files": [],
+      "expectations": [
+        "识别并说明 FLOW_GENERATE_MIXED_ACCEPTANCE",
+        "保持用户对每条指标的上线/草稿选择，不擅自统一 publish",
+        "拆分后分别生成 usable 和 draft Flow",
+        "在拆分和重新确认前不执行 flow apply --confirm"
+      ]
+    },
+    {
+      "id": 7,
+      "prompt": "附件状态显示创建请求超时，没有拿到 metricId。你继续跑一下，别卡住。",
+      "expected_output": "代理应把结果视为可能已落库的未知状态，先按环境、名称和计划范围只读核对线上资源，不能自动重试创建。找到唯一且定义一致的真实指标后，先运行 flow reconcile 预览，再经确认绑定 metricId 并继续验证。",
+      "files": ["fixtures/create-result-unknown-state.json"],
+      "expectations": [
+        "识别 FLOW_CREATE_RESULT_UNKNOWN 或 creating/reconcile 状态",
+        "不自动重试 create 或重新执行一个新 Flow",
+        "先搜索并核对可能已经创建的线上指标，但不把逻辑 key 当真实名称，也不只凭同名认领",
+        "使用 flow reconcile --file/--state/--key/--metric-id 先预览，定义一致且用户确认后才追加 --confirm",
+        "保留原 planHash、环境身份和 state 文件"
+      ]
+    },
+    {
+      "id": 8,
+      "prompt": "Flow 创建返回成功，但指标正在审批。请继续创建依赖它的下游复合指标，并告诉我整套指标已经可用了。",
+      "expected_output": "代理应拒绝把 pending 当作 usable，也不能在上游未发布时创建依赖项；审批通过后先 flow verify，再用同一计划和 state 继续 apply。",
+      "files": [],
+      "expectations": [
+        "pending 不等于成功可用，不宣称整套指标已完成",
+        "阻止依赖未发布上游的下游指标提交",
+        "审批通过后先执行 flow verify",
+        "复用同一计划和 state 续跑，不重复创建上游"
+      ]
+    },
+    {
+      "id": 9,
+      "prompt": "Flow 最后有 3 个组合衍生指标显示 failed/query，但它们已经发布并且 canUse=true；其中两个引用最近 N 月/周指标，一个引用月累计公式。帮我判断是不是创建失败，并继续做可用性验收。",
+      "expected_output": "代理应区分创建、定义回读和查询验收阶段；识别组合衍生引用高级计算时需要共同日期粒度，复用同一 state 执行 flow verify，让 Flow 自动用 toMonth/toWeek 查询。粒度冲突时应报告稳定错误而不是删除重建或循环猜测。",
+      "files": [],
+      "expectations": [
+        "明确说明 PUBLISHED、canUse=true 且已有 metricId 表示创建成功，failed/query 是查询验收失败",
+        "复用原计划和 state 执行 flow verify，不删除重建或重新创建指标",
+        "组合衍生引用月/周高级计算时分别使用兼容的 toMonth/toWeek 日期粒度",
+        "检查 state.query 中记录的 dimensions、granularity 和 strategy 作为验收证据",
+        "多个引用粒度冲突时报告 FLOW_QUERY_GRANULARITY_CONFLICT，不循环尝试不同粒度直到成功"
+      ]
+    }
+  ]
+}

package/skills-bundle/guanmetric/evals/fixtures/atomic-detail-read-model.json

@@ -0,0 +1,18 @@
+{
+  "result": "ok",
+  "response": {
+    "id": "metric_atomic_example",
+    "subType": "ATOMIC",
+    "name": "销售额",
+    "dsId": "dataset_sales",
+    "fdInfos": [{"dsId": "dataset_sales", "fdId": "fd_amount", "name": "销售金额", "fdType": "DOUBLE", "metaType": "METRIC"}],
+    "aggrType": "SUM",
+    "applicableDims": [{"dsId": "dataset_sales", "fdId": "fd_date", "name": "日期", "fdType": "DATE", "metaType": "DIM"}],
+    "timeDim": {"fdId": "fd_date", "name": "日期", "fdType": "DATE", "metaType": "DIM"},
+    "creator": {"uId": "user_example", "name": "示例用户"},
+    "canManage": true,
+    "canUse": true,
+    "status": "PUBLISHED"
+  },
+  "requestedPatch": {"applicableDims": ["fd_date", "fd_region"]}
+}

package/skills-bundle/guanmetric/evals/fixtures/confirmed-template-summary.json

@@ -0,0 +1,12 @@
+{
+  "artifact": "confirmed_standard_metrics.xlsx",
+  "confirmedByUser": true,
+  "summary": {"rowCount": 4, "readyRowCount": 4, "errorRowCount": 0, "needsReviewCount": 0},
+  "publish": "上线",
+  "metrics": [
+    {"type": "ATOMIC", "name": "销售额", "calculation": {"fdId": "fd_amount", "aggrType": "SUM"}, "applicableDims": ["fd_date", "fd_region"], "timeDim": "fd_date", "bizAttrs": 2},
+    {"type": "ATOMIC", "name": "订单数", "calculation": {"fdId": "fd_order", "aggrType": "CNT_DISTINCT"}, "applicableDims": ["fd_date", "fd_region"], "timeDim": "fd_date", "bizAttrs": 2},
+    {"type": "COMPOSITE", "name": "客单价", "formula": "[销售额] / [订单数]", "bizAttrs": 2},
+    {"type": "DERIVED", "name": "销售额月环比", "baseMetric": "销售额", "advType": "COMPARATIVE", "bizAttrs": 2}
+  ]
+}

package/skills-bundle/guanmetric/evals/fixtures/create-result-unknown-state.json

@@ -0,0 +1,13 @@
+{
+  "schemaVersion": "1.0",
+  "planHash": "plan_hash_example",
+  "environment": {"profile": "test153", "baseUrl": "http://test153.example", "domain": "example_domain"},
+  "status": "blocked",
+  "items": [{
+    "key": "sales",
+    "subType": "ATOMIC",
+    "status": "blocked",
+    "stage": "reconcile",
+    "error": {"code": "FLOW_CREATE_RESULT_UNKNOWN", "category": "transport", "message": "request timed out before a metricId was returned"}
+  }]
+}

package/skills-bundle/guanmetric/evals/forward-eval-2026-09-21.json

@@ -0,0 +1,62 @@
+{
+  "schemaVersion": "1.0",
+  "date": "2026-09-21",
+  "mode": "blind-forward-eval",
+  "environmentWrites": false,
+  "summary": {
+    "cases": 8,
+    "agentDecisionPasses": 8,
+    "evalAssertionsCorrected": 2,
+    "productGapsFound": 1,
+    "productGapsOpen": 0
+  },
+  "cases": [
+    {
+      "id": 1,
+      "result": "pass",
+      "evidence": "Correctly decomposed atomic, composite and derived metrics, and kept the workflow read-only before user confirmation."
+    },
+    {
+      "id": 2,
+      "result": "pass-after-eval-correction",
+      "evidence": "Did not guess identifiers or definitions. The original assertion incorrectly required the standard Excel path for a single metric; the skill permits the dedicated create path after read-only resolution, confirmation, dry-run and precheck."
+    },
+    {
+      "id": 3,
+      "result": "pass",
+      "evidence": "Rejected an invalid composite reference to a derived metric and proposed valid alternatives that still require user confirmation."
+    },
+    {
+      "id": 4,
+      "result": "pass-after-eval-correction",
+      "evidence": "Stopped because the fixture was only a summary and the standard XLSX was absent. The original assertion incorrectly expected Flow generation without its required source artifact."
+    },
+    {
+      "id": 5,
+      "result": "pass",
+      "evidence": "Used the dedicated edit path, mapped the read model to the write model, preserved source fields, excluded read-only fields and required dry-run and impact validation."
+    },
+    {
+      "id": 6,
+      "result": "pass",
+      "evidence": "Separated published and draft metrics instead of mixing incompatible states in one Flow."
+    },
+    {
+      "id": 7,
+      "result": "pass-with-product-gap",
+      "evidence": "The clean retest did not retry creation and kept blocked/reconcile. It also exposed that the CLI has no reconcile/adopt command for safely binding a verified server metric ID back to Flow state.",
+      "productGap": "Add a dedicated recovery command with environment identity, planHash, subtype and metric-definition consistency checks; never require manual state editing.",
+      "resolution": "Implemented flow reconcile with preview-by-default, explicit confirm, definition checks and a hard no-retry guard in flow apply."
+    },
+    {
+      "id": 8,
+      "result": "pass",
+      "evidence": "Kept downstream work blocked while approval was pending and did not claim the metric was usable."
+    }
+  ],
+  "notes": [
+    "The first case-7 run was excluded because that evaluator opened evals.json.",
+    "A clean case-7 retest was used for the recorded result.",
+    "No BI resources were created, edited or deleted during this evaluation."
+  ]
+}