From 78800511e0d0e3fca249b7efd9f18b2b5c9f056e Mon Sep 17 00:00:00 2001
From: ThearchyHelios <work@thearchyhelios.com>
Date: Wed, 21 Oct 2026 18:17:56 +0800
Subject: [PATCH] =?UTF-8?q?=E5=89=8D=E7=AB=AF=EF=BC=9A=E5=88=86=E6=9E=90?=
 =?UTF-8?q?=E5=8D=A1=E8=8A=82=E7=82=B9=EF=BC=88=E5=88=86=E6=9E=90=E8=83=BD?=
 =?UTF-8?q?=E5=8A=9B=E4=B8=80=E8=BD=AE=EF=BC=89=E3=80=82=E8=8A=82=E7=82=B9?=
 =?UTF-8?q?=E5=BA=93=E3=80=8C=E6=8A=8A=E5=85=B3=E3=80=8D=E4=B8=80=E7=B1=BB?=
 =?UTF-8?q?=E5=8A=A0=E3=80=8C=E5=88=86=E6=9E=90=E5=8D=A1=E3=80=8D=EF=BC=88?=
 =?UTF-8?q?=E7=B1=BB=E5=9E=8B=E8=89=B2=20--nt-analysis=EF=BC=8C=E4=BA=AE?=
 =?UTF-8?q?=E6=9A=97=E4=B8=A4=E5=A5=97=EF=BC=89=EF=BC=8C=E5=8D=A1=E7=89=87?=
 =?UTF-8?q?=E6=91=98=E8=A6=81=E5=86=99=E5=81=9A=E5=93=AA=E5=87=A0=E7=A7=8D?=
 =?UTF-8?q?=E5=88=86=E6=9E=90=EF=BC=88=E3=80=8C=E6=B3=A2=E5=8A=A8=E5=9F=BA?=
 =?UTF-8?q?=E7=BA=BF=E3=80=81=E5=8F=98=E5=8C=96=E6=8B=86=E8=A7=A3=EF=BC=88?=
 =?UTF-8?q?=E5=8F=AF=E5=8A=A0=EF=BC=89=E3=80=8D=EF=BC=8C=E5=8F=AB=E6=B3=95?=
 =?UTF-8?q?=E7=85=A7=E5=A5=91=E7=BA=A6=EF=BC=89=EF=BC=9B=E6=A3=80=E6=9F=A5?=
 =?UTF-8?q?=E5=99=A8=E7=9A=84=E5=88=86=E6=9E=90=E9=A1=B9=E7=BC=96=E8=BE=91?=
 =?UTF-8?q?=E5=99=A8=EF=BC=9A=E6=AF=8F=E9=A1=B9=E6=A0=87=E8=AF=86=E3=80=81?=
 =?UTF-8?q?=E5=88=86=E6=9E=90=E6=96=B9=E6=B3=95=EF=BC=88=E4=BA=94=E4=B8=AA?=
 =?UTF-8?q?=E7=AE=97=E5=AD=90=EF=BC=89=E3=80=81=E6=A8=A1=E5=BC=8F=EF=BC=88?=
 =?UTF-8?q?=E6=9C=89=E6=A8=A1=E5=BC=8F=E7=9A=84=E6=89=8D=E7=BB=99=EF=BC=89?=
 =?UTF-8?q?=E3=80=81=E8=BE=93=E5=85=A5=E5=8F=AA=E5=88=97=E4=B8=8A=E6=B8=B8?=
 =?UTF-8?q?=E7=9A=84=E6=9F=A5=E8=AF=A2=E8=8A=82=E7=82=B9=EF=BC=88=E8=B0=83?=
 =?UTF-8?q?=E7=94=A8=E6=95=B0=E6=8D=AE=E5=BA=93=E6=9F=A5=E8=AF=A2=E5=B7=A5?=
 =?UTF-8?q?=E5=85=B7=E7=9A=84=E3=80=8C=E8=B0=83=E7=94=A8=E5=B7=A5=E5=85=B7?=
 =?UTF-8?q?=E3=80=8D=E8=8A=82=E7=82=B9=E3=80=81=E5=90=88=E5=B9=B6=E6=9F=A5?=
 =?UTF-8?q?=E8=AF=A2=EF=BC=8C=E5=88=A0=E6=8E=89=E6=88=96=E6=8C=AA=E5=88=B0?=
 =?UTF-8?q?=E4=B8=8B=E6=B8=B8=E7=9A=84=E7=85=A7=E5=AE=9E=E5=88=97=E5=87=BA?=
 =?UTF-8?q?=EF=BC=89=EF=BC=8C=E6=8C=89=E6=96=B9=E6=B3=95=E6=91=86=E5=8F=82?=
 =?UTF-8?q?=E6=95=B0=E6=A0=BC=EF=BC=88analysisItems.paramSpecs=EF=BC=9A?=
 =?UTF-8?q?=E7=BB=B4=E5=BA=A6=E5=88=97=E3=80=81=E6=9C=9F=E6=AC=A1=E3=80=81?=
 =?UTF-8?q?=E5=9F=BA=E6=9C=9F=E3=80=81=E6=9C=AC=E6=9C=9F=E3=80=81=E5=BA=A6?=
 =?UTF-8?q?=E9=87=8F=E5=88=97=E6=88=96=E5=88=86=E5=AD=90=E5=88=86=E6=AF=8D?=
 =?UTF-8?q?=E3=80=81=E4=B9=98=E7=A7=AF=E5=92=8C=E5=9B=A0=E5=AD=90=E3=80=81?=
 =?UTF-8?q?=E6=AF=94=E8=BE=83=E6=96=B9=E5=BC=8F=E3=80=81=E7=B2=92=E5=BA=A6?=
 =?UTF-8?q?=E3=80=81=E5=AF=B9=E7=85=A7=E7=BB=84=E3=80=81=E5=87=A0=E5=AF=B9?=
 =?UTF-8?q?=E2=80=A6=E2=80=A6=EF=BC=8C=E6=9C=9F=E6=AC=A1=E5=8F=AF=E4=BB=A5?=
 =?UTF-8?q?=E5=86=99=20{{=20}}=EF=BC=89=EF=BC=8C=E5=8F=AF=E5=8A=A0?=
 =?UTF-8?q?=E3=80=81=E6=AF=94=E7=8E=87=E7=9A=84=E5=8F=98=E5=8C=96=E6=8B=86?=
 =?UTF-8?q?=E8=A7=A3=E5=8F=A6=E6=9C=89=E6=80=BB=E6=95=B0=E6=A0=B8=E5=AF=B9?=
 =?UTF-8?q?=E4=B8=A4=E6=A0=BC=E5=92=8C=E4=B8=8B=E9=92=BB=EF=BC=88=E4=B8=8A?=
 =?UTF-8?q?=E4=B8=80=E5=B1=82=E5=8F=AA=E5=88=97=E6=8E=92=E5=9C=A8=E5=89=8D?=
 =?UTF-8?q?=E9=9D=A2=E7=9A=84=E5=8F=98=E5=8C=96=E6=8B=86=E8=A7=A3=EF=BC=89?=
 =?UTF-8?q?=EF=BC=8C=E5=8F=98=E5=8C=96=E6=8B=86=E8=A7=A3=E3=80=81=E5=8F=AF?=
 =?UTF-8?q?=E6=AF=94=E6=80=A7=E6=A3=80=E6=9F=A5=E5=8F=AF=E4=BB=A5=E5=86=99?=
 =?UTF-8?q?=E6=88=90=E5=9F=BA=E6=9C=9F=E3=80=81=E6=9C=AC=E6=9C=9F=E4=B8=A4?=
 =?UTF-8?q?=E4=BB=BD=E6=9F=A5=E8=AF=A2=EF=BC=8C=E6=8D=A2=E6=96=B9=E6=B3=95?=
 =?UTF-8?q?=E3=80=81=E6=A8=A1=E5=BC=8F=E6=97=B6=E5=8F=AA=E7=95=99=E6=96=B0?=
 =?UTF-8?q?=E6=96=B9=E6=B3=95=E7=94=A8=E5=BE=97=E5=88=B0=E7=9A=84=E5=8F=82?=
 =?UTF-8?q?=E6=95=B0=EF=BC=9B=E5=9C=A8=E9=AA=8C=E8=AF=81=E4=BB=80=E4=B9=88?=
 =?UTF-8?q?=E4=B8=80=E5=8F=A5=E8=AF=9D=EF=BC=9B=E9=98=88=E5=80=BC=E8=A6=86?=
 =?UTF-8?q?=E7=9B=96=E5=8F=AA=E5=88=97=E5=86=99=E4=BA=86=E7=9A=84=E5=87=A0?=
 =?UTF-8?q?=E9=A1=B9=E3=80=81=E4=BB=8E=E6=B2=A1=E5=86=99=E7=9A=84=E9=87=8C?=
 =?UTF-8?q?=E9=80=89=EF=BC=8C=E8=8C=83=E5=9B=B4=E5=BD=93=E5=9C=BA=E6=8C=89?=
 =?UTF-8?q?=E5=A5=91=E7=BA=A6=E8=AF=B4=EF=BC=88=E4=B8=A4=E4=B8=A4=E4=B9=8B?=
 =?UTF-8?q?=E9=97=B4=E7=9A=84=E5=A4=A7=E5=B0=8F=E4=BA=A4=E7=BB=99=E6=9C=8D?=
 =?UTF-8?q?=E5=8A=A1=E7=AB=AF=EF=BC=89=EF=BC=8C=E5=85=A8=E5=88=A0=E6=8E=89?=
 =?UTF-8?q?=E4=B8=8D=E7=95=99=E7=A9=BA=E5=AF=B9=E8=B1=A1=E3=80=82=E6=9C=8D?=
 =?UTF-8?q?=E5=8A=A1=E7=AB=AF=E6=8C=89=20items[i].<=E9=94=AE>=20=E6=8A=A5?=
 =?UTF-8?q?=E7=9A=84=E9=97=AE=E9=A2=98=E8=90=BD=E5=9C=A8=E9=82=A3=E4=B8=80?=
 =?UTF-8?q?=E6=A0=BC=E4=B8=8B=E9=9D=A2=EF=BC=8C=E5=AE=9A=E4=BD=8D=E9=80=80?=
 =?UTF-8?q?=E4=B8=8D=E5=88=B0=E7=9A=84=E5=AD=90=E9=94=AE=EF=BC=88params.pa?=
 =?UTF-8?q?rent.key=EF=BC=89=E9=80=90=E6=AE=B5=E5=BE=80=E4=B8=8A=E9=80=80?=
 =?UTF-8?q?=EF=BC=9B=E6=9F=A5=E8=AF=A2=E8=8A=82=E7=82=B9=E8=BF=9E=E5=88=B0?=
 =?UTF-8?q?=E5=88=86=E6=9E=90=E5=8D=A1=E6=97=B6=E8=BF=98=E6=B2=A1=E9=80=89?=
 =?UTF-8?q?=E8=BE=93=E5=85=A5=E7=9A=84=E7=AC=AC=E4=B8=80=E9=A1=B9=E9=80=89?=
 =?UTF-8?q?=E4=B8=8A=E5=AE=83=EF=BC=88=E5=90=8C=E5=90=88=E5=B9=B6=E6=9F=A5?=
 =?UTF-8?q?=E8=AF=A2=EF=BC=8C=E5=92=8C=E8=BF=9E=E7=BA=BF=E5=90=8C=E4=B8=80?=
 =?UTF-8?q?=E6=AD=A5=E6=92=A4=E9=94=80=EF=BC=89=E3=80=82=E6=8A=A5=E5=91=8A?=
 =?UTF-8?q?=E8=8A=82=E7=82=B9=E5=8A=A0=E3=80=8C=E5=88=86=E6=9E=90=E8=AF=B4?=
 =?UTF-8?q?=E6=B3=95=E6=A0=B8=E5=AF=B9=E3=80=8D=EF=BC=88=E7=BC=BA=E7=9C=81?=
 =?UTF-8?q?=E9=87=8D=E5=86=99=E4=B8=80=E6=AC=A1=E5=90=8E=E6=A0=87=E5=87=BA?=
 =?UTF-8?q?=EF=BC=89=E3=80=81=E3=80=8C=E5=88=86=E6=9E=90=E8=8C=83=E5=9B=B4?=
 =?UTF-8?q?=E3=80=8D=EF=BC=88=E9=AB=98=E7=BA=A7=EF=BC=8C=E7=BC=BA=E7=9C=81?=
 =?UTF-8?q?=E6=9C=89=E5=88=86=E6=9E=90=E7=BB=93=E8=AE=BA=E6=97=B6=E9=99=84?=
 =?UTF-8?q?=E4=B8=8A=EF=BC=89=EF=BC=8C=E9=80=89=E9=A1=B9=E4=BB=A5=20engine?=
 =?UTF-8?q?/labels.py=20=E7=9A=84=E7=9F=AD=E8=AF=B4=E6=B3=95=E5=BC=80?=
 =?UTF-8?q?=E5=A4=B4=E3=80=82=E5=8F=AB=E6=B3=95=E3=80=81=E8=8C=83=E5=9B=B4?=
 =?UTF-8?q?=E5=8F=AA=E7=85=A7=E5=A5=91=E7=BA=A6=E4=B8=80=E4=BB=BD=EF=BC=88?=
 =?UTF-8?q?lib/termsAnalysis=EF=BC=9AOP=5FLABELS=20/=20MODE=5FLABELS=20/?=
 =?UTF-8?q?=20LABELS=20/=20RANGES=EF=BC=8Ccheck-studio=20=E8=AF=BB?=
 =?UTF-8?q?=E5=90=8E=E7=AB=AF=E6=BA=90=E7=A0=81=E9=80=90=E9=A1=B9=E6=AF=94?=
 =?UTF-8?q?=E5=AF=B9=EF=BC=89=E3=80=82check-studio=20=E5=8A=A0=E4=B8=A4?=
 =?UTF-8?q?=E8=8A=82=EF=BC=9A=E5=8F=AB=E6=B3=95=E4=B8=8E=E5=A5=91=E7=BA=A6?=
 =?UTF-8?q?=E4=B8=80=E8=87=B4=E3=80=81=E5=88=86=E6=9E=90=E5=8D=A1=E7=9A=84?=
 =?UTF-8?q?=E7=94=BB=E5=B8=83=E7=BC=96=E8=BE=91=E4=B8=8E=E6=A0=A1=E9=AA=8C?=
 =?UTF-8?q?=E5=AE=9A=E4=BD=8D?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 frontend/src/canvas/Inspector.tsx      | 380 ++++++++++++++++++++++--
 frontend/src/canvas/InspectorSheet.tsx |   9 +-
 frontend/src/canvas/NodeCard.tsx       |   4 +
 frontend/src/canvas/analysisItems.ts   | 165 +++++++++++
 frontend/src/canvas/nodeDefs.ts        |  41 ++-
 frontend/src/index.css                 |   4 +
 frontend/src/lib/terms.ts              |   2 +
 frontend/src/lib/termsAnalysis.ts      | 391 +++++++++++++++++++++++++
 frontend/src/store/studio.ts           |   6 +-
 frontend/src/types.ts                  | 258 +++++++++++++++-
 scripts/check-studio.mjs               | 226 ++++++++++++++
 11 files changed, 1454 insertions(+), 32 deletions(-)
 create mode 100644 frontend/src/canvas/analysisItems.ts
 create mode 100644 frontend/src/lib/termsAnalysis.ts

diff --git a/frontend/src/canvas/Inspector.tsx b/frontend/src/canvas/Inspector.tsx
index d0a099e8..e67938bc 100644
--- a/frontend/src/canvas/Inspector.tsx
+++ b/frontend/src/canvas/Inspector.tsx
@@ -15,7 +15,15 @@ import { formatShortcut } from '../lib/keys'
 import { INPUT_FIELD_TEXT, JUDGE_FIELD_LABEL, JUDGE_ON_UNSUPPORTED_HINT, JUDGE_ON_UNSUPPORTED_LABEL, JUDGE_SETTING_TEXT, MERGE_TEXT, CATALOG_UI_TEXT, NODE_TYPE_LABEL, PROVIDER_REF_TEXT, SUBGRAPH_UPGRADE_HELP, judgeByModelText, upgradeNewerText } from '../lib/terms'
 import { TemplateText } from './TemplateText'
 import { MERGE_ALIAS, isQueryNode, mergeAlias } from './mergeInputs'
-import type { NodeType, Provider, ValidationIssue, WorkflowVersion } from '../types'
+import {
+  ITEM_ID, TWO_QUERY_SKIP, allowsHeadline, allowsParent, allowsTwoQueries, itemsOf, nextItemId, paramSpecs, paramText,
+  paramValue, withMode, withOp, type ParamSpec,
+} from './analysisItems'
+import {
+  ANALYSIS_MODES, ANALYSIS_MODE_LABEL, ANALYSIS_OPS, ANALYSIS_OP_LABEL, ANALYSIS_SETTING_KEYS, ANALYSIS_SETTING_LABEL,
+  ANALYSIS_TEXT, isAnalysisOp, settingRangeText, settingValueProblem, type AnalysisSettingKey,
+} from '../lib/termsAnalysis'
+import type { AnalysisItem, NodeType, Provider, ValidationIssue, WorkflowVersion } from '../types'
 
 /** 检查器里一条落到字段上的问题 */
 type FieldIssue = ValidationIssue & { at: FieldRef | null }
@@ -151,6 +159,34 @@ function NodeRefsInput({ nodeId, type, value, onChange }: {
   )
 }
 
+/** 上游能当查询输入的节点（选了数据库查询工具的「调用工具」节点、合并查询）：合并查询、分析卡共用 */
+function useUpstreamQueries(nodeId: string): { value: string; label: string; hint?: string }[] {
+  const nodes = useStudio((s) => s.nodes)
+  const edges = useStudio((s) => s.edges)
+  return useMemo(() => {
+    const parents = new Map<string, string[]>()
+    for (const e of edges) parents.set(e.target, [...(parents.get(e.target) ?? []), e.source])
+    const seen = new Set<string>()
+    const stack = [...(parents.get(nodeId) ?? [])]
+    while (stack.length) {
+      const id = stack.pop()!
+      if (seen.has(id) || id === nodeId) continue
+      seen.add(id)
+      stack.push(...(parents.get(id) ?? []))
+    }
+    return nodes
+      .filter((n) => seen.has(n.id) && isQueryNode(n))
+      .map((n) => {
+        const label = String(n.data?.label || n.id)
+        const tool = String((n.data?.config as Record<string, any> | undefined)?.tool ?? '')
+        return {
+          value: n.id, label: label === n.id ? label : `${label}（${n.id}）`,
+          hint: n.data?.nodeType === 'merge' ? NODE_TYPE_LABEL.merge : tool,
+        }
+      })
+  }, [nodes, edges, nodeId])
+}
+
 interface MergeRow { alias: string; node: string }
 
 /** config.inputs（{别名: 节点 id}）→ 行。对象的键顺序就是行的顺序 */
@@ -178,28 +214,7 @@ function MergeInputList({ nodeId, value, issues, onChange }: {
   const nodes = useStudio((s) => s.nodes)
   const edges = useStudio((s) => s.edges)
   const uid = useId()
-  const options = useMemo(() => {
-    const parents = new Map<string, string[]>()
-    for (const e of edges) parents.set(e.target, [...(parents.get(e.target) ?? []), e.source])
-    const seen = new Set<string>()
-    const stack = [...(parents.get(nodeId) ?? [])]
-    while (stack.length) {
-      const id = stack.pop()!
-      if (seen.has(id) || id === nodeId) continue
-      seen.add(id)
-      stack.push(...(parents.get(id) ?? []))
-    }
-    return nodes
-      .filter((n) => seen.has(n.id) && isQueryNode(n))
-      .map((n) => {
-        const label = String(n.data?.label || n.id)
-        const tool = String((n.data?.config as Record<string, any> | undefined)?.tool ?? '')
-        return {
-          value: n.id, label: label === n.id ? label : `${label}（${n.id}）`,
-          hint: n.data?.nodeType === 'merge' ? NODE_TYPE_LABEL.merge : tool,
-        }
-      })
-  }, [nodes, edges, nodeId])
+  const options = useUpstreamQueries(nodeId)
 
   const external = useMemo(() => mergeRowsOf(value), [value])
   const [rows, setRows] = useState<MergeRow[]>(external)
@@ -295,6 +310,319 @@ function MergeInputList({ nodeId, value, issues, onChange }: {
   )
 }
 
+const AT = ANALYSIS_TEXT.node
+
+/** 选一个上游查询节点。已经选了、后来删掉或挪到下游的节点照样列出来并说明，好让人看见并改掉 */
+function QueryPick({ id, value, options, onChange }: {
+  id: string; value: string; options: { value: string; label: string }[]; onChange: (v: string) => void
+}) {
+  const known = options.some((o) => o.value === value)
+  return (
+    <select id={id} className="field text-xs" value={value} onChange={(e) => onChange(e.target.value)}>
+      <option value="" disabled>{AT.pickFrom}</option>
+      {options.map((o) => <option key={o.value} value={o.value}>{o.label}</option>)}
+      {value && !known && <option value={value} disabled>{AT.missingNode(value)}</option>}
+    </select>
+  )
+}
+
+/**
+ * 分析卡的分析项（engine/schema.analysis_issues 的配置）：每一项一个标识、一种分析方法（有模式的再选模式）、要分析的上游
+ * 查询节点（或基期、本期两份查询），再按方法摆参数格（analysisItems.paramSpecs）；可加、比率的变化拆解另有总数核对、下钻。
+ * 阈值不在这里写（统一写在阈值覆盖里）。服务端按 items[i].<键> 报的问题落在那一格下面，定位按 [data-item] [data-sub] 落进来
+ */
+function AnalysisItemList({ nodeId, value, issues, onChange }: {
+  nodeId: string; value: any; issues: FieldIssue[]; onChange: (v: any) => void
+}) {
+  const uid = useId()
+  const options = useUpstreamQueries(nodeId)
+  const items = itemsOf(value)
+  const commit = (next: AnalysisItem[]) => onChange(next)
+  const update = (i: number, next: AnalysisItem) => commit(items.map((it, j) => (j === i ? next : it)))
+  const setParam = (i: number, key: string, v: unknown) => {
+    const params = { ...(items[i].params ?? {}) }
+    if (v === undefined) delete params[key]
+    else params[key] = v
+    update(i, { ...items[i], params })
+  }
+  const add = () => commit([...items, { id: nextItemId(items), params: {} }])
+  return (
+    <div>
+      {items.map((item, i) => {
+        const own = issues.filter((x) => x.at?.index === i)
+        const at = (sub: string) => own.filter((x) => x.at?.sub === sub || x.at?.sub?.startsWith(`${sub}.`))
+        const lines = (sub: string) => at(sub).map((x, k) => <IssueLine key={k} issue={x} />)
+        const op = isAnalysisOp(item.op) ? item.op : null
+        const modes = op ? ANALYSIS_MODES[op] : undefined
+        const params = (item.params && typeof item.params === 'object' ? item.params : {}) as Record<string, any>
+        const two = !!op && allowsTwoQueries(op) && ('base_query' in params || 'target_query' in params)
+        const id = String(item.id ?? '')
+        const dup = !!id && items.some((other, j) => j !== i && String(other.id ?? '') === id)
+        const local = !id || !ITEM_ID.test(id) ? AT.idInvalid : dup ? AT.idDuplicate(id) : ''
+        const tone = own.some((x) => x.level === 'error') || local ? 'error' : own.length ? 'warning' : undefined
+        const before = items.slice(0, i).filter((it) => it.op === 'decompose' && it.id)
+        const parent = params.parent && typeof params.parent === 'object' ? params.parent as { of?: string; key?: unknown } : null
+        const field = (spec: ParamSpec) => {
+          const fid = `${uid}-${i}-p-${spec.key}`
+          const label = AT.param[spec.key] ?? spec.key
+          const text = paramText(spec.kind, params[spec.key])
+          const set = (t: string) => setParam(i, spec.key, paramValue(spec.kind, t))
+          let control
+          if (spec.kind === 'basis' || spec.kind === 'grain') {
+            const opts = spec.kind === 'basis' ? AT.basis : AT.grain
+            control = (
+              <select id={fid} className="field text-xs" value={text} onChange={(e) => set(e.target.value)}>
+                {Object.entries(opts).filter(([k]) => !(spec.kind === 'basis' && k === 'change')).map(([k, v]) => (
+                  <option key={k} value={k}>{v}</option>
+                ))}
+                {text && !(text in opts) && <option value={text} disabled>{text}</option>}
+              </select>
+            )
+          } else if (spec.kind === 'pairs') {
+            control = <textarea id={fid} className="field mono text-xs" rows={2} value={text} spellCheck={false}
+                                placeholder="华东、华南" onChange={(e) => set(e.target.value)} />
+          } else if (spec.kind === 'value') {
+            // 期次、对照组、叫法可以按运行输入写 {{ }}（执行时先渲染再分析）
+            control = <TemplateText id={fid} multiline={false} nodeId={nodeId} className="text-xs" value={text}
+                                    placeholder={AT.paramPlaceholder[spec.key]} onChange={set} />
+          } else {
+            control = <input id={fid} className="field mono text-xs" value={text} spellCheck={false}
+                             inputMode={spec.kind === 'int' ? 'numeric' : undefined}
+                             placeholder={AT.paramPlaceholder[spec.key]} onChange={(e) => set(e.target.value)} />
+          }
+          const hint = spec.kind === 'columns' ? AT.listHint : spec.kind === 'pairs' ? AT.pairsHint : AT.paramHint[spec.key]
+          return (
+            <div key={spec.key} className="min-w-0">
+              <Sub label={label} htmlFor={fid} sub={`params.${spec.key}`}>
+                {control}
+                {hint && <div className="mt-0.5 text-2xs leading-snug text-faint">{hint}</div>}
+              </Sub>
+              {lines(`params.${spec.key}`)}
+            </div>
+          )
+        }
+        const specs = op ? paramSpecs(op, item.mode).filter((p) => !(two && TWO_QUERY_SKIP.has(p.key))) : []
+        const headline = params.headline && typeof params.headline === 'object' ? params.headline as Record<string, string> : null
+        const setHeadline = (side: 'base' | 'target', t: string) => {
+          const next = { base: headline?.base ?? '', target: headline?.target ?? '', [side]: t.trim() }
+          setParam(i, 'headline', next.base || next.target ? next : undefined)
+        }
+        const toggleTwo = (on: boolean) => {
+          const next = { ...params }
+          if (on) {
+            next.base_query = item.from ?? ''
+            next.target_query = ''
+            for (const k of TWO_QUERY_SKIP) delete next[k]
+            update(i, { ...item, from: undefined, params: next })
+          } else {
+            const from = typeof next.base_query === 'string' ? next.base_query : ''
+            delete next.base_query
+            delete next.target_query
+            update(i, { ...item, from, params: next })
+          }
+        }
+        // 落在某一格下面的问题在那一格下面说，其余的（格式不对、多写了这种分析没有的参数……）放在这一项的最后
+        const placed = ['id', 'op', 'mode', 'hypothesis', ...(two ? ['params.base_query', 'params.target_query'] : ['from']),
+          ...specs.map((p) => `params.${p.key}`),
+          ...(op && allowsHeadline(op, item.mode) ? ['params.headline'] : []),
+          ...(op && allowsParent(op, item.mode) && (before.length > 0 || parent) ? ['params.parent'] : [])]
+        const rest = own.filter((x) => !placed.some((sub) => x.at?.sub === sub || x.at?.sub?.startsWith(`${sub}.`)))
+        return (
+          <Row key={i} item={i} removeLabel={AT.removeItem(id)} tone={tone}
+               onRemove={() => commit(items.filter((_, j) => j !== i))}>
+            <div className="flex gap-1.5" data-analysis-item={id || i}>
+              <Sub label={AT.id} htmlFor={`${uid}-${i}-id`} className="w-20 shrink-0" sub="id">
+                <input id={`${uid}-${i}-id`} className="field mono text-xs" value={id} spellCheck={false}
+                       aria-invalid={!!local || undefined}
+                       onChange={(e) => update(i, { ...item, id: e.target.value.trim() })} />
+              </Sub>
+              <Sub label={AT.op} htmlFor={`${uid}-${i}-op`} className="min-w-0 flex-1" sub="op">
+                <select id={`${uid}-${i}-op`} className="field text-xs" value={item.op ?? ''}
+                        onChange={(e) => update(i, withOp(item, e.target.value))}>
+                  <option value="" disabled>{AT.pickOp}</option>
+                  {ANALYSIS_OPS.map((o) => <option key={o} value={o}>{ANALYSIS_OP_LABEL[o]}</option>)}
+                  {!!item.op && !op && <option value={item.op} disabled>{AT.unknownOp(item.op)}</option>}
+                </select>
+              </Sub>
+              {modes && (
+                <Sub label={AT.mode} htmlFor={`${uid}-${i}-mode`} className="w-[5.5rem] shrink-0" sub="mode">
+                  <select id={`${uid}-${i}-mode`} className="field text-xs" value={item.mode ?? ''}
+                          onChange={(e) => update(i, withMode(item, e.target.value))}>
+                    <option value="" disabled>{AT.pickMode}</option>
+                    {modes.map((m) => <option key={m} value={m}>{ANALYSIS_MODE_LABEL[m]}</option>)}
+                    {!!item.mode && !modes.includes(item.mode) && <option value={item.mode} disabled>{AT.unknownMode(item.mode)}</option>}
+                  </select>
+                </Sub>
+              )}
+            </div>
+            {local && !at('id').length && (
+              <div className="flex items-start gap-1 text-2xs leading-snug" style={{ color: 'var(--err)' }}>
+                <XCircle size={11} className="mt-px shrink-0" aria-hidden />
+                <span className="min-w-0 [overflow-wrap:anywhere]">{local}</span>
+              </div>
+            )}
+            {lines('id')}
+            {lines('op')}
+            {lines('mode')}
+            {id && ITEM_ID.test(id) && (
+              <div className="mono truncate text-2xs text-faint" title={AT.idHint(nodeId, id)}>{AT.idHint(nodeId, id)}</div>
+            )}
+            {two ? (
+              <div className="grid grid-cols-2 gap-1.5">
+                <div className="min-w-0">
+                  <Sub label={AT.baseQuery} htmlFor={`${uid}-${i}-bq`} sub="params.base_query">
+                    <QueryPick id={`${uid}-${i}-bq`} value={String(params.base_query ?? '')} options={options}
+                               onChange={(v) => setParam(i, 'base_query', v)} />
+                  </Sub>
+                  {lines('params.base_query')}
+                </div>
+                <div className="min-w-0">
+                  <Sub label={AT.targetQuery} htmlFor={`${uid}-${i}-tq`} sub="params.target_query">
+                    <QueryPick id={`${uid}-${i}-tq`} value={String(params.target_query ?? '')} options={options}
+                               onChange={(v) => setParam(i, 'target_query', v)} />
+                  </Sub>
+                  {lines('params.target_query')}
+                </div>
+              </div>
+            ) : (
+              <div>
+                <Sub label={AT.from} htmlFor={`${uid}-${i}-from`} sub="from">
+                  <QueryPick id={`${uid}-${i}-from`} value={String(item.from ?? '')} options={options}
+                             onChange={(v) => update(i, { ...item, from: v })} />
+                </Sub>
+                {lines('from')}
+              </div>
+            )}
+            {op && allowsTwoQueries(op) && (
+              <label className="flex items-center gap-1.5 text-2xs text-dim">
+                <input type="checkbox" checked={two} onChange={(e) => toggleTwo(e.target.checked)} data-analysis-two="" />
+                {AT.twoQueries}
+              </label>
+            )}
+            {!options.length && <div className="text-2xs leading-snug text-faint">{AT.noUpstream}</div>}
+            {specs.length > 0 && (
+              <div className="grid grid-cols-2 gap-1.5" data-analysis-params="">
+                {specs.map(field)}
+              </div>
+            )}
+            {op && allowsHeadline(op, item.mode) && (
+              <div data-sub="params.headline">
+                <div className="mb-0.5 text-2xs text-faint">{AT.headline}</div>
+                <div className="grid grid-cols-2 gap-1.5">
+                  {(['base', 'target'] as const).map((side) => (
+                    <Sub key={side} label={side === 'base' ? AT.headlineBase : AT.headlineTarget}
+                         htmlFor={`${uid}-${i}-h-${side}`} sub={`params.headline.${side}`}>
+                      <input id={`${uid}-${i}-h-${side}`} className="field mono text-xs" spellCheck={false}
+                             value={headline?.[side] ?? ''} placeholder={side === 'base' ? 'weekly.r0.gmv' : 'weekly.r1.gmv'}
+                             onChange={(e) => setHeadline(side, e.target.value)} />
+                    </Sub>
+                  ))}
+                </div>
+                <div className="mt-0.5 text-2xs leading-snug text-faint">{AT.headlineHint}。{AT.cellHint}</div>
+                {lines('params.headline')}
+              </div>
+            )}
+            {op && allowsParent(op, item.mode) && (before.length > 0 || parent) && (
+              <div data-sub="params.parent">
+                <div className="grid grid-cols-2 gap-1.5">
+                  <Sub label={AT.parentOf} htmlFor={`${uid}-${i}-po`}>
+                    <select id={`${uid}-${i}-po`} className="field text-xs" value={parent?.of ?? ''}
+                            onChange={(e) => setParam(i, 'parent', e.target.value
+                              ? { of: e.target.value, key: Array.isArray(parent?.key) ? parent!.key : [] } : undefined)}>
+                      <option value="">{AT.noParent}</option>
+                      {before.map((it) => <option key={String(it.id)} value={String(it.id)}>{String(it.id)}</option>)}
+                      {parent?.of && !before.some((it) => it.id === parent.of) && <option value={parent.of} disabled>{parent.of}</option>}
+                    </select>
+                  </Sub>
+                  {parent && (
+                    <Sub label={AT.parentKey} htmlFor={`${uid}-${i}-pk`}>
+                      <input id={`${uid}-${i}-pk`} className="field mono text-xs" spellCheck={false} placeholder="华东"
+                             value={paramText('columns', parent.key)}
+                             onChange={(e) => setParam(i, 'parent', { of: parent.of, key: paramValue('columns', e.target.value) ?? [] })} />
+                    </Sub>
+                  )}
+                </div>
+                <div className="mt-0.5 text-2xs leading-snug text-faint">{AT.parentHint}</div>
+                {lines('params.parent')}
+              </div>
+            )}
+            <div>
+              <Sub label={AT.hypothesis} htmlFor={`${uid}-${i}-hyp`} sub="hypothesis">
+                <input id={`${uid}-${i}-hyp`} className="field text-xs" value={item.hypothesis ?? ''}
+                       placeholder={AT.hypothesisPlaceholder}
+                       onChange={(e) => update(i, { ...item, hypothesis: e.target.value || undefined })} />
+              </Sub>
+              <div className="mt-0.5 text-2xs leading-snug text-faint">{AT.hypothesisHint}</div>
+              {lines('hypothesis')}
+            </div>
+            {rest.map((x, k) => <IssueLine key={k} issue={x} />)}
+          </Row>
+        )
+      })}
+      <button type="button" className="btn btn-sm w-full justify-center" onClick={add} data-analysis-add="">
+        <Plus size={11} /> {AT.addItem}
+      </button>
+    </div>
+  )
+}
+
+/**
+ * 分析卡的阈值覆盖：只列写了的几项，「添加阈值覆盖」从没写的里面选。键和叫法同设置里的分析一组（contract.LABELS），范围先在
+ * 这里按 contract.RANGES 说，两两之间的大小（「明显超出」要大于「超出常规波动」……）由服务端核、问题落在这一栏下面。
+ * 全部删掉写回 undefined：没覆盖就不留一个空对象
+ */
+function AnalysisOverrides({ value, issues, onChange }: { value: any; issues: FieldIssue[]; onChange: (v: any) => void }) {
+  const uid = useId()
+  const current: Record<string, unknown> = value && typeof value === 'object' && !Array.isArray(value) ? value : {}
+  // 正在填的字先记在本地：「0.」这种还不是数的字写回去会被转成别的
+  const [drafts, setDrafts] = useState<Record<string, string>>({})
+  const keys = Object.keys(current)
+  const free = ANALYSIS_SETTING_KEYS.filter((k) => !(k in current))
+  const write = (next: Record<string, unknown>) => onChange(Object.keys(next).length ? next : undefined)
+  const set = (key: string, text: string) => {
+    setDrafts((d) => ({ ...d, [key]: text }))
+    const t = text.trim()
+    write({ ...current, [key]: t !== '' && Number.isFinite(Number(t)) ? Number(t) : t })
+  }
+  const remove = (key: string) => {
+    const next = { ...current }
+    delete next[key]
+    setDrafts((d) => { const c = { ...d }; delete c[key]; return c })
+    write(next)
+  }
+  return (
+    <div data-analysis-overrides="">
+      {keys.map((key, i) => {
+        const known = (ANALYSIS_SETTING_KEYS as readonly string[]).includes(key)
+        const label = known ? ANALYSIS_SETTING_LABEL[key as AnalysisSettingKey] : key
+        const text = drafts[key] ?? (current[key] == null ? '' : String(current[key]))
+        const problem = known ? settingValueProblem(key as AnalysisSettingKey, text) : null
+        const fid = `${uid}-${i}`
+        return (
+          <Row key={key} removeLabel={AT.removeOverride(label)} tone={problem ? 'error' : undefined} onRemove={() => remove(key)}>
+            <Sub label={label} htmlFor={fid} sub={key}>
+              <input id={fid} className="field tnum text-xs" inputMode="decimal" value={text} aria-invalid={!!problem || undefined}
+                     data-analysis-override={key} onChange={(e) => set(key, e.target.value)} />
+            </Sub>
+            <div className="text-2xs leading-snug" style={{ color: problem ? 'var(--err)' : 'var(--text-faint)' }}>
+              {problem ?? (known ? settingRangeText(key as AnalysisSettingKey) : '')}
+            </div>
+          </Row>
+        )
+      })}
+      {!keys.length && <div className="mb-1.5 text-2xs text-faint">{AT.noOverrides}</div>}
+      {issues.map((x, k) => <IssueLine key={k} issue={x} />)}
+      {free.length > 0 ? (
+        <select className="field text-xs" value="" aria-label={AT.addOverride} data-analysis-add-override=""
+                onChange={(e) => { if (e.target.value) set(e.target.value, '') }}>
+          <option value="">{AT.addOverride}</option>
+          {free.map((k) => <option key={k} value={k}>{ANALYSIS_SETTING_LABEL[k]}</option>)}
+        </select>
+      ) : <div className="text-2xs text-faint">{AT.allOverridden}</div>}
+    </div>
+  )
+}
+
 /**
  * 子工作流钉住版本后的升版处置：和口径卡的是同一个下拉——直接借口径卡的字段定义，选项、
  * 文案一字不差，只换说明里「和谁同一套」那半句。上游在钉住的那一版之后又发了版时，正式运行
@@ -790,6 +1118,12 @@ function FieldInput({ field, id, nodeId, syntax, value, config, invalid, describ
     case 'mergeInputs':
       return <MergeInputList nodeId={nodeId} value={value} issues={issues} onChange={onChange} />
 
+    case 'analysisItems':
+      return <AnalysisItemList nodeId={nodeId} value={value} issues={issues} onChange={onChange} />
+
+    case 'analysisOverrides':
+      return <AnalysisOverrides value={value} issues={issues} onChange={onChange} />
+
     case 'caliberFrom':
       return <CaliberFromPicker id={id} value={value} invalid={invalid} describedBy={describedBy} />
 
diff --git a/frontend/src/canvas/InspectorSheet.tsx b/frontend/src/canvas/InspectorSheet.tsx
index 98746721..23fec316 100644
--- a/frontend/src/canvas/InspectorSheet.tsx
+++ b/frontend/src/canvas/InspectorSheet.tsx
@@ -67,8 +67,13 @@ export function revealField(nodeId: string, field?: FieldRef | string | null, op
     const sheet = document.querySelector('[data-inspector-sheet]') ?? document
     const box = sheet.querySelector<HTMLElement>(`[data-field="${CSS.escape(at.key)}"]`)
     const item = at.index != null ? box?.querySelector<HTMLElement>(`[data-item="${at.index}"]`) : null
-    // 只有子键没有序号的（合并查询的 inputs.<别名>）：在整个字段里按子键找那一行
-    const part = at.sub ? (item ?? box)?.querySelector<HTMLElement>(`[data-sub="${CSS.escape(at.sub)}"]`) : null
+    // 只有子键没有序号的（合并查询的 inputs.<别名>）：在整个字段里按子键找那一行。子键带点号的（分析项的 params.headline.base、
+    // params.parent.key）找不到那一格时逐段往上退，落到它所在的那一块（params.parent）
+    const scope = item ?? box
+    let part: HTMLElement | null | undefined = null
+    for (let sub = at.sub; sub && !part; sub = sub.includes('.') ? sub.slice(0, sub.lastIndexOf('.')) : '') {
+      part = scope?.querySelector<HTMLElement>(`[data-sub="${CSS.escape(sub)}"]`)
+    }
     const target = part ?? item ?? box
     if (!target) return
     const still = window.matchMedia('(prefers-reduced-motion: reduce)').matches
diff --git a/frontend/src/canvas/NodeCard.tsx b/frontend/src/canvas/NodeCard.tsx
index 5798cc83..b6065491 100644
--- a/frontend/src/canvas/NodeCard.tsx
+++ b/frontend/src/canvas/NodeCard.tsx
@@ -21,6 +21,7 @@ import {
   CHALLENGE_TEXT, claimTally, DUAL_CHECK_TEXT, issuanceLabel, REJECTED_TEXT, reportStampText, RUN_CLASS_LABEL, UPGRADE_POLICY_LABEL,
 } from '../lib/terms'
 import { useReportStamp, type ReportStamp } from './reportStamp'
+import { analysisCardSummary } from './analysisItems'
 import { ApprovalCard } from '../run/RunPanel'
 import type { NodeState, NodeTrace } from '../run/trace'
 import { useRunClock } from '../run/useRunClock'
@@ -86,6 +87,9 @@ function summarize(type: NodeType, config: Record<string, any>): string {
       }
       return `${config.caliber || '口径'} ${config.caliber_version || 'v1'} · ${(config.metrics ?? []).length} 个指标`
     }
+    case 'analysis':
+      // 做哪几种分析：「波动基线、变化拆解（可加）」，叫法照契约（lib/termsAnalysis）
+      return analysisCardSummary(config)
     case 'report': {
       const from = Array.isArray(config.metrics_from) ? config.metrics_from.length : 0
       return `${first(config.instructions) || '未填写作要求'} · ${from ? `${from} 张口径卡` : '上游全部口径卡'}`
diff --git a/frontend/src/canvas/analysisItems.ts b/frontend/src/canvas/analysisItems.ts
new file mode 100644
index 00000000..5bee9ee2
--- /dev/null
+++ b/frontend/src/canvas/analysisItems.ts
@@ -0,0 +1,165 @@
+// ===========================================================================
+// 分析卡（节点类型 analysis）的一项：每种分析在检查器里给哪些参数格、参数怎么在格子和配置之间转换、卡片上的摘要。
+//
+// 参数名和服务端 tools/builtin/analysis.py 各参数模型里的字段一致（分析卡按同一份参数模型校验，engine/schema.analysis_issues）：
+// 主输入写在 from（对应 Agent 工具的 query），基期、本期是两份查询时写在 params 的 base_query、target_query；阈值不写在参数里，
+// 统一写在卡上的阈值覆盖。这里只决定界面上摆哪几格、按什么先后，规则以服务端校验为准（写错了检查器照服务端的说法提示）。
+// ===========================================================================
+
+import { ANALYSIS_MODES, ANALYSIS_TEXT, analysisMethodLabel, isAnalysisOp, type AnalysisOp } from '../lib/termsAnalysis'
+import type { AnalysisItem } from '../types'
+
+/** 一格参数的写法：列名、几列（顿号、逗号分隔）、期次的值（可以写 {{ }}）、整数、两种下拉、几对 */
+export type ParamKind = 'column' | 'columns' | 'value' | 'int' | 'basis' | 'grain' | 'pairs'
+
+export interface ParamSpec {
+  key: string
+  kind: ParamKind
+}
+
+const col = (key: string): ParamSpec => ({ key, kind: 'column' })
+const cols = (key: string): ParamSpec => ({ key, kind: 'columns' })
+const val = (key: string): ParamSpec => ({ key, kind: 'value' })
+
+/** 两期写在同一份查询里时的期次三格（期次列、基期、本期） */
+const PERIODS: ParamSpec[] = [col('period'), val('base'), val('target')]
+
+/**
+ * 每种分析（算子 + 模式）摆哪几格，按这个先后。总数核对（headline）、下钻（parent）、两份查询另有自己的一块，不在这里。
+ * label（度量的叫法）放在最后：留空用列名或数据目录里的名字
+ */
+export function paramSpecs(op: AnalysisOp, mode?: string | null): ParamSpec[] {
+  switch (op) {
+    case 'decompose':
+      if (mode === 'ratio') return [cols('by'), ...PERIODS, col('numerator'), col('denominator'), val('label')]
+      if (mode === 'factor') return [...PERIODS, col('product'), cols('factors'), val('label')]
+      return [cols('by'), ...PERIODS, col('measure'), val('label')]
+    case 'baseline':
+      return [col('period'), col('measure'), val('target'), { key: 'basis', kind: 'basis' }, { key: 'grain', kind: 'grain' },
+        { key: 'lag', kind: 'int' }, val('label')]
+    case 'trend':
+      return [col('period'), col('measure'), { key: 'grain', kind: 'grain' }, val('label')]
+    case 'differ': {
+      const values = mode === 'mean' ? [col('mean'), col('sd'), col('sum'), col('sumsq'), col('n')]
+        : mode === 'rows' ? [col('value')] : [col('numerator'), col('denominator')]
+      return [cols('by'), val('reference'), { key: 'pairs', kind: 'pairs' }, ...values, ...(mode === 'rows' ? [] : PERIODS),
+        val('label')]
+    }
+    case 'comparable':
+      return [...PERIODS, col('measure'), { key: 'grain', kind: 'grain' }]
+  }
+}
+
+/** 这种分析能不能写成基期、本期两份查询（DecomposeArgs、ComparableArgs 有 base_query / target_query） */
+export const allowsTwoQueries = (op: AnalysisOp): boolean => op === 'decompose' || op === 'comparable'
+/** 有总数核对（headline）的：可加、比率的变化拆解 */
+export const allowsHeadline = (op: AnalysisOp, mode?: string | null): boolean =>
+  op === 'decompose' && mode !== 'factor'
+/** 能下钻（parent）的：可加、比率的变化拆解（乘积拆解不按维度拆） */
+export const allowsParent = allowsHeadline
+
+/** 两份查询时期次三格不写（两份各自只有一期） */
+export const TWO_QUERY_SKIP = new Set(['period', 'base', 'target'])
+
+/** 这种分析会写到的全部参数键：换分析方法、模式时，这几样以外的参数去掉（服务端按多写的参数报错） */
+export function paramKeys(op: AnalysisOp, mode?: string | null): Set<string> {
+  const keys = new Set(paramSpecs(op, mode).map((p) => p.key))
+  if (allowsTwoQueries(op)) { keys.add('base_query'); keys.add('target_query') }
+  if (allowsHeadline(op, mode)) keys.add('headline')
+  if (allowsParent(op, mode)) keys.add('parent')
+  return keys
+}
+
+/** 「region、channel」「region, channel」→ ['region', 'channel'] */
+export function splitList(text: string): string[] {
+  return text.split(/[、,，\s]+/).map((s) => s.trim()).filter(Boolean)
+}
+
+/** 配置里的一格 → 格子里的字 */
+export function paramText(kind: ParamKind, value: unknown): string {
+  if (value == null) return ''
+  if (kind === 'columns') return Array.isArray(value) ? value.map(String).join('、') : String(value)
+  if (kind === 'pairs') {
+    return Array.isArray(value)
+      ? value.map((p) => (Array.isArray(p) ? p.map(String).join('、') : String(p))).join('\n')
+      : String(value)
+  }
+  return String(value)
+}
+
+/** 格子里的字 → 配置里的一格；空的返回 undefined（参数里不留这个键） */
+export function paramValue(kind: ParamKind, text: string): unknown {
+  const t = text.trim()
+  if (!t) return undefined
+  if (kind === 'columns') return splitList(t)
+  if (kind === 'pairs') return t.split('\n').map((line) => splitList(line)).filter((p) => p.length)
+  // 整数格写了不是整数的字照原样交给服务端校验（它会说清该写什么），不在这里悄悄丢掉
+  if (kind === 'int') return /^-?\d+$/.test(t) ? Number(t) : t
+  return t
+}
+
+/** 配置里的 items → 一项一项（不是对象的项原样留着，服务端会报「格式有误」） */
+export function itemsOf(value: unknown): AnalysisItem[] {
+  return Array.isArray(value) ? value.map((v) => (v && typeof v === 'object' ? v as AnalysisItem : {})) : []
+}
+
+/** 标识的写法（服务端 schema.ANALYSIS_ITEM_ID）：下游按 nodes.<卡>.results.<标识> 取判定 */
+export const ITEM_ID = /^[A-Za-z_][A-Za-z0-9_]*$/
+
+/** 新加一项的标识：a1、a2……不和已有的重名 */
+export function nextItemId(items: AnalysisItem[]): string {
+  const taken = new Set(items.map((i) => String(i.id ?? '')))
+  for (let k = items.length + 1; ; k += 1) if (!taken.has(`a${k}`)) return `a${k}`
+}
+
+/** 换分析方法：有模式的取第一种模式，参数只留新方法用得到的，输入和假设留着 */
+export function withOp(item: AnalysisItem, op: string): AnalysisItem {
+  if (!isAnalysisOp(op)) return { ...item, op, mode: undefined }
+  const mode = ANALYSIS_MODES[op]?.[0]
+  return withParams({ ...item, op, mode }, op, mode)
+}
+
+/** 换模式：参数只留这种模式用得到的 */
+export function withMode(item: AnalysisItem, mode: string): AnalysisItem {
+  return isAnalysisOp(item.op) ? withParams({ ...item, mode }, item.op, mode) : { ...item, mode }
+}
+
+function withParams(item: AnalysisItem, op: AnalysisOp, mode: string | null | undefined): AnalysisItem {
+  const keep = paramKeys(op, mode)
+  const params = Object.fromEntries(Object.entries(item.params ?? {}).filter(([k]) => keep.has(k)))
+  const next: AnalysisItem = { ...item, params }
+  if (!ANALYSIS_MODES[op]) delete next.mode
+  return next
+}
+
+/** 卡片正文的摘要：「波动基线、变化拆解（可加）」，超过三项写「等 N 项」 */
+export function analysisCardSummary(config: Record<string, any>): string {
+  const items = itemsOf(config.items)
+  if (!items.length) return ANALYSIS_TEXT.node.summaryEmpty
+  const names = items.map((i) => (i.op ? analysisMethodLabel(i.op, i.mode) : ANALYSIS_TEXT.node.summaryNoOp))
+  const head = names.slice(0, 3).join('、')
+  return names.length > 3 ? `${head}${ANALYSIS_TEXT.node.summaryMore(names.length)}` : head
+}
+
+interface CardNodeLike {
+  id: string
+  data?: { nodeType?: string; label?: string; config?: Record<string, any> }
+}
+
+/**
+ * 查询节点连到分析卡：第一项还没选输入（也不是两份查询）的分析项顺手选上它，和连线同一步撤销（同合并查询的 connectMergeInput）。
+ * 已经有项用着它、或者每一项都选好了输入时不动
+ */
+export function connectAnalysisInput<T extends CardNodeLike>(nodes: T[], source: string | null, target: string | null,
+  isQuery: (n: T | undefined) => boolean): T[] | null {
+  if (!source || !target || source === target) return null
+  const to = nodes.find((n) => n.id === target)
+  if (to?.data?.nodeType !== 'analysis' || !isQuery(nodes.find((n) => n.id === source))) return null
+  const items = itemsOf(to.data?.config?.items)
+  const uses = (it: AnalysisItem) => [it.from, it.params?.base_query, it.params?.target_query].includes(source)
+  if (items.some(uses)) return null
+  const i = items.findIndex((it) => !String(it.from ?? '').trim() && !('base_query' in (it.params ?? {})))
+  if (i < 0) return null
+  const next = items.map((it, j) => (j === i ? { ...it, from: source } : it))
+  return nodes.map((n) => (n.id === target ? { ...n, data: { ...n.data, config: { ...(n.data?.config ?? {}), items: next } } } : n))
+}
diff --git a/frontend/src/canvas/nodeDefs.ts b/frontend/src/canvas/nodeDefs.ts
index 3f3d5939..87d15d74 100644
--- a/frontend/src/canvas/nodeDefs.ts
+++ b/frontend/src/canvas/nodeDefs.ts
@@ -1,11 +1,12 @@
 import type { ReactNode } from 'react'
 import {
-  Bot, Braces, Brain, CheckCircle2, Code2, Combine, Database, FileInput, FileOutput, FileText,
+  Bot, Braces, Brain, ChartColumnStacked, CheckCircle2, Code2, Combine, Database, FileInput, FileOutput, FileText,
   Gauge, GitBranch, GitCompareArrows, Hand, Repeat, Search, ShieldAlert, Shuffle, Users, Wrench,
 } from 'lucide-react'
 import {
   APPROVAL_POLICY_LABEL, CHALLENGE_TEXT, DUAL_CHECK_TEXT, MERGE_TEXT, NODE_TYPE_LABEL, REJECTED_TEXT, UPGRADE_POLICY_LABEL,
 } from '../lib/terms'
+import { ANALYSIS_FIELD_LABEL, ANALYSIS_OPTION_LABEL, ANALYSIS_TEXT } from '../lib/termsAnalysis'
 import type { NodeType } from '../types'
 import { dualCheckFields, dualCheckSummary } from './DualCheckFields'
 import { HUMAN_PROCESS_FIELDS, TOOL_WRITE_FIELDS } from './humanFields'
@@ -15,6 +16,7 @@ export type FieldType =
   | 'text' | 'textarea' | 'prompt' | 'code' | 'number' | 'select' | 'switch'
   | 'json' | 'model' | 'tools' | 'skills' | 'collection'
   | 'ioFields' | 'cases' | 'agents' | 'metricsList' | 'nodeRefs' | 'caliberFrom' | 'judge' | 'mergeInputs'
+  | 'analysisItems' | 'analysisOverrides'
 
 /**
  * 这个字段里写的是什么语法。
@@ -654,6 +656,29 @@ export const NODE_DEFS: Record<NodeType, NodeDef> = {
           + '反引号中无从核对的名称标为「疑似不存在的名称」',
         unknownLabel: (v) => `${v}：无法识别的值，请从上方选项中选择`,
       },
+      {
+        // 后端认 off / check / strict（schema.ANALYSIS_CLAIMS_VALUES），没写按 check：缺省放第一项，没写时下拉正好落在它上面。
+        // 选项以 engine/labels.py 的短说法开头
+        key: 'analysis_claims', label: ANALYSIS_FIELD_LABEL.analysis_claims, type: 'select',
+        options: [
+          { value: 'check', label: `${ANALYSIS_OPTION_LABEL.analysis_claims.check}（默认）：仍有的在报告中标注，不影响出具档位` },
+          { value: 'strict', label: `${ANALYSIS_OPTION_LABEL.analysis_claims.strict}：按最多重写次数重写，出具随之降档` },
+          { value: 'off', label: ANALYSIS_OPTION_LABEL.analysis_claims.off },
+        ],
+        help: '「主要来自」「超出常规波动」「持续上升」这类判断要挂对应的分析结论且判定对得上；「由于」「导致」这类原因要有'
+          + '知识库原话或运行输入作依据，否则写成待核实的推测。受管级别发布须选择「重写后仍有的计入缺口」',
+        unknownLabel: (v) => `${v}：无法识别的值，请从上方选项中选择`,
+      },
+      {
+        // 后端认 auto / off（schema.ANALYSIS_SCOPE_VALUES），没写按 auto
+        key: 'analysis_scope', label: ANALYSIS_FIELD_LABEL.analysis_scope, type: 'select', advanced: true,
+        options: [
+          { value: 'auto', label: `${ANALYSIS_OPTION_LABEL.analysis_scope.auto}（默认）` },
+          { value: 'off', label: ANALYSIS_OPTION_LABEL.analysis_scope.off },
+        ],
+        help: '报告末尾由系统写一段：现象、拆过、排除、没拆、检验，写作者改不了，随报告封存',
+        unknownLabel: (v) => `${v}：无法识别的值，请从上方选项中选择`,
+      },
       ...MODEL_FIELDS,
       { key: 'system', label: '系统提示', type: 'textarea', advanced: true, placeholder: '你是…' },
       ...COMMON_TAIL,
@@ -698,6 +723,20 @@ export const NODE_DEFS: Record<NodeType, NodeDef> = {
       metrics: [{ id: 'total', name: '', unit: '', expression: '' }],
     },
   },
+  analysis: {
+    type: 'analysis', label: NODE_TYPE_LABEL.analysis, category: '把关', icon: ChartColumnStacked,
+    description: ANALYSIS_TEXT.node.description,
+    hasTarget: true, sources: [{ id: 'out', label: '' }],
+    // 配置的键照服务端 engine/schema.analysis_issues：items（每项 id、op、mode、from、params、hypothesis）、overrides（阈值覆盖，
+    // 键和范围同设置里的分析一组）。产出 nodes.<卡>.results.<项> = {status, verdict, direction, values, phrase, artifact}
+    fields: [
+      { key: 'items', label: ANALYSIS_FIELD_LABEL.items, type: 'analysisItems', help: ANALYSIS_TEXT.node.itemsHint },
+      { key: 'overrides', label: ANALYSIS_FIELD_LABEL.overrides, type: 'analysisOverrides', help: ANALYSIS_TEXT.node.overridesHint },
+      ...COMMON_TAIL,
+    ],
+    // 带一项空的：拖进画布就看得到一项要填什么（同口径卡带一个空指标）
+    defaults: { items: [{ id: 'a1', params: {} }] },
+  },
   dual_check: {
     type: 'dual_check', label: NODE_TYPE_LABEL.dual_check, category: '把关', icon: GitCompareArrows,
     description: DUAL_CHECK_TEXT.description,
diff --git a/frontend/src/index.css b/frontend/src/index.css
index ac5af16e..7a48c438 100644
--- a/frontend/src/index.css
+++ b/frontend/src/index.css
@@ -82,6 +82,7 @@
   --nt-human: #ce91e6;
   --nt-validate: #ebb1fd;
   --nt-metrics: #cbaddb;
+  --nt-analysis: #c09dd9;
   --nt-report: #a9b0ec;
   --nt-dual_check: #a98ee0;
   --nt-challenge: #dcb7ef;
@@ -151,6 +152,7 @@
   --nt-human: #6f3d81;
   --nt-validate: #885398;
   --nt-metrics: #6e557b;
+  --nt-analysis: #674a82;
   --nt-report: #4f5694;
   --nt-dual_check: #5a4a91;
   --nt-challenge: #7d4c8f;
@@ -214,6 +216,7 @@
     --nt-human: #6f3d81;
     --nt-validate: #885398;
     --nt-metrics: #6e557b;
+    --nt-analysis: #674a82;
     --nt-report: #4f5694;
     --nt-dual_check: #5a4a91;
     --nt-challenge: #7d4c8f;
@@ -428,6 +431,7 @@
 .nt-transform  { --nt: var(--nt-transform); }
 .nt-subgraph   { --nt: var(--nt-subgraph); }
 .nt-metrics    { --nt: var(--nt-metrics); }
+.nt-analysis   { --nt: var(--nt-analysis); }
 .nt-report     { --nt: var(--nt-report); }
 .nt-dual_check { --nt: var(--nt-dual_check); }
 .nt-challenge  { --nt: var(--nt-challenge); }
diff --git a/frontend/src/lib/terms.ts b/frontend/src/lib/terms.ts
index d36b59f5..e672c20a 100644
--- a/frontend/src/lib/terms.ts
+++ b/frontend/src/lib/terms.ts
@@ -189,6 +189,7 @@ export const NODE_TYPE_LABEL: Record<NodeType, string> = {
   human: '人工审批',
   validate: '结构校验',
   metrics: '口径卡',
+  analysis: '分析卡',
   report: '报告撰写',
   dual_check: '双盲复算',
   challenge: '反方审稿',
@@ -1164,6 +1165,7 @@ export const EVIDENCE_KIND_LABEL: Record<string, string> = {
   entity: '表或字段',
   column: '字段',
   quote: '引文',
+  analysis: '分析结论',
 }
 
 /**
diff --git a/frontend/src/lib/termsAnalysis.ts b/frontend/src/lib/termsAnalysis.ts
new file mode 100644
index 00000000..bdbbcf0f
--- /dev/null
+++ b/frontend/src/lib/termsAnalysis.ts
@@ -0,0 +1,391 @@
+/**
+ * 分析能力一轮（docs/design/analysis-round）的界面文案：分析卡的检查器和卡片、问数据时间线上的分析一步、证据面板的分析一步与
+ * 口径卡指标上的可比性、正文里的说法违规与系统写的「分析范围」、出具声明里的三项、设置「分析」一节、运行对比的分析卡一节。
+ *
+ * 叫法只有服务端契约一份（backend/app/analysis/contract.py 的 OP_LABELS / MODE_LABELS / LABELS，engine/labels.py 的分析卡字段与
+ * 报告节点两个新字段引用它们）：这里的 ANALYSIS_OP_LABEL、ANALYSIS_MODE_LABEL、ANALYSIS_SETTING_LABEL、ANALYSIS_FIELD_LABEL 和它们
+ * 一字不差（scripts/check-studio、check-manage 读后端源码逐项比对）。判定语、判定标准、算不了的原因、附带说明、代入式、闭合核对
+ * 的数、说法违规的原话都是服务端写好的，界面照原样显示，不按判定码、原因码自己拼。规矩同 terms.ts 开头的术语表。
+ *
+ * 术语：一次分析的结果叫「分析结论」，编号 A1、A2……；系统写的那一句叫「判定语」；判定的依据叫「判定标准」；做不成的叫「没有做成」
+ * （服务端判定语的写法），原因叫「算不了的原因」；写错的调用叫「调用有误」，不记成结论；Agent 写的那句叫「Agent 写的假设」。
+ */
+import { formatNumber } from './format'
+
+/** 五个算子（contract.OPS），按契约的先后 */
+export const ANALYSIS_OPS = ['decompose', 'baseline', 'trend', 'differ', 'comparable'] as const
+export type AnalysisOp = (typeof ANALYSIS_OPS)[number]
+
+/** 算子的叫法：contract.OP_LABELS */
+export const ANALYSIS_OP_LABEL: Record<AnalysisOp, string> = {
+  decompose: '变化拆解', baseline: '波动基线', trend: '趋势判断', differ: '差异检验', comparable: '可比性检查',
+}
+
+/** 模式的叫法：contract.MODE_LABELS（比率拆解和比率比较都叫「比率」） */
+export const ANALYSIS_MODE_LABEL: Record<string, string> = {
+  additive: '可加', ratio: '比率', factor: '乘积', rate: '比率', mean: '均值', rows: '逐行明细',
+}
+
+/** 有模式的算子和它们的模式：contract.MODES。其余算子没有模式 */
+export const ANALYSIS_MODES: Partial<Record<AnalysisOp, readonly string[]>> = {
+  decompose: ['additive', 'ratio', 'factor'],
+  differ: ['rate', 'mean', 'rows'],
+}
+
+export const isAnalysisOp = (op: unknown): op is AnalysisOp => typeof op === 'string' && (ANALYSIS_OPS as readonly string[]).includes(op)
+
+/**
+ * 「变化拆解（可加）」「波动基线」：同 contract.method_label。认不出的算子、模式原样写出来，不编叫法。证据面板、导出页用服务端
+ * 给的 method_label；事件里只有算子和模式的（时间线、运行对比）照这一份拼
+ */
+export function analysisMethodLabel(op: unknown, mode?: unknown): string {
+  const name = isAnalysisOp(op) ? ANALYSIS_OP_LABEL[op] : String(op ?? '')
+  if (mode == null || mode === '') return name
+  return `${name}（${ANALYSIS_MODE_LABEL[String(mode)] ?? String(mode)}）`
+}
+
+/** Agent 分析工具的名字（contract.TOOL_NAMES：analysis_<算子>）→ 算子；不是分析工具为 null */
+export function analysisToolOp(tool: unknown): AnalysisOp | null {
+  const m = typeof tool === 'string' ? /^analysis_([a-z]+)$/.exec(tool.trim()) : null
+  return m && isAnalysisOp(m[1]) ? m[1] : null
+}
+
+/** 分析卡、报告节点新字段的叫法：engine/labels.py 的 _FIELD_BY_TYPE['analysis'] 与 FIELD_LABEL */
+export const ANALYSIS_FIELD_LABEL = {
+  items: '分析项',
+  overrides: '阈值覆盖',
+  analysis_claims: '分析说法核对',
+  analysis_scope: '分析范围',
+} as const
+
+/** 报告节点两个新字段取值的短说法：engine/labels.py 的 _OPTION_LABEL（下拉的选项以它开头，后面可以带一句用途） */
+export const ANALYSIS_OPTION_LABEL = {
+  analysis_claims: { off: '不核对', check: '重写一次后标出', strict: '重写后仍有的计入缺口' },
+  analysis_scope: { auto: '有分析结论时附上', off: '不附' },
+} as const
+
+/** 分析设置的键（contract.DEFAULTS 的先后） */
+export const ANALYSIS_SETTING_KEYS = [
+  'main_share', 'concentration', 'offset_ratio', 'flat_pct', 'k_warn', 'k_alert', 'min_history', 'max_history',
+  'trend_min_points', 'trend_alpha', 'differ_alpha', 'min_n', 'min_expected', 'min_denominator', 'max_calls',
+  'max_drill_depth', 'table_top_k',
+] as const
+export type AnalysisSettingKey = (typeof ANALYSIS_SETTING_KEYS)[number]
+
+/** 每一项的叫法：contract.LABELS（设置页、阈值覆盖、服务端报错同一套） */
+export const ANALYSIS_SETTING_LABEL: Record<AnalysisSettingKey, string> = {
+  main_share: '「主要来源」的占比',
+  concentration: '「集中」的合计占比',
+  offset_ratio: '记为「抵消」的比例',
+  flat_pct: '「基本持平」的幅度',
+  k_warn: '「超出常规波动」的倍数',
+  k_alert: '「明显超出」的倍数',
+  min_history: '至少几期历史',
+  max_history: '最多用几期历史',
+  trend_min_points: '趋势至少几期',
+  trend_alpha: '趋势的显著性水平',
+  differ_alpha: '差异检验的显著性水平',
+  min_n: '均值比较的最小样本数',
+  min_expected: '比率比较的最小期望频数',
+  min_denominator: '比率拆解的最小分母',
+  max_calls: 'Agent 每次执行最多分析几次',
+  max_drill_depth: '最多下钻几层',
+  table_top_k: '列出前几段',
+}
+
+/** 每一项的类型和闭区间：contract.RANGES（整数项只收整数）。界面先按它说哪一项不对，两两之间的大小交给服务端核 */
+export const ANALYSIS_SETTING_RANGE: Record<AnalysisSettingKey, readonly ['float' | 'int', number, number]> = {
+  main_share: ['float', 0.3, 0.9],
+  concentration: ['float', 0.5, 0.99],
+  offset_ratio: ['float', 0.05, 1.0],
+  flat_pct: ['float', 0.0, 0.1],
+  k_warn: ['float', 1.0, 5.0],
+  k_alert: ['float', 1.5, 8.0],
+  min_history: ['int', 4, 104],
+  max_history: ['int', 8, 260],
+  trend_min_points: ['int', 4, 104],
+  trend_alpha: ['float', 0.001, 0.2],
+  differ_alpha: ['float', 0.001, 0.2],
+  min_n: ['int', 2, 10_000],
+  min_expected: ['int', 1, 50],
+  min_denominator: ['int', 1, 100_000],
+  max_calls: ['int', 1, 50],
+  max_drill_depth: ['int', 1, 6],
+  table_top_k: ['int', 3, 30],
+}
+
+/** 设置页把这几项分成几组（只是排版，键和叫法照上面） */
+export const ANALYSIS_SETTING_GROUPS: { title: string; keys: AnalysisSettingKey[] }[] = [
+  { title: '变化拆解', keys: ['main_share', 'concentration', 'offset_ratio', 'flat_pct', 'min_denominator', 'table_top_k'] },
+  { title: '波动基线与趋势判断', keys: ['k_warn', 'k_alert', 'min_history', 'max_history', 'trend_min_points', 'trend_alpha'] },
+  { title: '差异检验', keys: ['differ_alpha', 'min_n', 'min_expected'] },
+  { title: '次数与层数', keys: ['max_calls', 'max_drill_depth'] },
+]
+
+/** 范围写成人话：「0.3 到 0.9 之间」「4 到 104 之间的整数」 */
+export function settingRangeText(key: AnalysisSettingKey): string {
+  const [kind, lo, hi] = ANALYSIS_SETTING_RANGE[key]
+  return kind === 'int' ? `${formatNumber(lo)} 到 ${formatNumber(hi)} 之间的整数` : `${lo} 到 ${hi} 之间`
+}
+
+/** 一项填的值有什么问题（类型、范围），没问题为 null。说法同服务端 contract.settings_problem */
+export function settingValueProblem(key: AnalysisSettingKey, raw: string): string | null {
+  const text = raw.trim()
+  const [kind, lo, hi] = ANALYSIS_SETTING_RANGE[key]
+  const label = ANALYSIS_SETTING_LABEL[key]
+  if (!text || !Number.isFinite(Number(text))) return `${label}要填数字`
+  const n = Number(text)
+  if (kind === 'int' && !/^-?\d+$/.test(text)) return `${label}要填整数`
+  if (n < lo || n > hi) return `${label}要在 ${lo} 到 ${hi} 之间`
+  return null
+}
+
+const short = (text: string, n = 24) => (text.length > n ? `${text.slice(0, n - 1)}…` : text)
+
+export const ANALYSIS_TEXT = {
+  /** 记录页工件清单里分析结论那一行的说明（工件没记算子时） */
+  artifactHint: '分析结论：判定、方法、计算、输入和复算所需的一切',
+  // ---- 分析卡：节点库、检查器、卡片 ----
+  node: {
+    description: '对上游查询节点的结果固定做几项确定性分析：变化拆解、波动基线、趋势判断、差异检验、可比性检查。判定由平台按设置的阈值算出，报告可引用、可查看出处',
+    itemsHint: '每一项做一种分析：选分析方法、要分析的上游查询节点，再按方法填列名和期次。判定语由系统按平台的方法写好，报告引用它，不另写判断',
+    overridesHint: '只对这张卡生效，写进工作流、随版本固定；未覆盖的项跟随「设置 → 偏好设置 → 分析」。Agent 的分析工具没有阈值覆盖',
+    addItem: '添加分析项',
+    removeItem: (id: string) => `删除分析项 ${id || '（未命名）'}`,
+    id: '标识',
+    idHint: (node: string, id: string) => `下游条件分支按 nodes.${node}.results.${id || '<标识>'}.verdict 取判定`,
+    idInvalid: '标识只能用字母、数字和下划线，且不以数字开头',
+    idDuplicate: (id: string) => `标识「${id}」与另一项重复，请换一个`,
+    op: '分析方法',
+    pickOp: '选择分析方法',
+    mode: '模式',
+    pickMode: '选择模式',
+    unknownOp: (op: string) => `${op}：无法识别的分析方法，请从上方选项中选择`,
+    unknownMode: (mode: string) => `${mode}：无法识别的模式，请从上方选项中选择`,
+    from: '输入',
+    pickFrom: '选择上游的查询节点',
+    noUpstream: '上游没有查询节点。请先在本节点之前连接选择了数据库查询工具的「调用工具」节点，或「合并查询」节点',
+    missingNode: (id: string) => `节点「${id}」不在上游或已删除，请重新选择`,
+    twoQueries: '基期、本期是两份查询',
+    baseQuery: '基期查询',
+    targetQuery: '本期查询',
+    hypothesis: '在验证什么',
+    hypothesisHint: '一句话，只在证据面板显示，不进报告正文',
+    hypothesisPlaceholder: '下降是不是集中在某个地区',
+    params: '参数',
+    listHint: '多列用顿号或逗号分隔',
+    cellHint: '写成「节点 ID.r行号.列名」，行号从 0 数',
+    headline: '总数核对',
+    headlineHint: '给出基期、本期总数那两格，系统核对拆开后各段之和是否对得上',
+    headlineBase: '基期总数那一格',
+    headlineTarget: '本期总数那一格',
+    parent: '下钻',
+    parentHint: '这一项是前面某个变化拆解里某一段的拆解',
+    parentOf: '上一层分析项',
+    parentKey: '那一段的维度值',
+    noParent: '不下钻',
+    pairsHint: '每行一对，两个组之间用顿号或逗号分隔',
+    /** 参数的叫法（tools/builtin/analysis.py 各参数模型里的字段，分析卡按同一份校验） */
+    param: {
+      by: '维度列', period: '期次列', base: '基期', target: '本期', measure: '度量列', numerator: '分子列', denominator: '分母列',
+      product: '乘积列', factors: '因子列', label: '度量的叫法', basis: '比较方式', grain: '期次粒度', lag: '和往前第几期比',
+      pairs: '只比这几对', reference: '对照组', mean: '均值列', sd: '标准差列', sum: '合计列', sumsq: '平方和列', n: '样本数列',
+      value: '值列',
+    } as Record<string, string>,
+    paramPlaceholder: {
+      by: 'region', period: 'week', base: '2026-W41', target: '2026-W42', measure: 'gmv', numerator: 'refunds',
+      denominator: 'orders', product: 'gmv', factors: 'orders、aov', label: '销售额', lag: '1', reference: '华东',
+      mean: 'avg_amount', sd: 'sd_amount', sum: 'sum_amount', sumsq: 'sumsq_amount', n: 'orders', value: 'amount',
+    } as Record<string, string>,
+    paramHint: {
+      label: '系统写判定语时用；不写数字、百分号和单位，留空则用列名或数据目录中的名称',
+      target: '留空取最后一期',
+      lag: '日缺省 7（和上周同一天比），周、月缺省 1；同比写 52（周）或 12（月）',
+      reference: '其余各组各和它比；留空则各组与其余各组合计比',
+    } as Record<string, string>,
+    basis: { '': '和比较期比的变化率（缺省）', change: '和比较期比的变化率', level: '值本身（存量、比率用）' } as Record<string, string>,
+    grain: { '': '按期次的写法推断', day: '日', week: '周', month: '月', quarter: '季度' } as Record<string, string>,
+    /** differ 均值模式：给均值、标准差、样本数，或合计、平方和、样本数 */
+    meanShape: { sd: '均值、标准差、样本数', sumsq: '合计、平方和、样本数' } as Record<string, string>,
+    // ---- 阈值覆盖 ----
+    addOverride: '添加阈值覆盖',
+    pickOverride: '选择要覆盖的阈值',
+    removeOverride: (label: string) => `删除阈值覆盖「${label}」`,
+    noOverrides: '未覆盖：全部跟随设置',
+    allOverridden: '全部阈值都已覆盖',
+    // ---- 卡片 ----
+    summaryEmpty: '未配置分析项',
+    summaryNoOp: '未选择分析方法',
+    summaryMore: (n: number) => `等 ${formatNumber(n)} 项`,
+  },
+
+  // ---- 问数据时间线：Agent 的分析工具、分析卡的每一项 ----
+  step: {
+    /** 运行中的状态字（store/chat） */
+    running: (label: string) => `正在执行分析：${label}…`,
+    /** 一步的标题：「变化拆解（可加）· A1」；分析卡的一项：「变化拆解（可加）· where」 */
+    title: (method: string, alias?: string | null) => (alias ? `${method} · ${alias}` : method),
+    callError: '调用有误，没有记成结论',
+    callErrorNext: '这次调用没有记成结论，Agent 会按这条说明改正参数后重新调用',
+    refused: '没有做成',
+    refusedNext: '算不了也是一条结论：报告里可以照实写出原因，不能据此下判断',
+    hypothesis: 'Agent 写的假设',
+    itemHypothesis: '在验证什么',
+    args: '参数',
+    phrase: '判定语',
+    /** 分析卡那一项的标识 */
+    item: (id: string) => `分析项「${id}」`,
+  },
+
+  // ---- 时间线上的几条提醒（log 的 code：report_analysis_claims、analysis_refused、metric_comparability_skipped） ----
+  log: {
+    claims: (n: number | undefined, strict: boolean) => `报告${n ? `有 ${formatNumber(n)} 处` : '有'}分析说法未通过核对，`
+      + (strict ? '已计入出具缺口' : '已在报告中标注'),
+    refused: (method: string, item?: string) => `${method}${item ? `（分析项「${item}」）` : ''}没有做成`,
+    comparabilitySkipped: (metric: string) => `指标${metric ? `「${metric}」` : ''}用到的两次查询没有做成可比性检查`,
+    comparabilitySkippedNext: '指标照常计算，两期能不能直接比请自行核对',
+  },
+
+  // ---- 证据面板的分析一步（照导出页 report_export/runtime.js 的说法） ----
+  evidence: {
+    title: '分析结论',
+    titleOf: (alias: string, method: string) => `${['分析结论', alias].filter(Boolean).join(' ')} · ${method || '—'}`,
+    phrase: '判定语',
+    criterion: '判定标准',
+    subjects: '点名的段',
+    direction: '方向',
+    directionLabel: { up: '上升', down: '下降' } as Record<string, string>,
+    caveats: '附带说明',
+    reason: '算不了的原因',
+    citedPart: '这一处引用的是',
+    citedPhrase: '判定语',
+    citedValue: (key: string) => `具名的数「${key}」`,
+    citedCell: (row: number, column: string) => `结果表第 ${formatNumber(row + 1)} 行「${column}」`,
+    who: '谁做的',
+    origin: { card: '分析卡（模板定好的分析）', agent: 'Agent 现场做的分析', metric: '口径卡执行时自动做的可比性检查' } as Record<string, string>,
+    hypothesis: 'Agent 写的假设',
+    refused: '这次分析没有做成',
+    renderOk: '与按结论渲染的结果一致',
+    renderBad: '与按结论渲染的结果不一致',
+    method: '方法',
+    methodVersion: '方法与版本',
+    params: '参数',
+    settings: '实际用的阈值',
+    probes: '现查结果',
+    probeBasis: { live: '现查', profile: '剖析', observed: '取值观察' } as Record<string, string>,
+    probeBound: (expr: string, value: string) => `按运行当天现算的界 ${expr} = ${value}`,
+    probeEmpty: '这一列一个值都没有',
+    probeRange: (min: string, max: string) => `${min} 至 ${max}`,
+    calc: '计算',
+    values: '具名的数',
+    formulas: '关键数的代入式',
+    resultTable: '结果表',
+    resultRows: (shown: number, total: number) => `列出 ${formatNumber(shown)} 行，共 ${formatNumber(total)} 行`,
+    input: '输入',
+    inputOf: (ref: string) => `输入 ${ref}`.trim(),
+    role: '角色',
+    ref: '调用里写的名字',
+    query: '查询',
+    columnsUsed: '用到的列',
+    rowsUsed: '用到的行',
+    allRows: '全部行',
+    rowsList: (rows: number[]) => rows.map((r) => `第 ${formatNumber(r + 1)} 行`).join('、'),
+    headline: '总数那一格',
+    headlineRatio: '整体比率那一格',
+    cellAt: (row: number, column: string) => `第 ${formatNumber(row + 1)} 行「${column}」`,
+    closure: '闭合核对',
+    closureOk: (h: string, t: string) => `总数 ${h} 与各段之和 ${t} 闭合`,
+    closureBad: (h: string, t: string) => `总数 ${h} 与各段之和 ${t} 对不上`,
+    ratioClosureOk: (h: string, t: string) => `整体比率 ${h} 与各段分子之和 ÷ 分母之和 ${t} 闭合`,
+    ratioClosureBad: (h: string, t: string) => `整体比率 ${h} 与各段分子之和 ÷ 分母之和 ${t} 对不上`,
+    parent: '下钻自',
+    gotoInput: (alias: string) => `查看输入 ${alias}`,
+    inputOfAnalysis: (alias: string) => `分析结论 ${alias} 的输入`,
+    recompute: '复算',
+    seal: '封存',
+    artifact: '结论工件',
+    inLedger: '封存范围内的台账记过这份分析结论',
+    notInLedger: '封存范围内的台账里没有这份分析结论',
+    viaCard: '记在口径卡的指标集里，随卡封存（不进台账）',
+    chainPending: '正在加载判定、方法和输入…',
+    // ---- 口径卡指标上的两期可比性 ----
+    comparability: '两期可比性',
+    comparabilityLabel: { comparable: '两期可比', conditional: '两期有条件可比', incomparable: '两期不可比' } as Record<string, string>,
+    comparabilityOpen: '展开这份可比性结论（判定、方法、计算、输入、复算、封存）',
+    comparabilityClose: '收起可比性结论',
+    comparabilityReasons: '原因',
+    comparabilityUnsealed: '这份可比性结论不在封存范围内，不能作为证据',
+    comparabilityHash: '可比性结论与哈希不一致，疑似被修改',
+  },
+
+  // ---- 正文：说法违规的标注、系统写的「分析范围」 ----
+  claim: {
+    /** 标注的种类（服务端 analysis_claims.CODES）：给人看的短说法同审计的标签 */
+    label: {
+      analysis_claim_uncited: '判断词未挂分析结论',
+      analysis_claim_mismatch: '判断与分析结论不符',
+      causal_unsupported: '因果说法没有依据',
+    } as Record<string, string>,
+    tag: '说法待核',
+    hint: '这句话里的判断词没有挂对应的分析结论、判断与所挂结论对不上，或原因没有依据：系统无法核对，点开看原因',
+    panelTitle: '说法核对',
+    analyses: '所挂的分析结论',
+    noAnalyses: '这句没有挂分析结论',
+    count: (n: number) => `${formatNumber(n)} 处说法待核`,
+  },
+  scope: {
+    title: '分析范围',
+    hint: '系统按这次的分析结论写的一段，不是写作者写的，随报告封存；编号可点开看出处',
+    phenomena: '现象',
+    decomposed: '拆过',
+    excluded: '排除',
+    refused: '算不了',
+    notDecomposed: '没拆',
+    tests: '检验',
+    notDecomposedList: (dims: string[]) => `候选维度里没拆过的：${dims.join('、')}`,
+    notDecomposedNone: '候选维度都拆过了',
+    notDecomposedUnknown: '候选维度核对不了',
+    drill: (parent: string) => `下钻自 ${parent}`,
+    dims: (dims: string[]) => `按${dims.join('、')}`,
+    /** 服务端 ScopeBlock.corrected：只做了一次（一次里比了多对的已按次数校正）/ 不止一次（几次之间没有校正） */
+    testsText: (n: number, corrected: boolean | null | undefined) => `做了 ${formatNumber(n)} 次差异检验${corrected === true
+      ? '（一次里比了几对的已按次数校正）' : corrected === false ? '，几次检验之间没有按次数校正' : ''}`,
+    open: (alias: string) => `查看分析结论 ${alias} 的出处`,
+  },
+
+  // ---- 出具声明的三项：分析说法核对、指标的可比性、追不到原始数据的指标 ----
+  issuance: {
+    claimsTitle: (n: number, counted: boolean) => (counted
+      ? `${formatNumber(n)} 处分析说法未通过核对，已计入出具缺口`
+      : `${formatNumber(n)} 处分析说法未通过核对（已标注，不影响出具档位）`),
+    claimsStrictClean: '分析说法已按严格核对，没有未通过的',
+    claimsNext: '判断要挂对应的分析结论并照判定写；原因要有知识库原话或运行输入作依据，否则写成待核实的推测',
+    comparabilityTitle: '指标的两期可比性',
+    comparabilityItem: (name: string, phrase: string) => `指标「${name}」：${phrase}`,
+    untracedTitle: (counted: boolean) => (counted ? '报告引用的指标未追到原始数据（已计入出具缺口）' : '报告引用的指标未追到原始数据'),
+    untracedItem: (name: string, stops: string[]) => `指标「${name}」${stops.length ? `：${stops.join('；')}` : ''}`,
+  },
+
+  // ---- 设置「分析」一节 ----
+  settings: {
+    title: '分析',
+    hint: '分析工具和分析卡判定用的阈值、次数上限。发起运行时快照进这次运行；Agent 改不了这些阈值，分析卡上可以写阈值覆盖',
+    invalid: '分析的设置中有无效值，请修正后再保存',
+    current: (value: string) => `当前 ${value}`,
+  },
+
+  // ---- 运行对比 ----
+  compare: {
+    title: '分析卡',
+    hint: '两次运行的分析卡按节点和分析项对齐，只列出两次的判定和判定语，不拆变化来源',
+    change: {
+      same: '判定一致', verdict: '判定变了', phrase: '判定一致，判定语里的数变了', added: '这次新增', removed: '这次没有',
+    } as Record<string, string>,
+    base: '基准运行',
+    target: '这次运行',
+    none: '—',
+    item: (title: string, item: string) => `${title} · ${item}`,
+  },
+}
+
+export const shortPhrase = short
diff --git a/frontend/src/store/studio.ts b/frontend/src/store/studio.ts
index ec90d9d3..dbd55e7f 100644
--- a/frontend/src/store/studio.ts
+++ b/frontend/src/store/studio.ts
@@ -5,7 +5,8 @@ import {
 } from '@xyflow/react'
 import { api, streamCopilot, streamRun } from '../api/client'
 import { copilotReceipt, mergeNodeConfig, toolChangesOf } from '../canvas/copilotMerge'
-import { connectMergeInput } from '../canvas/mergeInputs'
+import { connectMergeInput, isQueryNode } from '../canvas/mergeInputs'
+import { connectAnalysisInput } from '../canvas/analysisItems'
 import { contentSig, isMissingEndpoint, normalizeUpgrade } from '../canvas/issues'
 import { NODE_DEFS, sourceHandles } from '../canvas/nodeDefs'
 import { toast } from '../components/ui'
@@ -1420,8 +1421,9 @@ export const useStudio = create<StudioState>((set, get) => ({
   onConnect: (conn) => {
     if (locked(get)) return
     commit(set, get, '连线')
-    // 查询节点连到合并查询：顺手把它加成合并的输入（别名取有意义的短名），和连线同一步撤销
+    // 查询节点连到合并查询：顺手把它加成合并的输入（别名取有意义的短名）；连到分析卡：还没选输入的第一项选上它。和连线同一步撤销
     const nodes = connectMergeInput(get().nodes, conn.source, conn.target)
+      ?? connectAnalysisInput(get().nodes, conn.source, conn.target, isQueryNode)
     set({
       edges: addEdge({ ...conn, type: 'flow' }, get().edges),
       ...(nodes ? { nodes } : {}),
diff --git a/frontend/src/types.ts b/frontend/src/types.ts
index c1d22a1d..aa45b19a 100644
--- a/frontend/src/types.ts
+++ b/frontend/src/types.ts
@@ -1,7 +1,7 @@
 export type NodeType =
   | 'input' | 'output' | 'llm' | 'agent' | 'supervisor' | 'tool' | 'code' | 'merge'
   | 'branch' | 'loop' | 'subgraph' | 'memory' | 'retrieve' | 'transform'
-  | 'human' | 'validate' | 'metrics' | 'report' | 'dual_check' | 'challenge'
+  | 'human' | 'validate' | 'metrics' | 'analysis' | 'report' | 'dual_check' | 'challenge'
 
 export interface GraphNode {
   id: string
@@ -1771,9 +1771,9 @@ export interface EvidenceCitation {
   ref?: string
   /** 目录里的别名，如 m:gmv、i:week、Q1 */
   alias?: string
-  /** 引文：hit 是检索快照里第几条命中，start / end 是引文在那条原文里的码点位置 */
+  /** 引文：hit 是检索快照里第几条命中，start / end 是引文在那条原文里的码点位置；分析结论的具名的数：key（[[a:A1.top_share]]） */
   locator?: { metric?: string; field?: string; row?: number; column?: string; table?: string
-              hit?: number; start?: number; end?: number }
+              hit?: number; start?: number; end?: number; key?: string }
   eid?: string
   /** metric / input / cell / table / column / quote / query */
   kind?: string
@@ -2008,8 +2008,15 @@ export interface EvidenceBlock {
 }
 
 export interface EvidenceViolation {
-  /** uncited_number / unresolved_ref / render_mismatch / eid_mismatch / … */
+  /**
+   * uncited_number / unresolved_ref / render_mismatch / eid_mismatch / …；分析能力一轮的说法违规：analysis_claim_uncited
+   * （判断词未挂分析结论）、analysis_claim_mismatch（判断与分析结论不符）、causal_unsupported（因果说法没有依据）
+   */
   code: string
+  /** 说法违规：判断词的类（source、anomaly、causal……）、要挂的算子、所挂的分析结论编号 */
+  claim?: string
+  op?: string
+  analyses?: string[]
   message?: string
   span?: [number, number]
   text?: string
@@ -2021,6 +2028,8 @@ export interface EvidenceViolation {
 }
 
 export interface EvidenceStats {
+  /** 说法违规几处（只有记了分析说法核对的文档有） */
+  analysis_claims?: number
   units?: number
   segments?: number
   claims?: number
@@ -2062,6 +2071,10 @@ export interface EvidenceDocData {
   violations?: EvidenceViolation[]
   /** 四期：写作时结论句策略是 judge 才有 */
   judge?: EvidenceJudgeSummary
+  /** 分析能力一轮：报告节点核分析说法时记下的处置（check / strict）和说法表版本；老文档没有 */
+  analysis_claims?: { mode?: string; lexicon?: number }
+  /** 系统写的「分析范围」：这次有分析结论、报告节点没关掉时才有 */
+  analysis_scope?: AnalysisScope
 }
 
 /**
@@ -2371,6 +2384,207 @@ export interface EvidenceUntraced {
   type: string | null
 }
 
+// ---------------------------------------------------------------------------
+// 分析能力一轮（docs/design/analysis-round）：分析结论在证据里的样子。形状照服务端 api/evidence.py 的 _analysis_chain、
+// _comparability_view，报告文档的 analysis_scope（analysis.contract.ScopeBlock），出具声明的 analysis_claims / comparability /
+// untraced（engine/nodes/io.py）。判定语、判定标准、原因、附带说明、代入式、闭合核对的数都是服务端写好的字，界面照原样显示
+// ---------------------------------------------------------------------------
+
+/** 分析结论的判定：判定码、点名的段（一段一个列表，多维时几个值；乘积拆解是因子名）、方向 */
+export interface AnalysisVerdict {
+  code?: string
+  subjects?: (string | number | null)[][]
+  direction?: 'up' | 'down' | null
+}
+
+/** 附带说明：一种说法，服务端渲染好的一句 */
+export interface AnalysisCaveat { code?: string; text?: string }
+
+/** 可比性检查现查的一条：一列的实际范围（table、column、min、max），或按运行当天现算的界（expr、value） */
+export interface AnalysisProbe {
+  source?: string
+  table?: string
+  column?: string
+  min?: string | null
+  max?: string | null
+  basis?: 'live' | 'profile' | 'observed' | string
+  at?: string
+  expr?: string
+  value?: string | null
+}
+
+/**
+ * 出处链里分析结论的六步（step 依次是 analysis、analysis_method、analysis_calc、analysis_input（每份输入一步）、
+ * analysis_recompute、analysis_seal）共用的形状：每一步只带自己那几项
+ */
+export interface EvidenceAnalysisStep {
+  step: string
+  alias?: string
+  // ---- analysis：判定 ----
+  op?: string
+  mode?: string | null
+  /** 「变化拆解（可加）」：服务端 contract.method_label */
+  method_label?: string
+  /** card 分析卡、agent Agent 现场做的、metric 口径卡执行时自动做的可比性检查 */
+  origin?: 'card' | 'agent' | 'metric' | string
+  item?: string
+  call_id?: string
+  node_id?: string | null
+  label?: string | null
+  artifact?: string | null
+  eid?: string
+  eid_ok?: boolean
+  /** 这一处引用的是结论的哪一项：判定语、具名的数（key）、结果表的一格（row、column） */
+  cited?: { what?: 'phrase' | 'value' | 'cell' | string; key?: string; row?: number; column?: string; raw?: unknown; conv?: string }
+  rendered?: string
+  render_ok?: boolean
+  hash_ok?: boolean | null
+  status?: string
+  reason_code?: string | null
+  reason?: string | null
+  phrase?: string
+  criterion?: string
+  verdict?: AnalysisVerdict | null
+  caveats?: AnalysisCaveat[]
+  hypothesis?: string
+  hypothesis_note?: string
+  note?: string
+  // ---- analysis_method：方法、参数、实际用的阈值、现查结果 ----
+  method?: string
+  params?: Record<string, unknown>
+  settings?: { key: string; label?: string; value: unknown }[]
+  probes?: AnalysisProbe[]
+  // ---- analysis_calc：具名的数、代入式、结果表 ----
+  values?: { key: string; raw?: unknown; rendered?: string; cited?: boolean }[]
+  formulas?: { label?: string; text?: string }[]
+  table?: {
+    columns: string[]
+    labels?: Record<string, string>
+    total_rows?: number
+    row_index?: number[]
+    rows: unknown[][]
+    masked?: string[]
+    highlight?: { cells?: [number, string][] }
+    window_truncated?: boolean
+  }
+  // ---- analysis_input：一份输入 ----
+  /** 这一步是哪份结论的输入 */
+  of?: string
+  ref?: string
+  role?: string
+  role_label?: string
+  columns?: string[]
+  rows?: 'all' | number[]
+  row?: number
+  column?: string
+  value?: string | null
+  /** 输入落到报告目录的哪个 Q */
+  query?: string | null
+  sealed?: boolean
+  /** headline 那一格和结论里总数的闭合核对：mode 可加对各段之和，比率对各段分子之和 ÷ 分母之和 */
+  closure?: { mode?: string; total_key?: string; total?: string; headline?: string; closed?: boolean }
+  // ---- analysis_recompute：复算（服务端算好的那一句） ----
+  text?: string
+  // ---- analysis_seal：封存 ----
+  ledger?: boolean
+  /** metric_set：口径卡指标上的可比性结论，记在指标集里、随卡封存 */
+  via?: string
+}
+
+/** 指标步骤上的两期可比性（api/evidence._comparability_view）：判定、判定语、原因，连同那份可比性结论的出处链 */
+export interface EvidenceComparability {
+  verdict?: 'comparable' | 'conditional' | 'incomparable' | string
+  phrase?: string
+  reasons?: string[]
+  artifact?: string | null
+  hash_ok?: boolean | null
+  sealed?: boolean
+  method?: string
+  criterion?: string
+  caveats?: AnalysisCaveat[]
+  probes?: AnalysisProbe[]
+  /** 判定 → 方法 → 计算 → 输入 → 复算 → 封存（两份输入快照的查询步骤指标链里已经有了，不再给） */
+  steps?: EvidenceAnalysisStep[]
+}
+
+/** 报告文档里系统写的「分析范围」的一条（contract.ScopeItem） */
+export interface AnalysisScopeItem {
+  alias?: string
+  op?: string
+  mode?: string | null
+  /** 拆解的维度（给人看的叫法） */
+  dims?: string[]
+  verdict?: string | null
+  phrase?: string
+  /** 下钻：上一层的编号 */
+  parent?: string | null
+  reason?: string | null
+}
+
+/** 报告文档的 analysis_scope（contract.ScopeBlock）：不进正文、不过裸数字和说法核对，随文档封存 */
+export interface AnalysisScope {
+  phenomena?: AnalysisScopeItem[]
+  decomposed?: AnalysisScopeItem[]
+  excluded?: AnalysisScopeItem[]
+  refused?: AnalysisScopeItem[]
+  /** 候选维度里没拆过的叫法；null 是候选维度核对不了（看 note） */
+  not_decomposed?: string[] | null
+  note?: string | null
+  /** 这次做了几次差异检验 */
+  tests?: number
+  /** 检验校正过没有：只做了一次为 true，不止一次为 false，没做为 null */
+  corrected?: boolean | null
+}
+
+/** 出具声明的 analysis_claims：处置、几处、记没记缺口、按种类的计数、前几处（同报告文档的说法违规） */
+export interface IssuanceAnalysisClaims {
+  mode?: 'check' | 'strict' | string
+  count?: number
+  counts?: Record<string, number>
+  counted?: boolean
+  violations?: EvidenceViolation[]
+}
+
+/** 出具声明的 comparability：两期有条件可比、不可比的指标，一条一个 */
+export interface IssuanceComparability {
+  metric?: string
+  name?: string
+  verdict?: string
+  phrase?: string
+  reasons?: string[]
+  artifact?: string
+}
+
+/** 出具声明的 untraced：报告引用的指标里输入追不到原始数据的，counted 是记了缺口（受管级别的正式运行） */
+export interface IssuanceUntraced {
+  metrics?: { metric?: string; name?: string; stops?: { path?: string; node_id?: string | null; label?: string | null; type?: string | null }[] }[]
+  counted?: boolean
+}
+
+/** Agent 分析工具 tool.end 的 analysis（服务端 tools/analysis.event_view）：preview 只是结论的前一截，判定语从这里读 */
+export interface AnalysisEventView {
+  /** 这份结论在本节点的编号（「【分析 An】」那个）；不记台账的执行没有 */
+  alias?: string | null
+  op?: string
+  mode?: string | null
+  status?: 'ok' | 'refused' | string
+  verdict?: string | null
+  direction?: string | null
+  phrase?: string
+  reason_code?: string | null
+}
+
+/** 分析卡的一项（节点配置 items[i]，engine/schema.analysis_issues 校验） */
+export interface AnalysisItem {
+  id?: string
+  op?: string
+  mode?: string | null
+  /** 要分析的上游查询节点（调用数据库查询工具的「调用工具」节点，或「合并查询」）；两份查询时写在 params 的 base_query、target_query */
+  from?: string
+  params?: Record<string, any>
+  hypothesis?: string
+}
+
 export interface EvidenceStep extends Omit<EvidenceInput, 'status'> {
   step: string
   /**
@@ -2477,6 +2691,10 @@ export interface EvidenceStep extends Omit<EvidenceInput, 'status'> {
   merge?: EvidenceMerge
   /** 这一步是哪次合并查询的输入（合并结果在目录里的编号）：排在那个合并步骤后面，高亮的是追到的格 */
   merged_into?: string
+  /** query：这一步是哪份分析结论的输入（结论在目录里的编号 A…）：接在那份结论的输入后面 */
+  input_of?: string
+  /** metric：跨两份同形查询算的环比、同比，口径卡执行时自动做的可比性检查。没做过检查的指标没有这个键 */
+  comparability?: EvidenceComparability
   // ---- entity：表或字段（三期）----
   /** entity：table / column */
   kind?: string
@@ -4371,12 +4589,38 @@ export interface RunCompareQuery {
 }
 
 /** GET /runs/{id}/compare?base=。任一未封存或封存核对不一致 409；不是同一工作流、是同一次运行 422 */
+/** 运行对比里分析卡的一边（analysis.end 那一项）：没有这一项为 null */
+export interface RunCompareAnalysisSide {
+  status?: string | null
+  verdict?: string | null
+  direction?: string | null
+  phrase?: string
+  artifact?: string | null
+}
+
+/**
+ * 运行对比里分析卡的一项（engine/compare._analyses）：两次运行的分析卡按节点和项的标识对齐。change：same 判定和判定语都一样、
+ * verdict 状态或判定变了、phrase 判定一样而判定语里的数变了、added / removed 只在一边有
+ */
+export interface RunCompareAnalysis {
+  node_id: string
+  title?: string
+  item: string
+  op?: string | null
+  mode?: string | null
+  base: RunCompareAnalysisSide | null
+  target: RunCompareAnalysisSide | null
+  change: 'same' | 'verdict' | 'phrase' | 'added' | 'removed' | string
+}
+
 export interface RunComparison {
   target: RunCompareRef
   base: RunCompareRef
   summary: RunCompareSummary
   template: RunCompareTemplate
   metrics: RunCompareMetric[]
+  /** 分析卡按项对齐（分析能力一轮）；老服务端没有 */
+  analyses?: RunCompareAnalysis[]
   queries: RunCompareQuery[]
   notes: string[]
 }
@@ -4505,6 +4749,12 @@ export interface DualCheckConfig {
   enabled?: boolean | null
 }
 
+/**
+ * 设置里的 analysis 一组（GET /settings 给的是补齐缺省之后的完整一组）：判定用的阈值、Agent 每次执行最多分析几次、最多下钻
+ * 几层、列出前几段。键、叫法、范围见 lib/termsAnalysis（照服务端 analysis/contract.py）
+ */
+export type AnalysisSettings = Record<string, number>
+
 /** 设置里的 verify 一组（GET /settings 给的是补齐默认之后的完整一组；上限为 null 表示不限） */
 export interface VerifySettings {
   provider: string | null
diff --git a/scripts/check-studio.mjs b/scripts/check-studio.mjs
index 87a80fd4..843e4008 100644
--- a/scripts/check-studio.mjs
+++ b/scripts/check-studio.mjs
@@ -131,6 +131,31 @@ const MERGE_GRAPH = {
     edge('q_visits', 'merge'), edge('think', 'merge'), edge('merge', 'done')],
 }
 
+// 分析卡（分析能力一轮）：整体周销售、按地区两周各查一次，分析卡做两项（波动基线、按地区的变化拆解，带总数核对），报告引用它的
+// 结论。上游另有一个模型调用节点，检查器的输入下拉里不该有它。全是虚构零售库的通用名
+const TOTALS_SQL = 'SELECT week, SUM(amount) AS gmv FROM orders GROUP BY week ORDER BY week'
+const REGION_SQL = 'SELECT region, week, SUM(amount) AS gmv FROM orders GROUP BY region, week ORDER BY region, week'
+const ANALYSIS_GRAPH = {
+  nodes: [
+    node('start', 'input', 0, 200, '输入', { fields: [{ name: 'week', required: true }] }),
+    node('weekly', 'tool', 300, 60, '整体周销售', { tool: 'db_query__shop', args: { sql: TOTALS_SQL } }),
+    node('by_region', 'tool', 300, 220, '按地区两周', { tool: 'db_query__shop', args: { sql: REGION_SQL } }),
+    node('think', 'llm', 300, 400, '先想一想', { prompt: '{{ input.week }}' }),
+    node('card', 'analysis', 640, 200, '异动分析', {
+      items: [
+        { id: 'wow', op: 'baseline', from: 'weekly', params: { period: 'week', measure: 'gmv', label: '销售额' } },
+        { id: 'where', op: 'decompose', mode: 'additive', from: 'by_region', hypothesis: '下降是不是集中在某个地区',
+          params: { by: ['region'], period: 'week', base: '2026-W41', target: '2026-W42', measure: 'gmv',
+                    headline: { base: 'weekly.r0.gmv', target: 'weekly.r1.gmv' }, label: '销售额' } },
+      ],
+    }),
+    node('write', 'report', 980, 200, '写周报', { instructions: '写本周销售的异动' }),
+    node('done', 'output', 1300, 200, '成果', { fields: [{ name: 'answer', value: '{{ nodes.write.text }}' }] }),
+  ],
+  edges: [edge('start', 'weekly'), edge('start', 'by_region'), edge('start', 'think'), edge('weekly', 'card'),
+    edge('by_region', 'card'), edge('think', 'card'), edge('card', 'write'), edge('write', 'done')],
+}
+
 const wf = (id, extra) => ({
   id, name: extra.name, description: extra.description ?? '检查脚本伪造的工作流', graph: extra.graph ?? GRAPH,
   tags: extra.tags ?? [], version: extra.version ?? 3, is_template: !!extra.is_template,
@@ -147,6 +172,7 @@ const FAKES = {
   'st-lint': wf('st-lint', { name: '__studio_check_lint__', graph: TEAM_GRAPH, version: 1 }),
   'st-ev': wf('st-ev', { name: '__studio_check_ev__', graph: EV_GRAPH, version: 2 }),
   'st-merge': wf('st-merge', { name: '__studio_check_merge__', graph: MERGE_GRAPH, version: 1 }),
+  'st-analysis': wf('st-analysis', { name: '__studio_check_analysis__', graph: ANALYSIS_GRAPH, version: 1 }),
 }
 const VERSIONS = [
   { id: 'v3', version: 3, note: '', created_at: '2026-09-26T02:00:00Z' },
@@ -3142,6 +3168,206 @@ await section('受管工作流在画布上改过之后：页头、版本历史
   }
 })
 
+// ================================================================ 分析卡
+await section('分析卡：叫法和服务端契约一字不差（算子、模式、阈值、字段、报告节点两个新字段的取值）', async () => {
+  // 叫法只有服务端契约一份（analysis/contract.py 的 OP_LABELS / MODE_LABELS / LABELS，engine/labels.py 引用它们）：界面上的
+  // 字和校验报错里的字不一样，人就找不到报错说的是哪一项。不开浏览器也能比，terms 借页面的模块加载
+  const py = readFileSync(new URL('../backend/app/analysis/contract.py', import.meta.url), 'utf8')
+  const dict = (name) => Object.fromEntries([...(py.match(new RegExp(`^${name}: dict\\[str, [^\\]]+\\] = \\{([\\s\\S]*?)^\\}`, 'm'))?.[1] ?? '')
+    .matchAll(/"([^"]+)":\s*"([^"]+)"/g)].map((m) => [m[1], m[2]]))
+  const ranges = Object.fromEntries([...(py.match(/^RANGES: dict\[[^\]]+\]\] = \{([\s\S]*?)^\}/m)?.[1] ?? '')
+    .matchAll(/"([^"]+)":\s*\((float|int),\s*([\d._]+),\s*([\d._]+)\)/g)].map((m) => [m[1], [m[2], Number(m[3].replace(/_/g, '')), Number(m[4].replace(/_/g, ''))]]))
+  const labelsPy = readFileSync(new URL('../backend/app/engine/labels.py', import.meta.url), 'utf8')
+  const fieldBlock = labelsPy.match(/"analysis": \{([^}]*)\}/)?.[1] ?? ''
+  const fields = Object.fromEntries([...fieldBlock.matchAll(/"([^"]+)":\s*"([^"]+)"/g)].map((m) => [m[1], m[2]]))
+  const opt = (key) => Object.fromEntries([...(labelsPy.match(new RegExp(`"${key}": \\{([^}]*)\\}`))?.[1] ?? '')
+    .matchAll(/"([^"]+)":\s*"([^"]+)"/g)].map((m) => [m[1], m[2]]))
+  const top = Object.fromEntries([...(labelsPy.match(/^FIELD_LABEL: dict\[str, str\] = \{([\s\S]*?)^\}/m)?.[1] ?? '')
+    .matchAll(/"(analysis_claims|analysis_scope)":\s*"([^"]+)"/g)].map((m) => [m[1], m[2]]))
+  const { ctx, page } = await open({ path: '/studio/st-analysis' })
+  const t = await page.evaluate(async () => {
+    const m = await window.__appImport('/src/lib/termsAnalysis.ts')
+    return { op: m.ANALYSIS_OP_LABEL, mode: m.ANALYSIS_MODE_LABEL, setting: m.ANALYSIS_SETTING_LABEL, range: m.ANALYSIS_SETTING_RANGE,
+             field: m.ANALYSIS_FIELD_LABEL, option: m.ANALYSIS_OPTION_LABEL, keys: m.ANALYSIS_SETTING_KEYS }
+  })
+  const same = (a, b) => JSON.stringify(Object.entries(a).sort()) === JSON.stringify(Object.entries(b).sort())
+  const diff = (a, b) => Object.keys({ ...a, ...b }).filter((k) => a[k] !== b[k]).map((k) => `${k}：${a[k]} ≠ ${b[k]}`).join('；')
+  check('从 contract.py 读出了 OP_LABELS、MODE_LABELS、LABELS、RANGES', Object.keys(dict('OP_LABELS')).length === 5
+    && Object.keys(dict('MODE_LABELS')).length === 6 && Object.keys(dict('LABELS')).length >= 17 && Object.keys(ranges).length >= 17)
+  check('算子的叫法和 OP_LABELS 一字不差', same(t.op, dict('OP_LABELS')), diff(t.op, dict('OP_LABELS')))
+  check('模式的叫法和 MODE_LABELS 一字不差', same(t.mode, dict('MODE_LABELS')), diff(t.mode, dict('MODE_LABELS')))
+  check('阈值的叫法和 LABELS 一字不差、键的先后同 DEFAULTS', same(t.setting, dict('LABELS'))
+    && t.keys.join(',') === Object.keys(dict('LABELS')).join(','), diff(t.setting, dict('LABELS')))
+  const r = Object.fromEntries(Object.entries(t.range).map(([k, v]) => [k, v.join(',')]))
+  const rp = Object.fromEntries(Object.entries(ranges).map(([k, v]) => [k, v.join(',')]))
+  check('阈值的范围和 RANGES 一致', same(r, rp), diff(r, rp))
+  check('分析卡字段的叫法同 engine/labels.py（分析项、阈值覆盖）', t.field.items === fields.items && t.field.overrides === fields.overrides,
+    JSON.stringify(fields))
+  check('报告节点两个新字段的叫法同 FIELD_LABEL', t.field.analysis_claims === top.analysis_claims && t.field.analysis_scope === top.analysis_scope,
+    JSON.stringify(top))
+  check('……取值的短说法同 _OPTION_LABEL', same(t.option.analysis_claims, opt('analysis_claims')) && same(t.option.analysis_scope, opt('analysis_scope')),
+    `${JSON.stringify(opt('analysis_claims'))} ${JSON.stringify(opt('analysis_scope'))}`)
+  await ctx.close()
+})
+
+await section('分析卡：节点库认得、卡片摘要、检查器按方法摆参数格、只列上游查询节点、阈值覆盖、服务端校验落到那一项、定位、连线带出输入', async () => {
+  const { page, errors } = await open({ path: '/studio/st-analysis' })
+  const entry = page.locator('[aria-label="节点库"] button[data-node-type="analysis"]')
+  check('节点库里有「分析卡」', await entry.count() === 1 && (await entry.innerText()).includes('分析卡'))
+  const card = page.locator('.react-flow__node[data-id="card"]')
+  const summary = (await card.innerText().catch(() => '')).replace(/\s+/g, ' ')
+  check('卡片摘要写做哪几种分析（契约的叫法）', summary.includes('波动基线、变化拆解（可加）'), summary.slice(0, 80))
+  const tint = await card.evaluate((el) => getComputedStyle(el.querySelector('.nt-analysis') ?? el).getPropertyValue('--nt').trim()).catch(() => '')
+  check('分析卡有自己的类型色（--nt-analysis）', !!tint, tint)
+
+  await S(page, () => window.__studio.getState().select('card'))
+  const box = page.locator('[data-inspector-sheet] [data-field="items"]')
+  await box.waitFor({ timeout: 4000 }).catch(() => {})
+  if (SHOTS) {
+    for (const theme of ['dark', 'light']) {
+      await page.evaluate((th) => document.documentElement.setAttribute('data-theme', th), theme)
+      await page.waitForTimeout(250)
+      await page.screenshot({ path: `${SHOTS}/studio-analysis-${theme}.png` })
+    }
+    await page.evaluate(() => document.documentElement.setAttribute('data-theme', 'dark'))
+  }
+  const items = box.locator('[data-item]')
+  check('两项各一行，标识 wow、where', await items.count() === 2
+    && (await items.nth(0).locator('[data-sub="id"] input').inputValue()) === 'wow'
+    && (await items.nth(1).locator('[data-sub="id"] input').inputValue()) === 'where')
+  const ops = await items.nth(0).locator('[data-sub="op"] option:not([disabled])').allInnerTexts()
+  check('分析方法下拉是五个算子的叫法', ops.join('、') === '变化拆解、波动基线、趋势判断、差异检验、可比性检查', ops.join('、'))
+  check('波动基线没有模式；变化拆解有模式（可加 / 比率 / 乘积）', await items.nth(0).locator('[data-sub="mode"]').count() === 0
+    && (await items.nth(1).locator('[data-sub="mode"] option:not([disabled])').allInnerTexts()).join('、') === '可加、比率、乘积')
+  const from = await items.nth(1).locator('[data-sub="from"] option:not([disabled])').evaluateAll((els) => els.map((e) => e.value))
+  check('输入只列上游的查询节点，不列模型调用', from.sort().join(',') === 'by_region,weekly', from.join(','))
+  const params = await items.nth(1).locator('[data-analysis-params] [data-sub]').evaluateAll((els) => els.map((e) => e.dataset.sub))
+  check('可加拆解摆维度列、期次、基期、本期、度量列、度量的叫法', params.join(',') === 'params.by,params.period,params.base,params.target,params.measure,params.label',
+    params.join(','))
+  check('……多列用顿号写', (await items.nth(1).locator('[data-sub="params.by"] input').inputValue()) === 'region')
+  check('……总数核对两格照原样', (await items.nth(1).locator('[data-sub="params.headline.base"] input').inputValue()) === 'weekly.r0.gmv'
+    && (await items.nth(1).locator('[data-sub="params.headline.target"] input').inputValue()) === 'weekly.r1.gmv')
+  check('……在验证什么写在这一项上', (await items.nth(1).locator('[data-sub="hypothesis"] input').inputValue()) === '下降是不是集中在某个地区')
+  check('下游取判定的写法写在标识下面', (await items.nth(0).innerText()).includes('nodes.card.results.wow.verdict'))
+  const config = () => S(page, () => JSON.stringify(window.__studio.getState().nodes.find((n) => n.id === 'card')?.data.config))
+  const cfg = async () => JSON.parse(await config())
+
+  // 改模式：参数只留这种模式用得到的（度量列去掉，补分子、分母两格）
+  await items.nth(1).locator('[data-sub="mode"] select').selectOption('ratio')
+  await page.waitForTimeout(150)
+  const ratio = (await cfg()).items[1]
+  check('换成比率模式：度量列去掉，参数格换成分子、分母', ratio.mode === 'ratio' && !('measure' in ratio.params)
+    && await items.nth(1).locator('[data-sub="params.numerator"]').count() === 1, JSON.stringify(ratio.params))
+  await items.nth(1).locator('[data-sub="params.numerator"] input').fill('refunds')
+  await items.nth(1).locator('[data-sub="params.by"] input').fill('region、channel')
+  await page.waitForTimeout(150)
+  const typed = (await cfg()).items[1].params
+  check('……填的列写回配置（几列写成列表）', typed.numerator === 'refunds' && JSON.stringify(typed.by) === '["region","channel"]', JSON.stringify(typed))
+  // 换分析方法：基线换成趋势判断，基线才有的参数去掉
+  await items.nth(0).locator('[data-sub="op"] select').selectOption('trend')
+  await page.waitForTimeout(150)
+  const trend = (await cfg()).items[0]
+  check('换成趋势判断：留下期次列、度量列、叫法，输入不动', trend.op === 'trend' && trend.from === 'weekly'
+    && Object.keys(trend.params).sort().join(',') === 'label,measure,period', JSON.stringify(trend))
+  // 两份查询：基期、本期各选一个查询节点，期次三格不写
+  await items.nth(1).locator('[data-analysis-two]').check()
+  await page.waitForTimeout(150)
+  const two = (await cfg()).items[1]
+  check('勾「基期、本期是两份查询」：输入挪到基期查询，期次三格去掉', two.params.base_query === 'by_region' && !two.from
+    && !('period' in two.params) && await items.nth(1).locator('[data-sub="params.target_query"] select').count() === 1, JSON.stringify(two))
+  await items.nth(1).locator('[data-analysis-two]').uncheck()
+  await page.waitForTimeout(150)
+  check('……取消勾选：基期查询回到输入', (await cfg()).items[1].from === 'by_region')
+
+  // 添加一项：新标识不重名；变化拆解才能下钻，下钻的上一层只列排在前面的变化拆解
+  await box.locator('[data-analysis-add]').click()
+  await page.waitForTimeout(150)
+  const added = (await cfg()).items
+  check('添加分析项：标识 a3，还没选分析方法', added.length === 3 && added[2].id === 'a3' && !added[2].op, JSON.stringify(added[2]))
+  await items.nth(2).locator('[data-sub="op"] select').selectOption('decompose')
+  await page.waitForTimeout(150)
+  const parentOpts = await items.nth(2).locator('[data-sub="params.parent"] select option').evaluateAll((els) => els.map((e) => e.value))
+  check('……选了变化拆解：下钻的上一层可以选前面那一项拆解（where）', parentOpts.includes('where') && !parentOpts.includes('wow'), parentOpts.join(','))
+  await items.nth(2).locator('[data-sub="params.parent"] select').selectOption('where')
+  await items.nth(2).locator('[data-sub="params.parent"] input').fill('华东')
+  await page.waitForTimeout(150)
+  check('……下钻写成上一层的标识和那一段的维度值', JSON.stringify((await cfg()).items[2].params.parent) === '{"of":"where","key":["华东"]}',
+    JSON.stringify((await cfg()).items[2].params))
+  await items.nth(2).locator('[data-sub="id"] input').fill('wow')
+  await page.waitForTimeout(150)
+  check('标识和另一项重复：当场说明', (await items.nth(2).innerText()).includes('与另一项重复'))
+  await items.nth(2).locator('[data-sub="id"] input').fill('east')
+
+  // 阈值覆盖：从没写的里面选一项，范围当场说
+  const ov = page.locator('[data-inspector-sheet] [data-field="overrides"]')
+  check('没写覆盖时写明全部跟随设置', (await ov.innerText()).includes('跟随设置'))
+  await ov.locator('[data-analysis-add-override]').selectOption('main_share')
+  await page.waitForTimeout(150)
+  await ov.locator('[data-analysis-override="main_share"]').fill('0.95')
+  await page.waitForTimeout(150)
+  check('……超出范围当场说（说法同服务端）', (await ov.innerText()).includes('「主要来源」的占比要在 0.3 到 0.9 之间'), (await ov.innerText()).slice(0, 120))
+  await ov.locator('[data-analysis-override="main_share"]').fill('0.6')
+  await page.waitForTimeout(150)
+  check('……填对了写回成数', JSON.stringify((await cfg()).overrides) === '{"main_share":0.6}', JSON.stringify((await cfg()).overrides))
+  await ov.getByRole('button', { name: '删除阈值覆盖「「主要来源」的占比」' }).click()
+  await page.waitForTimeout(150)
+  check('……删掉最后一项：配置里不留空对象', (await cfg()).overrides === undefined, String((await cfg()).overrides))
+
+  // 服务端校验：参数里写了阈值、下钻的上一层排在后面——问题落在那一项的那一格下面（校验是真的服务端算的）
+  await S(page, () => {
+    const st = window.__studio.getState()
+    const n = st.nodes.find((x) => x.id === 'card')
+    const next = n.data.config.items.map((it) => (it.id === 'where' ? { ...it, params: { ...it.params, k_warn: 3 } } : it))
+    st.updateNode('card', { config: { ...n.data.config, items: next } })
+  })
+  await S(page, () => window.__studio.getState().validate())
+  await page.waitForTimeout(800)
+  const whereText = await items.nth(1).innerText()
+  check('参数里写了阈值：服务端的说法落在这一项下面', whereText.includes('阈值不写在分析项里'), whereText.slice(-160))
+  const r = await page.evaluate(async () => {
+    const m = await window.__appImport('/src/canvas/InspectorSheet.tsx')
+    window.__studio.getState().select(null)
+    await new Promise((res) => setTimeout(res, 100))
+    m.revealField('card', 'items[1].params.headline.target', { focus: true })
+    await new Promise((res) => setTimeout(res, 600))
+    const a = document.activeElement
+    return { at: a?.closest('[data-sub]')?.dataset.sub ?? null, value: (a && 'value' in a) ? a.value : null }
+  }).catch((e) => ({ error: e.message.split('\n')[0] }))
+  check('问题落在 items[1].params.headline.target：定位到那一格、光标落进去', r.at === 'params.headline.target' && r.value === 'weekly.r1.gmv',
+    JSON.stringify(r))
+  const r2 = await page.evaluate(async () => {
+    const m = await window.__appImport('/src/canvas/InspectorSheet.tsx')
+    m.revealField('card', 'items[2].params.parent.key', { focus: true })
+    await new Promise((res) => setTimeout(res, 600))
+    return document.activeElement?.closest('[data-sub]')?.dataset.sub ?? null
+  }).catch((e) => e.message.split('\n')[0])
+  check('……落到没有自己一格的子键（params.parent.key）：退到所在的那一块', r2 === 'params.parent', String(r2))
+
+  // 报告节点：分析说法核对缺省是「重写一次后标出」，分析范围在高级选项里
+  await S(page, () => window.__studio.getState().select('write'))
+  await page.waitForSelector('[data-inspector-sheet] [data-field="analysis_claims"]', { timeout: 4000 }).catch(() => {})
+  const claims = page.locator('[data-inspector-sheet] [data-field="analysis_claims"] select')
+  const shown = await claims.evaluate((el) => el.options[el.selectedIndex]?.textContent ?? '').catch(() => '')
+  check('报告节点「分析说法核对」没写时显示缺省（重写一次后标出）', shown.startsWith('重写一次后标出'), shown)
+  const opts = await claims.locator('option').allInnerTexts().catch(() => [])
+  check('……三个取值以服务端的短说法开头', opts.length === 3 && opts[1].startsWith('重写后仍有的计入缺口') && opts[2] === '不核对', opts.join(' | '))
+
+  // 连线带出输入：查询节点连到新拖进来的分析卡，第一项（还没选输入的）选上它；模型调用不算
+  const mid = await S(page, () => {
+    window.__studio.getState().addNode('analysis', { x: 640, y: 620 })
+    return window.__studio.getState().selectedId
+  })
+  const connect = (a, b) => S(page, ([x, y]) => window.__studio.getState().onConnect({ source: x, target: y, sourceHandle: 'out', targetHandle: null }), [a, b])
+  const firstFrom = () => S(page, (x) => window.__studio.getState().nodes.find((n) => n.id === x)?.data.config.items?.[0]?.from ?? '', mid)
+  await connect('think', mid)
+  check('新拖进来的分析卡带一项空的；模型调用连进来不算输入', await firstFrom() === '')
+  await connect('weekly', mid)
+  check('……查询节点连进来：那一项的输入选上它', await firstFrom() === 'weekly', await firstFrom())
+  await S(page, () => window.__studio.getState().undo())
+  check('……撤销连线时一起退回', await firstFrom() === '')
+  check('全程页面没有报错', errors.length === 0, errors.slice(0, 2).join('；'))
+})
+
 check('整个检查没有弹出原生 confirm / prompt', dialogs === 0, `${dialogs} 次`)
 check('检查没有写库（兜底拦下的写请求）', writes.length === 0, writes.slice(0, 4).join('、'))