diff --git a/frontend/src/lib/termsEval.ts b/frontend/src/lib/termsEval.ts
index cfc9f652..04e268cd 100644
--- a/frontend/src/lib/termsEval.ts
+++ b/frontend/src/lib/termsEval.ts
@@ -91,6 +91,12 @@ export const EVAL_TEXT = {
     upgradeHint: '每道题三选一，没有缺省：改为固定新版本（历史在这里断开）、新旧两版各算一份期望、或旧题归档另建新题',
     runsEmpty: '还没有评测记录',
     runsLabel: '历次评测',
+    cleanup: '清理更早的明细',
+    cleanupHint: '评测收尾时会按保留规则自动清理，这里可以手动再清一次',
+    cleanupTitle: (name: string) => `清理「${name}」更早的评测明细？`,
+    cleanupConsequences: ['按设置里的保留次数清理更早的评测：逐题结果、作答和算期望的运行一起删除，不可恢复',
+      '设为基线的那次、A/B 的另几臂留着；每次评测的总表永久保留'],
+    cleaned: (runs: number, files: number) => (runs ? `已清理 ${runs} 次评测的明细，回收了 ${files} 个工件文件` : '没有要清理的评测'),
     schedules: '定时评测',
     noSchedules: '这个题集还没有定时任务',
     budget: '预算',
@@ -112,6 +118,7 @@ export const EVAL_TEXT = {
   caseRow: {
     seq: (n: number) => `第 ${n} 题`,
     kind: { ask: '问数据题', workflow: '工作流题' } as Record<'ask' | 'workflow', string>,
+    workflow: (name: string, v: number | null) => `工作流 ${name} v${v ?? '?'}`,
     disabled: '未启用',
     archived: '已归档',
     supersededBy: (n: string) => `已被新题 ${n} 取代`,
@@ -172,6 +179,17 @@ export const EVAL_TEXT = {
       special: '问法问的是没有的东西（期望答「没有数据」），或者问法含糊（期望助手反问）',
     } as Record<'sql' | 'caliber' | 'from_run' | 'special', string>,
     question: '问法',
+    workflowQuestion: '这道题考什么（可选）',
+    workflowQuestionHint: '工作流题不需要问法：写一句说明，列表里认得出是哪道题',
+    caseKind: '题目种类',
+    caseKindHint: {
+      ask: '问法交给助手：建图、运行、复核，和问数据页同一条路',
+      workflow: '固定一个工作流的已发布版本，按输入运行一次（不建图、不复核），考的是这一版本身',
+    } as Record<'ask' | 'workflow', string>,
+    workflow: '工作流',
+    workflowPlaceholder: '选择已发布的工作流',
+    workflowNone: '还没有已发布的工作流',
+    workflowVersion: '固定的版本',
     questionPlaceholder: '例如：上周按支付日统计、不含退款的销售额是多少？',
     questionHint: '可以写 {{ schedule.last_week_start }} 这类时间变量，按评测的截至时间计算',
     scope: '限定的数据源',
@@ -473,6 +491,9 @@ export const EVAL_TEXT = {
     filterAll: '全部',
     filter: '只看',
     repeat: (n: number) => `第 ${n + 1} 次`,
+    attempts: (n: number) => `这一题自动重来了 ${n} 次`,
+    turnReason: '作答的经过',
+    answer: '答案原文',
     expand: '展开',
     collapse: '收起',
   },
diff --git a/frontend/src/pages/evals/CaseEditor.tsx b/frontend/src/pages/evals/CaseEditor.tsx
index e89fac44..12a27a9c 100644
--- a/frontend/src/pages/evals/CaseEditor.tsx
+++ b/frontend/src/pages/evals/CaseEditor.tsx
@@ -10,7 +10,7 @@ import { EVAL_TEXT } from '../../lib/termsEval'
 import { useCatalog } from '../../store/catalog'
 import type { Run } from '../../types'
 import type {
-  EvalCaliber, EvalCase, EvalCaseBody, EvalExpectation, EvalGold, EvalPreview, EvalRunQuery, EvalSuite, EvalTrial,
+  EvalCaliber, EvalCase, EvalCaseBody, EvalCaseKind, EvalExpectation, EvalGold, EvalPreview, EvalRunQuery, EvalSuite, EvalTrial,
 } from '../../typesEval'
 import { graphFields, type InputField } from '../schedules/model'
 import { cellText, expectationText, goldValueText } from './parts'
@@ -69,6 +69,11 @@ export function CaseEditor({ open, onClose, suite, editing, initialQuestion, onS
   // 从运行认定
   const [runId, setRunId] = useState('')
   const [queryId, setQueryId] = useState('')
+  // 题目种类：问数据题（问法交给助手作答）或工作流题（固定一个已发布版本，按输入跑一次，考的是这一版本身）
+  const [caseKind, setCaseKind] = useState<EvalCaseKind>('ask')
+  const [workflowId, setWorkflowId] = useState('')
+  const [workflowVersion, setWorkflowVersion] = useState<number | null>(null)
+  const [workflowInputs, setWorkflowInputs] = useState<Record<string, string>>({})
   // 试算、保存
   const [trial, setTrial] = useState<EvalTrial | null>(null)
   const [trialing, setTrialing] = useState(false)
@@ -101,6 +106,10 @@ export function CaseEditor({ open, onClose, suite, editing, initialQuestion, onS
     setInputs(c?.gold?.kind === 'caliber' ? { ...(c.gold.inputs ?? {}) } : {})
     setRunId(c?.gold?.kind === 'from_run' ? c.gold.run_id ?? '' : '')
     setQueryId(c?.gold?.kind === 'from_run' ? c.gold.query ?? '' : '')
+    setCaseKind(c?.kind ?? 'ask')
+    setWorkflowId(c?.workflow_id ?? '')
+    setWorkflowVersion(c?.workflow_version ?? null)
+    setWorkflowInputs({ ...(c?.input_template ?? {}) })
     setTrial(null)
     setError(null)
   }, [open, editing?.id, editing?.version, initialQuestion])
@@ -144,7 +153,13 @@ export function CaseEditor({ open, onClose, suite, editing, initialQuestion, onS
       tags: tags.split(/[,，、\n]/).map((t) => t.trim()).filter(Boolean),
     }
     if (!fixedFromRun) out.gold = gold()
-    if (!editing) out.kind = 'ask'
+    const workflow = caseKind === 'workflow' && (mode === 'sql' || mode === 'caliber')
+    if (!editing) out.kind = workflow ? 'workflow' : 'ask'
+    if (workflow) {
+      out.workflow_id = workflowId
+      out.workflow_version = workflowVersion
+      out.input_template = Object.fromEntries(Object.entries(workflowInputs).filter(([, v]) => v.trim() !== ''))
+    }
     return out
   }
 
@@ -215,12 +230,16 @@ export function CaseEditor({ open, onClose, suite, editing, initialQuestion, onS
 
   const fieldError = (...fields: string[]) => (error && error.field && fields.some((f) => error.field === f || error.field!.startsWith(`${f}.`))
     ? error.message : undefined)
-  const generalError = error && !(error.field && ['question', 'gold.sql', 'gold.datasource_id', 'gold.inputs', 'query', 'tags']
+  const generalError = error && !(error.field && ['question', 'gold.sql', 'gold.datasource_id', 'gold.inputs', 'query', 'tags',
+    'workflow_id', 'workflow_version', 'input_template']
     .some((f) => error.field === f || error.field!.startsWith(`${f}.`))) ? error.message : null
   const needsMain = mode !== 'special'
+  /** 工作流题只配「写 SQL 定期望」「选口径」两种：反问是助手的事，工作流不会反问；从运行认定只认问数据运行里的查询 */
+  const workflowCase = caseKind === 'workflow' && (mode === 'sql' || mode === 'caliber')
   // 关掉之前要不要先问：新建时填了东西、编辑时改了要试算的那几样
   const dirty = editing ? changed() : !!(question.trim() || sql.trim() || main || points.trim() || runId)
   const ready = (!needsMain || !!main) && (mode !== 'from_run' || fixedFromRun || (!!runId && !!queryId))
+    && (!workflowCase || (!!workflowId && !!workflowVersion))
     && (mode !== 'caliber' || !!caliber) && (!(mode === 'sql' || (mode === 'special' && special === 'none')) || (!!sourceId && !!sql.trim()))
 
   return (
@@ -256,8 +275,16 @@ export function CaseEditor({ open, onClose, suite, editing, initialQuestion, onS
           <p className="mt-1 text-2xs leading-relaxed text-faint">{T.modeHint[mode]}</p>
         </div>
 
-        <Field label={T.question} htmlFor="eval-case-question" hint={T.questionHint} error={fieldError('question')}
-               required={mode !== 'from_run'}>
+        {(mode === 'sql' || mode === 'caliber') && (
+          <CaseKindPart kind={caseKind} setKind={setCaseKind} disabled={!!editing} workflowId={workflowId}
+                        setWorkflowId={(v) => { setWorkflowId(v); setWorkflowInputs({}) }} version={workflowVersion}
+                        setVersion={setWorkflowVersion} inputs={workflowInputs} setInputs={setWorkflowInputs}
+                        error={fieldError('workflow_id', 'workflow_version', 'input_template')} />
+        )}
+
+        <Field label={workflowCase ? T.workflowQuestion : T.question} htmlFor="eval-case-question"
+               hint={workflowCase ? T.workflowQuestionHint : T.questionHint} error={fieldError('question')}
+               required={mode !== 'from_run' && !workflowCase}>
           <textarea id="eval-case-question" className="field min-h-16" value={question} maxLength={2000}
                     placeholder={T.questionPlaceholder} onChange={(e) => setQuestion(e.target.value)} data-eval-case-question="" />
         </Field>
@@ -342,6 +369,72 @@ export function CaseEditor({ open, onClose, suite, editing, initialQuestion, onS
   )
 }
 
+// ---------------------------------------------------------------------------
+// 题目种类：问数据题 / 工作流题
+// ---------------------------------------------------------------------------
+
+/**
+ * 工作流题：固定一个工作流的已发布版本（服务端核对已发布、未退役），按输入模板跑一次探索运行，考的是这一版本身（不建图、
+ * 不复核）。输入模板按那一版入口节点的字段填，可以写时间变量
+ */
+function CaseKindPart({ kind, setKind, disabled, workflowId, setWorkflowId, version, setVersion, inputs, setInputs, error }: {
+  kind: EvalCaseKind; setKind: (k: EvalCaseKind) => void; disabled: boolean; workflowId: string; setWorkflowId: (v: string) => void
+  version: number | null; setVersion: (v: number | null) => void; inputs: Record<string, string>
+  setInputs: (v: Record<string, string>) => void; error?: string
+}) {
+  const workflows = useCatalog((s) => s.workflows)
+  const published = workflows.filter((w) => w.published_version)
+  const [fields, setFields] = useState<InputField[] | null>(null)
+  useEffect(() => {
+    if (kind !== 'workflow' || !workflowId || !version) { setFields(null); return }
+    setFields(null)
+    api.workflows.version(workflowId, version).then((v) => setFields(graphFields(v.graph)), () => setFields([]))
+  }, [kind, workflowId, version])
+  return (
+    <div className="space-y-2" data-eval-case-kind={kind}>
+      <div className="label">{T.caseKind}</div>
+      <div className="flex flex-wrap gap-3 text-xs">
+        {(['ask', 'workflow'] as const).map((k) => (
+          <label key={k} className="flex items-center gap-1.5">
+            <input type="radio" name="eval-case-kind" checked={kind === k} disabled={disabled} onChange={() => setKind(k)}
+                   data-eval-case-kind-pick={k} />
+            {EVAL_TEXT.caseRow.kind[k]}
+          </label>
+        ))}
+      </div>
+      <p className="text-2xs leading-relaxed text-faint">{T.caseKindHint[kind]}</p>
+      {kind === 'workflow' && (
+        <div className="grid gap-3 rounded-md border p-2.5 sm:grid-cols-[minmax(0,1fr)_8rem]">
+          <Field label={T.workflow} htmlFor="eval-case-workflow" required>
+            <select id="eval-case-workflow" className="field" value={workflowId} disabled={disabled} data-eval-case-workflow=""
+                    onChange={(e) => {
+                      const w = published.find((x) => x.id === e.target.value)
+                      setWorkflowId(e.target.value)
+                      setVersion(w?.published_version ?? null)
+                    }}>
+              <option value="">{published.length ? T.workflowPlaceholder : T.workflowNone}</option>
+              {published.map((w) => <option key={w.id} value={w.id}>{w.name}</option>)}
+            </select>
+          </Field>
+          <Field label={T.workflowVersion} htmlFor="eval-case-workflow-version" required>
+            <input id="eval-case-workflow-version" className="field tnum" inputMode="numeric" value={version ?? ''} disabled={disabled}
+                   onChange={(e) => setVersion(e.target.value.trim() ? Number(e.target.value) : null)} />
+          </Field>
+          {fields === null && workflowId && version && <p className="text-2xs text-dim sm:col-span-2"><Spinner size={10} /> {T.caliberInputsLoading}</p>}
+          {fields?.map((f) => (
+            <Field key={f.name} label={f.name} htmlFor={`eval-case-input-${f.name}`} required={!!f.required} hint={f.description}
+                   className="sm:col-span-2">
+              <input id={`eval-case-input-${f.name}`} className="field mono text-xs" value={inputs[f.name] ?? ''}
+                     onChange={(e) => setInputs({ ...inputs, [f.name]: e.target.value })} data-eval-case-input={f.name} />
+            </Field>
+          ))}
+          {error && <p className="text-2xs text-[var(--err)] sm:col-span-2" role="alert">{error}</p>}
+        </div>
+      )}
+    </div>
+  )
+}
+
 // ---------------------------------------------------------------------------
 // 写 SQL 定期望
 // ---------------------------------------------------------------------------
diff --git a/frontend/src/pages/evals/EvaluationPage.tsx b/frontend/src/pages/evals/EvaluationPage.tsx
index c1784b63..11d473df 100644
--- a/frontend/src/pages/evals/EvaluationPage.tsx
+++ b/frontend/src/pages/evals/EvaluationPage.tsx
@@ -414,6 +414,7 @@ function Results({ id, live, focusCase }: { id: string; live: boolean; focusCase
                         </p>
                       ))}
                       {r.sql_match?.reason && <p className="text-faint">{r.sql_match.reason}</p>}
+                      <ResultDetail evaluationId={id} resultId={r.id} />
                       <div className="flex flex-wrap gap-2">
                         {r.run_id && <Link className="btn btn-sm" to={`/runs/${encodeURIComponent(r.run_id)}`} data-eval-result-run="">{T.run.openRun}</Link>}
                         {r.gold_run_ids.slice(0, 2).map((g) => (
@@ -432,6 +433,31 @@ function Results({ id, live, focusCase }: { id: string; live: boolean; focusCase
   )
 }
 
+/** 一题的明细（GET …/results/{id} 的明细工件）：这一轮作答的经过（重来了几次、原因）和读出来的答案原文。明细清理了的不显示 */
+function ResultDetail({ evaluationId, resultId }: { evaluationId: string; resultId: string }) {
+  const [details, setDetails] = useState<Record<string, any> | null | undefined>(undefined)
+  useEffect(() => {
+    api.evals.result(evaluationId, resultId).then((r) => setDetails(r.details ?? null), () => setDetails(null))
+  }, [evaluationId, resultId])
+  if (details === undefined) return <Spinner size={11} />
+  if (!details) return null
+  const turn = details.turn ?? {}
+  const answer = String(details.answer?.text ?? turn.reply ?? '').trim()
+  const R = T.results
+  return (
+    <div className="space-y-1" data-eval-result-details="">
+      {(turn.attempts ?? 0) > 0 && <p className="text-faint">{R.attempts(turn.attempts)}</p>}
+      {turn.reason && <p className="text-dim"><span className="text-faint">{R.turnReason}：</span>{turn.reason}</p>}
+      {answer && (
+        <details>
+          <summary className="cursor-pointer text-2xs text-dim">{R.answer}</summary>
+          <p className="mt-1 max-h-48 overflow-auto whitespace-pre-wrap break-words text-dim">{answer}</p>
+        </details>
+      )}
+    </div>
+  )
+}
+
 // ---------------------------------------------------------------------------
 // 和上一次比
 // ---------------------------------------------------------------------------
diff --git a/frontend/src/pages/evals/SuiteDetail.tsx b/frontend/src/pages/evals/SuiteDetail.tsx
index 7faeda59..36d5e250 100644
--- a/frontend/src/pages/evals/SuiteDetail.tsx
+++ b/frontend/src/pages/evals/SuiteDetail.tsx
@@ -11,7 +11,7 @@ import { formatDateTime, formatTime } from '../../lib/format'
 import { timezoneOptions } from '../../lib/freq'
 import { leavePass } from '../../lib/leave'
 import { EVAL_TEXT, pct, rateText } from '../../lib/termsEval'
-import { useOnReconnect } from '../../store/catalog'
+import { useCatalog, useOnReconnect } from '../../store/catalog'
 import type {
   EvalCase, EvalRun, EvalSnapshot, EvalSuiteDetail, EvalTrial, EvalUpgradePolicy, EvalUpgradeState, EvalVariant,
 } from '../../typesEval'
@@ -111,6 +111,14 @@ export function SuiteDetail({ id }: { id: string }) {
     toast.ok(T.suite.deleted(suite.name))
     navigate('/evals', leavePass({ replace: true }))
   })
+  const cleanup = () => act('cleanup', async () => {
+    if (!suite) return
+    if (!(await confirmDialog({ title: T.suite.cleanupTitle(suite.name), consequences: T.suite.cleanupConsequences,
+      confirmLabel: T.suite.cleanup, danger: true }))) return
+    const r = await api.evals.cleanup(suite.id) as { runs?: number; files?: number }
+    toast.ok(T.suite.cleaned(Number(r.runs ?? 0), Number(r.files ?? 0)))
+    await load()
+  })
   const toggleCase = (c: EvalCase, enabled: boolean) => act(`toggle:${c.id}`, async () => {
     if (!suite) return
     await api.evals.updateCase(suite.id, c.id, { enabled, version: c.version })
@@ -210,7 +218,14 @@ export function SuiteDetail({ id }: { id: string }) {
               )}
             </div>
             <div role="tabpanel" id="eval-suite-panel-runs" aria-labelledby="eval-suite-tab-runs" hidden={tab !== 'runs'}
-                 className="mx-auto max-w-6xl p-4">
+                 className="mx-auto max-w-6xl space-y-3 p-4">
+              <div className="flex flex-wrap items-center justify-end gap-2">
+                <span className="text-2xs text-faint">{T.suite.cleanupHint}</span>
+                <button type="button" className="btn btn-sm" onClick={() => void cleanup()} disabled={!!busy || !runs?.length}
+                        data-eval-suite-cleanup="">
+                  {busy === 'cleanup' && <Spinner size={11} />} {T.suite.cleanup}
+                </button>
+              </div>
               <RunList runs={runs ?? []} />
             </div>
           </div>
@@ -242,6 +257,8 @@ function CaseRow({ c, busy, trial, archivedSuite, focused, onEdit, onToggle, onT
 }) {
   const last = c.trial
   const pending = !!c.upgrade?.pending
+  const workflows = useCatalog((s) => s.workflows)
+  const workflowName = (id: string | null) => workflows.find((w) => w.id === id)?.name ?? id ?? ''
   return (
     <li className={clsx('rounded-lg border bg-panel px-3 py-2.5', (!c.enabled || c.archived) && 'opacity-75',
       focused && 'border-[var(--accent)]')} data-eval-case={c.id} data-eval-case-focused={focused ? '' : undefined}
@@ -257,6 +274,9 @@ function CaseRow({ c, busy, trial, archivedSuite, focused, onEdit, onToggle, onT
           </div>
           <div className="mt-1 break-words text-sm" data-eval-case-text="">{c.question || '—'}</div>
           <div className="mt-1 text-2xs text-dim">
+            {c.kind === 'workflow' && (
+              <span data-eval-case-workflow={c.workflow_id ?? ''}>{T.caseRow.workflow(workflowName(c.workflow_id), c.workflow_version)} · </span>
+            )}
             <span title={T.caseRow.goldSource}>{goldText(c)}</span>
             {c.datasources.length > 0 && <span className="text-faint"> · {T.caseRow.datasources}：{c.datasources.map((d) => d.name ?? d.id).join('、')}</span>}
           </div>
diff --git a/scripts/check-evals.mjs b/scripts/check-evals.mjs
index d65f5594..50e238e7 100644
--- a/scripts/check-evals.mjs
+++ b/scripts/check-evals.mjs
@@ -231,6 +231,9 @@ const EVAL_RUN = { id: 'run-eval-ans', workflow_id: null, workflow_name: '评测
   origin_kind: 'eval', origin: { suite_id: 'suite-a', eval_run_id: 'ev-1', case_id: 'case-2', role: 'answer', repeat: 0 }, schedule_id: null,
   created_at: iso(NOW - 90 * 60e3), started_at: iso(NOW - 90 * 60e3), finished_at: iso(NOW - 89 * 60e3) }
 const RUN_GRAPH = { nodes: [{ id: 'out', type: 'output', position: { x: 0, y: 0 }, data: { label: '成果', config: {} } }], edges: [] }
+/** 已发布的工作流（工作流题要固定一个已发布版本） */
+const WF_WEEK = { id: 'wf-week', name: '周销售', description: '', graph: VERSION_WEEK.graph, tags: [], version: 4, is_template: false,
+  status: 'published', published_version: 3, published_level: 'governed', created_at: iso(NOW - 30 * 24 * H), updated_at: iso(NOW - 24 * H) }
 const BUILDER = { mode: 'local', username: 'jiang', display_name: '检查员甲', role: 'builder', groups: [], verified: true }
 const USER = { mode: 'local', username: 'li', display_name: '使用者乙', role: 'user', groups: [], verified: true }
 
@@ -370,6 +373,14 @@ function reply(state, method, path, url, body) {
       return { json: ev }
     }
   }
+  m = path.match(/^\/evals\/evaluations\/([^/]+)\/results\/([^/]+)$/)
+  if (method === 'GET' && m) {
+    const r = RESULTS.find((x) => x.id === m[2])
+    return r ? { json: { ...r, details: { turn: { attempts: 1, reason: null, reply: null }, answer: { text: `答案原文：上周销售额 ${r.scores[0]?.got ?? ''} 元` } } } }
+      : { status: 404, json: { detail: '这一题的结果不存在' } }
+  }
+  if (method === 'POST' && path === '/evals/suites/suite-a/cleanup') { write(); return { json: { runs: 2, results: 6, deleted_runs: 8, files: 3 } } }
+  if (method === 'GET' && path === '/workflows') return { json: [WF_WEEK] }
   if (method === 'GET' && path === '/evals/pairs/ev-pa') return { json: PAIR }
   if (method === 'GET' && path === '/evals/calibers') return { json: { calibers: CALIBERS } }
   if (method === 'GET' && path === '/workflows/wf-week/versions/3') return { json: VERSION_WEEK }
@@ -565,6 +576,10 @@ await section('题集详情：题目表、口径升版处置、启用开关、
   await page.locator('[data-eval-runs]').count()
   await page.getByRole('tab', { name: '评测记录' }).click()
   check('评测记录：历次评测链到详情', await page.locator('[data-eval-run-row="ev-1"]').getAttribute('href') === '/evals/runs/ev-1')
+  await page.locator('[data-eval-suite-cleanup]').click()
+  await page.locator('[role="dialog"]').getByRole('button', { name: '清理更早的明细' }).click()
+  check('清理更早的明细：先确认、再发一次，提示清了几次', await until(() => writes(state, /cleanup$/).length === 1)
+    && await until(async () => (await page.locator('[data-toast]').allInnerTexts()).some((x) => x.includes('已清理 2 次评测的明细'))))
   check('没有页面报错', !errors.length, errors.join(' ｜ '))
 })
 
@@ -652,6 +667,20 @@ await section('加题：选口径、从运行认定、期望反问', async () =>
   const ask = lastWrite(state, /^POST \/evals\/suites\/suite-b\/cases$/)?.body
   check('期望反问：期望只有「反问」，不带期望答案的来源', JSON.stringify(ask?.expectations) === JSON.stringify([{ kind: 'ask' }]) && ask?.gold === null,
     JSON.stringify(ask))
+  await page.locator('[data-eval-add-case]').click()
+  await page.locator('[data-eval-case-kind-pick="workflow"]').check()
+  await page.locator('[data-eval-case-workflow]').selectOption('wf-week')
+  await page.locator('[data-eval-case-input="week"]').fill('{{ schedule.last_week_start }}')
+  await page.locator('[data-eval-case-source]').selectOption('ds-shop')
+  await page.locator('[data-eval-case-sql]').fill('SELECT region AS 地区, SUM(amount) AS gmv FROM orders GROUP BY region')
+  await page.locator('[data-eval-case-preview]').click()
+  await page.locator('[data-eval-preview-cell="0:gmv"]').click()
+  await page.locator('[data-eval-case-save]').click()
+  await until(() => writes(state, /^POST \/evals\/suites\/suite-b\/cases$/).length === 3)
+  const wf = lastWrite(state, /^POST \/evals\/suites\/suite-b\/cases$/)?.body
+  check('工作流题：固定已发布的 v3、按入口字段填输入模板，期望照样由 SQL 算', wf?.kind === 'workflow' && wf.workflow_id === 'wf-week'
+    && wf.workflow_version === 3 && wf.input_template?.week === '{{ schedule.last_week_start }}' && wf.gold?.kind === 'sql',
+    JSON.stringify(wf))
   check('没有页面报错', !errors.length, errors.join(' ｜ '))
 })
 
@@ -744,6 +773,9 @@ await section('逐题结果、和上一次比、A/B 配对表', async () => {
   check('逐题：结论、期望和答案里的数、SQL 差在哪', t2.includes('不通过') && t2.includes('330') && t2.includes('310') && t2.includes('时间口径'), t2)
   check('一题的花费照服务端加好的写', (await page.locator('[data-eval-result="r-1"]').innerText()).includes('$0.014'))
   check('点开去运行详情', await page.locator('[data-eval-result-detail="r-2"] [data-eval-result-run]').getAttribute('href') === '/runs/run-answer-2')
+  await page.locator('[data-eval-result-details]').waitFor()
+  check('展开的明细写作答的经过和答案原文（明细工件）', (await text(page, '[data-eval-result-details]')).includes('自动重来了 1 次')
+    && (await page.locator('[data-eval-result-details] details').count()) === 1)
 
   await page.getByRole('tab', { name: '和上一次比' }).click()
   await page.locator('[data-eval-transitions]').waitFor()