diff --git a/.gitea/workflows/ci.yml b/.gitea/workflows/ci.yml
index 8f1a41c..95b4004 100644
--- a/.gitea/workflows/ci.yml
+++ b/.gitea/workflows/ci.yml
@@ -13,7 +13,7 @@ jobs:
         env:
           TOKEN: ${{ secrets.GITHUB_TOKEN }}
           UI_DEPLOY_KEY: ${{ secrets.UI_DEPLOY_KEY }}
-          UI_SHA: 316891b13acf4c35434a8d7e5e1a5f1d03b081cd
+          UI_SHA: 67beff6c45f2887b6634677843fae1485acbee4f
         run: |
           set -euo pipefail
           printf 'context server=%s repository=%s sha=%s\n' "${GITHUB_SERVER_URL}" "${GITHUB_REPOSITORY}" "${GITHUB_SHA}"
diff --git a/e2e/work-checkpoints.playwright.ts b/e2e/work-checkpoints.playwright.ts
index 85dca3e..f3cd632 100644
--- a/e2e/work-checkpoints.playwright.ts
+++ b/e2e/work-checkpoints.playwright.ts
@@ -33,20 +33,17 @@ async function setup(request:any,scenario:Scenario='normal',awaitTask=true,optio
  }).toBe(scenario.endsWith('failure')?'failed':'completed')
  return {project,conv,ids,sendRunId:sendBody.run.id}
 }
-async function waitForFailureNotice(request:any,seed:any){
+async function waitForFailedWorkCard(request:any,seed:any){
  await expect.poll(async()=>{
-  const response=await request.get(manifest.base+`/api/conversations/${seed.conv.id}`)
-  if(!response.ok())return false
-  const current=await response.json()
-  const children=current.runs.filter((r:any)=>r.kind==='subagent')
-  const noticeText=current.messages.filter((m:any)=>m.role==='system').flatMap((m:any)=>m.parts??[])
-   .filter((p:any)=>p.type==='text').map((p:any)=>p.text??'').join('\n')
-  const noticeCount=(noticeText.match(/后台子代理「/g)??[]).length
-  const noticeRuns=current.runs.filter((r:any)=>r.kind==='conversation'&&r.id!==seed.sendRunId)
-  return children.length>0&&noticeCount>=children.length&&noticeRuns.length>0&&
-   noticeRuns.every((r:any)=>['completed','failed','interrupted'].includes(r.status))
+  const out=await(await request.get(manifest.base+`/api/conversations/${seed.conv.id}/cards`)).json()
+  const card=out.cards.find((c:any)=>c.task)
+  return card?.task.executionState==='failed'&&card.status==='failed'&&Boolean(card.work)
  }).toBe(true)
 }
+async function expectOnlySendRun(request:any,seed:any){
+ const current=await(await request.get(manifest.base+`/api/conversations/${seed.conv.id}`)).json()
+ expect(current.runs.filter((r:any)=>r.kind==='conversation').map((r:any)=>r.id)).toEqual([seed.sendRunId])
+}
 async function open(page:any,seed:any){
  await page.addInitScript(({id}:{id:string})=>{localStorage.setItem('as.conv',id);localStorage.setItem('as.ui',JSON.stringify({theme:'light',rightCollapsed:true,fontSize:14}))},{id:seed.conv.id})
  await page.goto('/')
@@ -145,25 +142,27 @@ test('two sourced stages appear while the material task is still running',async(
 for(const scenario of ['zero-failure','partial-failure'] as const){
  test(`${scenario} shows the saved scope without a deliverable choice`,async({page,request})=>{
   const seed=await setup(request,scenario)
-  await waitForFailureNotice(request,seed)
+  await waitForFailedWorkCard(request,seed)
   await page.addInitScript(({id}:{id:string})=>{localStorage.setItem('as.conv',id);localStorage.setItem('as.ui',JSON.stringify({theme:'light',rightCollapsed:true,fontSize:14}))},{id:seed.conv.id})
   await page.goto('/')
   const work=page.locator('[data-conversation-work]')
   await expect(work).toContainText('执行失败')
   if(scenario==='zero-failure')await expect(work).toContainText('暂无阶段结论记录')
   else await expect(work).toContainText('上线六月')
-  await expect(page.getByText('后台整理状态已回投工作卡；请按卡片中的结果继续。').last()).toBeVisible()
+  await expect(page.locator('.work-choice')).toHaveCount(0)
   await expect(page.locator('.hitl-choice')).toHaveCount(0)
+  await expectOnlySendRun(request,seed)
  })
 }
 test('completed sourced work offers only its grounded suggestions',async({page,request})=>{
  const seed=await setup(request,'choice')
  await open(page,seed)
- await expect(page.locator('.hitl-choice')).toHaveCount(3)
- for(const label of ['会议摘要','验收清单','联调计划'])await expect(page.locator('.hitl-choice').filter({hasText:label})).toHaveCount(1)
- await expect(page.locator('.hitl-card')).toContainText('基于整理结果，接下来生成什么？')
+ const work=page.locator('[data-conversation-work]')
+ await expect(work.locator('.work-choice')).toHaveCount(3)
+ for(const label of ['会议摘要','验收清单','联调计划'])await expect(work.locator('.work-choice').filter({hasText:label})).toHaveCount(1)
+ await expect(work.locator('.work-decision legend')).toContainText('接下来需要做什么？')
  await page.reload()
- await expect(page.locator('.hitl-choice')).toHaveCount(3)
+ await expect(work.locator('.work-choice')).toHaveCount(3)
 })
 test('failed material registration reports no started task or unsupported finding',async({page,request})=>{
  const seed=await setup(request,'normal',false,{startFailure:true})