agent-insight 0.2.0-beta → 0.2.2-beta

package/.next/standalone/.next/server/chunks/ssr/src_app_(main)_eval_trajectory_page_tsx_052751f5._.js

@@ -1,3 +1,3 @@
-module.exports=[783210,a=>{"use strict";var b=a.i(187924),c=a.i(572131),d=a.i(50944),e=a.i(729936),f=a.i(575110),g=a.i(943408),h=a.i(767961),i=a.i(97289),j=a.i(434465);function k(a){return String(a).padStart(2,"0")}function l(a=new Date){return`评测执行 ${a.getFullYear()}-${k(a.getMonth()+1)}-${k(a.getDate())} ${k(a.getHours())}:${k(a.getMinutes())}:${k(a.getSeconds())}`}function m(a,b=new Date){let c=String(a.title||"").trim(),d=String(a.description||"").trim();return{title:c||l(b),description:d}}var n=a.i(882524);let o=a.i(478614).presetEvaluators.map(a=>({id:a.id,name:a.name})),p=new Set(o.map(a=>a.id)),q="#4F46E5",r="#EEF2FF",s="#C7D2FE",t="#16A34A",u="#DC2626",v="#D97706",w="#E4E4E7",x="#F4F4F5",y="#FAFAFA",z="#18181B",A="#52525B",B="#71717A",C="#A1A1AA";function D(a,b){let c=Array.isArray(a?.selectedEvaluators)?a.selectedEvaluators:[];return 0===c.length?"preset-agent-trace-quality"===b:c.includes(b)}function E(a){return!0===a.auto_eval_ready||!0===a.autoEvalReady}function F(){var a,e,k;let{user:F}=(0,g.useAuth)(),N=(0,d.useRouter)(),O=(0,d.useSearchParams)(),[P,Q]=(0,c.useState)([]),[R,S]=(0,c.useState)(()=>{let a=O?.get("evaluatorId");return a&&p.has(a)?[a]:[]}),[T,U]=(0,c.useState)(""),[V,W]=(0,c.useState)(""),[X,Y]=(0,c.useState)(""),[Z,$]=(0,c.useState)([]),[_,aa]=(0,c.useState)([]),[ab,ac]=(0,c.useState)(0),[ad,ae]=(0,c.useState)(1),[af,ag]=(0,c.useState)(!1),[ah,ai]=(0,c.useState)(!1),[aj,ak]=(0,c.useState)([]),[al,am]=(0,c.useState)(new Set),[an,ao]=(0,c.useState)(!0),[ap,aq]=(0,c.useState)(""),[ar,as]=(0,c.useState)(""),[at,au]=(0,c.useState)(!1),[av,aw]=(0,c.useState)(!1),[ax,ay]=(0,c.useState)(""),[az,aA]=(0,c.useState)([]),[aB,aC]=(0,c.useState)(""),[aD,aE]=(0,c.useState)([]),[aF,aG]=(0,c.useState)(""),[aH,aI]=(0,c.useState)(""),[aJ,aK]=(0,c.useState)(""),aL=(0,c.useMemo)(()=>`trajectory-eval-task-draft:${F||"anonymous"}`,[F]),aM=(0,c.useRef)(new Set),aN=(0,c.useRef)(!1);(0,c.useEffect)(()=>{F&&(0,f.apiFetch)(`/api/eval/settings?user=${encodeURIComponent(F)}`).then(a=>a.json()).then(a=>{let b=Array.isArray(a?.configs)?a.configs.map(a=>({id:a.id,name:a.name||a.id,model:a.model||"",baseUrl:a.baseUrl})):[];aA(b);let c=a?.activeConfigId;c&&b.find(a=>a.id===c)?aC(c):b.length>0&&aC(b[0].id)}).catch(()=>void 0)},[F]);let aO=async()=>{try{let a=await fetch("/api/agents");if(a.ok){let b=(await a.json()).agents.map(a=>({id:a.id,name:a.name,ownership:"user",layer:"main",platform:a.platform,version:"v1.0",framework:"Custom",status:"idle",todayCalls:a.todayCalls||"0",lastExecutedAt:a.lastExecutedAt||a.createdAt}));aE(b)}}catch(a){console.error("Failed to fetch DB agents",a)}};(0,c.useEffect)(()=>{aO()},[]),(0,c.useEffect)(()=>{},[aL]);let aP=(0,c.useMemo)(()=>[{id:"demo-agent",name:"demo-agent",ownership:"system",layer:"main",platform:"opencode",version:"v1.0",framework:"opencode",status:"running",successRate:"—",todayCalls:"—",lastExecutedAt:"2026-05-06T12:00:00"},{id:"customer-service-agent",name:"customer-service-agent",ownership:"system",layer:"main",platform:"opencode",version:"v2.1",framework:"LangChain",status:"running",successRate:"87.3%",todayCalls:"3,847",lastExecutedAt:"2026-05-06T11:48:00"},{id:"data-analyzer-v2",name:"data-analyzer-v2",ownership:"user",layer:"main",platform:"openclaw",version:"v1.0",framework:"Custom",status:"idle",successRate:"94.1%",todayCalls:"128",lastExecutedAt:"2026-05-05T20:10:00"},{id:"order-executor",name:"order-executor",ownership:"system",layer:"subagent",platform:"opencode",version:"v1.1",framework:"LangGraph",status:"running",todayCalls:"412",p99:"4.3s",parentAgent:"customer-service-agent",lastExecutedAt:"2026-05-06T11:43:00"},{id:"email-dispatcher",name:"email-dispatcher",ownership:"user",layer:"main",platform:"hermes",version:"v0.9",framework:"AutoGPT",status:"running",successRate:"99.2%",todayCalls:"1,024",lastExecutedAt:"2026-05-06T09:24:00"},{id:"security-guard",name:"security-guard",ownership:"system",layer:"main",platform:"opencode",version:"v3.0",framework:"Internal",status:"running",successRate:"100%",todayCalls:"45,201",lastExecutedAt:"2026-05-06T11:56:00"},{id:"trace-quality-evaluator",name:"trace-quality-evaluator",ownership:"system",layer:"main",platform:"opencode",version:"v1.0",framework:"opencode",status:"running",successRate:"—",todayCalls:"—",lastExecutedAt:"2026-05-06T12:00:00"},{id:"task-completion-evaluator",name:"task-completion-evaluator",ownership:"system",layer:"main",platform:"opencode",version:"v1.0",framework:"opencode",status:"running",successRate:"—",todayCalls:"—",lastExecutedAt:"2026-05-06T12:00:00"}],[]),aQ=(0,c.useMemo)(()=>{let a=new Set(aD.map(a=>`${a.platform}-${a.name}`));return[...aD,...aP.filter(b=>!a.has(`${b.platform}-${b.name}`))]},[aD,aP]);(0,c.useEffect)(()=>{F&&(0,f.apiFetch)(`/api/agent-datasets?user=${encodeURIComponent(F)}`).then(a=>a.json()).then(a=>{Q((Array.isArray(a)?a:[]).filter(a=>"trajectory"===a.datasetKind))}).catch(a=>aq(`加载数据集失败：${a?.message||a}`))},[F]),(0,c.useEffect)(()=>{F&&(0,f.apiFetch)(`/api/observe/data?user=${encodeURIComponent(F)}&summary=agents`).then(a=>a.json()).then(a=>{$(Array.isArray(a?.agents)?a.agents:[])}).catch(a=>aq(`加载执行 Agent 失败：${a?.message||a}`)).finally(()=>ao(!1))},[F]),(0,c.useEffect)(()=>{if(!F)return;let a=!1,b=async()=>{try{let b=await (0,f.apiFetch)(`/api/eval/trajectory/results?user=${encodeURIComponent(F)}&limit=200`),c=await b.json();a||ak(Array.isArray(c?.results)?c.results:[])}catch{}};b();let c=setInterval(b,3e3);return()=>{a=!0,clearInterval(c)}},[F]);let aR=(0,c.useCallback)(async()=>{if(!F)return;let a=await (0,f.apiFetch)(`/api/eval/trajectory/results?user=${encodeURIComponent(F)}&limit=200`),b=await a.json();ak(Array.isArray(b?.results)?b.results:[])},[F]),aS=(0,c.useCallback)(async(a,b)=>{if(!F||!a){aa([]),ac(0);return}ag(!0);try{let c=await (0,f.apiFetch)(`/api/observe/data?user=${encodeURIComponent(F)}&paginated=1&includeEvaluations=0&agentName=${encodeURIComponent(a)}&page=${b}&pageSize=20`),d=await c.json();aa(Array.isArray(d?.records)?d.records:[]),ac("number"==typeof d?.total?d.total:0)}catch(a){aq(`加载执行记录失败：${a?.message||a}`),aa([]),ac(0)}finally{ag(!1)}},[F]),aT=(0,c.useMemo)(()=>{let a=aQ.filter(a=>"main"===a.layer).filter(a=>!(0,j.isEvaluatorAgent)(a)).map(a=>a.name),b=Z.filter(a=>!!a),c=new Set;return[...a,...b].filter(a=>!(!a||c.has(a))&&(c.add(a),!0)).sort()},[aQ,Z]);(0,c.useEffect)(()=>{!T&&aT.length>0&&U(aT[0])},[aT,T]),(0,c.useEffect)(()=>{!X&&P.length>0&&Y(P[0].id)},[P,X]);let aU=(0,c.useMemo)(()=>{let a=new Map;for(let b of aj){if(b.watchPlaceholder)continue;let c=b.taskId||b.executionId||"";if(!c)continue;let d=a.get(c);(!d||new Date(b.createdAt).getTime()>new Date(d.createdAt).getTime())&&a.set(c,b)}return a},[aj]);async function aV(){if(F&&T){ai(!0),aq("");try{let a=await (0,f.apiFetch)(`/api/observe/data?user=${encodeURIComponent(F)}&summary=traceIds&agentName=${encodeURIComponent(T)}`),b=await a.json(),c=Array.isArray(b?.traceIds)?b.traceIds.filter(a=>"string"==typeof a&&a.trim().length>0):[];am(new Set(c)),as(c.length>0?`已全选 ${c.length} 条 trace`:"当前 Agent 没有可选 trace")}catch(a){aq(`全选失败：${a?.message||a}`)}finally{ai(!1)}}}async function aW(){if(!F)return void aq("请先登录");if(0===al.size)return void aq("请至少勾选一条执行记录");if(!T)return void aq("请选择执行 Agent");if(0===R.length)return void aq("请至少选择一个评估器");aq(""),as("");let a=m({title:aF,description:aH}),b=Array.from(al);if(0===b.length)return void aq("没有可评测的执行记录");au(!0);try{let c=await (0,f.apiFetch)("/api/eval/trajectory/run",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({user:F,taskIds:b,evaluators:R,taskTitle:a.title,taskDescription:a.description})}),d=await c.json();if(c.ok){let b=Array.isArray(d?.evaluatorNames)?d.evaluatorNames.join("、"):R.join("、");as(`已发起 ${d.created?.length??0} 条评测：${a.title}（${b}）`),am(new Set);let c={title:l(),description:""},e=JSON.stringify(c);aG(c.title),aI(c.description),aK(e),(0,f.apiFetch)(`/api/eval/trajectory/results?user=${encodeURIComponent(F)}&limit=200`).then(a=>a.json()).then(a=>ak(Array.isArray(a?.results)?a.results:[]))}else aq(`提交失败：${d?.error||c.statusText}`)}catch(a){aq(`提交失败：${a?.message||a}`)}finally{au(!1)}}async function aX(){if(!F)return void aq("请先登录");if(!T)return void aq("请选择执行 Agent");if(0===R.length)return void aq("请至少选择一个评估器");if(al.size>0)return void aq("自动观测只追踪未来新来的 trace。请先取消当前勾选的 trace。");aq(""),as(""),aw(!0);let a=m({title:aF,description:aH});try{let b=await (0,f.apiFetch)("/api/eval/trajectory/run",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({user:F,taskIds:[],evaluators:R,taskTitle:a.title,taskDescription:a.description,autoWatch:!0,watchedAgent:T})}),c=await b.json();if(!b.ok)return void aq(`自动观测启动失败：${c?.error||b.statusText}`);aM.current=new Set(_.map(a=>a.task_id||"").filter(Boolean)),ay(c.evaluatorRunId||""),am(new Set),as(`已开启自动观测：后续 ${T} 的新 trace 会自动追加到「${a.title}」`),await aR()}catch(a){aq(`自动观测启动失败：${a instanceof Error?a.message:String(a)}`)}finally{aw(!1)}}function aY(a){a&&N.push(`/eval/trajectory/${encodeURIComponent(a)}`)}if((0,c.useEffect)(()=>{am(new Set),ae(1)},[T]),(0,c.useEffect)(()=>{if(!T){aa([]),ac(0);return}aS(T,ad)},[ad,aS,T]),(0,c.useEffect)(()=>{if(!F||!ax||aN.current)return;let a=new Set(aj.filter(a=>a.evaluatorRunId===ax&&!a.watchPlaceholder).map(a=>a.taskId||a.executionId||"").filter(Boolean)),b=_.filter(E).map(a=>a.task_id||"").filter(Boolean).filter(a=>!aM.current.has(a)).filter(b=>!a.has(b));0!==b.length&&(aN.current=!0,(0,f.apiFetch)("/api/eval/trajectory/run",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({user:F,evaluatorRunId:ax,taskIds:b,evaluators:R,autoWatch:!0,watchedAgent:T})}).then(async a=>{let c=await a.json();a.ok?(as(`自动观测已追加 ${c.created?.length??b.length} 条新 trace`),await aR()):aq(`自动追加失败：${c?.error||a.statusText}`)}).catch(a=>aq(`自动追加失败：${a instanceof Error?a.message:String(a)}`)).finally(()=>{aN.current=!1}))},[ax,_,aR,aj,T,R,F]),an)return(0,b.jsx)("div",{className:"loading",style:{padding:24},children:"正在加载评测视图..."});let aZ=JSON.stringify(m({title:aF,description:aH}))!==aJ,a$=az.find(a=>a.id===aB);return(0,b.jsxs)("div",{style:{padding:"18px 22px 28px",maxWidth:1480,margin:"0 auto",color:z},children:[(0,b.jsxs)("div",{style:{display:"flex",alignItems:"center",gap:10,marginBottom:4},children:[(0,b.jsx)("span",{style:(a=r,{display:"inline-block",padding:"2px 8px",background:a,color:q,borderRadius:4,fontSize:11,fontWeight:500,whiteSpace:"nowrap"}),children:"评测执行"}),(0,b.jsx)("span",{style:{fontSize:11,color:C},children:"选执行 Agent 与评估器 → 多选 trace → 一键发起评测；可同时勾选结果评测与过程评测"})]}),(0,b.jsxs)("div",{style:{display:"flex",gap:12,alignItems:"flex-start",marginTop:10,padding:14,borderRadius:12,border:`1px solid ${w}`,background:"#fff",boxShadow:"0 1px 2px rgba(24, 24, 27, 0.04)"},children:[(0,b.jsxs)("div",{style:{minWidth:0,flex:1,display:"grid",gap:10},children:[(0,b.jsx)("input",{value:aF,onChange:a=>aG(a.target.value),placeholder:l(),style:{width:"100%",minHeight:42,border:`1px solid ${s}`,borderRadius:9,outline:"none",background:"#fff",padding:"0 14px",margin:0,fontSize:20,fontWeight:700,lineHeight:1.3,color:z,boxShadow:"0 0 0 3px rgba(79, 70, 229, 0.06)"}}),(0,b.jsx)("input",{value:aH,onChange:a=>aI(a.target.value),placeholder:"任务描述（可选）…",style:{width:"100%",minHeight:46,border:`1px solid ${w}`,borderRadius:9,outline:"none",background:"#fff",padding:"0 14px",marginTop:0,fontSize:14,lineHeight:1.65,color:A}})]}),(0,b.jsx)("button",{type:"button",onClick:function(){},style:{...I(aZ?q:C,aZ?r:y),marginTop:2,height:34,padding:"0 14px",fontSize:12},children:"保存"})]}),ap&&(0,b.jsx)("div",{style:K(u,"#FFEBEB","#FFD4D4"),children:ap}),ar&&(0,b.jsx)("div",{style:K(t,"#F0FDF4","#C6E5D9"),children:ar}),(0,b.jsxs)("div",{style:{marginTop:12,padding:"8px 14px",background:az.length>0?y:"#FFF7E6",border:`1px solid ${az.length>0?w:"#FFD591"}`,borderRadius:7,display:"flex",alignItems:"center",gap:8,fontSize:12},children:[(0,b.jsx)("span",{style:{color:B},children:"评测器运行模型："}),az.length>0?(0,b.jsxs)(b.Fragment,{children:[(0,b.jsx)("span",{style:{color:z,fontSize:12,fontWeight:500,lineHeight:"20px"},children:a$?`${a$.name} \xb7 ${a$.model||"未填写 model"}`:"未找到当前默认模型"}),(0,b.jsx)("span",{style:{color:C,fontSize:11},children:"（所有评估器共用）"})]}):(0,b.jsx)("span",{style:{color:v,fontWeight:500},children:"未配置 — 评测会失败"}),(0,b.jsx)("span",{style:{flex:1}}),(0,b.jsx)("button",{type:"button",onClick:()=>N.push("/modelconfig/registry"),style:{padding:"3px 10px",background:"#fff",color:A,border:`1px solid ${w}`,borderRadius:5,fontSize:11,cursor:"pointer"},children:az.length>0?"管理模型":"去配置 →"})]}),(0,b.jsxs)("div",{style:{background:"#fff",border:`1px solid ${w}`,borderRadius:9,padding:"14px 16px",marginTop:14,display:"flex",gap:18,alignItems:"center",flexWrap:"wrap"},children:[(0,b.jsx)(G,{label:"执行 Agent",children:(0,b.jsxs)("select",{value:T,onChange:a=>U(a.target.value),style:J(),children:[0===aT.length&&(0,b.jsx)("option",{value:"",children:"(暂无 Agent)"}),aT.map(a=>(0,b.jsx)("option",{value:a,children:a},a))]})}),(0,b.jsx)(G,{label:"评估器",children:(0,b.jsxs)("div",{style:{minWidth:360,maxWidth:560,border:`1px solid ${w}`,borderRadius:10,background:y,padding:8,display:"flex",flexDirection:"column",gap:8},children:[(0,b.jsx)("div",{style:{display:"flex",alignItems:"center",gap:8},children:(0,b.jsxs)("select",{value:V,onChange:a=>{var b;(b=a.target.value)&&p.has(b)&&(S(a=>a.includes(b)?a:[...a,b]),W(""))},style:{...J(),minWidth:0,width:"100%",height:32,background:"#fff",borderColor:s,fontSize:12,boxShadow:"0 1px 2px rgba(26, 26, 24, 0.04)"},children:[(0,b.jsx)("option",{value:"",children:"添加评估器..."}),o.map(a=>(0,b.jsx)("option",{value:a.id,disabled:R.includes(a.id),children:a.name},a.id))]})}),(0,b.jsx)("div",{style:{display:"flex",alignItems:"center",gap:8,flexWrap:"wrap",minHeight:40,padding:0===R.length?"0 4px":"2px 0"},children:0===R.length?(0,b.jsx)("span",{style:{fontSize:11,color:C},children:"已选评估器会显示在这里"}):R.map(a=>{let c=o.find(b=>b.id===a);return c?(0,b.jsxs)("span",{style:{display:"inline-flex",alignItems:"center",gap:8,padding:"5px 8px 5px 10px",border:`1px solid ${s}`,borderRadius:6,background:"#fff",color:A,fontSize:12,lineHeight:1,boxShadow:"0 1px 2px rgba(26, 26, 24, 0.03)"},children:[(0,b.jsx)("span",{children:c.name}),(0,b.jsx)("button",{type:"button",onClick:()=>{var a;return a=c.id,void S(b=>b.filter(b=>b!==a))},style:{width:16,height:16,borderRadius:4,border:`1px solid ${s}`,background:r,color:q,cursor:"pointer",fontSize:10,lineHeight:"14px",padding:0,flex:"0 0 auto"},"aria-label":`删除${c.name}`,title:`删除 ${c.name}`,children:"×"})]},c.id):null})})]})}),(0,b.jsx)("div",{style:{flex:1}}),(0,b.jsx)("button",{onClick:aW,disabled:at||av||0===al.size||!T||0===R.length,style:{padding:"6px 14px",background:(e=at||av||0===al.size||!T||0===R.length)?"#bfb6f0":q,color:"#fff",border:"none",borderRadius:6,fontSize:12,fontWeight:500,cursor:e?"not-allowed":"pointer"},children:at?"提交中…":`开始评测 (${al.size})`}),(0,b.jsx)("button",{onClick:aX,disabled:at||av||al.size>0||!T||0===R.length,style:H(at||av||al.size>0||!T||0===R.length),title:"开启后，本页面会自动追踪该执行 Agent 新上报的 trace，并追加到本次评测任务",children:av?"启动中…":"自动观测"})]}),(0,b.jsxs)("div",{style:{marginTop:16},children:[(0,b.jsxs)("div",{style:{display:"flex",alignItems:"baseline",gap:8,marginBottom:8},children:[(0,b.jsx)("div",{style:{fontSize:13,fontWeight:600},children:"执行记录清单"}),(0,b.jsx)("div",{style:{fontSize:11,color:B},children:T?(0,b.jsxs)(b.Fragment,{children:["当前 Agent ",(0,b.jsx)("code",{style:{color:q},children:T})," 共 ",ab," 条 trace，20 条/页"]}):"请先选 Agent"}),(0,b.jsxs)("div",{style:{marginLeft:"auto",display:"flex",alignItems:"center",gap:8},children:[(0,b.jsx)("button",{type:"button",onClick:aV,disabled:!T||ah||0===ab,style:H(!T||ah||0===ab),children:ah?"全选中…":`全选全部 (${ab})`}),(0,b.jsx)("button",{type:"button",onClick:function(){am(new Set)},disabled:0===al.size,style:{padding:"6px 14px",background:"transparent",color:(k=0===al.size)?C:A,border:`1px solid ${w}`,borderRadius:6,fontSize:12,fontWeight:500,cursor:k?"not-allowed":"pointer"},children:"清空选择"})]})]}),(0,b.jsx)("div",{style:{border:`1px solid ${w}`,borderRadius:9,overflow:"hidden",background:"#fff"},children:af?(0,b.jsx)("div",{style:{padding:24,textAlign:"center",color:B,fontSize:12},children:"正在加载执行记录..."}):0===_.length?(0,b.jsx)("div",{style:{padding:24,textAlign:"center",color:B,fontSize:12},children:"该 Agent 没有可评测的执行记录。"}):(0,b.jsxs)(b.Fragment,{children:[(0,b.jsxs)("table",{style:{width:"100%",borderCollapse:"collapse",fontSize:12},children:[(0,b.jsx)("thead",{children:(0,b.jsxs)("tr",{style:{background:y,borderBottom:`1px solid ${w}`},children:[(0,b.jsx)("th",{style:L(36),children:(0,b.jsx)("input",{type:"checkbox",checked:_.length>0&&_.every(a=>a.task_id&&al.has(a.task_id)),onChange:function(){let a=_.map(a=>a.task_id||"").filter(Boolean),b=a.length>0&&a.every(a=>al.has(a));am(c=>{let d=new Set(c);return b?a.forEach(a=>d.delete(a)):a.forEach(a=>d.add(a)),d})}})}),(0,b.jsx)("th",{style:L(170,"left"),children:"TRACE ID"}),(0,b.jsx)("th",{style:L(void 0,"left"),children:"Trace 实际输入"}),(0,b.jsx)("th",{style:L(void 0,"left"),children:"Trace 实际输出"}),(0,b.jsx)("th",{style:L(80,"center"),children:"评测状态"}),(0,b.jsx)("th",{style:L(76,"right"),children:"得分"}),(0,b.jsx)("th",{style:L(80,"left"),children:"评测器"}),(0,b.jsx)("th",{style:L(60,"right"),children:"耗时"}),(0,b.jsx)("th",{style:L(80,"center")})]})}),(0,b.jsx)("tbody",{children:_.map(a=>{var c,d;let e,f,g,j,k,l=a.task_id||"",m=aU.get(l),n=al.has(l),o=m&&("done"===(c=m.status)||"failed"===c)?(e=D(m,"preset-agent-trace-quality"),f=D(m,"preset-agent-task-completion"),0===(g=[e?m.trajectoryScore:null,f&&"number"==typeof m.resultEvaluationScore?m.resultEvaluationScore:null].filter(a=>"number"==typeof a&&Number.isFinite(a))).length?null:g.reduce((a,b)=>a+b,0)/g.length):null;return(0,b.jsxs)("tr",{style:{borderBottom:`1px solid ${x}`,background:n?"#F8F7FE":"transparent",cursor:"pointer"},onClick:()=>aY(l),children:[(0,b.jsx)("td",{style:M("center"),onClick:a=>a.stopPropagation(),children:(0,b.jsx)("input",{type:"checkbox",checked:n,onChange:()=>{am(a=>{let b=new Set(a);return b.has(l)?b.delete(l):b.add(l),b})}})}),(0,b.jsx)("td",{style:{...M("left"),fontFamily:"monospace",color:A,whiteSpace:"nowrap"},title:l,children:l&&l.slice(0,14)||"--"}),(0,b.jsx)("td",{style:{...M("left"),maxWidth:280},children:(0,b.jsx)("div",{style:{overflow:"hidden",textOverflow:"ellipsis",whiteSpace:"nowrap"},title:a.query||"",children:a.query||"—"})}),(0,b.jsxs)("td",{style:{...M("left"),maxWidth:280},children:[(0,b.jsx)("div",{style:{overflow:"hidden",textOverflow:"ellipsis",whiteSpace:"nowrap"},title:a.final_result||"",children:a.final_result||"—"}),(0,b.jsxs)("div",{style:{fontSize:10,color:C,marginTop:2},children:[a.framework||""," · ",a.model||""," · ",function(a){if(!a)return"--";try{let b=new Date(a),c=Date.now()-b.getTime();if(c<6e4)return"刚刚";if(c<36e5)return`${Math.floor(c/6e4)} 分钟前`;if(c<864e5)return`${Math.floor(c/36e5)} 小时前`;return`${Math.floor(c/864e5)} 天前`}catch{return a}}(a.timestamp)]})]}),(0,b.jsx)("td",{style:M("center"),children:m?(0,b.jsx)("span",{style:{color:"warning"===(j=(0,i.getTrajectoryStatusTone)(m))?v:"danger"===j?u:"running"===j?"#1677ff":"success"===j?t:C,fontWeight:600,fontSize:11},title:(0,i.getTrajectoryStatusTitle)(m),children:(0,i.getTrajectoryStatusLabel)(m)}):(0,b.jsx)("span",{style:{color:C},children:"待评测"})}),(0,b.jsx)("td",{style:{...M("right"),color:null!=o?q:C,fontWeight:600,whiteSpace:"nowrap"},children:null!=o?`${(0,h.fmtPercentScore)(o)} 分`:"--"}),(0,b.jsx)("td",{style:M("left"),children:(0,b.jsx)("span",{style:{color:m?A:C},children:m?(k=Array.isArray(m?.selectedEvaluatorNames)?m.selectedEvaluatorNames:[]).length>0?k.join("、"):"Agent 轨迹质量":"—"})}),(0,b.jsx)("td",{style:{...M("right"),color:B},children:null==(d=a.latency)||Number.isNaN(d)?"--":d>=60?`${(d/60).toFixed(1)}m`:`${d.toFixed(2)}s`}),(0,b.jsx)("td",{style:M("center"),onClick:a=>a.stopPropagation(),children:(0,b.jsx)("button",{onClick:()=>aY(l),style:I(),children:"详情"})})]},l)})})]}),(0,b.jsx)("div",{style:{padding:"10px 12px",borderTop:`1px solid ${x}`},children:(0,b.jsx)(n.Pagination,{page:ad,pageSize:20,total:ab,onPageChange:ae,summary:(a,b,c)=>`${a}-${b} / ${c}`})})]})})]})]})}function G({label:a,children:c}){return(0,b.jsxs)("div",{style:{display:"flex",alignItems:"center",gap:8},children:[(0,b.jsxs)("span",{style:{fontSize:11.5,color:B,fontWeight:500,whiteSpace:"nowrap"},children:[a,"："]}),c]})}function H(a){return{padding:"6px 14px",background:a?y:"#fff",color:a?C:q,border:`1px solid ${a?w:q}`,borderRadius:6,fontSize:12,fontWeight:500,cursor:a?"not-allowed":"pointer"}}function I(a=A,b="#fff"){return{padding:"3px 9px",background:b,color:a,border:`1px solid ${w}`,borderRadius:5,fontSize:11,cursor:"pointer"}}function J(){return{padding:"4px 12px",height:30,borderRadius:7,border:`1px solid ${w}`,fontSize:12,background:"#fff",color:z,fontWeight:500}}function K(a,b,c){return{padding:10,marginTop:12,borderRadius:6,background:b,border:`1px solid ${c}`,color:a,fontSize:12}}function L(a,b="left"){return{padding:"8px 10px",textAlign:b,fontWeight:500,color:B,fontSize:11,...a?{width:a}:{}}}function M(a="left"){return{padding:"8px 10px",textAlign:a,verticalAlign:"middle"}}function N(){let a=(0,d.useRouter)();return(0,b.jsxs)(b.Fragment,{children:[(0,b.jsx)(e.AppTopBar,{title:"发起新评测",showDefaultActions:!1}),(0,b.jsxs)("div",{style:{flex:1,overflowY:"auto"},children:[(0,b.jsx)("div",{style:{padding:"10px 16px 0"},children:(0,b.jsx)("button",{type:"button",onClick:()=>a.push("/eval"),style:{padding:"4px 10px",borderRadius:6,border:"1px solid var(--border)",background:"var(--background)",color:"var(--foreground)",fontSize:12,cursor:"pointer"},children:"< 返回评测执行"})}),(0,b.jsx)(c.Suspense,{fallback:(0,b.jsx)("div",{children:"Loading..."}),children:(0,b.jsx)(F,{})})]})]})}a.s(["default",()=>N],783210)}];
+module.exports=[783210,a=>{"use strict";var b=a.i(187924),c=a.i(572131),d=a.i(50944),e=a.i(729936),f=a.i(575110),g=a.i(943408),h=a.i(767961),i=a.i(97289),j=a.i(434465);function k(a){return String(a).padStart(2,"0")}function l(a=new Date){return`评测执行 ${a.getFullYear()}-${k(a.getMonth()+1)}-${k(a.getDate())} ${k(a.getHours())}:${k(a.getMinutes())}:${k(a.getSeconds())}`}function m(a,b=new Date){let c=String(a.title||"").trim(),d=String(a.description||"").trim(),e=String(a.scope||"").trim(),f=String(a.skillName||"").trim(),g=a.skillVersion,h=null==g||""===g?null:Number(g);return{title:c||l(b),description:d,...e?{scope:e}:{},...f?{skillName:f}:{},...null!=h&&Number.isFinite(h)?{skillVersion:h}:{}}}var n=a.i(882524);let o=a.i(478614).presetEvaluators.map(a=>({id:a.id,name:a.name})),p=new Set(o.map(a=>a.id)),q="#4F46E5",r="#EEF2FF",s="#C7D2FE",t="#16A34A",u="#DC2626",v="#D97706",w="#E4E4E7",x="#F4F4F5",y="#FAFAFA",z="#18181B",A="#52525B",B="#71717A",C="#A1A1AA";function D(a,b){let c=Array.isArray(a?.selectedEvaluators)?a.selectedEvaluators:[];return 0===c.length?"preset-agent-trace-quality"===b:c.includes(b)}function E(a){return!0===a.auto_eval_ready||!0===a.autoEvalReady}function F(){var a,e,k;let{user:F}=(0,g.useAuth)(),N=(0,d.useRouter)(),O=(0,d.useSearchParams)(),[P,Q]=(0,c.useState)([]),[R,S]=(0,c.useState)(()=>{let a=O?.get("evaluatorId");return a&&p.has(a)?[a]:[]}),[T,U]=(0,c.useState)(""),[V,W]=(0,c.useState)(""),[X,Y]=(0,c.useState)(""),[Z,$]=(0,c.useState)([]),[_,aa]=(0,c.useState)([]),[ab,ac]=(0,c.useState)(0),[ad,ae]=(0,c.useState)(1),[af,ag]=(0,c.useState)(!1),[ah,ai]=(0,c.useState)(!1),[aj,ak]=(0,c.useState)([]),[al,am]=(0,c.useState)(new Set),[an,ao]=(0,c.useState)(!0),[ap,aq]=(0,c.useState)(""),[ar,as]=(0,c.useState)(""),[at,au]=(0,c.useState)(!1),[av,aw]=(0,c.useState)(!1),[ax,ay]=(0,c.useState)(""),[az,aA]=(0,c.useState)([]),[aB,aC]=(0,c.useState)(""),[aD,aE]=(0,c.useState)([]),[aF,aG]=(0,c.useState)(""),[aH,aI]=(0,c.useState)(""),[aJ,aK]=(0,c.useState)(""),aL=(0,c.useMemo)(()=>`trajectory-eval-task-draft:${F||"anonymous"}`,[F]),aM=(0,c.useRef)(new Set),aN=(0,c.useRef)(!1);(0,c.useEffect)(()=>{F&&(0,f.apiFetch)(`/api/eval/settings?user=${encodeURIComponent(F)}`).then(a=>a.json()).then(a=>{let b=Array.isArray(a?.configs)?a.configs.map(a=>({id:a.id,name:a.name||a.id,model:a.model||"",baseUrl:a.baseUrl})):[];aA(b);let c=a?.activeConfigId;c&&b.find(a=>a.id===c)?aC(c):b.length>0&&aC(b[0].id)}).catch(()=>void 0)},[F]);let aO=async()=>{try{let a=await fetch("/api/agents");if(a.ok){let b=(await a.json()).agents.map(a=>({id:a.id,name:a.name,ownership:"user",layer:"main",platform:a.platform,version:"v1.0",framework:"Custom",status:"idle",todayCalls:a.todayCalls||"0",lastExecutedAt:a.lastExecutedAt||a.createdAt}));aE(b)}}catch(a){console.error("Failed to fetch DB agents",a)}};(0,c.useEffect)(()=>{aO()},[]),(0,c.useEffect)(()=>{},[aL]);let aP=(0,c.useMemo)(()=>[{id:"demo-agent",name:"demo-agent",ownership:"system",layer:"main",platform:"opencode",version:"v1.0",framework:"opencode",status:"running",successRate:"—",todayCalls:"—",lastExecutedAt:"2026-05-06T12:00:00"},{id:"customer-service-agent",name:"customer-service-agent",ownership:"system",layer:"main",platform:"opencode",version:"v2.1",framework:"LangChain",status:"running",successRate:"87.3%",todayCalls:"3,847",lastExecutedAt:"2026-05-06T11:48:00"},{id:"data-analyzer-v2",name:"data-analyzer-v2",ownership:"user",layer:"main",platform:"openclaw",version:"v1.0",framework:"Custom",status:"idle",successRate:"94.1%",todayCalls:"128",lastExecutedAt:"2026-05-05T20:10:00"},{id:"order-executor",name:"order-executor",ownership:"system",layer:"subagent",platform:"opencode",version:"v1.1",framework:"LangGraph",status:"running",todayCalls:"412",p99:"4.3s",parentAgent:"customer-service-agent",lastExecutedAt:"2026-05-06T11:43:00"},{id:"email-dispatcher",name:"email-dispatcher",ownership:"user",layer:"main",platform:"hermes",version:"v0.9",framework:"AutoGPT",status:"running",successRate:"99.2%",todayCalls:"1,024",lastExecutedAt:"2026-05-06T09:24:00"},{id:"security-guard",name:"security-guard",ownership:"system",layer:"main",platform:"opencode",version:"v3.0",framework:"Internal",status:"running",successRate:"100%",todayCalls:"45,201",lastExecutedAt:"2026-05-06T11:56:00"},{id:"trace-quality-evaluator",name:"trace-quality-evaluator",ownership:"system",layer:"main",platform:"opencode",version:"v1.0",framework:"opencode",status:"running",successRate:"—",todayCalls:"—",lastExecutedAt:"2026-05-06T12:00:00"},{id:"task-completion-evaluator",name:"task-completion-evaluator",ownership:"system",layer:"main",platform:"opencode",version:"v1.0",framework:"opencode",status:"running",successRate:"—",todayCalls:"—",lastExecutedAt:"2026-05-06T12:00:00"}],[]),aQ=(0,c.useMemo)(()=>{let a=new Set(aD.map(a=>`${a.platform}-${a.name}`));return[...aD,...aP.filter(b=>!a.has(`${b.platform}-${b.name}`))]},[aD,aP]);(0,c.useEffect)(()=>{F&&(0,f.apiFetch)(`/api/agent-datasets?user=${encodeURIComponent(F)}`).then(a=>a.json()).then(a=>{Q((Array.isArray(a)?a:[]).filter(a=>"trajectory"===a.datasetKind))}).catch(a=>aq(`加载数据集失败：${a?.message||a}`))},[F]),(0,c.useEffect)(()=>{F&&(0,f.apiFetch)(`/api/observe/data?user=${encodeURIComponent(F)}&summary=agents`).then(a=>a.json()).then(a=>{$(Array.isArray(a?.agents)?a.agents:[])}).catch(a=>aq(`加载执行 Agent 失败：${a?.message||a}`)).finally(()=>ao(!1))},[F]),(0,c.useEffect)(()=>{if(!F)return;let a=!1,b=async()=>{try{let b=await (0,f.apiFetch)(`/api/eval/trajectory/results?user=${encodeURIComponent(F)}&limit=200`),c=await b.json();a||ak(Array.isArray(c?.results)?c.results:[])}catch{}};b();let c=setInterval(b,3e3);return()=>{a=!0,clearInterval(c)}},[F]);let aR=(0,c.useCallback)(async()=>{if(!F)return;let a=await (0,f.apiFetch)(`/api/eval/trajectory/results?user=${encodeURIComponent(F)}&limit=200`),b=await a.json();ak(Array.isArray(b?.results)?b.results:[])},[F]),aS=(0,c.useCallback)(async(a,b)=>{if(!F||!a){aa([]),ac(0);return}ag(!0);try{let c=await (0,f.apiFetch)(`/api/observe/data?user=${encodeURIComponent(F)}&paginated=1&includeEvaluations=0&agentName=${encodeURIComponent(a)}&page=${b}&pageSize=20`),d=await c.json();aa(Array.isArray(d?.records)?d.records:[]),ac("number"==typeof d?.total?d.total:0)}catch(a){aq(`加载执行记录失败：${a?.message||a}`),aa([]),ac(0)}finally{ag(!1)}},[F]),aT=(0,c.useMemo)(()=>{let a=aQ.filter(a=>"main"===a.layer).filter(a=>!(0,j.isEvaluatorAgent)(a)).map(a=>a.name),b=Z.filter(a=>!!a),c=new Set;return[...a,...b].filter(a=>!(!a||c.has(a))&&(c.add(a),!0)).sort()},[aQ,Z]);(0,c.useEffect)(()=>{!T&&aT.length>0&&U(aT[0])},[aT,T]),(0,c.useEffect)(()=>{!X&&P.length>0&&Y(P[0].id)},[P,X]);let aU=(0,c.useMemo)(()=>{let a=new Map;for(let b of aj){if(b.watchPlaceholder)continue;let c=b.taskId||b.executionId||"";if(!c)continue;let d=a.get(c);(!d||new Date(b.createdAt).getTime()>new Date(d.createdAt).getTime())&&a.set(c,b)}return a},[aj]);async function aV(){if(F&&T){ai(!0),aq("");try{let a=await (0,f.apiFetch)(`/api/observe/data?user=${encodeURIComponent(F)}&summary=traceIds&agentName=${encodeURIComponent(T)}`),b=await a.json(),c=Array.isArray(b?.traceIds)?b.traceIds.filter(a=>"string"==typeof a&&a.trim().length>0):[];am(new Set(c)),as(c.length>0?`已全选 ${c.length} 条 trace`:"当前 Agent 没有可选 trace")}catch(a){aq(`全选失败：${a?.message||a}`)}finally{ai(!1)}}}async function aW(){if(!F)return void aq("请先登录");if(0===al.size)return void aq("请至少勾选一条执行记录");if(!T)return void aq("请选择执行 Agent");if(0===R.length)return void aq("请至少选择一个评估器");aq(""),as("");let a=m({title:aF,description:aH}),b=Array.from(al);if(0===b.length)return void aq("没有可评测的执行记录");au(!0);try{let c=await (0,f.apiFetch)("/api/eval/trajectory/run",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({user:F,taskIds:b,evaluators:R,taskTitle:a.title,taskDescription:a.description})}),d=await c.json();if(c.ok){let b=Array.isArray(d?.evaluatorNames)?d.evaluatorNames.join("、"):R.join("、");as(`已发起 ${d.created?.length??0} 条评测：${a.title}（${b}）`),am(new Set);let c={title:l(),description:""},e=JSON.stringify(c);aG(c.title),aI(c.description),aK(e),(0,f.apiFetch)(`/api/eval/trajectory/results?user=${encodeURIComponent(F)}&limit=200`).then(a=>a.json()).then(a=>ak(Array.isArray(a?.results)?a.results:[]))}else aq(`提交失败：${d?.error||c.statusText}`)}catch(a){aq(`提交失败：${a?.message||a}`)}finally{au(!1)}}async function aX(){if(!F)return void aq("请先登录");if(!T)return void aq("请选择执行 Agent");if(0===R.length)return void aq("请至少选择一个评估器");if(al.size>0)return void aq("自动观测只追踪未来新来的 trace。请先取消当前勾选的 trace。");aq(""),as(""),aw(!0);let a=m({title:aF,description:aH});try{let b=await (0,f.apiFetch)("/api/eval/trajectory/run",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({user:F,taskIds:[],evaluators:R,taskTitle:a.title,taskDescription:a.description,autoWatch:!0,watchedAgent:T})}),c=await b.json();if(!b.ok)return void aq(`自动观测启动失败：${c?.error||b.statusText}`);aM.current=new Set(_.map(a=>a.task_id||"").filter(Boolean)),ay(c.evaluatorRunId||""),am(new Set),as(`已开启自动观测：后续 ${T} 的新 trace 会自动追加到「${a.title}」`),await aR()}catch(a){aq(`自动观测启动失败：${a instanceof Error?a.message:String(a)}`)}finally{aw(!1)}}function aY(a){a&&N.push(`/eval/trajectory/${encodeURIComponent(a)}`)}if((0,c.useEffect)(()=>{am(new Set),ae(1)},[T]),(0,c.useEffect)(()=>{if(!T){aa([]),ac(0);return}aS(T,ad)},[ad,aS,T]),(0,c.useEffect)(()=>{if(!F||!ax||aN.current)return;let a=new Set(aj.filter(a=>a.evaluatorRunId===ax&&!a.watchPlaceholder).map(a=>a.taskId||a.executionId||"").filter(Boolean)),b=_.filter(E).map(a=>a.task_id||"").filter(Boolean).filter(a=>!aM.current.has(a)).filter(b=>!a.has(b));0!==b.length&&(aN.current=!0,(0,f.apiFetch)("/api/eval/trajectory/run",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({user:F,evaluatorRunId:ax,taskIds:b,evaluators:R,autoWatch:!0,watchedAgent:T})}).then(async a=>{let c=await a.json();a.ok?(as(`自动观测已追加 ${c.created?.length??b.length} 条新 trace`),await aR()):aq(`自动追加失败：${c?.error||a.statusText}`)}).catch(a=>aq(`自动追加失败：${a instanceof Error?a.message:String(a)}`)).finally(()=>{aN.current=!1}))},[ax,_,aR,aj,T,R,F]),an)return(0,b.jsx)("div",{className:"loading",style:{padding:24},children:"正在加载评测视图..."});let aZ=JSON.stringify(m({title:aF,description:aH}))!==aJ,a$=az.find(a=>a.id===aB);return(0,b.jsxs)("div",{style:{padding:"18px 22px 28px",maxWidth:1480,margin:"0 auto",color:z},children:[(0,b.jsxs)("div",{style:{display:"flex",alignItems:"center",gap:10,marginBottom:4},children:[(0,b.jsx)("span",{style:(a=r,{display:"inline-block",padding:"2px 8px",background:a,color:q,borderRadius:4,fontSize:11,fontWeight:500,whiteSpace:"nowrap"}),children:"评测执行"}),(0,b.jsx)("span",{style:{fontSize:11,color:C},children:"选执行 Agent 与评估器 → 多选 trace → 一键发起评测；可同时勾选结果评测与过程评测"})]}),(0,b.jsxs)("div",{style:{display:"flex",gap:12,alignItems:"flex-start",marginTop:10,padding:14,borderRadius:12,border:`1px solid ${w}`,background:"#fff",boxShadow:"0 1px 2px rgba(24, 24, 27, 0.04)"},children:[(0,b.jsxs)("div",{style:{minWidth:0,flex:1,display:"grid",gap:10},children:[(0,b.jsx)("input",{value:aF,onChange:a=>aG(a.target.value),placeholder:l(),style:{width:"100%",minHeight:42,border:`1px solid ${s}`,borderRadius:9,outline:"none",background:"#fff",padding:"0 14px",margin:0,fontSize:20,fontWeight:700,lineHeight:1.3,color:z,boxShadow:"0 0 0 3px rgba(79, 70, 229, 0.06)"}}),(0,b.jsx)("input",{value:aH,onChange:a=>aI(a.target.value),placeholder:"任务描述（可选）…",style:{width:"100%",minHeight:46,border:`1px solid ${w}`,borderRadius:9,outline:"none",background:"#fff",padding:"0 14px",marginTop:0,fontSize:14,lineHeight:1.65,color:A}})]}),(0,b.jsx)("button",{type:"button",onClick:function(){},style:{...I(aZ?q:C,aZ?r:y),marginTop:2,height:34,padding:"0 14px",fontSize:12},children:"保存"})]}),ap&&(0,b.jsx)("div",{style:K(u,"#FFEBEB","#FFD4D4"),children:ap}),ar&&(0,b.jsx)("div",{style:K(t,"#F0FDF4","#C6E5D9"),children:ar}),(0,b.jsxs)("div",{style:{marginTop:12,padding:"8px 14px",background:az.length>0?y:"#FFF7E6",border:`1px solid ${az.length>0?w:"#FFD591"}`,borderRadius:7,display:"flex",alignItems:"center",gap:8,fontSize:12},children:[(0,b.jsx)("span",{style:{color:B},children:"评测器运行模型："}),az.length>0?(0,b.jsxs)(b.Fragment,{children:[(0,b.jsx)("span",{style:{color:z,fontSize:12,fontWeight:500,lineHeight:"20px"},children:a$?`${a$.name} \xb7 ${a$.model||"未填写 model"}`:"未找到当前默认模型"}),(0,b.jsx)("span",{style:{color:C,fontSize:11},children:"（所有评估器共用）"})]}):(0,b.jsx)("span",{style:{color:v,fontWeight:500},children:"未配置 — 评测会失败"}),(0,b.jsx)("span",{style:{flex:1}}),(0,b.jsx)("button",{type:"button",onClick:()=>N.push("/modelconfig/registry"),style:{padding:"3px 10px",background:"#fff",color:A,border:`1px solid ${w}`,borderRadius:5,fontSize:11,cursor:"pointer"},children:az.length>0?"管理模型":"去配置 →"})]}),(0,b.jsxs)("div",{style:{background:"#fff",border:`1px solid ${w}`,borderRadius:9,padding:"14px 16px",marginTop:14,display:"flex",gap:18,alignItems:"center",flexWrap:"wrap"},children:[(0,b.jsx)(G,{label:"执行 Agent",children:(0,b.jsxs)("select",{value:T,onChange:a=>U(a.target.value),style:J(),children:[0===aT.length&&(0,b.jsx)("option",{value:"",children:"(暂无 Agent)"}),aT.map(a=>(0,b.jsx)("option",{value:a,children:a},a))]})}),(0,b.jsx)(G,{label:"评估器",children:(0,b.jsxs)("div",{style:{minWidth:360,maxWidth:560,border:`1px solid ${w}`,borderRadius:10,background:y,padding:8,display:"flex",flexDirection:"column",gap:8},children:[(0,b.jsx)("div",{style:{display:"flex",alignItems:"center",gap:8},children:(0,b.jsxs)("select",{value:V,onChange:a=>{var b;(b=a.target.value)&&p.has(b)&&(S(a=>a.includes(b)?a:[...a,b]),W(""))},style:{...J(),minWidth:0,width:"100%",height:32,background:"#fff",borderColor:s,fontSize:12,boxShadow:"0 1px 2px rgba(26, 26, 24, 0.04)"},children:[(0,b.jsx)("option",{value:"",children:"添加评估器..."}),o.map(a=>(0,b.jsx)("option",{value:a.id,disabled:R.includes(a.id),children:a.name},a.id))]})}),(0,b.jsx)("div",{style:{display:"flex",alignItems:"center",gap:8,flexWrap:"wrap",minHeight:40,padding:0===R.length?"0 4px":"2px 0"},children:0===R.length?(0,b.jsx)("span",{style:{fontSize:11,color:C},children:"已选评估器会显示在这里"}):R.map(a=>{let c=o.find(b=>b.id===a);return c?(0,b.jsxs)("span",{style:{display:"inline-flex",alignItems:"center",gap:8,padding:"5px 8px 5px 10px",border:`1px solid ${s}`,borderRadius:6,background:"#fff",color:A,fontSize:12,lineHeight:1,boxShadow:"0 1px 2px rgba(26, 26, 24, 0.03)"},children:[(0,b.jsx)("span",{children:c.name}),(0,b.jsx)("button",{type:"button",onClick:()=>{var a;return a=c.id,void S(b=>b.filter(b=>b!==a))},style:{width:16,height:16,borderRadius:4,border:`1px solid ${s}`,background:r,color:q,cursor:"pointer",fontSize:10,lineHeight:"14px",padding:0,flex:"0 0 auto"},"aria-label":`删除${c.name}`,title:`删除 ${c.name}`,children:"×"})]},c.id):null})})]})}),(0,b.jsx)("div",{style:{flex:1}}),(0,b.jsx)("button",{onClick:aW,disabled:at||av||0===al.size||!T||0===R.length,style:{padding:"6px 14px",background:(e=at||av||0===al.size||!T||0===R.length)?"#bfb6f0":q,color:"#fff",border:"none",borderRadius:6,fontSize:12,fontWeight:500,cursor:e?"not-allowed":"pointer"},children:at?"提交中…":`开始评测 (${al.size})`}),(0,b.jsx)("button",{onClick:aX,disabled:at||av||al.size>0||!T||0===R.length,style:H(at||av||al.size>0||!T||0===R.length),title:"开启后，本页面会自动追踪该执行 Agent 新上报的 trace，并追加到本次评测任务",children:av?"启动中…":"自动观测"})]}),(0,b.jsxs)("div",{style:{marginTop:16},children:[(0,b.jsxs)("div",{style:{display:"flex",alignItems:"baseline",gap:8,marginBottom:8},children:[(0,b.jsx)("div",{style:{fontSize:13,fontWeight:600},children:"执行记录清单"}),(0,b.jsx)("div",{style:{fontSize:11,color:B},children:T?(0,b.jsxs)(b.Fragment,{children:["当前 Agent ",(0,b.jsx)("code",{style:{color:q},children:T})," 共 ",ab," 条 trace，20 条/页"]}):"请先选 Agent"}),(0,b.jsxs)("div",{style:{marginLeft:"auto",display:"flex",alignItems:"center",gap:8},children:[(0,b.jsx)("button",{type:"button",onClick:aV,disabled:!T||ah||0===ab,style:H(!T||ah||0===ab),children:ah?"全选中…":`全选全部 (${ab})`}),(0,b.jsx)("button",{type:"button",onClick:function(){am(new Set)},disabled:0===al.size,style:{padding:"6px 14px",background:"transparent",color:(k=0===al.size)?C:A,border:`1px solid ${w}`,borderRadius:6,fontSize:12,fontWeight:500,cursor:k?"not-allowed":"pointer"},children:"清空选择"})]})]}),(0,b.jsx)("div",{style:{border:`1px solid ${w}`,borderRadius:9,overflow:"hidden",background:"#fff"},children:af?(0,b.jsx)("div",{style:{padding:24,textAlign:"center",color:B,fontSize:12},children:"正在加载执行记录..."}):0===_.length?(0,b.jsx)("div",{style:{padding:24,textAlign:"center",color:B,fontSize:12},children:"该 Agent 没有可评测的执行记录。"}):(0,b.jsxs)(b.Fragment,{children:[(0,b.jsxs)("table",{style:{width:"100%",borderCollapse:"collapse",fontSize:12},children:[(0,b.jsx)("thead",{children:(0,b.jsxs)("tr",{style:{background:y,borderBottom:`1px solid ${w}`},children:[(0,b.jsx)("th",{style:L(36),children:(0,b.jsx)("input",{type:"checkbox",checked:_.length>0&&_.every(a=>a.task_id&&al.has(a.task_id)),onChange:function(){let a=_.map(a=>a.task_id||"").filter(Boolean),b=a.length>0&&a.every(a=>al.has(a));am(c=>{let d=new Set(c);return b?a.forEach(a=>d.delete(a)):a.forEach(a=>d.add(a)),d})}})}),(0,b.jsx)("th",{style:L(170,"left"),children:"TRACE ID"}),(0,b.jsx)("th",{style:L(void 0,"left"),children:"Trace 实际输入"}),(0,b.jsx)("th",{style:L(void 0,"left"),children:"Trace 实际输出"}),(0,b.jsx)("th",{style:L(80,"center"),children:"评测状态"}),(0,b.jsx)("th",{style:L(76,"right"),children:"得分"}),(0,b.jsx)("th",{style:L(80,"left"),children:"评测器"}),(0,b.jsx)("th",{style:L(60,"right"),children:"耗时"}),(0,b.jsx)("th",{style:L(80,"center")})]})}),(0,b.jsx)("tbody",{children:_.map(a=>{var c,d;let e,f,g,j,k,l=a.task_id||"",m=aU.get(l),n=al.has(l),o=m&&("done"===(c=m.status)||"failed"===c)?(e=D(m,"preset-agent-trace-quality"),f=D(m,"preset-agent-task-completion"),0===(g=[e?m.trajectoryScore:null,f&&"number"==typeof m.resultEvaluationScore?m.resultEvaluationScore:null].filter(a=>"number"==typeof a&&Number.isFinite(a))).length?null:g.reduce((a,b)=>a+b,0)/g.length):null;return(0,b.jsxs)("tr",{style:{borderBottom:`1px solid ${x}`,background:n?"#F8F7FE":"transparent",cursor:"pointer"},onClick:()=>aY(l),children:[(0,b.jsx)("td",{style:M("center"),onClick:a=>a.stopPropagation(),children:(0,b.jsx)("input",{type:"checkbox",checked:n,onChange:()=>{am(a=>{let b=new Set(a);return b.has(l)?b.delete(l):b.add(l),b})}})}),(0,b.jsx)("td",{style:{...M("left"),fontFamily:"monospace",color:A,whiteSpace:"nowrap"},title:l,children:l&&l.slice(0,14)||"--"}),(0,b.jsx)("td",{style:{...M("left"),maxWidth:280},children:(0,b.jsx)("div",{style:{overflow:"hidden",textOverflow:"ellipsis",whiteSpace:"nowrap"},title:a.query||"",children:a.query||"—"})}),(0,b.jsxs)("td",{style:{...M("left"),maxWidth:280},children:[(0,b.jsx)("div",{style:{overflow:"hidden",textOverflow:"ellipsis",whiteSpace:"nowrap"},title:a.final_result||"",children:a.final_result||"—"}),(0,b.jsxs)("div",{style:{fontSize:10,color:C,marginTop:2},children:[a.framework||""," · ",a.model||""," · ",function(a){if(!a)return"--";try{let b=new Date(a),c=Date.now()-b.getTime();if(c<6e4)return"刚刚";if(c<36e5)return`${Math.floor(c/6e4)} 分钟前`;if(c<864e5)return`${Math.floor(c/36e5)} 小时前`;return`${Math.floor(c/864e5)} 天前`}catch{return a}}(a.timestamp)]})]}),(0,b.jsx)("td",{style:M("center"),children:m?(0,b.jsx)("span",{style:{color:"warning"===(j=(0,i.getTrajectoryStatusTone)(m))?v:"danger"===j?u:"running"===j?"#1677ff":"success"===j?t:C,fontWeight:600,fontSize:11},title:(0,i.getTrajectoryStatusTitle)(m),children:(0,i.getTrajectoryStatusLabel)(m)}):(0,b.jsx)("span",{style:{color:C},children:"待评测"})}),(0,b.jsx)("td",{style:{...M("right"),color:null!=o?q:C,fontWeight:600,whiteSpace:"nowrap"},children:null!=o?`${(0,h.fmtPercentScore)(o)} 分`:"--"}),(0,b.jsx)("td",{style:M("left"),children:(0,b.jsx)("span",{style:{color:m?A:C},children:m?(k=Array.isArray(m?.selectedEvaluatorNames)?m.selectedEvaluatorNames:[]).length>0?k.join("、"):"Agent 轨迹质量":"—"})}),(0,b.jsx)("td",{style:{...M("right"),color:B},children:null==(d=a.latency)||Number.isNaN(d)?"--":d>=60?`${(d/60).toFixed(1)}m`:`${d.toFixed(2)}s`}),(0,b.jsx)("td",{style:M("center"),onClick:a=>a.stopPropagation(),children:(0,b.jsx)("button",{onClick:()=>aY(l),style:I(),children:"详情"})})]},l)})})]}),(0,b.jsx)("div",{style:{padding:"10px 12px",borderTop:`1px solid ${x}`},children:(0,b.jsx)(n.Pagination,{page:ad,pageSize:20,total:ab,onPageChange:ae,summary:(a,b,c)=>`${a}-${b} / ${c}`})})]})})]})]})}function G({label:a,children:c}){return(0,b.jsxs)("div",{style:{display:"flex",alignItems:"center",gap:8},children:[(0,b.jsxs)("span",{style:{fontSize:11.5,color:B,fontWeight:500,whiteSpace:"nowrap"},children:[a,"："]}),c]})}function H(a){return{padding:"6px 14px",background:a?y:"#fff",color:a?C:q,border:`1px solid ${a?w:q}`,borderRadius:6,fontSize:12,fontWeight:500,cursor:a?"not-allowed":"pointer"}}function I(a=A,b="#fff"){return{padding:"3px 9px",background:b,color:a,border:`1px solid ${w}`,borderRadius:5,fontSize:11,cursor:"pointer"}}function J(){return{padding:"4px 12px",height:30,borderRadius:7,border:`1px solid ${w}`,fontSize:12,background:"#fff",color:z,fontWeight:500}}function K(a,b,c){return{padding:10,marginTop:12,borderRadius:6,background:b,border:`1px solid ${c}`,color:a,fontSize:12}}function L(a,b="left"){return{padding:"8px 10px",textAlign:b,fontWeight:500,color:B,fontSize:11,...a?{width:a}:{}}}function M(a="left"){return{padding:"8px 10px",textAlign:a,verticalAlign:"middle"}}function N(){let a=(0,d.useRouter)();return(0,b.jsxs)(b.Fragment,{children:[(0,b.jsx)(e.AppTopBar,{title:"发起新评测",showDefaultActions:!1}),(0,b.jsxs)("div",{style:{flex:1,overflowY:"auto"},children:[(0,b.jsx)("div",{style:{padding:"10px 16px 0"},children:(0,b.jsx)("button",{type:"button",onClick:()=>a.push("/eval"),style:{padding:"4px 10px",borderRadius:6,border:"1px solid var(--border)",background:"var(--background)",color:"var(--foreground)",fontSize:12,cursor:"pointer"},children:"< 返回评测执行"})}),(0,b.jsx)(c.Suspense,{fallback:(0,b.jsx)("div",{children:"Loading..."}),children:(0,b.jsx)(F,{})})]})]})}a.s(["default",()=>N],783210)}];
 
 //# sourceMappingURL=src_app_%28main%29_eval_trajectory_page_tsx_052751f5._.js.map