agent-insight 0.2.0-beta → 0.2.2-beta

package/.next/standalone/.next/server/chunks/_73cb0d03._.js

@@ -85,7 +85,7 @@ ${t}
 - 每条关键观点都必须输出显式 score，范围 0.0～1.0。
 - covered=true 的关键观点必须给 coverage_reason。
 - covered=false / partial / wrong 的关键观点必须给 missing_reason、trace_root_cause.failure_stage、trace_root_cause.failure_reason。
-`}function h(e){return!!(void 0!==e.score||void 0!==e.is_correct||Array.isArray(e.key_point_findings)||Array.isArray(e.result_issues))}async function w(e,t){if("none"===e.precomputedRootCauseSource)return{rootCauses:[],source:"none"};if("dataset-cache"===e.precomputedRootCauseSource)return{rootCauses:Array.isArray(e.precomputedRootCauses)?e.precomputedRootCauses:[],source:"dataset-cache"};if(!String(e.expectedOutput||"").trim())return{rootCauses:[],source:"none"};try{return{rootCauses:await (0,p.extractRootCausesFromExpected)(e.caseInput,e.expectedOutput,t),source:"live-extract"}}catch{return{rootCauses:[],source:"none"}}}function _(e,t="skill-aware"){var a;let r,n="no-skill"===t?{...e,key_point_findings:(0,m.stripSkillAttributionFromKeyPointFindings)(e.key_point_findings)}:e,s=(0,m.deriveTaskCompletionScoreFromFindings)(n.key_point_findings),i=s.score??0,o=i>=.8,l=String(e.reason||"").trim()||"任务完成度评测已完成，但未返回理由。",u=void 0===e.score?void 0:(a=e.score,r="number"==typeof a?a:"string"==typeof a?Number(a):NaN,Number.isFinite(r)?Math.max(0,Math.min(1,r)):0),{result_issues:c,resultIssues:d,...p}=n;return{isCorrect:o,score:i,reason:l,rawAnalysis:{...p,score:i,is_correct:o,llm_reported_score:u,score_computation:{method:"equal_weight_average",item_count:s.itemCount},skillAttributionMode:t,key_point_findings:s.findings}}}async function v(e,t,a,p){return(0,s.withBackgroundOpencodeSlot)(async()=>(0,n.runWithEphemeralOpencodeServer)({user:t||void 0,verbose:!1,isolateHome:!0},async a=>{var n;let s,p,m,{rootCauses:g,source:v}=await w(e,t),S=e.skillAttributionMode||"skill-aware",E=await (0,o.getActiveConfig)(t);if(!E)return{isCorrect:!1,score:0,reason:"请先在模型配置中激活一个评测模型，才能执行结果评测。",rawAnalysis:{root_cause_source:v}};let x=t?await (0,l.loadServerModelForUser)(t):null,k=x?.providerID||(0,l.normalizeProviderID)(E.provider||(0,l.inferProviderFromBaseUrl)(E.baseUrl)),b=x?.modelID||E.model||"deepseek-chat",R={text:(p=g.length>0?g.map((e,t)=>`${t+1}. ${e.content}`).join("\n"):"（未提取到关键观点，可仅按任务完成度评判）",n=e.skillContext,s=Array.isArray(n?.invokedSkills)?n.invokedSkills:[],m=0===s.length?"（无）":s.map((e,t)=>[`### Skill ${t+1}: ${e.name}${null!=e.version?` v${e.version}`:""}`,e.content?.trim()||"（无 SKILL.md 内容）"].join("\n")).join("\n\n"),`# Agent 任务完成度评测输入
+`}function h(e){return!!(void 0!==e.score||void 0!==e.is_correct||Array.isArray(e.key_point_findings)||Array.isArray(e.result_issues))}async function w(e,t){if("none"===e.precomputedRootCauseSource)return{rootCauses:[],source:"none"};if("dataset-cache"===e.precomputedRootCauseSource)return{rootCauses:Array.isArray(e.precomputedRootCauses)?e.precomputedRootCauses:[],source:"dataset-cache"};if(!String(e.expectedOutput||"").trim())return{rootCauses:[],source:"none"};try{return{rootCauses:await (0,p.extractRootCausesFromExpected)(e.caseInput,e.expectedOutput,t),source:"live-extract"}}catch{return{rootCauses:[],source:"none"}}}function _(e,t="skill-aware"){var a;let r,n="no-skill"===t?{...e,key_point_findings:(0,m.stripSkillAttributionFromKeyPointFindings)(e.key_point_findings)}:e,s=(0,m.deriveTaskCompletionScoreFromFindings)(n.key_point_findings),i=s.score??0,o=i>=.8,l=String(e.reason||"").trim()||"任务完成度评测已完成，但未返回理由。",u=void 0===e.score?void 0:(a=e.score,r="number"==typeof a?a:"string"==typeof a?Number(a):NaN,Number.isFinite(r)?Math.max(0,Math.min(1,r)):0),{result_issues:c,resultIssues:d,...p}=n;return{isCorrect:o,score:i,reason:l,rawAnalysis:{...p,score:i,is_correct:o,llm_reported_score:u,score_computation:{method:"equal_weight_average",item_count:s.itemCount},skillAttributionMode:t,key_point_findings:s.findings}}}async function v(e,t,a,p){return(0,s.withBackgroundOpencodeSlot)(async()=>(0,n.runWithEphemeralOpencodeServer)({user:t||void 0,verbose:!1,isolateHome:!0},async a=>{var n;let s,p,m,{rootCauses:g,source:v}=await w(e,t),S=e.skillAttributionMode||"skill-aware",E=await (0,o.getActiveConfig)(t);if(!E)return{isCorrect:!1,score:0,reason:"请先在模型配置中激活一个评测模型，才能执行结果评测。",rawAnalysis:{root_cause_source:v}};let k=t?await (0,l.loadServerModelForUser)(t):null,x=k?.providerID||(0,l.normalizeProviderID)(E.provider||(0,l.inferProviderFromBaseUrl)(E.baseUrl)),b=k?.modelID||E.model||"deepseek-chat",R={text:(p=g.length>0?g.map((e,t)=>`${t+1}. ${e.content}`).join("\n"):"（未提取到关键观点，可仅按任务完成度评判）",n=e.skillContext,s=Array.isArray(n?.invokedSkills)?n.invokedSkills:[],m=0===s.length?"（无）":s.map((e,t)=>[`### Skill ${t+1}: ${e.name}${null!=e.version?` v${e.version}`:""}`,e.content?.trim()||"（无 SKILL.md 内容）"].join("\n")).join("\n\n"),`# Agent 任务完成度评测输入
 
 ## Skill 归因模式
 ${"skill-aware"===e.skillAttributionMode?"skill-aware：本 trace 关联了 Skill，允许基于 Skill 上下文输出 Skill 归因和改进建议。":"no-skill：本 trace 未关联任何 Skill，禁止输出 Skill 归因和 Skill 改进建议。"}
@@ -108,7 +108,7 @@ ${e.traceSummaryText?.trim()||"（无可用轨迹）"}
 ## Skill 上下文
 ${m}
 
-请你自行逐条检查关键观点覆盖情况，并在不派发任何子代理的前提下完成任务完成度评测和关键观点根因分析。`),agent:"build",model:{providerID:k,modelID:b,apiKey:x?.apiKey||E.apiKey,baseURL:x?.baseURL||E.baseUrl},modelOptions:{temperature:0,top_p:1,seed:42},system:y(S),permission:(0,i.buildEvaluatorPermissions)()},I="",C=null,N="",$=null,O=null;try{O=new r.AgentInsight({baseURL:a,logLevel:"warn"});let n=await O.createSession({title:`${A}-${Date.now()}`,directory:"/tmp"}),s=String(n?.id||n?.ID||"");if(!s)throw Error("Failed to create opencode session for task completion evaluation");N=s;let i=await (0,c.getSystemAgentId)("opencode",A),o=(0,c.findSystemAgentDefinition)("opencode",A);if((0,u.tagOpencodeSession)(s,{agentName:A,agentId:i,skill:o?.traceSkill,displayQuery:e.caseInput,user:t||void 0}),I=(await O.chat(s,R,{onText:e=>{I+=e.delta},onError:e=>{C=e},onSubagent:e=>{$=e.agent||e.sessionID||"unknown-subagent",console.warn(`[opencode-task-completion] unexpected subagent spawned: ${$}`)}},{streamTimeoutMs:6e5,idleTimeoutMs:18e4})).text||I,await (0,d.recordEvaluatorExecution)(O,{taskId:s,agentName:A,user:t,query:e.caseInput}),$)throw Error(`任务完成度评估器不允许派发子代理，但实际派发了：${$}`);let l=(0,f.parseLooseJson)(I);if(l&&h(l)){let e=_(l,S);return{...e,rawAnalysis:{...e.rawAnalysis||{},evaluatorSessionId:N,root_cause_source:v}}}}catch(e){C=e instanceof Error?e:Error(String(e))}if(N&&O)try{await (0,d.recordEvaluatorExecution)(O,{taskId:N,agentName:A,user:t,query:e.caseInput})}catch(e){console.warn("[opencode-task-completion] failed to persist evaluator execution:",e?.message||e)}if($)throw C||Error(`任务完成度评估器不允许派发子代理，但实际派发了：${$}`);let j=function(e,...t){for(let a of t){let t=(0,f.parseLooseJson)(String(a||""));if(t&&h(t))return _(t,e)}return null}(S,I,C?.message);if(j)return{...j,rawAnalysis:{...j.rawAnalysis||{},evaluatorSessionId:N||void 0,root_cause_source:v}};let M=C?.message||`Agent 输出前 800 字符：${I.slice(0,800)}`;throw Error(`任务完成度评估器未产出有效 JSON。opencode 评测失败：${M}`)}),{taskType:"task-completion-eval",user:t??void 0,skill:a??void 0,skillVersion:p??null,label:`task-completion: ${(e.caseInput||"").slice(0,40)}`,silent:!0})}new Date().toISOString(),y("skill-aware"),y("no-skill"),e.s(["evaluateTaskCompletionViaOpencode",()=>v]),a()}catch(e){a(e)}},!1),508439,e=>e.a(async(t,a)=>{try{e.i(889228);var r=e.i(88045),n=e.i(322632),s=e.i(637351);e.i(384660);var i=e.i(106411),o=t([i]);[i]=o.then?(await o)():o;let f=n.z.object({normalized_input:n.z.string().default(""),confidence:n.z.number().min(0).max(1).default(0),ignored_parts:n.z.array(n.z.string()).default([]),reason:n.z.string().default("")}),g=n.z.object({best_case_id:n.z.string().nullable().default(null),best_confidence:n.z.number().min(0).max(1).default(0),reason:n.z.string().default("")});function l(e){let t=e.trim(),a=t.match(/```(?:json)?\s*([\s\S]*?)\s*```/i);if(a)t=a[1];else{let e=t.indexOf("{"),a=t.lastIndexOf("}");-1!==e&&-1!==a&&a>=e&&(t=t.substring(e,a+1))}return JSON.parse(t)}async function u(e){var t;let a,n=await (0,i.getActiveConfig)(e);if(!n)return{client:null,model:null};let o=n.apiKey||"no-api-key-required",l=(t=n.baseUrl||"https://api.deepseek.com",String(t||"").replace(/\/chat\/completions\/?$/,"")),{customFetch:u}=(0,s.getProxyConfig)();return{client:new r.OpenAI({apiKey:o,baseURL:l,fetch:u,timeout:(a=Number(process.env.SEMANTIC_MATCH_TIMEOUT_MS||process.env.JUDGMENT_TIMEOUT_MS||3e5),Number.isFinite(a)&&a>0?a:3e5)}),model:n.model||"deepseek-chat"}}async function c(e){let{client:t,model:a}=await u(e);if(!t||!a)return"No active evaluation model configured";try{let e=await t.chat.completions.create({messages:[{role:"user",content:"Hi"}],model:a,max_tokens:5});if(e?.choices)return null;return"No response from model"}catch(a){let e,t;return e=a?.message||"LLM request failed",t=a?.cause?.message?` (${a.cause.message})`:"",`${e}${t}`}}async function d(e,t){let a={normalized_input:String(e||"").trim(),confidence:0,ignored_parts:[],reason:"LLM extraction unavailable, fallback to raw input"};if(!a.normalized_input)return a;let{client:r,model:n}=await u(t);if(!r||!n)return a;try{var s;let e=await r.chat.completions.create({model:n,temperature:0,top_p:1,seed:42,messages:[{role:"user",content:(s=a.normalized_input,`你是一个 agent 评测系统中的“真实用户输入抽取器”。
+请你自行逐条检查关键观点覆盖情况，并在不派发任何子代理的前提下完成任务完成度评测和关键观点根因分析。`),agent:"build",model:{providerID:x,modelID:b,apiKey:k?.apiKey||E.apiKey,baseURL:k?.baseURL||E.baseUrl},modelOptions:{temperature:0,top_p:1,seed:42},system:y(S),permission:(0,i.buildEvaluatorPermissions)()},I="",C=null,N="",$=null,O=null;try{O=new r.AgentInsight({baseURL:a,logLevel:"warn"});let n=await O.createSession({title:`${A}-${Date.now()}`,directory:"/tmp"}),s=String(n?.id||n?.ID||"");if(!s)throw Error("Failed to create opencode session for task completion evaluation");N=s;let i=await (0,c.getSystemAgentId)("opencode",A),o=(0,c.findSystemAgentDefinition)("opencode",A);if((0,u.tagOpencodeSession)(s,{agentName:A,agentId:i,skill:o?.traceSkill,displayQuery:e.caseInput,user:t||void 0}),I=(await O.chat(s,R,{onText:e=>{I+=e.delta},onError:e=>{C=e},onSubagent:e=>{$=e.agent||e.sessionID||"unknown-subagent",console.warn(`[opencode-task-completion] unexpected subagent spawned: ${$}`)}},{streamTimeoutMs:6e5,idleTimeoutMs:18e4})).text||I,await (0,d.recordEvaluatorExecution)(O,{taskId:s,agentName:A,user:t,query:e.caseInput}),$)throw Error(`任务完成度评估器不允许派发子代理，但实际派发了：${$}`);let l=(0,f.parseLooseJson)(I);if(l&&h(l)){let e=_(l,S);return{...e,rawAnalysis:{...e.rawAnalysis||{},evaluatorSessionId:N,root_cause_source:v}}}}catch(e){C=e instanceof Error?e:Error(String(e))}if(N&&O)try{await (0,d.recordEvaluatorExecution)(O,{taskId:N,agentName:A,user:t,query:e.caseInput})}catch(e){console.warn("[opencode-task-completion] failed to persist evaluator execution:",e?.message||e)}if($)throw C||Error(`任务完成度评估器不允许派发子代理，但实际派发了：${$}`);let j=function(e,...t){for(let a of t){let t=(0,f.parseLooseJson)(String(a||""));if(t&&h(t))return _(t,e)}return null}(S,I,C?.message);if(j)return{...j,rawAnalysis:{...j.rawAnalysis||{},evaluatorSessionId:N||void 0,root_cause_source:v}};let M=C?.message||`Agent 输出前 800 字符：${I.slice(0,800)}`;throw Error(`任务完成度评估器未产出有效 JSON。opencode 评测失败：${M}`)}),{taskType:"task-completion-eval",user:t??void 0,skill:a??void 0,skillVersion:p??null,label:`task-completion: ${(e.caseInput||"").slice(0,40)}`,silent:!0})}new Date().toISOString(),y("skill-aware"),y("no-skill"),e.s(["evaluateTaskCompletionViaOpencode",()=>v]),a()}catch(e){a(e)}},!1),508439,e=>e.a(async(t,a)=>{try{e.i(889228);var r=e.i(88045),n=e.i(322632),s=e.i(637351);e.i(384660);var i=e.i(106411),o=t([i]);[i]=o.then?(await o)():o;let f=n.z.object({normalized_input:n.z.string().default(""),confidence:n.z.number().min(0).max(1).default(0),ignored_parts:n.z.array(n.z.string()).default([]),reason:n.z.string().default("")}),g=n.z.object({best_case_id:n.z.string().nullable().default(null),best_confidence:n.z.number().min(0).max(1).default(0),reason:n.z.string().default("")});function l(e){let t=e.trim(),a=t.match(/```(?:json)?\s*([\s\S]*?)\s*```/i);if(a)t=a[1];else{let e=t.indexOf("{"),a=t.lastIndexOf("}");-1!==e&&-1!==a&&a>=e&&(t=t.substring(e,a+1))}return JSON.parse(t)}async function u(e){var t;let a,n=await (0,i.getActiveConfig)(e);if(!n)return{client:null,model:null};let o=n.apiKey||"no-api-key-required",l=(t=n.baseUrl||"https://api.deepseek.com",String(t||"").replace(/\/chat\/completions\/?$/,"")),{customFetch:u}=(0,s.getProxyConfig)();return{client:new r.OpenAI({apiKey:o,baseURL:l,fetch:u,timeout:(a=Number(process.env.SEMANTIC_MATCH_TIMEOUT_MS||process.env.JUDGMENT_TIMEOUT_MS||3e5),Number.isFinite(a)&&a>0?a:3e5)}),model:n.model||"deepseek-chat"}}async function c(e){let{client:t,model:a}=await u(e);if(!t||!a)return"No active evaluation model configured";try{let e=await t.chat.completions.create({messages:[{role:"user",content:"Hi"}],model:a,max_tokens:5});if(e?.choices)return null;return"No response from model"}catch(a){let e,t;return e=a?.message||"LLM request failed",t=a?.cause?.message?` (${a.cause.message})`:"",`${e}${t}`}}async function d(e,t){let a={normalized_input:String(e||"").trim(),confidence:0,ignored_parts:[],reason:"LLM extraction unavailable, fallback to raw input"};if(!a.normalized_input)return a;let{client:r,model:n}=await u(t);if(!r||!n)return a;try{var s;let e=await r.chat.completions.create({model:n,temperature:0,top_p:1,seed:42,messages:[{role:"user",content:(s=a.normalized_input,`你是一个 agent 评测系统中的“真实用户输入抽取器”。
 
 你的任务：
 从 raw_input 中抽取真正应该用于评测数据集匹配的用户输入。
@@ -170,9 +170,9 @@ normalized_input:
 ${e}
 
 candidate_cases:
-${JSON.stringify(s,null,2)}`)}]}),i=a.choices?.[0]?.message?.content?.trim();if(!i)return{confidence:0,reason:"Empty LLM batch match response",error:"Empty LLM batch match response"};let o=g.parse(l(i)),u=t.find(e=>e.id===o.best_case_id);if(!u)return{confidence:0,reason:"LLM returned unknown case id",error:"LLM returned unknown case id"};return{id:u.id,confidence:o.best_confidence,reason:o.reason}}catch(t){console.warn("[SemanticMatch] Failed to compare candidate batch:",t);let e=t?.message||String(t);return{confidence:0,reason:"LLM batch match failed",error:e}}}async function m(e,t,a){let r=a?.batchSize??100,n=a?.threshold??.75,s=a?.requireModelAvailable??!1,i=e.map(e=>({id:e.id,query:e.input})).filter(e=>e.id&&e.query.trim()),o=await d(t,a?.user),l=o.normalized_input.trim()||String(t||"").trim();if(!l||0===i.length)return{normalizedInput:l,extractionConfidence:o.confidence,matchConfidence:0,ignoredParts:o.ignored_parts,extractReason:o.reason,matchReason:0===i.length?"No candidate cases available":"Normalized input is empty",matchedBy:"none"};if(s){let e=await c(a?.user);if(e)return{normalizedInput:l,extractionConfidence:o.confidence,matchConfidence:0,ignoredParts:o.ignored_parts,extractReason:o.reason,matchReason:"Semantic matcher model connection failed",matchedBy:"none",error:e}}let u=function(e,t){let a=[];for(let r=0;r<e.length;r+=t)a.push(e.slice(r,r+t));return a}(i,Math.max(1,r)),m=await Promise.all(u.map(e=>p(l,e,a?.user))),f=m.map(e=>e.error).filter(e=>!!e),g=m.reduce((e,t)=>{if(!t.id)return e;if(!e.id)return t;if(t.confidence!==e.confidence)return t.confidence>e.confidence?t:e;let a=i.find(e=>e.id===t.id)?.query.length??0,r=i.find(t=>t.id===e.id)?.query.length??0;return a>r?t:e},{confidence:0,reason:"No semantic match found"});return!g.id||g.confidence<n?{normalizedInput:l,extractionConfidence:o.confidence,matchConfidence:g.confidence||0,ignoredParts:o.ignored_parts,extractReason:o.reason,matchReason:g.reason||"Best semantic confidence below threshold",matchedBy:"none",...f.length===m.length&&f.length>0?{error:f.join("; ")}:{}}:{caseId:g.id,normalizedInput:l,extractionConfidence:o.confidence,matchConfidence:g.confidence,ignoredParts:o.ignored_parts,extractReason:o.reason,matchReason:g.reason,matchedBy:"semantic"}}e.s(["extractRealUserInput",()=>d,"findBestSemanticCaseMatch",()=>m]),a()}catch(e){a(e)}},!1),670497,e=>e.a(async(t,a)=>{try{e.i(889228);var r=e.i(88045),n=e.i(322632),s=e.i(637351),i=e.i(106411),o=t([i]);[i]=o.then?(await o)():o;let E=n.z.object({has_task_output:n.z.boolean().default(!1),confidence:n.z.number().min(0).max(1).default(0),selected_source_id:n.z.string().nullable().default(null),needs_followup_context:n.z.boolean().default(!1),followup_anchor_llm_call_id:n.z.string().nullable().default(null),reason:n.z.string().default("")}),x=n.z.object({has_task_output:n.z.boolean().default(!1),confidence:n.z.number().min(0).max(1).default(0),selected_source_id:n.z.string().nullable().default(null),artifact_paths:n.z.array(n.z.string()).default([]),reason:n.z.string().default("")});async function l(e){var t;let a,n=await (0,i.getActiveConfig)(e);if(!n)return{client:null,model:null};let{customFetch:o}=(0,s.getProxyConfig)();return{client:new r.OpenAI({apiKey:n.apiKey||"no-api-key-required",baseURL:(t=n.baseUrl||"https://api.deepseek.com",String(t||"").replace(/\/chat\/completions\/?$/,"")),fetch:o,timeout:(a=Number(process.env.RESULT_ARTIFACT_TIMEOUT_MS||process.env.JUDGMENT_TIMEOUT_MS||3e5),Number.isFinite(a)&&a>0?a:3e5)}),model:n.model||"deepseek-chat"}}function u(e){let t=e.trim();if(!t)return null;let a=t.match(/```(?:json)?\s*([\s\S]*?)\s*```/i);if(a)t=a[1];else{let e=t.indexOf("{"),a=t.lastIndexOf("}");-1!==e&&-1!==a&&a>=e&&(t=t.substring(e,a+1))}try{return JSON.parse(t)}catch{return null}}function c(e){return e&&"object"==typeof e&&!Array.isArray(e)?e:null}function d(e){if(null==e)return"";if("string"==typeof e)return e;if("number"==typeof e||"boolean"==typeof e)return String(e);try{return JSON.stringify(e,null,2)}catch{return String(e)}}function p(e,t){let a=String(e||"").trim();if(a.length<=t)return a;let r=Math.floor(.65*t),n=Math.max(0,t-r-80);return`${a.slice(0,r)}
+${JSON.stringify(s,null,2)}`)}]}),i=a.choices?.[0]?.message?.content?.trim();if(!i)return{confidence:0,reason:"Empty LLM batch match response",error:"Empty LLM batch match response"};let o=g.parse(l(i)),u=t.find(e=>e.id===o.best_case_id);if(!u)return{confidence:0,reason:"LLM returned unknown case id",error:"LLM returned unknown case id"};return{id:u.id,confidence:o.best_confidence,reason:o.reason}}catch(t){console.warn("[SemanticMatch] Failed to compare candidate batch:",t);let e=t?.message||String(t);return{confidence:0,reason:"LLM batch match failed",error:e}}}async function m(e,t,a){let r=a?.batchSize??100,n=a?.threshold??.75,s=a?.requireModelAvailable??!1,i=e.map(e=>({id:e.id,query:e.input})).filter(e=>e.id&&e.query.trim()),o=await d(t,a?.user),l=o.normalized_input.trim()||String(t||"").trim();if(!l||0===i.length)return{normalizedInput:l,extractionConfidence:o.confidence,matchConfidence:0,ignoredParts:o.ignored_parts,extractReason:o.reason,matchReason:0===i.length?"No candidate cases available":"Normalized input is empty",matchedBy:"none"};if(s){let e=await c(a?.user);if(e)return{normalizedInput:l,extractionConfidence:o.confidence,matchConfidence:0,ignoredParts:o.ignored_parts,extractReason:o.reason,matchReason:"Semantic matcher model connection failed",matchedBy:"none",error:e}}let u=function(e,t){let a=[];for(let r=0;r<e.length;r+=t)a.push(e.slice(r,r+t));return a}(i,Math.max(1,r)),m=await Promise.all(u.map(e=>p(l,e,a?.user))),f=m.map(e=>e.error).filter(e=>!!e),g=m.reduce((e,t)=>{if(!t.id)return e;if(!e.id)return t;if(t.confidence!==e.confidence)return t.confidence>e.confidence?t:e;let a=i.find(e=>e.id===t.id)?.query.length??0,r=i.find(t=>t.id===e.id)?.query.length??0;return a>r?t:e},{confidence:0,reason:"No semantic match found"});return!g.id||g.confidence<n?{normalizedInput:l,extractionConfidence:o.confidence,matchConfidence:g.confidence||0,ignoredParts:o.ignored_parts,extractReason:o.reason,matchReason:g.reason||"Best semantic confidence below threshold",matchedBy:"none",...f.length===m.length&&f.length>0?{error:f.join("; ")}:{}}:{caseId:g.id,normalizedInput:l,extractionConfidence:o.confidence,matchConfidence:g.confidence,ignoredParts:o.ignored_parts,extractReason:o.reason,matchReason:g.reason,matchedBy:"semantic"}}e.s(["extractRealUserInput",()=>d,"findBestSemanticCaseMatch",()=>m]),a()}catch(e){a(e)}},!1),670497,e=>e.a(async(t,a)=>{try{e.i(889228);var r=e.i(88045),n=e.i(322632),s=e.i(637351),i=e.i(106411),o=t([i]);[i]=o.then?(await o)():o;let E=n.z.object({has_task_output:n.z.boolean().default(!1),confidence:n.z.number().min(0).max(1).default(0),selected_source_id:n.z.string().nullable().default(null),needs_followup_context:n.z.boolean().default(!1),followup_anchor_llm_call_id:n.z.string().nullable().default(null),reason:n.z.string().default("")}),k=n.z.object({has_task_output:n.z.boolean().default(!1),confidence:n.z.number().min(0).max(1).default(0),selected_source_id:n.z.string().nullable().default(null),artifact_paths:n.z.array(n.z.string()).default([]),reason:n.z.string().default("")});async function l(e){var t;let a,n=await (0,i.getActiveConfig)(e);if(!n)return{client:null,model:null};let{customFetch:o}=(0,s.getProxyConfig)();return{client:new r.OpenAI({apiKey:n.apiKey||"no-api-key-required",baseURL:(t=n.baseUrl||"https://api.deepseek.com",String(t||"").replace(/\/chat\/completions\/?$/,"")),fetch:o,timeout:(a=Number(process.env.RESULT_ARTIFACT_TIMEOUT_MS||process.env.JUDGMENT_TIMEOUT_MS||3e5),Number.isFinite(a)&&a>0?a:3e5)}),model:n.model||"deepseek-chat"}}function u(e){let t=e.trim();if(!t)return null;let a=t.match(/```(?:json)?\s*([\s\S]*?)\s*```/i);if(a)t=a[1];else{let e=t.indexOf("{"),a=t.lastIndexOf("}");-1!==e&&-1!==a&&a>=e&&(t=t.substring(e,a+1))}try{return JSON.parse(t)}catch{return null}}function c(e){return e&&"object"==typeof e&&!Array.isArray(e)?e:null}function d(e){if(null==e)return"";if("string"==typeof e)return e;if("number"==typeof e||"boolean"==typeof e)return String(e);try{return JSON.stringify(e,null,2)}catch{return String(e)}}function p(e,t){let a=String(e||"").trim();if(a.length<=t)return a;let r=Math.floor(.65*t),n=Math.max(0,t-r-80);return`${a.slice(0,r)}
 ...[中间内容已截断 ${a.length-r-n} 字]...
-${a.slice(-n)}`}function m(e){return"string"==typeof e?e:Array.isArray(e)?e.map(e=>{let t=c(e);return t?d(t.text??t.content??t.input??t):d(e)}).filter(Boolean).join("\n"):d(e)}function f(e){let t=e?.tool_calls??e?.toolCalls;return Array.isArray(t)?t.map(c).filter(e=>!!e):[]}function g(e){let t=c(e.function);return String(t?.name||e.name||e.type||"tool").trim()}function y(e){let t=c(e.function);return d(t?.arguments??e.arguments??e.input??"")}function h(e){return d(e.output??e.result??e.content??"")}function w(e,t){return t&&e.find(e=>e.id===t)?.text?.trim()||""}function _(e,t){return!!t&&e.some(e=>e.id===t)}function v(e,t,a,r,n){return{status:"found",outputForEvaluation:e,confidence:t,reason:a,sourceRefs:r?[r]:[],rawAnalysis:n}}async function A(e,t,a){let r=await e.chat.completions.create({model:t,temperature:0,top_p:1,seed:42,messages:[{role:"user",content:a}]});return r.choices?.[0]?.message?.content?.trim()||""}async function S(e){let t=e.threshold??.75,a=e.maxLlmCalls??10,r=String(e.fallbackOutput||"").trim(),n=(e,t={})=>r?function(e,t,a={}){return v(e,0,`未能从 trace 中定位真实任务输出，已回退使用 execution.finalResult。原始原因：${t}`,"execution.finalResult",{...a,fallback:{used:!0,source:"execution.finalResult",reason:t}})}(r,e,t):function(e,t={}){return{status:"missing",outputForEvaluation:null,confidence:0,reason:e,sourceRefs:[],rawAnalysis:t}}(e,t),{client:s,model:i}=await l(e.user);if(!s||!i)return n("未配置可用评测模型，无法提取真实输出。");let o=function(e){let t=[],a=0,r=0,n=0;for(let s=0;s<e.length;s+=1){let i=c(e[s]);if(!i)continue;let o=Array.isArray(i.requestMessages)?i.requestMessages:[];for(let e of o){let r=c(e);r&&"tool"===String(r.role||"").toLowerCase()&&(n+=1,t.push({id:`tool-${n}`,type:"tool",order:a++,interactionIndex:s,name:String(r.name||r.tool_call_id||"tool"),input:"",output:m(r.content)}))}let l=c(i.responseMessage);if(l){let e=m(l.content);(e||f(l).length>0)&&(r+=1,t.push({id:`llm-${r}`,type:"llm",order:a++,interactionIndex:s,agent:String(i.agent||i.subagent_name||""),input:function(e,t){if(!Array.isArray(e))return"";let a=e.map((e,t)=>{let a=c(e);if(!a)return"";let r=String(a.role||"message"),n=a.name?`/${String(a.name)}`:"";return`[${t+1}] ${r}${n}: ${m(a.content)}`}).filter(Boolean).join("\n");return p(a,1800)}(o,0),inputParts:Array.isArray(o)?o.map(e=>{let t=c(e);if(!t)return null;let a=m(t.content).trim();return a?{role:String(t.role||"message"),text:a,...t.name?{name:String(t.name)}:{}}:null}).filter(e=>!!e):[],output:e}));for(let e of f(l))n+=1,t.push({id:`tool-${n}`,type:"tool",order:a++,interactionIndex:s,name:g(e),input:y(e),output:h(e)});continue}let u=String(i.role||"").toLowerCase();for(let e of("tool"===u?(n+=1,t.push({id:`tool-${n}`,type:"tool",order:a++,interactionIndex:s,name:String(i.name||i.tool_call_id||"tool"),input:"",output:m(i.content)})):("assistant"===u||"subagent"===u||"agent"===u||"opencode"===u)&&(r+=1,t.push({id:`llm-${r}`,type:"llm",order:a++,interactionIndex:s,agent:String(i.agent||i.subagent_name||""),input:"",output:m(i.content)})),f(i)))n+=1,t.push({id:`tool-${n}`,type:"tool",order:a++,interactionIndex:s,name:g(e),input:y(e),output:h(e)})}return t}(e.interactions||[]),d=o.filter(e=>"llm"===e.type&&e.output?.trim()).slice(-Math.max(1,a));if(0===d.length)return n("链路中没有可用于输出提取的 LLM 调用。",{event_count:o.length});try{var S,k;let a,r,l=d.filter(e=>"llm"===e.type&&e.output?.trim()).map(e=>({id:`${e.id}.output`,eventId:e.id,kind:"llm_output",text:e.output.trim(),agent:e.agent,name:e.name,interactionIndex:e.interactionIndex})),c=await A(s,i,(S=e.userTask,a=l.map(e=>({id:e.id,event_id:e.eventId,kind:e.kind,agent:e.agent||"",name:e.name||"",interaction_index:e.interactionIndex,text_preview:p(e.text,"llm_input"===e.kind?1800:3600)})),`你是 agent 评测系统中的“真实任务输出定位器”。
+${a.slice(-n)}`}function m(e){return"string"==typeof e?e:Array.isArray(e)?e.map(e=>{let t=c(e);return t?d(t.text??t.content??t.input??t):d(e)}).filter(Boolean).join("\n"):d(e)}function f(e){let t=e?.tool_calls??e?.toolCalls;return Array.isArray(t)?t.map(c).filter(e=>!!e):[]}function g(e){let t=c(e.function);return String(t?.name||e.name||e.type||"tool").trim()}function y(e){let t=c(e.function);return d(t?.arguments??e.arguments??e.input??"")}function h(e){return d(e.output??e.result??e.content??"")}function w(e,t){return t&&e.find(e=>e.id===t)?.text?.trim()||""}function _(e,t){return!!t&&e.some(e=>e.id===t)}function v(e,t,a,r,n){return{status:"found",outputForEvaluation:e,confidence:t,reason:a,sourceRefs:r?[r]:[],rawAnalysis:n}}async function A(e,t,a){let r=await e.chat.completions.create({model:t,temperature:0,top_p:1,seed:42,messages:[{role:"user",content:a}]});return r.choices?.[0]?.message?.content?.trim()||""}async function S(e){let t=e.threshold??.75,a=e.maxLlmCalls??10,r=String(e.fallbackOutput||"").trim(),n=(e,t={})=>r?function(e,t,a={}){return v(e,0,`未能从 trace 中定位真实任务输出，已回退使用 execution.finalResult。原始原因：${t}`,"execution.finalResult",{...a,fallback:{used:!0,source:"execution.finalResult",reason:t}})}(r,e,t):function(e,t={}){return{status:"missing",outputForEvaluation:null,confidence:0,reason:e,sourceRefs:[],rawAnalysis:t}}(e,t),{client:s,model:i}=await l(e.user);if(!s||!i)return n("未配置可用评测模型，无法提取真实输出。");let o=function(e){let t=[],a=0,r=0,n=0;for(let s=0;s<e.length;s+=1){let i=c(e[s]);if(!i)continue;let o=Array.isArray(i.requestMessages)?i.requestMessages:[];for(let e of o){let r=c(e);r&&"tool"===String(r.role||"").toLowerCase()&&(n+=1,t.push({id:`tool-${n}`,type:"tool",order:a++,interactionIndex:s,name:String(r.name||r.tool_call_id||"tool"),input:"",output:m(r.content)}))}let l=c(i.responseMessage);if(l){let e=m(l.content);(e||f(l).length>0)&&(r+=1,t.push({id:`llm-${r}`,type:"llm",order:a++,interactionIndex:s,agent:String(i.agent||i.subagent_name||""),input:function(e,t){if(!Array.isArray(e))return"";let a=e.map((e,t)=>{let a=c(e);if(!a)return"";let r=String(a.role||"message"),n=a.name?`/${String(a.name)}`:"";return`[${t+1}] ${r}${n}: ${m(a.content)}`}).filter(Boolean).join("\n");return p(a,1800)}(o,0),inputParts:Array.isArray(o)?o.map(e=>{let t=c(e);if(!t)return null;let a=m(t.content).trim();return a?{role:String(t.role||"message"),text:a,...t.name?{name:String(t.name)}:{}}:null}).filter(e=>!!e):[],output:e}));for(let e of f(l))n+=1,t.push({id:`tool-${n}`,type:"tool",order:a++,interactionIndex:s,name:g(e),input:y(e),output:h(e)});continue}let u=String(i.role||"").toLowerCase();for(let e of("tool"===u?(n+=1,t.push({id:`tool-${n}`,type:"tool",order:a++,interactionIndex:s,name:String(i.name||i.tool_call_id||"tool"),input:"",output:m(i.content)})):("assistant"===u||"subagent"===u||"agent"===u||"opencode"===u)&&(r+=1,t.push({id:`llm-${r}`,type:"llm",order:a++,interactionIndex:s,agent:String(i.agent||i.subagent_name||""),input:"",output:m(i.content)})),f(i)))n+=1,t.push({id:`tool-${n}`,type:"tool",order:a++,interactionIndex:s,name:g(e),input:y(e),output:h(e)})}return t}(e.interactions||[]),d=o.filter(e=>"llm"===e.type&&e.output?.trim()).slice(-Math.max(1,a));if(0===d.length)return n("链路中没有可用于输出提取的 LLM 调用。",{event_count:o.length});try{var S,x;let a,r,l=d.filter(e=>"llm"===e.type&&e.output?.trim()).map(e=>({id:`${e.id}.output`,eventId:e.id,kind:"llm_output",text:e.output.trim(),agent:e.agent,name:e.name,interactionIndex:e.interactionIndex})),c=await A(s,i,(S=e.userTask,a=l.map(e=>({id:e.id,event_id:e.eventId,kind:e.kind,agent:e.agent||"",name:e.name||"",interaction_index:e.interactionIndex,text_preview:p(e.text,"llm_input"===e.kind?1800:3600)})),`你是 agent 评测系统中的“真实任务输出定位器”。
 
 你的任务：
 基于用户真实输入和最近的 LLM 调用，从 source_candidates 中选择“一条最像用户任务真实输出的完整原始数据”。
@@ -203,7 +203,7 @@ user_task:
 ${S}
 
 source_candidates:
-${JSON.stringify(a,null,2)}`)),m=u(c);if(!m)return n("第一轮输出提取模型未返回合法 JSON。",{pass:"first",raw_response:c,threshold:t});let f=E.parse(m),g=w(l,f.selected_source_id),y={pass:"first",response:f,raw_response:c,source_candidate_ids:l.map(e=>e.id),threshold:t};if(f.has_task_output&&g)return v(g,f.confidence,f.reason,f.selected_source_id,{...y,accepted_below_threshold:f.confidence<t});let h=f.followup_anchor_llm_call_id||"";if(!f.needs_followup_context||!h){let e=f.selected_source_id&&!_(l,f.selected_source_id)?`第一轮返回的 selected_source_id=${f.selected_source_id} 不在候选列表中。`:"";return n(e||f.reason||`输出定位置信度 ${f.confidence} 低于阈值 ${t}，且无需二次上下文。`,y)}let b=o.find(e=>"llm"===e.type&&e.id===h);if(!b)return n(`第一轮返回的 followup_anchor_llm_call_id=${h} 不存在。`,y);let R=o.find(e=>"llm"===e.type&&e.order>b.order),I=o.filter(e=>e.order>=b.order&&(!R||e.order<R.order)),C=function(e){let t=function(e){let t=[];for(let a of e)a.input?.trim()&&("llm"===a.type&&a.inputParts?.length?a.inputParts.forEach((e,r)=>{t.push({id:`${a.id}.input.${r+1}`,eventId:a.id,kind:"llm_input",text:e.text,agent:a.agent,name:a.name,interactionIndex:a.interactionIndex})}):t.push({id:`${a.id}.input`,eventId:a.id,kind:"llm"===a.type?"llm_input":"tool_input",text:a.input,agent:a.agent,name:a.name,interactionIndex:a.interactionIndex})),a.output?.trim()&&t.push({id:`${a.id}.output`,eventId:a.id,kind:"llm"===a.type?"llm_output":"tool_output",text:a.output,agent:a.agent,name:a.name,interactionIndex:a.interactionIndex});return t}(e),a=[],r=0;for(let e of t){let t="tool_output"===e.kind?4e3:1200,n=p(e.text,t);if(r+n.length>24e3)break;a.push(e),r+=n.length}return a}(I),N=await A(s,i,(k=e.userTask,r=C.map(e=>({id:e.id,event_id:e.eventId,kind:e.kind,agent:e.agent||"",name:e.name||"",interaction_index:e.interactionIndex,text_preview:p(e.text,"tool_output"===e.kind?4e3:1200)})),`你是 agent 评测系统中的“工具上下文产物定位器”。
+${JSON.stringify(a,null,2)}`)),m=u(c);if(!m)return n("第一轮输出提取模型未返回合法 JSON。",{pass:"first",raw_response:c,threshold:t});let f=E.parse(m),g=w(l,f.selected_source_id),y={pass:"first",response:f,raw_response:c,source_candidate_ids:l.map(e=>e.id),threshold:t};if(f.has_task_output&&g)return v(g,f.confidence,f.reason,f.selected_source_id,{...y,accepted_below_threshold:f.confidence<t});let h=f.followup_anchor_llm_call_id||"";if(!f.needs_followup_context||!h){let e=f.selected_source_id&&!_(l,f.selected_source_id)?`第一轮返回的 selected_source_id=${f.selected_source_id} 不在候选列表中。`:"";return n(e||f.reason||`输出定位置信度 ${f.confidence} 低于阈值 ${t}，且无需二次上下文。`,y)}let b=o.find(e=>"llm"===e.type&&e.id===h);if(!b)return n(`第一轮返回的 followup_anchor_llm_call_id=${h} 不存在。`,y);let R=o.find(e=>"llm"===e.type&&e.order>b.order),I=o.filter(e=>e.order>=b.order&&(!R||e.order<R.order)),C=function(e){let t=function(e){let t=[];for(let a of e)a.input?.trim()&&("llm"===a.type&&a.inputParts?.length?a.inputParts.forEach((e,r)=>{t.push({id:`${a.id}.input.${r+1}`,eventId:a.id,kind:"llm_input",text:e.text,agent:a.agent,name:a.name,interactionIndex:a.interactionIndex})}):t.push({id:`${a.id}.input`,eventId:a.id,kind:"llm"===a.type?"llm_input":"tool_input",text:a.input,agent:a.agent,name:a.name,interactionIndex:a.interactionIndex})),a.output?.trim()&&t.push({id:`${a.id}.output`,eventId:a.id,kind:"llm"===a.type?"llm_output":"tool_output",text:a.output,agent:a.agent,name:a.name,interactionIndex:a.interactionIndex});return t}(e),a=[],r=0;for(let e of t){let t="tool_output"===e.kind?4e3:1200,n=p(e.text,t);if(r+n.length>24e3)break;a.push(e),r+=n.length}return a}(I),N=await A(s,i,(x=e.userTask,r=C.map(e=>({id:e.id,event_id:e.eventId,kind:e.kind,agent:e.agent||"",name:e.name||"",interaction_index:e.interactionIndex,text_preview:p(e.text,"tool_output"===e.kind?4e3:1200)})),`你是 agent 评测系统中的“工具上下文产物定位器”。
 
 你的任务：
 从 source_candidates 中选择“一条最像用户任务真实输出的原始数据”。
@@ -227,9 +227,9 @@ ${JSON.stringify(a,null,2)}`)),m=u(c);if(!m)return n("第一轮输出提取模
 }
 
 user_task:
-${k}
+${x}
 
 source_candidates:
-${JSON.stringify(r,null,2)}`)),$=u(N);if(!$)return n("第二轮输出提取模型未返回合法 JSON。",{pass:"second",first:f,raw_first_response:c,raw_second_response:N,followup_anchor_llm_call_id:h,followup_event_ids:I.map(e=>e.id),threshold:t});let O=x.parse($),j=w(C,O.selected_source_id),M={pass:"second",first:f,second:O,raw_first_response:c,raw_second_response:N,followup_anchor_llm_call_id:h,followup_event_ids:I.map(e=>e.id),source_candidate_ids:C.map(e=>e.id),threshold:t};if(O.has_task_output&&j)return v(j,O.confidence,O.reason,O.selected_source_id,{...M,accepted_below_threshold:O.confidence<t});let T=O.selected_source_id&&!_(C,O.selected_source_id)?`第二轮返回的 selected_source_id=${O.selected_source_id} 不在候选列表中。`:"";return n(T||O.reason||`二次输出定位置信度 ${O.confidence} 低于阈值 ${t}。`,M)}catch(e){return n(`真实输出提取 LLM 调用失败：${e.message||String(e)}`,{error:e.message||String(e)})}}e.s(["extractTaskResultArtifact",()=>S]),a()}catch(e){a(e)}},!1),830484,e=>{"use strict";let t={trace_missing_input:{errorCategory:"trace",shortStatus:"Trace 缺输入",userMessage:"Trace 缺少实际输入，无法匹配评测集 case。",reason:"系统没有从本次执行记录中提取到可用于匹配 case 的任务输入。",nextAction:"检查 trace 采集是否包含用户任务输入；必要时重新采集或重新执行该任务。",retryable:!1,stage:"case-match"},trace_missing_output:{errorCategory:"trace",shortStatus:"Trace 缺输出",userMessage:"Trace 缺少实际输出，无法做任务完成度评测。",reason:"本次 trace 没有提取到 Agent 最终输出，也没有可用的 execution.finalResult fallback。",nextAction:"检查 Agent 是否产出最终回复；确认 trace 采集包含 assistant final message；必要时重新跑该 case。",retryable:!1,stage:"result-artifact"},task_output_extraction_failed:{errorCategory:"trace",shortStatus:"任务输出提取失败",userMessage:"无法从 trace 中提取任务输出。",reason:"任务输出提取流程没有定位到可用于任务完成度评测的最终输出，且没有 fallback 可用。",nextAction:"检查 trace 中是否包含最终回复或关键工具输出；必要时重试评测或重新采集 trace。",retryable:!0,stage:"result-artifact"},trace_data_empty:{errorCategory:"trace",shortStatus:"Trace 数据为空",userMessage:"Trace 数据为空，无法评测。",reason:"系统没有找到可用的 session interactions，也没有足够的 query/finalResult 生成兜底交互。",nextAction:"检查 trace 采集、上传或执行记录入库是否完整。",retryable:!1,stage:"trace-load"},trace_parse_failed:{errorCategory:"trace",shortStatus:"Trace 解析失败",userMessage:"Trace 数据解析失败，无法评测。",reason:"Session.interactions 不是合法 JSON 或结构不兼容。",nextAction:"检查 trace 采集数据格式，必要时重新上传或修复采集链路。",retryable:!1,stage:"trace-parse"},case_not_matched:{errorCategory:"case",shortStatus:"无匹配 case",userMessage:"Trace 输入没有匹配到可用于评测的 case。",reason:"本次任务输入与当前评测集中的 case.input 没有达到可用匹配。",nextAction:"补充对应 case，或确认本次 trace 的任务输入与评测集 case.input 是否一致。",retryable:!1,stage:"case-match"},case_match_failed:{errorCategory:"case",shortStatus:"匹配 case 失败",userMessage:"无法根据 trace 输入自动定位可用于评测的 case。",reason:"trace 输入没有精确匹配到评测集 case，后台自动匹配也没有得到可用结果。",nextAction:"检查本次 trace 的任务输入是否与评测集 case.input 对齐；必要时补充 case 或重新选择评测集。",retryable:!0,stage:"case-match"},dataset_missing:{errorCategory:"dataset",shortStatus:"评测集不存在",userMessage:"评测集不存在，无法匹配可评测 case。",reason:"评测结果关联的数据集已经不存在，或当前用户不可访问。",nextAction:"重新选择评测集后再发起评测。",retryable:!1,stage:"dataset-load"},dataset_empty:{errorCategory:"dataset",shortStatus:"评测集为空",userMessage:"评测集没有可评测 case。",reason:"当前匹配范围内没有可用于评测的 case。",nextAction:"补充评测集 case 后重新评测。",retryable:!1,stage:"dataset-load"},case_missing_expected_output:{errorCategory:"case",shortStatus:"Case 缺预期结果",userMessage:"Case 缺少预期结果，无法做任务完成度评测。",reason:"已匹配到 case，但 expectedOutput 为空。",nextAction:"为该 case 补充 expectedOutput 后重新评测。",retryable:!1,stage:"case-match"},custom_reference_missing:{errorCategory:"case",shortStatus:"缺 reference_output",userMessage:"自定义评估器缺少 reference_output。",reason:"自定义评估器 prompt 引用了 reference_output，但没有匹配到带预期结果的 case。",nextAction:"补充评测集 expectedOutput，或调整自定义评估器 prompt。",retryable:!1,stage:"custom-evaluator"},trace_step_analysis_failed:{errorCategory:"model",shortStatus:"步骤抽取失败",userMessage:"Trace 步骤动态分析失败。",reason:"系统尝试从 trace 中抽取实际执行步骤时失败，通常与模型配置或调用异常有关。",nextAction:"检查评测模型配置和服务状态后重试。",retryable:!0,stage:"trace-step-analysis"},trace_no_steps:{errorCategory:"trace",shortStatus:"Trace 无步骤",userMessage:"Trace 无可提取执行步骤，无法做轨迹评测。",reason:"系统没有从本次 trace 中提取到可用于轨迹评测的实际执行步骤。",nextAction:"检查 trace 质量、采集粒度或执行记录是否完整。",retryable:!1,stage:"trace-step-analysis"},evaluator_timeout:{errorCategory:"evaluator",shortStatus:"评测超时",userMessage:"评测器运行超时。",reason:"评测器在限定时间内没有返回结果。",nextAction:"重试评测；如果持续出现，请检查模型服务响应时间。",retryable:!0,stage:"evaluator"},evaluator_output_invalid:{errorCategory:"evaluator",shortStatus:"评测器输出异常",userMessage:"评测器未返回有效结构化结果。",reason:"评测器输出无法解析为系统需要的 JSON/评分结构。",nextAction:"重试评测；如果持续出现，请检查评测器模型输出。",retryable:!0,stage:"evaluator"},result_evaluator_failed:{errorCategory:"evaluator",shortStatus:"结果评测失败",userMessage:"任务完成度评测失败。",reason:"任务完成度评测器没有产出可用评分或评测 session。",nextAction:"重试评测；如果持续出现，请检查模型配置和评测器日志。",retryable:!0,stage:"result-evaluator"},persist_failed:{errorCategory:"system",shortStatus:"保存失败",userMessage:"评测结果保存失败。",reason:"评测已执行，但结果写入数据库时失败。",nextAction:"检查数据库或服务状态后重试。",retryable:!0,stage:"persist"},evaluator_recovered_failed:{errorCategory:"system",shortStatus:"评测进程异常",userMessage:"评测进程异常中断。",reason:"后台恢复逻辑发现该评测长期停留在 pending/running 状态。",nextAction:"重试评测，并查看服务日志确认是否有进程异常。",retryable:!0,stage:"recover"}};function a(e,r={}){let n=t[e]??{errorCategory:"system",shortStatus:"评测失败",userMessage:"评测执行失败。",reason:"评测执行过程中出现未分类异常。",nextAction:"重试评测；如果持续出现，请查看服务日志。",retryable:!0,stage:"unknown"};return{errorCode:e,errorCategory:n.errorCategory,shortStatus:n.shortStatus,userMessage:r.userMessage||n.userMessage,reason:r.reason||n.reason,nextAction:r.nextAction||n.nextAction,retryable:"boolean"==typeof r.retryable?r.retryable:n.retryable,stage:r.stage||n.stage,...r.details?{details:r.details}:{},...r.internal?{internal:r.internal}:{}}}e.s(["buildEvaluationDiagnostic",()=>a])},652192,e=>{"use strict";function t(e){switch(e.status){case"ok":return{state:"ok",code:"ok",message:"skill 关键步骤与 trace 实际步骤已完整对比"};case"no-skill-targets":return{state:"not-applicable",code:"no-skill-targets",message:"trace 未关联任何 skill，本次评测不做 skill 归因"};case"missing-skill":return{state:"degraded",code:"missing-skill",message:`trace 关联的 skill 在 Skills Hub 中找不到：${e.missingSkills.join("、")}`};case"missing-parsed-flow":return{state:"degraded",code:"missing-parsed-flow",message:`skill 尚未生成可用的解析流程，无法提取关键步骤：${e.missingSkills.join("、")}`};case"no-key-actions":return{state:"degraded",code:"no-key-actions",message:"skill flow 解析成功但未识别出任何关键步骤"};case"dynamic-analysis-failed":return{state:"degraded",code:"dynamic-analysis-failed",message:"trace 步骤动态分析失败（通常是 LLM 未配置或调用失败）"};case"no-extracted-steps":return{state:"degraded",code:"no-extracted-steps",message:"未能从 trace 中提取出实际执行步骤"}}}e.s(["buildSkillAttributionStatus",()=>t])},600461,e=>e.a(async(t,a)=>{try{let z;var r=e.i(89171),n=e.i(854763),s=e.i(948594),i=e.i(833821),o=e.i(511284),l=e.i(841212),u=e.i(664503),c=e.i(826917),d=e.i(418494),p=e.i(556602),m=e.i(926307),f=e.i(190014),g=e.i(508439),y=e.i(670497),h=e.i(794538),w=e.i(374995),_=e.i(830484),v=e.i(228356),A=e.i(240317),S=e.i(652192),E=e.i(450579),x=e.i(837926),k=t([n,s,i,l,u,d,f,g,y,v,E]);[n,s,i,l,u,d,f,g,y,v,E]=k.then?(await k)():k;let W=new Set(["trace-quality-evaluator","preset-agent-trace-quality","preset-agent-task-completion"]),B="preset-agent-trace-quality",K="preset-agent-task-completion",V=[1500,3e3,5e3,8e3],H={[B]:"Agent 轨迹质量",[K]:"Agent 任务完成度"},X=/\{\{\s*(input|output|reference_output|trajectory)\s*\}\}/g;function b(e){return String(e||"").trim().replace(/\s+/g," ")}function R(e){let t=(Array.isArray(e)?e:null==e?[]:[e]).map(e=>String(e||"").trim()).map(e=>"trace-quality-evaluator"===e?B:e).filter(e=>W.has(e)||(0,v.isCustomEvaluatorId)(e));return Array.from(new Set(t))}function I(e,t={}){return{selectedEvaluators:e,selectedEvaluatorNames:e.map(e=>{let a=H[e];if(a)return a;let r=t.customNameResolver?.(e);return r||e}),...t.autoWatch?{autoWatch:!0}:{},...t.watchedAgent?{watchedAgent:t.watchedAgent}:{},...t.autoWatch&&t.autoWatchEnabledAt?{autoWatchEnabledAt:t.autoWatchEnabledAt}:{}}}function C(e){let t=e?(0,h.parseLooseJson)(e):null,a=R(t?.selectedEvaluators),r=Array.isArray(t?.selectedEvaluatorNames)?t.selectedEvaluatorNames.filter(e=>"string"==typeof e&&!!e.trim()):[],n=t?.autoWatch===!0,s="string"==typeof t?.watchedAgent?t.watchedAgent.trim():"",i=(0,E.normalizeAutoWatchEnabledAt)(t?.autoWatchEnabledAt);return a.length>0?r.length===a.length?{selectedEvaluators:a,selectedEvaluatorNames:r,...n?{autoWatch:!0}:{},...s?{watchedAgent:s}:{},...n&&i?{autoWatchEnabledAt:i}:{}}:I(a,{autoWatch:n,watchedAgent:s,autoWatchEnabledAt:i}):I([B],{autoWatch:n,watchedAgent:s,autoWatchEnabledAt:i})}function N(e){let t=(0,h.parseLooseJson)(e||"");return t&&"object"==typeof t?t:{}}function $(e){if(!e||"object"!=typeof e||Array.isArray(e))return null;let t=String(e.source||"").trim(),a=String(e.grayscaleTaskId||"").trim(),r=String(e.caseId||"").trim(),n=String(e.side||"").trim(),s=String(e.executionTraceId||"").trim(),i=String(e.evaluationClaimId||"").trim(),o="number"==typeof e.runIndex&&Number.isFinite(e.runIndex)?Math.floor(e.runIndex):NaN,l="number"==typeof e.roundIndex&&Number.isFinite(e.roundIndex)?Math.floor(e.roundIndex):NaN,u="number"==typeof e.evaluationAttempt&&Number.isFinite(e.evaluationAttempt)?Math.floor(e.evaluationAttempt):NaN;return"grayscale-ab"===t&&a&&r&&("a"===n||"b"===n)&&s&&i&&Number.isFinite(o)&&Number.isFinite(l)&&Number.isFinite(u)?{source:"grayscale-ab",grayscaleTaskId:a,caseId:r,side:n,runIndex:o,roundIndex:l,executionTraceId:s,evaluationClaimId:i,evaluationAttempt:u}:null}async function O(e,t,a,r,n,s,i,o,l,c,d){let p=await (0,u.evaluateTaskCompletionViaOpencode)({caseInput:e,expectedOutput:t,actualOutput:a,precomputedRootCauses:r,precomputedRootCauseSource:n,traceSummaryText:s,skillContext:i,skillAttributionMode:o},l,c,d);return{isCorrect:p.isCorrect,score:p.score,reason:p.reason,rawAnalysis:p.rawAnalysis}}async function j(e,t,a){return Promise.race([e,new Promise((e,r)=>setTimeout(()=>r(a),t))])}function M(e,t){let a=t.now??new Date;return{attemptCount:e,maxAttempts:5,retrying:t.retrying,exhausted:t.exhausted,lastAttemptAt:a.toISOString(),...t.lastError?{lastError:t.lastError}:{},...t.retrying&&t.delayMs?{nextRetryAt:new Date(a.getTime()+t.delayMs).toISOString()}:{}}}async function T(e,t,a){let r=a.rawAnalysis&&"object"==typeof a.rawAnalysis?a.rawAnalysis:null,s=r?.key_point_findings;if(!("number"==typeof a.score&&Number.isFinite(a.score)&&String(a.reason||"").trim()&&Array.isArray(s)))return void console.warn("[trajectory-eval] skip persisting incomplete result-evaluation payload to execution");let i={isAnswerCorrect:a.isCorrect,answerScore:a.score,judgmentReason:a.reason};e?.id?await n.prismaRaw.execution.update({where:{id:e.id},data:i}):t&&await n.prismaRaw.execution.updateMany({where:{taskId:t},data:i})}async function L(e,t,a={}){let r=b(t);if(!r)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} trace 没有实际输入，无法匹配评估数据集 case`);let n=!0===a.requireExpectedOutput,s=!0===a.includeAllDatasetKinds,o=Array.isArray(a.allowedDatasetIds)?a.allowedDatasetIds.map(e=>String(e).trim()).filter(Boolean):[],l=(await (0,i.readAllAgentDatasets)()).filter(t=>t.user===e).filter(e=>0===o.length||o.includes(e.id)).filter(e=>s||n||"trajectory"===e.datasetKind);if(0===l.length)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} ${s?"未找到可用于评测的数据集":n?"未找到可用于结果评测的数据集":"未找到轨迹评测数据集"}`);for(let e of l){let t=e.cases.find(e=>b(e.input)===r&&(!n||!!b(e.expectedOutput)));if(t)return{dataset:e,caseEntry:t}}let u=[],c=new Map;for(let e of l)for(let t of e.cases){if(n&&!b(t.expectedOutput))continue;let a=`${e.id}::${t.id}`;u.push({id:a,input:t.input}),c.set(a,{dataset:e,caseEntry:t})}let d=await (0,g.findBestSemanticCaseMatch)(u,t,{user:e,requireModelAvailable:!0});if(d.error)throw new Q("semantic-match-llm",`${p.NO_EVALUABLE_CASE_PREFIX} 语义匹配调用评测模型失败：${d.error}`);if(d.caseId){let e=c.get(d.caseId);if(e)return e}throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} trace 实际输入未匹配到${s?"评测数据集中的":n?"带预期结果的":"轨迹评测数据集中的"} case 输入`)}async function J(e){try{let t=await e.json(),a=String(t.user||"").trim(),o=String(t.evaluatorRunId||t.runId||"").trim(),l=$(t.grayscaleBinding),u=Array.isArray(t.evaluators)&&t.evaluators.length>0?t.evaluators:t.evaluator,c=R(u),d=!0===t.autoWatch,p=(0,E.normalizeAutoWatchAgent)(t.watchedAgent||t.agent||""),f=p,g=new Date().toISOString(),y=!0===t.placeholderOnly;if(!a)return r.NextResponse.json({error:"user is required"},{status:400});if(d&&p&&(0,x.isEvaluatorAgentName)(p))return r.NextResponse.json({error:"autoWatch requires a non-evaluator execution agent"},{status:400});let _=c.filter(v.isCustomEvaluatorId),A=new Set,S=new Map;if(_.length>0){let e=await (0,v.listCustomEvaluatorIds)(a),t=[];for(let a of _)e.has(a)?A.add(a):t.push(a);if(t.length>0)return r.NextResponse.json({error:`unknown custom evaluators: ${t.join(", ")}`},{status:400});for(let e of A){let t=await (0,v.loadCustomEvaluator)(a,e);t&&S.set(e,t.name)}}let k=c.filter(e=>(0,v.isCustomEvaluatorId)(e)?A.has(e):W.has(e)),b=I(k.length>0?k:[B],{autoWatch:d,watchedAgent:f,customNameResolver:e=>S.get(e)});if(u&&0===k.length||0===b.selectedEvaluators.length)return r.NextResponse.json({error:"unsupported evaluators"},{status:400});let N=Array.isArray(t.taskIds)?t.taskIds.map(e=>String(e).trim()).filter(Boolean):[],O=String(t.datasetId||"").trim(),j=Array.isArray(t.datasetIds)?t.datasetIds.map(e=>String(e).trim()).filter(Boolean):[],M=Array.isArray(t.pairs)?t.pairs:[],T=(0,w.normalizeTrajectoryTaskMeta)({title:t.taskTitle,description:t.taskDescription}),L=new Set,J="",F="";if(o){let e=await n.prismaRaw.trajectoryEvalResult.findMany({where:{user:a,evaluatorRunId:o},orderBy:{createdAt:"asc"}});if(0===e.length)return r.NextResponse.json({error:"evaluatorRunId not found"},{status:404});let t=e[0],s=[...e].reverse().map(e=>(function(e){let t=e?(0,h.parseLooseJson)(e):null,a=R(t?.selectedEvaluators);if(0===a.length)return null;let r=Array.isArray(t?.selectedEvaluatorNames)?t.selectedEvaluatorNames.filter(e=>"string"==typeof e&&!!e.trim()):[],n=t?.autoWatch===!0,s="string"==typeof t?.watchedAgent?t.watchedAgent.trim():"",i=(0,E.normalizeAutoWatchEnabledAt)(t?.autoWatchEnabledAt);return r.length===a.length?{selectedEvaluators:a,selectedEvaluatorNames:r,...n?{autoWatch:!0}:{},...s?{watchedAgent:s}:{},...n&&i?{autoWatchEnabledAt:i}:{}}:I(a,{autoWatch:n,watchedAgent:s,autoWatchEnabledAt:i})})(e.rawAnalysisJson)).find(e=>!!e);b=k.length>0?I(k,{autoWatch:d,watchedAgent:f,customNameResolver:e=>S.get(e)}):s||C(t.rawAnalysisJson),T=(0,w.extractTrajectoryTaskMeta)(t.rawAnalysisJson,t.createdAt),O=O||t.datasetId||"",J=(0,E.normalizeAutoWatchAgent)(s?.watchedAgent||await (0,E.resolveWatchedAgentForRunRows)(a,e)),F=(0,E.normalizeAutoWatchEnabledAt)(s?.autoWatchEnabledAt),L=new Set(e.filter(e=>"failed"!==e.status).map(e=>e.taskId||e.executionId||"").filter(Boolean))}if(d&&N.length>0){let e=await (0,E.resolveSingleTrajectoryCandidateAgent)(a,N);if(e.missingTaskIds.length>0)return r.NextResponse.json({error:`autoWatch only accepts traces that appear in 发起新评测: ${e.missingTaskIds.join(", ")}`},{status:400});if(!e.agent)return r.NextResponse.json({error:`autoWatch requires a single execution agent, got: ${e.distinctAgents.join(", ")||"none"}`},{status:400});if(J&&!(0,E.autoWatchAgentNamesMatch)(J,e.agent))return r.NextResponse.json({error:`autoWatch agent mismatch: run watches ${J}, new traces belong to ${e.agent}`},{status:400});f=e.agent}else{if(d&&!o&&!f)return r.NextResponse.json({error:"watchedAgent is required when starting autoWatch without seed traces"},{status:400});d&&J&&(f=J)}b=I(k.length>0?k:b.selectedEvaluators,{autoWatch:d,watchedAgent:d?f:"",autoWatchEnabledAt:d?o&&F||g:"",customNameResolver:e=>S.get(e)});let P=o||`trun_${Date.now()}_${Math.random().toString(36).slice(2,8)}`,D=[],U=[];if(N.length>0)for(let e of N){if(L.has(e)){U.push({caseId:"",reason:`taskId ${e} already exists in this run`});continue}let t=await n.prismaRaw.trajectoryEvalResult.create({data:{user:a,evaluatorRunId:P,datasetId:"",caseId:"",executionId:null,taskId:e,status:"pending",rawAnalysisJson:JSON.stringify({...b,taskMeta:T,...l?{grayscaleBinding:l}:{},...j.length>0?{allowedDatasetIds:j}:{}})}});D.push({id:t.id,caseId:"",taskId:e})}else if((d||y)&&!o){if(O){let e=await (0,i.findAgentDataset)(a,O);if(!e)return r.NextResponse.json({error:"dataset not found"},{status:404});if("trajectory"!==e.datasetKind)return r.NextResponse.json({error:`dataset kind must be 'trajectory', got '${e.datasetKind}'`},{status:400})}let e=await n.prismaRaw.trajectoryEvalResult.create({data:{user:a,evaluatorRunId:P,datasetId:O,caseId:"",executionId:null,taskId:null,status:"pending",rawAnalysisJson:JSON.stringify({...b,taskMeta:T,...l?{grayscaleBinding:l}:{},watchPlaceholder:!0,...y?{placeholderOnly:!0}:{}})}});D.push({id:e.id,caseId:""})}else{if(!O||!(M.length>0))return r.NextResponse.json({error:"provide taskIds, datasetId+pairs, or placeholderOnly:true"},{status:400});let e=await (0,i.findAgentDataset)(a,O);if(!e)return r.NextResponse.json({error:"dataset not found"},{status:404});let t=new Map(e.cases.map(e=>[e.id,e]));for(let e of M){let r=String(e.caseId||"").trim(),s=e.executionId?String(e.executionId).trim():void 0,i=e.taskId?String(e.taskId).trim():void 0,o=i||s||"";if(!l&&o&&L.has(o)){U.push({caseId:r,reason:`${o} already exists in this run`});continue}if(!t.get(r)){U.push({caseId:r,reason:"case not found in dataset"});continue}if(!s&&!i){U.push({caseId:r,reason:"executionId or taskId is required"});continue}let u=await n.prismaRaw.trajectoryEvalResult.create({data:{user:a,evaluatorRunId:P,datasetId:O,caseId:r,executionId:s||null,taskId:i||null,status:"pending",rawAnalysisJson:JSON.stringify({...b,taskMeta:T,...l?{grayscaleBinding:l}:{}})}});D.push({id:u.id,caseId:r,executionId:s,taskId:i})}}if(0===D.length)return r.NextResponse.json({error:"no valid tasks to run",skipped:U},{status:400});Promise.all(D.map(e=>(0,s.reattributeServiceTraceOwner)((e.taskId||e.executionId||"").trim(),a).catch(()=>!1)));let z=D.filter(e=>e.taskId||e.executionId).map(e=>e.id);return z.length>0&&((0,m.registerTrajectoryEvalRun)(P,a),q(a,P,z).catch(e=>{console.error("[trajectory-eval] background run crashed:",e)}).finally(()=>(0,m.unregisterTrajectoryEvalRun)(P))),r.NextResponse.json({success:!0,evaluatorRunId:P,evaluators:b.selectedEvaluators,evaluatorNames:b.selectedEvaluatorNames,datasetId:O,created:D,skipped:U})}catch(t){console.error("trajectory/run POST error:",t);let e=t instanceof Error?t.message:"failed to start trajectory evaluation";return r.NextResponse.json({error:e},{status:500})}}async function F(e){try{let t=await e.json(),a=String(t.user||"").trim(),s=String(t.evaluatorRunId||t.runId||"").trim(),i=!0===t.autoWatch;if(!a)return r.NextResponse.json({error:"user is required"},{status:400});if(!s)return r.NextResponse.json({error:"evaluatorRunId is required"},{status:400});let o=await n.prismaRaw.trajectoryEvalResult.findMany({where:{user:a,evaluatorRunId:s},orderBy:{createdAt:"asc"}});if(0===o.length)return r.NextResponse.json({error:"evaluatorRunId not found"},{status:404});let l=C(o[0].rawAnalysisJson),u=(0,E.normalizeAutoWatchAgent)(l.watchedAgent),c=u&&!(0,x.isEvaluatorAgentName)(u)?u:await (0,E.resolveWatchedAgentForRunRows)(a,o);if(i&&!c)return r.NextResponse.json({error:"autoWatch requires existing traces from a single non-evaluator execution agent"},{status:400});let d=I(l.selectedEvaluators,{autoWatch:i,watchedAgent:c,autoWatchEnabledAt:i?new Date().toISOString():""});return await n.prismaRaw.$transaction(o.map(e=>{var t;let a;return n.prismaRaw.trajectoryEvalResult.update({where:{id:e.id},data:{rawAnalysisJson:(0,E.mergeWatchedAgentIntoRawAnalysis)((t=e.rawAnalysisJson,a={...N(t),selectedEvaluators:d.selectedEvaluators,selectedEvaluatorNames:d.selectedEvaluatorNames,watchedAgent:d.watchedAgent||"",autoWatchEnabledAt:d.autoWatchEnabledAt||""},d.autoWatch?a.autoWatch=!0:(delete a.autoWatch,delete a.autoWatchEnabledAt),d.watchedAgent||delete a.watchedAgent,d.autoWatchEnabledAt||delete a.autoWatchEnabledAt,JSON.stringify(a)),d.watchedAgent||"")}})})),r.NextResponse.json({success:!0,evaluatorRunId:s,autoWatch:i,watchedAgent:d.watchedAgent||""})}catch(t){console.error("trajectory/run PATCH error:",t);let e=t instanceof Error?t.message:"failed to update auto watch state";return r.NextResponse.json({error:e},{status:500})}}let G=(z=Number(process.env.MAX_EVAL_ROW_CONCURRENCY),Number.isFinite(z)&&z>0?Math.min(20,Math.floor(z)):4),Z=Symbol.for("@witty-insight/trajectory-eval-row-limiter");class Q extends Error{stage;cause;constructor(e,t,a){super(`[${e}] ${t}`),this.stage=e,this.cause=a,this.name="StagedEvaluationError"}}async function P(e,t){let a,r=((a=globalThis)[Z]||(a[Z]=(0,p.createSimpleAsyncLimiter)(G)),a[Z]);await r.acquire();try{await D(e,t)}finally{r.release()}}async function D(e,t){let a=null,r=null;try{var s,u,d,h;let E,x,k,R;await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:"running",errorMessage:null}});let I=await n.prismaRaw.trajectoryEvalResult.findUnique({where:{id:t}});if(!I)throw new Q("lookup",`result row ${t} disappeared`);if((0,m.getTrajectoryEvalRunSignal)(I.evaluatorRunId)?.aborted)return void await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:"failed",errorMessage:"已终止：用户中止了该批次评测"}}).catch(()=>{});I.taskId&&(a=I.taskId,r=(0,c.startOrReplace)(e,I.taskId,"rejudge").runId);let J=null;I.executionId&&(J=await n.prismaRaw.execution.findUnique({where:{id:I.executionId},select:{skill:!0,skillVersion:!0}}).catch(()=>null)),!J&&I.taskId&&(J=await n.prismaRaw.execution.findFirst({where:{taskId:I.taskId},orderBy:{timestamp:"desc"},select:{skill:!0,skillVersion:!0}}).catch(()=>null));let F=J?.skill||null,D="number"==typeof J?.skillVersion?J.skillVersion:null,q=C(I.rawAnalysisJson),z=(0,w.extractTrajectoryTaskMeta)(I.rawAnalysisJson,I.createdAt),W=(s=I.rawAnalysisJson,x=N(s),$(x.grayscaleBinding)),H=q.selectedEvaluators.includes(B),G=q.selectedEvaluators.includes(K),Z=q.selectedEvaluators.filter(v.isCustomEvaluatorId),Y=Z.length>0,ee=null,et=null,ea=null,er=null,en=0,es=null,ei="",eo=null,el=null,eu=[],ec=Y?await Promise.all(Z.map(async t=>({evaluatorId:t,bundle:await (0,v.loadCustomEvaluator)(e,t)}))):[],ed=function(e){let t=new Set;for(let a of e){let e;for(X.lastIndex=0;null!==(e=X.exec(a||""));)t.add(e[1])}return t}(ec.flatMap(e=>[e.bundle?.config.systemPrompt||"",e.bundle?.config.userPrompt||""]).filter(Boolean)),ep=ed.has("reference_output"),em=ed.has("output"),ef=ed.has("trajectory"),eg=I.taskId,ey=I.executionId?await n.prismaRaw.execution.findUnique({where:{id:I.executionId}}):null;!eg&&I.executionId&&ey?.taskId&&(eg=ey.taskId),!ey&&eg&&(ey=await n.prismaRaw.execution.findFirst({where:{taskId:eg}}));let eh=ey?.query||"",ew=await (0,g.extractRealUserInput)(eh,e),e_=ew.normalized_input.trim()||eh,ev=String(ey?.finalResult||"").trim(),eA=(k=N(I.rawAnalysisJson),Array.isArray(k?.allowedDatasetIds)?k.allowedDatasetIds.map(e=>String(e).trim()).filter(Boolean):[]);await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...N(I.rawAnalysisJson),...q,taskMeta:z,caseSnapshot:{taskInput:e_,rawTaskInput:eh,taskInputExtraction:ew}})}});let eS=[];if(eg){let e=await n.prismaRaw.session.findUnique({where:{taskId:eg}});if(e?.interactions)try{let t=JSON.parse(e.interactions);eS=Array.isArray(t)?t:[]}catch(e){throw new Q("trace-parse",`Session.interactions JSON 解析失败: ${e.message}`,e)}}if(0===eS.length&&(e_||ev)){let e=new Date().toISOString();eS=[...e_?[{role:"user",content:e_,timestamp:e}]:[],...ev?[{role:"assistant",content:ev,timestamp:e}]:[]]}let eE=null,ex=null,ek="no-dataset";if(I.datasetId&&I.caseId){try{let t=await (0,i.findAgentDataset)(e,I.datasetId);if(!t)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 评估数据集不存在，无法匹配可评测 case`);ex={id:t.id,name:t.name};let a=b(e_);if(!a&&!I.caseId)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} trace 没有实际输入，无法匹配评估数据集 case`);let r=(I.caseId?t.cases.find(e=>e.id===I.caseId):t.cases.find(e=>b(e.input)===a))||null;if(r&&(ek=I.caseId?"explicit-pair":"exact-input"),!r&&!I.caseId){let a=await (0,g.findBestSemanticCaseMatch)(t.cases.map(e=>({id:e.id,input:e.input})),e_,{user:e,requireModelAvailable:!0});if(a.error)throw new Q("semantic-match-llm",`${p.NO_EVALUABLE_CASE_PREFIX} 语义匹配调用评测模型失败：${a.error}`);a.caseId&&(r=t.cases.find(e=>e.id===a.caseId)??null)&&(ek="semantic")}if(!r)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} trace 实际输入未匹配到当前数据集中的 case 输入`);eE=r}catch(e){if(e instanceof Q)throw e;throw new Q("dataset-load",`读取数据集失败: ${e.message}`,e)}if(!eE)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 未匹配到可评测 case`)}else if(G||H)try{let a=await L(e,e_,{requireExpectedOutput:G,allowedDatasetIds:eA});eE=a.caseEntry,ex={id:a.dataset.id,name:a.dataset.name},ek="auto-match",await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{datasetId:a.dataset.id,caseId:a.caseEntry.id}})}catch(e){if(e instanceof Q&&"no-evaluable-case"===e.stage&&!G)console.warn(`[trajectory/run] no matching dataset, falling back to empty case: ${e.message}`),eE={input:e_,expectedOutput:"",trajectory:"",evaluationFocus:"",rootCauses:[]},ek="fallback";else if(e instanceof Q)throw e;else throw new Q("dataset-auto-match",`自动匹配数据集失败: ${e.message}`,e)}else if(Y)if(ep)try{let a=await L(e,e_,{requireExpectedOutput:!0,includeAllDatasetKinds:!0,allowedDatasetIds:eA});eE=a.caseEntry,ex={id:a.dataset.id,name:a.dataset.name},ek="auto-match",await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{datasetId:a.dataset.id,caseId:a.caseEntry.id}})}catch(e){if(e instanceof Q)throw e;throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 自定义评估器需要 reference_output，但 trace 输入未匹配到带预期结果的数据集 case`,e)}else eE={input:e_,expectedOutput:"",trajectory:"",evaluationFocus:"",rootCauses:[]};else eE={input:e_,expectedOutput:"",trajectory:"",evaluationFocus:"",rootCauses:[]};if(G&&!b(eE.expectedOutput))throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 已匹配到 case，但该 case 缺少预期结果 expectedOutput，无法执行结果评测`);if(Y&&ep&&!b(eE.expectedOutput))throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 自定义评估器需要 reference_output，但未找到可用的预期结果 expectedOutput`);let eb=!!(eE?.rootCauseMeta&&(0,o.canReuseRootCauseCache)(eE.expectedOutput,eE.rootCauseMeta)),eR=eb&&eE?.rootCauseMeta?.status==="ready"?eE.rootCauses||[]:void 0,eI=eb?eE?.rootCauseMeta?.status==="empty"?"none":eE?.rootCauseMeta?.status==="ready"?"dataset-cache":void 0:void 0,eC=e_||eE.input||"",eN={id:eE.id||I.caseId||"",input:eE.input,taskInput:eC,rawTaskInput:eh,taskInputExtraction:ew,expectedOutput:eE.expectedOutput,trajectory:eE.trajectory,evaluationFocus:eE.evaluationFocus,rootCauseCacheStatus:eE.rootCauseMeta?.status||null,rootCauseCacheUpdatedAt:eE.rootCauseMeta?.updatedAt||null,datasetId:ex?.id||I.datasetId||"",datasetName:ex?.name||"",matchKind:ek},e$="trajectory",eO="",ej="",eM=[],eT=[],eL=eE.evaluationFocus,eJ=eE.trajectory,eF=eS.length>0?(u=eS,{summary:R=(0,A.summarizeTrace)(u,{maxSteps:80,maxTextLen:400}),text:(0,A.formatTraceForLLM)(R),steps:R.steps.map(e=>({...e,step_index:e.index,stepIndex:e.index}))}):null,eP=await (0,f.buildSkillKeyActionReference)(ey,e,eS),eD=(0,f.getPrimaryExecutionSkillTargets)(ey,eS).length>0?"skill-aware":"no-skill",eU="ok"===eP.status?{status:"ok",referenceKeyActionsText:eP.referenceKeyActionsText,actualExtractedStepsText:eF?.text||"",referenceKeyActions:eP.referenceKeyActions,actualExtractedSteps:eF?.steps||[]}:eP,eq=(0,S.buildSkillAttributionStatus)(eU);if("ok"===eP.status&&eF&&eF.steps.length>0)e$="skill_key_actions",eO=eP.referenceKeyActionsText,ej=eF.text,eM=eP.referenceKeyActions??[],eT=eF.steps,eL=[b(eE.evaluationFocus),"逐条比较 Skill 自动提取的关键动作与 trace 事件级步骤（user/llm/tool/skill/task），并直接生成每个关键动作的 Skill 改进建议。"].filter(Boolean).join(" ");else if(H){if(!eF||0===eF.steps.length)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 当前 trace 无法提取可评测事件步骤`);e$="trace_only",ej=eF.text,eT=eF.steps,eL=[b(eE.evaluationFocus),"当前 trace 未获得可用 Skill 关键动作；不评估完整性，不生成 Skill 关键动作建议，仅评估工具选择与冗余。"].filter(Boolean).join(" ")}let ez={...q,taskMeta:z,...W?{grayscaleBinding:W}:{},caseSnapshot:eN,skillAttribution:eq,resultSkillMode:eD,comparisonMode:e$,actualTraceExtraction:eF?{source:"summarizeTrace",totalSteps:eF.summary.totalSteps,shownSteps:eF.steps.length,truncated:eF.summary.truncated,totalLlmCalls:eF.summary.totalLlmCalls,totalToolCalls:eF.summary.totalToolCalls,totalSkillCalls:eF.summary.totalSkillCalls,totalTaskCalls:eF.summary.totalTaskCalls}:null,skillKeyActionComparison:"ok"===eP.status?{referenceKeyActionsText:eP.referenceKeyActionsText,actualExtractedStepsText:ej,referenceKeyActions:eP.referenceKeyActions??[],actualExtractedSteps:eT}:null};if((H||G)&&0===eS.length)throw new Q("trace-empty",`taskId=${eg||"N/A"} 对应的 Session 不存在或 interactions 为空`);if(H&&!G){let a=String(ey?.finalResult||"").trim();try{let t=await j((0,y.extractTaskResultArtifact)({userTask:eE.input,interactions:eS,fallbackOutput:a,user:e}),6e5,new Q("result-artifact-timeout",`实际输出提取超过 600s 未完成`));es=t.rawAnalysis,ei=String(t.outputForEvaluation||a||"").trim()}catch(e){console.warn("[trajectory-eval] result artifact extraction failed:",e),ei=a}await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:null,resultEvaluationError:null})}})}if(G){let a="skill-aware"===eD?await (0,f.loadTaskCompletionSkillContext)(ey,eS,e):void 0,r=eF?.text||"",s=Math.min(function(e){let t=N(e).resultEvaluationRetry;if(!t||"object"!=typeof t)return 0;let a=t.attemptCount;return"number"==typeof a&&Number.isFinite(a)&&a>0?Math.floor(a):0}(I.rawAnalysisJson)+1,5),i=new Date;try{await j((async()=>{let o=await (0,y.extractTaskResultArtifact)({userTask:eE.input,interactions:eS,fallbackOutput:ev,user:e});if(es=o.rawAnalysis,!(ei=String(o.outputForEvaluation||"").trim())){et=`结果输出提取失败：${o.reason}`,eo=(0,_.buildEvaluationDiagnostic)(String(o.reason||"").includes("没有可用于输出提取的 LLM 调用")?"trace_missing_output":"task_output_extraction_failed",{details:et,reason:et}),ea=M(s,{retrying:!1,exhausted:!0,lastError:et,now:i}),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:"",resultEvaluation:null,resultEvaluationError:et,resultEvaluationRetry:ea,diagnostic:eo})}});return}ea=M(s,{retrying:!1,exhausted:!1,now:i}),eo=null,await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:null,resultEvaluationError:null,resultEvaluationRetry:ea})}});let l=await O(eE.input,eE.expectedOutput,ei,eR,eI,r,a,eD,e,F,D);ee={...l.rawAnalysis||{},score:l.score,is_correct:l.isCorrect,reason:l.reason,result_artifact_extraction:{status:o.rawAnalysis?.fallback?"fallback_final_result":o.status,confidence:o.confidence,reason:o.reason,source_refs:o.sourceRefs,fallback_source:o.rawAnalysis?.fallback?"execution.finalResult":null}},await T(ey,eg,l),ea=M(s,{retrying:!1,exhausted:!1,now:i}),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:ee,resultEvaluationError:null,resultEvaluationRetry:ea})}}),et=null,eo=null})(),6e5,new Q("result-timeout",`结果评测超过 600s 未完成`))}catch(a){console.warn(`[trajectory-eval] result-based evaluation attempt ${s}/5 failed:`,a);let e=a.message||String(a);s<5&&(0,p.isRetryableResultEvaluationFailure)(a)?(er=V[Math.max(0,s-1)]??V[V.length-1]??8e3,en=s,ea=M(s,{retrying:!0,exhausted:!1,lastError:e,delayMs:er,now:i}),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:"running",errorMessage:null,rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:null,resultEvaluationError:null,resultEvaluationRetry:ea})}})):(et=`结果评测失败：${e}`,eo=(0,_.buildEvaluationDiagnostic)(a instanceof Q&&("result-timeout"===a.stage||"result-artifact-timeout"===a.stage)?"evaluator_timeout":"result_evaluator_failed",{details:et,reason:et}),ea=M(s,{retrying:!1,exhausted:!0,lastError:e,now:i}),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:ee,resultEvaluationError:et,resultEvaluationRetry:ea,diagnostic:eo})}}))}}if(null!=er)return void(d=er,h=en,setTimeout(()=>{P(e,t).catch(e=>{console.error(`[trajectory-eval] scheduled result retry failed row=${t} attempt=${h+1}/5:`,e)})},d));if(Y){if(em&&!ei){let t=String(ey?.finalResult||"").trim();if(eS.length>0)try{let a=await j((0,y.extractTaskResultArtifact)({userTask:eC,interactions:eS,fallbackOutput:t,user:e}),6e5,new Q("result-artifact-timeout",`实际输出提取超过 600s 未完成`));es=a.rawAnalysis,ei=String(a.outputForEvaluation||t||"").trim()}catch(e){console.warn("[trajectory-eval] custom evaluator output extraction failed:",e),ei=t}else ei=t}let t=em?ei:"",a=ef&&eF?.text||"",r=await Promise.all(Z.map(r=>(0,v.runCustomLlmEvaluator)(e,r,{caseInput:eC,expectedOutput:ep?eE.expectedOutput:"",actualOutput:t,traceText:a},F,D).catch(e=>({evaluatorId:r,evaluatorName:r,score:null,reason:"",rawResponse:"",model:"",durationMs:0,error:`运行异常：${e?.message||String(e)}`})))),n={};for(let e of r)n[e.evaluatorId]={evaluatorId:e.evaluatorId,evaluatorName:e.evaluatorName,score:e.score,reason:e.reason,model:e.model,durationMs:e.durationMs,error:e.error},"number"==typeof e.score&&eu.push(e.score);el=n}if(!H){let a=eu.length>0?eu.reduce((e,t)=>e+t,0)/eu.length:null,r=ee?.evaluatorSessionId,s="string"==typeof r?r.trim():"",i=G&&(!ee||!s),o=i?"failed":"done",l=i?et||"结果评测未产出有效结果或评测 session":null;i&&!eo&&(eo=(0,_.buildEvaluationDiagnostic)("result_evaluator_failed",{details:l||void 0,reason:l||void 0})),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:o,trajectoryScore:null,dimensionScoresJson:null,deviationStepsJson:JSON.stringify([]),rootCauseStep:null,reasonText:null,rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:ee,resultEvaluationError:et,resultEvaluationRetry:ea,customEvaluations:el,customVariableNeeds:Array.from(ed),customEvaluationScore:a,...eo?{diagnostic:eo}:{}}),errorMessage:l}}),"done"===o&&await U(e,t,ey,eS);return}let eW={caseId:eE.id||I.caseId||"",caseInput:eE.input,referenceOutput:eE.expectedOutput,referenceTrajectory:eJ,referenceKeyActionsText:eO,actualExtractedStepsText:ej,referenceKeyActions:eM,actualExtractedSteps:eT,skillContext:await (0,f.loadTaskCompletionSkillContext)(ey,eS,e),comparisonMode:e$,evaluationFocus:eL,actualInteractions:eS,taskId:eg||void 0,executionId:I.executionId||ey?.id||void 0};try{E=await Promise.race([(0,l.evaluateTrajectoryViaOpencode)(eW,e,F,D),new Promise((e,t)=>setTimeout(()=>t(new Q("timeout",`单条评测超过 600s 未完成`)),6e5))])}catch(e){if(e instanceof Q)throw e;if(e instanceof l.TrajectoryEvalConfigError)throw new Q("config",e.message,e);throw new Q("llm-or-agent",e.message||String(e),e)}try{let e={...E.rawAnalysis&&"object"==typeof E.rawAnalysis?E.rawAnalysis:{},...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:ee,resultEvaluationError:et,resultEvaluationRetry:ea,customEvaluations:el,customVariableNeeds:Array.from(ed),customEvaluationScore:eu.length>0?eu.reduce((e,t)=>e+t,0)/eu.length:null,...eo?{diagnostic:eo}:{}};await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:"done",trajectoryScore:E.trajectoryScore,dimensionScoresJson:JSON.stringify(E.dimensionScores),deviationStepsJson:JSON.stringify(E.deviationSteps),rootCauseStep:E.rootCauseStep||null,reasonText:E.reasonText||null,rawAnalysisJson:JSON.stringify(e),errorMessage:null}})}catch(e){throw new Q("persist",`写入评测结果失败: ${e.message}`,e)}await U(e,t,ey,eS)}finally{if(a&&r)try{(0,c.finish)(e,a,r)}catch(t){console.warn(`[trajectory-eval] finishEvalTask failed user=${e} task=${a}: ${t?.message}`)}}}async function U(e,t,a,r){try{let s=await n.prismaRaw.trajectoryEvalResult.findUnique({where:{id:t},select:{id:!0,taskId:!0,evaluatorRunId:!0,deviationStepsJson:!0,rootCauseStep:!0,reasonText:!0,rawAnalysisJson:!0}});if(!s)return;let i=(0,f.getPrimaryExecutionSkillTargets)(a,r).map(e=>({name:e.skill,version:e.version}));if(0===i.length)return;let o=await (0,d.deriveAndPersistOptPoints)({user:e,taskId:s.taskId,runId:s.evaluatorRunId,trajectoryRow:s,skills:i});o>0&&console.log(`[trajectory-eval]   • derived ${o} skill optimization point(s) for row=${t}`)}catch(e){console.warn("[trajectory-eval] derive opt points failed (non-fatal):",e.message)}}async function q(e,t,a){let r=Date.now();console.log(`[trajectory-eval] run ${t} started: ${a.length} traces, concurrency=3`);let s=[...a],i=0,o=0,l=0;await new Promise(r=>{let u=()=>{for((0,m.getTrajectoryEvalRunSignal)(t)?.aborted&&(s.length=0);i<3&&s.length>0;){let c=s.shift();i++,P(e,c).then(()=>{o++,console.log(`[trajectory-eval]   ✓ ${c} done (${o}+${l}/${a.length})`)}).catch(async e=>{l++;let t=e instanceof Q?e.stage:"unknown",a=e?.message||"unknown error",r=function(e){let t=e instanceof Q?e.stage:"unknown",a=(e instanceof Error?e.message:String(e||"unknown error")).replace(/^\[[^\]]+\]\s*/,"").replace(p.NO_EVALUABLE_CASE_PREFIX,"").trim();if("trace-parse"===t)return(0,_.buildEvaluationDiagnostic)("trace_parse_failed",{details:a,reason:a});if("trace-empty"===t)return(0,_.buildEvaluationDiagnostic)("trace_data_empty",{details:a,reason:a});if("semantic-match-llm"===t||"dataset-auto-match"===t)return(0,_.buildEvaluationDiagnostic)("case_match_failed",{details:a,internal:{stage:"semantic_match",reason:a}});if("result-artifact-timeout"===t||"result-timeout"===t||"timeout"===t)return(0,_.buildEvaluationDiagnostic)("evaluator_timeout",{details:a,reason:a});if("persist"===t)return(0,_.buildEvaluationDiagnostic)("persist_failed",{details:a,reason:a});if("llm-or-agent"===t||"config"===t)return(0,_.buildEvaluationDiagnostic)("evaluator_output_invalid",{details:a,reason:a});if("no-evaluable-case"===t)return a.includes("没有实际输入")?(0,_.buildEvaluationDiagnostic)("trace_missing_input",{details:a,reason:a}):a.includes("评估数据集不存在")?(0,_.buildEvaluationDiagnostic)("dataset_missing",{details:a,reason:a}):a.includes("未找到可用于评测的数据集")||a.includes("未找到可用于结果评测的数据集")||a.includes("未找到轨迹评测数据集")?(0,_.buildEvaluationDiagnostic)("dataset_empty",{details:a,reason:a}):a.includes("reference_output")?(0,_.buildEvaluationDiagnostic)("custom_reference_missing",{details:a,reason:a}):a.includes("expectedOutput")||a.includes("预期结果")?(0,_.buildEvaluationDiagnostic)("case_missing_expected_output",{details:a,reason:a}):a.includes("动态分析失败")?(0,_.buildEvaluationDiagnostic)("trace_step_analysis_failed",{details:a,reason:a}):a.includes("无法提取可评测步骤")||a.includes("提取出实际执行步骤")||a.includes("无可提取执行步骤")?(0,_.buildEvaluationDiagnostic)("trace_no_steps",{details:a,reason:a}):(0,_.buildEvaluationDiagnostic)("case_not_matched",{details:a,reason:a});return(0,_.buildEvaluationDiagnostic)("evaluator_recovered_failed",{details:a,reason:a})}(e);console.error(`[trajectory-eval]   ✗ ${c} failed [${t}]: ${a}`);let s=await n.prismaRaw.trajectoryEvalResult.findUnique({where:{id:c},select:{rawAnalysisJson:!0}}).catch(()=>null);await n.prismaRaw.trajectoryEvalResult.update({where:{id:c},data:{status:"failed",errorMessage:a,rawAnalysisJson:function(e,t,a={}){return JSON.stringify({...N(e),...a,diagnostic:t})}(s?.rawAnalysisJson,r)}}).catch(e=>console.error(`[trajectory-eval] failed to mark ${c} as failed: ${e.message}`))}).finally(()=>{i--,(0,m.getTrajectoryEvalRunSignal)(t)?.aborted&&(s.length=0),0===s.length&&0===i?r():u()})}};u()});try{let e=await n.prismaRaw.trajectoryEvalResult.updateMany({where:{id:{in:a},evaluatorRunId:t,updatedAt:{lt:new Date(r)},status:{in:["pending","running"]}},data:{status:"failed",errorMessage:"[recover] 进程异常或评测器未返回，已自动标 failed"}});e.count>0&&(console.warn(`[trajectory-eval] run ${t} recovered ${e.count} stuck rows.`),l+=e.count)}catch(e){console.error(`[trajectory-eval] failed to recover stuck rows: ${e.message}`)}let u=((Date.now()-r)/1e3).toFixed(1);console.log(`[trajectory-eval] run ${t} finished: ${o} done, ${l} failed, ${u}s`)}e.s(["PATCH",()=>F,"POST",()=>J,"dynamic",0,"force-dynamic","runEvaluations",()=>q]),a()}catch(e){a(e)}},!1),656687,e=>e.a(async(t,a)=>{try{var r=e.i(747909),n=e.i(174017),s=e.i(996250),i=e.i(759756),o=e.i(561916),l=e.i(174677),u=e.i(869741),c=e.i(316795),d=e.i(487718),p=e.i(995169),m=e.i(47587),f=e.i(666012),g=e.i(570101),y=e.i(626937),h=e.i(10372),w=e.i(193695);e.i(52474);var _=e.i(600220),v=e.i(600461),A=t([v]);[v]=A.then?(await A)():A;let x=new r.AppRouteRouteModule({definition:{kind:n.RouteKind.APP_ROUTE,page:"/api/eval/trajectory/run/route",pathname:"/api/eval/trajectory/run",filename:"route",bundlePath:""},distDir:".next",relativeProjectDir:"",resolvedPagePath:"[project]/src/app/api/eval/trajectory/run/route.ts",nextConfigOutput:"standalone",userland:v}),{workAsyncStorage:k,workUnitAsyncStorage:b,serverHooks:R}=x;function S(){return(0,s.patchFetch)({workAsyncStorage:k,workUnitAsyncStorage:b})}async function E(e,t,a){x.isDev&&(0,i.addRequestMeta)(e,"devRequestTimingInternalsEnd",process.hrtime.bigint());let r="/api/eval/trajectory/run/route";r=r.replace(/\/index$/,"")||"/";let s=await x.prepare(e,t,{srcPage:r,multiZoneDraftMode:!1});if(!s)return t.statusCode=400,t.end("Bad Request"),null==a.waitUntil||a.waitUntil.call(a,Promise.resolve()),null;let{buildId:v,params:A,nextConfig:S,parsedUrl:E,isDraftMode:k,prerenderManifest:b,routerServerContext:R,isOnDemandRevalidate:I,revalidateOnlyGenerated:C,resolvedPathname:N,clientReferenceManifest:$,serverActionsManifest:O}=s,j=(0,u.normalizeAppPath)(r),M=!!(b.dynamicRoutes[j]||b.routes[N]),T=async()=>((null==R?void 0:R.render404)?await R.render404(e,t,E,!1):t.end("This page could not be found"),null);if(M&&!k){let e=!!b.routes[N],t=b.dynamicRoutes[j];if(t&&!1===t.fallback&&!e){if(S.experimental.adapterPath)return await T();throw new w.NoFallbackError}}let L=null;!M||x.isDev||k||(L=N,L="/index"===L?"/":L);let J=!0===x.isDev||!M,F=M&&!J;O&&$&&(0,l.setManifestsSingleton)({page:r,clientReferenceManifest:$,serverActionsManifest:O});let P=e.method||"GET",D=(0,o.getTracer)(),U=D.getActiveScopeSpan(),q={params:A,prerenderManifest:b,renderOpts:{experimental:{authInterrupts:!!S.experimental.authInterrupts},cacheComponents:!!S.cacheComponents,supportsDynamicResponse:J,incrementalCache:(0,i.getRequestMeta)(e,"incrementalCache"),cacheLifeProfiles:S.cacheLife,waitUntil:a.waitUntil,onClose:e=>{t.on("close",e)},onAfterTaskError:void 0,onInstrumentationRequestError:(t,a,r,n)=>x.onRequestError(e,t,r,n,R)},sharedContext:{buildId:v}},z=new c.NodeNextRequest(e),W=new c.NodeNextResponse(t),B=d.NextRequestAdapter.fromNodeNextRequest(z,(0,d.signalFromNodeResponse)(t));try{let s=async e=>x.handle(B,q).finally(()=>{if(!e)return;e.setAttributes({"http.status_code":t.statusCode,"next.rsc":!1});let a=D.getRootSpanAttributes();if(!a)return;if(a.get("next.span_type")!==p.BaseServerSpan.handleRequest)return void console.warn(`Unexpected root span type '${a.get("next.span_type")}'. Please report this Next.js issue https://github.com/vercel/next.js`);let n=a.get("next.route");if(n){let t=`${P} ${n}`;e.setAttributes({"next.route":n,"http.route":n,"next.span_name":t}),e.updateName(t)}else e.updateName(`${P} ${r}`)}),l=!!(0,i.getRequestMeta)(e,"minimalMode"),u=async i=>{var o,u;let c=async({previousCacheEntry:n})=>{try{if(!l&&I&&C&&!n)return t.statusCode=404,t.setHeader("x-nextjs-cache","REVALIDATED"),t.end("This page could not be found"),null;let r=await s(i);e.fetchMetrics=q.renderOpts.fetchMetrics;let o=q.renderOpts.pendingWaitUntil;o&&a.waitUntil&&(a.waitUntil(o),o=void 0);let u=q.renderOpts.collectedTags;if(!M)return await (0,f.sendResponse)(z,W,r,q.renderOpts.pendingWaitUntil),null;{let e=await r.blob(),t=(0,g.toNodeOutgoingHttpHeaders)(r.headers);u&&(t[h.NEXT_CACHE_TAGS_HEADER]=u),!t["content-type"]&&e.type&&(t["content-type"]=e.type);let a=void 0!==q.renderOpts.collectedRevalidate&&!(q.renderOpts.collectedRevalidate>=h.INFINITE_CACHE)&&q.renderOpts.collectedRevalidate,n=void 0===q.renderOpts.collectedExpire||q.renderOpts.collectedExpire>=h.INFINITE_CACHE?void 0:q.renderOpts.collectedExpire;return{value:{kind:_.CachedRouteKind.APP_ROUTE,status:r.status,body:Buffer.from(await e.arrayBuffer()),headers:t},cacheControl:{revalidate:a,expire:n}}}}catch(t){throw(null==n?void 0:n.isStale)&&await x.onRequestError(e,t,{routerKind:"App Router",routePath:r,routeType:"route",revalidateReason:(0,m.getRevalidateReason)({isStaticGeneration:F,isOnDemandRevalidate:I})},!1,R),t}},d=await x.handleResponse({req:e,nextConfig:S,cacheKey:L,routeKind:n.RouteKind.APP_ROUTE,isFallback:!1,prerenderManifest:b,isRoutePPREnabled:!1,isOnDemandRevalidate:I,revalidateOnlyGenerated:C,responseGenerator:c,waitUntil:a.waitUntil,isMinimalMode:l});if(!M)return null;if((null==d||null==(o=d.value)?void 0:o.kind)!==_.CachedRouteKind.APP_ROUTE)throw Object.defineProperty(Error(`Invariant: app-route received invalid cache entry ${null==d||null==(u=d.value)?void 0:u.kind}`),"__NEXT_ERROR_CODE",{value:"E701",enumerable:!1,configurable:!0});l||t.setHeader("x-nextjs-cache",I?"REVALIDATED":d.isMiss?"MISS":d.isStale?"STALE":"HIT"),k&&t.setHeader("Cache-Control","private, no-cache, no-store, max-age=0, must-revalidate");let p=(0,g.fromNodeOutgoingHttpHeaders)(d.value.headers);return l&&M||p.delete(h.NEXT_CACHE_TAGS_HEADER),!d.cacheControl||t.getHeader("Cache-Control")||p.get("Cache-Control")||p.set("Cache-Control",(0,y.getCacheControlHeader)(d.cacheControl)),await (0,f.sendResponse)(z,W,new Response(d.value.body,{headers:p,status:d.value.status||200})),null};U?await u(U):await D.withPropagatedContext(e.headers,()=>D.trace(p.BaseServerSpan.handleRequest,{spanName:`${P} ${r}`,kind:o.SpanKind.SERVER,attributes:{"http.method":P,"http.target":e.url}},u))}catch(t){if(t instanceof w.NoFallbackError||await x.onRequestError(e,t,{routerKind:"App Router",routePath:j,routeType:"route",revalidateReason:(0,m.getRevalidateReason)({isStaticGeneration:F,isOnDemandRevalidate:I})},!1,R),M)throw t;return await (0,f.sendResponse)(z,W,new Response(null,{status:500})),null}}e.s(["handler",()=>E,"patchFetch",()=>S,"routeModule",()=>x,"serverHooks",()=>R,"workAsyncStorage",()=>k,"workUnitAsyncStorage",()=>b]),a()}catch(e){a(e)}},!1)];
+${JSON.stringify(r,null,2)}`)),$=u(N);if(!$)return n("第二轮输出提取模型未返回合法 JSON。",{pass:"second",first:f,raw_first_response:c,raw_second_response:N,followup_anchor_llm_call_id:h,followup_event_ids:I.map(e=>e.id),threshold:t});let O=k.parse($),j=w(C,O.selected_source_id),M={pass:"second",first:f,second:O,raw_first_response:c,raw_second_response:N,followup_anchor_llm_call_id:h,followup_event_ids:I.map(e=>e.id),source_candidate_ids:C.map(e=>e.id),threshold:t};if(O.has_task_output&&j)return v(j,O.confidence,O.reason,O.selected_source_id,{...M,accepted_below_threshold:O.confidence<t});let T=O.selected_source_id&&!_(C,O.selected_source_id)?`第二轮返回的 selected_source_id=${O.selected_source_id} 不在候选列表中。`:"";return n(T||O.reason||`二次输出定位置信度 ${O.confidence} 低于阈值 ${t}。`,M)}catch(e){return n(`真实输出提取 LLM 调用失败：${e.message||String(e)}`,{error:e.message||String(e)})}}e.s(["extractTaskResultArtifact",()=>S]),a()}catch(e){a(e)}},!1),830484,e=>{"use strict";let t={trace_missing_input:{errorCategory:"trace",shortStatus:"Trace 缺输入",userMessage:"Trace 缺少实际输入，无法匹配评测集 case。",reason:"系统没有从本次执行记录中提取到可用于匹配 case 的任务输入。",nextAction:"检查 trace 采集是否包含用户任务输入；必要时重新采集或重新执行该任务。",retryable:!1,stage:"case-match"},trace_missing_output:{errorCategory:"trace",shortStatus:"Trace 缺输出",userMessage:"Trace 缺少实际输出，无法做任务完成度评测。",reason:"本次 trace 没有提取到 Agent 最终输出，也没有可用的 execution.finalResult fallback。",nextAction:"检查 Agent 是否产出最终回复；确认 trace 采集包含 assistant final message；必要时重新跑该 case。",retryable:!1,stage:"result-artifact"},task_output_extraction_failed:{errorCategory:"trace",shortStatus:"任务输出提取失败",userMessage:"无法从 trace 中提取任务输出。",reason:"任务输出提取流程没有定位到可用于任务完成度评测的最终输出，且没有 fallback 可用。",nextAction:"检查 trace 中是否包含最终回复或关键工具输出；必要时重试评测或重新采集 trace。",retryable:!0,stage:"result-artifact"},trace_data_empty:{errorCategory:"trace",shortStatus:"Trace 数据为空",userMessage:"Trace 数据为空，无法评测。",reason:"系统没有找到可用的 session interactions，也没有足够的 query/finalResult 生成兜底交互。",nextAction:"检查 trace 采集、上传或执行记录入库是否完整。",retryable:!1,stage:"trace-load"},trace_parse_failed:{errorCategory:"trace",shortStatus:"Trace 解析失败",userMessage:"Trace 数据解析失败，无法评测。",reason:"Session.interactions 不是合法 JSON 或结构不兼容。",nextAction:"检查 trace 采集数据格式，必要时重新上传或修复采集链路。",retryable:!1,stage:"trace-parse"},case_not_matched:{errorCategory:"case",shortStatus:"无匹配 case",userMessage:"Trace 输入没有匹配到可用于评测的 case。",reason:"本次任务输入与当前评测集中的 case.input 没有达到可用匹配。",nextAction:"补充对应 case，或确认本次 trace 的任务输入与评测集 case.input 是否一致。",retryable:!1,stage:"case-match"},case_match_failed:{errorCategory:"case",shortStatus:"匹配 case 失败",userMessage:"无法根据 trace 输入自动定位可用于评测的 case。",reason:"trace 输入没有精确匹配到评测集 case，后台自动匹配也没有得到可用结果。",nextAction:"检查本次 trace 的任务输入是否与评测集 case.input 对齐；必要时补充 case 或重新选择评测集。",retryable:!0,stage:"case-match"},dataset_missing:{errorCategory:"dataset",shortStatus:"评测集不存在",userMessage:"评测集不存在，无法匹配可评测 case。",reason:"评测结果关联的数据集已经不存在，或当前用户不可访问。",nextAction:"重新选择评测集后再发起评测。",retryable:!1,stage:"dataset-load"},dataset_empty:{errorCategory:"dataset",shortStatus:"评测集为空",userMessage:"评测集没有可评测 case。",reason:"当前匹配范围内没有可用于评测的 case。",nextAction:"补充评测集 case 后重新评测。",retryable:!1,stage:"dataset-load"},case_missing_expected_output:{errorCategory:"case",shortStatus:"Case 缺预期结果",userMessage:"Case 缺少预期结果，无法做任务完成度评测。",reason:"已匹配到 case，但 expectedOutput 为空。",nextAction:"为该 case 补充 expectedOutput 后重新评测。",retryable:!1,stage:"case-match"},custom_reference_missing:{errorCategory:"case",shortStatus:"缺 reference_output",userMessage:"自定义评估器缺少 reference_output。",reason:"自定义评估器 prompt 引用了 reference_output，但没有匹配到带预期结果的 case。",nextAction:"补充评测集 expectedOutput，或调整自定义评估器 prompt。",retryable:!1,stage:"custom-evaluator"},trace_step_analysis_failed:{errorCategory:"model",shortStatus:"步骤抽取失败",userMessage:"Trace 步骤动态分析失败。",reason:"系统尝试从 trace 中抽取实际执行步骤时失败，通常与模型配置或调用异常有关。",nextAction:"检查评测模型配置和服务状态后重试。",retryable:!0,stage:"trace-step-analysis"},trace_no_steps:{errorCategory:"trace",shortStatus:"Trace 无步骤",userMessage:"Trace 无可提取执行步骤，无法做轨迹评测。",reason:"系统没有从本次 trace 中提取到可用于轨迹评测的实际执行步骤。",nextAction:"检查 trace 质量、采集粒度或执行记录是否完整。",retryable:!1,stage:"trace-step-analysis"},evaluator_timeout:{errorCategory:"evaluator",shortStatus:"评测超时",userMessage:"评测器运行超时。",reason:"评测器在限定时间内没有返回结果。",nextAction:"重试评测；如果持续出现，请检查模型服务响应时间。",retryable:!0,stage:"evaluator"},evaluator_output_invalid:{errorCategory:"evaluator",shortStatus:"评测器输出异常",userMessage:"评测器未返回有效结构化结果。",reason:"评测器输出无法解析为系统需要的 JSON/评分结构。",nextAction:"重试评测；如果持续出现，请检查评测器模型输出。",retryable:!0,stage:"evaluator"},result_evaluator_failed:{errorCategory:"evaluator",shortStatus:"结果评测失败",userMessage:"任务完成度评测失败。",reason:"任务完成度评测器没有产出可用评分或评测 session。",nextAction:"重试评测；如果持续出现，请检查模型配置和评测器日志。",retryable:!0,stage:"result-evaluator"},persist_failed:{errorCategory:"system",shortStatus:"保存失败",userMessage:"评测结果保存失败。",reason:"评测已执行，但结果写入数据库时失败。",nextAction:"检查数据库或服务状态后重试。",retryable:!0,stage:"persist"},evaluator_recovered_failed:{errorCategory:"system",shortStatus:"评测进程异常",userMessage:"评测进程异常中断。",reason:"后台恢复逻辑发现该评测长期停留在 pending/running 状态。",nextAction:"重试评测，并查看服务日志确认是否有进程异常。",retryable:!0,stage:"recover"}};function a(e,r={}){let n=t[e]??{errorCategory:"system",shortStatus:"评测失败",userMessage:"评测执行失败。",reason:"评测执行过程中出现未分类异常。",nextAction:"重试评测；如果持续出现，请查看服务日志。",retryable:!0,stage:"unknown"};return{errorCode:e,errorCategory:n.errorCategory,shortStatus:n.shortStatus,userMessage:r.userMessage||n.userMessage,reason:r.reason||n.reason,nextAction:r.nextAction||n.nextAction,retryable:"boolean"==typeof r.retryable?r.retryable:n.retryable,stage:r.stage||n.stage,...r.details?{details:r.details}:{},...r.internal?{internal:r.internal}:{}}}e.s(["buildEvaluationDiagnostic",()=>a])},652192,e=>{"use strict";function t(e){switch(e.status){case"ok":return{state:"ok",code:"ok",message:"skill 关键步骤与 trace 实际步骤已完整对比"};case"no-skill-targets":return{state:"not-applicable",code:"no-skill-targets",message:"trace 未关联任何 skill，本次评测不做 skill 归因"};case"missing-skill":return{state:"degraded",code:"missing-skill",message:`trace 关联的 skill 在 Skills Hub 中找不到：${e.missingSkills.join("、")}`};case"missing-parsed-flow":return{state:"degraded",code:"missing-parsed-flow",message:`skill 尚未生成可用的解析流程，无法提取关键步骤：${e.missingSkills.join("、")}`};case"no-key-actions":return{state:"degraded",code:"no-key-actions",message:"skill flow 解析成功但未识别出任何关键步骤"};case"dynamic-analysis-failed":return{state:"degraded",code:"dynamic-analysis-failed",message:"trace 步骤动态分析失败（通常是 LLM 未配置或调用失败）"};case"no-extracted-steps":return{state:"degraded",code:"no-extracted-steps",message:"未能从 trace 中提取出实际执行步骤"}}}e.s(["buildSkillAttributionStatus",()=>t])},600461,e=>e.a(async(t,a)=>{try{let z;var r=e.i(89171),n=e.i(854763),s=e.i(948594),i=e.i(833821),o=e.i(511284),l=e.i(841212),u=e.i(664503),c=e.i(826917),d=e.i(418494),p=e.i(556602),m=e.i(926307),f=e.i(190014),g=e.i(508439),y=e.i(670497),h=e.i(794538),w=e.i(374995),_=e.i(830484),v=e.i(228356),A=e.i(240317),S=e.i(652192),E=e.i(450579),k=e.i(837926),x=t([n,s,i,l,u,d,f,g,y,v,E]);[n,s,i,l,u,d,f,g,y,v,E]=x.then?(await x)():x;let W=new Set(["trace-quality-evaluator","preset-agent-trace-quality","preset-agent-task-completion"]),B="preset-agent-trace-quality",K="preset-agent-task-completion",V=[1500,3e3,5e3,8e3],H={[B]:"Agent 轨迹质量",[K]:"Agent 任务完成度"},X=/\{\{\s*(input|output|reference_output|trajectory)\s*\}\}/g;function b(e){return String(e||"").trim().replace(/\s+/g," ")}function R(e){let t=(Array.isArray(e)?e:null==e?[]:[e]).map(e=>String(e||"").trim()).map(e=>"trace-quality-evaluator"===e?B:e).filter(e=>W.has(e)||(0,v.isCustomEvaluatorId)(e));return Array.from(new Set(t))}function I(e,t={}){return{selectedEvaluators:e,selectedEvaluatorNames:e.map(e=>{let a=H[e];if(a)return a;let r=t.customNameResolver?.(e);return r||e}),...t.autoWatch?{autoWatch:!0}:{},...t.watchedAgent?{watchedAgent:t.watchedAgent}:{},...t.autoWatch&&t.autoWatchEnabledAt?{autoWatchEnabledAt:t.autoWatchEnabledAt}:{}}}function C(e){let t=e?(0,h.parseLooseJson)(e):null,a=R(t?.selectedEvaluators),r=Array.isArray(t?.selectedEvaluatorNames)?t.selectedEvaluatorNames.filter(e=>"string"==typeof e&&!!e.trim()):[],n=t?.autoWatch===!0,s="string"==typeof t?.watchedAgent?t.watchedAgent.trim():"",i=(0,E.normalizeAutoWatchEnabledAt)(t?.autoWatchEnabledAt);return a.length>0?r.length===a.length?{selectedEvaluators:a,selectedEvaluatorNames:r,...n?{autoWatch:!0}:{},...s?{watchedAgent:s}:{},...n&&i?{autoWatchEnabledAt:i}:{}}:I(a,{autoWatch:n,watchedAgent:s,autoWatchEnabledAt:i}):I([B],{autoWatch:n,watchedAgent:s,autoWatchEnabledAt:i})}function N(e){let t=(0,h.parseLooseJson)(e||"");return t&&"object"==typeof t?t:{}}function $(e){if(!e||"object"!=typeof e||Array.isArray(e))return null;let t=String(e.source||"").trim(),a=String(e.grayscaleTaskId||"").trim(),r=String(e.caseId||"").trim(),n=String(e.side||"").trim(),s=String(e.executionTraceId||"").trim(),i=String(e.evaluationClaimId||"").trim(),o="number"==typeof e.runIndex&&Number.isFinite(e.runIndex)?Math.floor(e.runIndex):NaN,l="number"==typeof e.roundIndex&&Number.isFinite(e.roundIndex)?Math.floor(e.roundIndex):NaN,u="number"==typeof e.evaluationAttempt&&Number.isFinite(e.evaluationAttempt)?Math.floor(e.evaluationAttempt):NaN;return"grayscale-ab"===t&&a&&r&&("a"===n||"b"===n)&&s&&i&&Number.isFinite(o)&&Number.isFinite(l)&&Number.isFinite(u)?{source:"grayscale-ab",grayscaleTaskId:a,caseId:r,side:n,runIndex:o,roundIndex:l,executionTraceId:s,evaluationClaimId:i,evaluationAttempt:u}:null}async function O(e,t,a,r,n,s,i,o,l,c,d){let p=await (0,u.evaluateTaskCompletionViaOpencode)({caseInput:e,expectedOutput:t,actualOutput:a,precomputedRootCauses:r,precomputedRootCauseSource:n,traceSummaryText:s,skillContext:i,skillAttributionMode:o},l,c,d);return{isCorrect:p.isCorrect,score:p.score,reason:p.reason,rawAnalysis:p.rawAnalysis}}async function j(e,t,a){return Promise.race([e,new Promise((e,r)=>setTimeout(()=>r(a),t))])}function M(e,t){let a=t.now??new Date;return{attemptCount:e,maxAttempts:5,retrying:t.retrying,exhausted:t.exhausted,lastAttemptAt:a.toISOString(),...t.lastError?{lastError:t.lastError}:{},...t.retrying&&t.delayMs?{nextRetryAt:new Date(a.getTime()+t.delayMs).toISOString()}:{}}}async function T(e,t,a){let r=a.rawAnalysis&&"object"==typeof a.rawAnalysis?a.rawAnalysis:null,s=r?.key_point_findings;if(!("number"==typeof a.score&&Number.isFinite(a.score)&&String(a.reason||"").trim()&&Array.isArray(s)))return void console.warn("[trajectory-eval] skip persisting incomplete result-evaluation payload to execution");let i={isAnswerCorrect:a.isCorrect,answerScore:a.score,judgmentReason:a.reason};e?.id?await n.prismaRaw.execution.update({where:{id:e.id},data:i}):t&&await n.prismaRaw.execution.updateMany({where:{taskId:t},data:i})}async function L(e,t,a={}){let r=b(t);if(!r)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} trace 没有实际输入，无法匹配评估数据集 case`);let n=!0===a.requireExpectedOutput,s=!0===a.includeAllDatasetKinds,o=Array.isArray(a.allowedDatasetIds)?a.allowedDatasetIds.map(e=>String(e).trim()).filter(Boolean):[],l=(await (0,i.readAllAgentDatasets)()).filter(t=>t.user===e).filter(e=>0===o.length||o.includes(e.id)).filter(e=>s||n||"trajectory"===e.datasetKind);if(0===l.length)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} ${s?"未找到可用于评测的数据集":n?"未找到可用于结果评测的数据集":"未找到轨迹评测数据集"}`);for(let e of l){let t=e.cases.find(e=>b(e.input)===r&&(!n||!!b(e.expectedOutput)));if(t)return{dataset:e,caseEntry:t}}let u=[],c=new Map;for(let e of l)for(let t of e.cases){if(n&&!b(t.expectedOutput))continue;let a=`${e.id}::${t.id}`;u.push({id:a,input:t.input}),c.set(a,{dataset:e,caseEntry:t})}let d=await (0,g.findBestSemanticCaseMatch)(u,t,{user:e,requireModelAvailable:!0});if(d.error)throw new Q("semantic-match-llm",`${p.NO_EVALUABLE_CASE_PREFIX} 语义匹配调用评测模型失败：${d.error}`);if(d.caseId){let e=c.get(d.caseId);if(e)return e}throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} trace 实际输入未匹配到${s?"评测数据集中的":n?"带预期结果的":"轨迹评测数据集中的"} case 输入`)}async function J(e){try{let t=await e.json(),a=String(t.user||"").trim(),o=String(t.evaluatorRunId||t.runId||"").trim(),l=$(t.grayscaleBinding),u=Array.isArray(t.evaluators)&&t.evaluators.length>0?t.evaluators:t.evaluator,c=R(u),d=!0===t.autoWatch,p=(0,E.normalizeAutoWatchAgent)(t.watchedAgent||t.agent||""),f=p,g=new Date().toISOString(),y=!0===t.placeholderOnly,_=String(t.taskScope||"").trim(),A=String(t.skillName||"").trim(),S=t.skillVersion,x=null==S||""===S?null:Number(S);if(!a)return r.NextResponse.json({error:"user is required"},{status:400});if(d&&p&&(0,k.isEvaluatorAgentName)(p))return r.NextResponse.json({error:"autoWatch requires a non-evaluator execution agent"},{status:400});if("skill-case-analysis"===_){if(!A)return r.NextResponse.json({error:"skillName is required for skill-case-analysis task"},{status:400});if(null==x||!Number.isFinite(x))return r.NextResponse.json({error:"skillVersion is required for skill-case-analysis task"},{status:400})}let b=c.filter(v.isCustomEvaluatorId),N=new Set,O=new Map;if(b.length>0){let e=await (0,v.listCustomEvaluatorIds)(a),t=[];for(let a of b)e.has(a)?N.add(a):t.push(a);if(t.length>0)return r.NextResponse.json({error:`unknown custom evaluators: ${t.join(", ")}`},{status:400});for(let e of N){let t=await (0,v.loadCustomEvaluator)(a,e);t&&O.set(e,t.name)}}let j=c.filter(e=>(0,v.isCustomEvaluatorId)(e)?N.has(e):W.has(e)),M=I(j.length>0?j:[B],{autoWatch:d,watchedAgent:f,customNameResolver:e=>O.get(e)});if(u&&0===j.length||0===M.selectedEvaluators.length)return r.NextResponse.json({error:"unsupported evaluators"},{status:400});let T=Array.isArray(t.taskIds)?t.taskIds.map(e=>String(e).trim()).filter(Boolean):[],L=String(t.datasetId||"").trim(),J=Array.isArray(t.datasetIds)?t.datasetIds.map(e=>String(e).trim()).filter(Boolean):[],F=Array.isArray(t.pairs)?t.pairs:[],P=(0,w.normalizeTrajectoryTaskMeta)({title:t.taskTitle,description:t.taskDescription,scope:_,skillName:"skill-case-analysis"===_?A:void 0,skillVersion:"skill-case-analysis"===_?x:void 0}),D=new Set,U="",z="";if(o){let e=await n.prismaRaw.trajectoryEvalResult.findMany({where:{user:a,evaluatorRunId:o},orderBy:{createdAt:"asc"}});if(0===e.length)return r.NextResponse.json({error:"evaluatorRunId not found"},{status:404});let t=e[0],s=[...e].reverse().map(e=>(function(e){let t=e?(0,h.parseLooseJson)(e):null,a=R(t?.selectedEvaluators);if(0===a.length)return null;let r=Array.isArray(t?.selectedEvaluatorNames)?t.selectedEvaluatorNames.filter(e=>"string"==typeof e&&!!e.trim()):[],n=t?.autoWatch===!0,s="string"==typeof t?.watchedAgent?t.watchedAgent.trim():"",i=(0,E.normalizeAutoWatchEnabledAt)(t?.autoWatchEnabledAt);return r.length===a.length?{selectedEvaluators:a,selectedEvaluatorNames:r,...n?{autoWatch:!0}:{},...s?{watchedAgent:s}:{},...n&&i?{autoWatchEnabledAt:i}:{}}:I(a,{autoWatch:n,watchedAgent:s,autoWatchEnabledAt:i})})(e.rawAnalysisJson)).find(e=>!!e);M=j.length>0?I(j,{autoWatch:d,watchedAgent:f,customNameResolver:e=>O.get(e)}):s||C(t.rawAnalysisJson),P=(0,w.extractTrajectoryTaskMeta)(t.rawAnalysisJson,t.createdAt),L=L||t.datasetId||"",U=(0,E.normalizeAutoWatchAgent)(s?.watchedAgent||await (0,E.resolveWatchedAgentForRunRows)(a,e)),z=(0,E.normalizeAutoWatchEnabledAt)(s?.autoWatchEnabledAt),D=new Set(e.filter(e=>"failed"!==e.status).map(e=>e.taskId||e.executionId||"").filter(Boolean))}if(d&&T.length>0){let e=await (0,E.resolveSingleTrajectoryCandidateAgent)(a,T);if(e.missingTaskIds.length>0)return r.NextResponse.json({error:`autoWatch only accepts traces that appear in 发起新评测: ${e.missingTaskIds.join(", ")}`},{status:400});if(!e.agent)return r.NextResponse.json({error:`autoWatch requires a single execution agent, got: ${e.distinctAgents.join(", ")||"none"}`},{status:400});if(U&&!(0,E.autoWatchAgentNamesMatch)(U,e.agent))return r.NextResponse.json({error:`autoWatch agent mismatch: run watches ${U}, new traces belong to ${e.agent}`},{status:400});f=e.agent}else{if(d&&!o&&!f)return r.NextResponse.json({error:"watchedAgent is required when starting autoWatch without seed traces"},{status:400});d&&U&&(f=U)}M=I(j.length>0?j:M.selectedEvaluators,{autoWatch:d,watchedAgent:d?f:"",autoWatchEnabledAt:d?o&&z||g:"",customNameResolver:e=>O.get(e)});let K=o||`trun_${Date.now()}_${Math.random().toString(36).slice(2,8)}`,V=[],H=[];if(T.length>0)for(let e of T){if(D.has(e)){H.push({caseId:"",reason:`taskId ${e} already exists in this run`});continue}let t=await n.prismaRaw.trajectoryEvalResult.create({data:{user:a,evaluatorRunId:K,datasetId:"",caseId:"",executionId:null,taskId:e,status:"pending",rawAnalysisJson:JSON.stringify({...M,taskMeta:P,...l?{grayscaleBinding:l}:{},...J.length>0?{allowedDatasetIds:J}:{}})}});V.push({id:t.id,caseId:"",taskId:e})}else if((d||y)&&!o){if(L){let e=await (0,i.findAgentDataset)(a,L);if(!e)return r.NextResponse.json({error:"dataset not found"},{status:404});if("trajectory"!==e.datasetKind)return r.NextResponse.json({error:`dataset kind must be 'trajectory', got '${e.datasetKind}'`},{status:400})}let e=await n.prismaRaw.trajectoryEvalResult.create({data:{user:a,evaluatorRunId:K,datasetId:L,caseId:"",executionId:null,taskId:null,status:"pending",rawAnalysisJson:JSON.stringify({...M,taskMeta:P,...l?{grayscaleBinding:l}:{},watchPlaceholder:!0,...y?{placeholderOnly:!0}:{}})}});V.push({id:e.id,caseId:""})}else{if(!L||!(F.length>0))return r.NextResponse.json({error:"provide taskIds, datasetId+pairs, or placeholderOnly:true"},{status:400});let e=await (0,i.findAgentDataset)(a,L);if(!e)return r.NextResponse.json({error:"dataset not found"},{status:404});let t=new Map(e.cases.map(e=>[e.id,e]));for(let e of F){let r=String(e.caseId||"").trim(),s=e.executionId?String(e.executionId).trim():void 0,i=e.taskId?String(e.taskId).trim():void 0,o=i||s||"";if(!l&&o&&D.has(o)){H.push({caseId:r,reason:`${o} already exists in this run`});continue}if(!t.get(r)){H.push({caseId:r,reason:"case not found in dataset"});continue}if(!s&&!i){H.push({caseId:r,reason:"executionId or taskId is required"});continue}let u=await n.prismaRaw.trajectoryEvalResult.create({data:{user:a,evaluatorRunId:K,datasetId:L,caseId:r,executionId:s||null,taskId:i||null,status:"pending",rawAnalysisJson:JSON.stringify({...M,taskMeta:P,...l?{grayscaleBinding:l}:{}})}});V.push({id:u.id,caseId:r,executionId:s,taskId:i})}}if(0===V.length)return r.NextResponse.json({error:"no valid tasks to run",skipped:H},{status:400});Promise.all(V.map(e=>(0,s.reattributeServiceTraceOwner)((e.taskId||e.executionId||"").trim(),a).catch(()=>!1)));let X=V.filter(e=>e.taskId||e.executionId).map(e=>e.id);return X.length>0&&((0,m.registerTrajectoryEvalRun)(K,a),q(a,K,X).catch(e=>{console.error("[trajectory-eval] background run crashed:",e)}).finally(()=>(0,m.unregisterTrajectoryEvalRun)(K))),r.NextResponse.json({success:!0,evaluatorRunId:K,evaluators:M.selectedEvaluators,evaluatorNames:M.selectedEvaluatorNames,datasetId:L,created:V,skipped:H})}catch(t){console.error("trajectory/run POST error:",t);let e=t instanceof Error?t.message:"failed to start trajectory evaluation";return r.NextResponse.json({error:e},{status:500})}}async function F(e){try{let t=await e.json(),a=String(t.user||"").trim(),s=String(t.evaluatorRunId||t.runId||"").trim(),i=!0===t.autoWatch;if(!a)return r.NextResponse.json({error:"user is required"},{status:400});if(!s)return r.NextResponse.json({error:"evaluatorRunId is required"},{status:400});let o=await n.prismaRaw.trajectoryEvalResult.findMany({where:{user:a,evaluatorRunId:s},orderBy:{createdAt:"asc"}});if(0===o.length)return r.NextResponse.json({error:"evaluatorRunId not found"},{status:404});let l=C(o[0].rawAnalysisJson),u=(0,E.normalizeAutoWatchAgent)(l.watchedAgent),c=u&&!(0,k.isEvaluatorAgentName)(u)?u:await (0,E.resolveWatchedAgentForRunRows)(a,o);if(i&&!c)return r.NextResponse.json({error:"autoWatch requires existing traces from a single non-evaluator execution agent"},{status:400});let d=I(l.selectedEvaluators,{autoWatch:i,watchedAgent:c,autoWatchEnabledAt:i?new Date().toISOString():""});return await n.prismaRaw.$transaction(o.map(e=>{var t;let a;return n.prismaRaw.trajectoryEvalResult.update({where:{id:e.id},data:{rawAnalysisJson:(0,E.mergeWatchedAgentIntoRawAnalysis)((t=e.rawAnalysisJson,a={...N(t),selectedEvaluators:d.selectedEvaluators,selectedEvaluatorNames:d.selectedEvaluatorNames,watchedAgent:d.watchedAgent||"",autoWatchEnabledAt:d.autoWatchEnabledAt||""},d.autoWatch?a.autoWatch=!0:(delete a.autoWatch,delete a.autoWatchEnabledAt),d.watchedAgent||delete a.watchedAgent,d.autoWatchEnabledAt||delete a.autoWatchEnabledAt,JSON.stringify(a)),d.watchedAgent||"")}})})),r.NextResponse.json({success:!0,evaluatorRunId:s,autoWatch:i,watchedAgent:d.watchedAgent||""})}catch(t){console.error("trajectory/run PATCH error:",t);let e=t instanceof Error?t.message:"failed to update auto watch state";return r.NextResponse.json({error:e},{status:500})}}let G=(z=Number(process.env.MAX_EVAL_ROW_CONCURRENCY),Number.isFinite(z)&&z>0?Math.min(20,Math.floor(z)):4),Z=Symbol.for("@witty-insight/trajectory-eval-row-limiter");class Q extends Error{stage;cause;constructor(e,t,a){super(`[${e}] ${t}`),this.stage=e,this.cause=a,this.name="StagedEvaluationError"}}async function P(e,t){let a,r=((a=globalThis)[Z]||(a[Z]=(0,p.createSimpleAsyncLimiter)(G)),a[Z]);await r.acquire();try{await D(e,t)}finally{r.release()}}async function D(e,t){let a=null,r=null;try{var s,u,d,h;let E,k,x,R;await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:"running",errorMessage:null}});let I=await n.prismaRaw.trajectoryEvalResult.findUnique({where:{id:t}});if(!I)throw new Q("lookup",`result row ${t} disappeared`);if((0,m.getTrajectoryEvalRunSignal)(I.evaluatorRunId)?.aborted)return void await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:"failed",errorMessage:"已终止：用户中止了该批次评测"}}).catch(()=>{});I.taskId&&(a=I.taskId,r=(0,c.startOrReplace)(e,I.taskId,"rejudge").runId);let J=null;I.executionId&&(J=await n.prismaRaw.execution.findUnique({where:{id:I.executionId},select:{skill:!0,skillVersion:!0}}).catch(()=>null)),!J&&I.taskId&&(J=await n.prismaRaw.execution.findFirst({where:{taskId:I.taskId},orderBy:{timestamp:"desc"},select:{skill:!0,skillVersion:!0}}).catch(()=>null));let F=J?.skill||null,D="number"==typeof J?.skillVersion?J.skillVersion:null,q=C(I.rawAnalysisJson),z=(0,w.extractTrajectoryTaskMeta)(I.rawAnalysisJson,I.createdAt),W=(s=I.rawAnalysisJson,k=N(s),$(k.grayscaleBinding)),H=q.selectedEvaluators.includes(B),G=q.selectedEvaluators.includes(K),Z=q.selectedEvaluators.filter(v.isCustomEvaluatorId),Y=Z.length>0,ee=null,et=null,ea=null,er=null,en=0,es=null,ei="",eo=null,el=null,eu=[],ec=Y?await Promise.all(Z.map(async t=>({evaluatorId:t,bundle:await (0,v.loadCustomEvaluator)(e,t)}))):[],ed=function(e){let t=new Set;for(let a of e){let e;for(X.lastIndex=0;null!==(e=X.exec(a||""));)t.add(e[1])}return t}(ec.flatMap(e=>[e.bundle?.config.systemPrompt||"",e.bundle?.config.userPrompt||""]).filter(Boolean)),ep=ed.has("reference_output"),em=ed.has("output"),ef=ed.has("trajectory"),eg=I.taskId,ey=I.executionId?await n.prismaRaw.execution.findUnique({where:{id:I.executionId}}):null;!eg&&I.executionId&&ey?.taskId&&(eg=ey.taskId),!ey&&eg&&(ey=await n.prismaRaw.execution.findFirst({where:{taskId:eg}}));let eh=ey?.query||"",ew=await (0,g.extractRealUserInput)(eh,e),e_=ew.normalized_input.trim()||eh,ev=String(ey?.finalResult||"").trim(),eA=(x=N(I.rawAnalysisJson),Array.isArray(x?.allowedDatasetIds)?x.allowedDatasetIds.map(e=>String(e).trim()).filter(Boolean):[]);await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...N(I.rawAnalysisJson),...q,taskMeta:z,caseSnapshot:{taskInput:e_,rawTaskInput:eh,taskInputExtraction:ew}})}});let eS=[];if(eg){let e=await n.prismaRaw.session.findUnique({where:{taskId:eg}});if(e?.interactions)try{let t=JSON.parse(e.interactions);eS=Array.isArray(t)?t:[]}catch(e){throw new Q("trace-parse",`Session.interactions JSON 解析失败: ${e.message}`,e)}}if(0===eS.length&&(e_||ev)){let e=new Date().toISOString();eS=[...e_?[{role:"user",content:e_,timestamp:e}]:[],...ev?[{role:"assistant",content:ev,timestamp:e}]:[]]}let eE=null,ek=null,ex="no-dataset";if(I.datasetId&&I.caseId){try{let t=await (0,i.findAgentDataset)(e,I.datasetId);if(!t)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 评估数据集不存在，无法匹配可评测 case`);ek={id:t.id,name:t.name};let a=b(e_);if(!a&&!I.caseId)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} trace 没有实际输入，无法匹配评估数据集 case`);let r=(I.caseId?t.cases.find(e=>e.id===I.caseId):t.cases.find(e=>b(e.input)===a))||null;if(r&&(ex=I.caseId?"explicit-pair":"exact-input"),!r&&!I.caseId){let a=await (0,g.findBestSemanticCaseMatch)(t.cases.map(e=>({id:e.id,input:e.input})),e_,{user:e,requireModelAvailable:!0});if(a.error)throw new Q("semantic-match-llm",`${p.NO_EVALUABLE_CASE_PREFIX} 语义匹配调用评测模型失败：${a.error}`);a.caseId&&(r=t.cases.find(e=>e.id===a.caseId)??null)&&(ex="semantic")}if(!r)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} trace 实际输入未匹配到当前数据集中的 case 输入`);eE=r}catch(e){if(e instanceof Q)throw e;throw new Q("dataset-load",`读取数据集失败: ${e.message}`,e)}if(!eE)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 未匹配到可评测 case`)}else if(G||H)try{let a=await L(e,e_,{requireExpectedOutput:G,allowedDatasetIds:eA});eE=a.caseEntry,ek={id:a.dataset.id,name:a.dataset.name},ex="auto-match",await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{datasetId:a.dataset.id,caseId:a.caseEntry.id}})}catch(e){if(e instanceof Q&&"no-evaluable-case"===e.stage&&!G)console.warn(`[trajectory/run] no matching dataset, falling back to empty case: ${e.message}`),eE={input:e_,expectedOutput:"",trajectory:"",evaluationFocus:"",rootCauses:[]},ex="fallback";else if(e instanceof Q)throw e;else throw new Q("dataset-auto-match",`自动匹配数据集失败: ${e.message}`,e)}else if(Y)if(ep)try{let a=await L(e,e_,{requireExpectedOutput:!0,includeAllDatasetKinds:!0,allowedDatasetIds:eA});eE=a.caseEntry,ek={id:a.dataset.id,name:a.dataset.name},ex="auto-match",await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{datasetId:a.dataset.id,caseId:a.caseEntry.id}})}catch(e){if(e instanceof Q)throw e;throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 自定义评估器需要 reference_output，但 trace 输入未匹配到带预期结果的数据集 case`,e)}else eE={input:e_,expectedOutput:"",trajectory:"",evaluationFocus:"",rootCauses:[]};else eE={input:e_,expectedOutput:"",trajectory:"",evaluationFocus:"",rootCauses:[]};if(G&&!b(eE.expectedOutput))throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 已匹配到 case，但该 case 缺少预期结果 expectedOutput，无法执行结果评测`);if(Y&&ep&&!b(eE.expectedOutput))throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 自定义评估器需要 reference_output，但未找到可用的预期结果 expectedOutput`);let eb=!!(eE?.rootCauseMeta&&(0,o.canReuseRootCauseCache)(eE.expectedOutput,eE.rootCauseMeta)),eR=eb&&eE?.rootCauseMeta?.status==="ready"?eE.rootCauses||[]:void 0,eI=eb?eE?.rootCauseMeta?.status==="empty"?"none":eE?.rootCauseMeta?.status==="ready"?"dataset-cache":void 0:void 0,eC=e_||eE.input||"",eN={id:eE.id||I.caseId||"",input:eE.input,taskInput:eC,rawTaskInput:eh,taskInputExtraction:ew,expectedOutput:eE.expectedOutput,trajectory:eE.trajectory,evaluationFocus:eE.evaluationFocus,rootCauseCacheStatus:eE.rootCauseMeta?.status||null,rootCauseCacheUpdatedAt:eE.rootCauseMeta?.updatedAt||null,datasetId:ek?.id||I.datasetId||"",datasetName:ek?.name||"",matchKind:ex},e$="trajectory",eO="",ej="",eM=[],eT=[],eL=eE.evaluationFocus,eJ=eE.trajectory,eF=eS.length>0?(u=eS,{summary:R=(0,A.summarizeTrace)(u,{maxSteps:80,maxTextLen:400}),text:(0,A.formatTraceForLLM)(R),steps:R.steps.map(e=>({...e,step_index:e.index,stepIndex:e.index}))}):null,eP=await (0,f.buildSkillKeyActionReference)(ey,e,eS),eD=(0,f.getPrimaryExecutionSkillTargets)(ey,eS).length>0?"skill-aware":"no-skill",eU="ok"===eP.status?{status:"ok",referenceKeyActionsText:eP.referenceKeyActionsText,actualExtractedStepsText:eF?.text||"",referenceKeyActions:eP.referenceKeyActions,actualExtractedSteps:eF?.steps||[]}:eP,eq=(0,S.buildSkillAttributionStatus)(eU);if("ok"===eP.status&&eF&&eF.steps.length>0)e$="skill_key_actions",eO=eP.referenceKeyActionsText,ej=eF.text,eM=eP.referenceKeyActions??[],eT=eF.steps,eL=[b(eE.evaluationFocus),"逐条比较 Skill 自动提取的关键动作与 trace 事件级步骤（user/llm/tool/skill/task），并直接生成每个关键动作的 Skill 改进建议。"].filter(Boolean).join(" ");else if(H){if(!eF||0===eF.steps.length)throw new Q("no-evaluable-case",`${p.NO_EVALUABLE_CASE_PREFIX} 当前 trace 无法提取可评测事件步骤`);e$="trace_only",ej=eF.text,eT=eF.steps,eL=[b(eE.evaluationFocus),"当前 trace 未获得可用 Skill 关键动作；不评估完整性，不生成 Skill 关键动作建议，仅评估工具选择与冗余。"].filter(Boolean).join(" ")}let ez={...q,taskMeta:z,...W?{grayscaleBinding:W}:{},caseSnapshot:eN,skillAttribution:eq,resultSkillMode:eD,comparisonMode:e$,actualTraceExtraction:eF?{source:"summarizeTrace",totalSteps:eF.summary.totalSteps,shownSteps:eF.steps.length,truncated:eF.summary.truncated,totalLlmCalls:eF.summary.totalLlmCalls,totalToolCalls:eF.summary.totalToolCalls,totalSkillCalls:eF.summary.totalSkillCalls,totalTaskCalls:eF.summary.totalTaskCalls}:null,skillKeyActionComparison:"ok"===eP.status?{referenceKeyActionsText:eP.referenceKeyActionsText,actualExtractedStepsText:ej,referenceKeyActions:eP.referenceKeyActions??[],actualExtractedSteps:eT}:null};if((H||G)&&0===eS.length)throw new Q("trace-empty",`taskId=${eg||"N/A"} 对应的 Session 不存在或 interactions 为空`);if(H&&!G){let a=String(ey?.finalResult||"").trim();try{let t=await j((0,y.extractTaskResultArtifact)({userTask:eE.input,interactions:eS,fallbackOutput:a,user:e}),6e5,new Q("result-artifact-timeout",`实际输出提取超过 600s 未完成`));es=t.rawAnalysis,ei=String(t.outputForEvaluation||a||"").trim()}catch(e){console.warn("[trajectory-eval] result artifact extraction failed:",e),ei=a}await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:null,resultEvaluationError:null})}})}if(G){let a="skill-aware"===eD?await (0,f.loadTaskCompletionSkillContext)(ey,eS,e):void 0,r=eF?.text||"",s=Math.min(function(e){let t=N(e).resultEvaluationRetry;if(!t||"object"!=typeof t)return 0;let a=t.attemptCount;return"number"==typeof a&&Number.isFinite(a)&&a>0?Math.floor(a):0}(I.rawAnalysisJson)+1,5),i=new Date;try{await j((async()=>{let o=await (0,y.extractTaskResultArtifact)({userTask:eE.input,interactions:eS,fallbackOutput:ev,user:e});if(es=o.rawAnalysis,!(ei=String(o.outputForEvaluation||"").trim())){et=`结果输出提取失败：${o.reason}`,eo=(0,_.buildEvaluationDiagnostic)(String(o.reason||"").includes("没有可用于输出提取的 LLM 调用")?"trace_missing_output":"task_output_extraction_failed",{details:et,reason:et}),ea=M(s,{retrying:!1,exhausted:!0,lastError:et,now:i}),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:"",resultEvaluation:null,resultEvaluationError:et,resultEvaluationRetry:ea,diagnostic:eo})}});return}ea=M(s,{retrying:!1,exhausted:!1,now:i}),eo=null,await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:null,resultEvaluationError:null,resultEvaluationRetry:ea})}});let l=await O(eE.input,eE.expectedOutput,ei,eR,eI,r,a,eD,e,F,D);ee={...l.rawAnalysis||{},score:l.score,is_correct:l.isCorrect,reason:l.reason,result_artifact_extraction:{status:o.rawAnalysis?.fallback?"fallback_final_result":o.status,confidence:o.confidence,reason:o.reason,source_refs:o.sourceRefs,fallback_source:o.rawAnalysis?.fallback?"execution.finalResult":null}},await T(ey,eg,l),ea=M(s,{retrying:!1,exhausted:!1,now:i}),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:ee,resultEvaluationError:null,resultEvaluationRetry:ea})}}),et=null,eo=null})(),6e5,new Q("result-timeout",`结果评测超过 600s 未完成`))}catch(a){console.warn(`[trajectory-eval] result-based evaluation attempt ${s}/5 failed:`,a);let e=a.message||String(a);s<5&&(0,p.isRetryableResultEvaluationFailure)(a)?(er=V[Math.max(0,s-1)]??V[V.length-1]??8e3,en=s,ea=M(s,{retrying:!0,exhausted:!1,lastError:e,delayMs:er,now:i}),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:"running",errorMessage:null,rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:null,resultEvaluationError:null,resultEvaluationRetry:ea})}})):(et=`结果评测失败：${e}`,eo=(0,_.buildEvaluationDiagnostic)(a instanceof Q&&("result-timeout"===a.stage||"result-artifact-timeout"===a.stage)?"evaluator_timeout":"result_evaluator_failed",{details:et,reason:et}),ea=M(s,{retrying:!1,exhausted:!0,lastError:e,now:i}),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:ee,resultEvaluationError:et,resultEvaluationRetry:ea,diagnostic:eo})}}))}}if(null!=er)return void(d=er,h=en,setTimeout(()=>{P(e,t).catch(e=>{console.error(`[trajectory-eval] scheduled result retry failed row=${t} attempt=${h+1}/5:`,e)})},d));if(Y){if(em&&!ei){let t=String(ey?.finalResult||"").trim();if(eS.length>0)try{let a=await j((0,y.extractTaskResultArtifact)({userTask:eC,interactions:eS,fallbackOutput:t,user:e}),6e5,new Q("result-artifact-timeout",`实际输出提取超过 600s 未完成`));es=a.rawAnalysis,ei=String(a.outputForEvaluation||t||"").trim()}catch(e){console.warn("[trajectory-eval] custom evaluator output extraction failed:",e),ei=t}else ei=t}let t=em?ei:"",a=ef&&eF?.text||"",r=await Promise.all(Z.map(r=>(0,v.runCustomLlmEvaluator)(e,r,{caseInput:eC,expectedOutput:ep?eE.expectedOutput:"",actualOutput:t,traceText:a},F,D).catch(e=>({evaluatorId:r,evaluatorName:r,score:null,reason:"",rawResponse:"",model:"",durationMs:0,error:`运行异常：${e?.message||String(e)}`})))),n={};for(let e of r)n[e.evaluatorId]={evaluatorId:e.evaluatorId,evaluatorName:e.evaluatorName,score:e.score,reason:e.reason,model:e.model,durationMs:e.durationMs,error:e.error},"number"==typeof e.score&&eu.push(e.score);el=n}if(!H){let a=eu.length>0?eu.reduce((e,t)=>e+t,0)/eu.length:null,r=ee?.evaluatorSessionId,s="string"==typeof r?r.trim():"",i=G&&(!ee||!s),o=i?"failed":"done",l=i?et||"结果评测未产出有效结果或评测 session":null;i&&!eo&&(eo=(0,_.buildEvaluationDiagnostic)("result_evaluator_failed",{details:l||void 0,reason:l||void 0})),await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:o,trajectoryScore:null,dimensionScoresJson:null,deviationStepsJson:JSON.stringify([]),rootCauseStep:null,reasonText:null,rawAnalysisJson:JSON.stringify({...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:ee,resultEvaluationError:et,resultEvaluationRetry:ea,customEvaluations:el,customVariableNeeds:Array.from(ed),customEvaluationScore:a,...eo?{diagnostic:eo}:{}}),errorMessage:l}}),"done"===o&&await U(e,t,ey,eS);return}let eW={caseId:eE.id||I.caseId||"",caseInput:eE.input,referenceOutput:eE.expectedOutput,referenceTrajectory:eJ,referenceKeyActionsText:eO,actualExtractedStepsText:ej,referenceKeyActions:eM,actualExtractedSteps:eT,skillContext:await (0,f.loadTaskCompletionSkillContext)(ey,eS,e),comparisonMode:e$,evaluationFocus:eL,actualInteractions:eS,taskId:eg||void 0,executionId:I.executionId||ey?.id||void 0};try{E=await Promise.race([(0,l.evaluateTrajectoryViaOpencode)(eW,e,F,D),new Promise((e,t)=>setTimeout(()=>t(new Q("timeout",`单条评测超过 600s 未完成`)),6e5))])}catch(e){if(e instanceof Q)throw e;if(e instanceof l.TrajectoryEvalConfigError)throw new Q("config",e.message,e);throw new Q("llm-or-agent",e.message||String(e),e)}try{let e={...E.rawAnalysis&&"object"==typeof E.rawAnalysis?E.rawAnalysis:{},...ez,resultArtifactExtraction:es,resultActualOutput:ei,resultEvaluation:ee,resultEvaluationError:et,resultEvaluationRetry:ea,customEvaluations:el,customVariableNeeds:Array.from(ed),customEvaluationScore:eu.length>0?eu.reduce((e,t)=>e+t,0)/eu.length:null,...eo?{diagnostic:eo}:{}};await n.prismaRaw.trajectoryEvalResult.update({where:{id:t},data:{status:"done",trajectoryScore:E.trajectoryScore,dimensionScoresJson:JSON.stringify(E.dimensionScores),deviationStepsJson:JSON.stringify(E.deviationSteps),rootCauseStep:E.rootCauseStep||null,reasonText:E.reasonText||null,rawAnalysisJson:JSON.stringify(e),errorMessage:null}})}catch(e){throw new Q("persist",`写入评测结果失败: ${e.message}`,e)}await U(e,t,ey,eS)}finally{if(a&&r)try{(0,c.finish)(e,a,r)}catch(t){console.warn(`[trajectory-eval] finishEvalTask failed user=${e} task=${a}: ${t?.message}`)}}}async function U(e,t,a,r){try{let s=await n.prismaRaw.trajectoryEvalResult.findUnique({where:{id:t},select:{id:!0,taskId:!0,evaluatorRunId:!0,deviationStepsJson:!0,rootCauseStep:!0,reasonText:!0,rawAnalysisJson:!0}});if(!s)return;let i=(0,f.getPrimaryExecutionSkillTargets)(a,r).map(e=>({name:e.skill,version:e.version}));if(0===i.length)return;let o=await (0,d.deriveAndPersistOptPoints)({user:e,taskId:s.taskId,runId:s.evaluatorRunId,trajectoryRow:s,skills:i});o>0&&console.log(`[trajectory-eval]   • derived ${o} skill optimization point(s) for row=${t}`)}catch(e){console.warn("[trajectory-eval] derive opt points failed (non-fatal):",e.message)}}async function q(e,t,a){let r=Date.now();console.log(`[trajectory-eval] run ${t} started: ${a.length} traces, concurrency=3`);let s=[...a],i=0,o=0,l=0;await new Promise(r=>{let u=()=>{for((0,m.getTrajectoryEvalRunSignal)(t)?.aborted&&(s.length=0);i<3&&s.length>0;){let c=s.shift();i++,P(e,c).then(()=>{o++,console.log(`[trajectory-eval]   ✓ ${c} done (${o}+${l}/${a.length})`)}).catch(async e=>{l++;let t=e instanceof Q?e.stage:"unknown",a=e?.message||"unknown error",r=function(e){let t=e instanceof Q?e.stage:"unknown",a=(e instanceof Error?e.message:String(e||"unknown error")).replace(/^\[[^\]]+\]\s*/,"").replace(p.NO_EVALUABLE_CASE_PREFIX,"").trim();if("trace-parse"===t)return(0,_.buildEvaluationDiagnostic)("trace_parse_failed",{details:a,reason:a});if("trace-empty"===t)return(0,_.buildEvaluationDiagnostic)("trace_data_empty",{details:a,reason:a});if("semantic-match-llm"===t||"dataset-auto-match"===t)return(0,_.buildEvaluationDiagnostic)("case_match_failed",{details:a,internal:{stage:"semantic_match",reason:a}});if("result-artifact-timeout"===t||"result-timeout"===t||"timeout"===t)return(0,_.buildEvaluationDiagnostic)("evaluator_timeout",{details:a,reason:a});if("persist"===t)return(0,_.buildEvaluationDiagnostic)("persist_failed",{details:a,reason:a});if("llm-or-agent"===t||"config"===t)return(0,_.buildEvaluationDiagnostic)("evaluator_output_invalid",{details:a,reason:a});if("no-evaluable-case"===t)return a.includes("没有实际输入")?(0,_.buildEvaluationDiagnostic)("trace_missing_input",{details:a,reason:a}):a.includes("评估数据集不存在")?(0,_.buildEvaluationDiagnostic)("dataset_missing",{details:a,reason:a}):a.includes("未找到可用于评测的数据集")||a.includes("未找到可用于结果评测的数据集")||a.includes("未找到轨迹评测数据集")?(0,_.buildEvaluationDiagnostic)("dataset_empty",{details:a,reason:a}):a.includes("reference_output")?(0,_.buildEvaluationDiagnostic)("custom_reference_missing",{details:a,reason:a}):a.includes("expectedOutput")||a.includes("预期结果")?(0,_.buildEvaluationDiagnostic)("case_missing_expected_output",{details:a,reason:a}):a.includes("动态分析失败")?(0,_.buildEvaluationDiagnostic)("trace_step_analysis_failed",{details:a,reason:a}):a.includes("无法提取可评测步骤")||a.includes("提取出实际执行步骤")||a.includes("无可提取执行步骤")?(0,_.buildEvaluationDiagnostic)("trace_no_steps",{details:a,reason:a}):(0,_.buildEvaluationDiagnostic)("case_not_matched",{details:a,reason:a});return(0,_.buildEvaluationDiagnostic)("evaluator_recovered_failed",{details:a,reason:a})}(e);console.error(`[trajectory-eval]   ✗ ${c} failed [${t}]: ${a}`);let s=await n.prismaRaw.trajectoryEvalResult.findUnique({where:{id:c},select:{rawAnalysisJson:!0}}).catch(()=>null);await n.prismaRaw.trajectoryEvalResult.update({where:{id:c},data:{status:"failed",errorMessage:a,rawAnalysisJson:function(e,t,a={}){return JSON.stringify({...N(e),...a,diagnostic:t})}(s?.rawAnalysisJson,r)}}).catch(e=>console.error(`[trajectory-eval] failed to mark ${c} as failed: ${e.message}`))}).finally(()=>{i--,(0,m.getTrajectoryEvalRunSignal)(t)?.aborted&&(s.length=0),0===s.length&&0===i?r():u()})}};u()});try{let e=await n.prismaRaw.trajectoryEvalResult.updateMany({where:{id:{in:a},evaluatorRunId:t,updatedAt:{lt:new Date(r)},status:{in:["pending","running"]}},data:{status:"failed",errorMessage:"[recover] 进程异常或评测器未返回，已自动标 failed"}});e.count>0&&(console.warn(`[trajectory-eval] run ${t} recovered ${e.count} stuck rows.`),l+=e.count)}catch(e){console.error(`[trajectory-eval] failed to recover stuck rows: ${e.message}`)}let u=((Date.now()-r)/1e3).toFixed(1);console.log(`[trajectory-eval] run ${t} finished: ${o} done, ${l} failed, ${u}s`)}e.s(["PATCH",()=>F,"POST",()=>J,"dynamic",0,"force-dynamic","runEvaluations",()=>q]),a()}catch(e){a(e)}},!1),656687,e=>e.a(async(t,a)=>{try{var r=e.i(747909),n=e.i(174017),s=e.i(996250),i=e.i(759756),o=e.i(561916),l=e.i(174677),u=e.i(869741),c=e.i(316795),d=e.i(487718),p=e.i(995169),m=e.i(47587),f=e.i(666012),g=e.i(570101),y=e.i(626937),h=e.i(10372),w=e.i(193695);e.i(52474);var _=e.i(600220),v=e.i(600461),A=t([v]);[v]=A.then?(await A)():A;let k=new r.AppRouteRouteModule({definition:{kind:n.RouteKind.APP_ROUTE,page:"/api/eval/trajectory/run/route",pathname:"/api/eval/trajectory/run",filename:"route",bundlePath:""},distDir:".next",relativeProjectDir:"",resolvedPagePath:"[project]/src/app/api/eval/trajectory/run/route.ts",nextConfigOutput:"standalone",userland:v}),{workAsyncStorage:x,workUnitAsyncStorage:b,serverHooks:R}=k;function S(){return(0,s.patchFetch)({workAsyncStorage:x,workUnitAsyncStorage:b})}async function E(e,t,a){k.isDev&&(0,i.addRequestMeta)(e,"devRequestTimingInternalsEnd",process.hrtime.bigint());let r="/api/eval/trajectory/run/route";r=r.replace(/\/index$/,"")||"/";let s=await k.prepare(e,t,{srcPage:r,multiZoneDraftMode:!1});if(!s)return t.statusCode=400,t.end("Bad Request"),null==a.waitUntil||a.waitUntil.call(a,Promise.resolve()),null;let{buildId:v,params:A,nextConfig:S,parsedUrl:E,isDraftMode:x,prerenderManifest:b,routerServerContext:R,isOnDemandRevalidate:I,revalidateOnlyGenerated:C,resolvedPathname:N,clientReferenceManifest:$,serverActionsManifest:O}=s,j=(0,u.normalizeAppPath)(r),M=!!(b.dynamicRoutes[j]||b.routes[N]),T=async()=>((null==R?void 0:R.render404)?await R.render404(e,t,E,!1):t.end("This page could not be found"),null);if(M&&!x){let e=!!b.routes[N],t=b.dynamicRoutes[j];if(t&&!1===t.fallback&&!e){if(S.experimental.adapterPath)return await T();throw new w.NoFallbackError}}let L=null;!M||k.isDev||x||(L=N,L="/index"===L?"/":L);let J=!0===k.isDev||!M,F=M&&!J;O&&$&&(0,l.setManifestsSingleton)({page:r,clientReferenceManifest:$,serverActionsManifest:O});let P=e.method||"GET",D=(0,o.getTracer)(),U=D.getActiveScopeSpan(),q={params:A,prerenderManifest:b,renderOpts:{experimental:{authInterrupts:!!S.experimental.authInterrupts},cacheComponents:!!S.cacheComponents,supportsDynamicResponse:J,incrementalCache:(0,i.getRequestMeta)(e,"incrementalCache"),cacheLifeProfiles:S.cacheLife,waitUntil:a.waitUntil,onClose:e=>{t.on("close",e)},onAfterTaskError:void 0,onInstrumentationRequestError:(t,a,r,n)=>k.onRequestError(e,t,r,n,R)},sharedContext:{buildId:v}},z=new c.NodeNextRequest(e),W=new c.NodeNextResponse(t),B=d.NextRequestAdapter.fromNodeNextRequest(z,(0,d.signalFromNodeResponse)(t));try{let s=async e=>k.handle(B,q).finally(()=>{if(!e)return;e.setAttributes({"http.status_code":t.statusCode,"next.rsc":!1});let a=D.getRootSpanAttributes();if(!a)return;if(a.get("next.span_type")!==p.BaseServerSpan.handleRequest)return void console.warn(`Unexpected root span type '${a.get("next.span_type")}'. Please report this Next.js issue https://github.com/vercel/next.js`);let n=a.get("next.route");if(n){let t=`${P} ${n}`;e.setAttributes({"next.route":n,"http.route":n,"next.span_name":t}),e.updateName(t)}else e.updateName(`${P} ${r}`)}),l=!!(0,i.getRequestMeta)(e,"minimalMode"),u=async i=>{var o,u;let c=async({previousCacheEntry:n})=>{try{if(!l&&I&&C&&!n)return t.statusCode=404,t.setHeader("x-nextjs-cache","REVALIDATED"),t.end("This page could not be found"),null;let r=await s(i);e.fetchMetrics=q.renderOpts.fetchMetrics;let o=q.renderOpts.pendingWaitUntil;o&&a.waitUntil&&(a.waitUntil(o),o=void 0);let u=q.renderOpts.collectedTags;if(!M)return await (0,f.sendResponse)(z,W,r,q.renderOpts.pendingWaitUntil),null;{let e=await r.blob(),t=(0,g.toNodeOutgoingHttpHeaders)(r.headers);u&&(t[h.NEXT_CACHE_TAGS_HEADER]=u),!t["content-type"]&&e.type&&(t["content-type"]=e.type);let a=void 0!==q.renderOpts.collectedRevalidate&&!(q.renderOpts.collectedRevalidate>=h.INFINITE_CACHE)&&q.renderOpts.collectedRevalidate,n=void 0===q.renderOpts.collectedExpire||q.renderOpts.collectedExpire>=h.INFINITE_CACHE?void 0:q.renderOpts.collectedExpire;return{value:{kind:_.CachedRouteKind.APP_ROUTE,status:r.status,body:Buffer.from(await e.arrayBuffer()),headers:t},cacheControl:{revalidate:a,expire:n}}}}catch(t){throw(null==n?void 0:n.isStale)&&await k.onRequestError(e,t,{routerKind:"App Router",routePath:r,routeType:"route",revalidateReason:(0,m.getRevalidateReason)({isStaticGeneration:F,isOnDemandRevalidate:I})},!1,R),t}},d=await k.handleResponse({req:e,nextConfig:S,cacheKey:L,routeKind:n.RouteKind.APP_ROUTE,isFallback:!1,prerenderManifest:b,isRoutePPREnabled:!1,isOnDemandRevalidate:I,revalidateOnlyGenerated:C,responseGenerator:c,waitUntil:a.waitUntil,isMinimalMode:l});if(!M)return null;if((null==d||null==(o=d.value)?void 0:o.kind)!==_.CachedRouteKind.APP_ROUTE)throw Object.defineProperty(Error(`Invariant: app-route received invalid cache entry ${null==d||null==(u=d.value)?void 0:u.kind}`),"__NEXT_ERROR_CODE",{value:"E701",enumerable:!1,configurable:!0});l||t.setHeader("x-nextjs-cache",I?"REVALIDATED":d.isMiss?"MISS":d.isStale?"STALE":"HIT"),x&&t.setHeader("Cache-Control","private, no-cache, no-store, max-age=0, must-revalidate");let p=(0,g.fromNodeOutgoingHttpHeaders)(d.value.headers);return l&&M||p.delete(h.NEXT_CACHE_TAGS_HEADER),!d.cacheControl||t.getHeader("Cache-Control")||p.get("Cache-Control")||p.set("Cache-Control",(0,y.getCacheControlHeader)(d.cacheControl)),await (0,f.sendResponse)(z,W,new Response(d.value.body,{headers:p,status:d.value.status||200})),null};U?await u(U):await D.withPropagatedContext(e.headers,()=>D.trace(p.BaseServerSpan.handleRequest,{spanName:`${P} ${r}`,kind:o.SpanKind.SERVER,attributes:{"http.method":P,"http.target":e.url}},u))}catch(t){if(t instanceof w.NoFallbackError||await k.onRequestError(e,t,{routerKind:"App Router",routePath:j,routeType:"route",revalidateReason:(0,m.getRevalidateReason)({isStaticGeneration:F,isOnDemandRevalidate:I})},!1,R),M)throw t;return await (0,f.sendResponse)(z,W,new Response(null,{status:500})),null}}e.s(["handler",()=>E,"patchFetch",()=>S,"routeModule",()=>k,"serverHooks",()=>R,"workAsyncStorage",()=>x,"workUnitAsyncStorage",()=>b]),a()}catch(e){a(e)}},!1)];
 
 //# sourceMappingURL=_73cb0d03._.js.map