openpond 0.2.55 → 0.2.57

package/dist/chunks/{chunk-QRAKXZOC.js → chunk-YOBN7KTF.js}

@@ -1,5 +1,5 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{Mb as nr,Pb as or,Qb as sr,V as Mt,Wa as ar,ab as et,ba as Bt,bb as tt,db as at,q as Ze,v as er,x as tr}from"./chunk-YUPBAWBJ.js";import{a as rr}from"./chunk-QKD5OMDI.js";import{i as Ya}from"./chunk-T4QGDQHP.js";import{a as Ye}from"./chunk-OSIO7PLX.js";import{f as Za,h as He}from"./chunk-STDMBJOQ.js";import{E as Ja,L as Ga,M as _e,Ma as Xa,N as Ka,O as Oe,Pa as Xe,Ua as Qe,Va as se,Wa as Qa,a as We,b as Ia,c as Sa,d as Ha,e as Ea,f as Aa,i as Wa}from"./chunk-2DFR7PJ4.js";import{g as he}from"./chunk-5CMKGD43.js";import{e as wa,g as Va}from"./chunk-L4235TRQ.js";import{Ed as Na,Gd as Se,Hd as Je,Jd as Da,Od as Fa,Sd as $a,Td as za,Xa as Be,_c as qa,be as Ua,fe as Ge,g as Le,h as sa,he as F,i as ia,ie as Ke,j as da,k as ca,kf as Ct,q as la,s as ma,t as ua}from"./chunk-JIMU3KDA.js";import{Dc as La,Ka as Me,Na as Tt,ra as Oa,wa as ja}from"./chunk-YDT7L3VZ.js";import{$ as xa,A as Ra,B as Ve,G as ba,V as ya,ca as Pa,k as pa,ka as Ta,ma as Ca}from"./chunk-55G6B54U.js";import{hb as Ie,ib as ka,k as fa,l as ha,r as ga,ra as va}from"./chunk-KD3QVRIJ.js";import{Z as _a}from"./chunk-ARQEJOEK.js";import{P as Ma}from"./chunk-JUJSA7KI.js";import{h as oa,k as re,l as fe,m as v}from"./chunk-KGNCKIFP.js";import{b as M}from"./chunk-ES2ISFJC.js";import{n as Ba}from"./chunk-H7IWJN7L.js";import{a as h}from"./chunk-FTQLQJQV.js";import{execFile as Wr}from"node:child_process";import{readFile as Jr}from"node:fs/promises";import{promisify as Gr}from"node:util";var Kr=5,Xr=20*1024*1024,ir=4e4,Ot=Gr(Wr);function Zn(e){return async({grader:t,task:a,attempt:n})=>{let s=[],o=0,i=await Qr({store:e.store,attempt:n}),d=n.modelRef?.providerId===t.judge.providerId&&n.modelRef.modelId===t.judge.modelId?tr(n.metadata.hostedTokenPricing)??void 0:void 0,g=await e.modelText({model:t.judge,hostedTokenPricing:d,signal:new AbortController().signal,requestId:`task-judge:${n.id}:${t.id}`,onUsage:h((m,r)=>{s.push(m),typeof r=="number"&&Number.isFinite(r)&&r>=0&&(o+=r)},"onUsage"),messages:[{role:"system",content:`Apply this rubric and return JSON only with score (0..1), passed, and feedback.
+import{b as tr,s as Mt}from"./chunk-7HPDGZOR.js";import{a as sr}from"./chunk-PF5R6CMP.js";import{a as tt}from"./chunk-X4OSOOSU.js";import{b as at}from"./chunk-3K4VVB7S.js";import{a as Qa}from"./chunk-FEAKTSRS.js";import{a as Xe}from"./chunk-OSIO7PLX.js";import{d as Ze}from"./chunk-FY5KHLMW.js";import{b as et}from"./chunk-773TH7FN.js";import{a as ar,f as nr,i as or}from"./chunk-HDL5YSAA.js";import{a as rr}from"./chunk-HIOKCIL7.js";import{a as Bt}from"./chunk-IZ6DLEVY.js";import{a as er}from"./chunk-7KWD2I6W.js";import{g as he}from"./chunk-KG325WTN.js";import{c as Ya,e as He}from"./chunk-FUOEIDMA.js";import{a as Ja,b as Be,c as Ga,d as _e,g as Ge,l as Ke,m as se,n as Xa}from"./chunk-QNG27G3T.js";import{c as Va,y as Wa}from"./chunk-EIRYMSEG.js";import{q as Oe}from"./chunk-WJ43MJ3F.js";import{o as Za,s as Qe,u as F,v as Ye}from"./chunk-QOGJAQDO.js";import{a as Da,c as Se,d as Je,f as Fa,l as $a,p as za,r as Ua}from"./chunk-EYPCBBMS.js";import{a as Ct}from"./chunk-7ED24PP6.js";import{a as Ka}from"./chunk-2IHO4GDJ.js";import{b as la,d as ma,e as ua}from"./chunk-N4FFQPLR.js";import{a as Le,e as sa,f as ia,g as da,h as ca}from"./chunk-GJ2ZJFDD.js";import{e as wa}from"./chunk-H6IW2R22.js";import{b as La}from"./chunk-L3KS6MOV.js";import{r as Na}from"./chunk-PRGBLQVH.js";import{n as qa}from"./chunk-XDQMGUWL.js";import{ca as Oa,la as ja,ua as Me,xa as Tt}from"./chunk-25MBPJVY.js";import{f as ya}from"./chunk-YBT2NB32.js";import{k as pa}from"./chunk-74ZXXBGN.js";import{Z as _a}from"./chunk-I7AQMAQI.js";import{M as Ma}from"./chunk-Z4Z5UOZM.js";import{G as We,H as Ia,L as Sa,R as Ha,S as Ea,T as Aa,ba as xa,e as Ra,f as Ve,ga as Pa,n as ba,pa as Ta,ra as Ca}from"./chunk-Y4OWBBDD.js";import{Cb as Ie,Db as ka,b as ga,qa as va}from"./chunk-ACYFMLWG.js";import{r as fa,s as ha}from"./chunk-TVQAPBLZ.js";import{h as oa,k as re,l as fe,m as v}from"./chunk-KGNCKIFP.js";import{b as M}from"./chunk-ES2ISFJC.js";import{n as Ba}from"./chunk-H7IWJN7L.js";import{a as h}from"./chunk-FTQLQJQV.js";import{execFile as Wr}from"node:child_process";import{readFile as Jr}from"node:fs/promises";import{promisify as Gr}from"node:util";var Kr=5,Xr=20*1024*1024,ir=4e4,Ot=Gr(Wr);function Zn(e){return async({grader:t,task:a,attempt:n})=>{let s=[],o=0,i=await Qr({store:e.store,attempt:n}),d=n.modelRef?.providerId===t.judge.providerId&&n.modelRef.modelId===t.judge.modelId?tr(n.metadata.hostedTokenPricing)??void 0:void 0,g=await e.modelText({model:t.judge,hostedTokenPricing:d,signal:new AbortController().signal,requestId:`task-judge:${n.id}:${t.id}`,onUsage:h((m,r)=>{s.push(m),typeof r=="number"&&Number.isFinite(r)&&r>=0&&(o+=r)},"onUsage"),messages:[{role:"system",content:`Apply this rubric and return JSON only with score (0..1), passed, and feedback.
 
 ${t.rubric}`},{role:"user",content:JSON.stringify({input:a.input,expectedOutput:a.expectedOutput,output:n.output,artifactEvidence:i})}]}),f=ar(g);if(!f)throw new Error("Model judge returned invalid structured output.");return{...f,usage:s,...o>0?{costUsd:o}:{}}}}h(Zn,"createTaskAttemptModelJudge");async function Qr(e){let t=(await e.store.listTaskAttemptArtifacts({attemptId:e.attempt.id})).filter(a=>a.kind==="output_artifact").slice(0,Kr);return Promise.all(t.map(async a=>{let n={artifactId:a.id,mediaType:a.mediaType??null,sha256:a.sha256,sizeBytes:a.sizeBytes,requiredOutputPath:typeof a.metadata.requiredOutputPath=="string"?a.metadata.requiredOutputPath:null};if(a.sizeBytes>Xr)return{...n,extraction:"omitted",reason:"artifact exceeds inspection bound"};try{if(a.mediaType==="application/pdf"){let{stdout:s}=await Ot("pdftotext",[a.path,"-"],{env:{...process.env,PATH:Bt()},maxBuffer:8388608,timeout:15e3});return{...n,extraction:"pdf_text",content:_t(s)}}if(Zr(a.mediaType)){let s=await en(a.path,a.mediaType);return{...n,extraction:"open_xml_text",content:_t(s.content),structure:s.structure}}return Yr(a.mediaType)?{...n,extraction:"text",content:_t(await Jr(a.path,"utf8"))}:{...n,extraction:"metadata_only"}}catch(s){return{...n,extraction:"unavailable",reason:s instanceof Error?s.message.slice(0,500):String(s).slice(0,500)}}}))}h(Qr,"loadTaskAttemptGraderEvidence");function Yr(e){return!!(e?.startsWith("text/")||e==="application/json"||e==="application/javascript"||e==="application/sql"||e==="application/xml"||e==="application/yaml")}h(Yr,"isTextArtifact");function Zr(e){return e==="application/vnd.openxmlformats-officedocument.wordprocessingml.document"||e==="application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"||e==="application/vnd.openxmlformats-officedocument.presentationml.presentation"}h(Zr,"isOpenXmlArtifact");async function en(e,t){let a={env:{...process.env,PATH:Bt()},maxBuffer:16777216,timeout:15e3},{stdout:n}=await Ot("unzip",["-Z1",e],a),s=n.split(/\r?\n/).filter(Boolean),o=tn(s,t).slice(0,250),i=[];for(let d of o){let{stdout:g}=await Ot("unzip",["-p",e,d],a),f=rn(g);f&&i.push(`[${d}]
 ${f}`)}return{content:i.join(`
@@ -10,10 +10,10 @@ ${f}`)}return{content:i.join(`
 [truncated]`}h(_t,"boundedText");var jt=2,lr=18e4;function ot(e){if(e.seeds.length!==1||e.repetitions!==1)throw new Error("Sequential Harness Refiner benchmarks require one admitted seed and one trajectory repetition.");let t=e.taskset.benchmark;if(!t)throw new Error("Harness Refiner Taskset has no benchmark definition.");let a=e.selectedPopulation,n=a?.holdoutTaskIds??cr(e.taskset,t.evaluationSplit),s=a?.adaptationTaskIds??cr(e.taskset,t.adaptationSplit);if(a){if(a.adaptationSplit!==t.adaptationSplit||a.holdoutSplit!==t.evaluationSplit)throw new Error("The advanced plan changed its exact split binding.");let i=new Set;for(let d of s){let g=e.taskset.tasks.find(f=>f.id===d);if(!g||g.split!==a.adaptationSplit)throw new Error("An admitted adaptation task changed its split.");i.add(g.clusterKey)}for(let d of n){let g=e.taskset.tasks.find(f=>f.id===d);if(!g||g.split!==a.holdoutSplit||i.has(g.clusterKey))throw new Error("An admitted private holdout overlaps adaptation families.")}}let o=e.seeds.length*e.repetitions;return[rt("adaptation",t.adaptationSplit,s,o),rt("baseline",t.evaluationSplit,n,o),rt("candidate_adaptation",t.adaptationSplit,s,o),rt("candidate",t.evaluationSplit,n,o)]}h(ot,"createHarnessRefinerExecutionPlan");function qe(e){return e.reduce((t,a)=>t+a.attemptCount,0)}h(qe,"totalPlannedAttempts");function st(e){return new Set(e.flatMap(t=>t.taskIds)).size}h(st,"totalPlannedTasks");function ne(e,t){let a=e.findIndex(n=>n.stage===t);if(a<0)throw new Error(`Benchmark execution plan has no ${t} stage.`);return e.slice(0,a).reduce((n,s)=>n+s.attemptCount,0)}h(ne,"completedBeforeStage");var nt=class{static{h(this,"BenchmarkSpendBudget")}maximumSpendUsd;#e;constructor(t,a=0){if(!Number.isFinite(t)||t<=0)throw new Error("Benchmark maximum spend must be a positive finite number.");if(!Number.isFinite(a)||a<0)throw new Error("Observed benchmark spend must be a non-negative finite number.");this.maximumSpendUsd=t,this.#e=a}get observedSpendUsd(){return this.#e}assertAvailable(t){if(this.#e>=this.maximumSpendUsd)throw new Error(`Benchmark maximum spend of $${this.maximumSpendUsd.toFixed(2)} was reached before ${t}.`)}charge(t,a){if(typeof t!="number"||!Number.isFinite(t)||t<0)throw new Error(`Benchmark cost for ${a} is unavailable; the spend ceiling cannot be enforced.`);let n=this.#e+t;if(n>this.maximumSpendUsd+1e-9)throw new Error(`Benchmark maximum spend of $${this.maximumSpendUsd.toFixed(2)} was exceeded during ${a}.`);this.#e=n}},je=class extends Error{constructor(a,n,s){super(`Frozen web evidence is unavailable for ${a} ${n} call ${s+1}.`);this.taskId=a;this.toolName=n;this.ordinal=s;this.name="FrozenToolEvidenceExhaustedError"}taskId;toolName;ordinal;static{h(this,"FrozenToolEvidenceExhaustedError")}},ge=class{static{h(this,"BenchmarkEvidenceSnapshot")}#e=new Map;#t=new Map;constructor(t=[]){for(let a of t){let n=dr(a);if(this.#e.has(n))throw new Error(`Frozen web evidence contains duplicate observation ${n}.`);this.#e.set(n,structuredClone(a))}}async execute(t){if(t.toolName!=="web_search"&&t.toolName!=="web_fetch")return t.execute();let a=v(t.args),n=`${t.mode}:${t.cohort}:${t.taskId}:${t.toolName}`,s=this.#t.get(n)??0;this.#t.set(n,s+1);let o=dr({cohort:t.cohort,taskId:t.taskId,toolName:t.toolName,ordinal:s});if(t.mode==="replay"){let g=this.#e.get(o);if(!g)throw new je(t.taskId,t.toolName,s);return structuredClone(g.result)}let i=await t.execute(),d={cohort:t.cohort,taskId:t.taskId,toolName:t.toolName,argumentsHash:a,ordinal:s,result:structuredClone(i)};return this.#e.set(o,d),i}manifest(){let t=[...this.#e.values()].sort((n,s)=>JSON.stringify(n).localeCompare(JSON.stringify(s))),a={schemaVersion:"openpond.benchmarkEvidenceSnapshot.v1",id:`benchmark-evidence-${v(t).slice(0,24)}`,observations:t};return{...a,contentHash:v(a)}}};function dr(e){return`${e.cohort}:${e.taskId}:${e.toolName}:${e.ordinal}`}h(dr,"observationKey");function it(e){let t=e.baselineTokens-e.candidateTokens,a=e.refinerTokens+e.graderTokens,n=t-a,s=t>0?Math.ceil(a/t):null,o=e.amortizedReuseCount??10;return{grossForegroundTokenSavings:t,overheadTokens:a,firstPassNetTokenSavings:n,breakEvenReuseCount:s,amortizedTokenSavings:t*o-a,amortizedReuseCount:o}}h(it,"benchmarkEfficiency");function dt(e){let t=new Set(["grader_failure","environment_failure","infrastructure_failure","timeout","cancelled"]);return e.every(a=>typeof a.grade.score=="number"&&!t.has(a.grade.failureClass??""))}h(dt,"benchmarkAttemptsInfrastructureValid");function cr(e,t){let a=e.tasks.filter(n=>n.split===t).map(n=>n.id);if(!a.length)throw new Error(`Harness Refiner split ${t} has no cases.`);return a}h(cr,"taskIdsForSplit");function rt(e,t,a,n){return{stage:e,split:t,taskIds:a,attemptCount:a.length*n}}h(rt,"planItem");import{promises as qt}from"node:fs";import mr from"node:path";function Ft(e){return{run:e.run,attempts:e.attempts.map(t=>({attempt:t.attempt,grade:t.grade,artifacts:t.artifacts,receiptContentHash:t.portable.receipt.contentHash,artifactManifest:t.portable.artifactManifest,rewardReceipt:t.portable.rewardReceipt}))}}h(Ft,"completedStage");function Nt(e){return`${e.taskId}\0${e.seed}\0${e.attempt}`}h(Nt,"attemptKey");function on(e,t){let{contentHash:a,...n}=Ve.parse(e),s=Ra.parse({...n,runManifest:t});return Ve.parse({...s,contentHash:v(s)})}h(on,"rebaseAttemptReceipt");async function ur(e){if(!e.retries.length)return e.prior;let t=e.retries[0].portable.runManifest,a=new Map(e.retries.map(d=>[Nt(d.attempt),{evidence:{attempt:d.attempt,grade:d.grade,artifacts:d.artifacts,receiptContentHash:d.portable.receipt.contentHash,artifactManifest:d.portable.artifactManifest,rewardReceipt:d.portable.rewardReceipt},receipt:d.portable.receipt}])),n=e.prior.attempts.map(d=>{let g=a.get(Nt(d.attempt)),f=Ve.parse(d.attempt.metadata.portableAttemptReceipt);return g??{evidence:d,receipt:f}});if(n.length!==e.prior.run.attemptCount)throw new Error("Retried benchmark evidence changed the admitted attempt count.");let s=n.map(({receipt:d})=>on(d,{id:t.id,contentHash:t.contentHash})),o=ba({id:`benchmark-evaluation-${v({manifest:t.contentHash,phase:e.prior.run.phase,receipts:s.map(d=>d.contentHash)}).slice(0,24)}`,manifest:t,receipts:s,metadata:{kind:"benchmark",phase:e.prior.run.phase,split:e.prior.run.protocol.split,reasoningEffort:e.prior.run.reasoningEffort,sourceTasksetId:e.tasksetId,benchmarkDefinitionId:"harness-refiner",parentModelRunId:e.parentModelRunId,recoveredInfrastructureAttempts:e.retries.length}});await e.store.saveEvaluationResult({tasksetId:e.tasksetId,kind:e.prior.run.phase,result:o,createdAt:e.createdAt});let i=Ea({id:`benchmark-run-${o.contentHash.slice(0,24)}`,phase:e.prior.run.phase,evaluation:o,receipts:s,reasoningEffort:e.prior.run.reasoningEffort,protocol:e.prior.run.protocol,createdAt:e.createdAt,metadata:{...e.prior.run.metadata,recoveredInfrastructureAttempts:e.retries.length}});return await e.store.saveBenchmarkRun({tasksetId:e.tasksetId,run:i}),{run:i,attempts:n.map(({evidence:d})=>d)}}h(ur,"combineRetriedBenchmarkStage");async function ve(e){let[t,a,n]=await Promise.all([e.store.listBenchmarkRuns(e.tasksetId),e.store.listTaskAttempts(e.tasksetId),e.store.listGradeResultsForTaskset(e.tasksetId)]),s=e.plan.stage==="baseline"||e.plan.stage==="adaptation"?"baseline":"candidate",o=t.find(r=>r.metadata.parentModelRunId===e.modelRunId&&r.phase===s&&r.protocol.split===e.plan.split&&v(r.protocol.taskIds)===v(e.plan.taskIds));if(!o)throw new Error(`Completed ${e.plan.stage} benchmark evidence is unavailable.`);let i=new Map(e.plan.taskIds.map((r,p)=>[r,p])),d=new Map;for(let r of a)r.metadata.parentModelRunId===e.modelRunId&&i.has(r.taskId)&&sn(r)===o.harnessRelease.contentHash&&d.set(Nt(r),r);let g=[...d.values()].sort((r,p)=>(i.get(r.taskId)??0)-(i.get(p.taskId)??0)||r.seed-p.seed||r.attempt-p.attempt);if(g.length!==e.plan.attemptCount)throw new Error(`Completed ${e.plan.stage} evidence has ${g.length}/${e.plan.attemptCount} attempts.`);let f=new Map(n.map(r=>[r.attemptId,r])),m=await Promise.all(g.map(async r=>{let p=f.get(r.id);if(!p)throw new Error(`Attempt ${r.id} has no durable grade.`);return{attempt:r,grade:p,artifacts:await e.store.listTaskAttemptArtifacts({attemptId:r.id}),receiptContentHash:pr(r),...fr(r)}}));return{run:o,attempts:m}}h(ve,"loadCompletedBenchmarkStage");async function ct(e){let t=new Set(e.attemptIds),[a,n]=await Promise.all([e.store.listTaskAttempts(e.tasksetId),e.store.listGradeResultsForTaskset(e.tasksetId)]),s=a.filter(d=>t.has(d.id));if(s.length!==t.size)throw new Error(`Sequential adaptation evidence has ${s.length}/${t.size} attempts.`);let o=new Map(s.map(d=>[d.id,d])),i=new Map(n.map(d=>[d.attemptId,d]));return Promise.all(e.attemptIds.map(async d=>{let g=o.get(d);if(!g)throw new Error(`Sequential adaptation attempt ${d} is unavailable.`);let f=i.get(g.id);if(!f)throw new Error(`Attempt ${g.id} has no durable grade.`);return{attempt:g,grade:f,artifacts:await e.store.listTaskAttemptArtifacts({attemptId:g.id}),receiptContentHash:pr(g),...fr(g)}}))}h(ct,"loadBenchmarkAttemptEvidenceByIds");function sn(e){let t=oe(e.metadata.harnessCapabilityReceipt),a=oe(t?.harnessRelease);return typeof a?.contentHash=="string"?a.contentHash:null}h(sn,"attemptHarnessReleaseHash");function pr(e){let a=oe(e.metadata.portableAttemptReceipt)?.contentHash;if(typeof a!="string"||!/^[a-f0-9]{64}$/.test(a))throw new Error(`Attempt ${e.id} has no durable portable receipt.`);return a}h(pr,"portableReceiptContentHash");function fr(e){try{let t=xa.parse(e.metadata.portableArtifactManifest),a=Pa.parse(e.metadata.portableRewardReceipt);if(!Ta(t)||!Ca(a))throw new Error("Canonical reward evidence failed content-hash verification.");return{artifactManifest:t,rewardReceipt:a}}catch(t){throw new Error(`Attempt ${e.id} has no valid durable canonical reward evidence.`,{cause:t})}}h(fr,"portableCanonicalReceipts");function oe(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}h(oe,"objectRecord");function ke(e,t){return{agentSnapshot:e.agentSnapshot,harnessRelease:e.harnessRelease,instructionContext:t}}h(ke,"releasedHarness");function we(e,t){let a=e[t];return typeof a=="string"&&a.trim()?a:null}h(we,"stringMetadata");function hr(e){return typeof e.input.prompt=="string"&&e.input.prompt.trim()?e.input.prompt:(Array.isArray(e.input.messages)?e.input.messages:[]).flatMap(a=>{if(!a||typeof a!="object")return[];let n=a.content;return typeof n=="string"&&n.trim()?[n]:[]}).join(`
 
 `)}h(hr,"taskPrompt");async function lt(e,t,a){let n=await e.getModelProject(t);if(!n||n.profileId!==a)throw new Error("A Model in the active Profile is required.");return n}h(lt,"requireModelProject");async function ie(e,t,a){let n=await e.getModelRun(t);if(!n||n.kind!=="evaluation")throw new Error(`Evaluation Model Run ${t} is unavailable.`);return n.status!=="running"?n:e.saveModelRun(F.parse({...n,evaluationProgress:{...n.evaluationProgress,...a},updatedAt:new Date().toISOString()}))}h(ie,"updateProgress");function de(){return{usage:{baseline:Y(),adaptation:Y(),candidateAdaptation:Y(),candidate:Y(),refiner:Y(),grader:Y()},observedSpendUsd:0,attempts:[]}}h(de,"emptyEvaluationAccounting");async function mt(e,t,a){let n=await e.getModelRun(t);if(!n||n.status!=="running")return n;let s=structuredClone(n.evaluationProgress?.accounting??de()),o=Y(),i=a.result.attempt.metadata.usage;for(let f of Array.isArray(i)?i:[i])Ne(o,f);a.result.attempt.costUsd!==null&&(o.costUsd=a.result.attempt.costUsd);let d=a.stage==="candidate_adaptation"?"candidateAdaptation":a.stage;Dt(s.usage[d],o),Dt(s.usage.grader,a.grader),s.observedSpendUsd=a.observedSpendUsd;let g=ln(a.stage,[a.result])[0];return s.attempts=[...s.attempts.filter(f=>f.attemptId!==g.attemptId),g],ie(e,t,{stage:a.stage,completedAttempts:a.completedAttempts,totalAttempts:a.totalAttempts,accounting:s})}h(mt,"checkpointAttempt");async function gr(e,t,a,n){let s=await e.getModelRun(t);if(!s||s.status!=="running"||!s.evaluationProgress)return s;let o=structuredClone(s.evaluationProgress.accounting??de());return Dt(o.usage.refiner,a),o.observedSpendUsd=n,ie(e,t,{stage:s.evaluationProgress.stage,completedAttempts:s.evaluationProgress.completedAttempts,totalAttempts:s.evaluationProgress.totalAttempts,accounting:o})}h(gr,"checkpointRefinerUsage");async function ut(e,t,a,n){let s=mr.join(t,"training","harness-refiner-benchmarks",a);await qt.mkdir(s,{recursive:!0});let o=mr.join(s,"evidence-snapshot.json");await qt.writeFile(o,`${JSON.stringify(n,null,2)}
-`,"utf8");let i=await e.getModelRun(a);return!i||i.status!=="running"||!i.evaluationProgress?i:e.saveModelRun(F.parse({...i,evaluationProgress:{...i.evaluationProgress,evidenceSnapshot:{id:n.id,contentHash:n.contentHash,artifactPath:o}},updatedAt:new Date().toISOString()}))}h(ut,"checkpointEvidenceSnapshot");async function pt(e){let t=e.modelRun.evaluationProgress?.evidenceSnapshot;if(t){let s=dn(JSON.parse(await qt.readFile(_a(e.storeDir,t.artifactPath),"utf8")));if(s.id!==t.id||s.contentHash!==t.contentHash)throw new Error("Durable benchmark evidence snapshot does not match its admitted ref.");return new ge(s.observations)}let a=await cn(e.store,e.attempts),n=new ge(a);return await ut(e.store,e.storeDir,e.modelRun.id,n.manifest()),n}h(pt,"loadOrReconstructEvidenceSnapshot");function dn(e){let t=oe(e);if(t?.schemaVersion!=="openpond.benchmarkEvidenceSnapshot.v1"||typeof t.id!="string"||typeof t.contentHash!="string"||!Array.isArray(t.observations))throw new Error("Durable benchmark evidence snapshot is malformed.");let a=t.observations.map(s=>{let o=oe(s),i=oe(o?.result);if(o?.cohort!=="adaptation"&&o?.cohort!=="held_out"||typeof o.taskId!="string"||o.toolName!=="web_search"&&o.toolName!=="web_fetch"||typeof o.argumentsHash!="string"||!Number.isInteger(o.ordinal)||o.ordinal<0||!i||typeof i.toolCallId!="string"||typeof i.name!="string"||typeof i.ok!="boolean"||typeof i.contentText!="string")throw new Error("Durable benchmark evidence snapshot contains a malformed observation.");return o}),n={schemaVersion:"openpond.benchmarkEvidenceSnapshot.v1",id:t.id,observations:a};if(v(n)!==t.contentHash)throw new Error("Durable benchmark evidence snapshot failed content-hash validation.");return{...n,contentHash:t.contentHash}}h(dn,"parseEvidenceSnapshotManifest");async function cn(e,t){let a=[];for(let{result:n,cohort:s}of t){let o=we(n.attempt.metadata,"sessionId"),i=we(n.attempt.metadata,"turnId");if(!o||!i)continue;let d=(await e.runtimeEventsForSession(o,{names:["tool.started","tool.completed"],limit:1e5})).filter(m=>m.turnId===i),g=new Map,f=new Map;for(let m of d){let r=oe(m.data),p=typeof r?.toolCallId=="string"?r.toolCallId:null,l=m.action==="web_search"||m.action==="web_fetch"?m.action:null;if(!p||!l)continue;if(m.name==="tool.started"){let b=f.get(l)??0;f.set(l,b+1),g.set(p,{toolName:l,args:oe(m.args)??{},ordinal:b});continue}let c=g.get(p),w=oe(r?.result);if(!c||!w)continue;let u=typeof w.ok=="boolean"?w.ok:m.status==="completed",k=typeof w.output=="string"?w.output:typeof m.output=="string"?m.output:"",R=Object.fromEntries(Object.entries(w).filter(([b])=>b!=="ok"&&b!=="output"));a.push({cohort:s,taskId:n.attempt.taskId,toolName:c.toolName,argumentsHash:v(c.args),ordinal:c.ordinal,result:{toolCallId:p,name:c.toolName,ok:u,contentText:k,...Object.keys(R).length?{data:R}:{}}})}}return a}h(cn,"reconstructFrozenToolObservations");function ln(e,t){return t.map(a=>{let n=Y(),s=a.attempt.metadata.usage;for(let o of Array.isArray(s)?s:[s])Ne(n,o);return{phase:e,taskId:a.attempt.taskId,attemptId:a.attempt.id,sessionId:we(a.attempt.metadata,"sessionId"),turnId:we(a.attempt.metadata,"turnId"),passed:a.grade.passed,score:a.grade.score,failureClass:a.grade.failureClass,inputTokens:n.inputTokens,outputTokens:n.outputTokens,totalTokens:n.totalTokens,latencyMs:a.attempt.latencyMs,costUsd:a.attempt.costUsd,startedAt:a.attempt.startedAt}})}h(ln,"modelEvaluationAttempts");function wr(e){let t=Y();for(let a of Array.isArray(e)?e:[e])Ne(t,a);return{inputTokens:t.inputTokens,outputTokens:t.outputTokens,totalTokens:t.totalTokens}}h(wr,"attemptUsageSummary");function Y(){return{inputTokens:0,outputTokens:0,totalTokens:0,costUsd:null}}h(Y,"emptyUsageCategory");function Ne(e,t,a){let n=er(t);e.inputTokens+=n.promptTokens??0,e.outputTokens+=n.completionTokens??0,e.totalTokens+=n.totalTokens??0,typeof a=="number"&&Number.isFinite(a)&&a>=0&&(e.costUsd=(e.costUsd??0)+a)}h(Ne,"addUsage");function Dt(e,t){e.inputTokens+=t.inputTokens,e.outputTokens+=t.outputTokens,e.totalTokens+=t.totalTokens,t.costUsd!==null&&(e.costUsd=(e.costUsd??0)+t.costUsd)}h(Dt,"mergeUsage");function Ee(e,t){let a=e.find(n=>n.stage===t);if(!a)throw new Error(`Harness Refiner plan is missing ${t}.`);return a}h(Ee,"requirePlanStage");function ce(e,t,a){return{execute:h(async n=>{try{return{...await e.execute({...n,mode:t,cohort:a}),toolCallId:n.callId}}catch(s){if(t!=="replay"||!(s instanceof je))throw s;return{toolCallId:n.callId,name:n.toolName,ok:!1,contentText:JSON.stringify({ok:!1,action:n.toolName,output:"The benchmark's frozen external-evidence budget is exhausted for this task.",repairHint:"Do not call web tools again. Finish the task using the frozen evidence already returned."})}}},"execute")}}h(ce,"frozenToolEvidence");function ft(e){let t=[];for(let[a,n]of[["held-out baseline",e.baseline],["adaptation baseline",e.adaptation],["sequential adaptation treatment",e.candidateAdaptation],["held-out candidate",e.candidate]])n.terminalCount!==n.attemptCount&&t.push(`${a} did not terminalize all admitted attempts.`);return e.infrastructureValid||t.push("At least one attempt has infrastructure-invalid outcome evidence."),e.harnessChanged||t.push("The candidate Harness is unchanged."),e.lineageValid||t.push("Candidate Harness lineage is incomplete."),e.candidateAdaptation.passedCount!==e.candidateAdaptation.attemptCount&&t.push("Sequential adaptation treatment did not pass every case."),e.candidate.passedCount!==e.candidate.attemptCount&&t.push("Candidate held-out quality did not pass every case."),t}h(ft,"comparisonInvalidReasons");function ht(e){return!e.infrastructureValid||e.baseline.terminalCount!==e.baseline.attemptCount||e.adaptation.terminalCount!==e.adaptation.attemptCount||e.candidate.terminalCount!==e.candidate.attemptCount||e.candidateAdaptation.terminalCount!==e.candidateAdaptation.attemptCount?"infrastructure_failure":!e.harnessChanged||!e.lineageValid?"inconclusive":!e.comparison.qualityPassed||e.candidate.passedCount!==e.candidate.attemptCount||e.candidateAdaptation.passedCount!==e.candidateAdaptation.attemptCount?"regressed":e.comparison.improved?"improved":"no_improvement"}h(ht,"classifyComparison");function vr(e){return`model_version_${v({modelId:e,version:0}).slice(0,24)}`}h(vr,"modelVersionId");function $t(e){return(e instanceof Error?e.message:String(e)).trim().slice(0,5e3)||"Harness Refiner benchmark failed."}h($t,"safeError");async function zt(e,t,a){if(!Number.isFinite(t.maximumSpendUsd)||t.maximumSpendUsd<=0)throw new Error("Harness Refiner benchmark maximum spend must be greater than zero.");let n=e.resolveEvaluationProject?await e.resolveEvaluationProject(t.modelId,t.profileId):await lt(e.store,t.modelId,t.profileId);if(n.id!==t.modelId||n.profileId!==t.profileId)throw new Error("The actual evaluation configuration owner changed.");let s=t.advancedEvaluation?Fa(t.advancedEvaluation):null;if(s){if(!e.advancedBoundary||!t.selectedTaskset)throw new Error("The advanced evaluation owner and exact selected Dataset adapter are unavailable.");if(await e.advancedBoundary.authorize(s),s.profileRef.profileId!==t.profileId||s.maximumCostUsd!==t.maximumSpendUsd)throw new Error("Advanced evaluation changed its admitted source or budget.")}let o=s?await e.store.getTaskset(t.selectedTaskset.id):await e.benchmarkTasksets.ensureHarnessRefiner({profileId:t.profileId});if(!o)throw new Error("The exact selected Taskset is unavailable.");if(s&&(o.revision!==t.selectedTaskset.revision||o.contentHash!==t.selectedTaskset.contentHash||o.metadata.importedPackageHash!==s.externalDatasetBinding.packageHash))throw new Error("The advanced selected Dataset package changed.");if(!o.benchmark)throw new Error("Harness Refiner Taskset is unavailable.");if(o.graders.some(c=>c.kind==="model_judge"&&c.rewardEligible&&c.calibrationStatus!=="passed")){if(s)throw new Error("The selected model judge needs its own completed calibration before advanced evaluation.");let c=await e.evaluation.calibrateModelJudges(o.id);if(!c.passed)throw new Error("Harness Refiner model judge did not pass its declared calibration fixtures.");o=c.taskset}let i=ot({taskset:o,seeds:t.seeds,repetitions:t.repetitions,...s?{selectedPopulation:s}:{}}),d=s?await e.store.getHarnessReleaseRecord(s.baselineRelease.contentHash):await Za(e.store);if(s&&d?.harnessRelease.id!==s.baselineRelease.id)throw new Error("The admitted baseline source release is unavailable.");if(!d)throw new Error("A selected local Harness is required.");let g=await e.resolveUpstreamModel(t.model),f=a(),m=s?`advanced-refiner-${v([s.actorId,s.teamId,s.operationId,s.contentHash,t.model,t.seeds,t.repetitions]).slice(0,40)}`:`model_run_${v({modelId:n.id,taskset:o.contentHash,model:t.model,reasoningEffort:t.reasoningEffort,seeds:t.seeds,repetitions:t.repetitions,startedAt:f}).slice(0,24)}`;if(s){let c=await e.store.getModelRun(m);if(c){if(c.evaluation?.benchmarkId!=="harness-refiner"||c.evaluation.advancedEvaluation?.contentHash!==s.contentHash)throw new Error("The advanced operation retains another immutable intent.");return{prepared:c,project:n,existing:!0}}}let r=await Qe({store:e.store,storeDir:e.storeDir,id:`benchmark-${m}`,ownerId:`benchmark:${m}`,name:`Harness Refiner ${m}`,sourceRelease:{id:d.harnessRelease.id,contentHash:d.harnessRelease.contentHash},now:a});await e.store.setHarnessBackgroundReviewSettings({workspaceId:r.workspace.id,enabled:!1,updatedAt:a()});let p=vr(n.id);return{prepared:await e.store.saveModelRun(F.parse({schemaVersion:"openpond.modelRun.v1",id:m,modelId:n.id,modelVersionId:p,profileId:t.profileId,kind:"evaluation",status:"running",method:null,destinationId:null,taskset:{id:o.id,revision:o.revision,contentHash:o.contentHash},harnessRelease:{id:r.release.harnessRelease.id,contentHash:r.release.harnessRelease.contentHash},quote:null,evaluation:{benchmarkId:"harness-refiner",...s?{advancedEvaluation:s}:{},model:t.model,upstreamModel:g,reasoningEffort:t.reasoningEffort,seeds:t.seeds,repetitions:t.repetitions,maximumSpendUsd:t.maximumSpendUsd,attemptPlan:i},evaluationProgress:{stage:"adaptation",completedAttempts:0,totalAttempts:qe(i),accounting:de()},reward:null,receipt:null,adapterArtifactLineageId:null,failure:null,startedAt:f,completedAt:null,updatedAt:f})),project:n,existing:!1}}h(zt,"prepareHarnessRefinerBenchmarkAdmission");import{promises as Ae}from"node:fs";import le from"node:path";function gt(e){let t={schemaVersion:"openpond.harnessRefinerBenchmarkResult.v2",id:`benchmark-result-${e.modelRunId}`,modelRunId:e.modelRunId,benchmarkId:"harness-refiner",model:e.model,upstreamModel:e.upstreamModel,reasoningEffort:e.reasoningEffort,tasksetRelease:e.baseline.tasksetRelease,baseline:e.baseline,adaptation:e.adaptation,refiner:e.refiner,candidateAdaptation:e.candidateAdaptation,candidate:e.candidate,comparison:e.comparison,executionPlan:e.executionPlan,evidenceSnapshot:e.evidenceSnapshot,lineage:e.lineage,publicationPolicy:{primaryReward:"deterministic_output_contract",supplementaryJudge:"not_executed_uncalibrated",diagnosticPasses:1,confirmationPasses:1,uncertainty:"paired_per_case_descriptive"},harness:{baseline:e.baseline.harnessRelease,candidate:e.candidate.harnessRelease},createdAt:e.createdAt};return{...t,contentHash:v(t)}}h(gt,"createResultManifest");async function wt(e,t,a){let n=le.join(e,"training","model-runs",t,"benchmark");await Ae.mkdir(n,{recursive:!0});let s=le.join(n,`${a.contentHash}.json`);return await Ae.writeFile(s,`${JSON.stringify(a,null,2)}
-`,"utf8"),le.relative(e,s).replaceAll(le.sep,"/")}h(wt,"writeManagedResult");async function vt(e,t){let a=le.join(e,"training","model-runs",t,"benchmark"),n;try{n=(await Ae.readdir(a)).filter(o=>o.endsWith(".json"))}catch(o){if(o.code==="ENOENT")return null;throw o}return(await Promise.all(n.map(async o=>{let i=JSON.parse(await Ae.readFile(le.join(a,o),"utf8"));return i.schemaVersion==="openpond.harnessRefinerBenchmarkResult.v2"&&i.modelRunId===t?i:null}))).filter(o=>o!==null).sort((o,i)=>i.createdAt.localeCompare(o.createdAt))[0]??null}h(vt,"loadLatestManagedResult");async function kt(e){if(e.profile.mode!=="local"||!e.profile.repoPath||!e.profile.git?.head)return null;let t=Xe(e.storeDir,e.workspaceId).source,a=await kr(t),n=`benchmarks/harness-refiner/runs/${e.modelRunId}`;return Wa({repoPath:e.profile.repoPath,runId:e.modelRunId,baseCommit:e.profile.git.head,message:`Preserve Harness Refiner benchmark ${e.modelRunId}`,files:[{path:`${n}/result.json`,contents:`${JSON.stringify(e.manifest,null,2)}
-`},...await Promise.all(a.map(async s=>({path:`${n}/candidate-harness/${s}`,contents:await Ae.readFile(le.join(t,...s.split("/")))})))]})}h(kt,"preserveProfileResult");async function kr(e,t=""){let a=le.join(e,...t.split("/").filter(Boolean)),n=await Ae.readdir(a,{withFileTypes:!0}),s=[];for(let o of n.sort((i,d)=>i.name.localeCompare(d.name))){let i=t?`${t}/${o.name}`:o.name;o.isDirectory()?s.push(...await kr(e,i)):o.isFile()&&s.push(i)}return s}h(kr,"listFiles");async function Ut(e){if(!e.modelRun.modelVersionId)throw new Error("A base-Version materialization Run requires a Model Version identity.");let t=await e.store.getModelVersion(e.modelRun.modelVersionId);if(t)return t;if(!("portable"in e.baseline))throw new Error("A portable baseline attempt is required to create the base Model Version.");let a=e.baseline.portable.runManifest,n=e.baseline.portable.tasksetRelease,s={resolvedBundleHash:v({tasksetRelease:n.contentHash,runManifest:a.contentHash}),profileRelease:{id:`profile-release-${e.modelRun.profileId}`,revision:1,contentHash:v({profileId:e.modelRun.profileId})},harnessRelease:a.harnessRelease,agentRelease:{id:e.baseline.portable.agentSnapshot.id,contentHash:e.baseline.portable.agentSnapshot.contentHash},grader:{id:`grader-${n.id}`,contentHash:v(n.graders)}},o={schemaVersion:"openpond.modelVersion.v1",id:e.modelRun.modelVersionId,modelId:e.project.id,profileId:e.project.profileId,version:0,kind:"base_reference",status:"available",baseModel:{schemaVersion:"openpond.baseModelPreference.v1",modelId:`${e.model.providerId}/${e.model.modelId}`,revision:a.model.revision,tokenizerRevision:a.model.tokenizerRevision,chatTemplateHash:a.model.chatTemplateHash,modelAssetId:null,source:"managed"},taskset:e.modelRun.taskset,releaseGraph:s,artifactLineageId:null,adapterStatus:"not_trained",createdAt:e.modelRun.startedAt};return e.store.saveModelVersion(Ua.parse({...o,contentHash:v(o)}))}h(Ut,"ensureBaseVersion");async function Rr(e){let{deps:t,benchmarkInput:a,modelRun:n,signal:s,workspace:o,taskset:i,executionPlan:d,admittedPricing:g,budget:f,totalAttempts:m}=e,[r,p,l]=await Promise.all([ve({store:t.store,modelRunId:n.id,tasksetId:i.id,plan:e.baselinePlan}),ve({store:t.store,modelRunId:n.id,tasksetId:i.id,plan:e.adaptationPlan}),ve({store:t.store,modelRunId:n.id,tasksetId:i.id,plan:e.candidatePlan})]),c=await vt(t.storeDir,n.id);if(!c)throw new Error("Comparison recovery requires the durable benchmark result manifest.");let w=n.evaluationProgress?.accounting?.attempts.filter(G=>G.phase==="candidate_adaptation").map(G=>G.attemptId)??[],u=await ct({store:t.store,tasksetId:i.id,attemptIds:w});if(c.tasksetRelease.id!==r.run.tasksetRelease.id||c.tasksetRelease.contentHash!==r.run.tasksetRelease.contentHash||c.harness.baseline.contentHash!==r.run.harnessRelease.contentHash)throw new Error("Comparison recovery manifest drifted from the admitted benchmark.");let k=o.currentChannel.release;if(!k)throw new Error("Comparison recovery has no candidate Harness release.");let R=await t.store.getHarnessReleaseRecord(k.contentHash);if(!R||R.workspaceId!==o.id)throw new Error("Comparison recovery candidate Harness is unavailable.");let b=await He({workspace:o,release:R}),y=ke(b.release,b.instructionContext),H=await pt({store:t.store,storeDir:t.storeDir,modelRun:n,attempts:[...r.attempts.map(G=>({result:G,cohort:"held_out"})),...p.attempts.map(G=>({result:G,cohort:"adaptation"}))]}),I=await h(async(G,Ue,ye,j)=>{let ue=G.attempts.filter(ee=>ee.grade.score===null||ee.grade.failureClass==="infrastructure_failure").map(ee=>ee.attempt.taskId);if(!ue.length)return G;let At=await t.evaluation.executeBenchmark({tasksetId:i.id,phase:"candidate",model:a.model,reasoningEffort:a.reasoningEffort,seeds:a.seeds,repetitions:a.repetitions,split:Ue.split,taskIds:ue,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:y,hostedTokenPricing:g,parentModelRunId:n.id,signal:s,toolEvidence:ce(H,"replay",j),onAttemptComplete:h(async ee=>{s.throwIfAborted(),f.assertAvailable(`${ye} infrastructure retry`),f.charge(ee.attempt.costUsd,`${ye} infrastructure retry foreground work`);let pe=t.evaluation.consumeGraderUsage([ee.attempt.id]);f.charge(pe.costUsd,`${ye} infrastructure retry grading`),await mt(t.store,n.id,{stage:ye,completedAttempts:m,totalAttempts:m,result:ee,grader:pe,observedSpendUsd:f.observedSpendUsd})},"onAttemptComplete")});return ur({store:t.store,tasksetId:i.id,parentModelRunId:n.id,prior:G,retries:At.attempts,createdAt:t.now()})},"retryStage")(l,e.candidatePlan,"candidate","held_out"),P=Aa({id:`benchmark-comparison-${v([r.run.contentHash,I.run.contentHash]).slice(0,24)}`,baseline:r.run,candidate:I.run,primaryMetric:i.benchmark.primaryMetric,qualityGate:i.benchmark.qualityGate,createdAt:t.now(),metadata:{sourceTasksetId:i.id,benchmarkDefinitionId:i.benchmark.definitionId,recoveredInfrastructureAttempts:!0}});await t.store.saveBenchmarkComparison({tasksetId:i.id,comparison:P});let T={...c.lineage,candidateRelease:{id:c.lineage.candidateRelease.id,contentHash:c.lineage.candidateRelease.contentHash}},S=H.manifest(),D=gt({modelRunId:n.id,model:a.model,upstreamModel:c.upstreamModel,reasoningEffort:a.reasoningEffort,baseline:r.run,adaptation:p.run,refiner:c.refiner,candidateAdaptation:c.candidateAdaptation,candidate:I.run,comparison:P,executionPlan:d,evidenceSnapshot:S,lineage:T,createdAt:t.now()}),A=await wt(t.storeDir,n.id,D),U=await t.loadProfileState(),O=await kt({profile:U,modelRunId:n.id,workspaceId:o.id,storeDir:t.storeDir,manifest:D}),W=dt([...r.attempts,...p.attempts,...u,...I.attempts]),L=r.run.harnessRelease.contentHash!==I.run.harnessRelease.contentHash,E=ht({comparison:P,baseline:r.run,adaptation:p.run,candidate:I.run,candidateAdaptation:c.candidateAdaptation,harnessChanged:L,lineageValid:T.valid,infrastructureValid:W}),Q=ft({baseline:r.run,adaptation:p.run,candidateAdaptation:c.candidateAdaptation,candidate:I.run,harnessChanged:L,lineageValid:T.valid,infrastructureValid:W}),X=(await t.store.getModelRun(n.id)??n).evaluationProgress?.accounting??de(),C=Ke({attempts:X.attempts,targetTaskCount:st(d)}).summary,B=it({baselineTokens:r.run.usage.totalTokens,candidateTokens:I.run.usage.totalTokens,refinerTokens:X.usage.refiner.totalTokens,graderTokens:X.usage.grader.totalTokens}),J={schemaVersion:"openpond.modelEvaluationReceipt.v1",benchmarkId:"harness-refiner",resultManifest:{id:D.id,contentHash:D.contentHash,artifactPath:A},stages:{baseline:{id:r.run.id,contentHash:r.run.contentHash},adaptation:{id:p.run.id,contentHash:p.run.contentHash},candidateAdaptation:{id:c.candidateAdaptation.id,contentHash:c.candidateAdaptation.contentHash},refiner:{id:c.refiner.id,contentHash:c.refiner.contentHash},candidate:{id:I.run.id,contentHash:I.run.contentHash},comparison:{id:P.id,contentHash:P.contentHash}},usage:X.usage,quality:{baselinePassRate:P.baselinePassRate,candidatePassRate:P.candidatePassRate,adaptationBaselinePassRate:p.run.passedCount/p.run.attemptCount,adaptationCandidatePassRate:c.candidateAdaptation.passedCount/c.candidateAdaptation.attemptCount,adaptationCandidatePassed:c.candidateAdaptation.passedCount===c.candidateAdaptation.attemptCount,heldOutCandidatePassed:I.run.passedCount===I.run.attemptCount,passed:P.qualityPassed&&c.candidateAdaptation.passedCount===c.candidateAdaptation.attemptCount&&I.run.passedCount===I.run.attemptCount&&W&&E!=="infrastructure_failure"},foregroundTokenDelta:P.foregroundTokenDelta,foregroundTokenDeltaPercent:P.foregroundTokenDeltaPercent,taskEfficiency:C,efficiency:B,budget:{maximumSpendUsd:f.maximumSpendUsd,observedSpendUsd:f.observedSpendUsd,enforced:!0},evidenceSnapshot:{id:S.id,contentHash:S.contentHash},lineage:T,invalidReasons:Q,attempts:X.attempts,terminalClassification:E,profileGit:O?{ref:O.ref,commit:O.commit,baseCommit:O.baseCommit}:null},_=Ge.parse({...J,contentHash:v(J)}),Z=t.now(),V=await t.store.getModelRun(n.id)??n,ae=E==="infrastructure_failure";return t.store.saveModelRun(F.parse({...V,status:ae?"failed":"succeeded",harnessRelease:r.run.harnessRelease,receipt:_,failure:ae?"Benchmark attempts did not reach the model because the Work runtime failed.":null,evaluationProgress:{...V.evaluationProgress,stage:"comparison",completedAttempts:m,totalAttempts:m},completedAt:Z,updatedAt:Z}))}h(Rr,"resumeHarnessRefinerComparison");async function br(e){let[t,a,n,s]=await Promise.all([e.store.listHarnessImprovementArtifacts(e.workspaceId,"refiner_outcome",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"proposal",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"targeted_validation",1e4),e.store.listHarnessImprovementArtifacts(e.workspaceId,"apply_receipt",1e3)]),o=a,i=n,d=v(e.adaptationAttempts.map(p=>({attempt:p.attempt.id,receipt:p.receiptContentHash,grade:v(p.grade)}))),g=o.every(p=>p.validationPlan.filter(l=>l.required).every(l=>i.some(c=>c.validationId===l.id&&c.status==="passed"))),f=new Set(o.map(p=>p.contentHash)),m=new Set(s.flatMap(p=>{let l="proposal"in p?p.proposal:null;return l&&typeof l=="object"&&"contentHash"in l?[String(l.contentHash)]:[]})),r=[...f].every(p=>m.has(p));return{adaptationEvidenceHash:d,refinerInputHash:e.refinerInputHash,refinerOutcomeHash:v(t),validationHash:v(i),applyReceiptHash:v(s),candidateRelease:{id:e.candidateRelease.id,contentHash:e.candidateRelease.contentHash},valid:t.length===e.completedSteps.filter(p=>p.outcome).length&&g&&r}}h(br,"benchmarkLineage");var mn=100,De=class extends Error{constructor(a,n,s){super(a,s);this.details=n}details;static{h(this,"BenchmarkRefinerInvocationError")}name="BenchmarkRefinerInvocationError"};function un(e){let t=e.artifactManifest.entries.slice(0,mn);return{schemaVersion:"openpond.refinerRewardPacket.v1",attemptRef:e.rewardReceipt.attemptRef,artifactManifest:{ref:{id:e.artifactManifest.id,contentHash:e.artifactManifest.contentHash},entries:t,entryCount:e.artifactManifest.entries.length,truncated:t.length<e.artifactManifest.entries.length},rewardReceipt:e.rewardReceipt,attempt:{status:e.rewardReceipt.outcomeClass,infrastructureError:e.attempt.infrastructureError,outputPresent:typeof e.attempt.output.text=="string"&&e.attempt.output.text.trim().length>0,artifactCount:e.artifactCount,runtimeEventCount:e.attempt.runtimeEventRefs.length,modelRequestCount:Array.isArray(e.attempt.metadata.usage)?e.attempt.metadata.usage.length:e.attempt.metadata.usage?1:0,latencyMs:e.attempt.latencyMs,usage:wr(e.attempt.metadata.usage)}}}h(un,"benchmarkRefinerRewardPacket");async function pn(e){let t=e.result.attempt,a=we(t.metadata,"sessionId"),n=we(t.metadata,"turnId");if(!a||!n)return null;let s=await e.store.getSession(a);if(!s)return null;let o=await et({store:e.store,runId:s.id,workspace:e.runtime.workspace,harnessRelease:{id:e.runtime.release.harnessRelease.id,contentHash:e.runtime.release.harnessRelease.contentHash},admittedAt:t.startedAt}),i=e.taskset.tasks.find(l=>l.id===t.taskId);if(!i)return null;let d=Be.parse({id:n,sessionId:s.id,providerTurnId:null,modelRef:e.model,prompt:hr(i),startedAt:t.startedAt,completedAt:t.completedAt,status:"completed",error:null,metadata:{automatedTasksetWorkAttempt:!0,benchmarkId:"harness-refiner",modelRunId:e.modelRun.id,attemptId:t.id},harnessSnapshot:{schemaVersion:"openpond.harnessTurnSnapshot.v1",workspaceId:e.runtime.workspace.id,workspaceRevision:e.runtime.workspace.revision,sourceRevision:e.runtime.workspace.sourceRevision,channelName:e.runtime.workspace.currentChannel.name,channelRevision:e.runtime.workspace.currentChannel.revision,harnessRelease:o.baseHarnessRelease,overlay:{id:o.id,revision:o.revision,contentHash:o.contentHash}}});await e.store.getTurn(d.id)?await e.store.updateTurn(d.id,l=>Be.parse({...l,modelRef:d.modelRef,metadata:{...l.metadata,...d.metadata},harnessSnapshot:d.harnessSnapshot})):await e.store.insertTurn(d);let f=t.output.text;typeof f=="string"&&f.trim()&&await e.store.appendRuntimeEvent(he({sessionId:s.id,turnId:d.id,name:"assistant.delta",source:"server",appId:s.appId,status:"completed",output:f}));let m=un({attempt:t,artifactManifest:e.result.artifactManifest,rewardReceipt:e.result.rewardReceipt,artifactCount:e.result.artifacts.length}),r=JSON.stringify(m),p=e.result.rewardReceipt.status==="scored"&&e.result.rewardReceipt.passed;return await e.store.appendRuntimeEvent(he({sessionId:s.id,turnId:d.id,name:"diagnostic",source:"server",appId:s.appId,action:"taskset_grade",status:p?"completed":"failed",output:r,error:p?void 0:r,data:{result:{output:r,passed:p,status:e.result.rewardReceipt.status,reward:e.result.rewardReceipt.reward,rewardReceiptRef:{id:e.result.rewardReceipt.id,contentHash:e.result.rewardReceipt.contentHash},artifactManifestRef:{id:e.result.artifactManifest.id,contentHash:e.result.artifactManifest.contentHash}}}})),{session:s,turn:d,result:e.result}}h(pn,"materializeBenchmarkRefinerBoundary");async function yr(e){let t=await pn(e);if(!t)throw new Error(`Adaptation attempt ${e.result.attempt.id} has no Refiner boundary.`);let a=await fn({store:e.store,workspaceId:e.runtime.workspace.id,turnId:t.turn.id}),n=a?{observations:[],trigger:a}:await tt({store:e.store,session:t.session,turn:t.turn,boundaryKind:"turn_completed",now:e.now});if(!n)throw new Error(`Adaptation attempt ${e.result.attempt.id} produced no Refiner detection.`);if(n.trigger.decision!=="queue_refiner")return{detection:n,result:null};e.budget.assertAvailable(`Refiner turn ${n.trigger.id}`);let s=Y(),o=!1,i=0,d=0,g=0,f=e.now(),m=null,r=null;try{m=await at({store:e.store,storeDir:e.storeDir,trigger:n.trigger,...Mt()?{loadActiveRefinerRelease:h(async()=>{let w=Mt().pin,k=(await Ze(e.storeDir)).releases.find(R=>R.id===w.refinerRelease.id&&R.contentHash===w.refinerRelease.contentHash);if(!k)throw new Error("The admitted Refiner release is unavailable.");return k},"loadActiveRefinerRelease")}:{},stream:h(async function*(w){o=!0,i+=1;let u=!1;g+=hn(w.messages,e.admittedPricing);try{for await(let k of e.refinerStream({...w,model:e.model,pricing:e.admittedPricing,requestId:`advanced-refiner:${e.modelRun.id}:${n.trigger.id}:${i}`}))typeof k.costUsd=="number"&&(u=!0),(k.usage!==void 0||k.costUsd!==void 0)&&Ne(s,k.usage,k.costUsd),k.text&&(yield{text:k.text})}finally{u&&(d+=1)}},"stream"),signal:e.signal,timeoutMs:lr,now:e.now})}catch(w){r=w}let p=e.now(),l=o?s.costUsd!==null&&d===i?"authoritative":"estimated":"none",c=l==="authoritative"?s.costUsd:l==="estimated"?Math.max(s.costUsd??0,g):null;if(o)try{e.budget.charge(c,"Refiner"),await gr(e.store,e.modelRun.id,s,e.budget.observedSpendUsd)}catch(w){r??=w}if(r){let w=r instanceof Error?r.message:String(r),u=gn(r,e.signal),k=u!=="cancelled";throw await e.store.appendRuntimeEvent(he({sessionId:t.session.id,turnId:t.turn.id,name:"harness.refiner.failed",source:"server",appId:t.session.appId,status:"failed",output:w,data:{trigger:{id:n.trigger.id,contentHash:n.trigger.contentHash},benchmark:!0,modelRunId:e.modelRun.id,attemptId:e.result.attempt.id,failureKind:u,retryable:k,costBasis:l,estimatedCostUsd:l==="estimated"?g:null}})).catch(()=>{}),new De(w,{trigger:n.trigger,usage:s,costBasis:l,estimatedCostUsd:l==="estimated"?g:null,failureKind:u,retryable:k,startedAt:f,completedAt:p},{cause:r})}return{detection:n,result:m,invocation:{usage:s,costBasis:l,estimatedCostUsd:l==="estimated"?g:null,startedAt:f,completedAt:p}}}h(yr,"runBenchmarkRefinerAfterAttempt");async function fn(e){return(await e.store.listHarnessImprovementArtifacts(e.workspaceId,"trigger_decision",1e3)).find(a=>a.turnId===e.turnId&&a.decision==="queue_refiner")??null}h(fn,"queuedTriggerForTurn");function hn(e,t){let a=e.reduce((s,o)=>s+o.content.length,0);return(Math.ceil(a/3)*t.inputUsdPerMillionTokens+pa*t.outputUsdPerMillionTokens)/1e6}h(hn,"conservativeRefinerRequestCost");function gn(e,t){if(t.aborted)return"cancelled";let a=e instanceof Error?e.message:String(e);return/timed out|timeout/i.test(a)?"timeout":/invalid structured output|response limit|JSON/i.test(a)?"invalid_output":/provider|rate limit|billing|signed out|network|fetch/i.test(a)?"provider_failure":"unknown"}h(gn,"benchmarkRefinerFailureKind");import{promises as Rt}from"node:fs";import Ir from"node:path";function Lt(e){let t={schemaVersion:"openpond.benchmarkRefinerInvocation.v1",id:`benchmark-refiner-invocation-${v({taskId:e.taskId,attemptId:e.attemptId,invocationOrdinal:e.invocationOrdinal,trigger:e.trigger,status:e.status,startedAt:e.startedAt}).slice(0,24)}`,...e};return{...t,contentHash:v(t)}}h(Lt,"createBenchmarkRefinerInvocationReceipt");async function bt(e){let t=Er(e.storeDir,e.modelRunId),a;try{a=JSON.parse(await Rt.readFile(t,"utf8"))}catch(n){if(n.code==="ENOENT")return null;throw n}return wn(a,e.modelRunId)}h(bt,"loadSequentialAdaptationCheckpoint");async function Vt(e){let t=await bt(e);if(t){if(!vn(t.initialHarness,e.initialHarness))throw new Error("Sequential adaptation checkpoint changed its initial Harness release.");return t}return Hr({storeDir:e.storeDir,checkpoint:Sr({modelRunId:e.modelRunId,initialHarness:e.initialHarness,steps:[],invocations:[],updatedAt:e.now})})}h(Vt,"initializeSequentialAdaptationCheckpoint");async function Wt(e){let t=await Vt({storeDir:e.storeDir,modelRunId:e.modelRunId,initialHarness:e.initialHarness,now:e.now});if(!e.invocation&&!e.step)throw new Error("A sequential checkpoint append requires an invocation or step.");let a=e.invocation?[...t.invocations.filter(s=>s.id!==e.invocation.id),e.invocation]:t.invocations,n=e.step?[...t.steps.filter(s=>s.taskId!==e.step.taskId),e.step].sort((s,o)=>s.ordinal-o.ordinal):t.steps;return Hr({storeDir:e.storeDir,checkpoint:Sr({modelRunId:t.modelRunId,initialHarness:t.initialHarness,steps:n,invocations:a,updatedAt:e.now})})}h(Wt,"appendSequentialAdaptationInvocation");function Sr(e){let t={schemaVersion:"openpond.sequentialAdaptationCheckpoint.v1",...e};return{...t,contentHash:v(t)}}h(Sr,"checkpoint");async function Hr(e){let t=Er(e.storeDir,e.checkpoint.modelRunId),a=Ir.dirname(t);await Rt.mkdir(a,{recursive:!0});let n=`${t}.${process.pid}.tmp`;return await Rt.writeFile(n,`${JSON.stringify(e.checkpoint,null,2)}
-`,"utf8"),await Rt.rename(n,t),e.checkpoint}h(Hr,"writeSequentialAdaptationCheckpoint");function wn(e,t){if(!e||typeof e!="object"||Array.isArray(e))throw new Error("Sequential adaptation checkpoint is malformed.");let a=e;if(a.schemaVersion!=="openpond.sequentialAdaptationCheckpoint.v1"||a.modelRunId!==t||!a.initialHarness||!Array.isArray(a.steps)||!Array.isArray(a.invocations)||typeof a.updatedAt!="string"||typeof a.contentHash!="string")throw new Error("Sequential adaptation checkpoint is malformed.");let{contentHash:n,...s}=a;if(v(s)!==n)throw new Error("Sequential adaptation checkpoint failed content-hash validation.");let o=a;for(let i of o.invocations){let{contentHash:d,...g}=i;if(v(g)!==d)throw new Error(`Refiner invocation ${i.id} failed content-hash validation.`)}return o}h(wn,"parseCheckpoint");function Er(e,t){return Ir.join(e,"training","harness-refiner-benchmarks",t,"sequential-adaptation.json")}h(Er,"checkpointPath");function vn(e,t){return e.id===t.id&&e.contentHash===t.contentHash}h(vn,"sameRef");async function Ar(e){let t=e.adapters?.runRefinerAfterAttempt??yr,a=e.adapters?.loadRuntimeFromRelease??He,n=e.adapters?.buildLineage??br,s=e.adapters?.loadAttemptEvidenceByIds??ct;await e.store.setHarnessBackgroundReviewSettings({workspaceId:e.initialRuntime.workspace.id,enabled:!0,updatedAt:e.now()});let o={id:e.initialRuntime.release.harnessRelease.id,contentHash:e.initialRuntime.release.harnessRelease.contentHash},i=await Vt({storeDir:e.storeDir,modelRunId:e.modelRun.id,initialHarness:o,now:e.now()});kn(i.steps,e.taskIds);let d=await Rn({store:e.store,initialRuntime:e.initialRuntime,steps:i.steps,loadRuntimeFromRelease:a}),g=i.steps.length?await s({store:e.store,tasksetId:e.taskset.id,attemptIds:i.steps.map(y=>y.attemptId)}):[],f=new Map(g.map(y=>[y.attempt.id,y])),m=[],r=[...i.steps];try{for(let[y,H]of e.taskIds.entries()){if(y<r.length)continue;e.signal.throwIfAborted();let x={id:d.release.harnessRelease.id,contentHash:d.release.harnessRelease.contentHash},I=i.invocations.filter(E=>E.taskId===H),P=I.at(-1);if(P?.status==="failed"&&I.length>=jt)throw new Error(`Sequential Refiner exhausted ${jt} admitted invocations for ${H}.`);let T=P?.status==="failed"?f.get(P.attemptId):void 0;if(!T&&P?.status==="failed"&&([T]=await s({store:e.store,tasksetId:e.taskset.id,attemptIds:[P.attemptId]}),T&&(g.push(T),f.set(T.attempt.id,T))),!T){let E=await e.evaluation.execute({tasksetId:e.taskset.id,taskId:H,model:e.model,reasoningEffort:e.reasoningEffort,seed:e.seed,attempt:0,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:ke(d.release,d.instructionContext),hostedTokenPricing:e.admittedPricing,parentModelRunId:e.modelRun.id,signal:e.signal,toolEvidence:ce(e.evidenceSnapshot,"replay","adaptation")});await e.onAttemptComplete(E),T={attempt:E.attempt,grade:E.grade,artifacts:E.artifacts,receiptContentHash:E.portable.receipt.contentHash,artifactManifest:E.portable.artifactManifest,rewardReceipt:E.portable.rewardReceipt},g.push(T),f.set(T.attempt.id,T)}let S;try{S=await t({store:e.store,storeDir:e.storeDir,modelRun:e.modelRun,model:e.model,taskset:e.taskset,runtime:d,result:T,budget:e.budget,admittedPricing:e.admittedPricing,refinerStream:e.refinerStream,signal:e.signal,now:e.now})}catch(E){throw E instanceof De&&(i=await Wt({storeDir:e.storeDir,modelRunId:e.modelRun.id,initialHarness:o,invocation:Lt({taskId:H,attemptId:T.attempt.id,invocationOrdinal:I.length,trigger:{id:E.details.trigger.id,contentHash:E.details.trigger.contentHash},status:"failed",outcome:null,failure:{kind:E.details.failureKind,message:E.message.slice(0,2e3),retryable:E.details.retryable},inputHarness:x,outputHarness:x,usage:E.details.usage,costBasis:E.details.costBasis,estimatedCostUsd:E.details.estimatedCostUsd,startedAt:E.details.startedAt,completedAt:E.details.completedAt}),now:E.details.completedAt})),E}S.result&&m.push(S.result);let D=await e.store.getHarnessWorkspace(d.workspace.id),A=D?.currentChannel.release;if(!D||!A)throw new Error("Sequential Refiner did not retain a current Harness release.");let U=await e.store.getHarnessReleaseRecord(A.contentHash);if(!U)throw new Error("Sequential Refiner release is unavailable.");d=await a({workspace:D,release:U});let O={id:d.release.harnessRelease.id,contentHash:d.release.harnessRelease.contentHash},W={ordinal:y,taskId:H,attemptId:T.attempt.id,inputHarness:x,outputHarness:O,trigger:{id:S.detection.trigger.id,contentHash:S.detection.trigger.contentHash,decision:S.detection.trigger.decision},outcome:S.result?{id:S.result.outcome.id,contentHash:S.result.outcome.contentHash,decision:S.result.outcome.decision}:null,changed:x.contentHash!==O.contentHash};r.push(W);let L=S.result&&S.invocation?Lt({taskId:H,attemptId:T.attempt.id,invocationOrdinal:I.length,trigger:{id:S.detection.trigger.id,contentHash:S.detection.trigger.contentHash},status:"completed",outcome:{id:S.result.outcome.id,contentHash:S.result.outcome.contentHash,decision:S.result.outcome.decision},failure:null,inputHarness:x,outputHarness:O,usage:S.invocation.usage,costBasis:S.invocation.costBasis,estimatedCostUsd:S.invocation.estimatedCostUsd,startedAt:S.invocation.startedAt,completedAt:S.invocation.completedAt}):void 0;i=await Wt({storeDir:e.storeDir,modelRunId:e.modelRun.id,initialHarness:o,invocation:L,step:W,now:e.now()})}}finally{await e.store.setHarnessBackgroundReviewSettings({workspaceId:e.initialRuntime.workspace.id,enabled:!1,updatedAt:e.now()})}let p=g.reduce((y,H)=>{let x=Array.isArray(H.attempt.metadata.usage)?H.attempt.metadata.usage:[H.attempt.metadata.usage];for(let I of x){if(!I||typeof I!="object"||Array.isArray(I))continue;let P=I,T=Jt(P,["promptTokens","prompt_tokens","inputTokens","input_tokens"]),S=Jt(P,["completionTokens","completion_tokens","outputTokens","output_tokens"]);y.inputTokens+=T,y.outputTokens+=S,y.totalTokens+=Jt(P,["totalTokens","total_tokens"])||T+S}return y},{inputTokens:0,outputTokens:0,totalTokens:0}),l=g.flatMap(y=>typeof y.attempt.costUsd=="number"?[y.attempt.costUsd]:[]),c={id:d.release.harnessRelease.id,contentHash:d.release.harnessRelease.contentHash},w={schemaVersion:"openpond.sequentialHarnessAdaptation.v1",id:`sequential-adaptation-${e.modelRun.id}`,initialHarness:o,finalHarness:c,attemptCount:g.length,passedCount:g.filter(y=>y.grade.passed).length,terminalCount:g.filter(y=>!y.attempt.infrastructureError).length,usage:p,costUsd:l.length?l.reduce((y,H)=>y+H,0):null,latencyMs:g.reduce((y,H)=>y+H.attempt.latencyMs,0),steps:r,createdAt:e.now()},u={...w,contentHash:v(w)},k=await e.store.listHarnessImprovementArtifacts(e.initialRuntime.workspace.id,"refiner_outcome",1e3),R={id:`benchmark-refiner-${e.modelRun.id}`,contentHash:v(k),outcomeCount:k.length},b=await n({store:e.store,workspaceId:e.initialRuntime.workspace.id,adaptationAttempts:g,completedSteps:r,candidateRelease:c,refinerInputHash:v(g.map(y=>({attempt:y.attempt.id,receipt:y.receiptContentHash,grade:v(y.grade)})))});return{attempts:g,runtime:d,summary:u,refinerStage:R,lineage:b,refinerResults:m}}h(Ar,"runSequentialHarnessAdaptation");function kn(e,t){for(let[a,n]of e.entries())if(n.ordinal!==a||n.taskId!==t[a])throw new Error("Sequential adaptation checkpoint drifted from the admitted task order.")}h(kn,"assertCheckpointTaskOrder");async function Rn(e){let t=e.steps.at(-1);if(!t)return e.initialRuntime;let a=await e.store.getHarnessWorkspace(e.initialRuntime.workspace.id);if(!a?.currentChannel.release||a.currentChannel.release.id!==t.outputHarness.id||a.currentChannel.release.contentHash!==t.outputHarness.contentHash)throw new Error("Sequential adaptation workspace drifted from its durable checkpoint.");let n=await e.store.getHarnessReleaseRecord(t.outputHarness.contentHash);if(!n||n.harnessRelease.id!==t.outputHarness.id)throw new Error("Sequential adaptation checkpoint release is unavailable.");return e.loadRuntimeFromRelease({workspace:a,release:n})}h(Rn,"restoreCheckpointRuntime");function Jt(e,t){for(let a of t){let n=e[a];if(typeof n=="number"&&Number.isFinite(n)&&n>=0)return Math.trunc(n)}return 0}h(Jt,"numericToken");var Gt=class extends Error{static{h(this,"BenchmarkRunCancelledError")}constructor(){super("Benchmark cancelled by operator."),this.name="BenchmarkRunCancelledError"}},me=new Map;function Ts(e){let t=e.now??(()=>new Date().toISOString());async function a(f){let{prepared:m,project:r,existing:p}=await zt(e,f,t);if(p&&m.status!=="prepared")return m;let l=p?await e.store.saveModelRun(F.parse({...m,status:"running",updatedAt:t()})):m,c=l.id,w=new AbortController,u=d({input:f,project:r,modelRun:l,signal:w.signal}).finally(()=>me.delete(c));return me.set(c,{controller:w,execution:u}),u.catch(()=>{}),l}h(a,"start");async function n(f){let m=await zt(e,f,t);return m.existing?m.prepared:e.store.saveModelRun(F.parse({...m.prepared,status:"prepared",updatedAt:t()}))}h(n,"prepare");async function s(f){let m=await e.store.getModelRun(f);if(!m||m.kind!=="evaluation"||m.evaluation?.benchmarkId!=="harness-refiner")throw new Error("No Harness Refiner evaluation exists for this Model Run.");if(m.status!=="running")return m;me.get(f)?.controller.abort(new Gt);let r=t();return e.store.saveModelRun(F.parse({...m,status:"cancelled",failure:"Benchmark cancelled by operator.",completedAt:r,updatedAt:r}))}h(s,"cancel");async function o(f){let m=await e.store.getModelRun(f);if(!m||m.kind!=="evaluation"||m.evaluation?.benchmarkId!=="harness-refiner")throw new Error("No Harness Refiner evaluation exists for this Model Run.");if(me.has(f)||m.status==="running")return m;let r=m.evaluationProgress?.stage==="refiner"&&m.evaluationProgress?.completedAttempts===ne(m.evaluation.attemptPlan,"candidate_adaptation"),p=ne(m.evaluation.attemptPlan,"candidate_adaptation"),l=p+Ee(m.evaluation.attemptPlan,"candidate_adaptation").attemptCount,c=await bt({storeDir:e.storeDir,modelRunId:m.id}),w=m.evaluationProgress?.stage==="candidate_adaptation"&&m.evaluationProgress.completedAttempts>=p&&m.evaluationProgress.completedAttempts<=l&&!!c,k=["candidate_adaptation","candidate","comparison"].includes(m.evaluationProgress?.stage??"")&&m.evaluationProgress?.completedAttempts===qe(m.evaluation.attemptPlan)&&!!await vt(e.storeDir,m.id);if(!["failed","cancelled"].includes(m.status)||m.evaluation.advancedEvaluation?.mode!=="review_quality"&&!r&&!w&&!k)throw new Error("Only a Harness Refiner run with a durable Refiner, sequential adaptation, or comparison checkpoint can resume.");let R=e.resolveEvaluationProject?await e.resolveEvaluationProject(m.modelId,m.profileId):await lt(e.store,m.modelId,m.profileId),b=await e.store.saveModelRun(F.parse({...m,status:"running",receipt:null,failure:null,completedAt:null,updatedAt:t()})),y={modelId:m.modelId,profileId:m.profileId,model:m.evaluation.model,reasoningEffort:m.evaluation.reasoningEffort==="none"?"none":m.evaluation.reasoningEffort===null?null:Ma.parse(m.evaluation.reasoningEffort),seeds:m.evaluation.seeds,repetitions:m.evaluation.repetitions,maximumSpendUsd:m.evaluation.maximumSpendUsd,...m.evaluation.advancedEvaluation?{advancedEvaluation:m.evaluation.advancedEvaluation,selectedTaskset:m.taskset}:{}},H=new AbortController,x=d({input:y,project:R,modelRun:b,signal:H.signal}).finally(()=>me.delete(f));return me.set(f,{controller:H,execution:x}),x.catch(()=>{}),b}h(o,"resume");async function i(){let f=(await e.store.listModelRuns()).filter(r=>r.kind==="evaluation"&&r.evaluation?.benchmarkId==="harness-refiner"&&r.status==="running"),m=t();for(let r of f)await e.store.saveModelRun(F.parse({...r,status:"failed",failure:"Benchmark execution was interrupted when the app server stopped. Resume it when an exact durable Refiner checkpoint is available; otherwise start a new run to preserve a complete causal record.",completedAt:m,updatedAt:m}));return f.length}h(i,"reconcileInterrupted");async function d(f){let m=f.modelRun.evaluation?.benchmarkId==="harness-refiner"?f.modelRun.evaluation.advancedEvaluation:void 0;if(!m)return g(f);if(!e.advancedBoundary)throw new Error("Advanced paid admission is unavailable.");let r=AbortSignal.any([f.signal,AbortSignal.timeout(m.maximumDurationMs)]);try{let p=await e.advancedBoundary.run(f.modelRun.id,m,()=>m.mode==="review_quality"?e.reviewQuality?.({modelRun:f.modelRun,signal:r})??Promise.reject(new Error("The actual review-quality evaluator is unavailable.")):g({...f,signal:r}));return e.store.saveModelRun(F.parse({...p,evaluationProgress:p.evaluationProgress?{...p.evaluationProgress,advancedAccounting:e.advancedBoundary.accounting(p.id,m)}:null}))}catch(p){let l=await e.store.getModelRun(f.modelRun.id);if(!l)throw p;let c=t();return e.store.saveModelRun(F.parse({...l,status:r.aborted||l.status==="cancelled"?"cancelled":"failed",failure:$t(p),evaluationProgress:l.evaluationProgress?{...l.evaluationProgress,advancedAccounting:e.advancedBoundary.accounting(l.id,m)}:null,completedAt:c,updatedAt:c}))}}h(d,"executePass");async function g(f){let{input:m,modelRun:r}=f;try{f.signal.throwIfAborted();let p=r.evaluation;if(p?.benchmarkId!=="harness-refiner")throw new Error("Harness Refiner Model Run has the wrong evaluation configuration.");if(!r.harnessRelease)throw new Error("Harness Refiner Model Run has no admitted Harness release.");let l=`benchmark-${r.id}`,[c,w]=await Promise.all([e.store.getHarnessWorkspace(l),e.store.getHarnessReleaseRecord(r.harnessRelease.contentHash)]);if(!c||!w||w.workspaceId!==c.id)throw new Error("The Harness pinned at admission is unavailable.");let u={workspace:c,release:w},k=await e.store.getTaskset(r.taskset.id);if(!k?.benchmark)throw new Error("Harness Refiner Taskset changed before execution.");if(k.revision!==r.taskset.revision||k.contentHash!==r.taskset.contentHash)throw new Error("Harness Refiner Taskset release drifted after admission.");let R=ot({taskset:k,seeds:m.seeds,repetitions:m.repetitions,...p.advancedEvaluation?{selectedPopulation:p.advancedEvaluation}:{}});if(v(R)!==v(p.attemptPlan))throw new Error("Harness Refiner execution plan changed after admission.");let b=p.upstreamModel;if(!b?.pricing)throw new Error("Harness Refiner run has no admitted upstream pricing.");let y=b.pricing,H=qe(R),x=Ee(R,"baseline"),I=Ee(R,"adaptation"),P=Ee(R,"candidate_adaptation"),T=Ee(R,"candidate"),S=r.evaluationProgress?.stage==="refiner"&&r.evaluationProgress.completedAttempts===ne(R,"candidate_adaptation"),D=ne(R,"candidate_adaptation"),A=D+P.attemptCount,U=await bt({storeDir:e.storeDir,modelRunId:r.id}),O=r.evaluationProgress?.stage==="candidate_adaptation"&&r.evaluationProgress.completedAttempts>=D&&r.evaluationProgress.completedAttempts<=A&&!!U,W=["candidate_adaptation","candidate","comparison"].includes(r.evaluationProgress?.stage??"")&&r.evaluationProgress?.completedAttempts===H,L=new nt(m.maximumSpendUsd,S||O||W?r.evaluationProgress?.accounting?.observedSpendUsd??0:0);if(W)return Rr({deps:{store:e.store,storeDir:e.storeDir,evaluation:e.evaluation,loadProfileState:e.loadProfileState,now:t},benchmarkInput:m,modelRun:r,signal:f.signal,workspace:u.workspace,taskset:k,executionPlan:R,baselinePlan:x,adaptationPlan:I,candidatePlan:T,admittedPricing:y,budget:L,totalAttempts:H});let E=new ge,Q=h((te,Te)=>{let ra=r.evaluationProgress?.accounting?.attempts.filter(Ce=>Ce.phase===te).length??0;return async Ce=>{f.signal.throwIfAborted(),L.assertAvailable(Te),L.charge(Ce.attempt.costUsd,`${Te} foreground work`);let na=e.evaluation.consumeGraderUsage([Ce.attempt.id]);L.charge(na.costUsd,`${Te} grading`),ra+=1,await mt(e.store,r.id,{stage:te,completedAttempts:ne(R,te)+ra,totalAttempts:H,result:Ce,grader:na,observedSpendUsd:L.observedSpendUsd})}},"observeStageAttempts"),$=await He({workspace:u.workspace,release:u.release}),X=ke($.release,$.instructionContext),C,B,J;if(S||O)C=await ve({store:e.store,modelRunId:r.id,tasksetId:k.id,plan:x}),B=await ve({store:e.store,modelRunId:r.id,tasksetId:k.id,plan:I}),J=B.attempts,E=await pt({store:e.store,storeDir:e.storeDir,modelRun:r,attempts:[...C.attempts.map(te=>({result:te,cohort:"held_out"})),...B.attempts.map(te=>({result:te,cohort:"adaptation"}))]});else{let te=await e.evaluation.executeBenchmark({tasksetId:k.id,phase:"baseline",model:m.model,reasoningEffort:m.reasoningEffort,seeds:m.seeds,repetitions:m.repetitions,split:I.split,taskIds:I.taskIds,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:X,hostedTokenPricing:y,parentModelRunId:r.id,signal:f.signal,toolEvidence:ce(E,"record","adaptation"),onAttemptComplete:Q("adaptation","adaptation baseline")});if(B=Ft(te),J=B.attempts,await Ut({store:e.store,project:f.project,modelRun:r,model:m.model,baseline:te.attempts[0]}),B.run.harnessRelease.id!==r.harnessRelease.id||B.run.harnessRelease.contentHash!==r.harnessRelease.contentHash)throw new Error("Adaptation baseline drifted from the admitted Harness release.");await ut(e.store,e.storeDir,r.id,E.manifest()),await ie(e.store,r.id,{stage:"baseline",completedAttempts:ne(R,"baseline"),totalAttempts:H});let Te=await e.evaluation.executeBenchmark({tasksetId:k.id,phase:"baseline",model:m.model,reasoningEffort:m.reasoningEffort,seeds:m.seeds,repetitions:m.repetitions,split:x.split,taskIds:x.taskIds,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:X,hostedTokenPricing:y,parentModelRunId:r.id,signal:f.signal,toolEvidence:ce(E,"record","held_out"),onAttemptComplete:Q("baseline","held-out baseline")});if(C=Ft(Te),C.run.harnessRelease.id!==r.harnessRelease.id||C.run.harnessRelease.contentHash!==r.harnessRelease.contentHash)throw new Error("Held-out baseline drifted from the admitted Harness release.");await ut(e.store,e.storeDir,r.id,E.manifest()),await ie(e.store,r.id,{stage:"refiner",completedAttempts:ne(R,"candidate_adaptation"),totalAttempts:H})}if((S||O)&&!r.evaluationProgress?.evidenceSnapshot)throw new Error("Sequential adaptation cannot resume because the interrupted run did not preserve its exact frozen evidence snapshot. Start a new run to preserve a valid causal comparison.");O||await ie(e.store,r.id,{stage:"candidate_adaptation",completedAttempts:D,totalAttempts:H});let _=await Ar({store:e.store,storeDir:e.storeDir,evaluation:e.evaluation,modelRun:r,model:m.model,reasoningEffort:m.reasoningEffort,taskset:k,taskIds:P.taskIds,seed:m.seeds[0],initialRuntime:$,evidenceSnapshot:E,budget:L,admittedPricing:y,refinerStream:e.refinerStream,signal:f.signal,onAttemptComplete:Q("candidate_adaptation","sequential adaptation treatment"),now:t}),Z=_.runtime,V=_.refinerStage,ae=_.lineage,G=E.manifest(),Ue=C.run.harnessRelease.contentHash!==_.summary.finalHarness.contentHash,ye=ke(Z.release,Z.instructionContext);await ie(e.store,r.id,{stage:"candidate",completedAttempts:ne(R,"candidate"),totalAttempts:H});let j=await e.evaluation.executeBenchmark({tasksetId:k.id,phase:"candidate",model:m.model,reasoningEffort:m.reasoningEffort,seeds:m.seeds,repetitions:m.repetitions,split:T.split,taskIds:T.taskIds,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:ye,hostedTokenPricing:y,parentModelRunId:r.id,signal:f.signal,toolEvidence:ce(E,"replay","held_out"),onAttemptComplete:Q("candidate","held-out candidate")});if(!j.comparison)throw new Error("Paired benchmark comparison was not produced.");await ie(e.store,r.id,{stage:"comparison",completedAttempts:H,totalAttempts:H});let ue=gt({modelRunId:r.id,model:m.model,upstreamModel:p.upstreamModel,reasoningEffort:m.reasoningEffort,baseline:C.run,adaptation:B.run,refiner:V,candidateAdaptation:_.summary,candidate:j.run,comparison:j.comparison,executionPlan:R,evidenceSnapshot:G,lineage:ae,createdAt:t()}),At=await wt(e.storeDir,r.id,ue),ee=await e.loadProfileState(),pe=await kt({profile:ee,modelRunId:r.id,workspaceId:u.workspace.id,storeDir:e.storeDir,manifest:ue});await Ut({store:e.store,project:f.project,modelRun:r,model:m.model,baseline:C.attempts[0]});let xt=dt([...C.attempts,...J,..._.attempts,...j.attempts]),Pt=ht({comparison:j.comparison,baseline:C.run,adaptation:B.run,candidate:j.run,harnessChanged:Ue,candidateAdaptation:_.summary,lineageValid:ae.valid,infrastructureValid:xt}),zr=ft({baseline:C.run,adaptation:B.run,candidateAdaptation:_.summary,candidate:j.run,harnessChanged:Ue,lineageValid:ae.valid,infrastructureValid:xt}),Pe=(await e.store.getModelRun(r.id)??r).evaluationProgress?.accounting??de(),Ur=Ke({attempts:Pe.attempts,targetTaskCount:st(R)}).summary,Lr=it({baselineTokens:C.run.usage.totalTokens,candidateTokens:j.run.usage.totalTokens,refinerTokens:Pe.usage.refiner.totalTokens,graderTokens:Pe.usage.grader.totalTokens}),Zt={schemaVersion:"openpond.modelEvaluationReceipt.v1",benchmarkId:"harness-refiner",resultManifest:{id:ue.id,contentHash:ue.contentHash,artifactPath:At},stages:{baseline:{id:C.run.id,contentHash:C.run.contentHash},adaptation:{id:B.run.id,contentHash:B.run.contentHash},candidateAdaptation:{id:_.summary.id,contentHash:_.summary.contentHash},refiner:{id:V.id,contentHash:V.contentHash},candidate:{id:j.run.id,contentHash:j.run.contentHash},comparison:{id:j.comparison.id,contentHash:j.comparison.contentHash}},usage:Pe.usage,quality:{baselinePassRate:j.comparison.baselinePassRate,candidatePassRate:j.comparison.candidatePassRate,adaptationBaselinePassRate:B.run.passedCount/B.run.attemptCount,adaptationCandidatePassRate:_.summary.passedCount/_.summary.attemptCount,adaptationCandidatePassed:_.summary.passedCount===_.summary.attemptCount,heldOutCandidatePassed:j.run.passedCount===j.run.attemptCount,passed:j.comparison.qualityPassed&&_.summary.passedCount===_.summary.attemptCount&&j.run.passedCount===j.run.attemptCount&&xt&&Pt!=="infrastructure_failure"},foregroundTokenDelta:j.comparison.foregroundTokenDelta,foregroundTokenDeltaPercent:j.comparison.foregroundTokenDeltaPercent,taskEfficiency:Ur,efficiency:Lr,budget:{maximumSpendUsd:L.maximumSpendUsd,observedSpendUsd:L.observedSpendUsd,enforced:!0},evidenceSnapshot:{id:G.id,contentHash:G.contentHash},lineage:ae,invalidReasons:zr,attempts:Pe.attempts,terminalClassification:Pt,profileGit:pe?{ref:pe.ref,commit:pe.commit,baseCommit:pe.baseCommit}:null},Vr=Ge.parse({...Zt,contentHash:v(Zt)}),ea=t(),ta=await e.store.getModelRun(r.id)??r,aa=Pt==="infrastructure_failure";return e.store.saveModelRun(F.parse({...ta,status:aa?"failed":"succeeded",harnessRelease:C.run.harnessRelease,receipt:Vr,failure:aa?"Benchmark attempts did not reach the model because the Work runtime failed.":null,evaluationProgress:{...ta.evaluationProgress,stage:"comparison",completedAttempts:H,totalAttempts:H},completedAt:ea,updatedAt:ea}))}catch(p){let l=t(),c=await e.store.getModelRun(r.id)??r,w=f.signal.aborted||c.status==="cancelled";return e.store.saveModelRun(F.parse({...c,status:w?"cancelled":"failed",failure:w?"Benchmark cancelled by operator.":$t(p),completedAt:l,updatedAt:l}))}}return h(g,"executePassInternal"),{async close(){let f=[...me.values()];for(let m of f)m.controller.abort(new Error("The Refiner evaluation owner is closing."));await Promise.allSettled(f.map(m=>m.execution))},reconcileInterrupted:i,prepare:n,cancel:s,resume:o,start:a,wait:h(async f=>me.get(f)?.execution??e.store.getModelRun(f),"wait")}}h(Ts,"createHarnessRefinerBenchmarkService");import{mkdir as bn,readFile as yn,rename as In,writeFile as Sn}from"node:fs/promises";import xr from"node:path";import{randomUUID as Hn}from"node:crypto";var En=M.object({outcome:ma,proposal:ia.nullable(),validations:M.array(da),overlay:sa,workspace:Le,advanceReceipt:ca.nullable(),applyReceipt:ua.nullable()}),Pr=M.object({runId:M.string(),taskId:M.string(),actorId:M.string(),teamId:M.string(),pinHash:M.string(),sourceCaseHash:M.string(),sessionId:M.string(),turnId:M.string(),workspaceId:M.string(),trigger:la,result:En.nullable(),receipt:$a.nullable(),cleanupComplete:M.boolean(),contentHash:M.string()});function Tr(e){let t=h((s,o)=>xr.join(e,"training","advanced-refiner-reviews",v([s,o])+".json"),"file");async function a(s,o,i){let d;try{d=JSON.parse(await yn(t(s,o),"utf8"))}catch(r){if(r.code==="ENOENT")return null;throw r}let g=Pr.parse(d),{contentHash:f,...m}=g;if(f!==v(m)||g.runId!==s||g.taskId!==o||g.actorId!==i.actorId||g.teamId!==i.teamId||g.pinHash!==i.contentHash)throw new Error("The retained review case belongs to another immutable admission.");return g}h(a,"read");async function n(s){let o=Pr.parse({...s,contentHash:v(s)}),i=t(o.runId,o.taskId),d=i+"."+Hn();return await bn(xr.dirname(i),{recursive:!0}),await Sn(d,JSON.stringify(o),{mode:384}),await In(d,i),o}return h(n,"write"),{read:a,write:n}}h(Tr,"createAdvancedReviewCheckpoints");function Ys(e){let t=Tr(e.benchmark.storeDir);return async a=>{let n=a.modelRun,s=n.evaluation?.benchmarkId==="harness-refiner"?n.evaluation.advancedEvaluation:void 0;if(!s||s.mode!=="review_quality")throw new Error("Review-quality admission is unavailable.");let o=await e.source.read(s),i=await e.benchmark.store.getTaskset(n.taskset.id);if(!i)throw new Error("The exact review Taskset is unavailable.");let d=o.packageValue.taskset.metadata.advancedRefinerReview;if(d?.outputConvention!=="openpond.refinerReviewOutput.v1"||v(d.contextEvidence)!==v(s.evidence))throw new Error("Publish a review-quality Dataset with independently grounded labels and the exact retained context-evidence binding before starting this target.");let g=await e.source.evidence(s),f=(await Ze(e.benchmark.storeDir)).releases.find(u=>u.id===s.refinerRelease.id&&u.contentHash===s.refinerRelease.contentHash);if(!f)throw new Error("The admitted Refiner release is unavailable.");let m=[];for(let u of s.holdoutTaskIds){a.signal.throwIfAborted(),await o.authorize();let k=i.tasks.find(S=>S.id===u),R=g.result.cases.find(S=>S.identity.caseId===u);if(!k||!R||R.status!=="completed")throw new Error("The exact independent review context is unavailable.");let b=await t.read(n.id,u,s);if(b&&b.sourceCaseHash!==v(R))throw new Error("The retained independent review context changed.");if(b?.receipt&&b.cleanupComplete){m.push(b.receipt);continue}if(b?.cleanupComplete&&!b.result&&b.trigger.decision==="queue_refiner")throw new Error("This interrupted Refiner case cleaned its overlay before receiving a settled outcome; reconcile the retained paid request before resuming.");let y=await e.source.reviewContext(s,u),H=await e.benchmark.store.getHarnessWorkspace(`review-${v([n.id,u]).slice(0,40)}`),x=H?.currentChannel.release&&await e.benchmark.store.getHarnessReleaseRecord(H.currentChannel.release.contentHash),I=H&&x?{workspace:H,release:x}:await Qe({store:e.benchmark.store,storeDir:e.benchmark.storeDir,id:`review-${v([n.id,u]).slice(0,40)}`,ownerId:`benchmark:${n.id}`,name:`Review ${u}`,sourceRelease:s.baselineRelease}),P=null,T=!1;try{let S=await e.benchmark.store.sessionShells(),A=(b?await e.benchmark.store.getSession(b.sessionId):S.find(C=>C.metadata?.advancedRunId===n.id&&C.metadata?.advancedTaskId===u&&C.metadata?.actorId===s.actorId&&C.cloudTeamId===s.teamId))??await e.createSession({experience:"work",provider:"openpond",modelRef:n.evaluation.model,hiddenFromDefaultSidebar:!0,openPondCommandAccessMode:"disabled",cwd:null,title:`Evaluate Refiner ${u}`,cloudTeamId:s.teamId,metadata:{source:"advanced-refiner-evaluation",automatedTasksetWorkAttempt:!0,advancedRunId:n.id,advancedTaskId:u,actorId:s.actorId,sourceEvidence:s.evidence,sourceCaseHash:v(R)}});P=A.id;let U=new Date().toISOString(),O=`review-turn-${v([n.id,u]).slice(0,40)}`,W=await et({store:e.benchmark.store,runId:A.id,workspace:I.workspace,harnessRelease:{id:I.release.harnessRelease.id,contentHash:I.release.harnessRelease.contentHash},admittedAt:U}),L=Be.parse({id:O,sessionId:A.id,providerTurnId:null,modelRef:n.evaluation.model,prompt:y.turn.prompt,startedAt:U,completedAt:U,status:"completed",error:null,metadata:{source:"advanced-refiner-evaluation",automatedTasksetWorkAttempt:!0,sourceCaseHash:v(R)},createImproveRun:null,profileSnapshot:null,harnessSnapshot:{schemaVersion:"openpond.harnessTurnSnapshot.v1",workspaceId:I.workspace.id,workspaceRevision:I.workspace.revision,sourceRevision:I.workspace.sourceRevision,channelName:I.workspace.currentChannel.name,channelRevision:I.workspace.currentChannel.revision,harnessRelease:W.baseHarnessRelease,overlay:{id:W.id,revision:W.revision,contentHash:W.contentHash}}});if(!b){await e.benchmark.store.getTurn(O)||await e.benchmark.store.insertTurn(L);let C=await e.benchmark.store.runtimeEventsForTurn(O);for(let B of y.events){if(C.some(ae=>ae.data?.reviewContextOrigin?.eventId===B.id))continue;let{id:J,timestamp:_,sequence:Z,...V}=B;await e.benchmark.store.appendRuntimeEvent(he({...V,sessionId:A.id,turnId:O,appId:A.appId,data:{...typeof V.data=="object"&&V.data?V.data:{},reviewContextOrigin:{eventId:J,eventHash:v(B),traceHash:y.traceHash,sourceEvidence:s.evidence}}}))}C.some(B=>B.name==="assistant.delta"&&B.data?.sourceCaseHash===v(R))||await e.benchmark.store.appendRuntimeEvent(he({sessionId:A.id,turnId:O,name:"assistant.delta",source:"server",appId:A.appId,status:"completed",output:y.case.output??void 0,data:{sourceEvidence:s.evidence,sourceCaseHash:v(R)}}))}await e.benchmark.store.setHarnessBackgroundReviewSettings({workspaceId:I.workspace.id,enabled:!0,updatedAt:new Date().toISOString()});let E=b?{trigger:b.trigger}:await tt({store:e.benchmark.store,session:A,turn:L,boundaryKind:"turn_completed"});if(!E)throw new Error("The actual context produced no retained trigger decision.");let Q=h(async C=>{let B=b?(({contentHash:J,..._})=>_)(b):{runId:n.id,taskId:u,actorId:s.actorId,teamId:s.teamId,pinHash:s.contentHash,sourceCaseHash:v(R),sessionId:A.id,turnId:O,workspaceId:I.workspace.id,trigger:E.trigger,result:null,receipt:null,cleanupComplete:!1};b=await t.write({...B,...C})},"saveCheckpoint");b||await Q({});let $=b?.result??null,X=!1;try{!$&&E.trigger.decision==="queue_refiner"&&($=await at({store:e.benchmark.store,storeDir:e.benchmark.storeDir,trigger:E.trigger,loadActiveRefinerRelease:h(async()=>f,"loadActiveRefinerRelease"),signal:a.signal,stream:h(async function*(V){yield*e.benchmark.refinerStream({...V,model:n.evaluation.model,pricing:n.evaluation.upstreamModel.pricing,requestId:`review-refiner:${n.id}:${u}`})},"stream")})),await Q({result:$});let C={schemaVersion:"openpond.refinerReviewOutput.v1",decision:E.trigger.decision,trigger:E.trigger,outcome:$?.outcome??null,proposal:$?.proposal??null,validations:$?.validations??[],sourceContext:{evidence:s.evidence,caseHash:v(R)}},B=qa.parse({schemaVersion:"openpond.taskAttempt.v1",id:`review-attempt-${v([n.id,u]).slice(0,40)}`,tasksetId:i.id,taskId:u,split:k.split,attempt:0,modelRef:n.evaluation.model,seed:n.evaluation.seeds[0],output:{text:JSON.stringify(C),json:C},runtimeEventRefs:[],artifactRefs:[],privilegedOutcomeRef:null,infrastructureError:null,latencyMs:Date.now()-Date.parse(U),costUsd:null,startedAt:U,completedAt:new Date().toISOString(),metadata:{execution:"advanced_refiner_review",runId:n.id,sourceCaseHash:v(R),refinerRelease:s.refinerRelease}}),J=await e.benchmark.evaluation.grade({tasksetId:i.id,taskId:u,attempt:B,hostedTokenPricing:n.evaluation.upstreamModel.pricing}),_=await e.benchmark.store.getHarnessRunOverlay(A.id);if(!_)throw new Error("The actual isolated review overlay is unavailable.");_.status!=="abandoned"&&await e.benchmark.store.abandonHarnessRunOverlayAtomically({runId:A.id,expectedRevision:_.revision,updatedAt:new Date().toISOString()}),X=!0;let Z={taskId:u,sourceCaseHash:v(R),reviewSessionId:A.id,reviewTurnId:O,inputHarness:W.baseHarnessRelease,trigger:{id:E.trigger.id,contentHash:E.trigger.contentHash},decision:E.trigger.decision,outcome:$?{id:$.outcome.id,contentHash:$.outcome.contentHash}:null,proposal:$?.proposal?{id:$.proposal.id,contentHash:$.proposal.contentHash}:null,validations:($?.validations??[]).map(V=>({id:V.id,contentHash:V.contentHash})),grade:{id:J.id,contentHash:v(J),score:J.score,passed:J.passed,failureClass:J.failureClass},resetComplete:X};await Q({receipt:Z,cleanupComplete:!0}),m.push(Z)}finally{if(await e.benchmark.store.setHarnessBackgroundReviewSettings({workspaceId:I.workspace.id,enabled:!1,updatedAt:new Date().toISOString()}),!X){let C=await e.benchmark.store.getHarnessRunOverlay(A.id);C&&C.status!=="abandoned"&&await e.benchmark.store.abandonHarnessRunOverlayAtomically({runId:A.id,expectedRevision:C.revision,updatedAt:new Date().toISOString()})}await Q({cleanupComplete:!0}),T=!0}}finally{if(!T&&(await e.benchmark.store.setHarnessBackgroundReviewSettings({workspaceId:I.workspace.id,enabled:!1,updatedAt:new Date().toISOString()}),P)){let S=await e.benchmark.store.getHarnessRunOverlay(P);S&&S.status!=="abandoned"&&await e.benchmark.store.abandonHarnessRunOverlayAtomically({runId:P,expectedRevision:S.revision,updatedAt:new Date().toISOString()})}}}let r=e.boundary.read(n.id,s)??[],p=r.some(u=>u.state!=="completed"&&u.state!=="not_dispatched")?null:r.reduce((u,k)=>u+Number(k.actual_spend??0),0),l=m.map(u=>u.grade.score).filter(u=>u!==null),c=new Date().toISOString(),w=za({schemaVersion:"openpond.advancedRefinerReviewReceipt.v1",id:`review-receipt-${n.id}`,runId:n.id,pinHash:s.contentHash,ownerActorId:s.actorId,teamId:s.teamId,privateClosureHash:s.privateClosureHash,refinerRelease:s.refinerRelease,cases:m,reviewScore:l.length===m.length?l.reduce((u,k)=>u+k,0)/l.length:null,completedCases:l.length,noActionCases:m.filter(u=>u.decision==="no_action").length,routedCases:m.filter(u=>u.decision==="route_deterministically").length,proposalCases:m.filter(u=>u.proposal!==null).length,costUsd:p,cleanupComplete:m.every(u=>u.resetComplete),createdAt:c});return e.benchmark.store.saveModelRun(F.parse({...n,status:w.reviewScore!==null&&w.cleanupComplete&&p!==null?"succeeded":"failed",receipt:w,reward:null,failure:w.reviewScore===null?"Review grade coverage is incomplete.":null,completedAt:c,updatedAt:c}))}}h(Ys,"createAdvancedRefinerReviewQuality");import{createHash as An}from"node:crypto";import{mkdir as Re,readFile as xn,writeFile as xe}from"node:fs/promises";import N from"node:path";var Pn="harness-refiner";function ci(e){let t=e.now??(()=>new Date().toISOString());async function a(o){let i=Mn(o.profileId),d=await e.store.getTaskset(i);if(d?.benchmark?.releaseId===We.id&&d.benchmark.releaseHash===We.contentHash)return d;let g=N.join(e.storeDir,"training","tasksets",i),f=await Cn({release:We,assets:Ia,tasksetRoot:g,tasksetId:i,profileId:o.profileId,revision:d?d.revision+1:1,timestamp:t()});return await e.store.upsertTaskset(f),f}h(a,"ensureHarnessRefiner");async function n(o){let i=ja(o.package),d=`advanced-refiner-${v([o.profileId,i.contentHash,o.adaptationSplit,o.holdoutSplit]).slice(0,40)}`;if(o.adaptationSplit===o.holdoutSplit)throw new Error("Adaptation cannot include its held-out split.");let g=i.taskset.tasks.filter(u=>u.split===o.adaptationSplit),f=i.taskset.tasks.filter(u=>u.split===o.holdoutSplit);if(!g.length||!f.length)throw new Error("The selected release needs both adaptation and private holdout cases.");let m=new Set(g.map(u=>u.clusterKey));if(f.some(u=>m.has(u.clusterKey)))throw new Error("Private holdout families overlap adaptation data.");let r=await e.store.getTaskset(d);if(r){if(r.metadata.importedPackageHash!==i.contentHash)throw new Error("The selected package binding changed.");return r}let p=La({package:i,profileId:o.profileId,name:i.taskset.id,createdAt:t()}),l=N.join(e.storeDir,"training","tasksets",d);await Re(N.join(l,"benchmark"),{recursive:!0});for(let u of i.files){let k=N.resolve(l,u.asset.path);if(!k.startsWith(l+N.sep))throw new Error("Selected evaluator asset escapes its private store.");await Re(N.dirname(k),{recursive:!0}),await xe(k,Oa(u))}for(let u of p.generatedFiles){let k=N.resolve(l,u.path);if(!k.startsWith(l+N.sep))throw new Error("Projected verifier escapes its evaluator store.");await Re(N.dirname(k),{recursive:!0}),await xe(k,u.content)}await xe(N.join(l,"benchmark","taskset.release.json"),JSON.stringify(i.taskset),"utf8");let c=Me.parse({...p.taskset,id:d,environment:{...p.taskset.environment,metadata:{...p.taskset.environment.metadata,runtimeSourceTasksetId:d}},purpose:"benchmark",benchmark:{schemaVersion:"openpond.tasksetBenchmark.v1",definitionId:d,releaseId:i.taskset.id,releaseHash:i.taskset.contentHash,managedReleasePath:"benchmark/taskset.release.json",adaptationSplit:o.adaptationSplit,evaluationSplit:o.holdoutSplit,primaryMetric:"success_rate",qualityGate:"non_regression",source:"imported",metadata:{packageHash:i.contentHash}},metadata:{...p.taskset.metadata,advancedRefiner:!0}}),w=Me.parse({...c,contentHash:Tt(c)});return await e.store.upsertTaskset(w),w}h(n,"projectSelected");async function s(o){if(o.purpose!=="benchmark"||!o.benchmark)return null;let i=N.resolve(e.storeDir,"training","tasksets",o.id,o.benchmark.managedReleasePath),d=va.parse(JSON.parse(await xn(i,"utf8")));if(d.id!==o.benchmark.releaseId||d.contentHash!==o.benchmark.releaseHash)throw new Error("Managed benchmark release does not match its Taskset binding.");return d}return h(s,"releaseForTaskset"),{ensureHarnessRefiner:a,projectSelected:n,releaseForTaskset:s}}h(ci,"createBenchmarkTasksetService");function Tn(e){return Ha({schemaVersion:"openpond.benchmarkDefinition.v1",id:Pn,title:"Harness Refiner 20260818 v2",description:"Measures whether sequential, evidence-driven Harness improvements change verified reward on a frozen cohort; foreground tokens are secondary.",tasksetRelease:{id:e.id,contentHash:e.contentHash},adaptationSplit:"validation",evaluationSplit:"frozen_eval",primaryMetric:"success_rate",qualityGate:"non_regression",caseCounts:{adaptation:e.tasks.filter(t=>t.split==="validation").length,evaluation:e.tasks.filter(t=>t.split==="frozen_eval").length},metadata:{builtin:!0,protocol:"sequential_product_lifecycle_v3",orderSeed:"harness-refiner-20260818-order-v1",secondaryMetrics:["foreground_tokens"]}})}h(Tn,"definitionForRelease");async function Cn(e){let t=Sa.parse(Tn(e.release)),a=_n(e.release.tasks.flatMap(l=>l.artifactRefs)),n=new Map;await Re(N.join(e.tasksetRoot,"assets"),{recursive:!0}),await Re(N.join(e.tasksetRoot,"benchmark"),{recursive:!0}),await Re(N.join(e.tasksetRoot,"rubrics"),{recursive:!0}),await Re(N.join(e.tasksetRoot,"verifiers"),{recursive:!0}),await xe(N.join(e.tasksetRoot,"benchmark","taskset.release.json"),JSON.stringify(e.release),"utf8");for(let l of a){let c=Mr(e.assets,l.path);if(v(c)!==l.contentHash)throw new Error(`Benchmark asset ${l.path} failed its content hash.`);let w=`${l.id}-${N.basename(l.path)}`,u=N.posix.join("assets",w);await xe(N.join(e.tasksetRoot,u),c,"utf8"),n.set(l.id,{artifactRef:u,fileName:w,sha256:An("sha256").update(c).digest("hex"),sizeBytes:Buffer.byteLength(c),mediaType:l.mediaType})}let s=e.release.graders.find(l=>l.kind==="custom_verifier"),o=On(e.release);if(!s)throw new Error("Benchmark release requires a custom verifier.");await Cr(e.assets,e.tasksetRoot,o.rubricRef),await Cr(e.assets,e.tasksetRoot,s.verifierRef);let i=`source-${e.tasksetId}`,d=e.release.tasks.map(l=>({schemaVersion:"openpond.taskData.v1",id:l.id,clusterKey:l.clusterKey,split:l.split,input:l.input,expectedOutput:l.expectedOutput,policyVisibleContext:l.policyVisibleContext,privilegedContextRef:l.privilegedContextRef,sourceRefs:[i],assets:l.artifactRefs.map(c=>{let w=n.get(c.id);if(!w)throw new Error(`Managed benchmark asset ${c.id} is missing.`);return{id:c.id,sourceRefId:i,artifactRef:w.artifactRef,fileName:w.fileName,mediaType:w.mediaType,sha256:w.sha256,sizeBytes:w.sizeBytes,split:l.split,metadata:{portableContentHash:c.contentHash}}}),requiredOutputs:(l.requiredOutputs??[]).map(c=>({path:c.path,mediaType:c.mediaType,schemaRef:c.schemaRef?.id??null,maxBytes:c.maxBytes??void 0,metadata:c.metadata})),tags:l.tags,metadata:{portableTaskId:l.id,portableTaskRecord:l}})),g=[...n.values()].map(l=>l.sha256),f=[{id:s.id,version:s.version,label:"Output contract",kind:"custom_verifier",weight:s.weight,hardGate:s.hardGate,rewardEligible:s.rewardEligible,privileged:s.privileged,module:s.verifierRef.path,exportName:"verify",timeoutMs:s.timeoutMs,networkPolicy:"none",metadata:{portableVerifierHash:s.verifierRef.contentHash,portableVerifierRef:s.verifierRef}}],m=d.find(l=>l.id==="adaptation-launch-delay-email")??d[0],r=Bn(m.id),p=Me.parse({schemaVersion:"openpond.taskset.v1",id:e.tasksetId,revision:e.revision,profileId:e.profileId,profileRelease:null,createImproveRunId:null,name:t.title,objective:t.description,purpose:"benchmark",benchmark:{schemaVersion:"openpond.tasksetBenchmark.v1",definitionId:t.id,releaseId:e.release.id,releaseHash:e.release.contentHash,managedReleasePath:"benchmark/taskset.release.json",adaptationSplit:t.adaptationSplit,evaluationSplit:t.evaluationSplit,primaryMetric:t.primaryMetric,qualityGate:t.qualityGate,source:"builtin",metadata:{}},status:"ready",sourceRefs:[{schemaVersion:"openpond.uploadedFileDatasetSource.v1",kind:"uploaded_file",id:i,profileId:e.profileId,title:"Harness Refiner benchmark fixtures",sourceHash:e.release.contentHash,occurredAt:e.timestamp,licensingStatus:"approved",secretScanStatus:"passed",piiScanStatus:"passed",originalFileNames:[...n.values()].map(l=>l.fileName),mediaTypes:[...new Set([...n.values()].map(l=>l.mediaType))],sourceFileHashes:g,totalBytes:[...n.values()].reduce((l,c)=>l+c.sizeBytes,0),parserVersion:"openpond-builtin-benchmark-v1",metadata:{generatedPublicFixture:!0}}],policy:e.release.policy,environment:{protocolVersion:"openpond.taskEnvironment.v1",kind:"work",entrypoint:"openpond-work-v1",stateful:!0,deterministicSeeds:e.release.environment.deterministicSeeds,toolNames:e.release.tools.map(l=>l.name),lifecycle:["create","reset","step","grade","cleanup"],defaultTimeoutMs:e.release.environment.defaultTimeoutMs,networkPolicy:e.release.environment.networkPolicy,metadata:{portableEnvironmentHash:v(e.release.environment),portableTools:e.release.tools,maxToolTurns:40}},capabilities:{schemaVersion:"openpond.tasksetCapabilities.v1",taskKind:"single_agent",supportedSignals:[],compatibleMethods:["none"],rewardKinds:["deterministic"],requiresTools:!0,requiresState:!0,requiresPrivilegedGrading:!0,environmentPlacements:["local","remote"],exportable:!0,portabilityBlockers:[]},tasks:d,graders:f,graderFixtures:r,learningSignals:{demonstrations:[],preferences:[],corrections:[],feedback:[],rewards:[],labels:[]},authoringProvenance:{schemaVersion:"openpond.taskAuthoringProvenance.v1",model:null,modelConfig:{},skillHash:e.release.contentHash,promptTemplateVersion:"public-benchmark-release-v1",buildIntent:"rubric",buildSpecification:null,evidenceHashes:[e.release.contentHash,...g],tasksetSdkVersion:"portable-builtin-release-v1",sourceCommit:null,repairHistory:[],createdAt:e.timestamp},readiness:null,contentHash:"00000000",createdAt:e.timestamp,updatedAt:e.timestamp,metadata:{portableTasksetRelease:{id:e.release.id,contentHash:e.release.contentHash},supplementaryModelJudge:{id:o.id,version:o.version,rubricHash:o.rubricRef.contentHash,calibrationStatus:o.calibrationStatus,executable:!1,rewardEligible:!1},portableCapabilities:e.release.capabilities}});return Me.parse({...p,contentHash:Tt(p)})}h(Cn,"projectRelease");function Mn(e){return`benchmark-harness-refiner-${v(e).slice(0,16)}`}h(Mn,"tasksetIdForProfile");async function Cr(e,t,a){let n=Mr(e,a.path);if(v(n)!==a.contentHash)throw new Error(`Benchmark verifier asset ${a.path} failed its content hash.`);return await xe(N.join(t,a.path),n,"utf8"),n}h(Cr,"copyVerifierAsset");function Mr(e,t){let a=e[t];if(a===void 0)throw new Error(`Built-in benchmark asset ${t} is unavailable.`);return a}h(Mr,"builtinAsset");function Bn(e){return[{id:"benchmark-positive",taskId:e,label:"positive",output:{text:`Subject: Acme pilot launch update
+`,"utf8");let i=await e.getModelRun(a);return!i||i.status!=="running"||!i.evaluationProgress?i:e.saveModelRun(F.parse({...i,evaluationProgress:{...i.evaluationProgress,evidenceSnapshot:{id:n.id,contentHash:n.contentHash,artifactPath:o}},updatedAt:new Date().toISOString()}))}h(ut,"checkpointEvidenceSnapshot");async function pt(e){let t=e.modelRun.evaluationProgress?.evidenceSnapshot;if(t){let s=dn(JSON.parse(await qt.readFile(_a(e.storeDir,t.artifactPath),"utf8")));if(s.id!==t.id||s.contentHash!==t.contentHash)throw new Error("Durable benchmark evidence snapshot does not match its admitted ref.");return new ge(s.observations)}let a=await cn(e.store,e.attempts),n=new ge(a);return await ut(e.store,e.storeDir,e.modelRun.id,n.manifest()),n}h(pt,"loadOrReconstructEvidenceSnapshot");function dn(e){let t=oe(e);if(t?.schemaVersion!=="openpond.benchmarkEvidenceSnapshot.v1"||typeof t.id!="string"||typeof t.contentHash!="string"||!Array.isArray(t.observations))throw new Error("Durable benchmark evidence snapshot is malformed.");let a=t.observations.map(s=>{let o=oe(s),i=oe(o?.result);if(o?.cohort!=="adaptation"&&o?.cohort!=="held_out"||typeof o.taskId!="string"||o.toolName!=="web_search"&&o.toolName!=="web_fetch"||typeof o.argumentsHash!="string"||!Number.isInteger(o.ordinal)||o.ordinal<0||!i||typeof i.toolCallId!="string"||typeof i.name!="string"||typeof i.ok!="boolean"||typeof i.contentText!="string")throw new Error("Durable benchmark evidence snapshot contains a malformed observation.");return o}),n={schemaVersion:"openpond.benchmarkEvidenceSnapshot.v1",id:t.id,observations:a};if(v(n)!==t.contentHash)throw new Error("Durable benchmark evidence snapshot failed content-hash validation.");return{...n,contentHash:t.contentHash}}h(dn,"parseEvidenceSnapshotManifest");async function cn(e,t){let a=[];for(let{result:n,cohort:s}of t){let o=we(n.attempt.metadata,"sessionId"),i=we(n.attempt.metadata,"turnId");if(!o||!i)continue;let d=(await e.runtimeEventsForSession(o,{names:["tool.started","tool.completed"],limit:1e5})).filter(m=>m.turnId===i),g=new Map,f=new Map;for(let m of d){let r=oe(m.data),p=typeof r?.toolCallId=="string"?r.toolCallId:null,l=m.action==="web_search"||m.action==="web_fetch"?m.action:null;if(!p||!l)continue;if(m.name==="tool.started"){let b=f.get(l)??0;f.set(l,b+1),g.set(p,{toolName:l,args:oe(m.args)??{},ordinal:b});continue}let c=g.get(p),w=oe(r?.result);if(!c||!w)continue;let u=typeof w.ok=="boolean"?w.ok:m.status==="completed",k=typeof w.output=="string"?w.output:typeof m.output=="string"?m.output:"",R=Object.fromEntries(Object.entries(w).filter(([b])=>b!=="ok"&&b!=="output"));a.push({cohort:s,taskId:n.attempt.taskId,toolName:c.toolName,argumentsHash:v(c.args),ordinal:c.ordinal,result:{toolCallId:p,name:c.toolName,ok:u,contentText:k,...Object.keys(R).length?{data:R}:{}}})}}return a}h(cn,"reconstructFrozenToolObservations");function ln(e,t){return t.map(a=>{let n=Y(),s=a.attempt.metadata.usage;for(let o of Array.isArray(s)?s:[s])Ne(n,o);return{phase:e,taskId:a.attempt.taskId,attemptId:a.attempt.id,sessionId:we(a.attempt.metadata,"sessionId"),turnId:we(a.attempt.metadata,"turnId"),passed:a.grade.passed,score:a.grade.score,failureClass:a.grade.failureClass,inputTokens:n.inputTokens,outputTokens:n.outputTokens,totalTokens:n.totalTokens,latencyMs:a.attempt.latencyMs,costUsd:a.attempt.costUsd,startedAt:a.attempt.startedAt}})}h(ln,"modelEvaluationAttempts");function wr(e){let t=Y();for(let a of Array.isArray(e)?e:[e])Ne(t,a);return{inputTokens:t.inputTokens,outputTokens:t.outputTokens,totalTokens:t.totalTokens}}h(wr,"attemptUsageSummary");function Y(){return{inputTokens:0,outputTokens:0,totalTokens:0,costUsd:null}}h(Y,"emptyUsageCategory");function Ne(e,t,a){let n=er(t);e.inputTokens+=n.promptTokens??0,e.outputTokens+=n.completionTokens??0,e.totalTokens+=n.totalTokens??0,typeof a=="number"&&Number.isFinite(a)&&a>=0&&(e.costUsd=(e.costUsd??0)+a)}h(Ne,"addUsage");function Dt(e,t){e.inputTokens+=t.inputTokens,e.outputTokens+=t.outputTokens,e.totalTokens+=t.totalTokens,t.costUsd!==null&&(e.costUsd=(e.costUsd??0)+t.costUsd)}h(Dt,"mergeUsage");function Ee(e,t){let a=e.find(n=>n.stage===t);if(!a)throw new Error(`Harness Refiner plan is missing ${t}.`);return a}h(Ee,"requirePlanStage");function ce(e,t,a){return{execute:h(async n=>{try{return{...await e.execute({...n,mode:t,cohort:a}),toolCallId:n.callId}}catch(s){if(t!=="replay"||!(s instanceof je))throw s;return{toolCallId:n.callId,name:n.toolName,ok:!1,contentText:JSON.stringify({ok:!1,action:n.toolName,output:"The benchmark's frozen external-evidence budget is exhausted for this task.",repairHint:"Do not call web tools again. Finish the task using the frozen evidence already returned."})}}},"execute")}}h(ce,"frozenToolEvidence");function ft(e){let t=[];for(let[a,n]of[["held-out baseline",e.baseline],["adaptation baseline",e.adaptation],["sequential adaptation treatment",e.candidateAdaptation],["held-out candidate",e.candidate]])n.terminalCount!==n.attemptCount&&t.push(`${a} did not terminalize all admitted attempts.`);return e.infrastructureValid||t.push("At least one attempt has infrastructure-invalid outcome evidence."),e.harnessChanged||t.push("The candidate Harness is unchanged."),e.lineageValid||t.push("Candidate Harness lineage is incomplete."),e.candidateAdaptation.passedCount!==e.candidateAdaptation.attemptCount&&t.push("Sequential adaptation treatment did not pass every case."),e.candidate.passedCount!==e.candidate.attemptCount&&t.push("Candidate held-out quality did not pass every case."),t}h(ft,"comparisonInvalidReasons");function ht(e){return!e.infrastructureValid||e.baseline.terminalCount!==e.baseline.attemptCount||e.adaptation.terminalCount!==e.adaptation.attemptCount||e.candidate.terminalCount!==e.candidate.attemptCount||e.candidateAdaptation.terminalCount!==e.candidateAdaptation.attemptCount?"infrastructure_failure":!e.harnessChanged||!e.lineageValid?"inconclusive":!e.comparison.qualityPassed||e.candidate.passedCount!==e.candidate.attemptCount||e.candidateAdaptation.passedCount!==e.candidateAdaptation.attemptCount?"regressed":e.comparison.improved?"improved":"no_improvement"}h(ht,"classifyComparison");function vr(e){return`model_version_${v({modelId:e,version:0}).slice(0,24)}`}h(vr,"modelVersionId");function $t(e){return(e instanceof Error?e.message:String(e)).trim().slice(0,5e3)||"Harness Refiner benchmark failed."}h($t,"safeError");async function zt(e,t,a){if(!Number.isFinite(t.maximumSpendUsd)||t.maximumSpendUsd<=0)throw new Error("Harness Refiner benchmark maximum spend must be greater than zero.");let n=e.resolveEvaluationProject?await e.resolveEvaluationProject(t.modelId,t.profileId):await lt(e.store,t.modelId,t.profileId);if(n.id!==t.modelId||n.profileId!==t.profileId)throw new Error("The actual evaluation configuration owner changed.");let s=t.advancedEvaluation?$a(t.advancedEvaluation):null;if(s){if(!e.advancedBoundary||!t.selectedTaskset)throw new Error("The advanced evaluation owner and exact selected Dataset adapter are unavailable.");if(await e.advancedBoundary.authorize(s),s.profileRef.profileId!==t.profileId||s.maximumCostUsd!==t.maximumSpendUsd)throw new Error("Advanced evaluation changed its admitted source or budget.")}let o=s?await e.store.getTaskset(t.selectedTaskset.id):await e.benchmarkTasksets.ensureHarnessRefiner({profileId:t.profileId});if(!o)throw new Error("The exact selected Taskset is unavailable.");if(s&&(o.revision!==t.selectedTaskset.revision||o.contentHash!==t.selectedTaskset.contentHash||o.metadata.importedPackageHash!==s.externalDatasetBinding.packageHash))throw new Error("The advanced selected Dataset package changed.");if(!o.benchmark)throw new Error("Harness Refiner Taskset is unavailable.");if(o.graders.some(c=>c.kind==="model_judge"&&c.rewardEligible&&c.calibrationStatus!=="passed")){if(s)throw new Error("The selected model judge needs its own completed calibration before advanced evaluation.");let c=await e.evaluation.calibrateModelJudges(o.id);if(!c.passed)throw new Error("Harness Refiner model judge did not pass its declared calibration fixtures.");o=c.taskset}let i=ot({taskset:o,seeds:t.seeds,repetitions:t.repetitions,...s?{selectedPopulation:s}:{}}),d=s?await e.store.getHarnessReleaseRecord(s.baselineRelease.contentHash):await Ya(e.store);if(s&&d?.harnessRelease.id!==s.baselineRelease.id)throw new Error("The admitted baseline source release is unavailable.");if(!d)throw new Error("A selected local Harness is required.");let g=await e.resolveUpstreamModel(t.model),f=a(),m=s?`advanced-refiner-${v([s.actorId,s.teamId,s.operationId,s.contentHash,t.model,t.seeds,t.repetitions]).slice(0,40)}`:`model_run_${v({modelId:n.id,taskset:o.contentHash,model:t.model,reasoningEffort:t.reasoningEffort,seeds:t.seeds,repetitions:t.repetitions,startedAt:f}).slice(0,24)}`;if(s){let c=await e.store.getModelRun(m);if(c){if(c.evaluation?.benchmarkId!=="harness-refiner"||c.evaluation.advancedEvaluation?.contentHash!==s.contentHash)throw new Error("The advanced operation retains another immutable intent.");return{prepared:c,project:n,existing:!0}}}let r=await Ke({store:e.store,storeDir:e.storeDir,id:`benchmark-${m}`,ownerId:`benchmark:${m}`,name:`Harness Refiner ${m}`,sourceRelease:{id:d.harnessRelease.id,contentHash:d.harnessRelease.contentHash},now:a});await e.store.setHarnessBackgroundReviewSettings({workspaceId:r.workspace.id,enabled:!1,updatedAt:a()});let p=vr(n.id);return{prepared:await e.store.saveModelRun(F.parse({schemaVersion:"openpond.modelRun.v1",id:m,modelId:n.id,modelVersionId:p,profileId:t.profileId,kind:"evaluation",status:"running",method:null,destinationId:null,taskset:{id:o.id,revision:o.revision,contentHash:o.contentHash},harnessRelease:{id:r.release.harnessRelease.id,contentHash:r.release.harnessRelease.contentHash},quote:null,evaluation:{benchmarkId:"harness-refiner",...s?{advancedEvaluation:s}:{},model:t.model,upstreamModel:g,reasoningEffort:t.reasoningEffort,seeds:t.seeds,repetitions:t.repetitions,maximumSpendUsd:t.maximumSpendUsd,attemptPlan:i},evaluationProgress:{stage:"adaptation",completedAttempts:0,totalAttempts:qe(i),accounting:de()},reward:null,receipt:null,adapterArtifactLineageId:null,failure:null,startedAt:f,completedAt:null,updatedAt:f})),project:n,existing:!1}}h(zt,"prepareHarnessRefinerBenchmarkAdmission");import{promises as Ae}from"node:fs";import le from"node:path";function gt(e){let t={schemaVersion:"openpond.harnessRefinerBenchmarkResult.v2",id:`benchmark-result-${e.modelRunId}`,modelRunId:e.modelRunId,benchmarkId:"harness-refiner",model:e.model,upstreamModel:e.upstreamModel,reasoningEffort:e.reasoningEffort,tasksetRelease:e.baseline.tasksetRelease,baseline:e.baseline,adaptation:e.adaptation,refiner:e.refiner,candidateAdaptation:e.candidateAdaptation,candidate:e.candidate,comparison:e.comparison,executionPlan:e.executionPlan,evidenceSnapshot:e.evidenceSnapshot,lineage:e.lineage,publicationPolicy:{primaryReward:"deterministic_output_contract",supplementaryJudge:"not_executed_uncalibrated",diagnosticPasses:1,confirmationPasses:1,uncertainty:"paired_per_case_descriptive"},harness:{baseline:e.baseline.harnessRelease,candidate:e.candidate.harnessRelease},createdAt:e.createdAt};return{...t,contentHash:v(t)}}h(gt,"createResultManifest");async function wt(e,t,a){let n=le.join(e,"training","model-runs",t,"benchmark");await Ae.mkdir(n,{recursive:!0});let s=le.join(n,`${a.contentHash}.json`);return await Ae.writeFile(s,`${JSON.stringify(a,null,2)}
+`,"utf8"),le.relative(e,s).replaceAll(le.sep,"/")}h(wt,"writeManagedResult");async function vt(e,t){let a=le.join(e,"training","model-runs",t,"benchmark"),n;try{n=(await Ae.readdir(a)).filter(o=>o.endsWith(".json"))}catch(o){if(o.code==="ENOENT")return null;throw o}return(await Promise.all(n.map(async o=>{let i=JSON.parse(await Ae.readFile(le.join(a,o),"utf8"));return i.schemaVersion==="openpond.harnessRefinerBenchmarkResult.v2"&&i.modelRunId===t?i:null}))).filter(o=>o!==null).sort((o,i)=>i.createdAt.localeCompare(o.createdAt))[0]??null}h(vt,"loadLatestManagedResult");async function kt(e){if(e.profile.mode!=="local"||!e.profile.repoPath||!e.profile.git?.head)return null;let t=Ge(e.storeDir,e.workspaceId).source,a=await kr(t),n=`benchmarks/harness-refiner/runs/${e.modelRunId}`;return Va({repoPath:e.profile.repoPath,runId:e.modelRunId,baseCommit:e.profile.git.head,message:`Preserve Harness Refiner benchmark ${e.modelRunId}`,files:[{path:`${n}/result.json`,contents:`${JSON.stringify(e.manifest,null,2)}
+`},...await Promise.all(a.map(async s=>({path:`${n}/candidate-harness/${s}`,contents:await Ae.readFile(le.join(t,...s.split("/")))})))]})}h(kt,"preserveProfileResult");async function kr(e,t=""){let a=le.join(e,...t.split("/").filter(Boolean)),n=await Ae.readdir(a,{withFileTypes:!0}),s=[];for(let o of n.sort((i,d)=>i.name.localeCompare(d.name))){let i=t?`${t}/${o.name}`:o.name;o.isDirectory()?s.push(...await kr(e,i)):o.isFile()&&s.push(i)}return s}h(kr,"listFiles");async function Ut(e){if(!e.modelRun.modelVersionId)throw new Error("A base-Version materialization Run requires a Model Version identity.");let t=await e.store.getModelVersion(e.modelRun.modelVersionId);if(t)return t;if(!("portable"in e.baseline))throw new Error("A portable baseline attempt is required to create the base Model Version.");let a=e.baseline.portable.runManifest,n=e.baseline.portable.tasksetRelease,s={resolvedBundleHash:v({tasksetRelease:n.contentHash,runManifest:a.contentHash}),profileRelease:{id:`profile-release-${e.modelRun.profileId}`,revision:1,contentHash:v({profileId:e.modelRun.profileId})},harnessRelease:a.harnessRelease,agentRelease:{id:e.baseline.portable.agentSnapshot.id,contentHash:e.baseline.portable.agentSnapshot.contentHash},grader:{id:`grader-${n.id}`,contentHash:v(n.graders)}},o={schemaVersion:"openpond.modelVersion.v1",id:e.modelRun.modelVersionId,modelId:e.project.id,profileId:e.project.profileId,version:0,kind:"base_reference",status:"available",baseModel:{schemaVersion:"openpond.baseModelPreference.v1",modelId:`${e.model.providerId}/${e.model.modelId}`,revision:a.model.revision,tokenizerRevision:a.model.tokenizerRevision,chatTemplateHash:a.model.chatTemplateHash,modelAssetId:null,source:"managed"},taskset:e.modelRun.taskset,releaseGraph:s,artifactLineageId:null,adapterStatus:"not_trained",createdAt:e.modelRun.startedAt};return e.store.saveModelVersion(Za.parse({...o,contentHash:v(o)}))}h(Ut,"ensureBaseVersion");async function Rr(e){let{deps:t,benchmarkInput:a,modelRun:n,signal:s,workspace:o,taskset:i,executionPlan:d,admittedPricing:g,budget:f,totalAttempts:m}=e,[r,p,l]=await Promise.all([ve({store:t.store,modelRunId:n.id,tasksetId:i.id,plan:e.baselinePlan}),ve({store:t.store,modelRunId:n.id,tasksetId:i.id,plan:e.adaptationPlan}),ve({store:t.store,modelRunId:n.id,tasksetId:i.id,plan:e.candidatePlan})]),c=await vt(t.storeDir,n.id);if(!c)throw new Error("Comparison recovery requires the durable benchmark result manifest.");let w=n.evaluationProgress?.accounting?.attempts.filter(G=>G.phase==="candidate_adaptation").map(G=>G.attemptId)??[],u=await ct({store:t.store,tasksetId:i.id,attemptIds:w});if(c.tasksetRelease.id!==r.run.tasksetRelease.id||c.tasksetRelease.contentHash!==r.run.tasksetRelease.contentHash||c.harness.baseline.contentHash!==r.run.harnessRelease.contentHash)throw new Error("Comparison recovery manifest drifted from the admitted benchmark.");let k=o.currentChannel.release;if(!k)throw new Error("Comparison recovery has no candidate Harness release.");let R=await t.store.getHarnessReleaseRecord(k.contentHash);if(!R||R.workspaceId!==o.id)throw new Error("Comparison recovery candidate Harness is unavailable.");let b=await He({workspace:o,release:R}),y=ke(b.release,b.instructionContext),H=await pt({store:t.store,storeDir:t.storeDir,modelRun:n,attempts:[...r.attempts.map(G=>({result:G,cohort:"held_out"})),...p.attempts.map(G=>({result:G,cohort:"adaptation"}))]}),I=await h(async(G,Ue,ye,j)=>{let ue=G.attempts.filter(ee=>ee.grade.score===null||ee.grade.failureClass==="infrastructure_failure").map(ee=>ee.attempt.taskId);if(!ue.length)return G;let At=await t.evaluation.executeBenchmark({tasksetId:i.id,phase:"candidate",model:a.model,reasoningEffort:a.reasoningEffort,seeds:a.seeds,repetitions:a.repetitions,split:Ue.split,taskIds:ue,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:y,hostedTokenPricing:g,parentModelRunId:n.id,signal:s,toolEvidence:ce(H,"replay",j),onAttemptComplete:h(async ee=>{s.throwIfAborted(),f.assertAvailable(`${ye} infrastructure retry`),f.charge(ee.attempt.costUsd,`${ye} infrastructure retry foreground work`);let pe=t.evaluation.consumeGraderUsage([ee.attempt.id]);f.charge(pe.costUsd,`${ye} infrastructure retry grading`),await mt(t.store,n.id,{stage:ye,completedAttempts:m,totalAttempts:m,result:ee,grader:pe,observedSpendUsd:f.observedSpendUsd})},"onAttemptComplete")});return ur({store:t.store,tasksetId:i.id,parentModelRunId:n.id,prior:G,retries:At.attempts,createdAt:t.now()})},"retryStage")(l,e.candidatePlan,"candidate","held_out"),P=Aa({id:`benchmark-comparison-${v([r.run.contentHash,I.run.contentHash]).slice(0,24)}`,baseline:r.run,candidate:I.run,primaryMetric:i.benchmark.primaryMetric,qualityGate:i.benchmark.qualityGate,createdAt:t.now(),metadata:{sourceTasksetId:i.id,benchmarkDefinitionId:i.benchmark.definitionId,recoveredInfrastructureAttempts:!0}});await t.store.saveBenchmarkComparison({tasksetId:i.id,comparison:P});let T={...c.lineage,candidateRelease:{id:c.lineage.candidateRelease.id,contentHash:c.lineage.candidateRelease.contentHash}},S=H.manifest(),D=gt({modelRunId:n.id,model:a.model,upstreamModel:c.upstreamModel,reasoningEffort:a.reasoningEffort,baseline:r.run,adaptation:p.run,refiner:c.refiner,candidateAdaptation:c.candidateAdaptation,candidate:I.run,comparison:P,executionPlan:d,evidenceSnapshot:S,lineage:T,createdAt:t.now()}),A=await wt(t.storeDir,n.id,D),U=await t.loadProfileState(),O=await kt({profile:U,modelRunId:n.id,workspaceId:o.id,storeDir:t.storeDir,manifest:D}),W=dt([...r.attempts,...p.attempts,...u,...I.attempts]),L=r.run.harnessRelease.contentHash!==I.run.harnessRelease.contentHash,E=ht({comparison:P,baseline:r.run,adaptation:p.run,candidate:I.run,candidateAdaptation:c.candidateAdaptation,harnessChanged:L,lineageValid:T.valid,infrastructureValid:W}),Q=ft({baseline:r.run,adaptation:p.run,candidateAdaptation:c.candidateAdaptation,candidate:I.run,harnessChanged:L,lineageValid:T.valid,infrastructureValid:W}),X=(await t.store.getModelRun(n.id)??n).evaluationProgress?.accounting??de(),C=Ye({attempts:X.attempts,targetTaskCount:st(d)}).summary,B=it({baselineTokens:r.run.usage.totalTokens,candidateTokens:I.run.usage.totalTokens,refinerTokens:X.usage.refiner.totalTokens,graderTokens:X.usage.grader.totalTokens}),J={schemaVersion:"openpond.modelEvaluationReceipt.v1",benchmarkId:"harness-refiner",resultManifest:{id:D.id,contentHash:D.contentHash,artifactPath:A},stages:{baseline:{id:r.run.id,contentHash:r.run.contentHash},adaptation:{id:p.run.id,contentHash:p.run.contentHash},candidateAdaptation:{id:c.candidateAdaptation.id,contentHash:c.candidateAdaptation.contentHash},refiner:{id:c.refiner.id,contentHash:c.refiner.contentHash},candidate:{id:I.run.id,contentHash:I.run.contentHash},comparison:{id:P.id,contentHash:P.contentHash}},usage:X.usage,quality:{baselinePassRate:P.baselinePassRate,candidatePassRate:P.candidatePassRate,adaptationBaselinePassRate:p.run.passedCount/p.run.attemptCount,adaptationCandidatePassRate:c.candidateAdaptation.passedCount/c.candidateAdaptation.attemptCount,adaptationCandidatePassed:c.candidateAdaptation.passedCount===c.candidateAdaptation.attemptCount,heldOutCandidatePassed:I.run.passedCount===I.run.attemptCount,passed:P.qualityPassed&&c.candidateAdaptation.passedCount===c.candidateAdaptation.attemptCount&&I.run.passedCount===I.run.attemptCount&&W&&E!=="infrastructure_failure"},foregroundTokenDelta:P.foregroundTokenDelta,foregroundTokenDeltaPercent:P.foregroundTokenDeltaPercent,taskEfficiency:C,efficiency:B,budget:{maximumSpendUsd:f.maximumSpendUsd,observedSpendUsd:f.observedSpendUsd,enforced:!0},evidenceSnapshot:{id:S.id,contentHash:S.contentHash},lineage:T,invalidReasons:Q,attempts:X.attempts,terminalClassification:E,profileGit:O?{ref:O.ref,commit:O.commit,baseCommit:O.baseCommit}:null},_=Qe.parse({...J,contentHash:v(J)}),Z=t.now(),V=await t.store.getModelRun(n.id)??n,ae=E==="infrastructure_failure";return t.store.saveModelRun(F.parse({...V,status:ae?"failed":"succeeded",harnessRelease:r.run.harnessRelease,receipt:_,failure:ae?"Benchmark attempts did not reach the model because the Work runtime failed.":null,evaluationProgress:{...V.evaluationProgress,stage:"comparison",completedAttempts:m,totalAttempts:m},completedAt:Z,updatedAt:Z}))}h(Rr,"resumeHarnessRefinerComparison");async function br(e){let[t,a,n,s]=await Promise.all([e.store.listHarnessImprovementArtifacts(e.workspaceId,"refiner_outcome",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"proposal",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"targeted_validation",1e4),e.store.listHarnessImprovementArtifacts(e.workspaceId,"apply_receipt",1e3)]),o=a,i=n,d=v(e.adaptationAttempts.map(p=>({attempt:p.attempt.id,receipt:p.receiptContentHash,grade:v(p.grade)}))),g=o.every(p=>p.validationPlan.filter(l=>l.required).every(l=>i.some(c=>c.validationId===l.id&&c.status==="passed"))),f=new Set(o.map(p=>p.contentHash)),m=new Set(s.flatMap(p=>{let l="proposal"in p?p.proposal:null;return l&&typeof l=="object"&&"contentHash"in l?[String(l.contentHash)]:[]})),r=[...f].every(p=>m.has(p));return{adaptationEvidenceHash:d,refinerInputHash:e.refinerInputHash,refinerOutcomeHash:v(t),validationHash:v(i),applyReceiptHash:v(s),candidateRelease:{id:e.candidateRelease.id,contentHash:e.candidateRelease.contentHash},valid:t.length===e.completedSteps.filter(p=>p.outcome).length&&g&&r}}h(br,"benchmarkLineage");var mn=100,De=class extends Error{constructor(a,n,s){super(a,s);this.details=n}details;static{h(this,"BenchmarkRefinerInvocationError")}name="BenchmarkRefinerInvocationError"};function un(e){let t=e.artifactManifest.entries.slice(0,mn);return{schemaVersion:"openpond.refinerRewardPacket.v1",attemptRef:e.rewardReceipt.attemptRef,artifactManifest:{ref:{id:e.artifactManifest.id,contentHash:e.artifactManifest.contentHash},entries:t,entryCount:e.artifactManifest.entries.length,truncated:t.length<e.artifactManifest.entries.length},rewardReceipt:e.rewardReceipt,attempt:{status:e.rewardReceipt.outcomeClass,infrastructureError:e.attempt.infrastructureError,outputPresent:typeof e.attempt.output.text=="string"&&e.attempt.output.text.trim().length>0,artifactCount:e.artifactCount,runtimeEventCount:e.attempt.runtimeEventRefs.length,modelRequestCount:Array.isArray(e.attempt.metadata.usage)?e.attempt.metadata.usage.length:e.attempt.metadata.usage?1:0,latencyMs:e.attempt.latencyMs,usage:wr(e.attempt.metadata.usage)}}}h(un,"benchmarkRefinerRewardPacket");async function pn(e){let t=e.result.attempt,a=we(t.metadata,"sessionId"),n=we(t.metadata,"turnId");if(!a||!n)return null;let s=await e.store.getSession(a);if(!s)return null;let o=await et({store:e.store,runId:s.id,workspace:e.runtime.workspace,harnessRelease:{id:e.runtime.release.harnessRelease.id,contentHash:e.runtime.release.harnessRelease.contentHash},admittedAt:t.startedAt}),i=e.taskset.tasks.find(l=>l.id===t.taskId);if(!i)return null;let d=Oe.parse({id:n,sessionId:s.id,providerTurnId:null,modelRef:e.model,prompt:hr(i),startedAt:t.startedAt,completedAt:t.completedAt,status:"completed",error:null,metadata:{automatedTasksetWorkAttempt:!0,benchmarkId:"harness-refiner",modelRunId:e.modelRun.id,attemptId:t.id},harnessSnapshot:{schemaVersion:"openpond.harnessTurnSnapshot.v1",workspaceId:e.runtime.workspace.id,workspaceRevision:e.runtime.workspace.revision,sourceRevision:e.runtime.workspace.sourceRevision,channelName:e.runtime.workspace.currentChannel.name,channelRevision:e.runtime.workspace.currentChannel.revision,harnessRelease:o.baseHarnessRelease,overlay:{id:o.id,revision:o.revision,contentHash:o.contentHash}}});await e.store.getTurn(d.id)?await e.store.updateTurn(d.id,l=>Oe.parse({...l,modelRef:d.modelRef,metadata:{...l.metadata,...d.metadata},harnessSnapshot:d.harnessSnapshot})):await e.store.insertTurn(d);let f=t.output.text;typeof f=="string"&&f.trim()&&await e.store.appendRuntimeEvent(he({sessionId:s.id,turnId:d.id,name:"assistant.delta",source:"server",appId:s.appId,status:"completed",output:f}));let m=un({attempt:t,artifactManifest:e.result.artifactManifest,rewardReceipt:e.result.rewardReceipt,artifactCount:e.result.artifacts.length}),r=JSON.stringify(m),p=e.result.rewardReceipt.status==="scored"&&e.result.rewardReceipt.passed;return await e.store.appendRuntimeEvent(he({sessionId:s.id,turnId:d.id,name:"diagnostic",source:"server",appId:s.appId,action:"taskset_grade",status:p?"completed":"failed",output:r,error:p?void 0:r,data:{result:{output:r,passed:p,status:e.result.rewardReceipt.status,reward:e.result.rewardReceipt.reward,rewardReceiptRef:{id:e.result.rewardReceipt.id,contentHash:e.result.rewardReceipt.contentHash},artifactManifestRef:{id:e.result.artifactManifest.id,contentHash:e.result.artifactManifest.contentHash}}}})),{session:s,turn:d,result:e.result}}h(pn,"materializeBenchmarkRefinerBoundary");async function yr(e){let t=await pn(e);if(!t)throw new Error(`Adaptation attempt ${e.result.attempt.id} has no Refiner boundary.`);let a=await fn({store:e.store,workspaceId:e.runtime.workspace.id,turnId:t.turn.id}),n=a?{observations:[],trigger:a}:await tt({store:e.store,session:t.session,turn:t.turn,boundaryKind:"turn_completed",now:e.now});if(!n)throw new Error(`Adaptation attempt ${e.result.attempt.id} produced no Refiner detection.`);if(n.trigger.decision!=="queue_refiner")return{detection:n,result:null};e.budget.assertAvailable(`Refiner turn ${n.trigger.id}`);let s=Y(),o=!1,i=0,d=0,g=0,f=e.now(),m=null,r=null;try{m=await at({store:e.store,storeDir:e.storeDir,trigger:n.trigger,...Mt()?{loadActiveRefinerRelease:h(async()=>{let w=Mt().pin,k=(await Ze(e.storeDir)).releases.find(R=>R.id===w.refinerRelease.id&&R.contentHash===w.refinerRelease.contentHash);if(!k)throw new Error("The admitted Refiner release is unavailable.");return k},"loadActiveRefinerRelease")}:{},stream:h(async function*(w){o=!0,i+=1;let u=!1;g+=hn(w.messages,e.admittedPricing);try{for await(let k of e.refinerStream({...w,model:e.model,pricing:e.admittedPricing,requestId:`advanced-refiner:${e.modelRun.id}:${n.trigger.id}:${i}`}))typeof k.costUsd=="number"&&(u=!0),(k.usage!==void 0||k.costUsd!==void 0)&&Ne(s,k.usage,k.costUsd),k.text&&(yield{text:k.text})}finally{u&&(d+=1)}},"stream"),signal:e.signal,timeoutMs:lr,now:e.now})}catch(w){r=w}let p=e.now(),l=o?s.costUsd!==null&&d===i?"authoritative":"estimated":"none",c=l==="authoritative"?s.costUsd:l==="estimated"?Math.max(s.costUsd??0,g):null;if(o)try{e.budget.charge(c,"Refiner"),await gr(e.store,e.modelRun.id,s,e.budget.observedSpendUsd)}catch(w){r??=w}if(r){let w=r instanceof Error?r.message:String(r),u=gn(r,e.signal),k=u!=="cancelled";throw await e.store.appendRuntimeEvent(he({sessionId:t.session.id,turnId:t.turn.id,name:"harness.refiner.failed",source:"server",appId:t.session.appId,status:"failed",output:w,data:{trigger:{id:n.trigger.id,contentHash:n.trigger.contentHash},benchmark:!0,modelRunId:e.modelRun.id,attemptId:e.result.attempt.id,failureKind:u,retryable:k,costBasis:l,estimatedCostUsd:l==="estimated"?g:null}})).catch(()=>{}),new De(w,{trigger:n.trigger,usage:s,costBasis:l,estimatedCostUsd:l==="estimated"?g:null,failureKind:u,retryable:k,startedAt:f,completedAt:p},{cause:r})}return{detection:n,result:m,invocation:{usage:s,costBasis:l,estimatedCostUsd:l==="estimated"?g:null,startedAt:f,completedAt:p}}}h(yr,"runBenchmarkRefinerAfterAttempt");async function fn(e){return(await e.store.listHarnessImprovementArtifacts(e.workspaceId,"trigger_decision",1e3)).find(a=>a.turnId===e.turnId&&a.decision==="queue_refiner")??null}h(fn,"queuedTriggerForTurn");function hn(e,t){let a=e.reduce((s,o)=>s+o.content.length,0);return(Math.ceil(a/3)*t.inputUsdPerMillionTokens+pa*t.outputUsdPerMillionTokens)/1e6}h(hn,"conservativeRefinerRequestCost");function gn(e,t){if(t.aborted)return"cancelled";let a=e instanceof Error?e.message:String(e);return/timed out|timeout/i.test(a)?"timeout":/invalid structured output|response limit|JSON/i.test(a)?"invalid_output":/provider|rate limit|billing|signed out|network|fetch/i.test(a)?"provider_failure":"unknown"}h(gn,"benchmarkRefinerFailureKind");import{promises as Rt}from"node:fs";import Ir from"node:path";function Lt(e){let t={schemaVersion:"openpond.benchmarkRefinerInvocation.v1",id:`benchmark-refiner-invocation-${v({taskId:e.taskId,attemptId:e.attemptId,invocationOrdinal:e.invocationOrdinal,trigger:e.trigger,status:e.status,startedAt:e.startedAt}).slice(0,24)}`,...e};return{...t,contentHash:v(t)}}h(Lt,"createBenchmarkRefinerInvocationReceipt");async function bt(e){let t=Er(e.storeDir,e.modelRunId),a;try{a=JSON.parse(await Rt.readFile(t,"utf8"))}catch(n){if(n.code==="ENOENT")return null;throw n}return wn(a,e.modelRunId)}h(bt,"loadSequentialAdaptationCheckpoint");async function Vt(e){let t=await bt(e);if(t){if(!vn(t.initialHarness,e.initialHarness))throw new Error("Sequential adaptation checkpoint changed its initial Harness release.");return t}return Hr({storeDir:e.storeDir,checkpoint:Sr({modelRunId:e.modelRunId,initialHarness:e.initialHarness,steps:[],invocations:[],updatedAt:e.now})})}h(Vt,"initializeSequentialAdaptationCheckpoint");async function Wt(e){let t=await Vt({storeDir:e.storeDir,modelRunId:e.modelRunId,initialHarness:e.initialHarness,now:e.now});if(!e.invocation&&!e.step)throw new Error("A sequential checkpoint append requires an invocation or step.");let a=e.invocation?[...t.invocations.filter(s=>s.id!==e.invocation.id),e.invocation]:t.invocations,n=e.step?[...t.steps.filter(s=>s.taskId!==e.step.taskId),e.step].sort((s,o)=>s.ordinal-o.ordinal):t.steps;return Hr({storeDir:e.storeDir,checkpoint:Sr({modelRunId:t.modelRunId,initialHarness:t.initialHarness,steps:n,invocations:a,updatedAt:e.now})})}h(Wt,"appendSequentialAdaptationInvocation");function Sr(e){let t={schemaVersion:"openpond.sequentialAdaptationCheckpoint.v1",...e};return{...t,contentHash:v(t)}}h(Sr,"checkpoint");async function Hr(e){let t=Er(e.storeDir,e.checkpoint.modelRunId),a=Ir.dirname(t);await Rt.mkdir(a,{recursive:!0});let n=`${t}.${process.pid}.tmp`;return await Rt.writeFile(n,`${JSON.stringify(e.checkpoint,null,2)}
+`,"utf8"),await Rt.rename(n,t),e.checkpoint}h(Hr,"writeSequentialAdaptationCheckpoint");function wn(e,t){if(!e||typeof e!="object"||Array.isArray(e))throw new Error("Sequential adaptation checkpoint is malformed.");let a=e;if(a.schemaVersion!=="openpond.sequentialAdaptationCheckpoint.v1"||a.modelRunId!==t||!a.initialHarness||!Array.isArray(a.steps)||!Array.isArray(a.invocations)||typeof a.updatedAt!="string"||typeof a.contentHash!="string")throw new Error("Sequential adaptation checkpoint is malformed.");let{contentHash:n,...s}=a;if(v(s)!==n)throw new Error("Sequential adaptation checkpoint failed content-hash validation.");let o=a;for(let i of o.invocations){let{contentHash:d,...g}=i;if(v(g)!==d)throw new Error(`Refiner invocation ${i.id} failed content-hash validation.`)}return o}h(wn,"parseCheckpoint");function Er(e,t){return Ir.join(e,"training","harness-refiner-benchmarks",t,"sequential-adaptation.json")}h(Er,"checkpointPath");function vn(e,t){return e.id===t.id&&e.contentHash===t.contentHash}h(vn,"sameRef");async function Ar(e){let t=e.adapters?.runRefinerAfterAttempt??yr,a=e.adapters?.loadRuntimeFromRelease??He,n=e.adapters?.buildLineage??br,s=e.adapters?.loadAttemptEvidenceByIds??ct;await e.store.setHarnessBackgroundReviewSettings({workspaceId:e.initialRuntime.workspace.id,enabled:!0,updatedAt:e.now()});let o={id:e.initialRuntime.release.harnessRelease.id,contentHash:e.initialRuntime.release.harnessRelease.contentHash},i=await Vt({storeDir:e.storeDir,modelRunId:e.modelRun.id,initialHarness:o,now:e.now()});kn(i.steps,e.taskIds);let d=await Rn({store:e.store,initialRuntime:e.initialRuntime,steps:i.steps,loadRuntimeFromRelease:a}),g=i.steps.length?await s({store:e.store,tasksetId:e.taskset.id,attemptIds:i.steps.map(y=>y.attemptId)}):[],f=new Map(g.map(y=>[y.attempt.id,y])),m=[],r=[...i.steps];try{for(let[y,H]of e.taskIds.entries()){if(y<r.length)continue;e.signal.throwIfAborted();let x={id:d.release.harnessRelease.id,contentHash:d.release.harnessRelease.contentHash},I=i.invocations.filter(E=>E.taskId===H),P=I.at(-1);if(P?.status==="failed"&&I.length>=jt)throw new Error(`Sequential Refiner exhausted ${jt} admitted invocations for ${H}.`);let T=P?.status==="failed"?f.get(P.attemptId):void 0;if(!T&&P?.status==="failed"&&([T]=await s({store:e.store,tasksetId:e.taskset.id,attemptIds:[P.attemptId]}),T&&(g.push(T),f.set(T.attempt.id,T))),!T){let E=await e.evaluation.execute({tasksetId:e.taskset.id,taskId:H,model:e.model,reasoningEffort:e.reasoningEffort,seed:e.seed,attempt:0,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:ke(d.release,d.instructionContext),hostedTokenPricing:e.admittedPricing,parentModelRunId:e.modelRun.id,signal:e.signal,toolEvidence:ce(e.evidenceSnapshot,"replay","adaptation")});await e.onAttemptComplete(E),T={attempt:E.attempt,grade:E.grade,artifacts:E.artifacts,receiptContentHash:E.portable.receipt.contentHash,artifactManifest:E.portable.artifactManifest,rewardReceipt:E.portable.rewardReceipt},g.push(T),f.set(T.attempt.id,T)}let S;try{S=await t({store:e.store,storeDir:e.storeDir,modelRun:e.modelRun,model:e.model,taskset:e.taskset,runtime:d,result:T,budget:e.budget,admittedPricing:e.admittedPricing,refinerStream:e.refinerStream,signal:e.signal,now:e.now})}catch(E){throw E instanceof De&&(i=await Wt({storeDir:e.storeDir,modelRunId:e.modelRun.id,initialHarness:o,invocation:Lt({taskId:H,attemptId:T.attempt.id,invocationOrdinal:I.length,trigger:{id:E.details.trigger.id,contentHash:E.details.trigger.contentHash},status:"failed",outcome:null,failure:{kind:E.details.failureKind,message:E.message.slice(0,2e3),retryable:E.details.retryable},inputHarness:x,outputHarness:x,usage:E.details.usage,costBasis:E.details.costBasis,estimatedCostUsd:E.details.estimatedCostUsd,startedAt:E.details.startedAt,completedAt:E.details.completedAt}),now:E.details.completedAt})),E}S.result&&m.push(S.result);let D=await e.store.getHarnessWorkspace(d.workspace.id),A=D?.currentChannel.release;if(!D||!A)throw new Error("Sequential Refiner did not retain a current Harness release.");let U=await e.store.getHarnessReleaseRecord(A.contentHash);if(!U)throw new Error("Sequential Refiner release is unavailable.");d=await a({workspace:D,release:U});let O={id:d.release.harnessRelease.id,contentHash:d.release.harnessRelease.contentHash},W={ordinal:y,taskId:H,attemptId:T.attempt.id,inputHarness:x,outputHarness:O,trigger:{id:S.detection.trigger.id,contentHash:S.detection.trigger.contentHash,decision:S.detection.trigger.decision},outcome:S.result?{id:S.result.outcome.id,contentHash:S.result.outcome.contentHash,decision:S.result.outcome.decision}:null,changed:x.contentHash!==O.contentHash};r.push(W);let L=S.result&&S.invocation?Lt({taskId:H,attemptId:T.attempt.id,invocationOrdinal:I.length,trigger:{id:S.detection.trigger.id,contentHash:S.detection.trigger.contentHash},status:"completed",outcome:{id:S.result.outcome.id,contentHash:S.result.outcome.contentHash,decision:S.result.outcome.decision},failure:null,inputHarness:x,outputHarness:O,usage:S.invocation.usage,costBasis:S.invocation.costBasis,estimatedCostUsd:S.invocation.estimatedCostUsd,startedAt:S.invocation.startedAt,completedAt:S.invocation.completedAt}):void 0;i=await Wt({storeDir:e.storeDir,modelRunId:e.modelRun.id,initialHarness:o,invocation:L,step:W,now:e.now()})}}finally{await e.store.setHarnessBackgroundReviewSettings({workspaceId:e.initialRuntime.workspace.id,enabled:!1,updatedAt:e.now()})}let p=g.reduce((y,H)=>{let x=Array.isArray(H.attempt.metadata.usage)?H.attempt.metadata.usage:[H.attempt.metadata.usage];for(let I of x){if(!I||typeof I!="object"||Array.isArray(I))continue;let P=I,T=Jt(P,["promptTokens","prompt_tokens","inputTokens","input_tokens"]),S=Jt(P,["completionTokens","completion_tokens","outputTokens","output_tokens"]);y.inputTokens+=T,y.outputTokens+=S,y.totalTokens+=Jt(P,["totalTokens","total_tokens"])||T+S}return y},{inputTokens:0,outputTokens:0,totalTokens:0}),l=g.flatMap(y=>typeof y.attempt.costUsd=="number"?[y.attempt.costUsd]:[]),c={id:d.release.harnessRelease.id,contentHash:d.release.harnessRelease.contentHash},w={schemaVersion:"openpond.sequentialHarnessAdaptation.v1",id:`sequential-adaptation-${e.modelRun.id}`,initialHarness:o,finalHarness:c,attemptCount:g.length,passedCount:g.filter(y=>y.grade.passed).length,terminalCount:g.filter(y=>!y.attempt.infrastructureError).length,usage:p,costUsd:l.length?l.reduce((y,H)=>y+H,0):null,latencyMs:g.reduce((y,H)=>y+H.attempt.latencyMs,0),steps:r,createdAt:e.now()},u={...w,contentHash:v(w)},k=await e.store.listHarnessImprovementArtifacts(e.initialRuntime.workspace.id,"refiner_outcome",1e3),R={id:`benchmark-refiner-${e.modelRun.id}`,contentHash:v(k),outcomeCount:k.length},b=await n({store:e.store,workspaceId:e.initialRuntime.workspace.id,adaptationAttempts:g,completedSteps:r,candidateRelease:c,refinerInputHash:v(g.map(y=>({attempt:y.attempt.id,receipt:y.receiptContentHash,grade:v(y.grade)})))});return{attempts:g,runtime:d,summary:u,refinerStage:R,lineage:b,refinerResults:m}}h(Ar,"runSequentialHarnessAdaptation");function kn(e,t){for(let[a,n]of e.entries())if(n.ordinal!==a||n.taskId!==t[a])throw new Error("Sequential adaptation checkpoint drifted from the admitted task order.")}h(kn,"assertCheckpointTaskOrder");async function Rn(e){let t=e.steps.at(-1);if(!t)return e.initialRuntime;let a=await e.store.getHarnessWorkspace(e.initialRuntime.workspace.id);if(!a?.currentChannel.release||a.currentChannel.release.id!==t.outputHarness.id||a.currentChannel.release.contentHash!==t.outputHarness.contentHash)throw new Error("Sequential adaptation workspace drifted from its durable checkpoint.");let n=await e.store.getHarnessReleaseRecord(t.outputHarness.contentHash);if(!n||n.harnessRelease.id!==t.outputHarness.id)throw new Error("Sequential adaptation checkpoint release is unavailable.");return e.loadRuntimeFromRelease({workspace:a,release:n})}h(Rn,"restoreCheckpointRuntime");function Jt(e,t){for(let a of t){let n=e[a];if(typeof n=="number"&&Number.isFinite(n)&&n>=0)return Math.trunc(n)}return 0}h(Jt,"numericToken");var Gt=class extends Error{static{h(this,"BenchmarkRunCancelledError")}constructor(){super("Benchmark cancelled by operator."),this.name="BenchmarkRunCancelledError"}},me=new Map;function Ts(e){let t=e.now??(()=>new Date().toISOString());async function a(f){let{prepared:m,project:r,existing:p}=await zt(e,f,t);if(p&&m.status!=="prepared")return m;let l=p?await e.store.saveModelRun(F.parse({...m,status:"running",updatedAt:t()})):m,c=l.id,w=new AbortController,u=d({input:f,project:r,modelRun:l,signal:w.signal}).finally(()=>me.delete(c));return me.set(c,{controller:w,execution:u}),u.catch(()=>{}),l}h(a,"start");async function n(f){let m=await zt(e,f,t);return m.existing?m.prepared:e.store.saveModelRun(F.parse({...m.prepared,status:"prepared",updatedAt:t()}))}h(n,"prepare");async function s(f){let m=await e.store.getModelRun(f);if(!m||m.kind!=="evaluation"||m.evaluation?.benchmarkId!=="harness-refiner")throw new Error("No Harness Refiner evaluation exists for this Model Run.");if(m.status!=="running")return m;me.get(f)?.controller.abort(new Gt);let r=t();return e.store.saveModelRun(F.parse({...m,status:"cancelled",failure:"Benchmark cancelled by operator.",completedAt:r,updatedAt:r}))}h(s,"cancel");async function o(f){let m=await e.store.getModelRun(f);if(!m||m.kind!=="evaluation"||m.evaluation?.benchmarkId!=="harness-refiner")throw new Error("No Harness Refiner evaluation exists for this Model Run.");if(me.has(f)||m.status==="running")return m;let r=m.evaluationProgress?.stage==="refiner"&&m.evaluationProgress?.completedAttempts===ne(m.evaluation.attemptPlan,"candidate_adaptation"),p=ne(m.evaluation.attemptPlan,"candidate_adaptation"),l=p+Ee(m.evaluation.attemptPlan,"candidate_adaptation").attemptCount,c=await bt({storeDir:e.storeDir,modelRunId:m.id}),w=m.evaluationProgress?.stage==="candidate_adaptation"&&m.evaluationProgress.completedAttempts>=p&&m.evaluationProgress.completedAttempts<=l&&!!c,k=["candidate_adaptation","candidate","comparison"].includes(m.evaluationProgress?.stage??"")&&m.evaluationProgress?.completedAttempts===qe(m.evaluation.attemptPlan)&&!!await vt(e.storeDir,m.id);if(!["failed","cancelled"].includes(m.status)||m.evaluation.advancedEvaluation?.mode!=="review_quality"&&!r&&!w&&!k)throw new Error("Only a Harness Refiner run with a durable Refiner, sequential adaptation, or comparison checkpoint can resume.");let R=e.resolveEvaluationProject?await e.resolveEvaluationProject(m.modelId,m.profileId):await lt(e.store,m.modelId,m.profileId),b=await e.store.saveModelRun(F.parse({...m,status:"running",receipt:null,failure:null,completedAt:null,updatedAt:t()})),y={modelId:m.modelId,profileId:m.profileId,model:m.evaluation.model,reasoningEffort:m.evaluation.reasoningEffort==="none"?"none":m.evaluation.reasoningEffort===null?null:Ma.parse(m.evaluation.reasoningEffort),seeds:m.evaluation.seeds,repetitions:m.evaluation.repetitions,maximumSpendUsd:m.evaluation.maximumSpendUsd,...m.evaluation.advancedEvaluation?{advancedEvaluation:m.evaluation.advancedEvaluation,selectedTaskset:m.taskset}:{}},H=new AbortController,x=d({input:y,project:R,modelRun:b,signal:H.signal}).finally(()=>me.delete(f));return me.set(f,{controller:H,execution:x}),x.catch(()=>{}),b}h(o,"resume");async function i(){let f=(await e.store.listModelRuns()).filter(r=>r.kind==="evaluation"&&r.evaluation?.benchmarkId==="harness-refiner"&&r.status==="running"),m=t();for(let r of f)await e.store.saveModelRun(F.parse({...r,status:"failed",failure:"Benchmark execution was interrupted when the app server stopped. Resume it when an exact durable Refiner checkpoint is available; otherwise start a new run to preserve a complete causal record.",completedAt:m,updatedAt:m}));return f.length}h(i,"reconcileInterrupted");async function d(f){let m=f.modelRun.evaluation?.benchmarkId==="harness-refiner"?f.modelRun.evaluation.advancedEvaluation:void 0;if(!m)return g(f);if(!e.advancedBoundary)throw new Error("Advanced paid admission is unavailable.");let r=AbortSignal.any([f.signal,AbortSignal.timeout(m.maximumDurationMs)]);try{let p=await e.advancedBoundary.run(f.modelRun.id,m,()=>m.mode==="review_quality"?e.reviewQuality?.({modelRun:f.modelRun,signal:r})??Promise.reject(new Error("The actual review-quality evaluator is unavailable.")):g({...f,signal:r}));return e.store.saveModelRun(F.parse({...p,evaluationProgress:p.evaluationProgress?{...p.evaluationProgress,advancedAccounting:e.advancedBoundary.accounting(p.id,m)}:null}))}catch(p){let l=await e.store.getModelRun(f.modelRun.id);if(!l)throw p;let c=t();return e.store.saveModelRun(F.parse({...l,status:r.aborted||l.status==="cancelled"?"cancelled":"failed",failure:$t(p),evaluationProgress:l.evaluationProgress?{...l.evaluationProgress,advancedAccounting:e.advancedBoundary.accounting(l.id,m)}:null,completedAt:c,updatedAt:c}))}}h(d,"executePass");async function g(f){let{input:m,modelRun:r}=f;try{f.signal.throwIfAborted();let p=r.evaluation;if(p?.benchmarkId!=="harness-refiner")throw new Error("Harness Refiner Model Run has the wrong evaluation configuration.");if(!r.harnessRelease)throw new Error("Harness Refiner Model Run has no admitted Harness release.");let l=`benchmark-${r.id}`,[c,w]=await Promise.all([e.store.getHarnessWorkspace(l),e.store.getHarnessReleaseRecord(r.harnessRelease.contentHash)]);if(!c||!w||w.workspaceId!==c.id)throw new Error("The Harness pinned at admission is unavailable.");let u={workspace:c,release:w},k=await e.store.getTaskset(r.taskset.id);if(!k?.benchmark)throw new Error("Harness Refiner Taskset changed before execution.");if(k.revision!==r.taskset.revision||k.contentHash!==r.taskset.contentHash)throw new Error("Harness Refiner Taskset release drifted after admission.");let R=ot({taskset:k,seeds:m.seeds,repetitions:m.repetitions,...p.advancedEvaluation?{selectedPopulation:p.advancedEvaluation}:{}});if(v(R)!==v(p.attemptPlan))throw new Error("Harness Refiner execution plan changed after admission.");let b=p.upstreamModel;if(!b?.pricing)throw new Error("Harness Refiner run has no admitted upstream pricing.");let y=b.pricing,H=qe(R),x=Ee(R,"baseline"),I=Ee(R,"adaptation"),P=Ee(R,"candidate_adaptation"),T=Ee(R,"candidate"),S=r.evaluationProgress?.stage==="refiner"&&r.evaluationProgress.completedAttempts===ne(R,"candidate_adaptation"),D=ne(R,"candidate_adaptation"),A=D+P.attemptCount,U=await bt({storeDir:e.storeDir,modelRunId:r.id}),O=r.evaluationProgress?.stage==="candidate_adaptation"&&r.evaluationProgress.completedAttempts>=D&&r.evaluationProgress.completedAttempts<=A&&!!U,W=["candidate_adaptation","candidate","comparison"].includes(r.evaluationProgress?.stage??"")&&r.evaluationProgress?.completedAttempts===H,L=new nt(m.maximumSpendUsd,S||O||W?r.evaluationProgress?.accounting?.observedSpendUsd??0:0);if(W)return Rr({deps:{store:e.store,storeDir:e.storeDir,evaluation:e.evaluation,loadProfileState:e.loadProfileState,now:t},benchmarkInput:m,modelRun:r,signal:f.signal,workspace:u.workspace,taskset:k,executionPlan:R,baselinePlan:x,adaptationPlan:I,candidatePlan:T,admittedPricing:y,budget:L,totalAttempts:H});let E=new ge,Q=h((te,Te)=>{let ra=r.evaluationProgress?.accounting?.attempts.filter(Ce=>Ce.phase===te).length??0;return async Ce=>{f.signal.throwIfAborted(),L.assertAvailable(Te),L.charge(Ce.attempt.costUsd,`${Te} foreground work`);let na=e.evaluation.consumeGraderUsage([Ce.attempt.id]);L.charge(na.costUsd,`${Te} grading`),ra+=1,await mt(e.store,r.id,{stage:te,completedAttempts:ne(R,te)+ra,totalAttempts:H,result:Ce,grader:na,observedSpendUsd:L.observedSpendUsd})}},"observeStageAttempts"),$=await He({workspace:u.workspace,release:u.release}),X=ke($.release,$.instructionContext),C,B,J;if(S||O)C=await ve({store:e.store,modelRunId:r.id,tasksetId:k.id,plan:x}),B=await ve({store:e.store,modelRunId:r.id,tasksetId:k.id,plan:I}),J=B.attempts,E=await pt({store:e.store,storeDir:e.storeDir,modelRun:r,attempts:[...C.attempts.map(te=>({result:te,cohort:"held_out"})),...B.attempts.map(te=>({result:te,cohort:"adaptation"}))]});else{let te=await e.evaluation.executeBenchmark({tasksetId:k.id,phase:"baseline",model:m.model,reasoningEffort:m.reasoningEffort,seeds:m.seeds,repetitions:m.repetitions,split:I.split,taskIds:I.taskIds,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:X,hostedTokenPricing:y,parentModelRunId:r.id,signal:f.signal,toolEvidence:ce(E,"record","adaptation"),onAttemptComplete:Q("adaptation","adaptation baseline")});if(B=Ft(te),J=B.attempts,await Ut({store:e.store,project:f.project,modelRun:r,model:m.model,baseline:te.attempts[0]}),B.run.harnessRelease.id!==r.harnessRelease.id||B.run.harnessRelease.contentHash!==r.harnessRelease.contentHash)throw new Error("Adaptation baseline drifted from the admitted Harness release.");await ut(e.store,e.storeDir,r.id,E.manifest()),await ie(e.store,r.id,{stage:"baseline",completedAttempts:ne(R,"baseline"),totalAttempts:H});let Te=await e.evaluation.executeBenchmark({tasksetId:k.id,phase:"baseline",model:m.model,reasoningEffort:m.reasoningEffort,seeds:m.seeds,repetitions:m.repetitions,split:x.split,taskIds:x.taskIds,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:X,hostedTokenPricing:y,parentModelRunId:r.id,signal:f.signal,toolEvidence:ce(E,"record","held_out"),onAttemptComplete:Q("baseline","held-out baseline")});if(C=Ft(Te),C.run.harnessRelease.id!==r.harnessRelease.id||C.run.harnessRelease.contentHash!==r.harnessRelease.contentHash)throw new Error("Held-out baseline drifted from the admitted Harness release.");await ut(e.store,e.storeDir,r.id,E.manifest()),await ie(e.store,r.id,{stage:"refiner",completedAttempts:ne(R,"candidate_adaptation"),totalAttempts:H})}if((S||O)&&!r.evaluationProgress?.evidenceSnapshot)throw new Error("Sequential adaptation cannot resume because the interrupted run did not preserve its exact frozen evidence snapshot. Start a new run to preserve a valid causal comparison.");O||await ie(e.store,r.id,{stage:"candidate_adaptation",completedAttempts:D,totalAttempts:H});let _=await Ar({store:e.store,storeDir:e.storeDir,evaluation:e.evaluation,modelRun:r,model:m.model,reasoningEffort:m.reasoningEffort,taskset:k,taskIds:P.taskIds,seed:m.seeds[0],initialRuntime:$,evidenceSnapshot:E,budget:L,admittedPricing:y,refinerStream:e.refinerStream,signal:f.signal,onAttemptComplete:Q("candidate_adaptation","sequential adaptation treatment"),now:t}),Z=_.runtime,V=_.refinerStage,ae=_.lineage,G=E.manifest(),Ue=C.run.harnessRelease.contentHash!==_.summary.finalHarness.contentHash,ye=ke(Z.release,Z.instructionContext);await ie(e.store,r.id,{stage:"candidate",completedAttempts:ne(R,"candidate"),totalAttempts:H});let j=await e.evaluation.executeBenchmark({tasksetId:k.id,phase:"candidate",model:m.model,reasoningEffort:m.reasoningEffort,seeds:m.seeds,repetitions:m.repetitions,split:T.split,taskIds:T.taskIds,sampling:{maxOutputTokens:4096,temperature:0,topP:1},releasedHarness:ye,hostedTokenPricing:y,parentModelRunId:r.id,signal:f.signal,toolEvidence:ce(E,"replay","held_out"),onAttemptComplete:Q("candidate","held-out candidate")});if(!j.comparison)throw new Error("Paired benchmark comparison was not produced.");await ie(e.store,r.id,{stage:"comparison",completedAttempts:H,totalAttempts:H});let ue=gt({modelRunId:r.id,model:m.model,upstreamModel:p.upstreamModel,reasoningEffort:m.reasoningEffort,baseline:C.run,adaptation:B.run,refiner:V,candidateAdaptation:_.summary,candidate:j.run,comparison:j.comparison,executionPlan:R,evidenceSnapshot:G,lineage:ae,createdAt:t()}),At=await wt(e.storeDir,r.id,ue),ee=await e.loadProfileState(),pe=await kt({profile:ee,modelRunId:r.id,workspaceId:u.workspace.id,storeDir:e.storeDir,manifest:ue});await Ut({store:e.store,project:f.project,modelRun:r,model:m.model,baseline:C.attempts[0]});let xt=dt([...C.attempts,...J,..._.attempts,...j.attempts]),Pt=ht({comparison:j.comparison,baseline:C.run,adaptation:B.run,candidate:j.run,harnessChanged:Ue,candidateAdaptation:_.summary,lineageValid:ae.valid,infrastructureValid:xt}),zr=ft({baseline:C.run,adaptation:B.run,candidateAdaptation:_.summary,candidate:j.run,harnessChanged:Ue,lineageValid:ae.valid,infrastructureValid:xt}),Pe=(await e.store.getModelRun(r.id)??r).evaluationProgress?.accounting??de(),Ur=Ye({attempts:Pe.attempts,targetTaskCount:st(R)}).summary,Lr=it({baselineTokens:C.run.usage.totalTokens,candidateTokens:j.run.usage.totalTokens,refinerTokens:Pe.usage.refiner.totalTokens,graderTokens:Pe.usage.grader.totalTokens}),Zt={schemaVersion:"openpond.modelEvaluationReceipt.v1",benchmarkId:"harness-refiner",resultManifest:{id:ue.id,contentHash:ue.contentHash,artifactPath:At},stages:{baseline:{id:C.run.id,contentHash:C.run.contentHash},adaptation:{id:B.run.id,contentHash:B.run.contentHash},candidateAdaptation:{id:_.summary.id,contentHash:_.summary.contentHash},refiner:{id:V.id,contentHash:V.contentHash},candidate:{id:j.run.id,contentHash:j.run.contentHash},comparison:{id:j.comparison.id,contentHash:j.comparison.contentHash}},usage:Pe.usage,quality:{baselinePassRate:j.comparison.baselinePassRate,candidatePassRate:j.comparison.candidatePassRate,adaptationBaselinePassRate:B.run.passedCount/B.run.attemptCount,adaptationCandidatePassRate:_.summary.passedCount/_.summary.attemptCount,adaptationCandidatePassed:_.summary.passedCount===_.summary.attemptCount,heldOutCandidatePassed:j.run.passedCount===j.run.attemptCount,passed:j.comparison.qualityPassed&&_.summary.passedCount===_.summary.attemptCount&&j.run.passedCount===j.run.attemptCount&&xt&&Pt!=="infrastructure_failure"},foregroundTokenDelta:j.comparison.foregroundTokenDelta,foregroundTokenDeltaPercent:j.comparison.foregroundTokenDeltaPercent,taskEfficiency:Ur,efficiency:Lr,budget:{maximumSpendUsd:L.maximumSpendUsd,observedSpendUsd:L.observedSpendUsd,enforced:!0},evidenceSnapshot:{id:G.id,contentHash:G.contentHash},lineage:ae,invalidReasons:zr,attempts:Pe.attempts,terminalClassification:Pt,profileGit:pe?{ref:pe.ref,commit:pe.commit,baseCommit:pe.baseCommit}:null},Vr=Qe.parse({...Zt,contentHash:v(Zt)}),ea=t(),ta=await e.store.getModelRun(r.id)??r,aa=Pt==="infrastructure_failure";return e.store.saveModelRun(F.parse({...ta,status:aa?"failed":"succeeded",harnessRelease:C.run.harnessRelease,receipt:Vr,failure:aa?"Benchmark attempts did not reach the model because the Work runtime failed.":null,evaluationProgress:{...ta.evaluationProgress,stage:"comparison",completedAttempts:H,totalAttempts:H},completedAt:ea,updatedAt:ea}))}catch(p){let l=t(),c=await e.store.getModelRun(r.id)??r,w=f.signal.aborted||c.status==="cancelled";return e.store.saveModelRun(F.parse({...c,status:w?"cancelled":"failed",failure:w?"Benchmark cancelled by operator.":$t(p),completedAt:l,updatedAt:l}))}}return h(g,"executePassInternal"),{async close(){let f=[...me.values()];for(let m of f)m.controller.abort(new Error("The Refiner evaluation owner is closing."));await Promise.allSettled(f.map(m=>m.execution))},reconcileInterrupted:i,prepare:n,cancel:s,resume:o,start:a,wait:h(async f=>me.get(f)?.execution??e.store.getModelRun(f),"wait")}}h(Ts,"createHarnessRefinerBenchmarkService");import{mkdir as bn,readFile as yn,rename as In,writeFile as Sn}from"node:fs/promises";import xr from"node:path";import{randomUUID as Hn}from"node:crypto";var En=M.object({outcome:ma,proposal:ia.nullable(),validations:M.array(da),overlay:sa,workspace:Le,advanceReceipt:ca.nullable(),applyReceipt:ua.nullable()}),Pr=M.object({runId:M.string(),taskId:M.string(),actorId:M.string(),teamId:M.string(),pinHash:M.string(),sourceCaseHash:M.string(),sessionId:M.string(),turnId:M.string(),workspaceId:M.string(),trigger:la,result:En.nullable(),receipt:za.nullable(),cleanupComplete:M.boolean(),contentHash:M.string()});function Tr(e){let t=h((s,o)=>xr.join(e,"training","advanced-refiner-reviews",v([s,o])+".json"),"file");async function a(s,o,i){let d;try{d=JSON.parse(await yn(t(s,o),"utf8"))}catch(r){if(r.code==="ENOENT")return null;throw r}let g=Pr.parse(d),{contentHash:f,...m}=g;if(f!==v(m)||g.runId!==s||g.taskId!==o||g.actorId!==i.actorId||g.teamId!==i.teamId||g.pinHash!==i.contentHash)throw new Error("The retained review case belongs to another immutable admission.");return g}h(a,"read");async function n(s){let o=Pr.parse({...s,contentHash:v(s)}),i=t(o.runId,o.taskId),d=i+"."+Hn();return await bn(xr.dirname(i),{recursive:!0}),await Sn(d,JSON.stringify(o),{mode:384}),await In(d,i),o}return h(n,"write"),{read:a,write:n}}h(Tr,"createAdvancedReviewCheckpoints");function Ys(e){let t=Tr(e.benchmark.storeDir);return async a=>{let n=a.modelRun,s=n.evaluation?.benchmarkId==="harness-refiner"?n.evaluation.advancedEvaluation:void 0;if(!s||s.mode!=="review_quality")throw new Error("Review-quality admission is unavailable.");let o=await e.source.read(s),i=await e.benchmark.store.getTaskset(n.taskset.id);if(!i)throw new Error("The exact review Taskset is unavailable.");let d=o.packageValue.taskset.metadata.advancedRefinerReview;if(d?.outputConvention!=="openpond.refinerReviewOutput.v1"||v(d.contextEvidence)!==v(s.evidence))throw new Error("Publish a review-quality Dataset with independently grounded labels and the exact retained context-evidence binding before starting this target.");let g=await e.source.evidence(s),f=(await Ze(e.benchmark.storeDir)).releases.find(u=>u.id===s.refinerRelease.id&&u.contentHash===s.refinerRelease.contentHash);if(!f)throw new Error("The admitted Refiner release is unavailable.");let m=[];for(let u of s.holdoutTaskIds){a.signal.throwIfAborted(),await o.authorize();let k=i.tasks.find(S=>S.id===u),R=g.result.cases.find(S=>S.identity.caseId===u);if(!k||!R||R.status!=="completed")throw new Error("The exact independent review context is unavailable.");let b=await t.read(n.id,u,s);if(b&&b.sourceCaseHash!==v(R))throw new Error("The retained independent review context changed.");if(b?.receipt&&b.cleanupComplete){m.push(b.receipt);continue}if(b?.cleanupComplete&&!b.result&&b.trigger.decision==="queue_refiner")throw new Error("This interrupted Refiner case cleaned its overlay before receiving a settled outcome; reconcile the retained paid request before resuming.");let y=await e.source.reviewContext(s,u),H=await e.benchmark.store.getHarnessWorkspace(`review-${v([n.id,u]).slice(0,40)}`),x=H?.currentChannel.release&&await e.benchmark.store.getHarnessReleaseRecord(H.currentChannel.release.contentHash),I=H&&x?{workspace:H,release:x}:await Ke({store:e.benchmark.store,storeDir:e.benchmark.storeDir,id:`review-${v([n.id,u]).slice(0,40)}`,ownerId:`benchmark:${n.id}`,name:`Review ${u}`,sourceRelease:s.baselineRelease}),P=null,T=!1;try{let S=await e.benchmark.store.sessionShells(),A=(b?await e.benchmark.store.getSession(b.sessionId):S.find(C=>C.metadata?.advancedRunId===n.id&&C.metadata?.advancedTaskId===u&&C.metadata?.actorId===s.actorId&&C.cloudTeamId===s.teamId))??await e.createSession({experience:"work",provider:"openpond",modelRef:n.evaluation.model,hiddenFromDefaultSidebar:!0,openPondCommandAccessMode:"disabled",cwd:null,title:`Evaluate Refiner ${u}`,cloudTeamId:s.teamId,metadata:{source:"advanced-refiner-evaluation",automatedTasksetWorkAttempt:!0,advancedRunId:n.id,advancedTaskId:u,actorId:s.actorId,sourceEvidence:s.evidence,sourceCaseHash:v(R)}});P=A.id;let U=new Date().toISOString(),O=`review-turn-${v([n.id,u]).slice(0,40)}`,W=await et({store:e.benchmark.store,runId:A.id,workspace:I.workspace,harnessRelease:{id:I.release.harnessRelease.id,contentHash:I.release.harnessRelease.contentHash},admittedAt:U}),L=Oe.parse({id:O,sessionId:A.id,providerTurnId:null,modelRef:n.evaluation.model,prompt:y.turn.prompt,startedAt:U,completedAt:U,status:"completed",error:null,metadata:{source:"advanced-refiner-evaluation",automatedTasksetWorkAttempt:!0,sourceCaseHash:v(R)},createImproveRun:null,profileSnapshot:null,harnessSnapshot:{schemaVersion:"openpond.harnessTurnSnapshot.v1",workspaceId:I.workspace.id,workspaceRevision:I.workspace.revision,sourceRevision:I.workspace.sourceRevision,channelName:I.workspace.currentChannel.name,channelRevision:I.workspace.currentChannel.revision,harnessRelease:W.baseHarnessRelease,overlay:{id:W.id,revision:W.revision,contentHash:W.contentHash}}});if(!b){await e.benchmark.store.getTurn(O)||await e.benchmark.store.insertTurn(L);let C=await e.benchmark.store.runtimeEventsForTurn(O);for(let B of y.events){if(C.some(ae=>ae.data?.reviewContextOrigin?.eventId===B.id))continue;let{id:J,timestamp:_,sequence:Z,...V}=B;await e.benchmark.store.appendRuntimeEvent(he({...V,sessionId:A.id,turnId:O,appId:A.appId,data:{...typeof V.data=="object"&&V.data?V.data:{},reviewContextOrigin:{eventId:J,eventHash:v(B),traceHash:y.traceHash,sourceEvidence:s.evidence}}}))}C.some(B=>B.name==="assistant.delta"&&B.data?.sourceCaseHash===v(R))||await e.benchmark.store.appendRuntimeEvent(he({sessionId:A.id,turnId:O,name:"assistant.delta",source:"server",appId:A.appId,status:"completed",output:y.case.output??void 0,data:{sourceEvidence:s.evidence,sourceCaseHash:v(R)}}))}await e.benchmark.store.setHarnessBackgroundReviewSettings({workspaceId:I.workspace.id,enabled:!0,updatedAt:new Date().toISOString()});let E=b?{trigger:b.trigger}:await tt({store:e.benchmark.store,session:A,turn:L,boundaryKind:"turn_completed"});if(!E)throw new Error("The actual context produced no retained trigger decision.");let Q=h(async C=>{let B=b?(({contentHash:J,..._})=>_)(b):{runId:n.id,taskId:u,actorId:s.actorId,teamId:s.teamId,pinHash:s.contentHash,sourceCaseHash:v(R),sessionId:A.id,turnId:O,workspaceId:I.workspace.id,trigger:E.trigger,result:null,receipt:null,cleanupComplete:!1};b=await t.write({...B,...C})},"saveCheckpoint");b||await Q({});let $=b?.result??null,X=!1;try{!$&&E.trigger.decision==="queue_refiner"&&($=await at({store:e.benchmark.store,storeDir:e.benchmark.storeDir,trigger:E.trigger,loadActiveRefinerRelease:h(async()=>f,"loadActiveRefinerRelease"),signal:a.signal,stream:h(async function*(V){yield*e.benchmark.refinerStream({...V,model:n.evaluation.model,pricing:n.evaluation.upstreamModel.pricing,requestId:`review-refiner:${n.id}:${u}`})},"stream")})),await Q({result:$});let C={schemaVersion:"openpond.refinerReviewOutput.v1",decision:E.trigger.decision,trigger:E.trigger,outcome:$?.outcome??null,proposal:$?.proposal??null,validations:$?.validations??[],sourceContext:{evidence:s.evidence,caseHash:v(R)}},B=Na.parse({schemaVersion:"openpond.taskAttempt.v1",id:`review-attempt-${v([n.id,u]).slice(0,40)}`,tasksetId:i.id,taskId:u,split:k.split,attempt:0,modelRef:n.evaluation.model,seed:n.evaluation.seeds[0],output:{text:JSON.stringify(C),json:C},runtimeEventRefs:[],artifactRefs:[],privilegedOutcomeRef:null,infrastructureError:null,latencyMs:Date.now()-Date.parse(U),costUsd:null,startedAt:U,completedAt:new Date().toISOString(),metadata:{execution:"advanced_refiner_review",runId:n.id,sourceCaseHash:v(R),refinerRelease:s.refinerRelease}}),J=await e.benchmark.evaluation.grade({tasksetId:i.id,taskId:u,attempt:B,hostedTokenPricing:n.evaluation.upstreamModel.pricing}),_=await e.benchmark.store.getHarnessRunOverlay(A.id);if(!_)throw new Error("The actual isolated review overlay is unavailable.");_.status!=="abandoned"&&await e.benchmark.store.abandonHarnessRunOverlayAtomically({runId:A.id,expectedRevision:_.revision,updatedAt:new Date().toISOString()}),X=!0;let Z={taskId:u,sourceCaseHash:v(R),reviewSessionId:A.id,reviewTurnId:O,inputHarness:W.baseHarnessRelease,trigger:{id:E.trigger.id,contentHash:E.trigger.contentHash},decision:E.trigger.decision,outcome:$?{id:$.outcome.id,contentHash:$.outcome.contentHash}:null,proposal:$?.proposal?{id:$.proposal.id,contentHash:$.proposal.contentHash}:null,validations:($?.validations??[]).map(V=>({id:V.id,contentHash:V.contentHash})),grade:{id:J.id,contentHash:v(J),score:J.score,passed:J.passed,failureClass:J.failureClass},resetComplete:X};await Q({receipt:Z,cleanupComplete:!0}),m.push(Z)}finally{if(await e.benchmark.store.setHarnessBackgroundReviewSettings({workspaceId:I.workspace.id,enabled:!1,updatedAt:new Date().toISOString()}),!X){let C=await e.benchmark.store.getHarnessRunOverlay(A.id);C&&C.status!=="abandoned"&&await e.benchmark.store.abandonHarnessRunOverlayAtomically({runId:A.id,expectedRevision:C.revision,updatedAt:new Date().toISOString()})}await Q({cleanupComplete:!0}),T=!0}}finally{if(!T&&(await e.benchmark.store.setHarnessBackgroundReviewSettings({workspaceId:I.workspace.id,enabled:!1,updatedAt:new Date().toISOString()}),P)){let S=await e.benchmark.store.getHarnessRunOverlay(P);S&&S.status!=="abandoned"&&await e.benchmark.store.abandonHarnessRunOverlayAtomically({runId:P,expectedRevision:S.revision,updatedAt:new Date().toISOString()})}}}let r=e.boundary.read(n.id,s)??[],p=r.some(u=>u.state!=="completed"&&u.state!=="not_dispatched")?null:r.reduce((u,k)=>u+Number(k.actual_spend??0),0),l=m.map(u=>u.grade.score).filter(u=>u!==null),c=new Date().toISOString(),w=Ua({schemaVersion:"openpond.advancedRefinerReviewReceipt.v1",id:`review-receipt-${n.id}`,runId:n.id,pinHash:s.contentHash,ownerActorId:s.actorId,teamId:s.teamId,privateClosureHash:s.privateClosureHash,refinerRelease:s.refinerRelease,cases:m,reviewScore:l.length===m.length?l.reduce((u,k)=>u+k,0)/l.length:null,completedCases:l.length,noActionCases:m.filter(u=>u.decision==="no_action").length,routedCases:m.filter(u=>u.decision==="route_deterministically").length,proposalCases:m.filter(u=>u.proposal!==null).length,costUsd:p,cleanupComplete:m.every(u=>u.resetComplete),createdAt:c});return e.benchmark.store.saveModelRun(F.parse({...n,status:w.reviewScore!==null&&w.cleanupComplete&&p!==null?"succeeded":"failed",receipt:w,reward:null,failure:w.reviewScore===null?"Review grade coverage is incomplete.":null,completedAt:c,updatedAt:c}))}}h(Ys,"createAdvancedRefinerReviewQuality");import{createHash as An}from"node:crypto";import{mkdir as Re,readFile as xn,writeFile as xe}from"node:fs/promises";import N from"node:path";var Pn="harness-refiner";function ci(e){let t=e.now??(()=>new Date().toISOString());async function a(o){let i=Mn(o.profileId),d=await e.store.getTaskset(i);if(d?.benchmark?.releaseId===We.id&&d.benchmark.releaseHash===We.contentHash)return d;let g=N.join(e.storeDir,"training","tasksets",i),f=await Cn({release:We,assets:Ia,tasksetRoot:g,tasksetId:i,profileId:o.profileId,revision:d?d.revision+1:1,timestamp:t()});return await e.store.upsertTaskset(f),f}h(a,"ensureHarnessRefiner");async function n(o){let i=ja(o.package),d=`advanced-refiner-${v([o.profileId,i.contentHash,o.adaptationSplit,o.holdoutSplit]).slice(0,40)}`;if(o.adaptationSplit===o.holdoutSplit)throw new Error("Adaptation cannot include its held-out split.");let g=i.taskset.tasks.filter(u=>u.split===o.adaptationSplit),f=i.taskset.tasks.filter(u=>u.split===o.holdoutSplit);if(!g.length||!f.length)throw new Error("The selected release needs both adaptation and private holdout cases.");let m=new Set(g.map(u=>u.clusterKey));if(f.some(u=>m.has(u.clusterKey)))throw new Error("Private holdout families overlap adaptation data.");let r=await e.store.getTaskset(d);if(r){if(r.metadata.importedPackageHash!==i.contentHash)throw new Error("The selected package binding changed.");return r}let p=qa({package:i,profileId:o.profileId,name:i.taskset.id,createdAt:t()}),l=N.join(e.storeDir,"training","tasksets",d);await Re(N.join(l,"benchmark"),{recursive:!0});for(let u of i.files){let k=N.resolve(l,u.asset.path);if(!k.startsWith(l+N.sep))throw new Error("Selected evaluator asset escapes its private store.");await Re(N.dirname(k),{recursive:!0}),await xe(k,Oa(u))}for(let u of p.generatedFiles){let k=N.resolve(l,u.path);if(!k.startsWith(l+N.sep))throw new Error("Projected verifier escapes its evaluator store.");await Re(N.dirname(k),{recursive:!0}),await xe(k,u.content)}await xe(N.join(l,"benchmark","taskset.release.json"),JSON.stringify(i.taskset),"utf8");let c=Me.parse({...p.taskset,id:d,environment:{...p.taskset.environment,metadata:{...p.taskset.environment.metadata,runtimeSourceTasksetId:d}},purpose:"benchmark",benchmark:{schemaVersion:"openpond.tasksetBenchmark.v1",definitionId:d,releaseId:i.taskset.id,releaseHash:i.taskset.contentHash,managedReleasePath:"benchmark/taskset.release.json",adaptationSplit:o.adaptationSplit,evaluationSplit:o.holdoutSplit,primaryMetric:"success_rate",qualityGate:"non_regression",source:"imported",metadata:{packageHash:i.contentHash}},metadata:{...p.taskset.metadata,advancedRefiner:!0}}),w=Me.parse({...c,contentHash:Tt(c)});return await e.store.upsertTaskset(w),w}h(n,"projectSelected");async function s(o){if(o.purpose!=="benchmark"||!o.benchmark)return null;let i=N.resolve(e.storeDir,"training","tasksets",o.id,o.benchmark.managedReleasePath),d=va.parse(JSON.parse(await xn(i,"utf8")));if(d.id!==o.benchmark.releaseId||d.contentHash!==o.benchmark.releaseHash)throw new Error("Managed benchmark release does not match its Taskset binding.");return d}return h(s,"releaseForTaskset"),{ensureHarnessRefiner:a,projectSelected:n,releaseForTaskset:s}}h(ci,"createBenchmarkTasksetService");function Tn(e){return Ha({schemaVersion:"openpond.benchmarkDefinition.v1",id:Pn,title:"Harness Refiner 20260818 v2",description:"Measures whether sequential, evidence-driven Harness improvements change verified reward on a frozen cohort; foreground tokens are secondary.",tasksetRelease:{id:e.id,contentHash:e.contentHash},adaptationSplit:"validation",evaluationSplit:"frozen_eval",primaryMetric:"success_rate",qualityGate:"non_regression",caseCounts:{adaptation:e.tasks.filter(t=>t.split==="validation").length,evaluation:e.tasks.filter(t=>t.split==="frozen_eval").length},metadata:{builtin:!0,protocol:"sequential_product_lifecycle_v3",orderSeed:"harness-refiner-20260818-order-v1",secondaryMetrics:["foreground_tokens"]}})}h(Tn,"definitionForRelease");async function Cn(e){let t=Sa.parse(Tn(e.release)),a=_n(e.release.tasks.flatMap(l=>l.artifactRefs)),n=new Map;await Re(N.join(e.tasksetRoot,"assets"),{recursive:!0}),await Re(N.join(e.tasksetRoot,"benchmark"),{recursive:!0}),await Re(N.join(e.tasksetRoot,"rubrics"),{recursive:!0}),await Re(N.join(e.tasksetRoot,"verifiers"),{recursive:!0}),await xe(N.join(e.tasksetRoot,"benchmark","taskset.release.json"),JSON.stringify(e.release),"utf8");for(let l of a){let c=Mr(e.assets,l.path);if(v(c)!==l.contentHash)throw new Error(`Benchmark asset ${l.path} failed its content hash.`);let w=`${l.id}-${N.basename(l.path)}`,u=N.posix.join("assets",w);await xe(N.join(e.tasksetRoot,u),c,"utf8"),n.set(l.id,{artifactRef:u,fileName:w,sha256:An("sha256").update(c).digest("hex"),sizeBytes:Buffer.byteLength(c),mediaType:l.mediaType})}let s=e.release.graders.find(l=>l.kind==="custom_verifier"),o=On(e.release);if(!s)throw new Error("Benchmark release requires a custom verifier.");await Cr(e.assets,e.tasksetRoot,o.rubricRef),await Cr(e.assets,e.tasksetRoot,s.verifierRef);let i=`source-${e.tasksetId}`,d=e.release.tasks.map(l=>({schemaVersion:"openpond.taskData.v1",id:l.id,clusterKey:l.clusterKey,split:l.split,input:l.input,expectedOutput:l.expectedOutput,policyVisibleContext:l.policyVisibleContext,privilegedContextRef:l.privilegedContextRef,sourceRefs:[i],assets:l.artifactRefs.map(c=>{let w=n.get(c.id);if(!w)throw new Error(`Managed benchmark asset ${c.id} is missing.`);return{id:c.id,sourceRefId:i,artifactRef:w.artifactRef,fileName:w.fileName,mediaType:w.mediaType,sha256:w.sha256,sizeBytes:w.sizeBytes,split:l.split,metadata:{portableContentHash:c.contentHash}}}),requiredOutputs:(l.requiredOutputs??[]).map(c=>({path:c.path,mediaType:c.mediaType,schemaRef:c.schemaRef?.id??null,maxBytes:c.maxBytes??void 0,metadata:c.metadata})),tags:l.tags,metadata:{portableTaskId:l.id,portableTaskRecord:l}})),g=[...n.values()].map(l=>l.sha256),f=[{id:s.id,version:s.version,label:"Output contract",kind:"custom_verifier",weight:s.weight,hardGate:s.hardGate,rewardEligible:s.rewardEligible,privileged:s.privileged,module:s.verifierRef.path,exportName:"verify",timeoutMs:s.timeoutMs,networkPolicy:"none",metadata:{portableVerifierHash:s.verifierRef.contentHash,portableVerifierRef:s.verifierRef}}],m=d.find(l=>l.id==="adaptation-launch-delay-email")??d[0],r=Bn(m.id),p=Me.parse({schemaVersion:"openpond.taskset.v1",id:e.tasksetId,revision:e.revision,profileId:e.profileId,profileRelease:null,createImproveRunId:null,name:t.title,objective:t.description,purpose:"benchmark",benchmark:{schemaVersion:"openpond.tasksetBenchmark.v1",definitionId:t.id,releaseId:e.release.id,releaseHash:e.release.contentHash,managedReleasePath:"benchmark/taskset.release.json",adaptationSplit:t.adaptationSplit,evaluationSplit:t.evaluationSplit,primaryMetric:t.primaryMetric,qualityGate:t.qualityGate,source:"builtin",metadata:{}},status:"ready",sourceRefs:[{schemaVersion:"openpond.uploadedFileDatasetSource.v1",kind:"uploaded_file",id:i,profileId:e.profileId,title:"Harness Refiner benchmark fixtures",sourceHash:e.release.contentHash,occurredAt:e.timestamp,licensingStatus:"approved",secretScanStatus:"passed",piiScanStatus:"passed",originalFileNames:[...n.values()].map(l=>l.fileName),mediaTypes:[...new Set([...n.values()].map(l=>l.mediaType))],sourceFileHashes:g,totalBytes:[...n.values()].reduce((l,c)=>l+c.sizeBytes,0),parserVersion:"openpond-builtin-benchmark-v1",metadata:{generatedPublicFixture:!0}}],policy:e.release.policy,environment:{protocolVersion:"openpond.taskEnvironment.v1",kind:"work",entrypoint:"openpond-work-v1",stateful:!0,deterministicSeeds:e.release.environment.deterministicSeeds,toolNames:e.release.tools.map(l=>l.name),lifecycle:["create","reset","step","grade","cleanup"],defaultTimeoutMs:e.release.environment.defaultTimeoutMs,networkPolicy:e.release.environment.networkPolicy,metadata:{portableEnvironmentHash:v(e.release.environment),portableTools:e.release.tools,maxToolTurns:40}},capabilities:{schemaVersion:"openpond.tasksetCapabilities.v1",taskKind:"single_agent",supportedSignals:[],compatibleMethods:["none"],rewardKinds:["deterministic"],requiresTools:!0,requiresState:!0,requiresPrivilegedGrading:!0,environmentPlacements:["local","remote"],exportable:!0,portabilityBlockers:[]},tasks:d,graders:f,graderFixtures:r,learningSignals:{demonstrations:[],preferences:[],corrections:[],feedback:[],rewards:[],labels:[]},authoringProvenance:{schemaVersion:"openpond.taskAuthoringProvenance.v1",model:null,modelConfig:{},skillHash:e.release.contentHash,promptTemplateVersion:"public-benchmark-release-v1",buildIntent:"rubric",buildSpecification:null,evidenceHashes:[e.release.contentHash,...g],tasksetSdkVersion:"portable-builtin-release-v1",sourceCommit:null,repairHistory:[],createdAt:e.timestamp},readiness:null,contentHash:"00000000",createdAt:e.timestamp,updatedAt:e.timestamp,metadata:{portableTasksetRelease:{id:e.release.id,contentHash:e.release.contentHash},supplementaryModelJudge:{id:o.id,version:o.version,rubricHash:o.rubricRef.contentHash,calibrationStatus:o.calibrationStatus,executable:!1,rewardEligible:!1},portableCapabilities:e.release.capabilities}});return Me.parse({...p,contentHash:Tt(p)})}h(Cn,"projectRelease");function Mn(e){return`benchmark-harness-refiner-${v(e).slice(0,16)}`}h(Mn,"tasksetIdForProfile");async function Cr(e,t,a){let n=Mr(e,a.path);if(v(n)!==a.contentHash)throw new Error(`Benchmark verifier asset ${a.path} failed its content hash.`);return await xe(N.join(t,a.path),n,"utf8"),n}h(Cr,"copyVerifierAsset");function Mr(e,t){let a=e[t];if(a===void 0)throw new Error(`Built-in benchmark asset ${t} is unavailable.`);return a}h(Mr,"builtinAsset");function Bn(e){return[{id:"benchmark-positive",taskId:e,label:"positive",output:{text:`Subject: Acme pilot launch update
 
 The launch is moving from August 20 to August 27 because final accessibility testing is not complete. Testing is expected to finish August 22, and existing pilot access remains available. Please send questions to pilot-support@example.com.`},infrastructureError:null,expectedPassed:!0,expectedRewardEligible:!0,metadata:{}},{id:"benchmark-negative",taskId:e,label:"negative",output:{text:"Draft saved to /workspace/outputs/acme-launch-email.md. Checklist: new date included; accessibility testing mentioned; pilot access preserved; support address included; under 140 words."},infrastructureError:null,expectedPassed:!1,expectedRewardEligible:!0,metadata:{}},{id:"benchmark-boundary",taskId:e,label:"boundary",output:{text:"The launch is moving to August 27 while testing finishes."},infrastructureError:null,expectedPassed:!1,expectedRewardEligible:!0,metadata:{}}]}h(Bn,"benchmarkFixtures");function _n(e){return[...new Map(e.map(t=>[t.id,t])).values()]}h(_n,"uniqueAssets");function On(e){let t=e.metadata.supplementaryModelJudge;if(!t||typeof t!="object"||Array.isArray(t))throw new Error("Benchmark release has no supplementary model-judge metadata.");let a=t;if(typeof a.id!="string"||typeof a.version!="string"||a.calibrationStatus!=="pending"||a.executable!==!1||a.rewardEligible!==!1)throw new Error("Benchmark supplementary model-judge metadata is invalid.");return{id:a.id,version:a.version,calibrationStatus:a.calibrationStatus,rubricRef:oa.parse(a.rubricRef)}}h(On,"supplementaryModelJudge");import{promises as ze}from"node:fs";import Dr from"node:path";import jn from"node:path";function Br(e){let t=Ba(jn.join(e,"library","harnesses","experiment-improvements.sqlite"));t.exec(`CREATE TABLE IF NOT EXISTS experiment_improvements(id TEXT PRIMARY KEY, actor_id TEXT NOT NULL, team_id TEXT NOT NULL, operation_id TEXT NOT NULL, request_hash TEXT NOT NULL, revision INTEGER NOT NULL, payload TEXT NOT NULL, UNIQUE(actor_id,team_id,operation_id));
-    CREATE TABLE IF NOT EXISTS experiment_improvement_revisions(id TEXT NOT NULL, revision INTEGER NOT NULL, payload TEXT NOT NULL, PRIMARY KEY(id,revision));`);let a=Promise.resolve(),n=h(i=>{let d=a.then(i);return a=d.catch(()=>{}),d},"serialize");function s(i,d,g){let f=t.prepare("SELECT payload FROM experiment_improvements WHERE id=? AND actor_id=? AND team_id=?").get(i,d,g);return f?Se(JSON.parse(String(f.payload))):null}h(s,"row");function o(i){let d=Se(i);return t.prepare("INSERT INTO experiment_improvement_revisions(id,revision,payload) VALUES(?,?,?)").run(i.id,i.revision,JSON.stringify(d)),t.prepare("INSERT INTO experiment_improvements(id,actor_id,team_id,operation_id,request_hash,revision,payload) VALUES(?,?,?,?,?,?,?) ON CONFLICT(id) DO UPDATE SET revision=excluded.revision,payload=excluded.payload").run(i.id,i.actorId,i.teamId,i.operationId,i.requestHash,i.revision,JSON.stringify(d)),d}return h(o,"persist"),{async find(i,d,g){return await a,s(i,d,g)},async read(i,d,g){await a;let f=s(i,d,g);if(!f)throw new Error("Improvement is unavailable to this owner.");return f},async list(i,d,g=30,f){if(await a,g<1||g>50)throw new Error("Improvement list limit is invalid.");return t.prepare("SELECT payload FROM experiment_improvements WHERE actor_id=? AND team_id=? AND id>? ORDER BY id LIMIT ?").all(i,d,f??"",g).map(m=>Se(JSON.parse(String(m.payload))))},create(i){return n(async()=>{t.exec("BEGIN IMMEDIATE");try{let d=t.prepare("SELECT payload FROM experiment_improvements WHERE actor_id=? AND team_id=? AND operation_id=?").get(i.actorId,i.teamId,i.operationId);if(d){let f=Se(JSON.parse(String(d.payload)));if(f.requestHash!==i.requestHash)throw new Error("Improvement operation was reused with different authority or evidence.");return t.exec("COMMIT"),f}let g=o(i);return t.exec("COMMIT"),g}catch(d){throw t.exec("ROLLBACK"),d}})},mutate(i,d,g,f,m){return n(async()=>{t.exec("BEGIN IMMEDIATE");try{let r=s(i,d,g);if(!r||r.revision!==f)throw new Error("Improvement changed; reload its current revision.");let p=await m(r);if(p.id!==r.id||p.actorId!==d||p.teamId!==g||p.operationId!==r.operationId||p.requestHash!==r.requestHash||p.revision!==r.revision+1)throw new Error("Improvement transition changed its immutable owner or operation authority.");let l=o(p);return t.exec("COMMIT"),l}catch(r){throw t.exec("ROLLBACK"),r}})},async close(){await a,t.close()}}}h(Br,"createExperimentImprovementStore");import{randomUUID as _r}from"node:crypto";import{promises as q}from"node:fs";import z from"node:path";function yt(e){if(!e||z.isAbsolute(e)||e.includes("\\")||e.includes("\0")||e.split("/").some(t=>!t||t==="."||t===".."))throw new Error("Candidate source path is not contained.");return e}h(yt,"safe");var be=h((e,t)=>e.writablePaths.some(a=>t===a||t.startsWith(`${a}/`)),"selected");function It(e,t){return z.join(e,"library","harnesses","experiment-candidates",v(t))}h(It,"candidateRoot");async function Or(e){let t=Xa.parse(e),a=z.join(t.bundlePath,"source"),n=await se({workspaceId:t.agentSnapshot.metadata.workspaceId??t.workspaceId,sourceDir:a});if(n.sourceRevision!==t.sourceRevision||n.harnessRelease.contentHash!==t.harnessRelease.contentHash||n.agentSnapshot.contentHash!==t.agentSnapshot.contentHash)throw new Error("Candidate baseline source differs from the immutable released bytes.");return n}h(Or,"verifiedBase");async function jr(e){let t=await Or(e.base),a={...e.component,path:yt(e.component.path)},n=t.manifest.files.find(u=>u.path===a.path);if(!n||n.kind!==a.kind||n.visibility!=="policy")throw new Error("Select a released policy component for candidate authoring.");if(a.kind==="workflow"&&!a.workflowId)throw new Error("Select the exact Workflow within its released catalog.");let s=new Set(t.sourceFiles.filter(u=>u.asset.visibility!=="policy").map(u=>u.asset.contentHash)),o=t.sourceFiles.filter(u=>u.asset.visibility==="policy"&&!s.has(u.asset.contentHash));if(!o.some(u=>u.path===a.path))throw new Error("The selected component duplicates protected private source and cannot be disclosed.");let i=a.kind==="workflow"?`workflows/${yt(a.workflowId)}`:null,g=i&&o.some(u=>u.path===`${i}/PROMPT.md`||u.path===`${i}/ACTION.json`)?[a.path,i]:a.kind==="agent"?[a.path.split("/").slice(0,2).join("/")]:a.kind==="skill"?[z.posix.dirname(a.path)]:[a.path];if(a.kind==="agent"&&(!a.path.startsWith("agents/")||a.path.split("/").length<3))throw new Error("Select the exact released Agent source prefix.");if(g.some(u=>u==="."))throw new Error("Candidate authoring needs a focused component source directory.");let f=new Set(o.map(u=>u.path)),m=Ct.parse({...t.manifest,files:t.manifest.files.filter(u=>f.has(u.path)),metadata:{...t.manifest.metadata,profileSourceBindings:Fe(t.manifest).filter(u=>f.has(u.releasedPath)),experimentCandidate:!0}}),r=It(e.storeDir,e.candidateId),p=z.join(r,"authoring"),l=`${r}.preparing-${_r()}`,c={sourceRoot:p,base:{id:e.base.harnessRelease.id,contentHash:e.base.harnessRelease.contentHash,sourceRevision:e.base.sourceRevision},component:a,writablePaths:g,publicManifest:m,publicHashes:Object.fromEntries(o.map(u=>[u.path,u.asset.contentHash])),protectedHashes:Object.fromEntries(t.sourceFiles.filter(u=>!f.has(u.path)).map(u=>[u.path,u.asset.contentHash]))},w={...c,contentHash:v(c)};try{await q.mkdir(z.join(l,"authoring"),{recursive:!0,mode:448}),await q.writeFile(z.join(l,"partition.json"),re(w),{mode:384}),await q.writeFile(z.join(l,"authoring","harness.json"),re(m),{mode:256});for(let u of o){let k=z.join(l,"authoring",...u.path.split("/"));await q.mkdir(z.dirname(k),{recursive:!0,mode:448}),await q.writeFile(k,u.bytes,{mode:be(w,u.path)?384:256})}await q.mkdir(z.dirname(r),{recursive:!0,mode:448});try{await q.rename(l,r)}catch(u){let k=await $e(e.storeDir,e.candidateId).catch(()=>null);if(!k||k.contentHash!==w.contentHash)throw u}return w}finally{await q.rm(l,{recursive:!0,force:!0})}}h(jr,"materializeExperimentCandidateSource");async function $e(e,t){let a=JSON.parse(await q.readFile(z.join(It(e,t),"partition.json"),"utf8")),{contentHash:n,...s}=a;if(v(s)!==n||a.sourceRoot!==z.join(It(e,t),"authoring"))throw new Error("Candidate partition integrity failed.");return Ct.parse(a.publicManifest),a}h($e,"readExperimentCandidatePartition");async function qr(e){if(Buffer.byteLength(e.text)>25e4)throw new Error("Manual instructions exceed the component size limit.");let t=await $e(e.storeDir,e.candidateId);if(t.component.kind==="agent")throw new Error("Manual instruction editing requires an instruction-bearing component; use ordinary Work tools for Agent code.");let a=await _e(t.sourceRoot,t.component.path),n=await q.readFile(a);if(fe(n)!==e.expectedHash)throw new Error("Candidate instructions changed; reload before saving.");if(t.component.kind==="workflow"){let s=JSON.parse(n.toString("utf8")),o=s.workflows?.find(d=>d.id===t.component.workflowId);if(!o||o.invocation?.kind!=="instructions")throw new Error("This Workflow does not expose manual instruction editing.");o.invocation.instructions=e.text;let i=`workflows/${t.component.workflowId}/PROMPT.md`;if(t.publicHashes[i]){let d=await _e(t.sourceRoot,i),g=await q.readFile(d,"utf8"),f=/^---\r?\n[\s\S]*?\r?\n---\r?\n/.exec(g)?.[0]??"";await q.writeFile(d,f+e.text)}await q.writeFile(a,re(s))}else await q.writeFile(a,e.text);return{path:t.component.path,contentHash:fe(await q.readFile(a))}}h(qr,"writeManualCandidateInstructions");async function Nr(e){let t=await $e(e.storeDir,e.candidateId),a=await Or(e.base);if(t.base.contentHash!==e.base.harnessRelease.contentHash||t.base.sourceRevision!==e.base.sourceRevision)throw new Error("Candidate baseline release mismatch.");let n=(await Ka(t.sourceRoot)).filter(m=>m!=="harness.json"),s=await q.readFile(await _e(t.sourceRoot,"harness.json"),"utf8");if(v(JSON.parse(s))!==v(t.publicManifest))throw new Error("Candidate authoring changed the protected source manifest.");let o=new Map;for(let m of n){let r=await q.readFile(await _e(t.sourceRoot,m));if(r.byteLength>1e7)throw new Error("Candidate source exceeds the bounded component file size.");if(!be(t,m)&&fe(r)!==t.publicHashes[m])throw new Error("Candidate changed source outside its selected component.");if(t.protectedHashes[m])throw new Error("Candidate attempted to introduce protected private source.");o.set(m,r)}for(let m of Object.keys(t.publicHashes))if(!o.has(m)&&!be(t,m))throw new Error("Candidate removed source outside its selected component.");for(let m of Fe(a.manifest).filter(r=>be(t,r.releasedPath)))for(let r of Fe(a.manifest).filter(p=>p.profileRelativePath===m.profileRelativePath)){if(t.protectedHashes[r.releasedPath])throw new Error("A selected public file aliases protected source.");let p=o.get(m.releasedPath);p?o.set(r.releasedPath,p):o.delete(r.releasedPath)}let i=new Set(Fe(a.manifest).filter(m=>Fe(a.manifest).some(r=>be(t,r.releasedPath)&&r.profileRelativePath===m.profileRelativePath)).map(m=>m.releasedPath)),d=[...o.keys()].filter(m=>/^workflows\/[^/]+\/(PROMPT\.md|ACTION\.json)$/.test(m));if(t.component.kind==="workflow"&&d.length){let m=new Map([...o].filter(([l])=>l.startsWith("workflows/")&&l!=="workflows/catalog.json"&&l!=="workflows/actions.json").map(([l,c])=>[l,c.toString("utf8")])),r=JSON.parse(o.get("workflows/actions.json").toString("utf8")),p=wa({files:m,skillPaths:new Set(a.manifest.files.filter(l=>l.kind==="skill").map(l=>l.path)),actionIds:new Set(r.actions.map(l=>l.id))});o.set("workflows/catalog.json",Buffer.from(re(p.catalog)))}if(t.component.kind==="workflow"){let m=JSON.parse(Buffer.from(a.sourceFiles.find(l=>l.path===t.component.path).bytes).toString("utf8")),r=JSON.parse(o.get(t.component.path).toString("utf8")),p=h(l=>({...l,workflows:l.workflows.filter(c=>c.id!==t.component.workflowId)}),"withoutSelected");if(v(p(m))!==v(p(r))||r.workflows.filter(l=>l.id===t.component.workflowId).length!==1)throw new Error("Candidate changed another Workflow or the protected catalog structure.")}let g=a.manifest.files.filter(m=>m.visibility!=="policy"||!be(t,m.path)||o.has(m.path));for(let m of n)g.some(r=>r.path===m)||g.push({id:`candidate-asset-${v(m).slice(0,24)}`,path:m,kind:t.component.kind==="skill"?"skill_resource":"asset",parentId:t.component.kind==="skill"?a.manifest.files.find(r=>r.path===t.component.path).id:null,mediaType:Ga(m),visibility:"policy",portability:"portable"});let f=z.join(It(e.storeDir,e.candidateId),`.freeze-${_r()}`);try{await q.mkdir(f,{mode:448});let m=v([...o.entries()].map(([u,k])=>({path:u,contentHash:fe(k)}))),r=m.slice(0,40);for(let u of g){let k=a.sourceFiles.find(y=>y.path===u.path),R=be(t,u.path)||i.has(u.path)?o.get(u.path):k?.bytes;if(u.path==="dependency-lock/profile-import.json"&&R){let y=JSON.parse(Buffer.from(R).toString("utf8"));if(y.source!=="openpond.profile"||y.profileGitHead!==a.manifest.metadata.profileGitHead)throw new Error("Generated provenance differs from the baseline compiler.");R=Buffer.from(re({...y,profileGitHead:r}))}if(!R)throw new Error("Candidate source cannot replace its selected primary or protected closure.");let b=z.join(f,...u.path.split("/"));await q.mkdir(z.dirname(b),{recursive:!0,mode:448}),await q.writeFile(b,R)}await q.writeFile(z.join(f,"harness.json"),re({...a.manifest,files:g,metadata:{...a.manifest.metadata,...a.manifest.metadata.importedFrom==="openpond.profile"?{profileGitHead:r}:{}}}));let p=await se({workspaceId:e.workspaceId,sourceDir:f});for(let u of p.manifest.files.filter(k=>k.kind==="skill")){let k=Va(Buffer.from(p.sourceFiles.find(R=>R.path===u.path).bytes).toString("utf8"));if(k.messages.length>0)throw new Error(`Candidate Skill is invalid: ${k.messages.join(" ")}`)}let l=p.sourceFiles.filter(u=>u.asset.contentHash!==a.sourceFiles.find(k=>k.path===u.path)?.asset.contentHash).map(u=>({path:u.path,before:t.publicHashes[u.path]??null,after:u.asset.contentHash})),c=a.sourceFiles.filter(u=>!p.sourceFiles.some(k=>k.path===u.path)).map(u=>({path:u.path,before:u.asset.contentHash,after:null}));return{release:await Qa({storeDir:e.storeDir,workspaceId:e.workspaceId,compiled:p,createdAt:e.createdAt}),diff:[...l,...c],authoringHash:m,profileSourceRevision:r,executableIdentity:Ye(p),partitionHash:t.contentHash}}finally{await q.rm(f,{recursive:!0,force:!0})}}h(Nr,"freezeExperimentCandidateSource");function Fe(e){let t=e.metadata.profileSourceBindings;if(t===void 0)return[];if(!Array.isArray(t))throw new Error("Profile source binding metadata is invalid.");return t.map(a=>{if(!a||typeof a!="object"||typeof a.releasedPath!="string"||typeof a.profileRelativePath!="string")throw new Error("Profile source binding is invalid.");return{releasedPath:yt(a.releasedPath),profileRelativePath:yt(a.profileRelativePath)}})}h(Fe,"bindings");function Ui(e){let t=Br(e.storeDir),a=h(()=>new Date().toISOString(),"now");async function n(r,p,l=!1){let c=await e.store.getHarnessWorkspace(p.ownerWorkspaceId);if(!c)throw new Error("Improvement owner is unavailable.");if(l&&await e.authorizeOwner(r,c,p.profileRef,p.projectId,{requireCurrentBase:!0}),c.revision!==p.ownerRevision||c.sourceRevision!==p.ownerSourceRevision||v(c.currentChannel.release)!==v(p.baseRelease))throw new Error("The live owner changed; reconcile and qualify against its new base before adoption.");return c}h(n,"owner");async function s(r){let p=await e.store.getHarnessReleaseRecord(r.baseRelease.contentHash);if(!p||p.harnessRelease.id!==r.baseRelease.id||p.workspaceId!==r.ownerWorkspaceId)throw new Error("Improvement baseline release is unavailable.");return p}h(s,"base");async function o(r,p,l,c,w){if(!e.projectHumanComparison||!p.projectId)throw new Error("The current accepted Human comparison owner is unavailable.");let u=ya(await e.projectHumanComparison(r,p,l,c,w),{baseline:c,candidate:w});if(u.teamId!==r.teamId||u.projectId!==p.projectId||v(u.selections)!==v(l)||u.original.baseline.id!==c.manifest.id||u.original.baseline.manifestHash!==c.manifest.contentHash||u.original.baseline.resultHash!==c.result.contentHash||u.original.candidate.id!==w.manifest.id||u.original.candidate.manifestHash!==w.manifest.contentHash||u.original.candidate.resultHash!==w.result.contentHash)throw new Error("The accepted Human projection changed its original receipts.");return u}h(o,"humanProjection");async function i(r,p){let l=p.comparison?.humanProjection;if(!l)return;let c=Ie(await e.loadEvidence(r,{id:l.original.baseline.id,contentHash:l.original.baseline.resultHash})),w=Ie(await e.loadEvidence(r,{id:l.original.candidate.id,contentHash:l.original.candidate.resultHash}));if((await o(r,p,l.selections,c,w)).contentHash!==l.contentHash)throw new Error("The accepted Human comparison changed; requalify before adoption.")}h(i,"reauthorizeHuman");async function d(r,p){let l=await t.read(p,r.actorId,r.teamId),c=await e.store.getHarnessWorkspace(l.ownerWorkspaceId);if(!c)throw new Error("Improvement owner is unavailable.");return await e.authorizeOwner(r,c,l.profileRef,l.projectId,{requireCurrentBase:!1}),await e.loadEvidence(r,{id:l.evidence.manifest.id,contentHash:l.evidence.result.contentHash}),await i(r,l),l}h(d,"read");async function g(r,p,l,c){let w=await d(r,p);return await n(r,w,!0),t.mutate(p,r.actorId,r.teamId,l,c)}h(g,"mutateOwned");let f=h((r,p)=>{let{contentHash:l,...c}=r;return Je({...c,...p,revision:r.revision+1,updatedAt:a()})},"transition");return{read:d,list:h(async(r,p=30,l)=>{let c=await t.list(r.actorId,r.teamId,p,l),w=[];for(let u of c)try{w.push(await d(r,u.id))}catch{}return{states:w,nextCursor:c.length===p?c.at(-1).id:null}},"list"),async start(r,p){let l=Da.parse(p),c=v(l),w=`improve-${v([r.actorId,r.teamId,l.operationId]).slice(0,40)}`,u=await t.find(w,r.actorId,r.teamId);if(u){if(u.requestHash!==c)throw new Error("The original creation operation was reused with another request.");return d(r,u.id)}if(l.mode==="manual"&&l.component.kind==="agent")throw new Error("Manual Improve edits instruction-bearing components; use ordinary Work authoring for Agent code.");let k=await e.store.getHarnessWorkspace(l.ownerWorkspaceId);if(!k||k.location!=="local")throw new Error("Select an admitted local Profile owner.");if(await e.authorizeOwner(r,k,l.profileRef,l.projectId,{requireCurrentBase:!0}),k.revision!==l.expectedOwnerRevision||v(k.currentChannel.release)!==v(l.baseRelease))throw new Error("Improvement base owner changed.");let R=Ie(await e.loadEvidence(r,l.evidence));if(R.result.contentHash!==l.evidence.contentHash||R.manifest.id!==l.evidence.id)throw new Error("Improvement evidence identity changed.");if((await e.store.getHarnessReleaseRecord(l.baseRelease.contentHash))?.harnessRelease.metadata.profile?.sourceRevision!==l.profileSourceRevision)throw new Error("Improvement source revision differs from its exact owner Profile release.");let H=a(),x=await t.create(Je({schemaVersion:"openpond.experimentImprovement.v1",id:w,operationId:l.operationId,requestHash:c,...r,projectId:l.projectId,ownerWorkspaceId:k.id,ownerRevision:k.revision,ownerSourceRevision:k.sourceRevision,baseRelease:l.baseRelease,profileRef:l.profileRef,baseProfileSourceRevision:l.profileSourceRevision,evidence:R,component:l.component,mode:l.mode,status:"draft",revision:1,limits:l.limits,authoringSteps:0,work:null,frozen:null,test:null,comparison:null,adoptionIntent:null,acceptance:null,rollback:null,error:null,createdAt:H,updatedAt:H}));return await jr({storeDir:e.storeDir,candidateId:w,base:await s(x),component:x.component}),await n(r,x),x},async edit(r,p,l,c,w){return g(r,p,l,async u=>{if(await n(r,u),["authoring_queued","authoring","testing","adopting","accepted","rolled_back","declined","cancelled"].includes(u.status))throw new Error("Stop active authoring/evaluation before editing this candidate.");return await qr({storeDir:e.storeDir,candidateId:p,expectedHash:c,text:w}),f(u,{status:"draft",frozen:null,test:null,comparison:null,error:null})})},async setMode(r,p,l,c){return g(r,p,l,async w=>{if(w.status!=="draft")throw new Error("Change mode only in a settled authoring draft.");if(c==="manual"&&w.component.kind==="agent")throw new Error("Use ordinary Work for Agent source authoring.");return w.mode===c?w:f(w,{mode:c})})},async queueAuthoringTurn(r,p,l,c,w,u){return g(r,p,l,async k=>{if(k.mode!=="llm_assisted"||k.status!=="draft")throw new Error("Queue only a settled LLM authoring draft.");return f(k,{status:"authoring_queued",work:{sessionId:c,turnId:w,revision:k.revision+2,startedAt:u},frozen:null,test:null,comparison:null,error:null})})},async cancelQueuedAuthoring(r,p,l){return g(r,p,l,async c=>{if(c.status!=="authoring_queued")throw new Error("Cancel only a queued candidate turn after actual host cancellation.");return f(c,{status:"cancelled",error:"The owner cancelled queued candidate Work before provider dispatch."})})},async bindAuthoringTurn(r,p,l,c,w){return g(r,p,l,async u=>{if(await n(r,u),u.mode!=="llm_assisted"||u.status!=="draft"&&u.status!=="authoring_queued"||u.status==="authoring_queued"&&(u.work?.sessionId!==c||u.work.turnId!==w))throw new Error("An explicit LLM authoring action is required.");let k=await e.store.getSession(c),R=await e.store.getTurn(w);if(!k||k.experience!=="work"||!R||R.sessionId!==c||R.status!=="in_progress")throw new Error("Candidate authoring requires its actual ordinary Work session and turn.");let b={candidateId:p,authoringRevision:l+1,actorId:r.actorId,teamId:r.teamId};if(v(R.metadata.refinementCandidate)!==v(b)||R.metadata.source!=="experiment-improvement")throw new Error("The Work turn has no trusted candidate admission marker.");return f(u,{status:"authoring",work:{sessionId:c,turnId:w,revision:l+1,startedAt:a()},authoringSteps:0,frozen:null,test:null,comparison:null,error:null})})},async authoringStep(r,p,l){return g(r,p,l,async c=>{if(await n(r,c),c.status!=="authoring"||!c.work)throw new Error("Candidate Work is not authoring.");if(c.authoringSteps>=c.limits.maximumAuthoringSteps||Date.now()-Date.parse(c.work.startedAt)>c.limits.maximumDurationMs)throw new Error("Candidate authoring reached its admitted step or time limit.");return f(c,{authoringSteps:c.authoringSteps+1})})},async beginTestDispatch(r,p,l,c){return g(r,p,l,async w=>{if(w.status!=="testing"||!w.test||w.test.cancelRequestedAt||w.test.dispatching)throw new Error("The test is cancelled or another exact dispatch owns admission.");return f(w,{test:{...w.test,dispatching:c}})})},async endTestDispatch(r,p,l,c){return g(r,p,l,async w=>{if(w.status!=="testing"||!w.test||w.test.dispatching?.token!==c)throw new Error("The test dispatch ownership changed.");return f(w,{test:{...w.test,dispatching:null}})})},async cancelTesting(r,p,l){return g(r,p,l,async c=>{if(c.status!=="testing"||!c.test)throw new Error("Cancel only an admitted candidate test.");return c.test.cancelRequestedAt?c:f(c,{test:{...c.test,cancelRequestedAt:a()}})})},async finishAuthoring(r,p,l){return g(r,p,l,async c=>{if(await n(r,c),c.status!=="authoring"||!c.work)throw new Error("Candidate is not authoring.");let w=await e.store.getTurn(c.work.turnId);if(!w||w.status==="in_progress")throw new Error("Wait for the admitted Work turn to finish or cancel it.");return f(c,{status:w.status==="completed"?"draft":"failed",error:w.status==="completed"?null:"Candidate Work did not complete successfully."})})},async freeze(r,p,l){return g(r,p,l,async c=>{if(await n(r,c),c.status!=="draft")throw new Error("Freeze a settled draft before testing.");let w=`candidate-${v([p,c.revision]).slice(0,40)}`,u=await Nr({storeDir:e.storeDir,candidateId:p,base:await s(c),workspaceId:w,createdAt:a()});e.freezeProfile&&(u=await e.freezeProfile(r,c,u));let k={release:{id:u.release.harnessRelease.id,contentHash:u.release.harnessRelease.contentHash},sourceRevision:u.release.sourceRevision,profileSourceRevision:u.profileSourceRevision,authoringHash:u.authoringHash,partitionHash:u.partitionHash,executableSourceHash:u.executableIdentity.contentHash,protectedClosureHash:u.executableIdentity.protectedClosureHash,candidateRevision:c.revision+1,diff:u.diff},R={schemaVersion:"openpond.experimentCandidateOrigin.v1",candidateId:p,actorId:r.actorId,teamId:r.teamId,baseWorkspaceId:c.ownerWorkspaceId,baseProfileRef:c.profileRef,baseProfileSourceRevision:c.baseProfileSourceRevision,baseRelease:c.baseRelease,frozenRevision:k.candidateRevision,frozenHash:v(k)},b=await e.store.getHarnessWorkspace(w);if(b){if(v(b.currentChannel.release)!==v(k.release)||v(b.metadata.experimentCandidateOrigin)!==v(R))throw new Error("Candidate workspace identity conflict.")}else{let y=Xe(e.storeDir,w);if(await ze.mkdir(y.root,{recursive:!0,mode:448}),await ze.stat(y.source).catch(()=>null)){if((await se({workspaceId:typeof u.release.agentSnapshot.metadata.workspaceId=="string"?u.release.agentSnapshot.metadata.workspaceId:w,sourceDir:y.source})).harnessRelease.contentHash!==k.release.contentHash)throw new Error("Interrupted candidate workspace source differs from its frozen release.")}else{let x=`${y.source}.install-${v(k).slice(0,20)}`;await ze.rm(x,{recursive:!0,force:!0}),await ze.cp(Dr.join(u.release.bundlePath,"source"),x,{recursive:!0,errorOnExist:!0,force:!1}),await ze.rename(x,y.source)}await e.store.createHarnessWorkspaceWithRelease({workspace:Le.parse({schemaVersion:"openpond.harnessWorkspace.v1",id:w,ownerScope:{kind:"personal",id:r.actorId},name:`Candidate ${p}`,location:"local",sourceRevision:u.release.sourceRevision,revision:0,dirty:!1,currentChannel:{name:"candidate",release:k.release,revision:1},createdAt:a(),updatedAt:a(),metadata:{sourceLayout:"openpond.harnessSourceManifest.v1",selectionEligible:!1,experimentCandidateOrigin:R}}),release:u.release})}return await n(r,c,!0),f(c,{status:"frozen",frozen:k,test:null,comparison:null,error:null})})},async beginTesting(r,p,l,c){return g(r,p,l,async w=>{await n(r,w);let u=Na.parse(c),k=u.externalDatasetBinding;if(w.status==="testing"&&w.test){if(v(u)!==w.test.recipeHash)throw new Error("The admitted test recipe cannot change during execution.");return w}if(w.status!=="frozen"||!w.frozen)throw new Error("Test a frozen candidate revision.");let R=w.evidence.manifest;if(k.profileId!==w.profileRef.profileId||v(k.dataset)!==v(R.dataset)||!u.scoring&&v(k.evaluators)!==v(R.evaluators)||u.scoring&&u.scoring.sourcePass.id!==R.lineage?.scoringPassId)throw new Error("The qualification changed the original Dataset or evaluator selection.");let b=k.recordedOrigin?k.recordedOrigin.members.map(I=>({caseId:I.caseId,seed:I.seed,fixtureId:I.fixtureId})):k.population.map(I=>({caseId:I.taskId,seed:I.seed,fixtureId:I.fixtureId}));if(v(b)!==v(R.population)||k.recordedOrigin&&(k.recordedOrigin.originalEvidence.id!==R.id||k.recordedOrigin.originalEvidence.contentHash!==w.evidence.result.contentHash))throw new Error("The qualification changed the exact original ordered population or recorded cutoff.");if(k.protectedProfileClosureHash!==w.frozen.protectedClosureHash)throw new Error("The test recipe changed the protected Profile evaluator closure.");let y=k.target;if(w.component.kind==="workflow"?y.kind!=="workflow"||y.workflowId!==w.component.workflowId:w.component.kind==="skill"?y.kind!=="skill"||y.skillPath!==w.component.path:w.component.kind==="instruction"?y.kind!=="profile":y.kind!=="agent_action")throw new Error("Test the selected authored Profile component.");if(w.component.kind==="agent"){let I=await s(w),P=await se({workspaceId:typeof I.agentSnapshot.metadata.workspaceId=="string"?I.agentSnapshot.metadata.workspaceId:I.workspaceId,sourceDir:Dr.join(I.bundlePath,"source")}),T=P.sourceFiles.find(D=>D.path==="workflows/actions.json")?.bytes,S=T&&ga.parse(JSON.parse(Buffer.from(T).toString("utf8"))).actions.find(D=>y.kind==="agent_action"&&D.id===y.actionId);if(!S||!w.component.path.startsWith(`agents/${S.agentId}/`))throw new Error("The tested Agent action belongs to a different authored component.")}let H=v(u),x=h(I=>`improve-${I}-${v([p,w.frozen,H]).slice(0,40)}`,"operation");return f(w,{status:"testing",test:{recipe:u,recipeHash:H,baselineOperationId:x("base"),candidateOperationId:x("candidate"),cancelRequestedAt:null,dispatching:null}})})},async failTesting(r,p,l,c,w=!1){return g(r,p,l,async u=>{if(u.status!=="testing")throw new Error("The candidate has no active test.");return f(u,{status:w?"cancelled":"failed",error:c.slice(0,2e3)})})},async completeComparison(r,p,l,c,w,u){let k=await d(r,p),R=k.comparison?.humanProjection;if(u&&k.status==="ready_for_review"&&k.revision===l+1&&R){if(v(R.selections)!==v(u)||R.original.baseline.id!==c.manifest.id||R.original.baseline.resultHash!==c.result.contentHash||R.original.candidate.id!==w.manifest.id||R.original.candidate.resultHash!==w.result.contentHash)throw new Error("The original Human comparison command changed.");let b=await d(r,p);if(b.contentHash!==k.contentHash)throw new Error("The retained comparison changed during recovery.");return b}return g(r,p,l,async b=>{if(await n(r,b),b.status!=="testing"||!b.frozen||b.test?.cancelRequestedAt||b.test?.dispatching)throw new Error("Candidate has no settled uncancelled admitted test.");let y=Ie(await e.loadEvidence(r,{id:c.manifest.id,contentHash:c.result.contentHash})),H=Ie(await e.loadEvidence(r,{id:w.manifest.id,contentHash:w.result.contentHash}));if(!b.test)throw new Error("The comparison has no retained test operations.");for(let[A,U]of[["baseline",y],["candidate",H]]){let O=b.test.recipe.scoring?`improve-grade-${v([b.id,b.test.recipeHash,A]).slice(0,40)}`:A==="baseline"?b.test.baselineOperationId:b.test.candidateOperationId;if(U.manifest.operationId!==O)throw new Error("Compare only the exact originally admitted baseline and candidate operations.")}let x=u?await o(r,b,u,y,H):void 0,I=x?.baseline??y,P=x?.candidate??H,T=h(A=>A.target.kind==="harness"||A.target.kind==="agent"?A.target.source.harnessRelease:null,"releaseOf");if(v(T(I.manifest))!==v(b.baseRelease)||v(T(P.manifest))!==v(b.frozen.release))throw new Error("The comparison did not execute the exact baseline and frozen candidate releases.");if(I.result.status!=="completed"||P.result.status!=="completed"||[...I.result.cases,...P.result.cases].some(A=>A.status!=="completed"||A.feedback.some(U=>U.status!=="scored")))throw new Error("Complete all admitted case execution and grading before adopting this candidate.");if(!ka(I,P).comparable)throw new Error("Candidate and baseline require a matched population, environment and graders.");if(!b.test)throw new Error("The comparison has no admitted immutable rerun recipe.");let S=b.test.recipe.externalDatasetBinding,D=S.population.map(A=>({caseId:A.taskId,seed:A.seed,fixtureId:A.fixtureId}));if(v(I.manifest.dataset)!==v(S.dataset)||v(I.manifest.population)!==v(D)||v(y.manifest.evaluators)!==v(b.test.recipe.scoring?b.evidence.manifest.evaluators:S.evaluators))throw new Error("The qualification changed its admitted rerun population or evaluator selection.");for(let A of[I,P])if(A.manifest.target.kind!=="harness"&&A.manifest.target.kind!=="agent"||!("profileId"in A.manifest.target.source)||v(A.manifest.target.source.externalDatasetBinding)!==v(S))throw new Error("The executed Profile omitted its actual immutable external Dataset recipe.");return f(b,{status:"ready_for_review",comparison:{baseline:I,candidate:P,frozenHash:v(b.frozen),...x?{humanProjection:x}:{}}})})},async beginAdoption(r,p,l){let c=await d(r,p);return c.status==="adopting"||c.status==="accepted"?c:g(r,p,l,async w=>{if(await n(r,w),w.status!=="ready_for_review"||!w.frozen||!w.comparison)throw new Error("Use candidate requires its complete exact tested revision.");return f(w,{status:"adopting",adoptionIntent:{operationId:`adopt-${v([p,w.frozen]).slice(0,40)}`,expectedOwnerRevision:w.ownerRevision,frozenHash:v(w.frozen)}})})},async useCandidate(r,p,l){let c=await d(r,p);if(c.status==="accepted")return c;if(c.status!=="adopting"&&(c=await this.beginAdoption(r,p,l)),!c.frozen||!c.adoptionIntent)throw new Error("Adoption has no durable tested intent.");let w=await e.store.getHarnessReleaseRecord(c.frozen.release.contentHash);if(!w||w.harnessRelease.id!==c.frozen.release.id)throw new Error("Tested candidate release is unavailable.");await i(r,c);let u=await e.readAdoption(r,c,c.adoptionIntent.operationId)??await e.adopt(r,c,w,c.adoptionIntent.operationId);if(v(u.testedCandidateRelease)!==v(c.frozen.release)||u.executableSourceHash!==c.frozen.executableSourceHash||u.protectedClosureHash!==c.frozen.protectedClosureHash)throw new Error("Owner adoption differs from tested executable source or protected closure.");return await i(r,c),t.mutate(p,r.actorId,r.teamId,c.revision,async k=>{if(k.status!=="adopting"||v(k.adoptionIntent)!==v(c.adoptionIntent))throw new Error("Adoption intent changed during owner readback.");return f(k,{status:"accepted",acceptance:u.receipt})})},async rollback(r,p,l){let c=await d(r,p);if(c.status==="rolled_back")return c;if(c.status!=="accepted"||c.revision!==l||!c.acceptance)throw new Error("Rollback requires the current accepted Profile adoption.");let w=await e.rollback(r,c,`rollback-${c.acceptance.contentHash.slice(0,40)}`);if(v(w.testedCandidateRelease)!==v(c.baseRelease))throw new Error("Profile rollback did not restore the original executable release.");return t.mutate(p,r.actorId,r.teamId,l,async u=>f(u,{status:"rolled_back",rollback:w.receipt}))},async discard(r,p,l,c=!1){return g(r,p,l,async w=>{if(await n(r,w),["authoring_queued","authoring","testing"].includes(w.status))throw new Error("Cancel the actual Work/evaluation execution before discarding the candidate.");if(w.status==="accepted"||w.status==="rolled_back"||w.status==="adopting")throw new Error("Recover adoption or use the owner rollback operation for an adopted release.");return f(w,{status:c?"cancelled":"declined"})})},async partition(r,p){let l=await d(r,p);return await n(r,l),$e(e.storeDir,p)},states:t,close:h(()=>t.close(),"close")}}h(Ui,"createLocalExperimentImprovementService");import K from"node:path";import{promises as Kt}from"node:fs";import{tmpdir as qn}from"node:os";var St=h(e=>`'${e.replaceAll("'","'\\''")}'`,"quote");function td(e){async function t(o,i){let d=await e.store.getSession(o),g=await e.store.getTurn(i);if(!d||!g||g.sessionId!==o)throw new Error("Candidate tool has no actual Work turn authority.");let f=g.metadata?.refinementCandidate;if(g.metadata?.source!=="experiment-improvement"||!f||typeof f!="object"||!("candidateId"in f)||typeof f.candidateId!="string"){if(d.metadata?.source==="experiment-improvement"||d.metadata?.refinementCandidate!==void 0)throw new Error("This candidate Session requires a fresh private stored-turn admission before any continuation.");return null}let m={actorId:await e.actorId(),teamId:await e.teamId()},r=await e.improvements.read(m,f.candidateId);if(r.status!=="authoring"||!r.work||r.work.sessionId!==o||r.work.turnId!==i||g.status!=="in_progress"||d.experience!=="work"||v(f)!==v({candidateId:r.id,authoringRevision:r.work.revision,...m})||v(d.currentProfile)!==v(r.profileRef))throw new Error("Candidate tool binding changed or its Work turn has ended.");let p=await e.improvements.partition(m,r.id);return{actor:m,state:r,partition:p}}h(t,"authority");let a=e.agentRuntime,n=Ya({bwrapPath:e.bwrapPath,authorize:h(async o=>{let i=await t(o.sessionId,o.turnId);if(!i||i.state.id!==o.candidateId||i.state.revision!==o.expectedRevision)throw new Error("Candidate command changed before filesystem admission.");return{candidateId:i.state.id,candidateRevision:i.state.revision,ownerId:i.actor.actorId,sessionId:o.sessionId,turnId:o.turnId,sourceRoot:i.partition.sourceRoot,writablePaths:i.partition.writablePaths,...a?{runtimeMounts:[{source:a.source,destination:"/runtime/agent"},{source:K.join(a.source,"node_modules"),destination:"/workspace/node_modules"}]}:{}}},"authorize")});async function s(o,i,d,g={}){e.operationSignal?.throwIfAborted();let f=await t(o,i);if(!f)throw new Error("This command is not admitted candidate Work.");let m=await e.improvements.authoringStep(f.actor,f.state.id,f.state.revision),r=g.cwd?.startsWith("/workspace/work/")?K.resolve(f.partition.sourceRoot,g.cwd.slice(16)):g.cwd==="/workspace/work"?f.partition.sourceRoot:g.cwd;return n({candidateId:m.id,sessionId:o,turnId:i,expectedRevision:m.revision,command:d,...g,signal:g.signal&&e.operationSignal?AbortSignal.any([g.signal,e.operationSignal]):g.signal??e.operationSignal,cwd:r})}return h(s,"command"),{async candidateAuthoringForTurn(o,i){return await t(o,i)!==null},async authorizeCandidateTool(o){if(!await t(o.sessionId,o.turnId))return!1;if(!nr.has(o.name))throw new Error("This tool has no authority over isolated candidate Work.");return!0},async resolveCandidateProfile(o){let i=await t(o.session.id,o.turnId);if(!i)throw new Error("Candidate Profile requires a trusted Work admission.");let d=await e.loadProfile(i.state.profileRef),g=i.partition.publicManifest,f={...d,repoPath:i.partition.sourceRoot,sourcePath:i.partition.sourceRoot,skills:d.skills.filter(m=>g.files.some(r=>r.kind==="skill"&&r.path===`skills/${Oe(m.name)}/SKILL.md`)).map(m=>({...m,path:`skills/${Oe(m.name)}/SKILL.md`,sourcePath:i.partition.sourceRoot})),agents:d.agents.filter(m=>g.files.some(r=>r.kind==="agent"&&r.path.startsWith(`agents/${Oe(m.id)}/`))).map(m=>({...m,path:g.files.find(r=>r.kind==="agent"&&r.path.startsWith(`agents/${Oe(m.id)}/`)).path})),evals:[],git:d.git?{...d.git,dirty:!0,files:[]}:null};return await t(o.session.id,o.turnId),f},async executeCandidateCommand(o,i){let d=await s(o.session.id,o.turnId,i.command,{cwd:i.cwd??void 0,timeoutMs:(i.timeoutSeconds??120)*1e3,signal:i.signal});return{ok:d.code===0&&!d.timedOut,command:i.command,cwd:i.cwd??o.session.cwd,exitCode:d.code,stdout:d.stdout,stderr:d.stderr,timedOut:d.timedOut,timeoutSeconds:Math.min(i.timeoutSeconds??120,180),truncated:d.stdoutTruncated||d.stderrTruncated,blockedReason:null}},async executeCandidateImage(o){let i=await t(o.session.id,o.turnId);if(!i)throw new Error("Image inspection is not admitted candidate Work.");let d=o.args.path;if(typeof d!="string"||!d.trim())throw new Error("Select a candidate image path.");let g=d.startsWith("/workspace/work/")?K.resolve(i.partition.sourceRoot,d.slice(16)):K.resolve(i.partition.sourceRoot,d),f=K.relative(i.partition.sourceRoot,g);if(!f||f.startsWith("..")||K.isAbsolute(f))throw new Error("The image path escapes this candidate source.");let m=`/workspace/work/${f.split(K.sep).join("/")}`,r=h(async(y,H,x)=>{let I=await s(o.session.id,o.turnId,[y,...H].map(St).join(" "),{signal:x,maxOutputBytes:75e4,timeoutMs:3e4});if(I.code!==0||I.timedOut||I.stdoutTruncated)throw new Error("The confined image decoder could not inspect this candidate image.");return I.stdout},"run"),p=await or(m,o.signal,r),l={PNG:"png",JPEG:"jpg",GIF:"gif",WEBP:"webp",AVIF:"avif"}[p.format];if(!l)throw new Error("Select a PNG, JPEG, GIF, WebP or AVIF image.");let c=await r("/runtime/node",["-e","const f=require('fs'),p=process.argv[1],b=f.readFileSync(p);if(b.length>500000)throw Error('Create a contact sheet under 500000 bytes for candidate inspection');process.stdout.write(b.toString('base64'))",m],o.signal);if(!/^[A-Za-z0-9+/]*={0,2}$/.test(c)||c.length>67e4)throw new Error("Candidate image bytes exceed the inspection boundary.");let w=await Kt.mkdtemp(K.join(qn(),"candidate-image-"));await Kt.chmod(w,448);let u=K.join(w,`preview.${l}`),k=Buffer.from(c,"base64");await Kt.writeFile(u,k,{mode:256}),await t(o.session.id,o.turnId);let R=`image/${l==="jpg"?"jpeg":l}`,b={path:m,contentType:R,sizeBytes:k.length,...p};return{toolCallId:o.callId,name:"view_image",ok:!0,contentText:JSON.stringify({ok:!0,action:"view_image",data:b}),data:{...b,openpondImagePreviewPath:u}}},async executeCandidateAgentCommand(o){let i=await t(o.context.session.id,o.context.turnId);if(!i)throw new Error("Agent command is not admitted candidate Work.");if(a??=await e.loadAgentRuntime?.(),!a)throw new Error("The server has not staged a dependency-complete confined Agent runtime; build and install that public runtime before authoring Agent components.");let d=a.cliRelativePath;if(K.isAbsolute(d)||d.split(/[\\/]/).some(r=>!r||r===".."||r==="."))throw new Error("Agent runtime CLI binding is invalid.");let g=K.relative(i.partition.sourceRoot,o.cwd);if(g.startsWith("..")||K.isAbsolute(g))throw new Error("Agent command source escapes its candidate.");let f=`/workspace/work${g?`/${g.split(K.sep).join("/")}`:""}`,m=o.command==="run"?[o.command,...o.args??[]]:[o.command,"--cwd",f,...o.args??[]];return s(o.context.session.id,o.context.turnId,`/runtime/node ${St(`/runtime/agent/${d}`)} ${m.map(St).join(" ")}`,{cwd:o.cwd,timeoutMs:o.timeoutMs,maxOutputBytes:o.maxOutputBytes,signal:o.context.signal})},executeCandidateWorkspaceTool:h((async o=>{if(!o.turnId){if(o.session.metadata?.source==="experiment-improvement"||o.session.metadata?.refinementCandidate!==void 0)throw new Error("Candidate files require a trusted actual turn.");return null}let i=await t(o.session.id,o.turnId);if(!i)return null;if(!new Set(["list_files","read_files","search_files","workspace_status","preview_write_file","preview_write_files","preview_edit_file","preview_delete_file","write_file","write_files","edit_file","delete_file"]).has(o.request.action))throw new Error("Use the candidate file tools or confined exec_command/Agent tools; this operation has no authority over the live Profile.");let g=await s(o.session.id,o.turnId,`/runtime/node -e ${St(sr)}`,{stdin:JSON.stringify({action:o.request.action,args:o.request.args,writablePaths:i.partition.writablePaths}),maxOutputBytes:8e5});if(g.code!==0||g.timedOut||g.stdoutTruncated)throw new Error(g.stderr.slice(0,2e3)||"Candidate source operation did not complete.");return{ok:!0,action:o.request.action,output:"Completed in the isolated candidate source.",data:JSON.parse(g.stdout)}}),"executeCandidateWorkspaceTool")}}h(td,"createExperimentCandidateTools");import{createHash as $r}from"node:crypto";import{mkdir as Xt,writeFile as Nn,rm as Dn,rename as Fn,readdir as $n,readFile as zn,lstat as Un}from"node:fs/promises";import{randomUUID as Ln}from"node:crypto";import Ht from"node:path";var Vn=M.object({path:M.string(),type:M.enum(["file","directory"]),sizeBytes:M.number().nullable(),oid:M.string().nullable(),mode:M.string().nullable()}),Fr=M.object({commitSha:M.string(),entries:M.array(Vn).max(5e3),truncated:M.boolean(),selectedFile:M.object({path:M.string(),sizeBytes:M.number(),contents:M.string().nullable(),encoding:M.literal("utf8").nullable(),isBinary:M.boolean(),truncated:M.boolean()}).nullable()});async function ld(e){let t=Fr.parse(await e.read({branch:e.revision,path:"__openpond_profile_origin_tree_only__",maxEntries:5e3}));if(t.commitSha!==e.revision||t.truncated)throw new Error("Profile origin tree is incomplete or changed its commit.");let a=t.entries.filter(i=>i.type==="file");if(!a.length||a.length>5e3)throw new Error("Profile origin has no complete bounded source tree.");let n=0,s=new Set;for(let i of a){if(i.path.startsWith("/")||i.path.includes("\\")||i.path.split("/").some(d=>!d||d==="."||d==="..")||s.has(i.path)||!["100644","100755"].includes(i.mode??"")||!Number.isSafeInteger(i.sizeBytes)||i.sizeBytes<0||i.sizeBytes>2e6||!i.oid?.match(/^[a-f0-9]{40}$/))throw new Error("Profile origin contains unsupported paths, modes, or bytes.");if(i.path.split("/").some(d=>d===".git"||d.startsWith(".env")||/\.(sql|sqlite|sqlite3|db)$/i.test(d)))throw new Error("Profile origin contains an unsupported secret or database source path.");s.add(i.path),n+=i.sizeBytes}if(n>32e6)throw new Error("Profile origin exceeds the bounded source size.");let o=`${e.root}.partial-${Ln()}`;await Xt(o,{recursive:!0,mode:448});try{let i=[];for(let d of a){let g=Fr.parse(await e.read({branch:e.revision,path:d.path,maxEntries:5e3})),f=g.selectedFile;if(g.commitSha!==e.revision||!f||f.path!==d.path||f.isBinary||f.truncated||f.encoding!=="utf8"||f.contents===null)throw new Error("Profile origin cannot supply complete exact source bytes.");let m=Buffer.from(f.contents,"utf8"),r=$r("sha1").update(`blob ${m.length}\0`).update(m).digest("hex");if(m.length!==d.sizeBytes||f.sizeBytes!==d.sizeBytes||r!==d.oid)throw new Error("Profile origin file differs from its committed Git blob.");let p=Ht.join(o,...d.path.split("/"));await Xt(Ht.dirname(p),{recursive:!0,mode:448}),await Nn(p,m,{mode:384,flag:"wx"}),i.push({path:d.path,oid:r,sizeBytes:m.length})}return await Xt(Ht.dirname(e.root),{recursive:!0,mode:448}),await Fn(o,e.root),i.sort((d,g)=>d.path<g.path?-1:d.path>g.path?1:0)}catch(i){throw await Dn(o,{recursive:!0,force:!0}),i}}h(ld,"materializeProfileOrigin");async function Qt(e){let t=[],a=0;async function n(s,o){for(let i of await $n(s,{withFileTypes:!0})){let d=o?`${o}/${i.name}`:i.name,g=Ht.join(s,i.name),f=await Un(g);if(f.isSymbolicLink())throw new Error("Profile origin cache contains a symlink.");if(f.isDirectory()){await n(g,d);continue}if(!f.isFile()||f.size>2e6||t.length>=5e3||(a+=f.size)>32e6)throw new Error("Profile origin cache exceeds its bounds.");let m=await zn(g);t.push({path:d,oid:$r("sha1").update(`blob ${m.length}\0`).update(m).digest("hex"),sizeBytes:m.length})}}return h(n,"visit"),await n(e,""),v(t.sort((s,o)=>s.path<o.path?-1:s.path>o.path?1:0))}h(Qt,"profileOriginFilesHash");import{promises as Et}from"node:fs";import Yt from"node:path";async function wd(e,t){let a=await e.store.getHarnessWorkspace(e.state.ownerWorkspaceId),n=a?.metadata.profileExperimentOrigin,s=Yt.join(e.storeDir,"library","profile-experiment-origins",e.state.ownerWorkspaceId);if(e.state.profileRef.source!=="openpond_git"||a?.location!=="local"||a.metadata.selectionEligible!==!1||n?.actorId!==e.state.actorId||n.teamId!==e.state.teamId||n.sourceRevision!==e.state.baseProfileSourceRevision||n.sourcePath!==s||typeof n.sourceFilesHash!="string"||await Qt(s)!==n.sourceFilesHash)throw new Error("The exact authorized committed Profile snapshot is unavailable.");let o=Yt.join(e.storeDir,"library","harnesses","profile-candidate-staging");await Et.mkdir(o,{recursive:!0,mode:448});let i=await Et.mkdtemp(Yt.join(o,"remote-source-"));try{await Et.cp(s,i,{recursive:!0,errorOnExist:!0,force:!1});let d=await Ja(i,e.state.profileRef.profileId);if(d.error||!d.sourcePath)throw new Error("The actual committed Profile archive cannot compile.");let g=await t({root:i,profile:d,source:d.sourcePath});if(await Qt(s)!==n.sourceFilesHash)throw new Error("The committed Profile cache changed during candidate compilation.");return g}finally{await Et.rm(i,{recursive:!0,force:!0})}}h(wd,"withRemoteProfileCandidateSnapshot");async function Sd(e,t,a="owner_attested_local"){if(t.profileRef.source!=="openpond_git"||!t.frozen||!t.comparison||!t.test||!t.adoptionIntent||t.status!=="adopting")throw new Error("Export only the exact owner-approved qualified hosted Profile candidate.");let n=await e.getHarnessReleaseRecord(t.frozen.release.contentHash);if(!n||n.harnessRelease.id!==t.frozen.release.id)throw new Error("The tested native candidate is unavailable.");let s=n.agentSnapshot.metadata.workspaceId;if(typeof s!="string")throw new Error("The candidate compiler identity is absent.");let o=await se({workspaceId:s,sourceDir:`${n.bundlePath}/source`}),i=Ye(o);if(o.harnessRelease.contentHash!==n.harnessRelease.contentHash||i.contentHash!==t.frozen.executableSourceHash||i.protectedClosureHash!==t.frozen.protectedClosureHash||t.comparison.frozenHash!==v(t.frozen))throw new Error("The actual tested source differs from its qualified immutable candidate.");let d=o.manifest.metadata.profileRepositoryId;if(d!==t.profileRef.repositoryId)throw new Error("The candidate source compiler belongs to another Profile owner.");let g=Buffer.from(re(o.manifest)),f=ha(fa({agentSnapshot:o.agentSnapshot,harnessRelease:o.harnessRelease,files:new Map(o.sourceFiles.map(p=>[p.path,p.bytes]))})),m={schemaVersion:"openpond.hostedProfileCandidateExport.v1",actorId:t.actorId,teamId:t.teamId,candidateId:t.id,candidateRevision:t.revision,projectId:t.projectId,profileRef:{...t.profileRef,source:"openpond_git"},expectedCommit:t.baseProfileSourceRevision,component:t.component,baseRelease:t.baseRelease,frozenHash:v(t.frozen),executableSourceHash:i.contentHash,protectedClosureHash:i.protectedClosureHash,compiler:{workspaceId:s,name:o.manifest.name,repositoryId:d},sourceManifest:{base64:g.toString("base64"),contentHash:fe(g),sizeBytes:g.length},policySource:f,qualification:{kind:a,ownerActorId:t.actorId,teamId:t.teamId,originalEvidence:{id:t.evidence.manifest.id,contentHash:t.evidence.result.contentHash},baseline:{id:t.comparison.baseline.manifest.id,contentHash:t.comparison.baseline.result.contentHash},candidate:{id:t.comparison.candidate.manifest.id,contentHash:t.comparison.candidate.result.contentHash},comparisonHash:v(t.comparison),recipe:t.test.recipe},createdAt:t.adoptionIntent?t.updatedAt:t.createdAt},r=rr({...m,contentHash:v(m)});if(Buffer.byteLength(JSON.stringify(r))>8*1024*1024)throw new Error("Candidate policy export exceeds the bounded hosted admission size.");return r}h(Sd,"exportHostedProfileCandidate");export{ci as a,Zn as b,Ts as c,Ys as d,Ui as e,td as f,ld as g,Qt as h,wd as i,Sd as j};
+    CREATE TABLE IF NOT EXISTS experiment_improvement_revisions(id TEXT NOT NULL, revision INTEGER NOT NULL, payload TEXT NOT NULL, PRIMARY KEY(id,revision));`);let a=Promise.resolve(),n=h(i=>{let d=a.then(i);return a=d.catch(()=>{}),d},"serialize");function s(i,d,g){let f=t.prepare("SELECT payload FROM experiment_improvements WHERE id=? AND actor_id=? AND team_id=?").get(i,d,g);return f?Se(JSON.parse(String(f.payload))):null}h(s,"row");function o(i){let d=Se(i);return t.prepare("INSERT INTO experiment_improvement_revisions(id,revision,payload) VALUES(?,?,?)").run(i.id,i.revision,JSON.stringify(d)),t.prepare("INSERT INTO experiment_improvements(id,actor_id,team_id,operation_id,request_hash,revision,payload) VALUES(?,?,?,?,?,?,?) ON CONFLICT(id) DO UPDATE SET revision=excluded.revision,payload=excluded.payload").run(i.id,i.actorId,i.teamId,i.operationId,i.requestHash,i.revision,JSON.stringify(d)),d}return h(o,"persist"),{async find(i,d,g){return await a,s(i,d,g)},async read(i,d,g){await a;let f=s(i,d,g);if(!f)throw new Error("Improvement is unavailable to this owner.");return f},async list(i,d,g=30,f){if(await a,g<1||g>50)throw new Error("Improvement list limit is invalid.");return t.prepare("SELECT payload FROM experiment_improvements WHERE actor_id=? AND team_id=? AND id>? ORDER BY id LIMIT ?").all(i,d,f??"",g).map(m=>Se(JSON.parse(String(m.payload))))},create(i){return n(async()=>{t.exec("BEGIN IMMEDIATE");try{let d=t.prepare("SELECT payload FROM experiment_improvements WHERE actor_id=? AND team_id=? AND operation_id=?").get(i.actorId,i.teamId,i.operationId);if(d){let f=Se(JSON.parse(String(d.payload)));if(f.requestHash!==i.requestHash)throw new Error("Improvement operation was reused with different authority or evidence.");return t.exec("COMMIT"),f}let g=o(i);return t.exec("COMMIT"),g}catch(d){throw t.exec("ROLLBACK"),d}})},mutate(i,d,g,f,m){return n(async()=>{t.exec("BEGIN IMMEDIATE");try{let r=s(i,d,g);if(!r||r.revision!==f)throw new Error("Improvement changed; reload its current revision.");let p=await m(r);if(p.id!==r.id||p.actorId!==d||p.teamId!==g||p.operationId!==r.operationId||p.requestHash!==r.requestHash||p.revision!==r.revision+1)throw new Error("Improvement transition changed its immutable owner or operation authority.");let l=o(p);return t.exec("COMMIT"),l}catch(r){throw t.exec("ROLLBACK"),r}})},async close(){await a,t.close()}}}h(Br,"createExperimentImprovementStore");import{randomUUID as _r}from"node:crypto";import{promises as q}from"node:fs";import z from"node:path";function yt(e){if(!e||z.isAbsolute(e)||e.includes("\\")||e.includes("\0")||e.split("/").some(t=>!t||t==="."||t===".."))throw new Error("Candidate source path is not contained.");return e}h(yt,"safe");var be=h((e,t)=>e.writablePaths.some(a=>t===a||t.startsWith(`${a}/`)),"selected");function It(e,t){return z.join(e,"library","harnesses","experiment-candidates",v(t))}h(It,"candidateRoot");async function Or(e){let t=Ka.parse(e),a=z.join(t.bundlePath,"source"),n=await se({workspaceId:t.agentSnapshot.metadata.workspaceId??t.workspaceId,sourceDir:a});if(n.sourceRevision!==t.sourceRevision||n.harnessRelease.contentHash!==t.harnessRelease.contentHash||n.agentSnapshot.contentHash!==t.agentSnapshot.contentHash)throw new Error("Candidate baseline source differs from the immutable released bytes.");return n}h(Or,"verifiedBase");async function jr(e){let t=await Or(e.base),a={...e.component,path:yt(e.component.path)},n=t.manifest.files.find(u=>u.path===a.path);if(!n||n.kind!==a.kind||n.visibility!=="policy")throw new Error("Select a released policy component for candidate authoring.");if(a.kind==="workflow"&&!a.workflowId)throw new Error("Select the exact Workflow within its released catalog.");let s=new Set(t.sourceFiles.filter(u=>u.asset.visibility!=="policy").map(u=>u.asset.contentHash)),o=t.sourceFiles.filter(u=>u.asset.visibility==="policy"&&!s.has(u.asset.contentHash));if(!o.some(u=>u.path===a.path))throw new Error("The selected component duplicates protected private source and cannot be disclosed.");let i=a.kind==="workflow"?`workflows/${yt(a.workflowId)}`:null,g=i&&o.some(u=>u.path===`${i}/PROMPT.md`||u.path===`${i}/ACTION.json`)?[a.path,i]:a.kind==="agent"?[a.path.split("/").slice(0,2).join("/")]:a.kind==="skill"?[z.posix.dirname(a.path)]:[a.path];if(a.kind==="agent"&&(!a.path.startsWith("agents/")||a.path.split("/").length<3))throw new Error("Select the exact released Agent source prefix.");if(g.some(u=>u==="."))throw new Error("Candidate authoring needs a focused component source directory.");let f=new Set(o.map(u=>u.path)),m=Ct.parse({...t.manifest,files:t.manifest.files.filter(u=>f.has(u.path)),metadata:{...t.manifest.metadata,profileSourceBindings:Fe(t.manifest).filter(u=>f.has(u.releasedPath)),experimentCandidate:!0}}),r=It(e.storeDir,e.candidateId),p=z.join(r,"authoring"),l=`${r}.preparing-${_r()}`,c={sourceRoot:p,base:{id:e.base.harnessRelease.id,contentHash:e.base.harnessRelease.contentHash,sourceRevision:e.base.sourceRevision},component:a,writablePaths:g,publicManifest:m,publicHashes:Object.fromEntries(o.map(u=>[u.path,u.asset.contentHash])),protectedHashes:Object.fromEntries(t.sourceFiles.filter(u=>!f.has(u.path)).map(u=>[u.path,u.asset.contentHash]))},w={...c,contentHash:v(c)};try{await q.mkdir(z.join(l,"authoring"),{recursive:!0,mode:448}),await q.writeFile(z.join(l,"partition.json"),re(w),{mode:384}),await q.writeFile(z.join(l,"authoring","harness.json"),re(m),{mode:256});for(let u of o){let k=z.join(l,"authoring",...u.path.split("/"));await q.mkdir(z.dirname(k),{recursive:!0,mode:448}),await q.writeFile(k,u.bytes,{mode:be(w,u.path)?384:256})}await q.mkdir(z.dirname(r),{recursive:!0,mode:448});try{await q.rename(l,r)}catch(u){let k=await $e(e.storeDir,e.candidateId).catch(()=>null);if(!k||k.contentHash!==w.contentHash)throw u}return w}finally{await q.rm(l,{recursive:!0,force:!0})}}h(jr,"materializeExperimentCandidateSource");async function $e(e,t){let a=JSON.parse(await q.readFile(z.join(It(e,t),"partition.json"),"utf8")),{contentHash:n,...s}=a;if(v(s)!==n||a.sourceRoot!==z.join(It(e,t),"authoring"))throw new Error("Candidate partition integrity failed.");return Ct.parse(a.publicManifest),a}h($e,"readExperimentCandidatePartition");async function qr(e){if(Buffer.byteLength(e.text)>25e4)throw new Error("Manual instructions exceed the component size limit.");let t=await $e(e.storeDir,e.candidateId);if(t.component.kind==="agent")throw new Error("Manual instruction editing requires an instruction-bearing component; use ordinary Work tools for Agent code.");let a=await Be(t.sourceRoot,t.component.path),n=await q.readFile(a);if(fe(n)!==e.expectedHash)throw new Error("Candidate instructions changed; reload before saving.");if(t.component.kind==="workflow"){let s=JSON.parse(n.toString("utf8")),o=s.workflows?.find(d=>d.id===t.component.workflowId);if(!o||o.invocation?.kind!=="instructions")throw new Error("This Workflow does not expose manual instruction editing.");o.invocation.instructions=e.text;let i=`workflows/${t.component.workflowId}/PROMPT.md`;if(t.publicHashes[i]){let d=await Be(t.sourceRoot,i),g=await q.readFile(d,"utf8"),f=/^---\r?\n[\s\S]*?\r?\n---\r?\n/.exec(g)?.[0]??"";await q.writeFile(d,f+e.text)}await q.writeFile(a,re(s))}else await q.writeFile(a,e.text);return{path:t.component.path,contentHash:fe(await q.readFile(a))}}h(qr,"writeManualCandidateInstructions");async function Nr(e){let t=await $e(e.storeDir,e.candidateId),a=await Or(e.base);if(t.base.contentHash!==e.base.harnessRelease.contentHash||t.base.sourceRevision!==e.base.sourceRevision)throw new Error("Candidate baseline release mismatch.");let n=(await Ga(t.sourceRoot)).filter(m=>m!=="harness.json"),s=await q.readFile(await Be(t.sourceRoot,"harness.json"),"utf8");if(v(JSON.parse(s))!==v(t.publicManifest))throw new Error("Candidate authoring changed the protected source manifest.");let o=new Map;for(let m of n){let r=await q.readFile(await Be(t.sourceRoot,m));if(r.byteLength>1e7)throw new Error("Candidate source exceeds the bounded component file size.");if(!be(t,m)&&fe(r)!==t.publicHashes[m])throw new Error("Candidate changed source outside its selected component.");if(t.protectedHashes[m])throw new Error("Candidate attempted to introduce protected private source.");o.set(m,r)}for(let m of Object.keys(t.publicHashes))if(!o.has(m)&&!be(t,m))throw new Error("Candidate removed source outside its selected component.");for(let m of Fe(a.manifest).filter(r=>be(t,r.releasedPath)))for(let r of Fe(a.manifest).filter(p=>p.profileRelativePath===m.profileRelativePath)){if(t.protectedHashes[r.releasedPath])throw new Error("A selected public file aliases protected source.");let p=o.get(m.releasedPath);p?o.set(r.releasedPath,p):o.delete(r.releasedPath)}let i=new Set(Fe(a.manifest).filter(m=>Fe(a.manifest).some(r=>be(t,r.releasedPath)&&r.profileRelativePath===m.profileRelativePath)).map(m=>m.releasedPath)),d=[...o.keys()].filter(m=>/^workflows\/[^/]+\/(PROMPT\.md|ACTION\.json)$/.test(m));if(t.component.kind==="workflow"&&d.length){let m=new Map([...o].filter(([l])=>l.startsWith("workflows/")&&l!=="workflows/catalog.json"&&l!=="workflows/actions.json").map(([l,c])=>[l,c.toString("utf8")])),r=JSON.parse(o.get("workflows/actions.json").toString("utf8")),p=wa({files:m,skillPaths:new Set(a.manifest.files.filter(l=>l.kind==="skill").map(l=>l.path)),actionIds:new Set(r.actions.map(l=>l.id))});o.set("workflows/catalog.json",Buffer.from(re(p.catalog)))}if(t.component.kind==="workflow"){let m=JSON.parse(Buffer.from(a.sourceFiles.find(l=>l.path===t.component.path).bytes).toString("utf8")),r=JSON.parse(o.get(t.component.path).toString("utf8")),p=h(l=>({...l,workflows:l.workflows.filter(c=>c.id!==t.component.workflowId)}),"withoutSelected");if(v(p(m))!==v(p(r))||r.workflows.filter(l=>l.id===t.component.workflowId).length!==1)throw new Error("Candidate changed another Workflow or the protected catalog structure.")}let g=a.manifest.files.filter(m=>m.visibility!=="policy"||!be(t,m.path)||o.has(m.path));for(let m of n)g.some(r=>r.path===m)||g.push({id:`candidate-asset-${v(m).slice(0,24)}`,path:m,kind:t.component.kind==="skill"?"skill_resource":"asset",parentId:t.component.kind==="skill"?a.manifest.files.find(r=>r.path===t.component.path).id:null,mediaType:Ja(m),visibility:"policy",portability:"portable"});let f=z.join(It(e.storeDir,e.candidateId),`.freeze-${_r()}`);try{await q.mkdir(f,{mode:448});let m=v([...o.entries()].map(([u,k])=>({path:u,contentHash:fe(k)}))),r=m.slice(0,40);for(let u of g){let k=a.sourceFiles.find(y=>y.path===u.path),R=be(t,u.path)||i.has(u.path)?o.get(u.path):k?.bytes;if(u.path==="dependency-lock/profile-import.json"&&R){let y=JSON.parse(Buffer.from(R).toString("utf8"));if(y.source!=="openpond.profile"||y.profileGitHead!==a.manifest.metadata.profileGitHead)throw new Error("Generated provenance differs from the baseline compiler.");R=Buffer.from(re({...y,profileGitHead:r}))}if(!R)throw new Error("Candidate source cannot replace its selected primary or protected closure.");let b=z.join(f,...u.path.split("/"));await q.mkdir(z.dirname(b),{recursive:!0,mode:448}),await q.writeFile(b,R)}await q.writeFile(z.join(f,"harness.json"),re({...a.manifest,files:g,metadata:{...a.manifest.metadata,...a.manifest.metadata.importedFrom==="openpond.profile"?{profileGitHead:r}:{}}}));let p=await se({workspaceId:e.workspaceId,sourceDir:f});for(let u of p.manifest.files.filter(k=>k.kind==="skill")){let k=La(Buffer.from(p.sourceFiles.find(R=>R.path===u.path).bytes).toString("utf8"));if(k.messages.length>0)throw new Error(`Candidate Skill is invalid: ${k.messages.join(" ")}`)}let l=p.sourceFiles.filter(u=>u.asset.contentHash!==a.sourceFiles.find(k=>k.path===u.path)?.asset.contentHash).map(u=>({path:u.path,before:t.publicHashes[u.path]??null,after:u.asset.contentHash})),c=a.sourceFiles.filter(u=>!p.sourceFiles.some(k=>k.path===u.path)).map(u=>({path:u.path,before:u.asset.contentHash,after:null}));return{release:await Xa({storeDir:e.storeDir,workspaceId:e.workspaceId,compiled:p,createdAt:e.createdAt}),diff:[...l,...c],authoringHash:m,profileSourceRevision:r,executableIdentity:Xe(p),partitionHash:t.contentHash}}finally{await q.rm(f,{recursive:!0,force:!0})}}h(Nr,"freezeExperimentCandidateSource");function Fe(e){let t=e.metadata.profileSourceBindings;if(t===void 0)return[];if(!Array.isArray(t))throw new Error("Profile source binding metadata is invalid.");return t.map(a=>{if(!a||typeof a!="object"||typeof a.releasedPath!="string"||typeof a.profileRelativePath!="string")throw new Error("Profile source binding is invalid.");return{releasedPath:yt(a.releasedPath),profileRelativePath:yt(a.profileRelativePath)}})}h(Fe,"bindings");function Ui(e){let t=Br(e.storeDir),a=h(()=>new Date().toISOString(),"now");async function n(r,p,l=!1){let c=await e.store.getHarnessWorkspace(p.ownerWorkspaceId);if(!c)throw new Error("Improvement owner is unavailable.");if(l&&await e.authorizeOwner(r,c,p.profileRef,p.projectId,{requireCurrentBase:!0}),c.revision!==p.ownerRevision||c.sourceRevision!==p.ownerSourceRevision||v(c.currentChannel.release)!==v(p.baseRelease))throw new Error("The live owner changed; reconcile and qualify against its new base before adoption.");return c}h(n,"owner");async function s(r){let p=await e.store.getHarnessReleaseRecord(r.baseRelease.contentHash);if(!p||p.harnessRelease.id!==r.baseRelease.id||p.workspaceId!==r.ownerWorkspaceId)throw new Error("Improvement baseline release is unavailable.");return p}h(s,"base");async function o(r,p,l,c,w){if(!e.projectHumanComparison||!p.projectId)throw new Error("The current accepted Human comparison owner is unavailable.");let u=ya(await e.projectHumanComparison(r,p,l,c,w),{baseline:c,candidate:w});if(u.teamId!==r.teamId||u.projectId!==p.projectId||v(u.selections)!==v(l)||u.original.baseline.id!==c.manifest.id||u.original.baseline.manifestHash!==c.manifest.contentHash||u.original.baseline.resultHash!==c.result.contentHash||u.original.candidate.id!==w.manifest.id||u.original.candidate.manifestHash!==w.manifest.contentHash||u.original.candidate.resultHash!==w.result.contentHash)throw new Error("The accepted Human projection changed its original receipts.");return u}h(o,"humanProjection");async function i(r,p){let l=p.comparison?.humanProjection;if(!l)return;let c=Ie(await e.loadEvidence(r,{id:l.original.baseline.id,contentHash:l.original.baseline.resultHash})),w=Ie(await e.loadEvidence(r,{id:l.original.candidate.id,contentHash:l.original.candidate.resultHash}));if((await o(r,p,l.selections,c,w)).contentHash!==l.contentHash)throw new Error("The accepted Human comparison changed; requalify before adoption.")}h(i,"reauthorizeHuman");async function d(r,p){let l=await t.read(p,r.actorId,r.teamId),c=await e.store.getHarnessWorkspace(l.ownerWorkspaceId);if(!c)throw new Error("Improvement owner is unavailable.");return await e.authorizeOwner(r,c,l.profileRef,l.projectId,{requireCurrentBase:!1}),await e.loadEvidence(r,{id:l.evidence.manifest.id,contentHash:l.evidence.result.contentHash}),await i(r,l),l}h(d,"read");async function g(r,p,l,c){let w=await d(r,p);return await n(r,w,!0),t.mutate(p,r.actorId,r.teamId,l,c)}h(g,"mutateOwned");let f=h((r,p)=>{let{contentHash:l,...c}=r;return Je({...c,...p,revision:r.revision+1,updatedAt:a()})},"transition");return{read:d,list:h(async(r,p=30,l)=>{let c=await t.list(r.actorId,r.teamId,p,l),w=[];for(let u of c)try{w.push(await d(r,u.id))}catch{}return{states:w,nextCursor:c.length===p?c.at(-1).id:null}},"list"),async start(r,p){let l=Fa.parse(p),c=v(l),w=`improve-${v([r.actorId,r.teamId,l.operationId]).slice(0,40)}`,u=await t.find(w,r.actorId,r.teamId);if(u){if(u.requestHash!==c)throw new Error("The original creation operation was reused with another request.");return d(r,u.id)}if(l.mode==="manual"&&l.component.kind==="agent")throw new Error("Manual Improve edits instruction-bearing components; use ordinary Work authoring for Agent code.");let k=await e.store.getHarnessWorkspace(l.ownerWorkspaceId);if(!k||k.location!=="local")throw new Error("Select an admitted local Profile owner.");if(await e.authorizeOwner(r,k,l.profileRef,l.projectId,{requireCurrentBase:!0}),k.revision!==l.expectedOwnerRevision||v(k.currentChannel.release)!==v(l.baseRelease))throw new Error("Improvement base owner changed.");let R=Ie(await e.loadEvidence(r,l.evidence));if(R.result.contentHash!==l.evidence.contentHash||R.manifest.id!==l.evidence.id)throw new Error("Improvement evidence identity changed.");if((await e.store.getHarnessReleaseRecord(l.baseRelease.contentHash))?.harnessRelease.metadata.profile?.sourceRevision!==l.profileSourceRevision)throw new Error("Improvement source revision differs from its exact owner Profile release.");let H=a(),x=await t.create(Je({schemaVersion:"openpond.experimentImprovement.v1",id:w,operationId:l.operationId,requestHash:c,...r,projectId:l.projectId,ownerWorkspaceId:k.id,ownerRevision:k.revision,ownerSourceRevision:k.sourceRevision,baseRelease:l.baseRelease,profileRef:l.profileRef,baseProfileSourceRevision:l.profileSourceRevision,evidence:R,component:l.component,mode:l.mode,status:"draft",revision:1,limits:l.limits,authoringSteps:0,work:null,frozen:null,test:null,comparison:null,adoptionIntent:null,acceptance:null,rollback:null,error:null,createdAt:H,updatedAt:H}));return await jr({storeDir:e.storeDir,candidateId:w,base:await s(x),component:x.component}),await n(r,x),x},async edit(r,p,l,c,w){return g(r,p,l,async u=>{if(await n(r,u),["authoring_queued","authoring","testing","adopting","accepted","rolled_back","declined","cancelled"].includes(u.status))throw new Error("Stop active authoring/evaluation before editing this candidate.");return await qr({storeDir:e.storeDir,candidateId:p,expectedHash:c,text:w}),f(u,{status:"draft",frozen:null,test:null,comparison:null,error:null})})},async setMode(r,p,l,c){return g(r,p,l,async w=>{if(w.status!=="draft")throw new Error("Change mode only in a settled authoring draft.");if(c==="manual"&&w.component.kind==="agent")throw new Error("Use ordinary Work for Agent source authoring.");return w.mode===c?w:f(w,{mode:c})})},async queueAuthoringTurn(r,p,l,c,w,u){return g(r,p,l,async k=>{if(k.mode!=="llm_assisted"||k.status!=="draft")throw new Error("Queue only a settled LLM authoring draft.");return f(k,{status:"authoring_queued",work:{sessionId:c,turnId:w,revision:k.revision+2,startedAt:u},frozen:null,test:null,comparison:null,error:null})})},async cancelQueuedAuthoring(r,p,l){return g(r,p,l,async c=>{if(c.status!=="authoring_queued")throw new Error("Cancel only a queued candidate turn after actual host cancellation.");return f(c,{status:"cancelled",error:"The owner cancelled queued candidate Work before provider dispatch."})})},async bindAuthoringTurn(r,p,l,c,w){return g(r,p,l,async u=>{if(await n(r,u),u.mode!=="llm_assisted"||u.status!=="draft"&&u.status!=="authoring_queued"||u.status==="authoring_queued"&&(u.work?.sessionId!==c||u.work.turnId!==w))throw new Error("An explicit LLM authoring action is required.");let k=await e.store.getSession(c),R=await e.store.getTurn(w);if(!k||k.experience!=="work"||!R||R.sessionId!==c||R.status!=="in_progress")throw new Error("Candidate authoring requires its actual ordinary Work session and turn.");let b={candidateId:p,authoringRevision:l+1,actorId:r.actorId,teamId:r.teamId};if(v(R.metadata.refinementCandidate)!==v(b)||R.metadata.source!=="experiment-improvement")throw new Error("The Work turn has no trusted candidate admission marker.");return f(u,{status:"authoring",work:{sessionId:c,turnId:w,revision:l+1,startedAt:a()},authoringSteps:0,frozen:null,test:null,comparison:null,error:null})})},async authoringStep(r,p,l){return g(r,p,l,async c=>{if(await n(r,c),c.status!=="authoring"||!c.work)throw new Error("Candidate Work is not authoring.");if(c.authoringSteps>=c.limits.maximumAuthoringSteps||Date.now()-Date.parse(c.work.startedAt)>c.limits.maximumDurationMs)throw new Error("Candidate authoring reached its admitted step or time limit.");return f(c,{authoringSteps:c.authoringSteps+1})})},async beginTestDispatch(r,p,l,c){return g(r,p,l,async w=>{if(w.status!=="testing"||!w.test||w.test.cancelRequestedAt||w.test.dispatching)throw new Error("The test is cancelled or another exact dispatch owns admission.");return f(w,{test:{...w.test,dispatching:c}})})},async endTestDispatch(r,p,l,c){return g(r,p,l,async w=>{if(w.status!=="testing"||!w.test||w.test.dispatching?.token!==c)throw new Error("The test dispatch ownership changed.");return f(w,{test:{...w.test,dispatching:null}})})},async cancelTesting(r,p,l){return g(r,p,l,async c=>{if(c.status!=="testing"||!c.test)throw new Error("Cancel only an admitted candidate test.");return c.test.cancelRequestedAt?c:f(c,{test:{...c.test,cancelRequestedAt:a()}})})},async finishAuthoring(r,p,l){return g(r,p,l,async c=>{if(await n(r,c),c.status!=="authoring"||!c.work)throw new Error("Candidate is not authoring.");let w=await e.store.getTurn(c.work.turnId);if(!w||w.status==="in_progress")throw new Error("Wait for the admitted Work turn to finish or cancel it.");return f(c,{status:w.status==="completed"?"draft":"failed",error:w.status==="completed"?null:"Candidate Work did not complete successfully."})})},async freeze(r,p,l){return g(r,p,l,async c=>{if(await n(r,c),c.status!=="draft")throw new Error("Freeze a settled draft before testing.");let w=`candidate-${v([p,c.revision]).slice(0,40)}`,u=await Nr({storeDir:e.storeDir,candidateId:p,base:await s(c),workspaceId:w,createdAt:a()});e.freezeProfile&&(u=await e.freezeProfile(r,c,u));let k={release:{id:u.release.harnessRelease.id,contentHash:u.release.harnessRelease.contentHash},sourceRevision:u.release.sourceRevision,profileSourceRevision:u.profileSourceRevision,authoringHash:u.authoringHash,partitionHash:u.partitionHash,executableSourceHash:u.executableIdentity.contentHash,protectedClosureHash:u.executableIdentity.protectedClosureHash,candidateRevision:c.revision+1,diff:u.diff},R={schemaVersion:"openpond.experimentCandidateOrigin.v1",candidateId:p,actorId:r.actorId,teamId:r.teamId,baseWorkspaceId:c.ownerWorkspaceId,baseProfileRef:c.profileRef,baseProfileSourceRevision:c.baseProfileSourceRevision,baseRelease:c.baseRelease,frozenRevision:k.candidateRevision,frozenHash:v(k)},b=await e.store.getHarnessWorkspace(w);if(b){if(v(b.currentChannel.release)!==v(k.release)||v(b.metadata.experimentCandidateOrigin)!==v(R))throw new Error("Candidate workspace identity conflict.")}else{let y=Ge(e.storeDir,w);if(await ze.mkdir(y.root,{recursive:!0,mode:448}),await ze.stat(y.source).catch(()=>null)){if((await se({workspaceId:typeof u.release.agentSnapshot.metadata.workspaceId=="string"?u.release.agentSnapshot.metadata.workspaceId:w,sourceDir:y.source})).harnessRelease.contentHash!==k.release.contentHash)throw new Error("Interrupted candidate workspace source differs from its frozen release.")}else{let x=`${y.source}.install-${v(k).slice(0,20)}`;await ze.rm(x,{recursive:!0,force:!0}),await ze.cp(Dr.join(u.release.bundlePath,"source"),x,{recursive:!0,errorOnExist:!0,force:!1}),await ze.rename(x,y.source)}await e.store.createHarnessWorkspaceWithRelease({workspace:Le.parse({schemaVersion:"openpond.harnessWorkspace.v1",id:w,ownerScope:{kind:"personal",id:r.actorId},name:`Candidate ${p}`,location:"local",sourceRevision:u.release.sourceRevision,revision:0,dirty:!1,currentChannel:{name:"candidate",release:k.release,revision:1},createdAt:a(),updatedAt:a(),metadata:{sourceLayout:"openpond.harnessSourceManifest.v1",selectionEligible:!1,experimentCandidateOrigin:R}}),release:u.release})}return await n(r,c,!0),f(c,{status:"frozen",frozen:k,test:null,comparison:null,error:null})})},async beginTesting(r,p,l,c){return g(r,p,l,async w=>{await n(r,w);let u=Da.parse(c),k=u.externalDatasetBinding;if(w.status==="testing"&&w.test){if(v(u)!==w.test.recipeHash)throw new Error("The admitted test recipe cannot change during execution.");return w}if(w.status!=="frozen"||!w.frozen)throw new Error("Test a frozen candidate revision.");let R=w.evidence.manifest;if(k.profileId!==w.profileRef.profileId||v(k.dataset)!==v(R.dataset)||!u.scoring&&v(k.evaluators)!==v(R.evaluators)||u.scoring&&u.scoring.sourcePass.id!==R.lineage?.scoringPassId)throw new Error("The qualification changed the original Dataset or evaluator selection.");let b=k.recordedOrigin?k.recordedOrigin.members.map(I=>({caseId:I.caseId,seed:I.seed,fixtureId:I.fixtureId})):k.population.map(I=>({caseId:I.taskId,seed:I.seed,fixtureId:I.fixtureId}));if(v(b)!==v(R.population)||k.recordedOrigin&&(k.recordedOrigin.originalEvidence.id!==R.id||k.recordedOrigin.originalEvidence.contentHash!==w.evidence.result.contentHash))throw new Error("The qualification changed the exact original ordered population or recorded cutoff.");if(k.protectedProfileClosureHash!==w.frozen.protectedClosureHash)throw new Error("The test recipe changed the protected Profile evaluator closure.");let y=k.target;if(w.component.kind==="workflow"?y.kind!=="workflow"||y.workflowId!==w.component.workflowId:w.component.kind==="skill"?y.kind!=="skill"||y.skillPath!==w.component.path:w.component.kind==="instruction"?y.kind!=="profile":y.kind!=="agent_action")throw new Error("Test the selected authored Profile component.");if(w.component.kind==="agent"){let I=await s(w),P=await se({workspaceId:typeof I.agentSnapshot.metadata.workspaceId=="string"?I.agentSnapshot.metadata.workspaceId:I.workspaceId,sourceDir:Dr.join(I.bundlePath,"source")}),T=P.sourceFiles.find(D=>D.path==="workflows/actions.json")?.bytes,S=T&&ga.parse(JSON.parse(Buffer.from(T).toString("utf8"))).actions.find(D=>y.kind==="agent_action"&&D.id===y.actionId);if(!S||!w.component.path.startsWith(`agents/${S.agentId}/`))throw new Error("The tested Agent action belongs to a different authored component.")}let H=v(u),x=h(I=>`improve-${I}-${v([p,w.frozen,H]).slice(0,40)}`,"operation");return f(w,{status:"testing",test:{recipe:u,recipeHash:H,baselineOperationId:x("base"),candidateOperationId:x("candidate"),cancelRequestedAt:null,dispatching:null}})})},async failTesting(r,p,l,c,w=!1){return g(r,p,l,async u=>{if(u.status!=="testing")throw new Error("The candidate has no active test.");return f(u,{status:w?"cancelled":"failed",error:c.slice(0,2e3)})})},async completeComparison(r,p,l,c,w,u){let k=await d(r,p),R=k.comparison?.humanProjection;if(u&&k.status==="ready_for_review"&&k.revision===l+1&&R){if(v(R.selections)!==v(u)||R.original.baseline.id!==c.manifest.id||R.original.baseline.resultHash!==c.result.contentHash||R.original.candidate.id!==w.manifest.id||R.original.candidate.resultHash!==w.result.contentHash)throw new Error("The original Human comparison command changed.");let b=await d(r,p);if(b.contentHash!==k.contentHash)throw new Error("The retained comparison changed during recovery.");return b}return g(r,p,l,async b=>{if(await n(r,b),b.status!=="testing"||!b.frozen||b.test?.cancelRequestedAt||b.test?.dispatching)throw new Error("Candidate has no settled uncancelled admitted test.");let y=Ie(await e.loadEvidence(r,{id:c.manifest.id,contentHash:c.result.contentHash})),H=Ie(await e.loadEvidence(r,{id:w.manifest.id,contentHash:w.result.contentHash}));if(!b.test)throw new Error("The comparison has no retained test operations.");for(let[A,U]of[["baseline",y],["candidate",H]]){let O=b.test.recipe.scoring?`improve-grade-${v([b.id,b.test.recipeHash,A]).slice(0,40)}`:A==="baseline"?b.test.baselineOperationId:b.test.candidateOperationId;if(U.manifest.operationId!==O)throw new Error("Compare only the exact originally admitted baseline and candidate operations.")}let x=u?await o(r,b,u,y,H):void 0,I=x?.baseline??y,P=x?.candidate??H,T=h(A=>A.target.kind==="harness"||A.target.kind==="agent"?A.target.source.harnessRelease:null,"releaseOf");if(v(T(I.manifest))!==v(b.baseRelease)||v(T(P.manifest))!==v(b.frozen.release))throw new Error("The comparison did not execute the exact baseline and frozen candidate releases.");if(I.result.status!=="completed"||P.result.status!=="completed"||[...I.result.cases,...P.result.cases].some(A=>A.status!=="completed"||A.feedback.some(U=>U.status!=="scored")))throw new Error("Complete all admitted case execution and grading before adopting this candidate.");if(!ka(I,P).comparable)throw new Error("Candidate and baseline require a matched population, environment and graders.");if(!b.test)throw new Error("The comparison has no admitted immutable rerun recipe.");let S=b.test.recipe.externalDatasetBinding,D=S.population.map(A=>({caseId:A.taskId,seed:A.seed,fixtureId:A.fixtureId}));if(v(I.manifest.dataset)!==v(S.dataset)||v(I.manifest.population)!==v(D)||v(y.manifest.evaluators)!==v(b.test.recipe.scoring?b.evidence.manifest.evaluators:S.evaluators))throw new Error("The qualification changed its admitted rerun population or evaluator selection.");for(let A of[I,P])if(A.manifest.target.kind!=="harness"&&A.manifest.target.kind!=="agent"||!("profileId"in A.manifest.target.source)||v(A.manifest.target.source.externalDatasetBinding)!==v(S))throw new Error("The executed Profile omitted its actual immutable external Dataset recipe.");return f(b,{status:"ready_for_review",comparison:{baseline:I,candidate:P,frozenHash:v(b.frozen),...x?{humanProjection:x}:{}}})})},async beginAdoption(r,p,l){let c=await d(r,p);return c.status==="adopting"||c.status==="accepted"?c:g(r,p,l,async w=>{if(await n(r,w),w.status!=="ready_for_review"||!w.frozen||!w.comparison)throw new Error("Use candidate requires its complete exact tested revision.");return f(w,{status:"adopting",adoptionIntent:{operationId:`adopt-${v([p,w.frozen]).slice(0,40)}`,expectedOwnerRevision:w.ownerRevision,frozenHash:v(w.frozen)}})})},async useCandidate(r,p,l){let c=await d(r,p);if(c.status==="accepted")return c;if(c.status!=="adopting"&&(c=await this.beginAdoption(r,p,l)),!c.frozen||!c.adoptionIntent)throw new Error("Adoption has no durable tested intent.");let w=await e.store.getHarnessReleaseRecord(c.frozen.release.contentHash);if(!w||w.harnessRelease.id!==c.frozen.release.id)throw new Error("Tested candidate release is unavailable.");await i(r,c);let u=await e.readAdoption(r,c,c.adoptionIntent.operationId)??await e.adopt(r,c,w,c.adoptionIntent.operationId);if(v(u.testedCandidateRelease)!==v(c.frozen.release)||u.executableSourceHash!==c.frozen.executableSourceHash||u.protectedClosureHash!==c.frozen.protectedClosureHash)throw new Error("Owner adoption differs from tested executable source or protected closure.");return await i(r,c),t.mutate(p,r.actorId,r.teamId,c.revision,async k=>{if(k.status!=="adopting"||v(k.adoptionIntent)!==v(c.adoptionIntent))throw new Error("Adoption intent changed during owner readback.");return f(k,{status:"accepted",acceptance:u.receipt})})},async rollback(r,p,l){let c=await d(r,p);if(c.status==="rolled_back")return c;if(c.status!=="accepted"||c.revision!==l||!c.acceptance)throw new Error("Rollback requires the current accepted Profile adoption.");let w=await e.rollback(r,c,`rollback-${c.acceptance.contentHash.slice(0,40)}`);if(v(w.testedCandidateRelease)!==v(c.baseRelease))throw new Error("Profile rollback did not restore the original executable release.");return t.mutate(p,r.actorId,r.teamId,l,async u=>f(u,{status:"rolled_back",rollback:w.receipt}))},async discard(r,p,l,c=!1){return g(r,p,l,async w=>{if(await n(r,w),["authoring_queued","authoring","testing"].includes(w.status))throw new Error("Cancel the actual Work/evaluation execution before discarding the candidate.");if(w.status==="accepted"||w.status==="rolled_back"||w.status==="adopting")throw new Error("Recover adoption or use the owner rollback operation for an adopted release.");return f(w,{status:c?"cancelled":"declined"})})},async partition(r,p){let l=await d(r,p);return await n(r,l),$e(e.storeDir,p)},states:t,close:h(()=>t.close(),"close")}}h(Ui,"createLocalExperimentImprovementService");import K from"node:path";import{promises as Kt}from"node:fs";import{tmpdir as qn}from"node:os";var St=h(e=>`'${e.replaceAll("'","'\\''")}'`,"quote");function td(e){async function t(o,i){let d=await e.store.getSession(o),g=await e.store.getTurn(i);if(!d||!g||g.sessionId!==o)throw new Error("Candidate tool has no actual Work turn authority.");let f=g.metadata?.refinementCandidate;if(g.metadata?.source!=="experiment-improvement"||!f||typeof f!="object"||!("candidateId"in f)||typeof f.candidateId!="string"){if(d.metadata?.source==="experiment-improvement"||d.metadata?.refinementCandidate!==void 0)throw new Error("This candidate Session requires a fresh private stored-turn admission before any continuation.");return null}let m={actorId:await e.actorId(),teamId:await e.teamId()},r=await e.improvements.read(m,f.candidateId);if(r.status!=="authoring"||!r.work||r.work.sessionId!==o||r.work.turnId!==i||g.status!=="in_progress"||d.experience!=="work"||v(f)!==v({candidateId:r.id,authoringRevision:r.work.revision,...m})||v(d.currentProfile)!==v(r.profileRef))throw new Error("Candidate tool binding changed or its Work turn has ended.");let p=await e.improvements.partition(m,r.id);return{actor:m,state:r,partition:p}}h(t,"authority");let a=e.agentRuntime,n=Qa({bwrapPath:e.bwrapPath,authorize:h(async o=>{let i=await t(o.sessionId,o.turnId);if(!i||i.state.id!==o.candidateId||i.state.revision!==o.expectedRevision)throw new Error("Candidate command changed before filesystem admission.");return{candidateId:i.state.id,candidateRevision:i.state.revision,ownerId:i.actor.actorId,sessionId:o.sessionId,turnId:o.turnId,sourceRoot:i.partition.sourceRoot,writablePaths:i.partition.writablePaths,...a?{runtimeMounts:[{source:a.source,destination:"/runtime/agent"},{source:K.join(a.source,"node_modules"),destination:"/workspace/node_modules"}]}:{}}},"authorize")});async function s(o,i,d,g={}){e.operationSignal?.throwIfAborted();let f=await t(o,i);if(!f)throw new Error("This command is not admitted candidate Work.");let m=await e.improvements.authoringStep(f.actor,f.state.id,f.state.revision),r=g.cwd?.startsWith("/workspace/work/")?K.resolve(f.partition.sourceRoot,g.cwd.slice(16)):g.cwd==="/workspace/work"?f.partition.sourceRoot:g.cwd;return n({candidateId:m.id,sessionId:o,turnId:i,expectedRevision:m.revision,command:d,...g,signal:g.signal&&e.operationSignal?AbortSignal.any([g.signal,e.operationSignal]):g.signal??e.operationSignal,cwd:r})}return h(s,"command"),{async candidateAuthoringForTurn(o,i){return await t(o,i)!==null},async authorizeCandidateTool(o){if(!await t(o.sessionId,o.turnId))return!1;if(!nr.has(o.name))throw new Error("This tool has no authority over isolated candidate Work.");return!0},async resolveCandidateProfile(o){let i=await t(o.session.id,o.turnId);if(!i)throw new Error("Candidate Profile requires a trusted Work admission.");let d=await e.loadProfile(i.state.profileRef),g=i.partition.publicManifest,f={...d,repoPath:i.partition.sourceRoot,sourcePath:i.partition.sourceRoot,skills:d.skills.filter(m=>g.files.some(r=>r.kind==="skill"&&r.path===`skills/${_e(m.name)}/SKILL.md`)).map(m=>({...m,path:`skills/${_e(m.name)}/SKILL.md`,sourcePath:i.partition.sourceRoot})),agents:d.agents.filter(m=>g.files.some(r=>r.kind==="agent"&&r.path.startsWith(`agents/${_e(m.id)}/`))).map(m=>({...m,path:g.files.find(r=>r.kind==="agent"&&r.path.startsWith(`agents/${_e(m.id)}/`)).path})),evals:[],git:d.git?{...d.git,dirty:!0,files:[]}:null};return await t(o.session.id,o.turnId),f},async executeCandidateCommand(o,i){let d=await s(o.session.id,o.turnId,i.command,{cwd:i.cwd??void 0,timeoutMs:(i.timeoutSeconds??120)*1e3,signal:i.signal});return{ok:d.code===0&&!d.timedOut,command:i.command,cwd:i.cwd??o.session.cwd,exitCode:d.code,stdout:d.stdout,stderr:d.stderr,timedOut:d.timedOut,timeoutSeconds:Math.min(i.timeoutSeconds??120,180),truncated:d.stdoutTruncated||d.stderrTruncated,blockedReason:null}},async executeCandidateImage(o){let i=await t(o.session.id,o.turnId);if(!i)throw new Error("Image inspection is not admitted candidate Work.");let d=o.args.path;if(typeof d!="string"||!d.trim())throw new Error("Select a candidate image path.");let g=d.startsWith("/workspace/work/")?K.resolve(i.partition.sourceRoot,d.slice(16)):K.resolve(i.partition.sourceRoot,d),f=K.relative(i.partition.sourceRoot,g);if(!f||f.startsWith("..")||K.isAbsolute(f))throw new Error("The image path escapes this candidate source.");let m=`/workspace/work/${f.split(K.sep).join("/")}`,r=h(async(y,H,x)=>{let I=await s(o.session.id,o.turnId,[y,...H].map(St).join(" "),{signal:x,maxOutputBytes:75e4,timeoutMs:3e4});if(I.code!==0||I.timedOut||I.stdoutTruncated)throw new Error("The confined image decoder could not inspect this candidate image.");return I.stdout},"run"),p=await or(m,o.signal,r),l={PNG:"png",JPEG:"jpg",GIF:"gif",WEBP:"webp",AVIF:"avif"}[p.format];if(!l)throw new Error("Select a PNG, JPEG, GIF, WebP or AVIF image.");let c=await r("/runtime/node",["-e","const f=require('fs'),p=process.argv[1],b=f.readFileSync(p);if(b.length>500000)throw Error('Create a contact sheet under 500000 bytes for candidate inspection');process.stdout.write(b.toString('base64'))",m],o.signal);if(!/^[A-Za-z0-9+/]*={0,2}$/.test(c)||c.length>67e4)throw new Error("Candidate image bytes exceed the inspection boundary.");let w=await Kt.mkdtemp(K.join(qn(),"candidate-image-"));await Kt.chmod(w,448);let u=K.join(w,`preview.${l}`),k=Buffer.from(c,"base64");await Kt.writeFile(u,k,{mode:256}),await t(o.session.id,o.turnId);let R=`image/${l==="jpg"?"jpeg":l}`,b={path:m,contentType:R,sizeBytes:k.length,...p};return{toolCallId:o.callId,name:"view_image",ok:!0,contentText:JSON.stringify({ok:!0,action:"view_image",data:b}),data:{...b,openpondImagePreviewPath:u}}},async executeCandidateAgentCommand(o){let i=await t(o.context.session.id,o.context.turnId);if(!i)throw new Error("Agent command is not admitted candidate Work.");if(a??=await e.loadAgentRuntime?.(),!a)throw new Error("The server has not staged a dependency-complete confined Agent runtime; build and install that public runtime before authoring Agent components.");let d=a.cliRelativePath;if(K.isAbsolute(d)||d.split(/[\\/]/).some(r=>!r||r===".."||r==="."))throw new Error("Agent runtime CLI binding is invalid.");let g=K.relative(i.partition.sourceRoot,o.cwd);if(g.startsWith("..")||K.isAbsolute(g))throw new Error("Agent command source escapes its candidate.");let f=`/workspace/work${g?`/${g.split(K.sep).join("/")}`:""}`,m=o.command==="run"?[o.command,...o.args??[]]:[o.command,"--cwd",f,...o.args??[]];return s(o.context.session.id,o.context.turnId,`/runtime/node ${St(`/runtime/agent/${d}`)} ${m.map(St).join(" ")}`,{cwd:o.cwd,timeoutMs:o.timeoutMs,maxOutputBytes:o.maxOutputBytes,signal:o.context.signal})},executeCandidateWorkspaceTool:h((async o=>{if(!o.turnId){if(o.session.metadata?.source==="experiment-improvement"||o.session.metadata?.refinementCandidate!==void 0)throw new Error("Candidate files require a trusted actual turn.");return null}let i=await t(o.session.id,o.turnId);if(!i)return null;if(!new Set(["list_files","read_files","search_files","workspace_status","preview_write_file","preview_write_files","preview_edit_file","preview_delete_file","write_file","write_files","edit_file","delete_file"]).has(o.request.action))throw new Error("Use the candidate file tools or confined exec_command/Agent tools; this operation has no authority over the live Profile.");let g=await s(o.session.id,o.turnId,`/runtime/node -e ${St(sr)}`,{stdin:JSON.stringify({action:o.request.action,args:o.request.args,writablePaths:i.partition.writablePaths}),maxOutputBytes:8e5});if(g.code!==0||g.timedOut||g.stdoutTruncated)throw new Error(g.stderr.slice(0,2e3)||"Candidate source operation did not complete.");return{ok:!0,action:o.request.action,output:"Completed in the isolated candidate source.",data:JSON.parse(g.stdout)}}),"executeCandidateWorkspaceTool")}}h(td,"createExperimentCandidateTools");import{createHash as $r}from"node:crypto";import{mkdir as Xt,writeFile as Nn,rm as Dn,rename as Fn,readdir as $n,readFile as zn,lstat as Un}from"node:fs/promises";import{randomUUID as Ln}from"node:crypto";import Ht from"node:path";var Vn=M.object({path:M.string(),type:M.enum(["file","directory"]),sizeBytes:M.number().nullable(),oid:M.string().nullable(),mode:M.string().nullable()}),Fr=M.object({commitSha:M.string(),entries:M.array(Vn).max(5e3),truncated:M.boolean(),selectedFile:M.object({path:M.string(),sizeBytes:M.number(),contents:M.string().nullable(),encoding:M.literal("utf8").nullable(),isBinary:M.boolean(),truncated:M.boolean()}).nullable()});async function ld(e){let t=Fr.parse(await e.read({branch:e.revision,path:"__openpond_profile_origin_tree_only__",maxEntries:5e3}));if(t.commitSha!==e.revision||t.truncated)throw new Error("Profile origin tree is incomplete or changed its commit.");let a=t.entries.filter(i=>i.type==="file");if(!a.length||a.length>5e3)throw new Error("Profile origin has no complete bounded source tree.");let n=0,s=new Set;for(let i of a){if(i.path.startsWith("/")||i.path.includes("\\")||i.path.split("/").some(d=>!d||d==="."||d==="..")||s.has(i.path)||!["100644","100755"].includes(i.mode??"")||!Number.isSafeInteger(i.sizeBytes)||i.sizeBytes<0||i.sizeBytes>2e6||!i.oid?.match(/^[a-f0-9]{40}$/))throw new Error("Profile origin contains unsupported paths, modes, or bytes.");if(i.path.split("/").some(d=>d===".git"||d.startsWith(".env")||/\.(sql|sqlite|sqlite3|db)$/i.test(d)))throw new Error("Profile origin contains an unsupported secret or database source path.");s.add(i.path),n+=i.sizeBytes}if(n>32e6)throw new Error("Profile origin exceeds the bounded source size.");let o=`${e.root}.partial-${Ln()}`;await Xt(o,{recursive:!0,mode:448});try{let i=[];for(let d of a){let g=Fr.parse(await e.read({branch:e.revision,path:d.path,maxEntries:5e3})),f=g.selectedFile;if(g.commitSha!==e.revision||!f||f.path!==d.path||f.isBinary||f.truncated||f.encoding!=="utf8"||f.contents===null)throw new Error("Profile origin cannot supply complete exact source bytes.");let m=Buffer.from(f.contents,"utf8"),r=$r("sha1").update(`blob ${m.length}\0`).update(m).digest("hex");if(m.length!==d.sizeBytes||f.sizeBytes!==d.sizeBytes||r!==d.oid)throw new Error("Profile origin file differs from its committed Git blob.");let p=Ht.join(o,...d.path.split("/"));await Xt(Ht.dirname(p),{recursive:!0,mode:448}),await Nn(p,m,{mode:384,flag:"wx"}),i.push({path:d.path,oid:r,sizeBytes:m.length})}return await Xt(Ht.dirname(e.root),{recursive:!0,mode:448}),await Fn(o,e.root),i.sort((d,g)=>d.path<g.path?-1:d.path>g.path?1:0)}catch(i){throw await Dn(o,{recursive:!0,force:!0}),i}}h(ld,"materializeProfileOrigin");async function Qt(e){let t=[],a=0;async function n(s,o){for(let i of await $n(s,{withFileTypes:!0})){let d=o?`${o}/${i.name}`:i.name,g=Ht.join(s,i.name),f=await Un(g);if(f.isSymbolicLink())throw new Error("Profile origin cache contains a symlink.");if(f.isDirectory()){await n(g,d);continue}if(!f.isFile()||f.size>2e6||t.length>=5e3||(a+=f.size)>32e6)throw new Error("Profile origin cache exceeds its bounds.");let m=await zn(g);t.push({path:d,oid:$r("sha1").update(`blob ${m.length}\0`).update(m).digest("hex"),sizeBytes:m.length})}}return h(n,"visit"),await n(e,""),v(t.sort((s,o)=>s.path<o.path?-1:s.path>o.path?1:0))}h(Qt,"profileOriginFilesHash");import{promises as Et}from"node:fs";import Yt from"node:path";async function wd(e,t){let a=await e.store.getHarnessWorkspace(e.state.ownerWorkspaceId),n=a?.metadata.profileExperimentOrigin,s=Yt.join(e.storeDir,"library","profile-experiment-origins",e.state.ownerWorkspaceId);if(e.state.profileRef.source!=="openpond_git"||a?.location!=="local"||a.metadata.selectionEligible!==!1||n?.actorId!==e.state.actorId||n.teamId!==e.state.teamId||n.sourceRevision!==e.state.baseProfileSourceRevision||n.sourcePath!==s||typeof n.sourceFilesHash!="string"||await Qt(s)!==n.sourceFilesHash)throw new Error("The exact authorized committed Profile snapshot is unavailable.");let o=Yt.join(e.storeDir,"library","harnesses","profile-candidate-staging");await Et.mkdir(o,{recursive:!0,mode:448});let i=await Et.mkdtemp(Yt.join(o,"remote-source-"));try{await Et.cp(s,i,{recursive:!0,errorOnExist:!0,force:!1});let d=await Wa(i,e.state.profileRef.profileId);if(d.error||!d.sourcePath)throw new Error("The actual committed Profile archive cannot compile.");let g=await t({root:i,profile:d,source:d.sourcePath});if(await Qt(s)!==n.sourceFilesHash)throw new Error("The committed Profile cache changed during candidate compilation.");return g}finally{await Et.rm(i,{recursive:!0,force:!0})}}h(wd,"withRemoteProfileCandidateSnapshot");async function Sd(e,t,a="owner_attested_local"){if(t.profileRef.source!=="openpond_git"||!t.frozen||!t.comparison||!t.test||!t.adoptionIntent||t.status!=="adopting")throw new Error("Export only the exact owner-approved qualified hosted Profile candidate.");let n=await e.getHarnessReleaseRecord(t.frozen.release.contentHash);if(!n||n.harnessRelease.id!==t.frozen.release.id)throw new Error("The tested native candidate is unavailable.");let s=n.agentSnapshot.metadata.workspaceId;if(typeof s!="string")throw new Error("The candidate compiler identity is absent.");let o=await se({workspaceId:s,sourceDir:`${n.bundlePath}/source`}),i=Xe(o);if(o.harnessRelease.contentHash!==n.harnessRelease.contentHash||i.contentHash!==t.frozen.executableSourceHash||i.protectedClosureHash!==t.frozen.protectedClosureHash||t.comparison.frozenHash!==v(t.frozen))throw new Error("The actual tested source differs from its qualified immutable candidate.");let d=o.manifest.metadata.profileRepositoryId;if(d!==t.profileRef.repositoryId)throw new Error("The candidate source compiler belongs to another Profile owner.");let g=Buffer.from(re(o.manifest)),f=ha(fa({agentSnapshot:o.agentSnapshot,harnessRelease:o.harnessRelease,files:new Map(o.sourceFiles.map(p=>[p.path,p.bytes]))})),m={schemaVersion:"openpond.hostedProfileCandidateExport.v1",actorId:t.actorId,teamId:t.teamId,candidateId:t.id,candidateRevision:t.revision,projectId:t.projectId,profileRef:{...t.profileRef,source:"openpond_git"},expectedCommit:t.baseProfileSourceRevision,component:t.component,baseRelease:t.baseRelease,frozenHash:v(t.frozen),executableSourceHash:i.contentHash,protectedClosureHash:i.protectedClosureHash,compiler:{workspaceId:s,name:o.manifest.name,repositoryId:d},sourceManifest:{base64:g.toString("base64"),contentHash:fe(g),sizeBytes:g.length},policySource:f,qualification:{kind:a,ownerActorId:t.actorId,teamId:t.teamId,originalEvidence:{id:t.evidence.manifest.id,contentHash:t.evidence.result.contentHash},baseline:{id:t.comparison.baseline.manifest.id,contentHash:t.comparison.baseline.result.contentHash},candidate:{id:t.comparison.candidate.manifest.id,contentHash:t.comparison.candidate.result.contentHash},comparisonHash:v(t.comparison),recipe:t.test.recipe},createdAt:t.adoptionIntent?t.updatedAt:t.createdAt},r=rr({...m,contentHash:v(m)});if(Buffer.byteLength(JSON.stringify(r))>8*1024*1024)throw new Error("Candidate policy export exceeds the bounded hosted admission size.");return r}h(Sd,"exportHostedProfileCandidate");export{ci as a,Zn as b,Ts as c,Ys as d,Ui as e,td as f,ld as g,Qt as h,wd as i,Sd as j};