openpond 0.2.65 → 0.2.67

package/dist/chunks/{chunk-LI3FTPAH.js → chunk-VEXJV5H5.js}

@@ -1,5 +1,5 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as As,b as Ms}from"./chunk-4D3CICBG.js";import{a as Es,e as $e,f as Cs,g as Tt}from"./chunk-OW3EOC5Q.js";import{a as xs,c as Is}from"./chunk-XVDFJ3ZF.js";import{h as ne}from"./chunk-3MVKXAH7.js";import{a as ee,b as _s}from"./chunk-67HXM6AL.js";import{b as ys,c as bs,d as yt}from"./chunk-JLW5BQVO.js";import{a as ws,b as vs,c as Ts,d as qe,e as Ss,g as vt}from"./chunk-ANG2YMMD.js";import{a as wt,c as Rs}from"./chunk-6TRVEA3T.js";import{b as Kr,e as Xr,f as lt,g as ut}from"./chunk-LYVHYT3L.js";import{F as gt,G as fs,H as ht,I as kt,J as gs,K as hs,N as ks,x as ps}from"./chunk-66HFW4OU.js";import{g as pt,i as ft}from"./chunk-MCBRB5D6.js";import{a as ve}from"./chunk-7KWD2I6W.js";import{s as Pr}from"./chunk-4TTXNVCE.js";import{d as Ne}from"./chunk-BJXPQIVD.js";import{ja as rs,qa as ss,sa as os,x as ts}from"./chunk-MVPKU72J.js";import{M as V,N as as,O as ns,P as is,Q as mt,R as ds,S as q,T as cs,U as je,V as ls,X as us,Y as Vt,Z as ms}from"./chunk-QCJY6FTC.js";import{l as Ft}from"./chunk-6S2MWJVD.js";import{d as Zr,f as W,g as Qr,i as es}from"./chunk-VS3AMJEK.js";import{f as ct}from"./chunk-5KJMFK65.js";import{$a as Yr,Ba as Jr,Ca as Re,K as Nr,L as qr,M as $r,N as Lr,Na as Gr,Oa as De,P as Ur,aa as Br,ba as Wr,ia as Fr,ua as Vr,wa as we,xa as zr}from"./chunk-WLZJOCLA.js";import{q as ae}from"./chunk-ZWW227M3.js";import{E as Or,F as Hr,Tb as jr,xb as Dr}from"./chunk-QAFXQHHG.js";import{f as Sr}from"./chunk-AJXWJLZA.js";import{K as L,L as Er,Lb as Ar,Q as Wt,R as oe,W as dt,X as Cr,aa as He,d as xr,g as Ir,mb as ke,nb as ye}from"./chunk-G4E3WGLM.js";import{bb as be,cb as _r,ea as it,v as Ge}from"./chunk-ZY32SZ2P.js";import{l as br,m as Rr}from"./chunk-OTDDXEQ6.js";import{i as Tr,m as se,n as k}from"./chunk-RCT3O3HQ.js";import{A as P}from"./chunk-X2J2P5ZT.js";import{m as Mr}from"./chunk-UGOK7VTC.js";import{a as d}from"./chunk-FTQLQJQV.js";var zt={buildIntent:"verifiable_reward",methodHint:"grpo"};async function ii(e){let t=await aa(e.store,e.workspaceId,e.reviewRef),r=`harness-review-${k({review:t.contentHash,strategy:zt}).slice(0,24)}`,s=(await e.store.listTaskCreationSnapshots(e.profileId)).find(c=>c.request.candidateId===r);if(s)return s;let[a,o,n]=await Promise.all([e.store.listHarnessImprovementArtifacts(e.workspaceId,"route_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"trigger_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"observation",1e3)]),i=new Map(a.map(c=>[_e(c),c])),u=new Map(o.map(c=>[_e(c),c])),f=new Map(n.map(c=>[_e(c),c])),p=new Map;for(let c of t.selectedEvidence){if(c.kind!=="route_decision")continue;let m=i.get(_e(c.evidence)),g=m?u.get(_e(m.trigger)):null;if(!m||!g)throw new Error(`Harness review evidence ${c.evidence.id} cannot be reconstructed.`);let h=g.observations.map(y=>f.get(_e(y))??null).filter(y=>y?.runRef===c.sourceRef);if(!h.length)throw new Error(`Harness review source ${c.sourceRef} has no exact observation evidence.`);let w=p.get(c.sourceRef)??{turnIds:new Set,evidenceRefs:[],sourcePolicies:[]};for(let y of h)w.turnIds.add(y.turnId);w.evidenceRefs.push(c.evidence),w.sourcePolicies.push(c.sourcePolicy),p.set(c.sourceRef,w)}if(!p.size)throw new Error("Harness Taskset review has no reconstructable source runs.");let l=[];for(let[c,m]of[...p.entries()].sort(([g],[h])=>g.localeCompare(h))){let g=await e.taskCreator.addSessionSource({profileId:e.profileId,sessionId:c,turnIds:[...m.turnIds].sort(),consentScope:"selected_turns"}),h=Array.isArray(g.metadata.harnessEvaluationReviews)?g.metadata.harnessEvaluationReviews:[],w={...g,metadata:{...g.metadata,harnessEvaluationReviews:Gt([...h,{id:t.id,contentHash:t.contentHash,workspaceId:t.workspaceRef,harnessRelease:t.harnessRelease,claimFingerprint:t.claim?.fingerprint??null,evidenceRefs:m.evidenceRefs,sourcePolicies:m.sourcePolicies}])}};l.push(await e.store.upsertTrainingSource(w))}return e.startCreation({profileId:e.profileId,sourceIds:l.map(c=>c.id),surface:"task_candidate",mode:"customize",entryMode:"automated",resourceIntent:"workproduct",buildIntent:zt.buildIntent,buildSpecification:null,objective:t.claim?.statement??t.reason,methodHint:zt.methodHint,preferredBaseModelId:null,preferredBaseModel:null,candidateId:r,analysisModel:e.analysisModel??null,analysisReasoningEffort:e.analysisReasoningEffort??null,createImproveRunId:null,targetIntent:{kind:"model",id:null,displayName:null,operation:"create"}})}d(ii,"startHarnessReviewTasksetAuthoring");function di(e){let t=e.flatMap(o=>Array.isArray(o.metadata.harnessEvaluationReviews)?o.metadata.harnessEvaluationReviews.filter(Jt):[]);if(!t.length)return null;let r=`${String(t[0].id)}:${String(t[0].contentHash)}`,s=t.filter(o=>`${String(o.id)}:${String(o.contentHash)}`===r);if(s.length!==e.length)return null;let a=s[0];return{review:{id:String(a.id),contentHash:String(a.contentHash),workspaceId:String(a.workspaceId),harnessRelease:a.harnessRelease,claimFingerprint:typeof a.claimFingerprint=="string"?a.claimFingerprint:null},evidenceRefs:Gt(s.flatMap(o=>Array.isArray(o.evidenceRefs)?o.evidenceRefs:[])),sourcePolicies:Gt(s.flatMap(o=>Array.isArray(o.sourcePolicies)?o.sourcePolicies:[]))}}d(di,"harnessReviewLineageFromSources");async function Ps(e){let t=e.taskset.metadata.harnessEvaluationLineage;if(!Jt(t)||!Jt(t.review))return null;let r=String(t.review.workspaceId??""),s={id:String(t.review.id??""),contentHash:String(t.review.contentHash??"")};if(!r||!s.id||!s.contentHash)throw new Error("Taskset Harness Evaluation lineage is incomplete.");let a=await e.store.listHarnessImprovementArtifacts(r,"evaluation_review",1e3),o=a.find(c=>_e(c)===_e(s));if(!o)throw new Error("Originating Harness Evaluation review was not found.");let n=e.evaluationResult?{id:e.evaluationResult.id,contentHash:e.evaluationResult.contentHash}:null,i={id:e.taskset.id,contentHash:e.taskset.contentHash},u=k({originRef:s,tasksetRef:i,evaluationRef:n}),f=a.find(c=>c.metadata.downstreamLinkKey===u);if(f)return f;let p=(e.now??(()=>new Date().toISOString()))(),l=Sr({schemaVersion:"openpond.harnessEvaluationReviewReceipt.v1",id:`evaluation-review-link-${u.slice(0,24)}`,ownerScope:o.ownerScope,workspaceRef:o.workspaceRef,harnessRelease:o.harnessRelease,previousWatermark:o.nextWatermark,nextWatermark:{cursor:k({cursor:o.nextWatermark.cursor,tasksetRef:i,evaluationRef:n}),throughCreatedAt:p},selectedEvidence:o.selectedEvidence,excludedEvidence:o.excludedEvidence,claim:o.claim,classification:"taskset",triage:o.triage,reason:n?"The approved Taskset completed a real baseline Evaluation.":"The approved Harness review materialized one immutable Taskset.",nextAuthority:"human_review",maxEstimatedCostUsd:o.maxEstimatedCostUsd,tasksetProposal:i,evaluation:n,trainingQualification:null,policyVersion:o.policyVersion,createdAt:p,metadata:{originReview:s,downstreamLinkKey:u,tasksetRevision:e.taskset.revision}});return await e.store.saveHarnessImprovementArtifact(r,"evaluation_review",l),l}d(Ps,"linkHarnessReviewTaskset");async function aa(e,t,r){let s=await e.getHarnessWorkspace(t);if(!s)throw new Error(`Harness workspace ${t} was not found.`);let a=(await e.listHarnessImprovementArtifacts(t,"evaluation_review",1e3)).find(n=>n.id===r.id&&n.contentHash===r.contentHash);if(!a||a.schemaVersion!=="openpond.harnessEvaluationReviewReceipt.v1")throw new Error("Harness Evaluation review was not found.");let o=a;if(o.classification!=="taskset"||o.nextAuthority!=="human_review")throw new Error("Only a human-review Taskset recommendation can enter Taskset authoring.");if(o.tasksetProposal)throw new Error("Harness Evaluation review already has a materialized Taskset link.");if(o.ownerScope.kind!==s.ownerScope.kind||o.ownerScope.id!==s.ownerScope.id)throw new Error("Harness Evaluation review owner scope does not match its workspace.");return o}d(aa,"requireTasksetReview");function _e(e){return`${e.id}:${e.contentHash}`}d(_e,"artifactKey");function Jt(e){return!!e&&typeof e=="object"&&!Array.isArray(e)}d(Jt,"isRecord");function Gt(e){let t=new Set;return e.filter(r=>{let s=JSON.stringify(r);return t.has(s)?!1:(t.add(s),!0)})}d(Gt,"uniqueObjects");import{randomUUID as na}from"node:crypto";function bt(e){let t=St(St(St(e.metadata).billing).pricing);return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}d(bt,"hostedTokenPricingFromCatalog");function mi(e){let t=St(e);try{return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}catch{return null}}d(mi,"hostedTokenPricingFromValue");function xe(e,t){let r=ve(e);if(r.promptTokens===null&&r.completionTokens===null)return null;let s=r.cachedPromptTokens??0;return((r.uncachedPromptTokens??r.promptTokens??0)*t.inputUsdPerMillionTokens+s*t.cachedInputUsdPerMillionTokens+(r.completionTokens??0)*t.outputUsdPerMillionTokens)/1e6}d(xe,"hostedUsageCostUsd");function St(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(St,"record");function Le(e,t){if(typeof e=="string"&&e.trim())return e.trim();throw new Error(`Hosted model catalog is missing ${t}.`)}d(Le,"requiredString");function Ue(e,t){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return e;throw new Error(`Hosted model catalog is missing ${t}.`)}d(Ue,"requiredRate");function Ds(e={}){let t=e.catalog??pt,r=e.stream??ft;return{async prepare(s){if(s.providerId!=="openpond")throw new Error("The selected judge provider is not connected to the hosted judge transport.");if(s.revision!==null)throw new Error("This judge transport cannot enforce a model revision; select a provider with revision support.");let a=await t();if(a.error)throw new Error(`Judge model catalog unavailable: ${a.error}`);let o=a.models.find(l=>l.id===s.modelId);if(!o)throw new Error("The selected judge model is unavailable.");let n=Os(o.raw),i=bt(n),u=Hs(n.context_window,"context window"),f=Math.min(4096,Hs(n.output_limit,"output limit")),p=(u*Math.max(i.inputUsdPerMillionTokens,i.cachedInputUsdPerMillionTokens)+f*i.outputUsdPerMillionTokens)/1e6;if(!(p>0))throw new Error("The judge model must have a positive known charge ceiling.");return{maximumChargeUsd:p,async dispatch(l){l?.throwIfAborted();let c="",m=null,g=null,h=null,w=!1;for await(let _ of r({model:s.modelId,messages:[{role:"system",content:s.system},{role:"user",content:s.data}],requestId:`learning-judge-${na()}`,maxTokens:f,temperature:s.temperature,signal:l})){let x=Os(_.raw);if(typeof x.model=="string"){if(m!==null&&x.model!==m)throw new Error("Judge response model identity changed during streaming.");m=x.model}if(typeof x.id=="string"){if(g!==null&&g!==x.id)throw new Error("Judge response identity changed during streaming.");g=x.id}if(_.type==="text_delta"&&(c+=_.text),c.length>1e5)throw new Error("Judge response exceeds the retained response limit.");_.type==="usage"&&(h=_.usage),_.type==="finish"&&(w=!0)}if(!w||!m||!g)throw new Error("Judge response did not include a complete model receipt.");let y=ve(h),S=y.promptTokens!==null&&y.completionTokens!==null?xe(h,i):null;return{text:c,modelId:m,modelRevision:null,responseId:g,inputTokens:y.promptTokens,outputTokens:y.completionTokens,costUsd:S}}}},async cancel({run:s}){return!s.judgeCalls?.some(a=>a.status==="reserved")}}}d(Ds,"createLearningHostedJudgeProvider");function Os(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Os,"record");function Hs(e,t){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new Error(`Judge catalog is missing its ${t}.`);return e}d(Hs,"positiveInteger");function xi(e,t=Ds()){return{execute(r){if(r.definition.execution.environment.entrypoint==="openpond.javascript-environment.v1")throw new Error("Tool Tasksets require grading an owner-recorded environment attempt.");let s=Ar({repository:e,scope:r.scope,run:r.run});return _r({binding:r.binding,rewards:r.rewards,task:dt(r.evidence,r.definition),evidence:Cr(r.evidence,r.run.output),evaluatorContext:r.evidence.submission.evaluatorContext,signal:r.signal,modelJudge:xr({readRubric:d(a=>e.transaction(r.scope,async o=>oe(await ke(o,"asset",a.id,1),a)),"readRubric"),async executeBudgeted(a,o){let n=await t.prepare(a,{scope:r.scope,run:r.run});return Ir({store:s,maximumCharge:d(()=>n.maximumChargeUsd,"maximumCharge"),dispatch:d((u,f)=>n.dispatch(f),"dispatch")})(`grade-${k(a)}`,a,o)}}),customVerifier:d(async({grader:a,task:o,evidence:n})=>{let i=await Ye({repository:e,scope:r.scope,grader:a,task:o,evidence:n,evaluatorContext:r.evidence.submission.evaluatorContext,timeoutMs:r.run.timeoutMs,signal:r.signal});return{score:i.score,passed:i.passed,rewardEligible:a.rewardEligible,failureClass:null,feedback:[i.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:i.evidenceRefs}},"customVerifier")})},async cancel({scope:r,run:s}){let a=await e.transaction(r,o=>ke(o,"grade",s.id));return a.judgeCalls?.some(o=>o.status==="reserved")?!1:t.cancel({scope:r,run:a})}}}d(xi,"createLocalTaskGradeExecutor");async function Ye(e){let t=await e.repository.transaction(e.scope,async r=>oe(await ke(r,"asset",e.grader.verifierRef.id,1),e.grader.verifierRef));return $e({source:t,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:Math.min(e.grader.timeoutMs,e.timeoutMs??e.grader.timeoutMs),signal:e.signal,value:{task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,expectedOutput:e.task.expectedOutput,evaluatorContext:e.evaluatorContext,infrastructureError:e.evidence.infrastructureError??null}})}d(Ye,"executeLocalLearningVerifier");import ia from"node:path";async function js(e,t,r){if(!r)throw new Error("Ordinary tool execution requires its owner's package directory.");let s=De({taskset:e,adapterId:qe(e)}),a=await vt({root:ia.join(r,"training","tasksets",ee(e)),content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:s.tasksetRelease,environment:s.environmentRelease,verifierSet:s.verifierSetRelease},sources:ct.parse(e.metadata.portableFileInventory)}),o=Wr(a);if(!o)throw new Error("The Taskset does not declare a JavaScript execution graph.");let n=a.taskset.tasks.find(f=>f.id===t.id);if(!n)throw new Error("The selected task is missing from its executable package.");let i=o.assets.find(f=>f.id===o.execution.javascript.module.id),u=o.assets.find(f=>f.id===n.privilegedContextRef);return{execution:o.execution,module:i,initialState:JSON.parse(u.text),definition:L(a.taskset),instructions:Fr(a),task:n}}d(js,"loadOrdinaryToolEnvironment");var J="openpond.javascript-environment.v1";async function Be(e,t,r,s){let a=t.tasks.find(o=>o.id===r.id);if(!a||k(a)!==k(r))throw new Error("Tool task differs from the selected immutable Taskset.");return t.metadata.taskDefinition===void 0?js(t,r,s):e.learningRepository().transaction(t.profileId,async o=>{let n=await ye(o,"definition",ae.parse(t.metadata.taskDefinition)),i=n.execution.environment;if(i.entrypoint!==J||t.environment.entrypoint!==J||t.environment.kind!=="agent"||t.environment.stateful!==i.stateful||t.environment.defaultTimeoutMs!==i.defaultTimeoutMs||t.environment.networkPolicy!==i.networkPolicy||t.environment.deterministicSeeds!==i.deterministicSeeds||k(t.environment.toolNames)!==k(n.execution.tools.map(v=>v.name))||k(t.environment.metadata.portableTools)!==k(n.execution.tools)||k(t.policy)!==k(n.execution.policy)||k(t.metadata.rewardBinding)!==k(n.rewardBinding))throw new Error("Tool Taskset context differs from its pinned task definition.");let u=await ye(o,"binding",n.rewardBinding),f=await Promise.all(u.sources.map(v=>ye(o,"reward",v.reward))),p=it.parse({id:r.id,clusterKey:r.clusterKey,split:r.split,input:r.input,expectedOutput:r.expectedOutput,policyVisibleContext:r.policyVisibleContext,privilegedContextRef:r.privilegedContextRef,artifactRefs:[],tags:r.tags}),l={...n.execution,graders:be(u,f),tasks:[p]},c=await ke(o,"asset",Lr(l),1),m=qr.parse(JSON.parse(oe(c,c.asset))),g=[m.javascript.module,m.environment.actionSchemaRef,m.environment.observationSchemaRef,m.environment.stateSchemaRef].filter(v=>v!==null),h=[...new Set([...g.map(v=>v.id),r.privilegedContextRef].filter(v=>v!==null))],w=await Promise.all(h.map(v=>ke(o,"asset",v,1))),y=Ur(l,c,w);if(k(t.environment.metadata.portableExecutionResources)!==k({environment:y.environment,verifierSet:y.verifierSet}))throw new Error("Tool Taskset publication differs from its private execution resources.");let S=w.find(v=>v.id===y.javascript.module.id),_=w.find(v=>v.id===r.privilegedContextRef),x=JSON.parse(oe(_,_.asset));return{execution:y,module:S,initialState:x,definition:L(n),instructions:n.instructions,task:p}})}d(Be,"loadStarterToolEnvironment");function Ns(e){let t=Wt({text:e.instructions,path:"policy/instructions.txt",mediaType:"text/plain",visibility:"policy"}),r=Wt({text:JSON.stringify({environment:L(e.execution.environment),verifierSet:L(e.execution.verifierSet),javascript:L(e.execution.javascript)}),path:"environment/dependencies.json",mediaType:"application/json",visibility:"host_private"}),s=br({schemaVersion:"openpond.agentSnapshot.v2",id:`starter-agent-${e.definition.contentHash}`,sourceRelease:{id:e.definition.id,contentHash:e.definition.contentHash},instructions:[t.asset],skills:[],agents:[],toolDeclarations:e.execution.javascript.tools,capabilityRequirements:[],dependencyLock:r.asset,portability:{portable:!0,blockers:[],localOnlyAssetRefs:[],hostPrivateAssetRefs:[e.module.id,r.id]},metadata:{execution:J}}),a=Rr({schemaVersion:"openpond.harnessRelease.v2",id:`starter-harness-${s.contentHash}`,agentSnapshot:{id:s.id,contentHash:s.contentHash},program:e.module.asset,tools:e.execution.javascript.tools,lifecycle:{create:!0,reset:!0,step:!0,collect:!0,destroy:!0,resetScope:"attempt"},graderInterface:{visibleEvidence:["output"],privilegedEvidence:["environment_state","private_verifier"],privateVerifierIsolation:!0},files:[t.asset,r.asset,e.module.asset],metadata:{execution:J}});return{agentSnapshot:s,harnessRelease:a,instructionContext:void 0}}d(Ns,"compileStarterToolHarness");import Rt from"node:path";import{constants as qs}from"node:fs";import{open as da,realpath as $s}from"node:fs/promises";var ca=P.string().regex(/^[a-f0-9]{64}$/),la=P.object({schemaVersion:P.literal("openpond.starterToolAttempt.v1"),requestId:P.string(),taskset:ae,taskHash:ca,model:P.unknown(),seed:P.number().int(),attempt:P.number().int(),definition:ae,policySource:P.enum(["model","fixture"]),environment:ae,verifierSet:ae,result:Nr}).strict();async function _t(e){let{store:t,taskset:r,task:s,attempt:a}=e,o=await Be(t,r,s,e.storeDir),i=(await t.listTaskAttemptArtifacts({attemptId:a.id})).filter(S=>S.kind==="environment_state"&&S.tasksetId===r.id&&S.taskId===s.id&&S.attemptId===a.id&&a.artifactRefs.includes(S.id)&&a.privilegedOutcomeRef===S.id);if(i.length!==1)throw new Error("Tool grading requires exactly one owner-recorded environment artifact.");let u=i[0],f=await $s(Rt.join(e.storeDir,"training","evaluation-artifacts")),p=await $s(Rt.dirname(u.path)),l=Rt.relative(f,p);if(!l||l.startsWith("..")||Rt.isAbsolute(l))throw new Error("Tool evidence is outside the execution owner's artifact directory.");let c=await da(u.path,qs.O_RDONLY|qs.O_NOFOLLOW),m;try{let S=await c.stat();if(!S.isFile()||S.size>33554432||S.size!==u.sizeBytes)throw new Error("Tool evidence exceeds its declared byte boundary.");m=await c.readFile()}finally{await c.close()}if(m.byteLength!==u.sizeBytes||se(m)!==u.sha256)throw new Error("Owner-recorded tool evidence bytes changed.");let g=la.parse(JSON.parse(m.toString("utf8"))),{contentHash:h,...w}=g.result,y=g.result;if(k(w)!==h||!y.snapshot||!y.environmentCleanupComplete||!["completed","budget_exhausted","policy_failure"].includes(y.status)||k(g.taskset)!==k(L(r))||g.taskHash!==k(s)||k(g.model)!==k(a.modelRef)||g.seed!==a.seed||g.attempt!==a.attempt||a.tasksetId!==r.id||a.taskId!==s.id||a.split!==s.split||k(g.definition)!==k(o.definition)||k(g.environment)!==k(L(o.execution.environment))||k(g.verifierSet)!==k(L(o.execution.verifierSet))||y.taskId!==s.id||y.snapshot.seed!==a.seed||y.snapshot.inputHash!==k(s.input)||k(y.snapshot.definition)!==k(L(o.execution.javascript))||y.snapshot.finalStateHash!==k(y.snapshot.state)||k(a.output)!==k({text:y.output??""})||a.metadata.environmentAttemptHash!==h)throw new Error("Tool evidence differs from the selected task, model or recorded attempt.");if(a.metadata.policySource!==g.policySource)throw new Error("Tool evidence policy source changed.");return{environment:{status:y.status,collected:y.collected,definition:y.snapshot.definition,initialStateHash:y.snapshot.initialStateHash,finalStateHash:y.snapshot.finalStateHash,finalState:y.snapshot.state,events:y.snapshot.events}}}d(_t,"readStarterToolEvidence");async function Te(e,t){if(t.metadata.learning!==void 0)return He.parse(t.metadata.learning);if(t.metadata.rewardBinding===void 0)return;let r=ae.parse(t.metadata.rewardBinding);return e.learningRepository().transaction(t.profileId,async s=>{let a=await ye(s,"binding",r),o=[...new Map(a.sources.map(i=>[`${i.reward.id}:${i.reward.revision}:${i.reward.contentHash}`,i.reward])).values()],n=await Promise.all(o.map(i=>ye(s,"reward",i)));return{binding:a,rewards:n}})}d(Te,"resolveTasksetRewardBinding");async function ua(e,t,r){let s=await Te(e,t);if(!s)return{rewardExecution:void 0,verifierAssets:[]};let a=[...s.rewards.flatMap(i=>i.assets),...be(s.binding,s.rewards).flatMap(i=>i.kind==="custom_verifier"?[i.verifierRef]:i.kind==="model_judge"?[i.rubricRef]:[])],o=t.metadata.learning===void 0?void 0:await wt(r,t);if(o){let i=[...new Map(a.map(u=>{let f=o.learningResources.assets.find(p=>p.id===u.id);if(!f)throw new Error("Imported batch verifier is missing from its package.");return oe(f,u),[f.id,f]})).values()];return{rewardExecution:s,verifierAssets:i}}let n=await e.learningRepository().transaction(t.profileId,async i=>{let u=new Map;for(let f of a){let p=await ke(i,"asset",f.id,1);oe(p,f),u.set(p.id,p)}return[...u.values()]});return{rewardExecution:s,verifierAssets:n}}d(ua,"resolveTasksetTrainingReward");async function gd(e,t,r){let s=await ua(e,t,r.storeDir);if(!s.rewardExecution)return s;let a=be(s.rewardExecution.binding,s.rewardExecution.rewards);if(!s.rewardExecution.binding.sources.some(o=>o.role==="evaluation"&&o.weight>0))throw new Error("Managed training requires a positively weighted evaluation Reward source for retained validation.");if(a.some(o=>o.kind==="custom_verifier"&&o.runtime==="sandbox_process"))throw new Error("Managed preparation requires a qualified isolated sandbox for process verifiers.");if(r.placement!=="remote"||t.environment.kind==="work"||t.capabilities.requiresState||t.capabilities.requiresTools||a.some(o=>o.kind==="human"||o.kind==="model_judge"&&(!o.model||o.calibrationStatus!=="passed"))||t.tasks.some(o=>o.privilegedContextRef!==null)||!s.rewardExecution.binding.sources.some(o=>o.role==="training"&&o.weight>0)||r.hasLearnedPreferenceReward)throw new Error("This Reward binding requires an additional managed execution adapter.");return s}d(gd,"resolveManagedTasksetReward");async function Ls(e,t,r){let s=await Te(e,t);if(!s)throw new Error("The Taskset has no published Reward binding.");let a=be(s.binding,s.rewards);return async({grader:o,task:n,attempt:i,signal:u})=>{let f=t.tasks.find(m=>m.id===n.id);if(!f||k(f)!==k(n))throw new Error("Task differs from the selected immutable Taskset.");let p=null;if(t.environment.entrypoint===J){if(!r)throw new Error("Tool grading requires the execution owner's artifact directory.");p=await _t({store:e,storeDir:r,taskset:t,task:n,attempt:i})}else if(n.privilegedContextRef!==null)throw new Error("This authored Taskset requires a private evaluator-context resolver.");let l=a.find(m=>m.id===o.id);if(l?.kind!=="custom_verifier")throw new Error("The verifier is absent from the published Reward binding.");let c=it.parse({id:n.id,clusterKey:n.clusterKey,split:n.split,input:n.input,expectedOutput:n.expectedOutput,policyVisibleContext:n.policyVisibleContext,privilegedContextRef:n.privilegedContextRef,artifactRefs:[],tags:n.tags});return Ye({repository:e.learningRepository(),scope:t.profileId,grader:l,task:c,evidence:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},evaluatorContext:p,signal:u})}}d(Ls,"createTasksetBindingVerifier");async function Ie(e,t,r){let s=r?ae.parse(r):null;if(s&&s.id!==t)throw new Error("Taskset execution reference differs from the requested Taskset.");let a=s?await e.getTasksetRevision(s.id,s.revision):await e.getTaskset(t);if(!a||s&&(a.contentHash!==s.contentHash||Re(a)!==s.contentHash))throw new Error("Taskset execution revision is unavailable or differs from its immutable hash.");return a}d(Ie,"requireLocalTasksetRevision");async function Sd(e,t,r){let s=await e.releaseForTaskset(t);return s||De({taskset:t,rewardExecution:r?await Te(r,t):void 0,adapterId:qe(t)}).tasksetRelease}d(Sd,"requireReleasedTaskset");import{createHash as ma}from"node:crypto";function Us(e,t,r){if(t.metadata.importedPackageHash!==e.contentHash)throw new Error("The Work projection differs from the admitted published package.");let s=e.taskset.tasks.find(n=>n.id===r.id);if(!s||s.split!==r.split||k(r.metadata.portableTaskRecord)!==k(s))throw new Error("The Work task differs from its exact released population.");let a=0,o=(r.assets??[]).map(n=>{let i=s.artifactRefs.find(l=>l.id===n.id&&l.contentHash===n.sha256&&l.path===n.artifactRef),u=i&&e.files.find(l=>k(l.asset)===k(i));if(!i||!u||i.visibility!=="policy")throw new Error("The Work task requests an unavailable or private evaluator file.");let f=Buffer.from(Br(u));if(ma("sha256").update(f).digest("hex")!==n.sha256||f.length!==n.sizeBytes||n.split!==r.split)throw new Error("The published Work input manifest changed.");if(a+=f.length,a>25e7)throw new Error("Published Work inputs exceed the admitted 250 MB ceiling.");return{...n,assetId:n.id,storageName:n.stagingPath??n.fileName,bytes:f,sizeBytes:f.length}});if(o.length!==s.artifactRefs.length)throw new Error("The exact policy-visible Work asset population is incomplete.");return o}d(Us,"advancedPublishedWorkAssets");async function Yt(e,t=pt,r=!1){if(e.kind==="hosted_harness")return Yt({kind:"hosted_chat",modelId:e.modelId,...Yr(e)},t);if(e.kind!=="hosted_chat")throw new ne("local_target_not_qualified","Select an explicit model target for local execution.",422);if(e.harness&&!r)throw new ne("local_harness_not_qualified","This released Harness requires its native turn executor; it cannot run as a model-only case.",422);let s=await t();if(s.error)throw new ne("local_model_catalog_unavailable",s.error,422);let a=s.models.find(c=>c.id===e.modelId);if(!a)throw new ne("local_model_unavailable","The selected model is unavailable.",422);let o=Bs(a.raw),n=bt(o),i=Bs(o.capabilities).samplingParameters===!0;if(!i&&(e.temperature!==0||e.topP!==1))throw new ne("local_sampling_unsupported","This model uses provider-managed sampling. Custom temperature or top P requires an explicit supported capability.",422);let u=Ws(o.context_window),f=Ws(o.output_limit);if(e.maxOutputTokens>f)throw new ne("local_model_output_limit","The model cannot accept the configured output limit.",422);let p={providerId:"openpond",modelId:e.modelId,maxOutputTokens:e.maxOutputTokens,...i?{temperature:e.temperature,topP:e.topP}:{},messages:e.messages??[]},l=(u*Math.max(n.inputUsdPerMillionTokens,n.cachedInputUsdPerMillionTokens)+e.maxOutputTokens*n.outputUsdPerMillionTokens)/1e6;if(!(l>0))throw new ne("local_model_charge_unknown","The model must publish an enforceable positive request ceiling.",422);return{model:{...p,configurationHash:k({configuration:p,contextWindow:u,outputLimit:f,pricing:n})},pricing:n,maximumChargeUsd:l}}d(Yt,"prepareLocalExperimentModel");function Nd(e){let t=e.admission.streamFactory?.(e)??e.admission.stream??e.stream??ft;return async function*(r){if(!r.requestId)throw new ne("local_request_identity_missing","Every local model dispatch requires a stable request identity.");let s=e.admission.model;if(r.model!==s.modelId||r.maxTokens!==s.maxOutputTokens||r.temperature!==s.temperature||r.topP!==s.topP)throw new ne("local_model_configuration_conflict","The native model request changed its admitted model, output limit or sampling configuration.",422);let a=xs();a?.assertExecution(e.executionId);let o={teamId:e.teamId,id:e.executionId,ownerId:e.ownerId,requestId:r.requestId};await e.store.reserveLocalCharge({...o,caseId:e.caseId,maximumUsd:e.admission.maximumChargeUsd}),r.signal?.aborted&&(await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),r.signal.throwIfAborted());try{await e.store.markLocalChargeDispatched(o.teamId,o.id,o.requestId,o.ownerId)}catch(p){throw a&&await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),p}await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:"model.dispatch",payload:{requestId:o.requestId,model:r.model,messages:r.messages}});let n=null,i=!1,u=!1,f;try{for await(let p of t(r)){u=!0,p.type==="usage"&&(n=p.usage),p.type==="finish"&&(i=!0);let l=p.type==="text_delta"?{requestId:o.requestId,text:p.text}:p.type==="tool_call_delta"?{requestId:o.requestId,toolCalls:p.toolCalls}:p.type==="usage"?{requestId:o.requestId,usage:ve(p.usage)}:p.type==="finish"?{requestId:o.requestId,finishReason:p.finishReason}:null;l&&await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:`model.${p.type}`,payload:l}),yield p}}catch(p){throw!u&&p instanceof Is&&p.executionId===o.id&&p.requestId===o.requestId&&(f=p),p}finally{let p=ve(n),l=e.admission.pricing&&i&&p.promptTokens!==null&&p.completionTokens!==null?xe(n,e.admission.pricing):null;await e.store.settleLocalCharge({...o,costUsd:l,confirmedTransportNotInvoked:f,usage:n===null?null:{promptTokens:p.promptTokens,completionTokens:p.completionTokens,cachedPromptTokens:p.cachedPromptTokens,uncachedPromptTokens:p.uncachedPromptTokens}})}}}d(Nd,"createLocalBudgetedModelStream");function Bs(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Bs,"record");function Ws(e){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new ne("local_model_bound_unknown","The model catalog is missing an enforceable context/output bound.",422);return e}d(Ws,"positive");import{AsyncLocalStorage as pa}from"node:async_hooks";import fa from"node:path";var Fs=new pa;function Vs(){return Fs.getStore()}d(Vs,"currentAdvancedRefinerBoundary");function Jd(e){let t=Mr(fa.join(e.storeDir,"library","harnesses","advanced-refiner-paid-calls.sqlite"));t.exec(`CREATE TABLE IF NOT EXISTS advanced_refiner_budgets(id TEXT PRIMARY KEY,actor_id TEXT NOT NULL,team_id TEXT NOT NULL,pin_hash TEXT NOT NULL,maximum_spend REAL NOT NULL,maximum_steps INTEGER NOT NULL,expires_at TEXT NOT NULL);
+import{a as As,b as Ms}from"./chunk-ATFV2NA4.js";import{a as Es,e as $e,f as Cs,g as Tt}from"./chunk-WGISITX6.js";import{a as xs,c as Is}from"./chunk-XVDFJ3ZF.js";import{h as ne}from"./chunk-R5GRESW3.js";import{a as ee,b as _s}from"./chunk-K3LS5SK7.js";import{b as ys,c as bs,d as yt}from"./chunk-CWAQOFLP.js";import{a as ws,b as vs,c as Ts,d as qe,e as Ss,g as vt}from"./chunk-B7XJBP4U.js";import{a as wt,c as Rs}from"./chunk-FVK4W27G.js";import{b as Kr,e as Xr,f as lt,g as ut}from"./chunk-YONDU5TN.js";import{F as gt,G as fs,H as ht,I as kt,J as gs,K as hs,N as ks,x as ps}from"./chunk-4VZJQRQI.js";import{g as pt,i as ft}from"./chunk-MCBRB5D6.js";import{a as ve}from"./chunk-7KWD2I6W.js";import{s as Pr}from"./chunk-4TTXNVCE.js";import{d as Ne}from"./chunk-BJXPQIVD.js";import{ja as rs,qa as ss,sa as os,x as ts}from"./chunk-CJDGLPNX.js";import{M as V,N as as,O as ns,P as is,Q as mt,R as ds,S as q,T as cs,U as je,V as ls,X as us,Y as Vt,Z as ms}from"./chunk-FCXDPVGJ.js";import{l as Ft}from"./chunk-5TPYBXWM.js";import{d as Zr,f as W,g as Qr,i as es}from"./chunk-6LFTQBXP.js";import{f as ct}from"./chunk-AXYAEWWA.js";import{$a as Yr,Ba as Jr,Ca as Re,K as Nr,L as qr,M as $r,N as Lr,Na as Gr,Oa as De,P as Ur,aa as Br,ba as Wr,ia as Fr,ua as Vr,wa as we,xa as zr}from"./chunk-5ZNE43XC.js";import{q as ae}from"./chunk-263WS3TB.js";import{E as Or,F as Hr,Tb as jr,xb as Dr}from"./chunk-QAFXQHHG.js";import{f as Sr}from"./chunk-AJXWJLZA.js";import{K as L,L as Er,Nb as Ar,Q as Wt,R as oe,W as dt,X as Cr,aa as He,d as xr,g as Ir,ob as ke,pb as ye}from"./chunk-XHU4BZQY.js";import{bb as be,cb as _r,ea as it,v as Ge}from"./chunk-ZY32SZ2P.js";import{l as br,m as Rr}from"./chunk-OTDDXEQ6.js";import{i as Tr,m as se,n as k}from"./chunk-RCT3O3HQ.js";import{A as P}from"./chunk-X2J2P5ZT.js";import{m as Mr}from"./chunk-UGOK7VTC.js";import{a as d}from"./chunk-FTQLQJQV.js";var zt={buildIntent:"verifiable_reward",methodHint:"grpo"};async function ii(e){let t=await aa(e.store,e.workspaceId,e.reviewRef),r=`harness-review-${k({review:t.contentHash,strategy:zt}).slice(0,24)}`,s=(await e.store.listTaskCreationSnapshots(e.profileId)).find(c=>c.request.candidateId===r);if(s)return s;let[a,o,n]=await Promise.all([e.store.listHarnessImprovementArtifacts(e.workspaceId,"route_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"trigger_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"observation",1e3)]),i=new Map(a.map(c=>[_e(c),c])),u=new Map(o.map(c=>[_e(c),c])),f=new Map(n.map(c=>[_e(c),c])),p=new Map;for(let c of t.selectedEvidence){if(c.kind!=="route_decision")continue;let m=i.get(_e(c.evidence)),g=m?u.get(_e(m.trigger)):null;if(!m||!g)throw new Error(`Harness review evidence ${c.evidence.id} cannot be reconstructed.`);let h=g.observations.map(y=>f.get(_e(y))??null).filter(y=>y?.runRef===c.sourceRef);if(!h.length)throw new Error(`Harness review source ${c.sourceRef} has no exact observation evidence.`);let w=p.get(c.sourceRef)??{turnIds:new Set,evidenceRefs:[],sourcePolicies:[]};for(let y of h)w.turnIds.add(y.turnId);w.evidenceRefs.push(c.evidence),w.sourcePolicies.push(c.sourcePolicy),p.set(c.sourceRef,w)}if(!p.size)throw new Error("Harness Taskset review has no reconstructable source runs.");let l=[];for(let[c,m]of[...p.entries()].sort(([g],[h])=>g.localeCompare(h))){let g=await e.taskCreator.addSessionSource({profileId:e.profileId,sessionId:c,turnIds:[...m.turnIds].sort(),consentScope:"selected_turns"}),h=Array.isArray(g.metadata.harnessEvaluationReviews)?g.metadata.harnessEvaluationReviews:[],w={...g,metadata:{...g.metadata,harnessEvaluationReviews:Gt([...h,{id:t.id,contentHash:t.contentHash,workspaceId:t.workspaceRef,harnessRelease:t.harnessRelease,claimFingerprint:t.claim?.fingerprint??null,evidenceRefs:m.evidenceRefs,sourcePolicies:m.sourcePolicies}])}};l.push(await e.store.upsertTrainingSource(w))}return e.startCreation({profileId:e.profileId,sourceIds:l.map(c=>c.id),surface:"task_candidate",mode:"customize",entryMode:"automated",resourceIntent:"workproduct",buildIntent:zt.buildIntent,buildSpecification:null,objective:t.claim?.statement??t.reason,methodHint:zt.methodHint,preferredBaseModelId:null,preferredBaseModel:null,candidateId:r,analysisModel:e.analysisModel??null,analysisReasoningEffort:e.analysisReasoningEffort??null,createImproveRunId:null,targetIntent:{kind:"model",id:null,displayName:null,operation:"create"}})}d(ii,"startHarnessReviewTasksetAuthoring");function di(e){let t=e.flatMap(o=>Array.isArray(o.metadata.harnessEvaluationReviews)?o.metadata.harnessEvaluationReviews.filter(Jt):[]);if(!t.length)return null;let r=`${String(t[0].id)}:${String(t[0].contentHash)}`,s=t.filter(o=>`${String(o.id)}:${String(o.contentHash)}`===r);if(s.length!==e.length)return null;let a=s[0];return{review:{id:String(a.id),contentHash:String(a.contentHash),workspaceId:String(a.workspaceId),harnessRelease:a.harnessRelease,claimFingerprint:typeof a.claimFingerprint=="string"?a.claimFingerprint:null},evidenceRefs:Gt(s.flatMap(o=>Array.isArray(o.evidenceRefs)?o.evidenceRefs:[])),sourcePolicies:Gt(s.flatMap(o=>Array.isArray(o.sourcePolicies)?o.sourcePolicies:[]))}}d(di,"harnessReviewLineageFromSources");async function Ps(e){let t=e.taskset.metadata.harnessEvaluationLineage;if(!Jt(t)||!Jt(t.review))return null;let r=String(t.review.workspaceId??""),s={id:String(t.review.id??""),contentHash:String(t.review.contentHash??"")};if(!r||!s.id||!s.contentHash)throw new Error("Taskset Harness Evaluation lineage is incomplete.");let a=await e.store.listHarnessImprovementArtifacts(r,"evaluation_review",1e3),o=a.find(c=>_e(c)===_e(s));if(!o)throw new Error("Originating Harness Evaluation review was not found.");let n=e.evaluationResult?{id:e.evaluationResult.id,contentHash:e.evaluationResult.contentHash}:null,i={id:e.taskset.id,contentHash:e.taskset.contentHash},u=k({originRef:s,tasksetRef:i,evaluationRef:n}),f=a.find(c=>c.metadata.downstreamLinkKey===u);if(f)return f;let p=(e.now??(()=>new Date().toISOString()))(),l=Sr({schemaVersion:"openpond.harnessEvaluationReviewReceipt.v1",id:`evaluation-review-link-${u.slice(0,24)}`,ownerScope:o.ownerScope,workspaceRef:o.workspaceRef,harnessRelease:o.harnessRelease,previousWatermark:o.nextWatermark,nextWatermark:{cursor:k({cursor:o.nextWatermark.cursor,tasksetRef:i,evaluationRef:n}),throughCreatedAt:p},selectedEvidence:o.selectedEvidence,excludedEvidence:o.excludedEvidence,claim:o.claim,classification:"taskset",triage:o.triage,reason:n?"The approved Taskset completed a real baseline Evaluation.":"The approved Harness review materialized one immutable Taskset.",nextAuthority:"human_review",maxEstimatedCostUsd:o.maxEstimatedCostUsd,tasksetProposal:i,evaluation:n,trainingQualification:null,policyVersion:o.policyVersion,createdAt:p,metadata:{originReview:s,downstreamLinkKey:u,tasksetRevision:e.taskset.revision}});return await e.store.saveHarnessImprovementArtifact(r,"evaluation_review",l),l}d(Ps,"linkHarnessReviewTaskset");async function aa(e,t,r){let s=await e.getHarnessWorkspace(t);if(!s)throw new Error(`Harness workspace ${t} was not found.`);let a=(await e.listHarnessImprovementArtifacts(t,"evaluation_review",1e3)).find(n=>n.id===r.id&&n.contentHash===r.contentHash);if(!a||a.schemaVersion!=="openpond.harnessEvaluationReviewReceipt.v1")throw new Error("Harness Evaluation review was not found.");let o=a;if(o.classification!=="taskset"||o.nextAuthority!=="human_review")throw new Error("Only a human-review Taskset recommendation can enter Taskset authoring.");if(o.tasksetProposal)throw new Error("Harness Evaluation review already has a materialized Taskset link.");if(o.ownerScope.kind!==s.ownerScope.kind||o.ownerScope.id!==s.ownerScope.id)throw new Error("Harness Evaluation review owner scope does not match its workspace.");return o}d(aa,"requireTasksetReview");function _e(e){return`${e.id}:${e.contentHash}`}d(_e,"artifactKey");function Jt(e){return!!e&&typeof e=="object"&&!Array.isArray(e)}d(Jt,"isRecord");function Gt(e){let t=new Set;return e.filter(r=>{let s=JSON.stringify(r);return t.has(s)?!1:(t.add(s),!0)})}d(Gt,"uniqueObjects");import{randomUUID as na}from"node:crypto";function bt(e){let t=St(St(St(e.metadata).billing).pricing);return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}d(bt,"hostedTokenPricingFromCatalog");function mi(e){let t=St(e);try{return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}catch{return null}}d(mi,"hostedTokenPricingFromValue");function xe(e,t){let r=ve(e);if(r.promptTokens===null&&r.completionTokens===null)return null;let s=r.cachedPromptTokens??0;return((r.uncachedPromptTokens??r.promptTokens??0)*t.inputUsdPerMillionTokens+s*t.cachedInputUsdPerMillionTokens+(r.completionTokens??0)*t.outputUsdPerMillionTokens)/1e6}d(xe,"hostedUsageCostUsd");function St(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(St,"record");function Le(e,t){if(typeof e=="string"&&e.trim())return e.trim();throw new Error(`Hosted model catalog is missing ${t}.`)}d(Le,"requiredString");function Ue(e,t){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return e;throw new Error(`Hosted model catalog is missing ${t}.`)}d(Ue,"requiredRate");function Ds(e={}){let t=e.catalog??pt,r=e.stream??ft;return{async prepare(s){if(s.providerId!=="openpond")throw new Error("The selected judge provider is not connected to the hosted judge transport.");if(s.revision!==null)throw new Error("This judge transport cannot enforce a model revision; select a provider with revision support.");let a=await t();if(a.error)throw new Error(`Judge model catalog unavailable: ${a.error}`);let o=a.models.find(l=>l.id===s.modelId);if(!o)throw new Error("The selected judge model is unavailable.");let n=Os(o.raw),i=bt(n),u=Hs(n.context_window,"context window"),f=Math.min(4096,Hs(n.output_limit,"output limit")),p=(u*Math.max(i.inputUsdPerMillionTokens,i.cachedInputUsdPerMillionTokens)+f*i.outputUsdPerMillionTokens)/1e6;if(!(p>0))throw new Error("The judge model must have a positive known charge ceiling.");return{maximumChargeUsd:p,async dispatch(l){l?.throwIfAborted();let c="",m=null,g=null,h=null,w=!1;for await(let _ of r({model:s.modelId,messages:[{role:"system",content:s.system},{role:"user",content:s.data}],requestId:`learning-judge-${na()}`,maxTokens:f,temperature:s.temperature,signal:l})){let x=Os(_.raw);if(typeof x.model=="string"){if(m!==null&&x.model!==m)throw new Error("Judge response model identity changed during streaming.");m=x.model}if(typeof x.id=="string"){if(g!==null&&g!==x.id)throw new Error("Judge response identity changed during streaming.");g=x.id}if(_.type==="text_delta"&&(c+=_.text),c.length>1e5)throw new Error("Judge response exceeds the retained response limit.");_.type==="usage"&&(h=_.usage),_.type==="finish"&&(w=!0)}if(!w||!m||!g)throw new Error("Judge response did not include a complete model receipt.");let y=ve(h),S=y.promptTokens!==null&&y.completionTokens!==null?xe(h,i):null;return{text:c,modelId:m,modelRevision:null,responseId:g,inputTokens:y.promptTokens,outputTokens:y.completionTokens,costUsd:S}}}},async cancel({run:s}){return!s.judgeCalls?.some(a=>a.status==="reserved")}}}d(Ds,"createLearningHostedJudgeProvider");function Os(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Os,"record");function Hs(e,t){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new Error(`Judge catalog is missing its ${t}.`);return e}d(Hs,"positiveInteger");function xi(e,t=Ds()){return{execute(r){if(r.definition.execution.environment.entrypoint==="openpond.javascript-environment.v1")throw new Error("Tool Tasksets require grading an owner-recorded environment attempt.");let s=Ar({repository:e,scope:r.scope,run:r.run});return _r({binding:r.binding,rewards:r.rewards,task:dt(r.evidence,r.definition),evidence:Cr(r.evidence,r.run.output),evaluatorContext:r.evidence.submission.evaluatorContext,signal:r.signal,modelJudge:xr({readRubric:d(a=>e.transaction(r.scope,async o=>oe(await ke(o,"asset",a.id,1),a)),"readRubric"),async executeBudgeted(a,o){let n=await t.prepare(a,{scope:r.scope,run:r.run});return Ir({store:s,maximumCharge:d(()=>n.maximumChargeUsd,"maximumCharge"),dispatch:d((u,f)=>n.dispatch(f),"dispatch")})(`grade-${k(a)}`,a,o)}}),customVerifier:d(async({grader:a,task:o,evidence:n})=>{let i=await Ye({repository:e,scope:r.scope,grader:a,task:o,evidence:n,evaluatorContext:r.evidence.submission.evaluatorContext,timeoutMs:r.run.timeoutMs,signal:r.signal});return{score:i.score,passed:i.passed,rewardEligible:a.rewardEligible,failureClass:null,feedback:[i.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:i.evidenceRefs}},"customVerifier")})},async cancel({scope:r,run:s}){let a=await e.transaction(r,o=>ke(o,"grade",s.id));return a.judgeCalls?.some(o=>o.status==="reserved")?!1:t.cancel({scope:r,run:a})}}}d(xi,"createLocalTaskGradeExecutor");async function Ye(e){let t=await e.repository.transaction(e.scope,async r=>oe(await ke(r,"asset",e.grader.verifierRef.id,1),e.grader.verifierRef));return $e({source:t,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:Math.min(e.grader.timeoutMs,e.timeoutMs??e.grader.timeoutMs),signal:e.signal,value:{task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,expectedOutput:e.task.expectedOutput,evaluatorContext:e.evaluatorContext,infrastructureError:e.evidence.infrastructureError??null}})}d(Ye,"executeLocalLearningVerifier");import ia from"node:path";async function js(e,t,r){if(!r)throw new Error("Ordinary tool execution requires its owner's package directory.");let s=De({taskset:e,adapterId:qe(e)}),a=await vt({root:ia.join(r,"training","tasksets",ee(e)),content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:s.tasksetRelease,environment:s.environmentRelease,verifierSet:s.verifierSetRelease},sources:ct.parse(e.metadata.portableFileInventory)}),o=Wr(a);if(!o)throw new Error("The Taskset does not declare a JavaScript execution graph.");let n=a.taskset.tasks.find(f=>f.id===t.id);if(!n)throw new Error("The selected task is missing from its executable package.");let i=o.assets.find(f=>f.id===o.execution.javascript.module.id),u=o.assets.find(f=>f.id===n.privilegedContextRef);return{execution:o.execution,module:i,initialState:JSON.parse(u.text),definition:L(a.taskset),instructions:Fr(a),task:n}}d(js,"loadOrdinaryToolEnvironment");var J="openpond.javascript-environment.v1";async function Be(e,t,r,s){let a=t.tasks.find(o=>o.id===r.id);if(!a||k(a)!==k(r))throw new Error("Tool task differs from the selected immutable Taskset.");return t.metadata.taskDefinition===void 0?js(t,r,s):e.learningRepository().transaction(t.profileId,async o=>{let n=await ye(o,"definition",ae.parse(t.metadata.taskDefinition)),i=n.execution.environment;if(i.entrypoint!==J||t.environment.entrypoint!==J||t.environment.kind!=="agent"||t.environment.stateful!==i.stateful||t.environment.defaultTimeoutMs!==i.defaultTimeoutMs||t.environment.networkPolicy!==i.networkPolicy||t.environment.deterministicSeeds!==i.deterministicSeeds||k(t.environment.toolNames)!==k(n.execution.tools.map(v=>v.name))||k(t.environment.metadata.portableTools)!==k(n.execution.tools)||k(t.policy)!==k(n.execution.policy)||k(t.metadata.rewardBinding)!==k(n.rewardBinding))throw new Error("Tool Taskset context differs from its pinned task definition.");let u=await ye(o,"binding",n.rewardBinding),f=await Promise.all(u.sources.map(v=>ye(o,"reward",v.reward))),p=it.parse({id:r.id,clusterKey:r.clusterKey,split:r.split,input:r.input,expectedOutput:r.expectedOutput,policyVisibleContext:r.policyVisibleContext,privilegedContextRef:r.privilegedContextRef,artifactRefs:[],tags:r.tags}),l={...n.execution,graders:be(u,f),tasks:[p]},c=await ke(o,"asset",Lr(l),1),m=qr.parse(JSON.parse(oe(c,c.asset))),g=[m.javascript.module,m.environment.actionSchemaRef,m.environment.observationSchemaRef,m.environment.stateSchemaRef].filter(v=>v!==null),h=[...new Set([...g.map(v=>v.id),r.privilegedContextRef].filter(v=>v!==null))],w=await Promise.all(h.map(v=>ke(o,"asset",v,1))),y=Ur(l,c,w);if(k(t.environment.metadata.portableExecutionResources)!==k({environment:y.environment,verifierSet:y.verifierSet}))throw new Error("Tool Taskset publication differs from its private execution resources.");let S=w.find(v=>v.id===y.javascript.module.id),_=w.find(v=>v.id===r.privilegedContextRef),x=JSON.parse(oe(_,_.asset));return{execution:y,module:S,initialState:x,definition:L(n),instructions:n.instructions,task:p}})}d(Be,"loadStarterToolEnvironment");function Ns(e){let t=Wt({text:e.instructions,path:"policy/instructions.txt",mediaType:"text/plain",visibility:"policy"}),r=Wt({text:JSON.stringify({environment:L(e.execution.environment),verifierSet:L(e.execution.verifierSet),javascript:L(e.execution.javascript)}),path:"environment/dependencies.json",mediaType:"application/json",visibility:"host_private"}),s=br({schemaVersion:"openpond.agentSnapshot.v2",id:`starter-agent-${e.definition.contentHash}`,sourceRelease:{id:e.definition.id,contentHash:e.definition.contentHash},instructions:[t.asset],skills:[],agents:[],toolDeclarations:e.execution.javascript.tools,capabilityRequirements:[],dependencyLock:r.asset,portability:{portable:!0,blockers:[],localOnlyAssetRefs:[],hostPrivateAssetRefs:[e.module.id,r.id]},metadata:{execution:J}}),a=Rr({schemaVersion:"openpond.harnessRelease.v2",id:`starter-harness-${s.contentHash}`,agentSnapshot:{id:s.id,contentHash:s.contentHash},program:e.module.asset,tools:e.execution.javascript.tools,lifecycle:{create:!0,reset:!0,step:!0,collect:!0,destroy:!0,resetScope:"attempt"},graderInterface:{visibleEvidence:["output"],privilegedEvidence:["environment_state","private_verifier"],privateVerifierIsolation:!0},files:[t.asset,r.asset,e.module.asset],metadata:{execution:J}});return{agentSnapshot:s,harnessRelease:a,instructionContext:void 0}}d(Ns,"compileStarterToolHarness");import Rt from"node:path";import{constants as qs}from"node:fs";import{open as da,realpath as $s}from"node:fs/promises";var ca=P.string().regex(/^[a-f0-9]{64}$/),la=P.object({schemaVersion:P.literal("openpond.starterToolAttempt.v1"),requestId:P.string(),taskset:ae,taskHash:ca,model:P.unknown(),seed:P.number().int(),attempt:P.number().int(),definition:ae,policySource:P.enum(["model","fixture"]),environment:ae,verifierSet:ae,result:Nr}).strict();async function _t(e){let{store:t,taskset:r,task:s,attempt:a}=e,o=await Be(t,r,s,e.storeDir),i=(await t.listTaskAttemptArtifacts({attemptId:a.id})).filter(S=>S.kind==="environment_state"&&S.tasksetId===r.id&&S.taskId===s.id&&S.attemptId===a.id&&a.artifactRefs.includes(S.id)&&a.privilegedOutcomeRef===S.id);if(i.length!==1)throw new Error("Tool grading requires exactly one owner-recorded environment artifact.");let u=i[0],f=await $s(Rt.join(e.storeDir,"training","evaluation-artifacts")),p=await $s(Rt.dirname(u.path)),l=Rt.relative(f,p);if(!l||l.startsWith("..")||Rt.isAbsolute(l))throw new Error("Tool evidence is outside the execution owner's artifact directory.");let c=await da(u.path,qs.O_RDONLY|qs.O_NOFOLLOW),m;try{let S=await c.stat();if(!S.isFile()||S.size>33554432||S.size!==u.sizeBytes)throw new Error("Tool evidence exceeds its declared byte boundary.");m=await c.readFile()}finally{await c.close()}if(m.byteLength!==u.sizeBytes||se(m)!==u.sha256)throw new Error("Owner-recorded tool evidence bytes changed.");let g=la.parse(JSON.parse(m.toString("utf8"))),{contentHash:h,...w}=g.result,y=g.result;if(k(w)!==h||!y.snapshot||!y.environmentCleanupComplete||!["completed","budget_exhausted","policy_failure"].includes(y.status)||k(g.taskset)!==k(L(r))||g.taskHash!==k(s)||k(g.model)!==k(a.modelRef)||g.seed!==a.seed||g.attempt!==a.attempt||a.tasksetId!==r.id||a.taskId!==s.id||a.split!==s.split||k(g.definition)!==k(o.definition)||k(g.environment)!==k(L(o.execution.environment))||k(g.verifierSet)!==k(L(o.execution.verifierSet))||y.taskId!==s.id||y.snapshot.seed!==a.seed||y.snapshot.inputHash!==k(s.input)||k(y.snapshot.definition)!==k(L(o.execution.javascript))||y.snapshot.finalStateHash!==k(y.snapshot.state)||k(a.output)!==k({text:y.output??""})||a.metadata.environmentAttemptHash!==h)throw new Error("Tool evidence differs from the selected task, model or recorded attempt.");if(a.metadata.policySource!==g.policySource)throw new Error("Tool evidence policy source changed.");return{environment:{status:y.status,collected:y.collected,definition:y.snapshot.definition,initialStateHash:y.snapshot.initialStateHash,finalStateHash:y.snapshot.finalStateHash,finalState:y.snapshot.state,events:y.snapshot.events}}}d(_t,"readStarterToolEvidence");async function Te(e,t){if(t.metadata.learning!==void 0)return He.parse(t.metadata.learning);if(t.metadata.rewardBinding===void 0)return;let r=ae.parse(t.metadata.rewardBinding);return e.learningRepository().transaction(t.profileId,async s=>{let a=await ye(s,"binding",r),o=[...new Map(a.sources.map(i=>[`${i.reward.id}:${i.reward.revision}:${i.reward.contentHash}`,i.reward])).values()],n=await Promise.all(o.map(i=>ye(s,"reward",i)));return{binding:a,rewards:n}})}d(Te,"resolveTasksetRewardBinding");async function ua(e,t,r){let s=await Te(e,t);if(!s)return{rewardExecution:void 0,verifierAssets:[]};let a=[...s.rewards.flatMap(i=>i.assets),...be(s.binding,s.rewards).flatMap(i=>i.kind==="custom_verifier"?[i.verifierRef]:i.kind==="model_judge"?[i.rubricRef]:[])],o=t.metadata.learning===void 0?void 0:await wt(r,t);if(o){let i=[...new Map(a.map(u=>{let f=o.learningResources.assets.find(p=>p.id===u.id);if(!f)throw new Error("Imported batch verifier is missing from its package.");return oe(f,u),[f.id,f]})).values()];return{rewardExecution:s,verifierAssets:i}}let n=await e.learningRepository().transaction(t.profileId,async i=>{let u=new Map;for(let f of a){let p=await ke(i,"asset",f.id,1);oe(p,f),u.set(p.id,p)}return[...u.values()]});return{rewardExecution:s,verifierAssets:n}}d(ua,"resolveTasksetTrainingReward");async function gd(e,t,r){let s=await ua(e,t,r.storeDir);if(!s.rewardExecution)return s;let a=be(s.rewardExecution.binding,s.rewardExecution.rewards);if(!s.rewardExecution.binding.sources.some(o=>o.role==="evaluation"&&o.weight>0))throw new Error("Managed training requires a positively weighted evaluation Reward source for retained validation.");if(a.some(o=>o.kind==="custom_verifier"&&o.runtime==="sandbox_process"))throw new Error("Managed preparation requires a qualified isolated sandbox for process verifiers.");if(r.placement!=="remote"||t.environment.kind==="work"||t.capabilities.requiresState||t.capabilities.requiresTools||a.some(o=>o.kind==="human"||o.kind==="model_judge"&&(!o.model||o.calibrationStatus!=="passed"))||t.tasks.some(o=>o.privilegedContextRef!==null)||!s.rewardExecution.binding.sources.some(o=>o.role==="training"&&o.weight>0)||r.hasLearnedPreferenceReward)throw new Error("This Reward binding requires an additional managed execution adapter.");return s}d(gd,"resolveManagedTasksetReward");async function Ls(e,t,r){let s=await Te(e,t);if(!s)throw new Error("The Taskset has no published Reward binding.");let a=be(s.binding,s.rewards);return async({grader:o,task:n,attempt:i,signal:u})=>{let f=t.tasks.find(m=>m.id===n.id);if(!f||k(f)!==k(n))throw new Error("Task differs from the selected immutable Taskset.");let p=null;if(t.environment.entrypoint===J){if(!r)throw new Error("Tool grading requires the execution owner's artifact directory.");p=await _t({store:e,storeDir:r,taskset:t,task:n,attempt:i})}else if(n.privilegedContextRef!==null)throw new Error("This authored Taskset requires a private evaluator-context resolver.");let l=a.find(m=>m.id===o.id);if(l?.kind!=="custom_verifier")throw new Error("The verifier is absent from the published Reward binding.");let c=it.parse({id:n.id,clusterKey:n.clusterKey,split:n.split,input:n.input,expectedOutput:n.expectedOutput,policyVisibleContext:n.policyVisibleContext,privilegedContextRef:n.privilegedContextRef,artifactRefs:[],tags:n.tags});return Ye({repository:e.learningRepository(),scope:t.profileId,grader:l,task:c,evidence:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},evaluatorContext:p,signal:u})}}d(Ls,"createTasksetBindingVerifier");async function Ie(e,t,r){let s=r?ae.parse(r):null;if(s&&s.id!==t)throw new Error("Taskset execution reference differs from the requested Taskset.");let a=s?await e.getTasksetRevision(s.id,s.revision):await e.getTaskset(t);if(!a||s&&(a.contentHash!==s.contentHash||Re(a)!==s.contentHash))throw new Error("Taskset execution revision is unavailable or differs from its immutable hash.");return a}d(Ie,"requireLocalTasksetRevision");async function Sd(e,t,r){let s=await e.releaseForTaskset(t);return s||De({taskset:t,rewardExecution:r?await Te(r,t):void 0,adapterId:qe(t)}).tasksetRelease}d(Sd,"requireReleasedTaskset");import{createHash as ma}from"node:crypto";function Us(e,t,r){if(t.metadata.importedPackageHash!==e.contentHash)throw new Error("The Work projection differs from the admitted published package.");let s=e.taskset.tasks.find(n=>n.id===r.id);if(!s||s.split!==r.split||k(r.metadata.portableTaskRecord)!==k(s))throw new Error("The Work task differs from its exact released population.");let a=0,o=(r.assets??[]).map(n=>{let i=s.artifactRefs.find(l=>l.id===n.id&&l.contentHash===n.sha256&&l.path===n.artifactRef),u=i&&e.files.find(l=>k(l.asset)===k(i));if(!i||!u||i.visibility!=="policy")throw new Error("The Work task requests an unavailable or private evaluator file.");let f=Buffer.from(Br(u));if(ma("sha256").update(f).digest("hex")!==n.sha256||f.length!==n.sizeBytes||n.split!==r.split)throw new Error("The published Work input manifest changed.");if(a+=f.length,a>25e7)throw new Error("Published Work inputs exceed the admitted 250 MB ceiling.");return{...n,assetId:n.id,storageName:n.stagingPath??n.fileName,bytes:f,sizeBytes:f.length}});if(o.length!==s.artifactRefs.length)throw new Error("The exact policy-visible Work asset population is incomplete.");return o}d(Us,"advancedPublishedWorkAssets");async function Yt(e,t=pt,r=!1){if(e.kind==="hosted_harness")return Yt({kind:"hosted_chat",modelId:e.modelId,...Yr(e)},t);if(e.kind!=="hosted_chat")throw new ne("local_target_not_qualified","Select an explicit model target for local execution.",422);if(e.harness&&!r)throw new ne("local_harness_not_qualified","This released Harness requires its native turn executor; it cannot run as a model-only case.",422);let s=await t();if(s.error)throw new ne("local_model_catalog_unavailable",s.error,422);let a=s.models.find(c=>c.id===e.modelId);if(!a)throw new ne("local_model_unavailable","The selected model is unavailable.",422);let o=Bs(a.raw),n=bt(o),i=Bs(o.capabilities).samplingParameters===!0;if(!i&&(e.temperature!==0||e.topP!==1))throw new ne("local_sampling_unsupported","This model uses provider-managed sampling. Custom temperature or top P requires an explicit supported capability.",422);let u=Ws(o.context_window),f=Ws(o.output_limit);if(e.maxOutputTokens>f)throw new ne("local_model_output_limit","The model cannot accept the configured output limit.",422);let p={providerId:"openpond",modelId:e.modelId,maxOutputTokens:e.maxOutputTokens,...i?{temperature:e.temperature,topP:e.topP}:{},messages:e.messages??[]},l=(u*Math.max(n.inputUsdPerMillionTokens,n.cachedInputUsdPerMillionTokens)+e.maxOutputTokens*n.outputUsdPerMillionTokens)/1e6;if(!(l>0))throw new ne("local_model_charge_unknown","The model must publish an enforceable positive request ceiling.",422);return{model:{...p,configurationHash:k({configuration:p,contextWindow:u,outputLimit:f,pricing:n})},pricing:n,maximumChargeUsd:l}}d(Yt,"prepareLocalExperimentModel");function Nd(e){let t=e.admission.streamFactory?.(e)??e.admission.stream??e.stream??ft;return async function*(r){if(!r.requestId)throw new ne("local_request_identity_missing","Every local model dispatch requires a stable request identity.");let s=e.admission.model;if(r.model!==s.modelId||r.maxTokens!==s.maxOutputTokens||r.temperature!==s.temperature||r.topP!==s.topP)throw new ne("local_model_configuration_conflict","The native model request changed its admitted model, output limit or sampling configuration.",422);let a=xs();a?.assertExecution(e.executionId);let o={teamId:e.teamId,id:e.executionId,ownerId:e.ownerId,requestId:r.requestId};await e.store.reserveLocalCharge({...o,caseId:e.caseId,maximumUsd:e.admission.maximumChargeUsd}),r.signal?.aborted&&(await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),r.signal.throwIfAborted());try{await e.store.markLocalChargeDispatched(o.teamId,o.id,o.requestId,o.ownerId)}catch(p){throw a&&await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),p}await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:"model.dispatch",payload:{requestId:o.requestId,model:r.model,messages:r.messages}});let n=null,i=!1,u=!1,f;try{for await(let p of t(r)){u=!0,p.type==="usage"&&(n=p.usage),p.type==="finish"&&(i=!0);let l=p.type==="text_delta"?{requestId:o.requestId,text:p.text}:p.type==="tool_call_delta"?{requestId:o.requestId,toolCalls:p.toolCalls}:p.type==="usage"?{requestId:o.requestId,usage:ve(p.usage)}:p.type==="finish"?{requestId:o.requestId,finishReason:p.finishReason}:null;l&&await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:`model.${p.type}`,payload:l}),yield p}}catch(p){throw!u&&p instanceof Is&&p.executionId===o.id&&p.requestId===o.requestId&&(f=p),p}finally{let p=ve(n),l=e.admission.pricing&&i&&p.promptTokens!==null&&p.completionTokens!==null?xe(n,e.admission.pricing):null;await e.store.settleLocalCharge({...o,costUsd:l,confirmedTransportNotInvoked:f,usage:n===null?null:{promptTokens:p.promptTokens,completionTokens:p.completionTokens,cachedPromptTokens:p.cachedPromptTokens,uncachedPromptTokens:p.uncachedPromptTokens}})}}}d(Nd,"createLocalBudgetedModelStream");function Bs(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Bs,"record");function Ws(e){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new ne("local_model_bound_unknown","The model catalog is missing an enforceable context/output bound.",422);return e}d(Ws,"positive");import{AsyncLocalStorage as pa}from"node:async_hooks";import fa from"node:path";var Fs=new pa;function Vs(){return Fs.getStore()}d(Vs,"currentAdvancedRefinerBoundary");function Jd(e){let t=Mr(fa.join(e.storeDir,"library","harnesses","advanced-refiner-paid-calls.sqlite"));t.exec(`CREATE TABLE IF NOT EXISTS advanced_refiner_budgets(id TEXT PRIMARY KEY,actor_id TEXT NOT NULL,team_id TEXT NOT NULL,pin_hash TEXT NOT NULL,maximum_spend REAL NOT NULL,maximum_steps INTEGER NOT NULL,expires_at TEXT NOT NULL);
  CREATE TABLE IF NOT EXISTS advanced_refiner_calls(run_id TEXT NOT NULL,id TEXT NOT NULL,intent_hash TEXT NOT NULL,ceiling REAL NOT NULL,state TEXT NOT NULL,actual_spend REAL,result TEXT,PRIMARY KEY(run_id,id));`);function r(a){t.exec("BEGIN IMMEDIATE");try{let o=a();return t.exec("COMMIT"),o}catch(o){throw t.exec("ROLLBACK"),o}}d(r,"transaction");async function s(a,o,n){let i=Ft(o);await e.authorize(i),r(()=>{let l=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(l){if(l.actor_id!==i.actorId||l.team_id!==i.teamId||l.pin_hash!==i.contentHash)throw new Error("Advanced Run budget belongs to another immutable admission.");return}t.prepare("INSERT INTO advanced_refiner_budgets VALUES(?,?,?,?,?,?,?)").run(a,i.actorId,i.teamId,i.contentHash,i.maximumCostUsd,i.maximumModelSteps,new Date(Date.now()+i.maximumDurationMs).toISOString())});let u=t.prepare("SELECT expires_at FROM advanced_refiner_budgets WHERE id=?").get(a),f=Date.parse(String(u.expires_at))-Date.now();if(f<=0)throw new Error("The immutable advanced evaluation duration is exhausted.");let p={runId:a,pin:i,signal:AbortSignal.timeout(f),authorize:d(()=>e.authorize(i),"authorize"),async resolveWorkAssets(l,c){await e.authorize(i),p.signal.throwIfAborted();let m=await e.packageForPin(i),g=Us(m,l,c);return await e.authorize(i),p.signal.throwIfAborted(),g},async call(l,c,m,g,h){if(await e.authorize(i),!Number.isInteger(g)||g<1)throw new Error("Advanced paid requests require a finite enforced token ceiling.");let w=c.model;if(w?.providerId!=="openpond"||!w.modelId)throw new Error("Advanced paid calls require an explicitly qualified hosted model.");let y=w.modelId,S=e.prepareModel?await e.prepareModel(y,g,m):await(async()=>{let C=await Yt({kind:"hosted_chat",modelId:y,maxOutputTokens:g,temperature:0,topP:1});if(k(C.pricing)!==k(m))throw new Error("The actual model pricing differs from its reviewed request pin.");return{configurationHash:C.model.configurationHash,maximumChargeUsd:C.maximumChargeUsd}})();if(!/^[a-f0-9]{64}$/.test(S.configurationHash)||!Number.isFinite(S.maximumChargeUsd)||S.maximumChargeUsd<=0)throw new Error("The actual owner model ceiling is unavailable.");let _=k({intent:c,pricing:m,maximumOutputTokens:g,modelConfigurationHash:S.configurationHash}),x=S.maximumChargeUsd,v=r(()=>{let C=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(!C||String(C.expires_at)<=new Date().toISOString())throw new Error("The immutable advanced evaluation duration is exhausted.");let E=t.prepare("SELECT * FROM advanced_refiner_calls WHERE run_id=? AND id=?").get(a,l);if(E){if(E.intent_hash!==_)throw new Error("Advanced provider request ID retains another exact payload.");if(E.state!=="completed")throw new Error("This paid request has an unknown outcome; recover its actual provider receipt before resuming.");return JSON.parse(String(E.result))}let T=t.prepare("SELECT COUNT(*) AS steps,COALESCE(SUM(COALESCE(actual_spend,ceiling)),0) AS spent,SUM(CASE WHEN state NOT IN ('completed','not_dispatched') THEN 1 ELSE 0 END) AS unknown FROM advanced_refiner_calls WHERE run_id=?").get(a);if(Number(T.unknown)>0)throw new Error("A prior advanced paid request is unsettled; no fresh request is allowed.");if(Number(T.steps)>=i.maximumModelSteps||Number(T.spent)+x>i.maximumCostUsd+1e-9)throw new Error("The advanced evaluation step or cumulative spend ceiling is exhausted.");t.prepare("INSERT INTO advanced_refiner_calls(run_id,id,intent_hash,ceiling,state) VALUES(?,?,?,?,?)").run(a,l,_,x,"reserved")});if(v!==void 0)return await e.authorize(i),p.signal.throwIfAborted(),v;let M=!1;try{if(await e.authorize(i),p.signal.throwIfAborted(),t.prepare("UPDATE advanced_refiner_calls SET state='dispatched' WHERE run_id=? AND id=? AND intent_hash=? AND state='reserved'").run(a,l,_).changes!==1)throw new Error("The retained advanced request admission changed.");M=!0;let E=await h();if(E.costUsd===null||!Number.isFinite(E.costUsd)||E.costUsd<0||E.costUsd>x+1e-9)throw new Error("The actual provider receipt cannot settle its admitted request ceiling.");let T=JSON.stringify(E.value);if(Buffer.byteLength(T)>8*1024*1024)throw new Error("The advanced provider response exceeds the retained output bound.");return r(()=>{if(t.prepare("UPDATE advanced_refiner_calls SET state='completed',actual_spend=?,result=? WHERE run_id=? AND id=? AND intent_hash=? AND state='dispatched'").run(E.costUsd,T,a,l,_).changes!==1)throw new Error("The advanced request settlement changed concurrently.")}),await e.authorize(i),E.value}catch(C){throw M?t.prepare("UPDATE advanced_refiner_calls SET state='unknown' WHERE run_id=? AND id=? AND state='dispatched'").run(a,l):t.prepare("UPDATE advanced_refiner_calls SET state='not_dispatched',actual_spend=0 WHERE run_id=? AND id=? AND state='reserved'").run(a,l),C}}};return Fs.run(p,n)}return d(s,"run"),{run:s,async authorize(a){await e.authorize(Ft(a))},read(a,o){return t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=? AND actor_id=? AND team_id=? AND pin_hash=?").get(a,o.actorId,o.teamId,o.contentHash)?t.prepare("SELECT id,intent_hash,ceiling,state,actual_spend FROM advanced_refiner_calls WHERE run_id=? ORDER BY id").all(a):null},accounting(a,o){let n=this.read(a,o)??[],i=n.filter(f=>f.state!=="completed"&&f.state!=="not_dispatched"),u=n.reduce((f,p)=>f+Number(p.actual_spend??0),0);return{modelRequests:n.filter(f=>f.state!=="not_dispatched").length,knownSpendUsd:u,totalSpendUsd:i.length?null:u,heldUsd:i.reduce((f,p)=>f+Number(p.ceiling),0),uncertainRequests:i.filter(f=>f.state!=="reserved").length}},close(){t.close()}}}d(Jd,"createAdvancedRefinerPaidBoundary");import ra from"node:path";function xt(e,t,r){let s=new Date().toISOString();return W.parse({schemaVersion:"openpond.taskAttempt.v1",id:`fixture_attempt_${t.id}_${r}`,tasksetId:e,taskId:t.taskId,split:Ke(t)??"frozen_eval",attempt:r,seed:0,modelRef:null,startedAt:s,completedAt:s,output:t.output,runtimeEventRefs:[],artifactRefs:[],privilegedOutcomeRef:null,infrastructureError:t.infrastructureError,costUsd:0,latencyMs:0,userInterventions:0,metadata:{fixtureLabel:t.label}})}d(xt,"fixtureAttempt");function Ke(e){let t=e&&typeof e=="object"&&!Array.isArray(e)&&"metadata"in e&&e.metadata&&typeof e.metadata=="object"&&!Array.isArray(e.metadata)?e.metadata.artifactSplit:null;return t==="train"||t==="validation"||t==="test"||t==="frozen_eval"?t:null}d(Ke,"artifactSplit");function Js(e){let t=Jr(e.taskset),r=t.issues.filter(b=>b.severity==="error").map(b=>({code:b.code,message:b.message,path:b.path})),s=t.issues.filter(b=>b.severity==="warning").map(b=>({code:b.code,message:b.message,path:b.path})),a=typeof e.taskset.metadata.trainingMethod=="string"?e.taskset.metadata.trainingMethod:null,o=ga(a),n=o==="sft",i=o==="dpo",u=o==="grpo"||o==="ppo",f=new Set(e.taskset.learningSignals.demonstrations.filter(b=>b.approved&&b.taskId).map(b=>b.taskId)),p=e.taskset.metadata.learning===void 0?null:He.parse(e.taskset.metadata.learning),l=d(b=>p?p.admissions.some(B=>B.taskId===b.id&&B.supervisedTarget!==null):b.expectedOutput!=null,"hasSupervisedTarget"),c=e.taskset.datasetArtifact,m=new Set(zs(e.taskset.metadata.approvedArtifactSignals)),g=!!c?.schema.fields.some(b=>b.semanticRole==="demonstration"||b.semanticRole==="expected_output"||b.semanticRole==="chosen"),h=e.taskset.tasks.filter(b=>b.split==="train"&&(l(b)||u)),w=n?h.filter(b=>f.has(b.id)):h,y=e.taskset.tasks.filter(b=>b.split==="train"&&l(b)&&!f.has(b.id)),S=e.taskset.tasks.filter(b=>b.split==="frozen_eval"&&(b.expectedOutput||u)),_=c&&g?c.splitCounts.train??0:0,x=c&&g?c.splitCounts.frozen_eval??0:0,v=c?m.has(o??""):!1,M=c?v?_:0:w.length,C=e.taskset.learningSignals.preferences.filter(b=>b.approved).length,E=Gr(e.taskset),T=E?E.binding.sources.filter(b=>b.role==="training"&&b.weight>0&&(!b.taskIds||w.some(B=>b.taskIds.includes(B.id)))).length:e.taskset.learningSignals.rewards.filter(b=>b.approved&&b.executable).length,A=i?C:u?T?M:0:M,R=c?x:S.length;A===0&&r.push(n?{code:"sft_demonstrations_missing",message:"At least one explicitly approved training demonstration is required.",path:"learningSignals.demonstrations"}:i?{code:"dpo_preferences_missing",message:"At least one approved chosen/rejected response pair is required for DPO.",path:"learningSignals.preferences"}:{code:"online_reward_missing",message:`${o?.toUpperCase()??"Online training"} requires an executable reward-bearing task.`,path:"learningSignals.rewards"});let j=c?n&&!v?_:0:y.length;n&&j&&r.push({code:"sft_demonstrations_unapproved",message:`${j} training example${j===1?" is":"s are"} not explicitly approved.`,path:c?"metadata.approvedArtifactSignals":"learningSignals.demonstrations"}),R===0&&s.push({code:"frozen_eval_missing",message:"At least one independent evaluation example is recommended before making quality claims.",path:c?"datasetArtifact.splitCounts.frozen_eval":"tasks"});let $=new Set(w.map(b=>b.clusterKey)),z=new Set(S.map(b=>b.clusterKey)),U=c&&e.taskset.metadata.splitIsolationVerified===!0;(c?!U||A===0||R===0:$.size===0||z.size===0||[...$].some(b=>z.has(b)))&&s.push({code:"independent_evaluation_missing",message:"Independent source clusters are recommended before using evaluation results for quality claims.",path:c?"metadata.splitIsolationVerified":"tasks"}),(c?e.taskset.metadata.artifactRowsVerified!==!0:e.taskset.tasks.some(b=>typeof b.metadata.exampleOrigin!="string"))&&r.push({code:"example_provenance_missing",message:"Every training and evaluation example must record whether it was extracted, corrected, synthetic, or expert-authored.",path:c?"metadata.artifactRowsVerified":"tasks.metadata.exampleOrigin"});let Q=ha(e.taskset.metadata.diagnosis);(!Q||typeof Q.summary!="string"||!Array.isArray(Q.stableBehavior))&&s.push({code:"capability_diagnosis_missing",message:"A capability diagnosis is recommended to separate stable behavior from changing knowledge.",path:"metadata.diagnosis"}),Q?.trainingEligible===!1&&s.push({code:"training_not_recommended",message:"The capability diagnosis does not recommend storing this behavior in model weights.",path:"metadata.diagnosis.trainingEligible"});let me=e.taskset.graders.some(b=>b.rewardEligible&&(b.kind!=="model_judge"||b.metadata.calibrationIsAdvisory!==!0));me&&!e.graderAudit&&s.push({code:"grader_audit_missing",message:"Optional grader calibration evidence has not been recorded.",path:"graderFixtures"}),me&&e.graderAudit&&e.graderAudit.tasksetHash!==e.taskset.contentHash&&s.push({code:"grader_audit_stale",message:"Optional grader calibration evidence predates this immutable Taskset revision.",path:"graderFixtures"}),me&&e.graderAudit&&!e.graderAudit.passed&&s.push({code:"grader_audit_failed",message:"The grader did not match every optional calibration fixture's expected outcome.",path:"graderFixtures"}),me&&e.graderAudit&&!e.graderAudit.hackingChecksPassed&&s.push({code:"grader_adversarial_calibration_failed",message:"Adversarial or prompt-injection calibration fixtures did not match their expected grader outcomes.",path:"graders"}),me&&e.graderAudit&&!e.graderAudit.leakageChecksPassed&&r.push({code:"environment_leakage",message:"Environment or privileged-state leakage checks failed.",path:"environment"}),me&&e.graderAudit&&!e.graderAudit.infrastructureSafetyPassed&&r.push({code:"infrastructure_reward",message:"An infrastructure failure produced a score or eligible reward.",path:"graderFixtures"});let tt=e.taskset.graders.some(b=>b.rewardEligible);u&&!tt&&r.push({code:"online_reward_grader_missing",message:`${o?.toUpperCase()} requires a user-selected reward-eligible grader.`,path:"graders"});let ie=o??(A>0&&e.taskset.capabilities.compatibleMethods.includes("sft")?"sft":"none"),I=e.taskset.learningSignals.demonstrations.filter(b=>b.approved).map(b=>b.id),N=ie==="none"?null:{primaryMethod:ie,bootstrap:ie==="grpo"&&I.length?{method:"sft",purpose:"trajectory_bootstrap",demonstrationRefs:I,limitations:["The SFT bootstrap imitates approved trajectories; it does not optimize verifier reward.","Completing the bootstrap does not satisfy the primary GRPO recommendation."]}:null},K=r.length===0,pe=r.length>0,rt=["sft","dpo","grpo","ppo"].map(b=>{let B=[],X=[];b==="sft"&&I.length===0&&(B.push("demonstrations_missing"),X.push("Add at least one approved prompt/response demonstration.")),b==="dpo"&&C===0&&(B.push("preference_pairs_missing"),X.push("Add at least one approved chosen/rejected response pair.")),(b==="grpo"||b==="ppo")&&T===0&&(B.push("executable_reward_missing"),X.push("Implement and validate an executable scalar reward.")),b==="ppo"&&(B.push("value_model_required"),X.push("Choose and bind a value/critic model before PPO can execute.")),R===0&&(B.push("frozen_eval_missing"),X.push("Add an independent frozen-evaluation split.")),pe&&B.length===0&&(B.push("taskset_not_ready"),X.push("Resolve the Dataset validation and grader-audit blockers."));let Me=B.filter(fe=>fe!=="value_model_required");return{method:b,status:Me.length?"needs_dataset_work":b===ie?"recommended":"compatible",reasonCodes:B,reasons:X}});return Vr.parse({schemaVersion:"openpond.tasksetReadiness.v1",tasksetId:e.taskset.id,tasksetHash:e.taskset.contentHash,ready:K,recommendedMethod:ie,trainingPath:N,methodReadiness:rt,compatibleDestinationClasses:K?ie==="grpo"?["export","custom","hosted_managed"]:["export","custom"]:["export"],blockers:r,advisories:s,warnings:[...t.issues.filter(b=>b.severity==="warning").map(b=>b.message),...ie==="grpo"?["GRPO is readiness-compatible but not executable in the local CPU fixture."]:[],...zs(e.taskset.metadata.warnings)],generatedAt:e.generatedAt??new Date().toISOString(),metadata:{reportHash:k([e.taskset.contentHash,r])}})}d(Js,"buildTasksetReadiness");function ga(e){return e==="sft"||e==="dpo"||e==="grpo"||e==="ppo"||e==="sdft"||e==="opsd"||e==="sdpo"?e:null}d(ga,"trainingPathMethod");function ha(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}d(ha,"metadataRecord");function zs(e){return Array.isArray(e)?e.filter(t=>typeof t=="string"&&!!t.trim()):[]}d(zs,"metadataStrings");import eo from"node:path";function Gs(e,t,r){let s=He.parse(t.metadata.learning),a=be(s.binding,s.rewards);return async({grader:o,task:n,attempt:i,signal:u})=>{let f=s.admissions.find(h=>h.taskId===n.id),p=a.find(h=>h.id===o.id);if(!f||p?.kind!=="custom_verifier")throw new Error("Prepared batch has no matching verifier admission.");let l=await wt(r,t),c=l?l.learningResources?.evidence.find(h=>Er(L(h),f.evidence)):await e.learningRepository().transaction(t.profileId,h=>ye(h,"evidence",f.evidence));if(!c)throw new Error("Imported batch evidence is missing from its package.");let m=dt(c,s.definition),g=d(h=>({id:h.id,input:h.input,expectedOutput:h.expectedOutput,policyVisibleContext:h.policyVisibleContext,privilegedContextRef:h.privilegedContextRef}),"selected");if(k(g(n))!==k(g(m)))throw new Error("Prepared task differs from its admitted evaluator context.");if(l){let h=l.learningResources?.assets.find(w=>w.id===p.verifierRef.id);if(!h)throw new Error("Imported batch verifier is missing from its package.");return $e({source:oe(h,p.verifierRef),exportName:p.exportName,timeoutMs:p.timeoutMs,signal:u,value:{task:m,attempt:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},input:m.input,output:i.output,expectedOutput:m.expectedOutput,evaluatorContext:c.submission.evaluatorContext,infrastructureError:i.infrastructureError??null}})}return Ye({repository:e.learningRepository(),scope:t.profileId,grader:p,task:m,evidence:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},evaluatorContext:c.submission.evaluatorContext,signal:u})}}d(Gs,"createLearningBatchVerifier");import{readFile as ka,realpath as Ys}from"node:fs/promises";import Ks from"node:path";async function Xs(e){let t=await Ys(e.allowedRoot),r=await Ys(Ks.resolve(t,e.grader.module));if(r!==t&&!r.startsWith(`${t}${Ks.sep}`))throw new Error("Verifier module is outside the approved Taskset root.");let s=await ka(r);if(e.grader.metadata.portableVerifierRef!==void 0){let o=Tr.parse(e.grader.metadata.portableVerifierRef);if(o.sizeBytes!==s.length||o.contentHash!==se(s))throw new Error("Verifier source differs from its immutable Taskset release.")}let a=new TextDecoder("utf-8",{fatal:!0}).decode(s);return $e({source:a,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:e.grader.timeoutMs,signal:e.signal,value:{task:e.task,attempt:e.attempt,input:e.task.input,expectedOutput:e.task.expectedOutput,output:e.attempt.output,infrastructureError:e.attempt.infrastructureError,...e.evaluatorContext?{evaluatorContext:e.evaluatorContext}:{}}})}d(Xs,"runSandboxedVerifier");import{constants as Kt}from"node:fs";import{open as ya,realpath as Zs}from"node:fs/promises";import We from"node:path";async function It(e,t){let r=De({taskset:t,adapterId:qe(t)});if(typeof t.metadata.importedPackageHash=="string"){let s=await Rs(e,t.metadata.importedPackageHash);if(s.taskset.contentHash!==r.tasksetRelease.contentHash)throw new Error("Evaluation Taskset differs from its pinned portable package.");let a=ct.parse(t.metadata.portableFileInventory??[]),{files:o,contentHash:n,...i}=s,u=await vt({root:We.join(e,"training","tasksets",ee(t)),content:i,sources:o.map(f=>({asset:f.asset,sourcePath:a.find(p=>p.asset.id===f.asset.id)?.sourcePath??f.asset.path})),fileOrder:o.map(f=>f.asset.id)});if(u.contentHash!==s.contentHash)throw new Error("Evaluation package changed during capture.");return u}return _s({storeDir:e,taskset:t,content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:r.tasksetRelease,environment:r.environmentRelease,verifierSet:r.verifierSetRelease}})}d(It,"captureTasksetEvaluationPackage");function Qs(e){return async({grader:t,task:r,attempt:s,signal:a})=>{let o=e.taskset.tasks.find(l=>l.id===r.id),n=e.taskset.graders.find(l=>l.id===t.id);if(!o||k(o)!==k(r)||!n||k(n)!==k(t)||s.tasksetId!==e.taskset.id||s.taskId!==r.id||s.split!==r.split)throw new Error("Process grading differs from the selected Taskset revision or attempt.");let i=e.packageValue.taskset.tasks.find(l=>l.id===r.id),u=e.packageValue.taskset.graders.find(l=>l.id===t.id);if(!i||u?.kind!=="custom_verifier"||u.runtime!=="sandbox_process")throw new Error("The process verifier is absent from the immutable package.");let f=e.fixtures?.get(s.id);if(f&&f.attemptHash!==k(s))throw new Error("Fixture attempt changed before private grading.");let p=f?.artifacts??await wa(e.store,e.storeDir,e.taskset,s);return Es({packageValue:e.packageValue,grader:u,task:i,evidence:{output:s.output,artifactRefs:s.artifactRefs,runtimeEventRefs:s.runtimeEventRefs,infrastructureError:s.infrastructureError,artifacts:p},signal:a,authorize:d(async()=>{a?.throwIfAborted()},"authorize")})}}d(Qs,"createTasksetProcessVerifier");async function wa(e,t,r,s){let a=await e.listTaskAttemptArtifacts({attemptId:s.id}),o=[],n=0;if(new Set(s.artifactRefs).size!==s.artifactRefs.length)throw new Error("Attempt contains duplicate artifact references.");for(let i of s.artifactRefs){let u=a.filter(g=>g.id===i&&g.tasksetId===r.id&&g.taskId===s.taskId&&g.attemptId===s.id);if(u.length!==1)throw new Error("Attempt artifact is not recorded by its execution owner.");let f=u[0];if(f.kind!=="output_artifact")continue;if(o.length>=16||(n+=f.sizeBytes)>2e7)throw new Error("Attempt artifacts exceed their byte boundary.");let p=await Zs(t),l=await Zs(We.dirname(f.path));if(![We.join(p,"training/evaluation-artifacts"),We.join(p,"work/outputs")].some(g=>l===g||l.startsWith(`${g}${We.sep}`)))throw new Error("Attempt artifact is outside the execution owner's output directories.");let c=await ya(f.path,Kt.O_RDONLY|Kt.O_NOFOLLOW|Kt.O_NONBLOCK),m;try{let g=await c.stat();if(!g.isFile()||g.size!==f.sizeBytes)throw new Error("Owner-recorded artifact size changed.");m=Buffer.alloc(g.size);let h=0;for(;h<m.length;){let w=await c.read(m,h,m.length-h,h);if(!w.bytesRead)throw new Error("Owner-recorded artifact changed during capture.");h+=w.bytesRead}if((await c.read(Buffer.alloc(1),0,1,h)).bytesRead)throw new Error("Owner-recorded artifact changed during capture.")}finally{await c.close()}if(se(m)!==f.sha256)throw new Error("Owner-recorded artifact bytes changed.");o.push({reference:{id:f.id,contentHash:f.sha256,mediaType:f.mediaType??"application/octet-stream",sizeBytes:f.sizeBytes},title:We.basename(typeof f.metadata.requiredOutputPath=="string"?f.metadata.requiredOutputPath:f.path),contentType:f.mediaType??"application/octet-stream",base64:m.toString("base64")})}return o}d(wa,"readOwnedOutputArtifacts");async function Et(e,t,r){if(!t.graders.some(p=>p.kind==="custom_verifier"))return;let s=t.graders.some(p=>p.kind==="custom_verifier"&&p.runtime==="sandbox_process"),a=t.metadata.learning!==void 0?await Gs(e.store,t,e.storeDir):t.metadata.rewardBinding!==void 0?await Ls(e.store,t,e.storeDir):void 0;if(a&&!s)return a;let o=e.storeDir?null:await(e.loadProfileState??Pr)(),n=e.storeDir?eo.join(e.storeDir,"training","tasksets",ee(t)):o?.sourcePath?eo.join(o.sourcePath,"tasksets",t.id):null,i=typeof t.metadata.creationSnapshotId=="string"?t.metadata.creationSnapshotId:null,u=i?await e.store.getTaskDesignProposal(i):null;if(n&&t.purpose!=="benchmark"&&t.environment.metadata.runtimeSourceTasksetId===void 0&&await ut(t,n,{generatedFiles:u?.generatedFiles??[]}),s&&!e.storeDir)throw new Error("Process grading requires the execution owner's package directory.");let f=s?Qs({store:e.store,storeDir:e.storeDir,taskset:t,packageValue:r?.packageValue??await It(e.storeDir,t),fixtures:r?.fixtures}):void 0;return n?async({grader:p,task:l,attempt:c,signal:m})=>{let g=t.graders.find(w=>w.id===p.id);if(!g||g.kind!=="custom_verifier")throw new Error("Verifier is not declared by this Taskset revision.");if(g.runtime==="sandbox_process")return f({grader:p,task:l,attempt:c,signal:m});if(a)return a({grader:g,task:l,attempt:c,signal:m});let h;if(t.environment.entrypoint===J&&!c.infrastructureError){if(!e.storeDir)throw new Error("Tool grading requires the execution owner's artifact directory.");h=await _t({store:e.store,storeDir:e.storeDir,taskset:t,task:l,attempt:c})}return Xs({grader:g,task:l,attempt:c,allowedRoot:n,signal:m,evaluatorContext:h})}:void 0}d(Et,"createTasksetEvaluationVerifier");var va=new Set(["correct","incorrect","parse_failure","budget_exhausted","tool_schema_violation","infrastructure_failure","cancelled"]);async function to(e){let t=Sa(e.attempt);if(t){let f=k(e.graders),p=`Taskset Work attempt ended with ${t.failureClass}: `+t.message;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${k([e.attempt.id,f,t]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:f,score:null,passed:!1,components:e.graders.map(l=>({graderId:l.id,graderVersion:l.version,score:0,passed:!1,hardGate:l.hardGate,rewardEligible:!1,feedback:p,evidenceRefs:e.attempt.artifactRefs,judge:l.kind==="model_judge"?l.judge:null,calibrationStatus:l.kind==="model_judge"?l.calibrationStatus:"not_applicable"})),failureClass:t.failureClass,feedback:[p],rewardEligible:!1,createdAt:e.now?.()??new Date().toISOString()}}let r=e.learning?null:ba(e.attempt.metadata);if(!r)try{return await lt(e)}catch(f){return Ta({attempt:e.attempt,graders:e.graders,now:e.now,error:f})}let s=k(e.graders),a=r.rewardEligible?r.outcome==="correct"?1:0:null,o=r.outcome==="correct",n=Ra(r.outcome),i=`Cross-System trajectory verifier outcome: ${r.outcome}.`,u=e.graders.map(f=>({graderId:f.id,graderVersion:f.version,score:a??0,passed:o,hardGate:f.hardGate,rewardEligible:r.rewardEligible&&f.rewardEligible,feedback:i,evidenceRefs:e.attempt.artifactRefs,judge:f.kind==="model_judge"?f.judge:null,calibrationStatus:f.kind==="model_judge"?f.calibrationStatus:"not_applicable"}));return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${k([e.attempt.id,s,r]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:s,score:a,passed:o,components:u,failureClass:n,feedback:[i],rewardEligible:r.rewardEligible,createdAt:e.now?.()??new Date().toISOString()}}d(to,"gradeTasksetEvaluationAttempt");function Ta(e){let t=k(e.graders),r=(e.error instanceof Error?e.error.message:String(e.error)).slice(0,19e3),s=`Taskset grader failed: ${r}`;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${k([e.attempt.id,t,"grader_failure",r]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:t,score:null,passed:!1,components:e.graders.map(a=>({graderId:a.id,graderVersion:a.version,score:0,passed:!1,hardGate:a.hardGate,rewardEligible:!1,feedback:s,evidenceRefs:e.attempt.artifactRefs,judge:a.kind==="model_judge"?a.judge:null,calibrationStatus:a.kind==="model_judge"?a.calibrationStatus:"not_applicable"})),failureClass:"grader_failure",feedback:[s],rewardEligible:!1,createdAt:e.now?.()??new Date().toISOString()}}d(Ta,"graderFailureResult");function Sa(e){if(e.metadata.execution!=="taskset_work"||!e.infrastructureError)return null;let t=Zr.safeParse(e.metadata.failureClass);if(!t.success||t.data==="policy_failure"||t.data==="grader_failure")throw new Error("Taskset Work attempt is missing a trusted execution failure class.");return{failureClass:t.data,message:e.infrastructureError}}d(Sa,"tasksetWorkFailure");function ba(e){if(e.execution!=="post_training_evaluation_tool_loop")return null;let t=e.verifierOutcome,r=e.verifierReward,s=e.verifierRewardEligible;if(typeof t!="string"||!va.has(t)||typeof s!="boolean"||r!==null&&typeof r!="number")throw new Error("Cross-System evaluation attempt is missing trusted verifier metadata.");if(s!==(t==="correct"||t==="incorrect"||t==="parse_failure")||s&&typeof r!="number"||!s&&r!==null)throw new Error("Cross-System evaluation verifier metadata is internally inconsistent.");return{outcome:t,reward:r,rewardEligible:s}}d(ba,"crossSystemGrade");function Ra(e){return e==="correct"?null:e==="infrastructure_failure"?"infrastructure_failure":e==="cancelled"?"cancelled":"policy_failure"}d(Ra,"failureClassFor");import Xe from"node:path";import{mkdir as ro,readFile as _a,writeFile as so}from"node:fs/promises";async function te(e){let t=Xe.join(e.storeDir,"training","evaluation-artifacts",e.tasksetId),r=Xe.join(t,`${e.attemptId}-${oo(e.fileLabel??e.kind)}.json`),s=Buffer.from(`${JSON.stringify({schemaVersion:"openpond.rawEvaluationArtifact.v1",requestId:e.requestId,...e.payload},null,2)}
 `,"utf8");return await ro(t,{recursive:!0}),await so(r,s,{mode:384}),Xt({...e,file:r,bytes:s,mediaType:"application/json"})}d(te,"persistJsonTaskAttemptArtifact");async function Zc(e){let t=e.extension.replace(/^\.+/,"");if(!/^[a-z0-9]{1,12}$/i.test(t))throw new Error("Task Attempt artifact extension is invalid.");let r=Xe.join(e.storeDir,"training","evaluation-artifacts",e.tasksetId);await ro(r,{recursive:!0});let s=Xe.join(r,`${e.attemptId}-${oo(e.fileLabel)}.${t}`);return await so(s,e.bytes,{mode:384}),Xt({...e,file:s})}d(Zc,"persistBinaryTaskAttemptArtifact");function oo(e){let t=e.trim().replace(/[^a-zA-Z0-9._-]+/g,"-").replace(/^-+|-+$/g,"");if(!t)throw new Error("Evaluation artifact file label is empty.");return t}d(oo,"safeFileLabel");async function ao(e){let t=await _a(e.path),r=se(t);if(t.byteLength!==e.expectedSizeBytes||r!==e.expectedSha256)throw new Error(`Saved Work output ${Xe.basename(e.path)} no longer matches its OutputRef.`);return Xt({...e,file:e.path,bytes:t,kind:"output_artifact",metadata:{...e.metadata,source:"openpond-work",outputValidated:!0}})}d(ao,"persistTaskAttemptOutputArtifact");async function Xt(e){let t=se(e.bytes),r=Qr.parse({schemaVersion:"openpond.taskAttemptArtifact.v1",id:`attempt_artifact_${k([e.attemptId,e.kind,e.file,t]).slice(0,24)}`,tasksetId:e.tasksetId,taskId:e.taskId,attemptId:e.attemptId,kind:e.kind,path:e.file,mediaType:e.mediaType,sha256:t,sizeBytes:e.bytes.byteLength,createdAt:e.timestamp(),metadata:{requestId:e.requestId,localOnly:!0,containsPrivilegedOutcome:!1,...e.metadata}});return e.store.saveTaskAttemptArtifact(r)}d(Xt,"saveArtifact");function Zt(e){let t=new Set(e.approvedTrajectoryIds),r=new Map(e.tasks.map(a=>[a.id,a])),s=new Map(e.results.map(a=>[a.trajectoryId,a]));return e.trajectories.flatMap(a=>{let o=s.get(a.id),n=r.get(a.taskId);return!t.has(a.id)||!n||o?.outcome!=="correct"||!o.rewardEligible?[]:[us.parse({schemaVersion:V,id:`cso_bootstrap_${a.id}`,taskId:n.id,worldId:a.worldId,trajectoryId:a.id,toolContractHash:q,approval:{status:"approved",approvedBy:e.approvedBy,approvedAt:e.approvedAt},messages:xa(n,a)})]})}d(Zt,"buildCrossSystemBootstrapDataset");function xa(e,t){let r=[{role:"system",content:cs},{role:"user",content:e.prompt}];for(let s of t.steps)s.kind==="model"&&s.content.trim()&&r.push({role:"assistant",content:s.content}),s.kind==="tool_call"&&r.push({role:"assistant",content:null,tool_calls:[{id:s.callId,type:"function",function:{name:s.name,arguments:JSON.stringify(s.arguments)}}]}),s.kind==="tool_result"&&r.push({role:"tool",tool_call_id:s.callId,content:JSON.stringify({ok:s.ok,result:s.result,error:s.error})}),s.kind==="final"&&r.push({role:"assistant",content:s.content});return r}d(xa,"trajectoryMessages");import _n from"node:path";import{mkdir as Ia,mkdtemp as Ea,rm as Ca}from"node:fs/promises";import Aa from"node:path";async function no(e,t,r){await Ia(e,{recursive:!0});let s=await Ea(Aa.join(e,".authoring-"));try{return await ut(t,s,{generatedFiles:r}),await Qt(e,t,s)}finally{await Ca(s,{recursive:!0,force:!0})}}d(no,"materializeGeneratedTasksetPackage");async function Qt(e,t,r){let s=await Kr(r),a=await Ss(t,r),o=`authored-${k(a.taskset)}`,n=we.parse({...a.taskset,environment:{...a.taskset.environment,metadata:{...a.taskset.environment.metadata,runtimeSourceTasksetId:o}}}),i=we.parse({...n,contentHash:Re(n)});return await ys(e,{taskset:i,generatedFiles:a.generatedFiles},o,{source:{directory:r,packageHash:s},verify:d(u=>yt(u,i),"verify")}),i}d(Qt,"materializeTasksetRevisionFromSource");import{Buffer as tr}from"node:buffer";import{spawn as Ma}from"node:child_process";var io=String.raw`
 import contextlib, io, json, math, statistics, decimal, datetime, collections, itertools, functools, sys

package/dist/chunks/{chunk-AF3AVDPP.js → chunk-VPAEFG7X.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{A as h,C as I,D as f,G as u,H as S,I as k,f as R,h as y,ma as w,oa as n,qa as P,ra as g,sa as T,ua as v}from"./chunk-WLZJOCLA.js";import{l as p}from"./chunk-G4E3WGLM.js";import{A as e}from"./chunk-X2J2P5ZT.js";import{a as l}from"./chunk-FTQLQJQV.js";var c=e.string().trim().min(1).max(240),o=e.string().trim().min(1).max(5e3),E=e.enum(["no_training","prompting","retrieval","sft","preference","grpo_rft","sdft_opsd","sdpo","agentic_rl"]),C=e.object({id:c,taskIndex:e.number().int().nonnegative(),label:k,output:e.record(e.string(),e.unknown()),infrastructureError:e.string().trim().min(1).max(1e4).nullable(),expectedPassed:e.boolean(),expectedRewardEligible:e.boolean(),metadata:e.record(e.string(),e.unknown()).default({})}),j=e.object({schemaVersion:e.literal("openpond.capabilityDiagnosis.v1"),summary:e.string().trim().min(1).max(1e4),stableBehavior:e.array(o).max(100).default([]),changingKnowledge:e.array(o).max(100).default([]),requiredContext:e.array(o).max(100).default([]),requiredTools:e.array(c).max(100).default([]),intervention:E,trainingEligible:e.boolean(),rationale:e.array(o).min(1).max(100),confidence:e.number().min(0).max(1)}),F=e.object({id:c,sourceId:c,sourceTurnId:c.nullable(),split:h,origin:e.enum(["extracted","corrected","synthetic","expert_authored"]),inputPrompt:e.string().trim().min(1).max(1e5),expectedOutputText:e.string().trim().min(1).max(2e5).nullable(),rationale:o}),D=e.object({schemaVersion:e.literal("openpond.taskDesignProposal.v1"),id:c,name:e.string().trim().min(1).max(500),objective:e.string().trim().min(1).max(2e4),diagnosis:j.default({schemaVersion:"openpond.capabilityDiagnosis.v1",summary:"Reproduce the selected approved behavior.",stableBehavior:[],changingKnowledge:[],requiredContext:[],requiredTools:[],intervention:"sft",trainingEligible:!0,rationale:["The selected examples were supplied as demonstrations."],confidence:.5}),taskKind:u.shape.taskKind,sourceIds:e.array(c).min(1).max(1e5),assumptions:e.array(o).max(1e3),successCriteria:e.array(o).min(1).max(1e3),proposedGraders:e.array(S).max(1e3).default([]),graderFixtures:e.array(C).max(1e5).default([]),generatedFiles:e.array(g).max(1e3).default([]),proposedExamples:e.array(F).max(1e5).default([]),proposedMethod:v.shape.recommendedMethod,trainingPath:T.nullable().default(null),policy:f,warnings:e.array(e.string().trim().min(1)).default([]),createdAt:e.string().trim().min(1)});var q=e.string().min(1).max(240),_=e.string().regex(/^[a-f0-9]{64}$/),H=e.array(e.object({source:I,excerpts:e.array(e.object({role:e.enum(["user","assistant"]),text:e.string().max(2e5),turnId:q}).strict()).max(200)}).strict()).min(1).max(20),Ht=e.discriminatedUnion("action",[e.object({action:e.literal("messages"),id:q,evidence:H,instruction:e.string().max(2*1024*1024),buildSpecification:P,currentProposal:D.nullable().default(null)}).strict(),e.object({action:e.literal("validate"),evidence:H,content:e.string().max(2*1024*1024)}).strict()]),qt=e.object({messages:e.array(e.object({role:e.enum(["system","user"]),content:e.string()}).strict()).min(1).max(20),skillHash:_}).strict(),_t=e.object({proposal:D,skillHash:_}).strict(),r=n.shape.id,s=e.number().int().positive(),d=e.string().regex(/^[a-f0-9]{64}$/),m={teamId:r,modelId:r},At=e.object({schemaVersion:e.literal("openpond.tasksetDraftCreate.v1"),operationId:r,modelId:r,expectedModelRevision:s,name:n.shape.name}).strict(),Lt=e.object({schemaVersion:e.literal("openpond.tasksetDraftSourceDescriptor.v1"),...m,expectedModelRevision:s,sourcePackageHash:d}).strict(),Ut=e.object({schemaVersion:e.literal("openpond.tasksetDraftReadback.v1"),...m,draft:n,workspaceHash:d,projectEtag:d}).strict().refine(t=>t.draft.profileId===t.teamId&&t.draft.modelScope?.modelId===t.modelId,"The hosted draft must belong to its declared workspace and Model."),A=e.object({id:r,revision:s,name:n.shape.name,status:w,updatedAt:n.shape.updatedAt,sourceTasksetRef:p.nullable()}).strict(),Ot=e.object({schemaVersion:e.literal("openpond.tasksetDraftList.v1"),...m,drafts:e.array(A).max(100),nextCursor:e.string().min(1).max(2e3).nullable()}).strict(),Gt=e.object({expectedDraftRevision:s,draft:n}).strict(),Vt=e.object({expectedDraftRevision:s,expectedModelRevision:s}).strict(),L=e.object({expectedDraftRevision:s,expectedWorkspaceHash:d}).strict(),Bt=e.object({schemaVersion:e.literal("openpond.tasksetDraftValidation.v1"),...m,draftId:r,draftRevision:s,workspaceHash:d,valid:e.boolean(),issues:e.array(e.object({code:e.string().min(1),severity:e.enum(["warning","error"]),message:e.string(),path:e.string().nullable()}).strict()),packageHash:d.nullable(),tasksetRef:p.nullable()}).strict().refine(t=>t.valid?t.packageHash!==null&&t.tasksetRef!==null&&t.issues.every(a=>a.severity!=="error"):t.packageHash===null&&t.tasksetRef===null&&t.issues.some(a=>a.severity==="error"),"Validation must retain either a compiled package or publication errors."),zt=L.extend({operationId:r,expectedPackageHash:d,expectedTasksetRef:p}).strict(),Kt=e.object({schemaVersion:e.literal("openpond.tasksetDraftFileList.v1"),...m,draftId:r,draftRevision:s,files:e.array(R).max(10020)}).strict(),Wt=e.object({schemaVersion:e.literal("openpond.tasksetDraftFileReadback.v1"),...m,draftId:r,draftRevision:s,file:y}).strict(),Nt=e.object({schemaVersion:e.literal("openpond.tasksetDraftDeletion.v1"),...m,draftId:r,revision:s,deleted:e.literal(!0)}).strict(),Xt=n.shape.id,Yt=e.number().int().positive();var i=e.string().trim().min(1).max(500),Zt=e.object({projectId:i.optional(),tasksetId:i.optional(),draftId:i.optional(),taskId:i.optional(),query:e.string().trim().max(1e3).default(""),split:e.enum(["train","validation","test","frozen_eval"]).optional(),after:e.string().min(1).max(2e3).optional(),limit:e.number().int().min(1).max(100).default(30)}).strict(),U=e.object({id:i,method:e.enum(["code","judge","human"]),required:e.boolean()}),O=e.object({tasksetId:i,tasksetName:e.string(),tasksetRevision:e.number().int().positive(),tasksetHash:e.string(),taskId:i,ordinal:e.number().int().nonnegative(),title:e.string(),description:e.string(),split:e.string(),scoring:e.array(U),configuration:e.enum(["configured","needs_reward","needs_checks"]),draftId:i.nullable().default(null),modelId:i.nullable().default(null)}),ea=e.object({items:e.array(O).max(100),nextCursor:e.string().nullable()});function ta(t,a){let $=[t.instruction,t.prompt,t.request,t.question,t.message].find(b=>typeof b=="string"&&b.trim().length>0),M=typeof t.title=="string"?t.title.trim():"",x=$?.trim()??"";return{title:(M||x||a).slice(0,180),description:x.slice(0,2e3)}}l(ta,"describeTaskInput");function aa(t){return t.map(a=>({id:a.id,method:a.kind==="model_judge"?"judge":a.kind==="human"?"human":"code",required:a.hardGate}))}l(aa,"summarizeTaskScoring");function sa(t){return t.some(a=>a.kind!=="human")?t.some(a=>a.kind==="model_judge"&&a.calibrationStatus!=="passed")?"needs_checks":"configured":"needs_reward"}l(sa,"taskConfiguration");export{E as a,D as b,Ht as c,Zt as d,O as e,ea as f,ta as g,aa as h,sa as i};
+import{A as h,C as I,D as f,G as u,H as S,I as k,f as R,h as y,ma as w,oa as n,qa as P,ra as g,sa as T,ua as v}from"./chunk-5ZNE43XC.js";import{l as p}from"./chunk-XHU4BZQY.js";import{A as e}from"./chunk-X2J2P5ZT.js";import{a as l}from"./chunk-FTQLQJQV.js";var c=e.string().trim().min(1).max(240),o=e.string().trim().min(1).max(5e3),E=e.enum(["no_training","prompting","retrieval","sft","preference","grpo_rft","sdft_opsd","sdpo","agentic_rl"]),C=e.object({id:c,taskIndex:e.number().int().nonnegative(),label:k,output:e.record(e.string(),e.unknown()),infrastructureError:e.string().trim().min(1).max(1e4).nullable(),expectedPassed:e.boolean(),expectedRewardEligible:e.boolean(),metadata:e.record(e.string(),e.unknown()).default({})}),j=e.object({schemaVersion:e.literal("openpond.capabilityDiagnosis.v1"),summary:e.string().trim().min(1).max(1e4),stableBehavior:e.array(o).max(100).default([]),changingKnowledge:e.array(o).max(100).default([]),requiredContext:e.array(o).max(100).default([]),requiredTools:e.array(c).max(100).default([]),intervention:E,trainingEligible:e.boolean(),rationale:e.array(o).min(1).max(100),confidence:e.number().min(0).max(1)}),F=e.object({id:c,sourceId:c,sourceTurnId:c.nullable(),split:h,origin:e.enum(["extracted","corrected","synthetic","expert_authored"]),inputPrompt:e.string().trim().min(1).max(1e5),expectedOutputText:e.string().trim().min(1).max(2e5).nullable(),rationale:o}),D=e.object({schemaVersion:e.literal("openpond.taskDesignProposal.v1"),id:c,name:e.string().trim().min(1).max(500),objective:e.string().trim().min(1).max(2e4),diagnosis:j.default({schemaVersion:"openpond.capabilityDiagnosis.v1",summary:"Reproduce the selected approved behavior.",stableBehavior:[],changingKnowledge:[],requiredContext:[],requiredTools:[],intervention:"sft",trainingEligible:!0,rationale:["The selected examples were supplied as demonstrations."],confidence:.5}),taskKind:u.shape.taskKind,sourceIds:e.array(c).min(1).max(1e5),assumptions:e.array(o).max(1e3),successCriteria:e.array(o).min(1).max(1e3),proposedGraders:e.array(S).max(1e3).default([]),graderFixtures:e.array(C).max(1e5).default([]),generatedFiles:e.array(g).max(1e3).default([]),proposedExamples:e.array(F).max(1e5).default([]),proposedMethod:v.shape.recommendedMethod,trainingPath:T.nullable().default(null),policy:f,warnings:e.array(e.string().trim().min(1)).default([]),createdAt:e.string().trim().min(1)});var q=e.string().min(1).max(240),_=e.string().regex(/^[a-f0-9]{64}$/),H=e.array(e.object({source:I,excerpts:e.array(e.object({role:e.enum(["user","assistant"]),text:e.string().max(2e5),turnId:q}).strict()).max(200)}).strict()).min(1).max(20),Ht=e.discriminatedUnion("action",[e.object({action:e.literal("messages"),id:q,evidence:H,instruction:e.string().max(2*1024*1024),buildSpecification:P,currentProposal:D.nullable().default(null)}).strict(),e.object({action:e.literal("validate"),evidence:H,content:e.string().max(2*1024*1024)}).strict()]),qt=e.object({messages:e.array(e.object({role:e.enum(["system","user"]),content:e.string()}).strict()).min(1).max(20),skillHash:_}).strict(),_t=e.object({proposal:D,skillHash:_}).strict(),r=n.shape.id,s=e.number().int().positive(),d=e.string().regex(/^[a-f0-9]{64}$/),m={teamId:r,modelId:r},At=e.object({schemaVersion:e.literal("openpond.tasksetDraftCreate.v1"),operationId:r,modelId:r,expectedModelRevision:s,name:n.shape.name}).strict(),Lt=e.object({schemaVersion:e.literal("openpond.tasksetDraftSourceDescriptor.v1"),...m,expectedModelRevision:s,sourcePackageHash:d}).strict(),Ut=e.object({schemaVersion:e.literal("openpond.tasksetDraftReadback.v1"),...m,draft:n,workspaceHash:d,projectEtag:d}).strict().refine(t=>t.draft.profileId===t.teamId&&t.draft.modelScope?.modelId===t.modelId,"The hosted draft must belong to its declared workspace and Model."),A=e.object({id:r,revision:s,name:n.shape.name,status:w,updatedAt:n.shape.updatedAt,sourceTasksetRef:p.nullable()}).strict(),Ot=e.object({schemaVersion:e.literal("openpond.tasksetDraftList.v1"),...m,drafts:e.array(A).max(100),nextCursor:e.string().min(1).max(2e3).nullable()}).strict(),Gt=e.object({expectedDraftRevision:s,draft:n}).strict(),Vt=e.object({expectedDraftRevision:s,expectedModelRevision:s}).strict(),L=e.object({expectedDraftRevision:s,expectedWorkspaceHash:d}).strict(),Bt=e.object({schemaVersion:e.literal("openpond.tasksetDraftValidation.v1"),...m,draftId:r,draftRevision:s,workspaceHash:d,valid:e.boolean(),issues:e.array(e.object({code:e.string().min(1),severity:e.enum(["warning","error"]),message:e.string(),path:e.string().nullable()}).strict()),packageHash:d.nullable(),tasksetRef:p.nullable()}).strict().refine(t=>t.valid?t.packageHash!==null&&t.tasksetRef!==null&&t.issues.every(a=>a.severity!=="error"):t.packageHash===null&&t.tasksetRef===null&&t.issues.some(a=>a.severity==="error"),"Validation must retain either a compiled package or publication errors."),zt=L.extend({operationId:r,expectedPackageHash:d,expectedTasksetRef:p}).strict(),Kt=e.object({schemaVersion:e.literal("openpond.tasksetDraftFileList.v1"),...m,draftId:r,draftRevision:s,files:e.array(R).max(10020)}).strict(),Wt=e.object({schemaVersion:e.literal("openpond.tasksetDraftFileReadback.v1"),...m,draftId:r,draftRevision:s,file:y}).strict(),Nt=e.object({schemaVersion:e.literal("openpond.tasksetDraftDeletion.v1"),...m,draftId:r,revision:s,deleted:e.literal(!0)}).strict(),Xt=n.shape.id,Yt=e.number().int().positive();var i=e.string().trim().min(1).max(500),Zt=e.object({projectId:i.optional(),tasksetId:i.optional(),draftId:i.optional(),taskId:i.optional(),query:e.string().trim().max(1e3).default(""),split:e.enum(["train","validation","test","frozen_eval"]).optional(),after:e.string().min(1).max(2e3).optional(),limit:e.number().int().min(1).max(100).default(30)}).strict(),U=e.object({id:i,method:e.enum(["code","judge","human"]),required:e.boolean()}),O=e.object({tasksetId:i,tasksetName:e.string(),tasksetRevision:e.number().int().positive(),tasksetHash:e.string(),taskId:i,ordinal:e.number().int().nonnegative(),title:e.string(),description:e.string(),split:e.string(),scoring:e.array(U),configuration:e.enum(["configured","needs_reward","needs_checks"]),draftId:i.nullable().default(null),modelId:i.nullable().default(null)}),ea=e.object({items:e.array(O).max(100),nextCursor:e.string().nullable()});function ta(t,a){let $=[t.instruction,t.prompt,t.request,t.question,t.message].find(b=>typeof b=="string"&&b.trim().length>0),M=typeof t.title=="string"?t.title.trim():"",x=$?.trim()??"";return{title:(M||x||a).slice(0,180),description:x.slice(0,2e3)}}l(ta,"describeTaskInput");function aa(t){return t.map(a=>({id:a.id,method:a.kind==="model_judge"?"judge":a.kind==="human"?"human":"code",required:a.hardGate}))}l(aa,"summarizeTaskScoring");function sa(t){return t.some(a=>a.kind!=="human")?t.some(a=>a.kind==="model_judge"&&a.calibrationStatus!=="passed")?"needs_checks":"configured":"needs_reward"}l(sa,"taskConfiguration");export{E as a,D as b,Ht as c,Zt as d,O as e,ea as f,ta as g,aa as h,sa as i};

package/dist/chunks/{chunk-FGOTSCSE.js → chunk-W3ZILZUL.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{b as g,d as $,e as O,f as C}from"./chunk-E2S2EMYQ.js";import{a as w}from"./chunk-3M53YXEI.js";import{a as r}from"./chunk-FTQLQJQV.js";var L=Symbol("openpondCliOptionValues"),a=class extends Error{static{r(this,"CliUsageError")}exitCode=2;constructor(t){super(t),this.name="CliUsageError"}},K=new Set(["maxOutputBytes","timeoutSec"]);function k(n){let t=[...n],o="",i={},e={},u=[];for(;t.length>0;){let s=t.shift();if(s==="--"){u.push(...t);break}if(s.startsWith("--")){let f=s.indexOf("="),c=f>=0?s.slice(2,f):s.slice(2),p=N(c),d=f>=0?s.slice(f+1):P(t,p,o);h(i,e,p,d,o)}else _(s)?y(s,t,i,e):o?u.push(s):o=s}return Object.defineProperty(i,L,{value:e,enumerable:!1,configurable:!1,writable:!1}),V(o,i),{command:o,options:i,rest:u}}r(k,"parseArgs");function N(n){return n.replace(/-([a-z])/g,(t,o)=>o.toUpperCase())}r(N,"normalizeCliOptionKey");function _(n){return/^-[^-].*/.test(n)&&!S(n)}r(_,"isShortOptionToken");function y(n,t,o,i){let e=n.slice(1);if(!e)return;let u=e.indexOf("="),s=u>=0?e.slice(0,u):e,f=u>=0?e.slice(u+1):null;if(s.length!==1){for(let d of s){let m=g[d];if(!m)throw new a(`unknown short option -${d}`);if(C(m)!=="boolean")throw new a(`short option -${d} requires its own value`);h(o,i,m,"true")}return}let c=g[s];if(!c)throw new a(`unknown short option -${s}`);let p=f??P(t,c);h(o,i,c,p)}r(y,"parseShortOptionToken");function P(n,t,o){let i=I(t,o);if(i==="boolean")return n[0]&&x(n[0])?n.shift():"true";if(i){let e=n.shift();if(e===void 0||R(e)&&!S(e))throw e!==void 0&&n.unshift(e),new a(`${l(t)} must be ${b(i)}`);return e}return n[0]&&!n[0].startsWith("--")?n.shift():"true"}r(P,"readOptionValue");function h(n,t,o,i,e){T(o,i,e),n[o]=i,(t[o]??=[]).push(i)}r(h,"setParsedOption");function T(n,t,o){let i=I(n,o);if(i){if(i==="boolean"){if(!x(t))throw new a(`${l(n)} must be a boolean`);return}if(t.trim().length===0||t==="true")throw new a(`${l(n)} must be ${b(i)}`);if(i==="number"||i==="integer"){let e=Number(t);if(!Number.isFinite(e))throw new a(`${l(n)} must be ${b(i)}`);if(i==="integer"&&!Number.isInteger(e))throw new a(`${l(n)} must be an integer`);if(i==="integer"&&K.has(n)&&e<=0)throw new a(`${l(n)} must be a positive integer`);return}if(i==="json")try{w(t,l(n))}catch(e){throw new a(e instanceof Error?e.message:String(e))}}}r(T,"validateParsedOptionValue");function I(n,t){return t?O(t,n):C(n)}r(I,"typedOptionKind");function l(n){return n.replace(/[A-Z]/g,t=>`-${t.toLowerCase()}`)}r(l,"cliOptionLabel");function E(n){return n==="json"?"JSON value":n}r(E,"optionKindLabel");function b(n){let t=E(n);return/^[aeiou]/i.test(t)?`an ${t}`:`a ${t}`}r(b,"optionKindPhrase");function x(n){return/^(true|false|1|0|yes|no)$/i.test(n)}r(x,"isBooleanLiteral");function R(n){return n.startsWith("-")&&n!=="-"}r(R,"isOptionLike");function S(n){return/^-\d+(?:\.\d+)?$/.test(n)}r(S,"isNegativeNumber");function V(n,t){let o=n?$(n):null;for(let i of Object.keys(t))if(!O(n,i)){let e=l(i),u=o?` for ${o.name}`:"";throw new a(`unknown option --${e}${u}`)}}r(V,"validateCommandOptions");export{a,k as b};
+import{b as g,d as $,e as O,f as C}from"./chunk-BBIA24GD.js";import{a as w}from"./chunk-3M53YXEI.js";import{a as r}from"./chunk-FTQLQJQV.js";var L=Symbol("openpondCliOptionValues"),a=class extends Error{static{r(this,"CliUsageError")}exitCode=2;constructor(t){super(t),this.name="CliUsageError"}},K=new Set(["maxOutputBytes","timeoutSec"]);function k(n){let t=[...n],o="",i={},e={},u=[];for(;t.length>0;){let s=t.shift();if(s==="--"){u.push(...t);break}if(s.startsWith("--")){let f=s.indexOf("="),c=f>=0?s.slice(2,f):s.slice(2),p=N(c),d=f>=0?s.slice(f+1):P(t,p,o);h(i,e,p,d,o)}else _(s)?y(s,t,i,e):o?u.push(s):o=s}return Object.defineProperty(i,L,{value:e,enumerable:!1,configurable:!1,writable:!1}),V(o,i),{command:o,options:i,rest:u}}r(k,"parseArgs");function N(n){return n.replace(/-([a-z])/g,(t,o)=>o.toUpperCase())}r(N,"normalizeCliOptionKey");function _(n){return/^-[^-].*/.test(n)&&!S(n)}r(_,"isShortOptionToken");function y(n,t,o,i){let e=n.slice(1);if(!e)return;let u=e.indexOf("="),s=u>=0?e.slice(0,u):e,f=u>=0?e.slice(u+1):null;if(s.length!==1){for(let d of s){let m=g[d];if(!m)throw new a(`unknown short option -${d}`);if(C(m)!=="boolean")throw new a(`short option -${d} requires its own value`);h(o,i,m,"true")}return}let c=g[s];if(!c)throw new a(`unknown short option -${s}`);let p=f??P(t,c);h(o,i,c,p)}r(y,"parseShortOptionToken");function P(n,t,o){let i=I(t,o);if(i==="boolean")return n[0]&&x(n[0])?n.shift():"true";if(i){let e=n.shift();if(e===void 0||R(e)&&!S(e))throw e!==void 0&&n.unshift(e),new a(`${l(t)} must be ${b(i)}`);return e}return n[0]&&!n[0].startsWith("--")?n.shift():"true"}r(P,"readOptionValue");function h(n,t,o,i,e){T(o,i,e),n[o]=i,(t[o]??=[]).push(i)}r(h,"setParsedOption");function T(n,t,o){let i=I(n,o);if(i){if(i==="boolean"){if(!x(t))throw new a(`${l(n)} must be a boolean`);return}if(t.trim().length===0||t==="true")throw new a(`${l(n)} must be ${b(i)}`);if(i==="number"||i==="integer"){let e=Number(t);if(!Number.isFinite(e))throw new a(`${l(n)} must be ${b(i)}`);if(i==="integer"&&!Number.isInteger(e))throw new a(`${l(n)} must be an integer`);if(i==="integer"&&K.has(n)&&e<=0)throw new a(`${l(n)} must be a positive integer`);return}if(i==="json")try{w(t,l(n))}catch(e){throw new a(e instanceof Error?e.message:String(e))}}}r(T,"validateParsedOptionValue");function I(n,t){return t?O(t,n):C(n)}r(I,"typedOptionKind");function l(n){return n.replace(/[A-Z]/g,t=>`-${t.toLowerCase()}`)}r(l,"cliOptionLabel");function E(n){return n==="json"?"JSON value":n}r(E,"optionKindLabel");function b(n){let t=E(n);return/^[aeiou]/i.test(t)?`an ${t}`:`a ${t}`}r(b,"optionKindPhrase");function x(n){return/^(true|false|1|0|yes|no)$/i.test(n)}r(x,"isBooleanLiteral");function R(n){return n.startsWith("-")&&n!=="-"}r(R,"isOptionLike");function S(n){return/^-\d+(?:\.\d+)?$/.test(n)}r(S,"isNegativeNumber");function V(n,t){let o=n?$(n):null;for(let i of Object.keys(t))if(!O(n,i)){let e=l(i),u=o?` for ${o.name}`:"";throw new a(`unknown option --${e}${u}`)}}r(V,"validateCommandOptions");export{a,k as b};

package/dist/chunks/{chunk-U2BD545L.js → chunk-WD6ANOLL.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{b as f}from"./chunk-GDJ4X2ZN.js";import{Z as c}from"./chunk-6I5AUTVO.js";import{C as w,D as I,z as g}from"./chunk-MVPKU72J.js";import{b as m}from"./chunk-FYWQWKVS.js";import{f as R}from"./chunk-ZY32SZ2P.js";import{A as r}from"./chunk-X2J2P5ZT.js";import{a as d}from"./chunk-FTQLQJQV.js";import n from"node:path";import{readFile as y}from"node:fs/promises";var A=r.object({profileId:r.string().min(1),sourceRevision:r.string().min(1),repositoryId:r.string().min(1),harnessRelease:r.object({id:r.string().min(1),contentHash:r.string().regex(/^[a-f0-9]{64}$/)}).strict(),hostExecution:r.object({conversationId:r.string().min(1),turnId:r.string().min(1)}).strict().optional(),workspaceId:r.string().min(1),assetsPath:r.string().min(1).refine(n.isAbsolute)}).strict();async function _(p,e){let t=A.parse(e),s=await f(p,t.harnessRelease);if(!s||n.resolve(s.release.bundlePath)!==n.resolve(t.assetsPath))throw new Error("Hosted Profile release changed after admission.");let o=s.release.harnessRelease.metadata.profile;if(!o||typeof o!="object"||o.id!==t.profileId||o.sourceRevision!==t.sourceRevision||o.repositoryId!==void 0&&o.repositoryId!==t.repositoryId)throw new Error("Hosted Profile provenance differs from admitted release.");return{profile:I(),profileLibrary:w()}}d(_,"loadHostedProfileStateAndLibrary");async function B(p,e){let t=A.parse(e),s=await f(p,t.harnessRelease);if(!s||n.resolve(s.release.bundlePath)!==n.resolve(t.assetsPath))throw new Error("Hosted Profile workflow release is unavailable.");let o=s.release.harnessRelease.metadata.profile;if(o?.id!==t.profileId||o.sourceRevision!==t.sourceRevision||o.repositoryId!==void 0&&o.repositoryId!==t.repositoryId)throw new Error("Hosted Profile workflow provenance changed.");let a=t.harnessRelease,l={source:"openpond_git",repositoryId:t.repositoryId,profileId:t.profileId};if(!s.release.harnessRelease.files.some(u=>u.path==="workflows/catalog.json"))return{profileRef:l,sourceRevision:t.sourceRevision,harnessRelease:a,workflows:[]};let i=n.join(s.release.bundlePath,"source","workflows"),v=R({agentSnapshot:s.release.agentSnapshot,harnessRelease:s.release.harnessRelease,catalogBytes:await y(n.join(i,"catalog.json")),actionBytes:await y(n.join(i,"actions.json"))});return{profileRef:l,sourceRevision:t.sourceRevision,harnessRelease:a,workflows:v.catalog.workflows.map(u=>({workflow:u,binding:{schemaVersion:"openpond.profileWorkflowBinding.v1",profileId:t.profileId,sourceRevision:t.sourceRevision,harnessRelease:a,catalogHash:v.catalogHash,workflowId:u.id}}))}}d(B,"listHostedProfileWorkflows");import{createHash as U,randomUUID as h}from"node:crypto";var b=r.object({approval:g,revision:r.number().int().positive()}).strict(),P=r.object({record:m}).strict(),x=r.object({startedAt:r.string().min(1),requestOrdinal:r.number().int().nonnegative(),id:r.string().min(1)}).strict(),k=r.object({records:r.array(m).max(200),nextBefore:x.nullable()}).strict(),H=1e4,S=class{constructor(e){this.client=e}client;static{d(this,"HostedApprovalStorage")}async readApproval(e){let t=await this.client.request({contractVersion:c,requestId:h(),operation:"approval/get",params:{approvalId:e}});return t===null?null:b.parse(t)}async getApproval(e){return(await this.readApproval(e))?.approval??null}async putApproval(e){let t=g.parse(e.approval);if(!e.requestId.trim())throw new Error("Approval mutation requires a stable request ID.");return b.parse(await this.client.request({contractVersion:c,requestId:e.requestId,operation:"approval/upsert",params:{approval:t,expectedRevision:e.expectedRevision}}))}},M=class{constructor(e){this.client=e}client;static{d(this,"HostedModelUsageStorage")}async upsertModelUsageRecord(e){let t=m.parse(e),s=U("sha256").update(JSON.stringify(t)).digest("hex"),o=await this.client.request({contractVersion:c,requestId:`usage:${s}`,operation:"usage/upsert",params:{record:t}}),a=P.parse(o).record;if(a.requestId!==t.requestId)throw new Error("Host usage upsert changed request identity.");return a}async getModelUsageRecordByRequestId(e){let t=await this.client.request({contractVersion:c,requestId:h(),operation:"usage/getByRequestId",params:{requestId:e}});return t===null?null:P.parse(t).record}async listModelUsageRecords(e={}){let t=e.limit===void 0||e.limit>=H,s=e.limit===void 0?H:Math.max(1,Math.min(1e4,Math.trunc(e.limit))),o={...e.sessionId?{sessionId:e.sessionId}:{},...e.turnId?{turnId:e.turnId}:{},...e.provider?{provider:e.provider}:{},...e.model?{model:e.model}:{},...e.startedAtFrom?{startedAtFrom:e.startedAtFrom}:{},...e.startedAtTo?{startedAtTo:e.startedAtTo}:{},...e.visibility&&e.visibility!=="all"?{visibility:e.visibility}:{},...e.status&&e.status!=="all"?{status:e.status}:{}},a=[],l=null;for(;a.length<s+(t?1:0);){let i=k.parse(await this.client.request({contractVersion:c,requestId:h(),operation:"usage/page",params:{query:o,before:l,limit:Math.min(200,s+(t?1:0)-a.length)}}));if(a.push(...i.records),a.length>s)throw new Error(`Hosted usage query exceeds ${s} rows; narrow its filters or use a streamed aggregation.`);if(i.nextBefore===null)break;if(i.records.length===0||JSON.stringify(i.nextBefore)===JSON.stringify(l))throw new Error("Host usage cursor did not advance.");l=i.nextBefore}return a}async listModelUsageRecordsComplete(e={}){return this.listModelUsageRecords(e)}};export{A as a,_ as b,B as c,S as d,M as e};
+import{b as f}from"./chunk-IR2Y2VQY.js";import{Z as c}from"./chunk-6I5AUTVO.js";import{C as w,D as I,z as g}from"./chunk-CJDGLPNX.js";import{b as m}from"./chunk-FYWQWKVS.js";import{f as R}from"./chunk-ZY32SZ2P.js";import{A as r}from"./chunk-X2J2P5ZT.js";import{a as d}from"./chunk-FTQLQJQV.js";import n from"node:path";import{readFile as y}from"node:fs/promises";var A=r.object({profileId:r.string().min(1),sourceRevision:r.string().min(1),repositoryId:r.string().min(1),harnessRelease:r.object({id:r.string().min(1),contentHash:r.string().regex(/^[a-f0-9]{64}$/)}).strict(),hostExecution:r.object({conversationId:r.string().min(1),turnId:r.string().min(1)}).strict().optional(),workspaceId:r.string().min(1),assetsPath:r.string().min(1).refine(n.isAbsolute)}).strict();async function _(p,e){let t=A.parse(e),s=await f(p,t.harnessRelease);if(!s||n.resolve(s.release.bundlePath)!==n.resolve(t.assetsPath))throw new Error("Hosted Profile release changed after admission.");let o=s.release.harnessRelease.metadata.profile;if(!o||typeof o!="object"||o.id!==t.profileId||o.sourceRevision!==t.sourceRevision||o.repositoryId!==void 0&&o.repositoryId!==t.repositoryId)throw new Error("Hosted Profile provenance differs from admitted release.");return{profile:I(),profileLibrary:w()}}d(_,"loadHostedProfileStateAndLibrary");async function B(p,e){let t=A.parse(e),s=await f(p,t.harnessRelease);if(!s||n.resolve(s.release.bundlePath)!==n.resolve(t.assetsPath))throw new Error("Hosted Profile workflow release is unavailable.");let o=s.release.harnessRelease.metadata.profile;if(o?.id!==t.profileId||o.sourceRevision!==t.sourceRevision||o.repositoryId!==void 0&&o.repositoryId!==t.repositoryId)throw new Error("Hosted Profile workflow provenance changed.");let a=t.harnessRelease,l={source:"openpond_git",repositoryId:t.repositoryId,profileId:t.profileId};if(!s.release.harnessRelease.files.some(u=>u.path==="workflows/catalog.json"))return{profileRef:l,sourceRevision:t.sourceRevision,harnessRelease:a,workflows:[]};let i=n.join(s.release.bundlePath,"source","workflows"),v=R({agentSnapshot:s.release.agentSnapshot,harnessRelease:s.release.harnessRelease,catalogBytes:await y(n.join(i,"catalog.json")),actionBytes:await y(n.join(i,"actions.json"))});return{profileRef:l,sourceRevision:t.sourceRevision,harnessRelease:a,workflows:v.catalog.workflows.map(u=>({workflow:u,binding:{schemaVersion:"openpond.profileWorkflowBinding.v1",profileId:t.profileId,sourceRevision:t.sourceRevision,harnessRelease:a,catalogHash:v.catalogHash,workflowId:u.id}}))}}d(B,"listHostedProfileWorkflows");import{createHash as U,randomUUID as h}from"node:crypto";var b=r.object({approval:g,revision:r.number().int().positive()}).strict(),P=r.object({record:m}).strict(),x=r.object({startedAt:r.string().min(1),requestOrdinal:r.number().int().nonnegative(),id:r.string().min(1)}).strict(),k=r.object({records:r.array(m).max(200),nextBefore:x.nullable()}).strict(),H=1e4,S=class{constructor(e){this.client=e}client;static{d(this,"HostedApprovalStorage")}async readApproval(e){let t=await this.client.request({contractVersion:c,requestId:h(),operation:"approval/get",params:{approvalId:e}});return t===null?null:b.parse(t)}async getApproval(e){return(await this.readApproval(e))?.approval??null}async putApproval(e){let t=g.parse(e.approval);if(!e.requestId.trim())throw new Error("Approval mutation requires a stable request ID.");return b.parse(await this.client.request({contractVersion:c,requestId:e.requestId,operation:"approval/upsert",params:{approval:t,expectedRevision:e.expectedRevision}}))}},M=class{constructor(e){this.client=e}client;static{d(this,"HostedModelUsageStorage")}async upsertModelUsageRecord(e){let t=m.parse(e),s=U("sha256").update(JSON.stringify(t)).digest("hex"),o=await this.client.request({contractVersion:c,requestId:`usage:${s}`,operation:"usage/upsert",params:{record:t}}),a=P.parse(o).record;if(a.requestId!==t.requestId)throw new Error("Host usage upsert changed request identity.");return a}async getModelUsageRecordByRequestId(e){let t=await this.client.request({contractVersion:c,requestId:h(),operation:"usage/getByRequestId",params:{requestId:e}});return t===null?null:P.parse(t).record}async listModelUsageRecords(e={}){let t=e.limit===void 0||e.limit>=H,s=e.limit===void 0?H:Math.max(1,Math.min(1e4,Math.trunc(e.limit))),o={...e.sessionId?{sessionId:e.sessionId}:{},...e.turnId?{turnId:e.turnId}:{},...e.provider?{provider:e.provider}:{},...e.model?{model:e.model}:{},...e.startedAtFrom?{startedAtFrom:e.startedAtFrom}:{},...e.startedAtTo?{startedAtTo:e.startedAtTo}:{},...e.visibility&&e.visibility!=="all"?{visibility:e.visibility}:{},...e.status&&e.status!=="all"?{status:e.status}:{}},a=[],l=null;for(;a.length<s+(t?1:0);){let i=k.parse(await this.client.request({contractVersion:c,requestId:h(),operation:"usage/page",params:{query:o,before:l,limit:Math.min(200,s+(t?1:0)-a.length)}}));if(a.push(...i.records),a.length>s)throw new Error(`Hosted usage query exceeds ${s} rows; narrow its filters or use a streamed aggregation.`);if(i.nextBefore===null)break;if(i.records.length===0||JSON.stringify(i.nextBefore)===JSON.stringify(l))throw new Error("Host usage cursor did not advance.");l=i.nextBefore}return a}async listModelUsageRecordsComplete(e={}){return this.listModelUsageRecords(e)}};export{A as a,_ as b,B as c,S as d,M as e};

package/dist/chunks/{chunk-OTUS7NTQ.js → chunk-WDKZMXIU.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{va as t}from"./chunk-QCJY6FTC.js";import{A as e}from"./chunk-X2J2P5ZT.js";var n=e.string().trim().min(1).max(240),a=e.object({id:n,revision:e.number().int().positive(),ownerInstanceId:n,teamId:n,profileId:n,projectId:n,hostedModelId:n,role:t,roleTargetId:n,expectedBindingId:n,policyId:n.nullable(),authorizedConfigurationHash:e.string().length(64).nullable(),enabled:e.boolean(),updatedAt:e.string().datetime()}).strict(),o=e.object({id:n,grantId:n,revision:e.number().int().positive(),jobId:n,artifactHash:e.string().length(64),priorBindingId:n,bindingId:n.nullable(),restoredBindingId:n.nullable(),status:e.enum(["prepared","binding","canary","active","superseded","rolling_back","rolled_back","failed"]),leaseId:n,canarySeconds:e.number().int().min(30).max(600),activatedAt:e.string().datetime().nullable(),failure:e.string().max(2e3).nullable(),createdAt:e.string().datetime(),updatedAt:e.string().datetime()}).strict();export{a,o as b};
+import{va as t}from"./chunk-FCXDPVGJ.js";import{A as e}from"./chunk-X2J2P5ZT.js";var n=e.string().trim().min(1).max(240),a=e.object({id:n,revision:e.number().int().positive(),ownerInstanceId:n,teamId:n,profileId:n,projectId:n,hostedModelId:n,role:t,roleTargetId:n,expectedBindingId:n,policyId:n.nullable(),authorizedConfigurationHash:e.string().length(64).nullable(),enabled:e.boolean(),updatedAt:e.string().datetime()}).strict(),o=e.object({id:n,grantId:n,revision:e.number().int().positive(),jobId:n,artifactHash:e.string().length(64),priorBindingId:n,bindingId:n.nullable(),restoredBindingId:n.nullable(),status:e.enum(["prepared","binding","canary","active","superseded","rolling_back","rolled_back","failed"]),leaseId:n,canarySeconds:e.number().int().min(30).max(600),activatedAt:e.string().datetime().nullable(),failure:e.string().max(2e3).nullable(),createdAt:e.string().datetime(),updatedAt:e.string().datetime()}).strict();export{a,o as b};

package/dist/chunks/chunk-WGISITX6.js

@@ -0,0 +1,4 @@
+import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
+import{a as q}from"./chunk-FEAKTSRS.js";import{a as J,b as S}from"./chunk-R5GRESW3.js";import{$ as j,Z as $}from"./chunk-6I5AUTVO.js";import{aa as _,ja as H}from"./chunk-5ZNE43XC.js";import{yb as N}from"./chunk-QAFXQHHG.js";import{v as I}from"./chunk-ZY32SZ2P.js";import{n as k}from"./chunk-RCT3O3HQ.js";import{A as g}from"./chunk-X2J2P5ZT.js";import{a as o}from"./chunk-FTQLQJQV.js";import{randomUUID as O,createHash as W}from"node:crypto";import{mkdtemp as F,mkdir as G,writeFile as A,rm as U}from"node:fs/promises";import D from"node:os";import h from"node:path";var w=g.object({score:g.number().finite().min(0).max(1).nullable(),passed:g.boolean(),feedback:g.string().max(2e4),evidenceRefs:g.array(g.string().max(240)).max(1e3).default([])}).strict().refine(e=>e.score!==null||!e.passed,{message:"Unscorable verifiers cannot pass.",path:["passed"]});function L(e){if(e==="sandbox_process")throw new Error("verifier_sandbox_process_required")}o(L,"assertIsolatedVerifierRuntime");var ue=o(async e=>{let r=H(e.packageValue);if(e.grader.runtime!=="sandbox_process")throw new Error("Select the sandbox process runtime for a local file grader.");if(!r.taskset.graders.some(s=>k(s)===k(e.grader)))throw new Error("The private grader differs from its package.");if(!r.taskset.tasks.some(s=>k(s)===k(e.task)))throw new Error("The private case differs from its package.");await e.authorize(),e.signal?.throwIfAborted();let t=await F(h.join(D.tmpdir(),"profile-private-grade-")),c=O();try{let s=new Map;for(let a of r.files.filter(f=>f.asset.visibility!=="policy")){let f=a.asset.path;if(h.isAbsolute(f)||f.split(/[\\/]/).some(n=>!n||n==="."||n===".."))throw new Error("Private grader asset path invalid.");let l=s.get(f);if(l){if(l.contentHash!==a.asset.contentHash||l.sizeBytes!==a.asset.sizeBytes)throw new Error("Private grader assets declare conflicting bytes at the same path.");continue}s.set(f,{contentHash:a.asset.contentHash,sizeBytes:a.asset.sizeBytes});let x=h.join(t,"package",f);await G(h.dirname(x),{recursive:!0,mode:448}),await A(x,_(a),{mode:256})}await A(h.join(t,"request.json"),JSON.stringify({task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,artifacts:"artifacts"in e.evidence?e.evidence.artifacts:[],expectedOutput:e.task.expectedOutput}),{mode:256});let i=`import fs from 'node:fs/promises'; const module=await import(${JSON.stringify("./package/"+e.grader.verifierRef.path)}); const result=await module[${JSON.stringify(e.grader.exportName??"verify")}](JSON.parse(await fs.readFile('request.json','utf8'))); console.log(JSON.stringify(result));`;await A(h.join(t,"runner.mjs"),i,{mode:256});let m=await q({authorize:o(async()=>(await e.authorize(),e.signal?.throwIfAborted(),{candidateId:c,candidateRevision:1,ownerId:c,sessionId:c,turnId:c,sourceRoot:t,writablePaths:[]}),"authorize")})({candidateId:c,sessionId:c,turnId:c,expectedRevision:1,command:"/runtime/node --max-old-space-size=512 runner.mjs",timeoutMs:e.grader.timeoutMs,maxOutputBytes:7e4,signal:e.signal});if(await e.authorize(),e.signal?.throwIfAborted(),m.code!==0||m.timedOut||m.stdoutTruncated||m.stderrTruncated)throw new Error("Private grading sandbox did not complete.");return w.parse(JSON.parse(m.stdout))}finally{await U(t,{recursive:!0,force:!0})}},"executeLocalProfileVerifier");function pe(e){return async r=>{if(!r.manifest)throw new Error("The hosted grader requires its admitted manifest.");await r.authorize(),r.signal?.throwIfAborted();let t="artifacts"in r.evidence&&Array.isArray(r.evidence.artifacts)?r.evidence.artifacts.map(({reference:a,title:f,contentType:l})=>({reference:a,title:f,contentType:l})):[],c="caseOwner"in r.evidence?r.evidence.caseOwner:null;if(!c?.sessionId||!c.turnId)throw new Error("The private grader has no sealed case owner.");let s=Buffer.from(JSON.stringify({manifest:r.manifest,packageValue:r.packageValue,graderId:r.grader.id,taskId:r.task.id,evidence:{...r.evidence,artifacts:t}}));if(s.length>2e7)throw new Error("The private grading package exceeds its transport limit.");let i={callId:O(),sessionId:c.sessionId,turnId:c.turnId,sha256:W("sha256").update(s).digest("hex"),sizeBytes:s.length,timeoutMs:r.grader.timeoutMs},d=o((a,f=12e4)=>e.request(j.parse({contractVersion:$,requestId:O(),operation:"profile-evaluations/grade",params:a}),f),"request");for(let a=0;a<s.length;a+=98304)await r.authorize(),r.signal?.throwIfAborted(),await d({...i,action:"upload",offset:a,contentsBase64:s.subarray(a,a+98304).toString("base64")});let m=await d({...i,action:"execute"},r.grader.timeoutMs+12e4);return await r.authorize(),r.signal?.throwIfAborted(),w.parse(m)}}o(pe,"createHostedProfileVerifier");import{Worker as K}from"node:worker_threads";async function b(e){e.signal?.throwIfAborted(),S(e);let{javascriptVerifierWorkerSource:r}=await import("./javascript-verifier-worker-source-LGRP74W4.js");e.signal?.throwIfAborted();let t=e.signal,c={source:e.source,exportName:e.exportName,value:e.value,timeoutMs:e.timeoutMs,maxResultBytes:e.maxResultBytes,deterministic:e.deterministic===!0,errorPrefix:e.errorPrefix},s=o(i=>new Error(`${e.errorPrefix}_${i}`),"error");return new Promise((i,d)=>{let m=new K(r,{eval:!0,workerData:c,execArgv:[],env:{},resourceLimits:{maxOldGenerationSizeMb:64,maxYoungGenerationSizeMb:16,stackSizeMb:4}}),a=!1,f=o(n=>{a||(a=!0,clearTimeout(x),t?.removeEventListener("abort",l),m.terminate().then(()=>{"error"in n?d(n.error):i(n.result)},d))},"finish"),l=o(()=>f({error:t?.reason instanceof Error?t.reason:s("cancelled")}),"cancel"),x=setTimeout(()=>f({error:s("timeout")}),e.timeoutMs);t?.addEventListener("abort",l,{once:!0}),m.once("error",n=>f({error:n})),m.once("exit",()=>{a||f({error:s("worker_exited_without_result")})}),m.once("message",n=>{try{if(!n||typeof n!="object"||!("ok"in n))throw s("worker_invalid_response");if(n.ok!==!0)throw new Error("error"in n&&typeof n.error=="string"?n.error:`${e.errorPrefix}_execution_failed`);if(!("result"in n))throw s("worker_invalid_response");I(n.result,e.maxResultBytes),f({result:n.result})}catch(y){f({error:y instanceof Error?y:s("worker_invalid_response")})}}),t?.aborted&&l()})}o(b,"executeJavaScriptIsolateInWorker");import{spawn as Y}from"node:child_process";var B=class extends Error{static{o(this,"JavaScriptVerifierHostTimeoutError")}timeoutMs;elapsedMs;constructor(r,t){super("verifier_timeout"),this.timeoutMs=r,this.elapsedMs=t,this.name="JavaScriptVerifierHostTimeoutError"}};async function be(e){e.signal?.throwIfAborted(),S(e);let{javascriptIsolateProcessSource:r}=await import("./javascript-isolate-process-source-YAT3GPW4.js");e.signal?.throwIfAborted();let{signal:t,stripTypeScript:c,...s}=e,i=o(m=>new Error(`${e.errorPrefix}_${m}`),"error"),d=`globalThis.__openpondIsolateInput = JSON.parse(${JSON.stringify(JSON.stringify(s))});
+${c?`globalThis.__openpondIsolateInput.source = require("node:module").stripTypeScriptTypes(globalThis.__openpondIsolateInput.source, { mode: "strip" });
+`:""}${r}`;return new Promise((m,a)=>{let f=performance.now(),l=Y("node",["--max-old-space-size=64","--input-type=commonjs","-"],{stdio:["pipe","pipe","pipe","ipc"],env:{PATH:process.env.PATH},windowsHide:!0}),x,n="",y=0,M=0,v=o(u=>{x??=u,l.kill("SIGKILL")},"stop"),T=o(()=>v(t?.reason instanceof Error?t.reason:i("cancelled")),"cancel"),P=e.errorPrefix==="verifier",R=!P,V=o(()=>v(e.errorPrefix==="candidate"?new J("candidate_timeout"):e.errorPrefix==="verifier"?new B(e.timeoutMs,performance.now()-f):i("timeout")),"timedOut"),E=setTimeout(V,P?3e4:e.timeoutMs);l.on("message",u=>{if(!P||R||!u||typeof u!="object"||!("type"in u)||u.type!=="execution_ready"){v(i("process_invalid_startup_message"));return}R=!0,clearTimeout(E),E=setTimeout(V,e.timeoutMs)}),t?.addEventListener("abort",T,{once:!0}),l.stdout.setEncoding("utf8"),l.stdout.on("data",u=>{y+=Buffer.byteLength(u),y>e.maxResultBytes+65536?v(i("process_response_too_large")):n+=u}),l.stderr.on("data",u=>{M+=u.byteLength,M>65536&&v(i("process_stderr_too_large"))}),l.once("error",u=>{x??=u}),l.stdin.on("error",u=>v(u)),l.once("close",u=>{if(clearTimeout(E),t?.removeEventListener("abort",T),x){a(x);return}try{if(u!==0)throw i("process_exited_without_result");if(!R)throw i("process_exited_before_ready");let p=JSON.parse(n);if(!p||typeof p!="object"||!("ok"in p))throw i("process_invalid_response");if(p.ok!==!0){let z="error"in p&&typeof p.error=="string"?p.error:`${e.errorPrefix}_execution_failed`;throw"executionError"in p&&p.executionError===!0?new J(z):new Error(z)}if(!("result"in p))throw i("process_invalid_response");I(p.result,e.maxResultBytes),m(p.result)}catch(p){a(p instanceof Error?p:i("process_invalid_response"))}}),l.stdin.end(d),t?.aborted&&T()})}o(be,"executeJavaScriptIsolateInProcess");async function Je(e){return L(e.runtime),w.parse(await b({...e,exportName:e.exportName??"verify",maxResultBytes:65536,errorPrefix:"verifier"}))}o(Je,"executeJavaScriptVerifierInWorker");function Ne(e){return N(e,C)}o(Ne,"aggregateTasksetEvaluationInWorker");var C=o(async({source:e,module:r,exportName:t,scores:c,timeoutMs:s,signal:i})=>{let d=/\.[cm]?ts$/.test(r)?(await import("node:module")).stripTypeScriptTypes(e,{mode:"strip"}):e;return b({source:d,exportName:t,value:c,timeoutMs:s,signal:i,maxResultBytes:1024,deterministic:!0,errorPrefix:"metric"})},"executeMetric");var He=C;import Q from"node:path";function We(e,r){let t=g.array(g.string().min(1).max(240)).max(16).parse(r.metadata.artifactAssetIds??[]);if(new Set(t).size!==t.length)throw new Error("An artifact fixture contains duplicate assets.");let c=0;return t.map(s=>{let i=e.files.filter(a=>a.asset.id===s),d=i[0];if(i.length!==1||!d||d.asset.visibility!=="host_private")throw new Error("Artifact fixtures require a unique private package asset.");let m=_(d);if((c+=m.length)>2e7)throw new Error("Artifact fixtures exceed their aggregate byte limit.");return{reference:{id:`fixture/${r.id}/${d.asset.path}`,contentHash:d.asset.contentHash,mediaType:d.asset.mediaType,sizeBytes:d.asset.sizeBytes},title:Q.posix.basename(d.asset.path),contentType:d.asset.mediaType,base64:Buffer.from(m).toString("base64")}})}o(We,"localGraderFixtureArtifacts");export{ue as a,pe as b,b as c,be as d,Je as e,We as f,Ne as g,He as h};

package/dist/chunks/{chunk-55Q64WWF.js → chunk-XD4ZIQZ2.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{b}from"./chunk-5AU5Y6KS.js";import{ja as M}from"./chunk-MVPKU72J.js";import{Qb as x,m as T,qb as A}from"./chunk-QAFXQHHG.js";import{n as c,o as S}from"./chunk-RCT3O3HQ.js";import{A as I}from"./chunk-X2J2P5ZT.js";import{a as R}from"./chunk-FTQLQJQV.js";function j(e,a,i){if(a==="agent_action"){let n=e.find(u=>u.name==="workspace_action_result"&&u.action==="profile_workflow_action")?.output;if(typeof n!="string")throw new Error("Profile Agent action evaluation did not produce a released action result.");return n}if(i==="completed"){let n=[...e].reverse().find(s=>s.name==="turn.completed"),r=n?.data?.providerResponse,o=n?.data;if(!r&&o?.turn?.status==="completed"){let s=[...e].reverse().find(t=>P(t)!==null);if(!s)throw new Error("Native Profile evaluation has no retained final answer.");let d=P(s);return e.filter(t=>P(t)===d).map(t=>t.output??"").join("")}if(typeof r?.requestId!="string"||!r.requestId||typeof r.contentHash!="string")throw new Error("Profile evaluation has no retained final provider response.");let p=e.filter(s=>s.name==="assistant.delta"&&s.data?.providerRequestId===r.requestId).map(s=>s.output??"").join("");if(c(p)!==r.contentHash)throw new Error("Profile evaluation final provider response differs from its retained trace.");return b(e)??p}return e.filter(n=>n.name==="assistant.delta"&&typeof n.output=="string").map(n=>n.output??"").join("")}R(j,"profileEvaluationOutput");function P(e){let a=e.data;return e.name==="assistant.delta"&&a?.provider==="codex"&&a.phase==="final_answer"&&typeof a.itemId=="string"?a.itemId:null}R(P,"nativeFinalAnswerItemId");async function _(e){let{store:a,manifest:i,receipt:n}=e;if(!n.metadata.retainedEvidenceRef)return null;let u=x.parse(n.metadata.retainedEvidenceRef),[r,o]=await Promise.all([a.getSession(u.sessionId),a.getTurn(u.turnId)]),p=r?.metadata?.profileEvaluationRun,s=i.profileEvaluation;if(!r||!o||o.sessionId!==r.id||!s||!r.currentProfile||c(r.currentProfile)!==c(e.profileRef)||c(p)!==c({id:i.id,contentHash:i.contentHash})||r.metadata?.taskId!==n.taskId||r.metadata?.seed!==n.seed||!o.completedAt||o.completedAt!==n.completedAt||o.startedAt!==n.startedAt||!o.harnessSnapshot||o.harnessSnapshot.harnessRelease.id!==s.harnessRelease.id||o.harnessSnapshot.harnessRelease.contentHash!==s.harnessRelease.contentHash||i.policy.kind!=="model"||!o.modelRef||o.modelRef.providerId!==i.policy.model.provider||o.modelRef.modelId!==i.policy.model.model)throw new Error("Retained Profile policy turn differs from its admitted case.");let d=await a.runtimeEventsForTurn(o.id);if(d.some(v=>v.turnId!==o.id)||c(d)!==n.traceHash)throw new Error("Retained Profile policy trace differs from its immutable receipt.");let t=j(d,s.target.kind,o.status);if(n.outputHash!==null&&c({text:t})!==n.outputHash)throw new Error("Retained Profile output differs from its immutable receipt.");let m=n.artifactRefs.flatMap(v=>{let y=d.filter(l=>(l.action==="work_output_save"||l.action==="sandbox_save_output")&&l.status==="completed").flatMap(l=>{let C=l.data&&typeof l.data=="object"?l.data:{},k=M.safeParse(C.outputRef);return k.success?[k.data]:[]}).filter(l=>l.sourceTaskId===r.id&&l.sourceTurnId===o.id&&`${r.id}/${l.id}/${l.revision}/${l.title}`===v.id&&l.sha256===v.contentHash&&l.sizeBytes===v.sizeBytes&&l.contentType===v.mediaType);if(y.length!==1)throw new Error("The case output has no unique sealed saved-file identity.");let h=y[0];if(h.location.kind!=="managed")return[];let H=`/api/work/outputs/${encodeURIComponent(h.location.fileId)}`;if(h.location.downloadPath!==H)throw new Error("The saved output has no canonical managed download route.");return[{id:h.id,title:h.title,revision:h.revision,contentType:h.contentType,sizeBytes:h.sizeBytes,sha256:h.sha256,downloadPath:H}]}),f=await a.listModelUsageRecords({sessionId:r.id,turnId:o.id,limit:1e4}),E=e.eventAfterId?d.findIndex(v=>v.id===e.eventAfterId):-1;if(e.eventAfterId&&E<0)throw new Error("Profile trace cursor is unavailable in this case.");let g=E+1,q=Math.max(1,Math.min(250,e.eventLimit??100)),w=d.slice(g,g+q);return{ref:u,prompt:o.prompt,output:n.outputHash===null?null:{text:t},partialOutput:n.outputHash===null&&t?{text:t}:null,error:o.error,events:w,eventCount:d.length,nextEventCursor:g+w.length<d.length?w.at(-1).id:null,files:m,usage:f,usageCoverage:f.length===1e4?"partial":"complete"}}R(_,"readProfileEvaluationPolicyEvidence");var W=I.object({runId:I.string().trim().min(1).max(240),receiptId:I.string().trim().min(1).max(500).optional(),eventAfterId:I.string().trim().min(1).max(500).optional(),eventLimit:I.number().int().min(1).max(250).optional(),manifest:A.optional()}).strict().refine(e=>e.receiptId||!e.eventAfterId&&!e.eventLimit,"Select a case before requesting its trace page.").refine(e=>!e.manifest||e.receiptId,"Interrupted-run inspection requires one exact case.");async function D(e){let a=await e.store.getProfileEvaluationRun(e.runId);if(a&&c(a.profileRef)!==c(e.profileRef))throw new Error("Profile evaluation run is unavailable in the selected Profile.");let i=a?.manifest??e.manifest;if(!i||!a&&!e.receiptId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");if(a){if(e.manifest&&e.manifest.contentHash!==a.manifest.contentHash)throw new Error("Profile evaluation run differs from its pinned manifest.")}else if(S(i,"Interrupted Profile evaluation manifest"),i.id!==e.runId||i.profileEvaluation?.profileId!==e.profileRef.profileId||e.profileRef.source==="openpond_git"&&i.metadata.profileRepositoryId!==e.profileRef.repositoryId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");let n=i.population.map((r,o)=>({member:r,index:o})).filter(({member:r})=>!e.receiptId||r.receiptId===e.receiptId);if(e.receiptId&&n.length!==1)throw new Error("Profile evaluation case is unavailable in the selected run.");let u=await Promise.all(n.map(async({member:r,index:o})=>{let p=a?null:await e.store.getProfileEvaluationReceipt(r.receiptId),s=a?.receiptRefs[o]??(p?{id:p.id,contentHash:p.contentHash}:null),d=a?.gradeRefs[o]??p?.graderEvidenceRefs[0];if(!s||!d)throw new Error("Profile evaluation run has incomplete case evidence.");let[t,m]=await Promise.all([e.store.getProfileEvaluationReceipt(s.id),e.store.getProfileEvaluationGrade(d.contentHash)]);if(!t||!a&&!T(t)||t.contentHash!==s.contentHash||t.id!==r.receiptId||t.taskId!==r.taskId||t.seed!==r.seed||t.runManifest.id!==i.id||t.runManifest.contentHash!==i.contentHash||!m||m.contentHash!==d.contentHash||!t.graderEvidenceRefs.some(f=>f.id===d.id&&f.contentHash===d.contentHash))throw new Error("Profile evaluation case evidence differs from its retained run.");return{taskId:r.taskId,seed:r.seed,receiptId:t.id,receiptHash:t.contentHash,traceHash:t.traceHash,outputHash:t.outputHash,score:m.score,passed:m.passed,gradingStatus:m.gradingStatus,failureClass:m.failureClass,artifactCount:t.artifactRefs.length,latencyMs:t.latencyMs,costUsd:t.costUsd,terminal:t.terminal,feedback:m.components.map(f=>({graderId:f.graderId,graderVersion:f.graderVersion,score:f.score,passed:f.passed,feedback:f.feedback})),...e.receiptId?{receipt:t,evidence:await _({store:e.store,profileRef:e.profileRef,manifest:i,receipt:t,eventAfterId:e.eventAfterId,eventLimit:e.eventLimit})}:{}}}));return{runId:i.id,sourceRevision:i.profileEvaluation?.sourceRevision??null,cases:u}}R(D,"inspectProfileEvaluationRun");export{j as a,W as b,D as c};
+import{b}from"./chunk-YMCF5GR2.js";import{ja as M}from"./chunk-CJDGLPNX.js";import{Qb as x,m as T,qb as A}from"./chunk-QAFXQHHG.js";import{n as c,o as S}from"./chunk-RCT3O3HQ.js";import{A as I}from"./chunk-X2J2P5ZT.js";import{a as R}from"./chunk-FTQLQJQV.js";function j(e,a,i){if(a==="agent_action"){let n=e.find(u=>u.name==="workspace_action_result"&&u.action==="profile_workflow_action")?.output;if(typeof n!="string")throw new Error("Profile Agent action evaluation did not produce a released action result.");return n}if(i==="completed"){let n=[...e].reverse().find(s=>s.name==="turn.completed"),r=n?.data?.providerResponse,o=n?.data;if(!r&&o?.turn?.status==="completed"){let s=[...e].reverse().find(t=>P(t)!==null);if(!s)throw new Error("Native Profile evaluation has no retained final answer.");let d=P(s);return e.filter(t=>P(t)===d).map(t=>t.output??"").join("")}if(typeof r?.requestId!="string"||!r.requestId||typeof r.contentHash!="string")throw new Error("Profile evaluation has no retained final provider response.");let p=e.filter(s=>s.name==="assistant.delta"&&s.data?.providerRequestId===r.requestId).map(s=>s.output??"").join("");if(c(p)!==r.contentHash)throw new Error("Profile evaluation final provider response differs from its retained trace.");return b(e)??p}return e.filter(n=>n.name==="assistant.delta"&&typeof n.output=="string").map(n=>n.output??"").join("")}R(j,"profileEvaluationOutput");function P(e){let a=e.data;return e.name==="assistant.delta"&&a?.provider==="codex"&&a.phase==="final_answer"&&typeof a.itemId=="string"?a.itemId:null}R(P,"nativeFinalAnswerItemId");async function _(e){let{store:a,manifest:i,receipt:n}=e;if(!n.metadata.retainedEvidenceRef)return null;let u=x.parse(n.metadata.retainedEvidenceRef),[r,o]=await Promise.all([a.getSession(u.sessionId),a.getTurn(u.turnId)]),p=r?.metadata?.profileEvaluationRun,s=i.profileEvaluation;if(!r||!o||o.sessionId!==r.id||!s||!r.currentProfile||c(r.currentProfile)!==c(e.profileRef)||c(p)!==c({id:i.id,contentHash:i.contentHash})||r.metadata?.taskId!==n.taskId||r.metadata?.seed!==n.seed||!o.completedAt||o.completedAt!==n.completedAt||o.startedAt!==n.startedAt||!o.harnessSnapshot||o.harnessSnapshot.harnessRelease.id!==s.harnessRelease.id||o.harnessSnapshot.harnessRelease.contentHash!==s.harnessRelease.contentHash||i.policy.kind!=="model"||!o.modelRef||o.modelRef.providerId!==i.policy.model.provider||o.modelRef.modelId!==i.policy.model.model)throw new Error("Retained Profile policy turn differs from its admitted case.");let d=await a.runtimeEventsForTurn(o.id);if(d.some(v=>v.turnId!==o.id)||c(d)!==n.traceHash)throw new Error("Retained Profile policy trace differs from its immutable receipt.");let t=j(d,s.target.kind,o.status);if(n.outputHash!==null&&c({text:t})!==n.outputHash)throw new Error("Retained Profile output differs from its immutable receipt.");let m=n.artifactRefs.flatMap(v=>{let y=d.filter(l=>(l.action==="work_output_save"||l.action==="sandbox_save_output")&&l.status==="completed").flatMap(l=>{let C=l.data&&typeof l.data=="object"?l.data:{},k=M.safeParse(C.outputRef);return k.success?[k.data]:[]}).filter(l=>l.sourceTaskId===r.id&&l.sourceTurnId===o.id&&`${r.id}/${l.id}/${l.revision}/${l.title}`===v.id&&l.sha256===v.contentHash&&l.sizeBytes===v.sizeBytes&&l.contentType===v.mediaType);if(y.length!==1)throw new Error("The case output has no unique sealed saved-file identity.");let h=y[0];if(h.location.kind!=="managed")return[];let H=`/api/work/outputs/${encodeURIComponent(h.location.fileId)}`;if(h.location.downloadPath!==H)throw new Error("The saved output has no canonical managed download route.");return[{id:h.id,title:h.title,revision:h.revision,contentType:h.contentType,sizeBytes:h.sizeBytes,sha256:h.sha256,downloadPath:H}]}),f=await a.listModelUsageRecords({sessionId:r.id,turnId:o.id,limit:1e4}),E=e.eventAfterId?d.findIndex(v=>v.id===e.eventAfterId):-1;if(e.eventAfterId&&E<0)throw new Error("Profile trace cursor is unavailable in this case.");let g=E+1,q=Math.max(1,Math.min(250,e.eventLimit??100)),w=d.slice(g,g+q);return{ref:u,prompt:o.prompt,output:n.outputHash===null?null:{text:t},partialOutput:n.outputHash===null&&t?{text:t}:null,error:o.error,events:w,eventCount:d.length,nextEventCursor:g+w.length<d.length?w.at(-1).id:null,files:m,usage:f,usageCoverage:f.length===1e4?"partial":"complete"}}R(_,"readProfileEvaluationPolicyEvidence");var W=I.object({runId:I.string().trim().min(1).max(240),receiptId:I.string().trim().min(1).max(500).optional(),eventAfterId:I.string().trim().min(1).max(500).optional(),eventLimit:I.number().int().min(1).max(250).optional(),manifest:A.optional()}).strict().refine(e=>e.receiptId||!e.eventAfterId&&!e.eventLimit,"Select a case before requesting its trace page.").refine(e=>!e.manifest||e.receiptId,"Interrupted-run inspection requires one exact case.");async function D(e){let a=await e.store.getProfileEvaluationRun(e.runId);if(a&&c(a.profileRef)!==c(e.profileRef))throw new Error("Profile evaluation run is unavailable in the selected Profile.");let i=a?.manifest??e.manifest;if(!i||!a&&!e.receiptId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");if(a){if(e.manifest&&e.manifest.contentHash!==a.manifest.contentHash)throw new Error("Profile evaluation run differs from its pinned manifest.")}else if(S(i,"Interrupted Profile evaluation manifest"),i.id!==e.runId||i.profileEvaluation?.profileId!==e.profileRef.profileId||e.profileRef.source==="openpond_git"&&i.metadata.profileRepositoryId!==e.profileRef.repositoryId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");let n=i.population.map((r,o)=>({member:r,index:o})).filter(({member:r})=>!e.receiptId||r.receiptId===e.receiptId);if(e.receiptId&&n.length!==1)throw new Error("Profile evaluation case is unavailable in the selected run.");let u=await Promise.all(n.map(async({member:r,index:o})=>{let p=a?null:await e.store.getProfileEvaluationReceipt(r.receiptId),s=a?.receiptRefs[o]??(p?{id:p.id,contentHash:p.contentHash}:null),d=a?.gradeRefs[o]??p?.graderEvidenceRefs[0];if(!s||!d)throw new Error("Profile evaluation run has incomplete case evidence.");let[t,m]=await Promise.all([e.store.getProfileEvaluationReceipt(s.id),e.store.getProfileEvaluationGrade(d.contentHash)]);if(!t||!a&&!T(t)||t.contentHash!==s.contentHash||t.id!==r.receiptId||t.taskId!==r.taskId||t.seed!==r.seed||t.runManifest.id!==i.id||t.runManifest.contentHash!==i.contentHash||!m||m.contentHash!==d.contentHash||!t.graderEvidenceRefs.some(f=>f.id===d.id&&f.contentHash===d.contentHash))throw new Error("Profile evaluation case evidence differs from its retained run.");return{taskId:r.taskId,seed:r.seed,receiptId:t.id,receiptHash:t.contentHash,traceHash:t.traceHash,outputHash:t.outputHash,score:m.score,passed:m.passed,gradingStatus:m.gradingStatus,failureClass:m.failureClass,artifactCount:t.artifactRefs.length,latencyMs:t.latencyMs,costUsd:t.costUsd,terminal:t.terminal,feedback:m.components.map(f=>({graderId:f.graderId,graderVersion:f.graderVersion,score:f.score,passed:f.passed,feedback:f.feedback})),...e.receiptId?{receipt:t,evidence:await _({store:e.store,profileRef:e.profileRef,manifest:i,receipt:t,eventAfterId:e.eventAfterId,eventLimit:e.eventLimit})}:{}}}));return{runId:i.id,sourceRevision:i.profileEvaluation?.sourceRevision??null,cases:u}}R(D,"inspectProfileEvaluationRun");export{j as a,W as b,D as c};

package/dist/chunks/{chunk-W22KWKF6.js → chunk-XDFXKRTI.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{Ca as b,ja as w,wa as y}from"./chunk-WLZJOCLA.js";import{i as f,n}from"./chunk-RCT3O3HQ.js";import{a as h}from"./chunk-FTQLQJQV.js";function I(r,a){if(!a)return null;let g=w(a),m=(r.environment.resources??[]).filter(o=>o.kind==="file"&&o.visibility==="policy_visible");if(!m.length)return null;let k=m.map(o=>{let l=g.files.filter(s=>s.asset.path===o.path&&s.asset.visibility==="policy");if(l.length!==1)throw new Error("Reviewed shared resource differs from the exact owner-local package.");let c=f.parse(l[0].asset),t=r.sourceRefs.find(s=>"sourceFileHashes"in s&&s.sourceFileHashes?.includes(c.contentHash)&&"originalFileNames"in s&&s.originalFileNames?.includes(c.path.split("/").at(-1))&&"mediaTypes"in s&&s.mediaTypes?.includes(c.mediaType));if(!t)throw new Error("Shared workflow files require existing source review of their exact bytes.");return{asset:c,sourceRefId:t.id}});return new Map(k.map(o=>[o.asset.id,o]))}h(I,"reviewedModelSharedResources");function O(r,a,g){if(r.profileId!==a.profileId||r.tasks.length!==a.tasks.length)throw new Error("Model Taskset review differs from its compiled task scope.");let m=new Map(a.sourceRefs.map(t=>[t.id,t])),k=I(a,g),o=r.tasks.map(t=>{let s=a.tasks.find(e=>e.id===t.id);if(!s)throw new Error("Model Taskset review is missing a compiled task.");if(n(t.input)!==n(s.input)||n(t.expectedOutput)!==n(s.expectedOutput)||n(t.policyVisibleContext)!==n(s.policyVisibleContext))throw new Error("Model Taskset review differs from its compiled task content.");let T=(t.assets??[]).map(e=>{let i=s.assets?.find(u=>u.id===e.id);if(!i){let u=k?.get(e.id),d=u?.asset,R=f.safeParse(e.metadata.portableAsset);if(!d||!R.success||n(R.data)!==n(d)||e.artifactRef!==d.path||e.stagingPath!==d.path||e.sha256!==d.contentHash||e.sizeBytes!==d.sizeBytes||e.mediaType!==d.mediaType||e.fileName!==d.path.split("/").at(-1)||e.split!==t.split)throw new Error("Model Taskset review differs from its compiled shared resource.");return{...e,sourceRefId:u.sourceRefId}}let p=f.safeParse(i?.metadata.portableAsset),P=i?.artifactRef===e.artifactRef||p.success&&p.data.path===e.artifactRef&&p.data.id===e.id&&p.data.contentHash===e.sha256&&p.data.sizeBytes===e.sizeBytes&&p.data.mediaType===e.mediaType&&p.data.visibility==="policy";if(!i||i.sha256!==e.sha256||i.sizeBytes!==e.sizeBytes||!P||i.mediaType!==e.mediaType||i.fileName!==e.fileName||i.split!==t.split||!s.sourceRefs.includes(i.sourceRefId)||!m.has(i.sourceRefId))throw new Error("Model Taskset review differs from its compiled input asset.");return{...e,sourceRefId:i.sourceRefId,...i.stagingPath===void 0?{}:{stagingPath:i.stagingPath}}});if((s.assets??[]).some(e=>!T.some(i=>i.id===e.id)))throw new Error("Model Taskset review lost an input asset.");return{...t,sourceRefs:[...new Set([...s.sourceRefs,...T.map(e=>e.sourceRefId)])],assets:T}}),l=r.metadata.rewardBinding,c=y.parse({...r,sourceRefs:a.sourceRefs,tasks:o,metadata:{...r.metadata,...Object.fromEntries(["trainingMethod","diagnosis","buildIntent"].filter(t=>a.metadata[t]!==void 0).map(t=>[t,a.metadata[t]]))},capabilities:{...r.capabilities,supportedSignals:a.capabilities.supportedSignals,compatibleMethods:a.capabilities.compatibleMethods},graderFixtures:a.graderFixtures,learningSignals:{...a.learningSignals,rewards:a.learningSignals.rewards.map(t=>l?{...t,artifactRef:l.id,rules:[{id:l.id,points:1,condition:"Execute the published Reward binding with its declared normalization, weights and required gates."}]}:t)}});return y.parse({...c,contentHash:b(c)})}h(O,"retainModelTasksetInputReview");export{O as a};
+import{Ca as b,ja as w,wa as y}from"./chunk-5ZNE43XC.js";import{i as f,n}from"./chunk-RCT3O3HQ.js";import{a as h}from"./chunk-FTQLQJQV.js";function I(r,a){if(!a)return null;let g=w(a),m=(r.environment.resources??[]).filter(o=>o.kind==="file"&&o.visibility==="policy_visible");if(!m.length)return null;let k=m.map(o=>{let l=g.files.filter(s=>s.asset.path===o.path&&s.asset.visibility==="policy");if(l.length!==1)throw new Error("Reviewed shared resource differs from the exact owner-local package.");let c=f.parse(l[0].asset),t=r.sourceRefs.find(s=>"sourceFileHashes"in s&&s.sourceFileHashes?.includes(c.contentHash)&&"originalFileNames"in s&&s.originalFileNames?.includes(c.path.split("/").at(-1))&&"mediaTypes"in s&&s.mediaTypes?.includes(c.mediaType));if(!t)throw new Error("Shared workflow files require existing source review of their exact bytes.");return{asset:c,sourceRefId:t.id}});return new Map(k.map(o=>[o.asset.id,o]))}h(I,"reviewedModelSharedResources");function O(r,a,g){if(r.profileId!==a.profileId||r.tasks.length!==a.tasks.length)throw new Error("Model Taskset review differs from its compiled task scope.");let m=new Map(a.sourceRefs.map(t=>[t.id,t])),k=I(a,g),o=r.tasks.map(t=>{let s=a.tasks.find(e=>e.id===t.id);if(!s)throw new Error("Model Taskset review is missing a compiled task.");if(n(t.input)!==n(s.input)||n(t.expectedOutput)!==n(s.expectedOutput)||n(t.policyVisibleContext)!==n(s.policyVisibleContext))throw new Error("Model Taskset review differs from its compiled task content.");let T=(t.assets??[]).map(e=>{let i=s.assets?.find(u=>u.id===e.id);if(!i){let u=k?.get(e.id),d=u?.asset,R=f.safeParse(e.metadata.portableAsset);if(!d||!R.success||n(R.data)!==n(d)||e.artifactRef!==d.path||e.stagingPath!==d.path||e.sha256!==d.contentHash||e.sizeBytes!==d.sizeBytes||e.mediaType!==d.mediaType||e.fileName!==d.path.split("/").at(-1)||e.split!==t.split)throw new Error("Model Taskset review differs from its compiled shared resource.");return{...e,sourceRefId:u.sourceRefId}}let p=f.safeParse(i?.metadata.portableAsset),P=i?.artifactRef===e.artifactRef||p.success&&p.data.path===e.artifactRef&&p.data.id===e.id&&p.data.contentHash===e.sha256&&p.data.sizeBytes===e.sizeBytes&&p.data.mediaType===e.mediaType&&p.data.visibility==="policy";if(!i||i.sha256!==e.sha256||i.sizeBytes!==e.sizeBytes||!P||i.mediaType!==e.mediaType||i.fileName!==e.fileName||i.split!==t.split||!s.sourceRefs.includes(i.sourceRefId)||!m.has(i.sourceRefId))throw new Error("Model Taskset review differs from its compiled input asset.");return{...e,sourceRefId:i.sourceRefId,...i.stagingPath===void 0?{}:{stagingPath:i.stagingPath}}});if((s.assets??[]).some(e=>!T.some(i=>i.id===e.id)))throw new Error("Model Taskset review lost an input asset.");return{...t,sourceRefs:[...new Set([...s.sourceRefs,...T.map(e=>e.sourceRefId)])],assets:T}}),l=r.metadata.rewardBinding,c=y.parse({...r,sourceRefs:a.sourceRefs,tasks:o,metadata:{...r.metadata,...Object.fromEntries(["trainingMethod","diagnosis","buildIntent"].filter(t=>a.metadata[t]!==void 0).map(t=>[t,a.metadata[t]]))},capabilities:{...r.capabilities,supportedSignals:a.capabilities.supportedSignals,compatibleMethods:a.capabilities.compatibleMethods},graderFixtures:a.graderFixtures,learningSignals:{...a.learningSignals,rewards:a.learningSignals.rewards.map(t=>l?{...t,artifactRef:l.id,rules:[{id:l.id,points:1,condition:"Execute the published Reward binding with its declared normalization, weights and required gates."}]}:t)}});return y.parse({...c,contentHash:b(c)})}h(O,"retainModelTasksetInputReview");export{O as a};