openpond 0.2.58 → 0.2.59

package/dist/chunks/{chunk-NH6UUX5S.js → chunk-BRORSFJX.js}

@@ -1,6 +1,6 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as vs,b as Ts}from"./chunk-HDGTHRY7.js";import{a as ps,c as fs}from"./chunk-XVDFJ3ZF.js";import{a as oe}from"./chunk-NC27T4AB.js";import{i as ms,j as pt}from"./chunk-HCYX6RHP.js";import{d as je,e as it}from"./chunk-TRQQNFMO.js";import{a as ce}from"./chunk-I2NXNMRZ.js";import{a as Yr,b as Kr,c as Xr,d as dt,f as Zr,g as Qr,h as ct}from"./chunk-O72BSKSO.js";import{b as Br,e as at,f as nt}from"./chunk-DBK66CJP.js";import{E as ys,N as ws,a as ft,b as gs,c as gt,d as ht,e as hs,f as ks}from"./chunk-RNQF63ZR.js";import{g as ut,i as mt}from"./chunk-UKMA4R6X.js";import{a as ye}from"./chunk-7KWD2I6W.js";import{g as qe}from"./chunk-KG325WTN.js";import{s as Gr}from"./chunk-EIRYMSEG.js";import{A as rs,C as ss,t as ts}from"./chunk-QL4OIPEZ.js";import{q as es}from"./chunk-WJ43MJ3F.js";import{G as B,H as os,I as as,J as ns,K as lt,L as is,M as q,N as ds,O as Ne,P as cs,R as ls,S as Lt,T as us}from"./chunk-QOGJAQDO.js";import{l as $t}from"./chunk-EYPCBBMS.js";import{p as Vr,r as F,s as zr,u as Jr}from"./chunk-PRGBLQVH.js";import{f as Fr}from"./chunk-XDQMGUWL.js";import{Ib as ot,M as Pr,N as Or,O as Hr,P as Dr,R as jr,ca as Nr,da as qr,ka as $r,sa as Lr,ua as ke,va as Ur,wa as Wr,xa as Te}from"./chunk-25MBPJVY.js";import{n as se}from"./chunk-523TE4BI.js";import{C as Sr,S as Er,T as Cr,Tb as Ar}from"./chunk-Y4OWBBDD.js";import{f as yr}from"./chunk-JFXWI3EN.js";import{$ as De,Kb as Ir,N as U,O as _r,V as st,W as xr,c as qt,d as re,g as br,j as Rr,lb as ge,mb as he}from"./chunk-VYLGI3IH.js";import{La as ve,Ma as Tr,l as ze,oa as rt}from"./chunk-ACYFMLWG.js";import{l as wr,m as vr}from"./chunk-TVQAPBLZ.js";import{h as kr,l as fe,m as h}from"./chunk-KGNCKIFP.js";import{b as O}from"./chunk-ES2ISFJC.js";import{n as Mr}from"./chunk-H7IWJN7L.js";import{a as i}from"./chunk-FTQLQJQV.js";import{randomUUID as Vo}from"node:crypto";function yt(e){let t=kt(kt(kt(e.metadata).billing).pricing);return{version:$e(t.version,"pricing version"),source:$e(t.source,"pricing source"),effectiveAt:$e(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Le(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Le(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Le(t.outputUsdPerMillionTokens,"output token rate")}}i(yt,"hostedTokenPricingFromCatalog");function Fn(e){let t=kt(e);try{return{version:$e(t.version,"pricing version"),source:$e(t.source,"pricing source"),effectiveAt:$e(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Le(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Le(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Le(t.outputUsdPerMillionTokens,"output token rate")}}catch{return null}}i(Fn,"hostedTokenPricingFromValue");function Se(e,t){let r=ye(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}i(Se,"hostedUsageCostUsd");function kt(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(kt,"record");function $e(e,t){if(typeof e=="string"&&e.trim())return e.trim();throw new Error(`Hosted model catalog is missing ${t}.`)}i($e,"requiredString");function Le(e,t){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return e;throw new Error(`Hosted model catalog is missing ${t}.`)}i(Le,"requiredRate");function Rs(e={}){let t=e.catalog??ut,r=e.stream??mt;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: ${a.error}`);let o=a.models.find(l=>l.id===s.modelId);if(!o)throw new Error("The selected judge model is unavailable.");let n=Ss(o.raw),c=yt(n),u=bs(n.context_window,"context window"),g=Math.min(4096,bs(n.output_limit,"output limit")),p=(u*Math.max(c.inputUsdPerMillionTokens,c.cachedInputUsdPerMillionTokens)+g*c.outputUsdPerMillionTokens)/1e6;if(!(p>0))throw new Error("The judge model must have a positive known charge ceiling.");return{maximumChargeUsd:p,async dispatch(l){l?.throwIfAborted();let d="",m=null,f=null,T=null,y=!1;for await(let A of r({model:s.modelId,messages:[{role:"system",content:s.system},{role:"user",content:s.data}],requestId:`learning-judge-${Vo()}`,maxTokens:g,temperature:s.temperature,signal:l})){let I=Ss(A.raw);if(typeof I.model=="string"){if(m!==null&&I.model!==m)throw new Error("Judge response model identity changed during streaming.");m=I.model}if(typeof I.id=="string"){if(f!==null&&f!==I.id)throw new Error("Judge response identity changed during streaming.");f=I.id}if(A.type==="text_delta"&&(d+=A.text),d.length>1e5)throw new Error("Judge response exceeds the retained response limit.");A.type==="usage"&&(T=A.usage),A.type==="finish"&&(y=!0)}if(!y||!m||!f)throw new Error("Judge response did not include a complete model receipt.");let w=ye(T),_=w.promptTokens!==null&&w.completionTokens!==null?Se(T,c):null;return{text:d,modelId:m,modelRevision:null,responseId:f,inputTokens:w.promptTokens,outputTokens:w.completionTokens,costUsd:_}}}},async cancel({run:s}){return!s.judgeCalls?.some(a=>a.status==="reserved")}}}i(Rs,"createLearningHostedJudgeProvider");function Ss(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(Ss,"record");function bs(e,t){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new Error(`Judge catalog is missing its ${t}.`);return e}i(bs,"positiveInteger");function oi(e,t=Rs()){return{execute(r){if(r.definition.execution.environment.entrypoint==="openpond.javascript-environment.v1")throw new Error("Tool Tasksets require grading an owner-recorded environment attempt.");let s=Ir({repository:e,scope:r.scope,run:r.run});return Tr({binding:r.binding,rewards:r.rewards,task:st(r.evidence,r.definition),evidence:xr(r.evidence,r.run.output),evaluatorContext:r.evidence.submission.evaluatorContext,signal:r.signal,modelJudge:br({readRubric:i(a=>e.transaction(r.scope,async o=>re(await ge(o,"asset",a.id,1),a)),"readRubric"),async executeBudgeted(a,o){let n=await t.prepare(a,{scope:r.scope,run:r.run});return Rr({store:s,maximumCharge:i(()=>n.maximumChargeUsd,"maximumCharge"),dispatch:i((u,g)=>n.dispatch(g),"dispatch")})(`grade-${h(a)}`,a,o)}}),customVerifier:i(async({grader:a,task:o,evidence:n})=>{let c=await Je({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:c.score,passed:c.passed,rewardEligible:a.rewardEligible,failureClass:null,feedback:[c.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:c.evidenceRefs}},"customVerifier")})},async cancel({scope:r,run:s}){let a=await e.transaction(r,o=>ge(o,"grade",s.id));return a.judgeCalls?.some(o=>o.status==="reserved")?!1:t.cancel({scope:r,run:a})}}}i(oi,"createLocalTaskGradeExecutor");async function Je(e){let t=await e.repository.transaction(e.scope,async r=>re(await ge(r,"asset",e.grader.verifierRef.id,1),e.grader.verifierRef));return je({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}})}i(Je,"executeLocalLearningVerifier");import zo from"node:path";async function _s(e,t,r){if(!r)throw new Error("Ordinary tool execution requires its owner's package directory.");let s=ot({taskset:e,adapterId:dt(e)}),a=await Zr({root:zo.join(r,"training","tasksets",ce(e)),content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:s.tasksetRelease,environment:s.environmentRelease,verifierSet:s.verifierSetRelease},sources:Fr.parse(e.metadata.portableFileInventory)}),o=qr(a);if(!o)throw new Error("The Taskset does not declare a JavaScript execution graph.");let n=a.taskset.tasks.find(g=>g.id===t.id);if(!n)throw new Error("The selected task is missing from its executable package.");let c=o.assets.find(g=>g.id===o.execution.javascript.module.id),u=o.assets.find(g=>g.id===n.privilegedContextRef);return{execution:o.execution,module:c,initialState:JSON.parse(u.text),definition:U(a.taskset),instructions:$r(a),task:n}}i(_s,"loadOrdinaryToolEnvironment");var Y="openpond.javascript-environment.v1";async function Ue(e,t,r,s){let a=t.tasks.find(o=>o.id===r.id);if(!a||h(a)!==h(r))throw new Error("Tool task differs from the selected immutable Taskset.");return t.metadata.taskDefinition===void 0?_s(t,r,s):e.learningRepository().transaction(t.profileId,async o=>{let n=await he(o,"definition",se.parse(t.metadata.taskDefinition)),c=n.execution.environment;if(c.entrypoint!==Y||t.environment.entrypoint!==Y||t.environment.kind!=="agent"||t.environment.stateful!==c.stateful||t.environment.defaultTimeoutMs!==c.defaultTimeoutMs||t.environment.networkPolicy!==c.networkPolicy||t.environment.deterministicSeeds!==c.deterministicSeeds||h(t.environment.toolNames)!==h(n.execution.tools.map(S=>S.name))||h(t.environment.metadata.portableTools)!==h(n.execution.tools)||h(t.policy)!==h(n.execution.policy)||h(t.metadata.rewardBinding)!==h(n.rewardBinding))throw new Error("Tool Taskset context differs from its pinned task definition.");let u=await he(o,"binding",n.rewardBinding),g=await Promise.all(u.sources.map(S=>he(o,"reward",S.reward))),p=rt.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:ve(u,g),tasks:[p]},d=await ge(o,"asset",Dr(l),1),m=Or.parse(JSON.parse(re(d,d.asset))),f=[m.javascript.module,m.environment.actionSchemaRef,m.environment.observationSchemaRef,m.environment.stateSchemaRef].filter(S=>S!==null),T=[...new Set([...f.map(S=>S.id),r.privilegedContextRef].filter(S=>S!==null))],y=await Promise.all(T.map(S=>ge(o,"asset",S,1))),w=jr(l,d,y);if(h(t.environment.metadata.portableExecutionResources)!==h({environment:w.environment,verifierSet:w.verifierSet}))throw new Error("Tool Taskset publication differs from its private execution resources.");let _=y.find(S=>S.id===w.javascript.module.id),A=y.find(S=>S.id===r.privilegedContextRef),I=JSON.parse(re(A,A.asset));return{execution:w,module:_,initialState:I,definition:U(n),instructions:n.instructions,task:p}})}i(Ue,"loadStarterToolEnvironment");function xs(e){let t=qt({text:e.instructions,path:"policy/instructions.txt",mediaType:"text/plain",visibility:"policy"}),r=qt({text:JSON.stringify({environment:U(e.execution.environment),verifierSet:U(e.execution.verifierSet),javascript:U(e.execution.javascript)}),path:"environment/dependencies.json",mediaType:"application/json",visibility:"host_private"}),s=wr({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:Y}}),a=vr({schemaVersion:"openpond.harnessRelease.v2",id:`starter-harness-${s.contentHash}`,agentSnapshot:{id:s.id,contentHash:s.contentHash},program:e.module.asset,tools:e.execution.javascript.tools,lifecycle:{create:!0,reset:!0,step:!0,collect:!0,destroy:!0,resetScope:"attempt"},graderInterface:{visibleEvidence:["output"],privilegedEvidence:["environment_state","private_verifier"],privateVerifierIsolation:!0},files:[t.asset,r.asset,e.module.asset],metadata:{execution:Y}});return{agentSnapshot:s,harnessRelease:a,instructionContext:void 0}}i(xs,"compileStarterToolHarness");import wt from"node:path";import{constants as Is}from"node:fs";import{open as Jo,realpath as Es}from"node:fs/promises";var Go=O.string().regex(/^[a-f0-9]{64}$/),Yo=O.object({schemaVersion:O.literal("openpond.starterToolAttempt.v1"),requestId:O.string(),taskset:se,taskHash:Go,model:O.unknown(),seed:O.number().int(),attempt:O.number().int(),definition:se,policySource:O.enum(["model","fixture"]),environment:se,verifierSet:se,result:Pr}).strict();async function vt(e){let{store:t,taskset:r,task:s,attempt:a}=e,o=await Ue(t,r,s,e.storeDir),c=(await t.listTaskAttemptArtifacts({attemptId:a.id})).filter(_=>_.kind==="environment_state"&&_.tasksetId===r.id&&_.taskId===s.id&&_.attemptId===a.id&&a.artifactRefs.includes(_.id)&&a.privilegedOutcomeRef===_.id);if(c.length!==1)throw new Error("Tool grading requires exactly one owner-recorded environment artifact.");let u=c[0],g=await Es(wt.join(e.storeDir,"training","evaluation-artifacts")),p=await Es(wt.dirname(u.path)),l=wt.relative(g,p);if(!l||l.startsWith("..")||wt.isAbsolute(l))throw new Error("Tool evidence is outside the execution owner's artifact directory.");let d=await Jo(u.path,Is.O_RDONLY|Is.O_NOFOLLOW),m;try{let _=await d.stat();if(!_.isFile()||_.size>33554432||_.size!==u.sizeBytes)throw new Error("Tool evidence exceeds its declared byte boundary.");m=await d.readFile()}finally{await d.close()}if(m.byteLength!==u.sizeBytes||fe(m)!==u.sha256)throw new Error("Owner-recorded tool evidence bytes changed.");let f=Yo.parse(JSON.parse(m.toString("utf8"))),{contentHash:T,...y}=f.result,w=f.result;if(h(y)!==T||!w.snapshot||!w.environmentCleanupComplete||!["completed","budget_exhausted","policy_failure"].includes(w.status)||h(f.taskset)!==h(U(r))||f.taskHash!==h(s)||h(f.model)!==h(a.modelRef)||f.seed!==a.seed||f.attempt!==a.attempt||a.tasksetId!==r.id||a.taskId!==s.id||a.split!==s.split||h(f.definition)!==h(o.definition)||h(f.environment)!==h(U(o.execution.environment))||h(f.verifierSet)!==h(U(o.execution.verifierSet))||w.taskId!==s.id||w.snapshot.seed!==a.seed||w.snapshot.inputHash!==h(s.input)||h(w.snapshot.definition)!==h(U(o.execution.javascript))||w.snapshot.finalStateHash!==h(w.snapshot.state)||h(a.output)!==h({text:w.output??""})||a.metadata.environmentAttemptHash!==T)throw new Error("Tool evidence differs from the selected task, model or recorded attempt.");if(a.metadata.policySource!==f.policySource)throw new Error("Tool evidence policy source changed.");return{environment:{status:w.status,collected:w.collected,definition:w.snapshot.definition,initialStateHash:w.snapshot.initialStateHash,finalStateHash:w.snapshot.finalStateHash,finalState:w.snapshot.state,events:w.snapshot.events}}}i(vt,"readStarterToolEvidence");async function be(e,t){if(t.metadata.learning!==void 0)return De.parse(t.metadata.learning);if(t.metadata.rewardBinding===void 0)return;let r=se.parse(t.metadata.rewardBinding);return e.learningRepository().transaction(t.profileId,async s=>{let a=await he(s,"binding",r),o=[...new Map(a.sources.map(c=>[`${c.reward.id}:${c.reward.revision}:${c.reward.contentHash}`,c.reward])).values()],n=await Promise.all(o.map(c=>he(s,"reward",c)));return{binding:a,rewards:n}})}i(be,"resolveTasksetRewardBinding");async function Ko(e,t,r){let s=await be(e,t);if(!s)return{rewardExecution:void 0,verifierAssets:[]};let a=ve(s.binding,s.rewards).flatMap(c=>c.kind==="custom_verifier"?[c.verifierRef]:c.kind==="model_judge"?[c.rubricRef]:[]),o=t.metadata.learning===void 0?void 0:await ct(r,t);if(o){let c=[...new Map(a.map(u=>{let g=o.learningResources.assets.find(p=>p.id===u.id);if(!g)throw new Error("Imported batch verifier is missing from its package.");return re(g,u),[g.id,g]})).values()];return{rewardExecution:s,verifierAssets:c}}let n=await e.learningRepository().transaction(t.profileId,async c=>{let u=new Map;for(let g of a){let p=await ge(c,"asset",g.id,1);re(p,g),u.set(p.id,p)}return[...u.values()]});return{rewardExecution:s,verifierAssets:n}}i(Ko,"resolveTasksetTrainingReward");async function Ji(e,t,r){let s=await Ko(e,t,r.storeDir);if(!s.rewardExecution)return s;let a=ve(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}i(Ji,"resolveManagedTasksetReward");async function Cs(e,t,r){let s=await be(e,t);if(!s)throw new Error("The Taskset has no published Reward binding.");let a=ve(s.binding,s.rewards);return async({grader:o,task:n,attempt:c,signal:u})=>{let g=t.tasks.find(m=>m.id===n.id);if(!g||h(g)!==h(n))throw new Error("Task differs from the selected immutable Taskset.");let p=null;if(t.environment.entrypoint===Y){if(!r)throw new Error("Tool grading requires the execution owner's artifact directory.");p=await vt({store:e,storeDir:r,taskset:t,task:n,attempt:c})}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 d=rt.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 Je({repository:e.learningRepository(),scope:t.profileId,grader:l,task:d,evidence:{output:c.output,artifactRefs:c.artifactRefs,runtimeEventRefs:c.runtimeEventRefs,infrastructureError:c.infrastructureError},evaluatorContext:p,signal:u})}}i(Cs,"createTasksetBindingVerifier");async function Ee(e,t,r){let s=r?se.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||Te(a)!==s.contentHash))throw new Error("Taskset execution revision is unavailable or differs from its immutable hash.");return a}i(Ee,"requireLocalTasksetRevision");async function ed(e,t,r){let s=await e.releaseForTaskset(t);return s||ot({taskset:t,rewardExecution:r?await be(r,t):void 0,adapterId:dt(t)}).tasksetRelease}i(ed,"requireReleasedTaskset");import{createHash as Xo}from"node:crypto";function As(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||h(r.metadata.portableTaskRecord)!==h(s))throw new Error("The Work task differs from its exact released population.");let a=0,o=(r.assets??[]).map(n=>{let c=s.artifactRefs.find(l=>l.id===n.id&&l.contentHash===n.sha256&&l.path===n.artifactRef),u=c&&e.files.find(l=>h(l.asset)===h(c));if(!c||!u||c.visibility!=="policy")throw new Error("The Work task requests an unavailable or private evaluator file.");let g=Buffer.from(Nr(u));if(Xo("sha256").update(g).digest("hex")!==n.sha256||g.length!==n.sizeBytes||n.split!==r.split)throw new Error("The published Work input manifest changed.");if(a+=g.length,a>25e7)throw new Error("Published Work inputs exceed the admitted 250 MB ceiling.");return{...n,assetId:n.id,storageName:n.fileName,bytes:g,sizeBytes:g.length}});if(o.length!==s.artifactRefs.length)throw new Error("The exact policy-visible Work asset population is incomplete.");return o}i(As,"advancedPublishedWorkAssets");async function Ut(e,t=ut,r=!1){if(e.kind==="hosted_harness")return Ut({kind:"hosted_chat",modelId:e.modelId,maxOutputTokens:4096,temperature:0,topP:1},t);if(e.kind!=="hosted_chat")throw new oe("local_target_not_qualified","Select an explicit model target for local execution.",422);if(e.harness&&!r)throw new oe("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 oe("local_model_catalog_unavailable",s.error,422);let a=s.models.find(d=>d.id===e.modelId);if(!a)throw new oe("local_model_unavailable","The selected model is unavailable.",422);let o=Ms(a.raw),n=yt(o),c=Ms(o.capabilities).samplingParameters===!0;if(!c&&(e.temperature!==0||e.topP!==1))throw new oe("local_sampling_unsupported","This model uses provider-managed sampling. Custom temperature or top P requires an explicit supported capability.",422);let u=Ps(o.context_window),g=Ps(o.output_limit);if(e.maxOutputTokens>g)throw new oe("local_model_output_limit","The model cannot accept the configured output limit.",422);let p={providerId:"openpond",modelId:e.modelId,maxOutputTokens:e.maxOutputTokens,...c?{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 oe("local_model_charge_unknown","The model must publish an enforceable positive request ceiling.",422);return{model:{...p,configurationHash:h({configuration:p,contextWindow:u,outputLimit:g,pricing:n})},pricing:n,maximumChargeUsd:l}}i(Ut,"prepareLocalExperimentModel");function fd(e){let t=e.admission.streamFactory?.(e)??e.admission.stream??e.stream??mt;return async function*(r){if(!r.requestId)throw new oe("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 oe("local_model_configuration_conflict","The native model request changed its admitted model, output limit or sampling configuration.",422);let a=ps();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,c=!1,u=!1,g;try{for await(let p of t(r)){u=!0,p.type==="usage"&&(n=p.usage),p.type==="finish"&&(c=!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:ye(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 fs&&p.executionId===o.id&&p.requestId===o.requestId&&(g=p),p}finally{let p=ye(n),l=e.admission.pricing&&c&&p.promptTokens!==null&&p.completionTokens!==null?Se(n,e.admission.pricing):null;await e.store.settleLocalCharge({...o,costUsd:l,confirmedTransportNotInvoked:g,usage:n===null?null:{promptTokens:p.promptTokens,completionTokens:p.completionTokens,cachedPromptTokens:p.cachedPromptTokens,uncachedPromptTokens:p.uncachedPromptTokens}})}}}i(fd,"createLocalBudgetedModelStream");function Ms(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(Ms,"record");function Ps(e){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new oe("local_model_bound_unknown","The model catalog is missing an enforceable context/output bound.",422);return e}i(Ps,"positive");import{AsyncLocalStorage as Zo}from"node:async_hooks";import Qo from"node:path";var Os=new Zo;function Hs(){return Os.getStore()}i(Hs,"currentAdvancedRefinerBoundary");function Rd(e){let t=Mr(Qo.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}}i(r,"transaction");async function s(a,o,n){let c=$t(o);await e.authorize(c),r(()=>{let l=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(l){if(l.actor_id!==c.actorId||l.team_id!==c.teamId||l.pin_hash!==c.contentHash)throw new Error("Advanced Run budget belongs to another immutable admission.");return}t.prepare("INSERT INTO advanced_refiner_budgets VALUES(?,?,?,?,?,?,?)").run(a,c.actorId,c.teamId,c.contentHash,c.maximumCostUsd,c.maximumModelSteps,new Date(Date.now()+c.maximumDurationMs).toISOString())});let u=t.prepare("SELECT expires_at FROM advanced_refiner_budgets WHERE id=?").get(a),g=Date.parse(String(u.expires_at))-Date.now();if(g<=0)throw new Error("The immutable advanced evaluation duration is exhausted.");let p={runId:a,pin:c,signal:AbortSignal.timeout(g),authorize:i(()=>e.authorize(c),"authorize"),async resolveWorkAssets(l,d){await e.authorize(c),p.signal.throwIfAborted();let m=await e.packageForPin(c),f=As(m,l,d);return await e.authorize(c),p.signal.throwIfAborted(),f},async call(l,d,m,f,T){if(await e.authorize(c),!Number.isInteger(f)||f<1)throw new Error("Advanced paid requests require a finite enforced token ceiling.");let y=d.model;if(y?.providerId!=="openpond"||!y.modelId)throw new Error("Advanced paid calls require an explicitly qualified hosted model.");let w=y.modelId,_=e.prepareModel?await e.prepareModel(w,f,m):await(async()=>{let k=await Ut({kind:"hosted_chat",modelId:w,maxOutputTokens:f,temperature:0,topP:1});if(h(k.pricing)!==h(m))throw new Error("The actual model pricing differs from its reviewed request pin.");return{configurationHash:k.model.configurationHash,maximumChargeUsd:k.maximumChargeUsd}})();if(!/^[a-f0-9]{64}$/.test(_.configurationHash)||!Number.isFinite(_.maximumChargeUsd)||_.maximumChargeUsd<=0)throw new Error("The actual owner model ceiling is unavailable.");let A=h({intent:d,pricing:m,maximumOutputTokens:f,modelConfigurationHash:_.configurationHash}),I=_.maximumChargeUsd,S=r(()=>{let k=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(!k||String(k.expires_at)<=new Date().toISOString())throw new Error("The immutable advanced evaluation duration is exhausted.");let x=t.prepare("SELECT * FROM advanced_refiner_calls WHERE run_id=? AND id=?").get(a,l);if(x){if(x.intent_hash!==A)throw new Error("Advanced provider request ID retains another exact payload.");if(x.state!=="completed")throw new Error("This paid request has an unknown outcome; recover its actual provider receipt before resuming.");return JSON.parse(String(x.result))}let v=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(v.unknown)>0)throw new Error("A prior advanced paid request is unsettled; no fresh request is allowed.");if(Number(v.steps)>=c.maximumModelSteps||Number(v.spent)+I>c.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,A,I,"reserved")});if(S!==void 0)return await e.authorize(c),p.signal.throwIfAborted(),S;let M=!1;try{if(await e.authorize(c),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,A).changes!==1)throw new Error("The retained advanced request admission changed.");M=!0;let x=await T();if(x.costUsd===null||!Number.isFinite(x.costUsd)||x.costUsd<0||x.costUsd>I+1e-9)throw new Error("The actual provider receipt cannot settle its admitted request ceiling.");let v=JSON.stringify(x.value);if(Buffer.byteLength(v)>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(x.costUsd,v,a,l,A).changes!==1)throw new Error("The advanced request settlement changed concurrently.")}),await e.authorize(c),x.value}catch(k){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),k}}};return Os.run(p,n)}return i(s,"run"),{run:s,async authorize(a){await e.authorize($t(a))},read(a,o){return t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=? AND actor_id=? AND team_id=? AND pin_hash=?").get(a,o.actorId,o.teamId,o.contentHash)?t.prepare("SELECT id,intent_hash,ceiling,state,actual_spend FROM advanced_refiner_calls WHERE run_id=? ORDER BY id").all(a):null},accounting(a,o){let n=this.read(a,o)??[],c=n.filter(g=>g.state!=="completed"&&g.state!=="not_dispatched"),u=n.reduce((g,p)=>g+Number(p.actual_spend??0),0);return{modelRequests:n.filter(g=>g.state!=="not_dispatched").length,knownSpendUsd:u,totalSpendUsd:c.length?null:u,heldUsd:c.reduce((g,p)=>g+Number(p.ceiling),0),uncertainRequests:c.filter(g=>g.state!=="reserved").length}},close(){t.close()}}}i(Rd,"createAdvancedRefinerPaidBoundary");var Wt={buildIntent:"verifiable_reward",methodHint:"grpo"};async function Ed(e){let t=await ea(e.store,e.workspaceId,e.reviewRef),r=`harness-review-${h({review:t.contentHash,strategy:Wt}).slice(0,24)}`,s=(await e.store.listTaskCreationSnapshots(e.profileId)).find(d=>d.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)]),c=new Map(a.map(d=>[Re(d),d])),u=new Map(o.map(d=>[Re(d),d])),g=new Map(n.map(d=>[Re(d),d])),p=new Map;for(let d of t.selectedEvidence){if(d.kind!=="route_decision")continue;let m=c.get(Re(d.evidence)),f=m?u.get(Re(m.trigger)):null;if(!m||!f)throw new Error(`Harness review evidence ${d.evidence.id} cannot be reconstructed.`);let T=f.observations.map(w=>g.get(Re(w))??null).filter(w=>w?.runRef===d.sourceRef);if(!T.length)throw new Error(`Harness review source ${d.sourceRef} has no exact observation evidence.`);let y=p.get(d.sourceRef)??{turnIds:new Set,evidenceRefs:[],sourcePolicies:[]};for(let w of T)y.turnIds.add(w.turnId);y.evidenceRefs.push(d.evidence),y.sourcePolicies.push(d.sourcePolicy),p.set(d.sourceRef,y)}if(!p.size)throw new Error("Harness Taskset review has no reconstructable source runs.");let l=[];for(let[d,m]of[...p.entries()].sort(([f],[T])=>f.localeCompare(T))){let f=await e.taskCreator.addSessionSource({profileId:e.profileId,sessionId:d,turnIds:[...m.turnIds].sort(),consentScope:"selected_turns"}),T=Array.isArray(f.metadata.harnessEvaluationReviews)?f.metadata.harnessEvaluationReviews:[],y={...f,metadata:{...f.metadata,harnessEvaluationReviews:Ft([...T,{id:t.id,contentHash:t.contentHash,workspaceId:t.workspaceRef,harnessRelease:t.harnessRelease,claimFingerprint:t.claim?.fingerprint??null,evidenceRefs:m.evidenceRefs,sourcePolicies:m.sourcePolicies}])}};l.push(await e.store.upsertTrainingSource(y))}return e.startCreation({profileId:e.profileId,sourceIds:l.map(d=>d.id),surface:"task_candidate",mode:"customize",entryMode:"automated",resourceIntent:"workproduct",buildIntent:Wt.buildIntent,buildSpecification:null,objective:t.claim?.statement??t.reason,methodHint:Wt.methodHint,preferredBaseModelId:null,preferredBaseModel:null,candidateId:r,analysisModel:e.analysisModel??null,analysisReasoningEffort:e.analysisReasoningEffort??null,createImproveRunId:null,targetIntent:{kind:"model",id:null,displayName:null,operation:"create"}})}i(Ed,"startHarnessReviewTasksetAuthoring");function Cd(e){let t=e.flatMap(o=>Array.isArray(o.metadata.harnessEvaluationReviews)?o.metadata.harnessEvaluationReviews.filter(Bt):[]);if(!t.length)return null;let r=`${String(t[0].id)}:${String(t[0].contentHash)}`,s=t.filter(o=>`${String(o.id)}:${String(o.contentHash)}`===r);if(s.length!==e.length)return null;let a=s[0];return{review:{id:String(a.id),contentHash:String(a.contentHash),workspaceId:String(a.workspaceId),harnessRelease:a.harnessRelease,claimFingerprint:typeof a.claimFingerprint=="string"?a.claimFingerprint:null},evidenceRefs:Ft(s.flatMap(o=>Array.isArray(o.evidenceRefs)?o.evidenceRefs:[])),sourcePolicies:Ft(s.flatMap(o=>Array.isArray(o.sourcePolicies)?o.sourcePolicies:[]))}}i(Cd,"harnessReviewLineageFromSources");async function Ds(e){let t=e.taskset.metadata.harnessEvaluationLineage;if(!Bt(t)||!Bt(t.review))return null;let r=String(t.review.workspaceId??""),s={id:String(t.review.id??""),contentHash:String(t.review.contentHash??"")};if(!r||!s.id||!s.contentHash)throw new Error("Taskset Harness Evaluation lineage is incomplete.");let a=await e.store.listHarnessImprovementArtifacts(r,"evaluation_review",1e3),o=a.find(d=>Re(d)===Re(s));if(!o)throw new Error("Originating Harness Evaluation review was not found.");let n=e.evaluationResult?{id:e.evaluationResult.id,contentHash:e.evaluationResult.contentHash}:null,c={id:e.taskset.id,contentHash:e.taskset.contentHash},u=h({originRef:s,tasksetRef:c,evaluationRef:n}),g=a.find(d=>d.metadata.downstreamLinkKey===u);if(g)return g;let p=(e.now??(()=>new Date().toISOString()))(),l=yr({schemaVersion:"openpond.harnessEvaluationReviewReceipt.v1",id:`evaluation-review-link-${u.slice(0,24)}`,ownerScope:o.ownerScope,workspaceRef:o.workspaceRef,harnessRelease:o.harnessRelease,previousWatermark:o.nextWatermark,nextWatermark:{cursor:h({cursor:o.nextWatermark.cursor,tasksetRef:c,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:c,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}i(Ds,"linkHarnessReviewTaskset");async function ea(e,t,r){let s=await e.getHarnessWorkspace(t);if(!s)throw new Error(`Harness workspace ${t} was not found.`);let a=(await e.listHarnessImprovementArtifacts(t,"evaluation_review",1e3)).find(n=>n.id===r.id&&n.contentHash===r.contentHash);if(!a||a.schemaVersion!=="openpond.harnessEvaluationReviewReceipt.v1")throw new Error("Harness Evaluation review was not found.");let o=a;if(o.classification!=="taskset"||o.nextAuthority!=="human_review")throw new Error("Only a human-review Taskset recommendation can enter Taskset authoring.");if(o.tasksetProposal)throw new Error("Harness Evaluation review already has a materialized Taskset link.");if(o.ownerScope.kind!==s.ownerScope.kind||o.ownerScope.id!==s.ownerScope.id)throw new Error("Harness Evaluation review owner scope does not match its workspace.");return o}i(ea,"requireTasksetReview");function Re(e){return`${e.id}:${e.contentHash}`}i(Re,"artifactKey");function Bt(e){return!!e&&typeof e=="object"&&!Array.isArray(e)}i(Bt,"isRecord");function Ft(e){let t=new Set;return e.filter(r=>{let s=JSON.stringify(r);return t.has(s)?!1:(t.add(s),!0)})}i(Ft,"uniqueObjects");import Wo from"node:path";function Vt(e,t,r){let s=new Date().toISOString();return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:`fixture_attempt_${t.id}_${r}`,tasksetId:e,taskId:t.taskId,split:Tt(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}})}i(Vt,"fixtureAttempt");function Tt(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}i(Tt,"artifactSplit");function Ns(e){let t=Wr(e.taskset),r=t.issues.filter(R=>R.severity==="error").map(R=>({code:R.code,message:R.message,path:R.path})),s=t.issues.filter(R=>R.severity==="warning").map(R=>({code:R.code,message:R.message,path:R.path})),a=typeof e.taskset.metadata.trainingMethod=="string"?e.taskset.metadata.trainingMethod:null,o=ta(a),n=o==="sft",c=o==="dpo",u=o==="grpo"||o==="ppo",g=new Set(e.taskset.learningSignals.demonstrations.filter(R=>R.approved&&R.taskId).map(R=>R.taskId)),p=e.taskset.metadata.learning===void 0?null:De.parse(e.taskset.metadata.learning),l=i(R=>p?p.admissions.some(G=>G.taskId===R.id&&G.supervisedTarget!==null):R.expectedOutput!=null,"hasSupervisedTarget"),d=e.taskset.datasetArtifact,m=new Set(js(e.taskset.metadata.approvedArtifactSignals)),f=!!d?.schema.fields.some(R=>R.semanticRole==="demonstration"||R.semanticRole==="expected_output"||R.semanticRole==="chosen"),T=e.taskset.tasks.filter(R=>R.split==="train"&&(l(R)||u)),y=n?T.filter(R=>g.has(R.id)):T,w=e.taskset.tasks.filter(R=>R.split==="train"&&l(R)&&!g.has(R.id)),_=e.taskset.tasks.filter(R=>R.split==="frozen_eval"&&(R.expectedOutput||u)),A=d&&f?d.splitCounts.train??0:0,I=d&&f?d.splitCounts.frozen_eval??0:0,S=d?m.has(o??""):!1,M=d?S?A:0:y.length,k=e.taskset.learningSignals.preferences.filter(R=>R.approved).length,x=e.taskset.learningSignals.rewards.filter(R=>R.approved&&R.executable).length,v=c?k:u?x?M:0:M,E=d?I:_.length;v===0&&r.push(n?{code:"sft_demonstrations_missing",message:"At least one explicitly approved training demonstration is required.",path:"learningSignals.demonstrations"}:c?{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 b=d?n&&!S?A:0:w.length;n&&b&&r.push({code:"sft_demonstrations_unapproved",message:`${b} training example${b===1?" is":"s are"} not explicitly approved.`,path:d?"metadata.approvedArtifactSignals":"learningSignals.demonstrations"}),E===0&&s.push({code:"frozen_eval_missing",message:"At least one independent evaluation example is recommended before making quality claims.",path:d?"datasetArtifact.splitCounts.frozen_eval":"tasks"});let D=new Set(y.map(R=>R.clusterKey)),$=new Set(_.map(R=>R.clusterKey)),z=d&&e.taskset.metadata.splitIsolationVerified===!0;(d?!z||v===0||E===0:D.size===0||$.size===0||[...D].some(R=>$.has(R)))&&s.push({code:"independent_evaluation_missing",message:"Independent source clusters are recommended before using evaluation results for quality claims.",path:d?"metadata.splitIsolationVerified":"tasks"}),(d?e.taskset.metadata.artifactRowsVerified!==!0:e.taskset.tasks.some(R=>typeof R.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:d?"metadata.artifactRowsVerified":"tasks.metadata.exampleOrigin"});let L=ra(e.taskset.metadata.diagnosis);(!L||typeof L.summary!="string"||!Array.isArray(L.stableBehavior))&&s.push({code:"capability_diagnosis_missing",message:"A capability diagnosis is recommended to separate stable behavior from changing knowledge.",path:"metadata.diagnosis"}),L?.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 J=e.taskset.graders.some(R=>R.rewardEligible&&(R.kind!=="model_judge"||R.metadata.calibrationIsAdvisory!==!0));J&&!e.graderAudit&&s.push({code:"grader_audit_missing",message:"Optional grader calibration evidence has not been recorded.",path:"graderFixtures"}),J&&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"}),J&&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"}),J&&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"}),J&&e.graderAudit&&!e.graderAudit.leakageChecksPassed&&r.push({code:"environment_leakage",message:"Environment or privileged-state leakage checks failed.",path:"environment"}),J&&e.graderAudit&&!e.graderAudit.infrastructureSafetyPassed&&r.push({code:"infrastructure_reward",message:"An infrastructure failure produced a score or eligible reward.",path:"graderFixtures"});let Xe=e.taskset.graders.some(R=>R.rewardEligible);u&&!Xe&&r.push({code:"online_reward_grader_missing",message:`${o?.toUpperCase()} requires a user-selected reward-eligible grader.`,path:"graders"});let ae=o??(v>0&&e.taskset.capabilities.compatibleMethods.includes("sft")?"sft":"none"),Me=e.taskset.learningSignals.demonstrations.filter(R=>R.approved).map(R=>R.id),C=ae==="none"?null:{primaryMethod:ae,bootstrap:ae==="grpo"&&Me.length?{method:"sft",purpose:"trajectory_bootstrap",demonstrationRefs:Me,limitations:["The SFT bootstrap imitates approved trajectories; it does not optimize verifier reward.","Completing the bootstrap does not satisfy the primary GRPO recommendation."]}:null},j=r.length===0,Z=r.length>0,ue=["sft","dpo","grpo","ppo"].map(R=>{let G=[],te=[];R==="sft"&&Me.length===0&&(G.push("demonstrations_missing"),te.push("Add at least one approved prompt/response demonstration.")),R==="dpo"&&k===0&&(G.push("preference_pairs_missing"),te.push("Add at least one approved chosen/rejected response pair.")),(R==="grpo"||R==="ppo")&&x===0&&(G.push("executable_reward_missing"),te.push("Implement and validate an executable scalar reward.")),R==="ppo"&&(G.push("value_model_required"),te.push("Choose and bind a value/critic model before PPO can execute.")),E===0&&(G.push("frozen_eval_missing"),te.push("Add an independent frozen-evaluation split.")),Z&&G.length===0&&(G.push("taskset_not_ready"),te.push("Resolve the Dataset validation and grader-audit blockers."));let _e=G.filter(Pe=>Pe!=="value_model_required");return{method:R,status:_e.length?"needs_dataset_work":R===ae?"recommended":"compatible",reasonCodes:G,reasons:te}});return Lr.parse({schemaVersion:"openpond.tasksetReadiness.v1",tasksetId:e.taskset.id,tasksetHash:e.taskset.contentHash,ready:j,recommendedMethod:ae,trainingPath:C,methodReadiness:ue,compatibleDestinationClasses:j?ae==="grpo"?["export","custom","hosted_managed"]:["export","custom"]:["export"],blockers:r,advisories:s,warnings:[...t.issues.filter(R=>R.severity==="warning").map(R=>R.message),...ae==="grpo"?["GRPO is readiness-compatible but not executable in the local CPU fixture."]:[],...js(e.taskset.metadata.warnings)],generatedAt:e.generatedAt??new Date().toISOString(),metadata:{reportHash:h([e.taskset.contentHash,r])}})}i(Ns,"buildTasksetReadiness");function ta(e){return e==="sft"||e==="dpo"||e==="grpo"||e==="ppo"||e==="sdft"||e==="opsd"||e==="sdpo"?e:null}i(ta,"trainingPathMethod");function ra(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}i(ra,"metadataRecord");function js(e){return Array.isArray(e)?e.filter(t=>typeof t=="string"&&!!t.trim()):[]}i(js,"metadataStrings");import Ws from"node:path";function qs(e,t,r){let s=De.parse(t.metadata.learning),a=ve(s.binding,s.rewards);return async({grader:o,task:n,attempt:c,signal:u})=>{let g=s.admissions.find(T=>T.taskId===n.id),p=a.find(T=>T.id===o.id);if(!g||p?.kind!=="custom_verifier")throw new Error("Prepared batch has no matching verifier admission.");let l=await ct(r,t),d=l?l.learningResources?.evidence.find(T=>_r(U(T),g.evidence)):await e.learningRepository().transaction(t.profileId,T=>he(T,"evidence",g.evidence));if(!d)throw new Error("Imported batch evidence is missing from its package.");let m=st(d,s.definition),f=i(T=>({id:T.id,input:T.input,expectedOutput:T.expectedOutput,policyVisibleContext:T.policyVisibleContext,privilegedContextRef:T.privilegedContextRef}),"selected");if(h(f(n))!==h(f(m)))throw new Error("Prepared task differs from its admitted evaluator context.");if(l){let T=l.learningResources?.assets.find(y=>y.id===p.verifierRef.id);if(!T)throw new Error("Imported batch verifier is missing from its package.");return je({source:re(T,p.verifierRef),exportName:p.exportName,timeoutMs:p.timeoutMs,signal:u,value:{task:m,attempt:{output:c.output,artifactRefs:c.artifactRefs,runtimeEventRefs:c.runtimeEventRefs,infrastructureError:c.infrastructureError},input:m.input,output:c.output,expectedOutput:m.expectedOutput,evaluatorContext:d.submission.evaluatorContext,infrastructureError:c.infrastructureError??null}})}return Je({repository:e.learningRepository(),scope:t.profileId,grader:p,task:m,evidence:{output:c.output,artifactRefs:c.artifactRefs,runtimeEventRefs:c.runtimeEventRefs,infrastructureError:c.infrastructureError},evaluatorContext:d.submission.evaluatorContext,signal:u})}}i(qs,"createLearningBatchVerifier");import{readFile as sa,realpath as $s}from"node:fs/promises";import Ls from"node:path";async function Us(e){let t=await $s(e.allowedRoot),r=await $s(Ls.resolve(t,e.grader.module));if(r!==t&&!r.startsWith(`${t}${Ls.sep}`))throw new Error("Verifier module is outside the approved Taskset root.");let s=await sa(r);if(e.grader.metadata.portableVerifierRef!==void 0){let o=kr.parse(e.grader.metadata.portableVerifierRef);if(o.sizeBytes!==s.length||o.contentHash!==fe(s))throw new Error("Verifier source differs from its immutable Taskset release.")}let a=new TextDecoder("utf-8",{fatal:!0}).decode(s);return je({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}:{}}})}i(Us,"runSandboxedVerifier");async function zt(e,t){if(!t.graders.some(n=>n.kind==="custom_verifier"))return;if(t.metadata.learning!==void 0)return qs(e.store,t,e.storeDir);if(t.metadata.rewardBinding!==void 0)return Cs(e.store,t,e.storeDir);let r=e.storeDir?null:await(e.loadProfileState??Gr)(),s=e.storeDir?Ws.join(e.storeDir,"training","tasksets",ce(t)):r?.sourcePath?Ws.join(r.sourcePath,"tasksets",t.id):null,a=typeof t.metadata.creationSnapshotId=="string"?t.metadata.creationSnapshotId:null,o=a?await e.store.getTaskDesignProposal(a):null;return s&&t.purpose!=="benchmark"&&t.environment.metadata.runtimeSourceTasksetId===void 0&&await nt(t,s,{generatedFiles:o?.generatedFiles??[]}),s?async({grader:n,task:c,attempt:u,signal:g})=>{let p=t.graders.find(d=>d.id===n.id);if(!p||p.kind!=="custom_verifier")throw new Error("Verifier is not declared by this Taskset revision.");let l;if(t.environment.entrypoint===Y&&!u.infrastructureError){if(!e.storeDir)throw new Error("Tool grading requires the execution owner's artifact directory.");l=await vt({store:e.store,storeDir:e.storeDir,taskset:t,task:c,attempt:u})}return Us({grader:p,task:c,attempt:u,allowedRoot:s,signal:g,evaluatorContext:l})}:void 0}i(zt,"createTasksetEvaluationVerifier");var oa=new Set(["correct","incorrect","parse_failure","budget_exhausted","tool_schema_violation","infrastructure_failure","cancelled"]);async function Bs(e){let t=na(e.attempt);if(t){let g=h(e.graders),p=`Taskset Work attempt ended with ${t.failureClass}: `+t.message;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([e.attempt.id,g,t]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:g,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:ia(e.attempt.metadata);if(!r)try{return await at(e)}catch(g){return aa({attempt:e.attempt,graders:e.graders,now:e.now,error:g})}let s=h(e.graders),a=r.rewardEligible?r.outcome==="correct"?1:0:null,o=r.outcome==="correct",n=da(r.outcome),c=`Cross-System trajectory verifier outcome: ${r.outcome}.`,u=e.graders.map(g=>({graderId:g.id,graderVersion:g.version,score:a??0,passed:o,hardGate:g.hardGate,rewardEligible:r.rewardEligible&&g.rewardEligible,feedback:c,evidenceRefs:e.attempt.artifactRefs,judge:g.kind==="model_judge"?g.judge:null,calibrationStatus:g.kind==="model_judge"?g.calibrationStatus:"not_applicable"}));return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([e.attempt.id,s,r]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:s,score:a,passed:o,components:u,failureClass:n,feedback:[c],rewardEligible:r.rewardEligible,createdAt:e.now?.()??new Date().toISOString()}}i(Bs,"gradeTasksetEvaluationAttempt");function aa(e){let t=h(e.graders),r=(e.error instanceof Error?e.error.message:String(e.error)).slice(0,19e3),s=`Taskset grader failed: ${r}`;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([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()}}i(aa,"graderFailureResult");function na(e){if(e.metadata.execution!=="taskset_work"||!e.infrastructureError)return null;let t=Vr.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}}i(na,"tasksetWorkFailure");function ia(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"||!oa.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}}i(ia,"crossSystemGrade");function da(e){return e==="correct"?null:e==="infrastructure_failure"?"infrastructure_failure":e==="cancelled"?"cancelled":"policy_failure"}i(da,"failureClassFor");import Ge from"node:path";import{mkdir as Fs,readFile as ca,writeFile as Vs}from"node:fs/promises";async function Q(e){let t=Ge.join(e.storeDir,"training","evaluation-artifacts",e.tasksetId),r=Ge.join(t,`${e.attemptId}-${zs(e.fileLabel??e.kind)}.json`),s=Buffer.from(`${JSON.stringify({schemaVersion:"openpond.rawEvaluationArtifact.v1",requestId:e.requestId,...e.payload},null,2)}
+import{a as vs,b as Ts}from"./chunk-2G4BEISF.js";import{a as ps,c as fs}from"./chunk-XVDFJ3ZF.js";import{a as oe}from"./chunk-CKMSBAI2.js";import{i as ms,j as pt}from"./chunk-776WNDV2.js";import{d as je,e as it}from"./chunk-TRQQNFMO.js";import{a as ce}from"./chunk-I2NXNMRZ.js";import{a as Yr,b as Kr,c as Xr,d as dt,f as Zr,g as Qr,h as ct}from"./chunk-2L5XWJJ5.js";import{b as Br,e as at,f as nt}from"./chunk-JQKSXCAJ.js";import{E as ys,N as ws,a as ft,b as gs,c as gt,d as ht,e as hs,f as ks}from"./chunk-N6ENLEMY.js";import{g as ut,i as mt}from"./chunk-GOPFFQZU.js";import{a as ye}from"./chunk-7KWD2I6W.js";import{g as qe}from"./chunk-57TRXBLT.js";import{s as Gr}from"./chunk-FKJHEZCW.js";import{A as rs,C as ss,t as ts}from"./chunk-2RNQQNB7.js";import{q as es}from"./chunk-EUMXMT3S.js";import{G as B,H as os,I as as,J as ns,K as lt,L as is,M as q,N as ds,O as Ne,P as cs,R as ls,S as Lt,T as us}from"./chunk-TKLBP3MM.js";import{l as $t}from"./chunk-RBZH6A6M.js";import{p as Vr,r as F,s as zr,u as Jr}from"./chunk-JMBQYBPI.js";import{f as Fr}from"./chunk-H7ATCPPB.js";import{Ib as ot,M as Pr,N as Or,O as Hr,P as Dr,R as jr,ca as Nr,da as qr,ka as $r,sa as Lr,ua as ke,va as Ur,wa as Wr,xa as Te}from"./chunk-XMKCCLGB.js";import{n as se}from"./chunk-523TE4BI.js";import{C as br,S as Cr,T as Ar,Tb as Mr}from"./chunk-Y4OWBBDD.js";import{f as wr}from"./chunk-JFXWI3EN.js";import{n as yr}from"./chunk-H7IWJN7L.js";import{$ as De,Kb as Er,N as U,O as xr,V as st,W as Ir,c as qt,d as re,g as Rr,j as _r,lb as ge,mb as he}from"./chunk-VYLGI3IH.js";import{La as ve,Ma as Sr,l as ze,oa as rt}from"./chunk-ACYFMLWG.js";import{l as vr,m as Tr}from"./chunk-TVQAPBLZ.js";import{h as kr,l as fe,m as h}from"./chunk-KGNCKIFP.js";import{b as O}from"./chunk-ES2ISFJC.js";import{a as i}from"./chunk-FTQLQJQV.js";import{randomUUID as Vo}from"node:crypto";function yt(e){let t=kt(kt(kt(e.metadata).billing).pricing);return{version:$e(t.version,"pricing version"),source:$e(t.source,"pricing source"),effectiveAt:$e(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Le(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Le(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Le(t.outputUsdPerMillionTokens,"output token rate")}}i(yt,"hostedTokenPricingFromCatalog");function Fn(e){let t=kt(e);try{return{version:$e(t.version,"pricing version"),source:$e(t.source,"pricing source"),effectiveAt:$e(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Le(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Le(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Le(t.outputUsdPerMillionTokens,"output token rate")}}catch{return null}}i(Fn,"hostedTokenPricingFromValue");function Se(e,t){let r=ye(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}i(Se,"hostedUsageCostUsd");function kt(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(kt,"record");function $e(e,t){if(typeof e=="string"&&e.trim())return e.trim();throw new Error(`Hosted model catalog is missing ${t}.`)}i($e,"requiredString");function Le(e,t){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return e;throw new Error(`Hosted model catalog is missing ${t}.`)}i(Le,"requiredRate");function Rs(e={}){let t=e.catalog??ut,r=e.stream??mt;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: ${a.error}`);let o=a.models.find(l=>l.id===s.modelId);if(!o)throw new Error("The selected judge model is unavailable.");let n=Ss(o.raw),c=yt(n),u=bs(n.context_window,"context window"),g=Math.min(4096,bs(n.output_limit,"output limit")),p=(u*Math.max(c.inputUsdPerMillionTokens,c.cachedInputUsdPerMillionTokens)+g*c.outputUsdPerMillionTokens)/1e6;if(!(p>0))throw new Error("The judge model must have a positive known charge ceiling.");return{maximumChargeUsd:p,async dispatch(l){l?.throwIfAborted();let d="",m=null,f=null,T=null,y=!1;for await(let A of r({model:s.modelId,messages:[{role:"system",content:s.system},{role:"user",content:s.data}],requestId:`learning-judge-${Vo()}`,maxTokens:g,temperature:s.temperature,signal:l})){let I=Ss(A.raw);if(typeof I.model=="string"){if(m!==null&&I.model!==m)throw new Error("Judge response model identity changed during streaming.");m=I.model}if(typeof I.id=="string"){if(f!==null&&f!==I.id)throw new Error("Judge response identity changed during streaming.");f=I.id}if(A.type==="text_delta"&&(d+=A.text),d.length>1e5)throw new Error("Judge response exceeds the retained response limit.");A.type==="usage"&&(T=A.usage),A.type==="finish"&&(y=!0)}if(!y||!m||!f)throw new Error("Judge response did not include a complete model receipt.");let w=ye(T),_=w.promptTokens!==null&&w.completionTokens!==null?Se(T,c):null;return{text:d,modelId:m,modelRevision:null,responseId:f,inputTokens:w.promptTokens,outputTokens:w.completionTokens,costUsd:_}}}},async cancel({run:s}){return!s.judgeCalls?.some(a=>a.status==="reserved")}}}i(Rs,"createLearningHostedJudgeProvider");function Ss(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(Ss,"record");function bs(e,t){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new Error(`Judge catalog is missing its ${t}.`);return e}i(bs,"positiveInteger");function oi(e,t=Rs()){return{execute(r){if(r.definition.execution.environment.entrypoint==="openpond.javascript-environment.v1")throw new Error("Tool Tasksets require grading an owner-recorded environment attempt.");let s=Er({repository:e,scope:r.scope,run:r.run});return Sr({binding:r.binding,rewards:r.rewards,task:st(r.evidence,r.definition),evidence:Ir(r.evidence,r.run.output),evaluatorContext:r.evidence.submission.evaluatorContext,signal:r.signal,modelJudge:Rr({readRubric:i(a=>e.transaction(r.scope,async o=>re(await ge(o,"asset",a.id,1),a)),"readRubric"),async executeBudgeted(a,o){let n=await t.prepare(a,{scope:r.scope,run:r.run});return _r({store:s,maximumCharge:i(()=>n.maximumChargeUsd,"maximumCharge"),dispatch:i((u,g)=>n.dispatch(g),"dispatch")})(`grade-${h(a)}`,a,o)}}),customVerifier:i(async({grader:a,task:o,evidence:n})=>{let c=await Je({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:c.score,passed:c.passed,rewardEligible:a.rewardEligible,failureClass:null,feedback:[c.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:c.evidenceRefs}},"customVerifier")})},async cancel({scope:r,run:s}){let a=await e.transaction(r,o=>ge(o,"grade",s.id));return a.judgeCalls?.some(o=>o.status==="reserved")?!1:t.cancel({scope:r,run:a})}}}i(oi,"createLocalTaskGradeExecutor");async function Je(e){let t=await e.repository.transaction(e.scope,async r=>re(await ge(r,"asset",e.grader.verifierRef.id,1),e.grader.verifierRef));return je({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}})}i(Je,"executeLocalLearningVerifier");import zo from"node:path";async function _s(e,t,r){if(!r)throw new Error("Ordinary tool execution requires its owner's package directory.");let s=ot({taskset:e,adapterId:dt(e)}),a=await Zr({root:zo.join(r,"training","tasksets",ce(e)),content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:s.tasksetRelease,environment:s.environmentRelease,verifierSet:s.verifierSetRelease},sources:Fr.parse(e.metadata.portableFileInventory)}),o=qr(a);if(!o)throw new Error("The Taskset does not declare a JavaScript execution graph.");let n=a.taskset.tasks.find(g=>g.id===t.id);if(!n)throw new Error("The selected task is missing from its executable package.");let c=o.assets.find(g=>g.id===o.execution.javascript.module.id),u=o.assets.find(g=>g.id===n.privilegedContextRef);return{execution:o.execution,module:c,initialState:JSON.parse(u.text),definition:U(a.taskset),instructions:$r(a),task:n}}i(_s,"loadOrdinaryToolEnvironment");var Y="openpond.javascript-environment.v1";async function Ue(e,t,r,s){let a=t.tasks.find(o=>o.id===r.id);if(!a||h(a)!==h(r))throw new Error("Tool task differs from the selected immutable Taskset.");return t.metadata.taskDefinition===void 0?_s(t,r,s):e.learningRepository().transaction(t.profileId,async o=>{let n=await he(o,"definition",se.parse(t.metadata.taskDefinition)),c=n.execution.environment;if(c.entrypoint!==Y||t.environment.entrypoint!==Y||t.environment.kind!=="agent"||t.environment.stateful!==c.stateful||t.environment.defaultTimeoutMs!==c.defaultTimeoutMs||t.environment.networkPolicy!==c.networkPolicy||t.environment.deterministicSeeds!==c.deterministicSeeds||h(t.environment.toolNames)!==h(n.execution.tools.map(S=>S.name))||h(t.environment.metadata.portableTools)!==h(n.execution.tools)||h(t.policy)!==h(n.execution.policy)||h(t.metadata.rewardBinding)!==h(n.rewardBinding))throw new Error("Tool Taskset context differs from its pinned task definition.");let u=await he(o,"binding",n.rewardBinding),g=await Promise.all(u.sources.map(S=>he(o,"reward",S.reward))),p=rt.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:ve(u,g),tasks:[p]},d=await ge(o,"asset",Dr(l),1),m=Or.parse(JSON.parse(re(d,d.asset))),f=[m.javascript.module,m.environment.actionSchemaRef,m.environment.observationSchemaRef,m.environment.stateSchemaRef].filter(S=>S!==null),T=[...new Set([...f.map(S=>S.id),r.privilegedContextRef].filter(S=>S!==null))],y=await Promise.all(T.map(S=>ge(o,"asset",S,1))),w=jr(l,d,y);if(h(t.environment.metadata.portableExecutionResources)!==h({environment:w.environment,verifierSet:w.verifierSet}))throw new Error("Tool Taskset publication differs from its private execution resources.");let _=y.find(S=>S.id===w.javascript.module.id),A=y.find(S=>S.id===r.privilegedContextRef),I=JSON.parse(re(A,A.asset));return{execution:w,module:_,initialState:I,definition:U(n),instructions:n.instructions,task:p}})}i(Ue,"loadStarterToolEnvironment");function xs(e){let t=qt({text:e.instructions,path:"policy/instructions.txt",mediaType:"text/plain",visibility:"policy"}),r=qt({text:JSON.stringify({environment:U(e.execution.environment),verifierSet:U(e.execution.verifierSet),javascript:U(e.execution.javascript)}),path:"environment/dependencies.json",mediaType:"application/json",visibility:"host_private"}),s=vr({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:Y}}),a=Tr({schemaVersion:"openpond.harnessRelease.v2",id:`starter-harness-${s.contentHash}`,agentSnapshot:{id:s.id,contentHash:s.contentHash},program:e.module.asset,tools:e.execution.javascript.tools,lifecycle:{create:!0,reset:!0,step:!0,collect:!0,destroy:!0,resetScope:"attempt"},graderInterface:{visibleEvidence:["output"],privilegedEvidence:["environment_state","private_verifier"],privateVerifierIsolation:!0},files:[t.asset,r.asset,e.module.asset],metadata:{execution:Y}});return{agentSnapshot:s,harnessRelease:a,instructionContext:void 0}}i(xs,"compileStarterToolHarness");import wt from"node:path";import{constants as Is}from"node:fs";import{open as Jo,realpath as Es}from"node:fs/promises";var Go=O.string().regex(/^[a-f0-9]{64}$/),Yo=O.object({schemaVersion:O.literal("openpond.starterToolAttempt.v1"),requestId:O.string(),taskset:se,taskHash:Go,model:O.unknown(),seed:O.number().int(),attempt:O.number().int(),definition:se,policySource:O.enum(["model","fixture"]),environment:se,verifierSet:se,result:Pr}).strict();async function vt(e){let{store:t,taskset:r,task:s,attempt:a}=e,o=await Ue(t,r,s,e.storeDir),c=(await t.listTaskAttemptArtifacts({attemptId:a.id})).filter(_=>_.kind==="environment_state"&&_.tasksetId===r.id&&_.taskId===s.id&&_.attemptId===a.id&&a.artifactRefs.includes(_.id)&&a.privilegedOutcomeRef===_.id);if(c.length!==1)throw new Error("Tool grading requires exactly one owner-recorded environment artifact.");let u=c[0],g=await Es(wt.join(e.storeDir,"training","evaluation-artifacts")),p=await Es(wt.dirname(u.path)),l=wt.relative(g,p);if(!l||l.startsWith("..")||wt.isAbsolute(l))throw new Error("Tool evidence is outside the execution owner's artifact directory.");let d=await Jo(u.path,Is.O_RDONLY|Is.O_NOFOLLOW),m;try{let _=await d.stat();if(!_.isFile()||_.size>33554432||_.size!==u.sizeBytes)throw new Error("Tool evidence exceeds its declared byte boundary.");m=await d.readFile()}finally{await d.close()}if(m.byteLength!==u.sizeBytes||fe(m)!==u.sha256)throw new Error("Owner-recorded tool evidence bytes changed.");let f=Yo.parse(JSON.parse(m.toString("utf8"))),{contentHash:T,...y}=f.result,w=f.result;if(h(y)!==T||!w.snapshot||!w.environmentCleanupComplete||!["completed","budget_exhausted","policy_failure"].includes(w.status)||h(f.taskset)!==h(U(r))||f.taskHash!==h(s)||h(f.model)!==h(a.modelRef)||f.seed!==a.seed||f.attempt!==a.attempt||a.tasksetId!==r.id||a.taskId!==s.id||a.split!==s.split||h(f.definition)!==h(o.definition)||h(f.environment)!==h(U(o.execution.environment))||h(f.verifierSet)!==h(U(o.execution.verifierSet))||w.taskId!==s.id||w.snapshot.seed!==a.seed||w.snapshot.inputHash!==h(s.input)||h(w.snapshot.definition)!==h(U(o.execution.javascript))||w.snapshot.finalStateHash!==h(w.snapshot.state)||h(a.output)!==h({text:w.output??""})||a.metadata.environmentAttemptHash!==T)throw new Error("Tool evidence differs from the selected task, model or recorded attempt.");if(a.metadata.policySource!==f.policySource)throw new Error("Tool evidence policy source changed.");return{environment:{status:w.status,collected:w.collected,definition:w.snapshot.definition,initialStateHash:w.snapshot.initialStateHash,finalStateHash:w.snapshot.finalStateHash,finalState:w.snapshot.state,events:w.snapshot.events}}}i(vt,"readStarterToolEvidence");async function be(e,t){if(t.metadata.learning!==void 0)return De.parse(t.metadata.learning);if(t.metadata.rewardBinding===void 0)return;let r=se.parse(t.metadata.rewardBinding);return e.learningRepository().transaction(t.profileId,async s=>{let a=await he(s,"binding",r),o=[...new Map(a.sources.map(c=>[`${c.reward.id}:${c.reward.revision}:${c.reward.contentHash}`,c.reward])).values()],n=await Promise.all(o.map(c=>he(s,"reward",c)));return{binding:a,rewards:n}})}i(be,"resolveTasksetRewardBinding");async function Ko(e,t,r){let s=await be(e,t);if(!s)return{rewardExecution:void 0,verifierAssets:[]};let a=ve(s.binding,s.rewards).flatMap(c=>c.kind==="custom_verifier"?[c.verifierRef]:c.kind==="model_judge"?[c.rubricRef]:[]),o=t.metadata.learning===void 0?void 0:await ct(r,t);if(o){let c=[...new Map(a.map(u=>{let g=o.learningResources.assets.find(p=>p.id===u.id);if(!g)throw new Error("Imported batch verifier is missing from its package.");return re(g,u),[g.id,g]})).values()];return{rewardExecution:s,verifierAssets:c}}let n=await e.learningRepository().transaction(t.profileId,async c=>{let u=new Map;for(let g of a){let p=await ge(c,"asset",g.id,1);re(p,g),u.set(p.id,p)}return[...u.values()]});return{rewardExecution:s,verifierAssets:n}}i(Ko,"resolveTasksetTrainingReward");async function Ji(e,t,r){let s=await Ko(e,t,r.storeDir);if(!s.rewardExecution)return s;let a=ve(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}i(Ji,"resolveManagedTasksetReward");async function Cs(e,t,r){let s=await be(e,t);if(!s)throw new Error("The Taskset has no published Reward binding.");let a=ve(s.binding,s.rewards);return async({grader:o,task:n,attempt:c,signal:u})=>{let g=t.tasks.find(m=>m.id===n.id);if(!g||h(g)!==h(n))throw new Error("Task differs from the selected immutable Taskset.");let p=null;if(t.environment.entrypoint===Y){if(!r)throw new Error("Tool grading requires the execution owner's artifact directory.");p=await vt({store:e,storeDir:r,taskset:t,task:n,attempt:c})}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 d=rt.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 Je({repository:e.learningRepository(),scope:t.profileId,grader:l,task:d,evidence:{output:c.output,artifactRefs:c.artifactRefs,runtimeEventRefs:c.runtimeEventRefs,infrastructureError:c.infrastructureError},evaluatorContext:p,signal:u})}}i(Cs,"createTasksetBindingVerifier");async function Ee(e,t,r){let s=r?se.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||Te(a)!==s.contentHash))throw new Error("Taskset execution revision is unavailable or differs from its immutable hash.");return a}i(Ee,"requireLocalTasksetRevision");async function ed(e,t,r){let s=await e.releaseForTaskset(t);return s||ot({taskset:t,rewardExecution:r?await be(r,t):void 0,adapterId:dt(t)}).tasksetRelease}i(ed,"requireReleasedTaskset");import{createHash as Xo}from"node:crypto";function As(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||h(r.metadata.portableTaskRecord)!==h(s))throw new Error("The Work task differs from its exact released population.");let a=0,o=(r.assets??[]).map(n=>{let c=s.artifactRefs.find(l=>l.id===n.id&&l.contentHash===n.sha256&&l.path===n.artifactRef),u=c&&e.files.find(l=>h(l.asset)===h(c));if(!c||!u||c.visibility!=="policy")throw new Error("The Work task requests an unavailable or private evaluator file.");let g=Buffer.from(Nr(u));if(Xo("sha256").update(g).digest("hex")!==n.sha256||g.length!==n.sizeBytes||n.split!==r.split)throw new Error("The published Work input manifest changed.");if(a+=g.length,a>25e7)throw new Error("Published Work inputs exceed the admitted 250 MB ceiling.");return{...n,assetId:n.id,storageName:n.fileName,bytes:g,sizeBytes:g.length}});if(o.length!==s.artifactRefs.length)throw new Error("The exact policy-visible Work asset population is incomplete.");return o}i(As,"advancedPublishedWorkAssets");async function Ut(e,t=ut,r=!1){if(e.kind==="hosted_harness")return Ut({kind:"hosted_chat",modelId:e.modelId,maxOutputTokens:4096,temperature:0,topP:1},t);if(e.kind!=="hosted_chat")throw new oe("local_target_not_qualified","Select an explicit model target for local execution.",422);if(e.harness&&!r)throw new oe("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 oe("local_model_catalog_unavailable",s.error,422);let a=s.models.find(d=>d.id===e.modelId);if(!a)throw new oe("local_model_unavailable","The selected model is unavailable.",422);let o=Ms(a.raw),n=yt(o),c=Ms(o.capabilities).samplingParameters===!0;if(!c&&(e.temperature!==0||e.topP!==1))throw new oe("local_sampling_unsupported","This model uses provider-managed sampling. Custom temperature or top P requires an explicit supported capability.",422);let u=Ps(o.context_window),g=Ps(o.output_limit);if(e.maxOutputTokens>g)throw new oe("local_model_output_limit","The model cannot accept the configured output limit.",422);let p={providerId:"openpond",modelId:e.modelId,maxOutputTokens:e.maxOutputTokens,...c?{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 oe("local_model_charge_unknown","The model must publish an enforceable positive request ceiling.",422);return{model:{...p,configurationHash:h({configuration:p,contextWindow:u,outputLimit:g,pricing:n})},pricing:n,maximumChargeUsd:l}}i(Ut,"prepareLocalExperimentModel");function fd(e){let t=e.admission.streamFactory?.(e)??e.admission.stream??e.stream??mt;return async function*(r){if(!r.requestId)throw new oe("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 oe("local_model_configuration_conflict","The native model request changed its admitted model, output limit or sampling configuration.",422);let a=ps();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,c=!1,u=!1,g;try{for await(let p of t(r)){u=!0,p.type==="usage"&&(n=p.usage),p.type==="finish"&&(c=!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:ye(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 fs&&p.executionId===o.id&&p.requestId===o.requestId&&(g=p),p}finally{let p=ye(n),l=e.admission.pricing&&c&&p.promptTokens!==null&&p.completionTokens!==null?Se(n,e.admission.pricing):null;await e.store.settleLocalCharge({...o,costUsd:l,confirmedTransportNotInvoked:g,usage:n===null?null:{promptTokens:p.promptTokens,completionTokens:p.completionTokens,cachedPromptTokens:p.cachedPromptTokens,uncachedPromptTokens:p.uncachedPromptTokens}})}}}i(fd,"createLocalBudgetedModelStream");function Ms(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(Ms,"record");function Ps(e){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new oe("local_model_bound_unknown","The model catalog is missing an enforceable context/output bound.",422);return e}i(Ps,"positive");import{AsyncLocalStorage as Zo}from"node:async_hooks";import Qo from"node:path";var Os=new Zo;function Hs(){return Os.getStore()}i(Hs,"currentAdvancedRefinerBoundary");function Rd(e){let t=yr(Qo.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}}i(r,"transaction");async function s(a,o,n){let c=$t(o);await e.authorize(c),r(()=>{let l=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(l){if(l.actor_id!==c.actorId||l.team_id!==c.teamId||l.pin_hash!==c.contentHash)throw new Error("Advanced Run budget belongs to another immutable admission.");return}t.prepare("INSERT INTO advanced_refiner_budgets VALUES(?,?,?,?,?,?,?)").run(a,c.actorId,c.teamId,c.contentHash,c.maximumCostUsd,c.maximumModelSteps,new Date(Date.now()+c.maximumDurationMs).toISOString())});let u=t.prepare("SELECT expires_at FROM advanced_refiner_budgets WHERE id=?").get(a),g=Date.parse(String(u.expires_at))-Date.now();if(g<=0)throw new Error("The immutable advanced evaluation duration is exhausted.");let p={runId:a,pin:c,signal:AbortSignal.timeout(g),authorize:i(()=>e.authorize(c),"authorize"),async resolveWorkAssets(l,d){await e.authorize(c),p.signal.throwIfAborted();let m=await e.packageForPin(c),f=As(m,l,d);return await e.authorize(c),p.signal.throwIfAborted(),f},async call(l,d,m,f,T){if(await e.authorize(c),!Number.isInteger(f)||f<1)throw new Error("Advanced paid requests require a finite enforced token ceiling.");let y=d.model;if(y?.providerId!=="openpond"||!y.modelId)throw new Error("Advanced paid calls require an explicitly qualified hosted model.");let w=y.modelId,_=e.prepareModel?await e.prepareModel(w,f,m):await(async()=>{let k=await Ut({kind:"hosted_chat",modelId:w,maxOutputTokens:f,temperature:0,topP:1});if(h(k.pricing)!==h(m))throw new Error("The actual model pricing differs from its reviewed request pin.");return{configurationHash:k.model.configurationHash,maximumChargeUsd:k.maximumChargeUsd}})();if(!/^[a-f0-9]{64}$/.test(_.configurationHash)||!Number.isFinite(_.maximumChargeUsd)||_.maximumChargeUsd<=0)throw new Error("The actual owner model ceiling is unavailable.");let A=h({intent:d,pricing:m,maximumOutputTokens:f,modelConfigurationHash:_.configurationHash}),I=_.maximumChargeUsd,S=r(()=>{let k=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(!k||String(k.expires_at)<=new Date().toISOString())throw new Error("The immutable advanced evaluation duration is exhausted.");let x=t.prepare("SELECT * FROM advanced_refiner_calls WHERE run_id=? AND id=?").get(a,l);if(x){if(x.intent_hash!==A)throw new Error("Advanced provider request ID retains another exact payload.");if(x.state!=="completed")throw new Error("This paid request has an unknown outcome; recover its actual provider receipt before resuming.");return JSON.parse(String(x.result))}let v=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(v.unknown)>0)throw new Error("A prior advanced paid request is unsettled; no fresh request is allowed.");if(Number(v.steps)>=c.maximumModelSteps||Number(v.spent)+I>c.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,A,I,"reserved")});if(S!==void 0)return await e.authorize(c),p.signal.throwIfAborted(),S;let M=!1;try{if(await e.authorize(c),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,A).changes!==1)throw new Error("The retained advanced request admission changed.");M=!0;let x=await T();if(x.costUsd===null||!Number.isFinite(x.costUsd)||x.costUsd<0||x.costUsd>I+1e-9)throw new Error("The actual provider receipt cannot settle its admitted request ceiling.");let v=JSON.stringify(x.value);if(Buffer.byteLength(v)>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(x.costUsd,v,a,l,A).changes!==1)throw new Error("The advanced request settlement changed concurrently.")}),await e.authorize(c),x.value}catch(k){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),k}}};return Os.run(p,n)}return i(s,"run"),{run:s,async authorize(a){await e.authorize($t(a))},read(a,o){return t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=? AND actor_id=? AND team_id=? AND pin_hash=?").get(a,o.actorId,o.teamId,o.contentHash)?t.prepare("SELECT id,intent_hash,ceiling,state,actual_spend FROM advanced_refiner_calls WHERE run_id=? ORDER BY id").all(a):null},accounting(a,o){let n=this.read(a,o)??[],c=n.filter(g=>g.state!=="completed"&&g.state!=="not_dispatched"),u=n.reduce((g,p)=>g+Number(p.actual_spend??0),0);return{modelRequests:n.filter(g=>g.state!=="not_dispatched").length,knownSpendUsd:u,totalSpendUsd:c.length?null:u,heldUsd:c.reduce((g,p)=>g+Number(p.ceiling),0),uncertainRequests:c.filter(g=>g.state!=="reserved").length}},close(){t.close()}}}i(Rd,"createAdvancedRefinerPaidBoundary");var Wt={buildIntent:"verifiable_reward",methodHint:"grpo"};async function Ed(e){let t=await ea(e.store,e.workspaceId,e.reviewRef),r=`harness-review-${h({review:t.contentHash,strategy:Wt}).slice(0,24)}`,s=(await e.store.listTaskCreationSnapshots(e.profileId)).find(d=>d.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)]),c=new Map(a.map(d=>[Re(d),d])),u=new Map(o.map(d=>[Re(d),d])),g=new Map(n.map(d=>[Re(d),d])),p=new Map;for(let d of t.selectedEvidence){if(d.kind!=="route_decision")continue;let m=c.get(Re(d.evidence)),f=m?u.get(Re(m.trigger)):null;if(!m||!f)throw new Error(`Harness review evidence ${d.evidence.id} cannot be reconstructed.`);let T=f.observations.map(w=>g.get(Re(w))??null).filter(w=>w?.runRef===d.sourceRef);if(!T.length)throw new Error(`Harness review source ${d.sourceRef} has no exact observation evidence.`);let y=p.get(d.sourceRef)??{turnIds:new Set,evidenceRefs:[],sourcePolicies:[]};for(let w of T)y.turnIds.add(w.turnId);y.evidenceRefs.push(d.evidence),y.sourcePolicies.push(d.sourcePolicy),p.set(d.sourceRef,y)}if(!p.size)throw new Error("Harness Taskset review has no reconstructable source runs.");let l=[];for(let[d,m]of[...p.entries()].sort(([f],[T])=>f.localeCompare(T))){let f=await e.taskCreator.addSessionSource({profileId:e.profileId,sessionId:d,turnIds:[...m.turnIds].sort(),consentScope:"selected_turns"}),T=Array.isArray(f.metadata.harnessEvaluationReviews)?f.metadata.harnessEvaluationReviews:[],y={...f,metadata:{...f.metadata,harnessEvaluationReviews:Ft([...T,{id:t.id,contentHash:t.contentHash,workspaceId:t.workspaceRef,harnessRelease:t.harnessRelease,claimFingerprint:t.claim?.fingerprint??null,evidenceRefs:m.evidenceRefs,sourcePolicies:m.sourcePolicies}])}};l.push(await e.store.upsertTrainingSource(y))}return e.startCreation({profileId:e.profileId,sourceIds:l.map(d=>d.id),surface:"task_candidate",mode:"customize",entryMode:"automated",resourceIntent:"workproduct",buildIntent:Wt.buildIntent,buildSpecification:null,objective:t.claim?.statement??t.reason,methodHint:Wt.methodHint,preferredBaseModelId:null,preferredBaseModel:null,candidateId:r,analysisModel:e.analysisModel??null,analysisReasoningEffort:e.analysisReasoningEffort??null,createImproveRunId:null,targetIntent:{kind:"model",id:null,displayName:null,operation:"create"}})}i(Ed,"startHarnessReviewTasksetAuthoring");function Cd(e){let t=e.flatMap(o=>Array.isArray(o.metadata.harnessEvaluationReviews)?o.metadata.harnessEvaluationReviews.filter(Bt):[]);if(!t.length)return null;let r=`${String(t[0].id)}:${String(t[0].contentHash)}`,s=t.filter(o=>`${String(o.id)}:${String(o.contentHash)}`===r);if(s.length!==e.length)return null;let a=s[0];return{review:{id:String(a.id),contentHash:String(a.contentHash),workspaceId:String(a.workspaceId),harnessRelease:a.harnessRelease,claimFingerprint:typeof a.claimFingerprint=="string"?a.claimFingerprint:null},evidenceRefs:Ft(s.flatMap(o=>Array.isArray(o.evidenceRefs)?o.evidenceRefs:[])),sourcePolicies:Ft(s.flatMap(o=>Array.isArray(o.sourcePolicies)?o.sourcePolicies:[]))}}i(Cd,"harnessReviewLineageFromSources");async function Ds(e){let t=e.taskset.metadata.harnessEvaluationLineage;if(!Bt(t)||!Bt(t.review))return null;let r=String(t.review.workspaceId??""),s={id:String(t.review.id??""),contentHash:String(t.review.contentHash??"")};if(!r||!s.id||!s.contentHash)throw new Error("Taskset Harness Evaluation lineage is incomplete.");let a=await e.store.listHarnessImprovementArtifacts(r,"evaluation_review",1e3),o=a.find(d=>Re(d)===Re(s));if(!o)throw new Error("Originating Harness Evaluation review was not found.");let n=e.evaluationResult?{id:e.evaluationResult.id,contentHash:e.evaluationResult.contentHash}:null,c={id:e.taskset.id,contentHash:e.taskset.contentHash},u=h({originRef:s,tasksetRef:c,evaluationRef:n}),g=a.find(d=>d.metadata.downstreamLinkKey===u);if(g)return g;let p=(e.now??(()=>new Date().toISOString()))(),l=wr({schemaVersion:"openpond.harnessEvaluationReviewReceipt.v1",id:`evaluation-review-link-${u.slice(0,24)}`,ownerScope:o.ownerScope,workspaceRef:o.workspaceRef,harnessRelease:o.harnessRelease,previousWatermark:o.nextWatermark,nextWatermark:{cursor:h({cursor:o.nextWatermark.cursor,tasksetRef:c,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:c,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}i(Ds,"linkHarnessReviewTaskset");async function ea(e,t,r){let s=await e.getHarnessWorkspace(t);if(!s)throw new Error(`Harness workspace ${t} was not found.`);let a=(await e.listHarnessImprovementArtifacts(t,"evaluation_review",1e3)).find(n=>n.id===r.id&&n.contentHash===r.contentHash);if(!a||a.schemaVersion!=="openpond.harnessEvaluationReviewReceipt.v1")throw new Error("Harness Evaluation review was not found.");let o=a;if(o.classification!=="taskset"||o.nextAuthority!=="human_review")throw new Error("Only a human-review Taskset recommendation can enter Taskset authoring.");if(o.tasksetProposal)throw new Error("Harness Evaluation review already has a materialized Taskset link.");if(o.ownerScope.kind!==s.ownerScope.kind||o.ownerScope.id!==s.ownerScope.id)throw new Error("Harness Evaluation review owner scope does not match its workspace.");return o}i(ea,"requireTasksetReview");function Re(e){return`${e.id}:${e.contentHash}`}i(Re,"artifactKey");function Bt(e){return!!e&&typeof e=="object"&&!Array.isArray(e)}i(Bt,"isRecord");function Ft(e){let t=new Set;return e.filter(r=>{let s=JSON.stringify(r);return t.has(s)?!1:(t.add(s),!0)})}i(Ft,"uniqueObjects");import Wo from"node:path";function Vt(e,t,r){let s=new Date().toISOString();return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:`fixture_attempt_${t.id}_${r}`,tasksetId:e,taskId:t.taskId,split:Tt(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}})}i(Vt,"fixtureAttempt");function Tt(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}i(Tt,"artifactSplit");function Ns(e){let t=Wr(e.taskset),r=t.issues.filter(R=>R.severity==="error").map(R=>({code:R.code,message:R.message,path:R.path})),s=t.issues.filter(R=>R.severity==="warning").map(R=>({code:R.code,message:R.message,path:R.path})),a=typeof e.taskset.metadata.trainingMethod=="string"?e.taskset.metadata.trainingMethod:null,o=ta(a),n=o==="sft",c=o==="dpo",u=o==="grpo"||o==="ppo",g=new Set(e.taskset.learningSignals.demonstrations.filter(R=>R.approved&&R.taskId).map(R=>R.taskId)),p=e.taskset.metadata.learning===void 0?null:De.parse(e.taskset.metadata.learning),l=i(R=>p?p.admissions.some(G=>G.taskId===R.id&&G.supervisedTarget!==null):R.expectedOutput!=null,"hasSupervisedTarget"),d=e.taskset.datasetArtifact,m=new Set(js(e.taskset.metadata.approvedArtifactSignals)),f=!!d?.schema.fields.some(R=>R.semanticRole==="demonstration"||R.semanticRole==="expected_output"||R.semanticRole==="chosen"),T=e.taskset.tasks.filter(R=>R.split==="train"&&(l(R)||u)),y=n?T.filter(R=>g.has(R.id)):T,w=e.taskset.tasks.filter(R=>R.split==="train"&&l(R)&&!g.has(R.id)),_=e.taskset.tasks.filter(R=>R.split==="frozen_eval"&&(R.expectedOutput||u)),A=d&&f?d.splitCounts.train??0:0,I=d&&f?d.splitCounts.frozen_eval??0:0,S=d?m.has(o??""):!1,M=d?S?A:0:y.length,k=e.taskset.learningSignals.preferences.filter(R=>R.approved).length,x=e.taskset.learningSignals.rewards.filter(R=>R.approved&&R.executable).length,v=c?k:u?x?M:0:M,E=d?I:_.length;v===0&&r.push(n?{code:"sft_demonstrations_missing",message:"At least one explicitly approved training demonstration is required.",path:"learningSignals.demonstrations"}:c?{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 b=d?n&&!S?A:0:w.length;n&&b&&r.push({code:"sft_demonstrations_unapproved",message:`${b} training example${b===1?" is":"s are"} not explicitly approved.`,path:d?"metadata.approvedArtifactSignals":"learningSignals.demonstrations"}),E===0&&s.push({code:"frozen_eval_missing",message:"At least one independent evaluation example is recommended before making quality claims.",path:d?"datasetArtifact.splitCounts.frozen_eval":"tasks"});let D=new Set(y.map(R=>R.clusterKey)),$=new Set(_.map(R=>R.clusterKey)),z=d&&e.taskset.metadata.splitIsolationVerified===!0;(d?!z||v===0||E===0:D.size===0||$.size===0||[...D].some(R=>$.has(R)))&&s.push({code:"independent_evaluation_missing",message:"Independent source clusters are recommended before using evaluation results for quality claims.",path:d?"metadata.splitIsolationVerified":"tasks"}),(d?e.taskset.metadata.artifactRowsVerified!==!0:e.taskset.tasks.some(R=>typeof R.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:d?"metadata.artifactRowsVerified":"tasks.metadata.exampleOrigin"});let L=ra(e.taskset.metadata.diagnosis);(!L||typeof L.summary!="string"||!Array.isArray(L.stableBehavior))&&s.push({code:"capability_diagnosis_missing",message:"A capability diagnosis is recommended to separate stable behavior from changing knowledge.",path:"metadata.diagnosis"}),L?.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 J=e.taskset.graders.some(R=>R.rewardEligible&&(R.kind!=="model_judge"||R.metadata.calibrationIsAdvisory!==!0));J&&!e.graderAudit&&s.push({code:"grader_audit_missing",message:"Optional grader calibration evidence has not been recorded.",path:"graderFixtures"}),J&&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"}),J&&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"}),J&&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"}),J&&e.graderAudit&&!e.graderAudit.leakageChecksPassed&&r.push({code:"environment_leakage",message:"Environment or privileged-state leakage checks failed.",path:"environment"}),J&&e.graderAudit&&!e.graderAudit.infrastructureSafetyPassed&&r.push({code:"infrastructure_reward",message:"An infrastructure failure produced a score or eligible reward.",path:"graderFixtures"});let Xe=e.taskset.graders.some(R=>R.rewardEligible);u&&!Xe&&r.push({code:"online_reward_grader_missing",message:`${o?.toUpperCase()} requires a user-selected reward-eligible grader.`,path:"graders"});let ae=o??(v>0&&e.taskset.capabilities.compatibleMethods.includes("sft")?"sft":"none"),Me=e.taskset.learningSignals.demonstrations.filter(R=>R.approved).map(R=>R.id),C=ae==="none"?null:{primaryMethod:ae,bootstrap:ae==="grpo"&&Me.length?{method:"sft",purpose:"trajectory_bootstrap",demonstrationRefs:Me,limitations:["The SFT bootstrap imitates approved trajectories; it does not optimize verifier reward.","Completing the bootstrap does not satisfy the primary GRPO recommendation."]}:null},j=r.length===0,Z=r.length>0,ue=["sft","dpo","grpo","ppo"].map(R=>{let G=[],te=[];R==="sft"&&Me.length===0&&(G.push("demonstrations_missing"),te.push("Add at least one approved prompt/response demonstration.")),R==="dpo"&&k===0&&(G.push("preference_pairs_missing"),te.push("Add at least one approved chosen/rejected response pair.")),(R==="grpo"||R==="ppo")&&x===0&&(G.push("executable_reward_missing"),te.push("Implement and validate an executable scalar reward.")),R==="ppo"&&(G.push("value_model_required"),te.push("Choose and bind a value/critic model before PPO can execute.")),E===0&&(G.push("frozen_eval_missing"),te.push("Add an independent frozen-evaluation split.")),Z&&G.length===0&&(G.push("taskset_not_ready"),te.push("Resolve the Dataset validation and grader-audit blockers."));let _e=G.filter(Pe=>Pe!=="value_model_required");return{method:R,status:_e.length?"needs_dataset_work":R===ae?"recommended":"compatible",reasonCodes:G,reasons:te}});return Lr.parse({schemaVersion:"openpond.tasksetReadiness.v1",tasksetId:e.taskset.id,tasksetHash:e.taskset.contentHash,ready:j,recommendedMethod:ae,trainingPath:C,methodReadiness:ue,compatibleDestinationClasses:j?ae==="grpo"?["export","custom","hosted_managed"]:["export","custom"]:["export"],blockers:r,advisories:s,warnings:[...t.issues.filter(R=>R.severity==="warning").map(R=>R.message),...ae==="grpo"?["GRPO is readiness-compatible but not executable in the local CPU fixture."]:[],...js(e.taskset.metadata.warnings)],generatedAt:e.generatedAt??new Date().toISOString(),metadata:{reportHash:h([e.taskset.contentHash,r])}})}i(Ns,"buildTasksetReadiness");function ta(e){return e==="sft"||e==="dpo"||e==="grpo"||e==="ppo"||e==="sdft"||e==="opsd"||e==="sdpo"?e:null}i(ta,"trainingPathMethod");function ra(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}i(ra,"metadataRecord");function js(e){return Array.isArray(e)?e.filter(t=>typeof t=="string"&&!!t.trim()):[]}i(js,"metadataStrings");import Ws from"node:path";function qs(e,t,r){let s=De.parse(t.metadata.learning),a=ve(s.binding,s.rewards);return async({grader:o,task:n,attempt:c,signal:u})=>{let g=s.admissions.find(T=>T.taskId===n.id),p=a.find(T=>T.id===o.id);if(!g||p?.kind!=="custom_verifier")throw new Error("Prepared batch has no matching verifier admission.");let l=await ct(r,t),d=l?l.learningResources?.evidence.find(T=>xr(U(T),g.evidence)):await e.learningRepository().transaction(t.profileId,T=>he(T,"evidence",g.evidence));if(!d)throw new Error("Imported batch evidence is missing from its package.");let m=st(d,s.definition),f=i(T=>({id:T.id,input:T.input,expectedOutput:T.expectedOutput,policyVisibleContext:T.policyVisibleContext,privilegedContextRef:T.privilegedContextRef}),"selected");if(h(f(n))!==h(f(m)))throw new Error("Prepared task differs from its admitted evaluator context.");if(l){let T=l.learningResources?.assets.find(y=>y.id===p.verifierRef.id);if(!T)throw new Error("Imported batch verifier is missing from its package.");return je({source:re(T,p.verifierRef),exportName:p.exportName,timeoutMs:p.timeoutMs,signal:u,value:{task:m,attempt:{output:c.output,artifactRefs:c.artifactRefs,runtimeEventRefs:c.runtimeEventRefs,infrastructureError:c.infrastructureError},input:m.input,output:c.output,expectedOutput:m.expectedOutput,evaluatorContext:d.submission.evaluatorContext,infrastructureError:c.infrastructureError??null}})}return Je({repository:e.learningRepository(),scope:t.profileId,grader:p,task:m,evidence:{output:c.output,artifactRefs:c.artifactRefs,runtimeEventRefs:c.runtimeEventRefs,infrastructureError:c.infrastructureError},evaluatorContext:d.submission.evaluatorContext,signal:u})}}i(qs,"createLearningBatchVerifier");import{readFile as sa,realpath as $s}from"node:fs/promises";import Ls from"node:path";async function Us(e){let t=await $s(e.allowedRoot),r=await $s(Ls.resolve(t,e.grader.module));if(r!==t&&!r.startsWith(`${t}${Ls.sep}`))throw new Error("Verifier module is outside the approved Taskset root.");let s=await sa(r);if(e.grader.metadata.portableVerifierRef!==void 0){let o=kr.parse(e.grader.metadata.portableVerifierRef);if(o.sizeBytes!==s.length||o.contentHash!==fe(s))throw new Error("Verifier source differs from its immutable Taskset release.")}let a=new TextDecoder("utf-8",{fatal:!0}).decode(s);return je({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}:{}}})}i(Us,"runSandboxedVerifier");async function zt(e,t){if(!t.graders.some(n=>n.kind==="custom_verifier"))return;if(t.metadata.learning!==void 0)return qs(e.store,t,e.storeDir);if(t.metadata.rewardBinding!==void 0)return Cs(e.store,t,e.storeDir);let r=e.storeDir?null:await(e.loadProfileState??Gr)(),s=e.storeDir?Ws.join(e.storeDir,"training","tasksets",ce(t)):r?.sourcePath?Ws.join(r.sourcePath,"tasksets",t.id):null,a=typeof t.metadata.creationSnapshotId=="string"?t.metadata.creationSnapshotId:null,o=a?await e.store.getTaskDesignProposal(a):null;return s&&t.purpose!=="benchmark"&&t.environment.metadata.runtimeSourceTasksetId===void 0&&await nt(t,s,{generatedFiles:o?.generatedFiles??[]}),s?async({grader:n,task:c,attempt:u,signal:g})=>{let p=t.graders.find(d=>d.id===n.id);if(!p||p.kind!=="custom_verifier")throw new Error("Verifier is not declared by this Taskset revision.");let l;if(t.environment.entrypoint===Y&&!u.infrastructureError){if(!e.storeDir)throw new Error("Tool grading requires the execution owner's artifact directory.");l=await vt({store:e.store,storeDir:e.storeDir,taskset:t,task:c,attempt:u})}return Us({grader:p,task:c,attempt:u,allowedRoot:s,signal:g,evaluatorContext:l})}:void 0}i(zt,"createTasksetEvaluationVerifier");var oa=new Set(["correct","incorrect","parse_failure","budget_exhausted","tool_schema_violation","infrastructure_failure","cancelled"]);async function Bs(e){let t=na(e.attempt);if(t){let g=h(e.graders),p=`Taskset Work attempt ended with ${t.failureClass}: `+t.message;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([e.attempt.id,g,t]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:g,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:ia(e.attempt.metadata);if(!r)try{return await at(e)}catch(g){return aa({attempt:e.attempt,graders:e.graders,now:e.now,error:g})}let s=h(e.graders),a=r.rewardEligible?r.outcome==="correct"?1:0:null,o=r.outcome==="correct",n=da(r.outcome),c=`Cross-System trajectory verifier outcome: ${r.outcome}.`,u=e.graders.map(g=>({graderId:g.id,graderVersion:g.version,score:a??0,passed:o,hardGate:g.hardGate,rewardEligible:r.rewardEligible&&g.rewardEligible,feedback:c,evidenceRefs:e.attempt.artifactRefs,judge:g.kind==="model_judge"?g.judge:null,calibrationStatus:g.kind==="model_judge"?g.calibrationStatus:"not_applicable"}));return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([e.attempt.id,s,r]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:s,score:a,passed:o,components:u,failureClass:n,feedback:[c],rewardEligible:r.rewardEligible,createdAt:e.now?.()??new Date().toISOString()}}i(Bs,"gradeTasksetEvaluationAttempt");function aa(e){let t=h(e.graders),r=(e.error instanceof Error?e.error.message:String(e.error)).slice(0,19e3),s=`Taskset grader failed: ${r}`;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([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()}}i(aa,"graderFailureResult");function na(e){if(e.metadata.execution!=="taskset_work"||!e.infrastructureError)return null;let t=Vr.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}}i(na,"tasksetWorkFailure");function ia(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"||!oa.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}}i(ia,"crossSystemGrade");function da(e){return e==="correct"?null:e==="infrastructure_failure"?"infrastructure_failure":e==="cancelled"?"cancelled":"policy_failure"}i(da,"failureClassFor");import Ge from"node:path";import{mkdir as Fs,readFile as ca,writeFile as Vs}from"node:fs/promises";async function Q(e){let t=Ge.join(e.storeDir,"training","evaluation-artifacts",e.tasksetId),r=Ge.join(t,`${e.attemptId}-${zs(e.fileLabel??e.kind)}.json`),s=Buffer.from(`${JSON.stringify({schemaVersion:"openpond.rawEvaluationArtifact.v1",requestId:e.requestId,...e.payload},null,2)}
 `,"utf8");return await Fs(t,{recursive:!0}),await Vs(r,s,{mode:384}),Jt({...e,file:r,bytes:s,mediaType:"application/json"})}i(Q,"persistJsonTaskAttemptArtifact");async function wc(e){let t=e.extension.replace(/^\.+/,"");if(!/^[a-z0-9]{1,12}$/i.test(t))throw new Error("Task Attempt artifact extension is invalid.");let r=Ge.join(e.storeDir,"training","evaluation-artifacts",e.tasksetId);await Fs(r,{recursive:!0});let s=Ge.join(r,`${e.attemptId}-${zs(e.fileLabel)}.${t}`);return await Vs(s,e.bytes,{mode:384}),Jt({...e,file:s})}i(wc,"persistBinaryTaskAttemptArtifact");function zs(e){let t=e.trim().replace(/[^a-zA-Z0-9._-]+/g,"-").replace(/^-+|-+$/g,"");if(!t)throw new Error("Evaluation artifact file label is empty.");return t}i(zs,"safeFileLabel");async function Js(e){let t=await ca(e.path),r=fe(t);if(t.byteLength!==e.expectedSizeBytes||r!==e.expectedSha256)throw new Error(`Saved Work output ${Ge.basename(e.path)} no longer matches its OutputRef.`);return Jt({...e,file:e.path,bytes:t,kind:"output_artifact",metadata:{...e.metadata,source:"openpond-work",outputValidated:!0}})}i(Js,"persistTaskAttemptOutputArtifact");async function Jt(e){let t=fe(e.bytes),r=zr.parse({schemaVersion:"openpond.taskAttemptArtifact.v1",id:`attempt_artifact_${h([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)}i(Jt,"saveArtifact");function Gt(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?[]:[ls.parse({schemaVersion:B,id:`cso_bootstrap_${a.id}`,taskId:n.id,worldId:a.worldId,trajectoryId:a.id,toolContractHash:q,approval:{status:"approved",approvedBy:e.approvedBy,approvedAt:e.approvedAt},messages:la(n,a)})]})}i(Gt,"buildCrossSystemBootstrapDataset");function la(e,t){let r=[{role:"system",content:ds},{role:"user",content:e.prompt}];for(let s of t.steps)s.kind==="model"&&s.content.trim()&&r.push({role:"assistant",content:s.content}),s.kind==="tool_call"&&r.push({role:"assistant",content:null,tool_calls:[{id:s.callId,type:"function",function:{name:s.name,arguments:JSON.stringify(s.arguments)}}]}),s.kind==="tool_result"&&r.push({role:"tool",tool_call_id:s.callId,content:JSON.stringify({ok:s.ok,result:s.result,error:s.error})}),s.kind==="final"&&r.push({role:"assistant",content:s.content});return r}i(la,"trajectoryMessages");import cn from"node:path";import{mkdir as ua,mkdtemp as ma,rm as pa}from"node:fs/promises";import fa from"node:path";async function Gs(e,t,r){await ua(e,{recursive:!0});let s=await ma(fa.join(e,".authoring-"));try{return await nt(t,s,{generatedFiles:r}),await Yt(e,t,s)}finally{await pa(s,{recursive:!0,force:!0})}}i(Gs,"materializeGeneratedTasksetPackage");async function Yt(e,t,r){let s=await Br(r),a=await Qr(t,r),o=`authored-${h(a.taskset)}`,n=ke.parse({...a.taskset,environment:{...a.taskset.environment,metadata:{...a.taskset.environment.metadata,runtimeSourceTasksetId:o}}}),c=ke.parse({...n,contentHash:Te(n)});return await ms(e,{taskset:c,generatedFiles:a.generatedFiles},o,{source:{directory:r,packageHash:s},verify:i(u=>pt(u,c),"verify")}),c}i(Yt,"materializeTasksetRevisionFromSource");import{Buffer as Xt}from"node:buffer";import{spawn as ga}from"node:child_process";var Ys=String.raw`
 import contextlib, io, json, math, statistics, decimal, datetime, collections, itertools, functools, sys
 
@@ -83,4 +83,4 @@ _result = sum(values)`}})}i(Fa,"toolCallsFor");function Va(e){let t=new Date(`${
 
 `);throw new Error(`Evaluation task ${e.id} has no policy-visible prompt.`)}i(le,"tasksetWorkUserPrompt");function jn(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(jn,"asRecord");import So from"node:path";function bo(e){let t=[];e.data&&typeof e.data=="object"&&!Array.isArray(e.data)&&t.push(e.data);try{let r=JSON.parse(e.contentText);if(r&&typeof r=="object"&&!Array.isArray(r)){let s=r.data;s&&typeof s=="object"&&!Array.isArray(s)&&t.push(s)}}catch{}return[...new Set(t.flatMap(r=>(Array.isArray(r.files)?r.files:[]).flatMap(a=>{let o=typeof a=="string"?a:a&&typeof a=="object"&&!Array.isArray(a)?a.path:null;if(typeof o!="string")return[];let n=So.posix.normalize(o.replaceAll("\\","/"));return n.startsWith("outputs/")?[n.slice(8)]:n.startsWith("/")||n.startsWith("../")?[]:[n]})))]}i(bo,"listedWorkOutputPaths");function Ro(e){let t=e.availableOutputs.filter(r=>{if(e.claimedSourcePaths.has(r))return!1;let s=So.posix.extname(r).toLowerCase();return rs[s]===e.mediaType});return t.length===1?t[0]:null}i(Ro,"uniqueCompatibleOutputPath");function _o(e,t){if(t===null)return{signal:e,dispose(){}};let r=new AbortController,s=i(()=>r.abort(e.reason??new Error("The Work evaluation was cancelled.")),"abortFromParent");e.aborted?s():e.addEventListener("abort",s,{once:!0});let a=setTimeout(()=>{r.abort(new Error(`Work model request exceeded ${t} ms.`))},t);return a.unref?.(),{signal:r.signal,dispose(){clearTimeout(a),e.removeEventListener("abort",s)}}}i(_o,"boundedRequestSignal");async function xo(e,t){return Be(t),await new Promise((r,s)=>{let a=i(()=>s(t.reason instanceof Error?t.reason:new Error("The Work model request was cancelled.")),"rejectForAbort");t.addEventListener("abort",a,{once:!0}),e.next().then(r,s).finally(()=>{t.removeEventListener("abort",a)})})}i(xo,"nextTasksetWorkModelDelta");function Be(e){if(e.aborted)throw e.reason instanceof Error?e.reason:new Error("The Work evaluation was cancelled.")}i(Be,"throwIfAborted");function Ot(e){return e instanceof Error?e.message:String(e)}i(Ot,"errorMessage");function Io(e,t){return Math.max(0,Date.parse(t)-Date.parse(e))}i(Io,"elapsedMilliseconds");async function Co(e){if(e.taskset.environment.kind!=="work")throw new Error(`Taskset ${e.taskset.id} does not select Work.`);let t=e.timestamp??(()=>new Date().toISOString()),r=t(),s=`taskset-work:${h([e.taskset.id,e.taskset.contentHash,e.task.id,e.model,e.seed,e.attempt,r]).slice(0,40)}`,a=e.resultId??`attempt_${h([s,"work"]).slice(0,24)}`,o=`taskset_work_turn_${h([a,e.task.id,r]).slice(0,24)}`,n=new AbortController,c=!1,u=i(()=>n.abort(e.signal?.reason??new Error("The Work evaluation was cancelled.")),"abortFromParent");e.signal?.aborted?u():e.signal?.addEventListener("abort",u,{once:!0});let g=Math.max(1,Math.min(e.taskset.environment.defaultTimeoutMs,60*6e4)),p=setTimeout(()=>{c=!0,n.abort(new Error(`Work evaluation exceeded ${g} ms.`))},g);p.unref?.();let l=[],d=[],m=[],f=new Map,T=new Map,y=null,w=null,_=[],A="",I="completed",S=null,M=null,k=0,x=!1,v="assets";try{Be(n.signal);let C=Hs(),j=C?await C.resolveWorkAssets(e.taskset,e.task):await Mt({storeDir:e.storeDir,taskset:e.taskset,task:e.task});v="session",w=await e.runtime.createSession({experience:"work",provider:e.model.providerId,modelRef:e.model,openPondCommandAccessMode:"disabled",hiddenFromDefaultSidebar:!0,title:`Taskset Work attempt: ${e.task.id}`,cwd:null,metadata:{automatedTasksetWorkAttempt:!0,tasksetId:e.taskset.id,tasksetName:e.taskset.name,tasksetHash:e.taskset.contentHash,taskId:e.task.id,attemptId:a,requestId:s,parentModelRunId:e.parentModelRunId??null}}),await e.store.insertTurn(es.parse({id:o,sessionId:w.id,providerTurnId:null,modelRef:e.model,prompt:le(e.task),startedAt:r,completedAt:null,status:"in_progress",error:null,metadata:{automatedTasksetWorkAttempt:!0,tasksetId:e.taskset.id,taskId:e.task.id,attemptId:a,parentModelRunId:e.parentModelRunId??null},createImproveRun:null,profileSnapshot:null,harnessSnapshot:null})),await yo({store:e.store,session:w,turnId:o,prompt:le(e.task),tasksetId:e.taskset.id,taskId:e.task.id,attemptId:a}),v="environment",_=[...ws({executeWorkspaceTool:e.runtime.executeWorkspaceTool,inputs:j}),...e.additionalToolDefinitions??[]];let Z=new Map(_.map(P=>[P.name,P])),ue=e.taskset.environment.toolNames.filter(P=>!Z.has(P));if(ue.length)throw new Error(`Work environment declares unknown tools: ${ue.join(", ")}.`);if(!e.taskset.environment.toolNames.includes("work_save_output"))throw new Error("Work Tasksets must declare work_save_output.");let R=e.taskset.environment.toolNames.map(P=>Z.get(P)),G=R.map(ys),te=new Map(R.map(P=>[P.name,P])),_e=te.get("work_save_output");if(!_e)throw new Error("The Work output persistence tool is unavailable.");let Pe=te.get("work_list_files")??null,xe=To(e.task,j,e.harnessInstructionContext),mr=fo(e.taskset),Bo=go(e.taskset),Fo=ho(e.taskset),Nt=ko(e.taskset),Oe=new Set((e.task.requiredOutputs??[]).map(P=>P.path)),Fe=!1,pr=_.find(P=>P.name==="work_environment");if(!pr)throw new Error("The Work environment readiness tool is unavailable.");let fr=await Ae({definition:pr,runtime:e.runtime,sessionId:w.id,turnId:o,model:e.model,callId:`environment_${a}`,args:{},signal:n.signal,userPrompt:le(e.task)});if(!fr.ok)throw new Error(fr.contentText);v="model";for(let P=0;P<mr;P+=1){Be(n.signal);let ne=new ft,ie="",we=null,Ze=_o(n.signal,Fo);try{let H=e.stream({model:e.model,reasoningEffort:e.reasoningEffort??null,messages:xe,tools:G,toolChoice:"auto",requestId:`${s}:${P}`,signal:Ze.signal,maxOutputTokens:e.sampling?.maxOutputTokens??4096,temperature:e.sampling?.temperature??0,topP:e.sampling?.topP??1,seed:e.seed+e.attempt,hostedTokenPricing:e.hostedTokenPricing})[Symbol.asyncIterator]();for(;;){let me=await xo(H,Ze.signal);if(me.done)break;let N=me.value;N.text&&(ie+=N.text),N.continuation&&(we=N.continuation),N.toolCalls?.length&&ne.append(N.toolCalls),N.usage!==void 0&&d.push(N.usage),typeof N.costUsd=="number"&&Number.isFinite(N.costUsd)&&N.costUsd>=0&&m.push(N.costUsd)}}catch(H){throw Ze.signal.aborted&&!n.signal.aborted&&(x=!0),H}finally{Ze.dispose()}Be(n.signal);let Qe=ne.completed();if(await wo({store:e.store,session:w,turnId:o,text:ie}),l.push({kind:"model",turn:P,text:ie,toolCallCount:Qe.length}),!Qe.length){if(A=ie,xe.push({role:"assistant",content:ie,...we?{continuation:we}:{}}),Oe.size>0&&k<Bo){k+=1,xe.push({role:"user",content:["The turn ended before every declared deliverable was saved.","Resume the task now; this is not a request for a status summary.","Create and inspect the missing files, then call work_save_output for every exact required path.",`Required paths: ${JSON.stringify([...Oe])}`].join(`
 `)});continue}break}xe.push(gs(ie,Qe,{continuation:we,maxArgumentCharacters:6e3}));let gr=!1;for(let H of Qe){let me=te.get(H.name);if(!me){let W=ks(H);await Ke({store:e.store,session:w,turnId:o,callId:H.id,name:H.name,args:{},result:W}),l.push({kind:"tool",turn:P,callId:H.id,name:H.name,arguments:{},ok:!1,output:W.contentText}),xe.push(gt(W,{maxContentCharacters:6e3}));continue}let N;try{if(N=ht(H),H.name==="work_exec"&&Nt!==null){let W=typeof N.timeoutSeconds=="number"&&Number.isFinite(N.timeoutSeconds)&&N.timeoutSeconds>0?N.timeoutSeconds:Nt;N={...N,timeoutSeconds:Math.min(W,Nt)}}}catch(W){let He=Ot(W),Ie=hs(H,He);await Ke({store:e.store,session:w,turnId:o,callId:H.id,name:H.name,args:{},result:Ie}),l.push({kind:"tool",turn:P,callId:H.id,name:H.name,arguments:{},ok:!1,output:Ie.contentText}),xe.push(gt(Ie,{maxContentCharacters:6e3}));continue}await ar({store:e.store,session:w,turnId:o,callId:H.id,name:H.name,args:N});let Ve=i(()=>Ae({definition:me,runtime:e.runtime,sessionId:w.id,turnId:o,model:e.model,callId:H.id,args:N,signal:n.signal,userPrompt:le(e.task)}),"executeTool"),pe=e.toolEvidence?await e.toolEvidence.execute({taskId:e.task.id,callId:H.id,toolName:H.name,args:N,execute:Ve}):await Ve();await nr({store:e.store,session:w,turnId:o,callId:H.id,name:H.name,result:pe}),l.push({kind:"tool",turn:P,callId:H.id,name:H.name,arguments:N,ok:pe.ok,output:pe.contentText});let de=Pt(N,pe);if(de&&(f.set(de.relativePath,de),Fe=Oe.size>0&&[...Oe].every(W=>f.has(W))),pe.ok&&(H.name==="work_write_file"||H.name==="work_edit_file"||H.name==="work_exec")&&(gr=!0),xe.push(gt(pe,{maxContentCharacters:6e3})),Fe)break}if(!Fe&&gr&&Pe){let H=`discover_outputs_${a}_${P}`,me={area:"outputs",recursive:!0},N=await Ae({definition:Pe,runtime:e.runtime,sessionId:w.id,turnId:o,model:e.model,callId:H,args:me,signal:n.signal,userPrompt:le(e.task)});await Ke({store:e.store,session:w,turnId:o,callId:H,name:Pe.name,args:me,result:N}),l.push({kind:"tool",turn:P,callId:H,name:Pe.name,arguments:me,ok:N.ok,output:N.contentText});let Ve=bo(N),pe=new Set([...f.values()].map(de=>de.relativePath));for(let de of e.task.requiredOutputs??[]){let W=de.path;if(f.has(W))continue;let He=Ve.includes(W)?W:Ro({availableOutputs:Ve,claimedSourcePaths:pe,mediaType:de.mediaType});if(!He)continue;let Ie=`persist_output_${h([a,P,W,He]).slice(0,20)}`,et={path:He,suggestedName:Eo.posix.basename(W),validation:[]},tt=await Ae({definition:_e,runtime:e.runtime,sessionId:w.id,turnId:o,model:e.model,callId:Ie,args:et,signal:n.signal,userPrompt:le(e.task)});await Ke({store:e.store,session:w,turnId:o,callId:Ie,name:_e.name,args:et,result:tt}),l.push({kind:"tool",turn:P,callId:Ie,name:_e.name,arguments:et,ok:tt.ok,output:tt.contentText});let hr=Pt(et,tt);hr&&(f.set(W,{...hr,relativePath:W}),pe.add(He))}Fe=Oe.size>0&&[...Oe].every(de=>f.has(de))}if(Fe){clearTimeout(p);break}P===mr-1&&(I="budget_exhausted")}v="outputs",Be(n.signal);for(let P of e.task.requiredOutputs??[]){let ne=f.get(P.path)??null;if(!ne){let we=await Ae({definition:_e,runtime:e.runtime,sessionId:w.id,turnId:o,model:e.model,callId:`save_${h([a,P.path]).slice(0,20)}`,args:{path:P.path,suggestedName:Eo.posix.basename(P.path),validation:[]},signal:n.signal,userPrompt:le(e.task)});if(ne=Pt({path:P.path},we),ne&&f.set(ne.relativePath,ne),!we.ok||!ne){l.push({kind:"required_output",path:P.path,ok:!1,detail:we.contentText});continue}}let ie=await po({requiredOutput:P,saved:ne,validateRequiredOutput:e.validateRequiredOutput});T.set(P.path,ie),l.push({kind:"required_output",path:P.path,ok:ie.passed,detail:ie.detail})}}catch(C){S=Ot(C),n.signal.aborted?(I=c?"timeout":"cancelled",M=c?"timeout":"cancelled"):x?(I="timeout",M="timeout"):v==="assets"||v==="environment"?(I="environment_failure",M="environment_failure"):(I="infrastructure_failure",M="infrastructure_failure")}finally{if(clearTimeout(p),e.signal?.removeEventListener("abort",u),w){v="cleanup";let C=_.find(j=>j.name==="work_stop");if(C)try{let j=await Ae({definition:C,runtime:e.runtime,sessionId:w.id,turnId:o,model:e.model,callId:`cleanup_${a}`,args:{},signal:new AbortController().signal,userPrompt:le(e.task)});if(l.push({kind:"cleanup",ok:j.ok,detail:j.contentText}),j.ok&&(y=or(j.data),!y&&e.runtime.settleCostEvidence)){let Z=await e.runtime.settleCostEvidence(w.id,{turnId:o});Z.ok&&(y=or(Z.data))}j.ok||(S=`Work cleanup failed: ${j.contentText}`,I="infrastructure_failure",M="infrastructure_failure")}catch(j){S=`Work cleanup failed: ${Ot(j)}`,I="infrastructure_failure",M="infrastructure_failure",l.push({kind:"cleanup",ok:!1,detail:S})}}}let E=[],b=new Map;for(let C of f.values()){if(C.outputRef.location.kind!=="local")continue;let j=await Js({store:e.store,tasksetId:e.taskset.id,taskId:e.task.id,attemptId:a,requestId:s,path:C.artifactPath,mediaType:C.outputRef.contentType,expectedSha256:C.outputRef.sha256,expectedSizeBytes:C.outputRef.sizeBytes,timestamp:t,metadata:{requiredOutputPath:C.relativePath,outputRefId:C.outputRef.id,outputRefRevision:C.outputRef.revision}});E.push(j.id),b.set(C.relativePath,j.id)}let D=t();if(w){let C=I==="cancelled"||I==="timeout"?"interrupted":I==="environment_failure"||I==="infrastructure_failure"?"failed":"completed";await vo({store:e.store,session:w,turnId:o,status:C,error:C==="completed"?null:S}),await e.store.updateTurn(o,j=>({...j,completedAt:D,status:C,error:C==="completed"?null:S}))}let $=w?await e.runtime.runtimeEventsForSession(w.id):[],z=await Q({store:e.store,storeDir:e.storeDir,tasksetId:e.taskset.id,taskId:e.task.id,attemptId:a,requestId:s,kind:"runtime_trace",payload:{schemaVersion:"openpond.tasksetWorkTrace.v1",tasksetId:e.taskset.id,tasksetHash:e.taskset.contentHash,taskId:e.task.id,model:e.model,seed:e.seed,attempt:e.attempt,status:I,steps:l,runtimeEventRefs:$.map(C=>C.id),usage:d,infrastructureError:S,failureClass:M,startedAt:r,completedAt:D},timestamp:t});E.push(z.id);let L=(e.task.requiredOutputs??[]).map(C=>{let j=f.get(C.path),Z=[...l].reverse().find(ue=>ue.kind==="required_output"&&ue.path===C.path);return{path:C.path,mediaType:C.mediaType,validationKinds:Array.isArray(C.metadata.validationKinds)?C.metadata.validationKinds.filter(ue=>typeof ue=="string"):[],schemaRef:C.schemaRef??null,artifactId:j?b.get(C.path)??null:null,outputRefId:j?.outputRef.id??null,sha256:j?.outputRef.sha256??null,sizeBytes:j?.outputRef.sizeBytes??null,passed:Z?.kind==="required_output"?Z.ok:!1,detail:Z?.kind==="required_output"?Z.detail:"Required output was not saved.",...T.get(C.path)?.parsedJson!==void 0?{parsedJson:T.get(C.path)?.parsedJson}:{}}}),J=e.hostedTokenPricing?d.flatMap(C=>{let j=Se(C,e.hostedTokenPricing);return j===null?[]:[j]}):[],Xe=m.length?Ce(m):J.length?Ce(J):S&&d.length===0?0:null,ae=[Xe,y?.billableUsd??null].filter(C=>C!==null),Me=ae.length?Ce(ae):null;return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:a,tasksetId:e.taskset.id,taskId:e.task.id,split:e.task.split,attempt:e.attempt,seed:e.seed,modelRef:e.model,startedAt:r,completedAt:D,output:{text:A,requiredOutputs:L,outputsPassed:L.length>0&&L.every(C=>C.passed),toolFailureCount:l.filter(C=>C.kind==="tool"&&!C.ok).length,missingOutputRecoveryTurns:k},runtimeEventRefs:$.map(C=>C.id),artifactRefs:E,privilegedOutcomeRef:e.task.privilegedContextRef,infrastructureError:S,costUsd:Me,latencyMs:Io(r,D),userInterventions:0,metadata:{requestId:s,parentModelRunId:e.parentModelRunId??null,execution:"taskset_work",runtimeProfileId:e.taskset.environment.entrypoint,tasksetHash:e.taskset.contentHash,status:I,failureClass:M,sessionId:w?.id??null,turnId:o,toolNames:e.taskset.environment.toolNames,harnessCapabilityReceipt:e.harnessCapabilityReceipt?{...e.harnessCapabilityReceipt,executableToolNames:e.taskset.environment.toolNames}:null,hostedTokenPricing:e.hostedTokenPricing??null,assetHashes:(e.task.assets??[]).map(C=>C.sha256),requiredOutputPaths:(e.task.requiredOutputs??[]).map(C=>C.path),usage:d,costEvidence:{providerInferenceUsd:Xe,workRuntimeUsd:y?.totalUsd??null,workRuntimeBillableUsd:y?.billableUsd??null,workRuntimeSimulatedUsd:y?.simulatedUsd??null,combinedUsd:Me,workReceiptIds:y?.receiptIds??[],workDurationSeconds:y?.durationSeconds??null,settlementModes:y?.settlementModes??[]}}})}i(Co,"runTasksetWorkAttempt");function Ao(e){let t=0,r=null,s=new Map;return{policy:i(async({messages:o,tools:n,signal:c})=>{c.throwIfAborted();let u=0,g=o.map(f=>{if(f.role==="tool")return{role:"tool",tool_call_id:f.callId,name:f.name,content:JSON.stringify(f.observation)};if(f.role==="assistant"){let T=s.get(u++);return{role:"assistant",content:f.text||null,...T?{continuation:T}:{},tool_calls:f.toolCalls.map(y=>({id:y.id,type:"function",function:{name:y.name,arguments:JSON.stringify(y.arguments)}}))}}return{role:f.role,content:f.text}});ze(g,8388608);let p=t++,l=new ft,d="",m=0;for await(let f of e.stream({model:e.model,reasoningEffort:e.reasoningEffort??null,messages:g,tools:n.map(T=>({type:"function",function:{name:T.name,description:T.description,parameters:T.inputSchema}})),toolChoice:"auto",requestId:`${e.requestId}:${p}`,signal:c,maxOutputTokens:e.sampling?.maxOutputTokens??4096,temperature:e.sampling?.temperature??0,topP:e.sampling?.topP??1,seed:e.seed,hostedTokenPricing:e.hostedTokenPricing})){c.throwIfAborted();let T=ir({text:f.text,continuation:f.continuation,toolCalls:f.toolCalls?.map(y=>ir({...y,function:y.function?ir(y.function):void 0}))});if(ze(T,1048576),m+=new TextEncoder().encode(JSON.stringify(T)).byteLength,m>1048576)throw new Error("Model tool response exceeded its byte budget.");f.text&&(d+=f.text),f.toolCalls&&l.append(f.toolCalls),f.continuation&&s.set(p,structuredClone(f.continuation)),typeof f.costUsd=="number"&&Number.isFinite(f.costUsd)&&f.costUsd>=0&&(r=(r??0)+f.costUsd)}return c.throwIfAborted(),{text:d,toolCalls:l.completed({strict:!0}).map(f=>({id:f.id,name:f.name,arguments:ht(f)}))}},"policy"),costUsd:i(()=>r,"costUsd")}}i(Ao,"createStarterToolPolicy");function ir(e){return Object.fromEntries(Object.entries(e).filter(([,t])=>t!==void 0))}i(ir,"definedFields");async function Ht(e){let{attemptInput:t,taskset:r}=e,s=e.policySource==="fixture"?null:t.model,a=await Ue(e.store,r,t.task,e.storeDir),o=e.timestamp(),n=`starter-tool-${h({taskset:U(r),task:t.task.id,model:s,seed:t.seed,attempt:t.attempt,startedAt:o}).slice(0,32)}`,c=e.resultId??n;if(!/^[a-zA-Z0-9_-]{1,200}$/.test(r.id)||!/^[a-zA-Z0-9_-]{1,200}$/.test(c))throw new Error("Tool attempt identifiers must be safe artifact identities.");let u=Ao({stream:e.stream,model:t.model,reasoningEffort:t.reasoningEffort,requestId:n,seed:t.seed+t.attempt,sampling:t.sampling,hostedTokenPricing:e.hostedTokenPricing}),g=await vs({taskId:t.task.id,definition:a.execution.javascript,asset:a.module,initialState:a.initialState,input:t.task.input,instructions:`${a.instructions}
-${JSON.stringify({policyVisibleContext:t.task.policyVisibleContext})}`,seed:t.seed,timeoutMs:a.execution.environment.contract.defaultTimeoutMs,signal:t.signal,execute:Ts,policy:u.policy}),p=e.timestamp(),l=["environment_failure","timed_out","cancelled"].includes(g.status)?g.error??g.status:null,d={schemaVersion:"openpond.starterToolAttempt.v1",taskset:U(r),taskHash:h(t.task),model:s,seed:t.seed,attempt:t.attempt,policySource:e.policySource??"model",definition:a.definition,environment:U(a.execution.environment),verifierSet:U(a.execution.verifierSet),result:g};ze(d,12582912);let m=await Q({store:e.store,storeDir:e.storeDir,tasksetId:r.id,taskId:t.task.id,attemptId:c,requestId:n,kind:"environment_state",payload:d,timestamp:e.timestamp});return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:c,tasksetId:r.id,taskId:t.task.id,split:t.task.split,attempt:t.attempt,seed:t.seed,modelRef:s,startedAt:o,completedAt:p,output:{text:g.output??""},runtimeEventRefs:[],artifactRefs:[m.id],privilegedOutcomeRef:m.id,infrastructureError:l,costUsd:u.costUsd(),latencyMs:Math.max(0,Date.parse(p)-Date.parse(o)),userInterventions:0,metadata:{requestId:n,execution:"starter_tool_environment",policySource:e.policySource??"model",environmentStatus:g.status,environmentCleanupComplete:g.environmentCleanupComplete,environmentAttemptHash:g.contentHash}})}i(Ht,"runStarterToolAttempt");async function Po(e){let t=e.timestamp??(()=>new Date().toISOString()),r=await Ee(e.store,e.attemptInput.tasksetId,e.attemptInput.tasksetRef);if(r.environment.entrypoint===Y)return Ht({store:e.store,storeDir:e.storeDir,taskset:r,attemptInput:e.attemptInput,stream:e.crossSystemStream,timestamp:t,resultId:e.resultId,hostedTokenPricing:e.hostedTokenPricing});if(r.environment.kind==="work"){if(!e.work)throw new Error("Taskset Work execution is not configured for this evaluator.");return Co({store:e.store,storeDir:e.storeDir,taskset:r,task:e.attemptInput.task,model:e.attemptInput.model,reasoningEffort:e.attemptInput.reasoningEffort,seed:e.attemptInput.seed,attempt:e.attemptInput.attempt,sampling:e.attemptInput.sampling,signal:e.attemptInput.signal,stream:e.work.stream,runtime:e.work.runtime,timestamp:e.timestamp,resultId:e.resultId,parentModelRunId:e.parentModelRunId,harnessInstructionContext:e.harnessInstructionContext,validateRequiredOutput:e.work.validateRequiredOutput,additionalToolDefinitions:e.work.additionalToolDefinitions,toolEvidence:e.work.toolEvidence,harnessCapabilityReceipt:e.work.harnessCapabilityReceipt,hostedTokenPricing:e.work.hostedTokenPricing})}if(!Mo(r)&&(r.environment.kind!=="chat"||r.environment.stateful||r.environment.toolNames.length||r.capabilities.requiresTools||r.capabilities.requiresState))throw new Error("The Taskset environment has no registered evaluation adapter.");return Mo(r)?Nn({...e,timestamp:t,taskset:r}):qn({...e,timestamp:t})}i(Po,"runPostTrainingEvaluationAttempt");function Mo(e){let t=e.metadata.flagship==="cross-system-operations",r=e.environment.metadata.flagship==="cross-system-operations";return(t||r)&&e.environment.stateful&&e.environment.metadata.toolContractHash===q}i(Mo,"isCrossSystemTaskset");async function Nn(e){let{attemptInput:t,taskset:r}=e,s=e.timestamp(),a=Oo(t,s),o=uo(r,{taskId:t.task.id,prompt:typeof t.task.input.prompt=="string"?t.task.input.prompt:null}),n=new AbortController,c=i(()=>n.abort(t.signal?.reason??new Error("The evaluation was cancelled.")),"abortFromParent");t.signal?.addEventListener("abort",c,{once:!0});let u=Math.max(1,Math.min(r.environment.defaultTimeoutMs,10*6e4)),g=setTimeout(()=>n.abort(new Error(`Cross-System evaluation exceeded ${u} ms.`)),u);g.unref?.();try{let p=[],l=null,d=null;for(let w=0;w<=1;w+=1){let _=await io({world:o.world,task:o.generatedTask,model:t.model,reasoningEffort:null,stream:e.crossSystemStream,signal:n.signal,trajectoryId:`cso_post_training_eval_${h([a,w]).slice(0,24)}`,metadata:{evaluation:"post_training",execution:"post_training_evaluation_tool_loop",tasksetId:r.id,tasksetHash:r.contentHash,samplingSeed:t.seed,attempt:t.attempt,infrastructureRetryAttempt:w,priorInfrastructureErrors:p.map(A=>A.infrastructureError??"Unknown infrastructure failure.")}});if(l=_,d=xt({task:o.generatedTask,trajectory:_}),d.outcome!=="infrastructure_failure"||w===1)break;p.push(_)}if(!l||!d)throw new Error(`Cross-System task ${t.task.id} produced no trajectory.`);let m=[...l.steps].reverse().find(w=>w.kind==="final"),f=l.completedAt,T=e.resultId??`attempt_${h([a,l.id,d.outcome]).slice(0,24)}`,y=await Q({store:e.store,storeDir:e.storeDir,tasksetId:r.id,taskId:t.task.id,attemptId:T,requestId:a,kind:"runtime_trace",payload:{schemaVersion:B,model:t.model,seed:t.seed,attempt:t.attempt,trajectory:l,verifier:d,priorInfrastructureTrajectories:p},timestamp:e.timestamp});return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:T,tasksetId:r.id,taskId:t.task.id,split:t.task.split,attempt:t.attempt,seed:t.seed,modelRef:t.model,startedAt:l.startedAt,completedAt:f,output:{text:m?.content??"",trajectoryId:l.id},runtimeEventRefs:[],artifactRefs:[y.id],privilegedOutcomeRef:t.task.privilegedContextRef,infrastructureError:l.infrastructureError,costUsd:null,latencyMs:dr(l.startedAt,f),userInterventions:0,metadata:{requestId:a,execution:"post_training_evaluation_tool_loop",trajectoryId:l.id,worldId:l.worldId,toolContractHash:l.toolContractHash,verifierOutcome:d.outcome,verifierReward:d.reward,verifierRewardEligible:d.rewardEligible,infrastructureRetryAttempt:l.metadata.infrastructureRetryAttempt,priorInfrastructureErrors:l.metadata.priorInfrastructureErrors}})}finally{clearTimeout(g),t.signal?.removeEventListener("abort",c)}}i(Nn,"runCrossSystemAttempt");async function qn(e){let{attemptInput:t}=e,r=e.timestamp(),s=Oo(t,r);try{let a=await e.modelText({model:t.model,reasoningEffort:null,signal:t.signal??new AbortController().signal,requestId:s,messages:$n(t.task),maxOutputTokens:t.sampling?.maxOutputTokens??2048,temperature:t.sampling?.temperature??.8,topP:t.sampling?.topP??.95,seed:t.seed+t.attempt}),o=e.timestamp(),n=e.resultId??`attempt_${h([s,a]).slice(0,24)}`,c=await Q({store:e.store,storeDir:e.storeDir,tasksetId:t.tasksetId,taskId:t.task.id,attemptId:n,requestId:s,kind:"raw_model_response",payload:{model:t.model,seed:t.seed,attempt:t.attempt,output:{text:a},startedAt:r,completedAt:o},timestamp:e.timestamp});return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:n,tasksetId:t.tasksetId,taskId:t.task.id,split:t.task.split,attempt:t.attempt,seed:t.seed,modelRef:t.model,startedAt:r,completedAt:o,output:{text:a},runtimeEventRefs:[],artifactRefs:[c.id],privilegedOutcomeRef:t.task.privilegedContextRef,infrastructureError:null,costUsd:null,latencyMs:dr(r,o),userInterventions:0,metadata:{requestId:s,execution:"text_completion"}})}catch(a){if(t.signal?.aborted)throw t.signal.reason instanceof Error?t.signal.reason:new Error("The evaluation was cancelled.");let o=e.timestamp(),n=a instanceof Error?a.message:String(a),c=e.resultId??`attempt_${h([s,"failure"]).slice(0,24)}`,u=await Q({store:e.store,storeDir:e.storeDir,tasksetId:t.tasksetId,taskId:t.task.id,attemptId:c,requestId:s,kind:"raw_model_response",payload:{model:t.model,seed:t.seed,attempt:t.attempt,error:n,startedAt:r,completedAt:o},timestamp:e.timestamp});return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:c,tasksetId:t.tasksetId,taskId:t.task.id,split:t.task.split,attempt:t.attempt,seed:t.seed,modelRef:t.model,startedAt:r,completedAt:o,output:{},runtimeEventRefs:[],artifactRefs:[u.id],privilegedOutcomeRef:null,infrastructureError:n,costUsd:null,latencyMs:dr(r,o),userInterventions:0,metadata:{requestId:s,execution:"text_completion"}})}}i(qn,"runTextAttempt");function $n(e){let t=Array.isArray(e.input.messages)?e.input.messages.flatMap(r=>{if(!r||typeof r!="object"||Array.isArray(r))return[];let s=r;return s.role!=="system"&&s.role!=="user"||typeof s.content!="string"||!s.content.trim()?[]:[{role:s.role,content:s.content}]}):[];if(t.length)return t;if(typeof e.input.prompt=="string"&&e.input.prompt.trim())return[{role:"user",content:e.input.prompt}];throw new Error(`Evaluation task ${e.id} has no policy-visible prompt.`)}i($n,"policyMessages");function Oo(e,t){return`post-training-evaluation:${h([e.task.id,e.model,e.seed,e.attempt,t]).slice(0,40)}`}i(Oo,"evaluationRequestId");function dr(e,t){return Math.max(0,Date.parse(t)-Date.parse(e))}i(dr,"elapsedMilliseconds");async function Ho(e){let t=Kr({context:e.context,attempt:e.attempt,grade:e.grade,artifacts:e.artifacts}),r=Ar.safeParse(e.attempt.metadata.optimizerTrainingSample??e.attempt.metadata.trainingSample),s=Xr({context:e.context,attempt:e.attempt,artifacts:e.artifacts,canonical:t,optimizerSample:r.success?r.data:null}),a=await Promise.all([Q({store:e.store,storeDir:e.storeDir,tasksetId:e.taskset.id,taskId:e.task.id,attemptId:e.attempt.id,requestId:e.attempt.id,kind:"grader_evidence",fileLabel:"artifact-manifest",payload:t.artifactManifest,timestamp:i(()=>t.artifactManifest.createdAt,"timestamp")}),Q({store:e.store,storeDir:e.storeDir,tasksetId:e.taskset.id,taskId:e.task.id,attemptId:e.attempt.id,requestId:e.attempt.id,kind:"grader_evidence",fileLabel:"reward-receipt",payload:t.rewardReceipt,timestamp:i(()=>t.rewardReceipt.createdAt,"timestamp")}),Q({store:e.store,storeDir:e.storeDir,tasksetId:e.taskset.id,taskId:e.task.id,attemptId:e.attempt.id,requestId:e.attempt.id,kind:"grader_evidence",fileLabel:"canonical-rollout",payload:s,timestamp:i(()=>s.completedAt,"timestamp")})]),o=[...e.artifacts,...a],n=F.parse({...e.attempt,artifactRefs:[...new Set([...e.attempt.artifactRefs,...a.map(c=>c.id)])],metadata:{...e.attempt.metadata,portableRunManifestRef:{id:e.context.runManifest.id,contentHash:e.context.runManifest.contentHash},portableAttemptReceipt:t.attemptReceipt,portableArtifactManifest:t.artifactManifest,portableRewardReceipt:t.rewardReceipt,portableCanonicalRollout:s,portableEnvironmentReleaseRef:{id:e.context.environmentRelease.id,contentHash:e.context.environmentRelease.contentHash},portableVerifierSetReleaseRef:{id:e.context.verifierSetRelease.id,contentHash:e.context.verifierSetRelease.contentHash},canonicalEvidenceArtifactRefs:a.map(c=>c.id)}});return await e.store.saveTaskAttempt(n),{attempt:n,artifacts:o,attemptReceipt:t.attemptReceipt,artifactManifest:t.artifactManifest,rewardReceipt:t.rewardReceipt,rolloutRecord:s}}i(Ho,"persistCanonicalEvaluationEvidence");function Do(e){if(typeof e.costUsd=="number"&&Number.isFinite(e.costUsd)&&e.costUsd>=0)return e.costUsd;if(!e.pricing)return null;let r=(Array.isArray(e.usage)?e.usage:e.usage===void 0?[]:[e.usage]).map(s=>Se(s,e.pricing));return!r.length||!r.every(s=>s!==null)?null:r.reduce((s,a)=>s+a,0)}i(Do,"hostedModelJudgeCallCost");function jo(e,t){return!!(e&&typeof e=="object"&&!Array.isArray(e)&&e.id===t.id&&e.contentHash===t.contentHash)}i(jo,"reviewRefMatches");function No(e){if(e.length<2)return 0;let t=e.reduce((r,s)=>r+s,0)/e.length;return e.reduce((r,s)=>r+(s-t)**2,0)/e.length}i(No,"variance");async function qo(e){let t=Hr.parse(e.fixture.metadata.toolScript),r=0;return Ht({store:e.store,storeDir:e.storeDir,taskset:e.taskset,policySource:"fixture",timestamp:i(()=>new Date().toISOString(),"timestamp"),resultId:`tool_fixture_${h({taskset:e.taskset.contentHash,fixture:e.fixture,time:Date.now()}).slice(0,32)}`,attemptInput:{tasksetId:e.taskset.id,task:e.task,model:{providerId:"openpond",modelId:"scripted-tool-policy"},seed:0,attempt:0},stream:i(async function*(){let s=t.actions[r++];s?yield{toolCalls:[{id:`fixture-call-${r}`,type:"function",function:{name:s.name,arguments:JSON.stringify(s.arguments)}}]}:yield{text:typeof e.fixture.output.text=="string"?e.fixture.output.text:JSON.stringify(e.fixture.output)}},"stream")})}i(qo,"runStarterToolFixture");import{constants as cr}from"node:fs";import{lstat as Ln,open as Un,realpath as $o}from"node:fs/promises";import lr from"node:path";async function Dt(e,t,r){r?.throwIfAborted();let s=e.metrics?.customAggregator;if(!s)return;if(!t)throw new Error("Taskset metric execution requires its package storage directory.");let a=lr.join(t,"training","tasksets",ce(e)),o=await Ln(a);if(!o.isDirectory()||o.isSymbolicLink())throw new Error("Taskset metric package must be a regular directory.");let n=await $o(a),c=lr.resolve(n,s.module);if(!c.startsWith(`${n}${lr.sep}`)||await $o(c)!==c)throw new Error("Taskset metric module must be a regular file inside its pinned package.");let u=await Un(c,cr.O_RDONLY|cr.O_NOFOLLOW|cr.O_NONBLOCK);try{let g=await u.stat();if(!g.isFile()||g.size>524288)throw new Error("Taskset metric module must be a bounded regular file.");let p=Buffer.alloc(g.size),l=0;for(;l<p.length;){r?.throwIfAborted();let d=await u.read(p,l,p.length-l,l);if(!d.bytesRead)throw new Error("Taskset metric module changed during capture.");l+=d.bytesRead}if((await u.read(Buffer.alloc(1),0,1,l)).bytesRead||fe(p)!==s.contentHash)throw new Error("Taskset metric module differs from its pinned content hash.");return r?.throwIfAborted(),new TextDecoder("utf-8",{fatal:!0}).decode(p)}finally{await u.close()}}i(Dt,"captureTasksetMetricSource");function Lo(e,t,r){if(h(e.metrics??null)!==h(t.metrics??null))throw new Error("Local Taskset metric policy differs from the admitted portable release.");Sr(t,r)}i(Lo,"assertLocalTasksetMetric");async function Uo(e){let{run:t}=e,r=await e.store.listBenchmarkRuns(e.tasksetId),s=i(c=>t.phase===c?t:r.find(u=>u.phase===c&&u.tasksetRelease.contentHash===t.tasksetRelease.contentHash&&u.model.provider===t.model.provider&&u.model.model===t.model.model&&u.reasoningEffort===t.reasoningEffort&&h(u.protocol)===h(t.protocol)),"counterpart"),a=s("baseline"),o=s("candidate");if(!a||!o)return null;let n=Cr({id:`benchmark-comparison-${h([a.contentHash,o.contentHash]).slice(0,24)}`,baseline:a,candidate:o,primaryMetric:e.benchmark.primaryMetric,qualityGate:e.benchmark.qualityGate,createdAt:e.createdAt,metadata:{sourceTasksetId:e.tasksetId,benchmarkDefinitionId:e.benchmark.definitionId}});return await e.store.saveBenchmarkComparison({tasksetId:e.tasksetId,comparison:n}),n}i(Uo,"persistBenchmarkComparison");var Wn=O.number().finite().min(0).max(1).nullable().optional();function jt(e){return Wn.parse(e)}i(jt,"parseExpectedFixtureScore");function ur(e,t){return e===void 0||e===t||typeof e=="number"&&t!==null&&Math.abs(e-t)<=1e-12}i(ur,"fixtureScoreMatches");function ap(e){let t=new Map;async function r(d,m){if(!e.storeDir||!e.modelText||!e.modelStream)throw new Error("Taskset execution is not configured.");let f=await Ee(e.store,d.tasksetId,d.tasksetRef),T={id:f.id,revision:f.revision,contentHash:f.contentHash},y=await l(f,d.taskId),w=f.environment.entrypoint===Y;if(w&&d.releasedHarness)throw new Error("This tool starter executes its Taskset-owned policy; a selected Harness requires a separate execution adapter.");let _=w?xs(await Ue(e.store,f,y,e.storeDir)):d.releasedHarness??await e.resolveReleasedHarness?.()??null,A=await e.resolveTasksetRelease?.(f)??null,I=!_&&e.loadProfileState?await e.loadProfileState():null,S=Yr({taskset:f,rewardExecution:await be(e.store,f),selectedTask:y,profile:I,releasedHarness:_,tasksetRelease:A,reasoningEffort:d.reasoningEffort??null,model:d.model,now:d.admittedAt?()=>d.admittedAt:void 0}),M=m??await Dt(f,e.storeDir,d.signal);Lo(f,S.tasksetRelease,M);let k=await Po({store:e.store,storeDir:e.storeDir,modelText:e.modelText,crossSystemStream:e.modelStream,work:e.workRuntime?{stream:e.modelStream,runtime:e.workRuntime,validateRequiredOutput:e.validateWorkRequiredOutput,additionalToolDefinitions:e.additionalWorkToolDefinitions?.(),toolEvidence:d.toolEvidence,harnessCapabilityReceipt:_?{harnessRelease:{id:_.harnessRelease.id,contentHash:_.harnessRelease.contentHash},agentSnapshot:{id:_.agentSnapshot.id,contentHash:_.agentSnapshot.contentHash},instructionAssets:_.agentSnapshot.instructions,skillAssets:_.agentSnapshot.skills,agentAssets:_.agentSnapshot.agents,fileAssets:_.harnessRelease.files,declaredToolNames:_.agentSnapshot.toolDeclarations.map($=>$.name)}:void 0,hostedTokenPricing:d.hostedTokenPricing}:void 0,resultId:d.resultId,hostedTokenPricing:d.hostedTokenPricing,parentModelRunId:d.parentModelRunId,harnessInstructionContext:_?.instructionContext,attemptInput:{tasksetId:f.id,tasksetRef:T,task:y,model:d.model,reasoningEffort:d.reasoningEffort,seed:d.seed,attempt:d.attempt,sampling:d.sampling,signal:d.signal}}),x=await s({tasksetId:f.id,tasksetRef:T,taskId:y.id,attempt:k,hostedTokenPricing:d.hostedTokenPricing}),v=await e.store.listTaskAttemptArtifacts({attemptId:k.id}),E=await Ho({store:e.store,storeDir:e.storeDir,taskset:f,task:y,context:S,attempt:k,grade:x,artifacts:v}),b=E.attemptReceipt,D=await it({id:`evaluation-${b.id}`,manifest:S.runManifest,receipts:[b],taskset:S.tasksetRelease,source:M,signal:d.signal,metadata:{sourceTasksetId:f.id,sourceTasksetRevision:f.revision,sourceTasksetHash:f.contentHash,sourceAttemptId:k.id,sourceGradeId:x.id}});return{attempt:E.attempt,grade:x,artifacts:E.artifacts,portable:{agentSnapshot:S.agentSnapshot,harnessRelease:S.harnessRelease,tasksetRelease:S.tasksetRelease,environmentRelease:S.environmentRelease,verifierSetRelease:S.verifierSetRelease,runManifest:S.runManifest,receipt:b,artifactManifest:E.artifactManifest,rewardReceipt:E.rewardReceipt,rolloutRecord:E.rolloutRecord,evaluationResult:D}}}i(r,"execute");async function s(d){let m=await Ee(e.store,d.tasksetId,d.tasksetRef),f=F.parse(d.attempt),T=await l(m,d.taskId,f.split),y=await zt(e,m),w=[],_=0,A=0,I=0,S=await Bs({task:T,attempt:f,graders:m.graders,learning:await be(e.store,m),modelJudge:e.modelJudge?async k=>{A+=1;let x=await e.modelJudge(k);Array.isArray(x.usage)?w.push(...x.usage):x.usage!==void 0&&w.push(x.usage);let v=Do({costUsd:x.costUsd,usage:x.usage,pricing:d.hostedTokenPricing});return v!==null&&(_+=v,I+=1),x}:void 0,customVerifier:y,now:i(()=>new Date().toISOString(),"now")});await e.store.saveTaskAttempt(f),await e.store.saveGradeResult(S);let M=w.reduce((k,x)=>{let v=ye(x);return k.inputTokens+=v.promptTokens??0,k.outputTokens+=v.completionTokens??0,k.totalTokens+=v.totalTokens??0,k},{inputTokens:0,outputTokens:0,totalTokens:0});return t.set(f.id,{...M,costUsd:A===0?0:I===A?_:null}),S}i(s,"grade");function a(d){let m={inputTokens:0,outputTokens:0,totalTokens:0,costUsd:null};for(let f of d){let T=t.get(f);t.delete(f),T&&(m.inputTokens+=T.inputTokens,m.outputTokens+=T.outputTokens,m.totalTokens+=T.totalTokens,T.costUsd!==null&&(m.costUsd=(m.costUsd??0)+T.costUsd))}return m}i(a,"consumeGraderUsage");async function o(d){let m=await Ee(e.store,d.tasksetId,d.tasksetRef),f=await zt(e,m),T=d.fixtures?.length?d.fixtures.map((k,x)=>({...k,expectedScore:jt(k.expectedScore),id:`external_fixture_${x}`})):m.graderFixtures.map((k,x)=>({id:k.id,label:k.label,taskId:k.taskId,expectedPassed:k.expectedPassed,expectedRewardEligible:k.expectedRewardEligible,expectedScore:jt(k.metadata.expectedScore),attempt:Vt(m.id,k,x)})),y=[];for(let k of T){let x=await l(m,k.taskId,Tt(k)),v=d.fixtures?.length?void 0:m.graderFixtures.find(D=>D.id===k.id);if(m.environment.entrypoint===Y&&v&&!v.infrastructureError&&!e.storeDir)throw new Error("Tool fixture execution requires the owner's artifact directory.");let E=m.environment.entrypoint===Y&&v&&!v.infrastructureError?await qo({store:e.store,storeDir:e.storeDir,taskset:m,task:x,fixture:v}):F.parse(k.attempt),b=await at({task:x,attempt:E,graders:m.graders,learning:await be(e.store,m),modelJudge:e.modelJudge??void 0,customVerifier:f});await e.store.saveTaskAttempt(E),await e.store.saveGradeResult(b),y.push({id:k.id,label:k.label,expectedPassed:k.expectedPassed,expectedRewardEligible:k.expectedRewardEligible,expectedScore:k.expectedScore,result:b})}let w=y.filter(({label:k,expectedPassed:x,expectedRewardEligible:v,expectedScore:E,result:b})=>!ur(E,b.score)||k==="infrastructure_failure"&&b.score!==null||typeof x=="boolean"&&b.passed!==x?!0:typeof v=="boolean"&&b.rewardEligible!==v),_=y.filter(k=>k.label==="infrastructure_failure").every(k=>k.result.score===null&&k.result.rewardEligible===!1),A=!w.some(k=>k.label==="adversarial"||k.label==="prompt_injection")&&y.filter(k=>k.label==="adversarial"||k.label==="prompt_injection").every(k=>!k.result.passed&&(k.result.score??0)<.8),I=y.every(k=>!k.result.feedback.some(x=>/privileged.*leak|hidden grader.*leak/i.test(x))),S=w.map(k=>({fixtureId:k.id,label:k.label,gradeId:k.result.id,reason:(k.expectedScore===void 0?"":`Expected score=${String(k.expectedScore)}; received score=${String(k.result.score)}. `)+`Expected passed=${String(k.expectedPassed)} and rewardEligible=${String(k.expectedRewardEligible)}; received passed=${String(k.result.passed)} and rewardEligible=${String(k.result.rewardEligible)}.`})),M=Jr.parse({schemaVersion:"openpond.graderAuditReport.v1",id:`grader_audit_${h([m.contentHash,y.map(k=>k.result.id)]).slice(0,24)}`,tasksetId:m.id,tasksetHash:m.contentHash,fixtureRefs:y.map(k=>k.id),gradeRefs:y.map(k=>k.result.id),passed:S.length===0&&_&&A&&I,hackingChecksPassed:A,leakageChecksPassed:I,infrastructureSafetyPassed:_,failures:S,createdAt:new Date().toISOString()});return await e.store.saveGraderAuditReport(M),{report:M,passed:M.passed,results:y,failures:S.map(k=>({label:k.label,gradeId:k.gradeId}))}}i(o,"auditFixtures");async function n(d){if(!e.modelJudge)throw new Error("No model judge runner is configured.");let m=await p(d),f=m.graders.filter(S=>S.kind==="model_judge");if(!f.length)throw new Error("Taskset has no model judges to calibrate.");let T=new Map;for(let S of m.graderFixtures)T.set(S.id,(T.get(S.id)??0)+1);for(let S of f)for(let M of S.calibrationFixtureRefs)if(T.get(M)!==1)throw new Error(`Judge ${S.id} calibration requires exactly one fixture for ${M}.`);let y=[],w=[];for(let S of m.graders){if(S.kind!=="model_judge"){w.push(S);continue}let M=m.graderFixtures.filter(E=>S.calibrationFixtureRefs.includes(E.id)),k=[];for(let[E,b]of M.entries()){let D=jt(b.metadata.expectedScore),$=await l(m,b.taskId,Tt(b)),z=Vt(m.id,b,E);try{let L=await e.modelJudge({grader:S,task:$,attempt:z});k.push({fixtureId:b.id,expectedPassed:b.expectedPassed,expectedScore:D,passed:L.passed,score:L.score,feedback:L.feedback,matched:L.passed===b.expectedPassed&&ur(D,L.score)})}catch(L){k.push({fixtureId:b.id,expectedPassed:b.expectedPassed,expectedScore:D,passed:!1,score:0,feedback:L instanceof Error?L.message:String(L),matched:!1})}}let x=k.length>0&&k.every(E=>E.matched),v=h({graderId:S.id,graderVersion:S.version,judge:S.judge,rubric:S.rubric,temperature:S.temperature,results:k});w.push({...S,calibrationStatus:x?"passed":"failed",rewardEligible:x&&S.metadata.requestedRewardEligible===!0,metadata:{...S.metadata,calibrationEvidenceHash:v,calibrationAccuracy:k.length?k.filter(E=>E.matched).length/k.length:0,calibratedAt:new Date().toISOString()}}),y.push({graderId:S.id,passed:x,evidenceHash:v,results:k})}let _=new Date().toISOString(),A=ke.parse({...m,revision:m.revision+1,graders:w,status:"needs_review",readiness:null,contentHash:"00000000",updatedAt:_,metadata:{...m.metadata,judgeCalibration:{parentTasksetHash:m.contentHash,calibratedAt:_,graderIds:y.map(S=>S.graderId)}}}),I=ke.parse({...A,contentHash:Te(A)});if(!e.storeDir)throw new Error("Managed Taskset storage is required for judge calibration.");if(I.purpose!=="benchmark"){let S=Wo.join(e.storeDir,"training","tasksets"),M=Wo.join(S,ce(m));await pt(M,m),I=await Yt(S,I,M)}return await e.store.upsertTaskset(I),{taskset:I,results:y,passed:y.every(S=>S.passed)}}i(n,"calibrateModelJudges");async function c(d,m){let f=await Ee(e.store,d,m),y=(await e.store.listGraderAuditReports(d)).find(_=>_.tasksetHash===f.contentHash)??null;!y&&!f.graders.some(_=>_.kind==="model_judge")&&(y=(await o({tasksetId:d,tasksetRef:m})).report);let w=Ns({taskset:f,graderAudit:y});return await e.store.saveReadinessReport(w),await e.store.upsertTaskset({...f,status:w.ready?"ready":"needs_review",readiness:w,updatedAt:new Date().toISOString()}),w}i(c,"readiness");async function u(d){let m=await p(d.tasksetId);if(m.status!=="ready"||m.readiness?.ready!==!0)throw new Error("Taskset must pass grader audit and readiness before baseline Evaluation.");let f=m.metadata.harnessEvaluationReview;if(!f||typeof f!="object"||Array.isArray(f)||f.id!==d.reviewRef.id||f.contentHash!==d.reviewRef.contentHash)throw new Error("Taskset does not bind the supplied Harness Evaluation review.");let T=(await e.store.listEvaluationResults(m.id,"baseline")).find(b=>b.metadata.sourceTasksetHash===m.contentHash&&b.model.provider===d.model.providerId&&b.model.model===d.model.modelId&&(!m.metrics||b.authoredMetric?.policyHash===h(m.metrics))&&jo(b.metadata.harnessEvaluationReview,d.reviewRef));if(T)return{evaluationResult:T,attempts:[],reused:!0};let y=m.tasks.filter(b=>b.split==="frozen_eval");if(!y.length)throw new Error("Baseline Evaluation requires at least one isolated frozen-evaluation task.");let w=[...new Set(d.seeds?.length?d.seeds:[17])].slice(0,100),_=Math.max(1,Math.min(20,Math.trunc(d.attemptsPerTask??1))),A=new Date().toISOString(),I=await Dt(m,e.storeDir,d.signal),S=[];for(let b of y)for(let D of w)for(let $=0;$<_;$+=1){if(d.signal?.aborted)throw d.signal.reason;let z=await r({tasksetId:m.id,tasksetRef:{id:m.id,revision:m.revision,contentHash:m.contentHash},taskId:b.id,model:d.model,seed:D,attempt:$,sampling:d.sampling,signal:d.signal,admittedAt:A,resultId:`baseline-attempt-${h({reviewRef:d.reviewRef,tasksetHash:m.contentHash,model:d.model,taskId:b.id,seed:D,attempt:$}).slice(0,24)}`},I);S.push(z)}let M=S[0].portable.runManifest;if(S.some(b=>b.portable.runManifest.contentHash!==M.contentHash))throw new Error("Baseline attempts did not share one frozen Run Manifest.");let k=S.map(b=>b.portable.receipt),x=S[0].portable.tasksetRelease,v=k.flatMap(b=>typeof b.metadata.score=="number"?[b.metadata.score]:[]),E=await it({id:`baseline-evaluation-${h({manifest:M.contentHash,receipts:k.map(b=>b.contentHash),reviewRef:d.reviewRef}).slice(0,24)}`,manifest:M,receipts:k,taskset:x,source:I,signal:d.signal,metadata:{kind:"baseline",sourceTasksetId:m.id,sourceTasksetRevision:m.revision,sourceTasksetHash:m.contentHash,harnessEvaluationReview:d.reviewRef,environmentHash:h(x.environment),toolContractHash:h(x.tools),permissionContractHash:h({connectedAppScopes:x.policy.connectedAppScopes,networkPolicy:x.environment.networkPolicy}),policyHash:h(x.policy),verifierRef:{id:`verifier-${m.id}-r${m.revision}`,contentHash:h(x.graders)},scores:v,scoreVariance:No(v),admittedAt:A}});return await e.store.saveEvaluationResult({tasksetId:m.id,kind:"baseline",result:E,createdAt:A}),await Ds({store:e.store,taskset:m,evaluationResult:E}),{evaluationResult:E,attempts:S,reused:!1}}i(u,"executeBaseline");async function g(d){let m=await p(d.tasksetId);if(m.purpose!=="benchmark"||!m.benchmark)throw new Error("This Taskset does not define a benchmark protocol.");let f=d.split??m.benchmark.evaluationSplit,T=d.taskIds?.length?new Set(d.taskIds):null,y=m.tasks.filter(b=>b.split===f&&(!T||T.has(b.id)));if(T&&y.length!==T.size)throw new Error("Benchmark task selection contains an id outside the pinned split.");if(!y.length)throw new Error(`Benchmark split ${f} has no cases.`);let w=[...new Set(d.seeds?.length?d.seeds:[17])].slice(0,100),_=Math.max(1,Math.min(20,Math.trunc(d.repetitions??1))),A=new Date().toISOString(),I=[],S=await Dt(m,e.storeDir,d.signal);for(let b of y)for(let D of w)for(let $=0;$<_;$+=1){if(d.signal?.aborted)throw d.signal.reason;let z=await r({tasksetId:m.id,tasksetRef:{id:m.id,revision:m.revision,contentHash:m.contentHash},taskId:b.id,model:d.model,reasoningEffort:d.reasoningEffort??null,seed:D,attempt:$,sampling:d.sampling,signal:d.signal,releasedHarness:d.releasedHarness,admittedAt:A,parentModelRunId:d.parentModelRunId,toolEvidence:d.toolEvidence,hostedTokenPricing:d.hostedTokenPricing,resultId:`benchmark-attempt-${h({tasksetRelease:m.benchmark.releaseHash,phase:d.phase,model:d.model,reasoningEffort:d.reasoningEffort??null,split:f,taskId:b.id,seed:D,attempt:$,admittedAt:A}).slice(0,24)}`},S);I.push(z),await d.onAttemptComplete?.(z)}let M=I[0].portable.runManifest;if(I.some(b=>b.portable.runManifest.contentHash!==M.contentHash))throw new Error("Benchmark attempts did not share one pinned Run Manifest.");let k=I.map(b=>b.portable.receipt),x=await it({id:`benchmark-evaluation-${h({manifest:M.contentHash,phase:d.phase,receipts:k.map(b=>b.contentHash)}).slice(0,24)}`,manifest:M,receipts:k,taskset:I[0].portable.tasksetRelease,source:S,signal:d.signal,metadata:{kind:"benchmark",phase:d.phase,split:f,reasoningEffort:d.reasoningEffort??null,sourceTasksetId:m.id,sourceTasksetRevision:m.revision,sourceTasksetHash:m.contentHash,benchmarkDefinitionId:m.benchmark.definitionId,admittedAt:A,parentModelRunId:d.parentModelRunId??null}});await e.store.saveEvaluationResult({tasksetId:m.id,kind:d.phase,result:x,createdAt:A});let v=Er({id:`benchmark-run-${x.contentHash.slice(0,24)}`,phase:d.phase,evaluation:x,receipts:k,reasoningEffort:d.reasoningEffort??null,protocol:{split:f,taskIds:y.map(b=>b.id),seeds:w.map(String),repetitions:_,runtimeTargetHash:h(M.runtimeTarget),environmentHash:h(I[0].portable.tasksetRelease.environment),toolContractHash:h({tasksetTools:I[0].portable.tasksetRelease.tools,harnessTools:I[0].portable.harnessRelease.tools}),limitsHash:h(M.limits)},createdAt:A,metadata:{sourceTasksetId:m.id,split:f,seeds:w,repetitions:_,parentModelRunId:d.parentModelRunId??null}});await e.store.saveBenchmarkRun({tasksetId:m.id,run:v});let E=await Uo({store:e.store,tasksetId:m.id,benchmark:m.benchmark,run:v,createdAt:A});return{evaluationResult:x,run:v,comparison:E,attempts:I}}i(g,"executeBenchmark");async function p(d){let m=await e.store.getTaskset(d);if(!m)throw new Error("Taskset not found.");return m}i(p,"requireTaskset");async function l(d,m,f){let T=d.tasks.find(y=>y.id===m);if(T)return T;if(!d.datasetArtifact||!e.resolveTask)throw new Error(`Task ${m} not found.`);return e.resolveTask({tasksetId:d.id,taskId:m,split:f})}return i(l,"findTask"),{execute:r,grade:s,consumeGraderUsage:a,auditFixtures:o,calibrateModelJudges:n,readiness:c,executeBaseline:u,executeBenchmark:g,close:i(async()=>{},"close")}}i(ap,"createTaskEvaluationService");export{yt as a,Fn as b,Se as c,Rs as d,oi as e,be as f,Ko as g,Ji as h,ed as i,Ns as j,wc as k,Gs as l,It as m,Et as n,gl as o,Rl as p,Ut as q,fd as r,Hs as s,Rd as t,Kl as u,Xl as v,Ho as w,Ed as x,Cd as y,Ds as z,ap as A};
+${JSON.stringify({policyVisibleContext:t.task.policyVisibleContext})}`,seed:t.seed,timeoutMs:a.execution.environment.contract.defaultTimeoutMs,signal:t.signal,execute:Ts,policy:u.policy}),p=e.timestamp(),l=["environment_failure","timed_out","cancelled"].includes(g.status)?g.error??g.status:null,d={schemaVersion:"openpond.starterToolAttempt.v1",taskset:U(r),taskHash:h(t.task),model:s,seed:t.seed,attempt:t.attempt,policySource:e.policySource??"model",definition:a.definition,environment:U(a.execution.environment),verifierSet:U(a.execution.verifierSet),result:g};ze(d,12582912);let m=await Q({store:e.store,storeDir:e.storeDir,tasksetId:r.id,taskId:t.task.id,attemptId:c,requestId:n,kind:"environment_state",payload:d,timestamp:e.timestamp});return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:c,tasksetId:r.id,taskId:t.task.id,split:t.task.split,attempt:t.attempt,seed:t.seed,modelRef:s,startedAt:o,completedAt:p,output:{text:g.output??""},runtimeEventRefs:[],artifactRefs:[m.id],privilegedOutcomeRef:m.id,infrastructureError:l,costUsd:u.costUsd(),latencyMs:Math.max(0,Date.parse(p)-Date.parse(o)),userInterventions:0,metadata:{requestId:n,execution:"starter_tool_environment",policySource:e.policySource??"model",environmentStatus:g.status,environmentCleanupComplete:g.environmentCleanupComplete,environmentAttemptHash:g.contentHash}})}i(Ht,"runStarterToolAttempt");async function Po(e){let t=e.timestamp??(()=>new Date().toISOString()),r=await Ee(e.store,e.attemptInput.tasksetId,e.attemptInput.tasksetRef);if(r.environment.entrypoint===Y)return Ht({store:e.store,storeDir:e.storeDir,taskset:r,attemptInput:e.attemptInput,stream:e.crossSystemStream,timestamp:t,resultId:e.resultId,hostedTokenPricing:e.hostedTokenPricing});if(r.environment.kind==="work"){if(!e.work)throw new Error("Taskset Work execution is not configured for this evaluator.");return Co({store:e.store,storeDir:e.storeDir,taskset:r,task:e.attemptInput.task,model:e.attemptInput.model,reasoningEffort:e.attemptInput.reasoningEffort,seed:e.attemptInput.seed,attempt:e.attemptInput.attempt,sampling:e.attemptInput.sampling,signal:e.attemptInput.signal,stream:e.work.stream,runtime:e.work.runtime,timestamp:e.timestamp,resultId:e.resultId,parentModelRunId:e.parentModelRunId,harnessInstructionContext:e.harnessInstructionContext,validateRequiredOutput:e.work.validateRequiredOutput,additionalToolDefinitions:e.work.additionalToolDefinitions,toolEvidence:e.work.toolEvidence,harnessCapabilityReceipt:e.work.harnessCapabilityReceipt,hostedTokenPricing:e.work.hostedTokenPricing})}if(!Mo(r)&&(r.environment.kind!=="chat"||r.environment.stateful||r.environment.toolNames.length||r.capabilities.requiresTools||r.capabilities.requiresState))throw new Error("The Taskset environment has no registered evaluation adapter.");return Mo(r)?Nn({...e,timestamp:t,taskset:r}):qn({...e,timestamp:t})}i(Po,"runPostTrainingEvaluationAttempt");function Mo(e){let t=e.metadata.flagship==="cross-system-operations",r=e.environment.metadata.flagship==="cross-system-operations";return(t||r)&&e.environment.stateful&&e.environment.metadata.toolContractHash===q}i(Mo,"isCrossSystemTaskset");async function Nn(e){let{attemptInput:t,taskset:r}=e,s=e.timestamp(),a=Oo(t,s),o=uo(r,{taskId:t.task.id,prompt:typeof t.task.input.prompt=="string"?t.task.input.prompt:null}),n=new AbortController,c=i(()=>n.abort(t.signal?.reason??new Error("The evaluation was cancelled.")),"abortFromParent");t.signal?.addEventListener("abort",c,{once:!0});let u=Math.max(1,Math.min(r.environment.defaultTimeoutMs,10*6e4)),g=setTimeout(()=>n.abort(new Error(`Cross-System evaluation exceeded ${u} ms.`)),u);g.unref?.();try{let p=[],l=null,d=null;for(let w=0;w<=1;w+=1){let _=await io({world:o.world,task:o.generatedTask,model:t.model,reasoningEffort:null,stream:e.crossSystemStream,signal:n.signal,trajectoryId:`cso_post_training_eval_${h([a,w]).slice(0,24)}`,metadata:{evaluation:"post_training",execution:"post_training_evaluation_tool_loop",tasksetId:r.id,tasksetHash:r.contentHash,samplingSeed:t.seed,attempt:t.attempt,infrastructureRetryAttempt:w,priorInfrastructureErrors:p.map(A=>A.infrastructureError??"Unknown infrastructure failure.")}});if(l=_,d=xt({task:o.generatedTask,trajectory:_}),d.outcome!=="infrastructure_failure"||w===1)break;p.push(_)}if(!l||!d)throw new Error(`Cross-System task ${t.task.id} produced no trajectory.`);let m=[...l.steps].reverse().find(w=>w.kind==="final"),f=l.completedAt,T=e.resultId??`attempt_${h([a,l.id,d.outcome]).slice(0,24)}`,y=await Q({store:e.store,storeDir:e.storeDir,tasksetId:r.id,taskId:t.task.id,attemptId:T,requestId:a,kind:"runtime_trace",payload:{schemaVersion:B,model:t.model,seed:t.seed,attempt:t.attempt,trajectory:l,verifier:d,priorInfrastructureTrajectories:p},timestamp:e.timestamp});return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:T,tasksetId:r.id,taskId:t.task.id,split:t.task.split,attempt:t.attempt,seed:t.seed,modelRef:t.model,startedAt:l.startedAt,completedAt:f,output:{text:m?.content??"",trajectoryId:l.id},runtimeEventRefs:[],artifactRefs:[y.id],privilegedOutcomeRef:t.task.privilegedContextRef,infrastructureError:l.infrastructureError,costUsd:null,latencyMs:dr(l.startedAt,f),userInterventions:0,metadata:{requestId:a,execution:"post_training_evaluation_tool_loop",trajectoryId:l.id,worldId:l.worldId,toolContractHash:l.toolContractHash,verifierOutcome:d.outcome,verifierReward:d.reward,verifierRewardEligible:d.rewardEligible,infrastructureRetryAttempt:l.metadata.infrastructureRetryAttempt,priorInfrastructureErrors:l.metadata.priorInfrastructureErrors}})}finally{clearTimeout(g),t.signal?.removeEventListener("abort",c)}}i(Nn,"runCrossSystemAttempt");async function qn(e){let{attemptInput:t}=e,r=e.timestamp(),s=Oo(t,r);try{let a=await e.modelText({model:t.model,reasoningEffort:null,signal:t.signal??new AbortController().signal,requestId:s,messages:$n(t.task),maxOutputTokens:t.sampling?.maxOutputTokens??2048,temperature:t.sampling?.temperature??.8,topP:t.sampling?.topP??.95,seed:t.seed+t.attempt}),o=e.timestamp(),n=e.resultId??`attempt_${h([s,a]).slice(0,24)}`,c=await Q({store:e.store,storeDir:e.storeDir,tasksetId:t.tasksetId,taskId:t.task.id,attemptId:n,requestId:s,kind:"raw_model_response",payload:{model:t.model,seed:t.seed,attempt:t.attempt,output:{text:a},startedAt:r,completedAt:o},timestamp:e.timestamp});return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:n,tasksetId:t.tasksetId,taskId:t.task.id,split:t.task.split,attempt:t.attempt,seed:t.seed,modelRef:t.model,startedAt:r,completedAt:o,output:{text:a},runtimeEventRefs:[],artifactRefs:[c.id],privilegedOutcomeRef:t.task.privilegedContextRef,infrastructureError:null,costUsd:null,latencyMs:dr(r,o),userInterventions:0,metadata:{requestId:s,execution:"text_completion"}})}catch(a){if(t.signal?.aborted)throw t.signal.reason instanceof Error?t.signal.reason:new Error("The evaluation was cancelled.");let o=e.timestamp(),n=a instanceof Error?a.message:String(a),c=e.resultId??`attempt_${h([s,"failure"]).slice(0,24)}`,u=await Q({store:e.store,storeDir:e.storeDir,tasksetId:t.tasksetId,taskId:t.task.id,attemptId:c,requestId:s,kind:"raw_model_response",payload:{model:t.model,seed:t.seed,attempt:t.attempt,error:n,startedAt:r,completedAt:o},timestamp:e.timestamp});return F.parse({schemaVersion:"openpond.taskAttempt.v1",id:c,tasksetId:t.tasksetId,taskId:t.task.id,split:t.task.split,attempt:t.attempt,seed:t.seed,modelRef:t.model,startedAt:r,completedAt:o,output:{},runtimeEventRefs:[],artifactRefs:[u.id],privilegedOutcomeRef:null,infrastructureError:n,costUsd:null,latencyMs:dr(r,o),userInterventions:0,metadata:{requestId:s,execution:"text_completion"}})}}i(qn,"runTextAttempt");function $n(e){let t=Array.isArray(e.input.messages)?e.input.messages.flatMap(r=>{if(!r||typeof r!="object"||Array.isArray(r))return[];let s=r;return s.role!=="system"&&s.role!=="user"||typeof s.content!="string"||!s.content.trim()?[]:[{role:s.role,content:s.content}]}):[];if(t.length)return t;if(typeof e.input.prompt=="string"&&e.input.prompt.trim())return[{role:"user",content:e.input.prompt}];throw new Error(`Evaluation task ${e.id} has no policy-visible prompt.`)}i($n,"policyMessages");function Oo(e,t){return`post-training-evaluation:${h([e.task.id,e.model,e.seed,e.attempt,t]).slice(0,40)}`}i(Oo,"evaluationRequestId");function dr(e,t){return Math.max(0,Date.parse(t)-Date.parse(e))}i(dr,"elapsedMilliseconds");async function Ho(e){let t=Kr({context:e.context,attempt:e.attempt,grade:e.grade,artifacts:e.artifacts}),r=Mr.safeParse(e.attempt.metadata.optimizerTrainingSample??e.attempt.metadata.trainingSample),s=Xr({context:e.context,attempt:e.attempt,artifacts:e.artifacts,canonical:t,optimizerSample:r.success?r.data:null}),a=await Promise.all([Q({store:e.store,storeDir:e.storeDir,tasksetId:e.taskset.id,taskId:e.task.id,attemptId:e.attempt.id,requestId:e.attempt.id,kind:"grader_evidence",fileLabel:"artifact-manifest",payload:t.artifactManifest,timestamp:i(()=>t.artifactManifest.createdAt,"timestamp")}),Q({store:e.store,storeDir:e.storeDir,tasksetId:e.taskset.id,taskId:e.task.id,attemptId:e.attempt.id,requestId:e.attempt.id,kind:"grader_evidence",fileLabel:"reward-receipt",payload:t.rewardReceipt,timestamp:i(()=>t.rewardReceipt.createdAt,"timestamp")}),Q({store:e.store,storeDir:e.storeDir,tasksetId:e.taskset.id,taskId:e.task.id,attemptId:e.attempt.id,requestId:e.attempt.id,kind:"grader_evidence",fileLabel:"canonical-rollout",payload:s,timestamp:i(()=>s.completedAt,"timestamp")})]),o=[...e.artifacts,...a],n=F.parse({...e.attempt,artifactRefs:[...new Set([...e.attempt.artifactRefs,...a.map(c=>c.id)])],metadata:{...e.attempt.metadata,portableRunManifestRef:{id:e.context.runManifest.id,contentHash:e.context.runManifest.contentHash},portableAttemptReceipt:t.attemptReceipt,portableArtifactManifest:t.artifactManifest,portableRewardReceipt:t.rewardReceipt,portableCanonicalRollout:s,portableEnvironmentReleaseRef:{id:e.context.environmentRelease.id,contentHash:e.context.environmentRelease.contentHash},portableVerifierSetReleaseRef:{id:e.context.verifierSetRelease.id,contentHash:e.context.verifierSetRelease.contentHash},canonicalEvidenceArtifactRefs:a.map(c=>c.id)}});return await e.store.saveTaskAttempt(n),{attempt:n,artifacts:o,attemptReceipt:t.attemptReceipt,artifactManifest:t.artifactManifest,rewardReceipt:t.rewardReceipt,rolloutRecord:s}}i(Ho,"persistCanonicalEvaluationEvidence");function Do(e){if(typeof e.costUsd=="number"&&Number.isFinite(e.costUsd)&&e.costUsd>=0)return e.costUsd;if(!e.pricing)return null;let r=(Array.isArray(e.usage)?e.usage:e.usage===void 0?[]:[e.usage]).map(s=>Se(s,e.pricing));return!r.length||!r.every(s=>s!==null)?null:r.reduce((s,a)=>s+a,0)}i(Do,"hostedModelJudgeCallCost");function jo(e,t){return!!(e&&typeof e=="object"&&!Array.isArray(e)&&e.id===t.id&&e.contentHash===t.contentHash)}i(jo,"reviewRefMatches");function No(e){if(e.length<2)return 0;let t=e.reduce((r,s)=>r+s,0)/e.length;return e.reduce((r,s)=>r+(s-t)**2,0)/e.length}i(No,"variance");async function qo(e){let t=Hr.parse(e.fixture.metadata.toolScript),r=0;return Ht({store:e.store,storeDir:e.storeDir,taskset:e.taskset,policySource:"fixture",timestamp:i(()=>new Date().toISOString(),"timestamp"),resultId:`tool_fixture_${h({taskset:e.taskset.contentHash,fixture:e.fixture,time:Date.now()}).slice(0,32)}`,attemptInput:{tasksetId:e.taskset.id,task:e.task,model:{providerId:"openpond",modelId:"scripted-tool-policy"},seed:0,attempt:0},stream:i(async function*(){let s=t.actions[r++];s?yield{toolCalls:[{id:`fixture-call-${r}`,type:"function",function:{name:s.name,arguments:JSON.stringify(s.arguments)}}]}:yield{text:typeof e.fixture.output.text=="string"?e.fixture.output.text:JSON.stringify(e.fixture.output)}},"stream")})}i(qo,"runStarterToolFixture");import{constants as cr}from"node:fs";import{lstat as Ln,open as Un,realpath as $o}from"node:fs/promises";import lr from"node:path";async function Dt(e,t,r){r?.throwIfAborted();let s=e.metrics?.customAggregator;if(!s)return;if(!t)throw new Error("Taskset metric execution requires its package storage directory.");let a=lr.join(t,"training","tasksets",ce(e)),o=await Ln(a);if(!o.isDirectory()||o.isSymbolicLink())throw new Error("Taskset metric package must be a regular directory.");let n=await $o(a),c=lr.resolve(n,s.module);if(!c.startsWith(`${n}${lr.sep}`)||await $o(c)!==c)throw new Error("Taskset metric module must be a regular file inside its pinned package.");let u=await Un(c,cr.O_RDONLY|cr.O_NOFOLLOW|cr.O_NONBLOCK);try{let g=await u.stat();if(!g.isFile()||g.size>524288)throw new Error("Taskset metric module must be a bounded regular file.");let p=Buffer.alloc(g.size),l=0;for(;l<p.length;){r?.throwIfAborted();let d=await u.read(p,l,p.length-l,l);if(!d.bytesRead)throw new Error("Taskset metric module changed during capture.");l+=d.bytesRead}if((await u.read(Buffer.alloc(1),0,1,l)).bytesRead||fe(p)!==s.contentHash)throw new Error("Taskset metric module differs from its pinned content hash.");return r?.throwIfAborted(),new TextDecoder("utf-8",{fatal:!0}).decode(p)}finally{await u.close()}}i(Dt,"captureTasksetMetricSource");function Lo(e,t,r){if(h(e.metrics??null)!==h(t.metrics??null))throw new Error("Local Taskset metric policy differs from the admitted portable release.");br(t,r)}i(Lo,"assertLocalTasksetMetric");async function Uo(e){let{run:t}=e,r=await e.store.listBenchmarkRuns(e.tasksetId),s=i(c=>t.phase===c?t:r.find(u=>u.phase===c&&u.tasksetRelease.contentHash===t.tasksetRelease.contentHash&&u.model.provider===t.model.provider&&u.model.model===t.model.model&&u.reasoningEffort===t.reasoningEffort&&h(u.protocol)===h(t.protocol)),"counterpart"),a=s("baseline"),o=s("candidate");if(!a||!o)return null;let n=Ar({id:`benchmark-comparison-${h([a.contentHash,o.contentHash]).slice(0,24)}`,baseline:a,candidate:o,primaryMetric:e.benchmark.primaryMetric,qualityGate:e.benchmark.qualityGate,createdAt:e.createdAt,metadata:{sourceTasksetId:e.tasksetId,benchmarkDefinitionId:e.benchmark.definitionId}});return await e.store.saveBenchmarkComparison({tasksetId:e.tasksetId,comparison:n}),n}i(Uo,"persistBenchmarkComparison");var Wn=O.number().finite().min(0).max(1).nullable().optional();function jt(e){return Wn.parse(e)}i(jt,"parseExpectedFixtureScore");function ur(e,t){return e===void 0||e===t||typeof e=="number"&&t!==null&&Math.abs(e-t)<=1e-12}i(ur,"fixtureScoreMatches");function ap(e){let t=new Map;async function r(d,m){if(!e.storeDir||!e.modelText||!e.modelStream)throw new Error("Taskset execution is not configured.");let f=await Ee(e.store,d.tasksetId,d.tasksetRef),T={id:f.id,revision:f.revision,contentHash:f.contentHash},y=await l(f,d.taskId),w=f.environment.entrypoint===Y;if(w&&d.releasedHarness)throw new Error("This tool starter executes its Taskset-owned policy; a selected Harness requires a separate execution adapter.");let _=w?xs(await Ue(e.store,f,y,e.storeDir)):d.releasedHarness??await e.resolveReleasedHarness?.()??null,A=await e.resolveTasksetRelease?.(f)??null,I=!_&&e.loadProfileState?await e.loadProfileState():null,S=Yr({taskset:f,rewardExecution:await be(e.store,f),selectedTask:y,profile:I,releasedHarness:_,tasksetRelease:A,reasoningEffort:d.reasoningEffort??null,model:d.model,now:d.admittedAt?()=>d.admittedAt:void 0}),M=m??await Dt(f,e.storeDir,d.signal);Lo(f,S.tasksetRelease,M);let k=await Po({store:e.store,storeDir:e.storeDir,modelText:e.modelText,crossSystemStream:e.modelStream,work:e.workRuntime?{stream:e.modelStream,runtime:e.workRuntime,validateRequiredOutput:e.validateWorkRequiredOutput,additionalToolDefinitions:e.additionalWorkToolDefinitions?.(),toolEvidence:d.toolEvidence,harnessCapabilityReceipt:_?{harnessRelease:{id:_.harnessRelease.id,contentHash:_.harnessRelease.contentHash},agentSnapshot:{id:_.agentSnapshot.id,contentHash:_.agentSnapshot.contentHash},instructionAssets:_.agentSnapshot.instructions,skillAssets:_.agentSnapshot.skills,agentAssets:_.agentSnapshot.agents,fileAssets:_.harnessRelease.files,declaredToolNames:_.agentSnapshot.toolDeclarations.map($=>$.name)}:void 0,hostedTokenPricing:d.hostedTokenPricing}:void 0,resultId:d.resultId,hostedTokenPricing:d.hostedTokenPricing,parentModelRunId:d.parentModelRunId,harnessInstructionContext:_?.instructionContext,attemptInput:{tasksetId:f.id,tasksetRef:T,task:y,model:d.model,reasoningEffort:d.reasoningEffort,seed:d.seed,attempt:d.attempt,sampling:d.sampling,signal:d.signal}}),x=await s({tasksetId:f.id,tasksetRef:T,taskId:y.id,attempt:k,hostedTokenPricing:d.hostedTokenPricing}),v=await e.store.listTaskAttemptArtifacts({attemptId:k.id}),E=await Ho({store:e.store,storeDir:e.storeDir,taskset:f,task:y,context:S,attempt:k,grade:x,artifacts:v}),b=E.attemptReceipt,D=await it({id:`evaluation-${b.id}`,manifest:S.runManifest,receipts:[b],taskset:S.tasksetRelease,source:M,signal:d.signal,metadata:{sourceTasksetId:f.id,sourceTasksetRevision:f.revision,sourceTasksetHash:f.contentHash,sourceAttemptId:k.id,sourceGradeId:x.id}});return{attempt:E.attempt,grade:x,artifacts:E.artifacts,portable:{agentSnapshot:S.agentSnapshot,harnessRelease:S.harnessRelease,tasksetRelease:S.tasksetRelease,environmentRelease:S.environmentRelease,verifierSetRelease:S.verifierSetRelease,runManifest:S.runManifest,receipt:b,artifactManifest:E.artifactManifest,rewardReceipt:E.rewardReceipt,rolloutRecord:E.rolloutRecord,evaluationResult:D}}}i(r,"execute");async function s(d){let m=await Ee(e.store,d.tasksetId,d.tasksetRef),f=F.parse(d.attempt),T=await l(m,d.taskId,f.split),y=await zt(e,m),w=[],_=0,A=0,I=0,S=await Bs({task:T,attempt:f,graders:m.graders,learning:await be(e.store,m),modelJudge:e.modelJudge?async k=>{A+=1;let x=await e.modelJudge(k);Array.isArray(x.usage)?w.push(...x.usage):x.usage!==void 0&&w.push(x.usage);let v=Do({costUsd:x.costUsd,usage:x.usage,pricing:d.hostedTokenPricing});return v!==null&&(_+=v,I+=1),x}:void 0,customVerifier:y,now:i(()=>new Date().toISOString(),"now")});await e.store.saveTaskAttempt(f),await e.store.saveGradeResult(S);let M=w.reduce((k,x)=>{let v=ye(x);return k.inputTokens+=v.promptTokens??0,k.outputTokens+=v.completionTokens??0,k.totalTokens+=v.totalTokens??0,k},{inputTokens:0,outputTokens:0,totalTokens:0});return t.set(f.id,{...M,costUsd:A===0?0:I===A?_:null}),S}i(s,"grade");function a(d){let m={inputTokens:0,outputTokens:0,totalTokens:0,costUsd:null};for(let f of d){let T=t.get(f);t.delete(f),T&&(m.inputTokens+=T.inputTokens,m.outputTokens+=T.outputTokens,m.totalTokens+=T.totalTokens,T.costUsd!==null&&(m.costUsd=(m.costUsd??0)+T.costUsd))}return m}i(a,"consumeGraderUsage");async function o(d){let m=await Ee(e.store,d.tasksetId,d.tasksetRef),f=await zt(e,m),T=d.fixtures?.length?d.fixtures.map((k,x)=>({...k,expectedScore:jt(k.expectedScore),id:`external_fixture_${x}`})):m.graderFixtures.map((k,x)=>({id:k.id,label:k.label,taskId:k.taskId,expectedPassed:k.expectedPassed,expectedRewardEligible:k.expectedRewardEligible,expectedScore:jt(k.metadata.expectedScore),attempt:Vt(m.id,k,x)})),y=[];for(let k of T){let x=await l(m,k.taskId,Tt(k)),v=d.fixtures?.length?void 0:m.graderFixtures.find(D=>D.id===k.id);if(m.environment.entrypoint===Y&&v&&!v.infrastructureError&&!e.storeDir)throw new Error("Tool fixture execution requires the owner's artifact directory.");let E=m.environment.entrypoint===Y&&v&&!v.infrastructureError?await qo({store:e.store,storeDir:e.storeDir,taskset:m,task:x,fixture:v}):F.parse(k.attempt),b=await at({task:x,attempt:E,graders:m.graders,learning:await be(e.store,m),modelJudge:e.modelJudge??void 0,customVerifier:f});await e.store.saveTaskAttempt(E),await e.store.saveGradeResult(b),y.push({id:k.id,label:k.label,expectedPassed:k.expectedPassed,expectedRewardEligible:k.expectedRewardEligible,expectedScore:k.expectedScore,result:b})}let w=y.filter(({label:k,expectedPassed:x,expectedRewardEligible:v,expectedScore:E,result:b})=>!ur(E,b.score)||k==="infrastructure_failure"&&b.score!==null||typeof x=="boolean"&&b.passed!==x?!0:typeof v=="boolean"&&b.rewardEligible!==v),_=y.filter(k=>k.label==="infrastructure_failure").every(k=>k.result.score===null&&k.result.rewardEligible===!1),A=!w.some(k=>k.label==="adversarial"||k.label==="prompt_injection")&&y.filter(k=>k.label==="adversarial"||k.label==="prompt_injection").every(k=>!k.result.passed&&(k.result.score??0)<.8),I=y.every(k=>!k.result.feedback.some(x=>/privileged.*leak|hidden grader.*leak/i.test(x))),S=w.map(k=>({fixtureId:k.id,label:k.label,gradeId:k.result.id,reason:(k.expectedScore===void 0?"":`Expected score=${String(k.expectedScore)}; received score=${String(k.result.score)}. `)+`Expected passed=${String(k.expectedPassed)} and rewardEligible=${String(k.expectedRewardEligible)}; received passed=${String(k.result.passed)} and rewardEligible=${String(k.result.rewardEligible)}.`})),M=Jr.parse({schemaVersion:"openpond.graderAuditReport.v1",id:`grader_audit_${h([m.contentHash,y.map(k=>k.result.id)]).slice(0,24)}`,tasksetId:m.id,tasksetHash:m.contentHash,fixtureRefs:y.map(k=>k.id),gradeRefs:y.map(k=>k.result.id),passed:S.length===0&&_&&A&&I,hackingChecksPassed:A,leakageChecksPassed:I,infrastructureSafetyPassed:_,failures:S,createdAt:new Date().toISOString()});return await e.store.saveGraderAuditReport(M),{report:M,passed:M.passed,results:y,failures:S.map(k=>({label:k.label,gradeId:k.gradeId}))}}i(o,"auditFixtures");async function n(d){if(!e.modelJudge)throw new Error("No model judge runner is configured.");let m=await p(d),f=m.graders.filter(S=>S.kind==="model_judge");if(!f.length)throw new Error("Taskset has no model judges to calibrate.");let T=new Map;for(let S of m.graderFixtures)T.set(S.id,(T.get(S.id)??0)+1);for(let S of f)for(let M of S.calibrationFixtureRefs)if(T.get(M)!==1)throw new Error(`Judge ${S.id} calibration requires exactly one fixture for ${M}.`);let y=[],w=[];for(let S of m.graders){if(S.kind!=="model_judge"){w.push(S);continue}let M=m.graderFixtures.filter(E=>S.calibrationFixtureRefs.includes(E.id)),k=[];for(let[E,b]of M.entries()){let D=jt(b.metadata.expectedScore),$=await l(m,b.taskId,Tt(b)),z=Vt(m.id,b,E);try{let L=await e.modelJudge({grader:S,task:$,attempt:z});k.push({fixtureId:b.id,expectedPassed:b.expectedPassed,expectedScore:D,passed:L.passed,score:L.score,feedback:L.feedback,matched:L.passed===b.expectedPassed&&ur(D,L.score)})}catch(L){k.push({fixtureId:b.id,expectedPassed:b.expectedPassed,expectedScore:D,passed:!1,score:0,feedback:L instanceof Error?L.message:String(L),matched:!1})}}let x=k.length>0&&k.every(E=>E.matched),v=h({graderId:S.id,graderVersion:S.version,judge:S.judge,rubric:S.rubric,temperature:S.temperature,results:k});w.push({...S,calibrationStatus:x?"passed":"failed",rewardEligible:x&&S.metadata.requestedRewardEligible===!0,metadata:{...S.metadata,calibrationEvidenceHash:v,calibrationAccuracy:k.length?k.filter(E=>E.matched).length/k.length:0,calibratedAt:new Date().toISOString()}}),y.push({graderId:S.id,passed:x,evidenceHash:v,results:k})}let _=new Date().toISOString(),A=ke.parse({...m,revision:m.revision+1,graders:w,status:"needs_review",readiness:null,contentHash:"00000000",updatedAt:_,metadata:{...m.metadata,judgeCalibration:{parentTasksetHash:m.contentHash,calibratedAt:_,graderIds:y.map(S=>S.graderId)}}}),I=ke.parse({...A,contentHash:Te(A)});if(!e.storeDir)throw new Error("Managed Taskset storage is required for judge calibration.");if(I.purpose!=="benchmark"){let S=Wo.join(e.storeDir,"training","tasksets"),M=Wo.join(S,ce(m));await pt(M,m),I=await Yt(S,I,M)}return await e.store.upsertTaskset(I),{taskset:I,results:y,passed:y.every(S=>S.passed)}}i(n,"calibrateModelJudges");async function c(d,m){let f=await Ee(e.store,d,m),y=(await e.store.listGraderAuditReports(d)).find(_=>_.tasksetHash===f.contentHash)??null;!y&&!f.graders.some(_=>_.kind==="model_judge")&&(y=(await o({tasksetId:d,tasksetRef:m})).report);let w=Ns({taskset:f,graderAudit:y});return await e.store.saveReadinessReport(w),await e.store.upsertTaskset({...f,status:w.ready?"ready":"needs_review",readiness:w,updatedAt:new Date().toISOString()}),w}i(c,"readiness");async function u(d){let m=await p(d.tasksetId);if(m.status!=="ready"||m.readiness?.ready!==!0)throw new Error("Taskset must pass grader audit and readiness before baseline Evaluation.");let f=m.metadata.harnessEvaluationReview;if(!f||typeof f!="object"||Array.isArray(f)||f.id!==d.reviewRef.id||f.contentHash!==d.reviewRef.contentHash)throw new Error("Taskset does not bind the supplied Harness Evaluation review.");let T=(await e.store.listEvaluationResults(m.id,"baseline")).find(b=>b.metadata.sourceTasksetHash===m.contentHash&&b.model.provider===d.model.providerId&&b.model.model===d.model.modelId&&(!m.metrics||b.authoredMetric?.policyHash===h(m.metrics))&&jo(b.metadata.harnessEvaluationReview,d.reviewRef));if(T)return{evaluationResult:T,attempts:[],reused:!0};let y=m.tasks.filter(b=>b.split==="frozen_eval");if(!y.length)throw new Error("Baseline Evaluation requires at least one isolated frozen-evaluation task.");let w=[...new Set(d.seeds?.length?d.seeds:[17])].slice(0,100),_=Math.max(1,Math.min(20,Math.trunc(d.attemptsPerTask??1))),A=new Date().toISOString(),I=await Dt(m,e.storeDir,d.signal),S=[];for(let b of y)for(let D of w)for(let $=0;$<_;$+=1){if(d.signal?.aborted)throw d.signal.reason;let z=await r({tasksetId:m.id,tasksetRef:{id:m.id,revision:m.revision,contentHash:m.contentHash},taskId:b.id,model:d.model,seed:D,attempt:$,sampling:d.sampling,signal:d.signal,admittedAt:A,resultId:`baseline-attempt-${h({reviewRef:d.reviewRef,tasksetHash:m.contentHash,model:d.model,taskId:b.id,seed:D,attempt:$}).slice(0,24)}`},I);S.push(z)}let M=S[0].portable.runManifest;if(S.some(b=>b.portable.runManifest.contentHash!==M.contentHash))throw new Error("Baseline attempts did not share one frozen Run Manifest.");let k=S.map(b=>b.portable.receipt),x=S[0].portable.tasksetRelease,v=k.flatMap(b=>typeof b.metadata.score=="number"?[b.metadata.score]:[]),E=await it({id:`baseline-evaluation-${h({manifest:M.contentHash,receipts:k.map(b=>b.contentHash),reviewRef:d.reviewRef}).slice(0,24)}`,manifest:M,receipts:k,taskset:x,source:I,signal:d.signal,metadata:{kind:"baseline",sourceTasksetId:m.id,sourceTasksetRevision:m.revision,sourceTasksetHash:m.contentHash,harnessEvaluationReview:d.reviewRef,environmentHash:h(x.environment),toolContractHash:h(x.tools),permissionContractHash:h({connectedAppScopes:x.policy.connectedAppScopes,networkPolicy:x.environment.networkPolicy}),policyHash:h(x.policy),verifierRef:{id:`verifier-${m.id}-r${m.revision}`,contentHash:h(x.graders)},scores:v,scoreVariance:No(v),admittedAt:A}});return await e.store.saveEvaluationResult({tasksetId:m.id,kind:"baseline",result:E,createdAt:A}),await Ds({store:e.store,taskset:m,evaluationResult:E}),{evaluationResult:E,attempts:S,reused:!1}}i(u,"executeBaseline");async function g(d){let m=await p(d.tasksetId);if(m.purpose!=="benchmark"||!m.benchmark)throw new Error("This Taskset does not define a benchmark protocol.");let f=d.split??m.benchmark.evaluationSplit,T=d.taskIds?.length?new Set(d.taskIds):null,y=m.tasks.filter(b=>b.split===f&&(!T||T.has(b.id)));if(T&&y.length!==T.size)throw new Error("Benchmark task selection contains an id outside the pinned split.");if(!y.length)throw new Error(`Benchmark split ${f} has no cases.`);let w=[...new Set(d.seeds?.length?d.seeds:[17])].slice(0,100),_=Math.max(1,Math.min(20,Math.trunc(d.repetitions??1))),A=new Date().toISOString(),I=[],S=await Dt(m,e.storeDir,d.signal);for(let b of y)for(let D of w)for(let $=0;$<_;$+=1){if(d.signal?.aborted)throw d.signal.reason;let z=await r({tasksetId:m.id,tasksetRef:{id:m.id,revision:m.revision,contentHash:m.contentHash},taskId:b.id,model:d.model,reasoningEffort:d.reasoningEffort??null,seed:D,attempt:$,sampling:d.sampling,signal:d.signal,releasedHarness:d.releasedHarness,admittedAt:A,parentModelRunId:d.parentModelRunId,toolEvidence:d.toolEvidence,hostedTokenPricing:d.hostedTokenPricing,resultId:`benchmark-attempt-${h({tasksetRelease:m.benchmark.releaseHash,phase:d.phase,model:d.model,reasoningEffort:d.reasoningEffort??null,split:f,taskId:b.id,seed:D,attempt:$,admittedAt:A}).slice(0,24)}`},S);I.push(z),await d.onAttemptComplete?.(z)}let M=I[0].portable.runManifest;if(I.some(b=>b.portable.runManifest.contentHash!==M.contentHash))throw new Error("Benchmark attempts did not share one pinned Run Manifest.");let k=I.map(b=>b.portable.receipt),x=await it({id:`benchmark-evaluation-${h({manifest:M.contentHash,phase:d.phase,receipts:k.map(b=>b.contentHash)}).slice(0,24)}`,manifest:M,receipts:k,taskset:I[0].portable.tasksetRelease,source:S,signal:d.signal,metadata:{kind:"benchmark",phase:d.phase,split:f,reasoningEffort:d.reasoningEffort??null,sourceTasksetId:m.id,sourceTasksetRevision:m.revision,sourceTasksetHash:m.contentHash,benchmarkDefinitionId:m.benchmark.definitionId,admittedAt:A,parentModelRunId:d.parentModelRunId??null}});await e.store.saveEvaluationResult({tasksetId:m.id,kind:d.phase,result:x,createdAt:A});let v=Cr({id:`benchmark-run-${x.contentHash.slice(0,24)}`,phase:d.phase,evaluation:x,receipts:k,reasoningEffort:d.reasoningEffort??null,protocol:{split:f,taskIds:y.map(b=>b.id),seeds:w.map(String),repetitions:_,runtimeTargetHash:h(M.runtimeTarget),environmentHash:h(I[0].portable.tasksetRelease.environment),toolContractHash:h({tasksetTools:I[0].portable.tasksetRelease.tools,harnessTools:I[0].portable.harnessRelease.tools}),limitsHash:h(M.limits)},createdAt:A,metadata:{sourceTasksetId:m.id,split:f,seeds:w,repetitions:_,parentModelRunId:d.parentModelRunId??null}});await e.store.saveBenchmarkRun({tasksetId:m.id,run:v});let E=await Uo({store:e.store,tasksetId:m.id,benchmark:m.benchmark,run:v,createdAt:A});return{evaluationResult:x,run:v,comparison:E,attempts:I}}i(g,"executeBenchmark");async function p(d){let m=await e.store.getTaskset(d);if(!m)throw new Error("Taskset not found.");return m}i(p,"requireTaskset");async function l(d,m,f){let T=d.tasks.find(y=>y.id===m);if(T)return T;if(!d.datasetArtifact||!e.resolveTask)throw new Error(`Task ${m} not found.`);return e.resolveTask({tasksetId:d.id,taskId:m,split:f})}return i(l,"findTask"),{execute:r,grade:s,consumeGraderUsage:a,auditFixtures:o,calibrateModelJudges:n,readiness:c,executeBaseline:u,executeBenchmark:g,close:i(async()=>{},"close")}}i(ap,"createTaskEvaluationService");export{yt as a,Fn as b,Se as c,Rs as d,oi as e,be as f,Ko as g,Ji as h,ed as i,Ns as j,wc as k,Gs as l,It as m,Et as n,gl as o,Rl as p,Ut as q,fd as r,Hs as s,Rd as t,Kl as u,Xl as v,Ho as w,Ed as x,Cd as y,Ds as z,ap as A};

package/dist/chunks/{chunk-HPBQ4VTW.js → chunk-BVQWVIGK.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{v as r}from"./chunk-BLXEAZCQ.js";import{N as s,O as a,P as d}from"./chunk-QL4OIPEZ.js";import{a as n}from"./chunk-FTQLQJQV.js";async function u(){return{...s.parse(await r({path:"/saved-work"})),webBaseUrl:i()}}n(u,"listHostedSavedWork");async function m(e){let t=a.parse(e);return r({path:"/saved-work",method:"POST",body:t})}n(m,"createHostedSavedWork");async function h(e,t){let o=d.parse(t);return r({path:`/saved-work/schedules/${encodeURIComponent(e)}`,method:"PATCH",body:o})}n(h,"updateHostedSavedWork");async function k(e){return r({path:`/saved-work/schedules/${encodeURIComponent(e)}`,method:"DELETE"})}n(k,"deleteHostedSavedWork");async function P(e,t){return r({path:`/saved-work/schedules/${encodeURIComponent(e)}/run`,method:"POST",body:{clientRequestId:t}})}n(P,"runHostedSavedWork");function i(e=process.env){let t=e.OPENPOND_HOSTED_WEB_URL?.trim()||e.OPENPOND_SANDBOX_BASE_URL?.trim()||e.OPENPOND_API_URL?.trim()||e.OPENPOND_SANDBOX_API_URL?.trim()||"https://api.openpond.ai",o=new URL(t);return o.hostname==="api.openpond.ai"||o.hostname.endsWith(".api.openpond.ai")?"https://openpond.ai":(o.pathname="",o.search="",o.hash="",o.origin)}n(i,"hostedWebBaseUrl");export{u as a,m as b,h as c,k as d,P as e,i as f};
+import{v as r}from"./chunk-BL3YH2QU.js";import{N as s,O as a,P as d}from"./chunk-2RNQQNB7.js";import{a as n}from"./chunk-FTQLQJQV.js";async function u(){return{...s.parse(await r({path:"/saved-work"})),webBaseUrl:i()}}n(u,"listHostedSavedWork");async function m(e){let t=a.parse(e);return r({path:"/saved-work",method:"POST",body:t})}n(m,"createHostedSavedWork");async function h(e,t){let o=d.parse(t);return r({path:`/saved-work/schedules/${encodeURIComponent(e)}`,method:"PATCH",body:o})}n(h,"updateHostedSavedWork");async function k(e){return r({path:`/saved-work/schedules/${encodeURIComponent(e)}`,method:"DELETE"})}n(k,"deleteHostedSavedWork");async function P(e,t){return r({path:`/saved-work/schedules/${encodeURIComponent(e)}/run`,method:"POST",body:{clientRequestId:t}})}n(P,"runHostedSavedWork");function i(e=process.env){let t=e.OPENPOND_HOSTED_WEB_URL?.trim()||e.OPENPOND_SANDBOX_BASE_URL?.trim()||e.OPENPOND_API_URL?.trim()||e.OPENPOND_SANDBOX_API_URL?.trim()||"https://api.openpond.ai",o=new URL(t);return o.hostname==="api.openpond.ai"||o.hostname.endsWith(".api.openpond.ai")?"https://openpond.ai":(o.pathname="",o.search="",o.hash="",o.origin)}n(i,"hostedWebBaseUrl");export{u as a,m as b,h as c,k as d,P as e,i as f};

package/dist/chunks/{chunk-SWUEFLOH.js → chunk-C37PUHA6.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as m}from"./chunk-UKMA4R6X.js";import{e as h}from"./chunk-YR5ID2SS.js";import{a as s}from"./chunk-FTQLQJQV.js";var g=5,A=10,f=1200;function b(e){let r=y(e.endpoint),t=e.fetchImpl??fetch,n=e.apiKey?.trim()||null;return async(i,c)=>{let a=S(i),u=new Headers;u.set("Accept","application/json"),u.set("Content-Type","application/json"),u.set("x-openpond-client","openpond-app"),n&&(n.startsWith("opk_")&&u.set("openpond-api-key",n),u.set("Authorization",`Bearer ${n}`));let l=await t(r,{method:"POST",headers:u,body:JSON.stringify(a),signal:c?.signal});if(!l.ok)throw new Error(`Hosted web search failed: ${l.status} ${await O(l)}`);return w(await l.json(),a.query)}}s(b,"createHostedWebSearchExecutor");function D(e=process.env,r={}){let t=r.loadAccountContext??m;return async(n,i)=>{let c=p(e.OPENPOND_WEB_SEARCH_API_KEY,e.OPENPOND_SEARCH_API_KEY,e.OPENPOND_API_KEY),a=c?null:await t().catch(()=>null),u=E(e,a),l=c??a?.token??null;return b({endpoint:u,apiKey:l,fetchImpl:r.fetchImpl})(n,i)}}s(D,"createWebSearchExecutorFromEnv");function E(e=process.env,r){let t=p(e.OPENPOND_WEB_SEARCH_ENDPOINT);if(t)return y(t);let n=p(e.OPENPOND_SEARCH_API_URL);if(n)return d(n);let i=p(e.OPENPOND_PUBLIC_API_URL,e.OPENPOND_API_URL,r?.apiBaseUrl);if(i)return d(i);let c=p(e.OPENPOND_OPCHAT_API_URL,e.OPENPOND_CHAT_API_URL,r?.chatApiBaseUrl);return d(c||h)}s(E,"resolveWebSearchEndpoint");function d(e){let r=e.trim().replace(/\/+$/,"");if(!r)throw new Error("Hosted web search endpoint is required.");let t=new URL(r);if(t.protocol!=="http:"&&t.protocol!=="https:")throw new Error("Hosted web search endpoint must use http or https.");let n=t.pathname.split("/").filter(Boolean),i=n.findIndex(c=>c.toLowerCase()==="opchat");if(i>=0)t.pathname=`/${[...n.slice(0,i),"v1","search"].join("/")}`;else{let c=n.findIndex(a=>a.toLowerCase()==="search");if(c>=0)t.pathname=`/${n.slice(0,c+1).join("/")}`;else{let a=n.findIndex(u=>u.toLowerCase()==="v1");t.pathname=a>=0?`/${[...n.slice(0,a+1),"search"].join("/")}`:`${t.pathname.replace(/\/+$/,"")}/v1/search`.replace(/\/{2,}/g,"/")}}return t.search="",t.hash="",t.toString()}s(d,"normalizeSearchApiUrl");function S(e){let r=e.query.trim();if(!r)throw new Error("web_search query is required.");let t=_(e.limit),n=Array.isArray(e.domains)?e.domains.map(i=>i.trim()).filter(Boolean).slice(0,10):void 0;return{query:r,limit:t,...typeof e.recencyDays=="number"&&Number.isFinite(e.recencyDays)&&e.recencyDays>=0?{recencyDays:Math.floor(e.recencyDays)}:{},...n&&n.length>0?{domains:n}:{}}}s(S,"normalizeWebSearchRequest");function y(e){let r=e.trim();if(!r)throw new Error("Hosted web search endpoint is required.");let t=new URL(r);if(t.protocol!=="http:"&&t.protocol!=="https:")throw new Error("Hosted web search endpoint must use http or https.");return t.toString()}s(y,"normalizeEndpoint");function p(...e){for(let r of e){let t=r?.trim();if(t)return t}return null}s(p,"firstNonEmpty");function _(e){return typeof e!="number"||!Number.isFinite(e)||e<=0?g:Math.min(Math.floor(e),A)}s(_,"normalizeLimit");function w(e,r){let t=e&&typeof e=="object"?e:{},i=(Array.isArray(t.results)?t.results:Array.isArray(t.items)?t.items:[]).map(P).filter(c=>!!c);return{query:o(t.query)??r,provider:o(t.provider)??"hosted",searchedAt:o(t.searchedAt)??o(t.searched_at)??new Date().toISOString(),results:i,truncated:t.truncated===!0}}s(w,"normalizeWebSearchResult");function P(e,r){if(!e||typeof e!="object")return null;let t=e,n=o(t.url),i=o(t.title);return!n||!i?null:{id:o(t.id)??`result_${r+1}`,title:i,url:n,snippet:R(o(t.snippet)??o(t.description)??o(t.text)??o(t.content)),sourceName:o(t.sourceName)??o(t.source_name)??o(t.source)??I(n),faviconUrl:x(o(t.faviconUrl)??o(t.favicon_url)??o(t.iconUrl)??o(t.icon_url))??N(n),publishedAt:o(t.publishedAt)??o(t.published_at)??o(t.publishedDate)??o(t.published_date),updatedAt:o(t.updatedAt)??o(t.updated_at)??o(t.updatedDate)??o(t.updated_date)}}s(P,"normalizeWebSearchItem");function R(e){if(!e)return"";let r=e.replace(/\s+/g," ").trim();return r.length<=f?r:`${r.slice(0,f-1).trimEnd()}\u2026`}s(R,"normalizeSnippet");function x(e){if(!e)return null;try{let r=new URL(e);return r.protocol!=="http:"&&r.protocol!=="https:"?null:r.toString()}catch{return null}}s(x,"normalizeOptionalUrl");function N(e){try{return new URL("/favicon.ico",e).toString()}catch{return null}}s(N,"faviconUrlFromPageUrl");function I(e){try{return new URL(e).hostname.replace(/^www\./i,"")}catch{return null}}s(I,"hostnameFromUrl");async function O(e){let r=await e.text().catch(()=>"");if(!r)return e.statusText||`HTTP ${e.status}`;try{let t=JSON.parse(r);if(t&&typeof t=="object"){let n=t,i=n.error;return i&&typeof i=="object"?o(i.message)??JSON.stringify(t):o(n.message)??o(n.error)??JSON.stringify(t)}}catch{return r}return r}s(O,"readSearchError");function o(e){return typeof e=="string"&&e.trim()?e.trim():null}s(o,"stringValue");export{b as a,D as b,E as c,d,S as e};
+import{a as m}from"./chunk-GOPFFQZU.js";import{e as h}from"./chunk-YR5ID2SS.js";import{a as s}from"./chunk-FTQLQJQV.js";var g=5,A=10,f=1200;function b(e){let r=y(e.endpoint),t=e.fetchImpl??fetch,n=e.apiKey?.trim()||null;return async(i,c)=>{let a=S(i),u=new Headers;u.set("Accept","application/json"),u.set("Content-Type","application/json"),u.set("x-openpond-client","openpond-app"),n&&(n.startsWith("opk_")&&u.set("openpond-api-key",n),u.set("Authorization",`Bearer ${n}`));let l=await t(r,{method:"POST",headers:u,body:JSON.stringify(a),signal:c?.signal});if(!l.ok)throw new Error(`Hosted web search failed: ${l.status} ${await O(l)}`);return w(await l.json(),a.query)}}s(b,"createHostedWebSearchExecutor");function D(e=process.env,r={}){let t=r.loadAccountContext??m;return async(n,i)=>{let c=p(e.OPENPOND_WEB_SEARCH_API_KEY,e.OPENPOND_SEARCH_API_KEY,e.OPENPOND_API_KEY),a=c?null:await t().catch(()=>null),u=E(e,a),l=c??a?.token??null;return b({endpoint:u,apiKey:l,fetchImpl:r.fetchImpl})(n,i)}}s(D,"createWebSearchExecutorFromEnv");function E(e=process.env,r){let t=p(e.OPENPOND_WEB_SEARCH_ENDPOINT);if(t)return y(t);let n=p(e.OPENPOND_SEARCH_API_URL);if(n)return d(n);let i=p(e.OPENPOND_PUBLIC_API_URL,e.OPENPOND_API_URL,r?.apiBaseUrl);if(i)return d(i);let c=p(e.OPENPOND_OPCHAT_API_URL,e.OPENPOND_CHAT_API_URL,r?.chatApiBaseUrl);return d(c||h)}s(E,"resolveWebSearchEndpoint");function d(e){let r=e.trim().replace(/\/+$/,"");if(!r)throw new Error("Hosted web search endpoint is required.");let t=new URL(r);if(t.protocol!=="http:"&&t.protocol!=="https:")throw new Error("Hosted web search endpoint must use http or https.");let n=t.pathname.split("/").filter(Boolean),i=n.findIndex(c=>c.toLowerCase()==="opchat");if(i>=0)t.pathname=`/${[...n.slice(0,i),"v1","search"].join("/")}`;else{let c=n.findIndex(a=>a.toLowerCase()==="search");if(c>=0)t.pathname=`/${n.slice(0,c+1).join("/")}`;else{let a=n.findIndex(u=>u.toLowerCase()==="v1");t.pathname=a>=0?`/${[...n.slice(0,a+1),"search"].join("/")}`:`${t.pathname.replace(/\/+$/,"")}/v1/search`.replace(/\/{2,}/g,"/")}}return t.search="",t.hash="",t.toString()}s(d,"normalizeSearchApiUrl");function S(e){let r=e.query.trim();if(!r)throw new Error("web_search query is required.");let t=_(e.limit),n=Array.isArray(e.domains)?e.domains.map(i=>i.trim()).filter(Boolean).slice(0,10):void 0;return{query:r,limit:t,...typeof e.recencyDays=="number"&&Number.isFinite(e.recencyDays)&&e.recencyDays>=0?{recencyDays:Math.floor(e.recencyDays)}:{},...n&&n.length>0?{domains:n}:{}}}s(S,"normalizeWebSearchRequest");function y(e){let r=e.trim();if(!r)throw new Error("Hosted web search endpoint is required.");let t=new URL(r);if(t.protocol!=="http:"&&t.protocol!=="https:")throw new Error("Hosted web search endpoint must use http or https.");return t.toString()}s(y,"normalizeEndpoint");function p(...e){for(let r of e){let t=r?.trim();if(t)return t}return null}s(p,"firstNonEmpty");function _(e){return typeof e!="number"||!Number.isFinite(e)||e<=0?g:Math.min(Math.floor(e),A)}s(_,"normalizeLimit");function w(e,r){let t=e&&typeof e=="object"?e:{},i=(Array.isArray(t.results)?t.results:Array.isArray(t.items)?t.items:[]).map(P).filter(c=>!!c);return{query:o(t.query)??r,provider:o(t.provider)??"hosted",searchedAt:o(t.searchedAt)??o(t.searched_at)??new Date().toISOString(),results:i,truncated:t.truncated===!0}}s(w,"normalizeWebSearchResult");function P(e,r){if(!e||typeof e!="object")return null;let t=e,n=o(t.url),i=o(t.title);return!n||!i?null:{id:o(t.id)??`result_${r+1}`,title:i,url:n,snippet:R(o(t.snippet)??o(t.description)??o(t.text)??o(t.content)),sourceName:o(t.sourceName)??o(t.source_name)??o(t.source)??I(n),faviconUrl:x(o(t.faviconUrl)??o(t.favicon_url)??o(t.iconUrl)??o(t.icon_url))??N(n),publishedAt:o(t.publishedAt)??o(t.published_at)??o(t.publishedDate)??o(t.published_date),updatedAt:o(t.updatedAt)??o(t.updated_at)??o(t.updatedDate)??o(t.updated_date)}}s(P,"normalizeWebSearchItem");function R(e){if(!e)return"";let r=e.replace(/\s+/g," ").trim();return r.length<=f?r:`${r.slice(0,f-1).trimEnd()}\u2026`}s(R,"normalizeSnippet");function x(e){if(!e)return null;try{let r=new URL(e);return r.protocol!=="http:"&&r.protocol!=="https:"?null:r.toString()}catch{return null}}s(x,"normalizeOptionalUrl");function N(e){try{return new URL("/favicon.ico",e).toString()}catch{return null}}s(N,"faviconUrlFromPageUrl");function I(e){try{return new URL(e).hostname.replace(/^www\./i,"")}catch{return null}}s(I,"hostnameFromUrl");async function O(e){let r=await e.text().catch(()=>"");if(!r)return e.statusText||`HTTP ${e.status}`;try{let t=JSON.parse(r);if(t&&typeof t=="object"){let n=t,i=n.error;return i&&typeof i=="object"?o(i.message)??JSON.stringify(t):o(n.message)??o(n.error)??JSON.stringify(t)}}catch{return r}return r}s(O,"readSearchError");function o(e){return typeof e=="string"&&e.trim()?e.trim():null}s(o,"stringValue");export{b as a,D as b,E as c,d,S as e};