openpond 0.2.69 → 0.2.71

package/dist/chunks/{chunk-WCNPGKST.js → chunk-O4YHISU2.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{k as W}from"./chunk-WD4KAKEX.js";import{d as U,e as Y}from"./chunk-JVMSN7LH.js";import{Ia as K,Ka as V,La as N,Y as v,_ as w,ka as q}from"./chunk-BU6MW65V.js";import{V as z,W as B,Xb as j,f as _,k as M,l as $,la as G,ma as D,ta as F,ua as L}from"./chunk-D6L6UNUQ.js";import{S as P}from"./chunk-4PUOWSDE.js";import{e as E,i as I}from"./chunk-OTDDXEQ6.js";import{i as x,l as C,n as u}from"./chunk-RCT3O3HQ.js";import{a as c}from"./chunk-FTQLQJQV.js";import{constants as b}from"node:fs";import{lstat as oe,open as ne,readdir as ie}from"node:fs/promises";import S from"node:path";function pe(e){let t=N(e.taskset),{agentSnapshot:a,harnessRelease:s}=e.releasedHarness?{agentSnapshot:E.parse(e.releasedHarness.agentSnapshot),harnessRelease:I.parse(e.releasedHarness.harnessRelease)}:Q(e.taskset,e.profile);if(s.agentSnapshot.id!==a.id||s.agentSnapshot.contentHash!==a.contentHash)throw new Error("Released Harness does not bind the supplied Agent snapshot.");let d=V(e.taskset),n=e.taskset.tasks.length?e.taskset.tasks:e.selectedTask?[e.selectedTask]:[];if(!n.length)throw new Error("A portable Taskset release requires at least one resolved task.");let{environmentRelease:i,verifierSetRelease:o,tasksetRelease:l}=K({taskset:e.taskset,selectedTasks:n,adapterId:e.adapterId??A(e.taskset),admittedTasksetRelease:e.tasksetRelease,rewardExecution:e.rewardExecution}),R=e.now??(()=>new Date().toISOString()),f=M({schemaVersion:"openpond.runManifest.v1",id:`run-manifest-${u([s.contentHash,l.contentHash,e.model]).slice(0,24)}`,harnessRelease:{id:s.id,contentHash:s.contentHash},tasksetRelease:{id:l.id,contentHash:l.contentHash},model:{provider:e.model.providerId,model:e.model.modelId,revision:null,artifactHash:null,tokenizerRevision:null,chatTemplateHash:null},runtimeTarget:{adapterId:e.adapterId??A(e.taskset),placement:"local",runtimeVersion:"desktop-v1",capabilityReceipt:u({environment:d,tools:t,capabilities:e.taskset.capabilities})},limits:{maxTurns:128,timeoutMs:e.taskset.environment.defaultTimeoutMs,maxOutputBytes:25e7,maximumSpendUsd:null},approval:null,createdAt:R(),metadata:{sourceTasksetId:e.taskset.id,sourceTasksetRevision:e.taskset.revision,reasoningEffort:e.reasoningEffort??null}});return{agentSnapshot:a,harnessRelease:s,environmentRelease:i,verifierSetRelease:o,tasksetRelease:l,runManifest:f}}c(pe,"compileDesktopHarnessContext");function Q(e,t){let a=e.profileRelease?{id:e.profileRelease.id,contentHash:k(e.profileRelease.contentHash)}:t?.git?.head?{id:`source-${h(t.git.head)}`,contentHash:k(t.git.head)}:null;return W({sourceRelease:a,profileHead:t?.git?.head??null,skills:se(t),agents:re(t)})}c(Q,"compileTemporaryProfileHarness");function ee(e){let t=e.artifacts.map(T),a=e.artifacts.find(n=>n.kind==="environment_state")??e.artifacts.find(n=>n.kind==="runtime_trace")??e.artifacts.find(n=>n.kind==="raw_model_response"),s=J(e.attempt,e.grade),d=e.grade.components.map(n=>({id:`grader-${h(n.graderId)}-${e.grade.id}`,contentHash:u(n),mediaType:"application/vnd.openpond.grader-evidence+json",sizeBytes:Buffer.byteLength(C(n))}));return _.parse($({schemaVersion:"openpond.attemptReceipt.v1",id:`receipt-${u([e.manifest.contentHash,e.attempt.id,e.grade.id]).slice(0,24)}`,runManifest:{id:e.manifest.id,contentHash:e.manifest.contentHash},taskId:e.attempt.taskId,seed:String(e.attempt.seed),terminal:!["infrastructure_failure","timeout","cancelled"].includes(s??""),failureClass:s,outputHash:e.attempt.infrastructureError?null:u(e.attempt.output),traceHash:a?k(a.sha256):u({attemptId:e.attempt.id,output:e.attempt.output,runtimeEventRefs:e.attempt.runtimeEventRefs,artifactRefs:e.attempt.artifactRefs,metadata:e.attempt.metadata}),artifactRefs:t,graderEvidenceRefs:d,startedAt:e.attempt.startedAt,completedAt:e.attempt.completedAt,latencyMs:e.attempt.latencyMs,costUsd:e.attempt.costUsd,legacyAttemptRef:e.attempt.id,metadata:{legacyGradeRef:e.grade.id,rewardEligible:e.grade.rewardEligible,score:e.grade.score,passed:e.grade.passed,usage:e.attempt.metadata.usage??[]}}))}c(ee,"projectDesktopAttemptReceipt");function ue(e){let t=ee({manifest:e.context.runManifest,attempt:e.attempt,grade:e.grade,artifacts:e.artifacts}),a=e.context.tasksetRelease.tasks.find(r=>r.id===e.attempt.taskId);if(!a)throw new Error(`Task ${e.attempt.taskId} is absent from the admitted Taskset Release.`);let s=a.requiredOutputs??[],d=e.artifacts.filter(r=>r.kind==="output_artifact").map(r=>({path:typeof r.metadata.requiredOutputPath=="string"?r.metadata.requiredOutputPath:r.path,artifact:T(r),detectedMediaType:r.mediaType??null,status:"collected",parseStatus:X(r.metadata.parseStatus),schemaStatus:X(r.metadata.schemaStatus),evidenceRefs:[T(r)],metadata:{sourceArtifactId:r.id}})),n=F({id:`artifact-manifest-${t.id}`,attemptRef:{id:t.id,contentHash:t.contentHash},requiredOutputs:s,collectedArtifacts:d,createdAt:e.attempt.completedAt,metadata:{legacyAttemptRef:e.attempt.id}}),i=e.grade.components.map((r,O)=>te({component:r,grader:e.context.verifierSetRelease.graders.find(g=>g.id===r.graderId),evidenceRef:t.graderEvidenceRefs[O]??null,gradeScored:e.grade.score!==null,boundResult:e.grade.rewardComposition?.results.find(g=>g.graderId===r.graderId)})),o=L({requiredOutputs:s,manifest:n}),l=ae(e.attempt,e.grade),R=o.some(r=>r.status==="unscorable"&&r.failureOwner==="collector"),f=o.some(r=>r.status==="scored"&&!r.passed),p=l.outcomeClass==="completed"&&R?{outcomeClass:"collector_failure",failureOwner:"collector"}:l.outcomeClass==="completed"&&f?{outcomeClass:"incomplete_output",failureOwner:"policy"}:l,m=D({id:`reward-${u([t.contentHash,e.context.verifierSetRelease.contentHash,n.contentHash,e.grade.id]).slice(0,24)}`,attemptRef:{id:t.id,contentHash:t.contentHash},verifierSet:e.context.verifierSetRelease,artifactManifest:n,...p,components:[...i,...o],createdAt:e.grade.createdAt,metadata:{legacyGradeRef:e.grade.id,legacyAttemptRef:e.attempt.id,...e.grade.rewardComposition?{rewardComposition:e.grade.rewardComposition}:{}}});return{attemptReceipt:t,artifactManifest:n,rewardReceipt:m}}c(ue,"projectDesktopCanonicalReceipts");function Re(e){let t=e.artifacts.find(n=>n.kind==="runtime_trace")??e.artifacts.find(n=>n.kind==="raw_model_response"),a=t?T(t):{id:`trace-${e.canonical.attemptReceipt.id}`,contentHash:e.canonical.attemptReceipt.traceHash,mediaType:"application/vnd.openpond.attempt-trace+json",sizeBytes:null},s=e.canonical.rewardReceipt.outcomeClass==="task_deadline"||e.canonical.rewardReceipt.outcomeClass==="infrastructure_timeout",d=e.canonical.rewardReceipt.outcomeClass==="cancelled";return j({id:`rollout-${u([e.canonical.attemptReceipt.contentHash,e.canonical.rewardReceipt.contentHash,a.contentHash]).slice(0,24)}`,attemptReceipt:e.canonical.attemptReceipt,rewardReceipt:e.canonical.rewardReceipt,artifactManifestRef:{id:e.canonical.artifactManifest.id,contentHash:e.canonical.artifactManifest.contentHash},tasksetRelease:{id:e.context.tasksetRelease.id,contentHash:e.context.tasksetRelease.contentHash},environmentRelease:{id:e.context.environmentRelease.id,contentHash:e.context.environmentRelease.contentHash},harnessRelease:{id:e.context.harnessRelease.id,contentHash:e.context.harnessRelease.contentHash},taskId:e.attempt.taskId,split:e.attempt.split,model:e.context.runManifest.model,seed:String(e.attempt.seed),traceRef:a,optimizerSample:e.optimizerSample??null,environmentExecutions:[{id:`environment-execution-${e.attempt.id}`,environmentRelease:{id:e.context.environmentRelease.id,contentHash:e.context.environmentRelease.contentHash},status:d?"cancelled":s?"timed_out":e.canonical.rewardReceipt.status==="scored"?"completed":"failed",startedAt:e.attempt.startedAt,completedAt:e.attempt.completedAt,traceRefs:[a],metadata:{legacyAttemptRef:e.attempt.id,runtimeEventCount:e.attempt.runtimeEventRefs.length}}],startedAt:e.attempt.startedAt,completedAt:e.attempt.completedAt,metadata:{source:"openpond.desktop-local-runtime",optimizerSamplePresent:!!e.optimizerSample}})}c(Re,"projectDesktopCanonicalRollout");function te(e){if(!e.grader)throw new Error(`Grade component ${e.component.graderId} has no admitted Verifier definition.`);let t=(e.boundResult?e.boundResult.status==="scored":e.gradeScored)?"scored":"unscorable",a=t==="scored"&&e.component.rewardEligible,s=e.evidenceRef?[e.evidenceRef]:[];return{verifierId:e.component.graderId,verifierVersion:e.component.graderVersion,status:t,rawScore:t==="scored"?e.boundResult?.rawScore??e.component.score:null,normalizedScore:t==="scored"?e.boundResult?.normalizedScore??e.component.score:null,weight:e.grader.weight,passed:e.component.passed,hardGate:e.component.hardGate,rewardEligible:a,rewardContribution:a?e.component.score:null,failureOwner:e.component.passed?null:t==="scored"?"policy":"verifier",feedback:e.component.feedback?[e.component.feedback]:[],visibleEvidenceRefs:e.grader.privileged?[]:s,privilegedEvidenceRefs:e.grader.privileged?s:[],metadata:{judge:e.component.judge??null,calibrationStatus:e.component.calibrationStatus,legacyEvidenceRefs:e.component.evidenceRefs}}}c(te,"portableRewardComponent");function ae(e,t){let a=B.safeParse(e.metadata.outcomeClass);if(a.success){let d=e.metadata.failureOwner===null?null:z.safeParse(e.metadata.failureOwner);if(d!==null&&!d.success)throw new Error("Attempt has an invalid explicit failure owner.");return{outcomeClass:a.data,failureOwner:d===null?null:d.data}}let s=J(e,t);return s?G({failureClass:s,timeoutKind:e.metadata.timeoutKind==="task_deadline"?"task_deadline":"infrastructure_timeout"}):t.score===null?{outcomeClass:"verifier_failure",failureOwner:"verifier"}:t.passed?{outcomeClass:"completed",failureOwner:null}:{outcomeClass:"policy_failure",failureOwner:"policy"}}c(ae,"canonicalOutcome");function X(e){return e==="passed"||e==="failed"?e:"not_requested"}c(X,"artifactValidationStatus");function se(e){return(e?.skills??[]).filter(t=>t.enabled).map(t=>({id:`skill-${h(t.name)}`,path:`skills/${h(t.name)}/SKILL.md`,contentHash:k(t.sourceHash),sizeBytes:Math.max(0,t.charCount),mediaType:"text/markdown",visibility:"policy"}))}c(se,"portableSkills");function re(e){let t=new Map;for(let a of e?.agents??[])a.enabled&&t.set(a.id,u({name:a.name,path:a.path}));return[...t].map(([a,s])=>({id:`agent-${h(a)}`,path:`agents/${h(a)}/release.json`,contentHash:k(s),sizeBytes:0,mediaType:"application/vnd.openpond.agent-release+json",visibility:"policy"}))}c(re,"portableAgents");function A(e){let t=e.environment.metadata.runtimeAdapterId;if(typeof t=="string"&&t.trim())return t.trim();let a=e.environment.metadata.benchmark;if(a&&typeof a=="object"&&!Array.isArray(a)){let s=Reflect.get(a,"id");if(typeof s=="string"&&s.trim())return`conformance:${s.trim()}`}return e.environment.kind==="work"?"openpond:desktop-work":e.environment.kind==="chat"?"openpond:desktop-text":`openpond:${h(e.environment.entrypoint)}`}c(A,"desktopTasksetRuntimeAdapterId");function T(e){return{id:e.id,contentHash:k(e.sha256),mediaType:e.mediaType??null,sizeBytes:e.sizeBytes}}c(T,"portableArtifact");function J(e,t){if(e.infrastructureError)return"infrastructure_failure";if(t.failureClass)return t.failureClass;let a=e.metadata.failureClass;return a==="policy_failure"||a==="grader_failure"||a==="environment_failure"||a==="infrastructure_failure"||a==="timeout"||a==="cancelled"?a:null}c(J,"attemptFailureClass");function k(e){return/^[a-f0-9]{64}$/.test(e)?e:u(e)}c(k,"hash");function h(e){return e.trim().replace(/[^a-zA-Z0-9._-]+/g,"-").replace(/^-+|-+$/g,"")||"asset"}c(h,"segment");async function be(e,t){if(e.metadata.taskDefinition!==void 0||e.metadata.rewardBinding!==void 0)return{taskset:e,generatedFiles:[]};let a=await oe(t);if(!a.isDirectory()||a.isSymbolicLink())throw new Error("Taskset source must be a regular directory.");let s=new Map,d=0;async function n(i=""){for(let o of await ie(S.join(t,i),{withFileTypes:!0})){let l=S.posix.join(i,o.name);if(o.isSymbolicLink())throw new Error("Taskset files cannot contain symbolic links.");if(o.name.startsWith(".env"))throw new Error("Environment files cannot be included in a Taskset package.");if(o.isDirectory()){await n(l);continue}if(!o.isFile())throw new Error("Taskset packages require regular files.");if(U(l))continue;if(s.size>=1e4)throw new Error("Taskset source exceeds the file inventory limit.");let R=await ne(S.join(t,l),b.O_RDONLY|b.O_NOFOLLOW|b.O_NONBLOCK);try{let f=await R.stat();if(d+=f.size,!f.isFile()||d>w*3/4)throw new Error("Taskset source exceeds the package byte limit.");let p=Buffer.alloc(f.size),m=0;for(;m<p.length;){let r=await R.read(p,m,p.length-m,m);if(!r.bytesRead)throw new Error("Taskset source changed during capture.");m+=r.bytesRead}if((await R.read(Buffer.alloc(1),0,1,m)).bytesRead)throw new Error("Taskset source changed during capture.");s.set(l,p)}finally{await R.close()}}}return c(n,"visit"),await n(),Y(e,s,A(e))}c(be,"prepareAuthoredTasksetFiles");import{constants as H}from"node:fs";import{open as ce,realpath as Z}from"node:fs/promises";import y from"node:path";function Me(e,t){return t?new Set:new Set([...(e.modelResources?.assets??[]).map(a=>a.asset.id),...(e.learningResources?.assets??[]).map(a=>a.asset.id),...e.learningResources?v(e.learningResources).map(a=>a.asset.id):[]])}c(Me,"tasksetPackageInlineAssetIds");async function $e(e){let t=e.sources.length?await Z(e.root):y.resolve(e.root),a=[],s=new Set,d=0,n=c((i,o)=>{if(s.has(i.id))throw new Error(`Duplicate local Taskset asset: ${i.id}.`);if(s.add(i.id),d+=4*Math.ceil(o.byteLength/3),d>w)throw new Error("Local Taskset files exceed the package envelope.");a.push({asset:i,base64:Buffer.from(o).toString("base64")})},"append");for(let i of e.sources){let o=x.parse(i.asset);if(!i.sourcePath||y.isAbsolute(i.sourcePath))throw new Error(`Taskset asset ${o.id} requires a relative source path.`);let l=y.resolve(t,i.sourcePath);if(!l.startsWith(`${t}${y.sep}`))throw new Error(`Taskset asset ${o.id} escapes its local package.`);if(!(await Z(l)).startsWith(`${t}${y.sep}`))throw new Error(`Taskset asset ${o.id} escapes its local package.`);let f=await ce(l,H.O_RDONLY|H.O_NOFOLLOW|H.O_NONBLOCK);try{let p=await f.stat();if(!p.isFile()||p.size!==o.sizeBytes)throw new Error(`Taskset asset ${o.id} differs from its immutable file size.`);if(4*Math.ceil(p.size/3)+d>w)throw new Error("Local Taskset files exceed the package envelope.");let m=Buffer.alloc(p.size),r=0;for(;r<m.length;){let g=await f.read(m,r,m.length-r,r);if(!g.bytesRead)throw new Error(`Taskset asset ${o.id} changed during capture.`);r+=g.bytesRead}if((await f.read(Buffer.alloc(1),0,1,r)).bytesRead)throw new Error(`Taskset asset ${o.id} changed during capture.`);n(o,m)}finally{await f.close()}}for(let i of[...e.content.modelResources?.assets??[],...e.content.learningResources?.assets??[]])s.has(i.asset.id)||n(i.asset,Buffer.from(P(i,i.asset),"utf8"));if(e.content.learningResources)for(let i of v(e.content.learningResources))s.has(i.asset.id)||n(i.asset,Buffer.from(i.base64,"base64"));if(e.fileOrder){if(new Set(e.fileOrder).size!==a.length||e.fileOrder.length!==a.length||e.fileOrder.some(o=>!s.has(o)))throw new Error("Saved package file order differs from its inventory.");let i=new Map(e.fileOrder.map((o,l)=>[o,l]));a.sort((o,l)=>i.get(o.asset.id)-i.get(l.asset.id))}return q({...e.content,files:a})}c($e,"captureLocalTasksetPackage");export{pe as a,ue as b,Re as c,A as d,be as e,Me as f,$e as g};
+import{k as W}from"./chunk-64EDQ7UV.js";import{d as U,e as Y}from"./chunk-GR4MJJCE.js";import{Ia as K,Ka as V,La as N,Y as v,_ as w,ka as q}from"./chunk-BYVE4TMU.js";import{V as z,W as B,Xb as j,f as _,k as M,l as $,la as G,ma as D,ta as F,ua as L}from"./chunk-D6L6UNUQ.js";import{S as P}from"./chunk-ZEOYNUHL.js";import{e as E,i as I}from"./chunk-OTDDXEQ6.js";import{i as x,l as C,n as u}from"./chunk-RCT3O3HQ.js";import{a as c}from"./chunk-FTQLQJQV.js";import{constants as b}from"node:fs";import{lstat as oe,open as ne,readdir as ie}from"node:fs/promises";import S from"node:path";function pe(e){let t=N(e.taskset),{agentSnapshot:a,harnessRelease:s}=e.releasedHarness?{agentSnapshot:E.parse(e.releasedHarness.agentSnapshot),harnessRelease:I.parse(e.releasedHarness.harnessRelease)}:Q(e.taskset,e.profile);if(s.agentSnapshot.id!==a.id||s.agentSnapshot.contentHash!==a.contentHash)throw new Error("Released Harness does not bind the supplied Agent snapshot.");let d=V(e.taskset),n=e.taskset.tasks.length?e.taskset.tasks:e.selectedTask?[e.selectedTask]:[];if(!n.length)throw new Error("A portable Taskset release requires at least one resolved task.");let{environmentRelease:i,verifierSetRelease:o,tasksetRelease:l}=K({taskset:e.taskset,selectedTasks:n,adapterId:e.adapterId??A(e.taskset),admittedTasksetRelease:e.tasksetRelease,rewardExecution:e.rewardExecution}),R=e.now??(()=>new Date().toISOString()),f=M({schemaVersion:"openpond.runManifest.v1",id:`run-manifest-${u([s.contentHash,l.contentHash,e.model]).slice(0,24)}`,harnessRelease:{id:s.id,contentHash:s.contentHash},tasksetRelease:{id:l.id,contentHash:l.contentHash},model:{provider:e.model.providerId,model:e.model.modelId,revision:null,artifactHash:null,tokenizerRevision:null,chatTemplateHash:null},runtimeTarget:{adapterId:e.adapterId??A(e.taskset),placement:"local",runtimeVersion:"desktop-v1",capabilityReceipt:u({environment:d,tools:t,capabilities:e.taskset.capabilities})},limits:{maxTurns:128,timeoutMs:e.taskset.environment.defaultTimeoutMs,maxOutputBytes:25e7,maximumSpendUsd:null},approval:null,createdAt:R(),metadata:{sourceTasksetId:e.taskset.id,sourceTasksetRevision:e.taskset.revision,reasoningEffort:e.reasoningEffort??null}});return{agentSnapshot:a,harnessRelease:s,environmentRelease:i,verifierSetRelease:o,tasksetRelease:l,runManifest:f}}c(pe,"compileDesktopHarnessContext");function Q(e,t){let a=e.profileRelease?{id:e.profileRelease.id,contentHash:k(e.profileRelease.contentHash)}:t?.git?.head?{id:`source-${h(t.git.head)}`,contentHash:k(t.git.head)}:null;return W({sourceRelease:a,profileHead:t?.git?.head??null,skills:se(t),agents:re(t)})}c(Q,"compileTemporaryProfileHarness");function ee(e){let t=e.artifacts.map(T),a=e.artifacts.find(n=>n.kind==="environment_state")??e.artifacts.find(n=>n.kind==="runtime_trace")??e.artifacts.find(n=>n.kind==="raw_model_response"),s=J(e.attempt,e.grade),d=e.grade.components.map(n=>({id:`grader-${h(n.graderId)}-${e.grade.id}`,contentHash:u(n),mediaType:"application/vnd.openpond.grader-evidence+json",sizeBytes:Buffer.byteLength(C(n))}));return _.parse($({schemaVersion:"openpond.attemptReceipt.v1",id:`receipt-${u([e.manifest.contentHash,e.attempt.id,e.grade.id]).slice(0,24)}`,runManifest:{id:e.manifest.id,contentHash:e.manifest.contentHash},taskId:e.attempt.taskId,seed:String(e.attempt.seed),terminal:!["infrastructure_failure","timeout","cancelled"].includes(s??""),failureClass:s,outputHash:e.attempt.infrastructureError?null:u(e.attempt.output),traceHash:a?k(a.sha256):u({attemptId:e.attempt.id,output:e.attempt.output,runtimeEventRefs:e.attempt.runtimeEventRefs,artifactRefs:e.attempt.artifactRefs,metadata:e.attempt.metadata}),artifactRefs:t,graderEvidenceRefs:d,startedAt:e.attempt.startedAt,completedAt:e.attempt.completedAt,latencyMs:e.attempt.latencyMs,costUsd:e.attempt.costUsd,legacyAttemptRef:e.attempt.id,metadata:{legacyGradeRef:e.grade.id,rewardEligible:e.grade.rewardEligible,score:e.grade.score,passed:e.grade.passed,usage:e.attempt.metadata.usage??[]}}))}c(ee,"projectDesktopAttemptReceipt");function ue(e){let t=ee({manifest:e.context.runManifest,attempt:e.attempt,grade:e.grade,artifacts:e.artifacts}),a=e.context.tasksetRelease.tasks.find(r=>r.id===e.attempt.taskId);if(!a)throw new Error(`Task ${e.attempt.taskId} is absent from the admitted Taskset Release.`);let s=a.requiredOutputs??[],d=e.artifacts.filter(r=>r.kind==="output_artifact").map(r=>({path:typeof r.metadata.requiredOutputPath=="string"?r.metadata.requiredOutputPath:r.path,artifact:T(r),detectedMediaType:r.mediaType??null,status:"collected",parseStatus:X(r.metadata.parseStatus),schemaStatus:X(r.metadata.schemaStatus),evidenceRefs:[T(r)],metadata:{sourceArtifactId:r.id}})),n=F({id:`artifact-manifest-${t.id}`,attemptRef:{id:t.id,contentHash:t.contentHash},requiredOutputs:s,collectedArtifacts:d,createdAt:e.attempt.completedAt,metadata:{legacyAttemptRef:e.attempt.id}}),i=e.grade.components.map((r,O)=>te({component:r,grader:e.context.verifierSetRelease.graders.find(g=>g.id===r.graderId),evidenceRef:t.graderEvidenceRefs[O]??null,gradeScored:e.grade.score!==null,boundResult:e.grade.rewardComposition?.results.find(g=>g.graderId===r.graderId)})),o=L({requiredOutputs:s,manifest:n}),l=ae(e.attempt,e.grade),R=o.some(r=>r.status==="unscorable"&&r.failureOwner==="collector"),f=o.some(r=>r.status==="scored"&&!r.passed),p=l.outcomeClass==="completed"&&R?{outcomeClass:"collector_failure",failureOwner:"collector"}:l.outcomeClass==="completed"&&f?{outcomeClass:"incomplete_output",failureOwner:"policy"}:l,m=D({id:`reward-${u([t.contentHash,e.context.verifierSetRelease.contentHash,n.contentHash,e.grade.id]).slice(0,24)}`,attemptRef:{id:t.id,contentHash:t.contentHash},verifierSet:e.context.verifierSetRelease,artifactManifest:n,...p,components:[...i,...o],createdAt:e.grade.createdAt,metadata:{legacyGradeRef:e.grade.id,legacyAttemptRef:e.attempt.id,...e.grade.rewardComposition?{rewardComposition:e.grade.rewardComposition}:{}}});return{attemptReceipt:t,artifactManifest:n,rewardReceipt:m}}c(ue,"projectDesktopCanonicalReceipts");function Re(e){let t=e.artifacts.find(n=>n.kind==="runtime_trace")??e.artifacts.find(n=>n.kind==="raw_model_response"),a=t?T(t):{id:`trace-${e.canonical.attemptReceipt.id}`,contentHash:e.canonical.attemptReceipt.traceHash,mediaType:"application/vnd.openpond.attempt-trace+json",sizeBytes:null},s=e.canonical.rewardReceipt.outcomeClass==="task_deadline"||e.canonical.rewardReceipt.outcomeClass==="infrastructure_timeout",d=e.canonical.rewardReceipt.outcomeClass==="cancelled";return j({id:`rollout-${u([e.canonical.attemptReceipt.contentHash,e.canonical.rewardReceipt.contentHash,a.contentHash]).slice(0,24)}`,attemptReceipt:e.canonical.attemptReceipt,rewardReceipt:e.canonical.rewardReceipt,artifactManifestRef:{id:e.canonical.artifactManifest.id,contentHash:e.canonical.artifactManifest.contentHash},tasksetRelease:{id:e.context.tasksetRelease.id,contentHash:e.context.tasksetRelease.contentHash},environmentRelease:{id:e.context.environmentRelease.id,contentHash:e.context.environmentRelease.contentHash},harnessRelease:{id:e.context.harnessRelease.id,contentHash:e.context.harnessRelease.contentHash},taskId:e.attempt.taskId,split:e.attempt.split,model:e.context.runManifest.model,seed:String(e.attempt.seed),traceRef:a,optimizerSample:e.optimizerSample??null,environmentExecutions:[{id:`environment-execution-${e.attempt.id}`,environmentRelease:{id:e.context.environmentRelease.id,contentHash:e.context.environmentRelease.contentHash},status:d?"cancelled":s?"timed_out":e.canonical.rewardReceipt.status==="scored"?"completed":"failed",startedAt:e.attempt.startedAt,completedAt:e.attempt.completedAt,traceRefs:[a],metadata:{legacyAttemptRef:e.attempt.id,runtimeEventCount:e.attempt.runtimeEventRefs.length}}],startedAt:e.attempt.startedAt,completedAt:e.attempt.completedAt,metadata:{source:"openpond.desktop-local-runtime",optimizerSamplePresent:!!e.optimizerSample}})}c(Re,"projectDesktopCanonicalRollout");function te(e){if(!e.grader)throw new Error(`Grade component ${e.component.graderId} has no admitted Verifier definition.`);let t=(e.boundResult?e.boundResult.status==="scored":e.gradeScored)?"scored":"unscorable",a=t==="scored"&&e.component.rewardEligible,s=e.evidenceRef?[e.evidenceRef]:[];return{verifierId:e.component.graderId,verifierVersion:e.component.graderVersion,status:t,rawScore:t==="scored"?e.boundResult?.rawScore??e.component.score:null,normalizedScore:t==="scored"?e.boundResult?.normalizedScore??e.component.score:null,weight:e.grader.weight,passed:e.component.passed,hardGate:e.component.hardGate,rewardEligible:a,rewardContribution:a?e.component.score:null,failureOwner:e.component.passed?null:t==="scored"?"policy":"verifier",feedback:e.component.feedback?[e.component.feedback]:[],visibleEvidenceRefs:e.grader.privileged?[]:s,privilegedEvidenceRefs:e.grader.privileged?s:[],metadata:{judge:e.component.judge??null,calibrationStatus:e.component.calibrationStatus,legacyEvidenceRefs:e.component.evidenceRefs}}}c(te,"portableRewardComponent");function ae(e,t){let a=B.safeParse(e.metadata.outcomeClass);if(a.success){let d=e.metadata.failureOwner===null?null:z.safeParse(e.metadata.failureOwner);if(d!==null&&!d.success)throw new Error("Attempt has an invalid explicit failure owner.");return{outcomeClass:a.data,failureOwner:d===null?null:d.data}}let s=J(e,t);return s?G({failureClass:s,timeoutKind:e.metadata.timeoutKind==="task_deadline"?"task_deadline":"infrastructure_timeout"}):t.score===null?{outcomeClass:"verifier_failure",failureOwner:"verifier"}:t.passed?{outcomeClass:"completed",failureOwner:null}:{outcomeClass:"policy_failure",failureOwner:"policy"}}c(ae,"canonicalOutcome");function X(e){return e==="passed"||e==="failed"?e:"not_requested"}c(X,"artifactValidationStatus");function se(e){return(e?.skills??[]).filter(t=>t.enabled).map(t=>({id:`skill-${h(t.name)}`,path:`skills/${h(t.name)}/SKILL.md`,contentHash:k(t.sourceHash),sizeBytes:Math.max(0,t.charCount),mediaType:"text/markdown",visibility:"policy"}))}c(se,"portableSkills");function re(e){let t=new Map;for(let a of e?.agents??[])a.enabled&&t.set(a.id,u({name:a.name,path:a.path}));return[...t].map(([a,s])=>({id:`agent-${h(a)}`,path:`agents/${h(a)}/release.json`,contentHash:k(s),sizeBytes:0,mediaType:"application/vnd.openpond.agent-release+json",visibility:"policy"}))}c(re,"portableAgents");function A(e){let t=e.environment.metadata.runtimeAdapterId;if(typeof t=="string"&&t.trim())return t.trim();let a=e.environment.metadata.benchmark;if(a&&typeof a=="object"&&!Array.isArray(a)){let s=Reflect.get(a,"id");if(typeof s=="string"&&s.trim())return`conformance:${s.trim()}`}return e.environment.kind==="work"?"openpond:desktop-work":e.environment.kind==="chat"?"openpond:desktop-text":`openpond:${h(e.environment.entrypoint)}`}c(A,"desktopTasksetRuntimeAdapterId");function T(e){return{id:e.id,contentHash:k(e.sha256),mediaType:e.mediaType??null,sizeBytes:e.sizeBytes}}c(T,"portableArtifact");function J(e,t){if(e.infrastructureError)return"infrastructure_failure";if(t.failureClass)return t.failureClass;let a=e.metadata.failureClass;return a==="policy_failure"||a==="grader_failure"||a==="environment_failure"||a==="infrastructure_failure"||a==="timeout"||a==="cancelled"?a:null}c(J,"attemptFailureClass");function k(e){return/^[a-f0-9]{64}$/.test(e)?e:u(e)}c(k,"hash");function h(e){return e.trim().replace(/[^a-zA-Z0-9._-]+/g,"-").replace(/^-+|-+$/g,"")||"asset"}c(h,"segment");async function be(e,t){if(e.metadata.taskDefinition!==void 0||e.metadata.rewardBinding!==void 0)return{taskset:e,generatedFiles:[]};let a=await oe(t);if(!a.isDirectory()||a.isSymbolicLink())throw new Error("Taskset source must be a regular directory.");let s=new Map,d=0;async function n(i=""){for(let o of await ie(S.join(t,i),{withFileTypes:!0})){let l=S.posix.join(i,o.name);if(o.isSymbolicLink())throw new Error("Taskset files cannot contain symbolic links.");if(o.name.startsWith(".env"))throw new Error("Environment files cannot be included in a Taskset package.");if(o.isDirectory()){await n(l);continue}if(!o.isFile())throw new Error("Taskset packages require regular files.");if(U(l))continue;if(s.size>=1e4)throw new Error("Taskset source exceeds the file inventory limit.");let R=await ne(S.join(t,l),b.O_RDONLY|b.O_NOFOLLOW|b.O_NONBLOCK);try{let f=await R.stat();if(d+=f.size,!f.isFile()||d>w*3/4)throw new Error("Taskset source exceeds the package byte limit.");let p=Buffer.alloc(f.size),m=0;for(;m<p.length;){let r=await R.read(p,m,p.length-m,m);if(!r.bytesRead)throw new Error("Taskset source changed during capture.");m+=r.bytesRead}if((await R.read(Buffer.alloc(1),0,1,m)).bytesRead)throw new Error("Taskset source changed during capture.");s.set(l,p)}finally{await R.close()}}}return c(n,"visit"),await n(),Y(e,s,A(e))}c(be,"prepareAuthoredTasksetFiles");import{constants as H}from"node:fs";import{open as ce,realpath as Z}from"node:fs/promises";import y from"node:path";function Me(e,t){return t?new Set:new Set([...(e.modelResources?.assets??[]).map(a=>a.asset.id),...(e.learningResources?.assets??[]).map(a=>a.asset.id),...e.learningResources?v(e.learningResources).map(a=>a.asset.id):[]])}c(Me,"tasksetPackageInlineAssetIds");async function $e(e){let t=e.sources.length?await Z(e.root):y.resolve(e.root),a=[],s=new Set,d=0,n=c((i,o)=>{if(s.has(i.id))throw new Error(`Duplicate local Taskset asset: ${i.id}.`);if(s.add(i.id),d+=4*Math.ceil(o.byteLength/3),d>w)throw new Error("Local Taskset files exceed the package envelope.");a.push({asset:i,base64:Buffer.from(o).toString("base64")})},"append");for(let i of e.sources){let o=x.parse(i.asset);if(!i.sourcePath||y.isAbsolute(i.sourcePath))throw new Error(`Taskset asset ${o.id} requires a relative source path.`);let l=y.resolve(t,i.sourcePath);if(!l.startsWith(`${t}${y.sep}`))throw new Error(`Taskset asset ${o.id} escapes its local package.`);if(!(await Z(l)).startsWith(`${t}${y.sep}`))throw new Error(`Taskset asset ${o.id} escapes its local package.`);let f=await ce(l,H.O_RDONLY|H.O_NOFOLLOW|H.O_NONBLOCK);try{let p=await f.stat();if(!p.isFile()||p.size!==o.sizeBytes)throw new Error(`Taskset asset ${o.id} differs from its immutable file size.`);if(4*Math.ceil(p.size/3)+d>w)throw new Error("Local Taskset files exceed the package envelope.");let m=Buffer.alloc(p.size),r=0;for(;r<m.length;){let g=await f.read(m,r,m.length-r,r);if(!g.bytesRead)throw new Error(`Taskset asset ${o.id} changed during capture.`);r+=g.bytesRead}if((await f.read(Buffer.alloc(1),0,1,r)).bytesRead)throw new Error(`Taskset asset ${o.id} changed during capture.`);n(o,m)}finally{await f.close()}}for(let i of[...e.content.modelResources?.assets??[],...e.content.learningResources?.assets??[]])s.has(i.asset.id)||n(i.asset,Buffer.from(P(i,i.asset),"utf8"));if(e.content.learningResources)for(let i of v(e.content.learningResources))s.has(i.asset.id)||n(i.asset,Buffer.from(i.base64,"base64"));if(e.fileOrder){if(new Set(e.fileOrder).size!==a.length||e.fileOrder.length!==a.length||e.fileOrder.some(o=>!s.has(o)))throw new Error("Saved package file order differs from its inventory.");let i=new Map(e.fileOrder.map((o,l)=>[o,l]));a.sort((o,l)=>i.get(o.asset.id)-i.get(l.asset.id))}return q({...e.content,files:a})}c($e,"captureLocalTasksetPackage");export{pe as a,ue as b,Re as c,A as d,be as e,Me as f,$e as g};

package/dist/chunks/{chunk-RCQYEHHD.js → chunk-O5ZOW474.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{c as D}from"./chunk-IR3YJWPW.js";import{Wb as G}from"./chunk-HG2YFMVU.js";import{b as $,c as z,d as H,e as V,f as R,g as U}from"./chunk-I6VMRQF4.js";import{Q as P,S as Q}from"./chunk-4PUOWSDE.js";import{fb as q,v,x as T}from"./chunk-ZY32SZ2P.js";import{n as E}from"./chunk-RCT3O3HQ.js";import{A as n}from"./chunk-X2J2P5ZT.js";import{a}from"./chunk-FTQLQJQV.js";async function K(e){let t=F(await e),[o,c,{QuickJSWASMModule:w}]=await Promise.all([t.importModuleLoader().then(F),t.importFFI(),import("./module-ES6BEMUI-ZVL3F6NW.js").then(F)]),l=await o();l.type="sync";let s=new c(l);return new w(l,s)}a(K,"newQuickJSWASMModuleFromVariant");function F(e){return e&&"default"in e&&e.default?e.default&&"default"in e.default&&e.default.default?e.default.default:e.default:e}a(F,"smartUnwrap");var k=class extends Error{static{a(this,"JavaScriptIsolateExecutionError")}constructor(t){super(t),this.name="JavaScriptIsolateExecutionError"}};async function O(e,t){te(e),e.signal?.throwIfAborted();let o=a(s=>new Error(`${e.errorPrefix}_${s}`),"error"),c=Date.now()+e.timeoutMs,{default:w}=await import("./dist-SKYG5LZO.js");e.signal?.throwIfAborted();let l=await K(w);e.signal?.throwIfAborted();try{return D.withScope(s=>{let m=s.manage(l.newRuntime());m.setMemoryLimit(33554432),m.setMaxStackSize(262144),m.setInterruptHandler(()=>Date.now()>=c||e.signal?.aborted===!0);let i=s.manage(m.newContext());t&&(c=Date.now()+e.timeoutMs,t.onExecutionReady()),e.deterministic&&s.manage(i.unwrapResult(i.evalCode("globalThis.Date = undefined; Math.random = () => { throw new Error('environment_random_requires_explicit_seed'); };")));let r=s.manage(i.unwrapResult(i.evalCode("JSON.stringify"))),p=s.manage(i.unwrapResult(i.evalCode(`JSON.parse(${JSON.stringify(JSON.stringify(e.value))})`))),S=s.manage(i.unwrapResult(i.evalCode(e.source,`${e.errorPrefix}.mjs`,{type:"module"}))),M=s.manage(X(i,S,c,e.signal,o)),g=s.manage(i.getProp(M,e.exportName));if(i.typeof(g)!=="function")throw o("function_export_missing");let b=s.manage(i.unwrapResult(i.callFunction(g,i.undefined,p))),d=s.manage(X(i,b,c,e.signal,o)),h=s.manage(i.unwrapResult(i.callFunction(r,i.undefined,d)));if(i.typeof(h)!=="string")throw o("result_not_json");let y=i.getString(h);if(new TextEncoder().encode(y).byteLength>e.maxResultBytes)throw o("result_too_large");if(e.signal?.throwIfAborted(),Date.now()>=c)throw o("timeout");let f=JSON.parse(y);return v(f,e.maxResultBytes),f})}catch(s){throw e.signal?.throwIfAborted(),new k(s instanceof Error?s.message:"JavaScript execution failed.")}}a(O,"executeJavaScriptIsolate");function te(e){if(!Number.isInteger(e.timeoutMs)||e.timeoutMs<1||e.timeoutMs>3e5)throw new Error(`${e.errorPrefix}_timeout_invalid`);if(!Number.isInteger(e.maxResultBytes)||e.maxResultBytes<1||e.maxResultBytes>4194304)throw new Error(`${e.errorPrefix}_result_limit_invalid`);if(new TextEncoder().encode(e.source).byteLength>524288)throw new Error(`${e.errorPrefix}_source_too_large`);v(e.value,4194304)}a(te,"validateJavaScriptIsolateInput");function X(e,t,o,c,w){for(;;){if(c?.throwIfAborted(),Date.now()>=o)throw w("timeout");let l=e.getPromiseState(t);if(l.type==="fulfilled")return l.notAPromise?l.value.dup():l.value;if(l.type==="rejected")return e.unwrapResult(l);if(e.runtime.executePendingJobs(1).unwrap()===0)throw w("unresolved_promise")}}a(X,"settle");function I(e,t){let o=t.scope==="arguments"?e.action?.arguments:e[t.scope];for(let c of t.path){if(!o||typeof o!="object"||!Object.hasOwn(o,c))throw new Error("environment_service_reference_missing");o=o[c]}return o}a(I,"resolveValue");async function Y(e,t){if(!e.executionServices?.length)return t.controller(e);if(!Number.isSafeInteger(e.timeoutMs)||e.timeoutMs<1||e.timeoutMs>3e4)throw new Error("environment_timeout_invalid");v(e.value,4194304);let o=e.executionServices.map(m=>z.parse(m));if(o.length>16)throw new Error("environment_service_limit");if(new Set(o.map(m=>m.id)).size!==o.length)throw new Error("environment_service_identity_collision");let c=Date.now(),w=new AbortController,l=e.signal?AbortSignal.any([e.signal,w.signal]):w.signal,s=setTimeout(()=>w.abort(new Error("environment_timeout")),e.timeoutMs);try{let m=Object.create(null);for(let r of o){if(l.throwIfAborted(),r.operation!==e.operation||r.operation==="step"&&r.toolName!==e.value.action?.name)throw new Error("environment_service_operation_mismatch");let p=Math.min(r.timeoutMs,e.timeoutMs-(Date.now()-c));if(p<1)throw new Error("environment_timeout");let S=new AbortController,M=AbortSignal.any([l,S.signal]),g=setTimeout(()=>S.abort(new Error("environment_service_timeout")),p),b=Date.now();try{if(r.kind==="sqlite.v1"){if(!t.sql)throw new Error("environment_sql_executor_required");let d=I(e.value,r.sql);if(typeof d!="string"||!d.length||new TextEncoder().encode(d).byteLength>65536||d.includes("\0")){m[r.id]={status:"rejected",code:"invalid_submission"};continue}let h=$({schemaVersion:"openpond.sqlExecution.v1",snapshot:I(e.value,r.snapshot),sql:d,maxRows:r.maxRows,maxResultBytes:r.maxResultBytes});m[r.id]=await t.sql({request:h,timeoutMs:e.timeoutMs,signal:M})}else{let d=I(e.value,r.source),h=I(e.value,r.cases);if(typeof d!="string"||new TextEncoder().encode(d).byteLength>524288){m[r.id]={status:"rejected",code:"invalid_submission"};continue}if(!Array.isArray(h)||!h.length||h.length>r.maxCases||h.some(f=>!f||typeof f!="object"||!Object.hasOwn(f,"input")))throw new Error("environment_service_cases_invalid");let y=[];for(let f of h){let u=p-(Date.now()-b);if(u<1&&(S.abort(new Error("environment_service_timeout")),M.throwIfAborted()),y.push(await t.candidate({source:d,exportName:r.exportName,value:structuredClone(f.input),timeoutMs:Math.max(1,u),maxResultBytes:r.maxResultBytes,deterministic:!0,errorPrefix:"candidate",signal:M})),new TextEncoder().encode(JSON.stringify({status:"completed",cases:y})).byteLength>r.maxResultBytes)throw new k("candidate_result_too_large")}m[r.id]={status:"completed",cases:y}}}catch(d){if(l.throwIfAborted(),S.signal.aborted)m[r.id]={status:"rejected",code:"timeout"};else if(d instanceof k)m[r.id]={status:"rejected",code:"execution_failed"};else throw d}finally{clearTimeout(g)}}l.throwIfAborted();let i=e.timeoutMs-(Date.now()-c);if(i<1)throw new Error("environment_timeout");return await t.controller({...e,executionServices:void 0,value:{...e.value,services:m},timeoutMs:i,signal:l})}finally{clearTimeout(s)}}a(Y,"executeEnvironmentWithServices");async function re(e){let t=V.parse(e.operation);return Y(e,{candidate:O,controller:a(async o=>R.parse(await O({...o,exportName:t,maxResultBytes:1572864,deterministic:!0,errorPrefix:"environment"})),"controller")})}a(re,"executeJavaScriptEnvironment");var J=class extends Error{static{a(this,"JavaScriptEnvironmentActionError")}code;constructor(t){super(`environment_${t}`),this.code=t,this.name="JavaScriptEnvironmentActionError"}};async function Qe(e){let t=U(e.definition),o=Q(e.asset,t.module);if(!Number.isSafeInteger(e.seed))throw new Error("environment_seed_invalid");let c=e.seed;v(e.input,1048576),v(e.initialState,t.maxStateBytes);let w=structuredClone(e.input),l=structuredClone(e.initialState),s=e.execute??re,m=new AbortController,i=e.signal?AbortSignal.any([e.signal,m.signal]):m.signal,r={},p=0,S=!1,M=!1,g=null,b=[],d=null;async function h(f,u){if(S||M)throw new Error("environment_closed");if(g)throw new Error("environment_operation_in_progress");i.throwIfAborted(),u&&(v(u,262144),u=structuredClone(u));let L=E(r),_=u?t.tools.find(x=>x.name===u.name):null;if(f==="step"){if(p>=t.maxSteps)throw new J("step_budget_exhausted");p++}let j=a((x,C)=>b.push({sequence:b.length,operation:f,toolName:u?.name??null,arguments:u?structuredClone(u.arguments):null,observation:x,beforeStateHash:L,afterStateHash:E(r),error:C}),"record");try{if(u&&!_)throw new J("unknown_tool");if(u&&_&&!T(_.inputSchema,u.arguments).valid)throw new J("invalid_arguments");let x=t.executionServices?.filter(N=>N.operation===f&&(f!=="step"||N.toolName===u?.name)),C=s({source:o,operation:f,value:{input:w,initialState:l,seed:c,state:structuredClone(r),sequence:p,action:u?structuredClone(u):null},timeoutMs:Math.min(t.operationTimeoutMs,_?.timeoutMs??t.operationTimeoutMs),signal:i,...x?.length?{executionServices:x}:{}});g=C;let A=R.parse(await C);if(i.throwIfAborted(),v(A.state,t.maxStateBytes),v(A.observation,t.maxObservationBytes),(f==="collect"||_?.sideEffect==="read")&&E(A.state)!==L)throw new Error("environment_read_changed_state");r=structuredClone(A.state);let B=structuredClone(A.observation);return j(B,null),{state:structuredClone(r),observation:B}}catch(x){throw j(null,x instanceof Error?x.message:"environment_execution_failed"),x}finally{g=null}}a(h,"invoke");async function y(){return d||(M=!0,m.abort(new Error("environment_destroyed")),d=(async()=>{try{try{await g}catch{}await s({source:o,operation:"destroy",value:{input:w,initialState:l,seed:c,state:structuredClone(r),sequence:p,action:null},timeoutMs:t.operationTimeoutMs})}finally{r={},S=!0}})(),d)}a(y,"destroy");try{await h("create");let f=await h("reset"),u=E(r),L=a(()=>{if(S)throw new Error("environment_closed");return{definition:{id:t.id,revision:t.revision,contentHash:t.contentHash},inputHash:E(w),seed:c,initialStateHash:u,finalStateHash:E(r),state:structuredClone(r),events:structuredClone(b)}},"snapshot");return{definition:structuredClone(t),observation:f.observation,async step(_){return(await h("step",_)).observation},async collect(){let _=await h("collect");return{...L(),observation:_.observation}},snapshot:L,destroy:y}}catch(f){try{await y()}catch{}throw f}}a(Qe,"createJavaScriptEnvironmentSession");var ne=n.string().trim().min(1).max(500),oe=n.string().regex(/^[a-f0-9]{64}$/),W=n.object({kind:n.literal("model"),id:n.string().trim().min(1).max(191),taskId:ne,admissionHash:oe,model:G,instructions:n.string().max(262144),harness:q.optional(),input:n.record(n.string(),n.unknown()),policyVisibleContext:n.record(n.string(),n.unknown()),timeoutMs:n.number().int().positive().max(36e5),environment:n.discriminatedUnion("kind",[n.object({kind:n.literal("text")}).strict(),n.object({kind:n.literal("javascript"),definition:H,asset:P,initialState:n.record(n.string(),n.unknown()),seed:n.number().int().safe()}).strict()])}).strict(),Ye=n.discriminatedUnion("kind",[W,n.object({kind:n.literal("profile"),request:n.unknown()}).strict()]);var Z=n.string().trim().min(1).max(200),mt=n.object({receiptId:Z,taskId:Z,seed:n.string(),request:W}).strict(),ee=class extends Error{constructor(o,c,w=409){super(c);this.code=o;this.status=w}code;status;static{a(this,"LocalExperimentError")}};export{k as a,te as b,Y as c,J as d,Qe as e,W as f,Ye as g,ee as h};
+import{c as D}from"./chunk-IR3YJWPW.js";import{Wb as G}from"./chunk-WXG3G2O6.js";import{b as $,c as z,d as H,e as V,f as R,g as U}from"./chunk-I6VMRQF4.js";import{Q as P,S as Q}from"./chunk-ZEOYNUHL.js";import{fb as q,v,x as T}from"./chunk-ZY32SZ2P.js";import{n as E}from"./chunk-RCT3O3HQ.js";import{A as n}from"./chunk-X2J2P5ZT.js";import{a}from"./chunk-FTQLQJQV.js";async function K(e){let t=F(await e),[o,c,{QuickJSWASMModule:w}]=await Promise.all([t.importModuleLoader().then(F),t.importFFI(),import("./module-ES6BEMUI-ZVL3F6NW.js").then(F)]),l=await o();l.type="sync";let s=new c(l);return new w(l,s)}a(K,"newQuickJSWASMModuleFromVariant");function F(e){return e&&"default"in e&&e.default?e.default&&"default"in e.default&&e.default.default?e.default.default:e.default:e}a(F,"smartUnwrap");var k=class extends Error{static{a(this,"JavaScriptIsolateExecutionError")}constructor(t){super(t),this.name="JavaScriptIsolateExecutionError"}};async function O(e,t){te(e),e.signal?.throwIfAborted();let o=a(s=>new Error(`${e.errorPrefix}_${s}`),"error"),c=Date.now()+e.timeoutMs,{default:w}=await import("./dist-SKYG5LZO.js");e.signal?.throwIfAborted();let l=await K(w);e.signal?.throwIfAborted();try{return D.withScope(s=>{let m=s.manage(l.newRuntime());m.setMemoryLimit(33554432),m.setMaxStackSize(262144),m.setInterruptHandler(()=>Date.now()>=c||e.signal?.aborted===!0);let i=s.manage(m.newContext());t&&(c=Date.now()+e.timeoutMs,t.onExecutionReady()),e.deterministic&&s.manage(i.unwrapResult(i.evalCode("globalThis.Date = undefined; Math.random = () => { throw new Error('environment_random_requires_explicit_seed'); };")));let r=s.manage(i.unwrapResult(i.evalCode("JSON.stringify"))),p=s.manage(i.unwrapResult(i.evalCode(`JSON.parse(${JSON.stringify(JSON.stringify(e.value))})`))),S=s.manage(i.unwrapResult(i.evalCode(e.source,`${e.errorPrefix}.mjs`,{type:"module"}))),M=s.manage(X(i,S,c,e.signal,o)),g=s.manage(i.getProp(M,e.exportName));if(i.typeof(g)!=="function")throw o("function_export_missing");let b=s.manage(i.unwrapResult(i.callFunction(g,i.undefined,p))),d=s.manage(X(i,b,c,e.signal,o)),h=s.manage(i.unwrapResult(i.callFunction(r,i.undefined,d)));if(i.typeof(h)!=="string")throw o("result_not_json");let y=i.getString(h);if(new TextEncoder().encode(y).byteLength>e.maxResultBytes)throw o("result_too_large");if(e.signal?.throwIfAborted(),Date.now()>=c)throw o("timeout");let f=JSON.parse(y);return v(f,e.maxResultBytes),f})}catch(s){throw e.signal?.throwIfAborted(),new k(s instanceof Error?s.message:"JavaScript execution failed.")}}a(O,"executeJavaScriptIsolate");function te(e){if(!Number.isInteger(e.timeoutMs)||e.timeoutMs<1||e.timeoutMs>3e5)throw new Error(`${e.errorPrefix}_timeout_invalid`);if(!Number.isInteger(e.maxResultBytes)||e.maxResultBytes<1||e.maxResultBytes>4194304)throw new Error(`${e.errorPrefix}_result_limit_invalid`);if(new TextEncoder().encode(e.source).byteLength>524288)throw new Error(`${e.errorPrefix}_source_too_large`);v(e.value,4194304)}a(te,"validateJavaScriptIsolateInput");function X(e,t,o,c,w){for(;;){if(c?.throwIfAborted(),Date.now()>=o)throw w("timeout");let l=e.getPromiseState(t);if(l.type==="fulfilled")return l.notAPromise?l.value.dup():l.value;if(l.type==="rejected")return e.unwrapResult(l);if(e.runtime.executePendingJobs(1).unwrap()===0)throw w("unresolved_promise")}}a(X,"settle");function I(e,t){let o=t.scope==="arguments"?e.action?.arguments:e[t.scope];for(let c of t.path){if(!o||typeof o!="object"||!Object.hasOwn(o,c))throw new Error("environment_service_reference_missing");o=o[c]}return o}a(I,"resolveValue");async function Y(e,t){if(!e.executionServices?.length)return t.controller(e);if(!Number.isSafeInteger(e.timeoutMs)||e.timeoutMs<1||e.timeoutMs>3e4)throw new Error("environment_timeout_invalid");v(e.value,4194304);let o=e.executionServices.map(m=>z.parse(m));if(o.length>16)throw new Error("environment_service_limit");if(new Set(o.map(m=>m.id)).size!==o.length)throw new Error("environment_service_identity_collision");let c=Date.now(),w=new AbortController,l=e.signal?AbortSignal.any([e.signal,w.signal]):w.signal,s=setTimeout(()=>w.abort(new Error("environment_timeout")),e.timeoutMs);try{let m=Object.create(null);for(let r of o){if(l.throwIfAborted(),r.operation!==e.operation||r.operation==="step"&&r.toolName!==e.value.action?.name)throw new Error("environment_service_operation_mismatch");let p=Math.min(r.timeoutMs,e.timeoutMs-(Date.now()-c));if(p<1)throw new Error("environment_timeout");let S=new AbortController,M=AbortSignal.any([l,S.signal]),g=setTimeout(()=>S.abort(new Error("environment_service_timeout")),p),b=Date.now();try{if(r.kind==="sqlite.v1"){if(!t.sql)throw new Error("environment_sql_executor_required");let d=I(e.value,r.sql);if(typeof d!="string"||!d.length||new TextEncoder().encode(d).byteLength>65536||d.includes("\0")){m[r.id]={status:"rejected",code:"invalid_submission"};continue}let h=$({schemaVersion:"openpond.sqlExecution.v1",snapshot:I(e.value,r.snapshot),sql:d,maxRows:r.maxRows,maxResultBytes:r.maxResultBytes});m[r.id]=await t.sql({request:h,timeoutMs:e.timeoutMs,signal:M})}else{let d=I(e.value,r.source),h=I(e.value,r.cases);if(typeof d!="string"||new TextEncoder().encode(d).byteLength>524288){m[r.id]={status:"rejected",code:"invalid_submission"};continue}if(!Array.isArray(h)||!h.length||h.length>r.maxCases||h.some(f=>!f||typeof f!="object"||!Object.hasOwn(f,"input")))throw new Error("environment_service_cases_invalid");let y=[];for(let f of h){let u=p-(Date.now()-b);if(u<1&&(S.abort(new Error("environment_service_timeout")),M.throwIfAborted()),y.push(await t.candidate({source:d,exportName:r.exportName,value:structuredClone(f.input),timeoutMs:Math.max(1,u),maxResultBytes:r.maxResultBytes,deterministic:!0,errorPrefix:"candidate",signal:M})),new TextEncoder().encode(JSON.stringify({status:"completed",cases:y})).byteLength>r.maxResultBytes)throw new k("candidate_result_too_large")}m[r.id]={status:"completed",cases:y}}}catch(d){if(l.throwIfAborted(),S.signal.aborted)m[r.id]={status:"rejected",code:"timeout"};else if(d instanceof k)m[r.id]={status:"rejected",code:"execution_failed"};else throw d}finally{clearTimeout(g)}}l.throwIfAborted();let i=e.timeoutMs-(Date.now()-c);if(i<1)throw new Error("environment_timeout");return await t.controller({...e,executionServices:void 0,value:{...e.value,services:m},timeoutMs:i,signal:l})}finally{clearTimeout(s)}}a(Y,"executeEnvironmentWithServices");async function re(e){let t=V.parse(e.operation);return Y(e,{candidate:O,controller:a(async o=>R.parse(await O({...o,exportName:t,maxResultBytes:1572864,deterministic:!0,errorPrefix:"environment"})),"controller")})}a(re,"executeJavaScriptEnvironment");var J=class extends Error{static{a(this,"JavaScriptEnvironmentActionError")}code;constructor(t){super(`environment_${t}`),this.code=t,this.name="JavaScriptEnvironmentActionError"}};async function Qe(e){let t=U(e.definition),o=Q(e.asset,t.module);if(!Number.isSafeInteger(e.seed))throw new Error("environment_seed_invalid");let c=e.seed;v(e.input,1048576),v(e.initialState,t.maxStateBytes);let w=structuredClone(e.input),l=structuredClone(e.initialState),s=e.execute??re,m=new AbortController,i=e.signal?AbortSignal.any([e.signal,m.signal]):m.signal,r={},p=0,S=!1,M=!1,g=null,b=[],d=null;async function h(f,u){if(S||M)throw new Error("environment_closed");if(g)throw new Error("environment_operation_in_progress");i.throwIfAborted(),u&&(v(u,262144),u=structuredClone(u));let L=E(r),_=u?t.tools.find(x=>x.name===u.name):null;if(f==="step"){if(p>=t.maxSteps)throw new J("step_budget_exhausted");p++}let j=a((x,C)=>b.push({sequence:b.length,operation:f,toolName:u?.name??null,arguments:u?structuredClone(u.arguments):null,observation:x,beforeStateHash:L,afterStateHash:E(r),error:C}),"record");try{if(u&&!_)throw new J("unknown_tool");if(u&&_&&!T(_.inputSchema,u.arguments).valid)throw new J("invalid_arguments");let x=t.executionServices?.filter(N=>N.operation===f&&(f!=="step"||N.toolName===u?.name)),C=s({source:o,operation:f,value:{input:w,initialState:l,seed:c,state:structuredClone(r),sequence:p,action:u?structuredClone(u):null},timeoutMs:Math.min(t.operationTimeoutMs,_?.timeoutMs??t.operationTimeoutMs),signal:i,...x?.length?{executionServices:x}:{}});g=C;let A=R.parse(await C);if(i.throwIfAborted(),v(A.state,t.maxStateBytes),v(A.observation,t.maxObservationBytes),(f==="collect"||_?.sideEffect==="read")&&E(A.state)!==L)throw new Error("environment_read_changed_state");r=structuredClone(A.state);let B=structuredClone(A.observation);return j(B,null),{state:structuredClone(r),observation:B}}catch(x){throw j(null,x instanceof Error?x.message:"environment_execution_failed"),x}finally{g=null}}a(h,"invoke");async function y(){return d||(M=!0,m.abort(new Error("environment_destroyed")),d=(async()=>{try{try{await g}catch{}await s({source:o,operation:"destroy",value:{input:w,initialState:l,seed:c,state:structuredClone(r),sequence:p,action:null},timeoutMs:t.operationTimeoutMs})}finally{r={},S=!0}})(),d)}a(y,"destroy");try{await h("create");let f=await h("reset"),u=E(r),L=a(()=>{if(S)throw new Error("environment_closed");return{definition:{id:t.id,revision:t.revision,contentHash:t.contentHash},inputHash:E(w),seed:c,initialStateHash:u,finalStateHash:E(r),state:structuredClone(r),events:structuredClone(b)}},"snapshot");return{definition:structuredClone(t),observation:f.observation,async step(_){return(await h("step",_)).observation},async collect(){let _=await h("collect");return{...L(),observation:_.observation}},snapshot:L,destroy:y}}catch(f){try{await y()}catch{}throw f}}a(Qe,"createJavaScriptEnvironmentSession");var ne=n.string().trim().min(1).max(500),oe=n.string().regex(/^[a-f0-9]{64}$/),W=n.object({kind:n.literal("model"),id:n.string().trim().min(1).max(191),taskId:ne,admissionHash:oe,model:G,instructions:n.string().max(262144),harness:q.optional(),input:n.record(n.string(),n.unknown()),policyVisibleContext:n.record(n.string(),n.unknown()),timeoutMs:n.number().int().positive().max(36e5),environment:n.discriminatedUnion("kind",[n.object({kind:n.literal("text")}).strict(),n.object({kind:n.literal("javascript"),definition:H,asset:P,initialState:n.record(n.string(),n.unknown()),seed:n.number().int().safe()}).strict()])}).strict(),Ye=n.discriminatedUnion("kind",[W,n.object({kind:n.literal("profile"),request:n.unknown()}).strict()]);var Z=n.string().trim().min(1).max(200),mt=n.object({receiptId:Z,taskId:Z,seed:n.string(),request:W}).strict(),ee=class extends Error{constructor(o,c,w=409){super(c);this.code=o;this.status=w}code;status;static{a(this,"LocalExperimentError")}};export{k as a,te as b,Y as c,J as d,Qe as e,W as f,Ye as g,ee as h};

package/dist/chunks/{chunk-IIT66347.js → chunk-OEAKX5TS.js}

@@ -1,5 +1,5 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as As,b as Ms}from"./chunk-CO23RRN7.js";import{a as Es,e as $e,f as Cs,g as Tt}from"./chunk-3Z4VM2LF.js";import{a as xs,c as Is}from"./chunk-XVDFJ3ZF.js";import{h as ne}from"./chunk-RCQYEHHD.js";import{a as ee,b as _s}from"./chunk-SDHG6CKU.js";import{b as ys,c as bs,d as yt}from"./chunk-GYYJ47GX.js";import{a as ws,b as vs,c as Ts,d as qe,e as Ss,g as vt}from"./chunk-WCNPGKST.js";import{a as wt,c as Rs}from"./chunk-AFH462DO.js";import{b as Kr,e as Xr,f as lt,g as ut}from"./chunk-CIZI6Z7D.js";import{F as gt,G as fs,H as ht,I as kt,J as gs,K as hs,N as ks,x as ps}from"./chunk-MOXIGLN2.js";import{a as ve}from"./chunk-7KWD2I6W.js";import{g as pt,i as ft}from"./chunk-RY46UPBO.js";import{s as br}from"./chunk-OFJPOTES.js";import{d as Ne}from"./chunk-U7D6Q744.js";import{ja as rs,qa as ss,sa as os,x as ts}from"./chunk-HG2YFMVU.js";import{M as V,N as as,O as ns,P as is,Q as mt,R as ds,S as q,T as cs,U as je,V as ls,X as us,Y as Vt,Z as ms}from"./chunk-7OKZCJHS.js";import{l as Ft}from"./chunk-7NCM6GBN.js";import{d as Zr,f as W,g as Qr,i as es}from"./chunk-JT5YQ52I.js";import{k as Nr}from"./chunk-WGQ3FMRF.js";import{c as ct}from"./chunk-JVMSN7LH.js";import{Ba as Gr,Ca as Re,Ha as Yr,Ia as De,K as qr,L as $r,M as Lr,N as Ur,P as Br,aa as Wr,ba as Fr,ia as Vr,ua as zr,wa as we,xa as Jr}from"./chunk-BU6MW65V.js";import{x as ae}from"./chunk-DXJC3QW7.js";import{C as Or,S as Hr,T as Dr,Tb as jr}from"./chunk-D6L6UNUQ.js";import{f as Rr}from"./chunk-AJXWJLZA.js";import{L,M as Ar,Ob as Pr,R as Wt,S as oe,X as dt,Y as Mr,ba as He,d as Er,g as Cr,pb as ke,qb as ye}from"./chunk-4PUOWSDE.js";import{bb as be,cb as Ir,ea as it,v as Ge}from"./chunk-ZY32SZ2P.js";import{l as _r,m as xr}from"./chunk-OTDDXEQ6.js";import{i as Tr,m as se,n as k}from"./chunk-RCT3O3HQ.js";import{m as Sr}from"./chunk-UGOK7VTC.js";import{A as P}from"./chunk-X2J2P5ZT.js";import{a as d}from"./chunk-FTQLQJQV.js";var zt={buildIntent:"verifiable_reward",methodHint:"grpo"};async function ii(e){let t=await aa(e.store,e.workspaceId,e.reviewRef),r=`harness-review-${k({review:t.contentHash,strategy:zt}).slice(0,24)}`,s=(await e.store.listTaskCreationSnapshots(e.profileId)).find(c=>c.request.candidateId===r);if(s)return s;let[a,o,n]=await Promise.all([e.store.listHarnessImprovementArtifacts(e.workspaceId,"route_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"trigger_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"observation",1e3)]),i=new Map(a.map(c=>[_e(c),c])),u=new Map(o.map(c=>[_e(c),c])),f=new Map(n.map(c=>[_e(c),c])),p=new Map;for(let c of t.selectedEvidence){if(c.kind!=="route_decision")continue;let m=i.get(_e(c.evidence)),g=m?u.get(_e(m.trigger)):null;if(!m||!g)throw new Error(`Harness review evidence ${c.evidence.id} cannot be reconstructed.`);let h=g.observations.map(y=>f.get(_e(y))??null).filter(y=>y?.runRef===c.sourceRef);if(!h.length)throw new Error(`Harness review source ${c.sourceRef} has no exact observation evidence.`);let w=p.get(c.sourceRef)??{turnIds:new Set,evidenceRefs:[],sourcePolicies:[]};for(let y of h)w.turnIds.add(y.turnId);w.evidenceRefs.push(c.evidence),w.sourcePolicies.push(c.sourcePolicy),p.set(c.sourceRef,w)}if(!p.size)throw new Error("Harness Taskset review has no reconstructable source runs.");let l=[];for(let[c,m]of[...p.entries()].sort(([g],[h])=>g.localeCompare(h))){let g=await e.taskCreator.addSessionSource({profileId:e.profileId,sessionId:c,turnIds:[...m.turnIds].sort(),consentScope:"selected_turns"}),h=Array.isArray(g.metadata.harnessEvaluationReviews)?g.metadata.harnessEvaluationReviews:[],w={...g,metadata:{...g.metadata,harnessEvaluationReviews:Gt([...h,{id:t.id,contentHash:t.contentHash,workspaceId:t.workspaceRef,harnessRelease:t.harnessRelease,claimFingerprint:t.claim?.fingerprint??null,evidenceRefs:m.evidenceRefs,sourcePolicies:m.sourcePolicies}])}};l.push(await e.store.upsertTrainingSource(w))}return e.startCreation({profileId:e.profileId,sourceIds:l.map(c=>c.id),surface:"task_candidate",mode:"customize",entryMode:"automated",resourceIntent:"workproduct",buildIntent:zt.buildIntent,buildSpecification:null,objective:t.claim?.statement??t.reason,methodHint:zt.methodHint,preferredBaseModelId:null,preferredBaseModel:null,candidateId:r,analysisModel:e.analysisModel??null,analysisReasoningEffort:e.analysisReasoningEffort??null,createImproveRunId:null,targetIntent:{kind:"model",id:null,displayName:null,operation:"create"}})}d(ii,"startHarnessReviewTasksetAuthoring");function di(e){let t=e.flatMap(o=>Array.isArray(o.metadata.harnessEvaluationReviews)?o.metadata.harnessEvaluationReviews.filter(Jt):[]);if(!t.length)return null;let r=`${String(t[0].id)}:${String(t[0].contentHash)}`,s=t.filter(o=>`${String(o.id)}:${String(o.contentHash)}`===r);if(s.length!==e.length)return null;let a=s[0];return{review:{id:String(a.id),contentHash:String(a.contentHash),workspaceId:String(a.workspaceId),harnessRelease:a.harnessRelease,claimFingerprint:typeof a.claimFingerprint=="string"?a.claimFingerprint:null},evidenceRefs:Gt(s.flatMap(o=>Array.isArray(o.evidenceRefs)?o.evidenceRefs:[])),sourcePolicies:Gt(s.flatMap(o=>Array.isArray(o.sourcePolicies)?o.sourcePolicies:[]))}}d(di,"harnessReviewLineageFromSources");async function Ps(e){let t=e.taskset.metadata.harnessEvaluationLineage;if(!Jt(t)||!Jt(t.review))return null;let r=String(t.review.workspaceId??""),s={id:String(t.review.id??""),contentHash:String(t.review.contentHash??"")};if(!r||!s.id||!s.contentHash)throw new Error("Taskset Harness Evaluation lineage is incomplete.");let a=await e.store.listHarnessImprovementArtifacts(r,"evaluation_review",1e3),o=a.find(c=>_e(c)===_e(s));if(!o)throw new Error("Originating Harness Evaluation review was not found.");let n=e.evaluationResult?{id:e.evaluationResult.id,contentHash:e.evaluationResult.contentHash}:null,i={id:e.taskset.id,contentHash:e.taskset.contentHash},u=k({originRef:s,tasksetRef:i,evaluationRef:n}),f=a.find(c=>c.metadata.downstreamLinkKey===u);if(f)return f;let p=(e.now??(()=>new Date().toISOString()))(),l=Rr({schemaVersion:"openpond.harnessEvaluationReviewReceipt.v1",id:`evaluation-review-link-${u.slice(0,24)}`,ownerScope:o.ownerScope,workspaceRef:o.workspaceRef,harnessRelease:o.harnessRelease,previousWatermark:o.nextWatermark,nextWatermark:{cursor:k({cursor:o.nextWatermark.cursor,tasksetRef:i,evaluationRef:n}),throughCreatedAt:p},selectedEvidence:o.selectedEvidence,excludedEvidence:o.excludedEvidence,claim:o.claim,classification:"taskset",triage:o.triage,reason:n?"The approved Taskset completed a real baseline Evaluation.":"The approved Harness review materialized one immutable Taskset.",nextAuthority:"human_review",maxEstimatedCostUsd:o.maxEstimatedCostUsd,tasksetProposal:i,evaluation:n,trainingQualification:null,policyVersion:o.policyVersion,createdAt:p,metadata:{originReview:s,downstreamLinkKey:u,tasksetRevision:e.taskset.revision}});return await e.store.saveHarnessImprovementArtifact(r,"evaluation_review",l),l}d(Ps,"linkHarnessReviewTaskset");async function aa(e,t,r){let s=await e.getHarnessWorkspace(t);if(!s)throw new Error(`Harness workspace ${t} was not found.`);let a=(await e.listHarnessImprovementArtifacts(t,"evaluation_review",1e3)).find(n=>n.id===r.id&&n.contentHash===r.contentHash);if(!a||a.schemaVersion!=="openpond.harnessEvaluationReviewReceipt.v1")throw new Error("Harness Evaluation review was not found.");let o=a;if(o.classification!=="taskset"||o.nextAuthority!=="human_review")throw new Error("Only a human-review Taskset recommendation can enter Taskset authoring.");if(o.tasksetProposal)throw new Error("Harness Evaluation review already has a materialized Taskset link.");if(o.ownerScope.kind!==s.ownerScope.kind||o.ownerScope.id!==s.ownerScope.id)throw new Error("Harness Evaluation review owner scope does not match its workspace.");return o}d(aa,"requireTasksetReview");function _e(e){return`${e.id}:${e.contentHash}`}d(_e,"artifactKey");function Jt(e){return!!e&&typeof e=="object"&&!Array.isArray(e)}d(Jt,"isRecord");function Gt(e){let t=new Set;return e.filter(r=>{let s=JSON.stringify(r);return t.has(s)?!1:(t.add(s),!0)})}d(Gt,"uniqueObjects");import{randomUUID as na}from"node:crypto";function bt(e){let t=St(St(St(e.metadata).billing).pricing);return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}d(bt,"hostedTokenPricingFromCatalog");function mi(e){let t=St(e);try{return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}catch{return null}}d(mi,"hostedTokenPricingFromValue");function xe(e,t){let r=ve(e);if(r.promptTokens===null&&r.completionTokens===null)return null;let s=r.cachedPromptTokens??0;return((r.uncachedPromptTokens??r.promptTokens??0)*t.inputUsdPerMillionTokens+s*t.cachedInputUsdPerMillionTokens+(r.completionTokens??0)*t.outputUsdPerMillionTokens)/1e6}d(xe,"hostedUsageCostUsd");function St(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(St,"record");function Le(e,t){if(typeof e=="string"&&e.trim())return e.trim();throw new Error(`Hosted model catalog is missing ${t}.`)}d(Le,"requiredString");function Ue(e,t){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return e;throw new Error(`Hosted model catalog is missing ${t}.`)}d(Ue,"requiredRate");function Ds(e={}){let t=e.catalog??pt,r=e.stream??ft;return{async prepare(s){if(s.providerId!=="openpond")throw new Error("The selected judge provider is not connected to the hosted judge transport.");if(s.revision!==null)throw new Error("This judge transport cannot enforce a model revision; select a provider with revision support.");let a=await t();if(a.error)throw new Error(`Judge model catalog unavailable: ${a.error}`);let o=a.models.find(l=>l.id===s.modelId);if(!o)throw new Error("The selected judge model is unavailable.");let n=Os(o.raw),i=bt(n),u=Hs(n.context_window,"context window"),f=Math.min(4096,Hs(n.output_limit,"output limit")),p=(u*Math.max(i.inputUsdPerMillionTokens,i.cachedInputUsdPerMillionTokens)+f*i.outputUsdPerMillionTokens)/1e6;if(!(p>0))throw new Error("The judge model must have a positive known charge ceiling.");return{maximumChargeUsd:p,async dispatch(l){l?.throwIfAborted();let c="",m=null,g=null,h=null,w=!1;for await(let _ of r({model:s.modelId,messages:[{role:"system",content:s.system},{role:"user",content:s.data}],requestId:`learning-judge-${na()}`,maxTokens:f,temperature:s.temperature,signal:l})){let x=Os(_.raw);if(typeof x.model=="string"){if(m!==null&&x.model!==m)throw new Error("Judge response model identity changed during streaming.");m=x.model}if(typeof x.id=="string"){if(g!==null&&g!==x.id)throw new Error("Judge response identity changed during streaming.");g=x.id}if(_.type==="text_delta"&&(c+=_.text),c.length>1e5)throw new Error("Judge response exceeds the retained response limit.");_.type==="usage"&&(h=_.usage),_.type==="finish"&&(w=!0)}if(!w||!m||!g)throw new Error("Judge response did not include a complete model receipt.");let y=ve(h),S=y.promptTokens!==null&&y.completionTokens!==null?xe(h,i):null;return{text:c,modelId:m,modelRevision:null,responseId:g,inputTokens:y.promptTokens,outputTokens:y.completionTokens,costUsd:S}}}},async cancel({run:s}){return!s.judgeCalls?.some(a=>a.status==="reserved")}}}d(Ds,"createLearningHostedJudgeProvider");function Os(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Os,"record");function Hs(e,t){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new Error(`Judge catalog is missing its ${t}.`);return e}d(Hs,"positiveInteger");function xi(e,t=Ds()){return{execute(r){if(r.definition.execution.environment.entrypoint==="openpond.javascript-environment.v1")throw new Error("Tool Tasksets require grading an owner-recorded environment attempt.");let s=Pr({repository:e,scope:r.scope,run:r.run});return Ir({binding:r.binding,rewards:r.rewards,task:dt(r.evidence,r.definition),evidence:Mr(r.evidence,r.run.output),evaluatorContext:r.evidence.submission.evaluatorContext,signal:r.signal,modelJudge:Er({readRubric:d(a=>e.transaction(r.scope,async o=>oe(await ke(o,"asset",a.id,1),a)),"readRubric"),async executeBudgeted(a,o){let n=await t.prepare(a,{scope:r.scope,run:r.run});return Cr({store:s,maximumCharge:d(()=>n.maximumChargeUsd,"maximumCharge"),dispatch:d((u,f)=>n.dispatch(f),"dispatch")})(`grade-${k(a)}`,a,o)}}),customVerifier:d(async({grader:a,task:o,evidence:n})=>{let i=await Ye({repository:e,scope:r.scope,grader:a,task:o,evidence:n,evaluatorContext:r.evidence.submission.evaluatorContext,timeoutMs:r.run.timeoutMs,signal:r.signal});return{score:i.score,passed:i.passed,rewardEligible:a.rewardEligible,failureClass:null,feedback:[i.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:i.evidenceRefs}},"customVerifier")})},async cancel({scope:r,run:s}){let a=await e.transaction(r,o=>ke(o,"grade",s.id));return a.judgeCalls?.some(o=>o.status==="reserved")?!1:t.cancel({scope:r,run:a})}}}d(xi,"createLocalTaskGradeExecutor");async function Ye(e){let t=await e.repository.transaction(e.scope,async r=>oe(await ke(r,"asset",e.grader.verifierRef.id,1),e.grader.verifierRef));return $e({source:t,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:Math.min(e.grader.timeoutMs,e.timeoutMs??e.grader.timeoutMs),signal:e.signal,value:{task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,expectedOutput:e.task.expectedOutput,evaluatorContext:e.evaluatorContext,infrastructureError:e.evidence.infrastructureError??null}})}d(Ye,"executeLocalLearningVerifier");import ia from"node:path";async function js(e,t,r){if(!r)throw new Error("Ordinary tool execution requires its owner's package directory.");let s=De({taskset:e,adapterId:qe(e)}),a=await vt({root:ia.join(r,"training","tasksets",ee(e)),content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:s.tasksetRelease,environment:s.environmentRelease,verifierSet:s.verifierSetRelease},sources:ct.parse(e.metadata.portableFileInventory)}),o=Fr(a);if(!o)throw new Error("The Taskset does not declare a JavaScript execution graph.");let n=a.taskset.tasks.find(f=>f.id===t.id);if(!n)throw new Error("The selected task is missing from its executable package.");let i=o.assets.find(f=>f.id===o.execution.javascript.module.id),u=o.assets.find(f=>f.id===n.privilegedContextRef);return{execution:o.execution,module:i,initialState:JSON.parse(u.text),definition:L(a.taskset),instructions:Vr(a),task:n}}d(js,"loadOrdinaryToolEnvironment");var J="openpond.javascript-environment.v1";async function Be(e,t,r,s){let a=t.tasks.find(o=>o.id===r.id);if(!a||k(a)!==k(r))throw new Error("Tool task differs from the selected immutable Taskset.");return t.metadata.taskDefinition===void 0?js(t,r,s):e.learningRepository().transaction(t.profileId,async o=>{let n=await ye(o,"definition",ae.parse(t.metadata.taskDefinition)),i=n.execution.environment;if(i.entrypoint!==J||t.environment.entrypoint!==J||t.environment.kind!=="agent"||t.environment.stateful!==i.stateful||t.environment.defaultTimeoutMs!==i.defaultTimeoutMs||t.environment.networkPolicy!==i.networkPolicy||t.environment.deterministicSeeds!==i.deterministicSeeds||k(t.environment.toolNames)!==k(n.execution.tools.map(v=>v.name))||k(t.environment.metadata.portableTools)!==k(n.execution.tools)||k(t.policy)!==k(n.execution.policy)||k(t.metadata.rewardBinding)!==k(n.rewardBinding))throw new Error("Tool Taskset context differs from its pinned task definition.");let u=await ye(o,"binding",n.rewardBinding),f=await Promise.all(u.sources.map(v=>ye(o,"reward",v.reward))),p=it.parse({id:r.id,clusterKey:r.clusterKey,split:r.split,input:r.input,expectedOutput:r.expectedOutput,policyVisibleContext:r.policyVisibleContext,privilegedContextRef:r.privilegedContextRef,artifactRefs:[],tags:r.tags}),l={...n.execution,graders:be(u,f),tasks:[p]},c=await ke(o,"asset",Ur(l),1),m=$r.parse(JSON.parse(oe(c,c.asset))),g=[m.javascript.module,m.environment.actionSchemaRef,m.environment.observationSchemaRef,m.environment.stateSchemaRef].filter(v=>v!==null),h=[...new Set([...g.map(v=>v.id),r.privilegedContextRef].filter(v=>v!==null))],w=await Promise.all(h.map(v=>ke(o,"asset",v,1))),y=Br(l,c,w);if(k(t.environment.metadata.portableExecutionResources)!==k({environment:y.environment,verifierSet:y.verifierSet}))throw new Error("Tool Taskset publication differs from its private execution resources.");let S=w.find(v=>v.id===y.javascript.module.id),_=w.find(v=>v.id===r.privilegedContextRef),x=JSON.parse(oe(_,_.asset));return{execution:y,module:S,initialState:x,definition:L(n),instructions:n.instructions,task:p}})}d(Be,"loadStarterToolEnvironment");function Ns(e){let t=Wt({text:e.instructions,path:"policy/instructions.txt",mediaType:"text/plain",visibility:"policy"}),r=Wt({text:JSON.stringify({environment:L(e.execution.environment),verifierSet:L(e.execution.verifierSet),javascript:L(e.execution.javascript)}),path:"environment/dependencies.json",mediaType:"application/json",visibility:"host_private"}),s=_r({schemaVersion:"openpond.agentSnapshot.v2",id:`starter-agent-${e.definition.contentHash}`,sourceRelease:{id:e.definition.id,contentHash:e.definition.contentHash},instructions:[t.asset],skills:[],agents:[],toolDeclarations:e.execution.javascript.tools,capabilityRequirements:[],dependencyLock:r.asset,portability:{portable:!0,blockers:[],localOnlyAssetRefs:[],hostPrivateAssetRefs:[e.module.id,r.id]},metadata:{execution:J}}),a=xr({schemaVersion:"openpond.harnessRelease.v2",id:`starter-harness-${s.contentHash}`,agentSnapshot:{id:s.id,contentHash:s.contentHash},program:e.module.asset,tools:e.execution.javascript.tools,lifecycle:{create:!0,reset:!0,step:!0,collect:!0,destroy:!0,resetScope:"attempt"},graderInterface:{visibleEvidence:["output"],privilegedEvidence:["environment_state","private_verifier"],privateVerifierIsolation:!0},files:[t.asset,r.asset,e.module.asset],metadata:{execution:J}});return{agentSnapshot:s,harnessRelease:a,instructionContext:void 0}}d(Ns,"compileStarterToolHarness");import Rt from"node:path";import{constants as qs}from"node:fs";import{open as da,realpath as $s}from"node:fs/promises";var ca=P.string().regex(/^[a-f0-9]{64}$/),la=P.object({schemaVersion:P.literal("openpond.starterToolAttempt.v1"),requestId:P.string(),taskset:ae,taskHash:ca,model:P.unknown(),seed:P.number().int(),attempt:P.number().int(),definition:ae,policySource:P.enum(["model","fixture"]),environment:ae,verifierSet:ae,result:qr}).strict();async function _t(e){let{store:t,taskset:r,task:s,attempt:a}=e,o=await Be(t,r,s,e.storeDir),i=(await t.listTaskAttemptArtifacts({attemptId:a.id})).filter(S=>S.kind==="environment_state"&&S.tasksetId===r.id&&S.taskId===s.id&&S.attemptId===a.id&&a.artifactRefs.includes(S.id)&&a.privilegedOutcomeRef===S.id);if(i.length!==1)throw new Error("Tool grading requires exactly one owner-recorded environment artifact.");let u=i[0],f=await $s(Rt.join(e.storeDir,"training","evaluation-artifacts")),p=await $s(Rt.dirname(u.path)),l=Rt.relative(f,p);if(!l||l.startsWith("..")||Rt.isAbsolute(l))throw new Error("Tool evidence is outside the execution owner's artifact directory.");let c=await da(u.path,qs.O_RDONLY|qs.O_NOFOLLOW),m;try{let S=await c.stat();if(!S.isFile()||S.size>33554432||S.size!==u.sizeBytes)throw new Error("Tool evidence exceeds its declared byte boundary.");m=await c.readFile()}finally{await c.close()}if(m.byteLength!==u.sizeBytes||se(m)!==u.sha256)throw new Error("Owner-recorded tool evidence bytes changed.");let g=la.parse(JSON.parse(m.toString("utf8"))),{contentHash:h,...w}=g.result,y=g.result;if(k(w)!==h||!y.snapshot||!y.environmentCleanupComplete||!["completed","budget_exhausted","policy_failure"].includes(y.status)||k(g.taskset)!==k(L(r))||g.taskHash!==k(s)||k(g.model)!==k(a.modelRef)||g.seed!==a.seed||g.attempt!==a.attempt||a.tasksetId!==r.id||a.taskId!==s.id||a.split!==s.split||k(g.definition)!==k(o.definition)||k(g.environment)!==k(L(o.execution.environment))||k(g.verifierSet)!==k(L(o.execution.verifierSet))||y.taskId!==s.id||y.snapshot.seed!==a.seed||y.snapshot.inputHash!==k(s.input)||k(y.snapshot.definition)!==k(L(o.execution.javascript))||y.snapshot.finalStateHash!==k(y.snapshot.state)||k(a.output)!==k({text:y.output??""})||a.metadata.environmentAttemptHash!==h)throw new Error("Tool evidence differs from the selected task, model or recorded attempt.");if(a.metadata.policySource!==g.policySource)throw new Error("Tool evidence policy source changed.");return{environment:{status:y.status,collected:y.collected,definition:y.snapshot.definition,initialStateHash:y.snapshot.initialStateHash,finalStateHash:y.snapshot.finalStateHash,finalState:y.snapshot.state,events:y.snapshot.events}}}d(_t,"readStarterToolEvidence");async function Te(e,t){if(t.metadata.learning!==void 0)return He.parse(t.metadata.learning);if(t.metadata.rewardBinding===void 0)return;let r=ae.parse(t.metadata.rewardBinding);return e.learningRepository().transaction(t.profileId,async s=>{let a=await ye(s,"binding",r),o=[...new Map(a.sources.map(i=>[`${i.reward.id}:${i.reward.revision}:${i.reward.contentHash}`,i.reward])).values()],n=await Promise.all(o.map(i=>ye(s,"reward",i)));return{binding:a,rewards:n}})}d(Te,"resolveTasksetRewardBinding");async function ua(e,t,r){let s=await Te(e,t);if(!s)return{rewardExecution:void 0,verifierAssets:[]};let a=[...s.rewards.flatMap(i=>i.assets),...be(s.binding,s.rewards).flatMap(i=>i.kind==="custom_verifier"?[i.verifierRef]:i.kind==="model_judge"?[i.rubricRef]:[])],o=t.metadata.learning===void 0?void 0:await wt(r,t);if(o){let i=[...new Map(a.map(u=>{let f=o.learningResources.assets.find(p=>p.id===u.id);if(!f)throw new Error("Imported batch verifier is missing from its package.");return oe(f,u),[f.id,f]})).values()];return{rewardExecution:s,verifierAssets:i}}let n=await e.learningRepository().transaction(t.profileId,async i=>{let u=new Map;for(let f of a){let p=await ke(i,"asset",f.id,1);oe(p,f),u.set(p.id,p)}return[...u.values()]});return{rewardExecution:s,verifierAssets:n}}d(ua,"resolveTasksetTrainingReward");async function gd(e,t,r){let s=await ua(e,t,r.storeDir);if(!s.rewardExecution)return s;let a=be(s.rewardExecution.binding,s.rewardExecution.rewards);if(!s.rewardExecution.binding.sources.some(o=>o.role==="evaluation"&&o.weight>0))throw new Error("Managed training requires a positively weighted evaluation Reward source for retained validation.");if(a.some(o=>o.kind==="custom_verifier"&&o.runtime==="sandbox_process"))throw new Error("Managed preparation requires a qualified isolated sandbox for process verifiers.");if(r.placement!=="remote"||t.environment.kind==="work"||t.capabilities.requiresState||t.capabilities.requiresTools||a.some(o=>o.kind==="human"||o.kind==="model_judge"&&(!o.model||o.calibrationStatus!=="passed"))||t.tasks.some(o=>o.privilegedContextRef!==null)||!s.rewardExecution.binding.sources.some(o=>o.role==="training"&&o.weight>0)||r.hasLearnedPreferenceReward)throw new Error("This Reward binding requires an additional managed execution adapter.");return s}d(gd,"resolveManagedTasksetReward");async function Ls(e,t,r){let s=await Te(e,t);if(!s)throw new Error("The Taskset has no published Reward binding.");let a=be(s.binding,s.rewards);return async({grader:o,task:n,attempt:i,signal:u})=>{let f=t.tasks.find(m=>m.id===n.id);if(!f||k(f)!==k(n))throw new Error("Task differs from the selected immutable Taskset.");let p=null;if(t.environment.entrypoint===J){if(!r)throw new Error("Tool grading requires the execution owner's artifact directory.");p=await _t({store:e,storeDir:r,taskset:t,task:n,attempt:i})}else if(n.privilegedContextRef!==null)throw new Error("This authored Taskset requires a private evaluator-context resolver.");let l=a.find(m=>m.id===o.id);if(l?.kind!=="custom_verifier")throw new Error("The verifier is absent from the published Reward binding.");let c=it.parse({id:n.id,clusterKey:n.clusterKey,split:n.split,input:n.input,expectedOutput:n.expectedOutput,policyVisibleContext:n.policyVisibleContext,privilegedContextRef:n.privilegedContextRef,artifactRefs:[],tags:n.tags});return Ye({repository:e.learningRepository(),scope:t.profileId,grader:l,task:c,evidence:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},evaluatorContext:p,signal:u})}}d(Ls,"createTasksetBindingVerifier");async function Ie(e,t,r){let s=r?ae.parse(r):null;if(s&&s.id!==t)throw new Error("Taskset execution reference differs from the requested Taskset.");let a=s?await e.getTasksetRevision(s.id,s.revision):await e.getTaskset(t);if(!a||s&&(a.contentHash!==s.contentHash||Re(a)!==s.contentHash))throw new Error("Taskset execution revision is unavailable or differs from its immutable hash.");return a}d(Ie,"requireLocalTasksetRevision");async function Sd(e,t,r){let s=await e.releaseForTaskset(t);return s||De({taskset:t,rewardExecution:r?await Te(r,t):void 0,adapterId:qe(t)}).tasksetRelease}d(Sd,"requireReleasedTaskset");import{createHash as ma}from"node:crypto";function Us(e,t,r){if(t.metadata.importedPackageHash!==e.contentHash)throw new Error("The Work projection differs from the admitted published package.");let s=e.taskset.tasks.find(n=>n.id===r.id);if(!s||s.split!==r.split||k(r.metadata.portableTaskRecord)!==k(s))throw new Error("The Work task differs from its exact released population.");let a=0,o=(r.assets??[]).map(n=>{let i=s.artifactRefs.find(l=>l.id===n.id&&l.contentHash===n.sha256&&l.path===n.artifactRef),u=i&&e.files.find(l=>k(l.asset)===k(i));if(!i||!u||i.visibility!=="policy")throw new Error("The Work task requests an unavailable or private evaluator file.");let f=Buffer.from(Wr(u));if(ma("sha256").update(f).digest("hex")!==n.sha256||f.length!==n.sizeBytes||n.split!==r.split)throw new Error("The published Work input manifest changed.");if(a+=f.length,a>25e7)throw new Error("Published Work inputs exceed the admitted 250 MB ceiling.");return{...n,assetId:n.id,storageName:n.stagingPath??n.fileName,bytes:f,sizeBytes:f.length}});if(o.length!==s.artifactRefs.length)throw new Error("The exact policy-visible Work asset population is incomplete.");return o}d(Us,"advancedPublishedWorkAssets");async function Yt(e,t=pt,r=!1){if(e.kind==="hosted_harness")return Yt({kind:"hosted_chat",modelId:e.modelId,...Nr(e)},t);if(e.kind!=="hosted_chat")throw new ne("local_target_not_qualified","Select an explicit model target for local execution.",422);if(e.harness&&!r)throw new ne("local_harness_not_qualified","This released Harness requires its native turn executor; it cannot run as a model-only case.",422);let s=await t();if(s.error)throw new ne("local_model_catalog_unavailable",s.error,422);let a=s.models.find(c=>c.id===e.modelId);if(!a)throw new ne("local_model_unavailable","The selected model is unavailable.",422);let o=Bs(a.raw),n=bt(o),i=Bs(o.capabilities).samplingParameters===!0;if(!i&&(e.temperature!==0||e.topP!==1))throw new ne("local_sampling_unsupported","This model uses provider-managed sampling. Custom temperature or top P requires an explicit supported capability.",422);let u=Ws(o.context_window),f=Ws(o.output_limit);if(e.maxOutputTokens>f)throw new ne("local_model_output_limit","The model cannot accept the configured output limit.",422);let p={providerId:"openpond",modelId:e.modelId,maxOutputTokens:e.maxOutputTokens,...i?{temperature:e.temperature,topP:e.topP}:{},messages:e.messages??[]},l=(u*Math.max(n.inputUsdPerMillionTokens,n.cachedInputUsdPerMillionTokens)+e.maxOutputTokens*n.outputUsdPerMillionTokens)/1e6;if(!(l>0))throw new ne("local_model_charge_unknown","The model must publish an enforceable positive request ceiling.",422);return{model:{...p,configurationHash:k({configuration:p,contextWindow:u,outputLimit:f,pricing:n})},pricing:n,maximumChargeUsd:l}}d(Yt,"prepareLocalExperimentModel");function Nd(e){let t=e.admission.streamFactory?.(e)??e.admission.stream??e.stream??ft;return async function*(r){if(!r.requestId)throw new ne("local_request_identity_missing","Every local model dispatch requires a stable request identity.");let s=e.admission.model;if(r.model!==s.modelId||r.maxTokens!==s.maxOutputTokens||r.temperature!==s.temperature||r.topP!==s.topP)throw new ne("local_model_configuration_conflict","The native model request changed its admitted model, output limit or sampling configuration.",422);let a=xs();a?.assertExecution(e.executionId);let o={teamId:e.teamId,id:e.executionId,ownerId:e.ownerId,requestId:r.requestId};await e.store.reserveLocalCharge({...o,caseId:e.caseId,maximumUsd:e.admission.maximumChargeUsd}),r.signal?.aborted&&(await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),r.signal.throwIfAborted());try{await e.store.markLocalChargeDispatched(o.teamId,o.id,o.requestId,o.ownerId)}catch(p){throw a&&await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),p}await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:"model.dispatch",payload:{requestId:o.requestId,model:r.model,messages:r.messages}});let n=null,i=!1,u=!1,f;try{for await(let p of t(r)){u=!0,p.type==="usage"&&(n=p.usage),p.type==="finish"&&(i=!0);let l=p.type==="text_delta"?{requestId:o.requestId,text:p.text}:p.type==="tool_call_delta"?{requestId:o.requestId,toolCalls:p.toolCalls}:p.type==="usage"?{requestId:o.requestId,usage:ve(p.usage)}:p.type==="finish"?{requestId:o.requestId,finishReason:p.finishReason}:null;l&&await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:`model.${p.type}`,payload:l}),yield p}}catch(p){throw!u&&p instanceof Is&&p.executionId===o.id&&p.requestId===o.requestId&&(f=p),p}finally{let p=ve(n),l=e.admission.pricing&&i&&p.promptTokens!==null&&p.completionTokens!==null?xe(n,e.admission.pricing):null;await e.store.settleLocalCharge({...o,costUsd:l,confirmedTransportNotInvoked:f,usage:n===null?null:{promptTokens:p.promptTokens,completionTokens:p.completionTokens,cachedPromptTokens:p.cachedPromptTokens,uncachedPromptTokens:p.uncachedPromptTokens}})}}}d(Nd,"createLocalBudgetedModelStream");function Bs(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Bs,"record");function Ws(e){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new ne("local_model_bound_unknown","The model catalog is missing an enforceable context/output bound.",422);return e}d(Ws,"positive");import{AsyncLocalStorage as pa}from"node:async_hooks";import fa from"node:path";var Fs=new pa;function Vs(){return Fs.getStore()}d(Vs,"currentAdvancedRefinerBoundary");function Jd(e){let t=Sr(fa.join(e.storeDir,"library","harnesses","advanced-refiner-paid-calls.sqlite"));t.exec(`CREATE TABLE IF NOT EXISTS advanced_refiner_budgets(id TEXT PRIMARY KEY,actor_id TEXT NOT NULL,team_id TEXT NOT NULL,pin_hash TEXT NOT NULL,maximum_spend REAL NOT NULL,maximum_steps INTEGER NOT NULL,expires_at TEXT NOT NULL);
+import{a as As,b as Ms}from"./chunk-UMRQKE3W.js";import{a as Es,e as $e,f as Cs,g as Tt}from"./chunk-HUMNLSCV.js";import{a as xs,c as Is}from"./chunk-XVDFJ3ZF.js";import{h as ne}from"./chunk-O5ZOW474.js";import{a as ee,b as _s}from"./chunk-NDNSLCPE.js";import{b as ys,c as bs,d as yt}from"./chunk-CWE2N3JI.js";import{a as ws,b as vs,c as Ts,d as qe,e as Ss,g as vt}from"./chunk-O4YHISU2.js";import{a as wt,c as Rs}from"./chunk-HT6ILPD6.js";import{b as Kr,e as Xr,f as lt,g as ut}from"./chunk-VMW23TIZ.js";import{F as gt,G as fs,H as ht,I as kt,J as gs,K as hs,N as ks,x as ps}from"./chunk-23YW5PAQ.js";import{a as ve}from"./chunk-7KWD2I6W.js";import{g as pt,i as ft}from"./chunk-2ACX7IW5.js";import{s as br}from"./chunk-OFJPOTES.js";import{d as Ne}from"./chunk-U7D6Q744.js";import{ja as rs,qa as ss,sa as os,x as ts}from"./chunk-WXG3G2O6.js";import{M as V,N as as,O as ns,P as is,Q as mt,R as ds,S as q,T as cs,U as je,V as ls,X as us,Y as Vt,Z as ms}from"./chunk-OXRN2257.js";import{l as Ft}from"./chunk-TOWY6Y23.js";import{d as Zr,f as W,g as Qr,i as es}from"./chunk-Z6HWSTEP.js";import{k as Nr}from"./chunk-XKDOPNQ5.js";import{c as ct}from"./chunk-GR4MJJCE.js";import{Ba as Gr,Ca as Re,Ha as Yr,Ia as De,K as qr,L as $r,M as Lr,N as Ur,P as Br,aa as Wr,ba as Fr,ia as Vr,ua as zr,wa as we,xa as Jr}from"./chunk-BYVE4TMU.js";import{x as ae}from"./chunk-3VT2PPKW.js";import{C as Or,S as Hr,T as Dr,Tb as jr}from"./chunk-D6L6UNUQ.js";import{f as Rr}from"./chunk-AJXWJLZA.js";import{L,M as Ar,Ob as Pr,R as Wt,S as oe,X as dt,Y as Mr,ba as He,d as Er,g as Cr,pb as ke,qb as ye}from"./chunk-ZEOYNUHL.js";import{bb as be,cb as Ir,ea as it,v as Ge}from"./chunk-ZY32SZ2P.js";import{l as _r,m as xr}from"./chunk-OTDDXEQ6.js";import{i as Tr,m as se,n as k}from"./chunk-RCT3O3HQ.js";import{m as Sr}from"./chunk-UGOK7VTC.js";import{A as P}from"./chunk-X2J2P5ZT.js";import{a as d}from"./chunk-FTQLQJQV.js";var zt={buildIntent:"verifiable_reward",methodHint:"grpo"};async function ii(e){let t=await aa(e.store,e.workspaceId,e.reviewRef),r=`harness-review-${k({review:t.contentHash,strategy:zt}).slice(0,24)}`,s=(await e.store.listTaskCreationSnapshots(e.profileId)).find(c=>c.request.candidateId===r);if(s)return s;let[a,o,n]=await Promise.all([e.store.listHarnessImprovementArtifacts(e.workspaceId,"route_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"trigger_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"observation",1e3)]),i=new Map(a.map(c=>[_e(c),c])),u=new Map(o.map(c=>[_e(c),c])),f=new Map(n.map(c=>[_e(c),c])),p=new Map;for(let c of t.selectedEvidence){if(c.kind!=="route_decision")continue;let m=i.get(_e(c.evidence)),g=m?u.get(_e(m.trigger)):null;if(!m||!g)throw new Error(`Harness review evidence ${c.evidence.id} cannot be reconstructed.`);let h=g.observations.map(y=>f.get(_e(y))??null).filter(y=>y?.runRef===c.sourceRef);if(!h.length)throw new Error(`Harness review source ${c.sourceRef} has no exact observation evidence.`);let w=p.get(c.sourceRef)??{turnIds:new Set,evidenceRefs:[],sourcePolicies:[]};for(let y of h)w.turnIds.add(y.turnId);w.evidenceRefs.push(c.evidence),w.sourcePolicies.push(c.sourcePolicy),p.set(c.sourceRef,w)}if(!p.size)throw new Error("Harness Taskset review has no reconstructable source runs.");let l=[];for(let[c,m]of[...p.entries()].sort(([g],[h])=>g.localeCompare(h))){let g=await e.taskCreator.addSessionSource({profileId:e.profileId,sessionId:c,turnIds:[...m.turnIds].sort(),consentScope:"selected_turns"}),h=Array.isArray(g.metadata.harnessEvaluationReviews)?g.metadata.harnessEvaluationReviews:[],w={...g,metadata:{...g.metadata,harnessEvaluationReviews:Gt([...h,{id:t.id,contentHash:t.contentHash,workspaceId:t.workspaceRef,harnessRelease:t.harnessRelease,claimFingerprint:t.claim?.fingerprint??null,evidenceRefs:m.evidenceRefs,sourcePolicies:m.sourcePolicies}])}};l.push(await e.store.upsertTrainingSource(w))}return e.startCreation({profileId:e.profileId,sourceIds:l.map(c=>c.id),surface:"task_candidate",mode:"customize",entryMode:"automated",resourceIntent:"workproduct",buildIntent:zt.buildIntent,buildSpecification:null,objective:t.claim?.statement??t.reason,methodHint:zt.methodHint,preferredBaseModelId:null,preferredBaseModel:null,candidateId:r,analysisModel:e.analysisModel??null,analysisReasoningEffort:e.analysisReasoningEffort??null,createImproveRunId:null,targetIntent:{kind:"model",id:null,displayName:null,operation:"create"}})}d(ii,"startHarnessReviewTasksetAuthoring");function di(e){let t=e.flatMap(o=>Array.isArray(o.metadata.harnessEvaluationReviews)?o.metadata.harnessEvaluationReviews.filter(Jt):[]);if(!t.length)return null;let r=`${String(t[0].id)}:${String(t[0].contentHash)}`,s=t.filter(o=>`${String(o.id)}:${String(o.contentHash)}`===r);if(s.length!==e.length)return null;let a=s[0];return{review:{id:String(a.id),contentHash:String(a.contentHash),workspaceId:String(a.workspaceId),harnessRelease:a.harnessRelease,claimFingerprint:typeof a.claimFingerprint=="string"?a.claimFingerprint:null},evidenceRefs:Gt(s.flatMap(o=>Array.isArray(o.evidenceRefs)?o.evidenceRefs:[])),sourcePolicies:Gt(s.flatMap(o=>Array.isArray(o.sourcePolicies)?o.sourcePolicies:[]))}}d(di,"harnessReviewLineageFromSources");async function Ps(e){let t=e.taskset.metadata.harnessEvaluationLineage;if(!Jt(t)||!Jt(t.review))return null;let r=String(t.review.workspaceId??""),s={id:String(t.review.id??""),contentHash:String(t.review.contentHash??"")};if(!r||!s.id||!s.contentHash)throw new Error("Taskset Harness Evaluation lineage is incomplete.");let a=await e.store.listHarnessImprovementArtifacts(r,"evaluation_review",1e3),o=a.find(c=>_e(c)===_e(s));if(!o)throw new Error("Originating Harness Evaluation review was not found.");let n=e.evaluationResult?{id:e.evaluationResult.id,contentHash:e.evaluationResult.contentHash}:null,i={id:e.taskset.id,contentHash:e.taskset.contentHash},u=k({originRef:s,tasksetRef:i,evaluationRef:n}),f=a.find(c=>c.metadata.downstreamLinkKey===u);if(f)return f;let p=(e.now??(()=>new Date().toISOString()))(),l=Rr({schemaVersion:"openpond.harnessEvaluationReviewReceipt.v1",id:`evaluation-review-link-${u.slice(0,24)}`,ownerScope:o.ownerScope,workspaceRef:o.workspaceRef,harnessRelease:o.harnessRelease,previousWatermark:o.nextWatermark,nextWatermark:{cursor:k({cursor:o.nextWatermark.cursor,tasksetRef:i,evaluationRef:n}),throughCreatedAt:p},selectedEvidence:o.selectedEvidence,excludedEvidence:o.excludedEvidence,claim:o.claim,classification:"taskset",triage:o.triage,reason:n?"The approved Taskset completed a real baseline Evaluation.":"The approved Harness review materialized one immutable Taskset.",nextAuthority:"human_review",maxEstimatedCostUsd:o.maxEstimatedCostUsd,tasksetProposal:i,evaluation:n,trainingQualification:null,policyVersion:o.policyVersion,createdAt:p,metadata:{originReview:s,downstreamLinkKey:u,tasksetRevision:e.taskset.revision}});return await e.store.saveHarnessImprovementArtifact(r,"evaluation_review",l),l}d(Ps,"linkHarnessReviewTaskset");async function aa(e,t,r){let s=await e.getHarnessWorkspace(t);if(!s)throw new Error(`Harness workspace ${t} was not found.`);let a=(await e.listHarnessImprovementArtifacts(t,"evaluation_review",1e3)).find(n=>n.id===r.id&&n.contentHash===r.contentHash);if(!a||a.schemaVersion!=="openpond.harnessEvaluationReviewReceipt.v1")throw new Error("Harness Evaluation review was not found.");let o=a;if(o.classification!=="taskset"||o.nextAuthority!=="human_review")throw new Error("Only a human-review Taskset recommendation can enter Taskset authoring.");if(o.tasksetProposal)throw new Error("Harness Evaluation review already has a materialized Taskset link.");if(o.ownerScope.kind!==s.ownerScope.kind||o.ownerScope.id!==s.ownerScope.id)throw new Error("Harness Evaluation review owner scope does not match its workspace.");return o}d(aa,"requireTasksetReview");function _e(e){return`${e.id}:${e.contentHash}`}d(_e,"artifactKey");function Jt(e){return!!e&&typeof e=="object"&&!Array.isArray(e)}d(Jt,"isRecord");function Gt(e){let t=new Set;return e.filter(r=>{let s=JSON.stringify(r);return t.has(s)?!1:(t.add(s),!0)})}d(Gt,"uniqueObjects");import{randomUUID as na}from"node:crypto";function bt(e){let t=St(St(St(e.metadata).billing).pricing);return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}d(bt,"hostedTokenPricingFromCatalog");function mi(e){let t=St(e);try{return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}catch{return null}}d(mi,"hostedTokenPricingFromValue");function xe(e,t){let r=ve(e);if(r.promptTokens===null&&r.completionTokens===null)return null;let s=r.cachedPromptTokens??0;return((r.uncachedPromptTokens??r.promptTokens??0)*t.inputUsdPerMillionTokens+s*t.cachedInputUsdPerMillionTokens+(r.completionTokens??0)*t.outputUsdPerMillionTokens)/1e6}d(xe,"hostedUsageCostUsd");function St(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(St,"record");function Le(e,t){if(typeof e=="string"&&e.trim())return e.trim();throw new Error(`Hosted model catalog is missing ${t}.`)}d(Le,"requiredString");function Ue(e,t){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return e;throw new Error(`Hosted model catalog is missing ${t}.`)}d(Ue,"requiredRate");function Ds(e={}){let t=e.catalog??pt,r=e.stream??ft;return{async prepare(s){if(s.providerId!=="openpond")throw new Error("The selected judge provider is not connected to the hosted judge transport.");if(s.revision!==null)throw new Error("This judge transport cannot enforce a model revision; select a provider with revision support.");let a=await t();if(a.error)throw new Error(`Judge model catalog unavailable: ${a.error}`);let o=a.models.find(l=>l.id===s.modelId);if(!o)throw new Error("The selected judge model is unavailable.");let n=Os(o.raw),i=bt(n),u=Hs(n.context_window,"context window"),f=Math.min(4096,Hs(n.output_limit,"output limit")),p=(u*Math.max(i.inputUsdPerMillionTokens,i.cachedInputUsdPerMillionTokens)+f*i.outputUsdPerMillionTokens)/1e6;if(!(p>0))throw new Error("The judge model must have a positive known charge ceiling.");return{maximumChargeUsd:p,async dispatch(l){l?.throwIfAborted();let c="",m=null,g=null,h=null,w=!1;for await(let _ of r({model:s.modelId,messages:[{role:"system",content:s.system},{role:"user",content:s.data}],requestId:`learning-judge-${na()}`,maxTokens:f,temperature:s.temperature,signal:l})){let x=Os(_.raw);if(typeof x.model=="string"){if(m!==null&&x.model!==m)throw new Error("Judge response model identity changed during streaming.");m=x.model}if(typeof x.id=="string"){if(g!==null&&g!==x.id)throw new Error("Judge response identity changed during streaming.");g=x.id}if(_.type==="text_delta"&&(c+=_.text),c.length>1e5)throw new Error("Judge response exceeds the retained response limit.");_.type==="usage"&&(h=_.usage),_.type==="finish"&&(w=!0)}if(!w||!m||!g)throw new Error("Judge response did not include a complete model receipt.");let y=ve(h),S=y.promptTokens!==null&&y.completionTokens!==null?xe(h,i):null;return{text:c,modelId:m,modelRevision:null,responseId:g,inputTokens:y.promptTokens,outputTokens:y.completionTokens,costUsd:S}}}},async cancel({run:s}){return!s.judgeCalls?.some(a=>a.status==="reserved")}}}d(Ds,"createLearningHostedJudgeProvider");function Os(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Os,"record");function Hs(e,t){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new Error(`Judge catalog is missing its ${t}.`);return e}d(Hs,"positiveInteger");function xi(e,t=Ds()){return{execute(r){if(r.definition.execution.environment.entrypoint==="openpond.javascript-environment.v1")throw new Error("Tool Tasksets require grading an owner-recorded environment attempt.");let s=Pr({repository:e,scope:r.scope,run:r.run});return Ir({binding:r.binding,rewards:r.rewards,task:dt(r.evidence,r.definition),evidence:Mr(r.evidence,r.run.output),evaluatorContext:r.evidence.submission.evaluatorContext,signal:r.signal,modelJudge:Er({readRubric:d(a=>e.transaction(r.scope,async o=>oe(await ke(o,"asset",a.id,1),a)),"readRubric"),async executeBudgeted(a,o){let n=await t.prepare(a,{scope:r.scope,run:r.run});return Cr({store:s,maximumCharge:d(()=>n.maximumChargeUsd,"maximumCharge"),dispatch:d((u,f)=>n.dispatch(f),"dispatch")})(`grade-${k(a)}`,a,o)}}),customVerifier:d(async({grader:a,task:o,evidence:n})=>{let i=await Ye({repository:e,scope:r.scope,grader:a,task:o,evidence:n,evaluatorContext:r.evidence.submission.evaluatorContext,timeoutMs:r.run.timeoutMs,signal:r.signal});return{score:i.score,passed:i.passed,rewardEligible:a.rewardEligible,failureClass:null,feedback:[i.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:i.evidenceRefs}},"customVerifier")})},async cancel({scope:r,run:s}){let a=await e.transaction(r,o=>ke(o,"grade",s.id));return a.judgeCalls?.some(o=>o.status==="reserved")?!1:t.cancel({scope:r,run:a})}}}d(xi,"createLocalTaskGradeExecutor");async function Ye(e){let t=await e.repository.transaction(e.scope,async r=>oe(await ke(r,"asset",e.grader.verifierRef.id,1),e.grader.verifierRef));return $e({source:t,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:Math.min(e.grader.timeoutMs,e.timeoutMs??e.grader.timeoutMs),signal:e.signal,value:{task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,expectedOutput:e.task.expectedOutput,evaluatorContext:e.evaluatorContext,infrastructureError:e.evidence.infrastructureError??null}})}d(Ye,"executeLocalLearningVerifier");import ia from"node:path";async function js(e,t,r){if(!r)throw new Error("Ordinary tool execution requires its owner's package directory.");let s=De({taskset:e,adapterId:qe(e)}),a=await vt({root:ia.join(r,"training","tasksets",ee(e)),content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:s.tasksetRelease,environment:s.environmentRelease,verifierSet:s.verifierSetRelease},sources:ct.parse(e.metadata.portableFileInventory)}),o=Fr(a);if(!o)throw new Error("The Taskset does not declare a JavaScript execution graph.");let n=a.taskset.tasks.find(f=>f.id===t.id);if(!n)throw new Error("The selected task is missing from its executable package.");let i=o.assets.find(f=>f.id===o.execution.javascript.module.id),u=o.assets.find(f=>f.id===n.privilegedContextRef);return{execution:o.execution,module:i,initialState:JSON.parse(u.text),definition:L(a.taskset),instructions:Vr(a),task:n}}d(js,"loadOrdinaryToolEnvironment");var J="openpond.javascript-environment.v1";async function Be(e,t,r,s){let a=t.tasks.find(o=>o.id===r.id);if(!a||k(a)!==k(r))throw new Error("Tool task differs from the selected immutable Taskset.");return t.metadata.taskDefinition===void 0?js(t,r,s):e.learningRepository().transaction(t.profileId,async o=>{let n=await ye(o,"definition",ae.parse(t.metadata.taskDefinition)),i=n.execution.environment;if(i.entrypoint!==J||t.environment.entrypoint!==J||t.environment.kind!=="agent"||t.environment.stateful!==i.stateful||t.environment.defaultTimeoutMs!==i.defaultTimeoutMs||t.environment.networkPolicy!==i.networkPolicy||t.environment.deterministicSeeds!==i.deterministicSeeds||k(t.environment.toolNames)!==k(n.execution.tools.map(v=>v.name))||k(t.environment.metadata.portableTools)!==k(n.execution.tools)||k(t.policy)!==k(n.execution.policy)||k(t.metadata.rewardBinding)!==k(n.rewardBinding))throw new Error("Tool Taskset context differs from its pinned task definition.");let u=await ye(o,"binding",n.rewardBinding),f=await Promise.all(u.sources.map(v=>ye(o,"reward",v.reward))),p=it.parse({id:r.id,clusterKey:r.clusterKey,split:r.split,input:r.input,expectedOutput:r.expectedOutput,policyVisibleContext:r.policyVisibleContext,privilegedContextRef:r.privilegedContextRef,artifactRefs:[],tags:r.tags}),l={...n.execution,graders:be(u,f),tasks:[p]},c=await ke(o,"asset",Ur(l),1),m=$r.parse(JSON.parse(oe(c,c.asset))),g=[m.javascript.module,m.environment.actionSchemaRef,m.environment.observationSchemaRef,m.environment.stateSchemaRef].filter(v=>v!==null),h=[...new Set([...g.map(v=>v.id),r.privilegedContextRef].filter(v=>v!==null))],w=await Promise.all(h.map(v=>ke(o,"asset",v,1))),y=Br(l,c,w);if(k(t.environment.metadata.portableExecutionResources)!==k({environment:y.environment,verifierSet:y.verifierSet}))throw new Error("Tool Taskset publication differs from its private execution resources.");let S=w.find(v=>v.id===y.javascript.module.id),_=w.find(v=>v.id===r.privilegedContextRef),x=JSON.parse(oe(_,_.asset));return{execution:y,module:S,initialState:x,definition:L(n),instructions:n.instructions,task:p}})}d(Be,"loadStarterToolEnvironment");function Ns(e){let t=Wt({text:e.instructions,path:"policy/instructions.txt",mediaType:"text/plain",visibility:"policy"}),r=Wt({text:JSON.stringify({environment:L(e.execution.environment),verifierSet:L(e.execution.verifierSet),javascript:L(e.execution.javascript)}),path:"environment/dependencies.json",mediaType:"application/json",visibility:"host_private"}),s=_r({schemaVersion:"openpond.agentSnapshot.v2",id:`starter-agent-${e.definition.contentHash}`,sourceRelease:{id:e.definition.id,contentHash:e.definition.contentHash},instructions:[t.asset],skills:[],agents:[],toolDeclarations:e.execution.javascript.tools,capabilityRequirements:[],dependencyLock:r.asset,portability:{portable:!0,blockers:[],localOnlyAssetRefs:[],hostPrivateAssetRefs:[e.module.id,r.id]},metadata:{execution:J}}),a=xr({schemaVersion:"openpond.harnessRelease.v2",id:`starter-harness-${s.contentHash}`,agentSnapshot:{id:s.id,contentHash:s.contentHash},program:e.module.asset,tools:e.execution.javascript.tools,lifecycle:{create:!0,reset:!0,step:!0,collect:!0,destroy:!0,resetScope:"attempt"},graderInterface:{visibleEvidence:["output"],privilegedEvidence:["environment_state","private_verifier"],privateVerifierIsolation:!0},files:[t.asset,r.asset,e.module.asset],metadata:{execution:J}});return{agentSnapshot:s,harnessRelease:a,instructionContext:void 0}}d(Ns,"compileStarterToolHarness");import Rt from"node:path";import{constants as qs}from"node:fs";import{open as da,realpath as $s}from"node:fs/promises";var ca=P.string().regex(/^[a-f0-9]{64}$/),la=P.object({schemaVersion:P.literal("openpond.starterToolAttempt.v1"),requestId:P.string(),taskset:ae,taskHash:ca,model:P.unknown(),seed:P.number().int(),attempt:P.number().int(),definition:ae,policySource:P.enum(["model","fixture"]),environment:ae,verifierSet:ae,result:qr}).strict();async function _t(e){let{store:t,taskset:r,task:s,attempt:a}=e,o=await Be(t,r,s,e.storeDir),i=(await t.listTaskAttemptArtifacts({attemptId:a.id})).filter(S=>S.kind==="environment_state"&&S.tasksetId===r.id&&S.taskId===s.id&&S.attemptId===a.id&&a.artifactRefs.includes(S.id)&&a.privilegedOutcomeRef===S.id);if(i.length!==1)throw new Error("Tool grading requires exactly one owner-recorded environment artifact.");let u=i[0],f=await $s(Rt.join(e.storeDir,"training","evaluation-artifacts")),p=await $s(Rt.dirname(u.path)),l=Rt.relative(f,p);if(!l||l.startsWith("..")||Rt.isAbsolute(l))throw new Error("Tool evidence is outside the execution owner's artifact directory.");let c=await da(u.path,qs.O_RDONLY|qs.O_NOFOLLOW),m;try{let S=await c.stat();if(!S.isFile()||S.size>33554432||S.size!==u.sizeBytes)throw new Error("Tool evidence exceeds its declared byte boundary.");m=await c.readFile()}finally{await c.close()}if(m.byteLength!==u.sizeBytes||se(m)!==u.sha256)throw new Error("Owner-recorded tool evidence bytes changed.");let g=la.parse(JSON.parse(m.toString("utf8"))),{contentHash:h,...w}=g.result,y=g.result;if(k(w)!==h||!y.snapshot||!y.environmentCleanupComplete||!["completed","budget_exhausted","policy_failure"].includes(y.status)||k(g.taskset)!==k(L(r))||g.taskHash!==k(s)||k(g.model)!==k(a.modelRef)||g.seed!==a.seed||g.attempt!==a.attempt||a.tasksetId!==r.id||a.taskId!==s.id||a.split!==s.split||k(g.definition)!==k(o.definition)||k(g.environment)!==k(L(o.execution.environment))||k(g.verifierSet)!==k(L(o.execution.verifierSet))||y.taskId!==s.id||y.snapshot.seed!==a.seed||y.snapshot.inputHash!==k(s.input)||k(y.snapshot.definition)!==k(L(o.execution.javascript))||y.snapshot.finalStateHash!==k(y.snapshot.state)||k(a.output)!==k({text:y.output??""})||a.metadata.environmentAttemptHash!==h)throw new Error("Tool evidence differs from the selected task, model or recorded attempt.");if(a.metadata.policySource!==g.policySource)throw new Error("Tool evidence policy source changed.");return{environment:{status:y.status,collected:y.collected,definition:y.snapshot.definition,initialStateHash:y.snapshot.initialStateHash,finalStateHash:y.snapshot.finalStateHash,finalState:y.snapshot.state,events:y.snapshot.events}}}d(_t,"readStarterToolEvidence");async function Te(e,t){if(t.metadata.learning!==void 0)return He.parse(t.metadata.learning);if(t.metadata.rewardBinding===void 0)return;let r=ae.parse(t.metadata.rewardBinding);return e.learningRepository().transaction(t.profileId,async s=>{let a=await ye(s,"binding",r),o=[...new Map(a.sources.map(i=>[`${i.reward.id}:${i.reward.revision}:${i.reward.contentHash}`,i.reward])).values()],n=await Promise.all(o.map(i=>ye(s,"reward",i)));return{binding:a,rewards:n}})}d(Te,"resolveTasksetRewardBinding");async function ua(e,t,r){let s=await Te(e,t);if(!s)return{rewardExecution:void 0,verifierAssets:[]};let a=[...s.rewards.flatMap(i=>i.assets),...be(s.binding,s.rewards).flatMap(i=>i.kind==="custom_verifier"?[i.verifierRef]:i.kind==="model_judge"?[i.rubricRef]:[])],o=t.metadata.learning===void 0?void 0:await wt(r,t);if(o){let i=[...new Map(a.map(u=>{let f=o.learningResources.assets.find(p=>p.id===u.id);if(!f)throw new Error("Imported batch verifier is missing from its package.");return oe(f,u),[f.id,f]})).values()];return{rewardExecution:s,verifierAssets:i}}let n=await e.learningRepository().transaction(t.profileId,async i=>{let u=new Map;for(let f of a){let p=await ke(i,"asset",f.id,1);oe(p,f),u.set(p.id,p)}return[...u.values()]});return{rewardExecution:s,verifierAssets:n}}d(ua,"resolveTasksetTrainingReward");async function gd(e,t,r){let s=await ua(e,t,r.storeDir);if(!s.rewardExecution)return s;let a=be(s.rewardExecution.binding,s.rewardExecution.rewards);if(!s.rewardExecution.binding.sources.some(o=>o.role==="evaluation"&&o.weight>0))throw new Error("Managed training requires a positively weighted evaluation Reward source for retained validation.");if(a.some(o=>o.kind==="custom_verifier"&&o.runtime==="sandbox_process"))throw new Error("Managed preparation requires a qualified isolated sandbox for process verifiers.");if(r.placement!=="remote"||t.environment.kind==="work"||t.capabilities.requiresState||t.capabilities.requiresTools||a.some(o=>o.kind==="human"||o.kind==="model_judge"&&(!o.model||o.calibrationStatus!=="passed"))||t.tasks.some(o=>o.privilegedContextRef!==null)||!s.rewardExecution.binding.sources.some(o=>o.role==="training"&&o.weight>0)||r.hasLearnedPreferenceReward)throw new Error("This Reward binding requires an additional managed execution adapter.");return s}d(gd,"resolveManagedTasksetReward");async function Ls(e,t,r){let s=await Te(e,t);if(!s)throw new Error("The Taskset has no published Reward binding.");let a=be(s.binding,s.rewards);return async({grader:o,task:n,attempt:i,signal:u})=>{let f=t.tasks.find(m=>m.id===n.id);if(!f||k(f)!==k(n))throw new Error("Task differs from the selected immutable Taskset.");let p=null;if(t.environment.entrypoint===J){if(!r)throw new Error("Tool grading requires the execution owner's artifact directory.");p=await _t({store:e,storeDir:r,taskset:t,task:n,attempt:i})}else if(n.privilegedContextRef!==null)throw new Error("This authored Taskset requires a private evaluator-context resolver.");let l=a.find(m=>m.id===o.id);if(l?.kind!=="custom_verifier")throw new Error("The verifier is absent from the published Reward binding.");let c=it.parse({id:n.id,clusterKey:n.clusterKey,split:n.split,input:n.input,expectedOutput:n.expectedOutput,policyVisibleContext:n.policyVisibleContext,privilegedContextRef:n.privilegedContextRef,artifactRefs:[],tags:n.tags});return Ye({repository:e.learningRepository(),scope:t.profileId,grader:l,task:c,evidence:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},evaluatorContext:p,signal:u})}}d(Ls,"createTasksetBindingVerifier");async function Ie(e,t,r){let s=r?ae.parse(r):null;if(s&&s.id!==t)throw new Error("Taskset execution reference differs from the requested Taskset.");let a=s?await e.getTasksetRevision(s.id,s.revision):await e.getTaskset(t);if(!a||s&&(a.contentHash!==s.contentHash||Re(a)!==s.contentHash))throw new Error("Taskset execution revision is unavailable or differs from its immutable hash.");return a}d(Ie,"requireLocalTasksetRevision");async function Sd(e,t,r){let s=await e.releaseForTaskset(t);return s||De({taskset:t,rewardExecution:r?await Te(r,t):void 0,adapterId:qe(t)}).tasksetRelease}d(Sd,"requireReleasedTaskset");import{createHash as ma}from"node:crypto";function Us(e,t,r){if(t.metadata.importedPackageHash!==e.contentHash)throw new Error("The Work projection differs from the admitted published package.");let s=e.taskset.tasks.find(n=>n.id===r.id);if(!s||s.split!==r.split||k(r.metadata.portableTaskRecord)!==k(s))throw new Error("The Work task differs from its exact released population.");let a=0,o=(r.assets??[]).map(n=>{let i=s.artifactRefs.find(l=>l.id===n.id&&l.contentHash===n.sha256&&l.path===n.artifactRef),u=i&&e.files.find(l=>k(l.asset)===k(i));if(!i||!u||i.visibility!=="policy")throw new Error("The Work task requests an unavailable or private evaluator file.");let f=Buffer.from(Wr(u));if(ma("sha256").update(f).digest("hex")!==n.sha256||f.length!==n.sizeBytes||n.split!==r.split)throw new Error("The published Work input manifest changed.");if(a+=f.length,a>25e7)throw new Error("Published Work inputs exceed the admitted 250 MB ceiling.");return{...n,assetId:n.id,storageName:n.stagingPath??n.fileName,bytes:f,sizeBytes:f.length}});if(o.length!==s.artifactRefs.length)throw new Error("The exact policy-visible Work asset population is incomplete.");return o}d(Us,"advancedPublishedWorkAssets");async function Yt(e,t=pt,r=!1){if(e.kind==="hosted_harness")return Yt({kind:"hosted_chat",modelId:e.modelId,...Nr(e)},t);if(e.kind!=="hosted_chat")throw new ne("local_target_not_qualified","Select an explicit model target for local execution.",422);if(e.harness&&!r)throw new ne("local_harness_not_qualified","This released Harness requires its native turn executor; it cannot run as a model-only case.",422);let s=await t();if(s.error)throw new ne("local_model_catalog_unavailable",s.error,422);let a=s.models.find(c=>c.id===e.modelId);if(!a)throw new ne("local_model_unavailable","The selected model is unavailable.",422);let o=Bs(a.raw),n=bt(o),i=Bs(o.capabilities).samplingParameters===!0;if(!i&&(e.temperature!==0||e.topP!==1))throw new ne("local_sampling_unsupported","This model uses provider-managed sampling. Custom temperature or top P requires an explicit supported capability.",422);let u=Ws(o.context_window),f=Ws(o.output_limit);if(e.maxOutputTokens>f)throw new ne("local_model_output_limit","The model cannot accept the configured output limit.",422);let p={providerId:"openpond",modelId:e.modelId,maxOutputTokens:e.maxOutputTokens,...i?{temperature:e.temperature,topP:e.topP}:{},messages:e.messages??[]},l=(u*Math.max(n.inputUsdPerMillionTokens,n.cachedInputUsdPerMillionTokens)+e.maxOutputTokens*n.outputUsdPerMillionTokens)/1e6;if(!(l>0))throw new ne("local_model_charge_unknown","The model must publish an enforceable positive request ceiling.",422);return{model:{...p,configurationHash:k({configuration:p,contextWindow:u,outputLimit:f,pricing:n})},pricing:n,maximumChargeUsd:l}}d(Yt,"prepareLocalExperimentModel");function Nd(e){let t=e.admission.streamFactory?.(e)??e.admission.stream??e.stream??ft;return async function*(r){if(!r.requestId)throw new ne("local_request_identity_missing","Every local model dispatch requires a stable request identity.");let s=e.admission.model;if(r.model!==s.modelId||r.maxTokens!==s.maxOutputTokens||r.temperature!==s.temperature||r.topP!==s.topP)throw new ne("local_model_configuration_conflict","The native model request changed its admitted model, output limit or sampling configuration.",422);let a=xs();a?.assertExecution(e.executionId);let o={teamId:e.teamId,id:e.executionId,ownerId:e.ownerId,requestId:r.requestId};await e.store.reserveLocalCharge({...o,caseId:e.caseId,maximumUsd:e.admission.maximumChargeUsd}),r.signal?.aborted&&(await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),r.signal.throwIfAborted());try{await e.store.markLocalChargeDispatched(o.teamId,o.id,o.requestId,o.ownerId)}catch(p){throw a&&await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),p}await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:"model.dispatch",payload:{requestId:o.requestId,model:r.model,messages:r.messages}});let n=null,i=!1,u=!1,f;try{for await(let p of t(r)){u=!0,p.type==="usage"&&(n=p.usage),p.type==="finish"&&(i=!0);let l=p.type==="text_delta"?{requestId:o.requestId,text:p.text}:p.type==="tool_call_delta"?{requestId:o.requestId,toolCalls:p.toolCalls}:p.type==="usage"?{requestId:o.requestId,usage:ve(p.usage)}:p.type==="finish"?{requestId:o.requestId,finishReason:p.finishReason}:null;l&&await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:`model.${p.type}`,payload:l}),yield p}}catch(p){throw!u&&p instanceof Is&&p.executionId===o.id&&p.requestId===o.requestId&&(f=p),p}finally{let p=ve(n),l=e.admission.pricing&&i&&p.promptTokens!==null&&p.completionTokens!==null?xe(n,e.admission.pricing):null;await e.store.settleLocalCharge({...o,costUsd:l,confirmedTransportNotInvoked:f,usage:n===null?null:{promptTokens:p.promptTokens,completionTokens:p.completionTokens,cachedPromptTokens:p.cachedPromptTokens,uncachedPromptTokens:p.uncachedPromptTokens}})}}}d(Nd,"createLocalBudgetedModelStream");function Bs(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Bs,"record");function Ws(e){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new ne("local_model_bound_unknown","The model catalog is missing an enforceable context/output bound.",422);return e}d(Ws,"positive");import{AsyncLocalStorage as pa}from"node:async_hooks";import fa from"node:path";var Fs=new pa;function Vs(){return Fs.getStore()}d(Vs,"currentAdvancedRefinerBoundary");function Jd(e){let t=Sr(fa.join(e.storeDir,"library","harnesses","advanced-refiner-paid-calls.sqlite"));t.exec(`CREATE TABLE IF NOT EXISTS advanced_refiner_budgets(id TEXT PRIMARY KEY,actor_id TEXT NOT NULL,team_id TEXT NOT NULL,pin_hash TEXT NOT NULL,maximum_spend REAL NOT NULL,maximum_steps INTEGER NOT NULL,expires_at TEXT NOT NULL);
  CREATE TABLE IF NOT EXISTS advanced_refiner_calls(run_id TEXT NOT NULL,id TEXT NOT NULL,intent_hash TEXT NOT NULL,ceiling REAL NOT NULL,state TEXT NOT NULL,actual_spend REAL,result TEXT,PRIMARY KEY(run_id,id));`);function r(a){t.exec("BEGIN IMMEDIATE");try{let o=a();return t.exec("COMMIT"),o}catch(o){throw t.exec("ROLLBACK"),o}}d(r,"transaction");async function s(a,o,n){let i=Ft(o);await e.authorize(i),r(()=>{let l=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(l){if(l.actor_id!==i.actorId||l.team_id!==i.teamId||l.pin_hash!==i.contentHash)throw new Error("Advanced Run budget belongs to another immutable admission.");return}t.prepare("INSERT INTO advanced_refiner_budgets VALUES(?,?,?,?,?,?,?)").run(a,i.actorId,i.teamId,i.contentHash,i.maximumCostUsd,i.maximumModelSteps,new Date(Date.now()+i.maximumDurationMs).toISOString())});let u=t.prepare("SELECT expires_at FROM advanced_refiner_budgets WHERE id=?").get(a),f=Date.parse(String(u.expires_at))-Date.now();if(f<=0)throw new Error("The immutable advanced evaluation duration is exhausted.");let p={runId:a,pin:i,signal:AbortSignal.timeout(f),authorize:d(()=>e.authorize(i),"authorize"),async resolveWorkAssets(l,c){await e.authorize(i),p.signal.throwIfAborted();let m=await e.packageForPin(i),g=Us(m,l,c);return await e.authorize(i),p.signal.throwIfAborted(),g},async call(l,c,m,g,h){if(await e.authorize(i),!Number.isInteger(g)||g<1)throw new Error("Advanced paid requests require a finite enforced token ceiling.");let w=c.model;if(w?.providerId!=="openpond"||!w.modelId)throw new Error("Advanced paid calls require an explicitly qualified hosted model.");let y=w.modelId,S=e.prepareModel?await e.prepareModel(y,g,m):await(async()=>{let C=await Yt({kind:"hosted_chat",modelId:y,maxOutputTokens:g,temperature:0,topP:1});if(k(C.pricing)!==k(m))throw new Error("The actual model pricing differs from its reviewed request pin.");return{configurationHash:C.model.configurationHash,maximumChargeUsd:C.maximumChargeUsd}})();if(!/^[a-f0-9]{64}$/.test(S.configurationHash)||!Number.isFinite(S.maximumChargeUsd)||S.maximumChargeUsd<=0)throw new Error("The actual owner model ceiling is unavailable.");let _=k({intent:c,pricing:m,maximumOutputTokens:g,modelConfigurationHash:S.configurationHash}),x=S.maximumChargeUsd,v=r(()=>{let C=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(!C||String(C.expires_at)<=new Date().toISOString())throw new Error("The immutable advanced evaluation duration is exhausted.");let E=t.prepare("SELECT * FROM advanced_refiner_calls WHERE run_id=? AND id=?").get(a,l);if(E){if(E.intent_hash!==_)throw new Error("Advanced provider request ID retains another exact payload.");if(E.state!=="completed")throw new Error("This paid request has an unknown outcome; recover its actual provider receipt before resuming.");return JSON.parse(String(E.result))}let T=t.prepare("SELECT COUNT(*) AS steps,COALESCE(SUM(COALESCE(actual_spend,ceiling)),0) AS spent,SUM(CASE WHEN state NOT IN ('completed','not_dispatched') THEN 1 ELSE 0 END) AS unknown FROM advanced_refiner_calls WHERE run_id=?").get(a);if(Number(T.unknown)>0)throw new Error("A prior advanced paid request is unsettled; no fresh request is allowed.");if(Number(T.steps)>=i.maximumModelSteps||Number(T.spent)+x>i.maximumCostUsd+1e-9)throw new Error("The advanced evaluation step or cumulative spend ceiling is exhausted.");t.prepare("INSERT INTO advanced_refiner_calls(run_id,id,intent_hash,ceiling,state) VALUES(?,?,?,?,?)").run(a,l,_,x,"reserved")});if(v!==void 0)return await e.authorize(i),p.signal.throwIfAborted(),v;let M=!1;try{if(await e.authorize(i),p.signal.throwIfAborted(),t.prepare("UPDATE advanced_refiner_calls SET state='dispatched' WHERE run_id=? AND id=? AND intent_hash=? AND state='reserved'").run(a,l,_).changes!==1)throw new Error("The retained advanced request admission changed.");M=!0;let E=await h();if(E.costUsd===null||!Number.isFinite(E.costUsd)||E.costUsd<0||E.costUsd>x+1e-9)throw new Error("The actual provider receipt cannot settle its admitted request ceiling.");let T=JSON.stringify(E.value);if(Buffer.byteLength(T)>8*1024*1024)throw new Error("The advanced provider response exceeds the retained output bound.");return r(()=>{if(t.prepare("UPDATE advanced_refiner_calls SET state='completed',actual_spend=?,result=? WHERE run_id=? AND id=? AND intent_hash=? AND state='dispatched'").run(E.costUsd,T,a,l,_).changes!==1)throw new Error("The advanced request settlement changed concurrently.")}),await e.authorize(i),E.value}catch(C){throw M?t.prepare("UPDATE advanced_refiner_calls SET state='unknown' WHERE run_id=? AND id=? AND state='dispatched'").run(a,l):t.prepare("UPDATE advanced_refiner_calls SET state='not_dispatched',actual_spend=0 WHERE run_id=? AND id=? AND state='reserved'").run(a,l),C}}};return Fs.run(p,n)}return d(s,"run"),{run:s,async authorize(a){await e.authorize(Ft(a))},read(a,o){return t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=? AND actor_id=? AND team_id=? AND pin_hash=?").get(a,o.actorId,o.teamId,o.contentHash)?t.prepare("SELECT id,intent_hash,ceiling,state,actual_spend FROM advanced_refiner_calls WHERE run_id=? ORDER BY id").all(a):null},accounting(a,o){let n=this.read(a,o)??[],i=n.filter(f=>f.state!=="completed"&&f.state!=="not_dispatched"),u=n.reduce((f,p)=>f+Number(p.actual_spend??0),0);return{modelRequests:n.filter(f=>f.state!=="not_dispatched").length,knownSpendUsd:u,totalSpendUsd:i.length?null:u,heldUsd:i.reduce((f,p)=>f+Number(p.ceiling),0),uncertainRequests:i.filter(f=>f.state!=="reserved").length}},close(){t.close()}}}d(Jd,"createAdvancedRefinerPaidBoundary");import ra from"node:path";function xt(e,t,r){let s=new Date().toISOString();return W.parse({schemaVersion:"openpond.taskAttempt.v1",id:`fixture_attempt_${t.id}_${r}`,tasksetId:e,taskId:t.taskId,split:Ke(t)??"frozen_eval",attempt:r,seed:0,modelRef:null,startedAt:s,completedAt:s,output:t.output,runtimeEventRefs:[],artifactRefs:[],privilegedOutcomeRef:null,infrastructureError:t.infrastructureError,costUsd:0,latencyMs:0,userInterventions:0,metadata:{fixtureLabel:t.label}})}d(xt,"fixtureAttempt");function Ke(e){let t=e&&typeof e=="object"&&!Array.isArray(e)&&"metadata"in e&&e.metadata&&typeof e.metadata=="object"&&!Array.isArray(e.metadata)?e.metadata.artifactSplit:null;return t==="train"||t==="validation"||t==="test"||t==="frozen_eval"?t:null}d(Ke,"artifactSplit");function Js(e){let t=Gr(e.taskset),r=t.issues.filter(b=>b.severity==="error").map(b=>({code:b.code,message:b.message,path:b.path})),s=t.issues.filter(b=>b.severity==="warning").map(b=>({code:b.code,message:b.message,path:b.path})),a=typeof e.taskset.metadata.trainingMethod=="string"?e.taskset.metadata.trainingMethod:null,o=ga(a),n=o==="sft",i=o==="dpo",u=o==="grpo"||o==="ppo",f=new Set(e.taskset.learningSignals.demonstrations.filter(b=>b.approved&&b.taskId).map(b=>b.taskId)),p=e.taskset.metadata.learning===void 0?null:He.parse(e.taskset.metadata.learning),l=d(b=>p?p.admissions.some(B=>B.taskId===b.id&&B.supervisedTarget!==null):b.expectedOutput!=null,"hasSupervisedTarget"),c=e.taskset.datasetArtifact,m=new Set(zs(e.taskset.metadata.approvedArtifactSignals)),g=!!c?.schema.fields.some(b=>b.semanticRole==="demonstration"||b.semanticRole==="expected_output"||b.semanticRole==="chosen"),h=e.taskset.tasks.filter(b=>b.split==="train"&&(l(b)||u)),w=n?h.filter(b=>f.has(b.id)):h,y=e.taskset.tasks.filter(b=>b.split==="train"&&l(b)&&!f.has(b.id)),S=e.taskset.tasks.filter(b=>b.split==="frozen_eval"&&(b.expectedOutput||u)),_=c&&g?c.splitCounts.train??0:0,x=c&&g?c.splitCounts.frozen_eval??0:0,v=c?m.has(o??""):!1,M=c?v?_:0:w.length,C=e.taskset.learningSignals.preferences.filter(b=>b.approved).length,E=Yr(e.taskset),T=E?E.binding.sources.filter(b=>b.role==="training"&&b.weight>0&&(!b.taskIds||w.some(B=>b.taskIds.includes(B.id)))).length:e.taskset.learningSignals.rewards.filter(b=>b.approved&&b.executable).length,A=i?C:u?T?M:0:M,R=c?x:S.length;A===0&&r.push(n?{code:"sft_demonstrations_missing",message:"At least one explicitly approved training demonstration is required.",path:"learningSignals.demonstrations"}:i?{code:"dpo_preferences_missing",message:"At least one approved chosen/rejected response pair is required for DPO.",path:"learningSignals.preferences"}:{code:"online_reward_missing",message:`${o?.toUpperCase()??"Online training"} requires an executable reward-bearing task.`,path:"learningSignals.rewards"});let j=c?n&&!v?_:0:y.length;n&&j&&r.push({code:"sft_demonstrations_unapproved",message:`${j} training example${j===1?" is":"s are"} not explicitly approved.`,path:c?"metadata.approvedArtifactSignals":"learningSignals.demonstrations"}),R===0&&s.push({code:"frozen_eval_missing",message:"At least one independent evaluation example is recommended before making quality claims.",path:c?"datasetArtifact.splitCounts.frozen_eval":"tasks"});let $=new Set(w.map(b=>b.clusterKey)),z=new Set(S.map(b=>b.clusterKey)),U=c&&e.taskset.metadata.splitIsolationVerified===!0;(c?!U||A===0||R===0:$.size===0||z.size===0||[...$].some(b=>z.has(b)))&&s.push({code:"independent_evaluation_missing",message:"Independent source clusters are recommended before using evaluation results for quality claims.",path:c?"metadata.splitIsolationVerified":"tasks"}),(c?e.taskset.metadata.artifactRowsVerified!==!0:e.taskset.tasks.some(b=>typeof b.metadata.exampleOrigin!="string"))&&r.push({code:"example_provenance_missing",message:"Every training and evaluation example must record whether it was extracted, corrected, synthetic, or expert-authored.",path:c?"metadata.artifactRowsVerified":"tasks.metadata.exampleOrigin"});let Q=ha(e.taskset.metadata.diagnosis);(!Q||typeof Q.summary!="string"||!Array.isArray(Q.stableBehavior))&&s.push({code:"capability_diagnosis_missing",message:"A capability diagnosis is recommended to separate stable behavior from changing knowledge.",path:"metadata.diagnosis"}),Q?.trainingEligible===!1&&s.push({code:"training_not_recommended",message:"The capability diagnosis does not recommend storing this behavior in model weights.",path:"metadata.diagnosis.trainingEligible"});let me=e.taskset.graders.some(b=>b.rewardEligible&&(b.kind!=="model_judge"||b.metadata.calibrationIsAdvisory!==!0));me&&!e.graderAudit&&s.push({code:"grader_audit_missing",message:"Optional grader calibration evidence has not been recorded.",path:"graderFixtures"}),me&&e.graderAudit&&e.graderAudit.tasksetHash!==e.taskset.contentHash&&s.push({code:"grader_audit_stale",message:"Optional grader calibration evidence predates this immutable Taskset revision.",path:"graderFixtures"}),me&&e.graderAudit&&!e.graderAudit.passed&&s.push({code:"grader_audit_failed",message:"The grader did not match every optional calibration fixture's expected outcome.",path:"graderFixtures"}),me&&e.graderAudit&&!e.graderAudit.hackingChecksPassed&&s.push({code:"grader_adversarial_calibration_failed",message:"Adversarial or prompt-injection calibration fixtures did not match their expected grader outcomes.",path:"graders"}),me&&e.graderAudit&&!e.graderAudit.leakageChecksPassed&&r.push({code:"environment_leakage",message:"Environment or privileged-state leakage checks failed.",path:"environment"}),me&&e.graderAudit&&!e.graderAudit.infrastructureSafetyPassed&&r.push({code:"infrastructure_reward",message:"An infrastructure failure produced a score or eligible reward.",path:"graderFixtures"});let tt=e.taskset.graders.some(b=>b.rewardEligible);u&&!tt&&r.push({code:"online_reward_grader_missing",message:`${o?.toUpperCase()} requires a user-selected reward-eligible grader.`,path:"graders"});let ie=o??(A>0&&e.taskset.capabilities.compatibleMethods.includes("sft")?"sft":"none"),I=e.taskset.learningSignals.demonstrations.filter(b=>b.approved).map(b=>b.id),N=ie==="none"?null:{primaryMethod:ie,bootstrap:ie==="grpo"&&I.length?{method:"sft",purpose:"trajectory_bootstrap",demonstrationRefs:I,limitations:["The SFT bootstrap imitates approved trajectories; it does not optimize verifier reward.","Completing the bootstrap does not satisfy the primary GRPO recommendation."]}:null},K=r.length===0,pe=r.length>0,rt=["sft","dpo","grpo","ppo"].map(b=>{let B=[],X=[];b==="sft"&&I.length===0&&(B.push("demonstrations_missing"),X.push("Add at least one approved prompt/response demonstration.")),b==="dpo"&&C===0&&(B.push("preference_pairs_missing"),X.push("Add at least one approved chosen/rejected response pair.")),(b==="grpo"||b==="ppo")&&T===0&&(B.push("executable_reward_missing"),X.push("Implement and validate an executable scalar reward.")),b==="ppo"&&(B.push("value_model_required"),X.push("Choose and bind a value/critic model before PPO can execute.")),R===0&&(B.push("frozen_eval_missing"),X.push("Add an independent frozen-evaluation split.")),pe&&B.length===0&&(B.push("taskset_not_ready"),X.push("Resolve the Dataset validation and grader-audit blockers."));let Me=B.filter(fe=>fe!=="value_model_required");return{method:b,status:Me.length?"needs_dataset_work":b===ie?"recommended":"compatible",reasonCodes:B,reasons:X}});return zr.parse({schemaVersion:"openpond.tasksetReadiness.v1",tasksetId:e.taskset.id,tasksetHash:e.taskset.contentHash,ready:K,recommendedMethod:ie,trainingPath:N,methodReadiness:rt,compatibleDestinationClasses:K?ie==="grpo"?["export","custom","hosted_managed"]:["export","custom"]:["export"],blockers:r,advisories:s,warnings:[...t.issues.filter(b=>b.severity==="warning").map(b=>b.message),...ie==="grpo"?["GRPO is readiness-compatible but not executable in the local CPU fixture."]:[],...zs(e.taskset.metadata.warnings)],generatedAt:e.generatedAt??new Date().toISOString(),metadata:{reportHash:k([e.taskset.contentHash,r])}})}d(Js,"buildTasksetReadiness");function ga(e){return e==="sft"||e==="dpo"||e==="grpo"||e==="ppo"||e==="sdft"||e==="opsd"||e==="sdpo"?e:null}d(ga,"trainingPathMethod");function ha(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}d(ha,"metadataRecord");function zs(e){return Array.isArray(e)?e.filter(t=>typeof t=="string"&&!!t.trim()):[]}d(zs,"metadataStrings");import eo from"node:path";function Gs(e,t,r){let s=He.parse(t.metadata.learning),a=be(s.binding,s.rewards);return async({grader:o,task:n,attempt:i,signal:u})=>{let f=s.admissions.find(h=>h.taskId===n.id),p=a.find(h=>h.id===o.id);if(!f||p?.kind!=="custom_verifier")throw new Error("Prepared batch has no matching verifier admission.");let l=await wt(r,t),c=l?l.learningResources?.evidence.find(h=>Ar(L(h),f.evidence)):await e.learningRepository().transaction(t.profileId,h=>ye(h,"evidence",f.evidence));if(!c)throw new Error("Imported batch evidence is missing from its package.");let m=dt(c,s.definition),g=d(h=>({id:h.id,input:h.input,expectedOutput:h.expectedOutput,policyVisibleContext:h.policyVisibleContext,privilegedContextRef:h.privilegedContextRef}),"selected");if(k(g(n))!==k(g(m)))throw new Error("Prepared task differs from its admitted evaluator context.");if(l){let h=l.learningResources?.assets.find(w=>w.id===p.verifierRef.id);if(!h)throw new Error("Imported batch verifier is missing from its package.");return $e({source:oe(h,p.verifierRef),exportName:p.exportName,timeoutMs:p.timeoutMs,signal:u,value:{task:m,attempt:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},input:m.input,output:i.output,expectedOutput:m.expectedOutput,evaluatorContext:c.submission.evaluatorContext,infrastructureError:i.infrastructureError??null}})}return Ye({repository:e.learningRepository(),scope:t.profileId,grader:p,task:m,evidence:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},evaluatorContext:c.submission.evaluatorContext,signal:u})}}d(Gs,"createLearningBatchVerifier");import{readFile as ka,realpath as Ys}from"node:fs/promises";import Ks from"node:path";async function Xs(e){let t=await Ys(e.allowedRoot),r=await Ys(Ks.resolve(t,e.grader.module));if(r!==t&&!r.startsWith(`${t}${Ks.sep}`))throw new Error("Verifier module is outside the approved Taskset root.");let s=await ka(r);if(e.grader.metadata.portableVerifierRef!==void 0){let o=Tr.parse(e.grader.metadata.portableVerifierRef);if(o.sizeBytes!==s.length||o.contentHash!==se(s))throw new Error("Verifier source differs from its immutable Taskset release.")}let a=new TextDecoder("utf-8",{fatal:!0}).decode(s);return $e({source:a,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:e.grader.timeoutMs,signal:e.signal,value:{task:e.task,attempt:e.attempt,input:e.task.input,expectedOutput:e.task.expectedOutput,output:e.attempt.output,infrastructureError:e.attempt.infrastructureError,...e.evaluatorContext?{evaluatorContext:e.evaluatorContext}:{}}})}d(Xs,"runSandboxedVerifier");import{constants as Kt}from"node:fs";import{open as ya,realpath as Zs}from"node:fs/promises";import We from"node:path";async function It(e,t){let r=De({taskset:t,adapterId:qe(t)});if(typeof t.metadata.importedPackageHash=="string"){let s=await Rs(e,t.metadata.importedPackageHash);if(s.taskset.contentHash!==r.tasksetRelease.contentHash)throw new Error("Evaluation Taskset differs from its pinned portable package.");let a=ct.parse(t.metadata.portableFileInventory??[]),{files:o,contentHash:n,...i}=s,u=await vt({root:We.join(e,"training","tasksets",ee(t)),content:i,sources:o.map(f=>({asset:f.asset,sourcePath:a.find(p=>p.asset.id===f.asset.id)?.sourcePath??f.asset.path})),fileOrder:o.map(f=>f.asset.id)});if(u.contentHash!==s.contentHash)throw new Error("Evaluation package changed during capture.");return u}return _s({storeDir:e,taskset:t,content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:r.tasksetRelease,environment:r.environmentRelease,verifierSet:r.verifierSetRelease}})}d(It,"captureTasksetEvaluationPackage");function Qs(e){return async({grader:t,task:r,attempt:s,signal:a})=>{let o=e.taskset.tasks.find(l=>l.id===r.id),n=e.taskset.graders.find(l=>l.id===t.id);if(!o||k(o)!==k(r)||!n||k(n)!==k(t)||s.tasksetId!==e.taskset.id||s.taskId!==r.id||s.split!==r.split)throw new Error("Process grading differs from the selected Taskset revision or attempt.");let i=e.packageValue.taskset.tasks.find(l=>l.id===r.id),u=e.packageValue.taskset.graders.find(l=>l.id===t.id);if(!i||u?.kind!=="custom_verifier"||u.runtime!=="sandbox_process")throw new Error("The process verifier is absent from the immutable package.");let f=e.fixtures?.get(s.id);if(f&&f.attemptHash!==k(s))throw new Error("Fixture attempt changed before private grading.");let p=f?.artifacts??await wa(e.store,e.storeDir,e.taskset,s);return Es({packageValue:e.packageValue,grader:u,task:i,evidence:{output:s.output,artifactRefs:s.artifactRefs,runtimeEventRefs:s.runtimeEventRefs,infrastructureError:s.infrastructureError,artifacts:p},signal:a,authorize:d(async()=>{a?.throwIfAborted()},"authorize")})}}d(Qs,"createTasksetProcessVerifier");async function wa(e,t,r,s){let a=await e.listTaskAttemptArtifacts({attemptId:s.id}),o=[],n=0;if(new Set(s.artifactRefs).size!==s.artifactRefs.length)throw new Error("Attempt contains duplicate artifact references.");for(let i of s.artifactRefs){let u=a.filter(g=>g.id===i&&g.tasksetId===r.id&&g.taskId===s.taskId&&g.attemptId===s.id);if(u.length!==1)throw new Error("Attempt artifact is not recorded by its execution owner.");let f=u[0];if(f.kind!=="output_artifact")continue;if(o.length>=16||(n+=f.sizeBytes)>2e7)throw new Error("Attempt artifacts exceed their byte boundary.");let p=await Zs(t),l=await Zs(We.dirname(f.path));if(![We.join(p,"training/evaluation-artifacts"),We.join(p,"work/outputs")].some(g=>l===g||l.startsWith(`${g}${We.sep}`)))throw new Error("Attempt artifact is outside the execution owner's output directories.");let c=await ya(f.path,Kt.O_RDONLY|Kt.O_NOFOLLOW|Kt.O_NONBLOCK),m;try{let g=await c.stat();if(!g.isFile()||g.size!==f.sizeBytes)throw new Error("Owner-recorded artifact size changed.");m=Buffer.alloc(g.size);let h=0;for(;h<m.length;){let w=await c.read(m,h,m.length-h,h);if(!w.bytesRead)throw new Error("Owner-recorded artifact changed during capture.");h+=w.bytesRead}if((await c.read(Buffer.alloc(1),0,1,h)).bytesRead)throw new Error("Owner-recorded artifact changed during capture.")}finally{await c.close()}if(se(m)!==f.sha256)throw new Error("Owner-recorded artifact bytes changed.");o.push({reference:{id:f.id,contentHash:f.sha256,mediaType:f.mediaType??"application/octet-stream",sizeBytes:f.sizeBytes},title:We.basename(typeof f.metadata.requiredOutputPath=="string"?f.metadata.requiredOutputPath:f.path),contentType:f.mediaType??"application/octet-stream",base64:m.toString("base64")})}return o}d(wa,"readOwnedOutputArtifacts");async function Et(e,t,r){if(!t.graders.some(p=>p.kind==="custom_verifier"))return;let s=t.graders.some(p=>p.kind==="custom_verifier"&&p.runtime==="sandbox_process"),a=t.metadata.learning!==void 0?await Gs(e.store,t,e.storeDir):t.metadata.rewardBinding!==void 0?await Ls(e.store,t,e.storeDir):void 0;if(a&&!s)return a;let o=e.storeDir?null:await(e.loadProfileState??br)(),n=e.storeDir?eo.join(e.storeDir,"training","tasksets",ee(t)):o?.sourcePath?eo.join(o.sourcePath,"tasksets",t.id):null,i=typeof t.metadata.creationSnapshotId=="string"?t.metadata.creationSnapshotId:null,u=i?await e.store.getTaskDesignProposal(i):null;if(n&&t.purpose!=="benchmark"&&t.environment.metadata.runtimeSourceTasksetId===void 0&&await ut(t,n,{generatedFiles:u?.generatedFiles??[]}),s&&!e.storeDir)throw new Error("Process grading requires the execution owner's package directory.");let f=s?Qs({store:e.store,storeDir:e.storeDir,taskset:t,packageValue:r?.packageValue??await It(e.storeDir,t),fixtures:r?.fixtures}):void 0;return n?async({grader:p,task:l,attempt:c,signal:m})=>{let g=t.graders.find(w=>w.id===p.id);if(!g||g.kind!=="custom_verifier")throw new Error("Verifier is not declared by this Taskset revision.");if(g.runtime==="sandbox_process")return f({grader:p,task:l,attempt:c,signal:m});if(a)return a({grader:g,task:l,attempt:c,signal:m});let h;if(t.environment.entrypoint===J&&!c.infrastructureError){if(!e.storeDir)throw new Error("Tool grading requires the execution owner's artifact directory.");h=await _t({store:e.store,storeDir:e.storeDir,taskset:t,task:l,attempt:c})}return Xs({grader:g,task:l,attempt:c,allowedRoot:n,signal:m,evaluatorContext:h})}:void 0}d(Et,"createTasksetEvaluationVerifier");var va=new Set(["correct","incorrect","parse_failure","budget_exhausted","tool_schema_violation","infrastructure_failure","cancelled"]);async function to(e){let t=Sa(e.attempt);if(t){let f=k(e.graders),p=`Taskset Work attempt ended with ${t.failureClass}: `+t.message;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${k([e.attempt.id,f,t]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:f,score:null,passed:!1,components:e.graders.map(l=>({graderId:l.id,graderVersion:l.version,score:0,passed:!1,hardGate:l.hardGate,rewardEligible:!1,feedback:p,evidenceRefs:e.attempt.artifactRefs,judge:l.kind==="model_judge"?l.judge:null,calibrationStatus:l.kind==="model_judge"?l.calibrationStatus:"not_applicable"})),failureClass:t.failureClass,feedback:[p],rewardEligible:!1,createdAt:e.now?.()??new Date().toISOString()}}let r=e.learning?null:ba(e.attempt.metadata);if(!r)try{return await lt(e)}catch(f){return Ta({attempt:e.attempt,graders:e.graders,now:e.now,error:f})}let s=k(e.graders),a=r.rewardEligible?r.outcome==="correct"?1:0:null,o=r.outcome==="correct",n=Ra(r.outcome),i=`Cross-System trajectory verifier outcome: ${r.outcome}.`,u=e.graders.map(f=>({graderId:f.id,graderVersion:f.version,score:a??0,passed:o,hardGate:f.hardGate,rewardEligible:r.rewardEligible&&f.rewardEligible,feedback:i,evidenceRefs:e.attempt.artifactRefs,judge:f.kind==="model_judge"?f.judge:null,calibrationStatus:f.kind==="model_judge"?f.calibrationStatus:"not_applicable"}));return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${k([e.attempt.id,s,r]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:s,score:a,passed:o,components:u,failureClass:n,feedback:[i],rewardEligible:r.rewardEligible,createdAt:e.now?.()??new Date().toISOString()}}d(to,"gradeTasksetEvaluationAttempt");function Ta(e){let t=k(e.graders),r=(e.error instanceof Error?e.error.message:String(e.error)).slice(0,19e3),s=`Taskset grader failed: ${r}`;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${k([e.attempt.id,t,"grader_failure",r]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:t,score:null,passed:!1,components:e.graders.map(a=>({graderId:a.id,graderVersion:a.version,score:0,passed:!1,hardGate:a.hardGate,rewardEligible:!1,feedback:s,evidenceRefs:e.attempt.artifactRefs,judge:a.kind==="model_judge"?a.judge:null,calibrationStatus:a.kind==="model_judge"?a.calibrationStatus:"not_applicable"})),failureClass:"grader_failure",feedback:[s],rewardEligible:!1,createdAt:e.now?.()??new Date().toISOString()}}d(Ta,"graderFailureResult");function Sa(e){if(e.metadata.execution!=="taskset_work"||!e.infrastructureError)return null;let t=Zr.safeParse(e.metadata.failureClass);if(!t.success||t.data==="policy_failure"||t.data==="grader_failure")throw new Error("Taskset Work attempt is missing a trusted execution failure class.");return{failureClass:t.data,message:e.infrastructureError}}d(Sa,"tasksetWorkFailure");function ba(e){if(e.execution!=="post_training_evaluation_tool_loop")return null;let t=e.verifierOutcome,r=e.verifierReward,s=e.verifierRewardEligible;if(typeof t!="string"||!va.has(t)||typeof s!="boolean"||r!==null&&typeof r!="number")throw new Error("Cross-System evaluation attempt is missing trusted verifier metadata.");if(s!==(t==="correct"||t==="incorrect"||t==="parse_failure")||s&&typeof r!="number"||!s&&r!==null)throw new Error("Cross-System evaluation verifier metadata is internally inconsistent.");return{outcome:t,reward:r,rewardEligible:s}}d(ba,"crossSystemGrade");function Ra(e){return e==="correct"?null:e==="infrastructure_failure"?"infrastructure_failure":e==="cancelled"?"cancelled":"policy_failure"}d(Ra,"failureClassFor");import Xe from"node:path";import{mkdir as ro,readFile as _a,writeFile as so}from"node:fs/promises";async function te(e){let t=Xe.join(e.storeDir,"training","evaluation-artifacts",e.tasksetId),r=Xe.join(t,`${e.attemptId}-${oo(e.fileLabel??e.kind)}.json`),s=Buffer.from(`${JSON.stringify({schemaVersion:"openpond.rawEvaluationArtifact.v1",requestId:e.requestId,...e.payload},null,2)}
 `,"utf8");return await ro(t,{recursive:!0}),await so(r,s,{mode:384}),Xt({...e,file:r,bytes:s,mediaType:"application/json"})}d(te,"persistJsonTaskAttemptArtifact");async function Zc(e){let t=e.extension.replace(/^\.+/,"");if(!/^[a-z0-9]{1,12}$/i.test(t))throw new Error("Task Attempt artifact extension is invalid.");let r=Xe.join(e.storeDir,"training","evaluation-artifacts",e.tasksetId);await ro(r,{recursive:!0});let s=Xe.join(r,`${e.attemptId}-${oo(e.fileLabel)}.${t}`);return await so(s,e.bytes,{mode:384}),Xt({...e,file:s})}d(Zc,"persistBinaryTaskAttemptArtifact");function oo(e){let t=e.trim().replace(/[^a-zA-Z0-9._-]+/g,"-").replace(/^-+|-+$/g,"");if(!t)throw new Error("Evaluation artifact file label is empty.");return t}d(oo,"safeFileLabel");async function ao(e){let t=await _a(e.path),r=se(t);if(t.byteLength!==e.expectedSizeBytes||r!==e.expectedSha256)throw new Error(`Saved Work output ${Xe.basename(e.path)} no longer matches its OutputRef.`);return Xt({...e,file:e.path,bytes:t,kind:"output_artifact",metadata:{...e.metadata,source:"openpond-work",outputValidated:!0}})}d(ao,"persistTaskAttemptOutputArtifact");async function Xt(e){let t=se(e.bytes),r=Qr.parse({schemaVersion:"openpond.taskAttemptArtifact.v1",id:`attempt_artifact_${k([e.attemptId,e.kind,e.file,t]).slice(0,24)}`,tasksetId:e.tasksetId,taskId:e.taskId,attemptId:e.attemptId,kind:e.kind,path:e.file,mediaType:e.mediaType,sha256:t,sizeBytes:e.bytes.byteLength,createdAt:e.timestamp(),metadata:{requestId:e.requestId,localOnly:!0,containsPrivilegedOutcome:!1,...e.metadata}});return e.store.saveTaskAttemptArtifact(r)}d(Xt,"saveArtifact");function Zt(e){let t=new Set(e.approvedTrajectoryIds),r=new Map(e.tasks.map(a=>[a.id,a])),s=new Map(e.results.map(a=>[a.trajectoryId,a]));return e.trajectories.flatMap(a=>{let o=s.get(a.id),n=r.get(a.taskId);return!t.has(a.id)||!n||o?.outcome!=="correct"||!o.rewardEligible?[]:[us.parse({schemaVersion:V,id:`cso_bootstrap_${a.id}`,taskId:n.id,worldId:a.worldId,trajectoryId:a.id,toolContractHash:q,approval:{status:"approved",approvedBy:e.approvedBy,approvedAt:e.approvedAt},messages:xa(n,a)})]})}d(Zt,"buildCrossSystemBootstrapDataset");function xa(e,t){let r=[{role:"system",content:cs},{role:"user",content:e.prompt}];for(let s of t.steps)s.kind==="model"&&s.content.trim()&&r.push({role:"assistant",content:s.content}),s.kind==="tool_call"&&r.push({role:"assistant",content:null,tool_calls:[{id:s.callId,type:"function",function:{name:s.name,arguments:JSON.stringify(s.arguments)}}]}),s.kind==="tool_result"&&r.push({role:"tool",tool_call_id:s.callId,content:JSON.stringify({ok:s.ok,result:s.result,error:s.error})}),s.kind==="final"&&r.push({role:"assistant",content:s.content});return r}d(xa,"trajectoryMessages");import _n from"node:path";import{mkdir as Ia,mkdtemp as Ea,rm as Ca}from"node:fs/promises";import Aa from"node:path";async function no(e,t,r){await Ia(e,{recursive:!0});let s=await Ea(Aa.join(e,".authoring-"));try{return await ut(t,s,{generatedFiles:r}),await Qt(e,t,s)}finally{await Ca(s,{recursive:!0,force:!0})}}d(no,"materializeGeneratedTasksetPackage");async function Qt(e,t,r){let s=await Kr(r),a=await Ss(t,r),o=`authored-${k(a.taskset)}`,n=we.parse({...a.taskset,environment:{...a.taskset.environment,metadata:{...a.taskset.environment.metadata,runtimeSourceTasksetId:o}}}),i=we.parse({...n,contentHash:Re(n)});return await ys(e,{taskset:i,generatedFiles:a.generatedFiles},o,{source:{directory:r,packageHash:s},verify:d(u=>yt(u,i),"verify")}),i}d(Qt,"materializeTasksetRevisionFromSource");import{Buffer as tr}from"node:buffer";import{spawn as Ma}from"node:child_process";var io=String.raw`
 import contextlib, io, json, math, statistics, decimal, datetime, collections, itertools, functools, sys