openpond 0.2.48 → 0.2.50

package/dist/chunks/{chunk-DFM5JTS2.js → chunk-ZCWHVLJ4.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as H,f as te,h as ae,i as ie}from"./chunk-VOIH2BQA.js";import{g as se}from"./chunk-QJRXTIOC.js";import{F as re,H as oe}from"./chunk-7K72KS5C.js";import{U as X,ba as Y,ca as Z,ga as ee}from"./chunk-7MXZJQS7.js";import{L as K,da as b}from"./chunk-XDMUD4FW.js";import{Ca as g,Ta as U,pc as Q,xa as P}from"./chunk-QRWPEW3Z.js";import{s as G,t as F,u as T}from"./chunk-WV2APYWG.js";import{Qa as L,na as N,oa as y,qa as W,wa as $}from"./chunk-PJM2G7SG.js";import{Fa as _}from"./chunk-SDEDAHV6.js";import{b as R,c as v,d as J,l as S,m as n,o as j,r as q}from"./chunk-NNF6KCHD.js";import{b as h}from"./chunk-ES2ISFJC.js";import{a as d}from"./chunk-FTQLQJQV.js";import{randomUUID as V,createHash as fe}from"node:crypto";import{mkdtemp as ue,mkdir as pe,writeFile as D,rm as me}from"node:fs/promises";import he from"node:os";import E from"node:path";var ne=d(async e=>{let a=g(e.packageValue);if(e.grader.runtime!=="sandbox_process")throw new Error("Select the sandbox process runtime for a local file grader.");if(!a.taskset.graders.some(r=>n(r)===n(e.grader)))throw new Error("The private grader differs from its package.");if(!a.taskset.tasks.some(r=>n(r)===n(e.task)))throw new Error("The private case differs from its package.");await e.authorize(),e.signal?.throwIfAborted();let i=await ue(E.join(he.tmpdir(),"profile-private-grade-")),t=V();try{for(let f of a.files.filter(l=>l.asset.visibility!=="policy")){let l=f.asset.path;if(E.isAbsolute(l)||l.split(/[\\/]/).some(u=>!u||u==="."||u===".."))throw new Error("Private grader asset path invalid.");let m=E.join(i,"package",l);await pe(E.dirname(m),{recursive:!0,mode:448}),await D(m,P(f),{mode:256})}await D(E.join(i,"request.json"),JSON.stringify({task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,artifacts:"artifacts"in e.evidence?e.evidence.artifacts:[],expectedOutput:e.task.expectedOutput}),{mode:256});let r=`import fs from 'node:fs/promises'; const module=await import(${JSON.stringify("./package/"+e.grader.verifierRef.path)}); const result=await module[${JSON.stringify(e.grader.exportName??"verify")}](JSON.parse(await fs.readFile('request.json','utf8'))); console.log(JSON.stringify(result));`;await D(E.join(i,"runner.mjs"),r,{mode:256});let s=await ie({authorize:d(async()=>(await e.authorize(),e.signal?.throwIfAborted(),{candidateId:t,candidateRevision:1,ownerId:t,sessionId:t,turnId:t,sourceRoot:i,writablePaths:[]}),"authorize")})({candidateId:t,sessionId:t,turnId:t,expectedRevision:1,command:"/runtime/node --max-old-space-size=512 runner.mjs",timeoutMs:e.grader.timeoutMs,maxOutputBytes:7e4,signal:e.signal});if(await e.authorize(),e.signal?.throwIfAborted(),s.code!==0||s.timedOut||s.stdoutTruncated||s.stderrTruncated)throw new Error("Private grading sandbox did not complete.");return H.parse(JSON.parse(s.stdout))}finally{await me(i,{recursive:!0,force:!0})}},"executeLocalProfileVerifier");function Me(e){return async a=>{if(!a.manifest)throw new Error("The hosted grader requires its admitted manifest.");await a.authorize(),a.signal?.throwIfAborted();let i="artifacts"in a.evidence&&Array.isArray(a.evidence.artifacts)?a.evidence.artifacts.map(({reference:l,title:m,contentType:u})=>({reference:l,title:m,contentType:u})):[],t="caseOwner"in a.evidence?a.evidence.caseOwner:null;if(!t?.sessionId||!t.turnId)throw new Error("The private grader has no sealed case owner.");let r=Buffer.from(JSON.stringify({manifest:a.manifest,packageValue:a.packageValue,graderId:a.grader.id,taskId:a.task.id,evidence:{...a.evidence,artifacts:i}}));if(r.length>2e7)throw new Error("The private grading package exceeds its transport limit.");let o={callId:V(),sessionId:t.sessionId,turnId:t.turnId,sha256:fe("sha256").update(r).digest("hex"),sizeBytes:r.length,timeoutMs:a.grader.timeoutMs},s=d((l,m=12e4)=>e.request(oe.parse({contractVersion:re,requestId:V(),operation:"profile-evaluations/grade",params:l}),m),"request");for(let l=0;l<r.length;l+=98304)await a.authorize(),a.signal?.throwIfAborted(),await s({...o,action:"upload",offset:l,contentsBase64:r.subarray(l,l+98304).toString("base64")});let f=await s({...o,action:"execute"},a.grader.timeoutMs+12e4);return await a.authorize(),a.signal?.throwIfAborted(),H.parse(f)}}d(Me,"createHostedProfileVerifier");async function ge(e,a){let i=g(e);for(let r of i.taskset.graders){if(r.kind==="model_judge"&&(!a||r.calibrationStatus!=="passed"))throw new Error("This Profile requires its current calibrated judge and bounded private provider owner.");let o="verifierRef"in r?r.verifierRef:"rubricRef"in r?r.rubricRef:null;o&&A(i,o)}let t=i.taskset.metrics?.customAggregator;if(t){let r=i.files.find(o=>o.asset.path===t.module);if(!r||r.asset.visibility==="policy"||S(P(r))!==t.contentHash)throw new Error("The authored aggregate metric is not its exact private release.");new TextDecoder("utf8",{fatal:!0}).decode(P(r))}}d(ge,"preflightProfilePrivateGrading");function A(e,a){let i=e.files.find(t=>t.asset.id===a.id);if(!i||i.asset.visibility==="policy"||n(i.asset)!==n(a))throw new Error("The exact private grader asset is unavailable.");return new TextDecoder("utf8",{fatal:!0}).decode(P(i))}d(A,"readPrivate");async function $e(e){let a=g(e.packageValue);if(a.contentHash!==e.manifest.packageHash||a.taskset.contentHash!==e.manifest.tasksetRelease.contentHash)throw new Error("The private grading package changed its frozen Run identity.");await e.authorize(),await ge(a,!!e.executeBudgetedJudge),await e.authorize();let i=d(async({grader:o,task:s,evidence:f})=>{let l=e.signal;await e.authorize(),l?.throwIfAborted();let m={manifest:e.manifest,packageValue:a,grader:o,task:s,evidence:f,signal:l,authorize:e.authorize},u=e.executeSandboxVerifier?await e.executeSandboxVerifier(m):o.runtime==="sandbox_process"?await ne(m):await te({source:A(a,o.verifierRef),exportName:o.exportName,runtime:o.runtime,timeoutMs:o.timeoutMs,signal:l,value:{task:s,attempt:f,input:s.input,output:f.output,artifacts:"artifacts"in f?f.artifacts:[],expectedOutput:s.expectedOutput,infrastructureError:f.infrastructureError??null}});return await e.authorize(),l?.throwIfAborted(),{score:u.score,passed:u.passed,rewardEligible:o.rewardEligible,failureClass:null,feedback:[u.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:u.evidenceRefs}},"customVerifier"),t=a.taskset.metrics?.customAggregator,r=t?new TextDecoder("utf8",{fatal:!0}).decode(P(a.files.find(o=>o.asset.path===t.module))):void 0;return{authorize:e.authorize,customVerifier:i,...e.executeBudgetedJudge?{modelJudge:L({readRubric:d(async o=>{await e.authorize();let s=A(a,o);return await e.authorize(),s},"readRubric"),executeBudgeted:d(async(o,s)=>{await e.authorize(),s?.throwIfAborted();let f=await e.executeBudgetedJudge(o,s);return await e.authorize(),s?.throwIfAborted(),f},"executeBudgeted")})}:{},...r!==void 0?{metricSource:r,metricExecutor:d(async o=>{await e.authorize(),o.signal?.throwIfAborted();let s=await ae(o);return await e.authorize(),o.signal?.throwIfAborted(),s},"metricExecutor")}:{}}}d($e,"createProfilePrivateGradingOwner");import{promises as ve}from"node:fs";import ke from"node:path";async function at(e){let a=e.definition.tasksetRelease,i=await se(e.store,e.harnessRelease);if(!i)throw new Error("Evaluation Profile Harness release is unavailable.");let t=await Re({...e,runtime:i});if(t)return t;let r=await e.store.listTasksetRevisions(e.profileId);for(let o of r){if(a.id!==o.id&&a.id!==`taskset-release-${o.id}-r${o.revision}`)continue;let s=await we(e.storeDir,o);if(s.taskset.id===a.id&&s.taskset.contentHash===a.contentHash)return s}throw new Error(`Released Taskset ${a.id} is unavailable in Profile ${e.profileId}.`)}d(at,"loadLocalProfileEvaluationTaskset");async function we(e,a){let i=a.metadata.importedPackageHash;if(typeof i=="string"){let o=await Z(e,i);if(o.taskset.id!==a.id||o.taskset.revision!==a.revision)throw new Error("Imported evaluation Taskset differs from its local Profile revision.");return o}if(U(a)!==a.contentHash)throw new Error("Authored evaluation Taskset differs from its immutable revision.");let t=Q({taskset:a,adapterId:X(a)}),r=await ee({storeDir:e,taskset:a,content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:t.tasksetRelease,environment:t.environmentRelease,verifierSet:t.verifierSetRelease}});return await Y(e,r),r}d(we,"loadCandidate");async function Re(e){let a=e.definition.tasksetRelease,i=e.runtime;if(i.release.harnessRelease.id!==e.harnessRelease.id||i.release.harnessRelease.contentHash!==e.harnessRelease.contentHash)throw new Error("Evaluation Taskset runtime differs from its requested release.");let t=`evals/tasksets/${a.contentHash}.json`,r=i.release.harnessRelease.files.find(o=>o.path===t);if(r){if(r.visibility!=="verifier")throw new Error("Evaluation Taskset package is not verifier-private.");let o=await ve.readFile(ke.join(i.release.bundlePath,"source",...t.split("/")));if(S(o)!==r.contentHash)throw new Error("Released evaluation Taskset package bytes differ from their source receipt.");let s=g(JSON.parse(o.toString("utf8")));if(s.taskset.id!==a.id||s.taskset.contentHash!==a.contentHash)throw new Error("Released evaluation Taskset package differs from its definition.");return s}return null}d(Re,"loadReleasedProfileEvaluationTaskset");function C(e){return{id:W(e),label:"Selected Dataset evaluation",description:"Explicit sealed external Dataset evaluation recipe.",target:e.target,tasksetRelease:{id:e.dataset.id,contentHash:e.dataset.contentHash},split:e.split,taskIds:[...new Set(e.population.map(a=>a.taskId))],seeds:[...new Set(e.population.map(a=>a.seed))],criterion:e.criterion}}d(C,"externalProfileEvaluationDefinition");async function lt(e){let a=e.manifest.profileEvaluation,i=a?.externalDatasetBinding?y(a.externalDatasetBinding):null;if(!i)return null;if(!e.resolveExternalDataset)throw new Error("This owner has no authorized external Dataset resolver.");let t=await e.resolveExternalDataset(i,e.selected);await t.authorize();let r=g(t.packageValue),o=await e.loadCatalog({ref:e.selected.profileRef,sourceRevision:e.selected.sourceRevision,harnessRelease:e.selected.harnessRelease});if(n(t.binding)!==n(i)||i.profileId!==e.selected.profileRef.profileId||i.declaredProfileCatalogHash!==o.catalogHash||r.contentHash!==i.packageHash||n(r.taskset)!==n(e.taskset)||i.environmentHash!==n(r.taskset.environment)||i.privateDatasetClosureHash!==n(r.files.filter(l=>l.asset.visibility!=="policy").map(l=>l.asset)))throw new Error("External Dataset admission changed its actual Profile, package or private closure.");let s=C(i),f={schemaVersion:"openpond.profileEvaluations.v1",definitions:[s],suites:[]};if(T(e.manifest,e.taskset,f),!a||a.catalogHash!==n(f)||a.definitionHash!==n(s)||n(a.target)!==n(i.target))throw new Error("External Dataset manifest changed its derived definition.");return await t.authorize(),{...t,packageValue:r,definition:s,catalog:f}}d(lt,"admitProfileExternalDataset");var Pe=h.object({id:R,createdAt:J,definitionId:R,externalDatasetBinding:N.optional(),profileRef:K.optional(),profileSource:h.object({sourceRevision:h.string().min(1).max(500),harnessRelease:h.object({id:R,contentHash:v}).strict()}).strict().optional(),modelRef:q,hostModelConfigurationHash:v.optional(),hostArtifactHash:v.optional(),maximumSpendUsd:h.number().positive().max(1e4).optional(),expectedManifestHash:v.optional(),hostExperiment:h.object({definition:h.object({id:R,revision:h.number().int().positive(),contentHash:v}).strict().nullable(),configurationHash:v,project:h.object({id:R,revision:h.number().int().positive(),contentHash:v,targetId:R.nullable()}).strict().optional()}).strict().optional()}).strict();function vt(e){return async(a,i)=>{let t=Pe.parse(a);if(i?.requireExpectedManifestHash&&!t.expectedManifestHash)throw new Error("Profile evaluation execution requires the reviewed manifest hash.");if(t.profileSource&&!t.profileRef)throw new Error("An exact Profile source requires its accepted Profile reference.");let r=i?.selectedWorkflows??await e.selectedWorkflows(t.profileRef,t.profileSource);if(t.profileRef&&n(t.profileRef)!==n(r.profileRef))throw new Error("Profile preparation did not resolve the explicitly selected Profile.");if(t.profileSource&&(r.sourceRevision!==t.profileSource.sourceRevision||n(r.harnessRelease)!==n(t.profileSource.harnessRelease)))throw new Error("Profile preparation changed its explicitly pinned source.");let o=t.externalDatasetBinding?y(t.externalDatasetBinding):null,s=o?await(e.resolveExternalDataset?e.resolveExternalDataset(o,r):Promise.reject(new Error("This host has not qualified external Dataset Profile execution."))):null;if(s&&(await s.authorize(),n(y(s.binding))!==n(o)))throw new Error("External Dataset admission differs from the actual authorized recipe.");let f=await e.loadCatalog({ref:r.profileRef,sourceRevision:r.sourceRevision,harnessRelease:r.harnessRelease});if(o&&o.declaredProfileCatalogHash!==f.catalogHash)throw new Error("The original Profile verifier catalog changed.");let l=o?C(o):null;if(l&&t.definitionId!==l.id)throw new Error("Use the external recipe's own definition identity.");let m={schemaVersion:"openpond.profileEvaluations.v1",definitions:l?[l]:f.definitions,suites:l?[]:f.suites},u=l??f.definitions.find(c=>c.id===t.definitionId);if(!u)throw new Error(`Evaluation ${t.definitionId} is absent from the selected Profile release.`);let k=u.target,z=k.kind==="workflow"?r.workflows.find(c=>c.binding.workflowId===k.workflowId)?.binding:_.parse({schemaVersion:"openpond.profileComponentBinding.v1",profileId:r.profileRef.profileId,sourceRevision:r.sourceRevision,harnessRelease:r.harnessRelease,target:k});if(!z)throw new Error(`Evaluation workflow ${k.kind==="workflow"?k.workflowId:""} is absent from the selected Profile release.`);let w=g(s?.packageValue??await e.loadTasksetPackage(u,r.profileRef.profileId,r.harnessRelease)),p=w.taskset;if(p.id!==u.tasksetRelease.id||p.contentHash!==u.tasksetRelease.contentHash)throw new Error("Evaluation Taskset package differs from its released definition.");let le=p.environment.kind==="text"&&p.tools.length===0,M=p.environment.kind==="work"&&p.environment.entrypoint==="openpond-work-v1"&&p.tools.every(c=>["work_exec","work_save_output"].includes(c.name));if(!le&&!M)throw new Error("Profile evaluation supports text cases or Work cases with the built-in execution and output tools.");if(t.maximumSpendUsd!==void 0&&M&&!e.boundedWorkComputeAvailable)throw new Error("This host has not admitted budgeted Work compute.");if(p.tasks.some(c=>c.artifactRefs.length>b.maxAttachments||c.artifactRefs.some(x=>x.visibility!=="policy"||x.mediaType!=="application/pdf"||x.sizeBytes>b.maxAttachmentBytes)))throw new Error("Profile evaluation accepts only bounded policy-visible PDF task attachments.");if(!o&&!e.privateGradingPreflight&&(p.graders.some(c=>c.kind==="model_judge"||c.kind==="custom_verifier")||p.metrics?.aggregation==="custom"))throw new Error("Profile evaluation cannot run this Taskset's model judge, custom verifier, or custom metric in the current app-server runtime.");await e.privateGradingPreflight?.(w);let B={adapterId:k.kind==="workflow"?"openpond.profile-workflow":"openpond.profile-component",placement:e.placement,runtimeVersion:"app-server-v1",capabilityReceipt:n({adapterId:k.kind==="workflow"?"openpond.profile-workflow":"openpond.profile-component",packageHash:w.contentHash,connectedAppScopes:p.policy.connectedAppScopes})},ce={...$({catalog:m,definitionId:u.id,profileId:r.profileRef.profileId,sourceRevision:r.sourceRevision,harnessRelease:r.harnessRelease,environmentHash:n({packageHash:w.contentHash,runtimeTarget:B})}),...o?{externalDatasetBinding:o}:{},...e.sourceCandidate?{sourceCandidate:e.sourceCandidate}:{}};if(o&&(o.packageHash!==w.contentHash||o.environmentHash!==n(p.environment)||o.privateDatasetClosureHash!==n(w.files.filter(c=>c.asset.visibility!=="policy").map(c=>c.asset))))throw new Error("External Dataset package/environment/private closure changed.");let de=j.parse({provider:t.modelRef.providerId,model:t.modelRef.modelId,revision:null,artifactHash:t.hostArtifactHash??null,tokenizerRevision:null,chatTemplateHash:null}),O=await e.modelConfigurationHash(t.modelRef,t),I=G({schemaVersion:"openpond.tasksetRunManifest.v1",id:t.id,tasksetRelease:u.tasksetRelease,packageHash:w.contentHash,execution:{kind:"harness",harnessRelease:r.harnessRelease},profileEvaluation:ce,policy:{kind:"model",model:de,configurationHash:O},gradingRole:"evaluation",metricPolicy:F(p),population:o?o.population.map(c=>({...c,receiptId:`evaluation-${n([t.id,c.taskId,c.seed,c.fixtureId]).slice(0,32)}`})):u.taskIds.flatMap(c=>u.seeds.map(x=>({receiptId:`evaluation-${n([t.id,c,x]).slice(0,32)}`,taskId:c,seed:x,fixtureId:null}))),runtimeTarget:B,limits:{maxTurns:128,timeoutMs:p.environment.defaultTimeoutMs,maxOutputBytes:25e7,maximumSpendUsd:t.maximumSpendUsd??null},createdAt:t.createdAt,metadata:{sourceTasksetId:p.metadata.sourceTasksetId??p.id,...r.profileRef.source==="openpond_git"?{profileRepositoryId:r.profileRef.repositoryId}:{},...t.hostExperiment?{experimentDefinition:t.hostExperiment.definition,experimentConfigurationHash:t.hostExperiment.configurationHash,...t.hostExperiment.project?{project:t.hostExperiment.project}:{}}:{}}});if(T(I,p,m),s&&await s.authorize(),t.expectedManifestHash&&I.contentHash!==t.expectedManifestHash)throw new Error("Profile evaluation setup changed since preview; prepare the run again.");return{manifest:I,taskset:p,profileRef:r.profileRef,binding:z,modelRef:t.modelRef,modelConfigurationHash:O}}}d(vt,"createProfileEvaluationRunPreparationService");export{ne as a,Me as b,ge as c,$e as d,at as e,Re as f,lt as g,vt as h};
+import{a as H,f as te,h as ae,i as ie}from"./chunk-OE2VF6KU.js";import{g as se}from"./chunk-QJRXTIOC.js";import{F as re,H as oe}from"./chunk-7K72KS5C.js";import{U as X,ba as Y,ca as Z,ga as ee}from"./chunk-7MXZJQS7.js";import{L as K,da as b}from"./chunk-XDMUD4FW.js";import{Ca as g,Ta as U,pc as Q,xa as P}from"./chunk-QRWPEW3Z.js";import{s as G,t as F,u as T}from"./chunk-WV2APYWG.js";import{Qa as L,na as N,oa as y,qa as W,wa as $}from"./chunk-PJM2G7SG.js";import{Fa as _}from"./chunk-SDEDAHV6.js";import{b as R,c as v,d as J,l as S,m as n,o as j,r as q}from"./chunk-NNF6KCHD.js";import{b as h}from"./chunk-ES2ISFJC.js";import{a as d}from"./chunk-FTQLQJQV.js";import{randomUUID as V,createHash as fe}from"node:crypto";import{mkdtemp as ue,mkdir as pe,writeFile as D,rm as me}from"node:fs/promises";import he from"node:os";import E from"node:path";var ne=d(async e=>{let a=g(e.packageValue);if(e.grader.runtime!=="sandbox_process")throw new Error("Select the sandbox process runtime for a local file grader.");if(!a.taskset.graders.some(r=>n(r)===n(e.grader)))throw new Error("The private grader differs from its package.");if(!a.taskset.tasks.some(r=>n(r)===n(e.task)))throw new Error("The private case differs from its package.");await e.authorize(),e.signal?.throwIfAborted();let i=await ue(E.join(he.tmpdir(),"profile-private-grade-")),t=V();try{for(let f of a.files.filter(l=>l.asset.visibility!=="policy")){let l=f.asset.path;if(E.isAbsolute(l)||l.split(/[\\/]/).some(u=>!u||u==="."||u===".."))throw new Error("Private grader asset path invalid.");let m=E.join(i,"package",l);await pe(E.dirname(m),{recursive:!0,mode:448}),await D(m,P(f),{mode:256})}await D(E.join(i,"request.json"),JSON.stringify({task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,artifacts:"artifacts"in e.evidence?e.evidence.artifacts:[],expectedOutput:e.task.expectedOutput}),{mode:256});let r=`import fs from 'node:fs/promises'; const module=await import(${JSON.stringify("./package/"+e.grader.verifierRef.path)}); const result=await module[${JSON.stringify(e.grader.exportName??"verify")}](JSON.parse(await fs.readFile('request.json','utf8'))); console.log(JSON.stringify(result));`;await D(E.join(i,"runner.mjs"),r,{mode:256});let s=await ie({authorize:d(async()=>(await e.authorize(),e.signal?.throwIfAborted(),{candidateId:t,candidateRevision:1,ownerId:t,sessionId:t,turnId:t,sourceRoot:i,writablePaths:[]}),"authorize")})({candidateId:t,sessionId:t,turnId:t,expectedRevision:1,command:"/runtime/node --max-old-space-size=512 runner.mjs",timeoutMs:e.grader.timeoutMs,maxOutputBytes:7e4,signal:e.signal});if(await e.authorize(),e.signal?.throwIfAborted(),s.code!==0||s.timedOut||s.stdoutTruncated||s.stderrTruncated)throw new Error("Private grading sandbox did not complete.");return H.parse(JSON.parse(s.stdout))}finally{await me(i,{recursive:!0,force:!0})}},"executeLocalProfileVerifier");function Me(e){return async a=>{if(!a.manifest)throw new Error("The hosted grader requires its admitted manifest.");await a.authorize(),a.signal?.throwIfAborted();let i="artifacts"in a.evidence&&Array.isArray(a.evidence.artifacts)?a.evidence.artifacts.map(({reference:l,title:m,contentType:u})=>({reference:l,title:m,contentType:u})):[],t="caseOwner"in a.evidence?a.evidence.caseOwner:null;if(!t?.sessionId||!t.turnId)throw new Error("The private grader has no sealed case owner.");let r=Buffer.from(JSON.stringify({manifest:a.manifest,packageValue:a.packageValue,graderId:a.grader.id,taskId:a.task.id,evidence:{...a.evidence,artifacts:i}}));if(r.length>2e7)throw new Error("The private grading package exceeds its transport limit.");let o={callId:V(),sessionId:t.sessionId,turnId:t.turnId,sha256:fe("sha256").update(r).digest("hex"),sizeBytes:r.length,timeoutMs:a.grader.timeoutMs},s=d((l,m=12e4)=>e.request(oe.parse({contractVersion:re,requestId:V(),operation:"profile-evaluations/grade",params:l}),m),"request");for(let l=0;l<r.length;l+=98304)await a.authorize(),a.signal?.throwIfAborted(),await s({...o,action:"upload",offset:l,contentsBase64:r.subarray(l,l+98304).toString("base64")});let f=await s({...o,action:"execute"},a.grader.timeoutMs+12e4);return await a.authorize(),a.signal?.throwIfAborted(),H.parse(f)}}d(Me,"createHostedProfileVerifier");async function ge(e,a){let i=g(e);for(let r of i.taskset.graders){if(r.kind==="model_judge"&&(!a||r.calibrationStatus!=="passed"))throw new Error("This Profile requires its current calibrated judge and bounded private provider owner.");let o="verifierRef"in r?r.verifierRef:"rubricRef"in r?r.rubricRef:null;o&&A(i,o)}let t=i.taskset.metrics?.customAggregator;if(t){let r=i.files.find(o=>o.asset.path===t.module);if(!r||r.asset.visibility==="policy"||S(P(r))!==t.contentHash)throw new Error("The authored aggregate metric is not its exact private release.");new TextDecoder("utf8",{fatal:!0}).decode(P(r))}}d(ge,"preflightProfilePrivateGrading");function A(e,a){let i=e.files.find(t=>t.asset.id===a.id);if(!i||i.asset.visibility==="policy"||n(i.asset)!==n(a))throw new Error("The exact private grader asset is unavailable.");return new TextDecoder("utf8",{fatal:!0}).decode(P(i))}d(A,"readPrivate");async function $e(e){let a=g(e.packageValue);if(a.contentHash!==e.manifest.packageHash||a.taskset.contentHash!==e.manifest.tasksetRelease.contentHash)throw new Error("The private grading package changed its frozen Run identity.");await e.authorize(),await ge(a,!!e.executeBudgetedJudge),await e.authorize();let i=d(async({grader:o,task:s,evidence:f})=>{let l=e.signal;await e.authorize(),l?.throwIfAborted();let m={manifest:e.manifest,packageValue:a,grader:o,task:s,evidence:f,signal:l,authorize:e.authorize},u=e.executeSandboxVerifier?await e.executeSandboxVerifier(m):o.runtime==="sandbox_process"?await ne(m):await te({source:A(a,o.verifierRef),exportName:o.exportName,runtime:o.runtime,timeoutMs:o.timeoutMs,signal:l,value:{task:s,attempt:f,input:s.input,output:f.output,artifacts:"artifacts"in f?f.artifacts:[],expectedOutput:s.expectedOutput,infrastructureError:f.infrastructureError??null}});return await e.authorize(),l?.throwIfAborted(),{score:u.score,passed:u.passed,rewardEligible:o.rewardEligible,failureClass:null,feedback:[u.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:u.evidenceRefs}},"customVerifier"),t=a.taskset.metrics?.customAggregator,r=t?new TextDecoder("utf8",{fatal:!0}).decode(P(a.files.find(o=>o.asset.path===t.module))):void 0;return{authorize:e.authorize,customVerifier:i,...e.executeBudgetedJudge?{modelJudge:L({readRubric:d(async o=>{await e.authorize();let s=A(a,o);return await e.authorize(),s},"readRubric"),executeBudgeted:d(async(o,s)=>{await e.authorize(),s?.throwIfAborted();let f=await e.executeBudgetedJudge(o,s);return await e.authorize(),s?.throwIfAborted(),f},"executeBudgeted")})}:{},...r!==void 0?{metricSource:r,metricExecutor:d(async o=>{await e.authorize(),o.signal?.throwIfAborted();let s=await ae(o);return await e.authorize(),o.signal?.throwIfAborted(),s},"metricExecutor")}:{}}}d($e,"createProfilePrivateGradingOwner");import{promises as ve}from"node:fs";import ke from"node:path";async function at(e){let a=e.definition.tasksetRelease,i=await se(e.store,e.harnessRelease);if(!i)throw new Error("Evaluation Profile Harness release is unavailable.");let t=await Re({...e,runtime:i});if(t)return t;let r=await e.store.listTasksetRevisions(e.profileId);for(let o of r){if(a.id!==o.id&&a.id!==`taskset-release-${o.id}-r${o.revision}`)continue;let s=await we(e.storeDir,o);if(s.taskset.id===a.id&&s.taskset.contentHash===a.contentHash)return s}throw new Error(`Released Taskset ${a.id} is unavailable in Profile ${e.profileId}.`)}d(at,"loadLocalProfileEvaluationTaskset");async function we(e,a){let i=a.metadata.importedPackageHash;if(typeof i=="string"){let o=await Z(e,i);if(o.taskset.id!==a.id||o.taskset.revision!==a.revision)throw new Error("Imported evaluation Taskset differs from its local Profile revision.");return o}if(U(a)!==a.contentHash)throw new Error("Authored evaluation Taskset differs from its immutable revision.");let t=Q({taskset:a,adapterId:X(a)}),r=await ee({storeDir:e,taskset:a,content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:t.tasksetRelease,environment:t.environmentRelease,verifierSet:t.verifierSetRelease}});return await Y(e,r),r}d(we,"loadCandidate");async function Re(e){let a=e.definition.tasksetRelease,i=e.runtime;if(i.release.harnessRelease.id!==e.harnessRelease.id||i.release.harnessRelease.contentHash!==e.harnessRelease.contentHash)throw new Error("Evaluation Taskset runtime differs from its requested release.");let t=`evals/tasksets/${a.contentHash}.json`,r=i.release.harnessRelease.files.find(o=>o.path===t);if(r){if(r.visibility!=="verifier")throw new Error("Evaluation Taskset package is not verifier-private.");let o=await ve.readFile(ke.join(i.release.bundlePath,"source",...t.split("/")));if(S(o)!==r.contentHash)throw new Error("Released evaluation Taskset package bytes differ from their source receipt.");let s=g(JSON.parse(o.toString("utf8")));if(s.taskset.id!==a.id||s.taskset.contentHash!==a.contentHash)throw new Error("Released evaluation Taskset package differs from its definition.");return s}return null}d(Re,"loadReleasedProfileEvaluationTaskset");function C(e){return{id:W(e),label:"Selected Dataset evaluation",description:"Explicit sealed external Dataset evaluation recipe.",target:e.target,tasksetRelease:{id:e.dataset.id,contentHash:e.dataset.contentHash},split:e.split,taskIds:[...new Set(e.population.map(a=>a.taskId))],seeds:[...new Set(e.population.map(a=>a.seed))],criterion:e.criterion}}d(C,"externalProfileEvaluationDefinition");async function lt(e){let a=e.manifest.profileEvaluation,i=a?.externalDatasetBinding?y(a.externalDatasetBinding):null;if(!i)return null;if(!e.resolveExternalDataset)throw new Error("This owner has no authorized external Dataset resolver.");let t=await e.resolveExternalDataset(i,e.selected);await t.authorize();let r=g(t.packageValue),o=await e.loadCatalog({ref:e.selected.profileRef,sourceRevision:e.selected.sourceRevision,harnessRelease:e.selected.harnessRelease});if(n(t.binding)!==n(i)||i.profileId!==e.selected.profileRef.profileId||i.declaredProfileCatalogHash!==o.catalogHash||r.contentHash!==i.packageHash||n(r.taskset)!==n(e.taskset)||i.environmentHash!==n(r.taskset.environment)||i.privateDatasetClosureHash!==n(r.files.filter(l=>l.asset.visibility!=="policy").map(l=>l.asset)))throw new Error("External Dataset admission changed its actual Profile, package or private closure.");let s=C(i),f={schemaVersion:"openpond.profileEvaluations.v1",definitions:[s],suites:[]};if(T(e.manifest,e.taskset,f),!a||a.catalogHash!==n(f)||a.definitionHash!==n(s)||n(a.target)!==n(i.target))throw new Error("External Dataset manifest changed its derived definition.");return await t.authorize(),{...t,packageValue:r,definition:s,catalog:f}}d(lt,"admitProfileExternalDataset");var Pe=h.object({id:R,createdAt:J,definitionId:R,externalDatasetBinding:N.optional(),profileRef:K.optional(),profileSource:h.object({sourceRevision:h.string().min(1).max(500),harnessRelease:h.object({id:R,contentHash:v}).strict()}).strict().optional(),modelRef:q,hostModelConfigurationHash:v.optional(),hostArtifactHash:v.optional(),maximumSpendUsd:h.number().positive().max(1e4).optional(),expectedManifestHash:v.optional(),hostExperiment:h.object({definition:h.object({id:R,revision:h.number().int().positive(),contentHash:v}).strict().nullable(),configurationHash:v,project:h.object({id:R,revision:h.number().int().positive(),contentHash:v,targetId:R.nullable()}).strict().optional()}).strict().optional()}).strict();function vt(e){return async(a,i)=>{let t=Pe.parse(a);if(i?.requireExpectedManifestHash&&!t.expectedManifestHash)throw new Error("Profile evaluation execution requires the reviewed manifest hash.");if(t.profileSource&&!t.profileRef)throw new Error("An exact Profile source requires its accepted Profile reference.");let r=i?.selectedWorkflows??await e.selectedWorkflows(t.profileRef,t.profileSource);if(t.profileRef&&n(t.profileRef)!==n(r.profileRef))throw new Error("Profile preparation did not resolve the explicitly selected Profile.");if(t.profileSource&&(r.sourceRevision!==t.profileSource.sourceRevision||n(r.harnessRelease)!==n(t.profileSource.harnessRelease)))throw new Error("Profile preparation changed its explicitly pinned source.");let o=t.externalDatasetBinding?y(t.externalDatasetBinding):null,s=o?await(e.resolveExternalDataset?e.resolveExternalDataset(o,r):Promise.reject(new Error("This host has not qualified external Dataset Profile execution."))):null;if(s&&(await s.authorize(),n(y(s.binding))!==n(o)))throw new Error("External Dataset admission differs from the actual authorized recipe.");let f=await e.loadCatalog({ref:r.profileRef,sourceRevision:r.sourceRevision,harnessRelease:r.harnessRelease});if(o&&o.declaredProfileCatalogHash!==f.catalogHash)throw new Error("The original Profile verifier catalog changed.");let l=o?C(o):null;if(l&&t.definitionId!==l.id)throw new Error("Use the external recipe's own definition identity.");let m={schemaVersion:"openpond.profileEvaluations.v1",definitions:l?[l]:f.definitions,suites:l?[]:f.suites},u=l??f.definitions.find(c=>c.id===t.definitionId);if(!u)throw new Error(`Evaluation ${t.definitionId} is absent from the selected Profile release.`);let k=u.target,z=k.kind==="workflow"?r.workflows.find(c=>c.binding.workflowId===k.workflowId)?.binding:_.parse({schemaVersion:"openpond.profileComponentBinding.v1",profileId:r.profileRef.profileId,sourceRevision:r.sourceRevision,harnessRelease:r.harnessRelease,target:k});if(!z)throw new Error(`Evaluation workflow ${k.kind==="workflow"?k.workflowId:""} is absent from the selected Profile release.`);let w=g(s?.packageValue??await e.loadTasksetPackage(u,r.profileRef.profileId,r.harnessRelease)),p=w.taskset;if(p.id!==u.tasksetRelease.id||p.contentHash!==u.tasksetRelease.contentHash)throw new Error("Evaluation Taskset package differs from its released definition.");let le=p.environment.kind==="text"&&p.tools.length===0,M=p.environment.kind==="work"&&p.environment.entrypoint==="openpond-work-v1"&&p.tools.every(c=>["work_exec","work_save_output"].includes(c.name));if(!le&&!M)throw new Error("Profile evaluation supports text cases or Work cases with the built-in execution and output tools.");if(t.maximumSpendUsd!==void 0&&M&&!e.boundedWorkComputeAvailable)throw new Error("This host has not admitted budgeted Work compute.");if(p.tasks.some(c=>c.artifactRefs.length>b.maxAttachments||c.artifactRefs.some(x=>x.visibility!=="policy"||x.mediaType!=="application/pdf"||x.sizeBytes>b.maxAttachmentBytes)))throw new Error("Profile evaluation accepts only bounded policy-visible PDF task attachments.");if(!o&&!e.privateGradingPreflight&&(p.graders.some(c=>c.kind==="model_judge"||c.kind==="custom_verifier")||p.metrics?.aggregation==="custom"))throw new Error("Profile evaluation cannot run this Taskset's model judge, custom verifier, or custom metric in the current app-server runtime.");await e.privateGradingPreflight?.(w);let B={adapterId:k.kind==="workflow"?"openpond.profile-workflow":"openpond.profile-component",placement:e.placement,runtimeVersion:"app-server-v1",capabilityReceipt:n({adapterId:k.kind==="workflow"?"openpond.profile-workflow":"openpond.profile-component",packageHash:w.contentHash,connectedAppScopes:p.policy.connectedAppScopes})},ce={...$({catalog:m,definitionId:u.id,profileId:r.profileRef.profileId,sourceRevision:r.sourceRevision,harnessRelease:r.harnessRelease,environmentHash:n({packageHash:w.contentHash,runtimeTarget:B})}),...o?{externalDatasetBinding:o}:{},...e.sourceCandidate?{sourceCandidate:e.sourceCandidate}:{}};if(o&&(o.packageHash!==w.contentHash||o.environmentHash!==n(p.environment)||o.privateDatasetClosureHash!==n(w.files.filter(c=>c.asset.visibility!=="policy").map(c=>c.asset))))throw new Error("External Dataset package/environment/private closure changed.");let de=j.parse({provider:t.modelRef.providerId,model:t.modelRef.modelId,revision:null,artifactHash:t.hostArtifactHash??null,tokenizerRevision:null,chatTemplateHash:null}),O=await e.modelConfigurationHash(t.modelRef,t),I=G({schemaVersion:"openpond.tasksetRunManifest.v1",id:t.id,tasksetRelease:u.tasksetRelease,packageHash:w.contentHash,execution:{kind:"harness",harnessRelease:r.harnessRelease},profileEvaluation:ce,policy:{kind:"model",model:de,configurationHash:O},gradingRole:"evaluation",metricPolicy:F(p),population:o?o.population.map(c=>({...c,receiptId:`evaluation-${n([t.id,c.taskId,c.seed,c.fixtureId]).slice(0,32)}`})):u.taskIds.flatMap(c=>u.seeds.map(x=>({receiptId:`evaluation-${n([t.id,c,x]).slice(0,32)}`,taskId:c,seed:x,fixtureId:null}))),runtimeTarget:B,limits:{maxTurns:128,timeoutMs:p.environment.defaultTimeoutMs,maxOutputBytes:25e7,maximumSpendUsd:t.maximumSpendUsd??null},createdAt:t.createdAt,metadata:{sourceTasksetId:p.metadata.sourceTasksetId??p.id,...r.profileRef.source==="openpond_git"?{profileRepositoryId:r.profileRef.repositoryId}:{},...t.hostExperiment?{experimentDefinition:t.hostExperiment.definition,experimentConfigurationHash:t.hostExperiment.configurationHash,...t.hostExperiment.project?{project:t.hostExperiment.project}:{}}:{}}});if(T(I,p,m),s&&await s.authorize(),t.expectedManifestHash&&I.contentHash!==t.expectedManifestHash)throw new Error("Profile evaluation setup changed since preview; prepare the run again.");return{manifest:I,taskset:p,profileRef:r.profileRef,binding:z,modelRef:t.modelRef,modelConfigurationHash:O}}}d(vt,"createProfileEvaluationRunPreparationService");export{ne as a,Me as b,ge as c,$e as d,at as e,Re as f,lt as g,vt as h};

package/dist/chunks/{hosted-experiment-improvement-MGMZWTZY.js → hosted-experiment-improvement-NTYYSPQX.js}

@@ -1,3 +1,3 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as Ie,b as ye,c as Re,d as He,e as Ee,f as Pe,h as B,i as Ae,j as je}from"./chunk-TIUSKRBP.js";import{e as Ce,f as U}from"./chunk-IOQUF34U.js";import{V as be,Va as Se,W as ke,hb as xe,jb as Te,n as we,nb as G,o as ve,ob as De,pb as ze}from"./chunk-NHJ26KXF.js";import"./chunk-EMKZVP7Z.js";import"./chunk-DLZP22WV.js";import{k as ge}from"./chunk-VOIH2BQA.js";import{a as z}from"./chunk-T5LE6R5S.js";import"./chunk-QJRXTIOC.js";import"./chunk-SDO4FC4Q.js";import"./chunk-IR3YJWPW.js";import{h as me,ma as fe,qa as D,ra as he}from"./chunk-7MXZJQS7.js";import"./chunk-7YR6CBNU.js";import"./chunk-2ZJJFBIA.js";import"./chunk-5CMKGD43.js";import{M as W}from"./chunk-HA7EFTPF.js";import"./chunk-L3KS6MOV.js";import"./chunk-7QEE4PBI.js";import{Da as de,Ea as ce,M as C,N as $,a as se,lc as pe}from"./chunk-XDMUD4FW.js";import{aa as le,ga as V,ia as ue}from"./chunk-IKWDL3RB.js";import"./chunk-6MI4KKWF.js";import{Ca as ie,za as ae}from"./chunk-QRWPEW3Z.js";import"./chunk-6FA3GG5B.js";import"./chunk-Z3D4C7BG.js";import{B as re,C as oe,E as ne}from"./chunk-WV2APYWG.js";import{Ma as M}from"./chunk-PJM2G7SG.js";import{L as Y,P as ee,sa as te}from"./chunk-SDEDAHV6.js";import"./chunk-DALJMSJG.js";import"./chunk-TGIB24K4.js";import{l as J,m as R,r as X}from"./chunk-NNF6KCHD.js";import{b as e}from"./chunk-ES2ISFJC.js";import"./chunk-H7IWJN7L.js";import{a as i}from"./chunk-FTQLQJQV.js";import{promises as $e}from"node:fs";import Ue from"node:path";import Le from"node:http";var Q=e.object({socketPath:e.string().min(1).max(1e3),capability:e.string().regex(/^[a-f0-9]{64}$/),jobId:e.string().min(1),pinHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),Ne=e.object({jobId:e.string().min(1),pinHash:e.string().regex(/^[a-f0-9]{64}$/),requestId:e.string().min(1).max(500),model:X,reasoningEffort:e.enum(["off","low","medium","high","xhigh","max","none"]).nullable(),messages:e.array(e.unknown()).min(1).max(1e3),tools:e.array(e.unknown()).max(100),toolChoice:e.unknown(),maxOutputTokens:e.number().int().positive().max(131072),temperature:e.number().finite().optional(),topP:e.number().finite().optional(),seed:e.number().int().optional()}).strict(),Je=e.discriminatedUnion("kind",[e.object({kind:e.literal("chat_completions_reasoning"),reasoningContent:e.string()}).strict(),e.object({kind:e.literal("responses_reasoning_items"),items:e.array(e.record(e.string(),e.unknown()))}).strict()]),Ve=e.object({id:e.string().optional(),type:e.string(),function:e.object({name:e.string().optional(),arguments:e.string().optional()}).passthrough().optional()}).passthrough(),Ge=e.object({text:e.string().optional(),usage:e.unknown().optional(),costUsd:e.number().finite().nonnegative().optional(),continuation:Je.optional(),toolCalls:e.array(Ve).optional()}).strict();function qe(s,r){let t=Q.parse(s);async function n(u,v,I=r){I.throwIfAborted();let x=Buffer.from(JSON.stringify({action:u,jobId:t.jobId,pinHash:t.pinHash,value:v}));if(x.length>8*1024*1024)throw new Error("Hosted advanced request exceeds its private byte limit.");return new Promise((y,h)=>{let l=Le.request({socketPath:t.socketPath,path:"/",method:"POST",headers:{authorization:`Bearer ${t.capability}`,"content-type":"application/json","content-length":x.length},signal:I},H=>{let E=[],P=0;H.on("data",w=>{if(P+=w.length,P>8388608){H.destroy(new Error("Hosted advanced response exceeds its private byte limit."));return}E.push(w)}),H.on("error",h),H.on("end",()=>{try{let w=JSON.parse(Buffer.concat(E).toString("utf8"));if(H.statusCode!==200)throw new Error(typeof w?.error=="string"?w.error:"The actual hosted advanced owner refused this operation.");y(w)}catch(w){h(w)}})});l.on("error",h),l.end(x)})}i(n,"call");let d=i(async function*(u){let v=Ne.parse({jobId:t.jobId,pinHash:t.pinHash,requestId:u.requestId,model:u.model,reasoningEffort:u.reasoningEffort,messages:u.messages,tools:u.tools,toolChoice:u.toolChoice,maxOutputTokens:u.maxOutputTokens??4096,...u.temperature===void 0?{}:{temperature:u.temperature},...u.topP===void 0?{}:{topP:u.topP},...u.seed===void 0?{}:{seed:u.seed}}),I=e.object({deltas:e.array(Ge).max(1e5)}).strict().parse(await n("model",v,AbortSignal.any([r,u.signal])));for(let x of I.deltas)yield x},"stream");return{stream:d,text:i(async u=>{let v="";for await(let I of d({...u,reasoningEffort:u.reasoningEffort??null,tools:[],toolChoice:"none"}))I.text&&(v+=I.text),I.usage!==void 0&&u.onUsage?.(I.usage,I.costUsd);return v},"text"),authorize:i(async()=>{await n("authorize",{})},"authorize"),ceiling:i(async(u,v,I)=>e.object({configurationHash:e.string().regex(/^[a-f0-9]{64}$/),maximumChargeUsd:e.number().finite().positive()}).strict().parse(await n("ceiling",{modelId:u,pricing:v,maximumOutputTokens:I})),"ceiling"),compute:i((u,v=!1)=>n("compute",u,v?AbortSignal.timeout(6e4):r),"compute"),checkpoint:i(u=>n("checkpoint",u),"checkpoint")}}i(qe,"createHostedAdvancedModelBridge");import Qe from"node:path";var Me=e.string().regex(/^[a-f0-9]{64}$/),Ke=e.object({taskId:e.string().min(1),caseHash:Me,session:C,turn:$,events:e.array(se).max(1e4),traceHash:Me,output:e.string().nullable()}).strict(),K=e.object({packageValue:ae,contexts:e.array(Ke).max(1e4)}).strict();async function _e(s){let r=V(s.pin),t=K.parse(s.privateSource),n=ie(t.packageValue),d=M(e.object({manifest:e.unknown(),result:e.unknown()}).parse(s.evidence));async function m(){if(await s.authorize(),r.actorId!==s.actor.actorId||r.teamId!==s.actor.teamId||r.profileRef.source!=="openpond_git"||r.profileRef.repositoryId!==s.source.repositoryId||r.profileRef.profileId!==s.source.profileId||r.sourceRevision!==s.source.sourceRevision)throw new Error("The actual hosted advanced source owner changed.")}if(i(m,"authorize"),await m(),d.manifest.teamId!==r.teamId||d.result.status!=="completed"||d.manifest.id!==r.evidence.id||d.result.contentHash!==r.evidence.contentHash||R(d.manifest.dataset)!==R(r.externalDatasetBinding.dataset))throw new Error("The exact completed hosted advanced evidence changed.");if(n.contentHash!==r.externalDatasetBinding.packageHash||n.contentHash!==r.adaptationDatasetBinding.packageHash||n.environment.contentHash!==r.environmentPolicy.environmentHash||n.environment.contract.networkPolicy!=="none")throw new Error("The hosted advanced environment/package is not the exact reviewed closed-network release.");if(n.taskset.tools.some(l=>l.name==="web_search"||l.name==="web_fetch"))throw new Error("Live external tools require their own qualified frozen-observation and budget owner.");if(n.verifierSet.graders.some(l=>l.kind==="human"||l.kind==="model_judge"&&l.calibrationStatus!=="passed"))throw new Error("Advanced hosted evaluation needs executable, qualified independent graders.");let u=r.adaptationTaskIds.map(l=>n.taskset.tasks.find(H=>H.id===l)),v=r.holdoutTaskIds.map(l=>n.taskset.tasks.find(H=>H.id===l)),I=new Set(u.map(l=>l?.clusterKey));if(u.some(l=>!l||l.split!==r.adaptationSplit)||v.some(l=>!l||l.split!==r.holdoutSplit||I.has(l.clusterKey)))throw new Error("Hosted adaptation/holdout membership or independent families changed.");let x=await s.store.getHarnessReleaseRecord(r.baselineRelease.contentHash);if(!x||x.harnessRelease.id!==r.baselineRelease.id||s.workspace.currentChannel.release?.contentHash!==r.baselineRelease.contentHash)throw new Error("The exact hosted source baseline is unavailable.");let y=await D({workspaceId:typeof x.agentSnapshot.metadata.workspaceId=="string"?x.agentSnapshot.metadata.workspaceId:x.workspaceId,sourceDir:Qe.join(x.bundlePath,"source")});if(z(y).protectedClosureHash!==r.privateClosureHash)throw new Error("The protected hosted Profile closure changed.");let h={packageValue:n,authorize:m};return{read:i(async l=>{if(V(l).contentHash!==r.contentHash)throw new Error("The actual hosted advanced admission changed.");return await m(),h},"read"),evidence:i(async()=>(await m(),d),"evidence"),authorize:i(async()=>m(),"authorize"),reviewContext:i(async(l,H)=>{await m();let E=t.contexts.filter(p=>p.taskId===H),P=d.result.cases.find(p=>p.identity.caseId===H);if(E.length!==1||!P||P.status!=="completed")throw new Error("The actual retained hosted review case is unavailable.");let w=E[0];if(w.caseHash!==R(P)||w.session.cloudTeamId!==r.teamId||w.turn.sessionId!==w.session.id||w.turn.status==="in_progress"||w.events.some(p=>p.sessionId!==w.session.id||p.turnId!==w.turn.id))throw new Error("The actual hosted review session/turn/trace binding changed.");return await m(),{...w,case:{output:w.output}}},"reviewContext")}}i(_e,"createHostedAdvancedRefinerSource");import{randomUUID as Ze}from"node:crypto";function Oe(s){async function r(n){let d=await s.store.getSession(n);if(!d||d.cloudTeamId!==s.teamId||d.metadata?.hostedAdvancedTurnId!==s.hostedTurnId||d.metadata?.parentModelRunId!==s.modelRunId)throw new Error("The actual hosted advanced case owner changed.");return d}i(r,"read");async function t(n,d,m){let u=await r(n),v=e.object({action:e.string()}).passthrough().parse(d).action;return ce.parse(await s.bridge.compute({session:u,payload:d,turnId:m??null},v==="sandbox_stop"||v==="sandbox_receipts"))}return i(t,"tool"),{createSession:i(async n=>{await s.bridge.authorize();let d=C.partial().parse(n),m=new Date().toISOString();if(d.metadata?.parentModelRunId!==s.modelRunId)throw new Error("The foreground case belongs to another actual advanced Run.");let u=C.parse({...d,id:`hosted-advanced-case-${Ze()}`,provider:"openpond",currentProfile:s.profileRef,cloudTeamId:s.teamId,cwd:null,codexThreadId:null,appId:null,appName:null,createdAt:m,updatedAt:m,status:"active",pinned:!1,archived:!1,order:0,hiddenFromDefaultSidebar:!0,title:d.title??"Advanced hosted evaluation case",metadata:{...d.metadata,source:"hosted-advanced-evaluation",hostedAdvancedTurnId:s.hostedTurnId,actorId:s.actorId,workspaceTarget:"hosted",benchmarkRuntime:"hosted_managed_work"}});return await s.store.insertSessionAtFront(u),await s.bridge.authorize(),u},"createSession"),getSession:r,runtimeEventsForSession:i(n=>s.store.runtimeEventsForSession(n),"runtimeEventsForSession"),executeWorkspaceTool:i((n,d,m)=>t(n,d,m?.turnId),"executeWorkspaceTool"),settleCostEvidence:i((n,d)=>t(n,{action:"sandbox_receipts",args:{}},d?.turnId),"settleCostEvidence")}}i(Oe,"createHostedAdvancedWorkRuntime");var Xe=e.string().regex(/^[a-f0-9]{64}$/),Be=e.object({version:e.string().min(1),source:e.string().min(1),effectiveAt:e.string().min(1),inputUsdPerMillionTokens:e.number().finite().nonnegative(),cachedInputUsdPerMillionTokens:e.number().finite().nonnegative(),outputUsdPerMillionTokens:e.number().finite().nonnegative()}).strict(),Z=e.object({operation:e.enum(["prepare_run","start","read","resume","cancel"]),request:ue,configuration:e.object({id:e.string().min(1),profileId:e.string().min(1),etag:Xe}).strict(),refinerRelease:Y,privateSource:K,upstream:e.object({modelId:e.string().min(1),revision:e.string().min(1),pricing:Be}).strict(),judgeModels:e.array(e.object({modelId:e.string().min(1),revision:e.string().min(1),pricing:Be}).strict()).max(50),bridge:Q.optional(),hostedTurnId:e.string().min(1).optional()}).strict();async function Fe(s){let r=Z.parse(s.request),t=r.request.pin;if(r.request.model.providerId!=="openpond"||r.request.model.modelId!==r.upstream.modelId||r.configuration.id!==r.request.modelId||r.configuration.profileId!==t.profileRef.profileId||r.request.profileId!==t.profileRef.profileId||r.refinerRelease.id!==t.refinerRelease.id||r.refinerRelease.contentHash!==t.refinerRelease.contentHash)throw new Error("The actual hosted advanced configuration/model/Refiner admission changed.");let n=`advanced-refiner-${R([t.actorId,t.teamId,t.operationId,t.contentHash,r.request.model,[r.request.seed],1]).slice(0,40)}`;if((r.operation==="start"||r.operation==="resume")&&(!r.bridge||!r.hostedTurnId))throw new Error("The actual hosted advanced execution grant is unavailable.");let m=r.bridge?qe(r.bridge,s.signal):null;if(r.bridge&&(r.bridge.jobId!==n||r.bridge.pinHash!==t.contentHash))throw new Error("The hosted model bridge belongs to another actual Run.");let u=i(async()=>{s.signal.throwIfAborted(),await s.authorize(),m&&await m.authorize()},"authorize"),v=await _e({...s,pin:t,privateSource:r.privateSource,authorize:u}),I=we(ve(s.storeDir));await I.persistRelease(r.refinerRelease),await I.writeSource(ee(r.refinerRelease.profile)),I.binding()?.release.contentHash!==t.refinerRelease.contentHash&&await I.transition(r.refinerRelease,{operation:"activate",bindingChanged:!0,actor:s.actor.actorId,reason:"Pin the actual hosted evaluation fork to its admitted immutable Refiner release.",authoringSkillHash:null});let y=ke({storeDir:s.storeDir,authorize:i(async()=>u(),"authorize"),packageForPin:i(async()=>r.privateSource.packageValue,"packageForPin"),prepareModel:i(async(c,f,k)=>{if(!m)throw new Error("The actual hosted model ceiling owner is unavailable.");return m.ceiling(c,k,f)},"prepareModel")}),h=i(async function*(c){let f=be(),k=[r.upstream,...r.judgeModels].find(q=>q.modelId===c.model.modelId);if(!m||!f||!k)throw new Error("The actual hosted model and durable shared budget are unavailable.");let{signal:T,...j}=c,O=c.maxOutputTokens??4096,L=await f.call(c.requestId,j,k.pricing,O,async()=>{let q=[],F=null;for await(let N of m.stream({...c,signal:AbortSignal.any([T,f.signal])}))q.push(N),N.costUsd!==void 0&&(F=(F??0)+N.costUsd);return{value:q,costUsd:F}});for(let q of L)yield q},"stream"),l=i(async c=>{let f="";for await(let k of h({...c,reasoningEffort:c.reasoningEffort??null,tools:[],toolChoice:"none"}))k.text&&(f+=k.text),k.usage!==void 0&&c.onUsage?.(k.usage,k.costUsd);return f},"text"),H=Ie({store:s.store,storeDir:s.storeDir}),E=await H.projectSelected({profileId:t.profileRef.profileId,package:r.privateSource.packageValue,adaptationSplit:t.adaptationSplit,holdoutSplit:t.holdoutSplit}),P=i(async c=>{let f=C.partial().parse(c),k=new Date().toISOString(),T=C.parse({...f,id:`hosted-review-${R([n,f.metadata]).slice(0,40)}`,cloudTeamId:t.teamId,currentProfile:t.profileRef,appId:null,appName:null,cwd:null,codexThreadId:null,title:f.title??"Hosted Refiner review",provider:"openpond",createdAt:k,updatedAt:k,status:"active",pinned:!1,archived:!1,order:0,metadata:{...f.metadata,hostedAdvancedTurnId:r.hostedTurnId??null,actorId:t.actorId,source:"hosted-advanced-evaluation"}});return await s.store.insertSessionAtFront(T),T},"createSession"),w=m&&r.hostedTurnId?Oe({store:s.store,bridge:m,teamId:t.teamId,actorId:t.actorId,hostedTurnId:r.hostedTurnId,modelRunId:n,profileRef:t.profileRef}):void 0,p=Se({store:s.store,storeDir:s.storeDir,modelJudge:ye({store:s.store,modelText:l}),modelText:l,modelStream:h,workRuntime:w,resolveTasksetRelease:i(c=>H.releaseForTaskset(c),"resolveTasksetRelease")}),a={store:s.store,storeDir:s.storeDir,evaluation:p,benchmarkTasksets:H,loadProfileState:i(async()=>s.profile,"loadProfileState"),resolveEvaluationProject:i(async(c,f)=>{if(await u(),c!==r.configuration.id||f!==r.configuration.profileId)throw new Error("The actual hosted evaluation configuration changed.");return{id:c,profileId:f}},"resolveEvaluationProject"),resolveUpstreamModel:i(async()=>({providerId:"openpond",modelId:r.upstream.modelId,revision:r.upstream.revision,pricing:r.upstream.pricing}),"resolveUpstreamModel"),advancedBoundary:y,refinerStream:i(async function*(c){if(!c.requestId)throw new Error("A hosted Refiner call needs its actual retained ID.");yield*h({model:c.model,reasoningEffort:"low",messages:c.messages,tools:[],toolChoice:"none",requestId:c.requestId,maxOutputTokens:4096,signal:c.signal,hostedTokenPricing:c.pricing})},"refinerStream")},o=Re({...a,reviewQuality:He({source:v,boundary:y,benchmark:a,createSession:P})}),g={advancedEvaluation:t,selectedTaskset:{id:E.id,revision:E.revision,contentHash:E.contentHash},modelId:r.request.modelId,profileId:r.request.profileId,model:r.request.model,reasoningEffort:r.request.reasoningEffort,seeds:[r.request.seed],repetitions:1,maximumSpendUsd:t.maximumCostUsd},b,S=i(()=>{o.cancel(n)},"stop");s.signal.addEventListener("abort",S,{once:!0});try{if(r.operation==="prepare_run")return o.prepare(g);let c=await s.store.getModelRun(n);if(!c||c.evaluation?.benchmarkId!=="harness-refiner"||c.evaluation.advancedEvaluation?.contentHash!==t.contentHash)throw new Error("The original hosted advanced Run is unavailable.");if(r.operation==="read")return c;if(r.operation==="cancel")return o.cancel(n);if(m){let T=!1;b=setInterval(()=>{T||(T=!0,s.store.getModelRun(n).then(j=>j?m.checkpoint(j):void 0).catch(()=>S()).finally(()=>{T=!1}))},2e3)}r.operation==="start"?await o.start(g):await o.resume(n);let f=await o.wait(n);if(!f)throw new Error("The actual hosted advanced engine did not retain its Run.");let k=pe.parse(f);return m&&await m.checkpoint(k),k}finally{b&&clearInterval(b),s.signal.removeEventListener("abort",S),await o.close(),await y.close()}}i(Fe,"runHostedAdvancedRefinerOwner");import{promises as _}from"node:fs";import A from"node:path";var Ye=e.object({actor:e.object({actorId:e.string().min(1),teamId:e.string().min(1)}).strict(),storeDir:e.string().min(1),source:e.object({repoPath:e.string().min(1),repositoryId:e.string().min(1),profileId:e.string().min(1),sourceRevision:e.string().regex(/^[a-f0-9]{40,64}$/),filesHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),evidence:e.object({manifest:e.unknown(),result:e.unknown()}).strict(),operation:e.enum(["advanced","describe","command","queue","cancelQueued","bind","finish","beginTesting","cancelTesting","failTesting","comparison","beginAdoption","export","source","tool","toolCatalog"]),advanced:Z.optional(),command:le.optional(),candidateId:e.string().optional(),revision:e.number().int().positive().optional(),actualSession:C.optional(),actualTurn:$.optional(),recipe:e.unknown().optional(),baseline:e.object({manifest:e.unknown(),result:e.unknown()}).strict().optional(),candidate:e.object({manifest:e.unknown(),result:e.unknown()}).strict().optional(),humanProjection:oe.optional(),humanSelections:re.optional(),adoptionReadback:e.unknown().optional(),failure:e.object({message:e.string().max(2e3),cancelled:e.boolean()}).strict().optional(),tool:e.object({name:e.string(),args:e.record(e.string(),e.unknown()),callId:e.string()}).strict().optional(),agentRuntime:e.object({source:e.string(),cliRelativePath:e.string()}).strict().optional(),queued:e.object({sessionId:e.string(),turnId:e.string(),createdAt:e.string().datetime()}).strict().optional()}).strict();async function We(s,r=new AbortController().signal){r.throwIfAborted();let t=Ye.parse(s),{actor:n,source:d}=t;if(!A.isAbsolute(t.storeDir)||!A.isAbsolute(d.repoPath))throw new Error("Hosted candidate owner requires contained absolute private paths.");let m=M(t.evidence),u=[t.baseline,t.candidate].filter(l=>l!==void 0).map(l=>M(l));if(m.manifest.teamId!==n.teamId)throw new Error("Hosted candidate evidence belongs to another team.");if(await B(d.repoPath)!==d.filesHash)throw new Error("The admitted hosted Profile source bytes changed.");let v=`hosted-improve-source-${R([n,d.repositoryId,d.profileId,d.sourceRevision]).slice(0,40)}`,I=A.join(t.storeDir,"library","profile-experiment-origins",v);if(await _.mkdir(A.dirname(I),{recursive:!0,mode:448}),await _.stat(I).catch(()=>null)||await _.cp(d.repoPath,I,{recursive:!0,errorOnExist:!0,force:!1}),await B(I)!==d.filesHash)throw new Error("The private accepted Profile cache changed.");let x=await W(I,d.profileId);if(x.error||!x.sourcePath)throw new Error("The accepted hosted Profile does not compile.");let y=new xe(t.storeDir),h;try{let l=await y.getHarnessWorkspace(v);l||(l=(await fe({store:y,storeDir:t.storeDir,id:v,ownerId:n.actorId,name:d.profileId,profile:x,sourceRevision:d.sourceRevision,repositoryId:d.repositoryId,selectionEligible:!1,originMetadata:{profileExperimentOrigin:{schemaVersion:"openpond.profileExperimentOrigin.v1",...n,sourceRevision:d.sourceRevision,sourcePath:I,sourceFilesHash:d.filesHash},hostedImprovementOwner:!0}})).workspace);let H={source:"openpond_git",repositoryId:d.repositoryId,profileId:d.profileId},E=i(async()=>{if(r.throwIfAborted(),await B(I)!==d.filesHash||await B(d.repoPath)!==d.filesHash)throw new Error("Accepted source changed during private candidate I/O.")},"authorizer");if(t.operation==="advanced"){if(!t.advanced)throw new Error("Missing actual hosted advanced owner command.");return Fe({store:y,storeDir:t.storeDir,actor:n,profile:x,workspace:l,source:d,evidence:m,request:t.advanced,authorize:E,signal:r})}let P=i(()=>{if(!t.adoptionReadback)throw new Error("Actual hosted Git owner has not returned its sealed adoption receipt.");return e.object({receipt:e.object({id:e.string(),contentHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),activeRelease:e.object({id:e.string(),contentHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),testedCandidateRelease:e.object({id:e.string(),contentHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),executableSourceHash:e.string().regex(/^[a-f0-9]{64}$/),protectedClosureHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict().parse(t.adoptionReadback)},"readback");h=Ee({store:y,storeDir:t.storeDir,projectHumanComparison:i(async()=>{if(await E(),!t.humanProjection)throw new Error("The current private Human comparison owner is unavailable.");return ne(t.humanProjection)},"projectHumanComparison"),authorizeOwner:i(async(a,o,g,b,S)=>{if(R(a)!==R(n)||R(g)!==R(H)||o.id!==v)throw new Error("Hosted candidate source owner changed.");if(S.requireCurrentBase&&(o.currentChannel.release?.contentHash!==l.currentChannel.release?.contentHash||d.sourceRevision!==o.metadata.profileExperimentOrigin.sourceRevision))throw new Error("Hosted candidate accepted revision changed.");await E()},"authorizeOwner"),loadEvidence:i(async(a,o)=>{if(R(a)!==R(n))throw new Error("Hosted candidate evidence actor changed.");let g=[m,...u].find(b=>b.manifest.id===o.id&&b.result.contentHash===o.contentHash&&b.manifest.teamId===n.teamId);if(!g)throw new Error("Hosted candidate exact retained evidence changed.");return await E(),g},"loadEvidence"),freezeProfile:i(async(a,o,g)=>{let b=await y.getHarnessReleaseRecord(o.baseRelease.contentHash);if(!b)throw new Error("Hosted source baseline is missing.");let S=await D({workspaceId:typeof b.agentSnapshot.metadata.workspaceId=="string"?b.agentSnapshot.metadata.workspaceId:b.workspaceId,sourceDir:A.join(b.bundlePath,"source")}),c=await D({workspaceId:typeof g.release.agentSnapshot.metadata.workspaceId=="string"?g.release.agentSnapshot.metadata.workspaceId:g.release.workspaceId,sourceDir:A.join(g.release.bundlePath,"source")});return Ae({store:y,storeDir:t.storeDir,state:o},async f=>{await Ce({...f,state:o,base:S,candidate:c});let k=await W(f.root,o.profileRef.profileId),T=await U({storeDir:t.storeDir,state:o,profile:k,sourceRevision:g.profileSourceRevision,base:S}),j=z(T);if(j.protectedClosureHash!==z(S).protectedClosureHash)throw new Error("Hosted source candidate changed protected evaluation closure.");let O=A.join(t.storeDir,"library","profile-candidate-archives",o.id,j.contentHash);await _.mkdir(A.dirname(O),{recursive:!0,mode:448}),await _.stat(O).catch(()=>null)||await _.cp(f.root,O,{recursive:!0,errorOnExist:!0,force:!1});let L=await W(O,o.profileRef.profileId);if((await U({storeDir:t.storeDir,state:o,profile:L,sourceRevision:g.profileSourceRevision,base:S})).harnessRelease.contentHash!==T.harnessRelease.contentHash)throw new Error("Hosted frozen Profile archive readback differs.");let F=await he({storeDir:t.storeDir,workspaceId:g.release.workspaceId,compiled:T,createdAt:g.release.createdAt});return await E(),{...g,release:F,executableIdentity:j}})},"freezeProfile"),adopt:i(async()=>P(),"adopt"),readAdoption:i(async()=>t.adoptionReadback?P():null,"readAdoption"),rollback:i(async()=>P(),"rollback")});let w=Pe({store:y,improvements:h,actorId:i(async()=>n.actorId,"actorId"),teamId:i(async()=>n.teamId,"teamId"),loadProfile:i(async a=>{if(R(a)!==R(H))throw new Error("Candidate requested another source owner.");return x},"loadProfile"),agentRuntime:t.agentRuntime,operationSignal:r,loadAgentRuntime:De({storeDir:t.storeDir,loadArchive:ze({storeDir:t.storeDir})})});if(t.operation==="describe"){await E();let a=l.currentChannel.release,o=a?await y.getHarnessReleaseRecord(a.contentHash):null;if(!a||!o||o.workspaceId!==l.id||o.harnessRelease.id!==a.id||o.harnessRelease.contentHash!==a.contentHash||l.ownerScope.kind!=="personal"||l.ownerScope.id!==n.actorId)throw new Error("Hosted source release is unavailable to its admitted owner.");let g=await D({workspaceId:typeof o.agentSnapshot.metadata.workspaceId=="string"?o.agentSnapshot.metadata.workspaceId:v,sourceDir:A.join(o.bundlePath,"source")});if(g.harnessRelease.contentHash!==o.harnessRelease.contentHash)throw new Error("Hosted source compiler readback changed.");let b=await ge({runtime:{release:o},ref:H,sourceRevision:d.sourceRevision,harnessRelease:a}),S=i(c=>{let f=g.sourceFiles.find(k=>k.path===c)?.bytes;return f?JSON.parse(Buffer.from(f).toString("utf8")):null},"readJson");return await E(),{workspaceId:v,ownerRevision:l.revision,profileRef:H,baseRelease:l.currentChannel.release,profileSourceRevision:d.sourceRevision,manifest:g.manifest,executable:z(g),catalogHash:b.catalogHash,workflows:S("workflows/catalog.json"),actions:S("workflows/actions.json")}}if(t.operation==="command"){let a=t.command;if(!a)throw new Error("Missing canonical Improve command.");if(a.operation==="start")return h.start(n,a.request);if(a.operation==="list")return h.list(n,a.limit,a.cursor);if(!("id"in a))throw new Error("Discover exact hosted source choices through the authenticated host.");let o=await h.read(n,a.id);if(a.operation==="read")return o;if(a.operation==="instructions"){let g=await h.partition(n,o.id),b=await me(g.sourceRoot,o.component.path),S=await _.readFile(b);if(S.length>25e4)throw new Error("Instruction editor size limit exceeded.");let c=S.toString("utf8");if(o.component.kind==="workflow"){let f=JSON.parse(c).workflows?.find(k=>k.id===o.component.workflowId);if(f?.invocation?.kind!=="instructions")throw new Error("Use Work to author this executable Workflow.");c=f.invocation.instructions}return{candidateId:o.id,candidateRevision:o.revision,path:o.component.path,text:c,textHash:J(c),contentHash:J(S)}}if(a.operation==="humanComparisonOptions")throw new Error("Read the actual paired sources through their current hosted Experiment owner.");if(!("revision"in a)||o.revision!==a.revision)throw new Error("Candidate revision changed.");if(a.operation==="edit")return h.edit(n,o.id,o.revision,a.expectedFileHash,a.text);if(a.operation==="mode")return h.setMode(n,o.id,o.revision,a.mode);if(a.operation==="freeze")return h.freeze(n,o.id,o.revision);if(a.operation==="useCandidate")return h.useCandidate(n,o.id,o.revision);if(a.operation==="rollback")return h.rollback(n,o.id,o.revision);if(a.operation==="discard")return h.discard(n,o.id,o.revision);throw new Error("Hosted Work/test lifecycle requires the actual host turn/run admission.")}if(!t.candidateId)throw new Error("Missing actual hosted candidate ID.");let p=await h.read(n,t.candidateId);if(t.operation==="beginAdoption")return h.beginAdoption(n,p.id,t.revision);if(t.operation==="export")return je(y,p,"hosted_execution");if(t.operation==="source"){if(!p.frozen||p.status!=="frozen"&&p.status!=="testing"&&p.status!=="ready_for_review"&&p.status!=="adopting")throw new Error("Only the exact frozen source can enter hosted preparation.");let a=await y.getHarnessReleaseRecord(p.frozen.release.contentHash);if(!a||a.harnessRelease.id!==p.frozen.release.id)throw new Error("Frozen hosted source is unavailable.");let o=await D({workspaceId:typeof a.agentSnapshot.metadata.workspaceId=="string"?a.agentSnapshot.metadata.workspaceId:a.workspaceId,sourceDir:A.join(a.bundlePath,"source")}),g=z(o);if(o.harnessRelease.contentHash!==p.frozen.release.contentHash||g.contentHash!==p.frozen.executableSourceHash||g.protectedClosureHash!==p.frozen.protectedClosureHash)throw new Error("Frozen hosted executable changed.");let b=A.join(t.storeDir,"library","profile-candidate-archives",p.id,g.contentHash),S=await W(b,p.profileRef.profileId);if(S.error||!S.sourcePath)throw new Error("Frozen actual Profile archive is unavailable.");let c=await y.getHarnessReleaseRecord(p.baseRelease.contentHash);if(!c)throw new Error("Frozen baseline is unavailable.");let f=await D({workspaceId:typeof c.agentSnapshot.metadata.workspaceId=="string"?c.agentSnapshot.metadata.workspaceId:c.workspaceId,sourceDir:A.join(c.bundlePath,"source")}),k=await U({storeDir:t.storeDir,state:p,profile:S,sourceRevision:p.frozen.profileSourceRevision,base:f});if(k.harnessRelease.contentHash!==p.frozen.release.contentHash||z(k).contentHash!==g.contentHash)throw new Error("Actual frozen Profile archive differs from its compiled release.");return{state:p,profileSource:{repoPath:b,profileId:p.profileRef.profileId,sourceRevision:p.frozen.profileSourceRevision,repositoryId:p.profileRef.repositoryId,filesHash:await B(b)},sourcePackage:te({agentSnapshot:o.agentSnapshot,harnessRelease:o.harnessRelease,files:new Map(o.sourceFiles.map(T=>[T.path,T.bytes]))})}}if(t.operation==="comparison"){if(!t.baseline||!t.candidate)throw new Error("Missing actual hosted paired evidence.");return h.completeComparison(n,p.id,t.revision,M(t.baseline),M(t.candidate),t.humanSelections)}if(t.operation==="beginTesting")return h.beginTesting(n,p.id,t.revision,t.recipe);if(t.operation==="failTesting"){if(!t.failure)throw new Error("Missing actual hosted test failure.");return h.failTesting(n,p.id,t.revision,t.failure.message,t.failure.cancelled)}if(t.operation==="cancelTesting")return h.cancelTesting(n,p.id,t.revision);if(t.operation==="queue"){if(!t.queued)throw new Error("Missing actual hosted queue admission.");return h.queueAuthoringTurn(n,p.id,t.revision,t.queued.sessionId,t.queued.turnId,t.queued.createdAt)}if(t.operation==="cancelQueued")return h.cancelQueuedAuthoring(n,p.id,t.revision);if(t.operation==="bind"||t.operation==="finish"||t.operation==="tool"||t.operation==="toolCatalog"){if(!t.actualSession||!t.actualTurn||t.actualTurn.sessionId!==t.actualSession.id||t.actualSession.cloudTeamId!==n.teamId||R(t.actualSession.currentProfile)!==R(H))throw new Error("Actual hosted candidate Work binding changed.");let a=t.actualSession,o=t.actualTurn,g=await y.getSession(a.id);if(!g)await y.insertSessionAtFront(a);else if(R({...g,updatedAt:a.updatedAt,status:a.status})!==R(a))throw new Error("Retained hosted candidate session changed.");let b=await y.getTurn(o.id);if(!b)await y.insertTurn(o);else{if(b.sessionId!==o.sessionId||R(b.metadata.refinementCandidate)!==R(o.metadata.refinementCandidate)||b.prompt!==o.prompt)throw new Error("Actual hosted candidate turn identity changed.");await y.updateTurn(o.id,()=>o)}if(t.operation==="bind")return h.bindAuthoringTurn(n,p.id,t.revision,a.id,o.id);if(t.operation==="finish")return h.finishAuthoring(n,p.id,t.revision);if(t.operation==="toolCatalog")return await w.candidateAuthoringForTurn(a.id,o.id),[...G({loadProfileState:i(async()=>x,"loadProfileState"),resolveCandidateProfile:w.resolveCandidateProfile,executeCandidateAgentCommand:w.executeCandidateAgentCommand}),...Te(async()=>{throw new Error("Hosted file tools require a bound private call.")}),{name:"exec_command",description:"Run a bounded command confined to the selected public candidate component.",parameters:{type:"object",additionalProperties:!1,properties:{command:{type:"string",minLength:1,maxLength:2e4},cwd:{type:"string"}},required:["command"]}},{name:"view_image",description:"Inspect an image inside the candidate public source tree.",parameters:{type:"object",additionalProperties:!1,properties:{path:{type:"string"}},required:["path"]}}].map(({name:c,description:f,parameters:k})=>({name:c,description:f,parameters:k}));if(!t.tool)throw new Error("Missing candidate tool call.");await w.authorizeCandidateTool({sessionId:a.id,turnId:o.id,name:t.tool.name});let S={session:a,turnId:o.id,args:t.tool.args,callId:t.tool.callId,signal:r,workspaceDiffBaseline:null,turnPermissions:{approvalPolicy:"never",sandbox:"workspace-write",codexPermissionMode:"default"},provider:a.provider,model:o.modelRef?.modelId??a.modelRef?.modelId??"",mentionedApps:[],userPrompt:o.prompt,turnMetadata:o.metadata};if(t.tool.name==="get_profile"||t.tool.name.startsWith("agent_")||t.tool.name==="ask_user"){let c=G({loadProfileState:i(async()=>x,"loadProfileState"),resolveCandidateProfile:w.resolveCandidateProfile,executeCandidateAgentCommand:w.executeCandidateAgentCommand}).find(f=>f.name===t.tool.name);if(!c)throw new Error("Candidate authoring tool is unavailable.");return c.execute(S)}if(t.tool.name==="view_image")return w.executeCandidateImage(S);if(t.tool.name==="exec_command")return w.executeCandidateCommand(S,{session:a,turnId:o.id,command:e.string().min(1).max(2e4).parse(t.tool.args.command),cwd:typeof t.tool.args.cwd=="string"?t.tool.args.cwd:null,signal:S.signal,source:"model_tool"});if(["list_files","read_files","search_files","workspace_status","write_file","write_files","edit_file","delete_file"].includes(t.tool.name))return w.executeCandidateWorkspaceTool({session:a,turnId:o.id,request:de.parse({action:t.tool.name,args:t.tool.args,source:"chat_action"})});throw new Error("This hosted tool has no confined candidate transport.")}throw new Error("Unknown private hosted Improve operation.")}finally{await h?.close(),await y.close()}}i(We,"runHostedExperimentImprovementOwner");async function wr(s){if(s.length!==1||s[0].length>16e3||!/^[A-Za-z0-9_-]+$/.test(s[0]))throw new Error("Hosted Improve IPC argument is invalid.");let r=e.object({directory:e.string(),file:e.string().regex(/^[A-Za-z0-9_-]+\.json$/)}).strict().parse(JSON.parse(Buffer.from(s[0],"base64url").toString("utf8")));if(!Ue.isAbsolute(r.directory))throw new Error("Hosted Improve IPC requires an absolute private directory.");let t=Ue.join(r.directory,r.file),n=await $e.lstat(t);if(!n.isFile()||n.isSymbolicLink()||n.size>32*1024*1024)throw new Error("Hosted Improve private request exceeds its byte/type limit.");let d=new AbortController,m=i(()=>d.abort(new Error("The actual hosted owner stopped this private operation.")),"stop");process.once("SIGTERM",m),process.once("SIGINT",m);try{let u=await We(JSON.parse(await $e.readFile(t,"utf8")),d.signal),v=Buffer.from(JSON.stringify(u));if(v.length>32*1024*1024)throw new Error("Hosted Improve private response exceeds its byte limit.");process.stdout.write(`${v.toString("utf8")}
+import{a as Ie,b as ye,c as Re,d as He,e as Ee,f as Pe,h as B,i as Ae,j as je}from"./chunk-UBFRWC6I.js";import{e as Ce,f as U}from"./chunk-IOQUF34U.js";import{V as be,Va as Se,W as ke,hb as xe,jb as Te,n as we,nb as G,o as ve,ob as De,pb as ze}from"./chunk-TBNBCOW7.js";import"./chunk-EMKZVP7Z.js";import"./chunk-DLZP22WV.js";import{k as ge}from"./chunk-OE2VF6KU.js";import{a as z}from"./chunk-T5LE6R5S.js";import"./chunk-QJRXTIOC.js";import"./chunk-SDO4FC4Q.js";import"./chunk-IR3YJWPW.js";import{h as me,ma as fe,qa as D,ra as he}from"./chunk-7MXZJQS7.js";import"./chunk-7YR6CBNU.js";import"./chunk-2ZJJFBIA.js";import"./chunk-5CMKGD43.js";import{M as W}from"./chunk-HA7EFTPF.js";import"./chunk-L3KS6MOV.js";import"./chunk-7QEE4PBI.js";import{Da as de,Ea as ce,M as C,N as $,a as se,lc as pe}from"./chunk-XDMUD4FW.js";import{aa as le,ga as V,ia as ue}from"./chunk-IKWDL3RB.js";import"./chunk-6MI4KKWF.js";import{Ca as ie,za as ae}from"./chunk-QRWPEW3Z.js";import"./chunk-6FA3GG5B.js";import"./chunk-Z3D4C7BG.js";import{B as re,C as oe,E as ne}from"./chunk-WV2APYWG.js";import{Ma as M}from"./chunk-PJM2G7SG.js";import{L as Y,P as ee,sa as te}from"./chunk-SDEDAHV6.js";import"./chunk-DALJMSJG.js";import"./chunk-TGIB24K4.js";import{l as J,m as R,r as X}from"./chunk-NNF6KCHD.js";import{b as e}from"./chunk-ES2ISFJC.js";import"./chunk-H7IWJN7L.js";import{a as i}from"./chunk-FTQLQJQV.js";import{promises as $e}from"node:fs";import Ue from"node:path";import Le from"node:http";var Q=e.object({socketPath:e.string().min(1).max(1e3),capability:e.string().regex(/^[a-f0-9]{64}$/),jobId:e.string().min(1),pinHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),Ne=e.object({jobId:e.string().min(1),pinHash:e.string().regex(/^[a-f0-9]{64}$/),requestId:e.string().min(1).max(500),model:X,reasoningEffort:e.enum(["off","low","medium","high","xhigh","max","none"]).nullable(),messages:e.array(e.unknown()).min(1).max(1e3),tools:e.array(e.unknown()).max(100),toolChoice:e.unknown(),maxOutputTokens:e.number().int().positive().max(131072),temperature:e.number().finite().optional(),topP:e.number().finite().optional(),seed:e.number().int().optional()}).strict(),Je=e.discriminatedUnion("kind",[e.object({kind:e.literal("chat_completions_reasoning"),reasoningContent:e.string()}).strict(),e.object({kind:e.literal("responses_reasoning_items"),items:e.array(e.record(e.string(),e.unknown()))}).strict()]),Ve=e.object({id:e.string().optional(),type:e.string(),function:e.object({name:e.string().optional(),arguments:e.string().optional()}).passthrough().optional()}).passthrough(),Ge=e.object({text:e.string().optional(),usage:e.unknown().optional(),costUsd:e.number().finite().nonnegative().optional(),continuation:Je.optional(),toolCalls:e.array(Ve).optional()}).strict();function qe(s,r){let t=Q.parse(s);async function n(u,v,I=r){I.throwIfAborted();let x=Buffer.from(JSON.stringify({action:u,jobId:t.jobId,pinHash:t.pinHash,value:v}));if(x.length>8*1024*1024)throw new Error("Hosted advanced request exceeds its private byte limit.");return new Promise((y,h)=>{let l=Le.request({socketPath:t.socketPath,path:"/",method:"POST",headers:{authorization:`Bearer ${t.capability}`,"content-type":"application/json","content-length":x.length},signal:I},H=>{let E=[],P=0;H.on("data",w=>{if(P+=w.length,P>8388608){H.destroy(new Error("Hosted advanced response exceeds its private byte limit."));return}E.push(w)}),H.on("error",h),H.on("end",()=>{try{let w=JSON.parse(Buffer.concat(E).toString("utf8"));if(H.statusCode!==200)throw new Error(typeof w?.error=="string"?w.error:"The actual hosted advanced owner refused this operation.");y(w)}catch(w){h(w)}})});l.on("error",h),l.end(x)})}i(n,"call");let d=i(async function*(u){let v=Ne.parse({jobId:t.jobId,pinHash:t.pinHash,requestId:u.requestId,model:u.model,reasoningEffort:u.reasoningEffort,messages:u.messages,tools:u.tools,toolChoice:u.toolChoice,maxOutputTokens:u.maxOutputTokens??4096,...u.temperature===void 0?{}:{temperature:u.temperature},...u.topP===void 0?{}:{topP:u.topP},...u.seed===void 0?{}:{seed:u.seed}}),I=e.object({deltas:e.array(Ge).max(1e5)}).strict().parse(await n("model",v,AbortSignal.any([r,u.signal])));for(let x of I.deltas)yield x},"stream");return{stream:d,text:i(async u=>{let v="";for await(let I of d({...u,reasoningEffort:u.reasoningEffort??null,tools:[],toolChoice:"none"}))I.text&&(v+=I.text),I.usage!==void 0&&u.onUsage?.(I.usage,I.costUsd);return v},"text"),authorize:i(async()=>{await n("authorize",{})},"authorize"),ceiling:i(async(u,v,I)=>e.object({configurationHash:e.string().regex(/^[a-f0-9]{64}$/),maximumChargeUsd:e.number().finite().positive()}).strict().parse(await n("ceiling",{modelId:u,pricing:v,maximumOutputTokens:I})),"ceiling"),compute:i((u,v=!1)=>n("compute",u,v?AbortSignal.timeout(6e4):r),"compute"),checkpoint:i(u=>n("checkpoint",u),"checkpoint")}}i(qe,"createHostedAdvancedModelBridge");import Qe from"node:path";var Me=e.string().regex(/^[a-f0-9]{64}$/),Ke=e.object({taskId:e.string().min(1),caseHash:Me,session:C,turn:$,events:e.array(se).max(1e4),traceHash:Me,output:e.string().nullable()}).strict(),K=e.object({packageValue:ae,contexts:e.array(Ke).max(1e4)}).strict();async function _e(s){let r=V(s.pin),t=K.parse(s.privateSource),n=ie(t.packageValue),d=M(e.object({manifest:e.unknown(),result:e.unknown()}).parse(s.evidence));async function m(){if(await s.authorize(),r.actorId!==s.actor.actorId||r.teamId!==s.actor.teamId||r.profileRef.source!=="openpond_git"||r.profileRef.repositoryId!==s.source.repositoryId||r.profileRef.profileId!==s.source.profileId||r.sourceRevision!==s.source.sourceRevision)throw new Error("The actual hosted advanced source owner changed.")}if(i(m,"authorize"),await m(),d.manifest.teamId!==r.teamId||d.result.status!=="completed"||d.manifest.id!==r.evidence.id||d.result.contentHash!==r.evidence.contentHash||R(d.manifest.dataset)!==R(r.externalDatasetBinding.dataset))throw new Error("The exact completed hosted advanced evidence changed.");if(n.contentHash!==r.externalDatasetBinding.packageHash||n.contentHash!==r.adaptationDatasetBinding.packageHash||n.environment.contentHash!==r.environmentPolicy.environmentHash||n.environment.contract.networkPolicy!=="none")throw new Error("The hosted advanced environment/package is not the exact reviewed closed-network release.");if(n.taskset.tools.some(l=>l.name==="web_search"||l.name==="web_fetch"))throw new Error("Live external tools require their own qualified frozen-observation and budget owner.");if(n.verifierSet.graders.some(l=>l.kind==="human"||l.kind==="model_judge"&&l.calibrationStatus!=="passed"))throw new Error("Advanced hosted evaluation needs executable, qualified independent graders.");let u=r.adaptationTaskIds.map(l=>n.taskset.tasks.find(H=>H.id===l)),v=r.holdoutTaskIds.map(l=>n.taskset.tasks.find(H=>H.id===l)),I=new Set(u.map(l=>l?.clusterKey));if(u.some(l=>!l||l.split!==r.adaptationSplit)||v.some(l=>!l||l.split!==r.holdoutSplit||I.has(l.clusterKey)))throw new Error("Hosted adaptation/holdout membership or independent families changed.");let x=await s.store.getHarnessReleaseRecord(r.baselineRelease.contentHash);if(!x||x.harnessRelease.id!==r.baselineRelease.id||s.workspace.currentChannel.release?.contentHash!==r.baselineRelease.contentHash)throw new Error("The exact hosted source baseline is unavailable.");let y=await D({workspaceId:typeof x.agentSnapshot.metadata.workspaceId=="string"?x.agentSnapshot.metadata.workspaceId:x.workspaceId,sourceDir:Qe.join(x.bundlePath,"source")});if(z(y).protectedClosureHash!==r.privateClosureHash)throw new Error("The protected hosted Profile closure changed.");let h={packageValue:n,authorize:m};return{read:i(async l=>{if(V(l).contentHash!==r.contentHash)throw new Error("The actual hosted advanced admission changed.");return await m(),h},"read"),evidence:i(async()=>(await m(),d),"evidence"),authorize:i(async()=>m(),"authorize"),reviewContext:i(async(l,H)=>{await m();let E=t.contexts.filter(p=>p.taskId===H),P=d.result.cases.find(p=>p.identity.caseId===H);if(E.length!==1||!P||P.status!=="completed")throw new Error("The actual retained hosted review case is unavailable.");let w=E[0];if(w.caseHash!==R(P)||w.session.cloudTeamId!==r.teamId||w.turn.sessionId!==w.session.id||w.turn.status==="in_progress"||w.events.some(p=>p.sessionId!==w.session.id||p.turnId!==w.turn.id))throw new Error("The actual hosted review session/turn/trace binding changed.");return await m(),{...w,case:{output:w.output}}},"reviewContext")}}i(_e,"createHostedAdvancedRefinerSource");import{randomUUID as Ze}from"node:crypto";function Oe(s){async function r(n){let d=await s.store.getSession(n);if(!d||d.cloudTeamId!==s.teamId||d.metadata?.hostedAdvancedTurnId!==s.hostedTurnId||d.metadata?.parentModelRunId!==s.modelRunId)throw new Error("The actual hosted advanced case owner changed.");return d}i(r,"read");async function t(n,d,m){let u=await r(n),v=e.object({action:e.string()}).passthrough().parse(d).action;return ce.parse(await s.bridge.compute({session:u,payload:d,turnId:m??null},v==="sandbox_stop"||v==="sandbox_receipts"))}return i(t,"tool"),{createSession:i(async n=>{await s.bridge.authorize();let d=C.partial().parse(n),m=new Date().toISOString();if(d.metadata?.parentModelRunId!==s.modelRunId)throw new Error("The foreground case belongs to another actual advanced Run.");let u=C.parse({...d,id:`hosted-advanced-case-${Ze()}`,provider:"openpond",currentProfile:s.profileRef,cloudTeamId:s.teamId,cwd:null,codexThreadId:null,appId:null,appName:null,createdAt:m,updatedAt:m,status:"active",pinned:!1,archived:!1,order:0,hiddenFromDefaultSidebar:!0,title:d.title??"Advanced hosted evaluation case",metadata:{...d.metadata,source:"hosted-advanced-evaluation",hostedAdvancedTurnId:s.hostedTurnId,actorId:s.actorId,workspaceTarget:"hosted",benchmarkRuntime:"hosted_managed_work"}});return await s.store.insertSessionAtFront(u),await s.bridge.authorize(),u},"createSession"),getSession:r,runtimeEventsForSession:i(n=>s.store.runtimeEventsForSession(n),"runtimeEventsForSession"),executeWorkspaceTool:i((n,d,m)=>t(n,d,m?.turnId),"executeWorkspaceTool"),settleCostEvidence:i((n,d)=>t(n,{action:"sandbox_receipts",args:{}},d?.turnId),"settleCostEvidence")}}i(Oe,"createHostedAdvancedWorkRuntime");var Xe=e.string().regex(/^[a-f0-9]{64}$/),Be=e.object({version:e.string().min(1),source:e.string().min(1),effectiveAt:e.string().min(1),inputUsdPerMillionTokens:e.number().finite().nonnegative(),cachedInputUsdPerMillionTokens:e.number().finite().nonnegative(),outputUsdPerMillionTokens:e.number().finite().nonnegative()}).strict(),Z=e.object({operation:e.enum(["prepare_run","start","read","resume","cancel"]),request:ue,configuration:e.object({id:e.string().min(1),profileId:e.string().min(1),etag:Xe}).strict(),refinerRelease:Y,privateSource:K,upstream:e.object({modelId:e.string().min(1),revision:e.string().min(1),pricing:Be}).strict(),judgeModels:e.array(e.object({modelId:e.string().min(1),revision:e.string().min(1),pricing:Be}).strict()).max(50),bridge:Q.optional(),hostedTurnId:e.string().min(1).optional()}).strict();async function Fe(s){let r=Z.parse(s.request),t=r.request.pin;if(r.request.model.providerId!=="openpond"||r.request.model.modelId!==r.upstream.modelId||r.configuration.id!==r.request.modelId||r.configuration.profileId!==t.profileRef.profileId||r.request.profileId!==t.profileRef.profileId||r.refinerRelease.id!==t.refinerRelease.id||r.refinerRelease.contentHash!==t.refinerRelease.contentHash)throw new Error("The actual hosted advanced configuration/model/Refiner admission changed.");let n=`advanced-refiner-${R([t.actorId,t.teamId,t.operationId,t.contentHash,r.request.model,[r.request.seed],1]).slice(0,40)}`;if((r.operation==="start"||r.operation==="resume")&&(!r.bridge||!r.hostedTurnId))throw new Error("The actual hosted advanced execution grant is unavailable.");let m=r.bridge?qe(r.bridge,s.signal):null;if(r.bridge&&(r.bridge.jobId!==n||r.bridge.pinHash!==t.contentHash))throw new Error("The hosted model bridge belongs to another actual Run.");let u=i(async()=>{s.signal.throwIfAborted(),await s.authorize(),m&&await m.authorize()},"authorize"),v=await _e({...s,pin:t,privateSource:r.privateSource,authorize:u}),I=we(ve(s.storeDir));await I.persistRelease(r.refinerRelease),await I.writeSource(ee(r.refinerRelease.profile)),I.binding()?.release.contentHash!==t.refinerRelease.contentHash&&await I.transition(r.refinerRelease,{operation:"activate",bindingChanged:!0,actor:s.actor.actorId,reason:"Pin the actual hosted evaluation fork to its admitted immutable Refiner release.",authoringSkillHash:null});let y=ke({storeDir:s.storeDir,authorize:i(async()=>u(),"authorize"),packageForPin:i(async()=>r.privateSource.packageValue,"packageForPin"),prepareModel:i(async(c,f,k)=>{if(!m)throw new Error("The actual hosted model ceiling owner is unavailable.");return m.ceiling(c,k,f)},"prepareModel")}),h=i(async function*(c){let f=be(),k=[r.upstream,...r.judgeModels].find(q=>q.modelId===c.model.modelId);if(!m||!f||!k)throw new Error("The actual hosted model and durable shared budget are unavailable.");let{signal:T,...j}=c,O=c.maxOutputTokens??4096,L=await f.call(c.requestId,j,k.pricing,O,async()=>{let q=[],F=null;for await(let N of m.stream({...c,signal:AbortSignal.any([T,f.signal])}))q.push(N),N.costUsd!==void 0&&(F=(F??0)+N.costUsd);return{value:q,costUsd:F}});for(let q of L)yield q},"stream"),l=i(async c=>{let f="";for await(let k of h({...c,reasoningEffort:c.reasoningEffort??null,tools:[],toolChoice:"none"}))k.text&&(f+=k.text),k.usage!==void 0&&c.onUsage?.(k.usage,k.costUsd);return f},"text"),H=Ie({store:s.store,storeDir:s.storeDir}),E=await H.projectSelected({profileId:t.profileRef.profileId,package:r.privateSource.packageValue,adaptationSplit:t.adaptationSplit,holdoutSplit:t.holdoutSplit}),P=i(async c=>{let f=C.partial().parse(c),k=new Date().toISOString(),T=C.parse({...f,id:`hosted-review-${R([n,f.metadata]).slice(0,40)}`,cloudTeamId:t.teamId,currentProfile:t.profileRef,appId:null,appName:null,cwd:null,codexThreadId:null,title:f.title??"Hosted Refiner review",provider:"openpond",createdAt:k,updatedAt:k,status:"active",pinned:!1,archived:!1,order:0,metadata:{...f.metadata,hostedAdvancedTurnId:r.hostedTurnId??null,actorId:t.actorId,source:"hosted-advanced-evaluation"}});return await s.store.insertSessionAtFront(T),T},"createSession"),w=m&&r.hostedTurnId?Oe({store:s.store,bridge:m,teamId:t.teamId,actorId:t.actorId,hostedTurnId:r.hostedTurnId,modelRunId:n,profileRef:t.profileRef}):void 0,p=Se({store:s.store,storeDir:s.storeDir,modelJudge:ye({store:s.store,modelText:l}),modelText:l,modelStream:h,workRuntime:w,resolveTasksetRelease:i(c=>H.releaseForTaskset(c),"resolveTasksetRelease")}),a={store:s.store,storeDir:s.storeDir,evaluation:p,benchmarkTasksets:H,loadProfileState:i(async()=>s.profile,"loadProfileState"),resolveEvaluationProject:i(async(c,f)=>{if(await u(),c!==r.configuration.id||f!==r.configuration.profileId)throw new Error("The actual hosted evaluation configuration changed.");return{id:c,profileId:f}},"resolveEvaluationProject"),resolveUpstreamModel:i(async()=>({providerId:"openpond",modelId:r.upstream.modelId,revision:r.upstream.revision,pricing:r.upstream.pricing}),"resolveUpstreamModel"),advancedBoundary:y,refinerStream:i(async function*(c){if(!c.requestId)throw new Error("A hosted Refiner call needs its actual retained ID.");yield*h({model:c.model,reasoningEffort:"low",messages:c.messages,tools:[],toolChoice:"none",requestId:c.requestId,maxOutputTokens:4096,signal:c.signal,hostedTokenPricing:c.pricing})},"refinerStream")},o=Re({...a,reviewQuality:He({source:v,boundary:y,benchmark:a,createSession:P})}),g={advancedEvaluation:t,selectedTaskset:{id:E.id,revision:E.revision,contentHash:E.contentHash},modelId:r.request.modelId,profileId:r.request.profileId,model:r.request.model,reasoningEffort:r.request.reasoningEffort,seeds:[r.request.seed],repetitions:1,maximumSpendUsd:t.maximumCostUsd},b,S=i(()=>{o.cancel(n)},"stop");s.signal.addEventListener("abort",S,{once:!0});try{if(r.operation==="prepare_run")return o.prepare(g);let c=await s.store.getModelRun(n);if(!c||c.evaluation?.benchmarkId!=="harness-refiner"||c.evaluation.advancedEvaluation?.contentHash!==t.contentHash)throw new Error("The original hosted advanced Run is unavailable.");if(r.operation==="read")return c;if(r.operation==="cancel")return o.cancel(n);if(m){let T=!1;b=setInterval(()=>{T||(T=!0,s.store.getModelRun(n).then(j=>j?m.checkpoint(j):void 0).catch(()=>S()).finally(()=>{T=!1}))},2e3)}r.operation==="start"?await o.start(g):await o.resume(n);let f=await o.wait(n);if(!f)throw new Error("The actual hosted advanced engine did not retain its Run.");let k=pe.parse(f);return m&&await m.checkpoint(k),k}finally{b&&clearInterval(b),s.signal.removeEventListener("abort",S),await o.close(),await y.close()}}i(Fe,"runHostedAdvancedRefinerOwner");import{promises as _}from"node:fs";import A from"node:path";var Ye=e.object({actor:e.object({actorId:e.string().min(1),teamId:e.string().min(1)}).strict(),storeDir:e.string().min(1),source:e.object({repoPath:e.string().min(1),repositoryId:e.string().min(1),profileId:e.string().min(1),sourceRevision:e.string().regex(/^[a-f0-9]{40,64}$/),filesHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),evidence:e.object({manifest:e.unknown(),result:e.unknown()}).strict(),operation:e.enum(["advanced","describe","command","queue","cancelQueued","bind","finish","beginTesting","cancelTesting","failTesting","comparison","beginAdoption","export","source","tool","toolCatalog"]),advanced:Z.optional(),command:le.optional(),candidateId:e.string().optional(),revision:e.number().int().positive().optional(),actualSession:C.optional(),actualTurn:$.optional(),recipe:e.unknown().optional(),baseline:e.object({manifest:e.unknown(),result:e.unknown()}).strict().optional(),candidate:e.object({manifest:e.unknown(),result:e.unknown()}).strict().optional(),humanProjection:oe.optional(),humanSelections:re.optional(),adoptionReadback:e.unknown().optional(),failure:e.object({message:e.string().max(2e3),cancelled:e.boolean()}).strict().optional(),tool:e.object({name:e.string(),args:e.record(e.string(),e.unknown()),callId:e.string()}).strict().optional(),agentRuntime:e.object({source:e.string(),cliRelativePath:e.string()}).strict().optional(),queued:e.object({sessionId:e.string(),turnId:e.string(),createdAt:e.string().datetime()}).strict().optional()}).strict();async function We(s,r=new AbortController().signal){r.throwIfAborted();let t=Ye.parse(s),{actor:n,source:d}=t;if(!A.isAbsolute(t.storeDir)||!A.isAbsolute(d.repoPath))throw new Error("Hosted candidate owner requires contained absolute private paths.");let m=M(t.evidence),u=[t.baseline,t.candidate].filter(l=>l!==void 0).map(l=>M(l));if(m.manifest.teamId!==n.teamId)throw new Error("Hosted candidate evidence belongs to another team.");if(await B(d.repoPath)!==d.filesHash)throw new Error("The admitted hosted Profile source bytes changed.");let v=`hosted-improve-source-${R([n,d.repositoryId,d.profileId,d.sourceRevision]).slice(0,40)}`,I=A.join(t.storeDir,"library","profile-experiment-origins",v);if(await _.mkdir(A.dirname(I),{recursive:!0,mode:448}),await _.stat(I).catch(()=>null)||await _.cp(d.repoPath,I,{recursive:!0,errorOnExist:!0,force:!1}),await B(I)!==d.filesHash)throw new Error("The private accepted Profile cache changed.");let x=await W(I,d.profileId);if(x.error||!x.sourcePath)throw new Error("The accepted hosted Profile does not compile.");let y=new xe(t.storeDir),h;try{let l=await y.getHarnessWorkspace(v);l||(l=(await fe({store:y,storeDir:t.storeDir,id:v,ownerId:n.actorId,name:d.profileId,profile:x,sourceRevision:d.sourceRevision,repositoryId:d.repositoryId,selectionEligible:!1,originMetadata:{profileExperimentOrigin:{schemaVersion:"openpond.profileExperimentOrigin.v1",...n,sourceRevision:d.sourceRevision,sourcePath:I,sourceFilesHash:d.filesHash},hostedImprovementOwner:!0}})).workspace);let H={source:"openpond_git",repositoryId:d.repositoryId,profileId:d.profileId},E=i(async()=>{if(r.throwIfAborted(),await B(I)!==d.filesHash||await B(d.repoPath)!==d.filesHash)throw new Error("Accepted source changed during private candidate I/O.")},"authorizer");if(t.operation==="advanced"){if(!t.advanced)throw new Error("Missing actual hosted advanced owner command.");return Fe({store:y,storeDir:t.storeDir,actor:n,profile:x,workspace:l,source:d,evidence:m,request:t.advanced,authorize:E,signal:r})}let P=i(()=>{if(!t.adoptionReadback)throw new Error("Actual hosted Git owner has not returned its sealed adoption receipt.");return e.object({receipt:e.object({id:e.string(),contentHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),activeRelease:e.object({id:e.string(),contentHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),testedCandidateRelease:e.object({id:e.string(),contentHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict(),executableSourceHash:e.string().regex(/^[a-f0-9]{64}$/),protectedClosureHash:e.string().regex(/^[a-f0-9]{64}$/)}).strict().parse(t.adoptionReadback)},"readback");h=Ee({store:y,storeDir:t.storeDir,projectHumanComparison:i(async()=>{if(await E(),!t.humanProjection)throw new Error("The current private Human comparison owner is unavailable.");return ne(t.humanProjection)},"projectHumanComparison"),authorizeOwner:i(async(a,o,g,b,S)=>{if(R(a)!==R(n)||R(g)!==R(H)||o.id!==v)throw new Error("Hosted candidate source owner changed.");if(S.requireCurrentBase&&(o.currentChannel.release?.contentHash!==l.currentChannel.release?.contentHash||d.sourceRevision!==o.metadata.profileExperimentOrigin.sourceRevision))throw new Error("Hosted candidate accepted revision changed.");await E()},"authorizeOwner"),loadEvidence:i(async(a,o)=>{if(R(a)!==R(n))throw new Error("Hosted candidate evidence actor changed.");let g=[m,...u].find(b=>b.manifest.id===o.id&&b.result.contentHash===o.contentHash&&b.manifest.teamId===n.teamId);if(!g)throw new Error("Hosted candidate exact retained evidence changed.");return await E(),g},"loadEvidence"),freezeProfile:i(async(a,o,g)=>{let b=await y.getHarnessReleaseRecord(o.baseRelease.contentHash);if(!b)throw new Error("Hosted source baseline is missing.");let S=await D({workspaceId:typeof b.agentSnapshot.metadata.workspaceId=="string"?b.agentSnapshot.metadata.workspaceId:b.workspaceId,sourceDir:A.join(b.bundlePath,"source")}),c=await D({workspaceId:typeof g.release.agentSnapshot.metadata.workspaceId=="string"?g.release.agentSnapshot.metadata.workspaceId:g.release.workspaceId,sourceDir:A.join(g.release.bundlePath,"source")});return Ae({store:y,storeDir:t.storeDir,state:o},async f=>{await Ce({...f,state:o,base:S,candidate:c});let k=await W(f.root,o.profileRef.profileId),T=await U({storeDir:t.storeDir,state:o,profile:k,sourceRevision:g.profileSourceRevision,base:S}),j=z(T);if(j.protectedClosureHash!==z(S).protectedClosureHash)throw new Error("Hosted source candidate changed protected evaluation closure.");let O=A.join(t.storeDir,"library","profile-candidate-archives",o.id,j.contentHash);await _.mkdir(A.dirname(O),{recursive:!0,mode:448}),await _.stat(O).catch(()=>null)||await _.cp(f.root,O,{recursive:!0,errorOnExist:!0,force:!1});let L=await W(O,o.profileRef.profileId);if((await U({storeDir:t.storeDir,state:o,profile:L,sourceRevision:g.profileSourceRevision,base:S})).harnessRelease.contentHash!==T.harnessRelease.contentHash)throw new Error("Hosted frozen Profile archive readback differs.");let F=await he({storeDir:t.storeDir,workspaceId:g.release.workspaceId,compiled:T,createdAt:g.release.createdAt});return await E(),{...g,release:F,executableIdentity:j}})},"freezeProfile"),adopt:i(async()=>P(),"adopt"),readAdoption:i(async()=>t.adoptionReadback?P():null,"readAdoption"),rollback:i(async()=>P(),"rollback")});let w=Pe({store:y,improvements:h,actorId:i(async()=>n.actorId,"actorId"),teamId:i(async()=>n.teamId,"teamId"),loadProfile:i(async a=>{if(R(a)!==R(H))throw new Error("Candidate requested another source owner.");return x},"loadProfile"),agentRuntime:t.agentRuntime,operationSignal:r,loadAgentRuntime:De({storeDir:t.storeDir,loadArchive:ze({storeDir:t.storeDir})})});if(t.operation==="describe"){await E();let a=l.currentChannel.release,o=a?await y.getHarnessReleaseRecord(a.contentHash):null;if(!a||!o||o.workspaceId!==l.id||o.harnessRelease.id!==a.id||o.harnessRelease.contentHash!==a.contentHash||l.ownerScope.kind!=="personal"||l.ownerScope.id!==n.actorId)throw new Error("Hosted source release is unavailable to its admitted owner.");let g=await D({workspaceId:typeof o.agentSnapshot.metadata.workspaceId=="string"?o.agentSnapshot.metadata.workspaceId:v,sourceDir:A.join(o.bundlePath,"source")});if(g.harnessRelease.contentHash!==o.harnessRelease.contentHash)throw new Error("Hosted source compiler readback changed.");let b=await ge({runtime:{release:o},ref:H,sourceRevision:d.sourceRevision,harnessRelease:a}),S=i(c=>{let f=g.sourceFiles.find(k=>k.path===c)?.bytes;return f?JSON.parse(Buffer.from(f).toString("utf8")):null},"readJson");return await E(),{workspaceId:v,ownerRevision:l.revision,profileRef:H,baseRelease:l.currentChannel.release,profileSourceRevision:d.sourceRevision,manifest:g.manifest,executable:z(g),catalogHash:b.catalogHash,workflows:S("workflows/catalog.json"),actions:S("workflows/actions.json")}}if(t.operation==="command"){let a=t.command;if(!a)throw new Error("Missing canonical Improve command.");if(a.operation==="start")return h.start(n,a.request);if(a.operation==="list")return h.list(n,a.limit,a.cursor);if(!("id"in a))throw new Error("Discover exact hosted source choices through the authenticated host.");let o=await h.read(n,a.id);if(a.operation==="read")return o;if(a.operation==="instructions"){let g=await h.partition(n,o.id),b=await me(g.sourceRoot,o.component.path),S=await _.readFile(b);if(S.length>25e4)throw new Error("Instruction editor size limit exceeded.");let c=S.toString("utf8");if(o.component.kind==="workflow"){let f=JSON.parse(c).workflows?.find(k=>k.id===o.component.workflowId);if(f?.invocation?.kind!=="instructions")throw new Error("Use Work to author this executable Workflow.");c=f.invocation.instructions}return{candidateId:o.id,candidateRevision:o.revision,path:o.component.path,text:c,textHash:J(c),contentHash:J(S)}}if(a.operation==="humanComparisonOptions")throw new Error("Read the actual paired sources through their current hosted Experiment owner.");if(!("revision"in a)||o.revision!==a.revision)throw new Error("Candidate revision changed.");if(a.operation==="edit")return h.edit(n,o.id,o.revision,a.expectedFileHash,a.text);if(a.operation==="mode")return h.setMode(n,o.id,o.revision,a.mode);if(a.operation==="freeze")return h.freeze(n,o.id,o.revision);if(a.operation==="useCandidate")return h.useCandidate(n,o.id,o.revision);if(a.operation==="rollback")return h.rollback(n,o.id,o.revision);if(a.operation==="discard")return h.discard(n,o.id,o.revision);throw new Error("Hosted Work/test lifecycle requires the actual host turn/run admission.")}if(!t.candidateId)throw new Error("Missing actual hosted candidate ID.");let p=await h.read(n,t.candidateId);if(t.operation==="beginAdoption")return h.beginAdoption(n,p.id,t.revision);if(t.operation==="export")return je(y,p,"hosted_execution");if(t.operation==="source"){if(!p.frozen||p.status!=="frozen"&&p.status!=="testing"&&p.status!=="ready_for_review"&&p.status!=="adopting")throw new Error("Only the exact frozen source can enter hosted preparation.");let a=await y.getHarnessReleaseRecord(p.frozen.release.contentHash);if(!a||a.harnessRelease.id!==p.frozen.release.id)throw new Error("Frozen hosted source is unavailable.");let o=await D({workspaceId:typeof a.agentSnapshot.metadata.workspaceId=="string"?a.agentSnapshot.metadata.workspaceId:a.workspaceId,sourceDir:A.join(a.bundlePath,"source")}),g=z(o);if(o.harnessRelease.contentHash!==p.frozen.release.contentHash||g.contentHash!==p.frozen.executableSourceHash||g.protectedClosureHash!==p.frozen.protectedClosureHash)throw new Error("Frozen hosted executable changed.");let b=A.join(t.storeDir,"library","profile-candidate-archives",p.id,g.contentHash),S=await W(b,p.profileRef.profileId);if(S.error||!S.sourcePath)throw new Error("Frozen actual Profile archive is unavailable.");let c=await y.getHarnessReleaseRecord(p.baseRelease.contentHash);if(!c)throw new Error("Frozen baseline is unavailable.");let f=await D({workspaceId:typeof c.agentSnapshot.metadata.workspaceId=="string"?c.agentSnapshot.metadata.workspaceId:c.workspaceId,sourceDir:A.join(c.bundlePath,"source")}),k=await U({storeDir:t.storeDir,state:p,profile:S,sourceRevision:p.frozen.profileSourceRevision,base:f});if(k.harnessRelease.contentHash!==p.frozen.release.contentHash||z(k).contentHash!==g.contentHash)throw new Error("Actual frozen Profile archive differs from its compiled release.");return{state:p,profileSource:{repoPath:b,profileId:p.profileRef.profileId,sourceRevision:p.frozen.profileSourceRevision,repositoryId:p.profileRef.repositoryId,filesHash:await B(b)},sourcePackage:te({agentSnapshot:o.agentSnapshot,harnessRelease:o.harnessRelease,files:new Map(o.sourceFiles.map(T=>[T.path,T.bytes]))})}}if(t.operation==="comparison"){if(!t.baseline||!t.candidate)throw new Error("Missing actual hosted paired evidence.");return h.completeComparison(n,p.id,t.revision,M(t.baseline),M(t.candidate),t.humanSelections)}if(t.operation==="beginTesting")return h.beginTesting(n,p.id,t.revision,t.recipe);if(t.operation==="failTesting"){if(!t.failure)throw new Error("Missing actual hosted test failure.");return h.failTesting(n,p.id,t.revision,t.failure.message,t.failure.cancelled)}if(t.operation==="cancelTesting")return h.cancelTesting(n,p.id,t.revision);if(t.operation==="queue"){if(!t.queued)throw new Error("Missing actual hosted queue admission.");return h.queueAuthoringTurn(n,p.id,t.revision,t.queued.sessionId,t.queued.turnId,t.queued.createdAt)}if(t.operation==="cancelQueued")return h.cancelQueuedAuthoring(n,p.id,t.revision);if(t.operation==="bind"||t.operation==="finish"||t.operation==="tool"||t.operation==="toolCatalog"){if(!t.actualSession||!t.actualTurn||t.actualTurn.sessionId!==t.actualSession.id||t.actualSession.cloudTeamId!==n.teamId||R(t.actualSession.currentProfile)!==R(H))throw new Error("Actual hosted candidate Work binding changed.");let a=t.actualSession,o=t.actualTurn,g=await y.getSession(a.id);if(!g)await y.insertSessionAtFront(a);else if(R({...g,updatedAt:a.updatedAt,status:a.status})!==R(a))throw new Error("Retained hosted candidate session changed.");let b=await y.getTurn(o.id);if(!b)await y.insertTurn(o);else{if(b.sessionId!==o.sessionId||R(b.metadata.refinementCandidate)!==R(o.metadata.refinementCandidate)||b.prompt!==o.prompt)throw new Error("Actual hosted candidate turn identity changed.");await y.updateTurn(o.id,()=>o)}if(t.operation==="bind")return h.bindAuthoringTurn(n,p.id,t.revision,a.id,o.id);if(t.operation==="finish")return h.finishAuthoring(n,p.id,t.revision);if(t.operation==="toolCatalog")return await w.candidateAuthoringForTurn(a.id,o.id),[...G({loadProfileState:i(async()=>x,"loadProfileState"),resolveCandidateProfile:w.resolveCandidateProfile,executeCandidateAgentCommand:w.executeCandidateAgentCommand}),...Te(async()=>{throw new Error("Hosted file tools require a bound private call.")}),{name:"exec_command",description:"Run a bounded command confined to the selected public candidate component.",parameters:{type:"object",additionalProperties:!1,properties:{command:{type:"string",minLength:1,maxLength:2e4},cwd:{type:"string"}},required:["command"]}},{name:"view_image",description:"Inspect an image inside the candidate public source tree.",parameters:{type:"object",additionalProperties:!1,properties:{path:{type:"string"}},required:["path"]}}].map(({name:c,description:f,parameters:k})=>({name:c,description:f,parameters:k}));if(!t.tool)throw new Error("Missing candidate tool call.");await w.authorizeCandidateTool({sessionId:a.id,turnId:o.id,name:t.tool.name});let S={session:a,turnId:o.id,args:t.tool.args,callId:t.tool.callId,signal:r,workspaceDiffBaseline:null,turnPermissions:{approvalPolicy:"never",sandbox:"workspace-write",codexPermissionMode:"default"},provider:a.provider,model:o.modelRef?.modelId??a.modelRef?.modelId??"",mentionedApps:[],userPrompt:o.prompt,turnMetadata:o.metadata};if(t.tool.name==="get_profile"||t.tool.name.startsWith("agent_")||t.tool.name==="ask_user"){let c=G({loadProfileState:i(async()=>x,"loadProfileState"),resolveCandidateProfile:w.resolveCandidateProfile,executeCandidateAgentCommand:w.executeCandidateAgentCommand}).find(f=>f.name===t.tool.name);if(!c)throw new Error("Candidate authoring tool is unavailable.");return c.execute(S)}if(t.tool.name==="view_image")return w.executeCandidateImage(S);if(t.tool.name==="exec_command")return w.executeCandidateCommand(S,{session:a,turnId:o.id,command:e.string().min(1).max(2e4).parse(t.tool.args.command),cwd:typeof t.tool.args.cwd=="string"?t.tool.args.cwd:null,signal:S.signal,source:"model_tool"});if(["list_files","read_files","search_files","workspace_status","write_file","write_files","edit_file","delete_file"].includes(t.tool.name))return w.executeCandidateWorkspaceTool({session:a,turnId:o.id,request:de.parse({action:t.tool.name,args:t.tool.args,source:"chat_action"})});throw new Error("This hosted tool has no confined candidate transport.")}throw new Error("Unknown private hosted Improve operation.")}finally{await h?.close(),await y.close()}}i(We,"runHostedExperimentImprovementOwner");async function wr(s){if(s.length!==1||s[0].length>16e3||!/^[A-Za-z0-9_-]+$/.test(s[0]))throw new Error("Hosted Improve IPC argument is invalid.");let r=e.object({directory:e.string(),file:e.string().regex(/^[A-Za-z0-9_-]+\.json$/)}).strict().parse(JSON.parse(Buffer.from(s[0],"base64url").toString("utf8")));if(!Ue.isAbsolute(r.directory))throw new Error("Hosted Improve IPC requires an absolute private directory.");let t=Ue.join(r.directory,r.file),n=await $e.lstat(t);if(!n.isFile()||n.isSymbolicLink()||n.size>32*1024*1024)throw new Error("Hosted Improve private request exceeds its byte/type limit.");let d=new AbortController,m=i(()=>d.abort(new Error("The actual hosted owner stopped this private operation.")),"stop");process.once("SIGTERM",m),process.once("SIGINT",m);try{let u=await We(JSON.parse(await $e.readFile(t,"utf8")),d.signal),v=Buffer.from(JSON.stringify(u));if(v.length>32*1024*1024)throw new Error("Hosted Improve private response exceeds its byte limit.");process.stdout.write(`${v.toString("utf8")}
 `)}finally{process.removeListener("SIGTERM",m),process.removeListener("SIGINT",m)}}i(wr,"runHostedExperimentImprovementCli");export{wr as runHostedExperimentImprovementCli};

package/dist/chunks/{hosted-profile-compiler-PQFIWO5B.js → hosted-profile-compiler-JODGZCA7.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a,b,c,d}from"./chunk-LVL4TIAO.js";import"./chunk-7MXZJQS7.js";import"./chunk-7YR6CBNU.js";import"./chunk-5CMKGD43.js";import"./chunk-HA7EFTPF.js";import"./chunk-L3KS6MOV.js";import"./chunk-7QEE4PBI.js";import"./chunk-XDMUD4FW.js";import"./chunk-IKWDL3RB.js";import"./chunk-6MI4KKWF.js";import"./chunk-QRWPEW3Z.js";import"./chunk-6FA3GG5B.js";import"./chunk-Z3D4C7BG.js";import"./chunk-WV2APYWG.js";import"./chunk-PJM2G7SG.js";import"./chunk-SDEDAHV6.js";import"./chunk-DALJMSJG.js";import"./chunk-TGIB24K4.js";import"./chunk-NNF6KCHD.js";import"./chunk-ES2ISFJC.js";import"./chunk-H7IWJN7L.js";import"./chunk-FTQLQJQV.js";export{a as compileHostedProfileRelease,b as runHostedProfileCompilerCli,c as runHostedProfileEvaluationCli,d as runHostedProfileInspectionCli};
+import{a,b,c,d}from"./chunk-OC6UWKBL.js";import"./chunk-7MXZJQS7.js";import"./chunk-7YR6CBNU.js";import"./chunk-5CMKGD43.js";import"./chunk-HA7EFTPF.js";import"./chunk-L3KS6MOV.js";import"./chunk-7QEE4PBI.js";import"./chunk-XDMUD4FW.js";import"./chunk-IKWDL3RB.js";import"./chunk-6MI4KKWF.js";import"./chunk-QRWPEW3Z.js";import"./chunk-6FA3GG5B.js";import"./chunk-Z3D4C7BG.js";import"./chunk-WV2APYWG.js";import"./chunk-PJM2G7SG.js";import"./chunk-SDEDAHV6.js";import"./chunk-DALJMSJG.js";import"./chunk-TGIB24K4.js";import"./chunk-NNF6KCHD.js";import"./chunk-ES2ISFJC.js";import"./chunk-H7IWJN7L.js";import"./chunk-FTQLQJQV.js";export{a as compileHostedProfileRelease,b as runHostedProfileCompilerCli,c as runHostedProfileEvaluationCli,d as runHostedProfileInspectionCli};

package/dist/chunks/{hosted-profile-evaluation-preparation-JROREZH3.js → hosted-profile-evaluation-preparation-F2SPAEKK.js}

@@ -1,3 +1,3 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as V}from"./chunk-LVL4TIAO.js";import{c as E,f as j,h as D}from"./chunk-DFM5JTS2.js";import{k as x}from"./chunk-VOIH2BQA.js";import{a as b}from"./chunk-T5LE6R5S.js";import"./chunk-QJRXTIOC.js";import"./chunk-SDO4FC4Q.js";import"./chunk-IR3YJWPW.js";import"./chunk-7K72KS5C.js";import{ha as I,qa as C}from"./chunk-7MXZJQS7.js";import"./chunk-7YR6CBNU.js";import"./chunk-5CMKGD43.js";import"./chunk-HA7EFTPF.js";import"./chunk-L3KS6MOV.js";import"./chunk-7QEE4PBI.js";import"./chunk-XDMUD4FW.js";import{ta as m}from"./chunk-IKWDL3RB.js";import"./chunk-6MI4KKWF.js";import{Ca as S}from"./chunk-QRWPEW3Z.js";import"./chunk-6FA3GG5B.js";import"./chunk-Z3D4C7BG.js";import"./chunk-WV2APYWG.js";import{ma as H}from"./chunk-PJM2G7SG.js";import{Ga as y,j as P,m as R}from"./chunk-SDEDAHV6.js";import"./chunk-DALJMSJG.js";import"./chunk-TGIB24K4.js";import{m as d}from"./chunk-NNF6KCHD.js";import{b as e}from"./chunk-ES2ISFJC.js";import"./chunk-H7IWJN7L.js";import{a as s}from"./chunk-FTQLQJQV.js";import{readFile as h,realpath as z,lstat as A}from"node:fs/promises";import l from"node:path";var B=e.object({compile:e.object({repoPath:e.string(),profileId:e.string(),repositoryId:e.string(),sourceRevision:e.string(),workspaceId:e.string(),outputDir:e.string()}).strict(),externalDatasetPackage:e.string().regex(/^external-dataset-[a-f0-9-]+\.json$/).optional(),sourceCandidate:H.optional(),action:e.enum(["catalog","prepare"]),request:e.unknown().optional()}).strict();async function G(t){let o=await V(t.compile);if(t.sourceCandidate){let a=await C({workspaceId:typeof o.agentSnapshot.metadata.workspaceId=="string"?o.agentSnapshot.metadata.workspaceId:o.workspaceId,sourceDir:l.join(o.bundlePath,"source")});if(d(t.sourceCandidate.harnessRelease)!==d({id:o.harnessRelease.id,contentHash:o.harnessRelease.contentHash})||b(a).protectedClosureHash!==t.sourceCandidate.protectedClosureHash)throw new Error("Private candidate source differs from its actual frozen compiled release.")}let g={release:I.parse({schemaVersion:"openpond.localHarnessReleaseRecord.v1",workspaceId:o.workspaceId,sourceRevision:o.sourceRevision,agentSnapshot:P.parse(o.agentSnapshot),harnessRelease:R.parse(o.harnessRelease),bundlePath:o.bundlePath,createdAt:new Date().toISOString()})},p={source:"openpond_git",profileId:t.compile.profileId,repositoryId:t.compile.repositoryId},f=t.compile.sourceRevision,c={id:o.harnessRelease.id,contentHash:o.harnessRelease.contentHash},w=s(async a=>x({...a,runtime:g}),"loadCatalog"),q=await w({ref:p,sourceRevision:f,harnessRelease:c}),k=l.join(o.bundlePath,"source"),v=o.harnessRelease.files.some(a=>a.path==="workflows/catalog.json")?await(async()=>{let a=y({agentSnapshot:o.agentSnapshot,harnessRelease:o.harnessRelease,catalogBytes:await h(l.join(k,"workflows/catalog.json")),actionBytes:await h(l.join(k,"workflows/actions.json"))});return a.catalog.workflows.map(r=>({workflow:r,binding:{schemaVersion:"openpond.profileWorkflowBinding.v1",profileId:p.profileId,sourceRevision:f,harnessRelease:c,catalogHash:a.catalogHash,workflowId:r.id}}))})():[],T={profileRef:p,sourceRevision:f,harnessRelease:c,workflows:v};if(t.action==="catalog")return{...q,workflows:v};let u=t.externalDatasetPackage?await(async()=>{let a=await z(t.compile.outputDir),r=l.join(a,t.externalDatasetPackage),i=await A(r);if(!i.isFile()||i.isSymbolicLink()||i.size>32*1024*1024||await z(r)!==r)throw new Error("External Dataset cache is outside its bounded verifier storage.");let n=e.object({packageValue:e.unknown(),graders:e.array(e.object({id:e.string(),version:e.string(),contentHash:e.string(),feedbackKey:e.string(),release:e.object({id:e.string(),revision:e.union([e.string(),e.number()]),contentHash:e.string()}).strict().nullable()}).strict()).max(100)}).strict().parse(JSON.parse(await h(r,"utf8")));return{packageValue:S(n.packageValue),graders:n.graders}})():null;return D({privateGradingPreflight:s(a=>E(a,!1),"privateGradingPreflight"),sourceCandidate:t.sourceCandidate,resolveExternalDataset:u?async a=>{let r=m({binding:a,...u}),i=d(o.harnessRelease.files.filter(n=>n.visibility!=="policy").map(n=>({path:n.path,asset:n})));if(a.protectedProfileClosureHash!==i)throw new Error("The selected Profile's protected evaluator closure changed.");return{binding:a,packageValue:r,authorize:s(async()=>{m({binding:a,...u})},"authorize")}}:void 0,loadCatalog:w,selectedWorkflows:s(async()=>T,"selectedWorkflows"),placement:"remote",loadTasksetPackage:s(async a=>{let r=await j({runtime:g,definition:a,harnessRelease:c});if(!r)throw new Error("Hosted evaluation Taskset is absent from its published Profile release.");return r},"loadTasksetPackage"),modelConfigurationHash:s(async(a,r)=>{if(a.providerId!=="openpond"||!r.hostModelConfigurationHash)throw new Error("Hosted evaluation preparation requires a trusted model configuration receipt.");return r.hostModelConfigurationHash},"modelConfigurationHash")})(t.request)}s(G,"prepareHostedProfileEvaluation");async function ae(t){if(t.length!==1||t[0].length>48e3||!/^[A-Za-z0-9_-]+$/.test(t[0]))throw new Error("Hosted Profile evaluation preparation argument is invalid.");let o=B.parse(JSON.parse(Buffer.from(t[0],"base64url").toString("utf8")));process.stdout.write(`${JSON.stringify(await G(o))}
+import{a as V}from"./chunk-OC6UWKBL.js";import{c as I,f as j,h as D}from"./chunk-ZCWHVLJ4.js";import{k as x}from"./chunk-OE2VF6KU.js";import{a as E}from"./chunk-T5LE6R5S.js";import"./chunk-QJRXTIOC.js";import"./chunk-SDO4FC4Q.js";import"./chunk-IR3YJWPW.js";import"./chunk-7K72KS5C.js";import{ha as b,qa as C}from"./chunk-7MXZJQS7.js";import"./chunk-7YR6CBNU.js";import"./chunk-5CMKGD43.js";import"./chunk-HA7EFTPF.js";import"./chunk-L3KS6MOV.js";import"./chunk-7QEE4PBI.js";import"./chunk-XDMUD4FW.js";import{ta as m}from"./chunk-IKWDL3RB.js";import"./chunk-6MI4KKWF.js";import{Ca as S}from"./chunk-QRWPEW3Z.js";import"./chunk-6FA3GG5B.js";import"./chunk-Z3D4C7BG.js";import"./chunk-WV2APYWG.js";import{ma as H}from"./chunk-PJM2G7SG.js";import{Ga as y,j as P,m as R}from"./chunk-SDEDAHV6.js";import"./chunk-DALJMSJG.js";import"./chunk-TGIB24K4.js";import{m as d}from"./chunk-NNF6KCHD.js";import{b as e}from"./chunk-ES2ISFJC.js";import"./chunk-H7IWJN7L.js";import{a as s}from"./chunk-FTQLQJQV.js";import{readFile as h,realpath as z,lstat as T}from"node:fs/promises";import l from"node:path";var B=e.object({compile:e.object({repoPath:e.string(),profileId:e.string(),repositoryId:e.string(),sourceRevision:e.string(),workspaceId:e.string(),outputDir:e.string()}).strict(),externalDatasetPackage:e.string().regex(/^external-dataset-[a-f0-9-]+\.json$/).optional(),sourceCandidate:H.optional(),action:e.enum(["catalog","prepare"]),request:e.unknown().optional()}).strict();async function G(t){let o=await V(t.compile);if(t.sourceCandidate){let a=await C({workspaceId:typeof o.agentSnapshot.metadata.workspaceId=="string"?o.agentSnapshot.metadata.workspaceId:o.workspaceId,sourceDir:l.join(o.bundlePath,"source")});if(d(t.sourceCandidate.harnessRelease)!==d({id:o.harnessRelease.id,contentHash:o.harnessRelease.contentHash})||E(a).protectedClosureHash!==t.sourceCandidate.protectedClosureHash)throw new Error("Private candidate source differs from its actual frozen compiled release.")}let g={release:b.parse({schemaVersion:"openpond.localHarnessReleaseRecord.v1",workspaceId:o.workspaceId,sourceRevision:o.sourceRevision,agentSnapshot:P.parse(o.agentSnapshot),harnessRelease:R.parse(o.harnessRelease),bundlePath:o.bundlePath,createdAt:new Date().toISOString()})},p={source:"openpond_git",profileId:t.compile.profileId,repositoryId:t.compile.repositoryId},f=t.compile.sourceRevision,c={id:o.harnessRelease.id,contentHash:o.harnessRelease.contentHash},w=s(async a=>x({...a,runtime:g}),"loadCatalog"),q=await w({ref:p,sourceRevision:f,harnessRelease:c}),k=l.join(o.bundlePath,"source"),v=o.harnessRelease.files.some(a=>a.path==="workflows/catalog.json")?await(async()=>{let a=y({agentSnapshot:o.agentSnapshot,harnessRelease:o.harnessRelease,catalogBytes:await h(l.join(k,"workflows/catalog.json")),actionBytes:await h(l.join(k,"workflows/actions.json"))});return a.catalog.workflows.map(r=>({workflow:r,binding:{schemaVersion:"openpond.profileWorkflowBinding.v1",profileId:p.profileId,sourceRevision:f,harnessRelease:c,catalogHash:a.catalogHash,workflowId:r.id}}))})():[],A={profileRef:p,sourceRevision:f,harnessRelease:c,workflows:v};if(t.action==="catalog")return{...q,workflows:v};let u=t.externalDatasetPackage?await(async()=>{let a=await z(t.compile.outputDir),r=l.join(a,t.externalDatasetPackage),i=await T(r);if(!i.isFile()||i.isSymbolicLink()||i.size>32*1024*1024||await z(r)!==r)throw new Error("External Dataset cache is outside its bounded verifier storage.");let n=e.object({packageValue:e.unknown(),graders:e.array(e.object({id:e.string(),version:e.string(),contentHash:e.string(),feedbackKey:e.string(),release:e.object({id:e.string(),revision:e.union([e.string(),e.number()]),contentHash:e.string()}).strict().nullable()}).strict()).max(100)}).strict().parse(JSON.parse(await h(r,"utf8")));return{packageValue:S(n.packageValue),graders:n.graders}})():null;return D({privateGradingPreflight:s(a=>I(a,!1),"privateGradingPreflight"),sourceCandidate:t.sourceCandidate,resolveExternalDataset:u?async a=>{let r=m({binding:a,...u}),i=d(o.harnessRelease.files.filter(n=>n.visibility!=="policy").map(n=>({path:n.path,asset:n})));if(a.protectedProfileClosureHash!==i)throw new Error("The selected Profile's protected evaluator closure changed.");return{binding:a,packageValue:r,authorize:s(async()=>{m({binding:a,...u})},"authorize")}}:void 0,loadCatalog:w,selectedWorkflows:s(async()=>A,"selectedWorkflows"),placement:"remote",boundedWorkComputeAvailable:!0,loadTasksetPackage:s(async a=>{let r=await j({runtime:g,definition:a,harnessRelease:c});if(!r)throw new Error("Hosted evaluation Taskset is absent from its published Profile release.");return r},"loadTasksetPackage"),modelConfigurationHash:s(async(a,r)=>{if(a.providerId!=="openpond"||!r.hostModelConfigurationHash)throw new Error("Hosted evaluation preparation requires a trusted model configuration receipt.");return r.hostModelConfigurationHash},"modelConfigurationHash")})(t.request)}s(G,"prepareHostedProfileEvaluation");async function ae(t){if(t.length!==1||t[0].length>48e3||!/^[A-Za-z0-9_-]+$/.test(t[0]))throw new Error("Hosted Profile evaluation preparation argument is invalid.");let o=B.parse(JSON.parse(Buffer.from(t[0],"base64url").toString("utf8")));process.stdout.write(`${JSON.stringify(await G(o))}
 `)}s(ae,"runHostedProfileEvaluationCli");export{G as prepareHostedProfileEvaluation,ae as runHostedProfileEvaluationCli};