openpond 0.2.45 → 0.2.47

package/dist/chunks/chunk-3HBCQNRN.js

@@ -0,0 +1,2 @@
+import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
+import{k as H,na as R}from"./chunk-36YNUUZL.js";import{J as I,N as C}from"./chunk-HA7EFTPF.js";import{l as k,m as b}from"./chunk-NNF6KCHD.js";import{a as h}from"./chunk-FTQLQJQV.js";import{promises as s}from"node:fs";import n from"node:path";import{execFile as O}from"node:child_process";import{promisify as A}from"node:util";var T=A(O);async function v(e,i,a={}){return(await T("git",["-C",e,...i],{encoding:"utf8",timeout:3e4,maxBuffer:8388608,env:{...process.env,...a}})).stdout.trim()}h(v,"candidateGit");async function j(e,i=I,a=!1){if(e.profileRef.source!=="local"&&!a)throw new Error("This Git adapter requires the actual installed local Profile owner; hosted adoption needs its published-source owner.");let r=await i(e.profileRef);if(r.error||r.mode!=="local"||r.activeProfile!==e.profileRef.profileId||!r.repoPath||!r.sourcePath)throw new Error("The original Profile owner is unavailable.");let l=await s.realpath(r.repoPath),p=await s.realpath(r.sourcePath),f=n.relative(l,p);if(f.startsWith("..")||n.isAbsolute(f))throw new Error("Profile source escapes its actual Git owner.");return{profile:r,repo:l,source:p,sourceRelative:f}}h(j,"originalCandidateProfile");async function Pe(e,i,a=!0){if(!/^[a-f0-9]{40}$/.test(i)||await v(e,["rev-parse","HEAD"])!==i)throw new Error("The original Profile Git revision changed.");if(a&&await v(e,["status","--porcelain","--untracked-files=all"]))throw new Error("Commit or reconcile changes in the original Profile before this operation.")}h(Pe,"assertCandidateGitHead");async function ye(e,i){let a=await j(e.state,e.loadProfile,e.allowRemoteSnapshot),r=n.join(e.storeDir,"library","harnesses","profile-candidate-staging");await s.mkdir(r,{recursive:!0,mode:448});let l=await s.mkdtemp(n.join(r,"source-"));await s.rmdir(l),await v(a.repo,["worktree","add","--detach",l,e.commit]);try{let p=await C({repoPath:l,profileId:e.state.profileRef.profileId});if(p.error||!p.sourcePath)throw new Error("The staged original Profile cannot be loaded.");return await i({root:l,profile:p,repo:a.repo,source:p.sourcePath})}finally{await v(a.repo,["worktree","remove","--force",l]).catch(()=>{}),await s.rm(l,{recursive:!0,force:!0})}}h(ye,"withCandidateProfileWorktree");function y(e){if(!e||n.isAbsolute(e)||e.includes("\\")||e.split("/").some(i=>!i||i==="."||i===".."))throw new Error("Original Profile patch path is not contained.");return e}h(y,"portable");async function ve(e){let i=e.base.manifest.metadata.profileSourceBindings;if(!Array.isArray(i))throw new Error("The original release needs compiler source bindings before candidate adoption.");let a=e.state.component.path,r=e.state.component,l=new Map(i.map(t=>{if(!t||typeof t!="object"||typeof t.releasedPath!="string"||typeof t.profileRelativePath!="string")throw new Error("Original Profile source binding is invalid.");return[y(t.releasedPath),y(t.profileRelativePath)]})),p=r.kind==="instruction"&&a==="instructions/system.md"&&e.base.manifest.metadata.profileInstructionSourcePath==="instructions/system.md"&&!l.has(a),f=l.get(a)??(p?"instructions/system.md":void 0),S=r.kind==="workflow"?`workflows/${r.workflowId}/`:r.kind==="agent"?a.split("/").slice(0,2).join("/")+"/":n.posix.dirname(a)+"/",w=f?n.posix.dirname(f)+"/":null;if(r.kind==="workflow"&&!f&&(w=S),r.kind==="agent"){let t=e.profile.agents.find(o=>a.startsWith(`agents/${o.id}/`));if(!t)throw new Error("The selected Agent original source is unavailable.");let c=y(t.path);w=((await s.lstat(n.join(e.source,c))).isDirectory()?c:n.posix.dirname(c))+"/"}let E=new Map(e.base.sourceFiles.map(t=>[t.path,t])),x=new Map(e.candidate.sourceFiles.map(t=>[t.path,t])),g=new Map;for(let t of new Set([...E.keys(),...x.keys()])){let c=E.get(t),m=x.get(t);if(c?.asset.contentHash===m?.asset.contentHash)continue;if(c&&c.asset.visibility!=="policy"||m&&m.asset.visibility!=="policy")throw new Error("A candidate cannot change protected Profile source.");let o=l.get(t)??(p&&t===a?f:void 0);if(!o&&t.startsWith(S)&&w&&(o=w+t.slice(S.length)),!o&&t==="workflows/catalog.json"&&[...l.keys()].some(P=>/^workflows\/[^/]+\/(PROMPT\.md|ACTION\.json)$/.test(P))||!o&&t==="dependency-lock/profile-import.json")continue;if(!o)throw new Error(`The selected component has no original Profile source mapping: ${t}`);if(o=y(o),o!==f&&!(r.kind!=="instruction"&&w&&o.startsWith(w)))throw new Error("Candidate patch changes a different original Profile component.");if(H(o))throw new Error("Candidate patch crosses the protected Profile evaluation closure.");let u=g.get(o),d=m?.bytes??null;if(u&&(u.before!==(c?.asset.contentHash??null)||b(u.bytes===null?null:Array.from(u.bytes))!==b(d===null?null:Array.from(d))))throw new Error("Public component aliases disagree on the original source edit.");g.set(o,{before:p&&t===a?null:c?.asset.contentHash??null,bytes:d})}for(let[t,c]of g){let m=n.join(e.source,...t.split("/")),o=n.dirname(m);for(;!await s.stat(o).catch(()=>null);)o=n.dirname(o);let u=n.relative(await s.realpath(e.source),await s.realpath(o));if(u.startsWith("..")||n.isAbsolute(u))throw new Error("Candidate patch escapes the staged Profile source.");let d=await s.lstat(m).catch(()=>null);if(d?.isSymbolicLink()||d&&!d.isFile())throw new Error("Original source patch needs a regular file.");let P=d?await s.readFile(m):null;if((P?k(P):null)!==c.before)throw new Error("Original source bytes differ from the exact candidate baseline.");c.bytes===null?await s.rm(m):(await s.mkdir(n.dirname(m),{recursive:!0}),await s.writeFile(m,c.bytes,{mode:d?.mode??384}))}return[...g.keys()]}h(ve,"applyCandidateProfileSource");function Se(e){let i=e.base.manifest.metadata.profileRepositoryId;return R({storeDir:e.storeDir,workspaceId:e.state.ownerWorkspaceId,name:e.base.manifest.name,profile:e.profile,sourceRevision:e.sourceRevision,...typeof i=="string"?{repositoryId:i}:{}})}h(Se,"compileOriginalCandidateProfile");export{v as a,j as b,Pe as c,ye as d,ve as e,Se as f};

package/dist/chunks/chunk-4UTBLHFC.js

@@ -0,0 +1,2 @@
+import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
+import{n as b,o as E,r as x,u as T,w as C,x as _}from"./chunk-WV2APYWG.js";import{Y as M,h as S,i as A}from"./chunk-PJM2G7SG.js";import{i as k,m,n as y}from"./chunk-NNF6KCHD.js";import{b as p}from"./chunk-ES2ISFJC.js";import{a as g}from"./chunk-FTQLQJQV.js";var w=p.object({sessionId:p.string().trim().min(1).max(500),turnId:p.string().trim().min(1).max(500)}).strict();async function $(e){T(e.manifest,e.taskset,e.catalog),_(e.taskset,e.metricSource);let o=e.manifest.profileEvaluation,t=e.catalog.definitions.find(r=>r.id===o.definitionId),d=new Map(e.taskset.tasks.map(r=>[r.id,r])),u=[],s=[];for(let r of e.manifest.population){e.signal?.throwIfAborted();let l=d.get(r.taskId),a=await e.loadCompletedMember?.(r)??null;if(a){let{receipt:f,grade:I}=a,{contentHash:P,...U}=S.parse(I);if(!E(f)||P!==m(U)||f.id!==r.receiptId||f.runManifest.id!==e.manifest.id||f.runManifest.contentHash!==e.manifest.contentHash||f.taskId!==r.taskId||f.seed!==r.seed||f.graderEvidenceRefs.length!==1||f.graderEvidenceRefs[0]?.contentHash!==P||I.taskHash!==m(l)||I.graderSetHash!==m(e.taskset.graders)||f.metadata.score!==I.score||f.metadata.passed!==I.passed)throw new Error(`Retained Profile evaluation member ${r.receiptId} differs from its admitted evidence.`);s.push(I),u.push(f);continue}let i=await e.execute({task:M(l),seed:r.seed,source:o,...e.signal?{signal:e.signal}:{}});e.signal?.throwIfAborted();let c=await A({task:l,evidence:i.evidence,graders:e.taskset.graders,...e.modelJudge?{modelJudge:e.modelJudge}:{},...e.customVerifier?{customVerifier:e.customVerifier}:{},...e.signal?{signal:e.signal}:{}}),v=k.parse(await e.saveGrade(c));if(v.contentHash!==c.contentHash)throw new Error("Stored Profile evaluation grade differs from its graded evidence.");let H=i.failureClass??c.failureClass,R=b({schemaVersion:"openpond.attemptReceipt.v1",id:r.receiptId,runManifest:{id:e.manifest.id,contentHash:e.manifest.contentHash},taskId:r.taskId,seed:r.seed,terminal:i.terminal,failureClass:H,outputHash:i.evidence.infrastructureError?null:m(i.evidence.output),traceHash:i.traceHash,artifactRefs:i.artifactRefs,graderEvidenceRefs:[v],startedAt:i.startedAt,completedAt:i.completedAt,latencyMs:i.latencyMs,costUsd:i.costUsd,metadata:{gradingRole:"evaluation",score:c.score,passed:c.passed,rewardEligible:c.rewardEligible,sourceRevision:o.sourceRevision,definitionHash:o.definitionHash,environmentHash:o.environmentHash,...i.retainedEvidenceRef?{retainedEvidenceRef:w.parse(i.retainedEvidenceRef)}:{}}});await e.saveReceipt(R),s.push(c),u.push(R)}let n=await C({manifest:e.manifest,taskset:e.taskset,receipts:u,...e.metricSource!==void 0?{source:e.metricSource}:{},...e.signal?{signal:e.signal}:{}},e.metricExecutor),h=s.filter(r=>r.passed).length/s.length;return{receipts:u,grades:s,metric:n,passRate:h,passed:h>=t.criterion.minimumPassRate&&(!t.criterion.requireComplete||u.every(r=>r.terminal))}}g($,"executeProfileEvaluationRun");function V(e,o){if(o==="agent_action"){let t=e.find(d=>d.name==="workspace_action_result"&&d.action==="profile_workflow_action")?.output;if(typeof t!="string")throw new Error("Profile Agent action evaluation did not produce a released action result.");return t}return e.filter(t=>t.name==="assistant.delta"&&typeof t.output=="string").map(t=>t.output??"").join("")}g(V,"profileEvaluationOutput");async function L(e){let{store:o,manifest:t,receipt:d}=e;if(!d.metadata.retainedEvidenceRef)return null;let u=w.parse(d.metadata.retainedEvidenceRef),[s,n]=await Promise.all([o.getSession(u.sessionId),o.getTurn(u.turnId)]),h=s?.metadata?.profileEvaluationRun,r=t.profileEvaluation;if(!s||!n||n.sessionId!==s.id||!r||!s.currentProfile||m(s.currentProfile)!==m(e.profileRef)||m(h)!==m({id:t.id,contentHash:t.contentHash})||s.metadata?.taskId!==d.taskId||s.metadata?.seed!==d.seed||!n.completedAt||n.completedAt!==d.completedAt||n.startedAt!==d.startedAt||!n.harnessSnapshot||n.harnessSnapshot.harnessRelease.id!==r.harnessRelease.id||n.harnessSnapshot.harnessRelease.contentHash!==r.harnessRelease.contentHash||t.policy.kind!=="model"||!n.modelRef||n.modelRef.providerId!==t.policy.model.provider||n.modelRef.modelId!==t.policy.model.model)throw new Error("Retained Profile policy turn differs from its admitted case.");let l=await o.runtimeEventsForTurn(n.id);if(l.some(f=>f.turnId!==n.id)||m(l)!==d.traceHash)throw new Error("Retained Profile policy trace differs from its immutable receipt.");let a=V(l,r.target.kind);if(d.outputHash!==null&&m({text:a})!==d.outputHash)throw new Error("Retained Profile output differs from its immutable receipt.");let i=await o.listModelUsageRecords({sessionId:s.id,turnId:n.id,limit:1e4}),c=e.eventAfterId?l.findIndex(f=>f.id===e.eventAfterId):-1;if(e.eventAfterId&&c<0)throw new Error("Profile trace cursor is unavailable in this case.");let v=c+1,H=Math.max(1,Math.min(250,e.eventLimit??100)),R=l.slice(v,v+H);return{ref:u,prompt:n.prompt,output:d.outputHash===null?null:{text:a},partialOutput:d.outputHash===null&&a?{text:a}:null,error:n.error,events:R,eventCount:l.length,nextEventCursor:v+R.length<l.length?R.at(-1).id:null,usage:i,usageCoverage:i.length===1e4?"partial":"complete"}}g(L,"readProfileEvaluationPolicyEvidence");var te=p.object({runId:p.string().trim().min(1).max(240),receiptId:p.string().trim().min(1).max(500).optional(),eventAfterId:p.string().trim().min(1).max(500).optional(),eventLimit:p.number().int().min(1).max(250).optional(),manifest:x.optional()}).strict().refine(e=>e.receiptId||!e.eventAfterId&&!e.eventLimit,"Select a case before requesting its trace page.").refine(e=>!e.manifest||e.receiptId,"Interrupted-run inspection requires one exact case.");async function re(e){let o=await e.store.getProfileEvaluationRun(e.runId);if(o&&m(o.profileRef)!==m(e.profileRef))throw new Error("Profile evaluation run is unavailable in the selected Profile.");let t=o?.manifest??e.manifest;if(!t||!o&&!e.receiptId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");if(o){if(e.manifest&&e.manifest.contentHash!==o.manifest.contentHash)throw new Error("Profile evaluation run differs from its pinned manifest.")}else if(y(t,"Interrupted Profile evaluation manifest"),t.id!==e.runId||t.profileEvaluation?.profileId!==e.profileRef.profileId||e.profileRef.source==="openpond_git"&&t.metadata.profileRepositoryId!==e.profileRef.repositoryId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");let d=t.population.map((s,n)=>({member:s,index:n})).filter(({member:s})=>!e.receiptId||s.receiptId===e.receiptId);if(e.receiptId&&d.length!==1)throw new Error("Profile evaluation case is unavailable in the selected run.");let u=await Promise.all(d.map(async({member:s,index:n})=>{let h=o?null:await e.store.getProfileEvaluationReceipt(s.receiptId),r=o?.receiptRefs[n]??(h?{id:h.id,contentHash:h.contentHash}:null),l=o?.gradeRefs[n]??h?.graderEvidenceRefs[0];if(!r||!l)throw new Error("Profile evaluation run has incomplete case evidence.");let[a,i]=await Promise.all([e.store.getProfileEvaluationReceipt(r.id),e.store.getProfileEvaluationGrade(l.contentHash)]);if(!a||!o&&!E(a)||a.contentHash!==r.contentHash||a.id!==s.receiptId||a.taskId!==s.taskId||a.seed!==s.seed||a.runManifest.id!==t.id||a.runManifest.contentHash!==t.contentHash||!i||i.contentHash!==l.contentHash||!a.graderEvidenceRefs.some(c=>c.id===l.id&&c.contentHash===l.contentHash))throw new Error("Profile evaluation case evidence differs from its retained run.");return{taskId:s.taskId,seed:s.seed,receiptId:a.id,receiptHash:a.contentHash,traceHash:a.traceHash,outputHash:a.outputHash,score:i.score,passed:i.passed,gradingStatus:i.gradingStatus,failureClass:i.failureClass,artifactCount:a.artifactRefs.length,latencyMs:a.latencyMs,costUsd:a.costUsd,terminal:a.terminal,feedback:i.components.map(c=>({graderId:c.graderId,graderVersion:c.graderVersion,score:c.score,passed:c.passed,feedback:c.feedback})),...e.receiptId?{receipt:a,evidence:await L({store:e.store,profileRef:e.profileRef,manifest:t,receipt:a,eventAfterId:e.eventAfterId,eventLimit:e.eventLimit})}:{}}}));return{runId:t.id,sourceRevision:t.profileEvaluation?.sourceRevision??null,cases:u}}g(re,"inspectProfileEvaluationRun");export{$ as a,V as b,te as c,re as d};