openpond 0.2.58 → 0.2.60

package/dist/chunks/{chunk-DBK66CJP.js → chunk-OY5HC2P6.js}

@@ -1,5 +1,5 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{C as $,k as x,n as G,q as N,r as J,s as U,u as W}from"./chunk-FTEOCCPB.js";import{ua as z,wa as P}from"./chunk-25MBPJVY.js";import{N as L}from"./chunk-VYLGI3IH.js";import{Ka as M,Ma as B,o as C,oa as V,p as O,v as F}from"./chunk-ACYFMLWG.js";import{k as E,l as H,m as h}from"./chunk-KGNCKIFP.js";import{a as l}from"./chunk-FTQLQJQV.js";import{createHash as q}from"node:crypto";import{mkdir as K,readdir as ae,readFile as I,writeFile as X}from"node:fs/promises";import d from"node:path";async function je(e,t){let a=x.parse(e),r=["tasks","assets","environment","graders","rubrics","comparisons","metrics","fixtures"];await Promise.all(r.map(g=>K(d.join(t,g),{recursive:!0})));let o=await re(a,t);if(a.metrics.customAggregator){let g=d.join(t,a.metrics.customAggregator.module),k=q("sha256").update(await I(g)).digest("hex");a=x.parse({...a,metrics:{...a.metrics,customAggregator:{...a.metrics.customAggregator,contentHash:k}}})}let n=d.join(t,"taskset.json"),s=d.join(t,"tasks","tasks.jsonl"),f=d.join(t,"graders","graders.json"),i=d.join(t,"fixtures","grader-fixtures.json"),u=d.join(t,"metrics","policy.json"),c=d.join(t,"assets","manifest.json"),m=d.join(t,"environment","contract.json"),w=d.join(t,"rubrics","preference-review.md"),v=d.join(t,"comparisons","policy.json");return await Promise.all([...J(a)].map(([g,k])=>X(d.join(t,g),k,"utf8"))),{directory:t,files:[n,s,c,m,f,w,v,u,i,...o],draft:a}}l(je,"writeTasksetDraftPackage");async function Re(e){let t=await se(e),a=q("sha256");for(let r of t)a.update(d.relative(e,r).replaceAll(d.sep,"/")),a.update("\0"),a.update(await I(r)),a.update("\0");return a.digest("hex")}l(Re,"hashTasksetDraftPackage");async function Y(e,t){let a=d.basename(e)==="taskset.json"?e:d.join(e,"taskset.json"),r=d.dirname(a),o=l(async f=>{if(!t)return I(f,"utf8");let i=t.get(d.relative(r,f).replaceAll(d.sep,"/"));if(!i)throw Object.assign(new Error("Authored file is absent."),{code:"ENOENT"});return new TextDecoder("utf-8",{fatal:!0}).decode(i)},"read"),n=JSON.parse(await o(a)),s=x.safeParse(n);if(s.success){let[f,i,u,c,m,w]=await Promise.all([ne(d.join(r,"tasks","tasks.jsonl"),o),S(d.join(r,"graders","graders.json"),o),S(d.join(r,"fixtures","grader-fixtures.json"),o),S(d.join(r,"metrics","policy.json"),o),S(d.join(r,"environment","contract.json"),o),S(d.join(r,"comparisons","policy.json"),o)]);return x.parse({...s.data,...f===null?{}:{tasks:f},...i===null?{}:{graders:i},...u===null?{}:{graderFixtures:u},...c===null?{}:{metrics:c},...m===null?{}:{environment:m},...w===null?{}:{review:w}})}return $(n)}l(Y,"readTasksetDraftPackage");async function re(e,t){let a=N(e),r=[];for(let o of a){let n=d.join(t,o.relativePath);await K(d.dirname(n),{recursive:!0});try{await X(n,o.source,{encoding:"utf8",flag:"wx"})}catch(s){if(!oe(s))throw s}r.push(n)}return r}l(re,"writeConfiguredStarterCode");async function se(e){let t=[];async function a(r){let o=await ae(r,{withFileTypes:!0});o.sort((n,s)=>n.name.localeCompare(s.name));for(let n of o){let s=d.join(r,n.name);if(n.isSymbolicLink())throw new Error(`Taskset workspaces cannot contain symbolic links: ${s}`);n.isDirectory()?await a(s):n.isFile()&&t.push(s)}}return l(a,"visit"),await a(e),t}l(se,"packageFiles");async function S(e,t=a=>I(a,"utf8")){try{return JSON.parse(await t(e))}catch(a){if(Z(a))return null;throw a}}l(S,"readOptionalJson");async function ne(e,t=a=>I(a,"utf8")){try{return(await t(e)).split(/\r?\n/).map(r=>r.trim()).filter(Boolean).map(r=>JSON.parse(r))}catch(a){if(Z(a))return null;throw a}}l(ne,"readOptionalJsonLines");function Z(e){return e instanceof Error&&"code"in e&&e.code==="ENOENT"}l(Z,"isMissingFile");function oe(e){return e instanceof Error&&"code"in e&&e.code==="EEXIST"}l(oe,"isFileExists");import{constants as Q}from"node:fs";import{lstat as ie,open as ce,readdir as le}from"node:fs/promises";import A from"node:path";async function Oe(e){let t=A.resolve(e.directory);if(!(await ie(t)).isDirectory())throw new Error("Dataset upload requires a regular authored folder.");let a=new Map,r=0;async function o(c,m){if(m>30)throw new Error("Dataset folder nesting exceeds 30 levels.");for(let w of(await le(c,{withFileTypes:!0})).sort((v,g)=>v.name<g.name?-1:v.name>g.name?1:0)){if(w.name.startsWith("."))throw new Error("Dataset folders cannot include hidden files or directories.");if(w.isSymbolicLink()||!w.isDirectory()&&!w.isFile())throw new Error("Dataset folders can contain only regular files and directories.");let v=A.join(c,w.name);if(w.isDirectory()){await o(v,m+1);continue}if(a.size>=1e4)throw new Error("Dataset folder exceeds 10,000 files.");let g=await ce(v,Q.O_RDONLY|Q.O_NOFOLLOW);try{let k=await g.stat();if(!k.isFile()||k.size>U-r)throw new Error("Dataset folder exceeds the 64 MiB capture limit.");let j=new Uint8Array(k.size),R=0;for(;R<j.length;){let D=await g.read(j,R,j.length-R,R);if(!D.bytesRead)throw new Error("Dataset file changed during capture.");R+=D.bytesRead}let _=await g.stat();if(_.size!==k.size||_.mtimeMs!==k.mtimeMs||_.ctimeMs!==k.ctimeMs)throw new Error("Dataset file changed during capture.");r+=j.length,a.set(A.relative(t,v).replaceAll(A.sep,"/"),j)}finally{await g.close()}}}l(o,"visit"),await o(t,0);let n=await Y(t,a),s=[...a].map(([c,m])=>({path:c,contentHash:H(m),sizeBytes:m.byteLength})),f=h(s),i=x.parse({...n,id:e.datasetId??n.id,profileId:e.teamId,revision:e.expectedRevision+1,modelScope:null,status:"draft",publishedTasksetRef:null,metadata:{...n.metadata,importedTasksetPackage:{packageHash:f}}});return{workspace:W({schemaVersion:"openpond.tasksetDraftWorkspace.v1",draft:i,files:[...a].filter(([c])=>!G(c)).map(([c,m])=>({path:c,base64:Buffer.from(m).toString("base64"),contentHash:H(m),sizeBytes:m.byteLength}))}),sourceHash:f,sourceFiles:s,capturedBytes:r}}l(Oe,"captureTasksetDraftWorkspace");async function ee(e){let{binding:t,rewards:a}=e.learning,r=M(t,a);if(e.graders.length!==r.length||r.some(({source:i,reward:u})=>!e.graders.some(c=>c.id===i.graderId&&c.version===String(u.revision)&&h(c.metadata.rewardBinding)===h(L(t)))))throw new Error("Prepared Taskset graders do not match their immutable Reward binding.");let o=V.parse({id:e.task.id,clusterKey:e.task.clusterKey,split:e.task.split,input:e.task.input,expectedOutput:e.task.expectedOutput,policyVisibleContext:e.task.policyVisibleContext,privilegedContextRef:e.task.privilegedContextRef,artifactRefs:[],tags:e.task.tags}),n=await B({binding:t,rewards:a,task:o,signal:e.signal,evidence:{output:e.attempt.output,artifactRefs:e.attempt.artifactRefs,runtimeEventRefs:e.attempt.runtimeEventRefs,infrastructureError:e.attempt.infrastructureError},customVerifier:e.customVerifier?async({grader:i})=>{let u=e.graders.find(m=>m.id===i.id);if(u?.kind!=="custom_verifier")throw new Error("Bound verifier projection is missing.");let c=await e.customVerifier({grader:u,task:e.task,attempt:e.attempt,signal:e.signal});return{score:c.score,passed:c.passed,rewardEligible:i.rewardEligible,failureClass:null,feedback:[c.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:c.evidenceRefs??[]}}:void 0,modelJudge:e.modelJudge?async({grader:i})=>{let u=e.graders.find(m=>m.id===i.id);if(u?.kind!=="model_judge")throw new Error("Bound judge projection is missing.");let c=await e.modelJudge({grader:u,task:e.task,attempt:e.attempt});return{score:c.score,passed:c.passed,rewardEligible:i.rewardEligible,failureClass:null,feedback:[c.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:c.evidenceRefs??[]}}:void 0}),s=n.training.status==="not_configured"?n.evaluation:n.training,f=h(e.graders);return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([e.attempt.id,n.contentHash]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:f,score:s.score,passed:s.passed===!0,components:n.results.map(i=>{let u=t.sources.find(m=>m.graderId===i.graderId),c=e.graders.find(m=>m.id===i.graderId);return{graderId:i.graderId,graderVersion:c.version,score:i.normalizedScore,passed:i.passed===!0,hardGate:u.hardGate,rewardEligible:u.role==="training"&&i.status==="scored",feedback:i.message?.trim()||null,evidenceRefs:i.evidenceHashes,judge:c.kind==="model_judge"?c.judge:null,calibrationStatus:c.kind==="model_judge"?c.calibrationStatus:"not_applicable"}}),failureClass:e.attempt.infrastructureError?"infrastructure_failure":s.status!=="scored"?"grader_failure":s.passed?null:"policy_failure",feedback:n.results.flatMap(i=>i.message?.trim()?[i.message.trim()]:[]),rewardEligible:n.training.status==="scored",rewardComposition:n,createdAt:e.now?.()??new Date().toISOString()}}l(ee,"gradeLearningBatchAttempt");async function st(e){if(e.learning)return ee({...e,learning:e.learning});if(e.graders.some(o=>o.metadata.rewardBinding!==void 0))throw new Error("This Taskset requires its immutable public Reward binding to grade an attempt.");let t=h(e.graders),a=[];for(let o of e.graders)e.signal?.throwIfAborted(),a.push(e.attempt.infrastructureError?b(o,null,!1,"Infrastructure failure; no reward was produced.",[],!1):await me(o,e.task,e.attempt,e.modelJudge,e.customVerifier,e.signal));let r=F({graders:e.graders,infrastructureError:e.attempt.infrastructureError,components:a.map(o=>({...o,failureClass:o.score===null?"grader_failure":o.passed?null:"policy_failure"}))});return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([e.attempt.id,t,a]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:t,score:r.score,passed:r.passed,components:a,failureClass:r.failureClass,feedback:e.attempt.infrastructureError?[e.attempt.infrastructureError]:a.flatMap(o=>o.feedback?[o.feedback]:[]),rewardEligible:r.rewardEligible,createdAt:e.now?.()??new Date().toISOString()}}l(st,"gradeAttempt");async function me(e,t,a,r,o,n){if(e.kind==="model_judge"){if(!r)return b(e,null,!1,"Model judge runner is unavailable.",[],!1);if(e.calibrationStatus!=="passed")return b(e,null,!1,"Model judge calibration has not passed.",[],!1);let f=await r({grader:e,task:t,attempt:a});return b(e,f.score,f.passed,f.feedback,f.evidenceRefs??[])}if(e.kind==="human")return b(e,null,!1,"Human review is pending.",[],!1);if(e.kind==="custom_verifier"){if(!o)return b(e,null,!1,"Sandboxed verifier runner is unavailable.",[],!1);let f=await o({grader:e,task:t,attempt:a,signal:n});return b(e,f.score,f.passed,f.feedback,f.evidenceRefs??[])}let s=C({grader:O(e),task:t,evidence:a});return b(e,s.score,s.passed,s.feedback,[...a.artifactRefs,...a.runtimeEventRefs])}l(me,"runGrader");function b(e,t,a,r,o,n=e.rewardEligible){let s=t!==null&&Number.isFinite(t)?Math.max(0,Math.min(1,t)):null;return{graderId:e.id,graderVersion:e.version,score:s,passed:s!==null&&a,hardGate:e.hardGate,rewardEligible:s!==null&&n,feedback:r,evidenceRefs:o,judge:e.kind==="model_judge"?e.judge:null,calibrationStatus:e.kind==="model_judge"?e.calibrationStatus:"not_applicable"}}l(b,"component");import{access as ue,mkdir as T,readFile as Tt,writeFile as y}from"node:fs/promises";import p from"node:path";async function $t(e,t,a={}){let r=z.parse(e),o=P(r);if(!o.valid)throw new Error(o.issues.map(s=>`${s.code}: ${s.message}`).join(`
+import{C as $,k as x,n as G,q as N,r as J,s as U,u as W}from"./chunk-637UACKH.js";import{ua as z,wa as P}from"./chunk-A66CSUBJ.js";import{N as L}from"./chunk-SFNCAPCJ.js";import{A as O,G as F,Va as M,Xa as B,z as C,za as V}from"./chunk-RCA6NGVA.js";import{k as E,l as H,m as h}from"./chunk-BIPQCJ5V.js";import{a as l}from"./chunk-FTQLQJQV.js";import{createHash as q}from"node:crypto";import{mkdir as K,readdir as ae,readFile as I,writeFile as X}from"node:fs/promises";import d from"node:path";async function je(e,t){let a=x.parse(e),r=["tasks","assets","environment","graders","rubrics","comparisons","metrics","fixtures"];await Promise.all(r.map(g=>K(d.join(t,g),{recursive:!0})));let o=await re(a,t);if(a.metrics.customAggregator){let g=d.join(t,a.metrics.customAggregator.module),k=q("sha256").update(await I(g)).digest("hex");a=x.parse({...a,metrics:{...a.metrics,customAggregator:{...a.metrics.customAggregator,contentHash:k}}})}let n=d.join(t,"taskset.json"),s=d.join(t,"tasks","tasks.jsonl"),f=d.join(t,"graders","graders.json"),i=d.join(t,"fixtures","grader-fixtures.json"),u=d.join(t,"metrics","policy.json"),c=d.join(t,"assets","manifest.json"),m=d.join(t,"environment","contract.json"),w=d.join(t,"rubrics","preference-review.md"),v=d.join(t,"comparisons","policy.json");return await Promise.all([...J(a)].map(([g,k])=>X(d.join(t,g),k,"utf8"))),{directory:t,files:[n,s,c,m,f,w,v,u,i,...o],draft:a}}l(je,"writeTasksetDraftPackage");async function Re(e){let t=await se(e),a=q("sha256");for(let r of t)a.update(d.relative(e,r).replaceAll(d.sep,"/")),a.update("\0"),a.update(await I(r)),a.update("\0");return a.digest("hex")}l(Re,"hashTasksetDraftPackage");async function Y(e,t){let a=d.basename(e)==="taskset.json"?e:d.join(e,"taskset.json"),r=d.dirname(a),o=l(async f=>{if(!t)return I(f,"utf8");let i=t.get(d.relative(r,f).replaceAll(d.sep,"/"));if(!i)throw Object.assign(new Error("Authored file is absent."),{code:"ENOENT"});return new TextDecoder("utf-8",{fatal:!0}).decode(i)},"read"),n=JSON.parse(await o(a)),s=x.safeParse(n);if(s.success){let[f,i,u,c,m,w]=await Promise.all([ne(d.join(r,"tasks","tasks.jsonl"),o),S(d.join(r,"graders","graders.json"),o),S(d.join(r,"fixtures","grader-fixtures.json"),o),S(d.join(r,"metrics","policy.json"),o),S(d.join(r,"environment","contract.json"),o),S(d.join(r,"comparisons","policy.json"),o)]);return x.parse({...s.data,...f===null?{}:{tasks:f},...i===null?{}:{graders:i},...u===null?{}:{graderFixtures:u},...c===null?{}:{metrics:c},...m===null?{}:{environment:m},...w===null?{}:{review:w}})}return $(n)}l(Y,"readTasksetDraftPackage");async function re(e,t){let a=N(e),r=[];for(let o of a){let n=d.join(t,o.relativePath);await K(d.dirname(n),{recursive:!0});try{await X(n,o.source,{encoding:"utf8",flag:"wx"})}catch(s){if(!oe(s))throw s}r.push(n)}return r}l(re,"writeConfiguredStarterCode");async function se(e){let t=[];async function a(r){let o=await ae(r,{withFileTypes:!0});o.sort((n,s)=>n.name.localeCompare(s.name));for(let n of o){let s=d.join(r,n.name);if(n.isSymbolicLink())throw new Error(`Taskset workspaces cannot contain symbolic links: ${s}`);n.isDirectory()?await a(s):n.isFile()&&t.push(s)}}return l(a,"visit"),await a(e),t}l(se,"packageFiles");async function S(e,t=a=>I(a,"utf8")){try{return JSON.parse(await t(e))}catch(a){if(Z(a))return null;throw a}}l(S,"readOptionalJson");async function ne(e,t=a=>I(a,"utf8")){try{return(await t(e)).split(/\r?\n/).map(r=>r.trim()).filter(Boolean).map(r=>JSON.parse(r))}catch(a){if(Z(a))return null;throw a}}l(ne,"readOptionalJsonLines");function Z(e){return e instanceof Error&&"code"in e&&e.code==="ENOENT"}l(Z,"isMissingFile");function oe(e){return e instanceof Error&&"code"in e&&e.code==="EEXIST"}l(oe,"isFileExists");import{constants as Q}from"node:fs";import{lstat as ie,open as ce,readdir as le}from"node:fs/promises";import A from"node:path";async function Oe(e){let t=A.resolve(e.directory);if(!(await ie(t)).isDirectory())throw new Error("Dataset upload requires a regular authored folder.");let a=new Map,r=0;async function o(c,m){if(m>30)throw new Error("Dataset folder nesting exceeds 30 levels.");for(let w of(await le(c,{withFileTypes:!0})).sort((v,g)=>v.name<g.name?-1:v.name>g.name?1:0)){if(w.name.startsWith("."))throw new Error("Dataset folders cannot include hidden files or directories.");if(w.isSymbolicLink()||!w.isDirectory()&&!w.isFile())throw new Error("Dataset folders can contain only regular files and directories.");let v=A.join(c,w.name);if(w.isDirectory()){await o(v,m+1);continue}if(a.size>=1e4)throw new Error("Dataset folder exceeds 10,000 files.");let g=await ce(v,Q.O_RDONLY|Q.O_NOFOLLOW);try{let k=await g.stat();if(!k.isFile()||k.size>U-r)throw new Error("Dataset folder exceeds the 64 MiB capture limit.");let j=new Uint8Array(k.size),R=0;for(;R<j.length;){let D=await g.read(j,R,j.length-R,R);if(!D.bytesRead)throw new Error("Dataset file changed during capture.");R+=D.bytesRead}let _=await g.stat();if(_.size!==k.size||_.mtimeMs!==k.mtimeMs||_.ctimeMs!==k.ctimeMs)throw new Error("Dataset file changed during capture.");r+=j.length,a.set(A.relative(t,v).replaceAll(A.sep,"/"),j)}finally{await g.close()}}}l(o,"visit"),await o(t,0);let n=await Y(t,a),s=[...a].map(([c,m])=>({path:c,contentHash:H(m),sizeBytes:m.byteLength})),f=h(s),i=x.parse({...n,id:e.datasetId??n.id,profileId:e.teamId,revision:e.expectedRevision+1,modelScope:null,status:"draft",publishedTasksetRef:null,metadata:{...n.metadata,importedTasksetPackage:{packageHash:f}}});return{workspace:W({schemaVersion:"openpond.tasksetDraftWorkspace.v1",draft:i,files:[...a].filter(([c])=>!G(c)).map(([c,m])=>({path:c,base64:Buffer.from(m).toString("base64"),contentHash:H(m),sizeBytes:m.byteLength}))}),sourceHash:f,sourceFiles:s,capturedBytes:r}}l(Oe,"captureTasksetDraftWorkspace");async function ee(e){let{binding:t,rewards:a}=e.learning,r=M(t,a);if(e.graders.length!==r.length||r.some(({source:i,reward:u})=>!e.graders.some(c=>c.id===i.graderId&&c.version===String(u.revision)&&h(c.metadata.rewardBinding)===h(L(t)))))throw new Error("Prepared Taskset graders do not match their immutable Reward binding.");let o=V.parse({id:e.task.id,clusterKey:e.task.clusterKey,split:e.task.split,input:e.task.input,expectedOutput:e.task.expectedOutput,policyVisibleContext:e.task.policyVisibleContext,privilegedContextRef:e.task.privilegedContextRef,artifactRefs:[],tags:e.task.tags}),n=await B({binding:t,rewards:a,task:o,signal:e.signal,evidence:{output:e.attempt.output,artifactRefs:e.attempt.artifactRefs,runtimeEventRefs:e.attempt.runtimeEventRefs,infrastructureError:e.attempt.infrastructureError},customVerifier:e.customVerifier?async({grader:i})=>{let u=e.graders.find(m=>m.id===i.id);if(u?.kind!=="custom_verifier")throw new Error("Bound verifier projection is missing.");let c=await e.customVerifier({grader:u,task:e.task,attempt:e.attempt,signal:e.signal});return{score:c.score,passed:c.passed,rewardEligible:i.rewardEligible,failureClass:null,feedback:[c.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:c.evidenceRefs??[]}}:void 0,modelJudge:e.modelJudge?async({grader:i})=>{let u=e.graders.find(m=>m.id===i.id);if(u?.kind!=="model_judge")throw new Error("Bound judge projection is missing.");let c=await e.modelJudge({grader:u,task:e.task,attempt:e.attempt});return{score:c.score,passed:c.passed,rewardEligible:i.rewardEligible,failureClass:null,feedback:[c.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:c.evidenceRefs??[]}}:void 0}),s=n.training.status==="not_configured"?n.evaluation:n.training,f=h(e.graders);return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([e.attempt.id,n.contentHash]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:f,score:s.score,passed:s.passed===!0,components:n.results.map(i=>{let u=t.sources.find(m=>m.graderId===i.graderId),c=e.graders.find(m=>m.id===i.graderId);return{graderId:i.graderId,graderVersion:c.version,score:i.normalizedScore,passed:i.passed===!0,hardGate:u.hardGate,rewardEligible:u.role==="training"&&i.status==="scored",feedback:i.message?.trim()||null,evidenceRefs:i.evidenceHashes,judge:c.kind==="model_judge"?c.judge:null,calibrationStatus:c.kind==="model_judge"?c.calibrationStatus:"not_applicable"}}),failureClass:e.attempt.infrastructureError?"infrastructure_failure":s.status!=="scored"?"grader_failure":s.passed?null:"policy_failure",feedback:n.results.flatMap(i=>i.message?.trim()?[i.message.trim()]:[]),rewardEligible:n.training.status==="scored",rewardComposition:n,createdAt:e.now?.()??new Date().toISOString()}}l(ee,"gradeLearningBatchAttempt");async function st(e){if(e.learning)return ee({...e,learning:e.learning});if(e.graders.some(o=>o.metadata.rewardBinding!==void 0))throw new Error("This Taskset requires its immutable public Reward binding to grade an attempt.");let t=h(e.graders),a=[];for(let o of e.graders)e.signal?.throwIfAborted(),a.push(e.attempt.infrastructureError?b(o,null,!1,"Infrastructure failure; no reward was produced.",[],!1):await me(o,e.task,e.attempt,e.modelJudge,e.customVerifier,e.signal));let r=F({graders:e.graders,infrastructureError:e.attempt.infrastructureError,components:a.map(o=>({...o,failureClass:o.score===null?"grader_failure":o.passed?null:"policy_failure"}))});return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${h([e.attempt.id,t,a]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:t,score:r.score,passed:r.passed,components:a,failureClass:r.failureClass,feedback:e.attempt.infrastructureError?[e.attempt.infrastructureError]:a.flatMap(o=>o.feedback?[o.feedback]:[]),rewardEligible:r.rewardEligible,createdAt:e.now?.()??new Date().toISOString()}}l(st,"gradeAttempt");async function me(e,t,a,r,o,n){if(e.kind==="model_judge"){if(!r)return b(e,null,!1,"Model judge runner is unavailable.",[],!1);if(e.calibrationStatus!=="passed")return b(e,null,!1,"Model judge calibration has not passed.",[],!1);let f=await r({grader:e,task:t,attempt:a});return b(e,f.score,f.passed,f.feedback,f.evidenceRefs??[])}if(e.kind==="human")return b(e,null,!1,"Human review is pending.",[],!1);if(e.kind==="custom_verifier"){if(!o)return b(e,null,!1,"Sandboxed verifier runner is unavailable.",[],!1);let f=await o({grader:e,task:t,attempt:a,signal:n});return b(e,f.score,f.passed,f.feedback,f.evidenceRefs??[])}let s=C({grader:O(e),task:t,evidence:a});return b(e,s.score,s.passed,s.feedback,[...a.artifactRefs,...a.runtimeEventRefs])}l(me,"runGrader");function b(e,t,a,r,o,n=e.rewardEligible){let s=t!==null&&Number.isFinite(t)?Math.max(0,Math.min(1,t)):null;return{graderId:e.id,graderVersion:e.version,score:s,passed:s!==null&&a,hardGate:e.hardGate,rewardEligible:s!==null&&n,feedback:r,evidenceRefs:o,judge:e.kind==="model_judge"?e.judge:null,calibrationStatus:e.kind==="model_judge"?e.calibrationStatus:"not_applicable"}}l(b,"component");import{access as ue,mkdir as T,readFile as Tt,writeFile as y}from"node:fs/promises";import p from"node:path";async function $t(e,t,a={}){let r=z.parse(e),o=P(r);if(!o.valid)throw new Error(o.issues.map(s=>`${s.code}: ${s.message}`).join(`
 `));await T(p.join(t,"data"),{recursive:!0}),await T(p.join(t,"graders"),{recursive:!0}),await T(p.join(t,"environment"),{recursive:!0}),await T(p.join(t,"fixtures"),{recursive:!0});let n=[p.join(t,"taskset.json"),p.join(t,"capabilities.json"),p.join(t,"data","tasks.jsonl"),p.join(t,"graders","graders.json"),p.join(t,"fixtures","grader-fixtures.json"),p.join(t,"environment","taskset.ts")];await y(n[0],E(r),"utf8"),await y(n[1],E(r.capabilities),"utf8"),await y(n[2],r.tasks.map(s=>JSON.stringify(s)).join(`
 `)+`
 `,"utf8"),await y(n[3],E(r.graders),"utf8"),await y(n[4],E(r.graderFixtures),"utf8"),await y(n[5],pe(r),"utf8");for(let s of a.generatedFiles??[]){let f=te(s.path,s.role),i=p.join(t,f);await T(p.dirname(i),{recursive:!0}),await y(i,s.content,"utf8"),n.push(i)}for(let s of r.graders){if(s.kind!=="custom_verifier")continue;let f=te(s.module,"verifier");try{await ue(p.join(t,f))}catch{throw new Error(`Custom verifier ${s.id} is missing generated source ${f}.`)}}return{directory:t,files:n,taskset:r}}l($t,"buildTaskset");function te(e,t){let a=e.replaceAll("\\","/");if(p.posix.isAbsolute(a)||a.split("/").includes(".."))throw new Error(`Generated ${t} path must stay inside the Taskset: ${e}.`);let r=t==="verifier"?"graders/":t==="environment"?"environment/":"fixtures/";if(!a.startsWith(r))throw new Error(`Generated ${t} path must start with ${r}.`);return a}l(te,"safeGeneratedPath");function pe(e){return`// Generated by @openpond/taskset-sdk.

package/dist/chunks/{chunk-T5AACNJI.js → chunk-P7HRF4FM.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{b as e}from"./chunk-ES2ISFJC.js";import{a as o}from"./chunk-FTQLQJQV.js";var i=e.enum(["local","sandbox"]),s=e.enum(["pinned","saved_for_later"]),l=e.object({id:e.string().min(1),workspaceKind:i,workspaceId:e.string().min(1),workspaceName:e.string().min(1),path:e.string().min(1),status:s,order:e.number().int().nonnegative().nullable().default(null),sourceSessionId:e.string().min(1).nullable().default(null),available:e.boolean().default(!0),createdAt:e.string().min(1),updatedAt:e.string().min(1)}),p=e.object({items:e.array(l)}),d=e.object({workspaceKind:i,workspaceId:e.string().trim().min(1),workspaceName:e.string().trim().min(1),path:e.string().trim().min(1),status:e.enum(["pinned","saved_for_later","none"]),order:e.number().int().nonnegative().nullable().optional(),sourceSessionId:e.string().trim().min(1).nullable().optional()}),u=e.enum(["pin","save_for_later","remove","list"]);function c(n){let r=n.trim().replaceAll("\\","/");if(r=r.replace(/^(?:workspace|sandbox):file:/,""),r=r.replace(/^\/workspace\/app\//,""),r=r.replace(/^\.\//,""),!r||r.startsWith("/"))throw new Error("File path must be relative to the workspace.");let a=r.split("/").filter(t=>t&&t!==".");if(a.some(t=>t===".."))throw new Error("File path cannot leave the workspace.");if(a.length===0)throw new Error("File path is required.");return a.join("/")}o(c,"normalizeSidebarFilePath");function b(n){return["sidebar-file",n.workspaceKind,encodeURIComponent(n.workspaceId),encodeURIComponent(c(n.path))].join(":")}o(b,"sidebarFileBookmarkId");export{l as a,d as b,c,b as d};
+import{A as e}from"./chunk-PF56N7ZY.js";import{a as o}from"./chunk-FTQLQJQV.js";var i=e.enum(["local","sandbox"]),s=e.enum(["pinned","saved_for_later"]),l=e.object({id:e.string().min(1),workspaceKind:i,workspaceId:e.string().min(1),workspaceName:e.string().min(1),path:e.string().min(1),status:s,order:e.number().int().nonnegative().nullable().default(null),sourceSessionId:e.string().min(1).nullable().default(null),available:e.boolean().default(!0),createdAt:e.string().min(1),updatedAt:e.string().min(1)}),p=e.object({items:e.array(l)}),d=e.object({workspaceKind:i,workspaceId:e.string().trim().min(1),workspaceName:e.string().trim().min(1),path:e.string().trim().min(1),status:e.enum(["pinned","saved_for_later","none"]),order:e.number().int().nonnegative().nullable().optional(),sourceSessionId:e.string().trim().min(1).nullable().optional()}),u=e.enum(["pin","save_for_later","remove","list"]);function c(n){let r=n.trim().replaceAll("\\","/");if(r=r.replace(/^(?:workspace|sandbox):file:/,""),r=r.replace(/^\/workspace\/app\//,""),r=r.replace(/^\.\//,""),!r||r.startsWith("/"))throw new Error("File path must be relative to the workspace.");let a=r.split("/").filter(t=>t&&t!==".");if(a.some(t=>t===".."))throw new Error("File path cannot leave the workspace.");if(a.length===0)throw new Error("File path is required.");return a.join("/")}o(c,"normalizeSidebarFilePath");function b(n){return["sidebar-file",n.workspaceKind,encodeURIComponent(n.workspaceId),encodeURIComponent(c(n.path))].join(":")}o(b,"sidebarFileBookmarkId");export{l as a,d as b,c,b as d};

package/dist/chunks/{chunk-7ED24PP6.js → chunk-PCNLC5H3.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as d,b as c,f as m,g as p}from"./chunk-TVQAPBLZ.js";import{b as n}from"./chunk-KGNCKIFP.js";import{b as e}from"./chunk-ES2ISFJC.js";var u=e.string().trim().min(1).max(2e3).refine(i=>{let a=i.replaceAll("\\","/");return!a.startsWith("/")&&!a.includes("\0")&&!a.split("/").some(t=>!t||t==="."||t==="..")},"path must be a safe relative path"),f=e.enum(["instruction","skill","skill_resource","agent","workflow","dependency_lock","program","asset"]),h=e.object({id:n,kind:f,path:u,parentId:n.nullable(),mediaType:e.string().trim().min(1).max(200),visibility:e.enum(["policy","verifier","host_private"]),portability:e.enum(["portable","local_only"])}).strict().superRefine((i,a)=>{i.kind==="skill_resource"!=(i.parentId!==null)&&a.addIssue({code:"custom",message:"only Skill resources require parentId",path:["parentId"]})}),S=e.object({schemaVersion:e.literal("openpond.harnessSourceManifest.v1"),name:e.string().trim().min(1).max(240),files:e.array(h).min(2).max(1e5),toolDeclarations:e.array(d).max(200),capabilityRequirements:e.array(c).max(200),lifecycle:m,graderInterface:p,runtimeProtocol:e.string().trim().min(1).max(500),metadata:e.record(e.string(),e.unknown()).default({})}).strict().superRefine((i,a)=>{let t=new Set,l=new Map,o=new Set;for(let[r,s]of i.files.entries())t.has(s.id)&&a.addIssue({code:"custom",message:"file ids must be unique",path:["files",r,"id"]}),o.has(s.path)&&a.addIssue({code:"custom",message:"file paths must be unique",path:["files",r,"path"]}),t.add(s.id),l.set(s.id,s.kind),o.add(s.path);for(let[r,s]of i.files.entries())s.parentId&&!t.has(s.parentId)?a.addIssue({code:"custom",message:"parentId must reference a declared Skill",path:["files",r,"parentId"]}):s.parentId&&l.get(s.parentId)!=="skill"&&a.addIssue({code:"custom",message:"parentId must reference a primary Skill file",path:["files",r,"parentId"]});i.files.filter(r=>r.kind==="dependency_lock").length!==1&&a.addIssue({code:"custom",message:"exactly one dependency lock is required",path:["files"]}),i.files.filter(r=>r.kind==="program").length!==1&&a.addIssue({code:"custom",message:"exactly one Harness program is required",path:["files"]})});export{S as a};
+import{a as d,b as c,f as m,g as p}from"./chunk-VHSX3JPX.js";import{b as n}from"./chunk-BIPQCJ5V.js";import{A as e}from"./chunk-PF56N7ZY.js";var u=e.string().trim().min(1).max(2e3).refine(i=>{let a=i.replaceAll("\\","/");return!a.startsWith("/")&&!a.includes("\0")&&!a.split("/").some(t=>!t||t==="."||t==="..")},"path must be a safe relative path"),f=e.enum(["instruction","skill","skill_resource","agent","workflow","dependency_lock","program","asset"]),h=e.object({id:n,kind:f,path:u,parentId:n.nullable(),mediaType:e.string().trim().min(1).max(200),visibility:e.enum(["policy","verifier","host_private"]),portability:e.enum(["portable","local_only"])}).strict().superRefine((i,a)=>{i.kind==="skill_resource"!=(i.parentId!==null)&&a.addIssue({code:"custom",message:"only Skill resources require parentId",path:["parentId"]})}),S=e.object({schemaVersion:e.literal("openpond.harnessSourceManifest.v1"),name:e.string().trim().min(1).max(240),files:e.array(h).min(2).max(1e5),toolDeclarations:e.array(d).max(200),capabilityRequirements:e.array(c).max(200),lifecycle:m,graderInterface:p,runtimeProtocol:e.string().trim().min(1).max(500),metadata:e.record(e.string(),e.unknown()).default({})}).strict().superRefine((i,a)=>{let t=new Set,l=new Map,o=new Set;for(let[r,s]of i.files.entries())t.has(s.id)&&a.addIssue({code:"custom",message:"file ids must be unique",path:["files",r,"id"]}),o.has(s.path)&&a.addIssue({code:"custom",message:"file paths must be unique",path:["files",r,"path"]}),t.add(s.id),l.set(s.id,s.kind),o.add(s.path);for(let[r,s]of i.files.entries())s.parentId&&!t.has(s.parentId)?a.addIssue({code:"custom",message:"parentId must reference a declared Skill",path:["files",r,"parentId"]}):s.parentId&&l.get(s.parentId)!=="skill"&&a.addIssue({code:"custom",message:"parentId must reference a primary Skill file",path:["files",r,"parentId"]});i.files.filter(r=>r.kind==="dependency_lock").length!==1&&a.addIssue({code:"custom",message:"exactly one dependency lock is required",path:["files"]}),i.files.filter(r=>r.kind==="program").length!==1&&a.addIssue({code:"custom",message:"exactly one Harness program is required",path:["files"]})});export{S as a};