openpond 0.2.49 → 0.2.51

package/dist/chunks/{chunk-NHJ26KXF.js → chunk-TBNBCOW7.js}

@@ -1,5 +1,5 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{f as fn,h as hn}from"./chunk-DLZP22WV.js";import{b as Bo,c as Jo,d as Rc,e as pn,f as Vt,g as mn}from"./chunk-VOIH2BQA.js";import{a as Tc,c as vc,g as Ec,h as Sc}from"./chunk-QJRXTIOC.js";import{D as an,E as fr,F as Jt,G as hr,H as ic,I as cc,J as dc,K as cn,L as lc,M as Uo,N as uc,O as xe,P as pc,Q as mc,R as fc,S as hc,T as gc,U as dn,V as yc,W as ln,aa as un,e as oi,ea as wc,f as si,fa as Me,ja as kc,m as ac,n as Fo}from"./chunk-7MXZJQS7.js";import{b as gn,m as xc,p as bc,s as Ic,t as yn,y as Cc,z as wn}from"./chunk-2ZJJFBIA.js";import{d as Ve,g as Xe}from"./chunk-5CMKGD43.js";import{G as oc,P as sc}from"./chunk-HA7EFTPF.js";import{b as nc,e as on,f as sn}from"./chunk-7QEE4PBI.js";import{Ac as Ut,Ad as Ki,Bc as en,Bd as _i,Cc as tn,Cd as nn,Da as Hi,Dc as J,Ec as Xi,Ed as $i,Fa as _r,Fc as Bt,Fd as Rt,Gb as Fi,Gc as zi,Ib as re,Ic as Zi,Jb as Ui,Jc as jo,Kc as Yi,Kd as tc,La as Ho,Lb as Bi,Ld as tt,Ma as Wi,N as Mi,Oa as ji,Od as Se,Rd as mr,Sd as rc,Ta as qi,Y as Li,od as rn,pd as Qi,qd as Gi,rd as qo,va as Di,xc as $,yc as Ji,zc as Vi}from"./chunk-XDMUD4FW.js";import{c as ai,ga as Wo,z as Ni}from"./chunk-IKWDL3RB.js";import{Ba as Ci,Ca as pr,Oa as Ai,Qa as Je,Ra as Oi,Sa as Pi,Ta as et,ja as Ti,ka as vi,la as Ei,m as Ee,ma as Si,oa as xi,pc as $r,wc as ec,xa as bi,ya as Ii}from"./chunk-QRWPEW3Z.js";import{qa as ii,ra as Kr,sa as ci,ta as di,ua as lr,va as Ft,wa as li,x as _a}from"./chunk-WV2APYWG.js";import{Db as Gr,Eb as ri,Jb as qt,Qa as $a,Ra as Lo,Ta as ei,U as Qr,Vc as Ue,Wc as Be,a as Za,b as G,d as Ya,ga as $e,ha as Qa,ja as Ga,ka as Mo,la as ue,sd as ni,vb as Y,wb as ti,ya as Ka,z as Yr}from"./chunk-PJM2G7SG.js";import{A as ja,B as Ro,G as Jr,H as Vr,I as To,J as Fa,K as Ua,L as vo,M as Xr,N as Eo,O as So,P as cr,T as xo,U as Ba,V as Ja,W as Va,X as Xa,Y as za,Z as bo,_ as Io,e as Ma,ga as dr,ha as Co,ia as Ao,ja as Oo,ka as Po,la as zr,ma as Zr,n as La,na as jt,o as Da,oa as No,pa as Z,w as Ha,x as Wa}from"./chunk-SDEDAHV6.js";import{E as wi,F as ki,G as Ri,e as ui,f as wt}from"./chunk-DALJMSJG.js";import{B as pi,H as Do,N as mi,y as kt}from"./chunk-TGIB24K4.js";import{b as gt,c as he,d as yt,f as Aa,h as Oa,i as Pa,k as Na,l as ve,m as y,o as qa}from"./chunk-NNF6KCHD.js";import{b as g}from"./chunk-ES2ISFJC.js";import{h as ur,n as fi,o as hi,p as gi,q as yi}from"./chunk-H7IWJN7L.js";import{a as i,c as ko,e as Cp}from"./chunk-FTQLQJQV.js";var rd=ko((pv,td)=>{"use strict";var Im=["||","&&",";;","|&","<(","<<<",">>",">&","<&","&",";","(",")","|","<",">"],ed=/[\n\r\u2028\u2029]/,Cm=/[\s#!"$&'():;<=>@\\^`|]/g;td.exports=i(function(t){return t.map(function(r){if(r==="")return"''";if(r&&typeof r=="object"){if("op"in r&&r.op==="glob"){if(typeof r.pattern!="string")throw new TypeError("glob token requires a string `pattern`");if(ed.test(r.pattern))throw new TypeError("glob `pattern` must not contain line terminators");return r.pattern.replace(Cm,"\\$&")}if("op"in r&&typeof r.op=="string"){if(Im.indexOf(r.op)<0)throw new TypeError("invalid `op` value: "+JSON.stringify(r.op));return r.op.replace(/[\s\S]/g,"\\$&")}if("comment"in r&&typeof r.comment=="string"){if(ed.test(r.comment))throw new TypeError("`comment` must not contain line terminators");return"#"+r.comment}throw new TypeError("unrecognized object token shape")}return/["\s\\]/.test(r)&&!/'/.test(r)?"'"+r.replace(/(['])/g,"\\$1")+"'":/["'\s]/.test(r)?'"'+r.replace(/(["\\$`!])/g,"\\$1")+'"':String(r).replace(/([A-Za-z]:)?([#!"$&'()*,:;<=>?@[\\\]^`{|}~])/g,"$1\\$2")}).join(" ")},"quote")});var dd=ko((fv,cd)=>{"use strict";var id="(?:"+["\\|\\|","\\&\\&",";;","\\|\\&","\\<\\(","\\<\\<\\<",">>",">\\&","<\\&","[&;()|<>]"].join("|")+")",nd=new RegExp("^"+id+"$"),od="|&;()<> \\t",Am="'([^']*?)'",Om='"((\\\\"|[^"])*?)"',Pm=/^#$/,sd="'",ad='"',In="$",Et="",Nm=4294967296;for($o=0;$o<4;$o++)Et+=(Nm*Math.random()).toString(16);var $o,Mm=new RegExp("^"+Et);function Lm(e,t){for(var r=t.lastIndex,n=[],o;o=t.exec(e);)n[n.length]=o,t.lastIndex===o.index&&(t.lastIndex+=1);return t.lastIndex=r,n}i(Lm,"matchAll");function Dm(e,t,r){var n=typeof e=="function"?e(r):e[r];return typeof n>"u"&&r!=""?n="":typeof n>"u"&&(n="$"),typeof n=="object"?t+Et+JSON.stringify(n)+Et:t+n}i(Dm,"getVar");function Hm(e,t,r){r||(r={});var n=r.escape||"\\",o=r.splitUnquoted===!0?` 	
+import{f as fn,h as hn}from"./chunk-DLZP22WV.js";import{b as Bo,c as Jo,d as Rc,e as pn,f as Vt,g as mn}from"./chunk-OE2VF6KU.js";import{a as Tc,c as vc,g as Ec,h as Sc}from"./chunk-QJRXTIOC.js";import{D as an,E as fr,F as Jt,G as hr,H as ic,I as cc,J as dc,K as cn,L as lc,M as Uo,N as uc,O as xe,P as pc,Q as mc,R as fc,S as hc,T as gc,U as dn,V as yc,W as ln,aa as un,e as oi,ea as wc,f as si,fa as Me,ja as kc,m as ac,n as Fo}from"./chunk-7MXZJQS7.js";import{b as gn,m as xc,p as bc,s as Ic,t as yn,y as Cc,z as wn}from"./chunk-2ZJJFBIA.js";import{d as Ve,g as Xe}from"./chunk-5CMKGD43.js";import{G as oc,P as sc}from"./chunk-HA7EFTPF.js";import{b as nc,e as on,f as sn}from"./chunk-7QEE4PBI.js";import{Ac as Ut,Ad as Ki,Bc as en,Bd as _i,Cc as tn,Cd as nn,Da as Hi,Dc as J,Ec as Xi,Ed as $i,Fa as _r,Fc as Bt,Fd as Rt,Gb as Fi,Gc as zi,Ib as re,Ic as Zi,Jb as Ui,Jc as jo,Kc as Yi,Kd as tc,La as Ho,Lb as Bi,Ld as tt,Ma as Wi,N as Mi,Oa as ji,Od as Se,Rd as mr,Sd as rc,Ta as qi,Y as Li,od as rn,pd as Qi,qd as Gi,rd as qo,va as Di,xc as $,yc as Ji,zc as Vi}from"./chunk-XDMUD4FW.js";import{c as ai,ga as Wo,z as Ni}from"./chunk-IKWDL3RB.js";import{Ba as Ci,Ca as pr,Oa as Ai,Qa as Je,Ra as Oi,Sa as Pi,Ta as et,ja as Ti,ka as vi,la as Ei,m as Ee,ma as Si,oa as xi,pc as $r,wc as ec,xa as bi,ya as Ii}from"./chunk-QRWPEW3Z.js";import{qa as ii,ra as Kr,sa as ci,ta as di,ua as lr,va as Ft,wa as li,x as _a}from"./chunk-WV2APYWG.js";import{Db as Gr,Eb as ri,Jb as qt,Qa as $a,Ra as Lo,Ta as ei,U as Qr,Vc as Ue,Wc as Be,a as Za,b as G,d as Ya,ga as $e,ha as Qa,ja as Ga,ka as Mo,la as ue,sd as ni,vb as Y,wb as ti,ya as Ka,z as Yr}from"./chunk-PJM2G7SG.js";import{A as ja,B as Ro,G as Jr,H as Vr,I as To,J as Fa,K as Ua,L as vo,M as Xr,N as Eo,O as So,P as cr,T as xo,U as Ba,V as Ja,W as Va,X as Xa,Y as za,Z as bo,_ as Io,e as Ma,ga as dr,ha as Co,ia as Ao,ja as Oo,ka as Po,la as zr,ma as Zr,n as La,na as jt,o as Da,oa as No,pa as Z,w as Ha,x as Wa}from"./chunk-SDEDAHV6.js";import{E as wi,F as ki,G as Ri,e as ui,f as wt}from"./chunk-DALJMSJG.js";import{B as pi,H as Do,N as mi,y as kt}from"./chunk-TGIB24K4.js";import{b as gt,c as he,d as yt,f as Aa,h as Oa,i as Pa,k as Na,l as ve,m as y,o as qa}from"./chunk-NNF6KCHD.js";import{b as g}from"./chunk-ES2ISFJC.js";import{h as ur,n as fi,o as hi,p as gi,q as yi}from"./chunk-H7IWJN7L.js";import{a as i,c as ko,e as Cp}from"./chunk-FTQLQJQV.js";var rd=ko((pv,td)=>{"use strict";var Im=["||","&&",";;","|&","<(","<<<",">>",">&","<&","&",";","(",")","|","<",">"],ed=/[\n\r\u2028\u2029]/,Cm=/[\s#!"$&'():;<=>@\\^`|]/g;td.exports=i(function(t){return t.map(function(r){if(r==="")return"''";if(r&&typeof r=="object"){if("op"in r&&r.op==="glob"){if(typeof r.pattern!="string")throw new TypeError("glob token requires a string `pattern`");if(ed.test(r.pattern))throw new TypeError("glob `pattern` must not contain line terminators");return r.pattern.replace(Cm,"\\$&")}if("op"in r&&typeof r.op=="string"){if(Im.indexOf(r.op)<0)throw new TypeError("invalid `op` value: "+JSON.stringify(r.op));return r.op.replace(/[\s\S]/g,"\\$&")}if("comment"in r&&typeof r.comment=="string"){if(ed.test(r.comment))throw new TypeError("`comment` must not contain line terminators");return"#"+r.comment}throw new TypeError("unrecognized object token shape")}return/["\s\\]/.test(r)&&!/'/.test(r)?"'"+r.replace(/(['])/g,"\\$1")+"'":/["'\s]/.test(r)?'"'+r.replace(/(["\\$`!])/g,"\\$1")+'"':String(r).replace(/([A-Za-z]:)?([#!"$&'()*,:;<=>?@[\\\]^`{|}~])/g,"$1\\$2")}).join(" ")},"quote")});var dd=ko((fv,cd)=>{"use strict";var id="(?:"+["\\|\\|","\\&\\&",";;","\\|\\&","\\<\\(","\\<\\<\\<",">>",">\\&","<\\&","[&;()|<>]"].join("|")+")",nd=new RegExp("^"+id+"$"),od="|&;()<> \\t",Am="'([^']*?)'",Om='"((\\\\"|[^"])*?)"',Pm=/^#$/,sd="'",ad='"',In="$",Et="",Nm=4294967296;for($o=0;$o<4;$o++)Et+=(Nm*Math.random()).toString(16);var $o,Mm=new RegExp("^"+Et);function Lm(e,t){for(var r=t.lastIndex,n=[],o;o=t.exec(e);)n[n.length]=o,t.lastIndex===o.index&&(t.lastIndex+=1);return t.lastIndex=r,n}i(Lm,"matchAll");function Dm(e,t,r){var n=typeof e=="function"?e(r):e[r];return typeof n>"u"&&r!=""?n="":typeof n>"u"&&(n="$"),typeof n=="object"?t+Et+JSON.stringify(n)+Et:t+n}i(Dm,"getVar");function Hm(e,t,r){r||(r={});var n=r.escape||"\\",o=r.splitUnquoted===!0?` 	
 `:typeof r.splitUnquoted=="string"?r.splitUnquoted:"",s="(\\"+n+`['"`+od+`]|[^\\s'"`+od+"])+",a=new RegExp(["("+id+")","("+s+"|"+Om+"|"+Am+")+"].join("|"),"g"),c=Lm(e,a);if(c.length===0)return[];t||(t={});var d=!1;return c.map(function(l){var p=l[0];if(!p||d)return;if(nd.test(p))return{op:p};var u=!1,m=!1,f="",h=[],R=!1,w=null,k=!1,x;function C(){x+=1;var E,b,D=p.charAt(x);if(D==="{"){if(x+=1,p.charAt(x)==="}")throw new Error("Bad substitution: "+p.slice(x-2,x+1));var F=1;for(E=x;F>0&&E<p.length;)p.charAt(E)==="{"&&p.charAt(E-1)==="$"?F+=1:p.charAt(E)==="}"&&(F-=1),E+=1;if(F!==0)throw new Error("Bad substitution: "+p.slice(x));E-=1,b=p.slice(x,E),x=E}else if(/[*@#?$!_-]/.test(D))b=D,x+=1;else{var V=p.slice(x);E=V.match(/[^\w\d_]/),E?(b=V.slice(0,E.index),x+=E.index-1):(b=V,x=p.length)}return Dm(t,"",b)}i(C,"parseEnvVar");function A(){if(w!==null){if(w===0)f!==""&&(h[h.length]=f,f="");else{h[h.length]=f,f="";for(var E=1;E<w;E+=1)h[h.length]=""}w=null}}for(i(A,"flushRun"),x=0;x<p.length;x++){var T=p.charAt(x);if(o&&T!==In&&A(),k=k||!u&&(T==="*"||T==="?"),m)f+=T,m=!1;else if(u)T===u?u=!1:u==sd?f+=T:T===n?(x+=1,T=p.charAt(x),T===ad||T===n||T===In?f+=T:f+=n+T):T===In?f+=C():f+=T;else if(T===ad||T===sd)u=T,R=!0;else{if(nd.test(T))return{op:p};if(Pm.test(T)){d=!0;var P={comment:e.slice(l.index+x+1)};return f.length?[f,P]:[P]}else if(T===n)m=!0;else if(T===In){var v=C();if(!o)f+=v;else for(var I=0;I<v.length;I+=1){var S=v.charAt(I);o.indexOf(S)<0?(A(),f+=S):w===null?w=S===" "||S==="	"||S===`
 `?0:1:S!==" "&&S!=="	"&&S!==`
 `&&(w+=1)}}else f+=T}}if(k)return{op:"glob",pattern:f};if(o){if(w!==null&&w>0){h[h.length]=f,f="";for(var N=1;N<w;N+=1)h[h.length]=""}return(f!==""||R&&h.length===0)&&(h[h.length]=f),h}return f}).reduce(function(l,p){return typeof p>"u"||[].concat(p).forEach(function(u){l[l.length]=u}),l},[])}i(Hm,"parseInternal");cd.exports=i(function(t,r,n){var o=Hm(t,r,n);return typeof r!="function"?o:o.reduce(function(s,a){if(typeof a=="object")return s[s.length]=a,s;var c=a.split(RegExp("("+Et+".*?"+Et+")","g"));return c.length===1?(s[s.length]=c[0],s):(c.filter(Boolean).forEach(function(d){s[s.length]=Mm.test(d)?JSON.parse(d.split(Et)[1]):d}),s)},[])},"parse")});var ld=ko(es=>{"use strict";es.quote=rd();es.parse=dd()});import{promises as Lc}from"node:fs";import rt from"node:path";var Ap="openpond.workEvidenceReceipt.v1",Op="openpond.workProcessTrace.v1",Pp="openpond.workFeedbackReceipt.v1",hk="openpond.workEvidenceEligibility.v1",Np=g.string().regex(/^urn:openpond:work:[a-f0-9]{64}$/,"Work source references must be opaque SHA-256 URNs."),Mp=g.string().regex(/^urn:openpond:workspace:[a-f0-9]{64}$/,"Workspace references must be opaque SHA-256 URNs."),Lp=g.string().regex(/^urn:openpond:artifact:[a-f0-9]{64}$/,"Evidence artifacts must use content-addressed SHA-256 URNs."),ie=Pa.extend({id:Lp}).strict().superRefine((e,t)=>{e.id!==Pc(e.contentHash)&&t.addIssue({code:"custom",message:"Evidence artifact id must contain its content hash.",path:["id"]})}),Dp=g.enum(["tool","validation","state_transition","approval","question","artifact","cleanup"]),Hp=g.enum(["agent","environment"]),Wp=g.enum(["turn_started","turn_completed","turn_failed","turn_cancelled","turn_timed_out","tool_invoked","tool_completed","tool_failed","validation_completed","validation_failed","approval_requested","approval_resolved","question_asked","question_answered","question_dismissed","artifact_created","workspace_changed","environment_created","environment_reset","environment_destroyed","cleanup_completed","cleanup_failed"]),jp=g.enum(["started","completed","failed","cancelled"]),qp=g.enum(["filesystem","source_control","command","browser","connected_app","sandbox","agent","other"]),Fp=g.enum(["structural","visual","test","user_review","other"]),Up=g.enum(["running","completed","failed","cancelled","timeout"]),Bp=g.enum(["requested","approved","denied","answered","dismissed"]),Jp=g.enum(["policy","validation","environment","infrastructure","timeout","cancelled","unknown"]),Vp=g.enum(["missing_start","missing_terminal","source_truncated","unsupported_events_dropped","uncorrelated_environment_step","artifact_unavailable"]),Xp=g.object({sequence:g.number().int().nonnegative().max(1e6),timestamp:yt,layer:Hp,kind:Dp,action:Wp,status:jp,inputHash:he.nullable(),outputHash:he.nullable(),receiptHash:he.nullable(),parentReceiptHash:he.nullable(),artifactRefs:g.array(ie).max(1e3),attributes:g.object({toolCategory:qp.nullable(),validationKind:Fp.nullable(),transitionState:Up.nullable(),interventionOutcome:Bp.nullable(),artifactCount:g.number().int().nonnegative().max(1e5),exitCode:g.number().int().min(-1).max(255).nullable(),durationMs:g.number().int().nonnegative().max(31536e6).nullable(),cpuTimeMs:g.number().int().nonnegative().max(31536e6).nullable(),memoryPeakBytes:g.number().int().nonnegative().max(0x4000000000000).nullable(),errorClass:Jp.nullable()}).strict()}).strict().superRefine((e,t)=>{e.layer==="agent"&&e.parentReceiptHash!==null&&t.addIssue({code:"custom",message:"Agent-layer steps cannot have a parent environment correlation.",path:["parentReceiptHash"]}),e.layer==="environment"&&e.parentReceiptHash===null&&t.addIssue({code:"custom",message:"Environment steps must correlate to an Agent receipt.",path:["parentReceiptHash"]})}),Ac=g.object({schemaVersion:g.literal(Op),sourceRevisionHash:he,sanitationPolicyVersion:gt,incomplete:g.boolean(),incompleteReasons:g.array(Vp).max(10),droppedEventCount:g.number().int().nonnegative().max(1e6),steps:g.array(Xp).max(1e5)}).strict().superRefine((e,t)=>{e.incomplete!==e.incompleteReasons.length>0&&t.addIssue({code:"custom",message:"incomplete must match the presence of incompleteReasons.",path:["incomplete"]});for(let r=0;r<e.steps.length;r+=1)e.steps[r].sequence!==r&&t.addIssue({code:"custom",message:"Process trace sequences must be contiguous and zero-based.",path:["steps",r,"sequence"]})}),Xo=Ac.safeExtend({contentHash:he}).strict(),zp=g.enum(["policy_failure","validation_failure","model_failure","environment_failure","infrastructure_failure","timeout","cancelled","unknown"]),Zp=g.object({status:g.enum(["completed","failed","cancelled","timeout"]),failureClass:zp.nullable()}).strict().superRefine((e,t)=>{e.status==="completed"&&e.failureClass!==null&&t.addIssue({code:"custom",message:"Completed Work cannot carry a failure class.",path:["failureClass"]}),e.status!=="completed"&&e.failureClass===null&&t.addIssue({code:"custom",message:"Non-completed Work requires a failure class.",path:["failureClass"]}),e.status==="cancelled"&&e.failureClass!=="cancelled"&&t.addIssue({code:"custom",message:"Cancelled Work must use the cancelled failure class.",path:["failureClass"]}),e.status==="timeout"&&e.failureClass!=="timeout"&&t.addIssue({code:"custom",message:"Timed-out Work must use the timeout failure class.",path:["failureClass"]})}),zo=g.object({schemaVersion:g.literal(Ap),id:g.string().regex(/^work-evidence-[a-f0-9]{24}$/),source:g.object({surface:g.enum(["desktop","hosted"]),experience:g.enum(["work","development"]),opaqueRef:Np,revisionHash:he}).strict(),agentSnapshot:Aa.nullable(),model:qa,runtime:g.object({adapterId:gt,adapterVersion:gt,capabilityRef:ie.nullable()}).strict(),inputHash:he,terminal:Zp,trace:g.object({sanitizedRef:ie,traceHash:he,sanitationPolicyVersion:gt,incomplete:g.boolean()}).strict(),outputRefs:g.array(ie).max(1e4),artifactRefs:g.array(ie).max(1e5),validationEvidenceRefs:g.array(ie).max(1e4),interventions:g.object({approvals:g.number().int().nonnegative().max(1e6),questions:g.number().int().nonnegative().max(1e6),steeringEvents:g.number().int().nonnegative().max(1e6),otherUserInterventions:g.number().int().nonnegative().max(1e6)}).strict(),timing:g.object({startedAt:yt,completedAt:yt,latencyMs:g.number().int().nonnegative().max(31536e6)}).strict(),usage:g.object({promptTokens:g.number().int().nonnegative().nullable(),completionTokens:g.number().int().nonnegative().nullable(),totalTokens:g.number().int().nonnegative().nullable()}).strict(),costUsd:g.number().nonnegative().max(1e6).nullable(),provenance:g.object({consentReceiptRef:ie,consentScope:g.literal("work_process_and_artifacts"),consentGrantedAt:yt,policyVersion:gt,projectorVersion:gt,disclosure:g.literal("portable_sanitized"),ownershipScope:g.enum(["personal","workspace"]),workspaceRef:Mp.nullable(),participantPolicy:g.enum(["creator_only","all_participants"]),retention:g.object({policy:g.literal("source_bound"),deleteWithSource:g.literal(!0),expiresAt:yt.nullable()}).strict()}).strict()}).strict().superRefine((e,t)=>{e.trace.traceHash!==e.trace.sanitizedRef.contentHash&&t.addIssue({code:"custom",message:"Sanitized trace reference must bind traceHash.",path:["trace","traceHash"]});let r=Date.parse(e.timing.completedAt)-Date.parse(e.timing.startedAt);(r<0||r!==e.timing.latencyMs)&&t.addIssue({code:"custom",message:"latencyMs must equal completedAt minus startedAt.",path:["timing","latencyMs"]}),e.provenance.ownershipScope==="workspace"&&e.provenance.workspaceRef===null&&t.addIssue({code:"custom",message:"Workspace-owned evidence requires workspaceRef.",path:["provenance","workspaceRef"]}),e.provenance.ownershipScope==="personal"&&e.provenance.workspaceRef!==null&&t.addIssue({code:"custom",message:"Personal evidence cannot carry workspaceRef.",path:["provenance","workspaceRef"]}),Vo(e.outputRefs,["outputRefs"],t),Vo(e.artifactRefs,["artifactRefs"],t),Vo(e.validationEvidenceRefs,["validationEvidenceRefs"],t)}),kn=zo.safeExtend({contentHash:he}).strict(),Yp=g.enum(["accepted","needs_correction","not_useful"]),Qp=g.enum(["correct","complete","high_quality","incorrect","incomplete","wrong_format","unsafe","stale","irrelevant","other"]),Oc=g.object({schemaVersion:g.literal(Pp),id:g.string().regex(/^work-feedback-[a-f0-9]{24}$/),evidenceReceiptRef:ie,outputRevisionRef:ie.nullable(),verdict:Yp,reasonCodes:g.array(Qp).max(16),correctionRef:ie.nullable(),correctedOutputRevisionRef:ie.nullable(),actor:g.literal("user"),createdAt:yt}).strict().superRefine((e,t)=>{new Set(e.reasonCodes).size!==e.reasonCodes.length&&t.addIssue({code:"custom",message:"Feedback reason codes must be unique.",path:["reasonCodes"]}),e.verdict!=="needs_correction"&&(e.correctionRef||e.correctedOutputRevisionRef)&&t.addIssue({code:"custom",message:"Only needs_correction feedback may bind correction artifacts.",path:["verdict"]})}),Zo=Oc.safeExtend({contentHash:he}).strict();function Pc(e){return`urn:openpond:artifact:${e}`}i(Pc,"evidenceArtifactId");function gk(e){return`urn:openpond:work:${y(e)}`}i(gk,"workSourceOpaqueRef");function yk(e){return`urn:openpond:workspace:${y(e)}`}i(yk,"workWorkspaceOpaqueRef");function Gp(e){return ie.parse({id:Pc(e.contentHash),contentHash:e.contentHash,mediaType:e.mediaType??null,sizeBytes:e.sizeBytes??null})}i(Gp,"evidenceArtifactRef");function wk(e){let t=Ac.parse(e);return Xo.parse({...t,contentHash:y(t)})}i(wk,"createWorkProcessTrace");function kk(e){let t=zo.parse(e);return kn.parse({...t,contentHash:y(t)})}i(kk,"createWorkEvidenceReceipt");function Rk(e,t){let r=Oc.parse(e),n=Zo.parse({...r,contentHash:y(r)});return t&&Kp(n,t),n}i(Rk,"createWorkFeedbackReceipt");function Tk(e){return _p(e,kn,zo)}i(Tk,"verifyWorkEvidenceReceipt");function vk(e){return Gp({contentHash:e.contentHash,mediaType:"application/vnd.openpond.work-evidence+json",sizeBytes:null})}i(vk,"workEvidenceReceiptRef");function Kp(e,t){if(e.evidenceReceiptRef.contentHash!==t.contentHash)throw new Error("Feedback targets a different Work evidence receipt.");if(e.outputRevisionRef&&!t.outputRefs.some(r=>r.contentHash===e.outputRevisionRef.contentHash))throw new Error("Feedback targets an output revision not bound by the Work evidence receipt.")}i(Kp,"assertFeedbackTargetsEvidence");function _p(e,t,r){let n=t.safeParse(e);if(!n.success)return!1;let{contentHash:o,...s}=n.data;return y(r.parse(s))===o}i(_p,"verifyHashedValue");function Vo(e,t,r){let n=e.map(o=>o.contentHash);new Set(n).size!==n.length&&r.addIssue({code:"custom",message:"Artifact content hashes must be unique.",path:t})}i(Vo,"addDuplicateHashIssues");import{randomUUID as $p}from"node:crypto";import{promises as Rn}from"node:fs";import Nc from"node:path";function Xt(e){return{binding:i(()=>{let t=gi(e.home,"refiner_bindings","active");return t?Xr.parse(t.value):null},"binding"),releases:i(async()=>{let t=(await Rn.readdir(e.releases)).filter(r=>r.endsWith(".json"));return Promise.all(t.map(async r=>vo.parse(JSON.parse(await Rn.readFile(Nc.join(e.releases,r),"utf8")))))},"releases"),transitions:i(()=>Object.values(yi(e.home,"refiner_transitions")).map(t=>Eo.parse(t.value)),"transitions"),release:i(async t=>{let r=vo.parse(JSON.parse(await Rn.readFile(Nc.join(e.releases,`${t.contentHash}.json`),"utf8")));if(r.id!==t.id||r.contentHash!==t.contentHash)throw new Error("Refiner release reference does not match immutable release content.");return r},"release"),persistRelease:i(async t=>Mc(e.releases+`/${t.contentHash}.json`,Na(t),!0),"persistRelease"),writeSource:i(async t=>Mc(e.source,t),"writeSource"),transition:i(async(t,r)=>{hi(e.home,n=>{n.exec("BEGIN IMMEDIATE");try{let o=n.prepare("SELECT payload FROM refiner_bindings WHERE id = 'active'").get(),s=o?Xr.parse(JSON.parse(o.payload)):null,a=new Date().toISOString(),c={id:t.id,contentHash:t.contentHash},d={schemaVersion:"openpond.refinerTransitionReceipt.v1",id:`refiner-transition-${$p()}`,operation:r.operation,bindingChanged:r.bindingChanged,previousRelease:s?.release??null,nextRelease:c,actor:r.actor,reason:r.reason,authoringSkillHash:r.authoringSkillHash,validation:{valid:!0,messages:[]},createdAt:a},l=Eo.parse({...d,contentHash:y(d)});if(n.prepare("INSERT INTO refiner_transitions VALUES (?, ?, 1)").run(l.contentHash,JSON.stringify(l)),r.bindingChanged){let p=Xr.parse({schemaVersion:"openpond.refinerBinding.v1",channel:"active",revision:(s?.revision??-1)+1,release:c,updatedAt:a});n.prepare("INSERT INTO refiner_bindings VALUES ('active', ?, 1) ON CONFLICT(id) DO UPDATE SET payload=excluded.payload, revision=revision+1").run(JSON.stringify(p))}n.exec("COMMIT")}catch(o){throw n.exec("ROLLBACK"),o}})},"transition")}}i(Xt,"createLocalRefinerProfileRepository");async function Mc(e,t,r=!1){await wt(e,async()=>{if(r){let n=await Rn.readFile(e,"utf8").catch(()=>null);if(n!==null){if(n!==t)throw new Error("Immutable Refiner release bytes changed.");return}}await ui(e,t)})}i(Mc,"atomicWrite");var Dc="OpenPond Refiner Core: evidence admission, privacy, ownership, validation, and immutable activation boundaries.";function gr(e){let t=rt.join(e,"library","refiners");return{home:e,root:t,source:rt.join(t,"source","openpond.review.json"),releases:rt.join(t,"releases")}}i(gr,"refinerProfilePaths");function qk(e){let t=i(()=>ze(e),"inspect"),r=i(s=>em(e,s),"update"),n=i(s=>rm(e,s),"activate"),o=i(s=>om(e,s),"rollback");return{refinerHistoryPayload:t,updateRefinerProfilePayload:r,activateRefinerReleasePayload:n,rollbackRefinerReleasePayload:o,inspectRefiner:t,updateRefiner:r,activateRefiner:n,rollbackRefiner:o}}i(qk,"createRefinerProfileRoutePayloads");async function ze(e){let t=gr(e),r=Xt(t),n=r.binding();if(!n)throw new Error("Active Refiner binding is unavailable.");let o=(await r.releases()).sort((c,d)=>d.createdAt.localeCompare(c.createdAt)),s=o.find(c=>c.id===n.release.id&&c.contentHash===n.release.contentHash);if(!s)throw new Error("Active Refiner release is unavailable.");let a=r.transitions().sort((c,d)=>d.createdAt.localeCompare(c.createdAt));return Qi.parse({rootPath:t.root,sourcePath:t.source,binding:n,currentRelease:s,releases:o,transitions:a})}i(ze,"inspectRefinerProfile");async function Hc(e){return(await ze(e)).currentRelease}i(Hc,"loadActiveRefinerRelease");async function em(e,t){return wt(rt.join(ur(e).runtime,"refiner-edit"),()=>tm(e,t))}i(em,"updateRefinerProfile");async function tm(e,t){let r=Gi.parse(t),n=gr(e),o=Xt(n);await Yo(n);let s=Fa.parse(r.profile),a=So({profile:s,coreVersion:xo,corePrompt:Dc}),d=(await o.releases()).find(p=>p.composedPromptHash===a.composedPromptHash&&p.profileHash===a.profileHash)??a;await o.persistRelease(d),await o.writeSource(cr(s));let l=o.binding();if(!l)throw new Error("Active Refiner binding is unavailable.");return r.activate&&l.release.contentHash===d.contentHash||!r.activate&&o.transitions().some(m=>!m.bindingChanged&&m.nextRelease.contentHash===d.contentHash&&m.actor===r.actor&&m.reason===r.reason&&m.authoringSkillHash===r.authoringSkillHash)||await o.transition(d,{operation:"update",bindingChanged:r.activate,actor:r.actor,reason:r.reason,authoringSkillHash:r.authoringSkillHash}),ze(e)}i(tm,"updateRefinerProfileUnlocked");async function rm(e,t){return wt(rt.join(ur(e).runtime,"refiner-edit"),()=>nm(e,t))}i(rm,"activateRefinerRelease");async function nm(e,t){let r=qo.parse(t),n=gr(e),o=Xt(n);await Yo(n);let s=await o.release(r.release),a=o.binding();if(!a)throw new Error("Active Refiner binding is unavailable.");return a.release.contentHash===s.contentHash||(await o.transition(s,{operation:"activate",bindingChanged:!0,actor:r.actor,reason:r.reason,authoringSkillHash:null}),await o.writeSource(cr(s.profile))),ze(e)}i(nm,"activateRefinerReleaseUnlocked");async function om(e,t){return wt(rt.join(ur(e).runtime,"refiner-edit"),()=>sm(e,t))}i(om,"rollbackRefinerRelease");async function sm(e,t){let r=qo.parse(t),n=gr(e),o=Xt(n);await Yo(n);let s=await o.release(r.release),a=o.binding();if(!a)throw new Error("Active Refiner binding is unavailable.");return a.release.contentHash===s.contentHash||(await o.transition(s,{operation:"rollback",bindingChanged:!0,actor:r.actor,reason:r.reason,authoringSkillHash:null}),await o.writeSource(cr(s.profile))),ze(e)}i(sm,"rollbackRefinerReleaseUnlocked");async function am(e){await wt(rt.join(ur(e).runtime,"refiner-initialize"),()=>im(gr(e)))}i(am,"initializeRefinerProfile");async function Yo(e){await am(e.home)}i(Yo,"ensureInitialized");async function im(e){let t=Xt(e);if(await Promise.all([Lc.mkdir(rt.dirname(e.source),{recursive:!0,mode:448}),Lc.mkdir(e.releases,{recursive:!0,mode:448})]),t.binding())return;let r=So({profile:Ua,coreVersion:xo,corePrompt:Dc});await t.persistRelease(r),await t.writeSource(cr(r.profile)),await t.transition(r,{operation:"initialize",bindingChanged:!0,actor:"openpond",reason:"Initialize the default Refiner Review Profile.",authoringSkillHash:null})}i(im,"initialize");import{randomUUID as fm}from"node:crypto";var cm=["prompt_tokens","promptTokens","input_tokens","inputTokens"],dm=["completion_tokens","completionTokens","output_tokens","outputTokens"],lm=["total_tokens","totalTokens"];function Ze(e){let t=e&&typeof e=="object"&&!Array.isArray(e)?e:{},r=Qo(t,cm),n=um(t,r),o=n.promptTokens,s=Qo(t,dm),c=Qo(t,lm)??(o!==null&&s!==null?o+s:null);return{promptTokens:o,cachedPromptTokens:n.cachedPromptTokens,uncachedPromptTokens:n.uncachedPromptTokens,cacheWritePromptTokens:n.cacheWritePromptTokens,cacheTelemetrySource:n.cacheTelemetrySource,completionTokens:s,totalTokens:c}}i(Ze,"normalizeModelUsageTokens");function um(e,t){let r=Tn(e,["cached_input_tokens","cachedInputTokens","cached_prompt_tokens","cachedPromptTokens","prompt_cache_hit_tokens","promptCacheHitTokens"]),n=Tn(e,["prompt_cache_miss_tokens","promptCacheMissTokens"]),o=Tn(e,["cache_read_input_tokens","cacheReadInputTokens"]),s=mm(e,[["prompt_tokens_details","cached_tokens"],["promptTokensDetails","cachedTokens"],["input_tokens_details","cached_tokens"],["inputTokensDetails","cachedTokens"]]),a=Tn(e,["cache_creation_input_tokens","cacheCreationInputTokens","cache_write_input_tokens","cacheWriteInputTokens"]),c=r.value!==null?r:o.value!==null?o:s,d=c.value!==null||n.value!==null||a.value!==null,l=d?pm(e):null;if(!d)return{promptTokens:t,cachedPromptTokens:null,uncachedPromptTokens:null,cacheWritePromptTokens:null,cacheTelemetrySource:null};let p=c.value,u=a.present?a.value:null,m=c===o&&o.value!==null||a.value!==null,f=t??(c.value!==null&&n.value!==null?c.value+n.value:null);if(f===null)return{promptTokens:null,cachedPromptTokens:p,uncachedPromptTokens:n.value,cacheWritePromptTokens:u,cacheTelemetrySource:l};if(m){let w=p??0,k=u??0,x=t??n.value??0;return{promptTokens:x+w+k,cachedPromptTokens:p,uncachedPromptTokens:x+k,cacheWritePromptTokens:u,cacheTelemetrySource:l}}let h=p===null?null:Math.min(f,p),R=n.value===null?h===null?null:f-h:Math.min(f-(h??0),n.value);return{promptTokens:f,cachedPromptTokens:h,uncachedPromptTokens:R,cacheWritePromptTokens:u,cacheTelemetrySource:l}}i(um,"normalizeCacheUsage");function pm(e){return(e.cache_telemetry_source??e.cacheTelemetrySource)==="provider_response_headers"?"provider_response_headers":"provider_usage_body"}i(pm,"cacheTelemetrySourceFromUsage");function Tn(e,t){for(let r of t)if(Object.prototype.hasOwnProperty.call(e,r))return{present:!0,value:Go(e[r])};return{present:!1,value:null}}i(Tn,"firstPresentTokenCount");function mm(e,t){for(let[r,n]of t){let o=e[r];if(!o||typeof o!="object"||Array.isArray(o))continue;let s=o;if(Object.prototype.hasOwnProperty.call(s,n))return{present:!0,value:Go(s[n])}}return{present:!1,value:null}}i(mm,"firstPresentNestedTokenCount");function Qo(e,t){for(let r of t){let n=Go(e[r]);if(n!==null)return n}return null}i(Qo,"firstTokenCount");function Go(e){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return Math.trunc(e);if(typeof e!="string")return null;let t=e.trim();if(!/^\d+$/.test(t))return null;let r=Number(t);return Number.isSafeInteger(r)?r:null}i(Go,"tokenCount");function En(e){let t=vn(vn(vn(e.metadata).billing).pricing);return{version:zt(t.version,"pricing version"),source:zt(t.source,"pricing source"),effectiveAt:zt(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Zt(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Zt(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Zt(t.outputUsdPerMillionTokens,"output token rate")}}i(En,"hostedTokenPricingFromCatalog");function Xk(e){let t=vn(e);try{return{version:zt(t.version,"pricing version"),source:zt(t.source,"pricing source"),effectiveAt:zt(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Zt(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Zt(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Zt(t.outputUsdPerMillionTokens,"output token rate")}}catch{return null}}i(Xk,"hostedTokenPricingFromValue");function nt(e,t){let r=Ze(e);if(r.promptTokens===null&&r.completionTokens===null)return null;let n=r.cachedPromptTokens??0;return((r.uncachedPromptTokens??r.promptTokens??0)*t.inputUsdPerMillionTokens+n*t.cachedInputUsdPerMillionTokens+(r.completionTokens??0)*t.outputUsdPerMillionTokens)/1e6}i(nt,"hostedUsageCostUsd");function vn(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(vn,"record");function zt(e,t){if(typeof e=="string"&&e.trim())return e.trim();throw new Error(`Hosted model catalog is missing ${t}.`)}i(zt,"requiredString");function Zt(e,t){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return e;throw new Error(`Hosted model catalog is missing ${t}.`)}i(Zt,"requiredRate");function qc(e={}){let t=e.catalog??fn,r=e.stream??hn;return{async prepare(n){if(n.providerId!=="openpond")throw new Error("The selected judge provider is not connected to the hosted judge transport.");if(n.revision!==null)throw new Error("This judge transport cannot enforce a model revision; select a provider with revision support.");let o=await t();if(o.error)throw new Error(`Judge model catalog unavailable: ${o.error}`);let s=o.models.find(u=>u.id===n.modelId);if(!s)throw new Error("The selected judge model is unavailable.");let a=Wc(s.raw),c=En(a),d=jc(a.context_window,"context window"),l=Math.min(4096,jc(a.output_limit,"output limit")),p=(d*Math.max(c.inputUsdPerMillionTokens,c.cachedInputUsdPerMillionTokens)+l*c.outputUsdPerMillionTokens)/1e6;if(!(p>0))throw new Error("The judge model must have a positive known charge ceiling.");return{maximumChargeUsd:p,async dispatch(u){u?.throwIfAborted();let m="",f=null,h=null,R=null,w=!1;for await(let C of r({model:n.modelId,messages:[{role:"system",content:n.system},{role:"user",content:n.data}],requestId:`learning-judge-${fm()}`,maxTokens:l,temperature:n.temperature,signal:u})){let A=Wc(C.raw);if(typeof A.model=="string"){if(f!==null&&A.model!==f)throw new Error("Judge response model identity changed during streaming.");f=A.model}if(typeof A.id=="string"){if(h!==null&&h!==A.id)throw new Error("Judge response identity changed during streaming.");h=A.id}if(C.type==="text_delta"&&(m+=C.text),m.length>1e5)throw new Error("Judge response exceeds the retained response limit.");C.type==="usage"&&(R=C.usage),C.type==="finish"&&(w=!0)}if(!w||!f||!h)throw new Error("Judge response did not include a complete model receipt.");let k=Ze(R),x=k.promptTokens!==null&&k.completionTokens!==null?nt(R,c):null;return{text:m,modelId:f,modelRevision:null,responseId:h,inputTokens:k.promptTokens,outputTokens:k.completionTokens,costUsd:x}}}},async cancel({run:n}){return!n.judgeCalls?.some(o=>o.status==="reserved")}}}i(qc,"createLearningHostedJudgeProvider");function Wc(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(Wc,"record");function jc(e,t){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new Error(`Judge catalog is missing its ${t}.`);return e}i(jc,"positiveInteger");function sR(e,t=qc()){return{execute(r){if(r.definition.execution.environment.entrypoint==="openpond.javascript-environment.v1")throw new Error("Tool Tasksets require grading an owner-recorded environment attempt.");let n=ni({repository:e,scope:r.scope,run:r.run});return Qa({binding:r.binding,rewards:r.rewards,task:Gr(r.evidence,r.definition),evidence:ri(r.evidence,r.run.output),evaluatorContext:r.evidence.submission.evaluatorContext,signal:r.signal,modelJudge:$a({readRubric:i(o=>e.transaction(r.scope,async s=>ue(await Ue(s,"asset",o.id,1),o)),"readRubric"),async executeBudgeted(o,s){let a=await t.prepare(o,{scope:r.scope,run:r.run});return ei({store:n,maximumCharge:i(()=>a.maximumChargeUsd,"maximumCharge"),dispatch:i((d,l)=>a.dispatch(l),"dispatch")})(`grade-${y(o)}`,o,s)}}),customVerifier:i(async({grader:o,task:s,evidence:a})=>{let c=await yr({repository:e,scope:r.scope,grader:o,task:s,evidence:a,evaluatorContext:r.evidence.submission.evaluatorContext,timeoutMs:r.run.timeoutMs,signal:r.signal});return{score:c.score,passed:c.passed,rewardEligible:o.rewardEligible,failureClass:null,feedback:[c.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:c.evidenceRefs}},"customVerifier")})},async cancel({scope:r,run:n}){let o=await e.transaction(r,s=>Ue(s,"grade",n.id));return o.judgeCalls?.some(s=>s.status==="reserved")?!1:t.cancel({scope:r,run:o})}}}i(sR,"createLocalTaskGradeExecutor");async function yr(e){let t=await e.repository.transaction(e.scope,async r=>ue(await Ue(r,"asset",e.grader.verifierRef.id,1),e.grader.verifierRef));return Vt({source:t,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:Math.min(e.grader.timeoutMs,e.timeoutMs??e.grader.timeoutMs),signal:e.signal,value:{task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,expectedOutput:e.task.expectedOutput,evaluatorContext:e.evaluatorContext,infrastructureError:e.evidence.infrastructureError??null}})}i(yr,"executeLocalLearningVerifier");import hm from"node:path";async function Fc(e,t,r){if(!r)throw new Error("Ordinary tool execution requires its owner's package directory.");let n=$r({taskset:e,adapterId:dn(e)}),o=await wc({root:hm.join(r,"training","tasksets",Me(e)),content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:n.tasksetRelease,environment:n.environmentRelease,verifierSet:n.verifierSetRelease},sources:ec.parse(e.metadata.portableFileInventory)}),s=Ii(o);if(!s)throw new Error("The Taskset does not declare a JavaScript execution graph.");let a=o.taskset.tasks.find(l=>l.id===t.id);if(!a)throw new Error("The selected task is missing from its executable package.");let c=s.assets.find(l=>l.id===s.execution.javascript.module.id),d=s.assets.find(l=>l.id===a.privilegedContextRef);return{execution:s.execution,module:c,initialState:JSON.parse(d.text),definition:Y(o.taskset),instructions:Ci(o),task:a}}i(Fc,"loadOrdinaryToolEnvironment");var ce="openpond.javascript-environment.v1";async function Yt(e,t,r,n){let o=t.tasks.find(s=>s.id===r.id);if(!o||y(o)!==y(r))throw new Error("Tool task differs from the selected immutable Taskset.");return t.metadata.taskDefinition===void 0?Fc(t,r,n):e.learningRepository().transaction(t.profileId,async s=>{let a=await Be(s,"definition",Ee.parse(t.metadata.taskDefinition)),c=a.execution.environment;if(c.entrypoint!==ce||t.environment.entrypoint!==ce||t.environment.kind!=="agent"||t.environment.stateful!==c.stateful||t.environment.defaultTimeoutMs!==c.defaultTimeoutMs||t.environment.networkPolicy!==c.networkPolicy||t.environment.deterministicSeeds!==c.deterministicSeeds||y(t.environment.toolNames)!==y(a.execution.tools.map(T=>T.name))||y(t.environment.metadata.portableTools)!==y(a.execution.tools)||y(t.policy)!==y(a.execution.policy)||y(t.metadata.rewardBinding)!==y(a.rewardBinding))throw new Error("Tool Taskset context differs from its pinned task definition.");let d=await Be(s,"binding",a.rewardBinding),l=await Promise.all(d.sources.map(T=>Be(s,"reward",T.reward))),p=Qr.parse({id:r.id,clusterKey:r.clusterKey,split:r.split,input:r.input,expectedOutput:r.expectedOutput,policyVisibleContext:r.policyVisibleContext,privilegedContextRef:r.privilegedContextRef,artifactRefs:[],tags:r.tags}),u={...a.execution,graders:$e(d,l),tasks:[p]},m=await Ue(s,"asset",Si(u),1),f=vi.parse(JSON.parse(ue(m,m.asset))),h=[f.javascript.module,f.environment.actionSchemaRef,f.environment.observationSchemaRef,f.environment.stateSchemaRef].filter(T=>T!==null),R=[...new Set([...h.map(T=>T.id),r.privilegedContextRef].filter(T=>T!==null))],w=await Promise.all(R.map(T=>Ue(s,"asset",T,1))),k=xi(u,m,w);if(y(t.environment.metadata.portableExecutionResources)!==y({environment:k.environment,verifierSet:k.verifierSet}))throw new Error("Tool Taskset publication differs from its private execution resources.");let x=w.find(T=>T.id===k.javascript.module.id),C=w.find(T=>T.id===r.privilegedContextRef),A=JSON.parse(ue(C,C.asset));return{execution:k,module:x,initialState:A,definition:Y(a),instructions:a.instructions,task:p}})}i(Yt,"loadStarterToolEnvironment");function Uc(e){let t=Mo({text:e.instructions,path:"policy/instructions.txt",mediaType:"text/plain",visibility:"policy"}),r=Mo({text:JSON.stringify({environment:Y(e.execution.environment),verifierSet:Y(e.execution.verifierSet),javascript:Y(e.execution.javascript)}),path:"environment/dependencies.json",mediaType:"application/json",visibility:"host_private"}),n=La({schemaVersion:"openpond.agentSnapshot.v2",id:`starter-agent-${e.definition.contentHash}`,sourceRelease:{id:e.definition.id,contentHash:e.definition.contentHash},instructions:[t.asset],skills:[],agents:[],toolDeclarations:e.execution.javascript.tools,capabilityRequirements:[],dependencyLock:r.asset,portability:{portable:!0,blockers:[],localOnlyAssetRefs:[],hostPrivateAssetRefs:[e.module.id,r.id]},metadata:{execution:ce}}),o=Da({schemaVersion:"openpond.harnessRelease.v2",id:`starter-harness-${n.contentHash}`,agentSnapshot:{id:n.id,contentHash:n.contentHash},program:e.module.asset,tools:e.execution.javascript.tools,lifecycle:{create:!0,reset:!0,step:!0,collect:!0,destroy:!0,resetScope:"attempt"},graderInterface:{visibleEvidence:["output"],privilegedEvidence:["environment_state","private_verifier"],privateVerifierIsolation:!0},files:[t.asset,r.asset,e.module.asset],metadata:{execution:ce}});return{agentSnapshot:n,harnessRelease:o,instructionContext:void 0}}i(Uc,"compileStarterToolHarness");import Sn from"node:path";import{constants as Bc}from"node:fs";import{open as gm,realpath as Jc}from"node:fs/promises";var ym=g.string().regex(/^[a-f0-9]{64}$/),wm=g.object({schemaVersion:g.literal("openpond.starterToolAttempt.v1"),requestId:g.string(),taskset:Ee,taskHash:ym,model:g.unknown(),seed:g.number().int(),attempt:g.number().int(),definition:Ee,policySource:g.enum(["model","fixture"]),environment:Ee,verifierSet:Ee,result:Ti}).strict();async function xn(e){let{store:t,taskset:r,task:n,attempt:o}=e,s=await Yt(t,r,n,e.storeDir),c=(await t.listTaskAttemptArtifacts({attemptId:o.id})).filter(x=>x.kind==="environment_state"&&x.tasksetId===r.id&&x.taskId===n.id&&x.attemptId===o.id&&o.artifactRefs.includes(x.id)&&o.privilegedOutcomeRef===x.id);if(c.length!==1)throw new Error("Tool grading requires exactly one owner-recorded environment artifact.");let d=c[0],l=await Jc(Sn.join(e.storeDir,"training","evaluation-artifacts")),p=await Jc(Sn.dirname(d.path)),u=Sn.relative(l,p);if(!u||u.startsWith("..")||Sn.isAbsolute(u))throw new Error("Tool evidence is outside the execution owner's artifact directory.");let m=await gm(d.path,Bc.O_RDONLY|Bc.O_NOFOLLOW),f;try{let x=await m.stat();if(!x.isFile()||x.size>33554432||x.size!==d.sizeBytes)throw new Error("Tool evidence exceeds its declared byte boundary.");f=await m.readFile()}finally{await m.close()}if(f.byteLength!==d.sizeBytes||ve(f)!==d.sha256)throw new Error("Owner-recorded tool evidence bytes changed.");let h=wm.parse(JSON.parse(f.toString("utf8"))),{contentHash:R,...w}=h.result,k=h.result;if(y(w)!==R||!k.snapshot||!k.environmentCleanupComplete||!["completed","budget_exhausted","policy_failure"].includes(k.status)||y(h.taskset)!==y(Y(r))||h.taskHash!==y(n)||y(h.model)!==y(o.modelRef)||h.seed!==o.seed||h.attempt!==o.attempt||o.tasksetId!==r.id||o.taskId!==n.id||o.split!==n.split||y(h.definition)!==y(s.definition)||y(h.environment)!==y(Y(s.execution.environment))||y(h.verifierSet)!==y(Y(s.execution.verifierSet))||k.taskId!==n.id||k.snapshot.seed!==o.seed||k.snapshot.inputHash!==y(n.input)||y(k.snapshot.definition)!==y(Y(s.execution.javascript))||k.snapshot.finalStateHash!==y(k.snapshot.state)||y(o.output)!==y({text:k.output??""})||o.metadata.environmentAttemptHash!==R)throw new Error("Tool evidence differs from the selected task, model or recorded attempt.");if(o.metadata.policySource!==h.policySource)throw new Error("Tool evidence policy source changed.");return{environment:{status:k.status,collected:k.collected,definition:k.snapshot.definition,initialStateHash:k.snapshot.initialStateHash,finalStateHash:k.snapshot.finalStateHash,finalState:k.snapshot.state,events:k.snapshot.events}}}i(xn,"readStarterToolEvidence");async function ot(e,t){if(t.metadata.learning!==void 0)return qt.parse(t.metadata.learning);if(t.metadata.rewardBinding===void 0)return;let r=Ee.parse(t.metadata.rewardBinding);return e.learningRepository().transaction(t.profileId,async n=>{let o=await Be(n,"binding",r),s=[...new Map(o.sources.map(c=>[`${c.reward.id}:${c.reward.revision}:${c.reward.contentHash}`,c.reward])).values()],a=await Promise.all(s.map(c=>Be(n,"reward",c)));return{binding:o,rewards:a}})}i(ot,"resolveTasksetRewardBinding");async function km(e,t,r){let n=await ot(e,t);if(!n)return{rewardExecution:void 0,verifierAssets:[]};let o=$e(n.binding,n.rewards).flatMap(c=>c.kind==="custom_verifier"?[c.verifierRef]:c.kind==="model_judge"?[c.rubricRef]:[]),s=t.metadata.learning===void 0?void 0:await un(r,t);if(s){let c=[...new Map(o.map(d=>{let l=s.learningResources.assets.find(p=>p.id===d.id);if(!l)throw new Error("Imported batch verifier is missing from its package.");return ue(l,d),[l.id,l]})).values()];return{rewardExecution:n,verifierAssets:c}}let a=await e.learningRepository().transaction(t.profileId,async c=>{let d=new Map;for(let l of o){let p=await Ue(c,"asset",l.id,1);ue(p,l),d.set(p.id,p)}return[...d.values()]});return{rewardExecution:n,verifierAssets:a}}i(km,"resolveTasksetTrainingReward");async function YR(e,t,r){let n=await km(e,t,r.storeDir);if(!n.rewardExecution)return n;let o=$e(n.rewardExecution.binding,n.rewardExecution.rewards);if(!n.rewardExecution.binding.sources.some(s=>s.role==="evaluation"&&s.weight>0))throw new Error("Managed training requires a positively weighted evaluation Reward source for retained validation.");if(o.some(s=>s.kind==="custom_verifier"&&s.runtime==="sandbox_process"))throw new Error("Managed preparation requires a qualified isolated sandbox for process verifiers.");if(r.placement!=="remote"||t.environment.kind==="work"||t.capabilities.requiresState||t.capabilities.requiresTools||o.some(s=>s.kind==="human"||s.kind==="model_judge"&&(!s.model||s.calibrationStatus!=="passed"))||t.tasks.some(s=>s.privilegedContextRef!==null)||!n.rewardExecution.binding.sources.some(s=>s.role==="training"&&s.weight>0)||r.hasLearnedPreferenceReward)throw new Error("This Reward binding requires an additional managed execution adapter.");return n}i(YR,"resolveManagedTasksetReward");async function Vc(e,t,r){let n=await ot(e,t);if(!n)throw new Error("The Taskset has no published Reward binding.");let o=$e(n.binding,n.rewards);return async({grader:s,task:a,attempt:c,signal:d})=>{let l=t.tasks.find(f=>f.id===a.id);if(!l||y(l)!==y(a))throw new Error("Task differs from the selected immutable Taskset.");let p=null;if(t.environment.entrypoint===ce){if(!r)throw new Error("Tool grading requires the execution owner's artifact directory.");p=await xn({store:e,storeDir:r,taskset:t,task:a,attempt:c})}else if(a.privilegedContextRef!==null)throw new Error("This authored Taskset requires a private evaluator-context resolver.");let u=o.find(f=>f.id===s.id);if(u?.kind!=="custom_verifier")throw new Error("The verifier is absent from the published Reward binding.");let m=Qr.parse({id:a.id,clusterKey:a.clusterKey,split:a.split,input:a.input,expectedOutput:a.expectedOutput,policyVisibleContext:a.policyVisibleContext,privilegedContextRef:a.privilegedContextRef,artifactRefs:[],tags:a.tags});return yr({repository:e.learningRepository(),scope:t.profileId,grader:u,task:m,evidence:{output:c.output,artifactRefs:c.artifactRefs,runtimeEventRefs:c.runtimeEventRefs,infrastructureError:c.infrastructureError},evaluatorContext:p,signal:d})}}i(Vc,"createTasksetBindingVerifier");async function Tt(e,t,r){let n=r?Ee.parse(r):null;if(n&&n.id!==t)throw new Error("Taskset execution reference differs from the requested Taskset.");let o=n?await e.getTasksetRevision(n.id,n.revision):await e.getTaskset(t);if(!o||n&&(o.contentHash!==n.contentHash||et(o)!==n.contentHash))throw new Error("Taskset execution revision is unavailable or differs from its immutable hash.");return o}i(Tt,"requireLocalTasksetRevision");async function tT(e,t,r){let n=await e.releaseForTaskset(t);return n||$r({taskset:t,rewardExecution:r?await ot(r,t):void 0,adapterId:dn(t)}).tasksetRelease}i(tT,"requireReleasedTaskset");import{createHash as Rm}from"node:crypto";function Xc(e,t,r){if(t.metadata.importedPackageHash!==e.contentHash)throw new Error("The Work projection differs from the admitted published package.");let n=e.taskset.tasks.find(a=>a.id===r.id);if(!n||n.split!==r.split||y(r.metadata.portableTaskRecord)!==y(n))throw new Error("The Work task differs from its exact released population.");let o=0,s=(r.assets??[]).map(a=>{let c=n.artifactRefs.find(u=>u.id===a.id&&u.contentHash===a.sha256&&u.path===a.artifactRef),d=c&&e.files.find(u=>y(u.asset)===y(c));if(!c||!d||c.visibility!=="policy")throw new Error("The Work task requests an unavailable or private evaluator file.");let l=Buffer.from(bi(d));if(Rm("sha256").update(l).digest("hex")!==a.sha256||l.length!==a.sizeBytes||a.split!==r.split)throw new Error("The published Work input manifest changed.");if(o+=l.length,o>25e7)throw new Error("Published Work inputs exceed the admitted 250 MB ceiling.");return{...a,assetId:a.id,storageName:a.fileName,bytes:l,sizeBytes:l.length}});if(s.length!==n.artifactRefs.length)throw new Error("The exact policy-visible Work asset population is incomplete.");return s}i(Xc,"advancedPublishedWorkAssets");import{AsyncLocalStorage as Tm}from"node:async_hooks";var Zc=new Tm,wr=i(()=>Zc.getStore(),"currentScheduledAdmissionGuard");function lT(e,t){return Zc.run(e,t)}i(lT,"withScheduledAdmission");var Qt=class extends Error{constructor(r,n,o){super("The scheduled request was refused before provider transport.",{cause:o});this.executionId=r;this.requestId=n}executionId;requestId;static{i(this,"ScheduledTransportNotInvokedError")}},zc=class extends Error{constructor(r,n,o){super("The scheduled admission lost its exact owner lease before Run insertion.");this.scheduleId=r;this.operationId=n;this.configurationHash=o}scheduleId;operationId;configurationHash;static{i(this,"ScheduledAdmissionRejectedError")}};function bn(e,t){let r=t.scope==="arguments"?e.action?.arguments:e[t.scope];for(let n of t.path){if(!r||typeof r!="object"||!Object.hasOwn(r,n))throw new Error("environment_service_reference_missing");r=r[n]}return r}i(bn,"resolveValue");async function kr(e,t){if(!e.executionServices?.length)return t.controller(e);if(!Number.isSafeInteger(e.timeoutMs)||e.timeoutMs<1||e.timeoutMs>3e4)throw new Error("environment_timeout_invalid");G(e.value,4194304);let r=e.executionServices.map(c=>ci.parse(c));if(r.length>16)throw new Error("environment_service_limit");if(new Set(r.map(c=>c.id)).size!==r.length)throw new Error("environment_service_identity_collision");let n=Date.now(),o=new AbortController,s=e.signal?AbortSignal.any([e.signal,o.signal]):o.signal,a=setTimeout(()=>o.abort(new Error("environment_timeout")),e.timeoutMs);try{let c=Object.create(null);for(let l of r){if(s.throwIfAborted(),l.operation!==e.operation||l.operation==="step"&&l.toolName!==e.value.action?.name)throw new Error("environment_service_operation_mismatch");let p=Math.min(l.timeoutMs,e.timeoutMs-(Date.now()-n));if(p<1)throw new Error("environment_timeout");let u=new AbortController,m=AbortSignal.any([s,u.signal]),f=setTimeout(()=>u.abort(new Error("environment_service_timeout")),p),h=Date.now();try{if(l.kind==="sqlite.v1"){if(!t.sql)throw new Error("environment_sql_executor_required");let R=bn(e.value,l.sql);if(typeof R!="string"||!R.length||new TextEncoder().encode(R).byteLength>65536||R.includes("\0")){c[l.id]={status:"rejected",code:"invalid_submission"};continue}let w=Kr({schemaVersion:"openpond.sqlExecution.v1",snapshot:bn(e.value,l.snapshot),sql:R,maxRows:l.maxRows,maxResultBytes:l.maxResultBytes});c[l.id]=await t.sql({request:w,timeoutMs:e.timeoutMs,signal:m})}else{let R=bn(e.value,l.source),w=bn(e.value,l.cases);if(typeof R!="string"||new TextEncoder().encode(R).byteLength>524288){c[l.id]={status:"rejected",code:"invalid_submission"};continue}if(!Array.isArray(w)||!w.length||w.length>l.maxCases||w.some(x=>!x||typeof x!="object"||!Object.hasOwn(x,"input")))throw new Error("environment_service_cases_invalid");let k=[];for(let x of w){let C=p-(Date.now()-h);if(C<1&&(u.abort(new Error("environment_service_timeout")),m.throwIfAborted()),k.push(await t.candidate({source:R,exportName:l.exportName,value:structuredClone(x.input),timeoutMs:Math.max(1,C),maxResultBytes:l.maxResultBytes,deterministic:!0,errorPrefix:"candidate",signal:m})),new TextEncoder().encode(JSON.stringify({status:"completed",cases:k})).byteLength>l.maxResultBytes)throw new Bo("candidate_result_too_large")}c[l.id]={status:"completed",cases:k}}}catch(R){if(s.throwIfAborted(),u.signal.aborted)c[l.id]={status:"rejected",code:"timeout"};else if(R instanceof Bo)c[l.id]={status:"rejected",code:"execution_failed"};else throw R}finally{clearTimeout(f)}}s.throwIfAborted();let d=e.timeoutMs-(Date.now()-n);if(d<1)throw new Error("environment_timeout");return await t.controller({...e,executionServices:void 0,value:{...e.value,services:c},timeoutMs:d,signal:s})}finally{clearTimeout(a)}}i(kr,"executeEnvironmentWithServices");async function vm(e){let t=lr.parse(e.operation);return kr(e,{candidate:Jo,controller:i(async r=>Ft.parse(await Jo({...r,exportName:t,maxResultBytes:1572864,deterministic:!0,errorPrefix:"environment"})),"controller")})}i(vm,"executeJavaScriptEnvironment");var vt=class extends Error{static{i(this,"JavaScriptEnvironmentActionError")}code;constructor(t){super(`environment_${t}`),this.code=t,this.name="JavaScriptEnvironmentActionError"}};async function Yc(e){let t=li(e.definition),r=ue(e.asset,t.module);if(!Number.isSafeInteger(e.seed))throw new Error("environment_seed_invalid");let n=e.seed;G(e.input,1048576),G(e.initialState,t.maxStateBytes);let o=structuredClone(e.input),s=structuredClone(e.initialState),a=e.execute??vm,c=new AbortController,d=e.signal?AbortSignal.any([e.signal,c.signal]):c.signal,l={},p=0,u=!1,m=!1,f=null,h=[],R=null;async function w(x,C){if(u||m)throw new Error("environment_closed");if(f)throw new Error("environment_operation_in_progress");d.throwIfAborted(),C&&(G(C,262144),C=structuredClone(C));let A=y(l),T=C?t.tools.find(v=>v.name===C.name):null;if(x==="step"){if(p>=t.maxSteps)throw new vt("step_budget_exhausted");p++}let P=i((v,I)=>h.push({sequence:h.length,operation:x,toolName:C?.name??null,arguments:C?structuredClone(C.arguments):null,observation:v,beforeStateHash:A,afterStateHash:y(l),error:I}),"record");try{if(C&&!T)throw new vt("unknown_tool");if(C&&T&&!Ya(T.inputSchema,C.arguments).valid)throw new vt("invalid_arguments");let v=t.executionServices?.filter(E=>E.operation===x&&(x!=="step"||E.toolName===C?.name)),I=a({source:r,operation:x,value:{input:o,initialState:s,seed:n,state:structuredClone(l),sequence:p,action:C?structuredClone(C):null},timeoutMs:Math.min(t.operationTimeoutMs,T?.timeoutMs??t.operationTimeoutMs),signal:d,...v?.length?{executionServices:v}:{}});f=I;let S=Ft.parse(await I);if(d.throwIfAborted(),G(S.state,t.maxStateBytes),G(S.observation,t.maxObservationBytes),(x==="collect"||T?.sideEffect==="read")&&y(S.state)!==A)throw new Error("environment_read_changed_state");l=structuredClone(S.state);let N=structuredClone(S.observation);return P(N,null),{state:structuredClone(l),observation:N}}catch(v){throw P(null,v instanceof Error?v.message:"environment_execution_failed"),v}finally{f=null}}i(w,"invoke");async function k(){return R||(m=!0,c.abort(new Error("environment_destroyed")),R=(async()=>{try{try{await f}catch{}await a({source:r,operation:"destroy",value:{input:o,initialState:s,seed:n,state:structuredClone(l),sequence:p,action:null},timeoutMs:t.operationTimeoutMs})}finally{l={},u=!0}})(),R)}i(k,"destroy");try{await w("create");let x=await w("reset"),C=y(l),A=i(()=>{if(u)throw new Error("environment_closed");return{definition:{id:t.id,revision:t.revision,contentHash:t.contentHash},inputHash:y(o),seed:n,initialStateHash:C,finalStateHash:y(l),state:structuredClone(l),events:structuredClone(h)}},"snapshot");return{definition:structuredClone(t),observation:x.observation,async step(T){return(await w("step",T)).observation},async collect(){let T=await w("collect");return{...A(),observation:T.observation}},snapshot:A,destroy:k}}catch(x){try{await k()}catch{}throw x}}i(Yc,"createJavaScriptEnvironmentSession");var Em=g.string().trim().min(1).max(500),Sm=g.string().regex(/^[a-f0-9]{64}$/),Ko=g.object({kind:g.literal("model"),id:g.string().trim().min(1).max(191),taskId:Em,admissionHash:Sm,model:tc,instructions:g.string().max(262144),harness:Ka.optional(),input:g.record(g.string(),g.unknown()),policyVisibleContext:g.record(g.string(),g.unknown()),timeoutMs:g.number().int().positive().max(36e5),environment:g.discriminatedUnion("kind",[g.object({kind:g.literal("text")}).strict(),g.object({kind:g.literal("javascript"),definition:di,asset:Ga,initialState:g.record(g.string(),g.unknown()),seed:g.number().int().safe()}).strict()])}).strict(),MT=g.discriminatedUnion("kind",[Ko,g.object({kind:g.literal("profile"),request:g.unknown()}).strict()]);var Qc=g.string().trim().min(1).max(200),JT=g.object({receiptId:Qc,taskId:Qc,seed:g.string(),request:Ko}).strict(),M=class extends Error{constructor(r,n,o=409){super(n);this.code=r;this.status=o}code;status;static{i(this,"LocalExperimentError")}};async function _o(e,t=fn,r=!1){if(e.kind==="hosted_harness")return _o({kind:"hosted_chat",modelId:e.modelId,maxOutputTokens:4096,temperature:0,topP:1},t);if(e.kind!=="hosted_chat")throw new M("local_target_not_qualified","Select an explicit model target for local execution.",422);if(e.harness&&!r)throw new M("local_harness_not_qualified","This released Harness requires its native turn executor; it cannot run as a model-only case.",422);let n=await t();if(n.error)throw new M("local_model_catalog_unavailable",n.error,422);let o=n.models.find(m=>m.id===e.modelId);if(!o)throw new M("local_model_unavailable","The selected model is unavailable.",422);let s=Gc(o.raw),a=En(s),c=Gc(s.capabilities).samplingParameters===!0;if(!c&&(e.temperature!==0||e.topP!==1))throw new M("local_sampling_unsupported","This model uses provider-managed sampling. Custom temperature or top P requires an explicit supported capability.",422);let d=Kc(s.context_window),l=Kc(s.output_limit);if(e.maxOutputTokens>l)throw new M("local_model_output_limit","The model cannot accept the configured output limit.",422);let p={providerId:"openpond",modelId:e.modelId,maxOutputTokens:e.maxOutputTokens,...c?{temperature:e.temperature,topP:e.topP}:{},messages:e.messages??[]},u=(d*Math.max(a.inputUsdPerMillionTokens,a.cachedInputUsdPerMillionTokens)+e.maxOutputTokens*a.outputUsdPerMillionTokens)/1e6;if(!(u>0))throw new M("local_model_charge_unknown","The model must publish an enforceable positive request ceiling.",422);return{model:{...p,configurationHash:y({configuration:p,contextWindow:d,outputLimit:l,pricing:a})},pricing:a,maximumChargeUsd:u}}i(_o,"prepareLocalExperimentModel");function $T(e){let t=e.admission.streamFactory?.(e)??e.admission.stream??e.stream??hn;return async function*(r){if(!r.requestId)throw new M("local_request_identity_missing","Every local model dispatch requires a stable request identity.");let n=e.admission.model;if(r.model!==n.modelId||r.maxTokens!==n.maxOutputTokens||r.temperature!==n.temperature||r.topP!==n.topP)throw new M("local_model_configuration_conflict","The native model request changed its admitted model, output limit or sampling configuration.",422);let o=wr();o?.assertExecution(e.executionId);let s={teamId:e.teamId,id:e.executionId,ownerId:e.ownerId,requestId:r.requestId};await e.store.reserveLocalCharge({...s,caseId:e.caseId,maximumUsd:e.admission.maximumChargeUsd}),r.signal?.aborted&&(await e.store.settleLocalCharge({...s,costUsd:null,usage:null,notDispatched:!0}),r.signal.throwIfAborted());try{await e.store.markLocalChargeDispatched(s.teamId,s.id,s.requestId,s.ownerId)}catch(p){throw o&&await e.store.settleLocalCharge({...s,costUsd:null,usage:null,notDispatched:!0}),p}await e.store.appendLocalExperimentEvent({...s,caseId:e.caseId,type:"model.dispatch",payload:{requestId:s.requestId,model:r.model,messages:r.messages}});let a=null,c=!1,d=!1,l;try{for await(let p of t(r)){d=!0,p.type==="usage"&&(a=p.usage),p.type==="finish"&&(c=!0);let u=p.type==="text_delta"?{requestId:s.requestId,text:p.text}:p.type==="tool_call_delta"?{requestId:s.requestId,toolCalls:p.toolCalls}:p.type==="usage"?{requestId:s.requestId,usage:Ze(p.usage)}:p.type==="finish"?{requestId:s.requestId,finishReason:p.finishReason}:null;u&&await e.store.appendLocalExperimentEvent({...s,caseId:e.caseId,type:`model.${p.type}`,payload:u}),yield p}}catch(p){throw!d&&p instanceof Qt&&p.executionId===s.id&&p.requestId===s.requestId&&(l=p),p}finally{let p=Ze(a),u=e.admission.pricing&&c&&p.promptTokens!==null&&p.completionTokens!==null?nt(a,e.admission.pricing):null;await e.store.settleLocalCharge({...s,costUsd:u,confirmedTransportNotInvoked:l,usage:a===null?null:{promptTokens:p.promptTokens,completionTokens:p.completionTokens,cachedPromptTokens:p.cachedPromptTokens,uncachedPromptTokens:p.uncachedPromptTokens}})}}}i($T,"createLocalBudgetedModelStream");function Gc(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}i(Gc,"record");function Kc(e){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new M("local_model_bound_unknown","The model catalog is missing an enforceable context/output bound.",422);return e}i(Kc,"positive");import{AsyncLocalStorage as xm}from"node:async_hooks";import bm from"node:path";var _c=new xm;function $c(){return _c.getStore()}i($c,"currentAdvancedRefinerBoundary");function dv(e){let t=fi(bm.join(e.storeDir,"library","harnesses","advanced-refiner-paid-calls.sqlite"));t.exec(`CREATE TABLE IF NOT EXISTS advanced_refiner_budgets(id TEXT PRIMARY KEY,actor_id TEXT NOT NULL,team_id TEXT NOT NULL,pin_hash TEXT NOT NULL,maximum_spend REAL NOT NULL,maximum_steps INTEGER NOT NULL,expires_at TEXT NOT NULL);

package/dist/chunks/{chunk-TIUSKRBP.js → chunk-UBFRWC6I.js}

@@ -1,5 +1,5 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{V as Bt,Wa as ar,ab as et,ba as _t,bb as tt,db as at,ib as nr,lb as or,mb as sr,q as Ze,v as er,x as tr}from"./chunk-NHJ26KXF.js";import{a as rr}from"./chunk-EMKZVP7Z.js";import{i as Ya}from"./chunk-VOIH2BQA.js";import{a as Ye}from"./chunk-T5LE6R5S.js";import{f as Za,h as He}from"./chunk-QJRXTIOC.js";import{a as We,b as Sa,c as Ha,d as Ea,e as Aa,f as xa,g as Ga,h as _e,ha as Xa,i as Ka,j as Oe,ka as Xe,pa as Qe,qa as se,ra as Qa}from"./chunk-7MXZJQS7.js";import{g as he}from"./chunk-5CMKGD43.js";import{M as Ja,q as Wa}from"./chunk-HA7EFTPF.js";import{b as Va}from"./chunk-L3KS6MOV.js";import{Ib as Na,N as Be,fc as Ua,jc as Ge,lc as F,mc as Ke,od as Mt}from"./chunk-XDMUD4FW.js";import{$ as Ct,W as Da,Y as Se,Z as Je,ga as Fa,ka as $a,ma as za}from"./chunk-IKWDL3RB.js";import{Ca as qa,Dc as La,Qa as Me,Ta as Tt,xa as ja}from"./chunk-QRWPEW3Z.js";import{E as Ia,K as Pa,N as Ta,V as Ca,X as Ma,j as ba,k as Ve,p as ya}from"./chunk-WV2APYWG.js";import{Ma as Ie,Na as Ra,W as ka}from"./chunk-PJM2G7SG.js";import{B as ma,Ca as wa,D as ua,E as pa,Ma as va,S as fa,p as Le,r as ia,s as da,sa as ha,t as ca,ta as ga,u as la}from"./chunk-SDEDAHV6.js";import{Z as Oa}from"./chunk-DALJMSJG.js";import{L as Ba}from"./chunk-TGIB24K4.js";import{h as sa,k as re,l as fe,m as v}from"./chunk-NNF6KCHD.js";import{b as M}from"./chunk-ES2ISFJC.js";import{n as _a}from"./chunk-H7IWJN7L.js";import{a as h}from"./chunk-FTQLQJQV.js";import{execFile as Wr}from"node:child_process";import{readFile as Jr}from"node:fs/promises";import{promisify as Gr}from"node:util";var Kr=5,Xr=20*1024*1024,ir=4e4,jt=Gr(Wr);function Zn(e){return async({grader:t,task:a,attempt:n})=>{let s=[],o=0,i=await Qr({store:e.store,attempt:n}),d=n.modelRef?.providerId===t.judge.providerId&&n.modelRef.modelId===t.judge.modelId?tr(n.metadata.hostedTokenPricing)??void 0:void 0,g=await e.modelText({model:t.judge,hostedTokenPricing:d,signal:new AbortController().signal,requestId:`task-judge:${n.id}:${t.id}`,onUsage:h((m,r)=>{s.push(m),typeof r=="number"&&Number.isFinite(r)&&r>=0&&(o+=r)},"onUsage"),messages:[{role:"system",content:`Apply this rubric and return JSON only with score (0..1), passed, and feedback.
+import{V as Bt,Wa as ar,ab as et,ba as _t,bb as tt,db as at,ib as nr,lb as or,mb as sr,q as Ze,v as er,x as tr}from"./chunk-TBNBCOW7.js";import{a as rr}from"./chunk-EMKZVP7Z.js";import{i as Ya}from"./chunk-OE2VF6KU.js";import{a as Ye}from"./chunk-T5LE6R5S.js";import{f as Za,h as He}from"./chunk-QJRXTIOC.js";import{a as We,b as Sa,c as Ha,d as Ea,e as Aa,f as xa,g as Ga,h as _e,ha as Xa,i as Ka,j as Oe,ka as Xe,pa as Qe,qa as se,ra as Qa}from"./chunk-7MXZJQS7.js";import{g as he}from"./chunk-5CMKGD43.js";import{M as Ja,q as Wa}from"./chunk-HA7EFTPF.js";import{b as Va}from"./chunk-L3KS6MOV.js";import{Ib as Na,N as Be,fc as Ua,jc as Ge,lc as F,mc as Ke,od as Mt}from"./chunk-XDMUD4FW.js";import{$ as Ct,W as Da,Y as Se,Z as Je,ga as Fa,ka as $a,ma as za}from"./chunk-IKWDL3RB.js";import{Ca as qa,Dc as La,Qa as Me,Ta as Tt,xa as ja}from"./chunk-QRWPEW3Z.js";import{E as Ia,K as Pa,N as Ta,V as Ca,X as Ma,j as ba,k as Ve,p as ya}from"./chunk-WV2APYWG.js";import{Ma as Ie,Na as Ra,W as ka}from"./chunk-PJM2G7SG.js";import{B as ma,Ca as wa,D as ua,E as pa,Ma as va,S as fa,p as Le,r as ia,s as da,sa as ha,t as ca,ta as ga,u as la}from"./chunk-SDEDAHV6.js";import{Z as Oa}from"./chunk-DALJMSJG.js";import{L as Ba}from"./chunk-TGIB24K4.js";import{h as sa,k as re,l as fe,m as v}from"./chunk-NNF6KCHD.js";import{b as M}from"./chunk-ES2ISFJC.js";import{n as _a}from"./chunk-H7IWJN7L.js";import{a as h}from"./chunk-FTQLQJQV.js";import{execFile as Wr}from"node:child_process";import{readFile as Jr}from"node:fs/promises";import{promisify as Gr}from"node:util";var Kr=5,Xr=20*1024*1024,ir=4e4,jt=Gr(Wr);function Zn(e){return async({grader:t,task:a,attempt:n})=>{let s=[],o=0,i=await Qr({store:e.store,attempt:n}),d=n.modelRef?.providerId===t.judge.providerId&&n.modelRef.modelId===t.judge.modelId?tr(n.metadata.hostedTokenPricing)??void 0:void 0,g=await e.modelText({model:t.judge,hostedTokenPricing:d,signal:new AbortController().signal,requestId:`task-judge:${n.id}:${t.id}`,onUsage:h((m,r)=>{s.push(m),typeof r=="number"&&Number.isFinite(r)&&r>=0&&(o+=r)},"onUsage"),messages:[{role:"system",content:`Apply this rubric and return JSON only with score (0..1), passed, and feedback.
 
 ${t.rubric}`},{role:"user",content:JSON.stringify({input:a.input,expectedOutput:a.expectedOutput,output:n.output,artifactEvidence:i})}]}),f=ar(g);if(!f)throw new Error("Model judge returned invalid structured output.");return{...f,usage:s,...o>0?{costUsd:o}:{}}}}h(Zn,"createTaskAttemptModelJudge");async function Qr(e){let t=(await e.store.listTaskAttemptArtifacts({attemptId:e.attempt.id})).filter(a=>a.kind==="output_artifact").slice(0,Kr);return Promise.all(t.map(async a=>{let n={artifactId:a.id,mediaType:a.mediaType??null,sha256:a.sha256,sizeBytes:a.sizeBytes,requiredOutputPath:typeof a.metadata.requiredOutputPath=="string"?a.metadata.requiredOutputPath:null};if(a.sizeBytes>Xr)return{...n,extraction:"omitted",reason:"artifact exceeds inspection bound"};try{if(a.mediaType==="application/pdf"){let{stdout:s}=await jt("pdftotext",[a.path,"-"],{env:{...process.env,PATH:_t()},maxBuffer:8388608,timeout:15e3});return{...n,extraction:"pdf_text",content:Ot(s)}}if(Zr(a.mediaType)){let s=await en(a.path,a.mediaType);return{...n,extraction:"open_xml_text",content:Ot(s.content),structure:s.structure}}return Yr(a.mediaType)?{...n,extraction:"text",content:Ot(await Jr(a.path,"utf8"))}:{...n,extraction:"metadata_only"}}catch(s){return{...n,extraction:"unavailable",reason:s instanceof Error?s.message.slice(0,500):String(s).slice(0,500)}}}))}h(Qr,"loadTaskAttemptGraderEvidence");function Yr(e){return!!(e?.startsWith("text/")||e==="application/json"||e==="application/javascript"||e==="application/sql"||e==="application/xml"||e==="application/yaml")}h(Yr,"isTextArtifact");function Zr(e){return e==="application/vnd.openxmlformats-officedocument.wordprocessingml.document"||e==="application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"||e==="application/vnd.openxmlformats-officedocument.presentationml.presentation"}h(Zr,"isOpenXmlArtifact");async function en(e,t){let a={env:{...process.env,PATH:_t()},maxBuffer:16777216,timeout:15e3},{stdout:n}=await jt("unzip",["-Z1",e],a),s=n.split(/\r?\n/).filter(Boolean),o=tn(s,t).slice(0,250),i=[];for(let d of o){let{stdout:g}=await jt("unzip",["-p",e,d],a),f=rn(g);f&&i.push(`[${d}]
 ${f}`)}return{content:i.join(`

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-SI3EWWUM.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-HEPLG5FY.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-QHARNR4M.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-HEPLG5FY.js";import{c as I,f as j,h as D}from"./chunk-DFM5JTS2.js";import{k as x}from"./chunk-VOIH2BQA.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(`${JSON.stringify(await G(o))}
+import{a as V}from"./chunk-OC6UWKBL.js";import{c as I,f as j,h as D}from"./chunk-ZCWHVLJ4.js";import{k as x}from"./chunk-OE2VF6KU.js";import{a as E}from"./chunk-T5LE6R5S.js";import"./chunk-QJRXTIOC.js";import"./chunk-SDO4FC4Q.js";import"./chunk-IR3YJWPW.js";import"./chunk-7K72KS5C.js";import{ha as b,qa as C}from"./chunk-7MXZJQS7.js";import"./chunk-7YR6CBNU.js";import"./chunk-5CMKGD43.js";import"./chunk-HA7EFTPF.js";import"./chunk-L3KS6MOV.js";import"./chunk-7QEE4PBI.js";import"./chunk-XDMUD4FW.js";import{ta as m}from"./chunk-IKWDL3RB.js";import"./chunk-6MI4KKWF.js";import{Ca as S}from"./chunk-QRWPEW3Z.js";import"./chunk-6FA3GG5B.js";import"./chunk-Z3D4C7BG.js";import"./chunk-WV2APYWG.js";import{ma as H}from"./chunk-PJM2G7SG.js";import{Ga as y,j as P,m as R}from"./chunk-SDEDAHV6.js";import"./chunk-DALJMSJG.js";import"./chunk-TGIB24K4.js";import{m as d}from"./chunk-NNF6KCHD.js";import{b as e}from"./chunk-ES2ISFJC.js";import"./chunk-H7IWJN7L.js";import{a as s}from"./chunk-FTQLQJQV.js";import{readFile as h,realpath as z,lstat as T}from"node:fs/promises";import l from"node:path";var B=e.object({compile:e.object({repoPath:e.string(),profileId:e.string(),repositoryId:e.string(),sourceRevision:e.string(),workspaceId:e.string(),outputDir:e.string()}).strict(),externalDatasetPackage:e.string().regex(/^external-dataset-[a-f0-9-]+\.json$/).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};