openpond 0.2.70 → 0.2.72

package/dist/chunks/{chunk-7NCM6GBN.js → chunk-TOWY6Y23.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as j}from"./chunk-Y4I6CFZJ.js";import{j as y}from"./chunk-BU6MW65V.js";import{f as i,m as s}from"./chunk-HIQPGYVH.js";import{c as H,d as w,f as k}from"./chunk-LK77GI6M.js";import{Bb as h,Eb as I,Fb as E,oa as v,pa as g,wb as f}from"./chunk-ZY32SZ2P.js";import{A as e}from"./chunk-X2J2P5ZT.js";import{a as m}from"./chunk-FTQLQJQV.js";var A=e.object({sourcePass:i,graders:j.shape.graders,mappings:j.shape.mappings,maximumCostUsd:e.number().finite().positive().max(1e4)}).strict(),C=e.object({externalDatasetBinding:v,scoring:A.optional(),modelId:e.string().min(1).max(300),maximumCostUsd:e.number().finite().positive().max(1e4)}).strict(),R=e.object({source:e.enum(["local","github","openpond_git"]),repositoryId:e.string().trim().min(1),profileId:e.string().trim().min(1)}).strict(),n=e.string().min(1).max(240),c=e.string().regex(/^[a-f0-9]{64}$/),x=e.object({schemaVersion:e.literal("openpond.experimentImprovement.v1"),id:n,operationId:n,requestHash:c,actorId:n,teamId:n,projectId:n.nullable(),ownerWorkspaceId:n,ownerRevision:e.number().int().nonnegative(),ownerSourceRevision:e.string().min(1).max(500),baseRelease:i,profileRef:R,baseProfileSourceRevision:e.string().min(1).max(500),evidence:e.object({manifest:f,result:h}).strict(),component:e.object({kind:e.enum(["skill","workflow","agent","instruction"]),path:e.string().min(1).max(2e3),workflowId:n.optional()}).strict(),mode:e.enum(["manual","llm_assisted"]),status:e.enum(["draft","authoring_queued","authoring","frozen","testing","ready_for_review","adopting","accepted","rolled_back","declined","failed","cancelled"]),revision:e.number().int().positive(),limits:e.object({maximumCostUsd:e.number().finite().positive(),maximumAuthoringSteps:e.number().int().min(1).max(500),maximumDurationMs:e.number().int().min(1e3).max(864e5)}).strict(),authoringSteps:e.number().int().nonnegative().max(500),work:e.object({sessionId:n,turnId:n,revision:e.number().int().positive(),startedAt:e.string().datetime()}).strict().nullable(),frozen:e.object({release:i,sourceRevision:e.string().min(1).max(500),profileSourceRevision:e.string().min(1).max(500),authoringHash:c,partitionHash:c,executableSourceHash:c,protectedClosureHash:c,candidateRevision:e.number().int().positive(),diff:e.array(e.object({path:e.string(),before:c.nullable(),after:c.nullable()}).strict()).max(1e4)}).strict().nullable(),test:e.object({recipe:C,recipeHash:e.string().regex(/^[a-f0-9]{64}$/),baselineOperationId:e.string(),candidateOperationId:e.string(),cancelRequestedAt:e.string().datetime().nullable(),dispatching:e.object({kind:e.enum(["baseline","candidate"]),token:n,pid:e.number().int().positive(),processStart:e.string().nullable(),startedAt:e.string().datetime()}).strict().nullable()}).strict().nullable(),comparison:e.object({baseline:e.object({manifest:f,result:h}).strict(),candidate:e.object({manifest:f,result:h}).strict(),frozenHash:c,humanProjection:w.optional()}).strict().nullable(),adoptionIntent:e.object({operationId:n,expectedOwnerRevision:e.number().int().nonnegative(),frozenHash:c,humanProjection:w.optional()}).strict().nullable(),acceptance:i.nullable(),rollback:i.nullable(),error:e.string().max(2e3).nullable(),createdAt:e.string().datetime(),updatedAt:e.string().datetime(),contentHash:c}).strict();function D(r){let t=x.parse(r),{contentHash:l,...b}=t;if(s(b)!==l)throw new Error("Improvement state integrity failed.");if(I(t.evidence),t.evidence.manifest.teamId!==t.teamId||t.evidence.result.status!=="completed"||!t.evidence.result.cases.some(o=>o.status==="completed"))throw new Error("Improvement requires usable completed owner evidence.");if(t.test&&(s(t.test.recipe)!==t.test.recipeHash||g(t.test.recipe.externalDatasetBinding).profileId!==t.profileRef.profileId))throw new Error("The test recipe changed its exact selected Profile.");if(t.comparison?.humanProjection){let o=k(t.comparison.humanProjection);if(o.teamId!==t.teamId||o.projectId!==t.projectId||s(o.baseline)!==s(t.comparison.baseline)||s(o.candidate)!==s(t.comparison.candidate))throw new Error("Human comparison projection differs from its exact owner evidence.")}if(t.comparison&&(I(t.comparison.baseline),I(t.comparison.candidate),!t.frozen||t.comparison.frozenHash!==s(t.frozen)||!E(t.comparison.baseline,t.comparison.candidate).comparable))throw new Error("Improvement comparison is not bound to a comparable frozen candidate.");if(t.status==="ready_for_review"&&(!t.frozen||!t.comparison))throw new Error("Improvement is missing qualification.");if(t.status==="adopting"&&(!t.adoptionIntent||!t.frozen||t.adoptionIntent.frozenHash!==s(t.frozen)))throw new Error("Improvement adoption intent is not bound to its tested revision.");if(t.status==="accepted"&&(!t.acceptance||!t.comparison||!t.frozen))throw new Error("Improvement acceptance is missing its exact qualification receipt.");if(t.status==="rolled_back"&&(!t.acceptance||!t.rollback))throw new Error("Profile rollback is missing its durable owner receipt.");return t}m(D,"verifyExperimentImprovementState");function X(r){return D({...r,contentHash:s(r)})}m(X,"sealExperimentImprovementState");var Y=e.object({states:e.array(x).max(50),nextCursor:n.nullable()}).strict(),T=e.object({operationId:n,ownerWorkspaceId:n,expectedOwnerRevision:e.number().int().nonnegative(),baseRelease:i,profileRef:R,profileSourceRevision:e.string().min(1).max(500),projectId:n.nullable(),evidence:i,component:x.shape.component,mode:x.shape.mode,limits:x.shape.limits}).strict(),Z=e.discriminatedUnion("operation",[e.object({operation:e.literal("list"),limit:e.number().int().min(1).max(50).default(30),cursor:n.optional()}).strict(),e.object({operation:e.literal("options"),evidence:i}).strict(),e.object({operation:e.literal("start"),request:T}).strict(),e.object({operation:e.enum(["read","humanComparisonOptions"]),id:n}).strict(),e.object({operation:e.literal("instructions"),id:n}).strict(),e.object({operation:e.literal("edit"),id:n,revision:e.number().int().positive(),expectedFileHash:c,text:e.string().max(25e4)}).strict(),e.object({operation:e.literal("mode"),id:n,revision:e.number().int().positive(),mode:e.enum(["manual","llm_assisted"])}).strict(),e.object({operation:e.literal("author"),id:n,revision:e.number().int().positive(),modelId:n,prompt:e.string().min(1).max(2e4)}).strict(),e.object({operation:e.enum(["freeze","finishAuthoring","useCandidate","discard","cancel","rollback"]),id:n,revision:e.number().int().positive()}).strict(),e.object({operation:e.literal("test"),id:n,revision:e.number().int().positive(),recipe:C}).strict(),e.object({operation:e.literal("compare"),id:n,revision:e.number().int().positive(),baseline:i,candidate:i,humanSelections:H.optional()}).strict()]),S=e.object({teamId:n,evidence:i,scoring:A.optional(),profiles:e.array(e.object({id:c,name:e.string(),profileRef:R,ownerWorkspaceId:n,ownerRevision:e.number().int().nonnegative(),baseRelease:i,profileSourceRevision:e.string(),components:e.array(e.object({kind:e.enum(["skill","instruction","agent","workflow"]),path:e.string(),agentId:e.string().optional(),workflowId:e.string().optional(),label:e.string(),binding:v}).strict()).max(500)}).strict()).max(100)}).strict(),O=e.object({candidateId:n,actorId:n,teamId:n,projectId:n.nullable(),candidateRevision:e.number().int().positive(),recipeHash:c,baseline:e.object({manifest:f,result:h}).strict(),candidate:e.object({manifest:f,result:h}).strict(),contentHash:c}).strict();function U(r){let t=O.parse(r),{contentHash:l,...b}=t;if(s(b)!==l)throw new Error("Accepted Human discovery integrity failed.");for(let o of[t.baseline,t.candidate])if(I(o),o.manifest.teamId!==t.teamId||o.manifest.lineage?.humanProjection)throw new Error("Discovery must retain original owner evidence.");return t}m(U,"verifyExperimentImprovementHumanOptions");function ee(r,t){if(!r.test||r.test.cancelRequestedAt)throw new Error("The paired test is unavailable.");let l={candidateId:r.id,actorId:r.actorId,teamId:r.teamId,projectId:r.projectId,candidateRevision:r.revision,recipeHash:r.test.recipeHash,...t};return U({...l,contentHash:s(l)})}m(ee,"sealExperimentImprovementHumanOptions");var a=e.string().trim().min(1).max(240),d=e.string().regex(/^[a-f0-9]{64}$/),_=e.object({schemaVersion:e.literal("openpond.advancedRefinerEvaluation.v1"),operationId:a,actorId:a,teamId:a,projectId:a.nullable(),mode:e.enum(["review_quality","downstream_adaptation"]),profileRef:e.object({source:e.enum(["local","github","openpond_git"]),repositoryId:a,profileId:a}).strict(),sourceRevision:e.string().min(1).max(500),baselineRelease:i,refinerRelease:i,evidence:i,externalDatasetBinding:v,adaptationDatasetBinding:v,privateClosureHash:d,adaptationTaskIds:e.array(a).min(1).max(1e4),holdoutTaskIds:e.array(a).min(1).max(1e4),adaptationSplit:e.enum(["train","validation","test","frozen_eval"]),holdoutSplit:e.enum(["train","validation","test","frozen_eval"]),environmentPolicy:e.discriminatedUnion("kind",[e.object({kind:e.literal("resettable"),environmentHash:d}).strict(),e.object({kind:e.literal("frozen_observations"),environmentHash:d}).strict()]),maximumCostUsd:e.number().finite().positive().max(1e4),maximumDurationMs:e.number().int().min(1e3).max(864e5),maximumModelSteps:e.number().int().min(1).max(1e3),contentHash:d}).strict();function M(r){let t=_.parse(r),{contentHash:l,...b}=t;if(s(b)!==l)throw new Error("Advanced Refiner evaluation admission changed.");let o=g(t.externalDatasetBinding);if(o.profileId!==t.profileRef.profileId||o.protectedProfileClosureHash!==t.privateClosureHash)throw new Error("Advanced evaluation changed its Profile or protected evaluator closure.");if(t.adaptationSplit===t.holdoutSplit||new Set([...t.adaptationTaskIds,...t.holdoutTaskIds]).size!==t.adaptationTaskIds.length+t.holdoutTaskIds.length)throw new Error("Adaptation and private holdout must be disjoint ordered cases.");let p=g(t.adaptationDatasetBinding);if(p.profileId!==o.profileId||p.packageHash!==o.packageHash||s(p.dataset)!==s(o.dataset)||s(p.target)!==s(o.target)||p.protectedProfileClosureHash!==o.protectedProfileClosureHash)throw new Error("This native advanced adapter needs separate adaptation and holdout recipes from the same exact released package and target.");if(p.split!==t.adaptationSplit||o.split!==t.holdoutSplit)throw new Error("Advanced split policies differ from their exact admitted recipes.");let q=new Set(o.population.map(u=>u.taskId)),z=new Set(p.population.map(u=>u.taskId));if(t.holdoutTaskIds.some(u=>!q.has(u))||t.adaptationTaskIds.some(u=>!z.has(u)))throw new Error("Advanced evaluation selected cases outside its exact Dataset recipe.");return t}m(M,"verifyAdvancedRefinerEvaluationPin");function ce(r){return M({...r,contentHash:s(r)})}m(ce,"sealAdvancedRefinerEvaluationPin");var B=e.object({pin:_,modelId:a,profileId:a,model:y,seed:e.number().int(),reasoningEffort:e.enum(["none","off","low","medium","high","xhigh","max"]).nullable()}).strict(),V=e.discriminatedUnion("operation",[e.object({operation:e.literal("start"),request:B}).strict(),e.object({operation:e.enum(["read","cancel","resume"]),id:a}).strict(),e.object({operation:e.literal("list")}).strict()]),W=e.object({taskId:a,sourceCaseHash:d,reviewSessionId:a,reviewTurnId:a,inputHarness:i,trigger:i,decision:e.enum(["no_action","route_deterministically","queue_refiner"]),outcome:i.nullable(),proposal:i.nullable(),validations:e.array(i).max(100),grade:e.object({id:a,contentHash:d,score:e.number().min(0).max(1).nullable(),passed:e.boolean(),failureClass:e.string().nullable()}).strict(),resetComplete:e.boolean()}).strict(),$=e.object({schemaVersion:e.literal("openpond.advancedRefinerReviewReceipt.v1"),id:a,runId:a,pinHash:d,ownerActorId:a,teamId:a,privateClosureHash:d,refinerRelease:i,cases:e.array(W).min(1).max(1e4),reviewScore:e.number().min(0).max(1).nullable(),completedCases:e.number().int().nonnegative(),noActionCases:e.number().int().nonnegative(),routedCases:e.number().int().nonnegative(),proposalCases:e.number().int().nonnegative(),costUsd:e.number().finite().nonnegative().nullable(),cleanupComplete:e.boolean(),createdAt:e.string().datetime(),contentHash:d}).strict();function de(r){return $.parse({...r,contentHash:s(r)})}m(de,"sealAdvancedRefinerReviewReceipt");var P=S.shape.profiles.element,me=S.extend({profiles:e.array(P.extend({components:e.array(P.shape.components.element.extend({adaptationSplits:e.array(e.object({split:e.enum(["train","validation","test","frozen_eval"]),taskIds:e.array(a)})),environmentHash:d,resettable:e.boolean(),reviewQuality:e.boolean()}))})),refinerReleases:e.array(i),activeRefiner:i.nullable(),models:e.array(e.object({id:a,profileId:a,name:e.string(),location:e.enum(["local","hosted"])}))}),L=e.object({operation:e.literal("prepare"),evidence:i,profileOptionId:a,componentLabel:a,mode:e.enum(["review_quality","downstream_adaptation"]),adaptationSplit:e.enum(["train","validation","test","frozen_eval"]),operationId:a,refinerRelease:i,maximumCostUsd:e.number().positive(),maximumDurationMs:e.number().int().positive(),maximumModelSteps:e.number().int().positive()}).strict(),le=e.union([V,L,e.object({operation:e.literal("options"),evidence:i}).strict()]),pe=e.object({modelRequests:e.number().int().nonnegative(),knownSpendUsd:e.number().nonnegative(),totalSpendUsd:e.number().nonnegative().nullable(),heldUsd:e.number().nonnegative(),uncertainRequests:e.number().int().nonnegative()}).strict();export{C as a,x as b,D as c,X as d,Y as e,T as f,Z as g,S as h,U as i,ee as j,_ as k,M as l,ce as m,B as n,V as o,W as p,$ as q,de as r,L as s,pe as t};
+import{a as j}from"./chunk-UGDBMRCT.js";import{j as y}from"./chunk-BYVE4TMU.js";import{f as i,m as s}from"./chunk-HIQPGYVH.js";import{c as H,d as w,f as k}from"./chunk-LK77GI6M.js";import{Bb as h,Eb as I,Fb as E,oa as v,pa as g,wb as f}from"./chunk-ZY32SZ2P.js";import{A as e}from"./chunk-X2J2P5ZT.js";import{a as m}from"./chunk-FTQLQJQV.js";var A=e.object({sourcePass:i,graders:j.shape.graders,mappings:j.shape.mappings,maximumCostUsd:e.number().finite().positive().max(1e4)}).strict(),C=e.object({externalDatasetBinding:v,scoring:A.optional(),modelId:e.string().min(1).max(300),maximumCostUsd:e.number().finite().positive().max(1e4)}).strict(),R=e.object({source:e.enum(["local","github","openpond_git"]),repositoryId:e.string().trim().min(1),profileId:e.string().trim().min(1)}).strict(),n=e.string().min(1).max(240),c=e.string().regex(/^[a-f0-9]{64}$/),x=e.object({schemaVersion:e.literal("openpond.experimentImprovement.v1"),id:n,operationId:n,requestHash:c,actorId:n,teamId:n,projectId:n.nullable(),ownerWorkspaceId:n,ownerRevision:e.number().int().nonnegative(),ownerSourceRevision:e.string().min(1).max(500),baseRelease:i,profileRef:R,baseProfileSourceRevision:e.string().min(1).max(500),evidence:e.object({manifest:f,result:h}).strict(),component:e.object({kind:e.enum(["skill","workflow","agent","instruction"]),path:e.string().min(1).max(2e3),workflowId:n.optional()}).strict(),mode:e.enum(["manual","llm_assisted"]),status:e.enum(["draft","authoring_queued","authoring","frozen","testing","ready_for_review","adopting","accepted","rolled_back","declined","failed","cancelled"]),revision:e.number().int().positive(),limits:e.object({maximumCostUsd:e.number().finite().positive(),maximumAuthoringSteps:e.number().int().min(1).max(500),maximumDurationMs:e.number().int().min(1e3).max(864e5)}).strict(),authoringSteps:e.number().int().nonnegative().max(500),work:e.object({sessionId:n,turnId:n,revision:e.number().int().positive(),startedAt:e.string().datetime()}).strict().nullable(),frozen:e.object({release:i,sourceRevision:e.string().min(1).max(500),profileSourceRevision:e.string().min(1).max(500),authoringHash:c,partitionHash:c,executableSourceHash:c,protectedClosureHash:c,candidateRevision:e.number().int().positive(),diff:e.array(e.object({path:e.string(),before:c.nullable(),after:c.nullable()}).strict()).max(1e4)}).strict().nullable(),test:e.object({recipe:C,recipeHash:e.string().regex(/^[a-f0-9]{64}$/),baselineOperationId:e.string(),candidateOperationId:e.string(),cancelRequestedAt:e.string().datetime().nullable(),dispatching:e.object({kind:e.enum(["baseline","candidate"]),token:n,pid:e.number().int().positive(),processStart:e.string().nullable(),startedAt:e.string().datetime()}).strict().nullable()}).strict().nullable(),comparison:e.object({baseline:e.object({manifest:f,result:h}).strict(),candidate:e.object({manifest:f,result:h}).strict(),frozenHash:c,humanProjection:w.optional()}).strict().nullable(),adoptionIntent:e.object({operationId:n,expectedOwnerRevision:e.number().int().nonnegative(),frozenHash:c,humanProjection:w.optional()}).strict().nullable(),acceptance:i.nullable(),rollback:i.nullable(),error:e.string().max(2e3).nullable(),createdAt:e.string().datetime(),updatedAt:e.string().datetime(),contentHash:c}).strict();function D(r){let t=x.parse(r),{contentHash:l,...b}=t;if(s(b)!==l)throw new Error("Improvement state integrity failed.");if(I(t.evidence),t.evidence.manifest.teamId!==t.teamId||t.evidence.result.status!=="completed"||!t.evidence.result.cases.some(o=>o.status==="completed"))throw new Error("Improvement requires usable completed owner evidence.");if(t.test&&(s(t.test.recipe)!==t.test.recipeHash||g(t.test.recipe.externalDatasetBinding).profileId!==t.profileRef.profileId))throw new Error("The test recipe changed its exact selected Profile.");if(t.comparison?.humanProjection){let o=k(t.comparison.humanProjection);if(o.teamId!==t.teamId||o.projectId!==t.projectId||s(o.baseline)!==s(t.comparison.baseline)||s(o.candidate)!==s(t.comparison.candidate))throw new Error("Human comparison projection differs from its exact owner evidence.")}if(t.comparison&&(I(t.comparison.baseline),I(t.comparison.candidate),!t.frozen||t.comparison.frozenHash!==s(t.frozen)||!E(t.comparison.baseline,t.comparison.candidate).comparable))throw new Error("Improvement comparison is not bound to a comparable frozen candidate.");if(t.status==="ready_for_review"&&(!t.frozen||!t.comparison))throw new Error("Improvement is missing qualification.");if(t.status==="adopting"&&(!t.adoptionIntent||!t.frozen||t.adoptionIntent.frozenHash!==s(t.frozen)))throw new Error("Improvement adoption intent is not bound to its tested revision.");if(t.status==="accepted"&&(!t.acceptance||!t.comparison||!t.frozen))throw new Error("Improvement acceptance is missing its exact qualification receipt.");if(t.status==="rolled_back"&&(!t.acceptance||!t.rollback))throw new Error("Profile rollback is missing its durable owner receipt.");return t}m(D,"verifyExperimentImprovementState");function X(r){return D({...r,contentHash:s(r)})}m(X,"sealExperimentImprovementState");var Y=e.object({states:e.array(x).max(50),nextCursor:n.nullable()}).strict(),T=e.object({operationId:n,ownerWorkspaceId:n,expectedOwnerRevision:e.number().int().nonnegative(),baseRelease:i,profileRef:R,profileSourceRevision:e.string().min(1).max(500),projectId:n.nullable(),evidence:i,component:x.shape.component,mode:x.shape.mode,limits:x.shape.limits}).strict(),Z=e.discriminatedUnion("operation",[e.object({operation:e.literal("list"),limit:e.number().int().min(1).max(50).default(30),cursor:n.optional()}).strict(),e.object({operation:e.literal("options"),evidence:i}).strict(),e.object({operation:e.literal("start"),request:T}).strict(),e.object({operation:e.enum(["read","humanComparisonOptions"]),id:n}).strict(),e.object({operation:e.literal("instructions"),id:n}).strict(),e.object({operation:e.literal("edit"),id:n,revision:e.number().int().positive(),expectedFileHash:c,text:e.string().max(25e4)}).strict(),e.object({operation:e.literal("mode"),id:n,revision:e.number().int().positive(),mode:e.enum(["manual","llm_assisted"])}).strict(),e.object({operation:e.literal("author"),id:n,revision:e.number().int().positive(),modelId:n,prompt:e.string().min(1).max(2e4)}).strict(),e.object({operation:e.enum(["freeze","finishAuthoring","useCandidate","discard","cancel","rollback"]),id:n,revision:e.number().int().positive()}).strict(),e.object({operation:e.literal("test"),id:n,revision:e.number().int().positive(),recipe:C}).strict(),e.object({operation:e.literal("compare"),id:n,revision:e.number().int().positive(),baseline:i,candidate:i,humanSelections:H.optional()}).strict()]),S=e.object({teamId:n,evidence:i,scoring:A.optional(),profiles:e.array(e.object({id:c,name:e.string(),profileRef:R,ownerWorkspaceId:n,ownerRevision:e.number().int().nonnegative(),baseRelease:i,profileSourceRevision:e.string(),components:e.array(e.object({kind:e.enum(["skill","instruction","agent","workflow"]),path:e.string(),agentId:e.string().optional(),workflowId:e.string().optional(),label:e.string(),binding:v}).strict()).max(500)}).strict()).max(100)}).strict(),O=e.object({candidateId:n,actorId:n,teamId:n,projectId:n.nullable(),candidateRevision:e.number().int().positive(),recipeHash:c,baseline:e.object({manifest:f,result:h}).strict(),candidate:e.object({manifest:f,result:h}).strict(),contentHash:c}).strict();function U(r){let t=O.parse(r),{contentHash:l,...b}=t;if(s(b)!==l)throw new Error("Accepted Human discovery integrity failed.");for(let o of[t.baseline,t.candidate])if(I(o),o.manifest.teamId!==t.teamId||o.manifest.lineage?.humanProjection)throw new Error("Discovery must retain original owner evidence.");return t}m(U,"verifyExperimentImprovementHumanOptions");function ee(r,t){if(!r.test||r.test.cancelRequestedAt)throw new Error("The paired test is unavailable.");let l={candidateId:r.id,actorId:r.actorId,teamId:r.teamId,projectId:r.projectId,candidateRevision:r.revision,recipeHash:r.test.recipeHash,...t};return U({...l,contentHash:s(l)})}m(ee,"sealExperimentImprovementHumanOptions");var a=e.string().trim().min(1).max(240),d=e.string().regex(/^[a-f0-9]{64}$/),_=e.object({schemaVersion:e.literal("openpond.advancedRefinerEvaluation.v1"),operationId:a,actorId:a,teamId:a,projectId:a.nullable(),mode:e.enum(["review_quality","downstream_adaptation"]),profileRef:e.object({source:e.enum(["local","github","openpond_git"]),repositoryId:a,profileId:a}).strict(),sourceRevision:e.string().min(1).max(500),baselineRelease:i,refinerRelease:i,evidence:i,externalDatasetBinding:v,adaptationDatasetBinding:v,privateClosureHash:d,adaptationTaskIds:e.array(a).min(1).max(1e4),holdoutTaskIds:e.array(a).min(1).max(1e4),adaptationSplit:e.enum(["train","validation","test","frozen_eval"]),holdoutSplit:e.enum(["train","validation","test","frozen_eval"]),environmentPolicy:e.discriminatedUnion("kind",[e.object({kind:e.literal("resettable"),environmentHash:d}).strict(),e.object({kind:e.literal("frozen_observations"),environmentHash:d}).strict()]),maximumCostUsd:e.number().finite().positive().max(1e4),maximumDurationMs:e.number().int().min(1e3).max(864e5),maximumModelSteps:e.number().int().min(1).max(1e3),contentHash:d}).strict();function M(r){let t=_.parse(r),{contentHash:l,...b}=t;if(s(b)!==l)throw new Error("Advanced Refiner evaluation admission changed.");let o=g(t.externalDatasetBinding);if(o.profileId!==t.profileRef.profileId||o.protectedProfileClosureHash!==t.privateClosureHash)throw new Error("Advanced evaluation changed its Profile or protected evaluator closure.");if(t.adaptationSplit===t.holdoutSplit||new Set([...t.adaptationTaskIds,...t.holdoutTaskIds]).size!==t.adaptationTaskIds.length+t.holdoutTaskIds.length)throw new Error("Adaptation and private holdout must be disjoint ordered cases.");let p=g(t.adaptationDatasetBinding);if(p.profileId!==o.profileId||p.packageHash!==o.packageHash||s(p.dataset)!==s(o.dataset)||s(p.target)!==s(o.target)||p.protectedProfileClosureHash!==o.protectedProfileClosureHash)throw new Error("This native advanced adapter needs separate adaptation and holdout recipes from the same exact released package and target.");if(p.split!==t.adaptationSplit||o.split!==t.holdoutSplit)throw new Error("Advanced split policies differ from their exact admitted recipes.");let q=new Set(o.population.map(u=>u.taskId)),z=new Set(p.population.map(u=>u.taskId));if(t.holdoutTaskIds.some(u=>!q.has(u))||t.adaptationTaskIds.some(u=>!z.has(u)))throw new Error("Advanced evaluation selected cases outside its exact Dataset recipe.");return t}m(M,"verifyAdvancedRefinerEvaluationPin");function ce(r){return M({...r,contentHash:s(r)})}m(ce,"sealAdvancedRefinerEvaluationPin");var B=e.object({pin:_,modelId:a,profileId:a,model:y,seed:e.number().int(),reasoningEffort:e.enum(["none","off","low","medium","high","xhigh","max"]).nullable()}).strict(),V=e.discriminatedUnion("operation",[e.object({operation:e.literal("start"),request:B}).strict(),e.object({operation:e.enum(["read","cancel","resume"]),id:a}).strict(),e.object({operation:e.literal("list")}).strict()]),W=e.object({taskId:a,sourceCaseHash:d,reviewSessionId:a,reviewTurnId:a,inputHarness:i,trigger:i,decision:e.enum(["no_action","route_deterministically","queue_refiner"]),outcome:i.nullable(),proposal:i.nullable(),validations:e.array(i).max(100),grade:e.object({id:a,contentHash:d,score:e.number().min(0).max(1).nullable(),passed:e.boolean(),failureClass:e.string().nullable()}).strict(),resetComplete:e.boolean()}).strict(),$=e.object({schemaVersion:e.literal("openpond.advancedRefinerReviewReceipt.v1"),id:a,runId:a,pinHash:d,ownerActorId:a,teamId:a,privateClosureHash:d,refinerRelease:i,cases:e.array(W).min(1).max(1e4),reviewScore:e.number().min(0).max(1).nullable(),completedCases:e.number().int().nonnegative(),noActionCases:e.number().int().nonnegative(),routedCases:e.number().int().nonnegative(),proposalCases:e.number().int().nonnegative(),costUsd:e.number().finite().nonnegative().nullable(),cleanupComplete:e.boolean(),createdAt:e.string().datetime(),contentHash:d}).strict();function de(r){return $.parse({...r,contentHash:s(r)})}m(de,"sealAdvancedRefinerReviewReceipt");var P=S.shape.profiles.element,me=S.extend({profiles:e.array(P.extend({components:e.array(P.shape.components.element.extend({adaptationSplits:e.array(e.object({split:e.enum(["train","validation","test","frozen_eval"]),taskIds:e.array(a)})),environmentHash:d,resettable:e.boolean(),reviewQuality:e.boolean()}))})),refinerReleases:e.array(i),activeRefiner:i.nullable(),models:e.array(e.object({id:a,profileId:a,name:e.string(),location:e.enum(["local","hosted"])}))}),L=e.object({operation:e.literal("prepare"),evidence:i,profileOptionId:a,componentLabel:a,mode:e.enum(["review_quality","downstream_adaptation"]),adaptationSplit:e.enum(["train","validation","test","frozen_eval"]),operationId:a,refinerRelease:i,maximumCostUsd:e.number().positive(),maximumDurationMs:e.number().int().positive(),maximumModelSteps:e.number().int().positive()}).strict(),le=e.union([V,L,e.object({operation:e.literal("options"),evidence:i}).strict()]),pe=e.object({modelRequests:e.number().int().nonnegative(),knownSpendUsd:e.number().nonnegative(),totalSpendUsd:e.number().nonnegative().nullable(),heldUsd:e.number().nonnegative(),uncertainRequests:e.number().int().nonnegative()}).strict();export{C as a,x as b,D as c,X as d,Y as e,T as f,Z as g,S as h,U as i,ee as j,_ as k,M as l,ce as m,B as n,V as o,W as p,$ as q,de as r,L as s,pe as t};

package/dist/chunks/{chunk-OXQ2WFIH.js → chunk-TR5UC5AF.js}

@@ -1,5 +1,5 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as As,b as Ms}from"./chunk-CO23RRN7.js";import{a as Es,e as $e,f as Cs,g as Tt}from"./chunk-3Z4VM2LF.js";import{a as xs,c as Is}from"./chunk-XVDFJ3ZF.js";import{h as ne}from"./chunk-RCQYEHHD.js";import{a as ee,b as _s}from"./chunk-SDHG6CKU.js";import{b as ys,c as bs,d as yt}from"./chunk-GYYJ47GX.js";import{a as ws,b as vs,c as Ts,d as qe,e as Ss,g as vt}from"./chunk-WCNPGKST.js";import{a as wt,c as Rs}from"./chunk-AFH462DO.js";import{b as Kr,e as Xr,f as lt,g as ut}from"./chunk-CIZI6Z7D.js";import{F as gt,G as fs,H as ht,I as kt,J as gs,K as hs,N as ks,x as ps}from"./chunk-MOXIGLN2.js";import{a as ve}from"./chunk-7KWD2I6W.js";import{g as pt,i as ft}from"./chunk-2ACX7IW5.js";import{s as br}from"./chunk-OFJPOTES.js";import{d as Ne}from"./chunk-U7D6Q744.js";import{ja as rs,qa as ss,sa as os,x as ts}from"./chunk-HG2YFMVU.js";import{M as V,N as as,O as ns,P as is,Q as mt,R as ds,S as q,T as cs,U as je,V as ls,X as us,Y as Vt,Z as ms}from"./chunk-7OKZCJHS.js";import{l as Ft}from"./chunk-7NCM6GBN.js";import{d as Zr,f as W,g as Qr,i as es}from"./chunk-JT5YQ52I.js";import{k as Nr}from"./chunk-WGQ3FMRF.js";import{c as ct}from"./chunk-JVMSN7LH.js";import{Ba as Gr,Ca as Re,Ha as Yr,Ia as De,K as qr,L as $r,M as Lr,N as Ur,P as Br,aa as Wr,ba as Fr,ia as Vr,ua as zr,wa as we,xa as Jr}from"./chunk-BU6MW65V.js";import{x as ae}from"./chunk-DXJC3QW7.js";import{C as Or,S as Hr,T as Dr,Tb as jr}from"./chunk-D6L6UNUQ.js";import{f as Rr}from"./chunk-AJXWJLZA.js";import{L,M as Ar,Ob as Pr,R as Wt,S as oe,X as dt,Y as Mr,ba as He,d as Er,g as Cr,pb as ke,qb as ye}from"./chunk-4PUOWSDE.js";import{bb as be,cb as Ir,ea as it,v as Ge}from"./chunk-ZY32SZ2P.js";import{l as _r,m as xr}from"./chunk-OTDDXEQ6.js";import{i as Tr,m as se,n as k}from"./chunk-RCT3O3HQ.js";import{m as Sr}from"./chunk-UGOK7VTC.js";import{A as P}from"./chunk-X2J2P5ZT.js";import{a as d}from"./chunk-FTQLQJQV.js";var zt={buildIntent:"verifiable_reward",methodHint:"grpo"};async function ii(e){let t=await aa(e.store,e.workspaceId,e.reviewRef),r=`harness-review-${k({review:t.contentHash,strategy:zt}).slice(0,24)}`,s=(await e.store.listTaskCreationSnapshots(e.profileId)).find(c=>c.request.candidateId===r);if(s)return s;let[a,o,n]=await Promise.all([e.store.listHarnessImprovementArtifacts(e.workspaceId,"route_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"trigger_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"observation",1e3)]),i=new Map(a.map(c=>[_e(c),c])),u=new Map(o.map(c=>[_e(c),c])),f=new Map(n.map(c=>[_e(c),c])),p=new Map;for(let c of t.selectedEvidence){if(c.kind!=="route_decision")continue;let m=i.get(_e(c.evidence)),g=m?u.get(_e(m.trigger)):null;if(!m||!g)throw new Error(`Harness review evidence ${c.evidence.id} cannot be reconstructed.`);let h=g.observations.map(y=>f.get(_e(y))??null).filter(y=>y?.runRef===c.sourceRef);if(!h.length)throw new Error(`Harness review source ${c.sourceRef} has no exact observation evidence.`);let w=p.get(c.sourceRef)??{turnIds:new Set,evidenceRefs:[],sourcePolicies:[]};for(let y of h)w.turnIds.add(y.turnId);w.evidenceRefs.push(c.evidence),w.sourcePolicies.push(c.sourcePolicy),p.set(c.sourceRef,w)}if(!p.size)throw new Error("Harness Taskset review has no reconstructable source runs.");let l=[];for(let[c,m]of[...p.entries()].sort(([g],[h])=>g.localeCompare(h))){let g=await e.taskCreator.addSessionSource({profileId:e.profileId,sessionId:c,turnIds:[...m.turnIds].sort(),consentScope:"selected_turns"}),h=Array.isArray(g.metadata.harnessEvaluationReviews)?g.metadata.harnessEvaluationReviews:[],w={...g,metadata:{...g.metadata,harnessEvaluationReviews:Gt([...h,{id:t.id,contentHash:t.contentHash,workspaceId:t.workspaceRef,harnessRelease:t.harnessRelease,claimFingerprint:t.claim?.fingerprint??null,evidenceRefs:m.evidenceRefs,sourcePolicies:m.sourcePolicies}])}};l.push(await e.store.upsertTrainingSource(w))}return e.startCreation({profileId:e.profileId,sourceIds:l.map(c=>c.id),surface:"task_candidate",mode:"customize",entryMode:"automated",resourceIntent:"workproduct",buildIntent:zt.buildIntent,buildSpecification:null,objective:t.claim?.statement??t.reason,methodHint:zt.methodHint,preferredBaseModelId:null,preferredBaseModel:null,candidateId:r,analysisModel:e.analysisModel??null,analysisReasoningEffort:e.analysisReasoningEffort??null,createImproveRunId:null,targetIntent:{kind:"model",id:null,displayName:null,operation:"create"}})}d(ii,"startHarnessReviewTasksetAuthoring");function di(e){let t=e.flatMap(o=>Array.isArray(o.metadata.harnessEvaluationReviews)?o.metadata.harnessEvaluationReviews.filter(Jt):[]);if(!t.length)return null;let r=`${String(t[0].id)}:${String(t[0].contentHash)}`,s=t.filter(o=>`${String(o.id)}:${String(o.contentHash)}`===r);if(s.length!==e.length)return null;let a=s[0];return{review:{id:String(a.id),contentHash:String(a.contentHash),workspaceId:String(a.workspaceId),harnessRelease:a.harnessRelease,claimFingerprint:typeof a.claimFingerprint=="string"?a.claimFingerprint:null},evidenceRefs:Gt(s.flatMap(o=>Array.isArray(o.evidenceRefs)?o.evidenceRefs:[])),sourcePolicies:Gt(s.flatMap(o=>Array.isArray(o.sourcePolicies)?o.sourcePolicies:[]))}}d(di,"harnessReviewLineageFromSources");async function Ps(e){let t=e.taskset.metadata.harnessEvaluationLineage;if(!Jt(t)||!Jt(t.review))return null;let r=String(t.review.workspaceId??""),s={id:String(t.review.id??""),contentHash:String(t.review.contentHash??"")};if(!r||!s.id||!s.contentHash)throw new Error("Taskset Harness Evaluation lineage is incomplete.");let a=await e.store.listHarnessImprovementArtifacts(r,"evaluation_review",1e3),o=a.find(c=>_e(c)===_e(s));if(!o)throw new Error("Originating Harness Evaluation review was not found.");let n=e.evaluationResult?{id:e.evaluationResult.id,contentHash:e.evaluationResult.contentHash}:null,i={id:e.taskset.id,contentHash:e.taskset.contentHash},u=k({originRef:s,tasksetRef:i,evaluationRef:n}),f=a.find(c=>c.metadata.downstreamLinkKey===u);if(f)return f;let p=(e.now??(()=>new Date().toISOString()))(),l=Rr({schemaVersion:"openpond.harnessEvaluationReviewReceipt.v1",id:`evaluation-review-link-${u.slice(0,24)}`,ownerScope:o.ownerScope,workspaceRef:o.workspaceRef,harnessRelease:o.harnessRelease,previousWatermark:o.nextWatermark,nextWatermark:{cursor:k({cursor:o.nextWatermark.cursor,tasksetRef:i,evaluationRef:n}),throughCreatedAt:p},selectedEvidence:o.selectedEvidence,excludedEvidence:o.excludedEvidence,claim:o.claim,classification:"taskset",triage:o.triage,reason:n?"The approved Taskset completed a real baseline Evaluation.":"The approved Harness review materialized one immutable Taskset.",nextAuthority:"human_review",maxEstimatedCostUsd:o.maxEstimatedCostUsd,tasksetProposal:i,evaluation:n,trainingQualification:null,policyVersion:o.policyVersion,createdAt:p,metadata:{originReview:s,downstreamLinkKey:u,tasksetRevision:e.taskset.revision}});return await e.store.saveHarnessImprovementArtifact(r,"evaluation_review",l),l}d(Ps,"linkHarnessReviewTaskset");async function aa(e,t,r){let s=await e.getHarnessWorkspace(t);if(!s)throw new Error(`Harness workspace ${t} was not found.`);let a=(await e.listHarnessImprovementArtifacts(t,"evaluation_review",1e3)).find(n=>n.id===r.id&&n.contentHash===r.contentHash);if(!a||a.schemaVersion!=="openpond.harnessEvaluationReviewReceipt.v1")throw new Error("Harness Evaluation review was not found.");let o=a;if(o.classification!=="taskset"||o.nextAuthority!=="human_review")throw new Error("Only a human-review Taskset recommendation can enter Taskset authoring.");if(o.tasksetProposal)throw new Error("Harness Evaluation review already has a materialized Taskset link.");if(o.ownerScope.kind!==s.ownerScope.kind||o.ownerScope.id!==s.ownerScope.id)throw new Error("Harness Evaluation review owner scope does not match its workspace.");return o}d(aa,"requireTasksetReview");function _e(e){return`${e.id}:${e.contentHash}`}d(_e,"artifactKey");function Jt(e){return!!e&&typeof e=="object"&&!Array.isArray(e)}d(Jt,"isRecord");function Gt(e){let t=new Set;return e.filter(r=>{let s=JSON.stringify(r);return t.has(s)?!1:(t.add(s),!0)})}d(Gt,"uniqueObjects");import{randomUUID as na}from"node:crypto";function bt(e){let t=St(St(St(e.metadata).billing).pricing);return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}d(bt,"hostedTokenPricingFromCatalog");function mi(e){let t=St(e);try{return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}catch{return null}}d(mi,"hostedTokenPricingFromValue");function xe(e,t){let r=ve(e);if(r.promptTokens===null&&r.completionTokens===null)return null;let s=r.cachedPromptTokens??0;return((r.uncachedPromptTokens??r.promptTokens??0)*t.inputUsdPerMillionTokens+s*t.cachedInputUsdPerMillionTokens+(r.completionTokens??0)*t.outputUsdPerMillionTokens)/1e6}d(xe,"hostedUsageCostUsd");function St(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(St,"record");function Le(e,t){if(typeof e=="string"&&e.trim())return e.trim();throw new Error(`Hosted model catalog is missing ${t}.`)}d(Le,"requiredString");function Ue(e,t){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return e;throw new Error(`Hosted model catalog is missing ${t}.`)}d(Ue,"requiredRate");function Ds(e={}){let t=e.catalog??pt,r=e.stream??ft;return{async prepare(s){if(s.providerId!=="openpond")throw new Error("The selected judge provider is not connected to the hosted judge transport.");if(s.revision!==null)throw new Error("This judge transport cannot enforce a model revision; select a provider with revision support.");let a=await t();if(a.error)throw new Error(`Judge model catalog unavailable: ${a.error}`);let o=a.models.find(l=>l.id===s.modelId);if(!o)throw new Error("The selected judge model is unavailable.");let n=Os(o.raw),i=bt(n),u=Hs(n.context_window,"context window"),f=Math.min(4096,Hs(n.output_limit,"output limit")),p=(u*Math.max(i.inputUsdPerMillionTokens,i.cachedInputUsdPerMillionTokens)+f*i.outputUsdPerMillionTokens)/1e6;if(!(p>0))throw new Error("The judge model must have a positive known charge ceiling.");return{maximumChargeUsd:p,async dispatch(l){l?.throwIfAborted();let c="",m=null,g=null,h=null,w=!1;for await(let _ of r({model:s.modelId,messages:[{role:"system",content:s.system},{role:"user",content:s.data}],requestId:`learning-judge-${na()}`,maxTokens:f,temperature:s.temperature,signal:l})){let x=Os(_.raw);if(typeof x.model=="string"){if(m!==null&&x.model!==m)throw new Error("Judge response model identity changed during streaming.");m=x.model}if(typeof x.id=="string"){if(g!==null&&g!==x.id)throw new Error("Judge response identity changed during streaming.");g=x.id}if(_.type==="text_delta"&&(c+=_.text),c.length>1e5)throw new Error("Judge response exceeds the retained response limit.");_.type==="usage"&&(h=_.usage),_.type==="finish"&&(w=!0)}if(!w||!m||!g)throw new Error("Judge response did not include a complete model receipt.");let y=ve(h),S=y.promptTokens!==null&&y.completionTokens!==null?xe(h,i):null;return{text:c,modelId:m,modelRevision:null,responseId:g,inputTokens:y.promptTokens,outputTokens:y.completionTokens,costUsd:S}}}},async cancel({run:s}){return!s.judgeCalls?.some(a=>a.status==="reserved")}}}d(Ds,"createLearningHostedJudgeProvider");function Os(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Os,"record");function Hs(e,t){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new Error(`Judge catalog is missing its ${t}.`);return e}d(Hs,"positiveInteger");function xi(e,t=Ds()){return{execute(r){if(r.definition.execution.environment.entrypoint==="openpond.javascript-environment.v1")throw new Error("Tool Tasksets require grading an owner-recorded environment attempt.");let s=Pr({repository:e,scope:r.scope,run:r.run});return Ir({binding:r.binding,rewards:r.rewards,task:dt(r.evidence,r.definition),evidence:Mr(r.evidence,r.run.output),evaluatorContext:r.evidence.submission.evaluatorContext,signal:r.signal,modelJudge:Er({readRubric:d(a=>e.transaction(r.scope,async o=>oe(await ke(o,"asset",a.id,1),a)),"readRubric"),async executeBudgeted(a,o){let n=await t.prepare(a,{scope:r.scope,run:r.run});return Cr({store:s,maximumCharge:d(()=>n.maximumChargeUsd,"maximumCharge"),dispatch:d((u,f)=>n.dispatch(f),"dispatch")})(`grade-${k(a)}`,a,o)}}),customVerifier:d(async({grader:a,task:o,evidence:n})=>{let i=await Ye({repository:e,scope:r.scope,grader:a,task:o,evidence:n,evaluatorContext:r.evidence.submission.evaluatorContext,timeoutMs:r.run.timeoutMs,signal:r.signal});return{score:i.score,passed:i.passed,rewardEligible:a.rewardEligible,failureClass:null,feedback:[i.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:i.evidenceRefs}},"customVerifier")})},async cancel({scope:r,run:s}){let a=await e.transaction(r,o=>ke(o,"grade",s.id));return a.judgeCalls?.some(o=>o.status==="reserved")?!1:t.cancel({scope:r,run:a})}}}d(xi,"createLocalTaskGradeExecutor");async function Ye(e){let t=await e.repository.transaction(e.scope,async r=>oe(await ke(r,"asset",e.grader.verifierRef.id,1),e.grader.verifierRef));return $e({source:t,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:Math.min(e.grader.timeoutMs,e.timeoutMs??e.grader.timeoutMs),signal:e.signal,value:{task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,expectedOutput:e.task.expectedOutput,evaluatorContext:e.evaluatorContext,infrastructureError:e.evidence.infrastructureError??null}})}d(Ye,"executeLocalLearningVerifier");import ia from"node:path";async function js(e,t,r){if(!r)throw new Error("Ordinary tool execution requires its owner's package directory.");let s=De({taskset:e,adapterId:qe(e)}),a=await vt({root:ia.join(r,"training","tasksets",ee(e)),content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:s.tasksetRelease,environment:s.environmentRelease,verifierSet:s.verifierSetRelease},sources:ct.parse(e.metadata.portableFileInventory)}),o=Fr(a);if(!o)throw new Error("The Taskset does not declare a JavaScript execution graph.");let n=a.taskset.tasks.find(f=>f.id===t.id);if(!n)throw new Error("The selected task is missing from its executable package.");let i=o.assets.find(f=>f.id===o.execution.javascript.module.id),u=o.assets.find(f=>f.id===n.privilegedContextRef);return{execution:o.execution,module:i,initialState:JSON.parse(u.text),definition:L(a.taskset),instructions:Vr(a),task:n}}d(js,"loadOrdinaryToolEnvironment");var J="openpond.javascript-environment.v1";async function Be(e,t,r,s){let a=t.tasks.find(o=>o.id===r.id);if(!a||k(a)!==k(r))throw new Error("Tool task differs from the selected immutable Taskset.");return t.metadata.taskDefinition===void 0?js(t,r,s):e.learningRepository().transaction(t.profileId,async o=>{let n=await ye(o,"definition",ae.parse(t.metadata.taskDefinition)),i=n.execution.environment;if(i.entrypoint!==J||t.environment.entrypoint!==J||t.environment.kind!=="agent"||t.environment.stateful!==i.stateful||t.environment.defaultTimeoutMs!==i.defaultTimeoutMs||t.environment.networkPolicy!==i.networkPolicy||t.environment.deterministicSeeds!==i.deterministicSeeds||k(t.environment.toolNames)!==k(n.execution.tools.map(v=>v.name))||k(t.environment.metadata.portableTools)!==k(n.execution.tools)||k(t.policy)!==k(n.execution.policy)||k(t.metadata.rewardBinding)!==k(n.rewardBinding))throw new Error("Tool Taskset context differs from its pinned task definition.");let u=await ye(o,"binding",n.rewardBinding),f=await Promise.all(u.sources.map(v=>ye(o,"reward",v.reward))),p=it.parse({id:r.id,clusterKey:r.clusterKey,split:r.split,input:r.input,expectedOutput:r.expectedOutput,policyVisibleContext:r.policyVisibleContext,privilegedContextRef:r.privilegedContextRef,artifactRefs:[],tags:r.tags}),l={...n.execution,graders:be(u,f),tasks:[p]},c=await ke(o,"asset",Ur(l),1),m=$r.parse(JSON.parse(oe(c,c.asset))),g=[m.javascript.module,m.environment.actionSchemaRef,m.environment.observationSchemaRef,m.environment.stateSchemaRef].filter(v=>v!==null),h=[...new Set([...g.map(v=>v.id),r.privilegedContextRef].filter(v=>v!==null))],w=await Promise.all(h.map(v=>ke(o,"asset",v,1))),y=Br(l,c,w);if(k(t.environment.metadata.portableExecutionResources)!==k({environment:y.environment,verifierSet:y.verifierSet}))throw new Error("Tool Taskset publication differs from its private execution resources.");let S=w.find(v=>v.id===y.javascript.module.id),_=w.find(v=>v.id===r.privilegedContextRef),x=JSON.parse(oe(_,_.asset));return{execution:y,module:S,initialState:x,definition:L(n),instructions:n.instructions,task:p}})}d(Be,"loadStarterToolEnvironment");function Ns(e){let t=Wt({text:e.instructions,path:"policy/instructions.txt",mediaType:"text/plain",visibility:"policy"}),r=Wt({text:JSON.stringify({environment:L(e.execution.environment),verifierSet:L(e.execution.verifierSet),javascript:L(e.execution.javascript)}),path:"environment/dependencies.json",mediaType:"application/json",visibility:"host_private"}),s=_r({schemaVersion:"openpond.agentSnapshot.v2",id:`starter-agent-${e.definition.contentHash}`,sourceRelease:{id:e.definition.id,contentHash:e.definition.contentHash},instructions:[t.asset],skills:[],agents:[],toolDeclarations:e.execution.javascript.tools,capabilityRequirements:[],dependencyLock:r.asset,portability:{portable:!0,blockers:[],localOnlyAssetRefs:[],hostPrivateAssetRefs:[e.module.id,r.id]},metadata:{execution:J}}),a=xr({schemaVersion:"openpond.harnessRelease.v2",id:`starter-harness-${s.contentHash}`,agentSnapshot:{id:s.id,contentHash:s.contentHash},program:e.module.asset,tools:e.execution.javascript.tools,lifecycle:{create:!0,reset:!0,step:!0,collect:!0,destroy:!0,resetScope:"attempt"},graderInterface:{visibleEvidence:["output"],privilegedEvidence:["environment_state","private_verifier"],privateVerifierIsolation:!0},files:[t.asset,r.asset,e.module.asset],metadata:{execution:J}});return{agentSnapshot:s,harnessRelease:a,instructionContext:void 0}}d(Ns,"compileStarterToolHarness");import Rt from"node:path";import{constants as qs}from"node:fs";import{open as da,realpath as $s}from"node:fs/promises";var ca=P.string().regex(/^[a-f0-9]{64}$/),la=P.object({schemaVersion:P.literal("openpond.starterToolAttempt.v1"),requestId:P.string(),taskset:ae,taskHash:ca,model:P.unknown(),seed:P.number().int(),attempt:P.number().int(),definition:ae,policySource:P.enum(["model","fixture"]),environment:ae,verifierSet:ae,result:qr}).strict();async function _t(e){let{store:t,taskset:r,task:s,attempt:a}=e,o=await Be(t,r,s,e.storeDir),i=(await t.listTaskAttemptArtifacts({attemptId:a.id})).filter(S=>S.kind==="environment_state"&&S.tasksetId===r.id&&S.taskId===s.id&&S.attemptId===a.id&&a.artifactRefs.includes(S.id)&&a.privilegedOutcomeRef===S.id);if(i.length!==1)throw new Error("Tool grading requires exactly one owner-recorded environment artifact.");let u=i[0],f=await $s(Rt.join(e.storeDir,"training","evaluation-artifacts")),p=await $s(Rt.dirname(u.path)),l=Rt.relative(f,p);if(!l||l.startsWith("..")||Rt.isAbsolute(l))throw new Error("Tool evidence is outside the execution owner's artifact directory.");let c=await da(u.path,qs.O_RDONLY|qs.O_NOFOLLOW),m;try{let S=await c.stat();if(!S.isFile()||S.size>33554432||S.size!==u.sizeBytes)throw new Error("Tool evidence exceeds its declared byte boundary.");m=await c.readFile()}finally{await c.close()}if(m.byteLength!==u.sizeBytes||se(m)!==u.sha256)throw new Error("Owner-recorded tool evidence bytes changed.");let g=la.parse(JSON.parse(m.toString("utf8"))),{contentHash:h,...w}=g.result,y=g.result;if(k(w)!==h||!y.snapshot||!y.environmentCleanupComplete||!["completed","budget_exhausted","policy_failure"].includes(y.status)||k(g.taskset)!==k(L(r))||g.taskHash!==k(s)||k(g.model)!==k(a.modelRef)||g.seed!==a.seed||g.attempt!==a.attempt||a.tasksetId!==r.id||a.taskId!==s.id||a.split!==s.split||k(g.definition)!==k(o.definition)||k(g.environment)!==k(L(o.execution.environment))||k(g.verifierSet)!==k(L(o.execution.verifierSet))||y.taskId!==s.id||y.snapshot.seed!==a.seed||y.snapshot.inputHash!==k(s.input)||k(y.snapshot.definition)!==k(L(o.execution.javascript))||y.snapshot.finalStateHash!==k(y.snapshot.state)||k(a.output)!==k({text:y.output??""})||a.metadata.environmentAttemptHash!==h)throw new Error("Tool evidence differs from the selected task, model or recorded attempt.");if(a.metadata.policySource!==g.policySource)throw new Error("Tool evidence policy source changed.");return{environment:{status:y.status,collected:y.collected,definition:y.snapshot.definition,initialStateHash:y.snapshot.initialStateHash,finalStateHash:y.snapshot.finalStateHash,finalState:y.snapshot.state,events:y.snapshot.events}}}d(_t,"readStarterToolEvidence");async function Te(e,t){if(t.metadata.learning!==void 0)return He.parse(t.metadata.learning);if(t.metadata.rewardBinding===void 0)return;let r=ae.parse(t.metadata.rewardBinding);return e.learningRepository().transaction(t.profileId,async s=>{let a=await ye(s,"binding",r),o=[...new Map(a.sources.map(i=>[`${i.reward.id}:${i.reward.revision}:${i.reward.contentHash}`,i.reward])).values()],n=await Promise.all(o.map(i=>ye(s,"reward",i)));return{binding:a,rewards:n}})}d(Te,"resolveTasksetRewardBinding");async function ua(e,t,r){let s=await Te(e,t);if(!s)return{rewardExecution:void 0,verifierAssets:[]};let a=[...s.rewards.flatMap(i=>i.assets),...be(s.binding,s.rewards).flatMap(i=>i.kind==="custom_verifier"?[i.verifierRef]:i.kind==="model_judge"?[i.rubricRef]:[])],o=t.metadata.learning===void 0?void 0:await wt(r,t);if(o){let i=[...new Map(a.map(u=>{let f=o.learningResources.assets.find(p=>p.id===u.id);if(!f)throw new Error("Imported batch verifier is missing from its package.");return oe(f,u),[f.id,f]})).values()];return{rewardExecution:s,verifierAssets:i}}let n=await e.learningRepository().transaction(t.profileId,async i=>{let u=new Map;for(let f of a){let p=await ke(i,"asset",f.id,1);oe(p,f),u.set(p.id,p)}return[...u.values()]});return{rewardExecution:s,verifierAssets:n}}d(ua,"resolveTasksetTrainingReward");async function gd(e,t,r){let s=await ua(e,t,r.storeDir);if(!s.rewardExecution)return s;let a=be(s.rewardExecution.binding,s.rewardExecution.rewards);if(!s.rewardExecution.binding.sources.some(o=>o.role==="evaluation"&&o.weight>0))throw new Error("Managed training requires a positively weighted evaluation Reward source for retained validation.");if(a.some(o=>o.kind==="custom_verifier"&&o.runtime==="sandbox_process"))throw new Error("Managed preparation requires a qualified isolated sandbox for process verifiers.");if(r.placement!=="remote"||t.environment.kind==="work"||t.capabilities.requiresState||t.capabilities.requiresTools||a.some(o=>o.kind==="human"||o.kind==="model_judge"&&(!o.model||o.calibrationStatus!=="passed"))||t.tasks.some(o=>o.privilegedContextRef!==null)||!s.rewardExecution.binding.sources.some(o=>o.role==="training"&&o.weight>0)||r.hasLearnedPreferenceReward)throw new Error("This Reward binding requires an additional managed execution adapter.");return s}d(gd,"resolveManagedTasksetReward");async function Ls(e,t,r){let s=await Te(e,t);if(!s)throw new Error("The Taskset has no published Reward binding.");let a=be(s.binding,s.rewards);return async({grader:o,task:n,attempt:i,signal:u})=>{let f=t.tasks.find(m=>m.id===n.id);if(!f||k(f)!==k(n))throw new Error("Task differs from the selected immutable Taskset.");let p=null;if(t.environment.entrypoint===J){if(!r)throw new Error("Tool grading requires the execution owner's artifact directory.");p=await _t({store:e,storeDir:r,taskset:t,task:n,attempt:i})}else if(n.privilegedContextRef!==null)throw new Error("This authored Taskset requires a private evaluator-context resolver.");let l=a.find(m=>m.id===o.id);if(l?.kind!=="custom_verifier")throw new Error("The verifier is absent from the published Reward binding.");let c=it.parse({id:n.id,clusterKey:n.clusterKey,split:n.split,input:n.input,expectedOutput:n.expectedOutput,policyVisibleContext:n.policyVisibleContext,privilegedContextRef:n.privilegedContextRef,artifactRefs:[],tags:n.tags});return Ye({repository:e.learningRepository(),scope:t.profileId,grader:l,task:c,evidence:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},evaluatorContext:p,signal:u})}}d(Ls,"createTasksetBindingVerifier");async function Ie(e,t,r){let s=r?ae.parse(r):null;if(s&&s.id!==t)throw new Error("Taskset execution reference differs from the requested Taskset.");let a=s?await e.getTasksetRevision(s.id,s.revision):await e.getTaskset(t);if(!a||s&&(a.contentHash!==s.contentHash||Re(a)!==s.contentHash))throw new Error("Taskset execution revision is unavailable or differs from its immutable hash.");return a}d(Ie,"requireLocalTasksetRevision");async function Sd(e,t,r){let s=await e.releaseForTaskset(t);return s||De({taskset:t,rewardExecution:r?await Te(r,t):void 0,adapterId:qe(t)}).tasksetRelease}d(Sd,"requireReleasedTaskset");import{createHash as ma}from"node:crypto";function Us(e,t,r){if(t.metadata.importedPackageHash!==e.contentHash)throw new Error("The Work projection differs from the admitted published package.");let s=e.taskset.tasks.find(n=>n.id===r.id);if(!s||s.split!==r.split||k(r.metadata.portableTaskRecord)!==k(s))throw new Error("The Work task differs from its exact released population.");let a=0,o=(r.assets??[]).map(n=>{let i=s.artifactRefs.find(l=>l.id===n.id&&l.contentHash===n.sha256&&l.path===n.artifactRef),u=i&&e.files.find(l=>k(l.asset)===k(i));if(!i||!u||i.visibility!=="policy")throw new Error("The Work task requests an unavailable or private evaluator file.");let f=Buffer.from(Wr(u));if(ma("sha256").update(f).digest("hex")!==n.sha256||f.length!==n.sizeBytes||n.split!==r.split)throw new Error("The published Work input manifest changed.");if(a+=f.length,a>25e7)throw new Error("Published Work inputs exceed the admitted 250 MB ceiling.");return{...n,assetId:n.id,storageName:n.stagingPath??n.fileName,bytes:f,sizeBytes:f.length}});if(o.length!==s.artifactRefs.length)throw new Error("The exact policy-visible Work asset population is incomplete.");return o}d(Us,"advancedPublishedWorkAssets");async function Yt(e,t=pt,r=!1){if(e.kind==="hosted_harness")return Yt({kind:"hosted_chat",modelId:e.modelId,...Nr(e)},t);if(e.kind!=="hosted_chat")throw new ne("local_target_not_qualified","Select an explicit model target for local execution.",422);if(e.harness&&!r)throw new ne("local_harness_not_qualified","This released Harness requires its native turn executor; it cannot run as a model-only case.",422);let s=await t();if(s.error)throw new ne("local_model_catalog_unavailable",s.error,422);let a=s.models.find(c=>c.id===e.modelId);if(!a)throw new ne("local_model_unavailable","The selected model is unavailable.",422);let o=Bs(a.raw),n=bt(o),i=Bs(o.capabilities).samplingParameters===!0;if(!i&&(e.temperature!==0||e.topP!==1))throw new ne("local_sampling_unsupported","This model uses provider-managed sampling. Custom temperature or top P requires an explicit supported capability.",422);let u=Ws(o.context_window),f=Ws(o.output_limit);if(e.maxOutputTokens>f)throw new ne("local_model_output_limit","The model cannot accept the configured output limit.",422);let p={providerId:"openpond",modelId:e.modelId,maxOutputTokens:e.maxOutputTokens,...i?{temperature:e.temperature,topP:e.topP}:{},messages:e.messages??[]},l=(u*Math.max(n.inputUsdPerMillionTokens,n.cachedInputUsdPerMillionTokens)+e.maxOutputTokens*n.outputUsdPerMillionTokens)/1e6;if(!(l>0))throw new ne("local_model_charge_unknown","The model must publish an enforceable positive request ceiling.",422);return{model:{...p,configurationHash:k({configuration:p,contextWindow:u,outputLimit:f,pricing:n})},pricing:n,maximumChargeUsd:l}}d(Yt,"prepareLocalExperimentModel");function Nd(e){let t=e.admission.streamFactory?.(e)??e.admission.stream??e.stream??ft;return async function*(r){if(!r.requestId)throw new ne("local_request_identity_missing","Every local model dispatch requires a stable request identity.");let s=e.admission.model;if(r.model!==s.modelId||r.maxTokens!==s.maxOutputTokens||r.temperature!==s.temperature||r.topP!==s.topP)throw new ne("local_model_configuration_conflict","The native model request changed its admitted model, output limit or sampling configuration.",422);let a=xs();a?.assertExecution(e.executionId);let o={teamId:e.teamId,id:e.executionId,ownerId:e.ownerId,requestId:r.requestId};await e.store.reserveLocalCharge({...o,caseId:e.caseId,maximumUsd:e.admission.maximumChargeUsd}),r.signal?.aborted&&(await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),r.signal.throwIfAborted());try{await e.store.markLocalChargeDispatched(o.teamId,o.id,o.requestId,o.ownerId)}catch(p){throw a&&await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),p}await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:"model.dispatch",payload:{requestId:o.requestId,model:r.model,messages:r.messages}});let n=null,i=!1,u=!1,f;try{for await(let p of t(r)){u=!0,p.type==="usage"&&(n=p.usage),p.type==="finish"&&(i=!0);let l=p.type==="text_delta"?{requestId:o.requestId,text:p.text}:p.type==="tool_call_delta"?{requestId:o.requestId,toolCalls:p.toolCalls}:p.type==="usage"?{requestId:o.requestId,usage:ve(p.usage)}:p.type==="finish"?{requestId:o.requestId,finishReason:p.finishReason}:null;l&&await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:`model.${p.type}`,payload:l}),yield p}}catch(p){throw!u&&p instanceof Is&&p.executionId===o.id&&p.requestId===o.requestId&&(f=p),p}finally{let p=ve(n),l=e.admission.pricing&&i&&p.promptTokens!==null&&p.completionTokens!==null?xe(n,e.admission.pricing):null;await e.store.settleLocalCharge({...o,costUsd:l,confirmedTransportNotInvoked:f,usage:n===null?null:{promptTokens:p.promptTokens,completionTokens:p.completionTokens,cachedPromptTokens:p.cachedPromptTokens,uncachedPromptTokens:p.uncachedPromptTokens}})}}}d(Nd,"createLocalBudgetedModelStream");function Bs(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Bs,"record");function Ws(e){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new ne("local_model_bound_unknown","The model catalog is missing an enforceable context/output bound.",422);return e}d(Ws,"positive");import{AsyncLocalStorage as pa}from"node:async_hooks";import fa from"node:path";var Fs=new pa;function Vs(){return Fs.getStore()}d(Vs,"currentAdvancedRefinerBoundary");function Jd(e){let t=Sr(fa.join(e.storeDir,"library","harnesses","advanced-refiner-paid-calls.sqlite"));t.exec(`CREATE TABLE IF NOT EXISTS advanced_refiner_budgets(id TEXT PRIMARY KEY,actor_id TEXT NOT NULL,team_id TEXT NOT NULL,pin_hash TEXT NOT NULL,maximum_spend REAL NOT NULL,maximum_steps INTEGER NOT NULL,expires_at TEXT NOT NULL);
+import{a as As,b as Ms}from"./chunk-TD5YYA5Z.js";import{a as Es,e as $e,f as Cs,g as Tt}from"./chunk-BUMCJHQP.js";import{a as xs,c as Is}from"./chunk-XVDFJ3ZF.js";import{h as ne}from"./chunk-5TTQCT3G.js";import{a as ee,b as _s}from"./chunk-NDNSLCPE.js";import{b as ys,c as bs,d as yt}from"./chunk-CWE2N3JI.js";import{a as ws,b as vs,c as Ts,d as qe,e as Ss,g as vt}from"./chunk-O4YHISU2.js";import{a as wt,c as Rs}from"./chunk-HT6ILPD6.js";import{b as Kr,e as Xr,f as lt,g as ut}from"./chunk-VMW23TIZ.js";import{F as gt,G as fs,H as ht,I as kt,J as gs,K as hs,N as ks,x as ps}from"./chunk-JALL5NPH.js";import{a as ve}from"./chunk-7KWD2I6W.js";import{g as pt,i as ft}from"./chunk-2ACX7IW5.js";import{s as br}from"./chunk-OFJPOTES.js";import{d as Ne}from"./chunk-L5L3H6CI.js";import{ja as rs,qa as ss,sa as os,x as ts}from"./chunk-2THZ6GZY.js";import{M as V,N as as,O as ns,P as is,Q as mt,R as ds,S as q,T as cs,U as je,V as ls,X as us,Y as Vt,Z as ms}from"./chunk-OXRN2257.js";import{l as Ft}from"./chunk-TOWY6Y23.js";import{d as Zr,f as W,g as Qr,i as es}from"./chunk-Z6HWSTEP.js";import{k as Nr}from"./chunk-XKDOPNQ5.js";import{c as ct}from"./chunk-GR4MJJCE.js";import{Ba as Gr,Ca as Re,Ha as Yr,Ia as De,K as qr,L as $r,M as Lr,N as Ur,P as Br,aa as Wr,ba as Fr,ia as Vr,ua as zr,wa as we,xa as Jr}from"./chunk-BYVE4TMU.js";import{x as ae}from"./chunk-3VT2PPKW.js";import{C as Or,S as Hr,T as Dr,Tb as jr}from"./chunk-D6L6UNUQ.js";import{f as Rr}from"./chunk-AJXWJLZA.js";import{L,M as Ar,Ob as Pr,R as Wt,S as oe,X as dt,Y as Mr,ba as He,d as Er,g as Cr,pb as ke,qb as ye}from"./chunk-ZEOYNUHL.js";import{bb as be,cb as Ir,ea as it,v as Ge}from"./chunk-ZY32SZ2P.js";import{l as _r,m as xr}from"./chunk-OTDDXEQ6.js";import{i as Tr,m as se,n as k}from"./chunk-RCT3O3HQ.js";import{m as Sr}from"./chunk-UGOK7VTC.js";import{A as P}from"./chunk-X2J2P5ZT.js";import{a as d}from"./chunk-FTQLQJQV.js";var zt={buildIntent:"verifiable_reward",methodHint:"grpo"};async function ii(e){let t=await aa(e.store,e.workspaceId,e.reviewRef),r=`harness-review-${k({review:t.contentHash,strategy:zt}).slice(0,24)}`,s=(await e.store.listTaskCreationSnapshots(e.profileId)).find(c=>c.request.candidateId===r);if(s)return s;let[a,o,n]=await Promise.all([e.store.listHarnessImprovementArtifacts(e.workspaceId,"route_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"trigger_decision",1e3),e.store.listHarnessImprovementArtifacts(e.workspaceId,"observation",1e3)]),i=new Map(a.map(c=>[_e(c),c])),u=new Map(o.map(c=>[_e(c),c])),f=new Map(n.map(c=>[_e(c),c])),p=new Map;for(let c of t.selectedEvidence){if(c.kind!=="route_decision")continue;let m=i.get(_e(c.evidence)),g=m?u.get(_e(m.trigger)):null;if(!m||!g)throw new Error(`Harness review evidence ${c.evidence.id} cannot be reconstructed.`);let h=g.observations.map(y=>f.get(_e(y))??null).filter(y=>y?.runRef===c.sourceRef);if(!h.length)throw new Error(`Harness review source ${c.sourceRef} has no exact observation evidence.`);let w=p.get(c.sourceRef)??{turnIds:new Set,evidenceRefs:[],sourcePolicies:[]};for(let y of h)w.turnIds.add(y.turnId);w.evidenceRefs.push(c.evidence),w.sourcePolicies.push(c.sourcePolicy),p.set(c.sourceRef,w)}if(!p.size)throw new Error("Harness Taskset review has no reconstructable source runs.");let l=[];for(let[c,m]of[...p.entries()].sort(([g],[h])=>g.localeCompare(h))){let g=await e.taskCreator.addSessionSource({profileId:e.profileId,sessionId:c,turnIds:[...m.turnIds].sort(),consentScope:"selected_turns"}),h=Array.isArray(g.metadata.harnessEvaluationReviews)?g.metadata.harnessEvaluationReviews:[],w={...g,metadata:{...g.metadata,harnessEvaluationReviews:Gt([...h,{id:t.id,contentHash:t.contentHash,workspaceId:t.workspaceRef,harnessRelease:t.harnessRelease,claimFingerprint:t.claim?.fingerprint??null,evidenceRefs:m.evidenceRefs,sourcePolicies:m.sourcePolicies}])}};l.push(await e.store.upsertTrainingSource(w))}return e.startCreation({profileId:e.profileId,sourceIds:l.map(c=>c.id),surface:"task_candidate",mode:"customize",entryMode:"automated",resourceIntent:"workproduct",buildIntent:zt.buildIntent,buildSpecification:null,objective:t.claim?.statement??t.reason,methodHint:zt.methodHint,preferredBaseModelId:null,preferredBaseModel:null,candidateId:r,analysisModel:e.analysisModel??null,analysisReasoningEffort:e.analysisReasoningEffort??null,createImproveRunId:null,targetIntent:{kind:"model",id:null,displayName:null,operation:"create"}})}d(ii,"startHarnessReviewTasksetAuthoring");function di(e){let t=e.flatMap(o=>Array.isArray(o.metadata.harnessEvaluationReviews)?o.metadata.harnessEvaluationReviews.filter(Jt):[]);if(!t.length)return null;let r=`${String(t[0].id)}:${String(t[0].contentHash)}`,s=t.filter(o=>`${String(o.id)}:${String(o.contentHash)}`===r);if(s.length!==e.length)return null;let a=s[0];return{review:{id:String(a.id),contentHash:String(a.contentHash),workspaceId:String(a.workspaceId),harnessRelease:a.harnessRelease,claimFingerprint:typeof a.claimFingerprint=="string"?a.claimFingerprint:null},evidenceRefs:Gt(s.flatMap(o=>Array.isArray(o.evidenceRefs)?o.evidenceRefs:[])),sourcePolicies:Gt(s.flatMap(o=>Array.isArray(o.sourcePolicies)?o.sourcePolicies:[]))}}d(di,"harnessReviewLineageFromSources");async function Ps(e){let t=e.taskset.metadata.harnessEvaluationLineage;if(!Jt(t)||!Jt(t.review))return null;let r=String(t.review.workspaceId??""),s={id:String(t.review.id??""),contentHash:String(t.review.contentHash??"")};if(!r||!s.id||!s.contentHash)throw new Error("Taskset Harness Evaluation lineage is incomplete.");let a=await e.store.listHarnessImprovementArtifacts(r,"evaluation_review",1e3),o=a.find(c=>_e(c)===_e(s));if(!o)throw new Error("Originating Harness Evaluation review was not found.");let n=e.evaluationResult?{id:e.evaluationResult.id,contentHash:e.evaluationResult.contentHash}:null,i={id:e.taskset.id,contentHash:e.taskset.contentHash},u=k({originRef:s,tasksetRef:i,evaluationRef:n}),f=a.find(c=>c.metadata.downstreamLinkKey===u);if(f)return f;let p=(e.now??(()=>new Date().toISOString()))(),l=Rr({schemaVersion:"openpond.harnessEvaluationReviewReceipt.v1",id:`evaluation-review-link-${u.slice(0,24)}`,ownerScope:o.ownerScope,workspaceRef:o.workspaceRef,harnessRelease:o.harnessRelease,previousWatermark:o.nextWatermark,nextWatermark:{cursor:k({cursor:o.nextWatermark.cursor,tasksetRef:i,evaluationRef:n}),throughCreatedAt:p},selectedEvidence:o.selectedEvidence,excludedEvidence:o.excludedEvidence,claim:o.claim,classification:"taskset",triage:o.triage,reason:n?"The approved Taskset completed a real baseline Evaluation.":"The approved Harness review materialized one immutable Taskset.",nextAuthority:"human_review",maxEstimatedCostUsd:o.maxEstimatedCostUsd,tasksetProposal:i,evaluation:n,trainingQualification:null,policyVersion:o.policyVersion,createdAt:p,metadata:{originReview:s,downstreamLinkKey:u,tasksetRevision:e.taskset.revision}});return await e.store.saveHarnessImprovementArtifact(r,"evaluation_review",l),l}d(Ps,"linkHarnessReviewTaskset");async function aa(e,t,r){let s=await e.getHarnessWorkspace(t);if(!s)throw new Error(`Harness workspace ${t} was not found.`);let a=(await e.listHarnessImprovementArtifacts(t,"evaluation_review",1e3)).find(n=>n.id===r.id&&n.contentHash===r.contentHash);if(!a||a.schemaVersion!=="openpond.harnessEvaluationReviewReceipt.v1")throw new Error("Harness Evaluation review was not found.");let o=a;if(o.classification!=="taskset"||o.nextAuthority!=="human_review")throw new Error("Only a human-review Taskset recommendation can enter Taskset authoring.");if(o.tasksetProposal)throw new Error("Harness Evaluation review already has a materialized Taskset link.");if(o.ownerScope.kind!==s.ownerScope.kind||o.ownerScope.id!==s.ownerScope.id)throw new Error("Harness Evaluation review owner scope does not match its workspace.");return o}d(aa,"requireTasksetReview");function _e(e){return`${e.id}:${e.contentHash}`}d(_e,"artifactKey");function Jt(e){return!!e&&typeof e=="object"&&!Array.isArray(e)}d(Jt,"isRecord");function Gt(e){let t=new Set;return e.filter(r=>{let s=JSON.stringify(r);return t.has(s)?!1:(t.add(s),!0)})}d(Gt,"uniqueObjects");import{randomUUID as na}from"node:crypto";function bt(e){let t=St(St(St(e.metadata).billing).pricing);return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}d(bt,"hostedTokenPricingFromCatalog");function mi(e){let t=St(e);try{return{version:Le(t.version,"pricing version"),source:Le(t.source,"pricing source"),effectiveAt:Le(t.effectiveAt,"pricing effective date"),inputUsdPerMillionTokens:Ue(t.inputUsdPerMillionTokens,"input token rate"),cachedInputUsdPerMillionTokens:Ue(t.cachedInputUsdPerMillionTokens,"cached input token rate"),outputUsdPerMillionTokens:Ue(t.outputUsdPerMillionTokens,"output token rate")}}catch{return null}}d(mi,"hostedTokenPricingFromValue");function xe(e,t){let r=ve(e);if(r.promptTokens===null&&r.completionTokens===null)return null;let s=r.cachedPromptTokens??0;return((r.uncachedPromptTokens??r.promptTokens??0)*t.inputUsdPerMillionTokens+s*t.cachedInputUsdPerMillionTokens+(r.completionTokens??0)*t.outputUsdPerMillionTokens)/1e6}d(xe,"hostedUsageCostUsd");function St(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(St,"record");function Le(e,t){if(typeof e=="string"&&e.trim())return e.trim();throw new Error(`Hosted model catalog is missing ${t}.`)}d(Le,"requiredString");function Ue(e,t){if(typeof e=="number"&&Number.isFinite(e)&&e>=0)return e;throw new Error(`Hosted model catalog is missing ${t}.`)}d(Ue,"requiredRate");function Ds(e={}){let t=e.catalog??pt,r=e.stream??ft;return{async prepare(s){if(s.providerId!=="openpond")throw new Error("The selected judge provider is not connected to the hosted judge transport.");if(s.revision!==null)throw new Error("This judge transport cannot enforce a model revision; select a provider with revision support.");let a=await t();if(a.error)throw new Error(`Judge model catalog unavailable: ${a.error}`);let o=a.models.find(l=>l.id===s.modelId);if(!o)throw new Error("The selected judge model is unavailable.");let n=Os(o.raw),i=bt(n),u=Hs(n.context_window,"context window"),f=Math.min(4096,Hs(n.output_limit,"output limit")),p=(u*Math.max(i.inputUsdPerMillionTokens,i.cachedInputUsdPerMillionTokens)+f*i.outputUsdPerMillionTokens)/1e6;if(!(p>0))throw new Error("The judge model must have a positive known charge ceiling.");return{maximumChargeUsd:p,async dispatch(l){l?.throwIfAborted();let c="",m=null,g=null,h=null,w=!1;for await(let _ of r({model:s.modelId,messages:[{role:"system",content:s.system},{role:"user",content:s.data}],requestId:`learning-judge-${na()}`,maxTokens:f,temperature:s.temperature,signal:l})){let x=Os(_.raw);if(typeof x.model=="string"){if(m!==null&&x.model!==m)throw new Error("Judge response model identity changed during streaming.");m=x.model}if(typeof x.id=="string"){if(g!==null&&g!==x.id)throw new Error("Judge response identity changed during streaming.");g=x.id}if(_.type==="text_delta"&&(c+=_.text),c.length>1e5)throw new Error("Judge response exceeds the retained response limit.");_.type==="usage"&&(h=_.usage),_.type==="finish"&&(w=!0)}if(!w||!m||!g)throw new Error("Judge response did not include a complete model receipt.");let y=ve(h),S=y.promptTokens!==null&&y.completionTokens!==null?xe(h,i):null;return{text:c,modelId:m,modelRevision:null,responseId:g,inputTokens:y.promptTokens,outputTokens:y.completionTokens,costUsd:S}}}},async cancel({run:s}){return!s.judgeCalls?.some(a=>a.status==="reserved")}}}d(Ds,"createLearningHostedJudgeProvider");function Os(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Os,"record");function Hs(e,t){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new Error(`Judge catalog is missing its ${t}.`);return e}d(Hs,"positiveInteger");function xi(e,t=Ds()){return{execute(r){if(r.definition.execution.environment.entrypoint==="openpond.javascript-environment.v1")throw new Error("Tool Tasksets require grading an owner-recorded environment attempt.");let s=Pr({repository:e,scope:r.scope,run:r.run});return Ir({binding:r.binding,rewards:r.rewards,task:dt(r.evidence,r.definition),evidence:Mr(r.evidence,r.run.output),evaluatorContext:r.evidence.submission.evaluatorContext,signal:r.signal,modelJudge:Er({readRubric:d(a=>e.transaction(r.scope,async o=>oe(await ke(o,"asset",a.id,1),a)),"readRubric"),async executeBudgeted(a,o){let n=await t.prepare(a,{scope:r.scope,run:r.run});return Cr({store:s,maximumCharge:d(()=>n.maximumChargeUsd,"maximumCharge"),dispatch:d((u,f)=>n.dispatch(f),"dispatch")})(`grade-${k(a)}`,a,o)}}),customVerifier:d(async({grader:a,task:o,evidence:n})=>{let i=await Ye({repository:e,scope:r.scope,grader:a,task:o,evidence:n,evaluatorContext:r.evidence.submission.evaluatorContext,timeoutMs:r.run.timeoutMs,signal:r.signal});return{score:i.score,passed:i.passed,rewardEligible:a.rewardEligible,failureClass:null,feedback:[i.feedback],visibleEvidenceRefs:[],privilegedEvidenceRefs:i.evidenceRefs}},"customVerifier")})},async cancel({scope:r,run:s}){let a=await e.transaction(r,o=>ke(o,"grade",s.id));return a.judgeCalls?.some(o=>o.status==="reserved")?!1:t.cancel({scope:r,run:a})}}}d(xi,"createLocalTaskGradeExecutor");async function Ye(e){let t=await e.repository.transaction(e.scope,async r=>oe(await ke(r,"asset",e.grader.verifierRef.id,1),e.grader.verifierRef));return $e({source:t,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:Math.min(e.grader.timeoutMs,e.timeoutMs??e.grader.timeoutMs),signal:e.signal,value:{task:e.task,attempt:e.evidence,input:e.task.input,output:e.evidence.output,expectedOutput:e.task.expectedOutput,evaluatorContext:e.evaluatorContext,infrastructureError:e.evidence.infrastructureError??null}})}d(Ye,"executeLocalLearningVerifier");import ia from"node:path";async function js(e,t,r){if(!r)throw new Error("Ordinary tool execution requires its owner's package directory.");let s=De({taskset:e,adapterId:qe(e)}),a=await vt({root:ia.join(r,"training","tasksets",ee(e)),content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:s.tasksetRelease,environment:s.environmentRelease,verifierSet:s.verifierSetRelease},sources:ct.parse(e.metadata.portableFileInventory)}),o=Fr(a);if(!o)throw new Error("The Taskset does not declare a JavaScript execution graph.");let n=a.taskset.tasks.find(f=>f.id===t.id);if(!n)throw new Error("The selected task is missing from its executable package.");let i=o.assets.find(f=>f.id===o.execution.javascript.module.id),u=o.assets.find(f=>f.id===n.privilegedContextRef);return{execution:o.execution,module:i,initialState:JSON.parse(u.text),definition:L(a.taskset),instructions:Vr(a),task:n}}d(js,"loadOrdinaryToolEnvironment");var J="openpond.javascript-environment.v1";async function Be(e,t,r,s){let a=t.tasks.find(o=>o.id===r.id);if(!a||k(a)!==k(r))throw new Error("Tool task differs from the selected immutable Taskset.");return t.metadata.taskDefinition===void 0?js(t,r,s):e.learningRepository().transaction(t.profileId,async o=>{let n=await ye(o,"definition",ae.parse(t.metadata.taskDefinition)),i=n.execution.environment;if(i.entrypoint!==J||t.environment.entrypoint!==J||t.environment.kind!=="agent"||t.environment.stateful!==i.stateful||t.environment.defaultTimeoutMs!==i.defaultTimeoutMs||t.environment.networkPolicy!==i.networkPolicy||t.environment.deterministicSeeds!==i.deterministicSeeds||k(t.environment.toolNames)!==k(n.execution.tools.map(v=>v.name))||k(t.environment.metadata.portableTools)!==k(n.execution.tools)||k(t.policy)!==k(n.execution.policy)||k(t.metadata.rewardBinding)!==k(n.rewardBinding))throw new Error("Tool Taskset context differs from its pinned task definition.");let u=await ye(o,"binding",n.rewardBinding),f=await Promise.all(u.sources.map(v=>ye(o,"reward",v.reward))),p=it.parse({id:r.id,clusterKey:r.clusterKey,split:r.split,input:r.input,expectedOutput:r.expectedOutput,policyVisibleContext:r.policyVisibleContext,privilegedContextRef:r.privilegedContextRef,artifactRefs:[],tags:r.tags}),l={...n.execution,graders:be(u,f),tasks:[p]},c=await ke(o,"asset",Ur(l),1),m=$r.parse(JSON.parse(oe(c,c.asset))),g=[m.javascript.module,m.environment.actionSchemaRef,m.environment.observationSchemaRef,m.environment.stateSchemaRef].filter(v=>v!==null),h=[...new Set([...g.map(v=>v.id),r.privilegedContextRef].filter(v=>v!==null))],w=await Promise.all(h.map(v=>ke(o,"asset",v,1))),y=Br(l,c,w);if(k(t.environment.metadata.portableExecutionResources)!==k({environment:y.environment,verifierSet:y.verifierSet}))throw new Error("Tool Taskset publication differs from its private execution resources.");let S=w.find(v=>v.id===y.javascript.module.id),_=w.find(v=>v.id===r.privilegedContextRef),x=JSON.parse(oe(_,_.asset));return{execution:y,module:S,initialState:x,definition:L(n),instructions:n.instructions,task:p}})}d(Be,"loadStarterToolEnvironment");function Ns(e){let t=Wt({text:e.instructions,path:"policy/instructions.txt",mediaType:"text/plain",visibility:"policy"}),r=Wt({text:JSON.stringify({environment:L(e.execution.environment),verifierSet:L(e.execution.verifierSet),javascript:L(e.execution.javascript)}),path:"environment/dependencies.json",mediaType:"application/json",visibility:"host_private"}),s=_r({schemaVersion:"openpond.agentSnapshot.v2",id:`starter-agent-${e.definition.contentHash}`,sourceRelease:{id:e.definition.id,contentHash:e.definition.contentHash},instructions:[t.asset],skills:[],agents:[],toolDeclarations:e.execution.javascript.tools,capabilityRequirements:[],dependencyLock:r.asset,portability:{portable:!0,blockers:[],localOnlyAssetRefs:[],hostPrivateAssetRefs:[e.module.id,r.id]},metadata:{execution:J}}),a=xr({schemaVersion:"openpond.harnessRelease.v2",id:`starter-harness-${s.contentHash}`,agentSnapshot:{id:s.id,contentHash:s.contentHash},program:e.module.asset,tools:e.execution.javascript.tools,lifecycle:{create:!0,reset:!0,step:!0,collect:!0,destroy:!0,resetScope:"attempt"},graderInterface:{visibleEvidence:["output"],privilegedEvidence:["environment_state","private_verifier"],privateVerifierIsolation:!0},files:[t.asset,r.asset,e.module.asset],metadata:{execution:J}});return{agentSnapshot:s,harnessRelease:a,instructionContext:void 0}}d(Ns,"compileStarterToolHarness");import Rt from"node:path";import{constants as qs}from"node:fs";import{open as da,realpath as $s}from"node:fs/promises";var ca=P.string().regex(/^[a-f0-9]{64}$/),la=P.object({schemaVersion:P.literal("openpond.starterToolAttempt.v1"),requestId:P.string(),taskset:ae,taskHash:ca,model:P.unknown(),seed:P.number().int(),attempt:P.number().int(),definition:ae,policySource:P.enum(["model","fixture"]),environment:ae,verifierSet:ae,result:qr}).strict();async function _t(e){let{store:t,taskset:r,task:s,attempt:a}=e,o=await Be(t,r,s,e.storeDir),i=(await t.listTaskAttemptArtifacts({attemptId:a.id})).filter(S=>S.kind==="environment_state"&&S.tasksetId===r.id&&S.taskId===s.id&&S.attemptId===a.id&&a.artifactRefs.includes(S.id)&&a.privilegedOutcomeRef===S.id);if(i.length!==1)throw new Error("Tool grading requires exactly one owner-recorded environment artifact.");let u=i[0],f=await $s(Rt.join(e.storeDir,"training","evaluation-artifacts")),p=await $s(Rt.dirname(u.path)),l=Rt.relative(f,p);if(!l||l.startsWith("..")||Rt.isAbsolute(l))throw new Error("Tool evidence is outside the execution owner's artifact directory.");let c=await da(u.path,qs.O_RDONLY|qs.O_NOFOLLOW),m;try{let S=await c.stat();if(!S.isFile()||S.size>33554432||S.size!==u.sizeBytes)throw new Error("Tool evidence exceeds its declared byte boundary.");m=await c.readFile()}finally{await c.close()}if(m.byteLength!==u.sizeBytes||se(m)!==u.sha256)throw new Error("Owner-recorded tool evidence bytes changed.");let g=la.parse(JSON.parse(m.toString("utf8"))),{contentHash:h,...w}=g.result,y=g.result;if(k(w)!==h||!y.snapshot||!y.environmentCleanupComplete||!["completed","budget_exhausted","policy_failure"].includes(y.status)||k(g.taskset)!==k(L(r))||g.taskHash!==k(s)||k(g.model)!==k(a.modelRef)||g.seed!==a.seed||g.attempt!==a.attempt||a.tasksetId!==r.id||a.taskId!==s.id||a.split!==s.split||k(g.definition)!==k(o.definition)||k(g.environment)!==k(L(o.execution.environment))||k(g.verifierSet)!==k(L(o.execution.verifierSet))||y.taskId!==s.id||y.snapshot.seed!==a.seed||y.snapshot.inputHash!==k(s.input)||k(y.snapshot.definition)!==k(L(o.execution.javascript))||y.snapshot.finalStateHash!==k(y.snapshot.state)||k(a.output)!==k({text:y.output??""})||a.metadata.environmentAttemptHash!==h)throw new Error("Tool evidence differs from the selected task, model or recorded attempt.");if(a.metadata.policySource!==g.policySource)throw new Error("Tool evidence policy source changed.");return{environment:{status:y.status,collected:y.collected,definition:y.snapshot.definition,initialStateHash:y.snapshot.initialStateHash,finalStateHash:y.snapshot.finalStateHash,finalState:y.snapshot.state,events:y.snapshot.events}}}d(_t,"readStarterToolEvidence");async function Te(e,t){if(t.metadata.learning!==void 0)return He.parse(t.metadata.learning);if(t.metadata.rewardBinding===void 0)return;let r=ae.parse(t.metadata.rewardBinding);return e.learningRepository().transaction(t.profileId,async s=>{let a=await ye(s,"binding",r),o=[...new Map(a.sources.map(i=>[`${i.reward.id}:${i.reward.revision}:${i.reward.contentHash}`,i.reward])).values()],n=await Promise.all(o.map(i=>ye(s,"reward",i)));return{binding:a,rewards:n}})}d(Te,"resolveTasksetRewardBinding");async function ua(e,t,r){let s=await Te(e,t);if(!s)return{rewardExecution:void 0,verifierAssets:[]};let a=[...s.rewards.flatMap(i=>i.assets),...be(s.binding,s.rewards).flatMap(i=>i.kind==="custom_verifier"?[i.verifierRef]:i.kind==="model_judge"?[i.rubricRef]:[])],o=t.metadata.learning===void 0?void 0:await wt(r,t);if(o){let i=[...new Map(a.map(u=>{let f=o.learningResources.assets.find(p=>p.id===u.id);if(!f)throw new Error("Imported batch verifier is missing from its package.");return oe(f,u),[f.id,f]})).values()];return{rewardExecution:s,verifierAssets:i}}let n=await e.learningRepository().transaction(t.profileId,async i=>{let u=new Map;for(let f of a){let p=await ke(i,"asset",f.id,1);oe(p,f),u.set(p.id,p)}return[...u.values()]});return{rewardExecution:s,verifierAssets:n}}d(ua,"resolveTasksetTrainingReward");async function gd(e,t,r){let s=await ua(e,t,r.storeDir);if(!s.rewardExecution)return s;let a=be(s.rewardExecution.binding,s.rewardExecution.rewards);if(!s.rewardExecution.binding.sources.some(o=>o.role==="evaluation"&&o.weight>0))throw new Error("Managed training requires a positively weighted evaluation Reward source for retained validation.");if(a.some(o=>o.kind==="custom_verifier"&&o.runtime==="sandbox_process"))throw new Error("Managed preparation requires a qualified isolated sandbox for process verifiers.");if(r.placement!=="remote"||t.environment.kind==="work"||t.capabilities.requiresState||t.capabilities.requiresTools||a.some(o=>o.kind==="human"||o.kind==="model_judge"&&(!o.model||o.calibrationStatus!=="passed"))||t.tasks.some(o=>o.privilegedContextRef!==null)||!s.rewardExecution.binding.sources.some(o=>o.role==="training"&&o.weight>0)||r.hasLearnedPreferenceReward)throw new Error("This Reward binding requires an additional managed execution adapter.");return s}d(gd,"resolveManagedTasksetReward");async function Ls(e,t,r){let s=await Te(e,t);if(!s)throw new Error("The Taskset has no published Reward binding.");let a=be(s.binding,s.rewards);return async({grader:o,task:n,attempt:i,signal:u})=>{let f=t.tasks.find(m=>m.id===n.id);if(!f||k(f)!==k(n))throw new Error("Task differs from the selected immutable Taskset.");let p=null;if(t.environment.entrypoint===J){if(!r)throw new Error("Tool grading requires the execution owner's artifact directory.");p=await _t({store:e,storeDir:r,taskset:t,task:n,attempt:i})}else if(n.privilegedContextRef!==null)throw new Error("This authored Taskset requires a private evaluator-context resolver.");let l=a.find(m=>m.id===o.id);if(l?.kind!=="custom_verifier")throw new Error("The verifier is absent from the published Reward binding.");let c=it.parse({id:n.id,clusterKey:n.clusterKey,split:n.split,input:n.input,expectedOutput:n.expectedOutput,policyVisibleContext:n.policyVisibleContext,privilegedContextRef:n.privilegedContextRef,artifactRefs:[],tags:n.tags});return Ye({repository:e.learningRepository(),scope:t.profileId,grader:l,task:c,evidence:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},evaluatorContext:p,signal:u})}}d(Ls,"createTasksetBindingVerifier");async function Ie(e,t,r){let s=r?ae.parse(r):null;if(s&&s.id!==t)throw new Error("Taskset execution reference differs from the requested Taskset.");let a=s?await e.getTasksetRevision(s.id,s.revision):await e.getTaskset(t);if(!a||s&&(a.contentHash!==s.contentHash||Re(a)!==s.contentHash))throw new Error("Taskset execution revision is unavailable or differs from its immutable hash.");return a}d(Ie,"requireLocalTasksetRevision");async function Sd(e,t,r){let s=await e.releaseForTaskset(t);return s||De({taskset:t,rewardExecution:r?await Te(r,t):void 0,adapterId:qe(t)}).tasksetRelease}d(Sd,"requireReleasedTaskset");import{createHash as ma}from"node:crypto";function Us(e,t,r){if(t.metadata.importedPackageHash!==e.contentHash)throw new Error("The Work projection differs from the admitted published package.");let s=e.taskset.tasks.find(n=>n.id===r.id);if(!s||s.split!==r.split||k(r.metadata.portableTaskRecord)!==k(s))throw new Error("The Work task differs from its exact released population.");let a=0,o=(r.assets??[]).map(n=>{let i=s.artifactRefs.find(l=>l.id===n.id&&l.contentHash===n.sha256&&l.path===n.artifactRef),u=i&&e.files.find(l=>k(l.asset)===k(i));if(!i||!u||i.visibility!=="policy")throw new Error("The Work task requests an unavailable or private evaluator file.");let f=Buffer.from(Wr(u));if(ma("sha256").update(f).digest("hex")!==n.sha256||f.length!==n.sizeBytes||n.split!==r.split)throw new Error("The published Work input manifest changed.");if(a+=f.length,a>25e7)throw new Error("Published Work inputs exceed the admitted 250 MB ceiling.");return{...n,assetId:n.id,storageName:n.stagingPath??n.fileName,bytes:f,sizeBytes:f.length}});if(o.length!==s.artifactRefs.length)throw new Error("The exact policy-visible Work asset population is incomplete.");return o}d(Us,"advancedPublishedWorkAssets");async function Yt(e,t=pt,r=!1){if(e.kind==="hosted_harness")return Yt({kind:"hosted_chat",modelId:e.modelId,...Nr(e)},t);if(e.kind!=="hosted_chat")throw new ne("local_target_not_qualified","Select an explicit model target for local execution.",422);if(e.harness&&!r)throw new ne("local_harness_not_qualified","This released Harness requires its native turn executor; it cannot run as a model-only case.",422);let s=await t();if(s.error)throw new ne("local_model_catalog_unavailable",s.error,422);let a=s.models.find(c=>c.id===e.modelId);if(!a)throw new ne("local_model_unavailable","The selected model is unavailable.",422);let o=Bs(a.raw),n=bt(o),i=Bs(o.capabilities).samplingParameters===!0;if(!i&&(e.temperature!==0||e.topP!==1))throw new ne("local_sampling_unsupported","This model uses provider-managed sampling. Custom temperature or top P requires an explicit supported capability.",422);let u=Ws(o.context_window),f=Ws(o.output_limit);if(e.maxOutputTokens>f)throw new ne("local_model_output_limit","The model cannot accept the configured output limit.",422);let p={providerId:"openpond",modelId:e.modelId,maxOutputTokens:e.maxOutputTokens,...i?{temperature:e.temperature,topP:e.topP}:{},messages:e.messages??[]},l=(u*Math.max(n.inputUsdPerMillionTokens,n.cachedInputUsdPerMillionTokens)+e.maxOutputTokens*n.outputUsdPerMillionTokens)/1e6;if(!(l>0))throw new ne("local_model_charge_unknown","The model must publish an enforceable positive request ceiling.",422);return{model:{...p,configurationHash:k({configuration:p,contextWindow:u,outputLimit:f,pricing:n})},pricing:n,maximumChargeUsd:l}}d(Yt,"prepareLocalExperimentModel");function Nd(e){let t=e.admission.streamFactory?.(e)??e.admission.stream??e.stream??ft;return async function*(r){if(!r.requestId)throw new ne("local_request_identity_missing","Every local model dispatch requires a stable request identity.");let s=e.admission.model;if(r.model!==s.modelId||r.maxTokens!==s.maxOutputTokens||r.temperature!==s.temperature||r.topP!==s.topP)throw new ne("local_model_configuration_conflict","The native model request changed its admitted model, output limit or sampling configuration.",422);let a=xs();a?.assertExecution(e.executionId);let o={teamId:e.teamId,id:e.executionId,ownerId:e.ownerId,requestId:r.requestId};await e.store.reserveLocalCharge({...o,caseId:e.caseId,maximumUsd:e.admission.maximumChargeUsd}),r.signal?.aborted&&(await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),r.signal.throwIfAborted());try{await e.store.markLocalChargeDispatched(o.teamId,o.id,o.requestId,o.ownerId)}catch(p){throw a&&await e.store.settleLocalCharge({...o,costUsd:null,usage:null,notDispatched:!0}),p}await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:"model.dispatch",payload:{requestId:o.requestId,model:r.model,messages:r.messages}});let n=null,i=!1,u=!1,f;try{for await(let p of t(r)){u=!0,p.type==="usage"&&(n=p.usage),p.type==="finish"&&(i=!0);let l=p.type==="text_delta"?{requestId:o.requestId,text:p.text}:p.type==="tool_call_delta"?{requestId:o.requestId,toolCalls:p.toolCalls}:p.type==="usage"?{requestId:o.requestId,usage:ve(p.usage)}:p.type==="finish"?{requestId:o.requestId,finishReason:p.finishReason}:null;l&&await e.store.appendLocalExperimentEvent({...o,caseId:e.caseId,type:`model.${p.type}`,payload:l}),yield p}}catch(p){throw!u&&p instanceof Is&&p.executionId===o.id&&p.requestId===o.requestId&&(f=p),p}finally{let p=ve(n),l=e.admission.pricing&&i&&p.promptTokens!==null&&p.completionTokens!==null?xe(n,e.admission.pricing):null;await e.store.settleLocalCharge({...o,costUsd:l,confirmedTransportNotInvoked:f,usage:n===null?null:{promptTokens:p.promptTokens,completionTokens:p.completionTokens,cachedPromptTokens:p.cachedPromptTokens,uncachedPromptTokens:p.uncachedPromptTokens}})}}}d(Nd,"createLocalBudgetedModelStream");function Bs(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}d(Bs,"record");function Ws(e){if(typeof e!="number"||!Number.isSafeInteger(e)||e<=0)throw new ne("local_model_bound_unknown","The model catalog is missing an enforceable context/output bound.",422);return e}d(Ws,"positive");import{AsyncLocalStorage as pa}from"node:async_hooks";import fa from"node:path";var Fs=new pa;function Vs(){return Fs.getStore()}d(Vs,"currentAdvancedRefinerBoundary");function Jd(e){let t=Sr(fa.join(e.storeDir,"library","harnesses","advanced-refiner-paid-calls.sqlite"));t.exec(`CREATE TABLE IF NOT EXISTS advanced_refiner_budgets(id TEXT PRIMARY KEY,actor_id TEXT NOT NULL,team_id TEXT NOT NULL,pin_hash TEXT NOT NULL,maximum_spend REAL NOT NULL,maximum_steps INTEGER NOT NULL,expires_at TEXT NOT NULL);
  CREATE TABLE IF NOT EXISTS advanced_refiner_calls(run_id TEXT NOT NULL,id TEXT NOT NULL,intent_hash TEXT NOT NULL,ceiling REAL NOT NULL,state TEXT NOT NULL,actual_spend REAL,result TEXT,PRIMARY KEY(run_id,id));`);function r(a){t.exec("BEGIN IMMEDIATE");try{let o=a();return t.exec("COMMIT"),o}catch(o){throw t.exec("ROLLBACK"),o}}d(r,"transaction");async function s(a,o,n){let i=Ft(o);await e.authorize(i),r(()=>{let l=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(l){if(l.actor_id!==i.actorId||l.team_id!==i.teamId||l.pin_hash!==i.contentHash)throw new Error("Advanced Run budget belongs to another immutable admission.");return}t.prepare("INSERT INTO advanced_refiner_budgets VALUES(?,?,?,?,?,?,?)").run(a,i.actorId,i.teamId,i.contentHash,i.maximumCostUsd,i.maximumModelSteps,new Date(Date.now()+i.maximumDurationMs).toISOString())});let u=t.prepare("SELECT expires_at FROM advanced_refiner_budgets WHERE id=?").get(a),f=Date.parse(String(u.expires_at))-Date.now();if(f<=0)throw new Error("The immutable advanced evaluation duration is exhausted.");let p={runId:a,pin:i,signal:AbortSignal.timeout(f),authorize:d(()=>e.authorize(i),"authorize"),async resolveWorkAssets(l,c){await e.authorize(i),p.signal.throwIfAborted();let m=await e.packageForPin(i),g=Us(m,l,c);return await e.authorize(i),p.signal.throwIfAborted(),g},async call(l,c,m,g,h){if(await e.authorize(i),!Number.isInteger(g)||g<1)throw new Error("Advanced paid requests require a finite enforced token ceiling.");let w=c.model;if(w?.providerId!=="openpond"||!w.modelId)throw new Error("Advanced paid calls require an explicitly qualified hosted model.");let y=w.modelId,S=e.prepareModel?await e.prepareModel(y,g,m):await(async()=>{let C=await Yt({kind:"hosted_chat",modelId:y,maxOutputTokens:g,temperature:0,topP:1});if(k(C.pricing)!==k(m))throw new Error("The actual model pricing differs from its reviewed request pin.");return{configurationHash:C.model.configurationHash,maximumChargeUsd:C.maximumChargeUsd}})();if(!/^[a-f0-9]{64}$/.test(S.configurationHash)||!Number.isFinite(S.maximumChargeUsd)||S.maximumChargeUsd<=0)throw new Error("The actual owner model ceiling is unavailable.");let _=k({intent:c,pricing:m,maximumOutputTokens:g,modelConfigurationHash:S.configurationHash}),x=S.maximumChargeUsd,v=r(()=>{let C=t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=?").get(a);if(!C||String(C.expires_at)<=new Date().toISOString())throw new Error("The immutable advanced evaluation duration is exhausted.");let E=t.prepare("SELECT * FROM advanced_refiner_calls WHERE run_id=? AND id=?").get(a,l);if(E){if(E.intent_hash!==_)throw new Error("Advanced provider request ID retains another exact payload.");if(E.state!=="completed")throw new Error("This paid request has an unknown outcome; recover its actual provider receipt before resuming.");return JSON.parse(String(E.result))}let T=t.prepare("SELECT COUNT(*) AS steps,COALESCE(SUM(COALESCE(actual_spend,ceiling)),0) AS spent,SUM(CASE WHEN state NOT IN ('completed','not_dispatched') THEN 1 ELSE 0 END) AS unknown FROM advanced_refiner_calls WHERE run_id=?").get(a);if(Number(T.unknown)>0)throw new Error("A prior advanced paid request is unsettled; no fresh request is allowed.");if(Number(T.steps)>=i.maximumModelSteps||Number(T.spent)+x>i.maximumCostUsd+1e-9)throw new Error("The advanced evaluation step or cumulative spend ceiling is exhausted.");t.prepare("INSERT INTO advanced_refiner_calls(run_id,id,intent_hash,ceiling,state) VALUES(?,?,?,?,?)").run(a,l,_,x,"reserved")});if(v!==void 0)return await e.authorize(i),p.signal.throwIfAborted(),v;let M=!1;try{if(await e.authorize(i),p.signal.throwIfAborted(),t.prepare("UPDATE advanced_refiner_calls SET state='dispatched' WHERE run_id=? AND id=? AND intent_hash=? AND state='reserved'").run(a,l,_).changes!==1)throw new Error("The retained advanced request admission changed.");M=!0;let E=await h();if(E.costUsd===null||!Number.isFinite(E.costUsd)||E.costUsd<0||E.costUsd>x+1e-9)throw new Error("The actual provider receipt cannot settle its admitted request ceiling.");let T=JSON.stringify(E.value);if(Buffer.byteLength(T)>8*1024*1024)throw new Error("The advanced provider response exceeds the retained output bound.");return r(()=>{if(t.prepare("UPDATE advanced_refiner_calls SET state='completed',actual_spend=?,result=? WHERE run_id=? AND id=? AND intent_hash=? AND state='dispatched'").run(E.costUsd,T,a,l,_).changes!==1)throw new Error("The advanced request settlement changed concurrently.")}),await e.authorize(i),E.value}catch(C){throw M?t.prepare("UPDATE advanced_refiner_calls SET state='unknown' WHERE run_id=? AND id=? AND state='dispatched'").run(a,l):t.prepare("UPDATE advanced_refiner_calls SET state='not_dispatched',actual_spend=0 WHERE run_id=? AND id=? AND state='reserved'").run(a,l),C}}};return Fs.run(p,n)}return d(s,"run"),{run:s,async authorize(a){await e.authorize(Ft(a))},read(a,o){return t.prepare("SELECT * FROM advanced_refiner_budgets WHERE id=? AND actor_id=? AND team_id=? AND pin_hash=?").get(a,o.actorId,o.teamId,o.contentHash)?t.prepare("SELECT id,intent_hash,ceiling,state,actual_spend FROM advanced_refiner_calls WHERE run_id=? ORDER BY id").all(a):null},accounting(a,o){let n=this.read(a,o)??[],i=n.filter(f=>f.state!=="completed"&&f.state!=="not_dispatched"),u=n.reduce((f,p)=>f+Number(p.actual_spend??0),0);return{modelRequests:n.filter(f=>f.state!=="not_dispatched").length,knownSpendUsd:u,totalSpendUsd:i.length?null:u,heldUsd:i.reduce((f,p)=>f+Number(p.ceiling),0),uncertainRequests:i.filter(f=>f.state!=="reserved").length}},close(){t.close()}}}d(Jd,"createAdvancedRefinerPaidBoundary");import ra from"node:path";function xt(e,t,r){let s=new Date().toISOString();return W.parse({schemaVersion:"openpond.taskAttempt.v1",id:`fixture_attempt_${t.id}_${r}`,tasksetId:e,taskId:t.taskId,split:Ke(t)??"frozen_eval",attempt:r,seed:0,modelRef:null,startedAt:s,completedAt:s,output:t.output,runtimeEventRefs:[],artifactRefs:[],privilegedOutcomeRef:null,infrastructureError:t.infrastructureError,costUsd:0,latencyMs:0,userInterventions:0,metadata:{fixtureLabel:t.label}})}d(xt,"fixtureAttempt");function Ke(e){let t=e&&typeof e=="object"&&!Array.isArray(e)&&"metadata"in e&&e.metadata&&typeof e.metadata=="object"&&!Array.isArray(e.metadata)?e.metadata.artifactSplit:null;return t==="train"||t==="validation"||t==="test"||t==="frozen_eval"?t:null}d(Ke,"artifactSplit");function Js(e){let t=Gr(e.taskset),r=t.issues.filter(b=>b.severity==="error").map(b=>({code:b.code,message:b.message,path:b.path})),s=t.issues.filter(b=>b.severity==="warning").map(b=>({code:b.code,message:b.message,path:b.path})),a=typeof e.taskset.metadata.trainingMethod=="string"?e.taskset.metadata.trainingMethod:null,o=ga(a),n=o==="sft",i=o==="dpo",u=o==="grpo"||o==="ppo",f=new Set(e.taskset.learningSignals.demonstrations.filter(b=>b.approved&&b.taskId).map(b=>b.taskId)),p=e.taskset.metadata.learning===void 0?null:He.parse(e.taskset.metadata.learning),l=d(b=>p?p.admissions.some(B=>B.taskId===b.id&&B.supervisedTarget!==null):b.expectedOutput!=null,"hasSupervisedTarget"),c=e.taskset.datasetArtifact,m=new Set(zs(e.taskset.metadata.approvedArtifactSignals)),g=!!c?.schema.fields.some(b=>b.semanticRole==="demonstration"||b.semanticRole==="expected_output"||b.semanticRole==="chosen"),h=e.taskset.tasks.filter(b=>b.split==="train"&&(l(b)||u)),w=n?h.filter(b=>f.has(b.id)):h,y=e.taskset.tasks.filter(b=>b.split==="train"&&l(b)&&!f.has(b.id)),S=e.taskset.tasks.filter(b=>b.split==="frozen_eval"&&(b.expectedOutput||u)),_=c&&g?c.splitCounts.train??0:0,x=c&&g?c.splitCounts.frozen_eval??0:0,v=c?m.has(o??""):!1,M=c?v?_:0:w.length,C=e.taskset.learningSignals.preferences.filter(b=>b.approved).length,E=Yr(e.taskset),T=E?E.binding.sources.filter(b=>b.role==="training"&&b.weight>0&&(!b.taskIds||w.some(B=>b.taskIds.includes(B.id)))).length:e.taskset.learningSignals.rewards.filter(b=>b.approved&&b.executable).length,A=i?C:u?T?M:0:M,R=c?x:S.length;A===0&&r.push(n?{code:"sft_demonstrations_missing",message:"At least one explicitly approved training demonstration is required.",path:"learningSignals.demonstrations"}:i?{code:"dpo_preferences_missing",message:"At least one approved chosen/rejected response pair is required for DPO.",path:"learningSignals.preferences"}:{code:"online_reward_missing",message:`${o?.toUpperCase()??"Online training"} requires an executable reward-bearing task.`,path:"learningSignals.rewards"});let j=c?n&&!v?_:0:y.length;n&&j&&r.push({code:"sft_demonstrations_unapproved",message:`${j} training example${j===1?" is":"s are"} not explicitly approved.`,path:c?"metadata.approvedArtifactSignals":"learningSignals.demonstrations"}),R===0&&s.push({code:"frozen_eval_missing",message:"At least one independent evaluation example is recommended before making quality claims.",path:c?"datasetArtifact.splitCounts.frozen_eval":"tasks"});let $=new Set(w.map(b=>b.clusterKey)),z=new Set(S.map(b=>b.clusterKey)),U=c&&e.taskset.metadata.splitIsolationVerified===!0;(c?!U||A===0||R===0:$.size===0||z.size===0||[...$].some(b=>z.has(b)))&&s.push({code:"independent_evaluation_missing",message:"Independent source clusters are recommended before using evaluation results for quality claims.",path:c?"metadata.splitIsolationVerified":"tasks"}),(c?e.taskset.metadata.artifactRowsVerified!==!0:e.taskset.tasks.some(b=>typeof b.metadata.exampleOrigin!="string"))&&r.push({code:"example_provenance_missing",message:"Every training and evaluation example must record whether it was extracted, corrected, synthetic, or expert-authored.",path:c?"metadata.artifactRowsVerified":"tasks.metadata.exampleOrigin"});let Q=ha(e.taskset.metadata.diagnosis);(!Q||typeof Q.summary!="string"||!Array.isArray(Q.stableBehavior))&&s.push({code:"capability_diagnosis_missing",message:"A capability diagnosis is recommended to separate stable behavior from changing knowledge.",path:"metadata.diagnosis"}),Q?.trainingEligible===!1&&s.push({code:"training_not_recommended",message:"The capability diagnosis does not recommend storing this behavior in model weights.",path:"metadata.diagnosis.trainingEligible"});let me=e.taskset.graders.some(b=>b.rewardEligible&&(b.kind!=="model_judge"||b.metadata.calibrationIsAdvisory!==!0));me&&!e.graderAudit&&s.push({code:"grader_audit_missing",message:"Optional grader calibration evidence has not been recorded.",path:"graderFixtures"}),me&&e.graderAudit&&e.graderAudit.tasksetHash!==e.taskset.contentHash&&s.push({code:"grader_audit_stale",message:"Optional grader calibration evidence predates this immutable Taskset revision.",path:"graderFixtures"}),me&&e.graderAudit&&!e.graderAudit.passed&&s.push({code:"grader_audit_failed",message:"The grader did not match every optional calibration fixture's expected outcome.",path:"graderFixtures"}),me&&e.graderAudit&&!e.graderAudit.hackingChecksPassed&&s.push({code:"grader_adversarial_calibration_failed",message:"Adversarial or prompt-injection calibration fixtures did not match their expected grader outcomes.",path:"graders"}),me&&e.graderAudit&&!e.graderAudit.leakageChecksPassed&&r.push({code:"environment_leakage",message:"Environment or privileged-state leakage checks failed.",path:"environment"}),me&&e.graderAudit&&!e.graderAudit.infrastructureSafetyPassed&&r.push({code:"infrastructure_reward",message:"An infrastructure failure produced a score or eligible reward.",path:"graderFixtures"});let tt=e.taskset.graders.some(b=>b.rewardEligible);u&&!tt&&r.push({code:"online_reward_grader_missing",message:`${o?.toUpperCase()} requires a user-selected reward-eligible grader.`,path:"graders"});let ie=o??(A>0&&e.taskset.capabilities.compatibleMethods.includes("sft")?"sft":"none"),I=e.taskset.learningSignals.demonstrations.filter(b=>b.approved).map(b=>b.id),N=ie==="none"?null:{primaryMethod:ie,bootstrap:ie==="grpo"&&I.length?{method:"sft",purpose:"trajectory_bootstrap",demonstrationRefs:I,limitations:["The SFT bootstrap imitates approved trajectories; it does not optimize verifier reward.","Completing the bootstrap does not satisfy the primary GRPO recommendation."]}:null},K=r.length===0,pe=r.length>0,rt=["sft","dpo","grpo","ppo"].map(b=>{let B=[],X=[];b==="sft"&&I.length===0&&(B.push("demonstrations_missing"),X.push("Add at least one approved prompt/response demonstration.")),b==="dpo"&&C===0&&(B.push("preference_pairs_missing"),X.push("Add at least one approved chosen/rejected response pair.")),(b==="grpo"||b==="ppo")&&T===0&&(B.push("executable_reward_missing"),X.push("Implement and validate an executable scalar reward.")),b==="ppo"&&(B.push("value_model_required"),X.push("Choose and bind a value/critic model before PPO can execute.")),R===0&&(B.push("frozen_eval_missing"),X.push("Add an independent frozen-evaluation split.")),pe&&B.length===0&&(B.push("taskset_not_ready"),X.push("Resolve the Dataset validation and grader-audit blockers."));let Me=B.filter(fe=>fe!=="value_model_required");return{method:b,status:Me.length?"needs_dataset_work":b===ie?"recommended":"compatible",reasonCodes:B,reasons:X}});return zr.parse({schemaVersion:"openpond.tasksetReadiness.v1",tasksetId:e.taskset.id,tasksetHash:e.taskset.contentHash,ready:K,recommendedMethod:ie,trainingPath:N,methodReadiness:rt,compatibleDestinationClasses:K?ie==="grpo"?["export","custom","hosted_managed"]:["export","custom"]:["export"],blockers:r,advisories:s,warnings:[...t.issues.filter(b=>b.severity==="warning").map(b=>b.message),...ie==="grpo"?["GRPO is readiness-compatible but not executable in the local CPU fixture."]:[],...zs(e.taskset.metadata.warnings)],generatedAt:e.generatedAt??new Date().toISOString(),metadata:{reportHash:k([e.taskset.contentHash,r])}})}d(Js,"buildTasksetReadiness");function ga(e){return e==="sft"||e==="dpo"||e==="grpo"||e==="ppo"||e==="sdft"||e==="opsd"||e==="sdpo"?e:null}d(ga,"trainingPathMethod");function ha(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}d(ha,"metadataRecord");function zs(e){return Array.isArray(e)?e.filter(t=>typeof t=="string"&&!!t.trim()):[]}d(zs,"metadataStrings");import eo from"node:path";function Gs(e,t,r){let s=He.parse(t.metadata.learning),a=be(s.binding,s.rewards);return async({grader:o,task:n,attempt:i,signal:u})=>{let f=s.admissions.find(h=>h.taskId===n.id),p=a.find(h=>h.id===o.id);if(!f||p?.kind!=="custom_verifier")throw new Error("Prepared batch has no matching verifier admission.");let l=await wt(r,t),c=l?l.learningResources?.evidence.find(h=>Ar(L(h),f.evidence)):await e.learningRepository().transaction(t.profileId,h=>ye(h,"evidence",f.evidence));if(!c)throw new Error("Imported batch evidence is missing from its package.");let m=dt(c,s.definition),g=d(h=>({id:h.id,input:h.input,expectedOutput:h.expectedOutput,policyVisibleContext:h.policyVisibleContext,privilegedContextRef:h.privilegedContextRef}),"selected");if(k(g(n))!==k(g(m)))throw new Error("Prepared task differs from its admitted evaluator context.");if(l){let h=l.learningResources?.assets.find(w=>w.id===p.verifierRef.id);if(!h)throw new Error("Imported batch verifier is missing from its package.");return $e({source:oe(h,p.verifierRef),exportName:p.exportName,timeoutMs:p.timeoutMs,signal:u,value:{task:m,attempt:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},input:m.input,output:i.output,expectedOutput:m.expectedOutput,evaluatorContext:c.submission.evaluatorContext,infrastructureError:i.infrastructureError??null}})}return Ye({repository:e.learningRepository(),scope:t.profileId,grader:p,task:m,evidence:{output:i.output,artifactRefs:i.artifactRefs,runtimeEventRefs:i.runtimeEventRefs,infrastructureError:i.infrastructureError},evaluatorContext:c.submission.evaluatorContext,signal:u})}}d(Gs,"createLearningBatchVerifier");import{readFile as ka,realpath as Ys}from"node:fs/promises";import Ks from"node:path";async function Xs(e){let t=await Ys(e.allowedRoot),r=await Ys(Ks.resolve(t,e.grader.module));if(r!==t&&!r.startsWith(`${t}${Ks.sep}`))throw new Error("Verifier module is outside the approved Taskset root.");let s=await ka(r);if(e.grader.metadata.portableVerifierRef!==void 0){let o=Tr.parse(e.grader.metadata.portableVerifierRef);if(o.sizeBytes!==s.length||o.contentHash!==se(s))throw new Error("Verifier source differs from its immutable Taskset release.")}let a=new TextDecoder("utf-8",{fatal:!0}).decode(s);return $e({source:a,exportName:e.grader.exportName,runtime:e.grader.runtime,timeoutMs:e.grader.timeoutMs,signal:e.signal,value:{task:e.task,attempt:e.attempt,input:e.task.input,expectedOutput:e.task.expectedOutput,output:e.attempt.output,infrastructureError:e.attempt.infrastructureError,...e.evaluatorContext?{evaluatorContext:e.evaluatorContext}:{}}})}d(Xs,"runSandboxedVerifier");import{constants as Kt}from"node:fs";import{open as ya,realpath as Zs}from"node:fs/promises";import We from"node:path";async function It(e,t){let r=De({taskset:t,adapterId:qe(t)});if(typeof t.metadata.importedPackageHash=="string"){let s=await Rs(e,t.metadata.importedPackageHash);if(s.taskset.contentHash!==r.tasksetRelease.contentHash)throw new Error("Evaluation Taskset differs from its pinned portable package.");let a=ct.parse(t.metadata.portableFileInventory??[]),{files:o,contentHash:n,...i}=s,u=await vt({root:We.join(e,"training","tasksets",ee(t)),content:i,sources:o.map(f=>({asset:f.asset,sourcePath:a.find(p=>p.asset.id===f.asset.id)?.sourcePath??f.asset.path})),fileOrder:o.map(f=>f.asset.id)});if(u.contentHash!==s.contentHash)throw new Error("Evaluation package changed during capture.");return u}return _s({storeDir:e,taskset:t,content:{schemaVersion:"openpond.tasksetPackage.v1",taskset:r.tasksetRelease,environment:r.environmentRelease,verifierSet:r.verifierSetRelease}})}d(It,"captureTasksetEvaluationPackage");function Qs(e){return async({grader:t,task:r,attempt:s,signal:a})=>{let o=e.taskset.tasks.find(l=>l.id===r.id),n=e.taskset.graders.find(l=>l.id===t.id);if(!o||k(o)!==k(r)||!n||k(n)!==k(t)||s.tasksetId!==e.taskset.id||s.taskId!==r.id||s.split!==r.split)throw new Error("Process grading differs from the selected Taskset revision or attempt.");let i=e.packageValue.taskset.tasks.find(l=>l.id===r.id),u=e.packageValue.taskset.graders.find(l=>l.id===t.id);if(!i||u?.kind!=="custom_verifier"||u.runtime!=="sandbox_process")throw new Error("The process verifier is absent from the immutable package.");let f=e.fixtures?.get(s.id);if(f&&f.attemptHash!==k(s))throw new Error("Fixture attempt changed before private grading.");let p=f?.artifacts??await wa(e.store,e.storeDir,e.taskset,s);return Es({packageValue:e.packageValue,grader:u,task:i,evidence:{output:s.output,artifactRefs:s.artifactRefs,runtimeEventRefs:s.runtimeEventRefs,infrastructureError:s.infrastructureError,artifacts:p},signal:a,authorize:d(async()=>{a?.throwIfAborted()},"authorize")})}}d(Qs,"createTasksetProcessVerifier");async function wa(e,t,r,s){let a=await e.listTaskAttemptArtifacts({attemptId:s.id}),o=[],n=0;if(new Set(s.artifactRefs).size!==s.artifactRefs.length)throw new Error("Attempt contains duplicate artifact references.");for(let i of s.artifactRefs){let u=a.filter(g=>g.id===i&&g.tasksetId===r.id&&g.taskId===s.taskId&&g.attemptId===s.id);if(u.length!==1)throw new Error("Attempt artifact is not recorded by its execution owner.");let f=u[0];if(f.kind!=="output_artifact")continue;if(o.length>=16||(n+=f.sizeBytes)>2e7)throw new Error("Attempt artifacts exceed their byte boundary.");let p=await Zs(t),l=await Zs(We.dirname(f.path));if(![We.join(p,"training/evaluation-artifacts"),We.join(p,"work/outputs")].some(g=>l===g||l.startsWith(`${g}${We.sep}`)))throw new Error("Attempt artifact is outside the execution owner's output directories.");let c=await ya(f.path,Kt.O_RDONLY|Kt.O_NOFOLLOW|Kt.O_NONBLOCK),m;try{let g=await c.stat();if(!g.isFile()||g.size!==f.sizeBytes)throw new Error("Owner-recorded artifact size changed.");m=Buffer.alloc(g.size);let h=0;for(;h<m.length;){let w=await c.read(m,h,m.length-h,h);if(!w.bytesRead)throw new Error("Owner-recorded artifact changed during capture.");h+=w.bytesRead}if((await c.read(Buffer.alloc(1),0,1,h)).bytesRead)throw new Error("Owner-recorded artifact changed during capture.")}finally{await c.close()}if(se(m)!==f.sha256)throw new Error("Owner-recorded artifact bytes changed.");o.push({reference:{id:f.id,contentHash:f.sha256,mediaType:f.mediaType??"application/octet-stream",sizeBytes:f.sizeBytes},title:We.basename(typeof f.metadata.requiredOutputPath=="string"?f.metadata.requiredOutputPath:f.path),contentType:f.mediaType??"application/octet-stream",base64:m.toString("base64")})}return o}d(wa,"readOwnedOutputArtifacts");async function Et(e,t,r){if(!t.graders.some(p=>p.kind==="custom_verifier"))return;let s=t.graders.some(p=>p.kind==="custom_verifier"&&p.runtime==="sandbox_process"),a=t.metadata.learning!==void 0?await Gs(e.store,t,e.storeDir):t.metadata.rewardBinding!==void 0?await Ls(e.store,t,e.storeDir):void 0;if(a&&!s)return a;let o=e.storeDir?null:await(e.loadProfileState??br)(),n=e.storeDir?eo.join(e.storeDir,"training","tasksets",ee(t)):o?.sourcePath?eo.join(o.sourcePath,"tasksets",t.id):null,i=typeof t.metadata.creationSnapshotId=="string"?t.metadata.creationSnapshotId:null,u=i?await e.store.getTaskDesignProposal(i):null;if(n&&t.purpose!=="benchmark"&&t.environment.metadata.runtimeSourceTasksetId===void 0&&await ut(t,n,{generatedFiles:u?.generatedFiles??[]}),s&&!e.storeDir)throw new Error("Process grading requires the execution owner's package directory.");let f=s?Qs({store:e.store,storeDir:e.storeDir,taskset:t,packageValue:r?.packageValue??await It(e.storeDir,t),fixtures:r?.fixtures}):void 0;return n?async({grader:p,task:l,attempt:c,signal:m})=>{let g=t.graders.find(w=>w.id===p.id);if(!g||g.kind!=="custom_verifier")throw new Error("Verifier is not declared by this Taskset revision.");if(g.runtime==="sandbox_process")return f({grader:p,task:l,attempt:c,signal:m});if(a)return a({grader:g,task:l,attempt:c,signal:m});let h;if(t.environment.entrypoint===J&&!c.infrastructureError){if(!e.storeDir)throw new Error("Tool grading requires the execution owner's artifact directory.");h=await _t({store:e.store,storeDir:e.storeDir,taskset:t,task:l,attempt:c})}return Xs({grader:g,task:l,attempt:c,allowedRoot:n,signal:m,evaluatorContext:h})}:void 0}d(Et,"createTasksetEvaluationVerifier");var va=new Set(["correct","incorrect","parse_failure","budget_exhausted","tool_schema_violation","infrastructure_failure","cancelled"]);async function to(e){let t=Sa(e.attempt);if(t){let f=k(e.graders),p=`Taskset Work attempt ended with ${t.failureClass}: `+t.message;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${k([e.attempt.id,f,t]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:f,score:null,passed:!1,components:e.graders.map(l=>({graderId:l.id,graderVersion:l.version,score:0,passed:!1,hardGate:l.hardGate,rewardEligible:!1,feedback:p,evidenceRefs:e.attempt.artifactRefs,judge:l.kind==="model_judge"?l.judge:null,calibrationStatus:l.kind==="model_judge"?l.calibrationStatus:"not_applicable"})),failureClass:t.failureClass,feedback:[p],rewardEligible:!1,createdAt:e.now?.()??new Date().toISOString()}}let r=e.learning?null:ba(e.attempt.metadata);if(!r)try{return await lt(e)}catch(f){return Ta({attempt:e.attempt,graders:e.graders,now:e.now,error:f})}let s=k(e.graders),a=r.rewardEligible?r.outcome==="correct"?1:0:null,o=r.outcome==="correct",n=Ra(r.outcome),i=`Cross-System trajectory verifier outcome: ${r.outcome}.`,u=e.graders.map(f=>({graderId:f.id,graderVersion:f.version,score:a??0,passed:o,hardGate:f.hardGate,rewardEligible:r.rewardEligible&&f.rewardEligible,feedback:i,evidenceRefs:e.attempt.artifactRefs,judge:f.kind==="model_judge"?f.judge:null,calibrationStatus:f.kind==="model_judge"?f.calibrationStatus:"not_applicable"}));return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${k([e.attempt.id,s,r]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:s,score:a,passed:o,components:u,failureClass:n,feedback:[i],rewardEligible:r.rewardEligible,createdAt:e.now?.()??new Date().toISOString()}}d(to,"gradeTasksetEvaluationAttempt");function Ta(e){let t=k(e.graders),r=(e.error instanceof Error?e.error.message:String(e.error)).slice(0,19e3),s=`Taskset grader failed: ${r}`;return{schemaVersion:"openpond.gradeResult.v1",id:`grade_${k([e.attempt.id,t,"grader_failure",r]).slice(0,24)}`,attemptId:e.attempt.id,graderSetHash:t,score:null,passed:!1,components:e.graders.map(a=>({graderId:a.id,graderVersion:a.version,score:0,passed:!1,hardGate:a.hardGate,rewardEligible:!1,feedback:s,evidenceRefs:e.attempt.artifactRefs,judge:a.kind==="model_judge"?a.judge:null,calibrationStatus:a.kind==="model_judge"?a.calibrationStatus:"not_applicable"})),failureClass:"grader_failure",feedback:[s],rewardEligible:!1,createdAt:e.now?.()??new Date().toISOString()}}d(Ta,"graderFailureResult");function Sa(e){if(e.metadata.execution!=="taskset_work"||!e.infrastructureError)return null;let t=Zr.safeParse(e.metadata.failureClass);if(!t.success||t.data==="policy_failure"||t.data==="grader_failure")throw new Error("Taskset Work attempt is missing a trusted execution failure class.");return{failureClass:t.data,message:e.infrastructureError}}d(Sa,"tasksetWorkFailure");function ba(e){if(e.execution!=="post_training_evaluation_tool_loop")return null;let t=e.verifierOutcome,r=e.verifierReward,s=e.verifierRewardEligible;if(typeof t!="string"||!va.has(t)||typeof s!="boolean"||r!==null&&typeof r!="number")throw new Error("Cross-System evaluation attempt is missing trusted verifier metadata.");if(s!==(t==="correct"||t==="incorrect"||t==="parse_failure")||s&&typeof r!="number"||!s&&r!==null)throw new Error("Cross-System evaluation verifier metadata is internally inconsistent.");return{outcome:t,reward:r,rewardEligible:s}}d(ba,"crossSystemGrade");function Ra(e){return e==="correct"?null:e==="infrastructure_failure"?"infrastructure_failure":e==="cancelled"?"cancelled":"policy_failure"}d(Ra,"failureClassFor");import Xe from"node:path";import{mkdir as ro,readFile as _a,writeFile as so}from"node:fs/promises";async function te(e){let t=Xe.join(e.storeDir,"training","evaluation-artifacts",e.tasksetId),r=Xe.join(t,`${e.attemptId}-${oo(e.fileLabel??e.kind)}.json`),s=Buffer.from(`${JSON.stringify({schemaVersion:"openpond.rawEvaluationArtifact.v1",requestId:e.requestId,...e.payload},null,2)}
 `,"utf8");return await ro(t,{recursive:!0}),await so(r,s,{mode:384}),Xt({...e,file:r,bytes:s,mediaType:"application/json"})}d(te,"persistJsonTaskAttemptArtifact");async function Zc(e){let t=e.extension.replace(/^\.+/,"");if(!/^[a-z0-9]{1,12}$/i.test(t))throw new Error("Task Attempt artifact extension is invalid.");let r=Xe.join(e.storeDir,"training","evaluation-artifacts",e.tasksetId);await ro(r,{recursive:!0});let s=Xe.join(r,`${e.attemptId}-${oo(e.fileLabel)}.${t}`);return await so(s,e.bytes,{mode:384}),Xt({...e,file:s})}d(Zc,"persistBinaryTaskAttemptArtifact");function oo(e){let t=e.trim().replace(/[^a-zA-Z0-9._-]+/g,"-").replace(/^-+|-+$/g,"");if(!t)throw new Error("Evaluation artifact file label is empty.");return t}d(oo,"safeFileLabel");async function ao(e){let t=await _a(e.path),r=se(t);if(t.byteLength!==e.expectedSizeBytes||r!==e.expectedSha256)throw new Error(`Saved Work output ${Xe.basename(e.path)} no longer matches its OutputRef.`);return Xt({...e,file:e.path,bytes:t,kind:"output_artifact",metadata:{...e.metadata,source:"openpond-work",outputValidated:!0}})}d(ao,"persistTaskAttemptOutputArtifact");async function Xt(e){let t=se(e.bytes),r=Qr.parse({schemaVersion:"openpond.taskAttemptArtifact.v1",id:`attempt_artifact_${k([e.attemptId,e.kind,e.file,t]).slice(0,24)}`,tasksetId:e.tasksetId,taskId:e.taskId,attemptId:e.attemptId,kind:e.kind,path:e.file,mediaType:e.mediaType,sha256:t,sizeBytes:e.bytes.byteLength,createdAt:e.timestamp(),metadata:{requestId:e.requestId,localOnly:!0,containsPrivilegedOutcome:!1,...e.metadata}});return e.store.saveTaskAttemptArtifact(r)}d(Xt,"saveArtifact");function Zt(e){let t=new Set(e.approvedTrajectoryIds),r=new Map(e.tasks.map(a=>[a.id,a])),s=new Map(e.results.map(a=>[a.trajectoryId,a]));return e.trajectories.flatMap(a=>{let o=s.get(a.id),n=r.get(a.taskId);return!t.has(a.id)||!n||o?.outcome!=="correct"||!o.rewardEligible?[]:[us.parse({schemaVersion:V,id:`cso_bootstrap_${a.id}`,taskId:n.id,worldId:a.worldId,trajectoryId:a.id,toolContractHash:q,approval:{status:"approved",approvedBy:e.approvedBy,approvedAt:e.approvedAt},messages:xa(n,a)})]})}d(Zt,"buildCrossSystemBootstrapDataset");function xa(e,t){let r=[{role:"system",content:cs},{role:"user",content:e.prompt}];for(let s of t.steps)s.kind==="model"&&s.content.trim()&&r.push({role:"assistant",content:s.content}),s.kind==="tool_call"&&r.push({role:"assistant",content:null,tool_calls:[{id:s.callId,type:"function",function:{name:s.name,arguments:JSON.stringify(s.arguments)}}]}),s.kind==="tool_result"&&r.push({role:"tool",tool_call_id:s.callId,content:JSON.stringify({ok:s.ok,result:s.result,error:s.error})}),s.kind==="final"&&r.push({role:"assistant",content:s.content});return r}d(xa,"trajectoryMessages");import _n from"node:path";import{mkdir as Ia,mkdtemp as Ea,rm as Ca}from"node:fs/promises";import Aa from"node:path";async function no(e,t,r){await Ia(e,{recursive:!0});let s=await Ea(Aa.join(e,".authoring-"));try{return await ut(t,s,{generatedFiles:r}),await Qt(e,t,s)}finally{await Ca(s,{recursive:!0,force:!0})}}d(no,"materializeGeneratedTasksetPackage");async function Qt(e,t,r){let s=await Kr(r),a=await Ss(t,r),o=`authored-${k(a.taskset)}`,n=we.parse({...a.taskset,environment:{...a.taskset.environment,metadata:{...a.taskset.environment.metadata,runtimeSourceTasksetId:o}}}),i=we.parse({...n,contentHash:Re(n)});return await ys(e,{taskset:i,generatedFiles:a.generatedFiles},o,{source:{directory:r,packageHash:s},verify:d(u=>yt(u,i),"verify")}),i}d(Qt,"materializeTasksetRevisionFromSource");import{Buffer as tr}from"node:buffer";import{spawn as Ma}from"node:child_process";var io=String.raw`
 import contextlib, io, json, math, statistics, decimal, datetime, collections, itertools, functools, sys

package/dist/chunks/{chunk-JQOVC35K.js → chunk-UCPD526R.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{ja as h}from"./chunk-HG2YFMVU.js";import{m as f}from"./chunk-RCT3O3HQ.js";import{A as e}from"./chunk-X2J2P5ZT.js";import{a as m}from"./chunk-FTQLQJQV.js";var g=e.object({schemaVersion:e.literal("openpond.profileWorkCompletion.v1"),sessionId:e.string().min(1),turnId:e.string().min(1),responsePath:e.string().min(1),responseSha256:e.string().regex(/^[a-f0-9]{64}$/),responseBytes:e.number().int().min(1).max(1048576),responseChunks:e.array(e.string().max(65536)).min(1).max(32),outputs:e.array(e.object({id:e.string(),revision:e.number().int(),sha256:e.string(),sizeBytes:e.number().int(),contentType:e.string(),sourceTaskId:e.string(),sourceTurnId:e.string(),title:e.string()}).strict()).min(1).max(100)}).strict();function W(s){let u=new TextDecoder("utf-8",{fatal:!0}).decode(s.response),i=[];for(let r=0;r<u.length;r+=65536)i.push(u.slice(r,r+65536));return g.parse({schemaVersion:"openpond.profileWorkCompletion.v1",sessionId:s.sessionId,turnId:s.turnId,responsePath:s.responsePath,responseSha256:f(s.response),responseBytes:s.response.length,responseChunks:i,outputs:s.outputs.map(({id:r,revision:t,sha256:a,sizeBytes:p,contentType:c,sourceTaskId:o,sourceTurnId:d,title:n})=>({id:r,revision:t,sha256:a,sizeBytes:p,contentType:c,sourceTaskId:o,sourceTurnId:d,title:n}))})}m(W,"createProfileWorkCompletion");function _(s){let u=s.filter(o=>o.name==="tool.completed"&&o.action==="work_finish"&&o.status==="completed");if(!u.length)return null;if(u.length!==1)throw new Error("The case has multiple Work completion receipts.");let i=u[0],r=i.data,t=g.parse(r?.result);if(t.sessionId!==i.sessionId||t.turnId!==i.turnId)throw new Error("Work completion belongs to another case or turn.");let a=t.responseChunks.join(""),p=Buffer.from(a,"utf8");if(p.length!==t.responseBytes||f(p)!==t.responseSha256)throw new Error("Work completion response differs from its sealed content.");if(new Set(t.outputs.map(o=>`${o.id}/${o.revision}`)).size!==t.outputs.length)throw new Error("Work completion repeats a saved output.");let c=s.slice(0,s.indexOf(i));for(let o of t.outputs){if(o.sourceTaskId!==t.sessionId||o.sourceTurnId!==t.turnId)throw new Error("Work completion output belongs to another case or turn.");if(c.filter(n=>n.name==="workspace_action_result"&&n.action==="work_output_save"&&n.status==="completed"&&n.sessionId===t.sessionId&&n.turnId===t.turnId).flatMap(n=>{let l=h.safeParse(n.data?.outputRef);return l.success&&Object.entries(o).every(([k,I])=>l.data[k]===I)?[l.data]:[]}).length!==1)throw new Error("Work completion lacks an exact prior saved-output receipt.")}return a}m(_,"completedProfileWorkResponse");export{W as a,_ as b};
+import{ja as h}from"./chunk-2THZ6GZY.js";import{m as f}from"./chunk-RCT3O3HQ.js";import{A as e}from"./chunk-X2J2P5ZT.js";import{a as m}from"./chunk-FTQLQJQV.js";var g=e.object({schemaVersion:e.literal("openpond.profileWorkCompletion.v1"),sessionId:e.string().min(1),turnId:e.string().min(1),responsePath:e.string().min(1),responseSha256:e.string().regex(/^[a-f0-9]{64}$/),responseBytes:e.number().int().min(1).max(1048576),responseChunks:e.array(e.string().max(65536)).min(1).max(32),outputs:e.array(e.object({id:e.string(),revision:e.number().int(),sha256:e.string(),sizeBytes:e.number().int(),contentType:e.string(),sourceTaskId:e.string(),sourceTurnId:e.string(),title:e.string()}).strict()).min(1).max(100)}).strict();function W(s){let u=new TextDecoder("utf-8",{fatal:!0}).decode(s.response),i=[];for(let r=0;r<u.length;r+=65536)i.push(u.slice(r,r+65536));return g.parse({schemaVersion:"openpond.profileWorkCompletion.v1",sessionId:s.sessionId,turnId:s.turnId,responsePath:s.responsePath,responseSha256:f(s.response),responseBytes:s.response.length,responseChunks:i,outputs:s.outputs.map(({id:r,revision:t,sha256:a,sizeBytes:p,contentType:c,sourceTaskId:o,sourceTurnId:d,title:n})=>({id:r,revision:t,sha256:a,sizeBytes:p,contentType:c,sourceTaskId:o,sourceTurnId:d,title:n}))})}m(W,"createProfileWorkCompletion");function _(s){let u=s.filter(o=>o.name==="tool.completed"&&o.action==="work_finish"&&o.status==="completed");if(!u.length)return null;if(u.length!==1)throw new Error("The case has multiple Work completion receipts.");let i=u[0],r=i.data,t=g.parse(r?.result);if(t.sessionId!==i.sessionId||t.turnId!==i.turnId)throw new Error("Work completion belongs to another case or turn.");let a=t.responseChunks.join(""),p=Buffer.from(a,"utf8");if(p.length!==t.responseBytes||f(p)!==t.responseSha256)throw new Error("Work completion response differs from its sealed content.");if(new Set(t.outputs.map(o=>`${o.id}/${o.revision}`)).size!==t.outputs.length)throw new Error("Work completion repeats a saved output.");let c=s.slice(0,s.indexOf(i));for(let o of t.outputs){if(o.sourceTaskId!==t.sessionId||o.sourceTurnId!==t.turnId)throw new Error("Work completion output belongs to another case or turn.");if(c.filter(n=>n.name==="workspace_action_result"&&n.action==="work_output_save"&&n.status==="completed"&&n.sessionId===t.sessionId&&n.turnId===t.turnId).flatMap(n=>{let l=h.safeParse(n.data?.outputRef);return l.success&&Object.entries(o).every(([k,I])=>l.data[k]===I)?[l.data]:[]}).length!==1)throw new Error("Work completion lacks an exact prior saved-output receipt.")}return a}m(_,"completedProfileWorkResponse");export{W as a,_ as b};

package/dist/chunks/{chunk-Y4I6CFZJ.js → chunk-UGDBMRCT.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{t as _,w as C}from"./chunk-WGQ3FMRF.js";import{f as R,m}from"./chunk-HIQPGYVH.js";import{Sa as A}from"./chunk-ZY32SZ2P.js";import{A as e}from"./chunk-X2J2P5ZT.js";import{a as l}from"./chunk-FTQLQJQV.js";var u=e.string().trim().min(1).max(200),P=e.object({operationId:u,execution:R,executionKind:e.literal("recorded_evidence").optional(),graders:e.array(A).min(1).max(100),maximumCostUsd:e.number().finite().min(1e-6).max(1e4),mappings:e.array(e.object({graderId:u,fields:_}).strict()).max(100).optional()}).strict().superRefine((a,n)=>{new Set(a.graders.map(o=>o.id)).size!==a.graders.length&&n.addIssue({code:"custom",path:["graders"],message:"Select each grader once."}),a.mappings&&(new Set(a.mappings.map(o=>o.graderId)).size!==a.mappings.length||a.mappings.some(o=>!a.graders.some(c=>c.id===o.graderId)))&&n.addIssue({code:"custom",path:["mappings"],message:"Map each selected grader at most once."})}),q=e.object({schemaVersion:e.literal("openpond.experimentScoringPass.v1"),id:u,teamId:u,revision:e.number().int().positive(),request:P,graders:e.array(C).min(1).max(100),status:e.enum(["queued","running","cancelling","completed","failed","cancelled"]),totalCount:e.number().int().min(1).max(1e4),counts:e.object({pending:e.number().int().nonnegative(),running:e.number().int().nonnegative(),scored:e.number().int().nonnegative(),unavailable:e.number().int().nonnegative(),failed:e.number().int().nonnegative(),cancelled:e.number().int().nonnegative()}).strict(),resultAvailable:e.boolean(),createdAt:e.iso.datetime(),startedAt:e.iso.datetime().nullable(),completedAt:e.iso.datetime().nullable(),error:e.object({code:u,message:e.string().max(2e3)}).strict().nullable()}).strict(),L=q.extend({contentHash:e.string().regex(/^[a-f0-9]{64}$/)});function O(a){let n=L.parse(a),{contentHash:o,...c}=n;if(m(c)!==o||Object.values(n.counts).reduce((r,d)=>r+d,0)!==n.totalCount)throw new Error("Scoring pass integrity or population accounting failed.");let t=["completed","failed","cancelled"].includes(n.status);if(t&&(!n.completedAt||n.counts.pending||n.counts.running)||!t&&(n.completedAt!==null||n.resultAvailable)||n.status==="completed"&&!n.resultAvailable)throw new Error("Scoring pass lifecycle differs from its retained result.");if(n.graders.length!==n.request.graders.length||new Set(n.graders.map(r=>r.feedbackKey)).size!==n.graders.length||n.graders.some((r,d)=>m(r.release)!==m(n.request.graders[d])||m(r.mappings??[])!==m(n.request.mappings?.find(f=>f.graderId===r.release?.id)?.fields??[])))throw new Error("Scoring pass graders differ from their exact admission.");return n}l(O,"verifyExperimentScoringPass");var I=8388608;async function B(a,n,o,c){let t=new AbortController,r=o.signal,d=l(()=>t.abort(r?.reason),"forward");r?.aborted?d():r?.addEventListener("abort",d,{once:!0});let f=setTimeout(()=>t.abort(new Error("Connected evidence request timed out.")),6e4),s,h=l(()=>{s&&s.cancel(t.signal.reason).catch(()=>{})},"abortReader");t.signal.addEventListener("abort",h,{once:!0});let v=new Promise((w,g)=>{let p=l(()=>g(t.signal.reason??new DOMException("Request aborted","AbortError")),"abort");t.signal.aborted?p():t.signal.addEventListener("abort",p,{once:!0})});try{let w=a(n,{...o,signal:t.signal}).then(i=>{if(t.signal.aborted)throw i.body?.cancel(t.signal.reason).catch(()=>{}),t.signal.reason;return i}),g=await Promise.race([w,v]),p=g.headers.get("content-length");if(s=g.body?.getReader(),t.signal.aborted)throw await s?.cancel(t.signal.reason),t.signal.reason;if(p&&Number(p)>I)throw await s?.cancel(),c(502,"connected_response_too_large","Connected evidence response exceeds 8 MiB.");let x=[],b=0;if(s)for(;;){let i=await Promise.race([s.read(),v]);if(i.done)break;if(b+=i.value.byteLength,b>I)throw await s.cancel(),c(502,"connected_response_too_large","Connected evidence response exceeds 8 MiB.");x.push(i.value)}let S=new Uint8Array(b),E=0;for(let i of x)S.set(i,E),E+=i.byteLength;let y;try{y=JSON.parse(new TextDecoder("utf-8",{fatal:!0}).decode(S))}catch{throw c(502,"connected_response_invalid","Connected evidence response is not valid JSON.")}return{response:g,value:y}}finally{clearTimeout(f),r?.removeEventListener("abort",d),t.signal.removeEventListener("abort",h),s?.releaseLock()}}l(B,"fetchConnectedJson");export{P as a,O as b,B as c};
+import{t as _,w as C}from"./chunk-XKDOPNQ5.js";import{f as R,m}from"./chunk-HIQPGYVH.js";import{Sa as A}from"./chunk-ZY32SZ2P.js";import{A as e}from"./chunk-X2J2P5ZT.js";import{a as l}from"./chunk-FTQLQJQV.js";var u=e.string().trim().min(1).max(200),P=e.object({operationId:u,execution:R,executionKind:e.literal("recorded_evidence").optional(),graders:e.array(A).min(1).max(100),maximumCostUsd:e.number().finite().min(1e-6).max(1e4),mappings:e.array(e.object({graderId:u,fields:_}).strict()).max(100).optional()}).strict().superRefine((a,n)=>{new Set(a.graders.map(o=>o.id)).size!==a.graders.length&&n.addIssue({code:"custom",path:["graders"],message:"Select each grader once."}),a.mappings&&(new Set(a.mappings.map(o=>o.graderId)).size!==a.mappings.length||a.mappings.some(o=>!a.graders.some(c=>c.id===o.graderId)))&&n.addIssue({code:"custom",path:["mappings"],message:"Map each selected grader at most once."})}),q=e.object({schemaVersion:e.literal("openpond.experimentScoringPass.v1"),id:u,teamId:u,revision:e.number().int().positive(),request:P,graders:e.array(C).min(1).max(100),status:e.enum(["queued","running","cancelling","completed","failed","cancelled"]),totalCount:e.number().int().min(1).max(1e4),counts:e.object({pending:e.number().int().nonnegative(),running:e.number().int().nonnegative(),scored:e.number().int().nonnegative(),unavailable:e.number().int().nonnegative(),failed:e.number().int().nonnegative(),cancelled:e.number().int().nonnegative()}).strict(),resultAvailable:e.boolean(),createdAt:e.iso.datetime(),startedAt:e.iso.datetime().nullable(),completedAt:e.iso.datetime().nullable(),error:e.object({code:u,message:e.string().max(2e3)}).strict().nullable()}).strict(),L=q.extend({contentHash:e.string().regex(/^[a-f0-9]{64}$/)});function O(a){let n=L.parse(a),{contentHash:o,...c}=n;if(m(c)!==o||Object.values(n.counts).reduce((r,d)=>r+d,0)!==n.totalCount)throw new Error("Scoring pass integrity or population accounting failed.");let t=["completed","failed","cancelled"].includes(n.status);if(t&&(!n.completedAt||n.counts.pending||n.counts.running)||!t&&(n.completedAt!==null||n.resultAvailable)||n.status==="completed"&&!n.resultAvailable)throw new Error("Scoring pass lifecycle differs from its retained result.");if(n.graders.length!==n.request.graders.length||new Set(n.graders.map(r=>r.feedbackKey)).size!==n.graders.length||n.graders.some((r,d)=>m(r.release)!==m(n.request.graders[d])||m(r.mappings??[])!==m(n.request.mappings?.find(f=>f.graderId===r.release?.id)?.fields??[])))throw new Error("Scoring pass graders differ from their exact admission.");return n}l(O,"verifyExperimentScoringPass");var I=8388608;async function B(a,n,o,c){let t=new AbortController,r=o.signal,d=l(()=>t.abort(r?.reason),"forward");r?.aborted?d():r?.addEventListener("abort",d,{once:!0});let f=setTimeout(()=>t.abort(new Error("Connected evidence request timed out.")),6e4),s,h=l(()=>{s&&s.cancel(t.signal.reason).catch(()=>{})},"abortReader");t.signal.addEventListener("abort",h,{once:!0});let v=new Promise((w,g)=>{let p=l(()=>g(t.signal.reason??new DOMException("Request aborted","AbortError")),"abort");t.signal.aborted?p():t.signal.addEventListener("abort",p,{once:!0})});try{let w=a(n,{...o,signal:t.signal}).then(i=>{if(t.signal.aborted)throw i.body?.cancel(t.signal.reason).catch(()=>{}),t.signal.reason;return i}),g=await Promise.race([w,v]),p=g.headers.get("content-length");if(s=g.body?.getReader(),t.signal.aborted)throw await s?.cancel(t.signal.reason),t.signal.reason;if(p&&Number(p)>I)throw await s?.cancel(),c(502,"connected_response_too_large","Connected evidence response exceeds 8 MiB.");let x=[],b=0;if(s)for(;;){let i=await Promise.race([s.read(),v]);if(i.done)break;if(b+=i.value.byteLength,b>I)throw await s.cancel(),c(502,"connected_response_too_large","Connected evidence response exceeds 8 MiB.");x.push(i.value)}let S=new Uint8Array(b),E=0;for(let i of x)S.set(i,E),E+=i.byteLength;let y;try{y=JSON.parse(new TextDecoder("utf-8",{fatal:!0}).decode(S))}catch{throw c(502,"connected_response_invalid","Connected evidence response is not valid JSON.")}return{response:g,value:y}}finally{clearTimeout(f),r?.removeEventListener("abort",d),t.signal.removeEventListener("abort",h),s?.releaseLock()}}l(B,"fetchConnectedJson");export{P as a,O as b,B as c};

package/dist/chunks/{chunk-E2JB5OCO.js → chunk-UZE3C2SF.js}

@@ -1,5 +1,5 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{a as Ae}from"./chunk-GW766ZSO.js";import{B as O,a as Ce,b as Pe}from"./chunk-OFJPOTES.js";import{b as be}from"./chunk-YU3C6U5V.js";import{d as H}from"./chunk-U7D6Q744.js";import{a as re,b as ye,j as A}from"./chunk-6ESBAHWK.js";import{f as Ie}from"./chunk-JT5YQ52I.js";import{D as we}from"./chunk-2MEOUS24.js";import{n as $}from"./chunk-RCT3O3HQ.js";import{a as o}from"./chunk-FTQLQJQV.js";import{existsSync as z}from"node:fs";import{cp as xe,mkdir as Ee,readFile as Oe}from"node:fs/promises";import k from"node:path";import{existsSync as L}from"node:fs";import{readFile as K}from"node:fs/promises";import q from"node:path";async function F(e){await V(e.target);let t={},a=new Set([`${e.target.sourceRootRelativePath}/.openpond/agent-inspect.json`,`${e.target.sourceRootRelativePath}/.openpond/action-registry.json`,`${e.target.sourceRootRelativePath}/.openpond/eval-results.json`]),n=[{name:"inspect",command:"inspect",args:["--json"]},{name:"build",command:"build"},{name:"validate",command:"validate",args:["--json"]},{name:"eval",command:"eval",args:["--json"]},{name:"direct-run",command:"run",args:[e.target.defaultAction,"--json","--input",JSON.stringify({prompt:e.snapshot.objective,channel:"openpond_chat"})]}];for(let r of n){let s=await O({cwd:e.target.sourceRoot,command:r.command,args:r.args,...r.command==="eval"&&e.requireEvalPass===!1?{throwOnFailure:!1}:{}});if(s.code!==0&&!(r.command==="eval"&&e.requireEvalPass===!1))throw new Error(s.stderr.trim()||s.stdout.trim()||`${r.command} failed with exit code ${s.code??"unknown"}`);let i=lt(s.stdout),d=mt(i);d&&a.add(q.join(e.target.sourceRootRelativePath,d)),t[r.name]={summary:pt(r.name,i),traceArtifactRef:d,stdout:s.stdout.trim().slice(0,4e3),stderr:s.stderr.trim().slice(0,4e3)}}return await dt(e.target,e.requireEvalPass!==!1),{checkRefs:[...a],metadata:t}}o(F,"runLocalCreatePipelineChecks");async function V(e){if(!L(e.sourceRoot))throw new Error(`Codex did not create the expected agent source root: ${e.sourceRoot}`);let t=q.join(e.sourceRoot,"agent","agent.ts");if(!L(t))throw new Error(`Codex did not create the expected SDK agent source: ${t}`);let a=q.join(e.repoPath,"openpond-profile.json");if(!L(a))throw new Error(`Codex did not update the profile repo manifest: ${a}`);let n=q.join(e.sourcePath,"settings","profile.yaml");if(!L(n))throw new Error(`Codex did not update the profile manifest: ${n}`);let s=JSON.parse(await K(a,"utf8")).profiles?.[e.activeProfile];if(!s)throw new Error(`openpond-profile.json does not register profile ${e.activeProfile}.`);if(!s.enabledAgents?.includes(e.agentId))throw new Error(`openpond-profile.json does not enable generated agent ${e.agentId}.`);let i=await K(n,"utf8");if(!i.includes(`id: ${e.agentId}`)&&!i.includes(`id: "${e.agentId}"`))throw new Error(`settings/profile.yaml does not register generated agent ${e.agentId}.`);let d=e.agentId==="default"?["agent/agent.ts","agent"]:[`agents/${e.agentId}`];if(!d.some(u=>i.includes(u)))throw new Error(`settings/profile.yaml does not point ${e.agentId} at ${d[0]}.`)}o(V,"assertLocalCreateSourceLayout");async function dt(e,t=!0){let a=q.join(e.sourceRoot,".openpond","action-registry.json"),n=q.join(e.sourceRoot,".openpond","eval-results.json");if(!L(a))throw new Error(`Generated source did not produce an action registry: ${a}`);if(!L(n))throw new Error(`Generated source did not produce eval results: ${n}`);let r=JSON.parse(await K(a,"utf8")),s=Array.isArray(r.actions)?r.actions:[];if(!new Set(s.map(c=>typeof c.id=="string"?c.id:null).filter(c=>!!c)).has(e.defaultAction))throw new Error(`Generated action registry does not expose default action ${e.defaultAction}.`);let d=s.filter(c=>typeof c.id=="string"&&(!W(c.label)||!W(c.description)||typeof c.timeoutSeconds!="number"));if(d.length>0)throw new Error(`Generated action registry has actions missing labels, descriptions, or timeout policies: ${d.map(c=>c.id).join(", ")}`);let u=Ce({actionCatalog:s.filter(c=>typeof c.id=="string").map(c=>({id:c.id,name:W(c.name),setupRequirements:ut(c.setupRequirements)??[]})),actionId:e.defaultAction});if(u.blockingRequirements.length>0)throw new Error(`Generated default action ${e.defaultAction} has unresolved required setup and cannot be ready_local: ${u.blockingRequirements.map(Pe).join("; ")}. Mark built-in/local fixture requirements ready or optional when they are satisfied, or leave the pipeline blocked until the setup is configured.`);let m=JSON.parse(await K(n,"utf8"));if(typeof m.summary?.total!="number"||m.summary.total<1)throw new Error("Generated source must define at least one deterministic SDK eval.");if(t&&m.summary.failed!==0)throw new Error("Generated source evals did not all pass.")}o(dt,"assertGeneratedSdkArtifacts");function lt(e){let t=e.trim();if(!t.startsWith("{")&&!t.startsWith("["))return null;try{return JSON.parse(t)}catch{return null}}o(lt,"parseJsonOutput");function E(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}o(E,"asRecord");function ut(e){return Array.isArray(e)?e.filter(t=>!!E(t)):null}o(ut,"recordArray");function mt(e){let t=E(e),a=W(t?.traceArtifactRef);return a?.startsWith(".openpond/")?a:null}o(mt,"traceRefFromCommandResult");function pt(e,t){let a=E(t);return a?e==="inspect"?{actionCount:Array.isArray(a.actionCatalog)?a.actionCatalog.length:null,defaultAction:E(a.agent)?.defaultAction??null,projectName:E(a.project)?.name??null}:e==="validate"?{status:a.status??null,errors:E(a.summary)?.errors??null,warnings:E(a.summary)?.warnings??null}:e==="eval"?E(a.summary):e==="direct-run"?{hasResult:!!E(a.result),resultKeys:Object.keys(E(a.result)??{})}:null:null}o(pt,"summaryFromCommandResult");function W(e){return typeof e=="string"&&e.trim()?e.trim():null}o(W,"stringValue");var gt="Approved local Create plans require model-backed SDK source application; no source mutation was performed.",ft=1800*1e3,Te=".openpond-harness-source-application.json",ht=300*1e3,vt=2*1e3,Rt=5*1e3;async function He(e,t={}){if(!kt(e))return e;let a=t.now??(()=>new Date().toISOString()),n=e.adapter;if(n.kind!=="local")return e;if(!t.session||!t.turn||!t.ensureCodexRuntime||!t.appendRuntimeEvent||!t.setProviderTurnId)return Y(e,{now:a(),reason:gt,metadata:{reason:"model_backed_source_application_required",sourcePath:n.sourcePath??null,repoPath:n.repoPath??null,activeProfile:n.activeProfile??null,operation:e.operation,requestedAgentId:e.target.kind==="agent"?e.target.id:null}});let r;try{r=se(e)}catch(i){return Y(e,{now:a(),reason:i instanceof Error?i.message:String(i),metadata:{reason:"local_create_target_unresolved"}})}try{await ie({snapshot:e,session:t.session,turn:t.turn,target:r,ensureCodexRuntime:t.ensureCodexRuntime,appendRuntimeEvent:t.appendRuntimeEvent,setProviderTurnId:t.setProviderTurnId,model:t.model??we,codexReasoningEffort:t.codexReasoningEffort})}catch(i){return Y(e,{now:a(),reason:i instanceof Error?i.message:String(i),metadata:{reason:"model_backed_source_application_failed",target:r}})}let s=oe(e,{state:"running_checks",now:a(),blockedReason:null,sourceRefs:[r.sourceRootRelativePath,k.join(r.profileRelativePath,"settings","profile.yaml"),"openpond-profile.json"],metadata:{localCreatePipeline:{status:"running_checks",target:r}},candidates:[{id:Se(e.id),target:{kind:"agent",id:r.agentId,displayName:e.target.displayName,defaultActionKey:r.defaultAction},status:"checking",git:null,parentCandidateId:null,tasksetRef:e.tasksetRef,authoringModelRef:t.model??null,allowedPaths:e.plan?.sourcePlan.map(i=>i.path)??[],sourceRefs:[r.sourceRootRelativePath,k.join(r.profileRelativePath,"settings","profile.yaml"),"openpond-profile.json"],artifactRefs:[],checkRefs:[],evaluationReceiptRefs:[],createdAt:e.createdAt,updatedAt:a(),metadata:{sourceAuthority:e.adapter.sourceAuthority}}]});await t.onSnapshot?.(s);try{let i=await(t.runChecks??F)({snapshot:s,target:r}),d=a(),u=Se(e.id),m=`agent_eval_${e.id}`;return oe(s,{state:"ready_local",now:d,blockedReason:null,checkRefs:i.checkRefs,candidates:s.candidates.map(c=>c.id===u?{...c,status:"accepted",checkRefs:X(c.checkRefs,i.checkRefs),evaluationReceiptRefs:X(c.evaluationReceiptRefs,[m]),updatedAt:d}:c),evaluationReceipts:[...s.evaluationReceipts.filter(c=>c.id!==m),{id:m,candidateId:u,target:{kind:"agent",id:r.agentId,displayName:e.target.displayName,defaultActionKey:r.defaultAction},evaluatorKind:"agent_sdk",subject:"candidate",sourceCommit:e.adapter.kind==="local"?e.adapter.localHead:null,sourceBranch:null,tasksetId:e.tasksetRef?.id??null,tasksetHash:e.tasksetRef?.contentHash??null,taskAttemptRefs:i.checkRefs.filter(c=>c.includes("eval")),status:"passed",publishGate:"passed",summaryCounts:null,evalRefs:i.checkRefs.filter(c=>c.includes("eval")),artifactRefs:i.checkRefs,summary:wt(i.metadata),createdAt:d,metadata:i.metadata??{}}],metadata:{localCreatePipeline:{status:"ready_local",target:r,checks:i.metadata??{}}}})}catch(i){return Y(s,{now:a(),reason:i instanceof Error?i.message:String(i),metadata:{reason:"local_create_checks_failed",target:r}})}}o(He,"applyApprovedLocalCreateImproveRun");function kt(e){return e.state==="applying_source"&&e.plan?.status==="approved"&&e.adapter.kind==="local"&&e.target.kind==="agent"}o(kt,"shouldApplyLocalCreatePipeline");function oe(e,t){return A(e,{state:t.state,blockedReason:t.blockedReason===void 0?e.blockedReason:t.blockedReason,checkRefs:X(e.checkRefs,t.checkRefs??[]),sourceRefs:X(e.sourceRefs,t.sourceRefs??[]),candidates:t.candidates??e.candidates,evaluationReceipts:t.evaluationReceipts??e.evaluationReceipts,metadata:{...e.metadata,...t.metadata??{}},updatedAt:t.now})}o(oe,"transitionCreatePipeline");function Y(e,t){return oe(e,{state:"blocked",now:t.now,blockedReason:t.reason,metadata:{localCreatePipeline:{status:"blocked",...t.metadata}}})}o(Y,"blockLocalCreatePipeline");function X(e,t){return Array.from(new Set([...e,...t].filter(Boolean)))}o(X,"mergeRefs");function Se(e){return`agent_candidate_${e}`}o(Se,"agentCandidateId");function wt(e){let t=M(e?.eval),a=typeof t?.total=="number"?t.total:null;return a===null?"Agent SDK checks and deterministic Evals passed.":`${a} Agent SDK Eval${a===1?"":"s"} passed.`}o(wt,"checkSummary");function se(e){let t=e.adapter;if(t.kind!=="local")throw new Error("Create/Improve target must use the local adapter.");if(e.target.kind!=="agent")throw new Error("Local agent source application requires an Agent target.");let a=t.sourcePath?.trim(),n=t.repoPath?.trim();if(!a)throw new Error("Local Create approval is missing a profile source path.");if(!n)throw new Error("Local Create approval is missing a profile repo path.");let r=t.activeProfile?.trim()||"default",s=k.resolve(a),i=k.resolve(n),d=qe(e.target.id)??xt(e)??St(e.objective),u=Tt(Et(e.plan?.defaultChatAction.key)??"chat",d),m=e.operation==="improve"&&d==="default"?s:k.join(s,"agents",d);return{activeProfile:r,agentId:d,defaultAction:u,repoPath:i,sourcePath:s,workspaceRoot:i,profileRelativePath:je(k.relative(i,s)),sourceRoot:m,sourceRootRelativePath:je(k.relative(i,m))}}o(se,"resolveLocalCreatePipelineTarget");async function ie(e){if(!z(e.target.workspaceRoot))throw new Error(`Profile repo path does not exist: ${e.target.workspaceRoot}`);if(!z(e.target.sourcePath))throw new Error(`Profile source path does not exist: ${e.target.sourcePath}`);if(await Ct(e.snapshot,e.target)){await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"server",appId:e.session.appId,status:"completed",output:`Applied prepared harness ${e.snapshot.operation==="improve"?"Improve":"Create"} source for ${e.target.agentId}; running local checks.`,data:{phase:"prepared_source_apply_completed",runId:e.snapshot.id,target:e.target}}));return}await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"server",appId:e.session.appId,status:"started",output:`Applying approved ${e.snapshot.operation==="improve"?"Improve":"Create"} plan with Codex in ${e.target.sourceRootRelativePath}.`,data:{phase:"source_apply_started",runId:e.snapshot.id,target:e.target}}));let t={...e.session,provider:"codex",cwd:e.target.workspaceRoot,workspaceKind:"local_project",workspaceId:e.session.workspaceId??`openpond-profile:${e.target.activeProfile}`,workspaceName:e.session.workspaceName??`OpenPond profile ${e.target.activeProfile}`},a=e.session.provider==="codex"?e.model:null,n=await e.ensureCodexRuntime(t,{approvalPolicy:"never",sandbox:"workspace-write",model:a,codexPermissionMode:"auto-review",codexReasoningEffort:e.codexReasoningEffort}),r=await n.client.startTurn({threadId:n.threadId,prompt:At(e.snapshot,e.target),cwd:e.target.workspaceRoot,model:a,approvalPolicy:"never",sandbox:"workspace-write"});await e.setProviderTurnId(r.turnId);try{await n.client.waitForTurn(r.turnId,ft)}catch(s){if(!$e(s,r.turnId)&&!It(s))throw s;let i=$e(s,r.turnId)?"timed out":"exited cleanly before returning a turn completion";if(await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"server",appId:e.session.appId,status:"started",output:`Codex source application ${i}; checking for completed generated source before blocking.`,data:{phase:"source_apply_recovery_started",runId:e.snapshot.id,providerTurnId:r.turnId,reason:i,target:e.target}})),!await yt(e.target))throw s;await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"server",appId:e.session.appId,status:"completed",output:`Recovered completed ${e.snapshot.operation==="improve"?"Improve":"Create"} source after Codex source application ${i} for ${e.target.agentId}; running local checks.`,data:{phase:"source_apply_recovered",runId:e.snapshot.id,providerTurnId:r.turnId,reason:i,target:e.target}}))}await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"provider",appId:e.session.appId,status:"completed",output:`Codex completed approved ${e.snapshot.operation==="improve"?"Improve":"Create"} source application for ${e.target.agentId}.`,data:{phase:"source_apply_completed",runId:e.snapshot.id,providerTurnId:r.turnId,target:e.target}}))}o(ie,"runModelBackedLocalCreateSourceApplication");async function Ct(e,t,a=process.env){if(!Ae(a))return!1;let n=k.join(t.workspaceRoot,Te);if(!z(n))return!1;let r=JSON.parse(await Oe(n,"utf8")),s=M(r);if(s?.schema!=="openpond.harnessPreparedSource.v1")throw new Error(`${Te} must use schema openpond.harnessPreparedSource.v1.`);let i=M(s.agents),d=M(i?.[t.agentId]),u=Pt(d?.[e.operation]);if(!u)return!1;let m=ne(t.workspaceRoot,u.source,"source");if(!z(m))throw new Error(`Prepared harness source does not exist: ${u.source}`);await Ee(t.sourceRoot,{recursive:!0}),await xe(m,t.sourceRoot,{recursive:!0,force:!0});for(let c of u.registrations??[]){let g=ne(t.workspaceRoot,c.source,"registration source"),l=ne(t.workspaceRoot,c.target,"registration target");if(!z(g))throw new Error(`Prepared harness registration source does not exist: ${c.source}`);await Ee(k.dirname(l),{recursive:!0}),await xe(g,l,{force:!0})}return!0}o(Ct,"applyPreparedHarnessSourceApplication");function Pt(e){let t=M(e);if(!t||typeof t.source!="string"||!t.source.trim())return null;if(t.registrations!==void 0&&!Array.isArray(t.registrations))throw new Error("Prepared harness source registrations must be an array.");let a=(t.registrations??[]).map(n=>{let r=M(n);if(!r||typeof r.source!="string"||!r.source.trim()||typeof r.target!="string"||!r.target.trim())throw new Error("Prepared harness source registrations require source and target paths.");return{source:r.source,target:r.target}});return{source:t.source,registrations:a}}o(Pt,"asHarnessPreparedSourceOperation");function ne(e,t,a){if(k.isAbsolute(t))throw new Error(`Prepared harness ${a} must be relative to the profile repo.`);let n=k.resolve(e),r=k.resolve(n,t);if(r!==n&&!r.startsWith(`${n}${k.sep}`))throw new Error(`Prepared harness ${a} must stay inside the profile repo.`);return r}o(ne,"resolveHarnessManifestPath");function $e(e,t){return(e instanceof Error?e.message:String(e))===`codex turn ${t} timed out`}o($e,"isCodexTurnTimeoutError");function It(e){return(e instanceof Error?e.message:String(e)).startsWith("codex app-server exited with code 0")}o(It,"isCodexAppServerCleanExitError");async function yt(e){let t=Date.now()+ht;for(;Date.now()<t;){if(await bt(e))return!0;await Le(vt)}return!1}o(yt,"waitForTimedOutSourceApplication");async function bt(e){try{await V(e);let t=await _e(e);await Le(Rt),await V(e);let a=await _e(e);return t===a}catch{return!1}}o(bt,"localCreateSourceLayoutIsStable");async function _e(e){let t=[k.join(e.sourceRoot,"agent","agent.ts"),k.join(e.sourceRoot,"package.json"),k.join(e.repoPath,"openpond-profile.json"),k.join(e.sourcePath,"settings","profile.yaml")];return(await Promise.all(t.map(async n=>{let r=await Oe(n,"utf8");return`${n}:${r.length}:${r}`}))).join(`
+import{a as Ae}from"./chunk-GW766ZSO.js";import{B as O,a as Ce,b as Pe}from"./chunk-OFJPOTES.js";import{b as be}from"./chunk-YU3C6U5V.js";import{d as H}from"./chunk-L5L3H6CI.js";import{a as re,b as ye,j as A}from"./chunk-6ESBAHWK.js";import{f as Ie}from"./chunk-Z6HWSTEP.js";import{D as we}from"./chunk-2MEOUS24.js";import{n as $}from"./chunk-RCT3O3HQ.js";import{a as o}from"./chunk-FTQLQJQV.js";import{existsSync as z}from"node:fs";import{cp as xe,mkdir as Ee,readFile as Oe}from"node:fs/promises";import k from"node:path";import{existsSync as L}from"node:fs";import{readFile as K}from"node:fs/promises";import q from"node:path";async function F(e){await V(e.target);let t={},a=new Set([`${e.target.sourceRootRelativePath}/.openpond/agent-inspect.json`,`${e.target.sourceRootRelativePath}/.openpond/action-registry.json`,`${e.target.sourceRootRelativePath}/.openpond/eval-results.json`]),n=[{name:"inspect",command:"inspect",args:["--json"]},{name:"build",command:"build"},{name:"validate",command:"validate",args:["--json"]},{name:"eval",command:"eval",args:["--json"]},{name:"direct-run",command:"run",args:[e.target.defaultAction,"--json","--input",JSON.stringify({prompt:e.snapshot.objective,channel:"openpond_chat"})]}];for(let r of n){let s=await O({cwd:e.target.sourceRoot,command:r.command,args:r.args,...r.command==="eval"&&e.requireEvalPass===!1?{throwOnFailure:!1}:{}});if(s.code!==0&&!(r.command==="eval"&&e.requireEvalPass===!1))throw new Error(s.stderr.trim()||s.stdout.trim()||`${r.command} failed with exit code ${s.code??"unknown"}`);let i=lt(s.stdout),d=mt(i);d&&a.add(q.join(e.target.sourceRootRelativePath,d)),t[r.name]={summary:pt(r.name,i),traceArtifactRef:d,stdout:s.stdout.trim().slice(0,4e3),stderr:s.stderr.trim().slice(0,4e3)}}return await dt(e.target,e.requireEvalPass!==!1),{checkRefs:[...a],metadata:t}}o(F,"runLocalCreatePipelineChecks");async function V(e){if(!L(e.sourceRoot))throw new Error(`Codex did not create the expected agent source root: ${e.sourceRoot}`);let t=q.join(e.sourceRoot,"agent","agent.ts");if(!L(t))throw new Error(`Codex did not create the expected SDK agent source: ${t}`);let a=q.join(e.repoPath,"openpond-profile.json");if(!L(a))throw new Error(`Codex did not update the profile repo manifest: ${a}`);let n=q.join(e.sourcePath,"settings","profile.yaml");if(!L(n))throw new Error(`Codex did not update the profile manifest: ${n}`);let s=JSON.parse(await K(a,"utf8")).profiles?.[e.activeProfile];if(!s)throw new Error(`openpond-profile.json does not register profile ${e.activeProfile}.`);if(!s.enabledAgents?.includes(e.agentId))throw new Error(`openpond-profile.json does not enable generated agent ${e.agentId}.`);let i=await K(n,"utf8");if(!i.includes(`id: ${e.agentId}`)&&!i.includes(`id: "${e.agentId}"`))throw new Error(`settings/profile.yaml does not register generated agent ${e.agentId}.`);let d=e.agentId==="default"?["agent/agent.ts","agent"]:[`agents/${e.agentId}`];if(!d.some(u=>i.includes(u)))throw new Error(`settings/profile.yaml does not point ${e.agentId} at ${d[0]}.`)}o(V,"assertLocalCreateSourceLayout");async function dt(e,t=!0){let a=q.join(e.sourceRoot,".openpond","action-registry.json"),n=q.join(e.sourceRoot,".openpond","eval-results.json");if(!L(a))throw new Error(`Generated source did not produce an action registry: ${a}`);if(!L(n))throw new Error(`Generated source did not produce eval results: ${n}`);let r=JSON.parse(await K(a,"utf8")),s=Array.isArray(r.actions)?r.actions:[];if(!new Set(s.map(c=>typeof c.id=="string"?c.id:null).filter(c=>!!c)).has(e.defaultAction))throw new Error(`Generated action registry does not expose default action ${e.defaultAction}.`);let d=s.filter(c=>typeof c.id=="string"&&(!W(c.label)||!W(c.description)||typeof c.timeoutSeconds!="number"));if(d.length>0)throw new Error(`Generated action registry has actions missing labels, descriptions, or timeout policies: ${d.map(c=>c.id).join(", ")}`);let u=Ce({actionCatalog:s.filter(c=>typeof c.id=="string").map(c=>({id:c.id,name:W(c.name),setupRequirements:ut(c.setupRequirements)??[]})),actionId:e.defaultAction});if(u.blockingRequirements.length>0)throw new Error(`Generated default action ${e.defaultAction} has unresolved required setup and cannot be ready_local: ${u.blockingRequirements.map(Pe).join("; ")}. Mark built-in/local fixture requirements ready or optional when they are satisfied, or leave the pipeline blocked until the setup is configured.`);let m=JSON.parse(await K(n,"utf8"));if(typeof m.summary?.total!="number"||m.summary.total<1)throw new Error("Generated source must define at least one deterministic SDK eval.");if(t&&m.summary.failed!==0)throw new Error("Generated source evals did not all pass.")}o(dt,"assertGeneratedSdkArtifacts");function lt(e){let t=e.trim();if(!t.startsWith("{")&&!t.startsWith("["))return null;try{return JSON.parse(t)}catch{return null}}o(lt,"parseJsonOutput");function E(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}o(E,"asRecord");function ut(e){return Array.isArray(e)?e.filter(t=>!!E(t)):null}o(ut,"recordArray");function mt(e){let t=E(e),a=W(t?.traceArtifactRef);return a?.startsWith(".openpond/")?a:null}o(mt,"traceRefFromCommandResult");function pt(e,t){let a=E(t);return a?e==="inspect"?{actionCount:Array.isArray(a.actionCatalog)?a.actionCatalog.length:null,defaultAction:E(a.agent)?.defaultAction??null,projectName:E(a.project)?.name??null}:e==="validate"?{status:a.status??null,errors:E(a.summary)?.errors??null,warnings:E(a.summary)?.warnings??null}:e==="eval"?E(a.summary):e==="direct-run"?{hasResult:!!E(a.result),resultKeys:Object.keys(E(a.result)??{})}:null:null}o(pt,"summaryFromCommandResult");function W(e){return typeof e=="string"&&e.trim()?e.trim():null}o(W,"stringValue");var gt="Approved local Create plans require model-backed SDK source application; no source mutation was performed.",ft=1800*1e3,Te=".openpond-harness-source-application.json",ht=300*1e3,vt=2*1e3,Rt=5*1e3;async function He(e,t={}){if(!kt(e))return e;let a=t.now??(()=>new Date().toISOString()),n=e.adapter;if(n.kind!=="local")return e;if(!t.session||!t.turn||!t.ensureCodexRuntime||!t.appendRuntimeEvent||!t.setProviderTurnId)return Y(e,{now:a(),reason:gt,metadata:{reason:"model_backed_source_application_required",sourcePath:n.sourcePath??null,repoPath:n.repoPath??null,activeProfile:n.activeProfile??null,operation:e.operation,requestedAgentId:e.target.kind==="agent"?e.target.id:null}});let r;try{r=se(e)}catch(i){return Y(e,{now:a(),reason:i instanceof Error?i.message:String(i),metadata:{reason:"local_create_target_unresolved"}})}try{await ie({snapshot:e,session:t.session,turn:t.turn,target:r,ensureCodexRuntime:t.ensureCodexRuntime,appendRuntimeEvent:t.appendRuntimeEvent,setProviderTurnId:t.setProviderTurnId,model:t.model??we,codexReasoningEffort:t.codexReasoningEffort})}catch(i){return Y(e,{now:a(),reason:i instanceof Error?i.message:String(i),metadata:{reason:"model_backed_source_application_failed",target:r}})}let s=oe(e,{state:"running_checks",now:a(),blockedReason:null,sourceRefs:[r.sourceRootRelativePath,k.join(r.profileRelativePath,"settings","profile.yaml"),"openpond-profile.json"],metadata:{localCreatePipeline:{status:"running_checks",target:r}},candidates:[{id:Se(e.id),target:{kind:"agent",id:r.agentId,displayName:e.target.displayName,defaultActionKey:r.defaultAction},status:"checking",git:null,parentCandidateId:null,tasksetRef:e.tasksetRef,authoringModelRef:t.model??null,allowedPaths:e.plan?.sourcePlan.map(i=>i.path)??[],sourceRefs:[r.sourceRootRelativePath,k.join(r.profileRelativePath,"settings","profile.yaml"),"openpond-profile.json"],artifactRefs:[],checkRefs:[],evaluationReceiptRefs:[],createdAt:e.createdAt,updatedAt:a(),metadata:{sourceAuthority:e.adapter.sourceAuthority}}]});await t.onSnapshot?.(s);try{let i=await(t.runChecks??F)({snapshot:s,target:r}),d=a(),u=Se(e.id),m=`agent_eval_${e.id}`;return oe(s,{state:"ready_local",now:d,blockedReason:null,checkRefs:i.checkRefs,candidates:s.candidates.map(c=>c.id===u?{...c,status:"accepted",checkRefs:X(c.checkRefs,i.checkRefs),evaluationReceiptRefs:X(c.evaluationReceiptRefs,[m]),updatedAt:d}:c),evaluationReceipts:[...s.evaluationReceipts.filter(c=>c.id!==m),{id:m,candidateId:u,target:{kind:"agent",id:r.agentId,displayName:e.target.displayName,defaultActionKey:r.defaultAction},evaluatorKind:"agent_sdk",subject:"candidate",sourceCommit:e.adapter.kind==="local"?e.adapter.localHead:null,sourceBranch:null,tasksetId:e.tasksetRef?.id??null,tasksetHash:e.tasksetRef?.contentHash??null,taskAttemptRefs:i.checkRefs.filter(c=>c.includes("eval")),status:"passed",publishGate:"passed",summaryCounts:null,evalRefs:i.checkRefs.filter(c=>c.includes("eval")),artifactRefs:i.checkRefs,summary:wt(i.metadata),createdAt:d,metadata:i.metadata??{}}],metadata:{localCreatePipeline:{status:"ready_local",target:r,checks:i.metadata??{}}}})}catch(i){return Y(s,{now:a(),reason:i instanceof Error?i.message:String(i),metadata:{reason:"local_create_checks_failed",target:r}})}}o(He,"applyApprovedLocalCreateImproveRun");function kt(e){return e.state==="applying_source"&&e.plan?.status==="approved"&&e.adapter.kind==="local"&&e.target.kind==="agent"}o(kt,"shouldApplyLocalCreatePipeline");function oe(e,t){return A(e,{state:t.state,blockedReason:t.blockedReason===void 0?e.blockedReason:t.blockedReason,checkRefs:X(e.checkRefs,t.checkRefs??[]),sourceRefs:X(e.sourceRefs,t.sourceRefs??[]),candidates:t.candidates??e.candidates,evaluationReceipts:t.evaluationReceipts??e.evaluationReceipts,metadata:{...e.metadata,...t.metadata??{}},updatedAt:t.now})}o(oe,"transitionCreatePipeline");function Y(e,t){return oe(e,{state:"blocked",now:t.now,blockedReason:t.reason,metadata:{localCreatePipeline:{status:"blocked",...t.metadata}}})}o(Y,"blockLocalCreatePipeline");function X(e,t){return Array.from(new Set([...e,...t].filter(Boolean)))}o(X,"mergeRefs");function Se(e){return`agent_candidate_${e}`}o(Se,"agentCandidateId");function wt(e){let t=M(e?.eval),a=typeof t?.total=="number"?t.total:null;return a===null?"Agent SDK checks and deterministic Evals passed.":`${a} Agent SDK Eval${a===1?"":"s"} passed.`}o(wt,"checkSummary");function se(e){let t=e.adapter;if(t.kind!=="local")throw new Error("Create/Improve target must use the local adapter.");if(e.target.kind!=="agent")throw new Error("Local agent source application requires an Agent target.");let a=t.sourcePath?.trim(),n=t.repoPath?.trim();if(!a)throw new Error("Local Create approval is missing a profile source path.");if(!n)throw new Error("Local Create approval is missing a profile repo path.");let r=t.activeProfile?.trim()||"default",s=k.resolve(a),i=k.resolve(n),d=qe(e.target.id)??xt(e)??St(e.objective),u=Tt(Et(e.plan?.defaultChatAction.key)??"chat",d),m=e.operation==="improve"&&d==="default"?s:k.join(s,"agents",d);return{activeProfile:r,agentId:d,defaultAction:u,repoPath:i,sourcePath:s,workspaceRoot:i,profileRelativePath:je(k.relative(i,s)),sourceRoot:m,sourceRootRelativePath:je(k.relative(i,m))}}o(se,"resolveLocalCreatePipelineTarget");async function ie(e){if(!z(e.target.workspaceRoot))throw new Error(`Profile repo path does not exist: ${e.target.workspaceRoot}`);if(!z(e.target.sourcePath))throw new Error(`Profile source path does not exist: ${e.target.sourcePath}`);if(await Ct(e.snapshot,e.target)){await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"server",appId:e.session.appId,status:"completed",output:`Applied prepared harness ${e.snapshot.operation==="improve"?"Improve":"Create"} source for ${e.target.agentId}; running local checks.`,data:{phase:"prepared_source_apply_completed",runId:e.snapshot.id,target:e.target}}));return}await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"server",appId:e.session.appId,status:"started",output:`Applying approved ${e.snapshot.operation==="improve"?"Improve":"Create"} plan with Codex in ${e.target.sourceRootRelativePath}.`,data:{phase:"source_apply_started",runId:e.snapshot.id,target:e.target}}));let t={...e.session,provider:"codex",cwd:e.target.workspaceRoot,workspaceKind:"local_project",workspaceId:e.session.workspaceId??`openpond-profile:${e.target.activeProfile}`,workspaceName:e.session.workspaceName??`OpenPond profile ${e.target.activeProfile}`},a=e.session.provider==="codex"?e.model:null,n=await e.ensureCodexRuntime(t,{approvalPolicy:"never",sandbox:"workspace-write",model:a,codexPermissionMode:"auto-review",codexReasoningEffort:e.codexReasoningEffort}),r=await n.client.startTurn({threadId:n.threadId,prompt:At(e.snapshot,e.target),cwd:e.target.workspaceRoot,model:a,approvalPolicy:"never",sandbox:"workspace-write"});await e.setProviderTurnId(r.turnId);try{await n.client.waitForTurn(r.turnId,ft)}catch(s){if(!$e(s,r.turnId)&&!It(s))throw s;let i=$e(s,r.turnId)?"timed out":"exited cleanly before returning a turn completion";if(await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"server",appId:e.session.appId,status:"started",output:`Codex source application ${i}; checking for completed generated source before blocking.`,data:{phase:"source_apply_recovery_started",runId:e.snapshot.id,providerTurnId:r.turnId,reason:i,target:e.target}})),!await yt(e.target))throw s;await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"server",appId:e.session.appId,status:"completed",output:`Recovered completed ${e.snapshot.operation==="improve"?"Improve":"Create"} source after Codex source application ${i} for ${e.target.agentId}; running local checks.`,data:{phase:"source_apply_recovered",runId:e.snapshot.id,providerTurnId:r.turnId,reason:i,target:e.target}}))}await e.appendRuntimeEvent(H({sessionId:e.session.id,turnId:e.turn.id,name:"create_improve.updated",source:"provider",appId:e.session.appId,status:"completed",output:`Codex completed approved ${e.snapshot.operation==="improve"?"Improve":"Create"} source application for ${e.target.agentId}.`,data:{phase:"source_apply_completed",runId:e.snapshot.id,providerTurnId:r.turnId,target:e.target}}))}o(ie,"runModelBackedLocalCreateSourceApplication");async function Ct(e,t,a=process.env){if(!Ae(a))return!1;let n=k.join(t.workspaceRoot,Te);if(!z(n))return!1;let r=JSON.parse(await Oe(n,"utf8")),s=M(r);if(s?.schema!=="openpond.harnessPreparedSource.v1")throw new Error(`${Te} must use schema openpond.harnessPreparedSource.v1.`);let i=M(s.agents),d=M(i?.[t.agentId]),u=Pt(d?.[e.operation]);if(!u)return!1;let m=ne(t.workspaceRoot,u.source,"source");if(!z(m))throw new Error(`Prepared harness source does not exist: ${u.source}`);await Ee(t.sourceRoot,{recursive:!0}),await xe(m,t.sourceRoot,{recursive:!0,force:!0});for(let c of u.registrations??[]){let g=ne(t.workspaceRoot,c.source,"registration source"),l=ne(t.workspaceRoot,c.target,"registration target");if(!z(g))throw new Error(`Prepared harness registration source does not exist: ${c.source}`);await Ee(k.dirname(l),{recursive:!0}),await xe(g,l,{force:!0})}return!0}o(Ct,"applyPreparedHarnessSourceApplication");function Pt(e){let t=M(e);if(!t||typeof t.source!="string"||!t.source.trim())return null;if(t.registrations!==void 0&&!Array.isArray(t.registrations))throw new Error("Prepared harness source registrations must be an array.");let a=(t.registrations??[]).map(n=>{let r=M(n);if(!r||typeof r.source!="string"||!r.source.trim()||typeof r.target!="string"||!r.target.trim())throw new Error("Prepared harness source registrations require source and target paths.");return{source:r.source,target:r.target}});return{source:t.source,registrations:a}}o(Pt,"asHarnessPreparedSourceOperation");function ne(e,t,a){if(k.isAbsolute(t))throw new Error(`Prepared harness ${a} must be relative to the profile repo.`);let n=k.resolve(e),r=k.resolve(n,t);if(r!==n&&!r.startsWith(`${n}${k.sep}`))throw new Error(`Prepared harness ${a} must stay inside the profile repo.`);return r}o(ne,"resolveHarnessManifestPath");function $e(e,t){return(e instanceof Error?e.message:String(e))===`codex turn ${t} timed out`}o($e,"isCodexTurnTimeoutError");function It(e){return(e instanceof Error?e.message:String(e)).startsWith("codex app-server exited with code 0")}o(It,"isCodexAppServerCleanExitError");async function yt(e){let t=Date.now()+ht;for(;Date.now()<t;){if(await bt(e))return!0;await Le(vt)}return!1}o(yt,"waitForTimedOutSourceApplication");async function bt(e){try{await V(e);let t=await _e(e);await Le(Rt),await V(e);let a=await _e(e);return t===a}catch{return!1}}o(bt,"localCreateSourceLayoutIsStable");async function _e(e){let t=[k.join(e.sourceRoot,"agent","agent.ts"),k.join(e.sourceRoot,"package.json"),k.join(e.repoPath,"openpond-profile.json"),k.join(e.sourcePath,"settings","profile.yaml")];return(await Promise.all(t.map(async n=>{let r=await Oe(n,"utf8");return`${n}:${r.length}:${r}`}))).join(`
 ---
 `)}o(_e,"sourceLayoutFingerprint");function Le(e){return new Promise(t=>setTimeout(t,e))}o(Le,"sleep");function At(e,t){return[`Apply this already-approved OpenPond ${e.operation==="improve"?"Improve":"Create"} plan by editing the local profile repo source.`,"","This is the source-materialization step, not another plan review. Make the file changes, inspect the existing profile shape first, and leave the repo with a valid openpond-agent-sdk source project.","","Hard requirements:",`- Current working directory is the profile repo root: ${t.workspaceRoot}`,`- Active profile source path relative to cwd: ${t.profileRelativePath}`,`- Target agent id: ${t.agentId}`,`- Target generated source root: ${t.sourceRootRelativePath}`,`- Default action to expose: ${t.defaultAction}`,`- For a create request, create the new SDK project under ${t.profileRelativePath}/agents/<agent-id>; do not overwrite ${t.profileRelativePath}/agent/agent.ts or unrelated existing agents.`,"- For an edit request, only edit the approved target agent source and shared profile registration files required for that edit.",`- Update openpond-profile.json and ${t.profileRelativePath}/settings/profile.yaml so the generated agent is registered, enabled, and catalog-discoverable.`,"- The generated source must be an actual openpond-agent-sdk project with package.json, tsconfig.json when needed, agent/agent.ts, deterministic evals, and action metadata.","- Preserve the approved objective and any captured conversation context in source metadata, instructions, fixture comments, or eval intent so source readback can prove what was captured.","- Treat plan.metadata.actionShape as the approved public action-shape decision when actionShapeDecisionSource is request_metadata or another explicit planner source.","- If plan.metadata.actionShapeDecisionSource is default_chat_fallback, use the approved objective and captured context to choose the actual generated public actions: keep chat-only for conversational assistants, create direct actions for explicit repeatable runs/workflows/artifact transforms, and expose both when the user needs normal chat plus a repeatable catalog action.","- When refining a default fallback during source materialization, encode the chosen public actions in generated SDK action metadata and evals; do not use business-example hardcoding.","- Every public action must have a stable action id, label, description, input/output schema when useful, timeout policy, artifact policy, setup requirement metadata, and deterministic eval coverage.","- The default action must be invocable through the local profile setup gate after checks. Do not leave built-in openpond_chat or committed local fixture/artifact requirements as unresolved required setup; mark those rows ready/satisfied or optional when the generated source already provides them.","- If the approved objective names multiple capabilities or direct action ids, implement each requested public action instead of collapsing everything into a single chat-only action.","- If a capability writes an artifact, declare it in the action artifact policy and assert it from at least one eval.","- Package scripts must be portable from the generated source root; do not hardcode absolute machine-local paths to this repository or CLI.","- Do not hardcode secrets, API keys, provider tokens, or user credentials in source, evals, traces, fixtures, or comments.","- Use realistic deterministic fixture-backed behavior when external systems are unavailable. Declare setup requirements rather than pretending live integrations exist.","- Run or prepare the source so these commands can pass from the target source root: inspect, build, validate, eval, and run of the default action.","","Approved Create request and plan:",JSON.stringify({run:e,plan:e.plan,questions:e.questions,workflowCapture:e.workflowCapture},null,2)].join(`
 `)}o(At,"localCreatePipelinePrompt");function xt(e){for(let t of e.plan?.sourcePlan??[]){let a=/^agents\/([^/]+)/.exec(t.path),n=qe(a?.[1]);if(n)return n}return null}o(xt,"agentIdFromSourcePlan");function qe(e){let t=e?.trim();return t&&t.toLowerCase().replace(/[^a-z0-9._-]+/g,"-").replace(/^-+|-+$/g,"")||null}o(qe,"normalizeAgentId");function Et(e){return e?.trim()||null}o(Et,"normalizeActionId");function M(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}o(M,"asRecord");function Tt(e,t){let a=`${t}.`;return e.startsWith(a)?e.slice(a.length)||"chat":e}o(Tt,"sourceActionIdForAgent");function St(e){return e.trim().toLowerCase().replace(/[^a-z0-9]+/g,"-").replace(/^-+|-+$/g,"").slice(0,64).replace(/-+$/g,"")||"generated-agent"}o(St,"slugFromObjective");function je(e){return e.replace(/\\/g,"/")||"."}o(je,"normalizeRelativePath");import de from"node:path";function Pa(e){let t=e.sources.length?e.sources.map(r=>({kind:"conversation",id:r.id,sourceHash:r.sourceHash,excerptRef:`training-source:${r.id}`})):[{kind:"manual_intent",id:`manual_intent_${$(e.objective).slice(0,24)}`,sourceHash:$({objective:e.objective}),excerptRef:null}],a={consent:{status:"granted",scope:e.sources.length?"selected_turns":"direct_intent",reviewedBy:"local_user",reviewedAt:e.timestamp},sources:t,reviewerIntent:e.objective},n=$(a);return re.parse({schemaVersion:"openpond.createImprove.evidenceSnapshot.v1",id:`evidence_snapshot_${n.slice(0,24)}`,contentHash:n,...a,createdAt:e.timestamp,metadata:{immutable:!0,sourceCount:t.length}})}o(Pa,"createEvidenceSnapshot");function Ia(e){let t={runId:e.run.id,candidateId:e.candidateId,tasksetId:e.run.tasksetRef?.id??null,tasksetRevision:e.run.tasksetRef?.revision??null,tasksetHash:e.run.tasksetRef?.contentHash??null,outcome:e.outcome,reason:e.reason,receiptRefs:e.receiptRefs},a=$(t);return re.parse({schemaVersion:"openpond.createImprove.evidenceSnapshot.v1",id:`evidence_outcome_${a.slice(0,24)}`,contentHash:a,consent:{status:"granted",scope:"direct_intent",reviewedBy:e.run.actor.id??"local_user",reviewedAt:e.timestamp},sources:[{kind:"artifact",id:`${e.outcome}:${e.run.id}:${e.candidateId}`,sourceHash:a,excerptRef:e.receiptRefs[0]??null}],reviewerIntent:e.reason,createdAt:e.timestamp,metadata:{immutable:!0,evidenceKind:"candidate_outcome",outcome:e.outcome,priorTasksetRevision:e.run.tasksetRef?.revision??null,recommendedNextTasksetRevision:e.run.tasksetRef?e.run.tasksetRef.revision+1:1,receiptRefs:e.receiptRefs}})}o(Ia,"createOutcomeEvidenceSnapshot");function ya(e){let t=e.proposal?$t(e.proposal):{kind:"model",rationale:["A reviewed training plan selected Model execution for this Taskset."],confidence:1};return ye.parse({schemaVersion:"openpond.createImprove.tasksetRef.v1",id:e.taskset.id,revision:e.taskset.revision,contentHash:e.taskset.contentHash,evidenceSnapshotIds:e.evidenceSnapshotIds,policyBoundary:e.taskset.policy,targetRecommendation:t,authoringSplitRefs:e.taskset.tasks.filter(a=>a.split==="train"||a.split==="validation").map(a=>a.id),privateSplitRefs:e.taskset.tasks.filter(a=>a.split==="test"||a.split==="frozen_eval").map(a=>a.id),approvedBy:"local_user",approvedAt:e.approvedAt,metadata:{immutable:!0,taskCreationProposalId:e.proposal?.id??null}})}o(ya,"createTasksetRef");function Me(e,t){if(e.id!==t.id||e.revision!==t.revision||e.contentHash!==t.contentHash)throw new Error(`Taskset lineage mismatch: the run approved ${e.id}@${e.revision} (${e.contentHash}), but execution resolved ${t.id}@${t.revision} (${t.contentHash}).`)}o(Me,"assertTasksetRefMatches");function $t(e){let t=_t(e);return{kind:t,rationale:e.diagnosis.rationale.length?e.diagnosis.rationale:[`${t} best matches the reviewed behavior boundary.`],confidence:e.diagnosis.confidence}}o($t,"recommendationForProposal");function _t(e){return e.diagnosis.trainingEligible?"model":e.diagnosis.intervention==="retrieval"?"configuration":"agent"}o(_t,"targetKindForProposal");async function Ge(e){let t=e.run.tasksetRef;if(!t)throw new Error("Agent Taskset execution requires an approved Taskset ref.");if(Me(t,e.taskset),e.taskset.environment.kind!=="agent")throw new Error(`Taskset ${e.taskset.id} is not an Agent-shaped Taskset.`);let a=new Set(t.privateSplitRefs),n=e.taskset.tasks.filter(c=>a.has(c.id));if(!n.length||n.length!==a.size)throw new Error("The approved Agent Taskset private split could not be resolved exactly.");let r={tasksetId:t.id,tasksetRevision:t.revision,tasksetHash:t.contentHash,taskRefs:n.map(c=>c.id),seed:0,timeoutMs:e.taskset.environment.defaultTimeoutMs,entrypoint:e.taskset.environment.entrypoint,evaluator:"openpond-agent-sdk-taskset.v1"},s=$(r),i=[],d=[],u=[];for(let c of n){let g=new Date().toISOString(),l=Date.now(),p=await(e.execute??O)({cwd:e.cwd,command:"run",args:[e.taskset.environment.entrypoint,"--json","--input",JSON.stringify({...c.input,channel:"manual"})],timeoutMs:e.taskset.environment.defaultTimeoutMs,throwOnFailure:!1}),x=p.code===0?jt(p.stdout):null,b=`agent_task_attempt_${$([e.run.id,e.subject,e.sourceCommit,t.contentHash,c.id,0]).slice(0,24)}`,f=Ot(x?.traceArtifactRef),P=Ie.parse({schemaVersion:"openpond.taskAttempt.v1",id:b,tasksetId:e.taskset.id,taskId:c.id,split:c.split,attempt:0,seed:0,modelRef:null,startedAt:g,completedAt:new Date().toISOString(),output:ce(x?.result),runtimeEventRefs:f?[f]:[],artifactRefs:f?[f]:[],privilegedOutcomeRef:c.privilegedContextRef,infrastructureError:p.code===0?null:p.stderr.trim().slice(0,1e4)||`Agent SDK action exited with code ${String(p.code)}.`,costUsd:0,latencyMs:Math.max(0,Date.now()-l),userInterventions:0,metadata:{executor:"openpond-agent-sdk",trustedTasksetExecution:!0,tasksetRevision:t.revision,tasksetHash:t.contentHash,sourceCommit:e.sourceCommit,subject:e.subject,executionContractHash:s}}),R=await e.gradeAttempt({tasksetId:e.taskset.id,taskId:c.id,attempt:P});d.push(P.id),i.push(R),f&&u.push(f)}let m=i.filter(c=>c.passed).length;return{taskAttemptRefs:d,gradeRefs:i.map(c=>c.id),taskRefs:n.map(c=>c.id),artifactRefs:[...new Set(u)],total:i.length,passed:m,failed:i.length-m,executionContractHash:s}}o(Ge,"executeAgentTasksetEvaluation");function jt(e){let t=e.trim();if(!t)return null;try{return ce(JSON.parse(t))}catch{let a=t.indexOf("{"),n=t.lastIndexOf("}");if(a<0||n<=a)return null;try{return ce(JSON.parse(t.slice(a,n+1)))}catch{return null}}}o(jt,"parseAgentRunPayload");function ce(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:{}}o(ce,"recordValue");function Ot(e){return typeof e=="string"&&e.trim()?e.trim():null}o(Ot,"stringValue");async function le(e){let t=await(e.execute??O)({cwd:e.cwd,command:"eval",args:["--json"],throwOnFailure:!1}),a=qt(t.stdout),n=G(a.summary?.total)??0,r=G(a.summary?.passed)??0,s=G(a.summary?.failed)??Math.max(0,n-r),i=a.publishGate?.status==="passed"?"passed":"failed",d=Array.isArray(a.results)?a.results:[],u=d.map(R=>I(R.name)).filter(R=>!!R),m=new Set([de.posix.join(e.sourceRef,".openpond/eval-results.json")]);for(let R of d){let _=I(R.traceArtifactRef);_&&m.add(de.posix.join(e.sourceRef,Ne(_)));for(let J of De(R.artifacts))m.add(de.posix.join(e.sourceRef,Ne(J)))}let c=e.run.tasksetRef?await Ge({run:e.run,taskset:Ht(e.taskset),cwd:e.cwd,sourceCommit:e.sourceCommit,subject:e.subject,gradeAttempt:Lt(e.gradeAttempt),execute:e.execute}):null;for(let R of c?.artifactRefs??[])m.add(R);let g=n+(c?.total??0),l=r+(c?.passed??0),p=s+(c?.failed??0),x=[...new Set([...u,...c?.taskRefs??[]])],b=i==="passed"&&(c?.failed??0)===0?"passed":"failed",f=t.code===0&&p===0&&b==="passed"?"passed":"failed",P=e.timestamp??new Date().toISOString();return{id:`agent_eval_${e.run.id}_${e.subject}_${e.sourceCommit.slice(0,12)}`,candidateId:e.candidateId,target:e.run.target,evaluatorKind:"agent_sdk",subject:e.subject,sourceCommit:e.sourceCommit,sourceBranch:e.sourceBranch,tasksetId:e.run.tasksetRef?.id??null,tasksetHash:e.run.tasksetRef?.contentHash??null,taskAttemptRefs:c?.taskAttemptRefs??u,status:f,publishGate:b,summaryCounts:{total:g,passed:l,failed:p},evalRefs:x,artifactRefs:[...m],summary:`${e.subject==="active"?"Active":e.subject==="candidate"?"Candidate":"Post-release"} Agent SDK and Taskset Evals: ${l}/${g} passed${b==="failed"?"; publish gate failed":""}.`,createdAt:P,metadata:{schemaVersion:I(a.schemaVersion),schema:I(a.schema),project:{name:I(a.project?.name),version:I(a.project?.version)},source:{configPath:I(a.source?.configPath),configHash:I(a.source?.configHash)},publishGate:{total:G(a.publishGate?.total),passed:G(a.publishGate?.passed),failed:G(a.publishGate?.failed),blockingFailures:De(a.publishGate?.blockingFailures)},command:{code:t.code,timedOut:t.timedOut,stderr:t.stderr.trim().slice(0,4e3)},trustedTasksetExecution:!!c,tasksetRevision:e.run.tasksetRef?.revision??null,executionContractHash:c?.executionContractHash??null,gradeRefs:c?.gradeRefs??[],results:d.map(R=>({name:I(R.name),status:I(R.status),error:I(R.error),traceArtifactRef:I(R.traceArtifactRef)}))}}}o(le,"runNormalizedAgentEvaluation");function Ht(e){if(!e)throw new Error("The approved Agent Taskset could not be resolved for evaluation.");return e}o(Ht,"requireTaskset");function Lt(e){if(!e)throw new Error("Trusted Taskset grading is unavailable for Agent evaluation.");return e}o(Lt,"requireGradeAttempt");function qt(e){let t=e.trim();if(!t)throw new Error("Agent SDK Eval returned no JSON output.");try{return JSON.parse(t)}catch{let a=t.indexOf("{"),n=t.lastIndexOf("}");if(a>=0&&n>a)try{return JSON.parse(t.slice(a,n+1))}catch{}throw new Error("Agent SDK Eval returned invalid JSON output.")}}o(qt,"parseAgentEvalPayload");function Ne(e){return e.replace(/^\.\//,"").replaceAll("\\","/")}o(Ne,"normalizeRef");function I(e){return typeof e=="string"&&e.trim()?e.trim():null}o(I,"stringValue");function G(e){return typeof e=="number"&&Number.isFinite(e)?e:null}o(G,"numberValue");function De(e){return Array.isArray(e)?e.map(I).filter(t=>!!t):[]}o(De,"stringArray");import{existsSync as Z}from"node:fs";import{mkdir as Xe,realpath as Qe,rm as he,writeFile as Ft}from"node:fs/promises";import Ze from"node:os";import S from"node:path";import{existsSync as ue}from"node:fs";import{cp as Mt,lstat as Gt,mkdir as Be,rm as Nt,symlink as Dt,writeFile as Bt}from"node:fs/promises";import w from"node:path";function Ue(e,t){let a=C(t.profileRelativePath),n=[C(t.sourceRootRelativePath),w.posix.join(a,"evals"),w.posix.join(a,"settings"),w.posix.join(a,"package.json"),w.posix.join(a,"tsconfig.json"),w.posix.join(a,"openpond.lock"),w.posix.join(a,".gitignore"),"openpond-profile.json"],r=(e.plan?.sourcePlan??[]).flatMap(s=>{let i=C(s.path);return i?i==="."||i.startsWith(`${a}/`)?[i]:[w.posix.join(a,i),i]:[]});return[...new Set([...n,...r])]}o(Ue,"allowedAgentImprovementPaths");function Je(e,t){let a=C(e);return t.some(n=>a===n||a.startsWith(`${n}/`))}o(Je,"pathAllowed");function me(e){return C(e).split("/").some(a=>a==="node_modules"||a===".openpond"||a==="artifacts")}o(me,"isCandidateRuntimePath");function Fe(e){return e.split(/\r?\n/).map(t=>t.trimEnd()).filter(Boolean).map(t=>{let a=t.slice(3),n=a.split(" -> ");return C(n.at(-1)??a)})}o(Fe,"parseGitStatusPaths");async function pe(e,t,a){let n=await e("git",["diff","--name-only","-z",a,"--","."],t);return h(n,"Unable to inspect Agent candidate paths."),[...new Set(n.stdout.split("\0").map(C).filter(r=>r!=="."&&!me(r)))]}o(pe,"gitChangedPaths");async function ge(e){for(let t of e.paths){let a=await B(e.command,e.repoPath,["rev-parse",`${e.commit}:${t}`]),n=w.join(e.repoPath,...t.split("/"));if(!a){if(ue(n))return!1;continue}if(!ue(n))return!1;let r=await B(e.command,e.repoPath,["hash-object","--",t]);if(!r||r!==a)return!1}return!0}o(ge,"workingTreeMatchesCommit");async function ze(e,t){let a=new Set(["node_modules",w.posix.join(C(e.profileRelativePath),"node_modules"),w.posix.join(C(e.sourceRootRelativePath),"node_modules")]);for(let n of a){let r=w.join(e.repoPath,...n.split("/")),s=w.join(t,...n.split("/"));!await Jt(r)||ue(s)||(await Be(w.dirname(s),{recursive:!0}),await Dt(r,s,process.platform==="win32"?"junction":"dir"))}}o(ze,"linkCandidateDependencies");async function We(e){let t=await e.command("git",["diff","--binary","HEAD","--",".",":(exclude)**/tasksets/**"],e.repoPath);if(h(t,"Unable to capture active Profile changes."),t.stdout){let i=w.join(e.workspaceRoot,"active-profile.patch");await Bt(i,t.stdout,"utf8");let d=await e.command("git",["apply","--whitespace=nowarn",i],e.worktreePath);h(d,"Unable to apply active Profile changes to the candidate baseline."),await Nt(i,{force:!0})}let a=await e.command("git",["ls-files","--others","--exclude-standard","-z"],e.repoPath);h(a,"Unable to inspect untracked Profile source.");for(let i of a.stdout.split("\0").filter(Ut)){let d=w.join(e.repoPath,...i.split("/")),u=w.join(e.worktreePath,...i.split("/"));await Be(w.dirname(u),{recursive:!0}),await Mt(d,u,{recursive:!0,preserveTimestamps:!0})}let n=await e.command("git",["add","-A"],e.worktreePath);h(n,"Unable to stage the active Profile snapshot.");let r=await e.command("git",["diff","--cached","--quiet"],e.worktreePath);if(r.code===0)return;r.code!==1&&h(r,"Unable to inspect the active Profile snapshot.");let s=await e.command("git",["commit","-m","Snapshot active Profile before OpenPond improvement"],e.worktreePath,Q());h(s,"Unable to commit the active Profile snapshot.")}o(We,"snapshotActiveProfileChanges");function Ut(e){return e?!C(e).split("/").some(a=>a==="node_modules"||a===".openpond"||a==="artifacts"||a==="tasksets"):!1}o(Ut,"snapshotEligiblePath");function Q(){return{GIT_AUTHOR_NAME:process.env.GIT_AUTHOR_NAME||"OpenPond",GIT_AUTHOR_EMAIL:process.env.GIT_AUTHOR_EMAIL||"openpond@example.invalid",GIT_COMMITTER_NAME:process.env.GIT_COMMITTER_NAME||"OpenPond",GIT_COMMITTER_EMAIL:process.env.GIT_COMMITTER_EMAIL||"openpond@example.invalid"}}o(Q,"gitIdentityEnv");async function Jt(e){try{return(await Gt(e)).isDirectory()}catch{return!1}}o(Jt,"directoryExists");async function Ke(e,t,a){let n=await e("gh",["pr","list","--head",a,"--state","all","--json","number,url,state,mergedAt,mergeCommit,baseRefName,headRefName","--limit","20"],t);h(n,"Unable to inspect existing Agent improvement pull requests.");let r=JSON.parse(n.stdout);return Array.isArray(r)?r.filter(s=>!!(s&&typeof s=="object")):[]}o(Ke,"listPullRequestsForBranch");function fe(e,t=new Date().toISOString(),a=new Date().toISOString()){let n=typeof e.number=="number"?e.number:Number(e.number),r=typeof e.state=="string"?e.state.toLowerCase():"",i=N(e.mergedAt)||r==="merged"?"merged":r==="open"?"open":"closed",d=e.mergeCommit&&typeof e.mergeCommit=="object"?e.mergeCommit:null,u=N(d?.oid)??N(e.mergeCommit),m=N(e.url),c=N(e.baseRefName),g=N(e.headRefName);if(!Number.isInteger(n)||n<=0||!m||!c||!g)throw new Error("GitHub returned an incomplete pull request payload.");return{provider:"github",number:n,url:m,state:i,baseBranch:c,headBranch:g,mergeCommit:u,openedAt:t,updatedAt:a}}o(fe,"normalizePullRequest");function Ve(e,t){return`openpond/improve/${D(t)}/${D(e.id).slice(0,36)}`}o(Ve,"candidateBranch");function U(e){if(!e.worktreePath)throw new Error("Agent candidate worktree is unavailable.");return e.worktreePath}o(U,"requireWorktreePath");function Ye(e){let t=e.replace(/^git@github\.com:/i,"").replace(/^https?:\/\/github\.com\//i,"").replace(/\.git$/i,"").replace(/^\/+|\/+$/g,"");return/^[^/]+\/[^/]+$/.test(t)?t:null}o(Ye,"githubRepoFromRemote");function C(e){return e.replaceAll("\\","/").replace(/^\.\//,"").replace(/^\/+|\/+$/g,"")||"."}o(C,"normalizeRepoPath");function D(e){return e.toLowerCase().replace(/[^a-z0-9._-]+/g,"-").replace(/^-+|-+$/g,"")||"candidate"}o(D,"safePathSegment");async function y(e,t,a,n){let r=await e("git",a,t);h(r,n);let s=r.stdout.trim();if(!s)throw new Error(n);return s}o(y,"requiredGitText");async function B(e,t,a){let n=await e("git",a,t);return n.code===0&&n.stdout.trim()?n.stdout.trim():null}o(B,"optionalGitText");function h(e,t){if(e.code!==0)throw new Error(e.stderr.trim()||e.stdout.trim()||t)}o(h,"assertCommand");function N(e){return typeof e=="string"&&e.trim()?e.trim():null}o(N,"stringValue");function T(e,t,a,n={}){return be(e,t,a,n)}o(T,"defaultCommandRunner");function et(e,t){let a=U(t);if(!Z(a))throw new Error("The existing Agent candidate worktree is no longer available.");let n=C(e.profileRelativePath),r=C(e.sourceRootRelativePath);return{target:{...e,repoPath:a,workspaceRoot:a,sourcePath:S.join(a,...n.split("/")),sourceRoot:S.join(a,...r.split("/"))},git:t}}o(et,"restoreAgentImprovementWorkspace");async function tt(e){let t=e.command??T,a=await y(t,e.target.repoPath,["rev-parse","--show-toplevel"],"Profile source is not a Git repository."),[n,r]=await Promise.all([Qe(a),Qe(e.target.repoPath)]);if(n!==r)throw new Error("Profile repo path does not match the Git repository root.");let s=await t("git",["status","--porcelain=v1","-uall"],a);h(s,"Unable to inspect Profile Git status.");let[i,d,u]=await Promise.all([y(t,a,["rev-parse","HEAD"],"Profile Git repo has no HEAD commit."),y(t,a,["branch","--show-current"],"Profile Git repo must be on a branch."),B(t,a,["remote","get-url","origin"])]);if(e.run.adapter.kind==="local"&&e.run.adapter.localHead&&e.run.adapter.localHead!==i)throw new Error(`Profile source changed from ${e.run.adapter.localHead.slice(0,12)} to ${i.slice(0,12)}. Re-plan the improvement against the current source.`);let m=Ve(e.run,e.target.agentId),c=S.join(Ze.tmpdir(),"openpond-agent-improvements",D(S.basename(a)||"profile"),D(e.run.id)),g=S.join(c,"repo");await Xe(c,{recursive:!0});let l=await t("git",["show-ref","--verify","--quiet",`refs/heads/${m}`],a);if(Z(g)&&await B(t,g,["branch","--show-current"])!==m&&(await t("git",["worktree","remove","--force",g],a),await he(c,{recursive:!0,force:!0}),await Xe(c,{recursive:!0})),!Z(g)){let P=l.code===0?["worktree","add",g,m]:["worktree","add","-b",m,g,i],R=await t("git",P,a);h(R,"Unable to create the Agent candidate worktree.")}s.stdout.trim()&&await We({command:t,repoPath:a,worktreePath:g,workspaceRoot:c});let p=await y(t,g,["rev-parse","HEAD"],"Agent candidate worktree has no baseline commit.");if(await y(t,g,["rev-parse","HEAD"],"Agent candidate worktree has no HEAD commit.")!==p)throw new Error("Agent candidate baseline could not be resolved.");await ze(e.target,g);let b=C(e.target.profileRelativePath),f=C(e.target.sourceRootRelativePath);return{target:{...e.target,repoPath:g,workspaceRoot:g,sourcePath:S.join(g,...b.split("/")),sourceRoot:S.join(g,...f.split("/"))},git:{baseBranch:d,baseCommit:p,branch:m,headCommit:null,remoteName:"origin",remoteUrl:u,worktreePath:g,changedPaths:[],diffStat:null,pullRequest:null}}}o(tt,"prepareAgentImprovementWorkspace");async function at(e){let t=e.command??T,a=U(e.workspace.git),n=await t("git",["status","--porcelain=v1","-uall"],a);h(n,"Unable to inspect Agent candidate changes.");let r=Fe(n.stdout).filter(g=>!me(g));if(r.length===0)throw new Error("The Agent improvement produced no source changes.");let s=Ue(e.run,e.activeTarget),i=r.filter(g=>!Je(g,s));if(i.length>0)throw new Error(`Agent candidate changed paths outside the approved scope: ${i.join(", ")}`);let d=await t("git",["add","--",...r],a);h(d,"Unable to stage Agent candidate changes.");let u=await t("git",["commit","-m",`Improve ${e.run.target.displayName??e.activeTarget.agentId} via OpenPond`],a,Q());h(u,"Unable to commit Agent candidate changes.");let m=await y(t,a,["rev-parse","HEAD"],"Agent candidate commit was not created."),c=await B(t,a,["diff","--stat",`${e.workspace.git.baseCommit}..${m}`]);return{...e.workspace.git,headCommit:m,changedPaths:r,diffStat:c}}o(at,"commitAgentImprovementCandidate");async function Va(e){let t=e.command??T,a=U(e.git);if(!e.git.headCommit)throw new Error("Agent candidate has no commit to push.");if(!e.git.remoteUrl||!Ye(e.git.remoteUrl))throw new Error("Agent candidate is ready locally. Configure a GitHub origin remote before opening a pull request.");let n=await t("git",["push","--set-upstream",e.git.remoteName,e.git.branch],a);h(n,"Unable to push the Agent candidate branch.");let r=await Ke(t,a,e.git.branch);if(r.length>0)return fe(r[0],e.timestamp);let s=`Improve ${e.run.target.displayName??e.run.target.id??"Agent"}`,i=["## OpenPond Agent improvement","",e.run.objective,"","## Evaluation","",e.evaluationSummary,"",`Create/Improve run: \`${e.run.id}\``,`Base commit: \`${e.git.baseCommit}\``,`Candidate commit: \`${e.git.headCommit}\``].join(`

package/dist/chunks/{chunk-PACLKDTN.js → chunk-V4VM2R6D.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{b}from"./chunk-JQOVC35K.js";import{ja as M}from"./chunk-HG2YFMVU.js";import{Qb as x,m as T,v as A}from"./chunk-D6L6UNUQ.js";import{n as c,o as S}from"./chunk-RCT3O3HQ.js";import{A as I}from"./chunk-X2J2P5ZT.js";import{a as R}from"./chunk-FTQLQJQV.js";function j(e,a,i){if(a==="agent_action"){let n=e.find(u=>u.name==="workspace_action_result"&&u.action==="profile_workflow_action")?.output;if(typeof n!="string")throw new Error("Profile Agent action evaluation did not produce a released action result.");return n}if(i==="completed"){let n=[...e].reverse().find(s=>s.name==="turn.completed"),r=n?.data?.providerResponse,o=n?.data;if(!r&&o?.turn?.status==="completed"){let s=[...e].reverse().find(t=>P(t)!==null);if(!s)throw new Error("Native Profile evaluation has no retained final answer.");let d=P(s);return e.filter(t=>P(t)===d).map(t=>t.output??"").join("")}if(typeof r?.requestId!="string"||!r.requestId||typeof r.contentHash!="string")throw new Error("Profile evaluation has no retained final provider response.");let p=e.filter(s=>s.name==="assistant.delta"&&s.data?.providerRequestId===r.requestId).map(s=>s.output??"").join("");if(c(p)!==r.contentHash)throw new Error("Profile evaluation final provider response differs from its retained trace.");return b(e)??p}return e.filter(n=>n.name==="assistant.delta"&&typeof n.output=="string").map(n=>n.output??"").join("")}R(j,"profileEvaluationOutput");function P(e){let a=e.data;return e.name==="assistant.delta"&&a?.provider==="codex"&&a.phase==="final_answer"&&typeof a.itemId=="string"?a.itemId:null}R(P,"nativeFinalAnswerItemId");async function _(e){let{store:a,manifest:i,receipt:n}=e;if(!n.metadata.retainedEvidenceRef)return null;let u=x.parse(n.metadata.retainedEvidenceRef),[r,o]=await Promise.all([a.getSession(u.sessionId),a.getTurn(u.turnId)]),p=r?.metadata?.profileEvaluationRun,s=i.profileEvaluation;if(!r||!o||o.sessionId!==r.id||!s||!r.currentProfile||c(r.currentProfile)!==c(e.profileRef)||c(p)!==c({id:i.id,contentHash:i.contentHash})||r.metadata?.taskId!==n.taskId||r.metadata?.seed!==n.seed||!o.completedAt||o.completedAt!==n.completedAt||o.startedAt!==n.startedAt||!o.harnessSnapshot||o.harnessSnapshot.harnessRelease.id!==s.harnessRelease.id||o.harnessSnapshot.harnessRelease.contentHash!==s.harnessRelease.contentHash||i.policy.kind!=="model"||!o.modelRef||o.modelRef.providerId!==i.policy.model.provider||o.modelRef.modelId!==i.policy.model.model)throw new Error("Retained Profile policy turn differs from its admitted case.");let d=await a.runtimeEventsForTurn(o.id);if(d.some(v=>v.turnId!==o.id)||c(d)!==n.traceHash)throw new Error("Retained Profile policy trace differs from its immutable receipt.");let t=j(d,s.target.kind,o.status);if(n.outputHash!==null&&c({text:t})!==n.outputHash)throw new Error("Retained Profile output differs from its immutable receipt.");let m=n.artifactRefs.flatMap(v=>{let y=d.filter(l=>(l.action==="work_output_save"||l.action==="sandbox_save_output")&&l.status==="completed").flatMap(l=>{let C=l.data&&typeof l.data=="object"?l.data:{},k=M.safeParse(C.outputRef);return k.success?[k.data]:[]}).filter(l=>l.sourceTaskId===r.id&&l.sourceTurnId===o.id&&`${r.id}/${l.id}/${l.revision}/${l.title}`===v.id&&l.sha256===v.contentHash&&l.sizeBytes===v.sizeBytes&&l.contentType===v.mediaType);if(y.length!==1)throw new Error("The case output has no unique sealed saved-file identity.");let h=y[0];if(h.location.kind!=="managed")return[];let H=`/api/work/outputs/${encodeURIComponent(h.location.fileId)}`;if(h.location.downloadPath!==H)throw new Error("The saved output has no canonical managed download route.");return[{id:h.id,title:h.title,revision:h.revision,contentType:h.contentType,sizeBytes:h.sizeBytes,sha256:h.sha256,downloadPath:H}]}),f=await a.listModelUsageRecords({sessionId:r.id,turnId:o.id,limit:1e4}),E=e.eventAfterId?d.findIndex(v=>v.id===e.eventAfterId):-1;if(e.eventAfterId&&E<0)throw new Error("Profile trace cursor is unavailable in this case.");let g=E+1,q=Math.max(1,Math.min(250,e.eventLimit??100)),w=d.slice(g,g+q);return{ref:u,prompt:o.prompt,output:n.outputHash===null?null:{text:t},partialOutput:n.outputHash===null&&t?{text:t}:null,error:o.error,events:w,eventCount:d.length,nextEventCursor:g+w.length<d.length?w.at(-1).id:null,files:m,usage:f,usageCoverage:f.length===1e4?"partial":"complete"}}R(_,"readProfileEvaluationPolicyEvidence");var W=I.object({runId:I.string().trim().min(1).max(240),receiptId:I.string().trim().min(1).max(500).optional(),eventAfterId:I.string().trim().min(1).max(500).optional(),eventLimit:I.number().int().min(1).max(250).optional(),manifest:A.optional()}).strict().refine(e=>e.receiptId||!e.eventAfterId&&!e.eventLimit,"Select a case before requesting its trace page.").refine(e=>!e.manifest||e.receiptId,"Interrupted-run inspection requires one exact case.");async function D(e){let a=await e.store.getProfileEvaluationRun(e.runId);if(a&&c(a.profileRef)!==c(e.profileRef))throw new Error("Profile evaluation run is unavailable in the selected Profile.");let i=a?.manifest??e.manifest;if(!i||!a&&!e.receiptId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");if(a){if(e.manifest&&e.manifest.contentHash!==a.manifest.contentHash)throw new Error("Profile evaluation run differs from its pinned manifest.")}else if(S(i,"Interrupted Profile evaluation manifest"),i.id!==e.runId||i.profileEvaluation?.profileId!==e.profileRef.profileId||e.profileRef.source==="openpond_git"&&i.metadata.profileRepositoryId!==e.profileRef.repositoryId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");let n=i.population.map((r,o)=>({member:r,index:o})).filter(({member:r})=>!e.receiptId||r.receiptId===e.receiptId);if(e.receiptId&&n.length!==1)throw new Error("Profile evaluation case is unavailable in the selected run.");let u=await Promise.all(n.map(async({member:r,index:o})=>{let p=a?null:await e.store.getProfileEvaluationReceipt(r.receiptId),s=a?.receiptRefs[o]??(p?{id:p.id,contentHash:p.contentHash}:null),d=a?.gradeRefs[o]??p?.graderEvidenceRefs[0];if(!s||!d)throw new Error("Profile evaluation run has incomplete case evidence.");let[t,m]=await Promise.all([e.store.getProfileEvaluationReceipt(s.id),e.store.getProfileEvaluationGrade(d.contentHash)]);if(!t||!a&&!T(t)||t.contentHash!==s.contentHash||t.id!==r.receiptId||t.taskId!==r.taskId||t.seed!==r.seed||t.runManifest.id!==i.id||t.runManifest.contentHash!==i.contentHash||!m||m.contentHash!==d.contentHash||!t.graderEvidenceRefs.some(f=>f.id===d.id&&f.contentHash===d.contentHash))throw new Error("Profile evaluation case evidence differs from its retained run.");return{taskId:r.taskId,seed:r.seed,receiptId:t.id,receiptHash:t.contentHash,traceHash:t.traceHash,outputHash:t.outputHash,score:m.score,passed:m.passed,gradingStatus:m.gradingStatus,failureClass:m.failureClass,artifactCount:t.artifactRefs.length,latencyMs:t.latencyMs,costUsd:t.costUsd,terminal:t.terminal,feedback:m.components.map(f=>({graderId:f.graderId,graderVersion:f.graderVersion,score:f.score,passed:f.passed,feedback:f.feedback})),...e.receiptId?{receipt:t,evidence:await _({store:e.store,profileRef:e.profileRef,manifest:i,receipt:t,eventAfterId:e.eventAfterId,eventLimit:e.eventLimit})}:{}}}));return{runId:i.id,sourceRevision:i.profileEvaluation?.sourceRevision??null,cases:u}}R(D,"inspectProfileEvaluationRun");export{j as a,W as b,D as c};
+import{b}from"./chunk-UCPD526R.js";import{ja as M}from"./chunk-2THZ6GZY.js";import{Qb as x,m as T,v as A}from"./chunk-D6L6UNUQ.js";import{n as c,o as S}from"./chunk-RCT3O3HQ.js";import{A as I}from"./chunk-X2J2P5ZT.js";import{a as R}from"./chunk-FTQLQJQV.js";function j(e,a,i){if(a==="agent_action"){let n=e.find(u=>u.name==="workspace_action_result"&&u.action==="profile_workflow_action")?.output;if(typeof n!="string")throw new Error("Profile Agent action evaluation did not produce a released action result.");return n}if(i==="completed"){let n=[...e].reverse().find(s=>s.name==="turn.completed"),r=n?.data?.providerResponse,o=n?.data;if(!r&&o?.turn?.status==="completed"){let s=[...e].reverse().find(t=>P(t)!==null);if(!s)throw new Error("Native Profile evaluation has no retained final answer.");let d=P(s);return e.filter(t=>P(t)===d).map(t=>t.output??"").join("")}if(typeof r?.requestId!="string"||!r.requestId||typeof r.contentHash!="string")throw new Error("Profile evaluation has no retained final provider response.");let p=e.filter(s=>s.name==="assistant.delta"&&s.data?.providerRequestId===r.requestId).map(s=>s.output??"").join("");if(c(p)!==r.contentHash)throw new Error("Profile evaluation final provider response differs from its retained trace.");return b(e)??p}return e.filter(n=>n.name==="assistant.delta"&&typeof n.output=="string").map(n=>n.output??"").join("")}R(j,"profileEvaluationOutput");function P(e){let a=e.data;return e.name==="assistant.delta"&&a?.provider==="codex"&&a.phase==="final_answer"&&typeof a.itemId=="string"?a.itemId:null}R(P,"nativeFinalAnswerItemId");async function _(e){let{store:a,manifest:i,receipt:n}=e;if(!n.metadata.retainedEvidenceRef)return null;let u=x.parse(n.metadata.retainedEvidenceRef),[r,o]=await Promise.all([a.getSession(u.sessionId),a.getTurn(u.turnId)]),p=r?.metadata?.profileEvaluationRun,s=i.profileEvaluation;if(!r||!o||o.sessionId!==r.id||!s||!r.currentProfile||c(r.currentProfile)!==c(e.profileRef)||c(p)!==c({id:i.id,contentHash:i.contentHash})||r.metadata?.taskId!==n.taskId||r.metadata?.seed!==n.seed||!o.completedAt||o.completedAt!==n.completedAt||o.startedAt!==n.startedAt||!o.harnessSnapshot||o.harnessSnapshot.harnessRelease.id!==s.harnessRelease.id||o.harnessSnapshot.harnessRelease.contentHash!==s.harnessRelease.contentHash||i.policy.kind!=="model"||!o.modelRef||o.modelRef.providerId!==i.policy.model.provider||o.modelRef.modelId!==i.policy.model.model)throw new Error("Retained Profile policy turn differs from its admitted case.");let d=await a.runtimeEventsForTurn(o.id);if(d.some(v=>v.turnId!==o.id)||c(d)!==n.traceHash)throw new Error("Retained Profile policy trace differs from its immutable receipt.");let t=j(d,s.target.kind,o.status);if(n.outputHash!==null&&c({text:t})!==n.outputHash)throw new Error("Retained Profile output differs from its immutable receipt.");let m=n.artifactRefs.flatMap(v=>{let y=d.filter(l=>(l.action==="work_output_save"||l.action==="sandbox_save_output")&&l.status==="completed").flatMap(l=>{let C=l.data&&typeof l.data=="object"?l.data:{},k=M.safeParse(C.outputRef);return k.success?[k.data]:[]}).filter(l=>l.sourceTaskId===r.id&&l.sourceTurnId===o.id&&`${r.id}/${l.id}/${l.revision}/${l.title}`===v.id&&l.sha256===v.contentHash&&l.sizeBytes===v.sizeBytes&&l.contentType===v.mediaType);if(y.length!==1)throw new Error("The case output has no unique sealed saved-file identity.");let h=y[0];if(h.location.kind!=="managed")return[];let H=`/api/work/outputs/${encodeURIComponent(h.location.fileId)}`;if(h.location.downloadPath!==H)throw new Error("The saved output has no canonical managed download route.");return[{id:h.id,title:h.title,revision:h.revision,contentType:h.contentType,sizeBytes:h.sizeBytes,sha256:h.sha256,downloadPath:H}]}),f=await a.listModelUsageRecords({sessionId:r.id,turnId:o.id,limit:1e4}),E=e.eventAfterId?d.findIndex(v=>v.id===e.eventAfterId):-1;if(e.eventAfterId&&E<0)throw new Error("Profile trace cursor is unavailable in this case.");let g=E+1,q=Math.max(1,Math.min(250,e.eventLimit??100)),w=d.slice(g,g+q);return{ref:u,prompt:o.prompt,output:n.outputHash===null?null:{text:t},partialOutput:n.outputHash===null&&t?{text:t}:null,error:o.error,events:w,eventCount:d.length,nextEventCursor:g+w.length<d.length?w.at(-1).id:null,files:m,usage:f,usageCoverage:f.length===1e4?"partial":"complete"}}R(_,"readProfileEvaluationPolicyEvidence");var W=I.object({runId:I.string().trim().min(1).max(240),receiptId:I.string().trim().min(1).max(500).optional(),eventAfterId:I.string().trim().min(1).max(500).optional(),eventLimit:I.number().int().min(1).max(250).optional(),manifest:A.optional()}).strict().refine(e=>e.receiptId||!e.eventAfterId&&!e.eventLimit,"Select a case before requesting its trace page.").refine(e=>!e.manifest||e.receiptId,"Interrupted-run inspection requires one exact case.");async function D(e){let a=await e.store.getProfileEvaluationRun(e.runId);if(a&&c(a.profileRef)!==c(e.profileRef))throw new Error("Profile evaluation run is unavailable in the selected Profile.");let i=a?.manifest??e.manifest;if(!i||!a&&!e.receiptId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");if(a){if(e.manifest&&e.manifest.contentHash!==a.manifest.contentHash)throw new Error("Profile evaluation run differs from its pinned manifest.")}else if(S(i,"Interrupted Profile evaluation manifest"),i.id!==e.runId||i.profileEvaluation?.profileId!==e.profileRef.profileId||e.profileRef.source==="openpond_git"&&i.metadata.profileRepositoryId!==e.profileRef.repositoryId)throw new Error("Profile evaluation run is unavailable in the selected Profile.");let n=i.population.map((r,o)=>({member:r,index:o})).filter(({member:r})=>!e.receiptId||r.receiptId===e.receiptId);if(e.receiptId&&n.length!==1)throw new Error("Profile evaluation case is unavailable in the selected run.");let u=await Promise.all(n.map(async({member:r,index:o})=>{let p=a?null:await e.store.getProfileEvaluationReceipt(r.receiptId),s=a?.receiptRefs[o]??(p?{id:p.id,contentHash:p.contentHash}:null),d=a?.gradeRefs[o]??p?.graderEvidenceRefs[0];if(!s||!d)throw new Error("Profile evaluation run has incomplete case evidence.");let[t,m]=await Promise.all([e.store.getProfileEvaluationReceipt(s.id),e.store.getProfileEvaluationGrade(d.contentHash)]);if(!t||!a&&!T(t)||t.contentHash!==s.contentHash||t.id!==r.receiptId||t.taskId!==r.taskId||t.seed!==r.seed||t.runManifest.id!==i.id||t.runManifest.contentHash!==i.contentHash||!m||m.contentHash!==d.contentHash||!t.graderEvidenceRefs.some(f=>f.id===d.id&&f.contentHash===d.contentHash))throw new Error("Profile evaluation case evidence differs from its retained run.");return{taskId:r.taskId,seed:r.seed,receiptId:t.id,receiptHash:t.contentHash,traceHash:t.traceHash,outputHash:t.outputHash,score:m.score,passed:m.passed,gradingStatus:m.gradingStatus,failureClass:m.failureClass,artifactCount:t.artifactRefs.length,latencyMs:t.latencyMs,costUsd:t.costUsd,terminal:t.terminal,feedback:m.components.map(f=>({graderId:f.graderId,graderVersion:f.graderVersion,score:f.score,passed:f.passed,feedback:f.feedback})),...e.receiptId?{receipt:t,evidence:await _({store:e.store,profileRef:e.profileRef,manifest:i,receipt:t,eventAfterId:e.eventAfterId,eventLimit:e.eventLimit})}:{}}}));return{runId:i.id,sourceRevision:i.profileEvaluation?.sourceRevision??null,cases:u}}R(D,"inspectProfileEvaluationRun");export{j as a,W as b,D as c};

package/dist/chunks/{chunk-XOPHFUUY.js → chunk-VMBWCIGG.js}

@@ -1,2 +1,2 @@
 import { createRequire as __openpondCreateRequire } from "node:module"; var require = __openpondCreateRequire(import.meta.url);
-import{Gg as l,Jg as h,Kg as E,Lg as w,Rg as f,aa as d,fe as u,ra as p}from"./chunk-6D3FI2XT.js";import{c as i,f as m,h as g}from"./chunk-3M53YXEI.js";import{a as c}from"./chunk-FTQLQJQV.js";var M=new Set(["hosted-engram-capability","hosted-compile-engram","hosted-prepare-engram","hosted-start-engram","hosted-cancel-engram","hosted-engram-execution","hosted-engram-executions","hosted-save-engram-model","hosted-control-engram-serving"]);async function k(x){let{action:r,target:s,teamId:n,options:o,client:t,readInput:y}=x;if(r==="hosted-engram-execution"){if(!s)throw new Error("Supply the original Engram execution ID.");return t.getEngramExecution(n,s)}if(r==="hosted-engram-executions"){if(s)throw new Error("List Engram executions using filters.");let e=f.parse(Object.fromEntries(Object.entries({id:i(o,"id")||void 0,projectId:i(o,"projectId")||void 0,phase:i(o,"phase")||void 0,cursor:i(o,"cursor")||void 0,limit:g(o.limit,"--limit")}).filter(([,I])=>I!==void 0)));return t.listEngramExecutions(n,e)}let a=await y();if(r==="hosted-control-engram-serving"){let e=p.parse(a);if(e.request.teamId!==n)throw new Error("Use the Model's original workspace.");if(e.request.action==="activate"&&!m(o.yes))throw new Error("Activating billed serving capacity requires --yes.");return t.controlEngramModelServing(e.modelProjectId,e.request)}if(r==="hosted-engram-capability"){let e=u.parse(a);if(e.teamId!==n)throw new Error("Use this input's original workspace.");return t.engramExecutionCapability(e)}if(r==="hosted-compile-engram"){let e=l.parse(a);if(e.teamId!==n)throw new Error("Use this input's original workspace.");return t.compileEngramExecution(e)}if(r==="hosted-prepare-engram"){let e=h.parse(a);if(e.teamId!==n)throw new Error("Use this input's original workspace.");return t.prepareEngramExecution(e)}if(r==="hosted-start-engram"){let e=E.parse(a);if(e.teamId!==n)throw new Error("Use this execution's original workspace.");if(!m(o.yes))throw new Error("Starting paid Engram training requires --yes.");return t.startEngramExecution(e)}if(r==="hosted-cancel-engram"){let e=w.parse(a);if(e.teamId!==n)throw new Error("Use this execution's original workspace.");return t.cancelEngramExecution(e)}if(r==="hosted-save-engram-model"){let e=d.parse(a);if(e.teamId!==n)throw new Error("Use this result's original workspace.");return t.saveEngramResultModel(e)}throw new Error("Unsupported Engram command.")}c(k,"runHostedEngramCommand");export{M as a,k as b};
+import{Hg as l,Kg as h,Lg as E,Mg as w,Sg as f,aa as d,fe as u,ra as p}from"./chunk-AWD72O3G.js";import{c as i,f as m,h as g}from"./chunk-3M53YXEI.js";import{a as c}from"./chunk-FTQLQJQV.js";var M=new Set(["hosted-engram-capability","hosted-compile-engram","hosted-prepare-engram","hosted-start-engram","hosted-cancel-engram","hosted-engram-execution","hosted-engram-executions","hosted-save-engram-model","hosted-control-engram-serving"]);async function k(x){let{action:r,target:s,teamId:n,options:o,client:t,readInput:y}=x;if(r==="hosted-engram-execution"){if(!s)throw new Error("Supply the original Engram execution ID.");return t.getEngramExecution(n,s)}if(r==="hosted-engram-executions"){if(s)throw new Error("List Engram executions using filters.");let e=f.parse(Object.fromEntries(Object.entries({id:i(o,"id")||void 0,projectId:i(o,"projectId")||void 0,phase:i(o,"phase")||void 0,cursor:i(o,"cursor")||void 0,limit:g(o.limit,"--limit")}).filter(([,I])=>I!==void 0)));return t.listEngramExecutions(n,e)}let a=await y();if(r==="hosted-control-engram-serving"){let e=p.parse(a);if(e.request.teamId!==n)throw new Error("Use the Model's original workspace.");if(e.request.action==="activate"&&!m(o.yes))throw new Error("Activating billed serving capacity requires --yes.");return t.controlEngramModelServing(e.modelProjectId,e.request)}if(r==="hosted-engram-capability"){let e=u.parse(a);if(e.teamId!==n)throw new Error("Use this input's original workspace.");return t.engramExecutionCapability(e)}if(r==="hosted-compile-engram"){let e=l.parse(a);if(e.teamId!==n)throw new Error("Use this input's original workspace.");return t.compileEngramExecution(e)}if(r==="hosted-prepare-engram"){let e=h.parse(a);if(e.teamId!==n)throw new Error("Use this input's original workspace.");return t.prepareEngramExecution(e)}if(r==="hosted-start-engram"){let e=E.parse(a);if(e.teamId!==n)throw new Error("Use this execution's original workspace.");if(!m(o.yes))throw new Error("Starting paid Engram training requires --yes.");return t.startEngramExecution(e)}if(r==="hosted-cancel-engram"){let e=w.parse(a);if(e.teamId!==n)throw new Error("Use this execution's original workspace.");return t.cancelEngramExecution(e)}if(r==="hosted-save-engram-model"){let e=d.parse(a);if(e.teamId!==n)throw new Error("Use this result's original workspace.");return t.saveEngramResultModel(e)}throw new Error("Unsupported Engram command.")}c(k,"runHostedEngramCommand");export{M as a,k as b};