openpond-sdk 0.9.2 → 0.9.4

package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts

@@ -1,4 +1,6 @@
 import { z } from "zod";
+import { TasksetRunManifestSchema } from "@openpond/evals/metrics";
+export { HarnessExperimentPolicySchema, type HarnessExperimentPolicy } from "./harness-experiment-policy.js";
 export declare const ModelTasksetRunPolicySchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
     kind: z.ZodLiteral<"hosted_chat">;
     modelId: z.ZodString;
@@ -15,6 +17,44 @@ export declare const ModelTasksetRunPolicySchema: z.ZodDiscriminatedUnion<[z.Zod
     topP: z.ZodDefault<z.ZodNumber>;
 }, z.core.$strict>, z.ZodObject<{
     kind: z.ZodLiteral<"fixture">;
+}, z.core.$strict>, z.ZodObject<{
+    kind: z.ZodLiteral<"hosted_harness">;
+    modelId: z.ZodString;
+    reasoningEffort: z.ZodOptional<z.ZodEnum<{
+        high: "high";
+        low: "low";
+        off: "off";
+        medium: "medium";
+        xhigh: "xhigh";
+        max: "max";
+    }>>;
+    profileRepositoryId: z.ZodString;
+    source: z.ZodObject<{
+        profileId: z.ZodString;
+        sourceRevision: z.ZodString;
+        harnessRelease: z.ZodObject<{
+            id: z.ZodString;
+            contentHash: z.ZodString;
+        }, z.core.$strict>;
+        catalogHash: z.ZodString;
+        definitionId: z.ZodString;
+        definitionHash: z.ZodString;
+        target: z.ZodDiscriminatedUnion<[z.ZodObject<{
+            kind: z.ZodLiteral<"profile">;
+        }, z.core.$strict>, z.ZodObject<{
+            kind: z.ZodLiteral<"workflow">;
+            workflowId: z.ZodString;
+        }, z.core.$strict>, z.ZodObject<{
+            kind: z.ZodLiteral<"skill">;
+            skillPath: z.ZodString;
+        }, z.core.$strict>, z.ZodObject<{
+            kind: z.ZodLiteral<"agent_action">;
+            actionId: z.ZodString;
+        }, z.core.$strict>], "kind">;
+        environmentHash: z.ZodString;
+    }, z.core.$strict>;
+    modelConfigurationHash: z.ZodString;
+    packageHash: z.ZodString;
 }, z.core.$strict>], "kind">;
 export declare const ModelTasksetRunRequestSchema: z.ZodObject<{
     schemaVersion: z.ZodLiteral<"openpond.modelTasksetRunRequest.v1">;
@@ -49,6 +89,44 @@ export declare const ModelTasksetRunRequestSchema: z.ZodObject<{
         topP: z.ZodDefault<z.ZodNumber>;
     }, z.core.$strict>, z.ZodObject<{
         kind: z.ZodLiteral<"fixture">;
+    }, z.core.$strict>, z.ZodObject<{
+        kind: z.ZodLiteral<"hosted_harness">;
+        modelId: z.ZodString;
+        reasoningEffort: z.ZodOptional<z.ZodEnum<{
+            high: "high";
+            low: "low";
+            off: "off";
+            medium: "medium";
+            xhigh: "xhigh";
+            max: "max";
+        }>>;
+        profileRepositoryId: z.ZodString;
+        source: z.ZodObject<{
+            profileId: z.ZodString;
+            sourceRevision: z.ZodString;
+            harnessRelease: z.ZodObject<{
+                id: z.ZodString;
+                contentHash: z.ZodString;
+            }, z.core.$strict>;
+            catalogHash: z.ZodString;
+            definitionId: z.ZodString;
+            definitionHash: z.ZodString;
+            target: z.ZodDiscriminatedUnion<[z.ZodObject<{
+                kind: z.ZodLiteral<"profile">;
+            }, z.core.$strict>, z.ZodObject<{
+                kind: z.ZodLiteral<"workflow">;
+                workflowId: z.ZodString;
+            }, z.core.$strict>, z.ZodObject<{
+                kind: z.ZodLiteral<"skill">;
+                skillPath: z.ZodString;
+            }, z.core.$strict>, z.ZodObject<{
+                kind: z.ZodLiteral<"agent_action">;
+                actionId: z.ZodString;
+            }, z.core.$strict>], "kind">;
+            environmentHash: z.ZodString;
+        }, z.core.$strict>;
+        modelConfigurationHash: z.ZodString;
+        packageHash: z.ZodString;
     }, z.core.$strict>], "kind">;
     population: z.ZodArray<z.ZodObject<{
         receiptId: z.ZodString;
@@ -80,6 +158,7 @@ export declare const ModelTasksetRunSummarySchema: z.ZodObject<{
     policyKind: z.ZodEnum<{
         fixture: "fixture";
         hosted_chat: "hosted_chat";
+        hosted_harness: "hosted_harness";
     }>;
     manifestHash: z.ZodString;
     status: z.ZodEnum<{
@@ -134,6 +213,7 @@ export declare const ModelTasksetRunDetailsSchema: z.ZodObject<{
         policyKind: z.ZodEnum<{
             fixture: "fixture";
             hosted_chat: "hosted_chat";
+            hosted_harness: "hosted_harness";
         }>;
         manifestHash: z.ZodString;
         status: z.ZodEnum<{
@@ -197,6 +277,44 @@ export declare const ModelTasksetRunDetailsSchema: z.ZodObject<{
             topP: z.ZodDefault<z.ZodNumber>;
         }, z.core.$strict>, z.ZodObject<{
             kind: z.ZodLiteral<"fixture">;
+        }, z.core.$strict>, z.ZodObject<{
+            kind: z.ZodLiteral<"hosted_harness">;
+            modelId: z.ZodString;
+            reasoningEffort: z.ZodOptional<z.ZodEnum<{
+                high: "high";
+                low: "low";
+                off: "off";
+                medium: "medium";
+                xhigh: "xhigh";
+                max: "max";
+            }>>;
+            profileRepositoryId: z.ZodString;
+            source: z.ZodObject<{
+                profileId: z.ZodString;
+                sourceRevision: z.ZodString;
+                harnessRelease: z.ZodObject<{
+                    id: z.ZodString;
+                    contentHash: z.ZodString;
+                }, z.core.$strict>;
+                catalogHash: z.ZodString;
+                definitionId: z.ZodString;
+                definitionHash: z.ZodString;
+                target: z.ZodDiscriminatedUnion<[z.ZodObject<{
+                    kind: z.ZodLiteral<"profile">;
+                }, z.core.$strict>, z.ZodObject<{
+                    kind: z.ZodLiteral<"workflow">;
+                    workflowId: z.ZodString;
+                }, z.core.$strict>, z.ZodObject<{
+                    kind: z.ZodLiteral<"skill">;
+                    skillPath: z.ZodString;
+                }, z.core.$strict>, z.ZodObject<{
+                    kind: z.ZodLiteral<"agent_action">;
+                    actionId: z.ZodString;
+                }, z.core.$strict>], "kind">;
+                environmentHash: z.ZodString;
+            }, z.core.$strict>;
+            modelConfigurationHash: z.ZodString;
+            packageHash: z.ZodString;
         }, z.core.$strict>], "kind">;
         population: z.ZodArray<z.ZodObject<{
             receiptId: z.ZodString;
@@ -356,6 +474,7 @@ export declare const ModelTasksetRunPageSchema: z.ZodObject<{
         policyKind: z.ZodEnum<{
             fixture: "fixture";
             hosted_chat: "hosted_chat";
+            hosted_harness: "hosted_harness";
         }>;
         manifestHash: z.ZodString;
         status: z.ZodEnum<{
@@ -413,6 +532,7 @@ export declare const ModelTasksetRunResultSchema: z.ZodObject<{
             policyKind: z.ZodEnum<{
                 fixture: "fixture";
                 hosted_chat: "hosted_chat";
+                hosted_harness: "hosted_harness";
             }>;
             manifestHash: z.ZodString;
             status: z.ZodEnum<{
@@ -476,6 +596,44 @@ export declare const ModelTasksetRunResultSchema: z.ZodObject<{
                 topP: z.ZodDefault<z.ZodNumber>;
             }, z.core.$strict>, z.ZodObject<{
                 kind: z.ZodLiteral<"fixture">;
+            }, z.core.$strict>, z.ZodObject<{
+                kind: z.ZodLiteral<"hosted_harness">;
+                modelId: z.ZodString;
+                reasoningEffort: z.ZodOptional<z.ZodEnum<{
+                    high: "high";
+                    low: "low";
+                    off: "off";
+                    medium: "medium";
+                    xhigh: "xhigh";
+                    max: "max";
+                }>>;
+                profileRepositoryId: z.ZodString;
+                source: z.ZodObject<{
+                    profileId: z.ZodString;
+                    sourceRevision: z.ZodString;
+                    harnessRelease: z.ZodObject<{
+                        id: z.ZodString;
+                        contentHash: z.ZodString;
+                    }, z.core.$strict>;
+                    catalogHash: z.ZodString;
+                    definitionId: z.ZodString;
+                    definitionHash: z.ZodString;
+                    target: z.ZodDiscriminatedUnion<[z.ZodObject<{
+                        kind: z.ZodLiteral<"profile">;
+                    }, z.core.$strict>, z.ZodObject<{
+                        kind: z.ZodLiteral<"workflow">;
+                        workflowId: z.ZodString;
+                    }, z.core.$strict>, z.ZodObject<{
+                        kind: z.ZodLiteral<"skill">;
+                        skillPath: z.ZodString;
+                    }, z.core.$strict>, z.ZodObject<{
+                        kind: z.ZodLiteral<"agent_action">;
+                        actionId: z.ZodString;
+                    }, z.core.$strict>], "kind">;
+                    environmentHash: z.ZodString;
+                }, z.core.$strict>;
+                modelConfigurationHash: z.ZodString;
+                packageHash: z.ZodString;
             }, z.core.$strict>], "kind">;
             population: z.ZodArray<z.ZodObject<{
                 receiptId: z.ZodString;
@@ -691,6 +849,7 @@ export declare const ModelTasksetRunResultSchema: z.ZodObject<{
     contentHash: z.ZodString;
 }, z.core.$strict>;
 export type ModelTasksetRunResult = z.infer<typeof ModelTasksetRunResultSchema>;
+export declare function verifyHarnessExperimentManifest(request: ModelTasksetRunRequest, manifest: z.infer<typeof TasksetRunManifestSchema>): void;
 /** Server artifact ownership authenticates the producer. These checks bind all
  * portable projections to the admitted request, population and policy. */
 export declare function verifyModelTasksetRunDetails(value: unknown): Promise<ModelTasksetRunDetails>;

package/dist/types/packages/sdk/src/model-taskset-runs-contracts.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model-taskset-runs-contracts.d.ts","sourceRoot":"","sources":["../../../../../src/model-taskset-runs-contracts.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAWxB,eAAO,MAAM,2BAA2B;;;;;;;;;;;;;;;;4BAGtC,CAAC;AACH,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAWvC,CAAC;AACH,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAqBvC,CAAC;AACH,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAK9B,CAAC;AACZ,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,8BAA8B;;;;;kBAA+M,CAAC;AAC3P,eAAO,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAAgH,CAAC;AACvJ,eAAO,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAM7B,CAAC;AACZ,MAAM,MAAM,qBAAqB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,2BAA2B,CAAC,CAAC;AAEhF;0EAC0E;AAC1E,wBAAsB,4BAA4B,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAAC,sBAAsB,CAAC,CAqBlG;AAED,wBAAsB,2BAA2B,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAAC,qBAAqB,CAAC,CAkBhG"}
+{"version":3,"file":"model-taskset-runs-contracts.d.ts","sourceRoot":"","sources":["../../../../../src/model-taskset-runs-contracts.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAGxB,OAAO,EAAE,wBAAwB,EAAyG,MAAM,yBAAyB,CAAC;AAM1K,OAAO,EAAE,6BAA6B,EAAE,KAAK,uBAAuB,EAAE,MAAM,gCAAgC,CAAC;AAI7G,eAAO,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;4BAItC,CAAC;AACH,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAgBvC,CAAC;AACH,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAqBvC,CAAC;AACH,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAK9B,CAAC;AACZ,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAClF,eAAO,MAAM,8BAA8B;;;;;kBAA+M,CAAC;AAC3P,eAAO,MAAM,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAAgH,CAAC;AACvJ,eAAO,MAAM,2BAA2B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kBAM7B,CAAC;AACZ,MAAM,MAAM,qBAAqB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,2BAA2B,CAAC,CAAC;AAEhF,wBAAgB,+BAA+B,CAAC,OAAO,EAAE,sBAAsB,EAAE,QAAQ,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,wBAAwB,CAAC,QAgBlI;AAED;0EAC0E;AAC1E,wBAAsB,4BAA4B,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAAC,sBAAsB,CAAC,CAwBlG;AAED,wBAAsB,2BAA2B,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAAC,qBAAqB,CAAC,CAkBhG"}

package/dist/types/packages/sdk/src/model-taskset-runs.d.ts

@@ -29,7 +29,7 @@ export declare class OpenPondModelTasksetRunsClient {
                 contentHash: string;
                 revision: number;
             };
-            policyKind: "fixture" | "hosted_chat";
+            policyKind: "fixture" | "hosted_chat" | "hosted_harness";
             manifestHash: string;
             status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
             totalCount: number;
@@ -79,6 +79,37 @@ export declare class OpenPondModelTasksetRunsClient {
                     role: "user" | "system" | "assistant";
                     content: string;
                 }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
             } | {
                 kind: "fixture";
             };
@@ -219,7 +250,7 @@ export declare class OpenPondModelTasksetRunsClient {
                 contentHash: string;
                 revision: number;
             };
-            policyKind: "fixture" | "hosted_chat";
+            policyKind: "fixture" | "hosted_chat" | "hosted_harness";
             manifestHash: string;
             status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
             totalCount: number;
@@ -269,6 +300,37 @@ export declare class OpenPondModelTasksetRunsClient {
                     role: "user" | "system" | "assistant";
                     content: string;
                 }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
             } | {
                 kind: "fixture";
             };
@@ -409,7 +471,7 @@ export declare class OpenPondModelTasksetRunsClient {
                 contentHash: string;
                 revision: number;
             };
-            policyKind: "fixture" | "hosted_chat";
+            policyKind: "fixture" | "hosted_chat" | "hosted_harness";
             manifestHash: string;
             status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
             totalCount: number;
@@ -459,6 +521,37 @@ export declare class OpenPondModelTasksetRunsClient {
                     role: "user" | "system" | "assistant";
                     content: string;
                 }[] | undefined;
+            } | {
+                kind: "hosted_harness";
+                modelId: string;
+                profileRepositoryId: string;
+                source: {
+                    profileId: string;
+                    sourceRevision: string;
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    catalogHash: string;
+                    definitionId: string;
+                    definitionHash: string;
+                    target: {
+                        kind: "profile";
+                    } | {
+                        kind: "workflow";
+                        workflowId: string;
+                    } | {
+                        kind: "skill";
+                        skillPath: string;
+                    } | {
+                        kind: "agent_action";
+                        actionId: string;
+                    };
+                    environmentHash: string;
+                };
+                modelConfigurationHash: string;
+                packageHash: string;
+                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
             } | {
                 kind: "fixture";
             };
@@ -601,7 +694,7 @@ export declare class OpenPondModelTasksetRunsClient {
                     contentHash: string;
                     revision: number;
                 };
-                policyKind: "fixture" | "hosted_chat";
+                policyKind: "fixture" | "hosted_chat" | "hosted_harness";
                 manifestHash: string;
                 status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
                 totalCount: number;
@@ -651,6 +744,37 @@ export declare class OpenPondModelTasksetRunsClient {
                         role: "user" | "system" | "assistant";
                         content: string;
                     }[] | undefined;
+                } | {
+                    kind: "hosted_harness";
+                    modelId: string;
+                    profileRepositoryId: string;
+                    source: {
+                        profileId: string;
+                        sourceRevision: string;
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        catalogHash: string;
+                        definitionId: string;
+                        definitionHash: string;
+                        target: {
+                            kind: "profile";
+                        } | {
+                            kind: "workflow";
+                            workflowId: string;
+                        } | {
+                            kind: "skill";
+                            skillPath: string;
+                        } | {
+                            kind: "agent_action";
+                            actionId: string;
+                        };
+                        environmentHash: string;
+                    };
+                    modelConfigurationHash: string;
+                    packageHash: string;
+                    reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
                 } | {
                     kind: "fixture";
                 };
@@ -861,7 +985,7 @@ export declare class OpenPondModelTasksetRunsClient {
                 contentHash: string;
                 revision: number;
             };
-            policyKind: "fixture" | "hosted_chat";
+            policyKind: "fixture" | "hosted_chat" | "hosted_harness";
             manifestHash: string;
             status: "failed" | "completed" | "cancelled" | "running" | "cancelling" | "queued";
             totalCount: number;

package/dist/types/packages/sdk/src/model-taskset-runs.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model-taskset-runs.d.ts","sourceRoot":"","sources":["../../../../../src/model-taskset-runs.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAExB,OAAO,EAAE,4BAA4B,EAAE,8BAA8B,EAAqH,MAAM,mCAAmC,CAAC;AACpO,cAAc,mCAAmC,CAAC;AAGlD,qBAAa,4BAA6B,SAAQ,KAAK;IACzC,QAAQ,CAAC,MAAM,EAAE,MAAM;IAAE,QAAQ,CAAC,IAAI,EAAE,MAAM;gBAArC,MAAM,EAAE,MAAM,EAAW,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM;CAC5E;AAED,qBAAa,8BAA8B;;gBAE7B,OAAO,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,OAAO,UAAU,CAAC,KAAK,CAAA;KAAE;IAKnG,MAAM,CAAC,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAQlG,GAAG,CAAC,EAAE,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAKtD,MAAM,CAAC,EAAE,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAKzD,MAAM,CAAC,EAAE,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAKzD,IAAI,CAAC,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,8BAA8B,CAAC,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA0BzG"}
+{"version":3,"file":"model-taskset-runs.d.ts","sourceRoot":"","sources":["../../../../../src/model-taskset-runs.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAExB,OAAO,EAAE,4BAA4B,EAAE,8BAA8B,EAAqH,MAAM,mCAAmC,CAAC;AACpO,cAAc,mCAAmC,CAAC;AAGlD,qBAAa,4BAA6B,SAAQ,KAAK;IACzC,QAAQ,CAAC,MAAM,EAAE,MAAM;IAAE,QAAQ,CAAC,IAAI,EAAE,MAAM;gBAArC,MAAM,EAAE,MAAM,EAAW,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM;CAC5E;AAED,qBAAa,8BAA8B;;gBAE7B,OAAO,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,OAAO,UAAU,CAAC,KAAK,CAAA;KAAE;IAKnG,MAAM,CAAC,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAQlG,GAAG,CAAC,EAAE,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAKtD,MAAM,CAAC,EAAE,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAKzD,MAAM,CAAC,EAAE,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;IAKzD,IAAI,CAAC,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,8BAA8B,CAAC,EAAE,OAAO,GAAE;QAAE,MAAM,CAAC,EAAE,WAAW,CAAA;KAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA0BzG"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openpond-sdk",
-  "version": "0.9.2",
+  "version": "0.9.4",
   "description": "Server-side TypeScript SDK for OpenPond sandboxes and agentic work",
   "license": "MIT",
   "type": "module",
@@ -106,6 +106,10 @@
     "./dataset-workspaces": {
       "types": "./dist/types/packages/sdk/src/dataset-workspaces.d.ts",
       "import": "./dist/dataset-workspaces.js"
+    },
+    "./experiments": {
+      "types": "./dist/types/packages/sdk/src/experiments.d.ts",
+      "import": "./dist/experiments.js"
     }
   },
   "publishConfig": {
@@ -147,7 +151,7 @@
     "vitest": "4.1.10"
   },
   "dependencies": {
-    "@openpond/evals": "^0.11.10",
+    "@openpond/evals": "^0.11.13",
     "@openpond/harness": "^0.4.2",
     "yaml": "^2.9.0",
     "esbuild": "0.28.1",

package/dist/chunk-IKTZ2CD4.js.map

@@ -1,7 +0,0 @@
-{
-  "version": 3,
-  "sources": ["../src/model-starter-attempts.ts", "../src/evaluation-project-context.ts"],
-  "sourcesContent": ["import { z } from \"zod\";\nimport { EvaluationProjectContextSchema, EvaluationMessagesSchema } from \"./evaluation-project-context.js\";\nexport { EvaluationProjectContextSchema, EvaluationMessagesSchema, type EvaluationProjectContext } from \"./evaluation-project-context.js\";\nimport { RewardCompositionSchema } from \"@openpond/evals/rewards\";\nimport { TaskGradeSchema } from \"@openpond/evals/graders\";\nimport { ModelProjectVersionedRefSchema } from \"./model-projects.js\";\nimport { canonicalJson, canonicalSha256 } from \"./protocol.js\";\n\nconst IdSchema = z.string().trim().min(1).max(200);\nconst HashSchema = z.string().regex(/^[a-f0-9]{64}$/);\nexport const ModelStarterAttemptPolicySchema = z.discriminatedUnion(\"kind\", [\n  z.object({ kind: z.literal(\"hosted_chat\"), modelId: IdSchema, messages: EvaluationMessagesSchema.optional(), maxOutputTokens: z.number().int().min(1).max(4_096).default(1_024), temperature: z.number().min(0).max(2).default(0), topP: z.number().gt(0).max(1).default(1) }).strict(),\n  z.object({ kind: z.literal(\"fixture\"), fixtureId: IdSchema }).strict(),\n]);\n/** The server resolves task input, private state, verifier and fixture script.\n * A caller can select releases, but cannot submit replacement execution bytes. */\nexport const ModelStarterAttemptRequestSchema = z.object({\n  schemaVersion: z.literal(\"openpond.modelStarterAttemptRequest.v1\"),\n  operationId: IdSchema, teamId: IdSchema, modelProjectId: IdSchema.nullable(),\n  project: EvaluationProjectContextSchema.optional(),\n  taskset: ModelProjectVersionedRefSchema, taskId: IdSchema,\n  environmentSeed: z.number().int().min(0).max(2_147_483_647).default(0),\n  policy: ModelStarterAttemptPolicySchema,\n}).strict();\nexport type ModelStarterAttemptRequest = z.infer<typeof ModelStarterAttemptRequestSchema>;\nexport const ModelStarterAttemptChoicesQuerySchema = z.object({ modelProjectId: IdSchema.nullable(), taskset: ModelProjectVersionedRefSchema, afterTaskId: IdSchema.optional(), limit: z.number().int().min(1).max(100).default(25) }).strict();\nexport const ModelStarterAttemptChoicesSchema = z.object({\n  modelProjectId: IdSchema.nullable(), taskset: ModelProjectVersionedRefSchema, available: z.boolean(), unavailableReason: z.string().max(1_000).nullable(),\n  models: z.array(z.object({ id: IdSchema, name: z.string().max(500) }).strict()).max(200),\n  tasks: z.array(z.object({ id: IdSchema, split: z.enum([\"train\", \"validation\", \"frozen_eval\"]), inputPreview: z.string().max(500), fixtures: z.array(z.object({ id: IdSchema, label: z.string().max(200) }).strict()).max(1_000) }).strict()).max(100),\n  nextCursor: IdSchema.nullable(),\n}).strict();\nexport const ModelStarterAttemptSummarySchema = z.object({\n  schemaVersion: z.literal(\"openpond.modelStarterAttemptSummary.v1\"),\n  id: IdSchema, revision: z.number().int().positive(), request: ModelStarterAttemptRequestSchema,\n  status: z.enum([\"queued\", \"running\", \"cancelling\", \"completed\", \"failed\", \"cancelled\"]),\n  policySnapshot: z.object({ modelId: IdSchema, provider: IdSchema, upstreamModelId: z.string().min(1).max(500), configurationHash: HashSchema }).strict().nullable(),\n  createdAt: z.iso.datetime(), startedAt: z.iso.datetime().nullable(), completedAt: z.iso.datetime().nullable(),\n  cleanupComplete: z.boolean(), resultAvailable: z.boolean(), score: z.number().min(0).max(1).nullable(),\n  gradingStatus: z.enum([\"not_started\", \"scored\", \"unscorable\", \"not_configured\"]).default(\"not_started\"),\n  passed: z.boolean().nullable(), outputPreview: z.string().max(500).nullable(),\n  error: z.object({ code: IdSchema, message: z.string().max(2_000) }).strict().nullable(),\n}).strict().superRefine((value, context) => {\n  if ((value.request.policy.kind === \"fixture\") !== (value.policySnapshot === null)) context.addIssue({ code: \"custom\", path: [\"policySnapshot\"], message: \"Fixture execution must not claim a model identity.\" });\n  if (value.status === \"completed\" && (!value.cleanupComplete || !value.resultAvailable || !value.completedAt)) context.addIssue({ code: \"custom\", message: \"Completed attempts require retained results and confirmed cleanup.\" });\n});\nexport type ModelStarterAttemptSummary = z.infer<typeof ModelStarterAttemptSummarySchema>;\nexport const ModelStarterAttemptListQuerySchema = z.object({ modelProjectId: IdSchema, afterId: IdSchema.optional(), limit: z.number().int().min(1).max(100).default(25) }).strict();\nexport const ModelStarterAttemptPageSchema = z.object({ items: z.array(ModelStarterAttemptSummarySchema).max(100), nextCursor: IdSchema.nullable() }).strict();\nexport const ModelStarterAttemptResultSchema = z.object({\n  schemaVersion: z.literal(\"openpond.modelStarterAttemptResult.v1\"),\n  attempt: ModelStarterAttemptSummarySchema,\n  output: z.string().max(1_048_576).nullable(),\n  // Only model-facing messages, never the private environment snapshot.\n  messages: z.array(z.record(z.string(), z.unknown())).max(4_002),\n  composition: RewardCompositionSchema.nullable(),\n  // Ordinary Tasksets have no Reward binding. Absence preserves the hash of\n  // already retained v1 results; new producers explicitly return null or a grade.\n  grade: TaskGradeSchema.nullable().optional(),\n  environment: z.object({ status: z.enum([\"completed\", \"budget_exhausted\", \"cancelled\", \"timed_out\", \"policy_failure\", \"environment_failure\"]), collected: z.boolean(), definition: ModelProjectVersionedRefSchema, initialStateHash: HashSchema.nullable(), finalStateHash: HashSchema.nullable(), attemptHash: HashSchema }).strict(),\n  providerRequestIds: z.array(IdSchema).max(1_001),\n  contentHash: HashSchema,\n}).strict().superRefine((value, context) => {\n  if (!value.grade) return;\n  if (value.composition) context.addIssue({ code: \"custom\", path: [\"grade\"], message: \"An ordinary grade cannot also claim a Reward composition.\" });\n  if (![\"completed\", \"failed\"].includes(value.attempt.status) || !value.attempt.resultAvailable || value.attempt.score !== value.grade.score || value.attempt.gradingStatus !== value.grade.gradingStatus || value.attempt.passed !== (value.grade.score === null ? null : value.grade.passed)) context.addIssue({ code: \"custom\", path: [\"grade\"], message: \"Ordinary grade differs from its terminal attempt summary.\" });\n});\nexport type ModelStarterAttemptResult = z.infer<typeof ModelStarterAttemptResultSchema>;\n\nexport class OpenPondModelStarterAttemptError extends Error {\n  constructor(readonly status: number, readonly code: string, message: string) { super(message); this.name = \"OpenPondModelStarterAttemptError\"; }\n}\n\nexport class OpenPondModelStarterAttemptsClient {\n  readonly #options: { baseUrl: string; apiKey: string; teamId: string; fetch?: typeof globalThis.fetch };\n  constructor(options: { baseUrl: string; apiKey: string; teamId: string; fetch?: typeof globalThis.fetch }) {\n    const url = new URL(options.baseUrl);\n    if (!options.apiKey.trim() || !options.teamId.trim() || ![\"https:\", \"http:\"].includes(url.protocol) || url.username || url.password || url.search || url.hash) throw new Error(\"Starter attempts require an HTTP(S) endpoint and workspace credentials.\");\n    this.#options = { ...options, baseUrl: url.toString().replace(/\\/+$/, \"\") };\n  }\n  async create(value: z.input<typeof ModelStarterAttemptRequestSchema>, options: { signal?: AbortSignal } = {}) {\n    const request = ModelStarterAttemptRequestSchema.parse(value);\n    if (request.teamId !== this.#options.teamId) throw new Error(\"Attempt belongs to another workspace.\");\n    const result = this.#summary(await this.#request(\"\", \"POST\", request, options.signal));\n    if (canonicalJson(result.request) !== canonicalJson(request)) throw new OpenPondModelStarterAttemptError(502, \"attempt_request_mismatch\", \"Attempt receipt differs from the submitted request.\");\n    return result;\n  }\n  async get(id: string, options: { signal?: AbortSignal } = {}) { return this.#summary(await this.#request(`/${encodeURIComponent(IdSchema.parse(id))}`, \"GET\", undefined, options.signal), id); }\n  async choices(value: z.input<typeof ModelStarterAttemptChoicesQuerySchema>, options: { signal?: AbortSignal } = {}) {\n    const query = ModelStarterAttemptChoicesQuerySchema.parse(value);\n    const params = new URLSearchParams({ ...(query.modelProjectId ? { modelProjectId: query.modelProjectId } : {}), tasksetId: query.taskset.id, revision: String(query.taskset.revision), contentHash: query.taskset.contentHash, limit: String(query.limit) });\n    if (query.afterTaskId) params.set(\"afterTaskId\", query.afterTaskId);\n    const result = ModelStarterAttemptChoicesSchema.parse(await this.#request(`/choices?${params}`, \"GET\", undefined, options.signal));\n    if (result.modelProjectId !== query.modelProjectId || canonicalJson(result.taskset) !== canonicalJson(query.taskset) || result.tasks.length > query.limit || new Set(result.tasks.map(task => task.id)).size !== result.tasks.length) throw new OpenPondModelStarterAttemptError(502, \"attempt_choices_mismatch\", \"Attempt choices differ from the selected model or Taskset.\");\n    return result;\n  }\n  async cancel(id: string, options: { signal?: AbortSignal } = {}) { return this.#summary(await this.#request(`/${encodeURIComponent(IdSchema.parse(id))}/cancel`, \"POST\", undefined, options.signal), id); }\n  async result(id: string, options: { signal?: AbortSignal } = {}) {\n    const value = ModelStarterAttemptResultSchema.parse(await this.#request(`/${encodeURIComponent(IdSchema.parse(id))}/result`, \"GET\", undefined, options.signal));\n    this.#summary(value.attempt, id);\n    const { contentHash, ...content } = value;\n    if (!value.attempt.resultAvailable || await canonicalSha256(content) !== contentHash) throw new OpenPondModelStarterAttemptError(502, \"attempt_result_integrity\", \"Attempt result integrity failed.\");\n    return value;\n  }\n  async list(value: z.input<typeof ModelStarterAttemptListQuerySchema>, options: { signal?: AbortSignal } = {}) {\n    const query = ModelStarterAttemptListQuerySchema.parse(value);\n    const params = new URLSearchParams({ modelProjectId: query.modelProjectId, limit: String(query.limit) });\n    if (query.afterId) params.set(\"afterId\", query.afterId);\n    const page = ModelStarterAttemptPageSchema.parse(await this.#request(`?${params}`, \"GET\", undefined, options.signal));\n    if (page.items.length > query.limit || new Set(page.items.map(item => item.id)).size !== page.items.length || page.items.some(item => this.#summary(item).request.modelProjectId !== query.modelProjectId)) throw new OpenPondModelStarterAttemptError(502, \"attempt_page_mismatch\", \"Attempt page differs from the selected model.\");\n    return page;\n  }\n  #summary(value: unknown, id?: string): ModelStarterAttemptSummary {\n    const result = ModelStarterAttemptSummarySchema.parse(value);\n    if (result.request.teamId !== this.#options.teamId || (id && result.id !== id)) throw new OpenPondModelStarterAttemptError(502, \"attempt_identity_mismatch\", \"Attempt receipt belongs to another identity or workspace.\");\n    return result;\n  }\n  async #request(path: string, method: string, body: unknown, signal?: AbortSignal): Promise<unknown> {\n    const response = await (this.#options.fetch ?? globalThis.fetch)(`${this.#options.baseUrl}/v1/model-starter-attempts${path}`, { method, headers: { Authorization: `Bearer ${this.#options.apiKey}`, \"X-OpenPond-Team-Id\": this.#options.teamId, \"Content-Type\": \"application/json\", Accept: \"application/json\" }, ...(body === undefined ? {} : { body: JSON.stringify(body) }), signal, redirect: \"error\" });\n    const reader = response.body?.getReader();\n    if (!reader) throw new OpenPondModelStarterAttemptError(response.status, \"attempt_empty_response\", \"Attempt response was empty.\");\n    let bytes = 0; let text = \"\"; const decoder = new TextDecoder();\n    try {\n      while (true) {\n        const chunk = await reader.read(); if (chunk.done) break;\n        bytes += chunk.value.byteLength;\n        if (bytes > 16_777_216) throw new OpenPondModelStarterAttemptError(502, \"attempt_response_too_large\", \"Attempt response exceeded its byte limit.\");\n        text += decoder.decode(chunk.value, { stream: true });\n      }\n      text += decoder.decode();\n    } finally { await reader.cancel(); reader.releaseLock(); }\n    const value: unknown = JSON.parse(text);\n    if (!response.ok) {\n      const error = z.object({ code: IdSchema, message: z.string().max(2_000) }).safeParse(value);\n      throw new OpenPondModelStarterAttemptError(response.status, error.success ? error.data.code : \"attempt_request_failed\", error.success ? error.data.message : \"Attempt request failed.\");\n    }\n    return value;\n  }\n}\n", "import { z } from \"zod\";\n\n/** The selected Project revision is retained independently of model artifacts.\n * Hosts resolve and authorize this reference before admitting execution. */\nexport const EvaluationProjectContextSchema = z.object({\n  id: z.string().trim().min(1).max(240),\n  revision: z.number().int().positive(),\n  contentHash: z.string().regex(/^[a-f0-9]{64}$/),\n  targetId: z.string().trim().min(1).max(240).nullable(),\n}).strict();\nexport type EvaluationProjectContext = z.infer<typeof EvaluationProjectContextSchema>;\n\nexport const EvaluationMessagesSchema = z.array(z.object({\n  role: z.enum([\"system\", \"user\", \"assistant\"]),\n  content: z.string().max(100_000),\n}).strict()).max(100).superRefine((messages, context) => {\n  if (messages.reduce((total, message) => total + message.content.length, 0) > 100_000)\n    context.addIssue({ code: \"custom\", message: \"Evaluation prompt exceeds 100,000 characters.\" });\n});\n"],
-  "mappings": ";;;;;;;AAAA,SAAS,KAAAA,UAAS;;;ACAlB,SAAS,SAAS;AAIX,IAAM,iCAAiC,EAAE,OAAO;AAAA,EACrD,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG;AAAA,EACpC,UAAU,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAAA,EACpC,aAAa,EAAE,OAAO,EAAE,MAAM,gBAAgB;AAAA,EAC9C,UAAU,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,SAAS;AACvD,CAAC,EAAE,OAAO;AAGH,IAAM,2BAA2B,EAAE,MAAM,EAAE,OAAO;AAAA,EACvD,MAAM,EAAE,KAAK,CAAC,UAAU,QAAQ,WAAW,CAAC;AAAA,EAC5C,SAAS,EAAE,OAAO,EAAE,IAAI,GAAO;AACjC,CAAC,EAAE,OAAO,CAAC,EAAE,IAAI,GAAG,EAAE,YAAY,CAAC,UAAU,YAAY;AACvD,MAAI,SAAS,OAAO,CAAC,OAAO,YAAY,QAAQ,QAAQ,QAAQ,QAAQ,CAAC,IAAI;AAC3E,YAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,gDAAgD,CAAC;AACjG,CAAC;;;ADfD,SAAS,+BAA+B;AACxC,SAAS,uBAAuB;AAIhC,IAAM,WAAWC,GAAE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG;AACjD,IAAM,aAAaA,GAAE,OAAO,EAAE,MAAM,gBAAgB;AAC7C,IAAM,kCAAkCA,GAAE,mBAAmB,QAAQ;AAAA,EAC1EA,GAAE,OAAO,EAAE,MAAMA,GAAE,QAAQ,aAAa,GAAG,SAAS,UAAU,UAAU,yBAAyB,SAAS,GAAG,iBAAiBA,GAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,IAAK,EAAE,QAAQ,IAAK,GAAG,aAAaA,GAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,QAAQ,CAAC,GAAG,MAAMA,GAAE,OAAO,EAAE,GAAG,CAAC,EAAE,IAAI,CAAC,EAAE,QAAQ,CAAC,EAAE,CAAC,EAAE,OAAO;AAAA,EACtRA,GAAE,OAAO,EAAE,MAAMA,GAAE,QAAQ,SAAS,GAAG,WAAW,SAAS,CAAC,EAAE,OAAO;AACvE,CAAC;AAGM,IAAM,mCAAmCA,GAAE,OAAO;AAAA,EACvD,eAAeA,GAAE,QAAQ,wCAAwC;AAAA,EACjE,aAAa;AAAA,EAAU,QAAQ;AAAA,EAAU,gBAAgB,SAAS,SAAS;AAAA,EAC3E,SAAS,+BAA+B,SAAS;AAAA,EACjD,SAAS;AAAA,EAAgC,QAAQ;AAAA,EACjD,iBAAiBA,GAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,UAAa,EAAE,QAAQ,CAAC;AAAA,EACrE,QAAQ;AACV,CAAC,EAAE,OAAO;AAEH,IAAM,wCAAwCA,GAAE,OAAO,EAAE,gBAAgB,SAAS,SAAS,GAAG,SAAS,gCAAgC,aAAa,SAAS,SAAS,GAAG,OAAOA,GAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,QAAQ,EAAE,EAAE,CAAC,EAAE,OAAO;AACvO,IAAM,mCAAmCA,GAAE,OAAO;AAAA,EACvD,gBAAgB,SAAS,SAAS;AAAA,EAAG,SAAS;AAAA,EAAgC,WAAWA,GAAE,QAAQ;AAAA,EAAG,mBAAmBA,GAAE,OAAO,EAAE,IAAI,GAAK,EAAE,SAAS;AAAA,EACxJ,QAAQA,GAAE,MAAMA,GAAE,OAAO,EAAE,IAAI,UAAU,MAAMA,GAAE,OAAO,EAAE,IAAI,GAAG,EAAE,CAAC,EAAE,OAAO,CAAC,EAAE,IAAI,GAAG;AAAA,EACvF,OAAOA,GAAE,MAAMA,GAAE,OAAO,EAAE,IAAI,UAAU,OAAOA,GAAE,KAAK,CAAC,SAAS,cAAc,aAAa,CAAC,GAAG,cAAcA,GAAE,OAAO,EAAE,IAAI,GAAG,GAAG,UAAUA,GAAE,MAAMA,GAAE,OAAO,EAAE,IAAI,UAAU,OAAOA,GAAE,OAAO,EAAE,IAAI,GAAG,EAAE,CAAC,EAAE,OAAO,CAAC,EAAE,IAAI,GAAK,EAAE,CAAC,EAAE,OAAO,CAAC,EAAE,IAAI,GAAG;AAAA,EACpP,YAAY,SAAS,SAAS;AAChC,CAAC,EAAE,OAAO;AACH,IAAM,mCAAmCA,GAAE,OAAO;AAAA,EACvD,eAAeA,GAAE,QAAQ,wCAAwC;AAAA,EACjE,IAAI;AAAA,EAAU,UAAUA,GAAE,OAAO,EAAE,IAAI,EAAE,SAAS;AAAA,EAAG,SAAS;AAAA,EAC9D,QAAQA,GAAE,KAAK,CAAC,UAAU,WAAW,cAAc,aAAa,UAAU,WAAW,CAAC;AAAA,EACtF,gBAAgBA,GAAE,OAAO,EAAE,SAAS,UAAU,UAAU,UAAU,iBAAiBA,GAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,GAAG,mBAAmB,WAAW,CAAC,EAAE,OAAO,EAAE,SAAS;AAAA,EAClK,WAAWA,GAAE,IAAI,SAAS;AAAA,EAAG,WAAWA,GAAE,IAAI,SAAS,EAAE,SAAS;AAAA,EAAG,aAAaA,GAAE,IAAI,SAAS,EAAE,SAAS;AAAA,EAC5G,iBAAiBA,GAAE,QAAQ;AAAA,EAAG,iBAAiBA,GAAE,QAAQ;AAAA,EAAG,OAAOA,GAAE,OAAO,EAAE,IAAI,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS;AAAA,EACrG,eAAeA,GAAE,KAAK,CAAC,eAAe,UAAU,cAAc,gBAAgB,CAAC,EAAE,QAAQ,aAAa;AAAA,EACtG,QAAQA,GAAE,QAAQ,EAAE,SAAS;AAAA,EAAG,eAAeA,GAAE,OAAO,EAAE,IAAI,GAAG,EAAE,SAAS;AAAA,EAC5E,OAAOA,GAAE,OAAO,EAAE,MAAM,UAAU,SAASA,GAAE,OAAO,EAAE,IAAI,GAAK,EAAE,CAAC,EAAE,OAAO,EAAE,SAAS;AACxF,CAAC,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,YAAY;AAC1C,MAAK,MAAM,QAAQ,OAAO,SAAS,eAAgB,MAAM,mBAAmB,MAAO,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,gBAAgB,GAAG,SAAS,qDAAqD,CAAC;AAC/M,MAAI,MAAM,WAAW,gBAAgB,CAAC,MAAM,mBAAmB,CAAC,MAAM,mBAAmB,CAAC,MAAM,aAAc,SAAQ,SAAS,EAAE,MAAM,UAAU,SAAS,qEAAqE,CAAC;AAClO,CAAC;AAEM,IAAM,qCAAqCA,GAAE,OAAO,EAAE,gBAAgB,UAAU,SAAS,SAAS,SAAS,GAAG,OAAOA,GAAE,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,EAAE,IAAI,GAAG,EAAE,QAAQ,EAAE,EAAE,CAAC,EAAE,OAAO;AAC5K,IAAM,gCAAgCA,GAAE,OAAO,EAAE,OAAOA,GAAE,MAAM,gCAAgC,EAAE,IAAI,GAAG,GAAG,YAAY,SAAS,SAAS,EAAE,CAAC,EAAE,OAAO;AACtJ,IAAM,kCAAkCA,GAAE,OAAO;AAAA,EACtD,eAAeA,GAAE,QAAQ,uCAAuC;AAAA,EAChE,SAAS;AAAA,EACT,QAAQA,GAAE,OAAO,EAAE,IAAI,OAAS,EAAE,SAAS;AAAA;AAAA,EAE3C,UAAUA,GAAE,MAAMA,GAAE,OAAOA,GAAE,OAAO,GAAGA,GAAE,QAAQ,CAAC,CAAC,EAAE,IAAI,IAAK;AAAA,EAC9D,aAAa,wBAAwB,SAAS;AAAA;AAAA;AAAA,EAG9C,OAAO,gBAAgB,SAAS,EAAE,SAAS;AAAA,EAC3C,aAAaA,GAAE,OAAO,EAAE,QAAQA,GAAE,KAAK,CAAC,aAAa,oBAAoB,aAAa,aAAa,kBAAkB,qBAAqB,CAAC,GAAG,WAAWA,GAAE,QAAQ,GAAG,YAAY,gCAAgC,kBAAkB,WAAW,SAAS,GAAG,gBAAgB,WAAW,SAAS,GAAG,aAAa,WAAW,CAAC,EAAE,OAAO;AAAA,EACpU,oBAAoBA,GAAE,MAAM,QAAQ,EAAE,IAAI,IAAK;AAAA,EAC/C,aAAa;AACf,CAAC,EAAE,OAAO,EAAE,YAAY,CAAC,OAAO,YAAY;AAC1C,MAAI,CAAC,MAAM,MAAO;AAClB,MAAI,MAAM,YAAa,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,OAAO,GAAG,SAAS,4DAA4D,CAAC;AACjJ,MAAI,CAAC,CAAC,aAAa,QAAQ,EAAE,SAAS,MAAM,QAAQ,MAAM,KAAK,CAAC,MAAM,QAAQ,mBAAmB,MAAM,QAAQ,UAAU,MAAM,MAAM,SAAS,MAAM,QAAQ,kBAAkB,MAAM,MAAM,iBAAiB,MAAM,QAAQ,YAAY,MAAM,MAAM,UAAU,OAAO,OAAO,MAAM,MAAM,QAAS,SAAQ,SAAS,EAAE,MAAM,UAAU,MAAM,CAAC,OAAO,GAAG,SAAS,4DAA4D,CAAC;AAC1Z,CAAC;AAGM,IAAM,mCAAN,cAA+C,MAAM;AAAA,EAC1D,YAAqB,QAAyB,MAAc,SAAiB;AAAE,UAAM,OAAO;AAAvE;AAAyB;AAAiD,SAAK,OAAO;AAAA,EAAoC;AAAA,EAA1H;AAAA,EAAyB;AAChD;AAEO,IAAM,qCAAN,MAAyC;AAAA,EACrC;AAAA,EACT,YAAY,SAA+F;AACzG,UAAM,MAAM,IAAI,IAAI,QAAQ,OAAO;AACnC,QAAI,CAAC,QAAQ,OAAO,KAAK,KAAK,CAAC,QAAQ,OAAO,KAAK,KAAK,CAAC,CAAC,UAAU,OAAO,EAAE,SAAS,IAAI,QAAQ,KAAK,IAAI,YAAY,IAAI,YAAY,IAAI,UAAU,IAAI,KAAM,OAAM,IAAI,MAAM,yEAAyE;AACxP,SAAK,WAAW,EAAE,GAAG,SAAS,SAAS,IAAI,SAAS,EAAE,QAAQ,QAAQ,EAAE,EAAE;AAAA,EAC5E;AAAA,EACA,MAAM,OAAO,OAAyD,UAAoC,CAAC,GAAG;AAC5G,UAAM,UAAU,iCAAiC,MAAM,KAAK;AAC5D,QAAI,QAAQ,WAAW,KAAK,SAAS,OAAQ,OAAM,IAAI,MAAM,uCAAuC;AACpG,UAAM,SAAS,KAAK,SAAS,MAAM,KAAK,SAAS,IAAI,QAAQ,SAAS,QAAQ,MAAM,CAAC;AACrF,QAAI,cAAc,OAAO,OAAO,MAAM,cAAc,OAAO,EAAG,OAAM,IAAI,iCAAiC,KAAK,4BAA4B,qDAAqD;AAC/L,WAAO;AAAA,EACT;AAAA,EACA,MAAM,IAAI,IAAY,UAAoC,CAAC,GAAG;AAAE,WAAO,KAAK,SAAS,MAAM,KAAK,SAAS,IAAI,mBAAmB,SAAS,MAAM,EAAE,CAAC,CAAC,IAAI,OAAO,QAAW,QAAQ,MAAM,GAAG,EAAE;AAAA,EAAG;AAAA,EAC/L,MAAM,QAAQ,OAA8D,UAAoC,CAAC,GAAG;AAClH,UAAM,QAAQ,sCAAsC,MAAM,KAAK;AAC/D,UAAM,SAAS,IAAI,gBAAgB,EAAE,GAAI,MAAM,iBAAiB,EAAE,gBAAgB,MAAM,eAAe,IAAI,CAAC,GAAI,WAAW,MAAM,QAAQ,IAAI,UAAU,OAAO,MAAM,QAAQ,QAAQ,GAAG,aAAa,MAAM,QAAQ,aAAa,OAAO,OAAO,MAAM,KAAK,EAAE,CAAC;AAC3P,QAAI,MAAM,YAAa,QAAO,IAAI,eAAe,MAAM,WAAW;AAClE,UAAM,SAAS,iCAAiC,MAAM,MAAM,KAAK,SAAS,YAAY,MAAM,IAAI,OAAO,QAAW,QAAQ,MAAM,CAAC;AACjI,QAAI,OAAO,mBAAmB,MAAM,kBAAkB,cAAc,OAAO,OAAO,MAAM,cAAc,MAAM,OAAO,KAAK,OAAO,MAAM,SAAS,MAAM,SAAS,IAAI,IAAI,OAAO,MAAM,IAAI,UAAQ,KAAK,EAAE,CAAC,EAAE,SAAS,OAAO,MAAM,OAAQ,OAAM,IAAI,iCAAiC,KAAK,4BAA4B,4DAA4D;AAC9W,WAAO;AAAA,EACT;AAAA,EACA,MAAM,OAAO,IAAY,UAAoC,CAAC,GAAG;AAAE,WAAO,KAAK,SAAS,MAAM,KAAK,SAAS,IAAI,mBAAmB,SAAS,MAAM,EAAE,CAAC,CAAC,WAAW,QAAQ,QAAW,QAAQ,MAAM,GAAG,EAAE;AAAA,EAAG;AAAA,EAC1M,MAAM,OAAO,IAAY,UAAoC,CAAC,GAAG;AAC/D,UAAM,QAAQ,gCAAgC,MAAM,MAAM,KAAK,SAAS,IAAI,mBAAmB,SAAS,MAAM,EAAE,CAAC,CAAC,WAAW,OAAO,QAAW,QAAQ,MAAM,CAAC;AAC9J,SAAK,SAAS,MAAM,SAAS,EAAE;AAC/B,UAAM,EAAE,aAAa,GAAG,QAAQ,IAAI;AACpC,QAAI,CAAC,MAAM,QAAQ,mBAAmB,MAAM,gBAAgB,OAAO,MAAM,YAAa,OAAM,IAAI,iCAAiC,KAAK,4BAA4B,kCAAkC;AACpM,WAAO;AAAA,EACT;AAAA,EACA,MAAM,KAAK,OAA2D,UAAoC,CAAC,GAAG;AAC5G,UAAM,QAAQ,mCAAmC,MAAM,KAAK;AAC5D,UAAM,SAAS,IAAI,gBAAgB,EAAE,gBAAgB,MAAM,gBAAgB,OAAO,OAAO,MAAM,KAAK,EAAE,CAAC;AACvG,QAAI,MAAM,QAAS,QAAO,IAAI,WAAW,MAAM,OAAO;AACtD,UAAM,OAAO,8BAA8B,MAAM,MAAM,KAAK,SAAS,IAAI,MAAM,IAAI,OAAO,QAAW,QAAQ,MAAM,CAAC;AACpH,QAAI,KAAK,MAAM,SAAS,MAAM,SAAS,IAAI,IAAI,KAAK,MAAM,IAAI,UAAQ,KAAK,EAAE,CAAC,EAAE,SAAS,KAAK,MAAM,UAAU,KAAK,MAAM,KAAK,UAAQ,KAAK,SAAS,IAAI,EAAE,QAAQ,mBAAmB,MAAM,cAAc,EAAG,OAAM,IAAI,iCAAiC,KAAK,yBAAyB,+CAA+C;AACpU,WAAO;AAAA,EACT;AAAA,EACA,SAAS,OAAgB,IAAyC;AAChE,UAAM,SAAS,iCAAiC,MAAM,KAAK;AAC3D,QAAI,OAAO,QAAQ,WAAW,KAAK,SAAS,UAAW,MAAM,OAAO,OAAO,GAAK,OAAM,IAAI,iCAAiC,KAAK,6BAA6B,2DAA2D;AACxN,WAAO;AAAA,EACT;AAAA,EACA,MAAM,SAAS,MAAc,QAAgB,MAAe,QAAwC;AAClG,UAAM,WAAW,OAAO,KAAK,SAAS,SAAS,WAAW,OAAO,GAAG,KAAK,SAAS,OAAO,6BAA6B,IAAI,IAAI,EAAE,QAAQ,SAAS,EAAE,eAAe,UAAU,KAAK,SAAS,MAAM,IAAI,sBAAsB,KAAK,SAAS,QAAQ,gBAAgB,oBAAoB,QAAQ,mBAAmB,GAAG,GAAI,SAAS,SAAY,CAAC,IAAI,EAAE,MAAM,KAAK,UAAU,IAAI,EAAE,GAAI,QAAQ,UAAU,QAAQ,CAAC;AAC5Y,UAAM,SAAS,SAAS,MAAM,UAAU;AACxC,QAAI,CAAC,OAAQ,OAAM,IAAI,iCAAiC,SAAS,QAAQ,0BAA0B,6BAA6B;AAChI,QAAI,QAAQ;AAAG,QAAI,OAAO;AAAI,UAAM,UAAU,IAAI,YAAY;AAC9D,QAAI;AACF,aAAO,MAAM;AACX,cAAM,QAAQ,MAAM,OAAO,KAAK;AAAG,YAAI,MAAM,KAAM;AACnD,iBAAS,MAAM,MAAM;AACrB,YAAI,QAAQ,SAAY,OAAM,IAAI,iCAAiC,KAAK,8BAA8B,2CAA2C;AACjJ,gBAAQ,QAAQ,OAAO,MAAM,OAAO,EAAE,QAAQ,KAAK,CAAC;AAAA,MACtD;AACA,cAAQ,QAAQ,OAAO;AAAA,IACzB,UAAE;AAAU,YAAM,OAAO,OAAO;AAAG,aAAO,YAAY;AAAA,IAAG;AACzD,UAAM,QAAiB,KAAK,MAAM,IAAI;AACtC,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,QAAQA,GAAE,OAAO,EAAE,MAAM,UAAU,SAASA,GAAE,OAAO,EAAE,IAAI,GAAK,EAAE,CAAC,EAAE,UAAU,KAAK;AAC1F,YAAM,IAAI,iCAAiC,SAAS,QAAQ,MAAM,UAAU,MAAM,KAAK,OAAO,0BAA0B,MAAM,UAAU,MAAM,KAAK,UAAU,yBAAyB;AAAA,IACxL;AACA,WAAO;AAAA,EACT;AACF;",
-  "names": ["z", "z"]
-}