openpond-sdk 0.9.9 → 0.10.0

package/dist/types/packages/sdk/src/experiment-client.d.ts

@@ -1,5 +1,6 @@
 import { z } from "zod";
-import { ExperimentListQuerySchema, SaveExperimentSchema, StartExperimentSchema, type ExperimentDefinitionRef } from "./experiment-contracts.js";
+import { ExperimentListQuerySchema } from "./experiment-contracts.js";
+import { RunExperimentSchema } from "./experiment-run-contracts.js";
 import { PrepareHarnessExperimentSchema } from "./experiment-contracts.js";
 import { type ExperimentScoringRequest } from "./experiment-scoring-contracts.js";
 export declare class OpenPondExperimentError extends Error {
@@ -7,7 +8,7 @@ export declare class OpenPondExperimentError extends Error {
     readonly code: string;
     constructor(status: number, code: string, message: string);
 }
-/** Saved configuration and execution identities share the host's service. */
+/** One Experiment is one run with its immutable configuration and results. */
 export declare class OpenPondExperimentsClient {
     private readonly options;
     private readonly baseUrl;
@@ -17,6 +18,30 @@ export declare class OpenPondExperimentsClient {
         teamId: string;
         fetch?: typeof fetch;
     });
+    /** Discover released workspace Harnesses independently of Project targets. */
+    harnessSources(options?: {
+        cursor?: string;
+        signal?: AbortSignal;
+    }): Promise<{
+        teamId: string;
+        items: {
+            name: string;
+            source: {
+                harnessRelease: {
+                    id: string;
+                    contentHash: string;
+                };
+                agentSnapshot: {
+                    id: string;
+                    contentHash: string;
+                };
+                sourcePackageHash: string;
+            };
+            ready: boolean;
+            reason: string | null;
+        }[];
+        nextCursor: string | null;
+    }>;
     /** Resolves a released target without saving a definition or dispatching it. */
     prepareHarness(value: z.input<typeof PrepareHarnessExperimentSchema>, signal?: AbortSignal): Promise<{
         request: {
@@ -35,6 +60,17 @@ export declare class OpenPondExperimentsClient {
                 maxOutputTokens: number;
                 temperature: number;
                 topP: number;
+                harness?: {
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    agentSnapshot: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    sourcePackageHash: string;
+                } | undefined;
                 messages?: {
                     role: "system" | "user" | "assistant";
                     content: string;
@@ -199,6 +235,7 @@ export declare class OpenPondExperimentsClient {
                 revision: number;
                 contentHash: string;
             } | null;
+            name?: string | undefined;
             mappings?: {
                 destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
                 path: string[];
@@ -207,71 +244,41 @@ export declare class OpenPondExperimentsClient {
             }[] | undefined;
         }[];
     }>;
-    save(value: z.input<typeof SaveExperimentSchema>, signal?: AbortSignal): Promise<{
-        schemaVersion: "openpond.experimentDefinition.v1";
-        id: string;
-        teamId: string;
-        revision: number;
-        request: {
-            schemaVersion: "openpond.modelTasksetRunRequest.v1";
-            operationId: string;
+    run(value: z.input<typeof RunExperimentSchema>, signal?: AbortSignal): Promise<{
+        summary: {
+            schemaVersion: "openpond.modelTasksetRunSummary.v1";
+            id: string;
+            revision: number;
             teamId: string;
             modelProjectId: string | null;
+            operationId: string;
             taskset: {
                 id: string;
                 contentHash: string;
                 revision: number;
             };
-            policy: {
-                kind: "hosted_chat";
-                modelId: string;
-                maxOutputTokens: number;
-                temperature: number;
-                topP: number;
-                messages?: {
-                    role: "system" | "user" | "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";
+            policyKind: "hosted_chat" | "fixture" | "hosted_harness";
+            manifestHash: string;
+            status: "failed" | "cancelled" | "queued" | "running" | "cancelling" | "completed";
+            totalCount: number;
+            counts: {
+                pending: number;
+                running: number;
+                completed: number;
+                failed: number;
+                cancelled: number;
             };
-            population: {
-                receiptId: string;
-                taskId: string;
-                seed: string;
-                fixtureId: string | null;
-            }[];
+            createdAt: string;
+            startedAt: string | null;
+            completedAt: string | null;
+            cleanupComplete: boolean;
+            resultAvailable: boolean;
+            score: number | null;
+            metricName: string;
+            error: {
+                code: string;
+                message: string;
+            } | null;
             name?: string | undefined;
             project?: {
                 id: string;
@@ -280,36 +287,6 @@ export declare class OpenPondExperimentsClient {
                 targetId: string | null;
             } | undefined;
         };
-        maximumCostUsd: number;
-        graders: {
-            id: string;
-            version: string;
-            contentHash: string;
-            feedbackKey: string;
-            release: {
-                id: string;
-                revision: number;
-                contentHash: string;
-            } | null;
-            mappings?: {
-                destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
-                path: string[];
-                source: "input" | "output" | "expectedOutput" | "evaluatorContext";
-                sourcePath: string[];
-            }[] | undefined;
-        }[];
-        createdAt: string;
-        updatedAt: string;
-        contentHash: string;
-    }>;
-    get(id: string, options?: {
-        reference?: ExperimentDefinitionRef;
-        signal?: AbortSignal;
-    }): Promise<{
-        schemaVersion: "openpond.experimentDefinition.v1";
-        id: string;
-        teamId: string;
-        revision: number;
         request: {
             schemaVersion: "openpond.modelTasksetRunRequest.v1";
             operationId: string;
@@ -326,6 +303,17 @@ export declare class OpenPondExperimentsClient {
                 maxOutputTokens: number;
                 temperature: number;
                 topP: number;
+                harness?: {
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    agentSnapshot: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    sourcePackageHash: string;
+                } | undefined;
                 messages?: {
                     role: "system" | "user" | "assistant";
                     content: string;
@@ -378,127 +366,220 @@ export declare class OpenPondExperimentsClient {
                 targetId: string | null;
             } | undefined;
         };
-        maximumCostUsd: number;
-        graders: {
+        manifest: {
+            schemaVersion: "openpond.tasksetRunManifest.v1";
             id: string;
-            version: string;
-            contentHash: string;
-            feedbackKey: string;
-            release: {
+            tasksetRelease: {
                 id: string;
-                revision: number;
                 contentHash: string;
-            } | null;
-            mappings?: {
-                destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
-                path: string[];
-                source: "input" | "output" | "expectedOutput" | "evaluatorContext";
-                sourcePath: string[];
-            }[] | undefined;
-        }[];
-        createdAt: string;
-        updatedAt: string;
-        contentHash: string;
-    }>;
-    list(value?: z.input<typeof ExperimentListQuerySchema>, signal?: AbortSignal): Promise<{
-        items: {
-            schemaVersion: "openpond.experimentDefinition.v1";
-            id: string;
-            teamId: string;
-            revision: number;
-            request: {
-                schemaVersion: "openpond.modelTasksetRunRequest.v1";
-                operationId: string;
-                teamId: string;
-                modelProjectId: string | null;
-                taskset: {
+            };
+            packageHash: string;
+            execution: {
+                kind: "taskset";
+                environmentRelease: {
                     id: string;
                     contentHash: string;
-                    revision: number;
-                };
-                policy: {
-                    kind: "hosted_chat";
-                    modelId: string;
-                    maxOutputTokens: number;
-                    temperature: number;
-                    topP: number;
-                    messages?: {
-                        role: "system" | "user" | "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";
                 };
-                population: {
-                    receiptId: string;
-                    taskId: string;
-                    seed: string;
-                    fixtureId: string | null;
-                }[];
-                name?: string | undefined;
-                project?: {
+                verifierSetRelease: {
                     id: string;
-                    revision: number;
                     contentHash: string;
-                    targetId: string | null;
-                } | undefined;
-            };
-            maximumCostUsd: number;
-            graders: {
-                id: string;
-                version: string;
-                contentHash: string;
-                feedbackKey: string;
-                release: {
+                };
+                policyHash: string;
+            } | {
+                kind: "harness";
+                harnessRelease: {
                     id: string;
-                    revision: number;
                     contentHash: string;
-                } | null;
-                mappings?: {
-                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
-                    path: string[];
-                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
-                    sourcePath: string[];
-                }[] | undefined;
+                };
+            };
+            policy: {
+                kind: "model";
+                model: {
+                    provider: string;
+                    model: string;
+                    revision: string | null;
+                    artifactHash: string | null;
+                    tokenizerRevision: string | null;
+                    chatTemplateHash: string | null;
+                };
+                configurationHash: string;
+            } | {
+                kind: "fixture";
+            };
+            gradingRole: "evaluation";
+            metricPolicy: {
+                schemaVersion: "openpond.tasksetMetricPolicy.v1";
+                primaryMetric: string;
+                aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
+                missingReward: "zero" | "exclude";
+                customAggregator: {
+                    module: string;
+                    exportName: string;
+                    contentHash: string;
+                    timeoutMs: number;
+                    networkPolicy: "none";
+                } | null;
+                taskWeights?: Record<string, number> | undefined;
+            };
+            population: {
+                receiptId: string;
+                taskId: string;
+                seed: string;
+                fixtureId: string | null;
             }[];
+            runtimeTarget: {
+                adapterId: string;
+                placement: "local" | "remote" | "colocated" | "provider_native";
+                runtimeVersion: string;
+                capabilityReceipt: string;
+            };
+            limits: {
+                maxTurns: number;
+                timeoutMs: number;
+                maxOutputBytes: number;
+                maximumSpendUsd: number | null;
+            };
             createdAt: string;
-            updatedAt: string;
+            metadata: Record<string, unknown>;
             contentHash: string;
-        }[];
-        nextCursor: string | null;
+            profileEvaluation?: {
+                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;
+            } | undefined;
+        };
+        policySnapshot: {
+            modelId: string;
+            provider: string;
+            upstreamModelId: string;
+            configurationHash: string;
+        } | null;
+        configuration: {
+            request: {
+                schemaVersion: "openpond.modelTasksetRunRequest.v1";
+                operationId: string;
+                teamId: string;
+                modelProjectId: string | null;
+                taskset: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                policy: {
+                    kind: "hosted_chat";
+                    modelId: string;
+                    maxOutputTokens: number;
+                    temperature: number;
+                    topP: number;
+                    harness?: {
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        agentSnapshot: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        sourcePackageHash: string;
+                    } | undefined;
+                    messages?: {
+                        role: "system" | "user" | "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";
+                };
+                population: {
+                    receiptId: string;
+                    taskId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                }[];
+                name?: string | undefined;
+                project?: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                    targetId: string | null;
+                } | undefined;
+            };
+            maximumCostUsd: number;
+            graders: {
+                id: string;
+                version: string;
+                contentHash: string;
+                feedbackKey: string;
+                release: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                } | null;
+                name?: string | undefined;
+                mappings?: {
+                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    path: string[];
+                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    sourcePath: string[];
+                }[] | undefined;
+            }[];
+            configurationHash: string;
+            admissionRequestHash?: string | undefined;
+            sourceExperimentId?: string | undefined;
+        };
     }>;
-    start(value: z.input<typeof StartExperimentSchema>, signal?: AbortSignal): Promise<{
+    get(id: string, options?: {
+        signal?: AbortSignal;
+    }): Promise<{
         summary: {
             schemaVersion: "openpond.modelTasksetRunSummary.v1";
             id: string;
@@ -557,6 +638,17 @@ export declare class OpenPondExperimentsClient {
                 maxOutputTokens: number;
                 temperature: number;
                 topP: number;
+                harness?: {
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    agentSnapshot: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    sourcePackageHash: string;
+                } | undefined;
                 messages?: {
                     role: "system" | "user" | "assistant";
                     content: string;
@@ -716,157 +808,111 @@ export declare class OpenPondExperimentsClient {
             upstreamModelId: string;
             configurationHash: string;
         } | null;
-    }>;
-    result(executionId: string, signal?: AbortSignal): Promise<{
-        executionId: string;
-        passId: string | null;
-        manifest: {
-            schemaVersion: "openpond.experimentManifest.v1";
-            id: string;
-            name: string;
-            teamId: string;
-            operationId: string;
-            maximumCostUsd: number | null;
-            dataset: {
-                id: string;
-                contentHash: string;
-                revision: number;
-            };
-            target: {
-                kind: "model";
-                modelId: string;
-                configurationHash: string;
-            } | {
-                kind: "harness";
-                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;
-                };
-                model: {
-                    modelId: string;
-                    configurationHash: string;
-                };
-            } | {
-                kind: "fixture";
-                configurationHash: string;
-            };
-            evaluators: {
-                release: {
+        configuration: {
+            request: {
+                schemaVersion: "openpond.modelTasksetRunRequest.v1";
+                operationId: string;
+                teamId: string;
+                modelProjectId: string | null;
+                taskset: {
                     id: string;
                     contentHash: string;
-                    revision: string | number;
+                    revision: number;
                 };
-                feedbackKey: string;
-                output: "boolean" | "score" | "category";
-                categories: string[];
-                configurationHash?: string | undefined;
-            }[];
-            population: {
-                caseId: string;
-                seed: string;
-                fixtureId: string | null;
-            }[];
-            createdAt: string;
-            contentHash: string;
-            execution?: {
-                packageHash: string;
-                runtimeTargetHash: string;
-                metricPolicyHash: string;
-            } | undefined;
-            lineage?: {
-                definition: {
+                policy: {
+                    kind: "hosted_chat";
+                    modelId: string;
+                    maxOutputTokens: number;
+                    temperature: number;
+                    topP: number;
+                    harness?: {
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        agentSnapshot: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        sourcePackageHash: string;
+                    } | undefined;
+                    messages?: {
+                        role: "system" | "user" | "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";
+                };
+                population: {
+                    receiptId: string;
+                    taskId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                }[];
+                name?: string | undefined;
+                project?: {
                     id: string;
-                    contentHash: string;
                     revision: number;
-                } | null;
-                execution: {
-                    id: string;
                     contentHash: string;
-                };
-                scoringPassId: string | null;
-            } | undefined;
-        };
-        result: {
-            schemaVersion: "openpond.experimentResult.v1";
-            manifest: {
+                    targetId: string | null;
+                } | undefined;
+            };
+            maximumCostUsd: number;
+            graders: {
                 id: string;
+                version: string;
                 contentHash: string;
-            };
-            status: "failed" | "completed" | "cancelled";
-            cases: {
-                identity: {
-                    caseId: string;
-                    seed: string;
-                    fixtureId: string | null;
-                };
-                status: "failed" | "completed" | "cancelled" | "unavailable";
-                output: unknown;
-                error: {
-                    code: string;
-                    message: string;
-                } | null;
-                feedback: {
-                    feedbackKey: string;
-                    evaluator: {
-                        id: string;
-                        contentHash: string;
-                        revision: string | number;
-                    };
-                    status: "pending" | "failed" | "scored" | "unavailable";
-                    value: string | number | boolean | null;
-                    reasoning: string | null;
-                    evidenceRefs: {
-                        id: string;
-                        contentHash: string;
-                        mediaType: string | null;
-                        sizeBytes: number | null;
-                    }[];
-                    passed?: boolean | null | undefined;
-                }[];
-                usage: {
-                    inputTokens: number | null;
-                    outputTokens: number | null;
-                    totalTokens: number | null;
-                    costUsd: number | null;
-                    latencyMs: number | null;
-                };
-                traceRef: {
+                feedbackKey: string;
+                release: {
                     id: string;
+                    revision: number;
                     contentHash: string;
-                    mediaType: string | null;
-                    sizeBytes: number | null;
                 } | null;
-                startedAt: string | null;
-                completedAt: string | null;
+                name?: string | undefined;
+                mappings?: {
+                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    path: string[];
+                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    sourcePath: string[];
+                }[] | undefined;
             }[];
-            completedAt: string;
-            contentHash: string;
+            configurationHash: string;
+            admissionRequestHash?: string | undefined;
+            sourceExperimentId?: string | undefined;
         };
     }>;
-    executions(definitionId: string, options?: {
-        afterId?: string;
-        signal?: AbortSignal;
-    }): Promise<{
+    list(value?: z.input<typeof ExperimentListQuerySchema>, signal?: AbortSignal): Promise<{
         items: {
             summary: {
                 schemaVersion: "openpond.modelTasksetRunSummary.v1";
@@ -926,6 +972,17 @@ export declare class OpenPondExperimentsClient {
                     maxOutputTokens: number;
                     temperature: number;
                     topP: number;
+                    harness?: {
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        agentSnapshot: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        sourcePackageHash: string;
+                    } | undefined;
                     messages?: {
                         role: "system" | "user" | "assistant";
                         content: string;
@@ -1085,76 +1142,133 @@ export declare class OpenPondExperimentsClient {
                 upstreamModelId: string;
                 configurationHash: string;
             } | null;
+            configuration: {
+                request: {
+                    schemaVersion: "openpond.modelTasksetRunRequest.v1";
+                    operationId: string;
+                    teamId: string;
+                    modelProjectId: string | null;
+                    taskset: {
+                        id: string;
+                        contentHash: string;
+                        revision: number;
+                    };
+                    policy: {
+                        kind: "hosted_chat";
+                        modelId: string;
+                        maxOutputTokens: number;
+                        temperature: number;
+                        topP: number;
+                        harness?: {
+                            harnessRelease: {
+                                id: string;
+                                contentHash: string;
+                            };
+                            agentSnapshot: {
+                                id: string;
+                                contentHash: string;
+                            };
+                            sourcePackageHash: string;
+                        } | undefined;
+                        messages?: {
+                            role: "system" | "user" | "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";
+                    };
+                    population: {
+                        receiptId: string;
+                        taskId: string;
+                        seed: string;
+                        fixtureId: string | null;
+                    }[];
+                    name?: string | undefined;
+                    project?: {
+                        id: string;
+                        revision: number;
+                        contentHash: string;
+                        targetId: string | null;
+                    } | undefined;
+                };
+                maximumCostUsd: number;
+                graders: {
+                    id: string;
+                    version: string;
+                    contentHash: string;
+                    feedbackKey: string;
+                    release: {
+                        id: string;
+                        revision: number;
+                        contentHash: string;
+                    } | null;
+                    name?: string | undefined;
+                    mappings?: {
+                        destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                        path: string[];
+                        source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                        sourcePath: string[];
+                    }[] | undefined;
+                }[];
+                configurationHash: string;
+                admissionRequestHash?: string | undefined;
+                sourceExperimentId?: string | undefined;
+            };
         }[];
         nextCursor: string | null;
     }>;
-    execution(id: string, signal?: AbortSignal): Promise<{
-        summary: {
-            schemaVersion: "openpond.modelTasksetRunSummary.v1";
+    result(executionId: string, signal?: AbortSignal): Promise<{
+        executionId: string;
+        passId: string | null;
+        manifest: {
+            schemaVersion: "openpond.experimentManifest.v1";
             id: string;
-            revision: number;
+            name: string;
             teamId: string;
-            modelProjectId: string | null;
             operationId: string;
-            taskset: {
-                id: string;
-                contentHash: string;
-                revision: number;
-            };
-            policyKind: "hosted_chat" | "fixture" | "hosted_harness";
-            manifestHash: string;
-            status: "failed" | "cancelled" | "queued" | "running" | "cancelling" | "completed";
-            totalCount: number;
-            counts: {
-                pending: number;
-                running: number;
-                completed: number;
-                failed: number;
-                cancelled: number;
-            };
-            createdAt: string;
-            startedAt: string | null;
-            completedAt: string | null;
-            cleanupComplete: boolean;
-            resultAvailable: boolean;
-            score: number | null;
-            metricName: string;
-            error: {
-                code: string;
-                message: string;
-            } | null;
-            name?: string | undefined;
-            project?: {
-                id: string;
-                revision: number;
-                contentHash: string;
-                targetId: string | null;
-            } | undefined;
-        };
-        request: {
-            schemaVersion: "openpond.modelTasksetRunRequest.v1";
-            operationId: string;
-            teamId: string;
-            modelProjectId: string | null;
-            taskset: {
+            maximumCostUsd: number | null;
+            dataset: {
                 id: string;
                 contentHash: string;
                 revision: number;
             };
-            policy: {
-                kind: "hosted_chat";
+            target: {
+                kind: "model";
                 modelId: string;
-                maxOutputTokens: number;
-                temperature: number;
-                topP: number;
-                messages?: {
-                    role: "system" | "user" | "assistant";
-                    content: string;
-                }[] | undefined;
+                configurationHash: string;
             } | {
-                kind: "hosted_harness";
-                modelId: string;
-                profileRepositoryId: string;
+                kind: "harness";
                 source: {
                     profileId: string;
                     sourceRevision: string;
@@ -1179,133 +1293,127 @@ export declare class OpenPondExperimentsClient {
                     };
                     environmentHash: string;
                 };
-                modelConfigurationHash: string;
-                packageHash: string;
-                reasoningEffort?: "high" | "low" | "off" | "medium" | "xhigh" | "max" | undefined;
-            } | {
-                kind: "fixture";
-            };
-            population: {
-                receiptId: string;
-                taskId: string;
-                seed: string;
-                fixtureId: string | null;
-            }[];
-            name?: string | undefined;
-            project?: {
-                id: string;
-                revision: number;
-                contentHash: string;
-                targetId: string | null;
-            } | undefined;
-        };
-        manifest: {
-            schemaVersion: "openpond.tasksetRunManifest.v1";
-            id: string;
-            tasksetRelease: {
-                id: string;
-                contentHash: string;
-            };
-            packageHash: string;
-            execution: {
-                kind: "taskset";
-                environmentRelease: {
-                    id: string;
-                    contentHash: string;
-                };
-                verifierSetRelease: {
-                    id: string;
-                    contentHash: string;
+                model: {
+                    modelId: string;
+                    configurationHash: string;
                 };
-                policyHash: string;
             } | {
-                kind: "harness";
-                harnessRelease: {
-                    id: string;
-                    contentHash: string;
+                kind: "agent";
+                source: {
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    agentSnapshot: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    sourcePackageHash: string;
                 };
-            };
-            policy: {
-                kind: "model";
                 model: {
-                    provider: string;
-                    model: string;
-                    revision: string | null;
-                    artifactHash: string | null;
-                    tokenizerRevision: string | null;
-                    chatTemplateHash: string | null;
+                    modelId: string;
+                    configurationHash: string;
                 };
-                configurationHash: string;
             } | {
                 kind: "fixture";
+                configurationHash: string;
             };
-            gradingRole: "evaluation";
-            metricPolicy: {
-                schemaVersion: "openpond.tasksetMetricPolicy.v1";
-                primaryMetric: string;
-                aggregation: "custom" | "mean_score" | "pass_rate" | "weighted_mean";
-                missingReward: "zero" | "exclude";
-                customAggregator: {
-                    module: string;
-                    exportName: string;
+            evaluators: {
+                release: {
+                    id: string;
                     contentHash: string;
-                    timeoutMs: number;
-                    networkPolicy: "none";
-                } | null;
-                taskWeights?: Record<string, number> | undefined;
-            };
+                    revision: string | number;
+                };
+                feedbackKey: string;
+                output: "boolean" | "score" | "category";
+                categories: string[];
+                configurationHash?: string | undefined;
+            }[];
             population: {
-                receiptId: string;
-                taskId: string;
+                caseId: string;
                 seed: string;
                 fixtureId: string | null;
             }[];
-            runtimeTarget: {
-                adapterId: string;
-                placement: "local" | "remote" | "colocated" | "provider_native";
-                runtimeVersion: string;
-                capabilityReceipt: string;
-            };
-            limits: {
-                maxTurns: number;
-                timeoutMs: number;
-                maxOutputBytes: number;
-                maximumSpendUsd: number | null;
-            };
             createdAt: string;
-            metadata: Record<string, unknown>;
             contentHash: string;
-            profileEvaluation?: {
-                profileId: string;
-                sourceRevision: string;
-                harnessRelease: {
+            execution?: {
+                packageHash: string;
+                runtimeTargetHash: string;
+                metricPolicyHash: string;
+                compatibility?: {
+                    protocol: "openpond.evaluation-execution.v1";
+                    targetKind: "agent" | "model" | "fixture" | "harness";
+                } | undefined;
+            } | undefined;
+            lineage?: {
+                definition: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                } | null;
+                execution: {
                     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;
+                scoringPassId: string | null;
             } | undefined;
         };
-        policySnapshot: {
-            modelId: string;
-            provider: string;
-            upstreamModelId: string;
-            configurationHash: string;
-        } | null;
+        result: {
+            schemaVersion: "openpond.experimentResult.v1";
+            manifest: {
+                id: string;
+                contentHash: string;
+            };
+            status: "failed" | "completed" | "cancelled";
+            cases: {
+                identity: {
+                    caseId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                };
+                status: "failed" | "completed" | "cancelled" | "unavailable";
+                output: unknown;
+                error: {
+                    code: string;
+                    message: string;
+                } | null;
+                feedback: {
+                    feedbackKey: string;
+                    evaluator: {
+                        id: string;
+                        contentHash: string;
+                        revision: string | number;
+                    };
+                    status: "pending" | "failed" | "scored" | "unavailable";
+                    value: string | number | boolean | null;
+                    reasoning: string | null;
+                    evidenceRefs: {
+                        id: string;
+                        contentHash: string;
+                        mediaType: string | null;
+                        sizeBytes: number | null;
+                    }[];
+                    passed?: boolean | null | undefined;
+                }[];
+                usage: {
+                    inputTokens: number | null;
+                    outputTokens: number | null;
+                    totalTokens: number | null;
+                    costUsd: number | null;
+                    latencyMs: number | null;
+                };
+                traceRef: {
+                    id: string;
+                    contentHash: string;
+                    mediaType: string | null;
+                    sizeBytes: number | null;
+                } | null;
+                startedAt: string | null;
+                completedAt: string | null;
+            }[];
+            completedAt: string;
+            contentHash: string;
+        };
     }>;
     cancel(id: string, signal?: AbortSignal): Promise<{
         summary: {
@@ -1366,6 +1474,17 @@ export declare class OpenPondExperimentsClient {
                 maxOutputTokens: number;
                 temperature: number;
                 topP: number;
+                harness?: {
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    agentSnapshot: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    sourcePackageHash: string;
+                } | undefined;
                 messages?: {
                     role: "system" | "user" | "assistant";
                     content: string;
@@ -1525,9 +1644,112 @@ export declare class OpenPondExperimentsClient {
             upstreamModelId: string;
             configurationHash: string;
         } | null;
+        configuration: {
+            request: {
+                schemaVersion: "openpond.modelTasksetRunRequest.v1";
+                operationId: string;
+                teamId: string;
+                modelProjectId: string | null;
+                taskset: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                policy: {
+                    kind: "hosted_chat";
+                    modelId: string;
+                    maxOutputTokens: number;
+                    temperature: number;
+                    topP: number;
+                    harness?: {
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        agentSnapshot: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        sourcePackageHash: string;
+                    } | undefined;
+                    messages?: {
+                        role: "system" | "user" | "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";
+                };
+                population: {
+                    receiptId: string;
+                    taskId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                }[];
+                name?: string | undefined;
+                project?: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                    targetId: string | null;
+                } | undefined;
+            };
+            maximumCostUsd: number;
+            graders: {
+                id: string;
+                version: string;
+                contentHash: string;
+                feedbackKey: string;
+                release: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                } | null;
+                name?: string | undefined;
+                mappings?: {
+                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    path: string[];
+                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    sourcePath: string[];
+                }[] | undefined;
+            }[];
+            configurationHash: string;
+            admissionRequestHash?: string | undefined;
+            sourceExperimentId?: string | undefined;
+        };
     }>;
-    /** Retry intentionally creates another execution; a transport retry reuses the same operation id. */
-    retry(id: string, operationId: string, signal?: AbortSignal): Promise<{
+    /** Duplicate uses the original run snapshot; edits are submitted with run(). */
+    duplicate(id: string, operationId: string, signal?: AbortSignal): Promise<{
         summary: {
             schemaVersion: "openpond.modelTasksetRunSummary.v1";
             id: string;
@@ -1586,6 +1808,17 @@ export declare class OpenPondExperimentsClient {
                 maxOutputTokens: number;
                 temperature: number;
                 topP: number;
+                harness?: {
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    agentSnapshot: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    sourcePackageHash: string;
+                } | undefined;
                 messages?: {
                     role: "system" | "user" | "assistant";
                     content: string;
@@ -1745,6 +1978,109 @@ export declare class OpenPondExperimentsClient {
             upstreamModelId: string;
             configurationHash: string;
         } | null;
+        configuration: {
+            request: {
+                schemaVersion: "openpond.modelTasksetRunRequest.v1";
+                operationId: string;
+                teamId: string;
+                modelProjectId: string | null;
+                taskset: {
+                    id: string;
+                    contentHash: string;
+                    revision: number;
+                };
+                policy: {
+                    kind: "hosted_chat";
+                    modelId: string;
+                    maxOutputTokens: number;
+                    temperature: number;
+                    topP: number;
+                    harness?: {
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        agentSnapshot: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        sourcePackageHash: string;
+                    } | undefined;
+                    messages?: {
+                        role: "system" | "user" | "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";
+                };
+                population: {
+                    receiptId: string;
+                    taskId: string;
+                    seed: string;
+                    fixtureId: string | null;
+                }[];
+                name?: string | undefined;
+                project?: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                    targetId: string | null;
+                } | undefined;
+            };
+            maximumCostUsd: number;
+            graders: {
+                id: string;
+                version: string;
+                contentHash: string;
+                feedbackKey: string;
+                release: {
+                    id: string;
+                    revision: number;
+                    contentHash: string;
+                } | null;
+                name?: string | undefined;
+                mappings?: {
+                    destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    path: string[];
+                    source: "input" | "output" | "expectedOutput" | "evaluatorContext";
+                    sourcePath: string[];
+                }[] | undefined;
+            }[];
+            configurationHash: string;
+            admissionRequestHash?: string | undefined;
+            sourceExperimentId?: string | undefined;
+        };
     }>;
     score(value: ExperimentScoringRequest, signal?: AbortSignal): Promise<{
         schemaVersion: "openpond.experimentScoringPass.v1";
@@ -1783,6 +2119,7 @@ export declare class OpenPondExperimentsClient {
                 revision: number;
                 contentHash: string;
             } | null;
+            name?: string | undefined;
             mappings?: {
                 destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
                 path: string[];
@@ -1847,6 +2184,7 @@ export declare class OpenPondExperimentsClient {
                 revision: number;
                 contentHash: string;
             } | null;
+            name?: string | undefined;
             mappings?: {
                 destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
                 path: string[];
@@ -1915,6 +2253,7 @@ export declare class OpenPondExperimentsClient {
                     revision: number;
                     contentHash: string;
                 } | null;
+                name?: string | undefined;
                 mappings?: {
                     destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
                     path: string[];
@@ -1981,6 +2320,7 @@ export declare class OpenPondExperimentsClient {
                 revision: number;
                 contentHash: string;
             } | null;
+            name?: string | undefined;
             mappings?: {
                 destination: "input" | "output" | "expectedOutput" | "evaluatorContext";
                 path: string[];
@@ -2057,6 +2397,23 @@ export declare class OpenPondExperimentsClient {
                     modelId: string;
                     configurationHash: string;
                 };
+            } | {
+                kind: "agent";
+                source: {
+                    harnessRelease: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    agentSnapshot: {
+                        id: string;
+                        contentHash: string;
+                    };
+                    sourcePackageHash: string;
+                };
+                model: {
+                    modelId: string;
+                    configurationHash: string;
+                };
             } | {
                 kind: "fixture";
                 configurationHash: string;
@@ -2083,6 +2440,10 @@ export declare class OpenPondExperimentsClient {
                 packageHash: string;
                 runtimeTargetHash: string;
                 metricPolicyHash: string;
+                compatibility?: {
+                    protocol: "openpond.evaluation-execution.v1";
+                    targetKind: "agent" | "model" | "fixture" | "harness";
+                } | undefined;
             } | undefined;
             lineage?: {
                 definition: {
@@ -2204,6 +2565,23 @@ export declare class OpenPondExperimentsClient {
                         modelId: string;
                         configurationHash: string;
                     };
+                } | {
+                    kind: "agent";
+                    source: {
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        agentSnapshot: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        sourcePackageHash: string;
+                    };
+                    model: {
+                        modelId: string;
+                        configurationHash: string;
+                    };
                 } | {
                     kind: "fixture";
                     configurationHash: string;
@@ -2230,6 +2608,10 @@ export declare class OpenPondExperimentsClient {
                     packageHash: string;
                     runtimeTargetHash: string;
                     metricPolicyHash: string;
+                    compatibility?: {
+                        protocol: "openpond.evaluation-execution.v1";
+                        targetKind: "agent" | "model" | "fixture" | "harness";
+                    } | undefined;
                 } | undefined;
                 lineage?: {
                     definition: {
@@ -2350,6 +2732,23 @@ export declare class OpenPondExperimentsClient {
                         modelId: string;
                         configurationHash: string;
                     };
+                } | {
+                    kind: "agent";
+                    source: {
+                        harnessRelease: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        agentSnapshot: {
+                            id: string;
+                            contentHash: string;
+                        };
+                        sourcePackageHash: string;
+                    };
+                    model: {
+                        modelId: string;
+                        configurationHash: string;
+                    };
                 } | {
                     kind: "fixture";
                     configurationHash: string;
@@ -2376,6 +2775,10 @@ export declare class OpenPondExperimentsClient {
                     packageHash: string;
                     runtimeTargetHash: string;
                     metricPolicyHash: string;
+                    compatibility?: {
+                        protocol: "openpond.evaluation-execution.v1";
+                        targetKind: "agent" | "model" | "fixture" | "harness";
+                    } | undefined;
                 } | undefined;
                 lineage?: {
                     definition: {
@@ -2449,8 +2852,7 @@ export declare class OpenPondExperimentsClient {
         };
         comparison: import("@openpond/evals/experiments").ExperimentComparison;
     }>;
-    private definition;
-    private executionDetails;
+    private details;
     private pass;
     private request;
 }