engineering-behavior-observatory 0.2.1

package/dist/src/index.d.ts

@@ -0,0 +1,71 @@
+export { main } from "./cli.js";
+export { assertUniqueArtifactIdentities, canonicalizeMetadata, digestBytes, digestMetadata, inspectRetainedArtifact, readVerifiedArtifact, SUPPORTED_ARTIFACT_SCHEMA_VERSIONS, validateArtifact, validateExportManifest, validateRunManifestEvidence, verifyDigest, writeArtifactAtomically, writeMetadataAtomically, } from "./artifacts.js";
+export type { ArtifactIdentity, ArtifactValidationError } from "./artifacts.js";
+export { assertAdmittedTaskPackets, assertArchiveMeasurements, assertControlledPerturbationDigest, assertDeclaredOrder, assertNoSelectedSymlinks, assertResolvedExperimentConfigurationDigests, declaredMatrixCells, isSafeArtifactRelativePath, MAX_CONFIGURATION_BYTES, resolveBundleConfiguration, resolveBundleArtifact, resolveBundleArtifactDigest, resolveTaskArchive, readTaskArchive, validateTaskArchive, } from "./contracts.js";
+export { admitTaskPacket, assertTaskPacketFreezeRecord, assertTaskPacketAdmitted, defaultFreezeLocator, formatErrors, freezeTaskPacket, inspectTaskPacket, modelVisibleTaskPacket, statusTaskPacket, MAX_TASK_PACKET_METADATA_BYTES, TASK_PACKET_FREEZE_SCHEMA_VERSION, } from "./task-packets.js";
+export type { TaskPacket, TaskPacketComponent, TaskPacketComponents, TaskPacketFreezeRecord, TaskPacketInspection, TaskPacketStatus } from "./task-packets.js";
+export type { ArtifactReference, AssessmentMode, ArchiveEntry, ArchiveLimits, ArchiveMeasurements, ControlledPerturbationDeclaration, DeclaredOrder, DeclaredMatrixCell, Digest, ExperimentConfiguration, ExperimentOrdering, ReferenceSolutionDeclaration, ResolvedTaskPacket, TaskCondition, TaskConditionSet, TaskArchiveEntry, } from "./contracts.js";
+export { assertValidRunQueue, compileRunQueue, expandMatrixCells, inspectRunQueue, LocalRunQueue, MAX_RUN_QUEUE_BYTES, MAX_RUN_QUEUE_ENTRIES, readRunQueue, validateRunQueue, writeRunQueue, } from "./scheduler.js";
+export type { CompileRunQueueOptions, FrozenTaskIdentity, FrozenTaskInput, PermutationAlgorithm, QueueOrderingStrategy, RunQueue, RunQueueEntry, RunQueueInspection, RunQueueMatrix, ValidateRunQueueOptions, } from "./scheduler.js";
+export { digestWorkspace, digestWorkspaceTree, executeVerifier, serializeVerifierResult, writeVerifierResult, } from "./verifiers.js";
+export type { CompleteVerifierResult, DiagnosticOrigin, DiagnosticReference, ExecuteVerifierOptions, VerifierAssertion, VerifierAssertionStatus, VerifierErrorResult, VerifierExecutionReference, VerifierFailedResult, VerifierNotRunResult, VerifierNotRunAssertion, VerifierPassedAssertion, VerifierPassedResult, VerifierResult, VerifierRunResult, VerifierWorkspace, } from "./verifiers.js";
+export { cleanupWorkspace, materializeWorkspace } from "./workspaces.js";
+export type { WorkspaceMaterialization, WorkspaceMaterializationOptions, WorkspaceSetupContext, WorkspaceSetupStep, } from "./workspaces.js";
+export { createAttemptIdentity, createRunIdentity, executeRunAttempt, InvalidLifecycleTransitionError, LifecycleController, retryAttempt, readAttemptRecord, writeAttemptRecord, } from "./lifecycle.js";
+export type { AttemptClassification, AttemptClassificationKind, AttemptIdentity, AttemptRecord, EvidenceSink, FailureClass, HarnessExecutionContext, HarnessExecutionResult, HarnessExecutor, HarnessShutdownResult, LifecycleSnapshot, LifecycleState, LifecycleTransition, RunAttemptOptions, RunAttemptResult, RunIdentity, StopReason, TerminalRecord, TerminalState, VerifierExecutionContext, VerifierExecutionResult, VerifierExecutor, VerifierShutdownResult, WorkspaceCleanupContext, WorkspaceCoordinator, WorkspaceExecutionResult, WorkspaceShutdownResult, WorkspaceSetupContext as LifecycleWorkspaceSetupContext, } from "./lifecycle.js";
+export { createRunBundleAssembler, qualifyRunBundle, RunBundleAssembler, } from "./run-bundles.js";
+export type { AgentSdkQualificationEvidence, CaptureQualificationDimension, CaptureQualificationOptions, CaptureQualificationReason, CaptureQualificationReasonCode, CaptureQualificationReport, CaptureQualificationStatus, CapturedWorkspaceOutcome, CaptureMissingEvidence, CaptureWorkspaceOutcomeOptions, RegisterRunBundleArtifact, RunBundleConfiguration, RunBundleDefinition, RunBundleEvidenceDescriptor, RunBundleRun, RunBundleRuntimeComponent, RunManifest, } from "./run-bundles.js";
+export { createPortableRunBundleExport, readPortableRunBundleExport } from "./exports.js";
+export type { CreatePortableRunBundleExportOptions, PortableExportArtifact, PortableExportManifest, PortableExportPolicy, } from "./exports.js";
+export { buildCorpusIndex, packPortableExport, queryCorpusIndex, readCorpusIndex, unpackPortableExport, validateCorpusIndex, writeCorpusIndex, } from "./corpus.js";
+export type { CorpusIndexEntry, CorpusIndexQuery, CorpusIndexValidationIssue, CorpusSourceIssue, } from "./corpus.js";
+export { BoundedDiagnosticCapture, JsonlEvidenceWriter, openJsonlEvidenceWriter, ProtocolEvidenceRecorder, ProtocolProcess, runProtocolProcess, spawnProtocolProcess, } from "./process-protocol.js";
+export { loadAtlas, queryAtlas, serveAtlas, shareAtlas, writeAtlas } from "./atlas.js";
+export type { AtlasCase, AtlasFilters, AtlasRequest, AtlasSource, AtlasView } from "./atlas.js";
+export { captureCodexAppServer, CODEX_ADAPTER_VERSION, CODEX_APP_SERVER_CAPABILITIES, CODEX_APP_SERVER_VERSION, CODEX_DEFAULT_SHUTDOWN_GRACE_MS, CODEX_HARNESS, createCodexHarnessAdapter, describeAndValidateCodexDataset, normalizeCodexCapture, } from "./codex.js";
+export type { CodexAppServerCapture, CodexAppServerCaptureRequest, CodexAppServerConfiguration, CodexApprovalPolicy, CodexCaptureGap, CodexOtlpRecord, CodexReasoningEffort, CodexSandbox, CodexTelemetryEvidence, CodexTelemetrySignal, } from "./codex.js";
+export { captureCodexAppServerRun, CODEX_CONFIG_SCHEMA_VERSION, CODEX_CONTRACT_DIGEST, resolveCodexConfigurationRecord, runCodexQueueEntry, } from "./codex-run.js";
+export type { CaptureCodexAppServerRunOptions, CaptureCodexAppServerRunResult, CaptureCodexAppServerVerifier, CodexCaptureProfileConfiguration, CodexConfigurationKind, CodexConfigurationRecord, CodexHarnessConfiguration, CodexModelConfiguration, CodexNativeLimitsConfiguration, CodexNativeToolPolicyConfiguration, CodexRunSummary, RunCodexQueueEntryOptions, } from "./codex-run.js";
+export { createDeepSeekHarnessAdapter, createDeepSeekRuntimeComposition, DEEPSEEK_ADAPTER_ID, DEEPSEEK_CAPABILITY_PROFILE, DEEPSEEK_HARNESS_ID, DEEPSEEK_SDK_VERSION, deepSeekCapabilities, deepSeekCompositionDigest, DeepSeekNativeCapture, executeDeepSeekHarness, normalizeDeepSeekCapture, qualifiedDeepSeekCapture, } from "./deepseek-adapter.js";
+export type { DeepSeekAdapterRequest, DeepSeekCapabilities, DeepSeekCaptureReport, DeepSeekFileReference, DeepSeekHarnessConfiguration, DeepSeekNativeObservation, DeepSeekRuntimeComposition, DeepSeekRuntimeCompositionInput, } from "./deepseek-adapter.js";
+export { buildClaudeAgentSdkEnvironment, executeClaudeAgentSdk, openClaudeAgentSdkHookCapture, openClaudeAgentSdkStreamCapture, probeClaudeAgentSdkCapabilities, } from "./agent-sdk.js";
+export type { ClaudeAgentSdkQuery, ClaudeAgentSdkQueryHandle, ClaudeAgentSdkEnvironment, ClaudeAgentSdkCaptureWarnings, ClaudeAgentSdkCaptureReport, ClaudeAgentSdkCapabilities, ClaudeAgentSdkConfiguration, ClaudeAgentSdkEvidenceSink, ClaudeAgentSdkHookCapture, ClaudeAgentSdkHookRecord, ClaudeAgentSdkLifecycleEvent, ClaudeAgentSdkMessageDiagnostic, ClaudeAgentSdkMessageRecord, ClaudeAgentSdkStreamCapture, ClaudeAgentSdkAttemptEvidence, ClaudeAgentSdkTelemetryConfiguration, ClaudeAgentSdkTelemetryCorrelation, ClaudeAgentSdkTelemetryEvidence, ClaudeAgentSdkTelemetryReceipt, ClaudeAgentSdkTelemetrySignal, } from "./agent-sdk.js";
+export { captureClaudeAgentSdkRun } from "./agent-sdk-run.js";
+export type { CaptureClaudeAgentSdkRunOptions, CaptureClaudeAgentSdkRunResult, CaptureClaudeAgentSdkVerifier, } from "./agent-sdk-run.js";
+export { AdapterRegistry, assertAdapterContract, UNIFORM_EVENT_FAMILIES, validateUniformEvents, } from "./uniform-events.js";
+export type { AdapterCapability, AdapterCapabilityProfile, CapturedNativeRecord, ContentReference, EvidenceUnavailable, EvidenceValue, HarnessAdapter, NativeCaptureAdapter, NativeEvidenceReference, NativeEvidenceResolution, NativeEvidenceResolver, NativeOrderEvidence, NormalizationInput, NormalizationResult, QualifiedNativeCapture, UniformAttributeScalar, UniformAttributeValue, UniformEvent, UniformEventFamily, UniformEventNormalizationAdapter, UnmappedNativeRecord, } from "./uniform-events.js";
+export { assessComparisonEligibility, assessLegacyComparisonEligibility, createCapturedNativeEvidenceResolver, describeNormalizedDataset, validateNormalizedCorpus, validateNormalizedDataset, } from "./normalization-integrity.js";
+export { CLAUDE_AGENT_SDK_NORMALIZATION_ADAPTER_VERSION, createAgentSdkStructuralObservationSet, createStructuralObservationSet, STRUCTURAL_EXTRACTOR_REGISTRY, STRUCTURAL_EXTRACTOR_VERSION, validateStructuralObservationSet, } from "./structural-observations.js";
+export type { StructuralObservation, StructuralObservationSet, } from "./structural-observations.js";
+export { createAgentSdkBehaviorEvidence, DEFAULT_BEHAVIOR_VOCABULARY, isConfirmedBehaviorAssertion, validateAgentSdkBehaviorAssertion, validateBehaviorAssertion, validateBehaviorReview, } from "./behavior-assertions.js";
+export type { AgentSdkBehaviorEvidence, BehaviorAssertion, BehaviorCitation, BehaviorReview, BehaviorVocabulary, ResolvedBehaviorCitation, } from "./behavior-assertions.js";
+export { assertCalibrationDestination, effectiveReviewOutcome, isDisputedReviewOutcome, importReviewDecision, selectReviewSample, summarizeCalibration, revalidateReviewSample, terminalVerifierOutcome, validateReviewSample, validateReviewHistory, writeReviewPacket, } from "./human-calibration.js";
+export type { CalibrationSummary, ReviewCandidate, ReviewDecision, ReviewFilters, ReviewHistory, ReviewSample, ReviewSampleCriteria, ReviewSourceSet, } from "./human-calibration.js";
+export { aggregateEvaluation } from "./aggregation.js";
+export type { AggregateMeasurement, AggregateMetric, AggregationDimension, AggregationInput, AggregationReport, AggregationRequest, AttemptSelectionPolicy, } from "./aggregation.js";
+export { CLAUDE_SEMANTIC_JUDGE_BACKEND_ID, packageSemanticJudgeInput, parseSemanticJudgeResponse, runAgentSdkSemanticJudge, runClaudeAgentSdkSemanticJudge, SEMANTIC_JUDGE_PROMPT_VERSION, } from "./semantic-judge.js";
+export type { RunAgentSdkSemanticJudgeOptions, SemanticJudgeBackend, SemanticJudgeBackendResult, SemanticJudgeEvidenceItem, SemanticJudgeInput, SemanticJudgeRequest, SemanticJudgmentRecord, SemanticJudgeUsage, } from "./semantic-judge.js";
+export type { AdapterCoverageReport, ComparisonCandidate, ComparisonCapability, ComparisonReason, ComparisonReport, ComparisonRequest, LegacyComparisonReport, LegacyComparisonRequest, NormalizedDataset, NormalizedDatasetInput, NormalizedContentReference, NormalizedNativeRecord, } from "./normalization-integrity.js";
+export { captureOpenHandsAgentServer, createOpenHandsHarnessAdapter, normalizeOpenHandsCapture, OPENHANDS_AGENT_SERVER_CAPABILITIES, OPENHANDS_AGENT_SERVER_VERSION, OPENHANDS_DEFAULT_MAX_RESPONSE_BYTES, OPENHANDS_MAX_RESPONSE_BYTES, OPENHANDS_DEFAULT_MAX_CAPTURE_BYTES, OPENHANDS_MAX_CAPTURE_BYTES, OPENHANDS_TYPESCRIPT_CLIENT_VERSION, } from "./openhands.js";
+export type { OpenHandsCapture, OpenHandsCaptureRequest, OpenHandsNativeRecord, OpenHandsWebSocket, } from "./openhands.js";
+export { captureOpenHandsAgentServerRun } from "./openhands-run.js";
+export type { CaptureOpenHandsAgentServerRunOptions, CaptureOpenHandsAgentServerRunResult, CaptureOpenHandsAgentServerVerifier, OpenHandsAgentServerRunConfiguration, } from "./openhands-run.js";
+export { CLAUDE_AGENT_SDK_HARNESS, CLAUDE_AGENT_SDK_NORMALIZATION_ADAPTER_ID, claudeAgentSdkNormalizationAdapter, createAgentSdkNativeEvidenceResolver, normalizeClaudeAgentSdkRunBundle, readQualifiedClaudeAgentSdkCapture, } from "./agent-sdk-normalizer.js";
+export type { AgentSdkNativeRecord, AgentSdkNormalizationResult, } from "./agent-sdk-normalizer.js";
+export { AGENT_SDK_CONFIG_SCHEMA_VERSION, resolveAgentSdkConfigurationRecord, runAgentSdkQueueEntry, } from "./agent-sdk-runner.js";
+export { assertPiNativeRecord, capturePiSdkRun, createPiPassiveObserver, describeAndValidatePiDataset, filterPiToolEnvironment, normalizePiCapture, piCapabilityProfile, piNativeType, qualifyRetainedPiCapture, resolvePiConfigurationRecord, runPiQueueEntry, PI_ADAPTER_VERSION, PI_CONFIG_SCHEMA_VERSION, PI_HARNESS, PINNED_PI_SDK_VERSION, } from "./pi.js";
+export type { CapturePiSdkRunOptions, CapturePiSdkRunResult, PiCaptureProfileConfiguration, PiConfigurationKind, PiConfigurationRecord, PiHarnessConfiguration, PiModelConfiguration, PiNativeLimitsConfiguration, PiNativeRecord, PiNativeToolPolicyConfiguration, PiRunSummary, PiSession, PiSessionFactory, PiSessionFactoryInput, RunPiQueueEntryOptions, } from "./pi.js";
+export type { AgentSdkCaptureProfileConfiguration, AgentSdkConfigurationKind, AgentSdkConfigurationRecord, AgentSdkHarnessConfiguration, AgentSdkModelConfiguration, AgentSdkNativeLimitsConfiguration, AgentSdkNativeToolPolicyConfiguration, AgentSdkRunSummary, RunAgentSdkQueueEntryOptions, } from "./agent-sdk-runner.js";
+export type { BoundedDiagnostic, JsonlEvidenceWriterOptions, ProcessTermination, ProtocolCapability, ProtocolCompletionEvidence, ProtocolIdentity, ProtocolObservation, ProtocolObservationInput, ProtocolObservationKind, ProtocolProcessOptions, ProtocolProcessResult, } from "./process-protocol.js";
+export { assertCursorWorkspaceIsolation, captureCursorSdkRun, createCursorSdkBehaviorEvidence, CURSOR_SDK_ADAPTER_ID, CURSOR_SDK_ADAPTER_VERSION, CURSOR_SDK_CAPABILITIES, CURSOR_SDK_DEFAULT_SHUTDOWN_GRACE_MS, CURSOR_SDK_HARNESS, CURSOR_SDK_VERSION, normalizeCursorSdkCapture, } from "./cursor-sdk.js";
+export type { CaptureCursorSdkRunOptions, CaptureCursorSdkRunResult, CaptureCursorSdkVerifier, CursorNativeRecord, CursorSdkAgentFactory, CursorSdkBehaviorEvidence, CursorSdkCaptureConfiguration, CursorSdkCaptureReport, CursorSdkToolPolicy, } from "./cursor-sdk.js";
+export { CURSOR_SDK_CONFIG_SCHEMA_VERSION, resolveCursorSdkConfigurationRecord, runCursorSdkQueueEntry, } from "./cursor-sdk-runner.js";
+export type { CursorSdkCaptureProfileConfiguration, CursorSdkConfigurationKind, CursorSdkConfigurationRecord, CursorSdkHarnessConfiguration, CursorSdkModelConfiguration, CursorSdkModelLister, CursorSdkNativeLimitsConfiguration, CursorSdkNativeToolPolicyConfiguration, CursorSdkRunSummary, RunCursorSdkQueueEntryOptions, } from "./cursor-sdk-runner.js";
+export { createRetainedBehaviorEvidence } from "./retained-evidence.js";
+export type { RetainedBehaviorEvidence } from "./retained-evidence.js";
+export { createRetainedStructuralObservationSet } from "./structural-observations.js";
+export { validateRetainedBehaviorAssertion } from "./behavior-assertions.js";
+export { runAgentSdkSemanticJudge as runRetainedSemanticJudge } from "./semantic-judge.js";
+export { runCodexSemanticJudge } from "./codex-judge.js";
+export type { BehaviorAggregate } from "./aggregation.js";
+export { selectAggregationAttempts } from "./aggregation.js";

package/dist/src/index.js

@@ -0,0 +1,38 @@
+export { main } from "./cli.js";
+export { assertUniqueArtifactIdentities, canonicalizeMetadata, digestBytes, digestMetadata, inspectRetainedArtifact, readVerifiedArtifact, SUPPORTED_ARTIFACT_SCHEMA_VERSIONS, validateArtifact, validateExportManifest, validateRunManifestEvidence, verifyDigest, writeArtifactAtomically, writeMetadataAtomically, } from "./artifacts.js";
+export { assertAdmittedTaskPackets, assertArchiveMeasurements, assertControlledPerturbationDigest, assertDeclaredOrder, assertNoSelectedSymlinks, assertResolvedExperimentConfigurationDigests, declaredMatrixCells, isSafeArtifactRelativePath, MAX_CONFIGURATION_BYTES, resolveBundleConfiguration, resolveBundleArtifact, resolveBundleArtifactDigest, resolveTaskArchive, readTaskArchive, validateTaskArchive, } from "./contracts.js";
+export { admitTaskPacket, assertTaskPacketFreezeRecord, assertTaskPacketAdmitted, defaultFreezeLocator, formatErrors, freezeTaskPacket, inspectTaskPacket, modelVisibleTaskPacket, statusTaskPacket, MAX_TASK_PACKET_METADATA_BYTES, TASK_PACKET_FREEZE_SCHEMA_VERSION, } from "./task-packets.js";
+export { assertValidRunQueue, compileRunQueue, expandMatrixCells, inspectRunQueue, LocalRunQueue, MAX_RUN_QUEUE_BYTES, MAX_RUN_QUEUE_ENTRIES, readRunQueue, validateRunQueue, writeRunQueue, } from "./scheduler.js";
+export { digestWorkspace, digestWorkspaceTree, executeVerifier, serializeVerifierResult, writeVerifierResult, } from "./verifiers.js";
+export { cleanupWorkspace, materializeWorkspace } from "./workspaces.js";
+export { createAttemptIdentity, createRunIdentity, executeRunAttempt, InvalidLifecycleTransitionError, LifecycleController, retryAttempt, readAttemptRecord, writeAttemptRecord, } from "./lifecycle.js";
+export { createRunBundleAssembler, qualifyRunBundle, RunBundleAssembler, } from "./run-bundles.js";
+export { createPortableRunBundleExport, readPortableRunBundleExport } from "./exports.js";
+export { buildCorpusIndex, packPortableExport, queryCorpusIndex, readCorpusIndex, unpackPortableExport, validateCorpusIndex, writeCorpusIndex, } from "./corpus.js";
+export { BoundedDiagnosticCapture, JsonlEvidenceWriter, openJsonlEvidenceWriter, ProtocolEvidenceRecorder, ProtocolProcess, runProtocolProcess, spawnProtocolProcess, } from "./process-protocol.js";
+export { loadAtlas, queryAtlas, serveAtlas, shareAtlas, writeAtlas } from "./atlas.js";
+export { captureCodexAppServer, CODEX_ADAPTER_VERSION, CODEX_APP_SERVER_CAPABILITIES, CODEX_APP_SERVER_VERSION, CODEX_DEFAULT_SHUTDOWN_GRACE_MS, CODEX_HARNESS, createCodexHarnessAdapter, describeAndValidateCodexDataset, normalizeCodexCapture, } from "./codex.js";
+export { captureCodexAppServerRun, CODEX_CONFIG_SCHEMA_VERSION, CODEX_CONTRACT_DIGEST, resolveCodexConfigurationRecord, runCodexQueueEntry, } from "./codex-run.js";
+export { createDeepSeekHarnessAdapter, createDeepSeekRuntimeComposition, DEEPSEEK_ADAPTER_ID, DEEPSEEK_CAPABILITY_PROFILE, DEEPSEEK_HARNESS_ID, DEEPSEEK_SDK_VERSION, deepSeekCapabilities, deepSeekCompositionDigest, DeepSeekNativeCapture, executeDeepSeekHarness, normalizeDeepSeekCapture, qualifiedDeepSeekCapture, } from "./deepseek-adapter.js";
+export { buildClaudeAgentSdkEnvironment, executeClaudeAgentSdk, openClaudeAgentSdkHookCapture, openClaudeAgentSdkStreamCapture, probeClaudeAgentSdkCapabilities, } from "./agent-sdk.js";
+export { captureClaudeAgentSdkRun } from "./agent-sdk-run.js";
+export { AdapterRegistry, assertAdapterContract, UNIFORM_EVENT_FAMILIES, validateUniformEvents, } from "./uniform-events.js";
+export { assessComparisonEligibility, assessLegacyComparisonEligibility, createCapturedNativeEvidenceResolver, describeNormalizedDataset, validateNormalizedCorpus, validateNormalizedDataset, } from "./normalization-integrity.js";
+export { CLAUDE_AGENT_SDK_NORMALIZATION_ADAPTER_VERSION, createAgentSdkStructuralObservationSet, createStructuralObservationSet, STRUCTURAL_EXTRACTOR_REGISTRY, STRUCTURAL_EXTRACTOR_VERSION, validateStructuralObservationSet, } from "./structural-observations.js";
+export { createAgentSdkBehaviorEvidence, DEFAULT_BEHAVIOR_VOCABULARY, isConfirmedBehaviorAssertion, validateAgentSdkBehaviorAssertion, validateBehaviorAssertion, validateBehaviorReview, } from "./behavior-assertions.js";
+export { assertCalibrationDestination, effectiveReviewOutcome, isDisputedReviewOutcome, importReviewDecision, selectReviewSample, summarizeCalibration, revalidateReviewSample, terminalVerifierOutcome, validateReviewSample, validateReviewHistory, writeReviewPacket, } from "./human-calibration.js";
+export { aggregateEvaluation } from "./aggregation.js";
+export { CLAUDE_SEMANTIC_JUDGE_BACKEND_ID, packageSemanticJudgeInput, parseSemanticJudgeResponse, runAgentSdkSemanticJudge, runClaudeAgentSdkSemanticJudge, SEMANTIC_JUDGE_PROMPT_VERSION, } from "./semantic-judge.js";
+export { captureOpenHandsAgentServer, createOpenHandsHarnessAdapter, normalizeOpenHandsCapture, OPENHANDS_AGENT_SERVER_CAPABILITIES, OPENHANDS_AGENT_SERVER_VERSION, OPENHANDS_DEFAULT_MAX_RESPONSE_BYTES, OPENHANDS_MAX_RESPONSE_BYTES, OPENHANDS_DEFAULT_MAX_CAPTURE_BYTES, OPENHANDS_MAX_CAPTURE_BYTES, OPENHANDS_TYPESCRIPT_CLIENT_VERSION, } from "./openhands.js";
+export { captureOpenHandsAgentServerRun } from "./openhands-run.js";
+export { CLAUDE_AGENT_SDK_HARNESS, CLAUDE_AGENT_SDK_NORMALIZATION_ADAPTER_ID, claudeAgentSdkNormalizationAdapter, createAgentSdkNativeEvidenceResolver, normalizeClaudeAgentSdkRunBundle, readQualifiedClaudeAgentSdkCapture, } from "./agent-sdk-normalizer.js";
+export { AGENT_SDK_CONFIG_SCHEMA_VERSION, resolveAgentSdkConfigurationRecord, runAgentSdkQueueEntry, } from "./agent-sdk-runner.js";
+export { assertPiNativeRecord, capturePiSdkRun, createPiPassiveObserver, describeAndValidatePiDataset, filterPiToolEnvironment, normalizePiCapture, piCapabilityProfile, piNativeType, qualifyRetainedPiCapture, resolvePiConfigurationRecord, runPiQueueEntry, PI_ADAPTER_VERSION, PI_CONFIG_SCHEMA_VERSION, PI_HARNESS, PINNED_PI_SDK_VERSION, } from "./pi.js";
+export { assertCursorWorkspaceIsolation, captureCursorSdkRun, createCursorSdkBehaviorEvidence, CURSOR_SDK_ADAPTER_ID, CURSOR_SDK_ADAPTER_VERSION, CURSOR_SDK_CAPABILITIES, CURSOR_SDK_DEFAULT_SHUTDOWN_GRACE_MS, CURSOR_SDK_HARNESS, CURSOR_SDK_VERSION, normalizeCursorSdkCapture, } from "./cursor-sdk.js";
+export { CURSOR_SDK_CONFIG_SCHEMA_VERSION, resolveCursorSdkConfigurationRecord, runCursorSdkQueueEntry, } from "./cursor-sdk-runner.js";
+export { createRetainedBehaviorEvidence } from "./retained-evidence.js";
+export { createRetainedStructuralObservationSet } from "./structural-observations.js";
+export { validateRetainedBehaviorAssertion } from "./behavior-assertions.js";
+export { runAgentSdkSemanticJudge as runRetainedSemanticJudge } from "./semantic-judge.js";
+export { runCodexSemanticJudge } from "./codex-judge.js";
+export { selectAggregationAttempts } from "./aggregation.js";

package/dist/src/lifecycle.d.ts

@@ -0,0 +1,196 @@
+import type { AssessmentMode } from "./contracts.js";
+export type RunIdentity = {
+    id: string;
+    taskId: string;
+    modelId: string;
+    harnessId: string;
+};
+export type AttemptIdentity = {
+    id: string;
+    number: number;
+    retryOf?: string;
+};
+export type LifecycleState = "created" | "setup" | "running" | "verifying" | "cleaning" | "terminal";
+export type LifecycleTransition = {
+    from: LifecycleState;
+    to: LifecycleState;
+    at: string;
+};
+export type LifecycleSnapshot = {
+    state: LifecycleState;
+    createdAt: string;
+    startedAt: string;
+    endedAt?: string;
+    timestamps: Partial<Record<LifecycleState, string>>;
+    transitions: LifecycleTransition[];
+};
+export type FailureClass = "none" | "infrastructure" | "task";
+export type StopReason = "none" | "budget" | "policy";
+export type TerminalState = "completed" | "failed" | "stopped" | "interrupted";
+export type TerminalRecord = {
+    state: TerminalState;
+    failureClass: FailureClass;
+    stopReason: StopReason;
+    workspaceArtifactId?: string;
+};
+export type AttemptClassificationKind = "completed" | "task-failure" | "infrastructure-failure" | "verifier-error" | "budget-stop" | "policy-stop" | "interrupted" | "capture-incomplete";
+export type AttemptClassification = {
+    kind: AttemptClassificationKind;
+    terminal: TerminalRecord;
+    reason?: string;
+    source?: "runner" | "workspace" | "harness" | "verifier" | "cleanup" | "capture";
+    underlying?: AttemptClassificationKind;
+    underlyingSource?: AttemptClassification["source"];
+};
+export type WorkspaceExecutionResult = {
+    status: "ready" | "failed";
+    path?: string;
+    artifactId?: string;
+    digest?: string;
+    retained?: boolean;
+    error?: string;
+    shutdownResult?: WorkspaceShutdownResult;
+    evidence?: unknown;
+};
+export type WorkspaceShutdownResult = {
+    status: "completed" | "failed" | "timed-out";
+    error?: string;
+};
+export type WorkspaceSetupContext = {
+    run: RunIdentity;
+    attempt: AttemptIdentity;
+    signal: AbortSignal;
+    registerShutdown: (shutdown: () => void | Promise<void>) => void;
+};
+export type WorkspaceCleanupContext = {
+    run: RunIdentity;
+    attempt: AttemptIdentity;
+    signal: AbortSignal;
+    outcome: TerminalRecord;
+    workspace?: WorkspaceExecutionResult;
+};
+export type WorkspaceCoordinator = {
+    setup: (context: WorkspaceSetupContext) => WorkspaceExecutionResult | Promise<WorkspaceExecutionResult>;
+    cleanup?: (context: WorkspaceCleanupContext) => void | Promise<void>;
+};
+export type HarnessExecutionContext = {
+    run: RunIdentity;
+    attempt: AttemptIdentity;
+    signal: AbortSignal;
+    budgetMs?: number;
+    workspace?: WorkspaceExecutionResult;
+    registerShutdown: (shutdown: () => void | Promise<void>) => void;
+};
+export type HarnessExecutionResult = {
+    status: "completed" | "failed" | "stopped" | "interrupted";
+    failureClass?: Exclude<FailureClass, "none">;
+    stopReason?: Exclude<StopReason, "none">;
+    reason?: string;
+    error?: string;
+    captureError?: string;
+    completionEvidence?: unknown;
+    evidence?: unknown;
+    shutdown?: () => void | Promise<void>;
+    shutdownResult?: HarnessShutdownResult;
+};
+export type HarnessShutdownResult = {
+    status: "completed" | "failed" | "timed-out";
+    error?: string;
+};
+export type HarnessExecutor = (context: HarnessExecutionContext) => HarnessExecutionResult | Promise<HarnessExecutionResult>;
+export type VerifierExecutionContext = {
+    run: RunIdentity;
+    attempt: AttemptIdentity;
+    signal: AbortSignal;
+    workspace?: WorkspaceExecutionResult;
+    registerShutdown: (shutdown: () => void | Promise<void>) => void;
+};
+export type VerifierExecutionResult = {
+    status: "passed" | "failed" | "error" | "not-run";
+    error?: string;
+    shutdownResult?: VerifierShutdownResult;
+    evidence?: unknown;
+};
+export type VerifierShutdownResult = {
+    status: "completed" | "failed" | "timed-out";
+    error?: string;
+};
+export type VerifierExecutor = (context: VerifierExecutionContext) => VerifierExecutionResult | Promise<VerifierExecutionResult>;
+export type EvidenceSink = {
+    flush?: () => void | Promise<void>;
+};
+export type RunAttemptOptions = {
+    run: RunIdentity;
+    assessmentMode?: AssessmentMode;
+    attempt?: AttemptIdentity;
+    workspace: WorkspaceCoordinator;
+    harness: HarnessExecutor;
+    verifier?: VerifierExecutor;
+    maxWallClockMs?: number;
+    harnessBudgetMs?: number;
+    signal?: AbortSignal;
+    evidence?: EvidenceSink;
+    recordPath?: string;
+    now?: () => string;
+    shutdownGraceMs?: number;
+};
+export type AttemptRecord = {
+    schemaVersion: "ebo.attempt/v1";
+    assessmentMode: AssessmentMode;
+    run: RunIdentity;
+    attempt: AttemptIdentity;
+    lifecycle: LifecycleSnapshot;
+    terminal?: TerminalRecord;
+    classification?: AttemptClassification;
+    workspace?: WorkspaceExecutionResult;
+    harness?: HarnessExecutionResult;
+    verifier?: VerifierExecutionResult;
+    harnessTerminationConfirmed?: boolean;
+    verifierTerminationConfirmed?: boolean;
+    cleanup?: {
+        status: "completed" | "failed" | "timed-out";
+        error?: string;
+    };
+    capture?: {
+        status: "complete" | "incomplete";
+        error?: string;
+    };
+    persistence?: {
+        status: "complete" | "incomplete";
+        error?: string;
+    };
+    partial: boolean;
+};
+export type RunAttemptResult = {
+    run: RunIdentity;
+    attempt: AttemptIdentity;
+    record: AttemptRecord;
+    terminal: TerminalRecord;
+    classification: AttemptClassification;
+};
+export declare class InvalidLifecycleTransitionError extends Error {
+    constructor(from: LifecycleState, to: LifecycleState);
+}
+export declare function createRunIdentity(input: {
+    id?: string;
+    taskId: string;
+    modelId: string;
+    harnessId: string;
+}): RunIdentity;
+export declare function createAttemptIdentity(runId: string, number?: number, id?: string, retryOf?: string): AttemptIdentity;
+export declare function retryAttempt(previous: AttemptIdentity | Pick<AttemptRecord, "attempt">, id?: string): AttemptIdentity;
+export declare class LifecycleController {
+    private readonly now;
+    private readonly transitions;
+    private readonly timestamps;
+    private current;
+    private readonly createdAt;
+    private endedAt;
+    constructor(now?: () => string);
+    get state(): LifecycleState;
+    transition(to: LifecycleState, at?: string): LifecycleSnapshot;
+    snapshot(): LifecycleSnapshot;
+}
+export declare function executeRunAttempt(options: RunAttemptOptions): Promise<RunAttemptResult>;
+export declare function writeAttemptRecord(path: string, record: AttemptRecord): Promise<void>;
+export declare function readAttemptRecord(path: string): Promise<AttemptRecord>;