@agentv/core 4.34.0-next.1 → 4.35.0-next.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/evaluation/validation/index.cjs +121 -1
- package/dist/evaluation/validation/index.cjs.map +1 -1
- package/dist/evaluation/validation/index.js +121 -1
- package/dist/evaluation/validation/index.js.map +1 -1
- package/dist/index.cjs +23 -21
- package/dist/index.cjs.map +1 -1
- package/dist/index.d.cts +20 -22
- package/dist/index.d.ts +20 -22
- package/dist/index.js +23 -21
- package/dist/index.js.map +1 -1
- package/package.json +1 -1
package/dist/index.d.cts
CHANGED
|
@@ -6051,19 +6051,18 @@ declare function getWorkspacePoolRoot(): string;
|
|
|
6051
6051
|
* projects:
|
|
6052
6052
|
* - id: my-app
|
|
6053
6053
|
* name: My App
|
|
6054
|
+
* repo_url: https://github.com/example/my-app.git
|
|
6054
6055
|
* path: /home/user/projects/my-app
|
|
6055
|
-
*
|
|
6056
|
-
* url: ${{ PROJECT_REPO_URL }}
|
|
6057
|
-
* ref: ${{ PROJECT_REPO_REF:-main }}
|
|
6056
|
+
* ref: main
|
|
6058
6057
|
* results:
|
|
6059
|
-
*
|
|
6060
|
-
* repo: example/my-app-results
|
|
6058
|
+
* repo_url: git@github.com:example/my-app-results.git
|
|
6061
6059
|
* path: /srv/agentv/results/my-app
|
|
6062
|
-
*
|
|
6060
|
+
* sync:
|
|
6061
|
+
* auto_push: true
|
|
6063
6062
|
* added_at: "2026-03-20T10:00:00Z"
|
|
6064
6063
|
* last_opened_at: "2026-03-30T14:00:00Z"
|
|
6065
6064
|
*
|
|
6066
|
-
* The optional `
|
|
6065
|
+
* The optional `repoUrl` field enables remote sync via syncProjects():
|
|
6067
6066
|
* first run — git clone --depth 1 --filter=blob:none
|
|
6068
6067
|
* subsequent runs — git pull --ff-only
|
|
6069
6068
|
*
|
|
@@ -6080,24 +6079,23 @@ declare function getWorkspacePoolRoot(): string;
|
|
|
6080
6079
|
* - discoverProjects() is a one-shot filesystem utility for bulk
|
|
6081
6080
|
* registration; it does not run in the request path.
|
|
6082
6081
|
*/
|
|
6082
|
+
interface ProjectResultsSyncConfig {
|
|
6083
|
+
autoPush?: boolean;
|
|
6084
|
+
}
|
|
6083
6085
|
interface ProjectResultsConfig {
|
|
6084
|
-
|
|
6085
|
-
repo: string;
|
|
6086
|
+
repoUrl: string;
|
|
6086
6087
|
path?: string;
|
|
6087
|
-
|
|
6088
|
+
sync?: ProjectResultsSyncConfig;
|
|
6088
6089
|
branchPrefix?: string;
|
|
6089
6090
|
}
|
|
6090
|
-
interface ProjectSource {
|
|
6091
|
-
url: string;
|
|
6092
|
-
ref: string;
|
|
6093
|
-
}
|
|
6094
6091
|
interface ProjectEntry {
|
|
6095
6092
|
id: string;
|
|
6096
6093
|
name: string;
|
|
6094
|
+
repoUrl?: string;
|
|
6097
6095
|
path: string;
|
|
6096
|
+
ref?: string;
|
|
6098
6097
|
addedAt: string;
|
|
6099
6098
|
lastOpenedAt: string;
|
|
6100
|
-
source?: ProjectSource;
|
|
6101
6099
|
results?: ProjectResultsConfig;
|
|
6102
6100
|
}
|
|
6103
6101
|
interface ProjectRegistry {
|
|
@@ -6143,13 +6141,13 @@ declare function touchProject(projectId: string): void;
|
|
|
6143
6141
|
declare function discoverProjects(rootDir: string, maxDepth?: number): string[];
|
|
6144
6142
|
|
|
6145
6143
|
/**
|
|
6146
|
-
* Project sync — pulls remote
|
|
6144
|
+
* Project sync — pulls remote Git repos to the local path declared in the
|
|
6147
6145
|
* project registry before Dashboard/eval startup.
|
|
6148
6146
|
*
|
|
6149
6147
|
* Sync is oneshot only, triggered by the Dashboard UI "Sync" button or the
|
|
6150
6148
|
* `agentv project sync` CLI command. There is no daemon or continuous mode.
|
|
6151
6149
|
*
|
|
6152
|
-
* First run — git clone --depth 1 --filter=blob:none --branch <ref> <
|
|
6150
|
+
* First run — git clone --depth 1 --filter=blob:none --branch <ref> <repoUrl> <path>
|
|
6153
6151
|
* Subsequent — git pull --ff-only (when <path>/.git already exists)
|
|
6154
6152
|
*
|
|
6155
6153
|
* Usage:
|
|
@@ -6158,15 +6156,15 @@ declare function discoverProjects(rootDir: string, maxDepth?: number): string[];
|
|
|
6158
6156
|
*/
|
|
6159
6157
|
|
|
6160
6158
|
/**
|
|
6161
|
-
* Clone or pull a single project entry from its declared
|
|
6159
|
+
* Clone or pull a single project entry from its declared repo URL.
|
|
6162
6160
|
* - No .git present: shallow clone into entry.path.
|
|
6163
6161
|
* - .git present: git pull --ff-only to update in place.
|
|
6164
|
-
* Throws on git error or missing
|
|
6162
|
+
* Throws on git error or missing repoUrl/ref.
|
|
6165
6163
|
*/
|
|
6166
6164
|
declare function syncProject(entry: ProjectEntry): Promise<void>;
|
|
6167
6165
|
/**
|
|
6168
|
-
* Iterate project entries and sync any that have a
|
|
6169
|
-
* Entries without
|
|
6166
|
+
* Iterate project entries and sync any that have a repo URL declared.
|
|
6167
|
+
* Entries without repoUrl are skipped silently.
|
|
6170
6168
|
*/
|
|
6171
6169
|
declare function syncProjects(entries: ProjectEntry[]): Promise<void>;
|
|
6172
6170
|
|
|
@@ -6742,4 +6740,4 @@ type AgentKernel = {
|
|
|
6742
6740
|
};
|
|
6743
6741
|
declare function createAgentKernel(): AgentKernel;
|
|
6744
6742
|
|
|
6745
|
-
export { type AcquireWorkspaceOptions, type AgentKernel, type AgentVConfig$1 as AgentVConfig, type AgentVResolvedConfig, type AgentVConfig as AgentVTsConfig, type AgentVConfig$1 as AgentVYamlConfig, type AnthropicResolvedConfig, type ApiFormat, type ArgsMatchMode, type AssertContext, type AssertEntry, type AssertFn, type AssertResult, type AssertionEntry, type AssertionResult, type AssistantTestMessage, type AzureResolvedConfig, COMMON_TARGET_SETTINGS, type CacheConfig, type CheckedOutResultsRepoBranch, type ChildGraderResult, type ClaudeDiscoverOptions, type ClaudeResolvedConfig, type ClaudeSession, type CliResolvedConfig, CodeGrader, type CodeGraderConfig, type CodeGraderOptions, type CodexDiscoverOptions, type CodexSession, type CommandExecutor, type CompositeAggregatorConfig, CompositeGrader, type CompositeGraderConfig, type CompositeGraderOptions, type ConfidenceIntervalAggregation, type ContainsAllGraderConfig, type ContainsAnyGraderConfig, type ContainsGraderConfig, type Content, type ContentFile, type ContentImage, type ContentPreprocessorConfig, type ContentText, type ConversationAggregation, type ConversationMode, type ConversationTurn, type ConversationTurnInput, type CopilotCliResolvedConfig, type DiscoverOptions as CopilotDiscoverOptions, type CopilotLogResolvedConfig, type CopilotSdkResolvedConfig, type CopilotSession, type CopilotSessionMeta, CostGrader, type CostGraderConfig, type CostGraderOptions, type CreateContainerOptions, DEFAULT_CATEGORY, DEFAULT_EVAL_PATTERNS, DEFAULT_EXPLORATION_TOOLS, DEFAULT_GRADER_TEMPLATE, DEFAULT_THRESHOLD, type DependencyFailurePolicy, type DependencyResult, type DepsScanResult, DeterministicAssertionGrader, type DockerWorkspaceConfig, DockerWorkspaceProvider, type EndsWithGraderConfig, type EnsureSubagentsOptions, type EnsureSubagentsResult, type EnvLookup, type EqualsGraderConfig, type EvalAssertionInput, type EvalCase, type EvalConfig, type EvalGraderSource, type EvalMetadata, type EvalRunResult, type EvalSourceReference, type EvalSuiteResult, type EvalSummary, type EvalTargetRef, type EvalTest, type EvalTestInput, type EvalTestSource, type EvalsJsonCase, type EvalsJsonFile, type EvaluationCache, type EvaluationContext, type EvaluationResult, type EvaluationScore, type EvaluationVerdict, type ExecInContainerOptions, type ExecResult, type ExecutionDefaults, type ExecutionError, type ExecutionMetrics, ExecutionMetricsGrader, type ExecutionMetricsGraderConfig, type ExecutionMetricsGraderOptions, type ExecutionStatus, type FailOnError, type FailureStage, FieldAccuracyGrader, type FieldAccuracyGraderConfig, type FieldAccuracyGraderOptions, type FieldAggregationType, type FieldConfig, type FieldMatchType, type GeminiResolvedConfig, type GenerateRubricsOptions, type GitListedRun, type Grader, type GraderConfig, type GraderDispatchContext, type GraderFactory, type GraderFactoryFn, type GraderKind, GraderRegistry, type GraderResult, type IcontainsAllGraderConfig, type IcontainsAnyGraderConfig, type IcontainsGraderConfig, type InlineAssertEvaluatorConfig, type IsJsonGraderConfig, type JsonObject, type JsonPrimitive, type JsonValue, LatencyGrader, type LatencyGraderConfig, type LatencyGraderOptions, LlmGrader, type LlmGraderConfig, type LlmGraderOptions, type LlmGraderPromptAssembly, type LocalPathValidationError, type MeanAggregation, type Message, type MockResolvedConfig, NORMALIZED_REDACTION_LEVELS, NORMALIZED_TOOL_STATUSES, NORMALIZED_TRACE_EVENT_TYPES, NORMALIZED_TRACE_SOURCE_KINDS, NORMALIZED_TRAJECTORY_SCHEMA_VERSION, type NormalizedRawEvidence, NormalizedRawEvidenceWireSchema, type NormalizedRedactionLevel, type NormalizedRedactionState, NormalizedRedactionStateWireSchema, type NormalizedToolStatus, type NormalizedTraceBranch, NormalizedTraceBranchWireSchema, type NormalizedTraceError, NormalizedTraceErrorWireSchema, type NormalizedTraceEvent, type NormalizedTraceEventType, type NormalizedTraceEventWire, NormalizedTraceEventWireSchema, type NormalizedTraceMessage, NormalizedTraceMessageWireSchema, type NormalizedTraceModel, NormalizedTraceModelWireSchema, type NormalizedTraceSession, NormalizedTraceSessionWireSchema, type NormalizedTraceSource, type NormalizedTraceSourceKind, type NormalizedTraceSourceRef, NormalizedTraceSourceRefWireSchema, NormalizedTraceSourceWireSchema, type NormalizedTraceTool, NormalizedTraceToolWireSchema, type NormalizedTrajectory, type NormalizedTrajectoryWire, NormalizedTrajectoryWireSchema, OTEL_BACKEND_PRESETS, type OpenAIResolvedConfig, type OpenRouterResolvedConfig, type OtelBackendPreset, type OtelExportOptions, OtelStreamingObserver, OtelTraceExporter, OtlpJsonFileExporter, type OutputMessage, PASS_THRESHOLD, type ParsedCopilotSession, type PassAtKAggregation, type PiCliResolvedConfig, type PiCodingAgentResolvedConfig, type PoolSlot, type PreparedResultsRepoBranch, type ProgressEvent, type ProjectEntry, type ProjectRegistry, type ProjectSource, type PromptInputs, type PromptScriptConfig, type Provider, type ProviderFactoryFn, type ProviderKind, ProviderRegistry, type ProviderRequest, type ProviderResponse, type ProviderStreamCallbacks, type ProviderTokenUsage, RUBRIC_OPERATOR_VALUES, type RegexGraderConfig, type RepoCheckout, type RepoClone, type RepoConfig, type RepoDep, RepoManager, type RepoSource, type ResolvedTarget, type ResolvedWorkspaceTemplate, ResponseCache, type ResultsConfig, type ResultsRepoLocalPaths, type ResultsRepoStatus, type ResultsRepoSyncStatus, type RubricItem, type RubricOperator, type RubricsEvaluatorConfig, RunBudgetTracker, type RunEvalCaseOptions, type RunEvaluationOptions, type ScoreRange, type ScriptExecutionContext, SkillTriggerGrader, type SkillTriggerGraderConfig, type StartsWithGraderConfig, type SystemTestMessage, TEST_MESSAGE_ROLES, type TargetAccessConfig, type TargetDefinition, type TargetHooksConfig, TemplateNotDirectoryError, TemplateNotFoundError, type TestMessage, type TestMessageContent, type TestMessageRole, type TokenUsage, TokenUsageGrader, type TokenUsageGraderConfig, type TokenUsageGraderOptions, type ToolCall, type ToolTestMessage, type ToolTrajectoryExpectedItem, ToolTrajectoryGrader, type ToolTrajectoryGraderConfig, type ToolTrajectoryGraderOptions, type TraceComputeResult, type TraceSummary, type TranscriptEntry, type TranscriptJsonLine, TranscriptProvider, type TranscriptReplayEntry, type TranscriptSource, type TranspileResult, type TrialAggregation, type TrialResult, type TrialStrategy, type TrialsConfig, type TsEvalResult, type TurnFailurePolicy, type UserTestMessage, type VSCodeResolvedConfig, type WorkspaceConfig, WorkspaceCreationError, type WorkspaceEnvConfig, type WorkspaceHookConfig, type WorkspaceHooksConfig, WorkspacePoolManager, type WorkspaceScriptConfig, addProject, assembleLlmGraderPrompt, avgToolDurationMs, buildDirectoryChain, buildOutputSchema, buildPromptInputs, buildRubricOutputSchema, buildScoreRangeOutputSchema, buildSearchRoots, calculateRubricScore, captureFileChanges, checkoutResultsRepoBranch, clampScore, cleanupEvalWorkspaces, cleanupWorkspace, commitAndPushResultsBranch, computeTraceSummary, computeTraceSummaryFromTrajectory, computeWorkspaceFingerprint, consumeClaudeLogEntries, consumeCodexLogEntries, consumeCopilotCliLogEntries, consumeCopilotSdkLogEntries, consumePiLogEntries, createAgentKernel, createBuiltinProviderRegistry, createBuiltinRegistry, createDraftResultsPr, createProvider, createTempWorkspace, deepEqual, defineConfig, deriveCategory, deriveProjectId, detectFormat, directPushResults, directorySizeBytes, discoverAssertions, discoverClaudeSessions, discoverCodexSessions, discoverCopilotSessions, discoverGraders, discoverProjects, discoverProviders, ensureResultsRepoClone, ensureVSCodeSubagents, evaluate, executeScript, executeWorkspaceScript, explorationRatio, extractCacheConfig, extractFailOnError, extractImageBlocks, extractJsonBlob, extractLastAssistantContent, extractTargetFromSuite, extractTargetRefsFromSuite, extractTargetsFromSuite, extractTargetsFromTestCase, extractThreshold, extractTrialsConfig, extractWorkersFromSuite, fileExists, findGitRoot, formatToolCalls, freeformEvaluationSchema, fromNormalizedTrajectoryWire, generateRubrics, getAgentvConfigDir, getAgentvDataDir, getAgentvHome, getOutputFilenames, getProject, getProjectForPath, getProjectsRegistryPath, getResultsRepoLocalPaths, getResultsRepoStatus, getResultsRepoSyncStatus, getSelectedTrajectoryEvents, getSubagentsRoot, getTextContent, getTraceStateRoot, getWorkspacePath, getWorkspacePoolRoot, getWorkspacesRoot, groupTranscriptJsonLines, initializeBaseline, isAgentSkillsFormat, isContent, isContentArray, isGraderKind, isJsonObject, isJsonValue, isNonEmptyString, isTestMessage, isTestMessageRole, killAllTrackedChildren, listGitRuns, listTargetNames, loadConfig, loadEvalCaseById, loadEvalCases, loadEvalSuite, loadProjectRegistry, loadTestById, loadTestSuite, loadTests, loadTsConfig, loadTsEvalFile, materializeGitRun, mergeExecutionMetrics, negateScore, normalizeLineEndings, normalizeResultsConfig, parseAgentSkillsEvals, parseClaudeSession, parseCodexSession, parseCopilotEvents, parseEnvOutput, parseJsonFromText, parseJsonSafe, parseYamlValue, prepareResultsRepoBranch, pushResultsRepoBranch, readJsonFile, readTargetDefinitions, readTestSuiteMetadata, readTextFile, readTranscriptFile, readTranscriptJsonl, removeProject, resolveAndCreateProvider, resolveDelegatedTargetDefinition, resolveFileReference, resolveResultsConfigForProject, resolveResultsRepoRunsDir, resolveResultsRepoUrl, resolveTargetDefinition, resolveWorkspaceTemplate, rubricEvaluationSchema, runBeforeSessionHook, runContainsAllAssertion, runContainsAnyAssertion, runContainsAssertion, runEndsWithAssertion, runEqualsAssertion, runEvalCase, runEvaluation, runIcontainsAllAssertion, runIcontainsAnyAssertion, runIcontainsAssertion, runIsJsonAssertion, runRegexAssertion, runStartsWithAssertion, saveProjectRegistry, scanRepoDeps, scoreRangeEvaluationSchema, scoreToVerdict, shouldEnableCache, shouldSkipCacheForTemperature, stageResultsArtifacts, subscribeToClaudeLogEntries, subscribeToCodexLogEntries, subscribeToCopilotCliLogEntries, subscribeToCopilotSdkLogEntries, subscribeToPiLogEntries, substituteVariables, syncProject, syncProjects, syncResultsRepo, syncResultsRepoForProject, toCamelCaseDeep, toNormalizedTrajectoryWire, toSnakeCaseDeep, toTranscriptJsonLines, tokensPerTool, touchProject, trackChild, trackedChildCount, transpileEvalYaml, transpileEvalYamlFile, trimBaselineResult };
|
|
6743
|
+
export { type AcquireWorkspaceOptions, type AgentKernel, type AgentVConfig$1 as AgentVConfig, type AgentVResolvedConfig, type AgentVConfig as AgentVTsConfig, type AgentVConfig$1 as AgentVYamlConfig, type AnthropicResolvedConfig, type ApiFormat, type ArgsMatchMode, type AssertContext, type AssertEntry, type AssertFn, type AssertResult, type AssertionEntry, type AssertionResult, type AssistantTestMessage, type AzureResolvedConfig, COMMON_TARGET_SETTINGS, type CacheConfig, type CheckedOutResultsRepoBranch, type ChildGraderResult, type ClaudeDiscoverOptions, type ClaudeResolvedConfig, type ClaudeSession, type CliResolvedConfig, CodeGrader, type CodeGraderConfig, type CodeGraderOptions, type CodexDiscoverOptions, type CodexSession, type CommandExecutor, type CompositeAggregatorConfig, CompositeGrader, type CompositeGraderConfig, type CompositeGraderOptions, type ConfidenceIntervalAggregation, type ContainsAllGraderConfig, type ContainsAnyGraderConfig, type ContainsGraderConfig, type Content, type ContentFile, type ContentImage, type ContentPreprocessorConfig, type ContentText, type ConversationAggregation, type ConversationMode, type ConversationTurn, type ConversationTurnInput, type CopilotCliResolvedConfig, type DiscoverOptions as CopilotDiscoverOptions, type CopilotLogResolvedConfig, type CopilotSdkResolvedConfig, type CopilotSession, type CopilotSessionMeta, CostGrader, type CostGraderConfig, type CostGraderOptions, type CreateContainerOptions, DEFAULT_CATEGORY, DEFAULT_EVAL_PATTERNS, DEFAULT_EXPLORATION_TOOLS, DEFAULT_GRADER_TEMPLATE, DEFAULT_THRESHOLD, type DependencyFailurePolicy, type DependencyResult, type DepsScanResult, DeterministicAssertionGrader, type DockerWorkspaceConfig, DockerWorkspaceProvider, type EndsWithGraderConfig, type EnsureSubagentsOptions, type EnsureSubagentsResult, type EnvLookup, type EqualsGraderConfig, type EvalAssertionInput, type EvalCase, type EvalConfig, type EvalGraderSource, type EvalMetadata, type EvalRunResult, type EvalSourceReference, type EvalSuiteResult, type EvalSummary, type EvalTargetRef, type EvalTest, type EvalTestInput, type EvalTestSource, type EvalsJsonCase, type EvalsJsonFile, type EvaluationCache, type EvaluationContext, type EvaluationResult, type EvaluationScore, type EvaluationVerdict, type ExecInContainerOptions, type ExecResult, type ExecutionDefaults, type ExecutionError, type ExecutionMetrics, ExecutionMetricsGrader, type ExecutionMetricsGraderConfig, type ExecutionMetricsGraderOptions, type ExecutionStatus, type FailOnError, type FailureStage, FieldAccuracyGrader, type FieldAccuracyGraderConfig, type FieldAccuracyGraderOptions, type FieldAggregationType, type FieldConfig, type FieldMatchType, type GeminiResolvedConfig, type GenerateRubricsOptions, type GitListedRun, type Grader, type GraderConfig, type GraderDispatchContext, type GraderFactory, type GraderFactoryFn, type GraderKind, GraderRegistry, type GraderResult, type IcontainsAllGraderConfig, type IcontainsAnyGraderConfig, type IcontainsGraderConfig, type InlineAssertEvaluatorConfig, type IsJsonGraderConfig, type JsonObject, type JsonPrimitive, type JsonValue, LatencyGrader, type LatencyGraderConfig, type LatencyGraderOptions, LlmGrader, type LlmGraderConfig, type LlmGraderOptions, type LlmGraderPromptAssembly, type LocalPathValidationError, type MeanAggregation, type Message, type MockResolvedConfig, NORMALIZED_REDACTION_LEVELS, NORMALIZED_TOOL_STATUSES, NORMALIZED_TRACE_EVENT_TYPES, NORMALIZED_TRACE_SOURCE_KINDS, NORMALIZED_TRAJECTORY_SCHEMA_VERSION, type NormalizedRawEvidence, NormalizedRawEvidenceWireSchema, type NormalizedRedactionLevel, type NormalizedRedactionState, NormalizedRedactionStateWireSchema, type NormalizedToolStatus, type NormalizedTraceBranch, NormalizedTraceBranchWireSchema, type NormalizedTraceError, NormalizedTraceErrorWireSchema, type NormalizedTraceEvent, type NormalizedTraceEventType, type NormalizedTraceEventWire, NormalizedTraceEventWireSchema, type NormalizedTraceMessage, NormalizedTraceMessageWireSchema, type NormalizedTraceModel, NormalizedTraceModelWireSchema, type NormalizedTraceSession, NormalizedTraceSessionWireSchema, type NormalizedTraceSource, type NormalizedTraceSourceKind, type NormalizedTraceSourceRef, NormalizedTraceSourceRefWireSchema, NormalizedTraceSourceWireSchema, type NormalizedTraceTool, NormalizedTraceToolWireSchema, type NormalizedTrajectory, type NormalizedTrajectoryWire, NormalizedTrajectoryWireSchema, OTEL_BACKEND_PRESETS, type OpenAIResolvedConfig, type OpenRouterResolvedConfig, type OtelBackendPreset, type OtelExportOptions, OtelStreamingObserver, OtelTraceExporter, OtlpJsonFileExporter, type OutputMessage, PASS_THRESHOLD, type ParsedCopilotSession, type PassAtKAggregation, type PiCliResolvedConfig, type PiCodingAgentResolvedConfig, type PoolSlot, type PreparedResultsRepoBranch, type ProgressEvent, type ProjectEntry, type ProjectRegistry, type PromptInputs, type PromptScriptConfig, type Provider, type ProviderFactoryFn, type ProviderKind, ProviderRegistry, type ProviderRequest, type ProviderResponse, type ProviderStreamCallbacks, type ProviderTokenUsage, RUBRIC_OPERATOR_VALUES, type RegexGraderConfig, type RepoCheckout, type RepoClone, type RepoConfig, type RepoDep, RepoManager, type RepoSource, type ResolvedTarget, type ResolvedWorkspaceTemplate, ResponseCache, type ResultsConfig, type ResultsRepoLocalPaths, type ResultsRepoStatus, type ResultsRepoSyncStatus, type RubricItem, type RubricOperator, type RubricsEvaluatorConfig, RunBudgetTracker, type RunEvalCaseOptions, type RunEvaluationOptions, type ScoreRange, type ScriptExecutionContext, SkillTriggerGrader, type SkillTriggerGraderConfig, type StartsWithGraderConfig, type SystemTestMessage, TEST_MESSAGE_ROLES, type TargetAccessConfig, type TargetDefinition, type TargetHooksConfig, TemplateNotDirectoryError, TemplateNotFoundError, type TestMessage, type TestMessageContent, type TestMessageRole, type TokenUsage, TokenUsageGrader, type TokenUsageGraderConfig, type TokenUsageGraderOptions, type ToolCall, type ToolTestMessage, type ToolTrajectoryExpectedItem, ToolTrajectoryGrader, type ToolTrajectoryGraderConfig, type ToolTrajectoryGraderOptions, type TraceComputeResult, type TraceSummary, type TranscriptEntry, type TranscriptJsonLine, TranscriptProvider, type TranscriptReplayEntry, type TranscriptSource, type TranspileResult, type TrialAggregation, type TrialResult, type TrialStrategy, type TrialsConfig, type TsEvalResult, type TurnFailurePolicy, type UserTestMessage, type VSCodeResolvedConfig, type WorkspaceConfig, WorkspaceCreationError, type WorkspaceEnvConfig, type WorkspaceHookConfig, type WorkspaceHooksConfig, WorkspacePoolManager, type WorkspaceScriptConfig, addProject, assembleLlmGraderPrompt, avgToolDurationMs, buildDirectoryChain, buildOutputSchema, buildPromptInputs, buildRubricOutputSchema, buildScoreRangeOutputSchema, buildSearchRoots, calculateRubricScore, captureFileChanges, checkoutResultsRepoBranch, clampScore, cleanupEvalWorkspaces, cleanupWorkspace, commitAndPushResultsBranch, computeTraceSummary, computeTraceSummaryFromTrajectory, computeWorkspaceFingerprint, consumeClaudeLogEntries, consumeCodexLogEntries, consumeCopilotCliLogEntries, consumeCopilotSdkLogEntries, consumePiLogEntries, createAgentKernel, createBuiltinProviderRegistry, createBuiltinRegistry, createDraftResultsPr, createProvider, createTempWorkspace, deepEqual, defineConfig, deriveCategory, deriveProjectId, detectFormat, directPushResults, directorySizeBytes, discoverAssertions, discoverClaudeSessions, discoverCodexSessions, discoverCopilotSessions, discoverGraders, discoverProjects, discoverProviders, ensureResultsRepoClone, ensureVSCodeSubagents, evaluate, executeScript, executeWorkspaceScript, explorationRatio, extractCacheConfig, extractFailOnError, extractImageBlocks, extractJsonBlob, extractLastAssistantContent, extractTargetFromSuite, extractTargetRefsFromSuite, extractTargetsFromSuite, extractTargetsFromTestCase, extractThreshold, extractTrialsConfig, extractWorkersFromSuite, fileExists, findGitRoot, formatToolCalls, freeformEvaluationSchema, fromNormalizedTrajectoryWire, generateRubrics, getAgentvConfigDir, getAgentvDataDir, getAgentvHome, getOutputFilenames, getProject, getProjectForPath, getProjectsRegistryPath, getResultsRepoLocalPaths, getResultsRepoStatus, getResultsRepoSyncStatus, getSelectedTrajectoryEvents, getSubagentsRoot, getTextContent, getTraceStateRoot, getWorkspacePath, getWorkspacePoolRoot, getWorkspacesRoot, groupTranscriptJsonLines, initializeBaseline, isAgentSkillsFormat, isContent, isContentArray, isGraderKind, isJsonObject, isJsonValue, isNonEmptyString, isTestMessage, isTestMessageRole, killAllTrackedChildren, listGitRuns, listTargetNames, loadConfig, loadEvalCaseById, loadEvalCases, loadEvalSuite, loadProjectRegistry, loadTestById, loadTestSuite, loadTests, loadTsConfig, loadTsEvalFile, materializeGitRun, mergeExecutionMetrics, negateScore, normalizeLineEndings, normalizeResultsConfig, parseAgentSkillsEvals, parseClaudeSession, parseCodexSession, parseCopilotEvents, parseEnvOutput, parseJsonFromText, parseJsonSafe, parseYamlValue, prepareResultsRepoBranch, pushResultsRepoBranch, readJsonFile, readTargetDefinitions, readTestSuiteMetadata, readTextFile, readTranscriptFile, readTranscriptJsonl, removeProject, resolveAndCreateProvider, resolveDelegatedTargetDefinition, resolveFileReference, resolveResultsConfigForProject, resolveResultsRepoRunsDir, resolveResultsRepoUrl, resolveTargetDefinition, resolveWorkspaceTemplate, rubricEvaluationSchema, runBeforeSessionHook, runContainsAllAssertion, runContainsAnyAssertion, runContainsAssertion, runEndsWithAssertion, runEqualsAssertion, runEvalCase, runEvaluation, runIcontainsAllAssertion, runIcontainsAnyAssertion, runIcontainsAssertion, runIsJsonAssertion, runRegexAssertion, runStartsWithAssertion, saveProjectRegistry, scanRepoDeps, scoreRangeEvaluationSchema, scoreToVerdict, shouldEnableCache, shouldSkipCacheForTemperature, stageResultsArtifacts, subscribeToClaudeLogEntries, subscribeToCodexLogEntries, subscribeToCopilotCliLogEntries, subscribeToCopilotSdkLogEntries, subscribeToPiLogEntries, substituteVariables, syncProject, syncProjects, syncResultsRepo, syncResultsRepoForProject, toCamelCaseDeep, toNormalizedTrajectoryWire, toSnakeCaseDeep, toTranscriptJsonLines, tokensPerTool, touchProject, trackChild, trackedChildCount, transpileEvalYaml, transpileEvalYamlFile, trimBaselineResult };
|
package/dist/index.d.ts
CHANGED
|
@@ -6051,19 +6051,18 @@ declare function getWorkspacePoolRoot(): string;
|
|
|
6051
6051
|
* projects:
|
|
6052
6052
|
* - id: my-app
|
|
6053
6053
|
* name: My App
|
|
6054
|
+
* repo_url: https://github.com/example/my-app.git
|
|
6054
6055
|
* path: /home/user/projects/my-app
|
|
6055
|
-
*
|
|
6056
|
-
* url: ${{ PROJECT_REPO_URL }}
|
|
6057
|
-
* ref: ${{ PROJECT_REPO_REF:-main }}
|
|
6056
|
+
* ref: main
|
|
6058
6057
|
* results:
|
|
6059
|
-
*
|
|
6060
|
-
* repo: example/my-app-results
|
|
6058
|
+
* repo_url: git@github.com:example/my-app-results.git
|
|
6061
6059
|
* path: /srv/agentv/results/my-app
|
|
6062
|
-
*
|
|
6060
|
+
* sync:
|
|
6061
|
+
* auto_push: true
|
|
6063
6062
|
* added_at: "2026-03-20T10:00:00Z"
|
|
6064
6063
|
* last_opened_at: "2026-03-30T14:00:00Z"
|
|
6065
6064
|
*
|
|
6066
|
-
* The optional `
|
|
6065
|
+
* The optional `repoUrl` field enables remote sync via syncProjects():
|
|
6067
6066
|
* first run — git clone --depth 1 --filter=blob:none
|
|
6068
6067
|
* subsequent runs — git pull --ff-only
|
|
6069
6068
|
*
|
|
@@ -6080,24 +6079,23 @@ declare function getWorkspacePoolRoot(): string;
|
|
|
6080
6079
|
* - discoverProjects() is a one-shot filesystem utility for bulk
|
|
6081
6080
|
* registration; it does not run in the request path.
|
|
6082
6081
|
*/
|
|
6082
|
+
interface ProjectResultsSyncConfig {
|
|
6083
|
+
autoPush?: boolean;
|
|
6084
|
+
}
|
|
6083
6085
|
interface ProjectResultsConfig {
|
|
6084
|
-
|
|
6085
|
-
repo: string;
|
|
6086
|
+
repoUrl: string;
|
|
6086
6087
|
path?: string;
|
|
6087
|
-
|
|
6088
|
+
sync?: ProjectResultsSyncConfig;
|
|
6088
6089
|
branchPrefix?: string;
|
|
6089
6090
|
}
|
|
6090
|
-
interface ProjectSource {
|
|
6091
|
-
url: string;
|
|
6092
|
-
ref: string;
|
|
6093
|
-
}
|
|
6094
6091
|
interface ProjectEntry {
|
|
6095
6092
|
id: string;
|
|
6096
6093
|
name: string;
|
|
6094
|
+
repoUrl?: string;
|
|
6097
6095
|
path: string;
|
|
6096
|
+
ref?: string;
|
|
6098
6097
|
addedAt: string;
|
|
6099
6098
|
lastOpenedAt: string;
|
|
6100
|
-
source?: ProjectSource;
|
|
6101
6099
|
results?: ProjectResultsConfig;
|
|
6102
6100
|
}
|
|
6103
6101
|
interface ProjectRegistry {
|
|
@@ -6143,13 +6141,13 @@ declare function touchProject(projectId: string): void;
|
|
|
6143
6141
|
declare function discoverProjects(rootDir: string, maxDepth?: number): string[];
|
|
6144
6142
|
|
|
6145
6143
|
/**
|
|
6146
|
-
* Project sync — pulls remote
|
|
6144
|
+
* Project sync — pulls remote Git repos to the local path declared in the
|
|
6147
6145
|
* project registry before Dashboard/eval startup.
|
|
6148
6146
|
*
|
|
6149
6147
|
* Sync is oneshot only, triggered by the Dashboard UI "Sync" button or the
|
|
6150
6148
|
* `agentv project sync` CLI command. There is no daemon or continuous mode.
|
|
6151
6149
|
*
|
|
6152
|
-
* First run — git clone --depth 1 --filter=blob:none --branch <ref> <
|
|
6150
|
+
* First run — git clone --depth 1 --filter=blob:none --branch <ref> <repoUrl> <path>
|
|
6153
6151
|
* Subsequent — git pull --ff-only (when <path>/.git already exists)
|
|
6154
6152
|
*
|
|
6155
6153
|
* Usage:
|
|
@@ -6158,15 +6156,15 @@ declare function discoverProjects(rootDir: string, maxDepth?: number): string[];
|
|
|
6158
6156
|
*/
|
|
6159
6157
|
|
|
6160
6158
|
/**
|
|
6161
|
-
* Clone or pull a single project entry from its declared
|
|
6159
|
+
* Clone or pull a single project entry from its declared repo URL.
|
|
6162
6160
|
* - No .git present: shallow clone into entry.path.
|
|
6163
6161
|
* - .git present: git pull --ff-only to update in place.
|
|
6164
|
-
* Throws on git error or missing
|
|
6162
|
+
* Throws on git error or missing repoUrl/ref.
|
|
6165
6163
|
*/
|
|
6166
6164
|
declare function syncProject(entry: ProjectEntry): Promise<void>;
|
|
6167
6165
|
/**
|
|
6168
|
-
* Iterate project entries and sync any that have a
|
|
6169
|
-
* Entries without
|
|
6166
|
+
* Iterate project entries and sync any that have a repo URL declared.
|
|
6167
|
+
* Entries without repoUrl are skipped silently.
|
|
6170
6168
|
*/
|
|
6171
6169
|
declare function syncProjects(entries: ProjectEntry[]): Promise<void>;
|
|
6172
6170
|
|
|
@@ -6742,4 +6740,4 @@ type AgentKernel = {
|
|
|
6742
6740
|
};
|
|
6743
6741
|
declare function createAgentKernel(): AgentKernel;
|
|
6744
6742
|
|
|
6745
|
-
export { type AcquireWorkspaceOptions, type AgentKernel, type AgentVConfig$1 as AgentVConfig, type AgentVResolvedConfig, type AgentVConfig as AgentVTsConfig, type AgentVConfig$1 as AgentVYamlConfig, type AnthropicResolvedConfig, type ApiFormat, type ArgsMatchMode, type AssertContext, type AssertEntry, type AssertFn, type AssertResult, type AssertionEntry, type AssertionResult, type AssistantTestMessage, type AzureResolvedConfig, COMMON_TARGET_SETTINGS, type CacheConfig, type CheckedOutResultsRepoBranch, type ChildGraderResult, type ClaudeDiscoverOptions, type ClaudeResolvedConfig, type ClaudeSession, type CliResolvedConfig, CodeGrader, type CodeGraderConfig, type CodeGraderOptions, type CodexDiscoverOptions, type CodexSession, type CommandExecutor, type CompositeAggregatorConfig, CompositeGrader, type CompositeGraderConfig, type CompositeGraderOptions, type ConfidenceIntervalAggregation, type ContainsAllGraderConfig, type ContainsAnyGraderConfig, type ContainsGraderConfig, type Content, type ContentFile, type ContentImage, type ContentPreprocessorConfig, type ContentText, type ConversationAggregation, type ConversationMode, type ConversationTurn, type ConversationTurnInput, type CopilotCliResolvedConfig, type DiscoverOptions as CopilotDiscoverOptions, type CopilotLogResolvedConfig, type CopilotSdkResolvedConfig, type CopilotSession, type CopilotSessionMeta, CostGrader, type CostGraderConfig, type CostGraderOptions, type CreateContainerOptions, DEFAULT_CATEGORY, DEFAULT_EVAL_PATTERNS, DEFAULT_EXPLORATION_TOOLS, DEFAULT_GRADER_TEMPLATE, DEFAULT_THRESHOLD, type DependencyFailurePolicy, type DependencyResult, type DepsScanResult, DeterministicAssertionGrader, type DockerWorkspaceConfig, DockerWorkspaceProvider, type EndsWithGraderConfig, type EnsureSubagentsOptions, type EnsureSubagentsResult, type EnvLookup, type EqualsGraderConfig, type EvalAssertionInput, type EvalCase, type EvalConfig, type EvalGraderSource, type EvalMetadata, type EvalRunResult, type EvalSourceReference, type EvalSuiteResult, type EvalSummary, type EvalTargetRef, type EvalTest, type EvalTestInput, type EvalTestSource, type EvalsJsonCase, type EvalsJsonFile, type EvaluationCache, type EvaluationContext, type EvaluationResult, type EvaluationScore, type EvaluationVerdict, type ExecInContainerOptions, type ExecResult, type ExecutionDefaults, type ExecutionError, type ExecutionMetrics, ExecutionMetricsGrader, type ExecutionMetricsGraderConfig, type ExecutionMetricsGraderOptions, type ExecutionStatus, type FailOnError, type FailureStage, FieldAccuracyGrader, type FieldAccuracyGraderConfig, type FieldAccuracyGraderOptions, type FieldAggregationType, type FieldConfig, type FieldMatchType, type GeminiResolvedConfig, type GenerateRubricsOptions, type GitListedRun, type Grader, type GraderConfig, type GraderDispatchContext, type GraderFactory, type GraderFactoryFn, type GraderKind, GraderRegistry, type GraderResult, type IcontainsAllGraderConfig, type IcontainsAnyGraderConfig, type IcontainsGraderConfig, type InlineAssertEvaluatorConfig, type IsJsonGraderConfig, type JsonObject, type JsonPrimitive, type JsonValue, LatencyGrader, type LatencyGraderConfig, type LatencyGraderOptions, LlmGrader, type LlmGraderConfig, type LlmGraderOptions, type LlmGraderPromptAssembly, type LocalPathValidationError, type MeanAggregation, type Message, type MockResolvedConfig, NORMALIZED_REDACTION_LEVELS, NORMALIZED_TOOL_STATUSES, NORMALIZED_TRACE_EVENT_TYPES, NORMALIZED_TRACE_SOURCE_KINDS, NORMALIZED_TRAJECTORY_SCHEMA_VERSION, type NormalizedRawEvidence, NormalizedRawEvidenceWireSchema, type NormalizedRedactionLevel, type NormalizedRedactionState, NormalizedRedactionStateWireSchema, type NormalizedToolStatus, type NormalizedTraceBranch, NormalizedTraceBranchWireSchema, type NormalizedTraceError, NormalizedTraceErrorWireSchema, type NormalizedTraceEvent, type NormalizedTraceEventType, type NormalizedTraceEventWire, NormalizedTraceEventWireSchema, type NormalizedTraceMessage, NormalizedTraceMessageWireSchema, type NormalizedTraceModel, NormalizedTraceModelWireSchema, type NormalizedTraceSession, NormalizedTraceSessionWireSchema, type NormalizedTraceSource, type NormalizedTraceSourceKind, type NormalizedTraceSourceRef, NormalizedTraceSourceRefWireSchema, NormalizedTraceSourceWireSchema, type NormalizedTraceTool, NormalizedTraceToolWireSchema, type NormalizedTrajectory, type NormalizedTrajectoryWire, NormalizedTrajectoryWireSchema, OTEL_BACKEND_PRESETS, type OpenAIResolvedConfig, type OpenRouterResolvedConfig, type OtelBackendPreset, type OtelExportOptions, OtelStreamingObserver, OtelTraceExporter, OtlpJsonFileExporter, type OutputMessage, PASS_THRESHOLD, type ParsedCopilotSession, type PassAtKAggregation, type PiCliResolvedConfig, type PiCodingAgentResolvedConfig, type PoolSlot, type PreparedResultsRepoBranch, type ProgressEvent, type ProjectEntry, type ProjectRegistry, type ProjectSource, type PromptInputs, type PromptScriptConfig, type Provider, type ProviderFactoryFn, type ProviderKind, ProviderRegistry, type ProviderRequest, type ProviderResponse, type ProviderStreamCallbacks, type ProviderTokenUsage, RUBRIC_OPERATOR_VALUES, type RegexGraderConfig, type RepoCheckout, type RepoClone, type RepoConfig, type RepoDep, RepoManager, type RepoSource, type ResolvedTarget, type ResolvedWorkspaceTemplate, ResponseCache, type ResultsConfig, type ResultsRepoLocalPaths, type ResultsRepoStatus, type ResultsRepoSyncStatus, type RubricItem, type RubricOperator, type RubricsEvaluatorConfig, RunBudgetTracker, type RunEvalCaseOptions, type RunEvaluationOptions, type ScoreRange, type ScriptExecutionContext, SkillTriggerGrader, type SkillTriggerGraderConfig, type StartsWithGraderConfig, type SystemTestMessage, TEST_MESSAGE_ROLES, type TargetAccessConfig, type TargetDefinition, type TargetHooksConfig, TemplateNotDirectoryError, TemplateNotFoundError, type TestMessage, type TestMessageContent, type TestMessageRole, type TokenUsage, TokenUsageGrader, type TokenUsageGraderConfig, type TokenUsageGraderOptions, type ToolCall, type ToolTestMessage, type ToolTrajectoryExpectedItem, ToolTrajectoryGrader, type ToolTrajectoryGraderConfig, type ToolTrajectoryGraderOptions, type TraceComputeResult, type TraceSummary, type TranscriptEntry, type TranscriptJsonLine, TranscriptProvider, type TranscriptReplayEntry, type TranscriptSource, type TranspileResult, type TrialAggregation, type TrialResult, type TrialStrategy, type TrialsConfig, type TsEvalResult, type TurnFailurePolicy, type UserTestMessage, type VSCodeResolvedConfig, type WorkspaceConfig, WorkspaceCreationError, type WorkspaceEnvConfig, type WorkspaceHookConfig, type WorkspaceHooksConfig, WorkspacePoolManager, type WorkspaceScriptConfig, addProject, assembleLlmGraderPrompt, avgToolDurationMs, buildDirectoryChain, buildOutputSchema, buildPromptInputs, buildRubricOutputSchema, buildScoreRangeOutputSchema, buildSearchRoots, calculateRubricScore, captureFileChanges, checkoutResultsRepoBranch, clampScore, cleanupEvalWorkspaces, cleanupWorkspace, commitAndPushResultsBranch, computeTraceSummary, computeTraceSummaryFromTrajectory, computeWorkspaceFingerprint, consumeClaudeLogEntries, consumeCodexLogEntries, consumeCopilotCliLogEntries, consumeCopilotSdkLogEntries, consumePiLogEntries, createAgentKernel, createBuiltinProviderRegistry, createBuiltinRegistry, createDraftResultsPr, createProvider, createTempWorkspace, deepEqual, defineConfig, deriveCategory, deriveProjectId, detectFormat, directPushResults, directorySizeBytes, discoverAssertions, discoverClaudeSessions, discoverCodexSessions, discoverCopilotSessions, discoverGraders, discoverProjects, discoverProviders, ensureResultsRepoClone, ensureVSCodeSubagents, evaluate, executeScript, executeWorkspaceScript, explorationRatio, extractCacheConfig, extractFailOnError, extractImageBlocks, extractJsonBlob, extractLastAssistantContent, extractTargetFromSuite, extractTargetRefsFromSuite, extractTargetsFromSuite, extractTargetsFromTestCase, extractThreshold, extractTrialsConfig, extractWorkersFromSuite, fileExists, findGitRoot, formatToolCalls, freeformEvaluationSchema, fromNormalizedTrajectoryWire, generateRubrics, getAgentvConfigDir, getAgentvDataDir, getAgentvHome, getOutputFilenames, getProject, getProjectForPath, getProjectsRegistryPath, getResultsRepoLocalPaths, getResultsRepoStatus, getResultsRepoSyncStatus, getSelectedTrajectoryEvents, getSubagentsRoot, getTextContent, getTraceStateRoot, getWorkspacePath, getWorkspacePoolRoot, getWorkspacesRoot, groupTranscriptJsonLines, initializeBaseline, isAgentSkillsFormat, isContent, isContentArray, isGraderKind, isJsonObject, isJsonValue, isNonEmptyString, isTestMessage, isTestMessageRole, killAllTrackedChildren, listGitRuns, listTargetNames, loadConfig, loadEvalCaseById, loadEvalCases, loadEvalSuite, loadProjectRegistry, loadTestById, loadTestSuite, loadTests, loadTsConfig, loadTsEvalFile, materializeGitRun, mergeExecutionMetrics, negateScore, normalizeLineEndings, normalizeResultsConfig, parseAgentSkillsEvals, parseClaudeSession, parseCodexSession, parseCopilotEvents, parseEnvOutput, parseJsonFromText, parseJsonSafe, parseYamlValue, prepareResultsRepoBranch, pushResultsRepoBranch, readJsonFile, readTargetDefinitions, readTestSuiteMetadata, readTextFile, readTranscriptFile, readTranscriptJsonl, removeProject, resolveAndCreateProvider, resolveDelegatedTargetDefinition, resolveFileReference, resolveResultsConfigForProject, resolveResultsRepoRunsDir, resolveResultsRepoUrl, resolveTargetDefinition, resolveWorkspaceTemplate, rubricEvaluationSchema, runBeforeSessionHook, runContainsAllAssertion, runContainsAnyAssertion, runContainsAssertion, runEndsWithAssertion, runEqualsAssertion, runEvalCase, runEvaluation, runIcontainsAllAssertion, runIcontainsAnyAssertion, runIcontainsAssertion, runIsJsonAssertion, runRegexAssertion, runStartsWithAssertion, saveProjectRegistry, scanRepoDeps, scoreRangeEvaluationSchema, scoreToVerdict, shouldEnableCache, shouldSkipCacheForTemperature, stageResultsArtifacts, subscribeToClaudeLogEntries, subscribeToCodexLogEntries, subscribeToCopilotCliLogEntries, subscribeToCopilotSdkLogEntries, subscribeToPiLogEntries, substituteVariables, syncProject, syncProjects, syncResultsRepo, syncResultsRepoForProject, toCamelCaseDeep, toNormalizedTrajectoryWire, toSnakeCaseDeep, toTranscriptJsonLines, tokensPerTool, touchProject, trackChild, trackedChildCount, transpileEvalYaml, transpileEvalYamlFile, trimBaselineResult };
|
|
6743
|
+
export { type AcquireWorkspaceOptions, type AgentKernel, type AgentVConfig$1 as AgentVConfig, type AgentVResolvedConfig, type AgentVConfig as AgentVTsConfig, type AgentVConfig$1 as AgentVYamlConfig, type AnthropicResolvedConfig, type ApiFormat, type ArgsMatchMode, type AssertContext, type AssertEntry, type AssertFn, type AssertResult, type AssertionEntry, type AssertionResult, type AssistantTestMessage, type AzureResolvedConfig, COMMON_TARGET_SETTINGS, type CacheConfig, type CheckedOutResultsRepoBranch, type ChildGraderResult, type ClaudeDiscoverOptions, type ClaudeResolvedConfig, type ClaudeSession, type CliResolvedConfig, CodeGrader, type CodeGraderConfig, type CodeGraderOptions, type CodexDiscoverOptions, type CodexSession, type CommandExecutor, type CompositeAggregatorConfig, CompositeGrader, type CompositeGraderConfig, type CompositeGraderOptions, type ConfidenceIntervalAggregation, type ContainsAllGraderConfig, type ContainsAnyGraderConfig, type ContainsGraderConfig, type Content, type ContentFile, type ContentImage, type ContentPreprocessorConfig, type ContentText, type ConversationAggregation, type ConversationMode, type ConversationTurn, type ConversationTurnInput, type CopilotCliResolvedConfig, type DiscoverOptions as CopilotDiscoverOptions, type CopilotLogResolvedConfig, type CopilotSdkResolvedConfig, type CopilotSession, type CopilotSessionMeta, CostGrader, type CostGraderConfig, type CostGraderOptions, type CreateContainerOptions, DEFAULT_CATEGORY, DEFAULT_EVAL_PATTERNS, DEFAULT_EXPLORATION_TOOLS, DEFAULT_GRADER_TEMPLATE, DEFAULT_THRESHOLD, type DependencyFailurePolicy, type DependencyResult, type DepsScanResult, DeterministicAssertionGrader, type DockerWorkspaceConfig, DockerWorkspaceProvider, type EndsWithGraderConfig, type EnsureSubagentsOptions, type EnsureSubagentsResult, type EnvLookup, type EqualsGraderConfig, type EvalAssertionInput, type EvalCase, type EvalConfig, type EvalGraderSource, type EvalMetadata, type EvalRunResult, type EvalSourceReference, type EvalSuiteResult, type EvalSummary, type EvalTargetRef, type EvalTest, type EvalTestInput, type EvalTestSource, type EvalsJsonCase, type EvalsJsonFile, type EvaluationCache, type EvaluationContext, type EvaluationResult, type EvaluationScore, type EvaluationVerdict, type ExecInContainerOptions, type ExecResult, type ExecutionDefaults, type ExecutionError, type ExecutionMetrics, ExecutionMetricsGrader, type ExecutionMetricsGraderConfig, type ExecutionMetricsGraderOptions, type ExecutionStatus, type FailOnError, type FailureStage, FieldAccuracyGrader, type FieldAccuracyGraderConfig, type FieldAccuracyGraderOptions, type FieldAggregationType, type FieldConfig, type FieldMatchType, type GeminiResolvedConfig, type GenerateRubricsOptions, type GitListedRun, type Grader, type GraderConfig, type GraderDispatchContext, type GraderFactory, type GraderFactoryFn, type GraderKind, GraderRegistry, type GraderResult, type IcontainsAllGraderConfig, type IcontainsAnyGraderConfig, type IcontainsGraderConfig, type InlineAssertEvaluatorConfig, type IsJsonGraderConfig, type JsonObject, type JsonPrimitive, type JsonValue, LatencyGrader, type LatencyGraderConfig, type LatencyGraderOptions, LlmGrader, type LlmGraderConfig, type LlmGraderOptions, type LlmGraderPromptAssembly, type LocalPathValidationError, type MeanAggregation, type Message, type MockResolvedConfig, NORMALIZED_REDACTION_LEVELS, NORMALIZED_TOOL_STATUSES, NORMALIZED_TRACE_EVENT_TYPES, NORMALIZED_TRACE_SOURCE_KINDS, NORMALIZED_TRAJECTORY_SCHEMA_VERSION, type NormalizedRawEvidence, NormalizedRawEvidenceWireSchema, type NormalizedRedactionLevel, type NormalizedRedactionState, NormalizedRedactionStateWireSchema, type NormalizedToolStatus, type NormalizedTraceBranch, NormalizedTraceBranchWireSchema, type NormalizedTraceError, NormalizedTraceErrorWireSchema, type NormalizedTraceEvent, type NormalizedTraceEventType, type NormalizedTraceEventWire, NormalizedTraceEventWireSchema, type NormalizedTraceMessage, NormalizedTraceMessageWireSchema, type NormalizedTraceModel, NormalizedTraceModelWireSchema, type NormalizedTraceSession, NormalizedTraceSessionWireSchema, type NormalizedTraceSource, type NormalizedTraceSourceKind, type NormalizedTraceSourceRef, NormalizedTraceSourceRefWireSchema, NormalizedTraceSourceWireSchema, type NormalizedTraceTool, NormalizedTraceToolWireSchema, type NormalizedTrajectory, type NormalizedTrajectoryWire, NormalizedTrajectoryWireSchema, OTEL_BACKEND_PRESETS, type OpenAIResolvedConfig, type OpenRouterResolvedConfig, type OtelBackendPreset, type OtelExportOptions, OtelStreamingObserver, OtelTraceExporter, OtlpJsonFileExporter, type OutputMessage, PASS_THRESHOLD, type ParsedCopilotSession, type PassAtKAggregation, type PiCliResolvedConfig, type PiCodingAgentResolvedConfig, type PoolSlot, type PreparedResultsRepoBranch, type ProgressEvent, type ProjectEntry, type ProjectRegistry, type PromptInputs, type PromptScriptConfig, type Provider, type ProviderFactoryFn, type ProviderKind, ProviderRegistry, type ProviderRequest, type ProviderResponse, type ProviderStreamCallbacks, type ProviderTokenUsage, RUBRIC_OPERATOR_VALUES, type RegexGraderConfig, type RepoCheckout, type RepoClone, type RepoConfig, type RepoDep, RepoManager, type RepoSource, type ResolvedTarget, type ResolvedWorkspaceTemplate, ResponseCache, type ResultsConfig, type ResultsRepoLocalPaths, type ResultsRepoStatus, type ResultsRepoSyncStatus, type RubricItem, type RubricOperator, type RubricsEvaluatorConfig, RunBudgetTracker, type RunEvalCaseOptions, type RunEvaluationOptions, type ScoreRange, type ScriptExecutionContext, SkillTriggerGrader, type SkillTriggerGraderConfig, type StartsWithGraderConfig, type SystemTestMessage, TEST_MESSAGE_ROLES, type TargetAccessConfig, type TargetDefinition, type TargetHooksConfig, TemplateNotDirectoryError, TemplateNotFoundError, type TestMessage, type TestMessageContent, type TestMessageRole, type TokenUsage, TokenUsageGrader, type TokenUsageGraderConfig, type TokenUsageGraderOptions, type ToolCall, type ToolTestMessage, type ToolTrajectoryExpectedItem, ToolTrajectoryGrader, type ToolTrajectoryGraderConfig, type ToolTrajectoryGraderOptions, type TraceComputeResult, type TraceSummary, type TranscriptEntry, type TranscriptJsonLine, TranscriptProvider, type TranscriptReplayEntry, type TranscriptSource, type TranspileResult, type TrialAggregation, type TrialResult, type TrialStrategy, type TrialsConfig, type TsEvalResult, type TurnFailurePolicy, type UserTestMessage, type VSCodeResolvedConfig, type WorkspaceConfig, WorkspaceCreationError, type WorkspaceEnvConfig, type WorkspaceHookConfig, type WorkspaceHooksConfig, WorkspacePoolManager, type WorkspaceScriptConfig, addProject, assembleLlmGraderPrompt, avgToolDurationMs, buildDirectoryChain, buildOutputSchema, buildPromptInputs, buildRubricOutputSchema, buildScoreRangeOutputSchema, buildSearchRoots, calculateRubricScore, captureFileChanges, checkoutResultsRepoBranch, clampScore, cleanupEvalWorkspaces, cleanupWorkspace, commitAndPushResultsBranch, computeTraceSummary, computeTraceSummaryFromTrajectory, computeWorkspaceFingerprint, consumeClaudeLogEntries, consumeCodexLogEntries, consumeCopilotCliLogEntries, consumeCopilotSdkLogEntries, consumePiLogEntries, createAgentKernel, createBuiltinProviderRegistry, createBuiltinRegistry, createDraftResultsPr, createProvider, createTempWorkspace, deepEqual, defineConfig, deriveCategory, deriveProjectId, detectFormat, directPushResults, directorySizeBytes, discoverAssertions, discoverClaudeSessions, discoverCodexSessions, discoverCopilotSessions, discoverGraders, discoverProjects, discoverProviders, ensureResultsRepoClone, ensureVSCodeSubagents, evaluate, executeScript, executeWorkspaceScript, explorationRatio, extractCacheConfig, extractFailOnError, extractImageBlocks, extractJsonBlob, extractLastAssistantContent, extractTargetFromSuite, extractTargetRefsFromSuite, extractTargetsFromSuite, extractTargetsFromTestCase, extractThreshold, extractTrialsConfig, extractWorkersFromSuite, fileExists, findGitRoot, formatToolCalls, freeformEvaluationSchema, fromNormalizedTrajectoryWire, generateRubrics, getAgentvConfigDir, getAgentvDataDir, getAgentvHome, getOutputFilenames, getProject, getProjectForPath, getProjectsRegistryPath, getResultsRepoLocalPaths, getResultsRepoStatus, getResultsRepoSyncStatus, getSelectedTrajectoryEvents, getSubagentsRoot, getTextContent, getTraceStateRoot, getWorkspacePath, getWorkspacePoolRoot, getWorkspacesRoot, groupTranscriptJsonLines, initializeBaseline, isAgentSkillsFormat, isContent, isContentArray, isGraderKind, isJsonObject, isJsonValue, isNonEmptyString, isTestMessage, isTestMessageRole, killAllTrackedChildren, listGitRuns, listTargetNames, loadConfig, loadEvalCaseById, loadEvalCases, loadEvalSuite, loadProjectRegistry, loadTestById, loadTestSuite, loadTests, loadTsConfig, loadTsEvalFile, materializeGitRun, mergeExecutionMetrics, negateScore, normalizeLineEndings, normalizeResultsConfig, parseAgentSkillsEvals, parseClaudeSession, parseCodexSession, parseCopilotEvents, parseEnvOutput, parseJsonFromText, parseJsonSafe, parseYamlValue, prepareResultsRepoBranch, pushResultsRepoBranch, readJsonFile, readTargetDefinitions, readTestSuiteMetadata, readTextFile, readTranscriptFile, readTranscriptJsonl, removeProject, resolveAndCreateProvider, resolveDelegatedTargetDefinition, resolveFileReference, resolveResultsConfigForProject, resolveResultsRepoRunsDir, resolveResultsRepoUrl, resolveTargetDefinition, resolveWorkspaceTemplate, rubricEvaluationSchema, runBeforeSessionHook, runContainsAllAssertion, runContainsAnyAssertion, runContainsAssertion, runEndsWithAssertion, runEqualsAssertion, runEvalCase, runEvaluation, runIcontainsAllAssertion, runIcontainsAnyAssertion, runIcontainsAssertion, runIsJsonAssertion, runRegexAssertion, runStartsWithAssertion, saveProjectRegistry, scanRepoDeps, scoreRangeEvaluationSchema, scoreToVerdict, shouldEnableCache, shouldSkipCacheForTemperature, stageResultsArtifacts, subscribeToClaudeLogEntries, subscribeToCodexLogEntries, subscribeToCopilotCliLogEntries, subscribeToCopilotSdkLogEntries, subscribeToPiLogEntries, substituteVariables, syncProject, syncProjects, syncResultsRepo, syncResultsRepoForProject, toCamelCaseDeep, toNormalizedTrajectoryWire, toSnakeCaseDeep, toTranscriptJsonLines, tokensPerTool, touchProject, trackChild, trackedChildCount, transpileEvalYaml, transpileEvalYamlFile, trimBaselineResult };
|
package/dist/index.js
CHANGED
|
@@ -1730,20 +1730,20 @@ function fromYaml(raw) {
|
|
|
1730
1730
|
addedAt: typeof e.added_at === "string" ? e.added_at : "",
|
|
1731
1731
|
lastOpenedAt: typeof e.last_opened_at === "string" ? e.last_opened_at : ""
|
|
1732
1732
|
};
|
|
1733
|
-
if (
|
|
1734
|
-
|
|
1735
|
-
|
|
1736
|
-
|
|
1737
|
-
|
|
1733
|
+
if (typeof e.repo_url === "string" && e.repo_url.trim().length > 0) {
|
|
1734
|
+
entry.repoUrl = e.repo_url.trim();
|
|
1735
|
+
}
|
|
1736
|
+
if (typeof e.ref === "string" && e.ref.trim().length > 0) {
|
|
1737
|
+
entry.ref = e.ref.trim();
|
|
1738
1738
|
}
|
|
1739
1739
|
if (e.results && typeof e.results === "object") {
|
|
1740
1740
|
const r = e.results;
|
|
1741
|
-
if (
|
|
1741
|
+
if (typeof r.repo_url === "string" && r.repo_url.trim().length > 0) {
|
|
1742
|
+
const sync = r.sync && typeof r.sync === "object" ? r.sync : void 0;
|
|
1742
1743
|
entry.results = {
|
|
1743
|
-
|
|
1744
|
-
repo: r.repo.trim(),
|
|
1744
|
+
repoUrl: r.repo_url.trim(),
|
|
1745
1745
|
...typeof r.path === "string" && r.path.trim().length > 0 ? { path: r.path.trim() } : {},
|
|
1746
|
-
...typeof
|
|
1746
|
+
...sync && typeof sync.auto_push === "boolean" ? { sync: { autoPush: sync.auto_push } } : {},
|
|
1747
1747
|
...typeof r.branch_prefix === "string" && r.branch_prefix.trim().length > 0 ? { branchPrefix: r.branch_prefix.trim() } : {}
|
|
1748
1748
|
};
|
|
1749
1749
|
}
|
|
@@ -1754,19 +1754,19 @@ function toYaml(entry) {
|
|
|
1754
1754
|
const yaml = {
|
|
1755
1755
|
id: entry.id,
|
|
1756
1756
|
name: entry.name,
|
|
1757
|
+
...entry.repoUrl !== void 0 && { repo_url: entry.repoUrl },
|
|
1757
1758
|
path: entry.path,
|
|
1759
|
+
...entry.ref !== void 0 && { ref: entry.ref },
|
|
1758
1760
|
added_at: entry.addedAt,
|
|
1759
1761
|
last_opened_at: entry.lastOpenedAt
|
|
1760
1762
|
};
|
|
1761
|
-
if (entry.source) {
|
|
1762
|
-
yaml.source = { url: entry.source.url, ref: entry.source.ref };
|
|
1763
|
-
}
|
|
1764
1763
|
if (entry.results) {
|
|
1765
1764
|
yaml.results = {
|
|
1766
|
-
|
|
1767
|
-
repo: entry.results.repo,
|
|
1765
|
+
repo_url: entry.results.repoUrl,
|
|
1768
1766
|
...entry.results.path !== void 0 && { path: entry.results.path },
|
|
1769
|
-
...entry.results.autoPush !== void 0 && {
|
|
1767
|
+
...entry.results.sync?.autoPush !== void 0 && {
|
|
1768
|
+
sync: { auto_push: entry.results.sync.autoPush }
|
|
1769
|
+
},
|
|
1770
1770
|
...entry.results.branchPrefix !== void 0 && {
|
|
1771
1771
|
branch_prefix: entry.results.branchPrefix
|
|
1772
1772
|
}
|
|
@@ -1906,25 +1906,27 @@ function discoverProjects(rootDir, maxDepth = 2) {
|
|
|
1906
1906
|
import * as childProcess from "node:child_process";
|
|
1907
1907
|
import { existsSync as existsSync3 } from "node:fs";
|
|
1908
1908
|
async function syncProject(entry) {
|
|
1909
|
-
if (!entry.
|
|
1910
|
-
throw new Error(`Project '${entry.id}' has no
|
|
1909
|
+
if (!entry.repoUrl) {
|
|
1910
|
+
throw new Error(`Project '${entry.id}' has no repo_url defined`);
|
|
1911
|
+
}
|
|
1912
|
+
if (!entry.ref) {
|
|
1913
|
+
throw new Error(`Project '${entry.id}' has no ref defined`);
|
|
1911
1914
|
}
|
|
1912
|
-
const { url, ref } = entry.source;
|
|
1913
1915
|
const dest = entry.path;
|
|
1914
1916
|
if (existsSync3(`${dest}/.git`)) {
|
|
1915
1917
|
childProcess.execFileSync("git", ["-C", dest, "pull", "--ff-only"], { stdio: "inherit" });
|
|
1916
1918
|
} else {
|
|
1917
1919
|
childProcess.execFileSync(
|
|
1918
1920
|
"git",
|
|
1919
|
-
["clone", "--depth", "1", "--filter=blob:none", "--branch", ref,
|
|
1921
|
+
["clone", "--depth", "1", "--filter=blob:none", "--branch", entry.ref, entry.repoUrl, dest],
|
|
1920
1922
|
{ stdio: "inherit" }
|
|
1921
1923
|
);
|
|
1922
1924
|
}
|
|
1923
1925
|
}
|
|
1924
1926
|
async function syncProjects(entries) {
|
|
1925
1927
|
for (const entry of entries) {
|
|
1926
|
-
if (!entry.
|
|
1927
|
-
console.log(`Syncing project '${entry.id}' from ${entry.
|
|
1928
|
+
if (!entry.repoUrl) continue;
|
|
1929
|
+
console.log(`Syncing project '${entry.id}' from ${entry.repoUrl}...`);
|
|
1928
1930
|
await syncProject(entry);
|
|
1929
1931
|
console.log(`Project '${entry.id}' synced.`);
|
|
1930
1932
|
}
|