@agentv/core 4.11.2-next.1 → 4.12.0-next.1

package/dist/index.cjs

@@ -1801,7 +1801,7 @@ __export(index_exports, {
   TranscriptProvider: () => TranscriptProvider,
   WorkspaceCreationError: () => WorkspaceCreationError,
   WorkspacePoolManager: () => WorkspacePoolManager,
-  addProject: () => addProject,
+  addBenchmark: () => addBenchmark,
   assembleLlmGraderPrompt: () => assembleLlmGraderPrompt,
   assembleLlmJudgePrompt: () => assembleLlmGraderPrompt,
   avgToolDurationMs: () => avgToolDurationMs,
@@ -1833,17 +1833,17 @@ __export(index_exports, {
   createTempWorkspace: () => createTempWorkspace,
   deepEqual: () => deepEqual,
   defineConfig: () => defineConfig,
+  deriveBenchmarkId: () => deriveBenchmarkId,
   deriveCategory: () => deriveCategory,
-  deriveProjectId: () => deriveProjectId,
   detectFormat: () => detectFormat,
   directorySizeBytes: () => directorySizeBytes,
   discoverAssertions: () => discoverAssertions,
+  discoverBenchmarks: () => discoverBenchmarks,
   discoverClaudeSessions: () => discoverClaudeSessions,
   discoverCodexSessions: () => discoverCodexSessions,
   discoverCopilotSessions: () => discoverCopilotSessions,
   discoverGraders: () => discoverGraders,
   discoverJudges: () => discoverGraders,
-  discoverProjects: () => discoverProjects,
   discoverProviders: () => discoverProviders,
   ensureResultsRepoClone: () => ensureResultsRepoClone,
   ensureVSCodeSubagents: () => ensureVSCodeSubagents,
@@ -1867,9 +1867,9 @@ __export(index_exports, {
   freeformEvaluationSchema: () => freeformEvaluationSchema,
   generateRubrics: () => generateRubrics,
   getAgentvHome: () => getAgentvHome,
+  getBenchmark: () => getBenchmark,
+  getBenchmarksRegistryPath: () => getBenchmarksRegistryPath,
   getOutputFilenames: () => getOutputFilenames,
-  getProject: () => getProject,
-  getProjectsRegistryPath: () => getProjectsRegistryPath,
   getResultsRepoCachePaths: () => getResultsRepoCachePaths,
   getResultsRepoStatus: () => getResultsRepoStatus,
   getSubagentsRoot: () => getSubagentsRoot,
@@ -1889,11 +1889,11 @@ __export(index_exports, {
   isTestMessage: () => isTestMessage,
   isTestMessageRole: () => isTestMessageRole,
   listTargetNames: () => listTargetNames,
+  loadBenchmarkRegistry: () => loadBenchmarkRegistry,
   loadConfig: () => loadConfig,
   loadEvalCaseById: () => loadEvalCaseById,
   loadEvalCases: () => loadEvalCases,
   loadEvalSuite: () => loadEvalSuite,
-  loadProjectRegistry: () => loadProjectRegistry,
   loadTestById: () => loadTestById,
   loadTestSuite: () => loadTestSuite,
   loadTests: () => loadTests,
@@ -1916,7 +1916,7 @@ __export(index_exports, {
   readTextFile: () => readTextFile,
   readTranscriptFile: () => readTranscriptFile,
   readTranscriptJsonl: () => readTranscriptJsonl,
-  removeProject: () => removeProject,
+  removeBenchmark: () => removeBenchmark,
   resolveAndCreateProvider: () => resolveAndCreateProvider,
   resolveDelegatedTargetDefinition: () => resolveDelegatedTargetDefinition,
   resolveFileReference: () => resolveFileReference3,
@@ -1938,7 +1938,7 @@ __export(index_exports, {
   runIsJsonAssertion: () => runIsJsonAssertion,
   runRegexAssertion: () => runRegexAssertion,
   runStartsWithAssertion: () => runStartsWithAssertion,
-  saveProjectRegistry: () => saveProjectRegistry,
+  saveBenchmarkRegistry: () => saveBenchmarkRegistry,
   scanRepoDeps: () => scanRepoDeps,
   scoreToVerdict: () => scoreToVerdict,
   shouldEnableCache: () => shouldEnableCache,
@@ -1955,7 +1955,7 @@ __export(index_exports, {
   toSnakeCaseDeep: () => toSnakeCaseDeep,
   toTranscriptJsonLine: () => toTranscriptJsonLine,
   tokensPerTool: () => tokensPerTool,
-  touchProject: () => touchProject,
+  touchBenchmark: () => touchBenchmark,
   transpileEvalYaml: () => transpileEvalYaml,
   transpileEvalYamlFile: () => transpileEvalYamlFile,
   trimBaselineResult: () => trimBaselineResult
@@ -23102,41 +23102,41 @@ async function createDraftResultsPr(params) {
   return stdout.trim();
 }
 
-// src/projects.ts
+// src/benchmarks.ts
 init_cjs_shims();
 var import_node_fs19 = require("fs");
 var import_node_path54 = __toESM(require("path"), 1);
 var import_yaml10 = require("yaml");
-function getProjectsRegistryPath() {
+function getBenchmarksRegistryPath() {
   return import_node_path54.default.join(getAgentvHome(), "projects.yaml");
 }
-function loadProjectRegistry() {
-  const registryPath = getProjectsRegistryPath();
+function loadBenchmarkRegistry() {
+  const registryPath = getBenchmarksRegistryPath();
   if (!(0, import_node_fs19.existsSync)(registryPath)) {
-    return { projects: [] };
+    return { benchmarks: [] };
   }
   try {
     const raw = (0, import_node_fs19.readFileSync)(registryPath, "utf-8");
     const parsed = (0, import_yaml10.parse)(raw);
-    if (!parsed || !Array.isArray(parsed.projects)) {
-      return { projects: [] };
+    if (!parsed || !Array.isArray(parsed.benchmarks)) {
+      return { benchmarks: [] };
     }
-    return { projects: parsed.projects };
+    return { benchmarks: parsed.benchmarks };
   } catch {
-    return { projects: [] };
+    return { benchmarks: [] };
   }
 }
-function saveProjectRegistry(registry) {
-  const registryPath = getProjectsRegistryPath();
+function saveBenchmarkRegistry(registry) {
+  const registryPath = getBenchmarksRegistryPath();
   const dir = import_node_path54.default.dirname(registryPath);
   if (!(0, import_node_fs19.existsSync)(dir)) {
     (0, import_node_fs19.mkdirSync)(dir, { recursive: true });
   }
-  (0, import_node_fs19.writeFileSync)(registryPath, (0, import_yaml10.stringify)(registry), "utf-8");
+  (0, import_node_fs19.writeFileSync)(registryPath, (0, import_yaml10.stringify)({ benchmarks: registry.benchmarks }), "utf-8");
 }
-function deriveProjectId(dirPath, existingIds) {
+function deriveBenchmarkId(dirPath, existingIds) {
   const base = import_node_path54.default.basename(dirPath).toLowerCase().replace(/[^a-z0-9-]/g, "-").replace(/-+/g, "-").replace(/^-|-$/g, "");
-  let candidate = base || "project";
+  let candidate = base || "benchmark";
   let suffix = 2;
   while (existingIds.includes(candidate)) {
     candidate = `${base}-${suffix}`;
@@ -23144,54 +23144,54 @@ function deriveProjectId(dirPath, existingIds) {
   }
   return candidate;
 }
-function addProject(projectPath) {
-  const absPath = import_node_path54.default.resolve(projectPath);
+function addBenchmark(benchmarkPath) {
+  const absPath = import_node_path54.default.resolve(benchmarkPath);
   if (!(0, import_node_fs19.existsSync)(absPath)) {
     throw new Error(`Directory not found: ${absPath}`);
   }
   if (!(0, import_node_fs19.existsSync)(import_node_path54.default.join(absPath, ".agentv"))) {
     throw new Error(`No .agentv/ directory found in ${absPath}. Run an evaluation first.`);
   }
-  const registry = loadProjectRegistry();
-  const existing = registry.projects.find((p) => p.path === absPath);
+  const registry = loadBenchmarkRegistry();
+  const existing = registry.benchmarks.find((p) => p.path === absPath);
   if (existing) {
     return existing;
   }
   const now = (/* @__PURE__ */ new Date()).toISOString();
   const entry = {
-    id: deriveProjectId(
+    id: deriveBenchmarkId(
       absPath,
-      registry.projects.map((p) => p.id)
+      registry.benchmarks.map((p) => p.id)
     ),
     name: import_node_path54.default.basename(absPath),
     path: absPath,
     addedAt: now,
     lastOpenedAt: now
   };
-  registry.projects.push(entry);
-  saveProjectRegistry(registry);
+  registry.benchmarks.push(entry);
+  saveBenchmarkRegistry(registry);
   return entry;
 }
-function removeProject(projectId) {
-  const registry = loadProjectRegistry();
-  const idx = registry.projects.findIndex((p) => p.id === projectId);
+function removeBenchmark(benchmarkId) {
+  const registry = loadBenchmarkRegistry();
+  const idx = registry.benchmarks.findIndex((p) => p.id === benchmarkId);
   if (idx < 0) return false;
-  registry.projects.splice(idx, 1);
-  saveProjectRegistry(registry);
+  registry.benchmarks.splice(idx, 1);
+  saveBenchmarkRegistry(registry);
   return true;
 }
-function getProject(projectId) {
-  return loadProjectRegistry().projects.find((p) => p.id === projectId);
+function getBenchmark(benchmarkId) {
+  return loadBenchmarkRegistry().benchmarks.find((p) => p.id === benchmarkId);
 }
-function touchProject(projectId) {
-  const registry = loadProjectRegistry();
-  const entry = registry.projects.find((p) => p.id === projectId);
+function touchBenchmark(benchmarkId) {
+  const registry = loadBenchmarkRegistry();
+  const entry = registry.benchmarks.find((p) => p.id === benchmarkId);
   if (entry) {
     entry.lastOpenedAt = (/* @__PURE__ */ new Date()).toISOString();
-    saveProjectRegistry(registry);
+    saveBenchmarkRegistry(registry);
   }
 }
-function discoverProjects(rootDir, maxDepth = 2) {
+function discoverBenchmarks(rootDir, maxDepth = 2) {
   const absRoot = import_node_path54.default.resolve(rootDir);
   if (!(0, import_node_fs19.existsSync)(absRoot) || !(0, import_node_fs19.statSync)(absRoot).isDirectory()) {
     return [];
@@ -24379,7 +24379,7 @@ function createAgentKernel() {
   TranscriptProvider,
   WorkspaceCreationError,
   WorkspacePoolManager,
-  addProject,
+  addBenchmark,
   assembleLlmGraderPrompt,
   assembleLlmJudgePrompt,
   avgToolDurationMs,
@@ -24411,17 +24411,17 @@ function createAgentKernel() {
   createTempWorkspace,
   deepEqual,
   defineConfig,
+  deriveBenchmarkId,
   deriveCategory,
-  deriveProjectId,
   detectFormat,
   directorySizeBytes,
   discoverAssertions,
+  discoverBenchmarks,
   discoverClaudeSessions,
   discoverCodexSessions,
   discoverCopilotSessions,
   discoverGraders,
   discoverJudges,
-  discoverProjects,
   discoverProviders,
   ensureResultsRepoClone,
   ensureVSCodeSubagents,
@@ -24445,9 +24445,9 @@ function createAgentKernel() {
   freeformEvaluationSchema,
   generateRubrics,
   getAgentvHome,
+  getBenchmark,
+  getBenchmarksRegistryPath,
   getOutputFilenames,
-  getProject,
-  getProjectsRegistryPath,
   getResultsRepoCachePaths,
   getResultsRepoStatus,
   getSubagentsRoot,
@@ -24467,11 +24467,11 @@ function createAgentKernel() {
   isTestMessage,
   isTestMessageRole,
   listTargetNames,
+  loadBenchmarkRegistry,
   loadConfig,
   loadEvalCaseById,
   loadEvalCases,
   loadEvalSuite,
-  loadProjectRegistry,
   loadTestById,
   loadTestSuite,
   loadTests,
@@ -24494,7 +24494,7 @@ function createAgentKernel() {
   readTextFile,
   readTranscriptFile,
   readTranscriptJsonl,
-  removeProject,
+  removeBenchmark,
   resolveAndCreateProvider,
   resolveDelegatedTargetDefinition,
   resolveFileReference,
@@ -24516,7 +24516,7 @@ function createAgentKernel() {
   runIsJsonAssertion,
   runRegexAssertion,
   runStartsWithAssertion,
-  saveProjectRegistry,
+  saveBenchmarkRegistry,
   scanRepoDeps,
   scoreToVerdict,
   shouldEnableCache,
@@ -24533,7 +24533,7 @@ function createAgentKernel() {
   toSnakeCaseDeep,
   toTranscriptJsonLine,
   tokensPerTool,
-  touchProject,
+  touchBenchmark,
   transpileEvalYaml,
   transpileEvalYamlFile,
   trimBaselineResult