@agentv/core 4.28.0 → 4.29.0-next.1

package/dist/index.js

@@ -1040,13 +1040,69 @@ async function directPushResults(params) {
   return false;
 }
 
-// src/benchmarks.ts
-import { existsSync as existsSync2, mkdirSync as mkdirSync2, readFileSync as readFileSync3, readdirSync, statSync, writeFileSync as writeFileSync2 } from "node:fs";
+// src/projects.ts
+import {
+  existsSync as existsSync2,
+  mkdirSync as mkdirSync2,
+  readFileSync as readFileSync3,
+  readdirSync,
+  renameSync,
+  statSync,
+  unlinkSync,
+  writeFileSync as writeFileSync2
+} from "node:fs";
 import path5 from "node:path";
 import { stringify as stringifyYaml } from "yaml";
-function getBenchmarksRegistryPath() {
+function getProjectsRegistryPath() {
+  return path5.join(getAgentvConfigDir(), "projects.yaml");
+}
+function getLegacyBenchmarksRegistryPath() {
   return path5.join(getAgentvConfigDir(), "benchmarks.yaml");
 }
+function migrateLegacyBenchmarksFile() {
+  const newPath = getProjectsRegistryPath();
+  const oldPath = getLegacyBenchmarksRegistryPath();
+  const newExists = existsSync2(newPath);
+  const oldExists = existsSync2(oldPath);
+  if (!oldExists) return;
+  if (newExists) {
+    console.warn(
+      `[agentv] Both ${oldPath} and ${newPath} exist. Using ${path5.basename(newPath)}; delete ${path5.basename(oldPath)} when you've confirmed the new file is correct.`
+    );
+    return;
+  }
+  let parsed = null;
+  try {
+    const raw = readFileSync3(oldPath, "utf-8");
+    parsed = parseYamlValue(raw);
+  } catch (err) {
+    console.warn(
+      `[agentv] Failed to read legacy ${path5.basename(oldPath)} for migration: ${err.message}. Leaving the file in place; you may need to migrate it manually.`
+    );
+    return;
+  }
+  const entries = parsed && typeof parsed === "object" && Array.isArray(parsed.benchmarks) ? parsed.benchmarks : [];
+  const newContent = stringifyYaml({ projects: entries });
+  const tempPath = `${newPath}.migrating`;
+  try {
+    mkdirSync2(path5.dirname(newPath), { recursive: true });
+    writeFileSync2(tempPath, newContent, "utf-8");
+    renameSync(tempPath, newPath);
+    unlinkSync(oldPath);
+  } catch (err) {
+    try {
+      if (existsSync2(tempPath)) unlinkSync(tempPath);
+    } catch {
+    }
+    console.warn(
+      `[agentv] Failed to migrate ${path5.basename(oldPath)} \u2192 ${path5.basename(newPath)}: ${err.message}. Legacy file left in place.`
+    );
+    return;
+  }
+  console.log(
+    `[agentv] Migrated registry: ${path5.basename(oldPath)} \u2192 ${path5.basename(newPath)} (${entries.length} entr${entries.length === 1 ? "y" : "ies"})`
+  );
+}
 function fromYaml(raw) {
   if (!raw || typeof raw !== "object") return null;
   const e = raw;
@@ -1081,36 +1137,37 @@ function toYaml(entry) {
   }
   return yaml;
 }
-function loadBenchmarkRegistry() {
-  const registryPath = getBenchmarksRegistryPath();
+function loadProjectRegistry() {
+  migrateLegacyBenchmarksFile();
+  const registryPath = getProjectsRegistryPath();
   if (!existsSync2(registryPath)) {
-    return { benchmarks: [] };
+    return { projects: [] };
   }
   try {
     const raw = readFileSync3(registryPath, "utf-8");
     const parsed = parseYamlValue(raw);
     if (!parsed || typeof parsed !== "object") {
-      return { benchmarks: [] };
+      return { projects: [] };
     }
     const env = process.env;
-    const benchmarks = Array.isArray(parsed.benchmarks) ? parsed.benchmarks.map((e) => fromYaml(interpolateEnv(e, env))).filter((e) => e !== null) : [];
-    return { benchmarks };
+    const projects = Array.isArray(parsed.projects) ? parsed.projects.map((e) => fromYaml(interpolateEnv(e, env))).filter((e) => e !== null) : [];
+    return { projects };
   } catch {
-    return { benchmarks: [] };
+    return { projects: [] };
   }
 }
-function saveBenchmarkRegistry(registry) {
-  const registryPath = getBenchmarksRegistryPath();
+function saveProjectRegistry(registry) {
+  const registryPath = getProjectsRegistryPath();
   const dir = path5.dirname(registryPath);
   if (!existsSync2(dir)) {
     mkdirSync2(dir, { recursive: true });
   }
-  const payload = { benchmarks: registry.benchmarks.map(toYaml) };
+  const payload = { projects: registry.projects.map(toYaml) };
   writeFileSync2(registryPath, stringifyYaml(payload), "utf-8");
 }
-function deriveBenchmarkId(dirPath, existingIds) {
+function deriveProjectId(dirPath, existingIds) {
   const base = path5.basename(dirPath).toLowerCase().replace(/[^a-z0-9-]/g, "-").replace(/-+/g, "-").replace(/^-|-$/g, "");
-  let candidate = base || "benchmark";
+  let candidate = base || "project";
   let suffix = 2;
   while (existingIds.includes(candidate)) {
     candidate = `${base}-${suffix}`;
@@ -1118,54 +1175,54 @@ function deriveBenchmarkId(dirPath, existingIds) {
   }
   return candidate;
 }
-function addBenchmark(benchmarkPath) {
-  const absPath = path5.resolve(benchmarkPath);
+function addProject(projectPath) {
+  const absPath = path5.resolve(projectPath);
   if (!existsSync2(absPath)) {
     throw new Error(`Directory not found: ${absPath}`);
   }
   if (!existsSync2(path5.join(absPath, ".agentv"))) {
     throw new Error(`No .agentv/ directory found in ${absPath}. Run an evaluation first.`);
   }
-  const registry = loadBenchmarkRegistry();
-  const existing = registry.benchmarks.find((p) => p.path === absPath);
+  const registry = loadProjectRegistry();
+  const existing = registry.projects.find((p) => p.path === absPath);
   if (existing) {
     return existing;
   }
   const now = (/* @__PURE__ */ new Date()).toISOString();
   const entry = {
-    id: deriveBenchmarkId(
+    id: deriveProjectId(
       absPath,
-      registry.benchmarks.map((p) => p.id)
+      registry.projects.map((p) => p.id)
     ),
     name: path5.basename(absPath),
     path: absPath,
     addedAt: now,
     lastOpenedAt: now
   };
-  registry.benchmarks.push(entry);
-  saveBenchmarkRegistry(registry);
+  registry.projects.push(entry);
+  saveProjectRegistry(registry);
   return entry;
 }
-function removeBenchmark(benchmarkId) {
-  const registry = loadBenchmarkRegistry();
-  const idx = registry.benchmarks.findIndex((p) => p.id === benchmarkId);
+function removeProject(projectId) {
+  const registry = loadProjectRegistry();
+  const idx = registry.projects.findIndex((p) => p.id === projectId);
   if (idx < 0) return false;
-  registry.benchmarks.splice(idx, 1);
-  saveBenchmarkRegistry(registry);
+  registry.projects.splice(idx, 1);
+  saveProjectRegistry(registry);
   return true;
 }
-function getBenchmark(benchmarkId) {
-  return loadBenchmarkRegistry().benchmarks.find((p) => p.id === benchmarkId);
+function getProject(projectId) {
+  return loadProjectRegistry().projects.find((p) => p.id === projectId);
 }
-function touchBenchmark(benchmarkId) {
-  const registry = loadBenchmarkRegistry();
-  const entry = registry.benchmarks.find((p) => p.id === benchmarkId);
+function touchProject(projectId) {
+  const registry = loadProjectRegistry();
+  const entry = registry.projects.find((p) => p.id === projectId);
   if (entry) {
     entry.lastOpenedAt = (/* @__PURE__ */ new Date()).toISOString();
-    saveBenchmarkRegistry(registry);
+    saveProjectRegistry(registry);
   }
 }
-function discoverBenchmarks(rootDir, maxDepth = 2) {
+function discoverProjects(rootDir, maxDepth = 2) {
   const absRoot = path5.resolve(rootDir);
   if (!existsSync2(absRoot) || !statSync(absRoot).isDirectory()) {
     return [];
@@ -1192,12 +1249,12 @@ function discoverBenchmarks(rootDir, maxDepth = 2) {
   return results.sort();
 }
 
-// src/benchmark-sync.ts
+// src/project-sync.ts
 import * as childProcess from "node:child_process";
 import { existsSync as existsSync3 } from "node:fs";
-async function syncBenchmark(entry) {
+async function syncProject(entry) {
   if (!entry.source) {
-    throw new Error(`Benchmark '${entry.id}' has no source defined`);
+    throw new Error(`Project '${entry.id}' has no source defined`);
   }
   const { url, ref } = entry.source;
   const dest = entry.path;
@@ -1211,12 +1268,12 @@ async function syncBenchmark(entry) {
     );
   }
 }
-async function syncBenchmarks(entries) {
+async function syncProjects(entries) {
   for (const entry of entries) {
     if (!entry.source) continue;
-    console.log(`Syncing benchmark '${entry.id}' from ${entry.source.url}...`);
-    await syncBenchmark(entry);
-    console.log(`Benchmark '${entry.id}' synced.`);
+    console.log(`Syncing project '${entry.id}' from ${entry.source.url}...`);
+    await syncProject(entry);
+    console.log(`Project '${entry.id}' synced.`);
   }
 }
 
@@ -2511,7 +2568,7 @@ export {
   TranscriptProvider,
   WorkspaceCreationError,
   WorkspacePoolManager,
-  addBenchmark,
+  addProject,
   assembleLlmGraderPrompt,
   avgToolDurationMs,
   buildDirectoryChain,
@@ -2542,17 +2599,17 @@ export {
   createTempWorkspace,
   deepEqual,
   defineConfig,
-  deriveBenchmarkId,
   deriveCategory,
+  deriveProjectId,
   detectFormat,
   directPushResults,
   directorySizeBytes,
   discoverAssertions,
-  discoverBenchmarks,
   discoverClaudeSessions,
   discoverCodexSessions,
   discoverCopilotSessions,
   discoverGraders,
+  discoverProjects,
   discoverProviders,
   ensureResultsRepoClone,
   ensureVSCodeSubagents,
@@ -2579,9 +2636,9 @@ export {
   generateRubrics,
   getAgentvConfigDir,
   getAgentvHome,
-  getBenchmark,
-  getBenchmarksRegistryPath,
   getOutputFilenames,
+  getProject,
+  getProjectsRegistryPath,
   getResultsRepoCachePaths,
   getResultsRepoStatus,
   getSubagentsRoot,
@@ -2603,11 +2660,11 @@ export {
   isTestMessageRole,
   killAllTrackedChildren,
   listTargetNames,
-  loadBenchmarkRegistry,
   loadConfig,
   loadEvalCaseById,
   loadEvalCases,
   loadEvalSuite,
+  loadProjectRegistry,
   loadTestById,
   loadTestSuite,
   loadTests,
@@ -2633,7 +2690,7 @@ export {
   readTextFile,
   readTranscriptFile,
   readTranscriptJsonl,
-  removeBenchmark,
+  removeProject,
   resolveAndCreateProvider,
   resolveDelegatedTargetDefinition,
   resolveFileReference,
@@ -2656,7 +2713,7 @@ export {
   runIsJsonAssertion,
   runRegexAssertion,
   runStartsWithAssertion,
-  saveBenchmarkRegistry,
+  saveProjectRegistry,
   scanRepoDeps,
   scoreRangeEvaluationSchema,
   scoreToVerdict,
@@ -2669,14 +2726,14 @@ export {
   subscribeToCopilotSdkLogEntries,
   subscribeToPiLogEntries,
   substituteVariables,
-  syncBenchmark,
-  syncBenchmarks,
+  syncProject,
+  syncProjects,
   syncResultsRepo,
   toCamelCaseDeep,
   toSnakeCaseDeep,
   toTranscriptJsonLines,
   tokensPerTool,
-  touchBenchmark,
+  touchProject,
   trackChild,
   trackedChildCount,
   transpileEvalYaml,