@agentv/core 4.3.4 → 4.4.0

package/dist/index.cjs

@@ -1358,13 +1358,13 @@ function serializeAttributeValue(value) {
   if (Array.isArray(value)) return { arrayValue: { values: value.map(serializeAttributeValue) } };
   return { stringValue: String(value) };
 }
-var import_promises35, import_node_path51, OtlpJsonFileExporter;
+var import_promises35, import_node_path52, OtlpJsonFileExporter;
 var init_otlp_json_file_exporter = __esm({
   "src/observability/otlp-json-file-exporter.ts"() {
     "use strict";
     init_cjs_shims();
     import_promises35 = require("fs/promises");
-    import_node_path51 = require("path");
+    import_node_path52 = require("path");
     OtlpJsonFileExporter = class {
       // biome-ignore lint/suspicious/noExplicitAny: serialized span data
       spans = [];
@@ -1403,7 +1403,7 @@ var init_otlp_json_file_exporter = __esm({
       }
       async flush() {
         if (this.spans.length === 0) return;
-        await (0, import_promises35.mkdir)((0, import_node_path51.dirname)(this.filePath), { recursive: true });
+        await (0, import_promises35.mkdir)((0, import_node_path52.dirname)(this.filePath), { recursive: true });
         const otlpJson = {
           resourceSpans: [
             {
@@ -1458,6 +1458,7 @@ __export(index_exports, {
   ToolTrajectoryEvaluator: () => ToolTrajectoryEvaluator,
   WorkspaceCreationError: () => WorkspaceCreationError,
   WorkspacePoolManager: () => WorkspacePoolManager,
+  addProject: () => addProject,
   assembleLlmGraderPrompt: () => assembleLlmGraderPrompt,
   assembleLlmJudgePrompt: () => assembleLlmGraderPrompt,
   avgToolDurationMs: () => avgToolDurationMs,
@@ -1487,11 +1488,13 @@ __export(index_exports, {
   deepEqual: () => deepEqual,
   defineConfig: () => defineConfig,
   deriveCategory: () => deriveCategory,
+  deriveProjectId: () => deriveProjectId,
   detectFormat: () => detectFormat,
   discoverAssertions: () => discoverAssertions,
   discoverCopilotSessions: () => discoverCopilotSessions,
   discoverGraders: () => discoverGraders,
   discoverJudges: () => discoverGraders,
+  discoverProjects: () => discoverProjects,
   discoverProviders: () => discoverProviders,
   ensureVSCodeSubagents: () => ensureVSCodeSubagents,
   evaluate: () => evaluate,
@@ -1515,6 +1518,8 @@ __export(index_exports, {
   generateRubrics: () => generateRubrics,
   getAgentvHome: () => getAgentvHome,
   getOutputFilenames: () => getOutputFilenames,
+  getProject: () => getProject,
+  getProjectsRegistryPath: () => getProjectsRegistryPath,
   getSubagentsRoot: () => getSubagentsRoot,
   getTextContent: () => getTextContent,
   getTraceStateRoot: () => getTraceStateRoot,
@@ -1536,6 +1541,7 @@ __export(index_exports, {
   loadEvalCaseById: () => loadEvalCaseById,
   loadEvalCases: () => loadEvalCases,
   loadEvalSuite: () => loadEvalSuite,
+  loadProjectRegistry: () => loadProjectRegistry,
   loadTestById: () => loadTestById,
   loadTestSuite: () => loadTestSuite,
   loadTests: () => loadTests,
@@ -1550,6 +1556,7 @@ __export(index_exports, {
   readTargetDefinitions: () => readTargetDefinitions,
   readTestSuiteMetadata: () => readTestSuiteMetadata,
   readTextFile: () => readTextFile,
+  removeProject: () => removeProject,
   resolveAndCreateProvider: () => resolveAndCreateProvider,
   resolveFileReference: () => resolveFileReference3,
   resolveTargetDefinition: () => resolveTargetDefinition,
@@ -1568,6 +1575,7 @@ __export(index_exports, {
   runIsJsonAssertion: () => runIsJsonAssertion,
   runRegexAssertion: () => runRegexAssertion,
   runStartsWithAssertion: () => runStartsWithAssertion,
+  saveProjectRegistry: () => saveProjectRegistry,
   scoreToVerdict: () => scoreToVerdict,
   shouldEnableCache: () => shouldEnableCache,
   shouldSkipCacheForTemperature: () => shouldSkipCacheForTemperature,
@@ -1580,6 +1588,7 @@ __export(index_exports, {
   toCamelCaseDeep: () => toCamelCaseDeep,
   toSnakeCaseDeep: () => toSnakeCaseDeep,
   tokensPerTool: () => tokensPerTool,
+  touchProject: () => touchProject,
   transpileEvalYaml: () => transpileEvalYaml,
   transpileEvalYamlFile: () => transpileEvalYamlFile,
   trimBaselineResult: () => trimBaselineResult
@@ -11556,8 +11565,8 @@ function resolveCliConfig(target, env, evalFilePath) {
   const parseResult = CliTargetInputSchema.safeParse(target, { errorMap: cliErrorMap });
   if (!parseResult.success) {
     const firstError = parseResult.error.errors[0];
-    const path50 = firstError?.path.join(".") || "";
-    const prefix = path50 ? `${target.name} ${path50}: ` : `${target.name}: `;
+    const path51 = firstError?.path.join(".") || "";
+    const prefix = path51 ? `${target.name} ${path51}: ` : `${target.name}: `;
     throw new Error(`${prefix}${firstError?.message}`);
   }
   const normalized = normalizeCliTargetInput(parseResult.data, env, evalFilePath);
@@ -13622,13 +13631,13 @@ async function execFileWithStdinNode(argv, stdinPayload, options) {
 async function execShellWithStdin(command, stdinPayload, options = {}) {
   const { mkdir: mkdir17, readFile: readFile16, rm: rm6, writeFile: writeFile9 } = await import("fs/promises");
   const { tmpdir: tmpdir3 } = await import("os");
-  const path50 = await import("path");
+  const path51 = await import("path");
   const { randomUUID: randomUUID10 } = await import("crypto");
-  const dir = path50.join(tmpdir3(), `agentv-exec-${randomUUID10()}`);
+  const dir = path51.join(tmpdir3(), `agentv-exec-${randomUUID10()}`);
   await mkdir17(dir, { recursive: true });
-  const stdinPath = path50.join(dir, "stdin.txt");
-  const stdoutPath = path50.join(dir, "stdout.txt");
-  const stderrPath = path50.join(dir, "stderr.txt");
+  const stdinPath = path51.join(dir, "stdin.txt");
+  const stdoutPath = path51.join(dir, "stdout.txt");
+  const stderrPath = path51.join(dir, "stderr.txt");
   await writeFile9(stdinPath, stdinPayload, "utf8");
   const wrappedCommand = process.platform === "win32" ? `(${command}) < ${shellEscapePath(stdinPath)} > ${shellEscapePath(stdoutPath)} 2> ${shellEscapePath(stderrPath)}` : `(${command}) < ${shellEscapePath(stdinPath)} > ${shellEscapePath(stdoutPath)} 2> ${shellEscapePath(stderrPath)}`;
   const { spawn: spawn5 } = await import("child_process");
@@ -15848,115 +15857,115 @@ var FieldAccuracyEvaluator = class {
    * Evaluate a single field against the expected value.
    */
   evaluateField(fieldConfig, candidateData, expectedData) {
-    const { path: path50, match, required = true, weight = 1 } = fieldConfig;
-    const candidateValue = resolvePath(candidateData, path50);
-    const expectedValue = resolvePath(expectedData, path50);
+    const { path: path51, match, required = true, weight = 1 } = fieldConfig;
+    const candidateValue = resolvePath(candidateData, path51);
+    const expectedValue = resolvePath(expectedData, path51);
     if (expectedValue === void 0) {
       return {
-        path: path50,
+        path: path51,
         score: 1,
         // No expected value means no comparison needed
         weight,
         hit: true,
-        message: `${path50}: no expected value`
+        message: `${path51}: no expected value`
       };
     }
     if (candidateValue === void 0) {
       if (required) {
         return {
-          path: path50,
+          path: path51,
           score: 0,
           weight,
           hit: false,
-          message: `${path50} (required, missing)`
+          message: `${path51} (required, missing)`
         };
       }
       return {
-        path: path50,
+        path: path51,
         score: 1,
         // Don't penalize missing optional fields
         weight: 0,
         // Zero weight means it won't affect the score
         hit: true,
-        message: `${path50}: optional field missing`
+        message: `${path51}: optional field missing`
       };
     }
     switch (match) {
       case "exact":
-        return this.compareExact(path50, candidateValue, expectedValue, weight);
+        return this.compareExact(path51, candidateValue, expectedValue, weight);
       case "numeric_tolerance":
         return this.compareNumericTolerance(
-          path50,
+          path51,
           candidateValue,
           expectedValue,
           fieldConfig,
           weight
         );
       case "date":
-        return this.compareDate(path50, candidateValue, expectedValue, fieldConfig, weight);
+        return this.compareDate(path51, candidateValue, expectedValue, fieldConfig, weight);
       default:
         return {
-          path: path50,
+          path: path51,
           score: 0,
           weight,
           hit: false,
-          message: `${path50}: unknown match type "${match}"`
+          message: `${path51}: unknown match type "${match}"`
         };
     }
   }
   /**
    * Exact equality comparison.
    */
-  compareExact(path50, candidateValue, expectedValue, weight) {
+  compareExact(path51, candidateValue, expectedValue, weight) {
     if (deepEqual(candidateValue, expectedValue)) {
       return {
-        path: path50,
+        path: path51,
         score: 1,
         weight,
         hit: true,
-        message: path50
+        message: path51
       };
     }
     if (typeof candidateValue !== typeof expectedValue) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
+        message: `${path51} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
       };
     }
     return {
-      path: path50,
+      path: path51,
       score: 0,
       weight,
       hit: false,
-      message: `${path50} (value mismatch)`
+      message: `${path51} (value mismatch)`
     };
   }
   /**
    * Numeric comparison with absolute or relative tolerance.
    */
-  compareNumericTolerance(path50, candidateValue, expectedValue, fieldConfig, weight) {
+  compareNumericTolerance(path51, candidateValue, expectedValue, fieldConfig, weight) {
     const { tolerance = 0, relative = false } = fieldConfig;
     const candidateNum = toNumber(candidateValue);
     const expectedNum = toNumber(expectedValue);
     if (candidateNum === null || expectedNum === null) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (non-numeric value)`
+        message: `${path51} (non-numeric value)`
       };
     }
     if (!Number.isFinite(candidateNum) || !Number.isFinite(expectedNum)) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (invalid numeric value)`
+        message: `${path51} (invalid numeric value)`
       };
     }
     const diff = Math.abs(candidateNum - expectedNum);
@@ -15969,61 +15978,61 @@ var FieldAccuracyEvaluator = class {
     }
     if (withinTolerance) {
       return {
-        path: path50,
+        path: path51,
         score: 1,
         weight,
         hit: true,
-        message: `${path50} (within tolerance: diff=${diff.toFixed(2)})`
+        message: `${path51} (within tolerance: diff=${diff.toFixed(2)})`
       };
     }
     return {
-      path: path50,
+      path: path51,
       score: 0,
       weight,
       hit: false,
-      message: `${path50} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
+      message: `${path51} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
     };
   }
   /**
    * Date comparison with format normalization.
    */
-  compareDate(path50, candidateValue, expectedValue, fieldConfig, weight) {
+  compareDate(path51, candidateValue, expectedValue, fieldConfig, weight) {
     const formats = fieldConfig.formats ?? DEFAULT_DATE_FORMATS;
     const candidateDate = parseDate(String(candidateValue), formats);
     const expectedDate = parseDate(String(expectedValue), formats);
     if (candidateDate === null) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (unparseable candidate date)`
+        message: `${path51} (unparseable candidate date)`
       };
     }
     if (expectedDate === null) {
       return {
-        path: path50,
+        path: path51,
         score: 0,
         weight,
         hit: false,
-        message: `${path50} (unparseable expected date)`
+        message: `${path51} (unparseable expected date)`
       };
     }
     if (candidateDate.getFullYear() === expectedDate.getFullYear() && candidateDate.getMonth() === expectedDate.getMonth() && candidateDate.getDate() === expectedDate.getDate()) {
       return {
-        path: path50,
+        path: path51,
         score: 1,
         weight,
         hit: true,
-        message: path50
+        message: path51
       };
     }
     return {
-      path: path50,
+      path: path51,
       score: 0,
       weight,
       hit: false,
-      message: `${path50} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
+      message: `${path51} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
     };
   }
   /**
@@ -16056,11 +16065,11 @@ var FieldAccuracyEvaluator = class {
     };
   }
 };
-function resolvePath(obj, path50) {
-  if (!path50 || !obj) {
+function resolvePath(obj, path51) {
+  if (!path51 || !obj) {
     return void 0;
   }
-  const parts = path50.split(/\.|\[|\]/).filter((p) => p.length > 0);
+  const parts = path51.split(/\.|\[|\]/).filter((p) => p.length > 0);
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0) {
@@ -16549,8 +16558,8 @@ var TokenUsageEvaluator = class {
 
 // src/evaluation/evaluators/tool-trajectory.ts
 init_cjs_shims();
-function getNestedValue(obj, path50) {
-  const parts = path50.split(".");
+function getNestedValue(obj, path51) {
+  const parts = path51.split(".");
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0 || typeof current !== "object") {
@@ -20817,7 +20826,7 @@ async function discoverDefaultTarget(repoRoot) {
   return null;
 }
 async function loadEnvHierarchy(repoRoot, startPath) {
-  const { readFileSync: readFileSync2 } = await import("fs");
+  const { readFileSync: readFileSync3 } = await import("fs");
   const chain = buildDirectoryChain2(startPath, repoRoot);
   const envFiles = [];
   for (const dir of chain) {
@@ -20826,7 +20835,7 @@ async function loadEnvHierarchy(repoRoot, startPath) {
   }
   for (let i = 0; i < envFiles.length; i++) {
     try {
-      const content = readFileSync2(envFiles[i], "utf8");
+      const content = readFileSync3(envFiles[i], "utf8");
       for (const line of content.split("\n")) {
         const trimmed = line.trim();
         if (!trimmed || trimmed.startsWith("#")) continue;
@@ -20899,12 +20908,12 @@ var CONFIG_FILE_NAMES = [
   ".agentv/config.js"
 ];
 async function loadTsConfig(projectRoot) {
-  const { existsSync: existsSync5 } = await import("fs");
+  const { existsSync: existsSync6 } = await import("fs");
   const { pathToFileURL } = await import("url");
   const { join: join2 } = await import("path");
   for (const fileName of CONFIG_FILE_NAMES) {
     const filePath = join2(projectRoot, fileName);
-    if (!existsSync5(filePath)) {
+    if (!existsSync6(filePath)) {
       continue;
     }
     try {
@@ -21049,6 +21058,122 @@ function shouldSkipCacheForTemperature(targetConfig) {
   return false;
 }
 
+// src/projects.ts
+init_cjs_shims();
+var import_node_fs16 = require("fs");
+var import_node_path51 = __toESM(require("path"), 1);
+var import_yaml8 = require("yaml");
+function getProjectsRegistryPath() {
+  return import_node_path51.default.join(getAgentvHome(), "projects.yaml");
+}
+function loadProjectRegistry() {
+  const registryPath = getProjectsRegistryPath();
+  if (!(0, import_node_fs16.existsSync)(registryPath)) {
+    return { projects: [] };
+  }
+  try {
+    const raw = (0, import_node_fs16.readFileSync)(registryPath, "utf-8");
+    const parsed = (0, import_yaml8.parse)(raw);
+    if (!parsed || !Array.isArray(parsed.projects)) {
+      return { projects: [] };
+    }
+    return { projects: parsed.projects };
+  } catch {
+    return { projects: [] };
+  }
+}
+function saveProjectRegistry(registry) {
+  const registryPath = getProjectsRegistryPath();
+  const dir = import_node_path51.default.dirname(registryPath);
+  if (!(0, import_node_fs16.existsSync)(dir)) {
+    (0, import_node_fs16.mkdirSync)(dir, { recursive: true });
+  }
+  (0, import_node_fs16.writeFileSync)(registryPath, (0, import_yaml8.stringify)(registry), "utf-8");
+}
+function deriveProjectId(dirPath, existingIds) {
+  const base = import_node_path51.default.basename(dirPath).toLowerCase().replace(/[^a-z0-9-]/g, "-").replace(/-+/g, "-").replace(/^-|-$/g, "");
+  let candidate = base || "project";
+  let suffix = 2;
+  while (existingIds.includes(candidate)) {
+    candidate = `${base}-${suffix}`;
+    suffix++;
+  }
+  return candidate;
+}
+function addProject(projectPath) {
+  const absPath = import_node_path51.default.resolve(projectPath);
+  if (!(0, import_node_fs16.existsSync)(absPath)) {
+    throw new Error(`Directory not found: ${absPath}`);
+  }
+  if (!(0, import_node_fs16.existsSync)(import_node_path51.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);
+  if (existing) {
+    return existing;
+  }
+  const now = (/* @__PURE__ */ new Date()).toISOString();
+  const entry = {
+    id: deriveProjectId(
+      absPath,
+      registry.projects.map((p) => p.id)
+    ),
+    name: import_node_path51.default.basename(absPath),
+    path: absPath,
+    addedAt: now,
+    lastOpenedAt: now
+  };
+  registry.projects.push(entry);
+  saveProjectRegistry(registry);
+  return entry;
+}
+function removeProject(projectId) {
+  const registry = loadProjectRegistry();
+  const idx = registry.projects.findIndex((p) => p.id === projectId);
+  if (idx < 0) return false;
+  registry.projects.splice(idx, 1);
+  saveProjectRegistry(registry);
+  return true;
+}
+function getProject(projectId) {
+  return loadProjectRegistry().projects.find((p) => p.id === projectId);
+}
+function touchProject(projectId) {
+  const registry = loadProjectRegistry();
+  const entry = registry.projects.find((p) => p.id === projectId);
+  if (entry) {
+    entry.lastOpenedAt = (/* @__PURE__ */ new Date()).toISOString();
+    saveProjectRegistry(registry);
+  }
+}
+function discoverProjects(rootDir, maxDepth = 2) {
+  const absRoot = import_node_path51.default.resolve(rootDir);
+  if (!(0, import_node_fs16.existsSync)(absRoot) || !(0, import_node_fs16.statSync)(absRoot).isDirectory()) {
+    return [];
+  }
+  const results = [];
+  function scan(dir, depth) {
+    if (depth > maxDepth) return;
+    if ((0, import_node_fs16.existsSync)(import_node_path51.default.join(dir, ".agentv"))) {
+      results.push(dir);
+      return;
+    }
+    if (depth === maxDepth) return;
+    try {
+      const entries = (0, import_node_fs16.readdirSync)(dir, { withFileTypes: true });
+      for (const entry of entries) {
+        if (!entry.isDirectory()) continue;
+        if (entry.name.startsWith(".") || entry.name === "node_modules") continue;
+        scan(import_node_path51.default.join(dir, entry.name), depth + 1);
+      }
+    } catch {
+    }
+  }
+  scan(absRoot, 0);
+  return results;
+}
+
 // src/evaluation/baseline.ts
 init_cjs_shims();
 var STRIPPED_TOP_LEVEL_FIELDS = /* @__PURE__ */ new Set([
@@ -21649,6 +21774,7 @@ function createAgentKernel() {
   ToolTrajectoryEvaluator,
   WorkspaceCreationError,
   WorkspacePoolManager,
+  addProject,
   assembleLlmGraderPrompt,
   assembleLlmJudgePrompt,
   avgToolDurationMs,
@@ -21678,11 +21804,13 @@ function createAgentKernel() {
   deepEqual,
   defineConfig,
   deriveCategory,
+  deriveProjectId,
   detectFormat,
   discoverAssertions,
   discoverCopilotSessions,
   discoverGraders,
   discoverJudges,
+  discoverProjects,
   discoverProviders,
   ensureVSCodeSubagents,
   evaluate,
@@ -21706,6 +21834,8 @@ function createAgentKernel() {
   generateRubrics,
   getAgentvHome,
   getOutputFilenames,
+  getProject,
+  getProjectsRegistryPath,
   getSubagentsRoot,
   getTextContent,
   getTraceStateRoot,
@@ -21727,6 +21857,7 @@ function createAgentKernel() {
   loadEvalCaseById,
   loadEvalCases,
   loadEvalSuite,
+  loadProjectRegistry,
   loadTestById,
   loadTestSuite,
   loadTests,
@@ -21741,6 +21872,7 @@ function createAgentKernel() {
   readTargetDefinitions,
   readTestSuiteMetadata,
   readTextFile,
+  removeProject,
   resolveAndCreateProvider,
   resolveFileReference,
   resolveTargetDefinition,
@@ -21759,6 +21891,7 @@ function createAgentKernel() {
   runIsJsonAssertion,
   runRegexAssertion,
   runStartsWithAssertion,
+  saveProjectRegistry,
   scoreToVerdict,
   shouldEnableCache,
   shouldSkipCacheForTemperature,
@@ -21771,6 +21904,7 @@ function createAgentKernel() {
   toCamelCaseDeep,
   toSnakeCaseDeep,
   tokensPerTool,
+  touchProject,
   transpileEvalYaml,
   transpileEvalYamlFile,
   trimBaselineResult