@agentv/core 4.3.4 → 4.4.0

package/dist/index.js

@@ -10342,13 +10342,13 @@ async function execFileWithStdinNode(argv, stdinPayload, options) {
 async function execShellWithStdin(command, stdinPayload, options = {}) {
   const { mkdir: mkdir16, readFile: readFile14, rm: rm6, writeFile: writeFile9 } = await import("node:fs/promises");
   const { tmpdir: tmpdir3 } = await import("node:os");
-  const path47 = await import("node:path");
+  const path48 = await import("node:path");
   const { randomUUID: randomUUID10 } = await import("node:crypto");
-  const dir = path47.join(tmpdir3(), `agentv-exec-${randomUUID10()}`);
+  const dir = path48.join(tmpdir3(), `agentv-exec-${randomUUID10()}`);
   await mkdir16(dir, { recursive: true });
-  const stdinPath = path47.join(dir, "stdin.txt");
-  const stdoutPath = path47.join(dir, "stdout.txt");
-  const stderrPath = path47.join(dir, "stderr.txt");
+  const stdinPath = path48.join(dir, "stdin.txt");
+  const stdoutPath = path48.join(dir, "stdout.txt");
+  const stderrPath = path48.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> ${shellEscapePath(stderrPath)}`;
   const { spawn: spawn5 } = await import("node:child_process");
@@ -12561,115 +12561,115 @@ var FieldAccuracyEvaluator = class {
    * Evaluate a single field against the expected value.
    */
   evaluateField(fieldConfig, candidateData, expectedData) {
-    const { path: path47, match, required = true, weight = 1 } = fieldConfig;
-    const candidateValue = resolvePath(candidateData, path47);
-    const expectedValue = resolvePath(expectedData, path47);
+    const { path: path48, match, required = true, weight = 1 } = fieldConfig;
+    const candidateValue = resolvePath(candidateData, path48);
+    const expectedValue = resolvePath(expectedData, path48);
     if (expectedValue === void 0) {
       return {
-        path: path47,
+        path: path48,
         score: 1,
         // No expected value means no comparison needed
         weight,
         hit: true,
-        message: `${path47}: no expected value`
+        message: `${path48}: no expected value`
       };
     }
     if (candidateValue === void 0) {
       if (required) {
         return {
-          path: path47,
+          path: path48,
           score: 0,
           weight,
           hit: false,
-          message: `${path47} (required, missing)`
+          message: `${path48} (required, missing)`
         };
       }
       return {
-        path: path47,
+        path: path48,
         score: 1,
         // Don't penalize missing optional fields
         weight: 0,
         // Zero weight means it won't affect the score
         hit: true,
-        message: `${path47}: optional field missing`
+        message: `${path48}: optional field missing`
       };
     }
     switch (match) {
       case "exact":
-        return this.compareExact(path47, candidateValue, expectedValue, weight);
+        return this.compareExact(path48, candidateValue, expectedValue, weight);
       case "numeric_tolerance":
         return this.compareNumericTolerance(
-          path47,
+          path48,
           candidateValue,
           expectedValue,
           fieldConfig,
           weight
         );
       case "date":
-        return this.compareDate(path47, candidateValue, expectedValue, fieldConfig, weight);
+        return this.compareDate(path48, candidateValue, expectedValue, fieldConfig, weight);
       default:
         return {
-          path: path47,
+          path: path48,
           score: 0,
           weight,
           hit: false,
-          message: `${path47}: unknown match type "${match}"`
+          message: `${path48}: unknown match type "${match}"`
         };
     }
   }
   /**
    * Exact equality comparison.
    */
-  compareExact(path47, candidateValue, expectedValue, weight) {
+  compareExact(path48, candidateValue, expectedValue, weight) {
     if (deepEqual(candidateValue, expectedValue)) {
       return {
-        path: path47,
+        path: path48,
         score: 1,
         weight,
         hit: true,
-        message: path47
+        message: path48
       };
     }
     if (typeof candidateValue !== typeof expectedValue) {
       return {
-        path: path47,
+        path: path48,
         score: 0,
         weight,
         hit: false,
-        message: `${path47} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
+        message: `${path48} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
       };
     }
     return {
-      path: path47,
+      path: path48,
       score: 0,
       weight,
       hit: false,
-      message: `${path47} (value mismatch)`
+      message: `${path48} (value mismatch)`
     };
   }
   /**
    * Numeric comparison with absolute or relative tolerance.
    */
-  compareNumericTolerance(path47, candidateValue, expectedValue, fieldConfig, weight) {
+  compareNumericTolerance(path48, candidateValue, expectedValue, fieldConfig, weight) {
     const { tolerance = 0, relative = false } = fieldConfig;
     const candidateNum = toNumber(candidateValue);
     const expectedNum = toNumber(expectedValue);
     if (candidateNum === null || expectedNum === null) {
       return {
-        path: path47,
+        path: path48,
         score: 0,
         weight,
         hit: false,
-        message: `${path47} (non-numeric value)`
+        message: `${path48} (non-numeric value)`
       };
     }
     if (!Number.isFinite(candidateNum) || !Number.isFinite(expectedNum)) {
       return {
-        path: path47,
+        path: path48,
         score: 0,
         weight,
         hit: false,
-        message: `${path47} (invalid numeric value)`
+        message: `${path48} (invalid numeric value)`
       };
     }
     const diff = Math.abs(candidateNum - expectedNum);
@@ -12682,61 +12682,61 @@ var FieldAccuracyEvaluator = class {
     }
     if (withinTolerance) {
       return {
-        path: path47,
+        path: path48,
         score: 1,
         weight,
         hit: true,
-        message: `${path47} (within tolerance: diff=${diff.toFixed(2)})`
+        message: `${path48} (within tolerance: diff=${diff.toFixed(2)})`
       };
     }
     return {
-      path: path47,
+      path: path48,
       score: 0,
       weight,
       hit: false,
-      message: `${path47} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
+      message: `${path48} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
     };
   }
   /**
    * Date comparison with format normalization.
    */
-  compareDate(path47, candidateValue, expectedValue, fieldConfig, weight) {
+  compareDate(path48, candidateValue, expectedValue, fieldConfig, weight) {
     const formats = fieldConfig.formats ?? DEFAULT_DATE_FORMATS;
     const candidateDate = parseDate(String(candidateValue), formats);
     const expectedDate = parseDate(String(expectedValue), formats);
     if (candidateDate === null) {
       return {
-        path: path47,
+        path: path48,
         score: 0,
         weight,
         hit: false,
-        message: `${path47} (unparseable candidate date)`
+        message: `${path48} (unparseable candidate date)`
       };
     }
     if (expectedDate === null) {
       return {
-        path: path47,
+        path: path48,
         score: 0,
         weight,
         hit: false,
-        message: `${path47} (unparseable expected date)`
+        message: `${path48} (unparseable expected date)`
       };
     }
     if (candidateDate.getFullYear() === expectedDate.getFullYear() && candidateDate.getMonth() === expectedDate.getMonth() && candidateDate.getDate() === expectedDate.getDate()) {
       return {
-        path: path47,
+        path: path48,
         score: 1,
         weight,
         hit: true,
-        message: path47
+        message: path48
       };
     }
     return {
-      path: path47,
+      path: path48,
       score: 0,
       weight,
       hit: false,
-      message: `${path47} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
+      message: `${path48} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
     };
   }
   /**
@@ -12769,11 +12769,11 @@ var FieldAccuracyEvaluator = class {
     };
   }
 };
-function resolvePath(obj, path47) {
-  if (!path47 || !obj) {
+function resolvePath(obj, path48) {
+  if (!path48 || !obj) {
     return void 0;
   }
-  const parts = path47.split(/\.|\[|\]/).filter((p) => p.length > 0);
+  const parts = path48.split(/\.|\[|\]/).filter((p) => p.length > 0);
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0) {
@@ -13257,8 +13257,8 @@ var TokenUsageEvaluator = class {
 };
 
 // src/evaluation/evaluators/tool-trajectory.ts
-function getNestedValue(obj, path47) {
-  const parts = path47.split(".");
+function getNestedValue(obj, path48) {
+  const parts = path48.split(".");
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0 || typeof current !== "object") {
@@ -17499,7 +17499,7 @@ async function discoverDefaultTarget(repoRoot) {
   return null;
 }
 async function loadEnvHierarchy(repoRoot, startPath) {
-  const { readFileSync: readFileSync2 } = await import("node:fs");
+  const { readFileSync: readFileSync3 } = await import("node:fs");
   const chain = buildDirectoryChain(startPath, repoRoot);
   const envFiles = [];
   for (const dir of chain) {
@@ -17508,7 +17508,7 @@ 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;
@@ -17580,12 +17580,12 @@ var CONFIG_FILE_NAMES = [
   ".agentv/config.js"
 ];
 async function loadTsConfig(projectRoot) {
-  const { existsSync: existsSync5 } = await import("node:fs");
+  const { existsSync: existsSync6 } = await import("node:fs");
   const { pathToFileURL } = await import("node:url");
   const { join: join2 } = await import("node:path");
   for (const fileName of CONFIG_FILE_NAMES) {
     const filePath = join2(projectRoot, fileName);
-    if (!existsSync5(filePath)) {
+    if (!existsSync6(filePath)) {
       continue;
     }
     try {
@@ -17722,6 +17722,121 @@ function shouldSkipCacheForTemperature(targetConfig) {
   return false;
 }
 
+// src/projects.ts
+import { existsSync as existsSync5, mkdirSync, readFileSync as readFileSync2, readdirSync as readdirSync3, statSync as statSync2, writeFileSync } from "node:fs";
+import path47 from "node:path";
+import { parse as parseYaml3, stringify as stringifyYaml } from "yaml";
+function getProjectsRegistryPath() {
+  return path47.join(getAgentvHome(), "projects.yaml");
+}
+function loadProjectRegistry() {
+  const registryPath = getProjectsRegistryPath();
+  if (!existsSync5(registryPath)) {
+    return { projects: [] };
+  }
+  try {
+    const raw = readFileSync2(registryPath, "utf-8");
+    const parsed = parseYaml3(raw);
+    if (!parsed || !Array.isArray(parsed.projects)) {
+      return { projects: [] };
+    }
+    return { projects: parsed.projects };
+  } catch {
+    return { projects: [] };
+  }
+}
+function saveProjectRegistry(registry) {
+  const registryPath = getProjectsRegistryPath();
+  const dir = path47.dirname(registryPath);
+  if (!existsSync5(dir)) {
+    mkdirSync(dir, { recursive: true });
+  }
+  writeFileSync(registryPath, stringifyYaml(registry), "utf-8");
+}
+function deriveProjectId(dirPath, existingIds) {
+  const base = path47.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 = path47.resolve(projectPath);
+  if (!existsSync5(absPath)) {
+    throw new Error(`Directory not found: ${absPath}`);
+  }
+  if (!existsSync5(path47.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: path47.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 = path47.resolve(rootDir);
+  if (!existsSync5(absRoot) || !statSync2(absRoot).isDirectory()) {
+    return [];
+  }
+  const results = [];
+  function scan(dir, depth) {
+    if (depth > maxDepth) return;
+    if (existsSync5(path47.join(dir, ".agentv"))) {
+      results.push(dir);
+      return;
+    }
+    if (depth === maxDepth) return;
+    try {
+      const entries = readdirSync3(dir, { withFileTypes: true });
+      for (const entry of entries) {
+        if (!entry.isDirectory()) continue;
+        if (entry.name.startsWith(".") || entry.name === "node_modules") continue;
+        scan(path47.join(dir, entry.name), depth + 1);
+      }
+    } catch {
+    }
+  }
+  scan(absRoot, 0);
+  return results;
+}
+
 // src/evaluation/baseline.ts
 var STRIPPED_TOP_LEVEL_FIELDS = /* @__PURE__ */ new Set([
   "requests",
@@ -18312,6 +18427,7 @@ export {
   ToolTrajectoryEvaluator,
   WorkspaceCreationError,
   WorkspacePoolManager,
+  addProject,
   assembleLlmGraderPrompt,
   assembleLlmGraderPrompt as assembleLlmJudgePrompt,
   avgToolDurationMs,
@@ -18341,11 +18457,13 @@ export {
   deepEqual,
   defineConfig,
   deriveCategory,
+  deriveProjectId,
   detectFormat,
   discoverAssertions,
   discoverCopilotSessions,
   discoverGraders,
   discoverGraders as discoverJudges,
+  discoverProjects,
   discoverProviders,
   ensureVSCodeSubagents,
   evaluate,
@@ -18369,6 +18487,8 @@ export {
   generateRubrics,
   getAgentvHome,
   getOutputFilenames,
+  getProject,
+  getProjectsRegistryPath,
   getSubagentsRoot,
   getTextContent,
   getTraceStateRoot,
@@ -18390,6 +18510,7 @@ export {
   loadEvalCaseById,
   loadEvalCases,
   loadEvalSuite,
+  loadProjectRegistry,
   loadTestById,
   loadTestSuite,
   loadTests,
@@ -18404,6 +18525,7 @@ export {
   readTargetDefinitions,
   readTestSuiteMetadata,
   readTextFile,
+  removeProject,
   resolveAndCreateProvider,
   resolveFileReference,
   resolveTargetDefinition,
@@ -18422,6 +18544,7 @@ export {
   runIsJsonAssertion,
   runRegexAssertion,
   runStartsWithAssertion,
+  saveProjectRegistry,
   scoreToVerdict,
   shouldEnableCache,
   shouldSkipCacheForTemperature,
@@ -18434,6 +18557,7 @@ export {
   toCamelCaseDeep,
   toSnakeCaseDeep,
   tokensPerTool,
+  touchProject,
   transpileEvalYaml,
   transpileEvalYamlFile,
   trimBaselineResult