agentv 3.12.0 → 3.13.1

package/dist/{chunk-2ELQ6F3C.js → chunk-UK7UMQOX.js}

@@ -1,8 +1,8 @@
 import { createRequire } from 'node:module'; const require = createRequire(import.meta.url);
 import {
   HtmlWriter,
-  LEGACY_RESULTS_FILENAME,
   RESULT_INDEX_FILENAME,
+  RESULT_RUNS_DIRNAME,
   detectFileType,
   findRepoRoot,
   loadLightweightResults,
@@ -22,7 +22,7 @@ import {
   validateFileReferences,
   validateTargetsFile,
   writeArtifactsFromResults
-} from "./chunk-VLOFRXH4.js";
+} from "./chunk-LSXO22CF.js";
 import {
   createBuiltinRegistry,
   executeScript,
@@ -39,7 +39,7 @@ import {
   toSnakeCaseDeep as toSnakeCaseDeep2,
   transpileEvalYamlFile,
   trimBaselineResult
-} from "./chunk-UYBLUYHN.js";
+} from "./chunk-K747KGDP.js";
 import {
   __commonJS,
   __esm,
@@ -3389,7 +3389,7 @@ function convertEvalsJsonToYaml(inputPath) {
       for (const assertion of test.assertions) {
         lines.push(`      - name: ${assertion.name}`);
         lines.push(`        type: ${assertion.type}`);
-        if ((assertion.type === "llm-grader" || assertion.type === "llm-judge") && "prompt" in assertion) {
+        if (assertion.type === "llm-grader" && "prompt" in assertion) {
           const prompt = assertion.prompt;
           lines.push(`        prompt: "${prompt.replace(/"/g, '\\"')}"`);
         }
@@ -3746,10 +3746,10 @@ async function getPromptEvalGradingBrief(evalPath, testId) {
           if (item.outcome) criteria.push(item.outcome);
         }
       }
-    } else if (type === "llm-grader" || type === "llm_grader" || type === "llm-judge" || type === "llm_judge") {
+    } else if (type === "llm-grader" || type === "llm_grader") {
       const prompt = entry.prompt ?? bag.prompt ?? bag.criteria;
       criteria.push(`[llm-grader] ${typeof prompt === "string" ? prompt : ""}`);
-    } else if (type === "code-grader" || type === "code_grader" || type === "code-judge" || type === "code_judge") {
+    } else if (type === "code-grader" || type === "code_grader") {
       const name = entry.name ?? type;
       const desc = bag.description ?? entry.description;
       criteria.push(`[code-grader] ${name}${desc ? `: ${desc}` : ""}`);
@@ -4126,11 +4126,6 @@ var evalRunCommand = command({
       long: "otel-file",
       description: "Write OTLP JSON trace to file (importable by OTel backends)"
     }),
-    traceFile: option({
-      type: optional(string),
-      long: "trace-file",
-      description: "Write human-readable trace JSONL to file"
-    }),
     exportOtel: flag({
       long: "export-otel",
       description: "Export evaluation traces via OTLP/HTTP to configured endpoint"
@@ -4185,7 +4180,7 @@ var evalRunCommand = command({
   },
   handler: async (args) => {
     if (args.evalPaths.length === 0 && process.stdin.isTTY) {
-      const { launchInteractiveWizard } = await import("./interactive-5X62YEEX.js");
+      const { launchInteractiveWizard } = await import("./interactive-76ZJVPI7.js");
       await launchInteractiveWizard();
       return;
     }
@@ -4211,7 +4206,6 @@ var evalRunCommand = command({
       workspacePath: args.workspacePath,
       trace: false,
       otelFile: args.otelFile,
-      traceFile: args.traceFile,
       exportOtel: args.exportOtel,
       otelBackend: args.otelBackend,
       otelCaptureContent: args.otelCaptureContent,
@@ -4767,7 +4761,7 @@ async function writeGraderConfigs(testDir, assertions, evalDir) {
   let hasCodeGraders = false;
   let hasLlmGraders = false;
   for (const assertion of assertions) {
-    if (assertion.type === "code-grader" || assertion.type === "code-judge") {
+    if (assertion.type === "code-grader") {
       if (!hasCodeGraders) {
         await mkdir3(codeGradersDir, { recursive: true });
         hasCodeGraders = true;
@@ -4780,7 +4774,7 @@ async function writeGraderConfigs(testDir, assertions, evalDir) {
         weight: config.weight ?? 1,
         config: config.config ?? {}
       });
-    } else if (assertion.type === "llm-grader" || assertion.type === "llm-judge") {
+    } else if (assertion.type === "llm-grader") {
       if (!hasLlmGraders) {
         await mkdir3(llmGradersDir, { recursive: true });
         hasLlmGraders = true;
@@ -4866,12 +4860,6 @@ function loadResultFile(filePath) {
   return loadJsonlRecords(resolvedFilePath);
 }
 function resolveTraceResultPath(filePath) {
-  if (path6.basename(filePath) === LEGACY_RESULTS_FILENAME) {
-    return filePath;
-  }
-  if (!filePath.endsWith(".jsonl") && !filePath.endsWith(".json")) {
-    return resolveWorkspaceOrFilePath(filePath);
-  }
   return resolveWorkspaceOrFilePath(filePath);
 }
 function loadJsonlRecords(filePath) {
@@ -4945,7 +4933,9 @@ function loadOtlpTraceFile(filePath) {
     }
   }
   const roots = spans.filter((span) => !span.parentSpanId || !spanMap.has(span.parentSpanId));
-  return roots.map((root, index) => {
+  const supportedRoots = roots.filter(isAgentvEvalRoot);
+  const candidateRoots = supportedRoots.length > 0 ? supportedRoots : roots;
+  return candidateRoots.map((root, index) => {
     const descendants = collectChildSpans(root.spanId, childMap);
     const rootAttrs = parseOtlpAttributes(root.attributes);
     const parsedDescendants = descendants.map((span) => ({
@@ -5032,18 +5022,24 @@ function loadOtlpTraceFile(filePath) {
       } : void 0,
       spans: traceSummary?.spans,
       output: stringAttr(rootAttrs.agentv_output_text),
-      scores: root.events?.filter((event) => event.name?.startsWith("agentv.evaluator.")).map((event) => {
+      scores: root.events?.filter(
+        (event) => event.name?.startsWith("agentv.grader.") || event.name?.startsWith("agentv.evaluator.")
+      ).map((event) => {
         const attrs = parseOtlpAttributes(event.attributes);
-        const name = event.name?.replace(/^agentv\.evaluator\./, "") ?? "unknown";
+        const name = event.name?.replace(/^agentv\.grader\./, "").replace(/^agentv\.evaluator\./, "") ?? "unknown";
         return {
           name,
-          type: stringAttr(attrs.agentv_evaluator_type) ?? "unknown",
-          score: numberAttr(attrs.agentv_evaluator_score) ?? 0
+          type: stringAttr(attrs.agentv_grader_type) ?? stringAttr(attrs.agentv_evaluator_type) ?? "unknown",
+          score: numberAttr(attrs.agentv_grader_score) ?? numberAttr(attrs.agentv_evaluator_score) ?? 0
         };
       })
     };
   });
 }
+function isAgentvEvalRoot(span) {
+  const attrs = parseOtlpAttributes(span.attributes);
+  return span.name === "agentv.eval" || numberAttr(attrs.agentv_score) !== void 0 || typeof stringAttr(attrs.agentv_test_id) === "string";
+}
 function collectChildSpans(spanId, childMap) {
   if (!spanId) return [];
   const direct = childMap.get(spanId) ?? [];
@@ -5138,13 +5134,13 @@ function toTraceSummary(result) {
 }
 function listResultFiles(cwd, limit) {
   const baseDir = path6.join(cwd, ".agentv", "results");
-  const rawDir = path6.join(baseDir, "raw");
+  const runsDir = path6.join(baseDir, RESULT_RUNS_DIRNAME);
   const files = [];
   try {
-    const entries2 = readdirSync2(rawDir, { withFileTypes: true });
+    const entries2 = readdirSync2(runsDir, { withFileTypes: true });
     for (const entry of entries2) {
       if (entry.isDirectory()) {
-        const primaryPath = resolveExistingRunPrimaryPath(path6.join(rawDir, entry.name));
+        const primaryPath = resolveExistingRunPrimaryPath(path6.join(runsDir, entry.name));
         if (primaryPath) {
           files.push({ filePath: primaryPath, displayName: entry.name });
         }
@@ -5152,7 +5148,7 @@ function listResultFiles(cwd, limit) {
     }
     for (const entry of entries2) {
       if (!entry.isDirectory() && entry.name.endsWith(".jsonl")) {
-        files.push({ filePath: path6.join(rawDir, entry.name), displayName: entry.name });
+        files.push({ filePath: path6.join(runsDir, entry.name), displayName: entry.name });
       }
     }
   } catch {
@@ -5317,8 +5313,7 @@ var resultsExportCommand = command({
       const { results } = await loadResults(source, cwd);
       const outputDir = out ? path7.isAbsolute(out) ? out : path7.resolve(cwd, out) : deriveOutputDir(cwd, sourceFile);
       await writeArtifactsFromResults(results, outputDir, {
-        evalFile: sourceFile,
-        writeLegacyResults: false
+        evalFile: sourceFile
       });
       console.log(`Exported ${results.length} test(s) to ${outputDir}`);
       for (const result of results) {
@@ -6640,7 +6635,7 @@ var traceScoreCommand = command({
       );
       if (!hasTrace) {
         console.error(
-          `${c2.red}Error:${c2.reset} Source lacks trace metrics. Export a trace file with ${c2.bold}--trace-file${c2.reset} or ${c2.bold}--otel-file${c2.reset}.`
+          `${c2.red}Error:${c2.reset} Source lacks trace metrics. Use an OTLP trace export via ${c2.bold}--otel-file${c2.reset} or a run manifest with summary metrics in ${c2.bold}index.jsonl${c2.reset}.`
         );
         process.exit(1);
       }
@@ -7761,4 +7756,4 @@ export {
   preprocessArgv,
   runCli
 };
-//# sourceMappingURL=chunk-2ELQ6F3C.js.map
+//# sourceMappingURL=chunk-UK7UMQOX.js.map