@agentv/core 3.7.0 → 3.9.0

package/dist/index.cjs

@@ -1315,12 +1315,12 @@ function serializeAttributeValue(value) {
   if (Array.isArray(value)) return { arrayValue: { values: value.map(serializeAttributeValue) } };
   return { stringValue: String(value) };
 }
-var import_promises33, import_node_path49, OtlpJsonFileExporter;
+var import_promises32, import_node_path48, OtlpJsonFileExporter;
 var init_otlp_json_file_exporter = __esm({
   "src/observability/otlp-json-file-exporter.ts"() {
     "use strict";
-    import_promises33 = require("fs/promises");
-    import_node_path49 = require("path");
+    import_promises32 = require("fs/promises");
+    import_node_path48 = require("path");
     OtlpJsonFileExporter = class {
       // biome-ignore lint/suspicious/noExplicitAny: serialized span data
       spans = [];
@@ -1359,7 +1359,7 @@ var init_otlp_json_file_exporter = __esm({
       }
       async flush() {
         if (this.spans.length === 0) return;
-        await (0, import_promises33.mkdir)((0, import_node_path49.dirname)(this.filePath), { recursive: true });
+        await (0, import_promises32.mkdir)((0, import_node_path48.dirname)(this.filePath), { recursive: true });
         const otlpJson = {
           resourceSpans: [
             {
@@ -1390,13 +1390,13 @@ function hrTimeDiffMs(start, end) {
   const diffNano = end[1] - start[1];
   return Math.round(diffSec * 1e3 + diffNano / 1e6);
 }
-var import_node_fs15, import_promises34, import_node_path50, SimpleTraceFileExporter;
+var import_node_fs15, import_promises33, import_node_path49, SimpleTraceFileExporter;
 var init_simple_trace_file_exporter = __esm({
   "src/observability/simple-trace-file-exporter.ts"() {
     "use strict";
     import_node_fs15 = require("fs");
-    import_promises34 = require("fs/promises");
-    import_node_path50 = require("path");
+    import_promises33 = require("fs/promises");
+    import_node_path49 = require("path");
     SimpleTraceFileExporter = class {
       stream = null;
       filePath;
@@ -1409,7 +1409,7 @@ var init_simple_trace_file_exporter = __esm({
       async ensureStream() {
         if (!this.streamReady) {
           this.streamReady = (async () => {
-            await (0, import_promises34.mkdir)((0, import_node_path50.dirname)(this.filePath), { recursive: true });
+            await (0, import_promises33.mkdir)((0, import_node_path49.dirname)(this.filePath), { recursive: true });
             this.stream = (0, import_node_fs15.createWriteStream)(this.filePath, { flags: "w" });
             return this.stream;
           })();
@@ -1589,7 +1589,6 @@ __export(index_exports, {
   initializeBaseline: () => initializeBaseline,
   isAgentSkillsFormat: () => isAgentSkillsFormat,
   isEvaluatorKind: () => isEvaluatorKind,
-  isGuidelineFile: () => isGuidelineFile,
   isJsonObject: () => isJsonObject,
   isJsonValue: () => isJsonValue,
   isNonEmptyString: () => isNonEmptyString,
@@ -1849,9 +1848,9 @@ function mergeExecutionMetrics(computed, metrics) {
 }
 
 // src/evaluation/yaml-parser.ts
-var import_promises9 = require("fs/promises");
-var import_node_path9 = __toESM(require("path"), 1);
-var import_micromatch3 = __toESM(require("micromatch"), 1);
+var import_promises8 = require("fs/promises");
+var import_node_path8 = __toESM(require("path"), 1);
+var import_micromatch2 = __toESM(require("micromatch"), 1);
 var import_yaml4 = require("yaml");
 
 // src/evaluation/interpolation.ts
@@ -1944,7 +1943,6 @@ function parseAgentSkillsEvals(parsed, source = "evals.json", baseDir) {
       input_segments: [{ type: "text", value: prompt }],
       expected_output: evalCase.expected_output ? [{ role: "assistant", content: evalCase.expected_output }] : [],
       reference_answer: evalCase.expected_output,
-      guideline_paths: [],
       file_paths: filePaths,
       criteria: evalCase.expected_output ?? "",
       assertions,
@@ -2076,7 +2074,6 @@ async function expandFileReferences(tests, evalFileDir) {
 // src/evaluation/loaders/config-loader.ts
 var import_promises4 = require("fs/promises");
 var import_node_path4 = __toESM(require("path"), 1);
-var import_micromatch = __toESM(require("micromatch"), 1);
 var import_yaml2 = require("yaml");
 
 // src/evaluation/loaders/file-resolver.ts
@@ -2207,15 +2204,6 @@ async function loadConfig(evalFilePath, repoRoot) {
         logWarning(`Invalid required_version in ${configPath}, expected string`);
         continue;
       }
-      const guidelinePatterns = config.guideline_patterns;
-      if (guidelinePatterns !== void 0 && !Array.isArray(guidelinePatterns)) {
-        logWarning(`Invalid guideline_patterns in ${configPath}, expected array`);
-        continue;
-      }
-      if (Array.isArray(guidelinePatterns) && !guidelinePatterns.every((p) => typeof p === "string")) {
-        logWarning(`Invalid guideline_patterns in ${configPath}, all entries must be strings`);
-        continue;
-      }
       const evalPatterns = config.eval_patterns;
       if (evalPatterns !== void 0 && !Array.isArray(evalPatterns)) {
         logWarning(`Invalid eval_patterns in ${configPath}, expected array`);
@@ -2231,7 +2219,6 @@ async function loadConfig(evalFilePath, repoRoot) {
       );
       return {
         required_version: requiredVersion,
-        guideline_patterns: guidelinePatterns,
         eval_patterns: evalPatterns,
         execution: executionDefaults
       };
@@ -2243,11 +2230,6 @@ async function loadConfig(evalFilePath, repoRoot) {
   }
   return null;
 }
-function isGuidelineFile(filePath, patterns) {
-  const normalized = filePath.split("\\").join("/");
-  const patternsToUse = patterns ?? [];
-  return import_micromatch.default.isMatch(normalized, patternsToUse);
-}
 function extractTargetFromSuite(suite) {
   const execution = suite.execution;
   if (execution && typeof execution === "object" && !Array.isArray(execution)) {
@@ -3823,7 +3805,7 @@ function parseInlineRubrics(rawRubrics) {
 // src/evaluation/loaders/jsonl-parser.ts
 var import_promises7 = require("fs/promises");
 var import_node_path7 = __toESM(require("path"), 1);
-var import_micromatch2 = __toESM(require("micromatch"), 1);
+var import_micromatch = __toESM(require("micromatch"), 1);
 var import_yaml3 = require("yaml");
 
 // src/evaluation/loaders/message-processor.ts
@@ -3850,10 +3832,6 @@ function formatSegment(segment, mode = "lm") {
   if (type === "text") {
     return asString2(segment.value);
   }
-  if (type === "guideline_ref") {
-    const refPath = asString2(segment.path);
-    return refPath ? `<Attached: ${refPath}>` : void 0;
-  }
   if (type === "file") {
     const filePath = asString2(segment.path);
     if (!filePath) {
@@ -3876,9 +3854,6 @@ function hasVisibleContent(segments) {
       const value = asString2(segment.value);
       return value !== void 0 && value.trim().length > 0;
     }
-    if (type === "guideline_ref") {
-      return false;
-    }
     if (type === "file") {
       const text = asString2(segment.text);
       return text !== void 0 && text.trim().length > 0;
@@ -3894,17 +3869,7 @@ function asString2(value) {
 var ANSI_YELLOW5 = "\x1B[33m";
 var ANSI_RESET6 = "\x1B[0m";
 async function processMessages(options) {
-  const {
-    messages,
-    searchRoots,
-    repoRootPath,
-    guidelinePatterns,
-    guidelinePaths,
-    treatFileSegmentsAsGuidelines,
-    textParts,
-    messageType,
-    verbose
-  } = options;
+  const { messages, searchRoots, repoRootPath, textParts, messageType, verbose } = options;
   const segments = [];
   for (const message of messages) {
     const content = message.content;
@@ -3948,21 +3913,6 @@ async function processMessages(options) {
         }
         try {
           const fileContent = (await (0, import_promises6.readFile)(resolvedPath, "utf8")).replace(/\r\n/g, "\n");
-          const classifyAsGuideline = shouldTreatAsGuideline({
-            messageType,
-            resolvedPath,
-            repoRootPath,
-            guidelinePatterns,
-            treatFileSegmentsAsGuidelines
-          });
-          if (classifyAsGuideline && guidelinePaths) {
-            guidelinePaths.push(import_node_path6.default.resolve(resolvedPath));
-            if (verbose) {
-              console.log(`  [Guideline] Found: ${displayPath}`);
-              console.log(`    Resolved to: ${resolvedPath}`);
-            }
-            continue;
-          }
           segments.push({
             type: "file",
             path: displayPath,
@@ -3990,26 +3940,6 @@ async function processMessages(options) {
   }
   return segments;
 }
-function shouldTreatAsGuideline(options) {
-  const {
-    messageType,
-    resolvedPath,
-    repoRootPath,
-    guidelinePatterns,
-    treatFileSegmentsAsGuidelines
-  } = options;
-  if (messageType !== "input") {
-    return false;
-  }
-  if (treatFileSegmentsAsGuidelines) {
-    return true;
-  }
-  if (!guidelinePatterns || guidelinePatterns.length === 0) {
-    return false;
-  }
-  const relativeToRepo = import_node_path6.default.relative(repoRootPath, resolvedPath);
-  return isGuidelineFile(relativeToRepo, guidelinePatterns);
-}
 function asString3(value) {
   return typeof value === "string" ? value : void 0;
 }
@@ -4210,7 +4140,7 @@ async function loadSidecarMetadata(jsonlPath, verbose) {
     }
     return {
       description: asString4(parsed.description),
-      dataset: asString4(parsed.dataset),
+      name: asString4(parsed.name),
       execution: isJsonObject(parsed.execution) ? parsed.execution : void 0,
       evaluator: parsed.evaluator
     };
@@ -4246,20 +4176,18 @@ async function loadTestsFromJsonl(evalFilePath, repoRoot, options) {
   const absoluteTestPath = import_node_path7.default.resolve(evalFilePath);
   const repoRootPath = resolveToAbsolutePath(repoRoot);
   const searchRoots = buildSearchRoots(absoluteTestPath, repoRootPath);
-  const config = await loadConfig(absoluteTestPath, repoRootPath);
-  const guidelinePatterns = config?.guideline_patterns;
   const sidecar = await loadSidecarMetadata(absoluteTestPath, verbose);
   const rawFile = await (0, import_promises7.readFile)(absoluteTestPath, "utf8");
   const rawCases = parseJsonlContent(rawFile, evalFilePath);
-  const fallbackDataset = import_node_path7.default.basename(absoluteTestPath, ".jsonl") || "eval";
-  const datasetName = sidecar.dataset && sidecar.dataset.trim().length > 0 ? sidecar.dataset : fallbackDataset;
+  const fallbackEvalSet = import_node_path7.default.basename(absoluteTestPath, ".jsonl") || "eval";
+  const evalSetName = sidecar.name && sidecar.name.trim().length > 0 ? sidecar.name : fallbackEvalSet;
   const globalEvaluator = coerceEvaluator(sidecar.evaluator, "sidecar") ?? "llm-grader";
   const globalExecution = sidecar.execution;
   if (verbose) {
     console.log(`
 [JSONL Dataset: ${evalFilePath}]`);
     console.log(`  Cases: ${rawCases.length}`);
-    console.log(`  Dataset name: ${datasetName}`);
+    console.log(`  Eval set: ${evalSetName}`);
     if (sidecar.description) {
       console.log(`  Description: ${sidecar.description}`);
     }
@@ -4269,7 +4197,7 @@ async function loadTestsFromJsonl(evalFilePath, repoRoot, options) {
     const evalcase = rawCases[lineIndex];
     const lineNumber = lineIndex + 1;
     const id = asString4(evalcase.id);
-    if (filterPattern && (!id || !import_micromatch2.default.isMatch(id, filterPattern))) {
+    if (filterPattern && (!id || !import_micromatch.default.isMatch(id, filterPattern))) {
       continue;
     }
     const conversationId = asString4(evalcase.conversation_id);
@@ -4292,14 +4220,11 @@ async function loadTestsFromJsonl(evalFilePath, repoRoot, options) {
       continue;
     }
     const hasExpectedMessages = expectedMessages.length > 0;
-    const guidelinePaths = [];
     const inputTextParts = [];
     const inputSegments = await processMessages({
       messages: inputMessages,
       searchRoots,
       repoRootPath,
-      guidelinePatterns,
-      guidelinePaths,
       textParts: inputTextParts,
       messageType: "input",
       verbose
@@ -4349,40 +4274,20 @@ async function loadTestsFromJsonl(evalFilePath, repoRoot, options) {
         userFilePaths.push(segment.resolvedPath);
       }
     }
-    const allFilePaths = [
-      ...guidelinePaths.map((guidelinePath) => import_node_path7.default.resolve(guidelinePath)),
-      ...userFilePaths
-    ];
     const testCase = {
       id,
-      dataset: datasetName,
+      eval_set: evalSetName,
       conversation_id: conversationId,
       question,
       input: inputMessages,
       input_segments: inputSegments,
       expected_output: outputSegments,
       reference_answer: referenceAnswer,
-      guideline_paths: guidelinePaths.map((guidelinePath) => import_node_path7.default.resolve(guidelinePath)),
-      guideline_patterns: guidelinePatterns,
-      file_paths: allFilePaths,
+      file_paths: userFilePaths,
       criteria: outcome ?? "",
       evaluator: evalCaseEvaluatorKind,
       assertions: evaluators
     };
-    if (verbose) {
-      console.log(`
-[Test: ${id}]`);
-      if (testCase.guideline_paths.length > 0) {
-        console.log(`  Guidelines used: ${testCase.guideline_paths.length}`);
-        for (const guidelinePath of testCase.guideline_paths) {
-          console.log(`    - ${guidelinePath}`);
-        }
-      } else if (!guidelinePatterns || guidelinePatterns.length === 0) {
-        console.log("  No guidelines found (guideline_patterns not configured)");
-      } else {
-        console.log("  No guidelines found");
-      }
-    }
     results.push(testCase);
   }
   return results;
@@ -4440,30 +4345,7 @@ function parseMetadata(suite) {
 }
 
 // src/evaluation/formatting/prompt-builder.ts
-var import_promises8 = require("fs/promises");
-var import_node_path8 = __toESM(require("path"), 1);
-var ANSI_YELLOW7 = "\x1B[33m";
-var ANSI_RESET8 = "\x1B[0m";
 async function buildPromptInputs(testCase, mode = "lm") {
-  const guidelineParts = [];
-  for (const rawPath of testCase.guideline_paths) {
-    const absolutePath = import_node_path8.default.resolve(rawPath);
-    if (!await fileExists(absolutePath)) {
-      logWarning5(`Could not read guideline file ${absolutePath}: file does not exist`);
-      continue;
-    }
-    try {
-      const content = (await (0, import_promises8.readFile)(absolutePath, "utf8")).replace(/\r\n/g, "\n").trim();
-      guidelineParts.push({
-        content,
-        isFile: true,
-        displayPath: import_node_path8.default.basename(absolutePath)
-      });
-    } catch (error) {
-      logWarning5(`Could not read guideline file ${absolutePath}: ${error.message}`);
-    }
-  }
-  const guidelines = formatFileContents(guidelineParts);
   const segmentsByMessage = [];
   const fileContentsByPath = /* @__PURE__ */ new Map();
   for (const segment of testCase.input_segments) {
@@ -4488,10 +4370,6 @@ async function buildPromptInputs(testCase, mode = "lm") {
           if (type === "file") {
             const value = asString5(segment.value);
             if (!value) continue;
-            if (testCase.guideline_patterns && isGuidelineFile(value, testCase.guideline_patterns)) {
-              messageSegments.push({ type: "guideline_ref", path: value });
-              continue;
-            }
             const fileText = fileContentsByPath.get(value);
             if (fileText !== void 0) {
               messageSegments.push({ type: "file", text: fileText, path: value });
@@ -4540,10 +4418,6 @@ ${messageContent}`);
   } else {
     const questionParts = [];
     for (const segment of testCase.input_segments) {
-      if (segment.type === "file" && typeof segment.path === "string" && testCase.guideline_patterns && isGuidelineFile(segment.path, testCase.guideline_patterns)) {
-        questionParts.push(`<Attached: ${segment.path}>`);
-        continue;
-      }
       const formattedContent = formatSegment(segment, mode);
       if (formattedContent) {
         questionParts.push(formattedContent);
@@ -4554,11 +4428,9 @@ ${messageContent}`);
   const chatPrompt = useRoleMarkers ? buildChatPromptFromSegments({
     messages: testCase.input,
     segmentsByMessage,
-    guidelinePatterns: testCase.guideline_patterns,
-    guidelineContent: guidelines,
     mode
   }) : void 0;
-  return { question, guidelines, chatPrompt };
+  return { question, chatPrompt };
 }
 function needsRoleMarkers(messages, processedSegmentsByMessage) {
   if (messages.some((msg) => msg.role === "assistant" || msg.role === "tool")) {
@@ -4573,14 +4445,7 @@ function needsRoleMarkers(messages, processedSegmentsByMessage) {
   return messagesWithContent > 1;
 }
 function buildChatPromptFromSegments(options) {
-  const {
-    messages,
-    segmentsByMessage,
-    guidelinePatterns,
-    guidelineContent,
-    systemPrompt,
-    mode = "lm"
-  } = options;
+  const { messages, segmentsByMessage, systemPrompt, mode = "lm" } = options;
   if (messages.length === 0) {
     return void 0;
   }
@@ -4588,11 +4453,6 @@ function buildChatPromptFromSegments(options) {
   if (systemPrompt && systemPrompt.trim().length > 0) {
     systemSegments.push(systemPrompt.trim());
   }
-  if (guidelineContent && guidelineContent.trim().length > 0) {
-    systemSegments.push(`[[ ## Guidelines ## ]]
-
-${guidelineContent.trim()}`);
-  }
   let startIndex = 0;
   while (startIndex < messages.length && messages[startIndex].role === "system") {
     const segments = segmentsByMessage[startIndex];
@@ -4628,15 +4488,8 @@ ${guidelineContent.trim()}`);
       contentParts.push("@[Tool]:");
     }
     for (const segment of segments) {
-      if (segment.type === "guideline_ref") {
-        continue;
-      }
       const formatted = formatSegment(segment, mode);
       if (formatted) {
-        const isGuidelineRef = segment.type === "file" && typeof segment.path === "string" && guidelinePatterns && isGuidelineFile(segment.path, guidelinePatterns);
-        if (isGuidelineRef) {
-          continue;
-        }
         contentParts.push(formatted);
       }
     }
@@ -4654,30 +4507,27 @@ ${guidelineContent.trim()}`);
 function asString5(value) {
   return typeof value === "string" ? value : void 0;
 }
-function logWarning5(message) {
-  console.warn(`${ANSI_YELLOW7}Warning: ${message}${ANSI_RESET8}`);
-}
 
 // src/evaluation/yaml-parser.ts
-var ANSI_YELLOW8 = "\x1B[33m";
+var ANSI_YELLOW7 = "\x1B[33m";
 var ANSI_RED3 = "\x1B[31m";
-var ANSI_RESET9 = "\x1B[0m";
+var ANSI_RESET8 = "\x1B[0m";
 function resolveTests(suite) {
   if (suite.tests !== void 0) return suite.tests;
   if (suite.eval_cases !== void 0) {
-    logWarning6("'eval_cases' is deprecated. Use 'tests' instead.");
+    logWarning5("'eval_cases' is deprecated. Use 'tests' instead.");
     return suite.eval_cases;
   }
   if (suite.evalcases !== void 0) {
-    logWarning6("'evalcases' is deprecated. Use 'tests' instead.");
+    logWarning5("'evalcases' is deprecated. Use 'tests' instead.");
     return suite.evalcases;
   }
   return void 0;
 }
 async function readTestSuiteMetadata(testFilePath) {
   try {
-    const absolutePath = import_node_path9.default.resolve(testFilePath);
-    const content = await (0, import_promises9.readFile)(absolutePath, "utf8");
+    const absolutePath = import_node_path8.default.resolve(testFilePath);
+    const content = await (0, import_promises8.readFile)(absolutePath, "utf8");
     const parsed = interpolateEnv((0, import_yaml4.parse)(content), process.env);
     if (!isJsonObject(parsed)) {
       return {};
@@ -4728,26 +4578,25 @@ var loadEvalCases = loadTests;
 async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
   const verbose = options?.verbose ?? false;
   const filterPattern = options?.filter;
-  const absoluteTestPath = import_node_path9.default.resolve(evalFilePath);
+  const absoluteTestPath = import_node_path8.default.resolve(evalFilePath);
   const repoRootPath = resolveToAbsolutePath(repoRoot);
   const searchRoots = buildSearchRoots(absoluteTestPath, repoRootPath);
   const config = await loadConfig(absoluteTestPath, repoRootPath);
-  const guidelinePatterns = config?.guideline_patterns;
-  const rawFile = await (0, import_promises9.readFile)(absoluteTestPath, "utf8");
+  const rawFile = await (0, import_promises8.readFile)(absoluteTestPath, "utf8");
   const interpolated = interpolateEnv((0, import_yaml4.parse)(rawFile), process.env);
   if (!isJsonObject(interpolated)) {
     throw new Error(`Invalid test file format: ${evalFilePath}`);
   }
   const suite = interpolated;
-  const datasetNameFromSuite = asString6(suite.dataset)?.trim();
-  const fallbackDataset = import_node_path9.default.basename(absoluteTestPath).replace(/\.ya?ml$/i, "") || "eval";
-  const datasetName = datasetNameFromSuite && datasetNameFromSuite.length > 0 ? datasetNameFromSuite : fallbackDataset;
+  const evalSetNameFromSuite = asString6(suite.name)?.trim();
+  const fallbackEvalSet = import_node_path8.default.basename(absoluteTestPath).replace(/\.ya?ml$/i, "") || "eval";
+  const evalSetName = evalSetNameFromSuite && evalSetNameFromSuite.length > 0 ? evalSetNameFromSuite : fallbackEvalSet;
   const rawTestcases = resolveTests(suite);
   const globalEvaluator = coerceEvaluator(suite.evaluator, "global") ?? "llm-grader";
-  const evalFileDir = import_node_path9.default.dirname(absoluteTestPath);
+  const evalFileDir = import_node_path8.default.dirname(absoluteTestPath);
   let expandedTestcases;
   if (typeof rawTestcases === "string") {
-    const externalPath = import_node_path9.default.resolve(evalFileDir, rawTestcases);
+    const externalPath = import_node_path8.default.resolve(evalFileDir, rawTestcases);
     expandedTestcases = await loadCasesFromFile(externalPath);
   } else if (Array.isArray(rawTestcases)) {
     expandedTestcases = await expandFileReferences(rawTestcases, evalFileDir);
@@ -4761,18 +4610,18 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
   const _globalTarget = asString6(rawGlobalExecution?.target) ?? asString6(suite.target);
   const suiteAssertions = suite.assertions ?? suite.assert;
   if (suite.assert !== void 0 && suite.assertions === void 0) {
-    logWarning6("'assert' is deprecated at the suite level. Use 'assertions' instead.");
+    logWarning5("'assert' is deprecated at the suite level. Use 'assertions' instead.");
   }
   const globalExecution = suiteAssertions !== void 0 ? { ...rawGlobalExecution ?? {}, assertions: suiteAssertions } : rawGlobalExecution;
   const results = [];
   for (const rawEvalcase of expandedTestcases) {
     if (!isJsonObject(rawEvalcase)) {
-      logWarning6("Skipping invalid test entry (expected object)");
+      logWarning5("Skipping invalid test entry (expected object)");
       continue;
     }
     const evalcase = rawEvalcase;
     const id = asString6(evalcase.id);
-    if (filterPattern && (!id || !import_micromatch3.default.isMatch(id, filterPattern))) {
+    if (filterPattern && (!id || !import_micromatch2.default.isMatch(id, filterPattern))) {
       continue;
     }
     const conversationId = asString6(evalcase.conversation_id);
@@ -4780,7 +4629,7 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
     if (!outcome && evalcase.expected_outcome !== void 0) {
       outcome = asString6(evalcase.expected_outcome);
       if (outcome) {
-        logWarning6(
+        logWarning5(
           `Test '${asString6(evalcase.id) ?? "unknown"}': 'expected_outcome' is deprecated. Use 'criteria' instead.`
         );
       }
@@ -4800,15 +4649,11 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
     const effectiveSuiteInputMessages = suiteInputMessages && !skipDefaults ? suiteInputMessages : void 0;
     const inputMessages = effectiveSuiteInputMessages ? [...effectiveSuiteInputMessages, ...testInputMessages] : testInputMessages;
     const hasExpectedMessages = expectedMessages.length > 0;
-    const guidelinePaths = [];
     const inputTextParts = [];
     const suiteInputSegments = effectiveSuiteInputMessages ? await processMessages({
       messages: effectiveSuiteInputMessages,
       searchRoots,
       repoRootPath,
-      guidelinePatterns,
-      guidelinePaths,
-      treatFileSegmentsAsGuidelines: true,
       textParts: inputTextParts,
       messageType: "input",
       verbose
@@ -4817,8 +4662,6 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
       messages: testInputMessages,
       searchRoots,
       repoRootPath,
-      guidelinePatterns,
-      guidelinePaths,
       textParts: inputTextParts,
       messageType: "input",
       verbose
@@ -4867,26 +4710,20 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
         userFilePaths.push(segment.resolvedPath);
       }
     }
-    const allFilePaths = [
-      ...guidelinePaths.map((guidelinePath) => import_node_path9.default.resolve(guidelinePath)),
-      ...userFilePaths
-    ];
     const caseWorkspace = await resolveWorkspaceConfig(evalcase.workspace, evalFileDir);
     const mergedWorkspace = mergeWorkspaceConfigs(suiteWorkspace, caseWorkspace);
     const metadata = isJsonObject(evalcase.metadata) ? evalcase.metadata : void 0;
     const caseTargets = extractTargetsFromTestCase(evalcase);
     const testCase = {
       id,
-      dataset: datasetName,
+      eval_set: evalSetName,
       conversation_id: conversationId,
       question,
       input: inputMessages,
       input_segments: inputSegments,
       expected_output: outputSegments,
       reference_answer: referenceAnswer,
-      guideline_paths: guidelinePaths.map((guidelinePath) => import_node_path9.default.resolve(guidelinePath)),
-      guideline_patterns: guidelinePatterns,
-      file_paths: allFilePaths,
+      file_paths: userFilePaths,
       criteria: outcome ?? "",
       evaluator: evalCaseEvaluatorKind,
       assertions: evaluators,
@@ -4894,20 +4731,6 @@ async function loadTestsFromYaml(evalFilePath, repoRoot, options) {
       metadata,
       targets: caseTargets
     };
-    if (verbose) {
-      console.log(`
-[Test: ${id}]`);
-      if (testCase.guideline_paths.length > 0) {
-        console.log(`  Guidelines used: ${testCase.guideline_paths.length}`);
-        for (const guidelinePath of testCase.guideline_paths) {
-          console.log(`    - ${guidelinePath}`);
-        }
-      } else if (!guidelinePatterns || guidelinePatterns.length === 0) {
-        console.log("  No guidelines found (guideline_patterns not configured)");
-      } else {
-        console.log("  No guidelines found");
-      }
-    }
     results.push(testCase);
   }
   return { tests: results, parsed: suite };
@@ -4926,7 +4749,7 @@ function parseWorkspaceScriptConfig(raw, evalFileDir) {
   if (!isJsonObject(raw)) return void 0;
   const obj = raw;
   if (obj.script !== void 0 && obj.command === void 0) {
-    logWarning6("'script' is deprecated. Use 'command' instead.");
+    logWarning5("'script' is deprecated. Use 'command' instead.");
   }
   const commandSource = obj.command ?? obj.script;
   if (!Array.isArray(commandSource) || commandSource.length === 0) return void 0;
@@ -4934,8 +4757,8 @@ function parseWorkspaceScriptConfig(raw, evalFileDir) {
   if (commandArr.length === 0) return void 0;
   const timeoutMs = typeof obj.timeout_ms === "number" ? obj.timeout_ms : void 0;
   let cwd = typeof obj.cwd === "string" ? obj.cwd : void 0;
-  if (cwd && !import_node_path9.default.isAbsolute(cwd)) {
-    cwd = import_node_path9.default.resolve(evalFileDir, cwd);
+  if (cwd && !import_node_path8.default.isAbsolute(cwd)) {
+    cwd = import_node_path8.default.resolve(evalFileDir, cwd);
   }
   const config = { command: commandArr };
   if (timeoutMs !== void 0) {
@@ -5025,10 +4848,10 @@ function parseWorkspaceHooksConfig(raw, evalFileDir) {
 }
 async function resolveWorkspaceConfig(raw, evalFileDir) {
   if (typeof raw === "string") {
-    const workspaceFilePath = import_node_path9.default.resolve(evalFileDir, raw);
+    const workspaceFilePath = import_node_path8.default.resolve(evalFileDir, raw);
     let content;
     try {
-      content = await (0, import_promises9.readFile)(workspaceFilePath, "utf8");
+      content = await (0, import_promises8.readFile)(workspaceFilePath, "utf8");
     } catch {
       throw new Error(`Workspace file not found: ${raw} (resolved to ${workspaceFilePath})`);
     }
@@ -5038,7 +4861,7 @@ async function resolveWorkspaceConfig(raw, evalFileDir) {
         `Invalid workspace file format: ${workspaceFilePath} (expected a YAML object)`
       );
     }
-    const workspaceFileDir = import_node_path9.default.dirname(workspaceFilePath);
+    const workspaceFileDir = import_node_path8.default.dirname(workspaceFilePath);
     return parseWorkspaceConfig(parsed, workspaceFileDir);
   }
   return parseWorkspaceConfig(raw, evalFileDir);
@@ -5058,8 +4881,8 @@ function parseWorkspaceConfig(raw, evalFileDir) {
     throw new Error("workspace.static has been removed. Use workspace.mode='static'.");
   }
   let template = typeof obj.template === "string" ? obj.template : void 0;
-  if (template && !import_node_path9.default.isAbsolute(template)) {
-    template = import_node_path9.default.resolve(evalFileDir, template);
+  if (template && !import_node_path8.default.isAbsolute(template)) {
+    template = import_node_path8.default.resolve(evalFileDir, template);
   }
   const isolation = obj.isolation === "shared" || obj.isolation === "per_test" ? obj.isolation : void 0;
   const repos = Array.isArray(obj.repos) ? obj.repos.map(parseRepoConfig).filter(Boolean) : void 0;
@@ -5109,28 +4932,28 @@ function mergeWorkspaceConfigs(suiteLevel, caseLevel) {
 function asString6(value) {
   return typeof value === "string" ? value : void 0;
 }
-function logWarning6(message, details) {
+function logWarning5(message, details) {
   if (details && details.length > 0) {
     const detailBlock = details.join("\n");
-    console.warn(`${ANSI_YELLOW8}Warning: ${message}
-${detailBlock}${ANSI_RESET9}`);
+    console.warn(`${ANSI_YELLOW7}Warning: ${message}
+${detailBlock}${ANSI_RESET8}`);
   } else {
-    console.warn(`${ANSI_YELLOW8}Warning: ${message}${ANSI_RESET9}`);
+    console.warn(`${ANSI_YELLOW7}Warning: ${message}${ANSI_RESET8}`);
   }
 }
 function logError3(message, details) {
   if (details && details.length > 0) {
     const detailBlock = details.join("\n");
     console.error(`${ANSI_RED3}Error: ${message}
-${detailBlock}${ANSI_RESET9}`);
+${detailBlock}${ANSI_RESET8}`);
   } else {
-    console.error(`${ANSI_RED3}Error: ${message}${ANSI_RESET9}`);
+    console.error(`${ANSI_RED3}Error: ${message}${ANSI_RESET8}`);
   }
 }
 
 // src/evaluation/loaders/eval-yaml-transpiler.ts
 var import_node_fs2 = require("fs");
-var import_node_path10 = __toESM(require("path"), 1);
+var import_node_path9 = __toESM(require("path"), 1);
 var import_yaml5 = require("yaml");
 function codeGraderInstruction(graderName, description) {
   const desc = description ? ` This grader: ${description}.` : "";
@@ -5375,7 +5198,7 @@ function transpileEvalYaml(suite, source = "EVAL.yaml") {
 function transpileEvalYamlFile(evalYamlPath) {
   const content = (0, import_node_fs2.readFileSync)(evalYamlPath, "utf8");
   const parsed = (0, import_yaml5.parse)(content);
-  return transpileEvalYaml(parsed, import_node_path10.default.basename(evalYamlPath));
+  return transpileEvalYaml(parsed, import_node_path9.default.basename(evalYamlPath));
 }
 function getOutputFilenames(result) {
   const names = /* @__PURE__ */ new Map();
@@ -5394,11 +5217,11 @@ function getOutputFilenames(result) {
 
 // src/evaluation/file-utils.ts
 var import_node_fs3 = require("fs");
-var import_promises10 = require("fs/promises");
-var import_node_path11 = __toESM(require("path"), 1);
+var import_promises9 = require("fs/promises");
+var import_node_path10 = __toESM(require("path"), 1);
 async function fileExists2(filePath) {
   try {
-    await (0, import_promises10.access)(filePath, import_node_fs3.constants.F_OK);
+    await (0, import_promises9.access)(filePath, import_node_fs3.constants.F_OK);
     return true;
   } catch {
     return false;
@@ -5408,22 +5231,22 @@ function normalizeLineEndings(content) {
   return content.replace(/\r\n/g, "\n");
 }
 async function readTextFile(filePath) {
-  const content = await (0, import_promises10.readFile)(filePath, "utf8");
+  const content = await (0, import_promises9.readFile)(filePath, "utf8");
   return normalizeLineEndings(content);
 }
 async function readJsonFile(filePath) {
-  const content = await (0, import_promises10.readFile)(filePath, "utf8");
+  const content = await (0, import_promises9.readFile)(filePath, "utf8");
   return JSON.parse(content);
 }
 async function findGitRoot(startPath) {
-  let currentDir = import_node_path11.default.dirname(import_node_path11.default.resolve(startPath));
-  const root = import_node_path11.default.parse(currentDir).root;
+  let currentDir = import_node_path10.default.dirname(import_node_path10.default.resolve(startPath));
+  const root = import_node_path10.default.parse(currentDir).root;
   while (currentDir !== root) {
-    const gitPath = import_node_path11.default.join(currentDir, ".git");
+    const gitPath = import_node_path10.default.join(currentDir, ".git");
     if (await fileExists2(gitPath)) {
       return currentDir;
     }
-    const parentDir = import_node_path11.default.dirname(currentDir);
+    const parentDir = import_node_path10.default.dirname(currentDir);
     if (parentDir === currentDir) {
       break;
     }
@@ -5434,8 +5257,8 @@ async function findGitRoot(startPath) {
 function buildDirectoryChain2(filePath, repoRoot) {
   const directories = [];
   const seen = /* @__PURE__ */ new Set();
-  const boundary = import_node_path11.default.resolve(repoRoot);
-  let current = import_node_path11.default.resolve(import_node_path11.default.dirname(filePath));
+  const boundary = import_node_path10.default.resolve(repoRoot);
+  let current = import_node_path10.default.resolve(import_node_path10.default.dirname(filePath));
   while (current !== void 0) {
     if (!seen.has(current)) {
       directories.push(current);
@@ -5444,7 +5267,7 @@ function buildDirectoryChain2(filePath, repoRoot) {
     if (current === boundary) {
       break;
     }
-    const parent = import_node_path11.default.dirname(current);
+    const parent = import_node_path10.default.dirname(current);
     if (parent === current) {
       break;
     }
@@ -5458,16 +5281,16 @@ function buildDirectoryChain2(filePath, repoRoot) {
 function buildSearchRoots2(evalPath, repoRoot) {
   const uniqueRoots = [];
   const addRoot = (root) => {
-    const normalized = import_node_path11.default.resolve(root);
+    const normalized = import_node_path10.default.resolve(root);
     if (!uniqueRoots.includes(normalized)) {
       uniqueRoots.push(normalized);
     }
   };
-  let currentDir = import_node_path11.default.dirname(evalPath);
+  let currentDir = import_node_path10.default.dirname(evalPath);
   let reachedBoundary = false;
   while (!reachedBoundary) {
     addRoot(currentDir);
-    const parentDir = import_node_path11.default.dirname(currentDir);
+    const parentDir = import_node_path10.default.dirname(currentDir);
     if (currentDir === repoRoot || parentDir === currentDir) {
       reachedBoundary = true;
     } else {
@@ -5485,16 +5308,16 @@ function trimLeadingSeparators2(value) {
 async function resolveFileReference3(rawValue, searchRoots) {
   const displayPath = trimLeadingSeparators2(rawValue);
   const potentialPaths = [];
-  if (import_node_path11.default.isAbsolute(rawValue)) {
-    potentialPaths.push(import_node_path11.default.normalize(rawValue));
+  if (import_node_path10.default.isAbsolute(rawValue)) {
+    potentialPaths.push(import_node_path10.default.normalize(rawValue));
   }
   for (const base of searchRoots) {
-    potentialPaths.push(import_node_path11.default.resolve(base, displayPath));
+    potentialPaths.push(import_node_path10.default.resolve(base, displayPath));
   }
   const attempted = [];
   const seen = /* @__PURE__ */ new Set();
   for (const candidate of potentialPaths) {
-    const absoluteCandidate = import_node_path11.default.resolve(candidate);
+    const absoluteCandidate = import_node_path10.default.resolve(candidate);
     if (seen.has(absoluteCandidate)) {
       continue;
     }
@@ -5728,10 +5551,10 @@ function buildChatPrompt(request) {
     if (hasSystemMessage) {
       return provided;
     }
-    const systemContent2 = resolveSystemContent(request, false);
+    const systemContent2 = resolveSystemContent(request);
     return [{ role: "system", content: systemContent2 }, ...provided];
   }
-  const systemContent = resolveSystemContent(request, true);
+  const systemContent = resolveSystemContent(request);
   const userContent = request.question.trim();
   const prompt = [
     { role: "system", content: systemContent },
@@ -5739,18 +5562,13 @@ function buildChatPrompt(request) {
   ];
   return prompt;
 }
-function resolveSystemContent(request, includeGuidelines) {
+function resolveSystemContent(request) {
   const systemSegments = [];
   if (request.systemPrompt && request.systemPrompt.trim().length > 0) {
     systemSegments.push(request.systemPrompt.trim());
   } else {
     systemSegments.push(DEFAULT_SYSTEM_PROMPT);
   }
-  if (includeGuidelines && request.guidelines && request.guidelines.trim().length > 0) {
-    systemSegments.push(`[[ ## Guidelines ## ]]
-
-${request.guidelines.trim()}`);
-  }
   return systemSegments.join("\n\n");
 }
 function toModelMessages(chatPrompt) {
@@ -5933,8 +5751,8 @@ async function withRetry(fn, retryConfig, signal) {
 var import_node_child_process = require("child_process");
 var import_node_crypto = require("crypto");
 var import_node_fs4 = require("fs");
-var import_promises11 = require("fs/promises");
-var import_node_path13 = __toESM(require("path"), 1);
+var import_promises10 = require("fs/promises");
+var import_node_path12 = __toESM(require("path"), 1);
 
 // src/evaluation/providers/claude-log-tracker.ts
 var GLOBAL_LOGS_KEY = Symbol.for("agentv.claudeLogs");
@@ -5990,17 +5808,11 @@ function subscribeToClaudeLogEntries(listener) {
 }
 
 // src/evaluation/providers/preread.ts
-var import_node_path12 = __toESM(require("path"), 1);
-function buildPromptDocument(request, inputFiles, options) {
+var import_node_path11 = __toESM(require("path"), 1);
+function buildPromptDocument(request, inputFiles) {
   const parts = [];
-  const guidelineFiles = collectGuidelineFiles(
-    inputFiles,
-    options?.guidelinePatterns ?? request.guideline_patterns,
-    options?.guidelineOverrides
-  );
   const inputFilesList = collectInputFiles(inputFiles);
-  const nonGuidelineInputFiles = inputFilesList.filter((file) => !guidelineFiles.includes(file));
-  const prereadBlock = buildMandatoryPrereadBlock(guidelineFiles, nonGuidelineInputFiles);
+  const prereadBlock = buildMandatoryPrereadBlock(inputFilesList);
   if (prereadBlock.length > 0) {
     parts.push("\n", prereadBlock);
   }
@@ -6013,62 +5825,36 @@ function normalizeInputFiles(inputFiles) {
   }
   const deduped = /* @__PURE__ */ new Map();
   for (const inputFile of inputFiles) {
-    const absolutePath = import_node_path12.default.resolve(inputFile);
+    const absolutePath = import_node_path11.default.resolve(inputFile);
     if (!deduped.has(absolutePath)) {
       deduped.set(absolutePath, absolutePath);
     }
   }
   return Array.from(deduped.values());
 }
-function collectGuidelineFiles(inputFiles, guidelinePatterns, overrides) {
-  if (!inputFiles || inputFiles.length === 0) {
-    return [];
-  }
-  const unique = /* @__PURE__ */ new Map();
-  for (const inputFile of inputFiles) {
-    const absolutePath = import_node_path12.default.resolve(inputFile);
-    if (overrides?.has(absolutePath)) {
-      if (!unique.has(absolutePath)) {
-        unique.set(absolutePath, absolutePath);
-      }
-      continue;
-    }
-    const normalized = absolutePath.split(import_node_path12.default.sep).join("/");
-    if (isGuidelineFile(normalized, guidelinePatterns)) {
-      if (!unique.has(absolutePath)) {
-        unique.set(absolutePath, absolutePath);
-      }
-    }
-  }
-  return Array.from(unique.values());
-}
 function collectInputFiles(inputFiles) {
   if (!inputFiles || inputFiles.length === 0) {
     return [];
   }
   const unique = /* @__PURE__ */ new Map();
   for (const inputFile of inputFiles) {
-    const absolutePath = import_node_path12.default.resolve(inputFile);
+    const absolutePath = import_node_path11.default.resolve(inputFile);
     if (!unique.has(absolutePath)) {
       unique.set(absolutePath, absolutePath);
     }
   }
   return Array.from(unique.values());
 }
-function buildMandatoryPrereadBlock(guidelineFiles, inputFiles) {
-  if (guidelineFiles.length === 0 && inputFiles.length === 0) {
+function buildMandatoryPrereadBlock(inputFiles) {
+  if (inputFiles.length === 0) {
     return "";
   }
   const buildList = (files) => files.map((absolutePath) => {
-    const fileName = import_node_path12.default.basename(absolutePath);
+    const fileName = import_node_path11.default.basename(absolutePath);
     const fileUri = pathToFileUri(absolutePath);
     return `* [${fileName}](${fileUri})`;
   });
   const sections = [];
-  if (guidelineFiles.length > 0) {
-    sections.push(`Read all guideline files:
-${buildList(guidelineFiles).join("\n")}.`);
-  }
   if (inputFiles.length > 0) {
     sections.push(`Read all input files:
 ${buildList(inputFiles).join("\n")}.`);
@@ -6080,7 +5866,7 @@ ${buildList(inputFiles).join("\n")}.`);
   return sections.join("\n");
 }
 function pathToFileUri(filePath) {
-  const absolutePath = import_node_path12.default.isAbsolute(filePath) ? filePath : import_node_path12.default.resolve(filePath);
+  const absolutePath = import_node_path11.default.isAbsolute(filePath) ? filePath : import_node_path11.default.resolve(filePath);
   const normalizedPath = absolutePath.replace(/\\/g, "/");
   if (/^[a-zA-Z]:\//.test(normalizedPath)) {
     return `file:///${normalizedPath}`;
@@ -6227,10 +6013,10 @@ var ClaudeCliProvider = class {
   }
   resolveCwd(cwdOverride) {
     if (cwdOverride) {
-      return import_node_path13.default.resolve(cwdOverride);
+      return import_node_path12.default.resolve(cwdOverride);
     }
     if (this.config.cwd) {
-      return import_node_path13.default.resolve(this.config.cwd);
+      return import_node_path12.default.resolve(this.config.cwd);
     }
     return void 0;
   }
@@ -6240,9 +6026,9 @@ var ClaudeCliProvider = class {
       return void 0;
     }
     if (this.config.logDir) {
-      return import_node_path13.default.resolve(this.config.logDir);
+      return import_node_path12.default.resolve(this.config.logDir);
     }
-    return import_node_path13.default.join(process.cwd(), ".agentv", "logs", "claude-cli");
+    return import_node_path12.default.join(process.cwd(), ".agentv", "logs", "claude-cli");
   }
   async createStreamLogger(request) {
     const logDir = this.resolveLogDirectory();
@@ -6250,13 +6036,13 @@ var ClaudeCliProvider = class {
       return void 0;
     }
     try {
-      await (0, import_promises11.mkdir)(logDir, { recursive: true });
+      await (0, import_promises10.mkdir)(logDir, { recursive: true });
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
       console.warn(`Skipping Claude CLI stream logging (could not create ${logDir}): ${message}`);
       return void 0;
     }
-    const filePath = import_node_path13.default.join(logDir, buildLogFilename(request, this.targetName));
+    const filePath = import_node_path12.default.join(logDir, buildLogFilename(request, this.targetName));
     try {
       const logger = await ClaudeCliStreamLogger.create({
         filePath,
@@ -6576,8 +6362,8 @@ function tryParseJson(line) {
 // src/evaluation/providers/claude-sdk.ts
 var import_node_crypto2 = require("crypto");
 var import_node_fs5 = require("fs");
-var import_promises12 = require("fs/promises");
-var import_node_path14 = __toESM(require("path"), 1);
+var import_promises11 = require("fs/promises");
+var import_node_path13 = __toESM(require("path"), 1);
 var claudeSdkModule = null;
 async function loadClaudeSdk() {
   if (!claudeSdkModule) {
@@ -6737,10 +6523,10 @@ var ClaudeSdkProvider = class {
   }
   resolveCwd(cwdOverride) {
     if (cwdOverride) {
-      return import_node_path14.default.resolve(cwdOverride);
+      return import_node_path13.default.resolve(cwdOverride);
     }
     if (this.config.cwd) {
-      return import_node_path14.default.resolve(this.config.cwd);
+      return import_node_path13.default.resolve(this.config.cwd);
     }
     return void 0;
   }
@@ -6750,9 +6536,9 @@ var ClaudeSdkProvider = class {
       return void 0;
     }
     if (this.config.logDir) {
-      return import_node_path14.default.resolve(this.config.logDir);
+      return import_node_path13.default.resolve(this.config.logDir);
     }
-    return import_node_path14.default.join(process.cwd(), ".agentv", "logs", "claude");
+    return import_node_path13.default.join(process.cwd(), ".agentv", "logs", "claude");
   }
   async createStreamLogger(request) {
     const logDir = this.resolveLogDirectory();
@@ -6760,13 +6546,13 @@ var ClaudeSdkProvider = class {
       return void 0;
     }
     try {
-      await (0, import_promises12.mkdir)(logDir, { recursive: true });
+      await (0, import_promises11.mkdir)(logDir, { recursive: true });
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
       console.warn(`Skipping Claude stream logging (could not create ${logDir}): ${message}`);
       return void 0;
     }
-    const filePath = import_node_path14.default.join(logDir, buildLogFilename2(request, this.targetName));
+    const filePath = import_node_path13.default.join(logDir, buildLogFilename2(request, this.targetName));
     try {
       const logger = await ClaudeStreamLogger.create({
         filePath,
@@ -6971,9 +6757,9 @@ function formatElapsed2(startedAt) {
 
 // src/evaluation/providers/cli.ts
 var import_node_child_process2 = require("child_process");
-var import_promises13 = __toESM(require("fs/promises"), 1);
+var import_promises12 = __toESM(require("fs/promises"), 1);
 var import_node_os = __toESM(require("os"), 1);
-var import_node_path15 = __toESM(require("path"), 1);
+var import_node_path14 = __toESM(require("path"), 1);
 var import_node_util = require("util");
 var import_zod2 = require("zod");
 var ToolCallSchema = import_zod2.z.object({
@@ -7182,7 +6968,6 @@ var CliProvider = class {
     const { values: templateValues, promptFilePath } = await buildTemplateValues(
       {
         question: "",
-        guidelines: "",
         inputFiles: batchInputFiles,
         evalCaseId: "batch",
         attempt: 0
@@ -7370,7 +7155,7 @@ var CliProvider = class {
       throw new Error(`Failed to read output file '${filePath}': ${errorMsg}`);
     } finally {
       if (!this.keepTempFiles) {
-        await import_promises13.default.unlink(filePath).catch(() => {
+        await import_promises12.default.unlink(filePath).catch(() => {
         });
       }
     }
@@ -7415,7 +7200,6 @@ var CliProvider = class {
     const { values: templateValues, promptFilePath } = await buildTemplateValues(
       {
         question: "",
-        guidelines: "",
         inputFiles: [],
         evalCaseId: "healthcheck",
         attempt: 0
@@ -7451,12 +7235,11 @@ var CliProvider = class {
 async function buildTemplateValues(request, config, outputFilePath) {
   const inputFiles = normalizeInputFiles2(request.inputFiles);
   const promptFilePath = generateOutputFilePath(request.evalCaseId, ".prompt.txt");
-  await import_promises13.default.writeFile(promptFilePath, request.question ?? "", "utf8");
+  await import_promises12.default.writeFile(promptFilePath, request.question ?? "", "utf8");
   return {
     values: {
       PROMPT: shellEscape(request.question ?? ""),
       PROMPT_FILE: shellEscape(promptFilePath),
-      GUIDELINES: shellEscape(request.guidelines ?? ""),
       EVAL_ID: shellEscape(request.evalCaseId ?? ""),
       ATTEMPT: shellEscape(String(request.attempt ?? 0)),
       FILES: formatFileList(inputFiles, config.filesFormat),
@@ -7469,7 +7252,7 @@ async function cleanupTempFile(filePath, keepTempFiles) {
   if (!filePath || keepTempFiles) {
     return;
   }
-  await import_promises13.default.unlink(filePath).catch(() => {
+  await import_promises12.default.unlink(filePath).catch(() => {
   });
 }
 function normalizeInputFiles2(inputFiles) {
@@ -7478,7 +7261,7 @@ function normalizeInputFiles2(inputFiles) {
   }
   const unique = /* @__PURE__ */ new Map();
   for (const inputFile of inputFiles) {
-    const absolutePath = import_node_path15.default.resolve(inputFile);
+    const absolutePath = import_node_path14.default.resolve(inputFile);
     if (!unique.has(absolutePath)) {
       unique.set(absolutePath, absolutePath);
     }
@@ -7492,7 +7275,7 @@ function formatFileList(files, template) {
   const formatter = template ?? "{path}";
   return files.map((filePath) => {
     const escapedPath = shellEscape(filePath);
-    const escapedName = shellEscape(import_node_path15.default.basename(filePath));
+    const escapedName = shellEscape(import_node_path14.default.basename(filePath));
     return formatter.replaceAll("{path}", escapedPath).replaceAll("{basename}", escapedName);
   }).join(" ");
 }
@@ -7516,7 +7299,7 @@ function generateOutputFilePath(evalCaseId, extension = ".json") {
   const safeEvalId = evalCaseId || "unknown";
   const timestamp = Date.now();
   const random = Math.random().toString(36).substring(2, 9);
-  return import_node_path15.default.join(import_node_os.default.tmpdir(), `agentv-${safeEvalId}-${timestamp}-${random}${extension}`);
+  return import_node_path14.default.join(import_node_os.default.tmpdir(), `agentv-${safeEvalId}-${timestamp}-${random}${extension}`);
 }
 function formatTimeoutSuffix2(timeoutMs) {
   if (!timeoutMs || timeoutMs <= 0) {
@@ -7529,8 +7312,8 @@ function formatTimeoutSuffix2(timeoutMs) {
 // src/evaluation/providers/codex.ts
 var import_node_crypto3 = require("crypto");
 var import_node_fs6 = require("fs");
-var import_promises14 = require("fs/promises");
-var import_node_path16 = __toESM(require("path"), 1);
+var import_promises13 = require("fs/promises");
+var import_node_path15 = __toESM(require("path"), 1);
 
 // src/evaluation/providers/codex-log-tracker.ts
 var GLOBAL_LOGS_KEY2 = Symbol.for("agentv.codexLogs");
@@ -7765,10 +7548,10 @@ ${basePrompt}` : basePrompt;
   }
   resolveCwd(cwdOverride) {
     if (cwdOverride) {
-      return import_node_path16.default.resolve(cwdOverride);
+      return import_node_path15.default.resolve(cwdOverride);
     }
     if (this.config.cwd) {
-      return import_node_path16.default.resolve(this.config.cwd);
+      return import_node_path15.default.resolve(this.config.cwd);
     }
     return void 0;
   }
@@ -7778,9 +7561,9 @@ ${basePrompt}` : basePrompt;
       return void 0;
     }
     if (this.config.logDir) {
-      return import_node_path16.default.resolve(this.config.logDir);
+      return import_node_path15.default.resolve(this.config.logDir);
     }
-    return import_node_path16.default.join(process.cwd(), ".agentv", "logs", "codex");
+    return import_node_path15.default.join(process.cwd(), ".agentv", "logs", "codex");
   }
   async createStreamLogger(request) {
     const logDir = this.resolveLogDirectory();
@@ -7788,13 +7571,13 @@ ${basePrompt}` : basePrompt;
       return void 0;
     }
     try {
-      await (0, import_promises14.mkdir)(logDir, { recursive: true });
+      await (0, import_promises13.mkdir)(logDir, { recursive: true });
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
       console.warn(`Skipping Codex SDK stream logging (could not create ${logDir}): ${message}`);
       return void 0;
     }
-    const filePath = import_node_path16.default.join(logDir, buildLogFilename3(request, this.targetName));
+    const filePath = import_node_path15.default.join(logDir, buildLogFilename3(request, this.targetName));
     try {
       const logger = await CodexSdkStreamLogger.create({
         filePath,
@@ -7937,8 +7720,8 @@ function formatElapsed3(startedAt) {
 
 // src/evaluation/providers/copilot-cli.ts
 var import_node_crypto5 = require("crypto");
-var import_promises15 = require("fs/promises");
-var import_node_path18 = __toESM(require("path"), 1);
+var import_promises14 = require("fs/promises");
+var import_node_path17 = __toESM(require("path"), 1);
 var import_node_stream = require("stream");
 var import_node_child_process3 = require("child_process");
 var acp = __toESM(require("@agentclientprotocol/sdk"), 1);
@@ -8000,7 +7783,7 @@ function subscribeToCopilotCliLogEntries(listener) {
 var import_node_crypto4 = require("crypto");
 var import_node_fs7 = require("fs");
 var import_node_os2 = require("os");
-var import_node_path17 = __toESM(require("path"), 1);
+var import_node_path16 = __toESM(require("path"), 1);
 var import_node_url2 = require("url");
 var import_meta = {};
 function resolvePlatformCliPath() {
@@ -8025,7 +7808,7 @@ function resolvePlatformCliPath() {
   try {
     const resolved = import_meta.resolve(`${packageName}/package.json`);
     const packageJsonPath = resolved.startsWith("file:") ? (0, import_node_url2.fileURLToPath)(resolved) : resolved;
-    const binaryPath = import_node_path17.default.join(import_node_path17.default.dirname(packageJsonPath), binaryName);
+    const binaryPath = import_node_path16.default.join(import_node_path16.default.dirname(packageJsonPath), binaryName);
     if ((0, import_node_fs7.existsSync)(binaryPath)) {
       return binaryPath;
     }
@@ -8033,7 +7816,7 @@ function resolvePlatformCliPath() {
   }
   let searchDir = process.cwd();
   for (let i = 0; i < 10; i++) {
-    const standardPath = import_node_path17.default.join(
+    const standardPath = import_node_path16.default.join(
       searchDir,
       "node_modules",
       ...packageName.split("/"),
@@ -8042,13 +7825,13 @@ function resolvePlatformCliPath() {
     if ((0, import_node_fs7.existsSync)(standardPath)) {
       return standardPath;
     }
-    const bunDir = import_node_path17.default.join(searchDir, "node_modules", ".bun");
+    const bunDir = import_node_path16.default.join(searchDir, "node_modules", ".bun");
     const prefix = `@github+copilot-${osPart}-${archPart}@`;
     try {
       const entries = (0, import_node_fs7.readdirSync)(bunDir);
       for (const entry of entries) {
         if (entry.startsWith(prefix)) {
-          const candidate = import_node_path17.default.join(
+          const candidate = import_node_path16.default.join(
             bunDir,
             entry,
             "node_modules",
@@ -8063,7 +7846,7 @@ function resolvePlatformCliPath() {
       }
     } catch {
     }
-    const parent = import_node_path17.default.dirname(searchDir);
+    const parent = import_node_path16.default.dirname(searchDir);
     if (parent === searchDir) break;
     searchDir = parent;
   }
@@ -8401,10 +8184,10 @@ var CopilotCliProvider = class {
   }
   resolveCwd(cwdOverride) {
     if (cwdOverride) {
-      return import_node_path18.default.resolve(cwdOverride);
+      return import_node_path17.default.resolve(cwdOverride);
     }
     if (this.config.cwd) {
-      return import_node_path18.default.resolve(this.config.cwd);
+      return import_node_path17.default.resolve(this.config.cwd);
     }
     return void 0;
   }
@@ -8423,9 +8206,9 @@ var CopilotCliProvider = class {
       return void 0;
     }
     if (this.config.logDir) {
-      return import_node_path18.default.resolve(this.config.logDir);
+      return import_node_path17.default.resolve(this.config.logDir);
     }
-    return import_node_path18.default.join(process.cwd(), ".agentv", "logs", "copilot-cli");
+    return import_node_path17.default.join(process.cwd(), ".agentv", "logs", "copilot-cli");
   }
   async createStreamLogger(request) {
     const logDir = this.resolveLogDirectory();
@@ -8433,13 +8216,13 @@ var CopilotCliProvider = class {
       return void 0;
     }
     try {
-      await (0, import_promises15.mkdir)(logDir, { recursive: true });
+      await (0, import_promises14.mkdir)(logDir, { recursive: true });
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
       console.warn(`Skipping Copilot CLI stream logging (could not create ${logDir}): ${message}`);
       return void 0;
     }
-    const filePath = import_node_path18.default.join(logDir, buildLogFilename4(request, this.targetName, "copilot-cli"));
+    const filePath = import_node_path17.default.join(logDir, buildLogFilename4(request, this.targetName, "copilot-cli"));
     try {
       const logger = await CopilotStreamLogger.create(
         {
@@ -8533,8 +8316,8 @@ function summarizeAcpEvent(eventType, data) {
 
 // src/evaluation/providers/copilot-sdk.ts
 var import_node_crypto6 = require("crypto");
-var import_promises16 = require("fs/promises");
-var import_node_path19 = __toESM(require("path"), 1);
+var import_promises15 = require("fs/promises");
+var import_node_path18 = __toESM(require("path"), 1);
 
 // src/evaluation/providers/copilot-sdk-log-tracker.ts
 var GLOBAL_LOGS_KEY4 = Symbol.for("agentv.copilotSdkLogs");
@@ -8813,10 +8596,10 @@ var CopilotSdkProvider = class {
   }
   resolveCwd(cwdOverride) {
     if (cwdOverride) {
-      return import_node_path19.default.resolve(cwdOverride);
+      return import_node_path18.default.resolve(cwdOverride);
     }
     if (this.config.cwd) {
-      return import_node_path19.default.resolve(this.config.cwd);
+      return import_node_path18.default.resolve(this.config.cwd);
     }
     return void 0;
   }
@@ -8825,9 +8608,9 @@ var CopilotSdkProvider = class {
       return void 0;
     }
     if (this.config.logDir) {
-      return import_node_path19.default.resolve(this.config.logDir);
+      return import_node_path18.default.resolve(this.config.logDir);
     }
-    return import_node_path19.default.join(process.cwd(), ".agentv", "logs", "copilot-sdk");
+    return import_node_path18.default.join(process.cwd(), ".agentv", "logs", "copilot-sdk");
   }
   async createStreamLogger(request) {
     const logDir = this.resolveLogDirectory();
@@ -8835,13 +8618,13 @@ var CopilotSdkProvider = class {
       return void 0;
     }
     try {
-      await (0, import_promises16.mkdir)(logDir, { recursive: true });
+      await (0, import_promises15.mkdir)(logDir, { recursive: true });
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
       console.warn(`Skipping Copilot SDK stream logging (could not create ${logDir}): ${message}`);
       return void 0;
     }
-    const filePath = import_node_path19.default.join(logDir, buildLogFilename4(request, this.targetName, "copilot-sdk"));
+    const filePath = import_node_path18.default.join(logDir, buildLogFilename4(request, this.targetName, "copilot-sdk"));
     try {
       const logger = await CopilotStreamLogger.create(
         {
@@ -8918,8 +8701,7 @@ var MockProvider = class {
     return {
       output: [{ role: "assistant", content: this.cannedResponse }],
       raw: {
-        question: request.question,
-        guidelines: request.guidelines
+        question: request.question
       }
     };
   }
@@ -9195,9 +8977,9 @@ function extractToolCalls3(content, toolTrackers, completedToolResults) {
 var import_node_child_process4 = require("child_process");
 var import_node_crypto7 = require("crypto");
 var import_node_fs8 = require("fs");
-var import_promises17 = require("fs/promises");
+var import_promises16 = require("fs/promises");
 var import_node_os3 = require("os");
-var import_node_path20 = __toESM(require("path"), 1);
+var import_node_path19 = __toESM(require("path"), 1);
 
 // src/evaluation/providers/pi-log-tracker.ts
 var GLOBAL_LOGS_KEY5 = Symbol.for("agentv.piLogs");
@@ -9278,8 +9060,8 @@ var PiCodingAgentProvider = class {
     const workspaceRoot = await this.createWorkspace();
     const logger = await this.createStreamLogger(request).catch(() => void 0);
     try {
-      const promptFile = import_node_path20.default.join(workspaceRoot, PROMPT_FILENAME);
-      await (0, import_promises17.writeFile)(promptFile, request.question, "utf8");
+      const promptFile = import_node_path19.default.join(workspaceRoot, PROMPT_FILENAME);
+      await (0, import_promises16.writeFile)(promptFile, request.question, "utf8");
       const args = this.buildPiArgs(request.question, inputFiles, request.captureFileChanges);
       const cwd = this.resolveCwd(workspaceRoot, request.cwd);
       const result = await this.executePi(args, cwd, request.signal, logger);
@@ -9340,12 +9122,12 @@ var PiCodingAgentProvider = class {
   }
   resolveCwd(workspaceRoot, cwdOverride) {
     if (cwdOverride) {
-      return import_node_path20.default.resolve(cwdOverride);
+      return import_node_path19.default.resolve(cwdOverride);
     }
     if (!this.config.cwd) {
       return workspaceRoot;
     }
-    return import_node_path20.default.resolve(this.config.cwd);
+    return import_node_path19.default.resolve(this.config.cwd);
   }
   buildPiArgs(prompt, inputFiles, _captureFileChanges) {
     const args = [];
@@ -9434,19 +9216,19 @@ ${prompt}` : prompt;
     return env;
   }
   async createWorkspace() {
-    return await (0, import_promises17.mkdtemp)(import_node_path20.default.join((0, import_node_os3.tmpdir)(), WORKSPACE_PREFIX));
+    return await (0, import_promises16.mkdtemp)(import_node_path19.default.join((0, import_node_os3.tmpdir)(), WORKSPACE_PREFIX));
   }
   async cleanupWorkspace(workspaceRoot) {
     try {
-      await (0, import_promises17.rm)(workspaceRoot, { recursive: true, force: true });
+      await (0, import_promises16.rm)(workspaceRoot, { recursive: true, force: true });
     } catch {
     }
   }
   resolveLogDirectory() {
     if (this.config.logDir) {
-      return import_node_path20.default.resolve(this.config.logDir);
+      return import_node_path19.default.resolve(this.config.logDir);
     }
-    return import_node_path20.default.join(process.cwd(), ".agentv", "logs", "pi-coding-agent");
+    return import_node_path19.default.join(process.cwd(), ".agentv", "logs", "pi-coding-agent");
   }
   async createStreamLogger(request) {
     const logDir = this.resolveLogDirectory();
@@ -9454,13 +9236,13 @@ ${prompt}` : prompt;
       return void 0;
     }
     try {
-      await (0, import_promises17.mkdir)(logDir, { recursive: true });
+      await (0, import_promises16.mkdir)(logDir, { recursive: true });
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
       console.warn(`Skipping Pi stream logging (could not create ${logDir}): ${message}`);
       return void 0;
     }
-    const filePath = import_node_path20.default.join(logDir, buildLogFilename5(request, this.targetName));
+    const filePath = import_node_path19.default.join(logDir, buildLogFilename5(request, this.targetName));
     try {
       const logger = await PiStreamLogger.create({
         filePath,
@@ -9959,7 +9741,7 @@ var ProviderRegistry = class {
 };
 
 // src/evaluation/providers/targets.ts
-var import_node_path21 = __toESM(require("path"), 1);
+var import_node_path20 = __toESM(require("path"), 1);
 var import_zod3 = require("zod");
 var CliHealthcheckHttpInputSchema = import_zod3.z.object({
   url: import_zod3.z.string().min(1, "healthcheck URL is required"),
@@ -10056,11 +9838,11 @@ function normalizeCliHealthcheck(input, env, targetName, evalFilePath) {
     allowLiteral: true,
     optionalEnv: true
   });
-  if (cwd && evalFilePath && !import_node_path21.default.isAbsolute(cwd)) {
-    cwd = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), cwd);
+  if (cwd && evalFilePath && !import_node_path20.default.isAbsolute(cwd)) {
+    cwd = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), cwd);
   }
   if (!cwd && evalFilePath) {
-    cwd = import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath));
+    cwd = import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath));
   }
   return {
     command,
@@ -10083,15 +9865,15 @@ function normalizeCliTargetInput(input, env, evalFilePath) {
       optionalEnv: true
     }
   );
-  if (workspaceTemplate && evalFilePath && !import_node_path21.default.isAbsolute(workspaceTemplate)) {
-    workspaceTemplate = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), workspaceTemplate);
+  if (workspaceTemplate && evalFilePath && !import_node_path20.default.isAbsolute(workspaceTemplate)) {
+    workspaceTemplate = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), workspaceTemplate);
   }
   let cwd = resolveOptionalString(input.cwd, env, `${targetName} working directory`, {
     allowLiteral: true,
     optionalEnv: true
   });
-  if (cwd && evalFilePath && !import_node_path21.default.isAbsolute(cwd)) {
-    cwd = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), cwd);
+  if (cwd && evalFilePath && !import_node_path20.default.isAbsolute(cwd)) {
+    cwd = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), cwd);
   }
   if (cwd && workspaceTemplate) {
     throw new Error(
@@ -10099,7 +9881,7 @@ function normalizeCliTargetInput(input, env, evalFilePath) {
     );
   }
   if (!cwd && !workspaceTemplate && evalFilePath) {
-    cwd = import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath));
+    cwd = import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath));
   }
   const timeoutSeconds = input.timeout_seconds ?? input.timeoutSeconds;
   const timeoutMs = timeoutSeconds !== void 0 ? Math.floor(timeoutSeconds * 1e3) : void 0;
@@ -10122,7 +9904,6 @@ function normalizeCliTargetInput(input, env, evalFilePath) {
 var CLI_PLACEHOLDERS = /* @__PURE__ */ new Set([
   "PROMPT",
   "PROMPT_FILE",
-  "GUIDELINES",
   "EVAL_ID",
   "ATTEMPT",
   "FILES",
@@ -10517,8 +10298,8 @@ function resolveCodexConfig(target, env, evalFilePath) {
       optionalEnv: true
     }
   );
-  if (workspaceTemplate && evalFilePath && !import_node_path21.default.isAbsolute(workspaceTemplate)) {
-    workspaceTemplate = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), workspaceTemplate);
+  if (workspaceTemplate && evalFilePath && !import_node_path20.default.isAbsolute(workspaceTemplate)) {
+    workspaceTemplate = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), workspaceTemplate);
   }
   if (cwd && workspaceTemplate) {
     throw new Error(
@@ -10602,8 +10383,8 @@ function resolveCopilotSdkConfig(target, env, evalFilePath) {
       optionalEnv: true
     }
   );
-  if (workspaceTemplate && evalFilePath && !import_node_path21.default.isAbsolute(workspaceTemplate)) {
-    workspaceTemplate = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), workspaceTemplate);
+  if (workspaceTemplate && evalFilePath && !import_node_path20.default.isAbsolute(workspaceTemplate)) {
+    workspaceTemplate = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), workspaceTemplate);
   }
   if (cwd && workspaceTemplate) {
     throw new Error(
@@ -10667,8 +10448,8 @@ function resolveCopilotCliConfig(target, env, evalFilePath) {
       optionalEnv: true
     }
   );
-  if (workspaceTemplate && evalFilePath && !import_node_path21.default.isAbsolute(workspaceTemplate)) {
-    workspaceTemplate = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), workspaceTemplate);
+  if (workspaceTemplate && evalFilePath && !import_node_path20.default.isAbsolute(workspaceTemplate)) {
+    workspaceTemplate = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), workspaceTemplate);
   }
   if (cwd && workspaceTemplate) {
     throw new Error(
@@ -10758,8 +10539,8 @@ function resolvePiCodingAgentConfig(target, env, evalFilePath) {
       optionalEnv: true
     }
   );
-  if (workspaceTemplate && evalFilePath && !import_node_path21.default.isAbsolute(workspaceTemplate)) {
-    workspaceTemplate = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), workspaceTemplate);
+  if (workspaceTemplate && evalFilePath && !import_node_path20.default.isAbsolute(workspaceTemplate)) {
+    workspaceTemplate = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), workspaceTemplate);
   }
   if (cwd && workspaceTemplate) {
     throw new Error(
@@ -10847,8 +10628,8 @@ function resolveClaudeConfig(target, env, evalFilePath) {
       optionalEnv: true
     }
   );
-  if (workspaceTemplate && evalFilePath && !import_node_path21.default.isAbsolute(workspaceTemplate)) {
-    workspaceTemplate = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), workspaceTemplate);
+  if (workspaceTemplate && evalFilePath && !import_node_path20.default.isAbsolute(workspaceTemplate)) {
+    workspaceTemplate = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), workspaceTemplate);
   }
   if (cwd && workspaceTemplate) {
     throw new Error(
@@ -10906,8 +10687,8 @@ function resolveVSCodeConfig(target, env, insiders, evalFilePath) {
       optionalEnv: true
     }
   ) : void 0;
-  if (workspaceTemplate && evalFilePath && !import_node_path21.default.isAbsolute(workspaceTemplate)) {
-    workspaceTemplate = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), workspaceTemplate);
+  if (workspaceTemplate && evalFilePath && !import_node_path20.default.isAbsolute(workspaceTemplate)) {
+    workspaceTemplate = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), workspaceTemplate);
   }
   const executableSource = target.executable;
   const waitSource = target.wait;
@@ -10948,8 +10729,8 @@ function resolveCliConfig(target, env, evalFilePath) {
   const parseResult = CliTargetInputSchema.safeParse(target, { errorMap: cliErrorMap });
   if (!parseResult.success) {
     const firstError = parseResult.error.errors[0];
-    const path48 = firstError?.path.join(".") || "";
-    const prefix = path48 ? `${target.name} ${path48}: ` : `${target.name}: `;
+    const path47 = firstError?.path.join(".") || "";
+    const prefix = path47 ? `${target.name} ${path47}: ` : `${target.name}: `;
     throw new Error(`${prefix}${firstError?.message}`);
   }
   const normalized = normalizeCliTargetInput(parseResult.data, env, evalFilePath);
@@ -10970,11 +10751,11 @@ function resolveDiscoveredProviderConfig(target, providerKind, env, evalFilePath
     allowLiteral: true,
     optionalEnv: true
   });
-  if (cwd && evalFilePath && !import_node_path21.default.isAbsolute(cwd)) {
-    cwd = import_node_path21.default.resolve(import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath)), cwd);
+  if (cwd && evalFilePath && !import_node_path20.default.isAbsolute(cwd)) {
+    cwd = import_node_path20.default.resolve(import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath)), cwd);
   }
   if (!cwd && evalFilePath) {
-    cwd = import_node_path21.default.dirname(import_node_path21.default.resolve(evalFilePath));
+    cwd = import_node_path20.default.dirname(import_node_path20.default.resolve(evalFilePath));
   }
   return {
     command,
@@ -11158,40 +10939,40 @@ function resolveOptionalNumberArray(source, description) {
 
 // src/evaluation/providers/vscode-provider.ts
 var import_node_child_process6 = require("child_process");
-var import_promises24 = require("fs/promises");
-var import_node_path33 = __toESM(require("path"), 1);
+var import_promises23 = require("fs/promises");
+var import_node_path32 = __toESM(require("path"), 1);
 var import_node_util3 = require("util");
 
 // src/evaluation/providers/vscode/dispatch/agentDispatch.ts
-var import_promises22 = require("fs/promises");
-var import_node_path31 = __toESM(require("path"), 1);
+var import_promises21 = require("fs/promises");
+var import_node_path30 = __toESM(require("path"), 1);
 
 // src/evaluation/providers/vscode/utils/fs.ts
 var import_node_fs9 = require("fs");
-var import_promises18 = require("fs/promises");
-var import_node_path22 = __toESM(require("path"), 1);
+var import_promises17 = require("fs/promises");
+var import_node_path21 = __toESM(require("path"), 1);
 async function pathExists(target) {
   try {
-    await (0, import_promises18.access)(target, import_node_fs9.constants.F_OK);
+    await (0, import_promises17.access)(target, import_node_fs9.constants.F_OK);
     return true;
   } catch {
     return false;
   }
 }
 async function ensureDir(target) {
-  await (0, import_promises18.mkdir)(target, { recursive: true });
+  await (0, import_promises17.mkdir)(target, { recursive: true });
 }
 async function readDirEntries(target) {
-  const entries = await (0, import_promises18.readdir)(target, { withFileTypes: true });
+  const entries = await (0, import_promises17.readdir)(target, { withFileTypes: true });
   return entries.map((entry) => ({
     name: entry.name,
-    absolutePath: import_node_path22.default.join(target, entry.name),
+    absolutePath: import_node_path21.default.join(target, entry.name),
     isDirectory: entry.isDirectory()
   }));
 }
 async function removeIfExists(target) {
   try {
-    await (0, import_promises18.rm)(target, { force: true, recursive: false });
+    await (0, import_promises17.rm)(target, { force: true, recursive: false });
   } catch (error) {
     if (error.code !== "ENOENT") {
       throw error;
@@ -11200,9 +10981,9 @@ async function removeIfExists(target) {
 }
 
 // src/evaluation/providers/vscode/utils/path.ts
-var import_node_path23 = __toESM(require("path"), 1);
+var import_node_path22 = __toESM(require("path"), 1);
 function pathToFileUri2(filePath) {
-  const absolutePath = import_node_path23.default.isAbsolute(filePath) ? filePath : import_node_path23.default.resolve(filePath);
+  const absolutePath = import_node_path22.default.isAbsolute(filePath) ? filePath : import_node_path22.default.resolve(filePath);
   const normalizedPath = absolutePath.replace(/\\/g, "/");
   if (/^[a-zA-Z]:\//.test(normalizedPath)) {
     return `file:///${normalizedPath}`;
@@ -11211,7 +10992,7 @@ function pathToFileUri2(filePath) {
 }
 
 // src/evaluation/providers/vscode/dispatch/promptBuilder.ts
-var import_node_path24 = __toESM(require("path"), 1);
+var import_node_path23 = __toESM(require("path"), 1);
 
 // src/evaluation/providers/vscode/utils/template.ts
 function renderTemplate2(content, variables) {
@@ -11303,8 +11084,8 @@ function createBatchRequestPrompt(userQuery, responseFileTmp, responseFileFinal,
   });
 }
 function createBatchOrchestratorPrompt(requestFiles, responseFiles, templateContent) {
-  const requestLines = requestFiles.map((file, index) => `${index + 1}. messages/${import_node_path24.default.basename(file)}`).join("\n");
-  const responseList = responseFiles.map((file) => `"${import_node_path24.default.basename(file)}"`).join(", ");
+  const requestLines = requestFiles.map((file, index) => `${index + 1}. messages/${import_node_path23.default.basename(file)}`).join("\n");
+  const responseList = responseFiles.map((file) => `"${import_node_path23.default.basename(file)}"`).join(", ");
   return renderTemplate2(templateContent, {
     requestFiles: requestLines,
     responseList
@@ -11312,8 +11093,8 @@ function createBatchOrchestratorPrompt(requestFiles, responseFiles, templateCont
 }
 
 // src/evaluation/providers/vscode/dispatch/responseWaiter.ts
-var import_promises19 = require("fs/promises");
-var import_node_path25 = __toESM(require("path"), 1);
+var import_promises18 = require("fs/promises");
+var import_node_path24 = __toESM(require("path"), 1);
 
 // src/evaluation/providers/vscode/utils/time.ts
 function sleep2(ms) {
@@ -11351,7 +11132,7 @@ async function waitForResponseOutput(responseFileFinal, pollInterval = 1e3, sile
   const maxAttempts = 10;
   while (attempts < maxAttempts) {
     try {
-      const content = await (0, import_promises19.readFile)(responseFileFinal, { encoding: "utf8" });
+      const content = await (0, import_promises18.readFile)(responseFileFinal, { encoding: "utf8" });
       if (!silent) {
         process.stdout.write(`${content}
 `);
@@ -11372,7 +11153,7 @@ async function waitForResponseOutput(responseFileFinal, pollInterval = 1e3, sile
 }
 async function waitForBatchResponses(responseFilesFinal, pollInterval = 1e3, silent = false, timeoutMs = DEFAULT_TIMEOUT_MS) {
   if (!silent) {
-    const fileList = responseFilesFinal.map((file) => import_node_path25.default.basename(file)).join(", ");
+    const fileList = responseFilesFinal.map((file) => import_node_path24.default.basename(file)).join(", ");
     console.error(`waiting for ${responseFilesFinal.length} batch response(s): ${fileList}`);
   }
   const deadline = Date.now() + timeoutMs;
@@ -11381,7 +11162,7 @@ async function waitForBatchResponses(responseFilesFinal, pollInterval = 1e3, sil
     while (pending.size > 0) {
       if (Date.now() >= deadline) {
         if (!silent) {
-          const remaining = [...pending].map((f) => import_node_path25.default.basename(f)).join(", ");
+          const remaining = [...pending].map((f) => import_node_path24.default.basename(f)).join(", ");
           console.error(
             `error: timed out after ${Math.round(timeoutMs / 1e3)}s waiting for batch responses. Still pending: ${remaining}`
           );
@@ -11408,7 +11189,7 @@ async function waitForBatchResponses(responseFilesFinal, pollInterval = 1e3, sil
     const maxAttempts = 10;
     while (attempts < maxAttempts) {
       try {
-        const content = await (0, import_promises19.readFile)(file, { encoding: "utf8" });
+        const content = await (0, import_promises18.readFile)(file, { encoding: "utf8" });
         if (!silent) {
           process.stdout.write(`${content}
 `);
@@ -11431,16 +11212,16 @@ async function waitForBatchResponses(responseFilesFinal, pollInterval = 1e3, sil
 
 // src/evaluation/providers/vscode/dispatch/vscodeProcess.ts
 var import_node_child_process5 = require("child_process");
-var import_promises20 = require("fs/promises");
-var import_node_path28 = __toESM(require("path"), 1);
+var import_promises19 = require("fs/promises");
+var import_node_path27 = __toESM(require("path"), 1);
 var import_node_util2 = require("util");
 
 // src/evaluation/providers/vscode/dispatch/constants.ts
-var import_node_path27 = __toESM(require("path"), 1);
+var import_node_path26 = __toESM(require("path"), 1);
 
 // src/paths.ts
 var import_node_os4 = __toESM(require("os"), 1);
-var import_node_path26 = __toESM(require("path"), 1);
+var import_node_path25 = __toESM(require("path"), 1);
 var logged = false;
 function getAgentvHome() {
   const envHome = process.env.AGENTV_HOME;
@@ -11451,19 +11232,19 @@ function getAgentvHome() {
     }
     return envHome;
   }
-  return import_node_path26.default.join(import_node_os4.default.homedir(), ".agentv");
+  return import_node_path25.default.join(import_node_os4.default.homedir(), ".agentv");
 }
 function getWorkspacesRoot() {
-  return import_node_path26.default.join(getAgentvHome(), "workspaces");
+  return import_node_path25.default.join(getAgentvHome(), "workspaces");
 }
 function getSubagentsRoot() {
-  return import_node_path26.default.join(getAgentvHome(), "subagents");
+  return import_node_path25.default.join(getAgentvHome(), "subagents");
 }
 function getTraceStateRoot() {
-  return import_node_path26.default.join(getAgentvHome(), "trace-state");
+  return import_node_path25.default.join(getAgentvHome(), "trace-state");
 }
 function getWorkspacePoolRoot() {
-  return import_node_path26.default.join(getAgentvHome(), "workspace-pool");
+  return import_node_path25.default.join(getAgentvHome(), "workspace-pool");
 }
 
 // src/evaluation/providers/vscode/dispatch/constants.ts
@@ -11471,7 +11252,7 @@ var DEFAULT_LOCK_NAME = "subagent.lock";
 var DEFAULT_ALIVE_FILENAME = ".alive";
 function getDefaultSubagentRoot(vscodeCmd = "code") {
   const folder = vscodeCmd === "code-insiders" ? "vscode-insiders-agents" : "vscode-agents";
-  return import_node_path27.default.join(getSubagentsRoot(), folder);
+  return import_node_path26.default.join(getSubagentsRoot(), folder);
 }
 var DEFAULT_SUBAGENT_ROOT = getDefaultSubagentRoot();
 
@@ -11538,12 +11319,12 @@ async function ensureWorkspaceFocused(workspacePath, workspaceName, subagentDir,
     await raceSpawnError(child);
     return true;
   }
-  const aliveFile = import_node_path28.default.join(subagentDir, DEFAULT_ALIVE_FILENAME);
+  const aliveFile = import_node_path27.default.join(subagentDir, DEFAULT_ALIVE_FILENAME);
   await removeIfExists(aliveFile);
-  const githubAgentsDir = import_node_path28.default.join(subagentDir, ".github", "agents");
-  await (0, import_promises20.mkdir)(githubAgentsDir, { recursive: true });
-  const wakeupDst = import_node_path28.default.join(githubAgentsDir, "wakeup.md");
-  await (0, import_promises20.writeFile)(wakeupDst, DEFAULT_WAKEUP_CONTENT, "utf8");
+  const githubAgentsDir = import_node_path27.default.join(subagentDir, ".github", "agents");
+  await (0, import_promises19.mkdir)(githubAgentsDir, { recursive: true });
+  const wakeupDst = import_node_path27.default.join(githubAgentsDir, "wakeup.md");
+  await (0, import_promises19.writeFile)(wakeupDst, DEFAULT_WAKEUP_CONTENT, "utf8");
   const workspaceChild = spawnVsCode(vscodeCmd, [workspacePath], {
     label: "open-workspace"
   });
@@ -11555,7 +11336,7 @@ async function ensureWorkspaceFocused(workspacePath, workspaceName, subagentDir,
     "chat",
     "-m",
     wakeupChatId,
-    `create a file named .alive in the ${import_node_path28.default.basename(subagentDir)} folder`
+    `create a file named .alive in the ${import_node_path27.default.basename(subagentDir)} folder`
   ];
   const wakeupChild = spawnVsCode(vscodeCmd, chatArgs, { label: "send-wakeup-chat" });
   await raceSpawnError(wakeupChild);
@@ -11570,27 +11351,27 @@ async function ensureWorkspaceFocused(workspacePath, workspaceName, subagentDir,
   return true;
 }
 async function launchVsCodeWithChat(subagentDir, chatId, attachmentPaths, requestInstructions, timestamp, vscodeCmd) {
-  const workspacePath = import_node_path28.default.join(subagentDir, `${import_node_path28.default.basename(subagentDir)}.code-workspace`);
-  const messagesDir = import_node_path28.default.join(subagentDir, "messages");
-  await (0, import_promises20.mkdir)(messagesDir, { recursive: true });
-  const reqFile = import_node_path28.default.join(messagesDir, `${timestamp}_req.md`);
-  await (0, import_promises20.writeFile)(reqFile, requestInstructions, { encoding: "utf8" });
+  const workspacePath = import_node_path27.default.join(subagentDir, `${import_node_path27.default.basename(subagentDir)}.code-workspace`);
+  const messagesDir = import_node_path27.default.join(subagentDir, "messages");
+  await (0, import_promises19.mkdir)(messagesDir, { recursive: true });
+  const reqFile = import_node_path27.default.join(messagesDir, `${timestamp}_req.md`);
+  await (0, import_promises19.writeFile)(reqFile, requestInstructions, { encoding: "utf8" });
   const reqUri = pathToFileUri2(reqFile);
   const chatArgs = ["-r", "chat", "-m", chatId];
   for (const attachment of attachmentPaths) {
     chatArgs.push("-a", attachment);
   }
   chatArgs.push("-a", reqFile);
-  chatArgs.push(`Follow instructions in [${import_node_path28.default.basename(reqFile)}](${reqUri})`);
+  chatArgs.push(`Follow instructions in [${import_node_path27.default.basename(reqFile)}](${reqUri})`);
   const workspaceReady = await ensureWorkspaceFocused(
     workspacePath,
-    import_node_path28.default.basename(subagentDir),
+    import_node_path27.default.basename(subagentDir),
     subagentDir,
     vscodeCmd
   );
   if (!workspaceReady) {
     throw new Error(
-      `VS Code workspace '${import_node_path28.default.basename(subagentDir)}' failed to become ready within the timeout. Check that '${vscodeCmd}' can open workspaces.`
+      `VS Code workspace '${import_node_path27.default.basename(subagentDir)}' failed to become ready within the timeout. Check that '${vscodeCmd}' can open workspaces.`
     );
   }
   await sleep2(500);
@@ -11598,9 +11379,9 @@ async function launchVsCodeWithChat(subagentDir, chatId, attachmentPaths, reques
   await raceSpawnError(child);
 }
 async function launchVsCodeWithBatchChat(subagentDir, chatId, attachmentPaths, chatInstruction, vscodeCmd) {
-  const workspacePath = import_node_path28.default.join(subagentDir, `${import_node_path28.default.basename(subagentDir)}.code-workspace`);
-  const messagesDir = import_node_path28.default.join(subagentDir, "messages");
-  await (0, import_promises20.mkdir)(messagesDir, { recursive: true });
+  const workspacePath = import_node_path27.default.join(subagentDir, `${import_node_path27.default.basename(subagentDir)}.code-workspace`);
+  const messagesDir = import_node_path27.default.join(subagentDir, "messages");
+  await (0, import_promises19.mkdir)(messagesDir, { recursive: true });
   const chatArgs = ["-r", "chat", "-m", chatId];
   for (const attachment of attachmentPaths) {
     chatArgs.push("-a", attachment);
@@ -11608,13 +11389,13 @@ async function launchVsCodeWithBatchChat(subagentDir, chatId, attachmentPaths, c
   chatArgs.push(chatInstruction);
   const workspaceReady = await ensureWorkspaceFocused(
     workspacePath,
-    import_node_path28.default.basename(subagentDir),
+    import_node_path27.default.basename(subagentDir),
     subagentDir,
     vscodeCmd
   );
   if (!workspaceReady) {
     throw new Error(
-      `VS Code workspace '${import_node_path28.default.basename(subagentDir)}' failed to become ready within the timeout. Check that '${vscodeCmd}' can open workspaces.`
+      `VS Code workspace '${import_node_path27.default.basename(subagentDir)}' failed to become ready within the timeout. Check that '${vscodeCmd}' can open workspaces.`
     );
   }
   await sleep2(500);
@@ -11623,11 +11404,11 @@ async function launchVsCodeWithBatchChat(subagentDir, chatId, attachmentPaths, c
 }
 
 // src/evaluation/providers/vscode/dispatch/workspaceManager.ts
-var import_promises21 = require("fs/promises");
-var import_node_path30 = __toESM(require("path"), 1);
+var import_promises20 = require("fs/promises");
+var import_node_path29 = __toESM(require("path"), 1);
 
 // src/evaluation/providers/vscode/utils/workspace.ts
-var import_node_path29 = __toESM(require("path"), 1);
+var import_node_path28 = __toESM(require("path"), 1);
 var import_json5 = __toESM(require("json5"), 1);
 function transformWorkspacePaths(workspaceContent, templateDir) {
   let workspace;
@@ -11644,10 +11425,10 @@ function transformWorkspacePaths(workspaceContent, templateDir) {
   }
   const transformedFolders = workspace.folders.map((folder) => {
     const folderPath = folder.path;
-    if (import_node_path29.default.isAbsolute(folderPath)) {
+    if (import_node_path28.default.isAbsolute(folderPath)) {
       return folder;
     }
-    const absolutePath = import_node_path29.default.resolve(templateDir, folderPath);
+    const absolutePath = import_node_path28.default.resolve(templateDir, folderPath);
     return {
       ...folder,
       path: absolutePath
@@ -11669,19 +11450,19 @@ function transformWorkspacePaths(workspaceContent, templateDir) {
       if (locationMap && typeof locationMap === "object") {
         const transformedMap = {};
         for (const [locationPath, value] of Object.entries(locationMap)) {
-          const isAbsolute = import_node_path29.default.isAbsolute(locationPath);
+          const isAbsolute = import_node_path28.default.isAbsolute(locationPath);
           if (isAbsolute) {
             transformedMap[locationPath] = value;
           } else {
             const firstGlobIndex = locationPath.search(/[*]/);
             if (firstGlobIndex === -1) {
-              const resolvedPath = import_node_path29.default.resolve(templateDir, locationPath).replace(/\\/g, "/");
+              const resolvedPath = import_node_path28.default.resolve(templateDir, locationPath).replace(/\\/g, "/");
               transformedMap[resolvedPath] = value;
             } else {
               const basePathEnd = locationPath.lastIndexOf("/", firstGlobIndex);
               const basePath = basePathEnd !== -1 ? locationPath.substring(0, basePathEnd) : ".";
               const patternPath = locationPath.substring(basePathEnd !== -1 ? basePathEnd : 0);
-              const resolvedPath = (import_node_path29.default.resolve(templateDir, basePath) + patternPath).replace(
+              const resolvedPath = (import_node_path28.default.resolve(templateDir, basePath) + patternPath).replace(
                 /\\/g,
                 "/"
               );
@@ -11722,7 +11503,7 @@ async function findUnlockedSubagent(subagentRoot) {
     number: Number.parseInt(entry.name.split("-")[1] ?? "", 10)
   })).filter((entry) => Number.isInteger(entry.number)).sort((a, b) => a.number - b.number);
   for (const subagent of subagents) {
-    const lockFile = import_node_path30.default.join(subagent.absolutePath, DEFAULT_LOCK_NAME);
+    const lockFile = import_node_path29.default.join(subagent.absolutePath, DEFAULT_LOCK_NAME);
     if (!await pathExists(lockFile)) {
       return subagent.absolutePath;
     }
@@ -11732,26 +11513,26 @@ async function findUnlockedSubagent(subagentRoot) {
 async function copyAgentConfig(subagentDir, workspaceTemplate, cwd) {
   let workspaceContent;
   if (workspaceTemplate) {
-    const workspaceSrc = import_node_path30.default.resolve(workspaceTemplate);
+    const workspaceSrc = import_node_path29.default.resolve(workspaceTemplate);
     if (!await pathExists(workspaceSrc)) {
       throw new Error(`workspace template not found: ${workspaceSrc}`);
     }
-    const stats = await (0, import_promises21.stat)(workspaceSrc);
+    const stats = await (0, import_promises20.stat)(workspaceSrc);
     if (!stats.isFile()) {
       throw new Error(`workspace template must be a file, not a directory: ${workspaceSrc}`);
     }
-    const templateText = await (0, import_promises21.readFile)(workspaceSrc, "utf8");
+    const templateText = await (0, import_promises20.readFile)(workspaceSrc, "utf8");
     workspaceContent = JSON.parse(templateText);
   } else {
     workspaceContent = DEFAULT_WORKSPACE_TEMPLATE;
   }
-  const workspaceName = `${import_node_path30.default.basename(subagentDir)}.code-workspace`;
-  const workspaceDst = import_node_path30.default.join(subagentDir, workspaceName);
-  const templateDir = workspaceTemplate ? import_node_path30.default.dirname(import_node_path30.default.resolve(workspaceTemplate)) : subagentDir;
+  const workspaceName = `${import_node_path29.default.basename(subagentDir)}.code-workspace`;
+  const workspaceDst = import_node_path29.default.join(subagentDir, workspaceName);
+  const templateDir = workspaceTemplate ? import_node_path29.default.dirname(import_node_path29.default.resolve(workspaceTemplate)) : subagentDir;
   const workspaceJson = JSON.stringify(workspaceContent, null, 2);
   let transformedContent = transformWorkspacePaths(workspaceJson, templateDir);
   if (cwd) {
-    const absCwd = import_node_path30.default.resolve(cwd);
+    const absCwd = import_node_path29.default.resolve(cwd);
     const parsed = JSON.parse(transformedContent);
     const alreadyPresent = parsed.folders.some((f) => f.path === absCwd);
     if (!alreadyPresent) {
@@ -11759,36 +11540,36 @@ async function copyAgentConfig(subagentDir, workspaceTemplate, cwd) {
       transformedContent = JSON.stringify(parsed, null, 2);
     }
   }
-  await (0, import_promises21.writeFile)(workspaceDst, transformedContent, "utf8");
-  const messagesDir = import_node_path30.default.join(subagentDir, "messages");
-  await (0, import_promises21.mkdir)(messagesDir, { recursive: true });
+  await (0, import_promises20.writeFile)(workspaceDst, transformedContent, "utf8");
+  const messagesDir = import_node_path29.default.join(subagentDir, "messages");
+  await (0, import_promises20.mkdir)(messagesDir, { recursive: true });
   return { workspace: workspaceDst, messagesDir };
 }
 async function createSubagentLock(subagentDir) {
-  const messagesDir = import_node_path30.default.join(subagentDir, "messages");
+  const messagesDir = import_node_path29.default.join(subagentDir, "messages");
   if (await pathExists(messagesDir)) {
-    const files = await (0, import_promises21.readdir)(messagesDir);
+    const files = await (0, import_promises20.readdir)(messagesDir);
     await Promise.all(
       files.map(async (file) => {
-        const target = import_node_path30.default.join(messagesDir, file);
+        const target = import_node_path29.default.join(messagesDir, file);
         await removeIfExists(target);
       })
     );
   }
-  const githubAgentsDir = import_node_path30.default.join(subagentDir, ".github", "agents");
+  const githubAgentsDir = import_node_path29.default.join(subagentDir, ".github", "agents");
   if (await pathExists(githubAgentsDir)) {
-    const agentFiles = await (0, import_promises21.readdir)(githubAgentsDir);
+    const agentFiles = await (0, import_promises20.readdir)(githubAgentsDir);
     const preservedFiles = /* @__PURE__ */ new Set(["wakeup.md", "subagent.md"]);
     await Promise.all(
-      agentFiles.filter((file) => file.endsWith(".md") && !preservedFiles.has(file)).map((file) => removeIfExists(import_node_path30.default.join(githubAgentsDir, file)))
+      agentFiles.filter((file) => file.endsWith(".md") && !preservedFiles.has(file)).map((file) => removeIfExists(import_node_path29.default.join(githubAgentsDir, file)))
     );
   }
-  const lockFile = import_node_path30.default.join(subagentDir, DEFAULT_LOCK_NAME);
-  await (0, import_promises21.writeFile)(lockFile, "", { encoding: "utf8" });
+  const lockFile = import_node_path29.default.join(subagentDir, DEFAULT_LOCK_NAME);
+  await (0, import_promises20.writeFile)(lockFile, "", { encoding: "utf8" });
   return lockFile;
 }
 async function removeSubagentLock(subagentDir) {
-  const lockFile = import_node_path30.default.join(subagentDir, DEFAULT_LOCK_NAME);
+  const lockFile = import_node_path29.default.join(subagentDir, DEFAULT_LOCK_NAME);
   await removeIfExists(lockFile);
 }
 async function prepareSubagentDirectory(subagentDir, promptFile, chatId, workspaceTemplate, dryRun, cwd) {
@@ -11808,11 +11589,11 @@ async function prepareSubagentDirectory(subagentDir, promptFile, chatId, workspa
     return 1;
   }
   if (promptFile) {
-    const githubAgentsDir = import_node_path30.default.join(subagentDir, ".github", "agents");
-    await (0, import_promises21.mkdir)(githubAgentsDir, { recursive: true });
-    const agentFile = import_node_path30.default.join(githubAgentsDir, `${chatId}.md`);
+    const githubAgentsDir = import_node_path29.default.join(subagentDir, ".github", "agents");
+    await (0, import_promises20.mkdir)(githubAgentsDir, { recursive: true });
+    const agentFile = import_node_path29.default.join(githubAgentsDir, `${chatId}.md`);
     try {
-      await (0, import_promises21.copyFile)(promptFile, agentFile);
+      await (0, import_promises20.copyFile)(promptFile, agentFile);
     } catch (error) {
       console.error(`error: Failed to copy prompt file to agent mode: ${error.message}`);
       return 1;
@@ -11829,11 +11610,11 @@ async function resolvePromptFile(promptFile) {
   if (!promptFile) {
     return void 0;
   }
-  const resolvedPrompt = import_node_path31.default.resolve(promptFile);
+  const resolvedPrompt = import_node_path30.default.resolve(promptFile);
   if (!await pathExists(resolvedPrompt)) {
     throw new Error(`Prompt file not found: ${resolvedPrompt}`);
   }
-  const promptStats = await (0, import_promises22.stat)(resolvedPrompt);
+  const promptStats = await (0, import_promises21.stat)(resolvedPrompt);
   if (!promptStats.isFile()) {
     throw new Error(`Prompt file must be a file, not a directory: ${resolvedPrompt}`);
   }
@@ -11845,7 +11626,7 @@ async function resolveAttachments(extraAttachments) {
   }
   const resolved = [];
   for (const attachment of extraAttachments) {
-    const resolvedPath = import_node_path31.default.resolve(attachment);
+    const resolvedPath = import_node_path30.default.resolve(attachment);
     if (!await pathExists(resolvedPath)) {
       throw new Error(`Attachment not found: ${resolvedPath}`);
     }
@@ -11887,7 +11668,7 @@ async function dispatchAgentSession(options) {
         error: "No unlocked subagents available. Provision additional subagents with: subagent code provision --subagents <desired_total>"
       };
     }
-    const subagentName = import_node_path31.default.basename(subagentDir);
+    const subagentName = import_node_path30.default.basename(subagentDir);
     const chatId = Math.random().toString(16).slice(2, 10);
     const preparationResult = await prepareSubagentDirectory(
       subagentDir,
@@ -11915,9 +11696,9 @@ async function dispatchAgentSession(options) {
       };
     }
     const timestamp = generateTimestamp();
-    const messagesDir = import_node_path31.default.join(subagentDir, "messages");
-    const responseFileTmp = import_node_path31.default.join(messagesDir, `${timestamp}_res.tmp.md`);
-    const responseFileFinal = import_node_path31.default.join(messagesDir, `${timestamp}_res.md`);
+    const messagesDir = import_node_path30.default.join(subagentDir, "messages");
+    const responseFileTmp = import_node_path30.default.join(messagesDir, `${timestamp}_res.tmp.md`);
+    const responseFileFinal = import_node_path30.default.join(messagesDir, `${timestamp}_res.md`);
     const requestInstructions = createRequestPrompt(
       userQuery,
       responseFileTmp,
@@ -12022,7 +11803,7 @@ async function dispatchBatchAgent(options) {
         error: "No unlocked subagents available. Provision additional subagents with: subagent code provision --subagents <desired_total>"
       };
     }
-    subagentName = import_node_path31.default.basename(subagentDir);
+    subagentName = import_node_path30.default.basename(subagentDir);
     const chatId = Math.random().toString(16).slice(2, 10);
     const preparationResult = await prepareSubagentDirectory(
       subagentDir,
@@ -12053,24 +11834,24 @@ async function dispatchBatchAgent(options) {
       };
     }
     const timestamp = generateTimestamp();
-    const messagesDir = import_node_path31.default.join(subagentDir, "messages");
+    const messagesDir = import_node_path30.default.join(subagentDir, "messages");
     requestFiles = userQueries.map(
-      (_, index) => import_node_path31.default.join(messagesDir, `${timestamp}_${index}_req.md`)
+      (_, index) => import_node_path30.default.join(messagesDir, `${timestamp}_${index}_req.md`)
     );
     const responseTmpFiles = userQueries.map(
-      (_, index) => import_node_path31.default.join(messagesDir, `${timestamp}_${index}_res.tmp.md`)
+      (_, index) => import_node_path30.default.join(messagesDir, `${timestamp}_${index}_res.tmp.md`)
     );
     responseFilesFinal = userQueries.map(
-      (_, index) => import_node_path31.default.join(messagesDir, `${timestamp}_${index}_res.md`)
+      (_, index) => import_node_path30.default.join(messagesDir, `${timestamp}_${index}_res.md`)
     );
-    const orchestratorFile = import_node_path31.default.join(messagesDir, `${timestamp}_orchestrator.md`);
+    const orchestratorFile = import_node_path30.default.join(messagesDir, `${timestamp}_orchestrator.md`);
     if (!dryRun) {
       await Promise.all(
         userQueries.map((query, index) => {
           const reqFile = requestFiles[index];
           const tmpFile = responseTmpFiles[index];
           const finalFile = responseFilesFinal[index];
-          return (0, import_promises22.writeFile)(
+          return (0, import_promises21.writeFile)(
             reqFile,
             createBatchRequestPrompt(query, tmpFile, finalFile, batchRequestTemplateContent),
             { encoding: "utf8" }
@@ -12082,7 +11863,7 @@ async function dispatchBatchAgent(options) {
         responseFilesFinal,
         orchestratorTemplateContent
       );
-      await (0, import_promises22.writeFile)(orchestratorFile, orchestratorContent, { encoding: "utf8" });
+      await (0, import_promises21.writeFile)(orchestratorFile, orchestratorContent, { encoding: "utf8" });
     }
     const chatAttachments = [orchestratorFile, ...attachments];
     const orchestratorUri = pathToFileUri2(orchestratorFile);
@@ -12148,8 +11929,8 @@ async function dispatchBatchAgent(options) {
 }
 
 // src/evaluation/providers/vscode/dispatch/provision.ts
-var import_promises23 = require("fs/promises");
-var import_node_path32 = __toESM(require("path"), 1);
+var import_promises22 = require("fs/promises");
+var import_node_path31 = __toESM(require("path"), 1);
 var DEFAULT_WORKSPACE_TEMPLATE2 = {
   folders: [
     {
@@ -12180,7 +11961,7 @@ async function provisionSubagents(options) {
   if (!Number.isInteger(subagents) || subagents < 1) {
     throw new Error("subagents must be a positive integer");
   }
-  const targetPath = import_node_path32.default.resolve(targetRoot);
+  const targetPath = import_node_path31.default.resolve(targetRoot);
   if (!dryRun) {
     await ensureDir(targetPath);
   }
@@ -12200,7 +11981,7 @@ async function provisionSubagents(options) {
         continue;
       }
       highestNumber = Math.max(highestNumber, parsed);
-      const lockFile = import_node_path32.default.join(entry.absolutePath, lockName);
+      const lockFile = import_node_path31.default.join(entry.absolutePath, lockName);
       const locked = await pathExists(lockFile);
       if (locked) {
         lockedSubagents.add(entry.absolutePath);
@@ -12217,10 +11998,10 @@ async function provisionSubagents(options) {
       break;
     }
     const subagentDir = subagent.absolutePath;
-    const githubAgentsDir = import_node_path32.default.join(subagentDir, ".github", "agents");
-    const lockFile = import_node_path32.default.join(subagentDir, lockName);
-    const workspaceDst = import_node_path32.default.join(subagentDir, `${import_node_path32.default.basename(subagentDir)}.code-workspace`);
-    const wakeupDst = import_node_path32.default.join(githubAgentsDir, "wakeup.md");
+    const githubAgentsDir = import_node_path31.default.join(subagentDir, ".github", "agents");
+    const lockFile = import_node_path31.default.join(subagentDir, lockName);
+    const workspaceDst = import_node_path31.default.join(subagentDir, `${import_node_path31.default.basename(subagentDir)}.code-workspace`);
+    const wakeupDst = import_node_path31.default.join(githubAgentsDir, "wakeup.md");
     const isLocked = await pathExists(lockFile);
     if (isLocked && !force) {
       continue;
@@ -12229,8 +12010,8 @@ async function provisionSubagents(options) {
       if (!dryRun) {
         await removeIfExists(lockFile);
         await ensureDir(githubAgentsDir);
-        await (0, import_promises23.writeFile)(workspaceDst, JSON.stringify(workspaceTemplate, null, 2), "utf8");
-        await (0, import_promises23.writeFile)(wakeupDst, wakeupContent, "utf8");
+        await (0, import_promises22.writeFile)(workspaceDst, JSON.stringify(workspaceTemplate, null, 2), "utf8");
+        await (0, import_promises22.writeFile)(wakeupDst, wakeupContent, "utf8");
       }
       created.push(subagentDir);
       lockedSubagents.delete(subagentDir);
@@ -12240,8 +12021,8 @@ async function provisionSubagents(options) {
     if (!isLocked && force) {
       if (!dryRun) {
         await ensureDir(githubAgentsDir);
-        await (0, import_promises23.writeFile)(workspaceDst, JSON.stringify(workspaceTemplate, null, 2), "utf8");
-        await (0, import_promises23.writeFile)(wakeupDst, wakeupContent, "utf8");
+        await (0, import_promises22.writeFile)(workspaceDst, JSON.stringify(workspaceTemplate, null, 2), "utf8");
+        await (0, import_promises22.writeFile)(wakeupDst, wakeupContent, "utf8");
       }
       created.push(subagentDir);
       subagentsProvisioned += 1;
@@ -12249,8 +12030,8 @@ async function provisionSubagents(options) {
     }
     if (!dryRun && !await pathExists(workspaceDst)) {
       await ensureDir(githubAgentsDir);
-      await (0, import_promises23.writeFile)(workspaceDst, JSON.stringify(workspaceTemplate, null, 2), "utf8");
-      await (0, import_promises23.writeFile)(wakeupDst, wakeupContent, "utf8");
+      await (0, import_promises22.writeFile)(workspaceDst, JSON.stringify(workspaceTemplate, null, 2), "utf8");
+      await (0, import_promises22.writeFile)(wakeupDst, wakeupContent, "utf8");
     }
     skippedExisting.push(subagentDir);
     subagentsProvisioned += 1;
@@ -12258,15 +12039,15 @@ async function provisionSubagents(options) {
   let nextIndex = highestNumber;
   while (subagentsProvisioned < subagents) {
     nextIndex += 1;
-    const subagentDir = import_node_path32.default.join(targetPath, `subagent-${nextIndex}`);
-    const githubAgentsDir = import_node_path32.default.join(subagentDir, ".github", "agents");
-    const workspaceDst = import_node_path32.default.join(subagentDir, `${import_node_path32.default.basename(subagentDir)}.code-workspace`);
-    const wakeupDst = import_node_path32.default.join(githubAgentsDir, "wakeup.md");
+    const subagentDir = import_node_path31.default.join(targetPath, `subagent-${nextIndex}`);
+    const githubAgentsDir = import_node_path31.default.join(subagentDir, ".github", "agents");
+    const workspaceDst = import_node_path31.default.join(subagentDir, `${import_node_path31.default.basename(subagentDir)}.code-workspace`);
+    const wakeupDst = import_node_path31.default.join(githubAgentsDir, "wakeup.md");
     if (!dryRun) {
       await ensureDir(subagentDir);
       await ensureDir(githubAgentsDir);
-      await (0, import_promises23.writeFile)(workspaceDst, JSON.stringify(workspaceTemplate, null, 2), "utf8");
-      await (0, import_promises23.writeFile)(wakeupDst, wakeupContent, "utf8");
+      await (0, import_promises22.writeFile)(workspaceDst, JSON.stringify(workspaceTemplate, null, 2), "utf8");
+      await (0, import_promises22.writeFile)(wakeupDst, wakeupContent, "utf8");
     }
     created.push(subagentDir);
     subagentsProvisioned += 1;
@@ -12328,7 +12109,7 @@ var VSCodeProvider = class {
     }
     await this.ensureEnvironmentReady();
     const inputFiles = normalizeAttachments(request.inputFiles);
-    const promptContent = buildPromptDocument2(request, inputFiles, request.guideline_patterns);
+    const promptContent = buildPromptDocument2(request, inputFiles);
     const workspaceTemplate = request.workspaceFile ?? await resolveWorkspaceTemplateFile(this.config.workspaceTemplate);
     const startTime = Date.now();
     const session = await dispatchAgentSession({
@@ -12382,7 +12163,7 @@ var VSCodeProvider = class {
       normalizedRequests.map(({ inputFiles }) => inputFiles)
     );
     const userQueries = normalizedRequests.map(
-      ({ request, inputFiles }) => buildPromptDocument2(request, inputFiles, request.guideline_patterns)
+      ({ request, inputFiles }) => buildPromptDocument2(request, inputFiles)
     );
     const batchWorkspaceTemplate = await resolveWorkspaceTemplateFile(
       this.config.workspaceTemplate
@@ -12451,9 +12232,9 @@ var VSCodeProvider = class {
 async function locateVSCodeExecutable(candidate) {
   const includesPathSeparator = candidate.includes("/") || candidate.includes("\\");
   if (includesPathSeparator) {
-    const resolved = import_node_path33.default.isAbsolute(candidate) ? candidate : import_node_path33.default.resolve(candidate);
+    const resolved = import_node_path32.default.isAbsolute(candidate) ? candidate : import_node_path32.default.resolve(candidate);
     try {
-      await (0, import_promises24.access)(resolved, import_promises24.constants.F_OK);
+      await (0, import_promises23.access)(resolved, import_promises23.constants.F_OK);
       return resolved;
     } catch {
       throw new Error(
@@ -12466,7 +12247,7 @@ async function locateVSCodeExecutable(candidate) {
     const { stdout } = await execAsync3(`${locator} ${candidate}`);
     const lines = stdout.split(/\r?\n/).map((line) => line.trim()).filter((line) => line.length > 0);
     if (lines.length > 0 && lines[0]) {
-      await (0, import_promises24.access)(lines[0], import_promises24.constants.F_OK);
+      await (0, import_promises23.access)(lines[0], import_promises23.constants.F_OK);
       return lines[0];
     }
   } catch {
@@ -12480,41 +12261,35 @@ async function resolveWorkspaceTemplateFile(template) {
     return void 0;
   }
   try {
-    const stats = await (0, import_promises24.stat)(import_node_path33.default.resolve(template));
+    const stats = await (0, import_promises23.stat)(import_node_path32.default.resolve(template));
     return stats.isFile() ? template : void 0;
   } catch {
     return template;
   }
 }
-function buildPromptDocument2(request, attachments, guidelinePatterns) {
+function buildPromptDocument2(request, attachments) {
   const parts = [];
   if (request.systemPrompt && request.systemPrompt.trim().length > 0) {
     parts.push(request.systemPrompt.trim());
   }
-  const guidelineFiles = collectGuidelineFiles2(attachments, guidelinePatterns);
   const attachmentFiles = collectAttachmentFiles(attachments);
-  const nonGuidelineAttachments = attachmentFiles.filter((file) => !guidelineFiles.includes(file));
-  const prereadBlock = buildMandatoryPrereadBlock2(guidelineFiles, nonGuidelineAttachments);
+  const prereadBlock = buildMandatoryPrereadBlock2(attachmentFiles);
   if (prereadBlock.length > 0) {
     parts.push("\n", prereadBlock);
   }
   parts.push("\n[[ ## user_query ## ]]\n", request.question.trim());
   return parts.join("\n").trim();
 }
-function buildMandatoryPrereadBlock2(guidelineFiles, attachmentFiles) {
-  if (guidelineFiles.length === 0 && attachmentFiles.length === 0) {
+function buildMandatoryPrereadBlock2(attachmentFiles) {
+  if (attachmentFiles.length === 0) {
     return "";
   }
   const buildList = (files) => files.map((absolutePath) => {
-    const fileName = import_node_path33.default.basename(absolutePath);
+    const fileName = import_node_path32.default.basename(absolutePath);
     const fileUri = pathToFileUri3(absolutePath);
     return `* [${fileName}](${fileUri})`;
   });
   const sections = [];
-  if (guidelineFiles.length > 0) {
-    sections.push(`Read all guideline files:
-${buildList(guidelineFiles).join("\n")}.`);
-  }
   if (attachmentFiles.length > 0) {
     sections.push(`Read all attachment files:
 ${buildList(attachmentFiles).join("\n")}.`);
@@ -12525,29 +12300,13 @@ ${buildList(attachmentFiles).join("\n")}.`);
   );
   return sections.join("\n");
 }
-function collectGuidelineFiles2(attachments, guidelinePatterns) {
-  if (!attachments || attachments.length === 0) {
-    return [];
-  }
-  const unique = /* @__PURE__ */ new Map();
-  for (const attachment of attachments) {
-    const absolutePath = import_node_path33.default.resolve(attachment);
-    const normalized = absolutePath.split(import_node_path33.default.sep).join("/");
-    if (isGuidelineFile(normalized, guidelinePatterns)) {
-      if (!unique.has(absolutePath)) {
-        unique.set(absolutePath, absolutePath);
-      }
-    }
-  }
-  return Array.from(unique.values());
-}
 function collectAttachmentFiles(attachments) {
   if (!attachments || attachments.length === 0) {
     return [];
   }
   const unique = /* @__PURE__ */ new Map();
   for (const attachment of attachments) {
-    const absolutePath = import_node_path33.default.resolve(attachment);
+    const absolutePath = import_node_path32.default.resolve(attachment);
     if (!unique.has(absolutePath)) {
       unique.set(absolutePath, absolutePath);
     }
@@ -12555,7 +12314,7 @@ function collectAttachmentFiles(attachments) {
   return Array.from(unique.values());
 }
 function pathToFileUri3(filePath) {
-  const absolutePath = import_node_path33.default.isAbsolute(filePath) ? filePath : import_node_path33.default.resolve(filePath);
+  const absolutePath = import_node_path32.default.isAbsolute(filePath) ? filePath : import_node_path32.default.resolve(filePath);
   const normalizedPath = absolutePath.replace(/\\/g, "/");
   if (/^[a-zA-Z]:\//.test(normalizedPath)) {
     return `file:///${normalizedPath}`;
@@ -12568,7 +12327,7 @@ function normalizeAttachments(attachments) {
   }
   const deduped = /* @__PURE__ */ new Set();
   for (const attachment of attachments) {
-    deduped.add(import_node_path33.default.resolve(attachment));
+    deduped.add(import_node_path32.default.resolve(attachment));
   }
   return Array.from(deduped);
 }
@@ -12577,7 +12336,7 @@ function mergeAttachments(all) {
   for (const list of all) {
     if (!list) continue;
     for (const inputFile of list) {
-      deduped.add(import_node_path33.default.resolve(inputFile));
+      deduped.add(import_node_path32.default.resolve(inputFile));
     }
   }
   return deduped.size > 0 ? Array.from(deduped) : void 0;
@@ -12625,8 +12384,8 @@ total unlocked subagents available: ${result.created.length + result.skippedExis
 
 // src/evaluation/providers/targets-file.ts
 var import_node_fs10 = require("fs");
-var import_promises25 = require("fs/promises");
-var import_node_path34 = __toESM(require("path"), 1);
+var import_promises24 = require("fs/promises");
+var import_node_path33 = __toESM(require("path"), 1);
 var import_yaml6 = require("yaml");
 function isRecord(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
@@ -12656,18 +12415,18 @@ function assertTargetDefinition(value, index, filePath) {
 }
 async function fileExists3(filePath) {
   try {
-    await (0, import_promises25.access)(filePath, import_node_fs10.constants.F_OK);
+    await (0, import_promises24.access)(filePath, import_node_fs10.constants.F_OK);
     return true;
   } catch {
     return false;
   }
 }
 async function readTargetDefinitions(filePath) {
-  const absolutePath = import_node_path34.default.resolve(filePath);
+  const absolutePath = import_node_path33.default.resolve(filePath);
   if (!await fileExists3(absolutePath)) {
     throw new Error(`targets.yaml not found at ${absolutePath}`);
   }
-  const raw = await (0, import_promises25.readFile)(absolutePath, "utf8");
+  const raw = await (0, import_promises24.readFile)(absolutePath, "utf8");
   const parsed = (0, import_yaml6.parse)(raw);
   if (!isRecord(parsed)) {
     throw new Error(`targets.yaml at ${absolutePath} must be a YAML object with a 'targets' field`);
@@ -12683,16 +12442,16 @@ function listTargetNames(definitions) {
 }
 
 // src/evaluation/providers/provider-discovery.ts
-var import_node_path35 = __toESM(require("path"), 1);
+var import_node_path34 = __toESM(require("path"), 1);
 var import_fast_glob2 = __toESM(require("fast-glob"), 1);
 async function discoverProviders(registry, baseDir) {
   const patterns = ["*.ts", "*.js", "*.mts", "*.mjs"];
   const candidateDirs = [];
-  let dir = import_node_path35.default.resolve(baseDir);
-  const root = import_node_path35.default.parse(dir).root;
+  let dir = import_node_path34.default.resolve(baseDir);
+  const root = import_node_path34.default.parse(dir).root;
   while (dir !== root) {
-    candidateDirs.push(import_node_path35.default.join(dir, ".agentv", "providers"));
-    dir = import_node_path35.default.dirname(dir);
+    candidateDirs.push(import_node_path34.default.join(dir, ".agentv", "providers"));
+    dir = import_node_path34.default.dirname(dir);
   }
   let files = [];
   for (const providersDir of candidateDirs) {
@@ -12708,7 +12467,7 @@ async function discoverProviders(registry, baseDir) {
   }
   const discoveredKinds = [];
   for (const filePath of files) {
-    const basename = import_node_path35.default.basename(filePath);
+    const basename = import_node_path34.default.basename(filePath);
     const kindName = basename.replace(/\.(ts|js|mts|mjs)$/, "");
     if (registry.has(kindName)) {
       continue;
@@ -12815,9 +12574,9 @@ function negateScore(score) {
 }
 
 // src/evaluation/evaluators/code-evaluator.ts
-var import_promises26 = require("fs/promises");
+var import_promises25 = require("fs/promises");
 var import_node_os5 = require("os");
-var import_node_path36 = require("path");
+var import_node_path35 = require("path");
 
 // src/runtime/exec.ts
 function shellEscapePath(value) {
@@ -12917,15 +12676,15 @@ async function execFileWithStdinNode(argv, stdinPayload, options) {
   });
 }
 async function execShellWithStdin(command, stdinPayload, options = {}) {
-  const { mkdir: mkdir17, readFile: readFile15, rm: rm6, writeFile: writeFile9 } = await import("fs/promises");
+  const { mkdir: mkdir17, readFile: readFile14, rm: rm6, writeFile: writeFile9 } = await import("fs/promises");
   const { tmpdir: tmpdir3 } = await import("os");
-  const path48 = await import("path");
+  const path47 = await import("path");
   const { randomUUID: randomUUID9 } = await import("crypto");
-  const dir = path48.join(tmpdir3(), `agentv-exec-${randomUUID9()}`);
+  const dir = path47.join(tmpdir3(), `agentv-exec-${randomUUID9()}`);
   await mkdir17(dir, { recursive: true });
-  const stdinPath = path48.join(dir, "stdin.txt");
-  const stdoutPath = path48.join(dir, "stdout.txt");
-  const stderrPath = path48.join(dir, "stderr.txt");
+  const stdinPath = path47.join(dir, "stdin.txt");
+  const stdoutPath = path47.join(dir, "stdout.txt");
+  const stderrPath = path47.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");
@@ -12955,8 +12714,8 @@ async function execShellWithStdin(command, stdinPayload, options = {}) {
         resolve(code ?? 0);
       });
     });
-    const stdout = (await readFile15(stdoutPath, "utf8")).replace(/\r\n/g, "\n");
-    const stderr = (await readFile15(stderrPath, "utf8")).replace(/\r\n/g, "\n");
+    const stdout = (await readFile14(stdoutPath, "utf8")).replace(/\r\n/g, "\n");
+    const stderr = (await readFile14(stderrPath, "utf8")).replace(/\r\n/g, "\n");
     return { stdout, stderr, exitCode };
   } finally {
     await rm6(dir, { recursive: true, force: true });
@@ -13261,9 +13020,9 @@ var CodeEvaluator = class {
     if (outputForPayload) {
       const serialized = JSON.stringify(outputForPayload);
       if (serialized.length > FILE_BACKED_OUTPUT_THRESHOLD) {
-        const tmpDir = await (0, import_promises26.mkdtemp)((0, import_node_path36.join)((0, import_node_os5.tmpdir)(), "agentv-judge-"));
-        outputPath = (0, import_node_path36.join)(tmpDir, "output.json");
-        await (0, import_promises26.writeFile)(outputPath, serialized);
+        const tmpDir = await (0, import_promises25.mkdtemp)((0, import_node_path35.join)((0, import_node_os5.tmpdir)(), "agentv-judge-"));
+        outputPath = (0, import_node_path35.join)(tmpDir, "output.json");
+        await (0, import_promises25.writeFile)(outputPath, serialized);
         outputForPayload = null;
       }
     }
@@ -13273,10 +13032,7 @@ var CodeEvaluator = class {
       outputText: context2.candidate,
       output: outputForPayload,
       outputPath,
-      guidelineFiles: context2.evalCase.guideline_paths,
-      inputFiles: context2.evalCase.file_paths.filter(
-        (path48) => !context2.evalCase.guideline_paths.includes(path48)
-      ),
+      inputFiles: context2.evalCase.file_paths,
       input: context2.evalCase.input,
       trace: context2.trace ?? null,
       tokenUsage: context2.tokenUsage ?? null,
@@ -13375,7 +13131,7 @@ var CodeEvaluator = class {
         await proxyShutdown();
       }
       if (outputPath) {
-        await (0, import_promises26.rm)((0, import_node_path36.dirname)(outputPath), { recursive: true, force: true }).catch(() => {
+        await (0, import_promises25.rm)((0, import_node_path35.dirname)(outputPath), { recursive: true, force: true }).catch(() => {
         });
       }
     }
@@ -13438,8 +13194,8 @@ function isAgentProvider(provider) {
 }
 
 // src/evaluation/evaluators/llm-grader.ts
-var import_promises27 = __toESM(require("fs/promises"), 1);
-var import_node_path37 = __toESM(require("path"), 1);
+var import_promises26 = __toESM(require("fs/promises"), 1);
+var import_node_path36 = __toESM(require("path"), 1);
 var import_ai2 = require("ai");
 var import_zod4 = require("zod");
 var DEFAULT_MAX_STEPS = 10;
@@ -14271,8 +14027,8 @@ function calculateScoreRangeResult(result, rubrics) {
   };
 }
 function resolveSandboxed(basePath, relativePath) {
-  const resolved = import_node_path37.default.resolve(basePath, relativePath);
-  if (!resolved.startsWith(basePath + import_node_path37.default.sep) && resolved !== basePath) {
+  const resolved = import_node_path36.default.resolve(basePath, relativePath);
+  if (!resolved.startsWith(basePath + import_node_path36.default.sep) && resolved !== basePath) {
     throw new Error(`Path '${relativePath}' is outside the workspace`);
   }
   return resolved;
@@ -14287,7 +14043,7 @@ function createFilesystemTools(workspacePath) {
       execute: async (input) => {
         try {
           const resolved = resolveSandboxed(workspacePath, input.path);
-          const entries = await import_promises27.default.readdir(resolved, { withFileTypes: true });
+          const entries = await import_promises26.default.readdir(resolved, { withFileTypes: true });
           return entries.map((e) => ({
             name: e.name,
             type: e.isDirectory() ? "directory" : "file"
@@ -14305,12 +14061,12 @@ function createFilesystemTools(workspacePath) {
       execute: async (input) => {
         try {
           const resolved = resolveSandboxed(workspacePath, input.path);
-          const stat8 = await import_promises27.default.stat(resolved);
+          const stat8 = await import_promises26.default.stat(resolved);
           if (stat8.isDirectory()) {
             return { error: `'${input.path}' is a directory, not a file` };
           }
           const buffer = Buffer.alloc(Math.min(stat8.size, MAX_FILE_SIZE));
-          const fd = await import_promises27.default.open(resolved, "r");
+          const fd = await import_promises26.default.open(resolved, "r");
           try {
             await fd.read(buffer, 0, buffer.length, 0);
           } finally {
@@ -14355,30 +14111,30 @@ async function searchDirectory(dirPath, workspacePath, regex, matches) {
   if (matches.length >= MAX_SEARCH_MATCHES) return;
   let entries;
   try {
-    entries = await import_promises27.default.readdir(dirPath, { withFileTypes: true });
+    entries = await import_promises26.default.readdir(dirPath, { withFileTypes: true });
   } catch {
     return;
   }
   for (const entry of entries) {
     if (matches.length >= MAX_SEARCH_MATCHES) return;
     if (SEARCH_SKIP_DIRS.has(entry.name)) continue;
-    const fullPath = import_node_path37.default.join(dirPath, entry.name);
+    const fullPath = import_node_path36.default.join(dirPath, entry.name);
     if (entry.isDirectory()) {
       await searchDirectory(fullPath, workspacePath, regex, matches);
     } else if (entry.isFile()) {
-      const ext = import_node_path37.default.extname(entry.name).toLowerCase();
+      const ext = import_node_path36.default.extname(entry.name).toLowerCase();
       if (BINARY_EXTENSIONS.has(ext)) continue;
       try {
-        const stat8 = await import_promises27.default.stat(fullPath);
+        const stat8 = await import_promises26.default.stat(fullPath);
         if (stat8.size > MAX_FILE_SIZE) continue;
-        const content = await import_promises27.default.readFile(fullPath, "utf-8");
+        const content = await import_promises26.default.readFile(fullPath, "utf-8");
         const lines = content.split("\n");
         for (let i = 0; i < lines.length; i++) {
           if (matches.length >= MAX_SEARCH_MATCHES) return;
           regex.lastIndex = 0;
           if (regex.test(lines[i])) {
             matches.push({
-              file: import_node_path37.default.relative(workspacePath, fullPath),
+              file: import_node_path36.default.relative(workspacePath, fullPath),
               line: i + 1,
               text: lines[i].substring(0, 200)
             });
@@ -15013,115 +14769,115 @@ var FieldAccuracyEvaluator = class {
    * Evaluate a single field against the expected value.
    */
   evaluateField(fieldConfig, candidateData, expectedData) {
-    const { path: path48, match, required = true, weight = 1 } = fieldConfig;
-    const candidateValue = resolvePath(candidateData, path48);
-    const expectedValue = resolvePath(expectedData, path48);
+    const { path: path47, match, required = true, weight = 1 } = fieldConfig;
+    const candidateValue = resolvePath(candidateData, path47);
+    const expectedValue = resolvePath(expectedData, path47);
     if (expectedValue === void 0) {
       return {
-        path: path48,
+        path: path47,
         score: 1,
         // No expected value means no comparison needed
         weight,
         hit: true,
-        message: `${path48}: no expected value`
+        message: `${path47}: no expected value`
       };
     }
     if (candidateValue === void 0) {
       if (required) {
         return {
-          path: path48,
+          path: path47,
           score: 0,
           weight,
           hit: false,
-          message: `${path48} (required, missing)`
+          message: `${path47} (required, missing)`
         };
       }
       return {
-        path: path48,
+        path: path47,
         score: 1,
         // Don't penalize missing optional fields
         weight: 0,
         // Zero weight means it won't affect the score
         hit: true,
-        message: `${path48}: optional field missing`
+        message: `${path47}: optional field missing`
       };
     }
     switch (match) {
       case "exact":
-        return this.compareExact(path48, candidateValue, expectedValue, weight);
+        return this.compareExact(path47, candidateValue, expectedValue, weight);
       case "numeric_tolerance":
         return this.compareNumericTolerance(
-          path48,
+          path47,
           candidateValue,
           expectedValue,
           fieldConfig,
           weight
         );
       case "date":
-        return this.compareDate(path48, candidateValue, expectedValue, fieldConfig, weight);
+        return this.compareDate(path47, candidateValue, expectedValue, fieldConfig, weight);
       default:
         return {
-          path: path48,
+          path: path47,
           score: 0,
           weight,
           hit: false,
-          message: `${path48}: unknown match type "${match}"`
+          message: `${path47}: unknown match type "${match}"`
         };
     }
   }
   /**
    * Exact equality comparison.
    */
-  compareExact(path48, candidateValue, expectedValue, weight) {
+  compareExact(path47, candidateValue, expectedValue, weight) {
     if (deepEqual(candidateValue, expectedValue)) {
       return {
-        path: path48,
+        path: path47,
         score: 1,
         weight,
         hit: true,
-        message: path48
+        message: path47
       };
     }
     if (typeof candidateValue !== typeof expectedValue) {
       return {
-        path: path48,
+        path: path47,
         score: 0,
         weight,
         hit: false,
-        message: `${path48} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
+        message: `${path47} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
       };
     }
     return {
-      path: path48,
+      path: path47,
       score: 0,
       weight,
       hit: false,
-      message: `${path48} (value mismatch)`
+      message: `${path47} (value mismatch)`
     };
   }
   /**
    * Numeric comparison with absolute or relative tolerance.
    */
-  compareNumericTolerance(path48, candidateValue, expectedValue, fieldConfig, weight) {
+  compareNumericTolerance(path47, candidateValue, expectedValue, fieldConfig, weight) {
     const { tolerance = 0, relative = false } = fieldConfig;
     const candidateNum = toNumber(candidateValue);
     const expectedNum = toNumber(expectedValue);
     if (candidateNum === null || expectedNum === null) {
       return {
-        path: path48,
+        path: path47,
         score: 0,
         weight,
         hit: false,
-        message: `${path48} (non-numeric value)`
+        message: `${path47} (non-numeric value)`
       };
     }
     if (!Number.isFinite(candidateNum) || !Number.isFinite(expectedNum)) {
       return {
-        path: path48,
+        path: path47,
         score: 0,
         weight,
         hit: false,
-        message: `${path48} (invalid numeric value)`
+        message: `${path47} (invalid numeric value)`
       };
     }
     const diff = Math.abs(candidateNum - expectedNum);
@@ -15134,61 +14890,61 @@ var FieldAccuracyEvaluator = class {
     }
     if (withinTolerance) {
       return {
-        path: path48,
+        path: path47,
         score: 1,
         weight,
         hit: true,
-        message: `${path48} (within tolerance: diff=${diff.toFixed(2)})`
+        message: `${path47} (within tolerance: diff=${diff.toFixed(2)})`
       };
     }
     return {
-      path: path48,
+      path: path47,
       score: 0,
       weight,
       hit: false,
-      message: `${path48} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
+      message: `${path47} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
     };
   }
   /**
    * Date comparison with format normalization.
    */
-  compareDate(path48, candidateValue, expectedValue, fieldConfig, weight) {
+  compareDate(path47, candidateValue, expectedValue, fieldConfig, weight) {
     const formats = fieldConfig.formats ?? DEFAULT_DATE_FORMATS;
     const candidateDate = parseDate(String(candidateValue), formats);
     const expectedDate = parseDate(String(expectedValue), formats);
     if (candidateDate === null) {
       return {
-        path: path48,
+        path: path47,
         score: 0,
         weight,
         hit: false,
-        message: `${path48} (unparseable candidate date)`
+        message: `${path47} (unparseable candidate date)`
       };
     }
     if (expectedDate === null) {
       return {
-        path: path48,
+        path: path47,
         score: 0,
         weight,
         hit: false,
-        message: `${path48} (unparseable expected date)`
+        message: `${path47} (unparseable expected date)`
       };
     }
     if (candidateDate.getFullYear() === expectedDate.getFullYear() && candidateDate.getMonth() === expectedDate.getMonth() && candidateDate.getDate() === expectedDate.getDate()) {
       return {
-        path: path48,
+        path: path47,
         score: 1,
         weight,
         hit: true,
-        message: path48
+        message: path47
       };
     }
     return {
-      path: path48,
+      path: path47,
       score: 0,
       weight,
       hit: false,
-      message: `${path48} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
+      message: `${path47} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
     };
   }
   /**
@@ -15221,11 +14977,11 @@ var FieldAccuracyEvaluator = class {
     };
   }
 };
-function resolvePath(obj, path48) {
-  if (!path48 || !obj) {
+function resolvePath(obj, path47) {
+  if (!path47 || !obj) {
     return void 0;
   }
-  const parts = path48.split(/\.|\[|\]/).filter((p) => p.length > 0);
+  const parts = path47.split(/\.|\[|\]/).filter((p) => p.length > 0);
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0) {
@@ -15685,8 +15441,8 @@ var TokenUsageEvaluator = class {
 };
 
 // src/evaluation/evaluators/tool-trajectory.ts
-function getNestedValue(obj, path48) {
-  const parts = path48.split(".");
+function getNestedValue(obj, path47) {
+  const parts = path47.split(".");
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0 || typeof current !== "object") {
@@ -16307,9 +16063,9 @@ function runEqualsAssertion(output, value) {
 
 // src/evaluation/orchestrator.ts
 var import_node_crypto10 = require("crypto");
-var import_promises31 = require("fs/promises");
-var import_node_path46 = __toESM(require("path"), 1);
-var import_micromatch4 = __toESM(require("micromatch"), 1);
+var import_promises30 = require("fs/promises");
+var import_node_path45 = __toESM(require("path"), 1);
+var import_micromatch3 = __toESM(require("micromatch"), 1);
 
 // ../../node_modules/.bun/yocto-queue@1.2.2/node_modules/yocto-queue/index.js
 var Node = class {
@@ -16522,7 +16278,7 @@ var InlineAssertEvaluator = class {
 };
 
 // src/evaluation/evaluators/prompt-resolution.ts
-var import_node_path38 = __toESM(require("path"), 1);
+var import_node_path37 = __toESM(require("path"), 1);
 async function resolveCustomPrompt(promptConfig, context2, timeoutMs) {
   if (promptConfig.resolvedPromptScript && promptConfig.resolvedPromptScript.length > 0) {
     if (!context2) {
@@ -16557,10 +16313,7 @@ async function executePromptTemplate(script, context2, config, timeoutMs) {
     expectedOutput: context2.evalCase.expected_output,
     outputText: context2.candidate,
     output: context2.output ?? null,
-    guidelineFiles: context2.evalCase.guideline_paths,
-    inputFiles: context2.evalCase.file_paths.filter(
-      (p) => !context2.evalCase.guideline_paths.includes(p)
-    ),
+    inputFiles: context2.evalCase.file_paths,
     input: context2.evalCase.input,
     trace: context2.trace ?? null,
     fileChanges: context2.fileChanges ?? null,
@@ -16571,7 +16324,7 @@ async function executePromptTemplate(script, context2, config, timeoutMs) {
   };
   const inputJson = JSON.stringify(toSnakeCaseDeep(payload), null, 2);
   const scriptPath = script[script.length - 1];
-  const cwd = import_node_path38.default.dirname(scriptPath);
+  const cwd = import_node_path37.default.dirname(scriptPath);
   try {
     const stdout = await executeScript(script, inputJson, timeoutMs, cwd);
     const prompt = stdout.trim();
@@ -16843,16 +16596,16 @@ function createBuiltinRegistry() {
 }
 
 // src/evaluation/registry/assertion-discovery.ts
-var import_node_path39 = __toESM(require("path"), 1);
+var import_node_path38 = __toESM(require("path"), 1);
 var import_fast_glob3 = __toESM(require("fast-glob"), 1);
 async function discoverAssertions(registry, baseDir) {
   const patterns = ["*.ts", "*.js", "*.mts", "*.mjs"];
   const candidateDirs = [];
-  let dir = import_node_path39.default.resolve(baseDir);
-  const root = import_node_path39.default.parse(dir).root;
+  let dir = import_node_path38.default.resolve(baseDir);
+  const root = import_node_path38.default.parse(dir).root;
   while (dir !== root) {
-    candidateDirs.push(import_node_path39.default.join(dir, ".agentv", "assertions"));
-    dir = import_node_path39.default.dirname(dir);
+    candidateDirs.push(import_node_path38.default.join(dir, ".agentv", "assertions"));
+    dir = import_node_path38.default.dirname(dir);
   }
   let files = [];
   for (const assertionsDir of candidateDirs) {
@@ -16868,7 +16621,7 @@ async function discoverAssertions(registry, baseDir) {
   }
   const discoveredTypes = [];
   for (const filePath of files) {
-    const basename = import_node_path39.default.basename(filePath);
+    const basename = import_node_path38.default.basename(filePath);
     const typeName = basename.replace(/\.(ts|js|mts|mjs)$/, "");
     if (registry.has(typeName)) {
       continue;
@@ -16886,17 +16639,17 @@ async function discoverAssertions(registry, baseDir) {
 }
 
 // src/evaluation/registry/grader-discovery.ts
-var import_node_path40 = __toESM(require("path"), 1);
+var import_node_path39 = __toESM(require("path"), 1);
 var import_fast_glob4 = __toESM(require("fast-glob"), 1);
 async function discoverGraders(registry, baseDir) {
   const patterns = ["*.ts", "*.js", "*.mts", "*.mjs"];
   const candidateDirs = [];
-  let dir = import_node_path40.default.resolve(baseDir);
-  const root = import_node_path40.default.parse(dir).root;
+  let dir = import_node_path39.default.resolve(baseDir);
+  const root = import_node_path39.default.parse(dir).root;
   while (dir !== root) {
-    candidateDirs.push(import_node_path40.default.join(dir, ".agentv", "graders"));
-    candidateDirs.push(import_node_path40.default.join(dir, ".agentv", "judges"));
-    dir = import_node_path40.default.dirname(dir);
+    candidateDirs.push(import_node_path39.default.join(dir, ".agentv", "graders"));
+    candidateDirs.push(import_node_path39.default.join(dir, ".agentv", "judges"));
+    dir = import_node_path39.default.dirname(dir);
   }
   let files = [];
   for (const gradersDir of candidateDirs) {
@@ -16912,7 +16665,7 @@ async function discoverGraders(registry, baseDir) {
   }
   const discoveredTypes = [];
   for (const filePath of files) {
-    const basename = import_node_path40.default.basename(filePath);
+    const basename = import_node_path39.default.basename(filePath);
     const typeName = basename.replace(/\.(ts|js|mts|mjs)$/, "");
     if (registry.has(typeName)) {
       continue;
@@ -17072,7 +16825,7 @@ function getTCritical(df) {
 // src/evaluation/workspace/file-changes.ts
 var import_node_child_process7 = require("child_process");
 var import_node_fs11 = require("fs");
-var import_node_path41 = __toESM(require("path"), 1);
+var import_node_path40 = __toESM(require("path"), 1);
 var import_node_util4 = require("util");
 var execAsync4 = (0, import_node_util4.promisify)(import_node_child_process7.exec);
 function gitExecOpts(workspacePath) {
@@ -17106,10 +16859,10 @@ async function stageNestedRepoChanges(workspacePath) {
   }
   for (const entry of entries) {
     if (entry === ".git" || entry === "node_modules") continue;
-    const childPath = import_node_path41.default.join(workspacePath, entry);
+    const childPath = import_node_path40.default.join(workspacePath, entry);
     try {
       if (!(0, import_node_fs11.statSync)(childPath).isDirectory()) continue;
-      if (!(0, import_node_fs11.statSync)(import_node_path41.default.join(childPath, ".git")).isDirectory()) continue;
+      if (!(0, import_node_fs11.statSync)(import_node_path40.default.join(childPath, ".git")).isDirectory()) continue;
     } catch {
       continue;
     }
@@ -17119,8 +16872,8 @@ async function stageNestedRepoChanges(workspacePath) {
 }
 
 // src/evaluation/workspace/manager.ts
-var import_promises28 = require("fs/promises");
-var import_node_path42 = __toESM(require("path"), 1);
+var import_promises27 = require("fs/promises");
+var import_node_path41 = __toESM(require("path"), 1);
 var TemplateNotFoundError = class extends Error {
   constructor(templatePath) {
     super(`Workspace template not found: ${templatePath}`);
@@ -17142,7 +16895,7 @@ var WorkspaceCreationError = class extends Error {
 };
 async function isDirectory(filePath) {
   try {
-    const stats = await (0, import_promises28.stat)(filePath);
+    const stats = await (0, import_promises27.stat)(filePath);
     return stats.isDirectory();
   } catch {
     return false;
@@ -17150,26 +16903,26 @@ async function isDirectory(filePath) {
 }
 function getWorkspacePath(evalRunId, caseId, workspaceRoot) {
   const root = workspaceRoot ?? getWorkspacesRoot();
-  return import_node_path42.default.join(root, evalRunId, caseId);
+  return import_node_path41.default.join(root, evalRunId, caseId);
 }
 async function copyDirectoryRecursive(src, dest) {
-  await (0, import_promises28.mkdir)(dest, { recursive: true });
-  const entries = await (0, import_promises28.readdir)(src, { withFileTypes: true });
+  await (0, import_promises27.mkdir)(dest, { recursive: true });
+  const entries = await (0, import_promises27.readdir)(src, { withFileTypes: true });
   for (const entry of entries) {
-    const srcPath = import_node_path42.default.join(src, entry.name);
-    const destPath = import_node_path42.default.join(dest, entry.name);
+    const srcPath = import_node_path41.default.join(src, entry.name);
+    const destPath = import_node_path41.default.join(dest, entry.name);
     if (entry.name === ".git") {
       continue;
     }
     if (entry.isDirectory()) {
       await copyDirectoryRecursive(srcPath, destPath);
     } else {
-      await (0, import_promises28.cp)(srcPath, destPath, { preserveTimestamps: true });
+      await (0, import_promises27.cp)(srcPath, destPath, { preserveTimestamps: true });
     }
   }
 }
 async function createTempWorkspace(templatePath, evalRunId, caseId, workspaceRoot) {
-  const resolvedTemplatePath = import_node_path42.default.resolve(templatePath);
+  const resolvedTemplatePath = import_node_path41.default.resolve(templatePath);
   if (!await fileExists2(resolvedTemplatePath)) {
     throw new TemplateNotFoundError(resolvedTemplatePath);
   }
@@ -17179,7 +16932,7 @@ async function createTempWorkspace(templatePath, evalRunId, caseId, workspaceRoo
   const workspacePath = getWorkspacePath(evalRunId, caseId, workspaceRoot);
   try {
     if (await fileExists2(workspacePath)) {
-      await (0, import_promises28.rm)(workspacePath, { recursive: true, force: true });
+      await (0, import_promises27.rm)(workspacePath, { recursive: true, force: true });
     }
     await copyDirectoryRecursive(resolvedTemplatePath, workspacePath);
     return workspacePath;
@@ -17213,14 +16966,14 @@ async function createTempWorkspace(templatePath, evalRunId, caseId, workspaceRoo
 }
 async function cleanupWorkspace(workspacePath) {
   if (await fileExists2(workspacePath)) {
-    await (0, import_promises28.rm)(workspacePath, { recursive: true, force: true });
+    await (0, import_promises27.rm)(workspacePath, { recursive: true, force: true });
   }
 }
 async function cleanupEvalWorkspaces(evalRunId, workspaceRoot) {
   const root = workspaceRoot ?? getWorkspacesRoot();
-  const evalDir = import_node_path42.default.join(root, evalRunId);
+  const evalDir = import_node_path41.default.join(root, evalRunId);
   if (await fileExists2(evalDir)) {
-    await (0, import_promises28.rm)(evalDir, { recursive: true, force: true });
+    await (0, import_promises27.rm)(evalDir, { recursive: true, force: true });
   }
 }
 
@@ -17228,8 +16981,8 @@ async function cleanupEvalWorkspaces(evalRunId, workspaceRoot) {
 var import_node_child_process8 = require("child_process");
 var import_node_crypto9 = require("crypto");
 var import_node_fs12 = require("fs");
-var import_promises29 = require("fs/promises");
-var import_node_path43 = __toESM(require("path"), 1);
+var import_promises28 = require("fs/promises");
+var import_node_path42 = __toESM(require("path"), 1);
 var import_node_util5 = require("util");
 var execFileAsync = (0, import_node_util5.promisify)(import_node_child_process8.execFile);
 function gitEnv() {
@@ -17280,11 +17033,11 @@ function computeWorkspaceFingerprint(repos) {
   return (0, import_node_crypto9.createHash)("sha256").update(JSON.stringify(canonical)).digest("hex");
 }
 async function copyDirectoryRecursive2(src, dest, skipDirs) {
-  await (0, import_promises29.mkdir)(dest, { recursive: true });
-  const entries = await (0, import_promises29.readdir)(src, { withFileTypes: true });
+  await (0, import_promises28.mkdir)(dest, { recursive: true });
+  const entries = await (0, import_promises28.readdir)(src, { withFileTypes: true });
   for (const entry of entries) {
-    const srcPath = import_node_path43.default.join(src, entry.name);
-    const destPath = import_node_path43.default.join(dest, entry.name);
+    const srcPath = import_node_path42.default.join(src, entry.name);
+    const destPath = import_node_path42.default.join(dest, entry.name);
     if (entry.name === ".git") {
       continue;
     }
@@ -17294,7 +17047,7 @@ async function copyDirectoryRecursive2(src, dest, skipDirs) {
       }
       await copyDirectoryRecursive2(srcPath, destPath, skipDirs);
     } else {
-      await (0, import_promises29.cp)(srcPath, destPath, { preserveTimestamps: true, force: true });
+      await (0, import_promises28.cp)(srcPath, destPath, { preserveTimestamps: true, force: true });
     }
   }
 }
@@ -17317,8 +17070,8 @@ var WorkspacePoolManager = class {
   async acquireWorkspace(options) {
     const { templatePath, repos, maxSlots, repoManager, poolReset } = options;
     const fingerprint = computeWorkspaceFingerprint(repos);
-    const poolDir = import_node_path43.default.join(this.poolRoot, fingerprint);
-    await (0, import_promises29.mkdir)(poolDir, { recursive: true });
+    const poolDir = import_node_path42.default.join(this.poolRoot, fingerprint);
+    await (0, import_promises28.mkdir)(poolDir, { recursive: true });
     const drifted = await this.checkDrift(poolDir, fingerprint);
     if (drifted) {
       console.warn(
@@ -17327,7 +17080,7 @@ var WorkspacePoolManager = class {
       await this.removeAllSlots(poolDir);
     }
     for (let i = 0; i < maxSlots; i++) {
-      const slotPath = import_node_path43.default.join(poolDir, `slot-${i}`);
+      const slotPath = import_node_path42.default.join(poolDir, `slot-${i}`);
       const lockPath = `${slotPath}.lock`;
       const locked = await this.tryLock(lockPath);
       if (!locked) {
@@ -17345,7 +17098,7 @@ var WorkspacePoolManager = class {
           poolDir
         };
       }
-      await (0, import_promises29.mkdir)(slotPath, { recursive: true });
+      await (0, import_promises28.mkdir)(slotPath, { recursive: true });
       if (templatePath) {
         await copyDirectoryRecursive2(templatePath, slotPath);
       }
@@ -17369,7 +17122,7 @@ var WorkspacePoolManager = class {
   /** Remove lock file to release a slot. */
   async releaseSlot(slot) {
     try {
-      await (0, import_promises29.unlink)(slot.lockPath);
+      await (0, import_promises28.unlink)(slot.lockPath);
     } catch {
     }
   }
@@ -17382,21 +17135,21 @@ var WorkspacePoolManager = class {
   async tryLock(lockPath) {
     for (let attempt = 0; attempt < 3; attempt++) {
       try {
-        await (0, import_promises29.writeFile)(lockPath, String(process.pid), { flag: "wx" });
+        await (0, import_promises28.writeFile)(lockPath, String(process.pid), { flag: "wx" });
         return true;
       } catch (err) {
         if (err.code !== "EEXIST") {
           throw err;
         }
         try {
-          const pidStr = await (0, import_promises29.readFile)(lockPath, "utf-8");
+          const pidStr = await (0, import_promises28.readFile)(lockPath, "utf-8");
           const pid = Number.parseInt(pidStr.trim(), 10);
           if (!Number.isNaN(pid)) {
             try {
               process.kill(pid, 0);
               return false;
             } catch {
-              await (0, import_promises29.unlink)(lockPath).catch(() => {
+              await (0, import_promises28.unlink)(lockPath).catch(() => {
               });
               continue;
             }
@@ -17414,9 +17167,9 @@ var WorkspacePoolManager = class {
    * Returns false (no drift) if metadata.json doesn't exist (first use).
    */
   async checkDrift(poolDir, fingerprint) {
-    const metadataPath = import_node_path43.default.join(poolDir, "metadata.json");
+    const metadataPath = import_node_path42.default.join(poolDir, "metadata.json");
     try {
-      const raw = await (0, import_promises29.readFile)(metadataPath, "utf-8");
+      const raw = await (0, import_promises28.readFile)(metadataPath, "utf-8");
       const metadata = JSON.parse(raw);
       return metadata.fingerprint !== fingerprint;
     } catch {
@@ -17431,17 +17184,17 @@ var WorkspacePoolManager = class {
       repos,
       createdAt: (/* @__PURE__ */ new Date()).toISOString()
     };
-    await (0, import_promises29.writeFile)(import_node_path43.default.join(poolDir, "metadata.json"), JSON.stringify(metadata, null, 2));
+    await (0, import_promises28.writeFile)(import_node_path42.default.join(poolDir, "metadata.json"), JSON.stringify(metadata, null, 2));
   }
   /** Remove all slot directories and their lock files from a pool directory. */
   async removeAllSlots(poolDir) {
-    const entries = await (0, import_promises29.readdir)(poolDir);
+    const entries = await (0, import_promises28.readdir)(poolDir);
     for (const entry of entries) {
       if (entry.startsWith("slot-") && !entry.endsWith(".lock")) {
-        const lockPath = import_node_path43.default.join(poolDir, `${entry}.lock`);
+        const lockPath = import_node_path42.default.join(poolDir, `${entry}.lock`);
         if ((0, import_node_fs12.existsSync)(lockPath)) {
           try {
-            const pidStr = await (0, import_promises29.readFile)(lockPath, "utf-8");
+            const pidStr = await (0, import_promises28.readFile)(lockPath, "utf-8");
             const pid = Number.parseInt(pidStr.trim(), 10);
             if (!Number.isNaN(pid)) {
               try {
@@ -17454,12 +17207,12 @@ var WorkspacePoolManager = class {
           } catch {
           }
         }
-        await (0, import_promises29.rm)(import_node_path43.default.join(poolDir, entry), { recursive: true, force: true });
-        await (0, import_promises29.rm)(lockPath, { force: true }).catch(() => {
+        await (0, import_promises28.rm)(import_node_path42.default.join(poolDir, entry), { recursive: true, force: true });
+        await (0, import_promises28.rm)(lockPath, { force: true }).catch(() => {
         });
       }
     }
-    await (0, import_promises29.rm)(import_node_path43.default.join(poolDir, "metadata.json"), { force: true }).catch(() => {
+    await (0, import_promises28.rm)(import_node_path42.default.join(poolDir, "metadata.json"), { force: true }).catch(() => {
     });
   }
   /**
@@ -17469,7 +17222,7 @@ var WorkspacePoolManager = class {
    */
   async resetSlot(slotPath, templatePath, repos, poolReset = "fast") {
     for (const repo of repos) {
-      const repoDir = import_node_path43.default.join(slotPath, repo.path);
+      const repoDir = import_node_path42.default.join(slotPath, repo.path);
       if (!(0, import_node_fs12.existsSync)(repoDir)) {
         continue;
       }
@@ -17496,7 +17249,7 @@ var WorkspacePoolManager = class {
 // src/evaluation/workspace/repo-manager.ts
 var import_node_child_process9 = require("child_process");
 var import_node_fs13 = require("fs");
-var import_node_path44 = __toESM(require("path"), 1);
+var import_node_path43 = __toESM(require("path"), 1);
 var import_node_util6 = require("util");
 var execFileAsync2 = (0, import_node_util6.promisify)(import_node_child_process9.execFile);
 var DEFAULT_TIMEOUT_MS2 = 3e5;
@@ -17596,7 +17349,7 @@ ${lines.join("\n")}`;
    * Handles checkout, ref resolution, ancestor walking, shallow clone, sparse checkout.
    */
   async materialize(repo, workspacePath) {
-    const targetDir = import_node_path44.default.join(workspacePath, repo.path);
+    const targetDir = import_node_path43.default.join(workspacePath, repo.path);
     const sourceUrl = getSourceUrl(repo.source);
     const startedAt = Date.now();
     if (this.verbose) {
@@ -17687,7 +17440,7 @@ ${lines.join("\n")}`;
   async reset(repos, workspacePath, reset) {
     const cleanFlag = reset === "strict" ? "-fdx" : "-fd";
     for (const repo of repos) {
-      const targetDir = import_node_path44.default.join(workspacePath, repo.path);
+      const targetDir = import_node_path43.default.join(workspacePath, repo.path);
       await this.runGit(["reset", "--hard", "HEAD"], { cwd: targetDir });
       await this.runGit(["clean", cleanFlag], { cwd: targetDir });
     }
@@ -17695,36 +17448,36 @@ ${lines.join("\n")}`;
 };
 
 // src/evaluation/workspace/resolve.ts
-var import_promises30 = require("fs/promises");
-var import_node_path45 = __toESM(require("path"), 1);
+var import_promises29 = require("fs/promises");
+var import_node_path44 = __toESM(require("path"), 1);
 async function resolveWorkspaceTemplate(templatePath) {
   if (!templatePath) {
     return void 0;
   }
-  const resolved = import_node_path45.default.resolve(templatePath);
-  const stats = await (0, import_promises30.stat)(resolved);
+  const resolved = import_node_path44.default.resolve(templatePath);
+  const stats = await (0, import_promises29.stat)(resolved);
   if (stats.isFile()) {
     return {
-      dir: import_node_path45.default.dirname(resolved),
+      dir: import_node_path44.default.dirname(resolved),
       workspaceFile: resolved
     };
   }
   if (!stats.isDirectory()) {
     throw new Error(`workspace template is neither a file nor a directory: ${resolved}`);
   }
-  const entries = await (0, import_promises30.readdir)(resolved);
+  const entries = await (0, import_promises29.readdir)(resolved);
   const workspaceFiles = entries.filter((e) => e.endsWith(".code-workspace"));
   if (workspaceFiles.length === 1) {
     return {
       dir: resolved,
-      workspaceFile: import_node_path45.default.join(resolved, workspaceFiles[0])
+      workspaceFile: import_node_path44.default.join(resolved, workspaceFiles[0])
     };
   }
   if (workspaceFiles.length > 1) {
     const conventionFile = workspaceFiles.find((f) => f === "template.code-workspace");
     return {
       dir: resolved,
-      workspaceFile: conventionFile ? import_node_path45.default.join(resolved, conventionFile) : void 0
+      workspaceFile: conventionFile ? import_node_path44.default.join(resolved, conventionFile) : void 0
     };
   }
   return { dir: resolved };
@@ -17928,7 +17681,7 @@ async function runEvaluation(options) {
   ];
   const evaluatorRegistry = buildEvaluatorRegistry(evaluators, resolveGraderProvider);
   const typeRegistry = createBuiltinRegistry();
-  const discoveryBaseDir = evalFilePath ? import_node_path46.default.dirname(import_node_path46.default.resolve(evalFilePath)) : process.cwd();
+  const discoveryBaseDir = evalFilePath ? import_node_path45.default.dirname(import_node_path45.default.resolve(evalFilePath)) : process.cwd();
   const evalDir = discoveryBaseDir;
   await discoverAssertions(typeRegistry, discoveryBaseDir);
   await discoverGraders(typeRegistry, discoveryBaseDir);
@@ -18060,14 +17813,14 @@ async function runEvaluation(options) {
   let staticMaterialised = false;
   if (useStaticWorkspace && configuredStaticPath) {
     const isYamlConfiguredPath = !cliWorkspacePath && !!yamlWorkspacePath;
-    const dirExists = await (0, import_promises31.stat)(configuredStaticPath).then(
+    const dirExists = await (0, import_promises30.stat)(configuredStaticPath).then(
       (s) => s.isDirectory(),
       () => false
     );
-    const isEmpty = dirExists ? (await (0, import_promises31.readdir)(configuredStaticPath)).length === 0 : false;
+    const isEmpty = dirExists ? (await (0, import_promises30.readdir)(configuredStaticPath)).length === 0 : false;
     if (isYamlConfiguredPath && (!dirExists || isEmpty)) {
       if (!dirExists) {
-        await (0, import_promises31.mkdir)(configuredStaticPath, { recursive: true });
+        await (0, import_promises30.mkdir)(configuredStaticPath, { recursive: true });
       }
       if (workspaceTemplate) {
         await copyDirectoryRecursive(workspaceTemplate, configuredStaticPath);
@@ -18112,14 +17865,14 @@ async function runEvaluation(options) {
     }
   } else if (suiteWorkspace?.hooks || suiteWorkspace?.repos?.length && !isPerTestIsolation) {
     sharedWorkspacePath = getWorkspacePath(evalRunId, "shared");
-    await (0, import_promises31.mkdir)(sharedWorkspacePath, { recursive: true });
+    await (0, import_promises30.mkdir)(sharedWorkspacePath, { recursive: true });
     setupLog(`created empty shared workspace at: ${sharedWorkspacePath}`);
   }
   try {
     if (suiteWorkspaceFile && sharedWorkspacePath) {
-      const copiedWorkspaceFile = import_node_path46.default.join(sharedWorkspacePath, import_node_path46.default.basename(suiteWorkspaceFile));
+      const copiedWorkspaceFile = import_node_path45.default.join(sharedWorkspacePath, import_node_path45.default.basename(suiteWorkspaceFile));
       try {
-        await (0, import_promises31.stat)(copiedWorkspaceFile);
+        await (0, import_promises30.stat)(copiedWorkspaceFile);
         suiteWorkspaceFile = copiedWorkspaceFile;
       } catch {
       }
@@ -18227,10 +17980,10 @@ async function runEvaluation(options) {
           const budgetResult = {
             timestamp: (now ?? (() => /* @__PURE__ */ new Date()))().toISOString(),
             testId: evalCase.id,
-            dataset: evalCase.dataset,
+            eval_set: evalCase.eval_set,
             score: 0,
             assertions: [],
-            outputText: "",
+            output: [],
             target: target.name,
             error: `Suite budget exceeded ($${cumulativeBudgetCost.toFixed(4)} / $${totalBudgetUsd.toFixed(4)})`,
             budgetExceeded: true,
@@ -18263,10 +18016,10 @@ async function runEvaluation(options) {
           const haltResult = {
             timestamp: (now ?? (() => /* @__PURE__ */ new Date()))().toISOString(),
             testId: evalCase.id,
-            dataset: evalCase.dataset,
+            eval_set: evalCase.eval_set,
             score: 0,
             assertions: [],
-            outputText: "",
+            output: [],
             target: target.name,
             error: errorMsg,
             executionStatus: "execution_error",
@@ -18495,8 +18248,6 @@ async function runBatchEvaluation(options) {
     const promptInputs = promptInputsList[index];
     return {
       question: promptInputs.question,
-      guidelines: promptInputs.guidelines,
-      guideline_patterns: evalCase.guideline_patterns,
       inputFiles: evalCase.file_paths,
       evalCaseId: evalCase.id,
       metadata: {
@@ -18694,9 +18445,9 @@ async function runEvalCase(options) {
         );
       }
       if (caseWorkspaceFile && workspacePath) {
-        const copiedFile = import_node_path46.default.join(workspacePath, import_node_path46.default.basename(caseWorkspaceFile));
+        const copiedFile = import_node_path45.default.join(workspacePath, import_node_path45.default.basename(caseWorkspaceFile));
         try {
-          await (0, import_promises31.stat)(copiedFile);
+          await (0, import_promises30.stat)(copiedFile);
           caseWorkspaceFile = copiedFile;
         } catch {
         }
@@ -18704,7 +18455,7 @@ async function runEvalCase(options) {
     }
     if (!workspacePath && (evalCase.workspace?.hooks || evalCase.workspace?.repos?.length) && evalRunId) {
       workspacePath = getWorkspacePath(evalRunId, evalCase.id);
-      await (0, import_promises31.mkdir)(workspacePath, { recursive: true });
+      await (0, import_promises30.mkdir)(workspacePath, { recursive: true });
     }
     if (evalCase.workspace?.repos?.length && workspacePath) {
       const localPathErrors = RepoManager.validateLocalPaths(evalCase.workspace.repos);
@@ -18754,11 +18505,11 @@ async function runEvalCase(options) {
       const files = evalCase.metadata.agent_skills_files;
       if (baseDir && files.length > 0) {
         for (const relPath of files) {
-          const srcPath = import_node_path46.default.resolve(baseDir, relPath);
-          const destPath = import_node_path46.default.resolve(workspacePath, relPath);
+          const srcPath = import_node_path45.default.resolve(baseDir, relPath);
+          const destPath = import_node_path45.default.resolve(workspacePath, relPath);
           try {
-            await (0, import_promises31.mkdir)(import_node_path46.default.dirname(destPath), { recursive: true });
-            await (0, import_promises31.copyFile)(srcPath, destPath);
+            await (0, import_promises30.mkdir)(import_node_path45.default.dirname(destPath), { recursive: true });
+            await (0, import_promises30.copyFile)(srcPath, destPath);
           } catch (error) {
             const message = error instanceof Error ? error.message : String(error);
             return buildErrorResult(
@@ -19204,8 +18955,7 @@ async function evaluateCandidate(options) {
   let lmRequest;
   if (isAgentProvider(provider)) {
     agentRequest = {
-      question: promptInputs.question,
-      guideline_paths: evalCase.guideline_paths
+      question: promptInputs.question
     };
   } else {
     if (promptInputs.chatPrompt) {
@@ -19214,8 +18964,7 @@ async function evaluateCandidate(options) {
       };
     } else {
       lmRequest = {
-        question: promptInputs.question,
-        guidelines: promptInputs.guidelines
+        question: promptInputs.question
       };
     }
   }
@@ -19229,11 +18978,10 @@ async function evaluateCandidate(options) {
   return {
     timestamp: completedAt.toISOString(),
     testId: evalCase.id,
-    dataset: evalCase.dataset,
+    eval_set: evalCase.eval_set,
     conversationId: evalCase.conversation_id,
     score: score.score,
     assertions: score.assertions,
-    outputText: candidate,
     target: target.name,
     tokenUsage,
     costUsd,
@@ -19244,7 +18992,7 @@ async function evaluateCandidate(options) {
     input,
     scores,
     trace: trace2,
-    output,
+    output: output ?? [{ role: "assistant", content: candidate }],
     fileChanges,
     executionStatus: classifyQualityStatus(score.score)
   };
@@ -19378,7 +19126,7 @@ async function runEvaluatorList(options) {
     fileChanges,
     workspacePath
   };
-  const evalFileDir = evalCase.guideline_paths[0] ? import_node_path46.default.dirname(evalCase.guideline_paths[0]) : process.cwd();
+  const evalFileDir = evalCase.file_paths[0] ? import_node_path45.default.dirname(evalCase.file_paths[0]) : process.cwd();
   const dispatchContext = {
     graderProvider,
     targetResolver,
@@ -19409,7 +19157,7 @@ async function runEvaluatorList(options) {
         weight,
         verdict: score2.verdict,
         assertions: score2.assertions,
-        evaluatorProviderRequest: score2.evaluatorRawRequest,
+        input: score2.evaluatorRawRequest,
         details: score2.details,
         scores: mapChildResults(score2.scores),
         tokenUsage: score2.tokenUsage,
@@ -19492,7 +19240,7 @@ function filterEvalCases(evalCases, filter) {
   if (!filter) {
     return evalCases;
   }
-  return evalCases.filter((evalCase) => import_micromatch4.default.isMatch(evalCase.id, filter));
+  return evalCases.filter((evalCase) => import_micromatch3.default.isMatch(evalCase.id, filter));
 }
 function buildEvaluatorRegistry(overrides, resolveGraderProvider) {
   const llmGrader = overrides?.["llm-grader"] ?? overrides?.["llm-judge"] ?? new LlmGraderEvaluator({
@@ -19529,8 +19277,6 @@ async function invokeProvider(provider, options) {
     const braintrustSpanIds = streamCallbacks?.getActiveSpanIds?.() ?? void 0;
     return await provider.invoke({
       question: promptInputs.question,
-      guidelines: promptInputs.guidelines,
-      guideline_patterns: evalCase.guideline_patterns,
       chatPrompt: promptInputs.chatPrompt,
       inputFiles: evalCase.file_paths,
       evalCaseId: evalCase.id,
@@ -19558,21 +19304,17 @@ function buildErrorResult(evalCase, targetName, timestamp, error, promptInputs,
   if (isAgentProvider(provider)) {
     agentRequest = {
       question: promptInputs.question,
-      guideline_paths: evalCase.guideline_paths,
       error: message
     };
   } else {
     if (promptInputs.chatPrompt) {
       lmRequest = {
         chat_prompt: promptInputs.chatPrompt,
-        guideline_paths: evalCase.guideline_paths,
         error: message
       };
     } else {
       lmRequest = {
         question: promptInputs.question,
-        guidelines: promptInputs.guidelines,
-        guideline_paths: evalCase.guideline_paths,
         error: message
       };
     }
@@ -19585,11 +19327,11 @@ function buildErrorResult(evalCase, targetName, timestamp, error, promptInputs,
   return {
     timestamp: timestamp.toISOString(),
     testId: evalCase.id,
-    dataset: evalCase.dataset,
+    eval_set: evalCase.eval_set,
     conversationId: evalCase.conversation_id,
     score: 0,
     assertions: [{ text: `Error: ${message}`, passed: false }],
-    outputText: `Error occurred: ${message}`,
+    output: [{ role: "assistant", content: `Error occurred: ${message}` }],
     target: targetName,
     requests,
     input,
@@ -19618,7 +19360,6 @@ function createCacheKey(provider, target, evalCase, promptInputs) {
   hash.update(target.name);
   hash.update(evalCase.id);
   hash.update(promptInputs.question);
-  hash.update(promptInputs.guidelines);
   hash.update(promptInputs.systemMessage ?? "");
   if (promptInputs.chatPrompt) {
     hash.update(JSON.stringify(promptInputs.chatPrompt));
@@ -19633,7 +19374,7 @@ function buildResultInput(promptInputs) {
       content: message.content
     }));
   }
-  return promptInputs.question;
+  return [{ role: "user", content: promptInputs.question }];
 }
 function aggregateEvaluatorTokenUsage(scores) {
   if (!scores || scores.length === 0) return void 0;
@@ -19699,7 +19440,7 @@ function mapChildResults(children) {
     weight: child.weight,
     verdict: child.verdict,
     assertions: child.assertions,
-    evaluatorProviderRequest: child.evaluatorRawRequest,
+    input: child.evaluatorRawRequest,
     scores: mapChildResults(child.scores),
     details: child.details,
     tokenUsage: child.tokenUsage
@@ -19718,7 +19459,7 @@ function computeWeightedMean(entries) {
 
 // src/evaluation/evaluate.ts
 var import_node_fs14 = require("fs");
-var import_node_path47 = __toESM(require("path"), 1);
+var import_node_path46 = __toESM(require("path"), 1);
 
 // src/evaluation/providers/function-provider.ts
 function createFunctionProvider(taskFn) {
@@ -19755,7 +19496,7 @@ async function evaluate(config) {
   }
   const gitRoot = await findGitRoot(process.cwd());
   const repoRoot = gitRoot ?? process.cwd();
-  const testFilePath = config.specFile ? import_node_path47.default.resolve(config.specFile) : import_node_path47.default.join(process.cwd(), "__programmatic__.yaml");
+  const testFilePath = config.specFile ? import_node_path46.default.resolve(config.specFile) : import_node_path46.default.join(process.cwd(), "__programmatic__.yaml");
   await loadEnvHierarchy(repoRoot, testFilePath);
   let resolvedTarget;
   let taskProvider;
@@ -19821,8 +19562,6 @@ async function evaluate(config) {
         input_segments: inputSegments,
         expected_output: expectedOutput,
         reference_answer: expectedOutputValue,
-        guideline_paths: [],
-        guideline_patterns: [],
         file_paths: [],
         assertions: assertConfigs.length > 0 ? assertConfigs : void 0,
         metadata: test.metadata
@@ -19884,10 +19623,10 @@ function computeSummary(results, durationMs) {
 var TARGET_FILE_CANDIDATES = [".agentv/targets.yaml", ".agentv/targets.yml"];
 async function discoverDefaultTarget(repoRoot) {
   const cwd = process.cwd();
-  const chain = buildDirectoryChain2(import_node_path47.default.join(cwd, "_placeholder"), repoRoot);
+  const chain = buildDirectoryChain2(import_node_path46.default.join(cwd, "_placeholder"), repoRoot);
   for (const dir of chain) {
     for (const candidate of TARGET_FILE_CANDIDATES) {
-      const targetsPath = import_node_path47.default.join(dir, candidate);
+      const targetsPath = import_node_path46.default.join(dir, candidate);
       if (!(0, import_node_fs14.existsSync)(targetsPath)) continue;
       try {
         const definitions = await readTargetDefinitions(targetsPath);
@@ -19904,7 +19643,7 @@ async function loadEnvHierarchy(repoRoot, startPath) {
   const chain = buildDirectoryChain2(startPath, repoRoot);
   const envFiles = [];
   for (const dir of chain) {
-    const envPath = import_node_path47.default.join(dir, ".env");
+    const envPath = import_node_path46.default.join(dir, ".env");
     if ((0, import_node_fs14.existsSync)(envPath)) envFiles.push(envPath);
   }
   for (let i = 0; i < envFiles.length; i++) {
@@ -20085,8 +19824,8 @@ function buildPrompt(criteria, question, referenceAnswer) {
 }
 
 // src/evaluation/cache/response-cache.ts
-var import_promises32 = require("fs/promises");
-var import_node_path48 = __toESM(require("path"), 1);
+var import_promises31 = require("fs/promises");
+var import_node_path47 = __toESM(require("path"), 1);
 var DEFAULT_CACHE_PATH = ".agentv/cache";
 var ResponseCache = class {
   cachePath;
@@ -20096,7 +19835,7 @@ var ResponseCache = class {
   async get(key) {
     const filePath = this.keyToPath(key);
     try {
-      const data = await (0, import_promises32.readFile)(filePath, "utf8");
+      const data = await (0, import_promises31.readFile)(filePath, "utf8");
       return JSON.parse(data);
     } catch {
       return void 0;
@@ -20104,13 +19843,13 @@ var ResponseCache = class {
   }
   async set(key, value) {
     const filePath = this.keyToPath(key);
-    const dir = import_node_path48.default.dirname(filePath);
-    await (0, import_promises32.mkdir)(dir, { recursive: true });
-    await (0, import_promises32.writeFile)(filePath, JSON.stringify(value, null, 2), "utf8");
+    const dir = import_node_path47.default.dirname(filePath);
+    await (0, import_promises31.mkdir)(dir, { recursive: true });
+    await (0, import_promises31.writeFile)(filePath, JSON.stringify(value, null, 2), "utf8");
   }
   keyToPath(key) {
     const prefix = key.slice(0, 2);
-    return import_node_path48.default.join(this.cachePath, prefix, `${key}.json`);
+    return import_node_path47.default.join(this.cachePath, prefix, `${key}.json`);
   }
 };
 function shouldEnableCache(params) {
@@ -20127,7 +19866,6 @@ function shouldSkipCacheForTemperature(targetConfig) {
 
 // src/evaluation/baseline.ts
 var STRIPPED_TOP_LEVEL_FIELDS = /* @__PURE__ */ new Set([
-  "outputText",
   "requests",
   "trace",
   "workspacePath",
@@ -20144,7 +19882,7 @@ var STRIPPED_TOP_LEVEL_FIELDS = /* @__PURE__ */ new Set([
   "startTime",
   "endTime"
 ]);
-var STRIPPED_EVALUATOR_FIELDS = /* @__PURE__ */ new Set(["rawRequest", "evaluatorProviderRequest"]);
+var STRIPPED_EVALUATOR_FIELDS = /* @__PURE__ */ new Set(["rawRequest", "input"]);
 function trimEvaluatorResult(result) {
   const trimmed = {};
   for (const [key, value] of Object.entries(result)) {
@@ -20299,9 +20037,13 @@ var OtelTraceExporter = class {
         rootSpan.setAttribute("gen_ai.system", "agentv");
         rootSpan.setAttribute("agentv.test_id", result.testId);
         rootSpan.setAttribute("agentv.target", result.target);
-        if (result.dataset) rootSpan.setAttribute("agentv.dataset", result.dataset);
+        if (result.eval_set) rootSpan.setAttribute("agentv.eval_set", result.eval_set);
         rootSpan.setAttribute("agentv.score", result.score);
-        if (captureContent) rootSpan.setAttribute("agentv.output_text", result.outputText);
+        if (captureContent && result.output.length > 0) {
+          const lastMsg = result.output[result.output.length - 1];
+          const text = typeof lastMsg.content === "string" ? lastMsg.content : JSON.stringify(lastMsg.content);
+          rootSpan.setAttribute("agentv.output_text", text);
+        }
         if (result.durationMs != null)
           rootSpan.setAttribute("agentv.trace.duration_ms", result.durationMs);
         if (result.costUsd != null) rootSpan.setAttribute("agentv.trace.cost_usd", result.costUsd);
@@ -20480,14 +20222,14 @@ var OtelStreamingObserver = class {
   // biome-ignore lint/suspicious/noExplicitAny: OTel context loaded dynamically
   rootCtx = null;
   /** Create root eval span immediately (visible in backend right away) */
-  startEvalCase(testId, target, dataset) {
+  startEvalCase(testId, target, evalSet) {
     const ctx = this.parentCtx ?? this.api.context.active();
     this.rootSpan = this.tracer.startSpan("agentv.eval", void 0, ctx);
     this.rootSpan.setAttribute("gen_ai.operation.name", "evaluate");
     this.rootSpan.setAttribute("gen_ai.system", "agentv");
     this.rootSpan.setAttribute("agentv.test_id", testId);
     this.rootSpan.setAttribute("agentv.target", target);
-    if (dataset) this.rootSpan.setAttribute("agentv.dataset", dataset);
+    if (evalSet) this.rootSpan.setAttribute("agentv.eval_set", evalSet);
     this.rootCtx = this.api.trace.setSpan(this.api.context.active(), this.rootSpan);
   }
   /** Create and immediately export a tool span */
@@ -20680,7 +20422,6 @@ function createAgentKernel() {
   initializeBaseline,
   isAgentSkillsFormat,
   isEvaluatorKind,
-  isGuidelineFile,
   isJsonObject,
   isJsonValue,
   isNonEmptyString,