@agentv/core 2.1.1 → 2.5.1

package/dist/index.js

@@ -10,7 +10,7 @@ import {
   readTextFile,
   resolveFileReference,
   resolveTargetDefinition
-} from "./chunk-KDEP4I7G.js";
+} from "./chunk-RP3M7COZ.js";
 
 // src/evaluation/types.ts
 var TEST_MESSAGE_ROLE_VALUES = ["system", "user", "assistant", "tool"];
@@ -146,8 +146,9 @@ function mergeExecutionMetrics(summary, metrics) {
 }
 
 // src/evaluation/yaml-parser.ts
-import { readFile as readFile5 } from "node:fs/promises";
-import path6 from "node:path";
+import { readFile as readFile6 } from "node:fs/promises";
+import path7 from "node:path";
+import micromatch3 from "micromatch";
 import { parse as parse2 } from "yaml";
 
 // src/evaluation/loaders/config-loader.ts
@@ -462,11 +463,11 @@ async function parseEvaluators(rawEvalCase, globalExecution, searchRoots, evalId
           );
         }
       }
-      const knownProps = /* @__PURE__ */ new Set(["name", "type", "script", "cwd", "weight", "target"]);
-      const config = {};
+      const knownProps2 = /* @__PURE__ */ new Set(["name", "type", "script", "cwd", "weight", "target"]);
+      const config2 = {};
       for (const [key, value] of Object.entries(rawEvaluator)) {
-        if (!knownProps.has(key) && value !== void 0) {
-          config[key] = value;
+        if (!knownProps2.has(key) && value !== void 0) {
+          config2[key] = value;
         }
       }
       evaluators.push({
@@ -476,7 +477,7 @@ async function parseEvaluators(rawEvalCase, globalExecution, searchRoots, evalId
         cwd,
         resolvedCwd,
         ...weight2 !== void 0 ? { weight: weight2 } : {},
-        ...Object.keys(config).length > 0 ? { config } : {},
+        ...Object.keys(config2).length > 0 ? { config: config2 } : {},
         ...targetConfig !== void 0 ? { target: targetConfig } : {}
       });
       continue;
@@ -641,7 +642,7 @@ async function parseEvaluators(rawEvalCase, globalExecution, searchRoots, evalId
         continue;
       }
       const weight2 = validateWeight(rawEvaluator.weight, name, evalId);
-      const config = {
+      const config2 = {
         name,
         type: "tool_trajectory",
         mode,
@@ -649,7 +650,7 @@ async function parseEvaluators(rawEvalCase, globalExecution, searchRoots, evalId
         ...expected ? { expected } : {},
         ...weight2 !== void 0 ? { weight: weight2 } : {}
       };
-      evaluators.push(config);
+      evaluators.push(config2);
       continue;
     }
     if (typeValue === "field_accuracy") {
@@ -786,9 +787,33 @@ async function parseEvaluators(rawEvalCase, globalExecution, searchRoots, evalId
       });
       continue;
     }
-    const prompt = asString(rawEvaluator.prompt);
+    const rawPrompt = rawEvaluator.prompt;
+    let prompt;
     let promptPath;
-    if (prompt) {
+    let resolvedPromptScript;
+    let promptScriptConfig;
+    if (isJsonObject2(rawPrompt)) {
+      const scriptArray = asStringArray(
+        rawPrompt.script,
+        `prompt.script for evaluator '${name}' in '${evalId}'`
+      );
+      if (!scriptArray) {
+        throw new Error(`Evaluator '${name}' in '${evalId}': prompt object requires script array`);
+      }
+      const scriptPath = scriptArray[scriptArray.length - 1];
+      const resolved = await resolveFileReference2(scriptPath, searchRoots);
+      if (resolved.resolvedPath) {
+        resolvedPromptScript = [...scriptArray.slice(0, -1), path3.resolve(resolved.resolvedPath)];
+      } else {
+        throw new Error(
+          `Evaluator '${name}' in '${evalId}': prompt script file not found: ${resolved.displayPath}`
+        );
+      }
+      if (isJsonObject2(rawPrompt.config)) {
+        promptScriptConfig = rawPrompt.config;
+      }
+    } else if (typeof rawPrompt === "string") {
+      prompt = rawPrompt;
       const resolved = await resolveFileReference2(prompt, searchRoots);
       if (resolved.resolvedPath) {
         promptPath = path3.resolve(resolved.resolvedPath);
@@ -807,12 +832,7 @@ async function parseEvaluators(rawEvalCase, globalExecution, searchRoots, evalId
     }
     const _model = asString(rawEvaluator.model);
     const rawRubrics = rawEvaluator.rubrics;
-    const parsedRubrics = Array.isArray(rawRubrics) ? rawRubrics.filter((r) => isJsonObject2(r)).map((rubric, index) => ({
-      id: asString(rubric.id) ?? `rubric-${index + 1}`,
-      description: asString(rubric.description) ?? "",
-      weight: typeof rubric.weight === "number" ? rubric.weight : 1,
-      required: typeof rubric.required === "boolean" ? rubric.required : true
-    })).filter((r) => r.description.length > 0) : void 0;
+    const parsedRubrics = Array.isArray(rawRubrics) ? parseRubricItems(rawRubrics, name, evalId) : void 0;
     if (typeValue === "rubric") {
       if (!parsedRubrics) {
         logWarning2(`Skipping rubric evaluator '${name}' in '${evalId}': missing rubrics array`);
@@ -832,13 +852,26 @@ async function parseEvaluators(rawEvalCase, globalExecution, searchRoots, evalId
       continue;
     }
     const weight = validateWeight(rawEvaluator.weight, name, evalId);
+    const knownProps = /* @__PURE__ */ new Set(["name", "type", "prompt", "model", "rubrics", "weight", "config"]);
+    const config = {};
+    for (const [key, value] of Object.entries(rawEvaluator)) {
+      if (!knownProps.has(key) && value !== void 0) {
+        config[key] = value;
+      }
+    }
+    const topLevelConfig = isJsonObject2(rawEvaluator.config) ? rawEvaluator.config : {};
+    const mergedConfig = { ...config, ...topLevelConfig };
+    const finalConfig = promptScriptConfig ?? (Object.keys(mergedConfig).length > 0 ? mergedConfig : void 0);
     evaluators.push({
       name,
       type: "llm_judge",
       prompt,
       promptPath,
+      ...promptPath ? { resolvedPromptPath: promptPath } : {},
+      ...resolvedPromptScript ? { resolvedPromptScript } : {},
       ...parsedRubrics && parsedRubrics.length > 0 ? { rubrics: parsedRubrics } : {},
-      ...weight !== void 0 ? { weight } : {}
+      ...weight !== void 0 ? { weight } : {},
+      ...finalConfig ? { config: finalConfig } : {}
     });
   }
   return evaluators.length > 0 ? evaluators : void 0;
@@ -925,6 +958,191 @@ var VALID_FIELD_AGGREGATION_TYPES = /* @__PURE__ */ new Set(["weighted_average",
 function isValidFieldAggregationType(value) {
   return typeof value === "string" && VALID_FIELD_AGGREGATION_TYPES.has(value);
 }
+function parseRubricItems(rawRubrics, evaluatorName, evalId) {
+  const items = [];
+  for (const [index, rawRubric] of rawRubrics.entries()) {
+    if (!isJsonObject2(rawRubric)) {
+      logWarning2(
+        `Skipping invalid rubric entry at index ${index} in evaluator '${evaluatorName}' (expected object)`
+      );
+      continue;
+    }
+    const id = asString(rawRubric.id) ?? `rubric-${index + 1}`;
+    const expectedOutcome = asString(rawRubric.expected_outcome) ?? asString(rawRubric.description) ?? "";
+    const weight = typeof rawRubric.weight === "number" ? rawRubric.weight : 1;
+    let requiredMinScore;
+    let required;
+    if (typeof rawRubric.required_min_score === "number") {
+      const minScore = rawRubric.required_min_score;
+      if (!Number.isInteger(minScore) || minScore < 0 || minScore > 10) {
+        throw new Error(
+          `Invalid required_min_score for rubric '${id}' in evaluator '${evaluatorName}' in '${evalId}': must be an integer 0-10 (got ${minScore})`
+        );
+      }
+      requiredMinScore = minScore;
+    }
+    if (typeof rawRubric.required === "boolean") {
+      required = rawRubric.required;
+    }
+    let scoreRanges;
+    const rawScoreRanges = rawRubric.score_ranges;
+    if (rawScoreRanges !== void 0) {
+      if (!Array.isArray(rawScoreRanges)) {
+        throw new Error(
+          `Invalid score_ranges for rubric '${id}' in evaluator '${evaluatorName}' in '${evalId}': must be an array`
+        );
+      }
+      scoreRanges = parseScoreRanges(rawScoreRanges, id, evaluatorName, evalId);
+      items.push({
+        id,
+        weight,
+        ...expectedOutcome.length > 0 ? { expected_outcome: expectedOutcome } : {},
+        ...required !== void 0 ? { required } : {},
+        ...requiredMinScore !== void 0 ? { required_min_score: requiredMinScore } : {},
+        score_ranges: scoreRanges
+      });
+    } else {
+      if (expectedOutcome.length === 0) {
+        logWarning2(
+          `Skipping rubric '${id}' in evaluator '${evaluatorName}' in '${evalId}': missing expected_outcome`
+        );
+        continue;
+      }
+      items.push({
+        id,
+        expected_outcome: expectedOutcome,
+        weight,
+        // Default to required: true if not specified (backward compatibility)
+        required: required ?? true,
+        ...requiredMinScore !== void 0 ? { required_min_score: requiredMinScore } : {}
+      });
+    }
+  }
+  return items.length > 0 ? items : void 0;
+}
+function parseScoreRanges(rawRanges, rubricId, evaluatorName, evalId) {
+  const ranges = [];
+  for (const [index, rawRange] of rawRanges.entries()) {
+    if (!isJsonObject2(rawRange)) {
+      throw new Error(
+        `Invalid score_range entry at index ${index} for rubric '${rubricId}' in evaluator '${evaluatorName}' in '${evalId}': expected object`
+      );
+    }
+    const scoreRangeValue = rawRange.score_range;
+    if (!Array.isArray(scoreRangeValue) || scoreRangeValue.length !== 2 || typeof scoreRangeValue[0] !== "number" || typeof scoreRangeValue[1] !== "number") {
+      throw new Error(
+        `Invalid score_range at index ${index} for rubric '${rubricId}' in evaluator '${evaluatorName}' in '${evalId}': must be [min, max] array of two numbers`
+      );
+    }
+    const [min, max] = scoreRangeValue;
+    if (!Number.isInteger(min) || !Number.isInteger(max)) {
+      throw new Error(
+        `Invalid score_range at index ${index} for rubric '${rubricId}' in evaluator '${evaluatorName}' in '${evalId}': values must be integers (got [${min}, ${max}])`
+      );
+    }
+    if (min < 0 || min > 10 || max < 0 || max > 10) {
+      throw new Error(
+        `Invalid score_range at index ${index} for rubric '${rubricId}' in evaluator '${evaluatorName}' in '${evalId}': values must be 0-10 (got [${min}, ${max}])`
+      );
+    }
+    if (min > max) {
+      throw new Error(
+        `Invalid score_range at index ${index} for rubric '${rubricId}' in evaluator '${evaluatorName}' in '${evalId}': min must be <= max (got [${min}, ${max}])`
+      );
+    }
+    const expectedOutcome = asString(rawRange.expected_outcome) ?? asString(rawRange.description) ?? "";
+    if (expectedOutcome.length === 0) {
+      throw new Error(
+        `Missing expected_outcome for score_range [${min}, ${max}] in rubric '${rubricId}' in evaluator '${evaluatorName}' in '${evalId}'`
+      );
+    }
+    ranges.push({
+      score_range: [min, max],
+      expected_outcome: expectedOutcome
+    });
+  }
+  const sortedRanges = [...ranges].sort((a, b) => a.score_range[0] - b.score_range[0]);
+  for (let i = 1; i < sortedRanges.length; i++) {
+    const prev = sortedRanges[i - 1];
+    const curr = sortedRanges[i];
+    if (curr.score_range[0] <= prev.score_range[1]) {
+      throw new Error(
+        `Overlapping score_ranges in rubric '${rubricId}' in evaluator '${evaluatorName}' in '${evalId}': [${prev.score_range[0]}, ${prev.score_range[1]}] overlaps with [${curr.score_range[0]}, ${curr.score_range[1]}]`
+      );
+    }
+  }
+  const covered = /* @__PURE__ */ new Set();
+  for (const range of ranges) {
+    for (let i = range.score_range[0]; i <= range.score_range[1]; i++) {
+      covered.add(i);
+    }
+  }
+  const missing = [];
+  for (let i = 0; i <= 10; i++) {
+    if (!covered.has(i)) {
+      missing.push(i);
+    }
+  }
+  if (missing.length > 0) {
+    throw new Error(
+      `Incomplete score_ranges coverage in rubric '${rubricId}' in evaluator '${evaluatorName}' in '${evalId}': missing coverage for scores: ${missing.join(", ")}. Ranges must cover all integers 0-10.`
+    );
+  }
+  return ranges;
+}
+function parseInlineRubrics(rawRubrics) {
+  const rubricItems = rawRubrics.filter((r) => isJsonObject2(r) || typeof r === "string").map((rubric, index) => {
+    if (typeof rubric === "string") {
+      return {
+        id: `rubric-${index + 1}`,
+        expected_outcome: rubric,
+        weight: 1,
+        required: true
+      };
+    }
+    const expectedOutcome = asString(rubric.expected_outcome) ?? asString(rubric.description) ?? "";
+    const rawScoreRanges = rubric.score_ranges;
+    const scoreRanges = Array.isArray(rawScoreRanges) && rawScoreRanges.length > 0 ? rawScoreRanges.filter((r) => isJsonObject2(r)).map((range) => ({
+      score_range: Array.isArray(range.score_range) ? range.score_range : [0, 10],
+      expected_outcome: asString(range.expected_outcome) ?? asString(range.description) ?? ""
+    })).filter((r) => r.expected_outcome.length > 0) : void 0;
+    const baseRubric = {
+      id: asString(rubric.id) ?? `rubric-${index + 1}`,
+      weight: typeof rubric.weight === "number" ? rubric.weight : 1
+    };
+    if (scoreRanges && scoreRanges.length > 0) {
+      return {
+        ...baseRubric,
+        ...expectedOutcome.length > 0 ? { expected_outcome: expectedOutcome } : {},
+        ...typeof rubric.required === "boolean" ? { required: rubric.required } : {},
+        ...typeof rubric.required_min_score === "number" ? { required_min_score: rubric.required_min_score } : {},
+        score_ranges: scoreRanges
+      };
+    }
+    return {
+      ...baseRubric,
+      expected_outcome: expectedOutcome,
+      required: typeof rubric.required === "boolean" ? rubric.required : true,
+      ...typeof rubric.required_min_score === "number" ? { required_min_score: rubric.required_min_score } : {}
+    };
+  }).filter(
+    (r) => r.expected_outcome && r.expected_outcome.length > 0 || "score_ranges" in r && r.score_ranges
+  );
+  if (rubricItems.length === 0) {
+    return void 0;
+  }
+  return {
+    name: "rubric",
+    type: "llm_judge",
+    rubrics: rubricItems
+  };
+}
+
+// src/evaluation/loaders/jsonl-parser.ts
+import { readFile as readFile4 } from "node:fs/promises";
+import path5 from "node:path";
+import micromatch2 from "micromatch";
+import { parse as parseYaml } from "yaml";
 
 // src/evaluation/loaders/message-processor.ts
 import { readFile as readFile3 } from "node:fs/promises";
@@ -1186,28 +1404,302 @@ async function processExpectedMessages(options) {
   return segments;
 }
 
-// src/evaluation/formatting/prompt-builder.ts
-import { readFile as readFile4 } from "node:fs/promises";
-import path5 from "node:path";
+// src/evaluation/loaders/shorthand-expansion.ts
+function expandInputShorthand(value) {
+  if (value === void 0 || value === null) {
+    return void 0;
+  }
+  if (typeof value === "string") {
+    return [{ role: "user", content: value }];
+  }
+  if (Array.isArray(value)) {
+    const messages = value.filter((msg) => isTestMessage(msg));
+    return messages.length > 0 ? messages : void 0;
+  }
+  return void 0;
+}
+function expandExpectedOutputShorthand(value) {
+  if (value === void 0 || value === null) {
+    return void 0;
+  }
+  if (typeof value === "string") {
+    return [{ role: "assistant", content: value }];
+  }
+  if (Array.isArray(value)) {
+    if (value.length > 0 && isJsonObject(value[0]) && "role" in value[0]) {
+      const messages = value.filter((msg) => isTestMessage(msg));
+      return messages.length > 0 ? messages : void 0;
+    }
+    return [{ role: "assistant", content: value }];
+  }
+  if (isJsonObject(value)) {
+    if ("role" in value) {
+      return isTestMessage(value) ? [value] : void 0;
+    }
+    return [{ role: "assistant", content: value }];
+  }
+  return void 0;
+}
+function resolveInputMessages(raw) {
+  if (raw.input_messages !== void 0) {
+    if (Array.isArray(raw.input_messages)) {
+      const messages = raw.input_messages.filter((msg) => isTestMessage(msg));
+      return messages.length > 0 ? messages : void 0;
+    }
+    return void 0;
+  }
+  return expandInputShorthand(raw.input);
+}
+function resolveExpectedMessages(raw) {
+  if (raw.expected_messages !== void 0) {
+    if (Array.isArray(raw.expected_messages)) {
+      const messages = raw.expected_messages.filter(
+        (msg) => isTestMessage(msg)
+      );
+      return messages.length > 0 ? messages : void 0;
+    }
+    return void 0;
+  }
+  return expandExpectedOutputShorthand(raw.expected_output);
+}
+
+// src/evaluation/loaders/jsonl-parser.ts
 var ANSI_YELLOW5 = "\x1B[33m";
+var ANSI_RED = "\x1B[31m";
 var ANSI_RESET5 = "\x1B[0m";
+function detectFormat(filePath) {
+  const ext = path5.extname(filePath).toLowerCase();
+  if (ext === ".jsonl") return "jsonl";
+  if (ext === ".yaml" || ext === ".yml") return "yaml";
+  throw new Error(`Unsupported file format: '${ext}'. Supported formats: .yaml, .yml, .jsonl`);
+}
+async function loadSidecarMetadata(jsonlPath, verbose) {
+  const dir = path5.dirname(jsonlPath);
+  const base = path5.basename(jsonlPath, ".jsonl");
+  const sidecarPath = path5.join(dir, `${base}.yaml`);
+  if (!await fileExists2(sidecarPath)) {
+    if (verbose) {
+      logWarning4(`Sidecar metadata file not found: ${sidecarPath} (using defaults)`);
+    }
+    return {};
+  }
+  try {
+    const content = await readFile4(sidecarPath, "utf8");
+    const parsed = parseYaml(content);
+    if (!isJsonObject(parsed)) {
+      logWarning4(`Invalid sidecar metadata format in ${sidecarPath}`);
+      return {};
+    }
+    return {
+      description: asString4(parsed.description),
+      dataset: asString4(parsed.dataset),
+      execution: isJsonObject(parsed.execution) ? parsed.execution : void 0,
+      evaluator: parsed.evaluator
+    };
+  } catch (error) {
+    logWarning4(`Could not read sidecar metadata from ${sidecarPath}: ${error.message}`);
+    return {};
+  }
+}
+function parseJsonlContent(content, filePath) {
+  const lines = content.split("\n");
+  const cases = [];
+  for (let i = 0; i < lines.length; i++) {
+    const line = lines[i].trim();
+    if (line === "") continue;
+    try {
+      const parsed = JSON.parse(line);
+      if (!isJsonObject(parsed)) {
+        throw new Error("Expected JSON object");
+      }
+      cases.push(parsed);
+    } catch (error) {
+      const message = error instanceof Error ? error.message : String(error);
+      throw new Error(`Line ${i + 1}: Invalid JSON - ${message}
+  File: ${filePath}`);
+    }
+  }
+  return cases;
+}
+async function loadEvalCasesFromJsonl(evalFilePath, repoRoot, options) {
+  const verbose = options?.verbose ?? false;
+  const filterPattern = options?.filter;
+  const absoluteTestPath = path5.resolve(evalFilePath);
+  const repoRootPath = resolveToAbsolutePath(repoRoot);
+  const searchRoots = buildSearchRoots2(absoluteTestPath, repoRootPath);
+  const config = await loadConfig(absoluteTestPath, repoRootPath);
+  const guidelinePatterns = config?.guideline_patterns;
+  const sidecar = await loadSidecarMetadata(absoluteTestPath, verbose);
+  const rawFile = await readFile4(absoluteTestPath, "utf8");
+  const rawCases = parseJsonlContent(rawFile, evalFilePath);
+  const fallbackDataset = path5.basename(absoluteTestPath, ".jsonl") || "eval";
+  const datasetName = sidecar.dataset && sidecar.dataset.trim().length > 0 ? sidecar.dataset : fallbackDataset;
+  const globalEvaluator = coerceEvaluator(sidecar.evaluator, "sidecar") ?? "llm_judge";
+  const globalExecution = sidecar.execution;
+  if (verbose) {
+    console.log(`
+[JSONL Dataset: ${evalFilePath}]`);
+    console.log(`  Cases: ${rawCases.length}`);
+    console.log(`  Dataset name: ${datasetName}`);
+    if (sidecar.description) {
+      console.log(`  Description: ${sidecar.description}`);
+    }
+  }
+  const results = [];
+  for (let lineIndex = 0; lineIndex < rawCases.length; lineIndex++) {
+    const evalcase = rawCases[lineIndex];
+    const lineNumber = lineIndex + 1;
+    const id = asString4(evalcase.id);
+    if (filterPattern && (!id || !micromatch2.isMatch(id, filterPattern))) {
+      continue;
+    }
+    const conversationId = asString4(evalcase.conversation_id);
+    const outcome = asString4(evalcase.expected_outcome) ?? asString4(evalcase.outcome);
+    const inputMessages = resolveInputMessages(evalcase);
+    const expectedMessages = resolveExpectedMessages(evalcase) ?? [];
+    if (!id || !outcome || !inputMessages || inputMessages.length === 0) {
+      logError(
+        `Skipping incomplete eval case at line ${lineNumber}: ${id ?? "unknown"}. Missing required fields: id, expected_outcome, and/or input_messages (or input)`
+      );
+      continue;
+    }
+    const hasExpectedMessages = expectedMessages.length > 0;
+    const guidelinePaths = [];
+    const inputTextParts = [];
+    const inputSegments = await processMessages({
+      messages: inputMessages,
+      searchRoots,
+      repoRootPath,
+      guidelinePatterns,
+      guidelinePaths,
+      textParts: inputTextParts,
+      messageType: "input",
+      verbose
+    });
+    const outputSegments = hasExpectedMessages ? await processExpectedMessages({
+      messages: expectedMessages,
+      searchRoots,
+      repoRootPath,
+      verbose
+    }) : [];
+    let referenceAnswer = "";
+    if (outputSegments.length > 0) {
+      const lastMessage = outputSegments[outputSegments.length - 1];
+      const content = lastMessage.content;
+      const toolCalls = lastMessage.tool_calls;
+      if (typeof content === "string") {
+        referenceAnswer = content;
+      } else if (content !== void 0 && content !== null) {
+        referenceAnswer = JSON.stringify(content, null, 2);
+      } else if (toolCalls !== void 0 && toolCalls !== null) {
+        referenceAnswer = JSON.stringify(toolCalls, null, 2);
+      }
+    }
+    const question = inputTextParts.map((part) => part.trim()).filter((part) => part.length > 0).join(" ");
+    const caseExecution = isJsonObject(evalcase.execution) ? evalcase.execution : void 0;
+    const mergedExecution = caseExecution ?? globalExecution;
+    const evalCaseEvaluatorKind = coerceEvaluator(evalcase.evaluator, id) ?? globalEvaluator;
+    let evaluators;
+    try {
+      evaluators = await parseEvaluators(evalcase, mergedExecution, searchRoots, id ?? "unknown");
+    } catch (error) {
+      const message = error instanceof Error ? error.message : String(error);
+      logError(`Skipping eval case '${id}' at line ${lineNumber}: ${message}`);
+      continue;
+    }
+    const inlineRubrics = evalcase.rubrics;
+    if (inlineRubrics !== void 0 && Array.isArray(inlineRubrics)) {
+      const rubricEvaluator = parseInlineRubrics(inlineRubrics);
+      if (rubricEvaluator) {
+        evaluators = evaluators ? [rubricEvaluator, ...evaluators] : [rubricEvaluator];
+      }
+    }
+    const userFilePaths = [];
+    for (const segment of inputSegments) {
+      if (segment.type === "file" && typeof segment.resolvedPath === "string") {
+        userFilePaths.push(segment.resolvedPath);
+      }
+    }
+    const allFilePaths = [
+      ...guidelinePaths.map((guidelinePath) => path5.resolve(guidelinePath)),
+      ...userFilePaths
+    ];
+    const testCase = {
+      id,
+      dataset: datasetName,
+      conversation_id: conversationId,
+      question,
+      input_messages: inputMessages,
+      input_segments: inputSegments,
+      expected_messages: outputSegments,
+      reference_answer: referenceAnswer,
+      guideline_paths: guidelinePaths.map((guidelinePath) => path5.resolve(guidelinePath)),
+      guideline_patterns: guidelinePatterns,
+      file_paths: allFilePaths,
+      expected_outcome: outcome,
+      evaluator: evalCaseEvaluatorKind,
+      evaluators
+    };
+    if (verbose) {
+      console.log(`
+[Eval Case: ${id}]`);
+      if (testCase.guideline_paths.length > 0) {
+        console.log(`  Guidelines used: ${testCase.guideline_paths.length}`);
+        for (const guidelinePath of testCase.guideline_paths) {
+          console.log(`    - ${guidelinePath}`);
+        }
+      } else {
+        console.log("  No guidelines found");
+      }
+    }
+    results.push(testCase);
+  }
+  return results;
+}
+function asString4(value) {
+  return typeof value === "string" ? value : void 0;
+}
+function logWarning4(message, details) {
+  if (details && details.length > 0) {
+    const detailBlock = details.join("\n");
+    console.warn(`${ANSI_YELLOW5}Warning: ${message}
+${detailBlock}${ANSI_RESET5}`);
+  } else {
+    console.warn(`${ANSI_YELLOW5}Warning: ${message}${ANSI_RESET5}`);
+  }
+}
+function logError(message, details) {
+  if (details && details.length > 0) {
+    const detailBlock = details.join("\n");
+    console.error(`${ANSI_RED}Error: ${message}
+${detailBlock}${ANSI_RESET5}`);
+  } else {
+    console.error(`${ANSI_RED}Error: ${message}${ANSI_RESET5}`);
+  }
+}
+
+// src/evaluation/formatting/prompt-builder.ts
+import { readFile as readFile5 } from "node:fs/promises";
+import path6 from "node:path";
+var ANSI_YELLOW6 = "\x1B[33m";
+var ANSI_RESET6 = "\x1B[0m";
 async function buildPromptInputs(testCase, mode = "lm") {
   const guidelineParts = [];
   for (const rawPath of testCase.guideline_paths) {
-    const absolutePath = path5.resolve(rawPath);
+    const absolutePath = path6.resolve(rawPath);
     if (!await fileExists2(absolutePath)) {
-      logWarning4(`Could not read guideline file ${absolutePath}: file does not exist`);
+      logWarning5(`Could not read guideline file ${absolutePath}: file does not exist`);
       continue;
     }
     try {
-      const content = (await readFile4(absolutePath, "utf8")).replace(/\r\n/g, "\n").trim();
+      const content = (await readFile5(absolutePath, "utf8")).replace(/\r\n/g, "\n").trim();
       guidelineParts.push({
         content,
         isFile: true,
-        displayPath: path5.basename(absolutePath)
+        displayPath: path6.basename(absolutePath)
       });
     } catch (error) {
-      logWarning4(`Could not read guideline file ${absolutePath}: ${error.message}`);
+      logWarning5(`Could not read guideline file ${absolutePath}: ${error.message}`);
     }
   }
   const guidelines = formatFileContents(guidelineParts);
@@ -1231,9 +1723,9 @@ async function buildPromptInputs(testCase, mode = "lm") {
             messageSegments.push({ type: "text", value: segment });
           }
         } else if (isJsonObject(segment)) {
-          const type = asString4(segment.type);
+          const type = asString5(segment.type);
           if (type === "file") {
-            const value = asString4(segment.value);
+            const value = asString5(segment.value);
             if (!value) continue;
             if (testCase.guideline_patterns && isGuidelineFile(value, testCase.guideline_patterns)) {
               messageSegments.push({ type: "guideline_ref", path: value });
@@ -1244,7 +1736,7 @@ async function buildPromptInputs(testCase, mode = "lm") {
               messageSegments.push({ type: "file", text: fileText, path: value });
             }
           } else if (type === "text") {
-            const textValue = asString4(segment.value);
+            const textValue = asString5(segment.value);
             if (textValue && textValue.trim().length > 0) {
               messageSegments.push({ type: "text", value: textValue });
             }
@@ -1398,21 +1890,21 @@ ${guidelineContent.trim()}`);
   }
   return chatPrompt.length > 0 ? chatPrompt : void 0;
 }
-function asString4(value) {
+function asString5(value) {
   return typeof value === "string" ? value : void 0;
 }
-function logWarning4(message) {
-  console.warn(`${ANSI_YELLOW5}Warning: ${message}${ANSI_RESET5}`);
+function logWarning5(message) {
+  console.warn(`${ANSI_YELLOW6}Warning: ${message}${ANSI_RESET6}`);
 }
 
 // src/evaluation/yaml-parser.ts
-var ANSI_YELLOW6 = "\x1B[33m";
-var ANSI_RED = "\x1B[31m";
-var ANSI_RESET6 = "\x1B[0m";
+var ANSI_YELLOW7 = "\x1B[33m";
+var ANSI_RED2 = "\x1B[31m";
+var ANSI_RESET7 = "\x1B[0m";
 async function readTestSuiteMetadata(testFilePath) {
   try {
-    const absolutePath = path6.resolve(testFilePath);
-    const content = await readFile5(absolutePath, "utf8");
+    const absolutePath = path7.resolve(testFilePath);
+    const content = await readFile6(absolutePath, "utf8");
     const parsed = parse2(content);
     if (!isJsonObject(parsed)) {
       return {};
@@ -1423,21 +1915,25 @@ async function readTestSuiteMetadata(testFilePath) {
   }
 }
 async function loadEvalCases(evalFilePath, repoRoot, options) {
+  const format = detectFormat(evalFilePath);
+  if (format === "jsonl") {
+    return loadEvalCasesFromJsonl(evalFilePath, repoRoot, options);
+  }
   const verbose = options?.verbose ?? false;
-  const evalIdFilter = options?.evalId;
-  const absoluteTestPath = path6.resolve(evalFilePath);
+  const filterPattern = options?.filter;
+  const absoluteTestPath = path7.resolve(evalFilePath);
   const repoRootPath = resolveToAbsolutePath(repoRoot);
   const searchRoots = buildSearchRoots2(absoluteTestPath, repoRootPath);
   const config = await loadConfig(absoluteTestPath, repoRootPath);
   const guidelinePatterns = config?.guideline_patterns;
-  const rawFile = await readFile5(absoluteTestPath, "utf8");
+  const rawFile = await readFile6(absoluteTestPath, "utf8");
   const parsed = parse2(rawFile);
   if (!isJsonObject(parsed)) {
     throw new Error(`Invalid test file format: ${evalFilePath}`);
   }
   const suite = parsed;
-  const datasetNameFromSuite = asString5(suite.dataset)?.trim();
-  const fallbackDataset = path6.basename(absoluteTestPath).replace(/\.ya?ml$/i, "") || "eval";
+  const datasetNameFromSuite = asString6(suite.dataset)?.trim();
+  const fallbackDataset = path7.basename(absoluteTestPath).replace(/\.ya?ml$/i, "") || "eval";
   const datasetName = datasetNameFromSuite && datasetNameFromSuite.length > 0 ? datasetNameFromSuite : fallbackDataset;
   const rawTestcases = suite.evalcases;
   if (!Array.isArray(rawTestcases)) {
@@ -1445,37 +1941,29 @@ async function loadEvalCases(evalFilePath, repoRoot, options) {
   }
   const globalEvaluator = coerceEvaluator(suite.evaluator, "global") ?? "llm_judge";
   const globalExecution = isJsonObject(suite.execution) ? suite.execution : void 0;
-  const _globalTarget = asString5(globalExecution?.target) ?? asString5(suite.target);
+  const _globalTarget = asString6(globalExecution?.target) ?? asString6(suite.target);
   const results = [];
   for (const rawEvalcase of rawTestcases) {
     if (!isJsonObject(rawEvalcase)) {
-      logWarning5("Skipping invalid eval case entry (expected object)");
+      logWarning6("Skipping invalid eval case entry (expected object)");
       continue;
     }
     const evalcase = rawEvalcase;
-    const id = asString5(evalcase.id);
-    if (evalIdFilter && id !== evalIdFilter) {
+    const id = asString6(evalcase.id);
+    if (filterPattern && (!id || !micromatch3.isMatch(id, filterPattern))) {
       continue;
     }
-    const conversationId = asString5(evalcase.conversation_id);
-    const outcome = asString5(evalcase.expected_outcome) ?? asString5(evalcase.outcome);
-    const inputMessagesValue = evalcase.input_messages;
-    const expectedMessagesValue = evalcase.expected_messages;
-    if (!id || !outcome || !Array.isArray(inputMessagesValue)) {
-      logError(
-        `Skipping incomplete eval case: ${id ?? "unknown"}. Missing required fields: id, outcome, and/or input_messages`
+    const conversationId = asString6(evalcase.conversation_id);
+    const outcome = asString6(evalcase.expected_outcome) ?? asString6(evalcase.outcome);
+    const inputMessages = resolveInputMessages(evalcase);
+    const expectedMessages = resolveExpectedMessages(evalcase) ?? [];
+    if (!id || !outcome || !inputMessages || inputMessages.length === 0) {
+      logError2(
+        `Skipping incomplete eval case: ${id ?? "unknown"}. Missing required fields: id, outcome, and/or input_messages (or input)`
       );
       continue;
     }
-    const hasExpectedMessages = Array.isArray(expectedMessagesValue) && expectedMessagesValue.length > 0;
-    const inputMessages = inputMessagesValue.filter(
-      (msg) => isTestMessage(msg)
-    );
-    const expectedMessages = hasExpectedMessages ? expectedMessagesValue.filter((msg) => isTestMessage(msg)) : [];
-    if (hasExpectedMessages && expectedMessages.length === 0) {
-      logError(`No valid expected message found for eval case: ${id}`);
-      continue;
-    }
+    const hasExpectedMessages = expectedMessages.length > 0;
     const guidelinePaths = [];
     const inputTextParts = [];
     const inputSegments = await processMessages({
@@ -1514,33 +2002,13 @@ async function loadEvalCases(evalFilePath, repoRoot, options) {
       evaluators = await parseEvaluators(evalcase, globalExecution, searchRoots, id ?? "unknown");
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
-      logError(`Skipping eval case '${id}': ${message}`);
+      logError2(`Skipping eval case '${id}': ${message}`);
       continue;
     }
     const inlineRubrics = evalcase.rubrics;
     if (inlineRubrics !== void 0 && Array.isArray(inlineRubrics)) {
-      const rubricItems = inlineRubrics.filter((r) => isJsonObject(r) || typeof r === "string").map((rubric, index) => {
-        if (typeof rubric === "string") {
-          return {
-            id: `rubric-${index + 1}`,
-            description: rubric,
-            weight: 1,
-            required: true
-          };
-        }
-        return {
-          id: asString5(rubric.id) ?? `rubric-${index + 1}`,
-          description: asString5(rubric.description) ?? "",
-          weight: typeof rubric.weight === "number" ? rubric.weight : 1,
-          required: typeof rubric.required === "boolean" ? rubric.required : true
-        };
-      }).filter((r) => r.description.length > 0);
-      if (rubricItems.length > 0) {
-        const rubricEvaluator = {
-          name: "rubric",
-          type: "llm_judge",
-          rubrics: rubricItems
-        };
+      const rubricEvaluator = parseInlineRubrics(inlineRubrics);
+      if (rubricEvaluator) {
         evaluators = evaluators ? [rubricEvaluator, ...evaluators] : [rubricEvaluator];
       }
     }
@@ -1551,7 +2019,7 @@ async function loadEvalCases(evalFilePath, repoRoot, options) {
       }
     }
     const allFilePaths = [
-      ...guidelinePaths.map((guidelinePath) => path6.resolve(guidelinePath)),
+      ...guidelinePaths.map((guidelinePath) => path7.resolve(guidelinePath)),
       ...userFilePaths
     ];
     const testCase = {
@@ -1563,7 +2031,7 @@ async function loadEvalCases(evalFilePath, repoRoot, options) {
       input_segments: inputSegments,
       expected_messages: outputSegments,
       reference_answer: referenceAnswer,
-      guideline_paths: guidelinePaths.map((guidelinePath) => path6.resolve(guidelinePath)),
+      guideline_paths: guidelinePaths.map((guidelinePath) => path7.resolve(guidelinePath)),
       guideline_patterns: guidelinePatterns,
       file_paths: allFilePaths,
       expected_outcome: outcome,
@@ -1586,25 +2054,25 @@ async function loadEvalCases(evalFilePath, repoRoot, options) {
   }
   return results;
 }
-function asString5(value) {
+function asString6(value) {
   return typeof value === "string" ? value : void 0;
 }
-function logWarning5(message, details) {
+function logWarning6(message, details) {
   if (details && details.length > 0) {
     const detailBlock = details.join("\n");
-    console.warn(`${ANSI_YELLOW6}Warning: ${message}
-${detailBlock}${ANSI_RESET6}`);
+    console.warn(`${ANSI_YELLOW7}Warning: ${message}
+${detailBlock}${ANSI_RESET7}`);
   } else {
-    console.warn(`${ANSI_YELLOW6}Warning: ${message}${ANSI_RESET6}`);
+    console.warn(`${ANSI_YELLOW7}Warning: ${message}${ANSI_RESET7}`);
   }
 }
-function logError(message, details) {
+function logError2(message, details) {
   if (details && details.length > 0) {
     const detailBlock = details.join("\n");
-    console.error(`${ANSI_RED}Error: ${message}
-${detailBlock}${ANSI_RESET6}`);
+    console.error(`${ANSI_RED2}Error: ${message}
+${detailBlock}${ANSI_RESET7}`);
   } else {
-    console.error(`${ANSI_RED}Error: ${message}${ANSI_RESET6}`);
+    console.error(`${ANSI_RED2}Error: ${message}${ANSI_RESET7}`);
   }
 }
 
@@ -1947,7 +2415,7 @@ import { randomUUID } from "node:crypto";
 import { createWriteStream } from "node:fs";
 import { mkdir, mkdtemp, rm, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
-import path8 from "node:path";
+import path9 from "node:path";
 
 // src/evaluation/providers/claude-code-log-tracker.ts
 var GLOBAL_LOGS_KEY = Symbol.for("agentv.claudeCodeLogs");
@@ -2003,7 +2471,7 @@ function subscribeToClaudeCodeLogEntries(listener) {
 }
 
 // src/evaluation/providers/preread.ts
-import path7 from "node:path";
+import path8 from "node:path";
 function buildPromptDocument(request, inputFiles, options) {
   const parts = [];
   const guidelineFiles = collectGuidelineFiles(
@@ -2026,7 +2494,7 @@ function normalizeInputFiles(inputFiles) {
   }
   const deduped = /* @__PURE__ */ new Map();
   for (const inputFile of inputFiles) {
-    const absolutePath = path7.resolve(inputFile);
+    const absolutePath = path8.resolve(inputFile);
     if (!deduped.has(absolutePath)) {
       deduped.set(absolutePath, absolutePath);
     }
@@ -2039,14 +2507,14 @@ function collectGuidelineFiles(inputFiles, guidelinePatterns, overrides) {
   }
   const unique = /* @__PURE__ */ new Map();
   for (const inputFile of inputFiles) {
-    const absolutePath = path7.resolve(inputFile);
+    const absolutePath = path8.resolve(inputFile);
     if (overrides?.has(absolutePath)) {
       if (!unique.has(absolutePath)) {
         unique.set(absolutePath, absolutePath);
       }
       continue;
     }
-    const normalized = absolutePath.split(path7.sep).join("/");
+    const normalized = absolutePath.split(path8.sep).join("/");
     if (isGuidelineFile(normalized, guidelinePatterns)) {
       if (!unique.has(absolutePath)) {
         unique.set(absolutePath, absolutePath);
@@ -2061,7 +2529,7 @@ function collectInputFiles(inputFiles) {
   }
   const unique = /* @__PURE__ */ new Map();
   for (const inputFile of inputFiles) {
-    const absolutePath = path7.resolve(inputFile);
+    const absolutePath = path8.resolve(inputFile);
     if (!unique.has(absolutePath)) {
       unique.set(absolutePath, absolutePath);
     }
@@ -2073,7 +2541,7 @@ function buildMandatoryPrereadBlock(guidelineFiles, inputFiles) {
     return "";
   }
   const buildList = (files) => files.map((absolutePath) => {
-    const fileName = path7.basename(absolutePath);
+    const fileName = path8.basename(absolutePath);
     const fileUri = pathToFileUri(absolutePath);
     return `* [${fileName}](${fileUri})`;
   });
@@ -2093,7 +2561,7 @@ ${buildList(inputFiles).join("\n")}.`);
   return sections.join("\n");
 }
 function pathToFileUri(filePath) {
-  const absolutePath = path7.isAbsolute(filePath) ? filePath : path7.resolve(filePath);
+  const absolutePath = path8.isAbsolute(filePath) ? filePath : path8.resolve(filePath);
   const normalizedPath = absolutePath.replace(/\\/g, "/");
   if (/^[a-zA-Z]:\//.test(normalizedPath)) {
     return `file:///${normalizedPath}`;
@@ -2130,7 +2598,7 @@ var ClaudeCodeProvider = class {
     const workspaceRoot = await this.createWorkspace();
     const logger = await this.createStreamLogger(request).catch(() => void 0);
     try {
-      const promptFile = path8.join(workspaceRoot, PROMPT_FILENAME);
+      const promptFile = path9.join(workspaceRoot, PROMPT_FILENAME);
       await writeFile(promptFile, request.question, "utf8");
       const args = this.buildClaudeCodeArgs(request.question, inputFiles);
       const cwd = this.resolveCwd();
@@ -2178,7 +2646,7 @@ var ClaudeCodeProvider = class {
     if (!this.config.cwd) {
       return process.cwd();
     }
-    return path8.resolve(this.config.cwd);
+    return path9.resolve(this.config.cwd);
   }
   buildClaudeCodeArgs(prompt, inputFiles) {
     const args = [];
@@ -2235,7 +2703,7 @@ ${filesContext}`;
     }
   }
   async createWorkspace() {
-    return await mkdtemp(path8.join(tmpdir(), WORKSPACE_PREFIX));
+    return await mkdtemp(path9.join(tmpdir(), WORKSPACE_PREFIX));
   }
   async cleanupWorkspace(workspaceRoot) {
     try {
@@ -2249,9 +2717,9 @@ ${filesContext}`;
       return void 0;
     }
     if (this.config.logDir) {
-      return path8.resolve(this.config.logDir);
+      return path9.resolve(this.config.logDir);
     }
-    return path8.join(process.cwd(), ".agentv", "logs", "claude-code");
+    return path9.join(process.cwd(), ".agentv", "logs", "claude-code");
   }
   async createStreamLogger(request) {
     const logDir = this.resolveLogDirectory();
@@ -2265,7 +2733,7 @@ ${filesContext}`;
       console.warn(`Skipping Claude Code stream logging (could not create ${logDir}): ${message}`);
       return void 0;
     }
-    const filePath = path8.join(logDir, buildLogFilename(request, this.targetName));
+    const filePath = path9.join(logDir, buildLogFilename(request, this.targetName));
     try {
       const logger = await ClaudeCodeStreamLogger.create({
         filePath,
@@ -2670,10 +3138,10 @@ function escapeShellArg(arg) {
 }
 async function defaultClaudeCodeRunner(options) {
   const tempId = randomUUID();
-  const stdoutFile = path8.join(tmpdir(), `agentv-cc-${tempId}-stdout`);
-  const stderrFile = path8.join(tmpdir(), `agentv-cc-${tempId}-stderr`);
-  const exitFile = path8.join(tmpdir(), `agentv-cc-${tempId}-exit`);
-  const pidFile = path8.join(tmpdir(), `agentv-cc-${tempId}-pid`);
+  const stdoutFile = path9.join(tmpdir(), `agentv-cc-${tempId}-stdout`);
+  const stderrFile = path9.join(tmpdir(), `agentv-cc-${tempId}-stderr`);
+  const exitFile = path9.join(tmpdir(), `agentv-cc-${tempId}-exit`);
+  const pidFile = path9.join(tmpdir(), `agentv-cc-${tempId}-pid`);
   try {
     return await runClaudeCodeWithTempFiles(options, stdoutFile, stderrFile, exitFile, pidFile);
   } finally {
@@ -2713,8 +3181,8 @@ async function runClaudeCodeWithTempFiles(options, stdoutFile, stderrFile, exitF
   let lastStdoutSize = 0;
   const readFileIfExists = async (filePath) => {
     try {
-      const { readFile: readFile7 } = await import("node:fs/promises");
-      return await readFile7(filePath, "utf8");
+      const { readFile: readFile8 } = await import("node:fs/promises");
+      return await readFile8(filePath, "utf8");
     } catch {
       return "";
     }
@@ -2789,7 +3257,7 @@ async function runClaudeCodeWithTempFiles(options, stdoutFile, stderrFile, exitF
 import { exec as execWithCallback } from "node:child_process";
 import fs from "node:fs/promises";
 import os from "node:os";
-import path9 from "node:path";
+import path10 from "node:path";
 import { promisify } from "node:util";
 import { z } from "zod";
 var ToolCallSchema = z.object({
@@ -2797,7 +3265,8 @@ var ToolCallSchema = z.object({
   input: z.unknown().optional(),
   output: z.unknown().optional(),
   id: z.string().optional(),
-  timestamp: z.string().optional()
+  timestamp: z.string().optional(),
+  duration_ms: z.number().optional()
 });
 var OutputMessageInputSchema = z.object({
   role: z.string(),
@@ -2805,6 +3274,7 @@ var OutputMessageInputSchema = z.object({
   content: z.unknown().optional(),
   tool_calls: z.array(ToolCallSchema).optional(),
   timestamp: z.string().optional(),
+  duration_ms: z.number().optional(),
   metadata: z.record(z.unknown()).optional()
 });
 var TokenUsageSchema = z.object({
@@ -2843,8 +3313,16 @@ function convertOutputMessages(messages) {
     role: msg.role,
     name: msg.name,
     content: msg.content,
-    toolCalls: msg.tool_calls,
+    toolCalls: msg.tool_calls?.map((tc) => ({
+      tool: tc.tool,
+      input: tc.input,
+      output: tc.output,
+      id: tc.id,
+      timestamp: tc.timestamp,
+      durationMs: tc.duration_ms
+    })),
     timestamp: msg.timestamp,
+    durationMs: msg.duration_ms,
     metadata: msg.metadata
   }));
 }
@@ -3246,7 +3724,7 @@ function normalizeInputFiles2(inputFiles) {
   }
   const unique = /* @__PURE__ */ new Map();
   for (const inputFile of inputFiles) {
-    const absolutePath = path9.resolve(inputFile);
+    const absolutePath = path10.resolve(inputFile);
     if (!unique.has(absolutePath)) {
       unique.set(absolutePath, absolutePath);
     }
@@ -3260,7 +3738,7 @@ function formatFileList(files, template) {
   const formatter = template ?? "{path}";
   return files.map((filePath) => {
     const escapedPath = shellEscape(filePath);
-    const escapedName = shellEscape(path9.basename(filePath));
+    const escapedName = shellEscape(path10.basename(filePath));
     return formatter.replaceAll("{path}", escapedPath).replaceAll("{basename}", escapedName);
   }).join(" ");
 }
@@ -3284,7 +3762,7 @@ function generateOutputFilePath(evalCaseId, extension = ".json") {
   const safeEvalId = evalCaseId || "unknown";
   const timestamp = Date.now();
   const random = Math.random().toString(36).substring(2, 9);
-  return path9.join(os.tmpdir(), `agentv-${safeEvalId}-${timestamp}-${random}${extension}`);
+  return path10.join(os.tmpdir(), `agentv-${safeEvalId}-${timestamp}-${random}${extension}`);
 }
 function formatTimeoutSuffix2(timeoutMs) {
   if (!timeoutMs || timeoutMs <= 0) {
@@ -3300,7 +3778,7 @@ import { randomUUID as randomUUID2 } from "node:crypto";
 import { constants as constants2, createWriteStream as createWriteStream2 } from "node:fs";
 import { access as access2, mkdir as mkdir2, mkdtemp as mkdtemp2, rm as rm2, writeFile as writeFile2 } from "node:fs/promises";
 import { tmpdir as tmpdir2 } from "node:os";
-import path10 from "node:path";
+import path11 from "node:path";
 import { promisify as promisify2 } from "node:util";
 
 // src/evaluation/providers/codex-log-tracker.ts
@@ -3395,7 +3873,7 @@ var CodexProvider = class {
       const promptContent = `${systemPrompt}
 
 ${basePrompt}`;
-      const promptFile = path10.join(workspaceRoot, PROMPT_FILENAME2);
+      const promptFile = path11.join(workspaceRoot, PROMPT_FILENAME2);
       await writeFile2(promptFile, promptContent, "utf8");
       const args = this.buildCodexArgs();
       const cwd = this.resolveCwd(workspaceRoot);
@@ -3445,7 +3923,7 @@ ${basePrompt}`;
     if (!this.config.cwd) {
       return workspaceRoot;
     }
-    return path10.resolve(this.config.cwd);
+    return path11.resolve(this.config.cwd);
   }
   buildCodexArgs() {
     const args = [
@@ -3487,7 +3965,7 @@ ${basePrompt}`;
     }
   }
   async createWorkspace() {
-    return await mkdtemp2(path10.join(tmpdir2(), WORKSPACE_PREFIX2));
+    return await mkdtemp2(path11.join(tmpdir2(), WORKSPACE_PREFIX2));
   }
   async cleanupWorkspace(workspaceRoot) {
     try {
@@ -3501,9 +3979,9 @@ ${basePrompt}`;
       return void 0;
     }
     if (this.config.logDir) {
-      return path10.resolve(this.config.logDir);
+      return path11.resolve(this.config.logDir);
     }
-    return path10.join(process.cwd(), ".agentv", "logs", "codex");
+    return path11.join(process.cwd(), ".agentv", "logs", "codex");
   }
   async createStreamLogger(request) {
     const logDir = this.resolveLogDirectory();
@@ -3517,7 +3995,7 @@ ${basePrompt}`;
       console.warn(`Skipping Codex stream logging (could not create ${logDir}): ${message}`);
       return void 0;
     }
-    const filePath = path10.join(logDir, buildLogFilename2(request, this.targetName));
+    const filePath = path11.join(logDir, buildLogFilename2(request, this.targetName));
     try {
       const logger = await CodexStreamLogger.create({
         filePath,
@@ -3732,7 +4210,7 @@ function tryParseJsonValue2(rawLine) {
 async function locateExecutable(candidate) {
   const includesPathSeparator = candidate.includes("/") || candidate.includes("\\");
   if (includesPathSeparator) {
-    const resolved = path10.isAbsolute(candidate) ? candidate : path10.resolve(candidate);
+    const resolved = path11.isAbsolute(candidate) ? candidate : path11.resolve(candidate);
     const executablePath = await ensureWindowsExecutableVariant(resolved);
     await access2(executablePath, constants2.F_OK);
     return executablePath;
@@ -4245,7 +4723,7 @@ import { randomUUID as randomUUID3 } from "node:crypto";
 import { createWriteStream as createWriteStream3 } from "node:fs";
 import { mkdir as mkdir3, mkdtemp as mkdtemp3, rm as rm3, writeFile as writeFile3 } from "node:fs/promises";
 import { tmpdir as tmpdir3 } from "node:os";
-import path11 from "node:path";
+import path12 from "node:path";
 
 // src/evaluation/providers/pi-log-tracker.ts
 var GLOBAL_LOGS_KEY3 = Symbol.for("agentv.piLogs");
@@ -4329,7 +4807,7 @@ var PiCodingAgentProvider = class {
     const workspaceRoot = await this.createWorkspace();
     const logger = await this.createStreamLogger(request).catch(() => void 0);
     try {
-      const promptFile = path11.join(workspaceRoot, PROMPT_FILENAME3);
+      const promptFile = path12.join(workspaceRoot, PROMPT_FILENAME3);
       await writeFile3(promptFile, request.question, "utf8");
       const args = this.buildPiArgs(request.question, inputFiles);
       const cwd = this.resolveCwd(workspaceRoot);
@@ -4371,7 +4849,7 @@ var PiCodingAgentProvider = class {
     if (!this.config.cwd) {
       return workspaceRoot;
     }
-    return path11.resolve(this.config.cwd);
+    return path12.resolve(this.config.cwd);
   }
   buildPiArgs(prompt, inputFiles) {
     const args = [];
@@ -4460,7 +4938,7 @@ ${prompt}`;
     return env;
   }
   async createWorkspace() {
-    return await mkdtemp3(path11.join(tmpdir3(), WORKSPACE_PREFIX3));
+    return await mkdtemp3(path12.join(tmpdir3(), WORKSPACE_PREFIX3));
   }
   async cleanupWorkspace(workspaceRoot) {
     try {
@@ -4470,9 +4948,9 @@ ${prompt}`;
   }
   resolveLogDirectory() {
     if (this.config.logDir) {
-      return path11.resolve(this.config.logDir);
+      return path12.resolve(this.config.logDir);
     }
-    return path11.join(process.cwd(), ".agentv", "logs", "pi-coding-agent");
+    return path12.join(process.cwd(), ".agentv", "logs", "pi-coding-agent");
   }
   async createStreamLogger(request) {
     const logDir = this.resolveLogDirectory();
@@ -4486,7 +4964,7 @@ ${prompt}`;
       console.warn(`Skipping Pi stream logging (could not create ${logDir}): ${message}`);
       return void 0;
     }
-    const filePath = path11.join(logDir, buildLogFilename3(request, this.targetName));
+    const filePath = path12.join(logDir, buildLogFilename3(request, this.targetName));
     try {
       const logger = await PiStreamLogger.create({
         filePath,
@@ -4919,7 +5397,7 @@ async function defaultPiRunner(options) {
 }
 
 // src/evaluation/providers/vscode.ts
-import path12 from "node:path";
+import path13 from "node:path";
 import {
   dispatchAgentSession,
   dispatchBatchAgent,
@@ -5094,7 +5572,7 @@ function buildMandatoryPrereadBlock2(guidelineFiles, attachmentFiles) {
     return "";
   }
   const buildList = (files) => files.map((absolutePath) => {
-    const fileName = path12.basename(absolutePath);
+    const fileName = path13.basename(absolutePath);
     const fileUri = pathToFileUri2(absolutePath);
     return `* [${fileName}](${fileUri})`;
   });
@@ -5119,8 +5597,8 @@ function collectGuidelineFiles2(attachments, guidelinePatterns) {
   }
   const unique = /* @__PURE__ */ new Map();
   for (const attachment of attachments) {
-    const absolutePath = path12.resolve(attachment);
-    const normalized = absolutePath.split(path12.sep).join("/");
+    const absolutePath = path13.resolve(attachment);
+    const normalized = absolutePath.split(path13.sep).join("/");
     if (isGuidelineFile(normalized, guidelinePatterns)) {
       if (!unique.has(absolutePath)) {
         unique.set(absolutePath, absolutePath);
@@ -5135,7 +5613,7 @@ function collectAttachmentFiles(attachments) {
   }
   const unique = /* @__PURE__ */ new Map();
   for (const attachment of attachments) {
-    const absolutePath = path12.resolve(attachment);
+    const absolutePath = path13.resolve(attachment);
     if (!unique.has(absolutePath)) {
       unique.set(absolutePath, absolutePath);
     }
@@ -5143,7 +5621,7 @@ function collectAttachmentFiles(attachments) {
   return Array.from(unique.values());
 }
 function pathToFileUri2(filePath) {
-  const absolutePath = path12.isAbsolute(filePath) ? filePath : path12.resolve(filePath);
+  const absolutePath = path13.isAbsolute(filePath) ? filePath : path13.resolve(filePath);
   const normalizedPath = absolutePath.replace(/\\/g, "/");
   if (/^[a-zA-Z]:\//.test(normalizedPath)) {
     return `file:///${normalizedPath}`;
@@ -5156,7 +5634,7 @@ function normalizeAttachments(attachments) {
   }
   const deduped = /* @__PURE__ */ new Set();
   for (const attachment of attachments) {
-    deduped.add(path12.resolve(attachment));
+    deduped.add(path13.resolve(attachment));
   }
   return Array.from(deduped);
 }
@@ -5165,7 +5643,7 @@ function mergeAttachments(all) {
   for (const list of all) {
     if (!list) continue;
     for (const inputFile of list) {
-      deduped.add(path12.resolve(inputFile));
+      deduped.add(path13.resolve(inputFile));
     }
   }
   return deduped.size > 0 ? Array.from(deduped) : void 0;
@@ -5213,8 +5691,8 @@ total unlocked subagents available: ${result.created.length + result.skippedExis
 
 // src/evaluation/providers/targets-file.ts
 import { constants as constants3 } from "node:fs";
-import { access as access3, readFile as readFile6 } from "node:fs/promises";
-import path13 from "node:path";
+import { access as access3, readFile as readFile7 } from "node:fs/promises";
+import path14 from "node:path";
 import { parse as parse3 } from "yaml";
 function isRecord(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
@@ -5251,11 +5729,11 @@ async function fileExists3(filePath) {
   }
 }
 async function readTargetDefinitions(filePath) {
-  const absolutePath = path13.resolve(filePath);
+  const absolutePath = path14.resolve(filePath);
   if (!await fileExists3(absolutePath)) {
     throw new Error(`targets.yaml not found at ${absolutePath}`);
   }
-  const raw = await readFile6(absolutePath, "utf8");
+  const raw = await readFile7(absolutePath, "utf8");
   const parsed = parse3(raw);
   if (!isRecord(parsed)) {
     throw new Error(`targets.yaml at ${absolutePath} must be a YAML object with a 'targets' field`);
@@ -5462,15 +5940,15 @@ async function execFileWithStdinNode(argv, stdinPayload, options) {
   });
 }
 async function execShellWithStdin(command, stdinPayload, options = {}) {
-  const { mkdir: mkdir4, readFile: readFile7, rm: rm4, writeFile: writeFile4 } = await import("node:fs/promises");
+  const { mkdir: mkdir4, readFile: readFile8, rm: rm4, writeFile: writeFile4 } = await import("node:fs/promises");
   const { tmpdir: tmpdir4 } = await import("node:os");
-  const path15 = await import("node:path");
+  const path16 = await import("node:path");
   const { randomUUID: randomUUID4 } = await import("node:crypto");
-  const dir = path15.join(tmpdir4(), `agentv-exec-${randomUUID4()}`);
+  const dir = path16.join(tmpdir4(), `agentv-exec-${randomUUID4()}`);
   await mkdir4(dir, { recursive: true });
-  const stdinPath = path15.join(dir, "stdin.txt");
-  const stdoutPath = path15.join(dir, "stdout.txt");
-  const stderrPath = path15.join(dir, "stderr.txt");
+  const stdinPath = path16.join(dir, "stdin.txt");
+  const stdoutPath = path16.join(dir, "stdout.txt");
+  const stderrPath = path16.join(dir, "stderr.txt");
   await writeFile4(stdinPath, stdinPayload, "utf8");
   const wrappedCommand = process.platform === "win32" ? `(${command}) < ${shellEscapePath(stdinPath)} > ${shellEscapePath(stdoutPath)} 2> ${shellEscapePath(stderrPath)}` : `(${command}) < ${shellEscapePath(stdinPath)} > ${shellEscapePath(stdoutPath)} 2> ${shellEscapePath(stderrPath)}`;
   const { spawn: spawn4 } = await import("node:child_process");
@@ -5500,8 +5978,8 @@ async function execShellWithStdin(command, stdinPayload, options = {}) {
         resolve(code ?? 0);
       });
     });
-    const stdout = (await readFile7(stdoutPath, "utf8")).replace(/\r\n/g, "\n");
-    const stderr = (await readFile7(stderrPath, "utf8")).replace(/\r\n/g, "\n");
+    const stdout = (await readFile8(stdoutPath, "utf8")).replace(/\r\n/g, "\n");
+    const stderr = (await readFile8(stderrPath, "utf8")).replace(/\r\n/g, "\n");
     return { stdout, stderr, exitCode };
   } finally {
     await rm4(dir, { recursive: true, force: true });
@@ -5773,7 +6251,7 @@ var CodeEvaluator = class {
       outputMessages: context.outputMessages ?? null,
       guidelineFiles: context.evalCase.guideline_paths,
       inputFiles: context.evalCase.file_paths.filter(
-        (path15) => !context.evalCase.guideline_paths.includes(path15)
+        (path16) => !context.evalCase.guideline_paths.includes(path16)
       ),
       inputMessages: context.evalCase.input_messages,
       traceSummary: context.traceSummary ?? null,
@@ -5921,6 +6399,15 @@ var rubricEvaluationSchema = z2.object({
   checks: z2.array(rubricCheckResultSchema).describe("Results for each rubric item"),
   overall_reasoning: z2.string().describe("Overall assessment summary (1-2 sentences)")
 });
+var scoreRangeCheckResultSchema = z2.object({
+  id: z2.string().describe("The ID of the rubric criterion being scored"),
+  score: z2.number().int().min(0).max(10).describe("Integer score 0-10 for this criterion"),
+  reasoning: z2.string().describe("Brief explanation (1-2 sentences) for this score").optional()
+});
+var scoreRangeEvaluationSchema = z2.object({
+  checks: z2.array(scoreRangeCheckResultSchema).describe("Scores for each rubric criterion"),
+  overall_reasoning: z2.string().describe("Overall assessment summary (1-2 sentences)").optional()
+});
 var LlmJudgeEvaluator = class {
   kind = "llm_judge";
   resolveJudgeProvider;
@@ -6006,6 +6493,10 @@ var LlmJudgeEvaluator = class {
         `No rubrics found for evaluator "${context.evaluator?.name ?? "llm_judge"}". Run "agentv generate rubrics" first.`
       );
     }
+    const hasScoreRanges = rubrics.some((r) => r.score_ranges && r.score_ranges.length > 0);
+    if (hasScoreRanges) {
+      return this.evaluateWithScoreRanges(context, judgeProvider, rubrics);
+    }
     const prompt = this.buildRubricPrompt(context, rubrics);
     const systemPrompt = buildRubricOutputSchema();
     const evaluatorRawRequest = {
@@ -6031,6 +6522,84 @@ var LlmJudgeEvaluator = class {
       evaluatorRawRequest
     };
   }
+  /**
+   * Evaluate using score-range rubrics (analytic rubric scoring).
+   * Each criterion is scored 0-10 and normalized to 0-1.
+   */
+  async evaluateWithScoreRanges(context, judgeProvider, rubrics) {
+    const prompt = this.buildScoreRangePrompt(context, rubrics);
+    const systemPrompt = buildScoreRangeOutputSchema();
+    const evaluatorRawRequest = {
+      userPrompt: prompt,
+      systemPrompt,
+      target: judgeProvider.targetName
+    };
+    const { data } = await this.runWithRetry({
+      context,
+      judgeProvider,
+      systemPrompt,
+      userPrompt: prompt,
+      schema: scoreRangeEvaluationSchema
+    });
+    const { score, verdict, hits, misses, details } = calculateScoreRangeResult(data, rubrics);
+    return {
+      score,
+      verdict,
+      hits,
+      misses,
+      expectedAspectCount: rubrics.length,
+      reasoning: data.overall_reasoning,
+      evaluatorRawRequest,
+      details
+    };
+  }
+  /**
+   * Build prompt for score-range rubric evaluation.
+   */
+  buildScoreRangePrompt(context, rubrics) {
+    const formattedQuestion = context.promptInputs.question && context.promptInputs.question.trim().length > 0 ? context.promptInputs.question : context.evalCase.question;
+    const parts = [
+      "You are an expert evaluator. Score the candidate answer on each criterion below using the provided score ranges.",
+      "For each criterion, output an integer score from 0 to 10 based on which score range best matches the answer.",
+      "",
+      "[[ ## question ## ]]",
+      formattedQuestion,
+      "",
+      "[[ ## expected_outcome ## ]]",
+      context.evalCase.expected_outcome,
+      ""
+    ];
+    if (context.evalCase.reference_answer && context.evalCase.reference_answer.trim().length > 0) {
+      parts.push("[[ ## reference_answer ## ]]", context.evalCase.reference_answer, "");
+    }
+    parts.push(
+      "[[ ## candidate_answer ## ]]",
+      context.candidate,
+      "",
+      "[[ ## scoring_criteria ## ]]"
+    );
+    for (const rubric of rubrics) {
+      const weightLabel = rubric.weight !== 1 ? ` (weight: ${rubric.weight})` : "";
+      const minScoreLabel = rubric.required_min_score !== void 0 ? ` [REQUIRED: min score ${rubric.required_min_score}]` : "";
+      parts.push("", `### Criterion: ${rubric.id}${weightLabel}${minScoreLabel}`);
+      if (rubric.expected_outcome) {
+        parts.push(`Description: ${rubric.expected_outcome}`);
+      }
+      if (rubric.score_ranges && rubric.score_ranges.length > 0) {
+        parts.push("Score ranges:");
+        for (const range of rubric.score_ranges) {
+          const [min, max] = range.score_range;
+          const rangeLabel = min === max ? `${min}` : `${min}-${max}`;
+          parts.push(`  - Score ${rangeLabel}: ${range.expected_outcome}`);
+        }
+      }
+    }
+    parts.push(
+      "",
+      "For each criterion, provide an integer score 0-10 that matches one of its defined score ranges."
+    );
+    return parts.join("\n");
+  }
   buildRubricPrompt(context, rubrics) {
     const formattedQuestion = context.promptInputs.question && context.promptInputs.question.trim().length > 0 ? context.promptInputs.question : context.evalCase.question;
     const parts = [
@@ -6050,7 +6619,7 @@ var LlmJudgeEvaluator = class {
     for (const rubric of rubrics) {
       const requiredLabel = rubric.required ? " (REQUIRED)" : "";
       const weightLabel = rubric.weight !== 1 ? ` (weight: ${rubric.weight})` : "";
-      parts.push(`- [${rubric.id}]${requiredLabel}${weightLabel}: ${rubric.description}`);
+      parts.push(`- [${rubric.id}]${requiredLabel}${weightLabel}: ${rubric.expected_outcome}`);
     }
     parts.push("", "For each rubric, determine if it is satisfied and provide brief reasoning.");
     return parts.join("\n");
@@ -6137,9 +6706,9 @@ function calculateRubricScore(result, rubrics) {
     totalWeight += rubric.weight;
     if (check.satisfied) {
       earnedWeight += rubric.weight;
-      hits.push(`[${rubric.id}] ${rubric.description}: ${check.reasoning}`);
+      hits.push(`[${rubric.id}] ${rubric.expected_outcome}: ${check.reasoning}`);
     } else {
-      misses.push(`[${rubric.id}] ${rubric.description}: ${check.reasoning}`);
+      misses.push(`[${rubric.id}] ${rubric.expected_outcome}: ${check.reasoning}`);
       if (rubric.required) {
         failedRequired = true;
       }
@@ -6149,6 +6718,76 @@ function calculateRubricScore(result, rubrics) {
   const verdict = failedRequired ? "fail" : scoreToVerdict(score);
   return { score, verdict, hits, misses };
 }
+function buildScoreRangeOutputSchema() {
+  return `You are an expert evaluator. Score the candidate answer on each criterion.
+You must return a valid JSON object matching this schema:
+{
+  "checks": [
+    {
+      "id": "string (criterion id)",
+      "score": integer (0-10),
+      "reasoning": "string (brief explanation for score)"
+    }
+  ],
+  "overall_reasoning": "string (summary, optional)"
+}
+
+Important: The "score" must be an integer from 0 to 10 that falls within one of the defined score ranges for that criterion.`;
+}
+function calculateScoreRangeResult(result, rubrics) {
+  const rubricMap = new Map(rubrics.map((rubric) => [rubric.id, rubric]));
+  const hits = [];
+  const misses = [];
+  const rawScores = {};
+  let totalWeight = 0;
+  let weightedScoreSum = 0;
+  let failedRequired = false;
+  for (const check of result.checks) {
+    const rubric = rubricMap.get(check.id);
+    if (!rubric) {
+      continue;
+    }
+    const rawScore = Math.max(0, Math.min(10, check.score));
+    const normalizedScore = rawScore / 10;
+    rawScores[rubric.id] = rawScore;
+    totalWeight += rubric.weight;
+    weightedScoreSum += normalizedScore * rubric.weight;
+    let requiredMinScore;
+    if (rubric.required_min_score !== void 0) {
+      requiredMinScore = rubric.required_min_score;
+    } else if (rubric.required === true) {
+      requiredMinScore = 10;
+    }
+    const matchingRange = rubric.score_ranges?.find(
+      (r) => rawScore >= r.score_range[0] && rawScore <= r.score_range[1]
+    );
+    const rangeDescription = matchingRange?.expected_outcome ?? "";
+    const criterionLabel = rubric.expected_outcome ?? rubric.id;
+    const reasoningText = check.reasoning ? `: ${check.reasoning}` : "";
+    const scoreInfo = `[${rubric.id}] ${criterionLabel} - Score: ${rawScore}/10 (${rangeDescription})${reasoningText}`;
+    if (requiredMinScore !== void 0 && rawScore < requiredMinScore) {
+      failedRequired = true;
+      misses.push(scoreInfo);
+    } else if (rawScore >= 7) {
+      hits.push(scoreInfo);
+    } else {
+      misses.push(scoreInfo);
+    }
+  }
+  const score = totalWeight > 0 ? Math.min(1, Math.max(0, weightedScoreSum / totalWeight)) : 0;
+  const verdict = failedRequired ? "fail" : scoreToVerdict(score);
+  return {
+    score,
+    verdict,
+    hits,
+    misses,
+    details: {
+      raw_scores: rawScores,
+      normalization: "score / 10",
+      aggregation: "weighted_average"
+    }
+  };
+}
 
 // src/evaluation/evaluators/composite.ts
 var DEFAULT_COMPOSITE_AGGREGATOR_PROMPT = `Review the following evaluation results:
@@ -6532,115 +7171,115 @@ var FieldAccuracyEvaluator = class {
    * Evaluate a single field against the expected value.
    */
   evaluateField(fieldConfig, candidateData, expectedData) {
-    const { path: path15, match, required = true, weight = 1 } = fieldConfig;
-    const candidateValue = resolvePath(candidateData, path15);
-    const expectedValue = resolvePath(expectedData, path15);
+    const { path: path16, match, required = true, weight = 1 } = fieldConfig;
+    const candidateValue = resolvePath(candidateData, path16);
+    const expectedValue = resolvePath(expectedData, path16);
     if (expectedValue === void 0) {
       return {
-        path: path15,
+        path: path16,
         score: 1,
         // No expected value means no comparison needed
         weight,
         hit: true,
-        message: `${path15}: no expected value`
+        message: `${path16}: no expected value`
       };
     }
     if (candidateValue === void 0) {
       if (required) {
         return {
-          path: path15,
+          path: path16,
           score: 0,
           weight,
           hit: false,
-          message: `${path15} (required, missing)`
+          message: `${path16} (required, missing)`
         };
       }
       return {
-        path: path15,
+        path: path16,
         score: 1,
         // Don't penalize missing optional fields
         weight: 0,
         // Zero weight means it won't affect the score
         hit: true,
-        message: `${path15}: optional field missing`
+        message: `${path16}: optional field missing`
       };
     }
     switch (match) {
       case "exact":
-        return this.compareExact(path15, candidateValue, expectedValue, weight);
+        return this.compareExact(path16, candidateValue, expectedValue, weight);
       case "numeric_tolerance":
         return this.compareNumericTolerance(
-          path15,
+          path16,
           candidateValue,
           expectedValue,
           fieldConfig,
           weight
         );
       case "date":
-        return this.compareDate(path15, candidateValue, expectedValue, fieldConfig, weight);
+        return this.compareDate(path16, candidateValue, expectedValue, fieldConfig, weight);
       default:
         return {
-          path: path15,
+          path: path16,
           score: 0,
           weight,
           hit: false,
-          message: `${path15}: unknown match type "${match}"`
+          message: `${path16}: unknown match type "${match}"`
         };
     }
   }
   /**
    * Exact equality comparison.
    */
-  compareExact(path15, candidateValue, expectedValue, weight) {
+  compareExact(path16, candidateValue, expectedValue, weight) {
     if (deepEqual(candidateValue, expectedValue)) {
       return {
-        path: path15,
+        path: path16,
         score: 1,
         weight,
         hit: true,
-        message: path15
+        message: path16
       };
     }
     if (typeof candidateValue !== typeof expectedValue) {
       return {
-        path: path15,
+        path: path16,
         score: 0,
         weight,
         hit: false,
-        message: `${path15} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
+        message: `${path16} (type mismatch: got ${typeof candidateValue}, expected ${typeof expectedValue})`
       };
     }
     return {
-      path: path15,
+      path: path16,
       score: 0,
       weight,
       hit: false,
-      message: `${path15} (value mismatch)`
+      message: `${path16} (value mismatch)`
     };
   }
   /**
    * Numeric comparison with absolute or relative tolerance.
    */
-  compareNumericTolerance(path15, candidateValue, expectedValue, fieldConfig, weight) {
+  compareNumericTolerance(path16, candidateValue, expectedValue, fieldConfig, weight) {
     const { tolerance = 0, relative = false } = fieldConfig;
     const candidateNum = toNumber(candidateValue);
     const expectedNum = toNumber(expectedValue);
     if (candidateNum === null || expectedNum === null) {
       return {
-        path: path15,
+        path: path16,
         score: 0,
         weight,
         hit: false,
-        message: `${path15} (non-numeric value)`
+        message: `${path16} (non-numeric value)`
       };
     }
     if (!Number.isFinite(candidateNum) || !Number.isFinite(expectedNum)) {
       return {
-        path: path15,
+        path: path16,
         score: 0,
         weight,
         hit: false,
-        message: `${path15} (invalid numeric value)`
+        message: `${path16} (invalid numeric value)`
       };
     }
     const diff = Math.abs(candidateNum - expectedNum);
@@ -6653,61 +7292,61 @@ var FieldAccuracyEvaluator = class {
     }
     if (withinTolerance) {
       return {
-        path: path15,
+        path: path16,
         score: 1,
         weight,
         hit: true,
-        message: `${path15} (within tolerance: diff=${diff.toFixed(2)})`
+        message: `${path16} (within tolerance: diff=${diff.toFixed(2)})`
       };
     }
     return {
-      path: path15,
+      path: path16,
       score: 0,
       weight,
       hit: false,
-      message: `${path15} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
+      message: `${path16} (outside tolerance: diff=${diff.toFixed(2)}, tolerance=${tolerance})`
     };
   }
   /**
    * Date comparison with format normalization.
    */
-  compareDate(path15, candidateValue, expectedValue, fieldConfig, weight) {
+  compareDate(path16, candidateValue, expectedValue, fieldConfig, weight) {
     const formats = fieldConfig.formats ?? DEFAULT_DATE_FORMATS;
     const candidateDate = parseDate(String(candidateValue), formats);
     const expectedDate = parseDate(String(expectedValue), formats);
     if (candidateDate === null) {
       return {
-        path: path15,
+        path: path16,
         score: 0,
         weight,
         hit: false,
-        message: `${path15} (unparseable candidate date)`
+        message: `${path16} (unparseable candidate date)`
       };
     }
     if (expectedDate === null) {
       return {
-        path: path15,
+        path: path16,
         score: 0,
         weight,
         hit: false,
-        message: `${path15} (unparseable expected date)`
+        message: `${path16} (unparseable expected date)`
       };
     }
     if (candidateDate.getFullYear() === expectedDate.getFullYear() && candidateDate.getMonth() === expectedDate.getMonth() && candidateDate.getDate() === expectedDate.getDate()) {
       return {
-        path: path15,
+        path: path16,
         score: 1,
         weight,
         hit: true,
-        message: path15
+        message: path16
       };
     }
     return {
-      path: path15,
+      path: path16,
       score: 0,
       weight,
       hit: false,
-      message: `${path15} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
+      message: `${path16} (date mismatch: got ${formatDateISO(candidateDate)}, expected ${formatDateISO(expectedDate)})`
     };
   }
   /**
@@ -6747,11 +7386,11 @@ var FieldAccuracyEvaluator = class {
     };
   }
 };
-function resolvePath(obj, path15) {
-  if (!path15 || !obj) {
+function resolvePath(obj, path16) {
+  if (!path16 || !obj) {
     return void 0;
   }
-  const parts = path15.split(/\.|\[|\]/).filter((p) => p.length > 0);
+  const parts = path16.split(/\.|\[|\]/).filter((p) => p.length > 0);
   let current = obj;
   for (const part of parts) {
     if (current === null || current === void 0) {
@@ -6976,6 +7615,27 @@ function argsMatch(expected, actual) {
   }
   return true;
 }
+function checkLatency(toolName, maxDurationMs, actualDurationMs) {
+  if (maxDurationMs === void 0) {
+    return { status: "skip", message: "" };
+  }
+  if (actualDurationMs === void 0) {
+    return {
+      status: "skip",
+      message: `No duration data for ${toolName}; latency assertion skipped`
+    };
+  }
+  if (actualDurationMs <= maxDurationMs) {
+    return {
+      status: "pass",
+      message: `${toolName} completed in ${actualDurationMs}ms (max: ${maxDurationMs}ms)`
+    };
+  }
+  return {
+    status: "fail",
+    message: `${toolName} took ${actualDurationMs}ms (max: ${maxDurationMs}ms)`
+  };
+}
 var ToolTrajectoryEvaluator = class {
   kind = "tool_trajectory";
   config;
@@ -7034,7 +7694,8 @@ var ToolTrajectoryEvaluator = class {
         for (const call of message.toolCalls) {
           toolCalls.push({
             name: call.tool,
-            args: call.input
+            args: call.input,
+            durationMs: call.durationMs
           });
         }
       }
@@ -7102,17 +7763,27 @@ var ToolTrajectoryEvaluator = class {
     }
     const hits = [];
     const misses = [];
+    const warnings = [];
     let actualIndex = 0;
+    let sequenceHits = 0;
+    let latencyHits = 0;
+    let latencySkips = 0;
+    const latencyAssertionCount = expected.filter(
+      (item) => item.maxDurationMs !== void 0
+    ).length;
     for (let i = 0; i < expected.length; i++) {
       const expectedItem = expected[i];
       const expectedTool = expectedItem.tool;
       let found = false;
       let argsMismatch = false;
+      let matchedCall;
       while (actualIndex < toolCalls.length) {
         const actualCall = toolCalls[actualIndex];
         if (actualCall.name === expectedTool) {
           if (argsMatch(expectedItem.args, actualCall.args)) {
             hits.push(`Found ${expectedTool} at position ${actualIndex}`);
+            sequenceHits++;
+            matchedCall = actualCall;
             actualIndex++;
             found = true;
             break;
@@ -7129,14 +7800,35 @@ var ToolTrajectoryEvaluator = class {
       if (!found && !argsMismatch) {
         misses.push(`Expected ${expectedTool} at position ${i}, not found in remaining trace`);
       }
+      if (found && matchedCall) {
+        const latencyResult = checkLatency(
+          expectedTool,
+          expectedItem.maxDurationMs,
+          matchedCall.durationMs
+        );
+        if (latencyResult.status === "pass") {
+          hits.push(latencyResult.message);
+          latencyHits++;
+        } else if (latencyResult.status === "fail") {
+          misses.push(latencyResult.message);
+        } else if (latencyResult.message) {
+          warnings.push(latencyResult.message);
+          latencySkips++;
+        }
+      }
     }
-    const score = hits.length / expected.length;
+    for (const warning of warnings) {
+      console.warn(`[tool_trajectory] ${warning}`);
+    }
+    const effectiveLatencyAssertions = latencyAssertionCount - latencySkips;
+    const totalAssertions = expected.length + effectiveLatencyAssertions;
+    const score = totalAssertions > 0 ? (sequenceHits + latencyHits) / totalAssertions : 1;
     return {
       score,
       verdict: scoreToVerdict(score),
       hits,
       misses,
-      expectedAspectCount: expected.length
+      expectedAspectCount: totalAssertions
     };
   }
   evaluateExact(toolCalls) {
@@ -7152,6 +7844,13 @@ var ToolTrajectoryEvaluator = class {
     }
     const hits = [];
     const misses = [];
+    const warnings = [];
+    let sequenceHits = 0;
+    let latencyHits = 0;
+    let latencySkips = 0;
+    const latencyAssertionCount = expected.filter(
+      (item) => item.maxDurationMs !== void 0
+    ).length;
     if (toolCalls.length !== expected.length) {
       misses.push(`Expected ${expected.length} tool calls, got ${toolCalls.length}`);
     }
@@ -7161,33 +7860,58 @@ var ToolTrajectoryEvaluator = class {
       const expectedTool = expectedItem.tool;
       const actualCall = toolCalls[i];
       const actualTool = actualCall.name;
+      let sequenceMatched = false;
       if (actualTool === expectedTool) {
         if (argsMatch(expectedItem.args, actualCall.args)) {
           hits.push(`Position ${i}: ${expectedTool}`);
+          sequenceHits++;
+          sequenceMatched = true;
         } else {
           misses.push(`Position ${i}: ${expectedTool} args mismatch`);
         }
       } else {
         misses.push(`Position ${i}: expected ${expectedTool}, got ${actualTool}`);
       }
+      if (sequenceMatched) {
+        const latencyResult = checkLatency(
+          expectedTool,
+          expectedItem.maxDurationMs,
+          actualCall.durationMs
+        );
+        if (latencyResult.status === "pass") {
+          hits.push(latencyResult.message);
+          latencyHits++;
+        } else if (latencyResult.status === "fail") {
+          misses.push(latencyResult.message);
+        } else if (latencyResult.message) {
+          warnings.push(latencyResult.message);
+          latencySkips++;
+        }
+      }
     }
     for (let i = checkLength; i < expected.length; i++) {
       misses.push(`Position ${i}: expected ${expected[i].tool}, got nothing`);
     }
-    const score = hits.length / expected.length;
+    for (const warning of warnings) {
+      console.warn(`[tool_trajectory] ${warning}`);
+    }
+    const effectiveLatencyAssertions = latencyAssertionCount - latencySkips;
+    const totalAssertions = expected.length + effectiveLatencyAssertions;
+    const score = totalAssertions > 0 ? (sequenceHits + latencyHits) / totalAssertions : 1;
     return {
       score,
       verdict: scoreToVerdict(score),
       hits,
       misses,
-      expectedAspectCount: expected.length
+      expectedAspectCount: totalAssertions
     };
   }
 };
 
 // src/evaluation/orchestrator.ts
 import { createHash } from "node:crypto";
-import path14 from "node:path";
+import path15 from "node:path";
+import micromatch4 from "micromatch";
 
 // ../../node_modules/.bun/yocto-queue@1.2.2/node_modules/yocto-queue/index.js
 var Node = class {
@@ -7346,17 +8070,17 @@ async function runEvaluation(options) {
     cache,
     useCache,
     now,
-    evalId,
+    filter,
     verbose,
     evalCases: preloadedEvalCases,
     onResult,
     onProgress
   } = options;
-  const evalCases = preloadedEvalCases ?? await loadEvalCases(evalFilePath, repoRoot, { verbose, evalId });
-  const filteredEvalCases = filterEvalCases(evalCases, evalId);
+  const evalCases = preloadedEvalCases ?? await loadEvalCases(evalFilePath, repoRoot, { verbose, filter });
+  const filteredEvalCases = filterEvalCases(evalCases, filter);
   if (filteredEvalCases.length === 0) {
-    if (evalId) {
-      throw new Error(`Eval case with id '${evalId}' not found in ${evalFilePath}`);
+    if (filter) {
+      throw new Error(`No eval cases matched filter '${filter}' in ${evalFilePath}`);
     }
     return [];
   }
@@ -7932,7 +8656,10 @@ async function runEvaluatorList(options) {
           attempt,
           promptInputs,
           now,
-          judgeProvider
+          judgeProvider,
+          outputMessages,
+          traceSummary,
+          agentTimeoutMs
         });
         const weight = evaluator.weight ?? 1;
         scored.push({ score: score2, name: evaluator.name, type: evaluator.type, weight });
@@ -7986,7 +8713,7 @@ async function runEvaluatorList(options) {
         });
       }
       if (evaluator.type === "composite") {
-        const evalFileDir = evalCase.guideline_paths[0] ? path14.dirname(evalCase.guideline_paths[0]) : process.cwd();
+        const evalFileDir = evalCase.guideline_paths[0] ? path15.dirname(evalCase.guideline_paths[0]) : process.cwd();
         const createEvaluator = (memberConfig) => {
           switch (memberConfig.type) {
             case "llm_judge":
@@ -8267,9 +8994,22 @@ async function runLlmJudgeEvaluator(options) {
     attempt,
     promptInputs,
     now,
-    judgeProvider
+    judgeProvider,
+    outputMessages,
+    traceSummary,
+    agentTimeoutMs
   } = options;
-  const customPrompt = await resolveCustomPrompt(config);
+  const customPrompt = await resolveCustomPrompt(
+    config,
+    {
+      evalCase,
+      candidate,
+      outputMessages,
+      traceSummary,
+      config: config.config
+    },
+    agentTimeoutMs
+  );
   return evaluatorRegistry.llm_judge.evaluate({
     evalCase,
     candidate,
@@ -8283,23 +9023,70 @@ async function runLlmJudgeEvaluator(options) {
     evaluator: config
   });
 }
-async function resolveCustomPrompt(config) {
-  if (config.promptPath) {
+async function resolveCustomPrompt(promptConfig, context, timeoutMs) {
+  if (promptConfig.resolvedPromptScript && promptConfig.resolvedPromptScript.length > 0) {
+    if (!context) {
+      throw new Error("Context required for executable prompt templates");
+    }
+    return executePromptTemplate(
+      promptConfig.resolvedPromptScript,
+      context,
+      promptConfig.config,
+      timeoutMs
+    );
+  }
+  const promptPath = promptConfig.resolvedPromptPath ?? promptConfig.promptPath;
+  if (promptPath) {
     try {
-      const content = await readTextFile(config.promptPath);
+      const content = await readTextFile(promptPath);
       return content;
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
-      console.warn(`Could not read custom prompt at ${config.promptPath}: ${message}`);
+      console.warn(`Could not read custom prompt at ${promptPath}: ${message}`);
     }
   }
-  return config.prompt;
+  const promptValue = promptConfig.prompt;
+  if (typeof promptValue === "string") {
+    return promptValue;
+  }
+  return void 0;
+}
+async function executePromptTemplate(script, context, config, timeoutMs) {
+  const payload = {
+    question: context.evalCase.question,
+    expectedOutcome: context.evalCase.expected_outcome,
+    expectedMessages: context.evalCase.expected_messages,
+    referenceAnswer: context.evalCase.reference_answer,
+    candidateAnswer: context.candidate,
+    outputMessages: context.outputMessages ?? null,
+    guidelineFiles: context.evalCase.guideline_paths,
+    inputFiles: context.evalCase.file_paths.filter(
+      (p) => !context.evalCase.guideline_paths.includes(p)
+    ),
+    inputMessages: context.evalCase.input_messages,
+    traceSummary: context.traceSummary ?? null,
+    config: config ?? context.config ?? null
+  };
+  const inputJson = JSON.stringify(toSnakeCaseDeep(payload), null, 2);
+  const scriptPath = script[script.length - 1];
+  const cwd = path15.dirname(scriptPath);
+  try {
+    const stdout = await executeScript(script, inputJson, timeoutMs, cwd);
+    const prompt = stdout.trim();
+    if (!prompt) {
+      throw new Error("Prompt template produced empty output");
+    }
+    return prompt;
+  } catch (error) {
+    const message = error instanceof Error ? error.message : String(error);
+    throw new Error(`Prompt template execution failed: ${message}`);
+  }
 }
-function filterEvalCases(evalCases, evalId) {
-  if (!evalId) {
+function filterEvalCases(evalCases, filter) {
+  if (!filter) {
     return evalCases;
   }
-  return evalCases.filter((evalCase) => evalCase.id === evalId);
+  return evalCases.filter((evalCase) => micromatch4.isMatch(evalCase.id, filter));
 }
 function buildEvaluatorRegistry(overrides, resolveJudgeProvider) {
   const llmJudge = overrides?.llm_judge ?? new LlmJudgeEvaluator({
@@ -8457,7 +9244,7 @@ import { generateText as generateText4 } from "ai";
 import { z as z3 } from "zod";
 var rubricItemSchema = z3.object({
   id: z3.string().describe("Short identifier for this rubric (e.g., clarity, completeness)"),
-  description: z3.string().describe("What this rubric checks for"),
+  expected_outcome: z3.string().describe("Concrete expected outcome for this rubric item"),
   weight: z3.number().default(1).describe("Relative importance (default 1.0)"),
   required: z3.boolean().default(true).describe("Whether this is a mandatory requirement")
 });
@@ -8477,7 +9264,7 @@ You must return a valid JSON object matching this schema:
   "rubrics": [
     {
       "id": "string (short identifier)",
-      "description": "string (what to check)",
+      "expected_outcome": "string (concrete expected outcome for this rubric item)",
       "weight": number (default 1.0),
       "required": boolean (default true)
     }
@@ -8513,7 +9300,7 @@ function buildPrompt(expectedOutcome, question, referenceAnswer) {
     "Each rubric should:",
     "- Be specific and testable",
     "- Have a short, descriptive ID",
-    "- Include a clear description of what to check",
+    "- Include a clear expected outcome statement (what a good answer must demonstrate for this rubric)",
     "- Indicate if it is required (mandatory) or optional",
     "- Have an appropriate weight (default 1.0, use higher values for more important aspects)",
     "",
@@ -8560,6 +9347,7 @@ export {
   createAgentKernel,
   createProvider,
   deepEqual,
+  detectFormat,
   ensureVSCodeSubagents,
   executeScript,
   explorationRatio,