@axlsdk/eval 0.16.1 → 0.17.0

package/dist/cli.js

@@ -4,7 +4,7 @@ import {
   evalCompare,
   rescore,
   runEval
-} from "./chunk-CV6VE5VT.js";
+} from "./chunk-NPKIP4SB.js";
 
 // src/cli.ts
 import { readdirSync, statSync } from "fs";
@@ -12,65 +12,52 @@ import { readFile as readFileAsync, writeFile as writeFileAsync, mkdir } from "f
 import * as path from "path";
 
 // src/cli-utils.ts
-import { resolve } from "path";
-import { existsSync } from "fs";
-import { pathToFileURL } from "url";
-var CONFIG_CANDIDATES = [
-  "axl.config.mts",
-  "axl.config.ts",
-  "axl.config.mjs",
-  "axl.config.js"
-];
-function findConfig(cwd) {
-  for (const name of CONFIG_CANDIDATES) {
-    const p = resolve(cwd, name);
-    if (existsSync(p)) return p;
+import {
+  resolveRuntime,
+  pickDefault,
+  pickExport,
+  CONFIG_CANDIDATES,
+  findConfig,
+  needsTsxLoader,
+  importModule,
+  expandGlob,
+  registerConditions
+} from "@axlsdk/axl";
+
+// src/cli-validate.ts
+function validateEvalConfig(cfg) {
+  if (!cfg || typeof cfg !== "object") {
+    return `does not export a valid eval config \u2014 got ${cfg === null ? "null" : typeof cfg}.`;
   }
-  return void 0;
-}
-function needsTsxLoader(filePath) {
-  return /\.[mc]?tsx?$/.test(filePath);
-}
-function resolveRuntime(mod) {
-  const def = mod.default;
-  return def?.default ?? def ?? mod.runtime;
-}
-var tsImportFn;
-async function importModule(filePath, parentURL) {
-  if (needsTsxLoader(filePath)) {
-    if (tsImportFn === void 0) {
-      try {
-        const mod = await import("tsx/esm/api");
-        tsImportFn = mod.tsImport ?? null;
-      } catch {
-        tsImportFn = null;
-        console.warn(
-          "[axl-eval] Warning: tsx is not installed. TypeScript files require tsx.\n  Install it with: npm install -D tsx"
-        );
-      }
-    }
-    if (tsImportFn) {
-      return await tsImportFn(pathToFileURL(filePath).href, parentURL);
-    }
+  const c = cfg;
+  const missing = [];
+  if (!c.workflow) missing.push("workflow");
+  if (!c.dataset) missing.push("dataset");
+  if (!c.scorers) missing.push("scorers");
+  if (missing.length > 0) {
+    const keys = Object.keys(c).slice(0, 10);
+    const got = keys.length > 0 ? ` Got: { ${keys.join(", ")} }.` : "";
+    return `does not export a valid eval config (missing ${missing.join(", ")}).${got}`;
   }
-  return await import(pathToFileURL(filePath).href);
-}
-async function registerConditions(conditions) {
-  try {
-    const nodeModule = await import("module");
-    const hookCode = [
-      `const extra = ${JSON.stringify(conditions)};`,
-      `export async function resolve(specifier, context, nextResolve) {`,
-      `  return nextResolve(specifier, {`,
-      `    ...context,`,
-      `    conditions: [...new Set([...context.conditions, ...extra])],`,
-      `  });`,
-      `}`
-    ].join("\n");
-    nodeModule.register(`data:text/javascript,${encodeURIComponent(hookCode)}`);
-  } catch {
-    console.warn("[axl-eval] Warning: --conditions requires Node.js 20.6+");
+  if (typeof c.workflow !== "string") {
+    return `exports a non-string workflow (got ${typeof c.workflow}) \u2014 workflow must be the registered workflow name.`;
+  }
+  if (typeof c.dataset !== "object" || Array.isArray(c.dataset)) {
+    return `exports a non-object dataset (got ${Array.isArray(c.dataset) ? "array" : typeof c.dataset}) \u2014 use the dataset() factory.`;
   }
+  if (typeof c.dataset.getItems !== "function") {
+    return `exports a dataset without a getItems() method \u2014 use the dataset() factory.`;
+  }
+  if (!Array.isArray(c.scorers) || c.scorers.length === 0) {
+    return `exports an empty scorers array \u2014 at least one scorer is required.`;
+  }
+  const badScorerIdx = c.scorers.findIndex(
+    (s) => !s || typeof s.score !== "function"
+  );
+  if (badScorerIdx !== -1) {
+    return `exports a scorers array with a non-scorer entry at index ${badScorerIdx} \u2014 each scorer must have a score() method (use the scorer() factory).`;
+  }
+  return void 0;
 }
 
 // src/cli.ts
@@ -226,9 +213,15 @@ async function runRescore(args) {
   const runtime = await getRuntime(configArg, conditions);
   try {
     const mod = await importModule(path.resolve(evalFilePath), import.meta.url);
-    const evalConfig = mod.default?.default ?? mod.default ?? mod.config ?? mod;
-    if (!evalConfig.scorers || evalConfig.scorers.length === 0) {
-      console.error(`Error: ${evalFilePath} does not export scorers`);
+    const evalConfig = pickDefault(mod);
+    if (!evalConfig || typeof evalConfig !== "object") {
+      console.error(
+        `Error: ${evalFilePath} does not export an eval config object \u2014 got ${evalConfig === null ? "null" : typeof evalConfig}.`
+      );
+      process.exit(1);
+    }
+    if (!Array.isArray(evalConfig.scorers) || evalConfig.scorers.length === 0) {
+      console.error(`Error: ${evalFilePath} does not export a non-empty scorers array.`);
       process.exit(1);
     }
     const rescored = [];
@@ -251,6 +244,7 @@ async function runRescore(args) {
   }
 }
 function collectEvalFiles(p) {
+  if (p.includes("*")) return expandGlob(p, process.cwd());
   const resolved = path.resolve(p);
   try {
     const stat = statSync(resolved);
@@ -461,26 +455,38 @@ async function runEvalCommand(args) {
   }
   const runtime = await getRuntime(configArg, conditions);
   const results = [];
+  let failedFiles = 0;
   try {
     for (const filePath of evalFiles) {
       try {
         const mod = await importModule(path.resolve(filePath), import.meta.url);
-        const evalConfig = mod.default?.default ?? mod.default ?? mod.config ?? mod;
-        if (!evalConfig.workflow || !evalConfig.dataset || !evalConfig.scorers) {
+        const evalConfig = pickDefault(mod);
+        const validationError = validateEvalConfig(evalConfig);
+        if (validationError) {
+          console.error(`Error: ${filePath} ${validationError}`);
+          failedFiles++;
+          continue;
+        }
+        const exported = pickExport(mod, "executeWorkflow");
+        let customExecute;
+        if (typeof exported === "function") {
+          customExecute = exported;
+        } else if (exported !== void 0) {
           console.error(
-            `Error: ${filePath} does not export a valid eval config (missing workflow, dataset, or scorers)`
+            `[axl-eval] Error: ${filePath} exports executeWorkflow but it is ${typeof exported}, not a function.`
           );
+          failedFiles++;
           continue;
         }
         let executeWorkflow;
-        if (mod.executeWorkflow) {
+        if (customExecute) {
           executeWorkflow = async (input, rt) => {
             const {
               result,
               cost: trackedCost,
               metadata
             } = await runtime.trackExecution(async () => {
-              return mod.executeWorkflow(input, rt);
+              return customExecute(input, rt);
             });
             return {
               output: result.output,
@@ -496,26 +502,69 @@ async function runEvalCommand(args) {
             return { output: result, cost, metadata };
           };
         } else {
-          console.warn(
-            `[axl-eval] Warning: ${filePath} does not export executeWorkflow \u2014 using input passthrough`
+          const exportShape = (() => {
+            const top = Object.keys(mod).filter((k) => k !== "default");
+            const fromDefault = mod.default && typeof mod.default === "object" ? Object.keys(mod.default) : [];
+            const all = [.../* @__PURE__ */ new Set([...top, ...fromDefault])];
+            return all.length > 0 ? `Found exports: [${all.join(", ")}].` : "";
+          })();
+          console.error(
+            `[axl-eval] Error: ${filePath} does not export an executeWorkflow function and no workflow named "${evalConfig.workflow}" is registered on the runtime.
+` + (exportShape ? `  ${exportShape}
+` : "") + `  Add either:
+    export async function executeWorkflow(input, runtime) { ... }
+  or register the workflow on your runtime via runtime.registerWorkflow(workflow).
+  Tip: if your eval previously worked, your package.json may be missing
+    "type": "module" \u2014 tsx loads .ts as CJS, putting named exports on
+    mod.default.<name>. The CLI handles that interop, but if you renamed
+    or removed the export the file path is your fix.`
           );
-          executeWorkflow = async (input) => ({ output: input });
+          failedFiles++;
+          continue;
         }
         const runOptions = captureTraces ? { captureTraces: true } : void 0;
         if (runs > 1) {
           const { randomUUID } = await import("crypto");
           const runGroupId = randomUUID();
           const runResults = [];
+          let runFailure;
           for (let r = 0; r < runs; r++) {
             console.error(`[axl-eval] Run ${r + 1}/${runs}...`);
-            const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);
-            result.metadata.runGroupId = runGroupId;
-            result.metadata.runIndex = r;
-            runResults.push(result);
-            results.push(result);
+            try {
+              const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);
+              result.metadata.runGroupId = runGroupId;
+              result.metadata.runIndex = r;
+              runResults.push(result);
+            } catch (err) {
+              runFailure = err instanceof Error ? err : new Error(String(err));
+              console.error(`[axl-eval] Run ${r + 1}/${runs} failed: ${runFailure.message}`);
+              break;
+            }
+          }
+          if (runResults.length === 0) {
+            console.error(
+              `[axl-eval] Aborting ${filePath}: no runs completed (${runFailure?.message ?? "unknown failure"}).`
+            );
+            failedFiles++;
+            continue;
+          }
+          const partial = runResults.length < runs;
+          if (partial) {
+            const failureMsg = runFailure ? runFailure.message || String(runFailure) || void 0 : void 0;
+            for (const r of runResults) {
+              r.metadata.fromPartialBatch = true;
+              r.metadata.batchCompleted = runResults.length;
+              r.metadata.batchAttempted = runs;
+              if (failureMsg) r.metadata.batchFailure = failureMsg;
+            }
+            console.error(
+              `[axl-eval] PARTIAL: ${runResults.length} of ${runs} runs completed for ${filePath}.`
+            );
+            failedFiles++;
           }
           const summary = aggregateRuns(runResults);
           console.log("\n" + formatMultiRunTable(summary) + "\n");
+          for (const r of runResults) results.push(r);
         } else {
           const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);
           results.push(result);
@@ -525,6 +574,7 @@ async function runEvalCommand(args) {
         console.error(
           `Error running eval ${filePath}: ${err instanceof Error ? err.message : String(err)}`
         );
+        failedFiles++;
       }
     }
     if (outputPath && results.length > 0) {
@@ -538,7 +588,7 @@ async function runEvalCommand(args) {
     await runtime.shutdown().catch(() => {
     });
   }
-  if (results.length === 0) {
+  if (failedFiles > 0 || results.length === 0) {
     process.exit(1);
   }
 }

package/dist/cli.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/cli.ts","../src/cli-utils.ts"],"sourcesContent":["#!/usr/bin/env node\n\nimport { readdirSync, statSync } from 'node:fs';\nimport { readFile as readFileAsync, writeFile as writeFileAsync, mkdir } from 'node:fs/promises';\nimport * as path from 'node:path';\nimport type { AxlRuntime } from '@axlsdk/axl';\nimport { evalCompare } from './compare.js';\nimport { runEval } from './runner.js';\nimport { rescore } from './rescore.js';\nimport { aggregateRuns } from './multi-run.js';\nimport type { MultiRunSummary } from './multi-run.js';\nimport type { EvalConfig, EvalResult } from './types.js';\nimport {\n  findConfig,\n  resolveRuntime,\n  importModule,\n  registerConditions,\n  CONFIG_CANDIDATES,\n} from './cli-utils.js';\n\nconst KNOWN_FLAGS = new Set([\n  '--output',\n  '--config',\n  '--conditions',\n  '--fail-on-regression',\n  '--threshold',\n  '--runs',\n  '--capture-traces',\n]);\n\nasync function main() {\n  const args = process.argv.slice(2);\n\n  if (args.length === 0 || args[0] === '--help' || args[0] === '-h') {\n    console.log(`\nUsage:\n  axl-eval <path>                         Run eval file(s)\n  axl-eval <path> --runs <n>              Run eval N times (multi-run)\n  axl-eval <path> --output <file>         Save results to JSON\n  axl-eval <path> --config <file>         Use config file for runtime\n  axl-eval <path> --conditions <list>     Node.js import conditions (comma-separated)\n  axl-eval <path> --capture-traces        Populate EvalItem.traces on every item\n                                          (success + failure). Adds memory\n                                          overhead proportional to dataset size\n                                          x turns x agents; off by default.\n  axl-eval rescore <results> <eval-file>  Re-run scorers on saved outputs\n  axl-eval compare <a> <b>                Compare two eval result files\n  axl-eval compare <a> <b> --threshold <v>  Set regression threshold (global or per-scorer)\n  axl-eval compare <a> <b> --fail-on-regression  Exit 1 if regressions\n\nConfig auto-detection (when --config is not specified):\n  ${CONFIG_CANDIDATES.join(' -> ')}\n\nWhen a config is found, the exported AxlRuntime is passed to executeWorkflow\nand cost is tracked automatically via runtime.trackCost().\nWhen no config is found, a bare AxlRuntime is created (providers from env vars).\n`);\n    process.exit(0);\n  }\n\n  if (args[0] === 'compare') {\n    await runCompare(args.slice(1));\n    return;\n  }\n\n  if (args[0] === 'rescore') {\n    await runRescore(args.slice(1));\n    return;\n  }\n\n  await runEvalCommand(args);\n}\n\nfunction parseThresholdArg(args: string[]): Record<string, number> | number | undefined {\n  const idx = args.indexOf('--threshold');\n  if (idx === -1 || idx + 1 >= args.length) return undefined;\n  const value = args[idx + 1];\n\n  // Global numeric threshold: --threshold 0.05\n  if (/^[\\d.]+$/.test(value)) return parseFloat(value);\n\n  // Per-scorer map: --threshold accuracy=0,tone=0.1\n  const map: Record<string, number> = {};\n  for (const pair of value.split(',')) {\n    const [key, val] = pair.split('=');\n    if (key && val != null) map[key.trim()] = parseFloat(val.trim());\n  }\n  return Object.keys(map).length > 0 ? map : undefined;\n}\n\nasync function runCompare(args: string[]) {\n  const failOnRegression = args.includes('--fail-on-regression');\n  const thresholds = parseThresholdArg(args);\n  const files = args.filter(\n    (a, i) => !a.startsWith('--') && !(i > 0 && args[i - 1] === '--threshold'),\n  );\n\n  if (files.length !== 2) {\n    console.error('Usage: axl-eval compare <baseline.json> <candidate.json> [--threshold <value>]');\n    process.exit(1);\n  }\n\n  const baselineRaw = JSON.parse(await readFileAsync(files[0], 'utf-8'));\n  const candidateRaw = JSON.parse(await readFileAsync(files[1], 'utf-8'));\n  const baseline: EvalResult | EvalResult[] = baselineRaw;\n  const candidate: EvalResult | EvalResult[] = candidateRaw;\n  const compareOptions = thresholds != null ? { thresholds } : undefined;\n  const comparison = evalCompare(baseline, candidate, compareOptions);\n\n  console.log(\n    `\\nCompare: baseline (${comparison.baseline.id.slice(0, 8)}) -> candidate (${comparison.candidate.id.slice(0, 8)})\\n`,\n  );\n\n  const scorerNames = Object.keys(comparison.scorers);\n  const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 6);\n  const hasCI = scorerNames.some((n) => comparison.scorers[n].ci != null);\n\n  const ciHeader = hasCI ? '  CI 95%            Sig' : '';\n  const ciRule = hasCI ? '  ----------------  ---' : '';\n  console.log(`  ${'Scorer'.padEnd(maxNameLen)}  Baseline  Candidate  Delta     Change${ciHeader}`);\n  console.log(`  ${''.padEnd(maxNameLen, '-')}  --------  ---------  --------  ------${ciRule}`);\n\n  for (const name of scorerNames) {\n    const s = comparison.scorers[name];\n    const sign = s.delta > 0 ? '+' : '';\n    let line = `  ${name.padEnd(maxNameLen)}  ${s.baselineMean.toFixed(3).padStart(8)}  ${s.candidateMean.toFixed(3).padStart(9)}  ${(sign + s.delta.toFixed(3)).padStart(8)}  ${(sign + s.deltaPercent.toFixed(1) + '%').padStart(6)}`;\n    if (hasCI) {\n      if (s.ci) {\n        const lo = (s.ci.lower >= 0 ? '+' : '') + s.ci.lower.toFixed(4);\n        const hi = (s.ci.upper >= 0 ? '+' : '') + s.ci.upper.toFixed(4);\n        const sig = s.significant ? '  *' : '   ';\n        line += `  [${lo}, ${hi}]${sig}`;\n      } else {\n        line += '  —'.padEnd(22);\n      }\n    }\n    console.log(line);\n  }\n\n  if (comparison.timing) {\n    const t = comparison.timing;\n    const sign = t.delta > 0 ? '+' : '';\n    console.log(\n      `\\n  Timing: baseline ${(t.baselineMean / 1000).toFixed(2)}s -> candidate ${(t.candidateMean / 1000).toFixed(2)}s (${sign}${t.deltaPercent.toFixed(1)}%)`,\n    );\n  }\n  if (comparison.cost) {\n    const c = comparison.cost;\n    const sign = c.delta > 0 ? '+' : '';\n    console.log(\n      `  Cost: baseline $${c.baselineTotal.toFixed(2)} -> candidate $${c.candidateTotal.toFixed(2)} (${sign}${c.deltaPercent.toFixed(1)}%)`,\n    );\n  }\n\n  const baselineRef = Array.isArray(baseline) ? baseline[0] : baseline;\n  const stable = Math.max(\n    0,\n    baselineRef.items.length - comparison.regressions.length - comparison.improvements.length,\n  );\n  console.log(\n    `\\n  Regressions: ${comparison.regressions.length} | Improvements: ${comparison.improvements.length} | Stable: ${stable}\\n`,\n  );\n\n  if (failOnRegression && comparison.regressions.length > 0) {\n    // When CI is available, only fail on significant regressions\n    const hasSignificance = scorerNames.some((n) => comparison.scorers[n].significant != null);\n    if (hasSignificance) {\n      const hasSignificantRegression = scorerNames.some(\n        (n) => comparison.scorers[n].significant === true && comparison.scorers[n].delta < 0,\n      );\n      if (hasSignificantRegression) process.exit(1);\n    } else {\n      process.exit(1);\n    }\n  }\n}\n\nasync function runRescore(args: string[]) {\n  const { outputPath, configArg, conditions, paths } = parseEvalArgs(args);\n\n  if (paths.length < 2) {\n    console.error('Usage: axl-eval rescore <results.json> <eval-file> [--output <file>]');\n    process.exit(1);\n  }\n\n  const [resultsPath, evalFilePath] = paths;\n  const raw = JSON.parse(await readFileAsync(resultsPath, 'utf-8'));\n  const results: EvalResult[] = Array.isArray(raw) ? raw : [raw];\n\n  const runtime = await getRuntime(configArg, conditions);\n  try {\n    const mod = await importModule(path.resolve(evalFilePath), import.meta.url);\n    const evalConfig: EvalConfig = mod.default?.default ?? mod.default ?? mod.config ?? mod;\n\n    if (!evalConfig.scorers || evalConfig.scorers.length === 0) {\n      console.error(`Error: ${evalFilePath} does not export scorers`);\n      process.exit(1);\n    }\n\n    const rescored: EvalResult[] = [];\n    for (const resultData of results) {\n      rescored.push(await rescore(resultData, evalConfig.scorers, runtime));\n    }\n\n    for (const r of rescored) {\n      console.log('\\n' + formatTable(r) + '\\n');\n    }\n\n    if (outputPath) {\n      const output = Array.isArray(raw) ? rescored : rescored[0];\n      const outputDir = path.dirname(path.resolve(outputPath));\n      await mkdir(outputDir, { recursive: true });\n      await writeFileAsync(outputPath, JSON.stringify(output, null, 2), 'utf-8');\n      console.log(`Rescored results saved to ${outputPath}`);\n    }\n  } finally {\n    await runtime.shutdown().catch(() => {});\n  }\n}\n\nfunction collectEvalFiles(p: string): string[] {\n  const resolved = path.resolve(p);\n  try {\n    const stat = statSync(resolved);\n    if (stat.isDirectory()) {\n      const entries = readdirSync(resolved);\n      return entries\n        .filter((e) => /\\.eval\\.[mc]?[jt]sx?$/.test(e))\n        .map((e) => path.join(resolved, e));\n    }\n    return [resolved];\n  } catch {\n    return [resolved];\n  }\n}\n\n/** Render observed workflows for a terse header line. Multiple → comma-separated. */\nfunction formatWorkflows(workflows: unknown): string {\n  if (!Array.isArray(workflows) || workflows.length === 0) return '(unknown)';\n  return (workflows as unknown[]).filter((w): w is string => typeof w === 'string').join(', ');\n}\n\nfunction formatTable(result: EvalResult): string {\n  const lines: string[] = [];\n  const scorerNames = Object.keys(result.summary.scorers);\n  const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 'Scorer'.length);\n  const colWidth = 8;\n\n  lines.push(\n    `Eval: ${formatWorkflows(result.metadata.workflows)} x ${result.dataset} (${result.summary.count} items)`,\n  );\n  lines.push(\n    `  ${'Scorer'.padEnd(maxNameLen)}  ${'Mean'.padStart(colWidth)}  ${'Min'.padStart(colWidth)}  ${'Max'.padStart(colWidth)}  ${'p50'.padStart(colWidth)}  ${'p95'.padStart(colWidth)}`,\n  );\n\n  const ruleLen = maxNameLen + 2 + (colWidth + 2) * 5;\n  lines.push('  ' + '\\u2500'.repeat(ruleLen));\n\n  for (const name of scorerNames) {\n    const s = result.summary.scorers[name];\n    // Detect scorers with no valid scores (all items errored → computeStats([]) → all zeros)\n    const validScoreCount = result.items.filter((i) => !i.error && i.scores[name] != null).length;\n    if (validScoreCount === 0) {\n      lines.push(\n        `  ${name.padEnd(maxNameLen)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}`,\n      );\n    } else {\n      lines.push(\n        `  ${name.padEnd(maxNameLen)}  ${s.mean.toFixed(2).padStart(colWidth)}  ${s.min.toFixed(2).padStart(colWidth)}  ${s.max.toFixed(2).padStart(colWidth)}  ${s.p50.toFixed(2).padStart(colWidth)}  ${s.p95.toFixed(2).padStart(colWidth)}`,\n      );\n    }\n  }\n\n  if (result.summary.timing) {\n    const t = result.summary.timing;\n    lines.push(\n      `  ${'Timing'.padEnd(maxNameLen)}  ${(t.mean / 1000).toFixed(2).padStart(colWidth)}s ${(t.p50 / 1000).toFixed(2).padStart(colWidth)}s ${(t.p95 / 1000).toFixed(2).padStart(colWidth)}s`,\n    );\n  }\n\n  const durationSec = (result.duration / 1000).toFixed(1);\n  const costStr = result.totalCost > 0 ? `$${result.totalCost.toFixed(2)}` : '$0.00';\n  lines.push('');\n  lines.push(\n    `  Failures: ${result.summary.failures}/${result.summary.count} | Cost: ${costStr} | Duration: ${durationSec}s`,\n  );\n\n  const itemsWithErrors = result.items.filter((i) => i.scorerErrors?.length);\n  if (itemsWithErrors.length > 0) {\n    const uniqueErrors = [...new Set(itemsWithErrors.flatMap((i) => i.scorerErrors!))];\n    lines.push('');\n    lines.push(\n      `  Scorer errors (${itemsWithErrors.length}/${result.summary.count} items affected):`,\n    );\n    for (const err of uniqueErrors.slice(0, 5)) {\n      lines.push(`    - ${err}`);\n    }\n    if (uniqueErrors.length > 5) {\n      lines.push(`    ... and ${uniqueErrors.length - 5} more`);\n    }\n  }\n\n  return lines.join('\\n');\n}\n\nfunction formatMultiRunTable(summary: MultiRunSummary): string {\n  const lines: string[] = [];\n  const scorerNames = Object.keys(summary.scorers);\n  const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 'Scorer'.length);\n  const colWidth = 16;\n\n  lines.push(\n    `Eval: ${formatWorkflows(summary.workflows)} x ${summary.dataset} \\u2014 ${summary.runCount} runs`,\n  );\n  lines.push(\n    `  ${'Scorer'.padEnd(maxNameLen)}  ${'Mean \\u00b1 Std'.padStart(colWidth)}  ${'Min'.padStart(8)}  ${'Max'.padStart(8)}`,\n  );\n  const ruleLen = maxNameLen + 2 + colWidth + 2 + 8 + 2 + 8;\n  lines.push('  ' + '\\u2500'.repeat(ruleLen));\n\n  for (const name of scorerNames) {\n    const s = summary.scorers[name];\n    const meanStd = `${s.mean.toFixed(3)} \\u00b1 ${s.std.toFixed(3)}`;\n    lines.push(\n      `  ${name.padEnd(maxNameLen)}  ${meanStd.padStart(colWidth)}  ${s.min.toFixed(3).padStart(8)}  ${s.max.toFixed(3).padStart(8)}`,\n    );\n  }\n\n  if (summary.timing) {\n    lines.push(\n      `  ${'Timing'.padEnd(maxNameLen)}  ${((summary.timing.mean / 1000).toFixed(2) + ' \\u00b1 ' + (summary.timing.std / 1000).toFixed(2) + 's').padStart(colWidth)}`,\n    );\n  }\n\n  const costStr = summary.totalCost > 0 ? `$${summary.totalCost.toFixed(2)}` : '$0.00';\n  const durationStr = (summary.totalDuration / 1000).toFixed(1);\n  lines.push('');\n  lines.push(`  Total Cost: ${costStr} | Total Duration: ${durationStr}s`);\n\n  return lines.join('\\n');\n}\n\n// ── Runtime resolution ─────────────────────────────────────────────\n\nasync function resolveRuntimeFromConfig(configPath: string): Promise<AxlRuntime> {\n  try {\n    const mod = await importModule(configPath, import.meta.url);\n    const runtime = resolveRuntime(mod) as AxlRuntime;\n\n    if (!runtime || typeof runtime.execute !== 'function') {\n      console.error(`Config must export a default AxlRuntime instance.`);\n      if (runtime) {\n        const keys = Object.keys(runtime as object)\n          .slice(0, 5)\n          .join(', ');\n        console.error(`  Got: ${typeof runtime}${keys ? ` with keys: { ${keys} }` : ''}`);\n      }\n      console.error(\n        `Example:\\n  import { AxlRuntime } from '@axlsdk/axl';\\n  export default new AxlRuntime({ ... });`,\n      );\n      process.exit(1);\n    }\n\n    return runtime;\n  } catch (err) {\n    const msg = err instanceof Error ? err.message : String(err);\n    if (\n      /Cannot use import statement|Unexpected reserved word|top-level await|exports is not defined/.test(\n        msg,\n      )\n    ) {\n      const ext = path.extname(configPath);\n      console.error(`[axl-eval] Config failed to load due to a CJS/ESM compatibility issue.`);\n      if (ext === '.ts' || ext === '.tsx') {\n        console.error(\n          `  Tip: try renaming to .mts to force ESM format, or ensure tsx is installed and up to date.`,\n        );\n      } else {\n        console.error(`  Tip: add \"type\": \"module\" to your package.json.`);\n      }\n      console.error();\n    }\n    console.error(`Failed to load config:`, err);\n    process.exit(1);\n  }\n}\n\nasync function getRuntime(configArg?: string, conditions?: string[]): Promise<AxlRuntime> {\n  // Register import conditions before any config loading\n  if (conditions && conditions.length > 0) {\n    await registerConditions(conditions);\n  }\n\n  // 1. Explicit --config\n  if (configArg) {\n    const configPath = path.resolve(process.cwd(), configArg);\n    const stat = statSync(configPath, { throwIfNoEntry: false });\n    if (!stat?.isFile()) {\n      console.error(`Config file not found: ${configPath}`);\n      process.exit(1);\n    }\n    console.error(`[axl-eval] Loading config from ${configPath}`);\n    return resolveRuntimeFromConfig(configPath);\n  }\n\n  // 2. Auto-detect\n  const found = findConfig(process.cwd());\n  if (found) {\n    console.error(`[axl-eval] Auto-detected config: ${found}`);\n    return resolveRuntimeFromConfig(found);\n  }\n\n  // 3. Bare runtime (providers from env vars)\n  const { AxlRuntime } = await import('@axlsdk/axl');\n  return new AxlRuntime();\n}\n\n// ── Arg parsing ────────────────────────────────────────────────────\n\nfunction parseEvalArgs(args: string[]): {\n  outputPath?: string;\n  configArg?: string;\n  conditions: string[];\n  runs: number;\n  captureTraces: boolean;\n  paths: string[];\n} {\n  let outputPath: string | undefined;\n  let configArg: string | undefined;\n  let conditions: string[] = [];\n  let runs = 1;\n  let captureTraces = false;\n  const paths: string[] = [];\n\n  for (let i = 0; i < args.length; i++) {\n    const arg = args[i];\n    if (arg === '--output' || arg === '--config' || arg === '--conditions' || arg === '--runs') {\n      if (i + 1 >= args.length) {\n        console.error(`Error: ${arg} requires a value`);\n        process.exit(1);\n      }\n      const value = args[++i];\n      if (arg === '--output') outputPath = value;\n      else if (arg === '--config') configArg = value;\n      else if (arg === '--runs') runs = Math.max(1, parseInt(value, 10) || 1);\n      else\n        conditions = value\n          .split(',')\n          .map((c) => c.trim())\n          .filter(Boolean);\n    } else if (arg === '--capture-traces') {\n      // Boolean flag — no value consumed.\n      captureTraces = true;\n    } else if (arg.startsWith('--')) {\n      if (!KNOWN_FLAGS.has(arg)) {\n        console.error(`Unknown flag: ${arg}`);\n        process.exit(1);\n      }\n    } else {\n      paths.push(arg);\n    }\n  }\n\n  return { outputPath, configArg, conditions, runs, captureTraces, paths };\n}\n\n// ── Main eval command ──────────────────────────────────────────────\n\nasync function runEvalCommand(args: string[]) {\n  const { outputPath, configArg, conditions, runs, captureTraces, paths } = parseEvalArgs(args);\n\n  if (paths.length === 0) {\n    console.error('Error: No eval file path provided');\n    process.exit(1);\n  }\n\n  const evalFiles: string[] = [];\n  for (const p of paths) {\n    evalFiles.push(...collectEvalFiles(p));\n  }\n\n  if (evalFiles.length === 0) {\n    console.error('Error: No eval files found');\n    process.exit(1);\n  }\n\n  const runtime = await getRuntime(configArg, conditions);\n  const results: EvalResult[] = [];\n\n  try {\n    for (const filePath of evalFiles) {\n      try {\n        const mod = await importModule(path.resolve(filePath), import.meta.url);\n        const evalConfig: EvalConfig = mod.default?.default ?? mod.default ?? mod.config ?? mod;\n\n        if (!evalConfig.workflow || !evalConfig.dataset || !evalConfig.scorers) {\n          console.error(\n            `Error: ${filePath} does not export a valid eval config (missing workflow, dataset, or scorers)`,\n          );\n          continue;\n        }\n\n        // Resolve executeWorkflow: custom export > registered workflow > passthrough\n        let executeWorkflow: (\n          input: unknown,\n          rt?: unknown,\n        ) => Promise<{ output: unknown; cost?: number; metadata?: Record<string, unknown> }>;\n\n        if (mod.executeWorkflow) {\n          // Wrap custom executeWorkflow with trackExecution for cost + metadata attribution\n          executeWorkflow = async (input, rt) => {\n            const {\n              result,\n              cost: trackedCost,\n              metadata,\n            } = await runtime.trackExecution(async () => {\n              return mod.executeWorkflow(input, rt);\n            });\n            return {\n              output: result.output,\n              cost: result.cost ?? trackedCost,\n              metadata: result.metadata ?? metadata,\n            };\n          };\n        } else if (runtime.getWorkflow(evalConfig.workflow)) {\n          // No executeWorkflow exported but workflow is registered — use runtime.execute()\n          executeWorkflow = async (input) => {\n            const { result, cost, metadata } = await runtime.trackExecution(async () => {\n              return runtime.execute(evalConfig.workflow, input);\n            });\n            return { output: result, cost, metadata };\n          };\n        } else {\n          console.warn(\n            `[axl-eval] Warning: ${filePath} does not export executeWorkflow — using input passthrough`,\n          );\n          executeWorkflow = async (input) => ({ output: input });\n        }\n\n        // When --capture-traces is set, forward it to runEval so the runner\n        // wraps each item's execution in a nested trackExecution({ captureTraces: true }).\n        // This populates EvalItem.traces on both success (via tracked.traces)\n        // and failure (via the axlCapturedTraces side-channel on the thrown\n        // error). Nested trackExecution scopes walk the AsyncLocalStorage\n        // parent chain, so the outer trackExecution above still observes\n        // cost/metadata correctly.\n        const runOptions = captureTraces ? { captureTraces: true } : undefined;\n\n        if (runs > 1) {\n          // Multi-run mode\n          const { randomUUID } = await import('node:crypto');\n          const runGroupId = randomUUID();\n          const runResults: EvalResult[] = [];\n\n          for (let r = 0; r < runs; r++) {\n            console.error(`[axl-eval] Run ${r + 1}/${runs}...`);\n            const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);\n            result.metadata.runGroupId = runGroupId;\n            result.metadata.runIndex = r;\n            runResults.push(result);\n            results.push(result);\n          }\n\n          const summary = aggregateRuns(runResults);\n          console.log('\\n' + formatMultiRunTable(summary) + '\\n');\n        } else {\n          const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);\n          results.push(result);\n\n          console.log('\\n' + formatTable(result) + '\\n');\n        }\n      } catch (err) {\n        console.error(\n          `Error running eval ${filePath}: ${err instanceof Error ? err.message : String(err)}`,\n        );\n      }\n    }\n\n    if (outputPath && results.length > 0) {\n      const output = results.length === 1 ? results[0] : results;\n      const outputDir = path.dirname(path.resolve(outputPath));\n      await mkdir(outputDir, { recursive: true });\n      await writeFileAsync(outputPath, JSON.stringify(output, null, 2), 'utf-8');\n      console.log(`Results saved to ${outputPath}`);\n    }\n  } finally {\n    await runtime.shutdown().catch(() => {});\n  }\n\n  // Exit with non-zero code if no evals succeeded\n  if (results.length === 0) {\n    process.exit(1);\n  }\n}\n\nmain().catch((err) => {\n  console.error(err);\n  process.exit(1);\n});\n","/**\n * CLI utilities for config detection, runtime resolution, and loader registration.\n *\n * These are duplicated from @axlsdk/studio (cli-utils.ts, resolve-runtime.ts,\n * eval-loader.ts, cli.ts) because @axlsdk/eval cannot depend on @axlsdk/studio.\n * Keep in sync with the studio versions if either changes.\n */\n\nimport { resolve } from 'node:path';\nimport { existsSync } from 'node:fs';\nimport { pathToFileURL } from 'node:url';\n\n// ── Config auto-detection ──────────────────────────────────────────\n\nexport const CONFIG_CANDIDATES = [\n  'axl.config.mts',\n  'axl.config.ts',\n  'axl.config.mjs',\n  'axl.config.js',\n];\n\nexport function findConfig(cwd: string): string | undefined {\n  for (const name of CONFIG_CANDIDATES) {\n    const p = resolve(cwd, name);\n    if (existsSync(p)) return p;\n  }\n  return undefined;\n}\n\n// ── Extension helpers ──────────────────────────────────────────────\n\n/** Returns true if the file is TypeScript and needs tsx to load. */\nexport function needsTsxLoader(filePath: string): boolean {\n  return /\\.[mc]?tsx?$/.test(filePath);\n}\n\n// ── Runtime resolution ─────────────────────────────────────────────\n\n/**\n * Resolve the AxlRuntime from a dynamically imported config module.\n *\n * Module shapes handled:\n * - ESM `export default runtime` → mod.default is the runtime\n * - CJS compiled from `export default runtime` → mod.default.default is the runtime\n * - CJS `module.exports = runtime` → mod.default is the runtime\n * - Named `export { runtime }` → mod.runtime is the runtime\n */\nexport function resolveRuntime(mod: Record<string, any>): unknown {\n  const def = mod.default as Record<string, any> | undefined;\n  return def?.default ?? def ?? mod.runtime;\n}\n\n// ── Module loading ────────────────────────────────────────────────\n\n// Lazily resolved tsImport function from tsx. `undefined` = not yet checked,\n// `null` = tsx not available.\nlet tsImportFn:\n  | ((specifier: string, parentURL: string) => Promise<Record<string, any>>)\n  | null\n  | undefined;\n\n/**\n * Import a module, using tsx's `tsImport()` for TypeScript files.\n *\n * `tsImport()` handles ESM/CJS format correctly without process-wide side effects —\n * no need for `register()` hooks or ESM-forcing workarounds. Falls back to regular\n * `import()` for non-TypeScript files or when tsx is not installed.\n */\nexport async function importModule(\n  filePath: string,\n  parentURL: string,\n): Promise<Record<string, any>> {\n  if (needsTsxLoader(filePath)) {\n    if (tsImportFn === undefined) {\n      try {\n        // @ts-expect-error — tsx is an optional runtime dependency\n        const mod = await import('tsx/esm/api');\n        tsImportFn = mod.tsImport ?? null;\n      } catch {\n        tsImportFn = null;\n        console.warn(\n          '[axl-eval] Warning: tsx is not installed. TypeScript files require tsx.\\n' +\n            '  Install it with: npm install -D tsx',\n        );\n      }\n    }\n    if (tsImportFn) {\n      return (await tsImportFn(pathToFileURL(filePath).href, parentURL)) as Record<string, any>;\n    }\n  }\n  return await import(pathToFileURL(filePath).href);\n}\n\n// ── Conditions ────────────────────────────────────────────────────\n\nexport async function registerConditions(conditions: string[]): Promise<void> {\n  try {\n    const nodeModule = await import('node:module');\n    const hookCode = [\n      `const extra = ${JSON.stringify(conditions)};`,\n      `export async function resolve(specifier, context, nextResolve) {`,\n      `  return nextResolve(specifier, {`,\n      `    ...context,`,\n      `    conditions: [...new Set([...context.conditions, ...extra])],`,\n      `  });`,\n      `}`,\n    ].join('\\n');\n    nodeModule.register(`data:text/javascript,${encodeURIComponent(hookCode)}`);\n  } catch {\n    console.warn('[axl-eval] Warning: --conditions requires Node.js 20.6+');\n  }\n}\n"],"mappings":";;;;;;;;;AAEA,SAAS,aAAa,gBAAgB;AACtC,SAAS,YAAY,eAAe,aAAa,gBAAgB,aAAa;AAC9E,YAAY,UAAU;;;ACItB,SAAS,eAAe;AACxB,SAAS,kBAAkB;AAC3B,SAAS,qBAAqB;AAIvB,IAAM,oBAAoB;AAAA,EAC/B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEO,SAAS,WAAW,KAAiC;AAC1D,aAAW,QAAQ,mBAAmB;AACpC,UAAM,IAAI,QAAQ,KAAK,IAAI;AAC3B,QAAI,WAAW,CAAC,EAAG,QAAO;AAAA,EAC5B;AACA,SAAO;AACT;AAKO,SAAS,eAAe,UAA2B;AACxD,SAAO,eAAe,KAAK,QAAQ;AACrC;AAaO,SAAS,eAAe,KAAmC;AAChE,QAAM,MAAM,IAAI;AAChB,SAAO,KAAK,WAAW,OAAO,IAAI;AACpC;AAMA,IAAI;AAYJ,eAAsB,aACpB,UACA,WAC8B;AAC9B,MAAI,eAAe,QAAQ,GAAG;AAC5B,QAAI,eAAe,QAAW;AAC5B,UAAI;AAEF,cAAM,MAAM,MAAM,OAAO,aAAa;AACtC,qBAAa,IAAI,YAAY;AAAA,MAC/B,QAAQ;AACN,qBAAa;AACb,gBAAQ;AAAA,UACN;AAAA,QAEF;AAAA,MACF;AAAA,IACF;AACA,QAAI,YAAY;AACd,aAAQ,MAAM,WAAW,cAAc,QAAQ,EAAE,MAAM,SAAS;AAAA,IAClE;AAAA,EACF;AACA,SAAO,MAAM,OAAO,cAAc,QAAQ,EAAE;AAC9C;AAIA,eAAsB,mBAAmB,YAAqC;AAC5E,MAAI;AACF,UAAM,aAAa,MAAM,OAAO,QAAa;AAC7C,UAAM,WAAW;AAAA,MACf,iBAAiB,KAAK,UAAU,UAAU,CAAC;AAAA,MAC3C;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AACX,eAAW,SAAS,wBAAwB,mBAAmB,QAAQ,CAAC,EAAE;AAAA,EAC5E,QAAQ;AACN,YAAQ,KAAK,yDAAyD;AAAA,EACxE;AACF;;;AD3FA,IAAM,cAAc,oBAAI,IAAI;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,eAAe,OAAO;AACpB,QAAM,OAAO,QAAQ,KAAK,MAAM,CAAC;AAEjC,MAAI,KAAK,WAAW,KAAK,KAAK,CAAC,MAAM,YAAY,KAAK,CAAC,MAAM,MAAM;AACjE,YAAQ,IAAI;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAiBZ,kBAAkB,KAAK,MAAM,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,CAKjC;AACG,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAEA,QAAM,eAAe,IAAI;AAC3B;AAEA,SAAS,kBAAkB,MAA6D;AACtF,QAAM,MAAM,KAAK,QAAQ,aAAa;AACtC,MAAI,QAAQ,MAAM,MAAM,KAAK,KAAK,OAAQ,QAAO;AACjD,QAAM,QAAQ,KAAK,MAAM,CAAC;AAG1B,MAAI,WAAW,KAAK,KAAK,EAAG,QAAO,WAAW,KAAK;AAGnD,QAAM,MAA8B,CAAC;AACrC,aAAW,QAAQ,MAAM,MAAM,GAAG,GAAG;AACnC,UAAM,CAAC,KAAK,GAAG,IAAI,KAAK,MAAM,GAAG;AACjC,QAAI,OAAO,OAAO,KAAM,KAAI,IAAI,KAAK,CAAC,IAAI,WAAW,IAAI,KAAK,CAAC;AAAA,EACjE;AACA,SAAO,OAAO,KAAK,GAAG,EAAE,SAAS,IAAI,MAAM;AAC7C;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,mBAAmB,KAAK,SAAS,sBAAsB;AAC7D,QAAM,aAAa,kBAAkB,IAAI;AACzC,QAAM,QAAQ,KAAK;AAAA,IACjB,CAAC,GAAG,MAAM,CAAC,EAAE,WAAW,IAAI,KAAK,EAAE,IAAI,KAAK,KAAK,IAAI,CAAC,MAAM;AAAA,EAC9D;AAEA,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ,MAAM,gFAAgF;AAC9F,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,cAAc,KAAK,MAAM,MAAM,cAAc,MAAM,CAAC,GAAG,OAAO,CAAC;AACrE,QAAM,eAAe,KAAK,MAAM,MAAM,cAAc,MAAM,CAAC,GAAG,OAAO,CAAC;AACtE,QAAM,WAAsC;AAC5C,QAAM,YAAuC;AAC7C,QAAM,iBAAiB,cAAc,OAAO,EAAE,WAAW,IAAI;AAC7D,QAAM,aAAa,YAAY,UAAU,WAAW,cAAc;AAElE,UAAQ;AAAA,IACN;AAAA,qBAAwB,WAAW,SAAS,GAAG,MAAM,GAAG,CAAC,CAAC,mBAAmB,WAAW,UAAU,GAAG,MAAM,GAAG,CAAC,CAAC;AAAA;AAAA,EAClH;AAEA,QAAM,cAAc,OAAO,KAAK,WAAW,OAAO;AAClD,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,CAAC;AAClE,QAAM,QAAQ,YAAY,KAAK,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,MAAM,IAAI;AAEtE,QAAM,WAAW,QAAQ,4BAA4B;AACrD,QAAM,SAAS,QAAQ,4BAA4B;AACnD,UAAQ,IAAI,KAAK,SAAS,OAAO,UAAU,CAAC,0CAA0C,QAAQ,EAAE;AAChG,UAAQ,IAAI,KAAK,GAAG,OAAO,YAAY,GAAG,CAAC,0CAA0C,MAAM,EAAE;AAE7F,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,WAAW,QAAQ,IAAI;AACjC,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,QAAI,OAAO,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,EAAE,aAAa,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,EAAE,cAAc,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,MAAM,OAAO,EAAE,MAAM,QAAQ,CAAC,GAAG,SAAS,CAAC,CAAC,MAAM,OAAO,EAAE,aAAa,QAAQ,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC;AACjO,QAAI,OAAO;AACT,UAAI,EAAE,IAAI;AACR,cAAM,MAAM,EAAE,GAAG,SAAS,IAAI,MAAM,MAAM,EAAE,GAAG,MAAM,QAAQ,CAAC;AAC9D,cAAM,MAAM,EAAE,GAAG,SAAS,IAAI,MAAM,MAAM,EAAE,GAAG,MAAM,QAAQ,CAAC;AAC9D,cAAM,MAAM,EAAE,cAAc,QAAQ;AACpC,gBAAQ,MAAM,EAAE,KAAK,EAAE,IAAI,GAAG;AAAA,MAChC,OAAO;AACL,gBAAQ,WAAM,OAAO,EAAE;AAAA,MACzB;AAAA,IACF;AACA,YAAQ,IAAI,IAAI;AAAA,EAClB;AAEA,MAAI,WAAW,QAAQ;AACrB,UAAM,IAAI,WAAW;AACrB,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,YAAQ;AAAA,MACN;AAAA,sBAAyB,EAAE,eAAe,KAAM,QAAQ,CAAC,CAAC,mBAAmB,EAAE,gBAAgB,KAAM,QAAQ,CAAC,CAAC,MAAM,IAAI,GAAG,EAAE,aAAa,QAAQ,CAAC,CAAC;AAAA,IACvJ;AAAA,EACF;AACA,MAAI,WAAW,MAAM;AACnB,UAAM,IAAI,WAAW;AACrB,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,YAAQ;AAAA,MACN,qBAAqB,EAAE,cAAc,QAAQ,CAAC,CAAC,kBAAkB,EAAE,eAAe,QAAQ,CAAC,CAAC,KAAK,IAAI,GAAG,EAAE,aAAa,QAAQ,CAAC,CAAC;AAAA,IACnI;AAAA,EACF;AAEA,QAAM,cAAc,MAAM,QAAQ,QAAQ,IAAI,SAAS,CAAC,IAAI;AAC5D,QAAM,SAAS,KAAK;AAAA,IAClB;AAAA,IACA,YAAY,MAAM,SAAS,WAAW,YAAY,SAAS,WAAW,aAAa;AAAA,EACrF;AACA,UAAQ;AAAA,IACN;AAAA,iBAAoB,WAAW,YAAY,MAAM,oBAAoB,WAAW,aAAa,MAAM,cAAc,MAAM;AAAA;AAAA,EACzH;AAEA,MAAI,oBAAoB,WAAW,YAAY,SAAS,GAAG;AAEzD,UAAM,kBAAkB,YAAY,KAAK,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,eAAe,IAAI;AACzF,QAAI,iBAAiB;AACnB,YAAM,2BAA2B,YAAY;AAAA,QAC3C,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,gBAAgB,QAAQ,WAAW,QAAQ,CAAC,EAAE,QAAQ;AAAA,MACrF;AACA,UAAI,yBAA0B,SAAQ,KAAK,CAAC;AAAA,IAC9C,OAAO;AACL,cAAQ,KAAK,CAAC;AAAA,IAChB;AAAA,EACF;AACF;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,EAAE,YAAY,WAAW,YAAY,MAAM,IAAI,cAAc,IAAI;AAEvE,MAAI,MAAM,SAAS,GAAG;AACpB,YAAQ,MAAM,sEAAsE;AACpF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,CAAC,aAAa,YAAY,IAAI;AACpC,QAAM,MAAM,KAAK,MAAM,MAAM,cAAc,aAAa,OAAO,CAAC;AAChE,QAAM,UAAwB,MAAM,QAAQ,GAAG,IAAI,MAAM,CAAC,GAAG;AAE7D,QAAM,UAAU,MAAM,WAAW,WAAW,UAAU;AACtD,MAAI;AACF,UAAM,MAAM,MAAM,aAAkB,aAAQ,YAAY,GAAG,YAAY,GAAG;AAC1E,UAAM,aAAyB,IAAI,SAAS,WAAW,IAAI,WAAW,IAAI,UAAU;AAEpF,QAAI,CAAC,WAAW,WAAW,WAAW,QAAQ,WAAW,GAAG;AAC1D,cAAQ,MAAM,UAAU,YAAY,0BAA0B;AAC9D,cAAQ,KAAK,CAAC;AAAA,IAChB;AAEA,UAAM,WAAyB,CAAC;AAChC,eAAW,cAAc,SAAS;AAChC,eAAS,KAAK,MAAM,QAAQ,YAAY,WAAW,SAAS,OAAO,CAAC;AAAA,IACtE;AAEA,eAAW,KAAK,UAAU;AACxB,cAAQ,IAAI,OAAO,YAAY,CAAC,IAAI,IAAI;AAAA,IAC1C;AAEA,QAAI,YAAY;AACd,YAAM,SAAS,MAAM,QAAQ,GAAG,IAAI,WAAW,SAAS,CAAC;AACzD,YAAM,YAAiB,aAAa,aAAQ,UAAU,CAAC;AACvD,YAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,YAAM,eAAe,YAAY,KAAK,UAAU,QAAQ,MAAM,CAAC,GAAG,OAAO;AACzE,cAAQ,IAAI,6BAA6B,UAAU,EAAE;AAAA,IACvD;AAAA,EACF,UAAE;AACA,UAAM,QAAQ,SAAS,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EACzC;AACF;AAEA,SAAS,iBAAiB,GAAqB;AAC7C,QAAM,WAAgB,aAAQ,CAAC;AAC/B,MAAI;AACF,UAAM,OAAO,SAAS,QAAQ;AAC9B,QAAI,KAAK,YAAY,GAAG;AACtB,YAAM,UAAU,YAAY,QAAQ;AACpC,aAAO,QACJ,OAAO,CAAC,MAAM,wBAAwB,KAAK,CAAC,CAAC,EAC7C,IAAI,CAAC,MAAW,UAAK,UAAU,CAAC,CAAC;AAAA,IACtC;AACA,WAAO,CAAC,QAAQ;AAAA,EAClB,QAAQ;AACN,WAAO,CAAC,QAAQ;AAAA,EAClB;AACF;AAGA,SAAS,gBAAgB,WAA4B;AACnD,MAAI,CAAC,MAAM,QAAQ,SAAS,KAAK,UAAU,WAAW,EAAG,QAAO;AAChE,SAAQ,UAAwB,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ,EAAE,KAAK,IAAI;AAC7F;AAEA,SAAS,YAAY,QAA4B;AAC/C,QAAM,QAAkB,CAAC;AACzB,QAAM,cAAc,OAAO,KAAK,OAAO,QAAQ,OAAO;AACtD,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,MAAM;AAChF,QAAM,WAAW;AAEjB,QAAM;AAAA,IACJ,SAAS,gBAAgB,OAAO,SAAS,SAAS,CAAC,MAAM,OAAO,OAAO,KAAK,OAAO,QAAQ,KAAK;AAAA,EAClG;AACA,QAAM;AAAA,IACJ,KAAK,SAAS,OAAO,UAAU,CAAC,KAAK,OAAO,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC;AAAA,EACpL;AAEA,QAAM,UAAU,aAAa,KAAK,WAAW,KAAK;AAClD,QAAM,KAAK,OAAO,SAAS,OAAO,OAAO,CAAC;AAE1C,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,OAAO,QAAQ,QAAQ,IAAI;AAErC,UAAM,kBAAkB,OAAO,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,SAAS,EAAE,OAAO,IAAI,KAAK,IAAI,EAAE;AACvF,QAAI,oBAAoB,GAAG;AACzB,YAAM;AAAA,QACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC;AAAA,MAC1K;AAAA,IACF,OAAO;AACL,YAAM;AAAA,QACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,EAAE,KAAK,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC;AAAA,MACvO;AAAA,IACF;AAAA,EACF;AAEA,MAAI,OAAO,QAAQ,QAAQ;AACzB,UAAM,IAAI,OAAO,QAAQ;AACzB,UAAM;AAAA,MACJ,KAAK,SAAS,OAAO,UAAU,CAAC,MAAM,EAAE,OAAO,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,MAAM,EAAE,MAAM,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,MAAM,EAAE,MAAM,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC;AAAA,IACtL;AAAA,EACF;AAEA,QAAM,eAAe,OAAO,WAAW,KAAM,QAAQ,CAAC;AACtD,QAAM,UAAU,OAAO,YAAY,IAAI,IAAI,OAAO,UAAU,QAAQ,CAAC,CAAC,KAAK;AAC3E,QAAM,KAAK,EAAE;AACb,QAAM;AAAA,IACJ,eAAe,OAAO,QAAQ,QAAQ,IAAI,OAAO,QAAQ,KAAK,YAAY,OAAO,gBAAgB,WAAW;AAAA,EAC9G;AAEA,QAAM,kBAAkB,OAAO,MAAM,OAAO,CAAC,MAAM,EAAE,cAAc,MAAM;AACzE,MAAI,gBAAgB,SAAS,GAAG;AAC9B,UAAM,eAAe,CAAC,GAAG,IAAI,IAAI,gBAAgB,QAAQ,CAAC,MAAM,EAAE,YAAa,CAAC,CAAC;AACjF,UAAM,KAAK,EAAE;AACb,UAAM;AAAA,MACJ,oBAAoB,gBAAgB,MAAM,IAAI,OAAO,QAAQ,KAAK;AAAA,IACpE;AACA,eAAW,OAAO,aAAa,MAAM,GAAG,CAAC,GAAG;AAC1C,YAAM,KAAK,SAAS,GAAG,EAAE;AAAA,IAC3B;AACA,QAAI,aAAa,SAAS,GAAG;AAC3B,YAAM,KAAK,eAAe,aAAa,SAAS,CAAC,OAAO;AAAA,IAC1D;AAAA,EACF;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,oBAAoB,SAAkC;AAC7D,QAAM,QAAkB,CAAC;AACzB,QAAM,cAAc,OAAO,KAAK,QAAQ,OAAO;AAC/C,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,MAAM;AAChF,QAAM,WAAW;AAEjB,QAAM;AAAA,IACJ,SAAS,gBAAgB,QAAQ,SAAS,CAAC,MAAM,QAAQ,OAAO,WAAW,QAAQ,QAAQ;AAAA,EAC7F;AACA,QAAM;AAAA,IACJ,KAAK,SAAS,OAAO,UAAU,CAAC,KAAK,gBAAkB,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,CAAC,CAAC,KAAK,MAAM,SAAS,CAAC,CAAC;AAAA,EACvH;AACA,QAAM,UAAU,aAAa,IAAI,WAAW,IAAI,IAAI,IAAI;AACxD,QAAM,KAAK,OAAO,SAAS,OAAO,OAAO,CAAC;AAE1C,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,QAAQ,QAAQ,IAAI;AAC9B,UAAM,UAAU,GAAG,EAAE,KAAK,QAAQ,CAAC,CAAC,SAAW,EAAE,IAAI,QAAQ,CAAC,CAAC;AAC/D,UAAM;AAAA,MACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,QAAQ,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC;AAAA,IAC/H;AAAA,EACF;AAEA,MAAI,QAAQ,QAAQ;AAClB,UAAM;AAAA,MACJ,KAAK,SAAS,OAAO,UAAU,CAAC,OAAO,QAAQ,OAAO,OAAO,KAAM,QAAQ,CAAC,IAAI,YAAc,QAAQ,OAAO,MAAM,KAAM,QAAQ,CAAC,IAAI,KAAK,SAAS,QAAQ,CAAC;AAAA,IAC/J;AAAA,EACF;AAEA,QAAM,UAAU,QAAQ,YAAY,IAAI,IAAI,QAAQ,UAAU,QAAQ,CAAC,CAAC,KAAK;AAC7E,QAAM,eAAe,QAAQ,gBAAgB,KAAM,QAAQ,CAAC;AAC5D,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,iBAAiB,OAAO,sBAAsB,WAAW,GAAG;AAEvE,SAAO,MAAM,KAAK,IAAI;AACxB;AAIA,eAAe,yBAAyB,YAAyC;AAC/E,MAAI;AACF,UAAM,MAAM,MAAM,aAAa,YAAY,YAAY,GAAG;AAC1D,UAAM,UAAU,eAAe,GAAG;AAElC,QAAI,CAAC,WAAW,OAAO,QAAQ,YAAY,YAAY;AACrD,cAAQ,MAAM,mDAAmD;AACjE,UAAI,SAAS;AACX,cAAM,OAAO,OAAO,KAAK,OAAiB,EACvC,MAAM,GAAG,CAAC,EACV,KAAK,IAAI;AACZ,gBAAQ,MAAM,UAAU,OAAO,OAAO,GAAG,OAAO,iBAAiB,IAAI,OAAO,EAAE,EAAE;AAAA,MAClF;AACA,cAAQ;AAAA,QACN;AAAA;AAAA;AAAA,MACF;AACA,cAAQ,KAAK,CAAC;AAAA,IAChB;AAEA,WAAO;AAAA,EACT,SAAS,KAAK;AACZ,UAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC3D,QACE,8FAA8F;AAAA,MAC5F;AAAA,IACF,GACA;AACA,YAAM,MAAW,aAAQ,UAAU;AACnC,cAAQ,MAAM,wEAAwE;AACtF,UAAI,QAAQ,SAAS,QAAQ,QAAQ;AACnC,gBAAQ;AAAA,UACN;AAAA,QACF;AAAA,MACF,OAAO;AACL,gBAAQ,MAAM,mDAAmD;AAAA,MACnE;AACA,cAAQ,MAAM;AAAA,IAChB;AACA,YAAQ,MAAM,0BAA0B,GAAG;AAC3C,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAEA,eAAe,WAAW,WAAoB,YAA4C;AAExF,MAAI,cAAc,WAAW,SAAS,GAAG;AACvC,UAAM,mBAAmB,UAAU;AAAA,EACrC;AAGA,MAAI,WAAW;AACb,UAAM,aAAkB,aAAQ,QAAQ,IAAI,GAAG,SAAS;AACxD,UAAM,OAAO,SAAS,YAAY,EAAE,gBAAgB,MAAM,CAAC;AAC3D,QAAI,CAAC,MAAM,OAAO,GAAG;AACnB,cAAQ,MAAM,0BAA0B,UAAU,EAAE;AACpD,cAAQ,KAAK,CAAC;AAAA,IAChB;AACA,YAAQ,MAAM,kCAAkC,UAAU,EAAE;AAC5D,WAAO,yBAAyB,UAAU;AAAA,EAC5C;AAGA,QAAM,QAAQ,WAAW,QAAQ,IAAI,CAAC;AACtC,MAAI,OAAO;AACT,YAAQ,MAAM,oCAAoC,KAAK,EAAE;AACzD,WAAO,yBAAyB,KAAK;AAAA,EACvC;AAGA,QAAM,EAAE,WAAW,IAAI,MAAM,OAAO,aAAa;AACjD,SAAO,IAAI,WAAW;AACxB;AAIA,SAAS,cAAc,MAOrB;AACA,MAAI;AACJ,MAAI;AACJ,MAAI,aAAuB,CAAC;AAC5B,MAAI,OAAO;AACX,MAAI,gBAAgB;AACpB,QAAM,QAAkB,CAAC;AAEzB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAClB,QAAI,QAAQ,cAAc,QAAQ,cAAc,QAAQ,kBAAkB,QAAQ,UAAU;AAC1F,UAAI,IAAI,KAAK,KAAK,QAAQ;AACxB,gBAAQ,MAAM,UAAU,GAAG,mBAAmB;AAC9C,gBAAQ,KAAK,CAAC;AAAA,MAChB;AACA,YAAM,QAAQ,KAAK,EAAE,CAAC;AACtB,UAAI,QAAQ,WAAY,cAAa;AAAA,eAC5B,QAAQ,WAAY,aAAY;AAAA,eAChC,QAAQ,SAAU,QAAO,KAAK,IAAI,GAAG,SAAS,OAAO,EAAE,KAAK,CAAC;AAAA;AAEpE,qBAAa,MACV,MAAM,GAAG,EACT,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,OAAO;AAAA,IACrB,WAAW,QAAQ,oBAAoB;AAErC,sBAAgB;AAAA,IAClB,WAAW,IAAI,WAAW,IAAI,GAAG;AAC/B,UAAI,CAAC,YAAY,IAAI,GAAG,GAAG;AACzB,gBAAQ,MAAM,iBAAiB,GAAG,EAAE;AACpC,gBAAQ,KAAK,CAAC;AAAA,MAChB;AAAA,IACF,OAAO;AACL,YAAM,KAAK,GAAG;AAAA,IAChB;AAAA,EACF;AAEA,SAAO,EAAE,YAAY,WAAW,YAAY,MAAM,eAAe,MAAM;AACzE;AAIA,eAAe,eAAe,MAAgB;AAC5C,QAAM,EAAE,YAAY,WAAW,YAAY,MAAM,eAAe,MAAM,IAAI,cAAc,IAAI;AAE5F,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ,MAAM,mCAAmC;AACjD,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,YAAsB,CAAC;AAC7B,aAAW,KAAK,OAAO;AACrB,cAAU,KAAK,GAAG,iBAAiB,CAAC,CAAC;AAAA,EACvC;AAEA,MAAI,UAAU,WAAW,GAAG;AAC1B,YAAQ,MAAM,4BAA4B;AAC1C,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,UAAU,MAAM,WAAW,WAAW,UAAU;AACtD,QAAM,UAAwB,CAAC;AAE/B,MAAI;AACF,eAAW,YAAY,WAAW;AAChC,UAAI;AACF,cAAM,MAAM,MAAM,aAAkB,aAAQ,QAAQ,GAAG,YAAY,GAAG;AACtE,cAAM,aAAyB,IAAI,SAAS,WAAW,IAAI,WAAW,IAAI,UAAU;AAEpF,YAAI,CAAC,WAAW,YAAY,CAAC,WAAW,WAAW,CAAC,WAAW,SAAS;AACtE,kBAAQ;AAAA,YACN,UAAU,QAAQ;AAAA,UACpB;AACA;AAAA,QACF;AAGA,YAAI;AAKJ,YAAI,IAAI,iBAAiB;AAEvB,4BAAkB,OAAO,OAAO,OAAO;AACrC,kBAAM;AAAA,cACJ;AAAA,cACA,MAAM;AAAA,cACN;AAAA,YACF,IAAI,MAAM,QAAQ,eAAe,YAAY;AAC3C,qBAAO,IAAI,gBAAgB,OAAO,EAAE;AAAA,YACtC,CAAC;AACD,mBAAO;AAAA,cACL,QAAQ,OAAO;AAAA,cACf,MAAM,OAAO,QAAQ;AAAA,cACrB,UAAU,OAAO,YAAY;AAAA,YAC/B;AAAA,UACF;AAAA,QACF,WAAW,QAAQ,YAAY,WAAW,QAAQ,GAAG;AAEnD,4BAAkB,OAAO,UAAU;AACjC,kBAAM,EAAE,QAAQ,MAAM,SAAS,IAAI,MAAM,QAAQ,eAAe,YAAY;AAC1E,qBAAO,QAAQ,QAAQ,WAAW,UAAU,KAAK;AAAA,YACnD,CAAC;AACD,mBAAO,EAAE,QAAQ,QAAQ,MAAM,SAAS;AAAA,UAC1C;AAAA,QACF,OAAO;AACL,kBAAQ;AAAA,YACN,uBAAuB,QAAQ;AAAA,UACjC;AACA,4BAAkB,OAAO,WAAW,EAAE,QAAQ,MAAM;AAAA,QACtD;AASA,cAAM,aAAa,gBAAgB,EAAE,eAAe,KAAK,IAAI;AAE7D,YAAI,OAAO,GAAG;AAEZ,gBAAM,EAAE,WAAW,IAAI,MAAM,OAAO,QAAa;AACjD,gBAAM,aAAa,WAAW;AAC9B,gBAAM,aAA2B,CAAC;AAElC,mBAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,oBAAQ,MAAM,kBAAkB,IAAI,CAAC,IAAI,IAAI,KAAK;AAClD,kBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,SAAS,UAAU;AAC7E,mBAAO,SAAS,aAAa;AAC7B,mBAAO,SAAS,WAAW;AAC3B,uBAAW,KAAK,MAAM;AACtB,oBAAQ,KAAK,MAAM;AAAA,UACrB;AAEA,gBAAM,UAAU,cAAc,UAAU;AACxC,kBAAQ,IAAI,OAAO,oBAAoB,OAAO,IAAI,IAAI;AAAA,QACxD,OAAO;AACL,gBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,SAAS,UAAU;AAC7E,kBAAQ,KAAK,MAAM;AAEnB,kBAAQ,IAAI,OAAO,YAAY,MAAM,IAAI,IAAI;AAAA,QAC/C;AAAA,MACF,SAAS,KAAK;AACZ,gBAAQ;AAAA,UACN,sBAAsB,QAAQ,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,QACrF;AAAA,MACF;AAAA,IACF;AAEA,QAAI,cAAc,QAAQ,SAAS,GAAG;AACpC,YAAM,SAAS,QAAQ,WAAW,IAAI,QAAQ,CAAC,IAAI;AACnD,YAAM,YAAiB,aAAa,aAAQ,UAAU,CAAC;AACvD,YAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,YAAM,eAAe,YAAY,KAAK,UAAU,QAAQ,MAAM,CAAC,GAAG,OAAO;AACzE,cAAQ,IAAI,oBAAoB,UAAU,EAAE;AAAA,IAC9C;AAAA,EACF,UAAE;AACA,UAAM,QAAQ,SAAS,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EACzC;AAGA,MAAI,QAAQ,WAAW,GAAG;AACxB,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAEA,KAAK,EAAE,MAAM,CAAC,QAAQ;AACpB,UAAQ,MAAM,GAAG;AACjB,UAAQ,KAAK,CAAC;AAChB,CAAC;","names":[]}
+{"version":3,"sources":["../src/cli.ts","../src/cli-utils.ts","../src/cli-validate.ts"],"sourcesContent":["#!/usr/bin/env node\n\nimport { readdirSync, statSync } from 'node:fs';\nimport { readFile as readFileAsync, writeFile as writeFileAsync, mkdir } from 'node:fs/promises';\nimport * as path from 'node:path';\nimport type { AxlRuntime, EvalExecuteWorkflow } from '@axlsdk/axl';\nimport { evalCompare } from './compare.js';\nimport { runEval } from './runner.js';\nimport { rescore } from './rescore.js';\nimport { aggregateRuns } from './multi-run.js';\nimport type { MultiRunSummary } from './multi-run.js';\nimport type { EvalConfig, EvalResult } from './types.js';\nimport {\n  findConfig,\n  resolveRuntime,\n  importModule,\n  registerConditions,\n  pickDefault,\n  pickExport,\n  expandGlob,\n  CONFIG_CANDIDATES,\n} from './cli-utils.js';\nimport { validateEvalConfig } from './cli-validate.js';\n\nconst KNOWN_FLAGS = new Set([\n  '--output',\n  '--config',\n  '--conditions',\n  '--fail-on-regression',\n  '--threshold',\n  '--runs',\n  '--capture-traces',\n]);\n\nasync function main() {\n  const args = process.argv.slice(2);\n\n  if (args.length === 0 || args[0] === '--help' || args[0] === '-h') {\n    console.log(`\nUsage:\n  axl-eval <path>                         Run eval file(s)\n  axl-eval <path> --runs <n>              Run eval N times (multi-run)\n  axl-eval <path> --output <file>         Save results to JSON\n  axl-eval <path> --config <file>         Use config file for runtime\n  axl-eval <path> --conditions <list>     Node.js import conditions (comma-separated)\n  axl-eval <path> --capture-traces        Populate EvalItem.traces on every item\n                                          (success + failure). Adds memory\n                                          overhead proportional to dataset size\n                                          x turns x agents; off by default.\n  axl-eval rescore <results> <eval-file>  Re-run scorers on saved outputs\n  axl-eval compare <a> <b>                Compare two eval result files\n  axl-eval compare <a> <b> --threshold <v>  Set regression threshold (global or per-scorer)\n  axl-eval compare <a> <b> --fail-on-regression  Exit 1 if regressions\n\nConfig auto-detection (when --config is not specified):\n  ${CONFIG_CANDIDATES.join(' -> ')}\n\nWhen a config is found, the exported AxlRuntime is passed to executeWorkflow\nand cost is tracked automatically via runtime.trackCost().\nWhen no config is found, a bare AxlRuntime is created (providers from env vars).\n`);\n    process.exit(0);\n  }\n\n  if (args[0] === 'compare') {\n    await runCompare(args.slice(1));\n    return;\n  }\n\n  if (args[0] === 'rescore') {\n    await runRescore(args.slice(1));\n    return;\n  }\n\n  await runEvalCommand(args);\n}\n\nfunction parseThresholdArg(args: string[]): Record<string, number> | number | undefined {\n  const idx = args.indexOf('--threshold');\n  if (idx === -1 || idx + 1 >= args.length) return undefined;\n  const value = args[idx + 1];\n\n  // Global numeric threshold: --threshold 0.05\n  if (/^[\\d.]+$/.test(value)) return parseFloat(value);\n\n  // Per-scorer map: --threshold accuracy=0,tone=0.1\n  const map: Record<string, number> = {};\n  for (const pair of value.split(',')) {\n    const [key, val] = pair.split('=');\n    if (key && val != null) map[key.trim()] = parseFloat(val.trim());\n  }\n  return Object.keys(map).length > 0 ? map : undefined;\n}\n\nasync function runCompare(args: string[]) {\n  const failOnRegression = args.includes('--fail-on-regression');\n  const thresholds = parseThresholdArg(args);\n  const files = args.filter(\n    (a, i) => !a.startsWith('--') && !(i > 0 && args[i - 1] === '--threshold'),\n  );\n\n  if (files.length !== 2) {\n    console.error('Usage: axl-eval compare <baseline.json> <candidate.json> [--threshold <value>]');\n    process.exit(1);\n  }\n\n  const baselineRaw = JSON.parse(await readFileAsync(files[0], 'utf-8'));\n  const candidateRaw = JSON.parse(await readFileAsync(files[1], 'utf-8'));\n  const baseline: EvalResult | EvalResult[] = baselineRaw;\n  const candidate: EvalResult | EvalResult[] = candidateRaw;\n  const compareOptions = thresholds != null ? { thresholds } : undefined;\n  const comparison = evalCompare(baseline, candidate, compareOptions);\n\n  console.log(\n    `\\nCompare: baseline (${comparison.baseline.id.slice(0, 8)}) -> candidate (${comparison.candidate.id.slice(0, 8)})\\n`,\n  );\n\n  const scorerNames = Object.keys(comparison.scorers);\n  const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 6);\n  const hasCI = scorerNames.some((n) => comparison.scorers[n].ci != null);\n\n  const ciHeader = hasCI ? '  CI 95%            Sig' : '';\n  const ciRule = hasCI ? '  ----------------  ---' : '';\n  console.log(`  ${'Scorer'.padEnd(maxNameLen)}  Baseline  Candidate  Delta     Change${ciHeader}`);\n  console.log(`  ${''.padEnd(maxNameLen, '-')}  --------  ---------  --------  ------${ciRule}`);\n\n  for (const name of scorerNames) {\n    const s = comparison.scorers[name];\n    const sign = s.delta > 0 ? '+' : '';\n    let line = `  ${name.padEnd(maxNameLen)}  ${s.baselineMean.toFixed(3).padStart(8)}  ${s.candidateMean.toFixed(3).padStart(9)}  ${(sign + s.delta.toFixed(3)).padStart(8)}  ${(sign + s.deltaPercent.toFixed(1) + '%').padStart(6)}`;\n    if (hasCI) {\n      if (s.ci) {\n        const lo = (s.ci.lower >= 0 ? '+' : '') + s.ci.lower.toFixed(4);\n        const hi = (s.ci.upper >= 0 ? '+' : '') + s.ci.upper.toFixed(4);\n        const sig = s.significant ? '  *' : '   ';\n        line += `  [${lo}, ${hi}]${sig}`;\n      } else {\n        line += '  —'.padEnd(22);\n      }\n    }\n    console.log(line);\n  }\n\n  if (comparison.timing) {\n    const t = comparison.timing;\n    const sign = t.delta > 0 ? '+' : '';\n    console.log(\n      `\\n  Timing: baseline ${(t.baselineMean / 1000).toFixed(2)}s -> candidate ${(t.candidateMean / 1000).toFixed(2)}s (${sign}${t.deltaPercent.toFixed(1)}%)`,\n    );\n  }\n  if (comparison.cost) {\n    const c = comparison.cost;\n    const sign = c.delta > 0 ? '+' : '';\n    console.log(\n      `  Cost: baseline $${c.baselineTotal.toFixed(2)} -> candidate $${c.candidateTotal.toFixed(2)} (${sign}${c.deltaPercent.toFixed(1)}%)`,\n    );\n  }\n\n  const baselineRef = Array.isArray(baseline) ? baseline[0] : baseline;\n  const stable = Math.max(\n    0,\n    baselineRef.items.length - comparison.regressions.length - comparison.improvements.length,\n  );\n  console.log(\n    `\\n  Regressions: ${comparison.regressions.length} | Improvements: ${comparison.improvements.length} | Stable: ${stable}\\n`,\n  );\n\n  if (failOnRegression && comparison.regressions.length > 0) {\n    // When CI is available, only fail on significant regressions\n    const hasSignificance = scorerNames.some((n) => comparison.scorers[n].significant != null);\n    if (hasSignificance) {\n      const hasSignificantRegression = scorerNames.some(\n        (n) => comparison.scorers[n].significant === true && comparison.scorers[n].delta < 0,\n      );\n      if (hasSignificantRegression) process.exit(1);\n    } else {\n      process.exit(1);\n    }\n  }\n}\n\nasync function runRescore(args: string[]) {\n  const { outputPath, configArg, conditions, paths } = parseEvalArgs(args);\n\n  if (paths.length < 2) {\n    console.error('Usage: axl-eval rescore <results.json> <eval-file> [--output <file>]');\n    process.exit(1);\n  }\n\n  const [resultsPath, evalFilePath] = paths;\n  const raw = JSON.parse(await readFileAsync(resultsPath, 'utf-8'));\n  const results: EvalResult[] = Array.isArray(raw) ? raw : [raw];\n\n  const runtime = await getRuntime(configArg, conditions);\n  try {\n    const mod = await importModule(path.resolve(evalFilePath), import.meta.url);\n    const evalConfig = pickDefault<EvalConfig>(mod);\n\n    if (!evalConfig || typeof evalConfig !== 'object') {\n      console.error(\n        `Error: ${evalFilePath} does not export an eval config object — got ${evalConfig === null ? 'null' : typeof evalConfig}.`,\n      );\n      process.exit(1);\n    }\n    if (!Array.isArray(evalConfig.scorers) || evalConfig.scorers.length === 0) {\n      console.error(`Error: ${evalFilePath} does not export a non-empty scorers array.`);\n      process.exit(1);\n    }\n\n    const rescored: EvalResult[] = [];\n    for (const resultData of results) {\n      rescored.push(await rescore(resultData, evalConfig.scorers, runtime));\n    }\n\n    for (const r of rescored) {\n      console.log('\\n' + formatTable(r) + '\\n');\n    }\n\n    if (outputPath) {\n      const output = Array.isArray(raw) ? rescored : rescored[0];\n      const outputDir = path.dirname(path.resolve(outputPath));\n      await mkdir(outputDir, { recursive: true });\n      await writeFileAsync(outputPath, JSON.stringify(output, null, 2), 'utf-8');\n      console.log(`Rescored results saved to ${outputPath}`);\n    }\n  } finally {\n    await runtime.shutdown().catch(() => {});\n  }\n}\n\n/**\n * Resolve a CLI path argument into a list of eval files.\n *\n * Accepts:\n * - An explicit file path (`path/to/foo.eval.ts`)\n * - A directory (lists `*.eval.[mc]?[jt]sx?` inside, non-recursive)\n * - A glob pattern: `evals/*.eval.ts`, `evals/**\\/*.eval.ts`, `**\\/*.eval.ts`\n *\n * Glob expansion mirrors the studio eval-loader's so users get the same\n * behavior in both places. Without it, users running on Windows or with\n * quoted patterns hit confusing \"file not found\" errors when the shell\n * couldn't expand the glob.\n */\nfunction collectEvalFiles(p: string): string[] {\n  if (p.includes('*')) return expandGlob(p, process.cwd());\n  const resolved = path.resolve(p);\n  try {\n    const stat = statSync(resolved);\n    if (stat.isDirectory()) {\n      const entries = readdirSync(resolved);\n      return entries\n        .filter((e) => /\\.eval\\.[mc]?[jt]sx?$/.test(e))\n        .map((e) => path.join(resolved, e));\n    }\n    return [resolved];\n  } catch {\n    return [resolved];\n  }\n}\n\n/** Render observed workflows for a terse header line. Multiple → comma-separated. */\nfunction formatWorkflows(workflows: unknown): string {\n  if (!Array.isArray(workflows) || workflows.length === 0) return '(unknown)';\n  return (workflows as unknown[]).filter((w): w is string => typeof w === 'string').join(', ');\n}\n\nfunction formatTable(result: EvalResult): string {\n  const lines: string[] = [];\n  const scorerNames = Object.keys(result.summary.scorers);\n  const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 'Scorer'.length);\n  const colWidth = 8;\n\n  lines.push(\n    `Eval: ${formatWorkflows(result.metadata.workflows)} x ${result.dataset} (${result.summary.count} items)`,\n  );\n  lines.push(\n    `  ${'Scorer'.padEnd(maxNameLen)}  ${'Mean'.padStart(colWidth)}  ${'Min'.padStart(colWidth)}  ${'Max'.padStart(colWidth)}  ${'p50'.padStart(colWidth)}  ${'p95'.padStart(colWidth)}`,\n  );\n\n  const ruleLen = maxNameLen + 2 + (colWidth + 2) * 5;\n  lines.push('  ' + '\\u2500'.repeat(ruleLen));\n\n  for (const name of scorerNames) {\n    const s = result.summary.scorers[name];\n    // Detect scorers with no valid scores (all items errored → computeStats([]) → all zeros)\n    const validScoreCount = result.items.filter((i) => !i.error && i.scores[name] != null).length;\n    if (validScoreCount === 0) {\n      lines.push(\n        `  ${name.padEnd(maxNameLen)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}`,\n      );\n    } else {\n      lines.push(\n        `  ${name.padEnd(maxNameLen)}  ${s.mean.toFixed(2).padStart(colWidth)}  ${s.min.toFixed(2).padStart(colWidth)}  ${s.max.toFixed(2).padStart(colWidth)}  ${s.p50.toFixed(2).padStart(colWidth)}  ${s.p95.toFixed(2).padStart(colWidth)}`,\n      );\n    }\n  }\n\n  if (result.summary.timing) {\n    const t = result.summary.timing;\n    lines.push(\n      `  ${'Timing'.padEnd(maxNameLen)}  ${(t.mean / 1000).toFixed(2).padStart(colWidth)}s ${(t.p50 / 1000).toFixed(2).padStart(colWidth)}s ${(t.p95 / 1000).toFixed(2).padStart(colWidth)}s`,\n    );\n  }\n\n  const durationSec = (result.duration / 1000).toFixed(1);\n  const costStr = result.totalCost > 0 ? `$${result.totalCost.toFixed(2)}` : '$0.00';\n  lines.push('');\n  lines.push(\n    `  Failures: ${result.summary.failures}/${result.summary.count} | Cost: ${costStr} | Duration: ${durationSec}s`,\n  );\n\n  const itemsWithErrors = result.items.filter((i) => i.scorerErrors?.length);\n  if (itemsWithErrors.length > 0) {\n    const uniqueErrors = [...new Set(itemsWithErrors.flatMap((i) => i.scorerErrors!))];\n    lines.push('');\n    lines.push(\n      `  Scorer errors (${itemsWithErrors.length}/${result.summary.count} items affected):`,\n    );\n    for (const err of uniqueErrors.slice(0, 5)) {\n      lines.push(`    - ${err}`);\n    }\n    if (uniqueErrors.length > 5) {\n      lines.push(`    ... and ${uniqueErrors.length - 5} more`);\n    }\n  }\n\n  return lines.join('\\n');\n}\n\nfunction formatMultiRunTable(summary: MultiRunSummary): string {\n  const lines: string[] = [];\n  const scorerNames = Object.keys(summary.scorers);\n  const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 'Scorer'.length);\n  const colWidth = 16;\n\n  lines.push(\n    `Eval: ${formatWorkflows(summary.workflows)} x ${summary.dataset} \\u2014 ${summary.runCount} runs`,\n  );\n  lines.push(\n    `  ${'Scorer'.padEnd(maxNameLen)}  ${'Mean \\u00b1 Std'.padStart(colWidth)}  ${'Min'.padStart(8)}  ${'Max'.padStart(8)}`,\n  );\n  const ruleLen = maxNameLen + 2 + colWidth + 2 + 8 + 2 + 8;\n  lines.push('  ' + '\\u2500'.repeat(ruleLen));\n\n  for (const name of scorerNames) {\n    const s = summary.scorers[name];\n    const meanStd = `${s.mean.toFixed(3)} \\u00b1 ${s.std.toFixed(3)}`;\n    lines.push(\n      `  ${name.padEnd(maxNameLen)}  ${meanStd.padStart(colWidth)}  ${s.min.toFixed(3).padStart(8)}  ${s.max.toFixed(3).padStart(8)}`,\n    );\n  }\n\n  if (summary.timing) {\n    lines.push(\n      `  ${'Timing'.padEnd(maxNameLen)}  ${((summary.timing.mean / 1000).toFixed(2) + ' \\u00b1 ' + (summary.timing.std / 1000).toFixed(2) + 's').padStart(colWidth)}`,\n    );\n  }\n\n  const costStr = summary.totalCost > 0 ? `$${summary.totalCost.toFixed(2)}` : '$0.00';\n  const durationStr = (summary.totalDuration / 1000).toFixed(1);\n  lines.push('');\n  lines.push(`  Total Cost: ${costStr} | Total Duration: ${durationStr}s`);\n\n  return lines.join('\\n');\n}\n\n// ── Runtime resolution ─────────────────────────────────────────────\n\nasync function resolveRuntimeFromConfig(configPath: string): Promise<AxlRuntime> {\n  try {\n    const mod = await importModule(configPath, import.meta.url);\n    const runtime = resolveRuntime(mod) as AxlRuntime;\n\n    if (!runtime || typeof runtime.execute !== 'function') {\n      console.error(`Config must export a default AxlRuntime instance.`);\n      if (runtime) {\n        const keys = Object.keys(runtime as object)\n          .slice(0, 5)\n          .join(', ');\n        console.error(`  Got: ${typeof runtime}${keys ? ` with keys: { ${keys} }` : ''}`);\n      }\n      console.error(\n        `Example:\\n  import { AxlRuntime } from '@axlsdk/axl';\\n  export default new AxlRuntime({ ... });`,\n      );\n      process.exit(1);\n    }\n\n    return runtime;\n  } catch (err) {\n    const msg = err instanceof Error ? err.message : String(err);\n    if (\n      /Cannot use import statement|Unexpected reserved word|top-level await|exports is not defined/.test(\n        msg,\n      )\n    ) {\n      const ext = path.extname(configPath);\n      console.error(`[axl-eval] Config failed to load due to a CJS/ESM compatibility issue.`);\n      if (ext === '.ts' || ext === '.tsx') {\n        console.error(\n          `  Tip: try renaming to .mts to force ESM format, or ensure tsx is installed and up to date.`,\n        );\n      } else {\n        console.error(`  Tip: add \"type\": \"module\" to your package.json.`);\n      }\n      console.error();\n    }\n    console.error(`Failed to load config:`, err);\n    process.exit(1);\n  }\n}\n\nasync function getRuntime(configArg?: string, conditions?: string[]): Promise<AxlRuntime> {\n  // Register import conditions before any config loading\n  if (conditions && conditions.length > 0) {\n    await registerConditions(conditions);\n  }\n\n  // 1. Explicit --config\n  if (configArg) {\n    const configPath = path.resolve(process.cwd(), configArg);\n    const stat = statSync(configPath, { throwIfNoEntry: false });\n    if (!stat?.isFile()) {\n      console.error(`Config file not found: ${configPath}`);\n      process.exit(1);\n    }\n    console.error(`[axl-eval] Loading config from ${configPath}`);\n    return resolveRuntimeFromConfig(configPath);\n  }\n\n  // 2. Auto-detect\n  const found = findConfig(process.cwd());\n  if (found) {\n    console.error(`[axl-eval] Auto-detected config: ${found}`);\n    return resolveRuntimeFromConfig(found);\n  }\n\n  // 3. Bare runtime (providers from env vars)\n  const { AxlRuntime } = await import('@axlsdk/axl');\n  return new AxlRuntime();\n}\n\n// ── Arg parsing ────────────────────────────────────────────────────\n\nfunction parseEvalArgs(args: string[]): {\n  outputPath?: string;\n  configArg?: string;\n  conditions: string[];\n  runs: number;\n  captureTraces: boolean;\n  paths: string[];\n} {\n  let outputPath: string | undefined;\n  let configArg: string | undefined;\n  let conditions: string[] = [];\n  let runs = 1;\n  let captureTraces = false;\n  const paths: string[] = [];\n\n  for (let i = 0; i < args.length; i++) {\n    const arg = args[i];\n    if (arg === '--output' || arg === '--config' || arg === '--conditions' || arg === '--runs') {\n      if (i + 1 >= args.length) {\n        console.error(`Error: ${arg} requires a value`);\n        process.exit(1);\n      }\n      const value = args[++i];\n      if (arg === '--output') outputPath = value;\n      else if (arg === '--config') configArg = value;\n      else if (arg === '--runs') runs = Math.max(1, parseInt(value, 10) || 1);\n      else\n        conditions = value\n          .split(',')\n          .map((c) => c.trim())\n          .filter(Boolean);\n    } else if (arg === '--capture-traces') {\n      // Boolean flag — no value consumed.\n      captureTraces = true;\n    } else if (arg.startsWith('--')) {\n      if (!KNOWN_FLAGS.has(arg)) {\n        console.error(`Unknown flag: ${arg}`);\n        process.exit(1);\n      }\n    } else {\n      paths.push(arg);\n    }\n  }\n\n  return { outputPath, configArg, conditions, runs, captureTraces, paths };\n}\n\n// ── Main eval command ──────────────────────────────────────────────\n\nasync function runEvalCommand(args: string[]) {\n  const { outputPath, configArg, conditions, runs, captureTraces, paths } = parseEvalArgs(args);\n\n  if (paths.length === 0) {\n    console.error('Error: No eval file path provided');\n    process.exit(1);\n  }\n\n  const evalFiles: string[] = [];\n  for (const p of paths) {\n    evalFiles.push(...collectEvalFiles(p));\n  }\n\n  if (evalFiles.length === 0) {\n    console.error('Error: No eval files found');\n    process.exit(1);\n  }\n\n  const runtime = await getRuntime(configArg, conditions);\n  const results: EvalResult[] = [];\n  let failedFiles = 0;\n\n  try {\n    for (const filePath of evalFiles) {\n      try {\n        const mod = await importModule(path.resolve(filePath), import.meta.url);\n        const evalConfig = pickDefault<EvalConfig>(mod);\n\n        const validationError = validateEvalConfig(evalConfig);\n        if (validationError) {\n          console.error(`Error: ${filePath} ${validationError}`);\n          failedFiles++;\n          continue;\n        }\n\n        // Resolve executeWorkflow: custom export > registered workflow > error\n        // Walk the ESM/CJS interop chain symmetrically with the default export\n        // so named exports stay visible when tsx loads the eval module as CJS.\n        const exported = pickExport<unknown>(mod, 'executeWorkflow');\n\n        // Validate that the export, if present, is callable. Catches the\n        // `export const executeWorkflow = 'foo'` typo at the boundary instead\n        // of crashing deep inside trackExecution with a confusing stack.\n        let customExecute: EvalExecuteWorkflow | undefined;\n        if (typeof exported === 'function') {\n          customExecute = exported as EvalExecuteWorkflow;\n        } else if (exported !== undefined) {\n          console.error(\n            `[axl-eval] Error: ${filePath} exports executeWorkflow but it is ${typeof exported}, not a function.`,\n          );\n          failedFiles++;\n          continue;\n        }\n\n        let executeWorkflow: EvalExecuteWorkflow;\n\n        if (customExecute) {\n          // Wrap custom executeWorkflow with trackExecution for cost + metadata attribution\n          executeWorkflow = async (input, rt) => {\n            const {\n              result,\n              cost: trackedCost,\n              metadata,\n            } = await runtime.trackExecution(async () => {\n              return customExecute(input, rt);\n            });\n            return {\n              output: result.output,\n              cost: result.cost ?? trackedCost,\n              metadata: result.metadata ?? metadata,\n            };\n          };\n        } else if (runtime.getWorkflow(evalConfig.workflow)) {\n          // No executeWorkflow exported but workflow is registered — use runtime.execute()\n          executeWorkflow = async (input) => {\n            const { result, cost, metadata } = await runtime.trackExecution(async () => {\n              return runtime.execute(evalConfig.workflow, input);\n            });\n            return { output: result, cost, metadata };\n          };\n        } else {\n          // Fail loudly. The previous identity-passthrough fallback silently\n          // produced all-zero scores in CI — exactly the kind of footgun the\n          // \"fail loudly, not silently\" rule is meant to prevent. If you\n          // genuinely want identity, export it explicitly:\n          //   export const executeWorkflow = async (input) => ({ output: input });\n          //\n          // Surface what WAS exported (top-level + .default) so users\n          // recognize a typo like `execute_workflow` and can spot the\n          // CJS-interop case that motivates this whole code path: when tsx\n          // loads a `.ts` file as CJS, all named exports land on\n          // `mod.default.<name>`. The chain walk above already handled this,\n          // but a user who removed `executeWorkflow` will see exactly what\n          // export names ARE present.\n          const exportShape = (() => {\n            const top = Object.keys(mod).filter((k) => k !== 'default');\n            const fromDefault =\n              mod.default && typeof mod.default === 'object'\n                ? Object.keys(mod.default as Record<string, unknown>)\n                : [];\n            const all = [...new Set([...top, ...fromDefault])];\n            return all.length > 0 ? `Found exports: [${all.join(', ')}].` : '';\n          })();\n          console.error(\n            `[axl-eval] Error: ${filePath} does not export an executeWorkflow function ` +\n              `and no workflow named \"${evalConfig.workflow}\" is registered on the runtime.\\n` +\n              (exportShape ? `  ${exportShape}\\n` : '') +\n              `  Add either:\\n` +\n              `    export async function executeWorkflow(input, runtime) { ... }\\n` +\n              `  or register the workflow on your runtime via runtime.registerWorkflow(workflow).\\n` +\n              `  Tip: if your eval previously worked, your package.json may be missing\\n` +\n              `    \"type\": \"module\" — tsx loads .ts as CJS, putting named exports on\\n` +\n              `    mod.default.<name>. The CLI handles that interop, but if you renamed\\n` +\n              `    or removed the export the file path is your fix.`,\n          );\n          failedFiles++;\n          continue;\n        }\n\n        // When --capture-traces is set, forward it to runEval so the runner\n        // wraps each item's execution in a nested trackExecution({ captureTraces: true }).\n        // This populates EvalItem.traces on both success (via tracked.traces)\n        // and failure (via the axlCapturedTraces side-channel on the thrown\n        // error). Nested trackExecution scopes walk the AsyncLocalStorage\n        // parent chain, so the outer trackExecution above still observes\n        // cost/metadata correctly.\n        const runOptions = captureTraces ? { captureTraces: true } : undefined;\n\n        if (runs > 1) {\n          // Multi-run mode. Buffer per-run results locally so we can mark the\n          // batch correctly before committing to outer `results[]`. The\n          // original bug was the opposite: results pushed eagerly per-run, a\n          // throw mid-batch left partial results posing as a complete run\n          // (no `fromPartialBatch` marker, aggregate never ran). The fix ISN'T\n          // to throw the partials away — those runs cost money and have\n          // statistical signal. Instead we preserve them with explicit\n          // partial-batch metadata, aggregate over what completed, and exit\n          // non-zero so CI knows the batch wasn't clean.\n          const { randomUUID } = await import('node:crypto');\n          const runGroupId = randomUUID();\n          const runResults: EvalResult[] = [];\n          let runFailure: Error | undefined;\n\n          for (let r = 0; r < runs; r++) {\n            console.error(`[axl-eval] Run ${r + 1}/${runs}...`);\n            try {\n              const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);\n              result.metadata.runGroupId = runGroupId;\n              result.metadata.runIndex = r;\n              runResults.push(result);\n            } catch (err) {\n              runFailure = err instanceof Error ? err : new Error(String(err));\n              console.error(`[axl-eval] Run ${r + 1}/${runs} failed: ${runFailure.message}`);\n              // Stop attempting further runs — we don't know whether this is\n              // transient (rate limit) or permanent (auth failure), and\n              // burning more API calls speculatively is hostile. Preserve\n              // what completed.\n              break;\n            }\n          }\n\n          if (runResults.length === 0) {\n            // Nothing completed — surface as a per-file failure. No partial\n            // artifact to write; the outer try/catch error message would\n            // duplicate the \"Run 1/N failed\" line we already printed, so\n            // increment and continue without re-throwing.\n            console.error(\n              `[axl-eval] Aborting ${filePath}: no runs completed (${runFailure?.message ?? 'unknown failure'}).`,\n            );\n            failedFiles++;\n            continue;\n          }\n\n          const partial = runResults.length < runs;\n          if (partial) {\n            // Mark every completed run so any consumer reading\n            // `result.metadata` sees the partial-batch context. `aggregateRuns`\n            // computes statistics over `runResults.length`, which is the\n            // honest sample size to report.\n            //\n            // The flag is `fromPartialBatch` — a per-run marker meaning\n            // \"this run came from a batch that did not complete.\" It is\n            // NOT named `partialBatch` because that read as \"this run is\n            // partial,\" and a defensive consumer wrote\n            // `if (run.metadata.partialBatch) skipThisRun()`, throwing\n            // away successful work that cost real money. The flag describes\n            // the BATCH the run belongs to, not the run itself.\n            //\n            // `runFailure?.message` can be empty (e.g. `new Error('')` from a\n            // misbehaving provider); fall back to `String(runFailure)` so the\n            // banner never renders a blank \"Stopped after:\" line. Skip the\n            // field entirely if both are empty so `buildMultiRunResult`'s\n            // empty-string filter doesn't have to defensively guess.\n            const failureMsg = runFailure\n              ? runFailure.message || String(runFailure) || undefined\n              : undefined;\n            for (const r of runResults) {\n              r.metadata.fromPartialBatch = true;\n              r.metadata.batchCompleted = runResults.length;\n              r.metadata.batchAttempted = runs;\n              if (failureMsg) r.metadata.batchFailure = failureMsg;\n            }\n            console.error(\n              `[axl-eval] PARTIAL: ${runResults.length} of ${runs} runs completed for ${filePath}.`,\n            );\n            failedFiles++;\n          }\n\n          const summary = aggregateRuns(runResults);\n          console.log('\\n' + formatMultiRunTable(summary) + '\\n');\n          for (const r of runResults) results.push(r);\n        } else {\n          const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);\n          results.push(result);\n\n          console.log('\\n' + formatTable(result) + '\\n');\n        }\n      } catch (err) {\n        console.error(\n          `Error running eval ${filePath}: ${err instanceof Error ? err.message : String(err)}`,\n        );\n        failedFiles++;\n      }\n    }\n\n    if (outputPath && results.length > 0) {\n      const output = results.length === 1 ? results[0] : results;\n      const outputDir = path.dirname(path.resolve(outputPath));\n      await mkdir(outputDir, { recursive: true });\n      await writeFileAsync(outputPath, JSON.stringify(output, null, 2), 'utf-8');\n      console.log(`Results saved to ${outputPath}`);\n    }\n  } finally {\n    await runtime.shutdown().catch(() => {});\n  }\n\n  // Exit non-zero if any eval file failed to load/run. Previously we only\n  // exited non-zero when zero evals succeeded, which silently masked\n  // misconfigured files in mixed-success runs (a CI footgun).\n  if (failedFiles > 0 || results.length === 0) {\n    process.exit(1);\n  }\n}\n\nmain().catch((err) => {\n  console.error(err);\n  process.exit(1);\n});\n","/**\n * CLI utilities for `axl-eval`.\n *\n * The actual implementations of the loader, glob, and config-detection\n * helpers live in `@axlsdk/axl`'s `cli-internals` so a fix to (e.g.) the\n * tsx loader registration applies to both axl-eval and axl-studio\n * atomically. This file is now a thin re-export plus the symmetric\n * module-resolution helpers from `@axlsdk/axl`.\n */\n\nexport {\n  resolveRuntime,\n  pickDefault,\n  pickExport,\n  CONFIG_CANDIDATES,\n  findConfig,\n  needsTsxLoader,\n  importModule,\n  expandGlob,\n  registerConditions,\n} from '@axlsdk/axl';\n","import type { EvalConfig } from './types.js';\n\n/**\n * Validate an eval config's required fields and produce a diagnostic string\n * if invalid. Returns `undefined` when the config passes.\n *\n * Returns a single-line message tail (caller prepends the file path) so the\n * full error reads like: `Error: <file> <message>`.\n *\n * Goes beyond a falsy check to:\n *   - catch `scorers: []` which previously silently ran with no scorers\n *   - reject scorers arrays that contain non-callable entries (e.g. typo\n *     `scorers: [null, undefined]` would crash deep inside `runEval`)\n *   - reject non-string `workflow` (downstream `runtime.getWorkflow(...)`\n *     would silently miss it)\n *   - reject non-object `dataset` (must be a `dataset()` factory result)\n *   - provide a `Got: { keys: [...] }` hint so users can spot typos like\n *     `scorerS` vs `scorers` immediately\n *\n * Lives in its own module so unit tests can import it without triggering\n * `cli.ts`'s `main()` at module load time.\n */\nexport function validateEvalConfig(cfg: unknown): string | undefined {\n  if (!cfg || typeof cfg !== 'object') {\n    return `does not export a valid eval config — got ${cfg === null ? 'null' : typeof cfg}.`;\n  }\n  const c = cfg as Partial<EvalConfig>;\n  const missing: string[] = [];\n  if (!c.workflow) missing.push('workflow');\n  if (!c.dataset) missing.push('dataset');\n  if (!c.scorers) missing.push('scorers');\n\n  if (missing.length > 0) {\n    const keys = Object.keys(c as object).slice(0, 10);\n    const got = keys.length > 0 ? ` Got: { ${keys.join(', ')} }.` : '';\n    return `does not export a valid eval config (missing ${missing.join(', ')}).${got}`;\n  }\n\n  if (typeof c.workflow !== 'string') {\n    return `exports a non-string workflow (got ${typeof c.workflow}) — workflow must be the registered workflow name.`;\n  }\n  if (typeof c.dataset !== 'object' || Array.isArray(c.dataset)) {\n    return `exports a non-object dataset (got ${Array.isArray(c.dataset) ? 'array' : typeof c.dataset}) — use the dataset() factory.`;\n  }\n  if (typeof (c.dataset as { getItems?: unknown }).getItems !== 'function') {\n    return `exports a dataset without a getItems() method — use the dataset() factory.`;\n  }\n  if (!Array.isArray(c.scorers) || c.scorers.length === 0) {\n    return `exports an empty scorers array — at least one scorer is required.`;\n  }\n  const badScorerIdx = c.scorers.findIndex(\n    (s) => !s || typeof (s as { score?: unknown }).score !== 'function',\n  );\n  if (badScorerIdx !== -1) {\n    return `exports a scorers array with a non-scorer entry at index ${badScorerIdx} — each scorer must have a score() method (use the scorer() factory).`;\n  }\n  return undefined;\n}\n"],"mappings":";;;;;;;;;AAEA,SAAS,aAAa,gBAAgB;AACtC,SAAS,YAAY,eAAe,aAAa,gBAAgB,aAAa;AAC9E,YAAY,UAAU;;;ACMtB;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;;;ACEA,SAAS,mBAAmB,KAAkC;AACnE,MAAI,CAAC,OAAO,OAAO,QAAQ,UAAU;AACnC,WAAO,kDAA6C,QAAQ,OAAO,SAAS,OAAO,GAAG;AAAA,EACxF;AACA,QAAM,IAAI;AACV,QAAM,UAAoB,CAAC;AAC3B,MAAI,CAAC,EAAE,SAAU,SAAQ,KAAK,UAAU;AACxC,MAAI,CAAC,EAAE,QAAS,SAAQ,KAAK,SAAS;AACtC,MAAI,CAAC,EAAE,QAAS,SAAQ,KAAK,SAAS;AAEtC,MAAI,QAAQ,SAAS,GAAG;AACtB,UAAM,OAAO,OAAO,KAAK,CAAW,EAAE,MAAM,GAAG,EAAE;AACjD,UAAM,MAAM,KAAK,SAAS,IAAI,WAAW,KAAK,KAAK,IAAI,CAAC,QAAQ;AAChE,WAAO,gDAAgD,QAAQ,KAAK,IAAI,CAAC,KAAK,GAAG;AAAA,EACnF;AAEA,MAAI,OAAO,EAAE,aAAa,UAAU;AAClC,WAAO,sCAAsC,OAAO,EAAE,QAAQ;AAAA,EAChE;AACA,MAAI,OAAO,EAAE,YAAY,YAAY,MAAM,QAAQ,EAAE,OAAO,GAAG;AAC7D,WAAO,qCAAqC,MAAM,QAAQ,EAAE,OAAO,IAAI,UAAU,OAAO,EAAE,OAAO;AAAA,EACnG;AACA,MAAI,OAAQ,EAAE,QAAmC,aAAa,YAAY;AACxE,WAAO;AAAA,EACT;AACA,MAAI,CAAC,MAAM,QAAQ,EAAE,OAAO,KAAK,EAAE,QAAQ,WAAW,GAAG;AACvD,WAAO;AAAA,EACT;AACA,QAAM,eAAe,EAAE,QAAQ;AAAA,IAC7B,CAAC,MAAM,CAAC,KAAK,OAAQ,EAA0B,UAAU;AAAA,EAC3D;AACA,MAAI,iBAAiB,IAAI;AACvB,WAAO,4DAA4D,YAAY;AAAA,EACjF;AACA,SAAO;AACT;;;AFjCA,IAAM,cAAc,oBAAI,IAAI;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,eAAe,OAAO;AACpB,QAAM,OAAO,QAAQ,KAAK,MAAM,CAAC;AAEjC,MAAI,KAAK,WAAW,KAAK,KAAK,CAAC,MAAM,YAAY,KAAK,CAAC,MAAM,MAAM;AACjE,YAAQ,IAAI;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAiBZ,kBAAkB,KAAK,MAAM,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,CAKjC;AACG,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAEA,QAAM,eAAe,IAAI;AAC3B;AAEA,SAAS,kBAAkB,MAA6D;AACtF,QAAM,MAAM,KAAK,QAAQ,aAAa;AACtC,MAAI,QAAQ,MAAM,MAAM,KAAK,KAAK,OAAQ,QAAO;AACjD,QAAM,QAAQ,KAAK,MAAM,CAAC;AAG1B,MAAI,WAAW,KAAK,KAAK,EAAG,QAAO,WAAW,KAAK;AAGnD,QAAM,MAA8B,CAAC;AACrC,aAAW,QAAQ,MAAM,MAAM,GAAG,GAAG;AACnC,UAAM,CAAC,KAAK,GAAG,IAAI,KAAK,MAAM,GAAG;AACjC,QAAI,OAAO,OAAO,KAAM,KAAI,IAAI,KAAK,CAAC,IAAI,WAAW,IAAI,KAAK,CAAC;AAAA,EACjE;AACA,SAAO,OAAO,KAAK,GAAG,EAAE,SAAS,IAAI,MAAM;AAC7C;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,mBAAmB,KAAK,SAAS,sBAAsB;AAC7D,QAAM,aAAa,kBAAkB,IAAI;AACzC,QAAM,QAAQ,KAAK;AAAA,IACjB,CAAC,GAAG,MAAM,CAAC,EAAE,WAAW,IAAI,KAAK,EAAE,IAAI,KAAK,KAAK,IAAI,CAAC,MAAM;AAAA,EAC9D;AAEA,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ,MAAM,gFAAgF;AAC9F,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,cAAc,KAAK,MAAM,MAAM,cAAc,MAAM,CAAC,GAAG,OAAO,CAAC;AACrE,QAAM,eAAe,KAAK,MAAM,MAAM,cAAc,MAAM,CAAC,GAAG,OAAO,CAAC;AACtE,QAAM,WAAsC;AAC5C,QAAM,YAAuC;AAC7C,QAAM,iBAAiB,cAAc,OAAO,EAAE,WAAW,IAAI;AAC7D,QAAM,aAAa,YAAY,UAAU,WAAW,cAAc;AAElE,UAAQ;AAAA,IACN;AAAA,qBAAwB,WAAW,SAAS,GAAG,MAAM,GAAG,CAAC,CAAC,mBAAmB,WAAW,UAAU,GAAG,MAAM,GAAG,CAAC,CAAC;AAAA;AAAA,EAClH;AAEA,QAAM,cAAc,OAAO,KAAK,WAAW,OAAO;AAClD,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,CAAC;AAClE,QAAM,QAAQ,YAAY,KAAK,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,MAAM,IAAI;AAEtE,QAAM,WAAW,QAAQ,4BAA4B;AACrD,QAAM,SAAS,QAAQ,4BAA4B;AACnD,UAAQ,IAAI,KAAK,SAAS,OAAO,UAAU,CAAC,0CAA0C,QAAQ,EAAE;AAChG,UAAQ,IAAI,KAAK,GAAG,OAAO,YAAY,GAAG,CAAC,0CAA0C,MAAM,EAAE;AAE7F,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,WAAW,QAAQ,IAAI;AACjC,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,QAAI,OAAO,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,EAAE,aAAa,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,EAAE,cAAc,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,MAAM,OAAO,EAAE,MAAM,QAAQ,CAAC,GAAG,SAAS,CAAC,CAAC,MAAM,OAAO,EAAE,aAAa,QAAQ,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC;AACjO,QAAI,OAAO;AACT,UAAI,EAAE,IAAI;AACR,cAAM,MAAM,EAAE,GAAG,SAAS,IAAI,MAAM,MAAM,EAAE,GAAG,MAAM,QAAQ,CAAC;AAC9D,cAAM,MAAM,EAAE,GAAG,SAAS,IAAI,MAAM,MAAM,EAAE,GAAG,MAAM,QAAQ,CAAC;AAC9D,cAAM,MAAM,EAAE,cAAc,QAAQ;AACpC,gBAAQ,MAAM,EAAE,KAAK,EAAE,IAAI,GAAG;AAAA,MAChC,OAAO;AACL,gBAAQ,WAAM,OAAO,EAAE;AAAA,MACzB;AAAA,IACF;AACA,YAAQ,IAAI,IAAI;AAAA,EAClB;AAEA,MAAI,WAAW,QAAQ;AACrB,UAAM,IAAI,WAAW;AACrB,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,YAAQ;AAAA,MACN;AAAA,sBAAyB,EAAE,eAAe,KAAM,QAAQ,CAAC,CAAC,mBAAmB,EAAE,gBAAgB,KAAM,QAAQ,CAAC,CAAC,MAAM,IAAI,GAAG,EAAE,aAAa,QAAQ,CAAC,CAAC;AAAA,IACvJ;AAAA,EACF;AACA,MAAI,WAAW,MAAM;AACnB,UAAM,IAAI,WAAW;AACrB,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,YAAQ;AAAA,MACN,qBAAqB,EAAE,cAAc,QAAQ,CAAC,CAAC,kBAAkB,EAAE,eAAe,QAAQ,CAAC,CAAC,KAAK,IAAI,GAAG,EAAE,aAAa,QAAQ,CAAC,CAAC;AAAA,IACnI;AAAA,EACF;AAEA,QAAM,cAAc,MAAM,QAAQ,QAAQ,IAAI,SAAS,CAAC,IAAI;AAC5D,QAAM,SAAS,KAAK;AAAA,IAClB;AAAA,IACA,YAAY,MAAM,SAAS,WAAW,YAAY,SAAS,WAAW,aAAa;AAAA,EACrF;AACA,UAAQ;AAAA,IACN;AAAA,iBAAoB,WAAW,YAAY,MAAM,oBAAoB,WAAW,aAAa,MAAM,cAAc,MAAM;AAAA;AAAA,EACzH;AAEA,MAAI,oBAAoB,WAAW,YAAY,SAAS,GAAG;AAEzD,UAAM,kBAAkB,YAAY,KAAK,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,eAAe,IAAI;AACzF,QAAI,iBAAiB;AACnB,YAAM,2BAA2B,YAAY;AAAA,QAC3C,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,gBAAgB,QAAQ,WAAW,QAAQ,CAAC,EAAE,QAAQ;AAAA,MACrF;AACA,UAAI,yBAA0B,SAAQ,KAAK,CAAC;AAAA,IAC9C,OAAO;AACL,cAAQ,KAAK,CAAC;AAAA,IAChB;AAAA,EACF;AACF;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,EAAE,YAAY,WAAW,YAAY,MAAM,IAAI,cAAc,IAAI;AAEvE,MAAI,MAAM,SAAS,GAAG;AACpB,YAAQ,MAAM,sEAAsE;AACpF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,CAAC,aAAa,YAAY,IAAI;AACpC,QAAM,MAAM,KAAK,MAAM,MAAM,cAAc,aAAa,OAAO,CAAC;AAChE,QAAM,UAAwB,MAAM,QAAQ,GAAG,IAAI,MAAM,CAAC,GAAG;AAE7D,QAAM,UAAU,MAAM,WAAW,WAAW,UAAU;AACtD,MAAI;AACF,UAAM,MAAM,MAAM,aAAkB,aAAQ,YAAY,GAAG,YAAY,GAAG;AAC1E,UAAM,aAAa,YAAwB,GAAG;AAE9C,QAAI,CAAC,cAAc,OAAO,eAAe,UAAU;AACjD,cAAQ;AAAA,QACN,UAAU,YAAY,qDAAgD,eAAe,OAAO,SAAS,OAAO,UAAU;AAAA,MACxH;AACA,cAAQ,KAAK,CAAC;AAAA,IAChB;AACA,QAAI,CAAC,MAAM,QAAQ,WAAW,OAAO,KAAK,WAAW,QAAQ,WAAW,GAAG;AACzE,cAAQ,MAAM,UAAU,YAAY,6CAA6C;AACjF,cAAQ,KAAK,CAAC;AAAA,IAChB;AAEA,UAAM,WAAyB,CAAC;AAChC,eAAW,cAAc,SAAS;AAChC,eAAS,KAAK,MAAM,QAAQ,YAAY,WAAW,SAAS,OAAO,CAAC;AAAA,IACtE;AAEA,eAAW,KAAK,UAAU;AACxB,cAAQ,IAAI,OAAO,YAAY,CAAC,IAAI,IAAI;AAAA,IAC1C;AAEA,QAAI,YAAY;AACd,YAAM,SAAS,MAAM,QAAQ,GAAG,IAAI,WAAW,SAAS,CAAC;AACzD,YAAM,YAAiB,aAAa,aAAQ,UAAU,CAAC;AACvD,YAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,YAAM,eAAe,YAAY,KAAK,UAAU,QAAQ,MAAM,CAAC,GAAG,OAAO;AACzE,cAAQ,IAAI,6BAA6B,UAAU,EAAE;AAAA,IACvD;AAAA,EACF,UAAE;AACA,UAAM,QAAQ,SAAS,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EACzC;AACF;AAeA,SAAS,iBAAiB,GAAqB;AAC7C,MAAI,EAAE,SAAS,GAAG,EAAG,QAAO,WAAW,GAAG,QAAQ,IAAI,CAAC;AACvD,QAAM,WAAgB,aAAQ,CAAC;AAC/B,MAAI;AACF,UAAM,OAAO,SAAS,QAAQ;AAC9B,QAAI,KAAK,YAAY,GAAG;AACtB,YAAM,UAAU,YAAY,QAAQ;AACpC,aAAO,QACJ,OAAO,CAAC,MAAM,wBAAwB,KAAK,CAAC,CAAC,EAC7C,IAAI,CAAC,MAAW,UAAK,UAAU,CAAC,CAAC;AAAA,IACtC;AACA,WAAO,CAAC,QAAQ;AAAA,EAClB,QAAQ;AACN,WAAO,CAAC,QAAQ;AAAA,EAClB;AACF;AAGA,SAAS,gBAAgB,WAA4B;AACnD,MAAI,CAAC,MAAM,QAAQ,SAAS,KAAK,UAAU,WAAW,EAAG,QAAO;AAChE,SAAQ,UAAwB,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ,EAAE,KAAK,IAAI;AAC7F;AAEA,SAAS,YAAY,QAA4B;AAC/C,QAAM,QAAkB,CAAC;AACzB,QAAM,cAAc,OAAO,KAAK,OAAO,QAAQ,OAAO;AACtD,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,MAAM;AAChF,QAAM,WAAW;AAEjB,QAAM;AAAA,IACJ,SAAS,gBAAgB,OAAO,SAAS,SAAS,CAAC,MAAM,OAAO,OAAO,KAAK,OAAO,QAAQ,KAAK;AAAA,EAClG;AACA,QAAM;AAAA,IACJ,KAAK,SAAS,OAAO,UAAU,CAAC,KAAK,OAAO,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC;AAAA,EACpL;AAEA,QAAM,UAAU,aAAa,KAAK,WAAW,KAAK;AAClD,QAAM,KAAK,OAAO,SAAS,OAAO,OAAO,CAAC;AAE1C,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,OAAO,QAAQ,QAAQ,IAAI;AAErC,UAAM,kBAAkB,OAAO,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,SAAS,EAAE,OAAO,IAAI,KAAK,IAAI,EAAE;AACvF,QAAI,oBAAoB,GAAG;AACzB,YAAM;AAAA,QACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC;AAAA,MAC1K;AAAA,IACF,OAAO;AACL,YAAM;AAAA,QACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,EAAE,KAAK,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC;AAAA,MACvO;AAAA,IACF;AAAA,EACF;AAEA,MAAI,OAAO,QAAQ,QAAQ;AACzB,UAAM,IAAI,OAAO,QAAQ;AACzB,UAAM;AAAA,MACJ,KAAK,SAAS,OAAO,UAAU,CAAC,MAAM,EAAE,OAAO,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,MAAM,EAAE,MAAM,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,MAAM,EAAE,MAAM,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC;AAAA,IACtL;AAAA,EACF;AAEA,QAAM,eAAe,OAAO,WAAW,KAAM,QAAQ,CAAC;AACtD,QAAM,UAAU,OAAO,YAAY,IAAI,IAAI,OAAO,UAAU,QAAQ,CAAC,CAAC,KAAK;AAC3E,QAAM,KAAK,EAAE;AACb,QAAM;AAAA,IACJ,eAAe,OAAO,QAAQ,QAAQ,IAAI,OAAO,QAAQ,KAAK,YAAY,OAAO,gBAAgB,WAAW;AAAA,EAC9G;AAEA,QAAM,kBAAkB,OAAO,MAAM,OAAO,CAAC,MAAM,EAAE,cAAc,MAAM;AACzE,MAAI,gBAAgB,SAAS,GAAG;AAC9B,UAAM,eAAe,CAAC,GAAG,IAAI,IAAI,gBAAgB,QAAQ,CAAC,MAAM,EAAE,YAAa,CAAC,CAAC;AACjF,UAAM,KAAK,EAAE;AACb,UAAM;AAAA,MACJ,oBAAoB,gBAAgB,MAAM,IAAI,OAAO,QAAQ,KAAK;AAAA,IACpE;AACA,eAAW,OAAO,aAAa,MAAM,GAAG,CAAC,GAAG;AAC1C,YAAM,KAAK,SAAS,GAAG,EAAE;AAAA,IAC3B;AACA,QAAI,aAAa,SAAS,GAAG;AAC3B,YAAM,KAAK,eAAe,aAAa,SAAS,CAAC,OAAO;AAAA,IAC1D;AAAA,EACF;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,oBAAoB,SAAkC;AAC7D,QAAM,QAAkB,CAAC;AACzB,QAAM,cAAc,OAAO,KAAK,QAAQ,OAAO;AAC/C,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,MAAM;AAChF,QAAM,WAAW;AAEjB,QAAM;AAAA,IACJ,SAAS,gBAAgB,QAAQ,SAAS,CAAC,MAAM,QAAQ,OAAO,WAAW,QAAQ,QAAQ;AAAA,EAC7F;AACA,QAAM;AAAA,IACJ,KAAK,SAAS,OAAO,UAAU,CAAC,KAAK,gBAAkB,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,CAAC,CAAC,KAAK,MAAM,SAAS,CAAC,CAAC;AAAA,EACvH;AACA,QAAM,UAAU,aAAa,IAAI,WAAW,IAAI,IAAI,IAAI;AACxD,QAAM,KAAK,OAAO,SAAS,OAAO,OAAO,CAAC;AAE1C,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,QAAQ,QAAQ,IAAI;AAC9B,UAAM,UAAU,GAAG,EAAE,KAAK,QAAQ,CAAC,CAAC,SAAW,EAAE,IAAI,QAAQ,CAAC,CAAC;AAC/D,UAAM;AAAA,MACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,QAAQ,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC;AAAA,IAC/H;AAAA,EACF;AAEA,MAAI,QAAQ,QAAQ;AAClB,UAAM;AAAA,MACJ,KAAK,SAAS,OAAO,UAAU,CAAC,OAAO,QAAQ,OAAO,OAAO,KAAM,QAAQ,CAAC,IAAI,YAAc,QAAQ,OAAO,MAAM,KAAM,QAAQ,CAAC,IAAI,KAAK,SAAS,QAAQ,CAAC;AAAA,IAC/J;AAAA,EACF;AAEA,QAAM,UAAU,QAAQ,YAAY,IAAI,IAAI,QAAQ,UAAU,QAAQ,CAAC,CAAC,KAAK;AAC7E,QAAM,eAAe,QAAQ,gBAAgB,KAAM,QAAQ,CAAC;AAC5D,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,iBAAiB,OAAO,sBAAsB,WAAW,GAAG;AAEvE,SAAO,MAAM,KAAK,IAAI;AACxB;AAIA,eAAe,yBAAyB,YAAyC;AAC/E,MAAI;AACF,UAAM,MAAM,MAAM,aAAa,YAAY,YAAY,GAAG;AAC1D,UAAM,UAAU,eAAe,GAAG;AAElC,QAAI,CAAC,WAAW,OAAO,QAAQ,YAAY,YAAY;AACrD,cAAQ,MAAM,mDAAmD;AACjE,UAAI,SAAS;AACX,cAAM,OAAO,OAAO,KAAK,OAAiB,EACvC,MAAM,GAAG,CAAC,EACV,KAAK,IAAI;AACZ,gBAAQ,MAAM,UAAU,OAAO,OAAO,GAAG,OAAO,iBAAiB,IAAI,OAAO,EAAE,EAAE;AAAA,MAClF;AACA,cAAQ;AAAA,QACN;AAAA;AAAA;AAAA,MACF;AACA,cAAQ,KAAK,CAAC;AAAA,IAChB;AAEA,WAAO;AAAA,EACT,SAAS,KAAK;AACZ,UAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC3D,QACE,8FAA8F;AAAA,MAC5F;AAAA,IACF,GACA;AACA,YAAM,MAAW,aAAQ,UAAU;AACnC,cAAQ,MAAM,wEAAwE;AACtF,UAAI,QAAQ,SAAS,QAAQ,QAAQ;AACnC,gBAAQ;AAAA,UACN;AAAA,QACF;AAAA,MACF,OAAO;AACL,gBAAQ,MAAM,mDAAmD;AAAA,MACnE;AACA,cAAQ,MAAM;AAAA,IAChB;AACA,YAAQ,MAAM,0BAA0B,GAAG;AAC3C,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAEA,eAAe,WAAW,WAAoB,YAA4C;AAExF,MAAI,cAAc,WAAW,SAAS,GAAG;AACvC,UAAM,mBAAmB,UAAU;AAAA,EACrC;AAGA,MAAI,WAAW;AACb,UAAM,aAAkB,aAAQ,QAAQ,IAAI,GAAG,SAAS;AACxD,UAAM,OAAO,SAAS,YAAY,EAAE,gBAAgB,MAAM,CAAC;AAC3D,QAAI,CAAC,MAAM,OAAO,GAAG;AACnB,cAAQ,MAAM,0BAA0B,UAAU,EAAE;AACpD,cAAQ,KAAK,CAAC;AAAA,IAChB;AACA,YAAQ,MAAM,kCAAkC,UAAU,EAAE;AAC5D,WAAO,yBAAyB,UAAU;AAAA,EAC5C;AAGA,QAAM,QAAQ,WAAW,QAAQ,IAAI,CAAC;AACtC,MAAI,OAAO;AACT,YAAQ,MAAM,oCAAoC,KAAK,EAAE;AACzD,WAAO,yBAAyB,KAAK;AAAA,EACvC;AAGA,QAAM,EAAE,WAAW,IAAI,MAAM,OAAO,aAAa;AACjD,SAAO,IAAI,WAAW;AACxB;AAIA,SAAS,cAAc,MAOrB;AACA,MAAI;AACJ,MAAI;AACJ,MAAI,aAAuB,CAAC;AAC5B,MAAI,OAAO;AACX,MAAI,gBAAgB;AACpB,QAAM,QAAkB,CAAC;AAEzB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAClB,QAAI,QAAQ,cAAc,QAAQ,cAAc,QAAQ,kBAAkB,QAAQ,UAAU;AAC1F,UAAI,IAAI,KAAK,KAAK,QAAQ;AACxB,gBAAQ,MAAM,UAAU,GAAG,mBAAmB;AAC9C,gBAAQ,KAAK,CAAC;AAAA,MAChB;AACA,YAAM,QAAQ,KAAK,EAAE,CAAC;AACtB,UAAI,QAAQ,WAAY,cAAa;AAAA,eAC5B,QAAQ,WAAY,aAAY;AAAA,eAChC,QAAQ,SAAU,QAAO,KAAK,IAAI,GAAG,SAAS,OAAO,EAAE,KAAK,CAAC;AAAA;AAEpE,qBAAa,MACV,MAAM,GAAG,EACT,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,OAAO;AAAA,IACrB,WAAW,QAAQ,oBAAoB;AAErC,sBAAgB;AAAA,IAClB,WAAW,IAAI,WAAW,IAAI,GAAG;AAC/B,UAAI,CAAC,YAAY,IAAI,GAAG,GAAG;AACzB,gBAAQ,MAAM,iBAAiB,GAAG,EAAE;AACpC,gBAAQ,KAAK,CAAC;AAAA,MAChB;AAAA,IACF,OAAO;AACL,YAAM,KAAK,GAAG;AAAA,IAChB;AAAA,EACF;AAEA,SAAO,EAAE,YAAY,WAAW,YAAY,MAAM,eAAe,MAAM;AACzE;AAIA,eAAe,eAAe,MAAgB;AAC5C,QAAM,EAAE,YAAY,WAAW,YAAY,MAAM,eAAe,MAAM,IAAI,cAAc,IAAI;AAE5F,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ,MAAM,mCAAmC;AACjD,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,YAAsB,CAAC;AAC7B,aAAW,KAAK,OAAO;AACrB,cAAU,KAAK,GAAG,iBAAiB,CAAC,CAAC;AAAA,EACvC;AAEA,MAAI,UAAU,WAAW,GAAG;AAC1B,YAAQ,MAAM,4BAA4B;AAC1C,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,UAAU,MAAM,WAAW,WAAW,UAAU;AACtD,QAAM,UAAwB,CAAC;AAC/B,MAAI,cAAc;AAElB,MAAI;AACF,eAAW,YAAY,WAAW;AAChC,UAAI;AACF,cAAM,MAAM,MAAM,aAAkB,aAAQ,QAAQ,GAAG,YAAY,GAAG;AACtE,cAAM,aAAa,YAAwB,GAAG;AAE9C,cAAM,kBAAkB,mBAAmB,UAAU;AACrD,YAAI,iBAAiB;AACnB,kBAAQ,MAAM,UAAU,QAAQ,IAAI,eAAe,EAAE;AACrD;AACA;AAAA,QACF;AAKA,cAAM,WAAW,WAAoB,KAAK,iBAAiB;AAK3D,YAAI;AACJ,YAAI,OAAO,aAAa,YAAY;AAClC,0BAAgB;AAAA,QAClB,WAAW,aAAa,QAAW;AACjC,kBAAQ;AAAA,YACN,qBAAqB,QAAQ,sCAAsC,OAAO,QAAQ;AAAA,UACpF;AACA;AACA;AAAA,QACF;AAEA,YAAI;AAEJ,YAAI,eAAe;AAEjB,4BAAkB,OAAO,OAAO,OAAO;AACrC,kBAAM;AAAA,cACJ;AAAA,cACA,MAAM;AAAA,cACN;AAAA,YACF,IAAI,MAAM,QAAQ,eAAe,YAAY;AAC3C,qBAAO,cAAc,OAAO,EAAE;AAAA,YAChC,CAAC;AACD,mBAAO;AAAA,cACL,QAAQ,OAAO;AAAA,cACf,MAAM,OAAO,QAAQ;AAAA,cACrB,UAAU,OAAO,YAAY;AAAA,YAC/B;AAAA,UACF;AAAA,QACF,WAAW,QAAQ,YAAY,WAAW,QAAQ,GAAG;AAEnD,4BAAkB,OAAO,UAAU;AACjC,kBAAM,EAAE,QAAQ,MAAM,SAAS,IAAI,MAAM,QAAQ,eAAe,YAAY;AAC1E,qBAAO,QAAQ,QAAQ,WAAW,UAAU,KAAK;AAAA,YACnD,CAAC;AACD,mBAAO,EAAE,QAAQ,QAAQ,MAAM,SAAS;AAAA,UAC1C;AAAA,QACF,OAAO;AAcL,gBAAM,eAAe,MAAM;AACzB,kBAAM,MAAM,OAAO,KAAK,GAAG,EAAE,OAAO,CAAC,MAAM,MAAM,SAAS;AAC1D,kBAAM,cACJ,IAAI,WAAW,OAAO,IAAI,YAAY,WAClC,OAAO,KAAK,IAAI,OAAkC,IAClD,CAAC;AACP,kBAAM,MAAM,CAAC,GAAG,oBAAI,IAAI,CAAC,GAAG,KAAK,GAAG,WAAW,CAAC,CAAC;AACjD,mBAAO,IAAI,SAAS,IAAI,mBAAmB,IAAI,KAAK,IAAI,CAAC,OAAO;AAAA,UAClE,GAAG;AACH,kBAAQ;AAAA,YACN,qBAAqB,QAAQ,uEACD,WAAW,QAAQ;AAAA,KAC5C,cAAc,KAAK,WAAW;AAAA,IAAO,MACtC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,UAOJ;AACA;AACA;AAAA,QACF;AASA,cAAM,aAAa,gBAAgB,EAAE,eAAe,KAAK,IAAI;AAE7D,YAAI,OAAO,GAAG;AAUZ,gBAAM,EAAE,WAAW,IAAI,MAAM,OAAO,QAAa;AACjD,gBAAM,aAAa,WAAW;AAC9B,gBAAM,aAA2B,CAAC;AAClC,cAAI;AAEJ,mBAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,oBAAQ,MAAM,kBAAkB,IAAI,CAAC,IAAI,IAAI,KAAK;AAClD,gBAAI;AACF,oBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,SAAS,UAAU;AAC7E,qBAAO,SAAS,aAAa;AAC7B,qBAAO,SAAS,WAAW;AAC3B,yBAAW,KAAK,MAAM;AAAA,YACxB,SAAS,KAAK;AACZ,2BAAa,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,GAAG,CAAC;AAC/D,sBAAQ,MAAM,kBAAkB,IAAI,CAAC,IAAI,IAAI,YAAY,WAAW,OAAO,EAAE;AAK7E;AAAA,YACF;AAAA,UACF;AAEA,cAAI,WAAW,WAAW,GAAG;AAK3B,oBAAQ;AAAA,cACN,uBAAuB,QAAQ,wBAAwB,YAAY,WAAW,iBAAiB;AAAA,YACjG;AACA;AACA;AAAA,UACF;AAEA,gBAAM,UAAU,WAAW,SAAS;AACpC,cAAI,SAAS;AAmBX,kBAAM,aAAa,aACf,WAAW,WAAW,OAAO,UAAU,KAAK,SAC5C;AACJ,uBAAW,KAAK,YAAY;AAC1B,gBAAE,SAAS,mBAAmB;AAC9B,gBAAE,SAAS,iBAAiB,WAAW;AACvC,gBAAE,SAAS,iBAAiB;AAC5B,kBAAI,WAAY,GAAE,SAAS,eAAe;AAAA,YAC5C;AACA,oBAAQ;AAAA,cACN,uBAAuB,WAAW,MAAM,OAAO,IAAI,uBAAuB,QAAQ;AAAA,YACpF;AACA;AAAA,UACF;AAEA,gBAAM,UAAU,cAAc,UAAU;AACxC,kBAAQ,IAAI,OAAO,oBAAoB,OAAO,IAAI,IAAI;AACtD,qBAAW,KAAK,WAAY,SAAQ,KAAK,CAAC;AAAA,QAC5C,OAAO;AACL,gBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,SAAS,UAAU;AAC7E,kBAAQ,KAAK,MAAM;AAEnB,kBAAQ,IAAI,OAAO,YAAY,MAAM,IAAI,IAAI;AAAA,QAC/C;AAAA,MACF,SAAS,KAAK;AACZ,gBAAQ;AAAA,UACN,sBAAsB,QAAQ,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,QACrF;AACA;AAAA,MACF;AAAA,IACF;AAEA,QAAI,cAAc,QAAQ,SAAS,GAAG;AACpC,YAAM,SAAS,QAAQ,WAAW,IAAI,QAAQ,CAAC,IAAI;AACnD,YAAM,YAAiB,aAAa,aAAQ,UAAU,CAAC;AACvD,YAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,YAAM,eAAe,YAAY,KAAK,UAAU,QAAQ,MAAM,CAAC,GAAG,OAAO;AACzE,cAAQ,IAAI,oBAAoB,UAAU,EAAE;AAAA,IAC9C;AAAA,EACF,UAAE;AACA,UAAM,QAAQ,SAAS,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EACzC;AAKA,MAAI,cAAc,KAAK,QAAQ,WAAW,GAAG;AAC3C,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAEA,KAAK,EAAE,MAAM,CAAC,QAAQ;AACpB,UAAQ,MAAM,GAAG;AACjB,UAAQ,KAAK,CAAC;AAChB,CAAC;","names":[]}

package/dist/index.cjs

@@ -542,13 +542,16 @@ function resolveThreshold(scorerName, options, metadata) {
   return LEGACY_THRESHOLD;
 }
 function evalCompare(baseline, candidate, options) {
-  const baselineRuns = Array.isArray(baseline) ? baseline : [baseline];
-  const candidateRuns = Array.isArray(candidate) ? candidate : [candidate];
-  if (baselineRuns.length === 0 || candidateRuns.length === 0) {
+  const baselineRunsRaw = Array.isArray(baseline) ? baseline : [baseline];
+  const candidateRunsRaw = Array.isArray(candidate) ? candidate : [candidate];
+  if (baselineRunsRaw.length === 0 || candidateRunsRaw.length === 0) {
     throw new Error("Cannot compare empty eval result arrays");
   }
-  const baselineRef = baselineRuns[0];
-  const candidateRef = candidateRuns[0];
+  const runCount = Math.min(baselineRunsRaw.length, candidateRunsRaw.length);
+  const baselineRuns = baselineRunsRaw.slice(0, runCount);
+  const candidateRuns = candidateRunsRaw.slice(0, runCount);
+  const baselineRef = baselineRunsRaw[0];
+  const candidateRef = candidateRunsRaw[0];
   if (baselineRef.dataset !== candidateRef.dataset) {
     throw new Error(
       `Cannot compare evals from different datasets: "${baselineRef.dataset}" vs "${candidateRef.dataset}"`
@@ -563,7 +566,6 @@ function evalCompare(baseline, candidate, options) {
   for (const name of baselineScorerNames) {
     pairedDiffs[name] = [];
   }
-  const runCount = Math.min(baselineRuns.length, candidateRuns.length);
   for (let r = 0; r < runCount; r++) {
     const bRun = baselineRuns[r];
     const cRun = candidateRuns[r];
@@ -700,9 +702,21 @@ function evalCompare(baseline, candidate, options) {
     parts.push(`${Math.abs(cost.deltaPercent).toFixed(0)}% ${dir}`);
   }
   const summaryStr = `candidate ${parts.length > 0 ? parts.join(", ") : "no meaningful changes"} with ${regressions.length} regressions and ${improvements.length} improvements`;
+  const baselinePartial = detectPartial(baselineRunsRaw);
+  const candidatePartial = detectPartial(candidateRunsRaw);
   return {
-    baseline: { id: baselineRef.id, metadata: baselineRef.metadata },
-    candidate: { id: candidateRef.id, metadata: candidateRef.metadata },
+    baseline: {
+      id: baselineRef.id,
+      metadata: baselineRef.metadata,
+      runCount,
+      ...baselinePartial ? { partial: baselinePartial } : {}
+    },
+    candidate: {
+      id: candidateRef.id,
+      metadata: candidateRef.metadata,
+      runCount,
+      ...candidatePartial ? { partial: candidatePartial } : {}
+    },
     scorers,
     timing,
     cost,
@@ -711,6 +725,20 @@ function evalCompare(baseline, candidate, options) {
     summary: summaryStr
   };
 }
+function detectPartial(runs) {
+  if (runs.length === 0) return void 0;
+  let attempted;
+  for (const r of runs) {
+    const v = r.metadata?.batchAttempted;
+    if (typeof v === "number" && Number.isFinite(v)) {
+      attempted = v;
+      break;
+    }
+  }
+  if (attempted === void 0) return void 0;
+  if (runs.length >= attempted) return void 0;
+  return { completed: runs.length, attempted };
+}
 
 // src/rescore.ts
 var import_node_crypto2 = require("crypto");