@axlsdk/eval 0.22.3 → 0.23.0

package/dist/cli.js

@@ -6,7 +6,7 @@ import {
   rescore,
   runEval,
   scorerCounts
-} from "./chunk-C7CICWIA.js";
+} from "./chunk-FQHKBSZD.js";
 
 // src/cli.ts
 import { readdirSync, statSync } from "fs";
@@ -161,6 +161,24 @@ function parseEvalArgs(args) {
   };
 }
 
+// src/cli-format.ts
+function formatModelTimingLines(modelTiming, nameWidth) {
+  if (!modelTiming) return [];
+  return Object.entries(modelTiming).map(([model, t]) => {
+    const parts = [`wire ${formatMs(t.wireMs.mean)}/${formatMs(t.wireMs.p95)}`];
+    if (t.firstTokenMs) {
+      parts.push(`first token ${formatMs(t.firstTokenMs.mean)}/${formatMs(t.firstTokenMs.p95)}`);
+    }
+    parts.push(`queued ${formatMs(t.queuedMs.mean)}`);
+    parts.push(`retries ${formatMs(t.retryMs.mean)}`);
+    const calls = `${t.calls} call${t.calls === 1 ? "" : "s"}`;
+    return `    ${model.padEnd(Math.max(nameWidth - 2, 0))}  ${parts.join(" \xB7 ")}  (${calls}, mean/p95 per call)`;
+  });
+}
+function formatMs(ms) {
+  return `${Math.round(ms)}ms`;
+}
+
 // src/cli.ts
 function refuseToGate(reason) {
   console.error(`[axl-eval] Refusing to gate: ${reason}`);
@@ -521,6 +539,7 @@ function formatTable(result) {
       `  ${"Timing".padEnd(maxNameLen)}  ${(t.mean / 1e3).toFixed(2).padStart(colWidth)}s ${(t.p50 / 1e3).toFixed(2).padStart(colWidth)}s ${(t.p95 / 1e3).toFixed(2).padStart(colWidth)}s`
     );
   }
+  lines.push(...formatModelTimingLines(result.summary.modelTiming, maxNameLen));
   const durationSec = (result.duration / 1e3).toFixed(1);
   const costStr = result.totalCost > 0 ? `$${result.totalCost.toFixed(2)}` : "$0.00";
   lines.push("");

package/dist/cli.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/cli.ts","../src/cli-utils.ts","../src/cli-validate.ts","../src/cli-args.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, evaluateScorerErrorRateGate } 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';\nimport { parseEvalArgs, envInt } from './cli-args.js';\nimport { scorerCounts } from './utils.js';\n\n/**\n * Refuse to certify a comparison and exit non-zero, with one consistent\n * message format. Routes both the scorer-filtered guard and the\n * `--max-scorer-error-rate` failure-rate gate through a single path so the\n * \"why CI failed\" line always reads the same.\n */\nfunction refuseToGate(reason: string): never {\n  console.error(`[axl-eval] Refusing to gate: ${reason}`);\n  process.exit(1);\n}\n\n/**\n * Read a scorer's `scored`/`failed` counts from a summary entry, falling back\n * to recomputing from `items` when the fields are absent (pre-0.18.0\n * artifacts). `items` is omitted for multi-run summaries that carry no items —\n * there the fields are authoritative (summed by `aggregateRuns`).\n */\nfunction scorerSampleCounts(\n  entry: { scored?: number; failed?: number; skipped?: number },\n  name: string,\n  items?: EvalResult['items'],\n): { scored: number; failed: number; skipped: number } {\n  if (entry.scored != null && entry.failed != null) {\n    // `scored`/`failed` are authoritative; `skipped` was added later, so an\n    // 0.18.0-era artifact may carry the first two without it — recompute that one\n    // field from items when we can, else 0.\n    const skipped = entry.skipped ?? (items ? scorerCounts(items, name).skipped : 0);\n    return { scored: entry.scored, failed: entry.failed, skipped };\n  }\n  if (items) return scorerCounts(items, name);\n  return { scored: entry.scored ?? 0, failed: entry.failed ?? 0, skipped: entry.skipped ?? 0 };\n}\n\n/** Install a SIGINT/SIGTERM handler that aborts an in-flight eval gracefully.\n *  First signal: aborts the controller so runEval / rescore propagate cancellation\n *  through ScorerContext.signal into in-flight provider.chat calls. The eval\n *  unwinds normally and runtime.shutdown() runs in the `finally` block.\n *  Second signal: hard exit — the user has signalled twice; cleanup is hung\n *  or taking too long, so we abandon gracefulness rather than appear frozen. */\nfunction installAbortHandler(controller: AbortController): void {\n  let signalled = false;\n  const onSignal = (sig: NodeJS.Signals) => {\n    if (signalled) {\n      console.error(`\\n[axl-eval] Second ${sig} — forcing exit.`);\n      process.exit(130);\n    }\n    signalled = true;\n    console.error(`\\n[axl-eval] ${sig} received — cancelling eval (press again to force exit)...`);\n    controller.abort();\n  };\n  process.on('SIGINT', onSignal);\n  process.on('SIGTERM', onSignal);\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> --concurrency <n>       Override item concurrency (flag > env\n                                          AXL_EVAL_CONCURRENCY > config > 5).\n                                          Max concurrent scorer calls is\n                                          concurrency x scorerConcurrency (the\n                                          per-eval scorerConcurrency defaults to 5).\n  axl-eval <path> --scorers <a,b>         Run only the named scorers (single eval\n                                          file only). Stamps the result as a\n                                          subset so compare won't treat it as a\n                                          full baseline.\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  axl-eval compare <a> <b> --max-scorer-error-rate <0..1>  Exit 1 if a scorer's\n                                          failure rate exceeds the limit on\n                                          either side (deterministic scorers\n                                          tolerate zero failures). Always warns\n                                          when either side rests on a thinned\n                                          sample, even without this flag.\n\nConfig auto-detection (when --config is not specified):\n  ${CONFIG_CANDIDATES.join(' -> ')}\n\nWhen a config is found, the exported AxlRuntime is passed to executeWorkflow\nand cost is tracked automatically via runtime.trackCost().\nWhen no config is found, a bare AxlRuntime is created (providers from env vars).\n`);\n    process.exit(0);\n  }\n\n  if (args[0] === 'compare') {\n    // `compare` is pure computation — no LLM calls, no AbortSignal plumbing needed.\n    await runCompare(args.slice(1));\n    return;\n  }\n\n  // SIGINT/SIGTERM aborts in-flight LLM calls (eval + rescore) instead of leaving\n  // the process to hard-exit mid-request, which would skip runtime.shutdown() and\n  // leak streaming-buffer state.\n  const controller = new AbortController();\n  installAbortHandler(controller);\n\n  if (args[0] === 'rescore') {\n    await runRescore(args.slice(1), controller.signal);\n    return;\n  }\n\n  await runEvalCommand(args, controller.signal);\n}\n\nfunction parseThresholdArg(args: string[]): Record<string, number> | number | undefined {\n  const idx = args.indexOf('--threshold');\n  if (idx === -1 || idx + 1 >= args.length) return undefined;\n  const value = args[idx + 1];\n\n  // Global numeric threshold: --threshold 0.05\n  if (/^[\\d.]+$/.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\n/**\n * Parse `--max-scorer-error-rate <0..1>`. Returns the numeric limit, or\n * `undefined` when the flag is absent. Exits non-zero on a malformed value\n * (fail loud — a typo'd gate threshold must not silently disable the gate).\n */\nfunction parseMaxScorerErrorRate(args: string[]): number | undefined {\n  const idx = args.indexOf('--max-scorer-error-rate');\n  if (idx === -1) return undefined;\n  if (idx + 1 >= args.length) {\n    console.error('Error: --max-scorer-error-rate requires a value in [0, 1]');\n    process.exit(1);\n  }\n  const raw = args[idx + 1];\n  // Strict numeric parse — `parseFloat('0.5abc')` returns 0.5 and `parseFloat('0,5')`\n  // returns 0, both of which would silently set an unintended gate threshold. Reject\n  // anything that isn't a clean decimal (matches `parseThresholdArg`'s guard).\n  const n = /^\\d*\\.?\\d+$/.test(raw.trim()) ? Number(raw) : NaN;\n  if (!Number.isFinite(n) || n < 0 || n > 1) {\n    console.error(`Error: --max-scorer-error-rate must be a number in [0, 1], got \"${raw}\"`);\n    process.exit(1);\n  }\n  return n;\n}\n\n/**\n * If a raw compare input (single result or multi-run array) was produced by a\n * `--scorers` filtered run, return the scorer names it ran; otherwise undefined.\n * Reads the first run's metadata — a multi-run group shares the same stamp.\n */\nfunction scorerFilteredScorers(r: EvalResult | EvalResult[]): string[] | undefined {\n  const first = Array.isArray(r) ? r[0] : r;\n  const meta = first?.metadata;\n  if (!meta || meta.scorerFiltered !== true) return undefined;\n  return Array.isArray(meta.scorersRun) ? (meta.scorersRun as string[]) : [];\n}\n\nasync function runCompare(args: string[]) {\n  const failOnRegression = args.includes('--fail-on-regression');\n  const thresholds = parseThresholdArg(args);\n  const maxScorerErrorRate = parseMaxScorerErrorRate(args);\n  // Exclude flags and the values consumed by value-taking flags so neither a\n  // threshold nor an error-rate value is mistaken for a result file path.\n  const valueFlags = new Set(['--threshold', '--max-scorer-error-rate']);\n  const files = args.filter(\n    (a, i) => !a.startsWith('--') && !(i > 0 && valueFlags.has(args[i - 1])),\n  );\n\n  if (files.length !== 2) {\n    console.error(\n      'Usage: axl-eval compare <baseline.json> <candidate.json> [--threshold <value>] [--fail-on-regression] [--max-scorer-error-rate <0..1>]',\n    );\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\n  // A scorer-filtered result (CLI `--scorers`) tested only a subset of scorers,\n  // so it isn't a sound baseline/candidate for a full comparison. Check this\n  // BEFORE evalCompare(): a subset-vs-full comparison has mismatched scorer sets\n  // and evalCompare() would throw a generic \"different scorers\" error that never\n  // mentions --scorers. Warn loudly here so the cause is actionable; under\n  // `--fail-on-regression`, refuse outright (a swallowed stderr line wouldn't\n  // actually prevent the footgun).\n  const filteredSides = (\n    [\n      ['baseline', baseline],\n      ['candidate', candidate],\n    ] as const\n  ).filter(([, r]) => scorerFilteredScorers(r) != null);\n  for (const [side, r] of filteredSides) {\n    const ran = scorerFilteredScorers(r);\n    const ranStr = ran && ran.length ? ` (ran: ${ran.join(', ')})` : '';\n    console.error(\n      `[axl-eval] WARNING: ${side} was run with --scorers — it tested only a subset${ranStr}, so this comparison is incomplete.`,\n    );\n  }\n  if (failOnRegression && filteredSides.length > 0) {\n    refuseToGate(\n      `${filteredSides.map(([s]) => s).join(', ')} was run with --scorers (a scorer subset); re-run without --scorers for a complete baseline.`,\n    );\n  }\n\n  const compareOptions = thresholds != null ? { thresholds } : undefined;\n  const comparison = evalCompare(baseline, candidate, compareOptions);\n\n  console.log(\n    `\\nCompare: baseline (${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: ${stable}\\n`,\n  );\n\n  // Scorer failure-rate signal. Always WARN loudly when either side rests on a\n  // thinned sample (a scorer ran and failed on some items) — a regression check\n  // computed over survivors can look green while half the judges errored. Only\n  // REFUSE to gate when the user opted in with `--max-scorer-error-rate` AND a\n  // scorer is over tolerance (type-aware: deterministic scorers tolerate zero\n  // failures, LLM judges use the flag value), or a scorer produced no scored\n  // items at all on a side (a zero-sample mean can't be certified).\n  let sawThinnedSample = false;\n  for (const name of scorerNames) {\n    const s = comparison.scorers[name];\n    const bFailed = s.baselineFailed ?? 0;\n    const cFailed = s.candidateFailed ?? 0;\n    if (bFailed > 0 || cFailed > 0) {\n      sawThinnedSample = true;\n      const bAtt = (s.baselineScored ?? 0) + bFailed;\n      const cAtt = (s.candidateScored ?? 0) + cFailed;\n      console.error(\n        `[axl-eval] WARNING: scorer \"${name}\" failed on some items ` +\n          `(baseline ${bFailed}/${bAtt}, candidate ${cFailed}/${cAtt}); ` +\n          `the means above are computed over the surviving sample.`,\n      );\n    }\n    // Conditional scorers (`applies`) skip N/A items. Skips never move the\n    // delta (excluded from both means), but they make the two per-side means\n    // cover different applicable subsets when the sides skipped different COUNTS,\n    // OR the same count of DIFFERENT items — the latter shows up as a paired\n    // shortfall: fewer items scored on BOTH sides (`n`) than either side scored\n    // alone. (`bSkipped !== cSkipped` alone would miss the disjoint-but-equal\n    // case.) Note it (advisory only; not a gate; failures are covered by the\n    // separate WARNING above).\n    const bSkipped = s.baselineSkipped ?? 0;\n    const cSkipped = s.candidateSkipped ?? 0;\n    const pairedN = s.n ?? 0;\n    const minScored = Math.min(s.baselineScored ?? 0, s.candidateScored ?? 0);\n    const skipsDiverge =\n      (bSkipped > 0 || cSkipped > 0) && (bSkipped !== cSkipped || pairedN < minScored);\n    if (skipsDiverge) {\n      console.error(\n        `[axl-eval] NOTE: scorer \"${name}\" applies to different subsets per side ` +\n          `(baseline ${bSkipped} N/A, candidate ${cSkipped} N/A; ${pairedN} item(s) scored on both); ` +\n          `the two means cover different applicable subsets, so compare the delta with care.`,\n      );\n    }\n  }\n  // Tell the user how to make the advisory warning blocking (when they haven't).\n  if (sawThinnedSample && maxScorerErrorRate == null) {\n    console.error(\n      `[axl-eval] (advisory — pass --max-scorer-error-rate <0..1> to fail CI on a thinned sample)`,\n    );\n  }\n\n  // Gate-side refusal (opt-in). The decision lives in evaluateScorerErrorRateGate\n  // (pure + unit-tested) so the same type-aware tolerance + zero-sample logic\n  // isn't re-implemented here and at the source-side gate in runEval.\n  if (maxScorerErrorRate != null) {\n    const reason = evaluateScorerErrorRateGate(comparison, maxScorerErrorRate);\n    if (reason) refuseToGate(reason);\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[], signal: AbortSignal) {\n  const { outputPath, configArg, conditions, concurrency, scorerNames, paths } =\n    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  // --scorers is a run-command-only filter — rescore always re-runs the full\n  // scorer set from the eval file. Fail loudly rather than silently ignoring it\n  // (it parses as a known flag, so it wouldn't trip the unknown-flag guard).\n  if (scorerNames?.length) {\n    console.error(\n      'Error: --scorers is not supported with rescore; pass a scorer subset in the eval file instead.',\n    );\n    process.exit(1);\n  }\n\n  // Honor the same item-concurrency override as the run command (flag > env).\n  // scorerConcurrency keeps its rescore default (5).\n  const itemConcurrency = concurrency ?? envInt('AXL_EVAL_CONCURRENCY');\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      if (signal.aborted) break;\n      rescored.push(\n        await rescore(resultData, evalConfig.scorers, runtime, {\n          signal,\n          ...(itemConcurrency != null ? { concurrency: itemConcurrency } : {}),\n        }),\n      );\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)}  ${'p95'.padStart(colWidth)}`,\n  );\n\n  const ruleLen = maxNameLen + 2 + (colWidth + 2) * 5;\n  lines.push('  ' + '\\u2500'.repeat(ruleLen));\n\n  for (const name of scorerNames) {\n    const s = result.summary.scorers[name];\n    const { scored, failed, skipped } = scorerSampleCounts(s, name, result.items);\n    // Loud trailing annotation: `failed > 0` means the mean rests on a thinned\n    // sample; `skipped > 0` (an `applies`-conditional scorer) labels items the\n    // scorer deliberately didn't run — crucial so a fully-skipped scorer (all\n    // `--`) reads as \"N/A\" rather than \"broken\". Both shown only when non-zero so\n    // clean, unconditional runs stay uncluttered.\n    const annotationParts: string[] = [];\n    if (failed > 0) annotationParts.push(`${scored}/${scored + failed} scored, ${failed} failed`);\n    if (skipped > 0) annotationParts.push(`${skipped} N/A`);\n    const annotation = annotationParts.length > 0 ? `   (${annotationParts.join(', ')})` : '';\n    // Detect scorers with no valid scores (all items errored → computeStats([]) → all zeros)\n    if (scored === 0) {\n      lines.push(\n        `  ${name.padEnd(maxNameLen)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}${annotation}`,\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)}  ${s.p95.toFixed(2).padStart(colWidth)}${annotation}`,\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    // Multi-run summaries have no items, so the summed scored/failed fields are\n    // authoritative (no fallback recompute possible). Annotate when failed > 0.\n    const { scored, failed } = scorerSampleCounts(s, name);\n    const annotation =\n      failed > 0 ? `   (${scored}/${scored + failed} scored, ${failed} failed)` : '';\n    lines.push(\n      `  ${name.padEnd(maxNameLen)}  ${meanStd.padStart(colWidth)}  ${s.min.toFixed(3).padStart(8)}  ${s.max.toFixed(3).padStart(8)}${annotation}`,\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: ${durationStr}s`);\n\n  return lines.join('\\n');\n}\n\n/**\n * Print each degraded scorer (set by `runEval` when `failOnScorerErrorRate`\n * tripped) to stderr and return whether the run was degraded, so the caller can\n * count the file as failed (→ non-zero exit). Quiet when nothing degraded.\n */\nfunction reportDegraded(result: EvalResult, label: string): boolean {\n  const d = result.summary.degraded;\n  if (!d || d.length === 0) return false;\n  for (const x of d) {\n    const attempted = x.scored + x.failed;\n    const limitStr =\n      x.type === 'deterministic'\n        ? 'deterministic zero-tolerance'\n        : `${(x.limit * 100).toFixed(1)}% limit`;\n    console.error(\n      `[axl-eval] DEGRADED: ${label} scorer \"${x.scorer}\" (${x.type}) failed ` +\n        `${(x.rate * 100).toFixed(1)}% (${x.failed}/${attempted}), over the ${limitStr}.`,\n    );\n  }\n  return true;\n}\n\n/**\n * Advisory (non-fatal): a conditional scorer (`applies`) that ended up applicable\n * to ZERO items — every item was skipped as N/A, nothing scored or failed. Its\n * mean is a meaningless empty-sample 0 and the failure-rate gate can't assess it\n * (a zero-sample scorer is never `degraded`). That's almost always a\n * misconfiguration: an over-strict predicate, a dataset missing the flagged\n * items, or a predicate that forgot to `return`. The `(N N/A)` table annotation\n * hints at it, but a fully-skipped GATE scorer would otherwise exit 0 with no\n * stderr signal — this closes that silent-green hole without changing the exit\n * code (skips are legitimately conditional; we inform, we don't fail).\n */\nfunction reportFullySkippedScorers(result: EvalResult, label: string): void {\n  for (const [name, entry] of Object.entries(result.summary.scorers)) {\n    const { scored, failed, skipped } = scorerSampleCounts(entry, name, result.items);\n    if (skipped > 0 && scored + failed === 0) {\n      console.error(\n        `[axl-eval] NOTE: ${label} scorer \"${name}\" was not applicable to any item ` +\n          `(${skipped} N/A, 0 scored) — its mean is not meaningful. Check the scorer's ` +\n          `\\`applies\\` predicate and dataset if you expected it to run.`,\n      );\n    }\n  }\n}\n\n/**\n * A run where EVERY item errored in the workflow (0 succeeded) produced no valid\n * output to score — the eval is meaningless and must never pass CI green. This\n * is distinct from (and complementary to) the scorer failure-rate gate: that one\n * is about a flaky/failing *scorer*, this is about a broken *workflow*. The\n * scorer gate deliberately ignores a zero-sample scorer (nothing ran), so without\n * this guard a 100%-workflow-error run exits 0 — a silent-green trap.\n *\n * Non-configurable on purpose: a 0%-success eval is unambiguously broken, so this\n * always fails. (A configurable `failOnItemErrorRate` for PARTIAL workflow-failure\n * gating is a reasonable future opt-in; the per-item failure count is already shown\n * loudly in the table either way.) Returns whether the run was a total wipeout.\n */\nfunction reportTotalWipeout(result: EvalResult, label: string): boolean {\n  const { count, failures } = result.summary;\n  if (count === 0 || failures < count) return false;\n  console.error(\n    `[axl-eval] FAILED: ${label} — all ${count} item(s) errored in the workflow ` +\n      `(0 succeeded); the eval produced no scorable output.`,\n  );\n  return true;\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// ── Main eval command ──────────────────────────────────────────────\n\nasync function runEvalCommand(args: string[], signal: AbortSignal) {\n  const {\n    outputPath,\n    configArg,\n    conditions,\n    runs,\n    captureTraces,\n    concurrency,\n    scorerNames,\n    paths,\n  } = 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  // --scorers is single-file only — filtering by name across a glob/dir of evals\n  // with different scorer sets is ambiguous. Fail fast, before importing any\n  // module, rather than discovering the conflict mid-loop.\n  if (scorerNames?.length && evalFiles.length > 1) {\n    console.error(`Error: --scorers requires a single eval file (matched ${evalFiles.length}).`);\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      // Short-circuit if the user aborted between eval files. The in-flight eval\n      // unwinds on its own via runEval({ signal }); this prevents starting the next one.\n      if (signal.aborted) break;\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        // Concurrency override precedence: flag > env > config > default.\n        // Mutating evalConfig in place mirrors how this loop already stamps\n        // result.metadata. `concurrency` is pure scheduling — never changes\n        // results — so a per-run override is always safe.\n        evalConfig.concurrency =\n          concurrency ?? envInt('AXL_EVAL_CONCURRENCY') ?? evalConfig.concurrency ?? 5;\n\n        // --scorers: run a subset of scorers for a focused iteration loop.\n        // (The single-file guard already ran before the loop.) Validate-then-\n        // filter so the error lists every available name.\n        if (scorerNames?.length) {\n          const available = evalConfig.scorers.map((s) => s.name);\n          const unknown = scorerNames.filter((n) => !available.includes(n));\n          if (unknown.length) {\n            console.error(\n              `Error: --scorers: unknown scorer(s): ${unknown.join(', ')}. Available: ${available.join(', ')}`,\n            );\n            process.exit(1);\n          }\n          evalConfig.scorers = evalConfig.scorers.filter((s) => scorerNames.includes(s.name));\n          if (evalConfig.scorers.length === 0) {\n            console.error('Error: --scorers matched no scorers.');\n            process.exit(1);\n          }\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        // Always pass the SIGINT-driven signal so Ctrl+C aborts in-flight\n        // workflows + scorer LLM calls instead of dropping the connection mid-request.\n        const runOptions = { signal, ...(captureTraces ? { captureTraces: true } : {}) };\n\n        // Stamp filtered runs so a scorer-subset result can't be silently used\n        // as a full baseline. `compare` warns (and refuses to gate) on it. Must\n        // be applied to EVERY run in a multi-run batch — Studio's multi-run\n        // builder spreads only run[0]'s metadata.\n        const scorersRun = scorerNames?.length ? evalConfig.scorers.map((s) => s.name) : undefined;\n        const stampFiltered = (result: EvalResult) => {\n          if (scorersRun) {\n            result.metadata.scorerFiltered = true;\n            result.metadata.scorersRun = scorersRun;\n          }\n        };\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            if (signal.aborted) break;\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              stampFiltered(result);\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\n          // Failure-rate gate (opt-in via failOnScorerErrorRate) + total-wipeout\n          // guard (always). Report every run; count the file as failed unless it\n          // was already counted as a partial batch above (avoids double-count).\n          let anyFailing = false;\n          for (const r of runResults) {\n            // Call both (no short-circuit) so each prints its own diagnostic.\n            const wipeout = reportTotalWipeout(r, filePath);\n            const degraded = reportDegraded(r, filePath);\n            reportFullySkippedScorers(r, filePath); // advisory only\n            if (wipeout || degraded) anyFailing = true;\n          }\n          if (anyFailing && !partial) failedFiles++;\n        } else {\n          const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);\n          stampFiltered(result);\n          results.push(result);\n\n          console.log('\\n' + formatTable(result) + '\\n');\n          // Total-wipeout guard (always) + scorer failure-rate gate (opt-in). They\n          // are mutually exclusive (a wipeout has no scored items, so no scorer can\n          // be degraded), but report both for clarity; either fails the file.\n          const wipeout = reportTotalWipeout(result, filePath);\n          const degraded = reportDegraded(result, filePath);\n          reportFullySkippedScorers(result, filePath); // advisory only\n          if (wipeout || degraded) failedFiles++;\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  // Concurrency knobs must be positive integers when present. The worker pool\n  // defensively clamps non-positive/non-finite values to 1 at runtime, but a\n  // config that wrote `concurrency: 0` (or NaN via `Number(env)`) almost\n  // certainly has a bug — surface it loudly, matching the `--concurrency` flag.\n  for (const key of ['concurrency', 'scorerConcurrency'] as const) {\n    const v = (c as Record<string, unknown>)[key];\n    if (v !== undefined && (typeof v !== 'number' || !Number.isInteger(v) || v < 1)) {\n      return `exports an invalid ${key} (${typeof v === 'number' ? v : typeof v}) — must be a positive integer.`;\n    }\n  }\n  // The failure-rate gate must fail loud on a bad value, symmetric with the\n  // `--max-scorer-error-rate` flag. Without this, a typo (e.g. `5` meaning \"5%\")\n  // only `console.warn`s at runtime and silently disables the gate the user\n  // explicitly opted into — the exact silent-no-op the feature exists to catch.\n  {\n    const v = (c as Record<string, unknown>).failOnScorerErrorRate;\n    if (v !== undefined && (typeof v !== 'number' || !Number.isFinite(v) || v < 0 || v > 1)) {\n      return `exports an invalid failOnScorerErrorRate (${typeof v === 'number' ? v : typeof v}) — must be a number between 0 and 1.`;\n    }\n  }\n  return undefined;\n}\n","/**\n * Argument parsing for the `axl-eval` CLI, factored out of `cli.ts` so it can be\n * unit-tested without importing `cli.ts` (whose module body runs `main()` on\n * import).\n */\n\nexport const KNOWN_FLAGS = new Set([\n  '--output',\n  '--config',\n  '--conditions',\n  '--fail-on-regression',\n  '--threshold',\n  '--runs',\n  '--capture-traces',\n  '--concurrency',\n  '--scorers',\n]);\n\n/**\n * The subset of flags `parseEvalArgs` consumes a following value for. NOT every\n * value-taking flag in the CLI: `--threshold` takes a value too but is parsed\n * by `runCompare`'s own parser, not here. Every entry MUST also be in\n * {@link KNOWN_FLAGS} — a value-flag missing from KNOWN_FLAGS would be rejected\n * as \"unknown\", and one missing from VALUE_FLAGS would silently swallow its\n * value into `paths`. A test asserts `VALUE_FLAGS ⊆ KNOWN_FLAGS`.\n */\nexport const VALUE_FLAGS = new Set([\n  '--output',\n  '--config',\n  '--conditions',\n  '--runs',\n  '--concurrency',\n  '--scorers',\n]);\n\nexport type ParsedEvalArgs = {\n  outputPath?: string;\n  configArg?: string;\n  conditions: string[];\n  runs: number;\n  captureTraces: boolean;\n  /** Item-level concurrency override (flag value). Clamped to >= 1. */\n  concurrency?: number;\n  /** Scorer names from `--scorers` (deduped, in first-seen order). */\n  scorerNames?: string[];\n  paths: string[];\n};\n\n/**\n * Read a positive integer from an environment variable. Returns `undefined`\n * when the var is absent, non-numeric, or `<= 0`, so a malformed value falls\n * through to the next precedence tier rather than poisoning the run.\n */\nexport function envInt(name: string): number | undefined {\n  const raw = process.env[name];\n  if (raw == null || raw === '') return undefined;\n  const n = parseInt(raw, 10);\n  if (!Number.isFinite(n) || n <= 0) return undefined;\n  return n;\n}\n\nexport function parseEvalArgs(args: string[]): ParsedEvalArgs {\n  let outputPath: string | undefined;\n  let configArg: string | undefined;\n  let conditions: string[] = [];\n  let runs = 1;\n  let captureTraces = false;\n  let concurrency: number | undefined;\n  let scorerNames: string[] | undefined;\n  const paths: string[] = [];\n\n  for (let i = 0; i < args.length; i++) {\n    const arg = args[i];\n    if (VALUE_FLAGS.has(arg)) {\n      if (i + 1 >= args.length) {\n        console.error(`Error: ${arg} requires a value`);\n        process.exit(1);\n      }\n      const value = args[++i];\n      if (arg === '--output') outputPath = value;\n      else if (arg === '--config') configArg = value;\n      else if (arg === '--runs') runs = Math.max(1, parseInt(value, 10) || 1);\n      else if (arg === '--concurrency') {\n        // Clamp to a floor of 1 with a warning (matches `--runs`' clamp ethos\n        // but surfaces the typo). A `0` here would otherwise spawn zero workers.\n        // Note the intentional asymmetry with `envInt` above: an explicit flag\n        // typo is worth a warning, whereas a malformed ambient env var falls\n        // through silently to the next precedence tier.\n        const n = parseInt(value, 10);\n        if (!Number.isFinite(n) || n <= 0) {\n          console.error(`[axl-eval] Ignoring invalid --concurrency \"${value}\"; using 1.`);\n          concurrency = 1;\n        } else {\n          concurrency = n;\n        }\n      } else if (arg === '--scorers') {\n        scorerNames = [\n          ...new Set(\n            value\n              .split(',')\n              .map((s) => s.trim())\n              .filter(Boolean),\n          ),\n        ];\n      } else {\n        conditions = value\n          .split(',')\n          .map((c) => c.trim())\n          .filter(Boolean);\n      }\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 {\n    outputPath,\n    configArg,\n    conditions,\n    runs,\n    captureTraces,\n    concurrency,\n    scorerNames,\n    paths,\n  };\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;AAKA,aAAW,OAAO,CAAC,eAAe,mBAAmB,GAAY;AAC/D,UAAM,IAAK,EAA8B,GAAG;AAC5C,QAAI,MAAM,WAAc,OAAO,MAAM,YAAY,CAAC,OAAO,UAAU,CAAC,KAAK,IAAI,IAAI;AAC/E,aAAO,sBAAsB,GAAG,KAAK,OAAO,MAAM,WAAW,IAAI,OAAO,CAAC;AAAA,IAC3E;AAAA,EACF;AAKA;AACE,UAAM,IAAK,EAA8B;AACzC,QAAI,MAAM,WAAc,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,KAAK,IAAI,KAAK,IAAI,IAAI;AACvF,aAAO,6CAA6C,OAAO,MAAM,WAAW,IAAI,OAAO,CAAC;AAAA,IAC1F;AAAA,EACF;AACA,SAAO;AACT;;;ACvEO,IAAM,cAAc,oBAAI,IAAI;AAAA,EACjC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUM,IAAM,cAAc,oBAAI,IAAI;AAAA,EACjC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAoBM,SAAS,OAAO,MAAkC;AACvD,QAAM,MAAM,QAAQ,IAAI,IAAI;AAC5B,MAAI,OAAO,QAAQ,QAAQ,GAAI,QAAO;AACtC,QAAM,IAAI,SAAS,KAAK,EAAE;AAC1B,MAAI,CAAC,OAAO,SAAS,CAAC,KAAK,KAAK,EAAG,QAAO;AAC1C,SAAO;AACT;AAEO,SAAS,cAAc,MAAgC;AAC5D,MAAI;AACJ,MAAI;AACJ,MAAI,aAAuB,CAAC;AAC5B,MAAI,OAAO;AACX,MAAI,gBAAgB;AACpB,MAAI;AACJ,MAAI;AACJ,QAAM,QAAkB,CAAC;AAEzB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAClB,QAAI,YAAY,IAAI,GAAG,GAAG;AACxB,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,eAC7D,QAAQ,iBAAiB;AAMhC,cAAM,IAAI,SAAS,OAAO,EAAE;AAC5B,YAAI,CAAC,OAAO,SAAS,CAAC,KAAK,KAAK,GAAG;AACjC,kBAAQ,MAAM,8CAA8C,KAAK,aAAa;AAC9E,wBAAc;AAAA,QAChB,OAAO;AACL,wBAAc;AAAA,QAChB;AAAA,MACF,WAAW,QAAQ,aAAa;AAC9B,sBAAc;AAAA,UACZ,GAAG,IAAI;AAAA,YACL,MACG,MAAM,GAAG,EACT,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,OAAO;AAAA,UACnB;AAAA,QACF;AAAA,MACF,OAAO;AACL,qBAAa,MACV,MAAM,GAAG,EACT,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,OAAO;AAAA,MACnB;AAAA,IACF,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;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;;;AHrGA,SAAS,aAAa,QAAuB;AAC3C,UAAQ,MAAM,gCAAgC,MAAM,EAAE;AACtD,UAAQ,KAAK,CAAC;AAChB;AAQA,SAAS,mBACP,OACA,MACA,OACqD;AACrD,MAAI,MAAM,UAAU,QAAQ,MAAM,UAAU,MAAM;AAIhD,UAAM,UAAU,MAAM,YAAY,QAAQ,aAAa,OAAO,IAAI,EAAE,UAAU;AAC9E,WAAO,EAAE,QAAQ,MAAM,QAAQ,QAAQ,MAAM,QAAQ,QAAQ;AAAA,EAC/D;AACA,MAAI,MAAO,QAAO,aAAa,OAAO,IAAI;AAC1C,SAAO,EAAE,QAAQ,MAAM,UAAU,GAAG,QAAQ,MAAM,UAAU,GAAG,SAAS,MAAM,WAAW,EAAE;AAC7F;AAQA,SAAS,oBAAoB,YAAmC;AAC9D,MAAI,YAAY;AAChB,QAAM,WAAW,CAAC,QAAwB;AACxC,QAAI,WAAW;AACb,cAAQ,MAAM;AAAA,oBAAuB,GAAG,uBAAkB;AAC1D,cAAQ,KAAK,GAAG;AAAA,IAClB;AACA,gBAAY;AACZ,YAAQ,MAAM;AAAA,aAAgB,GAAG,iEAA4D;AAC7F,eAAW,MAAM;AAAA,EACnB;AACA,UAAQ,GAAG,UAAU,QAAQ;AAC7B,UAAQ,GAAG,WAAW,QAAQ;AAChC;AAEA,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;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAgCZ,kBAAkB,KAAK,MAAM,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,CAKjC;AACG,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AAEzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAKA,QAAM,aAAa,IAAI,gBAAgB;AACvC,sBAAoB,UAAU;AAE9B,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,GAAG,WAAW,MAAM;AACjD;AAAA,EACF;AAEA,QAAM,eAAe,MAAM,WAAW,MAAM;AAC9C;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;AAOA,SAAS,wBAAwB,MAAoC;AACnE,QAAM,MAAM,KAAK,QAAQ,yBAAyB;AAClD,MAAI,QAAQ,GAAI,QAAO;AACvB,MAAI,MAAM,KAAK,KAAK,QAAQ;AAC1B,YAAQ,MAAM,2DAA2D;AACzE,YAAQ,KAAK,CAAC;AAAA,EAChB;AACA,QAAM,MAAM,KAAK,MAAM,CAAC;AAIxB,QAAM,IAAI,cAAc,KAAK,IAAI,KAAK,CAAC,IAAI,OAAO,GAAG,IAAI;AACzD,MAAI,CAAC,OAAO,SAAS,CAAC,KAAK,IAAI,KAAK,IAAI,GAAG;AACzC,YAAQ,MAAM,mEAAmE,GAAG,GAAG;AACvF,YAAQ,KAAK,CAAC;AAAA,EAChB;AACA,SAAO;AACT;AAOA,SAAS,sBAAsB,GAAoD;AACjF,QAAM,QAAQ,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC,IAAI;AACxC,QAAM,OAAO,OAAO;AACpB,MAAI,CAAC,QAAQ,KAAK,mBAAmB,KAAM,QAAO;AAClD,SAAO,MAAM,QAAQ,KAAK,UAAU,IAAK,KAAK,aAA0B,CAAC;AAC3E;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,mBAAmB,KAAK,SAAS,sBAAsB;AAC7D,QAAM,aAAa,kBAAkB,IAAI;AACzC,QAAM,qBAAqB,wBAAwB,IAAI;AAGvD,QAAM,aAAa,oBAAI,IAAI,CAAC,eAAe,yBAAyB,CAAC;AACrE,QAAM,QAAQ,KAAK;AAAA,IACjB,CAAC,GAAG,MAAM,CAAC,EAAE,WAAW,IAAI,KAAK,EAAE,IAAI,KAAK,WAAW,IAAI,KAAK,IAAI,CAAC,CAAC;AAAA,EACxE;AAEA,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ;AAAA,MACN;AAAA,IACF;AACA,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;AAS7C,QAAM,gBACJ;AAAA,IACE,CAAC,YAAY,QAAQ;AAAA,IACrB,CAAC,aAAa,SAAS;AAAA,EACzB,EACA,OAAO,CAAC,CAAC,EAAE,CAAC,MAAM,sBAAsB,CAAC,KAAK,IAAI;AACpD,aAAW,CAAC,MAAM,CAAC,KAAK,eAAe;AACrC,UAAM,MAAM,sBAAsB,CAAC;AACnC,UAAM,SAAS,OAAO,IAAI,SAAS,UAAU,IAAI,KAAK,IAAI,CAAC,MAAM;AACjE,YAAQ;AAAA,MACN,uBAAuB,IAAI,yDAAoD,MAAM;AAAA,IACvF;AAAA,EACF;AACA,MAAI,oBAAoB,cAAc,SAAS,GAAG;AAChD;AAAA,MACE,GAAG,cAAc,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,EAAE,KAAK,IAAI,CAAC;AAAA,IAC7C;AAAA,EACF;AAEA,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;AASA,MAAI,mBAAmB;AACvB,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,WAAW,QAAQ,IAAI;AACjC,UAAM,UAAU,EAAE,kBAAkB;AACpC,UAAM,UAAU,EAAE,mBAAmB;AACrC,QAAI,UAAU,KAAK,UAAU,GAAG;AAC9B,yBAAmB;AACnB,YAAM,QAAQ,EAAE,kBAAkB,KAAK;AACvC,YAAM,QAAQ,EAAE,mBAAmB,KAAK;AACxC,cAAQ;AAAA,QACN,+BAA+B,IAAI,oCACpB,OAAO,IAAI,IAAI,eAAe,OAAO,IAAI,IAAI;AAAA,MAE9D;AAAA,IACF;AASA,UAAM,WAAW,EAAE,mBAAmB;AACtC,UAAM,WAAW,EAAE,oBAAoB;AACvC,UAAM,UAAU,EAAE,KAAK;AACvB,UAAM,YAAY,KAAK,IAAI,EAAE,kBAAkB,GAAG,EAAE,mBAAmB,CAAC;AACxE,UAAM,gBACH,WAAW,KAAK,WAAW,OAAO,aAAa,YAAY,UAAU;AACxE,QAAI,cAAc;AAChB,cAAQ;AAAA,QACN,4BAA4B,IAAI,qDACjB,QAAQ,mBAAmB,QAAQ,SAAS,OAAO;AAAA,MAEpE;AAAA,IACF;AAAA,EACF;AAEA,MAAI,oBAAoB,sBAAsB,MAAM;AAClD,YAAQ;AAAA,MACN;AAAA,IACF;AAAA,EACF;AAKA,MAAI,sBAAsB,MAAM;AAC9B,UAAM,SAAS,4BAA4B,YAAY,kBAAkB;AACzE,QAAI,OAAQ,cAAa,MAAM;AAAA,EACjC;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,QAAqB;AAC7D,QAAM,EAAE,YAAY,WAAW,YAAY,aAAa,aAAa,MAAM,IACzE,cAAc,IAAI;AAEpB,MAAI,MAAM,SAAS,GAAG;AACpB,YAAQ,MAAM,sEAAsE;AACpF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAKA,MAAI,aAAa,QAAQ;AACvB,YAAQ;AAAA,MACN;AAAA,IACF;AACA,YAAQ,KAAK,CAAC;AAAA,EAChB;AAIA,QAAM,kBAAkB,eAAe,OAAO,sBAAsB;AAEpE,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,UAAI,OAAO,QAAS;AACpB,eAAS;AAAA,QACP,MAAM,QAAQ,YAAY,WAAW,SAAS,SAAS;AAAA,UACrD;AAAA,UACA,GAAI,mBAAmB,OAAO,EAAE,aAAa,gBAAgB,IAAI,CAAC;AAAA,QACpE,CAAC;AAAA,MACH;AAAA,IACF;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;AACrC,UAAM,EAAE,QAAQ,QAAQ,QAAQ,IAAI,mBAAmB,GAAG,MAAM,OAAO,KAAK;AAM5E,UAAM,kBAA4B,CAAC;AACnC,QAAI,SAAS,EAAG,iBAAgB,KAAK,GAAG,MAAM,IAAI,SAAS,MAAM,YAAY,MAAM,SAAS;AAC5F,QAAI,UAAU,EAAG,iBAAgB,KAAK,GAAG,OAAO,MAAM;AACtD,UAAM,aAAa,gBAAgB,SAAS,IAAI,OAAO,gBAAgB,KAAK,IAAI,CAAC,MAAM;AAEvF,QAAI,WAAW,GAAG;AAChB,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,GAAG,UAAU;AAAA,MACvL;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,GAAG,UAAU;AAAA,MACpP;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;AAG/D,UAAM,EAAE,QAAQ,OAAO,IAAI,mBAAmB,GAAG,IAAI;AACrD,UAAM,aACJ,SAAS,IAAI,OAAO,MAAM,IAAI,SAAS,MAAM,YAAY,MAAM,aAAa;AAC9E,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,GAAG,UAAU;AAAA,IAC5I;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;AAOA,SAAS,eAAe,QAAoB,OAAwB;AAClE,QAAM,IAAI,OAAO,QAAQ;AACzB,MAAI,CAAC,KAAK,EAAE,WAAW,EAAG,QAAO;AACjC,aAAW,KAAK,GAAG;AACjB,UAAM,YAAY,EAAE,SAAS,EAAE;AAC/B,UAAM,WACJ,EAAE,SAAS,kBACP,iCACA,IAAI,EAAE,QAAQ,KAAK,QAAQ,CAAC,CAAC;AACnC,YAAQ;AAAA,MACN,wBAAwB,KAAK,YAAY,EAAE,MAAM,MAAM,EAAE,IAAI,aACvD,EAAE,OAAO,KAAK,QAAQ,CAAC,CAAC,MAAM,EAAE,MAAM,IAAI,SAAS,eAAe,QAAQ;AAAA,IAClF;AAAA,EACF;AACA,SAAO;AACT;AAaA,SAAS,0BAA0B,QAAoB,OAAqB;AAC1E,aAAW,CAAC,MAAM,KAAK,KAAK,OAAO,QAAQ,OAAO,QAAQ,OAAO,GAAG;AAClE,UAAM,EAAE,QAAQ,QAAQ,QAAQ,IAAI,mBAAmB,OAAO,MAAM,OAAO,KAAK;AAChF,QAAI,UAAU,KAAK,SAAS,WAAW,GAAG;AACxC,cAAQ;AAAA,QACN,oBAAoB,KAAK,YAAY,IAAI,qCACnC,OAAO;AAAA,MAEf;AAAA,IACF;AAAA,EACF;AACF;AAeA,SAAS,mBAAmB,QAAoB,OAAwB;AACtE,QAAM,EAAE,OAAO,SAAS,IAAI,OAAO;AACnC,MAAI,UAAU,KAAK,WAAW,MAAO,QAAO;AAC5C,UAAQ;AAAA,IACN,sBAAsB,KAAK,eAAU,KAAK;AAAA,EAE5C;AACA,SAAO;AACT;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,eAAe,eAAe,MAAgB,QAAqB;AACjE,QAAM;AAAA,IACJ;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,IAAI,cAAc,IAAI;AAEtB,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;AAKA,MAAI,aAAa,UAAU,UAAU,SAAS,GAAG;AAC/C,YAAQ,MAAM,yDAAyD,UAAU,MAAM,IAAI;AAC3F,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;AAGhC,UAAI,OAAO,QAAS;AACpB,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;AAMA,mBAAW,cACT,eAAe,OAAO,sBAAsB,KAAK,WAAW,eAAe;AAK7E,YAAI,aAAa,QAAQ;AACvB,gBAAM,YAAY,WAAW,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI;AACtD,gBAAM,UAAU,YAAY,OAAO,CAAC,MAAM,CAAC,UAAU,SAAS,CAAC,CAAC;AAChE,cAAI,QAAQ,QAAQ;AAClB,oBAAQ;AAAA,cACN,wCAAwC,QAAQ,KAAK,IAAI,CAAC,gBAAgB,UAAU,KAAK,IAAI,CAAC;AAAA,YAChG;AACA,oBAAQ,KAAK,CAAC;AAAA,UAChB;AACA,qBAAW,UAAU,WAAW,QAAQ,OAAO,CAAC,MAAM,YAAY,SAAS,EAAE,IAAI,CAAC;AAClF,cAAI,WAAW,QAAQ,WAAW,GAAG;AACnC,oBAAQ,MAAM,sCAAsC;AACpD,oBAAQ,KAAK,CAAC;AAAA,UAChB;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;AAWA,cAAM,aAAa,EAAE,QAAQ,GAAI,gBAAgB,EAAE,eAAe,KAAK,IAAI,CAAC,EAAG;AAM/E,cAAM,aAAa,aAAa,SAAS,WAAW,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI,IAAI;AACjF,cAAM,gBAAgB,CAAC,WAAuB;AAC5C,cAAI,YAAY;AACd,mBAAO,SAAS,iBAAiB;AACjC,mBAAO,SAAS,aAAa;AAAA,UAC/B;AAAA,QACF;AAEA,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,gBAAI,OAAO,QAAS;AACpB,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,4BAAc,MAAM;AACpB,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;AAK1C,cAAI,aAAa;AACjB,qBAAW,KAAK,YAAY;AAE1B,kBAAM,UAAU,mBAAmB,GAAG,QAAQ;AAC9C,kBAAM,WAAW,eAAe,GAAG,QAAQ;AAC3C,sCAA0B,GAAG,QAAQ;AACrC,gBAAI,WAAW,SAAU,cAAa;AAAA,UACxC;AACA,cAAI,cAAc,CAAC,QAAS;AAAA,QAC9B,OAAO;AACL,gBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,SAAS,UAAU;AAC7E,wBAAc,MAAM;AACpB,kBAAQ,KAAK,MAAM;AAEnB,kBAAQ,IAAI,OAAO,YAAY,MAAM,IAAI,IAAI;AAI7C,gBAAM,UAAU,mBAAmB,QAAQ,QAAQ;AACnD,gBAAM,WAAW,eAAe,QAAQ,QAAQ;AAChD,oCAA0B,QAAQ,QAAQ;AAC1C,cAAI,WAAW,SAAU;AAAA,QAC3B;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":[]}
+{"version":3,"sources":["../src/cli.ts","../src/cli-utils.ts","../src/cli-validate.ts","../src/cli-args.ts","../src/cli-format.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, evaluateScorerErrorRateGate } 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';\nimport { parseEvalArgs, envInt } from './cli-args.js';\nimport { formatModelTimingLines } from './cli-format.js';\nimport { scorerCounts } from './utils.js';\n\n/**\n * Refuse to certify a comparison and exit non-zero, with one consistent\n * message format. Routes both the scorer-filtered guard and the\n * `--max-scorer-error-rate` failure-rate gate through a single path so the\n * \"why CI failed\" line always reads the same.\n */\nfunction refuseToGate(reason: string): never {\n  console.error(`[axl-eval] Refusing to gate: ${reason}`);\n  process.exit(1);\n}\n\n/**\n * Read a scorer's `scored`/`failed` counts from a summary entry, falling back\n * to recomputing from `items` when the fields are absent (pre-0.18.0\n * artifacts). `items` is omitted for multi-run summaries that carry no items —\n * there the fields are authoritative (summed by `aggregateRuns`).\n */\nfunction scorerSampleCounts(\n  entry: { scored?: number; failed?: number; skipped?: number },\n  name: string,\n  items?: EvalResult['items'],\n): { scored: number; failed: number; skipped: number } {\n  if (entry.scored != null && entry.failed != null) {\n    // `scored`/`failed` are authoritative; `skipped` was added later, so an\n    // 0.18.0-era artifact may carry the first two without it — recompute that one\n    // field from items when we can, else 0.\n    const skipped = entry.skipped ?? (items ? scorerCounts(items, name).skipped : 0);\n    return { scored: entry.scored, failed: entry.failed, skipped };\n  }\n  if (items) return scorerCounts(items, name);\n  return { scored: entry.scored ?? 0, failed: entry.failed ?? 0, skipped: entry.skipped ?? 0 };\n}\n\n/** Install a SIGINT/SIGTERM handler that aborts an in-flight eval gracefully.\n *  First signal: aborts the controller so runEval / rescore propagate cancellation\n *  through ScorerContext.signal into in-flight provider.chat calls. The eval\n *  unwinds normally and runtime.shutdown() runs in the `finally` block.\n *  Second signal: hard exit — the user has signalled twice; cleanup is hung\n *  or taking too long, so we abandon gracefulness rather than appear frozen. */\nfunction installAbortHandler(controller: AbortController): void {\n  let signalled = false;\n  const onSignal = (sig: NodeJS.Signals) => {\n    if (signalled) {\n      console.error(`\\n[axl-eval] Second ${sig} — forcing exit.`);\n      process.exit(130);\n    }\n    signalled = true;\n    console.error(`\\n[axl-eval] ${sig} received — cancelling eval (press again to force exit)...`);\n    controller.abort();\n  };\n  process.on('SIGINT', onSignal);\n  process.on('SIGTERM', onSignal);\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> --concurrency <n>       Override item concurrency (flag > env\n                                          AXL_EVAL_CONCURRENCY > config > 5).\n                                          Max concurrent scorer calls is\n                                          concurrency x scorerConcurrency (the\n                                          per-eval scorerConcurrency defaults to 5).\n  axl-eval <path> --scorers <a,b>         Run only the named scorers (single eval\n                                          file only). Stamps the result as a\n                                          subset so compare won't treat it as a\n                                          full baseline.\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  axl-eval compare <a> <b> --max-scorer-error-rate <0..1>  Exit 1 if a scorer's\n                                          failure rate exceeds the limit on\n                                          either side (deterministic scorers\n                                          tolerate zero failures). Always warns\n                                          when either side rests on a thinned\n                                          sample, even without this flag.\n\nConfig auto-detection (when --config is not specified):\n  ${CONFIG_CANDIDATES.join(' -> ')}\n\nWhen a config is found, the exported AxlRuntime is passed to executeWorkflow\nand cost is tracked automatically via runtime.trackCost().\nWhen no config is found, a bare AxlRuntime is created (providers from env vars).\n`);\n    process.exit(0);\n  }\n\n  if (args[0] === 'compare') {\n    // `compare` is pure computation — no LLM calls, no AbortSignal plumbing needed.\n    await runCompare(args.slice(1));\n    return;\n  }\n\n  // SIGINT/SIGTERM aborts in-flight LLM calls (eval + rescore) instead of leaving\n  // the process to hard-exit mid-request, which would skip runtime.shutdown() and\n  // leak streaming-buffer state.\n  const controller = new AbortController();\n  installAbortHandler(controller);\n\n  if (args[0] === 'rescore') {\n    await runRescore(args.slice(1), controller.signal);\n    return;\n  }\n\n  await runEvalCommand(args, controller.signal);\n}\n\nfunction parseThresholdArg(args: string[]): Record<string, number> | number | undefined {\n  const idx = args.indexOf('--threshold');\n  if (idx === -1 || idx + 1 >= args.length) return undefined;\n  const value = args[idx + 1];\n\n  // Global numeric threshold: --threshold 0.05\n  if (/^[\\d.]+$/.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\n/**\n * Parse `--max-scorer-error-rate <0..1>`. Returns the numeric limit, or\n * `undefined` when the flag is absent. Exits non-zero on a malformed value\n * (fail loud — a typo'd gate threshold must not silently disable the gate).\n */\nfunction parseMaxScorerErrorRate(args: string[]): number | undefined {\n  const idx = args.indexOf('--max-scorer-error-rate');\n  if (idx === -1) return undefined;\n  if (idx + 1 >= args.length) {\n    console.error('Error: --max-scorer-error-rate requires a value in [0, 1]');\n    process.exit(1);\n  }\n  const raw = args[idx + 1];\n  // Strict numeric parse — `parseFloat('0.5abc')` returns 0.5 and `parseFloat('0,5')`\n  // returns 0, both of which would silently set an unintended gate threshold. Reject\n  // anything that isn't a clean decimal (matches `parseThresholdArg`'s guard).\n  const n = /^\\d*\\.?\\d+$/.test(raw.trim()) ? Number(raw) : NaN;\n  if (!Number.isFinite(n) || n < 0 || n > 1) {\n    console.error(`Error: --max-scorer-error-rate must be a number in [0, 1], got \"${raw}\"`);\n    process.exit(1);\n  }\n  return n;\n}\n\n/**\n * If a raw compare input (single result or multi-run array) was produced by a\n * `--scorers` filtered run, return the scorer names it ran; otherwise undefined.\n * Reads the first run's metadata — a multi-run group shares the same stamp.\n */\nfunction scorerFilteredScorers(r: EvalResult | EvalResult[]): string[] | undefined {\n  const first = Array.isArray(r) ? r[0] : r;\n  const meta = first?.metadata;\n  if (!meta || meta.scorerFiltered !== true) return undefined;\n  return Array.isArray(meta.scorersRun) ? (meta.scorersRun as string[]) : [];\n}\n\nasync function runCompare(args: string[]) {\n  const failOnRegression = args.includes('--fail-on-regression');\n  const thresholds = parseThresholdArg(args);\n  const maxScorerErrorRate = parseMaxScorerErrorRate(args);\n  // Exclude flags and the values consumed by value-taking flags so neither a\n  // threshold nor an error-rate value is mistaken for a result file path.\n  const valueFlags = new Set(['--threshold', '--max-scorer-error-rate']);\n  const files = args.filter(\n    (a, i) => !a.startsWith('--') && !(i > 0 && valueFlags.has(args[i - 1])),\n  );\n\n  if (files.length !== 2) {\n    console.error(\n      'Usage: axl-eval compare <baseline.json> <candidate.json> [--threshold <value>] [--fail-on-regression] [--max-scorer-error-rate <0..1>]',\n    );\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\n  // A scorer-filtered result (CLI `--scorers`) tested only a subset of scorers,\n  // so it isn't a sound baseline/candidate for a full comparison. Check this\n  // BEFORE evalCompare(): a subset-vs-full comparison has mismatched scorer sets\n  // and evalCompare() would throw a generic \"different scorers\" error that never\n  // mentions --scorers. Warn loudly here so the cause is actionable; under\n  // `--fail-on-regression`, refuse outright (a swallowed stderr line wouldn't\n  // actually prevent the footgun).\n  const filteredSides = (\n    [\n      ['baseline', baseline],\n      ['candidate', candidate],\n    ] as const\n  ).filter(([, r]) => scorerFilteredScorers(r) != null);\n  for (const [side, r] of filteredSides) {\n    const ran = scorerFilteredScorers(r);\n    const ranStr = ran && ran.length ? ` (ran: ${ran.join(', ')})` : '';\n    console.error(\n      `[axl-eval] WARNING: ${side} was run with --scorers — it tested only a subset${ranStr}, so this comparison is incomplete.`,\n    );\n  }\n  if (failOnRegression && filteredSides.length > 0) {\n    refuseToGate(\n      `${filteredSides.map(([s]) => s).join(', ')} was run with --scorers (a scorer subset); re-run without --scorers for a complete baseline.`,\n    );\n  }\n\n  const compareOptions = thresholds != null ? { thresholds } : undefined;\n  const comparison = evalCompare(baseline, candidate, compareOptions);\n\n  console.log(\n    `\\nCompare: baseline (${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: ${stable}\\n`,\n  );\n\n  // Scorer failure-rate signal. Always WARN loudly when either side rests on a\n  // thinned sample (a scorer ran and failed on some items) — a regression check\n  // computed over survivors can look green while half the judges errored. Only\n  // REFUSE to gate when the user opted in with `--max-scorer-error-rate` AND a\n  // scorer is over tolerance (type-aware: deterministic scorers tolerate zero\n  // failures, LLM judges use the flag value), or a scorer produced no scored\n  // items at all on a side (a zero-sample mean can't be certified).\n  let sawThinnedSample = false;\n  for (const name of scorerNames) {\n    const s = comparison.scorers[name];\n    const bFailed = s.baselineFailed ?? 0;\n    const cFailed = s.candidateFailed ?? 0;\n    if (bFailed > 0 || cFailed > 0) {\n      sawThinnedSample = true;\n      const bAtt = (s.baselineScored ?? 0) + bFailed;\n      const cAtt = (s.candidateScored ?? 0) + cFailed;\n      console.error(\n        `[axl-eval] WARNING: scorer \"${name}\" failed on some items ` +\n          `(baseline ${bFailed}/${bAtt}, candidate ${cFailed}/${cAtt}); ` +\n          `the means above are computed over the surviving sample.`,\n      );\n    }\n    // Conditional scorers (`applies`) skip N/A items. Skips never move the\n    // delta (excluded from both means), but they make the two per-side means\n    // cover different applicable subsets when the sides skipped different COUNTS,\n    // OR the same count of DIFFERENT items — the latter shows up as a paired\n    // shortfall: fewer items scored on BOTH sides (`n`) than either side scored\n    // alone. (`bSkipped !== cSkipped` alone would miss the disjoint-but-equal\n    // case.) Note it (advisory only; not a gate; failures are covered by the\n    // separate WARNING above).\n    const bSkipped = s.baselineSkipped ?? 0;\n    const cSkipped = s.candidateSkipped ?? 0;\n    const pairedN = s.n ?? 0;\n    const minScored = Math.min(s.baselineScored ?? 0, s.candidateScored ?? 0);\n    const skipsDiverge =\n      (bSkipped > 0 || cSkipped > 0) && (bSkipped !== cSkipped || pairedN < minScored);\n    if (skipsDiverge) {\n      console.error(\n        `[axl-eval] NOTE: scorer \"${name}\" applies to different subsets per side ` +\n          `(baseline ${bSkipped} N/A, candidate ${cSkipped} N/A; ${pairedN} item(s) scored on both); ` +\n          `the two means cover different applicable subsets, so compare the delta with care.`,\n      );\n    }\n  }\n  // Tell the user how to make the advisory warning blocking (when they haven't).\n  if (sawThinnedSample && maxScorerErrorRate == null) {\n    console.error(\n      `[axl-eval] (advisory — pass --max-scorer-error-rate <0..1> to fail CI on a thinned sample)`,\n    );\n  }\n\n  // Gate-side refusal (opt-in). The decision lives in evaluateScorerErrorRateGate\n  // (pure + unit-tested) so the same type-aware tolerance + zero-sample logic\n  // isn't re-implemented here and at the source-side gate in runEval.\n  if (maxScorerErrorRate != null) {\n    const reason = evaluateScorerErrorRateGate(comparison, maxScorerErrorRate);\n    if (reason) refuseToGate(reason);\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[], signal: AbortSignal) {\n  const { outputPath, configArg, conditions, concurrency, scorerNames, paths } =\n    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  // --scorers is a run-command-only filter — rescore always re-runs the full\n  // scorer set from the eval file. Fail loudly rather than silently ignoring it\n  // (it parses as a known flag, so it wouldn't trip the unknown-flag guard).\n  if (scorerNames?.length) {\n    console.error(\n      'Error: --scorers is not supported with rescore; pass a scorer subset in the eval file instead.',\n    );\n    process.exit(1);\n  }\n\n  // Honor the same item-concurrency override as the run command (flag > env).\n  // scorerConcurrency keeps its rescore default (5).\n  const itemConcurrency = concurrency ?? envInt('AXL_EVAL_CONCURRENCY');\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      if (signal.aborted) break;\n      rescored.push(\n        await rescore(resultData, evalConfig.scorers, runtime, {\n          signal,\n          ...(itemConcurrency != null ? { concurrency: itemConcurrency } : {}),\n        }),\n      );\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)}  ${'p95'.padStart(colWidth)}`,\n  );\n\n  const ruleLen = maxNameLen + 2 + (colWidth + 2) * 5;\n  lines.push('  ' + '\\u2500'.repeat(ruleLen));\n\n  for (const name of scorerNames) {\n    const s = result.summary.scorers[name];\n    const { scored, failed, skipped } = scorerSampleCounts(s, name, result.items);\n    // Loud trailing annotation: `failed > 0` means the mean rests on a thinned\n    // sample; `skipped > 0` (an `applies`-conditional scorer) labels items the\n    // scorer deliberately didn't run — crucial so a fully-skipped scorer (all\n    // `--`) reads as \"N/A\" rather than \"broken\". Both shown only when non-zero so\n    // clean, unconditional runs stay uncluttered.\n    const annotationParts: string[] = [];\n    if (failed > 0) annotationParts.push(`${scored}/${scored + failed} scored, ${failed} failed`);\n    if (skipped > 0) annotationParts.push(`${skipped} N/A`);\n    const annotation = annotationParts.length > 0 ? `   (${annotationParts.join(', ')})` : '';\n    // Detect scorers with no valid scores (all items errored → computeStats([]) → all zeros)\n    if (scored === 0) {\n      lines.push(\n        `  ${name.padEnd(maxNameLen)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}  ${'--'.padStart(colWidth)}${annotation}`,\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)}  ${s.p95.toFixed(2).padStart(colWidth)}${annotation}`,\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  // Per-model provider latency, indented under the wall-clock Timing row. All\n  // four buckets are shown because their separation is the point: `wire` is the\n  // provider, `queued` is the SDK's own rate limiter, `retries` is the\n  // provider's throttling that day, and `first token` is the figure that\n  // actually discriminates between models. See `formatModelTimingLines` for why\n  // these are the call-weighted means and not the per-item distributions.\n  lines.push(...formatModelTimingLines(result.summary.modelTiming, maxNameLen));\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    // Multi-run summaries have no items, so the summed scored/failed fields are\n    // authoritative (no fallback recompute possible). Annotate when failed > 0.\n    const { scored, failed } = scorerSampleCounts(s, name);\n    const annotation =\n      failed > 0 ? `   (${scored}/${scored + failed} scored, ${failed} failed)` : '';\n    lines.push(\n      `  ${name.padEnd(maxNameLen)}  ${meanStd.padStart(colWidth)}  ${s.min.toFixed(3).padStart(8)}  ${s.max.toFixed(3).padStart(8)}${annotation}`,\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: ${durationStr}s`);\n\n  return lines.join('\\n');\n}\n\n/**\n * Print each degraded scorer (set by `runEval` when `failOnScorerErrorRate`\n * tripped) to stderr and return whether the run was degraded, so the caller can\n * count the file as failed (→ non-zero exit). Quiet when nothing degraded.\n */\nfunction reportDegraded(result: EvalResult, label: string): boolean {\n  const d = result.summary.degraded;\n  if (!d || d.length === 0) return false;\n  for (const x of d) {\n    const attempted = x.scored + x.failed;\n    const limitStr =\n      x.type === 'deterministic'\n        ? 'deterministic zero-tolerance'\n        : `${(x.limit * 100).toFixed(1)}% limit`;\n    console.error(\n      `[axl-eval] DEGRADED: ${label} scorer \"${x.scorer}\" (${x.type}) failed ` +\n        `${(x.rate * 100).toFixed(1)}% (${x.failed}/${attempted}), over the ${limitStr}.`,\n    );\n  }\n  return true;\n}\n\n/**\n * Advisory (non-fatal): a conditional scorer (`applies`) that ended up applicable\n * to ZERO items — every item was skipped as N/A, nothing scored or failed. Its\n * mean is a meaningless empty-sample 0 and the failure-rate gate can't assess it\n * (a zero-sample scorer is never `degraded`). That's almost always a\n * misconfiguration: an over-strict predicate, a dataset missing the flagged\n * items, or a predicate that forgot to `return`. The `(N N/A)` table annotation\n * hints at it, but a fully-skipped GATE scorer would otherwise exit 0 with no\n * stderr signal — this closes that silent-green hole without changing the exit\n * code (skips are legitimately conditional; we inform, we don't fail).\n */\nfunction reportFullySkippedScorers(result: EvalResult, label: string): void {\n  for (const [name, entry] of Object.entries(result.summary.scorers)) {\n    const { scored, failed, skipped } = scorerSampleCounts(entry, name, result.items);\n    if (skipped > 0 && scored + failed === 0) {\n      console.error(\n        `[axl-eval] NOTE: ${label} scorer \"${name}\" was not applicable to any item ` +\n          `(${skipped} N/A, 0 scored) — its mean is not meaningful. Check the scorer's ` +\n          `\\`applies\\` predicate and dataset if you expected it to run.`,\n      );\n    }\n  }\n}\n\n/**\n * A run where EVERY item errored in the workflow (0 succeeded) produced no valid\n * output to score — the eval is meaningless and must never pass CI green. This\n * is distinct from (and complementary to) the scorer failure-rate gate: that one\n * is about a flaky/failing *scorer*, this is about a broken *workflow*. The\n * scorer gate deliberately ignores a zero-sample scorer (nothing ran), so without\n * this guard a 100%-workflow-error run exits 0 — a silent-green trap.\n *\n * Non-configurable on purpose: a 0%-success eval is unambiguously broken, so this\n * always fails. (A configurable `failOnItemErrorRate` for PARTIAL workflow-failure\n * gating is a reasonable future opt-in; the per-item failure count is already shown\n * loudly in the table either way.) Returns whether the run was a total wipeout.\n */\nfunction reportTotalWipeout(result: EvalResult, label: string): boolean {\n  const { count, failures } = result.summary;\n  if (count === 0 || failures < count) return false;\n  console.error(\n    `[axl-eval] FAILED: ${label} — all ${count} item(s) errored in the workflow ` +\n      `(0 succeeded); the eval produced no scorable output.`,\n  );\n  return true;\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// ── Main eval command ──────────────────────────────────────────────\n\nasync function runEvalCommand(args: string[], signal: AbortSignal) {\n  const {\n    outputPath,\n    configArg,\n    conditions,\n    runs,\n    captureTraces,\n    concurrency,\n    scorerNames,\n    paths,\n  } = 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  // --scorers is single-file only — filtering by name across a glob/dir of evals\n  // with different scorer sets is ambiguous. Fail fast, before importing any\n  // module, rather than discovering the conflict mid-loop.\n  if (scorerNames?.length && evalFiles.length > 1) {\n    console.error(`Error: --scorers requires a single eval file (matched ${evalFiles.length}).`);\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      // Short-circuit if the user aborted between eval files. The in-flight eval\n      // unwinds on its own via runEval({ signal }); this prevents starting the next one.\n      if (signal.aborted) break;\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        // Concurrency override precedence: flag > env > config > default.\n        // Mutating evalConfig in place mirrors how this loop already stamps\n        // result.metadata. `concurrency` is pure scheduling — never changes\n        // results — so a per-run override is always safe.\n        evalConfig.concurrency =\n          concurrency ?? envInt('AXL_EVAL_CONCURRENCY') ?? evalConfig.concurrency ?? 5;\n\n        // --scorers: run a subset of scorers for a focused iteration loop.\n        // (The single-file guard already ran before the loop.) Validate-then-\n        // filter so the error lists every available name.\n        if (scorerNames?.length) {\n          const available = evalConfig.scorers.map((s) => s.name);\n          const unknown = scorerNames.filter((n) => !available.includes(n));\n          if (unknown.length) {\n            console.error(\n              `Error: --scorers: unknown scorer(s): ${unknown.join(', ')}. Available: ${available.join(', ')}`,\n            );\n            process.exit(1);\n          }\n          evalConfig.scorers = evalConfig.scorers.filter((s) => scorerNames.includes(s.name));\n          if (evalConfig.scorers.length === 0) {\n            console.error('Error: --scorers matched no scorers.');\n            process.exit(1);\n          }\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        // Always pass the SIGINT-driven signal so Ctrl+C aborts in-flight\n        // workflows + scorer LLM calls instead of dropping the connection mid-request.\n        const runOptions = { signal, ...(captureTraces ? { captureTraces: true } : {}) };\n\n        // Stamp filtered runs so a scorer-subset result can't be silently used\n        // as a full baseline. `compare` warns (and refuses to gate) on it. Must\n        // be applied to EVERY run in a multi-run batch — Studio's multi-run\n        // builder spreads only run[0]'s metadata.\n        const scorersRun = scorerNames?.length ? evalConfig.scorers.map((s) => s.name) : undefined;\n        const stampFiltered = (result: EvalResult) => {\n          if (scorersRun) {\n            result.metadata.scorerFiltered = true;\n            result.metadata.scorersRun = scorersRun;\n          }\n        };\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            if (signal.aborted) break;\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              stampFiltered(result);\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\n          // Failure-rate gate (opt-in via failOnScorerErrorRate) + total-wipeout\n          // guard (always). Report every run; count the file as failed unless it\n          // was already counted as a partial batch above (avoids double-count).\n          let anyFailing = false;\n          for (const r of runResults) {\n            // Call both (no short-circuit) so each prints its own diagnostic.\n            const wipeout = reportTotalWipeout(r, filePath);\n            const degraded = reportDegraded(r, filePath);\n            reportFullySkippedScorers(r, filePath); // advisory only\n            if (wipeout || degraded) anyFailing = true;\n          }\n          if (anyFailing && !partial) failedFiles++;\n        } else {\n          const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);\n          stampFiltered(result);\n          results.push(result);\n\n          console.log('\\n' + formatTable(result) + '\\n');\n          // Total-wipeout guard (always) + scorer failure-rate gate (opt-in). They\n          // are mutually exclusive (a wipeout has no scored items, so no scorer can\n          // be degraded), but report both for clarity; either fails the file.\n          const wipeout = reportTotalWipeout(result, filePath);\n          const degraded = reportDegraded(result, filePath);\n          reportFullySkippedScorers(result, filePath); // advisory only\n          if (wipeout || degraded) failedFiles++;\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  // Concurrency knobs must be positive integers when present. The worker pool\n  // defensively clamps non-positive/non-finite values to 1 at runtime, but a\n  // config that wrote `concurrency: 0` (or NaN via `Number(env)`) almost\n  // certainly has a bug — surface it loudly, matching the `--concurrency` flag.\n  for (const key of ['concurrency', 'scorerConcurrency'] as const) {\n    const v = (c as Record<string, unknown>)[key];\n    if (v !== undefined && (typeof v !== 'number' || !Number.isInteger(v) || v < 1)) {\n      return `exports an invalid ${key} (${typeof v === 'number' ? v : typeof v}) — must be a positive integer.`;\n    }\n  }\n  // The failure-rate gate must fail loud on a bad value, symmetric with the\n  // `--max-scorer-error-rate` flag. Without this, a typo (e.g. `5` meaning \"5%\")\n  // only `console.warn`s at runtime and silently disables the gate the user\n  // explicitly opted into — the exact silent-no-op the feature exists to catch.\n  {\n    const v = (c as Record<string, unknown>).failOnScorerErrorRate;\n    if (v !== undefined && (typeof v !== 'number' || !Number.isFinite(v) || v < 0 || v > 1)) {\n      return `exports an invalid failOnScorerErrorRate (${typeof v === 'number' ? v : typeof v}) — must be a number between 0 and 1.`;\n    }\n  }\n  return undefined;\n}\n","/**\n * Argument parsing for the `axl-eval` CLI, factored out of `cli.ts` so it can be\n * unit-tested without importing `cli.ts` (whose module body runs `main()` on\n * import).\n */\n\nexport const KNOWN_FLAGS = new Set([\n  '--output',\n  '--config',\n  '--conditions',\n  '--fail-on-regression',\n  '--threshold',\n  '--runs',\n  '--capture-traces',\n  '--concurrency',\n  '--scorers',\n]);\n\n/**\n * The subset of flags `parseEvalArgs` consumes a following value for. NOT every\n * value-taking flag in the CLI: `--threshold` takes a value too but is parsed\n * by `runCompare`'s own parser, not here. Every entry MUST also be in\n * {@link KNOWN_FLAGS} — a value-flag missing from KNOWN_FLAGS would be rejected\n * as \"unknown\", and one missing from VALUE_FLAGS would silently swallow its\n * value into `paths`. A test asserts `VALUE_FLAGS ⊆ KNOWN_FLAGS`.\n */\nexport const VALUE_FLAGS = new Set([\n  '--output',\n  '--config',\n  '--conditions',\n  '--runs',\n  '--concurrency',\n  '--scorers',\n]);\n\nexport type ParsedEvalArgs = {\n  outputPath?: string;\n  configArg?: string;\n  conditions: string[];\n  runs: number;\n  captureTraces: boolean;\n  /** Item-level concurrency override (flag value). Clamped to >= 1. */\n  concurrency?: number;\n  /** Scorer names from `--scorers` (deduped, in first-seen order). */\n  scorerNames?: string[];\n  paths: string[];\n};\n\n/**\n * Read a positive integer from an environment variable. Returns `undefined`\n * when the var is absent, non-numeric, or `<= 0`, so a malformed value falls\n * through to the next precedence tier rather than poisoning the run.\n */\nexport function envInt(name: string): number | undefined {\n  const raw = process.env[name];\n  if (raw == null || raw === '') return undefined;\n  const n = parseInt(raw, 10);\n  if (!Number.isFinite(n) || n <= 0) return undefined;\n  return n;\n}\n\nexport function parseEvalArgs(args: string[]): ParsedEvalArgs {\n  let outputPath: string | undefined;\n  let configArg: string | undefined;\n  let conditions: string[] = [];\n  let runs = 1;\n  let captureTraces = false;\n  let concurrency: number | undefined;\n  let scorerNames: string[] | undefined;\n  const paths: string[] = [];\n\n  for (let i = 0; i < args.length; i++) {\n    const arg = args[i];\n    if (VALUE_FLAGS.has(arg)) {\n      if (i + 1 >= args.length) {\n        console.error(`Error: ${arg} requires a value`);\n        process.exit(1);\n      }\n      const value = args[++i];\n      if (arg === '--output') outputPath = value;\n      else if (arg === '--config') configArg = value;\n      else if (arg === '--runs') runs = Math.max(1, parseInt(value, 10) || 1);\n      else if (arg === '--concurrency') {\n        // Clamp to a floor of 1 with a warning (matches `--runs`' clamp ethos\n        // but surfaces the typo). A `0` here would otherwise spawn zero workers.\n        // Note the intentional asymmetry with `envInt` above: an explicit flag\n        // typo is worth a warning, whereas a malformed ambient env var falls\n        // through silently to the next precedence tier.\n        const n = parseInt(value, 10);\n        if (!Number.isFinite(n) || n <= 0) {\n          console.error(`[axl-eval] Ignoring invalid --concurrency \"${value}\"; using 1.`);\n          concurrency = 1;\n        } else {\n          concurrency = n;\n        }\n      } else if (arg === '--scorers') {\n        scorerNames = [\n          ...new Set(\n            value\n              .split(',')\n              .map((s) => s.trim())\n              .filter(Boolean),\n          ),\n        ];\n      } else {\n        conditions = value\n          .split(',')\n          .map((c) => c.trim())\n          .filter(Boolean);\n      }\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 {\n    outputPath,\n    configArg,\n    conditions,\n    runs,\n    captureTraces,\n    concurrency,\n    scorerNames,\n    paths,\n  };\n}\n","/**\n * Terminal rendering helpers for the `axl-eval` CLI, factored out of `cli.ts` so\n * they can be unit-tested without importing `cli.ts` (whose module body runs\n * `main()` on import) — same reason as `cli-args.ts`.\n */\n\nimport type { ModelTimingStats } from './types.js';\n\n/**\n * Render the per-model provider-latency rows that sit under the wall-clock\n * `Timing` row.\n *\n * Every figure is per PROVIDER CALL — the whole `modelTiming` surface is now\n * one kind of average, so there is nothing here a reader can accidentally\n * divide by the call count and get wrong.\n *\n * `wire` and `first token` carry a `mean/p95` pair because those are the two\n * model-comparison figures and a tail matters for both. `queued` and `retries`\n * print a mean only: they describe Axl's own limiter and the provider's\n * throttling on the day of the run, so a p95 would lengthen every row without\n * changing a model choice. The full distributions stay on the JSON artifact.\n *\n * Units are milliseconds and every number is suffixed, because these are\n * routinely sub-second and the `Timing` row above renders seconds — an\n * unlabelled column would silently mix the two.\n *\n * `first token` appears only when at least one call actually streamed one, so a\n * non-streaming run shows no misleading `0ms`.\n *\n * @param modelTiming `EvalSummary.modelTiming`, or undefined for no rows.\n * @param nameWidth   The scorer-name column width, so model names line up under it.\n */\nexport function formatModelTimingLines(\n  modelTiming: Record<string, ModelTimingStats> | undefined,\n  nameWidth: number,\n): string[] {\n  if (!modelTiming) return [];\n  return Object.entries(modelTiming).map(([model, t]) => {\n    const parts = [`wire ${formatMs(t.wireMs.mean)}/${formatMs(t.wireMs.p95)}`];\n    if (t.firstTokenMs) {\n      parts.push(`first token ${formatMs(t.firstTokenMs.mean)}/${formatMs(t.firstTokenMs.p95)}`);\n    }\n    parts.push(`queued ${formatMs(t.queuedMs.mean)}`);\n    parts.push(`retries ${formatMs(t.retryMs.mean)}`);\n    const calls = `${t.calls} call${t.calls === 1 ? '' : 's'}`;\n    return `    ${model.padEnd(Math.max(nameWidth - 2, 0))}  ${parts.join(' · ')}  (${calls}, mean/p95 per call)`;\n  });\n}\n\n/** Whole milliseconds with an explicit unit. Latencies here are per call, so a\n *  seconds-with-one-decimal rendering would collapse most of them to `0.0s`. */\nfunction formatMs(ms: number): string {\n  return `${Math.round(ms)}ms`;\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;AAKA,aAAW,OAAO,CAAC,eAAe,mBAAmB,GAAY;AAC/D,UAAM,IAAK,EAA8B,GAAG;AAC5C,QAAI,MAAM,WAAc,OAAO,MAAM,YAAY,CAAC,OAAO,UAAU,CAAC,KAAK,IAAI,IAAI;AAC/E,aAAO,sBAAsB,GAAG,KAAK,OAAO,MAAM,WAAW,IAAI,OAAO,CAAC;AAAA,IAC3E;AAAA,EACF;AAKA;AACE,UAAM,IAAK,EAA8B;AACzC,QAAI,MAAM,WAAc,OAAO,MAAM,YAAY,CAAC,OAAO,SAAS,CAAC,KAAK,IAAI,KAAK,IAAI,IAAI;AACvF,aAAO,6CAA6C,OAAO,MAAM,WAAW,IAAI,OAAO,CAAC;AAAA,IAC1F;AAAA,EACF;AACA,SAAO;AACT;;;ACvEO,IAAM,cAAc,oBAAI,IAAI;AAAA,EACjC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUM,IAAM,cAAc,oBAAI,IAAI;AAAA,EACjC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAoBM,SAAS,OAAO,MAAkC;AACvD,QAAM,MAAM,QAAQ,IAAI,IAAI;AAC5B,MAAI,OAAO,QAAQ,QAAQ,GAAI,QAAO;AACtC,QAAM,IAAI,SAAS,KAAK,EAAE;AAC1B,MAAI,CAAC,OAAO,SAAS,CAAC,KAAK,KAAK,EAAG,QAAO;AAC1C,SAAO;AACT;AAEO,SAAS,cAAc,MAAgC;AAC5D,MAAI;AACJ,MAAI;AACJ,MAAI,aAAuB,CAAC;AAC5B,MAAI,OAAO;AACX,MAAI,gBAAgB;AACpB,MAAI;AACJ,MAAI;AACJ,QAAM,QAAkB,CAAC;AAEzB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAClB,QAAI,YAAY,IAAI,GAAG,GAAG;AACxB,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,eAC7D,QAAQ,iBAAiB;AAMhC,cAAM,IAAI,SAAS,OAAO,EAAE;AAC5B,YAAI,CAAC,OAAO,SAAS,CAAC,KAAK,KAAK,GAAG;AACjC,kBAAQ,MAAM,8CAA8C,KAAK,aAAa;AAC9E,wBAAc;AAAA,QAChB,OAAO;AACL,wBAAc;AAAA,QAChB;AAAA,MACF,WAAW,QAAQ,aAAa;AAC9B,sBAAc;AAAA,UACZ,GAAG,IAAI;AAAA,YACL,MACG,MAAM,GAAG,EACT,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,OAAO;AAAA,UACnB;AAAA,QACF;AAAA,MACF,OAAO;AACL,qBAAa,MACV,MAAM,GAAG,EACT,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,OAAO;AAAA,MACnB;AAAA,IACF,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;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;;;ACrGO,SAAS,uBACd,aACA,WACU;AACV,MAAI,CAAC,YAAa,QAAO,CAAC;AAC1B,SAAO,OAAO,QAAQ,WAAW,EAAE,IAAI,CAAC,CAAC,OAAO,CAAC,MAAM;AACrD,UAAM,QAAQ,CAAC,QAAQ,SAAS,EAAE,OAAO,IAAI,CAAC,IAAI,SAAS,EAAE,OAAO,GAAG,CAAC,EAAE;AAC1E,QAAI,EAAE,cAAc;AAClB,YAAM,KAAK,eAAe,SAAS,EAAE,aAAa,IAAI,CAAC,IAAI,SAAS,EAAE,aAAa,GAAG,CAAC,EAAE;AAAA,IAC3F;AACA,UAAM,KAAK,UAAU,SAAS,EAAE,SAAS,IAAI,CAAC,EAAE;AAChD,UAAM,KAAK,WAAW,SAAS,EAAE,QAAQ,IAAI,CAAC,EAAE;AAChD,UAAM,QAAQ,GAAG,EAAE,KAAK,QAAQ,EAAE,UAAU,IAAI,KAAK,GAAG;AACxD,WAAO,OAAO,MAAM,OAAO,KAAK,IAAI,YAAY,GAAG,CAAC,CAAC,CAAC,KAAK,MAAM,KAAK,QAAK,CAAC,MAAM,KAAK;AAAA,EACzF,CAAC;AACH;AAIA,SAAS,SAAS,IAAoB;AACpC,SAAO,GAAG,KAAK,MAAM,EAAE,CAAC;AAC1B;;;AJpBA,SAAS,aAAa,QAAuB;AAC3C,UAAQ,MAAM,gCAAgC,MAAM,EAAE;AACtD,UAAQ,KAAK,CAAC;AAChB;AAQA,SAAS,mBACP,OACA,MACA,OACqD;AACrD,MAAI,MAAM,UAAU,QAAQ,MAAM,UAAU,MAAM;AAIhD,UAAM,UAAU,MAAM,YAAY,QAAQ,aAAa,OAAO,IAAI,EAAE,UAAU;AAC9E,WAAO,EAAE,QAAQ,MAAM,QAAQ,QAAQ,MAAM,QAAQ,QAAQ;AAAA,EAC/D;AACA,MAAI,MAAO,QAAO,aAAa,OAAO,IAAI;AAC1C,SAAO,EAAE,QAAQ,MAAM,UAAU,GAAG,QAAQ,MAAM,UAAU,GAAG,SAAS,MAAM,WAAW,EAAE;AAC7F;AAQA,SAAS,oBAAoB,YAAmC;AAC9D,MAAI,YAAY;AAChB,QAAM,WAAW,CAAC,QAAwB;AACxC,QAAI,WAAW;AACb,cAAQ,MAAM;AAAA,oBAAuB,GAAG,uBAAkB;AAC1D,cAAQ,KAAK,GAAG;AAAA,IAClB;AACA,gBAAY;AACZ,YAAQ,MAAM;AAAA,aAAgB,GAAG,iEAA4D;AAC7F,eAAW,MAAM;AAAA,EACnB;AACA,UAAQ,GAAG,UAAU,QAAQ;AAC7B,UAAQ,GAAG,WAAW,QAAQ;AAChC;AAEA,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;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAgCZ,kBAAkB,KAAK,MAAM,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,CAKjC;AACG,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AAEzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAKA,QAAM,aAAa,IAAI,gBAAgB;AACvC,sBAAoB,UAAU;AAE9B,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,GAAG,WAAW,MAAM;AACjD;AAAA,EACF;AAEA,QAAM,eAAe,MAAM,WAAW,MAAM;AAC9C;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;AAOA,SAAS,wBAAwB,MAAoC;AACnE,QAAM,MAAM,KAAK,QAAQ,yBAAyB;AAClD,MAAI,QAAQ,GAAI,QAAO;AACvB,MAAI,MAAM,KAAK,KAAK,QAAQ;AAC1B,YAAQ,MAAM,2DAA2D;AACzE,YAAQ,KAAK,CAAC;AAAA,EAChB;AACA,QAAM,MAAM,KAAK,MAAM,CAAC;AAIxB,QAAM,IAAI,cAAc,KAAK,IAAI,KAAK,CAAC,IAAI,OAAO,GAAG,IAAI;AACzD,MAAI,CAAC,OAAO,SAAS,CAAC,KAAK,IAAI,KAAK,IAAI,GAAG;AACzC,YAAQ,MAAM,mEAAmE,GAAG,GAAG;AACvF,YAAQ,KAAK,CAAC;AAAA,EAChB;AACA,SAAO;AACT;AAOA,SAAS,sBAAsB,GAAoD;AACjF,QAAM,QAAQ,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC,IAAI;AACxC,QAAM,OAAO,OAAO;AACpB,MAAI,CAAC,QAAQ,KAAK,mBAAmB,KAAM,QAAO;AAClD,SAAO,MAAM,QAAQ,KAAK,UAAU,IAAK,KAAK,aAA0B,CAAC;AAC3E;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,mBAAmB,KAAK,SAAS,sBAAsB;AAC7D,QAAM,aAAa,kBAAkB,IAAI;AACzC,QAAM,qBAAqB,wBAAwB,IAAI;AAGvD,QAAM,aAAa,oBAAI,IAAI,CAAC,eAAe,yBAAyB,CAAC;AACrE,QAAM,QAAQ,KAAK;AAAA,IACjB,CAAC,GAAG,MAAM,CAAC,EAAE,WAAW,IAAI,KAAK,EAAE,IAAI,KAAK,WAAW,IAAI,KAAK,IAAI,CAAC,CAAC;AAAA,EACxE;AAEA,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ;AAAA,MACN;AAAA,IACF;AACA,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;AAS7C,QAAM,gBACJ;AAAA,IACE,CAAC,YAAY,QAAQ;AAAA,IACrB,CAAC,aAAa,SAAS;AAAA,EACzB,EACA,OAAO,CAAC,CAAC,EAAE,CAAC,MAAM,sBAAsB,CAAC,KAAK,IAAI;AACpD,aAAW,CAAC,MAAM,CAAC,KAAK,eAAe;AACrC,UAAM,MAAM,sBAAsB,CAAC;AACnC,UAAM,SAAS,OAAO,IAAI,SAAS,UAAU,IAAI,KAAK,IAAI,CAAC,MAAM;AACjE,YAAQ;AAAA,MACN,uBAAuB,IAAI,yDAAoD,MAAM;AAAA,IACvF;AAAA,EACF;AACA,MAAI,oBAAoB,cAAc,SAAS,GAAG;AAChD;AAAA,MACE,GAAG,cAAc,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,EAAE,KAAK,IAAI,CAAC;AAAA,IAC7C;AAAA,EACF;AAEA,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;AASA,MAAI,mBAAmB;AACvB,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,WAAW,QAAQ,IAAI;AACjC,UAAM,UAAU,EAAE,kBAAkB;AACpC,UAAM,UAAU,EAAE,mBAAmB;AACrC,QAAI,UAAU,KAAK,UAAU,GAAG;AAC9B,yBAAmB;AACnB,YAAM,QAAQ,EAAE,kBAAkB,KAAK;AACvC,YAAM,QAAQ,EAAE,mBAAmB,KAAK;AACxC,cAAQ;AAAA,QACN,+BAA+B,IAAI,oCACpB,OAAO,IAAI,IAAI,eAAe,OAAO,IAAI,IAAI;AAAA,MAE9D;AAAA,IACF;AASA,UAAM,WAAW,EAAE,mBAAmB;AACtC,UAAM,WAAW,EAAE,oBAAoB;AACvC,UAAM,UAAU,EAAE,KAAK;AACvB,UAAM,YAAY,KAAK,IAAI,EAAE,kBAAkB,GAAG,EAAE,mBAAmB,CAAC;AACxE,UAAM,gBACH,WAAW,KAAK,WAAW,OAAO,aAAa,YAAY,UAAU;AACxE,QAAI,cAAc;AAChB,cAAQ;AAAA,QACN,4BAA4B,IAAI,qDACjB,QAAQ,mBAAmB,QAAQ,SAAS,OAAO;AAAA,MAEpE;AAAA,IACF;AAAA,EACF;AAEA,MAAI,oBAAoB,sBAAsB,MAAM;AAClD,YAAQ;AAAA,MACN;AAAA,IACF;AAAA,EACF;AAKA,MAAI,sBAAsB,MAAM;AAC9B,UAAM,SAAS,4BAA4B,YAAY,kBAAkB;AACzE,QAAI,OAAQ,cAAa,MAAM;AAAA,EACjC;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,QAAqB;AAC7D,QAAM,EAAE,YAAY,WAAW,YAAY,aAAa,aAAa,MAAM,IACzE,cAAc,IAAI;AAEpB,MAAI,MAAM,SAAS,GAAG;AACpB,YAAQ,MAAM,sEAAsE;AACpF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAKA,MAAI,aAAa,QAAQ;AACvB,YAAQ;AAAA,MACN;AAAA,IACF;AACA,YAAQ,KAAK,CAAC;AAAA,EAChB;AAIA,QAAM,kBAAkB,eAAe,OAAO,sBAAsB;AAEpE,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,UAAI,OAAO,QAAS;AACpB,eAAS;AAAA,QACP,MAAM,QAAQ,YAAY,WAAW,SAAS,SAAS;AAAA,UACrD;AAAA,UACA,GAAI,mBAAmB,OAAO,EAAE,aAAa,gBAAgB,IAAI,CAAC;AAAA,QACpE,CAAC;AAAA,MACH;AAAA,IACF;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;AACrC,UAAM,EAAE,QAAQ,QAAQ,QAAQ,IAAI,mBAAmB,GAAG,MAAM,OAAO,KAAK;AAM5E,UAAM,kBAA4B,CAAC;AACnC,QAAI,SAAS,EAAG,iBAAgB,KAAK,GAAG,MAAM,IAAI,SAAS,MAAM,YAAY,MAAM,SAAS;AAC5F,QAAI,UAAU,EAAG,iBAAgB,KAAK,GAAG,OAAO,MAAM;AACtD,UAAM,aAAa,gBAAgB,SAAS,IAAI,OAAO,gBAAgB,KAAK,IAAI,CAAC,MAAM;AAEvF,QAAI,WAAW,GAAG;AAChB,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,GAAG,UAAU;AAAA,MACvL;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,GAAG,UAAU;AAAA,MACpP;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;AAQA,QAAM,KAAK,GAAG,uBAAuB,OAAO,QAAQ,aAAa,UAAU,CAAC;AAE5E,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;AAG/D,UAAM,EAAE,QAAQ,OAAO,IAAI,mBAAmB,GAAG,IAAI;AACrD,UAAM,aACJ,SAAS,IAAI,OAAO,MAAM,IAAI,SAAS,MAAM,YAAY,MAAM,aAAa;AAC9E,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,GAAG,UAAU;AAAA,IAC5I;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;AAOA,SAAS,eAAe,QAAoB,OAAwB;AAClE,QAAM,IAAI,OAAO,QAAQ;AACzB,MAAI,CAAC,KAAK,EAAE,WAAW,EAAG,QAAO;AACjC,aAAW,KAAK,GAAG;AACjB,UAAM,YAAY,EAAE,SAAS,EAAE;AAC/B,UAAM,WACJ,EAAE,SAAS,kBACP,iCACA,IAAI,EAAE,QAAQ,KAAK,QAAQ,CAAC,CAAC;AACnC,YAAQ;AAAA,MACN,wBAAwB,KAAK,YAAY,EAAE,MAAM,MAAM,EAAE,IAAI,aACvD,EAAE,OAAO,KAAK,QAAQ,CAAC,CAAC,MAAM,EAAE,MAAM,IAAI,SAAS,eAAe,QAAQ;AAAA,IAClF;AAAA,EACF;AACA,SAAO;AACT;AAaA,SAAS,0BAA0B,QAAoB,OAAqB;AAC1E,aAAW,CAAC,MAAM,KAAK,KAAK,OAAO,QAAQ,OAAO,QAAQ,OAAO,GAAG;AAClE,UAAM,EAAE,QAAQ,QAAQ,QAAQ,IAAI,mBAAmB,OAAO,MAAM,OAAO,KAAK;AAChF,QAAI,UAAU,KAAK,SAAS,WAAW,GAAG;AACxC,cAAQ;AAAA,QACN,oBAAoB,KAAK,YAAY,IAAI,qCACnC,OAAO;AAAA,MAEf;AAAA,IACF;AAAA,EACF;AACF;AAeA,SAAS,mBAAmB,QAAoB,OAAwB;AACtE,QAAM,EAAE,OAAO,SAAS,IAAI,OAAO;AACnC,MAAI,UAAU,KAAK,WAAW,MAAO,QAAO;AAC5C,UAAQ;AAAA,IACN,sBAAsB,KAAK,eAAU,KAAK;AAAA,EAE5C;AACA,SAAO;AACT;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,eAAe,eAAe,MAAgB,QAAqB;AACjE,QAAM;AAAA,IACJ;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,IAAI,cAAc,IAAI;AAEtB,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;AAKA,MAAI,aAAa,UAAU,UAAU,SAAS,GAAG;AAC/C,YAAQ,MAAM,yDAAyD,UAAU,MAAM,IAAI;AAC3F,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;AAGhC,UAAI,OAAO,QAAS;AACpB,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;AAMA,mBAAW,cACT,eAAe,OAAO,sBAAsB,KAAK,WAAW,eAAe;AAK7E,YAAI,aAAa,QAAQ;AACvB,gBAAM,YAAY,WAAW,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI;AACtD,gBAAM,UAAU,YAAY,OAAO,CAAC,MAAM,CAAC,UAAU,SAAS,CAAC,CAAC;AAChE,cAAI,QAAQ,QAAQ;AAClB,oBAAQ;AAAA,cACN,wCAAwC,QAAQ,KAAK,IAAI,CAAC,gBAAgB,UAAU,KAAK,IAAI,CAAC;AAAA,YAChG;AACA,oBAAQ,KAAK,CAAC;AAAA,UAChB;AACA,qBAAW,UAAU,WAAW,QAAQ,OAAO,CAAC,MAAM,YAAY,SAAS,EAAE,IAAI,CAAC;AAClF,cAAI,WAAW,QAAQ,WAAW,GAAG;AACnC,oBAAQ,MAAM,sCAAsC;AACpD,oBAAQ,KAAK,CAAC;AAAA,UAChB;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;AAWA,cAAM,aAAa,EAAE,QAAQ,GAAI,gBAAgB,EAAE,eAAe,KAAK,IAAI,CAAC,EAAG;AAM/E,cAAM,aAAa,aAAa,SAAS,WAAW,QAAQ,IAAI,CAAC,MAAM,EAAE,IAAI,IAAI;AACjF,cAAM,gBAAgB,CAAC,WAAuB;AAC5C,cAAI,YAAY;AACd,mBAAO,SAAS,iBAAiB;AACjC,mBAAO,SAAS,aAAa;AAAA,UAC/B;AAAA,QACF;AAEA,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,gBAAI,OAAO,QAAS;AACpB,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,4BAAc,MAAM;AACpB,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;AAK1C,cAAI,aAAa;AACjB,qBAAW,KAAK,YAAY;AAE1B,kBAAM,UAAU,mBAAmB,GAAG,QAAQ;AAC9C,kBAAM,WAAW,eAAe,GAAG,QAAQ;AAC3C,sCAA0B,GAAG,QAAQ;AACrC,gBAAI,WAAW,SAAU,cAAa;AAAA,UACxC;AACA,cAAI,cAAc,CAAC,QAAS;AAAA,QAC9B,OAAO;AACL,gBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,SAAS,UAAU;AAC7E,wBAAc,MAAM;AACpB,kBAAQ,KAAK,MAAM;AAEnB,kBAAQ,IAAI,OAAO,YAAY,MAAM,IAAI,IAAI;AAI7C,gBAAM,UAAU,mBAAmB,QAAQ,QAAQ;AACnD,gBAAM,WAAW,eAAe,QAAQ,QAAQ;AAChD,oCAA0B,QAAQ,QAAQ;AAC1C,cAAI,WAAW,SAAU;AAAA,QAC3B;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

@@ -415,6 +415,27 @@ async function runEval(config, executeWorkflow, runtime, options) {
   const evalItems = new Array(items.length);
   let totalCost = 0;
   let budgetExceeded = false;
+  const runCallSamples = /* @__PURE__ */ new Map();
+  function absorbTiming(rollup, evalItem) {
+    if (!rollup) return;
+    const perItem = {};
+    for (const [model, bucket] of Object.entries(rollup)) {
+      const sums = {
+        calls: bucket.calls,
+        queuedMs: bucket.queuedMs,
+        retryMs: bucket.retryMs,
+        wireMs: bucket.wireMs,
+        ...bucket.firstTokenMs !== void 0 && bucket.firstTokenCalls !== void 0 ? { firstTokenMs: bucket.firstTokenMs, firstTokenCalls: bucket.firstTokenCalls } : {}
+      };
+      perItem[model] = sums;
+      const samples = bucket.samples;
+      if (!samples) continue;
+      const pooled = runCallSamples.get(model) ?? [];
+      for (const sample of samples) pooled.push(sample);
+      runCallSamples.set(model, pooled);
+    }
+    evalItem.timing = perItem;
+  }
   async function processItem(item, itemIndex) {
     if (options?.signal?.aborted) {
       evalItems[itemIndex] = {
@@ -449,13 +470,14 @@ async function runEval(config, executeWorkflow, runtime, options) {
       if (options?.captureTraces) {
         const tracked = await runtime.trackExecution(
           async () => executeWorkflow(item.input, runtime),
-          { captureTraces: true }
+          { captureTraces: true, captureTimingSamples: true }
         );
         evalItem.duration = Date.now() - itemStart;
         evalItem.output = tracked.result.output;
         evalItem.cost = extractUserCost(tracked.result) ?? tracked.cost;
         if (tracked.unpriced) evalItem.unpriced = true;
         evalItem.metadata = extractUserMetadata(tracked.result) ?? tracked.metadata;
+        absorbTiming(tracked.modelTiming, evalItem);
         if (tracked.traces && tracked.traces.length > 0) {
           evalItem.traces = tracked.traces;
         }
@@ -463,12 +485,24 @@ async function runEval(config, executeWorkflow, runtime, options) {
           totalCost += evalItem.cost;
         }
       } else {
-        const result = await executeWorkflow(item.input, runtime);
+        let trackedTiming;
+        let result;
+        if (typeof runtime.trackExecution === "function") {
+          const tracked = await runtime.trackExecution(
+            async () => executeWorkflow(item.input, runtime),
+            { captureTraces: false, captureTimingSamples: true }
+          );
+          result = tracked.result;
+          trackedTiming = tracked.modelTiming;
+        } else {
+          result = await executeWorkflow(item.input, runtime);
+        }
         evalItem.duration = Date.now() - itemStart;
         evalItem.output = result.output;
         evalItem.cost = extractUserCost(result);
         const meta = extractUserMetadata(result);
         if (meta) evalItem.metadata = meta;
+        absorbTiming(trackedTiming, evalItem);
         if (evalItem.cost != null) {
           totalCost += evalItem.cost;
         }
@@ -536,6 +570,19 @@ async function runEval(config, executeWorkflow, runtime, options) {
   }
   const durations = evalItems.filter((i) => !i.error && i.duration != null).map((i) => i.duration);
   const timing = durations.length > 0 ? computeStats(durations) : void 0;
+  const modelTiming = runCallSamples.size > 0 ? Object.fromEntries(
+    [...runCallSamples.entries()].map(([model, samples]) => {
+      const firstToken = samples.map((s) => s.firstTokenMs).filter((v) => v != null);
+      const stats = {
+        calls: samples.length,
+        wireMs: computeStats(samples.map((s) => s.wireMs)),
+        queuedMs: computeStats(samples.map((s) => s.queuedMs)),
+        retryMs: computeStats(samples.map((s) => s.retryMs)),
+        ...firstToken.length > 0 ? { firstTokenMs: computeStats(firstToken), firstTokenCalls: firstToken.length } : {}
+      };
+      return [model, stats];
+    })
+  ) : void 0;
   const totalModelCalls = /* @__PURE__ */ new Map();
   for (const item of evalItems) {
     const itemCounts = item.metadata?.modelCallCounts;
@@ -607,6 +654,7 @@ async function runEval(config, executeWorkflow, runtime, options) {
       failures,
       scorers: scorerStats,
       timing,
+      ...modelTiming ? { modelTiming } : {},
       ...degraded ? { degraded } : {}
     }
   };