@axlsdk/eval 0.13.8 → 0.15.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/cli.js.map CHANGED
@@ -1 +1 @@
1
- {"version":3,"sources":["../src/cli.ts","../src/cli-utils.ts"],"sourcesContent":["#!/usr/bin/env node\n\nimport { readdirSync, statSync } from 'node:fs';\nimport { readFile as readFileAsync, writeFile as writeFileAsync, mkdir } from 'node:fs/promises';\nimport * as path from 'node:path';\nimport type { AxlRuntime } from '@axlsdk/axl';\nimport { evalCompare } from './compare.js';\nimport { runEval } from './runner.js';\nimport { rescore } from './rescore.js';\nimport { aggregateRuns } from './multi-run.js';\nimport type { MultiRunSummary } from './multi-run.js';\nimport type { EvalConfig, EvalResult } from './types.js';\nimport {\n findConfig,\n resolveRuntime,\n importModule,\n registerConditions,\n CONFIG_CANDIDATES,\n} from './cli-utils.js';\n\nconst KNOWN_FLAGS = new Set([\n '--output',\n '--config',\n '--conditions',\n '--fail-on-regression',\n '--threshold',\n '--runs',\n]);\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 rescore <results> <eval-file> Re-run scorers on saved outputs\n axl-eval compare <a> <b> Compare two eval result files\n axl-eval compare <a> <b> --threshold <v> Set regression threshold (global or per-scorer)\n axl-eval compare <a> <b> --fail-on-regression Exit 1 if regressions\n\nConfig auto-detection (when --config is not specified):\n ${CONFIG_CANDIDATES.join(' -> ')}\n\nWhen a config is found, the exported AxlRuntime is passed to executeWorkflow\nand cost is tracked automatically via runtime.trackCost().\nWhen no config is found, a bare AxlRuntime is created (providers from env vars).\n`);\n process.exit(0);\n }\n\n if (args[0] === 'compare') {\n await runCompare(args.slice(1));\n return;\n }\n\n if (args[0] === 'rescore') {\n await runRescore(args.slice(1));\n return;\n }\n\n await runEvalCommand(args);\n}\n\nfunction parseThresholdArg(args: string[]): Record<string, number> | number | undefined {\n const idx = args.indexOf('--threshold');\n if (idx === -1 || idx + 1 >= args.length) return undefined;\n const value = args[idx + 1];\n\n // Global numeric threshold: --threshold 0.05\n if (/^[\\d.]+$/.test(value)) return parseFloat(value);\n\n // Per-scorer map: --threshold accuracy=0,tone=0.1\n const map: Record<string, number> = {};\n for (const pair of value.split(',')) {\n const [key, val] = pair.split('=');\n if (key && val != null) map[key.trim()] = parseFloat(val.trim());\n }\n return Object.keys(map).length > 0 ? map : undefined;\n}\n\nasync function runCompare(args: string[]) {\n const failOnRegression = args.includes('--fail-on-regression');\n const thresholds = parseThresholdArg(args);\n const files = args.filter(\n (a, i) => !a.startsWith('--') && !(i > 0 && args[i - 1] === '--threshold'),\n );\n\n if (files.length !== 2) {\n console.error('Usage: axl-eval compare <baseline.json> <candidate.json> [--threshold <value>]');\n process.exit(1);\n }\n\n const baselineRaw = JSON.parse(await readFileAsync(files[0], 'utf-8'));\n const candidateRaw = JSON.parse(await readFileAsync(files[1], 'utf-8'));\n const baseline: EvalResult | EvalResult[] = baselineRaw;\n const candidate: EvalResult | EvalResult[] = candidateRaw;\n const compareOptions = thresholds != null ? { thresholds } : undefined;\n const comparison = evalCompare(baseline, candidate, compareOptions);\n\n console.log(\n `\\nCompare: baseline (${comparison.baseline.id.slice(0, 8)}) -> candidate (${comparison.candidate.id.slice(0, 8)})\\n`,\n );\n\n const scorerNames = Object.keys(comparison.scorers);\n const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 6);\n const hasCI = scorerNames.some((n) => comparison.scorers[n].ci != null);\n\n const ciHeader = hasCI ? ' CI 95% Sig' : '';\n const ciRule = hasCI ? ' ---------------- ---' : '';\n console.log(` ${'Scorer'.padEnd(maxNameLen)} Baseline Candidate Delta Change${ciHeader}`);\n console.log(` ${''.padEnd(maxNameLen, '-')} -------- --------- -------- ------${ciRule}`);\n\n for (const name of scorerNames) {\n const s = comparison.scorers[name];\n const sign = s.delta > 0 ? '+' : '';\n let line = ` ${name.padEnd(maxNameLen)} ${s.baselineMean.toFixed(3).padStart(8)} ${s.candidateMean.toFixed(3).padStart(9)} ${(sign + s.delta.toFixed(3)).padStart(8)} ${(sign + s.deltaPercent.toFixed(1) + '%').padStart(6)}`;\n if (hasCI) {\n if (s.ci) {\n const lo = (s.ci.lower >= 0 ? '+' : '') + s.ci.lower.toFixed(4);\n const hi = (s.ci.upper >= 0 ? '+' : '') + s.ci.upper.toFixed(4);\n const sig = s.significant ? ' *' : ' ';\n line += ` [${lo}, ${hi}]${sig}`;\n } else {\n line += ' —'.padEnd(22);\n }\n }\n console.log(line);\n }\n\n if (comparison.timing) {\n const t = comparison.timing;\n const sign = t.delta > 0 ? '+' : '';\n console.log(\n `\\n Timing: baseline ${(t.baselineMean / 1000).toFixed(2)}s -> candidate ${(t.candidateMean / 1000).toFixed(2)}s (${sign}${t.deltaPercent.toFixed(1)}%)`,\n );\n }\n if (comparison.cost) {\n const c = comparison.cost;\n const sign = c.delta > 0 ? '+' : '';\n console.log(\n ` Cost: baseline $${c.baselineTotal.toFixed(2)} -> candidate $${c.candidateTotal.toFixed(2)} (${sign}${c.deltaPercent.toFixed(1)}%)`,\n );\n }\n\n const baselineRef = Array.isArray(baseline) ? baseline[0] : baseline;\n const stable = Math.max(\n 0,\n baselineRef.items.length - comparison.regressions.length - comparison.improvements.length,\n );\n console.log(\n `\\n Regressions: ${comparison.regressions.length} | Improvements: ${comparison.improvements.length} | Stable: ${stable}\\n`,\n );\n\n if (failOnRegression && comparison.regressions.length > 0) {\n // When CI is available, only fail on significant regressions\n const hasSignificance = scorerNames.some((n) => comparison.scorers[n].significant != null);\n if (hasSignificance) {\n const hasSignificantRegression = scorerNames.some(\n (n) => comparison.scorers[n].significant === true && comparison.scorers[n].delta < 0,\n );\n if (hasSignificantRegression) process.exit(1);\n } else {\n process.exit(1);\n }\n }\n}\n\nasync function runRescore(args: string[]) {\n const { outputPath, configArg, conditions, paths } = parseEvalArgs(args);\n\n if (paths.length < 2) {\n console.error('Usage: axl-eval rescore <results.json> <eval-file> [--output <file>]');\n process.exit(1);\n }\n\n const [resultsPath, evalFilePath] = paths;\n const raw = JSON.parse(await readFileAsync(resultsPath, 'utf-8'));\n const results: EvalResult[] = Array.isArray(raw) ? raw : [raw];\n\n const runtime = await getRuntime(configArg, conditions);\n try {\n const mod = await importModule(path.resolve(evalFilePath), import.meta.url);\n const evalConfig: EvalConfig = mod.default?.default ?? mod.default ?? mod.config ?? mod;\n\n if (!evalConfig.scorers || evalConfig.scorers.length === 0) {\n console.error(`Error: ${evalFilePath} does not export scorers`);\n process.exit(1);\n }\n\n const rescored: EvalResult[] = [];\n for (const resultData of results) {\n rescored.push(await rescore(resultData, evalConfig.scorers, runtime));\n }\n\n for (const r of rescored) {\n console.log('\\n' + formatTable(r) + '\\n');\n }\n\n if (outputPath) {\n const output = Array.isArray(raw) ? rescored : rescored[0];\n const outputDir = path.dirname(path.resolve(outputPath));\n await mkdir(outputDir, { recursive: true });\n await writeFileAsync(outputPath, JSON.stringify(output, null, 2), 'utf-8');\n console.log(`Rescored results saved to ${outputPath}`);\n }\n } finally {\n await runtime.shutdown().catch(() => {});\n }\n}\n\nfunction collectEvalFiles(p: string): string[] {\n const resolved = path.resolve(p);\n try {\n const stat = statSync(resolved);\n if (stat.isDirectory()) {\n const entries = readdirSync(resolved);\n return entries\n .filter((e) => /\\.eval\\.[mc]?[jt]sx?$/.test(e))\n .map((e) => path.join(resolved, e));\n }\n return [resolved];\n } catch {\n return [resolved];\n }\n}\n\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(`Eval: ${result.workflow} x ${result.dataset} (${result.summary.count} items)`);\n lines.push(\n ` ${'Scorer'.padEnd(maxNameLen)} ${'Mean'.padStart(colWidth)} ${'Min'.padStart(colWidth)} ${'Max'.padStart(colWidth)} ${'p50'.padStart(colWidth)} ${'p95'.padStart(colWidth)}`,\n );\n\n const ruleLen = maxNameLen + 2 + (colWidth + 2) * 5;\n lines.push(' ' + '\\u2500'.repeat(ruleLen));\n\n for (const name of scorerNames) {\n const s = result.summary.scorers[name];\n // Detect scorers with no valid scores (all items errored → computeStats([]) → all zeros)\n const validScoreCount = result.items.filter((i) => !i.error && i.scores[name] != null).length;\n if (validScoreCount === 0) {\n lines.push(\n ` ${name.padEnd(maxNameLen)} ${'--'.padStart(colWidth)} ${'--'.padStart(colWidth)} ${'--'.padStart(colWidth)} ${'--'.padStart(colWidth)} ${'--'.padStart(colWidth)}`,\n );\n } else {\n lines.push(\n ` ${name.padEnd(maxNameLen)} ${s.mean.toFixed(2).padStart(colWidth)} ${s.min.toFixed(2).padStart(colWidth)} ${s.max.toFixed(2).padStart(colWidth)} ${s.p50.toFixed(2).padStart(colWidth)} ${s.p95.toFixed(2).padStart(colWidth)}`,\n );\n }\n }\n\n if (result.summary.timing) {\n const t = result.summary.timing;\n lines.push(\n ` ${'Timing'.padEnd(maxNameLen)} ${(t.mean / 1000).toFixed(2).padStart(colWidth)}s ${(t.p50 / 1000).toFixed(2).padStart(colWidth)}s ${(t.p95 / 1000).toFixed(2).padStart(colWidth)}s`,\n );\n }\n\n const durationSec = (result.duration / 1000).toFixed(1);\n const costStr = result.totalCost > 0 ? `$${result.totalCost.toFixed(2)}` : '$0.00';\n lines.push('');\n lines.push(\n ` Failures: ${result.summary.failures}/${result.summary.count} | Cost: ${costStr} | Duration: ${durationSec}s`,\n );\n\n const itemsWithErrors = result.items.filter((i) => i.scorerErrors?.length);\n if (itemsWithErrors.length > 0) {\n const uniqueErrors = [...new Set(itemsWithErrors.flatMap((i) => i.scorerErrors!))];\n lines.push('');\n lines.push(\n ` Scorer errors (${itemsWithErrors.length}/${result.summary.count} items affected):`,\n );\n for (const err of uniqueErrors.slice(0, 5)) {\n lines.push(` - ${err}`);\n }\n if (uniqueErrors.length > 5) {\n lines.push(` ... and ${uniqueErrors.length - 5} more`);\n }\n }\n\n return lines.join('\\n');\n}\n\nfunction formatMultiRunTable(summary: MultiRunSummary): string {\n const lines: string[] = [];\n const scorerNames = Object.keys(summary.scorers);\n const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 'Scorer'.length);\n const colWidth = 16;\n\n lines.push(`Eval: ${summary.workflow} x ${summary.dataset} \\u2014 ${summary.runCount} runs`);\n lines.push(\n ` ${'Scorer'.padEnd(maxNameLen)} ${'Mean \\u00b1 Std'.padStart(colWidth)} ${'Min'.padStart(8)} ${'Max'.padStart(8)}`,\n );\n const ruleLen = maxNameLen + 2 + colWidth + 2 + 8 + 2 + 8;\n lines.push(' ' + '\\u2500'.repeat(ruleLen));\n\n for (const name of scorerNames) {\n const s = summary.scorers[name];\n const meanStd = `${s.mean.toFixed(3)} \\u00b1 ${s.std.toFixed(3)}`;\n lines.push(\n ` ${name.padEnd(maxNameLen)} ${meanStd.padStart(colWidth)} ${s.min.toFixed(3).padStart(8)} ${s.max.toFixed(3).padStart(8)}`,\n );\n }\n\n if (summary.timing) {\n lines.push(\n ` ${'Timing'.padEnd(maxNameLen)} ${((summary.timing.mean / 1000).toFixed(2) + ' \\u00b1 ' + (summary.timing.std / 1000).toFixed(2) + 's').padStart(colWidth)}`,\n );\n }\n\n const costStr = summary.totalCost > 0 ? `$${summary.totalCost.toFixed(2)}` : '$0.00';\n const durationStr = (summary.totalDuration / 1000).toFixed(1);\n lines.push('');\n lines.push(` Total Cost: ${costStr} | Total Duration: ${durationStr}s`);\n\n return lines.join('\\n');\n}\n\n// ── Runtime resolution ─────────────────────────────────────────────\n\nasync function resolveRuntimeFromConfig(configPath: string): Promise<AxlRuntime> {\n try {\n const mod = await importModule(configPath, import.meta.url);\n const runtime = resolveRuntime(mod) as AxlRuntime;\n\n if (!runtime || typeof runtime.execute !== 'function') {\n console.error(`Config must export a default AxlRuntime instance.`);\n if (runtime) {\n const keys = Object.keys(runtime as object)\n .slice(0, 5)\n .join(', ');\n console.error(` Got: ${typeof runtime}${keys ? ` with keys: { ${keys} }` : ''}`);\n }\n console.error(\n `Example:\\n import { AxlRuntime } from '@axlsdk/axl';\\n export default new AxlRuntime({ ... });`,\n );\n process.exit(1);\n }\n\n return runtime;\n } catch (err) {\n const msg = err instanceof Error ? err.message : String(err);\n if (\n /Cannot use import statement|Unexpected reserved word|top-level await|exports is not defined/.test(\n msg,\n )\n ) {\n const ext = path.extname(configPath);\n console.error(`[axl-eval] Config failed to load due to a CJS/ESM compatibility issue.`);\n if (ext === '.ts' || ext === '.tsx') {\n console.error(\n ` Tip: try renaming to .mts to force ESM format, or ensure tsx is installed and up to date.`,\n );\n } else {\n console.error(` Tip: add \"type\": \"module\" to your package.json.`);\n }\n console.error();\n }\n console.error(`Failed to load config:`, err);\n process.exit(1);\n }\n}\n\nasync function getRuntime(configArg?: string, conditions?: string[]): Promise<AxlRuntime> {\n // Register import conditions before any config loading\n if (conditions && conditions.length > 0) {\n await registerConditions(conditions);\n }\n\n // 1. Explicit --config\n if (configArg) {\n const configPath = path.resolve(process.cwd(), configArg);\n const stat = statSync(configPath, { throwIfNoEntry: false });\n if (!stat?.isFile()) {\n console.error(`Config file not found: ${configPath}`);\n process.exit(1);\n }\n console.error(`[axl-eval] Loading config from ${configPath}`);\n return resolveRuntimeFromConfig(configPath);\n }\n\n // 2. Auto-detect\n const found = findConfig(process.cwd());\n if (found) {\n console.error(`[axl-eval] Auto-detected config: ${found}`);\n return resolveRuntimeFromConfig(found);\n }\n\n // 3. Bare runtime (providers from env vars)\n const { AxlRuntime } = await import('@axlsdk/axl');\n return new AxlRuntime();\n}\n\n// ── Arg parsing ────────────────────────────────────────────────────\n\nfunction parseEvalArgs(args: string[]): {\n outputPath?: string;\n configArg?: string;\n conditions: string[];\n runs: number;\n paths: string[];\n} {\n let outputPath: string | undefined;\n let configArg: string | undefined;\n let conditions: string[] = [];\n let runs = 1;\n const paths: string[] = [];\n\n for (let i = 0; i < args.length; i++) {\n const arg = args[i];\n if (arg === '--output' || arg === '--config' || arg === '--conditions' || arg === '--runs') {\n if (i + 1 >= args.length) {\n console.error(`Error: ${arg} requires a value`);\n process.exit(1);\n }\n const value = args[++i];\n if (arg === '--output') outputPath = value;\n else if (arg === '--config') configArg = value;\n else if (arg === '--runs') runs = Math.max(1, parseInt(value, 10) || 1);\n else\n conditions = value\n .split(',')\n .map((c) => c.trim())\n .filter(Boolean);\n } else if (arg.startsWith('--')) {\n if (!KNOWN_FLAGS.has(arg)) {\n console.error(`Unknown flag: ${arg}`);\n process.exit(1);\n }\n } else {\n paths.push(arg);\n }\n }\n\n return { outputPath, configArg, conditions, runs, paths };\n}\n\n// ── Main eval command ──────────────────────────────────────────────\n\nasync function runEvalCommand(args: string[]) {\n const { outputPath, configArg, conditions, runs, paths } = parseEvalArgs(args);\n\n if (paths.length === 0) {\n console.error('Error: No eval file path provided');\n process.exit(1);\n }\n\n const evalFiles: string[] = [];\n for (const p of paths) {\n evalFiles.push(...collectEvalFiles(p));\n }\n\n if (evalFiles.length === 0) {\n console.error('Error: No eval files found');\n process.exit(1);\n }\n\n const runtime = await getRuntime(configArg, conditions);\n const results: EvalResult[] = [];\n\n try {\n for (const filePath of evalFiles) {\n try {\n const mod = await importModule(path.resolve(filePath), import.meta.url);\n const evalConfig: EvalConfig = mod.default?.default ?? mod.default ?? mod.config ?? mod;\n\n if (!evalConfig.workflow || !evalConfig.dataset || !evalConfig.scorers) {\n console.error(\n `Error: ${filePath} does not export a valid eval config (missing workflow, dataset, or scorers)`,\n );\n continue;\n }\n\n // Resolve executeWorkflow: custom export > registered workflow > passthrough\n let executeWorkflow: (\n input: unknown,\n rt?: unknown,\n ) => Promise<{ output: unknown; cost?: number; metadata?: Record<string, unknown> }>;\n\n if (mod.executeWorkflow) {\n // Wrap custom executeWorkflow with trackExecution for cost + metadata attribution\n executeWorkflow = async (input, rt) => {\n const {\n result,\n cost: trackedCost,\n metadata,\n } = await runtime.trackExecution(async () => {\n return mod.executeWorkflow(input, rt);\n });\n return {\n output: result.output,\n cost: result.cost ?? trackedCost,\n metadata: result.metadata ?? metadata,\n };\n };\n } else if (runtime.getWorkflow(evalConfig.workflow)) {\n // No executeWorkflow exported but workflow is registered — use runtime.execute()\n executeWorkflow = async (input) => {\n const { result, cost, metadata } = await runtime.trackExecution(async () => {\n return runtime.execute(evalConfig.workflow, input);\n });\n return { output: result, cost, metadata };\n };\n } else {\n console.warn(\n `[axl-eval] Warning: ${filePath} does not export executeWorkflow — using input passthrough`,\n );\n executeWorkflow = async (input) => ({ output: input });\n }\n\n if (runs > 1) {\n // Multi-run mode\n const { randomUUID } = await import('node:crypto');\n const runGroupId = randomUUID();\n const runResults: EvalResult[] = [];\n\n for (let r = 0; r < runs; r++) {\n console.error(`[axl-eval] Run ${r + 1}/${runs}...`);\n const result = await runEval(evalConfig, executeWorkflow, runtime);\n result.metadata.runGroupId = runGroupId;\n result.metadata.runIndex = r;\n runResults.push(result);\n results.push(result);\n }\n\n const summary = aggregateRuns(runResults);\n console.log('\\n' + formatMultiRunTable(summary) + '\\n');\n } else {\n const result = await runEval(evalConfig, executeWorkflow, runtime);\n results.push(result);\n\n console.log('\\n' + formatTable(result) + '\\n');\n }\n } catch (err) {\n console.error(\n `Error running eval ${filePath}: ${err instanceof Error ? err.message : String(err)}`,\n );\n }\n }\n\n if (outputPath && results.length > 0) {\n const output = results.length === 1 ? results[0] : results;\n const outputDir = path.dirname(path.resolve(outputPath));\n await mkdir(outputDir, { recursive: true });\n await writeFileAsync(outputPath, JSON.stringify(output, null, 2), 'utf-8');\n console.log(`Results saved to ${outputPath}`);\n }\n } finally {\n await runtime.shutdown().catch(() => {});\n }\n\n // Exit with non-zero code if no evals succeeded\n if (results.length === 0) {\n process.exit(1);\n }\n}\n\nmain().catch((err) => {\n console.error(err);\n process.exit(1);\n});\n","/**\n * CLI utilities for config detection, runtime resolution, and loader registration.\n *\n * These are duplicated from @axlsdk/studio (cli-utils.ts, resolve-runtime.ts,\n * eval-loader.ts, cli.ts) because @axlsdk/eval cannot depend on @axlsdk/studio.\n * Keep in sync with the studio versions if either changes.\n */\n\nimport { resolve } from 'node:path';\nimport { existsSync } from 'node:fs';\nimport { pathToFileURL } from 'node:url';\n\n// ── Config auto-detection ──────────────────────────────────────────\n\nexport const CONFIG_CANDIDATES = [\n 'axl.config.mts',\n 'axl.config.ts',\n 'axl.config.mjs',\n 'axl.config.js',\n];\n\nexport function findConfig(cwd: string): string | undefined {\n for (const name of CONFIG_CANDIDATES) {\n const p = resolve(cwd, name);\n if (existsSync(p)) return p;\n }\n return undefined;\n}\n\n// ── Extension helpers ──────────────────────────────────────────────\n\n/** Returns true if the file is TypeScript and needs tsx to load. */\nexport function needsTsxLoader(filePath: string): boolean {\n return /\\.[mc]?tsx?$/.test(filePath);\n}\n\n// ── Runtime resolution ─────────────────────────────────────────────\n\n/**\n * Resolve the AxlRuntime from a dynamically imported config module.\n *\n * Module shapes handled:\n * - ESM `export default runtime` → mod.default is the runtime\n * - CJS compiled from `export default runtime` → mod.default.default is the runtime\n * - CJS `module.exports = runtime` → mod.default is the runtime\n * - Named `export { runtime }` → mod.runtime is the runtime\n */\nexport function resolveRuntime(mod: Record<string, any>): unknown {\n const def = mod.default as Record<string, any> | undefined;\n return def?.default ?? def ?? mod.runtime;\n}\n\n// ── Module loading ────────────────────────────────────────────────\n\n// Lazily resolved tsImport function from tsx. `undefined` = not yet checked,\n// `null` = tsx not available.\nlet tsImportFn:\n | ((specifier: string, parentURL: string) => Promise<Record<string, any>>)\n | null\n | undefined;\n\n/**\n * Import a module, using tsx's `tsImport()` for TypeScript files.\n *\n * `tsImport()` handles ESM/CJS format correctly without process-wide side effects —\n * no need for `register()` hooks or ESM-forcing workarounds. Falls back to regular\n * `import()` for non-TypeScript files or when tsx is not installed.\n */\nexport async function importModule(\n filePath: string,\n parentURL: string,\n): Promise<Record<string, any>> {\n if (needsTsxLoader(filePath)) {\n if (tsImportFn === undefined) {\n try {\n // @ts-expect-error — tsx is an optional runtime dependency\n const mod = await import('tsx/esm/api');\n tsImportFn = mod.tsImport ?? null;\n } catch {\n tsImportFn = null;\n console.warn(\n '[axl-eval] Warning: tsx is not installed. TypeScript files require tsx.\\n' +\n ' Install it with: npm install -D tsx',\n );\n }\n }\n if (tsImportFn) {\n return (await tsImportFn(pathToFileURL(filePath).href, parentURL)) as Record<string, any>;\n }\n }\n return await import(pathToFileURL(filePath).href);\n}\n\n// ── Conditions ────────────────────────────────────────────────────\n\nexport async function registerConditions(conditions: string[]): Promise<void> {\n try {\n const nodeModule = await import('node:module');\n const hookCode = [\n `const extra = ${JSON.stringify(conditions)};`,\n `export async function resolve(specifier, context, nextResolve) {`,\n ` return nextResolve(specifier, {`,\n ` ...context,`,\n ` conditions: [...new Set([...context.conditions, ...extra])],`,\n ` });`,\n `}`,\n ].join('\\n');\n nodeModule.register(`data:text/javascript,${encodeURIComponent(hookCode)}`);\n } catch {\n console.warn('[axl-eval] Warning: --conditions requires Node.js 20.6+');\n }\n}\n"],"mappings":";;;;;;;;;AAEA,SAAS,aAAa,gBAAgB;AACtC,SAAS,YAAY,eAAe,aAAa,gBAAgB,aAAa;AAC9E,YAAY,UAAU;;;ACItB,SAAS,eAAe;AACxB,SAAS,kBAAkB;AAC3B,SAAS,qBAAqB;AAIvB,IAAM,oBAAoB;AAAA,EAC/B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEO,SAAS,WAAW,KAAiC;AAC1D,aAAW,QAAQ,mBAAmB;AACpC,UAAM,IAAI,QAAQ,KAAK,IAAI;AAC3B,QAAI,WAAW,CAAC,EAAG,QAAO;AAAA,EAC5B;AACA,SAAO;AACT;AAKO,SAAS,eAAe,UAA2B;AACxD,SAAO,eAAe,KAAK,QAAQ;AACrC;AAaO,SAAS,eAAe,KAAmC;AAChE,QAAM,MAAM,IAAI;AAChB,SAAO,KAAK,WAAW,OAAO,IAAI;AACpC;AAMA,IAAI;AAYJ,eAAsB,aACpB,UACA,WAC8B;AAC9B,MAAI,eAAe,QAAQ,GAAG;AAC5B,QAAI,eAAe,QAAW;AAC5B,UAAI;AAEF,cAAM,MAAM,MAAM,OAAO,aAAa;AACtC,qBAAa,IAAI,YAAY;AAAA,MAC/B,QAAQ;AACN,qBAAa;AACb,gBAAQ;AAAA,UACN;AAAA,QAEF;AAAA,MACF;AAAA,IACF;AACA,QAAI,YAAY;AACd,aAAQ,MAAM,WAAW,cAAc,QAAQ,EAAE,MAAM,SAAS;AAAA,IAClE;AAAA,EACF;AACA,SAAO,MAAM,OAAO,cAAc,QAAQ,EAAE;AAC9C;AAIA,eAAsB,mBAAmB,YAAqC;AAC5E,MAAI;AACF,UAAM,aAAa,MAAM,OAAO,QAAa;AAC7C,UAAM,WAAW;AAAA,MACf,iBAAiB,KAAK,UAAU,UAAU,CAAC;AAAA,MAC3C;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AACX,eAAW,SAAS,wBAAwB,mBAAmB,QAAQ,CAAC,EAAE;AAAA,EAC5E,QAAQ;AACN,YAAQ,KAAK,yDAAyD;AAAA,EACxE;AACF;;;AD3FA,IAAM,cAAc,oBAAI,IAAI;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,eAAe,OAAO;AACpB,QAAM,OAAO,QAAQ,KAAK,MAAM,CAAC;AAEjC,MAAI,KAAK,WAAW,KAAK,KAAK,CAAC,MAAM,YAAY,KAAK,CAAC,MAAM,MAAM;AACjE,YAAQ,IAAI;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAaZ,kBAAkB,KAAK,MAAM,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,CAKjC;AACG,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAEA,QAAM,eAAe,IAAI;AAC3B;AAEA,SAAS,kBAAkB,MAA6D;AACtF,QAAM,MAAM,KAAK,QAAQ,aAAa;AACtC,MAAI,QAAQ,MAAM,MAAM,KAAK,KAAK,OAAQ,QAAO;AACjD,QAAM,QAAQ,KAAK,MAAM,CAAC;AAG1B,MAAI,WAAW,KAAK,KAAK,EAAG,QAAO,WAAW,KAAK;AAGnD,QAAM,MAA8B,CAAC;AACrC,aAAW,QAAQ,MAAM,MAAM,GAAG,GAAG;AACnC,UAAM,CAAC,KAAK,GAAG,IAAI,KAAK,MAAM,GAAG;AACjC,QAAI,OAAO,OAAO,KAAM,KAAI,IAAI,KAAK,CAAC,IAAI,WAAW,IAAI,KAAK,CAAC;AAAA,EACjE;AACA,SAAO,OAAO,KAAK,GAAG,EAAE,SAAS,IAAI,MAAM;AAC7C;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,mBAAmB,KAAK,SAAS,sBAAsB;AAC7D,QAAM,aAAa,kBAAkB,IAAI;AACzC,QAAM,QAAQ,KAAK;AAAA,IACjB,CAAC,GAAG,MAAM,CAAC,EAAE,WAAW,IAAI,KAAK,EAAE,IAAI,KAAK,KAAK,IAAI,CAAC,MAAM;AAAA,EAC9D;AAEA,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ,MAAM,gFAAgF;AAC9F,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,cAAc,KAAK,MAAM,MAAM,cAAc,MAAM,CAAC,GAAG,OAAO,CAAC;AACrE,QAAM,eAAe,KAAK,MAAM,MAAM,cAAc,MAAM,CAAC,GAAG,OAAO,CAAC;AACtE,QAAM,WAAsC;AAC5C,QAAM,YAAuC;AAC7C,QAAM,iBAAiB,cAAc,OAAO,EAAE,WAAW,IAAI;AAC7D,QAAM,aAAa,YAAY,UAAU,WAAW,cAAc;AAElE,UAAQ;AAAA,IACN;AAAA,qBAAwB,WAAW,SAAS,GAAG,MAAM,GAAG,CAAC,CAAC,mBAAmB,WAAW,UAAU,GAAG,MAAM,GAAG,CAAC,CAAC;AAAA;AAAA,EAClH;AAEA,QAAM,cAAc,OAAO,KAAK,WAAW,OAAO;AAClD,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,CAAC;AAClE,QAAM,QAAQ,YAAY,KAAK,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,MAAM,IAAI;AAEtE,QAAM,WAAW,QAAQ,4BAA4B;AACrD,QAAM,SAAS,QAAQ,4BAA4B;AACnD,UAAQ,IAAI,KAAK,SAAS,OAAO,UAAU,CAAC,0CAA0C,QAAQ,EAAE;AAChG,UAAQ,IAAI,KAAK,GAAG,OAAO,YAAY,GAAG,CAAC,0CAA0C,MAAM,EAAE;AAE7F,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,WAAW,QAAQ,IAAI;AACjC,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,QAAI,OAAO,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,EAAE,aAAa,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,EAAE,cAAc,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,MAAM,OAAO,EAAE,MAAM,QAAQ,CAAC,GAAG,SAAS,CAAC,CAAC,MAAM,OAAO,EAAE,aAAa,QAAQ,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC;AACjO,QAAI,OAAO;AACT,UAAI,EAAE,IAAI;AACR,cAAM,MAAM,EAAE,GAAG,SAAS,IAAI,MAAM,MAAM,EAAE,GAAG,MAAM,QAAQ,CAAC;AAC9D,cAAM,MAAM,EAAE,GAAG,SAAS,IAAI,MAAM,MAAM,EAAE,GAAG,MAAM,QAAQ,CAAC;AAC9D,cAAM,MAAM,EAAE,cAAc,QAAQ;AACpC,gBAAQ,MAAM,EAAE,KAAK,EAAE,IAAI,GAAG;AAAA,MAChC,OAAO;AACL,gBAAQ,WAAM,OAAO,EAAE;AAAA,MACzB;AAAA,IACF;AACA,YAAQ,IAAI,IAAI;AAAA,EAClB;AAEA,MAAI,WAAW,QAAQ;AACrB,UAAM,IAAI,WAAW;AACrB,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,YAAQ;AAAA,MACN;AAAA,sBAAyB,EAAE,eAAe,KAAM,QAAQ,CAAC,CAAC,mBAAmB,EAAE,gBAAgB,KAAM,QAAQ,CAAC,CAAC,MAAM,IAAI,GAAG,EAAE,aAAa,QAAQ,CAAC,CAAC;AAAA,IACvJ;AAAA,EACF;AACA,MAAI,WAAW,MAAM;AACnB,UAAM,IAAI,WAAW;AACrB,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,YAAQ;AAAA,MACN,qBAAqB,EAAE,cAAc,QAAQ,CAAC,CAAC,kBAAkB,EAAE,eAAe,QAAQ,CAAC,CAAC,KAAK,IAAI,GAAG,EAAE,aAAa,QAAQ,CAAC,CAAC;AAAA,IACnI;AAAA,EACF;AAEA,QAAM,cAAc,MAAM,QAAQ,QAAQ,IAAI,SAAS,CAAC,IAAI;AAC5D,QAAM,SAAS,KAAK;AAAA,IAClB;AAAA,IACA,YAAY,MAAM,SAAS,WAAW,YAAY,SAAS,WAAW,aAAa;AAAA,EACrF;AACA,UAAQ;AAAA,IACN;AAAA,iBAAoB,WAAW,YAAY,MAAM,oBAAoB,WAAW,aAAa,MAAM,cAAc,MAAM;AAAA;AAAA,EACzH;AAEA,MAAI,oBAAoB,WAAW,YAAY,SAAS,GAAG;AAEzD,UAAM,kBAAkB,YAAY,KAAK,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,eAAe,IAAI;AACzF,QAAI,iBAAiB;AACnB,YAAM,2BAA2B,YAAY;AAAA,QAC3C,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,gBAAgB,QAAQ,WAAW,QAAQ,CAAC,EAAE,QAAQ;AAAA,MACrF;AACA,UAAI,yBAA0B,SAAQ,KAAK,CAAC;AAAA,IAC9C,OAAO;AACL,cAAQ,KAAK,CAAC;AAAA,IAChB;AAAA,EACF;AACF;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,EAAE,YAAY,WAAW,YAAY,MAAM,IAAI,cAAc,IAAI;AAEvE,MAAI,MAAM,SAAS,GAAG;AACpB,YAAQ,MAAM,sEAAsE;AACpF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,CAAC,aAAa,YAAY,IAAI;AACpC,QAAM,MAAM,KAAK,MAAM,MAAM,cAAc,aAAa,OAAO,CAAC;AAChE,QAAM,UAAwB,MAAM,QAAQ,GAAG,IAAI,MAAM,CAAC,GAAG;AAE7D,QAAM,UAAU,MAAM,WAAW,WAAW,UAAU;AACtD,MAAI;AACF,UAAM,MAAM,MAAM,aAAkB,aAAQ,YAAY,GAAG,YAAY,GAAG;AAC1E,UAAM,aAAyB,IAAI,SAAS,WAAW,IAAI,WAAW,IAAI,UAAU;AAEpF,QAAI,CAAC,WAAW,WAAW,WAAW,QAAQ,WAAW,GAAG;AAC1D,cAAQ,MAAM,UAAU,YAAY,0BAA0B;AAC9D,cAAQ,KAAK,CAAC;AAAA,IAChB;AAEA,UAAM,WAAyB,CAAC;AAChC,eAAW,cAAc,SAAS;AAChC,eAAS,KAAK,MAAM,QAAQ,YAAY,WAAW,SAAS,OAAO,CAAC;AAAA,IACtE;AAEA,eAAW,KAAK,UAAU;AACxB,cAAQ,IAAI,OAAO,YAAY,CAAC,IAAI,IAAI;AAAA,IAC1C;AAEA,QAAI,YAAY;AACd,YAAM,SAAS,MAAM,QAAQ,GAAG,IAAI,WAAW,SAAS,CAAC;AACzD,YAAM,YAAiB,aAAa,aAAQ,UAAU,CAAC;AACvD,YAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,YAAM,eAAe,YAAY,KAAK,UAAU,QAAQ,MAAM,CAAC,GAAG,OAAO;AACzE,cAAQ,IAAI,6BAA6B,UAAU,EAAE;AAAA,IACvD;AAAA,EACF,UAAE;AACA,UAAM,QAAQ,SAAS,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EACzC;AACF;AAEA,SAAS,iBAAiB,GAAqB;AAC7C,QAAM,WAAgB,aAAQ,CAAC;AAC/B,MAAI;AACF,UAAM,OAAO,SAAS,QAAQ;AAC9B,QAAI,KAAK,YAAY,GAAG;AACtB,YAAM,UAAU,YAAY,QAAQ;AACpC,aAAO,QACJ,OAAO,CAAC,MAAM,wBAAwB,KAAK,CAAC,CAAC,EAC7C,IAAI,CAAC,MAAW,UAAK,UAAU,CAAC,CAAC;AAAA,IACtC;AACA,WAAO,CAAC,QAAQ;AAAA,EAClB,QAAQ;AACN,WAAO,CAAC,QAAQ;AAAA,EAClB;AACF;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,KAAK,SAAS,OAAO,QAAQ,MAAM,OAAO,OAAO,KAAK,OAAO,QAAQ,KAAK,SAAS;AACzF,QAAM;AAAA,IACJ,KAAK,SAAS,OAAO,UAAU,CAAC,KAAK,OAAO,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC;AAAA,EACpL;AAEA,QAAM,UAAU,aAAa,KAAK,WAAW,KAAK;AAClD,QAAM,KAAK,OAAO,SAAS,OAAO,OAAO,CAAC;AAE1C,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,OAAO,QAAQ,QAAQ,IAAI;AAErC,UAAM,kBAAkB,OAAO,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,SAAS,EAAE,OAAO,IAAI,KAAK,IAAI,EAAE;AACvF,QAAI,oBAAoB,GAAG;AACzB,YAAM;AAAA,QACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC;AAAA,MAC1K;AAAA,IACF,OAAO;AACL,YAAM;AAAA,QACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,EAAE,KAAK,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC;AAAA,MACvO;AAAA,IACF;AAAA,EACF;AAEA,MAAI,OAAO,QAAQ,QAAQ;AACzB,UAAM,IAAI,OAAO,QAAQ;AACzB,UAAM;AAAA,MACJ,KAAK,SAAS,OAAO,UAAU,CAAC,MAAM,EAAE,OAAO,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,MAAM,EAAE,MAAM,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,MAAM,EAAE,MAAM,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC;AAAA,IACtL;AAAA,EACF;AAEA,QAAM,eAAe,OAAO,WAAW,KAAM,QAAQ,CAAC;AACtD,QAAM,UAAU,OAAO,YAAY,IAAI,IAAI,OAAO,UAAU,QAAQ,CAAC,CAAC,KAAK;AAC3E,QAAM,KAAK,EAAE;AACb,QAAM;AAAA,IACJ,eAAe,OAAO,QAAQ,QAAQ,IAAI,OAAO,QAAQ,KAAK,YAAY,OAAO,gBAAgB,WAAW;AAAA,EAC9G;AAEA,QAAM,kBAAkB,OAAO,MAAM,OAAO,CAAC,MAAM,EAAE,cAAc,MAAM;AACzE,MAAI,gBAAgB,SAAS,GAAG;AAC9B,UAAM,eAAe,CAAC,GAAG,IAAI,IAAI,gBAAgB,QAAQ,CAAC,MAAM,EAAE,YAAa,CAAC,CAAC;AACjF,UAAM,KAAK,EAAE;AACb,UAAM;AAAA,MACJ,oBAAoB,gBAAgB,MAAM,IAAI,OAAO,QAAQ,KAAK;AAAA,IACpE;AACA,eAAW,OAAO,aAAa,MAAM,GAAG,CAAC,GAAG;AAC1C,YAAM,KAAK,SAAS,GAAG,EAAE;AAAA,IAC3B;AACA,QAAI,aAAa,SAAS,GAAG;AAC3B,YAAM,KAAK,eAAe,aAAa,SAAS,CAAC,OAAO;AAAA,IAC1D;AAAA,EACF;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,oBAAoB,SAAkC;AAC7D,QAAM,QAAkB,CAAC;AACzB,QAAM,cAAc,OAAO,KAAK,QAAQ,OAAO;AAC/C,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,MAAM;AAChF,QAAM,WAAW;AAEjB,QAAM,KAAK,SAAS,QAAQ,QAAQ,MAAM,QAAQ,OAAO,WAAW,QAAQ,QAAQ,OAAO;AAC3F,QAAM;AAAA,IACJ,KAAK,SAAS,OAAO,UAAU,CAAC,KAAK,gBAAkB,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,CAAC,CAAC,KAAK,MAAM,SAAS,CAAC,CAAC;AAAA,EACvH;AACA,QAAM,UAAU,aAAa,IAAI,WAAW,IAAI,IAAI,IAAI;AACxD,QAAM,KAAK,OAAO,SAAS,OAAO,OAAO,CAAC;AAE1C,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,QAAQ,QAAQ,IAAI;AAC9B,UAAM,UAAU,GAAG,EAAE,KAAK,QAAQ,CAAC,CAAC,SAAW,EAAE,IAAI,QAAQ,CAAC,CAAC;AAC/D,UAAM;AAAA,MACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,QAAQ,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC;AAAA,IAC/H;AAAA,EACF;AAEA,MAAI,QAAQ,QAAQ;AAClB,UAAM;AAAA,MACJ,KAAK,SAAS,OAAO,UAAU,CAAC,OAAO,QAAQ,OAAO,OAAO,KAAM,QAAQ,CAAC,IAAI,YAAc,QAAQ,OAAO,MAAM,KAAM,QAAQ,CAAC,IAAI,KAAK,SAAS,QAAQ,CAAC;AAAA,IAC/J;AAAA,EACF;AAEA,QAAM,UAAU,QAAQ,YAAY,IAAI,IAAI,QAAQ,UAAU,QAAQ,CAAC,CAAC,KAAK;AAC7E,QAAM,eAAe,QAAQ,gBAAgB,KAAM,QAAQ,CAAC;AAC5D,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,iBAAiB,OAAO,sBAAsB,WAAW,GAAG;AAEvE,SAAO,MAAM,KAAK,IAAI;AACxB;AAIA,eAAe,yBAAyB,YAAyC;AAC/E,MAAI;AACF,UAAM,MAAM,MAAM,aAAa,YAAY,YAAY,GAAG;AAC1D,UAAM,UAAU,eAAe,GAAG;AAElC,QAAI,CAAC,WAAW,OAAO,QAAQ,YAAY,YAAY;AACrD,cAAQ,MAAM,mDAAmD;AACjE,UAAI,SAAS;AACX,cAAM,OAAO,OAAO,KAAK,OAAiB,EACvC,MAAM,GAAG,CAAC,EACV,KAAK,IAAI;AACZ,gBAAQ,MAAM,UAAU,OAAO,OAAO,GAAG,OAAO,iBAAiB,IAAI,OAAO,EAAE,EAAE;AAAA,MAClF;AACA,cAAQ;AAAA,QACN;AAAA;AAAA;AAAA,MACF;AACA,cAAQ,KAAK,CAAC;AAAA,IAChB;AAEA,WAAO;AAAA,EACT,SAAS,KAAK;AACZ,UAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC3D,QACE,8FAA8F;AAAA,MAC5F;AAAA,IACF,GACA;AACA,YAAM,MAAW,aAAQ,UAAU;AACnC,cAAQ,MAAM,wEAAwE;AACtF,UAAI,QAAQ,SAAS,QAAQ,QAAQ;AACnC,gBAAQ;AAAA,UACN;AAAA,QACF;AAAA,MACF,OAAO;AACL,gBAAQ,MAAM,mDAAmD;AAAA,MACnE;AACA,cAAQ,MAAM;AAAA,IAChB;AACA,YAAQ,MAAM,0BAA0B,GAAG;AAC3C,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAEA,eAAe,WAAW,WAAoB,YAA4C;AAExF,MAAI,cAAc,WAAW,SAAS,GAAG;AACvC,UAAM,mBAAmB,UAAU;AAAA,EACrC;AAGA,MAAI,WAAW;AACb,UAAM,aAAkB,aAAQ,QAAQ,IAAI,GAAG,SAAS;AACxD,UAAM,OAAO,SAAS,YAAY,EAAE,gBAAgB,MAAM,CAAC;AAC3D,QAAI,CAAC,MAAM,OAAO,GAAG;AACnB,cAAQ,MAAM,0BAA0B,UAAU,EAAE;AACpD,cAAQ,KAAK,CAAC;AAAA,IAChB;AACA,YAAQ,MAAM,kCAAkC,UAAU,EAAE;AAC5D,WAAO,yBAAyB,UAAU;AAAA,EAC5C;AAGA,QAAM,QAAQ,WAAW,QAAQ,IAAI,CAAC;AACtC,MAAI,OAAO;AACT,YAAQ,MAAM,oCAAoC,KAAK,EAAE;AACzD,WAAO,yBAAyB,KAAK;AAAA,EACvC;AAGA,QAAM,EAAE,WAAW,IAAI,MAAM,OAAO,aAAa;AACjD,SAAO,IAAI,WAAW;AACxB;AAIA,SAAS,cAAc,MAMrB;AACA,MAAI;AACJ,MAAI;AACJ,MAAI,aAAuB,CAAC;AAC5B,MAAI,OAAO;AACX,QAAM,QAAkB,CAAC;AAEzB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAClB,QAAI,QAAQ,cAAc,QAAQ,cAAc,QAAQ,kBAAkB,QAAQ,UAAU;AAC1F,UAAI,IAAI,KAAK,KAAK,QAAQ;AACxB,gBAAQ,MAAM,UAAU,GAAG,mBAAmB;AAC9C,gBAAQ,KAAK,CAAC;AAAA,MAChB;AACA,YAAM,QAAQ,KAAK,EAAE,CAAC;AACtB,UAAI,QAAQ,WAAY,cAAa;AAAA,eAC5B,QAAQ,WAAY,aAAY;AAAA,eAChC,QAAQ,SAAU,QAAO,KAAK,IAAI,GAAG,SAAS,OAAO,EAAE,KAAK,CAAC;AAAA;AAEpE,qBAAa,MACV,MAAM,GAAG,EACT,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,OAAO;AAAA,IACrB,WAAW,IAAI,WAAW,IAAI,GAAG;AAC/B,UAAI,CAAC,YAAY,IAAI,GAAG,GAAG;AACzB,gBAAQ,MAAM,iBAAiB,GAAG,EAAE;AACpC,gBAAQ,KAAK,CAAC;AAAA,MAChB;AAAA,IACF,OAAO;AACL,YAAM,KAAK,GAAG;AAAA,IAChB;AAAA,EACF;AAEA,SAAO,EAAE,YAAY,WAAW,YAAY,MAAM,MAAM;AAC1D;AAIA,eAAe,eAAe,MAAgB;AAC5C,QAAM,EAAE,YAAY,WAAW,YAAY,MAAM,MAAM,IAAI,cAAc,IAAI;AAE7E,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ,MAAM,mCAAmC;AACjD,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,YAAsB,CAAC;AAC7B,aAAW,KAAK,OAAO;AACrB,cAAU,KAAK,GAAG,iBAAiB,CAAC,CAAC;AAAA,EACvC;AAEA,MAAI,UAAU,WAAW,GAAG;AAC1B,YAAQ,MAAM,4BAA4B;AAC1C,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,UAAU,MAAM,WAAW,WAAW,UAAU;AACtD,QAAM,UAAwB,CAAC;AAE/B,MAAI;AACF,eAAW,YAAY,WAAW;AAChC,UAAI;AACF,cAAM,MAAM,MAAM,aAAkB,aAAQ,QAAQ,GAAG,YAAY,GAAG;AACtE,cAAM,aAAyB,IAAI,SAAS,WAAW,IAAI,WAAW,IAAI,UAAU;AAEpF,YAAI,CAAC,WAAW,YAAY,CAAC,WAAW,WAAW,CAAC,WAAW,SAAS;AACtE,kBAAQ;AAAA,YACN,UAAU,QAAQ;AAAA,UACpB;AACA;AAAA,QACF;AAGA,YAAI;AAKJ,YAAI,IAAI,iBAAiB;AAEvB,4BAAkB,OAAO,OAAO,OAAO;AACrC,kBAAM;AAAA,cACJ;AAAA,cACA,MAAM;AAAA,cACN;AAAA,YACF,IAAI,MAAM,QAAQ,eAAe,YAAY;AAC3C,qBAAO,IAAI,gBAAgB,OAAO,EAAE;AAAA,YACtC,CAAC;AACD,mBAAO;AAAA,cACL,QAAQ,OAAO;AAAA,cACf,MAAM,OAAO,QAAQ;AAAA,cACrB,UAAU,OAAO,YAAY;AAAA,YAC/B;AAAA,UACF;AAAA,QACF,WAAW,QAAQ,YAAY,WAAW,QAAQ,GAAG;AAEnD,4BAAkB,OAAO,UAAU;AACjC,kBAAM,EAAE,QAAQ,MAAM,SAAS,IAAI,MAAM,QAAQ,eAAe,YAAY;AAC1E,qBAAO,QAAQ,QAAQ,WAAW,UAAU,KAAK;AAAA,YACnD,CAAC;AACD,mBAAO,EAAE,QAAQ,QAAQ,MAAM,SAAS;AAAA,UAC1C;AAAA,QACF,OAAO;AACL,kBAAQ;AAAA,YACN,uBAAuB,QAAQ;AAAA,UACjC;AACA,4BAAkB,OAAO,WAAW,EAAE,QAAQ,MAAM;AAAA,QACtD;AAEA,YAAI,OAAO,GAAG;AAEZ,gBAAM,EAAE,WAAW,IAAI,MAAM,OAAO,QAAa;AACjD,gBAAM,aAAa,WAAW;AAC9B,gBAAM,aAA2B,CAAC;AAElC,mBAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,oBAAQ,MAAM,kBAAkB,IAAI,CAAC,IAAI,IAAI,KAAK;AAClD,kBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,OAAO;AACjE,mBAAO,SAAS,aAAa;AAC7B,mBAAO,SAAS,WAAW;AAC3B,uBAAW,KAAK,MAAM;AACtB,oBAAQ,KAAK,MAAM;AAAA,UACrB;AAEA,gBAAM,UAAU,cAAc,UAAU;AACxC,kBAAQ,IAAI,OAAO,oBAAoB,OAAO,IAAI,IAAI;AAAA,QACxD,OAAO;AACL,gBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,OAAO;AACjE,kBAAQ,KAAK,MAAM;AAEnB,kBAAQ,IAAI,OAAO,YAAY,MAAM,IAAI,IAAI;AAAA,QAC/C;AAAA,MACF,SAAS,KAAK;AACZ,gBAAQ;AAAA,UACN,sBAAsB,QAAQ,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,QACrF;AAAA,MACF;AAAA,IACF;AAEA,QAAI,cAAc,QAAQ,SAAS,GAAG;AACpC,YAAM,SAAS,QAAQ,WAAW,IAAI,QAAQ,CAAC,IAAI;AACnD,YAAM,YAAiB,aAAa,aAAQ,UAAU,CAAC;AACvD,YAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,YAAM,eAAe,YAAY,KAAK,UAAU,QAAQ,MAAM,CAAC,GAAG,OAAO;AACzE,cAAQ,IAAI,oBAAoB,UAAU,EAAE;AAAA,IAC9C;AAAA,EACF,UAAE;AACA,UAAM,QAAQ,SAAS,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EACzC;AAGA,MAAI,QAAQ,WAAW,GAAG;AACxB,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAEA,KAAK,EAAE,MAAM,CAAC,QAAQ;AACpB,UAAQ,MAAM,GAAG;AACjB,UAAQ,KAAK,CAAC;AAChB,CAAC;","names":[]}
1
+ {"version":3,"sources":["../src/cli.ts","../src/cli-utils.ts"],"sourcesContent":["#!/usr/bin/env node\n\nimport { readdirSync, statSync } from 'node:fs';\nimport { readFile as readFileAsync, writeFile as writeFileAsync, mkdir } from 'node:fs/promises';\nimport * as path from 'node:path';\nimport type { AxlRuntime } from '@axlsdk/axl';\nimport { evalCompare } from './compare.js';\nimport { runEval } from './runner.js';\nimport { rescore } from './rescore.js';\nimport { aggregateRuns } from './multi-run.js';\nimport type { MultiRunSummary } from './multi-run.js';\nimport type { EvalConfig, EvalResult } from './types.js';\nimport {\n findConfig,\n resolveRuntime,\n importModule,\n registerConditions,\n CONFIG_CANDIDATES,\n} from './cli-utils.js';\n\nconst KNOWN_FLAGS = new Set([\n '--output',\n '--config',\n '--conditions',\n '--fail-on-regression',\n '--threshold',\n '--runs',\n '--capture-traces',\n]);\n\nasync function main() {\n const args = process.argv.slice(2);\n\n if (args.length === 0 || args[0] === '--help' || args[0] === '-h') {\n console.log(`\nUsage:\n axl-eval <path> Run eval file(s)\n axl-eval <path> --runs <n> Run eval N times (multi-run)\n axl-eval <path> --output <file> Save results to JSON\n axl-eval <path> --config <file> Use config file for runtime\n axl-eval <path> --conditions <list> Node.js import conditions (comma-separated)\n axl-eval <path> --capture-traces Populate EvalItem.traces on every item\n (success + failure). Adds memory\n overhead proportional to dataset size\n x turns x agents; off by default.\n axl-eval rescore <results> <eval-file> Re-run scorers on saved outputs\n axl-eval compare <a> <b> Compare two eval result files\n axl-eval compare <a> <b> --threshold <v> Set regression threshold (global or per-scorer)\n axl-eval compare <a> <b> --fail-on-regression Exit 1 if regressions\n\nConfig auto-detection (when --config is not specified):\n ${CONFIG_CANDIDATES.join(' -> ')}\n\nWhen a config is found, the exported AxlRuntime is passed to executeWorkflow\nand cost is tracked automatically via runtime.trackCost().\nWhen no config is found, a bare AxlRuntime is created (providers from env vars).\n`);\n process.exit(0);\n }\n\n if (args[0] === 'compare') {\n await runCompare(args.slice(1));\n return;\n }\n\n if (args[0] === 'rescore') {\n await runRescore(args.slice(1));\n return;\n }\n\n await runEvalCommand(args);\n}\n\nfunction parseThresholdArg(args: string[]): Record<string, number> | number | undefined {\n const idx = args.indexOf('--threshold');\n if (idx === -1 || idx + 1 >= args.length) return undefined;\n const value = args[idx + 1];\n\n // Global numeric threshold: --threshold 0.05\n if (/^[\\d.]+$/.test(value)) return parseFloat(value);\n\n // Per-scorer map: --threshold accuracy=0,tone=0.1\n const map: Record<string, number> = {};\n for (const pair of value.split(',')) {\n const [key, val] = pair.split('=');\n if (key && val != null) map[key.trim()] = parseFloat(val.trim());\n }\n return Object.keys(map).length > 0 ? map : undefined;\n}\n\nasync function runCompare(args: string[]) {\n const failOnRegression = args.includes('--fail-on-regression');\n const thresholds = parseThresholdArg(args);\n const files = args.filter(\n (a, i) => !a.startsWith('--') && !(i > 0 && args[i - 1] === '--threshold'),\n );\n\n if (files.length !== 2) {\n console.error('Usage: axl-eval compare <baseline.json> <candidate.json> [--threshold <value>]');\n process.exit(1);\n }\n\n const baselineRaw = JSON.parse(await readFileAsync(files[0], 'utf-8'));\n const candidateRaw = JSON.parse(await readFileAsync(files[1], 'utf-8'));\n const baseline: EvalResult | EvalResult[] = baselineRaw;\n const candidate: EvalResult | EvalResult[] = candidateRaw;\n const compareOptions = thresholds != null ? { thresholds } : undefined;\n const comparison = evalCompare(baseline, candidate, compareOptions);\n\n console.log(\n `\\nCompare: baseline (${comparison.baseline.id.slice(0, 8)}) -> candidate (${comparison.candidate.id.slice(0, 8)})\\n`,\n );\n\n const scorerNames = Object.keys(comparison.scorers);\n const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 6);\n const hasCI = scorerNames.some((n) => comparison.scorers[n].ci != null);\n\n const ciHeader = hasCI ? ' CI 95% Sig' : '';\n const ciRule = hasCI ? ' ---------------- ---' : '';\n console.log(` ${'Scorer'.padEnd(maxNameLen)} Baseline Candidate Delta Change${ciHeader}`);\n console.log(` ${''.padEnd(maxNameLen, '-')} -------- --------- -------- ------${ciRule}`);\n\n for (const name of scorerNames) {\n const s = comparison.scorers[name];\n const sign = s.delta > 0 ? '+' : '';\n let line = ` ${name.padEnd(maxNameLen)} ${s.baselineMean.toFixed(3).padStart(8)} ${s.candidateMean.toFixed(3).padStart(9)} ${(sign + s.delta.toFixed(3)).padStart(8)} ${(sign + s.deltaPercent.toFixed(1) + '%').padStart(6)}`;\n if (hasCI) {\n if (s.ci) {\n const lo = (s.ci.lower >= 0 ? '+' : '') + s.ci.lower.toFixed(4);\n const hi = (s.ci.upper >= 0 ? '+' : '') + s.ci.upper.toFixed(4);\n const sig = s.significant ? ' *' : ' ';\n line += ` [${lo}, ${hi}]${sig}`;\n } else {\n line += ' —'.padEnd(22);\n }\n }\n console.log(line);\n }\n\n if (comparison.timing) {\n const t = comparison.timing;\n const sign = t.delta > 0 ? '+' : '';\n console.log(\n `\\n Timing: baseline ${(t.baselineMean / 1000).toFixed(2)}s -> candidate ${(t.candidateMean / 1000).toFixed(2)}s (${sign}${t.deltaPercent.toFixed(1)}%)`,\n );\n }\n if (comparison.cost) {\n const c = comparison.cost;\n const sign = c.delta > 0 ? '+' : '';\n console.log(\n ` Cost: baseline $${c.baselineTotal.toFixed(2)} -> candidate $${c.candidateTotal.toFixed(2)} (${sign}${c.deltaPercent.toFixed(1)}%)`,\n );\n }\n\n const baselineRef = Array.isArray(baseline) ? baseline[0] : baseline;\n const stable = Math.max(\n 0,\n baselineRef.items.length - comparison.regressions.length - comparison.improvements.length,\n );\n console.log(\n `\\n Regressions: ${comparison.regressions.length} | Improvements: ${comparison.improvements.length} | Stable: ${stable}\\n`,\n );\n\n if (failOnRegression && comparison.regressions.length > 0) {\n // When CI is available, only fail on significant regressions\n const hasSignificance = scorerNames.some((n) => comparison.scorers[n].significant != null);\n if (hasSignificance) {\n const hasSignificantRegression = scorerNames.some(\n (n) => comparison.scorers[n].significant === true && comparison.scorers[n].delta < 0,\n );\n if (hasSignificantRegression) process.exit(1);\n } else {\n process.exit(1);\n }\n }\n}\n\nasync function runRescore(args: string[]) {\n const { outputPath, configArg, conditions, paths } = parseEvalArgs(args);\n\n if (paths.length < 2) {\n console.error('Usage: axl-eval rescore <results.json> <eval-file> [--output <file>]');\n process.exit(1);\n }\n\n const [resultsPath, evalFilePath] = paths;\n const raw = JSON.parse(await readFileAsync(resultsPath, 'utf-8'));\n const results: EvalResult[] = Array.isArray(raw) ? raw : [raw];\n\n const runtime = await getRuntime(configArg, conditions);\n try {\n const mod = await importModule(path.resolve(evalFilePath), import.meta.url);\n const evalConfig: EvalConfig = mod.default?.default ?? mod.default ?? mod.config ?? mod;\n\n if (!evalConfig.scorers || evalConfig.scorers.length === 0) {\n console.error(`Error: ${evalFilePath} does not export scorers`);\n process.exit(1);\n }\n\n const rescored: EvalResult[] = [];\n for (const resultData of results) {\n rescored.push(await rescore(resultData, evalConfig.scorers, runtime));\n }\n\n for (const r of rescored) {\n console.log('\\n' + formatTable(r) + '\\n');\n }\n\n if (outputPath) {\n const output = Array.isArray(raw) ? rescored : rescored[0];\n const outputDir = path.dirname(path.resolve(outputPath));\n await mkdir(outputDir, { recursive: true });\n await writeFileAsync(outputPath, JSON.stringify(output, null, 2), 'utf-8');\n console.log(`Rescored results saved to ${outputPath}`);\n }\n } finally {\n await runtime.shutdown().catch(() => {});\n }\n}\n\nfunction collectEvalFiles(p: string): string[] {\n const resolved = path.resolve(p);\n try {\n const stat = statSync(resolved);\n if (stat.isDirectory()) {\n const entries = readdirSync(resolved);\n return entries\n .filter((e) => /\\.eval\\.[mc]?[jt]sx?$/.test(e))\n .map((e) => path.join(resolved, e));\n }\n return [resolved];\n } catch {\n return [resolved];\n }\n}\n\n/** Render observed workflows for a terse header line. Multiple → comma-separated. */\nfunction formatWorkflows(workflows: unknown): string {\n if (!Array.isArray(workflows) || workflows.length === 0) return '(unknown)';\n return (workflows as unknown[]).filter((w): w is string => typeof w === 'string').join(', ');\n}\n\nfunction formatTable(result: EvalResult): string {\n const lines: string[] = [];\n const scorerNames = Object.keys(result.summary.scorers);\n const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 'Scorer'.length);\n const colWidth = 8;\n\n lines.push(\n `Eval: ${formatWorkflows(result.metadata.workflows)} x ${result.dataset} (${result.summary.count} items)`,\n );\n lines.push(\n ` ${'Scorer'.padEnd(maxNameLen)} ${'Mean'.padStart(colWidth)} ${'Min'.padStart(colWidth)} ${'Max'.padStart(colWidth)} ${'p50'.padStart(colWidth)} ${'p95'.padStart(colWidth)}`,\n );\n\n const ruleLen = maxNameLen + 2 + (colWidth + 2) * 5;\n lines.push(' ' + '\\u2500'.repeat(ruleLen));\n\n for (const name of scorerNames) {\n const s = result.summary.scorers[name];\n // Detect scorers with no valid scores (all items errored → computeStats([]) → all zeros)\n const validScoreCount = result.items.filter((i) => !i.error && i.scores[name] != null).length;\n if (validScoreCount === 0) {\n lines.push(\n ` ${name.padEnd(maxNameLen)} ${'--'.padStart(colWidth)} ${'--'.padStart(colWidth)} ${'--'.padStart(colWidth)} ${'--'.padStart(colWidth)} ${'--'.padStart(colWidth)}`,\n );\n } else {\n lines.push(\n ` ${name.padEnd(maxNameLen)} ${s.mean.toFixed(2).padStart(colWidth)} ${s.min.toFixed(2).padStart(colWidth)} ${s.max.toFixed(2).padStart(colWidth)} ${s.p50.toFixed(2).padStart(colWidth)} ${s.p95.toFixed(2).padStart(colWidth)}`,\n );\n }\n }\n\n if (result.summary.timing) {\n const t = result.summary.timing;\n lines.push(\n ` ${'Timing'.padEnd(maxNameLen)} ${(t.mean / 1000).toFixed(2).padStart(colWidth)}s ${(t.p50 / 1000).toFixed(2).padStart(colWidth)}s ${(t.p95 / 1000).toFixed(2).padStart(colWidth)}s`,\n );\n }\n\n const durationSec = (result.duration / 1000).toFixed(1);\n const costStr = result.totalCost > 0 ? `$${result.totalCost.toFixed(2)}` : '$0.00';\n lines.push('');\n lines.push(\n ` Failures: ${result.summary.failures}/${result.summary.count} | Cost: ${costStr} | Duration: ${durationSec}s`,\n );\n\n const itemsWithErrors = result.items.filter((i) => i.scorerErrors?.length);\n if (itemsWithErrors.length > 0) {\n const uniqueErrors = [...new Set(itemsWithErrors.flatMap((i) => i.scorerErrors!))];\n lines.push('');\n lines.push(\n ` Scorer errors (${itemsWithErrors.length}/${result.summary.count} items affected):`,\n );\n for (const err of uniqueErrors.slice(0, 5)) {\n lines.push(` - ${err}`);\n }\n if (uniqueErrors.length > 5) {\n lines.push(` ... and ${uniqueErrors.length - 5} more`);\n }\n }\n\n return lines.join('\\n');\n}\n\nfunction formatMultiRunTable(summary: MultiRunSummary): string {\n const lines: string[] = [];\n const scorerNames = Object.keys(summary.scorers);\n const maxNameLen = Math.max(...scorerNames.map((n) => n.length), 'Scorer'.length);\n const colWidth = 16;\n\n lines.push(\n `Eval: ${formatWorkflows(summary.workflows)} x ${summary.dataset} \\u2014 ${summary.runCount} runs`,\n );\n lines.push(\n ` ${'Scorer'.padEnd(maxNameLen)} ${'Mean \\u00b1 Std'.padStart(colWidth)} ${'Min'.padStart(8)} ${'Max'.padStart(8)}`,\n );\n const ruleLen = maxNameLen + 2 + colWidth + 2 + 8 + 2 + 8;\n lines.push(' ' + '\\u2500'.repeat(ruleLen));\n\n for (const name of scorerNames) {\n const s = summary.scorers[name];\n const meanStd = `${s.mean.toFixed(3)} \\u00b1 ${s.std.toFixed(3)}`;\n lines.push(\n ` ${name.padEnd(maxNameLen)} ${meanStd.padStart(colWidth)} ${s.min.toFixed(3).padStart(8)} ${s.max.toFixed(3).padStart(8)}`,\n );\n }\n\n if (summary.timing) {\n lines.push(\n ` ${'Timing'.padEnd(maxNameLen)} ${((summary.timing.mean / 1000).toFixed(2) + ' \\u00b1 ' + (summary.timing.std / 1000).toFixed(2) + 's').padStart(colWidth)}`,\n );\n }\n\n const costStr = summary.totalCost > 0 ? `$${summary.totalCost.toFixed(2)}` : '$0.00';\n const durationStr = (summary.totalDuration / 1000).toFixed(1);\n lines.push('');\n lines.push(` Total Cost: ${costStr} | Total Duration: ${durationStr}s`);\n\n return lines.join('\\n');\n}\n\n// ── Runtime resolution ─────────────────────────────────────────────\n\nasync function resolveRuntimeFromConfig(configPath: string): Promise<AxlRuntime> {\n try {\n const mod = await importModule(configPath, import.meta.url);\n const runtime = resolveRuntime(mod) as AxlRuntime;\n\n if (!runtime || typeof runtime.execute !== 'function') {\n console.error(`Config must export a default AxlRuntime instance.`);\n if (runtime) {\n const keys = Object.keys(runtime as object)\n .slice(0, 5)\n .join(', ');\n console.error(` Got: ${typeof runtime}${keys ? ` with keys: { ${keys} }` : ''}`);\n }\n console.error(\n `Example:\\n import { AxlRuntime } from '@axlsdk/axl';\\n export default new AxlRuntime({ ... });`,\n );\n process.exit(1);\n }\n\n return runtime;\n } catch (err) {\n const msg = err instanceof Error ? err.message : String(err);\n if (\n /Cannot use import statement|Unexpected reserved word|top-level await|exports is not defined/.test(\n msg,\n )\n ) {\n const ext = path.extname(configPath);\n console.error(`[axl-eval] Config failed to load due to a CJS/ESM compatibility issue.`);\n if (ext === '.ts' || ext === '.tsx') {\n console.error(\n ` Tip: try renaming to .mts to force ESM format, or ensure tsx is installed and up to date.`,\n );\n } else {\n console.error(` Tip: add \"type\": \"module\" to your package.json.`);\n }\n console.error();\n }\n console.error(`Failed to load config:`, err);\n process.exit(1);\n }\n}\n\nasync function getRuntime(configArg?: string, conditions?: string[]): Promise<AxlRuntime> {\n // Register import conditions before any config loading\n if (conditions && conditions.length > 0) {\n await registerConditions(conditions);\n }\n\n // 1. Explicit --config\n if (configArg) {\n const configPath = path.resolve(process.cwd(), configArg);\n const stat = statSync(configPath, { throwIfNoEntry: false });\n if (!stat?.isFile()) {\n console.error(`Config file not found: ${configPath}`);\n process.exit(1);\n }\n console.error(`[axl-eval] Loading config from ${configPath}`);\n return resolveRuntimeFromConfig(configPath);\n }\n\n // 2. Auto-detect\n const found = findConfig(process.cwd());\n if (found) {\n console.error(`[axl-eval] Auto-detected config: ${found}`);\n return resolveRuntimeFromConfig(found);\n }\n\n // 3. Bare runtime (providers from env vars)\n const { AxlRuntime } = await import('@axlsdk/axl');\n return new AxlRuntime();\n}\n\n// ── Arg parsing ────────────────────────────────────────────────────\n\nfunction parseEvalArgs(args: string[]): {\n outputPath?: string;\n configArg?: string;\n conditions: string[];\n runs: number;\n captureTraces: boolean;\n paths: string[];\n} {\n let outputPath: string | undefined;\n let configArg: string | undefined;\n let conditions: string[] = [];\n let runs = 1;\n let captureTraces = false;\n const paths: string[] = [];\n\n for (let i = 0; i < args.length; i++) {\n const arg = args[i];\n if (arg === '--output' || arg === '--config' || arg === '--conditions' || arg === '--runs') {\n if (i + 1 >= args.length) {\n console.error(`Error: ${arg} requires a value`);\n process.exit(1);\n }\n const value = args[++i];\n if (arg === '--output') outputPath = value;\n else if (arg === '--config') configArg = value;\n else if (arg === '--runs') runs = Math.max(1, parseInt(value, 10) || 1);\n else\n conditions = value\n .split(',')\n .map((c) => c.trim())\n .filter(Boolean);\n } else if (arg === '--capture-traces') {\n // Boolean flag — no value consumed.\n captureTraces = true;\n } else if (arg.startsWith('--')) {\n if (!KNOWN_FLAGS.has(arg)) {\n console.error(`Unknown flag: ${arg}`);\n process.exit(1);\n }\n } else {\n paths.push(arg);\n }\n }\n\n return { outputPath, configArg, conditions, runs, captureTraces, paths };\n}\n\n// ── Main eval command ──────────────────────────────────────────────\n\nasync function runEvalCommand(args: string[]) {\n const { outputPath, configArg, conditions, runs, captureTraces, paths } = parseEvalArgs(args);\n\n if (paths.length === 0) {\n console.error('Error: No eval file path provided');\n process.exit(1);\n }\n\n const evalFiles: string[] = [];\n for (const p of paths) {\n evalFiles.push(...collectEvalFiles(p));\n }\n\n if (evalFiles.length === 0) {\n console.error('Error: No eval files found');\n process.exit(1);\n }\n\n const runtime = await getRuntime(configArg, conditions);\n const results: EvalResult[] = [];\n\n try {\n for (const filePath of evalFiles) {\n try {\n const mod = await importModule(path.resolve(filePath), import.meta.url);\n const evalConfig: EvalConfig = mod.default?.default ?? mod.default ?? mod.config ?? mod;\n\n if (!evalConfig.workflow || !evalConfig.dataset || !evalConfig.scorers) {\n console.error(\n `Error: ${filePath} does not export a valid eval config (missing workflow, dataset, or scorers)`,\n );\n continue;\n }\n\n // Resolve executeWorkflow: custom export > registered workflow > passthrough\n let executeWorkflow: (\n input: unknown,\n rt?: unknown,\n ) => Promise<{ output: unknown; cost?: number; metadata?: Record<string, unknown> }>;\n\n if (mod.executeWorkflow) {\n // Wrap custom executeWorkflow with trackExecution for cost + metadata attribution\n executeWorkflow = async (input, rt) => {\n const {\n result,\n cost: trackedCost,\n metadata,\n } = await runtime.trackExecution(async () => {\n return mod.executeWorkflow(input, rt);\n });\n return {\n output: result.output,\n cost: result.cost ?? trackedCost,\n metadata: result.metadata ?? metadata,\n };\n };\n } else if (runtime.getWorkflow(evalConfig.workflow)) {\n // No executeWorkflow exported but workflow is registered — use runtime.execute()\n executeWorkflow = async (input) => {\n const { result, cost, metadata } = await runtime.trackExecution(async () => {\n return runtime.execute(evalConfig.workflow, input);\n });\n return { output: result, cost, metadata };\n };\n } else {\n console.warn(\n `[axl-eval] Warning: ${filePath} does not export executeWorkflow — using input passthrough`,\n );\n executeWorkflow = async (input) => ({ output: input });\n }\n\n // When --capture-traces is set, forward it to runEval so the runner\n // wraps each item's execution in a nested trackExecution({ captureTraces: true }).\n // This populates EvalItem.traces on both success (via tracked.traces)\n // and failure (via the axlCapturedTraces side-channel on the thrown\n // error). Nested trackExecution scopes walk the AsyncLocalStorage\n // parent chain, so the outer trackExecution above still observes\n // cost/metadata correctly.\n const runOptions = captureTraces ? { captureTraces: true } : undefined;\n\n if (runs > 1) {\n // Multi-run mode\n const { randomUUID } = await import('node:crypto');\n const runGroupId = randomUUID();\n const runResults: EvalResult[] = [];\n\n for (let r = 0; r < runs; r++) {\n console.error(`[axl-eval] Run ${r + 1}/${runs}...`);\n const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);\n result.metadata.runGroupId = runGroupId;\n result.metadata.runIndex = r;\n runResults.push(result);\n results.push(result);\n }\n\n const summary = aggregateRuns(runResults);\n console.log('\\n' + formatMultiRunTable(summary) + '\\n');\n } else {\n const result = await runEval(evalConfig, executeWorkflow, runtime, runOptions);\n results.push(result);\n\n console.log('\\n' + formatTable(result) + '\\n');\n }\n } catch (err) {\n console.error(\n `Error running eval ${filePath}: ${err instanceof Error ? err.message : String(err)}`,\n );\n }\n }\n\n if (outputPath && results.length > 0) {\n const output = results.length === 1 ? results[0] : results;\n const outputDir = path.dirname(path.resolve(outputPath));\n await mkdir(outputDir, { recursive: true });\n await writeFileAsync(outputPath, JSON.stringify(output, null, 2), 'utf-8');\n console.log(`Results saved to ${outputPath}`);\n }\n } finally {\n await runtime.shutdown().catch(() => {});\n }\n\n // Exit with non-zero code if no evals succeeded\n if (results.length === 0) {\n process.exit(1);\n }\n}\n\nmain().catch((err) => {\n console.error(err);\n process.exit(1);\n});\n","/**\n * CLI utilities for config detection, runtime resolution, and loader registration.\n *\n * These are duplicated from @axlsdk/studio (cli-utils.ts, resolve-runtime.ts,\n * eval-loader.ts, cli.ts) because @axlsdk/eval cannot depend on @axlsdk/studio.\n * Keep in sync with the studio versions if either changes.\n */\n\nimport { resolve } from 'node:path';\nimport { existsSync } from 'node:fs';\nimport { pathToFileURL } from 'node:url';\n\n// ── Config auto-detection ──────────────────────────────────────────\n\nexport const CONFIG_CANDIDATES = [\n 'axl.config.mts',\n 'axl.config.ts',\n 'axl.config.mjs',\n 'axl.config.js',\n];\n\nexport function findConfig(cwd: string): string | undefined {\n for (const name of CONFIG_CANDIDATES) {\n const p = resolve(cwd, name);\n if (existsSync(p)) return p;\n }\n return undefined;\n}\n\n// ── Extension helpers ──────────────────────────────────────────────\n\n/** Returns true if the file is TypeScript and needs tsx to load. */\nexport function needsTsxLoader(filePath: string): boolean {\n return /\\.[mc]?tsx?$/.test(filePath);\n}\n\n// ── Runtime resolution ─────────────────────────────────────────────\n\n/**\n * Resolve the AxlRuntime from a dynamically imported config module.\n *\n * Module shapes handled:\n * - ESM `export default runtime` → mod.default is the runtime\n * - CJS compiled from `export default runtime` → mod.default.default is the runtime\n * - CJS `module.exports = runtime` → mod.default is the runtime\n * - Named `export { runtime }` → mod.runtime is the runtime\n */\nexport function resolveRuntime(mod: Record<string, any>): unknown {\n const def = mod.default as Record<string, any> | undefined;\n return def?.default ?? def ?? mod.runtime;\n}\n\n// ── Module loading ────────────────────────────────────────────────\n\n// Lazily resolved tsImport function from tsx. `undefined` = not yet checked,\n// `null` = tsx not available.\nlet tsImportFn:\n | ((specifier: string, parentURL: string) => Promise<Record<string, any>>)\n | null\n | undefined;\n\n/**\n * Import a module, using tsx's `tsImport()` for TypeScript files.\n *\n * `tsImport()` handles ESM/CJS format correctly without process-wide side effects —\n * no need for `register()` hooks or ESM-forcing workarounds. Falls back to regular\n * `import()` for non-TypeScript files or when tsx is not installed.\n */\nexport async function importModule(\n filePath: string,\n parentURL: string,\n): Promise<Record<string, any>> {\n if (needsTsxLoader(filePath)) {\n if (tsImportFn === undefined) {\n try {\n // @ts-expect-error — tsx is an optional runtime dependency\n const mod = await import('tsx/esm/api');\n tsImportFn = mod.tsImport ?? null;\n } catch {\n tsImportFn = null;\n console.warn(\n '[axl-eval] Warning: tsx is not installed. TypeScript files require tsx.\\n' +\n ' Install it with: npm install -D tsx',\n );\n }\n }\n if (tsImportFn) {\n return (await tsImportFn(pathToFileURL(filePath).href, parentURL)) as Record<string, any>;\n }\n }\n return await import(pathToFileURL(filePath).href);\n}\n\n// ── Conditions ────────────────────────────────────────────────────\n\nexport async function registerConditions(conditions: string[]): Promise<void> {\n try {\n const nodeModule = await import('node:module');\n const hookCode = [\n `const extra = ${JSON.stringify(conditions)};`,\n `export async function resolve(specifier, context, nextResolve) {`,\n ` return nextResolve(specifier, {`,\n ` ...context,`,\n ` conditions: [...new Set([...context.conditions, ...extra])],`,\n ` });`,\n `}`,\n ].join('\\n');\n nodeModule.register(`data:text/javascript,${encodeURIComponent(hookCode)}`);\n } catch {\n console.warn('[axl-eval] Warning: --conditions requires Node.js 20.6+');\n }\n}\n"],"mappings":";;;;;;;;;AAEA,SAAS,aAAa,gBAAgB;AACtC,SAAS,YAAY,eAAe,aAAa,gBAAgB,aAAa;AAC9E,YAAY,UAAU;;;ACItB,SAAS,eAAe;AACxB,SAAS,kBAAkB;AAC3B,SAAS,qBAAqB;AAIvB,IAAM,oBAAoB;AAAA,EAC/B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEO,SAAS,WAAW,KAAiC;AAC1D,aAAW,QAAQ,mBAAmB;AACpC,UAAM,IAAI,QAAQ,KAAK,IAAI;AAC3B,QAAI,WAAW,CAAC,EAAG,QAAO;AAAA,EAC5B;AACA,SAAO;AACT;AAKO,SAAS,eAAe,UAA2B;AACxD,SAAO,eAAe,KAAK,QAAQ;AACrC;AAaO,SAAS,eAAe,KAAmC;AAChE,QAAM,MAAM,IAAI;AAChB,SAAO,KAAK,WAAW,OAAO,IAAI;AACpC;AAMA,IAAI;AAYJ,eAAsB,aACpB,UACA,WAC8B;AAC9B,MAAI,eAAe,QAAQ,GAAG;AAC5B,QAAI,eAAe,QAAW;AAC5B,UAAI;AAEF,cAAM,MAAM,MAAM,OAAO,aAAa;AACtC,qBAAa,IAAI,YAAY;AAAA,MAC/B,QAAQ;AACN,qBAAa;AACb,gBAAQ;AAAA,UACN;AAAA,QAEF;AAAA,MACF;AAAA,IACF;AACA,QAAI,YAAY;AACd,aAAQ,MAAM,WAAW,cAAc,QAAQ,EAAE,MAAM,SAAS;AAAA,IAClE;AAAA,EACF;AACA,SAAO,MAAM,OAAO,cAAc,QAAQ,EAAE;AAC9C;AAIA,eAAsB,mBAAmB,YAAqC;AAC5E,MAAI;AACF,UAAM,aAAa,MAAM,OAAO,QAAa;AAC7C,UAAM,WAAW;AAAA,MACf,iBAAiB,KAAK,UAAU,UAAU,CAAC;AAAA,MAC3C;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AACX,eAAW,SAAS,wBAAwB,mBAAmB,QAAQ,CAAC,EAAE;AAAA,EAC5E,QAAQ;AACN,YAAQ,KAAK,yDAAyD;AAAA,EACxE;AACF;;;AD3FA,IAAM,cAAc,oBAAI,IAAI;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,eAAe,OAAO;AACpB,QAAM,OAAO,QAAQ,KAAK,MAAM,CAAC;AAEjC,MAAI,KAAK,WAAW,KAAK,KAAK,CAAC,MAAM,YAAY,KAAK,CAAC,MAAM,MAAM;AACjE,YAAQ,IAAI;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAiBZ,kBAAkB,KAAK,MAAM,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,CAKjC;AACG,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAEA,MAAI,KAAK,CAAC,MAAM,WAAW;AACzB,UAAM,WAAW,KAAK,MAAM,CAAC,CAAC;AAC9B;AAAA,EACF;AAEA,QAAM,eAAe,IAAI;AAC3B;AAEA,SAAS,kBAAkB,MAA6D;AACtF,QAAM,MAAM,KAAK,QAAQ,aAAa;AACtC,MAAI,QAAQ,MAAM,MAAM,KAAK,KAAK,OAAQ,QAAO;AACjD,QAAM,QAAQ,KAAK,MAAM,CAAC;AAG1B,MAAI,WAAW,KAAK,KAAK,EAAG,QAAO,WAAW,KAAK;AAGnD,QAAM,MAA8B,CAAC;AACrC,aAAW,QAAQ,MAAM,MAAM,GAAG,GAAG;AACnC,UAAM,CAAC,KAAK,GAAG,IAAI,KAAK,MAAM,GAAG;AACjC,QAAI,OAAO,OAAO,KAAM,KAAI,IAAI,KAAK,CAAC,IAAI,WAAW,IAAI,KAAK,CAAC;AAAA,EACjE;AACA,SAAO,OAAO,KAAK,GAAG,EAAE,SAAS,IAAI,MAAM;AAC7C;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,mBAAmB,KAAK,SAAS,sBAAsB;AAC7D,QAAM,aAAa,kBAAkB,IAAI;AACzC,QAAM,QAAQ,KAAK;AAAA,IACjB,CAAC,GAAG,MAAM,CAAC,EAAE,WAAW,IAAI,KAAK,EAAE,IAAI,KAAK,KAAK,IAAI,CAAC,MAAM;AAAA,EAC9D;AAEA,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ,MAAM,gFAAgF;AAC9F,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,cAAc,KAAK,MAAM,MAAM,cAAc,MAAM,CAAC,GAAG,OAAO,CAAC;AACrE,QAAM,eAAe,KAAK,MAAM,MAAM,cAAc,MAAM,CAAC,GAAG,OAAO,CAAC;AACtE,QAAM,WAAsC;AAC5C,QAAM,YAAuC;AAC7C,QAAM,iBAAiB,cAAc,OAAO,EAAE,WAAW,IAAI;AAC7D,QAAM,aAAa,YAAY,UAAU,WAAW,cAAc;AAElE,UAAQ;AAAA,IACN;AAAA,qBAAwB,WAAW,SAAS,GAAG,MAAM,GAAG,CAAC,CAAC,mBAAmB,WAAW,UAAU,GAAG,MAAM,GAAG,CAAC,CAAC;AAAA;AAAA,EAClH;AAEA,QAAM,cAAc,OAAO,KAAK,WAAW,OAAO;AAClD,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,CAAC;AAClE,QAAM,QAAQ,YAAY,KAAK,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,MAAM,IAAI;AAEtE,QAAM,WAAW,QAAQ,4BAA4B;AACrD,QAAM,SAAS,QAAQ,4BAA4B;AACnD,UAAQ,IAAI,KAAK,SAAS,OAAO,UAAU,CAAC,0CAA0C,QAAQ,EAAE;AAChG,UAAQ,IAAI,KAAK,GAAG,OAAO,YAAY,GAAG,CAAC,0CAA0C,MAAM,EAAE;AAE7F,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,WAAW,QAAQ,IAAI;AACjC,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,QAAI,OAAO,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,EAAE,aAAa,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,EAAE,cAAc,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,MAAM,OAAO,EAAE,MAAM,QAAQ,CAAC,GAAG,SAAS,CAAC,CAAC,MAAM,OAAO,EAAE,aAAa,QAAQ,CAAC,IAAI,KAAK,SAAS,CAAC,CAAC;AACjO,QAAI,OAAO;AACT,UAAI,EAAE,IAAI;AACR,cAAM,MAAM,EAAE,GAAG,SAAS,IAAI,MAAM,MAAM,EAAE,GAAG,MAAM,QAAQ,CAAC;AAC9D,cAAM,MAAM,EAAE,GAAG,SAAS,IAAI,MAAM,MAAM,EAAE,GAAG,MAAM,QAAQ,CAAC;AAC9D,cAAM,MAAM,EAAE,cAAc,QAAQ;AACpC,gBAAQ,MAAM,EAAE,KAAK,EAAE,IAAI,GAAG;AAAA,MAChC,OAAO;AACL,gBAAQ,WAAM,OAAO,EAAE;AAAA,MACzB;AAAA,IACF;AACA,YAAQ,IAAI,IAAI;AAAA,EAClB;AAEA,MAAI,WAAW,QAAQ;AACrB,UAAM,IAAI,WAAW;AACrB,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,YAAQ;AAAA,MACN;AAAA,sBAAyB,EAAE,eAAe,KAAM,QAAQ,CAAC,CAAC,mBAAmB,EAAE,gBAAgB,KAAM,QAAQ,CAAC,CAAC,MAAM,IAAI,GAAG,EAAE,aAAa,QAAQ,CAAC,CAAC;AAAA,IACvJ;AAAA,EACF;AACA,MAAI,WAAW,MAAM;AACnB,UAAM,IAAI,WAAW;AACrB,UAAM,OAAO,EAAE,QAAQ,IAAI,MAAM;AACjC,YAAQ;AAAA,MACN,qBAAqB,EAAE,cAAc,QAAQ,CAAC,CAAC,kBAAkB,EAAE,eAAe,QAAQ,CAAC,CAAC,KAAK,IAAI,GAAG,EAAE,aAAa,QAAQ,CAAC,CAAC;AAAA,IACnI;AAAA,EACF;AAEA,QAAM,cAAc,MAAM,QAAQ,QAAQ,IAAI,SAAS,CAAC,IAAI;AAC5D,QAAM,SAAS,KAAK;AAAA,IAClB;AAAA,IACA,YAAY,MAAM,SAAS,WAAW,YAAY,SAAS,WAAW,aAAa;AAAA,EACrF;AACA,UAAQ;AAAA,IACN;AAAA,iBAAoB,WAAW,YAAY,MAAM,oBAAoB,WAAW,aAAa,MAAM,cAAc,MAAM;AAAA;AAAA,EACzH;AAEA,MAAI,oBAAoB,WAAW,YAAY,SAAS,GAAG;AAEzD,UAAM,kBAAkB,YAAY,KAAK,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,eAAe,IAAI;AACzF,QAAI,iBAAiB;AACnB,YAAM,2BAA2B,YAAY;AAAA,QAC3C,CAAC,MAAM,WAAW,QAAQ,CAAC,EAAE,gBAAgB,QAAQ,WAAW,QAAQ,CAAC,EAAE,QAAQ;AAAA,MACrF;AACA,UAAI,yBAA0B,SAAQ,KAAK,CAAC;AAAA,IAC9C,OAAO;AACL,cAAQ,KAAK,CAAC;AAAA,IAChB;AAAA,EACF;AACF;AAEA,eAAe,WAAW,MAAgB;AACxC,QAAM,EAAE,YAAY,WAAW,YAAY,MAAM,IAAI,cAAc,IAAI;AAEvE,MAAI,MAAM,SAAS,GAAG;AACpB,YAAQ,MAAM,sEAAsE;AACpF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,CAAC,aAAa,YAAY,IAAI;AACpC,QAAM,MAAM,KAAK,MAAM,MAAM,cAAc,aAAa,OAAO,CAAC;AAChE,QAAM,UAAwB,MAAM,QAAQ,GAAG,IAAI,MAAM,CAAC,GAAG;AAE7D,QAAM,UAAU,MAAM,WAAW,WAAW,UAAU;AACtD,MAAI;AACF,UAAM,MAAM,MAAM,aAAkB,aAAQ,YAAY,GAAG,YAAY,GAAG;AAC1E,UAAM,aAAyB,IAAI,SAAS,WAAW,IAAI,WAAW,IAAI,UAAU;AAEpF,QAAI,CAAC,WAAW,WAAW,WAAW,QAAQ,WAAW,GAAG;AAC1D,cAAQ,MAAM,UAAU,YAAY,0BAA0B;AAC9D,cAAQ,KAAK,CAAC;AAAA,IAChB;AAEA,UAAM,WAAyB,CAAC;AAChC,eAAW,cAAc,SAAS;AAChC,eAAS,KAAK,MAAM,QAAQ,YAAY,WAAW,SAAS,OAAO,CAAC;AAAA,IACtE;AAEA,eAAW,KAAK,UAAU;AACxB,cAAQ,IAAI,OAAO,YAAY,CAAC,IAAI,IAAI;AAAA,IAC1C;AAEA,QAAI,YAAY;AACd,YAAM,SAAS,MAAM,QAAQ,GAAG,IAAI,WAAW,SAAS,CAAC;AACzD,YAAM,YAAiB,aAAa,aAAQ,UAAU,CAAC;AACvD,YAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,YAAM,eAAe,YAAY,KAAK,UAAU,QAAQ,MAAM,CAAC,GAAG,OAAO;AACzE,cAAQ,IAAI,6BAA6B,UAAU,EAAE;AAAA,IACvD;AAAA,EACF,UAAE;AACA,UAAM,QAAQ,SAAS,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EACzC;AACF;AAEA,SAAS,iBAAiB,GAAqB;AAC7C,QAAM,WAAgB,aAAQ,CAAC;AAC/B,MAAI;AACF,UAAM,OAAO,SAAS,QAAQ;AAC9B,QAAI,KAAK,YAAY,GAAG;AACtB,YAAM,UAAU,YAAY,QAAQ;AACpC,aAAO,QACJ,OAAO,CAAC,MAAM,wBAAwB,KAAK,CAAC,CAAC,EAC7C,IAAI,CAAC,MAAW,UAAK,UAAU,CAAC,CAAC;AAAA,IACtC;AACA,WAAO,CAAC,QAAQ;AAAA,EAClB,QAAQ;AACN,WAAO,CAAC,QAAQ;AAAA,EAClB;AACF;AAGA,SAAS,gBAAgB,WAA4B;AACnD,MAAI,CAAC,MAAM,QAAQ,SAAS,KAAK,UAAU,WAAW,EAAG,QAAO;AAChE,SAAQ,UAAwB,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ,EAAE,KAAK,IAAI;AAC7F;AAEA,SAAS,YAAY,QAA4B;AAC/C,QAAM,QAAkB,CAAC;AACzB,QAAM,cAAc,OAAO,KAAK,OAAO,QAAQ,OAAO;AACtD,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,MAAM;AAChF,QAAM,WAAW;AAEjB,QAAM;AAAA,IACJ,SAAS,gBAAgB,OAAO,SAAS,SAAS,CAAC,MAAM,OAAO,OAAO,KAAK,OAAO,QAAQ,KAAK;AAAA,EAClG;AACA,QAAM;AAAA,IACJ,KAAK,SAAS,OAAO,UAAU,CAAC,KAAK,OAAO,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,QAAQ,CAAC;AAAA,EACpL;AAEA,QAAM,UAAU,aAAa,KAAK,WAAW,KAAK;AAClD,QAAM,KAAK,OAAO,SAAS,OAAO,OAAO,CAAC;AAE1C,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,OAAO,QAAQ,QAAQ,IAAI;AAErC,UAAM,kBAAkB,OAAO,MAAM,OAAO,CAAC,MAAM,CAAC,EAAE,SAAS,EAAE,OAAO,IAAI,KAAK,IAAI,EAAE;AACvF,QAAI,oBAAoB,GAAG;AACzB,YAAM;AAAA,QACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC,KAAK,KAAK,SAAS,QAAQ,CAAC;AAAA,MAC1K;AAAA,IACF,OAAO;AACL,YAAM;AAAA,QACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,EAAE,KAAK,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC;AAAA,MACvO;AAAA,IACF;AAAA,EACF;AAEA,MAAI,OAAO,QAAQ,QAAQ;AACzB,UAAM,IAAI,OAAO,QAAQ;AACzB,UAAM;AAAA,MACJ,KAAK,SAAS,OAAO,UAAU,CAAC,MAAM,EAAE,OAAO,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,MAAM,EAAE,MAAM,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC,MAAM,EAAE,MAAM,KAAM,QAAQ,CAAC,EAAE,SAAS,QAAQ,CAAC;AAAA,IACtL;AAAA,EACF;AAEA,QAAM,eAAe,OAAO,WAAW,KAAM,QAAQ,CAAC;AACtD,QAAM,UAAU,OAAO,YAAY,IAAI,IAAI,OAAO,UAAU,QAAQ,CAAC,CAAC,KAAK;AAC3E,QAAM,KAAK,EAAE;AACb,QAAM;AAAA,IACJ,eAAe,OAAO,QAAQ,QAAQ,IAAI,OAAO,QAAQ,KAAK,YAAY,OAAO,gBAAgB,WAAW;AAAA,EAC9G;AAEA,QAAM,kBAAkB,OAAO,MAAM,OAAO,CAAC,MAAM,EAAE,cAAc,MAAM;AACzE,MAAI,gBAAgB,SAAS,GAAG;AAC9B,UAAM,eAAe,CAAC,GAAG,IAAI,IAAI,gBAAgB,QAAQ,CAAC,MAAM,EAAE,YAAa,CAAC,CAAC;AACjF,UAAM,KAAK,EAAE;AACb,UAAM;AAAA,MACJ,oBAAoB,gBAAgB,MAAM,IAAI,OAAO,QAAQ,KAAK;AAAA,IACpE;AACA,eAAW,OAAO,aAAa,MAAM,GAAG,CAAC,GAAG;AAC1C,YAAM,KAAK,SAAS,GAAG,EAAE;AAAA,IAC3B;AACA,QAAI,aAAa,SAAS,GAAG;AAC3B,YAAM,KAAK,eAAe,aAAa,SAAS,CAAC,OAAO;AAAA,IAC1D;AAAA,EACF;AAEA,SAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,oBAAoB,SAAkC;AAC7D,QAAM,QAAkB,CAAC;AACzB,QAAM,cAAc,OAAO,KAAK,QAAQ,OAAO;AAC/C,QAAM,aAAa,KAAK,IAAI,GAAG,YAAY,IAAI,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,MAAM;AAChF,QAAM,WAAW;AAEjB,QAAM;AAAA,IACJ,SAAS,gBAAgB,QAAQ,SAAS,CAAC,MAAM,QAAQ,OAAO,WAAW,QAAQ,QAAQ;AAAA,EAC7F;AACA,QAAM;AAAA,IACJ,KAAK,SAAS,OAAO,UAAU,CAAC,KAAK,gBAAkB,SAAS,QAAQ,CAAC,KAAK,MAAM,SAAS,CAAC,CAAC,KAAK,MAAM,SAAS,CAAC,CAAC;AAAA,EACvH;AACA,QAAM,UAAU,aAAa,IAAI,WAAW,IAAI,IAAI,IAAI;AACxD,QAAM,KAAK,OAAO,SAAS,OAAO,OAAO,CAAC;AAE1C,aAAW,QAAQ,aAAa;AAC9B,UAAM,IAAI,QAAQ,QAAQ,IAAI;AAC9B,UAAM,UAAU,GAAG,EAAE,KAAK,QAAQ,CAAC,CAAC,SAAW,EAAE,IAAI,QAAQ,CAAC,CAAC;AAC/D,UAAM;AAAA,MACJ,KAAK,KAAK,OAAO,UAAU,CAAC,KAAK,QAAQ,SAAS,QAAQ,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,EAAE,IAAI,QAAQ,CAAC,EAAE,SAAS,CAAC,CAAC;AAAA,IAC/H;AAAA,EACF;AAEA,MAAI,QAAQ,QAAQ;AAClB,UAAM;AAAA,MACJ,KAAK,SAAS,OAAO,UAAU,CAAC,OAAO,QAAQ,OAAO,OAAO,KAAM,QAAQ,CAAC,IAAI,YAAc,QAAQ,OAAO,MAAM,KAAM,QAAQ,CAAC,IAAI,KAAK,SAAS,QAAQ,CAAC;AAAA,IAC/J;AAAA,EACF;AAEA,QAAM,UAAU,QAAQ,YAAY,IAAI,IAAI,QAAQ,UAAU,QAAQ,CAAC,CAAC,KAAK;AAC7E,QAAM,eAAe,QAAQ,gBAAgB,KAAM,QAAQ,CAAC;AAC5D,QAAM,KAAK,EAAE;AACb,QAAM,KAAK,iBAAiB,OAAO,sBAAsB,WAAW,GAAG;AAEvE,SAAO,MAAM,KAAK,IAAI;AACxB;AAIA,eAAe,yBAAyB,YAAyC;AAC/E,MAAI;AACF,UAAM,MAAM,MAAM,aAAa,YAAY,YAAY,GAAG;AAC1D,UAAM,UAAU,eAAe,GAAG;AAElC,QAAI,CAAC,WAAW,OAAO,QAAQ,YAAY,YAAY;AACrD,cAAQ,MAAM,mDAAmD;AACjE,UAAI,SAAS;AACX,cAAM,OAAO,OAAO,KAAK,OAAiB,EACvC,MAAM,GAAG,CAAC,EACV,KAAK,IAAI;AACZ,gBAAQ,MAAM,UAAU,OAAO,OAAO,GAAG,OAAO,iBAAiB,IAAI,OAAO,EAAE,EAAE;AAAA,MAClF;AACA,cAAQ;AAAA,QACN;AAAA;AAAA;AAAA,MACF;AACA,cAAQ,KAAK,CAAC;AAAA,IAChB;AAEA,WAAO;AAAA,EACT,SAAS,KAAK;AACZ,UAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC3D,QACE,8FAA8F;AAAA,MAC5F;AAAA,IACF,GACA;AACA,YAAM,MAAW,aAAQ,UAAU;AACnC,cAAQ,MAAM,wEAAwE;AACtF,UAAI,QAAQ,SAAS,QAAQ,QAAQ;AACnC,gBAAQ;AAAA,UACN;AAAA,QACF;AAAA,MACF,OAAO;AACL,gBAAQ,MAAM,mDAAmD;AAAA,MACnE;AACA,cAAQ,MAAM;AAAA,IAChB;AACA,YAAQ,MAAM,0BAA0B,GAAG;AAC3C,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAEA,eAAe,WAAW,WAAoB,YAA4C;AAExF,MAAI,cAAc,WAAW,SAAS,GAAG;AACvC,UAAM,mBAAmB,UAAU;AAAA,EACrC;AAGA,MAAI,WAAW;AACb,UAAM,aAAkB,aAAQ,QAAQ,IAAI,GAAG,SAAS;AACxD,UAAM,OAAO,SAAS,YAAY,EAAE,gBAAgB,MAAM,CAAC;AAC3D,QAAI,CAAC,MAAM,OAAO,GAAG;AACnB,cAAQ,MAAM,0BAA0B,UAAU,EAAE;AACpD,cAAQ,KAAK,CAAC;AAAA,IAChB;AACA,YAAQ,MAAM,kCAAkC,UAAU,EAAE;AAC5D,WAAO,yBAAyB,UAAU;AAAA,EAC5C;AAGA,QAAM,QAAQ,WAAW,QAAQ,IAAI,CAAC;AACtC,MAAI,OAAO;AACT,YAAQ,MAAM,oCAAoC,KAAK,EAAE;AACzD,WAAO,yBAAyB,KAAK;AAAA,EACvC;AAGA,QAAM,EAAE,WAAW,IAAI,MAAM,OAAO,aAAa;AACjD,SAAO,IAAI,WAAW;AACxB;AAIA,SAAS,cAAc,MAOrB;AACA,MAAI;AACJ,MAAI;AACJ,MAAI,aAAuB,CAAC;AAC5B,MAAI,OAAO;AACX,MAAI,gBAAgB;AACpB,QAAM,QAAkB,CAAC;AAEzB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAClB,QAAI,QAAQ,cAAc,QAAQ,cAAc,QAAQ,kBAAkB,QAAQ,UAAU;AAC1F,UAAI,IAAI,KAAK,KAAK,QAAQ;AACxB,gBAAQ,MAAM,UAAU,GAAG,mBAAmB;AAC9C,gBAAQ,KAAK,CAAC;AAAA,MAChB;AACA,YAAM,QAAQ,KAAK,EAAE,CAAC;AACtB,UAAI,QAAQ,WAAY,cAAa;AAAA,eAC5B,QAAQ,WAAY,aAAY;AAAA,eAChC,QAAQ,SAAU,QAAO,KAAK,IAAI,GAAG,SAAS,OAAO,EAAE,KAAK,CAAC;AAAA;AAEpE,qBAAa,MACV,MAAM,GAAG,EACT,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,OAAO;AAAA,IACrB,WAAW,QAAQ,oBAAoB;AAErC,sBAAgB;AAAA,IAClB,WAAW,IAAI,WAAW,IAAI,GAAG;AAC/B,UAAI,CAAC,YAAY,IAAI,GAAG,GAAG;AACzB,gBAAQ,MAAM,iBAAiB,GAAG,EAAE;AACpC,gBAAQ,KAAK,CAAC;AAAA,MAChB;AAAA,IACF,OAAO;AACL,YAAM,KAAK,GAAG;AAAA,IAChB;AAAA,EACF;AAEA,SAAO,EAAE,YAAY,WAAW,YAAY,MAAM,eAAe,MAAM;AACzE;AAIA,eAAe,eAAe,MAAgB;AAC5C,QAAM,EAAE,YAAY,WAAW,YAAY,MAAM,eAAe,MAAM,IAAI,cAAc,IAAI;AAE5F,MAAI,MAAM,WAAW,GAAG;AACtB,YAAQ,MAAM,mCAAmC;AACjD,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,YAAsB,CAAC;AAC7B,aAAW,KAAK,OAAO;AACrB,cAAU,KAAK,GAAG,iBAAiB,CAAC,CAAC;AAAA,EACvC;AAEA,MAAI,UAAU,WAAW,GAAG;AAC1B,YAAQ,MAAM,4BAA4B;AAC1C,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,QAAM,UAAU,MAAM,WAAW,WAAW,UAAU;AACtD,QAAM,UAAwB,CAAC;AAE/B,MAAI;AACF,eAAW,YAAY,WAAW;AAChC,UAAI;AACF,cAAM,MAAM,MAAM,aAAkB,aAAQ,QAAQ,GAAG,YAAY,GAAG;AACtE,cAAM,aAAyB,IAAI,SAAS,WAAW,IAAI,WAAW,IAAI,UAAU;AAEpF,YAAI,CAAC,WAAW,YAAY,CAAC,WAAW,WAAW,CAAC,WAAW,SAAS;AACtE,kBAAQ;AAAA,YACN,UAAU,QAAQ;AAAA,UACpB;AACA;AAAA,QACF;AAGA,YAAI;AAKJ,YAAI,IAAI,iBAAiB;AAEvB,4BAAkB,OAAO,OAAO,OAAO;AACrC,kBAAM;AAAA,cACJ;AAAA,cACA,MAAM;AAAA,cACN;AAAA,YACF,IAAI,MAAM,QAAQ,eAAe,YAAY;AAC3C,qBAAO,IAAI,gBAAgB,OAAO,EAAE;AAAA,YACtC,CAAC;AACD,mBAAO;AAAA,cACL,QAAQ,OAAO;AAAA,cACf,MAAM,OAAO,QAAQ;AAAA,cACrB,UAAU,OAAO,YAAY;AAAA,YAC/B;AAAA,UACF;AAAA,QACF,WAAW,QAAQ,YAAY,WAAW,QAAQ,GAAG;AAEnD,4BAAkB,OAAO,UAAU;AACjC,kBAAM,EAAE,QAAQ,MAAM,SAAS,IAAI,MAAM,QAAQ,eAAe,YAAY;AAC1E,qBAAO,QAAQ,QAAQ,WAAW,UAAU,KAAK;AAAA,YACnD,CAAC;AACD,mBAAO,EAAE,QAAQ,QAAQ,MAAM,SAAS;AAAA,UAC1C;AAAA,QACF,OAAO;AACL,kBAAQ;AAAA,YACN,uBAAuB,QAAQ;AAAA,UACjC;AACA,4BAAkB,OAAO,WAAW,EAAE,QAAQ,MAAM;AAAA,QACtD;AASA,cAAM,aAAa,gBAAgB,EAAE,eAAe,KAAK,IAAI;AAE7D,YAAI,OAAO,GAAG;AAEZ,gBAAM,EAAE,WAAW,IAAI,MAAM,OAAO,QAAa;AACjD,gBAAM,aAAa,WAAW;AAC9B,gBAAM,aAA2B,CAAC;AAElC,mBAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,oBAAQ,MAAM,kBAAkB,IAAI,CAAC,IAAI,IAAI,KAAK;AAClD,kBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,SAAS,UAAU;AAC7E,mBAAO,SAAS,aAAa;AAC7B,mBAAO,SAAS,WAAW;AAC3B,uBAAW,KAAK,MAAM;AACtB,oBAAQ,KAAK,MAAM;AAAA,UACrB;AAEA,gBAAM,UAAU,cAAc,UAAU;AACxC,kBAAQ,IAAI,OAAO,oBAAoB,OAAO,IAAI,IAAI;AAAA,QACxD,OAAO;AACL,gBAAM,SAAS,MAAM,QAAQ,YAAY,iBAAiB,SAAS,UAAU;AAC7E,kBAAQ,KAAK,MAAM;AAEnB,kBAAQ,IAAI,OAAO,YAAY,MAAM,IAAI,IAAI;AAAA,QAC/C;AAAA,MACF,SAAS,KAAK;AACZ,gBAAQ;AAAA,UACN,sBAAsB,QAAQ,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,QACrF;AAAA,MACF;AAAA,IACF;AAEA,QAAI,cAAc,QAAQ,SAAS,GAAG;AACpC,YAAM,SAAS,QAAQ,WAAW,IAAI,QAAQ,CAAC,IAAI;AACnD,YAAM,YAAiB,aAAa,aAAQ,UAAU,CAAC;AACvD,YAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,YAAM,eAAe,YAAY,KAAK,UAAU,QAAQ,MAAM,CAAC,GAAG,OAAO;AACzE,cAAQ,IAAI,oBAAoB,UAAU,EAAE;AAAA,IAC9C;AAAA,EACF,UAAE;AACA,UAAM,QAAQ,SAAS,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EACzC;AAGA,MAAI,QAAQ,WAAW,GAAG;AACxB,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAEA,KAAK,EAAE,MAAM,CAAC,QAAQ;AACpB,UAAQ,MAAM,GAAG;AACjB,UAAQ,KAAK,CAAC;AAChB,CAAC;","names":[]}
package/dist/index.cjs CHANGED
@@ -184,7 +184,36 @@ function parseCost(cost) {
184
184
  if (!match) throw new Error(`Invalid cost format: "${cost}"`);
185
185
  return parseFloat(match[1]);
186
186
  }
187
- async function runEval(config, executeWorkflow, runtime) {
187
+ function extractUserCost(result, label = "executeWorkflow") {
188
+ if (result === null || typeof result !== "object") return void 0;
189
+ const raw = result.cost;
190
+ if (raw === void 0) return void 0;
191
+ if (typeof raw === "number" && Number.isFinite(raw) && raw >= 0) return raw;
192
+ console.warn(
193
+ `[axl-eval] Ignoring invalid \`cost\` from ${label} return: expected non-negative finite number, got ${typeof raw === "number" ? String(raw) : typeof raw}. Falling back to tracked cost (or undefined).`
194
+ );
195
+ return void 0;
196
+ }
197
+ function extractUserMetadata(result, label = "executeWorkflow") {
198
+ if (result === null || typeof result !== "object") return void 0;
199
+ const meta = result.metadata;
200
+ if (meta === void 0) return void 0;
201
+ if (meta === null || typeof meta !== "object" || Array.isArray(meta)) {
202
+ console.warn(
203
+ `[axl-eval] Ignoring invalid \`metadata\` from ${label} return: expected plain object, got ${meta === null ? "null" : Array.isArray(meta) ? "array" : typeof meta}.`
204
+ );
205
+ return void 0;
206
+ }
207
+ const proto = Object.getPrototypeOf(meta);
208
+ if (proto !== Object.prototype && proto !== null) {
209
+ console.warn(
210
+ `[axl-eval] Ignoring invalid \`metadata\` from ${label} return: expected plain object, got ${proto?.constructor?.name ?? "exotic object"}.`
211
+ );
212
+ return void 0;
213
+ }
214
+ return meta;
215
+ }
216
+ async function runEval(config, executeWorkflow, runtime, options) {
188
217
  const startTime = Date.now();
189
218
  const id = (0, import_node_crypto.randomUUID)();
190
219
  const items = await config.dataset.getItems();
@@ -204,6 +233,17 @@ async function runEval(config, executeWorkflow, runtime) {
204
233
  let totalCost = 0;
205
234
  let budgetExceeded = false;
206
235
  async function processItem(item, itemIndex) {
236
+ if (options?.signal?.aborted) {
237
+ evalItems[itemIndex] = {
238
+ input: item.input,
239
+ annotations: item.annotations,
240
+ output: null,
241
+ error: "Cancelled",
242
+ scores: {}
243
+ };
244
+ options?.onProgress?.({ type: "item_done", itemIndex, totalItems: items.length });
245
+ return;
246
+ }
207
247
  if (budgetExceeded) {
208
248
  evalItems[itemIndex] = {
209
249
  input: item.input,
@@ -212,6 +252,7 @@ async function runEval(config, executeWorkflow, runtime) {
212
252
  error: "Budget exceeded",
213
253
  scores: {}
214
254
  };
255
+ options?.onProgress?.({ type: "item_done", itemIndex, totalItems: items.length });
215
256
  return;
216
257
  }
217
258
  const evalItem = {
@@ -222,18 +263,43 @@ async function runEval(config, executeWorkflow, runtime) {
222
263
  };
223
264
  const itemStart = Date.now();
224
265
  try {
225
- const result = await executeWorkflow(item.input, runtime);
226
- evalItem.duration = Date.now() - itemStart;
227
- evalItem.output = result.output;
228
- evalItem.cost = result.cost;
229
- if (result.metadata) evalItem.metadata = result.metadata;
230
- if (result.cost != null) {
231
- totalCost += result.cost;
266
+ if (options?.captureTraces) {
267
+ const tracked = await runtime.trackExecution(
268
+ async () => executeWorkflow(item.input, runtime),
269
+ { captureTraces: true }
270
+ );
271
+ evalItem.duration = Date.now() - itemStart;
272
+ evalItem.output = tracked.result.output;
273
+ evalItem.cost = extractUserCost(tracked.result) ?? tracked.cost;
274
+ evalItem.metadata = extractUserMetadata(tracked.result) ?? tracked.metadata;
275
+ if (tracked.traces && tracked.traces.length > 0) {
276
+ evalItem.traces = tracked.traces;
277
+ }
278
+ if (evalItem.cost != null) {
279
+ totalCost += evalItem.cost;
280
+ }
281
+ } else {
282
+ const result = await executeWorkflow(item.input, runtime);
283
+ evalItem.duration = Date.now() - itemStart;
284
+ evalItem.output = result.output;
285
+ evalItem.cost = extractUserCost(result);
286
+ const meta = extractUserMetadata(result);
287
+ if (meta) evalItem.metadata = meta;
288
+ if (evalItem.cost != null) {
289
+ totalCost += evalItem.cost;
290
+ }
232
291
  }
233
292
  } catch (err) {
234
293
  evalItem.duration = Date.now() - itemStart;
235
294
  evalItem.error = err instanceof Error ? err.message : String(err);
295
+ if (options?.captureTraces) {
296
+ const captured = err.axlCapturedTraces;
297
+ if (Array.isArray(captured) && captured.length > 0) {
298
+ evalItem.traces = captured;
299
+ }
300
+ }
236
301
  evalItems[itemIndex] = evalItem;
302
+ options?.onProgress?.({ type: "item_done", itemIndex, totalItems: items.length });
237
303
  return;
238
304
  }
239
305
  if (budgetLimit != null && totalCost > budgetLimit) {
@@ -242,6 +308,7 @@ async function runEval(config, executeWorkflow, runtime) {
242
308
  evalItem.scoreDetails = {};
243
309
  let itemScorerCost = 0;
244
310
  for (const scorer2 of config.scorers) {
311
+ if (options?.signal?.aborted) break;
245
312
  const scorerStart = Date.now();
246
313
  try {
247
314
  const raw = await scorer2.score(
@@ -297,6 +364,7 @@ async function runEval(config, executeWorkflow, runtime) {
297
364
  }
298
365
  evalItem.scorerCost = itemScorerCost > 0 ? itemScorerCost : void 0;
299
366
  evalItems[itemIndex] = evalItem;
367
+ options?.onProgress?.({ type: "item_done", itemIndex, totalItems: items.length });
300
368
  }
301
369
  let index = 0;
302
370
  async function runNext() {
@@ -343,14 +411,42 @@ async function runEval(config, executeWorkflow, runtime) {
343
411
  modelsMeta.models = sorted.map(([m]) => m);
344
412
  modelsMeta.modelCounts = Object.fromEntries(sorted);
345
413
  }
414
+ const totalWorkflowCalls = /* @__PURE__ */ new Map();
415
+ for (const item of evalItems) {
416
+ const itemCounts = item.metadata?.workflowCallCounts;
417
+ if (itemCounts && typeof itemCounts === "object") {
418
+ for (const [w, count] of Object.entries(itemCounts)) {
419
+ if (typeof count === "number")
420
+ totalWorkflowCalls.set(w, (totalWorkflowCalls.get(w) ?? 0) + count);
421
+ }
422
+ } else {
423
+ const itemWorkflows = item.metadata?.workflows;
424
+ if (Array.isArray(itemWorkflows)) {
425
+ for (const w of itemWorkflows) {
426
+ if (typeof w === "string")
427
+ totalWorkflowCalls.set(w, (totalWorkflowCalls.get(w) ?? 0) + 1);
428
+ }
429
+ }
430
+ }
431
+ }
432
+ const workflowsMeta = {};
433
+ if (totalWorkflowCalls.size > 0) {
434
+ const sorted = [...totalWorkflowCalls.entries()].sort((a, b) => b[1] - a[1]);
435
+ workflowsMeta.workflows = sorted.map(([w]) => w);
436
+ workflowsMeta.workflowCounts = Object.fromEntries(sorted);
437
+ } else if (config.workflow) {
438
+ workflowsMeta.workflows = [config.workflow];
439
+ workflowsMeta.workflowCounts = { [config.workflow]: items.length };
440
+ }
441
+ options?.onProgress?.({ type: "run_done", totalItems: items.length, failures });
346
442
  return {
347
443
  id,
348
- workflow: config.workflow,
349
444
  dataset: config.dataset.name,
350
445
  metadata: {
351
446
  ...config.metadata,
352
447
  scorerTypes,
353
- ...modelsMeta
448
+ ...modelsMeta,
449
+ ...workflowsMeta
354
450
  },
355
451
  timestamp: (/* @__PURE__ */ new Date()).toISOString(),
356
452
  totalCost,
@@ -649,6 +745,7 @@ async function rescore(result, scorers, runtime, options) {
649
745
  annotations: original.annotations,
650
746
  output: original.output,
651
747
  metadata: original.metadata,
748
+ traces: original.traces,
652
749
  scores: {},
653
750
  scoreDetails: {}
654
751
  };
@@ -730,11 +827,23 @@ async function rescore(result, scorers, runtime, options) {
730
827
  }
731
828
  return {
732
829
  id: (0, import_node_crypto2.randomUUID)(),
733
- workflow: result.workflow,
734
830
  dataset: result.dataset,
735
831
  metadata: (() => {
736
832
  const { runGroupId: _, runIndex: __, ...rest } = result.metadata;
737
- return { ...rest, rescored: true, originalId: result.id, scorerTypes };
833
+ const merged = {
834
+ ...rest,
835
+ rescored: true,
836
+ originalId: result.id,
837
+ scorerTypes
838
+ };
839
+ if (!Array.isArray(merged.workflows)) {
840
+ const legacyWorkflow = result.workflow;
841
+ if (typeof legacyWorkflow === "string" && legacyWorkflow) {
842
+ merged.workflows = [legacyWorkflow];
843
+ merged.workflowCounts = { [legacyWorkflow]: result.items.length };
844
+ }
845
+ }
846
+ return merged;
738
847
  })(),
739
848
  timestamp: (/* @__PURE__ */ new Date()).toISOString(),
740
849
  totalCost,
@@ -754,7 +863,20 @@ function std(values, mean) {
754
863
  function aggregateRuns(runs) {
755
864
  if (runs.length === 0) throw new Error("Cannot aggregate zero runs");
756
865
  const runGroupId = runs[0].metadata.runGroupId ?? (0, import_node_crypto3.randomUUID)();
757
- const { workflow, dataset: dataset2 } = runs[0];
866
+ const { dataset: dataset2 } = runs[0];
867
+ const seenWorkflows = /* @__PURE__ */ new Set();
868
+ const workflows = [];
869
+ for (const run of runs) {
870
+ const list = run.metadata.workflows;
871
+ if (Array.isArray(list)) {
872
+ for (const w of list) {
873
+ if (typeof w === "string" && !seenWorkflows.has(w)) {
874
+ seenWorkflows.add(w);
875
+ workflows.push(w);
876
+ }
877
+ }
878
+ }
879
+ }
758
880
  const totalCost = runs.reduce((sum, r) => sum + r.totalCost, 0);
759
881
  const totalDuration = runs.reduce((sum, r) => sum + r.duration, 0);
760
882
  const scorerNames = Object.keys(runs[0].summary.scorers);
@@ -783,7 +905,7 @@ function aggregateRuns(runs) {
783
905
  return {
784
906
  runGroupId,
785
907
  runCount: runs.length,
786
- workflow,
908
+ workflows,
787
909
  dataset: dataset2,
788
910
  totalCost: round(totalCost),
789
911
  totalDuration: round(totalDuration),