@exercode/problem-utils 2.4.3 → 3.0.0

package/dist/presets/evaluation.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"evaluation.cjs","names":["parseArgs","resolveCwds","passesIsolationCheckInDebugMode","path","judgeByStaticAnalysis","readProblemMarkdownFrontMatter","isSafeSubmissionOutputPath","fs","DecisionCode","parseCsvRecords","CsvParseError","parseCsv"],"sources":["../../src/presets/evaluation.ts"],"sourcesContent":["import fs from 'node:fs/promises';\nimport path from 'node:path';\n\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { CsvParseError, type CsvParseErrorReason, parseCsv, parseCsvRecords } from '../helpers/parseCsv.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { isSafeSubmissionOutputPath } from '../helpers/readOutputFiles.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst DEFAULT_SUBMISSION_FILE_PATH = 'submission.csv';\nconst TEST_CASE_ID = 'evaluation';\nconst MAX_LISTED_IDS = 10;\n// Learner-controlled cells are echoed into the feedback, so clip them to keep the result line small.\nconst MAX_ECHOED_CELL_LENGTH = 50;\nconst CSV_PARSE_ERROR_MESSAGES: Record<CsvParseErrorReason, string> = {\n  unterminated_quote: '引用符（`\"`）が閉じられていません',\n  unexpected_quote: '引用符（`\"`）は値全体を囲む位置にしか書けません',\n  text_after_quoted_field: '閉じ引用符の後に区切り文字以外の文字があります',\n};\n\nexport interface EvaluationMetric {\n  /** Label shown with the score, e.g. `RMSLE`. */\n  label: string;\n  isHigherBetter: boolean;\n  /**\n   * Returns an error message when a predicted value cannot be scored (e.g. it is not a number).\n   * The message is appended to the row's line number and the offending value, so phrase it as a\n   * predicate such as `数値ではありません`. It also validates the ground truth when it is loaded.\n   */\n  validatePrediction?: (prediction: string) => string | undefined;\n  /** Computes the score from predictions and answers aligned by id. Both are raw CSV cell strings. */\n  compute: (predictions: readonly string[], answers: readonly string[]) => number;\n}\n\nexport interface EvaluationJudgePresetOptions {\n  /** CSV file holding the hidden ground truth, relative to the problem directory. */\n  answerFilePath: string;\n  /** CSV file the submission must contain. Defaults to `submission.csv`. */\n  submissionFilePath?: string;\n  /** Column that identifies each row in both CSV files. */\n  idColumn: string;\n  /** Column holding the answer in the ground truth and the prediction in the submission. */\n  targetColumn: string;\n  metric: EvaluationMetric;\n  /**\n   * Score required for acceptance: the submission is accepted when its score is at least this value\n   * for a higher-is-better metric and at most this value otherwise. Every valid submission is\n   * accepted when omitted.\n   */\n  acceptableScore?: number;\n}\n\nexport const evaluationMetrics = {\n  rmsle: {\n    label: 'RMSLE',\n    isHigherBetter: false,\n    validatePrediction: (prediction: string) =>\n      validateFiniteNumber(prediction) ?? (Number(prediction) < 0 ? '負の値です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(\n        mean(\n          predictions.map(\n            (prediction, index) => (Math.log1p(Number(prediction)) - Math.log1p(Number(answers[index]))) ** 2\n          )\n        )\n      ),\n  },\n  rmse: {\n    label: 'RMSE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(mean(predictions.map((prediction, index) => (Number(prediction) - Number(answers[index])) ** 2))),\n  },\n  mae: {\n    label: 'MAE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => Math.abs(Number(prediction) - Number(answers[index])))),\n  },\n  accuracy: {\n    label: 'Accuracy',\n    isHigherBetter: true,\n    validatePrediction: (prediction: string) => (prediction.trim() === '' ? '空です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => (prediction.trim() === answers[index]?.trim() ? 1 : 0))),\n  },\n} as const satisfies Record<string, EvaluationMetric>;\n\n/**\n * A preset for Kaggle-style model evaluation problems: the submission is a CSV of predictions for a\n * hidden test set, and the judge reports the metric value as `score` instead of comparing outputs.\n * The problem directory keeps the ground truth CSV, so it is never shipped to learners.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { evaluationJudgePreset, evaluationMetrics } from '@exercode/problem-utils/presets/evaluation';\n *\n * await evaluationJudgePreset(import.meta.dirname, {\n *   answerFilePath: 'evaluation/answer.csv',\n *   idColumn: 'Id',\n *   targetColumn: 'TradePrice',\n *   metric: evaluationMetrics.rmsle,\n *   acceptableScore: 0.5,\n * });\n * ```\n *\n * Run `bun judge.ts model_answers/default` to judge one submission, or `bun judge.ts` to judge every\n * `<problemDir>/model_answers/*` and `<problemDir>/model_answers.fails/*` directory for debugging.\n */\nexport async function evaluationJudgePreset(problemDir: string, options: EvaluationJudgePresetOptions): Promise<void> {\n  const args = parseArgs(process.argv);\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n\n  if (isDebugMode && !(await passesIsolationCheckInDebugMode(problemDir, cwds, args.params))) {\n    process.exitCode = 1;\n    return;\n  }\n\n  const answer = await readAnswer(path.join(problemDir, options.answerFilePath), options);\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await judgeSubmission(problemDir, resolvedCwd.cwd, answer, options);\n    printTestCaseResult({ testCaseId: TEST_CASE_ID, ...result });\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function judgeSubmission(\n  problemDir: string,\n  cwd: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Promise<Omit<TestCaseResult, 'testCaseId'>> {\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, await readProblemMarkdownFrontMatter(problemDir));\n  if (staticAnalysisResult) return staticAnalysisResult;\n\n  const submissionFilePath = options.submissionFilePath ?? DEFAULT_SUBMISSION_FILE_PATH;\n  let submissionText: string;\n  try {\n    const resolvedPath = path.join(cwd, submissionFilePath);\n    // The harness reads the file as the trusted user, so refuse symlinks escaping the submission.\n    if (!(await isSafeSubmissionOutputPath(cwd, resolvedPath))) throw new Error('unsafe path');\n    submissionText = await fs.readFile(resolvedPath, 'utf8');\n  } catch {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      feedbackMarkdown: `\\`${submissionFilePath}\\`が見つかりません。予測結果のCSVファイルを\\`${submissionFilePath}\\`という名前で提出してください。`,\n    };\n  }\n\n  const predictions = readPredictions(submissionText, submissionFilePath, answer, options);\n  if (typeof predictions === 'string') {\n    return { decisionCode: DecisionCode.WRONG_ANSWER, feedbackMarkdown: predictions };\n  }\n\n  const answerValues = [...answer.values()];\n  const score = options.metric.compute(\n    [...answer.keys()].map((id) => predictions.get(id) as string),\n    answerValues\n  );\n  if (!Number.isFinite(score)) {\n    return {\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      feedbackMarkdown: `予測値が極端なため、${options.metric.label}を計算できませんでした。予測値を見直してください。`,\n    };\n  }\n  const isAccepted =\n    options.acceptableScore === undefined ||\n    (options.metric.isHigherBetter ? score >= options.acceptableScore : score <= options.acceptableScore);\n  return {\n    decisionCode: isAccepted ? DecisionCode.ACCEPTED : DecisionCode.WRONG_ANSWER,\n    score,\n    scoreLabel: options.metric.label,\n    feedbackMarkdown: buildScoreMarkdown(score, answerValues.length, isAccepted, options),\n  };\n}\n\nfunction buildScoreMarkdown(\n  score: number,\n  rowCount: number,\n  isAccepted: boolean,\n  options: EvaluationJudgePresetOptions\n): string {\n  const { label, isHigherBetter } = options.metric;\n  const precision = findDistinguishingPrecision(score, options.acceptableScore);\n  const acceptanceLine =\n    options.acceptableScore === undefined\n      ? ''\n      : `\\n合格基準: ${label} ${isHigherBetter ? '≧' : '≦'} ${formatScore(options.acceptableScore, precision)}（${isAccepted ? '達成' : '未達成'}）`;\n  return `| 指標 | スコア |\n| ---- | ------ |\n| ${label} | ${formatScore(score, precision)} |\n\n評価件数: ${rowCount.toLocaleString('en-US')}件${acceptanceLine}\n`;\n}\n\n/**\n * Finds the number of significant digits at which the score and the threshold render differently\n * (or stop changing), so the feedback never shows a pair that contradicts the decision.\n */\nfunction findDistinguishingPrecision(score: number, threshold: number | undefined): number {\n  const minimumPrecision = 6;\n  if (threshold === undefined || score === threshold) return minimumPrecision;\n  for (let precision = minimumPrecision; precision < 17; precision++) {\n    if (formatScore(score, precision) !== formatScore(threshold, precision)) return precision;\n  }\n  return 17;\n}\n\nfunction formatScore(score: number, precision: number): string {\n  return Number(score.toPrecision(precision)).toString();\n}\n\nfunction readPredictions(\n  submissionText: string,\n  submissionFilePath: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Map<string, string> | string {\n  let records: ReturnType<typeof parseCsvRecords>;\n  try {\n    records = parseCsvRecords(submissionText);\n  } catch (error) {\n    const detail =\n      error instanceof CsvParseError ? `${error.lineNumber}行目付近: ${CSV_PARSE_ERROR_MESSAGES[error.reason]}` : '';\n    return `\\`${submissionFilePath}\\`をCSVとして読み込めませんでした。${detail}`;\n  }\n  const columns = findColumns(records[0]?.cells ?? [], options);\n  if (!columns) {\n    return `\\`${submissionFilePath}\\`の1行目には\\`${options.idColumn}\\`列と\\`${options.targetColumn}\\`列のヘッダーが必要です。`;\n  }\n\n  const predictions = new Map<string, string>();\n  const problems: string[] = [];\n  let problemCount = 0;\n  const reportProblem = (problem: string): void => {\n    problemCount++;\n    if (problems.length < MAX_LISTED_IDS) problems.push(problem);\n  };\n  for (const { cells: row, lineNumber } of records.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const prediction = row[columns.targetIndex] ?? '';\n    if (!answer.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行は評価対象ではありません`);\n    } else if (predictions.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行が重複しています`);\n    } else {\n      const validationError = options.metric.validatePrediction?.(prediction);\n      if (validationError) {\n        reportProblem(`${lineNumber}行目: \\`${options.targetColumn}\\`の値\\`${clip(prediction)}\\`は${validationError}`);\n      }\n      predictions.set(id, prediction);\n    }\n  }\n  const missingIds = [...answer.keys()].filter((id) => !predictions.has(id));\n  if (missingIds.length > 0) {\n    // Missing rows are the most actionable problem, so they must survive the listing cap.\n    problemCount++;\n    problems.unshift(\n      `\\`${options.idColumn}\\`が${listIds(missingIds)}の行がありません（不足 ${missingIds.length.toLocaleString('en-US')}件）`\n    );\n  }\n  if (problemCount > 0) {\n    return `\\`${submissionFilePath}\\`の内容に問題があります。\n\n${problems.map((problem) => `- ${problem}`).join('\\n')}${problemCount > problems.length ? `\\n- ほか${(problemCount - problems.length).toLocaleString('en-US')}件` : ''}\n`;\n  }\n  return predictions;\n}\n\nasync function readAnswer(answerFilePath: string, options: EvaluationJudgePresetOptions): Promise<Map<string, string>> {\n  const rows = parseCsv(await fs.readFile(answerFilePath, 'utf8'));\n  const columns = findColumns(rows[0] ?? [], options);\n  if (!columns) {\n    throw new Error(`answer file must have ${options.idColumn} and ${options.targetColumn} columns: ${answerFilePath}`);\n  }\n  const answer = new Map<string, string>();\n  for (const row of rows.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const value = row[columns.targetIndex] ?? '';\n    if (!id) throw new Error(`empty ${options.idColumn} in answer file: ${answerFilePath}`);\n    if (answer.has(id)) throw new Error(`duplicate ${options.idColumn} in answer file: ${id}`);\n    const validationError = options.metric.validatePrediction?.(value);\n    if (validationError)\n      throw new Error(`invalid ${options.targetColumn} in answer file for ${id}: ${validationError}`);\n    answer.set(id, value);\n  }\n  if (answer.size === 0) throw new Error(`answer file has no rows: ${answerFilePath}`);\n  return answer;\n}\n\nfunction findColumns(\n  header: readonly string[],\n  options: EvaluationJudgePresetOptions\n): { idIndex: number; targetIndex: number } | undefined {\n  const names = header.map((name) => name.trim());\n  const idIndex = names.indexOf(options.idColumn);\n  const targetIndex = names.indexOf(options.targetColumn);\n  return idIndex === -1 || targetIndex === -1 ? undefined : { idIndex, targetIndex };\n}\n\nfunction isBlankRow(row: readonly string[]): boolean {\n  return row.every((cell) => cell.trim() === '');\n}\n\nfunction listIds(ids: readonly string[]): string {\n  const listed = ids.slice(0, MAX_LISTED_IDS).map((id) => `\\`${clip(id)}\\``);\n  return ids.length > MAX_LISTED_IDS ? `${listed.join(', ')}, ...` : listed.join(', ');\n}\n\nfunction clip(value: string): string {\n  // Control characters and backticks would break the markdown list item and inline code the value is echoed in.\n  const printable = value.replaceAll(/[\\p{Cc}]/gu, ' ').replaceAll('`', \"'\");\n  if (printable.trim() === '') return '（空）';\n  return printable.length > MAX_ECHOED_CELL_LENGTH ? `${printable.slice(0, MAX_ECHOED_CELL_LENGTH)}…` : printable;\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: Pick<TestCaseResult, 'decisionCode'>): boolean {\n  return (result.decisionCode === DecisionCode.ACCEPTED) === (resolvedCwd.expectedResult === 'accepted');\n}\n\nfunction validateFiniteNumber(prediction: string): string | undefined {\n  return Number.isFinite(Number(prediction)) && prediction.trim() !== '' ? undefined : '数値ではありません';\n}\n\nfunction mean(values: readonly number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length;\n}\n"],"mappings":"goBAmBA,MAKM,EAAgE,CACpE,mBAAoB,qBACpB,iBAAkB,4BAClB,wBAAyB,yBAC3B,EAkCa,EAAoB,CAC/B,MAAO,CACL,MAAO,QACP,eAAgB,GAChB,mBAAqB,GACnB,EAAqB,CAAU,IAAM,OAAO,CAAU,EAAI,EAAI,QAAU,IAAA,IAC1E,SAAU,EAAgC,IACxC,KAAK,KACH,EACE,EAAY,KACT,EAAY,KAAW,KAAK,MAAM,OAAO,CAAU,CAAC,EAAI,KAAK,MAAM,OAAO,EAAQ,EAAM,CAAC,IAAM,CAClG,CACF,CACF,CACJ,EACA,KAAM,CACJ,MAAO,OACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,KAAK,KAAK,EAAK,EAAY,KAAK,EAAY,KAAW,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,IAAM,CAAC,CAAC,CAAC,CAC9G,EACA,IAAK,CACH,MAAO,MACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAU,KAAK,IAAI,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,CAAC,CAAC,CAAC,CACtG,EACA,SAAU,CACR,MAAO,WACP,eAAgB,GAChB,mBAAqB,GAAwB,EAAW,KAAK,IAAM,GAAK,MAAQ,IAAA,GAChF,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAW,IAAW,KAAK,IAAM,EAAQ,EAAM,EAAE,KAAK,EAAU,CAAC,CACvG,CACF,EAwBA,eAAsB,EAAsB,EAAoB,EAAsD,CACpH,IAAM,EAAOA,EAAAA,UAAU,QAAQ,IAAI,EAC7B,CAAE,OAAM,eAAgB,MAAMC,EAAAA,YAAY,EAAY,EAAK,GAAG,EAEpE,GAAI,GAAe,CAAE,MAAMC,EAAAA,gCAAgC,EAAY,EAAM,EAAK,MAAM,EAAI,CAC1F,QAAQ,SAAW,EACnB,MACF,CAEA,IAAM,EAAS,MAAM,EAAWC,EAAAA,QAAK,KAAK,EAAY,EAAQ,cAAc,EAAG,CAAO,EACtF,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAA,oBAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAAgB,EAAY,EAAY,IAAK,EAAQ,CAAO,EACjF,EAAA,oBAAoB,CAAE,WAAY,aAAc,GAAG,CAAO,CAAC,EACvD,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAA,mCAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EACb,EACA,EACA,EACA,EAC6C,CAC7C,IAAM,EAAuB,MAAMC,EAAAA,sBAAsB,EAAK,MAAMC,EAAAA,+BAA+B,CAAU,CAAC,EAC9G,GAAI,EAAsB,OAAO,EAEjC,IAAM,EAAqB,EAAQ,oBAAsB,iBACrD,EACJ,GAAI,CACF,IAAM,EAAeF,EAAAA,QAAK,KAAK,EAAK,CAAkB,EAEtD,GAAI,CAAE,MAAMG,EAAAA,2BAA2B,EAAK,CAAY,EAAI,MAAU,MAAM,aAAa,EACzF,EAAiB,MAAMC,EAAAA,QAAG,SAAS,EAAc,MAAM,CACzD,MAAQ,CACN,MAAO,CACL,aAAcC,EAAAA,aAAa,uCAC3B,iBAAkB,KAAK,EAAmB,4BAA4B,EAAmB,kBAC3F,CACF,CAEA,IAAM,EAAc,EAAgB,EAAgB,EAAoB,EAAQ,CAAO,EACvF,GAAI,OAAO,GAAgB,SACzB,MAAO,CAAE,aAAcA,EAAAA,aAAa,aAAc,iBAAkB,CAAY,EAGlF,IAAM,EAAe,CAAC,GAAG,EAAO,OAAO,CAAC,EAClC,EAAQ,EAAQ,OAAO,QAC3B,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,IAAK,GAAO,EAAY,IAAI,CAAE,CAAW,EAC5D,CACF,EACA,GAAI,CAAC,OAAO,SAAS,CAAK,EACxB,MAAO,CACL,aAAcA,EAAAA,aAAa,aAC3B,iBAAkB,aAAa,EAAQ,OAAO,MAAM,0BACtD,EAEF,IAAM,EACJ,EAAQ,kBAAoB,IAAA,KAC3B,EAAQ,OAAO,eAAiB,GAAS,EAAQ,gBAAkB,GAAS,EAAQ,iBACvF,MAAO,CACL,aAAc,EAAaA,EAAAA,aAAa,SAAWA,EAAAA,aAAa,aAChE,QACA,WAAY,EAAQ,OAAO,MAC3B,iBAAkB,EAAmB,EAAO,EAAa,OAAQ,EAAY,CAAO,CACtF,CACF,CAEA,SAAS,EACP,EACA,EACA,EACA,EACQ,CACR,GAAM,CAAE,QAAO,kBAAmB,EAAQ,OACpC,EAAY,EAA4B,EAAO,EAAQ,eAAe,EACtE,EACJ,EAAQ,kBAAoB,IAAA,GACxB,GACA,WAAW,EAAM,GAAG,EAAiB,IAAM,IAAI,GAAG,EAAY,EAAQ,gBAAiB,CAAS,EAAE,GAAG,EAAa,KAAO,MAAM,GACrI,MAAO;;IAEL,EAAM,KAAK,EAAY,EAAO,CAAS,EAAE;;QAErC,EAAS,eAAe,OAAO,EAAE,GAAG,EAAe;CAE3D,CAMA,SAAS,EAA4B,EAAe,EAAuC,CAEzF,GAAI,IAAc,IAAA,IAAa,IAAU,EAAW,MAAO,GAC3D,IAAK,IAAI,EAAY,EAAkB,EAAY,GAAI,IACrD,GAAI,EAAY,EAAO,CAAS,IAAM,EAAY,EAAW,CAAS,EAAG,OAAO,EAElF,MAAO,GACT,CAEA,SAAS,EAAY,EAAe,EAA2B,CAC7D,OAAO,OAAO,EAAM,YAAY,CAAS,CAAC,CAAC,CAAC,SAAS,CACvD,CAEA,SAAS,EACP,EACA,EACA,EACA,EAC8B,CAC9B,IAAI,EACJ,GAAI,CACF,EAAUC,EAAAA,gBAAgB,CAAc,CAC1C,OAAS,EAAO,CAGd,MAAO,KAAK,EAAmB,sBAD7B,aAAiBC,EAAAA,cAAgB,GAAG,EAAM,WAAW,QAAQ,EAAyB,EAAM,UAAY,IAE5G,CACA,IAAM,EAAU,EAAY,EAAQ,EAAE,EAAE,OAAS,CAAC,EAAG,CAAO,EAC5D,GAAI,CAAC,EACH,MAAO,KAAK,EAAmB,YAAY,EAAQ,SAAS,QAAQ,EAAQ,aAAa,gBAG3F,IAAM,EAAc,IAAI,IAClB,EAAqB,CAAC,EACxB,EAAe,EACb,EAAiB,GAA0B,CAC/C,IACI,EAAS,OAAS,IAAgB,EAAS,KAAK,CAAO,CAC7D,EACA,IAAK,GAAM,CAAE,MAAO,EAAK,gBAAgB,EAAQ,MAAM,CAAC,EAAG,CACzD,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAa,EAAI,EAAQ,cAAgB,GAC/C,GAAI,CAAC,EAAO,IAAI,CAAE,EAChB,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,iBAAiB,OACjF,GAAI,EAAY,IAAI,CAAE,EAC3B,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,aAAa,MAC7E,CACL,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAU,EAClE,GACF,EAAc,GAAG,EAAW,QAAQ,EAAQ,aAAa,QAAQ,EAAK,CAAU,EAAE,KAAK,GAAiB,EAE1G,EAAY,IAAI,EAAI,CAAU,CAChC,CACF,CACA,IAAM,EAAa,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,OAAQ,GAAO,CAAC,EAAY,IAAI,CAAE,CAAC,EAczE,OAbI,EAAW,OAAS,IAEtB,IACA,EAAS,QACP,KAAK,EAAQ,SAAS,KAAK,EAAQ,CAAU,EAAE,cAAc,EAAW,OAAO,eAAe,OAAO,EAAE,GACzG,GAEE,EAAe,EACV,KAAK,EAAmB;;EAEjC,EAAS,IAAK,GAAY,KAAK,GAAS,CAAC,CAAC,KAAK;CAAI,IAAI,EAAe,EAAS,OAAS,UAAU,EAAe,EAAS,OAAA,CAAQ,eAAe,OAAO,EAAE,GAAK,GAAG;EAG3J,CACT,CAEA,eAAe,EAAW,EAAwB,EAAqE,CACrH,IAAM,EAAOC,EAAAA,SAAS,MAAMJ,EAAAA,QAAG,SAAS,EAAgB,MAAM,CAAC,EACzD,EAAU,EAAY,EAAK,IAAM,CAAC,EAAG,CAAO,EAClD,GAAI,CAAC,EACH,MAAU,MAAM,yBAAyB,EAAQ,SAAS,OAAO,EAAQ,aAAa,YAAY,GAAgB,EAEpH,IAAM,EAAS,IAAI,IACnB,IAAK,IAAM,KAAO,EAAK,MAAM,CAAC,EAAG,CAC/B,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAQ,EAAI,EAAQ,cAAgB,GAC1C,GAAI,CAAC,EAAI,MAAU,MAAM,SAAS,EAAQ,SAAS,mBAAmB,GAAgB,EACtF,GAAI,EAAO,IAAI,CAAE,EAAG,MAAU,MAAM,aAAa,EAAQ,SAAS,mBAAmB,GAAI,EACzF,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAK,EACjE,GAAI,EACF,MAAU,MAAM,WAAW,EAAQ,aAAa,sBAAsB,EAAG,IAAI,GAAiB,EAChG,EAAO,IAAI,EAAI,CAAK,CACtB,CACA,GAAI,EAAO,OAAS,EAAG,MAAU,MAAM,4BAA4B,GAAgB,EACnF,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACsD,CACtD,IAAM,EAAQ,EAAO,IAAK,GAAS,EAAK,KAAK,CAAC,EACxC,EAAU,EAAM,QAAQ,EAAQ,QAAQ,EACxC,EAAc,EAAM,QAAQ,EAAQ,YAAY,EACtD,OAAO,IAAY,IAAM,IAAgB,GAAK,IAAA,GAAY,CAAE,UAAS,aAAY,CACnF,CAEA,SAAS,EAAW,EAAiC,CACnD,OAAO,EAAI,MAAO,GAAS,EAAK,KAAK,IAAM,EAAE,CAC/C,CAEA,SAAS,EAAQ,EAAgC,CAC/C,IAAM,EAAS,EAAI,MAAM,EAAG,EAAc,CAAC,CAAC,IAAK,GAAO,KAAK,EAAK,CAAE,EAAE,GAAG,EACzE,OAAO,EAAI,OAAS,GAAiB,GAAG,EAAO,KAAK,IAAI,EAAE,OAAS,EAAO,KAAK,IAAI,CACrF,CAEA,SAAS,EAAK,EAAuB,CAEnC,IAAM,EAAY,EAAM,WAAW,aAAc,GAAG,CAAC,CAAC,WAAW,IAAK,GAAG,EAEzE,OADI,EAAU,KAAK,IAAM,GAAW,MAC7B,EAAU,OAAS,GAAyB,GAAG,EAAU,MAAM,EAAG,EAAsB,EAAE,GAAK,CACxG,CAEA,SAAS,EAAsB,EAA0B,EAAuD,CAC9G,OAAQ,EAAO,eAAiBC,EAAAA,aAAa,WAAe,EAAY,iBAAmB,WAC7F,CAEA,SAAS,EAAqB,EAAwC,CACpE,OAAO,OAAO,SAAS,OAAO,CAAU,CAAC,GAAK,EAAW,KAAK,IAAM,GAAK,IAAA,GAAY,WACvF,CAEA,SAAS,EAAK,EAAmC,CAC/C,OAAO,EAAO,QAAQ,EAAK,IAAU,EAAM,EAAO,CAAC,EAAI,EAAO,MAChE"}
+{"version":3,"file":"evaluation.cjs","names":["parseArgs","resolveCwds","passesIsolationCheckInDebugMode","path","judgeByStaticAnalysis","readProblemMarkdownFrontMatter","fs","DecisionCode","parseCsvRecords","CsvParseError","parseCsv"],"sources":["../../src/presets/evaluation.ts"],"sourcesContent":["import fs from 'node:fs/promises';\nimport path from 'node:path';\n\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { CsvParseError, type CsvParseErrorReason, parseCsv, parseCsvRecords } from '../helpers/parseCsv.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst DEFAULT_SUBMISSION_FILE_PATH = 'submission.csv';\nconst TEST_CASE_ID = 'evaluation';\nconst MAX_LISTED_IDS = 10;\n// Learner-controlled cells are echoed into the feedback, so clip them to keep the result line small.\nconst MAX_ECHOED_CELL_LENGTH = 50;\nconst CSV_PARSE_ERROR_MESSAGES: Record<CsvParseErrorReason, string> = {\n  unterminated_quote: '引用符（`\"`）が閉じられていません',\n  unexpected_quote: '引用符（`\"`）は値全体を囲む位置にしか書けません',\n  text_after_quoted_field: '閉じ引用符の後に区切り文字以外の文字があります',\n};\n\nexport interface EvaluationMetric {\n  /** Label shown with the score, e.g. `RMSLE`. */\n  label: string;\n  isHigherBetter: boolean;\n  /**\n   * Returns an error message when a predicted value cannot be scored (e.g. it is not a number).\n   * The message is appended to the row's line number and the offending value, so phrase it as a\n   * predicate such as `数値ではありません`. It also validates the ground truth when it is loaded.\n   */\n  validatePrediction?: (prediction: string) => string | undefined;\n  /** Computes the score from predictions and answers aligned by id. Both are raw CSV cell strings. */\n  compute: (predictions: readonly string[], answers: readonly string[]) => number;\n}\n\nexport interface EvaluationJudgePresetOptions {\n  /** CSV file holding the hidden ground truth, relative to the problem directory. */\n  answerFilePath: string;\n  /** CSV file the submission must contain. Defaults to `submission.csv`. */\n  submissionFilePath?: string;\n  /** Column that identifies each row in both CSV files. */\n  idColumn: string;\n  /** Column holding the answer in the ground truth and the prediction in the submission. */\n  targetColumn: string;\n  metric: EvaluationMetric;\n  /**\n   * Score required for acceptance: the submission is accepted when its score is at least this value\n   * for a higher-is-better metric and at most this value otherwise. Every valid submission is\n   * accepted when omitted.\n   */\n  acceptableScore?: number;\n}\n\nexport const evaluationMetrics = {\n  rmsle: {\n    label: 'RMSLE',\n    isHigherBetter: false,\n    validatePrediction: (prediction: string) =>\n      validateFiniteNumber(prediction) ?? (Number(prediction) < 0 ? '負の値です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(\n        mean(\n          predictions.map(\n            (prediction, index) => (Math.log1p(Number(prediction)) - Math.log1p(Number(answers[index]))) ** 2\n          )\n        )\n      ),\n  },\n  rmse: {\n    label: 'RMSE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(mean(predictions.map((prediction, index) => (Number(prediction) - Number(answers[index])) ** 2))),\n  },\n  mae: {\n    label: 'MAE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => Math.abs(Number(prediction) - Number(answers[index])))),\n  },\n  accuracy: {\n    label: 'Accuracy',\n    isHigherBetter: true,\n    validatePrediction: (prediction: string) => (prediction.trim() === '' ? '空です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => (prediction.trim() === answers[index]?.trim() ? 1 : 0))),\n  },\n} as const satisfies Record<string, EvaluationMetric>;\n\n/**\n * A preset for Kaggle-style model evaluation problems: the submission is a CSV of predictions for a\n * hidden test set, and the judge reports the metric value as `score` instead of comparing outputs.\n * The problem directory keeps the ground truth CSV, so it is never shipped to learners.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { evaluationJudgePreset, evaluationMetrics } from '@exercode/problem-utils/presets/evaluation';\n *\n * await evaluationJudgePreset(import.meta.dirname, {\n *   answerFilePath: 'evaluation/answer.csv',\n *   idColumn: 'Id',\n *   targetColumn: 'TradePrice',\n *   metric: evaluationMetrics.rmsle,\n *   acceptableScore: 0.5,\n * });\n * ```\n *\n * Run `bun judge.ts model_answers/default` to judge one submission, or `bun judge.ts` to judge every\n * `<problemDir>/model_answers/*` and `<problemDir>/model_answers.fails/*` directory for debugging.\n */\nexport async function evaluationJudgePreset(problemDir: string, options: EvaluationJudgePresetOptions): Promise<void> {\n  const args = parseArgs(process.argv);\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n\n  if (isDebugMode && !(await passesIsolationCheckInDebugMode(problemDir, cwds, args.params))) {\n    process.exitCode = 1;\n    return;\n  }\n\n  const answer = await readAnswer(path.join(problemDir, options.answerFilePath), options);\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await judgeSubmission(problemDir, resolvedCwd.cwd, answer, options);\n    printTestCaseResult({ testCaseId: TEST_CASE_ID, ...result });\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function judgeSubmission(\n  problemDir: string,\n  cwd: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Promise<Omit<TestCaseResult, 'testCaseId'>> {\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, await readProblemMarkdownFrontMatter(problemDir));\n  if (staticAnalysisResult) return staticAnalysisResult;\n\n  const submissionFilePath = options.submissionFilePath ?? DEFAULT_SUBMISSION_FILE_PATH;\n  let submissionText: string;\n  try {\n    const resolvedPath = path.join(cwd, submissionFilePath);\n    submissionText = await fs.readFile(resolvedPath, 'utf8');\n  } catch {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      feedbackMarkdown: `\\`${submissionFilePath}\\`が見つかりません。予測結果のCSVファイルを\\`${submissionFilePath}\\`という名前で提出してください。`,\n    };\n  }\n\n  const predictions = readPredictions(submissionText, submissionFilePath, answer, options);\n  if (typeof predictions === 'string') {\n    return { decisionCode: DecisionCode.WRONG_ANSWER, feedbackMarkdown: predictions };\n  }\n\n  const answerValues = [...answer.values()];\n  const score = options.metric.compute(\n    [...answer.keys()].map((id) => predictions.get(id) as string),\n    answerValues\n  );\n  if (!Number.isFinite(score)) {\n    return {\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      feedbackMarkdown: `予測値が極端なため、${options.metric.label}を計算できませんでした。予測値を見直してください。`,\n    };\n  }\n  const isAccepted =\n    options.acceptableScore === undefined ||\n    (options.metric.isHigherBetter ? score >= options.acceptableScore : score <= options.acceptableScore);\n  return {\n    decisionCode: isAccepted ? DecisionCode.ACCEPTED : DecisionCode.WRONG_ANSWER,\n    score,\n    scoreLabel: options.metric.label,\n    feedbackMarkdown: buildScoreMarkdown(score, answerValues.length, isAccepted, options),\n  };\n}\n\nfunction buildScoreMarkdown(\n  score: number,\n  rowCount: number,\n  isAccepted: boolean,\n  options: EvaluationJudgePresetOptions\n): string {\n  const { label, isHigherBetter } = options.metric;\n  const precision = findDistinguishingPrecision(score, options.acceptableScore);\n  const acceptanceLine =\n    options.acceptableScore === undefined\n      ? ''\n      : `\\n合格基準: ${label} ${isHigherBetter ? '≧' : '≦'} ${formatScore(options.acceptableScore, precision)}（${isAccepted ? '達成' : '未達成'}）`;\n  return `| 指標 | スコア |\n| ---- | ------ |\n| ${label} | ${formatScore(score, precision)} |\n\n評価件数: ${rowCount.toLocaleString('en-US')}件${acceptanceLine}\n`;\n}\n\n/**\n * Finds the number of significant digits at which the score and the threshold render differently\n * (or stop changing), so the feedback never shows a pair that contradicts the decision.\n */\nfunction findDistinguishingPrecision(score: number, threshold: number | undefined): number {\n  const minimumPrecision = 6;\n  if (threshold === undefined || score === threshold) return minimumPrecision;\n  for (let precision = minimumPrecision; precision < 17; precision++) {\n    if (formatScore(score, precision) !== formatScore(threshold, precision)) return precision;\n  }\n  return 17;\n}\n\nfunction formatScore(score: number, precision: number): string {\n  return Number(score.toPrecision(precision)).toString();\n}\n\nfunction readPredictions(\n  submissionText: string,\n  submissionFilePath: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Map<string, string> | string {\n  let records: ReturnType<typeof parseCsvRecords>;\n  try {\n    records = parseCsvRecords(submissionText);\n  } catch (error) {\n    const detail =\n      error instanceof CsvParseError ? `${error.lineNumber}行目付近: ${CSV_PARSE_ERROR_MESSAGES[error.reason]}` : '';\n    return `\\`${submissionFilePath}\\`をCSVとして読み込めませんでした。${detail}`;\n  }\n  const columns = findColumns(records[0]?.cells ?? [], options);\n  if (!columns) {\n    return `\\`${submissionFilePath}\\`の1行目には\\`${options.idColumn}\\`列と\\`${options.targetColumn}\\`列のヘッダーが必要です。`;\n  }\n\n  const predictions = new Map<string, string>();\n  const problems: string[] = [];\n  let problemCount = 0;\n  const reportProblem = (problem: string): void => {\n    problemCount++;\n    if (problems.length < MAX_LISTED_IDS) problems.push(problem);\n  };\n  for (const { cells: row, lineNumber } of records.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const prediction = row[columns.targetIndex] ?? '';\n    if (!answer.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行は評価対象ではありません`);\n    } else if (predictions.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行が重複しています`);\n    } else {\n      const validationError = options.metric.validatePrediction?.(prediction);\n      if (validationError) {\n        reportProblem(`${lineNumber}行目: \\`${options.targetColumn}\\`の値\\`${clip(prediction)}\\`は${validationError}`);\n      }\n      predictions.set(id, prediction);\n    }\n  }\n  const missingIds = [...answer.keys()].filter((id) => !predictions.has(id));\n  if (missingIds.length > 0) {\n    // Missing rows are the most actionable problem, so they must survive the listing cap.\n    problemCount++;\n    problems.unshift(\n      `\\`${options.idColumn}\\`が${listIds(missingIds)}の行がありません（不足 ${missingIds.length.toLocaleString('en-US')}件）`\n    );\n  }\n  if (problemCount > 0) {\n    return `\\`${submissionFilePath}\\`の内容に問題があります。\n\n${problems.map((problem) => `- ${problem}`).join('\\n')}${problemCount > problems.length ? `\\n- ほか${(problemCount - problems.length).toLocaleString('en-US')}件` : ''}\n`;\n  }\n  return predictions;\n}\n\nasync function readAnswer(answerFilePath: string, options: EvaluationJudgePresetOptions): Promise<Map<string, string>> {\n  const rows = parseCsv(await fs.readFile(answerFilePath, 'utf8'));\n  const columns = findColumns(rows[0] ?? [], options);\n  if (!columns) {\n    throw new Error(`answer file must have ${options.idColumn} and ${options.targetColumn} columns: ${answerFilePath}`);\n  }\n  const answer = new Map<string, string>();\n  for (const row of rows.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const value = row[columns.targetIndex] ?? '';\n    if (!id) throw new Error(`empty ${options.idColumn} in answer file: ${answerFilePath}`);\n    if (answer.has(id)) throw new Error(`duplicate ${options.idColumn} in answer file: ${id}`);\n    const validationError = options.metric.validatePrediction?.(value);\n    if (validationError)\n      throw new Error(`invalid ${options.targetColumn} in answer file for ${id}: ${validationError}`);\n    answer.set(id, value);\n  }\n  if (answer.size === 0) throw new Error(`answer file has no rows: ${answerFilePath}`);\n  return answer;\n}\n\nfunction findColumns(\n  header: readonly string[],\n  options: EvaluationJudgePresetOptions\n): { idIndex: number; targetIndex: number } | undefined {\n  const names = header.map((name) => name.trim());\n  const idIndex = names.indexOf(options.idColumn);\n  const targetIndex = names.indexOf(options.targetColumn);\n  return idIndex === -1 || targetIndex === -1 ? undefined : { idIndex, targetIndex };\n}\n\nfunction isBlankRow(row: readonly string[]): boolean {\n  return row.every((cell) => cell.trim() === '');\n}\n\nfunction listIds(ids: readonly string[]): string {\n  const listed = ids.slice(0, MAX_LISTED_IDS).map((id) => `\\`${clip(id)}\\``);\n  return ids.length > MAX_LISTED_IDS ? `${listed.join(', ')}, ...` : listed.join(', ');\n}\n\nfunction clip(value: string): string {\n  // Control characters and backticks would break the markdown list item and inline code the value is echoed in.\n  const printable = value.replaceAll(/[\\p{Cc}]/gu, ' ').replaceAll('`', \"'\");\n  if (printable.trim() === '') return '（空）';\n  return printable.length > MAX_ECHOED_CELL_LENGTH ? `${printable.slice(0, MAX_ECHOED_CELL_LENGTH)}…` : printable;\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: Pick<TestCaseResult, 'decisionCode'>): boolean {\n  return (result.decisionCode === DecisionCode.ACCEPTED) === (resolvedCwd.expectedResult === 'accepted');\n}\n\nfunction validateFiniteNumber(prediction: string): string | undefined {\n  return Number.isFinite(Number(prediction)) && prediction.trim() !== '' ? undefined : '数値ではありません';\n}\n\nfunction mean(values: readonly number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length;\n}\n"],"mappings":"olBAkBA,MAKM,EAAgE,CACpE,mBAAoB,qBACpB,iBAAkB,4BAClB,wBAAyB,yBAC3B,EAkCa,EAAoB,CAC/B,MAAO,CACL,MAAO,QACP,eAAgB,GAChB,mBAAqB,GACnB,EAAqB,CAAU,IAAM,OAAO,CAAU,EAAI,EAAI,QAAU,IAAA,IAC1E,SAAU,EAAgC,IACxC,KAAK,KACH,EACE,EAAY,KACT,EAAY,KAAW,KAAK,MAAM,OAAO,CAAU,CAAC,EAAI,KAAK,MAAM,OAAO,EAAQ,EAAM,CAAC,IAAM,CAClG,CACF,CACF,CACJ,EACA,KAAM,CACJ,MAAO,OACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,KAAK,KAAK,EAAK,EAAY,KAAK,EAAY,KAAW,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,IAAM,CAAC,CAAC,CAAC,CAC9G,EACA,IAAK,CACH,MAAO,MACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAU,KAAK,IAAI,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,CAAC,CAAC,CAAC,CACtG,EACA,SAAU,CACR,MAAO,WACP,eAAgB,GAChB,mBAAqB,GAAwB,EAAW,KAAK,IAAM,GAAK,MAAQ,IAAA,GAChF,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAW,IAAW,KAAK,IAAM,EAAQ,EAAM,EAAE,KAAK,EAAU,CAAC,CACvG,CACF,EAwBA,eAAsB,EAAsB,EAAoB,EAAsD,CACpH,IAAM,EAAOA,EAAAA,UAAU,QAAQ,IAAI,EAC7B,CAAE,OAAM,eAAgB,MAAMC,EAAAA,YAAY,EAAY,EAAK,GAAG,EAEpE,GAAI,GAAe,CAAE,MAAMC,EAAAA,gCAAgC,EAAY,EAAM,EAAK,MAAM,EAAI,CAC1F,QAAQ,SAAW,EACnB,MACF,CAEA,IAAM,EAAS,MAAM,EAAWC,EAAAA,QAAK,KAAK,EAAY,EAAQ,cAAc,EAAG,CAAO,EACtF,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAA,oBAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAAgB,EAAY,EAAY,IAAK,EAAQ,CAAO,EACjF,EAAA,oBAAoB,CAAE,WAAY,aAAc,GAAG,CAAO,CAAC,EACvD,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAA,mCAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EACb,EACA,EACA,EACA,EAC6C,CAC7C,IAAM,EAAuB,MAAMC,EAAAA,sBAAsB,EAAK,MAAMC,EAAAA,+BAA+B,CAAU,CAAC,EAC9G,GAAI,EAAsB,OAAO,EAEjC,IAAM,EAAqB,EAAQ,oBAAsB,iBACrD,EACJ,GAAI,CACF,IAAM,EAAeF,EAAAA,QAAK,KAAK,EAAK,CAAkB,EACtD,EAAiB,MAAMG,EAAAA,QAAG,SAAS,EAAc,MAAM,CACzD,MAAQ,CACN,MAAO,CACL,aAAcC,EAAAA,aAAa,uCAC3B,iBAAkB,KAAK,EAAmB,4BAA4B,EAAmB,kBAC3F,CACF,CAEA,IAAM,EAAc,EAAgB,EAAgB,EAAoB,EAAQ,CAAO,EACvF,GAAI,OAAO,GAAgB,SACzB,MAAO,CAAE,aAAcA,EAAAA,aAAa,aAAc,iBAAkB,CAAY,EAGlF,IAAM,EAAe,CAAC,GAAG,EAAO,OAAO,CAAC,EAClC,EAAQ,EAAQ,OAAO,QAC3B,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,IAAK,GAAO,EAAY,IAAI,CAAE,CAAW,EAC5D,CACF,EACA,GAAI,CAAC,OAAO,SAAS,CAAK,EACxB,MAAO,CACL,aAAcA,EAAAA,aAAa,aAC3B,iBAAkB,aAAa,EAAQ,OAAO,MAAM,0BACtD,EAEF,IAAM,EACJ,EAAQ,kBAAoB,IAAA,KAC3B,EAAQ,OAAO,eAAiB,GAAS,EAAQ,gBAAkB,GAAS,EAAQ,iBACvF,MAAO,CACL,aAAc,EAAaA,EAAAA,aAAa,SAAWA,EAAAA,aAAa,aAChE,QACA,WAAY,EAAQ,OAAO,MAC3B,iBAAkB,EAAmB,EAAO,EAAa,OAAQ,EAAY,CAAO,CACtF,CACF,CAEA,SAAS,EACP,EACA,EACA,EACA,EACQ,CACR,GAAM,CAAE,QAAO,kBAAmB,EAAQ,OACpC,EAAY,EAA4B,EAAO,EAAQ,eAAe,EACtE,EACJ,EAAQ,kBAAoB,IAAA,GACxB,GACA,WAAW,EAAM,GAAG,EAAiB,IAAM,IAAI,GAAG,EAAY,EAAQ,gBAAiB,CAAS,EAAE,GAAG,EAAa,KAAO,MAAM,GACrI,MAAO;;IAEL,EAAM,KAAK,EAAY,EAAO,CAAS,EAAE;;QAErC,EAAS,eAAe,OAAO,EAAE,GAAG,EAAe;CAE3D,CAMA,SAAS,EAA4B,EAAe,EAAuC,CAEzF,GAAI,IAAc,IAAA,IAAa,IAAU,EAAW,MAAO,GAC3D,IAAK,IAAI,EAAY,EAAkB,EAAY,GAAI,IACrD,GAAI,EAAY,EAAO,CAAS,IAAM,EAAY,EAAW,CAAS,EAAG,OAAO,EAElF,MAAO,GACT,CAEA,SAAS,EAAY,EAAe,EAA2B,CAC7D,OAAO,OAAO,EAAM,YAAY,CAAS,CAAC,CAAC,CAAC,SAAS,CACvD,CAEA,SAAS,EACP,EACA,EACA,EACA,EAC8B,CAC9B,IAAI,EACJ,GAAI,CACF,EAAUC,EAAAA,gBAAgB,CAAc,CAC1C,OAAS,EAAO,CAGd,MAAO,KAAK,EAAmB,sBAD7B,aAAiBC,EAAAA,cAAgB,GAAG,EAAM,WAAW,QAAQ,EAAyB,EAAM,UAAY,IAE5G,CACA,IAAM,EAAU,EAAY,EAAQ,EAAE,EAAE,OAAS,CAAC,EAAG,CAAO,EAC5D,GAAI,CAAC,EACH,MAAO,KAAK,EAAmB,YAAY,EAAQ,SAAS,QAAQ,EAAQ,aAAa,gBAG3F,IAAM,EAAc,IAAI,IAClB,EAAqB,CAAC,EACxB,EAAe,EACb,EAAiB,GAA0B,CAC/C,IACI,EAAS,OAAS,IAAgB,EAAS,KAAK,CAAO,CAC7D,EACA,IAAK,GAAM,CAAE,MAAO,EAAK,gBAAgB,EAAQ,MAAM,CAAC,EAAG,CACzD,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAa,EAAI,EAAQ,cAAgB,GAC/C,GAAI,CAAC,EAAO,IAAI,CAAE,EAChB,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,iBAAiB,OACjF,GAAI,EAAY,IAAI,CAAE,EAC3B,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,aAAa,MAC7E,CACL,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAU,EAClE,GACF,EAAc,GAAG,EAAW,QAAQ,EAAQ,aAAa,QAAQ,EAAK,CAAU,EAAE,KAAK,GAAiB,EAE1G,EAAY,IAAI,EAAI,CAAU,CAChC,CACF,CACA,IAAM,EAAa,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,OAAQ,GAAO,CAAC,EAAY,IAAI,CAAE,CAAC,EAczE,OAbI,EAAW,OAAS,IAEtB,IACA,EAAS,QACP,KAAK,EAAQ,SAAS,KAAK,EAAQ,CAAU,EAAE,cAAc,EAAW,OAAO,eAAe,OAAO,EAAE,GACzG,GAEE,EAAe,EACV,KAAK,EAAmB;;EAEjC,EAAS,IAAK,GAAY,KAAK,GAAS,CAAC,CAAC,KAAK;CAAI,IAAI,EAAe,EAAS,OAAS,UAAU,EAAe,EAAS,OAAA,CAAQ,eAAe,OAAO,EAAE,GAAK,GAAG;EAG3J,CACT,CAEA,eAAe,EAAW,EAAwB,EAAqE,CACrH,IAAM,EAAOC,EAAAA,SAAS,MAAMJ,EAAAA,QAAG,SAAS,EAAgB,MAAM,CAAC,EACzD,EAAU,EAAY,EAAK,IAAM,CAAC,EAAG,CAAO,EAClD,GAAI,CAAC,EACH,MAAU,MAAM,yBAAyB,EAAQ,SAAS,OAAO,EAAQ,aAAa,YAAY,GAAgB,EAEpH,IAAM,EAAS,IAAI,IACnB,IAAK,IAAM,KAAO,EAAK,MAAM,CAAC,EAAG,CAC/B,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAQ,EAAI,EAAQ,cAAgB,GAC1C,GAAI,CAAC,EAAI,MAAU,MAAM,SAAS,EAAQ,SAAS,mBAAmB,GAAgB,EACtF,GAAI,EAAO,IAAI,CAAE,EAAG,MAAU,MAAM,aAAa,EAAQ,SAAS,mBAAmB,GAAI,EACzF,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAK,EACjE,GAAI,EACF,MAAU,MAAM,WAAW,EAAQ,aAAa,sBAAsB,EAAG,IAAI,GAAiB,EAChG,EAAO,IAAI,EAAI,CAAK,CACtB,CACA,GAAI,EAAO,OAAS,EAAG,MAAU,MAAM,4BAA4B,GAAgB,EACnF,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACsD,CACtD,IAAM,EAAQ,EAAO,IAAK,GAAS,EAAK,KAAK,CAAC,EACxC,EAAU,EAAM,QAAQ,EAAQ,QAAQ,EACxC,EAAc,EAAM,QAAQ,EAAQ,YAAY,EACtD,OAAO,IAAY,IAAM,IAAgB,GAAK,IAAA,GAAY,CAAE,UAAS,aAAY,CACnF,CAEA,SAAS,EAAW,EAAiC,CACnD,OAAO,EAAI,MAAO,GAAS,EAAK,KAAK,IAAM,EAAE,CAC/C,CAEA,SAAS,EAAQ,EAAgC,CAC/C,IAAM,EAAS,EAAI,MAAM,EAAG,EAAc,CAAC,CAAC,IAAK,GAAO,KAAK,EAAK,CAAE,EAAE,GAAG,EACzE,OAAO,EAAI,OAAS,GAAiB,GAAG,EAAO,KAAK,IAAI,EAAE,OAAS,EAAO,KAAK,IAAI,CACrF,CAEA,SAAS,EAAK,EAAuB,CAEnC,IAAM,EAAY,EAAM,WAAW,aAAc,GAAG,CAAC,CAAC,WAAW,IAAK,GAAG,EAEzE,OADI,EAAU,KAAK,IAAM,GAAW,MAC7B,EAAU,OAAS,GAAyB,GAAG,EAAU,MAAM,EAAG,EAAsB,EAAE,GAAK,CACxG,CAEA,SAAS,EAAsB,EAA0B,EAAuD,CAC9G,OAAQ,EAAO,eAAiBC,EAAAA,aAAa,WAAe,EAAY,iBAAmB,WAC7F,CAEA,SAAS,EAAqB,EAAwC,CACpE,OAAO,OAAO,SAAS,OAAO,CAAU,CAAC,GAAK,EAAW,KAAK,IAAM,GAAK,IAAA,GAAY,WACvF,CAEA,SAAS,EAAK,EAAmC,CAC/C,OAAO,EAAO,QAAQ,EAAK,IAAU,EAAM,EAAO,CAAC,EAAI,EAAO,MAChE"}

package/dist/presets/evaluation.js

@@ -1,11 +1,11 @@
-import{printTestCaseResult as e}from"../helpers/printTestCaseResult.js";import{isSafeSubmissionOutputPath as t}from"../helpers/readOutputFiles.js";import{parseArgs as n}from"../helpers/parseArgs.js";import{CsvParseError as r,parseCsv as i,parseCsvRecords as a}from"../helpers/parseCsv.js";import{DecisionCode as o}from"../types/decisionCode.js";import{judgeByStaticAnalysis as s}from"../helpers/judgeByStaticAnalysis.js";import{passesIsolationCheckInDebugMode as c}from"../helpers/runIsolationCheckInDebugMode.js";import{readProblemMarkdownFrontMatter as l}from"../helpers/readProblemMarkdownFrontMatter.js";import{printDebugCwdBanner as u,printDebugExpectationFailureBanner as d,resolveCwds as f}from"../helpers/resolveCwds.js";import p from"node:path";import m from"node:fs/promises";const h={unterminated_quote:'引用符（`"`）が閉じられていません',unexpected_quote:'引用符（`"`）は値全体を囲む位置にしか書けません',text_after_quoted_field:`閉じ引用符の後に区切り文字以外の文字があります`},g={rmsle:{label:`RMSLE`,isHigherBetter:!1,validatePrediction:e=>k(e)??(Number(e)<0?`負の値です`:void 0),compute:(e,t)=>Math.sqrt(A(e.map((e,n)=>(Math.log1p(Number(e))-Math.log1p(Number(t[n])))**2)))},rmse:{label:`RMSE`,isHigherBetter:!1,validatePrediction:k,compute:(e,t)=>Math.sqrt(A(e.map((e,n)=>(Number(e)-Number(t[n]))**2)))},mae:{label:`MAE`,isHigherBetter:!1,validatePrediction:k,compute:(e,t)=>A(e.map((e,n)=>Math.abs(Number(e)-Number(t[n]))))},accuracy:{label:`Accuracy`,isHigherBetter:!0,validatePrediction:e=>e.trim()===``?`空です`:void 0,compute:(e,t)=>A(e.map((e,n)=>+(e.trim()===t[n]?.trim())))}};async function _(t,r){let i=n(process.argv),{cwds:a,isDebugMode:o}=await f(t,i.cwd);if(o&&!await c(t,a,i.params)){process.exitCode=1;return}let s=await C(p.join(t,r.answerFilePath),r);for(let n of a){o&&u(t,n);let i=await v(t,n.cwd,s,r);e({testCaseId:`evaluation`,...i}),o&&!O(n,i)&&(process.exitCode=1,d(t,n))}}async function v(e,n,r,i){let a=await s(n,await l(e));if(a)return a;let c=i.submissionFilePath??`submission.csv`,u;try{let e=p.join(n,c);if(!await t(n,e))throw Error(`unsafe path`);u=await m.readFile(e,`utf8`)}catch{return{decisionCode:o.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,feedbackMarkdown:`\`${c}\`が見つかりません。予測結果のCSVファイルを\`${c}\`という名前で提出してください。`}}let d=S(u,c,r,i);if(typeof d==`string`)return{decisionCode:o.WRONG_ANSWER,feedbackMarkdown:d};let f=[...r.values()],h=i.metric.compute([...r.keys()].map(e=>d.get(e)),f);if(!Number.isFinite(h))return{decisionCode:o.WRONG_ANSWER,feedbackMarkdown:`予測値が極端なため、${i.metric.label}を計算できませんでした。予測値を見直してください。`};let g=i.acceptableScore===void 0||(i.metric.isHigherBetter?h>=i.acceptableScore:h<=i.acceptableScore);return{decisionCode:g?o.ACCEPTED:o.WRONG_ANSWER,score:h,scoreLabel:i.metric.label,feedbackMarkdown:y(h,f.length,g,i)}}function y(e,t,n,r){let{label:i,isHigherBetter:a}=r.metric,o=b(e,r.acceptableScore),s=r.acceptableScore===void 0?``:`\n合格基準: ${i} ${a?`≧`:`≦`} ${x(r.acceptableScore,o)}（${n?`達成`:`未達成`}）`;return`| 指標 | スコア |
+import{printTestCaseResult as e}from"../helpers/printTestCaseResult.js";import{parseArgs as t}from"../helpers/parseArgs.js";import{CsvParseError as n,parseCsv as r,parseCsvRecords as i}from"../helpers/parseCsv.js";import{DecisionCode as a}from"../types/decisionCode.js";import{judgeByStaticAnalysis as o}from"../helpers/judgeByStaticAnalysis.js";import{passesIsolationCheckInDebugMode as s}from"../helpers/runIsolationCheckInDebugMode.js";import{readProblemMarkdownFrontMatter as c}from"../helpers/readProblemMarkdownFrontMatter.js";import{printDebugCwdBanner as l,printDebugExpectationFailureBanner as u,resolveCwds as d}from"../helpers/resolveCwds.js";import f from"node:path";import p from"node:fs/promises";const m={unterminated_quote:'引用符（`"`）が閉じられていません',unexpected_quote:'引用符（`"`）は値全体を囲む位置にしか書けません',text_after_quoted_field:`閉じ引用符の後に区切り文字以外の文字があります`},h={rmsle:{label:`RMSLE`,isHigherBetter:!1,validatePrediction:e=>O(e)??(Number(e)<0?`負の値です`:void 0),compute:(e,t)=>Math.sqrt(k(e.map((e,n)=>(Math.log1p(Number(e))-Math.log1p(Number(t[n])))**2)))},rmse:{label:`RMSE`,isHigherBetter:!1,validatePrediction:O,compute:(e,t)=>Math.sqrt(k(e.map((e,n)=>(Number(e)-Number(t[n]))**2)))},mae:{label:`MAE`,isHigherBetter:!1,validatePrediction:O,compute:(e,t)=>k(e.map((e,n)=>Math.abs(Number(e)-Number(t[n]))))},accuracy:{label:`Accuracy`,isHigherBetter:!0,validatePrediction:e=>e.trim()===``?`空です`:void 0,compute:(e,t)=>k(e.map((e,n)=>+(e.trim()===t[n]?.trim())))}};async function g(n,r){let i=t(process.argv),{cwds:a,isDebugMode:o}=await d(n,i.cwd);if(o&&!await s(n,a,i.params)){process.exitCode=1;return}let c=await S(f.join(n,r.answerFilePath),r);for(let t of a){o&&l(n,t);let i=await _(n,t.cwd,c,r);e({testCaseId:`evaluation`,...i}),o&&!D(t,i)&&(process.exitCode=1,u(n,t))}}async function _(e,t,n,r){let i=await o(t,await c(e));if(i)return i;let s=r.submissionFilePath??`submission.csv`,l;try{let e=f.join(t,s);l=await p.readFile(e,`utf8`)}catch{return{decisionCode:a.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,feedbackMarkdown:`\`${s}\`が見つかりません。予測結果のCSVファイルを\`${s}\`という名前で提出してください。`}}let u=x(l,s,n,r);if(typeof u==`string`)return{decisionCode:a.WRONG_ANSWER,feedbackMarkdown:u};let d=[...n.values()],m=r.metric.compute([...n.keys()].map(e=>u.get(e)),d);if(!Number.isFinite(m))return{decisionCode:a.WRONG_ANSWER,feedbackMarkdown:`予測値が極端なため、${r.metric.label}を計算できませんでした。予測値を見直してください。`};let h=r.acceptableScore===void 0||(r.metric.isHigherBetter?m>=r.acceptableScore:m<=r.acceptableScore);return{decisionCode:h?a.ACCEPTED:a.WRONG_ANSWER,score:m,scoreLabel:r.metric.label,feedbackMarkdown:v(m,d.length,h,r)}}function v(e,t,n,r){let{label:i,isHigherBetter:a}=r.metric,o=y(e,r.acceptableScore),s=r.acceptableScore===void 0?``:`\n合格基準: ${i} ${a?`≧`:`≦`} ${b(r.acceptableScore,o)}（${n?`達成`:`未達成`}）`;return`| 指標 | スコア |
 | ---- | ------ |
-| ${i} | ${x(e,o)} |
+| ${i} | ${b(e,o)} |
 
 評価件数: ${t.toLocaleString(`en-US`)}件${s}
-`}function b(e,t){if(t===void 0||e===t)return 6;for(let n=6;n<17;n++)if(x(e,n)!==x(t,n))return n;return 17}function x(e,t){return Number(e.toPrecision(t)).toString()}function S(e,t,n,i){let o;try{o=a(e)}catch(e){return`\`${t}\`をCSVとして読み込めませんでした。${e instanceof r?`${e.lineNumber}行目付近: ${h[e.reason]}`:``}`}let s=w(o[0]?.cells??[],i);if(!s)return`\`${t}\`の1行目には\`${i.idColumn}\`列と\`${i.targetColumn}\`列のヘッダーが必要です。`;let c=new Map,l=[],u=0,d=e=>{u++,l.length<10&&l.push(e)};for(let{cells:e,lineNumber:t}of o.slice(1)){if(T(e))continue;let r=e[s.idIndex]?.trim()??``,a=e[s.targetIndex]??``;if(!n.has(r))d(`${t}行目: \`${i.idColumn}\`が\`${D(r)}\`の行は評価対象ではありません`);else if(c.has(r))d(`${t}行目: \`${i.idColumn}\`が\`${D(r)}\`の行が重複しています`);else{let e=i.metric.validatePrediction?.(a);e&&d(`${t}行目: \`${i.targetColumn}\`の値\`${D(a)}\`は${e}`),c.set(r,a)}}let f=[...n.keys()].filter(e=>!c.has(e));return f.length>0&&(u++,l.unshift(`\`${i.idColumn}\`が${E(f)}の行がありません（不足 ${f.length.toLocaleString(`en-US`)}件）`)),u>0?`\`${t}\`の内容に問題があります。
+`}function y(e,t){if(t===void 0||e===t)return 6;for(let n=6;n<17;n++)if(b(e,n)!==b(t,n))return n;return 17}function b(e,t){return Number(e.toPrecision(t)).toString()}function x(e,t,r,a){let o;try{o=i(e)}catch(e){return`\`${t}\`をCSVとして読み込めませんでした。${e instanceof n?`${e.lineNumber}行目付近: ${m[e.reason]}`:``}`}let s=C(o[0]?.cells??[],a);if(!s)return`\`${t}\`の1行目には\`${a.idColumn}\`列と\`${a.targetColumn}\`列のヘッダーが必要です。`;let c=new Map,l=[],u=0,d=e=>{u++,l.length<10&&l.push(e)};for(let{cells:e,lineNumber:t}of o.slice(1)){if(w(e))continue;let n=e[s.idIndex]?.trim()??``,i=e[s.targetIndex]??``;if(!r.has(n))d(`${t}行目: \`${a.idColumn}\`が\`${E(n)}\`の行は評価対象ではありません`);else if(c.has(n))d(`${t}行目: \`${a.idColumn}\`が\`${E(n)}\`の行が重複しています`);else{let e=a.metric.validatePrediction?.(i);e&&d(`${t}行目: \`${a.targetColumn}\`の値\`${E(i)}\`は${e}`),c.set(n,i)}}let f=[...r.keys()].filter(e=>!c.has(e));return f.length>0&&(u++,l.unshift(`\`${a.idColumn}\`が${T(f)}の行がありません（不足 ${f.length.toLocaleString(`en-US`)}件）`)),u>0?`\`${t}\`の内容に問題があります。
 
 ${l.map(e=>`- ${e}`).join(`
 `)}${u>l.length?`\n- ほか${(u-l.length).toLocaleString(`en-US`)}件`:``}
-`:c}async function C(e,t){let n=i(await m.readFile(e,`utf8`)),r=w(n[0]??[],t);if(!r)throw Error(`answer file must have ${t.idColumn} and ${t.targetColumn} columns: ${e}`);let a=new Map;for(let i of n.slice(1)){if(T(i))continue;let n=i[r.idIndex]?.trim()??``,o=i[r.targetIndex]??``;if(!n)throw Error(`empty ${t.idColumn} in answer file: ${e}`);if(a.has(n))throw Error(`duplicate ${t.idColumn} in answer file: ${n}`);let s=t.metric.validatePrediction?.(o);if(s)throw Error(`invalid ${t.targetColumn} in answer file for ${n}: ${s}`);a.set(n,o)}if(a.size===0)throw Error(`answer file has no rows: ${e}`);return a}function w(e,t){let n=e.map(e=>e.trim()),r=n.indexOf(t.idColumn),i=n.indexOf(t.targetColumn);return r===-1||i===-1?void 0:{idIndex:r,targetIndex:i}}function T(e){return e.every(e=>e.trim()===``)}function E(e){let t=e.slice(0,10).map(e=>`\`${D(e)}\``);return e.length>10?`${t.join(`, `)}, ...`:t.join(`, `)}function D(e){let t=e.replaceAll(/[\p{Cc}]/gu,` `).replaceAll("`",`'`);return t.trim()===``?`（空）`:t.length>50?`${t.slice(0,50)}…`:t}function O(e,t){return t.decisionCode===o.ACCEPTED==(e.expectedResult===`accepted`)}function k(e){return Number.isFinite(Number(e))&&e.trim()!==``?void 0:`数値ではありません`}function A(e){return e.reduce((e,t)=>e+t,0)/e.length}export{_ as evaluationJudgePreset,g as evaluationMetrics};
+`:c}async function S(e,t){let n=r(await p.readFile(e,`utf8`)),i=C(n[0]??[],t);if(!i)throw Error(`answer file must have ${t.idColumn} and ${t.targetColumn} columns: ${e}`);let a=new Map;for(let r of n.slice(1)){if(w(r))continue;let n=r[i.idIndex]?.trim()??``,o=r[i.targetIndex]??``;if(!n)throw Error(`empty ${t.idColumn} in answer file: ${e}`);if(a.has(n))throw Error(`duplicate ${t.idColumn} in answer file: ${n}`);let s=t.metric.validatePrediction?.(o);if(s)throw Error(`invalid ${t.targetColumn} in answer file for ${n}: ${s}`);a.set(n,o)}if(a.size===0)throw Error(`answer file has no rows: ${e}`);return a}function C(e,t){let n=e.map(e=>e.trim()),r=n.indexOf(t.idColumn),i=n.indexOf(t.targetColumn);return r===-1||i===-1?void 0:{idIndex:r,targetIndex:i}}function w(e){return e.every(e=>e.trim()===``)}function T(e){let t=e.slice(0,10).map(e=>`\`${E(e)}\``);return e.length>10?`${t.join(`, `)}, ...`:t.join(`, `)}function E(e){let t=e.replaceAll(/[\p{Cc}]/gu,` `).replaceAll("`",`'`);return t.trim()===``?`（空）`:t.length>50?`${t.slice(0,50)}…`:t}function D(e,t){return t.decisionCode===a.ACCEPTED==(e.expectedResult===`accepted`)}function O(e){return Number.isFinite(Number(e))&&e.trim()!==``?void 0:`数値ではありません`}function k(e){return e.reduce((e,t)=>e+t,0)/e.length}export{g as evaluationJudgePreset,h as evaluationMetrics};
 //# sourceMappingURL=evaluation.js.map

package/dist/presets/evaluation.js.map

@@ -1 +1 @@
-{"version":3,"file":"evaluation.js","names":[],"sources":["../../src/presets/evaluation.ts"],"sourcesContent":["import fs from 'node:fs/promises';\nimport path from 'node:path';\n\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { CsvParseError, type CsvParseErrorReason, parseCsv, parseCsvRecords } from '../helpers/parseCsv.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { isSafeSubmissionOutputPath } from '../helpers/readOutputFiles.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst DEFAULT_SUBMISSION_FILE_PATH = 'submission.csv';\nconst TEST_CASE_ID = 'evaluation';\nconst MAX_LISTED_IDS = 10;\n// Learner-controlled cells are echoed into the feedback, so clip them to keep the result line small.\nconst MAX_ECHOED_CELL_LENGTH = 50;\nconst CSV_PARSE_ERROR_MESSAGES: Record<CsvParseErrorReason, string> = {\n  unterminated_quote: '引用符（`\"`）が閉じられていません',\n  unexpected_quote: '引用符（`\"`）は値全体を囲む位置にしか書けません',\n  text_after_quoted_field: '閉じ引用符の後に区切り文字以外の文字があります',\n};\n\nexport interface EvaluationMetric {\n  /** Label shown with the score, e.g. `RMSLE`. */\n  label: string;\n  isHigherBetter: boolean;\n  /**\n   * Returns an error message when a predicted value cannot be scored (e.g. it is not a number).\n   * The message is appended to the row's line number and the offending value, so phrase it as a\n   * predicate such as `数値ではありません`. It also validates the ground truth when it is loaded.\n   */\n  validatePrediction?: (prediction: string) => string | undefined;\n  /** Computes the score from predictions and answers aligned by id. Both are raw CSV cell strings. */\n  compute: (predictions: readonly string[], answers: readonly string[]) => number;\n}\n\nexport interface EvaluationJudgePresetOptions {\n  /** CSV file holding the hidden ground truth, relative to the problem directory. */\n  answerFilePath: string;\n  /** CSV file the submission must contain. Defaults to `submission.csv`. */\n  submissionFilePath?: string;\n  /** Column that identifies each row in both CSV files. */\n  idColumn: string;\n  /** Column holding the answer in the ground truth and the prediction in the submission. */\n  targetColumn: string;\n  metric: EvaluationMetric;\n  /**\n   * Score required for acceptance: the submission is accepted when its score is at least this value\n   * for a higher-is-better metric and at most this value otherwise. Every valid submission is\n   * accepted when omitted.\n   */\n  acceptableScore?: number;\n}\n\nexport const evaluationMetrics = {\n  rmsle: {\n    label: 'RMSLE',\n    isHigherBetter: false,\n    validatePrediction: (prediction: string) =>\n      validateFiniteNumber(prediction) ?? (Number(prediction) < 0 ? '負の値です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(\n        mean(\n          predictions.map(\n            (prediction, index) => (Math.log1p(Number(prediction)) - Math.log1p(Number(answers[index]))) ** 2\n          )\n        )\n      ),\n  },\n  rmse: {\n    label: 'RMSE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(mean(predictions.map((prediction, index) => (Number(prediction) - Number(answers[index])) ** 2))),\n  },\n  mae: {\n    label: 'MAE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => Math.abs(Number(prediction) - Number(answers[index])))),\n  },\n  accuracy: {\n    label: 'Accuracy',\n    isHigherBetter: true,\n    validatePrediction: (prediction: string) => (prediction.trim() === '' ? '空です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => (prediction.trim() === answers[index]?.trim() ? 1 : 0))),\n  },\n} as const satisfies Record<string, EvaluationMetric>;\n\n/**\n * A preset for Kaggle-style model evaluation problems: the submission is a CSV of predictions for a\n * hidden test set, and the judge reports the metric value as `score` instead of comparing outputs.\n * The problem directory keeps the ground truth CSV, so it is never shipped to learners.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { evaluationJudgePreset, evaluationMetrics } from '@exercode/problem-utils/presets/evaluation';\n *\n * await evaluationJudgePreset(import.meta.dirname, {\n *   answerFilePath: 'evaluation/answer.csv',\n *   idColumn: 'Id',\n *   targetColumn: 'TradePrice',\n *   metric: evaluationMetrics.rmsle,\n *   acceptableScore: 0.5,\n * });\n * ```\n *\n * Run `bun judge.ts model_answers/default` to judge one submission, or `bun judge.ts` to judge every\n * `<problemDir>/model_answers/*` and `<problemDir>/model_answers.fails/*` directory for debugging.\n */\nexport async function evaluationJudgePreset(problemDir: string, options: EvaluationJudgePresetOptions): Promise<void> {\n  const args = parseArgs(process.argv);\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n\n  if (isDebugMode && !(await passesIsolationCheckInDebugMode(problemDir, cwds, args.params))) {\n    process.exitCode = 1;\n    return;\n  }\n\n  const answer = await readAnswer(path.join(problemDir, options.answerFilePath), options);\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await judgeSubmission(problemDir, resolvedCwd.cwd, answer, options);\n    printTestCaseResult({ testCaseId: TEST_CASE_ID, ...result });\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function judgeSubmission(\n  problemDir: string,\n  cwd: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Promise<Omit<TestCaseResult, 'testCaseId'>> {\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, await readProblemMarkdownFrontMatter(problemDir));\n  if (staticAnalysisResult) return staticAnalysisResult;\n\n  const submissionFilePath = options.submissionFilePath ?? DEFAULT_SUBMISSION_FILE_PATH;\n  let submissionText: string;\n  try {\n    const resolvedPath = path.join(cwd, submissionFilePath);\n    // The harness reads the file as the trusted user, so refuse symlinks escaping the submission.\n    if (!(await isSafeSubmissionOutputPath(cwd, resolvedPath))) throw new Error('unsafe path');\n    submissionText = await fs.readFile(resolvedPath, 'utf8');\n  } catch {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      feedbackMarkdown: `\\`${submissionFilePath}\\`が見つかりません。予測結果のCSVファイルを\\`${submissionFilePath}\\`という名前で提出してください。`,\n    };\n  }\n\n  const predictions = readPredictions(submissionText, submissionFilePath, answer, options);\n  if (typeof predictions === 'string') {\n    return { decisionCode: DecisionCode.WRONG_ANSWER, feedbackMarkdown: predictions };\n  }\n\n  const answerValues = [...answer.values()];\n  const score = options.metric.compute(\n    [...answer.keys()].map((id) => predictions.get(id) as string),\n    answerValues\n  );\n  if (!Number.isFinite(score)) {\n    return {\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      feedbackMarkdown: `予測値が極端なため、${options.metric.label}を計算できませんでした。予測値を見直してください。`,\n    };\n  }\n  const isAccepted =\n    options.acceptableScore === undefined ||\n    (options.metric.isHigherBetter ? score >= options.acceptableScore : score <= options.acceptableScore);\n  return {\n    decisionCode: isAccepted ? DecisionCode.ACCEPTED : DecisionCode.WRONG_ANSWER,\n    score,\n    scoreLabel: options.metric.label,\n    feedbackMarkdown: buildScoreMarkdown(score, answerValues.length, isAccepted, options),\n  };\n}\n\nfunction buildScoreMarkdown(\n  score: number,\n  rowCount: number,\n  isAccepted: boolean,\n  options: EvaluationJudgePresetOptions\n): string {\n  const { label, isHigherBetter } = options.metric;\n  const precision = findDistinguishingPrecision(score, options.acceptableScore);\n  const acceptanceLine =\n    options.acceptableScore === undefined\n      ? ''\n      : `\\n合格基準: ${label} ${isHigherBetter ? '≧' : '≦'} ${formatScore(options.acceptableScore, precision)}（${isAccepted ? '達成' : '未達成'}）`;\n  return `| 指標 | スコア |\n| ---- | ------ |\n| ${label} | ${formatScore(score, precision)} |\n\n評価件数: ${rowCount.toLocaleString('en-US')}件${acceptanceLine}\n`;\n}\n\n/**\n * Finds the number of significant digits at which the score and the threshold render differently\n * (or stop changing), so the feedback never shows a pair that contradicts the decision.\n */\nfunction findDistinguishingPrecision(score: number, threshold: number | undefined): number {\n  const minimumPrecision = 6;\n  if (threshold === undefined || score === threshold) return minimumPrecision;\n  for (let precision = minimumPrecision; precision < 17; precision++) {\n    if (formatScore(score, precision) !== formatScore(threshold, precision)) return precision;\n  }\n  return 17;\n}\n\nfunction formatScore(score: number, precision: number): string {\n  return Number(score.toPrecision(precision)).toString();\n}\n\nfunction readPredictions(\n  submissionText: string,\n  submissionFilePath: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Map<string, string> | string {\n  let records: ReturnType<typeof parseCsvRecords>;\n  try {\n    records = parseCsvRecords(submissionText);\n  } catch (error) {\n    const detail =\n      error instanceof CsvParseError ? `${error.lineNumber}行目付近: ${CSV_PARSE_ERROR_MESSAGES[error.reason]}` : '';\n    return `\\`${submissionFilePath}\\`をCSVとして読み込めませんでした。${detail}`;\n  }\n  const columns = findColumns(records[0]?.cells ?? [], options);\n  if (!columns) {\n    return `\\`${submissionFilePath}\\`の1行目には\\`${options.idColumn}\\`列と\\`${options.targetColumn}\\`列のヘッダーが必要です。`;\n  }\n\n  const predictions = new Map<string, string>();\n  const problems: string[] = [];\n  let problemCount = 0;\n  const reportProblem = (problem: string): void => {\n    problemCount++;\n    if (problems.length < MAX_LISTED_IDS) problems.push(problem);\n  };\n  for (const { cells: row, lineNumber } of records.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const prediction = row[columns.targetIndex] ?? '';\n    if (!answer.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行は評価対象ではありません`);\n    } else if (predictions.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行が重複しています`);\n    } else {\n      const validationError = options.metric.validatePrediction?.(prediction);\n      if (validationError) {\n        reportProblem(`${lineNumber}行目: \\`${options.targetColumn}\\`の値\\`${clip(prediction)}\\`は${validationError}`);\n      }\n      predictions.set(id, prediction);\n    }\n  }\n  const missingIds = [...answer.keys()].filter((id) => !predictions.has(id));\n  if (missingIds.length > 0) {\n    // Missing rows are the most actionable problem, so they must survive the listing cap.\n    problemCount++;\n    problems.unshift(\n      `\\`${options.idColumn}\\`が${listIds(missingIds)}の行がありません（不足 ${missingIds.length.toLocaleString('en-US')}件）`\n    );\n  }\n  if (problemCount > 0) {\n    return `\\`${submissionFilePath}\\`の内容に問題があります。\n\n${problems.map((problem) => `- ${problem}`).join('\\n')}${problemCount > problems.length ? `\\n- ほか${(problemCount - problems.length).toLocaleString('en-US')}件` : ''}\n`;\n  }\n  return predictions;\n}\n\nasync function readAnswer(answerFilePath: string, options: EvaluationJudgePresetOptions): Promise<Map<string, string>> {\n  const rows = parseCsv(await fs.readFile(answerFilePath, 'utf8'));\n  const columns = findColumns(rows[0] ?? [], options);\n  if (!columns) {\n    throw new Error(`answer file must have ${options.idColumn} and ${options.targetColumn} columns: ${answerFilePath}`);\n  }\n  const answer = new Map<string, string>();\n  for (const row of rows.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const value = row[columns.targetIndex] ?? '';\n    if (!id) throw new Error(`empty ${options.idColumn} in answer file: ${answerFilePath}`);\n    if (answer.has(id)) throw new Error(`duplicate ${options.idColumn} in answer file: ${id}`);\n    const validationError = options.metric.validatePrediction?.(value);\n    if (validationError)\n      throw new Error(`invalid ${options.targetColumn} in answer file for ${id}: ${validationError}`);\n    answer.set(id, value);\n  }\n  if (answer.size === 0) throw new Error(`answer file has no rows: ${answerFilePath}`);\n  return answer;\n}\n\nfunction findColumns(\n  header: readonly string[],\n  options: EvaluationJudgePresetOptions\n): { idIndex: number; targetIndex: number } | undefined {\n  const names = header.map((name) => name.trim());\n  const idIndex = names.indexOf(options.idColumn);\n  const targetIndex = names.indexOf(options.targetColumn);\n  return idIndex === -1 || targetIndex === -1 ? undefined : { idIndex, targetIndex };\n}\n\nfunction isBlankRow(row: readonly string[]): boolean {\n  return row.every((cell) => cell.trim() === '');\n}\n\nfunction listIds(ids: readonly string[]): string {\n  const listed = ids.slice(0, MAX_LISTED_IDS).map((id) => `\\`${clip(id)}\\``);\n  return ids.length > MAX_LISTED_IDS ? `${listed.join(', ')}, ...` : listed.join(', ');\n}\n\nfunction clip(value: string): string {\n  // Control characters and backticks would break the markdown list item and inline code the value is echoed in.\n  const printable = value.replaceAll(/[\\p{Cc}]/gu, ' ').replaceAll('`', \"'\");\n  if (printable.trim() === '') return '（空）';\n  return printable.length > MAX_ECHOED_CELL_LENGTH ? `${printable.slice(0, MAX_ECHOED_CELL_LENGTH)}…` : printable;\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: Pick<TestCaseResult, 'decisionCode'>): boolean {\n  return (result.decisionCode === DecisionCode.ACCEPTED) === (resolvedCwd.expectedResult === 'accepted');\n}\n\nfunction validateFiniteNumber(prediction: string): string | undefined {\n  return Number.isFinite(Number(prediction)) && prediction.trim() !== '' ? undefined : '数値ではありません';\n}\n\nfunction mean(values: readonly number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length;\n}\n"],"mappings":"kxBAmBA,MAKM,EAAgE,CACpE,mBAAoB,qBACpB,iBAAkB,4BAClB,wBAAyB,yBAC3B,EAkCa,EAAoB,CAC/B,MAAO,CACL,MAAO,QACP,eAAgB,GAChB,mBAAqB,GACnB,EAAqB,CAAU,IAAM,OAAO,CAAU,EAAI,EAAI,QAAU,IAAA,IAC1E,SAAU,EAAgC,IACxC,KAAK,KACH,EACE,EAAY,KACT,EAAY,KAAW,KAAK,MAAM,OAAO,CAAU,CAAC,EAAI,KAAK,MAAM,OAAO,EAAQ,EAAM,CAAC,IAAM,CAClG,CACF,CACF,CACJ,EACA,KAAM,CACJ,MAAO,OACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,KAAK,KAAK,EAAK,EAAY,KAAK,EAAY,KAAW,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,IAAM,CAAC,CAAC,CAAC,CAC9G,EACA,IAAK,CACH,MAAO,MACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAU,KAAK,IAAI,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,CAAC,CAAC,CAAC,CACtG,EACA,SAAU,CACR,MAAO,WACP,eAAgB,GAChB,mBAAqB,GAAwB,EAAW,KAAK,IAAM,GAAK,MAAQ,IAAA,GAChF,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAW,IAAW,KAAK,IAAM,EAAQ,EAAM,EAAE,KAAK,EAAU,CAAC,CACvG,CACF,EAwBA,eAAsB,EAAsB,EAAoB,EAAsD,CACpH,IAAM,EAAO,EAAU,QAAQ,IAAI,EAC7B,CAAE,OAAM,eAAgB,MAAM,EAAY,EAAY,EAAK,GAAG,EAEpE,GAAI,GAAe,CAAE,MAAM,EAAgC,EAAY,EAAM,EAAK,MAAM,EAAI,CAC1F,QAAQ,SAAW,EACnB,MACF,CAEA,IAAM,EAAS,MAAM,EAAW,EAAK,KAAK,EAAY,EAAQ,cAAc,EAAG,CAAO,EACtF,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAAgB,EAAY,EAAY,IAAK,EAAQ,CAAO,EACjF,EAAoB,CAAE,WAAY,aAAc,GAAG,CAAO,CAAC,EACvD,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EACb,EACA,EACA,EACA,EAC6C,CAC7C,IAAM,EAAuB,MAAM,EAAsB,EAAK,MAAM,EAA+B,CAAU,CAAC,EAC9G,GAAI,EAAsB,OAAO,EAEjC,IAAM,EAAqB,EAAQ,oBAAsB,iBACrD,EACJ,GAAI,CACF,IAAM,EAAe,EAAK,KAAK,EAAK,CAAkB,EAEtD,GAAI,CAAE,MAAM,EAA2B,EAAK,CAAY,EAAI,MAAU,MAAM,aAAa,EACzF,EAAiB,MAAM,EAAG,SAAS,EAAc,MAAM,CACzD,MAAQ,CACN,MAAO,CACL,aAAc,EAAa,uCAC3B,iBAAkB,KAAK,EAAmB,4BAA4B,EAAmB,kBAC3F,CACF,CAEA,IAAM,EAAc,EAAgB,EAAgB,EAAoB,EAAQ,CAAO,EACvF,GAAI,OAAO,GAAgB,SACzB,MAAO,CAAE,aAAc,EAAa,aAAc,iBAAkB,CAAY,EAGlF,IAAM,EAAe,CAAC,GAAG,EAAO,OAAO,CAAC,EAClC,EAAQ,EAAQ,OAAO,QAC3B,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,IAAK,GAAO,EAAY,IAAI,CAAE,CAAW,EAC5D,CACF,EACA,GAAI,CAAC,OAAO,SAAS,CAAK,EACxB,MAAO,CACL,aAAc,EAAa,aAC3B,iBAAkB,aAAa,EAAQ,OAAO,MAAM,0BACtD,EAEF,IAAM,EACJ,EAAQ,kBAAoB,IAAA,KAC3B,EAAQ,OAAO,eAAiB,GAAS,EAAQ,gBAAkB,GAAS,EAAQ,iBACvF,MAAO,CACL,aAAc,EAAa,EAAa,SAAW,EAAa,aAChE,QACA,WAAY,EAAQ,OAAO,MAC3B,iBAAkB,EAAmB,EAAO,EAAa,OAAQ,EAAY,CAAO,CACtF,CACF,CAEA,SAAS,EACP,EACA,EACA,EACA,EACQ,CACR,GAAM,CAAE,QAAO,kBAAmB,EAAQ,OACpC,EAAY,EAA4B,EAAO,EAAQ,eAAe,EACtE,EACJ,EAAQ,kBAAoB,IAAA,GACxB,GACA,WAAW,EAAM,GAAG,EAAiB,IAAM,IAAI,GAAG,EAAY,EAAQ,gBAAiB,CAAS,EAAE,GAAG,EAAa,KAAO,MAAM,GACrI,MAAO;;IAEL,EAAM,KAAK,EAAY,EAAO,CAAS,EAAE;;QAErC,EAAS,eAAe,OAAO,EAAE,GAAG,EAAe;CAE3D,CAMA,SAAS,EAA4B,EAAe,EAAuC,CAEzF,GAAI,IAAc,IAAA,IAAa,IAAU,EAAW,MAAO,GAC3D,IAAK,IAAI,EAAY,EAAkB,EAAY,GAAI,IACrD,GAAI,EAAY,EAAO,CAAS,IAAM,EAAY,EAAW,CAAS,EAAG,OAAO,EAElF,MAAO,GACT,CAEA,SAAS,EAAY,EAAe,EAA2B,CAC7D,OAAO,OAAO,EAAM,YAAY,CAAS,CAAC,CAAC,CAAC,SAAS,CACvD,CAEA,SAAS,EACP,EACA,EACA,EACA,EAC8B,CAC9B,IAAI,EACJ,GAAI,CACF,EAAU,EAAgB,CAAc,CAC1C,OAAS,EAAO,CAGd,MAAO,KAAK,EAAmB,sBAD7B,aAAiB,EAAgB,GAAG,EAAM,WAAW,QAAQ,EAAyB,EAAM,UAAY,IAE5G,CACA,IAAM,EAAU,EAAY,EAAQ,EAAE,EAAE,OAAS,CAAC,EAAG,CAAO,EAC5D,GAAI,CAAC,EACH,MAAO,KAAK,EAAmB,YAAY,EAAQ,SAAS,QAAQ,EAAQ,aAAa,gBAG3F,IAAM,EAAc,IAAI,IAClB,EAAqB,CAAC,EACxB,EAAe,EACb,EAAiB,GAA0B,CAC/C,IACI,EAAS,OAAS,IAAgB,EAAS,KAAK,CAAO,CAC7D,EACA,IAAK,GAAM,CAAE,MAAO,EAAK,gBAAgB,EAAQ,MAAM,CAAC,EAAG,CACzD,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAa,EAAI,EAAQ,cAAgB,GAC/C,GAAI,CAAC,EAAO,IAAI,CAAE,EAChB,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,iBAAiB,OACjF,GAAI,EAAY,IAAI,CAAE,EAC3B,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,aAAa,MAC7E,CACL,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAU,EAClE,GACF,EAAc,GAAG,EAAW,QAAQ,EAAQ,aAAa,QAAQ,EAAK,CAAU,EAAE,KAAK,GAAiB,EAE1G,EAAY,IAAI,EAAI,CAAU,CAChC,CACF,CACA,IAAM,EAAa,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,OAAQ,GAAO,CAAC,EAAY,IAAI,CAAE,CAAC,EAczE,OAbI,EAAW,OAAS,IAEtB,IACA,EAAS,QACP,KAAK,EAAQ,SAAS,KAAK,EAAQ,CAAU,EAAE,cAAc,EAAW,OAAO,eAAe,OAAO,EAAE,GACzG,GAEE,EAAe,EACV,KAAK,EAAmB;;EAEjC,EAAS,IAAK,GAAY,KAAK,GAAS,CAAC,CAAC,KAAK;CAAI,IAAI,EAAe,EAAS,OAAS,UAAU,EAAe,EAAS,OAAA,CAAQ,eAAe,OAAO,EAAE,GAAK,GAAG;EAG3J,CACT,CAEA,eAAe,EAAW,EAAwB,EAAqE,CACrH,IAAM,EAAO,EAAS,MAAM,EAAG,SAAS,EAAgB,MAAM,CAAC,EACzD,EAAU,EAAY,EAAK,IAAM,CAAC,EAAG,CAAO,EAClD,GAAI,CAAC,EACH,MAAU,MAAM,yBAAyB,EAAQ,SAAS,OAAO,EAAQ,aAAa,YAAY,GAAgB,EAEpH,IAAM,EAAS,IAAI,IACnB,IAAK,IAAM,KAAO,EAAK,MAAM,CAAC,EAAG,CAC/B,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAQ,EAAI,EAAQ,cAAgB,GAC1C,GAAI,CAAC,EAAI,MAAU,MAAM,SAAS,EAAQ,SAAS,mBAAmB,GAAgB,EACtF,GAAI,EAAO,IAAI,CAAE,EAAG,MAAU,MAAM,aAAa,EAAQ,SAAS,mBAAmB,GAAI,EACzF,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAK,EACjE,GAAI,EACF,MAAU,MAAM,WAAW,EAAQ,aAAa,sBAAsB,EAAG,IAAI,GAAiB,EAChG,EAAO,IAAI,EAAI,CAAK,CACtB,CACA,GAAI,EAAO,OAAS,EAAG,MAAU,MAAM,4BAA4B,GAAgB,EACnF,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACsD,CACtD,IAAM,EAAQ,EAAO,IAAK,GAAS,EAAK,KAAK,CAAC,EACxC,EAAU,EAAM,QAAQ,EAAQ,QAAQ,EACxC,EAAc,EAAM,QAAQ,EAAQ,YAAY,EACtD,OAAO,IAAY,IAAM,IAAgB,GAAK,IAAA,GAAY,CAAE,UAAS,aAAY,CACnF,CAEA,SAAS,EAAW,EAAiC,CACnD,OAAO,EAAI,MAAO,GAAS,EAAK,KAAK,IAAM,EAAE,CAC/C,CAEA,SAAS,EAAQ,EAAgC,CAC/C,IAAM,EAAS,EAAI,MAAM,EAAG,EAAc,CAAC,CAAC,IAAK,GAAO,KAAK,EAAK,CAAE,EAAE,GAAG,EACzE,OAAO,EAAI,OAAS,GAAiB,GAAG,EAAO,KAAK,IAAI,EAAE,OAAS,EAAO,KAAK,IAAI,CACrF,CAEA,SAAS,EAAK,EAAuB,CAEnC,IAAM,EAAY,EAAM,WAAW,aAAc,GAAG,CAAC,CAAC,WAAW,IAAK,GAAG,EAEzE,OADI,EAAU,KAAK,IAAM,GAAW,MAC7B,EAAU,OAAS,GAAyB,GAAG,EAAU,MAAM,EAAG,EAAsB,EAAE,GAAK,CACxG,CAEA,SAAS,EAAsB,EAA0B,EAAuD,CAC9G,OAAQ,EAAO,eAAiB,EAAa,WAAe,EAAY,iBAAmB,WAC7F,CAEA,SAAS,EAAqB,EAAwC,CACpE,OAAO,OAAO,SAAS,OAAO,CAAU,CAAC,GAAK,EAAW,KAAK,IAAM,GAAK,IAAA,GAAY,WACvF,CAEA,SAAS,EAAK,EAAmC,CAC/C,OAAO,EAAO,QAAQ,EAAK,IAAU,EAAM,EAAO,CAAC,EAAI,EAAO,MAChE"}
+{"version":3,"file":"evaluation.js","names":[],"sources":["../../src/presets/evaluation.ts"],"sourcesContent":["import fs from 'node:fs/promises';\nimport path from 'node:path';\n\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { CsvParseError, type CsvParseErrorReason, parseCsv, parseCsvRecords } from '../helpers/parseCsv.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst DEFAULT_SUBMISSION_FILE_PATH = 'submission.csv';\nconst TEST_CASE_ID = 'evaluation';\nconst MAX_LISTED_IDS = 10;\n// Learner-controlled cells are echoed into the feedback, so clip them to keep the result line small.\nconst MAX_ECHOED_CELL_LENGTH = 50;\nconst CSV_PARSE_ERROR_MESSAGES: Record<CsvParseErrorReason, string> = {\n  unterminated_quote: '引用符（`\"`）が閉じられていません',\n  unexpected_quote: '引用符（`\"`）は値全体を囲む位置にしか書けません',\n  text_after_quoted_field: '閉じ引用符の後に区切り文字以外の文字があります',\n};\n\nexport interface EvaluationMetric {\n  /** Label shown with the score, e.g. `RMSLE`. */\n  label: string;\n  isHigherBetter: boolean;\n  /**\n   * Returns an error message when a predicted value cannot be scored (e.g. it is not a number).\n   * The message is appended to the row's line number and the offending value, so phrase it as a\n   * predicate such as `数値ではありません`. It also validates the ground truth when it is loaded.\n   */\n  validatePrediction?: (prediction: string) => string | undefined;\n  /** Computes the score from predictions and answers aligned by id. Both are raw CSV cell strings. */\n  compute: (predictions: readonly string[], answers: readonly string[]) => number;\n}\n\nexport interface EvaluationJudgePresetOptions {\n  /** CSV file holding the hidden ground truth, relative to the problem directory. */\n  answerFilePath: string;\n  /** CSV file the submission must contain. Defaults to `submission.csv`. */\n  submissionFilePath?: string;\n  /** Column that identifies each row in both CSV files. */\n  idColumn: string;\n  /** Column holding the answer in the ground truth and the prediction in the submission. */\n  targetColumn: string;\n  metric: EvaluationMetric;\n  /**\n   * Score required for acceptance: the submission is accepted when its score is at least this value\n   * for a higher-is-better metric and at most this value otherwise. Every valid submission is\n   * accepted when omitted.\n   */\n  acceptableScore?: number;\n}\n\nexport const evaluationMetrics = {\n  rmsle: {\n    label: 'RMSLE',\n    isHigherBetter: false,\n    validatePrediction: (prediction: string) =>\n      validateFiniteNumber(prediction) ?? (Number(prediction) < 0 ? '負の値です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(\n        mean(\n          predictions.map(\n            (prediction, index) => (Math.log1p(Number(prediction)) - Math.log1p(Number(answers[index]))) ** 2\n          )\n        )\n      ),\n  },\n  rmse: {\n    label: 'RMSE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      Math.sqrt(mean(predictions.map((prediction, index) => (Number(prediction) - Number(answers[index])) ** 2))),\n  },\n  mae: {\n    label: 'MAE',\n    isHigherBetter: false,\n    validatePrediction: validateFiniteNumber,\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => Math.abs(Number(prediction) - Number(answers[index])))),\n  },\n  accuracy: {\n    label: 'Accuracy',\n    isHigherBetter: true,\n    validatePrediction: (prediction: string) => (prediction.trim() === '' ? '空です' : undefined),\n    compute: (predictions: readonly string[], answers: readonly string[]) =>\n      mean(predictions.map((prediction, index) => (prediction.trim() === answers[index]?.trim() ? 1 : 0))),\n  },\n} as const satisfies Record<string, EvaluationMetric>;\n\n/**\n * A preset for Kaggle-style model evaluation problems: the submission is a CSV of predictions for a\n * hidden test set, and the judge reports the metric value as `score` instead of comparing outputs.\n * The problem directory keeps the ground truth CSV, so it is never shipped to learners.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { evaluationJudgePreset, evaluationMetrics } from '@exercode/problem-utils/presets/evaluation';\n *\n * await evaluationJudgePreset(import.meta.dirname, {\n *   answerFilePath: 'evaluation/answer.csv',\n *   idColumn: 'Id',\n *   targetColumn: 'TradePrice',\n *   metric: evaluationMetrics.rmsle,\n *   acceptableScore: 0.5,\n * });\n * ```\n *\n * Run `bun judge.ts model_answers/default` to judge one submission, or `bun judge.ts` to judge every\n * `<problemDir>/model_answers/*` and `<problemDir>/model_answers.fails/*` directory for debugging.\n */\nexport async function evaluationJudgePreset(problemDir: string, options: EvaluationJudgePresetOptions): Promise<void> {\n  const args = parseArgs(process.argv);\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n\n  if (isDebugMode && !(await passesIsolationCheckInDebugMode(problemDir, cwds, args.params))) {\n    process.exitCode = 1;\n    return;\n  }\n\n  const answer = await readAnswer(path.join(problemDir, options.answerFilePath), options);\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await judgeSubmission(problemDir, resolvedCwd.cwd, answer, options);\n    printTestCaseResult({ testCaseId: TEST_CASE_ID, ...result });\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function judgeSubmission(\n  problemDir: string,\n  cwd: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Promise<Omit<TestCaseResult, 'testCaseId'>> {\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, await readProblemMarkdownFrontMatter(problemDir));\n  if (staticAnalysisResult) return staticAnalysisResult;\n\n  const submissionFilePath = options.submissionFilePath ?? DEFAULT_SUBMISSION_FILE_PATH;\n  let submissionText: string;\n  try {\n    const resolvedPath = path.join(cwd, submissionFilePath);\n    submissionText = await fs.readFile(resolvedPath, 'utf8');\n  } catch {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      feedbackMarkdown: `\\`${submissionFilePath}\\`が見つかりません。予測結果のCSVファイルを\\`${submissionFilePath}\\`という名前で提出してください。`,\n    };\n  }\n\n  const predictions = readPredictions(submissionText, submissionFilePath, answer, options);\n  if (typeof predictions === 'string') {\n    return { decisionCode: DecisionCode.WRONG_ANSWER, feedbackMarkdown: predictions };\n  }\n\n  const answerValues = [...answer.values()];\n  const score = options.metric.compute(\n    [...answer.keys()].map((id) => predictions.get(id) as string),\n    answerValues\n  );\n  if (!Number.isFinite(score)) {\n    return {\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      feedbackMarkdown: `予測値が極端なため、${options.metric.label}を計算できませんでした。予測値を見直してください。`,\n    };\n  }\n  const isAccepted =\n    options.acceptableScore === undefined ||\n    (options.metric.isHigherBetter ? score >= options.acceptableScore : score <= options.acceptableScore);\n  return {\n    decisionCode: isAccepted ? DecisionCode.ACCEPTED : DecisionCode.WRONG_ANSWER,\n    score,\n    scoreLabel: options.metric.label,\n    feedbackMarkdown: buildScoreMarkdown(score, answerValues.length, isAccepted, options),\n  };\n}\n\nfunction buildScoreMarkdown(\n  score: number,\n  rowCount: number,\n  isAccepted: boolean,\n  options: EvaluationJudgePresetOptions\n): string {\n  const { label, isHigherBetter } = options.metric;\n  const precision = findDistinguishingPrecision(score, options.acceptableScore);\n  const acceptanceLine =\n    options.acceptableScore === undefined\n      ? ''\n      : `\\n合格基準: ${label} ${isHigherBetter ? '≧' : '≦'} ${formatScore(options.acceptableScore, precision)}（${isAccepted ? '達成' : '未達成'}）`;\n  return `| 指標 | スコア |\n| ---- | ------ |\n| ${label} | ${formatScore(score, precision)} |\n\n評価件数: ${rowCount.toLocaleString('en-US')}件${acceptanceLine}\n`;\n}\n\n/**\n * Finds the number of significant digits at which the score and the threshold render differently\n * (or stop changing), so the feedback never shows a pair that contradicts the decision.\n */\nfunction findDistinguishingPrecision(score: number, threshold: number | undefined): number {\n  const minimumPrecision = 6;\n  if (threshold === undefined || score === threshold) return minimumPrecision;\n  for (let precision = minimumPrecision; precision < 17; precision++) {\n    if (formatScore(score, precision) !== formatScore(threshold, precision)) return precision;\n  }\n  return 17;\n}\n\nfunction formatScore(score: number, precision: number): string {\n  return Number(score.toPrecision(precision)).toString();\n}\n\nfunction readPredictions(\n  submissionText: string,\n  submissionFilePath: string,\n  answer: ReadonlyMap<string, string>,\n  options: EvaluationJudgePresetOptions\n): Map<string, string> | string {\n  let records: ReturnType<typeof parseCsvRecords>;\n  try {\n    records = parseCsvRecords(submissionText);\n  } catch (error) {\n    const detail =\n      error instanceof CsvParseError ? `${error.lineNumber}行目付近: ${CSV_PARSE_ERROR_MESSAGES[error.reason]}` : '';\n    return `\\`${submissionFilePath}\\`をCSVとして読み込めませんでした。${detail}`;\n  }\n  const columns = findColumns(records[0]?.cells ?? [], options);\n  if (!columns) {\n    return `\\`${submissionFilePath}\\`の1行目には\\`${options.idColumn}\\`列と\\`${options.targetColumn}\\`列のヘッダーが必要です。`;\n  }\n\n  const predictions = new Map<string, string>();\n  const problems: string[] = [];\n  let problemCount = 0;\n  const reportProblem = (problem: string): void => {\n    problemCount++;\n    if (problems.length < MAX_LISTED_IDS) problems.push(problem);\n  };\n  for (const { cells: row, lineNumber } of records.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const prediction = row[columns.targetIndex] ?? '';\n    if (!answer.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行は評価対象ではありません`);\n    } else if (predictions.has(id)) {\n      reportProblem(`${lineNumber}行目: \\`${options.idColumn}\\`が\\`${clip(id)}\\`の行が重複しています`);\n    } else {\n      const validationError = options.metric.validatePrediction?.(prediction);\n      if (validationError) {\n        reportProblem(`${lineNumber}行目: \\`${options.targetColumn}\\`の値\\`${clip(prediction)}\\`は${validationError}`);\n      }\n      predictions.set(id, prediction);\n    }\n  }\n  const missingIds = [...answer.keys()].filter((id) => !predictions.has(id));\n  if (missingIds.length > 0) {\n    // Missing rows are the most actionable problem, so they must survive the listing cap.\n    problemCount++;\n    problems.unshift(\n      `\\`${options.idColumn}\\`が${listIds(missingIds)}の行がありません（不足 ${missingIds.length.toLocaleString('en-US')}件）`\n    );\n  }\n  if (problemCount > 0) {\n    return `\\`${submissionFilePath}\\`の内容に問題があります。\n\n${problems.map((problem) => `- ${problem}`).join('\\n')}${problemCount > problems.length ? `\\n- ほか${(problemCount - problems.length).toLocaleString('en-US')}件` : ''}\n`;\n  }\n  return predictions;\n}\n\nasync function readAnswer(answerFilePath: string, options: EvaluationJudgePresetOptions): Promise<Map<string, string>> {\n  const rows = parseCsv(await fs.readFile(answerFilePath, 'utf8'));\n  const columns = findColumns(rows[0] ?? [], options);\n  if (!columns) {\n    throw new Error(`answer file must have ${options.idColumn} and ${options.targetColumn} columns: ${answerFilePath}`);\n  }\n  const answer = new Map<string, string>();\n  for (const row of rows.slice(1)) {\n    if (isBlankRow(row)) continue;\n    const id = row[columns.idIndex]?.trim() ?? '';\n    const value = row[columns.targetIndex] ?? '';\n    if (!id) throw new Error(`empty ${options.idColumn} in answer file: ${answerFilePath}`);\n    if (answer.has(id)) throw new Error(`duplicate ${options.idColumn} in answer file: ${id}`);\n    const validationError = options.metric.validatePrediction?.(value);\n    if (validationError)\n      throw new Error(`invalid ${options.targetColumn} in answer file for ${id}: ${validationError}`);\n    answer.set(id, value);\n  }\n  if (answer.size === 0) throw new Error(`answer file has no rows: ${answerFilePath}`);\n  return answer;\n}\n\nfunction findColumns(\n  header: readonly string[],\n  options: EvaluationJudgePresetOptions\n): { idIndex: number; targetIndex: number } | undefined {\n  const names = header.map((name) => name.trim());\n  const idIndex = names.indexOf(options.idColumn);\n  const targetIndex = names.indexOf(options.targetColumn);\n  return idIndex === -1 || targetIndex === -1 ? undefined : { idIndex, targetIndex };\n}\n\nfunction isBlankRow(row: readonly string[]): boolean {\n  return row.every((cell) => cell.trim() === '');\n}\n\nfunction listIds(ids: readonly string[]): string {\n  const listed = ids.slice(0, MAX_LISTED_IDS).map((id) => `\\`${clip(id)}\\``);\n  return ids.length > MAX_LISTED_IDS ? `${listed.join(', ')}, ...` : listed.join(', ');\n}\n\nfunction clip(value: string): string {\n  // Control characters and backticks would break the markdown list item and inline code the value is echoed in.\n  const printable = value.replaceAll(/[\\p{Cc}]/gu, ' ').replaceAll('`', \"'\");\n  if (printable.trim() === '') return '（空）';\n  return printable.length > MAX_ECHOED_CELL_LENGTH ? `${printable.slice(0, MAX_ECHOED_CELL_LENGTH)}…` : printable;\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: Pick<TestCaseResult, 'decisionCode'>): boolean {\n  return (result.decisionCode === DecisionCode.ACCEPTED) === (resolvedCwd.expectedResult === 'accepted');\n}\n\nfunction validateFiniteNumber(prediction: string): string | undefined {\n  return Number.isFinite(Number(prediction)) && prediction.trim() !== '' ? undefined : '数値ではありません';\n}\n\nfunction mean(values: readonly number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length;\n}\n"],"mappings":"usBAkBA,MAKM,EAAgE,CACpE,mBAAoB,qBACpB,iBAAkB,4BAClB,wBAAyB,yBAC3B,EAkCa,EAAoB,CAC/B,MAAO,CACL,MAAO,QACP,eAAgB,GAChB,mBAAqB,GACnB,EAAqB,CAAU,IAAM,OAAO,CAAU,EAAI,EAAI,QAAU,IAAA,IAC1E,SAAU,EAAgC,IACxC,KAAK,KACH,EACE,EAAY,KACT,EAAY,KAAW,KAAK,MAAM,OAAO,CAAU,CAAC,EAAI,KAAK,MAAM,OAAO,EAAQ,EAAM,CAAC,IAAM,CAClG,CACF,CACF,CACJ,EACA,KAAM,CACJ,MAAO,OACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,KAAK,KAAK,EAAK,EAAY,KAAK,EAAY,KAAW,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,IAAM,CAAC,CAAC,CAAC,CAC9G,EACA,IAAK,CACH,MAAO,MACP,eAAgB,GAChB,mBAAoB,EACpB,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAU,KAAK,IAAI,OAAO,CAAU,EAAI,OAAO,EAAQ,EAAM,CAAC,CAAC,CAAC,CACtG,EACA,SAAU,CACR,MAAO,WACP,eAAgB,GAChB,mBAAqB,GAAwB,EAAW,KAAK,IAAM,GAAK,MAAQ,IAAA,GAChF,SAAU,EAAgC,IACxC,EAAK,EAAY,KAAK,EAAY,IAAW,IAAW,KAAK,IAAM,EAAQ,EAAM,EAAE,KAAK,EAAU,CAAC,CACvG,CACF,EAwBA,eAAsB,EAAsB,EAAoB,EAAsD,CACpH,IAAM,EAAO,EAAU,QAAQ,IAAI,EAC7B,CAAE,OAAM,eAAgB,MAAM,EAAY,EAAY,EAAK,GAAG,EAEpE,GAAI,GAAe,CAAE,MAAM,EAAgC,EAAY,EAAM,EAAK,MAAM,EAAI,CAC1F,QAAQ,SAAW,EACnB,MACF,CAEA,IAAM,EAAS,MAAM,EAAW,EAAK,KAAK,EAAY,EAAQ,cAAc,EAAG,CAAO,EACtF,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAAgB,EAAY,EAAY,IAAK,EAAQ,CAAO,EACjF,EAAoB,CAAE,WAAY,aAAc,GAAG,CAAO,CAAC,EACvD,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EACb,EACA,EACA,EACA,EAC6C,CAC7C,IAAM,EAAuB,MAAM,EAAsB,EAAK,MAAM,EAA+B,CAAU,CAAC,EAC9G,GAAI,EAAsB,OAAO,EAEjC,IAAM,EAAqB,EAAQ,oBAAsB,iBACrD,EACJ,GAAI,CACF,IAAM,EAAe,EAAK,KAAK,EAAK,CAAkB,EACtD,EAAiB,MAAM,EAAG,SAAS,EAAc,MAAM,CACzD,MAAQ,CACN,MAAO,CACL,aAAc,EAAa,uCAC3B,iBAAkB,KAAK,EAAmB,4BAA4B,EAAmB,kBAC3F,CACF,CAEA,IAAM,EAAc,EAAgB,EAAgB,EAAoB,EAAQ,CAAO,EACvF,GAAI,OAAO,GAAgB,SACzB,MAAO,CAAE,aAAc,EAAa,aAAc,iBAAkB,CAAY,EAGlF,IAAM,EAAe,CAAC,GAAG,EAAO,OAAO,CAAC,EAClC,EAAQ,EAAQ,OAAO,QAC3B,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,IAAK,GAAO,EAAY,IAAI,CAAE,CAAW,EAC5D,CACF,EACA,GAAI,CAAC,OAAO,SAAS,CAAK,EACxB,MAAO,CACL,aAAc,EAAa,aAC3B,iBAAkB,aAAa,EAAQ,OAAO,MAAM,0BACtD,EAEF,IAAM,EACJ,EAAQ,kBAAoB,IAAA,KAC3B,EAAQ,OAAO,eAAiB,GAAS,EAAQ,gBAAkB,GAAS,EAAQ,iBACvF,MAAO,CACL,aAAc,EAAa,EAAa,SAAW,EAAa,aAChE,QACA,WAAY,EAAQ,OAAO,MAC3B,iBAAkB,EAAmB,EAAO,EAAa,OAAQ,EAAY,CAAO,CACtF,CACF,CAEA,SAAS,EACP,EACA,EACA,EACA,EACQ,CACR,GAAM,CAAE,QAAO,kBAAmB,EAAQ,OACpC,EAAY,EAA4B,EAAO,EAAQ,eAAe,EACtE,EACJ,EAAQ,kBAAoB,IAAA,GACxB,GACA,WAAW,EAAM,GAAG,EAAiB,IAAM,IAAI,GAAG,EAAY,EAAQ,gBAAiB,CAAS,EAAE,GAAG,EAAa,KAAO,MAAM,GACrI,MAAO;;IAEL,EAAM,KAAK,EAAY,EAAO,CAAS,EAAE;;QAErC,EAAS,eAAe,OAAO,EAAE,GAAG,EAAe;CAE3D,CAMA,SAAS,EAA4B,EAAe,EAAuC,CAEzF,GAAI,IAAc,IAAA,IAAa,IAAU,EAAW,MAAO,GAC3D,IAAK,IAAI,EAAY,EAAkB,EAAY,GAAI,IACrD,GAAI,EAAY,EAAO,CAAS,IAAM,EAAY,EAAW,CAAS,EAAG,OAAO,EAElF,MAAO,GACT,CAEA,SAAS,EAAY,EAAe,EAA2B,CAC7D,OAAO,OAAO,EAAM,YAAY,CAAS,CAAC,CAAC,CAAC,SAAS,CACvD,CAEA,SAAS,EACP,EACA,EACA,EACA,EAC8B,CAC9B,IAAI,EACJ,GAAI,CACF,EAAU,EAAgB,CAAc,CAC1C,OAAS,EAAO,CAGd,MAAO,KAAK,EAAmB,sBAD7B,aAAiB,EAAgB,GAAG,EAAM,WAAW,QAAQ,EAAyB,EAAM,UAAY,IAE5G,CACA,IAAM,EAAU,EAAY,EAAQ,EAAE,EAAE,OAAS,CAAC,EAAG,CAAO,EAC5D,GAAI,CAAC,EACH,MAAO,KAAK,EAAmB,YAAY,EAAQ,SAAS,QAAQ,EAAQ,aAAa,gBAG3F,IAAM,EAAc,IAAI,IAClB,EAAqB,CAAC,EACxB,EAAe,EACb,EAAiB,GAA0B,CAC/C,IACI,EAAS,OAAS,IAAgB,EAAS,KAAK,CAAO,CAC7D,EACA,IAAK,GAAM,CAAE,MAAO,EAAK,gBAAgB,EAAQ,MAAM,CAAC,EAAG,CACzD,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAa,EAAI,EAAQ,cAAgB,GAC/C,GAAI,CAAC,EAAO,IAAI,CAAE,EAChB,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,iBAAiB,OACjF,GAAI,EAAY,IAAI,CAAE,EAC3B,EAAc,GAAG,EAAW,QAAQ,EAAQ,SAAS,OAAO,EAAK,CAAE,EAAE,aAAa,MAC7E,CACL,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAU,EAClE,GACF,EAAc,GAAG,EAAW,QAAQ,EAAQ,aAAa,QAAQ,EAAK,CAAU,EAAE,KAAK,GAAiB,EAE1G,EAAY,IAAI,EAAI,CAAU,CAChC,CACF,CACA,IAAM,EAAa,CAAC,GAAG,EAAO,KAAK,CAAC,CAAC,CAAC,OAAQ,GAAO,CAAC,EAAY,IAAI,CAAE,CAAC,EAczE,OAbI,EAAW,OAAS,IAEtB,IACA,EAAS,QACP,KAAK,EAAQ,SAAS,KAAK,EAAQ,CAAU,EAAE,cAAc,EAAW,OAAO,eAAe,OAAO,EAAE,GACzG,GAEE,EAAe,EACV,KAAK,EAAmB;;EAEjC,EAAS,IAAK,GAAY,KAAK,GAAS,CAAC,CAAC,KAAK;CAAI,IAAI,EAAe,EAAS,OAAS,UAAU,EAAe,EAAS,OAAA,CAAQ,eAAe,OAAO,EAAE,GAAK,GAAG;EAG3J,CACT,CAEA,eAAe,EAAW,EAAwB,EAAqE,CACrH,IAAM,EAAO,EAAS,MAAM,EAAG,SAAS,EAAgB,MAAM,CAAC,EACzD,EAAU,EAAY,EAAK,IAAM,CAAC,EAAG,CAAO,EAClD,GAAI,CAAC,EACH,MAAU,MAAM,yBAAyB,EAAQ,SAAS,OAAO,EAAQ,aAAa,YAAY,GAAgB,EAEpH,IAAM,EAAS,IAAI,IACnB,IAAK,IAAM,KAAO,EAAK,MAAM,CAAC,EAAG,CAC/B,GAAI,EAAW,CAAG,EAAG,SACrB,IAAM,EAAK,EAAI,EAAQ,QAAQ,EAAE,KAAK,GAAK,GACrC,EAAQ,EAAI,EAAQ,cAAgB,GAC1C,GAAI,CAAC,EAAI,MAAU,MAAM,SAAS,EAAQ,SAAS,mBAAmB,GAAgB,EACtF,GAAI,EAAO,IAAI,CAAE,EAAG,MAAU,MAAM,aAAa,EAAQ,SAAS,mBAAmB,GAAI,EACzF,IAAM,EAAkB,EAAQ,OAAO,qBAAqB,CAAK,EACjE,GAAI,EACF,MAAU,MAAM,WAAW,EAAQ,aAAa,sBAAsB,EAAG,IAAI,GAAiB,EAChG,EAAO,IAAI,EAAI,CAAK,CACtB,CACA,GAAI,EAAO,OAAS,EAAG,MAAU,MAAM,4BAA4B,GAAgB,EACnF,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACsD,CACtD,IAAM,EAAQ,EAAO,IAAK,GAAS,EAAK,KAAK,CAAC,EACxC,EAAU,EAAM,QAAQ,EAAQ,QAAQ,EACxC,EAAc,EAAM,QAAQ,EAAQ,YAAY,EACtD,OAAO,IAAY,IAAM,IAAgB,GAAK,IAAA,GAAY,CAAE,UAAS,aAAY,CACnF,CAEA,SAAS,EAAW,EAAiC,CACnD,OAAO,EAAI,MAAO,GAAS,EAAK,KAAK,IAAM,EAAE,CAC/C,CAEA,SAAS,EAAQ,EAAgC,CAC/C,IAAM,EAAS,EAAI,MAAM,EAAG,EAAc,CAAC,CAAC,IAAK,GAAO,KAAK,EAAK,CAAE,EAAE,GAAG,EACzE,OAAO,EAAI,OAAS,GAAiB,GAAG,EAAO,KAAK,IAAI,EAAE,OAAS,EAAO,KAAK,IAAI,CACrF,CAEA,SAAS,EAAK,EAAuB,CAEnC,IAAM,EAAY,EAAM,WAAW,aAAc,GAAG,CAAC,CAAC,WAAW,IAAK,GAAG,EAEzE,OADI,EAAU,KAAK,IAAM,GAAW,MAC7B,EAAU,OAAS,GAAyB,GAAG,EAAU,MAAM,EAAG,EAAsB,EAAE,GAAK,CACxG,CAEA,SAAS,EAAsB,EAA0B,EAAuD,CAC9G,OAAQ,EAAO,eAAiB,EAAa,WAAe,EAAY,iBAAmB,WAC7F,CAEA,SAAS,EAAqB,EAAwC,CACpE,OAAO,OAAO,SAAS,OAAO,CAAU,CAAC,GAAK,EAAW,KAAK,IAAM,GAAK,IAAA,GAAY,WACvF,CAEA,SAAS,EAAK,EAAmC,CAC/C,OAAO,EAAO,QAAQ,EAAK,IAAU,EAAM,EAAO,CAAC,EAAI,EAAO,MAChE"}

package/dist/presets/guiCommand.cjs

@@ -1,4 +1,4 @@
-"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:`Module`});const e=require("../_virtual/_rolldown/runtime.cjs"),t=require("../helpers/printTestCaseResult.cjs"),n=require("../helpers/sandboxUser.cjs"),r=require("../helpers/readOutputFiles.cjs"),i=require("../helpers/parseArgs.cjs"),a=require("../helpers/readTestCases.cjs"),o=require("../types/language.cjs"),s=require("../types/decisionCode.cjs"),c=require("../helpers/cleanWorkingDirectory.cjs"),l=require("../helpers/copyTestCaseFileInput.cjs"),u=require("../helpers/findEntryPointFile.cjs"),d=require("../helpers/findLanguageDefinitionByPath.cjs"),f=require("../helpers/judgeByStaticAnalysis.cjs"),p=require("../helpers/runCustomRunner.cjs"),m=require("../helpers/readProblemMarkdownFrontMatter.cjs"),h=require("../helpers/spawnSyncWithTimeout.cjs");let g=require("node:path");g=e.__toESM(g,1);let _=require("zod"),v=require("node:child_process");v=e.__toESM(v,1);let y=require("node:os");y=e.__toESM(y,1);let b=require("node:fs/promises");b=e.__toESM(b,1);let x=require("node:timers/promises");const S=.3,C=.2,w=[y.default.platform()===`darwin`?`gtime`:`/usr/bin/time`,`--format`,`%e %M`],T=_.z.object({language:_.z.union([_.z.string(),_.z.array(_.z.string())]).optional()});async function E(e,a){let o=i.parseArgs(process.argv);if(!o.cwd)throw Error(`cwd argument required`);let h=T.parse(o.params),g=o.cwd;n.makeAccessibleToSandboxUser(g);let _=await m.readProblemMarkdownFrontMatter(e),v=await(a.readTestCases??j)(e),y=v.length>0?v:[{id:`default`}],b=y[0]?.id??`prebuild`,x=await f.judgeByStaticAnalysis(o.cwd,_);if(x){t.printTestCaseResult({testCaseId:b,...x});return}let C=a.mainFilePath??await u.findEntryPointFile(o.cwd,h.language);if(!C){t.printTestCaseResult({testCaseId:b,decisionCode:s.DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,stderr:a.mainFilePath?`required main file not found: ${a.mainFilePath}`:`main file not found${h.language?`: language: ${h.language}`:``}`});return}let w=d.findLanguageDefinitionByPath(C);if(!w){t.printTestCaseResult({testCaseId:b,decisionCode:s.DecisionCode.WRONG_ANSWER,stderr:`unsupported language`});return}let E={...process.env,CI:``,FORCE_COLOR:`0`},N=await D({cwd:o.cwd,language:h.language,configuredMainFilePath:a.mainFilePath});if(w.prebuild)try{await w.prebuild(o.cwd);let e=await D({cwd:o.cwd,language:h.language??O(C),configuredMainFilePath:a.mainFilePath,allowConfiguredPathFallback:!0});e&&(N=e)}catch(e){t.printTestCaseResult({testCaseId:b,decisionCode:s.DecisionCode.BUILD_ERROR,stderr:V(e)});return}if(!N){t.printTestCaseResult({testCaseId:b,decisionCode:s.DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,stderr:a.mainFilePath?`required main file not found: ${a.mainFilePath}`:`main file not found${h.language?`: language: ${h.language}`:``}`});return}let P;if(a.prepare)try{P=await p.runCustomRunner(()=>a.prepare?.({cwd:g,env:{...E,...n.getSandboxUserEnvOverrides(E)},mainFilePath:N,problemMarkdownFrontMatter:_}))}catch(e){t.printTestCaseResult({testCaseId:b,decisionCode:s.DecisionCode.BUILD_ERROR,stderr:V(e)});return}let F=P??k({cwd:o.cwd,env:E,mainFilePath:N,languageDefinition:w});if(F){t.printTestCaseResult({testCaseId:b,decisionCode:F.decisionCode??s.DecisionCode.BUILD_ERROR,feedbackMarkdown:F.feedbackMarkdown,stderr:F.stderr,stdout:F.stdout,outputFiles:F.outputFiles});return}let L=await c.snapshotWorkingDirectory(o.cwd),R,z=b,B;try{R=a.runCommand?void 0:await I();let e=v.shared?.fileInputPath;for(let i of y){z=i.id,e&&await l.copyTestCaseFileInput(e,o.cwd),i.fileInputPath&&await l.copyTestCaseFileInput(i.fileInputPath,o.cwd);let u=typeof _.timeLimitMs==`number`?_.timeLimitMs/1e3:a.runTimeoutSeconds??5,d=R?{...E,DISPLAY:R.display}:E,f=await a.resolveInput?.({testCase:i,cwd:o.cwd,env:d})??i.input??``;B=f;let m=await a.command?.({testCase:i,cwd:o.cwd,env:d,mainFilePath:N})??w.command(N),h;try{h=a.runCommand?await p.runCustomRunner(()=>a.runCommand?.({testCase:i,command:n.wrapCommandWithSandboxUser(m),stdin:f,cwd:g,env:{...d,...n.getSandboxUserEnvOverrides(d)},timeLimitSeconds:u,screenshotWaitSeconds:a.screenshotWaitSeconds??S,stopDetectionThreshold:a.stopDetectionThreshold??5})):await M({command:m,stdin:f,cwd:o.cwd,env:d,timeLimitSeconds:u,screenshotWaitSeconds:a.screenshotWaitSeconds??S,stopDetectionThreshold:a.stopDetectionThreshold??5})}catch(e){t.printTestCaseResult({testCaseId:i.id,decisionCode:s.DecisionCode.RUNTIME_ERROR,stdin:f,stderr:V(e)}),await c.cleanWorkingDirectory(o.cwd,L);return}let v=await r.readOutputFiles(o.cwd,_.requiredOutputFilePaths??[]),y={timeLimitSeconds:u,problemMarkdownFrontMatter:{memoryLimitByte:_.memoryLimitByte,requiredOutputFilePaths:_.requiredOutputFilePaths}},b=A({runResult:h,outputFiles:v,context:y}),x=b;if(b.decisionCode===s.DecisionCode.ACCEPTED)try{let e=await a.test({testCase:i,runResult:h,outputFiles:v,cwd:o.cwd,context:y});x={decisionCode:e.decisionCode??b.decisionCode,feedbackMarkdown:e.feedbackMarkdown,stderr:e.stderr,stdout:e.stdout,outputFiles:e.outputFiles}}catch(e){x={decisionCode:s.DecisionCode.RUNTIME_ERROR,stderr:V(e)}}let C=x.decisionCode??s.DecisionCode.ACCEPTED,T=x.stdout??h.stdout,D=x.stderr??h.stderr;if(t.printTestCaseResult({testCaseId:i.id,decisionCode:C,exitStatus:h.status,stdin:h.stdin||void 0,stdout:T||void 0,stderr:D||void 0,timeSeconds:h.timeSeconds,memoryBytes:h.memoryBytes,feedbackMarkdown:x.feedbackMarkdown,outputFiles:x.outputFiles??(v.length>0?v:void 0)}),await c.cleanWorkingDirectory(o.cwd,L),C!==s.DecisionCode.ACCEPTED)break}}catch(e){t.printTestCaseResult({testCaseId:z,decisionCode:s.DecisionCode.RUNTIME_ERROR,stdin:B,stderr:V(e)}),await c.cleanWorkingDirectory(o.cwd,L)}finally{try{n.sandboxUserName&&n.killSandboxUserProcesses()}finally{await R?.dispose()}}}async function D(e){if(e.configuredMainFilePath){if(await B(g.default.join(e.cwd,e.configuredMainFilePath)))return e.configuredMainFilePath;if(!e.allowConfiguredPathFallback)return}return await u.findEntryPointFile(e.cwd,e.language)}function O(e){let t=Object.entries(o.languageIdToDefinition).filter(([,t])=>t.fileExtensions.some(t=>e.endsWith(t))).map(([e])=>e);return t.length>0?t:void 0}function k(e){let t=e.languageDefinition.buildCommand?.(e.mainFilePath);if(!t)return;let n=h.spawnSyncWithTimeout(t[0],t.slice(1),{cwd:e.cwd,encoding:`utf8`,env:e.env},10);if(n.timeSeconds>10)return{decisionCode:s.DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,stderr:n.stderr||void 0};if(n.status!==0)return{decisionCode:s.DecisionCode.BUILD_ERROR,stderr:n.stderr||n.stdout||void 0}}function A(e){if(e.runResult.stopReason===`timeout`)return{decisionCode:s.DecisionCode.TIME_LIMIT_EXCEEDED,stderr:e.runResult.stderr};if(e.runResult.status!==0)return{decisionCode:s.DecisionCode.RUNTIME_ERROR,stderr:e.runResult.stderr};if(e.runResult.memoryBytes>(e.context.problemMarkdownFrontMatter.memoryLimitByte??1/0))return{decisionCode:s.DecisionCode.MEMORY_LIMIT_EXCEEDED,stderr:e.runResult.stderr};let t=e.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length??0;return e.outputFiles.length<t?{decisionCode:s.DecisionCode.MISSING_REQUIRED_OUTPUT_FILE_ERROR}:{decisionCode:s.DecisionCode.ACCEPTED}}async function j(e){return await a.readTestCases(g.default.join(e,`test_cases`))}async function M(e){let t=n.wrapCommandWithSandboxUser([n.TIMEOUT_COMMAND,e.timeLimitSeconds.toFixed(3),...w,...e.command]),r=n.startSandboxTimeoutWatchdog(e.timeLimitSeconds),i;try{i=v.default.spawn(t[0],t.slice(1),{cwd:e.cwd,env:{...e.env,...n.getSandboxUserEnvOverrides(e.env)},detached:process.platform!==`win32`,stdio:[`pipe`,`pipe`,`pipe`]})}catch(e){throw r.cancel(),e}i.stdout.setEncoding(`utf8`),i.stderr.setEncoding(`utf8`);let a=``,o=``,s,c,l=`process_exit`,u=!1,d=[],f=[],p=Date.now()/1e3,m=0;i.stdout.on(`data`,e=>{a+=e}),i.stderr.on(`data`,e=>{o+=e}),i.on(`error`,e=>{c=e,s=1}),i.stdin.on(`error`,()=>{}),i.on(`close`,(e,t)=>{if(e===124){l=`timeout`,s=0;return}if(t){s=1,o=o?`${o}\nprocess terminated by signal: ${t}`:`process terminated by signal: ${t}`;return}s=e??1});try{for(e.stdin&&i.stdin.write(e.stdin),i.stdin.end();s===void 0;){if(await(0,x.setTimeout)(e.screenshotWaitSeconds*1e3),m=Math.max(m,R(i.pid)),f=N(e.env.DISPLAY).toSorted((e,t)=>e.data.length-t.data.length),f.length>0){let t=f.map(e=>e.data).toSorted();if(d.unshift(t),d.length=Math.min(d.length,e.stopDetectionThreshold),d.length===e.stopDetectionThreshold&&d.every(e=>e.length===t.length&&e.every((e,n)=>e===t[n]))){l=`stable_screenshot`,s=0;break}}if(Date.now()/1e3-p>e.timeLimitSeconds){l=`timeout`,s=0;break}}l!==`process_exit`&&i.exitCode===null&&(i.removeAllListeners(`close`),i.removeAllListeners(`error`));let t=l===`process_exit`&&r.fired();if(await L(i),u=!0,c)throw c;let{memoryBytes:n,stderr:h,timeSeconds:g}=F(o,p,m);return{stdin:e.stdin,stdout:a.trimEnd(),stderr:h,status:t?0:s,timeSeconds:g,memoryBytes:n,screenshots:f,stopReason:t?`timeout`:l}}finally{u&&r.cancel()}}function N(e){let t=n.getTrustedHelperEnv(e?{DISPLAY:e}:{DISPLAY:process.env.DISPLAY}),r=v.default.spawnSync(`xwininfo`,[`-root`,`-tree`],{encoding:`utf8`,env:t});if(r.error)throw r.error;if(r.status!==0||!r.stdout)return[];let i=[];for(let e of P(r.stdout)){let n=v.default.spawnSync(`maim`,[`-i`,e],{env:t});if(n.error)throw n.error;if(n.status!==0||n.stdout.length===0)continue;let r=v.default.spawnSync(`xdotool`,[`getwindowname`,e],{encoding:`utf8`,env:t});if(r.error)throw r.error;let a=r.stdout.trim().replaceAll(/[\s/]/g,`_`);i.push({path:`${a||`window`}_${e}.png`,data:n.stdout.toString(`base64`),encoding:`base64`})}return i}function P(e){let t=[],n=e.split(`
-`);for(let e of n){if(e.includes(`Root window id:`)||e.includes(`Parent window id:`)||e.includes(`()`))continue;let n=/^\s{5}(0x[\da-f]+) /.exec(e);n?.[1]&&t.push(Number.parseInt(n[1],16).toString())}return t}function F(e,t,n){let r=/(?:^|\n)(\d+\.\d+) (\d+)\s*$/.exec(e),i=r?e.slice(0,r.index).trimEnd():e.trimEnd(),a=Number(r?.[2])*1024||0;return{stderr:i,timeSeconds:Number(r?.[1])||Date.now()/1e3-t,memoryBytes:Math.max(a,n)}}async function I(){if(process.platform!==`linux`)throw Error(`GUI screenshot capture is supported only on Linux.`);for(let e=90;e<100;e++){let t=`:${e}`,r,i=v.default.spawn(`Xvfb`,[t,`-screen`,`0`,`1280x1024x24`,`-ac`],{stdio:`ignore`,env:n.getTrustedHelperEnv()});if(i.on(`error`,e=>{r=e}),await(0,x.setTimeout)(300),r)throw r;if(i.exitCode===null)return{display:t,dispose:async()=>{i.killed||(i.kill(`SIGTERM`),await(0,x.setTimeout)(100),i.exitCode===null&&i.kill(`SIGKILL`))}}}throw Error(`Xvfb could not be started.`)}async function L(e){if(e.pid){if(n.sandboxUserName){n.killSandboxUserProcesses([`TERM`]),await(0,x.setTimeout)(C*1e3),n.killSandboxUserProcesses([`KILL`]);return}z(e.pid,`SIGTERM`),await(0,x.setTimeout)(C*1e3),e.exitCode===null&&z(e.pid,`SIGKILL`)}}function R(e){if(!e||process.platform!==`linux`)return 0;let t=v.default.spawnSync(`ps`,[`-o`,`rss=`,`--no-headers`,`--pgroup`,String(e)],{encoding:`utf8`,env:n.getTrustedHelperEnv()});return t.error||t.status!==0||!t.stdout?0:t.stdout.split(`
-`).map(e=>Number(e.trim())).filter(e=>Number.isFinite(e)&&e>0).reduce((e,t)=>e+t*1024,0)}function z(e,t){try{if(process.platform!==`win32`){process.kill(-e,t);return}}catch{}try{process.kill(e,t)}catch{}}async function B(e){try{return await b.default.access(e),!0}catch{return!1}}function V(e){return e instanceof Error?e.message:String(e)}exports.guiCommandJudgePreset=E;
+"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:`Module`});const e=require("../_virtual/_rolldown/runtime.cjs"),t=require("../helpers/printTestCaseResult.cjs"),n=require("../helpers/parseArgs.cjs"),r=require("../helpers/readTestCases.cjs"),i=require("../types/language.cjs"),a=require("../types/decisionCode.cjs"),o=require("../helpers/cleanWorkingDirectory.cjs"),s=require("../helpers/copyTestCaseFileInput.cjs"),c=require("../helpers/findEntryPointFile.cjs"),l=require("../helpers/findLanguageDefinitionByPath.cjs"),u=require("../helpers/judgeByStaticAnalysis.cjs"),d=require("../helpers/readOutputFiles.cjs"),f=require("../helpers/readProblemMarkdownFrontMatter.cjs"),p=require("../helpers/spawnSyncWithTimeout.cjs");let m=require("node:path");m=e.__toESM(m,1);let h=require("zod"),g=require("node:fs/promises");g=e.__toESM(g,1);let _=require("node:child_process");_=e.__toESM(_,1);let v=require("node:os");v=e.__toESM(v,1);let y=require("node:timers/promises");const b=.3,x=[v.default.platform()===`darwin`?`gtime`:`/usr/bin/time`,`--format`,`%e %M`],S=h.z.object({language:h.z.union([h.z.string(),h.z.array(h.z.string())]).optional()});async function C(e,r){let i=n.parseArgs(process.argv);if(!i.cwd)throw Error(`cwd argument required`);let p=S.parse(i.params),m=i.cwd,h=await f.readProblemMarkdownFrontMatter(e),g=await(r.readTestCases??O)(e),_=g.length>0?g:[{id:`default`}],v=_[0]?.id??`prebuild`,y=await u.judgeByStaticAnalysis(i.cwd,h);if(y){t.printTestCaseResult({testCaseId:v,...y});return}let x=r.mainFilePath??await c.findEntryPointFile(i.cwd,p.language);if(!x){t.printTestCaseResult({testCaseId:v,decisionCode:a.DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,stderr:r.mainFilePath?`required main file not found: ${r.mainFilePath}`:`main file not found${p.language?`: language: ${p.language}`:``}`});return}let C=l.findLanguageDefinitionByPath(x);if(!C){t.printTestCaseResult({testCaseId:v,decisionCode:a.DecisionCode.WRONG_ANSWER,stderr:`unsupported language`});return}let A={...process.env,CI:``,FORCE_COLOR:`0`},j=await w({cwd:i.cwd,language:p.language,configuredMainFilePath:r.mainFilePath});if(C.prebuild)try{await C.prebuild(i.cwd);let e=await w({cwd:i.cwd,language:p.language??T(x),configuredMainFilePath:r.mainFilePath,allowConfiguredPathFallback:!0});e&&(j=e)}catch(e){t.printTestCaseResult({testCaseId:v,decisionCode:a.DecisionCode.BUILD_ERROR,stderr:R(e)});return}if(!j){t.printTestCaseResult({testCaseId:v,decisionCode:a.DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,stderr:r.mainFilePath?`required main file not found: ${r.mainFilePath}`:`main file not found${p.language?`: language: ${p.language}`:``}`});return}let M;if(r.prepare)try{M=await r.prepare({cwd:m,env:A,mainFilePath:j,problemMarkdownFrontMatter:h})}catch(e){t.printTestCaseResult({testCaseId:v,decisionCode:a.DecisionCode.BUILD_ERROR,stderr:R(e)});return}let P=M??E({cwd:i.cwd,env:A,mainFilePath:j,languageDefinition:C});if(P){t.printTestCaseResult({testCaseId:v,decisionCode:P.decisionCode??a.DecisionCode.BUILD_ERROR,feedbackMarkdown:P.feedbackMarkdown,stderr:P.stderr,stdout:P.stdout,outputFiles:P.outputFiles});return}let F=await o.snapshotWorkingDirectory(i.cwd),I,L=v,z;try{I=r.runCommand?void 0:await N();let e=g.shared?.fileInputPath;for(let n of _){L=n.id,e&&await s.copyTestCaseFileInput(e,i.cwd),n.fileInputPath&&await s.copyTestCaseFileInput(n.fileInputPath,i.cwd);let c=typeof h.timeLimitMs==`number`?h.timeLimitMs/1e3:r.runTimeoutSeconds??5,l=I?{...A,DISPLAY:I.display}:A,u=await r.resolveInput?.({testCase:n,cwd:i.cwd,env:l})??n.input??``;z=u;let f=await r.command?.({testCase:n,cwd:i.cwd,env:l,mainFilePath:j})??C.command(j),p;try{p=r.runCommand?await r.runCommand({testCase:n,command:f,stdin:u,cwd:m,env:l,timeLimitSeconds:c,screenshotWaitSeconds:r.screenshotWaitSeconds??b,stopDetectionThreshold:r.stopDetectionThreshold??5}):await k({command:f,stdin:u,cwd:i.cwd,env:l,timeLimitSeconds:c,screenshotWaitSeconds:r.screenshotWaitSeconds??b,stopDetectionThreshold:r.stopDetectionThreshold??5})}catch(e){t.printTestCaseResult({testCaseId:n.id,decisionCode:a.DecisionCode.RUNTIME_ERROR,stdin:u,stderr:R(e)}),await o.cleanWorkingDirectory(i.cwd,F);return}let g=await d.readOutputFiles(i.cwd,h.requiredOutputFilePaths??[]),_={timeLimitSeconds:c,problemMarkdownFrontMatter:{memoryLimitByte:h.memoryLimitByte,requiredOutputFilePaths:h.requiredOutputFilePaths}},v=D({runResult:p,outputFiles:g,context:_}),y=v;if(v.decisionCode===a.DecisionCode.ACCEPTED)try{let e=await r.test({testCase:n,runResult:p,outputFiles:g,cwd:i.cwd,context:_});y={decisionCode:e.decisionCode??v.decisionCode,feedbackMarkdown:e.feedbackMarkdown,stderr:e.stderr,stdout:e.stdout,outputFiles:e.outputFiles}}catch(e){y={decisionCode:a.DecisionCode.RUNTIME_ERROR,stderr:R(e)}}let x=y.decisionCode??a.DecisionCode.ACCEPTED,S=y.stdout??p.stdout,w=y.stderr??p.stderr;if(t.printTestCaseResult({testCaseId:n.id,decisionCode:x,exitStatus:p.status,stdin:p.stdin||void 0,stdout:S||void 0,stderr:w||void 0,timeSeconds:p.timeSeconds,memoryBytes:p.memoryBytes,feedbackMarkdown:y.feedbackMarkdown,outputFiles:y.outputFiles??(g.length>0?g:void 0)}),await o.cleanWorkingDirectory(i.cwd,F),x!==a.DecisionCode.ACCEPTED)break}}catch(e){t.printTestCaseResult({testCaseId:L,decisionCode:a.DecisionCode.RUNTIME_ERROR,stdin:z,stderr:R(e)}),await o.cleanWorkingDirectory(i.cwd,F)}finally{await I?.dispose()}}async function w(e){if(e.configuredMainFilePath){if(await L(m.default.join(e.cwd,e.configuredMainFilePath)))return e.configuredMainFilePath;if(!e.allowConfiguredPathFallback)return}return await c.findEntryPointFile(e.cwd,e.language)}function T(e){let t=Object.entries(i.languageIdToDefinition).filter(([,t])=>t.fileExtensions.some(t=>e.endsWith(t))).map(([e])=>e);return t.length>0?t:void 0}function E(e){let t=e.languageDefinition.buildCommand?.(e.mainFilePath);if(!t)return;let n=p.spawnSyncWithTimeout(t[0],t.slice(1),{cwd:e.cwd,encoding:`utf8`,env:e.env},10);if(n.timeSeconds>10)return{decisionCode:a.DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,stderr:n.stderr||void 0};if(n.status!==0)return{decisionCode:a.DecisionCode.BUILD_ERROR,stderr:n.stderr||n.stdout||void 0}}function D(e){if(e.runResult.stopReason===`timeout`)return{decisionCode:a.DecisionCode.TIME_LIMIT_EXCEEDED,stderr:e.runResult.stderr};if(e.runResult.status!==0)return{decisionCode:a.DecisionCode.RUNTIME_ERROR,stderr:e.runResult.stderr};if(e.runResult.memoryBytes>(e.context.problemMarkdownFrontMatter.memoryLimitByte??1/0))return{decisionCode:a.DecisionCode.MEMORY_LIMIT_EXCEEDED,stderr:e.runResult.stderr};let t=e.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length??0;return e.outputFiles.length<t?{decisionCode:a.DecisionCode.MISSING_REQUIRED_OUTPUT_FILE_ERROR}:{decisionCode:a.DecisionCode.ACCEPTED}}async function O(e){return await r.readTestCases(m.default.join(e,`test_cases`))}async function k(e){let t=_.default.spawn(`timeout`,[e.timeLimitSeconds.toFixed(3),...x,...e.command],{cwd:e.cwd,env:e.env,detached:process.platform!==`win32`,stdio:[`pipe`,`pipe`,`pipe`]});t.stdout.setEncoding(`utf8`),t.stderr.setEncoding(`utf8`);let n=``,r=``,i,a,o=`process_exit`,s=[],c=[],l=Date.now()/1e3,u=0;t.stdout.on(`data`,e=>{n+=e}),t.stderr.on(`data`,e=>{r+=e}),t.on(`error`,e=>{a=e,i=1}),t.stdin.on(`error`,()=>{}),t.on(`close`,(e,t)=>{if(e===124){o=`timeout`,i=0;return}if(t){i=1,r=r?`${r}\nprocess terminated by signal: ${t}`:`process terminated by signal: ${t}`;return}i=e??1});try{for(e.stdin&&t.stdin.write(e.stdin),t.stdin.end();i===void 0;){if(await(0,y.setTimeout)(e.screenshotWaitSeconds*1e3),u=Math.max(u,F(t.pid)),c=A(e.env.DISPLAY).toSorted((e,t)=>e.data.length-t.data.length),c.length>0){let t=c.map(e=>e.data).toSorted();if(s.unshift(t),s.length=Math.min(s.length,e.stopDetectionThreshold),s.length===e.stopDetectionThreshold&&s.every(e=>e.length===t.length&&e.every((e,n)=>e===t[n]))){o=`stable_screenshot`,i=0;break}}if(Date.now()/1e3-l>e.timeLimitSeconds){o=`timeout`,i=0;break}}o!==`process_exit`&&t.exitCode===null&&(t.removeAllListeners(`close`),t.removeAllListeners(`error`))}finally{await P(t)}if(a)throw a;let{memoryBytes:d,stderr:f,timeSeconds:p}=M(r,l,u);return{stdin:e.stdin,stdout:n.trimEnd(),stderr:f,status:i,timeSeconds:p,memoryBytes:d,screenshots:c,stopReason:o}}function A(e){let t={...process.env,DISPLAY:e??process.env.DISPLAY},n=_.default.spawnSync(`xwininfo`,[`-root`,`-tree`],{encoding:`utf8`,env:t});if(n.error)throw n.error;if(n.status!==0||!n.stdout)return[];let r=[];for(let e of j(n.stdout)){let n=_.default.spawnSync(`maim`,[`-i`,e],{env:t});if(n.error)throw n.error;if(n.status!==0||n.stdout.length===0)continue;let i=_.default.spawnSync(`xdotool`,[`getwindowname`,e],{encoding:`utf8`,env:t});if(i.error)throw i.error;let a=i.stdout.trim().replaceAll(/[\s/]/g,`_`);r.push({path:`${a||`window`}_${e}.png`,data:n.stdout.toString(`base64`),encoding:`base64`})}return r}function j(e){let t=[],n=e.split(`
+`);for(let e of n){if(e.includes(`Root window id:`)||e.includes(`Parent window id:`)||e.includes(`()`))continue;let n=/^\s{5}(0x[\da-f]+) /.exec(e);n?.[1]&&t.push(Number.parseInt(n[1],16).toString())}return t}function M(e,t,n){let r=/(?:^|\n)(\d+\.\d+) (\d+)\s*$/.exec(e),i=r?e.slice(0,r.index).trimEnd():e.trimEnd(),a=Number(r?.[2])*1024||0;return{stderr:i,timeSeconds:Number(r?.[1])||Date.now()/1e3-t,memoryBytes:Math.max(a,n)}}async function N(){if(process.platform!==`linux`)throw Error(`GUI screenshot capture is supported only on Linux.`);for(let e=90;e<100;e++){let t=`:${e}`,n,r=_.default.spawn(`Xvfb`,[t,`-screen`,`0`,`1280x1024x24`,`-ac`],{stdio:`ignore`});if(r.on(`error`,e=>{n=e}),await(0,y.setTimeout)(300),n)throw n;if(r.exitCode===null)return{display:t,dispose:async()=>{r.killed||(r.kill(`SIGTERM`),await(0,y.setTimeout)(100),r.exitCode===null&&r.kill(`SIGKILL`))}}}throw Error(`Xvfb could not be started.`)}async function P(e){e.pid&&(I(e.pid,`SIGTERM`),await(0,y.setTimeout)(200),e.exitCode===null&&I(e.pid,`SIGKILL`))}function F(e){if(!e||process.platform!==`linux`)return 0;let t=_.default.spawnSync(`ps`,[`-o`,`rss=`,`--no-headers`,`--pgroup`,String(e)],{encoding:`utf8`});return t.error||t.status!==0||!t.stdout?0:t.stdout.split(`
+`).map(e=>Number(e.trim())).filter(e=>Number.isFinite(e)&&e>0).reduce((e,t)=>e+t*1024,0)}function I(e,t){try{if(process.platform!==`win32`){process.kill(-e,t);return}}catch{}try{process.kill(e,t)}catch{}}async function L(e){try{return await g.default.access(e),!0}catch{return!1}}function R(e){return e instanceof Error?e.message:String(e)}exports.guiCommandJudgePreset=C;
 //# sourceMappingURL=guiCommand.cjs.map

package/dist/presets/guiCommand.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"guiCommand.cjs","names":["os","z","parseArgs","readProblemMarkdownFrontMatter","judgeByStaticAnalysis","findEntryPointFile","DecisionCode","findLanguageDefinitionByPath","runCustomRunner","getSandboxUserEnvOverrides","snapshotWorkingDirectory","copyTestCaseFileInput","wrapCommandWithSandboxUser","cleanWorkingDirectory","readOutputFiles","sandboxUserName","path","languageIdToDefinition","spawnSyncWithTimeout","readFileTestCases","TIMEOUT_COMMAND","startSandboxTimeoutWatchdog","childProcess","wait","getTrustedHelperEnv","fs"],"sources":["../../src/presets/guiCommand.ts"],"sourcesContent":["import childProcess from 'node:child_process';\nimport fs from 'node:fs/promises';\nimport os from 'node:os';\nimport path from 'node:path';\nimport { setTimeout as wait } from 'node:timers/promises';\n\nimport { z } from 'zod';\n\nimport { cleanWorkingDirectory, snapshotWorkingDirectory } from '../helpers/cleanWorkingDirectory.js';\nimport { copyTestCaseFileInput } from '../helpers/copyTestCaseFileInput.js';\nimport { findEntryPointFile } from '../helpers/findEntryPointFile.js';\nimport { findLanguageDefinitionByPath } from '../helpers/findLanguageDefinitionByPath.js';\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { readOutputFiles } from '../helpers/readOutputFiles.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { readTestCases as readFileTestCases } from '../helpers/readTestCases.js';\nimport { runCustomRunner } from '../helpers/runCustomRunner.js';\nimport {\n  getSandboxUserEnvOverrides,\n  getTrustedHelperEnv,\n  killSandboxUserProcesses,\n  makeAccessibleToSandboxUser,\n  sandboxUserName,\n  startSandboxTimeoutWatchdog,\n  TIMEOUT_COMMAND,\n  wrapCommandWithSandboxUser,\n} from '../helpers/sandboxUser.js';\nimport { spawnSyncWithTimeout } from '../helpers/spawnSyncWithTimeout.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport { languageIdToDefinition } from '../types/language.js';\nimport type { ProblemMarkdownFrontMatter } from '../types/problem.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst BUILD_TIMEOUT_SECONDS = 10;\nconst JUDGE_DEFAULT_TIMEOUT_SECONDS = 5;\nconst SCREENSHOT_WAIT_SECONDS = 0.3;\nconst XVFB_STARTUP_WAIT_SECONDS = 0.3;\nconst XVFB_SHUTDOWN_WAIT_SECONDS = 0.1;\nconst PROCESS_SHUTDOWN_WAIT_SECONDS = 0.2;\nconst STOP_DETECTION_THRESHOLD = 5;\nconst TIME_COMMAND = [os.platform() === 'darwin' ? 'gtime' : '/usr/bin/time', '--format', '%e %M'] as const;\n\nconst judgeParamsSchema = z.object({\n  language: z.union([z.string(), z.array(z.string())]).optional(),\n});\n\n/** What every GUI test case needs; a custom `readTestCases` may add any fields of its own. */\ninterface BaseGuiTestCase {\n  id: string;\n  /** Standard input (`test_cases/<id>.in`). */\n  input?: string;\n  /** Directory copied into the working directory before the run (`test_cases/<id>.fin/`). */\n  fileInputPath?: string;\n}\n\n/** A test case read from `test_cases/` by the default reader. */\nexport interface GuiTestCase extends BaseGuiTestCase {\n  /** Expected standard output (`test_cases/<id>.out`), for the problem's `test` to compare. */\n  output?: string;\n  /** Directory of expected output files (`test_cases/<id>.fout/`), for the problem's `test` to compare. */\n  fileOutputPath?: string;\n}\n\nexport interface GuiScreenshotFile {\n  path: string;\n  data: string;\n  encoding: 'base64';\n}\n\nexport interface GuiCommandRunResult {\n  stdin: string;\n  stdout: string;\n  stderr: string;\n  status: number | undefined;\n  timeSeconds: number;\n  memoryBytes: number;\n  screenshots: GuiScreenshotFile[];\n  stopReason: 'process_exit' | 'stable_screenshot' | 'timeout';\n}\n\ninterface GuiJudgeContext {\n  timeLimitSeconds: number;\n  problemMarkdownFrontMatter: Pick<ProblemMarkdownFrontMatter, 'memoryLimitByte' | 'requiredOutputFilePaths'>;\n}\n\ntype GuiJudgeCaseResult = Pick<\n  TestCaseResult,\n  'decisionCode' | 'feedbackMarkdown' | 'stderr' | 'stdout' | 'outputFiles'\n>;\n\nexport interface GuiCommandJudgePresetOptions<TTestCase extends BaseGuiTestCase = GuiTestCase> {\n  mainFilePath?: string;\n  runTimeoutSeconds?: number;\n  screenshotWaitSeconds?: number;\n  stopDetectionThreshold?: number;\n  readTestCases?: (problemDir: string) => Promise<readonly TTestCase[]>;\n  prepare?: (context: {\n    cwd: string;\n    /**\n     * Build the submission with it. Under `EXERCODE_SANDBOX_USER` delegation it already carries the\n     * sandbox user's overrides, but the handler must still wrap whatever it spawns with\n     * `wrapCommandWithSandboxUser` (both exported): a build runs the submission's own scripts\n     * (`package.json` lifecycle scripts, `build.gradle`, `build.rs`), which as the trusted harness\n     * user could read the problem's test cases and rewrite the harness. The preset terminates\n     * leftover sandbox processes after the handler returns.\n     */\n    env: NodeJS.ProcessEnv;\n    mainFilePath: string;\n    problemMarkdownFrontMatter: ProblemMarkdownFrontMatter;\n  }) => Promise<Partial<GuiJudgeCaseResult> | undefined> | Partial<GuiJudgeCaseResult> | undefined;\n  /**\n   * Runs as the trusted harness user with the submission's `cwd`. Fixture files it creates there\n   * must be written with `createDirectoryWithoutFollowingSymlinks`/`writeFileWithoutFollowingSymlinks`\n   * (both exported): a submission can plant a symlink at a fixture path, and a plain `fs.writeFile`\n   * would follow it into a file only the harness can write.\n   */\n  resolveInput?: (context: { testCase: TTestCase; cwd: string; env: NodeJS.ProcessEnv }) => Promise<string> | string;\n  command?: (context: {\n    testCase: TTestCase;\n    cwd: string;\n    env: NodeJS.ProcessEnv;\n    mainFilePath: string;\n  }) => Promise<readonly [string, ...string[]]> | readonly [string, ...string[]];\n  runCommand?: (context: {\n    testCase: TTestCase;\n    /**\n     * The command to run. Under `EXERCODE_SANDBOX_USER` delegation it is already wrapped so it\n     * executes as the sandbox user; spawn it as given (with the supplied `env`) instead of\n     * reconstructing it, or the submission runs as the trusted harness user.\n     *\n     * Its direct child is then a root-owned `sudo` whose descendants belong to the sandbox user, so\n     * the handler cannot signal them: enforce `timeLimitSeconds` with `startSandboxTimeoutWatchdog`\n     * and `killSandboxUserProcesses` (both exported) rather than `child.kill()` or an outer\n     * `timeout`. The preset terminates leftover sandbox processes after the handler returns.\n     */\n    command: readonly [string, ...string[]];\n    stdin: string;\n    cwd: string;\n    env: NodeJS.ProcessEnv;\n    timeLimitSeconds: number;\n    screenshotWaitSeconds: number;\n    stopDetectionThreshold: number;\n  }) => Promise<GuiCommandRunResult> | GuiCommandRunResult;\n  test: (context: {\n    testCase: TTestCase;\n    runResult: GuiCommandRunResult;\n    outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n    /** The submission's working directory, e.g. for `compareExpectedOutputFiles(cwd, testCase.fileOutputPath)`. */\n    cwd: string;\n    context: GuiJudgeContext;\n  }) => Promise<Partial<GuiJudgeCaseResult>> | Partial<GuiJudgeCaseResult>;\n}\n\n/**\n * A preset function for judging GUI programs by collecting screenshots while the program runs.\n *\n * Keep problem-specific logic in `prepare`, `command`, and `test`.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { DecisionCode } from '@exercode/problem-utils';\n * import { guiCommandJudgePreset } from '@exercode/problem-utils/presets/guiCommand';\n *\n * await guiCommandJudgePreset(import.meta.dirname, {\n *   mainFilePath: 'Main.java',\n *   readTestCases: async () => [{ id: 'default' }],\n *   test: ({ runResult }) => {\n *     return runResult.screenshots.length > 0\n *       ? { decisionCode: DecisionCode.ACCEPTED }\n *       : { decisionCode: DecisionCode.WRONG_ANSWER };\n *   },\n * });\n * ```\n */\nexport async function guiCommandJudgePreset<TTestCase extends BaseGuiTestCase = GuiTestCase>(\n  problemDir: string,\n  options: GuiCommandJudgePresetOptions<TTestCase>\n): Promise<void> {\n  const args = parseArgs(process.argv);\n  if (!args.cwd) throw new Error('cwd argument required');\n  const params = judgeParamsSchema.parse(args.params);\n  const submissionDir = args.cwd;\n\n  // The sandboxed submission must read its sources and write build/run outputs in its directory.\n  makeAccessibleToSandboxUser(submissionDir);\n\n  const problemMarkdownFrontMatter = await readProblemMarkdownFrontMatter(problemDir);\n  const configuredTestCases = await (options.readTestCases ?? readGuiTestCases<TTestCase>)(problemDir);\n  const testCases =\n    configuredTestCases.length > 0 ? configuredTestCases : ([{ id: 'default' }] as unknown as readonly TTestCase[]);\n  const prebuildTestCaseId = testCases[0]?.id ?? 'prebuild';\n\n  const staticAnalysisResult = await judgeByStaticAnalysis(args.cwd, problemMarkdownFrontMatter);\n  if (staticAnalysisResult) {\n    printTestCaseResult({ testCaseId: prebuildTestCaseId, ...staticAnalysisResult });\n    return;\n  }\n\n  const initialMainFilePath = options.mainFilePath ?? (await findEntryPointFile(args.cwd, params.language));\n  if (!initialMainFilePath) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      stderr: options.mainFilePath\n        ? `required main file not found: ${options.mainFilePath}`\n        : `main file not found${params.language ? `: language: ${params.language}` : ''}`,\n    });\n    return;\n  }\n\n  const languageDefinition = findLanguageDefinitionByPath(initialMainFilePath);\n  if (!languageDefinition) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      stderr: 'unsupported language',\n    });\n    return;\n  }\n\n  const env = { ...process.env, CI: '', FORCE_COLOR: '0' };\n\n  let resolvedMainFilePath = await resolveMainFilePath({\n    cwd: args.cwd,\n    language: params.language,\n    configuredMainFilePath: options.mainFilePath,\n  });\n  if (languageDefinition.prebuild) {\n    try {\n      await languageDefinition.prebuild(args.cwd);\n      const prebuiltMainFilePath = await resolveMainFilePath({\n        cwd: args.cwd,\n        language: params.language ?? inferLanguageIdsByPath(initialMainFilePath),\n        configuredMainFilePath: options.mainFilePath,\n        allowConfiguredPathFallback: true,\n      });\n      if (prebuiltMainFilePath) resolvedMainFilePath = prebuiltMainFilePath;\n    } catch (error) {\n      printTestCaseResult({\n        testCaseId: prebuildTestCaseId,\n        decisionCode: DecisionCode.BUILD_ERROR,\n        stderr: errorToMessage(error),\n      });\n      return;\n    }\n  }\n  if (!resolvedMainFilePath) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      stderr: options.mainFilePath\n        ? `required main file not found: ${options.mainFilePath}`\n        : `main file not found${params.language ? `: language: ${params.language}` : ''}`,\n    });\n    return;\n  }\n\n  let customPrepareResult: Partial<GuiJudgeCaseResult> | undefined;\n  if (options.prepare) {\n    try {\n      customPrepareResult = await runCustomRunner(() =>\n        options.prepare?.({\n          cwd: submissionDir,\n          env: { ...env, ...getSandboxUserEnvOverrides(env) },\n          mainFilePath: resolvedMainFilePath,\n          problemMarkdownFrontMatter,\n        })\n      );\n    } catch (error) {\n      // Like the `prebuild` step above: report the failed build rather than letting the throw (from\n      // the handler itself, or from the fail-closed sandbox sweep around it) end the run resultless.\n      printTestCaseResult({\n        testCaseId: prebuildTestCaseId,\n        decisionCode: DecisionCode.BUILD_ERROR,\n        stderr: errorToMessage(error),\n      });\n      return;\n    }\n  }\n  const prepareResult =\n    customPrepareResult ??\n    runDefaultPrepare({\n      cwd: args.cwd,\n      env,\n      mainFilePath: resolvedMainFilePath,\n      languageDefinition,\n    });\n  if (prepareResult) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: prepareResult.decisionCode ?? DecisionCode.BUILD_ERROR,\n      feedbackMarkdown: prepareResult.feedbackMarkdown,\n      stderr: prepareResult.stderr,\n      stdout: prepareResult.stdout,\n      outputFiles: prepareResult.outputFiles,\n    });\n    return;\n  }\n\n  const cwdSnapshot = await snapshotWorkingDirectory(args.cwd);\n  let displayServer: Awaited<ReturnType<typeof ensureDisplayServer>> | undefined;\n  let currentTestCaseId = prebuildTestCaseId;\n  let currentStdin: string | undefined;\n  try {\n    displayServer = options.runCommand ? undefined : await ensureDisplayServer();\n    const sharedFileInputPath = (configuredTestCases as { shared?: { fileInputPath?: string } }).shared?.fileInputPath;\n    for (const testCase of testCases) {\n      currentTestCaseId = testCase.id;\n      if (sharedFileInputPath) await copyTestCaseFileInput(sharedFileInputPath, args.cwd);\n      if (testCase.fileInputPath) await copyTestCaseFileInput(testCase.fileInputPath, args.cwd);\n\n      const timeLimitSeconds =\n        typeof problemMarkdownFrontMatter.timeLimitMs === 'number'\n          ? problemMarkdownFrontMatter.timeLimitMs / 1000\n          : (options.runTimeoutSeconds ?? JUDGE_DEFAULT_TIMEOUT_SECONDS);\n\n      const runEnv = displayServer ? { ...env, DISPLAY: displayServer.display } : env;\n      const stdin = (await options.resolveInput?.({ testCase, cwd: args.cwd, env: runEnv })) ?? testCase.input ?? '';\n      currentStdin = stdin;\n      const command =\n        (await options.command?.({ testCase, cwd: args.cwd, env: runEnv, mainFilePath: resolvedMainFilePath })) ??\n        languageDefinition.command(resolvedMainFilePath);\n\n      let runResult: GuiCommandRunResult;\n      try {\n        runResult = options.runCommand\n          ? await runCustomRunner(\n              () =>\n                // Hand custom runners a command that is already sandbox-wrapped, and the matching\n                // environment: they spawn it themselves, so an unwrapped command would run the\n                // submission as the trusted harness user and defeat the delegation boundary.\n                options.runCommand?.({\n                  testCase,\n                  command: wrapCommandWithSandboxUser(command),\n                  stdin,\n                  cwd: submissionDir,\n                  env: { ...runEnv, ...getSandboxUserEnvOverrides(runEnv) },\n                  timeLimitSeconds,\n                  screenshotWaitSeconds: options.screenshotWaitSeconds ?? SCREENSHOT_WAIT_SECONDS,\n                  stopDetectionThreshold: options.stopDetectionThreshold ?? STOP_DETECTION_THRESHOLD,\n                }) as Promise<GuiCommandRunResult>\n            )\n          : await spawnGuiProgram({\n              command,\n              stdin,\n              cwd: args.cwd,\n              env: runEnv,\n              timeLimitSeconds,\n              screenshotWaitSeconds: options.screenshotWaitSeconds ?? SCREENSHOT_WAIT_SECONDS,\n              stopDetectionThreshold: options.stopDetectionThreshold ?? STOP_DETECTION_THRESHOLD,\n            });\n      } catch (error) {\n        printTestCaseResult({\n          testCaseId: testCase.id,\n          decisionCode: DecisionCode.RUNTIME_ERROR,\n          stdin,\n          stderr: errorToMessage(error),\n        });\n        await cleanWorkingDirectory(args.cwd, cwdSnapshot);\n        return;\n      }\n\n      const outputFiles = await readOutputFiles(args.cwd, problemMarkdownFrontMatter.requiredOutputFilePaths ?? []);\n      const judgeContext: GuiJudgeContext = {\n        timeLimitSeconds,\n        problemMarkdownFrontMatter: {\n          memoryLimitByte: problemMarkdownFrontMatter.memoryLimitByte,\n          requiredOutputFilePaths: problemMarkdownFrontMatter.requiredOutputFilePaths,\n        },\n      };\n      const baseJudgeResult = evaluateGuiRunResult({ runResult, outputFiles, context: judgeContext });\n      let judgeResult = baseJudgeResult;\n      if (baseJudgeResult.decisionCode === DecisionCode.ACCEPTED) {\n        try {\n          const extendedJudgeResult = await options.test({\n            testCase,\n            runResult,\n            outputFiles,\n            cwd: args.cwd,\n            context: judgeContext,\n          });\n          judgeResult = {\n            decisionCode: extendedJudgeResult.decisionCode ?? baseJudgeResult.decisionCode,\n            feedbackMarkdown: extendedJudgeResult.feedbackMarkdown,\n            stderr: extendedJudgeResult.stderr,\n            stdout: extendedJudgeResult.stdout,\n            outputFiles: extendedJudgeResult.outputFiles,\n          };\n        } catch (error) {\n          judgeResult = {\n            decisionCode: DecisionCode.RUNTIME_ERROR,\n            stderr: errorToMessage(error),\n          };\n        }\n      }\n\n      const decisionCode = judgeResult.decisionCode ?? DecisionCode.ACCEPTED;\n      const stdout = judgeResult.stdout ?? runResult.stdout;\n      const stderr = judgeResult.stderr ?? runResult.stderr;\n      printTestCaseResult({\n        testCaseId: testCase.id,\n        decisionCode,\n        exitStatus: runResult.status,\n        stdin: runResult.stdin || undefined,\n        stdout: stdout || undefined,\n        stderr: stderr || undefined,\n        timeSeconds: runResult.timeSeconds,\n        memoryBytes: runResult.memoryBytes,\n        feedbackMarkdown: judgeResult.feedbackMarkdown,\n        outputFiles: judgeResult.outputFiles ?? (outputFiles.length > 0 ? outputFiles : undefined),\n      });\n\n      await cleanWorkingDirectory(args.cwd, cwdSnapshot);\n      if (decisionCode !== DecisionCode.ACCEPTED) break;\n    }\n  } catch (error) {\n    printTestCaseResult({\n      testCaseId: currentTestCaseId,\n      decisionCode: DecisionCode.RUNTIME_ERROR,\n      stdin: currentStdin,\n      stderr: errorToMessage(error),\n    });\n    await cleanWorkingDirectory(args.cwd, cwdSnapshot);\n  } finally {\n    try {\n      // Sweep any sandbox processes the run left behind (e.g. a SIGTERM-ignoring forked child) so\n      // they cannot race the harness's output reads or survive into the next test case/request.\n      if (sandboxUserName) killSandboxUserProcesses();\n    } finally {\n      // Must run even when the sweep fails closed, or Xvfb keeps holding its display number and\n      // later GUI requests exhaust the `:90`–`:99` range.\n      await displayServer?.dispose();\n    }\n  }\n}\n\nasync function resolveMainFilePath(context: {\n  cwd: string;\n  language?: string | string[];\n  configuredMainFilePath?: string;\n  allowConfiguredPathFallback?: boolean;\n}): Promise<string | undefined> {\n  if (context.configuredMainFilePath) {\n    const resolvedPath = path.join(context.cwd, context.configuredMainFilePath);\n    if (await pathExists(resolvedPath)) return context.configuredMainFilePath;\n    if (!context.allowConfiguredPathFallback) return undefined;\n  }\n\n  return await findEntryPointFile(context.cwd, context.language);\n}\n\nfunction inferLanguageIdsByPath(filePath: string): string[] | undefined {\n  const languageIds = Object.entries(languageIdToDefinition)\n    .filter(([, definition]) => definition.fileExtensions.some((ext) => filePath.endsWith(ext)))\n    .map(([languageId]) => languageId);\n  return languageIds.length > 0 ? languageIds : undefined;\n}\n\nfunction runDefaultPrepare(context: {\n  cwd: string;\n  env: NodeJS.ProcessEnv;\n  mainFilePath: string;\n  languageDefinition: NonNullable<ReturnType<typeof findLanguageDefinitionByPath>>;\n}): Partial<GuiJudgeCaseResult> | undefined {\n  const buildCommand = context.languageDefinition.buildCommand?.(context.mainFilePath);\n  if (!buildCommand) return undefined;\n\n  const buildResult = spawnSyncWithTimeout(\n    buildCommand[0],\n    buildCommand.slice(1),\n    { cwd: context.cwd, encoding: 'utf8', env: context.env },\n    BUILD_TIMEOUT_SECONDS\n  );\n\n  if (buildResult.timeSeconds > BUILD_TIMEOUT_SECONDS) {\n    return {\n      decisionCode: DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,\n      stderr: buildResult.stderr || undefined,\n    };\n  }\n\n  if (buildResult.status !== 0) {\n    return {\n      decisionCode: DecisionCode.BUILD_ERROR,\n      stderr: buildResult.stderr || buildResult.stdout || undefined,\n    };\n  }\n\n  return undefined;\n}\n\nfunction evaluateGuiRunResult(context: {\n  runResult: GuiCommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  context: GuiJudgeContext;\n}): Partial<GuiJudgeCaseResult> {\n  if (context.runResult.stopReason === 'timeout') {\n    return {\n      decisionCode: DecisionCode.TIME_LIMIT_EXCEEDED,\n      stderr: context.runResult.stderr,\n    };\n  }\n\n  if (context.runResult.status !== 0) {\n    return {\n      decisionCode: DecisionCode.RUNTIME_ERROR,\n      stderr: context.runResult.stderr,\n    };\n  }\n\n  if (\n    context.runResult.memoryBytes >\n    (context.context.problemMarkdownFrontMatter.memoryLimitByte ?? Number.POSITIVE_INFINITY)\n  ) {\n    return {\n      decisionCode: DecisionCode.MEMORY_LIMIT_EXCEEDED,\n      stderr: context.runResult.stderr,\n    };\n  }\n\n  const requiredOutputFilesCount = context.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length ?? 0;\n  if (context.outputFiles.length < requiredOutputFilesCount) {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_OUTPUT_FILE_ERROR,\n    };\n  }\n\n  return { decisionCode: DecisionCode.ACCEPTED };\n}\n\nasync function readGuiTestCases<TTestCase extends BaseGuiTestCase>(problemDir: string): Promise<readonly TTestCase[]> {\n  // The default reader yields the base shape; a narrower TTestCase must come from `options.readTestCases`.\n  return (await readFileTestCases(path.join(problemDir, 'test_cases'))) as unknown as readonly TTestCase[];\n}\n\nasync function spawnGuiProgram(context: {\n  command: readonly [string, ...string[]];\n  stdin: string;\n  cwd: string;\n  env: NodeJS.ProcessEnv;\n  timeLimitSeconds: number;\n  screenshotWaitSeconds: number;\n  stopDetectionThreshold: number;\n}): Promise<GuiCommandRunResult> {\n  const wrappedCommand = wrapCommandWithSandboxUser([\n    TIMEOUT_COMMAND,\n    context.timeLimitSeconds.toFixed(3),\n    ...TIME_COMMAND,\n    ...context.command,\n  ]);\n  // A submission can signal its own sandboxed `timeout` and then wedge the event loop (e.g.\n  // XGrabServer makes the synchronous `xwininfo`/`maim` block), so a harness-owned watchdog\n  // force-kills the sandbox user at the deadline; killing the client also releases its X grab.\n  // Armed before the submission starts: a submission that exhausts the PID cgroup would otherwise\n  // prevent the watchdog from spawning at all.\n  const watchdog = startSandboxTimeoutWatchdog(context.timeLimitSeconds);\n  let child: childProcess.ChildProcessWithoutNullStreams;\n  try {\n    child = childProcess.spawn(wrappedCommand[0], wrappedCommand.slice(1), {\n      cwd: context.cwd,\n      env: { ...context.env, ...getSandboxUserEnvOverrides(context.env) },\n      detached: process.platform !== 'win32',\n      stdio: ['pipe', 'pipe', 'pipe'],\n    });\n  } catch (error) {\n    watchdog.cancel();\n    throw error;\n  }\n\n  child.stdout.setEncoding('utf8');\n  child.stderr.setEncoding('utf8');\n\n  let stdout = '';\n  let stderr = '';\n  let exitCode: number | undefined;\n  let spawnError: Error | undefined;\n  let stopReason: GuiCommandRunResult['stopReason'] = 'process_exit';\n  let submissionStopped = false;\n  const screenshotSignaturesHistory: string[][] = [];\n  let screenshots: GuiScreenshotFile[] = [];\n  const startTimeSeconds = Date.now() / 1000;\n  let sampledMemoryBytes = 0;\n  child.stdout.on('data', (chunk: string) => {\n    stdout += chunk;\n  });\n  child.stderr.on('data', (chunk: string) => {\n    stderr += chunk;\n  });\n  child.on('error', (error) => {\n    spawnError = error;\n    exitCode = 1;\n  });\n  // A child that exits before the input is fully written makes the write fail; without a listener\n  // that EPIPE would be an unhandled stream error terminating the harness mid-run.\n  child.stdin.on('error', () => {\n    // The submission simply stopped reading its input; the exit handling below reports the result.\n  });\n  child.on('close', (code, signal) => {\n    if (code === 124) {\n      stopReason = 'timeout';\n      exitCode = 0;\n      return;\n    }\n    if (signal) {\n      exitCode = 1;\n      stderr = stderr\n        ? `${stderr}\\nprocess terminated by signal: ${signal}`\n        : `process terminated by signal: ${signal}`;\n      return;\n    }\n    exitCode = code ?? 1;\n  });\n\n  // Everything below must run under the `finally` that decides the watchdog's fate: it is cancelled\n  // only once the submission is demonstrably stopped. When a helper throws first — `takeScreenshots`\n  // rethrows spawn failures, which a submission can provoke by exhausting PIDs — the watchdog stays\n  // armed on purpose, since nothing else would end the submission. The cost is that a detached\n  // watchdog can outlive this harness and SIGKILL sandbox processes of a request that starts within\n  // its remaining deadline, which is the same trade-off the package-manager runner accepts.\n  try {\n    if (context.stdin) child.stdin.write(context.stdin);\n    child.stdin.end();\n\n    while (exitCode === undefined) {\n      await wait(context.screenshotWaitSeconds * 1000);\n      sampledMemoryBytes = Math.max(sampledMemoryBytes, readProcessGroupMemoryBytes(child.pid));\n      const currentScreenshots = takeScreenshots(context.env.DISPLAY);\n      screenshots = currentScreenshots.toSorted((a, b) => a.data.length - b.data.length);\n\n      if (screenshots.length > 0) {\n        const screenshotSignatures = screenshots.map((file) => file.data).toSorted();\n        screenshotSignaturesHistory.unshift(screenshotSignatures);\n        screenshotSignaturesHistory.length = Math.min(\n          screenshotSignaturesHistory.length,\n          context.stopDetectionThreshold\n        );\n        if (\n          screenshotSignaturesHistory.length === context.stopDetectionThreshold &&\n          screenshotSignaturesHistory.every(\n            (files) =>\n              files.length === screenshotSignatures.length &&\n              files.every((file, index) => file === screenshotSignatures[index])\n          )\n        ) {\n          stopReason = 'stable_screenshot';\n          exitCode = 0;\n          break;\n        }\n      }\n\n      if (Date.now() / 1000 - startTimeSeconds > context.timeLimitSeconds) {\n        stopReason = 'timeout';\n        exitCode = 0;\n        break;\n      }\n    }\n\n    if (stopReason !== 'process_exit' && child.exitCode === null) {\n      child.removeAllListeners('close');\n      child.removeAllListeners('error');\n    }\n    // A submission can kill its own same-UID `timeout` and be stopped by the watchdog instead,\n    // which surfaces as a SIGKILL rather than a deadline. Report that as the timeout it is, not as\n    // the runtime error the signal would otherwise imply — but only when the child's fate is what\n    // decided the verdict. A `stable_screenshot` (or the loop's own `timeout`) was already decided\n    // from the screenshots, so the watchdog must not overwrite it.\n    const watchdogFired = stopReason === 'process_exit' && watchdog.fired();\n    await stopProcess(child);\n    submissionStopped = true;\n    if (spawnError) throw spawnError;\n    const {\n      memoryBytes,\n      stderr: normalizedStderr,\n      timeSeconds,\n    } = parseTimedStderr(stderr, startTimeSeconds, sampledMemoryBytes);\n\n    return {\n      stdin: context.stdin,\n      stdout: stdout.trimEnd(),\n      stderr: normalizedStderr,\n      status: watchdogFired ? 0 : exitCode,\n      timeSeconds,\n      memoryBytes,\n      screenshots,\n      stopReason: watchdogFired ? 'timeout' : stopReason,\n    };\n  } finally {\n    // Only once the submission is demonstrably stopped: if termination itself failed (or never ran\n    // because a helper threw), the watchdog is the only thing left able to end it, and cancelling\n    // would leave it running unchecked.\n    if (submissionStopped) watchdog.cancel();\n  }\n}\n\nfunction takeScreenshots(display: string | undefined): GuiScreenshotFile[] {\n  // These helpers run as the trusted harness user, so they must not be resolved through a\n  // submission-influenced `PATH` (see `getTrustedHelperEnv`).\n  const env = getTrustedHelperEnv(display ? { DISPLAY: display } : { DISPLAY: process.env.DISPLAY });\n  const xwininfo = childProcess.spawnSync('xwininfo', ['-root', '-tree'], { encoding: 'utf8', env });\n  if (xwininfo.error) throw xwininfo.error;\n  if (xwininfo.status !== 0 || !xwininfo.stdout) return [];\n\n  const screenshots: GuiScreenshotFile[] = [];\n  for (const windowId of extractTopLevelWindowIds(xwininfo.stdout)) {\n    const screenshot = childProcess.spawnSync('maim', ['-i', windowId], { env });\n    if (screenshot.error) throw screenshot.error;\n    if (screenshot.status !== 0 || screenshot.stdout.length === 0) continue;\n\n    const windowNameResult = childProcess.spawnSync('xdotool', ['getwindowname', windowId], { encoding: 'utf8', env });\n    if (windowNameResult.error) throw windowNameResult.error;\n    const windowName = windowNameResult.stdout.trim().replaceAll(/[\\s/]/g, '_');\n\n    screenshots.push({\n      path: `${windowName || 'window'}_${windowId}.png`,\n      data: screenshot.stdout.toString('base64'),\n      encoding: 'base64',\n    });\n  }\n\n  return screenshots;\n}\n\nfunction extractTopLevelWindowIds(stdout: string): string[] {\n  const windowIds: string[] = [];\n  const lines = stdout.split('\\n');\n  for (const line of lines) {\n    if (line.includes('Root window id:') || line.includes('Parent window id:') || line.includes('()')) continue;\n\n    const match = /^\\s{5}(0x[\\da-f]+) /.exec(line);\n    if (!match?.[1]) continue;\n    windowIds.push(Number.parseInt(match[1], 16).toString());\n  }\n  return windowIds;\n}\n\nfunction parseTimedStderr(\n  stderr: string,\n  startTimeSeconds: number,\n  sampledMemoryBytes: number\n): Pick<GuiCommandRunResult, 'stderr' | 'timeSeconds' | 'memoryBytes'> {\n  const match = /(?:^|\\n)(\\d+\\.\\d+) (\\d+)\\s*$/.exec(stderr);\n  const normalizedStderr = match ? stderr.slice(0, match.index).trimEnd() : stderr.trimEnd();\n  const parsedMemoryBytes = Number(match?.[2]) * 1024 || 0;\n  return {\n    stderr: normalizedStderr,\n    timeSeconds: Number(match?.[1]) || Date.now() / 1000 - startTimeSeconds,\n    memoryBytes: Math.max(parsedMemoryBytes, sampledMemoryBytes),\n  };\n}\n\nasync function ensureDisplayServer(): Promise<{ display: string; dispose: () => Promise<void> }> {\n  if (process.platform !== 'linux') {\n    throw new Error('GUI screenshot capture is supported only on Linux.');\n  }\n\n  for (let displayNumber = 90; displayNumber < 100; displayNumber++) {\n    const display = `:${displayNumber}`;\n    let spawnError: Error | undefined;\n    const xvfb = childProcess.spawn('Xvfb', [display, '-screen', '0', '1280x1024x24', '-ac'], {\n      stdio: 'ignore',\n      // Runs as the trusted harness user, so pin the lookup to system paths.\n      env: getTrustedHelperEnv(),\n    });\n    xvfb.on('error', (error) => {\n      spawnError = error;\n    });\n\n    await wait(XVFB_STARTUP_WAIT_SECONDS * 1000);\n    if (spawnError) throw spawnError;\n    if (xvfb.exitCode !== null) continue;\n\n    return {\n      display,\n      dispose: async () => {\n        if (!xvfb.killed) {\n          xvfb.kill('SIGTERM');\n          await wait(XVFB_SHUTDOWN_WAIT_SECONDS * 1000);\n          if (xvfb.exitCode === null) xvfb.kill('SIGKILL');\n        }\n      },\n    };\n  }\n\n  throw new Error('Xvfb could not be started.');\n}\n\nasync function stopProcess(child: childProcess.ChildProcess): Promise<void> {\n  if (!child.pid) return;\n  // The current user cannot signal the root-owned sudo wrapper nor the sandbox user's processes, so\n  // go through sudo. SIGKILL is unconditional after the grace period: `child.exitCode` reflects only\n  // the sudo wrapper, so a daemonized child that ignored SIGTERM would survive if we gated on it.\n  if (sandboxUserName) {\n    killSandboxUserProcesses(['TERM']);\n    await wait(PROCESS_SHUTDOWN_WAIT_SECONDS * 1000);\n    killSandboxUserProcesses(['KILL']);\n    return;\n  }\n  killProcessGroup(child.pid, 'SIGTERM');\n  await wait(PROCESS_SHUTDOWN_WAIT_SECONDS * 1000);\n  if (child.exitCode === null) killProcessGroup(child.pid, 'SIGKILL');\n}\n\nfunction readProcessGroupMemoryBytes(processGroupId: number | undefined): number {\n  if (!processGroupId || process.platform !== 'linux') return 0;\n\n  // The delegation contract requires sudoers `!use_pty`, so sudo execs the command in place and it\n  // stays in the detached child's process group; `--pgroup` therefore still captures it there.\n  const result = childProcess.spawnSync('ps', ['-o', 'rss=', '--no-headers', '--pgroup', String(processGroupId)], {\n    encoding: 'utf8',\n    // Runs as the trusted harness user, so pin the lookup to system paths.\n    env: getTrustedHelperEnv(),\n  });\n  if (result.error || result.status !== 0 || !result.stdout) return 0;\n\n  return result.stdout\n    .split('\\n')\n    .map((line) => Number(line.trim()))\n    .filter((value) => Number.isFinite(value) && value > 0)\n    .reduce((sum, value) => sum + value * 1024, 0);\n}\n\nfunction killProcessGroup(processGroupId: number, signal: NodeJS.Signals): void {\n  try {\n    if (process.platform !== 'win32') {\n      process.kill(-processGroupId, signal);\n      return;\n    }\n  } catch {\n    // The process group may already be gone. Fall back to the direct PID below.\n  }\n\n  try {\n    process.kill(processGroupId, signal);\n  } catch {\n    // The direct child may also already be gone.\n  }\n}\n\nasync function pathExists(filePath: string): Promise<boolean> {\n  try {\n    await fs.access(filePath);\n    return true;\n  } catch {\n    return false;\n  }\n}\n\nfunction errorToMessage(error: unknown): string {\n  return error instanceof Error ? error.message : String(error);\n}\n"],"mappings":"8iCAmCA,MAEM,EAA0B,GAG1B,EAAgC,GAEhC,EAAe,CAACA,EAAAA,QAAG,SAAS,IAAM,SAAW,QAAU,gBAAiB,WAAY,OAAO,EAE3F,EAAoBC,EAAAA,EAAE,OAAO,CACjC,SAAUA,EAAAA,EAAE,MAAM,CAACA,EAAAA,EAAE,OAAO,EAAGA,EAAAA,EAAE,MAAMA,EAAAA,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAChE,CAAC,EAmID,eAAsB,EACpB,EACA,EACe,CACf,IAAM,EAAOC,EAAAA,UAAU,QAAQ,IAAI,EACnC,GAAI,CAAC,EAAK,IAAK,MAAU,MAAM,uBAAuB,EACtD,IAAM,EAAS,EAAkB,MAAM,EAAK,MAAM,EAC5C,EAAgB,EAAK,IAG3B,EAAA,4BAA4B,CAAa,EAEzC,IAAM,EAA6B,MAAMC,EAAAA,+BAA+B,CAAU,EAC5E,EAAsB,MAAO,EAAQ,eAAiB,EAAA,CAA6B,CAAU,EAC7F,EACJ,EAAoB,OAAS,EAAI,EAAuB,CAAC,CAAE,GAAI,SAAU,CAAC,EACtE,EAAqB,EAAU,EAAE,EAAE,IAAM,WAEzC,EAAuB,MAAMC,EAAAA,sBAAsB,EAAK,IAAK,CAA0B,EAC7F,GAAI,EAAsB,CACxB,EAAA,oBAAoB,CAAE,WAAY,EAAoB,GAAG,CAAqB,CAAC,EAC/E,MACF,CAEA,IAAM,EAAsB,EAAQ,cAAiB,MAAMC,EAAAA,mBAAmB,EAAK,IAAK,EAAO,QAAQ,EACvG,GAAI,CAAC,EAAqB,CACxB,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcC,EAAAA,aAAa,uCAC3B,OAAQ,EAAQ,aACZ,iCAAiC,EAAQ,eACzC,sBAAsB,EAAO,SAAW,eAAe,EAAO,WAAa,IACjF,CAAC,EACD,MACF,CAEA,IAAM,EAAqBC,EAAAA,6BAA6B,CAAmB,EAC3E,GAAI,CAAC,EAAoB,CACvB,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcD,EAAAA,aAAa,aAC3B,OAAQ,sBACV,CAAC,EACD,MACF,CAEA,IAAM,EAAM,CAAE,GAAG,QAAQ,IAAK,GAAI,GAAI,YAAa,GAAI,EAEnD,EAAuB,MAAM,EAAoB,CACnD,IAAK,EAAK,IACV,SAAU,EAAO,SACjB,uBAAwB,EAAQ,YAClC,CAAC,EACD,GAAI,EAAmB,SACrB,GAAI,CACF,MAAM,EAAmB,SAAS,EAAK,GAAG,EAC1C,IAAM,EAAuB,MAAM,EAAoB,CACrD,IAAK,EAAK,IACV,SAAU,EAAO,UAAY,EAAuB,CAAmB,EACvE,uBAAwB,EAAQ,aAChC,4BAA6B,EAC/B,CAAC,EACG,IAAsB,EAAuB,EACnD,OAAS,EAAO,CACd,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcA,EAAAA,aAAa,YAC3B,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MACF,CAEF,GAAI,CAAC,EAAsB,CACzB,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcA,EAAAA,aAAa,uCAC3B,OAAQ,EAAQ,aACZ,iCAAiC,EAAQ,eACzC,sBAAsB,EAAO,SAAW,eAAe,EAAO,WAAa,IACjF,CAAC,EACD,MACF,CAEA,IAAI,EACJ,GAAI,EAAQ,QACV,GAAI,CACF,EAAsB,MAAME,EAAAA,oBAC1B,EAAQ,UAAU,CAChB,IAAK,EACL,IAAK,CAAE,GAAG,EAAK,GAAGC,EAAAA,2BAA2B,CAAG,CAAE,EAClD,aAAc,EACd,4BACF,CAAC,CACH,CACF,OAAS,EAAO,CAGd,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcH,EAAAA,aAAa,YAC3B,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MACF,CAEF,IAAM,EACJ,GACA,EAAkB,CAChB,IAAK,EAAK,IACV,MACA,aAAc,EACd,oBACF,CAAC,EACH,GAAI,EAAe,CACjB,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAc,EAAc,cAAgBA,EAAAA,aAAa,YACzD,iBAAkB,EAAc,iBAChC,OAAQ,EAAc,OACtB,OAAQ,EAAc,OACtB,YAAa,EAAc,WAC7B,CAAC,EACD,MACF,CAEA,IAAM,EAAc,MAAMI,EAAAA,yBAAyB,EAAK,GAAG,EACvD,EACA,EAAoB,EACpB,EACJ,GAAI,CACF,EAAgB,EAAQ,WAAa,IAAA,GAAY,MAAM,EAAoB,EAC3E,IAAM,EAAuB,EAAgE,QAAQ,cACrG,IAAK,IAAM,KAAY,EAAW,CAChC,EAAoB,EAAS,GACzB,GAAqB,MAAMC,EAAAA,sBAAsB,EAAqB,EAAK,GAAG,EAC9E,EAAS,eAAe,MAAMA,EAAAA,sBAAsB,EAAS,cAAe,EAAK,GAAG,EAExF,IAAM,EACJ,OAAO,EAA2B,aAAgB,SAC9C,EAA2B,YAAc,IACxC,EAAQ,mBAAqB,EAE9B,EAAS,EAAgB,CAAE,GAAG,EAAK,QAAS,EAAc,OAAQ,EAAI,EACtE,EAAS,MAAM,EAAQ,eAAe,CAAE,WAAU,IAAK,EAAK,IAAK,IAAK,CAAO,CAAC,GAAM,EAAS,OAAS,GAC5G,EAAe,EACf,IAAM,EACH,MAAM,EAAQ,UAAU,CAAE,WAAU,IAAK,EAAK,IAAK,IAAK,EAAQ,aAAc,CAAqB,CAAC,GACrG,EAAmB,QAAQ,CAAoB,EAE7C,EACJ,GAAI,CACF,EAAY,EAAQ,WAChB,MAAMH,EAAAA,oBAKF,EAAQ,aAAa,CACnB,WACA,QAASI,EAAAA,2BAA2B,CAAO,EAC3C,QACA,IAAK,EACL,IAAK,CAAE,GAAG,EAAQ,GAAGH,EAAAA,2BAA2B,CAAM,CAAE,EACxD,mBACA,sBAAuB,EAAQ,uBAAyB,EACxD,uBAAwB,EAAQ,wBAA0B,CAC5D,CAAC,CACL,EACA,MAAM,EAAgB,CACpB,UACA,QACA,IAAK,EAAK,IACV,IAAK,EACL,mBACA,sBAAuB,EAAQ,uBAAyB,EACxD,uBAAwB,EAAQ,wBAA0B,CAC5D,CAAC,CACP,OAAS,EAAO,CACd,EAAA,oBAAoB,CAClB,WAAY,EAAS,GACrB,aAAcH,EAAAA,aAAa,cAC3B,QACA,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MAAMO,EAAAA,sBAAsB,EAAK,IAAK,CAAW,EACjD,MACF,CAEA,IAAM,EAAc,MAAMC,EAAAA,gBAAgB,EAAK,IAAK,EAA2B,yBAA2B,CAAC,CAAC,EACtG,EAAgC,CACpC,mBACA,2BAA4B,CAC1B,gBAAiB,EAA2B,gBAC5C,wBAAyB,EAA2B,uBACtD,CACF,EACM,EAAkB,EAAqB,CAAE,YAAW,cAAa,QAAS,CAAa,CAAC,EAC1F,EAAc,EAClB,GAAI,EAAgB,eAAiBR,EAAAA,aAAa,SAChD,GAAI,CACF,IAAM,EAAsB,MAAM,EAAQ,KAAK,CAC7C,WACA,YACA,cACA,IAAK,EAAK,IACV,QAAS,CACX,CAAC,EACD,EAAc,CACZ,aAAc,EAAoB,cAAgB,EAAgB,aAClE,iBAAkB,EAAoB,iBACtC,OAAQ,EAAoB,OAC5B,OAAQ,EAAoB,OAC5B,YAAa,EAAoB,WACnC,CACF,OAAS,EAAO,CACd,EAAc,CACZ,aAAcA,EAAAA,aAAa,cAC3B,OAAQ,EAAe,CAAK,CAC9B,CACF,CAGF,IAAM,EAAe,EAAY,cAAgBA,EAAAA,aAAa,SACxD,EAAS,EAAY,QAAU,EAAU,OACzC,EAAS,EAAY,QAAU,EAAU,OAe/C,GAdA,EAAA,oBAAoB,CAClB,WAAY,EAAS,GACrB,eACA,WAAY,EAAU,OACtB,MAAO,EAAU,OAAS,IAAA,GAC1B,OAAQ,GAAU,IAAA,GAClB,OAAQ,GAAU,IAAA,GAClB,YAAa,EAAU,YACvB,YAAa,EAAU,YACvB,iBAAkB,EAAY,iBAC9B,YAAa,EAAY,cAAgB,EAAY,OAAS,EAAI,EAAc,IAAA,GAClF,CAAC,EAED,MAAMO,EAAAA,sBAAsB,EAAK,IAAK,CAAW,EAC7C,IAAiBP,EAAAA,aAAa,SAAU,KAC9C,CACF,OAAS,EAAO,CACd,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcA,EAAAA,aAAa,cAC3B,MAAO,EACP,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MAAMO,EAAAA,sBAAsB,EAAK,IAAK,CAAW,CACnD,QAAU,CACR,GAAI,CAGEE,EAAAA,iBAAiB,EAAA,yBAAyB,CAChD,QAAU,CAGR,MAAM,GAAe,QAAQ,CAC/B,CACF,CACF,CAEA,eAAe,EAAoB,EAKH,CAC9B,GAAI,EAAQ,uBAAwB,CAElC,GAAI,MAAM,EADWC,EAAAA,QAAK,KAAK,EAAQ,IAAK,EAAQ,sBACpB,CAAC,EAAG,OAAO,EAAQ,uBACnD,GAAI,CAAC,EAAQ,4BAA6B,MAC5C,CAEA,OAAO,MAAMX,EAAAA,mBAAmB,EAAQ,IAAK,EAAQ,QAAQ,CAC/D,CAEA,SAAS,EAAuB,EAAwC,CACtE,IAAM,EAAc,OAAO,QAAQY,EAAAA,sBAAsB,CAAC,CACvD,QAAQ,EAAG,KAAgB,EAAW,eAAe,KAAM,GAAQ,EAAS,SAAS,CAAG,CAAC,CAAC,CAAC,CAC3F,KAAK,CAAC,KAAgB,CAAU,EACnC,OAAO,EAAY,OAAS,EAAI,EAAc,IAAA,EAChD,CAEA,SAAS,EAAkB,EAKiB,CAC1C,IAAM,EAAe,EAAQ,mBAAmB,eAAe,EAAQ,YAAY,EACnF,GAAI,CAAC,EAAc,OAEnB,IAAM,EAAcC,EAAAA,qBAClB,EAAa,GACb,EAAa,MAAM,CAAC,EACpB,CAAE,IAAK,EAAQ,IAAK,SAAU,OAAQ,IAAK,EAAQ,GAAI,EACvD,EACF,EAEA,GAAI,EAAY,YAAc,GAC5B,MAAO,CACL,aAAcZ,EAAAA,aAAa,0BAC3B,OAAQ,EAAY,QAAU,IAAA,EAChC,EAGF,GAAI,EAAY,SAAW,EACzB,MAAO,CACL,aAAcA,EAAAA,aAAa,YAC3B,OAAQ,EAAY,QAAU,EAAY,QAAU,IAAA,EACtD,CAIJ,CAEA,SAAS,EAAqB,EAIE,CAC9B,GAAI,EAAQ,UAAU,aAAe,UACnC,MAAO,CACL,aAAcA,EAAAA,aAAa,oBAC3B,OAAQ,EAAQ,UAAU,MAC5B,EAGF,GAAI,EAAQ,UAAU,SAAW,EAC/B,MAAO,CACL,aAAcA,EAAAA,aAAa,cAC3B,OAAQ,EAAQ,UAAU,MAC5B,EAGF,GACE,EAAQ,UAAU,aACjB,EAAQ,QAAQ,2BAA2B,iBAAmB,KAE/D,MAAO,CACL,aAAcA,EAAAA,aAAa,sBAC3B,OAAQ,EAAQ,UAAU,MAC5B,EAGF,IAAM,EAA2B,EAAQ,QAAQ,2BAA2B,yBAAyB,QAAU,EAO/G,OANI,EAAQ,YAAY,OAAS,EACxB,CACL,aAAcA,EAAAA,aAAa,kCAC7B,EAGK,CAAE,aAAcA,EAAAA,aAAa,QAAS,CAC/C,CAEA,eAAe,EAAoD,EAAmD,CAEpH,OAAQ,MAAMa,EAAAA,cAAkBH,EAAAA,QAAK,KAAK,EAAY,YAAY,CAAC,CACrE,CAEA,eAAe,EAAgB,EAQE,CAC/B,IAAM,EAAiBJ,EAAAA,2BAA2B,CAChDQ,EAAAA,gBACA,EAAQ,iBAAiB,QAAQ,CAAC,EAClC,GAAG,EACH,GAAG,EAAQ,OACb,CAAC,EAMK,EAAWC,EAAAA,4BAA4B,EAAQ,gBAAgB,EACjE,EACJ,GAAI,CACF,EAAQC,EAAAA,QAAa,MAAM,EAAe,GAAI,EAAe,MAAM,CAAC,EAAG,CACrE,IAAK,EAAQ,IACb,IAAK,CAAE,GAAG,EAAQ,IAAK,GAAGb,EAAAA,2BAA2B,EAAQ,GAAG,CAAE,EAClE,SAAU,QAAQ,WAAa,QAC/B,MAAO,CAAC,OAAQ,OAAQ,MAAM,CAChC,CAAC,CACH,OAAS,EAAO,CAEd,MADA,EAAS,OAAO,EACV,CACR,CAEA,EAAM,OAAO,YAAY,MAAM,EAC/B,EAAM,OAAO,YAAY,MAAM,EAE/B,IAAI,EAAS,GACT,EAAS,GACT,EACA,EACA,EAAgD,eAChD,EAAoB,GAClB,EAA0C,CAAC,EAC7C,EAAmC,CAAC,EAClC,EAAmB,KAAK,IAAI,EAAI,IAClC,EAAqB,EACzB,EAAM,OAAO,GAAG,OAAS,GAAkB,CACzC,GAAU,CACZ,CAAC,EACD,EAAM,OAAO,GAAG,OAAS,GAAkB,CACzC,GAAU,CACZ,CAAC,EACD,EAAM,GAAG,QAAU,GAAU,CAC3B,EAAa,EACb,EAAW,CACb,CAAC,EAGD,EAAM,MAAM,GAAG,YAAe,CAE9B,CAAC,EACD,EAAM,GAAG,SAAU,EAAM,IAAW,CAClC,GAAI,IAAS,IAAK,CAChB,EAAa,UACb,EAAW,EACX,MACF,CACA,GAAI,EAAQ,CACV,EAAW,EACX,EAAS,EACL,GAAG,EAAO,kCAAkC,IAC5C,iCAAiC,IACrC,MACF,CACA,EAAW,GAAQ,CACrB,CAAC,EAQD,GAAI,CAIF,IAHI,EAAQ,OAAO,EAAM,MAAM,MAAM,EAAQ,KAAK,EAClD,EAAM,MAAM,IAAI,EAET,IAAa,IAAA,IAAW,CAM7B,GALA,MAAA,EAAMc,EAAAA,WAAAA,CAAK,EAAQ,sBAAwB,GAAI,EAC/C,EAAqB,KAAK,IAAI,EAAoB,EAA4B,EAAM,GAAG,CAAC,EAExF,EAD2B,EAAgB,EAAQ,IAAI,OACxB,CAAC,CAAC,UAAU,EAAG,IAAM,EAAE,KAAK,OAAS,EAAE,KAAK,MAAM,EAE7E,EAAY,OAAS,EAAG,CAC1B,IAAM,EAAuB,EAAY,IAAK,GAAS,EAAK,IAAI,CAAC,CAAC,SAAS,EAM3E,GALA,EAA4B,QAAQ,CAAoB,EACxD,EAA4B,OAAS,KAAK,IACxC,EAA4B,OAC5B,EAAQ,sBACV,EAEE,EAA4B,SAAW,EAAQ,wBAC/C,EAA4B,MACzB,GACC,EAAM,SAAW,EAAqB,QACtC,EAAM,OAAO,EAAM,IAAU,IAAS,EAAqB,EAAM,CACrE,EACA,CACA,EAAa,oBACb,EAAW,EACX,KACF,CACF,CAEA,GAAI,KAAK,IAAI,EAAI,IAAO,EAAmB,EAAQ,iBAAkB,CACnE,EAAa,UACb,EAAW,EACX,KACF,CACF,CAEI,IAAe,gBAAkB,EAAM,WAAa,OACtD,EAAM,mBAAmB,OAAO,EAChC,EAAM,mBAAmB,OAAO,GAOlC,IAAM,EAAgB,IAAe,gBAAkB,EAAS,MAAM,EAGtE,GAFA,MAAM,EAAY,CAAK,EACvB,EAAoB,GAChB,EAAY,MAAM,EACtB,GAAM,CACJ,cACA,OAAQ,EACR,eACE,EAAiB,EAAQ,EAAkB,CAAkB,EAEjE,MAAO,CACL,MAAO,EAAQ,MACf,OAAQ,EAAO,QAAQ,EACvB,OAAQ,EACR,OAAQ,EAAgB,EAAI,EAC5B,cACA,cACA,cACA,WAAY,EAAgB,UAAY,CAC1C,CACF,QAAU,CAIJ,GAAmB,EAAS,OAAO,CACzC,CACF,CAEA,SAAS,EAAgB,EAAkD,CAGzE,IAAM,EAAMC,EAAAA,oBAAoB,EAAU,CAAE,QAAS,CAAQ,EAAI,CAAE,QAAS,QAAQ,IAAI,OAAQ,CAAC,EAC3F,EAAWF,EAAAA,QAAa,UAAU,WAAY,CAAC,QAAS,OAAO,EAAG,CAAE,SAAU,OAAQ,KAAI,CAAC,EACjG,GAAI,EAAS,MAAO,MAAM,EAAS,MACnC,GAAI,EAAS,SAAW,GAAK,CAAC,EAAS,OAAQ,MAAO,CAAC,EAEvD,IAAM,EAAmC,CAAC,EAC1C,IAAK,IAAM,KAAY,EAAyB,EAAS,MAAM,EAAG,CAChE,IAAM,EAAaA,EAAAA,QAAa,UAAU,OAAQ,CAAC,KAAM,CAAQ,EAAG,CAAE,KAAI,CAAC,EAC3E,GAAI,EAAW,MAAO,MAAM,EAAW,MACvC,GAAI,EAAW,SAAW,GAAK,EAAW,OAAO,SAAW,EAAG,SAE/D,IAAM,EAAmBA,EAAAA,QAAa,UAAU,UAAW,CAAC,gBAAiB,CAAQ,EAAG,CAAE,SAAU,OAAQ,KAAI,CAAC,EACjH,GAAI,EAAiB,MAAO,MAAM,EAAiB,MACnD,IAAM,EAAa,EAAiB,OAAO,KAAK,CAAC,CAAC,WAAW,SAAU,GAAG,EAE1E,EAAY,KAAK,CACf,KAAM,GAAG,GAAc,SAAS,GAAG,EAAS,MAC5C,KAAM,EAAW,OAAO,SAAS,QAAQ,EACzC,SAAU,QACZ,CAAC,CACH,CAEA,OAAO,CACT,CAEA,SAAS,EAAyB,EAA0B,CAC1D,IAAM,EAAsB,CAAC,EACvB,EAAQ,EAAO,MAAM;CAAI,EAC/B,IAAK,IAAM,KAAQ,EAAO,CACxB,GAAI,EAAK,SAAS,iBAAiB,GAAK,EAAK,SAAS,mBAAmB,GAAK,EAAK,SAAS,IAAI,EAAG,SAEnG,IAAM,EAAQ,sBAAsB,KAAK,CAAI,EACxC,IAAQ,IACb,EAAU,KAAK,OAAO,SAAS,EAAM,GAAI,EAAE,CAAC,CAAC,SAAS,CAAC,CACzD,CACA,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACA,EACqE,CACrE,IAAM,EAAQ,+BAA+B,KAAK,CAAM,EAClD,EAAmB,EAAQ,EAAO,MAAM,EAAG,EAAM,KAAK,CAAC,CAAC,QAAQ,EAAI,EAAO,QAAQ,EACnF,EAAoB,OAAO,IAAQ,EAAE,EAAI,MAAQ,EACvD,MAAO,CACL,OAAQ,EACR,YAAa,OAAO,IAAQ,EAAE,GAAK,KAAK,IAAI,EAAI,IAAO,EACvD,YAAa,KAAK,IAAI,EAAmB,CAAkB,CAC7D,CACF,CAEA,eAAe,GAAkF,CAC/F,GAAI,QAAQ,WAAa,QACvB,MAAU,MAAM,oDAAoD,EAGtE,IAAK,IAAI,EAAgB,GAAI,EAAgB,IAAK,IAAiB,CACjE,IAAM,EAAU,IAAI,IAChB,EACE,EAAOA,EAAAA,QAAa,MAAM,OAAQ,CAAC,EAAS,UAAW,IAAK,eAAgB,KAAK,EAAG,CACxF,MAAO,SAEP,IAAKE,EAAAA,oBAAoB,CAC3B,CAAC,EAMD,GALA,EAAK,GAAG,QAAU,GAAU,CAC1B,EAAa,CACf,CAAC,EAED,MAAA,EAAMD,EAAAA,WAAAA,CAAK,GAAgC,EACvC,EAAY,MAAM,EAClB,KAAK,WAAa,KAEtB,MAAO,CACL,UACA,QAAS,SAAY,CACd,EAAK,SACR,EAAK,KAAK,SAAS,EACnB,MAAA,EAAMA,EAAAA,WAAAA,CAAK,GAAiC,EACxC,EAAK,WAAa,MAAM,EAAK,KAAK,SAAS,EAEnD,CACF,CACF,CAEA,MAAU,MAAM,4BAA4B,CAC9C,CAEA,eAAe,EAAY,EAAiD,CACrE,KAAM,IAIX,IAAIR,EAAAA,gBAAiB,CACnB,EAAA,yBAAyB,CAAC,MAAM,CAAC,EACjC,MAAA,EAAMQ,EAAAA,WAAAA,CAAK,EAAgC,GAAI,EAC/C,EAAA,yBAAyB,CAAC,MAAM,CAAC,EACjC,MACF,CACA,EAAiB,EAAM,IAAK,SAAS,EACrC,MAAA,EAAMA,EAAAA,WAAAA,CAAK,EAAgC,GAAI,EAC3C,EAAM,WAAa,MAAM,EAAiB,EAAM,IAAK,SAAS,CAHlE,CAIF,CAEA,SAAS,EAA4B,EAA4C,CAC/E,GAAI,CAAC,GAAkB,QAAQ,WAAa,QAAS,MAAO,GAI5D,IAAM,EAASD,EAAAA,QAAa,UAAU,KAAM,CAAC,KAAM,OAAQ,eAAgB,WAAY,OAAO,CAAc,CAAC,EAAG,CAC9G,SAAU,OAEV,IAAKE,EAAAA,oBAAoB,CAC3B,CAAC,EAGD,OAFI,EAAO,OAAS,EAAO,SAAW,GAAK,CAAC,EAAO,OAAe,EAE3D,EAAO,OACX,MAAM;CAAI,CAAC,CACX,IAAK,GAAS,OAAO,EAAK,KAAK,CAAC,CAAC,CAAC,CAClC,OAAQ,GAAU,OAAO,SAAS,CAAK,GAAK,EAAQ,CAAC,CAAC,CACtD,QAAQ,EAAK,IAAU,EAAM,EAAQ,KAAM,CAAC,CACjD,CAEA,SAAS,EAAiB,EAAwB,EAA8B,CAC9E,GAAI,CACF,GAAI,QAAQ,WAAa,QAAS,CAChC,QAAQ,KAAK,CAAC,EAAgB,CAAM,EACpC,MACF,CACF,MAAQ,CAER,CAEA,GAAI,CACF,QAAQ,KAAK,EAAgB,CAAM,CACrC,MAAQ,CAER,CACF,CAEA,eAAe,EAAW,EAAoC,CAC5D,GAAI,CAEF,OADA,MAAMC,EAAAA,QAAG,OAAO,CAAQ,EACjB,EACT,MAAQ,CACN,MAAO,EACT,CACF,CAEA,SAAS,EAAe,EAAwB,CAC9C,OAAO,aAAiB,MAAQ,EAAM,QAAU,OAAO,CAAK,CAC9D"}
+{"version":3,"file":"guiCommand.cjs","names":["os","z","parseArgs","readProblemMarkdownFrontMatter","judgeByStaticAnalysis","findEntryPointFile","DecisionCode","findLanguageDefinitionByPath","snapshotWorkingDirectory","copyTestCaseFileInput","cleanWorkingDirectory","readOutputFiles","path","languageIdToDefinition","spawnSyncWithTimeout","readFileTestCases","childProcess","wait","fs"],"sources":["../../src/presets/guiCommand.ts"],"sourcesContent":["import childProcess from 'node:child_process';\nimport fs from 'node:fs/promises';\nimport os from 'node:os';\nimport path from 'node:path';\nimport { setTimeout as wait } from 'node:timers/promises';\n\nimport { z } from 'zod';\n\nimport { cleanWorkingDirectory, snapshotWorkingDirectory } from '../helpers/cleanWorkingDirectory.js';\nimport { copyTestCaseFileInput } from '../helpers/copyTestCaseFileInput.js';\nimport { findEntryPointFile } from '../helpers/findEntryPointFile.js';\nimport { findLanguageDefinitionByPath } from '../helpers/findLanguageDefinitionByPath.js';\nimport { judgeByStaticAnalysis } from '../helpers/judgeByStaticAnalysis.js';\nimport { parseArgs } from '../helpers/parseArgs.js';\nimport { printTestCaseResult } from '../helpers/printTestCaseResult.js';\nimport { readOutputFiles } from '../helpers/readOutputFiles.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { readTestCases as readFileTestCases } from '../helpers/readTestCases.js';\nimport { spawnSyncWithTimeout } from '../helpers/spawnSyncWithTimeout.js';\nimport { DecisionCode } from '../types/decisionCode.js';\nimport { languageIdToDefinition } from '../types/language.js';\nimport type { ProblemMarkdownFrontMatter } from '../types/problem.js';\nimport type { TestCaseResult } from '../types/testCaseResult.js';\n\nconst BUILD_TIMEOUT_SECONDS = 10;\nconst JUDGE_DEFAULT_TIMEOUT_SECONDS = 5;\nconst SCREENSHOT_WAIT_SECONDS = 0.3;\nconst XVFB_STARTUP_WAIT_SECONDS = 0.3;\nconst XVFB_SHUTDOWN_WAIT_SECONDS = 0.1;\nconst PROCESS_SHUTDOWN_WAIT_SECONDS = 0.2;\nconst STOP_DETECTION_THRESHOLD = 5;\nconst TIME_COMMAND = [os.platform() === 'darwin' ? 'gtime' : '/usr/bin/time', '--format', '%e %M'] as const;\n\nconst judgeParamsSchema = z.object({\n  language: z.union([z.string(), z.array(z.string())]).optional(),\n});\n\n/** What every GUI test case needs; a custom `readTestCases` may add any fields of its own. */\ninterface BaseGuiTestCase {\n  id: string;\n  /** Standard input (`test_cases/<id>.in`). */\n  input?: string;\n  /** Directory copied into the working directory before the run (`test_cases/<id>.fin/`). */\n  fileInputPath?: string;\n}\n\n/** A test case read from `test_cases/` by the default reader. */\nexport interface GuiTestCase extends BaseGuiTestCase {\n  /** Expected standard output (`test_cases/<id>.out`), for the problem's `test` to compare. */\n  output?: string;\n  /** Directory of expected output files (`test_cases/<id>.fout/`), for the problem's `test` to compare. */\n  fileOutputPath?: string;\n}\n\nexport interface GuiScreenshotFile {\n  path: string;\n  data: string;\n  encoding: 'base64';\n}\n\nexport interface GuiCommandRunResult {\n  stdin: string;\n  stdout: string;\n  stderr: string;\n  status: number | undefined;\n  timeSeconds: number;\n  memoryBytes: number;\n  screenshots: GuiScreenshotFile[];\n  stopReason: 'process_exit' | 'stable_screenshot' | 'timeout';\n}\n\ninterface GuiJudgeContext {\n  timeLimitSeconds: number;\n  problemMarkdownFrontMatter: Pick<ProblemMarkdownFrontMatter, 'memoryLimitByte' | 'requiredOutputFilePaths'>;\n}\n\ntype GuiJudgeCaseResult = Pick<\n  TestCaseResult,\n  'decisionCode' | 'feedbackMarkdown' | 'stderr' | 'stdout' | 'outputFiles'\n>;\n\nexport interface GuiCommandJudgePresetOptions<TTestCase extends BaseGuiTestCase = GuiTestCase> {\n  mainFilePath?: string;\n  runTimeoutSeconds?: number;\n  screenshotWaitSeconds?: number;\n  stopDetectionThreshold?: number;\n  readTestCases?: (problemDir: string) => Promise<readonly TTestCase[]>;\n  prepare?: (context: {\n    cwd: string;\n    env: NodeJS.ProcessEnv;\n    mainFilePath: string;\n    problemMarkdownFrontMatter: ProblemMarkdownFrontMatter;\n  }) => Promise<Partial<GuiJudgeCaseResult> | undefined> | Partial<GuiJudgeCaseResult> | undefined;\n  resolveInput?: (context: { testCase: TTestCase; cwd: string; env: NodeJS.ProcessEnv }) => Promise<string> | string;\n  command?: (context: {\n    testCase: TTestCase;\n    cwd: string;\n    env: NodeJS.ProcessEnv;\n    mainFilePath: string;\n  }) => Promise<readonly [string, ...string[]]> | readonly [string, ...string[]];\n  runCommand?: (context: {\n    testCase: TTestCase;\n    command: readonly [string, ...string[]];\n    stdin: string;\n    cwd: string;\n    env: NodeJS.ProcessEnv;\n    timeLimitSeconds: number;\n    screenshotWaitSeconds: number;\n    stopDetectionThreshold: number;\n  }) => Promise<GuiCommandRunResult> | GuiCommandRunResult;\n  test: (context: {\n    testCase: TTestCase;\n    runResult: GuiCommandRunResult;\n    outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n    /** The submission's working directory, e.g. for `compareExpectedOutputFiles(cwd, testCase.fileOutputPath)`. */\n    cwd: string;\n    context: GuiJudgeContext;\n  }) => Promise<Partial<GuiJudgeCaseResult>> | Partial<GuiJudgeCaseResult>;\n}\n\n/**\n * A preset function for judging GUI programs by collecting screenshots while the program runs.\n *\n * Keep problem-specific logic in `prepare`, `command`, and `test`.\n *\n * @example\n * Create `judge.ts`:\n * ```ts\n * import { DecisionCode } from '@exercode/problem-utils';\n * import { guiCommandJudgePreset } from '@exercode/problem-utils/presets/guiCommand';\n *\n * await guiCommandJudgePreset(import.meta.dirname, {\n *   mainFilePath: 'Main.java',\n *   readTestCases: async () => [{ id: 'default' }],\n *   test: ({ runResult }) => {\n *     return runResult.screenshots.length > 0\n *       ? { decisionCode: DecisionCode.ACCEPTED }\n *       : { decisionCode: DecisionCode.WRONG_ANSWER };\n *   },\n * });\n * ```\n */\nexport async function guiCommandJudgePreset<TTestCase extends BaseGuiTestCase = GuiTestCase>(\n  problemDir: string,\n  options: GuiCommandJudgePresetOptions<TTestCase>\n): Promise<void> {\n  const args = parseArgs(process.argv);\n  if (!args.cwd) throw new Error('cwd argument required');\n  const params = judgeParamsSchema.parse(args.params);\n  const submissionDir = args.cwd;\n\n  const problemMarkdownFrontMatter = await readProblemMarkdownFrontMatter(problemDir);\n  const configuredTestCases = await (options.readTestCases ?? readGuiTestCases<TTestCase>)(problemDir);\n  const testCases =\n    configuredTestCases.length > 0 ? configuredTestCases : ([{ id: 'default' }] as unknown as readonly TTestCase[]);\n  const prebuildTestCaseId = testCases[0]?.id ?? 'prebuild';\n\n  const staticAnalysisResult = await judgeByStaticAnalysis(args.cwd, problemMarkdownFrontMatter);\n  if (staticAnalysisResult) {\n    printTestCaseResult({ testCaseId: prebuildTestCaseId, ...staticAnalysisResult });\n    return;\n  }\n\n  const initialMainFilePath = options.mainFilePath ?? (await findEntryPointFile(args.cwd, params.language));\n  if (!initialMainFilePath) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      stderr: options.mainFilePath\n        ? `required main file not found: ${options.mainFilePath}`\n        : `main file not found${params.language ? `: language: ${params.language}` : ''}`,\n    });\n    return;\n  }\n\n  const languageDefinition = findLanguageDefinitionByPath(initialMainFilePath);\n  if (!languageDefinition) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      stderr: 'unsupported language',\n    });\n    return;\n  }\n\n  const env = { ...process.env, CI: '', FORCE_COLOR: '0' };\n\n  let resolvedMainFilePath = await resolveMainFilePath({\n    cwd: args.cwd,\n    language: params.language,\n    configuredMainFilePath: options.mainFilePath,\n  });\n  if (languageDefinition.prebuild) {\n    try {\n      await languageDefinition.prebuild(args.cwd);\n      const prebuiltMainFilePath = await resolveMainFilePath({\n        cwd: args.cwd,\n        language: params.language ?? inferLanguageIdsByPath(initialMainFilePath),\n        configuredMainFilePath: options.mainFilePath,\n        allowConfiguredPathFallback: true,\n      });\n      if (prebuiltMainFilePath) resolvedMainFilePath = prebuiltMainFilePath;\n    } catch (error) {\n      printTestCaseResult({\n        testCaseId: prebuildTestCaseId,\n        decisionCode: DecisionCode.BUILD_ERROR,\n        stderr: errorToMessage(error),\n      });\n      return;\n    }\n  }\n  if (!resolvedMainFilePath) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      stderr: options.mainFilePath\n        ? `required main file not found: ${options.mainFilePath}`\n        : `main file not found${params.language ? `: language: ${params.language}` : ''}`,\n    });\n    return;\n  }\n\n  let customPrepareResult: Partial<GuiJudgeCaseResult> | undefined;\n  if (options.prepare) {\n    try {\n      customPrepareResult = await options.prepare({\n        cwd: submissionDir,\n        env,\n        mainFilePath: resolvedMainFilePath,\n        problemMarkdownFrontMatter,\n      });\n    } catch (error) {\n      // Like the `prebuild` step above: report the failed build rather than letting the throw end the\n      // run resultless.\n      printTestCaseResult({\n        testCaseId: prebuildTestCaseId,\n        decisionCode: DecisionCode.BUILD_ERROR,\n        stderr: errorToMessage(error),\n      });\n      return;\n    }\n  }\n  const prepareResult =\n    customPrepareResult ??\n    runDefaultPrepare({\n      cwd: args.cwd,\n      env,\n      mainFilePath: resolvedMainFilePath,\n      languageDefinition,\n    });\n  if (prepareResult) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: prepareResult.decisionCode ?? DecisionCode.BUILD_ERROR,\n      feedbackMarkdown: prepareResult.feedbackMarkdown,\n      stderr: prepareResult.stderr,\n      stdout: prepareResult.stdout,\n      outputFiles: prepareResult.outputFiles,\n    });\n    return;\n  }\n\n  const cwdSnapshot = await snapshotWorkingDirectory(args.cwd);\n  let displayServer: Awaited<ReturnType<typeof ensureDisplayServer>> | undefined;\n  let currentTestCaseId = prebuildTestCaseId;\n  let currentStdin: string | undefined;\n  try {\n    displayServer = options.runCommand ? undefined : await ensureDisplayServer();\n    const sharedFileInputPath = (configuredTestCases as { shared?: { fileInputPath?: string } }).shared?.fileInputPath;\n    for (const testCase of testCases) {\n      currentTestCaseId = testCase.id;\n      if (sharedFileInputPath) await copyTestCaseFileInput(sharedFileInputPath, args.cwd);\n      if (testCase.fileInputPath) await copyTestCaseFileInput(testCase.fileInputPath, args.cwd);\n\n      const timeLimitSeconds =\n        typeof problemMarkdownFrontMatter.timeLimitMs === 'number'\n          ? problemMarkdownFrontMatter.timeLimitMs / 1000\n          : (options.runTimeoutSeconds ?? JUDGE_DEFAULT_TIMEOUT_SECONDS);\n\n      const runEnv = displayServer ? { ...env, DISPLAY: displayServer.display } : env;\n      const stdin = (await options.resolveInput?.({ testCase, cwd: args.cwd, env: runEnv })) ?? testCase.input ?? '';\n      currentStdin = stdin;\n      const command =\n        (await options.command?.({ testCase, cwd: args.cwd, env: runEnv, mainFilePath: resolvedMainFilePath })) ??\n        languageDefinition.command(resolvedMainFilePath);\n\n      let runResult: GuiCommandRunResult;\n      try {\n        runResult = options.runCommand\n          ? await options.runCommand({\n              testCase,\n              command,\n              stdin,\n              cwd: submissionDir,\n              env: runEnv,\n              timeLimitSeconds,\n              screenshotWaitSeconds: options.screenshotWaitSeconds ?? SCREENSHOT_WAIT_SECONDS,\n              stopDetectionThreshold: options.stopDetectionThreshold ?? STOP_DETECTION_THRESHOLD,\n            })\n          : await spawnGuiProgram({\n              command,\n              stdin,\n              cwd: args.cwd,\n              env: runEnv,\n              timeLimitSeconds,\n              screenshotWaitSeconds: options.screenshotWaitSeconds ?? SCREENSHOT_WAIT_SECONDS,\n              stopDetectionThreshold: options.stopDetectionThreshold ?? STOP_DETECTION_THRESHOLD,\n            });\n      } catch (error) {\n        printTestCaseResult({\n          testCaseId: testCase.id,\n          decisionCode: DecisionCode.RUNTIME_ERROR,\n          stdin,\n          stderr: errorToMessage(error),\n        });\n        await cleanWorkingDirectory(args.cwd, cwdSnapshot);\n        return;\n      }\n\n      const outputFiles = await readOutputFiles(args.cwd, problemMarkdownFrontMatter.requiredOutputFilePaths ?? []);\n      const judgeContext: GuiJudgeContext = {\n        timeLimitSeconds,\n        problemMarkdownFrontMatter: {\n          memoryLimitByte: problemMarkdownFrontMatter.memoryLimitByte,\n          requiredOutputFilePaths: problemMarkdownFrontMatter.requiredOutputFilePaths,\n        },\n      };\n      const baseJudgeResult = evaluateGuiRunResult({ runResult, outputFiles, context: judgeContext });\n      let judgeResult = baseJudgeResult;\n      if (baseJudgeResult.decisionCode === DecisionCode.ACCEPTED) {\n        try {\n          const extendedJudgeResult = await options.test({\n            testCase,\n            runResult,\n            outputFiles,\n            cwd: args.cwd,\n            context: judgeContext,\n          });\n          judgeResult = {\n            decisionCode: extendedJudgeResult.decisionCode ?? baseJudgeResult.decisionCode,\n            feedbackMarkdown: extendedJudgeResult.feedbackMarkdown,\n            stderr: extendedJudgeResult.stderr,\n            stdout: extendedJudgeResult.stdout,\n            outputFiles: extendedJudgeResult.outputFiles,\n          };\n        } catch (error) {\n          judgeResult = {\n            decisionCode: DecisionCode.RUNTIME_ERROR,\n            stderr: errorToMessage(error),\n          };\n        }\n      }\n\n      const decisionCode = judgeResult.decisionCode ?? DecisionCode.ACCEPTED;\n      const stdout = judgeResult.stdout ?? runResult.stdout;\n      const stderr = judgeResult.stderr ?? runResult.stderr;\n      printTestCaseResult({\n        testCaseId: testCase.id,\n        decisionCode,\n        exitStatus: runResult.status,\n        stdin: runResult.stdin || undefined,\n        stdout: stdout || undefined,\n        stderr: stderr || undefined,\n        timeSeconds: runResult.timeSeconds,\n        memoryBytes: runResult.memoryBytes,\n        feedbackMarkdown: judgeResult.feedbackMarkdown,\n        outputFiles: judgeResult.outputFiles ?? (outputFiles.length > 0 ? outputFiles : undefined),\n      });\n\n      await cleanWorkingDirectory(args.cwd, cwdSnapshot);\n      if (decisionCode !== DecisionCode.ACCEPTED) break;\n    }\n  } catch (error) {\n    printTestCaseResult({\n      testCaseId: currentTestCaseId,\n      decisionCode: DecisionCode.RUNTIME_ERROR,\n      stdin: currentStdin,\n      stderr: errorToMessage(error),\n    });\n    await cleanWorkingDirectory(args.cwd, cwdSnapshot);\n  } finally {\n    await displayServer?.dispose();\n  }\n}\n\nasync function resolveMainFilePath(context: {\n  cwd: string;\n  language?: string | string[];\n  configuredMainFilePath?: string;\n  allowConfiguredPathFallback?: boolean;\n}): Promise<string | undefined> {\n  if (context.configuredMainFilePath) {\n    const resolvedPath = path.join(context.cwd, context.configuredMainFilePath);\n    if (await pathExists(resolvedPath)) return context.configuredMainFilePath;\n    if (!context.allowConfiguredPathFallback) return undefined;\n  }\n\n  return await findEntryPointFile(context.cwd, context.language);\n}\n\nfunction inferLanguageIdsByPath(filePath: string): string[] | undefined {\n  const languageIds = Object.entries(languageIdToDefinition)\n    .filter(([, definition]) => definition.fileExtensions.some((ext) => filePath.endsWith(ext)))\n    .map(([languageId]) => languageId);\n  return languageIds.length > 0 ? languageIds : undefined;\n}\n\nfunction runDefaultPrepare(context: {\n  cwd: string;\n  env: NodeJS.ProcessEnv;\n  mainFilePath: string;\n  languageDefinition: NonNullable<ReturnType<typeof findLanguageDefinitionByPath>>;\n}): Partial<GuiJudgeCaseResult> | undefined {\n  const buildCommand = context.languageDefinition.buildCommand?.(context.mainFilePath);\n  if (!buildCommand) return undefined;\n\n  const buildResult = spawnSyncWithTimeout(\n    buildCommand[0],\n    buildCommand.slice(1),\n    { cwd: context.cwd, encoding: 'utf8', env: context.env },\n    BUILD_TIMEOUT_SECONDS\n  );\n\n  if (buildResult.timeSeconds > BUILD_TIMEOUT_SECONDS) {\n    return {\n      decisionCode: DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,\n      stderr: buildResult.stderr || undefined,\n    };\n  }\n\n  if (buildResult.status !== 0) {\n    return {\n      decisionCode: DecisionCode.BUILD_ERROR,\n      stderr: buildResult.stderr || buildResult.stdout || undefined,\n    };\n  }\n\n  return undefined;\n}\n\nfunction evaluateGuiRunResult(context: {\n  runResult: GuiCommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  context: GuiJudgeContext;\n}): Partial<GuiJudgeCaseResult> {\n  if (context.runResult.stopReason === 'timeout') {\n    return {\n      decisionCode: DecisionCode.TIME_LIMIT_EXCEEDED,\n      stderr: context.runResult.stderr,\n    };\n  }\n\n  if (context.runResult.status !== 0) {\n    return {\n      decisionCode: DecisionCode.RUNTIME_ERROR,\n      stderr: context.runResult.stderr,\n    };\n  }\n\n  if (\n    context.runResult.memoryBytes >\n    (context.context.problemMarkdownFrontMatter.memoryLimitByte ?? Number.POSITIVE_INFINITY)\n  ) {\n    return {\n      decisionCode: DecisionCode.MEMORY_LIMIT_EXCEEDED,\n      stderr: context.runResult.stderr,\n    };\n  }\n\n  const requiredOutputFilesCount = context.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length ?? 0;\n  if (context.outputFiles.length < requiredOutputFilesCount) {\n    return {\n      decisionCode: DecisionCode.MISSING_REQUIRED_OUTPUT_FILE_ERROR,\n    };\n  }\n\n  return { decisionCode: DecisionCode.ACCEPTED };\n}\n\nasync function readGuiTestCases<TTestCase extends BaseGuiTestCase>(problemDir: string): Promise<readonly TTestCase[]> {\n  // The default reader yields the base shape; a narrower TTestCase must come from `options.readTestCases`.\n  return (await readFileTestCases(path.join(problemDir, 'test_cases'))) as unknown as readonly TTestCase[];\n}\n\nasync function spawnGuiProgram(context: {\n  command: readonly [string, ...string[]];\n  stdin: string;\n  cwd: string;\n  env: NodeJS.ProcessEnv;\n  timeLimitSeconds: number;\n  screenshotWaitSeconds: number;\n  stopDetectionThreshold: number;\n}): Promise<GuiCommandRunResult> {\n  const child = childProcess.spawn(\n    'timeout',\n    [context.timeLimitSeconds.toFixed(3), ...TIME_COMMAND, ...context.command],\n    {\n      cwd: context.cwd,\n      env: context.env,\n      detached: process.platform !== 'win32',\n      stdio: ['pipe', 'pipe', 'pipe'],\n    }\n  );\n\n  child.stdout.setEncoding('utf8');\n  child.stderr.setEncoding('utf8');\n\n  let stdout = '';\n  let stderr = '';\n  let exitCode: number | undefined;\n  let spawnError: Error | undefined;\n  let stopReason: GuiCommandRunResult['stopReason'] = 'process_exit';\n  const screenshotSignaturesHistory: string[][] = [];\n  let screenshots: GuiScreenshotFile[] = [];\n  const startTimeSeconds = Date.now() / 1000;\n  let sampledMemoryBytes = 0;\n  child.stdout.on('data', (chunk: string) => {\n    stdout += chunk;\n  });\n  child.stderr.on('data', (chunk: string) => {\n    stderr += chunk;\n  });\n  child.on('error', (error) => {\n    spawnError = error;\n    exitCode = 1;\n  });\n  // A child that exits before the input is fully written makes the write fail; without a listener\n  // that EPIPE would be an unhandled stream error terminating the harness mid-run.\n  child.stdin.on('error', () => {\n    // The submission simply stopped reading its input; the exit handling below reports the result.\n  });\n  child.on('close', (code, signal) => {\n    if (code === 124) {\n      stopReason = 'timeout';\n      exitCode = 0;\n      return;\n    }\n    if (signal) {\n      exitCode = 1;\n      stderr = stderr\n        ? `${stderr}\\nprocess terminated by signal: ${signal}`\n        : `process terminated by signal: ${signal}`;\n      return;\n    }\n    exitCode = code ?? 1;\n  });\n\n  // Stop the program even when a helper throws (e.g. `xwininfo` missing); the spawned process\n  // would otherwise outlive this run.\n  try {\n    if (context.stdin) child.stdin.write(context.stdin);\n    child.stdin.end();\n\n    while (exitCode === undefined) {\n      await wait(context.screenshotWaitSeconds * 1000);\n      sampledMemoryBytes = Math.max(sampledMemoryBytes, readProcessGroupMemoryBytes(child.pid));\n      const currentScreenshots = takeScreenshots(context.env.DISPLAY);\n      screenshots = currentScreenshots.toSorted((a, b) => a.data.length - b.data.length);\n\n      if (screenshots.length > 0) {\n        const screenshotSignatures = screenshots.map((file) => file.data).toSorted();\n        screenshotSignaturesHistory.unshift(screenshotSignatures);\n        screenshotSignaturesHistory.length = Math.min(\n          screenshotSignaturesHistory.length,\n          context.stopDetectionThreshold\n        );\n        if (\n          screenshotSignaturesHistory.length === context.stopDetectionThreshold &&\n          screenshotSignaturesHistory.every(\n            (files) =>\n              files.length === screenshotSignatures.length &&\n              files.every((file, index) => file === screenshotSignatures[index])\n          )\n        ) {\n          stopReason = 'stable_screenshot';\n          exitCode = 0;\n          break;\n        }\n      }\n\n      if (Date.now() / 1000 - startTimeSeconds > context.timeLimitSeconds) {\n        stopReason = 'timeout';\n        exitCode = 0;\n        break;\n      }\n    }\n\n    if (stopReason !== 'process_exit' && child.exitCode === null) {\n      child.removeAllListeners('close');\n      child.removeAllListeners('error');\n    }\n  } finally {\n    await stopProcess(child);\n  }\n  if (spawnError) throw spawnError;\n  const {\n    memoryBytes,\n    stderr: normalizedStderr,\n    timeSeconds,\n  } = parseTimedStderr(stderr, startTimeSeconds, sampledMemoryBytes);\n\n  return {\n    stdin: context.stdin,\n    stdout: stdout.trimEnd(),\n    stderr: normalizedStderr,\n    status: exitCode,\n    timeSeconds,\n    memoryBytes,\n    screenshots,\n    stopReason,\n  };\n}\n\nfunction takeScreenshots(display: string | undefined): GuiScreenshotFile[] {\n  // Node omits environment entries whose value is undefined, so a missing display is simply not set.\n  const env = { ...process.env, DISPLAY: display ?? process.env.DISPLAY };\n  const xwininfo = childProcess.spawnSync('xwininfo', ['-root', '-tree'], { encoding: 'utf8', env });\n  if (xwininfo.error) throw xwininfo.error;\n  if (xwininfo.status !== 0 || !xwininfo.stdout) return [];\n\n  const screenshots: GuiScreenshotFile[] = [];\n  for (const windowId of extractTopLevelWindowIds(xwininfo.stdout)) {\n    const screenshot = childProcess.spawnSync('maim', ['-i', windowId], { env });\n    if (screenshot.error) throw screenshot.error;\n    if (screenshot.status !== 0 || screenshot.stdout.length === 0) continue;\n\n    const windowNameResult = childProcess.spawnSync('xdotool', ['getwindowname', windowId], { encoding: 'utf8', env });\n    if (windowNameResult.error) throw windowNameResult.error;\n    const windowName = windowNameResult.stdout.trim().replaceAll(/[\\s/]/g, '_');\n\n    screenshots.push({\n      path: `${windowName || 'window'}_${windowId}.png`,\n      data: screenshot.stdout.toString('base64'),\n      encoding: 'base64',\n    });\n  }\n\n  return screenshots;\n}\n\nfunction extractTopLevelWindowIds(stdout: string): string[] {\n  const windowIds: string[] = [];\n  const lines = stdout.split('\\n');\n  for (const line of lines) {\n    if (line.includes('Root window id:') || line.includes('Parent window id:') || line.includes('()')) continue;\n\n    const match = /^\\s{5}(0x[\\da-f]+) /.exec(line);\n    if (!match?.[1]) continue;\n    windowIds.push(Number.parseInt(match[1], 16).toString());\n  }\n  return windowIds;\n}\n\nfunction parseTimedStderr(\n  stderr: string,\n  startTimeSeconds: number,\n  sampledMemoryBytes: number\n): Pick<GuiCommandRunResult, 'stderr' | 'timeSeconds' | 'memoryBytes'> {\n  const match = /(?:^|\\n)(\\d+\\.\\d+) (\\d+)\\s*$/.exec(stderr);\n  const normalizedStderr = match ? stderr.slice(0, match.index).trimEnd() : stderr.trimEnd();\n  const parsedMemoryBytes = Number(match?.[2]) * 1024 || 0;\n  return {\n    stderr: normalizedStderr,\n    timeSeconds: Number(match?.[1]) || Date.now() / 1000 - startTimeSeconds,\n    memoryBytes: Math.max(parsedMemoryBytes, sampledMemoryBytes),\n  };\n}\n\nasync function ensureDisplayServer(): Promise<{ display: string; dispose: () => Promise<void> }> {\n  if (process.platform !== 'linux') {\n    throw new Error('GUI screenshot capture is supported only on Linux.');\n  }\n\n  for (let displayNumber = 90; displayNumber < 100; displayNumber++) {\n    const display = `:${displayNumber}`;\n    let spawnError: Error | undefined;\n    const xvfb = childProcess.spawn('Xvfb', [display, '-screen', '0', '1280x1024x24', '-ac'], { stdio: 'ignore' });\n    xvfb.on('error', (error) => {\n      spawnError = error;\n    });\n\n    await wait(XVFB_STARTUP_WAIT_SECONDS * 1000);\n    if (spawnError) throw spawnError;\n    if (xvfb.exitCode !== null) continue;\n\n    return {\n      display,\n      dispose: async () => {\n        if (!xvfb.killed) {\n          xvfb.kill('SIGTERM');\n          await wait(XVFB_SHUTDOWN_WAIT_SECONDS * 1000);\n          if (xvfb.exitCode === null) xvfb.kill('SIGKILL');\n        }\n      },\n    };\n  }\n\n  throw new Error('Xvfb could not be started.');\n}\n\nasync function stopProcess(child: childProcess.ChildProcess): Promise<void> {\n  if (!child.pid) return;\n  killProcessGroup(child.pid, 'SIGTERM');\n  await wait(PROCESS_SHUTDOWN_WAIT_SECONDS * 1000);\n  if (child.exitCode === null) killProcessGroup(child.pid, 'SIGKILL');\n}\n\nfunction readProcessGroupMemoryBytes(processGroupId: number | undefined): number {\n  if (!processGroupId || process.platform !== 'linux') return 0;\n\n  const result = childProcess.spawnSync('ps', ['-o', 'rss=', '--no-headers', '--pgroup', String(processGroupId)], {\n    encoding: 'utf8',\n  });\n  if (result.error || result.status !== 0 || !result.stdout) return 0;\n\n  return result.stdout\n    .split('\\n')\n    .map((line) => Number(line.trim()))\n    .filter((value) => Number.isFinite(value) && value > 0)\n    .reduce((sum, value) => sum + value * 1024, 0);\n}\n\nfunction killProcessGroup(processGroupId: number, signal: NodeJS.Signals): void {\n  try {\n    if (process.platform !== 'win32') {\n      process.kill(-processGroupId, signal);\n      return;\n    }\n  } catch {\n    // The process group may already be gone. Fall back to the direct PID below.\n  }\n\n  try {\n    process.kill(processGroupId, signal);\n  } catch {\n    // The direct child may also already be gone.\n  }\n}\n\nasync function pathExists(filePath: string): Promise<boolean> {\n  try {\n    await fs.access(filePath);\n    return true;\n  } catch {\n    return false;\n  }\n}\n\nfunction errorToMessage(error: unknown): string {\n  return error instanceof Error ? error.message : String(error);\n}\n"],"mappings":"09BAwBA,MAEM,EAA0B,GAK1B,EAAe,CAACA,EAAAA,QAAG,SAAS,IAAM,SAAW,QAAU,gBAAiB,WAAY,OAAO,EAE3F,EAAoBC,EAAAA,EAAE,OAAO,CACjC,SAAUA,EAAAA,EAAE,MAAM,CAACA,EAAAA,EAAE,OAAO,EAAGA,EAAAA,EAAE,MAAMA,EAAAA,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAChE,CAAC,EA2GD,eAAsB,EACpB,EACA,EACe,CACf,IAAM,EAAOC,EAAAA,UAAU,QAAQ,IAAI,EACnC,GAAI,CAAC,EAAK,IAAK,MAAU,MAAM,uBAAuB,EACtD,IAAM,EAAS,EAAkB,MAAM,EAAK,MAAM,EAC5C,EAAgB,EAAK,IAErB,EAA6B,MAAMC,EAAAA,+BAA+B,CAAU,EAC5E,EAAsB,MAAO,EAAQ,eAAiB,EAAA,CAA6B,CAAU,EAC7F,EACJ,EAAoB,OAAS,EAAI,EAAuB,CAAC,CAAE,GAAI,SAAU,CAAC,EACtE,EAAqB,EAAU,EAAE,EAAE,IAAM,WAEzC,EAAuB,MAAMC,EAAAA,sBAAsB,EAAK,IAAK,CAA0B,EAC7F,GAAI,EAAsB,CACxB,EAAA,oBAAoB,CAAE,WAAY,EAAoB,GAAG,CAAqB,CAAC,EAC/E,MACF,CAEA,IAAM,EAAsB,EAAQ,cAAiB,MAAMC,EAAAA,mBAAmB,EAAK,IAAK,EAAO,QAAQ,EACvG,GAAI,CAAC,EAAqB,CACxB,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcC,EAAAA,aAAa,uCAC3B,OAAQ,EAAQ,aACZ,iCAAiC,EAAQ,eACzC,sBAAsB,EAAO,SAAW,eAAe,EAAO,WAAa,IACjF,CAAC,EACD,MACF,CAEA,IAAM,EAAqBC,EAAAA,6BAA6B,CAAmB,EAC3E,GAAI,CAAC,EAAoB,CACvB,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcD,EAAAA,aAAa,aAC3B,OAAQ,sBACV,CAAC,EACD,MACF,CAEA,IAAM,EAAM,CAAE,GAAG,QAAQ,IAAK,GAAI,GAAI,YAAa,GAAI,EAEnD,EAAuB,MAAM,EAAoB,CACnD,IAAK,EAAK,IACV,SAAU,EAAO,SACjB,uBAAwB,EAAQ,YAClC,CAAC,EACD,GAAI,EAAmB,SACrB,GAAI,CACF,MAAM,EAAmB,SAAS,EAAK,GAAG,EAC1C,IAAM,EAAuB,MAAM,EAAoB,CACrD,IAAK,EAAK,IACV,SAAU,EAAO,UAAY,EAAuB,CAAmB,EACvE,uBAAwB,EAAQ,aAChC,4BAA6B,EAC/B,CAAC,EACG,IAAsB,EAAuB,EACnD,OAAS,EAAO,CACd,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcA,EAAAA,aAAa,YAC3B,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MACF,CAEF,GAAI,CAAC,EAAsB,CACzB,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcA,EAAAA,aAAa,uCAC3B,OAAQ,EAAQ,aACZ,iCAAiC,EAAQ,eACzC,sBAAsB,EAAO,SAAW,eAAe,EAAO,WAAa,IACjF,CAAC,EACD,MACF,CAEA,IAAI,EACJ,GAAI,EAAQ,QACV,GAAI,CACF,EAAsB,MAAM,EAAQ,QAAQ,CAC1C,IAAK,EACL,MACA,aAAc,EACd,4BACF,CAAC,CACH,OAAS,EAAO,CAGd,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcA,EAAAA,aAAa,YAC3B,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MACF,CAEF,IAAM,EACJ,GACA,EAAkB,CAChB,IAAK,EAAK,IACV,MACA,aAAc,EACd,oBACF,CAAC,EACH,GAAI,EAAe,CACjB,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAc,EAAc,cAAgBA,EAAAA,aAAa,YACzD,iBAAkB,EAAc,iBAChC,OAAQ,EAAc,OACtB,OAAQ,EAAc,OACtB,YAAa,EAAc,WAC7B,CAAC,EACD,MACF,CAEA,IAAM,EAAc,MAAME,EAAAA,yBAAyB,EAAK,GAAG,EACvD,EACA,EAAoB,EACpB,EACJ,GAAI,CACF,EAAgB,EAAQ,WAAa,IAAA,GAAY,MAAM,EAAoB,EAC3E,IAAM,EAAuB,EAAgE,QAAQ,cACrG,IAAK,IAAM,KAAY,EAAW,CAChC,EAAoB,EAAS,GACzB,GAAqB,MAAMC,EAAAA,sBAAsB,EAAqB,EAAK,GAAG,EAC9E,EAAS,eAAe,MAAMA,EAAAA,sBAAsB,EAAS,cAAe,EAAK,GAAG,EAExF,IAAM,EACJ,OAAO,EAA2B,aAAgB,SAC9C,EAA2B,YAAc,IACxC,EAAQ,mBAAqB,EAE9B,EAAS,EAAgB,CAAE,GAAG,EAAK,QAAS,EAAc,OAAQ,EAAI,EACtE,EAAS,MAAM,EAAQ,eAAe,CAAE,WAAU,IAAK,EAAK,IAAK,IAAK,CAAO,CAAC,GAAM,EAAS,OAAS,GAC5G,EAAe,EACf,IAAM,EACH,MAAM,EAAQ,UAAU,CAAE,WAAU,IAAK,EAAK,IAAK,IAAK,EAAQ,aAAc,CAAqB,CAAC,GACrG,EAAmB,QAAQ,CAAoB,EAE7C,EACJ,GAAI,CACF,EAAY,EAAQ,WAChB,MAAM,EAAQ,WAAW,CACvB,WACA,UACA,QACA,IAAK,EACL,IAAK,EACL,mBACA,sBAAuB,EAAQ,uBAAyB,EACxD,uBAAwB,EAAQ,wBAA0B,CAC5D,CAAC,EACD,MAAM,EAAgB,CACpB,UACA,QACA,IAAK,EAAK,IACV,IAAK,EACL,mBACA,sBAAuB,EAAQ,uBAAyB,EACxD,uBAAwB,EAAQ,wBAA0B,CAC5D,CAAC,CACP,OAAS,EAAO,CACd,EAAA,oBAAoB,CAClB,WAAY,EAAS,GACrB,aAAcH,EAAAA,aAAa,cAC3B,QACA,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MAAMI,EAAAA,sBAAsB,EAAK,IAAK,CAAW,EACjD,MACF,CAEA,IAAM,EAAc,MAAMC,EAAAA,gBAAgB,EAAK,IAAK,EAA2B,yBAA2B,CAAC,CAAC,EACtG,EAAgC,CACpC,mBACA,2BAA4B,CAC1B,gBAAiB,EAA2B,gBAC5C,wBAAyB,EAA2B,uBACtD,CACF,EACM,EAAkB,EAAqB,CAAE,YAAW,cAAa,QAAS,CAAa,CAAC,EAC1F,EAAc,EAClB,GAAI,EAAgB,eAAiBL,EAAAA,aAAa,SAChD,GAAI,CACF,IAAM,EAAsB,MAAM,EAAQ,KAAK,CAC7C,WACA,YACA,cACA,IAAK,EAAK,IACV,QAAS,CACX,CAAC,EACD,EAAc,CACZ,aAAc,EAAoB,cAAgB,EAAgB,aAClE,iBAAkB,EAAoB,iBACtC,OAAQ,EAAoB,OAC5B,OAAQ,EAAoB,OAC5B,YAAa,EAAoB,WACnC,CACF,OAAS,EAAO,CACd,EAAc,CACZ,aAAcA,EAAAA,aAAa,cAC3B,OAAQ,EAAe,CAAK,CAC9B,CACF,CAGF,IAAM,EAAe,EAAY,cAAgBA,EAAAA,aAAa,SACxD,EAAS,EAAY,QAAU,EAAU,OACzC,EAAS,EAAY,QAAU,EAAU,OAe/C,GAdA,EAAA,oBAAoB,CAClB,WAAY,EAAS,GACrB,eACA,WAAY,EAAU,OACtB,MAAO,EAAU,OAAS,IAAA,GAC1B,OAAQ,GAAU,IAAA,GAClB,OAAQ,GAAU,IAAA,GAClB,YAAa,EAAU,YACvB,YAAa,EAAU,YACvB,iBAAkB,EAAY,iBAC9B,YAAa,EAAY,cAAgB,EAAY,OAAS,EAAI,EAAc,IAAA,GAClF,CAAC,EAED,MAAMI,EAAAA,sBAAsB,EAAK,IAAK,CAAW,EAC7C,IAAiBJ,EAAAA,aAAa,SAAU,KAC9C,CACF,OAAS,EAAO,CACd,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcA,EAAAA,aAAa,cAC3B,MAAO,EACP,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MAAMI,EAAAA,sBAAsB,EAAK,IAAK,CAAW,CACnD,QAAU,CACR,MAAM,GAAe,QAAQ,CAC/B,CACF,CAEA,eAAe,EAAoB,EAKH,CAC9B,GAAI,EAAQ,uBAAwB,CAElC,GAAI,MAAM,EADWE,EAAAA,QAAK,KAAK,EAAQ,IAAK,EAAQ,sBACpB,CAAC,EAAG,OAAO,EAAQ,uBACnD,GAAI,CAAC,EAAQ,4BAA6B,MAC5C,CAEA,OAAO,MAAMP,EAAAA,mBAAmB,EAAQ,IAAK,EAAQ,QAAQ,CAC/D,CAEA,SAAS,EAAuB,EAAwC,CACtE,IAAM,EAAc,OAAO,QAAQQ,EAAAA,sBAAsB,CAAC,CACvD,QAAQ,EAAG,KAAgB,EAAW,eAAe,KAAM,GAAQ,EAAS,SAAS,CAAG,CAAC,CAAC,CAAC,CAC3F,KAAK,CAAC,KAAgB,CAAU,EACnC,OAAO,EAAY,OAAS,EAAI,EAAc,IAAA,EAChD,CAEA,SAAS,EAAkB,EAKiB,CAC1C,IAAM,EAAe,EAAQ,mBAAmB,eAAe,EAAQ,YAAY,EACnF,GAAI,CAAC,EAAc,OAEnB,IAAM,EAAcC,EAAAA,qBAClB,EAAa,GACb,EAAa,MAAM,CAAC,EACpB,CAAE,IAAK,EAAQ,IAAK,SAAU,OAAQ,IAAK,EAAQ,GAAI,EACvD,EACF,EAEA,GAAI,EAAY,YAAc,GAC5B,MAAO,CACL,aAAcR,EAAAA,aAAa,0BAC3B,OAAQ,EAAY,QAAU,IAAA,EAChC,EAGF,GAAI,EAAY,SAAW,EACzB,MAAO,CACL,aAAcA,EAAAA,aAAa,YAC3B,OAAQ,EAAY,QAAU,EAAY,QAAU,IAAA,EACtD,CAIJ,CAEA,SAAS,EAAqB,EAIE,CAC9B,GAAI,EAAQ,UAAU,aAAe,UACnC,MAAO,CACL,aAAcA,EAAAA,aAAa,oBAC3B,OAAQ,EAAQ,UAAU,MAC5B,EAGF,GAAI,EAAQ,UAAU,SAAW,EAC/B,MAAO,CACL,aAAcA,EAAAA,aAAa,cAC3B,OAAQ,EAAQ,UAAU,MAC5B,EAGF,GACE,EAAQ,UAAU,aACjB,EAAQ,QAAQ,2BAA2B,iBAAmB,KAE/D,MAAO,CACL,aAAcA,EAAAA,aAAa,sBAC3B,OAAQ,EAAQ,UAAU,MAC5B,EAGF,IAAM,EAA2B,EAAQ,QAAQ,2BAA2B,yBAAyB,QAAU,EAO/G,OANI,EAAQ,YAAY,OAAS,EACxB,CACL,aAAcA,EAAAA,aAAa,kCAC7B,EAGK,CAAE,aAAcA,EAAAA,aAAa,QAAS,CAC/C,CAEA,eAAe,EAAoD,EAAmD,CAEpH,OAAQ,MAAMS,EAAAA,cAAkBH,EAAAA,QAAK,KAAK,EAAY,YAAY,CAAC,CACrE,CAEA,eAAe,EAAgB,EAQE,CAC/B,IAAM,EAAQI,EAAAA,QAAa,MACzB,UACA,CAAC,EAAQ,iBAAiB,QAAQ,CAAC,EAAG,GAAG,EAAc,GAAG,EAAQ,OAAO,EACzE,CACE,IAAK,EAAQ,IACb,IAAK,EAAQ,IACb,SAAU,QAAQ,WAAa,QAC/B,MAAO,CAAC,OAAQ,OAAQ,MAAM,CAChC,CACF,EAEA,EAAM,OAAO,YAAY,MAAM,EAC/B,EAAM,OAAO,YAAY,MAAM,EAE/B,IAAI,EAAS,GACT,EAAS,GACT,EACA,EACA,EAAgD,eAC9C,EAA0C,CAAC,EAC7C,EAAmC,CAAC,EAClC,EAAmB,KAAK,IAAI,EAAI,IAClC,EAAqB,EACzB,EAAM,OAAO,GAAG,OAAS,GAAkB,CACzC,GAAU,CACZ,CAAC,EACD,EAAM,OAAO,GAAG,OAAS,GAAkB,CACzC,GAAU,CACZ,CAAC,EACD,EAAM,GAAG,QAAU,GAAU,CAC3B,EAAa,EACb,EAAW,CACb,CAAC,EAGD,EAAM,MAAM,GAAG,YAAe,CAE9B,CAAC,EACD,EAAM,GAAG,SAAU,EAAM,IAAW,CAClC,GAAI,IAAS,IAAK,CAChB,EAAa,UACb,EAAW,EACX,MACF,CACA,GAAI,EAAQ,CACV,EAAW,EACX,EAAS,EACL,GAAG,EAAO,kCAAkC,IAC5C,iCAAiC,IACrC,MACF,CACA,EAAW,GAAQ,CACrB,CAAC,EAID,GAAI,CAIF,IAHI,EAAQ,OAAO,EAAM,MAAM,MAAM,EAAQ,KAAK,EAClD,EAAM,MAAM,IAAI,EAET,IAAa,IAAA,IAAW,CAM7B,GALA,MAAA,EAAMC,EAAAA,WAAAA,CAAK,EAAQ,sBAAwB,GAAI,EAC/C,EAAqB,KAAK,IAAI,EAAoB,EAA4B,EAAM,GAAG,CAAC,EAExF,EAD2B,EAAgB,EAAQ,IAAI,OACxB,CAAC,CAAC,UAAU,EAAG,IAAM,EAAE,KAAK,OAAS,EAAE,KAAK,MAAM,EAE7E,EAAY,OAAS,EAAG,CAC1B,IAAM,EAAuB,EAAY,IAAK,GAAS,EAAK,IAAI,CAAC,CAAC,SAAS,EAM3E,GALA,EAA4B,QAAQ,CAAoB,EACxD,EAA4B,OAAS,KAAK,IACxC,EAA4B,OAC5B,EAAQ,sBACV,EAEE,EAA4B,SAAW,EAAQ,wBAC/C,EAA4B,MACzB,GACC,EAAM,SAAW,EAAqB,QACtC,EAAM,OAAO,EAAM,IAAU,IAAS,EAAqB,EAAM,CACrE,EACA,CACA,EAAa,oBACb,EAAW,EACX,KACF,CACF,CAEA,GAAI,KAAK,IAAI,EAAI,IAAO,EAAmB,EAAQ,iBAAkB,CACnE,EAAa,UACb,EAAW,EACX,KACF,CACF,CAEI,IAAe,gBAAkB,EAAM,WAAa,OACtD,EAAM,mBAAmB,OAAO,EAChC,EAAM,mBAAmB,OAAO,EAEpC,QAAU,CACR,MAAM,EAAY,CAAK,CACzB,CACA,GAAI,EAAY,MAAM,EACtB,GAAM,CACJ,cACA,OAAQ,EACR,eACE,EAAiB,EAAQ,EAAkB,CAAkB,EAEjE,MAAO,CACL,MAAO,EAAQ,MACf,OAAQ,EAAO,QAAQ,EACvB,OAAQ,EACR,OAAQ,EACR,cACA,cACA,cACA,YACF,CACF,CAEA,SAAS,EAAgB,EAAkD,CAEzE,IAAM,EAAM,CAAE,GAAG,QAAQ,IAAK,QAAS,GAAW,QAAQ,IAAI,OAAQ,EAChE,EAAWD,EAAAA,QAAa,UAAU,WAAY,CAAC,QAAS,OAAO,EAAG,CAAE,SAAU,OAAQ,KAAI,CAAC,EACjG,GAAI,EAAS,MAAO,MAAM,EAAS,MACnC,GAAI,EAAS,SAAW,GAAK,CAAC,EAAS,OAAQ,MAAO,CAAC,EAEvD,IAAM,EAAmC,CAAC,EAC1C,IAAK,IAAM,KAAY,EAAyB,EAAS,MAAM,EAAG,CAChE,IAAM,EAAaA,EAAAA,QAAa,UAAU,OAAQ,CAAC,KAAM,CAAQ,EAAG,CAAE,KAAI,CAAC,EAC3E,GAAI,EAAW,MAAO,MAAM,EAAW,MACvC,GAAI,EAAW,SAAW,GAAK,EAAW,OAAO,SAAW,EAAG,SAE/D,IAAM,EAAmBA,EAAAA,QAAa,UAAU,UAAW,CAAC,gBAAiB,CAAQ,EAAG,CAAE,SAAU,OAAQ,KAAI,CAAC,EACjH,GAAI,EAAiB,MAAO,MAAM,EAAiB,MACnD,IAAM,EAAa,EAAiB,OAAO,KAAK,CAAC,CAAC,WAAW,SAAU,GAAG,EAE1E,EAAY,KAAK,CACf,KAAM,GAAG,GAAc,SAAS,GAAG,EAAS,MAC5C,KAAM,EAAW,OAAO,SAAS,QAAQ,EACzC,SAAU,QACZ,CAAC,CACH,CAEA,OAAO,CACT,CAEA,SAAS,EAAyB,EAA0B,CAC1D,IAAM,EAAsB,CAAC,EACvB,EAAQ,EAAO,MAAM;CAAI,EAC/B,IAAK,IAAM,KAAQ,EAAO,CACxB,GAAI,EAAK,SAAS,iBAAiB,GAAK,EAAK,SAAS,mBAAmB,GAAK,EAAK,SAAS,IAAI,EAAG,SAEnG,IAAM,EAAQ,sBAAsB,KAAK,CAAI,EACxC,IAAQ,IACb,EAAU,KAAK,OAAO,SAAS,EAAM,GAAI,EAAE,CAAC,CAAC,SAAS,CAAC,CACzD,CACA,OAAO,CACT,CAEA,SAAS,EACP,EACA,EACA,EACqE,CACrE,IAAM,EAAQ,+BAA+B,KAAK,CAAM,EAClD,EAAmB,EAAQ,EAAO,MAAM,EAAG,EAAM,KAAK,CAAC,CAAC,QAAQ,EAAI,EAAO,QAAQ,EACnF,EAAoB,OAAO,IAAQ,EAAE,EAAI,MAAQ,EACvD,MAAO,CACL,OAAQ,EACR,YAAa,OAAO,IAAQ,EAAE,GAAK,KAAK,IAAI,EAAI,IAAO,EACvD,YAAa,KAAK,IAAI,EAAmB,CAAkB,CAC7D,CACF,CAEA,eAAe,GAAkF,CAC/F,GAAI,QAAQ,WAAa,QACvB,MAAU,MAAM,oDAAoD,EAGtE,IAAK,IAAI,EAAgB,GAAI,EAAgB,IAAK,IAAiB,CACjE,IAAM,EAAU,IAAI,IAChB,EACE,EAAOA,EAAAA,QAAa,MAAM,OAAQ,CAAC,EAAS,UAAW,IAAK,eAAgB,KAAK,EAAG,CAAE,MAAO,QAAS,CAAC,EAM7G,GALA,EAAK,GAAG,QAAU,GAAU,CAC1B,EAAa,CACf,CAAC,EAED,MAAA,EAAMC,EAAAA,WAAAA,CAAK,GAAgC,EACvC,EAAY,MAAM,EAClB,KAAK,WAAa,KAEtB,MAAO,CACL,UACA,QAAS,SAAY,CACd,EAAK,SACR,EAAK,KAAK,SAAS,EACnB,MAAA,EAAMA,EAAAA,WAAAA,CAAK,GAAiC,EACxC,EAAK,WAAa,MAAM,EAAK,KAAK,SAAS,EAEnD,CACF,CACF,CAEA,MAAU,MAAM,4BAA4B,CAC9C,CAEA,eAAe,EAAY,EAAiD,CACrE,EAAM,MACX,EAAiB,EAAM,IAAK,SAAS,EACrC,MAAA,EAAMA,EAAAA,WAAAA,CAAK,GAAoC,EAC3C,EAAM,WAAa,MAAM,EAAiB,EAAM,IAAK,SAAS,EACpE,CAEA,SAAS,EAA4B,EAA4C,CAC/E,GAAI,CAAC,GAAkB,QAAQ,WAAa,QAAS,MAAO,GAE5D,IAAM,EAASD,EAAAA,QAAa,UAAU,KAAM,CAAC,KAAM,OAAQ,eAAgB,WAAY,OAAO,CAAc,CAAC,EAAG,CAC9G,SAAU,MACZ,CAAC,EAGD,OAFI,EAAO,OAAS,EAAO,SAAW,GAAK,CAAC,EAAO,OAAe,EAE3D,EAAO,OACX,MAAM;CAAI,CAAC,CACX,IAAK,GAAS,OAAO,EAAK,KAAK,CAAC,CAAC,CAAC,CAClC,OAAQ,GAAU,OAAO,SAAS,CAAK,GAAK,EAAQ,CAAC,CAAC,CACtD,QAAQ,EAAK,IAAU,EAAM,EAAQ,KAAM,CAAC,CACjD,CAEA,SAAS,EAAiB,EAAwB,EAA8B,CAC9E,GAAI,CACF,GAAI,QAAQ,WAAa,QAAS,CAChC,QAAQ,KAAK,CAAC,EAAgB,CAAM,EACpC,MACF,CACF,MAAQ,CAER,CAEA,GAAI,CACF,QAAQ,KAAK,EAAgB,CAAM,CACrC,MAAQ,CAER,CACF,CAEA,eAAe,EAAW,EAAoC,CAC5D,GAAI,CAEF,OADA,MAAME,EAAAA,QAAG,OAAO,CAAQ,EACjB,EACT,MAAQ,CACN,MAAO,EACT,CACF,CAEA,SAAS,EAAe,EAAwB,CAC9C,OAAO,aAAiB,MAAQ,EAAM,QAAU,OAAO,CAAK,CAC9D"}

package/dist/presets/guiCommand.d.ts

@@ -43,24 +43,10 @@ export interface GuiCommandJudgePresetOptions<TTestCase extends BaseGuiTestCase
     readTestCases?: (problemDir: string) => Promise<readonly TTestCase[]>;
     prepare?: (context: {
         cwd: string;
-        /**
-         * Build the submission with it. Under `EXERCODE_SANDBOX_USER` delegation it already carries the
-         * sandbox user's overrides, but the handler must still wrap whatever it spawns with
-         * `wrapCommandWithSandboxUser` (both exported): a build runs the submission's own scripts
-         * (`package.json` lifecycle scripts, `build.gradle`, `build.rs`), which as the trusted harness
-         * user could read the problem's test cases and rewrite the harness. The preset terminates
-         * leftover sandbox processes after the handler returns.
-         */
         env: NodeJS.ProcessEnv;
         mainFilePath: string;
         problemMarkdownFrontMatter: ProblemMarkdownFrontMatter;
     }) => Promise<Partial<GuiJudgeCaseResult> | undefined> | Partial<GuiJudgeCaseResult> | undefined;
-    /**
-     * Runs as the trusted harness user with the submission's `cwd`. Fixture files it creates there
-     * must be written with `createDirectoryWithoutFollowingSymlinks`/`writeFileWithoutFollowingSymlinks`
-     * (both exported): a submission can plant a symlink at a fixture path, and a plain `fs.writeFile`
-     * would follow it into a file only the harness can write.
-     */
     resolveInput?: (context: {
         testCase: TTestCase;
         cwd: string;
@@ -74,16 +60,6 @@ export interface GuiCommandJudgePresetOptions<TTestCase extends BaseGuiTestCase
     }) => Promise<readonly [string, ...string[]]> | readonly [string, ...string[]];
     runCommand?: (context: {
         testCase: TTestCase;
-        /**
-         * The command to run. Under `EXERCODE_SANDBOX_USER` delegation it is already wrapped so it
-         * executes as the sandbox user; spawn it as given (with the supplied `env`) instead of
-         * reconstructing it, or the submission runs as the trusted harness user.
-         *
-         * Its direct child is then a root-owned `sudo` whose descendants belong to the sandbox user, so
-         * the handler cannot signal them: enforce `timeLimitSeconds` with `startSandboxTimeoutWatchdog`
-         * and `killSandboxUserProcesses` (both exported) rather than `child.kill()` or an outer
-         * `timeout`. The preset terminates leftover sandbox processes after the handler returns.
-         */
         command: readonly [string, ...string[]];
         stdin: string;
         cwd: string;