@exercode/problem-utils 4.2.6 → 4.3.0

package/dist/presets/command.cjs

@@ -1,2 +1,2 @@
-"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:`Module`});const e=require("../_virtual/_rolldown/runtime.cjs"),t=require("../helpers/printTestCaseResult.cjs"),n=require("../helpers/compareExpectedOutputFiles.cjs"),r=require("../helpers/parseArgs.cjs"),i=require("../helpers/readTestCases.cjs"),a=require("../helpers/findLanguageDefinitionByPath.cjs"),o=require("../helpers/readProblemMarkdownFrontMatter.cjs"),s=require("../types/decisionCode.cjs"),c=require("../helpers/cleanWorkingDirectory.cjs"),l=require("../helpers/copyTestCaseFileInput.cjs"),u=require("../helpers/findEntryPointFile.cjs"),d=require("../helpers/judgeByStaticAnalysis.cjs"),f=require("../helpers/readOutputFiles.cjs"),p=require("../helpers/runIsolationCheckInDebugMode.cjs"),m=require("../helpers/resolveCwds.cjs"),h=require("../helpers/spawnWithTimeout.cjs");let g=require("node:path");g=e.__toESM(g,1);let _=require("zod");const v=_.z.object({language:_.z.union([_.z.string(),_.z.array(_.z.string())]).optional()});async function y(e,t={}){let n=r.parseArgs(process.argv),i=v.parse(n.params),{cwds:a,isDebugMode:s}=await m.resolveCwds(e,n.cwd),c=await o.readProblemMarkdownFrontMatter(e),l=await(t.readTestCases??w)(e),u={buildTimeoutSeconds:t.limits?.buildTimeoutSeconds??10,maxOutputLength:t.limits?.maxOutputLength??5e4},d=typeof c.timeLimitMs==`number`?c.timeLimitMs/1e3:t.runTimeoutSeconds??2,f={problemMarkdownFrontMatter:c,testCases:l,limits:u,timeLimitSeconds:d};if(s){let t=(u.buildTimeoutSeconds+l.length*d)*1e3;if(!await p.passesIsolationCheckInDebugMode(e,a,i,{expectedMaxDurationMs:t})){process.exitCode=1;return}}for(let n of a){s&&m.printDebugCwdBanner(e,n);let r=await b(n.cwd,i,f,t);s&&!x(n,r)&&(process.exitCode=1,m.printDebugExpectationFailureBanner(e,n))}}async function b(e,n,r,i){let{problemMarkdownFrontMatter:o,testCases:p,limits:m,timeLimitSeconds:h}=r,g=p[0]?.id??`prebuild`,_=await d.judgeByStaticAnalysis(e,o);if(_)return t.printTestCaseResult({testCaseId:g,..._}),{allAccepted:!1};let v=await u.findEntryPointFile(e,n.language);if(!v)return t.printTestCaseResult({testCaseId:g,decisionCode:s.DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,stderr:`main file not found${n.language?`: language: ${n.language}`:``}`}),{allAccepted:!1};let y=a.findLanguageDefinitionByPath(v);if(!y)return t.printTestCaseResult({testCaseId:g,decisionCode:s.DecisionCode.WRONG_ANSWER,stderr:`unsupported language`}),{allAccepted:!1};let b={...process.env,CI:``,FORCE_COLOR:`0`},x=v;if(i.buildSubmission!==!1){if(y.prebuild)try{await y.prebuild(e);let t=await u.findEntryPointFile(e,n.language);t&&(x=t)}catch(e){return t.printTestCaseResult({testCaseId:g,decisionCode:s.DecisionCode.BUILD_ERROR,stderr:e instanceof Error?e.message:String(e)}),{allAccepted:!1}}let r=y.buildCommand?.(x);if(r){let n=await S(r,{cwd:e,env:b,testCaseId:g,limits:m});if(n)return t.printTestCaseResult(n),{allAccepted:!1}}}let w=await c.snapshotWorkingDirectory(e);if(p.length===0)return t.printTestCaseResult({testCaseId:`default`,decisionCode:s.DecisionCode.ACCEPTED}),{allAccepted:!0};let O=p.shared?.fileInputPath;for(let n of p){O&&await l.copyTestCaseFileInput(O,e),n.fileInputPath&&await l.copyTestCaseFileInput(n.fileInputPath,e);let r=y.command(x),a=n.input??``,u;try{i.resolveInput&&(a=await i.resolveInput({testCase:n,cwd:e,env:b})),i.runCommand&&(u=await i.runCommand({testCase:n,command:r,stdin:a,cwd:e,env:b,timeLimitSeconds:h}))}catch(r){return t.printTestCaseResult({testCaseId:n.id,decisionCode:s.DecisionCode.RUNTIME_ERROR,stdin:a,stderr:C(r)}),await c.cleanWorkingDirectory(e,w),{allAccepted:!1}}let d=u??await E(r,{stdin:a,cwd:e,env:b,timeLimitSeconds:h}),p=await f.readOutputFiles(e,o.requiredOutputFilePaths??[]),g={timeLimitSeconds:h,outputLimitLength:m.maxOutputLength,problemMarkdownFrontMatter:{memoryLimitByte:o.memoryLimitByte,requiredOutputFilePaths:o.requiredOutputFilePaths}},_=D({runResult:d,outputFiles:p,context:g}),v=_;if(_.decisionCode===s.DecisionCode.ACCEPTED)try{let t=i.test?await i.test({testCase:n,runResult:d,outputFiles:p,cwd:e,context:g}):await T({testCase:n,runResult:d,outputFiles:p,cwd:e});t&&(v={decisionCode:t.decisionCode??_.decisionCode,feedbackMarkdown:t.feedbackMarkdown,stderr:t.stderr})}catch(e){v={decisionCode:s.DecisionCode.RUNTIME_ERROR,stderr:C(e)}}if(t.printTestCaseResult({testCaseId:n.id,...v,exitStatus:d.status,stdin:d.stdin,stdout:d.stdout.slice(0,m.maxOutputLength)||void 0,stderr:(v.stderr??d.stderr).slice(0,m.maxOutputLength)||void 0,timeSeconds:d.timeSeconds,cpuTimeSeconds:d.cpuTimeSeconds,memoryBytes:d.memoryBytes,outputFiles:p.length>0?p:void 0}),await c.cleanWorkingDirectory(e,w),v.decisionCode!==s.DecisionCode.ACCEPTED)return{allAccepted:!1}}return{allAccepted:!0}}function x(e,t){return t.allAccepted===(e.expectedResult===`accepted`)}async function S(e,t){let n=await h.spawnWithTimeout(e[0],e.slice(1),{cwd:t.cwd,env:t.env},t.limits.buildTimeoutSeconds),r=n.status??void 0;if(n.status!==0)return{testCaseId:t.testCaseId,decisionCode:s.DecisionCode.BUILD_ERROR,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes};if(n.timeSeconds>t.limits.buildTimeoutSeconds)return{testCaseId:t.testCaseId,decisionCode:s.DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes};if(n.outputLimitExceeded||n.stdout.length>t.limits.maxOutputLength||n.stderr.length>t.limits.maxOutputLength)return{testCaseId:t.testCaseId,decisionCode:s.DecisionCode.BUILD_OUTPUT_SIZE_LIMIT_EXCEEDED,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes}}function C(e){return e instanceof Error?e.message:String(e)}async function w(e){return await i.readTestCases(g.default.join(e,`test_cases`))}async function T(e){let{testCase:t,runResult:r,outputFiles:i,cwd:a}=e,o=await n.judgeAgainstExpectations({stdout:r.stdout,expectedStdout:typeof t.output==`string`?t.output:void 0,fileOutputPath:typeof t.fileOutputPath==`string`?t.fileOutputPath:void 0,cwd:a,outputFiles:i});return i.splice(0,i.length,...o.outputFiles),o.matches?{}:{decisionCode:s.DecisionCode.WRONG_ANSWER}}async function E(e,t){let n=await h.spawnWithTimeout(e[0],e.slice(1),{cwd:t.cwd,stdin:t.stdin,env:t.env},t.timeLimitSeconds);return{stdin:t.stdin,stdout:n.stdout,stderr:n.stderr,status:n.status??void 0,timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes,outputLimitExceeded:n.outputLimitExceeded}}function D(e){if(e.runResult.status!==0)return{decisionCode:s.DecisionCode.RUNTIME_ERROR,stderr:e.runResult.stderr};if(e.runResult.timeSeconds>e.context.timeLimitSeconds)return{decisionCode:s.DecisionCode.TIME_LIMIT_EXCEEDED,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};if(e.runResult.outputLimitExceeded||e.runResult.stdout.length>e.context.outputLimitLength||e.runResult.stderr.length>e.context.outputLimitLength)return{decisionCode:s.DecisionCode.OUTPUT_SIZE_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}}exports.commandJudgePreset=y;
+"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:`Module`});const e=require("../_virtual/_rolldown/runtime.cjs"),t=require("../helpers/printTestCaseResult.cjs"),n=require("../helpers/compareExpectedOutputFiles.cjs"),r=require("../helpers/parseArgs.cjs"),i=require("../helpers/readTestCases.cjs"),a=require("../helpers/findLanguageDefinitionByPath.cjs"),o=require("../helpers/readProblemMarkdownFrontMatter.cjs"),s=require("../types/decisionCode.cjs"),c=require("../helpers/cleanWorkingDirectory.cjs"),l=require("../helpers/copyTestCaseFileInput.cjs"),u=require("../helpers/findEntryPointFile.cjs"),d=require("../helpers/judgeByStaticAnalysis.cjs"),f=require("../helpers/readOutputFiles.cjs"),p=require("../helpers/runIsolationCheckInDebugMode.cjs"),m=require("../helpers/resolveCwds.cjs"),h=require("../helpers/spawnWithTimeout.cjs");let g=require("node:path");g=e.__toESM(g,1);let _=require("zod"),v=require("@willbooster/shared-lib");const y=_.z.object({language:_.z.union([_.z.string(),_.z.array(_.z.string())]).optional()});async function b(e,t={}){let n=r.parseArgs(process.argv),i=y.parse(n.params),{cwds:a,isDebugMode:s}=await m.resolveCwds(e,n.cwd),c=await o.readProblemMarkdownFrontMatter(e),l=await(t.readTestCases??w)(e),u={buildTimeoutSeconds:t.limits?.buildTimeoutSeconds??10,maxOutputLength:t.limits?.maxOutputLength??5e4},d=typeof c.timeLimitMs==`number`?c.timeLimitMs/1e3:t.runTimeoutSeconds??2,f={problemMarkdownFrontMatter:c,testCases:l,limits:u,timeLimitSeconds:d};if(s){let t=(u.buildTimeoutSeconds+l.length*d)*1e3;if(!await p.passesIsolationCheckInDebugMode(e,a,i,{expectedMaxDurationMs:t})){process.exitCode=1;return}}for(let n of a){s&&m.printDebugCwdBanner(e,n);let r=await x(n.cwd,i,f,t);s&&!S(n,r)&&(process.exitCode=1,m.printDebugExpectationFailureBanner(e,n))}}async function x(e,n,r,i){let{problemMarkdownFrontMatter:o,testCases:p,limits:m,timeLimitSeconds:h}=r,g=p[0]?.id??`prebuild`,_=await d.judgeByStaticAnalysis(e,o);if(_)return t.printTestCaseResult({testCaseId:g,..._}),{allAccepted:!1};let y=await u.findEntryPointFile(e,n.language);if(!y)return t.printTestCaseResult({testCaseId:g,decisionCode:s.DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,stderr:`main file not found${n.language?`: language: ${n.language}`:``}`}),{allAccepted:!1};let b=a.findLanguageDefinitionByPath(y);if(!b)return t.printTestCaseResult({testCaseId:g,decisionCode:s.DecisionCode.WRONG_ANSWER,stderr:`unsupported language`}),{allAccepted:!1};let x={...process.env,CI:``,FORCE_COLOR:`0`},S=y;if(i.buildSubmission!==!1){if(b.prebuild)try{await b.prebuild(e);let t=await u.findEntryPointFile(e,n.language);t&&(S=t)}catch(e){return t.printTestCaseResult({testCaseId:g,decisionCode:s.DecisionCode.BUILD_ERROR,stderr:(0,v.getErrorMessage)(e)}),{allAccepted:!1}}let r=b.buildCommand?.(S);if(r){let n=await C(r,{cwd:e,env:x,testCaseId:g,limits:m});if(n)return t.printTestCaseResult(n),{allAccepted:!1}}}let w=await c.snapshotWorkingDirectory(e);if(p.length===0)return t.printTestCaseResult({testCaseId:`default`,decisionCode:s.DecisionCode.ACCEPTED}),{allAccepted:!0};let O=p.shared?.fileInputPath;for(let n of p){O&&await l.copyTestCaseFileInput(O,e),n.fileInputPath&&await l.copyTestCaseFileInput(n.fileInputPath,e);let r=b.command(S),a=n.input??``,u;try{i.resolveInput&&(a=await i.resolveInput({testCase:n,cwd:e,env:x})),i.runCommand&&(u=await i.runCommand({testCase:n,command:r,stdin:a,cwd:e,env:x,timeLimitSeconds:h}))}catch(r){return t.printTestCaseResult({testCaseId:n.id,decisionCode:s.DecisionCode.RUNTIME_ERROR,stdin:a,stderr:(0,v.getErrorMessage)(r)}),await c.cleanWorkingDirectory(e,w),{allAccepted:!1}}let d=u??await E(r,{stdin:a,cwd:e,env:x,timeLimitSeconds:h}),p=await f.readOutputFiles(e,o.requiredOutputFilePaths??[]),g={timeLimitSeconds:h,outputLimitLength:m.maxOutputLength,problemMarkdownFrontMatter:{memoryLimitByte:o.memoryLimitByte,requiredOutputFilePaths:o.requiredOutputFilePaths}},_=D({runResult:d,outputFiles:p,context:g}),y=_;if(_.decisionCode===s.DecisionCode.ACCEPTED)try{let t=i.test?await i.test({testCase:n,runResult:d,outputFiles:p,cwd:e,context:g}):await T({testCase:n,runResult:d,outputFiles:p,cwd:e});t&&(y={decisionCode:t.decisionCode??_.decisionCode,feedbackMarkdown:t.feedbackMarkdown,stderr:t.stderr})}catch(e){y={decisionCode:s.DecisionCode.RUNTIME_ERROR,stderr:(0,v.getErrorMessage)(e)}}if(t.printTestCaseResult({testCaseId:n.id,...y,exitStatus:d.status,stdin:d.stdin,stdout:d.stdout.slice(0,m.maxOutputLength)||void 0,stderr:(y.stderr??d.stderr).slice(0,m.maxOutputLength)||void 0,timeSeconds:d.timeSeconds,cpuTimeSeconds:d.cpuTimeSeconds,memoryBytes:d.memoryBytes,outputFiles:p.length>0?p:void 0}),await c.cleanWorkingDirectory(e,w),y.decisionCode!==s.DecisionCode.ACCEPTED)return{allAccepted:!1}}return{allAccepted:!0}}function S(e,t){return t.allAccepted===(e.expectedResult===`accepted`)}async function C(e,t){let n=await h.spawnWithTimeout(e[0],e.slice(1),{cwd:t.cwd,env:t.env},t.limits.buildTimeoutSeconds),r=n.status??void 0;if(n.status!==0)return{testCaseId:t.testCaseId,decisionCode:s.DecisionCode.BUILD_ERROR,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes};if(n.timeSeconds>t.limits.buildTimeoutSeconds)return{testCaseId:t.testCaseId,decisionCode:s.DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes};if(n.outputLimitExceeded||n.stdout.length>t.limits.maxOutputLength||n.stderr.length>t.limits.maxOutputLength)return{testCaseId:t.testCaseId,decisionCode:s.DecisionCode.BUILD_OUTPUT_SIZE_LIMIT_EXCEEDED,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes}}async function w(e){return await i.readTestCases(g.default.join(e,`test_cases`))}async function T(e){let{testCase:t,runResult:r,outputFiles:i,cwd:a}=e,o=await n.judgeAgainstExpectations({stdout:r.stdout,expectedStdout:typeof t.output==`string`?t.output:void 0,fileOutputPath:typeof t.fileOutputPath==`string`?t.fileOutputPath:void 0,cwd:a,outputFiles:i});return i.splice(0,i.length,...o.outputFiles),o.matches?{}:{decisionCode:s.DecisionCode.WRONG_ANSWER}}async function E(e,t){let n=await h.spawnWithTimeout(e[0],e.slice(1),{cwd:t.cwd,stdin:t.stdin,env:t.env},t.timeLimitSeconds);return{stdin:t.stdin,stdout:n.stdout,stderr:n.stderr,status:n.status??void 0,timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes,outputLimitExceeded:n.outputLimitExceeded}}function D(e){if(e.runResult.status!==0)return{decisionCode:s.DecisionCode.RUNTIME_ERROR,stderr:e.runResult.stderr};if(e.runResult.timeSeconds>e.context.timeLimitSeconds)return{decisionCode:s.DecisionCode.TIME_LIMIT_EXCEEDED,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};if(e.runResult.outputLimitExceeded||e.runResult.stdout.length>e.context.outputLimitLength||e.runResult.stderr.length>e.context.outputLimitLength)return{decisionCode:s.DecisionCode.OUTPUT_SIZE_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}}exports.commandJudgePreset=b;
 //# sourceMappingURL=command.cjs.map

package/dist/presets/command.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"command.cjs","names":["z","parseArgs","resolveCwds","readProblemMarkdownFrontMatter","passesIsolationCheckInDebugMode","judgeByStaticAnalysis","findEntryPointFile","DecisionCode","findLanguageDefinitionByPath","snapshotWorkingDirectory","copyTestCaseFileInput","cleanWorkingDirectory","readOutputFiles","spawnWithTimeout","readFileTestCases","path","judgeAgainstExpectations"],"sources":["../../src/presets/command.ts"],"sourcesContent":["import path from 'node:path';\n\nimport { z } from 'zod';\n\nimport { cleanWorkingDirectory, snapshotWorkingDirectory } from '../helpers/cleanWorkingDirectory.js';\nimport { judgeAgainstExpectations } from '../helpers/compareExpectedOutputFiles.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 { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { readTestCases as readFileTestCases } from '../helpers/readTestCases.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { spawnWithTimeout } from '../helpers/spawnWithTimeout.js';\nimport { DecisionCode } from '../types/decisionCode.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 = 2;\nconst MAX_STDOUT_LENGTH = 50_000;\n\nconst judgeParamsSchema = z.object({\n  language: z.union([z.string(), z.array(z.string())]).optional(),\n});\n\ntype JudgeParams = z.infer<typeof judgeParamsSchema>;\n\n/** What every command test case needs; a custom `readTestCases` may add any fields of its own. */\ninterface BaseCommandTestCase {\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 CommandTestCase extends BaseCommandTestCase {\n  /** Expected standard output (`test_cases/<id>.out`). */\n  output?: string;\n  /** Directory of expected output files compared after the run (`test_cases/<id>.fout/`). */\n  fileOutputPath?: string;\n}\n\ntype CommandJudgeCaseResult = Pick<TestCaseResult, 'decisionCode' | 'feedbackMarkdown' | 'stderr'>;\n\nexport interface CommandRunResult {\n  stdin: string;\n  stdout: string;\n  stderr: string;\n  status: number | undefined;\n  timeSeconds: number;\n  /** See `TestCaseResult.cpuTimeSeconds`; a custom `runCommand` may leave it out. */\n  cpuTimeSeconds?: number;\n  memoryBytes: number;\n  outputLimitExceeded?: boolean;\n}\n\ninterface CommandJudgeContext {\n  timeLimitSeconds: number;\n  outputLimitLength: number;\n  problemMarkdownFrontMatter: Pick<ProblemMarkdownFrontMatter, 'memoryLimitByte' | 'requiredOutputFilePaths'>;\n}\n\nexport interface CommandJudgeLimits {\n  buildTimeoutSeconds: number;\n  maxOutputLength: number;\n}\n\ninterface ResolvedCommandProblem<TTestCase extends BaseCommandTestCase> {\n  problemMarkdownFrontMatter: ProblemMarkdownFrontMatter;\n  testCases: readonly TTestCase[];\n  limits: CommandJudgeLimits;\n  timeLimitSeconds: number;\n}\n\nexport interface CommandJudgePresetOptions<\n  TTestCase extends BaseCommandTestCase = CommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n> {\n  /**\n   * Set to false when the custom runner builds the submitted source itself.\n   * Defaults to true.\n   */\n  buildSubmission?: boolean;\n  limits?: CommandJudgeLimits;\n  /** The per-case time limit used only when `problem.md` declares no `timeLimitMs`. */\n  runTimeoutSeconds?: number;\n  readTestCases?: (problemDir: string) => Promise<readonly TTestCase[]>;\n  resolveInput?: (context: { testCase: TTestCase; cwd: string; env: NodeJS.ProcessEnv }) => Promise<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  }) => Promise<TRunResult> | TRunResult;\n  /**\n   * Decides the verdict of a run that passed the limit checks. It replaces the default comparison\n   * of `testCase.output` (space-separated tokens) and `testCase.fileOutputPath` (see\n   * `compareStdoutAsSpaceSeparatedTokens` and `compareExpectedOutputFiles`, both exported), so\n   * call those yourself when the test case ships `.out` / `.fout` expectations to keep.\n   * `outputFiles` is the array printed with the result and may be edited in place.\n   */\n  test?: (context: {\n    testCase: TTestCase;\n    runResult: TRunResult;\n    outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n    /** The submission's working directory, e.g. for `compareExpectedOutputFiles(cwd, testCase.fileOutputPath)`. */\n    cwd: string;\n    context: CommandJudgeContext;\n  }) => Promise<Partial<CommandJudgeCaseResult>> | Partial<CommandJudgeCaseResult> | undefined;\n}\n\n/**\n * A preset function for judging by executable command.\n *\n * Without options, test cases come from `test_cases/` (`.in`, `.out`, `.fin/`, `.fout/`) and a run\n * within the limits is accepted when its stdout matches `.out` and its files match `.fout/`; a\n * case without either expectation only has to run within the limits.\n * Keep problem-specific logic in `resolveInput` and `test`.\n *\n * @example\n * Create `judge.ts` that judges `test_cases/` like the default stdio harness, with the command\n * preset's debug mode:\n * ```ts\n * import { commandJudgePreset } from '@exercode/problem-utils/presets/command';\n *\n * await commandJudgePreset(import.meta.dirname);\n * ```\n *\n * @example\n * Create `judge.ts` with an own verdict:\n * ```ts\n * import { commandJudgePreset } from '@exercode/problem-utils/presets/command';\n * import { DecisionCode } from '@exercode/problem-utils';\n *\n * await commandJudgePreset(import.meta.dirname, {\n *   readTestCases: async () => [\n *     { id: '01', input: '1 2' },\n *   ],\n *   test: ({ runResult }) => {\n *     return runResult.stdout.trim() === '3'\n *       ? { decisionCode: DecisionCode.ACCEPTED }\n *       : { decisionCode: DecisionCode.WRONG_ANSWER };\n *   },\n * });\n * ```\n *\n * Run with the required parameters:\n * ```bash\n * bun judge.ts model_answers/python '{ \"language\": \"python\" }'\n * ```\n *\n * Run without a cwd argument to judge each `<problemDir>/model_answers/*` directory\n * for debugging. A prominent banner is printed for each run.\n * ```bash\n * bun judge.ts\n * ```\n */\nexport async function commandJudgePreset<\n  TTestCase extends BaseCommandTestCase = CommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n>(problemDir: string, options: CommandJudgePresetOptions<TTestCase, TRunResult> = {}): Promise<void> {\n  const args = parseArgs(process.argv);\n  const params = judgeParamsSchema.parse(args.params);\n\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n  const problemMarkdownFrontMatter = await readProblemMarkdownFrontMatter(problemDir);\n  const testCases = await (options.readTestCases ?? readCommandTestCases)(problemDir);\n  const limits = {\n    buildTimeoutSeconds: options.limits?.buildTimeoutSeconds ?? BUILD_TIMEOUT_SECONDS,\n    maxOutputLength: options.limits?.maxOutputLength ?? MAX_STDOUT_LENGTH,\n  };\n  const timeLimitSeconds =\n    typeof problemMarkdownFrontMatter.timeLimitMs === 'number'\n      ? problemMarkdownFrontMatter.timeLimitMs / 1000\n      : (options.runTimeoutSeconds ?? JUDGE_DEFAULT_TIMEOUT_SECONDS);\n  const problem: ResolvedCommandProblem<TTestCase> = {\n    problemMarkdownFrontMatter,\n    testCases,\n    limits,\n    timeLimitSeconds,\n  };\n\n  if (isDebugMode) {\n    const expectedMaxDurationMs = (limits.buildTimeoutSeconds + testCases.length * timeLimitSeconds) * 1000;\n    if (!(await passesIsolationCheckInDebugMode(problemDir, cwds, params, { expectedMaxDurationMs }))) {\n      process.exitCode = 1;\n      return;\n    }\n  }\n\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await runCommandJudgeForCwd<TTestCase, TRunResult>(resolvedCwd.cwd, params, problem, options);\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function runCommandJudgeForCwd<\n  TTestCase extends BaseCommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n>(\n  cwd: string,\n  params: JudgeParams,\n  problem: ResolvedCommandProblem<TTestCase>,\n  options: CommandJudgePresetOptions<TTestCase, TRunResult>\n): Promise<{ allAccepted: boolean }> {\n  const { problemMarkdownFrontMatter, testCases, limits, timeLimitSeconds } = problem;\n  const prebuildTestCaseId = testCases[0]?.id ?? 'prebuild';\n\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, problemMarkdownFrontMatter);\n  if (staticAnalysisResult) {\n    printTestCaseResult({ testCaseId: prebuildTestCaseId, ...staticAnalysisResult });\n    return { allAccepted: false };\n  }\n\n  const originalMainFilePath = await findEntryPointFile(cwd, params.language);\n  if (!originalMainFilePath) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      stderr: `main file not found${params.language ? `: language: ${params.language}` : ''}`,\n    });\n    return { allAccepted: false };\n  }\n\n  const languageDefinition = findLanguageDefinitionByPath(originalMainFilePath);\n  if (!languageDefinition) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      stderr: 'unsupported language',\n    });\n    return { allAccepted: false };\n  }\n\n  // `CI` changes affects Chainlit. `FORCE_COLOR` affects Bun.\n  const env = { ...process.env, CI: '', FORCE_COLOR: '0' };\n\n  let mainFilePath = originalMainFilePath;\n  if (options.buildSubmission !== false) {\n    if (languageDefinition.prebuild) {\n      try {\n        await languageDefinition.prebuild(cwd);\n        const prebuiltMainFilePath = await findEntryPointFile(cwd, params.language);\n        if (prebuiltMainFilePath) mainFilePath = prebuiltMainFilePath;\n      } catch (error) {\n        printTestCaseResult({\n          testCaseId: prebuildTestCaseId,\n          decisionCode: DecisionCode.BUILD_ERROR,\n          stderr: error instanceof Error ? error.message : String(error),\n        });\n        return { allAccepted: false };\n      }\n    }\n\n    const buildCommand = languageDefinition.buildCommand?.(mainFilePath);\n    if (buildCommand) {\n      const buildResult = await runBuild(buildCommand, {\n        cwd,\n        env,\n        testCaseId: prebuildTestCaseId,\n        limits,\n      });\n      if (buildResult) {\n        printTestCaseResult(buildResult);\n        return { allAccepted: false };\n      }\n    }\n  }\n\n  const cwdSnapshot = await snapshotWorkingDirectory(cwd);\n\n  if (testCases.length === 0) {\n    printTestCaseResult({ testCaseId: 'default', decisionCode: DecisionCode.ACCEPTED });\n    return { allAccepted: true };\n  }\n\n  const sharedFileInputPath = (testCases as { shared?: { fileInputPath?: string } }).shared?.fileInputPath;\n\n  for (const testCase of testCases) {\n    if (sharedFileInputPath) await copyTestCaseFileInput(sharedFileInputPath, cwd);\n    if (testCase.fileInputPath) await copyTestCaseFileInput(testCase.fileInputPath, cwd);\n\n    const command = languageDefinition.command(mainFilePath);\n    let stdin = testCase.input ?? '';\n    let customRunResult: TRunResult | undefined;\n    try {\n      if (options.resolveInput) {\n        stdin = await options.resolveInput({ testCase, cwd, env });\n      }\n\n      if (options.runCommand) {\n        customRunResult = await options.runCommand({ testCase, command, stdin, cwd, env, timeLimitSeconds });\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(cwd, cwdSnapshot);\n      return { allAccepted: false };\n    }\n\n    // The default runner reports the submission's failures in its result; what it throws is the\n    // judge's own failure and ends the run without a verdict.\n    const runResult =\n      customRunResult ?? ((await runCommand(command, { stdin, cwd, env, timeLimitSeconds })) as TRunResult);\n\n    const outputFiles = await readOutputFiles(cwd, problemMarkdownFrontMatter.requiredOutputFilePaths ?? []);\n    const judgeContext: CommandJudgeContext = {\n      timeLimitSeconds,\n      outputLimitLength: limits.maxOutputLength,\n      problemMarkdownFrontMatter: {\n        memoryLimitByte: problemMarkdownFrontMatter.memoryLimitByte,\n        requiredOutputFilePaths: problemMarkdownFrontMatter.requiredOutputFilePaths,\n      },\n    };\n    const baseJudgeResult = evaluateByLimits({ runResult, outputFiles, context: judgeContext });\n    let judgeResult = baseJudgeResult;\n    if (baseJudgeResult.decisionCode === DecisionCode.ACCEPTED) {\n      try {\n        const extendedJudgeResult = options.test\n          ? await options.test({ testCase, runResult, outputFiles, cwd, context: judgeContext })\n          : await compareWithExpectedOutputs({ testCase, runResult, outputFiles, cwd });\n        if (extendedJudgeResult) {\n          judgeResult = {\n            decisionCode: extendedJudgeResult.decisionCode ?? baseJudgeResult.decisionCode,\n            feedbackMarkdown: extendedJudgeResult.feedbackMarkdown,\n            stderr: extendedJudgeResult.stderr,\n          };\n        }\n      } catch (error) {\n        judgeResult = {\n          decisionCode: DecisionCode.RUNTIME_ERROR,\n          stderr: errorToMessage(error),\n        };\n      }\n    }\n\n    printTestCaseResult({\n      testCaseId: testCase.id,\n      ...judgeResult,\n      exitStatus: runResult.status,\n      stdin: runResult.stdin,\n      stdout: runResult.stdout.slice(0, limits.maxOutputLength) || undefined,\n      stderr: (judgeResult.stderr ?? runResult.stderr).slice(0, limits.maxOutputLength) || undefined,\n      timeSeconds: runResult.timeSeconds,\n      cpuTimeSeconds: runResult.cpuTimeSeconds,\n      memoryBytes: runResult.memoryBytes,\n      outputFiles: outputFiles.length > 0 ? outputFiles : undefined,\n    });\n\n    await cleanWorkingDirectory(cwd, cwdSnapshot);\n    if (judgeResult.decisionCode !== DecisionCode.ACCEPTED) return { allAccepted: false };\n  }\n\n  return { allAccepted: true };\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: { allAccepted: boolean }): boolean {\n  return result.allAccepted === (resolvedCwd.expectedResult === 'accepted');\n}\n\nasync function runBuild(\n  buildCommand: readonly [string, ...string[]],\n  context: { cwd: string; env: NodeJS.ProcessEnv; testCaseId: string; limits: CommandJudgeLimits }\n): Promise<(Omit<TestCaseResult, 'testCaseId'> & { testCaseId: string }) | undefined> {\n  const spawnResult = await spawnWithTimeout(\n    buildCommand[0],\n    buildCommand.slice(1),\n    { cwd: context.cwd, env: context.env },\n    context.limits.buildTimeoutSeconds\n  );\n  const exitStatus = spawnResult.status ?? undefined;\n\n  if (spawnResult.status !== 0) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_ERROR,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  if (spawnResult.timeSeconds > context.limits.buildTimeoutSeconds) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  if (\n    spawnResult.outputLimitExceeded ||\n    spawnResult.stdout.length > context.limits.maxOutputLength ||\n    spawnResult.stderr.length > context.limits.maxOutputLength\n  ) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_OUTPUT_SIZE_LIMIT_EXCEEDED,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  return;\n}\n\nfunction errorToMessage(error: unknown): string {\n  return error instanceof Error ? error.message : String(error);\n}\n\nasync function readCommandTestCases<TTestCase extends BaseCommandTestCase = CommandTestCase>(\n  problemDir: string\n): 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\n/** The default verdict: a case exposing a string `output` / `fileOutputPath` is judged against it. */\nasync function compareWithExpectedOutputs(context: {\n  testCase: BaseCommandTestCase & Partial<CommandTestCase>;\n  runResult: CommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  cwd: string;\n}): Promise<Partial<CommandJudgeCaseResult>> {\n  const { testCase, runResult, outputFiles, cwd } = context;\n  const judgement = await judgeAgainstExpectations({\n    stdout: runResult.stdout,\n    expectedStdout: typeof testCase.output === 'string' ? testCase.output : undefined,\n    fileOutputPath: typeof testCase.fileOutputPath === 'string' ? testCase.fileOutputPath : undefined,\n    cwd,\n    outputFiles,\n  });\n  outputFiles.splice(0, outputFiles.length, ...judgement.outputFiles);\n  return judgement.matches ? {} : { decisionCode: DecisionCode.WRONG_ANSWER };\n}\n\nasync function runCommand(\n  command: readonly [string, ...string[]],\n  context: {\n    stdin: string;\n    cwd: string;\n    env: NodeJS.ProcessEnv;\n    timeLimitSeconds: number;\n  }\n): Promise<CommandRunResult> {\n  const spawnResult = await spawnWithTimeout(\n    command[0],\n    command.slice(1),\n    { cwd: context.cwd, stdin: context.stdin, env: context.env },\n    context.timeLimitSeconds\n  );\n\n  return {\n    stdin: context.stdin,\n    stdout: spawnResult.stdout,\n    stderr: spawnResult.stderr,\n    status: spawnResult.status ?? undefined,\n    timeSeconds: spawnResult.timeSeconds,\n    cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n    memoryBytes: spawnResult.memoryBytes,\n    outputLimitExceeded: spawnResult.outputLimitExceeded,\n  };\n}\n\nfunction evaluateByLimits(context: {\n  runResult: CommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  context: CommandJudgeContext;\n}): CommandJudgeCaseResult {\n  if (context.runResult.status !== 0) {\n    return { decisionCode: DecisionCode.RUNTIME_ERROR, stderr: context.runResult.stderr };\n  }\n\n  if (context.runResult.timeSeconds > context.context.timeLimitSeconds) {\n    return { decisionCode: DecisionCode.TIME_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  if (\n    context.runResult.memoryBytes >\n    (context.context.problemMarkdownFrontMatter.memoryLimitByte ?? Number.POSITIVE_INFINITY)\n  ) {\n    return { decisionCode: DecisionCode.MEMORY_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  if (\n    context.runResult.outputLimitExceeded ||\n    context.runResult.stdout.length > context.context.outputLimitLength ||\n    context.runResult.stderr.length > context.context.outputLimitLength\n  ) {\n    return { decisionCode: DecisionCode.OUTPUT_SIZE_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  const requiredOutputFileCount = context.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length ?? 0;\n  if (context.outputFiles.length < requiredOutputFileCount) {\n    return { decisionCode: DecisionCode.MISSING_REQUIRED_OUTPUT_FILE_ERROR };\n  }\n\n  return { decisionCode: DecisionCode.ACCEPTED };\n}\n"],"mappings":"u5BA2BA,MAIM,EAAoBA,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,EA0ID,eAAsB,EAGpB,EAAoB,EAA4D,CAAC,EAAkB,CACnG,IAAM,EAAOC,EAAAA,UAAU,QAAQ,IAAI,EAC7B,EAAS,EAAkB,MAAM,EAAK,MAAM,EAE5C,CAAE,OAAM,eAAgB,MAAMC,EAAAA,YAAY,EAAY,EAAK,GAAG,EAC9D,EAA6B,MAAMC,EAAAA,+BAA+B,CAAU,EAC5E,EAAY,MAAO,EAAQ,eAAiB,EAAA,CAAsB,CAAU,EAC5E,EAAS,CACb,oBAAqB,EAAQ,QAAQ,qBAAuB,GAC5D,gBAAiB,EAAQ,QAAQ,iBAAmB,GACtD,EACM,EACJ,OAAO,EAA2B,aAAgB,SAC9C,EAA2B,YAAc,IACxC,EAAQ,mBAAqB,EAC9B,EAA6C,CACjD,6BACA,YACA,SACA,kBACF,EAEA,GAAI,EAAa,CACf,IAAM,GAAyB,EAAO,oBAAsB,EAAU,OAAS,GAAoB,IACnG,GAAI,CAAE,MAAMC,EAAAA,gCAAgC,EAAY,EAAM,EAAQ,CAAE,uBAAsB,CAAC,EAAI,CACjG,QAAQ,SAAW,EACnB,MACF,CACF,CAEA,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAA,oBAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAA6C,EAAY,IAAK,EAAQ,EAAS,CAAO,EACvG,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAA,mCAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EAIb,EACA,EACA,EACA,EACmC,CACnC,GAAM,CAAE,6BAA4B,YAAW,SAAQ,oBAAqB,EACtE,EAAqB,EAAU,EAAE,EAAE,IAAM,WAEzC,EAAuB,MAAMC,EAAAA,sBAAsB,EAAK,CAA0B,EACxF,GAAI,EAEF,OADA,EAAA,oBAAoB,CAAE,WAAY,EAAoB,GAAG,CAAqB,CAAC,EACxE,CAAE,YAAa,EAAM,EAG9B,IAAM,EAAuB,MAAMC,EAAAA,mBAAmB,EAAK,EAAO,QAAQ,EAC1E,GAAI,CAAC,EAMH,OALA,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcC,EAAAA,aAAa,uCAC3B,OAAQ,sBAAsB,EAAO,SAAW,eAAe,EAAO,WAAa,IACrF,CAAC,EACM,CAAE,YAAa,EAAM,EAG9B,IAAM,EAAqBC,EAAAA,6BAA6B,CAAoB,EAC5E,GAAI,CAAC,EAMH,OALA,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcD,EAAAA,aAAa,aAC3B,OAAQ,sBACV,CAAC,EACM,CAAE,YAAa,EAAM,EAI9B,IAAM,EAAM,CAAE,GAAG,QAAQ,IAAK,GAAI,GAAI,YAAa,GAAI,EAEnD,EAAe,EACnB,GAAI,EAAQ,kBAAoB,GAAO,CACrC,GAAI,EAAmB,SACrB,GAAI,CACF,MAAM,EAAmB,SAAS,CAAG,EACrC,IAAM,EAAuB,MAAMD,EAAAA,mBAAmB,EAAK,EAAO,QAAQ,EACtE,IAAsB,EAAe,EAC3C,OAAS,EAAO,CAMd,OALA,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcC,EAAAA,aAAa,YAC3B,OAAQ,aAAiB,MAAQ,EAAM,QAAU,OAAO,CAAK,CAC/D,CAAC,EACM,CAAE,YAAa,EAAM,CAC9B,CAGF,IAAM,EAAe,EAAmB,eAAe,CAAY,EACnE,GAAI,EAAc,CAChB,IAAM,EAAc,MAAM,EAAS,EAAc,CAC/C,MACA,MACA,WAAY,EACZ,QACF,CAAC,EACD,GAAI,EAEF,OADA,EAAA,oBAAoB,CAAW,EACxB,CAAE,YAAa,EAAM,CAEhC,CACF,CAEA,IAAM,EAAc,MAAME,EAAAA,yBAAyB,CAAG,EAEtD,GAAI,EAAU,SAAW,EAEvB,OADA,EAAA,oBAAoB,CAAE,WAAY,UAAW,aAAcF,EAAAA,aAAa,QAAS,CAAC,EAC3E,CAAE,YAAa,EAAK,EAG7B,IAAM,EAAuB,EAAsD,QAAQ,cAE3F,IAAK,IAAM,KAAY,EAAW,CAC5B,GAAqB,MAAMG,EAAAA,sBAAsB,EAAqB,CAAG,EACzE,EAAS,eAAe,MAAMA,EAAAA,sBAAsB,EAAS,cAAe,CAAG,EAEnF,IAAM,EAAU,EAAmB,QAAQ,CAAY,EACnD,EAAQ,EAAS,OAAS,GAC1B,EACJ,GAAI,CACE,EAAQ,eACV,EAAQ,MAAM,EAAQ,aAAa,CAAE,WAAU,MAAK,KAAI,CAAC,GAGvD,EAAQ,aACV,EAAkB,MAAM,EAAQ,WAAW,CAAE,WAAU,UAAS,QAAO,MAAK,MAAK,kBAAiB,CAAC,EAEvG,OAAS,EAAO,CAQd,OAPA,EAAA,oBAAoB,CAClB,WAAY,EAAS,GACrB,aAAcH,EAAAA,aAAa,cAC3B,QACA,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MAAMI,EAAAA,sBAAsB,EAAK,CAAW,EACrC,CAAE,YAAa,EAAM,CAC9B,CAIA,IAAM,EACJ,GAAqB,MAAM,EAAW,EAAS,CAAE,QAAO,MAAK,MAAK,kBAAiB,CAAC,EAEhF,EAAc,MAAMC,EAAAA,gBAAgB,EAAK,EAA2B,yBAA2B,CAAC,CAAC,EACjG,EAAoC,CACxC,mBACA,kBAAmB,EAAO,gBAC1B,2BAA4B,CAC1B,gBAAiB,EAA2B,gBAC5C,wBAAyB,EAA2B,uBACtD,CACF,EACM,EAAkB,EAAiB,CAAE,YAAW,cAAa,QAAS,CAAa,CAAC,EACtF,EAAc,EAClB,GAAI,EAAgB,eAAiBL,EAAAA,aAAa,SAChD,GAAI,CACF,IAAM,EAAsB,EAAQ,KAChC,MAAM,EAAQ,KAAK,CAAE,WAAU,YAAW,cAAa,MAAK,QAAS,CAAa,CAAC,EACnF,MAAM,EAA2B,CAAE,WAAU,YAAW,cAAa,KAAI,CAAC,EAC1E,IACF,EAAc,CACZ,aAAc,EAAoB,cAAgB,EAAgB,aAClE,iBAAkB,EAAoB,iBACtC,OAAQ,EAAoB,MAC9B,EAEJ,OAAS,EAAO,CACd,EAAc,CACZ,aAAcA,EAAAA,aAAa,cAC3B,OAAQ,EAAe,CAAK,CAC9B,CACF,CAiBF,GAdA,EAAA,oBAAoB,CAClB,WAAY,EAAS,GACrB,GAAG,EACH,WAAY,EAAU,OACtB,MAAO,EAAU,MACjB,OAAQ,EAAU,OAAO,MAAM,EAAG,EAAO,eAAe,GAAK,IAAA,GAC7D,QAAS,EAAY,QAAU,EAAU,OAAA,CAAQ,MAAM,EAAG,EAAO,eAAe,GAAK,IAAA,GACrF,YAAa,EAAU,YACvB,eAAgB,EAAU,eAC1B,YAAa,EAAU,YACvB,YAAa,EAAY,OAAS,EAAI,EAAc,IAAA,EACtD,CAAC,EAED,MAAMI,EAAAA,sBAAsB,EAAK,CAAW,EACxC,EAAY,eAAiBJ,EAAAA,aAAa,SAAU,MAAO,CAAE,YAAa,EAAM,CACtF,CAEA,MAAO,CAAE,YAAa,EAAK,CAC7B,CAEA,SAAS,EAAsB,EAA0B,EAA2C,CAClG,OAAO,EAAO,eAAiB,EAAY,iBAAmB,WAChE,CAEA,eAAe,EACb,EACA,EACoF,CACpF,IAAM,EAAc,MAAMM,EAAAA,iBACxB,EAAa,GACb,EAAa,MAAM,CAAC,EACpB,CAAE,IAAK,EAAQ,IAAK,IAAK,EAAQ,GAAI,EACrC,EAAQ,OAAO,mBACjB,EACM,EAAa,EAAY,QAAU,IAAA,GAEzC,GAAI,EAAY,SAAW,EACzB,MAAO,CACL,WAAY,EAAQ,WACpB,aAAcN,EAAAA,aAAa,YAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,EAGF,GAAI,EAAY,YAAc,EAAQ,OAAO,oBAC3C,MAAO,CACL,WAAY,EAAQ,WACpB,aAAcA,EAAAA,aAAa,0BAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,EAGF,GACE,EAAY,qBACZ,EAAY,OAAO,OAAS,EAAQ,OAAO,iBAC3C,EAAY,OAAO,OAAS,EAAQ,OAAO,gBAE3C,MAAO,CACL,WAAY,EAAQ,WACpB,aAAcA,EAAAA,aAAa,iCAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,CAIJ,CAEA,SAAS,EAAe,EAAwB,CAC9C,OAAO,aAAiB,MAAQ,EAAM,QAAU,OAAO,CAAK,CAC9D,CAEA,eAAe,EACb,EAC+B,CAE/B,OAAQ,MAAMO,EAAAA,cAAkBC,EAAAA,QAAK,KAAK,EAAY,YAAY,CAAC,CACrE,CAGA,eAAe,EAA2B,EAKG,CAC3C,GAAM,CAAE,WAAU,YAAW,cAAa,OAAQ,EAC5C,EAAY,MAAMC,EAAAA,yBAAyB,CAC/C,OAAQ,EAAU,OAClB,eAAgB,OAAO,EAAS,QAAW,SAAW,EAAS,OAAS,IAAA,GACxE,eAAgB,OAAO,EAAS,gBAAmB,SAAW,EAAS,eAAiB,IAAA,GACxF,MACA,aACF,CAAC,EAED,OADA,EAAY,OAAO,EAAG,EAAY,OAAQ,GAAG,EAAU,WAAW,EAC3D,EAAU,QAAU,CAAC,EAAI,CAAE,aAAcT,EAAAA,aAAa,YAAa,CAC5E,CAEA,eAAe,EACb,EACA,EAM2B,CAC3B,IAAM,EAAc,MAAMM,EAAAA,iBACxB,EAAQ,GACR,EAAQ,MAAM,CAAC,EACf,CAAE,IAAK,EAAQ,IAAK,MAAO,EAAQ,MAAO,IAAK,EAAQ,GAAI,EAC3D,EAAQ,gBACV,EAEA,MAAO,CACL,MAAO,EAAQ,MACf,OAAQ,EAAY,OACpB,OAAQ,EAAY,OACpB,OAAQ,EAAY,QAAU,IAAA,GAC9B,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,YACzB,oBAAqB,EAAY,mBACnC,CACF,CAEA,SAAS,EAAiB,EAIC,CACzB,GAAI,EAAQ,UAAU,SAAW,EAC/B,MAAO,CAAE,aAAcN,EAAAA,aAAa,cAAe,OAAQ,EAAQ,UAAU,MAAO,EAGtF,GAAI,EAAQ,UAAU,YAAc,EAAQ,QAAQ,iBAClD,MAAO,CAAE,aAAcA,EAAAA,aAAa,oBAAqB,OAAQ,EAAQ,UAAU,MAAO,EAG5F,GACE,EAAQ,UAAU,aACjB,EAAQ,QAAQ,2BAA2B,iBAAmB,KAE/D,MAAO,CAAE,aAAcA,EAAAA,aAAa,sBAAuB,OAAQ,EAAQ,UAAU,MAAO,EAG9F,GACE,EAAQ,UAAU,qBAClB,EAAQ,UAAU,OAAO,OAAS,EAAQ,QAAQ,mBAClD,EAAQ,UAAU,OAAO,OAAS,EAAQ,QAAQ,kBAElD,MAAO,CAAE,aAAcA,EAAAA,aAAa,2BAA4B,OAAQ,EAAQ,UAAU,MAAO,EAGnG,IAAM,EAA0B,EAAQ,QAAQ,2BAA2B,yBAAyB,QAAU,EAK9G,OAJI,EAAQ,YAAY,OAAS,EACxB,CAAE,aAAcA,EAAAA,aAAa,kCAAmC,EAGlE,CAAE,aAAcA,EAAAA,aAAa,QAAS,CAC/C"}
+{"version":3,"file":"command.cjs","names":["z","parseArgs","resolveCwds","readProblemMarkdownFrontMatter","passesIsolationCheckInDebugMode","judgeByStaticAnalysis","findEntryPointFile","DecisionCode","findLanguageDefinitionByPath","getErrorMessage","snapshotWorkingDirectory","copyTestCaseFileInput","cleanWorkingDirectory","readOutputFiles","spawnWithTimeout","readFileTestCases","path","judgeAgainstExpectations"],"sources":["../../src/presets/command.ts"],"sourcesContent":["import path from 'node:path';\nimport { getErrorMessage } from '@willbooster/shared-lib';\n\nimport { z } from 'zod';\n\nimport { cleanWorkingDirectory, snapshotWorkingDirectory } from '../helpers/cleanWorkingDirectory.js';\nimport { judgeAgainstExpectations } from '../helpers/compareExpectedOutputFiles.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 { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { readTestCases as readFileTestCases } from '../helpers/readTestCases.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { spawnWithTimeout } from '../helpers/spawnWithTimeout.js';\nimport { DecisionCode } from '../types/decisionCode.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 = 2;\nconst MAX_STDOUT_LENGTH = 50_000;\n\nconst judgeParamsSchema = z.object({\n  language: z.union([z.string(), z.array(z.string())]).optional(),\n});\n\ntype JudgeParams = z.infer<typeof judgeParamsSchema>;\n\n/** What every command test case needs; a custom `readTestCases` may add any fields of its own. */\ninterface BaseCommandTestCase {\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 CommandTestCase extends BaseCommandTestCase {\n  /** Expected standard output (`test_cases/<id>.out`). */\n  output?: string;\n  /** Directory of expected output files compared after the run (`test_cases/<id>.fout/`). */\n  fileOutputPath?: string;\n}\n\ntype CommandJudgeCaseResult = Pick<TestCaseResult, 'decisionCode' | 'feedbackMarkdown' | 'stderr'>;\n\nexport interface CommandRunResult {\n  stdin: string;\n  stdout: string;\n  stderr: string;\n  status: number | undefined;\n  timeSeconds: number;\n  /** See `TestCaseResult.cpuTimeSeconds`; a custom `runCommand` may leave it out. */\n  cpuTimeSeconds?: number;\n  memoryBytes: number;\n  outputLimitExceeded?: boolean;\n}\n\ninterface CommandJudgeContext {\n  timeLimitSeconds: number;\n  outputLimitLength: number;\n  problemMarkdownFrontMatter: Pick<ProblemMarkdownFrontMatter, 'memoryLimitByte' | 'requiredOutputFilePaths'>;\n}\n\nexport interface CommandJudgeLimits {\n  buildTimeoutSeconds: number;\n  maxOutputLength: number;\n}\n\ninterface ResolvedCommandProblem<TTestCase extends BaseCommandTestCase> {\n  problemMarkdownFrontMatter: ProblemMarkdownFrontMatter;\n  testCases: readonly TTestCase[];\n  limits: CommandJudgeLimits;\n  timeLimitSeconds: number;\n}\n\nexport interface CommandJudgePresetOptions<\n  TTestCase extends BaseCommandTestCase = CommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n> {\n  /**\n   * Set to false when the custom runner builds the submitted source itself.\n   * Defaults to true.\n   */\n  buildSubmission?: boolean;\n  limits?: CommandJudgeLimits;\n  /** The per-case time limit used only when `problem.md` declares no `timeLimitMs`. */\n  runTimeoutSeconds?: number;\n  readTestCases?: (problemDir: string) => Promise<readonly TTestCase[]>;\n  resolveInput?: (context: { testCase: TTestCase; cwd: string; env: NodeJS.ProcessEnv }) => Promise<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  }) => Promise<TRunResult> | TRunResult;\n  /**\n   * Decides the verdict of a run that passed the limit checks. It replaces the default comparison\n   * of `testCase.output` (space-separated tokens) and `testCase.fileOutputPath` (see\n   * `compareStdoutAsSpaceSeparatedTokens` and `compareExpectedOutputFiles`, both exported), so\n   * call those yourself when the test case ships `.out` / `.fout` expectations to keep.\n   * `outputFiles` is the array printed with the result and may be edited in place.\n   */\n  test?: (context: {\n    testCase: TTestCase;\n    runResult: TRunResult;\n    outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n    /** The submission's working directory, e.g. for `compareExpectedOutputFiles(cwd, testCase.fileOutputPath)`. */\n    cwd: string;\n    context: CommandJudgeContext;\n  }) => Promise<Partial<CommandJudgeCaseResult>> | Partial<CommandJudgeCaseResult> | undefined;\n}\n\n/**\n * A preset function for judging by executable command.\n *\n * Without options, test cases come from `test_cases/` (`.in`, `.out`, `.fin/`, `.fout/`) and a run\n * within the limits is accepted when its stdout matches `.out` and its files match `.fout/`; a\n * case without either expectation only has to run within the limits.\n * Keep problem-specific logic in `resolveInput` and `test`.\n *\n * @example\n * Create `judge.ts` that judges `test_cases/` like the default stdio harness, with the command\n * preset's debug mode:\n * ```ts\n * import { commandJudgePreset } from '@exercode/problem-utils/presets/command';\n *\n * await commandJudgePreset(import.meta.dirname);\n * ```\n *\n * @example\n * Create `judge.ts` with an own verdict:\n * ```ts\n * import { commandJudgePreset } from '@exercode/problem-utils/presets/command';\n * import { DecisionCode } from '@exercode/problem-utils';\n *\n * await commandJudgePreset(import.meta.dirname, {\n *   readTestCases: async () => [\n *     { id: '01', input: '1 2' },\n *   ],\n *   test: ({ runResult }) => {\n *     return runResult.stdout.trim() === '3'\n *       ? { decisionCode: DecisionCode.ACCEPTED }\n *       : { decisionCode: DecisionCode.WRONG_ANSWER };\n *   },\n * });\n * ```\n *\n * Run with the required parameters:\n * ```bash\n * bun judge.ts model_answers/python '{ \"language\": \"python\" }'\n * ```\n *\n * Run without a cwd argument to judge each `<problemDir>/model_answers/*` directory\n * for debugging. A prominent banner is printed for each run.\n * ```bash\n * bun judge.ts\n * ```\n */\nexport async function commandJudgePreset<\n  TTestCase extends BaseCommandTestCase = CommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n>(problemDir: string, options: CommandJudgePresetOptions<TTestCase, TRunResult> = {}): Promise<void> {\n  const args = parseArgs(process.argv);\n  const params = judgeParamsSchema.parse(args.params);\n\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n  const problemMarkdownFrontMatter = await readProblemMarkdownFrontMatter(problemDir);\n  const testCases = await (options.readTestCases ?? readCommandTestCases)(problemDir);\n  const limits = {\n    buildTimeoutSeconds: options.limits?.buildTimeoutSeconds ?? BUILD_TIMEOUT_SECONDS,\n    maxOutputLength: options.limits?.maxOutputLength ?? MAX_STDOUT_LENGTH,\n  };\n  const timeLimitSeconds =\n    typeof problemMarkdownFrontMatter.timeLimitMs === 'number'\n      ? problemMarkdownFrontMatter.timeLimitMs / 1000\n      : (options.runTimeoutSeconds ?? JUDGE_DEFAULT_TIMEOUT_SECONDS);\n  const problem: ResolvedCommandProblem<TTestCase> = {\n    problemMarkdownFrontMatter,\n    testCases,\n    limits,\n    timeLimitSeconds,\n  };\n\n  if (isDebugMode) {\n    const expectedMaxDurationMs = (limits.buildTimeoutSeconds + testCases.length * timeLimitSeconds) * 1000;\n    if (!(await passesIsolationCheckInDebugMode(problemDir, cwds, params, { expectedMaxDurationMs }))) {\n      process.exitCode = 1;\n      return;\n    }\n  }\n\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await runCommandJudgeForCwd<TTestCase, TRunResult>(resolvedCwd.cwd, params, problem, options);\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function runCommandJudgeForCwd<\n  TTestCase extends BaseCommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n>(\n  cwd: string,\n  params: JudgeParams,\n  problem: ResolvedCommandProblem<TTestCase>,\n  options: CommandJudgePresetOptions<TTestCase, TRunResult>\n): Promise<{ allAccepted: boolean }> {\n  const { problemMarkdownFrontMatter, testCases, limits, timeLimitSeconds } = problem;\n  const prebuildTestCaseId = testCases[0]?.id ?? 'prebuild';\n\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, problemMarkdownFrontMatter);\n  if (staticAnalysisResult) {\n    printTestCaseResult({ testCaseId: prebuildTestCaseId, ...staticAnalysisResult });\n    return { allAccepted: false };\n  }\n\n  const originalMainFilePath = await findEntryPointFile(cwd, params.language);\n  if (!originalMainFilePath) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      stderr: `main file not found${params.language ? `: language: ${params.language}` : ''}`,\n    });\n    return { allAccepted: false };\n  }\n\n  const languageDefinition = findLanguageDefinitionByPath(originalMainFilePath);\n  if (!languageDefinition) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      stderr: 'unsupported language',\n    });\n    return { allAccepted: false };\n  }\n\n  // `CI` changes affects Chainlit. `FORCE_COLOR` affects Bun.\n  const env = { ...process.env, CI: '', FORCE_COLOR: '0' };\n\n  let mainFilePath = originalMainFilePath;\n  if (options.buildSubmission !== false) {\n    if (languageDefinition.prebuild) {\n      try {\n        await languageDefinition.prebuild(cwd);\n        const prebuiltMainFilePath = await findEntryPointFile(cwd, params.language);\n        if (prebuiltMainFilePath) mainFilePath = prebuiltMainFilePath;\n      } catch (error) {\n        printTestCaseResult({\n          testCaseId: prebuildTestCaseId,\n          decisionCode: DecisionCode.BUILD_ERROR,\n          stderr: getErrorMessage(error),\n        });\n        return { allAccepted: false };\n      }\n    }\n\n    const buildCommand = languageDefinition.buildCommand?.(mainFilePath);\n    if (buildCommand) {\n      const buildResult = await runBuild(buildCommand, {\n        cwd,\n        env,\n        testCaseId: prebuildTestCaseId,\n        limits,\n      });\n      if (buildResult) {\n        printTestCaseResult(buildResult);\n        return { allAccepted: false };\n      }\n    }\n  }\n\n  const cwdSnapshot = await snapshotWorkingDirectory(cwd);\n\n  if (testCases.length === 0) {\n    printTestCaseResult({ testCaseId: 'default', decisionCode: DecisionCode.ACCEPTED });\n    return { allAccepted: true };\n  }\n\n  const sharedFileInputPath = (testCases as { shared?: { fileInputPath?: string } }).shared?.fileInputPath;\n\n  for (const testCase of testCases) {\n    if (sharedFileInputPath) await copyTestCaseFileInput(sharedFileInputPath, cwd);\n    if (testCase.fileInputPath) await copyTestCaseFileInput(testCase.fileInputPath, cwd);\n\n    const command = languageDefinition.command(mainFilePath);\n    let stdin = testCase.input ?? '';\n    let customRunResult: TRunResult | undefined;\n    try {\n      if (options.resolveInput) {\n        stdin = await options.resolveInput({ testCase, cwd, env });\n      }\n\n      if (options.runCommand) {\n        customRunResult = await options.runCommand({ testCase, command, stdin, cwd, env, timeLimitSeconds });\n      }\n    } catch (error) {\n      printTestCaseResult({\n        testCaseId: testCase.id,\n        decisionCode: DecisionCode.RUNTIME_ERROR,\n        stdin,\n        stderr: getErrorMessage(error),\n      });\n      await cleanWorkingDirectory(cwd, cwdSnapshot);\n      return { allAccepted: false };\n    }\n\n    // The default runner reports the submission's failures in its result; what it throws is the\n    // judge's own failure and ends the run without a verdict.\n    const runResult =\n      customRunResult ?? ((await runCommand(command, { stdin, cwd, env, timeLimitSeconds })) as TRunResult);\n\n    const outputFiles = await readOutputFiles(cwd, problemMarkdownFrontMatter.requiredOutputFilePaths ?? []);\n    const judgeContext: CommandJudgeContext = {\n      timeLimitSeconds,\n      outputLimitLength: limits.maxOutputLength,\n      problemMarkdownFrontMatter: {\n        memoryLimitByte: problemMarkdownFrontMatter.memoryLimitByte,\n        requiredOutputFilePaths: problemMarkdownFrontMatter.requiredOutputFilePaths,\n      },\n    };\n    const baseJudgeResult = evaluateByLimits({ runResult, outputFiles, context: judgeContext });\n    let judgeResult = baseJudgeResult;\n    if (baseJudgeResult.decisionCode === DecisionCode.ACCEPTED) {\n      try {\n        const extendedJudgeResult = options.test\n          ? await options.test({ testCase, runResult, outputFiles, cwd, context: judgeContext })\n          : await compareWithExpectedOutputs({ testCase, runResult, outputFiles, cwd });\n        if (extendedJudgeResult) {\n          judgeResult = {\n            decisionCode: extendedJudgeResult.decisionCode ?? baseJudgeResult.decisionCode,\n            feedbackMarkdown: extendedJudgeResult.feedbackMarkdown,\n            stderr: extendedJudgeResult.stderr,\n          };\n        }\n      } catch (error) {\n        judgeResult = {\n          decisionCode: DecisionCode.RUNTIME_ERROR,\n          stderr: getErrorMessage(error),\n        };\n      }\n    }\n\n    printTestCaseResult({\n      testCaseId: testCase.id,\n      ...judgeResult,\n      exitStatus: runResult.status,\n      stdin: runResult.stdin,\n      stdout: runResult.stdout.slice(0, limits.maxOutputLength) || undefined,\n      stderr: (judgeResult.stderr ?? runResult.stderr).slice(0, limits.maxOutputLength) || undefined,\n      timeSeconds: runResult.timeSeconds,\n      cpuTimeSeconds: runResult.cpuTimeSeconds,\n      memoryBytes: runResult.memoryBytes,\n      outputFiles: outputFiles.length > 0 ? outputFiles : undefined,\n    });\n\n    await cleanWorkingDirectory(cwd, cwdSnapshot);\n    if (judgeResult.decisionCode !== DecisionCode.ACCEPTED) return { allAccepted: false };\n  }\n\n  return { allAccepted: true };\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: { allAccepted: boolean }): boolean {\n  return result.allAccepted === (resolvedCwd.expectedResult === 'accepted');\n}\n\nasync function runBuild(\n  buildCommand: readonly [string, ...string[]],\n  context: { cwd: string; env: NodeJS.ProcessEnv; testCaseId: string; limits: CommandJudgeLimits }\n): Promise<(Omit<TestCaseResult, 'testCaseId'> & { testCaseId: string }) | undefined> {\n  const spawnResult = await spawnWithTimeout(\n    buildCommand[0],\n    buildCommand.slice(1),\n    { cwd: context.cwd, env: context.env },\n    context.limits.buildTimeoutSeconds\n  );\n  const exitStatus = spawnResult.status ?? undefined;\n\n  if (spawnResult.status !== 0) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_ERROR,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  if (spawnResult.timeSeconds > context.limits.buildTimeoutSeconds) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  if (\n    spawnResult.outputLimitExceeded ||\n    spawnResult.stdout.length > context.limits.maxOutputLength ||\n    spawnResult.stderr.length > context.limits.maxOutputLength\n  ) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_OUTPUT_SIZE_LIMIT_EXCEEDED,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  return;\n}\n\nasync function readCommandTestCases<TTestCase extends BaseCommandTestCase = CommandTestCase>(\n  problemDir: string\n): 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\n/** The default verdict: a case exposing a string `output` / `fileOutputPath` is judged against it. */\nasync function compareWithExpectedOutputs(context: {\n  testCase: BaseCommandTestCase & Partial<CommandTestCase>;\n  runResult: CommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  cwd: string;\n}): Promise<Partial<CommandJudgeCaseResult>> {\n  const { testCase, runResult, outputFiles, cwd } = context;\n  const judgement = await judgeAgainstExpectations({\n    stdout: runResult.stdout,\n    expectedStdout: typeof testCase.output === 'string' ? testCase.output : undefined,\n    fileOutputPath: typeof testCase.fileOutputPath === 'string' ? testCase.fileOutputPath : undefined,\n    cwd,\n    outputFiles,\n  });\n  outputFiles.splice(0, outputFiles.length, ...judgement.outputFiles);\n  return judgement.matches ? {} : { decisionCode: DecisionCode.WRONG_ANSWER };\n}\n\nasync function runCommand(\n  command: readonly [string, ...string[]],\n  context: {\n    stdin: string;\n    cwd: string;\n    env: NodeJS.ProcessEnv;\n    timeLimitSeconds: number;\n  }\n): Promise<CommandRunResult> {\n  const spawnResult = await spawnWithTimeout(\n    command[0],\n    command.slice(1),\n    { cwd: context.cwd, stdin: context.stdin, env: context.env },\n    context.timeLimitSeconds\n  );\n\n  return {\n    stdin: context.stdin,\n    stdout: spawnResult.stdout,\n    stderr: spawnResult.stderr,\n    status: spawnResult.status ?? undefined,\n    timeSeconds: spawnResult.timeSeconds,\n    cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n    memoryBytes: spawnResult.memoryBytes,\n    outputLimitExceeded: spawnResult.outputLimitExceeded,\n  };\n}\n\nfunction evaluateByLimits(context: {\n  runResult: CommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  context: CommandJudgeContext;\n}): CommandJudgeCaseResult {\n  if (context.runResult.status !== 0) {\n    return { decisionCode: DecisionCode.RUNTIME_ERROR, stderr: context.runResult.stderr };\n  }\n\n  if (context.runResult.timeSeconds > context.context.timeLimitSeconds) {\n    return { decisionCode: DecisionCode.TIME_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  if (\n    context.runResult.memoryBytes >\n    (context.context.problemMarkdownFrontMatter.memoryLimitByte ?? Number.POSITIVE_INFINITY)\n  ) {\n    return { decisionCode: DecisionCode.MEMORY_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  if (\n    context.runResult.outputLimitExceeded ||\n    context.runResult.stdout.length > context.context.outputLimitLength ||\n    context.runResult.stderr.length > context.context.outputLimitLength\n  ) {\n    return { decisionCode: DecisionCode.OUTPUT_SIZE_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  const requiredOutputFileCount = context.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length ?? 0;\n  if (context.outputFiles.length < requiredOutputFileCount) {\n    return { decisionCode: DecisionCode.MISSING_REQUIRED_OUTPUT_FILE_ERROR };\n  }\n\n  return { decisionCode: DecisionCode.ACCEPTED };\n}\n"],"mappings":"47BA4BA,MAIM,EAAoBA,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,EA0ID,eAAsB,EAGpB,EAAoB,EAA4D,CAAC,EAAkB,CACnG,IAAM,EAAOC,EAAAA,UAAU,QAAQ,IAAI,EAC7B,EAAS,EAAkB,MAAM,EAAK,MAAM,EAE5C,CAAE,OAAM,eAAgB,MAAMC,EAAAA,YAAY,EAAY,EAAK,GAAG,EAC9D,EAA6B,MAAMC,EAAAA,+BAA+B,CAAU,EAC5E,EAAY,MAAO,EAAQ,eAAiB,EAAA,CAAsB,CAAU,EAC5E,EAAS,CACb,oBAAqB,EAAQ,QAAQ,qBAAuB,GAC5D,gBAAiB,EAAQ,QAAQ,iBAAmB,GACtD,EACM,EACJ,OAAO,EAA2B,aAAgB,SAC9C,EAA2B,YAAc,IACxC,EAAQ,mBAAqB,EAC9B,EAA6C,CACjD,6BACA,YACA,SACA,kBACF,EAEA,GAAI,EAAa,CACf,IAAM,GAAyB,EAAO,oBAAsB,EAAU,OAAS,GAAoB,IACnG,GAAI,CAAE,MAAMC,EAAAA,gCAAgC,EAAY,EAAM,EAAQ,CAAE,uBAAsB,CAAC,EAAI,CACjG,QAAQ,SAAW,EACnB,MACF,CACF,CAEA,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAA,oBAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAA6C,EAAY,IAAK,EAAQ,EAAS,CAAO,EACvG,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAA,mCAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EAIb,EACA,EACA,EACA,EACmC,CACnC,GAAM,CAAE,6BAA4B,YAAW,SAAQ,oBAAqB,EACtE,EAAqB,EAAU,EAAE,EAAE,IAAM,WAEzC,EAAuB,MAAMC,EAAAA,sBAAsB,EAAK,CAA0B,EACxF,GAAI,EAEF,OADA,EAAA,oBAAoB,CAAE,WAAY,EAAoB,GAAG,CAAqB,CAAC,EACxE,CAAE,YAAa,EAAM,EAG9B,IAAM,EAAuB,MAAMC,EAAAA,mBAAmB,EAAK,EAAO,QAAQ,EAC1E,GAAI,CAAC,EAMH,OALA,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcC,EAAAA,aAAa,uCAC3B,OAAQ,sBAAsB,EAAO,SAAW,eAAe,EAAO,WAAa,IACrF,CAAC,EACM,CAAE,YAAa,EAAM,EAG9B,IAAM,EAAqBC,EAAAA,6BAA6B,CAAoB,EAC5E,GAAI,CAAC,EAMH,OALA,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcD,EAAAA,aAAa,aAC3B,OAAQ,sBACV,CAAC,EACM,CAAE,YAAa,EAAM,EAI9B,IAAM,EAAM,CAAE,GAAG,QAAQ,IAAK,GAAI,GAAI,YAAa,GAAI,EAEnD,EAAe,EACnB,GAAI,EAAQ,kBAAoB,GAAO,CACrC,GAAI,EAAmB,SACrB,GAAI,CACF,MAAM,EAAmB,SAAS,CAAG,EACrC,IAAM,EAAuB,MAAMD,EAAAA,mBAAmB,EAAK,EAAO,QAAQ,EACtE,IAAsB,EAAe,EAC3C,OAAS,EAAO,CAMd,OALA,EAAA,oBAAoB,CAClB,WAAY,EACZ,aAAcC,EAAAA,aAAa,YAC3B,QAAA,EAAQE,EAAAA,gBAAAA,CAAgB,CAAK,CAC/B,CAAC,EACM,CAAE,YAAa,EAAM,CAC9B,CAGF,IAAM,EAAe,EAAmB,eAAe,CAAY,EACnE,GAAI,EAAc,CAChB,IAAM,EAAc,MAAM,EAAS,EAAc,CAC/C,MACA,MACA,WAAY,EACZ,QACF,CAAC,EACD,GAAI,EAEF,OADA,EAAA,oBAAoB,CAAW,EACxB,CAAE,YAAa,EAAM,CAEhC,CACF,CAEA,IAAM,EAAc,MAAMC,EAAAA,yBAAyB,CAAG,EAEtD,GAAI,EAAU,SAAW,EAEvB,OADA,EAAA,oBAAoB,CAAE,WAAY,UAAW,aAAcH,EAAAA,aAAa,QAAS,CAAC,EAC3E,CAAE,YAAa,EAAK,EAG7B,IAAM,EAAuB,EAAsD,QAAQ,cAE3F,IAAK,IAAM,KAAY,EAAW,CAC5B,GAAqB,MAAMI,EAAAA,sBAAsB,EAAqB,CAAG,EACzE,EAAS,eAAe,MAAMA,EAAAA,sBAAsB,EAAS,cAAe,CAAG,EAEnF,IAAM,EAAU,EAAmB,QAAQ,CAAY,EACnD,EAAQ,EAAS,OAAS,GAC1B,EACJ,GAAI,CACE,EAAQ,eACV,EAAQ,MAAM,EAAQ,aAAa,CAAE,WAAU,MAAK,KAAI,CAAC,GAGvD,EAAQ,aACV,EAAkB,MAAM,EAAQ,WAAW,CAAE,WAAU,UAAS,QAAO,MAAK,MAAK,kBAAiB,CAAC,EAEvG,OAAS,EAAO,CAQd,OAPA,EAAA,oBAAoB,CAClB,WAAY,EAAS,GACrB,aAAcJ,EAAAA,aAAa,cAC3B,QACA,QAAA,EAAQE,EAAAA,gBAAAA,CAAgB,CAAK,CAC/B,CAAC,EACD,MAAMG,EAAAA,sBAAsB,EAAK,CAAW,EACrC,CAAE,YAAa,EAAM,CAC9B,CAIA,IAAM,EACJ,GAAqB,MAAM,EAAW,EAAS,CAAE,QAAO,MAAK,MAAK,kBAAiB,CAAC,EAEhF,EAAc,MAAMC,EAAAA,gBAAgB,EAAK,EAA2B,yBAA2B,CAAC,CAAC,EACjG,EAAoC,CACxC,mBACA,kBAAmB,EAAO,gBAC1B,2BAA4B,CAC1B,gBAAiB,EAA2B,gBAC5C,wBAAyB,EAA2B,uBACtD,CACF,EACM,EAAkB,EAAiB,CAAE,YAAW,cAAa,QAAS,CAAa,CAAC,EACtF,EAAc,EAClB,GAAI,EAAgB,eAAiBN,EAAAA,aAAa,SAChD,GAAI,CACF,IAAM,EAAsB,EAAQ,KAChC,MAAM,EAAQ,KAAK,CAAE,WAAU,YAAW,cAAa,MAAK,QAAS,CAAa,CAAC,EACnF,MAAM,EAA2B,CAAE,WAAU,YAAW,cAAa,KAAI,CAAC,EAC1E,IACF,EAAc,CACZ,aAAc,EAAoB,cAAgB,EAAgB,aAClE,iBAAkB,EAAoB,iBACtC,OAAQ,EAAoB,MAC9B,EAEJ,OAAS,EAAO,CACd,EAAc,CACZ,aAAcA,EAAAA,aAAa,cAC3B,QAAA,EAAQE,EAAAA,gBAAAA,CAAgB,CAAK,CAC/B,CACF,CAiBF,GAdA,EAAA,oBAAoB,CAClB,WAAY,EAAS,GACrB,GAAG,EACH,WAAY,EAAU,OACtB,MAAO,EAAU,MACjB,OAAQ,EAAU,OAAO,MAAM,EAAG,EAAO,eAAe,GAAK,IAAA,GAC7D,QAAS,EAAY,QAAU,EAAU,OAAA,CAAQ,MAAM,EAAG,EAAO,eAAe,GAAK,IAAA,GACrF,YAAa,EAAU,YACvB,eAAgB,EAAU,eAC1B,YAAa,EAAU,YACvB,YAAa,EAAY,OAAS,EAAI,EAAc,IAAA,EACtD,CAAC,EAED,MAAMG,EAAAA,sBAAsB,EAAK,CAAW,EACxC,EAAY,eAAiBL,EAAAA,aAAa,SAAU,MAAO,CAAE,YAAa,EAAM,CACtF,CAEA,MAAO,CAAE,YAAa,EAAK,CAC7B,CAEA,SAAS,EAAsB,EAA0B,EAA2C,CAClG,OAAO,EAAO,eAAiB,EAAY,iBAAmB,WAChE,CAEA,eAAe,EACb,EACA,EACoF,CACpF,IAAM,EAAc,MAAMO,EAAAA,iBACxB,EAAa,GACb,EAAa,MAAM,CAAC,EACpB,CAAE,IAAK,EAAQ,IAAK,IAAK,EAAQ,GAAI,EACrC,EAAQ,OAAO,mBACjB,EACM,EAAa,EAAY,QAAU,IAAA,GAEzC,GAAI,EAAY,SAAW,EACzB,MAAO,CACL,WAAY,EAAQ,WACpB,aAAcP,EAAAA,aAAa,YAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,EAGF,GAAI,EAAY,YAAc,EAAQ,OAAO,oBAC3C,MAAO,CACL,WAAY,EAAQ,WACpB,aAAcA,EAAAA,aAAa,0BAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,EAGF,GACE,EAAY,qBACZ,EAAY,OAAO,OAAS,EAAQ,OAAO,iBAC3C,EAAY,OAAO,OAAS,EAAQ,OAAO,gBAE3C,MAAO,CACL,WAAY,EAAQ,WACpB,aAAcA,EAAAA,aAAa,iCAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,CAIJ,CAEA,eAAe,EACb,EAC+B,CAE/B,OAAQ,MAAMQ,EAAAA,cAAkBC,EAAAA,QAAK,KAAK,EAAY,YAAY,CAAC,CACrE,CAGA,eAAe,EAA2B,EAKG,CAC3C,GAAM,CAAE,WAAU,YAAW,cAAa,OAAQ,EAC5C,EAAY,MAAMC,EAAAA,yBAAyB,CAC/C,OAAQ,EAAU,OAClB,eAAgB,OAAO,EAAS,QAAW,SAAW,EAAS,OAAS,IAAA,GACxE,eAAgB,OAAO,EAAS,gBAAmB,SAAW,EAAS,eAAiB,IAAA,GACxF,MACA,aACF,CAAC,EAED,OADA,EAAY,OAAO,EAAG,EAAY,OAAQ,GAAG,EAAU,WAAW,EAC3D,EAAU,QAAU,CAAC,EAAI,CAAE,aAAcV,EAAAA,aAAa,YAAa,CAC5E,CAEA,eAAe,EACb,EACA,EAM2B,CAC3B,IAAM,EAAc,MAAMO,EAAAA,iBACxB,EAAQ,GACR,EAAQ,MAAM,CAAC,EACf,CAAE,IAAK,EAAQ,IAAK,MAAO,EAAQ,MAAO,IAAK,EAAQ,GAAI,EAC3D,EAAQ,gBACV,EAEA,MAAO,CACL,MAAO,EAAQ,MACf,OAAQ,EAAY,OACpB,OAAQ,EAAY,OACpB,OAAQ,EAAY,QAAU,IAAA,GAC9B,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,YACzB,oBAAqB,EAAY,mBACnC,CACF,CAEA,SAAS,EAAiB,EAIC,CACzB,GAAI,EAAQ,UAAU,SAAW,EAC/B,MAAO,CAAE,aAAcP,EAAAA,aAAa,cAAe,OAAQ,EAAQ,UAAU,MAAO,EAGtF,GAAI,EAAQ,UAAU,YAAc,EAAQ,QAAQ,iBAClD,MAAO,CAAE,aAAcA,EAAAA,aAAa,oBAAqB,OAAQ,EAAQ,UAAU,MAAO,EAG5F,GACE,EAAQ,UAAU,aACjB,EAAQ,QAAQ,2BAA2B,iBAAmB,KAE/D,MAAO,CAAE,aAAcA,EAAAA,aAAa,sBAAuB,OAAQ,EAAQ,UAAU,MAAO,EAG9F,GACE,EAAQ,UAAU,qBAClB,EAAQ,UAAU,OAAO,OAAS,EAAQ,QAAQ,mBAClD,EAAQ,UAAU,OAAO,OAAS,EAAQ,QAAQ,kBAElD,MAAO,CAAE,aAAcA,EAAAA,aAAa,2BAA4B,OAAQ,EAAQ,UAAU,MAAO,EAGnG,IAAM,EAA0B,EAAQ,QAAQ,2BAA2B,yBAAyB,QAAU,EAK9G,OAJI,EAAQ,YAAY,OAAS,EACxB,CAAE,aAAcA,EAAAA,aAAa,kCAAmC,EAGlE,CAAE,aAAcA,EAAAA,aAAa,QAAS,CAC/C"}

package/dist/presets/command.js

@@ -1,2 +1,2 @@
-import{printTestCaseResult as e}from"../helpers/printTestCaseResult.js";import{judgeAgainstExpectations as t}from"../helpers/compareExpectedOutputFiles.js";import{parseArgs as n}from"../helpers/parseArgs.js";import{readTestCases as r}from"../helpers/readTestCases.js";import{findLanguageDefinitionByPath as i}from"../helpers/findLanguageDefinitionByPath.js";import{readProblemMarkdownFrontMatter as a}from"../helpers/readProblemMarkdownFrontMatter.js";import{DecisionCode as o}from"../types/decisionCode.js";import{cleanWorkingDirectory as s,snapshotWorkingDirectory as c}from"../helpers/cleanWorkingDirectory.js";import{copyTestCaseFileInput as l}from"../helpers/copyTestCaseFileInput.js";import{findEntryPointFile as u}from"../helpers/findEntryPointFile.js";import{judgeByStaticAnalysis as d}from"../helpers/judgeByStaticAnalysis.js";import{readOutputFiles as f}from"../helpers/readOutputFiles.js";import{passesIsolationCheckInDebugMode as p}from"../helpers/runIsolationCheckInDebugMode.js";import{printDebugCwdBanner as m,printDebugExpectationFailureBanner as h,resolveCwds as g}from"../helpers/resolveCwds.js";import{spawnWithTimeout as _}from"../helpers/spawnWithTimeout.js";import v from"node:path";import{z as y}from"zod";const b=y.object({language:y.union([y.string(),y.array(y.string())]).optional()});async function x(e,t={}){let r=n(process.argv),i=b.parse(r.params),{cwds:o,isDebugMode:s}=await g(e,r.cwd),c=await a(e),l=await(t.readTestCases??E)(e),u={buildTimeoutSeconds:t.limits?.buildTimeoutSeconds??10,maxOutputLength:t.limits?.maxOutputLength??5e4},d=typeof c.timeLimitMs==`number`?c.timeLimitMs/1e3:t.runTimeoutSeconds??2,f={problemMarkdownFrontMatter:c,testCases:l,limits:u,timeLimitSeconds:d};if(s){let t=(u.buildTimeoutSeconds+l.length*d)*1e3;if(!await p(e,o,i,{expectedMaxDurationMs:t})){process.exitCode=1;return}}for(let n of o){s&&m(e,n);let r=await S(n.cwd,i,f,t);s&&!C(n,r)&&(process.exitCode=1,h(e,n))}}async function S(t,n,r,a){let{problemMarkdownFrontMatter:p,testCases:m,limits:h,timeLimitSeconds:g}=r,_=m[0]?.id??`prebuild`,v=await d(t,p);if(v)return e({testCaseId:_,...v}),{allAccepted:!1};let y=await u(t,n.language);if(!y)return e({testCaseId:_,decisionCode:o.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,stderr:`main file not found${n.language?`: language: ${n.language}`:``}`}),{allAccepted:!1};let b=i(y);if(!b)return e({testCaseId:_,decisionCode:o.WRONG_ANSWER,stderr:`unsupported language`}),{allAccepted:!1};let x={...process.env,CI:``,FORCE_COLOR:`0`},S=y;if(a.buildSubmission!==!1){if(b.prebuild)try{await b.prebuild(t);let e=await u(t,n.language);e&&(S=e)}catch(t){return e({testCaseId:_,decisionCode:o.BUILD_ERROR,stderr:t instanceof Error?t.message:String(t)}),{allAccepted:!1}}let r=b.buildCommand?.(S);if(r){let n=await w(r,{cwd:t,env:x,testCaseId:_,limits:h});if(n)return e(n),{allAccepted:!1}}}let C=await c(t);if(m.length===0)return e({testCaseId:`default`,decisionCode:o.ACCEPTED}),{allAccepted:!0};let E=m.shared?.fileInputPath;for(let n of m){E&&await l(E,t),n.fileInputPath&&await l(n.fileInputPath,t);let r=b.command(S),i=n.input??``,c;try{a.resolveInput&&(i=await a.resolveInput({testCase:n,cwd:t,env:x})),a.runCommand&&(c=await a.runCommand({testCase:n,command:r,stdin:i,cwd:t,env:x,timeLimitSeconds:g}))}catch(r){return e({testCaseId:n.id,decisionCode:o.RUNTIME_ERROR,stdin:i,stderr:T(r)}),await s(t,C),{allAccepted:!1}}let u=c??await O(r,{stdin:i,cwd:t,env:x,timeLimitSeconds:g}),d=await f(t,p.requiredOutputFilePaths??[]),m={timeLimitSeconds:g,outputLimitLength:h.maxOutputLength,problemMarkdownFrontMatter:{memoryLimitByte:p.memoryLimitByte,requiredOutputFilePaths:p.requiredOutputFilePaths}},_=k({runResult:u,outputFiles:d,context:m}),v=_;if(_.decisionCode===o.ACCEPTED)try{let e=a.test?await a.test({testCase:n,runResult:u,outputFiles:d,cwd:t,context:m}):await D({testCase:n,runResult:u,outputFiles:d,cwd:t});e&&(v={decisionCode:e.decisionCode??_.decisionCode,feedbackMarkdown:e.feedbackMarkdown,stderr:e.stderr})}catch(e){v={decisionCode:o.RUNTIME_ERROR,stderr:T(e)}}if(e({testCaseId:n.id,...v,exitStatus:u.status,stdin:u.stdin,stdout:u.stdout.slice(0,h.maxOutputLength)||void 0,stderr:(v.stderr??u.stderr).slice(0,h.maxOutputLength)||void 0,timeSeconds:u.timeSeconds,cpuTimeSeconds:u.cpuTimeSeconds,memoryBytes:u.memoryBytes,outputFiles:d.length>0?d:void 0}),await s(t,C),v.decisionCode!==o.ACCEPTED)return{allAccepted:!1}}return{allAccepted:!0}}function C(e,t){return t.allAccepted===(e.expectedResult===`accepted`)}async function w(e,t){let n=await _(e[0],e.slice(1),{cwd:t.cwd,env:t.env},t.limits.buildTimeoutSeconds),r=n.status??void 0;if(n.status!==0)return{testCaseId:t.testCaseId,decisionCode:o.BUILD_ERROR,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes};if(n.timeSeconds>t.limits.buildTimeoutSeconds)return{testCaseId:t.testCaseId,decisionCode:o.BUILD_TIME_LIMIT_EXCEEDED,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes};if(n.outputLimitExceeded||n.stdout.length>t.limits.maxOutputLength||n.stderr.length>t.limits.maxOutputLength)return{testCaseId:t.testCaseId,decisionCode:o.BUILD_OUTPUT_SIZE_LIMIT_EXCEEDED,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes}}function T(e){return e instanceof Error?e.message:String(e)}async function E(e){return await r(v.join(e,`test_cases`))}async function D(e){let{testCase:n,runResult:r,outputFiles:i,cwd:a}=e,s=await t({stdout:r.stdout,expectedStdout:typeof n.output==`string`?n.output:void 0,fileOutputPath:typeof n.fileOutputPath==`string`?n.fileOutputPath:void 0,cwd:a,outputFiles:i});return i.splice(0,i.length,...s.outputFiles),s.matches?{}:{decisionCode:o.WRONG_ANSWER}}async function O(e,t){let n=await _(e[0],e.slice(1),{cwd:t.cwd,stdin:t.stdin,env:t.env},t.timeLimitSeconds);return{stdin:t.stdin,stdout:n.stdout,stderr:n.stderr,status:n.status??void 0,timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes,outputLimitExceeded:n.outputLimitExceeded}}function k(e){if(e.runResult.status!==0)return{decisionCode:o.RUNTIME_ERROR,stderr:e.runResult.stderr};if(e.runResult.timeSeconds>e.context.timeLimitSeconds)return{decisionCode:o.TIME_LIMIT_EXCEEDED,stderr:e.runResult.stderr};if(e.runResult.memoryBytes>(e.context.problemMarkdownFrontMatter.memoryLimitByte??1/0))return{decisionCode:o.MEMORY_LIMIT_EXCEEDED,stderr:e.runResult.stderr};if(e.runResult.outputLimitExceeded||e.runResult.stdout.length>e.context.outputLimitLength||e.runResult.stderr.length>e.context.outputLimitLength)return{decisionCode:o.OUTPUT_SIZE_LIMIT_EXCEEDED,stderr:e.runResult.stderr};let t=e.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length??0;return e.outputFiles.length<t?{decisionCode:o.MISSING_REQUIRED_OUTPUT_FILE_ERROR}:{decisionCode:o.ACCEPTED}}export{x as commandJudgePreset};
+import{printTestCaseResult as e}from"../helpers/printTestCaseResult.js";import{judgeAgainstExpectations as t}from"../helpers/compareExpectedOutputFiles.js";import{parseArgs as n}from"../helpers/parseArgs.js";import{readTestCases as r}from"../helpers/readTestCases.js";import{findLanguageDefinitionByPath as i}from"../helpers/findLanguageDefinitionByPath.js";import{readProblemMarkdownFrontMatter as a}from"../helpers/readProblemMarkdownFrontMatter.js";import{DecisionCode as o}from"../types/decisionCode.js";import{cleanWorkingDirectory as s,snapshotWorkingDirectory as c}from"../helpers/cleanWorkingDirectory.js";import{copyTestCaseFileInput as l}from"../helpers/copyTestCaseFileInput.js";import{findEntryPointFile as u}from"../helpers/findEntryPointFile.js";import{judgeByStaticAnalysis as d}from"../helpers/judgeByStaticAnalysis.js";import{readOutputFiles as f}from"../helpers/readOutputFiles.js";import{passesIsolationCheckInDebugMode as p}from"../helpers/runIsolationCheckInDebugMode.js";import{printDebugCwdBanner as m,printDebugExpectationFailureBanner as h,resolveCwds as g}from"../helpers/resolveCwds.js";import{spawnWithTimeout as _}from"../helpers/spawnWithTimeout.js";import v from"node:path";import{z as y}from"zod";import{getErrorMessage as b}from"@willbooster/shared-lib";const x=y.object({language:y.union([y.string(),y.array(y.string())]).optional()});async function S(e,t={}){let r=n(process.argv),i=x.parse(r.params),{cwds:o,isDebugMode:s}=await g(e,r.cwd),c=await a(e),l=await(t.readTestCases??E)(e),u={buildTimeoutSeconds:t.limits?.buildTimeoutSeconds??10,maxOutputLength:t.limits?.maxOutputLength??5e4},d=typeof c.timeLimitMs==`number`?c.timeLimitMs/1e3:t.runTimeoutSeconds??2,f={problemMarkdownFrontMatter:c,testCases:l,limits:u,timeLimitSeconds:d};if(s){let t=(u.buildTimeoutSeconds+l.length*d)*1e3;if(!await p(e,o,i,{expectedMaxDurationMs:t})){process.exitCode=1;return}}for(let n of o){s&&m(e,n);let r=await C(n.cwd,i,f,t);s&&!w(n,r)&&(process.exitCode=1,h(e,n))}}async function C(t,n,r,a){let{problemMarkdownFrontMatter:p,testCases:m,limits:h,timeLimitSeconds:g}=r,_=m[0]?.id??`prebuild`,v=await d(t,p);if(v)return e({testCaseId:_,...v}),{allAccepted:!1};let y=await u(t,n.language);if(!y)return e({testCaseId:_,decisionCode:o.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,stderr:`main file not found${n.language?`: language: ${n.language}`:``}`}),{allAccepted:!1};let x=i(y);if(!x)return e({testCaseId:_,decisionCode:o.WRONG_ANSWER,stderr:`unsupported language`}),{allAccepted:!1};let S={...process.env,CI:``,FORCE_COLOR:`0`},C=y;if(a.buildSubmission!==!1){if(x.prebuild)try{await x.prebuild(t);let e=await u(t,n.language);e&&(C=e)}catch(t){return e({testCaseId:_,decisionCode:o.BUILD_ERROR,stderr:b(t)}),{allAccepted:!1}}let r=x.buildCommand?.(C);if(r){let n=await T(r,{cwd:t,env:S,testCaseId:_,limits:h});if(n)return e(n),{allAccepted:!1}}}let w=await c(t);if(m.length===0)return e({testCaseId:`default`,decisionCode:o.ACCEPTED}),{allAccepted:!0};let E=m.shared?.fileInputPath;for(let n of m){E&&await l(E,t),n.fileInputPath&&await l(n.fileInputPath,t);let r=x.command(C),i=n.input??``,c;try{a.resolveInput&&(i=await a.resolveInput({testCase:n,cwd:t,env:S})),a.runCommand&&(c=await a.runCommand({testCase:n,command:r,stdin:i,cwd:t,env:S,timeLimitSeconds:g}))}catch(r){return e({testCaseId:n.id,decisionCode:o.RUNTIME_ERROR,stdin:i,stderr:b(r)}),await s(t,w),{allAccepted:!1}}let u=c??await O(r,{stdin:i,cwd:t,env:S,timeLimitSeconds:g}),d=await f(t,p.requiredOutputFilePaths??[]),m={timeLimitSeconds:g,outputLimitLength:h.maxOutputLength,problemMarkdownFrontMatter:{memoryLimitByte:p.memoryLimitByte,requiredOutputFilePaths:p.requiredOutputFilePaths}},_=k({runResult:u,outputFiles:d,context:m}),v=_;if(_.decisionCode===o.ACCEPTED)try{let e=a.test?await a.test({testCase:n,runResult:u,outputFiles:d,cwd:t,context:m}):await D({testCase:n,runResult:u,outputFiles:d,cwd:t});e&&(v={decisionCode:e.decisionCode??_.decisionCode,feedbackMarkdown:e.feedbackMarkdown,stderr:e.stderr})}catch(e){v={decisionCode:o.RUNTIME_ERROR,stderr:b(e)}}if(e({testCaseId:n.id,...v,exitStatus:u.status,stdin:u.stdin,stdout:u.stdout.slice(0,h.maxOutputLength)||void 0,stderr:(v.stderr??u.stderr).slice(0,h.maxOutputLength)||void 0,timeSeconds:u.timeSeconds,cpuTimeSeconds:u.cpuTimeSeconds,memoryBytes:u.memoryBytes,outputFiles:d.length>0?d:void 0}),await s(t,w),v.decisionCode!==o.ACCEPTED)return{allAccepted:!1}}return{allAccepted:!0}}function w(e,t){return t.allAccepted===(e.expectedResult===`accepted`)}async function T(e,t){let n=await _(e[0],e.slice(1),{cwd:t.cwd,env:t.env},t.limits.buildTimeoutSeconds),r=n.status??void 0;if(n.status!==0)return{testCaseId:t.testCaseId,decisionCode:o.BUILD_ERROR,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes};if(n.timeSeconds>t.limits.buildTimeoutSeconds)return{testCaseId:t.testCaseId,decisionCode:o.BUILD_TIME_LIMIT_EXCEEDED,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes};if(n.outputLimitExceeded||n.stdout.length>t.limits.maxOutputLength||n.stderr.length>t.limits.maxOutputLength)return{testCaseId:t.testCaseId,decisionCode:o.BUILD_OUTPUT_SIZE_LIMIT_EXCEEDED,exitStatus:r,stdout:n.stdout.slice(0,t.limits.maxOutputLength),stderr:n.stderr.slice(0,t.limits.maxOutputLength),timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes}}async function E(e){return await r(v.join(e,`test_cases`))}async function D(e){let{testCase:n,runResult:r,outputFiles:i,cwd:a}=e,s=await t({stdout:r.stdout,expectedStdout:typeof n.output==`string`?n.output:void 0,fileOutputPath:typeof n.fileOutputPath==`string`?n.fileOutputPath:void 0,cwd:a,outputFiles:i});return i.splice(0,i.length,...s.outputFiles),s.matches?{}:{decisionCode:o.WRONG_ANSWER}}async function O(e,t){let n=await _(e[0],e.slice(1),{cwd:t.cwd,stdin:t.stdin,env:t.env},t.timeLimitSeconds);return{stdin:t.stdin,stdout:n.stdout,stderr:n.stderr,status:n.status??void 0,timeSeconds:n.timeSeconds,cpuTimeSeconds:n.cpuTimeSeconds,memoryBytes:n.memoryBytes,outputLimitExceeded:n.outputLimitExceeded}}function k(e){if(e.runResult.status!==0)return{decisionCode:o.RUNTIME_ERROR,stderr:e.runResult.stderr};if(e.runResult.timeSeconds>e.context.timeLimitSeconds)return{decisionCode:o.TIME_LIMIT_EXCEEDED,stderr:e.runResult.stderr};if(e.runResult.memoryBytes>(e.context.problemMarkdownFrontMatter.memoryLimitByte??1/0))return{decisionCode:o.MEMORY_LIMIT_EXCEEDED,stderr:e.runResult.stderr};if(e.runResult.outputLimitExceeded||e.runResult.stdout.length>e.context.outputLimitLength||e.runResult.stderr.length>e.context.outputLimitLength)return{decisionCode:o.OUTPUT_SIZE_LIMIT_EXCEEDED,stderr:e.runResult.stderr};let t=e.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length??0;return e.outputFiles.length<t?{decisionCode:o.MISSING_REQUIRED_OUTPUT_FILE_ERROR}:{decisionCode:o.ACCEPTED}}export{S as commandJudgePreset};
 //# sourceMappingURL=command.js.map

package/dist/presets/command.js.map

@@ -1 +1 @@
-{"version":3,"file":"command.js","names":["readFileTestCases"],"sources":["../../src/presets/command.ts"],"sourcesContent":["import path from 'node:path';\n\nimport { z } from 'zod';\n\nimport { cleanWorkingDirectory, snapshotWorkingDirectory } from '../helpers/cleanWorkingDirectory.js';\nimport { judgeAgainstExpectations } from '../helpers/compareExpectedOutputFiles.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 { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { readTestCases as readFileTestCases } from '../helpers/readTestCases.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { spawnWithTimeout } from '../helpers/spawnWithTimeout.js';\nimport { DecisionCode } from '../types/decisionCode.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 = 2;\nconst MAX_STDOUT_LENGTH = 50_000;\n\nconst judgeParamsSchema = z.object({\n  language: z.union([z.string(), z.array(z.string())]).optional(),\n});\n\ntype JudgeParams = z.infer<typeof judgeParamsSchema>;\n\n/** What every command test case needs; a custom `readTestCases` may add any fields of its own. */\ninterface BaseCommandTestCase {\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 CommandTestCase extends BaseCommandTestCase {\n  /** Expected standard output (`test_cases/<id>.out`). */\n  output?: string;\n  /** Directory of expected output files compared after the run (`test_cases/<id>.fout/`). */\n  fileOutputPath?: string;\n}\n\ntype CommandJudgeCaseResult = Pick<TestCaseResult, 'decisionCode' | 'feedbackMarkdown' | 'stderr'>;\n\nexport interface CommandRunResult {\n  stdin: string;\n  stdout: string;\n  stderr: string;\n  status: number | undefined;\n  timeSeconds: number;\n  /** See `TestCaseResult.cpuTimeSeconds`; a custom `runCommand` may leave it out. */\n  cpuTimeSeconds?: number;\n  memoryBytes: number;\n  outputLimitExceeded?: boolean;\n}\n\ninterface CommandJudgeContext {\n  timeLimitSeconds: number;\n  outputLimitLength: number;\n  problemMarkdownFrontMatter: Pick<ProblemMarkdownFrontMatter, 'memoryLimitByte' | 'requiredOutputFilePaths'>;\n}\n\nexport interface CommandJudgeLimits {\n  buildTimeoutSeconds: number;\n  maxOutputLength: number;\n}\n\ninterface ResolvedCommandProblem<TTestCase extends BaseCommandTestCase> {\n  problemMarkdownFrontMatter: ProblemMarkdownFrontMatter;\n  testCases: readonly TTestCase[];\n  limits: CommandJudgeLimits;\n  timeLimitSeconds: number;\n}\n\nexport interface CommandJudgePresetOptions<\n  TTestCase extends BaseCommandTestCase = CommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n> {\n  /**\n   * Set to false when the custom runner builds the submitted source itself.\n   * Defaults to true.\n   */\n  buildSubmission?: boolean;\n  limits?: CommandJudgeLimits;\n  /** The per-case time limit used only when `problem.md` declares no `timeLimitMs`. */\n  runTimeoutSeconds?: number;\n  readTestCases?: (problemDir: string) => Promise<readonly TTestCase[]>;\n  resolveInput?: (context: { testCase: TTestCase; cwd: string; env: NodeJS.ProcessEnv }) => Promise<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  }) => Promise<TRunResult> | TRunResult;\n  /**\n   * Decides the verdict of a run that passed the limit checks. It replaces the default comparison\n   * of `testCase.output` (space-separated tokens) and `testCase.fileOutputPath` (see\n   * `compareStdoutAsSpaceSeparatedTokens` and `compareExpectedOutputFiles`, both exported), so\n   * call those yourself when the test case ships `.out` / `.fout` expectations to keep.\n   * `outputFiles` is the array printed with the result and may be edited in place.\n   */\n  test?: (context: {\n    testCase: TTestCase;\n    runResult: TRunResult;\n    outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n    /** The submission's working directory, e.g. for `compareExpectedOutputFiles(cwd, testCase.fileOutputPath)`. */\n    cwd: string;\n    context: CommandJudgeContext;\n  }) => Promise<Partial<CommandJudgeCaseResult>> | Partial<CommandJudgeCaseResult> | undefined;\n}\n\n/**\n * A preset function for judging by executable command.\n *\n * Without options, test cases come from `test_cases/` (`.in`, `.out`, `.fin/`, `.fout/`) and a run\n * within the limits is accepted when its stdout matches `.out` and its files match `.fout/`; a\n * case without either expectation only has to run within the limits.\n * Keep problem-specific logic in `resolveInput` and `test`.\n *\n * @example\n * Create `judge.ts` that judges `test_cases/` like the default stdio harness, with the command\n * preset's debug mode:\n * ```ts\n * import { commandJudgePreset } from '@exercode/problem-utils/presets/command';\n *\n * await commandJudgePreset(import.meta.dirname);\n * ```\n *\n * @example\n * Create `judge.ts` with an own verdict:\n * ```ts\n * import { commandJudgePreset } from '@exercode/problem-utils/presets/command';\n * import { DecisionCode } from '@exercode/problem-utils';\n *\n * await commandJudgePreset(import.meta.dirname, {\n *   readTestCases: async () => [\n *     { id: '01', input: '1 2' },\n *   ],\n *   test: ({ runResult }) => {\n *     return runResult.stdout.trim() === '3'\n *       ? { decisionCode: DecisionCode.ACCEPTED }\n *       : { decisionCode: DecisionCode.WRONG_ANSWER };\n *   },\n * });\n * ```\n *\n * Run with the required parameters:\n * ```bash\n * bun judge.ts model_answers/python '{ \"language\": \"python\" }'\n * ```\n *\n * Run without a cwd argument to judge each `<problemDir>/model_answers/*` directory\n * for debugging. A prominent banner is printed for each run.\n * ```bash\n * bun judge.ts\n * ```\n */\nexport async function commandJudgePreset<\n  TTestCase extends BaseCommandTestCase = CommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n>(problemDir: string, options: CommandJudgePresetOptions<TTestCase, TRunResult> = {}): Promise<void> {\n  const args = parseArgs(process.argv);\n  const params = judgeParamsSchema.parse(args.params);\n\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n  const problemMarkdownFrontMatter = await readProblemMarkdownFrontMatter(problemDir);\n  const testCases = await (options.readTestCases ?? readCommandTestCases)(problemDir);\n  const limits = {\n    buildTimeoutSeconds: options.limits?.buildTimeoutSeconds ?? BUILD_TIMEOUT_SECONDS,\n    maxOutputLength: options.limits?.maxOutputLength ?? MAX_STDOUT_LENGTH,\n  };\n  const timeLimitSeconds =\n    typeof problemMarkdownFrontMatter.timeLimitMs === 'number'\n      ? problemMarkdownFrontMatter.timeLimitMs / 1000\n      : (options.runTimeoutSeconds ?? JUDGE_DEFAULT_TIMEOUT_SECONDS);\n  const problem: ResolvedCommandProblem<TTestCase> = {\n    problemMarkdownFrontMatter,\n    testCases,\n    limits,\n    timeLimitSeconds,\n  };\n\n  if (isDebugMode) {\n    const expectedMaxDurationMs = (limits.buildTimeoutSeconds + testCases.length * timeLimitSeconds) * 1000;\n    if (!(await passesIsolationCheckInDebugMode(problemDir, cwds, params, { expectedMaxDurationMs }))) {\n      process.exitCode = 1;\n      return;\n    }\n  }\n\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await runCommandJudgeForCwd<TTestCase, TRunResult>(resolvedCwd.cwd, params, problem, options);\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function runCommandJudgeForCwd<\n  TTestCase extends BaseCommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n>(\n  cwd: string,\n  params: JudgeParams,\n  problem: ResolvedCommandProblem<TTestCase>,\n  options: CommandJudgePresetOptions<TTestCase, TRunResult>\n): Promise<{ allAccepted: boolean }> {\n  const { problemMarkdownFrontMatter, testCases, limits, timeLimitSeconds } = problem;\n  const prebuildTestCaseId = testCases[0]?.id ?? 'prebuild';\n\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, problemMarkdownFrontMatter);\n  if (staticAnalysisResult) {\n    printTestCaseResult({ testCaseId: prebuildTestCaseId, ...staticAnalysisResult });\n    return { allAccepted: false };\n  }\n\n  const originalMainFilePath = await findEntryPointFile(cwd, params.language);\n  if (!originalMainFilePath) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      stderr: `main file not found${params.language ? `: language: ${params.language}` : ''}`,\n    });\n    return { allAccepted: false };\n  }\n\n  const languageDefinition = findLanguageDefinitionByPath(originalMainFilePath);\n  if (!languageDefinition) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      stderr: 'unsupported language',\n    });\n    return { allAccepted: false };\n  }\n\n  // `CI` changes affects Chainlit. `FORCE_COLOR` affects Bun.\n  const env = { ...process.env, CI: '', FORCE_COLOR: '0' };\n\n  let mainFilePath = originalMainFilePath;\n  if (options.buildSubmission !== false) {\n    if (languageDefinition.prebuild) {\n      try {\n        await languageDefinition.prebuild(cwd);\n        const prebuiltMainFilePath = await findEntryPointFile(cwd, params.language);\n        if (prebuiltMainFilePath) mainFilePath = prebuiltMainFilePath;\n      } catch (error) {\n        printTestCaseResult({\n          testCaseId: prebuildTestCaseId,\n          decisionCode: DecisionCode.BUILD_ERROR,\n          stderr: error instanceof Error ? error.message : String(error),\n        });\n        return { allAccepted: false };\n      }\n    }\n\n    const buildCommand = languageDefinition.buildCommand?.(mainFilePath);\n    if (buildCommand) {\n      const buildResult = await runBuild(buildCommand, {\n        cwd,\n        env,\n        testCaseId: prebuildTestCaseId,\n        limits,\n      });\n      if (buildResult) {\n        printTestCaseResult(buildResult);\n        return { allAccepted: false };\n      }\n    }\n  }\n\n  const cwdSnapshot = await snapshotWorkingDirectory(cwd);\n\n  if (testCases.length === 0) {\n    printTestCaseResult({ testCaseId: 'default', decisionCode: DecisionCode.ACCEPTED });\n    return { allAccepted: true };\n  }\n\n  const sharedFileInputPath = (testCases as { shared?: { fileInputPath?: string } }).shared?.fileInputPath;\n\n  for (const testCase of testCases) {\n    if (sharedFileInputPath) await copyTestCaseFileInput(sharedFileInputPath, cwd);\n    if (testCase.fileInputPath) await copyTestCaseFileInput(testCase.fileInputPath, cwd);\n\n    const command = languageDefinition.command(mainFilePath);\n    let stdin = testCase.input ?? '';\n    let customRunResult: TRunResult | undefined;\n    try {\n      if (options.resolveInput) {\n        stdin = await options.resolveInput({ testCase, cwd, env });\n      }\n\n      if (options.runCommand) {\n        customRunResult = await options.runCommand({ testCase, command, stdin, cwd, env, timeLimitSeconds });\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(cwd, cwdSnapshot);\n      return { allAccepted: false };\n    }\n\n    // The default runner reports the submission's failures in its result; what it throws is the\n    // judge's own failure and ends the run without a verdict.\n    const runResult =\n      customRunResult ?? ((await runCommand(command, { stdin, cwd, env, timeLimitSeconds })) as TRunResult);\n\n    const outputFiles = await readOutputFiles(cwd, problemMarkdownFrontMatter.requiredOutputFilePaths ?? []);\n    const judgeContext: CommandJudgeContext = {\n      timeLimitSeconds,\n      outputLimitLength: limits.maxOutputLength,\n      problemMarkdownFrontMatter: {\n        memoryLimitByte: problemMarkdownFrontMatter.memoryLimitByte,\n        requiredOutputFilePaths: problemMarkdownFrontMatter.requiredOutputFilePaths,\n      },\n    };\n    const baseJudgeResult = evaluateByLimits({ runResult, outputFiles, context: judgeContext });\n    let judgeResult = baseJudgeResult;\n    if (baseJudgeResult.decisionCode === DecisionCode.ACCEPTED) {\n      try {\n        const extendedJudgeResult = options.test\n          ? await options.test({ testCase, runResult, outputFiles, cwd, context: judgeContext })\n          : await compareWithExpectedOutputs({ testCase, runResult, outputFiles, cwd });\n        if (extendedJudgeResult) {\n          judgeResult = {\n            decisionCode: extendedJudgeResult.decisionCode ?? baseJudgeResult.decisionCode,\n            feedbackMarkdown: extendedJudgeResult.feedbackMarkdown,\n            stderr: extendedJudgeResult.stderr,\n          };\n        }\n      } catch (error) {\n        judgeResult = {\n          decisionCode: DecisionCode.RUNTIME_ERROR,\n          stderr: errorToMessage(error),\n        };\n      }\n    }\n\n    printTestCaseResult({\n      testCaseId: testCase.id,\n      ...judgeResult,\n      exitStatus: runResult.status,\n      stdin: runResult.stdin,\n      stdout: runResult.stdout.slice(0, limits.maxOutputLength) || undefined,\n      stderr: (judgeResult.stderr ?? runResult.stderr).slice(0, limits.maxOutputLength) || undefined,\n      timeSeconds: runResult.timeSeconds,\n      cpuTimeSeconds: runResult.cpuTimeSeconds,\n      memoryBytes: runResult.memoryBytes,\n      outputFiles: outputFiles.length > 0 ? outputFiles : undefined,\n    });\n\n    await cleanWorkingDirectory(cwd, cwdSnapshot);\n    if (judgeResult.decisionCode !== DecisionCode.ACCEPTED) return { allAccepted: false };\n  }\n\n  return { allAccepted: true };\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: { allAccepted: boolean }): boolean {\n  return result.allAccepted === (resolvedCwd.expectedResult === 'accepted');\n}\n\nasync function runBuild(\n  buildCommand: readonly [string, ...string[]],\n  context: { cwd: string; env: NodeJS.ProcessEnv; testCaseId: string; limits: CommandJudgeLimits }\n): Promise<(Omit<TestCaseResult, 'testCaseId'> & { testCaseId: string }) | undefined> {\n  const spawnResult = await spawnWithTimeout(\n    buildCommand[0],\n    buildCommand.slice(1),\n    { cwd: context.cwd, env: context.env },\n    context.limits.buildTimeoutSeconds\n  );\n  const exitStatus = spawnResult.status ?? undefined;\n\n  if (spawnResult.status !== 0) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_ERROR,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  if (spawnResult.timeSeconds > context.limits.buildTimeoutSeconds) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  if (\n    spawnResult.outputLimitExceeded ||\n    spawnResult.stdout.length > context.limits.maxOutputLength ||\n    spawnResult.stderr.length > context.limits.maxOutputLength\n  ) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_OUTPUT_SIZE_LIMIT_EXCEEDED,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  return;\n}\n\nfunction errorToMessage(error: unknown): string {\n  return error instanceof Error ? error.message : String(error);\n}\n\nasync function readCommandTestCases<TTestCase extends BaseCommandTestCase = CommandTestCase>(\n  problemDir: string\n): 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\n/** The default verdict: a case exposing a string `output` / `fileOutputPath` is judged against it. */\nasync function compareWithExpectedOutputs(context: {\n  testCase: BaseCommandTestCase & Partial<CommandTestCase>;\n  runResult: CommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  cwd: string;\n}): Promise<Partial<CommandJudgeCaseResult>> {\n  const { testCase, runResult, outputFiles, cwd } = context;\n  const judgement = await judgeAgainstExpectations({\n    stdout: runResult.stdout,\n    expectedStdout: typeof testCase.output === 'string' ? testCase.output : undefined,\n    fileOutputPath: typeof testCase.fileOutputPath === 'string' ? testCase.fileOutputPath : undefined,\n    cwd,\n    outputFiles,\n  });\n  outputFiles.splice(0, outputFiles.length, ...judgement.outputFiles);\n  return judgement.matches ? {} : { decisionCode: DecisionCode.WRONG_ANSWER };\n}\n\nasync function runCommand(\n  command: readonly [string, ...string[]],\n  context: {\n    stdin: string;\n    cwd: string;\n    env: NodeJS.ProcessEnv;\n    timeLimitSeconds: number;\n  }\n): Promise<CommandRunResult> {\n  const spawnResult = await spawnWithTimeout(\n    command[0],\n    command.slice(1),\n    { cwd: context.cwd, stdin: context.stdin, env: context.env },\n    context.timeLimitSeconds\n  );\n\n  return {\n    stdin: context.stdin,\n    stdout: spawnResult.stdout,\n    stderr: spawnResult.stderr,\n    status: spawnResult.status ?? undefined,\n    timeSeconds: spawnResult.timeSeconds,\n    cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n    memoryBytes: spawnResult.memoryBytes,\n    outputLimitExceeded: spawnResult.outputLimitExceeded,\n  };\n}\n\nfunction evaluateByLimits(context: {\n  runResult: CommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  context: CommandJudgeContext;\n}): CommandJudgeCaseResult {\n  if (context.runResult.status !== 0) {\n    return { decisionCode: DecisionCode.RUNTIME_ERROR, stderr: context.runResult.stderr };\n  }\n\n  if (context.runResult.timeSeconds > context.context.timeLimitSeconds) {\n    return { decisionCode: DecisionCode.TIME_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  if (\n    context.runResult.memoryBytes >\n    (context.context.problemMarkdownFrontMatter.memoryLimitByte ?? Number.POSITIVE_INFINITY)\n  ) {\n    return { decisionCode: DecisionCode.MEMORY_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  if (\n    context.runResult.outputLimitExceeded ||\n    context.runResult.stdout.length > context.context.outputLimitLength ||\n    context.runResult.stderr.length > context.context.outputLimitLength\n  ) {\n    return { decisionCode: DecisionCode.OUTPUT_SIZE_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  const requiredOutputFileCount = context.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length ?? 0;\n  if (context.outputFiles.length < requiredOutputFileCount) {\n    return { decisionCode: DecisionCode.MISSING_REQUIRED_OUTPUT_FILE_ERROR };\n  }\n\n  return { decisionCode: DecisionCode.ACCEPTED };\n}\n"],"mappings":"6sCA2BA,MAIM,EAAoB,EAAE,OAAO,CACjC,SAAU,EAAE,MAAM,CAAC,EAAE,OAAO,EAAG,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAChE,CAAC,EA0ID,eAAsB,EAGpB,EAAoB,EAA4D,CAAC,EAAkB,CACnG,IAAM,EAAO,EAAU,QAAQ,IAAI,EAC7B,EAAS,EAAkB,MAAM,EAAK,MAAM,EAE5C,CAAE,OAAM,eAAgB,MAAM,EAAY,EAAY,EAAK,GAAG,EAC9D,EAA6B,MAAM,EAA+B,CAAU,EAC5E,EAAY,MAAO,EAAQ,eAAiB,EAAA,CAAsB,CAAU,EAC5E,EAAS,CACb,oBAAqB,EAAQ,QAAQ,qBAAuB,GAC5D,gBAAiB,EAAQ,QAAQ,iBAAmB,GACtD,EACM,EACJ,OAAO,EAA2B,aAAgB,SAC9C,EAA2B,YAAc,IACxC,EAAQ,mBAAqB,EAC9B,EAA6C,CACjD,6BACA,YACA,SACA,kBACF,EAEA,GAAI,EAAa,CACf,IAAM,GAAyB,EAAO,oBAAsB,EAAU,OAAS,GAAoB,IACnG,GAAI,CAAE,MAAM,EAAgC,EAAY,EAAM,EAAQ,CAAE,uBAAsB,CAAC,EAAI,CACjG,QAAQ,SAAW,EACnB,MACF,CACF,CAEA,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAA6C,EAAY,IAAK,EAAQ,EAAS,CAAO,EACvG,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EAIb,EACA,EACA,EACA,EACmC,CACnC,GAAM,CAAE,6BAA4B,YAAW,SAAQ,oBAAqB,EACtE,EAAqB,EAAU,EAAE,EAAE,IAAM,WAEzC,EAAuB,MAAM,EAAsB,EAAK,CAA0B,EACxF,GAAI,EAEF,OADA,EAAoB,CAAE,WAAY,EAAoB,GAAG,CAAqB,CAAC,EACxE,CAAE,YAAa,EAAM,EAG9B,IAAM,EAAuB,MAAM,EAAmB,EAAK,EAAO,QAAQ,EAC1E,GAAI,CAAC,EAMH,OALA,EAAoB,CAClB,WAAY,EACZ,aAAc,EAAa,uCAC3B,OAAQ,sBAAsB,EAAO,SAAW,eAAe,EAAO,WAAa,IACrF,CAAC,EACM,CAAE,YAAa,EAAM,EAG9B,IAAM,EAAqB,EAA6B,CAAoB,EAC5E,GAAI,CAAC,EAMH,OALA,EAAoB,CAClB,WAAY,EACZ,aAAc,EAAa,aAC3B,OAAQ,sBACV,CAAC,EACM,CAAE,YAAa,EAAM,EAI9B,IAAM,EAAM,CAAE,GAAG,QAAQ,IAAK,GAAI,GAAI,YAAa,GAAI,EAEnD,EAAe,EACnB,GAAI,EAAQ,kBAAoB,GAAO,CACrC,GAAI,EAAmB,SACrB,GAAI,CACF,MAAM,EAAmB,SAAS,CAAG,EACrC,IAAM,EAAuB,MAAM,EAAmB,EAAK,EAAO,QAAQ,EACtE,IAAsB,EAAe,EAC3C,OAAS,EAAO,CAMd,OALA,EAAoB,CAClB,WAAY,EACZ,aAAc,EAAa,YAC3B,OAAQ,aAAiB,MAAQ,EAAM,QAAU,OAAO,CAAK,CAC/D,CAAC,EACM,CAAE,YAAa,EAAM,CAC9B,CAGF,IAAM,EAAe,EAAmB,eAAe,CAAY,EACnE,GAAI,EAAc,CAChB,IAAM,EAAc,MAAM,EAAS,EAAc,CAC/C,MACA,MACA,WAAY,EACZ,QACF,CAAC,EACD,GAAI,EAEF,OADA,EAAoB,CAAW,EACxB,CAAE,YAAa,EAAM,CAEhC,CACF,CAEA,IAAM,EAAc,MAAM,EAAyB,CAAG,EAEtD,GAAI,EAAU,SAAW,EAEvB,OADA,EAAoB,CAAE,WAAY,UAAW,aAAc,EAAa,QAAS,CAAC,EAC3E,CAAE,YAAa,EAAK,EAG7B,IAAM,EAAuB,EAAsD,QAAQ,cAE3F,IAAK,IAAM,KAAY,EAAW,CAC5B,GAAqB,MAAM,EAAsB,EAAqB,CAAG,EACzE,EAAS,eAAe,MAAM,EAAsB,EAAS,cAAe,CAAG,EAEnF,IAAM,EAAU,EAAmB,QAAQ,CAAY,EACnD,EAAQ,EAAS,OAAS,GAC1B,EACJ,GAAI,CACE,EAAQ,eACV,EAAQ,MAAM,EAAQ,aAAa,CAAE,WAAU,MAAK,KAAI,CAAC,GAGvD,EAAQ,aACV,EAAkB,MAAM,EAAQ,WAAW,CAAE,WAAU,UAAS,QAAO,MAAK,MAAK,kBAAiB,CAAC,EAEvG,OAAS,EAAO,CAQd,OAPA,EAAoB,CAClB,WAAY,EAAS,GACrB,aAAc,EAAa,cAC3B,QACA,OAAQ,EAAe,CAAK,CAC9B,CAAC,EACD,MAAM,EAAsB,EAAK,CAAW,EACrC,CAAE,YAAa,EAAM,CAC9B,CAIA,IAAM,EACJ,GAAqB,MAAM,EAAW,EAAS,CAAE,QAAO,MAAK,MAAK,kBAAiB,CAAC,EAEhF,EAAc,MAAM,EAAgB,EAAK,EAA2B,yBAA2B,CAAC,CAAC,EACjG,EAAoC,CACxC,mBACA,kBAAmB,EAAO,gBAC1B,2BAA4B,CAC1B,gBAAiB,EAA2B,gBAC5C,wBAAyB,EAA2B,uBACtD,CACF,EACM,EAAkB,EAAiB,CAAE,YAAW,cAAa,QAAS,CAAa,CAAC,EACtF,EAAc,EAClB,GAAI,EAAgB,eAAiB,EAAa,SAChD,GAAI,CACF,IAAM,EAAsB,EAAQ,KAChC,MAAM,EAAQ,KAAK,CAAE,WAAU,YAAW,cAAa,MAAK,QAAS,CAAa,CAAC,EACnF,MAAM,EAA2B,CAAE,WAAU,YAAW,cAAa,KAAI,CAAC,EAC1E,IACF,EAAc,CACZ,aAAc,EAAoB,cAAgB,EAAgB,aAClE,iBAAkB,EAAoB,iBACtC,OAAQ,EAAoB,MAC9B,EAEJ,OAAS,EAAO,CACd,EAAc,CACZ,aAAc,EAAa,cAC3B,OAAQ,EAAe,CAAK,CAC9B,CACF,CAiBF,GAdA,EAAoB,CAClB,WAAY,EAAS,GACrB,GAAG,EACH,WAAY,EAAU,OACtB,MAAO,EAAU,MACjB,OAAQ,EAAU,OAAO,MAAM,EAAG,EAAO,eAAe,GAAK,IAAA,GAC7D,QAAS,EAAY,QAAU,EAAU,OAAA,CAAQ,MAAM,EAAG,EAAO,eAAe,GAAK,IAAA,GACrF,YAAa,EAAU,YACvB,eAAgB,EAAU,eAC1B,YAAa,EAAU,YACvB,YAAa,EAAY,OAAS,EAAI,EAAc,IAAA,EACtD,CAAC,EAED,MAAM,EAAsB,EAAK,CAAW,EACxC,EAAY,eAAiB,EAAa,SAAU,MAAO,CAAE,YAAa,EAAM,CACtF,CAEA,MAAO,CAAE,YAAa,EAAK,CAC7B,CAEA,SAAS,EAAsB,EAA0B,EAA2C,CAClG,OAAO,EAAO,eAAiB,EAAY,iBAAmB,WAChE,CAEA,eAAe,EACb,EACA,EACoF,CACpF,IAAM,EAAc,MAAM,EACxB,EAAa,GACb,EAAa,MAAM,CAAC,EACpB,CAAE,IAAK,EAAQ,IAAK,IAAK,EAAQ,GAAI,EACrC,EAAQ,OAAO,mBACjB,EACM,EAAa,EAAY,QAAU,IAAA,GAEzC,GAAI,EAAY,SAAW,EACzB,MAAO,CACL,WAAY,EAAQ,WACpB,aAAc,EAAa,YAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,EAGF,GAAI,EAAY,YAAc,EAAQ,OAAO,oBAC3C,MAAO,CACL,WAAY,EAAQ,WACpB,aAAc,EAAa,0BAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,EAGF,GACE,EAAY,qBACZ,EAAY,OAAO,OAAS,EAAQ,OAAO,iBAC3C,EAAY,OAAO,OAAS,EAAQ,OAAO,gBAE3C,MAAO,CACL,WAAY,EAAQ,WACpB,aAAc,EAAa,iCAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,CAIJ,CAEA,SAAS,EAAe,EAAwB,CAC9C,OAAO,aAAiB,MAAQ,EAAM,QAAU,OAAO,CAAK,CAC9D,CAEA,eAAe,EACb,EAC+B,CAE/B,OAAQ,MAAMA,EAAkB,EAAK,KAAK,EAAY,YAAY,CAAC,CACrE,CAGA,eAAe,EAA2B,EAKG,CAC3C,GAAM,CAAE,WAAU,YAAW,cAAa,OAAQ,EAC5C,EAAY,MAAM,EAAyB,CAC/C,OAAQ,EAAU,OAClB,eAAgB,OAAO,EAAS,QAAW,SAAW,EAAS,OAAS,IAAA,GACxE,eAAgB,OAAO,EAAS,gBAAmB,SAAW,EAAS,eAAiB,IAAA,GACxF,MACA,aACF,CAAC,EAED,OADA,EAAY,OAAO,EAAG,EAAY,OAAQ,GAAG,EAAU,WAAW,EAC3D,EAAU,QAAU,CAAC,EAAI,CAAE,aAAc,EAAa,YAAa,CAC5E,CAEA,eAAe,EACb,EACA,EAM2B,CAC3B,IAAM,EAAc,MAAM,EACxB,EAAQ,GACR,EAAQ,MAAM,CAAC,EACf,CAAE,IAAK,EAAQ,IAAK,MAAO,EAAQ,MAAO,IAAK,EAAQ,GAAI,EAC3D,EAAQ,gBACV,EAEA,MAAO,CACL,MAAO,EAAQ,MACf,OAAQ,EAAY,OACpB,OAAQ,EAAY,OACpB,OAAQ,EAAY,QAAU,IAAA,GAC9B,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,YACzB,oBAAqB,EAAY,mBACnC,CACF,CAEA,SAAS,EAAiB,EAIC,CACzB,GAAI,EAAQ,UAAU,SAAW,EAC/B,MAAO,CAAE,aAAc,EAAa,cAAe,OAAQ,EAAQ,UAAU,MAAO,EAGtF,GAAI,EAAQ,UAAU,YAAc,EAAQ,QAAQ,iBAClD,MAAO,CAAE,aAAc,EAAa,oBAAqB,OAAQ,EAAQ,UAAU,MAAO,EAG5F,GACE,EAAQ,UAAU,aACjB,EAAQ,QAAQ,2BAA2B,iBAAmB,KAE/D,MAAO,CAAE,aAAc,EAAa,sBAAuB,OAAQ,EAAQ,UAAU,MAAO,EAG9F,GACE,EAAQ,UAAU,qBAClB,EAAQ,UAAU,OAAO,OAAS,EAAQ,QAAQ,mBAClD,EAAQ,UAAU,OAAO,OAAS,EAAQ,QAAQ,kBAElD,MAAO,CAAE,aAAc,EAAa,2BAA4B,OAAQ,EAAQ,UAAU,MAAO,EAGnG,IAAM,EAA0B,EAAQ,QAAQ,2BAA2B,yBAAyB,QAAU,EAK9G,OAJI,EAAQ,YAAY,OAAS,EACxB,CAAE,aAAc,EAAa,kCAAmC,EAGlE,CAAE,aAAc,EAAa,QAAS,CAC/C"}
+{"version":3,"file":"command.js","names":["readFileTestCases"],"sources":["../../src/presets/command.ts"],"sourcesContent":["import path from 'node:path';\nimport { getErrorMessage } from '@willbooster/shared-lib';\n\nimport { z } from 'zod';\n\nimport { cleanWorkingDirectory, snapshotWorkingDirectory } from '../helpers/cleanWorkingDirectory.js';\nimport { judgeAgainstExpectations } from '../helpers/compareExpectedOutputFiles.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 { passesIsolationCheckInDebugMode } from '../helpers/runIsolationCheckInDebugMode.js';\nimport { readProblemMarkdownFrontMatter } from '../helpers/readProblemMarkdownFrontMatter.js';\nimport { readTestCases as readFileTestCases } from '../helpers/readTestCases.js';\nimport {\n  printDebugCwdBanner,\n  printDebugExpectationFailureBanner,\n  resolveCwds,\n  type ResolvedCwd,\n} from '../helpers/resolveCwds.js';\nimport { spawnWithTimeout } from '../helpers/spawnWithTimeout.js';\nimport { DecisionCode } from '../types/decisionCode.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 = 2;\nconst MAX_STDOUT_LENGTH = 50_000;\n\nconst judgeParamsSchema = z.object({\n  language: z.union([z.string(), z.array(z.string())]).optional(),\n});\n\ntype JudgeParams = z.infer<typeof judgeParamsSchema>;\n\n/** What every command test case needs; a custom `readTestCases` may add any fields of its own. */\ninterface BaseCommandTestCase {\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 CommandTestCase extends BaseCommandTestCase {\n  /** Expected standard output (`test_cases/<id>.out`). */\n  output?: string;\n  /** Directory of expected output files compared after the run (`test_cases/<id>.fout/`). */\n  fileOutputPath?: string;\n}\n\ntype CommandJudgeCaseResult = Pick<TestCaseResult, 'decisionCode' | 'feedbackMarkdown' | 'stderr'>;\n\nexport interface CommandRunResult {\n  stdin: string;\n  stdout: string;\n  stderr: string;\n  status: number | undefined;\n  timeSeconds: number;\n  /** See `TestCaseResult.cpuTimeSeconds`; a custom `runCommand` may leave it out. */\n  cpuTimeSeconds?: number;\n  memoryBytes: number;\n  outputLimitExceeded?: boolean;\n}\n\ninterface CommandJudgeContext {\n  timeLimitSeconds: number;\n  outputLimitLength: number;\n  problemMarkdownFrontMatter: Pick<ProblemMarkdownFrontMatter, 'memoryLimitByte' | 'requiredOutputFilePaths'>;\n}\n\nexport interface CommandJudgeLimits {\n  buildTimeoutSeconds: number;\n  maxOutputLength: number;\n}\n\ninterface ResolvedCommandProblem<TTestCase extends BaseCommandTestCase> {\n  problemMarkdownFrontMatter: ProblemMarkdownFrontMatter;\n  testCases: readonly TTestCase[];\n  limits: CommandJudgeLimits;\n  timeLimitSeconds: number;\n}\n\nexport interface CommandJudgePresetOptions<\n  TTestCase extends BaseCommandTestCase = CommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n> {\n  /**\n   * Set to false when the custom runner builds the submitted source itself.\n   * Defaults to true.\n   */\n  buildSubmission?: boolean;\n  limits?: CommandJudgeLimits;\n  /** The per-case time limit used only when `problem.md` declares no `timeLimitMs`. */\n  runTimeoutSeconds?: number;\n  readTestCases?: (problemDir: string) => Promise<readonly TTestCase[]>;\n  resolveInput?: (context: { testCase: TTestCase; cwd: string; env: NodeJS.ProcessEnv }) => Promise<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  }) => Promise<TRunResult> | TRunResult;\n  /**\n   * Decides the verdict of a run that passed the limit checks. It replaces the default comparison\n   * of `testCase.output` (space-separated tokens) and `testCase.fileOutputPath` (see\n   * `compareStdoutAsSpaceSeparatedTokens` and `compareExpectedOutputFiles`, both exported), so\n   * call those yourself when the test case ships `.out` / `.fout` expectations to keep.\n   * `outputFiles` is the array printed with the result and may be edited in place.\n   */\n  test?: (context: {\n    testCase: TTestCase;\n    runResult: TRunResult;\n    outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n    /** The submission's working directory, e.g. for `compareExpectedOutputFiles(cwd, testCase.fileOutputPath)`. */\n    cwd: string;\n    context: CommandJudgeContext;\n  }) => Promise<Partial<CommandJudgeCaseResult>> | Partial<CommandJudgeCaseResult> | undefined;\n}\n\n/**\n * A preset function for judging by executable command.\n *\n * Without options, test cases come from `test_cases/` (`.in`, `.out`, `.fin/`, `.fout/`) and a run\n * within the limits is accepted when its stdout matches `.out` and its files match `.fout/`; a\n * case without either expectation only has to run within the limits.\n * Keep problem-specific logic in `resolveInput` and `test`.\n *\n * @example\n * Create `judge.ts` that judges `test_cases/` like the default stdio harness, with the command\n * preset's debug mode:\n * ```ts\n * import { commandJudgePreset } from '@exercode/problem-utils/presets/command';\n *\n * await commandJudgePreset(import.meta.dirname);\n * ```\n *\n * @example\n * Create `judge.ts` with an own verdict:\n * ```ts\n * import { commandJudgePreset } from '@exercode/problem-utils/presets/command';\n * import { DecisionCode } from '@exercode/problem-utils';\n *\n * await commandJudgePreset(import.meta.dirname, {\n *   readTestCases: async () => [\n *     { id: '01', input: '1 2' },\n *   ],\n *   test: ({ runResult }) => {\n *     return runResult.stdout.trim() === '3'\n *       ? { decisionCode: DecisionCode.ACCEPTED }\n *       : { decisionCode: DecisionCode.WRONG_ANSWER };\n *   },\n * });\n * ```\n *\n * Run with the required parameters:\n * ```bash\n * bun judge.ts model_answers/python '{ \"language\": \"python\" }'\n * ```\n *\n * Run without a cwd argument to judge each `<problemDir>/model_answers/*` directory\n * for debugging. A prominent banner is printed for each run.\n * ```bash\n * bun judge.ts\n * ```\n */\nexport async function commandJudgePreset<\n  TTestCase extends BaseCommandTestCase = CommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n>(problemDir: string, options: CommandJudgePresetOptions<TTestCase, TRunResult> = {}): Promise<void> {\n  const args = parseArgs(process.argv);\n  const params = judgeParamsSchema.parse(args.params);\n\n  const { cwds, isDebugMode } = await resolveCwds(problemDir, args.cwd);\n  const problemMarkdownFrontMatter = await readProblemMarkdownFrontMatter(problemDir);\n  const testCases = await (options.readTestCases ?? readCommandTestCases)(problemDir);\n  const limits = {\n    buildTimeoutSeconds: options.limits?.buildTimeoutSeconds ?? BUILD_TIMEOUT_SECONDS,\n    maxOutputLength: options.limits?.maxOutputLength ?? MAX_STDOUT_LENGTH,\n  };\n  const timeLimitSeconds =\n    typeof problemMarkdownFrontMatter.timeLimitMs === 'number'\n      ? problemMarkdownFrontMatter.timeLimitMs / 1000\n      : (options.runTimeoutSeconds ?? JUDGE_DEFAULT_TIMEOUT_SECONDS);\n  const problem: ResolvedCommandProblem<TTestCase> = {\n    problemMarkdownFrontMatter,\n    testCases,\n    limits,\n    timeLimitSeconds,\n  };\n\n  if (isDebugMode) {\n    const expectedMaxDurationMs = (limits.buildTimeoutSeconds + testCases.length * timeLimitSeconds) * 1000;\n    if (!(await passesIsolationCheckInDebugMode(problemDir, cwds, params, { expectedMaxDurationMs }))) {\n      process.exitCode = 1;\n      return;\n    }\n  }\n\n  for (const resolvedCwd of cwds) {\n    if (isDebugMode) printDebugCwdBanner(problemDir, resolvedCwd);\n    const result = await runCommandJudgeForCwd<TTestCase, TRunResult>(resolvedCwd.cwd, params, problem, options);\n    if (isDebugMode && !matchesExpectedResult(resolvedCwd, result)) {\n      process.exitCode = 1;\n      printDebugExpectationFailureBanner(problemDir, resolvedCwd);\n    }\n  }\n}\n\nasync function runCommandJudgeForCwd<\n  TTestCase extends BaseCommandTestCase,\n  TRunResult extends CommandRunResult = CommandRunResult,\n>(\n  cwd: string,\n  params: JudgeParams,\n  problem: ResolvedCommandProblem<TTestCase>,\n  options: CommandJudgePresetOptions<TTestCase, TRunResult>\n): Promise<{ allAccepted: boolean }> {\n  const { problemMarkdownFrontMatter, testCases, limits, timeLimitSeconds } = problem;\n  const prebuildTestCaseId = testCases[0]?.id ?? 'prebuild';\n\n  const staticAnalysisResult = await judgeByStaticAnalysis(cwd, problemMarkdownFrontMatter);\n  if (staticAnalysisResult) {\n    printTestCaseResult({ testCaseId: prebuildTestCaseId, ...staticAnalysisResult });\n    return { allAccepted: false };\n  }\n\n  const originalMainFilePath = await findEntryPointFile(cwd, params.language);\n  if (!originalMainFilePath) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.MISSING_REQUIRED_SUBMISSION_FILE_ERROR,\n      stderr: `main file not found${params.language ? `: language: ${params.language}` : ''}`,\n    });\n    return { allAccepted: false };\n  }\n\n  const languageDefinition = findLanguageDefinitionByPath(originalMainFilePath);\n  if (!languageDefinition) {\n    printTestCaseResult({\n      testCaseId: prebuildTestCaseId,\n      decisionCode: DecisionCode.WRONG_ANSWER,\n      stderr: 'unsupported language',\n    });\n    return { allAccepted: false };\n  }\n\n  // `CI` changes affects Chainlit. `FORCE_COLOR` affects Bun.\n  const env = { ...process.env, CI: '', FORCE_COLOR: '0' };\n\n  let mainFilePath = originalMainFilePath;\n  if (options.buildSubmission !== false) {\n    if (languageDefinition.prebuild) {\n      try {\n        await languageDefinition.prebuild(cwd);\n        const prebuiltMainFilePath = await findEntryPointFile(cwd, params.language);\n        if (prebuiltMainFilePath) mainFilePath = prebuiltMainFilePath;\n      } catch (error) {\n        printTestCaseResult({\n          testCaseId: prebuildTestCaseId,\n          decisionCode: DecisionCode.BUILD_ERROR,\n          stderr: getErrorMessage(error),\n        });\n        return { allAccepted: false };\n      }\n    }\n\n    const buildCommand = languageDefinition.buildCommand?.(mainFilePath);\n    if (buildCommand) {\n      const buildResult = await runBuild(buildCommand, {\n        cwd,\n        env,\n        testCaseId: prebuildTestCaseId,\n        limits,\n      });\n      if (buildResult) {\n        printTestCaseResult(buildResult);\n        return { allAccepted: false };\n      }\n    }\n  }\n\n  const cwdSnapshot = await snapshotWorkingDirectory(cwd);\n\n  if (testCases.length === 0) {\n    printTestCaseResult({ testCaseId: 'default', decisionCode: DecisionCode.ACCEPTED });\n    return { allAccepted: true };\n  }\n\n  const sharedFileInputPath = (testCases as { shared?: { fileInputPath?: string } }).shared?.fileInputPath;\n\n  for (const testCase of testCases) {\n    if (sharedFileInputPath) await copyTestCaseFileInput(sharedFileInputPath, cwd);\n    if (testCase.fileInputPath) await copyTestCaseFileInput(testCase.fileInputPath, cwd);\n\n    const command = languageDefinition.command(mainFilePath);\n    let stdin = testCase.input ?? '';\n    let customRunResult: TRunResult | undefined;\n    try {\n      if (options.resolveInput) {\n        stdin = await options.resolveInput({ testCase, cwd, env });\n      }\n\n      if (options.runCommand) {\n        customRunResult = await options.runCommand({ testCase, command, stdin, cwd, env, timeLimitSeconds });\n      }\n    } catch (error) {\n      printTestCaseResult({\n        testCaseId: testCase.id,\n        decisionCode: DecisionCode.RUNTIME_ERROR,\n        stdin,\n        stderr: getErrorMessage(error),\n      });\n      await cleanWorkingDirectory(cwd, cwdSnapshot);\n      return { allAccepted: false };\n    }\n\n    // The default runner reports the submission's failures in its result; what it throws is the\n    // judge's own failure and ends the run without a verdict.\n    const runResult =\n      customRunResult ?? ((await runCommand(command, { stdin, cwd, env, timeLimitSeconds })) as TRunResult);\n\n    const outputFiles = await readOutputFiles(cwd, problemMarkdownFrontMatter.requiredOutputFilePaths ?? []);\n    const judgeContext: CommandJudgeContext = {\n      timeLimitSeconds,\n      outputLimitLength: limits.maxOutputLength,\n      problemMarkdownFrontMatter: {\n        memoryLimitByte: problemMarkdownFrontMatter.memoryLimitByte,\n        requiredOutputFilePaths: problemMarkdownFrontMatter.requiredOutputFilePaths,\n      },\n    };\n    const baseJudgeResult = evaluateByLimits({ runResult, outputFiles, context: judgeContext });\n    let judgeResult = baseJudgeResult;\n    if (baseJudgeResult.decisionCode === DecisionCode.ACCEPTED) {\n      try {\n        const extendedJudgeResult = options.test\n          ? await options.test({ testCase, runResult, outputFiles, cwd, context: judgeContext })\n          : await compareWithExpectedOutputs({ testCase, runResult, outputFiles, cwd });\n        if (extendedJudgeResult) {\n          judgeResult = {\n            decisionCode: extendedJudgeResult.decisionCode ?? baseJudgeResult.decisionCode,\n            feedbackMarkdown: extendedJudgeResult.feedbackMarkdown,\n            stderr: extendedJudgeResult.stderr,\n          };\n        }\n      } catch (error) {\n        judgeResult = {\n          decisionCode: DecisionCode.RUNTIME_ERROR,\n          stderr: getErrorMessage(error),\n        };\n      }\n    }\n\n    printTestCaseResult({\n      testCaseId: testCase.id,\n      ...judgeResult,\n      exitStatus: runResult.status,\n      stdin: runResult.stdin,\n      stdout: runResult.stdout.slice(0, limits.maxOutputLength) || undefined,\n      stderr: (judgeResult.stderr ?? runResult.stderr).slice(0, limits.maxOutputLength) || undefined,\n      timeSeconds: runResult.timeSeconds,\n      cpuTimeSeconds: runResult.cpuTimeSeconds,\n      memoryBytes: runResult.memoryBytes,\n      outputFiles: outputFiles.length > 0 ? outputFiles : undefined,\n    });\n\n    await cleanWorkingDirectory(cwd, cwdSnapshot);\n    if (judgeResult.decisionCode !== DecisionCode.ACCEPTED) return { allAccepted: false };\n  }\n\n  return { allAccepted: true };\n}\n\nfunction matchesExpectedResult(resolvedCwd: ResolvedCwd, result: { allAccepted: boolean }): boolean {\n  return result.allAccepted === (resolvedCwd.expectedResult === 'accepted');\n}\n\nasync function runBuild(\n  buildCommand: readonly [string, ...string[]],\n  context: { cwd: string; env: NodeJS.ProcessEnv; testCaseId: string; limits: CommandJudgeLimits }\n): Promise<(Omit<TestCaseResult, 'testCaseId'> & { testCaseId: string }) | undefined> {\n  const spawnResult = await spawnWithTimeout(\n    buildCommand[0],\n    buildCommand.slice(1),\n    { cwd: context.cwd, env: context.env },\n    context.limits.buildTimeoutSeconds\n  );\n  const exitStatus = spawnResult.status ?? undefined;\n\n  if (spawnResult.status !== 0) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_ERROR,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  if (spawnResult.timeSeconds > context.limits.buildTimeoutSeconds) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_TIME_LIMIT_EXCEEDED,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  if (\n    spawnResult.outputLimitExceeded ||\n    spawnResult.stdout.length > context.limits.maxOutputLength ||\n    spawnResult.stderr.length > context.limits.maxOutputLength\n  ) {\n    return {\n      testCaseId: context.testCaseId,\n      decisionCode: DecisionCode.BUILD_OUTPUT_SIZE_LIMIT_EXCEEDED,\n      exitStatus,\n      stdout: spawnResult.stdout.slice(0, context.limits.maxOutputLength),\n      stderr: spawnResult.stderr.slice(0, context.limits.maxOutputLength),\n      timeSeconds: spawnResult.timeSeconds,\n      cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n      memoryBytes: spawnResult.memoryBytes,\n    };\n  }\n\n  return;\n}\n\nasync function readCommandTestCases<TTestCase extends BaseCommandTestCase = CommandTestCase>(\n  problemDir: string\n): 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\n/** The default verdict: a case exposing a string `output` / `fileOutputPath` is judged against it. */\nasync function compareWithExpectedOutputs(context: {\n  testCase: BaseCommandTestCase & Partial<CommandTestCase>;\n  runResult: CommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  cwd: string;\n}): Promise<Partial<CommandJudgeCaseResult>> {\n  const { testCase, runResult, outputFiles, cwd } = context;\n  const judgement = await judgeAgainstExpectations({\n    stdout: runResult.stdout,\n    expectedStdout: typeof testCase.output === 'string' ? testCase.output : undefined,\n    fileOutputPath: typeof testCase.fileOutputPath === 'string' ? testCase.fileOutputPath : undefined,\n    cwd,\n    outputFiles,\n  });\n  outputFiles.splice(0, outputFiles.length, ...judgement.outputFiles);\n  return judgement.matches ? {} : { decisionCode: DecisionCode.WRONG_ANSWER };\n}\n\nasync function runCommand(\n  command: readonly [string, ...string[]],\n  context: {\n    stdin: string;\n    cwd: string;\n    env: NodeJS.ProcessEnv;\n    timeLimitSeconds: number;\n  }\n): Promise<CommandRunResult> {\n  const spawnResult = await spawnWithTimeout(\n    command[0],\n    command.slice(1),\n    { cwd: context.cwd, stdin: context.stdin, env: context.env },\n    context.timeLimitSeconds\n  );\n\n  return {\n    stdin: context.stdin,\n    stdout: spawnResult.stdout,\n    stderr: spawnResult.stderr,\n    status: spawnResult.status ?? undefined,\n    timeSeconds: spawnResult.timeSeconds,\n    cpuTimeSeconds: spawnResult.cpuTimeSeconds,\n    memoryBytes: spawnResult.memoryBytes,\n    outputLimitExceeded: spawnResult.outputLimitExceeded,\n  };\n}\n\nfunction evaluateByLimits(context: {\n  runResult: CommandRunResult;\n  outputFiles: NonNullable<TestCaseResult['outputFiles']>;\n  context: CommandJudgeContext;\n}): CommandJudgeCaseResult {\n  if (context.runResult.status !== 0) {\n    return { decisionCode: DecisionCode.RUNTIME_ERROR, stderr: context.runResult.stderr };\n  }\n\n  if (context.runResult.timeSeconds > context.context.timeLimitSeconds) {\n    return { decisionCode: DecisionCode.TIME_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  if (\n    context.runResult.memoryBytes >\n    (context.context.problemMarkdownFrontMatter.memoryLimitByte ?? Number.POSITIVE_INFINITY)\n  ) {\n    return { decisionCode: DecisionCode.MEMORY_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  if (\n    context.runResult.outputLimitExceeded ||\n    context.runResult.stdout.length > context.context.outputLimitLength ||\n    context.runResult.stderr.length > context.context.outputLimitLength\n  ) {\n    return { decisionCode: DecisionCode.OUTPUT_SIZE_LIMIT_EXCEEDED, stderr: context.runResult.stderr };\n  }\n\n  const requiredOutputFileCount = context.context.problemMarkdownFrontMatter.requiredOutputFilePaths?.length ?? 0;\n  if (context.outputFiles.length < requiredOutputFileCount) {\n    return { decisionCode: DecisionCode.MISSING_REQUIRED_OUTPUT_FILE_ERROR };\n  }\n\n  return { decisionCode: DecisionCode.ACCEPTED };\n}\n"],"mappings":"uwCA4BA,MAIM,EAAoB,EAAE,OAAO,CACjC,SAAU,EAAE,MAAM,CAAC,EAAE,OAAO,EAAG,EAAE,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAChE,CAAC,EA0ID,eAAsB,EAGpB,EAAoB,EAA4D,CAAC,EAAkB,CACnG,IAAM,EAAO,EAAU,QAAQ,IAAI,EAC7B,EAAS,EAAkB,MAAM,EAAK,MAAM,EAE5C,CAAE,OAAM,eAAgB,MAAM,EAAY,EAAY,EAAK,GAAG,EAC9D,EAA6B,MAAM,EAA+B,CAAU,EAC5E,EAAY,MAAO,EAAQ,eAAiB,EAAA,CAAsB,CAAU,EAC5E,EAAS,CACb,oBAAqB,EAAQ,QAAQ,qBAAuB,GAC5D,gBAAiB,EAAQ,QAAQ,iBAAmB,GACtD,EACM,EACJ,OAAO,EAA2B,aAAgB,SAC9C,EAA2B,YAAc,IACxC,EAAQ,mBAAqB,EAC9B,EAA6C,CACjD,6BACA,YACA,SACA,kBACF,EAEA,GAAI,EAAa,CACf,IAAM,GAAyB,EAAO,oBAAsB,EAAU,OAAS,GAAoB,IACnG,GAAI,CAAE,MAAM,EAAgC,EAAY,EAAM,EAAQ,CAAE,uBAAsB,CAAC,EAAI,CACjG,QAAQ,SAAW,EACnB,MACF,CACF,CAEA,IAAK,IAAM,KAAe,EAAM,CAC1B,GAAa,EAAoB,EAAY,CAAW,EAC5D,IAAM,EAAS,MAAM,EAA6C,EAAY,IAAK,EAAQ,EAAS,CAAO,EACvG,GAAe,CAAC,EAAsB,EAAa,CAAM,IAC3D,QAAQ,SAAW,EACnB,EAAmC,EAAY,CAAW,EAE9D,CACF,CAEA,eAAe,EAIb,EACA,EACA,EACA,EACmC,CACnC,GAAM,CAAE,6BAA4B,YAAW,SAAQ,oBAAqB,EACtE,EAAqB,EAAU,EAAE,EAAE,IAAM,WAEzC,EAAuB,MAAM,EAAsB,EAAK,CAA0B,EACxF,GAAI,EAEF,OADA,EAAoB,CAAE,WAAY,EAAoB,GAAG,CAAqB,CAAC,EACxE,CAAE,YAAa,EAAM,EAG9B,IAAM,EAAuB,MAAM,EAAmB,EAAK,EAAO,QAAQ,EAC1E,GAAI,CAAC,EAMH,OALA,EAAoB,CAClB,WAAY,EACZ,aAAc,EAAa,uCAC3B,OAAQ,sBAAsB,EAAO,SAAW,eAAe,EAAO,WAAa,IACrF,CAAC,EACM,CAAE,YAAa,EAAM,EAG9B,IAAM,EAAqB,EAA6B,CAAoB,EAC5E,GAAI,CAAC,EAMH,OALA,EAAoB,CAClB,WAAY,EACZ,aAAc,EAAa,aAC3B,OAAQ,sBACV,CAAC,EACM,CAAE,YAAa,EAAM,EAI9B,IAAM,EAAM,CAAE,GAAG,QAAQ,IAAK,GAAI,GAAI,YAAa,GAAI,EAEnD,EAAe,EACnB,GAAI,EAAQ,kBAAoB,GAAO,CACrC,GAAI,EAAmB,SACrB,GAAI,CACF,MAAM,EAAmB,SAAS,CAAG,EACrC,IAAM,EAAuB,MAAM,EAAmB,EAAK,EAAO,QAAQ,EACtE,IAAsB,EAAe,EAC3C,OAAS,EAAO,CAMd,OALA,EAAoB,CAClB,WAAY,EACZ,aAAc,EAAa,YAC3B,OAAQ,EAAgB,CAAK,CAC/B,CAAC,EACM,CAAE,YAAa,EAAM,CAC9B,CAGF,IAAM,EAAe,EAAmB,eAAe,CAAY,EACnE,GAAI,EAAc,CAChB,IAAM,EAAc,MAAM,EAAS,EAAc,CAC/C,MACA,MACA,WAAY,EACZ,QACF,CAAC,EACD,GAAI,EAEF,OADA,EAAoB,CAAW,EACxB,CAAE,YAAa,EAAM,CAEhC,CACF,CAEA,IAAM,EAAc,MAAM,EAAyB,CAAG,EAEtD,GAAI,EAAU,SAAW,EAEvB,OADA,EAAoB,CAAE,WAAY,UAAW,aAAc,EAAa,QAAS,CAAC,EAC3E,CAAE,YAAa,EAAK,EAG7B,IAAM,EAAuB,EAAsD,QAAQ,cAE3F,IAAK,IAAM,KAAY,EAAW,CAC5B,GAAqB,MAAM,EAAsB,EAAqB,CAAG,EACzE,EAAS,eAAe,MAAM,EAAsB,EAAS,cAAe,CAAG,EAEnF,IAAM,EAAU,EAAmB,QAAQ,CAAY,EACnD,EAAQ,EAAS,OAAS,GAC1B,EACJ,GAAI,CACE,EAAQ,eACV,EAAQ,MAAM,EAAQ,aAAa,CAAE,WAAU,MAAK,KAAI,CAAC,GAGvD,EAAQ,aACV,EAAkB,MAAM,EAAQ,WAAW,CAAE,WAAU,UAAS,QAAO,MAAK,MAAK,kBAAiB,CAAC,EAEvG,OAAS,EAAO,CAQd,OAPA,EAAoB,CAClB,WAAY,EAAS,GACrB,aAAc,EAAa,cAC3B,QACA,OAAQ,EAAgB,CAAK,CAC/B,CAAC,EACD,MAAM,EAAsB,EAAK,CAAW,EACrC,CAAE,YAAa,EAAM,CAC9B,CAIA,IAAM,EACJ,GAAqB,MAAM,EAAW,EAAS,CAAE,QAAO,MAAK,MAAK,kBAAiB,CAAC,EAEhF,EAAc,MAAM,EAAgB,EAAK,EAA2B,yBAA2B,CAAC,CAAC,EACjG,EAAoC,CACxC,mBACA,kBAAmB,EAAO,gBAC1B,2BAA4B,CAC1B,gBAAiB,EAA2B,gBAC5C,wBAAyB,EAA2B,uBACtD,CACF,EACM,EAAkB,EAAiB,CAAE,YAAW,cAAa,QAAS,CAAa,CAAC,EACtF,EAAc,EAClB,GAAI,EAAgB,eAAiB,EAAa,SAChD,GAAI,CACF,IAAM,EAAsB,EAAQ,KAChC,MAAM,EAAQ,KAAK,CAAE,WAAU,YAAW,cAAa,MAAK,QAAS,CAAa,CAAC,EACnF,MAAM,EAA2B,CAAE,WAAU,YAAW,cAAa,KAAI,CAAC,EAC1E,IACF,EAAc,CACZ,aAAc,EAAoB,cAAgB,EAAgB,aAClE,iBAAkB,EAAoB,iBACtC,OAAQ,EAAoB,MAC9B,EAEJ,OAAS,EAAO,CACd,EAAc,CACZ,aAAc,EAAa,cAC3B,OAAQ,EAAgB,CAAK,CAC/B,CACF,CAiBF,GAdA,EAAoB,CAClB,WAAY,EAAS,GACrB,GAAG,EACH,WAAY,EAAU,OACtB,MAAO,EAAU,MACjB,OAAQ,EAAU,OAAO,MAAM,EAAG,EAAO,eAAe,GAAK,IAAA,GAC7D,QAAS,EAAY,QAAU,EAAU,OAAA,CAAQ,MAAM,EAAG,EAAO,eAAe,GAAK,IAAA,GACrF,YAAa,EAAU,YACvB,eAAgB,EAAU,eAC1B,YAAa,EAAU,YACvB,YAAa,EAAY,OAAS,EAAI,EAAc,IAAA,EACtD,CAAC,EAED,MAAM,EAAsB,EAAK,CAAW,EACxC,EAAY,eAAiB,EAAa,SAAU,MAAO,CAAE,YAAa,EAAM,CACtF,CAEA,MAAO,CAAE,YAAa,EAAK,CAC7B,CAEA,SAAS,EAAsB,EAA0B,EAA2C,CAClG,OAAO,EAAO,eAAiB,EAAY,iBAAmB,WAChE,CAEA,eAAe,EACb,EACA,EACoF,CACpF,IAAM,EAAc,MAAM,EACxB,EAAa,GACb,EAAa,MAAM,CAAC,EACpB,CAAE,IAAK,EAAQ,IAAK,IAAK,EAAQ,GAAI,EACrC,EAAQ,OAAO,mBACjB,EACM,EAAa,EAAY,QAAU,IAAA,GAEzC,GAAI,EAAY,SAAW,EACzB,MAAO,CACL,WAAY,EAAQ,WACpB,aAAc,EAAa,YAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,EAGF,GAAI,EAAY,YAAc,EAAQ,OAAO,oBAC3C,MAAO,CACL,WAAY,EAAQ,WACpB,aAAc,EAAa,0BAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,EAGF,GACE,EAAY,qBACZ,EAAY,OAAO,OAAS,EAAQ,OAAO,iBAC3C,EAAY,OAAO,OAAS,EAAQ,OAAO,gBAE3C,MAAO,CACL,WAAY,EAAQ,WACpB,aAAc,EAAa,iCAC3B,aACA,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,OAAQ,EAAY,OAAO,MAAM,EAAG,EAAQ,OAAO,eAAe,EAClE,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,WAC3B,CAIJ,CAEA,eAAe,EACb,EAC+B,CAE/B,OAAQ,MAAMA,EAAkB,EAAK,KAAK,EAAY,YAAY,CAAC,CACrE,CAGA,eAAe,EAA2B,EAKG,CAC3C,GAAM,CAAE,WAAU,YAAW,cAAa,OAAQ,EAC5C,EAAY,MAAM,EAAyB,CAC/C,OAAQ,EAAU,OAClB,eAAgB,OAAO,EAAS,QAAW,SAAW,EAAS,OAAS,IAAA,GACxE,eAAgB,OAAO,EAAS,gBAAmB,SAAW,EAAS,eAAiB,IAAA,GACxF,MACA,aACF,CAAC,EAED,OADA,EAAY,OAAO,EAAG,EAAY,OAAQ,GAAG,EAAU,WAAW,EAC3D,EAAU,QAAU,CAAC,EAAI,CAAE,aAAc,EAAa,YAAa,CAC5E,CAEA,eAAe,EACb,EACA,EAM2B,CAC3B,IAAM,EAAc,MAAM,EACxB,EAAQ,GACR,EAAQ,MAAM,CAAC,EACf,CAAE,IAAK,EAAQ,IAAK,MAAO,EAAQ,MAAO,IAAK,EAAQ,GAAI,EAC3D,EAAQ,gBACV,EAEA,MAAO,CACL,MAAO,EAAQ,MACf,OAAQ,EAAY,OACpB,OAAQ,EAAY,OACpB,OAAQ,EAAY,QAAU,IAAA,GAC9B,YAAa,EAAY,YACzB,eAAgB,EAAY,eAC5B,YAAa,EAAY,YACzB,oBAAqB,EAAY,mBACnC,CACF,CAEA,SAAS,EAAiB,EAIC,CACzB,GAAI,EAAQ,UAAU,SAAW,EAC/B,MAAO,CAAE,aAAc,EAAa,cAAe,OAAQ,EAAQ,UAAU,MAAO,EAGtF,GAAI,EAAQ,UAAU,YAAc,EAAQ,QAAQ,iBAClD,MAAO,CAAE,aAAc,EAAa,oBAAqB,OAAQ,EAAQ,UAAU,MAAO,EAG5F,GACE,EAAQ,UAAU,aACjB,EAAQ,QAAQ,2BAA2B,iBAAmB,KAE/D,MAAO,CAAE,aAAc,EAAa,sBAAuB,OAAQ,EAAQ,UAAU,MAAO,EAG9F,GACE,EAAQ,UAAU,qBAClB,EAAQ,UAAU,OAAO,OAAS,EAAQ,QAAQ,mBAClD,EAAQ,UAAU,OAAO,OAAS,EAAQ,QAAQ,kBAElD,MAAO,CAAE,aAAc,EAAa,2BAA4B,OAAQ,EAAQ,UAAU,MAAO,EAGnG,IAAM,EAA0B,EAAQ,QAAQ,2BAA2B,yBAAyB,QAAU,EAK9G,OAJI,EAAQ,YAAY,OAAS,EACxB,CAAE,aAAc,EAAa,kCAAmC,EAGlE,CAAE,aAAc,EAAa,QAAS,CAC/C"}

package/dist/presets/evaluation.cjs

@@ -3,9 +3,9 @@
 | ${i} | ${_(e,o)} |
 
 評価件数: ${t.toLocaleString(`en-US`)}件${s}
-`}function g(e,t){if(t===void 0||e===t)return 6;for(let n=6;n<17;n++)if(_(e,n)!==_(t,n))return n;return 17}function _(e,t){return Number(e.toPrecision(t)).toString()}function v(e,t,n,i){let a;try{a=r.parseCsvRecords(e)}catch(e){return`\`${t}\`をCSVとして読み込めませんでした。${e instanceof r.CsvParseError?`${e.lineNumber}行目付近: ${d[e.reason]}`:``}`}let o=b(a[0]?.cells??[],i);if(!o)return`\`${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 s=new Map,c=[],l=0,u=e=>{l++,c.length<10&&c.push(e)};for(let{cells:e,lineNumber:t}of a.slice(1)){if(x(e))continue;let r=e[o.idIndex]?.trim()??``,a=e[o.targetIndex]??``;if(!n.has(r))u(`${t}行目: \`${i.idColumn}\`が\`${C(r)}\`の行は評価対象ではありません`);else if(s.has(r))u(`${t}行目: \`${i.idColumn}\`が\`${C(r)}\`の行が重複しています`);else{let e=i.metric.validatePrediction?.(a);e&&u(`${t}行目: \`${i.targetColumn}\`の値\`${C(a)}\`は${e}`),s.set(r,a)}}let f=[...n.keys()].filter(e=>!s.has(e));return f.length>0&&(l++,c.unshift(`\`${i.idColumn}\`が${S(f)}の行がありません（不足 ${f.length.toLocaleString(`en-US`)}件）`)),l>0?`\`${t}\`の内容に問題があります。
+`}function g(e,t){if(t===void 0||e===t)return 6;for(let n=6;n<17;n++)if(_(e,n)!==_(t,n))return n;return 17}function _(e,t){return Number(e.toPrecision(t)).toString()}function v(e,t,n,i){let a;try{a=r.parseCsvRecords(e)}catch(e){return`\`${t}\`をCSVとして読み込めませんでした。${e instanceof r.CsvParseError?`${e.lineNumber}行目付近: ${d[e.reason]}`:``}`}let o=b(a[0]?.cells??[],i);if(!o)return`\`${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 s=/* @__PURE__ */ new Map,c=[],l=0,u=e=>{l++,c.length<10&&c.push(e)};for(let{cells:e,lineNumber:t}of a.slice(1)){if(x(e))continue;let r=e[o.idIndex]?.trim()??``,a=e[o.targetIndex]??``;if(!n.has(r))u(`${t}行目: \`${i.idColumn}\`が\`${C(r)}\`の行は評価対象ではありません`);else if(s.has(r))u(`${t}行目: \`${i.idColumn}\`が\`${C(r)}\`の行が重複しています`);else{let e=i.metric.validatePrediction?.(a);e&&u(`${t}行目: \`${i.targetColumn}\`の値\`${C(a)}\`は${e}`),s.set(r,a)}}let f=[...n.keys()].filter(e=>!s.has(e));return f.length>0&&(l++,c.unshift(`\`${i.idColumn}\`が${S(f)}の行がありません（不足 ${f.length.toLocaleString(`en-US`)}件）`)),l>0?`\`${t}\`の内容に問題があります。
 
 ${c.map(e=>`- ${e}`).join(`
 `)}${l>c.length?`\n- ほか${(l-c.length).toLocaleString(`en-US`)}件`:``}
-`:s}async function y(e,t){let n=r.parseCsv(await u.default.readFile(e,`utf8`)),i=b(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(x(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: ${e}`);return a}function b(e,t){let n=e.map(e=>e.trim()),r=n.indexOf(t.idColumn),i=n.indexOf(t.targetColumn);return r===-1||i===-1?void 0:{idIndex:r,targetIndex:i}}function x(e){return e.every(e=>e.trim()===``)}function S(e){let t=e.slice(0,10).map(e=>`\`${C(e)}\``);return e.length>10?`${t.join(`, `)}, ...`:t.join(`, `)}function C(e){let t=e.replaceAll(/[\p{Cc}]/gu,` `).replaceAll("`",`'`);return t.trim()===``?`（空）`:t.length>50?`${t.slice(0,50)}…`:t}function w(e,t){return t.decisionCode===a.DecisionCode.ACCEPTED==(e.expectedResult===`accepted`)}function T(e){return Number.isFinite(Number(e))&&e.trim()!==``?void 0:`数値ではありません`}function E(e){return e.reduce((e,t)=>e+t,0)/e.length}exports.evaluationJudgePreset=p,exports.evaluationMetrics=f;
+`:s}async function y(e,t){let n=r.parseCsv(await u.default.readFile(e,`utf8`)),i=b(n[0]??[],t);if(!i)throw Error(`answer file must have ${t.idColumn} and ${t.targetColumn} columns: ${e}`);let a=/* @__PURE__ */ new Map;for(let r of n.slice(1)){if(x(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: ${e}`);return a}function b(e,t){let n=e.map(e=>e.trim()),r=n.indexOf(t.idColumn),i=n.indexOf(t.targetColumn);return r===-1||i===-1?void 0:{idIndex:r,targetIndex:i}}function x(e){return e.every(e=>e.trim()===``)}function S(e){let t=e.slice(0,10).map(e=>`\`${C(e)}\``);return e.length>10?`${t.join(`, `)}, ...`:t.join(`, `)}function C(e){let t=e.replaceAll(/[\p{Cc}]/gu,` `).replaceAll("`",`'`);return t.trim()===``?`（空）`:t.length>50?`${t.slice(0,50)}…`:t}function w(e,t){return t.decisionCode===a.DecisionCode.ACCEPTED==(e.expectedResult===`accepted`)}function T(e){return Number.isFinite(Number(e))&&e.trim()!==``?void 0:`数値ではありません`}function E(e){return e.reduce((e,t)=>e+t,0)/e.length}exports.evaluationJudgePreset=p,exports.evaluationMetrics=f;
 //# sourceMappingURL=evaluation.cjs.map

package/dist/presets/evaluation.cjs.map

@@ -1 +1 @@
-{"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"}
+{"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,kBAAc,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,kBAAS,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

@@ -3,9 +3,9 @@ import{printTestCaseResult as e}from"../helpers/printTestCaseResult.js";import{p
 | ${i} | ${b(e,o)} |
 
 評価件数: ${t.toLocaleString(`en-US`)}件${s}
-`}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。
+`}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=/* @__PURE__ */ 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 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===o.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};
+`: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=/* @__PURE__ */ 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===o.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