bridgebench 3.1.0-alpha.0

package/dist/cli.js

@@ -0,0 +1,408 @@
+#!/usr/bin/env node
+import {
+  publishJournal,
+  publishTarget,
+  publishTasks,
+  resolveApiConfig,
+  runRemoteArena
+} from "./chunk-LFKEV2YL.js";
+import {
+  ArenaRunner,
+  ArenaStore,
+  ENGINE_VERSION,
+  PACKAGE_NAME,
+  categoryStoreConfig,
+  formatTriage,
+  triageJournal,
+  verifyJournal,
+  writeReports
+} from "./chunk-EQHRUV2I.js";
+import {
+  FileArenaLogger,
+  OpenRouterClient,
+  TaskLoader,
+  findProjectRoot,
+  listModels,
+  redactSecrets,
+  resolveCompetitorRoster,
+  sanitizeError,
+  validatePublicTaskFile
+} from "./chunk-JTVNKSMO.js";
+import {
+  BenchmarkCategorySchema,
+  CATEGORIES
+} from "./chunk-UECBSKTD.js";
+
+// src/commands.ts
+import path2 from "path";
+import { Command, InvalidArgumentError } from "commander";
+
+// src/env.ts
+import { existsSync } from "fs";
+import path from "path";
+var ENV_PATH = path.join(findProjectRoot(import.meta.url), ".env");
+function loadProjectEnv() {
+  if (!existsSync(ENV_PATH)) return { status: "missing" };
+  try {
+    process.loadEnvFile(ENV_PATH);
+    return { status: "loaded" };
+  } catch (error) {
+    return {
+      status: "error",
+      reason: error instanceof Error ? error.message : String(error)
+    };
+  }
+}
+
+// src/commands.ts
+var ROOT = process.cwd();
+var defaultDependencies = {
+  stdout: (message) => console.log(message),
+  stderr: (message) => console.error(message),
+  setExitCode: (code) => {
+    process.exitCode = code;
+  },
+  createStore: (category) => new ArenaStore(categoryStoreConfig(category)),
+  createTaskLoader: (category) => new TaskLoader(category),
+  createOpenRouter: (logger) => new OpenRouterClient(process.env.OPENROUTER_API_KEY ?? "", logger),
+  publishTasks,
+  publishJournal,
+  resolveApiConfig
+};
+function parseCategory(value) {
+  const parsed = BenchmarkCategorySchema.safeParse(value);
+  if (!parsed.success) {
+    throw new InvalidArgumentError(`expected one of ${CATEGORIES.join(", ")}, received ${value}`);
+  }
+  return parsed.data;
+}
+function positiveInteger(value) {
+  const parsed = Number(value);
+  if (!Number.isInteger(parsed) || parsed < 1) {
+    throw new InvalidArgumentError(`expected a positive integer, received ${value}`);
+  }
+  return parsed;
+}
+function positiveNumber(value) {
+  const parsed = Number(value);
+  if (!Number.isFinite(parsed) || parsed <= 0) {
+    throw new InvalidArgumentError(`expected a positive number, received ${value}`);
+  }
+  return parsed;
+}
+function collectOption(value, previous = []) {
+  return [...previous, value];
+}
+function arenaRunConfigFromOptions(options) {
+  const competitorIds = options.competitor?.length ? resolveCompetitorRoster(options.competitor).map((model) => model.id) : void 0;
+  return {
+    category: options.category,
+    seed: options.seed,
+    matches: options.matches,
+    maxCostUsd: options.maxCostUsd,
+    competitorIds,
+    resume: options.resume,
+    healthStop: options.healthStop
+  };
+}
+function selectedCategories(options, command) {
+  if (options.category && options.all) {
+    throw new Error(`${command} accepts either --category or --all, not both`);
+  }
+  if (options.category) return [options.category];
+  if (options.all) return [...CATEGORIES];
+  throw new Error(`${command} requires --category <category> or --all`);
+}
+function logRoot(category) {
+  return path2.join(ROOT, "results", category, "logs");
+}
+function displayPath(target) {
+  const relative = path2.relative(ROOT, target);
+  return relative.startsWith("..") ? path2.basename(target) : relative;
+}
+function buildProgram(overrides = {}) {
+  const dependencies = { ...defaultDependencies, ...overrides };
+  const program = new Command().name(PACKAGE_NAME).description("Autonomous BridgeBench arenas for reasoning and hallucination").version(ENGINE_VERSION).showHelpAfterError().configureHelp({ sortOptions: true, sortSubcommands: true }).configureOutput({
+    writeOut: (value) => dependencies.stdout(value.replace(/\n$/, "")),
+    writeErr: (value) => dependencies.stderr(value.replace(/\n$/, ""))
+  });
+  const models = program.command("models").description("Inspect or validate the pinned model registry");
+  models.command("list").description("List enabled competitors and judges").action(() => {
+    for (const model of listModels()) {
+      dependencies.stdout(`${model.role.padEnd(10)} ${model.id.padEnd(40)} ${model.displayName}`);
+    }
+  });
+  models.command("validate").description("Validate pinned model IDs and judge capabilities against OpenRouter").action(async () => {
+    const client = dependencies.createOpenRouter();
+    for (const model of listModels()) {
+      await client.validateModel(model);
+      dependencies.stdout(`\u2713 ${model.id} -> ${model.canonicalSlug}`);
+    }
+  });
+  const tasks = program.command("tasks").description("Validate or publish task packs");
+  tasks.command("validate").description("Validate public task schemas and pack balance").option(
+    "-c, --category <category>",
+    `limit to one category (${CATEGORIES.join(", ")})`,
+    parseCategory
+  ).option("--file <path>", "validate one public task without pack-balance checks").action(async (options) => {
+    if (options.file) {
+      if (options.category) {
+        throw new Error("tasks validate accepts --file or --category, not both");
+      }
+      const task = await validatePublicTaskFile(options.file);
+      dependencies.stdout(
+        `\u2713 ${task.public.id}: public task validated (${task.public.category}/${task.public.cluster})`
+      );
+      return;
+    }
+    const categories = options.category ? [options.category] : CATEGORIES;
+    for (const category of categories) {
+      const loader = dependencies.createTaskLoader(category);
+      const loaded = await loader.loadAll();
+      const mode = loader.hasPrivate ? "public + private halves" : "public halves only";
+      dependencies.stdout(`\u2713 ${category}: ${loaded.length} tasks validated (${mode})`);
+      for (const task of loaded) {
+        dependencies.stdout(`  ${task.public.cluster.padEnd(28)} ${task.public.id}`);
+      }
+    }
+  });
+  tasks.command("publish").description("Publish public and private task halves to the configured API").option(
+    "-c, --category <category>",
+    `category to publish (${CATEGORIES.join(", ")})`,
+    parseCategory
+  ).option("--all", "publish every category explicitly", false).action(async (options) => {
+    const categories = selectedCategories(options, "tasks publish");
+    const config = dependencies.resolveApiConfig();
+    dependencies.stdout(`Publishing ${categories.join(", ")} tasks to ${publishTarget(config)}`);
+    for (const category of categories) {
+      const result = await dependencies.publishTasks(category, config);
+      dependencies.stdout(`\u2713 ${category}: published ${result.imported} tasks`);
+    }
+  });
+  const arena = program.command("arena").description("Run and inspect arena matches");
+  arena.command("run").description("Run a paid deterministic match schedule").requiredOption(
+    "-c, --category <category>",
+    `arena to run (${CATEGORIES.join(", ")})`,
+    parseCategory
+  ).option("-n, --matches <count>", "number of task-level matches", positiveInteger, 12).option("-s, --seed <seed>", "reproducible scheduling seed", "bridgebench-v3-mvp").option(
+    "--max-cost-usd <amount>",
+    "hard stop before the next match after this spend",
+    positiveNumber,
+    25
+  ).option(
+    "--competitor <modelId>",
+    "limit the run roster; repeat for each competitor (minimum two)",
+    collectOption
+  ).option("--resume", "skip match IDs already present in the journal", false).option("--debug", "mirror structured log entries to the console", false).option("--no-health-stop", "do not halt when most matches have failed responses").action(async (options) => {
+    const config = arenaRunConfigFromOptions(options);
+    const logger = new FileArenaLogger({
+      dir: logRoot(config.category),
+      verbose: options.debug
+    });
+    const cancellation = new AbortController();
+    const requestCancellation = () => {
+      if (cancellation.signal.aborted) return;
+      dependencies.stderr("Cancellation requested; aborting active model calls.");
+      cancellation.abort();
+    };
+    process.once("SIGINT", requestCancellation);
+    try {
+      dependencies.stdout(`Run log: ${displayPath(logger.filePath)}`);
+      const store = dependencies.createStore(config.category);
+      const loaded = await dependencies.createTaskLoader(config.category).loadAll({ requirePrivate: true });
+      const progressOutput = (event) => {
+        if (event.type !== "match.completed") return;
+        dependencies.stdout(
+          `[${event.data.completed}/${event.data.total}] ${event.data.taskId}: ${event.data.winnerModelId ?? "no contest"} (${event.data.outcome}, $${event.data.costUsd.toFixed(4)})`
+        );
+      };
+      const runner = new ArenaRunner(
+        dependencies.createOpenRouter(logger),
+        store,
+        progressOutput,
+        logger
+      );
+      const result = await runner.run(config, loaded, { signal: cancellation.signal });
+      if (result.cancelled) {
+        dependencies.stdout(
+          `Cancelled after ${result.completed} new ${config.category} matches; journaled spend $${result.costUsd.toFixed(4)}. Completed matches remain journaled.`
+        );
+        dependencies.setExitCode(130);
+      } else {
+        dependencies.stdout(
+          `Completed ${result.completed} new ${config.category} matches; run spend $${result.costUsd.toFixed(4)}.`
+        );
+        if (result.stoppedForBudget) {
+          dependencies.stdout(
+            "Stopped at the cost boundary; rerun the same command with --resume."
+          );
+        }
+      }
+      const runMatches = store.readAll().filter((match) => match.runId === result.runId);
+      if (runMatches.length > 0) {
+        const triage = triageJournal(runMatches);
+        dependencies.stdout(`=== Run health ===
+${formatTriage(triage)}`);
+      }
+    } finally {
+      process.removeListener("SIGINT", requestCancellation);
+      dependencies.stdout(`Run log: ${displayPath(logger.filePath)}`);
+    }
+  });
+  arena.command("remote-run").description("Run a match schedule against the configured API and stream live arena events").requiredOption(
+    "-c, --category <category>",
+    `arena to run (${CATEGORIES.join(", ")})`,
+    parseCategory
+  ).option("-n, --matches <count>", "number of task-level matches", positiveInteger, 12).option("-s, --seed <seed>", "reproducible scheduling seed", "bridgebench-v3-mvp").option(
+    "--max-cost-usd <amount>",
+    "hard stop before the next match after this spend",
+    positiveNumber,
+    25
+  ).option(
+    "--competitor <modelId>",
+    "limit the run roster; repeat for each competitor (minimum two)",
+    collectOption
+  ).option("--resume", "skip match IDs already present in the journal", false).option("--debug", "mirror structured log entries to the console", false).option("--no-health-stop", "do not halt when most matches have failed responses").option("--mock", "use deterministic mock model completions instead of OpenRouter", false).option("--no-publish-matches", "skip publishing match results to the API as they complete").action(async (options) => {
+    const config = arenaRunConfigFromOptions(options);
+    const apiConfig = dependencies.resolveApiConfig();
+    const logger = new FileArenaLogger({
+      dir: logRoot(config.category),
+      verbose: options.debug
+    });
+    dependencies.stdout(
+      `Remote ${config.category} run against ${publishTarget(apiConfig)} (${options.mock ? "mock" : "openrouter"})`
+    );
+    dependencies.stdout(`Run log: ${displayPath(logger.filePath)}`);
+    try {
+      const result = await runRemoteArena(apiConfig, {
+        config,
+        mock: options.mock,
+        publishMatches: options.publishMatches,
+        logger
+      });
+      dependencies.stdout(
+        `Remote run ${result.runKey}: ${result.completed} matches, $${result.costUsd.toFixed(4)}`
+      );
+      if (result.cancelled) dependencies.setExitCode(130);
+    } finally {
+      dependencies.stdout(`Run log: ${displayPath(logger.filePath)}`);
+    }
+  });
+  arena.command("verify").description("Validate and replay outcomes, points, manifests, and Elo").option(
+    "-c, --category <category>",
+    `journal category (${CATEGORIES.join(", ")})`,
+    parseCategory,
+    "reasoning"
+  ).option("--journal <path>", "journal file to verify (defaults to the local category journal)").option("--manifests-dir <path>", "run-manifest directory (defaults to a runs/ sibling)").option("--json", "emit the machine-readable verification result", false).action(
+    (options) => {
+      const defaults = categoryStoreConfig(options.category);
+      const journalPath = options.journal ? path2.resolve(options.journal) : defaults.journalPath;
+      const runsDir = options.manifestsDir ? path2.resolve(options.manifestsDir) : options.journal ? path2.join(path2.dirname(journalPath), "runs") : defaults.runsDir;
+      const store = new ArenaStore({
+        ...defaults,
+        journalPath,
+        runsDir,
+        readOnly: true
+      });
+      const verified = verifyJournal(store.readAll(), options.category, {
+        manifestForRun: (runId) => store.readRunManifest(runId),
+        requireManifests: true
+      });
+      if (options.json) {
+        dependencies.stdout(JSON.stringify(verified, null, 2));
+        return;
+      }
+      dependencies.stdout(
+        `\u2713 ${options.category}: verified ${verified.matches.length} matches across ${verified.runs.length} runs`
+      );
+      for (const warning of verified.warnings) {
+        dependencies.stderr(`Warning: ${warning}`);
+      }
+    }
+  );
+  arena.command("publish").description("Publish a verified match journal to the configured API").option(
+    "-c, --category <category>",
+    `journal to publish (${CATEGORIES.join(", ")})`,
+    parseCategory
+  ).option("--all", "publish every category explicitly", false).option("--allow-empty", "treat an empty journal as a successful no-op", false).action(async (options) => {
+    const categories = selectedCategories(options, "arena publish");
+    const config = dependencies.resolveApiConfig();
+    dependencies.stdout(
+      `Publishing ${categories.join(", ")} journals to ${publishTarget(config)}`
+    );
+    for (const category of categories) {
+      const store = dependencies.createStore(category);
+      verifyJournal(store.readAll(), category, {
+        manifestForRun: (runId) => store.readRunManifest(runId),
+        requireManifests: true
+      });
+      const result = await dependencies.publishJournal(category, config);
+      if (result.matches === 0 && !options.allowEmpty) {
+        throw new Error(
+          `The ${category} journal is empty; pass --allow-empty for an intentional no-op`
+        );
+      }
+      dependencies.stdout(
+        `\u2713 ${category}: ${result.imported} new, ${result.skipped} already present (of ${result.matches} lines)`
+      );
+    }
+  });
+  arena.command("triage").description("Analyze failures, suspicious responses, and judging anomalies").option(
+    "-c, --category <category>",
+    `journal to analyze (${CATEGORIES.join(", ")})`,
+    parseCategory,
+    "reasoning"
+  ).option("--run <runId>", "limit the report to one run ID").option("--json", "emit the machine-readable report", false).option("--strict", "exit nonzero when any anomaly is found", false).option("--allow-empty", "treat no matching journal lines as success", false).action(
+    (options) => {
+      const results = dependencies.createStore(options.category).readAll().filter((match) => !options.run || match.runId === options.run);
+      if (results.length === 0) {
+        const message = options.run ? `No ${options.category} matches found for run ${options.run}.` : `The ${options.category} journal is empty.`;
+        if (!options.allowEmpty) throw new Error(message);
+        dependencies.stdout(message);
+        return;
+      }
+      const reports = triageJournal(results);
+      dependencies.stdout(
+        options.json ? JSON.stringify(reports, null, 2) : formatTriage(reports)
+      );
+      const anomalies = reports.reduce((sum, report) => sum + report.anomalies.length, 0);
+      if (options.strict && anomalies > 0) {
+        dependencies.stderr(`Strict mode: ${anomalies} anomalies found.`);
+        dependencies.setExitCode(1);
+      }
+    }
+  );
+  arena.command("generation <generationId>").description("Fetch OpenRouter's record for a journaled generation").action(async (generationId) => {
+    const data = await dependencies.createOpenRouter().fetchGeneration(generationId);
+    dependencies.stdout(JSON.stringify(data, null, 2));
+  });
+  program.command("report").description("Verify journals and rebuild JSON and Markdown reports").option(
+    "-c, --category <category>",
+    `limit to one category (${CATEGORIES.join(", ")})`,
+    parseCategory
+  ).action((options) => {
+    const categories = options.category ? [options.category] : CATEGORIES;
+    for (const category of categories) {
+      const snapshot = writeReports(dependencies.createStore(category));
+      dependencies.stdout(`\u2713 ${category}: wrote reports for ${snapshot.matches.length} matches`);
+    }
+  });
+  return program;
+}
+async function runCli(argv = process.argv, overrides = {}) {
+  const dependencies = { ...defaultDependencies, ...overrides };
+  const envResult = loadProjectEnv();
+  if (envResult.status === "error") {
+    dependencies.stderr(
+      `Warning: project environment file could not be loaded: ${redactSecrets(envResult.reason)}`
+    );
+  }
+  await buildProgram(dependencies).parseAsync(argv);
+}
+
+// src/cli.ts
+runCli().catch((error) => {
+  console.error(`BridgeBench: ${sanitizeError(error)}`);
+  process.exitCode = 1;
+});
+//# sourceMappingURL=cli.js.map

package/dist/cli.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/commands.ts","../src/env.ts","../src/cli.ts"],"sourcesContent":["import path from 'node:path';\n\nimport { Command, InvalidArgumentError } from 'commander';\n\nimport { ArenaRunner } from './arena.js';\nimport { loadProjectEnv } from './env.js';\nimport { FileArenaLogger, redactSecrets, type ArenaLogger } from './logger.js';\nimport { listModels, resolveCompetitorRoster } from './models.js';\nimport { OpenRouterClient } from './openrouter.js';\nimport { publishJournal, publishTarget, publishTasks, resolveApiConfig } from './publish.js';\nimport { runRemoteArena } from './remote-arena.js';\nimport { writeReports } from './report.js';\nimport { ArenaStore, categoryStoreConfig } from './store.js';\nimport { TaskLoader, validatePublicTaskFile } from './tasks.js';\nimport { formatTriage, triageJournal } from './triage.js';\nimport {\n  BenchmarkCategorySchema,\n  CATEGORIES,\n  type ArenaEventSink,\n  type ArenaRunConfig,\n  type BenchmarkCategory,\n} from './types.js';\nimport { verifyJournal } from './verification.js';\nimport { ENGINE_VERSION, PACKAGE_NAME } from './version.js';\n\n// CLI logs and path display anchor to the operator's working directory,\n// matching where categoryStoreConfig writes results.\nconst ROOT = process.cwd();\n\ninterface CategorySelection {\n  category?: BenchmarkCategory;\n  all?: boolean;\n}\n\nexport interface ArenaRunCliOptions {\n  category: BenchmarkCategory;\n  matches: number;\n  seed: string;\n  maxCostUsd: number;\n  competitor?: string[];\n  resume: boolean;\n  debug: boolean;\n  healthStop: boolean;\n}\n\nexport interface CliDependencies {\n  stdout: (message: string) => void;\n  stderr: (message: string) => void;\n  setExitCode: (code: number) => void;\n  createStore: (category: BenchmarkCategory) => ArenaStore;\n  createTaskLoader: (category: BenchmarkCategory) => TaskLoader;\n  createOpenRouter: (logger?: ArenaLogger) => OpenRouterClient;\n  publishTasks: typeof publishTasks;\n  publishJournal: typeof publishJournal;\n  resolveApiConfig: typeof resolveApiConfig;\n}\n\nconst defaultDependencies: CliDependencies = {\n  stdout: (message) => console.log(message),\n  stderr: (message) => console.error(message),\n  setExitCode: (code) => {\n    process.exitCode = code;\n  },\n  createStore: (category) => new ArenaStore(categoryStoreConfig(category)),\n  createTaskLoader: (category) => new TaskLoader(category),\n  createOpenRouter: (logger) => new OpenRouterClient(process.env.OPENROUTER_API_KEY ?? '', logger),\n  publishTasks,\n  publishJournal,\n  resolveApiConfig,\n};\n\nfunction parseCategory(value: string): BenchmarkCategory {\n  const parsed = BenchmarkCategorySchema.safeParse(value);\n  if (!parsed.success) {\n    throw new InvalidArgumentError(`expected one of ${CATEGORIES.join(', ')}, received ${value}`);\n  }\n  return parsed.data;\n}\n\nfunction positiveInteger(value: string): number {\n  const parsed = Number(value);\n  if (!Number.isInteger(parsed) || parsed < 1) {\n    throw new InvalidArgumentError(`expected a positive integer, received ${value}`);\n  }\n  return parsed;\n}\n\nfunction positiveNumber(value: string): number {\n  const parsed = Number(value);\n  if (!Number.isFinite(parsed) || parsed <= 0) {\n    throw new InvalidArgumentError(`expected a positive number, received ${value}`);\n  }\n  return parsed;\n}\n\nfunction collectOption(value: string, previous: string[] = []): string[] {\n  return [...previous, value];\n}\n\nexport function arenaRunConfigFromOptions(options: ArenaRunCliOptions): ArenaRunConfig {\n  const competitorIds = options.competitor?.length\n    ? resolveCompetitorRoster(options.competitor).map((model) => model.id)\n    : undefined;\n  return {\n    category: options.category,\n    seed: options.seed,\n    matches: options.matches,\n    maxCostUsd: options.maxCostUsd,\n    competitorIds,\n    resume: options.resume,\n    healthStop: options.healthStop,\n  };\n}\n\nfunction selectedCategories(options: CategorySelection, command: string): BenchmarkCategory[] {\n  if (options.category && options.all) {\n    throw new Error(`${command} accepts either --category or --all, not both`);\n  }\n  if (options.category) return [options.category];\n  if (options.all) return [...CATEGORIES];\n  throw new Error(`${command} requires --category <category> or --all`);\n}\n\nfunction logRoot(category: BenchmarkCategory): string {\n  return path.join(ROOT, 'results', category, 'logs');\n}\n\nfunction displayPath(target: string): string {\n  const relative = path.relative(ROOT, target);\n  return relative.startsWith('..') ? path.basename(target) : relative;\n}\n\nexport function buildProgram(overrides: Partial<CliDependencies> = {}): Command {\n  const dependencies: CliDependencies = { ...defaultDependencies, ...overrides };\n  const program = new Command()\n    .name(PACKAGE_NAME)\n    .description('Autonomous BridgeBench arenas for reasoning and hallucination')\n    .version(ENGINE_VERSION)\n    .showHelpAfterError()\n    .configureHelp({ sortOptions: true, sortSubcommands: true })\n    .configureOutput({\n      writeOut: (value) => dependencies.stdout(value.replace(/\\n$/, '')),\n      writeErr: (value) => dependencies.stderr(value.replace(/\\n$/, '')),\n    });\n\n  const models = program\n    .command('models')\n    .description('Inspect or validate the pinned model registry');\n  models\n    .command('list')\n    .description('List enabled competitors and judges')\n    .action(() => {\n      for (const model of listModels()) {\n        dependencies.stdout(`${model.role.padEnd(10)} ${model.id.padEnd(40)} ${model.displayName}`);\n      }\n    });\n  models\n    .command('validate')\n    .description('Validate pinned model IDs and judge capabilities against OpenRouter')\n    .action(async () => {\n      const client = dependencies.createOpenRouter();\n      for (const model of listModels()) {\n        await client.validateModel(model);\n        dependencies.stdout(`✓ ${model.id} -> ${model.canonicalSlug}`);\n      }\n    });\n\n  const tasks = program.command('tasks').description('Validate or publish task packs');\n  tasks\n    .command('validate')\n    .description('Validate public task schemas and pack balance')\n    .option(\n      '-c, --category <category>',\n      `limit to one category (${CATEGORIES.join(', ')})`,\n      parseCategory,\n    )\n    .option('--file <path>', 'validate one public task without pack-balance checks')\n    .action(async (options: { category?: BenchmarkCategory; file?: string }) => {\n      if (options.file) {\n        if (options.category) {\n          throw new Error('tasks validate accepts --file or --category, not both');\n        }\n        const task = await validatePublicTaskFile(options.file);\n        dependencies.stdout(\n          `✓ ${task.public.id}: public task validated (${task.public.category}/${task.public.cluster})`,\n        );\n        return;\n      }\n      const categories = options.category ? [options.category] : CATEGORIES;\n      for (const category of categories) {\n        const loader = dependencies.createTaskLoader(category);\n        const loaded = await loader.loadAll();\n        const mode = loader.hasPrivate ? 'public + private halves' : 'public halves only';\n        dependencies.stdout(`✓ ${category}: ${loaded.length} tasks validated (${mode})`);\n        for (const task of loaded) {\n          dependencies.stdout(`  ${task.public.cluster.padEnd(28)} ${task.public.id}`);\n        }\n      }\n    });\n\n  tasks\n    .command('publish')\n    .description('Publish public and private task halves to the configured API')\n    .option(\n      '-c, --category <category>',\n      `category to publish (${CATEGORIES.join(', ')})`,\n      parseCategory,\n    )\n    .option('--all', 'publish every category explicitly', false)\n    .action(async (options: CategorySelection) => {\n      const categories = selectedCategories(options, 'tasks publish');\n      const config = dependencies.resolveApiConfig();\n      dependencies.stdout(`Publishing ${categories.join(', ')} tasks to ${publishTarget(config)}`);\n      for (const category of categories) {\n        const result = await dependencies.publishTasks(category, config);\n        dependencies.stdout(`✓ ${category}: published ${result.imported} tasks`);\n      }\n    });\n\n  const arena = program.command('arena').description('Run and inspect arena matches');\n  arena\n    .command('run')\n    .description('Run a paid deterministic match schedule')\n    .requiredOption(\n      '-c, --category <category>',\n      `arena to run (${CATEGORIES.join(', ')})`,\n      parseCategory,\n    )\n    .option('-n, --matches <count>', 'number of task-level matches', positiveInteger, 12)\n    .option('-s, --seed <seed>', 'reproducible scheduling seed', 'bridgebench-v3-mvp')\n    .option(\n      '--max-cost-usd <amount>',\n      'hard stop before the next match after this spend',\n      positiveNumber,\n      25,\n    )\n    .option(\n      '--competitor <modelId>',\n      'limit the run roster; repeat for each competitor (minimum two)',\n      collectOption,\n    )\n    .option('--resume', 'skip match IDs already present in the journal', false)\n    .option('--debug', 'mirror structured log entries to the console', false)\n    .option('--no-health-stop', 'do not halt when most matches have failed responses')\n    .action(async (options: ArenaRunCliOptions) => {\n      const config = arenaRunConfigFromOptions(options);\n      const logger = new FileArenaLogger({\n        dir: logRoot(config.category),\n        verbose: options.debug,\n      });\n      const cancellation = new AbortController();\n      const requestCancellation = (): void => {\n        if (cancellation.signal.aborted) return;\n        dependencies.stderr('Cancellation requested; aborting active model calls.');\n        cancellation.abort();\n      };\n      process.once('SIGINT', requestCancellation);\n      try {\n        dependencies.stdout(`Run log: ${displayPath(logger.filePath)}`);\n        const store = dependencies.createStore(config.category);\n        const loaded = await dependencies\n          .createTaskLoader(config.category)\n          .loadAll({ requirePrivate: true });\n        const progressOutput: ArenaEventSink = (event) => {\n          if (event.type !== 'match.completed') return;\n          dependencies.stdout(\n            `[${event.data.completed}/${event.data.total}] ${event.data.taskId}: ${event.data.winnerModelId ?? 'no contest'} (${event.data.outcome}, $${event.data.costUsd.toFixed(4)})`,\n          );\n        };\n        const runner = new ArenaRunner(\n          dependencies.createOpenRouter(logger),\n          store,\n          progressOutput,\n          logger,\n        );\n        const result = await runner.run(config, loaded, { signal: cancellation.signal });\n        if (result.cancelled) {\n          dependencies.stdout(\n            `Cancelled after ${result.completed} new ${config.category} matches; journaled spend $${result.costUsd.toFixed(4)}. Completed matches remain journaled.`,\n          );\n          dependencies.setExitCode(130);\n        } else {\n          dependencies.stdout(\n            `Completed ${result.completed} new ${config.category} matches; run spend $${result.costUsd.toFixed(4)}.`,\n          );\n          if (result.stoppedForBudget) {\n            dependencies.stdout(\n              'Stopped at the cost boundary; rerun the same command with --resume.',\n            );\n          }\n        }\n        const runMatches = store.readAll().filter((match) => match.runId === result.runId);\n        if (runMatches.length > 0) {\n          const triage = triageJournal(runMatches);\n          dependencies.stdout(`=== Run health ===\\n${formatTriage(triage)}`);\n        }\n      } finally {\n        process.removeListener('SIGINT', requestCancellation);\n        dependencies.stdout(`Run log: ${displayPath(logger.filePath)}`);\n      }\n    });\n\n  arena\n    .command('remote-run')\n    .description('Run a match schedule against the configured API and stream live arena events')\n    .requiredOption(\n      '-c, --category <category>',\n      `arena to run (${CATEGORIES.join(', ')})`,\n      parseCategory,\n    )\n    .option('-n, --matches <count>', 'number of task-level matches', positiveInteger, 12)\n    .option('-s, --seed <seed>', 'reproducible scheduling seed', 'bridgebench-v3-mvp')\n    .option(\n      '--max-cost-usd <amount>',\n      'hard stop before the next match after this spend',\n      positiveNumber,\n      25,\n    )\n    .option(\n      '--competitor <modelId>',\n      'limit the run roster; repeat for each competitor (minimum two)',\n      collectOption,\n    )\n    .option('--resume', 'skip match IDs already present in the journal', false)\n    .option('--debug', 'mirror structured log entries to the console', false)\n    .option('--no-health-stop', 'do not halt when most matches have failed responses')\n    .option('--mock', 'use deterministic mock model completions instead of OpenRouter', false)\n    .option('--no-publish-matches', 'skip publishing match results to the API as they complete')\n    .action(async (options: ArenaRunCliOptions & { mock: boolean; publishMatches: boolean }) => {\n      const config = arenaRunConfigFromOptions(options);\n      const apiConfig = dependencies.resolveApiConfig();\n      const logger = new FileArenaLogger({\n        dir: logRoot(config.category),\n        verbose: options.debug,\n      });\n      dependencies.stdout(\n        `Remote ${config.category} run against ${publishTarget(apiConfig)} (${options.mock ? 'mock' : 'openrouter'})`,\n      );\n      dependencies.stdout(`Run log: ${displayPath(logger.filePath)}`);\n      try {\n        const result = await runRemoteArena(apiConfig, {\n          config,\n          mock: options.mock,\n          publishMatches: options.publishMatches,\n          logger,\n        });\n        dependencies.stdout(\n          `Remote run ${result.runKey}: ${result.completed} matches, $${result.costUsd.toFixed(4)}`,\n        );\n        if (result.cancelled) dependencies.setExitCode(130);\n      } finally {\n        dependencies.stdout(`Run log: ${displayPath(logger.filePath)}`);\n      }\n    });\n\n  arena\n    .command('verify')\n    .description('Validate and replay outcomes, points, manifests, and Elo')\n    .option(\n      '-c, --category <category>',\n      `journal category (${CATEGORIES.join(', ')})`,\n      parseCategory,\n      'reasoning',\n    )\n    .option('--journal <path>', 'journal file to verify (defaults to the local category journal)')\n    .option('--manifests-dir <path>', 'run-manifest directory (defaults to a runs/ sibling)')\n    .option('--json', 'emit the machine-readable verification result', false)\n    .action(\n      (options: {\n        category: BenchmarkCategory;\n        journal?: string;\n        manifestsDir?: string;\n        json: boolean;\n      }) => {\n        const defaults = categoryStoreConfig(options.category);\n        const journalPath = options.journal ? path.resolve(options.journal) : defaults.journalPath;\n        const runsDir = options.manifestsDir\n          ? path.resolve(options.manifestsDir)\n          : options.journal\n            ? path.join(path.dirname(journalPath), 'runs')\n            : defaults.runsDir;\n        const store = new ArenaStore({\n          ...defaults,\n          journalPath,\n          runsDir,\n          readOnly: true,\n        });\n        const verified = verifyJournal(store.readAll(), options.category, {\n          manifestForRun: (runId) => store.readRunManifest(runId),\n          requireManifests: true,\n        });\n        if (options.json) {\n          dependencies.stdout(JSON.stringify(verified, null, 2));\n          return;\n        }\n        dependencies.stdout(\n          `✓ ${options.category}: verified ${verified.matches.length} matches across ${verified.runs.length} runs`,\n        );\n        for (const warning of verified.warnings) {\n          dependencies.stderr(`Warning: ${warning}`);\n        }\n      },\n    );\n\n  arena\n    .command('publish')\n    .description('Publish a verified match journal to the configured API')\n    .option(\n      '-c, --category <category>',\n      `journal to publish (${CATEGORIES.join(', ')})`,\n      parseCategory,\n    )\n    .option('--all', 'publish every category explicitly', false)\n    .option('--allow-empty', 'treat an empty journal as a successful no-op', false)\n    .action(async (options: CategorySelection & { allowEmpty: boolean }) => {\n      const categories = selectedCategories(options, 'arena publish');\n      const config = dependencies.resolveApiConfig();\n      dependencies.stdout(\n        `Publishing ${categories.join(', ')} journals to ${publishTarget(config)}`,\n      );\n      for (const category of categories) {\n        const store = dependencies.createStore(category);\n        verifyJournal(store.readAll(), category, {\n          manifestForRun: (runId) => store.readRunManifest(runId),\n          requireManifests: true,\n        });\n        const result = await dependencies.publishJournal(category, config);\n        if (result.matches === 0 && !options.allowEmpty) {\n          throw new Error(\n            `The ${category} journal is empty; pass --allow-empty for an intentional no-op`,\n          );\n        }\n        dependencies.stdout(\n          `✓ ${category}: ${result.imported} new, ${result.skipped} already present (of ${result.matches} lines)`,\n        );\n      }\n    });\n\n  arena\n    .command('triage')\n    .description('Analyze failures, suspicious responses, and judging anomalies')\n    .option(\n      '-c, --category <category>',\n      `journal to analyze (${CATEGORIES.join(', ')})`,\n      parseCategory,\n      'reasoning',\n    )\n    .option('--run <runId>', 'limit the report to one run ID')\n    .option('--json', 'emit the machine-readable report', false)\n    .option('--strict', 'exit nonzero when any anomaly is found', false)\n    .option('--allow-empty', 'treat no matching journal lines as success', false)\n    .action(\n      (options: {\n        category: BenchmarkCategory;\n        run?: string;\n        json: boolean;\n        strict: boolean;\n        allowEmpty: boolean;\n      }) => {\n        const results = dependencies\n          .createStore(options.category)\n          .readAll()\n          .filter((match) => !options.run || match.runId === options.run);\n        if (results.length === 0) {\n          const message = options.run\n            ? `No ${options.category} matches found for run ${options.run}.`\n            : `The ${options.category} journal is empty.`;\n          if (!options.allowEmpty) throw new Error(message);\n          dependencies.stdout(message);\n          return;\n        }\n        const reports = triageJournal(results);\n        dependencies.stdout(\n          options.json ? JSON.stringify(reports, null, 2) : formatTriage(reports),\n        );\n        const anomalies = reports.reduce((sum, report) => sum + report.anomalies.length, 0);\n        if (options.strict && anomalies > 0) {\n          dependencies.stderr(`Strict mode: ${anomalies} anomalies found.`);\n          dependencies.setExitCode(1);\n        }\n      },\n    );\n\n  arena\n    .command('generation <generationId>')\n    .description(\"Fetch OpenRouter's record for a journaled generation\")\n    .action(async (generationId: string) => {\n      const data = await dependencies.createOpenRouter().fetchGeneration(generationId);\n      dependencies.stdout(JSON.stringify(data, null, 2));\n    });\n\n  program\n    .command('report')\n    .description('Verify journals and rebuild JSON and Markdown reports')\n    .option(\n      '-c, --category <category>',\n      `limit to one category (${CATEGORIES.join(', ')})`,\n      parseCategory,\n    )\n    .action((options: { category?: BenchmarkCategory }) => {\n      const categories = options.category ? [options.category] : CATEGORIES;\n      for (const category of categories) {\n        const snapshot = writeReports(dependencies.createStore(category));\n        dependencies.stdout(`✓ ${category}: wrote reports for ${snapshot.matches.length} matches`);\n      }\n    });\n\n  return program;\n}\n\nexport async function runCli(\n  argv: string[] = process.argv,\n  overrides: Partial<CliDependencies> = {},\n): Promise<void> {\n  const dependencies = { ...defaultDependencies, ...overrides };\n  const envResult = loadProjectEnv();\n  if (envResult.status === 'error') {\n    dependencies.stderr(\n      `Warning: project environment file could not be loaded: ${redactSecrets(envResult.reason)}`,\n    );\n  }\n  await buildProgram(dependencies).parseAsync(argv);\n}\n","import { existsSync } from 'node:fs';\nimport path from 'node:path';\n\nimport { findProjectRoot } from './paths.js';\n\nexport const ENV_PATH = path.join(findProjectRoot(import.meta.url), '.env');\n\nexport type ProjectEnvLoadResult =\n  { status: 'missing' } | { status: 'loaded' } | { status: 'error'; reason: string };\n\n/**\n * Loads the project-root .env into process.env regardless of the process cwd\n * or launch flags. Variables already present in the environment win, matching\n * node's --env-file precedence. Safe to call repeatedly: a key added to .env\n * after startup is picked up on the next call.\n */\nexport function loadProjectEnv(): ProjectEnvLoadResult {\n  if (!existsSync(ENV_PATH)) return { status: 'missing' };\n  try {\n    process.loadEnvFile(ENV_PATH);\n    return { status: 'loaded' };\n  } catch (error) {\n    return {\n      status: 'error',\n      reason: error instanceof Error ? error.message : String(error),\n    };\n  }\n}\n","#!/usr/bin/env node\n\nimport { runCli } from './commands.js';\nimport { sanitizeError } from './openrouter.js';\n\nrunCli().catch((error) => {\n  console.error(`BridgeBench: ${sanitizeError(error)}`);\n  process.exitCode = 1;\n});\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA,OAAOA,WAAU;AAEjB,SAAS,SAAS,4BAA4B;;;ACF9C,SAAS,kBAAkB;AAC3B,OAAO,UAAU;AAIV,IAAM,WAAW,KAAK,KAAK,gBAAgB,YAAY,GAAG,GAAG,MAAM;AAWnE,SAAS,iBAAuC;AACrD,MAAI,CAAC,WAAW,QAAQ,EAAG,QAAO,EAAE,QAAQ,UAAU;AACtD,MAAI;AACF,YAAQ,YAAY,QAAQ;AAC5B,WAAO,EAAE,QAAQ,SAAS;AAAA,EAC5B,SAAS,OAAO;AACd,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,QAAQ,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;AAAA,IAC/D;AAAA,EACF;AACF;;;ADAA,IAAM,OAAO,QAAQ,IAAI;AA8BzB,IAAM,sBAAuC;AAAA,EAC3C,QAAQ,CAAC,YAAY,QAAQ,IAAI,OAAO;AAAA,EACxC,QAAQ,CAAC,YAAY,QAAQ,MAAM,OAAO;AAAA,EAC1C,aAAa,CAAC,SAAS;AACrB,YAAQ,WAAW;AAAA,EACrB;AAAA,EACA,aAAa,CAAC,aAAa,IAAI,WAAW,oBAAoB,QAAQ,CAAC;AAAA,EACvE,kBAAkB,CAAC,aAAa,IAAI,WAAW,QAAQ;AAAA,EACvD,kBAAkB,CAAC,WAAW,IAAI,iBAAiB,QAAQ,IAAI,sBAAsB,IAAI,MAAM;AAAA,EAC/F;AAAA,EACA;AAAA,EACA;AACF;AAEA,SAAS,cAAc,OAAkC;AACvD,QAAM,SAAS,wBAAwB,UAAU,KAAK;AACtD,MAAI,CAAC,OAAO,SAAS;AACnB,UAAM,IAAI,qBAAqB,mBAAmB,WAAW,KAAK,IAAI,CAAC,cAAc,KAAK,EAAE;AAAA,EAC9F;AACA,SAAO,OAAO;AAChB;AAEA,SAAS,gBAAgB,OAAuB;AAC9C,QAAM,SAAS,OAAO,KAAK;AAC3B,MAAI,CAAC,OAAO,UAAU,MAAM,KAAK,SAAS,GAAG;AAC3C,UAAM,IAAI,qBAAqB,yCAAyC,KAAK,EAAE;AAAA,EACjF;AACA,SAAO;AACT;AAEA,SAAS,eAAe,OAAuB;AAC7C,QAAM,SAAS,OAAO,KAAK;AAC3B,MAAI,CAAC,OAAO,SAAS,MAAM,KAAK,UAAU,GAAG;AAC3C,UAAM,IAAI,qBAAqB,wCAAwC,KAAK,EAAE;AAAA,EAChF;AACA,SAAO;AACT;AAEA,SAAS,cAAc,OAAe,WAAqB,CAAC,GAAa;AACvE,SAAO,CAAC,GAAG,UAAU,KAAK;AAC5B;AAEO,SAAS,0BAA0B,SAA6C;AACrF,QAAM,gBAAgB,QAAQ,YAAY,SACtC,wBAAwB,QAAQ,UAAU,EAAE,IAAI,CAAC,UAAU,MAAM,EAAE,IACnE;AACJ,SAAO;AAAA,IACL,UAAU,QAAQ;AAAA,IAClB,MAAM,QAAQ;AAAA,IACd,SAAS,QAAQ;AAAA,IACjB,YAAY,QAAQ;AAAA,IACpB;AAAA,IACA,QAAQ,QAAQ;AAAA,IAChB,YAAY,QAAQ;AAAA,EACtB;AACF;AAEA,SAAS,mBAAmB,SAA4B,SAAsC;AAC5F,MAAI,QAAQ,YAAY,QAAQ,KAAK;AACnC,UAAM,IAAI,MAAM,GAAG,OAAO,+CAA+C;AAAA,EAC3E;AACA,MAAI,QAAQ,SAAU,QAAO,CAAC,QAAQ,QAAQ;AAC9C,MAAI,QAAQ,IAAK,QAAO,CAAC,GAAG,UAAU;AACtC,QAAM,IAAI,MAAM,GAAG,OAAO,0CAA0C;AACtE;AAEA,SAAS,QAAQ,UAAqC;AACpD,SAAOC,MAAK,KAAK,MAAM,WAAW,UAAU,MAAM;AACpD;AAEA,SAAS,YAAY,QAAwB;AAC3C,QAAM,WAAWA,MAAK,SAAS,MAAM,MAAM;AAC3C,SAAO,SAAS,WAAW,IAAI,IAAIA,MAAK,SAAS,MAAM,IAAI;AAC7D;AAEO,SAAS,aAAa,YAAsC,CAAC,GAAY;AAC9E,QAAM,eAAgC,EAAE,GAAG,qBAAqB,GAAG,UAAU;AAC7E,QAAM,UAAU,IAAI,QAAQ,EACzB,KAAK,YAAY,EACjB,YAAY,+DAA+D,EAC3E,QAAQ,cAAc,EACtB,mBAAmB,EACnB,cAAc,EAAE,aAAa,MAAM,iBAAiB,KAAK,CAAC,EAC1D,gBAAgB;AAAA,IACf,UAAU,CAAC,UAAU,aAAa,OAAO,MAAM,QAAQ,OAAO,EAAE,CAAC;AAAA,IACjE,UAAU,CAAC,UAAU,aAAa,OAAO,MAAM,QAAQ,OAAO,EAAE,CAAC;AAAA,EACnE,CAAC;AAEH,QAAM,SAAS,QACZ,QAAQ,QAAQ,EAChB,YAAY,+CAA+C;AAC9D,SACG,QAAQ,MAAM,EACd,YAAY,qCAAqC,EACjD,OAAO,MAAM;AACZ,eAAW,SAAS,WAAW,GAAG;AAChC,mBAAa,OAAO,GAAG,MAAM,KAAK,OAAO,EAAE,CAAC,IAAI,MAAM,GAAG,OAAO,EAAE,CAAC,IAAI,MAAM,WAAW,EAAE;AAAA,IAC5F;AAAA,EACF,CAAC;AACH,SACG,QAAQ,UAAU,EAClB,YAAY,qEAAqE,EACjF,OAAO,YAAY;AAClB,UAAM,SAAS,aAAa,iBAAiB;AAC7C,eAAW,SAAS,WAAW,GAAG;AAChC,YAAM,OAAO,cAAc,KAAK;AAChC,mBAAa,OAAO,UAAK,MAAM,EAAE,OAAO,MAAM,aAAa,EAAE;AAAA,IAC/D;AAAA,EACF,CAAC;AAEH,QAAM,QAAQ,QAAQ,QAAQ,OAAO,EAAE,YAAY,gCAAgC;AACnF,QACG,QAAQ,UAAU,EAClB,YAAY,+CAA+C,EAC3D;AAAA,IACC;AAAA,IACA,0BAA0B,WAAW,KAAK,IAAI,CAAC;AAAA,IAC/C;AAAA,EACF,EACC,OAAO,iBAAiB,sDAAsD,EAC9E,OAAO,OAAO,YAA6D;AAC1E,QAAI,QAAQ,MAAM;AAChB,UAAI,QAAQ,UAAU;AACpB,cAAM,IAAI,MAAM,uDAAuD;AAAA,MACzE;AACA,YAAM,OAAO,MAAM,uBAAuB,QAAQ,IAAI;AACtD,mBAAa;AAAA,QACX,UAAK,KAAK,OAAO,EAAE,4BAA4B,KAAK,OAAO,QAAQ,IAAI,KAAK,OAAO,OAAO;AAAA,MAC5F;AACA;AAAA,IACF;AACA,UAAM,aAAa,QAAQ,WAAW,CAAC,QAAQ,QAAQ,IAAI;AAC3D,eAAW,YAAY,YAAY;AACjC,YAAM,SAAS,aAAa,iBAAiB,QAAQ;AACrD,YAAM,SAAS,MAAM,OAAO,QAAQ;AACpC,YAAM,OAAO,OAAO,aAAa,4BAA4B;AAC7D,mBAAa,OAAO,UAAK,QAAQ,KAAK,OAAO,MAAM,qBAAqB,IAAI,GAAG;AAC/E,iBAAW,QAAQ,QAAQ;AACzB,qBAAa,OAAO,KAAK,KAAK,OAAO,QAAQ,OAAO,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,EAAE;AAAA,MAC7E;AAAA,IACF;AAAA,EACF,CAAC;AAEH,QACG,QAAQ,SAAS,EACjB,YAAY,8DAA8D,EAC1E;AAAA,IACC;AAAA,IACA,wBAAwB,WAAW,KAAK,IAAI,CAAC;AAAA,IAC7C;AAAA,EACF,EACC,OAAO,SAAS,qCAAqC,KAAK,EAC1D,OAAO,OAAO,YAA+B;AAC5C,UAAM,aAAa,mBAAmB,SAAS,eAAe;AAC9D,UAAM,SAAS,aAAa,iBAAiB;AAC7C,iBAAa,OAAO,cAAc,WAAW,KAAK,IAAI,CAAC,aAAa,cAAc,MAAM,CAAC,EAAE;AAC3F,eAAW,YAAY,YAAY;AACjC,YAAM,SAAS,MAAM,aAAa,aAAa,UAAU,MAAM;AAC/D,mBAAa,OAAO,UAAK,QAAQ,eAAe,OAAO,QAAQ,QAAQ;AAAA,IACzE;AAAA,EACF,CAAC;AAEH,QAAM,QAAQ,QAAQ,QAAQ,OAAO,EAAE,YAAY,+BAA+B;AAClF,QACG,QAAQ,KAAK,EACb,YAAY,yCAAyC,EACrD;AAAA,IACC;AAAA,IACA,iBAAiB,WAAW,KAAK,IAAI,CAAC;AAAA,IACtC;AAAA,EACF,EACC,OAAO,yBAAyB,gCAAgC,iBAAiB,EAAE,EACnF,OAAO,qBAAqB,gCAAgC,oBAAoB,EAChF;AAAA,IACC;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EACC;AAAA,IACC;AAAA,IACA;AAAA,IACA;AAAA,EACF,EACC,OAAO,YAAY,iDAAiD,KAAK,EACzE,OAAO,WAAW,gDAAgD,KAAK,EACvE,OAAO,oBAAoB,qDAAqD,EAChF,OAAO,OAAO,YAAgC;AAC7C,UAAM,SAAS,0BAA0B,OAAO;AAChD,UAAM,SAAS,IAAI,gBAAgB;AAAA,MACjC,KAAK,QAAQ,OAAO,QAAQ;AAAA,MAC5B,SAAS,QAAQ;AAAA,IACnB,CAAC;AACD,UAAM,eAAe,IAAI,gBAAgB;AACzC,UAAM,sBAAsB,MAAY;AACtC,UAAI,aAAa,OAAO,QAAS;AACjC,mBAAa,OAAO,sDAAsD;AAC1E,mBAAa,MAAM;AAAA,IACrB;AACA,YAAQ,KAAK,UAAU,mBAAmB;AAC1C,QAAI;AACF,mBAAa,OAAO,YAAY,YAAY,OAAO,QAAQ,CAAC,EAAE;AAC9D,YAAM,QAAQ,aAAa,YAAY,OAAO,QAAQ;AACtD,YAAM,SAAS,MAAM,aAClB,iBAAiB,OAAO,QAAQ,EAChC,QAAQ,EAAE,gBAAgB,KAAK,CAAC;AACnC,YAAM,iBAAiC,CAAC,UAAU;AAChD,YAAI,MAAM,SAAS,kBAAmB;AACtC,qBAAa;AAAA,UACX,IAAI,MAAM,KAAK,SAAS,IAAI,MAAM,KAAK,KAAK,KAAK,MAAM,KAAK,MAAM,KAAK,MAAM,KAAK,iBAAiB,YAAY,KAAK,MAAM,KAAK,OAAO,MAAM,MAAM,KAAK,QAAQ,QAAQ,CAAC,CAAC;AAAA,QAC3K;AAAA,MACF;AACA,YAAM,SAAS,IAAI;AAAA,QACjB,aAAa,iBAAiB,MAAM;AAAA,QACpC;AAAA,QACA;AAAA,QACA;AAAA,MACF;AACA,YAAM,SAAS,MAAM,OAAO,IAAI,QAAQ,QAAQ,EAAE,QAAQ,aAAa,OAAO,CAAC;AAC/E,UAAI,OAAO,WAAW;AACpB,qBAAa;AAAA,UACX,mBAAmB,OAAO,SAAS,QAAQ,OAAO,QAAQ,8BAA8B,OAAO,QAAQ,QAAQ,CAAC,CAAC;AAAA,QACnH;AACA,qBAAa,YAAY,GAAG;AAAA,MAC9B,OAAO;AACL,qBAAa;AAAA,UACX,aAAa,OAAO,SAAS,QAAQ,OAAO,QAAQ,wBAAwB,OAAO,QAAQ,QAAQ,CAAC,CAAC;AAAA,QACvG;AACA,YAAI,OAAO,kBAAkB;AAC3B,uBAAa;AAAA,YACX;AAAA,UACF;AAAA,QACF;AAAA,MACF;AACA,YAAM,aAAa,MAAM,QAAQ,EAAE,OAAO,CAAC,UAAU,MAAM,UAAU,OAAO,KAAK;AACjF,UAAI,WAAW,SAAS,GAAG;AACzB,cAAM,SAAS,cAAc,UAAU;AACvC,qBAAa,OAAO;AAAA,EAAuB,aAAa,MAAM,CAAC,EAAE;AAAA,MACnE;AAAA,IACF,UAAE;AACA,cAAQ,eAAe,UAAU,mBAAmB;AACpD,mBAAa,OAAO,YAAY,YAAY,OAAO,QAAQ,CAAC,EAAE;AAAA,IAChE;AAAA,EACF,CAAC;AAEH,QACG,QAAQ,YAAY,EACpB,YAAY,8EAA8E,EAC1F;AAAA,IACC;AAAA,IACA,iBAAiB,WAAW,KAAK,IAAI,CAAC;AAAA,IACtC;AAAA,EACF,EACC,OAAO,yBAAyB,gCAAgC,iBAAiB,EAAE,EACnF,OAAO,qBAAqB,gCAAgC,oBAAoB,EAChF;AAAA,IACC;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EACC;AAAA,IACC;AAAA,IACA;AAAA,IACA;AAAA,EACF,EACC,OAAO,YAAY,iDAAiD,KAAK,EACzE,OAAO,WAAW,gDAAgD,KAAK,EACvE,OAAO,oBAAoB,qDAAqD,EAChF,OAAO,UAAU,kEAAkE,KAAK,EACxF,OAAO,wBAAwB,2DAA2D,EAC1F,OAAO,OAAO,YAA6E;AAC1F,UAAM,SAAS,0BAA0B,OAAO;AAChD,UAAM,YAAY,aAAa,iBAAiB;AAChD,UAAM,SAAS,IAAI,gBAAgB;AAAA,MACjC,KAAK,QAAQ,OAAO,QAAQ;AAAA,MAC5B,SAAS,QAAQ;AAAA,IACnB,CAAC;AACD,iBAAa;AAAA,MACX,UAAU,OAAO,QAAQ,gBAAgB,cAAc,SAAS,CAAC,KAAK,QAAQ,OAAO,SAAS,YAAY;AAAA,IAC5G;AACA,iBAAa,OAAO,YAAY,YAAY,OAAO,QAAQ,CAAC,EAAE;AAC9D,QAAI;AACF,YAAM,SAAS,MAAM,eAAe,WAAW;AAAA,QAC7C;AAAA,QACA,MAAM,QAAQ;AAAA,QACd,gBAAgB,QAAQ;AAAA,QACxB;AAAA,MACF,CAAC;AACD,mBAAa;AAAA,QACX,cAAc,OAAO,MAAM,KAAK,OAAO,SAAS,cAAc,OAAO,QAAQ,QAAQ,CAAC,CAAC;AAAA,MACzF;AACA,UAAI,OAAO,UAAW,cAAa,YAAY,GAAG;AAAA,IACpD,UAAE;AACA,mBAAa,OAAO,YAAY,YAAY,OAAO,QAAQ,CAAC,EAAE;AAAA,IAChE;AAAA,EACF,CAAC;AAEH,QACG,QAAQ,QAAQ,EAChB,YAAY,0DAA0D,EACtE;AAAA,IACC;AAAA,IACA,qBAAqB,WAAW,KAAK,IAAI,CAAC;AAAA,IAC1C;AAAA,IACA;AAAA,EACF,EACC,OAAO,oBAAoB,iEAAiE,EAC5F,OAAO,0BAA0B,sDAAsD,EACvF,OAAO,UAAU,iDAAiD,KAAK,EACvE;AAAA,IACC,CAAC,YAKK;AACJ,YAAM,WAAW,oBAAoB,QAAQ,QAAQ;AACrD,YAAM,cAAc,QAAQ,UAAUA,MAAK,QAAQ,QAAQ,OAAO,IAAI,SAAS;AAC/E,YAAM,UAAU,QAAQ,eACpBA,MAAK,QAAQ,QAAQ,YAAY,IACjC,QAAQ,UACNA,MAAK,KAAKA,MAAK,QAAQ,WAAW,GAAG,MAAM,IAC3C,SAAS;AACf,YAAM,QAAQ,IAAI,WAAW;AAAA,QAC3B,GAAG;AAAA,QACH;AAAA,QACA;AAAA,QACA,UAAU;AAAA,MACZ,CAAC;AACD,YAAM,WAAW,cAAc,MAAM,QAAQ,GAAG,QAAQ,UAAU;AAAA,QAChE,gBAAgB,CAAC,UAAU,MAAM,gBAAgB,KAAK;AAAA,QACtD,kBAAkB;AAAA,MACpB,CAAC;AACD,UAAI,QAAQ,MAAM;AAChB,qBAAa,OAAO,KAAK,UAAU,UAAU,MAAM,CAAC,CAAC;AACrD;AAAA,MACF;AACA,mBAAa;AAAA,QACX,UAAK,QAAQ,QAAQ,cAAc,SAAS,QAAQ,MAAM,mBAAmB,SAAS,KAAK,MAAM;AAAA,MACnG;AACA,iBAAW,WAAW,SAAS,UAAU;AACvC,qBAAa,OAAO,YAAY,OAAO,EAAE;AAAA,MAC3C;AAAA,IACF;AAAA,EACF;AAEF,QACG,QAAQ,SAAS,EACjB,YAAY,wDAAwD,EACpE;AAAA,IACC;AAAA,IACA,uBAAuB,WAAW,KAAK,IAAI,CAAC;AAAA,IAC5C;AAAA,EACF,EACC,OAAO,SAAS,qCAAqC,KAAK,EAC1D,OAAO,iBAAiB,gDAAgD,KAAK,EAC7E,OAAO,OAAO,YAAyD;AACtE,UAAM,aAAa,mBAAmB,SAAS,eAAe;AAC9D,UAAM,SAAS,aAAa,iBAAiB;AAC7C,iBAAa;AAAA,MACX,cAAc,WAAW,KAAK,IAAI,CAAC,gBAAgB,cAAc,MAAM,CAAC;AAAA,IAC1E;AACA,eAAW,YAAY,YAAY;AACjC,YAAM,QAAQ,aAAa,YAAY,QAAQ;AAC/C,oBAAc,MAAM,QAAQ,GAAG,UAAU;AAAA,QACvC,gBAAgB,CAAC,UAAU,MAAM,gBAAgB,KAAK;AAAA,QACtD,kBAAkB;AAAA,MACpB,CAAC;AACD,YAAM,SAAS,MAAM,aAAa,eAAe,UAAU,MAAM;AACjE,UAAI,OAAO,YAAY,KAAK,CAAC,QAAQ,YAAY;AAC/C,cAAM,IAAI;AAAA,UACR,OAAO,QAAQ;AAAA,QACjB;AAAA,MACF;AACA,mBAAa;AAAA,QACX,UAAK,QAAQ,KAAK,OAAO,QAAQ,SAAS,OAAO,OAAO,wBAAwB,OAAO,OAAO;AAAA,MAChG;AAAA,IACF;AAAA,EACF,CAAC;AAEH,QACG,QAAQ,QAAQ,EAChB,YAAY,+DAA+D,EAC3E;AAAA,IACC;AAAA,IACA,uBAAuB,WAAW,KAAK,IAAI,CAAC;AAAA,IAC5C;AAAA,IACA;AAAA,EACF,EACC,OAAO,iBAAiB,gCAAgC,EACxD,OAAO,UAAU,oCAAoC,KAAK,EAC1D,OAAO,YAAY,0CAA0C,KAAK,EAClE,OAAO,iBAAiB,8CAA8C,KAAK,EAC3E;AAAA,IACC,CAAC,YAMK;AACJ,YAAM,UAAU,aACb,YAAY,QAAQ,QAAQ,EAC5B,QAAQ,EACR,OAAO,CAAC,UAAU,CAAC,QAAQ,OAAO,MAAM,UAAU,QAAQ,GAAG;AAChE,UAAI,QAAQ,WAAW,GAAG;AACxB,cAAM,UAAU,QAAQ,MACpB,MAAM,QAAQ,QAAQ,0BAA0B,QAAQ,GAAG,MAC3D,OAAO,QAAQ,QAAQ;AAC3B,YAAI,CAAC,QAAQ,WAAY,OAAM,IAAI,MAAM,OAAO;AAChD,qBAAa,OAAO,OAAO;AAC3B;AAAA,MACF;AACA,YAAM,UAAU,cAAc,OAAO;AACrC,mBAAa;AAAA,QACX,QAAQ,OAAO,KAAK,UAAU,SAAS,MAAM,CAAC,IAAI,aAAa,OAAO;AAAA,MACxE;AACA,YAAM,YAAY,QAAQ,OAAO,CAAC,KAAK,WAAW,MAAM,OAAO,UAAU,QAAQ,CAAC;AAClF,UAAI,QAAQ,UAAU,YAAY,GAAG;AACnC,qBAAa,OAAO,gBAAgB,SAAS,mBAAmB;AAChE,qBAAa,YAAY,CAAC;AAAA,MAC5B;AAAA,IACF;AAAA,EACF;AAEF,QACG,QAAQ,2BAA2B,EACnC,YAAY,sDAAsD,EAClE,OAAO,OAAO,iBAAyB;AACtC,UAAM,OAAO,MAAM,aAAa,iBAAiB,EAAE,gBAAgB,YAAY;AAC/E,iBAAa,OAAO,KAAK,UAAU,MAAM,MAAM,CAAC,CAAC;AAAA,EACnD,CAAC;AAEH,UACG,QAAQ,QAAQ,EAChB,YAAY,uDAAuD,EACnE;AAAA,IACC;AAAA,IACA,0BAA0B,WAAW,KAAK,IAAI,CAAC;AAAA,IAC/C;AAAA,EACF,EACC,OAAO,CAAC,YAA8C;AACrD,UAAM,aAAa,QAAQ,WAAW,CAAC,QAAQ,QAAQ,IAAI;AAC3D,eAAW,YAAY,YAAY;AACjC,YAAM,WAAW,aAAa,aAAa,YAAY,QAAQ,CAAC;AAChE,mBAAa,OAAO,UAAK,QAAQ,uBAAuB,SAAS,QAAQ,MAAM,UAAU;AAAA,IAC3F;AAAA,EACF,CAAC;AAEH,SAAO;AACT;AAEA,eAAsB,OACpB,OAAiB,QAAQ,MACzB,YAAsC,CAAC,GACxB;AACf,QAAM,eAAe,EAAE,GAAG,qBAAqB,GAAG,UAAU;AAC5D,QAAM,YAAY,eAAe;AACjC,MAAI,UAAU,WAAW,SAAS;AAChC,iBAAa;AAAA,MACX,0DAA0D,cAAc,UAAU,MAAM,CAAC;AAAA,IAC3F;AAAA,EACF;AACA,QAAM,aAAa,YAAY,EAAE,WAAW,IAAI;AAClD;;;AErgBA,OAAO,EAAE,MAAM,CAAC,UAAU;AACxB,UAAQ,MAAM,gBAAgB,cAAc,KAAK,CAAC,EAAE;AACpD,UAAQ,WAAW;AACrB,CAAC;","names":["path","path"]}

package/dist/client.cjs

@@ -0,0 +1,42 @@
+"use strict";Object.defineProperty(exports, "__esModule", {value: true});
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+var _chunk7YCJSOK7cjs = require('./chunk-7YCJSOK7.cjs');
+require('./chunk-NJTYVNP4.cjs');
+require('./chunk-4TWPCPRP.cjs');
+require('./chunk-CIXITJW6.cjs');
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+exports.ADMIN_KEY_HEADER = _chunk7YCJSOK7cjs.ADMIN_KEY_HEADER; exports.DEFAULT_API_TIMEOUT_MS = _chunk7YCJSOK7cjs.DEFAULT_API_TIMEOUT_MS; exports.REQUEST_SPACING_MS = _chunk7YCJSOK7cjs.REQUEST_SPACING_MS; exports.RemoteArenaEventSink = _chunk7YCJSOK7cjs.RemoteArenaEventSink; exports.chunk = _chunk7YCJSOK7cjs.chunk; exports.createRemoteRun = _chunk7YCJSOK7cjs.createRemoteRun; exports.delay = _chunk7YCJSOK7cjs.delay; exports.fetchExecutionPack = _chunk7YCJSOK7cjs.fetchExecutionPack; exports.getJson = _chunk7YCJSOK7cjs.getJson; exports.postChunks = _chunk7YCJSOK7cjs.postChunks; exports.postJson = _chunk7YCJSOK7cjs.postJson; exports.publishJournal = _chunk7YCJSOK7cjs.publishJournal; exports.publishJournalFromStore = _chunk7YCJSOK7cjs.publishJournalFromStore; exports.publishTarget = _chunk7YCJSOK7cjs.publishTarget; exports.publishTasks = _chunk7YCJSOK7cjs.publishTasks; exports.resolveApiConfig = _chunk7YCJSOK7cjs.resolveApiConfig; exports.runRemoteArena = _chunk7YCJSOK7cjs.runRemoteArena;
+//# sourceMappingURL=client.cjs.map

package/dist/client.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["/Users/matthewmiller/Desktop/bridgemind/bridgebench/dist/client.cjs"],"names":[],"mappings":"AAAA;AACE;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACF,wDAA6B;AAC7B,gCAA6B;AAC7B,gCAA6B;AAC7B,gCAA6B;AAC7B;AACE;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACF,s+BAAC","file":"/Users/matthewmiller/Desktop/bridgemind/bridgebench/dist/client.cjs"}

package/dist/client.d.cts

@@ -0,0 +1,93 @@
+import { z } from 'zod';
+import { A as ArenaStore, a as ArenaLogger } from './logger-CCR9Mg1c.cjs';
+import { B as BenchmarkCategory, d as CompleteArenaTask } from './tasks-CpaCJ6JE.cjs';
+import { f as ArenaRunConfig, A as ArenaEvent } from './reports-4CejmOHf.cjs';
+
+declare const ADMIN_KEY_HEADER = "x-bridgebench-admin-key";
+declare const DEFAULT_API_TIMEOUT_MS = 30000;
+declare const REQUEST_SPACING_MS = 700;
+interface ApiConfig {
+    baseUrl: string;
+    adminKey: string;
+    timeoutMs: number;
+}
+declare function resolveApiConfig(environment?: NodeJS.ProcessEnv): ApiConfig;
+declare function publishTarget(config: ApiConfig): string;
+declare function delay(ms: number): Promise<void>;
+declare function chunk<T>(items: readonly T[], size: number): T[][];
+declare function getJson<T>(config: ApiConfig, path: string, schema: z.ZodType<T>): Promise<T>;
+declare function postJson<T>(config: ApiConfig, path: string, body: unknown, schema: z.ZodType<T>): Promise<T>;
+declare function postChunks<T, R extends {
+    imported?: number;
+    skipped?: number;
+}>(config: ApiConfig, path: string, batches: readonly T[][], wrap: (batch: readonly T[]) => Record<string, readonly T[]>, schema: z.ZodType<R>): Promise<R[]>;
+
+/**
+ * Publish glue: pushes the engine's authored task pack and its match journal
+ * to the bridgebench.ai API. The local verified journal is the execution
+ * authority; the API is a one-way published replica and public read surface.
+ *
+ * Both API endpoints are admin-key guarded and idempotent (upsert on task
+ * key+version, insert-or-skip on match key), so re-publishing is always safe.
+ * Publishing tasks requires the private overlay (the API stores both halves;
+ * hidden references live in a table no public endpoint reads) — see
+ * docs/private-packs.md.
+ */
+/**
+ * Load, validate, and push a category's task pack. The API import contract
+ * requires both halves, so this fails closed (with a pointer to
+ * docs/private-packs.md) when the private overlay is absent.
+ */
+declare function publishTasks(category: BenchmarkCategory, config?: ApiConfig): Promise<{
+    imported: number;
+}>;
+/** Push a match journal from the given store (idempotent on the engine match key). */
+declare function publishJournalFromStore(store: ArenaStore, config?: ApiConfig): Promise<{
+    imported: number;
+    skipped: number;
+    matches: number;
+}>;
+/** Push a category's local match journal (idempotent on the engine match key). */
+declare function publishJournal(category: BenchmarkCategory, config?: ApiConfig): Promise<{
+    imported: number;
+    skipped: number;
+    matches: number;
+}>;
+
+interface RemoteArenaRunOptions {
+    config: ArenaRunConfig;
+    mock?: boolean;
+    publishMatches?: boolean;
+    logger?: ArenaLogger;
+}
+interface RemoteArenaRunResult {
+    runKey: string;
+    completed: number;
+    costUsd: number;
+    cancelled: boolean;
+    stoppedForBudget: boolean;
+}
+declare function fetchExecutionPack(config: ApiConfig, category: BenchmarkCategory): Promise<{
+    tasks: CompleteArenaTask[];
+}>;
+declare function createRemoteRun(apiConfig: ApiConfig, runConfig: ArenaRunConfig, tasks: readonly CompleteArenaTask[]): Promise<{
+    runKey: string;
+    created: boolean;
+}>;
+declare function runRemoteArena(apiConfig: ApiConfig, options: RemoteArenaRunOptions): Promise<RemoteArenaRunResult>;
+
+declare class RemoteArenaEventSink {
+    private readonly config;
+    private readonly runKey;
+    private readonly buffer;
+    private flushTimer;
+    private flushChain;
+    private closed;
+    constructor(config: ApiConfig, runKey: string);
+    readonly sink: (event: ArenaEvent) => void;
+    private scheduleFlush;
+    flush(): Promise<void>;
+    close(): Promise<void>;
+}
+
+export { ADMIN_KEY_HEADER, type ApiConfig, DEFAULT_API_TIMEOUT_MS, REQUEST_SPACING_MS, RemoteArenaEventSink, type RemoteArenaRunOptions, type RemoteArenaRunResult, chunk, createRemoteRun, delay, fetchExecutionPack, getJson, postChunks, postJson, publishJournal, publishJournalFromStore, publishTarget, publishTasks, resolveApiConfig, runRemoteArena };