@tangle-network/agent-eval 0.128.2 → 0.130.0

package/dist/chunk-EJGRPCO3.js

@@ -1,617 +0,0 @@
-import {
-  ROLLOUT_ROLES,
-  ROLLOUT_SCHEMA,
-  assertRolloutLine,
-  isTrainableSplit
-} from "./chunk-UWZZKKU7.js";
-
-// src/rollout/exporters.ts
-function toSftRows(lines, options = {}) {
-  return lines.filter((line) => isTrainingLineEligible(line, options) && line.messages.length > 0).map((line) => ({
-    messages: line.messages,
-    metadata: {
-      rollout_id: line.rollout_id,
-      run_id: line.run_id,
-      candidate_id: line.candidate_id ?? null,
-      instance_id: line.task.instance_id,
-      reward: line.outcome.reward
-    }
-  }));
-}
-function toRewardRows(lines, options = {}) {
-  return lines.filter(
-    (line) => isTrainingLineEligible(line, options) && line.messages.some(
-      (message) => message.role === "user" && typeof message.content === "string" && message.content.length > 0
-    )
-  ).map((line) => ({
-    prompt: line.messages.find((m) => m.role === "user")?.content ?? "",
-    steps: line.steps ?? [],
-    reward: line.outcome.reward,
-    metadata: {
-      rollout_id: line.rollout_id,
-      run_id: line.run_id,
-      candidate_id: line.candidate_id ?? null,
-      instance_id: line.task.instance_id,
-      split: line.task.split
-    }
-  }));
-}
-function firstAssistantIndex(messages) {
-  const index = messages.findIndex((m) => m.role === "assistant");
-  return index === -1 ? messages.length : index;
-}
-function toVerifiersRolloutOutput(line) {
-  const split = firstAssistantIndex(line.messages);
-  return {
-    prompt: line.messages.slice(0, split),
-    completion: line.messages.slice(split),
-    reward: line.outcome.reward,
-    metrics: line.outcome.metrics,
-    tool_defs: line.tool_defs,
-    token_usage: {
-      input_tokens: line.cost.tokens_in,
-      output_tokens: line.cost.tokens_out,
-      reasoning_tokens: line.cost.tokens_reasoning,
-      cache_read_tokens: line.cost.cache_read,
-      cache_write_tokens: line.cost.cache_write
-    },
-    info: {
-      task: line.task,
-      policy: line.policy,
-      rollout_id: line.rollout_id,
-      run_id: line.run_id,
-      experiment_id: line.experiment_id ?? null,
-      candidate_id: line.candidate_id ?? null,
-      generation: line.generation,
-      candidate_index: line.candidate_index,
-      role: line.role
-    }
-  };
-}
-function toVerifiersRolloutOutputs(lines, options = {}) {
-  return lines.filter(
-    (line) => isTrainingLineEligible(line, options) && firstAssistantIndex(line.messages) > 0 && firstAssistantIndex(line.messages) < line.messages.length
-  ).map(toVerifiersRolloutOutput);
-}
-function toRftItem(line) {
-  const split = firstAssistantIndex(line.messages);
-  return {
-    messages: line.messages.slice(0, split),
-    reference: {
-      reward: line.outcome.reward,
-      reward_source: line.outcome.reward_source,
-      verdict: line.outcome.verdict,
-      instance_id: line.task.instance_id,
-      suite: line.task.suite,
-      split: line.task.split,
-      rollout_id: line.rollout_id
-    }
-  };
-}
-function toRftItems(lines, options = {}) {
-  return lines.filter(
-    (line) => isTrainingLineEligible(line, options) && line.messages.length > 0 && firstAssistantIndex(line.messages) > 0
-  ).map(toRftItem);
-}
-function toJsonl(rows) {
-  return rows.map((r) => JSON.stringify(r)).join("\n") + (rows.length ? "\n" : "");
-}
-function isTrainingLineEligible(line, options) {
-  const minimumQualityExclusive = options.minimumQualityExclusive ?? 0;
-  if (!Number.isFinite(minimumQualityExclusive)) {
-    throw new Error("minimumQualityExclusive must be finite");
-  }
-  const reward = line.outcome.reward;
-  if (reward === null) return false;
-  if (!Number.isFinite(reward)) {
-    throw new Error(`training reward for rollout "${line.rollout_id}" must be finite`);
-  }
-  if (reward <= minimumQualityExclusive) return false;
-  if (!line.outcome.is_completed || line.outcome.is_truncated || line.outcome.error !== null) {
-    return false;
-  }
-  if (line.outcome.realness_gated === true) return false;
-  if (line.task.split === "search") return true;
-  return line.task.split === "holdout" && options.allowHeldOutTrainingData === true;
-}
-
-// src/rollout/ledger.ts
-import { appendFile, mkdir, readFile, writeFile } from "fs/promises";
-import { dirname } from "path";
-function serialize(lines) {
-  for (const [i, line] of lines.entries()) assertRolloutLine(line, `rollout line [${i}]`);
-  return lines.map((line) => JSON.stringify(line)).join("\n") + (lines.length > 0 ? "\n" : "");
-}
-async function writeRolloutLedger(path, lines) {
-  const payload = serialize(lines);
-  await mkdir(dirname(path), { recursive: true });
-  await writeFile(path, payload);
-}
-async function appendRolloutLines(path, lines) {
-  if (lines.length === 0) return;
-  const payload = serialize(lines);
-  await mkdir(dirname(path), { recursive: true });
-  await appendFile(path, payload);
-}
-async function readRolloutLedger(path) {
-  const raw = await readFile(path, "utf8");
-  const lines = [];
-  const rawLines = raw.split("\n");
-  for (let i = 0; i < rawLines.length; i++) {
-    const text = rawLines[i];
-    if (!text?.trim()) continue;
-    let parsed;
-    try {
-      parsed = JSON.parse(text);
-    } catch (error) {
-      throw new Error(
-        `${path}:${i + 1}: malformed JSON \u2014 ${error instanceof Error ? error.message : String(error)}`
-      );
-    }
-    assertRolloutLine(parsed, `${path}:${i + 1}`);
-    lines.push(parsed);
-  }
-  return lines;
-}
-
-// src/rollout/release/card.ts
-var RELEASE_FORMATS = ["sft", "verifiers", "rft", "raw"];
-var FORMAT_FILES = {
-  sft: "sft/train.jsonl",
-  verifiers: "verifiers/train.jsonl",
-  rft: "rft/train.jsonl",
-  raw: "raw/train.jsonl"
-};
-var FORMAT_DESCRIPTIONS = {
-  sft: "Successful trainable-split transcripts (`reward >= 1`, never realness-gated) as `{messages, metadata}` chat JSONL.",
-  verifiers: "Prime Intellect verifiers `RolloutOutput`: prompt/completion split at the first assistant turn, plus reward, metrics, tool defs, and token usage.",
-  rft: "OpenAI RFT items: prompt turns plus `reference.*` verdict fields for a grader (completions are re-sampled during RFT).",
-  raw: `Full \`${ROLLOUT_SCHEMA}\` ledger lines (scrubbed), one per agent invocation.`
-};
-function unique(values) {
-  return [...new Set(values.filter((v) => v !== null))].sort();
-}
-function formatReward(reward) {
-  if (reward === null) return "null";
-  return Number.isInteger(reward) ? String(reward) : reward.toFixed(4);
-}
-function markdownTable(header, rows) {
-  return [
-    `| ${header.join(" | ")} |`,
-    `| ${header.map(() => "---").join(" | ")} |`,
-    ...rows.map((row) => `| ${row.join(" | ")} |`)
-  ].join("\n");
-}
-function roleRewardRows(lines) {
-  const counts = /* @__PURE__ */ new Map();
-  for (const line of lines) {
-    const byReward = counts.get(line.role) ?? /* @__PURE__ */ new Map();
-    const key = formatReward(line.outcome.reward);
-    byReward.set(key, (byReward.get(key) ?? 0) + 1);
-    counts.set(line.role, byReward);
-  }
-  const rows = [];
-  for (const role of ROLLOUT_ROLES) {
-    const byReward = counts.get(role);
-    if (!byReward) continue;
-    const keys = [...byReward.keys()].sort((a, b) => {
-      if (a === "null") return 1;
-      if (b === "null") return -1;
-      return Number(b) - Number(a);
-    });
-    for (const key of keys) rows.push([role, key, String(byReward.get(key))]);
-  }
-  return rows;
-}
-function buildDatasetCard(inputs) {
-  const { lines, formats, includeProposers, sourceFiles, scrubTotals, excluded, formatCounts } = inputs;
-  const runIds = unique(lines.map((line) => line.run_id));
-  const generations = [...new Set(lines.map((line) => line.generation))].filter((g) => g !== null).sort((a, b) => a - b);
-  const models = unique(lines.map((line) => line.policy.model));
-  const harnesses = unique(lines.map((line) => line.policy.harness));
-  const captures = unique(lines.map((line) => line.provenance.capture));
-  const rewardSources = unique(lines.map((line) => line.outcome.reward_source));
-  const gapLines = lines.filter((line) => line.messages.length === 0).length;
-  const configs = formats.map(
-    (format) => [
-      `  - config_name: ${format}`,
-      "    data_files:",
-      "      - split: train",
-      `        path: ${FORMAT_FILES[format]}`
-    ].join("\n")
-  ).join("\n");
-  const frontmatter = [
-    "---",
-    "license: unknown",
-    "pretty_name: Tangle rollout ledger \u2014 agent trajectories",
-    "configs:",
-    configs,
-    "---"
-  ].join("\n");
-  const formatsTable = markdownTable(
-    ["config", "path", "rows", "contents"],
-    formats.map((format) => [
-      format,
-      `\`${FORMAT_FILES[format]}\``,
-      String(formatCounts[format] ?? 0),
-      FORMAT_DESCRIPTIONS[format]
-    ])
-  );
-  const countsTable = markdownTable(["role", "reward", "lines"], roleRewardRows(lines));
-  const scrubTable = markdownTable(
-    ["rule", "rewrites"],
-    Object.entries(scrubTotals).map(([rule, count]) => [rule, String(count)])
-  );
-  const proposerNote = includeProposers ? "Proposer sessions are INCLUDED (`--include-proposers`); their transcripts contain improvement-loop harness source." : `Proposer sessions are excluded by default (${excluded.proposers} lines dropped); they contain improvement-loop harness source. Rebuild with \`--include-proposers\` to keep them.`;
-  return `${frontmatter}
-
-# Tangle rollout ledger \u2014 agent trajectories
-
-One line per agent invocation (supervisor episode, worker session, proposer shot, judge call, analyst pass) captured by the \`${ROLLOUT_SCHEMA}\` rollout ledger, labeled with improvement-loop coordinates and the official-judge reward, with the full message transcript inline.
-
-This release contains the **trainable split only** (\`search\`). Holdout, dev, and canary splits are structurally excluded at build time, and the build additionally drops any non-trainable line as a fail-closed filter (${excluded.nonTrain} dropped here).
-
-## Formats
-
-${formatsTable}
-
-## Schema (\`${ROLLOUT_SCHEMA}\`)
-
-Each raw line carries:
-
-- \`rollout_id\` / \`parent_rollout_id\` \u2014 invocation identity; workers point at their spawning supervisor episode.
-- \`run_id\`, \`experiment_id\`, \`candidate_id\` \u2014 run/experiment/candidate identity from the producing RunRecord, when present.
-- \`generation\`, \`candidate_index\` \u2014 improvement-loop coordinates (\`-1\` = baseline campaign; \`null\` = not an improvement loop).
-- \`role\` \u2014 one of ${ROLLOUT_ROLES.map((role) => `\`${role}\``).join(", ")}.
-- \`task\` \u2014 suite, instance id, split, seed, replicate index.
-- \`policy\` \u2014 harness, model, provider, profile commit, prompt/config hashes, sampling params.
-- \`messages\` / \`tool_defs\` \u2014 full transcript in canonical OpenAI chat-with-tools form (including \`reasoning_content\`). An empty \`messages\` array is a labeled gap line; \`provenance.gap\` says why the transcript could not be recovered (${gapLines} gap lines in this release).
-- \`outcome\` \u2014 \`reward\` (the single scalar), \`reward_source\`, the verbatim judge \`verdict\`, non-scalar \`metrics\`, and \`realness_gated\` (anti-Goodhart flag: reward forced to 0, never SFT-eligible).
-- \`cost\` \u2014 usd, token counts, wall time.
-- \`artifacts\` / \`provenance\` \u2014 patch/run-dir/transcript pointers (scrubbed) and capture metadata.
-
-## Provenance
-
-- Source ledgers: ${sourceFiles.map((file) => `\`${file}\``).join(", ")}
-- Run ids: ${runIds.map((id) => `\`${id}\``).join(", ")}
-- Generations: ${generations.join(", ")} (\`-1\` = baseline campaign)
-- Models: ${models.map((m) => `\`${m}\``).join(", ")}
-- Harnesses: ${harnesses.map((h) => `\`${h}\``).join(", ")}
-- Capture modes: ${captures.join(", ")}
-- Every reward traces to a named source (reward sources in this release: ${rewardSources.map((s) => `\`${s}\``).join(", ")}).
-
-${proposerNote}
-
-## Reward semantics per role
-
-- **agent** \u2014 the producing RunRecord's holdout/search score, with the realness gate forcing gamed successes to 0.
-- **supervisor** \u2014 the official-judge verdict on the episode's delivered artifact (1 = resolved, 0 = not).
-- **worker** \u2014 INHERITED from the parent supervisor episode (\`\u2026/inherited\`). Caveat: reward 1 does not establish this worker's individual contribution (sibling workers in the same episode share the episode outcome), and reward 0 does not prove this worker failed.
-- **proposer** \u2014 the fraction of improvement-set instances the proposed candidate resolved (\`\u2026/candidate-resolved-fraction\`); a scalar in [0, 1], not a binary verdict.
-- **judge / analyst** \u2014 carry the episode verdict where one applies; otherwise \`reward: null\` (a labeled gap, never 0).
-
-## Counts
-
-${countsTable}
-
-Total lines: ${lines.length}
-
-## Scrubbing
-
-Absolute home paths were rewritten to \`$WORK\`, credential-shaped strings were replaced with \`[REDACTED:<kind>]\` markers, internal hostnames were normalized to \`*.internal.example\`, and username-bearing incidentals (\`ls -l\` owner columns, per-user pytest tmpdirs) were normalized to \`user\` / \`$USER\`. Rewrite counts for this release (full per-file breakdown in \`scrub-report.json\`):
-
-${scrubTable}
-
-## License
-
-\`license: unknown\` is a placeholder \u2014 the releasing operator must set the real SPDX license id in the frontmatter above before publishing.
-
-## Citation
-
-\`\`\`bibtex
-@misc{tangle_rollout_ledger,
-  title = {Tangle rollout ledger \u2014 agent trajectories},
-  author = {{Tangle Network}},
-  howpublished = {HuggingFace Datasets},
-  note = {Operator: fill in the repository URL, authors, and year before publishing}
-}
-\`\`\`
-`;
-}
-
-// src/rollout/release/scrub.ts
-var SCRUB_RULES = [
-  {
-    // /home/<user> and /Users/<user> prefixes → $WORK ($HOME-derived paths).
-    name: "home-path",
-    pattern: /\/(?:home|Users)\/[A-Za-z0-9._-]+/g,
-    rewrite: () => "$WORK"
-  },
-  {
-    // Harness stores encode cwd as a dashed path segment (-home-drew-…);
-    // strip the leading -home-<user> so the username never ships.
-    name: "home-path-encoded",
-    pattern: /(?<=\/)-(?:home|Users)-[A-Za-z0-9_.]+/g,
-    rewrite: () => "$WORK"
-  },
-  {
-    // pytest names its tmpdir after the invoking user (/tmp/pytest-of-drew).
-    name: "tmp-user-dir",
-    pattern: /\/tmp\/pytest-of-[A-Za-z0-9._-]+/g,
-    rewrite: () => "/tmp/pytest-of-$USER"
-  },
-  {
-    // `ls -l` owner/group columns leak the username in captured tool output.
-    // The `(?!user user…)` guard keeps a second pass at zero rewrites.
-    name: "ls-owner",
-    pattern: /([-bcdlps][-rwxsStT]{9}[.+@]?\s+\d+\s+)(?!user user(?=\s))[A-Za-z0-9._-]+\s+[A-Za-z0-9._-]+(?=\s)/g,
-    rewrite: (_match, prefix) => `${prefix}user user`
-  },
-  {
-    // Env-var-shaped secrets: NAME=value where NAME looks credential-bearing.
-    // The (?!\[REDACTED:) guard keeps the rule idempotent on its own output.
-    name: "env-secret",
-    pattern: /\b([A-Z][A-Z0-9_]*(?:KEY|TOKEN|SECRET|PASSWORD|PASSWD|CREDENTIALS?))=(?!\[REDACTED:)("[^"]*"|'[^']*'|[^\s"']+)/g,
-    rewrite: (_match, name) => `${name}=[REDACTED:env]`
-  },
-  {
-    name: "bearer-token",
-    pattern: /\b(Bearer|Basic)\s+(?!\[REDACTED:)[A-Za-z0-9\-._~+/=]{8,}/g,
-    rewrite: (_match, scheme) => `${scheme} [REDACTED:bearer]`
-  },
-  {
-    // Bare provider-prefixed keys (OpenAI/Anthropic sk-, GitHub ghp_/gho_/…,
-    // fine-grained PATs, HuggingFace hf_, Slack xox*, AWS AKIA).
-    name: "api-key",
-    pattern: /\b(?:sk-[A-Za-z0-9_-]{16,}|gh[pousr]_[A-Za-z0-9]{16,}|github_pat_[A-Za-z0-9_]{20,}|hf_[A-Za-z0-9]{16,}|xox[baprs]-[A-Za-z0-9-]{10,}|AKIA[0-9A-Z]{16})\b/g,
-    rewrite: () => "[REDACTED:api-key]"
-  },
-  {
-    // Our infra hostnames → placeholder domain, subdomain preserved
-    // (router.tangle.tools → router.internal.example).
-    name: "infra-host",
-    pattern: /(?<![A-Za-z0-9.-])((?:[A-Za-z0-9-]+\.)*)tangle\.(?:tools|network)(?![A-Za-z0-9-])/g,
-    rewrite: (_match, prefix) => `${prefix ?? ""}internal.example`
-  },
-  {
-    // Known workstation hostnames leak through email Message-IDs and fqdn
-    // lookups inside captured test output; extend the list as machines join.
-    name: "machine-host",
-    pattern: /\b[A-Za-z0-9]+-GTR-Pro\b/g,
-    rewrite: () => "workstation"
-  }
-];
-function emptyScrubCounts() {
-  return Object.fromEntries(SCRUB_RULES.map((rule) => [rule.name, 0]));
-}
-function addScrubCounts(into, from) {
-  for (const [name, count] of Object.entries(from)) into[name] = (into[name] ?? 0) + count;
-  return into;
-}
-function scrubText(text, counts) {
-  let out = text;
-  for (const rule of SCRUB_RULES) {
-    out = out.replace(rule.pattern, (match, ...rest) => {
-      counts[rule.name] = (counts[rule.name] ?? 0) + 1;
-      return rule.rewrite(match, typeof rest[0] === "string" ? rest[0] : void 0);
-    });
-  }
-  return out;
-}
-function scrubValue(value, counts) {
-  if (typeof value === "string") return scrubText(value, counts);
-  if (Array.isArray(value)) return value.map((item) => scrubValue(item, counts));
-  if (value !== null && typeof value === "object") {
-    const out = {};
-    for (const [key, item] of Object.entries(value)) out[key] = scrubValue(item, counts);
-    return out;
-  }
-  return value;
-}
-function scrubRolloutLine(line, counts) {
-  return scrubValue(line, counts);
-}
-function scrubLines(lines) {
-  const counts = emptyScrubCounts();
-  return { lines: lines.map((line) => scrubRolloutLine(line, counts)), counts };
-}
-function defaultRolloutScrubber(text) {
-  return scrubText(text, emptyScrubCounts());
-}
-
-// src/rollout/release/hf-dataset.ts
-import { spawnSync } from "child_process";
-import { mkdir as mkdir2, writeFile as writeFile2 } from "fs/promises";
-import { basename, dirname as dirname2, join } from "path";
-async function buildHfDataset(inputs, options) {
-  if (inputs.length === 0) throw new Error("no input ledgers given");
-  if (options.formats.length === 0) throw new Error("no formats selected");
-  const report = {
-    files: {},
-    totals: emptyScrubCounts(),
-    excluded: { proposers: 0, nonTrain: 0 }
-  };
-  const kept = [];
-  let read = 0;
-  for (const input of inputs) {
-    const lines = await readRolloutLedger(input);
-    read += lines.length;
-    const shippable = lines.filter((line) => {
-      if (!isTrainableSplit(line.task.split)) {
-        report.excluded.nonTrain += 1;
-        return false;
-      }
-      if (!options.includeProposers && line.role === "proposer") {
-        report.excluded.proposers += 1;
-        return false;
-      }
-      return true;
-    });
-    const scrubbed = scrubLines(shippable);
-    report.files[input] = scrubbed.counts;
-    addScrubCounts(report.totals, scrubbed.counts);
-    kept.push(...scrubbed.lines);
-  }
-  const formatCounts = {};
-  const files = [];
-  for (const format of options.formats) {
-    const path = join(options.out, FORMAT_FILES[format]);
-    await mkdir2(dirname2(path), { recursive: true });
-    if (format === "raw") {
-      await writeRolloutLedger(path, kept);
-      formatCounts.raw = kept.length;
-    } else if (format === "sft") {
-      const rows = toSftRows(kept);
-      await writeFile2(path, toJsonl(rows));
-      formatCounts.sft = rows.length;
-    } else if (format === "verifiers") {
-      const outputs = toVerifiersRolloutOutputs(kept);
-      await writeFile2(path, toJsonl(outputs));
-      formatCounts.verifiers = outputs.length;
-    } else {
-      const items = toRftItems(kept);
-      await writeFile2(path, toJsonl(items));
-      formatCounts.rft = items.length;
-    }
-    files.push(path);
-  }
-  const reportPath = join(options.out, "scrub-report.json");
-  await writeFile2(reportPath, `${JSON.stringify(report, null, 2)}
-`);
-  files.push(reportPath);
-  const cardPath = join(options.out, "README.md");
-  await writeFile2(
-    cardPath,
-    buildDatasetCard({
-      lines: kept,
-      formats: options.formats,
-      includeProposers: options.includeProposers,
-      sourceFiles: inputs.map((input) => basename(input)),
-      scrubTotals: report.totals,
-      excluded: report.excluded,
-      formatCounts
-    })
-  );
-  files.push(cardPath);
-  return { inputs, read, kept: kept.length, scrub: report, formatCounts, files };
-}
-function planPushCommand(repo, outDir) {
-  return ["huggingface-cli", "upload", repo, outDir, ".", "--repo-type", "dataset"];
-}
-function pushDataset(repo, outDir) {
-  const found = spawnSync("which", ["huggingface-cli"], { stdio: "ignore" });
-  if (found.status !== 0) {
-    throw new Error(
-      "huggingface-cli not found on PATH \u2014 install huggingface_hub[cli] before --push"
-    );
-  }
-  if (!process.env.HF_TOKEN) {
-    throw new Error("HF_TOKEN not present in env \u2014 refusing to push");
-  }
-  const [command, ...args] = planPushCommand(repo, outDir);
-  const run = spawnSync(command, args, { stdio: "inherit" });
-  if (run.status !== 0) throw new Error(`huggingface-cli upload exited ${String(run.status)}`);
-}
-var ROLLOUT_RELEASE_USAGE = "usage: agent-eval rollout-release <ledger.jsonl...> --out <dir> [--formats sft,verifiers,rft,raw] [--include-proposers] [--push <org/name>]";
-function parseRolloutReleaseArgs(argv) {
-  const args = {
-    inputs: [],
-    out: "",
-    formats: [...RELEASE_FORMATS],
-    includeProposers: false,
-    push: null
-  };
-  for (let i = 0; i < argv.length; i++) {
-    const arg = argv[i];
-    if (arg === "--out") {
-      args.out = argv[++i] ?? "";
-    } else if (arg === "--formats") {
-      const raw = (argv[++i] ?? "").split(",").filter(Boolean);
-      for (const format of raw) {
-        if (!RELEASE_FORMATS.includes(format)) {
-          throw new Error(
-            `unknown format "${format}" \u2014 expected one of ${RELEASE_FORMATS.join(",")}`
-          );
-        }
-      }
-      args.formats = raw;
-    } else if (arg === "--include-proposers") {
-      args.includeProposers = true;
-    } else if (arg === "--push") {
-      args.push = argv[++i] ?? null;
-    } else if (arg.startsWith("--")) {
-      throw new Error(`unknown flag "${arg}"`);
-    } else {
-      args.inputs.push(arg);
-    }
-  }
-  if (args.inputs.length === 0 || !args.out) {
-    throw new Error(ROLLOUT_RELEASE_USAGE);
-  }
-  if (args.push !== null && !/^[\w.-]+\/[\w.-]+$/.test(args.push)) {
-    throw new Error(`--push expects <org/name>, got "${args.push}"`);
-  }
-  return args;
-}
-async function runRolloutReleaseCli(argv) {
-  let args;
-  try {
-    args = parseRolloutReleaseArgs(argv);
-  } catch (error) {
-    process.stderr.write(`${error instanceof Error ? error.message : String(error)}
-`);
-    return 2;
-  }
-  const summary = await buildHfDataset(args.inputs, args);
-  process.stdout.write(
-    `${JSON.stringify(
-      {
-        read: summary.read,
-        kept: summary.kept,
-        formatCounts: summary.formatCounts,
-        scrub: summary.scrub
-      },
-      null,
-      2
-    )}
-`
-  );
-  process.stdout.write(`dataset \u2192 ${args.out} (${summary.files.length} files)
-`);
-  if (args.push !== null) {
-    pushDataset(args.push, args.out);
-    process.stdout.write(`pushed \u2192 ${args.push}
-`);
-  }
-  return 0;
-}
-
-export {
-  toSftRows,
-  toRewardRows,
-  toVerifiersRolloutOutput,
-  toVerifiersRolloutOutputs,
-  toRftItem,
-  toRftItems,
-  toJsonl,
-  writeRolloutLedger,
-  appendRolloutLines,
-  readRolloutLedger,
-  RELEASE_FORMATS,
-  FORMAT_FILES,
-  buildDatasetCard,
-  SCRUB_RULES,
-  emptyScrubCounts,
-  addScrubCounts,
-  scrubText,
-  scrubRolloutLine,
-  scrubLines,
-  defaultRolloutScrubber,
-  buildHfDataset,
-  planPushCommand,
-  pushDataset,
-  ROLLOUT_RELEASE_USAGE,
-  parseRolloutReleaseArgs,
-  runRolloutReleaseCli
-};
-//# sourceMappingURL=chunk-EJGRPCO3.js.map