@tangle-network/agent-bench 0.4.1 → 0.4.3

package/dist/chunk-BEN6IF2X.js

@@ -1,221 +0,0 @@
-import {
-  benchRoot,
-  preflightVenvImports,
-  readJsonReport,
-  runStagedJudge,
-  stageFile,
-  venvPython
-} from "./chunk-EIETHPD5.js";
-
-// src/benchmarks/aec-bench.ts
-import { readFile } from "fs/promises";
-import { join } from "path";
-var FIXTURES = join(benchRoot, "fixtures", "aec-bench.json");
-var REPO = "TheodoreGalanos/aec-bench";
-var RAW = `https://raw.githubusercontent.com/${REPO}/main`;
-var TREE = `https://api.github.com/repos/${REPO}/git/trees/main?recursive=1`;
-var verifyPathPattern = /^tasks\/(.+)\/tests\/verify\.py$/;
-var DEFAULT_LIMIT = 10;
-function recordToTask(rec) {
-  const meta = {
-    taskId: rec.id,
-    discipline: rec.discipline,
-    taskToml: rec.task_toml,
-    verifyPy: rec.verify_py,
-    goldenPassMd: rec.golden_pass_md
-  };
-  return {
-    id: rec.id,
-    split: rec.discipline,
-    // instruction.md is fully self-contained and already specifies the exact JSON
-    // output schema + the "write to /workspace/output.md" contract the verifier
-    // keys off — we pass it through verbatim so the verify.py extractor matches.
-    prompt: rec.instruction,
-    metadata: meta
-  };
-}
-function readMeta(task) {
-  const md = task.metadata;
-  if (!md || typeof md.verifyPy !== "string" || typeof md.taskId !== "string") {
-    throw new Error(`aec-bench task ${task.id} missing verifier metadata \u2014 loadTasks did not populate it`);
-  }
-  return md;
-}
-async function fetchText(url) {
-  const res = await fetch(url);
-  if (!res.ok) throw new Error(`aec-bench fetch ${res.status}: ${url}`);
-  return res.text();
-}
-async function fetchTextOrNull(url) {
-  const res = await fetch(url);
-  if (res.status === 404) return null;
-  if (!res.ok) throw new Error(`aec-bench fetch ${res.status}: ${url}`);
-  return res.text();
-}
-async function listAllInstances() {
-  const res = await fetch(TREE);
-  if (!res.ok) throw new Error(`aec-bench tree ${res.status}: ${TREE}`);
-  const { tree } = await res.json();
-  const ids = [];
-  for (const entry of tree) {
-    const m = verifyPathPattern.exec(entry.path);
-    if (m?.[1]) ids.push(m[1]);
-  }
-  return ids;
-}
-async function fetchInstance(id) {
-  const base = `${RAW}/tasks/${id}`;
-  const verify = await fetch(`${base}/tests/verify.py`);
-  if (verify.status === 404) return null;
-  if (!verify.ok) throw new Error(`aec-bench fetch ${verify.status}: ${id}/tests/verify.py`);
-  const [instruction, task_toml, golden_pass_md] = await Promise.all([
-    fetchText(`${base}/instruction.md`),
-    fetchText(`${base}/task.toml`),
-    fetchTextOrNull(`${base}/tests/fixtures/golden_pass.md`)
-  ]);
-  return {
-    id,
-    discipline: id.split("/")[0] ?? "",
-    instruction,
-    task_toml,
-    verify_py: await verify.text(),
-    golden_pass_md
-  };
-}
-function selectFixtures(records, opts) {
-  let tasks = records.map(recordToTask);
-  if (opts.split) tasks = tasks.filter((t) => t.split === opts.split);
-  if (opts.ids) {
-    const want = new Set(opts.ids);
-    tasks = tasks.filter((t) => want.has(t.id));
-  } else if (opts.limit !== void 0) {
-    tasks = tasks.slice(0, opts.limit);
-  }
-  return tasks;
-}
-async function loadFixtures(opts) {
-  const records = JSON.parse(await readFile(FIXTURES, "utf8"));
-  console.warn(
-    `[aec-bench] AEC_FIXTURES=1 \u2014 loading ${records.length} committed fixtures from ${FIXTURES} (no GitHub fetch)`
-  );
-  return selectFixtures(records, opts);
-}
-async function loadLive(opts) {
-  if (opts.ids) {
-    const records2 = [];
-    for (const id of opts.ids) {
-      const rec = await fetchInstance(id);
-      if (!rec) throw new Error(`aec-bench: ${id} has no tests/verify.py (seed-only or wrong id)`);
-      records2.push(rec);
-    }
-    return records2.map(recordToTask);
-  }
-  const limit = opts.limit ?? DEFAULT_LIMIT;
-  let ids = await listAllInstances();
-  if (opts.split) ids = ids.filter((id) => id.startsWith(`${opts.split}/`));
-  const records = [];
-  for (const id of ids) {
-    if (records.length >= limit) break;
-    try {
-      const rec = await fetchInstance(id);
-      if (rec) records.push(rec);
-    } catch (err) {
-      console.warn(`[aec-bench] skipping ${id}: ${err instanceof Error ? err.message : String(err)}`);
-    }
-  }
-  if (records.length === 0) {
-    throw new Error(
-      `aec-bench loadTasks found no runnable instances for ${JSON.stringify(opts)} (no tests/verify.py matched the requested split). Set AEC_FIXTURES=1 to run offline.`
-    );
-  }
-  return records.map(recordToTask);
-}
-async function runVerifier(meta, artifact) {
-  return runStagedJudge({
-    tmpPrefix: "aecbench-",
-    timeoutMs: 12e4,
-    async stage(dir) {
-      await Promise.all([
-        stageFile(join(dir, "output.md"), artifact),
-        stageFile(join(dir, "verify.py"), meta.verifyPy)
-      ]);
-    },
-    bin: venvPython,
-    argv: (dir) => [join(dir, "verify.py"), "--input", join(dir, "output.md"), "--output", join(dir, "reward.json")],
-    async parseReport(dir) {
-      const report = await readJsonReport(join(dir, "reward.json"));
-      const reward = report.reward;
-      if (typeof reward !== "number" || !Number.isFinite(reward)) {
-        throw new Error(`aec-bench verify.py wrote no numeric reward for ${meta.taskId}: ${JSON.stringify(report)}`);
-      }
-      const details = await readFile(join(dir, "details.json"), "utf8").then(
-        (s) => JSON.parse(s),
-        () => ({})
-      );
-      const score = Math.max(0, Math.min(1, reward));
-      return {
-        // resolved = full credit (all fields within tolerance), matching aec-bench's perfect_rate.
-        resolved: score >= 1,
-        score,
-        detail: JSON.stringify({ taskId: meta.taskId, discipline: meta.discipline, reward, fields: details })
-      };
-    }
-  });
-}
-function createAecBenchAdapter() {
-  const fixturesMode = process.env.AEC_FIXTURES === "1";
-  return {
-    name: "aec-bench",
-    async preflight() {
-      await preflightVenvImports({
-        modules: ["math", "json", "re"],
-        requireDocker: false,
-        fix: "Fix: python3 -m venv bench/.venv (verify.py only needs the stdlib \u2014 no pip install)."
-      });
-      if (fixturesMode) {
-        await readFile(FIXTURES, "utf8").catch((err) => {
-          throw new Error(`AEC_FIXTURES=1 but ${FIXTURES} unreadable: ${err instanceof Error ? err.message : err}`);
-        });
-        return;
-      }
-      const res = await fetch(`${RAW}/README.md`, { method: "HEAD" }).catch((err) => {
-        throw new Error(
-          `aec-bench preflight failed reaching ${RAW}: ${err instanceof Error ? err.message : err}
-Fix: ensure network access to raw.githubusercontent.com, or set AEC_FIXTURES=1 to run offline.`
-        );
-      });
-      if (!res.ok) {
-        throw new Error(
-          `aec-bench preflight: ${REPO} README HEAD ${res.status}. Set AEC_FIXTURES=1 to run against committed fixtures.`
-        );
-      }
-    },
-    async loadTasks(opts = {}) {
-      if (fixturesMode) return loadFixtures(opts);
-      try {
-        return await loadLive(opts);
-      } catch (err) {
-        if (err instanceof Error && /fetch \d|ENOTFOUND|getaddrinfo|network/i.test(err.message)) {
-          console.warn(
-            `[aec-bench] live fetch failed (${err.message.slice(0, 160)}); falling back to committed fixtures at ${FIXTURES}`
-          );
-          return loadFixtures(opts);
-        }
-        throw err;
-      }
-    },
-    async goldArtifact(task) {
-      const meta = readMeta(task);
-      return meta.goldenPassMd ?? void 0;
-    },
-    async judge(task, artifact) {
-      const meta = readMeta(task);
-      return runVerifier(meta, artifact);
-    }
-  };
-}
-
-export {
-  createAecBenchAdapter
-};
-//# sourceMappingURL=chunk-BEN6IF2X.js.map

package/dist/chunk-BEN6IF2X.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/benchmarks/aec-bench.ts"],"sourcesContent":["/**\n * AEC-Bench adapter (TheodoreGalanos/aec-bench, MIT) — closed-form\n * Architecture/Engineering/Construction calculation tasks. Worker artifact = a\n * markdown solution ending in a fenced ```json block with the required numeric\n * fields. Judge = the task's OWN `tests/verify.py`, run with python3: it\n * recomputes ground truth from the embedded engineering formulas, extracts the\n * last JSON block from the artifact, scores each field by math.isclose within a\n * per-field rel_tol, and writes {\"reward\": mean} + per-field details.json.\n * GRADED / partial-credit, FULLY DETERMINISTIC — no LLM judge.\n *\n * Distinct from the multimodal nomic-ai/aec-bench; this is the deterministic\n * calculation platform. The runnable-instance verify.py only needs python3 (no\n * Docker, no Harbor) for the pure-calc disciplines, so the local gate runs at\n * conc<=2 without a container backend.\n *\n * Requires for a live run: network to raw.githubusercontent.com /\n * api.github.com (the in-repo tasks tree) + a python3 interpreter (the bench\n * venv) to run verify.py. For offline/CI, set AEC_FIXTURES=1 to load the\n * committed fixtures (bench/fixtures/aec-bench.json) — never a silent fallback.\n */\n\nimport { readFile } from 'node:fs/promises'\nimport { join } from 'node:path'\nimport { benchRoot, preflightVenvImports, readJsonReport, runStagedJudge, stageFile, venvPython } from './_harness'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\n\nconst FIXTURES = join(benchRoot, 'fixtures', 'aec-bench.json')\n\nconst REPO = 'TheodoreGalanos/aec-bench'\nconst RAW = `https://raw.githubusercontent.com/${REPO}/main`\nconst TREE = `https://api.github.com/repos/${REPO}/git/trees/main?recursive=1`\n\n/** Matches every runnable-instance task id at ANY depth under tasks/. */\nconst verifyPathPattern = /^tasks\\/(.+)\\/tests\\/verify\\.py$/\n\n/** Default cap on tasks enumerated before a `limit` is applied. */\nconst DEFAULT_LIMIT = 10\n\ninterface AecRecord {\n  /** Canonical task id `<discipline>/<task>`, e.g. 'electrical/catenary-sag'. */\n  id: string\n  discipline: string\n  /** instruction.md — the self-contained prompt (table + required outputs + JSON schema). */\n  instruction: string\n  /** task.toml — metadata/difficulty/timeouts (carried for trace context). */\n  task_toml: string\n  /** tests/verify.py — the deterministic verifier (recomputes GT, scores fields). */\n  verify_py: string\n  /** tests/fixtures/golden_pass.md — the oracle artifact that scores reward 1.0,\n   *  when the task ships one. Null when the task only ships a non-md ground truth\n   *  (e.g. tests/ground_truth.json); the judge never needs it, only goldArtifact does. */\n  golden_pass_md: string | null\n}\n\ninterface AecMeta {\n  taskId: string\n  discipline: string\n  taskToml: string\n  verifyPy: string\n  goldenPassMd: string | null\n}\n\nfunction recordToTask(rec: AecRecord): BenchTask {\n  const meta: AecMeta = {\n    taskId: rec.id,\n    discipline: rec.discipline,\n    taskToml: rec.task_toml,\n    verifyPy: rec.verify_py,\n    goldenPassMd: rec.golden_pass_md,\n  }\n  return {\n    id: rec.id,\n    split: rec.discipline,\n    // instruction.md is fully self-contained and already specifies the exact JSON\n    // output schema + the \"write to /workspace/output.md\" contract the verifier\n    // keys off — we pass it through verbatim so the verify.py extractor matches.\n    prompt: rec.instruction,\n    metadata: meta as unknown as Record<string, unknown>,\n  }\n}\n\nfunction readMeta(task: BenchTask): AecMeta {\n  const md = task.metadata\n  // Gold is optional (goldenPassMd may be null) — only the verifier + id are required.\n  if (!md || typeof md.verifyPy !== 'string' || typeof md.taskId !== 'string') {\n    throw new Error(`aec-bench task ${task.id} missing verifier metadata — loadTasks did not populate it`)\n  }\n  return md as unknown as AecMeta\n}\n\nasync function fetchText(url: string): Promise<string> {\n  const res = await fetch(url)\n  if (!res.ok) throw new Error(`aec-bench fetch ${res.status}: ${url}`)\n  return res.text()\n}\n\n/** Like fetchText but returns null on a 404 (absent optional file); throws on any other non-OK. */\nasync function fetchTextOrNull(url: string): Promise<string | null> {\n  const res = await fetch(url)\n  if (res.status === 404) return null\n  if (!res.ok) throw new Error(`aec-bench fetch ${res.status}: ${url}`)\n  return res.text()\n}\n\ninterface GitTree {\n  tree: Array<{ path: string; type: string }>\n}\n\n/**\n * One recursive git-tree call enumerates EVERY runnable-instance id at any depth:\n * a task is runnable iff it ships tests/verify.py. The captured group is the id\n * `tasks/<id>/tests/verify.py` → `<id>` (e.g. 'electrical/pf-droop', or a deeper\n * '<discipline>/<family>/<task>'). Throws loud on a non-OK tree response.\n */\nasync function listAllInstances(): Promise<string[]> {\n  const res = await fetch(TREE)\n  if (!res.ok) throw new Error(`aec-bench tree ${res.status}: ${TREE}`)\n  const { tree } = (await res.json()) as GitTree\n  const ids: string[] = []\n  for (const entry of tree) {\n    const m = verifyPathPattern.exec(entry.path)\n    if (m?.[1]) ids.push(m[1])\n  }\n  return ids\n}\n\n/**\n * Fetch one task's instruction.md + task.toml + tests/verify.py + golden_pass.md.\n * Returns null when the dir is a SEED (no runnable verify.py) so enumeration can\n * skip it without faking a task. Gold is OPTIONAL — a task that ships a non-md\n * ground truth (e.g. tests/ground_truth.json) yields golden_pass_md=null; the\n * deterministic judge needs only verify.py.\n */\nasync function fetchInstance(id: string): Promise<AecRecord | null> {\n  const base = `${RAW}/tasks/${id}`\n  const verify = await fetch(`${base}/tests/verify.py`)\n  if (verify.status === 404) return null\n  if (!verify.ok) throw new Error(`aec-bench fetch ${verify.status}: ${id}/tests/verify.py`)\n  const [instruction, task_toml, golden_pass_md] = await Promise.all([\n    fetchText(`${base}/instruction.md`),\n    fetchText(`${base}/task.toml`),\n    fetchTextOrNull(`${base}/tests/fixtures/golden_pass.md`),\n  ])\n  return {\n    id,\n    discipline: id.split('/')[0] ?? '',\n    instruction,\n    task_toml,\n    verify_py: await verify.text(),\n    golden_pass_md,\n  }\n}\n\nfunction selectFixtures(records: AecRecord[], opts: LoadOptions): BenchTask[] {\n  let tasks = records.map(recordToTask)\n  if (opts.split) tasks = tasks.filter((t) => t.split === opts.split)\n  if (opts.ids) {\n    const want = new Set(opts.ids)\n    tasks = tasks.filter((t) => want.has(t.id))\n  } else if (opts.limit !== undefined) {\n    tasks = tasks.slice(0, opts.limit)\n  }\n  return tasks\n}\n\nasync function loadFixtures(opts: LoadOptions): Promise<BenchTask[]> {\n  const records = JSON.parse(await readFile(FIXTURES, 'utf8')) as AecRecord[]\n  console.warn(\n    `[aec-bench] AEC_FIXTURES=1 — loading ${records.length} committed fixtures from ${FIXTURES} (no GitHub fetch)`,\n  )\n  return selectFixtures(records, opts)\n}\n\n/** Enumerate live tasks: explicit ids (each required — throws on a bad id), or a\n *  capped slice of the recursive tree (optionally filtered to one split). Per-task\n *  resilient: a single fetch failure warns + SKIPS that task, never aborting the\n *  batch. Skips seed dirs (no verify.py) — never fabricates. */\nasync function loadLive(opts: LoadOptions): Promise<BenchTask[]> {\n  if (opts.ids) {\n    const records: AecRecord[] = []\n    for (const id of opts.ids) {\n      const rec = await fetchInstance(id)\n      if (!rec) throw new Error(`aec-bench: ${id} has no tests/verify.py (seed-only or wrong id)`)\n      records.push(rec)\n    }\n    return records.map(recordToTask)\n  }\n  const limit = opts.limit ?? DEFAULT_LIMIT\n  let ids = await listAllInstances()\n  if (opts.split) ids = ids.filter((id) => id.startsWith(`${opts.split}/`))\n  const records: AecRecord[] = []\n  for (const id of ids) {\n    if (records.length >= limit) break\n    try {\n      const rec = await fetchInstance(id)\n      if (rec) records.push(rec)\n    } catch (err) {\n      // One bad task must NEVER abort the batch — warn and skip it.\n      console.warn(`[aec-bench] skipping ${id}: ${err instanceof Error ? err.message : String(err)}`)\n    }\n  }\n  if (records.length === 0) {\n    throw new Error(\n      `aec-bench loadTasks found no runnable instances for ${JSON.stringify(opts)} ` +\n        `(no tests/verify.py matched the requested split). Set AEC_FIXTURES=1 to run offline.`,\n    )\n  }\n  return records.map(recordToTask)\n}\n\n/**\n * Run the task's own verify.py with python3 over the artifact via the shared\n * staged-judge spine (mkdtemp → stage → spawn → parseReport → cleanup). verify.py\n * writes {\"reward\": mean} to --output and per-field details.json as a sibling. We\n * read both: reward → graded score; details → sub-scores for the trace-analyst.\n * Fail loud if the verifier never wrote a numeric reward (a crashed verifier is\n * NOT a silent 0 — verify.py's own except-trap writes reward 0.0, so an absent /\n * non-numeric reward.json is a real bug, surfaced by readJsonReport).\n */\nasync function runVerifier(meta: AecMeta, artifact: string): Promise<BenchScore> {\n  return runStagedJudge({\n    tmpPrefix: 'aecbench-',\n    timeoutMs: 120_000,\n    async stage(dir) {\n      await Promise.all([\n        stageFile(join(dir, 'output.md'), artifact),\n        stageFile(join(dir, 'verify.py'), meta.verifyPy),\n      ])\n    },\n    bin: venvPython,\n    argv: (dir) => [join(dir, 'verify.py'), '--input', join(dir, 'output.md'), '--output', join(dir, 'reward.json')],\n    async parseReport(dir) {\n      const report = await readJsonReport<{ reward?: number }>(join(dir, 'reward.json'))\n      const reward = report.reward\n      if (typeof reward !== 'number' || !Number.isFinite(reward)) {\n        throw new Error(`aec-bench verify.py wrote no numeric reward for ${meta.taskId}: ${JSON.stringify(report)}`)\n      }\n      const details = await readFile(join(dir, 'details.json'), 'utf8').then(\n        (s) => JSON.parse(s) as Record<string, number>,\n        () => ({}),\n      )\n      const score = Math.max(0, Math.min(1, reward))\n      return {\n        // resolved = full credit (all fields within tolerance), matching aec-bench's perfect_rate.\n        resolved: score >= 1,\n        score,\n        detail: JSON.stringify({ taskId: meta.taskId, discipline: meta.discipline, reward, fields: details }),\n      }\n    },\n  })\n}\n\nexport function createAecBenchAdapter(): BenchmarkAdapter {\n  const fixturesMode = process.env.AEC_FIXTURES === '1'\n\n  return {\n    name: 'aec-bench',\n\n    async preflight() {\n      // The verifier is python3 over the stdlib (math/json/re) — no pip install.\n      // Reuse the shared import-probe so the bench venv interpreter is proven to\n      // exist + run before any judge spawns verify.py.\n      await preflightVenvImports({\n        modules: ['math', 'json', 're'],\n        requireDocker: false,\n        fix: 'Fix: python3 -m venv bench/.venv (verify.py only needs the stdlib — no pip install).',\n      })\n      if (fixturesMode) {\n        await readFile(FIXTURES, 'utf8').catch((err) => {\n          throw new Error(`AEC_FIXTURES=1 but ${FIXTURES} unreadable: ${err instanceof Error ? err.message : err}`)\n        })\n        return\n      }\n      const res = await fetch(`${RAW}/README.md`, { method: 'HEAD' }).catch((err) => {\n        throw new Error(\n          `aec-bench preflight failed reaching ${RAW}: ${err instanceof Error ? err.message : err}\\n` +\n            `Fix: ensure network access to raw.githubusercontent.com, or set AEC_FIXTURES=1 to run offline.`,\n        )\n      })\n      if (!res.ok) {\n        throw new Error(\n          `aec-bench preflight: ${REPO} README HEAD ${res.status}. Set AEC_FIXTURES=1 to run against committed fixtures.`,\n        )\n      }\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      if (fixturesMode) return loadFixtures(opts)\n      try {\n        return await loadLive(opts)\n      } catch (err) {\n        // A reachability failure falls back to fixtures with an explicit warn; a\n        // wrong-id / seed-only error (the loader's own throw) propagates.\n        if (err instanceof Error && /fetch \\d|ENOTFOUND|getaddrinfo|network/i.test(err.message)) {\n          console.warn(\n            `[aec-bench] live fetch failed (${err.message.slice(0, 160)}); falling back to committed fixtures at ${FIXTURES}`,\n          )\n          return loadFixtures(opts)\n        }\n        throw err\n      }\n    },\n\n    async goldArtifact(task: BenchTask) {\n      // Gold = the task's own golden_pass.md (scores reward 1.0 through the SAME\n      // verify.py the real artifact takes), proving the judge end-to-end. Tasks\n      // without a golden_pass.md (non-md ground truth) have no oracle artifact.\n      const meta = readMeta(task)\n      return meta.goldenPassMd ?? undefined\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const meta = readMeta(task)\n      // verify.py fail-closes an empty/unparseable artifact to reward 0.0 itself,\n      // so we pass it straight through (no pre-judging here).\n      return runVerifier(meta, artifact)\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;AAqBA,SAAS,gBAAgB;AACzB,SAAS,YAAY;AAIrB,IAAM,WAAW,KAAK,WAAW,YAAY,gBAAgB;AAE7D,IAAM,OAAO;AACb,IAAM,MAAM,qCAAqC,IAAI;AACrD,IAAM,OAAO,gCAAgC,IAAI;AAGjD,IAAM,oBAAoB;AAG1B,IAAM,gBAAgB;AA0BtB,SAAS,aAAa,KAA2B;AAC/C,QAAM,OAAgB;AAAA,IACpB,QAAQ,IAAI;AAAA,IACZ,YAAY,IAAI;AAAA,IAChB,UAAU,IAAI;AAAA,IACd,UAAU,IAAI;AAAA,IACd,cAAc,IAAI;AAAA,EACpB;AACA,SAAO;AAAA,IACL,IAAI,IAAI;AAAA,IACR,OAAO,IAAI;AAAA;AAAA;AAAA;AAAA,IAIX,QAAQ,IAAI;AAAA,IACZ,UAAU;AAAA,EACZ;AACF;AAEA,SAAS,SAAS,MAA0B;AAC1C,QAAM,KAAK,KAAK;AAEhB,MAAI,CAAC,MAAM,OAAO,GAAG,aAAa,YAAY,OAAO,GAAG,WAAW,UAAU;AAC3E,UAAM,IAAI,MAAM,kBAAkB,KAAK,EAAE,iEAA4D;AAAA,EACvG;AACA,SAAO;AACT;AAEA,eAAe,UAAU,KAA8B;AACrD,QAAM,MAAM,MAAM,MAAM,GAAG;AAC3B,MAAI,CAAC,IAAI,GAAI,OAAM,IAAI,MAAM,mBAAmB,IAAI,MAAM,KAAK,GAAG,EAAE;AACpE,SAAO,IAAI,KAAK;AAClB;AAGA,eAAe,gBAAgB,KAAqC;AAClE,QAAM,MAAM,MAAM,MAAM,GAAG;AAC3B,MAAI,IAAI,WAAW,IAAK,QAAO;AAC/B,MAAI,CAAC,IAAI,GAAI,OAAM,IAAI,MAAM,mBAAmB,IAAI,MAAM,KAAK,GAAG,EAAE;AACpE,SAAO,IAAI,KAAK;AAClB;AAYA,eAAe,mBAAsC;AACnD,QAAM,MAAM,MAAM,MAAM,IAAI;AAC5B,MAAI,CAAC,IAAI,GAAI,OAAM,IAAI,MAAM,kBAAkB,IAAI,MAAM,KAAK,IAAI,EAAE;AACpE,QAAM,EAAE,KAAK,IAAK,MAAM,IAAI,KAAK;AACjC,QAAM,MAAgB,CAAC;AACvB,aAAW,SAAS,MAAM;AACxB,UAAM,IAAI,kBAAkB,KAAK,MAAM,IAAI;AAC3C,QAAI,IAAI,CAAC,EAAG,KAAI,KAAK,EAAE,CAAC,CAAC;AAAA,EAC3B;AACA,SAAO;AACT;AASA,eAAe,cAAc,IAAuC;AAClE,QAAM,OAAO,GAAG,GAAG,UAAU,EAAE;AAC/B,QAAM,SAAS,MAAM,MAAM,GAAG,IAAI,kBAAkB;AACpD,MAAI,OAAO,WAAW,IAAK,QAAO;AAClC,MAAI,CAAC,OAAO,GAAI,OAAM,IAAI,MAAM,mBAAmB,OAAO,MAAM,KAAK,EAAE,kBAAkB;AACzF,QAAM,CAAC,aAAa,WAAW,cAAc,IAAI,MAAM,QAAQ,IAAI;AAAA,IACjE,UAAU,GAAG,IAAI,iBAAiB;AAAA,IAClC,UAAU,GAAG,IAAI,YAAY;AAAA,IAC7B,gBAAgB,GAAG,IAAI,gCAAgC;AAAA,EACzD,CAAC;AACD,SAAO;AAAA,IACL;AAAA,IACA,YAAY,GAAG,MAAM,GAAG,EAAE,CAAC,KAAK;AAAA,IAChC;AAAA,IACA;AAAA,IACA,WAAW,MAAM,OAAO,KAAK;AAAA,IAC7B;AAAA,EACF;AACF;AAEA,SAAS,eAAe,SAAsB,MAAgC;AAC5E,MAAI,QAAQ,QAAQ,IAAI,YAAY;AACpC,MAAI,KAAK,MAAO,SAAQ,MAAM,OAAO,CAAC,MAAM,EAAE,UAAU,KAAK,KAAK;AAClE,MAAI,KAAK,KAAK;AACZ,UAAM,OAAO,IAAI,IAAI,KAAK,GAAG;AAC7B,YAAQ,MAAM,OAAO,CAAC,MAAM,KAAK,IAAI,EAAE,EAAE,CAAC;AAAA,EAC5C,WAAW,KAAK,UAAU,QAAW;AACnC,YAAQ,MAAM,MAAM,GAAG,KAAK,KAAK;AAAA,EACnC;AACA,SAAO;AACT;AAEA,eAAe,aAAa,MAAyC;AACnE,QAAM,UAAU,KAAK,MAAM,MAAM,SAAS,UAAU,MAAM,CAAC;AAC3D,UAAQ;AAAA,IACN,6CAAwC,QAAQ,MAAM,4BAA4B,QAAQ;AAAA,EAC5F;AACA,SAAO,eAAe,SAAS,IAAI;AACrC;AAMA,eAAe,SAAS,MAAyC;AAC/D,MAAI,KAAK,KAAK;AACZ,UAAMA,WAAuB,CAAC;AAC9B,eAAW,MAAM,KAAK,KAAK;AACzB,YAAM,MAAM,MAAM,cAAc,EAAE;AAClC,UAAI,CAAC,IAAK,OAAM,IAAI,MAAM,cAAc,EAAE,iDAAiD;AAC3F,MAAAA,SAAQ,KAAK,GAAG;AAAA,IAClB;AACA,WAAOA,SAAQ,IAAI,YAAY;AAAA,EACjC;AACA,QAAM,QAAQ,KAAK,SAAS;AAC5B,MAAI,MAAM,MAAM,iBAAiB;AACjC,MAAI,KAAK,MAAO,OAAM,IAAI,OAAO,CAAC,OAAO,GAAG,WAAW,GAAG,KAAK,KAAK,GAAG,CAAC;AACxE,QAAM,UAAuB,CAAC;AAC9B,aAAW,MAAM,KAAK;AACpB,QAAI,QAAQ,UAAU,MAAO;AAC7B,QAAI;AACF,YAAM,MAAM,MAAM,cAAc,EAAE;AAClC,UAAI,IAAK,SAAQ,KAAK,GAAG;AAAA,IAC3B,SAAS,KAAK;AAEZ,cAAQ,KAAK,wBAAwB,EAAE,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC,EAAE;AAAA,IAChG;AAAA,EACF;AACA,MAAI,QAAQ,WAAW,GAAG;AACxB,UAAM,IAAI;AAAA,MACR,uDAAuD,KAAK,UAAU,IAAI,CAAC;AAAA,IAE7E;AAAA,EACF;AACA,SAAO,QAAQ,IAAI,YAAY;AACjC;AAWA,eAAe,YAAY,MAAe,UAAuC;AAC/E,SAAO,eAAe;AAAA,IACpB,WAAW;AAAA,IACX,WAAW;AAAA,IACX,MAAM,MAAM,KAAK;AACf,YAAM,QAAQ,IAAI;AAAA,QAChB,UAAU,KAAK,KAAK,WAAW,GAAG,QAAQ;AAAA,QAC1C,UAAU,KAAK,KAAK,WAAW,GAAG,KAAK,QAAQ;AAAA,MACjD,CAAC;AAAA,IACH;AAAA,IACA,KAAK;AAAA,IACL,MAAM,CAAC,QAAQ,CAAC,KAAK,KAAK,WAAW,GAAG,WAAW,KAAK,KAAK,WAAW,GAAG,YAAY,KAAK,KAAK,aAAa,CAAC;AAAA,IAC/G,MAAM,YAAY,KAAK;AACrB,YAAM,SAAS,MAAM,eAAoC,KAAK,KAAK,aAAa,CAAC;AACjF,YAAM,SAAS,OAAO;AACtB,UAAI,OAAO,WAAW,YAAY,CAAC,OAAO,SAAS,MAAM,GAAG;AAC1D,cAAM,IAAI,MAAM,mDAAmD,KAAK,MAAM,KAAK,KAAK,UAAU,MAAM,CAAC,EAAE;AAAA,MAC7G;AACA,YAAM,UAAU,MAAM,SAAS,KAAK,KAAK,cAAc,GAAG,MAAM,EAAE;AAAA,QAChE,CAAC,MAAM,KAAK,MAAM,CAAC;AAAA,QACnB,OAAO,CAAC;AAAA,MACV;AACA,YAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,MAAM,CAAC;AAC7C,aAAO;AAAA;AAAA,QAEL,UAAU,SAAS;AAAA,QACnB;AAAA,QACA,QAAQ,KAAK,UAAU,EAAE,QAAQ,KAAK,QAAQ,YAAY,KAAK,YAAY,QAAQ,QAAQ,QAAQ,CAAC;AAAA,MACtG;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAEO,SAAS,wBAA0C;AACxD,QAAM,eAAe,QAAQ,IAAI,iBAAiB;AAElD,SAAO;AAAA,IACL,MAAM;AAAA,IAEN,MAAM,YAAY;AAIhB,YAAM,qBAAqB;AAAA,QACzB,SAAS,CAAC,QAAQ,QAAQ,IAAI;AAAA,QAC9B,eAAe;AAAA,QACf,KAAK;AAAA,MACP,CAAC;AACD,UAAI,cAAc;AAChB,cAAM,SAAS,UAAU,MAAM,EAAE,MAAM,CAAC,QAAQ;AAC9C,gBAAM,IAAI,MAAM,sBAAsB,QAAQ,gBAAgB,eAAe,QAAQ,IAAI,UAAU,GAAG,EAAE;AAAA,QAC1G,CAAC;AACD;AAAA,MACF;AACA,YAAM,MAAM,MAAM,MAAM,GAAG,GAAG,cAAc,EAAE,QAAQ,OAAO,CAAC,EAAE,MAAM,CAAC,QAAQ;AAC7E,cAAM,IAAI;AAAA,UACR,uCAAuC,GAAG,KAAK,eAAe,QAAQ,IAAI,UAAU,GAAG;AAAA;AAAA,QAEzF;AAAA,MACF,CAAC;AACD,UAAI,CAAC,IAAI,IAAI;AACX,cAAM,IAAI;AAAA,UACR,wBAAwB,IAAI,gBAAgB,IAAI,MAAM;AAAA,QACxD;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;AACtC,UAAI,aAAc,QAAO,aAAa,IAAI;AAC1C,UAAI;AACF,eAAO,MAAM,SAAS,IAAI;AAAA,MAC5B,SAAS,KAAK;AAGZ,YAAI,eAAe,SAAS,0CAA0C,KAAK,IAAI,OAAO,GAAG;AACvF,kBAAQ;AAAA,YACN,kCAAkC,IAAI,QAAQ,MAAM,GAAG,GAAG,CAAC,4CAA4C,QAAQ;AAAA,UACjH;AACA,iBAAO,aAAa,IAAI;AAAA,QAC1B;AACA,cAAM;AAAA,MACR;AAAA,IACF;AAAA,IAEA,MAAM,aAAa,MAAiB;AAIlC,YAAM,OAAO,SAAS,IAAI;AAC1B,aAAO,KAAK,gBAAgB;AAAA,IAC9B;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,OAAO,SAAS,IAAI;AAG1B,aAAO,YAAY,MAAM,QAAQ;AAAA,IACnC;AAAA,EACF;AACF;","names":["records"]}

package/dist/chunk-BZY5QARD.js

@@ -1,120 +0,0 @@
-import {
-  FINAL_ANSWER_SENTINEL,
-  allStrings,
-  answerScoreToBenchScore,
-  contextBlock,
-  contextsFrom,
-  firstString,
-  isObject,
-  ragAnswerOutput,
-  readJsonRows,
-  scoreAnswerArtifact,
-  selectTasks,
-  stringFrom
-} from "./chunk-X3BTXCJ4.js";
-import {
-  benchRoot
-} from "./chunk-EIETHPD5.js";
-
-// src/benchmarks/open-rag-bench.ts
-import { readFile } from "fs/promises";
-import { join } from "path";
-var FIXTURES = join(benchRoot, "fixtures", "open-rag-bench.json");
-var dataFile = () => process.env.OPEN_RAG_BENCH_DATA_FILE;
-function rowToTask(raw, index) {
-  if (!isObject(raw)) throw new Error(`Open RAG Bench row ${index} must be an object`);
-  const query = firstString(raw, ["question", "query", "prompt"]);
-  const goldAnswers = allStrings(raw, ["answer", "answers", "reference", "reference_answer", "gold"]);
-  if (!query) throw new Error(`Open RAG Bench row ${index} missing question/query`);
-  if (goldAnswers.length === 0) throw new Error(`Open RAG Bench row ${index} missing answer`);
-  const contexts = contextsFrom(raw.contexts).length > 0 ? contextsFrom(raw.contexts) : contextsFrom(raw.chunks).length > 0 ? contextsFrom(raw.chunks) : contextsFrom(raw.pages);
-  const documentId = stringFrom(raw.document_id) ?? stringFrom(raw.doc_id) ?? stringFrom(raw.pdf_id) ?? stringFrom(raw.file_name) ?? stringFrom(raw.filename) ?? "unknown";
-  const modality = stringFrom(raw.modality) ?? stringFrom(raw.answer_modality) ?? "unknown";
-  const id = stringFrom(raw.id) ?? stringFrom(raw.query_id) ?? `open-rag-bench-${index}`;
-  const meta = {
-    benchmark: "open-rag-bench",
-    query,
-    goldAnswers,
-    contexts,
-    documentId,
-    modality
-  };
-  return {
-    id,
-    split: stringFrom(raw.split) ?? modality,
-    prompt: [
-      "Answer this Open RAG Bench PDF question using the supplied document context.",
-      "Use tables, text, and image-derived notes when present.",
-      "End with a single final line: `FINAL ANSWER: <answer>`.",
-      "",
-      `Question: ${query}`,
-      `Document: ${documentId}`,
-      `Modality: ${modality}`,
-      contexts.length > 0 ? `
-Document context:
-${contextBlock(contexts)}` : void 0
-    ].filter(Boolean).join("\n"),
-    metadata: meta
-  };
-}
-function readMeta(task) {
-  const md = task.metadata;
-  if (!md || !Array.isArray(md.goldAnswers)) {
-    throw new Error(`Open RAG Bench task ${task.id} missing metadata \u2014 loadTasks did not populate it`);
-  }
-  return md;
-}
-async function loadRows(path) {
-  const rows = await readJsonRows(path);
-  if (rows.length === 0) throw new Error(`Open RAG Bench: no rows in ${path}`);
-  return rows;
-}
-async function loadFixtures(opts) {
-  const rows = JSON.parse(await readFile(FIXTURES, "utf8"));
-  console.warn(`[open-rag-bench] OPEN_RAG_BENCH_FIXTURES=1 \u2014 loading ${rows.length} adapter fixtures`);
-  return selectTasks(rows.map(rowToTask), opts, "Open RAG Bench");
-}
-function createOpenRagBenchAdapter() {
-  const fixturesMode = process.env.OPEN_RAG_BENCH_FIXTURES === "1";
-  return {
-    name: "open-rag-bench",
-    output: ragAnswerOutput,
-    async preflight() {
-      if (fixturesMode) {
-        await readFile(FIXTURES, "utf8");
-        return;
-      }
-      const path = dataFile();
-      if (!path) {
-        throw new Error(
-          "OPEN_RAG_BENCH_DATA_FILE is required. Fix: export vectara/open-rag-bench rows to JSONL and set OPEN_RAG_BENCH_DATA_FILE=/path/to/open-rag-bench.jsonl, or set OPEN_RAG_BENCH_FIXTURES=1 for adapter plumbing."
-        );
-      }
-      await loadRows(path);
-    },
-    async loadTasks(opts = {}) {
-      if (fixturesMode) return loadFixtures(opts);
-      const path = dataFile();
-      if (!path) throw new Error("OPEN_RAG_BENCH_DATA_FILE is required to load Open RAG Bench tasks");
-      return selectTasks((await loadRows(path)).map(rowToTask), opts, "Open RAG Bench");
-    },
-    async goldArtifact(task) {
-      return `${FINAL_ANSWER_SENTINEL} ${readMeta(task).goldAnswers[0] ?? ""}`;
-    },
-    async judge(task, artifact) {
-      const meta = readMeta(task);
-      const score = scoreAnswerArtifact(artifact, meta.goldAnswers);
-      return answerScoreToBenchScore(score, {
-        benchmark: meta.benchmark,
-        documentId: meta.documentId,
-        modality: meta.modality,
-        contextCount: meta.contexts.length
-      });
-    }
-  };
-}
-
-export {
-  createOpenRagBenchAdapter
-};
-//# sourceMappingURL=chunk-BZY5QARD.js.map

package/dist/chunk-BZY5QARD.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/benchmarks/open-rag-bench.ts"],"sourcesContent":["/**\n * Open RAG Bench adapter.\n *\n * This targets Vectara-style Open RAG Bench exports over PDF-derived text,\n * table, and image contexts. The deterministic judge scores final-answer\n * agreement and surfaces modality/document metadata for diagnostics.\n */\n\nimport { readFile } from 'node:fs/promises'\nimport { join } from 'node:path'\nimport { benchRoot } from './_harness'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\nimport {\n  FINAL_ANSWER_SENTINEL,\n  allStrings,\n  answerScoreToBenchScore,\n  contextBlock,\n  contextsFrom,\n  firstString,\n  isObject,\n  ragAnswerOutput,\n  readJsonRows,\n  scoreAnswerArtifact,\n  selectTasks,\n  stringFrom,\n  type RagContext,\n} from './rag-shared'\n\nconst FIXTURES = join(benchRoot, 'fixtures', 'open-rag-bench.json')\n\ninterface OpenRagBenchMeta {\n  benchmark: 'open-rag-bench'\n  query: string\n  goldAnswers: string[]\n  contexts: RagContext[]\n  documentId: string\n  modality: string\n}\n\nconst dataFile = (): string | undefined => process.env.OPEN_RAG_BENCH_DATA_FILE\n\nfunction rowToTask(raw: unknown, index: number): BenchTask {\n  if (!isObject(raw)) throw new Error(`Open RAG Bench row ${index} must be an object`)\n  const query = firstString(raw, ['question', 'query', 'prompt'])\n  const goldAnswers = allStrings(raw, ['answer', 'answers', 'reference', 'reference_answer', 'gold'])\n  if (!query) throw new Error(`Open RAG Bench row ${index} missing question/query`)\n  if (goldAnswers.length === 0) throw new Error(`Open RAG Bench row ${index} missing answer`)\n  const contexts =\n    contextsFrom(raw.contexts).length > 0\n      ? contextsFrom(raw.contexts)\n      : contextsFrom(raw.chunks).length > 0\n        ? contextsFrom(raw.chunks)\n        : contextsFrom(raw.pages)\n  const documentId =\n    stringFrom(raw.document_id) ??\n    stringFrom(raw.doc_id) ??\n    stringFrom(raw.pdf_id) ??\n    stringFrom(raw.file_name) ??\n    stringFrom(raw.filename) ??\n    'unknown'\n  const modality = stringFrom(raw.modality) ?? stringFrom(raw.answer_modality) ?? 'unknown'\n  const id = stringFrom(raw.id) ?? stringFrom(raw.query_id) ?? `open-rag-bench-${index}`\n  const meta: OpenRagBenchMeta = {\n    benchmark: 'open-rag-bench',\n    query,\n    goldAnswers,\n    contexts,\n    documentId,\n    modality,\n  }\n  return {\n    id,\n    split: stringFrom(raw.split) ?? modality,\n    prompt: [\n      'Answer this Open RAG Bench PDF question using the supplied document context.',\n      'Use tables, text, and image-derived notes when present.',\n      'End with a single final line: `FINAL ANSWER: <answer>`.',\n      '',\n      `Question: ${query}`,\n      `Document: ${documentId}`,\n      `Modality: ${modality}`,\n      contexts.length > 0 ? `\\nDocument context:\\n${contextBlock(contexts)}` : undefined,\n    ]\n      .filter(Boolean)\n      .join('\\n'),\n    metadata: meta as unknown as Record<string, unknown>,\n  }\n}\n\nfunction readMeta(task: BenchTask): OpenRagBenchMeta {\n  const md = task.metadata\n  if (!md || !Array.isArray(md.goldAnswers)) {\n    throw new Error(`Open RAG Bench task ${task.id} missing metadata — loadTasks did not populate it`)\n  }\n  return md as unknown as OpenRagBenchMeta\n}\n\nasync function loadRows(path: string): Promise<unknown[]> {\n  const rows = await readJsonRows(path)\n  if (rows.length === 0) throw new Error(`Open RAG Bench: no rows in ${path}`)\n  return rows\n}\n\nasync function loadFixtures(opts: LoadOptions): Promise<BenchTask[]> {\n  const rows = JSON.parse(await readFile(FIXTURES, 'utf8')) as unknown[]\n  console.warn(`[open-rag-bench] OPEN_RAG_BENCH_FIXTURES=1 — loading ${rows.length} adapter fixtures`)\n  return selectTasks(rows.map(rowToTask), opts, 'Open RAG Bench')\n}\n\nexport function createOpenRagBenchAdapter(): BenchmarkAdapter {\n  const fixturesMode = process.env.OPEN_RAG_BENCH_FIXTURES === '1'\n\n  return {\n    name: 'open-rag-bench',\n    output: ragAnswerOutput,\n\n    async preflight() {\n      if (fixturesMode) {\n        await readFile(FIXTURES, 'utf8')\n        return\n      }\n      const path = dataFile()\n      if (!path) {\n        throw new Error(\n          'OPEN_RAG_BENCH_DATA_FILE is required. Fix: export vectara/open-rag-bench rows to JSONL and set OPEN_RAG_BENCH_DATA_FILE=/path/to/open-rag-bench.jsonl, or set OPEN_RAG_BENCH_FIXTURES=1 for adapter plumbing.',\n        )\n      }\n      await loadRows(path)\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      if (fixturesMode) return loadFixtures(opts)\n      const path = dataFile()\n      if (!path) throw new Error('OPEN_RAG_BENCH_DATA_FILE is required to load Open RAG Bench tasks')\n      return selectTasks((await loadRows(path)).map(rowToTask), opts, 'Open RAG Bench')\n    },\n\n    async goldArtifact(task: BenchTask) {\n      return `${FINAL_ANSWER_SENTINEL} ${readMeta(task).goldAnswers[0] ?? ''}`\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const meta = readMeta(task)\n      const score = scoreAnswerArtifact(artifact, meta.goldAnswers)\n      return answerScoreToBenchScore(score, {\n        benchmark: meta.benchmark,\n        documentId: meta.documentId,\n        modality: meta.modality,\n        contextCount: meta.contexts.length,\n      })\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;AAQA,SAAS,gBAAgB;AACzB,SAAS,YAAY;AAmBrB,IAAM,WAAW,KAAK,WAAW,YAAY,qBAAqB;AAWlE,IAAM,WAAW,MAA0B,QAAQ,IAAI;AAEvD,SAAS,UAAU,KAAc,OAA0B;AACzD,MAAI,CAAC,SAAS,GAAG,EAAG,OAAM,IAAI,MAAM,sBAAsB,KAAK,oBAAoB;AACnF,QAAM,QAAQ,YAAY,KAAK,CAAC,YAAY,SAAS,QAAQ,CAAC;AAC9D,QAAM,cAAc,WAAW,KAAK,CAAC,UAAU,WAAW,aAAa,oBAAoB,MAAM,CAAC;AAClG,MAAI,CAAC,MAAO,OAAM,IAAI,MAAM,sBAAsB,KAAK,yBAAyB;AAChF,MAAI,YAAY,WAAW,EAAG,OAAM,IAAI,MAAM,sBAAsB,KAAK,iBAAiB;AAC1F,QAAM,WACJ,aAAa,IAAI,QAAQ,EAAE,SAAS,IAChC,aAAa,IAAI,QAAQ,IACzB,aAAa,IAAI,MAAM,EAAE,SAAS,IAChC,aAAa,IAAI,MAAM,IACvB,aAAa,IAAI,KAAK;AAC9B,QAAM,aACJ,WAAW,IAAI,WAAW,KAC1B,WAAW,IAAI,MAAM,KACrB,WAAW,IAAI,MAAM,KACrB,WAAW,IAAI,SAAS,KACxB,WAAW,IAAI,QAAQ,KACvB;AACF,QAAM,WAAW,WAAW,IAAI,QAAQ,KAAK,WAAW,IAAI,eAAe,KAAK;AAChF,QAAM,KAAK,WAAW,IAAI,EAAE,KAAK,WAAW,IAAI,QAAQ,KAAK,kBAAkB,KAAK;AACpF,QAAM,OAAyB;AAAA,IAC7B,WAAW;AAAA,IACX;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,SAAO;AAAA,IACL;AAAA,IACA,OAAO,WAAW,IAAI,KAAK,KAAK;AAAA,IAChC,QAAQ;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,aAAa,KAAK;AAAA,MAClB,aAAa,UAAU;AAAA,MACvB,aAAa,QAAQ;AAAA,MACrB,SAAS,SAAS,IAAI;AAAA;AAAA,EAAwB,aAAa,QAAQ,CAAC,KAAK;AAAA,IAC3E,EACG,OAAO,OAAO,EACd,KAAK,IAAI;AAAA,IACZ,UAAU;AAAA,EACZ;AACF;AAEA,SAAS,SAAS,MAAmC;AACnD,QAAM,KAAK,KAAK;AAChB,MAAI,CAAC,MAAM,CAAC,MAAM,QAAQ,GAAG,WAAW,GAAG;AACzC,UAAM,IAAI,MAAM,uBAAuB,KAAK,EAAE,wDAAmD;AAAA,EACnG;AACA,SAAO;AACT;AAEA,eAAe,SAAS,MAAkC;AACxD,QAAM,OAAO,MAAM,aAAa,IAAI;AACpC,MAAI,KAAK,WAAW,EAAG,OAAM,IAAI,MAAM,8BAA8B,IAAI,EAAE;AAC3E,SAAO;AACT;AAEA,eAAe,aAAa,MAAyC;AACnE,QAAM,OAAO,KAAK,MAAM,MAAM,SAAS,UAAU,MAAM,CAAC;AACxD,UAAQ,KAAK,6DAAwD,KAAK,MAAM,mBAAmB;AACnG,SAAO,YAAY,KAAK,IAAI,SAAS,GAAG,MAAM,gBAAgB;AAChE;AAEO,SAAS,4BAA8C;AAC5D,QAAM,eAAe,QAAQ,IAAI,4BAA4B;AAE7D,SAAO;AAAA,IACL,MAAM;AAAA,IACN,QAAQ;AAAA,IAER,MAAM,YAAY;AAChB,UAAI,cAAc;AAChB,cAAM,SAAS,UAAU,MAAM;AAC/B;AAAA,MACF;AACA,YAAM,OAAO,SAAS;AACtB,UAAI,CAAC,MAAM;AACT,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AACA,YAAM,SAAS,IAAI;AAAA,IACrB;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;AACtC,UAAI,aAAc,QAAO,aAAa,IAAI;AAC1C,YAAM,OAAO,SAAS;AACtB,UAAI,CAAC,KAAM,OAAM,IAAI,MAAM,mEAAmE;AAC9F,aAAO,aAAa,MAAM,SAAS,IAAI,GAAG,IAAI,SAAS,GAAG,MAAM,gBAAgB;AAAA,IAClF;AAAA,IAEA,MAAM,aAAa,MAAiB;AAClC,aAAO,GAAG,qBAAqB,IAAI,SAAS,IAAI,EAAE,YAAY,CAAC,KAAK,EAAE;AAAA,IACxE;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,OAAO,SAAS,IAAI;AAC1B,YAAM,QAAQ,oBAAoB,UAAU,KAAK,WAAW;AAC5D,aAAO,wBAAwB,OAAO;AAAA,QACpC,WAAW,KAAK;AAAA,QAChB,YAAY,KAAK;AAAA,QACjB,UAAU,KAAK;AAAA,QACf,cAAc,KAAK,SAAS;AAAA,MAC9B,CAAC;AAAA,IACH;AAAA,EACF;AACF;","names":[]}

package/dist/chunk-C7T7WEK2.js

@@ -1,103 +0,0 @@
-// src/benchmarks/cadgenbench.ts
-import { execFile as execFile2 } from "child_process";
-import { mkdtemp as mkdtemp2, rm as rm2, writeFile as writeFile2 } from "fs/promises";
-import { tmpdir as tmpdir2 } from "os";
-import { join as join2 } from "path";
-import { promisify as promisify2 } from "util";
-
-// src/worker-build123d.ts
-import { execFile } from "child_process";
-import { existsSync, readFileSync } from "fs";
-import { mkdtemp, readFile, rm, writeFile } from "fs/promises";
-import { tmpdir } from "os";
-import { join } from "path";
-import { promisify } from "util";
-import { routerChatWithUsage } from "@tangle-network/agent-runtime/loops";
-var execFileAsync = promisify(execFile);
-var CGB_VENV_PY = process.env.CADGENBENCH_VENV ?? "/tmp/cgb-venv/bin/python";
-var CGB_DIR = process.env.CADGENBENCH_DIR ?? "/tmp/cadgenbench";
-
-// src/benchmarks/cadgenbench.ts
-var execFileAsync2 = promisify2(execFile2);
-var SCORE_PY = `
-import sys, json, tempfile, shutil
-from pathlib import Path
-from cadgenbench.eval.evaluate import evaluate_result
-cand, gt = Path(sys.argv[1]), Path(sys.argv[2])
-with tempfile.TemporaryDirectory() as rd, tempfile.TemporaryDirectory() as gd:
-    rd, gd = Path(rd), Path(gd)
-    (rd / 'result.json').write_text('{}')
-    shutil.copy(gt, gd / 'ground_truth.step')
-    try:
-        evaluate_result(rd, gd, candidate_step=cand)
-        d = json.loads((rd / 'result.json').read_text())
-        print('CGB_SCORE ' + json.dumps({'cad_score': d.get('cad_score', 0.0), 'status': d.get('status', 'unknown')}))
-    except Exception as e:
-        print('CGB_SCORE ' + json.dumps({'cad_score': 0.0, 'status': 'error', 'error': str(e)[:200]}))
-`.trim();
-function fixtureTasks() {
-  const g = join2(CGB_DIR, "tests/fixtures/geometry");
-  return [
-    { id: "box-10x20x30", prompt: "A rectangular solid box, 10 units wide (X), 20 units deep (Y), and 30 units tall (Z).", gtStep: join2(g, "box_10_20_30.step") },
-    { id: "cube-10", prompt: "A cube, 10 units on every side.", gtStep: join2(g, "box_10_10_10.step") },
-    { id: "sphere-10", prompt: "A sphere of radius 10 units, centered at the origin.", gtStep: join2(g, "sphere_10.step") }
-  ];
-}
-function createCadGenBenchAdapter() {
-  return {
-    name: "cadgenbench",
-    async preflight() {
-      const r = await execFileAsync2(CGB_VENV_PY, ["-c", 'import cadgenbench.eval.evaluate, build123d, trimesh, manifold3d; print("ok")'], { timeout: 6e4 }).catch(
-        (e) => ({ stdout: "", stderr: e instanceof Error ? e.message : String(e) })
-      );
-      if (!/ok/.test(r.stdout)) {
-        throw new Error(
-          `cadgenbench preflight failed (venv=${CGB_VENV_PY}): ${r.stderr.slice(0, 200)}
-Fix: git clone https://github.com/huggingface/cadgenbench ${CGB_DIR}; python3 -m venv $CADGENBENCH_VENV; $CADGENBENCH_VENV/bin/pip install -e ${CGB_DIR}`
-        );
-      }
-    },
-    async loadTasks(opts = {}) {
-      let tasks = fixtureTasks();
-      const hard = process.env.CGB_HARD_DIR;
-      if (hard) {
-        const { readFile: readFile2 } = await import("fs/promises");
-        tasks = JSON.parse(await readFile2(join2(hard, "tasks.json"), "utf8"));
-      }
-      if (opts.ids) tasks = tasks.filter((t) => opts.ids.includes(t.id));
-      if (opts.limit != null) tasks = tasks.slice(0, opts.limit);
-      const meta = (gtStep) => ({ gtStep, resolveThreshold: Number(process.env.CGB_RESOLVE_THRESHOLD ?? 0.9) });
-      return tasks.map((t) => ({ id: t.id, prompt: t.prompt, metadata: meta(t.gtStep) }));
-    },
-    async goldArtifact() {
-      return void 0;
-    },
-    async judge(task, artifact) {
-      const { gtStep, resolveThreshold } = task.metadata;
-      if (!artifact.includes("ISO-10303-21")) return { resolved: false, score: 0, detail: "artifact is not a STEP file" };
-      const dir = await mkdtemp2(join2(tmpdir2(), "cgb-judge-"));
-      const cand = join2(dir, "candidate.step");
-      const scorer = join2(dir, "score.py");
-      try {
-        await writeFile2(cand, artifact);
-        await writeFile2(scorer, SCORE_PY);
-        const r = await execFileAsync2("xvfb-run", ["-a", CGB_VENV_PY, scorer, cand, gtStep], { maxBuffer: 1 << 26, timeout: 18e4 }).catch(
-          (e) => ({ stdout: e.stdout ?? "", stderr: e instanceof Error ? e.message : String(e) })
-        );
-        const m = /CGB_SCORE (\{.*\})/.exec(r.stdout);
-        if (!m) return { resolved: false, score: 0, detail: `scorer produced no verdict: ${(r.stderr || r.stdout).slice(0, 160)}` };
-        const v = JSON.parse(m[1]);
-        const score = typeof v.cad_score === "number" ? v.cad_score : 0;
-        return { resolved: score >= resolveThreshold, score, detail: `cad_score=${score.toFixed(3)} status=${v.status}${v.error ? ` (${v.error})` : ""}` };
-      } finally {
-        await rm2(dir, { recursive: true, force: true }).catch(() => {
-        });
-      }
-    }
-  };
-}
-
-export {
-  createCadGenBenchAdapter
-};
-//# sourceMappingURL=chunk-C7T7WEK2.js.map

package/dist/chunk-C7T7WEK2.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/benchmarks/cadgenbench.ts","../src/worker-build123d.ts"],"sourcesContent":["/**\n * CADGenBench adapter (huggingface/cadgenbench, Apache-2.0). Task = a part\n * description → a STEP B-rep solid (output.step). Score = the benchmark's OWN\n * deterministic geometric metric (cad_score): validity gate → PCA/ICP align to\n * the ground truth → point-cloud F1 + volume IoU + edge F1 + topology match.\n * NOT an LLM judge, NOT self-defined checks — the published CAD kernel decides.\n *\n * The official task set (private GT, server-side graded) isn't released yet, so\n * tasks here are seeded from the repo's dimension-named geometry fixtures (real\n * GT STEPs scored by the real scorer). When CADGENBENCH_DATA_DIR is set, swap\n * loadTasks to read the published fixtures' description.yaml + ground_truth.step.\n *\n * Requires the CADGenBench venv (CADGENBENCH_VENV) + clone (CADGENBENCH_DIR) +\n * xvfb (the scorer's alignment renders need a display).\n */\n\nimport { execFile } from 'node:child_process'\nimport { mkdtemp, rm, writeFile } from 'node:fs/promises'\nimport { tmpdir } from 'node:os'\nimport { join } from 'node:path'\nimport { promisify } from 'node:util'\nimport type { BenchScore, BenchTask, BenchmarkAdapter, LoadOptions } from './types'\nimport { CGB_DIR, CGB_VENV_PY } from '../worker-build123d'\n\nconst execFileAsync = promisify(execFile)\n\n/** Self-contained scorer wrapper (written to a temp file, run in the venv).\n *  Scores a candidate STEP against a ground-truth STEP via the benchmark's own\n *  evaluate_result, printing the cad_score line. */\nconst SCORE_PY = `\nimport sys, json, tempfile, shutil\nfrom pathlib import Path\nfrom cadgenbench.eval.evaluate import evaluate_result\ncand, gt = Path(sys.argv[1]), Path(sys.argv[2])\nwith tempfile.TemporaryDirectory() as rd, tempfile.TemporaryDirectory() as gd:\n    rd, gd = Path(rd), Path(gd)\n    (rd / 'result.json').write_text('{}')\n    shutil.copy(gt, gd / 'ground_truth.step')\n    try:\n        evaluate_result(rd, gd, candidate_step=cand)\n        d = json.loads((rd / 'result.json').read_text())\n        print('CGB_SCORE ' + json.dumps({'cad_score': d.get('cad_score', 0.0), 'status': d.get('status', 'unknown')}))\n    except Exception as e:\n        print('CGB_SCORE ' + json.dumps({'cad_score': 0.0, 'status': 'error', 'error': str(e)[:200]}))\n`.trim()\n\ninterface CgbMeta {\n  gtStep: string\n  resolveThreshold: number\n}\n\n/** Fixture-seeded tasks (real GT STEPs from the repo, dim-named so the spec is\n *  exact). Replaced by the published dataset when CADGENBENCH_DATA_DIR is set. */\nfunction fixtureTasks(): Array<{ id: string; prompt: string; gtStep: string }> {\n  const g = join(CGB_DIR, 'tests/fixtures/geometry')\n  return [\n    { id: 'box-10x20x30', prompt: 'A rectangular solid box, 10 units wide (X), 20 units deep (Y), and 30 units tall (Z).', gtStep: join(g, 'box_10_20_30.step') },\n    { id: 'cube-10', prompt: 'A cube, 10 units on every side.', gtStep: join(g, 'box_10_10_10.step') },\n    { id: 'sphere-10', prompt: 'A sphere of radius 10 units, centered at the origin.', gtStep: join(g, 'sphere_10.step') },\n  ]\n}\n\nexport function createCadGenBenchAdapter(): BenchmarkAdapter {\n  return {\n    name: 'cadgenbench',\n\n    async preflight() {\n      const r = await execFileAsync(CGB_VENV_PY, ['-c', 'import cadgenbench.eval.evaluate, build123d, trimesh, manifold3d; print(\"ok\")'], { timeout: 60_000 }).catch(\n        (e) => ({ stdout: '', stderr: e instanceof Error ? e.message : String(e) }),\n      )\n      if (!/ok/.test(r.stdout)) {\n        throw new Error(\n          `cadgenbench preflight failed (venv=${CGB_VENV_PY}): ${r.stderr.slice(0, 200)}\\n` +\n            `Fix: git clone https://github.com/huggingface/cadgenbench ${CGB_DIR}; python3 -m venv $CADGENBENCH_VENV; $CADGENBENCH_VENV/bin/pip install -e ${CGB_DIR}`,\n        )\n      }\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      // CGB_HARD_DIR (a dir with tasks.json = [{id,prompt,gtStep}]) overrides the\n      // trivial fixture primitives with hard multi-feature parts (real headroom).\n      let tasks = fixtureTasks()\n      const hard = process.env.CGB_HARD_DIR\n      if (hard) {\n        const { readFile } = await import('node:fs/promises')\n        tasks = JSON.parse(await readFile(join(hard, 'tasks.json'), 'utf8')) as Array<{ id: string; prompt: string; gtStep: string }>\n      }\n      if (opts.ids) tasks = tasks.filter((t) => opts.ids!.includes(t.id))\n      if (opts.limit != null) tasks = tasks.slice(0, opts.limit)\n      const meta = (gtStep: string): CgbMeta => ({ gtStep, resolveThreshold: Number(process.env.CGB_RESOLVE_THRESHOLD ?? 0.9) })\n      return tasks.map((t): BenchTask => ({ id: t.id, prompt: t.prompt, metadata: meta(t.gtStep) as unknown as Record<string, unknown> }))\n    },\n\n    async goldArtifact() {\n      return undefined // GT is a STEP file scored by the kernel, not a returnable artifact\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const { gtStep, resolveThreshold } = task.metadata as unknown as CgbMeta\n      if (!artifact.includes('ISO-10303-21')) return { resolved: false, score: 0, detail: 'artifact is not a STEP file' }\n      const dir = await mkdtemp(join(tmpdir(), 'cgb-judge-'))\n      const cand = join(dir, 'candidate.step')\n      const scorer = join(dir, 'score.py')\n      try {\n        await writeFile(cand, artifact)\n        await writeFile(scorer, SCORE_PY)\n        // xvfb: the scorer's alignment step renders; needs a display.\n        const r = await execFileAsync('xvfb-run', ['-a', CGB_VENV_PY, scorer, cand, gtStep], { maxBuffer: 1 << 26, timeout: 180_000 }).catch(\n          (e) => ({ stdout: (e as { stdout?: string }).stdout ?? '', stderr: e instanceof Error ? e.message : String(e) }),\n        )\n        const m = /CGB_SCORE (\\{.*\\})/.exec(r.stdout)\n        if (!m) return { resolved: false, score: 0, detail: `scorer produced no verdict: ${(r.stderr || r.stdout).slice(0, 160)}` }\n        const v = JSON.parse(m[1]) as { cad_score: number; status: string; error?: string }\n        const score = typeof v.cad_score === 'number' ? v.cad_score : 0\n        return { resolved: score >= resolveThreshold, score, detail: `cad_score=${score.toFixed(3)} status=${v.status}${v.error ? ` (${v.error})` : ''}` }\n      } finally {\n        await rm(dir, { recursive: true, force: true }).catch(() => {})\n      }\n    },\n  }\n}\n","/**\n * CADGenBench worker. The deliverable is a STEP B-rep solid (output.step). We\n * author a build123d (Python on the OpenCascade kernel) script via the router,\n * execute it in the CADGenBench venv, and read back the produced output.step —\n * exactly the reference baseline's contract. The artifact returned IS the STEP\n * text, which the CADGenBench geometric scorer grades against the ground truth.\n *\n * The build123d authoring directive is the GEPA-optimizable surface; the\n * build123d API cheat sheet (shipped in the cadgenbench package) is appended as\n * fixed reference context. Requires the CADGenBench venv (CADGENBENCH_VENV) +\n * its clone (CADGENBENCH_DIR for the cheat sheet).\n */\n\nimport { execFile } from 'node:child_process'\nimport { existsSync, readFileSync } from 'node:fs'\nimport { mkdtemp, readFile, rm, writeFile } from 'node:fs/promises'\nimport { tmpdir } from 'node:os'\nimport { join } from 'node:path'\nimport { promisify } from 'node:util'\nimport type { Span } from '@tangle-network/agent-eval'\nimport type { BenchTask } from './benchmarks/types'\nimport { DEFAULT_BUILD123D_DIRECTIVE } from './directives'\nimport { runRefineLoop } from './refine-loop'\nimport { routerChatWithUsage } from '@tangle-network/agent-runtime/loops'\n\nexport { DEFAULT_BUILD123D_DIRECTIVE } from './directives'\n\nconst execFileAsync = promisify(execFile)\n\nexport const CGB_VENV_PY = process.env.CADGENBENCH_VENV ?? '/tmp/cgb-venv/bin/python'\nexport const CGB_DIR = process.env.CADGENBENCH_DIR ?? '/tmp/cadgenbench'\n\nasync function runLocal(cmd: string, args: string[], cwd: string, timeoutMs = 120_000): Promise<{ code: number; stdout: string; stderr: string }> {\n  try {\n    const { stdout, stderr } = await execFileAsync(cmd, args, { cwd, maxBuffer: 1 << 26, timeout: timeoutMs })\n    return { code: 0, stdout, stderr }\n  } catch (err) {\n    const e = err as { code?: number; stdout?: string; stderr?: string; message?: string }\n    return { code: typeof e.code === 'number' ? e.code : 1, stdout: e.stdout ?? '', stderr: e.stderr ?? e.message ?? String(err) }\n  }\n}\n\nfunction extractPy(text: string): string {\n  const fence = /```(?:python|py)?\\s*\\n([\\s\\S]*?)```/i.exec(text)\n  return (fence ? fence[1] : text).trim()\n}\n\nlet _cheat: string | null = null\nfunction cheatSheet(): string {\n  if (_cheat != null) return _cheat\n  const p = join(CGB_DIR, 'src/cadgenbench/baseline/build123d_cheat_sheet.md')\n  _cheat = existsSync(p) ? readFileSync(p, 'utf8').slice(0, 12000) : ''\n  return _cheat\n}\n\nexport interface Build123dConfig {\n  routerBaseUrl: string\n  routerKey: string\n  model: string\n  rounds?: number\n  /** The build123d authoring directive — the GEPA-optimizable surface. */\n  directive?: string\n}\n\nexport interface Build123dShot {\n  /** The produced STEP text (the artifact the CADGenBench scorer grades). */\n  artifact: string\n  /** The Python source the agent wrote. */\n  source: string\n  trace: Span[]\n  usage: { input: number; output: number }\n  ok: boolean\n  built: boolean\n  detail?: string\n}\n\n/** Author a build123d script via the router, execute it in the CADGenBench venv,\n *  read back output.step. Refine on execution error / missing STEP. */\nexport async function solveBuild123dLocal(task: BenchTask, cfg: Build123dConfig): Promise<Build123dShot> {\n  const rounds = Math.max(1, cfg.rounds ?? 2)\n  const directive = cfg.directive ?? DEFAULT_BUILD123D_DIRECTIVE\n  const sys = `${directive}\\n\\nbuild123d API reference:\\n${cheatSheet()}`\n  const trace: Span[] = []\n  const runId = `cadgenbench-${task.id}`\n  let ts = Date.now()\n  const tick = () => (ts += 1)\n  const usage = { input: 0, output: 0 }\n  // Carried across rounds in closures (the round Artifact is the Python source; the\n  // STEP text + built flag + lastErr persist outside the loop). usage is REAL.\n  let step = ''\n  let built = false\n  let lastErr = ''\n\n  trace.push({ spanId: 's-brief', runId, kind: 'llm', name: 'brief', model: cfg.model, messages: [{ role: 'user', content: task.prompt }], startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)\n\n  // Migrated onto runRefineLoop: the mkdtemp scratch dir is the Ctx; built (STEP\n  // produced) is the early-stop, modeled as a judge so default-decide stops the\n  // loop. The round-2+ steer carries lastErr + the prior source verbatim.\n  const res = await runRefineLoop<string, string>({\n    rounds,\n    setup: () => mkdtemp(join(tmpdir(), 'b123d-')),\n    prompt: (round, history) =>\n      round === 1\n        ? task.prompt\n        : `Your previous build123d script failed:\\n${lastErr}\\n\\nPrevious script:\\n${history[history.length - 1]?.artifact ?? ''}\\n\\nFix it so it runs in python and writes a valid output.step. Brief:\\n${task.prompt}`,\n    runShot: async (user, round, dir) => {\n      const scriptPath = join(dir, 'build.py')\n      const stepPath = join(dir, 'output.step')\n      const { content, usage: u } = await routerChatWithUsage(cfg, [\n        { role: 'system', content: sys },\n        { role: 'user', content: user },\n      ])\n      if (u) {\n        usage.input += u.input\n        usage.output += u.output\n      }\n      const source = extractPy(content)\n      trace.push({ spanId: `s-author-${round}`, runId, kind: 'llm', name: `author r${round}`, model: cfg.model, messages: [{ role: 'user', content: round === 1 ? task.prompt : 'refine' }], output: content.slice(0, 600), startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)\n      trace.push({ spanId: `s-write-${round}`, runId, kind: 'tool', name: 'write_file', toolName: 'create_file', args: { path: 'build.py', content: source }, startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)\n\n      await writeFile(scriptPath, source)\n      const run = await runLocal(CGB_VENV_PY, [scriptPath], dir)\n      const got = existsSync(stepPath) ? await readFile(stepPath, 'utf8').catch(() => '') : ''\n      built = got.includes('ISO-10303-21') && got.length > 200\n      lastErr = built ? '' : `${run.stdout}\\n${run.stderr}`.trim().slice(-800) || 'no output.step written'\n      if (built) step = got\n      trace.push({ spanId: `s-exec-${round}`, runId, kind: 'tool', name: `build123d r${round}`, toolName: 'shell.exec', args: 'python build.py', result: (built ? 'wrote output.step' : lastErr).slice(0, 1500), startedAt: tick(), endedAt: tick(), status: built ? 'ok' : 'error', error: built ? undefined : `exit ${run.code}` } as Span)\n      return { artifact: source }\n    },\n    judge: async () => ({ valid: built }),\n    teardown: (dir) => rm(dir, { recursive: true, force: true }).then(() => {}, () => {}),\n  })\n\n  return {\n    artifact: step,\n    source: res.final.artifact,\n    trace,\n    usage,\n    ok: res.final.artifact.trim().length > 0,\n    built,\n    detail: built ? 'exported output.step' : `did not produce a STEP in ${rounds} rounds${lastErr ? `; last: ${lastErr.slice(0, 140)}` : ''}`,\n  }\n}\n"],"mappings":";AAgBA,SAAS,YAAAA,iBAAgB;AACzB,SAAS,WAAAC,UAAS,MAAAC,KAAI,aAAAC,kBAAiB;AACvC,SAAS,UAAAC,eAAc;AACvB,SAAS,QAAAC,aAAY;AACrB,SAAS,aAAAC,kBAAiB;;;ACP1B,SAAS,gBAAgB;AACzB,SAAS,YAAY,oBAAoB;AACzC,SAAS,SAAS,UAAU,IAAI,iBAAiB;AACjD,SAAS,cAAc;AACvB,SAAS,YAAY;AACrB,SAAS,iBAAiB;AAK1B,SAAS,2BAA2B;AAIpC,IAAM,gBAAgB,UAAU,QAAQ;AAEjC,IAAM,cAAc,QAAQ,IAAI,oBAAoB;AACpD,IAAM,UAAU,QAAQ,IAAI,mBAAmB;;;ADNtD,IAAMC,iBAAgBC,WAAUC,SAAQ;AAKxC,IAAM,WAAW;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAef,KAAK;AASP,SAAS,eAAsE;AAC7E,QAAM,IAAIC,MAAK,SAAS,yBAAyB;AACjD,SAAO;AAAA,IACL,EAAE,IAAI,gBAAgB,QAAQ,yFAAyF,QAAQA,MAAK,GAAG,mBAAmB,EAAE;AAAA,IAC5J,EAAE,IAAI,WAAW,QAAQ,mCAAmC,QAAQA,MAAK,GAAG,mBAAmB,EAAE;AAAA,IACjG,EAAE,IAAI,aAAa,QAAQ,wDAAwD,QAAQA,MAAK,GAAG,gBAAgB,EAAE;AAAA,EACvH;AACF;AAEO,SAAS,2BAA6C;AAC3D,SAAO;AAAA,IACL,MAAM;AAAA,IAEN,MAAM,YAAY;AAChB,YAAM,IAAI,MAAMH,eAAc,aAAa,CAAC,MAAM,+EAA+E,GAAG,EAAE,SAAS,IAAO,CAAC,EAAE;AAAA,QACvJ,CAAC,OAAO,EAAE,QAAQ,IAAI,QAAQ,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,EAAE;AAAA,MAC3E;AACA,UAAI,CAAC,KAAK,KAAK,EAAE,MAAM,GAAG;AACxB,cAAM,IAAI;AAAA,UACR,sCAAsC,WAAW,MAAM,EAAE,OAAO,MAAM,GAAG,GAAG,CAAC;AAAA,4DACd,OAAO,6EAA6E,OAAO;AAAA,QAC5J;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;AAGtC,UAAI,QAAQ,aAAa;AACzB,YAAM,OAAO,QAAQ,IAAI;AACzB,UAAI,MAAM;AACR,cAAM,EAAE,UAAAI,UAAS,IAAI,MAAM,OAAO,aAAkB;AACpD,gBAAQ,KAAK,MAAM,MAAMA,UAASD,MAAK,MAAM,YAAY,GAAG,MAAM,CAAC;AAAA,MACrE;AACA,UAAI,KAAK,IAAK,SAAQ,MAAM,OAAO,CAAC,MAAM,KAAK,IAAK,SAAS,EAAE,EAAE,CAAC;AAClE,UAAI,KAAK,SAAS,KAAM,SAAQ,MAAM,MAAM,GAAG,KAAK,KAAK;AACzD,YAAM,OAAO,CAAC,YAA6B,EAAE,QAAQ,kBAAkB,OAAO,QAAQ,IAAI,yBAAyB,GAAG,EAAE;AACxH,aAAO,MAAM,IAAI,CAAC,OAAkB,EAAE,IAAI,EAAE,IAAI,QAAQ,EAAE,QAAQ,UAAU,KAAK,EAAE,MAAM,EAAwC,EAAE;AAAA,IACrI;AAAA,IAEA,MAAM,eAAe;AACnB,aAAO;AAAA,IACT;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,EAAE,QAAQ,iBAAiB,IAAI,KAAK;AAC1C,UAAI,CAAC,SAAS,SAAS,cAAc,EAAG,QAAO,EAAE,UAAU,OAAO,OAAO,GAAG,QAAQ,8BAA8B;AAClH,YAAM,MAAM,MAAME,SAAQF,MAAKG,QAAO,GAAG,YAAY,CAAC;AACtD,YAAM,OAAOH,MAAK,KAAK,gBAAgB;AACvC,YAAM,SAASA,MAAK,KAAK,UAAU;AACnC,UAAI;AACF,cAAMI,WAAU,MAAM,QAAQ;AAC9B,cAAMA,WAAU,QAAQ,QAAQ;AAEhC,cAAM,IAAI,MAAMP,eAAc,YAAY,CAAC,MAAM,aAAa,QAAQ,MAAM,MAAM,GAAG,EAAE,WAAW,KAAK,IAAI,SAAS,KAAQ,CAAC,EAAE;AAAA,UAC7H,CAAC,OAAO,EAAE,QAAS,EAA0B,UAAU,IAAI,QAAQ,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,EAAE;AAAA,QAChH;AACA,cAAM,IAAI,qBAAqB,KAAK,EAAE,MAAM;AAC5C,YAAI,CAAC,EAAG,QAAO,EAAE,UAAU,OAAO,OAAO,GAAG,QAAQ,gCAAgC,EAAE,UAAU,EAAE,QAAQ,MAAM,GAAG,GAAG,CAAC,GAAG;AAC1H,cAAM,IAAI,KAAK,MAAM,EAAE,CAAC,CAAC;AACzB,cAAM,QAAQ,OAAO,EAAE,cAAc,WAAW,EAAE,YAAY;AAC9D,eAAO,EAAE,UAAU,SAAS,kBAAkB,OAAO,QAAQ,aAAa,MAAM,QAAQ,CAAC,CAAC,WAAW,EAAE,MAAM,GAAG,EAAE,QAAQ,KAAK,EAAE,KAAK,MAAM,EAAE,GAAG;AAAA,MACnJ,UAAE;AACA,cAAMQ,IAAG,KAAK,EAAE,WAAW,MAAM,OAAO,KAAK,CAAC,EAAE,MAAM,MAAM;AAAA,QAAC,CAAC;AAAA,MAChE;AAAA,IACF;AAAA,EACF;AACF;","names":["execFile","mkdtemp","rm","writeFile","tmpdir","join","promisify","execFileAsync","promisify","execFile","join","readFile","mkdtemp","tmpdir","writeFile","rm"]}