@tangle-network/agent-bench 0.4.0 → 0.4.3

package/dist/chunk-TBKU5XQI.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/benchmarks/mind2web.ts"],"sourcesContent":["/**\n * Mind2Web adapter (osunlp/Multimodal-Mind2Web, split 'test_domain' by default).\n * Web-agent ACTION PREDICTION: each task is one step of a real web task — given\n * the natural-language goal and the page's candidate elements, the worker picks\n * the SINGLE element to act on and the action (CLICK / TYPE / SELECT [+ value]).\n *\n * Judge is the published Mind2Web step metric, FULLY DETERMINISTIC — no LLM:\n *   element correct  ⇔ predicted backend_node_id ∈ pos_candidates ids\n *   operation correct ⇔ op-type matches AND (TYPE/SELECT) the value matches\n *   resolved (Step-SR) ⇔ element correct AND operation correct\n *   score = 0.6·element + 0.4·operation  (a gradient for the optimizer; the right\n *           element is most of the credit, mirroring Element-Acc ≫ Op as the lever)\n * This is the low-noise reward a certifiable directive-lift needs: the number is a\n * programmatic match against human-verified ground truth, not a judge's opinion.\n *\n * The worker artifact is three sentinel lines (ELEMENT / ACTION / VALUE) so the\n * judge extracts deterministically; the prompt presents the candidate set as a\n * choice over backend_node_id ordered by id (position uncorrelated with the answer).\n *\n * Each step carries the dataset's OWN page screenshot (written to a temp file by\n * the loader); the worker drops it into a browser.<op> span so run-capsule's screen\n * capsule turns the run into a film — the real page, not a re-rendered DOM.\n *\n * Requires for a live run: a python with `datasets` + `pillow` and network to\n * Hugging Face. Point M2W_PYTHON at it (defaults to bench/.venv/bin/python).\n */\n\nimport { execFile } from 'node:child_process'\nimport { mkdir } from 'node:fs/promises'\nimport { tmpdir } from 'node:os'\nimport { join } from 'node:path'\nimport { fileURLToPath } from 'node:url'\nimport { promisify } from 'node:util'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\n\nconst execFileAsync = promisify(execFile)\nconst BENCH_ROOT = fileURLToPath(new URL('../..', import.meta.url))\nconst PY = process.env.M2W_PYTHON ?? join(BENCH_ROOT, '.venv', 'bin', 'python')\n\nconst DATASET = 'osunlp/Multimodal-Mind2Web'\nconst DEFAULT_SPLIT = process.env.M2W_SPLIT ?? 'test_domain'\nconst SHOTS_DIR = process.env.M2W_SHOTS ?? join(tmpdir(), 'm2w-shots')\n/** Candidate-set cap presented to the worker (always keeps all pos candidates). */\nconst CANDIDATE_CAP = Number(process.env.M2W_CANDIDATE_CAP ?? 30)\n\n/** The worker contract appended to every task prompt; the judge keys off these. */\nconst WORKER_CONTRACT = [\n  '',\n  'Pick the SINGLE next element to act on, then end your response with EXACTLY these three lines:',\n  'ELEMENT: <the [id] number of the chosen element>',\n  'ACTION: <CLICK | TYPE | SELECT>',\n  'VALUE: <text to type or option to select; leave empty for CLICK>',\n].join('\\n')\n\ninterface RawCandidate {\n  id: string\n  label: string\n}\ninterface Mind2WebRow {\n  id: string\n  task: string\n  website: string\n  domain: string\n  subdomain: string\n  op: string\n  value: string\n  goldIds: string[]\n  candidates: RawCandidate[]\n  screenshotPath: string\n  targetRepr: string\n}\ninterface Mind2WebMeta {\n  task: string\n  website: string\n  domain: string\n  op: string\n  value: string\n  goldIds: string[]\n  candidateIds: string[]\n  screenshotPath: string\n  targetRepr: string\n}\n\n/** Run the bench python with a script on stdin; return stdout (throws on nonzero). */\nasync function py(script: string, args: string[] = []): Promise<string> {\n  const { stdout } = await execFileAsync(PY, ['-c', script, ...args], { maxBuffer: 1024 * 1024 * 256 })\n  return stdout\n}\n\n/** The loader: stream the split, skip steps with no positive candidate (fail-loud,\n *  never a silent score-0), cap the candidate set, save each screenshot to a file. */\nconst LOADER = `\nimport json, sys, os\nfrom datasets import load_dataset\ncfg = json.loads(sys.argv[1])\nsplit = cfg[\"split\"]; limit = cfg.get(\"limit\"); cap = cfg.get(\"cap\", 30)\nshots = cfg[\"shotsDir\"]; ids = set(cfg[\"ids\"]) if cfg.get(\"ids\") else None\nos.makedirs(shots, exist_ok=True)\nKEEP = (\"aria_label\",\"aria-label\",\"role\",\"type\",\"name\",\"placeholder\",\"title\",\"alt\",\"value\",\"text\",\"id\",\"class\",\"href\")\ndef label(tag, attr):\n    try:\n        a = json.loads(attr) if isinstance(attr, str) else (attr or {})\n    except Exception:\n        a = {}\n    parts = []\n    for k in KEEP:\n        v = a.get(k)\n        if v:\n            sv = str(v).replace(\"\\\\n\", \" \").strip()\n            if k == \"class\": sv = sv[:40]\n            if k == \"href\": sv = sv[:50]\n            if sv: parts.append(k + \"=\" + sv[:60])\n    return \"<\" + str(tag) + \"> \" + \" \".join(parts[:6])\ndef cands(raw):\n    out = []\n    for c in raw or []:\n        try:\n            d = json.loads(c) if isinstance(c, str) else c\n        except Exception:\n            continue\n        bid = str(d.get(\"backend_node_id\", \"\"))\n        if not bid: continue\n        out.append({\"id\": bid, \"label\": label(d.get(\"tag\", \"?\"), d.get(\"attributes\"))})\n    return out\nds = load_dataset(${JSON.stringify(DATASET)}, split=split, streaming=True)\nemitted = []; skipped = 0; scanned = 0\nfor r in ds:\n    scanned += 1\n    if ids is not None and r[\"action_uid\"] not in ids:\n        if scanned > 8000: break\n        continue\n    pos = cands(r.get(\"pos_candidates\"))\n    if not pos:\n        skipped += 1\n        continue\n    neg = cands(r.get(\"neg_candidates\"))\n    seen = set(p[\"id\"] for p in pos); merged = list(pos)\n    for n in neg:\n        if len(merged) >= cap: break\n        if n[\"id\"] in seen: continue\n        seen.add(n[\"id\"]); merged.append(n)\n    merged.sort(key=lambda x: int(x[\"id\"]) if x[\"id\"].isdigit() else 0)\n    try:\n        op = json.loads(r[\"operation\"]) if isinstance(r[\"operation\"], str) else r[\"operation\"]\n    except Exception:\n        op = {}\n    sp = os.path.join(shots, str(r[\"action_uid\"]) + \".jpg\")\n    try:\n        r[\"screenshot\"].convert(\"RGB\").save(sp, \"JPEG\", quality=70)\n    except Exception:\n        sp = \"\"\n    emitted.append({\n        \"id\": r[\"action_uid\"], \"task\": r.get(\"confirmed_task\", \"\"),\n        \"website\": r.get(\"website\", \"\"), \"domain\": r.get(\"domain\", \"\"), \"subdomain\": r.get(\"subdomain\", \"\"),\n        \"op\": str(op.get(\"op\", \"\")).upper(), \"value\": str(op.get(\"value\", \"\")),\n        \"goldIds\": [p[\"id\"] for p in pos], \"candidates\": merged,\n        \"screenshotPath\": sp, \"targetRepr\": r.get(\"target_action_reprs\", \"\"),\n    })\n    if ids is None and limit is not None and len(emitted) >= limit: break\n    if ids is not None and len(emitted) >= len(ids): break\nsys.stderr.write(\"[mind2web] emitted=%d skipped_empty_pos=%d scanned=%d\\\\n\" % (len(emitted), skipped, scanned)); sys.stderr.flush()\nprint(json.dumps(emitted)); sys.stdout.flush()\nos._exit(0)\n`\n\nfunction buildPrompt(row: Mind2WebRow): string {\n  const choices = row.candidates.map((c) => `  [${c.id}] ${c.label}`).join('\\n')\n  return [\n    `Web task: ${row.task}`,\n    '',\n    'You are taking the NEXT single action on the current web page. Choose the one element',\n    'to act on from the candidates below (each line is \"[id] <tag> attributes\"):',\n    choices,\n    WORKER_CONTRACT,\n  ].join('\\n')\n}\n\nfunction rowToTask(row: Mind2WebRow): BenchTask {\n  const meta: Mind2WebMeta = {\n    task: row.task,\n    website: row.website,\n    domain: row.domain,\n    op: row.op,\n    value: row.value,\n    goldIds: row.goldIds,\n    candidateIds: row.candidates.map((c) => c.id),\n    screenshotPath: row.screenshotPath,\n    targetRepr: row.targetRepr,\n  }\n  return {\n    id: `mind2web-${row.id}`,\n    split: DEFAULT_SPLIT,\n    prompt: buildPrompt(row),\n    metadata: meta as unknown as Record<string, unknown>,\n  }\n}\n\nfunction readMeta(task: BenchTask): Mind2WebMeta {\n  const md = task.metadata\n  if (!md || !Array.isArray((md as { goldIds?: unknown }).goldIds)) {\n    throw new Error(`mind2web task ${task.id} missing metadata.goldIds — loadTasks did not populate it`)\n  }\n  return md as unknown as Mind2WebMeta\n}\n\n/** Normalize a TYPE/SELECT value for comparison: lowercase, collapse whitespace. */\nfunction normValue(s: string): string {\n  return s.toLowerCase().replace(/\\s+/g, ' ').trim()\n}\n\ninterface ParsedAction {\n  elementId: string\n  op: string\n  value: string\n}\n\n/** Parse the worker artifact's ELEMENT / ACTION / VALUE sentinel lines.\n *  Fail-closed: a missing ELEMENT or ACTION returns null (judge → resolved=false). */\nexport function parseAction(artifact: string): ParsedAction | null {\n  const elem = /ELEMENT:\\s*\\[?(\\d+)\\]?/i.exec(artifact)\n  const op = /ACTION:\\s*(CLICK|TYPE|SELECT)/i.exec(artifact)\n  if (!elem?.[1] || !op?.[1]) return null\n  const val = /VALUE:\\s*(.*)/i.exec(artifact)\n  return { elementId: elem[1], op: op[1].toUpperCase(), value: (val?.[1] ?? '').trim() }\n}\n\nexport function createMind2WebAdapter(): BenchmarkAdapter {\n  const split = DEFAULT_SPLIT\n\n  return {\n    name: 'mind2web',\n\n    async preflight() {\n      try {\n        await py(\n          `import os, sys\nfrom datasets import load_dataset\nimport PIL\nds = load_dataset(${JSON.stringify(DATASET)}, split=${JSON.stringify(split)}, streaming=True)\nnext(iter(ds))\nprint('ok'); sys.stdout.flush(); os._exit(0)`,\n        )\n      } catch (err) {\n        const msg = err instanceof Error ? err.message : String(err)\n        throw new Error(\n          `mind2web preflight failed: ${msg}\\n` +\n            `Fix: (1) a python with datasets + pillow (python3 -m venv bench/.venv && bench/.venv/bin/pip install datasets pillow) ; ` +\n            `(2) network access to Hugging Face for ${DATASET} (split ${split}) ; ` +\n            `point M2W_PYTHON at the python if not bench/.venv.`,\n        )\n      }\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      const cfg = {\n        split: opts.split ?? split,\n        limit: opts.ids ? undefined : (opts.limit ?? 10),\n        cap: CANDIDATE_CAP,\n        shotsDir: SHOTS_DIR,\n        ids: opts.ids ? opts.ids.map((id) => id.replace(/^mind2web-/, '')) : undefined,\n      }\n      await mkdir(SHOTS_DIR, { recursive: true })\n      const stdout = await py(LOADER, [JSON.stringify(cfg)])\n      const rows = JSON.parse(stdout) as Mind2WebRow[]\n      return rows.map(rowToTask)\n    },\n\n    async goldArtifact(task: BenchTask) {\n      // Gold = the worker-contract serialization of the ground-truth action, so\n      // verify-judge proves gold→resolved through the SAME parse path the real\n      // artifact takes. The first accepted positive id is the canonical target.\n      const meta = readMeta(task)\n      const id = meta.goldIds[0]\n      if (!id) return undefined\n      return `ELEMENT: ${id}\\nACTION: ${meta.op}\\nVALUE: ${meta.value}`\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const meta = readMeta(task)\n      const parsed = parseAction(artifact)\n      if (!parsed) {\n        return {\n          resolved: false,\n          score: 0,\n          detail: JSON.stringify({ reason: 'no parseable ELEMENT/ACTION', goldOp: meta.op, goldIds: meta.goldIds }),\n        }\n      }\n      const elementCorrect = meta.goldIds.includes(parsed.elementId)\n      const opMatch = parsed.op === meta.op\n      const valueMatch = meta.op === 'CLICK' ? true : normValue(parsed.value) === normValue(meta.value)\n      const operationCorrect = opMatch && valueMatch\n      const resolved = elementCorrect && operationCorrect\n      const score = 0.6 * (elementCorrect ? 1 : 0) + 0.4 * (operationCorrect ? 1 : 0)\n      return {\n        resolved,\n        score,\n        detail: JSON.stringify({\n          elementCorrect,\n          opMatch,\n          valueMatch,\n          predicted: parsed,\n          goldOp: meta.op,\n          goldValue: meta.value,\n          goldIds: meta.goldIds,\n          website: meta.website,\n          domain: meta.domain,\n        }),\n      }\n    },\n  }\n}\n"],"mappings":";AA2BA,SAAS,gBAAgB;AACzB,SAAS,aAAa;AACtB,SAAS,cAAc;AACvB,SAAS,YAAY;AACrB,SAAS,qBAAqB;AAC9B,SAAS,iBAAiB;AAG1B,IAAM,gBAAgB,UAAU,QAAQ;AACxC,IAAM,aAAa,cAAc,IAAI,IAAI,SAAS,YAAY,GAAG,CAAC;AAClE,IAAM,KAAK,QAAQ,IAAI,cAAc,KAAK,YAAY,SAAS,OAAO,QAAQ;AAE9E,IAAM,UAAU;AAChB,IAAM,gBAAgB,QAAQ,IAAI,aAAa;AAC/C,IAAM,YAAY,QAAQ,IAAI,aAAa,KAAK,OAAO,GAAG,WAAW;AAErE,IAAM,gBAAgB,OAAO,QAAQ,IAAI,qBAAqB,EAAE;AAGhE,IAAM,kBAAkB;AAAA,EACtB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,EAAE,KAAK,IAAI;AAgCX,eAAe,GAAG,QAAgB,OAAiB,CAAC,GAAoB;AACtE,QAAM,EAAE,OAAO,IAAI,MAAM,cAAc,IAAI,CAAC,MAAM,QAAQ,GAAG,IAAI,GAAG,EAAE,WAAW,OAAO,OAAO,IAAI,CAAC;AACpG,SAAO;AACT;AAIA,IAAM,SAAS;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,oBAiCK,KAAK,UAAU,OAAO,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAyC3C,SAAS,YAAY,KAA0B;AAC7C,QAAM,UAAU,IAAI,WAAW,IAAI,CAAC,MAAM,MAAM,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,EAAE,KAAK,IAAI;AAC7E,SAAO;AAAA,IACL,aAAa,IAAI,IAAI;AAAA,IACrB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AACb;AAEA,SAAS,UAAU,KAA6B;AAC9C,QAAM,OAAqB;AAAA,IACzB,MAAM,IAAI;AAAA,IACV,SAAS,IAAI;AAAA,IACb,QAAQ,IAAI;AAAA,IACZ,IAAI,IAAI;AAAA,IACR,OAAO,IAAI;AAAA,IACX,SAAS,IAAI;AAAA,IACb,cAAc,IAAI,WAAW,IAAI,CAAC,MAAM,EAAE,EAAE;AAAA,IAC5C,gBAAgB,IAAI;AAAA,IACpB,YAAY,IAAI;AAAA,EAClB;AACA,SAAO;AAAA,IACL,IAAI,YAAY,IAAI,EAAE;AAAA,IACtB,OAAO;AAAA,IACP,QAAQ,YAAY,GAAG;AAAA,IACvB,UAAU;AAAA,EACZ;AACF;AAEA,SAAS,SAAS,MAA+B;AAC/C,QAAM,KAAK,KAAK;AAChB,MAAI,CAAC,MAAM,CAAC,MAAM,QAAS,GAA6B,OAAO,GAAG;AAChE,UAAM,IAAI,MAAM,iBAAiB,KAAK,EAAE,gEAA2D;AAAA,EACrG;AACA,SAAO;AACT;AAGA,SAAS,UAAU,GAAmB;AACpC,SAAO,EAAE,YAAY,EAAE,QAAQ,QAAQ,GAAG,EAAE,KAAK;AACnD;AAUO,SAAS,YAAY,UAAuC;AACjE,QAAM,OAAO,0BAA0B,KAAK,QAAQ;AACpD,QAAM,KAAK,iCAAiC,KAAK,QAAQ;AACzD,MAAI,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC,EAAG,QAAO;AACnC,QAAM,MAAM,iBAAiB,KAAK,QAAQ;AAC1C,SAAO,EAAE,WAAW,KAAK,CAAC,GAAG,IAAI,GAAG,CAAC,EAAE,YAAY,GAAG,QAAQ,MAAM,CAAC,KAAK,IAAI,KAAK,EAAE;AACvF;AAEO,SAAS,wBAA0C;AACxD,QAAM,QAAQ;AAEd,SAAO;AAAA,IACL,MAAM;AAAA,IAEN,MAAM,YAAY;AAChB,UAAI;AACF,cAAM;AAAA,UACJ;AAAA;AAAA;AAAA,oBAGU,KAAK,UAAU,OAAO,CAAC,WAAW,KAAK,UAAU,KAAK,CAAC;AAAA;AAAA;AAAA,QAGnE;AAAA,MACF,SAAS,KAAK;AACZ,cAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC3D,cAAM,IAAI;AAAA,UACR,8BAA8B,GAAG;AAAA,iKAEW,OAAO,WAAW,KAAK;AAAA,QAErE;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;AACtC,YAAM,MAAM;AAAA,QACV,OAAO,KAAK,SAAS;AAAA,QACrB,OAAO,KAAK,MAAM,SAAa,KAAK,SAAS;AAAA,QAC7C,KAAK;AAAA,QACL,UAAU;AAAA,QACV,KAAK,KAAK,MAAM,KAAK,IAAI,IAAI,CAAC,OAAO,GAAG,QAAQ,cAAc,EAAE,CAAC,IAAI;AAAA,MACvE;AACA,YAAM,MAAM,WAAW,EAAE,WAAW,KAAK,CAAC;AAC1C,YAAM,SAAS,MAAM,GAAG,QAAQ,CAAC,KAAK,UAAU,GAAG,CAAC,CAAC;AACrD,YAAM,OAAO,KAAK,MAAM,MAAM;AAC9B,aAAO,KAAK,IAAI,SAAS;AAAA,IAC3B;AAAA,IAEA,MAAM,aAAa,MAAiB;AAIlC,YAAM,OAAO,SAAS,IAAI;AAC1B,YAAM,KAAK,KAAK,QAAQ,CAAC;AACzB,UAAI,CAAC,GAAI,QAAO;AAChB,aAAO,YAAY,EAAE;AAAA,UAAa,KAAK,EAAE;AAAA,SAAY,KAAK,KAAK;AAAA,IACjE;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,OAAO,SAAS,IAAI;AAC1B,YAAM,SAAS,YAAY,QAAQ;AACnC,UAAI,CAAC,QAAQ;AACX,eAAO;AAAA,UACL,UAAU;AAAA,UACV,OAAO;AAAA,UACP,QAAQ,KAAK,UAAU,EAAE,QAAQ,+BAA+B,QAAQ,KAAK,IAAI,SAAS,KAAK,QAAQ,CAAC;AAAA,QAC1G;AAAA,MACF;AACA,YAAM,iBAAiB,KAAK,QAAQ,SAAS,OAAO,SAAS;AAC7D,YAAM,UAAU,OAAO,OAAO,KAAK;AACnC,YAAM,aAAa,KAAK,OAAO,UAAU,OAAO,UAAU,OAAO,KAAK,MAAM,UAAU,KAAK,KAAK;AAChG,YAAM,mBAAmB,WAAW;AACpC,YAAM,WAAW,kBAAkB;AACnC,YAAM,QAAQ,OAAO,iBAAiB,IAAI,KAAK,OAAO,mBAAmB,IAAI;AAC7E,aAAO;AAAA,QACL;AAAA,QACA;AAAA,QACA,QAAQ,KAAK,UAAU;AAAA,UACrB;AAAA,UACA;AAAA,UACA;AAAA,UACA,WAAW;AAAA,UACX,QAAQ,KAAK;AAAA,UACb,WAAW,KAAK;AAAA,UAChB,SAAS,KAAK;AAAA,UACd,SAAS,KAAK;AAAA,UACd,QAAQ,KAAK;AAAA,QACf,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AACF;","names":[]}

package/dist/chunk-UPAMRDX4.js

@@ -1,233 +0,0 @@
-// src/benchmarks/finsearchcomp.ts
-import { readFile } from "fs/promises";
-import { join } from "path";
-import { fileURLToPath } from "url";
-var BENCH_ROOT = fileURLToPath(new URL("../..", import.meta.url));
-var FIXTURES = join(BENCH_ROOT, "fixtures", "finsearchcomp.json");
-var DATASET_URL = "https://raw.githubusercontent.com/randomtutu/FinSearchComp/main/data/finsearchcomp_data.json";
-var SUPPORTED_TYPES = /* @__PURE__ */ new Set(["T2", "T3"]);
-var WORKER_CONTRACT = [
-  "",
-  "Research this financial question using live web/market sources and answer it.",
-  "State your final answer explicitly, with the exact units and precision the question requests.",
-  "Respect any tolerance the question implies; show the value you settled on, not just intermediate work."
-].join("\n");
-function parseTaskType(promptId) {
-  const m = promptId.match(/^\((T[123])\)/);
-  if (!m) throw new Error(`FinSearchComp prompt_id has no (T1|T2|T3) prefix: ${JSON.stringify(promptId)}`);
-  return m[1];
-}
-function parseRegion(label) {
-  const m = label.match(/\(([^()]+)\)\s*$/);
-  return m ? m[1].trim() : "";
-}
-function recordToTask(rec) {
-  const taskType = parseTaskType(rec.prompt_id);
-  const meta = {
-    promptId: rec.prompt_id,
-    label: rec.label,
-    taskType,
-    region: parseRegion(rec.label),
-    responseReference: rec.response_reference,
-    judgePromptTemplate: rec.judge_prompt_template,
-    judgeSystemPrompt: rec.judge_system_prompt,
-    rawPrompt: rec.prompt
-  };
-  return {
-    id: rec.prompt_id,
-    split: taskType,
-    prompt: rec.prompt + WORKER_CONTRACT,
-    metadata: meta
-  };
-}
-function readMeta(task) {
-  const md = task.metadata;
-  if (!md || typeof md.responseReference !== "string" || typeof md.judgePromptTemplate !== "string" || typeof md.judgeSystemPrompt !== "string" || typeof md.rawPrompt !== "string") {
-    throw new Error(`FinSearchComp task ${task.id} missing judge metadata \u2014 loadTasks did not populate it`);
-  }
-  return md;
-}
-function fillJudgePrompt(meta, response) {
-  return meta.judgePromptTemplate.replaceAll("{prompt}", meta.rawPrompt).replaceAll("{response_reference}", meta.responseReference).replaceAll("{response}", response);
-}
-function judgeRouter() {
-  const key = process.env.TANGLE_API_KEY;
-  if (!key) throw new Error("TANGLE_API_KEY is required for the FinSearchComp per-record LLM judge");
-  const model = process.env.JUDGE_MODEL ?? "deepseek-v4-flash";
-  const baseUrl = process.env.ROUTER_BASE ?? "https://router.tangle.tools/v1";
-  return { baseUrl, key, model };
-}
-function readAnswerScore(value) {
-  let v = value;
-  while (Array.isArray(v)) v = v[0];
-  if (v === null || v === void 0 || v === "null") {
-    return { resolved: false, score: 0, raw: value };
-  }
-  const n = typeof v === "number" ? v : typeof v === "string" ? Number(v) : NaN;
-  if (!Number.isFinite(n)) {
-    throw new Error(`FinSearchComp judge answer_score not numeric: ${JSON.stringify(value)}`);
-  }
-  const resolved = n >= 1;
-  const score = n <= 0 ? 0 : n >= 1 ? 1 : n;
-  return { resolved, score, raw: value };
-}
-function parseJudgeOutput(content) {
-  const candidates = [];
-  for (const m of content.matchAll(/```(?:json)?\s*([\s\S]*?)```/g)) candidates.push(m[1].trim());
-  for (const m of content.matchAll(/\{[^{}]*"answer_score"[\s\S]*?\}/g)) candidates.push(m[0]);
-  candidates.push(content.trim());
-  for (const c of candidates) {
-    let parsed;
-    try {
-      parsed = JSON.parse(c);
-    } catch {
-      continue;
-    }
-    if ("answer_score" in parsed) return readAnswerScore(parsed.answer_score);
-  }
-  throw new Error(
-    `FinSearchComp judge produced no parseable {"answer_score": \u2026}: ${content.slice(0, 400)}`
-  );
-}
-async function runRecordJudge(meta, response, router) {
-  const res = await fetch(`${router.baseUrl}/chat/completions`, {
-    method: "POST",
-    headers: { "content-type": "application/json", authorization: `Bearer ${router.key}` },
-    body: JSON.stringify({
-      model: router.model,
-      temperature: 0,
-      messages: [
-        { role: "system", content: meta.judgeSystemPrompt },
-        { role: "user", content: fillJudgePrompt(meta, response) }
-      ]
-    })
-  });
-  if (!res.ok) {
-    throw new Error(`FinSearchComp judge HTTP ${res.status}: ${(await res.text()).slice(0, 300)}`);
-  }
-  const body = await res.json();
-  const content = body.choices?.[0]?.message?.content;
-  if (typeof content !== "string") {
-    throw new Error(`FinSearchComp judge returned no message content: ${JSON.stringify(body).slice(0, 300)}`);
-  }
-  const { resolved, score, raw } = parseJudgeOutput(content);
-  return {
-    resolved,
-    score,
-    detail: JSON.stringify({
-      taskType: meta.taskType,
-      region: meta.region,
-      answerScore: raw,
-      judgeModel: router.model
-    })
-  };
-}
-function selectRecords(records, opts) {
-  let tasks = records.filter((r) => SUPPORTED_TYPES.has(parseTaskType(r.prompt_id))).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 = balanceByType(tasks, opts.limit);
-  }
-  return tasks;
-}
-function balanceByType(tasks, limit) {
-  const queues = /* @__PURE__ */ new Map();
-  for (const t of tasks) {
-    const k = parseTaskType(t.id);
-    const q = queues.get(k);
-    if (q) q.push(t);
-    else queues.set(k, [t]);
-  }
-  const lanes = [...queues.values()];
-  const out = [];
-  for (let i = 0; out.length < limit; i++) {
-    const before = out.length;
-    for (const lane of lanes) {
-      if (i < lane.length) out.push(lane[i]);
-      if (out.length >= limit) break;
-    }
-    if (out.length === before) break;
-  }
-  return out;
-}
-async function loadFixtures(opts) {
-  const records = JSON.parse(await readFile(FIXTURES, "utf8"));
-  return selectRecords(records, opts);
-}
-async function fetchDataset() {
-  const res = await fetch(DATASET_URL);
-  if (!res.ok) throw new Error(`FinSearchComp dataset HTTP ${res.status} fetching ${DATASET_URL}`);
-  const data = await res.json();
-  if (!Array.isArray(data) || data.length === 0) {
-    throw new Error(`FinSearchComp dataset empty or not an array: ${DATASET_URL}`);
-  }
-  return data;
-}
-function createFinsearchcompAdapter() {
-  const fixturesMode = process.env.FINSEARCHCOMP_FIXTURES === "1";
-  return {
-    name: "finsearchcomp",
-    async preflight() {
-      judgeRouter();
-      console.warn(
-        "[finsearchcomp] T1 (Time_Sensitive_Data_Fetching) is NOT scored: it needs a live market snapshot / akshare ground truth for the judge {ground_truth} slot. loadTasks supports T2 + T3 only; wire T1Seam to add T1."
-      );
-      if (fixturesMode) {
-        await readFile(FIXTURES, "utf8").catch((err) => {
-          throw new Error(
-            `FINSEARCHCOMP_FIXTURES=1 but ${FIXTURES} unreadable: ${err instanceof Error ? err.message : err}`
-          );
-        });
-        return;
-      }
-      const res = await fetch(DATASET_URL, { method: "HEAD" }).catch((err) => {
-        throw new Error(
-          `finsearchcomp preflight failed reaching ${DATASET_URL}: ${err instanceof Error ? err.message : err}
-Fix: ensure network access to raw.githubusercontent.com, or set FINSEARCHCOMP_FIXTURES=1 to run against the committed fixtures offline.`
-        );
-      });
-      if (!res.ok) {
-        throw new Error(
-          `finsearchcomp preflight: dataset HEAD ${res.status} for ${DATASET_URL}. Set FINSEARCHCOMP_FIXTURES=1 to run against the committed fixtures offline.`
-        );
-      }
-    },
-    async loadTasks(opts = {}) {
-      if (fixturesMode) return loadFixtures(opts);
-      let records;
-      try {
-        records = await fetchDataset();
-      } catch (err) {
-        console.warn(
-          `[finsearchcomp] live dataset fetch failed (${err instanceof Error ? err.message : err}); falling back to committed fixtures at ${FIXTURES}`
-        );
-        return loadFixtures(opts);
-      }
-      return selectRecords(records, opts);
-    },
-    async goldArtifact(task) {
-      const meta = readMeta(task);
-      return meta.responseReference;
-    },
-    async judge(task, artifact) {
-      const meta = readMeta(task);
-      const response = artifact.trim();
-      if (response.length === 0) {
-        return {
-          resolved: false,
-          score: 0,
-          detail: JSON.stringify({ taskType: meta.taskType, region: meta.region, reason: "empty answer" })
-        };
-      }
-      return runRecordJudge(meta, response, judgeRouter());
-    }
-  };
-}
-
-export {
-  createFinsearchcompAdapter
-};
-//# sourceMappingURL=chunk-UPAMRDX4.js.map

package/dist/chunk-UPAMRDX4.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/benchmarks/finsearchcomp.ts"],"sourcesContent":["/**\n * FinSearchComp adapter (randomtutu/FinSearchComp). Agentic financial search +\n * reasoning. Worker artifact = the agent's free-form final answer text (zh/en).\n *\n * Each record is SELF-CONTAINED and carries its OWN judge: a per-record\n * `judge_system_prompt` + a `judge_prompt_template` with {prompt},\n * {response_reference} (gold answer WITH embedded tolerance/scoring-criteria),\n * and {response} (the student answer) placeholders. We faithfully REPLICATE the\n * benchmark's judge — fill the template, run it under the record's system prompt\n * via the pinned router (temperature 0), and parse the JSON the judge emits. The\n * judge returns an `answer_score` field; 1/positive = resolved, else not.\n * This matches the FinSearchComp leaderboard (Grok-4-web 68.9%).\n *\n * There is NO deterministic Tier-1 path here: the gold answer is prose with\n * region-specific units, embedded tolerances (\"允许1%的误差\"), and multi-point\n * scoring criteria. The benchmark intends its own LLM judge to be the arbiter,\n * so the judge IS the score — fail loud on unparseable judge output (never\n * default to resolved).\n *\n * SCOPE: T2 (Simple_Historical_Lookup) + T3 (Complex_Historical_Investigation)\n * are fully self-contained and supported. T1 (Time_Sensitive_Data_Fetching)\n * needs a live market snapshot / akshare ground truth to fill the judge's\n * {ground_truth} slot — it is SKIPPED in loadTasks and flagged in preflight.\n * See {@link T1Seam}.\n *\n * Requires for a live run: network access to the GitHub-hosted dataset JSON and\n * a TANGLE_API_KEY for the judge. For offline/CI verification, loadTasks falls back\n * to the committed fixtures (bench/fixtures/finsearchcomp.json) with an explicit\n * console.warn — never a silent fallback.\n */\n\nimport { readFile } from 'node:fs/promises'\nimport { join } from 'node:path'\nimport { fileURLToPath } from 'node:url'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\n\nconst BENCH_ROOT = fileURLToPath(new URL('../..', import.meta.url))\nconst FIXTURES = join(BENCH_ROOT, 'fixtures', 'finsearchcomp.json')\n\nconst DATASET_URL =\n  'https://raw.githubusercontent.com/randomtutu/FinSearchComp/main/data/finsearchcomp_data.json'\n\n/** Task-type prefix on prompt_id, e.g. \"(T2)Simple_Historical_Lookup_001\". */\ntype TaskType = 'T1' | 'T2' | 'T3'\n/** Only these are self-contained (gold + own judge, no external ground truth). */\nconst SUPPORTED_TYPES: ReadonlySet<TaskType> = new Set<TaskType>(['T2', 'T3'])\n\n/**\n * Worker contract appended to every task prompt. The judge is paraphrase- and\n * keyword-aware (\"答案为\", \"the answer is\"), so we do not impose a sentinel; we\n * pass the whole artifact through as {response}. We DO ask the worker to lead\n * with its final value so a truncated artifact still carries the answer.\n */\nconst WORKER_CONTRACT = [\n  '',\n  'Research this financial question using live web/market sources and answer it.',\n  'State your final answer explicitly, with the exact units and precision the question requests.',\n  'Respect any tolerance the question implies; show the value you settled on, not just intermediate work.',\n].join('\\n')\n\n/**\n * Typed seam for T1 (Time_Sensitive_Data_Fetching). T1 records score against a\n * live market snapshot the judge consumes via a {ground_truth} template slot;\n * the adapter does NOT fabricate that snapshot. Wiring T1 means supplying a\n * resolver that fetches the akshare/market value for the record at judge time\n * and filling {ground_truth}. Until then T1 is excluded from loadTasks.\n */\nexport interface T1Seam {\n  promptId: string\n  /** Resolver that returns the real-time ground-truth block for the {ground_truth} slot. */\n  resolveGroundTruth(promptId: string): Promise<string>\n}\n\ninterface FinSearchRecord {\n  prompt_id: string\n  prompt: string\n  response_reference: string\n  judge_prompt_template: string\n  judge_system_prompt: string\n  label: string\n}\n\ninterface FinSearchMeta {\n  promptId: string\n  label: string\n  taskType: TaskType\n  region: string\n  responseReference: string\n  judgePromptTemplate: string\n  judgeSystemPrompt: string\n  rawPrompt: string\n}\n\n/** \"(T2)Simple_Historical_Lookup_001\" → \"T2\"; throws on an unknown/missing prefix. */\nfunction parseTaskType(promptId: string): TaskType {\n  const m = promptId.match(/^\\((T[123])\\)/)\n  if (!m) throw new Error(`FinSearchComp prompt_id has no (T1|T2|T3) prefix: ${JSON.stringify(promptId)}`)\n  return m[1] as TaskType\n}\n\n/** \"Simple_Historical_Lookup(Greater China)\" → \"Greater China\"; '' when absent. */\nfunction parseRegion(label: string): string {\n  const m = label.match(/\\(([^()]+)\\)\\s*$/)\n  return m ? m[1].trim() : ''\n}\n\nfunction recordToTask(rec: FinSearchRecord): BenchTask {\n  const taskType = parseTaskType(rec.prompt_id)\n  const meta: FinSearchMeta = {\n    promptId: rec.prompt_id,\n    label: rec.label,\n    taskType,\n    region: parseRegion(rec.label),\n    responseReference: rec.response_reference,\n    judgePromptTemplate: rec.judge_prompt_template,\n    judgeSystemPrompt: rec.judge_system_prompt,\n    rawPrompt: rec.prompt,\n  }\n  return {\n    id: rec.prompt_id,\n    split: taskType,\n    prompt: rec.prompt + WORKER_CONTRACT,\n    metadata: meta as unknown as Record<string, unknown>,\n  }\n}\n\nfunction readMeta(task: BenchTask): FinSearchMeta {\n  const md = task.metadata\n  if (\n    !md ||\n    typeof md.responseReference !== 'string' ||\n    typeof md.judgePromptTemplate !== 'string' ||\n    typeof md.judgeSystemPrompt !== 'string' ||\n    typeof md.rawPrompt !== 'string'\n  ) {\n    throw new Error(`FinSearchComp task ${task.id} missing judge metadata — loadTasks did not populate it`)\n  }\n  return md as unknown as FinSearchMeta\n}\n\n/** Fill the per-record judge_prompt_template. {ground_truth} is T1-only and unused for T2/T3. */\nfunction fillJudgePrompt(meta: FinSearchMeta, response: string): string {\n  return meta.judgePromptTemplate\n    .replaceAll('{prompt}', meta.rawPrompt)\n    .replaceAll('{response_reference}', meta.responseReference)\n    .replaceAll('{response}', response)\n}\n\ninterface JudgeRouter {\n  baseUrl: string\n  key: string\n  model: string\n}\n\nfunction judgeRouter(): JudgeRouter {\n  const key = process.env.TANGLE_API_KEY\n  if (!key) throw new Error('TANGLE_API_KEY is required for the FinSearchComp per-record LLM judge')\n  const model = process.env.JUDGE_MODEL ?? 'deepseek-v4-flash'\n  const baseUrl = process.env.ROUTER_BASE ?? 'https://router.tangle.tools/v1'\n  return { baseUrl, key, model }\n}\n\n/**\n * The judge emits a JSON object with an `answer_score`. In this dataset it is a\n * flat integer ({\"answer_score\": 1}); we also accept the nested-array form\n * ([[1]] → read [0][0]) the upstream scorer can produce. 1/positive = resolved.\n * A \"null\" sentinel (judge could not obtain ground truth) is NOT a pass.\n */\nfunction readAnswerScore(value: unknown): { resolved: boolean; score: number; raw: unknown } {\n  let v: unknown = value\n  // unwrap nested-array form: [[s]] / [s]\n  while (Array.isArray(v)) v = v[0]\n  if (v === null || v === undefined || v === 'null') {\n    return { resolved: false, score: 0, raw: value }\n  }\n  const n = typeof v === 'number' ? v : typeof v === 'string' ? Number(v) : NaN\n  if (!Number.isFinite(n)) {\n    throw new Error(`FinSearchComp judge answer_score not numeric: ${JSON.stringify(value)}`)\n  }\n  // Their scale: 1 = correct, 0 = wrong. Treat any positive as resolved; clamp to 0..1.\n  const resolved = n >= 1\n  const score = n <= 0 ? 0 : n >= 1 ? 1 : n\n  return { resolved, score, raw: value }\n}\n\n/** Extract the first JSON object carrying `answer_score` from judge output; throw if none. */\nfunction parseJudgeOutput(content: string): { resolved: boolean; score: number; raw: unknown } {\n  // Prefer a fenced block, then fall back to scanning for a {...} with answer_score.\n  const candidates: string[] = []\n  for (const m of content.matchAll(/```(?:json)?\\s*([\\s\\S]*?)```/g)) candidates.push(m[1].trim())\n  for (const m of content.matchAll(/\\{[^{}]*\"answer_score\"[\\s\\S]*?\\}/g)) candidates.push(m[0])\n  candidates.push(content.trim())\n  for (const c of candidates) {\n    let parsed: { answer_score?: unknown }\n    try {\n      parsed = JSON.parse(c) as { answer_score?: unknown }\n    } catch {\n      continue\n    }\n    if ('answer_score' in parsed) return readAnswerScore(parsed.answer_score)\n  }\n  throw new Error(\n    `FinSearchComp judge produced no parseable {\"answer_score\": …}: ${content.slice(0, 400)}`,\n  )\n}\n\n/** Run the record's own judge via the router. Fail loud on transport/parse errors. */\nasync function runRecordJudge(meta: FinSearchMeta, response: string, router: JudgeRouter): Promise<BenchScore> {\n  const res = await fetch(`${router.baseUrl}/chat/completions`, {\n    method: 'POST',\n    headers: { 'content-type': 'application/json', authorization: `Bearer ${router.key}` },\n    body: JSON.stringify({\n      model: router.model,\n      temperature: 0,\n      messages: [\n        { role: 'system', content: meta.judgeSystemPrompt },\n        { role: 'user', content: fillJudgePrompt(meta, response) },\n      ],\n    }),\n  })\n  if (!res.ok) {\n    throw new Error(`FinSearchComp judge HTTP ${res.status}: ${(await res.text()).slice(0, 300)}`)\n  }\n  const body = (await res.json()) as { choices?: Array<{ message?: { content?: string } }> }\n  const content = body.choices?.[0]?.message?.content\n  if (typeof content !== 'string') {\n    throw new Error(`FinSearchComp judge returned no message content: ${JSON.stringify(body).slice(0, 300)}`)\n  }\n  const { resolved, score, raw } = parseJudgeOutput(content)\n  return {\n    resolved,\n    score,\n    detail: JSON.stringify({\n      taskType: meta.taskType,\n      region: meta.region,\n      answerScore: raw,\n      judgeModel: router.model,\n    }),\n  }\n}\n\nfunction selectRecords(records: FinSearchRecord[], opts: LoadOptions): BenchTask[] {\n  let tasks = records\n    .filter((r) => SUPPORTED_TYPES.has(parseTaskType(r.prompt_id)))\n    .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 = balanceByType(tasks, opts.limit)\n  }\n  return tasks\n}\n\n/** Round-robin across task types (T2/T3) so a `limit` pulls a balanced mix\n *  rather than slicing the type-ordered dataset's first N (which yields all\n *  T2). Deterministic — preserves dataset order within each type. */\nfunction balanceByType(tasks: BenchTask[], limit: number): BenchTask[] {\n  const queues = new Map<TaskType, BenchTask[]>()\n  for (const t of tasks) {\n    const k = parseTaskType(t.id)\n    const q = queues.get(k)\n    if (q) q.push(t)\n    else queues.set(k, [t])\n  }\n  const lanes = [...queues.values()]\n  const out: BenchTask[] = []\n  for (let i = 0; out.length < limit; i++) {\n    const before = out.length\n    for (const lane of lanes) {\n      if (i < lane.length) out.push(lane[i])\n      if (out.length >= limit) break\n    }\n    if (out.length === before) break // all lanes exhausted\n  }\n  return out\n}\n\nasync function loadFixtures(opts: LoadOptions): Promise<BenchTask[]> {\n  const records = JSON.parse(await readFile(FIXTURES, 'utf8')) as FinSearchRecord[]\n  return selectRecords(records, opts)\n}\n\nasync function fetchDataset(): Promise<FinSearchRecord[]> {\n  const res = await fetch(DATASET_URL)\n  if (!res.ok) throw new Error(`FinSearchComp dataset HTTP ${res.status} fetching ${DATASET_URL}`)\n  const data = (await res.json()) as FinSearchRecord[]\n  if (!Array.isArray(data) || data.length === 0) {\n    throw new Error(`FinSearchComp dataset empty or not an array: ${DATASET_URL}`)\n  }\n  return data\n}\n\nexport function createFinsearchcompAdapter(): BenchmarkAdapter {\n  const fixturesMode = process.env.FINSEARCHCOMP_FIXTURES === '1'\n\n  return {\n    name: 'finsearchcomp',\n\n    async preflight() {\n      // The per-record LLM judge is the arbiter in both modes — its router\n      // config must exist or no score can be produced.\n      judgeRouter()\n      console.warn(\n        '[finsearchcomp] T1 (Time_Sensitive_Data_Fetching) is NOT scored: it needs a live ' +\n          'market snapshot / akshare ground truth for the judge {ground_truth} slot. ' +\n          'loadTasks supports T2 + T3 only; wire T1Seam to add T1.',\n      )\n      if (fixturesMode) {\n        await readFile(FIXTURES, 'utf8').catch((err) => {\n          throw new Error(\n            `FINSEARCHCOMP_FIXTURES=1 but ${FIXTURES} unreadable: ${err instanceof Error ? err.message : err}`,\n          )\n        })\n        return\n      }\n      const res = await fetch(DATASET_URL, { method: 'HEAD' }).catch((err) => {\n        throw new Error(\n          `finsearchcomp preflight failed reaching ${DATASET_URL}: ${err instanceof Error ? err.message : err}\\n` +\n            `Fix: ensure network access to raw.githubusercontent.com, or set FINSEARCHCOMP_FIXTURES=1 ` +\n            `to run against the committed fixtures offline.`,\n        )\n      })\n      if (!res.ok) {\n        throw new Error(\n          `finsearchcomp preflight: dataset HEAD ${res.status} for ${DATASET_URL}. ` +\n            `Set FINSEARCHCOMP_FIXTURES=1 to run against the committed fixtures offline.`,\n        )\n      }\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      if (fixturesMode) return loadFixtures(opts)\n      let records: FinSearchRecord[]\n      try {\n        records = await fetchDataset()\n      } catch (err) {\n        console.warn(\n          `[finsearchcomp] live dataset fetch failed (${err instanceof Error ? err.message : err}); ` +\n            `falling back to committed fixtures at ${FIXTURES}`,\n        )\n        return loadFixtures(opts)\n      }\n      return selectRecords(records, opts)\n    },\n\n    async goldArtifact(task: BenchTask) {\n      // Gold artifact = the record's response_reference fed back as the student\n      // answer, so verify-judge proves gold→resolved through the SAME per-record\n      // judge the real artifact takes. The reference may carry tolerance/criteria\n      // prose; the judge is built to read the value out of it.\n      const meta = readMeta(task)\n      return meta.responseReference\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const meta = readMeta(task)\n      const response = artifact.trim()\n      if (response.length === 0) {\n        // Fail-closed — the benchmark's own judges score an empty answer 0.\n        return {\n          resolved: false,\n          score: 0,\n          detail: JSON.stringify({ taskType: meta.taskType, region: meta.region, reason: 'empty answer' }),\n        }\n      }\n      return runRecordJudge(meta, response, judgeRouter())\n    },\n  }\n}\n"],"mappings":";AA+BA,SAAS,gBAAgB;AACzB,SAAS,YAAY;AACrB,SAAS,qBAAqB;AAG9B,IAAM,aAAa,cAAc,IAAI,IAAI,SAAS,YAAY,GAAG,CAAC;AAClE,IAAM,WAAW,KAAK,YAAY,YAAY,oBAAoB;AAElE,IAAM,cACJ;AAKF,IAAM,kBAAyC,oBAAI,IAAc,CAAC,MAAM,IAAI,CAAC;AAQ7E,IAAM,kBAAkB;AAAA,EACtB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,EAAE,KAAK,IAAI;AAoCX,SAAS,cAAc,UAA4B;AACjD,QAAM,IAAI,SAAS,MAAM,eAAe;AACxC,MAAI,CAAC,EAAG,OAAM,IAAI,MAAM,qDAAqD,KAAK,UAAU,QAAQ,CAAC,EAAE;AACvG,SAAO,EAAE,CAAC;AACZ;AAGA,SAAS,YAAY,OAAuB;AAC1C,QAAM,IAAI,MAAM,MAAM,kBAAkB;AACxC,SAAO,IAAI,EAAE,CAAC,EAAE,KAAK,IAAI;AAC3B;AAEA,SAAS,aAAa,KAAiC;AACrD,QAAM,WAAW,cAAc,IAAI,SAAS;AAC5C,QAAM,OAAsB;AAAA,IAC1B,UAAU,IAAI;AAAA,IACd,OAAO,IAAI;AAAA,IACX;AAAA,IACA,QAAQ,YAAY,IAAI,KAAK;AAAA,IAC7B,mBAAmB,IAAI;AAAA,IACvB,qBAAqB,IAAI;AAAA,IACzB,mBAAmB,IAAI;AAAA,IACvB,WAAW,IAAI;AAAA,EACjB;AACA,SAAO;AAAA,IACL,IAAI,IAAI;AAAA,IACR,OAAO;AAAA,IACP,QAAQ,IAAI,SAAS;AAAA,IACrB,UAAU;AAAA,EACZ;AACF;AAEA,SAAS,SAAS,MAAgC;AAChD,QAAM,KAAK,KAAK;AAChB,MACE,CAAC,MACD,OAAO,GAAG,sBAAsB,YAChC,OAAO,GAAG,wBAAwB,YAClC,OAAO,GAAG,sBAAsB,YAChC,OAAO,GAAG,cAAc,UACxB;AACA,UAAM,IAAI,MAAM,sBAAsB,KAAK,EAAE,8DAAyD;AAAA,EACxG;AACA,SAAO;AACT;AAGA,SAAS,gBAAgB,MAAqB,UAA0B;AACtE,SAAO,KAAK,oBACT,WAAW,YAAY,KAAK,SAAS,EACrC,WAAW,wBAAwB,KAAK,iBAAiB,EACzD,WAAW,cAAc,QAAQ;AACtC;AAQA,SAAS,cAA2B;AAClC,QAAM,MAAM,QAAQ,IAAI;AACxB,MAAI,CAAC,IAAK,OAAM,IAAI,MAAM,uEAAuE;AACjG,QAAM,QAAQ,QAAQ,IAAI,eAAe;AACzC,QAAM,UAAU,QAAQ,IAAI,eAAe;AAC3C,SAAO,EAAE,SAAS,KAAK,MAAM;AAC/B;AAQA,SAAS,gBAAgB,OAAoE;AAC3F,MAAI,IAAa;AAEjB,SAAO,MAAM,QAAQ,CAAC,EAAG,KAAI,EAAE,CAAC;AAChC,MAAI,MAAM,QAAQ,MAAM,UAAa,MAAM,QAAQ;AACjD,WAAO,EAAE,UAAU,OAAO,OAAO,GAAG,KAAK,MAAM;AAAA,EACjD;AACA,QAAM,IAAI,OAAO,MAAM,WAAW,IAAI,OAAO,MAAM,WAAW,OAAO,CAAC,IAAI;AAC1E,MAAI,CAAC,OAAO,SAAS,CAAC,GAAG;AACvB,UAAM,IAAI,MAAM,iDAAiD,KAAK,UAAU,KAAK,CAAC,EAAE;AAAA,EAC1F;AAEA,QAAM,WAAW,KAAK;AACtB,QAAM,QAAQ,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI;AACxC,SAAO,EAAE,UAAU,OAAO,KAAK,MAAM;AACvC;AAGA,SAAS,iBAAiB,SAAqE;AAE7F,QAAM,aAAuB,CAAC;AAC9B,aAAW,KAAK,QAAQ,SAAS,+BAA+B,EAAG,YAAW,KAAK,EAAE,CAAC,EAAE,KAAK,CAAC;AAC9F,aAAW,KAAK,QAAQ,SAAS,mCAAmC,EAAG,YAAW,KAAK,EAAE,CAAC,CAAC;AAC3F,aAAW,KAAK,QAAQ,KAAK,CAAC;AAC9B,aAAW,KAAK,YAAY;AAC1B,QAAI;AACJ,QAAI;AACF,eAAS,KAAK,MAAM,CAAC;AAAA,IACvB,QAAQ;AACN;AAAA,IACF;AACA,QAAI,kBAAkB,OAAQ,QAAO,gBAAgB,OAAO,YAAY;AAAA,EAC1E;AACA,QAAM,IAAI;AAAA,IACR,uEAAkE,QAAQ,MAAM,GAAG,GAAG,CAAC;AAAA,EACzF;AACF;AAGA,eAAe,eAAe,MAAqB,UAAkB,QAA0C;AAC7G,QAAM,MAAM,MAAM,MAAM,GAAG,OAAO,OAAO,qBAAqB;AAAA,IAC5D,QAAQ;AAAA,IACR,SAAS,EAAE,gBAAgB,oBAAoB,eAAe,UAAU,OAAO,GAAG,GAAG;AAAA,IACrF,MAAM,KAAK,UAAU;AAAA,MACnB,OAAO,OAAO;AAAA,MACd,aAAa;AAAA,MACb,UAAU;AAAA,QACR,EAAE,MAAM,UAAU,SAAS,KAAK,kBAAkB;AAAA,QAClD,EAAE,MAAM,QAAQ,SAAS,gBAAgB,MAAM,QAAQ,EAAE;AAAA,MAC3D;AAAA,IACF,CAAC;AAAA,EACH,CAAC;AACD,MAAI,CAAC,IAAI,IAAI;AACX,UAAM,IAAI,MAAM,4BAA4B,IAAI,MAAM,MAAM,MAAM,IAAI,KAAK,GAAG,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EAC/F;AACA,QAAM,OAAQ,MAAM,IAAI,KAAK;AAC7B,QAAM,UAAU,KAAK,UAAU,CAAC,GAAG,SAAS;AAC5C,MAAI,OAAO,YAAY,UAAU;AAC/B,UAAM,IAAI,MAAM,oDAAoD,KAAK,UAAU,IAAI,EAAE,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EAC1G;AACA,QAAM,EAAE,UAAU,OAAO,IAAI,IAAI,iBAAiB,OAAO;AACzD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,QAAQ,KAAK,UAAU;AAAA,MACrB,UAAU,KAAK;AAAA,MACf,QAAQ,KAAK;AAAA,MACb,aAAa;AAAA,MACb,YAAY,OAAO;AAAA,IACrB,CAAC;AAAA,EACH;AACF;AAEA,SAAS,cAAc,SAA4B,MAAgC;AACjF,MAAI,QAAQ,QACT,OAAO,CAAC,MAAM,gBAAgB,IAAI,cAAc,EAAE,SAAS,CAAC,CAAC,EAC7D,IAAI,YAAY;AACnB,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,cAAc,OAAO,KAAK,KAAK;AAAA,EACzC;AACA,SAAO;AACT;AAKA,SAAS,cAAc,OAAoB,OAA4B;AACrE,QAAM,SAAS,oBAAI,IAA2B;AAC9C,aAAW,KAAK,OAAO;AACrB,UAAM,IAAI,cAAc,EAAE,EAAE;AAC5B,UAAM,IAAI,OAAO,IAAI,CAAC;AACtB,QAAI,EAAG,GAAE,KAAK,CAAC;AAAA,QACV,QAAO,IAAI,GAAG,CAAC,CAAC,CAAC;AAAA,EACxB;AACA,QAAM,QAAQ,CAAC,GAAG,OAAO,OAAO,CAAC;AACjC,QAAM,MAAmB,CAAC;AAC1B,WAAS,IAAI,GAAG,IAAI,SAAS,OAAO,KAAK;AACvC,UAAM,SAAS,IAAI;AACnB,eAAW,QAAQ,OAAO;AACxB,UAAI,IAAI,KAAK,OAAQ,KAAI,KAAK,KAAK,CAAC,CAAC;AACrC,UAAI,IAAI,UAAU,MAAO;AAAA,IAC3B;AACA,QAAI,IAAI,WAAW,OAAQ;AAAA,EAC7B;AACA,SAAO;AACT;AAEA,eAAe,aAAa,MAAyC;AACnE,QAAM,UAAU,KAAK,MAAM,MAAM,SAAS,UAAU,MAAM,CAAC;AAC3D,SAAO,cAAc,SAAS,IAAI;AACpC;AAEA,eAAe,eAA2C;AACxD,QAAM,MAAM,MAAM,MAAM,WAAW;AACnC,MAAI,CAAC,IAAI,GAAI,OAAM,IAAI,MAAM,8BAA8B,IAAI,MAAM,aAAa,WAAW,EAAE;AAC/F,QAAM,OAAQ,MAAM,IAAI,KAAK;AAC7B,MAAI,CAAC,MAAM,QAAQ,IAAI,KAAK,KAAK,WAAW,GAAG;AAC7C,UAAM,IAAI,MAAM,gDAAgD,WAAW,EAAE;AAAA,EAC/E;AACA,SAAO;AACT;AAEO,SAAS,6BAA+C;AAC7D,QAAM,eAAe,QAAQ,IAAI,2BAA2B;AAE5D,SAAO;AAAA,IACL,MAAM;AAAA,IAEN,MAAM,YAAY;AAGhB,kBAAY;AACZ,cAAQ;AAAA,QACN;AAAA,MAGF;AACA,UAAI,cAAc;AAChB,cAAM,SAAS,UAAU,MAAM,EAAE,MAAM,CAAC,QAAQ;AAC9C,gBAAM,IAAI;AAAA,YACR,gCAAgC,QAAQ,gBAAgB,eAAe,QAAQ,IAAI,UAAU,GAAG;AAAA,UAClG;AAAA,QACF,CAAC;AACD;AAAA,MACF;AACA,YAAM,MAAM,MAAM,MAAM,aAAa,EAAE,QAAQ,OAAO,CAAC,EAAE,MAAM,CAAC,QAAQ;AACtE,cAAM,IAAI;AAAA,UACR,2CAA2C,WAAW,KAAK,eAAe,QAAQ,IAAI,UAAU,GAAG;AAAA;AAAA,QAGrG;AAAA,MACF,CAAC;AACD,UAAI,CAAC,IAAI,IAAI;AACX,cAAM,IAAI;AAAA,UACR,yCAAyC,IAAI,MAAM,QAAQ,WAAW;AAAA,QAExE;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;AACtC,UAAI,aAAc,QAAO,aAAa,IAAI;AAC1C,UAAI;AACJ,UAAI;AACF,kBAAU,MAAM,aAAa;AAAA,MAC/B,SAAS,KAAK;AACZ,gBAAQ;AAAA,UACN,8CAA8C,eAAe,QAAQ,IAAI,UAAU,GAAG,4CAC3C,QAAQ;AAAA,QACrD;AACA,eAAO,aAAa,IAAI;AAAA,MAC1B;AACA,aAAO,cAAc,SAAS,IAAI;AAAA,IACpC;AAAA,IAEA,MAAM,aAAa,MAAiB;AAKlC,YAAM,OAAO,SAAS,IAAI;AAC1B,aAAO,KAAK;AAAA,IACd;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,OAAO,SAAS,IAAI;AAC1B,YAAM,WAAW,SAAS,KAAK;AAC/B,UAAI,SAAS,WAAW,GAAG;AAEzB,eAAO;AAAA,UACL,UAAU;AAAA,UACV,OAAO;AAAA,UACP,QAAQ,KAAK,UAAU,EAAE,UAAU,KAAK,UAAU,QAAQ,KAAK,QAAQ,QAAQ,eAAe,CAAC;AAAA,QACjG;AAAA,MACF;AACA,aAAO,eAAe,MAAM,UAAU,YAAY,CAAC;AAAA,IACrD;AAAA,EACF;AACF;","names":[]}

package/dist/chunk-V7AEBY6U.js

@@ -1,144 +0,0 @@
-import {
-  createToolLlmAdapter
-} from "./chunk-XKEFIFIC.js";
-import {
-  createTrataHedgeAdapter
-} from "./chunk-VQRS7VUC.js";
-import {
-  createWebArenaVerifiedAdapter
-} from "./chunk-HWST3SED.js";
-import {
-  createRagBenchAdapter
-} from "./chunk-IFVINJ4B.js";
-import {
-  createSimpleQaAdapter
-} from "./chunk-SEVJPLZC.js";
-import {
-  createSweBenchAdapter
-} from "./chunk-3U5TXJZS.js";
-import {
-  createT2RagBenchAdapter
-} from "./chunk-SHM6MRRF.js";
-import {
-  createTau2BenchAdapter
-} from "./chunk-RCYQEFNX.js";
-import {
-  createTau3BankingAdapter
-} from "./chunk-SFLA7OH3.js";
-import {
-  createTerminalBenchAdapter
-} from "./chunk-XYA4XSNU.js";
-import {
-  createFramesAdapter
-} from "./chunk-INNOYXCP.js";
-import {
-  createHotpotqaAdapter
-} from "./chunk-PA2ZKHJC.js";
-import {
-  createHumanEvalAdapter
-} from "./chunk-5H5XV76F.js";
-import {
-  createMind2WebAdapter
-} from "./chunk-TBKU5XQI.js";
-import {
-  createNoMiraclAdapter
-} from "./chunk-NQG5XDSB.js";
-import {
-  createOpenRagBenchAdapter
-} from "./chunk-SHYIRB7I.js";
-import {
-  createProgrambenchAdapter
-} from "./chunk-MQMRLGOG.js";
-import {
-  createCadBenchAdapter
-} from "./chunk-PUIRNYI7.js";
-import {
-  createCadGenBenchAdapter
-} from "./chunk-C7T7WEK2.js";
-import {
-  createCommit0Adapter
-} from "./chunk-Z4TZ76N7.js";
-import {
-  createCragAdapter
-} from "./chunk-SVR2LKYI.js";
-import {
-  createDabstepAdapter
-} from "./chunk-RH5F53JT.js";
-import {
-  createEnterpriseOpsGymAdapter
-} from "./chunk-YSMEKBTD.js";
-import {
-  createFinResearchBenchAdapter
-} from "./chunk-7GRVHU22.js";
-import {
-  createFinsearchcompAdapter
-} from "./chunk-UPAMRDX4.js";
-import {
-  createAecBenchAdapter
-} from "./chunk-K3BQGZCT.js";
-import {
-  createAgentBenchAdapter
-} from "./chunk-PB64GYIG.js";
-import {
-  createAppWorldAdapter,
-  createAppWorldReactAdapter
-} from "./chunk-IA2FBTWC.js";
-import {
-  createBfclAdapter
-} from "./chunk-KP5KD6EN.js";
-import {
-  createCadDesignAdapter
-} from "./chunk-53UPUNBZ.js";
-
-// src/adapters.ts
-var ADAPTERS = {
-  "swe-bench": createSweBenchAdapter,
-  "terminal-bench": createTerminalBenchAdapter,
-  // Code-benches sharing ./benchmarks/_harness (stage → external evaluator → report).
-  // loadTasks fetches the REAL dataset (committed fixtures fallback offline); judge
-  // delegates to the benchmark's own harness and fails loud when it/Docker is absent.
-  "aec-bench": createAecBenchAdapter,
-  commit0: createCommit0Adapter,
-  dabstep: createDabstepAdapter,
-  programbench: createProgrambenchAdapter,
-  "webarena-verified": createWebArenaVerifiedAdapter,
-  "tau2-bench": createTau2BenchAdapter,
-  "tau3-banking": createTau3BankingAdapter,
-  agentbench: createAgentBenchAdapter,
-  bfcl: createBfclAdapter,
-  toollm: createToolLlmAdapter,
-  appworld: createAppWorldAdapter,
-  // AppWorld's native interactive protocol — the worker is the in-engine ReAct
-  // episode (execution feedback every turn), the mode published baselines use.
-  "appworld-react": createAppWorldReactAdapter,
-  "enterpriseops-gym": createEnterpriseOpsGymAdapter,
-  "cad-design": createCadDesignAdapter,
-  cadbench: createCadBenchAdapter,
-  cadgenbench: createCadGenBenchAdapter,
-  frames: createFramesAdapter,
-  ragbench: createRagBenchAdapter,
-  crag: createCragAdapter,
-  nomiracl: createNoMiraclAdapter,
-  "open-rag-bench": createOpenRagBenchAdapter,
-  "t2-ragbench": createT2RagBenchAdapter,
-  finresearchbench: createFinResearchBenchAdapter,
-  finsearchcomp: createFinsearchcompAdapter,
-  simpleqa: createSimpleQaAdapter,
-  hotpotqa: createHotpotqaAdapter,
-  // Deployable-checker code domain: worker = router completion, judge = Docker test
-  // run (--network=none). The steering A/B counterpart to humaneval-gate.mts (selection).
-  humaneval: createHumanEvalAdapter,
-  mind2web: createMind2WebAdapter,
-  "trata-hedge": createTrataHedgeAdapter
-};
-function resolveAdapter(key) {
-  const make = ADAPTERS[key];
-  if (!make) throw new Error(`unknown benchmark ${JSON.stringify(key)} (have: ${Object.keys(ADAPTERS).join(", ")})`);
-  return make();
-}
-
-export {
-  ADAPTERS,
-  resolveAdapter
-};
-//# sourceMappingURL=chunk-V7AEBY6U.js.map

package/dist/chunk-V7AEBY6U.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/adapters.ts"],"sourcesContent":["/**\n * The benchmark registry — the single source of truth for every wired benchmark.\n * One key per benchmark; the value is its `BenchmarkAdapter` factory. `runBenchmarks`\n * (the unifier) maps over this; `run.ts`, `rsi.ts`, and `corpus-replay.mts` all read it\n * here rather than each keeping their own copy.\n */\n\nimport { createAecBenchAdapter } from './benchmarks/aec-bench'\nimport { createAgentBenchAdapter } from './benchmarks/agentbench'\nimport { createAppWorldAdapter, createAppWorldReactAdapter } from './benchmarks/appworld'\nimport { createBfclAdapter } from './benchmarks/bfcl'\nimport { createCadBenchAdapter } from './benchmarks/cadbench'\nimport { createCadDesignAdapter } from './benchmarks/cad-design'\nimport { createCadGenBenchAdapter } from './benchmarks/cadgenbench'\nimport { createCommit0Adapter } from './benchmarks/commit0'\nimport { createCragAdapter } from './benchmarks/crag'\nimport { createDabstepAdapter } from './benchmarks/dabstep'\nimport { createEnterpriseOpsGymAdapter } from './benchmarks/enterpriseops-gym'\nimport { createFinResearchBenchAdapter } from './benchmarks/finresearchbench'\nimport { createFinsearchcompAdapter } from './benchmarks/finsearchcomp'\nimport { createFramesAdapter } from './benchmarks/frames'\nimport { createHotpotqaAdapter } from './benchmarks/hotpotqa'\nimport { createHumanEvalAdapter } from './benchmarks/humaneval'\nimport { createMind2WebAdapter } from './benchmarks/mind2web'\nimport { createNoMiraclAdapter } from './benchmarks/nomiracl'\nimport { createOpenRagBenchAdapter } from './benchmarks/open-rag-bench'\nimport { createProgrambenchAdapter } from './benchmarks/programbench'\nimport { createRagBenchAdapter } from './benchmarks/ragbench'\nimport { createSimpleQaAdapter } from './benchmarks/simpleqa'\nimport { createSweBenchAdapter } from './benchmarks/swe-bench'\nimport { createT2RagBenchAdapter } from './benchmarks/t2-ragbench'\nimport { createTau2BenchAdapter } from './benchmarks/tau2-bench'\nimport { createTau3BankingAdapter } from './benchmarks/tau3-banking'\nimport { createTerminalBenchAdapter } from './benchmarks/terminal-bench'\nimport { createToolLlmAdapter } from './benchmarks/toollm'\nimport { createTrataHedgeAdapter } from './benchmarks/trata-hedge'\nimport { createWebArenaVerifiedAdapter } from './benchmarks/webarena-verified'\nimport type { BenchmarkAdapter } from './benchmarks/types'\n\nexport const ADAPTERS: Record<string, () => BenchmarkAdapter> = {\n  'swe-bench': createSweBenchAdapter,\n  'terminal-bench': createTerminalBenchAdapter,\n  // Code-benches sharing ./benchmarks/_harness (stage → external evaluator → report).\n  // loadTasks fetches the REAL dataset (committed fixtures fallback offline); judge\n  // delegates to the benchmark's own harness and fails loud when it/Docker is absent.\n  'aec-bench': createAecBenchAdapter,\n  commit0: createCommit0Adapter,\n  dabstep: createDabstepAdapter,\n  programbench: createProgrambenchAdapter,\n  'webarena-verified': createWebArenaVerifiedAdapter,\n  'tau2-bench': createTau2BenchAdapter,\n  'tau3-banking': createTau3BankingAdapter,\n  agentbench: createAgentBenchAdapter,\n  bfcl: createBfclAdapter,\n  toollm: createToolLlmAdapter,\n  appworld: createAppWorldAdapter,\n  // AppWorld's native interactive protocol — the worker is the in-engine ReAct\n  // episode (execution feedback every turn), the mode published baselines use.\n  'appworld-react': createAppWorldReactAdapter,\n  'enterpriseops-gym': createEnterpriseOpsGymAdapter,\n  'cad-design': createCadDesignAdapter,\n  cadbench: createCadBenchAdapter,\n  cadgenbench: createCadGenBenchAdapter,\n  frames: createFramesAdapter,\n  ragbench: createRagBenchAdapter,\n  crag: createCragAdapter,\n  nomiracl: createNoMiraclAdapter,\n  'open-rag-bench': createOpenRagBenchAdapter,\n  't2-ragbench': createT2RagBenchAdapter,\n  finresearchbench: createFinResearchBenchAdapter,\n  finsearchcomp: createFinsearchcompAdapter,\n  simpleqa: createSimpleQaAdapter,\n  hotpotqa: createHotpotqaAdapter,\n  // Deployable-checker code domain: worker = router completion, judge = Docker test\n  // run (--network=none). The steering A/B counterpart to humaneval-gate.mts (selection).\n  humaneval: createHumanEvalAdapter,\n  mind2web: createMind2WebAdapter,\n  'trata-hedge': createTrataHedgeAdapter,\n}\n\n/** Resolve a benchmark key to its adapter, failing loud with the known keys. */\nexport function resolveAdapter(key: string): BenchmarkAdapter {\n  const make = ADAPTERS[key]\n  if (!make) throw new Error(`unknown benchmark ${JSON.stringify(key)} (have: ${Object.keys(ADAPTERS).join(', ')})`)\n  return make()\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCO,IAAM,WAAmD;AAAA,EAC9D,aAAa;AAAA,EACb,kBAAkB;AAAA;AAAA;AAAA;AAAA,EAIlB,aAAa;AAAA,EACb,SAAS;AAAA,EACT,SAAS;AAAA,EACT,cAAc;AAAA,EACd,qBAAqB;AAAA,EACrB,cAAc;AAAA,EACd,gBAAgB;AAAA,EAChB,YAAY;AAAA,EACZ,MAAM;AAAA,EACN,QAAQ;AAAA,EACR,UAAU;AAAA;AAAA;AAAA,EAGV,kBAAkB;AAAA,EAClB,qBAAqB;AAAA,EACrB,cAAc;AAAA,EACd,UAAU;AAAA,EACV,aAAa;AAAA,EACb,QAAQ;AAAA,EACR,UAAU;AAAA,EACV,MAAM;AAAA,EACN,UAAU;AAAA,EACV,kBAAkB;AAAA,EAClB,eAAe;AAAA,EACf,kBAAkB;AAAA,EAClB,eAAe;AAAA,EACf,UAAU;AAAA,EACV,UAAU;AAAA;AAAA;AAAA,EAGV,WAAW;AAAA,EACX,UAAU;AAAA,EACV,eAAe;AACjB;AAGO,SAAS,eAAe,KAA+B;AAC5D,QAAM,OAAO,SAAS,GAAG;AACzB,MAAI,CAAC,KAAM,OAAM,IAAI,MAAM,qBAAqB,KAAK,UAAU,GAAG,CAAC,WAAW,OAAO,KAAK,QAAQ,EAAE,KAAK,IAAI,CAAC,GAAG;AACjH,SAAO,KAAK;AACd;","names":[]}