@tangle-network/agent-bench 0.4.1 → 0.4.3

package/dist/chunk-5H5XV76F.js

@@ -1,240 +0,0 @@
-// src/benchmarks/humaneval.ts
-import { execFile } from "child_process";
-import { chmodSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from "fs";
-import { tmpdir } from "os";
-import { join } from "path";
-import { gunzipSync } from "zlib";
-var humanevalUrl = "https://github.com/openai/human-eval/raw/master/data/HumanEval.jsonl.gz";
-var dockerImage = "python:3.12-slim";
-var dockerTimeoutMs = Number(process.env.DOCKER_TIMEOUT_MS ?? 2e4);
-async function loadHumanEval(limit, offset = 0) {
-  const localGz = process.env.HUMANEVAL_GZ;
-  let gz;
-  if (localGz) {
-    gz = readFileSync(localGz);
-  } else {
-    const res = await fetch(humanevalUrl);
-    if (!res.ok) throw new Error(`HumanEval fetch HTTP ${res.status}: ${humanevalUrl}`);
-    gz = Buffer.from(await res.arrayBuffer());
-  }
-  const text = gunzipSync(gz).toString("utf8");
-  const tasks = [];
-  for (const line of text.split("\n")) {
-    if (line.trim() === "") continue;
-    const d = JSON.parse(line);
-    if (!d.task_id || !d.prompt || !d.test || !d.entry_point) {
-      throw new Error(`malformed HumanEval record: ${line.slice(0, 120)}`);
-    }
-    tasks.push({
-      taskId: d.task_id,
-      prompt: d.prompt,
-      test: d.test,
-      entryPoint: d.entry_point,
-      ...d.canonical_solution ? { canonicalSolution: d.canonical_solution } : {}
-    });
-  }
-  if (tasks.length === 0) throw new Error("HumanEval parsed to 0 tasks");
-  if (offset >= tasks.length) throw new Error(`OFFSET ${offset} >= dataset size ${tasks.length}`);
-  return tasks.slice(offset, offset + limit);
-}
-var solveInstruction = "Complete the following Python function. Output the COMPLETE function definition (signature, docstring optional, body) inside a single ```python code block. Include any imports the function needs. Do not write tests or example calls.";
-function basePrompt(task) {
-  return `${solveInstruction}
-
-\`\`\`python
-${task.prompt}\`\`\``;
-}
-function extractCode(reply) {
-  const fenced = reply.match(/```(?:python|py)?\s*\n([\s\S]*?)```/i);
-  if (fenced && typeof fenced[1] === "string") return fenced[1].trim();
-  return reply.trim();
-}
-function buildProgram(task, candidate) {
-  return `${task.prompt}
-${candidate}
-
-${task.test}
-
-check(${task.entryPoint})
-`;
-}
-var dockerRunSeq = 0;
-function runPythonProgram(program, timeoutMs = dockerTimeoutMs) {
-  const dir = mkdtempSync(join(tmpdir(), "hev-"));
-  chmodSync(dir, 493);
-  writeFileSync(join(dir, "p.py"), program, { mode: 420 });
-  const name = `hev-${process.pid}-${dockerRunSeq++}`;
-  const startMarker = `__AGENT_RUNTIME_CANDIDATE_STARTED_${name}__`;
-  return new Promise((resolvePromise, reject) => {
-    let settled = false;
-    const cleanup = () => {
-      rmSync(dir, { recursive: true, force: true });
-      execFile("docker", ["rm", "-f", name], () => {
-      });
-    };
-    const finish = (res) => {
-      if (settled) return;
-      settled = true;
-      clearTimeout(backstop);
-      cleanup();
-      resolvePromise(res);
-    };
-    const fail = (e) => {
-      if (settled) return;
-      settled = true;
-      clearTimeout(backstop);
-      cleanup();
-      reject(e);
-    };
-    const outerTimeoutMs = timeoutMs + 3e3;
-    const backstop = setTimeout(
-      () => fail(new Error(`docker checker did not return within ${outerTimeoutMs + 3e3}ms`)),
-      outerTimeoutMs + 3e3
-    );
-    const inContainerSeconds = Math.max(1, Math.ceil(timeoutMs / 1e3));
-    execFile(
-      "docker",
-      [
-        "run",
-        "--rm",
-        "--pull=never",
-        "--name",
-        name,
-        "--network=none",
-        "--cpus=1",
-        "--memory=512m",
-        "--pids-limit=64",
-        "--cap-drop=ALL",
-        "--security-opt=no-new-privileges",
-        "--read-only",
-        "--tmpfs",
-        "/tmp:rw,nosuid,nodev,noexec,size=64m,mode=1777",
-        "--user",
-        "65534:65534",
-        "-v",
-        `${dir}:/w:ro`,
-        "-w",
-        "/w",
-        dockerImage,
-        "sh",
-        "-c",
-        'command -v timeout >/dev/null 2>&1 && command -v python >/dev/null 2>&1 && [ -r /w/p.py ] || exit 126; printf "%s\\n" "$1"; exec timeout -s KILL "$2" python /w/p.py',
-        "agent-runtime-checker",
-        startMarker,
-        `${inContainerSeconds}s`
-      ],
-      { timeout: outerTimeoutMs, killSignal: "SIGKILL", maxBuffer: 4 * 1024 * 1024 },
-      (err, stdout, stderr) => {
-        const markerLine = `${startMarker}
-`;
-        const candidateStarted = stdout?.startsWith(markerLine) === true;
-        const candidateStdout = candidateStarted ? stdout.slice(markerLine.length) : stdout ?? "";
-        if (err) {
-          const e = err;
-          if (e.code === "ENOENT") {
-            fail(new Error("docker binary not found on PATH \u2014 cannot run the deployable checker"));
-            return;
-          }
-          if (e.killed) {
-            fail(new Error(`docker checker exceeded its ${outerTimeoutMs}ms outer timeout`));
-            return;
-          }
-          if (!candidateStarted) {
-            if (/cannot connect to the docker daemon|is the docker daemon running|permission denied while trying to connect/i.test(
-              stderr
-            )) {
-              fail(new Error(`docker daemon unreachable: ${stderr.slice(0, 200)}`));
-              return;
-            }
-            if (/unable to find image|no such image|pull access denied|manifest unknown/i.test(stderr)) {
-              fail(new Error(`docker image ${dockerImage} unavailable: ${stderr.slice(0, 200)}`));
-              return;
-            }
-            fail(
-              new Error(
-                `docker checker did not start the candidate: ${stderr.slice(0, 200) || e.message}`
-              )
-            );
-            return;
-          }
-          const exitCode = typeof e.code === "number" ? e.code : 1;
-          finish({
-            exitCode,
-            stdout: candidateStdout,
-            stderr: stderr ?? ""
-          });
-          return;
-        }
-        if (!candidateStarted) {
-          fail(new Error("docker checker exited without starting the candidate"));
-          return;
-        }
-        finish({ exitCode: 0, stdout: candidateStdout, stderr: stderr ?? "" });
-      }
-    );
-  });
-}
-async function runChecker(task, candidate) {
-  const result = await runPythonProgram(buildProgram(task, candidate));
-  return result.exitCode === 0 ? { pass: 1 } : { pass: 0, detail: result.stderr.slice(-600) || "timed out (no output)" };
-}
-function toBenchTask(t) {
-  const metadata = {
-    promptHeader: t.prompt,
-    test: t.test,
-    entryPoint: t.entryPoint,
-    ...t.canonicalSolution ? { canonicalSolution: t.canonicalSolution } : {}
-  };
-  return { id: t.taskId, prompt: basePrompt(t), metadata };
-}
-function taskFromMeta(task) {
-  const m = task.metadata;
-  if (!m?.promptHeader || !m.test || !m.entryPoint) {
-    throw new Error(`HumanEval judge: task ${task.id} missing checker metadata`);
-  }
-  return { taskId: task.id, prompt: m.promptHeader, test: m.test, entryPoint: m.entryPoint };
-}
-function createHumanEvalAdapter() {
-  return {
-    name: "humaneval",
-    async preflight() {
-      await new Promise((resolve, reject) => {
-        execFile("docker", ["version", "--format", "{{.Server.Version}}"], (err) => {
-          if (err) reject(new Error("HumanEval judge needs a running Docker daemon (python:3.12-slim, --network=none)"));
-          else resolve();
-        });
-      });
-      await new Promise((resolve, reject) => {
-        execFile("docker", ["image", "inspect", dockerImage], (err) => {
-          if (err) reject(new Error(`HumanEval judge needs the cached Docker image ${dockerImage}`));
-          else resolve();
-        });
-      });
-    },
-    async loadTasks(opts) {
-      const offset = Number(process.env.OFFSET ?? 0);
-      const all = await loadHumanEval(opts?.ids ? 164 : opts?.limit ?? 8, offset);
-      const picked = opts?.ids ? all.filter((t) => opts.ids?.includes(t.taskId)) : all;
-      return picked.map(toBenchTask);
-    },
-    async judge(task, artifact) {
-      const { pass } = await runChecker(taskFromMeta(task), extractCode(artifact));
-      return { resolved: pass === 1, score: pass, detail: pass === 1 ? "tests passed" : "tests failed" };
-    },
-    async goldArtifact(task) {
-      const m = task.metadata;
-      const sol = m?.canonicalSolution;
-      return sol ? `${m.promptHeader}${sol}` : void 0;
-    }
-  };
-}
-
-export {
-  loadHumanEval,
-  basePrompt,
-  extractCode,
-  runPythonProgram,
-  runChecker,
-  createHumanEvalAdapter
-};
-//# sourceMappingURL=chunk-5H5XV76F.js.map

package/dist/chunk-5H5XV76F.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/benchmarks/humaneval.ts"],"sourcesContent":["/**\n * HumanEval adapter — the deployable-checker domain as a `BenchmarkAdapter`, so the\n * gate runner (`runGate`) can A/B the STEERING regime on it: a real rollout\n * through the `Supervisor` that self-corrects across rounds, vs blind\n * random@k. This is the experiment `humaneval-gate.mts` names as \"the next one\" —\n * the gate measures SELECTION over stateless single completions; this measures\n * whether observe→steer (self-correction) beats blind compute at equal k.\n *\n * Worker artifact = the model's reply (a Python function, fenced or raw). The\n * DETERMINISTIC judge runs the candidate against the task's own `test` in an\n * isolated `--network=none` python container — exit 0 = pass. No gold\n * `canonical_solution` is ever shown to the model; `goldArtifact` returns it only\n * to self-verify the judge before spending tokens.\n *\n * The primitives (loader / extractor / Docker checker) live here and are reused by\n * `humaneval-gate.mts` — one home, no duplication.\n */\n\nimport { execFile } from 'node:child_process'\nimport { chmodSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from 'node:fs'\nimport { tmpdir } from 'node:os'\nimport { join } from 'node:path'\nimport { gunzipSync } from 'node:zlib'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\n\nconst humanevalUrl = 'https://github.com/openai/human-eval/raw/master/data/HumanEval.jsonl.gz'\nconst dockerImage = 'python:3.12-slim'\nconst dockerTimeoutMs = Number(process.env.DOCKER_TIMEOUT_MS ?? 20000)\n\nexport interface HumanEvalTask {\n  taskId: string\n  prompt: string\n  test: string\n  entryPoint: string\n  /** Reference solution body — used ONLY to self-verify the judge, never shown to the worker. */\n  canonicalSolution?: string\n}\n\n/** Pull the 164-task HumanEval JSONL.gz and parse it. Fail loud on a non-OK fetch\n *  or a malformed line — a silently-short task set would poison the gate. `offset`\n *  selects a deeper slice (the later tasks are harder) so the worker has a\n *  correctable middle band rather than a saturated easy prefix. */\nexport async function loadHumanEval(limit: number, offset = 0): Promise<HumanEvalTask[]> {\n  // Prefer a locally-cached .jsonl.gz (HUMANEVAL_GZ) — the GitHub raw URL rate-limits\n  // (429) under repeated runs. Fall back to the network fetch when unset.\n  const localGz = process.env.HUMANEVAL_GZ\n  let gz: Buffer\n  if (localGz) {\n    gz = readFileSync(localGz)\n  } else {\n    const res = await fetch(humanevalUrl)\n    if (!res.ok) throw new Error(`HumanEval fetch HTTP ${res.status}: ${humanevalUrl}`)\n    gz = Buffer.from(await res.arrayBuffer())\n  }\n  const text = gunzipSync(gz).toString('utf8')\n  const tasks: HumanEvalTask[] = []\n  for (const line of text.split('\\n')) {\n    if (line.trim() === '') continue\n    const d = JSON.parse(line) as {\n      task_id?: string\n      prompt?: string\n      test?: string\n      entry_point?: string\n      canonical_solution?: string\n    }\n    if (!d.task_id || !d.prompt || !d.test || !d.entry_point) {\n      throw new Error(`malformed HumanEval record: ${line.slice(0, 120)}`)\n    }\n    tasks.push({\n      taskId: d.task_id,\n      prompt: d.prompt,\n      test: d.test,\n      entryPoint: d.entry_point,\n      ...(d.canonical_solution ? { canonicalSolution: d.canonical_solution } : {}),\n    })\n  }\n  if (tasks.length === 0) throw new Error('HumanEval parsed to 0 tasks')\n  if (offset >= tasks.length) throw new Error(`OFFSET ${offset} >= dataset size ${tasks.length}`)\n  return tasks.slice(offset, offset + limit)\n}\n\nconst solveInstruction =\n  'Complete the following Python function. Output the COMPLETE function definition (signature, docstring optional, body) inside a single ```python code block. Include any imports the function needs. Do not write tests or example calls.'\n\nexport function basePrompt(task: HumanEvalTask): string {\n  return `${solveInstruction}\\n\\n\\`\\`\\`python\\n${task.prompt}\\`\\`\\``\n}\n\n/** Extract the function source from a model reply: prefer a fenced ```python (or\n *  bare ```) block, else fall back to the raw text. The deployable program adds the\n *  prompt header (imports + signature context), so a candidate that returns only a\n *  body still runs; a candidate that re-defines the function shadows the header. */\nexport function extractCode(reply: string): string {\n  const fenced = reply.match(/```(?:python|py)?\\s*\\n([\\s\\S]*?)```/i)\n  if (fenced && typeof fenced[1] === 'string') return fenced[1].trim()\n  return reply.trim()\n}\n\n/** The deployable test program: the prompt header (imports + signature/docstring the\n *  model was given), then the candidate (its def shadows the header's stub), then the\n *  task's own check() suite and the call. No gold solution anywhere. */\nfunction buildProgram(task: HumanEvalTask, candidate: string): string {\n  return `${task.prompt}\\n${candidate}\\n\\n${task.test}\\n\\ncheck(${task.entryPoint})\\n`\n}\n\nexport interface CheckResult {\n  /** {0,1} pass-count for this candidate (1 = the check() suite passed). */\n  pass: number\n  /** On failure: the interpreter stderr tail (traceback / failing assertion). The\n   *  execution-grounded feedback a self-repair loop steers on; ignored by selection. */\n  detail?: string\n}\n\nexport interface PythonProgramResult {\n  exitCode: number\n  stdout: string\n  stderr: string\n}\n\n/** Run one candidate's deployable test program in an isolated container:\n *  read-only root, bounded writable `/tmp`, no network, one CPU, and 512 MiB.\n *  Exit 0 → pass. A docker invocation error (binary missing, daemon down, image\n *  unavailable) is NOT a test failure — it throws so the harness fails loud rather\n *  than scoring every candidate 0 from a broken checker. */\nlet dockerRunSeq = 0\n\n/** Run arbitrary Python in the same isolated container used by the HumanEval\n * checker. This is also the execution-tool primitive for experiments that let\n * a model test its own snippets; model-written code must never run on the host. */\nexport function runPythonProgram(\n  program: string,\n  timeoutMs = dockerTimeoutMs,\n): Promise<PythonProgramResult> {\n  const dir = mkdtempSync(join(tmpdir(), 'hev-'))\n  // Rootless Docker maps container root to a subordinate host uid that cannot\n  // traverse mkdtemp's default 0700 directory. The directory contains only the\n  // disposable candidate script and is mounted read-only into the container.\n  chmodSync(dir, 0o755)\n  writeFileSync(join(dir, 'p.py'), program, { mode: 0o644 })\n  // Unique container name so we can force-reap it regardless of the docker client's state.\n  const name = `hev-${process.pid}-${dockerRunSeq++}`\n  const startMarker = `__AGENT_RUNTIME_CANDIDATE_STARTED_${name}__`\n  return new Promise<PythonProgramResult>((resolvePromise, reject) => {\n    let settled = false\n    const cleanup = () => {\n      rmSync(dir, { recursive: true, force: true })\n      // `execFile`'s `timeout` kills the docker CLIENT, not the container — a hung\n      // `python` would otherwise pin a CPU forever. Force-reap by name (fire-and-forget;\n      // the name is unique, so no reuse race).\n      execFile('docker', ['rm', '-f', name], () => {})\n    }\n    const finish = (res: PythonProgramResult) => {\n      if (settled) return\n      settled = true\n      clearTimeout(backstop)\n      cleanup()\n      resolvePromise(res)\n    }\n    const fail = (e: Error) => {\n      if (settled) return\n      settled = true\n      clearTimeout(backstop)\n      cleanup()\n      reject(e)\n    }\n    // Candidate timeouts are enforced inside the container. If the outer Docker\n    // client or daemon misses this larger deadline, that is infrastructure failure,\n    // not a wrong answer.\n    const outerTimeoutMs = timeoutMs + 3_000\n    const backstop = setTimeout(\n      () => fail(new Error(`docker checker did not return within ${outerTimeoutMs + 3_000}ms`)),\n      outerTimeoutMs + 3_000,\n    )\n    const inContainerSeconds = Math.max(1, Math.ceil(timeoutMs / 1_000))\n    execFile(\n      'docker',\n      [\n        'run',\n        '--rm',\n        '--pull=never',\n        '--name',\n        name,\n        '--network=none',\n        '--cpus=1',\n        '--memory=512m',\n        '--pids-limit=64',\n        '--cap-drop=ALL',\n        '--security-opt=no-new-privileges',\n        '--read-only',\n        '--tmpfs',\n        '/tmp:rw,nosuid,nodev,noexec,size=64m,mode=1777',\n        '--user',\n        '65534:65534',\n        '-v',\n        `${dir}:/w:ro`,\n        '-w',\n        '/w',\n        dockerImage,\n        'sh',\n        '-c',\n        'command -v timeout >/dev/null 2>&1 && command -v python >/dev/null 2>&1 && [ -r /w/p.py ] || exit 126; printf \"%s\\\\n\" \"$1\"; exec timeout -s KILL \"$2\" python /w/p.py',\n        'agent-runtime-checker',\n        startMarker,\n        `${inContainerSeconds}s`,\n      ],\n      { timeout: outerTimeoutMs, killSignal: 'SIGKILL', maxBuffer: 4 * 1024 * 1024 },\n      (err, stdout, stderr) => {\n        const markerLine = `${startMarker}\\n`\n        const candidateStarted = stdout?.startsWith(markerLine) === true\n        const candidateStdout = candidateStarted ? stdout.slice(markerLine.length) : (stdout ?? '')\n        if (err) {\n          const e = err as Error & { killed?: boolean; code?: number | string }\n          if (e.code === 'ENOENT') {\n            fail(new Error('docker binary not found on PATH — cannot run the deployable checker'))\n            return\n          }\n          if (e.killed) {\n            fail(new Error(`docker checker exceeded its ${outerTimeoutMs}ms outer timeout`))\n            return\n          }\n          if (!candidateStarted) {\n            if (\n              /cannot connect to the docker daemon|is the docker daemon running|permission denied while trying to connect/i.test(\n                stderr,\n              )\n            ) {\n              fail(new Error(`docker daemon unreachable: ${stderr.slice(0, 200)}`))\n              return\n            }\n            if (/unable to find image|no such image|pull access denied|manifest unknown/i.test(stderr)) {\n              fail(new Error(`docker image ${dockerImage} unavailable: ${stderr.slice(0, 200)}`))\n              return\n            }\n            fail(\n              new Error(\n                `docker checker did not start the candidate: ${stderr.slice(0, 200) || e.message}`,\n              ),\n            )\n            return\n          }\n          const exitCode = typeof e.code === 'number' ? e.code : 1\n          finish({\n            exitCode,\n            stdout: candidateStdout,\n            stderr: stderr ?? '',\n          })\n          return\n        }\n        if (!candidateStarted) {\n          fail(new Error('docker checker exited without starting the candidate'))\n          return\n        }\n        finish({ exitCode: 0, stdout: candidateStdout, stderr: stderr ?? '' })\n      },\n    )\n  })\n}\n\nexport async function runChecker(task: HumanEvalTask, candidate: string): Promise<CheckResult> {\n  const result = await runPythonProgram(buildProgram(task, candidate))\n  return result.exitCode === 0\n    ? { pass: 1 }\n    : { pass: 0, detail: result.stderr.slice(-600) || 'timed out (no output)' }\n}\n\n/** A HumanEval task carries its checker inputs in metadata so the deterministic\n *  judge can rebuild the deployable program from a `BenchTask` alone. */\ninterface HumanEvalMeta extends Record<string, unknown> {\n  promptHeader: string\n  test: string\n  entryPoint: string\n  canonicalSolution?: string\n}\n\nfunction toBenchTask(t: HumanEvalTask): BenchTask {\n  const metadata: HumanEvalMeta = {\n    promptHeader: t.prompt,\n    test: t.test,\n    entryPoint: t.entryPoint,\n    ...(t.canonicalSolution ? { canonicalSolution: t.canonicalSolution } : {}),\n  }\n  return { id: t.taskId, prompt: basePrompt(t), metadata }\n}\n\nfunction taskFromMeta(task: BenchTask): HumanEvalTask {\n  const m = task.metadata as HumanEvalMeta | undefined\n  if (!m?.promptHeader || !m.test || !m.entryPoint) {\n    throw new Error(`HumanEval judge: task ${task.id} missing checker metadata`)\n  }\n  return { taskId: task.id, prompt: m.promptHeader, test: m.test, entryPoint: m.entryPoint }\n}\n\n/** The HumanEval `BenchmarkAdapter`. OFFSET (env) selects the correctable middle\n *  band; loadTasks honors `limit`/`ids`. The judge is the Docker deployable checker. */\nexport function createHumanEvalAdapter(): BenchmarkAdapter {\n  return {\n    name: 'humaneval',\n    async preflight() {\n      // Prove the daemon and exact local image before any model call. Scoring uses\n      // --pull=never so a registry or cold pull cannot become a candidate outcome.\n      await new Promise<void>((resolve, reject) => {\n        execFile('docker', ['version', '--format', '{{.Server.Version}}'], (err) => {\n          if (err) reject(new Error('HumanEval judge needs a running Docker daemon (python:3.12-slim, --network=none)'))\n          else resolve()\n        })\n      })\n      await new Promise<void>((resolve, reject) => {\n        execFile('docker', ['image', 'inspect', dockerImage], (err) => {\n          if (err) reject(new Error(`HumanEval judge needs the cached Docker image ${dockerImage}`))\n          else resolve()\n        })\n      })\n    },\n    async loadTasks(opts?: LoadOptions) {\n      const offset = Number(process.env.OFFSET ?? 0)\n      // Pull a generous window when filtering by id, else exactly `limit` from offset.\n      const all = await loadHumanEval(opts?.ids ? 164 : (opts?.limit ?? 8), offset)\n      const picked = opts?.ids ? all.filter((t) => opts.ids?.includes(t.taskId)) : all\n      return picked.map(toBenchTask)\n    },\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const { pass } = await runChecker(taskFromMeta(task), extractCode(artifact))\n      return { resolved: pass === 1, score: pass, detail: pass === 1 ? 'tests passed' : 'tests failed' }\n    },\n    async goldArtifact(task: BenchTask) {\n      const m = task.metadata as HumanEvalMeta | undefined\n      const sol = m?.canonicalSolution\n      // Return the COMPLETE function (signature header + canonical body), i.e. what a\n      // real worker emits — NOT the body alone. The judge runs `extractCode`, whose\n      // unfenced fallback is `reply.trim()`; trimming a body-only string strips its\n      // leading indent and breaks it, so a body-only gold fails its own judge. A full\n      // def starts at column 0, trims safely, and self-verifies.\n      return sol ? `${m!.promptHeader}${sol}` : undefined\n    },\n  }\n}\n"],"mappings":";AAkBA,SAAS,gBAAgB;AACzB,SAAS,WAAW,aAAa,cAAc,QAAQ,qBAAqB;AAC5E,SAAS,cAAc;AACvB,SAAS,YAAY;AACrB,SAAS,kBAAkB;AAG3B,IAAM,eAAe;AACrB,IAAM,cAAc;AACpB,IAAM,kBAAkB,OAAO,QAAQ,IAAI,qBAAqB,GAAK;AAerE,eAAsB,cAAc,OAAe,SAAS,GAA6B;AAGvF,QAAM,UAAU,QAAQ,IAAI;AAC5B,MAAI;AACJ,MAAI,SAAS;AACX,SAAK,aAAa,OAAO;AAAA,EAC3B,OAAO;AACL,UAAM,MAAM,MAAM,MAAM,YAAY;AACpC,QAAI,CAAC,IAAI,GAAI,OAAM,IAAI,MAAM,wBAAwB,IAAI,MAAM,KAAK,YAAY,EAAE;AAClF,SAAK,OAAO,KAAK,MAAM,IAAI,YAAY,CAAC;AAAA,EAC1C;AACA,QAAM,OAAO,WAAW,EAAE,EAAE,SAAS,MAAM;AAC3C,QAAM,QAAyB,CAAC;AAChC,aAAW,QAAQ,KAAK,MAAM,IAAI,GAAG;AACnC,QAAI,KAAK,KAAK,MAAM,GAAI;AACxB,UAAM,IAAI,KAAK,MAAM,IAAI;AAOzB,QAAI,CAAC,EAAE,WAAW,CAAC,EAAE,UAAU,CAAC,EAAE,QAAQ,CAAC,EAAE,aAAa;AACxD,YAAM,IAAI,MAAM,+BAA+B,KAAK,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,IACrE;AACA,UAAM,KAAK;AAAA,MACT,QAAQ,EAAE;AAAA,MACV,QAAQ,EAAE;AAAA,MACV,MAAM,EAAE;AAAA,MACR,YAAY,EAAE;AAAA,MACd,GAAI,EAAE,qBAAqB,EAAE,mBAAmB,EAAE,mBAAmB,IAAI,CAAC;AAAA,IAC5E,CAAC;AAAA,EACH;AACA,MAAI,MAAM,WAAW,EAAG,OAAM,IAAI,MAAM,6BAA6B;AACrE,MAAI,UAAU,MAAM,OAAQ,OAAM,IAAI,MAAM,UAAU,MAAM,oBAAoB,MAAM,MAAM,EAAE;AAC9F,SAAO,MAAM,MAAM,QAAQ,SAAS,KAAK;AAC3C;AAEA,IAAM,mBACJ;AAEK,SAAS,WAAW,MAA6B;AACtD,SAAO,GAAG,gBAAgB;AAAA;AAAA;AAAA,EAAqB,KAAK,MAAM;AAC5D;AAMO,SAAS,YAAY,OAAuB;AACjD,QAAM,SAAS,MAAM,MAAM,sCAAsC;AACjE,MAAI,UAAU,OAAO,OAAO,CAAC,MAAM,SAAU,QAAO,OAAO,CAAC,EAAE,KAAK;AACnE,SAAO,MAAM,KAAK;AACpB;AAKA,SAAS,aAAa,MAAqB,WAA2B;AACpE,SAAO,GAAG,KAAK,MAAM;AAAA,EAAK,SAAS;AAAA;AAAA,EAAO,KAAK,IAAI;AAAA;AAAA,QAAa,KAAK,UAAU;AAAA;AACjF;AAqBA,IAAI,eAAe;AAKZ,SAAS,iBACd,SACA,YAAY,iBACkB;AAC9B,QAAM,MAAM,YAAY,KAAK,OAAO,GAAG,MAAM,CAAC;AAI9C,YAAU,KAAK,GAAK;AACpB,gBAAc,KAAK,KAAK,MAAM,GAAG,SAAS,EAAE,MAAM,IAAM,CAAC;AAEzD,QAAM,OAAO,OAAO,QAAQ,GAAG,IAAI,cAAc;AACjD,QAAM,cAAc,qCAAqC,IAAI;AAC7D,SAAO,IAAI,QAA6B,CAAC,gBAAgB,WAAW;AAClE,QAAI,UAAU;AACd,UAAM,UAAU,MAAM;AACpB,aAAO,KAAK,EAAE,WAAW,MAAM,OAAO,KAAK,CAAC;AAI5C,eAAS,UAAU,CAAC,MAAM,MAAM,IAAI,GAAG,MAAM;AAAA,MAAC,CAAC;AAAA,IACjD;AACA,UAAM,SAAS,CAAC,QAA6B;AAC3C,UAAI,QAAS;AACb,gBAAU;AACV,mBAAa,QAAQ;AACrB,cAAQ;AACR,qBAAe,GAAG;AAAA,IACpB;AACA,UAAM,OAAO,CAAC,MAAa;AACzB,UAAI,QAAS;AACb,gBAAU;AACV,mBAAa,QAAQ;AACrB,cAAQ;AACR,aAAO,CAAC;AAAA,IACV;AAIA,UAAM,iBAAiB,YAAY;AACnC,UAAM,WAAW;AAAA,MACf,MAAM,KAAK,IAAI,MAAM,wCAAwC,iBAAiB,GAAK,IAAI,CAAC;AAAA,MACxF,iBAAiB;AAAA,IACnB;AACA,UAAM,qBAAqB,KAAK,IAAI,GAAG,KAAK,KAAK,YAAY,GAAK,CAAC;AACnE;AAAA,MACE;AAAA,MACA;AAAA,QACE;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,GAAG,GAAG;AAAA,QACN;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,GAAG,kBAAkB;AAAA,MACvB;AAAA,MACA,EAAE,SAAS,gBAAgB,YAAY,WAAW,WAAW,IAAI,OAAO,KAAK;AAAA,MAC7E,CAAC,KAAK,QAAQ,WAAW;AACvB,cAAM,aAAa,GAAG,WAAW;AAAA;AACjC,cAAM,mBAAmB,QAAQ,WAAW,UAAU,MAAM;AAC5D,cAAM,kBAAkB,mBAAmB,OAAO,MAAM,WAAW,MAAM,IAAK,UAAU;AACxF,YAAI,KAAK;AACP,gBAAM,IAAI;AACV,cAAI,EAAE,SAAS,UAAU;AACvB,iBAAK,IAAI,MAAM,0EAAqE,CAAC;AACrF;AAAA,UACF;AACA,cAAI,EAAE,QAAQ;AACZ,iBAAK,IAAI,MAAM,+BAA+B,cAAc,kBAAkB,CAAC;AAC/E;AAAA,UACF;AACA,cAAI,CAAC,kBAAkB;AACrB,gBACE,8GAA8G;AAAA,cAC5G;AAAA,YACF,GACA;AACA,mBAAK,IAAI,MAAM,8BAA8B,OAAO,MAAM,GAAG,GAAG,CAAC,EAAE,CAAC;AACpE;AAAA,YACF;AACA,gBAAI,0EAA0E,KAAK,MAAM,GAAG;AAC1F,mBAAK,IAAI,MAAM,gBAAgB,WAAW,iBAAiB,OAAO,MAAM,GAAG,GAAG,CAAC,EAAE,CAAC;AAClF;AAAA,YACF;AACA;AAAA,cACE,IAAI;AAAA,gBACF,+CAA+C,OAAO,MAAM,GAAG,GAAG,KAAK,EAAE,OAAO;AAAA,cAClF;AAAA,YACF;AACA;AAAA,UACF;AACA,gBAAM,WAAW,OAAO,EAAE,SAAS,WAAW,EAAE,OAAO;AACvD,iBAAO;AAAA,YACL;AAAA,YACA,QAAQ;AAAA,YACR,QAAQ,UAAU;AAAA,UACpB,CAAC;AACD;AAAA,QACF;AACA,YAAI,CAAC,kBAAkB;AACrB,eAAK,IAAI,MAAM,sDAAsD,CAAC;AACtE;AAAA,QACF;AACA,eAAO,EAAE,UAAU,GAAG,QAAQ,iBAAiB,QAAQ,UAAU,GAAG,CAAC;AAAA,MACvE;AAAA,IACF;AAAA,EACF,CAAC;AACH;AAEA,eAAsB,WAAW,MAAqB,WAAyC;AAC7F,QAAM,SAAS,MAAM,iBAAiB,aAAa,MAAM,SAAS,CAAC;AACnE,SAAO,OAAO,aAAa,IACvB,EAAE,MAAM,EAAE,IACV,EAAE,MAAM,GAAG,QAAQ,OAAO,OAAO,MAAM,IAAI,KAAK,wBAAwB;AAC9E;AAWA,SAAS,YAAY,GAA6B;AAChD,QAAM,WAA0B;AAAA,IAC9B,cAAc,EAAE;AAAA,IAChB,MAAM,EAAE;AAAA,IACR,YAAY,EAAE;AAAA,IACd,GAAI,EAAE,oBAAoB,EAAE,mBAAmB,EAAE,kBAAkB,IAAI,CAAC;AAAA,EAC1E;AACA,SAAO,EAAE,IAAI,EAAE,QAAQ,QAAQ,WAAW,CAAC,GAAG,SAAS;AACzD;AAEA,SAAS,aAAa,MAAgC;AACpD,QAAM,IAAI,KAAK;AACf,MAAI,CAAC,GAAG,gBAAgB,CAAC,EAAE,QAAQ,CAAC,EAAE,YAAY;AAChD,UAAM,IAAI,MAAM,yBAAyB,KAAK,EAAE,2BAA2B;AAAA,EAC7E;AACA,SAAO,EAAE,QAAQ,KAAK,IAAI,QAAQ,EAAE,cAAc,MAAM,EAAE,MAAM,YAAY,EAAE,WAAW;AAC3F;AAIO,SAAS,yBAA2C;AACzD,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,YAAY;AAGhB,YAAM,IAAI,QAAc,CAAC,SAAS,WAAW;AAC3C,iBAAS,UAAU,CAAC,WAAW,YAAY,qBAAqB,GAAG,CAAC,QAAQ;AAC1E,cAAI,IAAK,QAAO,IAAI,MAAM,kFAAkF,CAAC;AAAA,cACxG,SAAQ;AAAA,QACf,CAAC;AAAA,MACH,CAAC;AACD,YAAM,IAAI,QAAc,CAAC,SAAS,WAAW;AAC3C,iBAAS,UAAU,CAAC,SAAS,WAAW,WAAW,GAAG,CAAC,QAAQ;AAC7D,cAAI,IAAK,QAAO,IAAI,MAAM,iDAAiD,WAAW,EAAE,CAAC;AAAA,cACpF,SAAQ;AAAA,QACf,CAAC;AAAA,MACH,CAAC;AAAA,IACH;AAAA,IACA,MAAM,UAAU,MAAoB;AAClC,YAAM,SAAS,OAAO,QAAQ,IAAI,UAAU,CAAC;AAE7C,YAAM,MAAM,MAAM,cAAc,MAAM,MAAM,MAAO,MAAM,SAAS,GAAI,MAAM;AAC5E,YAAM,SAAS,MAAM,MAAM,IAAI,OAAO,CAAC,MAAM,KAAK,KAAK,SAAS,EAAE,MAAM,CAAC,IAAI;AAC7E,aAAO,OAAO,IAAI,WAAW;AAAA,IAC/B;AAAA,IACA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,EAAE,KAAK,IAAI,MAAM,WAAW,aAAa,IAAI,GAAG,YAAY,QAAQ,CAAC;AAC3E,aAAO,EAAE,UAAU,SAAS,GAAG,OAAO,MAAM,QAAQ,SAAS,IAAI,iBAAiB,eAAe;AAAA,IACnG;AAAA,IACA,MAAM,aAAa,MAAiB;AAClC,YAAM,IAAI,KAAK;AACf,YAAM,MAAM,GAAG;AAMf,aAAO,MAAM,GAAG,EAAG,YAAY,GAAG,GAAG,KAAK;AAAA,IAC5C;AAAA,EACF;AACF;","names":[]}

package/dist/chunk-67ACKDCX.js

@@ -1,118 +0,0 @@
-import {
-  benchRoot
-} from "./chunk-EIETHPD5.js";
-
-// src/benchmarks/agentbench.ts
-import { readFile } from "fs/promises";
-import { join } from "path";
-var FIXTURES = join(benchRoot, "fixtures", "agentbench-dbbench.json");
-var DEFAULT_SPLIT = "dev";
-var agentbenchDir = () => process.env.AGENTBENCH_DIR;
-var agentbenchAnswerOutput = {
-  parse(events) {
-    let text = "";
-    for (const ev of events) {
-      const d = ev?.data;
-      const t = d?.finalText ?? d?.text ?? d?.result;
-      if (typeof t === "string" && t.length > 0) text = t;
-    }
-    const fences = [...text.matchAll(/```(?:text|answer)?\s*\n([\s\S]*?)```/g)];
-    return (fences.at(-1)?.[1] ?? text).trim();
-  }
-};
-function normalizeAnswer(value) {
-  return value.toLowerCase().replace(/\s+/g, " ").replace(/^["'`]+|["'`]+$/g, "").trim();
-}
-function rowToTask(row, index, split) {
-  const columns = row.table?.table_info?.columns?.map((c) => `${c.name}${c.type ? ` (${c.type})` : ""}`).join(", ");
-  const sampleRows = row.table?.table_info?.rows?.slice(0, 40);
-  const meta = {
-    labels: row.label,
-    split,
-    subset: "dbbench",
-    table: row.table
-  };
-  return {
-    id: `dbbench-${split}-${index}`,
-    split,
-    prompt: [
-      "Answer this AgentBench DBBench question using the table below.",
-      "Return only the answer value.",
-      "",
-      `Question: ${row.description}`,
-      row.table?.table_name ? `Table: ${row.table.table_name}` : void 0,
-      columns ? `Columns: ${columns}` : void 0,
-      sampleRows ? `Rows JSON: ${JSON.stringify(sampleRows)}` : void 0
-    ].filter(Boolean).join("\n"),
-    metadata: meta
-  };
-}
-function readMeta(task) {
-  const md = task.metadata;
-  if (!md || !Array.isArray(md.labels)) {
-    throw new Error(`agentbench task ${task.id} missing metadata \u2014 loadTasks did not populate it`);
-  }
-  return md;
-}
-function selectRows(rows, opts, split) {
-  let tasks = rows.map((row, index) => rowToTask(row, index, split));
-  if (opts.ids) {
-    const want = new Set(opts.ids);
-    tasks = tasks.filter((task) => want.has(task.id));
-  } else if (opts.limit !== void 0) {
-    tasks = tasks.slice(0, opts.limit);
-  }
-  if (tasks.length === 0) throw new Error(`AgentBench DBBench: no tasks matched ${JSON.stringify(opts)}`);
-  return tasks;
-}
-async function loadJsonl(path) {
-  const raw = await readFile(path, "utf8");
-  return raw.split("\n").map((line) => line.trim()).filter(Boolean).map((line) => JSON.parse(line));
-}
-async function loadFixtures(opts, split) {
-  const rows = JSON.parse(await readFile(FIXTURES, "utf8"));
-  console.warn(`[agentbench] AGENTBENCH_FIXTURES=1 \u2014 loading ${rows.length} DBBench adapter fixtures`);
-  return selectRows(rows, opts, split);
-}
-function createAgentBenchAdapter() {
-  const fixturesMode = process.env.AGENTBENCH_FIXTURES === "1";
-  return {
-    name: "agentbench",
-    output: agentbenchAnswerOutput,
-    async preflight() {
-      if (fixturesMode) return;
-      const dir = agentbenchDir();
-      if (!dir) {
-        throw new Error("AGENTBENCH_DIR is required. Fix: clone https://github.com/THUDM/AgentBench and set AGENTBENCH_DIR=/path/to/AgentBench.");
-      }
-      await loadJsonl(join(dir, "data", "dbbench", `${DEFAULT_SPLIT}.jsonl`));
-    },
-    async loadTasks(opts = {}) {
-      const split = opts.split ?? DEFAULT_SPLIT;
-      if (fixturesMode) return loadFixtures(opts, split);
-      const dir = agentbenchDir();
-      if (!dir) throw new Error("AGENTBENCH_DIR is required to load AgentBench DBBench tasks");
-      return selectRows(await loadJsonl(join(dir, "data", "dbbench", `${split}.jsonl`)), opts, split);
-    },
-    async goldArtifact(task) {
-      return readMeta(task).labels[0];
-    },
-    async judge(task, artifact) {
-      const meta = readMeta(task);
-      const got = normalizeAnswer(artifact);
-      const expected = meta.labels.map(normalizeAnswer);
-      const resolved = expected.includes(got);
-      return {
-        resolved,
-        score: resolved ? 1 : 0,
-        detail: JSON.stringify({ subset: meta.subset, split: meta.split, expected: meta.labels, got: artifact })
-      };
-    }
-  };
-}
-
-export {
-  agentbenchAnswerOutput,
-  createAgentBenchAdapter
-};
-//# sourceMappingURL=chunk-67ACKDCX.js.map

package/dist/chunk-67ACKDCX.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/benchmarks/agentbench.ts"],"sourcesContent":["/**\n * AgentBench deterministic subset adapter.\n *\n * This targets AgentBench DBBench rows only: question + table + published label.\n * It does not wrap AgentBench's controller protocol or the non-deterministic game\n * environments. Worker artifact = final answer text. Judge = exact match against\n * the official DBBench label list after light whitespace/case normalization.\n */\n\nimport { readFile } from 'node:fs/promises'\nimport { join } from 'node:path'\nimport type { OutputAdapter } from '@tangle-network/agent-runtime/loops'\nimport { benchRoot } from './_harness'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\n\nconst FIXTURES = join(benchRoot, 'fixtures', 'agentbench-dbbench.json')\nconst DEFAULT_SPLIT = 'dev'\n\ninterface AgentBenchDbRow {\n  description: string\n  label: string[]\n  table?: {\n    table_name?: string\n    table_info?: {\n      columns?: Array<{ name: string; type?: string }>\n      rows?: unknown[][]\n    }\n  }\n}\n\ninterface AgentBenchMeta {\n  labels: string[]\n  split: string\n  subset: 'dbbench'\n  table?: AgentBenchDbRow['table']\n}\n\nconst agentbenchDir = (): string | undefined => process.env.AGENTBENCH_DIR\n\nexport const agentbenchAnswerOutput: OutputAdapter<string> = {\n  parse(events) {\n    let text = ''\n    for (const ev of events) {\n      const d = (ev as { data?: Record<string, unknown> })?.data\n      const t = d?.finalText ?? d?.text ?? d?.result\n      if (typeof t === 'string' && t.length > 0) text = t\n    }\n    const fences = [...text.matchAll(/```(?:text|answer)?\\s*\\n([\\s\\S]*?)```/g)]\n    return (fences.at(-1)?.[1] ?? text).trim()\n  },\n}\n\nfunction normalizeAnswer(value: string): string {\n  return value\n    .toLowerCase()\n    .replace(/\\s+/g, ' ')\n    .replace(/^[\"'`]+|[\"'`]+$/g, '')\n    .trim()\n}\n\nfunction rowToTask(row: AgentBenchDbRow, index: number, split: string): BenchTask {\n  const columns = row.table?.table_info?.columns?.map((c) => `${c.name}${c.type ? ` (${c.type})` : ''}`).join(', ')\n  const sampleRows = row.table?.table_info?.rows?.slice(0, 40)\n  const meta: AgentBenchMeta = {\n    labels: row.label,\n    split,\n    subset: 'dbbench',\n    table: row.table,\n  }\n  return {\n    id: `dbbench-${split}-${index}`,\n    split,\n    prompt: [\n      'Answer this AgentBench DBBench question using the table below.',\n      'Return only the answer value.',\n      '',\n      `Question: ${row.description}`,\n      row.table?.table_name ? `Table: ${row.table.table_name}` : undefined,\n      columns ? `Columns: ${columns}` : undefined,\n      sampleRows ? `Rows JSON: ${JSON.stringify(sampleRows)}` : undefined,\n    ]\n      .filter(Boolean)\n      .join('\\n'),\n    metadata: meta as unknown as Record<string, unknown>,\n  }\n}\n\nfunction readMeta(task: BenchTask): AgentBenchMeta {\n  const md = task.metadata\n  if (!md || !Array.isArray(md.labels)) {\n    throw new Error(`agentbench task ${task.id} missing metadata — loadTasks did not populate it`)\n  }\n  return md as unknown as AgentBenchMeta\n}\n\nfunction selectRows(rows: AgentBenchDbRow[], opts: LoadOptions, split: string): BenchTask[] {\n  let tasks = rows.map((row, index) => rowToTask(row, index, split))\n  if (opts.ids) {\n    const want = new Set(opts.ids)\n    tasks = tasks.filter((task) => want.has(task.id))\n  } else if (opts.limit !== undefined) {\n    tasks = tasks.slice(0, opts.limit)\n  }\n  if (tasks.length === 0) throw new Error(`AgentBench DBBench: no tasks matched ${JSON.stringify(opts)}`)\n  return tasks\n}\n\nasync function loadJsonl(path: string): Promise<AgentBenchDbRow[]> {\n  const raw = await readFile(path, 'utf8')\n  return raw\n    .split('\\n')\n    .map((line) => line.trim())\n    .filter(Boolean)\n    .map((line) => JSON.parse(line) as AgentBenchDbRow)\n}\n\nasync function loadFixtures(opts: LoadOptions, split: string): Promise<BenchTask[]> {\n  const rows = JSON.parse(await readFile(FIXTURES, 'utf8')) as AgentBenchDbRow[]\n  console.warn(`[agentbench] AGENTBENCH_FIXTURES=1 — loading ${rows.length} DBBench adapter fixtures`)\n  return selectRows(rows, opts, split)\n}\n\nexport function createAgentBenchAdapter(): BenchmarkAdapter {\n  const fixturesMode = process.env.AGENTBENCH_FIXTURES === '1'\n\n  return {\n    name: 'agentbench',\n    output: agentbenchAnswerOutput,\n\n    async preflight() {\n      if (fixturesMode) return\n      const dir = agentbenchDir()\n      if (!dir) {\n        throw new Error('AGENTBENCH_DIR is required. Fix: clone https://github.com/THUDM/AgentBench and set AGENTBENCH_DIR=/path/to/AgentBench.')\n      }\n      await loadJsonl(join(dir, 'data', 'dbbench', `${DEFAULT_SPLIT}.jsonl`))\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      const split = opts.split ?? DEFAULT_SPLIT\n      if (fixturesMode) return loadFixtures(opts, split)\n      const dir = agentbenchDir()\n      if (!dir) throw new Error('AGENTBENCH_DIR is required to load AgentBench DBBench tasks')\n      return selectRows(await loadJsonl(join(dir, 'data', 'dbbench', `${split}.jsonl`)), opts, split)\n    },\n\n    async goldArtifact(task: BenchTask) {\n      return readMeta(task).labels[0]\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const meta = readMeta(task)\n      const got = normalizeAnswer(artifact)\n      const expected = meta.labels.map(normalizeAnswer)\n      const resolved = expected.includes(got)\n      return {\n        resolved,\n        score: resolved ? 1 : 0,\n        detail: JSON.stringify({ subset: meta.subset, split: meta.split, expected: meta.labels, got: artifact }),\n      }\n    },\n  }\n}\n"],"mappings":";;;;;AASA,SAAS,gBAAgB;AACzB,SAAS,YAAY;AAKrB,IAAM,WAAW,KAAK,WAAW,YAAY,yBAAyB;AACtE,IAAM,gBAAgB;AAqBtB,IAAM,gBAAgB,MAA0B,QAAQ,IAAI;AAErD,IAAM,yBAAgD;AAAA,EAC3D,MAAM,QAAQ;AACZ,QAAI,OAAO;AACX,eAAW,MAAM,QAAQ;AACvB,YAAM,IAAK,IAA2C;AACtD,YAAM,IAAI,GAAG,aAAa,GAAG,QAAQ,GAAG;AACxC,UAAI,OAAO,MAAM,YAAY,EAAE,SAAS,EAAG,QAAO;AAAA,IACpD;AACA,UAAM,SAAS,CAAC,GAAG,KAAK,SAAS,wCAAwC,CAAC;AAC1E,YAAQ,OAAO,GAAG,EAAE,IAAI,CAAC,KAAK,MAAM,KAAK;AAAA,EAC3C;AACF;AAEA,SAAS,gBAAgB,OAAuB;AAC9C,SAAO,MACJ,YAAY,EACZ,QAAQ,QAAQ,GAAG,EACnB,QAAQ,oBAAoB,EAAE,EAC9B,KAAK;AACV;AAEA,SAAS,UAAU,KAAsB,OAAe,OAA0B;AAChF,QAAM,UAAU,IAAI,OAAO,YAAY,SAAS,IAAI,CAAC,MAAM,GAAG,EAAE,IAAI,GAAG,EAAE,OAAO,KAAK,EAAE,IAAI,MAAM,EAAE,EAAE,EAAE,KAAK,IAAI;AAChH,QAAM,aAAa,IAAI,OAAO,YAAY,MAAM,MAAM,GAAG,EAAE;AAC3D,QAAM,OAAuB;AAAA,IAC3B,QAAQ,IAAI;AAAA,IACZ;AAAA,IACA,QAAQ;AAAA,IACR,OAAO,IAAI;AAAA,EACb;AACA,SAAO;AAAA,IACL,IAAI,WAAW,KAAK,IAAI,KAAK;AAAA,IAC7B;AAAA,IACA,QAAQ;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA,aAAa,IAAI,WAAW;AAAA,MAC5B,IAAI,OAAO,aAAa,UAAU,IAAI,MAAM,UAAU,KAAK;AAAA,MAC3D,UAAU,YAAY,OAAO,KAAK;AAAA,MAClC,aAAa,cAAc,KAAK,UAAU,UAAU,CAAC,KAAK;AAAA,IAC5D,EACG,OAAO,OAAO,EACd,KAAK,IAAI;AAAA,IACZ,UAAU;AAAA,EACZ;AACF;AAEA,SAAS,SAAS,MAAiC;AACjD,QAAM,KAAK,KAAK;AAChB,MAAI,CAAC,MAAM,CAAC,MAAM,QAAQ,GAAG,MAAM,GAAG;AACpC,UAAM,IAAI,MAAM,mBAAmB,KAAK,EAAE,wDAAmD;AAAA,EAC/F;AACA,SAAO;AACT;AAEA,SAAS,WAAW,MAAyB,MAAmB,OAA4B;AAC1F,MAAI,QAAQ,KAAK,IAAI,CAAC,KAAK,UAAU,UAAU,KAAK,OAAO,KAAK,CAAC;AACjE,MAAI,KAAK,KAAK;AACZ,UAAM,OAAO,IAAI,IAAI,KAAK,GAAG;AAC7B,YAAQ,MAAM,OAAO,CAAC,SAAS,KAAK,IAAI,KAAK,EAAE,CAAC;AAAA,EAClD,WAAW,KAAK,UAAU,QAAW;AACnC,YAAQ,MAAM,MAAM,GAAG,KAAK,KAAK;AAAA,EACnC;AACA,MAAI,MAAM,WAAW,EAAG,OAAM,IAAI,MAAM,wCAAwC,KAAK,UAAU,IAAI,CAAC,EAAE;AACtG,SAAO;AACT;AAEA,eAAe,UAAU,MAA0C;AACjE,QAAM,MAAM,MAAM,SAAS,MAAM,MAAM;AACvC,SAAO,IACJ,MAAM,IAAI,EACV,IAAI,CAAC,SAAS,KAAK,KAAK,CAAC,EACzB,OAAO,OAAO,EACd,IAAI,CAAC,SAAS,KAAK,MAAM,IAAI,CAAoB;AACtD;AAEA,eAAe,aAAa,MAAmB,OAAqC;AAClF,QAAM,OAAO,KAAK,MAAM,MAAM,SAAS,UAAU,MAAM,CAAC;AACxD,UAAQ,KAAK,qDAAgD,KAAK,MAAM,2BAA2B;AACnG,SAAO,WAAW,MAAM,MAAM,KAAK;AACrC;AAEO,SAAS,0BAA4C;AAC1D,QAAM,eAAe,QAAQ,IAAI,wBAAwB;AAEzD,SAAO;AAAA,IACL,MAAM;AAAA,IACN,QAAQ;AAAA,IAER,MAAM,YAAY;AAChB,UAAI,aAAc;AAClB,YAAM,MAAM,cAAc;AAC1B,UAAI,CAAC,KAAK;AACR,cAAM,IAAI,MAAM,wHAAwH;AAAA,MAC1I;AACA,YAAM,UAAU,KAAK,KAAK,QAAQ,WAAW,GAAG,aAAa,QAAQ,CAAC;AAAA,IACxE;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;AACtC,YAAM,QAAQ,KAAK,SAAS;AAC5B,UAAI,aAAc,QAAO,aAAa,MAAM,KAAK;AACjD,YAAM,MAAM,cAAc;AAC1B,UAAI,CAAC,IAAK,OAAM,IAAI,MAAM,6DAA6D;AACvF,aAAO,WAAW,MAAM,UAAU,KAAK,KAAK,QAAQ,WAAW,GAAG,KAAK,QAAQ,CAAC,GAAG,MAAM,KAAK;AAAA,IAChG;AAAA,IAEA,MAAM,aAAa,MAAiB;AAClC,aAAO,SAAS,IAAI,EAAE,OAAO,CAAC;AAAA,IAChC;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,OAAO,SAAS,IAAI;AAC1B,YAAM,MAAM,gBAAgB,QAAQ;AACpC,YAAM,WAAW,KAAK,OAAO,IAAI,eAAe;AAChD,YAAM,WAAW,SAAS,SAAS,GAAG;AACtC,aAAO;AAAA,QACL;AAAA,QACA,OAAO,WAAW,IAAI;AAAA,QACtB,QAAQ,KAAK,UAAU,EAAE,QAAQ,KAAK,QAAQ,OAAO,KAAK,OAAO,UAAU,KAAK,QAAQ,KAAK,SAAS,CAAC;AAAA,MACzG;AAAA,IACF;AAAA,EACF;AACF;","names":[]}

package/dist/chunk-7FKBWOQT.js

@@ -1,147 +0,0 @@
-import {
-  contextBlock,
-  contextsFrom,
-  firstString,
-  isObject,
-  ragAnswerOutput,
-  readJsonRows,
-  selectTasks,
-  stringFrom
-} from "./chunk-X3BTXCJ4.js";
-import {
-  benchRoot
-} from "./chunk-EIETHPD5.js";
-
-// src/benchmarks/nomiracl.ts
-import { readFile } from "fs/promises";
-import { join } from "path";
-var FIXTURES = join(benchRoot, "fixtures", "nomiracl.json");
-var dataFile = () => process.env.NOMIRACL_DATA_FILE;
-function passagesFrom(value, relevant) {
-  return contextsFrom(value).map((context) => ({ ...context, relevant }));
-}
-function answerableFrom(raw, contexts) {
-  const direct = raw.answerable ?? raw.is_answerable ?? raw.has_answer ?? raw.label ?? raw.relevant;
-  if (typeof direct === "boolean") return direct;
-  if (typeof direct === "number") return direct > 0;
-  if (typeof direct === "string") {
-    const normalized = direct.toLowerCase();
-    if (["true", "t", "relevant", "answerable", "positive", "1", "yes"].includes(normalized)) return true;
-    if (["false", "f", "non-relevant", "non_relevant", "unanswerable", "negative", "0", "no"].includes(normalized)) {
-      return false;
-    }
-  }
-  return contexts.some((context) => context.relevant === true);
-}
-function rowToTask(raw, index) {
-  if (!isObject(raw)) throw new Error(`NoMIRACL row ${index} must be an object`);
-  const query = firstString(raw, ["query", "question"]);
-  if (!query) throw new Error(`NoMIRACL row ${index} missing query`);
-  const labelledPassages = [
-    ...passagesFrom(raw.positive_passages, true),
-    ...passagesFrom(raw.negative_passages, false)
-  ];
-  const contexts = labelledPassages.length > 0 ? labelledPassages : contextsFrom(raw.passages).length > 0 ? contextsFrom(raw.passages) : contextsFrom(raw.contexts).length > 0 ? contextsFrom(raw.contexts) : contextsFrom(raw.documents);
-  const answerable = answerableFrom(raw, contexts);
-  const language = stringFrom(raw.lang) ?? stringFrom(raw.language) ?? "unknown";
-  const subset = stringFrom(raw.subset) ?? (answerable ? "relevant" : "non-relevant");
-  const id = stringFrom(raw.id) ?? stringFrom(raw.query_id) ?? `nomiracl-${index}`;
-  const meta = {
-    benchmark: "nomiracl",
-    query,
-    answerable,
-    language,
-    subset,
-    contexts
-  };
-  return {
-    id,
-    split: stringFrom(raw.split) ?? language,
-    prompt: [
-      "Classify this NoMIRACL RAG relevance case.",
-      "Return exactly one word: ANSWERABLE if at least one supplied passage contains enough evidence to answer the query, otherwise UNANSWERABLE.",
-      "",
-      `Query: ${query}`,
-      contexts.length > 0 ? `
-Passages:
-${contextBlock(contexts)}` : void 0
-    ].filter(Boolean).join("\n"),
-    metadata: meta
-  };
-}
-function readMeta(task) {
-  const md = task.metadata;
-  if (!md || typeof md.answerable !== "boolean") {
-    throw new Error(`NoMIRACL task ${task.id} missing metadata \u2014 loadTasks did not populate it`);
-  }
-  return md;
-}
-function parseClassification(artifact) {
-  const normalized = artifact.toLowerCase().replace(/[^a-z0-9]+/g, " ").trim();
-  if (/\bunanswerable\b|\bnot answerable\b|\bno answer\b|\bnegative\b|\bfalse\b/.test(normalized)) {
-    return false;
-  }
-  if (/\banswerable\b|\brelevant\b|\bpositive\b|\btrue\b|\byes\b/.test(normalized)) return true;
-  return null;
-}
-async function loadRows(path) {
-  const rows = await readJsonRows(path);
-  if (rows.length === 0) throw new Error(`NoMIRACL: no rows in ${path}`);
-  return rows;
-}
-async function loadFixtures(opts) {
-  const rows = JSON.parse(await readFile(FIXTURES, "utf8"));
-  console.warn(`[nomiracl] NOMIRACL_FIXTURES=1 \u2014 loading ${rows.length} adapter fixtures`);
-  return selectTasks(rows.map(rowToTask), opts, "NoMIRACL");
-}
-function createNoMiraclAdapter() {
-  const fixturesMode = process.env.NOMIRACL_FIXTURES === "1";
-  return {
-    name: "nomiracl",
-    output: ragAnswerOutput,
-    async preflight() {
-      if (fixturesMode) {
-        await readFile(FIXTURES, "utf8");
-        return;
-      }
-      const path = dataFile();
-      if (!path) {
-        throw new Error(
-          "NOMIRACL_DATA_FILE is required. Fix: export project-miracl/nomiracl rows to JSONL and set NOMIRACL_DATA_FILE=/path/to/nomiracl.jsonl, or set NOMIRACL_FIXTURES=1 for adapter plumbing."
-        );
-      }
-      await loadRows(path);
-    },
-    async loadTasks(opts = {}) {
-      if (fixturesMode) return loadFixtures(opts);
-      const path = dataFile();
-      if (!path) throw new Error("NOMIRACL_DATA_FILE is required to load NoMIRACL tasks");
-      return selectTasks((await loadRows(path)).map(rowToTask), opts, "NoMIRACL");
-    },
-    async goldArtifact(task) {
-      return readMeta(task).answerable ? "ANSWERABLE" : "UNANSWERABLE";
-    },
-    async judge(task, artifact) {
-      const meta = readMeta(task);
-      const got = parseClassification(artifact);
-      const resolved = got === meta.answerable;
-      return {
-        resolved,
-        score: resolved ? 1 : 0,
-        detail: JSON.stringify({
-          benchmark: meta.benchmark,
-          language: meta.language,
-          subset: meta.subset,
-          expected: meta.answerable ? "ANSWERABLE" : "UNANSWERABLE",
-          got: got === null ? null : got ? "ANSWERABLE" : "UNANSWERABLE",
-          contextCount: meta.contexts.length
-        })
-      };
-    }
-  };
-}
-
-export {
-  createNoMiraclAdapter
-};
-//# sourceMappingURL=chunk-7FKBWOQT.js.map

package/dist/chunk-7FKBWOQT.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/benchmarks/nomiracl.ts"],"sourcesContent":["/**\n * NoMIRACL adapter.\n *\n * NoMIRACL tests robustness to irrelevant retrieved passages. The worker does\n * not generate an answer here; it classifies whether the supplied passages\n * contain enough evidence to answer the query. This directly measures false\n * positive / false negative behavior for RAG abstention.\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  contextBlock,\n  contextsFrom,\n  firstString,\n  isObject,\n  ragAnswerOutput,\n  readJsonRows,\n  selectTasks,\n  stringFrom,\n  type RagContext,\n} from './rag-shared'\n\nconst FIXTURES = join(benchRoot, 'fixtures', 'nomiracl.json')\n\ninterface NoMiraclMeta {\n  benchmark: 'nomiracl'\n  query: string\n  answerable: boolean\n  language: string\n  subset: string\n  contexts: RagContext[]\n}\n\nconst dataFile = (): string | undefined => process.env.NOMIRACL_DATA_FILE\n\nfunction passagesFrom(value: unknown, relevant: boolean): RagContext[] {\n  return contextsFrom(value).map((context) => ({ ...context, relevant }))\n}\n\nfunction answerableFrom(raw: Record<string, unknown>, contexts: readonly RagContext[]): boolean {\n  const direct = raw.answerable ?? raw.is_answerable ?? raw.has_answer ?? raw.label ?? raw.relevant\n  if (typeof direct === 'boolean') return direct\n  if (typeof direct === 'number') return direct > 0\n  if (typeof direct === 'string') {\n    const normalized = direct.toLowerCase()\n    if (['true', 't', 'relevant', 'answerable', 'positive', '1', 'yes'].includes(normalized)) return true\n    if (['false', 'f', 'non-relevant', 'non_relevant', 'unanswerable', 'negative', '0', 'no'].includes(normalized)) {\n      return false\n    }\n  }\n  return contexts.some((context) => context.relevant === true)\n}\n\nfunction rowToTask(raw: unknown, index: number): BenchTask {\n  if (!isObject(raw)) throw new Error(`NoMIRACL row ${index} must be an object`)\n  const query = firstString(raw, ['query', 'question'])\n  if (!query) throw new Error(`NoMIRACL row ${index} missing query`)\n  const labelledPassages = [\n    ...passagesFrom(raw.positive_passages, true),\n    ...passagesFrom(raw.negative_passages, false),\n  ]\n  const contexts =\n    labelledPassages.length > 0\n      ? labelledPassages\n      : contextsFrom(raw.passages).length > 0\n      ? contextsFrom(raw.passages)\n      : contextsFrom(raw.contexts).length > 0\n        ? contextsFrom(raw.contexts)\n        : contextsFrom(raw.documents)\n  const answerable = answerableFrom(raw, contexts)\n  const language = stringFrom(raw.lang) ?? stringFrom(raw.language) ?? 'unknown'\n  const subset = stringFrom(raw.subset) ?? (answerable ? 'relevant' : 'non-relevant')\n  const id = stringFrom(raw.id) ?? stringFrom(raw.query_id) ?? `nomiracl-${index}`\n  const meta: NoMiraclMeta = {\n    benchmark: 'nomiracl',\n    query,\n    answerable,\n    language,\n    subset,\n    contexts,\n  }\n  return {\n    id,\n    split: stringFrom(raw.split) ?? language,\n    prompt: [\n      'Classify this NoMIRACL RAG relevance case.',\n      'Return exactly one word: ANSWERABLE if at least one supplied passage contains enough evidence to answer the query, otherwise UNANSWERABLE.',\n      '',\n      `Query: ${query}`,\n      contexts.length > 0 ? `\\nPassages:\\n${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): NoMiraclMeta {\n  const md = task.metadata\n  if (!md || typeof md.answerable !== 'boolean') {\n    throw new Error(`NoMIRACL task ${task.id} missing metadata — loadTasks did not populate it`)\n  }\n  return md as unknown as NoMiraclMeta\n}\n\nfunction parseClassification(artifact: string): boolean | null {\n  const normalized = artifact.toLowerCase().replace(/[^a-z0-9]+/g, ' ').trim()\n  if (/\\bunanswerable\\b|\\bnot answerable\\b|\\bno answer\\b|\\bnegative\\b|\\bfalse\\b/.test(normalized)) {\n    return false\n  }\n  if (/\\banswerable\\b|\\brelevant\\b|\\bpositive\\b|\\btrue\\b|\\byes\\b/.test(normalized)) return true\n  return null\n}\n\nasync function loadRows(path: string): Promise<unknown[]> {\n  const rows = await readJsonRows(path)\n  if (rows.length === 0) throw new Error(`NoMIRACL: 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(`[nomiracl] NOMIRACL_FIXTURES=1 — loading ${rows.length} adapter fixtures`)\n  return selectTasks(rows.map(rowToTask), opts, 'NoMIRACL')\n}\n\nexport function createNoMiraclAdapter(): BenchmarkAdapter {\n  const fixturesMode = process.env.NOMIRACL_FIXTURES === '1'\n\n  return {\n    name: 'nomiracl',\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          'NOMIRACL_DATA_FILE is required. Fix: export project-miracl/nomiracl rows to JSONL and set NOMIRACL_DATA_FILE=/path/to/nomiracl.jsonl, or set NOMIRACL_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('NOMIRACL_DATA_FILE is required to load NoMIRACL tasks')\n      return selectTasks((await loadRows(path)).map(rowToTask), opts, 'NoMIRACL')\n    },\n\n    async goldArtifact(task: BenchTask) {\n      return readMeta(task).answerable ? 'ANSWERABLE' : 'UNANSWERABLE'\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const meta = readMeta(task)\n      const got = parseClassification(artifact)\n      const resolved = got === meta.answerable\n      return {\n        resolved,\n        score: resolved ? 1 : 0,\n        detail: JSON.stringify({\n          benchmark: meta.benchmark,\n          language: meta.language,\n          subset: meta.subset,\n          expected: meta.answerable ? 'ANSWERABLE' : 'UNANSWERABLE',\n          got: got === null ? null : got ? 'ANSWERABLE' : 'UNANSWERABLE',\n          contextCount: meta.contexts.length,\n        }),\n      }\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;AASA,SAAS,gBAAgB;AACzB,SAAS,YAAY;AAerB,IAAM,WAAW,KAAK,WAAW,YAAY,eAAe;AAW5D,IAAM,WAAW,MAA0B,QAAQ,IAAI;AAEvD,SAAS,aAAa,OAAgB,UAAiC;AACrE,SAAO,aAAa,KAAK,EAAE,IAAI,CAAC,aAAa,EAAE,GAAG,SAAS,SAAS,EAAE;AACxE;AAEA,SAAS,eAAe,KAA8B,UAA0C;AAC9F,QAAM,SAAS,IAAI,cAAc,IAAI,iBAAiB,IAAI,cAAc,IAAI,SAAS,IAAI;AACzF,MAAI,OAAO,WAAW,UAAW,QAAO;AACxC,MAAI,OAAO,WAAW,SAAU,QAAO,SAAS;AAChD,MAAI,OAAO,WAAW,UAAU;AAC9B,UAAM,aAAa,OAAO,YAAY;AACtC,QAAI,CAAC,QAAQ,KAAK,YAAY,cAAc,YAAY,KAAK,KAAK,EAAE,SAAS,UAAU,EAAG,QAAO;AACjG,QAAI,CAAC,SAAS,KAAK,gBAAgB,gBAAgB,gBAAgB,YAAY,KAAK,IAAI,EAAE,SAAS,UAAU,GAAG;AAC9G,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO,SAAS,KAAK,CAAC,YAAY,QAAQ,aAAa,IAAI;AAC7D;AAEA,SAAS,UAAU,KAAc,OAA0B;AACzD,MAAI,CAAC,SAAS,GAAG,EAAG,OAAM,IAAI,MAAM,gBAAgB,KAAK,oBAAoB;AAC7E,QAAM,QAAQ,YAAY,KAAK,CAAC,SAAS,UAAU,CAAC;AACpD,MAAI,CAAC,MAAO,OAAM,IAAI,MAAM,gBAAgB,KAAK,gBAAgB;AACjE,QAAM,mBAAmB;AAAA,IACvB,GAAG,aAAa,IAAI,mBAAmB,IAAI;AAAA,IAC3C,GAAG,aAAa,IAAI,mBAAmB,KAAK;AAAA,EAC9C;AACA,QAAM,WACJ,iBAAiB,SAAS,IACtB,mBACA,aAAa,IAAI,QAAQ,EAAE,SAAS,IACpC,aAAa,IAAI,QAAQ,IACzB,aAAa,IAAI,QAAQ,EAAE,SAAS,IAClC,aAAa,IAAI,QAAQ,IACzB,aAAa,IAAI,SAAS;AAClC,QAAM,aAAa,eAAe,KAAK,QAAQ;AAC/C,QAAM,WAAW,WAAW,IAAI,IAAI,KAAK,WAAW,IAAI,QAAQ,KAAK;AACrE,QAAM,SAAS,WAAW,IAAI,MAAM,MAAM,aAAa,aAAa;AACpE,QAAM,KAAK,WAAW,IAAI,EAAE,KAAK,WAAW,IAAI,QAAQ,KAAK,YAAY,KAAK;AAC9E,QAAM,OAAqB;AAAA,IACzB,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,UAAU,KAAK;AAAA,MACf,SAAS,SAAS,IAAI;AAAA;AAAA,EAAgB,aAAa,QAAQ,CAAC,KAAK;AAAA,IACnE,EACG,OAAO,OAAO,EACd,KAAK,IAAI;AAAA,IACZ,UAAU;AAAA,EACZ;AACF;AAEA,SAAS,SAAS,MAA+B;AAC/C,QAAM,KAAK,KAAK;AAChB,MAAI,CAAC,MAAM,OAAO,GAAG,eAAe,WAAW;AAC7C,UAAM,IAAI,MAAM,iBAAiB,KAAK,EAAE,wDAAmD;AAAA,EAC7F;AACA,SAAO;AACT;AAEA,SAAS,oBAAoB,UAAkC;AAC7D,QAAM,aAAa,SAAS,YAAY,EAAE,QAAQ,eAAe,GAAG,EAAE,KAAK;AAC3E,MAAI,2EAA2E,KAAK,UAAU,GAAG;AAC/F,WAAO;AAAA,EACT;AACA,MAAI,4DAA4D,KAAK,UAAU,EAAG,QAAO;AACzF,SAAO;AACT;AAEA,eAAe,SAAS,MAAkC;AACxD,QAAM,OAAO,MAAM,aAAa,IAAI;AACpC,MAAI,KAAK,WAAW,EAAG,OAAM,IAAI,MAAM,wBAAwB,IAAI,EAAE;AACrE,SAAO;AACT;AAEA,eAAe,aAAa,MAAyC;AACnE,QAAM,OAAO,KAAK,MAAM,MAAM,SAAS,UAAU,MAAM,CAAC;AACxD,UAAQ,KAAK,iDAA4C,KAAK,MAAM,mBAAmB;AACvF,SAAO,YAAY,KAAK,IAAI,SAAS,GAAG,MAAM,UAAU;AAC1D;AAEO,SAAS,wBAA0C;AACxD,QAAM,eAAe,QAAQ,IAAI,sBAAsB;AAEvD,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,uDAAuD;AAClF,aAAO,aAAa,MAAM,SAAS,IAAI,GAAG,IAAI,SAAS,GAAG,MAAM,UAAU;AAAA,IAC5E;AAAA,IAEA,MAAM,aAAa,MAAiB;AAClC,aAAO,SAAS,IAAI,EAAE,aAAa,eAAe;AAAA,IACpD;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,OAAO,SAAS,IAAI;AAC1B,YAAM,MAAM,oBAAoB,QAAQ;AACxC,YAAM,WAAW,QAAQ,KAAK;AAC9B,aAAO;AAAA,QACL;AAAA,QACA,OAAO,WAAW,IAAI;AAAA,QACtB,QAAQ,KAAK,UAAU;AAAA,UACrB,WAAW,KAAK;AAAA,UAChB,UAAU,KAAK;AAAA,UACf,QAAQ,KAAK;AAAA,UACb,UAAU,KAAK,aAAa,eAAe;AAAA,UAC3C,KAAK,QAAQ,OAAO,OAAO,MAAM,eAAe;AAAA,UAChD,cAAc,KAAK,SAAS;AAAA,QAC9B,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AACF;","names":[]}