@tangle-network/agent-bench 0.1.0 → 0.3.6

package/dist/chunk-5SBJCB6W.js

@@ -0,0 +1,144 @@
+import {
+  createToolLlmAdapter
+} from "./chunk-2PVVP7GN.js";
+import {
+  createTrataHedgeAdapter
+} from "./chunk-VQRS7VUC.js";
+import {
+  createWebArenaVerifiedAdapter
+} from "./chunk-Z7ML6L77.js";
+import {
+  createRagBenchAdapter
+} from "./chunk-R67DFVLO.js";
+import {
+  createSimpleQaAdapter
+} from "./chunk-SEVJPLZC.js";
+import {
+  createSweBenchAdapter
+} from "./chunk-CKUVRZ2T.js";
+import {
+  createT2RagBenchAdapter
+} from "./chunk-Y6O2OCUO.js";
+import {
+  createTau2BenchAdapter
+} from "./chunk-HBSWHQNJ.js";
+import {
+  createTau3BankingAdapter
+} from "./chunk-UAIOHCUK.js";
+import {
+  createTerminalBenchAdapter
+} from "./chunk-JRWWGMK7.js";
+import {
+  createFramesAdapter
+} from "./chunk-INNOYXCP.js";
+import {
+  createHotpotqaAdapter
+} from "./chunk-PA2ZKHJC.js";
+import {
+  createHumanEvalAdapter
+} from "./chunk-PPYSEKFM.js";
+import {
+  createMind2WebAdapter
+} from "./chunk-TBKU5XQI.js";
+import {
+  createNoMiraclAdapter
+} from "./chunk-TSWPNOYM.js";
+import {
+  createOpenRagBenchAdapter
+} from "./chunk-KDIKRJGB.js";
+import {
+  createProgrambenchAdapter
+} from "./chunk-ZEWMTR5M.js";
+import {
+  createCadBenchAdapter
+} from "./chunk-PUIRNYI7.js";
+import {
+  createCadGenBenchAdapter
+} from "./chunk-C7T7WEK2.js";
+import {
+  createCommit0Adapter
+} from "./chunk-2XU6OGEN.js";
+import {
+  createCragAdapter
+} from "./chunk-HHXFIHXC.js";
+import {
+  createDabstepAdapter
+} from "./chunk-J3KDJNX2.js";
+import {
+  createEnterpriseOpsGymAdapter
+} from "./chunk-X5YKXC6V.js";
+import {
+  createFinResearchBenchAdapter
+} from "./chunk-YCGY7UIZ.js";
+import {
+  createFinsearchcompAdapter
+} from "./chunk-UPAMRDX4.js";
+import {
+  createAecBenchAdapter
+} from "./chunk-ODT47UAY.js";
+import {
+  createAgentBenchAdapter
+} from "./chunk-7WSD27QQ.js";
+import {
+  createAppWorldAdapter,
+  createAppWorldReactAdapter
+} from "./chunk-SYDW647C.js";
+import {
+  createBfclAdapter
+} from "./chunk-IFAV6KEM.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-5SBJCB6W.js.map

package/dist/chunk-5SBJCB6W.js.map

@@ -0,0 +1 @@
+{"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":[]}

package/dist/chunk-7WSD27QQ.js

@@ -0,0 +1,118 @@
+import {
+  benchRoot
+} from "./chunk-LRRD7NAG.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-7WSD27QQ.js.map

package/dist/chunk-7WSD27QQ.js.map

@@ -0,0 +1 @@
+{"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-C7T7WEK2.js

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

package/dist/chunk-C7T7WEK2.js.map

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