@tangle-network/agent-bench 0.3.5 → 0.3.7

package/dist/chunk-PUIRNYI7.js

@@ -0,0 +1,189 @@
+// src/benchmarks/cadbench.ts
+import { readFile as readFile2 } from "fs/promises";
+
+// src/worker-blender.ts
+import { execFile } from "child_process";
+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);
+async function runLocal(cmd, args, cwd, timeoutMs = 18e4) {
+  try {
+    const { stdout, stderr } = await execFileAsync(cmd, args, { cwd, maxBuffer: 1 << 26, timeout: timeoutMs });
+    return { code: 0, stdout, stderr };
+  } catch (err) {
+    const e = err;
+    return { code: typeof e.code === "number" ? e.code : 1, stdout: e.stdout ?? "", stderr: e.stderr ?? e.message ?? String(err) };
+  }
+}
+var RUNNER_PY = `
+import bpy, sys, math, mathutils, traceback, os
+agent_script, outdir, nviews = sys.argv[-3], sys.argv[-2], int(sys.argv[-1])
+bpy.ops.object.select_all(action='SELECT'); bpy.ops.object.delete()
+ok=True
+try:
+    g={'bpy':bpy,'math':math,'mathutils':mathutils,'__name__':'__main__'}
+    exec(compile(open(agent_script).read(), agent_script, 'exec'), g)
+except Exception as e:
+    traceback.print_exc(); print('AGENT_SCRIPT_ERROR:'+repr(e)); ok=False
+meshes=[o for o in bpy.context.scene.objects if o.type=='MESH']
+if not meshes:
+    print('NO_MESH'); sys.exit(0 if ok else 3)
+mn=[1e18]*3; mx=[-1e18]*3
+for o in meshes:
+    for c in o.bound_box:
+        w=o.matrix_world @ mathutils.Vector(c)
+        for i in range(3): mn[i]=min(mn[i],w[i]); mx[i]=max(mx[i],w[i])
+center=mathutils.Vector(((mn[0]+mx[0])/2,(mn[1]+mx[1])/2,(mn[2]+mx[2])/2))
+size=max(mx[i]-mn[i] for i in range(3)) or 1.0
+# standardize: drop any agent-added cameras/lights
+for o in list(bpy.context.scene.objects):
+    if o.type in ('CAMERA','LIGHT'): bpy.data.objects.remove(o, do_unlink=True)
+w=bpy.context.scene.world or bpy.data.worlds.new('W'); bpy.context.scene.world=w
+w.use_nodes=True
+try: w.node_tree.nodes['Background'].inputs[1].default_value=0.6
+except Exception: pass
+bpy.ops.object.light_add(type='SUN'); sun=bpy.context.object; sun.data.energy=4.0; sun.rotation_euler=mathutils.Euler((0.6,0.2,0.4))
+bpy.ops.object.camera_add(); cam=bpy.context.object; bpy.context.scene.camera=cam
+sc=bpy.context.scene
+sc.render.engine='CYCLES'; sc.cycles.samples=20; sc.cycles.device='CPU'; sc.cycles.use_denoising=False
+sc.render.resolution_x=640; sc.render.resolution_y=640; sc.render.film_transparent=False
+dist=size*2.4
+el=math.radians(58)
+for v in range(nviews):
+    az=math.radians(40 + v*360.0/nviews)
+    cam.location=center+mathutils.Vector((math.cos(az)*math.sin(el), math.sin(az)*math.sin(el), math.cos(el)))*dist
+    d=(center-cam.location); cam.rotation_euler=d.to_track_quat('-Z','Y').to_euler()
+    sc.render.filepath=os.path.join(outdir, 'view_%d.png'%v); bpy.ops.render.render(write_still=True)
+print('RENDER_DONE')
+`.trim();
+async function renderBpy(script, opts = {}) {
+  const views = Math.max(1, opts.views ?? 4);
+  const dir = await mkdtemp(join(tmpdir(), "blender-judge-"));
+  const runnerPath = join(dir, "runner.py");
+  const scriptPath = join(dir, "model.py");
+  try {
+    await writeFile(runnerPath, RUNNER_PY);
+    await writeFile(scriptPath, script);
+    const run = await runLocal("xvfb-run", ["-a", "blender", "--background", "--python", runnerPath, "--", scriptPath, dir, String(views)], dir);
+    const out = `${run.stdout}
+${run.stderr}`;
+    const built = /RENDER_DONE/.test(out);
+    if (!built) return { built: false, renders: [], error: /(AGENT_SCRIPT_ERROR:.*|NO_MESH)/.exec(out)?.[0] ?? out.trim().slice(-400) };
+    const renders = [];
+    for (let v = 0; v < views; v++) {
+      const buf = await readFile(join(dir, `view_${v}.png`)).catch(() => void 0);
+      if (buf) renders.push(`data:image/png;base64,${buf.toString("base64")}`);
+    }
+    return { built: renders.length > 0, renders };
+  } finally {
+    await rm(dir, { recursive: true, force: true }).catch(() => {
+    });
+  }
+}
+
+// src/benchmarks/cadbench.ts
+function must(name) {
+  const v = process.env[name];
+  if (!v) throw new Error(`env ${name} is required for the CADBench judge`);
+  return v;
+}
+async function judgeCriteria(instruction, script, criteria, renders) {
+  const base = (process.env.ROUTER_BASE ?? "https://router.tangle.tools/v1").replace(/\/$/, "");
+  const key = must("TANGLE_API_KEY");
+  const model = process.env.JUDGE_MODEL ?? "deepseek-v4-flash";
+  const numbered = criteria.map((c, i) => `${i + 1}. ${c}`).join("\n");
+  const text = `You are strictly grading a 3D model that was built by a Blender bpy script for this instruction:
+"${instruction}"
+
+Below are ${renders.length} rendered views of the produced model, and the script that built it. For EACH numbered criterion, decide whether it is satisfied (judge geometry/shape/proportion/structure from the IMAGES; judge color/size/material reasonableness from the SCRIPT where the images are ambiguous). Return ONLY a JSON array of exactly ${criteria.length} booleans (true=satisfied, false=not), in order, no prose.
+
+CRITERIA:
+${numbered}
+
+SCRIPT:
+\`\`\`python
+${script.slice(0, 6e3)}
+\`\`\``;
+  const content = [{ type: "text", text }];
+  for (const url of renders) content.push({ type: "image_url", image_url: { url } });
+  const res = await fetch(`${base}/chat/completions`, {
+    method: "POST",
+    headers: { "content-type": "application/json", authorization: `Bearer ${key}` },
+    body: JSON.stringify({ model, max_tokens: 1500, temperature: 0, messages: [{ role: "user", content }] })
+  });
+  if (!res.ok) throw new Error(`judge ${model} ${res.status}: ${(await res.text()).slice(0, 200)}`);
+  const data = await res.json();
+  const raw = data.choices?.[0]?.message?.content ?? "";
+  const m = /\[\s*(?:true|false)[\s\S]*?\]/i.exec(raw);
+  if (!m) return { passed: criteria.map(() => false), note: `judge returned no parseable verdict: ${raw.slice(0, 80)}` };
+  let arr;
+  try {
+    arr = JSON.parse(m[0].toLowerCase());
+  } catch {
+    return { passed: criteria.map(() => false), note: "judge verdict not valid JSON" };
+  }
+  const bools = Array.isArray(arr) ? arr.map((x) => x === true) : [];
+  const passed = criteria.map((_, i) => bools[i] === true);
+  return { passed, note: `${passed.filter(Boolean).length}/${criteria.length} criteria` };
+}
+function createCadBenchAdapter() {
+  let cache = null;
+  async function load() {
+    if (cache) return cache;
+    const path = process.env.CADBENCH_PATH;
+    if (!path) throw new Error("CADBENCH_PATH must point at the cleaned CADBench JSONL ({id,instruction,type,criteria:[]} per line)");
+    const text = await readFile2(path, "utf8");
+    cache = text.split("\n").filter((l) => l.trim()).map((l) => {
+      const r = JSON.parse(l);
+      return { id: r.id, instruction: r.instruction, meta: { name: r.name ?? "", type: r.type ?? "", criteria: r.criteria } };
+    });
+    return cache;
+  }
+  return {
+    name: "cadbench",
+    async preflight() {
+      const { execFile: execFile2 } = await import("child_process");
+      const { promisify: promisify2 } = await import("util");
+      const exec = promisify2(execFile2);
+      try {
+        await exec("xvfb-run", ["-a", "blender", "--version"], { timeout: 3e4 });
+      } catch (err) {
+        throw new Error(
+          `cadbench preflight failed: ${(err instanceof Error ? err.message : String(err)).slice(0, 200)}
+Fix: install Blender + Xvfb (sudo apt-get install -y blender xvfb). The judge runs \`xvfb-run -a blender --background --python\`.`
+        );
+      }
+      await load();
+    },
+    async loadTasks(opts = {}) {
+      let rows = await load();
+      if (opts.ids) rows = rows.filter((r) => opts.ids.includes(r.id));
+      const t = process.env.CADBENCH_TYPE;
+      if (t) rows = rows.filter((r) => r.meta.type.toLowerCase() === t.toLowerCase());
+      if (opts.limit != null) rows = rows.slice(0, opts.limit);
+      return rows.map((r) => ({ id: r.id, prompt: r.instruction, metadata: r.meta }));
+    },
+    async goldArtifact() {
+      return void 0;
+    },
+    async judge(task, artifact) {
+      const meta = task.metadata;
+      const criteria = meta.criteria ?? [];
+      if (!artifact.trim()) return { resolved: false, score: 0, detail: "empty artifact" };
+      if (criteria.length === 0) return { resolved: false, score: 0, detail: "task has no criteria" };
+      const r = await renderBpy(artifact, { views: 4 });
+      if (!r.built) return { resolved: false, score: 0, detail: `did not build/render: ${r.error ?? "no mesh"}` };
+      const { passed, note } = await judgeCriteria(task.prompt, artifact, criteria, r.renders);
+      const score = passed.filter(Boolean).length / criteria.length;
+      return { resolved: score === 1, score, detail: note };
+    }
+  };
+}
+
+export {
+  createCadBenchAdapter
+};
+//# sourceMappingURL=chunk-PUIRNYI7.js.map

package/dist/chunk-PUIRNYI7.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/benchmarks/cadbench.ts","../src/worker-blender.ts"],"sourcesContent":["/**\n * CADBench / BlenderLLM adapter (FreedomIntelligence/CADBench, arXiv:2412.14203).\n * Task = NL instruction → a Blender `bpy` script. Score = the paper's criteria\n * eval: render the produced model to standardized views, then a vision judge\n * (GPT-4o-class) marks each per-task criterion bullet pass/fail against the\n * rendered images + the script text. score = fraction of criteria satisfied.\n *\n * Data: the published dataset's `criteria` flattened to a bullet list (700 tasks,\n * 500 Simulative + 200 Wild). Point CADBENCH_PATH at the cleaned JSONL\n * ({id,name,instruction,type,criteria:string[]} per line). Judge creds from\n * TANGLE_API_KEY / ROUTER_BASE / JUDGE_MODEL (default gpt-4o).\n */\n\nimport { readFile } from 'node:fs/promises'\nimport type { BenchScore, BenchTask, BenchmarkAdapter, LoadOptions } from './types'\nimport { renderBpy } from '../worker-blender'\n\ninterface CadBenchMeta {\n  name: string\n  type: string\n  criteria: string[]\n}\n\nfunction must(name: string): string {\n  const v = process.env[name]\n  if (!v) throw new Error(`env ${name} is required for the CADBench judge`)\n  return v\n}\n\n/** One batched vision call: rendered views + the bpy script + the numbered\n *  criteria → a JSON array of booleans (true = satisfied). Faithful to the\n *  paper's combined image+script evaluation. Throws on transport failure (never\n *  a silent zero); a parse miss falls back to \"all fail\" with a note. */\nasync function judgeCriteria(\n  instruction: string,\n  script: string,\n  criteria: string[],\n  renders: string[],\n): Promise<{ passed: boolean[]; note: string }> {\n  const base = (process.env.ROUTER_BASE ?? 'https://router.tangle.tools/v1').replace(/\\/$/, '')\n  const key = must('TANGLE_API_KEY')\n  const model = process.env.JUDGE_MODEL ?? 'deepseek-v4-flash'\n  const numbered = criteria.map((c, i) => `${i + 1}. ${c}`).join('\\n')\n  const text =\n    `You are strictly grading a 3D model that was built by a Blender bpy script for this instruction:\\n\"${instruction}\"\\n\\n` +\n    `Below are ${renders.length} rendered views of the produced model, and the script that built it. ` +\n    `For EACH numbered criterion, decide whether it is satisfied (judge geometry/shape/proportion/structure from the IMAGES; judge color/size/material reasonableness from the SCRIPT where the images are ambiguous). ` +\n    `Return ONLY a JSON array of exactly ${criteria.length} booleans (true=satisfied, false=not), in order, no prose.\\n\\nCRITERIA:\\n${numbered}\\n\\nSCRIPT:\\n\\`\\`\\`python\\n${script.slice(0, 6000)}\\n\\`\\`\\``\n  const content: unknown[] = [{ type: 'text', text }]\n  for (const url of renders) content.push({ type: 'image_url', image_url: { url } })\n  const res = await fetch(`${base}/chat/completions`, {\n    method: 'POST',\n    headers: { 'content-type': 'application/json', authorization: `Bearer ${key}` },\n    body: JSON.stringify({ model, max_tokens: 1500, temperature: 0, messages: [{ role: 'user', content }] }),\n  })\n  if (!res.ok) throw new Error(`judge ${model} ${res.status}: ${(await res.text()).slice(0, 200)}`)\n  const data = (await res.json()) as { choices?: Array<{ message?: { content?: string } }> }\n  const raw = data.choices?.[0]?.message?.content ?? ''\n  const m = /\\[\\s*(?:true|false)[\\s\\S]*?\\]/i.exec(raw)\n  if (!m) return { passed: criteria.map(() => false), note: `judge returned no parseable verdict: ${raw.slice(0, 80)}` }\n  let arr: unknown\n  try {\n    arr = JSON.parse(m[0].toLowerCase())\n  } catch {\n    return { passed: criteria.map(() => false), note: 'judge verdict not valid JSON' }\n  }\n  const bools = Array.isArray(arr) ? arr.map((x) => x === true) : []\n  // Pad/truncate to criteria length (a short array scores the missing as fail).\n  const passed = criteria.map((_, i) => bools[i] === true)\n  return { passed, note: `${passed.filter(Boolean).length}/${criteria.length} criteria` }\n}\n\nexport function createCadBenchAdapter(): BenchmarkAdapter {\n  let cache: Array<{ id: string; instruction: string; meta: CadBenchMeta }> | null = null\n\n  async function load(): Promise<typeof cache & object> {\n    if (cache) return cache\n    const path = process.env.CADBENCH_PATH\n    if (!path) throw new Error('CADBENCH_PATH must point at the cleaned CADBench JSONL ({id,instruction,type,criteria:[]} per line)')\n    const text = await readFile(path, 'utf8')\n    cache = text\n      .split('\\n')\n      .filter((l) => l.trim())\n      .map((l) => {\n        const r = JSON.parse(l) as { id: string; name?: string; instruction: string; type?: string; criteria: string[] }\n        return { id: r.id, instruction: r.instruction, meta: { name: r.name ?? '', type: r.type ?? '', criteria: r.criteria } }\n      })\n    return cache\n  }\n\n  return {\n    name: 'cadbench',\n\n    async preflight() {\n      const { execFile } = await import('node:child_process')\n      const { promisify } = await import('node:util')\n      const exec = promisify(execFile)\n      try {\n        await exec('xvfb-run', ['-a', 'blender', '--version'], { timeout: 30_000 })\n      } catch (err) {\n        throw new Error(\n          `cadbench preflight failed: ${(err instanceof Error ? err.message : String(err)).slice(0, 200)}\\n` +\n            `Fix: install Blender + Xvfb (sudo apt-get install -y blender xvfb). The judge runs \\`xvfb-run -a blender --background --python\\`.`,\n        )\n      }\n      await load()\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      let rows = await load()\n      if (opts.ids) rows = rows.filter((r) => opts.ids!.includes(r.id))\n      // TYPE filter (Simulative|Wild) via env, applied before limit.\n      const t = process.env.CADBENCH_TYPE\n      if (t) rows = rows.filter((r) => r.meta.type.toLowerCase() === t.toLowerCase())\n      if (opts.limit != null) rows = rows.slice(0, opts.limit)\n      return rows.map((r): BenchTask => ({ id: r.id, prompt: r.instruction, metadata: r.meta as unknown as Record<string, unknown> }))\n    },\n\n    async goldArtifact() {\n      return undefined // no reference bpy script ships with the benchmark\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const meta = task.metadata as unknown as CadBenchMeta\n      const criteria = meta.criteria ?? []\n      if (!artifact.trim()) return { resolved: false, score: 0, detail: 'empty artifact' }\n      if (criteria.length === 0) return { resolved: false, score: 0, detail: 'task has no criteria' }\n      const r = await renderBpy(artifact, { views: 4 })\n      if (!r.built) return { resolved: false, score: 0, detail: `did not build/render: ${r.error ?? 'no mesh'}` }\n      const { passed, note } = await judgeCriteria(task.prompt, artifact, criteria, r.renders)\n      const score = passed.filter(Boolean).length / criteria.length\n      return { resolved: score === 1, score, detail: note }\n    },\n  }\n}\n","/**\n * BlenderLLM / CADBench worker. The deliverable for a CADBench task is a Blender\n * `bpy` Python script that builds the described 3D model. We author it via the\n * router, execute it headless in Blender (Cycles CPU, no GPU), auto-frame the\n * produced geometry, and render N standardized views — the images the CADBench\n * criteria judge scores. The authoring directive is the GEPA-optimizable surface.\n *\n * Requires `blender` + `xvfb-run` on PATH (apt blender 4.x). No GPU: Cycles CPU\n * with denoising off (the apt build ships without OpenImageDenoise).\n */\n\nimport { execFile } from 'node:child_process'\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_BLENDER_DIRECTIVE } from './directives'\nimport { runRefineLoop } from './refine-loop'\nimport { routerChatWithUsage } from '@tangle-network/agent-runtime/loops'\n\nexport { DEFAULT_BLENDER_DIRECTIVE } from './directives'\n\nconst execFileAsync = promisify(execFile)\n\nasync function runLocal(cmd: string, args: string[], cwd: string, timeoutMs = 180_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\n\n/** Strip markdown fences so we keep just the Python. */\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\n/**\n * The standardized Blender runner (written to a temp file per run). It clears the\n * scene, executes the agent's bpy script, auto-frames the produced meshes, sets\n * up neutral lighting, and renders N azimuth views with Cycles CPU.\n */\nconst RUNNER_PY = `\nimport bpy, sys, math, mathutils, traceback, os\nagent_script, outdir, nviews = sys.argv[-3], sys.argv[-2], int(sys.argv[-1])\nbpy.ops.object.select_all(action='SELECT'); bpy.ops.object.delete()\nok=True\ntry:\n    g={'bpy':bpy,'math':math,'mathutils':mathutils,'__name__':'__main__'}\n    exec(compile(open(agent_script).read(), agent_script, 'exec'), g)\nexcept Exception as e:\n    traceback.print_exc(); print('AGENT_SCRIPT_ERROR:'+repr(e)); ok=False\nmeshes=[o for o in bpy.context.scene.objects if o.type=='MESH']\nif not meshes:\n    print('NO_MESH'); sys.exit(0 if ok else 3)\nmn=[1e18]*3; mx=[-1e18]*3\nfor o in meshes:\n    for c in o.bound_box:\n        w=o.matrix_world @ mathutils.Vector(c)\n        for i in range(3): mn[i]=min(mn[i],w[i]); mx[i]=max(mx[i],w[i])\ncenter=mathutils.Vector(((mn[0]+mx[0])/2,(mn[1]+mx[1])/2,(mn[2]+mx[2])/2))\nsize=max(mx[i]-mn[i] for i in range(3)) or 1.0\n# standardize: drop any agent-added cameras/lights\nfor o in list(bpy.context.scene.objects):\n    if o.type in ('CAMERA','LIGHT'): bpy.data.objects.remove(o, do_unlink=True)\nw=bpy.context.scene.world or bpy.data.worlds.new('W'); bpy.context.scene.world=w\nw.use_nodes=True\ntry: w.node_tree.nodes['Background'].inputs[1].default_value=0.6\nexcept Exception: pass\nbpy.ops.object.light_add(type='SUN'); sun=bpy.context.object; sun.data.energy=4.0; sun.rotation_euler=mathutils.Euler((0.6,0.2,0.4))\nbpy.ops.object.camera_add(); cam=bpy.context.object; bpy.context.scene.camera=cam\nsc=bpy.context.scene\nsc.render.engine='CYCLES'; sc.cycles.samples=20; sc.cycles.device='CPU'; sc.cycles.use_denoising=False\nsc.render.resolution_x=640; sc.render.resolution_y=640; sc.render.film_transparent=False\ndist=size*2.4\nel=math.radians(58)\nfor v in range(nviews):\n    az=math.radians(40 + v*360.0/nviews)\n    cam.location=center+mathutils.Vector((math.cos(az)*math.sin(el), math.sin(az)*math.sin(el), math.cos(el)))*dist\n    d=(center-cam.location); cam.rotation_euler=d.to_track_quat('-Z','Y').to_euler()\n    sc.render.filepath=os.path.join(outdir, 'view_%d.png'%v); bpy.ops.render.render(write_still=True)\nprint('RENDER_DONE')\n`.trim()\n\n/** Execute a bpy script headless + render N standardized views — no authoring.\n *  Used by the CADBench judge to render an artifact before vision-scoring it. */\nexport async function renderBpy(script: string, opts: { views?: number } = {}): Promise<{ built: boolean; renders: string[]; error?: string }> {\n  const views = Math.max(1, opts.views ?? 4)\n  const dir = await mkdtemp(join(tmpdir(), 'blender-judge-'))\n  const runnerPath = join(dir, 'runner.py')\n  const scriptPath = join(dir, 'model.py')\n  try {\n    await writeFile(runnerPath, RUNNER_PY)\n    await writeFile(scriptPath, script)\n    const run = await runLocal('xvfb-run', ['-a', 'blender', '--background', '--python', runnerPath, '--', scriptPath, dir, String(views)], dir)\n    const out = `${run.stdout}\\n${run.stderr}`\n    const built = /RENDER_DONE/.test(out)\n    if (!built) return { built: false, renders: [], error: (/(AGENT_SCRIPT_ERROR:.*|NO_MESH)/.exec(out)?.[0] ?? out.trim().slice(-400)) }\n    const renders: string[] = []\n    for (let v = 0; v < views; v++) {\n      const buf = await readFile(join(dir, `view_${v}.png`)).catch(() => undefined)\n      if (buf) renders.push(`data:image/png;base64,${buf.toString('base64')}`)\n    }\n    return { built: renders.length > 0, renders }\n  } finally {\n    await rm(dir, { recursive: true, force: true }).catch(() => {})\n  }\n}\n\nexport interface BlenderLocalConfig {\n  routerBaseUrl: string\n  routerKey: string\n  model: string\n  rounds?: number\n  /** N standardized views to render (CADBench uses 4). Default 4. */\n  views?: number\n  /** The bpy authoring directive — the GEPA-optimizable surface. */\n  directive?: string\n}\n\nexport interface BlenderShot {\n  /** The bpy script the agent wrote — the artifact. */\n  artifact: string\n  /** Rendered view PNGs as data URIs (the images the criteria judge scores). */\n  renders: string[]\n  trace: Span[]\n  usage: { input: number; output: number }\n  ok: boolean\n  /** True if the script executed and produced at least one mesh. */\n  built: boolean\n  detail?: string\n}\n\n/**\n * Author a bpy script for the task via the router, execute + render it headless\n * in Blender, refine on execution errors across rounds. Returns the script, the\n * rendered views, a screenshot-rich trace, and real token usage.\n */\nexport async function solveBlenderLocal(task: BenchTask, cfg: BlenderLocalConfig): Promise<BlenderShot> {\n  const rounds = Math.max(1, cfg.rounds ?? 2)\n  const views = Math.max(1, cfg.views ?? 4)\n  const directive = cfg.directive ?? DEFAULT_BLENDER_DIRECTIVE\n  const trace: Span[] = []\n  const runId = `cadbench-${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 bpy script; the\n  // render PNGs + built flag + lastErr persist outside the loop). usage is REAL.\n  let renders: string[] = []\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 dir (with the runner.py written once\n  // in setup) is the Ctx; built (RENDER_DONE + ≥1 collected view) is the early-stop,\n  // modeled as a judge so default-decide stops the loop. The round-2+ steer carries\n  // lastErr + the prior script verbatim.\n  const res = await runRefineLoop<string, string>({\n    rounds,\n    setup: async () => {\n      const dir = await mkdtemp(join(tmpdir(), 'blender-'))\n      await writeFile(join(dir, 'runner.py'), RUNNER_PY)\n      return dir\n    },\n    prompt: (round, history) =>\n      round === 1\n        ? task.prompt\n        : `Your previous bpy script failed:\\n${lastErr}\\n\\nPrevious script:\\n${history[history.length - 1]?.artifact ?? ''}\\n\\nFix it so it runs under \\`blender --background --python\\` and builds the object as mesh(es). Brief:\\n${task.prompt}`,\n    runShot: async (user, round, dir) => {\n      const runnerPath = join(dir, 'runner.py')\n      const scriptPath = join(dir, 'model.py')\n      const { content, usage: u } = await routerChatWithUsage(\n        cfg,\n        [\n          { role: 'system', content: directive },\n          { role: 'user', content: user },\n        ],\n        { temperature: 0.3 },\n      )\n      if (u) {\n        usage.input += u.input\n        usage.output += u.output\n      }\n      const script = extractPy(content)\n      trace.push({ spanId: `s-author-${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: 'model.py', content: script }, startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)\n\n      await writeFile(scriptPath, script)\n      const run = await runLocal('xvfb-run', ['-a', 'blender', '--background', '--python', runnerPath, '--', scriptPath, dir, String(views)], dir)\n      const out = `${run.stdout}\\n${run.stderr}`\n      built = /RENDER_DONE/.test(out)\n      lastErr = built ? '' : (/(AGENT_SCRIPT_ERROR:.*|NO_MESH)/.exec(out)?.[0] ?? out.trim().slice(-800))\n      trace.push({ spanId: `s-blender-${round}`, runId, kind: 'tool', name: `blender r${round}`, toolName: 'shell.exec', args: 'blender --background --python runner.py model.py', result: (built ? 'RENDER_DONE' : lastErr).slice(0, 1500), startedAt: tick(), endedAt: tick(), status: built ? 'ok' : 'error', error: built ? undefined : `exit ${run.code}` } as Span)\n\n      if (built) {\n        const collected: string[] = []\n        for (let v = 0; v < views; v++) {\n          const buf = await readFile(join(dir, `view_${v}.png`)).catch(() => undefined)\n          if (buf) collected.push(`data:image/png;base64,${buf.toString('base64')}`)\n        }\n        renders = collected\n        // first view carries the screen span (run-capsule reveal)\n        trace.push({ spanId: `s-render-${round}`, runId, kind: 'tool', name: 'render', toolName: 'render.screenshot', args: { action: 'rendered model', url: 'view_0.png' }, attributes: collected[0] ? { screenshot: collected[0] } : {}, startedAt: tick(), endedAt: tick(), status: collected.length ? 'ok' : 'error', error: collected.length ? undefined : 'render produced no image' } as Span)\n        built = collected.length > 0\n      }\n      return { artifact: script }\n    },\n    judge: async () => ({ valid: built }),\n    teardown: (dir) => rm(dir, { recursive: true, force: true }).then(() => {}, () => {}),\n  })\n\n  const script = res.final.artifact\n  return {\n    artifact: script,\n    renders,\n    trace,\n    usage,\n    ok: script.trim().length > 0,\n    built,\n    detail: built ? `built + rendered ${renders.length} views` : `did not build in ${rounds} rounds${lastErr ? `; last: ${lastErr.slice(0, 140)}` : ''}`,\n  }\n}\n"],"mappings":";AAaA,SAAS,YAAAA,iBAAgB;;;ACFzB,SAAS,gBAAgB;AACzB,SAAS,SAAS,UAAU,IAAI,iBAAiB;AACjD,SAAS,cAAc;AACvB,SAAS,YAAY;AACrB,SAAS,iBAAiB;AAK1B,SAAS,2BAA2B;AAIpC,IAAM,gBAAgB,UAAU,QAAQ;AAExC,eAAe,SAAS,KAAa,MAAgB,KAAa,YAAY,MAAoE;AAChJ,MAAI;AACF,UAAM,EAAE,QAAQ,OAAO,IAAI,MAAM,cAAc,KAAK,MAAM,EAAE,KAAK,WAAW,KAAK,IAAI,SAAS,UAAU,CAAC;AACzG,WAAO,EAAE,MAAM,GAAG,QAAQ,OAAO;AAAA,EACnC,SAAS,KAAK;AACZ,UAAM,IAAI;AACV,WAAO,EAAE,MAAM,OAAO,EAAE,SAAS,WAAW,EAAE,OAAO,GAAG,QAAQ,EAAE,UAAU,IAAI,QAAQ,EAAE,UAAU,EAAE,WAAW,OAAO,GAAG,EAAE;AAAA,EAC/H;AACF;AAcA,IAAM,YAAY;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAwChB,KAAK;AAIP,eAAsB,UAAU,QAAgB,OAA2B,CAAC,GAAmE;AAC7I,QAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,SAAS,CAAC;AACzC,QAAM,MAAM,MAAM,QAAQ,KAAK,OAAO,GAAG,gBAAgB,CAAC;AAC1D,QAAM,aAAa,KAAK,KAAK,WAAW;AACxC,QAAM,aAAa,KAAK,KAAK,UAAU;AACvC,MAAI;AACF,UAAM,UAAU,YAAY,SAAS;AACrC,UAAM,UAAU,YAAY,MAAM;AAClC,UAAM,MAAM,MAAM,SAAS,YAAY,CAAC,MAAM,WAAW,gBAAgB,YAAY,YAAY,MAAM,YAAY,KAAK,OAAO,KAAK,CAAC,GAAG,GAAG;AAC3I,UAAM,MAAM,GAAG,IAAI,MAAM;AAAA,EAAK,IAAI,MAAM;AACxC,UAAM,QAAQ,cAAc,KAAK,GAAG;AACpC,QAAI,CAAC,MAAO,QAAO,EAAE,OAAO,OAAO,SAAS,CAAC,GAAG,OAAQ,kCAAkC,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,KAAK,EAAE,MAAM,IAAI,EAAG;AACpI,UAAM,UAAoB,CAAC;AAC3B,aAAS,IAAI,GAAG,IAAI,OAAO,KAAK;AAC9B,YAAM,MAAM,MAAM,SAAS,KAAK,KAAK,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,MAAM,MAAS;AAC5E,UAAI,IAAK,SAAQ,KAAK,yBAAyB,IAAI,SAAS,QAAQ,CAAC,EAAE;AAAA,IACzE;AACA,WAAO,EAAE,OAAO,QAAQ,SAAS,GAAG,QAAQ;AAAA,EAC9C,UAAE;AACA,UAAM,GAAG,KAAK,EAAE,WAAW,MAAM,OAAO,KAAK,CAAC,EAAE,MAAM,MAAM;AAAA,IAAC,CAAC;AAAA,EAChE;AACF;;;AD1FA,SAAS,KAAK,MAAsB;AAClC,QAAM,IAAI,QAAQ,IAAI,IAAI;AAC1B,MAAI,CAAC,EAAG,OAAM,IAAI,MAAM,OAAO,IAAI,qCAAqC;AACxE,SAAO;AACT;AAMA,eAAe,cACb,aACA,QACA,UACA,SAC8C;AAC9C,QAAM,QAAQ,QAAQ,IAAI,eAAe,kCAAkC,QAAQ,OAAO,EAAE;AAC5F,QAAM,MAAM,KAAK,gBAAgB;AACjC,QAAM,QAAQ,QAAQ,IAAI,eAAe;AACzC,QAAM,WAAW,SAAS,IAAI,CAAC,GAAG,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,EAAE,EAAE,KAAK,IAAI;AACnE,QAAM,OACJ;AAAA,GAAsG,WAAW;AAAA;AAAA,YACpG,QAAQ,MAAM,8TAEY,SAAS,MAAM;AAAA;AAAA;AAAA,EAA4E,QAAQ;AAAA;AAAA;AAAA;AAAA,EAA8B,OAAO,MAAM,GAAG,GAAI,CAAC;AAAA;AAC/L,QAAM,UAAqB,CAAC,EAAE,MAAM,QAAQ,KAAK,CAAC;AAClD,aAAW,OAAO,QAAS,SAAQ,KAAK,EAAE,MAAM,aAAa,WAAW,EAAE,IAAI,EAAE,CAAC;AACjF,QAAM,MAAM,MAAM,MAAM,GAAG,IAAI,qBAAqB;AAAA,IAClD,QAAQ;AAAA,IACR,SAAS,EAAE,gBAAgB,oBAAoB,eAAe,UAAU,GAAG,GAAG;AAAA,IAC9E,MAAM,KAAK,UAAU,EAAE,OAAO,YAAY,MAAM,aAAa,GAAG,UAAU,CAAC,EAAE,MAAM,QAAQ,QAAQ,CAAC,EAAE,CAAC;AAAA,EACzG,CAAC;AACD,MAAI,CAAC,IAAI,GAAI,OAAM,IAAI,MAAM,SAAS,KAAK,IAAI,IAAI,MAAM,MAAM,MAAM,IAAI,KAAK,GAAG,MAAM,GAAG,GAAG,CAAC,EAAE;AAChG,QAAM,OAAQ,MAAM,IAAI,KAAK;AAC7B,QAAM,MAAM,KAAK,UAAU,CAAC,GAAG,SAAS,WAAW;AACnD,QAAM,IAAI,iCAAiC,KAAK,GAAG;AACnD,MAAI,CAAC,EAAG,QAAO,EAAE,QAAQ,SAAS,IAAI,MAAM,KAAK,GAAG,MAAM,wCAAwC,IAAI,MAAM,GAAG,EAAE,CAAC,GAAG;AACrH,MAAI;AACJ,MAAI;AACF,UAAM,KAAK,MAAM,EAAE,CAAC,EAAE,YAAY,CAAC;AAAA,EACrC,QAAQ;AACN,WAAO,EAAE,QAAQ,SAAS,IAAI,MAAM,KAAK,GAAG,MAAM,+BAA+B;AAAA,EACnF;AACA,QAAM,QAAQ,MAAM,QAAQ,GAAG,IAAI,IAAI,IAAI,CAAC,MAAM,MAAM,IAAI,IAAI,CAAC;AAEjE,QAAM,SAAS,SAAS,IAAI,CAAC,GAAG,MAAM,MAAM,CAAC,MAAM,IAAI;AACvD,SAAO,EAAE,QAAQ,MAAM,GAAG,OAAO,OAAO,OAAO,EAAE,MAAM,IAAI,SAAS,MAAM,YAAY;AACxF;AAEO,SAAS,wBAA0C;AACxD,MAAI,QAA+E;AAEnF,iBAAe,OAAuC;AACpD,QAAI,MAAO,QAAO;AAClB,UAAM,OAAO,QAAQ,IAAI;AACzB,QAAI,CAAC,KAAM,OAAM,IAAI,MAAM,qGAAqG;AAChI,UAAM,OAAO,MAAMC,UAAS,MAAM,MAAM;AACxC,YAAQ,KACL,MAAM,IAAI,EACV,OAAO,CAAC,MAAM,EAAE,KAAK,CAAC,EACtB,IAAI,CAAC,MAAM;AACV,YAAM,IAAI,KAAK,MAAM,CAAC;AACtB,aAAO,EAAE,IAAI,EAAE,IAAI,aAAa,EAAE,aAAa,MAAM,EAAE,MAAM,EAAE,QAAQ,IAAI,MAAM,EAAE,QAAQ,IAAI,UAAU,EAAE,SAAS,EAAE;AAAA,IACxH,CAAC;AACH,WAAO;AAAA,EACT;AAEA,SAAO;AAAA,IACL,MAAM;AAAA,IAEN,MAAM,YAAY;AAChB,YAAM,EAAE,UAAAC,UAAS,IAAI,MAAM,OAAO,eAAoB;AACtD,YAAM,EAAE,WAAAC,WAAU,IAAI,MAAM,OAAO,MAAW;AAC9C,YAAM,OAAOA,WAAUD,SAAQ;AAC/B,UAAI;AACF,cAAM,KAAK,YAAY,CAAC,MAAM,WAAW,WAAW,GAAG,EAAE,SAAS,IAAO,CAAC;AAAA,MAC5E,SAAS,KAAK;AACZ,cAAM,IAAI;AAAA,UACR,+BAA+B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAAG,MAAM,GAAG,GAAG,CAAC;AAAA;AAAA,QAEhG;AAAA,MACF;AACA,YAAM,KAAK;AAAA,IACb;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;AACtC,UAAI,OAAO,MAAM,KAAK;AACtB,UAAI,KAAK,IAAK,QAAO,KAAK,OAAO,CAAC,MAAM,KAAK,IAAK,SAAS,EAAE,EAAE,CAAC;AAEhE,YAAM,IAAI,QAAQ,IAAI;AACtB,UAAI,EAAG,QAAO,KAAK,OAAO,CAAC,MAAM,EAAE,KAAK,KAAK,YAAY,MAAM,EAAE,YAAY,CAAC;AAC9E,UAAI,KAAK,SAAS,KAAM,QAAO,KAAK,MAAM,GAAG,KAAK,KAAK;AACvD,aAAO,KAAK,IAAI,CAAC,OAAkB,EAAE,IAAI,EAAE,IAAI,QAAQ,EAAE,aAAa,UAAU,EAAE,KAA2C,EAAE;AAAA,IACjI;AAAA,IAEA,MAAM,eAAe;AACnB,aAAO;AAAA,IACT;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,OAAO,KAAK;AAClB,YAAM,WAAW,KAAK,YAAY,CAAC;AACnC,UAAI,CAAC,SAAS,KAAK,EAAG,QAAO,EAAE,UAAU,OAAO,OAAO,GAAG,QAAQ,iBAAiB;AACnF,UAAI,SAAS,WAAW,EAAG,QAAO,EAAE,UAAU,OAAO,OAAO,GAAG,QAAQ,uBAAuB;AAC9F,YAAM,IAAI,MAAM,UAAU,UAAU,EAAE,OAAO,EAAE,CAAC;AAChD,UAAI,CAAC,EAAE,MAAO,QAAO,EAAE,UAAU,OAAO,OAAO,GAAG,QAAQ,yBAAyB,EAAE,SAAS,SAAS,GAAG;AAC1G,YAAM,EAAE,QAAQ,KAAK,IAAI,MAAM,cAAc,KAAK,QAAQ,UAAU,UAAU,EAAE,OAAO;AACvF,YAAM,QAAQ,OAAO,OAAO,OAAO,EAAE,SAAS,SAAS;AACvD,aAAO,EAAE,UAAU,UAAU,GAAG,OAAO,QAAQ,KAAK;AAAA,IACtD;AAAA,EACF;AACF;","names":["readFile","readFile","execFile","promisify"]}

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

package/dist/chunk-PWQVGAJB.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-R36V2VP7.js

@@ -0,0 +1,169 @@
+import {
+  runVenvPython
+} from "./chunk-LRRD7NAG.js";
+
+// src/benchmarks/tau-bench-shared.ts
+import { readFile, stat } from "fs/promises";
+import { resolve } from "path";
+var tauResultsOutput = {
+  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|path|json)?\s*\n([\s\S]*?)```/g)];
+    return (fences.at(-1)?.[1] ?? text).trim();
+  }
+};
+async function assertPath(path, label, benchName) {
+  try {
+    await stat(path);
+  } catch (err) {
+    throw new Error(`${benchName}: missing ${label} at ${path} (${err instanceof Error ? err.message : err})`);
+  }
+}
+function benchDir(config) {
+  return process.env[config.dirEnv];
+}
+function benchDomain(config) {
+  return process.env[config.domainEnv] ?? config.defaultDomain;
+}
+function rowToTask(row, config, split) {
+  const meta = {
+    taskId: row.id,
+    domain: row.domain,
+    split,
+    userScenario: row.user_scenario,
+    description: row.description,
+    evaluationCriteria: row.evaluation_criteria
+  };
+  return {
+    id: row.id,
+    split,
+    prompt: [
+      config.taskIntro,
+      `Run this task in the official ${row.domain} domain.`,
+      "The benchmark is a simulated multi-turn user/tool conversation.",
+      "",
+      typeof row.user_scenario === "string" ? row.user_scenario : JSON.stringify(row.user_scenario ?? {}, null, 2),
+      "",
+      "Return the path to the official tau results.json or trajectory file containing this task run."
+    ].join("\n"),
+    metadata: meta
+  };
+}
+function readMeta(task, benchName) {
+  const md = task.metadata;
+  if (!md || typeof md.taskId !== "string" || typeof md.domain !== "string") {
+    throw new Error(`${benchName} task ${task.id} missing metadata \u2014 loadTasks did not populate it`);
+  }
+  return md;
+}
+function selectRows(rows, opts, config, split) {
+  let tasks = rows.map((row) => rowToTask(row, config, split));
+  if (opts.ids) {
+    const want = new Set(opts.ids);
+    tasks = tasks.filter((task) => want.has(task.id));
+  } else if (opts.limit !== void 0) {
+    tasks = tasks.slice(0, opts.limit);
+  }
+  if (tasks.length === 0) throw new Error(`${config.name}: no tasks matched ${JSON.stringify(opts)}`);
+  return tasks;
+}
+async function loadFixtures(config, opts) {
+  const rows = JSON.parse(await readFile(config.fixturePath, "utf8"));
+  console.warn(`[${config.name}] ${config.fixturesEnv}=1 \u2014 loading ${rows.length} adapter fixtures`);
+  return selectRows(rows, opts, config, opts.split);
+}
+async function loadOfficialTasks(config, root, opts) {
+  const domain = benchDomain(config);
+  const script = `
+import json, sys
+from pathlib import Path
+root = Path(sys.argv[1])
+domain = sys.argv[2]
+split = sys.argv[3] or None
+sys.path.insert(0, str(root / "src"))
+from tau2.registry import registry
+loader = registry.get_tasks_loader(domain)
+tasks = loader(split)
+rows = []
+for task in tasks:
+    row = task.model_dump(mode="json")
+    row["domain"] = domain
+    rows.append(row)
+print(json.dumps(rows))
+`;
+  const stdout = await runVenvPython(script, [root, domain, opts.split ?? ""]);
+  return selectRows(JSON.parse(stdout), opts, config, opts.split);
+}
+async function scoreOfficialTrajectory(root, meta, artifactPath) {
+  const script = `
+import json, sys
+from pathlib import Path
+root = Path(sys.argv[1])
+task_id = sys.argv[2]
+artifact = Path(sys.argv[3])
+sys.path.insert(0, str(root / "src"))
+from tau2.data_model.simulation import Results
+from tau2.scripts.evaluate_trajectories import compute_simulation_rewards
+results = Results.load(artifact)
+updated = compute_simulation_rewards(results)
+scores = []
+for sim in updated.simulations:
+    if sim.task_id == task_id and sim.reward_info is not None:
+        scores.append(float(sim.reward_info.reward))
+if not scores:
+    raise SystemExit(f"no scored simulations for task_id={task_id} in {artifact}")
+print(json.dumps({"count": len(scores), "score": sum(scores) / len(scores), "scores": scores}))
+`;
+  const stdout = await runVenvPython(script, [root, meta.taskId, artifactPath], 0);
+  return JSON.parse(stdout.trim().split("\n").at(-1) ?? "{}");
+}
+function createTauBenchAdapter(config) {
+  const fixturesMode = process.env[config.fixturesEnv] === "1";
+  return {
+    name: config.name,
+    output: tauResultsOutput,
+    async preflight() {
+      if (fixturesMode) return;
+      const dir = benchDir(config);
+      if (!dir) {
+        throw new Error(`${config.dirEnv} is required. Fix: ${config.installHint}`);
+      }
+      await assertPath(`${dir}/src/tau2/registry.py`, "tau registry", config.name);
+      await loadOfficialTasks(config, dir, { limit: 1 });
+    },
+    async loadTasks(opts = {}) {
+      if (fixturesMode) return loadFixtures(config, opts);
+      const dir = benchDir(config);
+      if (!dir) throw new Error(`${config.dirEnv} is required to load official ${config.name} tasks`);
+      return loadOfficialTasks(config, dir, opts);
+    },
+    async goldArtifact() {
+      return void 0;
+    },
+    async judge(task, artifact) {
+      const dir = benchDir(config);
+      if (!dir) throw new Error(`${config.dirEnv} is required to judge ${config.name} trajectory artifacts`);
+      const meta = readMeta(task, config.name);
+      const artifactPath = resolve(artifact.trim());
+      await assertPath(artifactPath, "tau results/trajectory artifact", config.name);
+      const report = await scoreOfficialTrajectory(dir, meta, artifactPath);
+      const score = typeof report.score === "number" ? report.score : 0;
+      return {
+        resolved: score === 1,
+        score,
+        detail: JSON.stringify({ taskId: meta.taskId, domain: meta.domain, count: report.count })
+      };
+    }
+  };
+}
+
+export {
+  tauResultsOutput,
+  createTauBenchAdapter
+};
+//# sourceMappingURL=chunk-R36V2VP7.js.map

package/dist/chunk-R36V2VP7.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/benchmarks/tau-bench-shared.ts"],"sourcesContent":["/**\n * Shared tau-bench adapter spine.\n *\n * tau2 and tau3 live in the same upstream repository/package namespace today:\n * `sierra-research/tau2-bench`, Python package `tau2`. Keep one implementation\n * for task loading and reward recomputation so the domain/version adapters only\n * choose env names, default domain, and fixture file.\n */\n\nimport { readFile, stat } from 'node:fs/promises'\nimport { resolve } from 'node:path'\nimport type { OutputAdapter } from '@tangle-network/agent-runtime/loops'\nimport { runVenvPython } from './_harness'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\n\nexport interface TauBenchConfig {\n  name: string\n  fixturePath: string\n  fixturesEnv: string\n  dirEnv: string\n  domainEnv: string\n  defaultDomain: string\n  installHint: string\n  taskIntro: string\n}\n\ninterface TauRow {\n  id: string\n  domain: string\n  user_scenario?: unknown\n  description?: unknown\n  evaluation_criteria?: unknown\n}\n\ninterface TauMeta {\n  taskId: string\n  domain: string\n  split?: string\n  userScenario?: unknown\n  description?: unknown\n  evaluationCriteria?: unknown\n}\n\nexport const tauResultsOutput: 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|path|json)?\\s*\\n([\\s\\S]*?)```/g)]\n    return (fences.at(-1)?.[1] ?? text).trim()\n  },\n}\n\nasync function assertPath(path: string, label: string, benchName: string): Promise<void> {\n  try {\n    await stat(path)\n  } catch (err) {\n    throw new Error(`${benchName}: missing ${label} at ${path} (${err instanceof Error ? err.message : err})`)\n  }\n}\n\nfunction benchDir(config: TauBenchConfig): string | undefined {\n  return process.env[config.dirEnv]\n}\n\nfunction benchDomain(config: TauBenchConfig): string {\n  return process.env[config.domainEnv] ?? config.defaultDomain\n}\n\nfunction rowToTask(row: TauRow, config: TauBenchConfig, split?: string): BenchTask {\n  const meta: TauMeta = {\n    taskId: row.id,\n    domain: row.domain,\n    split,\n    userScenario: row.user_scenario,\n    description: row.description,\n    evaluationCriteria: row.evaluation_criteria,\n  }\n  return {\n    id: row.id,\n    split,\n    prompt: [\n      config.taskIntro,\n      `Run this task in the official ${row.domain} domain.`,\n      'The benchmark is a simulated multi-turn user/tool conversation.',\n      '',\n      typeof row.user_scenario === 'string' ? row.user_scenario : JSON.stringify(row.user_scenario ?? {}, null, 2),\n      '',\n      'Return the path to the official tau results.json or trajectory file containing this task run.',\n    ].join('\\n'),\n    metadata: meta as unknown as Record<string, unknown>,\n  }\n}\n\nfunction readMeta(task: BenchTask, benchName: string): TauMeta {\n  const md = task.metadata\n  if (!md || typeof md.taskId !== 'string' || typeof md.domain !== 'string') {\n    throw new Error(`${benchName} task ${task.id} missing metadata — loadTasks did not populate it`)\n  }\n  return md as unknown as TauMeta\n}\n\nfunction selectRows(rows: TauRow[], opts: LoadOptions, config: TauBenchConfig, split?: string): BenchTask[] {\n  let tasks = rows.map((row) => rowToTask(row, config, split))\n  if (opts.ids) {\n    const want = new Set(opts.ids)\n    tasks = tasks.filter((task) => want.has(task.id))\n  } else if (opts.limit !== undefined) {\n    tasks = tasks.slice(0, opts.limit)\n  }\n  if (tasks.length === 0) throw new Error(`${config.name}: no tasks matched ${JSON.stringify(opts)}`)\n  return tasks\n}\n\nasync function loadFixtures(config: TauBenchConfig, opts: LoadOptions): Promise<BenchTask[]> {\n  const rows = JSON.parse(await readFile(config.fixturePath, 'utf8')) as TauRow[]\n  console.warn(`[${config.name}] ${config.fixturesEnv}=1 — loading ${rows.length} adapter fixtures`)\n  return selectRows(rows, opts, config, opts.split)\n}\n\nasync function loadOfficialTasks(config: TauBenchConfig, root: string, opts: LoadOptions): Promise<BenchTask[]> {\n  const domain = benchDomain(config)\n  const script = `\nimport json, sys\nfrom pathlib import Path\nroot = Path(sys.argv[1])\ndomain = sys.argv[2]\nsplit = sys.argv[3] or None\nsys.path.insert(0, str(root / \"src\"))\nfrom tau2.registry import registry\nloader = registry.get_tasks_loader(domain)\ntasks = loader(split)\nrows = []\nfor task in tasks:\n    row = task.model_dump(mode=\"json\")\n    row[\"domain\"] = domain\n    rows.append(row)\nprint(json.dumps(rows))\n`\n  const stdout = await runVenvPython(script, [root, domain, opts.split ?? ''])\n  return selectRows(JSON.parse(stdout) as TauRow[], opts, config, opts.split)\n}\n\nasync function scoreOfficialTrajectory(root: string, meta: TauMeta, artifactPath: string): Promise<Record<string, unknown>> {\n  const script = `\nimport json, sys\nfrom pathlib import Path\nroot = Path(sys.argv[1])\ntask_id = sys.argv[2]\nartifact = Path(sys.argv[3])\nsys.path.insert(0, str(root / \"src\"))\nfrom tau2.data_model.simulation import Results\nfrom tau2.scripts.evaluate_trajectories import compute_simulation_rewards\nresults = Results.load(artifact)\nupdated = compute_simulation_rewards(results)\nscores = []\nfor sim in updated.simulations:\n    if sim.task_id == task_id and sim.reward_info is not None:\n        scores.append(float(sim.reward_info.reward))\nif not scores:\n    raise SystemExit(f\"no scored simulations for task_id={task_id} in {artifact}\")\nprint(json.dumps({\"count\": len(scores), \"score\": sum(scores) / len(scores), \"scores\": scores}))\n`\n  const stdout = await runVenvPython(script, [root, meta.taskId, artifactPath], 0)\n  return JSON.parse(stdout.trim().split('\\n').at(-1) ?? '{}') as Record<string, unknown>\n}\n\nexport function createTauBenchAdapter(config: TauBenchConfig): BenchmarkAdapter {\n  const fixturesMode = process.env[config.fixturesEnv] === '1'\n\n  return {\n    name: config.name,\n    output: tauResultsOutput,\n\n    async preflight() {\n      if (fixturesMode) return\n      const dir = benchDir(config)\n      if (!dir) {\n        throw new Error(`${config.dirEnv} is required. Fix: ${config.installHint}`)\n      }\n      await assertPath(`${dir}/src/tau2/registry.py`, 'tau registry', config.name)\n      await loadOfficialTasks(config, dir, { limit: 1 })\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      if (fixturesMode) return loadFixtures(config, opts)\n      const dir = benchDir(config)\n      if (!dir) throw new Error(`${config.dirEnv} is required to load official ${config.name} tasks`)\n      return loadOfficialTasks(config, dir, opts)\n    },\n\n    async goldArtifact() {\n      return undefined\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const dir = benchDir(config)\n      if (!dir) throw new Error(`${config.dirEnv} is required to judge ${config.name} trajectory artifacts`)\n      const meta = readMeta(task, config.name)\n      const artifactPath = resolve(artifact.trim())\n      await assertPath(artifactPath, 'tau results/trajectory artifact', config.name)\n      const report = await scoreOfficialTrajectory(dir, meta, artifactPath)\n      const score = typeof report.score === 'number' ? report.score : 0\n      return {\n        resolved: score === 1,\n        score,\n        detail: JSON.stringify({ taskId: meta.taskId, domain: meta.domain, count: report.count }),\n      }\n    },\n  }\n}\n"],"mappings":";;;;;AASA,SAAS,UAAU,YAAY;AAC/B,SAAS,eAAe;AAiCjB,IAAM,mBAA0C;AAAA,EACrD,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,2CAA2C,CAAC;AAC7E,YAAQ,OAAO,GAAG,EAAE,IAAI,CAAC,KAAK,MAAM,KAAK;AAAA,EAC3C;AACF;AAEA,eAAe,WAAW,MAAc,OAAe,WAAkC;AACvF,MAAI;AACF,UAAM,KAAK,IAAI;AAAA,EACjB,SAAS,KAAK;AACZ,UAAM,IAAI,MAAM,GAAG,SAAS,aAAa,KAAK,OAAO,IAAI,KAAK,eAAe,QAAQ,IAAI,UAAU,GAAG,GAAG;AAAA,EAC3G;AACF;AAEA,SAAS,SAAS,QAA4C;AAC5D,SAAO,QAAQ,IAAI,OAAO,MAAM;AAClC;AAEA,SAAS,YAAY,QAAgC;AACnD,SAAO,QAAQ,IAAI,OAAO,SAAS,KAAK,OAAO;AACjD;AAEA,SAAS,UAAU,KAAa,QAAwB,OAA2B;AACjF,QAAM,OAAgB;AAAA,IACpB,QAAQ,IAAI;AAAA,IACZ,QAAQ,IAAI;AAAA,IACZ;AAAA,IACA,cAAc,IAAI;AAAA,IAClB,aAAa,IAAI;AAAA,IACjB,oBAAoB,IAAI;AAAA,EAC1B;AACA,SAAO;AAAA,IACL,IAAI,IAAI;AAAA,IACR;AAAA,IACA,QAAQ;AAAA,MACN,OAAO;AAAA,MACP,iCAAiC,IAAI,MAAM;AAAA,MAC3C;AAAA,MACA;AAAA,MACA,OAAO,IAAI,kBAAkB,WAAW,IAAI,gBAAgB,KAAK,UAAU,IAAI,iBAAiB,CAAC,GAAG,MAAM,CAAC;AAAA,MAC3G;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AAAA,IACX,UAAU;AAAA,EACZ;AACF;AAEA,SAAS,SAAS,MAAiB,WAA4B;AAC7D,QAAM,KAAK,KAAK;AAChB,MAAI,CAAC,MAAM,OAAO,GAAG,WAAW,YAAY,OAAO,GAAG,WAAW,UAAU;AACzE,UAAM,IAAI,MAAM,GAAG,SAAS,SAAS,KAAK,EAAE,wDAAmD;AAAA,EACjG;AACA,SAAO;AACT;AAEA,SAAS,WAAW,MAAgB,MAAmB,QAAwB,OAA6B;AAC1G,MAAI,QAAQ,KAAK,IAAI,CAAC,QAAQ,UAAU,KAAK,QAAQ,KAAK,CAAC;AAC3D,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,GAAG,OAAO,IAAI,sBAAsB,KAAK,UAAU,IAAI,CAAC,EAAE;AAClG,SAAO;AACT;AAEA,eAAe,aAAa,QAAwB,MAAyC;AAC3F,QAAM,OAAO,KAAK,MAAM,MAAM,SAAS,OAAO,aAAa,MAAM,CAAC;AAClE,UAAQ,KAAK,IAAI,OAAO,IAAI,KAAK,OAAO,WAAW,qBAAgB,KAAK,MAAM,mBAAmB;AACjG,SAAO,WAAW,MAAM,MAAM,QAAQ,KAAK,KAAK;AAClD;AAEA,eAAe,kBAAkB,QAAwB,MAAc,MAAyC;AAC9G,QAAM,SAAS,YAAY,MAAM;AACjC,QAAM,SAAS;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAiBf,QAAM,SAAS,MAAM,cAAc,QAAQ,CAAC,MAAM,QAAQ,KAAK,SAAS,EAAE,CAAC;AAC3E,SAAO,WAAW,KAAK,MAAM,MAAM,GAAe,MAAM,QAAQ,KAAK,KAAK;AAC5E;AAEA,eAAe,wBAAwB,MAAc,MAAe,cAAwD;AAC1H,QAAM,SAAS;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAmBf,QAAM,SAAS,MAAM,cAAc,QAAQ,CAAC,MAAM,KAAK,QAAQ,YAAY,GAAG,CAAC;AAC/E,SAAO,KAAK,MAAM,OAAO,KAAK,EAAE,MAAM,IAAI,EAAE,GAAG,EAAE,KAAK,IAAI;AAC5D;AAEO,SAAS,sBAAsB,QAA0C;AAC9E,QAAM,eAAe,QAAQ,IAAI,OAAO,WAAW,MAAM;AAEzD,SAAO;AAAA,IACL,MAAM,OAAO;AAAA,IACb,QAAQ;AAAA,IAER,MAAM,YAAY;AAChB,UAAI,aAAc;AAClB,YAAM,MAAM,SAAS,MAAM;AAC3B,UAAI,CAAC,KAAK;AACR,cAAM,IAAI,MAAM,GAAG,OAAO,MAAM,sBAAsB,OAAO,WAAW,EAAE;AAAA,MAC5E;AACA,YAAM,WAAW,GAAG,GAAG,yBAAyB,gBAAgB,OAAO,IAAI;AAC3E,YAAM,kBAAkB,QAAQ,KAAK,EAAE,OAAO,EAAE,CAAC;AAAA,IACnD;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;AACtC,UAAI,aAAc,QAAO,aAAa,QAAQ,IAAI;AAClD,YAAM,MAAM,SAAS,MAAM;AAC3B,UAAI,CAAC,IAAK,OAAM,IAAI,MAAM,GAAG,OAAO,MAAM,iCAAiC,OAAO,IAAI,QAAQ;AAC9F,aAAO,kBAAkB,QAAQ,KAAK,IAAI;AAAA,IAC5C;AAAA,IAEA,MAAM,eAAe;AACnB,aAAO;AAAA,IACT;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,MAAM,SAAS,MAAM;AAC3B,UAAI,CAAC,IAAK,OAAM,IAAI,MAAM,GAAG,OAAO,MAAM,yBAAyB,OAAO,IAAI,uBAAuB;AACrG,YAAM,OAAO,SAAS,MAAM,OAAO,IAAI;AACvC,YAAM,eAAe,QAAQ,SAAS,KAAK,CAAC;AAC5C,YAAM,WAAW,cAAc,mCAAmC,OAAO,IAAI;AAC7E,YAAM,SAAS,MAAM,wBAAwB,KAAK,MAAM,YAAY;AACpE,YAAM,QAAQ,OAAO,OAAO,UAAU,WAAW,OAAO,QAAQ;AAChE,aAAO;AAAA,QACL,UAAU,UAAU;AAAA,QACpB;AAAA,QACA,QAAQ,KAAK,UAAU,EAAE,QAAQ,KAAK,QAAQ,QAAQ,KAAK,QAAQ,OAAO,OAAO,MAAM,CAAC;AAAA,MAC1F;AAAA,IACF;AAAA,EACF;AACF;","names":[]}